page fault scalability patch V11 [0/7]: overview

page fault scalability patch V11 [0/7]: overview

[Christoph Lameter]

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Changes from V10->V11 of this patch:
- cmpxchg_i386: Optimize code generated after feedback from Linus. Various
  fixes.
- drop make_rss_atomic in favor of rss_sloppy
- generic: adapt to new changes in Linus tree, some fixes to fallback
  functions. Add generic ptep_xchg_flush based on xchg.
- S390: remove use of page_table_lock from ptep_xchg_flush (deadlock)
- x86_64: remove ptep_xchg
- i386: integrated Nick Piggin's changes for PAE mode. Create ptep_xchg_flush and
  various fixes.
- ia64: if necessary flush icache before ptep_cmpxchg. Remove ptep_xchg

This is a series of patches that increases the scalability of
the page fault handler for SMP. Here are some performance results
on a machine with 32 processors allocating 32 GB with an increasing
number of cpus.

Without the patches:
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32  10    1    3.966s    366.840s 370.082s 56556.456  56553.456
 32  10    2    3.604s    319.004s 172.058s 65006.086 121511.453
 32  10    4    3.705s    341.550s 106.007s 60741.936 197704.486
 32  10    8    3.597s    809.711s 119.021s 25785.427 175917.674
 32  10   16    5.886s   2238.122s 163.084s  9345.560 127998.973
 32  10   32   21.748s   5458.983s 201.062s  3826.409 104011.521

With the patches:
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32  10    1    3.772s    330.629s 334.042s 62713.587  62708.706
 32  10    2    3.767s    352.252s 185.077s 58905.502 112886.222
 32  10    4    3.549s    255.683s  77.000s 80898.177 272326.496
 32  10    8    3.522s    263.879s  52.030s 78427.083 400965.857
 32  10   16    5.193s    384.813s  42.076s 53772.158 490378.852
 32  10   32   15.806s    996.890s  54.077s 20708.587 382879.208

With a high number of CPUs the page fault rate improves more than
twofold and may reach 500000 faults/sec betweenr 16-512 cpus. The
fault rate drops if a process is running on all processors as also
here for the 32 cpu case.

Note that the measurements were done on a NUMA system and this
test uses off node memory. Variations may exist due to allocations in
memory areas in diverse distances to the local cpu. The slight drop
for 2 cpus is probably due to that effect.

The performance increase is accomplished by avoiding the use of the
page_table_lock spinlock (but not mm->mmap_sem!) through new atomic
operations on pte's (ptep_xchg, ptep_cmpxchg) and on pmd and pgd's
(pgd_test_and_populate, pmd_test_and_populate).

The page table lock can be avoided in the following situations:

1. An empty pte or pmd entry is populated

This is safe since the swapper may only depopulate them and the
swapper code has been changed to never set a pte to be empty until the
page has been evicted. The population of an empty pte is frequent
if a process touches newly allocated memory.

2. Modifications of flags in a pte entry (write/accessed).

These modifications are done by the CPU or by low level handlers
on various platforms also bypassing the page_table_lock. So this
seems to be safe too.

One essential change in the VM is the use of pte_cmpxchg (or its generic
emulation) on page table entries before doing an update_mmu_change without holding
the page table lock. However, we do similar things now with other atomic pte operations
such as ptep_get_and_clear and ptep_test_and_clear_dirty. These operations clear
a pte *after* doing an operation on it. The ptep_cmpxchg as used in this patch
operates on an *cleared* pte and replaces it with a pte pointing to valid memory.
The effect of this change on various architectures has to be thought through. Local
definitions of ptep_cmpxchg and ptep_xchg may be necessary.

For IA64 an icache coherency issue may arise that potentially requires the
flushing of the icache (as done via update_mmu_cache on IA64) prior
to the use of ptep_cmpxchg. Similar issues may arise on other platforms.

The patch uses sloppy rss handling. mm->rss is incremented without
proper locking because locking would introduce too much overhead. Rss
is not essential for vm operations (3 uses of rss in rmap.c were not necessary and
were removed). The difference in rss values has been found to be less than 1% in
our tests (see also the separate email to linux-mm and linux-ia64 on the subject
of "sloppy rss"). The move away from using atomic operations for rss in earlier versions
of this patch also increases the performance of the page fault handler in the single
thread case over an unpatched kernel.

Note that I have posted two other approaches of dealing with the rss problem:

A. make_rss_atomic. The earlier releases contained that patch but then another
   variable (such as anon_rss) was introduced that would have required additional
   atomic operations. Atomic rss operations are also causing slowdowns on
   machines with a high number of cpus due to memory contention.

B. remove_rss. Replace rss with a periodic scan over the vm to determine
   rss and additional numbers. This was also discussed on linux-mm and linux-ia64.
   The scans while displaying /proc data were undesirable.

The patchset is composed of 7 patches:

1/7: Sloppy rss

   Removes mm->rss usage from mm/rmap.c and insures that negative rss values
   are not displayed.

2/7: Avoid page_table_lock in handle_mm_fault

   This patch defers the acquisition of the page_table_lock as much as
   possible and uses atomic operations for allocating anonymous memory.
   These atomic operations are simulated by acquiring the page_table_lock
   for very small time frames if an architecture does not define
   __HAVE_ARCH_ATOMIC_TABLE_OPS. It also changes the swapper so that a
   pte will not be set to empty if a page is in transition to swap.

   If only the first two patches are applied then the time that the page_table_lock
   is held is simply reduced. The lock may then be acquired multiple
   times during a page fault.

   The remaining patches introduce the necessary atomic pte operations to avoid
   the page_table_lock.

3/7: Atomic pte operations for ia64

4/7: Make cmpxchg generally available on i386

   The atomic operations on the page table rely heavily on cmpxchg instructions.
   This patch adds emulations for cmpxchg and cmpxchg8b for old 80386 and 80486
   cpus. The emulations are only included if a kernel is build for these old
   cpus and are skipped for the real cmpxchg instructions if the kernel
   that is build for 386 or 486 is then run on a more recent cpu.

   This patch may be used independently of the other patches.

5/7: Atomic pte operations for i386

   A generally available cmpxchg (last patch) must be available for this patch to
   preserve the ability to build kernels for 386 and 486.

6/7: Atomic pte operation for x86_64

7/7: Atomic pte operations for s390

page fault scalability patch V11 [1/7]: sloppy rss

[Christoph Lameter]

Changelog
	* Enable the sloppy use of mm->rss and mm->anon_rss atomic without locking
	* Insure that negative rss values are not given out by the /proc filesystem
	* remove 3 checks of rss in mm/rmap.c
	* Prerequisite for page table scalability patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-11-15 11:13:39.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-11-18 13:04:30.000000000 -0800
@@ -216,7 +216,7 @@
 	atomic_t mm_count;			/* How many references to "struct mm_struct" (users count as 1) */
 	int map_count;				/* number of VMAs */
 	struct rw_semaphore mmap_sem;
-	spinlock_t page_table_lock;		/* Protects page tables, mm->rss, mm->anon_rss */
+	spinlock_t page_table_lock;		/* Protects page tables */

 	struct list_head mmlist;		/* List of maybe swapped mm's.  These are globally strung
 						 * together off init_mm.mmlist, and are protected
@@ -252,6 +252,21 @@
 	struct kioctx		default_kioctx;
 };

+/*
+ * rss and anon_rss are incremented and decremented in some locations without
+ * proper locking. This function insures that these values do not become negative.
+ */
+static long inline get_rss(struct mm_struct *mm)
+{
+	long rss = mm->rss;
+
+	if (rss < 0)
+		 mm->rss = rss = 0;
+	if ((long)mm->anon_rss < 0)
+		mm->anon_rss = 0;
+	return rss;
+}
+
 struct sighand_struct {
 	atomic_t		count;
 	struct k_sigaction	action[_NSIG];
Index: linux-2.6.9/fs/proc/task_mmu.c
===================================================================
--- linux-2.6.9.orig/fs/proc/task_mmu.c	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/fs/proc/task_mmu.c	2004-11-18 12:56:26.000000000 -0800
@@ -22,7 +22,7 @@
 		"VmPTE:\t%8lu kB\n",
 		(mm->total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),
-		mm->rss << (PAGE_SHIFT-10),
+		get_rss(mm) << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
@@ -37,7 +37,9 @@
 int task_statm(struct mm_struct *mm, int *shared, int *text,
 	       int *data, int *resident)
 {
-	*shared = mm->rss - mm->anon_rss;
+	*shared = get_rss(mm) - mm->anon_rss;
+	if (*shared <0)
+		*shared = 0;
 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
 								>> PAGE_SHIFT;
 	*data = mm->total_vm - mm->shared_vm;
Index: linux-2.6.9/fs/proc/array.c
===================================================================
--- linux-2.6.9.orig/fs/proc/array.c	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/fs/proc/array.c	2004-11-18 12:53:16.000000000 -0800
@@ -420,7 +420,7 @@
 		jiffies_to_clock_t(task->it_real_value),
 		start_time,
 		vsize,
-		mm ? mm->rss : 0, /* you might want to shift this left 3 */
+		mm ? get_rss(mm) : 0, /* you might want to shift this left 3 */
 	        rsslim,
 		mm ? mm->start_code : 0,
 		mm ? mm->end_code : 0,
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-11-15 11:13:40.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-11-18 12:26:45.000000000 -0800
@@ -263,8 +263,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -504,8 +502,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -788,8 +784,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;

page fault scalability patch V11 [2/7]: page fault handler optimizations

[Christoph Lameter]

Changelog
        * Increase parallelism in SMP configurations by deferring
          the acquisition of page_table_lock in handle_mm_fault
        * Anonymous memory page faults bypass the page_table_lock
          through the use of atomic page table operations
        * Swapper does not set pte to empty in transition to swap
        * Simulate atomic page table operations using the
          page_table_lock if an arch does not define
          __HAVE_ARCH_ATOMIC_TABLE_OPS. This still provides
          a performance benefit since the page_table_lock
          is held for shorter periods of time.

Signed-off-by: Christoph Lameter <clameter@sgi.com

Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-11-18 12:25:49.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-11-19 06:38:53.000000000 -0800
@@ -1330,8 +1330,7 @@
 }

 /*
- * We hold the mm semaphore and the page_table_lock on entry and
- * should release the pagetable lock on exit..
+ * We hold the mm semaphore
  */
 static int do_swap_page(struct mm_struct * mm,
 	struct vm_area_struct * vma, unsigned long address,
@@ -1343,15 +1342,13 @@
 	int ret = VM_FAULT_MINOR;

 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);
 	page = lookup_swap_cache(entry);
 	if (!page) {
  		swapin_readahead(entry, address, vma);
  		page = read_swap_cache_async(entry, vma, address);
 		if (!page) {
 			/*
-			 * Back out if somebody else faulted in this pte while
-			 * we released the page table lock.
+			 * Back out if somebody else faulted in this pte
 			 */
 			spin_lock(&mm->page_table_lock);
 			page_table = pte_offset_map(pmd, address);
@@ -1374,8 +1371,7 @@
 	lock_page(page);

 	/*
-	 * Back out if somebody else faulted in this pte while we
-	 * released the page table lock.
+	 * Back out if somebody else faulted in this pte
 	 */
 	spin_lock(&mm->page_table_lock);
 	page_table = pte_offset_map(pmd, address);
@@ -1422,14 +1418,12 @@
 }

 /*
- * We are called with the MM semaphore and page_table_lock
- * spinlock held to protect against concurrent faults in
- * multithreaded programs.
+ * We are called with the MM semaphore held.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 		pte_t *page_table, pmd_t *pmd, int write_access,
-		unsigned long addr)
+		unsigned long addr, pte_t orig_entry)
 {
 	pte_t entry;
 	struct page * page = ZERO_PAGE(addr);
@@ -1441,7 +1435,6 @@
 	if (write_access) {
 		/* Allocate our own private page. */
 		pte_unmap(page_table);
-		spin_unlock(&mm->page_table_lock);

 		if (unlikely(anon_vma_prepare(vma)))
 			goto no_mem;
@@ -1450,30 +1443,37 @@
 			goto no_mem;
 		clear_user_highpage(page, addr);

-		spin_lock(&mm->page_table_lock);
 		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
-			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
-		}
-		mm->rss++;
 		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
 							 vma->vm_page_prot)),
 				      vma);
-		lru_cache_add_active(page);
 		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}

-	set_pte(page_table, entry);
+	/* update the entry */
+	if (!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)) {
+		if (write_access) {
+			pte_unmap(page_table);
+			page_cache_release(page);
+		}
+		goto out;
+	}
+	if (write_access) {
+		/*
+		 * These two functions must come after the cmpxchg
+		 * because if the page is on the LRU then try_to_unmap may come
+		 * in and unmap the pte.
+		 */
+		lru_cache_add_active(page);
+		page_add_anon_rmap(page, vma, addr);
+		mm->rss++;
+
+	}
 	pte_unmap(page_table);

 	/* No need to invalidate - it was non-present before */
 	update_mmu_cache(vma, addr, entry);
-	spin_unlock(&mm->page_table_lock);
 out:
 	return VM_FAULT_MINOR;
 no_mem:
@@ -1489,12 +1489,12 @@
  * As this is called only for pages that do not currently exist, we
  * do not need to flush old virtual caches or the TLB.
  *
- * This is called with the MM semaphore held and the page table
- * spinlock held. Exit with the spinlock released.
+ * This is called with the MM semaphore held.
  */
 static int
 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-	unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *page_table,
+        pmd_t *pmd, pte_t orig_entry)
 {
 	struct page * new_page;
 	struct address_space *mapping = NULL;
@@ -1505,9 +1505,8 @@

 	if (!vma->vm_ops || !vma->vm_ops->nopage)
 		return do_anonymous_page(mm, vma, page_table,
-					pmd, write_access, address);
+					pmd, write_access, address, orig_entry);
 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);

 	if (vma->vm_file) {
 		mapping = vma->vm_file->f_mapping;
@@ -1605,7 +1604,7 @@
  * nonlinear vmas.
  */
 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
-	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd, pte_t entry)
 {
 	unsigned long pgoff;
 	int err;
@@ -1618,13 +1617,12 @@
 	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
-		return do_no_page(mm, vma, address, write_access, pte, pmd);
+		return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 	}

 	pgoff = pte_to_pgoff(*pte);

 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);

 	err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
 	if (err == -ENOMEM)
@@ -1643,49 +1641,40 @@
  * with external mmu caches can use to update those (ie the Sparc or
  * PowerPC hashed page tables that act as extended TLBs).
  *
- * Note the "page_table_lock". It is to protect against kswapd removing
- * pages from under us. Note that kswapd only ever _removes_ pages, never
- * adds them. As such, once we have noticed that the page is not present,
- * we can drop the lock early.
- *
- * The adding of pages is protected by the MM semaphore (which we hold),
- * so we don't need to worry about a page being suddenly been added into
- * our VM.
- *
- * We enter with the pagetable spinlock held, we are supposed to
- * release it when done.
+ * Note that kswapd only ever _removes_ pages, never adds them.
+ * We need to insure to handle that case properly.
  */
 static inline int handle_pte_fault(struct mm_struct *mm,
 	struct vm_area_struct * vma, unsigned long address,
 	int write_access, pte_t *pte, pmd_t *pmd)
 {
 	pte_t entry;
+	pte_t new_entry;

 	entry = *pte;
 	if (!pte_present(entry)) {
-		/*
-		 * If it truly wasn't present, we know that kswapd
-		 * and the PTE updates will not touch it later. So
-		 * drop the lock.
-		 */
 		if (pte_none(entry))
-			return do_no_page(mm, vma, address, write_access, pte, pmd);
+			return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 		if (pte_file(entry))
-			return do_file_page(mm, vma, address, write_access, pte, pmd);
+			return do_file_page(mm, vma, address, write_access, pte, pmd, entry);
 		return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
 	}

+	/*
+	 * This is the case in which we only update some bits in the pte.
+	 */
+	new_entry = pte_mkyoung(entry);
 	if (write_access) {
-		if (!pte_write(entry))
+		if (!pte_write(entry)) {
+			/* do_wp_page expects us to hold the page_table_lock */
+			spin_lock(&mm->page_table_lock);
 			return do_wp_page(mm, vma, address, pte, pmd, entry);
-
-		entry = pte_mkdirty(entry);
+		}
+		new_entry = pte_mkdirty(new_entry);
 	}
-	entry = pte_mkyoung(entry);
-	ptep_set_access_flags(vma, address, pte, entry, write_access);
-	update_mmu_cache(vma, address, entry);
+	if (ptep_cmpxchg(vma, address, pte, entry, new_entry))
+		update_mmu_cache(vma, address, new_entry);
 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_MINOR;
 }

@@ -1703,22 +1692,45 @@

 	inc_page_state(pgfault);

-	if (is_vm_hugetlb_page(vma))
+	if (unlikely(is_vm_hugetlb_page(vma)))
 		return VM_FAULT_SIGBUS;	/* mapping truncation does this. */

 	/*
-	 * We need the page table lock to synchronize with kswapd
-	 * and the SMP-safe atomic PTE updates.
+	 * We rely on the mmap_sem and the SMP-safe atomic PTE updates.
+	 * to synchronize with kswapd
 	 */
-	spin_lock(&mm->page_table_lock);
-	pmd = pmd_alloc(mm, pgd, address);
+	if (unlikely(pgd_none(*pgd))) {
+		pmd_t *new = pmd_alloc_one(mm, address);
+		if (!new)
+			return VM_FAULT_OOM;
+
+		/* Insure that the update is done in an atomic way */
+		if (!pgd_test_and_populate(mm, pgd, new))
+			pmd_free(new);
+	}
+
+	pmd = pmd_offset(pgd, address);
+
+	if (likely(pmd)) {
+		pte_t *pte;
+
+		if (!pmd_present(*pmd)) {
+			struct page *new;

-	if (pmd) {
-		pte_t * pte = pte_alloc_map(mm, pmd, address);
-		if (pte)
+			new = pte_alloc_one(mm, address);
+			if (!new)
+				return VM_FAULT_OOM;
+
+			if (!pmd_test_and_populate(mm, pmd, new))
+				pte_free(new);
+			else
+				inc_page_state(nr_page_table_pages);
+		}
+
+		pte = pte_offset_map(pmd, address);
+		if (likely(pte))
 			return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
 	}
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_OOM;
 }

Index: linux-2.6.9/include/asm-generic/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-generic/pgtable.h	2004-10-18 14:53:46.000000000 -0700
+++ linux-2.6.9/include/asm-generic/pgtable.h	2004-11-19 07:54:05.000000000 -0800
@@ -134,4 +134,60 @@
 #define pgd_offset_gate(mm, addr)	pgd_offset(mm, addr)
 #endif

+#ifndef __HAVE_ARCH_ATOMIC_TABLE_OPS
+/*
+ * If atomic page table operations are not available then use
+ * the page_table_lock to insure some form of locking.
+ * Note thought that low level operations as well as the
+ * page_table_handling of the cpu may bypass all locking.
+ */
+
+#ifndef __HAVE_ARCH_PTEP_CMPXCHG
+#define ptep_cmpxchg(__vma, __addr, __ptep, __oldval, __newval)		\
+({									\
+	int __rc;							\
+	spin_lock(&__vma->vm_mm->page_table_lock);			\
+	__rc = pte_same(*(__ptep), __oldval);				\
+	if (__rc) set_pte(__ptep, __newval);				\
+	spin_unlock(&__vma->vm_mm->page_table_lock);			\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_ARCH_PGP_TEST_AND_POPULATE
+#define pgd_test_and_populate(__mm, __pgd, __pmd)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pgd_present(*(__pgd));					\
+	if (__rc) pgd_populate(__mm, __pgd, __pmd);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_PMD_TEST_AND_POPULATE
+#define pmd_test_and_populate(__mm, __pmd, __page)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pmd_present(*(__pmd));					\
+	if (__rc) pmd_populate(__mm, __pmd, __page);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#endif
+
+#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH
+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)		\
+({									\
+	pte_t __p = __pte(xchg(&pte_val(*(__ptep)), pte_val(__pteval)));\
+	flush_tlb_page(__vma, __address);				\
+	__p;								\
+})
+
+#endif
+
 #endif /* _ASM_GENERIC_PGTABLE_H */
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-11-19 06:38:51.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-11-19 06:38:53.000000000 -0800
@@ -419,7 +419,10 @@
  * @vma:	the vm area in which the mapping is added
  * @address:	the user virtual address mapped
  *
- * The caller needs to hold the mm->page_table_lock.
+ * The caller needs to hold the mm->page_table_lock if page
+ * is pointing to something that is known by the vm.
+ * The lock does not need to be held if page is pointing
+ * to a newly allocated page.
  */
 void page_add_anon_rmap(struct page *page,
 	struct vm_area_struct *vma, unsigned long address)
@@ -561,11 +564,6 @@

 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address);
-	pteval = ptep_clear_flush(vma, address, pte);
-
-	/* Move the dirty bit to the physical page now the pte is gone. */
-	if (pte_dirty(pteval))
-		set_page_dirty(page);

 	if (PageAnon(page)) {
 		swp_entry_t entry = { .val = page->private };
@@ -580,11 +578,15 @@
 			list_add(&mm->mmlist, &init_mm.mmlist);
 			spin_unlock(&mmlist_lock);
 		}
-		set_pte(pte, swp_entry_to_pte(entry));
+		pteval = ptep_xchg_flush(vma, address, pte, swp_entry_to_pte(entry));
 		BUG_ON(pte_file(*pte));
 		mm->anon_rss--;
-	}
+	} else
+		pteval = ptep_clear_flush(vma, address, pte);

+	/* Move the dirty bit to the physical page now the pte is gone. */
+	if (pte_dirty(pteval))
+		set_page_dirty(page);
 	mm->rss--;
 	page_remove_rmap(page);
 	page_cache_release(page);
@@ -671,15 +673,21 @@
 		if (ptep_clear_flush_young(vma, address, pte))
 			continue;

-		/* Nuke the page table entry. */
 		flush_cache_page(vma, address);
-		pteval = ptep_clear_flush(vma, address, pte);
+		/*
+		 * There would be a race here with handle_mm_fault and do_anonymous_page
+		 * which  bypasses the page_table_lock if we would zap the pte before
+		 * putting something into it. On the other hand we need to
+		 * have the dirty flag setting at the time we replaced the value.
+		 */

 		/* If nonlinear, store the file page offset in the pte. */
 		if (page->index != linear_page_index(vma, address))
-			set_pte(pte, pgoff_to_pte(page->index));
+			pteval = ptep_xchg_flush(vma, address, pte, pgoff_to_pte(page->index));
+		else
+			pteval = ptep_get_and_clear(pte);

-		/* Move the dirty bit to the physical page now the pte is gone. */
+		/* Move the dirty bit to the physical page now that the pte is gone. */
 		if (pte_dirty(pteval))
 			set_page_dirty(page);

page fault scalability patch V11 [3/7]: ia64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for ia64
        * Enhanced parallelism in page fault handler if applied together
          with the generic patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-ia64/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-ia64/pgalloc.h	2004-10-18 14:53:06.000000000 -0700
+++ linux-2.6.9/include/asm-ia64/pgalloc.h	2004-11-19 07:54:19.000000000 -0800
@@ -34,6 +34,10 @@
 #define pmd_quicklist		(local_cpu_data->pmd_quick)
 #define pgtable_cache_size	(local_cpu_data->pgtable_cache_sz)

+/* Empty entries of PMD and PGD */
+#define PMD_NONE       0
+#define PGD_NONE       0
+
 static inline pgd_t*
 pgd_alloc_one_fast (struct mm_struct *mm)
 {
@@ -78,12 +82,19 @@
 	preempt_enable();
 }

+
 static inline void
 pgd_populate (struct mm_struct *mm, pgd_t *pgd_entry, pmd_t *pmd)
 {
 	pgd_val(*pgd_entry) = __pa(pmd);
 }

+/* Atomic populate */
+static inline int
+pgd_test_and_populate (struct mm_struct *mm, pgd_t *pgd_entry, pmd_t *pmd)
+{
+	return ia64_cmpxchg8_acq(pgd_entry,__pa(pmd), PGD_NONE) == PGD_NONE;
+}

 static inline pmd_t*
 pmd_alloc_one_fast (struct mm_struct *mm, unsigned long addr)
@@ -132,6 +143,13 @@
 	pmd_val(*pmd_entry) = page_to_phys(pte);
 }

+/* Atomic populate */
+static inline int
+pmd_test_and_populate (struct mm_struct *mm, pmd_t *pmd_entry, struct page *pte)
+{
+	return ia64_cmpxchg8_acq(pmd_entry, page_to_phys(pte), PMD_NONE) == PMD_NONE;
+}
+
 static inline void
 pmd_populate_kernel (struct mm_struct *mm, pmd_t *pmd_entry, pte_t *pte)
 {
Index: linux-2.6.9/include/asm-ia64/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-ia64/pgtable.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-ia64/pgtable.h	2004-11-19 07:55:35.000000000 -0800
@@ -414,6 +425,26 @@
 #endif
 }

+/*
+ * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
+ * information.  However, we use this routine to take care of any (delayed) i-cache
+ * flushing that may be necessary.
+ */
+extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
+
+static inline int
+ptep_cmpxchg (struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	/*
+	 * IA64 defers icache flushes. If the new pte is executable we may
+	 * have to flush the icache to insure cache coherency immediately
+	 * after the cmpxchg.
+	 */
+	if (pte_exec(newval))
+		update_mmu_cache(vma, addr, newval);
+	return ia64_cmpxchg8_acq(&ptep->pte, newval.pte, oldval.pte) == oldval.pte;
+}
+
 static inline int
 pte_same (pte_t a, pte_t b)
 {
@@ -476,13 +507,6 @@
 	struct vm_area_struct * prev, unsigned long start, unsigned long end);
 #endif

-/*
- * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
- * information.  However, we use this routine to take care of any (delayed) i-cache
- * flushing that may be necessary.
- */
-extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
-
 #define __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Update PTEP with ENTRY, which is guaranteed to be a less
@@ -560,6 +584,8 @@
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
 #define __HAVE_ARCH_PGD_OFFSET_GATE
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+#define __HAVE_ARCH_LOCK_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _ASM_IA64_PGTABLE_H */

page fault scalability patch V11 [4/7]: universal cmpxchg for i386

[Christoph Lameter]

Changelog
        * Make cmpxchg and cmpxchg8b generally available on the i386
	  platform.
        * Provide emulation of cmpxchg suitable for uniprocessor if
	  build and run on 386.
        * Provide emulation of cmpxchg8b suitable for uniprocessor systems
	  if build and run on 386 or 486.
	* Provide an inline function to atomically get a 64 bit value via
	  cmpxchg8b in an SMP system (courtesy of Nick Piggin)
	  (important for i386 PAE mode and other places where atomic 64 bit
	  operations are useful)

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/arch/i386/Kconfig
===================================================================
--- linux-2.6.9.orig/arch/i386/Kconfig	2004-11-15 11:13:34.000000000 -0800
+++ linux-2.6.9/arch/i386/Kconfig	2004-11-19 10:02:54.000000000 -0800
@@ -351,6 +351,11 @@
 	depends on !M386
 	default y

+config X86_CMPXCHG8B
+	bool
+	depends on !M386 && !M486
+	default y
+
 config X86_XADD
 	bool
 	depends on !M386
Index: linux-2.6.9/arch/i386/kernel/cpu/intel.c
===================================================================
--- linux-2.6.9.orig/arch/i386/kernel/cpu/intel.c	2004-11-15 11:13:34.000000000 -0800
+++ linux-2.6.9/arch/i386/kernel/cpu/intel.c	2004-11-19 10:38:26.000000000 -0800
@@ -6,6 +6,7 @@
 #include <linux/bitops.h>
 #include <linux/smp.h>
 #include <linux/thread_info.h>
+#include <linux/module.h>

 #include <asm/processor.h>
 #include <asm/msr.h>
@@ -287,5 +288,103 @@
 	return 0;
 }

+#ifndef CONFIG_X86_CMPXCHG
+unsigned long cmpxchg_386_u8(volatile void *ptr, u8 old, u8 new)
+{
+	u8 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u8));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u8 *)ptr;
+	if (prev == old)
+		*(u8 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u8);
+
+unsigned long cmpxchg_386_u16(volatile void *ptr, u16 old, u16 new)
+{
+	u16 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u16));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u16 *)ptr;
+	if (prev == old)
+		*(u16 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u16);
+
+unsigned long cmpxchg_386_u32(volatile void *ptr, u32 old, u32 new)
+{
+	u32 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u32));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u32 *)ptr;
+	if (prev == old)
+		*(u32 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u32);
+#endif
+
+#ifndef CONFIG_X86_CMPXCHG8B
+unsigned long long cmpxchg8b_486(volatile unsigned long long *ptr,
+	       unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	unsigned long flags;
+
+	/*
+	 * Check if the kernel was compiled for an old cpu but
+	 * we are running really on a cpu capable of cmpxchg8b
+	 */
+
+	if (cpu_has(cpu_data, X86_FEATURE_CX8))
+		return __cmpxchg8b(ptr, old, newv);
+
+	/* Poor mans cmpxchg8b for 386 and 486. Not suitable for SMP */
+	local_irq_save(flags);
+	prev = *ptr;
+	if (prev == old)
+		*ptr = newv;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg8b_486);
+#endif
+
 // arch_initcall(intel_cpu_init);

Index: linux-2.6.9/include/asm-i386/system.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/system.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-i386/system.h	2004-11-19 10:49:46.000000000 -0800
@@ -149,6 +149,9 @@
 #define __xg(x) ((struct __xchg_dummy *)(x))


+#define ll_low(x)	*(((unsigned int*)&(x))+0)
+#define ll_high(x)	*(((unsigned int*)&(x))+1)
+
 /*
  * The semantics of XCHGCMP8B are a bit strange, this is why
  * there is a loop and the loading of %%eax and %%edx has to
@@ -184,8 +187,6 @@
 {
 	__set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
 }
-#define ll_low(x)	*(((unsigned int*)&(x))+0)
-#define ll_high(x)	*(((unsigned int*)&(x))+1)

 static inline void __set_64bit_var (unsigned long long *ptr,
 			 unsigned long long value)
@@ -203,6 +204,26 @@
  __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
  __set_64bit(ptr, ll_low(value), ll_high(value)) )

+static inline unsigned long long __get_64bit(unsigned long long * ptr)
+{
+	unsigned long long ret;
+	__asm__ __volatile__ (
+		"\n1:\t"
+		"movl (%1), %%eax\n\t"
+		"movl 4(%1), %%edx\n\t"
+		"movl %%eax, %%ebx\n\t"
+		"movl %%edx, %%ecx\n\t"
+		LOCK_PREFIX "cmpxchg8b (%1)\n\t"
+		"jnz 1b"
+		:	"=A"(ret)
+		:	"D"(ptr)
+		:	"ebx", "ecx", "memory");
+	return ret;
+}
+
+#define get_64bit(ptr) __get_64bit(ptr)
+
+
 /*
  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
  * Note 2: xchg has side effect, so that attribute volatile is necessary,
@@ -240,7 +261,41 @@
  */

 #ifdef CONFIG_X86_CMPXCHG
+
 #define __HAVE_ARCH_CMPXCHG 1
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))__cmpxchg((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
+
+#else
+
+/*
+ * Building a kernel capable running on 80386. It may be necessary to
+ * simulate the cmpxchg on the 80386 CPU. For that purpose we define
+ * a function for each of the sizes we support.
+ */
+
+extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
+extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
+extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
+
+static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
+				      unsigned long new, int size)
+{
+	switch (size) {
+	case 1:
+		return cmpxchg_386_u8(ptr, old, new);
+	case 2:
+		return cmpxchg_386_u16(ptr, old, new);
+	case 4:
+		return cmpxchg_386_u32(ptr, old, new);
+	}
+	return old;
+}
+
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))cmpxchg_386((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
 #endif

 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
@@ -270,10 +325,32 @@
 	return old;
 }

-#define cmpxchg(ptr,o,n)\
-	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
-					(unsigned long)(n),sizeof(*(ptr))))
-
+static inline unsigned long long __cmpxchg8b(volatile unsigned long long *ptr,
+		unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	__asm__ __volatile__(
+	LOCK_PREFIX "cmpxchg8b (%4)"
+		: "=A" (prev)
+		: "0" (old), "c" ((unsigned long)(newv >> 32)),
+		  "b" ((unsigned long)(newv & 0xffffffffULL)), "D" (ptr)
+		: "memory");
+	return prev;
+}
+
+#ifdef CONFIG_X86_CMPXCHG8B
+#define cmpxchg8b __cmpxchg8b
+#else
+/*
+ * Building a kernel capable of running on 80486 and 80386. Both
+ * do not support cmpxchg8b. Call a function that emulates the
+ * instruction if necessary.
+ */
+extern unsigned long long cmpxchg8b_486(volatile unsigned long long *,
+		unsigned long long, unsigned long long);
+#define cmpxchg8b cmpxchg8b_486
+#endif
+
 #ifdef __KERNEL__
 struct alt_instr {
 	__u8 *instr; 		/* original instruction */

page fault scalability patch V11 [5/7]: i386 atomic pte operations

[Christoph Lameter]

Changelog
	* Atomic pte operations for i386 in regular and PAE modes

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-i386/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-i386/pgtable.h	2004-11-19 10:05:27.000000000 -0800
@@ -413,6 +413,7 @@
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _I386_PGTABLE_H */
Index: linux-2.6.9/include/asm-i386/pgtable-3level.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable-3level.h	2004-10-18 14:54:55.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgtable-3level.h	2004-11-19 10:10:06.000000000 -0800
@@ -6,7 +6,8 @@
  * tables on PPro+ CPUs.
  *
  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
- */
+ * August 26, 2004 added ptep_cmpxchg <christoph@lameter.com>
+*/

 #define pte_ERROR(e) \
 	printk("%s:%d: bad pte %p(%08lx%08lx).\n", __FILE__, __LINE__, &(e), (e).pte_high, (e).pte_low)
@@ -42,26 +43,15 @@
 	return pte_x(pte);
 }

-/* Rules for using set_pte: the pte being assigned *must* be
- * either not present or in a state where the hardware will
- * not attempt to update the pte.  In places where this is
- * not possible, use pte_get_and_clear to obtain the old pte
- * value and then use set_pte to update it.  -ben
- */
-static inline void set_pte(pte_t *ptep, pte_t pte)
-{
-	ptep->pte_high = pte.pte_high;
-	smp_wmb();
-	ptep->pte_low = pte.pte_low;
-}
-#define __HAVE_ARCH_SET_PTE_ATOMIC
-#define set_pte_atomic(pteptr,pteval) \
+#define set_pte(pteptr,pteval) \
 		set_64bit((unsigned long long *)(pteptr),pte_val(pteval))
 #define set_pmd(pmdptr,pmdval) \
 		set_64bit((unsigned long long *)(pmdptr),pmd_val(pmdval))
 #define set_pgd(pgdptr,pgdval) \
 		set_64bit((unsigned long long *)(pgdptr),pgd_val(pgdval))

+#define set_pte_atomic set_pte
+
 /*
  * Pentium-II erratum A13: in PAE mode we explicitly have to flush
  * the TLB via cr3 if the top-level pgd is changed...
@@ -142,4 +132,23 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t){ (pte).pte_high })
 #define __swp_entry_to_pte(x)		((pte_t){ 0, (x).val })

+/* Atomic PTE operations */
+#define ptep_xchg_flush(__vma, __addr, __ptep, __newval) \
+({	pte_t __r;							\
+	/* xchg acts as a barrier before the setting of the high bits. */\
+	__r.pte_low = xchg(&(__ptep)->pte_low, (__newval).pte_low);	\
+	__r.pte_high = (__ptep)->pte_high;				\
+	(__ptep)->pte_high = (__newval).pte_high;			\
+	flush_tlb_page(__vma, __addr);					\
+	(__r);								\
+})
+
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
+
+static inline int ptep_cmpxchg(struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg((unsigned int *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
+
 #endif /* _I386_PGTABLE_3LEVEL_H */
Index: linux-2.6.9/include/asm-i386/pgtable-2level.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable-2level.h	2004-10-18 14:54:31.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgtable-2level.h	2004-11-19 10:05:27.000000000 -0800
@@ -82,4 +82,7 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t) { (pte).pte_low })
 #define __swp_entry_to_pte(x)		((pte_t) { (x).val })

+/* Atomic PTE operations */
+#define ptep_cmpxchg(__vma,__a,__xp,__oldpte,__newpte) (cmpxchg(&(__xp)->pte_low, (__oldpte).pte_low, (__newpte).pte_low)==(__oldpte).pte_low)
+
 #endif /* _I386_PGTABLE_2LEVEL_H */
Index: linux-2.6.9/include/asm-i386/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgalloc.h	2004-10-18 14:53:10.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgalloc.h	2004-11-19 10:10:40.000000000 -0800
@@ -4,9 +4,12 @@
 #include <linux/config.h>
 #include <asm/processor.h>
 #include <asm/fixmap.h>
+#include <asm/system.h>
 #include <linux/threads.h>
 #include <linux/mm.h>		/* for struct page */

+#define PMD_NONE 0L
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte)))

@@ -16,6 +19,19 @@
 		((unsigned long long)page_to_pfn(pte) <<
 			(unsigned long long) PAGE_SHIFT)));
 }
+
+/* Atomic version */
+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+#ifdef CONFIG_X86_PAE
+	return cmpxchg8b( ((unsigned long long *)pmd), PMD_NONE, _PAGE_TABLE +
+		((unsigned long long)page_to_pfn(pte) <<
+			(unsigned long long) PAGE_SHIFT) ) == PMD_NONE;
+#else
+	return cmpxchg( (unsigned long *)pmd, PMD_NONE, _PAGE_TABLE + (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+#endif
+}
+
 /*
  * Allocate and free page tables.
  */
@@ -49,6 +65,7 @@
 #define pmd_free(x)			do { } while (0)
 #define __pmd_free_tlb(tlb,x)		do { } while (0)
 #define pgd_populate(mm, pmd, pte)	BUG()
+#define pgd_test_and_populate(mm, pmd, pte)	({ BUG(); 1; })

 #define check_pgt_cache()	do { } while (0)

page fault scalability patch V11 [6/7]: x86_64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for x86_64

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-x86_64/pgalloc.h	2004-10-18 14:54:30.000000000 -0700
+++ linux-2.6.9/include/asm-x86_64/pgalloc.h	2004-11-19 08:17:55.000000000 -0800
@@ -7,16 +7,26 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pgd_populate(mm, pgd, pmd) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pmd)))
+#define pgd_test_and_populate(mm, pgd, pmd) \
+		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pmd)) == PGD_NONE)

 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
 	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
 }

+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+}
+
 extern __inline__ pmd_t *get_pmd(void)
 {
 	return (pmd_t *)get_zeroed_page(GFP_KERNEL);
Index: linux-2.6.9/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-x86_64/pgtable.h	2004-11-15 11:13:39.000000000 -0800
+++ linux-2.6.9/include/asm-x86_64/pgtable.h	2004-11-19 08:18:52.000000000 -0800
@@ -437,6 +437,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

page fault scalability patch V11 [7/7]: s390 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for s390

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-s390/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-s390/pgtable.h	2004-10-18 14:54:55.000000000 -0700
+++ linux-2.6.9/include/asm-s390/pgtable.h	2004-11-19 11:35:08.000000000 -0800
@@ -567,6 +567,15 @@
 	return pte;
 }

+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)            \
+({                                                                     \
+	struct mm_struct *__mm = __vma->vm_mm;                          \
+	pte_t __pte;                                                    \
+	__pte = ptep_clear_flush(__vma, __address, __ptep);             \
+	set_pte(__ptep, __pteval);                                      \
+	__pte;                                                          \
+})
+
 static inline void ptep_set_wrprotect(pte_t *ptep)
 {
 	pte_t old_pte = *ptep;
@@ -778,6 +787,14 @@

 #define kern_addr_valid(addr)   (1)

+/* Atomic PTE operations */
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
+static inline int ptep_cmpxchg (struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg(ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
 /*
  * No page table caches to initialise
  */
@@ -791,6 +808,7 @@
 #define __HAVE_ARCH_PTEP_CLEAR_DIRTY_FLUSH
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR
 #define __HAVE_ARCH_PTEP_CLEAR_FLUSH
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
Index: linux-2.6.9/include/asm-s390/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-s390/pgalloc.h	2004-10-18 14:54:37.000000000 -0700
+++ linux-2.6.9/include/asm-s390/pgalloc.h	2004-11-19 11:33:25.000000000 -0800
@@ -97,6 +97,10 @@
 	pgd_val(*pgd) = _PGD_ENTRY | __pa(pmd);
 }

+static inline int pgd_test_and_populate(struct mm_struct *mm, pdg_t *pgd, pmd_t *pmd)
+{
+	return cmpxchg(pgd, _PAGE_TABLE_INV, _PGD_ENTRY | __pa(pmd)) == _PAGE_TABLE_INV;
+}
 #endif /* __s390x__ */

 static inline void
@@ -119,6 +123,18 @@
 	pmd_populate_kernel(mm, pmd, (pte_t *)((page-mem_map) << PAGE_SHIFT));
 }

+static inline int
+pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *page)
+{
+	int rc;
+	spin_lock(&mm->page_table_lock);
+
+	rc=pte_same(*pmd, _PAGE_INVALID_EMPTY);
+	if (rc) pmd_populate(mm, pmd, page);
+	spin_unlock(&mm->page_table_lock);
+	return rc;
+}
+
 /*
  * page table entry allocation/free routines.
  */

Re: page fault scalability patch V11 [0/7]: overview

[Linus Torvalds]

On Fri, 19 Nov 2004, Christoph Lameter wrote:
> 
> Note that I have posted two other approaches of dealing with the rss problem:

You could also make "rss" be a _signed_ integer per-thread.

When unmapping a page, you decrement one of the threads that shares the mm 
(doesn't matter which - which is why the per-thread rss may go negative), 
and when mapping a page you increment it.

Then, anybody who actually wants a global rss can just iterate over
threads and add it all up. If you do it under the mmap_sem, it's stable,
and if you do it outside the mmap_sem it's imprecise but stable in the
long term (ie errors never _accumulate_, like the non-atomic case will 
do).

Does anybody care enough? Maybe, maybe not. It certainly sounds a hell of 
a lot better than the periodic scan.

		Linus

Re: page fault scalability patch V11 [1/7]: sloppy rss

[Hugh Dickins]

Sorry, against what tree do these patches apply?
Apparently not linux-2.6.9, nor latest -bk, nor -mm?

Hugh

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

Linus Torvalds wrote:
> 
> On Fri, 19 Nov 2004, Christoph Lameter wrote:
> 
>>Note that I have posted two other approaches of dealing with the rss problem:
> 
> 
> You could also make "rss" be a _signed_ integer per-thread.
> 
> When unmapping a page, you decrement one of the threads that shares the mm 
> (doesn't matter which - which is why the per-thread rss may go negative), 
> and when mapping a page you increment it.
> 
> Then, anybody who actually wants a global rss can just iterate over
> threads and add it all up. If you do it under the mmap_sem, it's stable,
> and if you do it outside the mmap_sem it's imprecise but stable in the
> long term (ie errors never _accumulate_, like the non-atomic case will 
> do).
> 
> Does anybody care enough? Maybe, maybe not. It certainly sounds a hell of 
> a lot better than the periodic scan.
> 

I think this sounds like it might be a good idea. I prefer it to having
the unbounded error of sloppy rss (as improbable as it may be in practice).

The per thread rss may wrap (maybe not 64-bit counters), but even so,
the summation over all threads should still end up being correct I
think.

Re: page fault scalability patch V11 [0/7]: overview

[Christoph Lameter]

On Sat, 20 Nov 2004, Nick Piggin wrote:

> I think this sounds like it might be a good idea. I prefer it to having
> the unbounded error of sloppy rss (as improbable as it may be in practice).

It may also be faster since the processors can have exclusive cache lines.

This means we need to move rss into the task struct. But how does one get
from mm struct to task struct? current is likely available most of
the time. Is that always the case?

> The per thread rss may wrap (maybe not 64-bit counters), but even so,
> the summation over all threads should still end up being correct I
> think.

Note though that the mmap_sem is no protection. It is a read lock and may
be held by multiple processes while incrementing and decrementing rss.
This is likely reducing the number of collisions significantly but it wont
be a  guarantee like locking or atomic ops.

Re: page fault scalability patch V11 [1/7]: sloppy rss

[Christoph Lameter]

2.6.10-rc2-bk3

On Fri, 19 Nov 2004, Hugh Dickins wrote:

> Sorry, against what tree do these patches apply?
> Apparently not linux-2.6.9, nor latest -bk, nor -mm?

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

Christoph Lameter wrote:
> On Sat, 20 Nov 2004, Nick Piggin wrote:
> 
> 
>>I think this sounds like it might be a good idea. I prefer it to having
>>the unbounded error of sloppy rss (as improbable as it may be in practice).
> 
> 
> It may also be faster since the processors can have exclusive cache lines.
> 

Yep.

> This means we need to move rss into the task struct. But how does one get
> from mm struct to task struct? current is likely available most of
> the time. Is that always the case?
> 

It is available everywhere that mm_struct is, I guess. So yes, I
think `current` should be OK.

> 
>>The per thread rss may wrap (maybe not 64-bit counters), but even so,
>>the summation over all threads should still end up being correct I
>>think.
> 
> 
> Note though that the mmap_sem is no protection. It is a read lock and may
> be held by multiple processes while incrementing and decrementing rss.
> This is likely reducing the number of collisions significantly but it wont
> be a  guarantee like locking or atomic ops.
> 

Yeah the read lock won't do anything to serialise it. I think what Linus
is saying is that we _don't care_ most of the time (because the error will
be bounded). But if it happened that we really do care anywhere, then the
write lock should be sufficient.

Re: page fault scalability patch V11 [0/7]: overview

[Linus Torvalds]

On Sat, 20 Nov 2004, Nick Piggin wrote:
> 
> The per thread rss may wrap (maybe not 64-bit counters), but even so,
> the summation over all threads should still end up being correct I
> think.

Yes. As long as the total rss fits in an int, it doesn't matter if any of
them wrap. Addition is still associative in twos-complement arithmetic 
even in the presense of overflows. 

If you actually want to make it proper standard C, I guess you'd have to 
make the thing unsigned, which gives you the mod-2**n guarantees even if 
somebody were to ever make a non-twos-complement machine.

		Linus

Re: page fault scalability patch V11 [0/7]: overview

[Linus Torvalds]

On Fri, 19 Nov 2004, Christoph Lameter wrote:
> 
> Note though that the mmap_sem is no protection. It is a read lock and may
> be held by multiple processes while incrementing and decrementing rss.
> This is likely reducing the number of collisions significantly but it wont
> be a  guarantee like locking or atomic ops.

It is, though, if you hold it for a write.

The point being that you _can_ get an exact rss value if you want to.

Not that I really see any overwhelming evidence of anybody ever really 
caring, but it's nice to know that you have the option.

		Linus

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Fri, Nov 19, 2004 at 11:59:03AM -0800, Linus Torvalds wrote:
> You could also make "rss" be a _signed_ integer per-thread.
> When unmapping a page, you decrement one of the threads that shares the mm 
> (doesn't matter which - which is why the per-thread rss may go negative), 
> and when mapping a page you increment it.
> Then, anybody who actually wants a global rss can just iterate over
> threads and add it all up. If you do it under the mmap_sem, it's stable,
> and if you do it outside the mmap_sem it's imprecise but stable in the
> long term (ie errors never _accumulate_, like the non-atomic case will 
> do).
> Does anybody care enough? Maybe, maybe not. It certainly sounds a hell of 
> a lot better than the periodic scan.

Unprivileged triggers for full-tasklist scans are NMI oops material.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Fri, Nov 19, 2004 at 11:42:39AM -0800, Christoph Lameter wrote:
> A. make_rss_atomic. The earlier releases contained that patch but
> then another variable (such as anon_rss) was introduced that would
>    have required additional atomic operations. Atomic rss operations
>    are also causing slowdowns on machines with a high number of cpus
>    due to memory contention.
> B. remove_rss. Replace rss with a periodic scan over the vm to
>    determine rss and additional numbers. This was also discussed on
>    linux-mm and linux-ia64. The scans while displaying /proc data
>    were undesirable.

Split counters easily resolve the issues with both these approaches
(and apparently your co-workers are suggesting it too, and have
performance results backing it).


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Robin Holt]

On Fri, Nov 19, 2004 at 11:42:39AM -0800, Christoph Lameter wrote:
> Note that I have posted two other approaches of dealing with the rss problem:
> 
> A. make_rss_atomic. The earlier releases contained that patch but then another
>    variable (such as anon_rss) was introduced that would have required additional
>    atomic operations. Atomic rss operations are also causing slowdowns on
>    machines with a high number of cpus due to memory contention.
> 
> B. remove_rss. Replace rss with a periodic scan over the vm to determine
>    rss and additional numbers. This was also discussed on linux-mm and linux-ia64.
>    The scans while displaying /proc data were undesirable.

Can you run a comparison benchmark between atomic rss and anon_rss and
the sloppy rss with the rss and anon_rss in seperate cachelines.  I am not
sure that it is important to seperate the two into seperate lines, just
rss and anon_rss from the lock and sema.

If I have the time over the weekend, I might try this myself.  If not, can
you give it a try.

Thanks,
Robin

Re: page fault scalability patch V11 [0/7]: overview

[Linus Torvalds]

On Fri, 19 Nov 2004, Linus Torvalds wrote:
> 
> Not that I really see any overwhelming evidence of anybody ever really 
> caring, but it's nice to know that you have the option.

Btw, if you are going to look at doing this rss thing, you need to make 
sure that thread exit ends up adding its rss to _some_ remaining sibling. 

I guess that was obvious, but it's worth pointing out. That may actually
be the only case where we do _not_ have a nice SMP-safe access: we do have
a stable sibling (tsk->thread_leader), but we don't have any good
serialization _except_ for taking mmap_sem for writing. Which we currently
don't do: we take it for reading (and then we possibly upgrade it to a
write lock if we notice that there is a core-dump starting).

We can avoid this too by having a per-mm atomic rss "spill" counter. So 
exit_mm() would basically do:

	...
	tsk->mm = NULL;
	atomic_add(tsk->rss, &mm->rss_spill);
	...

and then the algorithm for getting rss would be:

	rss = atomic_read(mm->rss_spill);
	for_each_thread(..)
		rss += tsk->rss;

Or does anybody see any better approaches?

		Linus

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> On Fri, Nov 19, 2004 at 11:42:39AM -0800, Christoph Lameter wrote:
> 
>>A. make_rss_atomic. The earlier releases contained that patch but
>>then another variable (such as anon_rss) was introduced that would
>>   have required additional atomic operations. Atomic rss operations
>>   are also causing slowdowns on machines with a high number of cpus
>>   due to memory contention.
>>B. remove_rss. Replace rss with a periodic scan over the vm to
>>   determine rss and additional numbers. This was also discussed on
>>   linux-mm and linux-ia64. The scans while displaying /proc data
>>   were undesirable.
> 
> 
> Split counters easily resolve the issues with both these approaches
> (and apparently your co-workers are suggesting it too, and have
> performance results backing it).
> 

Split counters still require atomic operations though. This is what
Christoph's latest effort is directed at removing. And they'll still
bounce cachelines around. (I assume we've reached the conclusion
that per-cpu split counters per-mm won't fly?).

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> On Fri, Nov 19, 2004 at 11:59:03AM -0800, Linus Torvalds wrote:
> 
>>You could also make "rss" be a _signed_ integer per-thread.
>>When unmapping a page, you decrement one of the threads that shares the mm 
>>(doesn't matter which - which is why the per-thread rss may go negative), 
>>and when mapping a page you increment it.
>>Then, anybody who actually wants a global rss can just iterate over
>>threads and add it all up. If you do it under the mmap_sem, it's stable,
>>and if you do it outside the mmap_sem it's imprecise but stable in the
>>long term (ie errors never _accumulate_, like the non-atomic case will 
>>do).
>>Does anybody care enough? Maybe, maybe not. It certainly sounds a hell of 
>>a lot better than the periodic scan.
> 
> 
> Unprivileged triggers for full-tasklist scans are NMI oops material.
> 

What about pushing the per-thread rss delta back into the global atomic
rss counter in each schedule()?

Pros:
This would take the task exiting problem into its stride as a matter of
course.

Single atomic read to get rss.

Cons:
would just be moving the atomic op somewhere else if we don't get
many page faults per schedule.

Not really nice dependancies.

Assumes schedule (not context switch) must occur somewhat regularly.
At present this is not true for SCHED_FIFO tasks.


Too nasty?

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Split counters easily resolve the issues with both these approaches
>> (and apparently your co-workers are suggesting it too, and have
>> performance results backing it).

On Sat, Nov 20, 2004 at 01:18:22PM +1100, Nick Piggin wrote:
> Split counters still require atomic operations though. This is what
> Christoph's latest effort is directed at removing. And they'll still
> bounce cachelines around. (I assume we've reached the conclusion
> that per-cpu split counters per-mm won't fly?).

Split != per-cpu, though it may be. Counterexamples are
as simple as atomic_inc(&mm->rss[smp_processor_id()>>RSS_IDX_SHIFT]);
Furthermore, see Robin Holt's results regarding the performance of the
atomic operations and their relation to cacheline sharing.

And frankly, the argument that the space overhead of per-cpu counters
is problematic is not compelling. Even at 1024 cpus it's smaller than
an ia64 pagetable page, of which there are numerous instances attached
to each mm.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>Split counters easily resolve the issues with both these approaches
>>>(and apparently your co-workers are suggesting it too, and have
>>>performance results backing it).
> 
> 
> On Sat, Nov 20, 2004 at 01:18:22PM +1100, Nick Piggin wrote:
> 
>>Split counters still require atomic operations though. This is what
>>Christoph's latest effort is directed at removing. And they'll still
>>bounce cachelines around. (I assume we've reached the conclusion
>>that per-cpu split counters per-mm won't fly?).
> 
> 
> Split != per-cpu, though it may be. Counterexamples are
> as simple as atomic_inc(&mm->rss[smp_processor_id()>>RSS_IDX_SHIFT]);

Oh yes, I just meant that the only way split counters will relieve
the atomic ops and bouncing is by having them per-cpu. But you knew
that :)

> Furthermore, see Robin Holt's results regarding the performance of the
> atomic operations and their relation to cacheline sharing.
> 

Well yeah, but a. their patch isn't in 2.6 (or 2.4), and b. anon_rss
means another atomic op. While this doesn't immediately make it a
showstopper, it is gradually slowing down the single threaded page
fault path too, which is bad.

> And frankly, the argument that the space overhead of per-cpu counters
> is problematic is not compelling. Even at 1024 cpus it's smaller than
> an ia64 pagetable page, of which there are numerous instances attached
> to each mm.
> 

1024 CPUs * 64 byte cachelines == 64K, no? Well I'm sure they probably
don't even care about 64K on their large machines, but...

On i386 this would be maybe 32 * 128 byte == 4K per task for distro
kernels. Not so good.

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Unprivileged triggers for full-tasklist scans are NMI oops material.

On Sat, Nov 20, 2004 at 01:25:37PM +1100, Nick Piggin wrote:
> What about pushing the per-thread rss delta back into the global atomic
> rss counter in each schedule()?
> Pros:
> This would take the task exiting problem into its stride as a matter of
> course.
> Single atomic read to get rss.
> Cons:
> would just be moving the atomic op somewhere else if we don't get
> many page faults per schedule.
> Not really nice dependancies.
> Assumes schedule (not context switch) must occur somewhat regularly.
> At present this is not true for SCHED_FIFO tasks.
> Too nasty?

This doesn't sound too hot. There's enough accounting that can't be
done anywhere but schedule(), and this can be done elsewhere. Plus,
you're moving an already too-frequent operation to a more frequent
callsite.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>Unprivileged triggers for full-tasklist scans are NMI oops material.
> 
> 
> On Sat, Nov 20, 2004 at 01:25:37PM +1100, Nick Piggin wrote:
> 
>>What about pushing the per-thread rss delta back into the global atomic
>>rss counter in each schedule()?
>>Pros:
>>This would take the task exiting problem into its stride as a matter of
>>course.
>>Single atomic read to get rss.
>>Cons:
>>would just be moving the atomic op somewhere else if we don't get
>>many page faults per schedule.
>>Not really nice dependancies.
>>Assumes schedule (not context switch) must occur somewhat regularly.
>>At present this is not true for SCHED_FIFO tasks.
>>Too nasty?
> 
> 
> This doesn't sound too hot. There's enough accounting that can't be
> done anywhere but schedule(), and this can be done elsewhere. Plus,
> you're moving an already too-frequent operation to a more frequent
> callsite.
> 

No, it won't somehow increase the number of atomic rss operations
just because schedule is called more often. The number of ops will
be at _most_ the number of page faults.

But I agree with your overall evaluation of its 'hotness'. Just
another idea. Give this monkey another thousand years at the keys
and he'll come up with the perfect solution :P

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Furthermore, see Robin Holt's results regarding the performance of the
>> atomic operations and their relation to cacheline sharing.

On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> Well yeah, but a. their patch isn't in 2.6 (or 2.4), and b. anon_rss

Irrelevant. Unshare cachelines with hot mm-global ones, and the
"problem" goes away.

This stuff is going on and on about some purist "no atomic operations
anywhere" weirdness even though killing the last atomic operation
creates problems and doesn't improve performance.


On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> means another atomic op. While this doesn't immediately make it a
> showstopper, it is gradually slowing down the single threaded page
> fault path too, which is bad.

William Lee Irwin III wrote:
>> And frankly, the argument that the space overhead of per-cpu counters
>> is problematic is not compelling. Even at 1024 cpus it's smaller than
>> an ia64 pagetable page, of which there are numerous instances attached
>> to each mm.

On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> 1024 CPUs * 64 byte cachelines == 64K, no? Well I'm sure they probably
> don't even care about 64K on their large machines, but...
> On i386 this would be maybe 32 * 128 byte == 4K per task for distro
> kernels. Not so good.

Why the Hell would you bother giving each cpu a separate cacheline?
The odds of bouncing significantly merely amongst the counters are not
particularly high.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>Furthermore, see Robin Holt's results regarding the performance of the
>>>atomic operations and their relation to cacheline sharing.
> 
> 
> On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> 
>>Well yeah, but a. their patch isn't in 2.6 (or 2.4), and b. anon_rss
> 
> 
> Irrelevant. Unshare cachelines with hot mm-global ones, and the
> "problem" goes away.
> 

That's the idea.

> This stuff is going on and on about some purist "no atomic operations
> anywhere" weirdness even though killing the last atomic operation
> creates problems and doesn't improve performance.
> 

Huh? How is not wanting to impact single threaded performance being
"purist weirdness"? Practical, I'd call it.

> 
> On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> 
>>means another atomic op. While this doesn't immediately make it a
>>showstopper, it is gradually slowing down the single threaded page
>>fault path too, which is bad.
> 
> 
> William Lee Irwin III wrote:
> 
>>>And frankly, the argument that the space overhead of per-cpu counters
>>>is problematic is not compelling. Even at 1024 cpus it's smaller than
>>>an ia64 pagetable page, of which there are numerous instances attached
>>>to each mm.
> 
> 
> On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> 
>>1024 CPUs * 64 byte cachelines == 64K, no? Well I'm sure they probably
>>don't even care about 64K on their large machines, but...
>>On i386 this would be maybe 32 * 128 byte == 4K per task for distro
>>kernels. Not so good.
> 
> 
> Why the Hell would you bother giving each cpu a separate cacheline?
> The odds of bouncing significantly merely amongst the counters are not
> particularly high.
> 

Hmm yeah I guess wouldn't put them all on different cachelines.
As you can see though, Christoph ran into a wall at 8 CPUs, so
having them densly packed still might not be enough.

Re: page fault scalability patch V11 [0/7]: overview

[Robin Holt]

On Fri, Nov 19, 2004 at 07:04:25PM -0800, William Lee Irwin III wrote:
> Why the Hell would you bother giving each cpu a separate cacheline?
> The odds of bouncing significantly merely amongst the counters are not
> particularly high.

Agree, we are currently using atomic ops on a global rss on our 2.4
kernel with 512cpu systems and not seeing much cacheline contention.
I don't remember how little it ended up being, but it was very little.
We had gone to dropping the page_table_lock and only reaquiring it if
the pte was non-null when we went to insert our new one.  I think that
was how we had it working.  I would have to wake up and actually look
at that code as it was many months ago that Ray Bryant did that work.
We did make rss atomic.  Most of the contention is sorted out by the
mmap_sem.  Processes acquiring themselves off of mmap_sem were found
to have spaced themselves out enough that they were all approximately
equal time from doing their atomic_add and therefore had very little
contention for the cacheline.  At least it was not enough that we could
measure it as significant.

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> On Fri, Nov 19, 2004 at 11:59:03AM -0800, Linus Torvalds wrote:
> 
>>You could also make "rss" be a _signed_ integer per-thread.
>>When unmapping a page, you decrement one of the threads that shares the mm 
>>(doesn't matter which - which is why the per-thread rss may go negative), 
>>and when mapping a page you increment it.
>>Then, anybody who actually wants a global rss can just iterate over
>>threads and add it all up. If you do it under the mmap_sem, it's stable,
>>and if you do it outside the mmap_sem it's imprecise but stable in the
>>long term (ie errors never _accumulate_, like the non-atomic case will 
>>do).
>>Does anybody care enough? Maybe, maybe not. It certainly sounds a hell of 
>>a lot better than the periodic scan.
> 
> 
> Unprivileged triggers for full-tasklist scans are NMI oops material.
> 

Hang on, let's come back to this...

We already have unprivileged do-for-each-thread triggers in the proc
code. It's in do_task_stat, even. Rss reporting would basically just
involve one extra addition within that loop.

So... hmm, I can't see a problem with it.

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Irrelevant. Unshare cachelines with hot mm-global ones, and the
>> "problem" goes away.

On Sat, Nov 20, 2004 at 02:14:33PM +1100, Nick Piggin wrote:
> That's the idea.


William Lee Irwin III wrote:
>> This stuff is going on and on about some purist "no atomic operations
>> anywhere" weirdness even though killing the last atomic operation
>> creates problems and doesn't improve performance.

On Sat, Nov 20, 2004 at 02:14:33PM +1100, Nick Piggin wrote:
> Huh? How is not wanting to impact single threaded performance being
> "purist weirdness"? Practical, I'd call it.

Empirically demonstrate the impact on single-threaded performance.


On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
>> Why the Hell would you bother giving each cpu a separate cacheline?
>> The odds of bouncing significantly merely amongst the counters are not
>> particularly high.

On Sat, Nov 20, 2004 at 02:14:33PM +1100, Nick Piggin wrote:
> Hmm yeah I guess wouldn't put them all on different cachelines.
> As you can see though, Christoph ran into a wall at 8 CPUs, so
> having them densly packed still might not be enough.

Please be more specific about the result, and cite the Message-Id.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Unprivileged triggers for full-tasklist scans are NMI oops material.

On Sat, Nov 20, 2004 at 02:37:04PM +1100, Nick Piggin wrote:
> Hang on, let's come back to this...
> We already have unprivileged do-for-each-thread triggers in the proc
> code. It's in do_task_stat, even. Rss reporting would basically just
> involve one extra addition within that loop.
> So... hmm, I can't see a problem with it.

/proc/ triggering NMI oopses was a persistent problem even before that
code was merged. I've not bothered testing it as it at best aggravates it.

And thread groups can share mm's. do_for_each_thread() won't suffice.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>Irrelevant. Unshare cachelines with hot mm-global ones, and the
>>>"problem" goes away.
> 
> 
> On Sat, Nov 20, 2004 at 02:14:33PM +1100, Nick Piggin wrote:
> 
>>That's the idea.
> 
> 
> 
> William Lee Irwin III wrote:
> 
>>>This stuff is going on and on about some purist "no atomic operations
>>>anywhere" weirdness even though killing the last atomic operation
>>>creates problems and doesn't improve performance.
> 
> 
> On Sat, Nov 20, 2004 at 02:14:33PM +1100, Nick Piggin wrote:
> 
>>Huh? How is not wanting to impact single threaded performance being
>>"purist weirdness"? Practical, I'd call it.
> 
> 
> Empirically demonstrate the impact on single-threaded performance.
> 

I can tell you its worse. I don't have to demonstrate anything, more
atomic RMW ops in the page fault path is going to have an impact.

I'm not saying we must not compromise *anywhere*, but it would
just be nice to try to avoid making the path heavier, that's all.
I'm not being purist when I say I'd first rather explore all other
options before adding atomics.

But nevermind arguing, it appears Linus' suggested method will
be fine and *does* mean we don't have to compromise.

> 
> On Sat, Nov 20, 2004 at 01:40:40PM +1100, Nick Piggin wrote:
> 
>>>Why the Hell would you bother giving each cpu a separate cacheline?
>>>The odds of bouncing significantly merely amongst the counters are not
>>>particularly high.
> 
> 
> On Sat, Nov 20, 2004 at 02:14:33PM +1100, Nick Piggin wrote:
> 
>>Hmm yeah I guess wouldn't put them all on different cachelines.
>>As you can see though, Christoph ran into a wall at 8 CPUs, so
>>having them densly packed still might not be enough.
> 
> 
> Please be more specific about the result, and cite the Message-Id.
> 

Start of this thread.

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Please be more specific about the result, and cite the Message-Id.

On Sat, Nov 20, 2004 at 02:58:36PM +1100, Nick Piggin wrote:
> Start of this thread.

Those do not have testing results of different RSS counter
implementations in isolation.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>Unprivileged triggers for full-tasklist scans are NMI oops material.
> 
> 
> On Sat, Nov 20, 2004 at 02:37:04PM +1100, Nick Piggin wrote:
> 
>>Hang on, let's come back to this...
>>We already have unprivileged do-for-each-thread triggers in the proc
>>code. It's in do_task_stat, even. Rss reporting would basically just
>>involve one extra addition within that loop.
>>So... hmm, I can't see a problem with it.
> 
> 
> /proc/ triggering NMI oopses was a persistent problem even before that
> code was merged. I've not bothered testing it as it at best aggravates it.
> 

It isn't a problem. If it ever became a problem then we can just
touch the nmi oopser in the loop.

> And thread groups can share mm's. do_for_each_thread() won't suffice.
> 

I think it will be just fine.

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

Nick Piggin wrote:
> William Lee Irwin III wrote:

>> And thread groups can share mm's. do_for_each_thread() won't suffice.
>>
> 
> I think it will be just fine.
> 

Sorry, I misread. I think having per-thread rss counters will be
fine (regardless of whether or not do_for_each_thread itself will
suffice).

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> /proc/ triggering NMI oopses was a persistent problem even before that
>> code was merged. I've not bothered testing it as it at best aggravates it.

On Sat, Nov 20, 2004 at 03:03:17PM +1100, Nick Piggin wrote:
> It isn't a problem. If it ever became a problem then we can just
> touch the nmi oopser in the loop.

Very, very wrong. The tasklist scans hold the read side of the lock
and aren't even what's running with interrupts off. The contenders
on the write side are what the NMI oopser oopses.

And supposing the arch reenables interrupts in the write side's
spinloop, you just get a box that silently goes out of service for
extended periods of time, breaking cluster membership and more. The
NMI oopser is just the report of the problem, not the problem itself.
It's not a false report. The box is dead for > 5s at a time.


William Lee Irwin III wrote:
>> And thread groups can share mm's. do_for_each_thread() won't suffice.

On Sat, Nov 20, 2004 at 03:03:17PM +1100, Nick Piggin wrote:
> I think it will be just fine.

And that makes it wrong on both counts. The above fails any time
LD_ASSUME_KERNEL=2.4 is used, we well as when actual Linux features
are used directly.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Fri, Nov 19, 2004 at 09:33:12PM -0600, Robin Holt wrote:
> Agree, we are currently using atomic ops on a global rss on our 2.4
> kernel with 512cpu systems and not seeing much cacheline contention.
> I don't remember how little it ended up being, but it was very little.
> We had gone to dropping the page_table_lock and only reaquiring it if
> the pte was non-null when we went to insert our new one.  I think that
> was how we had it working.  I would have to wake up and actually look
> at that code as it was many months ago that Ray Bryant did that work.
> We did make rss atomic.  Most of the contention is sorted out by the
> mmap_sem.  Processes acquiring themselves off of mmap_sem were found
> to have spaced themselves out enough that they were all approximately
> equal time from doing their atomic_add and therefore had very little
> contention for the cacheline.  At least it was not enough that we could
> measure it as significant.

Also, the densely-packed split counter can only get 4-16 cpus to a
cacheline with cachelines <= 128B, so there are definite limitations to
the amount of cacheline contention in such schemes.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>/proc/ triggering NMI oopses was a persistent problem even before that
>>>code was merged. I've not bothered testing it as it at best aggravates it.
> 
> 
> On Sat, Nov 20, 2004 at 03:03:17PM +1100, Nick Piggin wrote:
> 
>>It isn't a problem. If it ever became a problem then we can just
>>touch the nmi oopser in the loop.
> 
> 
> Very, very wrong. The tasklist scans hold the read side of the lock
> and aren't even what's running with interrupts off. The contenders
> on the write side are what the NMI oopser oopses.
> 

*blinks*

So explain how this is "very very wrong", then?

> And supposing the arch reenables interrupts in the write side's
> spinloop, you just get a box that silently goes out of service for
> extended periods of time, breaking cluster membership and more. The
> NMI oopser is just the report of the problem, not the problem itself.
> It's not a false report. The box is dead for > 5s at a time.
> 

The point is, adding a for-each-thread loop or two in /proc isn't
going to cause a problem that isn't already there.

If you had zero for-each-thread loops then you might have a valid
complaint. Seeing as you have more than zero, with slim chances of
reducing that number, then there is no valid complaint.

> 
> William Lee Irwin III wrote:
> 
>>>And thread groups can share mm's. do_for_each_thread() won't suffice.
> 
> 
> On Sat, Nov 20, 2004 at 03:03:17PM +1100, Nick Piggin wrote:
> 
>>I think it will be just fine.
> 
> 
> And that makes it wrong on both counts. The above fails any time
> LD_ASSUME_KERNEL=2.4 is used, we well as when actual Linux features
> are used directly.
> 

See my followup.

Re: page fault scalability patch V11 [0/7]: overview

[Robin Holt]

On Sat, Nov 20, 2004 at 02:58:36PM +1100, Nick Piggin wrote:
> >Please be more specific about the result, and cite the Message-Id.
> >
> 
> Start of this thread.

Part of the impact was having the page table lock, the mmap_sem, and
these two atomic counters in the same cacheline.  What about seperating
the counters from the locks?

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> Very, very wrong. The tasklist scans hold the read side of the lock
>> and aren't even what's running with interrupts off. The contenders
>> on the write side are what the NMI oopser oopses.

On Sat, Nov 20, 2004 at 03:29:29PM +1100, Nick Piggin wrote:
> *blinks*
> So explain how this is "very very wrong", then?

There isn't anything left to explain. So if there's a question, be
specific about it.


William Lee Irwin III wrote:
>> And supposing the arch reenables interrupts in the write side's
>> spinloop, you just get a box that silently goes out of service for
>> extended periods of time, breaking cluster membership and more. The
>> NMI oopser is just the report of the problem, not the problem itself.
>> It's not a false report. The box is dead for > 5s at a time.

On Sat, Nov 20, 2004 at 03:29:29PM +1100, Nick Piggin wrote:
> The point is, adding a for-each-thread loop or two in /proc isn't
> going to cause a problem that isn't already there.
> If you had zero for-each-thread loops then you might have a valid
> complaint. Seeing as you have more than zero, with slim chances of
> reducing that number, then there is no valid complaint.

This entire line of argument is bogus. A preexisting bug of a similar
nature is not grounds for deliberately introducing any bug.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>Very, very wrong. The tasklist scans hold the read side of the lock
>>>and aren't even what's running with interrupts off. The contenders
>>>on the write side are what the NMI oopser oopses.
> 
> 
> On Sat, Nov 20, 2004 at 03:29:29PM +1100, Nick Piggin wrote:
> 
>>*blinks*
>>So explain how this is "very very wrong", then?
> 
> 
> There isn't anything left to explain. So if there's a question, be
> specific about it.
> 

Why am I very very wrong? Why won't touch_nmi_watchdog work from
the read loop?

And let's just be nice and try not to jump at the chance to point
out when people are very very wrong, and keep count of the times
they have been very very wrong. I'm trying to be constructive.

> 
> William Lee Irwin III wrote:
> 
>>>And supposing the arch reenables interrupts in the write side's
>>>spinloop, you just get a box that silently goes out of service for
>>>extended periods of time, breaking cluster membership and more. The
>>>NMI oopser is just the report of the problem, not the problem itself.
>>>It's not a false report. The box is dead for > 5s at a time.
> 
> 
> On Sat, Nov 20, 2004 at 03:29:29PM +1100, Nick Piggin wrote:
> 
>>The point is, adding a for-each-thread loop or two in /proc isn't
>>going to cause a problem that isn't already there.
>>If you had zero for-each-thread loops then you might have a valid
>>complaint. Seeing as you have more than zero, with slim chances of
>>reducing that number, then there is no valid complaint.
> 
> 
> This entire line of argument is bogus. A preexisting bug of a similar
> nature is not grounds for deliberately introducing any bug.
> 

Sure, if that is a bug and someone is just about to fix it then
yes you're right, we shouldn't introduce this. I didn't realise
it was a bug. Sounds like it would be causing you lots of problems
though - have you looked at how to fix it?

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> There isn't anything left to explain. So if there's a question, be
>> specific about it.

On Sat, Nov 20, 2004 at 04:50:25PM +1100, Nick Piggin wrote:
> Why am I very very wrong? Why won't touch_nmi_watchdog work from
> the read loop?
> And let's just be nice and try not to jump at the chance to point
> out when people are very very wrong, and keep count of the times
> they have been very very wrong. I'm trying to be constructive.

touch_nmi_watchdog() is only "protection" against local interrupt
disablement triggering the NMI oopser because alert_counter[]
increments are not atomic. Yet even supposing they were made so, the
net effect of "covering up" this gross deficiency is making the
user-observable problems it causes undiagnosable, as noted before.


William Lee Irwin III wrote:
>> This entire line of argument is bogus. A preexisting bug of a similar
>> nature is not grounds for deliberately introducing any bug.

On Sat, Nov 20, 2004 at 04:50:25PM +1100, Nick Piggin wrote:
> Sure, if that is a bug and someone is just about to fix it then
> yes you're right, we shouldn't introduce this. I didn't realise
> it was a bug. Sounds like it would be causing you lots of problems
> though - have you looked at how to fix it?

Kevin Marin was the first to report this issue to lkml. I had seen
instances of it in internal corporate bugreports and it was one of
the motivators for the work I did on pidhashing (one of the causes
of the timeouts was worst cases in pid allocation). Manfred Spraul
and myself wrote patches attempting to reduce read-side hold time
in /proc/ algorithms, Ingo Molnar wrote patches to hierarchically
subdivide the /proc/ iterations, and Dipankar Sarma and Maneesh
Soni wrote patches to carry out the long iterations in /proc/ locklessly.

The last several of these affecting /proc/ have not gained acceptance,
though the work has not been halted in any sense, as this problem
recurs quite regularly. A considerable amount of sustained effort has
gone toward mitigating and resolving rwlock starvation.

Aggravating the rwlock starvation destabilizes, not pessimizes,
and performance is secondary to stability.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>There isn't anything left to explain. So if there's a question, be
>>>specific about it.
> 
> 
> On Sat, Nov 20, 2004 at 04:50:25PM +1100, Nick Piggin wrote:
> 
>>Why am I very very wrong? Why won't touch_nmi_watchdog work from
>>the read loop?
>>And let's just be nice and try not to jump at the chance to point
>>out when people are very very wrong, and keep count of the times
>>they have been very very wrong. I'm trying to be constructive.
> 
> 
> touch_nmi_watchdog() is only "protection" against local interrupt
> disablement triggering the NMI oopser because alert_counter[]
> increments are not atomic. Yet even supposing they were made so, the

That would be a bug in touch_nmi_watchdog then, because you're
racy against your own NMI too.

So I'm actually not very very wrong at all. I'm technically wrong
because touch_nmi_watchdog has a theoretical 'bug'. In practice,
multiple races with the non atomic increments to the same counter,
and in an unbroken sequence would be about as likely as hardware
failure.

Anyway, this touch nmi thing is going off topic, sorry list.

> net effect of "covering up" this gross deficiency is making the
> user-observable problems it causes undiagnosable, as noted before.
> 

Well the loops that are in there now aren't covered up, and they
don't seem to be causing problems. Ergo there is no problem (we're
being _practical_ here, right?)

> 
> William Lee Irwin III wrote:
> 
>>>This entire line of argument is bogus. A preexisting bug of a similar
>>>nature is not grounds for deliberately introducing any bug.
> 
> 
> On Sat, Nov 20, 2004 at 04:50:25PM +1100, Nick Piggin wrote:
> 
>>Sure, if that is a bug and someone is just about to fix it then
>>yes you're right, we shouldn't introduce this. I didn't realise
>>it was a bug. Sounds like it would be causing you lots of problems
>>though - have you looked at how to fix it?
> 
> 
> Kevin Marin was the first to report this issue to lkml. I had seen
> instances of it in internal corporate bugreports and it was one of
> the motivators for the work I did on pidhashing (one of the causes
> of the timeouts was worst cases in pid allocation). Manfred Spraul
> and myself wrote patches attempting to reduce read-side hold time
> in /proc/ algorithms, Ingo Molnar wrote patches to hierarchically
> subdivide the /proc/ iterations, and Dipankar Sarma and Maneesh
> Soni wrote patches to carry out the long iterations in /proc/ locklessly.
> 
> The last several of these affecting /proc/ have not gained acceptance,
> though the work has not been halted in any sense, as this problem
> recurs quite regularly. A considerable amount of sustained effort has
> gone toward mitigating and resolving rwlock starvation.
> 

That's very nice. But there is no problem _now_, is there?

> Aggravating the rwlock starvation destabilizes, not pessimizes,
> and performance is secondary to stability.
> 

Well luckily we're not going to be aggravating the rwlock stavation.

If you found a problem with, and fixed do_task_stat: ?time, ???_flt,
et al, then you would apply the same solution to per thread rss to
fix it in the same way.

Re: page fault scalability patch V11 [0/7]: overview

[Andrew Morton]

Nick Piggin <nickpiggin@yahoo.com.au> wrote:
>
> per thread rss

Given that we have contention problems updating a single mm-wide rss and
given that the way to fix that up is to spread things out a bit, it seems
wildly arbitrary to me that the way in which we choose to spread the
counter out is to stick a bit of it into each task_struct.

I'd expect that just shoving a pointer into mm_struct which points at a
dynamically allocated array[NR_CPUS] of longs would suffice.  We probably
don't even need to spread them out on cachelines - having four or eight
cpus sharing the same cacheline probably isn't going to hurt much.

At least, that'd be my first attempt.  If it's still not good enough, try
something else.

Re: page fault scalability patch V11 [0/7]: overview

[Andrew Morton]

Andrew Morton <akpm@osdl.org> wrote:
>
> I'd expect that just shoving a pointer into mm_struct which points at a
>  dynamically allocated array[NR_CPUS] of longs would suffice.

One might even be able to use percpu_counter.h, although that might end up
hurting many-cpu fork times, due to all that work in __alloc_percpu().

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

Andrew Morton wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> 
>>per thread rss
> 
> 
> Given that we have contention problems updating a single mm-wide rss and
> given that the way to fix that up is to spread things out a bit, it seems
> wildly arbitrary to me that the way in which we choose to spread the
> counter out is to stick a bit of it into each task_struct.
> 
> I'd expect that just shoving a pointer into mm_struct which points at a
> dynamically allocated array[NR_CPUS] of longs would suffice.  We probably
> don't even need to spread them out on cachelines - having four or eight
> cpus sharing the same cacheline probably isn't going to hurt much.
> 
> At least, that'd be my first attempt.  If it's still not good enough, try
> something else.
> 
> 

That is what Bill thought too. I guess per-cpu and per-thread rss are
the leading candidates.

Per thread rss has the benefits of cacheline exclusivity, and not
causing task bloat in the common case.

Per CPU array has better worst case /proc properties, but shares
cachelines (or not, if using percpu_counter as you suggested).


I think I'd better leave it to others to finish off the arguments ;)

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

William Lee Irwin III wrote:
>> touch_nmi_watchdog() is only "protection" against local interrupt
>> disablement triggering the NMI oopser because alert_counter[]
>> increments are not atomic. Yet even supposing they were made so, the

On Sat, Nov 20, 2004 at 05:49:53PM +1100, Nick Piggin wrote:
> That would be a bug in touch_nmi_watchdog then, because you're
> racy against your own NMI too.
> So I'm actually not very very wrong at all. I'm technically wrong
> because touch_nmi_watchdog has a theoretical 'bug'. In practice,
> multiple races with the non atomic increments to the same counter,
> and in an unbroken sequence would be about as likely as hardware
> failure.
> Anyway, this touch nmi thing is going off topic, sorry list.

No, it's on-topic.
(1) The issue is not theoretical. e.g. sysrq t does trigger NMI oopses,
	merely not every time, and not on every system. It is not
	associated with hardware failure. It is, however, tolerable
	because sysrq's require privilege to trigger and are primarly
	used when the box is dying anyway.
(2) NMI's don't nest. There is no possibility of NMI's racing against
	themselves while the data is per-cpu.


William Lee Irwin III wrote:
>> net effect of "covering up" this gross deficiency is making the
>> user-observable problems it causes undiagnosable, as noted before.

On Sat, Nov 20, 2004 at 05:49:53PM +1100, Nick Piggin wrote:
> Well the loops that are in there now aren't covered up, and they
> don't seem to be causing problems. Ergo there is no problem (we're
> being _practical_ here, right?)

They are causing problems. They never stopped causing problems. None
of the above attempts to reduce rwlock starvation has been successful
in reducing it to untriggerable-in-the-field levels, and empirical
demonstrations of starvation recurring after those available at the
time of testing were put into place did in fact happen. Reduction of
frequency and making starvation more difficult to trigger are all that
they've achieved thus far.


William Lee Irwin III wrote:
>> Kevin Marin was the first to report this issue to lkml. I had seen
>> instances of it in internal corporate bugreports and it was one of
>> the motivators for the work I did on pidhashing (one of the causes
>> of the timeouts was worst cases in pid allocation). Manfred Spraul
>> and myself wrote patches attempting to reduce read-side hold time
>> in /proc/ algorithms, Ingo Molnar wrote patches to hierarchically
>> subdivide the /proc/ iterations, and Dipankar Sarma and Maneesh
>> Soni wrote patches to carry out the long iterations in /proc/ locklessly.
>> The last several of these affecting /proc/ have not gained acceptance,
>> though the work has not been halted in any sense, as this problem
>> recurs quite regularly. A considerable amount of sustained effort has
>> gone toward mitigating and resolving rwlock starvation.

On Sat, Nov 20, 2004 at 05:49:53PM +1100, Nick Piggin wrote:
> That's very nice. But there is no problem _now_, is there?

There is and has always been. All of the above merely mitigate the
issue, with the possible exception of the tasklist RCU patch, for
which I know of no testing results. Also note that almost none of
the work on /proc/ has been merged.


William Lee Irwin III wrote:
>> Aggravating the rwlock starvation destabilizes, not pessimizes,
>> and performance is secondary to stability.

On Sat, Nov 20, 2004 at 05:49:53PM +1100, Nick Piggin wrote:
> Well luckily we're not going to be aggravating the rwlock stavation.
> If you found a problem with, and fixed do_task_stat: ?time, ???_flt,
> et al, then you would apply the same solution to per thread rss to
> fix it in the same way.

You are aggravating the rwlock starvation by introducing gratuitous
full tasklist iterations. There is no solution to do_task_stat()
because it was recently introduced. There will be one as part of a port
of the usual mitigation patches when the perennial problem is reported
against a sufficiently recent kernel version, as usual. The already-
demonstrated problematic iterations have not been removed.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

William Lee Irwin III wrote:
> William Lee Irwin III wrote:
> 
>>>touch_nmi_watchdog() is only "protection" against local interrupt
>>>disablement triggering the NMI oopser because alert_counter[]
>>>increments are not atomic. Yet even supposing they were made so, the
> 
> 
> On Sat, Nov 20, 2004 at 05:49:53PM +1100, Nick Piggin wrote:
> 
>>That would be a bug in touch_nmi_watchdog then, because you're
>>racy against your own NMI too.
>>So I'm actually not very very wrong at all. I'm technically wrong
>>because touch_nmi_watchdog has a theoretical 'bug'. In practice,
>>multiple races with the non atomic increments to the same counter,
>>and in an unbroken sequence would be about as likely as hardware
>>failure.
>>Anyway, this touch nmi thing is going off topic, sorry list.
> 
> 
> No, it's on-topic.
> (1) The issue is not theoretical. e.g. sysrq t does trigger NMI oopses,
> 	merely not every time, and not on every system. It is not
> 	associated with hardware failure. It is, however, tolerable
> 	because sysrq's require privilege to trigger and are primarly
> 	used when the box is dying anyway.

OK then put a touch_nmi_watchdog in there if you must.

> (2) NMI's don't nest. There is no possibility of NMI's racing against
> 	themselves while the data is per-cpu.
> 

Your point was that touch_nmi_watchdog() which resets alert_counter,
is racy when resetting the counter of other CPUs. Yes it is racy.
It is also racy against the NMI on the _current_ CPU.

This has nothing whatsoever to do with NMIs racing against themselves,
I don't know how you got that idea when you were the one to bring up
this race anyway.

[ snip back-and-forth that is going nowhere ]

I'll bow out of the argument here. I grant you raise valid concens
WRT the /proc issues, of course.

touch_nmi_watchdog (was: page fault scalability patch V11 [0/7]: overview)

[Nick Piggin]

Nick Piggin wrote:

>> (2) NMI's don't nest. There is no possibility of NMI's racing against
>>     themselves while the data is per-cpu.
>>
> 
> Your point was that touch_nmi_watchdog() which resets alert_counter,
> is racy when resetting the counter of other CPUs. Yes it is racy.
> It is also racy against the NMI on the _current_ CPU.

Hmm no I think you're right in that it is only a problem WRT the remote
CPUs. However that would still be a problem, as the comment in i386
touch_nmi_watchdog attests.

Re: page fault scalability patch V11 [0/7]: overview

[Nick Piggin]

Nick Piggin wrote:
> William Lee Irwin III wrote:

>> No, it's on-topic.
>> (1) The issue is not theoretical. e.g. sysrq t does trigger NMI oopses,
>>     merely not every time, and not on every system. It is not
>>     associated with hardware failure. It is, however, tolerable
>>     because sysrq's require privilege to trigger and are primarly
>>     used when the box is dying anyway.
> 
> 
> OK then put a touch_nmi_watchdog in there if you must.
> 

Duh, there is one in there :\

Still, that doesn't really say much about a normal tasklist traversal
because this thing will spend ages writing stuff to serial console.

Now I know going over the whole tasklist is crap. Anything O(n) for
things like this is crap. I happen to just get frustrated to see
concessions being made to support more efficient /proc access. I know
you are one of the ones who has to deal with the practical realities
of that though. Sigh. Well try to bear with me... :|

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

Andrew Morton wrote:
>> Given that we have contention problems updating a single mm-wide rss and
>> given that the way to fix that up is to spread things out a bit, it seems
>> wildly arbitrary to me that the way in which we choose to spread the
>> counter out is to stick a bit of it into each task_struct.
>> I'd expect that just shoving a pointer into mm_struct which points at a
>> dynamically allocated array[NR_CPUS] of longs would suffice.  We probably
>> don't even need to spread them out on cachelines - having four or eight
>> cpus sharing the same cacheline probably isn't going to hurt much.
>> At least, that'd be my first attempt.  If it's still not good enough, try
>> something else.

On Sat, Nov 20, 2004 at 06:13:03PM +1100, Nick Piggin wrote:
> That is what Bill thought too. I guess per-cpu and per-thread rss are
> the leading candidates.
> Per thread rss has the benefits of cacheline exclusivity, and not
> causing task bloat in the common case.
> Per CPU array has better worst case /proc properties, but shares
> cachelines (or not, if using percpu_counter as you suggested).
> I think I'd better leave it to others to finish off the arguments ;)

(1) The "task bloat" is more than tolerable on the systems capable
	of having enough cpus to see significant per-process
	memory footprint, where "significant" is smaller than a
	pagetable page even for systems twice as large as now shipped.
(2) The cacheline exclusivity is not entirely gone in dense per-cpu
	arrays, it's merely "approximated" by sharing amongst small
	groups of adjacent cpus. This is fine for e.g. NUMA because
	those small groups of adjacent cpus will typically be on nearby
	nodes.
(3) The price paid to get "perfect exclusivity" instead of "approximate
	exclusivity" is unbounded tasklist_lock hold time, which takes
	boxen down outright in every known instance.

The properties are not for /proc/, they are for tasklist_lock. Every
read stops all other writes. When you hold tasklist_lock for an
extended period of time for read or write, (e.g. exhaustive tasklist
search) you stop all fork()'s and exit()'s and execve()'s on a running
system. The "worst case" analysis has nothing to do with speed. It has
everything to do with taking a box down outright, much like unplugging
power cables or dereferencing NULL. Unbounded tasklist_lock hold time
kills running boxen dead.

Read sides of rwlocks are not licenses to spin for aeons with locks held.

And the "question" of sufficiency has in fact already been answered.
SGI's own testing during the 2.4 out-of-tree patching cycle determined
that an mm-global atomic counter was already sufficient so long as the
cacheline was not shared with ->mmap_sem and the like. The "simplest"
optimization of moving the field out of the way of ->mmap_sem already
worked. The grander ones, if and *ONLY* if they don't have showstoppers
like unbounded tasklist_lock hold time or castrating workload monitoring
to unusability, will merely be more robust for future systems.
Reiterating, this is all just fine so long as they don't cause any
showstopping problems, like castrating the ability to monitor
processes, or introducing more tasklist_lock starvation.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

Nick Piggin wrote:
>> OK then put a touch_nmi_watchdog in there if you must.

On Sat, Nov 20, 2004 at 06:57:38PM +1100, Nick Piggin wrote:
> Duh, there is one in there :\
> Still, that doesn't really say much about a normal tasklist traversal
> because this thing will spend ages writing stuff to serial console.
> Now I know going over the whole tasklist is crap. Anything O(n) for
> things like this is crap. I happen to just get frustrated to see
> concessions being made to support more efficient /proc access. I know
> you are one of the ones who has to deal with the practical realities
> of that though. Sigh. Well try to bear with me... :|

I sure as Hell don't have any interest in /proc/ in and of itself,
but this stuff does really bite people, and hard, too.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Martin J. Bligh]

>> Given that we have contention problems updating a single mm-wide rss and
>> given that the way to fix that up is to spread things out a bit, it seems
>> wildly arbitrary to me that the way in which we choose to spread the
>> counter out is to stick a bit of it into each task_struct.
>> 
>> I'd expect that just shoving a pointer into mm_struct which points at a
>> dynamically allocated array[NR_CPUS] of longs would suffice.  We probably
>> don't even need to spread them out on cachelines - having four or eight
>> cpus sharing the same cacheline probably isn't going to hurt much.
>> 
>> At least, that'd be my first attempt.  If it's still not good enough, try
>> something else.
>> 
>> 
> 
> That is what Bill thought too. I guess per-cpu and per-thread rss are
> the leading candidates.
> 
> Per thread rss has the benefits of cacheline exclusivity, and not
> causing task bloat in the common case.
> 
> Per CPU array has better worst case /proc properties, but shares
> cachelines (or not, if using percpu_counter as you suggested).

Per thread seems much nicer to me - mainly because it degrades cleanly to 
a single counter for 99% of processes, which are single threaded.

M.

Re: page fault scalability patch V11 [0/7]: overview

[Linus Torvalds]

On Sat, 20 Nov 2004, Martin J. Bligh wrote:
> 
> Per thread seems much nicer to me - mainly because it degrades cleanly to 
> a single counter for 99% of processes, which are single threaded.

I will pretty much guarantee that if you put the per-thread patches next
to some abomination with per-cpu allocation for each mm, the choice will
be clear. Especially if the per-cpu/per-mm thing tries to avoid false
cacheline sharing, which sounds really "interesting" in itself.

And without the cacheline sharing avoidance, what's the point of this 
again? It sure wasn't to make the code simpler. It was about performance 
and scalability.

		Linus

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Sat, Nov 20, 2004 at 09:14:11AM -0800, Linus Torvalds wrote:
> I will pretty much guarantee that if you put the per-thread patches next
> to some abomination with per-cpu allocation for each mm, the choice will
> be clear. Especially if the per-cpu/per-mm thing tries to avoid false
> cacheline sharing, which sounds really "interesting" in itself.
> And without the cacheline sharing avoidance, what's the point of this 
> again? It sure wasn't to make the code simpler. It was about performance 
> and scalability.

"The perfect is the enemy of the good."

The "perfect" cacheline separation achieved that way is at the cost of
destabilizing the kernel. The dense per-cpu business is only really a
concession to the notion that the counter needs to be split up at all,
which has never been demonstrated with performance measurements. In fact,
Robin Holt has performance measurements demonstrating the opposite.

The "good" alternatives are negligibly different wrt. performance, and
don't carry the high cost of rwlock starvation that breaks boxen.


-- wli

Re: page fault scalability patch V11 [0/7]: overview

[Linus Torvalds]

On Sat, 20 Nov 2004, William Lee Irwin III wrote:
> 
> "The perfect is the enemy of the good."

Yes. But in this case, my suggestion _is_ the good. You seem to be pushing 
for a really horrid thing which allocates a per-cpu array for each 
mm_struct. 

What is it that you have against the per-thread rss? We already have 
several places that do the thread-looping, so it's not like "you can't do 
that" is a valid argument.

		Linus

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Sat, 20 Nov 2004, William Lee Irwin III wrote:
>> "The perfect is the enemy of the good."

On Sat, Nov 20, 2004 at 11:16:12AM -0800, Linus Torvalds wrote:
> Yes. But in this case, my suggestion _is_ the good. You seem to be pushing 
> for a really horrid thing which allocates a per-cpu array for each 
> mm_struct. 
> What is it that you have against the per-thread rss? We already have 
> several places that do the thread-looping, so it's not like "you can't do 
> that" is a valid argument.

Okay, first thread groups can share mm's, so it's worse than iterating
over a thread group. Second, the long loops under tasklist_lock didn't
stop causing rwlock starvation because what patches there were to do
something about them didn't get merged.

I'm not particularly "stuck on" the per-cpu business, it was merely the
most obvious method of splitting the RSS counter without catastrophes
elsewhere. Robin Holt's 2.4 performance studies actually show that
splitting the counter is not even essential.


-- wli

[OT] Re: page fault scalability patch V11 [0/7]: overview

[Adam Heath]

On Sat, 20 Nov 2004, William Lee Irwin III wrote:

> "The perfect is the enemy of the good."

"With the proper course of action made so explicit, we had merely to choose
between wisdom and folly.  Precisely how we chose folly in this instance is
not entirely clear."

Quote taken from Andrew Suffield on irc, who said he got it from Penny Arcade.

Re: page fault scalability patch V11 [1/7]: sloppy rss

[Hugh Dickins]

On Fri, 19 Nov 2004, Christoph Lameter wrote:
> On Fri, 19 Nov 2004, Hugh Dickins wrote:
> 
> > Sorry, against what tree do these patches apply?
> > Apparently not linux-2.6.9, nor latest -bk, nor -mm?
> 
> 2.6.10-rc2-bk3

Ah, thanks - got it patched now, but your mailer (or something else)
is eating trailing spaces.  Better than adding them, but we have to
apply this patch before your set:

--- 2.6.10-rc2-bk3/include/asm-i386/system.h	2004-11-15 16:21:12.000000000 +0000
+++ linux/include/asm-i386/system.h	2004-11-22 14:44:30.761904592 +0000
@@ -273,9 +273,9 @@ static inline unsigned long __cmpxchg(vo
 #define cmpxchg(ptr,o,n)\
 	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
 					(unsigned long)(n),sizeof(*(ptr))))
-    
+
 #ifdef __KERNEL__
-struct alt_instr { 
+struct alt_instr {
 	__u8 *instr; 		/* original instruction */
 	__u8 *replacement;
 	__u8  cpuid;		/* cpuid bit set for replacement */
--- 2.6.10-rc2-bk3/include/asm-s390/pgalloc.h	2004-05-10 03:33:39.000000000 +0100
+++ linux/include/asm-s390/pgalloc.h	2004-11-22 14:54:43.704723120 +0000
@@ -99,7 +99,7 @@ static inline void pgd_populate(struct m
 
 #endif /* __s390x__ */
 
-static inline void 
+static inline void
 pmd_populate_kernel(struct mm_struct *mm, pmd_t *pmd, pte_t *pte)
 {
 #ifndef __s390x__
--- 2.6.10-rc2-bk3/mm/memory.c	2004-11-18 17:56:11.000000000 +0000
+++ linux/mm/memory.c	2004-11-22 14:39:33.924030808 +0000
@@ -1424,7 +1424,7 @@ out:
 /*
  * We are called with the MM semaphore and page_table_lock
  * spinlock held to protect against concurrent faults in
- * multithreaded programs. 
+ * multithreaded programs.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
@@ -1615,7 +1615,7 @@ static int do_file_page(struct mm_struct
 	 * Fall back to the linear mapping if the fs does not support
 	 * ->populate:
 	 */
-	if (!vma->vm_ops || !vma->vm_ops->populate || 
+	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
 		return do_no_page(mm, vma, address, write_access, pte, pmd);

Re: page fault scalability patch V11 [0/7]: overview

[Christoph Lameter]

On Sat, 20 Nov 2004, William Lee Irwin III wrote:
> I'm not particularly "stuck on" the per-cpu business, it was merely the
> most obvious method of splitting the RSS counter without catastrophes
> elsewhere. Robin Holt's 2.4 performance studies actually show that
> splitting the counter is not even essential.

There is no problem moving back to the atomic approach that is if it is
okay to also make anon_rss atomic. But its a pretty significant
performance hit (comparison with some old data from V4 of patch which
makes this data a bit suspect since the test environment is likely
slightly different. I should really test this again. Note that the old
performance test was only run 3 times instead of 10):

atomic vs. sloppy rss performance 64G allocation:

sloppy rss:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 16  10    1    1.818s    131.556s 133.038s 78618.592  78615.672
 16  10    2    1.736s    121.167s  65.026s 85317.098 160656.362
 16  10    4    1.835s    120.444s  36.002s 85751.810 291074.998
 16  10    8    1.820s    131.068s  25.049s 78906.310 411304.895
 16  10   16    3.275s    194.971s  22.019s 52892.356 472497.962
 16  10   32   13.006s    496.628s  27.044s 20575.038 381999.865

atomic:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 16   3    1    0.610s     61.557s  62.016s 50600.438  50599.822
 16   3    2    0.640s     83.116s  43.016s 37557.847  72869.978
 16   3    4    0.621s     73.897s  26.023s 42214.002 119908.246
 16   3    8    0.596s     86.587s  14.098s 36081.229 209962.059
 16   3   16    0.646s     69.601s   7.000s 44780.269 448823.690
 16   3   32    0.903s    185.609s   8.085s 16866.018 355301.694

Lets go for the approach to move rss into the thread structure but
keep the rss in the mm structure as is (need to take page_table_lock
for update) to consolidate the values. This allows to keep most
of the code as is and the rss in the task struct is only used if
we are not holding page_table_lock.

Maybe we can then find some way to regularly update the rss in the mm
structure to avoid the loop over the tasklist in proc.

Re: page fault scalability patch V11 [0/7]: overview

[Bill Davidsen]

Linus Torvalds wrote:
> 
> On Sat, 20 Nov 2004, Nick Piggin wrote:
> 
>>The per thread rss may wrap (maybe not 64-bit counters), but even so,
>>the summation over all threads should still end up being correct I
>>think.
> 
> 
> Yes. As long as the total rss fits in an int, it doesn't matter if any of
> them wrap. Addition is still associative in twos-complement arithmetic 
> even in the presense of overflows. 
> 
> If you actually want to make it proper standard C, I guess you'd have to 
> make the thing unsigned, which gives you the mod-2**n guarantees even if 
> somebody were to ever make a non-twos-complement machine.

I think other stuff breaks as well, I think I saw you post some example 
code using something like (a & -a) or similar within the last few 
months. Fortunately neither 1's comp or BCD are likeliy to return in 
hardware. Big-end vs. little-end is still an issue, though.

-- 
    -bill davidsen (davidsen@tmr.com)
"The secret to procrastination is to put things off until the
  last possible moment - but no longer"  -me

deferred rss update instead of sloppy rss

[Christoph Lameter]

One way to solve the rss issues is--as discussed--to put rss into the
task structure and then have the page fault increment that rss.

The problem is then that the proc filesystem must do an extensive scan
over all threads to find users of a certain mm_struct.

The following patch does put the rss into task_struct. The page fault
handler is then incrementing current->rss if the page_table_lock is not
held.

The timer interrupt checks if task->rss is non zero (when doing
stime/utime updates. rss is defined near those so its hopefully on the
same cacheline and has a minimal impact).

If rss is non zero and the page_table_lock and the mmap_sem can be taken
then the mm->rss will be updated by the value of the task->rss and
task->rss will be zeroed. This avoids all proc issues. The only
disadvantage is that rss may be inaccurate for a couple of clock ticks.

This also adds some performance (sorry only a 4p system):

sloppy rss:

Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  4  10    1    0.593s     29.897s  30.050s 85973.585  85948.565
  4  10    2    0.616s     42.184s  22.045s 61247.450 116719.558
  4  10    4    0.559s     44.918s  14.076s 57641.255 177553.945

deferred rss:
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  4  10    1    0.565s     29.429s  30.000s 87395.518  87366.580
  4  10    2    0.500s     33.514s  18.002s 77067.935 145426.659
  4  10    4    0.533s     44.455s  14.085s 58269.368 176413.196

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-11-15 11:13:39.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-11-22 13:18:58.000000000 -0800
@@ -584,6 +584,10 @@
 	unsigned long it_real_incr, it_prof_incr, it_virt_incr;
 	struct timer_list real_timer;
 	unsigned long utime, stime;
+	long rss;	/* rss counter when mm->rss is not usable. mm->page_table_lock
+			 * not held but mm->mmap_sem must be held for sync with
+			 * the timer interrupt which clears rss and updates mm->rss.
+			 */
 	unsigned long nvcsw, nivcsw; /* context switch counts */
 	struct timespec start_time;
 /* mm fault and swap info: this can arguably be seen as either mm-specific or thread-specific */
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-11-22 09:51:58.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-11-22 11:16:02.000000000 -0800
@@ -263,8 +263,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -507,8 +505,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -791,8 +787,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;
Index: linux-2.6.9/kernel/fork.c
===================================================================
--- linux-2.6.9.orig/kernel/fork.c	2004-11-22 09:51:58.000000000 -0800
+++ linux-2.6.9/kernel/fork.c	2004-11-22 11:16:02.000000000 -0800
@@ -876,6 +876,7 @@
 	p->io_context = NULL;
 	p->io_wait = NULL;
 	p->audit_context = NULL;
+	p->rss = 0;
 #ifdef CONFIG_NUMA
  	p->mempolicy = mpol_copy(p->mempolicy);
  	if (IS_ERR(p->mempolicy)) {
Index: linux-2.6.9/kernel/exit.c
===================================================================
--- linux-2.6.9.orig/kernel/exit.c	2004-11-22 09:51:58.000000000 -0800
+++ linux-2.6.9/kernel/exit.c	2004-11-22 11:16:02.000000000 -0800
@@ -501,6 +501,9 @@
 	/* more a memory barrier than a real lock */
 	task_lock(tsk);
 	tsk->mm = NULL;
+	/* only holding mmap_sem here maybe get page_table_lock too? */
+	mm->rss += tsk->rss;
+	tsk->rss = 0;
 	up_read(&mm->mmap_sem);
 	enter_lazy_tlb(mm, current);
 	task_unlock(tsk);
Index: linux-2.6.9/kernel/timer.c
===================================================================
--- linux-2.6.9.orig/kernel/timer.c	2004-11-22 09:51:58.000000000 -0800
+++ linux-2.6.9/kernel/timer.c	2004-11-22 11:42:12.000000000 -0800
@@ -815,6 +815,15 @@
 		if (psecs / HZ >= p->signal->rlim[RLIMIT_CPU].rlim_max)
 			send_sig(SIGKILL, p, 1);
 	}
+	/* Update mm->rss if necessary */
+	if (p->rss && p->mm && down_write_trylock(&p->mm->mmap_sem)) {
+		if (spin_trylock(&p->mm->page_table_lock)) {
+			p->mm->rss += p->rss;
+			p->rss = 0;
+			spin_unlock(&p->mm->page_table_lock);
+		}
+		up_write(&p->mm->mmap_sem);
+	}
 }

 static inline void do_it_virt(struct task_struct * p, unsigned long ticks)

Re: deferred rss update instead of sloppy rss

[Andrew Morton]

Christoph Lameter <clameter@sgi.com> wrote:
>
> One way to solve the rss issues is--as discussed--to put rss into the
> task structure and then have the page fault increment that rss.
> 
> The problem is then that the proc filesystem must do an extensive scan
> over all threads to find users of a certain mm_struct.
> 
> The following patch does put the rss into task_struct. The page fault
> handler is then incrementing current->rss if the page_table_lock is not
> held.
> 
> The timer interrupt checks if task->rss is non zero (when doing
> stime/utime updates. rss is defined near those so its hopefully on the
> same cacheline and has a minimal impact).
> 
> If rss is non zero and the page_table_lock and the mmap_sem can be taken
> then the mm->rss will be updated by the value of the task->rss and
> task->rss will be zeroed. This avoids all proc issues. The only
> disadvantage is that rss may be inaccurate for a couple of clock ticks.
> 
> This also adds some performance (sorry only a 4p system):
> 
> sloppy rss:
> 
> Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>   4  10    1    0.593s     29.897s  30.050s 85973.585  85948.565
>   4  10    2    0.616s     42.184s  22.045s 61247.450 116719.558
>   4  10    4    0.559s     44.918s  14.076s 57641.255 177553.945
> 
> deferred rss:
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>   4  10    1    0.565s     29.429s  30.000s 87395.518  87366.580
>   4  10    2    0.500s     33.514s  18.002s 77067.935 145426.659
>   4  10    4    0.533s     44.455s  14.085s 58269.368 176413.196

hrm.  I cannot see anywhere in this patch where you update task_struct.rss.

> Index: linux-2.6.9/kernel/exit.c
> ===================================================================
> --- linux-2.6.9.orig/kernel/exit.c	2004-11-22 09:51:58.000000000 -0800
> +++ linux-2.6.9/kernel/exit.c	2004-11-22 11:16:02.000000000 -0800
> @@ -501,6 +501,9 @@
>  	/* more a memory barrier than a real lock */
>  	task_lock(tsk);
>  	tsk->mm = NULL;
> +	/* only holding mmap_sem here maybe get page_table_lock too? */
> +	mm->rss += tsk->rss;
> +	tsk->rss = 0;
>  	up_read(&mm->mmap_sem);

mmap_sem needs to be held for writing, surely?

> +	/* Update mm->rss if necessary */
> +	if (p->rss && p->mm && down_write_trylock(&p->mm->mmap_sem)) {
> +		if (spin_trylock(&p->mm->page_table_lock)) {
> +			p->mm->rss += p->rss;
> +			p->rss = 0;
> +			spin_unlock(&p->mm->page_table_lock);
> +		}
> +		up_write(&p->mm->mmap_sem);
> +	}
>  }

I'd also suggest that you do:

	tsk->rss++;
	if (tsk->rss > 16) {
		spin_lock(&mm->page_table_lock);
		mm->rss += tsk->rss;
		spin_unlock(&mm->page_table_lock);
		tsk->rss = 0;
	}

just to prevent transient gross inaccuracies.  For some value of "16".

Re: deferred rss update instead of sloppy rss

[Christoph Lameter]

On Mon, 22 Nov 2004, Andrew Morton wrote:

> hrm.  I cannot see anywhere in this patch where you update task_struct.rss.

This is just the piece around it dealing with rss. The updating of rss
happens in the generic code. The change to that is trivial. I can repost
the whole shebang if you want.

> > +	/* only holding mmap_sem here maybe get page_table_lock too? */
> > +	mm->rss += tsk->rss;
> > +	tsk->rss = 0;
> >  	up_read(&mm->mmap_sem);
>
> mmap_sem needs to be held for writing, surely?

If there are no page faults occurring anymore then we would not need to
get the lock. Q: Is it safe to assume that no faults occur
anymore at this point?

> just to prevent transient gross inaccuracies.  For some value of "16".

The page fault code only increments rss. For larger transactions that
increase / decrease rss significantly the page_table_lock is taken and
mm->rss is updated directly. So no
gross inaccuracies can result.

Re: deferred rss update instead of sloppy rss

[Benjamin Herrenschmidt]

On Mon, 2004-11-22 at 14:13 -0800, Christoph Lameter wrote:
> On Mon, 22 Nov 2004, Andrew Morton wrote:
> 
> > hrm.  I cannot see anywhere in this patch where you update task_struct.rss.
> 
> This is just the piece around it dealing with rss. The updating of rss
> happens in the generic code. The change to that is trivial. I can repost
> the whole shebang if you want.
> 
> > > +	/* only holding mmap_sem here maybe get page_table_lock too? */
> > > +	mm->rss += tsk->rss;
> > > +	tsk->rss = 0;
> > >  	up_read(&mm->mmap_sem);
> >
> > mmap_sem needs to be held for writing, surely?
> 
> If there are no page faults occurring anymore then we would not need to
> get the lock. Q: Is it safe to assume that no faults occur
> anymore at this point?

Why wouldn't the mm take faults on other CPUs ? (other threads)

> > just to prevent transient gross inaccuracies.  For some value of "16".
> 
> The page fault code only increments rss. For larger transactions that
> increase / decrease rss significantly the page_table_lock is taken and
> mm->rss is updated directly. So no
> gross inaccuracies can result.
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>
-- 
Benjamin Herrenschmidt <benh@kernel.crashing.org>

Re: deferred rss update instead of sloppy rss

[Linus Torvalds]

On Mon, 22 Nov 2004, Christoph Lameter wrote:
> 
> The problem is then that the proc filesystem must do an extensive scan
> over all threads to find users of a certain mm_struct.

The alternative is to just add a simple list into the task_struct and the
head of it into mm_struct. Then, at fork, you just finish the fork() with

	list_add(p->mm_list, p->mm->thread_list);

and do the proper list_del() in exit_mm() or wherever.

You'll still loop in /proc, but you'll do the minimal loop necessary.

		Linus

Re: deferred rss update instead of sloppy rss

[Christoph Lameter]

On Mon, 22 Nov 2004, Linus Torvalds wrote:

> The alternative is to just add a simple list into the task_struct and the
> head of it into mm_struct. Then, at fork, you just finish the fork() with
>
> 	list_add(p->mm_list, p->mm->thread_list);
>
> and do the proper list_del() in exit_mm() or wherever.
>
> You'll still loop in /proc, but you'll do the minimal loop necessary.

I think the approach that I posted is simpler unless there are other
benefits to be gained if it would be easy to figure out which tasks use an
mm.

Re: deferred rss update instead of sloppy rss

[Nick Piggin]

Linus Torvalds wrote:
> 
> On Mon, 22 Nov 2004, Christoph Lameter wrote:
> 
>>The problem is then that the proc filesystem must do an extensive scan
>>over all threads to find users of a certain mm_struct.
> 
> 
> The alternative is to just add a simple list into the task_struct and the
> head of it into mm_struct. Then, at fork, you just finish the fork() with
> 
> 	list_add(p->mm_list, p->mm->thread_list);
> 
> and do the proper list_del() in exit_mm() or wherever.
> 
> You'll still loop in /proc, but you'll do the minimal loop necessary.
> 

Yes, that was what I was thinking we'd have to resort to. Not a bad idea.

It would be nice if you could have it integrated with the locking that
is already there - for example mmap_sem, although that might mean you'd
have to take mmap_sem for writing which may limit scalability of thread
creation / destruction... maybe a seperate lock / semaphore for that list
itself would be OK.

Deferred rss might be a practical solution, but I'd prefer this if it can
be made workable.

Re: deferred rss update instead of sloppy rss

[Christoph Lameter]

On Tue, 23 Nov 2004, Nick Piggin wrote:

> Deferred rss might be a practical solution, but I'd prefer this if it can
> be made workable.

Both results in an additional field in task_struct that is going to be
incremented when the page_table_lock is not held. It would be possible
to switch to looping in procfs later. The main question with this patchset
is:

How and when can we get this get into the kernel?

Re: deferred rss update instead of sloppy rss

[Linus Torvalds]

On Mon, 22 Nov 2004, Christoph Lameter wrote:
> 
> I think the approach that I posted is simpler unless there are other
> benefits to be gained if it would be easy to figure out which tasks use an
> mm.

I'm just worried that your timer tick thing won't catch things in a timely 
manner. That said, if that isn't an issue, and people don't have problems 
with it. On the other hand, if /proc literally is the only real user, then 
I guess it really can't matter.

		Linus

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Sat, 20 Nov 2004, William Lee Irwin III wrote:
>> I'm not particularly "stuck on" the per-cpu business, it was merely the
>> most obvious method of splitting the RSS counter without catastrophes
>> elsewhere. Robin Holt's 2.4 performance studies actually show that
>> splitting the counter is not even essential.

On Mon, Nov 22, 2004 at 09:44:02AM -0800, Christoph Lameter wrote:
> There is no problem moving back to the atomic approach that is if it is
> okay to also make anon_rss atomic. But its a pretty significant
> performance hit (comparison with some old data from V4 of patch which
> makes this data a bit suspect since the test environment is likely
> slightly different. I should really test this again. Note that the old
> performance test was only run 3 times instead of 10):
> atomic vs. sloppy rss performance 64G allocation:

The specific patches you compared matter a great deal as there are
implementation blunders (e.g. poor placement of counters relative to
->mmap_sem) that can ruin the results. URL's to the specific patches
would rule out that source of error.


-- wli

Re: deferred rss update instead of sloppy rss

[Andrew Morton]

Christoph Lameter <clameter@sgi.com> wrote:
>
> > just to prevent transient gross inaccuracies.  For some value of "16".
> 
> The page fault code only increments rss. For larger transactions that
> increase / decrease rss significantly the page_table_lock is taken and
> mm->rss is updated directly. So no
> gross inaccuracies can result.

Sure.  Take a million successive pagefaults and mm->rss is grossly
inaccurate.  Hence my suggestion that it be spilled into mm->rss
periodically.

Re: deferred rss update instead of sloppy rss

[Christoph Lameter]

On Mon, 22 Nov 2004, Andrew Morton wrote:

> > The page fault code only increments rss. For larger transactions that
> > increase / decrease rss significantly the page_table_lock is taken and
> > mm->rss is updated directly. So no
> > gross inaccuracies can result.
>
> Sure.  Take a million successive pagefaults and mm->rss is grossly
> inaccurate.  Hence my suggestion that it be spilled into mm->rss
> periodically.

It is spilled into mm->rss periodically. That is the whole point of the
patch.

The timer tick occurs every 1 ms. The maximum pagefault frequency that I
have  seen is 500000 faults /second. The max deviation is therefore
less than 500 (could be greater if page table lock / mmap_sem always held
when the tick occurs).

Re: page fault scalability patch V11 [0/7]: overview

[Christoph Lameter]

On Mon, 22 Nov 2004, William Lee Irwin III wrote:

> The specific patches you compared matter a great deal as there are
> implementation blunders (e.g. poor placement of counters relative to
> ->mmap_sem) that can ruin the results. URL's to the specific patches
> would rule out that source of error.

I mentioned V4 of this patch which was posted to lkml. A simple search
should get you there.

Re: deferred rss update instead of sloppy rss

[Nick Piggin]

Christoph Lameter wrote:
> On Mon, 22 Nov 2004, Andrew Morton wrote:
> 
> 
>>>The page fault code only increments rss. For larger transactions that
>>>increase / decrease rss significantly the page_table_lock is taken and
>>>mm->rss is updated directly. So no
>>>gross inaccuracies can result.
>>
>>Sure.  Take a million successive pagefaults and mm->rss is grossly
>>inaccurate.  Hence my suggestion that it be spilled into mm->rss
>>periodically.
> 
> 
> It is spilled into mm->rss periodically. That is the whole point of the
> patch.
> 
> The timer tick occurs every 1 ms. The maximum pagefault frequency that I
> have  seen is 500000 faults /second. The max deviation is therefore
> less than 500 (could be greater if page table lock / mmap_sem always held
> when the tick occurs).


You could imagine a situation where something pagefaults and sleeps in
lock-step with the timer though. Theoretical problem only?

I think that by the time you get the spilling code in, the mm-list method
will be looking positively elegant!

Re: deferred rss update instead of sloppy rss

[Christoph Lameter]

On Tue, 23 Nov 2004, Nick Piggin wrote:

> > The timer tick occurs every 1 ms. The maximum pagefault frequency that I
> > have  seen is 500000 faults /second. The max deviation is therefore
> > less than 500 (could be greater if page table lock / mmap_sem always held
> > when the tick occurs).
> I think that by the time you get the spilling code in, the mm-list method
> will be looking positively elegant!

I do not care what gets in as long as something goes in to address the
performance issues. So far everyone seems to have their pet ideas. By all
means do the mm-list method and post it. But we have already seen
objections by other against loops in proc. So that will also cause
additional controversy.

Re: deferred rss update instead of sloppy rss

[Nick Piggin]

Christoph Lameter wrote:
> On Tue, 23 Nov 2004, Nick Piggin wrote:
> 
> 
>>Deferred rss might be a practical solution, but I'd prefer this if it can
>>be made workable.
> 
> 
> Both results in an additional field in task_struct that is going to be
> incremented when the page_table_lock is not held. It would be possible
> to switch to looping in procfs later. The main question with this patchset
> is:
> 

Sure.

> How and when can we get this get into the kernel?
> 

Well it is a good starting platform for the various PTL reduction patches
floating around.

I'd say Andrew could be convinced to stick it in -mm after 2.6.10, but we'd
probably need a clear path to one of the PTL patches before anything would
move into 2.6.

Re: deferred rss update instead of sloppy rss

[Andrew Morton]

Christoph Lameter <clameter@sgi.com> wrote:
>
> On Mon, 22 Nov 2004, Andrew Morton wrote:
> 
> > > The page fault code only increments rss. For larger transactions that
> > > increase / decrease rss significantly the page_table_lock is taken and
> > > mm->rss is updated directly. So no
> > > gross inaccuracies can result.
> >
> > Sure.  Take a million successive pagefaults and mm->rss is grossly
> > inaccurate.  Hence my suggestion that it be spilled into mm->rss
> > periodically.
> 
> It is spilled into mm->rss periodically. That is the whole point of the
> patch.
> 
> The timer tick occurs every 1 ms.

That only works if the task happens to have the CPU when the timer tick
occurs.  There remains no theoretical upper bound to the error in mm->rss,
and that's very easy to fix.

Re: deferred rss update instead of sloppy rss

[Christoph Lameter]

On Mon, 22 Nov 2004, Andrew Morton wrote:

> > The timer tick occurs every 1 ms.
>
> That only works if the task happens to have the CPU when the timer tick
> occurs.  There remains no theoretical upper bound to the error in mm->rss,
> and that's very easy to fix.

Page fault intensive programs typically hog the cpu. But you are in
principle right.

The "easy fix" that you propose is to add additional logic to some very
hot code paths. Then we are probably better off with another approach.

Re: page fault scalability patch V11 [0/7]: overview

[William Lee Irwin III]

On Mon, 22 Nov 2004, William Lee Irwin III wrote:
>> The specific patches you compared matter a great deal as there are
>> implementation blunders (e.g. poor placement of counters relative to
>> ->mmap_sem) that can ruin the results. URL's to the specific patches
>> would rule out that source of error.

On Mon, Nov 22, 2004 at 02:51:22PM -0800, Christoph Lameter wrote:
> I mentioned V4 of this patch which was posted to lkml. A simple search
> should get you there.

The counter's placement was poor in that version of the patch. The
results are very suspect and likely invalid. It would have been more
helpful if you provided some kind of unique identifier when requests
for complete disambiguation are made. For instance, the version tags of
your patches are not visible in Subject: lines.

There are, of course, other issues, e.g. where the arch sweeps went.
This discussion has degenerated into non-cooperation making it beyond
my power to help, and I'm in the midst of several rather urgent
bughunts, of which there are apparently more to come.


-- wli

page fault scalability patch V12 [0/7]: Overview and performance tests

[Christoph Lameter]

Changes from V11->V12 of this patch:
- dump sloppy_rss in favor of list_rss (Linus' proposal)
- keep up against current Linus tree (patch is based on 2.6.10-rc2-bk14)

This is a series of patches that increases the scalability of
the page fault handler for SMP. Here are some performance results
on a machine with 512 processors allocating 32 GB with an increasing
number of threads (that are assigned a processor each).

Without the patches:
Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32   3    1    1.416s    138.165s 139.050s 45073.831  45097.498
 32   3    2    1.397s    148.523s  78.044s 41965.149  80201.646
 32   3    4    1.390s    152.618s  44.044s 40851.258 141545.239
 32   3    8    1.500s    374.008s  53.001s 16754.519 118671.950
 32   3   16    1.415s   1051.759s  73.094s  5973.803  85087.358
 32   3   32    1.867s   3400.417s 117.003s  1849.186  53754.928
 32   3   64    5.361s  11633.040s 197.034s   540.577  31881.112
 32   3  128   23.387s  39386.390s 332.055s   159.642  18918.599
 32   3  256   15.409s  20031.450s 168.095s   313.837  37237.918
 32   3  512   18.720s  10338.511s  86.047s   607.446  72752.686

With the patches:
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32   3    1    1.451s    140.151s 141.060s 44430.367  44428.115
 32   3    2    1.399s    136.349s  73.041s 45673.303  85699.793
 32   3    4    1.321s    129.760s  39.027s 47996.303 160197.217
 32   3    8    1.279s    100.648s  20.039s 61724.641 308454.557
 32   3   16    1.414s    153.975s  15.090s 40488.236 395681.716
 32   3   32    2.534s    337.021s  17.016s 18528.487 366445.400
 32   3   64    4.271s    709.872s  18.057s  8809.787 338656.440
 32   3  128   18.734s   1805.094s  21.084s  3449.586 288005.644
 32   3  256   14.698s    963.787s  11.078s  6429.787 534077.540
 32   3  512   15.299s    453.990s   5.098s 13406.321 1050416.414

For more than 8 cpus the page fault rate increases by orders
of magnitude. For more than 64 cpus the improvement in performace
is 10 times better.

The performance increase is accomplished by avoiding the use of the
page_table_lock spinlock (but not mm->mmap_sem!) through new atomic
operations on pte's (ptep_xchg, ptep_cmpxchg) and on pmd and pgd's
(pgd_test_and_populate, pmd_test_and_populate).

The page table lock can be avoided in the following situations:

1. An empty pte or pmd entry is populated

This is safe since the swapper may only depopulate them and the
swapper code has been changed to never set a pte to be empty until the
page has been evicted. The population of an empty pte is frequent
if a process touches newly allocated memory.

2. Modifications of flags in a pte entry (write/accessed).

These modifications are done by the CPU or by low level handlers
on various platforms also bypassing the page_table_lock. So this
seems to be safe too.

One essential change in the VM is the use of pte_cmpxchg (or its
generic emulation) on page table entries before doing an
update_mmu_change without holding the page table lock. However, we do
similar things now with other atomic pte operations such as
ptep_get_and_clear and ptep_test_and_clear_dirty. These operations
clear a pte *after* doing an operation on it. The ptep_cmpxchg as used
in this patch operates on an *cleared* pte and replaces it with a pte
pointing to valid memory. The effect of this change on various
architectures has to be thought through. Local definitions of
ptep_cmpxchg and ptep_xchg may be necessary.

For IA64 an icache coherency issue may arise that potentially requires
the flushing of the icache (as done via update_mmu_cache on IA64) prior
to the use of ptep_cmpxchg. Similar issues may arise on other platforms.

The patch introduces a split counter for rss handling to avoid atomic
operations and locks currently necessary for rss modifications. In
addition to mm->rss, tsk->rss is introduced. tsk->rss is defined to be
in the same cache line as tsk->mm (which is already used by the fault
handler) and thus tsk->rss can be incremented without locks
in a fast way. The cache line does not need to be shared between
processors in the page table handler.

A tasklist is generated for each mm (rcu based). Values in that list
are added up to calculate rss or anon_rss values.

The patchset is composed of 7 patches:

1/7: Avoid page_table_lock in handle_mm_fault

   This patch defers the acquisition of the page_table_lock as much as
   possible and uses atomic operations for allocating anonymous memory.
   These atomic operations are simulated by acquiring the page_table_lock
   for very small time frames if an architecture does not define
   __HAVE_ARCH_ATOMIC_TABLE_OPS. It also changes the swapper so that a
   pte will not be set to empty if a page is in transition to swap.

   If only the first two patches are applied then the time that the
   page_table_lock is held is simply reduced. The lock may then be
   acquired multiple times during a page fault.

2/7: Atomic pte operations for ia64

3/7: Make cmpxchg generally available on i386

   The atomic operations on the page table rely heavily on cmpxchg
   instructions. This patch adds emulations for cmpxchg and cmpxchg8b
   for old 80386 and 80486 cpus. The emulations are only included if a
   kernel is build for these old cpus and are skipped for the real
   cmpxchg instructions if the kernel that is build for 386 or 486 is
   then run on a more recent cpu.

   This patch may be used independently of the other patches.

4/7: Atomic pte operations for i386

   A generally available cmpxchg (last patch) must be available for
   this patch to preserve the ability to build kernels for 386 and 486.

5/7: Atomic pte operation for x86_64

6/7: Atomic pte operations for s390

7/7: Split counter implementation for rss
  Add tsk->rss and tsk->anon_rss. Add tasklist. Add logic
  to calculate rss from tasklist.

There are some additional outstanding performance enhancements that are
not available yet but which require this patch. Those modifications
push the maximum page fault rate from ~ 1 mio faults per second as
shown above to above 3 mio faults per second.

The last editions of the sloppy rss, atomic rss and deferred rss patch
will be posted to linux-ia64 for archival purpose.

Signed-off-by: Christoph Lameter <clameter@sgi.com>

page fault scalability patch V12 [1/7]: Reduce use of thepage_table_lock

[Christoph Lameter]

Changelog
        * Increase parallelism in SMP configurations by deferring
          the acquisition of page_table_lock in handle_mm_fault
        * Anonymous memory page faults bypass the page_table_lock
          through the use of atomic page table operations
        * Swapper does not set pte to empty in transition to swap
        * Simulate atomic page table operations using the
          page_table_lock if an arch does not define
          __HAVE_ARCH_ATOMIC_TABLE_OPS. This still provides
          a performance benefit since the page_table_lock
          is held for shorter periods of time.

Signed-off-by: Christoph Lameter <clameter@sgi.com

Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-11-23 10:06:03.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-11-23 10:07:55.000000000 -0800
@@ -1330,8 +1330,7 @@
 }

 /*
- * We hold the mm semaphore and the page_table_lock on entry and
- * should release the pagetable lock on exit..
+ * We hold the mm semaphore
  */
 static int do_swap_page(struct mm_struct * mm,
 	struct vm_area_struct * vma, unsigned long address,
@@ -1343,15 +1342,13 @@
 	int ret = VM_FAULT_MINOR;

 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);
 	page = lookup_swap_cache(entry);
 	if (!page) {
  		swapin_readahead(entry, address, vma);
  		page = read_swap_cache_async(entry, vma, address);
 		if (!page) {
 			/*
-			 * Back out if somebody else faulted in this pte while
-			 * we released the page table lock.
+			 * Back out if somebody else faulted in this pte
 			 */
 			spin_lock(&mm->page_table_lock);
 			page_table = pte_offset_map(pmd, address);
@@ -1374,8 +1371,7 @@
 	lock_page(page);

 	/*
-	 * Back out if somebody else faulted in this pte while we
-	 * released the page table lock.
+	 * Back out if somebody else faulted in this pte
 	 */
 	spin_lock(&mm->page_table_lock);
 	page_table = pte_offset_map(pmd, address);
@@ -1422,14 +1418,12 @@
 }

 /*
- * We are called with the MM semaphore and page_table_lock
- * spinlock held to protect against concurrent faults in
- * multithreaded programs.
+ * We are called with the MM semaphore held.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 		pte_t *page_table, pmd_t *pmd, int write_access,
-		unsigned long addr)
+		unsigned long addr, pte_t orig_entry)
 {
 	pte_t entry;
 	struct page * page = ZERO_PAGE(addr);
@@ -1441,7 +1435,6 @@
 	if (write_access) {
 		/* Allocate our own private page. */
 		pte_unmap(page_table);
-		spin_unlock(&mm->page_table_lock);

 		if (unlikely(anon_vma_prepare(vma)))
 			goto no_mem;
@@ -1450,30 +1443,37 @@
 			goto no_mem;
 		clear_user_highpage(page, addr);

-		spin_lock(&mm->page_table_lock);
 		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
-			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
-		}
-		mm->rss++;
 		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
 							 vma->vm_page_prot)),
 				      vma);
-		lru_cache_add_active(page);
 		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}

-	set_pte(page_table, entry);
+	/* update the entry */
+	if (!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)) {
+		if (write_access) {
+			pte_unmap(page_table);
+			page_cache_release(page);
+		}
+		goto out;
+	}
+	if (write_access) {
+		/*
+		 * These two functions must come after the cmpxchg
+		 * because if the page is on the LRU then try_to_unmap may come
+		 * in and unmap the pte.
+		 */
+		lru_cache_add_active(page);
+		page_add_anon_rmap(page, vma, addr);
+		mm->rss++;
+
+	}
 	pte_unmap(page_table);

 	/* No need to invalidate - it was non-present before */
 	update_mmu_cache(vma, addr, entry);
-	spin_unlock(&mm->page_table_lock);
 out:
 	return VM_FAULT_MINOR;
 no_mem:
@@ -1489,12 +1489,12 @@
  * As this is called only for pages that do not currently exist, we
  * do not need to flush old virtual caches or the TLB.
  *
- * This is called with the MM semaphore held and the page table
- * spinlock held. Exit with the spinlock released.
+ * This is called with the MM semaphore held.
  */
 static int
 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-	unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *page_table,
+        pmd_t *pmd, pte_t orig_entry)
 {
 	struct page * new_page;
 	struct address_space *mapping = NULL;
@@ -1505,9 +1505,8 @@

 	if (!vma->vm_ops || !vma->vm_ops->nopage)
 		return do_anonymous_page(mm, vma, page_table,
-					pmd, write_access, address);
+					pmd, write_access, address, orig_entry);
 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);

 	if (vma->vm_file) {
 		mapping = vma->vm_file->f_mapping;
@@ -1605,7 +1604,7 @@
  * nonlinear vmas.
  */
 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
-	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd, pte_t entry)
 {
 	unsigned long pgoff;
 	int err;
@@ -1618,13 +1617,12 @@
 	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
-		return do_no_page(mm, vma, address, write_access, pte, pmd);
+		return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 	}

 	pgoff = pte_to_pgoff(*pte);

 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);

 	err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
 	if (err == -ENOMEM)
@@ -1643,49 +1641,40 @@
  * with external mmu caches can use to update those (ie the Sparc or
  * PowerPC hashed page tables that act as extended TLBs).
  *
- * Note the "page_table_lock". It is to protect against kswapd removing
- * pages from under us. Note that kswapd only ever _removes_ pages, never
- * adds them. As such, once we have noticed that the page is not present,
- * we can drop the lock early.
- *
- * The adding of pages is protected by the MM semaphore (which we hold),
- * so we don't need to worry about a page being suddenly been added into
- * our VM.
- *
- * We enter with the pagetable spinlock held, we are supposed to
- * release it when done.
+ * Note that kswapd only ever _removes_ pages, never adds them.
+ * We need to insure to handle that case properly.
  */
 static inline int handle_pte_fault(struct mm_struct *mm,
 	struct vm_area_struct * vma, unsigned long address,
 	int write_access, pte_t *pte, pmd_t *pmd)
 {
 	pte_t entry;
+	pte_t new_entry;

 	entry = *pte;
 	if (!pte_present(entry)) {
-		/*
-		 * If it truly wasn't present, we know that kswapd
-		 * and the PTE updates will not touch it later. So
-		 * drop the lock.
-		 */
 		if (pte_none(entry))
-			return do_no_page(mm, vma, address, write_access, pte, pmd);
+			return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 		if (pte_file(entry))
-			return do_file_page(mm, vma, address, write_access, pte, pmd);
+			return do_file_page(mm, vma, address, write_access, pte, pmd, entry);
 		return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
 	}

+	/*
+	 * This is the case in which we only update some bits in the pte.
+	 */
+	new_entry = pte_mkyoung(entry);
 	if (write_access) {
-		if (!pte_write(entry))
+		if (!pte_write(entry)) {
+			/* do_wp_page expects us to hold the page_table_lock */
+			spin_lock(&mm->page_table_lock);
 			return do_wp_page(mm, vma, address, pte, pmd, entry);
-
-		entry = pte_mkdirty(entry);
+		}
+		new_entry = pte_mkdirty(new_entry);
 	}
-	entry = pte_mkyoung(entry);
-	ptep_set_access_flags(vma, address, pte, entry, write_access);
-	update_mmu_cache(vma, address, entry);
+	if (ptep_cmpxchg(vma, address, pte, entry, new_entry))
+		update_mmu_cache(vma, address, new_entry);
 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_MINOR;
 }

@@ -1703,22 +1692,45 @@

 	inc_page_state(pgfault);

-	if (is_vm_hugetlb_page(vma))
+	if (unlikely(is_vm_hugetlb_page(vma)))
 		return VM_FAULT_SIGBUS;	/* mapping truncation does this. */

 	/*
-	 * We need the page table lock to synchronize with kswapd
-	 * and the SMP-safe atomic PTE updates.
+	 * We rely on the mmap_sem and the SMP-safe atomic PTE updates.
+	 * to synchronize with kswapd
 	 */
-	spin_lock(&mm->page_table_lock);
-	pmd = pmd_alloc(mm, pgd, address);
+	if (unlikely(pgd_none(*pgd))) {
+		pmd_t *new = pmd_alloc_one(mm, address);
+		if (!new)
+			return VM_FAULT_OOM;
+
+		/* Insure that the update is done in an atomic way */
+		if (!pgd_test_and_populate(mm, pgd, new))
+			pmd_free(new);
+	}
+
+	pmd = pmd_offset(pgd, address);
+
+	if (likely(pmd)) {
+		pte_t *pte;
+
+		if (!pmd_present(*pmd)) {
+			struct page *new;

-	if (pmd) {
-		pte_t * pte = pte_alloc_map(mm, pmd, address);
-		if (pte)
+			new = pte_alloc_one(mm, address);
+			if (!new)
+				return VM_FAULT_OOM;
+
+			if (!pmd_test_and_populate(mm, pmd, new))
+				pte_free(new);
+			else
+				inc_page_state(nr_page_table_pages);
+		}
+
+		pte = pte_offset_map(pmd, address);
+		if (likely(pte))
 			return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
 	}
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_OOM;
 }

Index: linux-2.6.9/include/asm-generic/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-generic/pgtable.h	2004-10-18 14:53:46.000000000 -0700
+++ linux-2.6.9/include/asm-generic/pgtable.h	2004-11-23 10:06:12.000000000 -0800
@@ -134,4 +134,60 @@
 #define pgd_offset_gate(mm, addr)	pgd_offset(mm, addr)
 #endif

+#ifndef __HAVE_ARCH_ATOMIC_TABLE_OPS
+/*
+ * If atomic page table operations are not available then use
+ * the page_table_lock to insure some form of locking.
+ * Note thought that low level operations as well as the
+ * page_table_handling of the cpu may bypass all locking.
+ */
+
+#ifndef __HAVE_ARCH_PTEP_CMPXCHG
+#define ptep_cmpxchg(__vma, __addr, __ptep, __oldval, __newval)		\
+({									\
+	int __rc;							\
+	spin_lock(&__vma->vm_mm->page_table_lock);			\
+	__rc = pte_same(*(__ptep), __oldval);				\
+	if (__rc) set_pte(__ptep, __newval);				\
+	spin_unlock(&__vma->vm_mm->page_table_lock);			\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_ARCH_PGP_TEST_AND_POPULATE
+#define pgd_test_and_populate(__mm, __pgd, __pmd)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pgd_present(*(__pgd));					\
+	if (__rc) pgd_populate(__mm, __pgd, __pmd);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_PMD_TEST_AND_POPULATE
+#define pmd_test_and_populate(__mm, __pmd, __page)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pmd_present(*(__pmd));					\
+	if (__rc) pmd_populate(__mm, __pmd, __page);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#endif
+
+#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH
+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)		\
+({									\
+	pte_t __p = __pte(xchg(&pte_val(*(__ptep)), pte_val(__pteval)));\
+	flush_tlb_page(__vma, __address);				\
+	__p;								\
+})
+
+#endif
+
 #endif /* _ASM_GENERIC_PGTABLE_H */
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-11-23 10:06:03.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-11-23 10:06:12.000000000 -0800
@@ -424,7 +424,10 @@
  * @vma:	the vm area in which the mapping is added
  * @address:	the user virtual address mapped
  *
- * The caller needs to hold the mm->page_table_lock.
+ * The caller needs to hold the mm->page_table_lock if page
+ * is pointing to something that is known by the vm.
+ * The lock does not need to be held if page is pointing
+ * to a newly allocated page.
  */
 void page_add_anon_rmap(struct page *page,
 	struct vm_area_struct *vma, unsigned long address)
@@ -568,11 +571,6 @@

 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address);
-	pteval = ptep_clear_flush(vma, address, pte);
-
-	/* Move the dirty bit to the physical page now the pte is gone. */
-	if (pte_dirty(pteval))
-		set_page_dirty(page);

 	if (PageAnon(page)) {
 		swp_entry_t entry = { .val = page->private };
@@ -587,11 +585,15 @@
 			list_add(&mm->mmlist, &init_mm.mmlist);
 			spin_unlock(&mmlist_lock);
 		}
-		set_pte(pte, swp_entry_to_pte(entry));
+		pteval = ptep_xchg_flush(vma, address, pte, swp_entry_to_pte(entry));
 		BUG_ON(pte_file(*pte));
 		mm->anon_rss--;
-	}
+	} else
+		pteval = ptep_clear_flush(vma, address, pte);

+	/* Move the dirty bit to the physical page now the pte is gone. */
+	if (pte_dirty(pteval))
+		set_page_dirty(page);
 	mm->rss--;
 	page_remove_rmap(page);
 	page_cache_release(page);
@@ -678,15 +680,21 @@
 		if (ptep_clear_flush_young(vma, address, pte))
 			continue;

-		/* Nuke the page table entry. */
 		flush_cache_page(vma, address);
-		pteval = ptep_clear_flush(vma, address, pte);
+		/*
+		 * There would be a race here with handle_mm_fault and do_anonymous_page
+		 * which  bypasses the page_table_lock if we would zap the pte before
+		 * putting something into it. On the other hand we need to
+		 * have the dirty flag setting at the time we replaced the value.
+		 */

 		/* If nonlinear, store the file page offset in the pte. */
 		if (page->index != linear_page_index(vma, address))
-			set_pte(pte, pgoff_to_pte(page->index));
+			pteval = ptep_xchg_flush(vma, address, pte, pgoff_to_pte(page->index));
+		else
+			pteval = ptep_get_and_clear(pte);

-		/* Move the dirty bit to the physical page now the pte is gone. */
+		/* Move the dirty bit to the physical page now that the pte is gone. */
 		if (pte_dirty(pteval))
 			set_page_dirty(page);

page fault scalability patch V12 [2/7]: atomic pte operations for ia64

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for ia64
        * Enhanced parallelism in page fault handler if applied together
          with the generic patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-ia64/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-ia64/pgalloc.h	2004-10-18 14:53:06.000000000 -0700
+++ linux-2.6.9/include/asm-ia64/pgalloc.h	2004-11-19 07:54:19.000000000 -0800
@@ -34,6 +34,10 @@
 #define pmd_quicklist		(local_cpu_data->pmd_quick)
 #define pgtable_cache_size	(local_cpu_data->pgtable_cache_sz)

+/* Empty entries of PMD and PGD */
+#define PMD_NONE       0
+#define PGD_NONE       0
+
 static inline pgd_t*
 pgd_alloc_one_fast (struct mm_struct *mm)
 {
@@ -78,12 +82,19 @@
 	preempt_enable();
 }

+
 static inline void
 pgd_populate (struct mm_struct *mm, pgd_t *pgd_entry, pmd_t *pmd)
 {
 	pgd_val(*pgd_entry) = __pa(pmd);
 }

+/* Atomic populate */
+static inline int
+pgd_test_and_populate (struct mm_struct *mm, pgd_t *pgd_entry, pmd_t *pmd)
+{
+	return ia64_cmpxchg8_acq(pgd_entry,__pa(pmd), PGD_NONE) == PGD_NONE;
+}

 static inline pmd_t*
 pmd_alloc_one_fast (struct mm_struct *mm, unsigned long addr)
@@ -132,6 +143,13 @@
 	pmd_val(*pmd_entry) = page_to_phys(pte);
 }

+/* Atomic populate */
+static inline int
+pmd_test_and_populate (struct mm_struct *mm, pmd_t *pmd_entry, struct page *pte)
+{
+	return ia64_cmpxchg8_acq(pmd_entry, page_to_phys(pte), PMD_NONE) == PMD_NONE;
+}
+
 static inline void
 pmd_populate_kernel (struct mm_struct *mm, pmd_t *pmd_entry, pte_t *pte)
 {
Index: linux-2.6.9/include/asm-ia64/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-ia64/pgtable.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-ia64/pgtable.h	2004-11-19 07:55:35.000000000 -0800
@@ -30,6 +30,8 @@
 #define _PAGE_P_BIT		0
 #define _PAGE_A_BIT		5
 #define _PAGE_D_BIT		6
+#define _PAGE_IG_BITS          53
+#define _PAGE_LOCK_BIT         (_PAGE_IG_BITS+3)       /* bit 56. Aligned to 8 bits */

 #define _PAGE_P			(1 << _PAGE_P_BIT)	/* page present bit */
 #define _PAGE_MA_WB		(0x0 <<  2)	/* write back memory attribute */
@@ -58,6 +60,7 @@
 #define _PAGE_PPN_MASK		(((__IA64_UL(1) << IA64_MAX_PHYS_BITS) - 1) & ~0xfffUL)
 #define _PAGE_ED		(__IA64_UL(1) << 52)	/* exception deferral */
 #define _PAGE_PROTNONE		(__IA64_UL(1) << 63)
+#define _PAGE_LOCK		(__IA64_UL(1) << _PAGE_LOCK_BIT)

 /* Valid only for a PTE with the present bit cleared: */
 #define _PAGE_FILE		(1 << 1)		/* see swap & file pte remarks below */
@@ -270,6 +273,8 @@
 #define pte_dirty(pte)		((pte_val(pte) & _PAGE_D) != 0)
 #define pte_young(pte)		((pte_val(pte) & _PAGE_A) != 0)
 #define pte_file(pte)		((pte_val(pte) & _PAGE_FILE) != 0)
+#define pte_locked(pte)		((pte_val(pte) & _PAGE_LOCK)!=0)
+
 /*
  * Note: we convert AR_RWX to AR_RX and AR_RW to AR_R by clearing the 2nd bit in the
  * access rights:
@@ -281,8 +286,15 @@
 #define pte_mkyoung(pte)	(__pte(pte_val(pte) | _PAGE_A))
 #define pte_mkclean(pte)	(__pte(pte_val(pte) & ~_PAGE_D))
 #define pte_mkdirty(pte)	(__pte(pte_val(pte) | _PAGE_D))
+#define pte_mkunlocked(pte)	(__pte(pte_val(pte) & ~_PAGE_LOCK))

 /*
+ * Lock functions for pte's
+ */
+#define ptep_lock(ptep)		test_and_set_bit(_PAGE_LOCK_BIT, ptep)
+#define ptep_unlock(ptep)	{ clear_bit(_PAGE_LOCK_BIT,ptep); smp_mb__after_clear_bit(); }
+#define ptep_unlock_set(ptep, val) set_pte(ptep, pte_mkunlocked(val))
+/*
  * Macro to a page protection value as "uncacheable".  Note that "protection" is really a
  * misnomer here as the protection value contains the memory attribute bits, dirty bits,
  * and various other bits as well.
@@ -342,7 +354,6 @@
 #define pte_unmap_nested(pte)		do { } while (0)

 /* atomic versions of the some PTE manipulations: */
-
 static inline int
 ptep_test_and_clear_young (pte_t *ptep)
 {
@@ -414,6 +425,26 @@
 #endif
 }

+/*
+ * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
+ * information.  However, we use this routine to take care of any (delayed) i-cache
+ * flushing that may be necessary.
+ */
+extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
+
+static inline int
+ptep_cmpxchg (struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	/*
+	 * IA64 defers icache flushes. If the new pte is executable we may
+	 * have to flush the icache to insure cache coherency immediately
+	 * after the cmpxchg.
+	 */
+	if (pte_exec(newval))
+		update_mmu_cache(vma, addr, newval);
+	return ia64_cmpxchg8_acq(&ptep->pte, newval.pte, oldval.pte) == oldval.pte;
+}
+
 static inline int
 pte_same (pte_t a, pte_t b)
 {
@@ -476,13 +507,6 @@
 	struct vm_area_struct * prev, unsigned long start, unsigned long end);
 #endif

-/*
- * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
- * information.  However, we use this routine to take care of any (delayed) i-cache
- * flushing that may be necessary.
- */
-extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
-
 #define __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Update PTEP with ENTRY, which is guaranteed to be a less
@@ -560,6 +584,8 @@
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
 #define __HAVE_ARCH_PGD_OFFSET_GATE
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+#define __HAVE_ARCH_LOCK_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _ASM_IA64_PGTABLE_H */

page fault scalability patch V12 [3/7]: universal cmpxchg for i386

[Christoph Lameter]

Changelog
        * Make cmpxchg and cmpxchg8b generally available on the i386
	  platform.
        * Provide emulation of cmpxchg suitable for uniprocessor if
	  build and run on 386.
        * Provide emulation of cmpxchg8b suitable for uniprocessor
	  systems if build and run on 386 or 486.
	* Provide an inline function to atomically get a 64 bit value
	  via cmpxchg8b in an SMP system (courtesy of Nick Piggin)
	  (important for i386 PAE mode and other places where atomic
	  64 bit operations are useful)

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/arch/i386/Kconfig
===================================================================
--- linux-2.6.9.orig/arch/i386/Kconfig	2004-11-15 11:13:34.000000000 -0800
+++ linux-2.6.9/arch/i386/Kconfig	2004-11-19 10:02:54.000000000 -0800
@@ -351,6 +351,11 @@
 	depends on !M386
 	default y

+config X86_CMPXCHG8B
+	bool
+	depends on !M386 && !M486
+	default y
+
 config X86_XADD
 	bool
 	depends on !M386
Index: linux-2.6.9/arch/i386/kernel/cpu/intel.c
===================================================================
--- linux-2.6.9.orig/arch/i386/kernel/cpu/intel.c	2004-11-15 11:13:34.000000000 -0800
+++ linux-2.6.9/arch/i386/kernel/cpu/intel.c	2004-11-19 10:38:26.000000000 -0800
@@ -6,6 +6,7 @@
 #include <linux/bitops.h>
 #include <linux/smp.h>
 #include <linux/thread_info.h>
+#include <linux/module.h>

 #include <asm/processor.h>
 #include <asm/msr.h>
@@ -287,5 +288,103 @@
 	return 0;
 }

+#ifndef CONFIG_X86_CMPXCHG
+unsigned long cmpxchg_386_u8(volatile void *ptr, u8 old, u8 new)
+{
+	u8 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u8));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u8 *)ptr;
+	if (prev == old)
+		*(u8 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u8);
+
+unsigned long cmpxchg_386_u16(volatile void *ptr, u16 old, u16 new)
+{
+	u16 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u16));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u16 *)ptr;
+	if (prev == old)
+		*(u16 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u16);
+
+unsigned long cmpxchg_386_u32(volatile void *ptr, u32 old, u32 new)
+{
+	u32 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u32));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u32 *)ptr;
+	if (prev == old)
+		*(u32 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u32);
+#endif
+
+#ifndef CONFIG_X86_CMPXCHG8B
+unsigned long long cmpxchg8b_486(volatile unsigned long long *ptr,
+	       unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	unsigned long flags;
+
+	/*
+	 * Check if the kernel was compiled for an old cpu but
+	 * we are running really on a cpu capable of cmpxchg8b
+	 */
+
+	if (cpu_has(cpu_data, X86_FEATURE_CX8))
+		return __cmpxchg8b(ptr, old, newv);
+
+	/* Poor mans cmpxchg8b for 386 and 486. Not suitable for SMP */
+	local_irq_save(flags);
+	prev = *ptr;
+	if (prev == old)
+		*ptr = newv;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg8b_486);
+#endif
+
 // arch_initcall(intel_cpu_init);

Index: linux-2.6.9/include/asm-i386/system.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/system.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-i386/system.h	2004-11-19 10:49:46.000000000 -0800
@@ -149,6 +149,9 @@
 #define __xg(x) ((struct __xchg_dummy *)(x))


+#define ll_low(x)	*(((unsigned int*)&(x))+0)
+#define ll_high(x)	*(((unsigned int*)&(x))+1)
+
 /*
  * The semantics of XCHGCMP8B are a bit strange, this is why
  * there is a loop and the loading of %%eax and %%edx has to
@@ -184,8 +187,6 @@
 {
 	__set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
 }
-#define ll_low(x)	*(((unsigned int*)&(x))+0)
-#define ll_high(x)	*(((unsigned int*)&(x))+1)

 static inline void __set_64bit_var (unsigned long long *ptr,
 			 unsigned long long value)
@@ -203,6 +204,26 @@
  __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
  __set_64bit(ptr, ll_low(value), ll_high(value)) )

+static inline unsigned long long __get_64bit(unsigned long long * ptr)
+{
+	unsigned long long ret;
+	__asm__ __volatile__ (
+		"\n1:\t"
+		"movl (%1), %%eax\n\t"
+		"movl 4(%1), %%edx\n\t"
+		"movl %%eax, %%ebx\n\t"
+		"movl %%edx, %%ecx\n\t"
+		LOCK_PREFIX "cmpxchg8b (%1)\n\t"
+		"jnz 1b"
+		:	"=A"(ret)
+		:	"D"(ptr)
+		:	"ebx", "ecx", "memory");
+	return ret;
+}
+
+#define get_64bit(ptr) __get_64bit(ptr)
+
+
 /*
  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
  * Note 2: xchg has side effect, so that attribute volatile is necessary,
@@ -240,7 +261,41 @@
  */

 #ifdef CONFIG_X86_CMPXCHG
+
 #define __HAVE_ARCH_CMPXCHG 1
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))__cmpxchg((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
+
+#else
+
+/*
+ * Building a kernel capable running on 80386. It may be necessary to
+ * simulate the cmpxchg on the 80386 CPU. For that purpose we define
+ * a function for each of the sizes we support.
+ */
+
+extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
+extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
+extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
+
+static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
+				      unsigned long new, int size)
+{
+	switch (size) {
+	case 1:
+		return cmpxchg_386_u8(ptr, old, new);
+	case 2:
+		return cmpxchg_386_u16(ptr, old, new);
+	case 4:
+		return cmpxchg_386_u32(ptr, old, new);
+	}
+	return old;
+}
+
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))cmpxchg_386((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
 #endif

 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
@@ -270,10 +325,32 @@
 	return old;
 }

-#define cmpxchg(ptr,o,n)\
-	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
-					(unsigned long)(n),sizeof(*(ptr))))
-
+static inline unsigned long long __cmpxchg8b(volatile unsigned long long *ptr,
+		unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	__asm__ __volatile__(
+	LOCK_PREFIX "cmpxchg8b (%4)"
+		: "=A" (prev)
+		: "0" (old), "c" ((unsigned long)(newv >> 32)),
+		  "b" ((unsigned long)(newv & 0xffffffffULL)), "D" (ptr)
+		: "memory");
+	return prev;
+}
+
+#ifdef CONFIG_X86_CMPXCHG8B
+#define cmpxchg8b __cmpxchg8b
+#else
+/*
+ * Building a kernel capable of running on 80486 and 80386. Both
+ * do not support cmpxchg8b. Call a function that emulates the
+ * instruction if necessary.
+ */
+extern unsigned long long cmpxchg8b_486(volatile unsigned long long *,
+		unsigned long long, unsigned long long);
+#define cmpxchg8b cmpxchg8b_486
+#endif
+
 #ifdef __KERNEL__
 struct alt_instr {
 	__u8 *instr; 		/* original instruction */

page fault scalability patch V12 [4/7]: atomic pte operations for i386

[Christoph Lameter]

Changelog
	* Atomic pte operations for i386 in regular and PAE modes

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-i386/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-i386/pgtable.h	2004-11-19 10:05:27.000000000 -0800
@@ -413,6 +413,7 @@
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _I386_PGTABLE_H */
Index: linux-2.6.9/include/asm-i386/pgtable-3level.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable-3level.h	2004-10-18 14:54:55.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgtable-3level.h	2004-11-19 10:10:06.000000000 -0800
@@ -6,7 +6,8 @@
  * tables on PPro+ CPUs.
  *
  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
- */
+ * August 26, 2004 added ptep_cmpxchg <christoph@lameter.com>
+*/

 #define pte_ERROR(e) \
 	printk("%s:%d: bad pte %p(%08lx%08lx).\n", __FILE__, __LINE__, &(e), (e).pte_high, (e).pte_low)
@@ -42,26 +43,15 @@
 	return pte_x(pte);
 }

-/* Rules for using set_pte: the pte being assigned *must* be
- * either not present or in a state where the hardware will
- * not attempt to update the pte.  In places where this is
- * not possible, use pte_get_and_clear to obtain the old pte
- * value and then use set_pte to update it.  -ben
- */
-static inline void set_pte(pte_t *ptep, pte_t pte)
-{
-	ptep->pte_high = pte.pte_high;
-	smp_wmb();
-	ptep->pte_low = pte.pte_low;
-}
-#define __HAVE_ARCH_SET_PTE_ATOMIC
-#define set_pte_atomic(pteptr,pteval) \
+#define set_pte(pteptr,pteval) \
 		set_64bit((unsigned long long *)(pteptr),pte_val(pteval))
 #define set_pmd(pmdptr,pmdval) \
 		set_64bit((unsigned long long *)(pmdptr),pmd_val(pmdval))
 #define set_pgd(pgdptr,pgdval) \
 		set_64bit((unsigned long long *)(pgdptr),pgd_val(pgdval))

+#define set_pte_atomic set_pte
+
 /*
  * Pentium-II erratum A13: in PAE mode we explicitly have to flush
  * the TLB via cr3 if the top-level pgd is changed...
@@ -142,4 +132,23 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t){ (pte).pte_high })
 #define __swp_entry_to_pte(x)		((pte_t){ 0, (x).val })

+/* Atomic PTE operations */
+#define ptep_xchg_flush(__vma, __addr, __ptep, __newval) \
+({	pte_t __r;							\
+	/* xchg acts as a barrier before the setting of the high bits. */\
+	__r.pte_low = xchg(&(__ptep)->pte_low, (__newval).pte_low);	\
+	__r.pte_high = (__ptep)->pte_high;				\
+	(__ptep)->pte_high = (__newval).pte_high;			\
+	flush_tlb_page(__vma, __addr);					\
+	(__r);								\
+})
+
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
+
+static inline int ptep_cmpxchg(struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg((unsigned int *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
+
 #endif /* _I386_PGTABLE_3LEVEL_H */
Index: linux-2.6.9/include/asm-i386/pgtable-2level.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable-2level.h	2004-10-18 14:54:31.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgtable-2level.h	2004-11-19 10:05:27.000000000 -0800
@@ -82,4 +82,7 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t) { (pte).pte_low })
 #define __swp_entry_to_pte(x)		((pte_t) { (x).val })

+/* Atomic PTE operations */
+#define ptep_cmpxchg(__vma,__a,__xp,__oldpte,__newpte) (cmpxchg(&(__xp)->pte_low, (__oldpte).pte_low, (__newpte).pte_low)==(__oldpte).pte_low)
+
 #endif /* _I386_PGTABLE_2LEVEL_H */
Index: linux-2.6.9/include/asm-i386/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgalloc.h	2004-10-18 14:53:10.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgalloc.h	2004-11-19 10:10:40.000000000 -0800
@@ -4,9 +4,12 @@
 #include <linux/config.h>
 #include <asm/processor.h>
 #include <asm/fixmap.h>
+#include <asm/system.h>
 #include <linux/threads.h>
 #include <linux/mm.h>		/* for struct page */

+#define PMD_NONE 0L
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte)))

@@ -16,6 +19,19 @@
 		((unsigned long long)page_to_pfn(pte) <<
 			(unsigned long long) PAGE_SHIFT)));
 }
+
+/* Atomic version */
+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+#ifdef CONFIG_X86_PAE
+	return cmpxchg8b( ((unsigned long long *)pmd), PMD_NONE, _PAGE_TABLE +
+		((unsigned long long)page_to_pfn(pte) <<
+			(unsigned long long) PAGE_SHIFT) ) == PMD_NONE;
+#else
+	return cmpxchg( (unsigned long *)pmd, PMD_NONE, _PAGE_TABLE + (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+#endif
+}
+
 /*
  * Allocate and free page tables.
  */
@@ -49,6 +65,7 @@
 #define pmd_free(x)			do { } while (0)
 #define __pmd_free_tlb(tlb,x)		do { } while (0)
 #define pgd_populate(mm, pmd, pte)	BUG()
+#define pgd_test_and_populate(mm, pmd, pte)	({ BUG(); 1; })

 #define check_pgt_cache()	do { } while (0)

page fault scalability patch V12 [5/7]: atomic pte operations for x86_64

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for x86_64

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-x86_64/pgalloc.h	2004-10-18 14:54:30.000000000 -0700
+++ linux-2.6.9/include/asm-x86_64/pgalloc.h	2004-11-23 10:59:01.000000000 -0800
@@ -7,16 +7,26 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pgd_populate(mm, pgd, pmd) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pmd)))
+#define pgd_test_and_populate(mm, pgd, pmd) \
+		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pmd)) == PGD_NONE)

 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
 	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
 }

+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+}
+
 extern __inline__ pmd_t *get_pmd(void)
 {
 	return (pmd_t *)get_zeroed_page(GFP_KERNEL);
Index: linux-2.6.9/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-x86_64/pgtable.h	2004-11-22 15:08:43.000000000 -0800
+++ linux-2.6.9/include/asm-x86_64/pgtable.h	2004-11-23 10:59:01.000000000 -0800
@@ -437,6 +437,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

page fault scalability patch V12 [6/7]: atomic pte operations for s390

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for s390

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-s390/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-s390/pgtable.h	2004-10-18 14:54:55.000000000 -0700
+++ linux-2.6.9/include/asm-s390/pgtable.h	2004-11-19 11:35:08.000000000 -0800
@@ -567,6 +567,15 @@
 	return pte;
 }

+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)            \
+({                                                                     \
+	struct mm_struct *__mm = __vma->vm_mm;                          \
+	pte_t __pte;                                                    \
+	__pte = ptep_clear_flush(__vma, __address, __ptep);             \
+	set_pte(__ptep, __pteval);                                      \
+	__pte;                                                          \
+})
+
 static inline void ptep_set_wrprotect(pte_t *ptep)
 {
 	pte_t old_pte = *ptep;
@@ -778,6 +787,14 @@

 #define kern_addr_valid(addr)   (1)

+/* Atomic PTE operations */
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
+static inline int ptep_cmpxchg (struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg(ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
 /*
  * No page table caches to initialise
  */
@@ -791,6 +808,7 @@
 #define __HAVE_ARCH_PTEP_CLEAR_DIRTY_FLUSH
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR
 #define __HAVE_ARCH_PTEP_CLEAR_FLUSH
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
Index: linux-2.6.9/include/asm-s390/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-s390/pgalloc.h	2004-10-18 14:54:37.000000000 -0700
+++ linux-2.6.9/include/asm-s390/pgalloc.h	2004-11-19 11:33:25.000000000 -0800
@@ -97,6 +97,10 @@
 	pgd_val(*pgd) = _PGD_ENTRY | __pa(pmd);
 }

+static inline int pgd_test_and_populate(struct mm_struct *mm, pdg_t *pgd, pmd_t *pmd)
+{
+	return cmpxchg(pgd, _PAGE_TABLE_INV, _PGD_ENTRY | __pa(pmd)) == _PAGE_TABLE_INV;
+}
 #endif /* __s390x__ */

 static inline void
@@ -119,6 +123,18 @@
 	pmd_populate_kernel(mm, pmd, (pte_t *)((page-mem_map) << PAGE_SHIFT));
 }

+static inline int
+pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *page)
+{
+	int rc;
+	spin_lock(&mm->page_table_lock);
+
+	rc=pte_same(*pmd, _PAGE_INVALID_EMPTY);
+	if (rc) pmd_populate(mm, pmd, page);
+	spin_unlock(&mm->page_table_lock);
+	return rc;
+}
+
 /*
  * page table entry allocation/free routines.
  */

page fault scalability patch V12 [7/7]: Split counter for rss

[Christoph Lameter]

Changelog
	* Split rss counter into the task structure
	* remove 3 checks of rss in mm/rmap.c
	* Prerequisite for page table scalability patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-11-30 20:33:31.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-11-30 20:33:50.000000000 -0800
@@ -30,6 +30,7 @@
 #include <linux/pid.h>
 #include <linux/percpu.h>
 #include <linux/topology.h>
+#include <linux/rcupdate.h>

 struct exec_domain;

@@ -217,6 +218,7 @@
 	int map_count;				/* number of VMAs */
 	struct rw_semaphore mmap_sem;
 	spinlock_t page_table_lock;		/* Protects page tables, mm->rss, mm->anon_rss */
+	long rss, anon_rss;

 	struct list_head mmlist;		/* List of maybe swapped mm's.  These are globally strung
 						 * together off init_mm.mmlist, and are protected
@@ -226,7 +228,7 @@
 	unsigned long start_code, end_code, start_data, end_data;
 	unsigned long start_brk, brk, start_stack;
 	unsigned long arg_start, arg_end, env_start, env_end;
-	unsigned long rss, anon_rss, total_vm, locked_vm, shared_vm;
+	unsigned long total_vm, locked_vm, shared_vm;
 	unsigned long exec_vm, stack_vm, reserved_vm, def_flags, nr_ptes;

 	unsigned long saved_auxv[42]; /* for /proc/PID/auxv */
@@ -236,6 +238,8 @@

 	/* Architecture-specific MM context */
 	mm_context_t context;
+	struct list_head task_list;		/* Tasks using this mm */
+	struct rcu_head rcu_head;		/* For freeing mm via rcu */

 	/* Token based thrashing protection. */
 	unsigned long swap_token_time;
@@ -545,6 +549,9 @@
 	struct list_head ptrace_list;

 	struct mm_struct *mm, *active_mm;
+	/* Split counters from mm */
+	long rss;
+	long anon_rss;

 /* task state */
 	struct linux_binfmt *binfmt;
@@ -578,6 +585,9 @@
 	struct completion *vfork_done;		/* for vfork() */
 	int __user *set_child_tid;		/* CLONE_CHILD_SETTID */
 	int __user *clear_child_tid;		/* CLONE_CHILD_CLEARTID */
+
+	/* List of other tasks using the same mm */
+	struct list_head mm_tasks;

 	unsigned long rt_priority;
 	unsigned long it_real_value, it_prof_value, it_virt_value;
@@ -1111,6 +1121,14 @@

 #endif

+unsigned long get_rss(struct mm_struct *mm);
+unsigned long get_anon_rss(struct mm_struct *mm);
+unsigned long get_shared(struct mm_struct *mm);
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk);
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk);
+
 #endif /* __KERNEL__ */

 #endif
+
Index: linux-2.6.9/fs/proc/task_mmu.c
===================================================================
--- linux-2.6.9.orig/fs/proc/task_mmu.c	2004-11-30 20:33:26.000000000 -0800
+++ linux-2.6.9/fs/proc/task_mmu.c	2004-11-30 20:33:50.000000000 -0800
@@ -22,7 +22,7 @@
 		"VmPTE:\t%8lu kB\n",
 		(mm->total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),
-		mm->rss << (PAGE_SHIFT-10),
+		get_rss(mm) << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
@@ -37,7 +37,7 @@
 int task_statm(struct mm_struct *mm, int *shared, int *text,
 	       int *data, int *resident)
 {
-	*shared = mm->rss - mm->anon_rss;
+	*shared = get_shared(mm);
 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
 								>> PAGE_SHIFT;
 	*data = mm->total_vm - mm->shared_vm;
Index: linux-2.6.9/fs/proc/array.c
===================================================================
--- linux-2.6.9.orig/fs/proc/array.c	2004-11-30 20:33:26.000000000 -0800
+++ linux-2.6.9/fs/proc/array.c	2004-11-30 20:33:50.000000000 -0800
@@ -420,7 +420,7 @@
 		jiffies_to_clock_t(task->it_real_value),
 		start_time,
 		vsize,
-		mm ? mm->rss : 0, /* you might want to shift this left 3 */
+		mm ? get_rss(mm) : 0, /* you might want to shift this left 3 */
 	        rsslim,
 		mm ? mm->start_code : 0,
 		mm ? mm->end_code : 0,
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-11-30 20:33:46.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-11-30 20:33:50.000000000 -0800
@@ -263,8 +263,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -438,7 +436,7 @@
 	BUG_ON(PageReserved(page));
 	BUG_ON(!anon_vma);

-	vma->vm_mm->anon_rss++;
+	current->anon_rss++;

 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 	index = (address - vma->vm_start) >> PAGE_SHIFT;
@@ -510,8 +508,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -799,8 +795,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;
Index: linux-2.6.9/kernel/fork.c
===================================================================
--- linux-2.6.9.orig/kernel/fork.c	2004-11-30 20:33:42.000000000 -0800
+++ linux-2.6.9/kernel/fork.c	2004-11-30 20:33:50.000000000 -0800
@@ -151,6 +151,7 @@
 	*tsk = *orig;
 	tsk->thread_info = ti;
 	ti->task = tsk;
+	tsk->rss = 0;

 	/* One for us, one for whoever does the "release_task()" (usually parent) */
 	atomic_set(&tsk->usage,2);
@@ -292,6 +293,7 @@
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);
 	INIT_LIST_HEAD(&mm->mmlist);
+	INIT_LIST_HEAD(&mm->task_list);
 	mm->core_waiters = 0;
 	mm->nr_ptes = 0;
 	spin_lock_init(&mm->page_table_lock);
@@ -323,6 +325,13 @@
 	return mm;
 }

+static void rcu_free_mm(struct rcu_head *head)
+{
+	struct mm_struct *mm = container_of(head ,struct mm_struct, rcu_head);
+
+	free_mm(mm);
+}
+
 /*
  * Called when the last reference to the mm
  * is dropped: either by a lazy thread or by
@@ -333,7 +342,7 @@
 	BUG_ON(mm == &init_mm);
 	mm_free_pgd(mm);
 	destroy_context(mm);
-	free_mm(mm);
+	call_rcu(&mm->rcu_head, rcu_free_mm);
 }

 /*
@@ -400,6 +409,8 @@

 	/* Get rid of any cached register state */
 	deactivate_mm(tsk, mm);
+	if (mm)
+		mm_remove_thread(mm, tsk);

 	/* notify parent sleeping on vfork() */
 	if (vfork_done) {
@@ -447,8 +458,8 @@
 		 * new threads start up in user mode using an mm, which
 		 * allows optimizing out ipis; the tlb_gather_mmu code
 		 * is an example.
+		 * (mm_add_thread does use the ptl .... )
 		 */
-		spin_unlock_wait(&oldmm->page_table_lock);
 		goto good_mm;
 	}

@@ -470,6 +481,7 @@
 		goto free_pt;

 good_mm:
+	mm_add_thread(mm, tsk);
 	tsk->mm = mm;
 	tsk->active_mm = mm;
 	return 0;
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-11-30 20:33:46.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-11-30 20:33:50.000000000 -0800
@@ -1467,7 +1467,7 @@
 		 */
 		lru_cache_add_active(page);
 		page_add_anon_rmap(page, vma, addr);
-		mm->rss++;
+		current->rss++;

 	}
 	pte_unmap(page_table);
@@ -1859,3 +1859,87 @@
 }

 #endif
+
+unsigned long get_rss(struct mm_struct *mm)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss;
+
+	if (!mm)
+		return 0;
+
+	rcu_read_lock();
+	rss = mm->rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss += t->rss;
+	}
+	if (rss < 0)
+		rss = 0;
+	rcu_read_unlock();
+	return rss;
+}
+
+unsigned long get_anon_rss(struct mm_struct *mm)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss;
+
+	if (!mm)
+		return 0;
+
+	rcu_read_lock();
+	rss = mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss += t->anon_rss;
+	}
+	if (rss < 0)
+		rss = 0;
+	rcu_read_unlock();
+	return rss;
+}
+
+unsigned long get_shared(struct mm_struct *mm)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss;
+
+	if (!mm)
+		return 0;
+
+	rcu_read_lock();
+	rss = mm->rss - mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss += t->rss - t->anon_rss;
+	}
+	if (rss < 0)
+		rss = 0;
+	rcu_read_unlock();
+	return rss;
+}
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	if (!mm)
+		return;
+
+	spin_lock(&mm->page_table_lock);
+	mm->rss += tsk->rss;
+	mm->anon_rss += tsk->anon_rss;
+	list_del_rcu(&tsk->mm_tasks);
+	spin_unlock(&mm->page_table_lock);
+}
+
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	spin_lock(&mm->page_table_lock);
+	list_add_rcu(&tsk->mm_tasks, &mm->task_list);
+	spin_unlock(&mm->page_table_lock);
+}
+
+
Index: linux-2.6.9/include/linux/init_task.h
===================================================================
--- linux-2.6.9.orig/include/linux/init_task.h	2004-11-30 20:33:30.000000000 -0800
+++ linux-2.6.9/include/linux/init_task.h	2004-11-30 20:33:50.000000000 -0800
@@ -42,6 +42,7 @@
 	.mmlist		= LIST_HEAD_INIT(name.mmlist),		\
 	.cpu_vm_mask	= CPU_MASK_ALL,				\
 	.default_kioctx = INIT_KIOCTX(name.default_kioctx, name),	\
+	.task_list	= LIST_HEAD_INIT(name.task_list),	\
 }

 #define INIT_SIGNALS(sig) {	\
@@ -112,6 +113,7 @@
 	.proc_lock	= SPIN_LOCK_UNLOCKED,				\
 	.switch_lock	= SPIN_LOCK_UNLOCKED,				\
 	.journal_info	= NULL,						\
+	.mm_tasks	= LIST_HEAD_INIT(tsk.mm_tasks),			\
 }


Index: linux-2.6.9/fs/exec.c
===================================================================
--- linux-2.6.9.orig/fs/exec.c	2004-11-30 20:33:41.000000000 -0800
+++ linux-2.6.9/fs/exec.c	2004-11-30 20:33:50.000000000 -0800
@@ -543,6 +543,7 @@
 	active_mm = tsk->active_mm;
 	tsk->mm = mm;
 	tsk->active_mm = mm;
+	mm_add_thread(mm, current);
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
 	arch_pick_mmap_layout(mm);

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Linus Torvalds]

On Wed, 1 Dec 2004, Christoph Lameter wrote:
>
> Changes from V11->V12 of this patch:
> - dump sloppy_rss in favor of list_rss (Linus' proposal)
> - keep up against current Linus tree (patch is based on 2.6.10-rc2-bk14)
> 
> This is a series of patches that increases the scalability of
> the page fault handler for SMP. Here are some performance results
> on a machine with 512 processors allocating 32 GB with an increasing
> number of threads (that are assigned a processor each).

Ok, consider me convinced. I don't want to apply this before I get 2.6.10 
out the door, but I'm happy with it. I assume Andrew has already picked up 
the previous version.

		Linus

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Linus Torvalds <torvalds@osdl.org> wrote:
>
> 
> 
> On Wed, 1 Dec 2004, Christoph Lameter wrote:
> >
> > Changes from V11->V12 of this patch:
> > - dump sloppy_rss in favor of list_rss (Linus' proposal)
> > - keep up against current Linus tree (patch is based on 2.6.10-rc2-bk14)
> > 
> > This is a series of patches that increases the scalability of
> > the page fault handler for SMP. Here are some performance results
> > on a machine with 512 processors allocating 32 GB with an increasing
> > number of threads (that are assigned a processor each).
> 
> Ok, consider me convinced. I don't want to apply this before I get 2.6.10 
> out the door, but I'm happy with it.

There were concerns about some architectures relying upon page_table_lock
for exclusivity within their own pte handling functions.  Have they all
been resolved?

> I assume Andrew has already picked up the previous version.

Nope.  It has major clashes with the 4-level-pagetable work.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Christoph Lameter]

On Wed, 1 Dec 2004, Andrew Morton wrote:

> > Ok, consider me convinced. I don't want to apply this before I get 2.6.10
> > out the door, but I'm happy with it.
>
> There were concerns about some architectures relying upon page_table_lock
> for exclusivity within their own pte handling functions.  Have they all
> been resolved?

The patch will fall back on the page_table_lock if an architecture cannot
provide atomic pte operations.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Linus Torvalds wrote:
> Ok, consider me convinced. I don't want to apply this before I get 2.6.10 
> out the door, but I'm happy with it. I assume Andrew has already picked up 
> the previous version.


Does that mean that 2.6.10 is actually close to the door?

/me runs...

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Jeff Garzik <jgarzik@pobox.com> wrote:
>
> Linus Torvalds wrote:
> > Ok, consider me convinced. I don't want to apply this before I get 2.6.10 
> > out the door, but I'm happy with it. I assume Andrew has already picked up 
> > the previous version.
> 
> 
> Does that mean that 2.6.10 is actually close to the door?
> 

We need an -rc3 yet.  And I need to do another pass through the
regressions-since-2.6.9 list.  We've made pretty good progress there
recently.  Mid to late December is looking like the 2.6.10 date.

We need to be be achieving higher-quality major releases than we did in
2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
stabilisation periods.

Of course, nobody will test -rc3 and a zillion people will test final
2.6.10, which is when we get lots of useful bug reports.  If this keeps on
happening then we'll need to get more serious about the 2.6.10.n process.

Or start alternating between stable and flakey releases, so 2.6.11 will be
a feature release with a 2-month development period and 2.6.12 will be a
bugfix-only release, with perhaps a 2-week development period, so people
know that the even-numbered releases are better stabilised.

We'll see.  It all depends on how many bugs you can fix in the next two
weeks ;)

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Andrew Morton wrote:
> We need to be be achieving higher-quality major releases than we did in
> 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
> stabilisation periods.


I'm still hoping that distros (like my employer) and orgs like OSDL will 
step up, and hook 2.6.x BK snapshots into daily test harnesses.

Something like John Cherry's reports to lkml on warnings and errors 
would be darned useful.  His reports are IMO an ideal model:  show 
day-to-day _changes_ in test results.  Don't just dump a huge list of 
testsuite results, results which are often clogged with expected 
failures and testsuite bug noise.

	Jeff

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Andrew Morton wrote:
> We need an -rc3 yet.  And I need to do another pass through the
> regressions-since-2.6.9 list.  We've made pretty good progress there
> recently.  Mid to late December is looking like the 2.6.10 date.


another for that list, BTW:

I am currently chasing a 2.6.8->2.6.9 SATA regression, which causes 
ata_piix (Intel ICH5/6/7) to not-find some SATA devices on x86-64 SMP, 
but works on UP.  Potentially related to >=4GB of RAM.



Details, in case anyone is interested:
Unless my code is screwed up (certainly possible), PIO data-in [using 
the insw() call] seems to return all zeroes on a true-blue SMP machine, 
for the identify-device command.  When this happens, libata (correctly) 
detects a bad id page and bails.  (problem doesn't show up on single CPU 
w/ HT)

What changed from 2.6.8 to 2.6.9 is

2.6.8:
	bitbang ATA taskfile registers (loads command)
	bitbang ATA data register (read id page)

2.6.9:
	bitbang ATA taskfile registers
	queue_work()
	workqueue thread bitbangs ATA data register (read id page)

So I wonder if <something> doesn't like CPU 0 sending I/O traffic to the 
on-board SATA PCI device, then immediately after that, CPU 1 sending I/O 
traffic.

Anyway, back to debugging...  :)

	Jeff

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Jeff Garzik <jgarzik@pobox.com> wrote:
>
> Andrew Morton wrote:
> > We need to be be achieving higher-quality major releases than we did in
> > 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
> > stabilisation periods.
> 
> 
> I'm still hoping that distros (like my employer) and orgs like OSDL will 
> step up, and hook 2.6.x BK snapshots into daily test harnesses.

I believe that both IBM and OSDL are doing this, or are getting geared up
to do this.  With both Linus bk and -mm.

However I have my doubts about how useful it will end up being.  These test
suites don't seem to pick up many regressions.  I've challenged Gerrit to
go back through a release cycle's bugfixes and work out how many of those
bugs would have been detected by the test suite.

My suspicion is that the answer will be "a very small proportion", and that
really is the bottom line.

We simply get far better coverage testing by releasing code, because of all
the wild, whacky and weird things which people do with their computers. 
Bless them.

> Something like John Cherry's reports to lkml on warnings and errors 
> would be darned useful.  His reports are IMO an ideal model:  show 
> day-to-day _changes_ in test results.  Don't just dump a huge list of 
> testsuite results, results which are often clogged with expected 
> failures and testsuite bug noise.
> 

Yes, we need humans between the tests and the developers.  Someone who has
good experience with the tests and who can say "hey, something changed
when I do X".  If nothing changed, we don't hear anything.

It's a developer role, not a testing role.   All testing is, really.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Benjamin Herrenschmidt]

On Thu, 2004-12-02 at 02:00 -0500, Jeff Garzik wrote:

> 
> 2.6.9:
> 	bitbang ATA taskfile registers
> 	queue_work()
> 	workqueue thread bitbangs ATA data register (read id page)
> 
> So I wonder if <something> doesn't like CPU 0 sending I/O traffic to the 
> on-board SATA PCI device, then immediately after that, CPU 1 sending I/O 
> traffic.
> 
> Anyway, back to debugging...  :)

They may not end up in order if they are stores (the stores to the
taskfile may be out of order vs; the loads/stores to/from the data
register) unless you have a spinlock protecting both or a full sync (on
ppc), but then, I don't know the ordering things on x86_64. This could
certainly be a problem on ppc & ppc64 too.

Ben.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Benjamin Herrenschmidt wrote:
> They may not end up in order if they are stores (the stores to the
> taskfile may be out of order vs; the loads/stores to/from the data
> register) unless you have a spinlock protecting both or a full sync (on
> ppc), but then, I don't know the ordering things on x86_64. This could
> certainly be a problem on ppc & ppc64 too.


Is synchronization beyond in[bwl] needed, do you think?

This specific problem is only on Intel ICHx AFAICS, which is PIO not 
MMIO and x86-only.  I presumed insw() by its very nature already has 
synchronization, but perhaps not...

	Jeff

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Martin J. Bligh]

--Andrew Morton <akpm@osdl.org> wrote (on Wednesday, December 01, 2004 23:02:17 -0800):

> Jeff Garzik <jgarzik@pobox.com> wrote:
>> 
>> Andrew Morton wrote:
>> > We need to be be achieving higher-quality major releases than we did in
>> > 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
>> > stabilisation periods.
>> 
>> 
>> I'm still hoping that distros (like my employer) and orgs like OSDL will 
>> step up, and hook 2.6.x BK snapshots into daily test harnesses.
> 
> I believe that both IBM and OSDL are doing this, or are getting geared up
> to do this.  With both Linus bk and -mm.

I already run a bunch of tests on a variety of machines for every new 
kernel ... but don't have an automated way to compare the results as yet, 
so don't actually look at them much ;-(. Sometime soon (quite possibly over 
Christmas) things will calm down enough I'll get a couple of days to write 
the appropriate perl script, and start publishing stuff.

> However I have my doubts about how useful it will end up being.  These test
> suites don't seem to pick up many regressions.  I've challenged Gerrit to
> go back through a release cycle's bugfixes and work out how many of those
> bugs would have been detected by the test suite.
> 
> My suspicion is that the answer will be "a very small proportion", and that
> really is the bottom line.

Yeah, probably. Though the stress tests catch a lot more than the 
functionality ones. The big pain in the ass is drivers, because I don't
have a hope in hell of testing more than 1% of them.

M.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Martin J. Bligh wrote:
> Yeah, probably. Though the stress tests catch a lot more than the 
> functionality ones. The big pain in the ass is drivers, because I don't
> have a hope in hell of testing more than 1% of them.

My dream is that hardware vendors rotate their current machines through 
a test shop :)  It would be nice to make sure that the popular drivers 
get daily test coverage.

	Jeff, dreaming on

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Benjamin Herrenschmidt]

On Thu, 2004-12-02 at 02:11 -0500, Jeff Garzik wrote:
> Benjamin Herrenschmidt wrote:
> > They may not end up in order if they are stores (the stores to the
> > taskfile may be out of order vs; the loads/stores to/from the data
> > register) unless you have a spinlock protecting both or a full sync (on
> > ppc), but then, I don't know the ordering things on x86_64. This could
> > certainly be a problem on ppc & ppc64 too.
> 
> 
> Is synchronization beyond in[bwl] needed, do you think?

Yes, when potentially hop'ing between CPUs, definitely.

> This specific problem is only on Intel ICHx AFAICS, which is PIO not 
> MMIO and x86-only.  I presumed insw() by its very nature already has 
> synchronization, but perhaps not...

Hrm... on "pure" x86, I would expect so at the HW level, not sure about
x86_64... but there would be definitely an issue on ppc with your
scheme. You need at least a full barrier before you trigger the
workqueue. That may not be the problem you are facing now, but it would
become one.

Ben.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andy Warner]

[-- Attachment #1: Type: text/plain, Size: 2854 bytes --]

Jeff Garzik wrote:
> [...]
> I am currently chasing a 2.6.8->2.6.9 SATA regression, which causes 
> ata_piix (Intel ICH5/6/7) to not-find some SATA devices on x86-64 SMP, 
> but works on UP.  Potentially related to >=4GB of RAM.
> 
> 
> 
> Details, in case anyone is interested:
> Unless my code is screwed up (certainly possible), PIO data-in [using 
> the insw() call] seems to return all zeroes on a true-blue SMP machine, 
> for the identify-device command.  When this happens, libata (correctly) 
> detects a bad id page and bails.  (problem doesn't show up on single CPU 
> w/ HT)

Ah, I might have been here recently, with the pass-thru stuff.

What I saw was that in an SMP machine:

1. queue_work() can result in the work running (on another
   CPU) instantly.

2. Having one CPU beat on PIO registers reading data from one port
   would significantly alter the timing of the CMD->BSY->DRQ sequence
   used in PIO. This behaviour was far worse for competing ports
   within one chip, which I put down to arbitration problems.

3. CPU utilisation would go through the roof. Effectively the
   entire pio_task state machine reduced to a busy spin loop.

4. The state machine needed some tweaks, especially in error
   handling cases.

I made some changes, which effectively solved the problem for promise
TX4-150 cards, and was going to test the results on other chipsets
next week before speaking up. Specifically, I have seen some
issues with SiI 3114 cards.

I was trying to explore using interrupts instead of polling state
but for some reason, I was not getting them for PIO data operations,
or I misunderstand the spec, after removing ata_qc_set_polling() - again
I saw a difference in behaviour between the Promise & SiI cards
here.

I'm about to go offline for 3 days, and hadn't prepared for this
yet. The best I can do is provide a patch (attached) that applies
against 2.6.9. It also seems to apply against libata-2.6, but
barfs a bit against libata-dev-2.6.

The changes boil down to these:

1. Minor changes in how status/error regs are read.
   Including attempts to use altstatus, while I was
   exploring interrupts.

2. State machine logic changes.

3. Replace calls to queue_work() with queue_delayed_work()
   to stop SMP machines going crazy.

With these changes, on a platform consisting of 2.6.9 and
Promise TX4-150 cards, I can move terabytes of parallel
PIO data, without error.

My gut says that the PIO mechanism should be overhauled, I
composed a "how much should we pay for this muffler" email
to linux-ide at least twice while working on this, but never
sent it - wanting to send a solution in rather than just
making more comments from the peanut gallery.

I'll pick up the thread on this next week, when I'm back online.
I hope this helps.
-- 
andyw@pobox.com

Andy Warner		Voice: (612) 801-8549	Fax: (208) 575-5634

[-- Attachment #2: 2.6.9-pio-smp.patch --]
[-- Type: text/plain, Size: 1862 bytes --]

diff -r -u -X dontdiff linux-2.6.9-vanilla/drivers/scsi/libata-core.c linux-2.6.9/drivers/scsi/libata-core.c
--- linux-2.6.9-vanilla/drivers/scsi/libata-core.c	2004-10-18 16:53:06.000000000 -0500
+++ linux-2.6.9/drivers/scsi/libata-core.c	2004-11-24 11:01:40.000000000 -0600
@@ -2099,7 +2099,7 @@
 	}
 
 	drv_stat = ata_wait_idle(ap);
-	if (!ata_ok(drv_stat)) {
+	if (drv_stat & (ATA_ERR | ATA_DF)) {
 		ap->pio_task_state = PIO_ST_ERR;
 		return;
 	}
@@ -2254,23 +2254,17 @@
 	 * chk-status again.  If still busy, fall back to
 	 * PIO_ST_POLL state.
 	 */
-	status = ata_busy_wait(ap, ATA_BUSY, 5);
-	if (status & ATA_BUSY) {
+	status = ata_altstatus(ap) ;
+	if (!(status & ATA_DRQ)) {
 		msleep(2);
-		status = ata_busy_wait(ap, ATA_BUSY, 10);
-		if (status & ATA_BUSY) {
+		status = ata_altstatus(ap) ;
+		if (!(status & ATA_DRQ)) {
 			ap->pio_task_state = PIO_ST_POLL;
 			ap->pio_task_timeout = jiffies + ATA_TMOUT_PIO;
 			return;
 		}
 	}
 
-	/* handle BSY=0, DRQ=0 as error */
-	if ((status & ATA_DRQ) == 0) {
-		ap->pio_task_state = PIO_ST_ERR;
-		return;
-	}
-
 	qc = ata_qc_from_tag(ap, ap->active_tag);
 	assert(qc != NULL);
 
@@ -2321,17 +2315,15 @@
 	case PIO_ST_TMOUT:
 	case PIO_ST_ERR:
 		ata_pio_error(ap);
-		break;
+		return ;
 	}
 
-	if ((ap->pio_task_state != PIO_ST_IDLE) &&
-	    (ap->pio_task_state != PIO_ST_TMOUT) &&
-	    (ap->pio_task_state != PIO_ST_ERR)) {
+	if (ap->pio_task_state != PIO_ST_IDLE) {
 		if (timeout)
 			queue_delayed_work(ata_wq, &ap->pio_task,
 					   timeout);
 		else
-			queue_work(ata_wq, &ap->pio_task);
+			queue_delayed_work(ata_wq, &ap->pio_task, 2);
 	}
 }
 
@@ -2624,7 +2616,7 @@
 		ata_qc_set_polling(qc);
 		ata_tf_to_host_nolock(ap, &qc->tf);
 		ap->pio_task_state = PIO_ST;
-		queue_work(ata_wq, &ap->pio_task);
+		queue_delayed_work(ata_wq, &ap->pio_task, 2);
 		break;
 
 	case ATA_PROT_ATAPI:

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Gerrit Huizenga]

On Wed, 01 Dec 2004 23:02:17 PST, Andrew Morton wrote:
> Jeff Garzik <jgarzik@pobox.com> wrote:
> >
> > Andrew Morton wrote:
> > > We need to be be achieving higher-quality major releases than we did in
> > > 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
> > > stabilisation periods.
> > 
> > 
> > I'm still hoping that distros (like my employer) and orgs like OSDL will 
> > step up, and hook 2.6.x BK snapshots into daily test harnesses.
> 
> I believe that both IBM and OSDL are doing this, or are getting geared up
> to do this.  With both Linus bk and -mm.
> 
> However I have my doubts about how useful it will end up being.  These test
> suites don't seem to pick up many regressions.  I've challenged Gerrit to
> go back through a release cycle's bugfixes and work out how many of those
> bugs would have been detected by the test suite.
> 
> My suspicion is that the answer will be "a very small proportion", and that
> really is the bottom line.
 
Yeah, sort of what Martin said.  LTP, for instance, doesn't find a lot
of what is in our internal bugzilla or the bugme database.  Automated
testing tends not to cover all the range of desktop peripherals and
drivers that make up a large quantity of the code but gets very little
coverage.  Our stress testing is extensive and was finding 3 year old
problems when we first ran it but it is pretty expensive to run those
types of tests (machines, people, data analysis) so we typically run
those tests on distros rather than mainline to help validate distro
quality.

However, that said, the LTP stuff is still *necessary* - it would
catch quite a number of regressions if we were to regress.  The good
thing is that most changes today haven't been leading to regressions.
That could change at any time, and one of the keys is to make sure that
when we do find regressions we get a test into LTP to make sure that
that particular regression never happens again.

I haven't looked at the code coverage for LTP in a while but it is
actually a high line count coverage test for core kernel.  I don't remember
if it was over 80% or not, but usually 85-88% is the point of diminishing
returns for a regression suite.  I think a more important proactive
step here is to understand what regressions we *do* have an whether
or not we can construct a test that in the future will catch that
regression (or better, a class of regressions).

And, maybe we need some kind of filter person or group for lkml that
can see what the key regressions are (e.g. akpm, if you know of a set
of regressions that you are working, maybe periodically sending those
to the ltp mailing list) we could focus on creating tests for those
regressions.

We are also working to set up large ISV applications in a couple of
spots - both inside IBM and there is a similar effort underway at OSDL.
Those ISV applications will catch a class of real world usage models
and also check for regressions.  I don't know if it is possible to set
up a better testing environment for the wild, whacky and weird things
that people do but, yes, Bless them.  ;-)

> We simply get far better coverage testing by releasing code, because of all
> the wild, whacky and weird things which people do with their computers. 
> Bless them.
> 
> > Something like John Cherry's reports to lkml on warnings and errors 
> > would be darned useful.  His reports are IMO an ideal model:  show 
> > day-to-day _changes_ in test results.  Don't just dump a huge list of 
> > testsuite results, results which are often clogged with expected 
> > failures and testsuite bug noise.
> 
> Yes, we need humans between the tests and the developers.  Someone who has
> good experience with the tests and who can say "hey, something changed
> when I do X".  If nothing changed, we don't hear anything.
> 
> It's a developer role, not a testing role.   All testing is, really.

Yep.  However, smart developers continue to write scripts to automate
the rote and mundane tasks that they hate doing.  Towards that end, there
was a recent effort at Bull on the NPTL work which serves as a very
interesting model:

http://nptl.bullopensource.org/Tests/results/run-browse.php

Basically, you can compare results from any test run with any other
and get a summary of differences.  That helps give a quick status
check and helps you focus on the correct issues when tracking down
defects.

gerrit

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[cliff white]

On Wed, 1 Dec 2004 23:02:17 -0800
Andrew Morton <akpm@osdl.org> wrote:

> Jeff Garzik <jgarzik@pobox.com> wrote:
> >
> > Andrew Morton wrote:
> > > We need to be be achieving higher-quality major releases than we did in
> > > 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
> > > stabilisation periods.
> > 
> > 
> > I'm still hoping that distros (like my employer) and orgs like OSDL will 
> > step up, and hook 2.6.x BK snapshots into daily test harnesses.
> 
> I believe that both IBM and OSDL are doing this, or are getting geared up
> to do this.  With both Linus bk and -mm.

Gee, OSDL has been doing this sort of testing for > 1 years now. Getting
bandwidth to look at the results has been a problem. We need more eyeballs
and community support badly, i'm very glad Marcelo has shown recent interest. 
> 
> However I have my doubts about how useful it will end up being.  These test
> suites don't seem to pick up many regressions.  I've challenged Gerrit to
> go back through a release cycle's bugfixes and work out how many of those
> bugs would have been detected by the test suite.

> 
> My suspicion is that the answer will be "a very small proportion", and that
> really is the bottom line.
> 
> We simply get far better coverage testing by releasing code, because of all
> the wild, whacky and weird things which people do with their computers. 
> Bless them.
> 
> > Something like John Cherry's reports to lkml on warnings and errors 
> > would be darned useful.  His reports are IMO an ideal model:  show 
> > day-to-day _changes_ in test results.  Don't just dump a huge list of 
> > testsuite results, results which are often clogged with expected 
> > failures and testsuite bug noise.
> > 
> 
> Yes, we need humans between the tests and the developers.  Someone who has
> good experience with the tests and who can say "hey, something changed
> when I do X".  If nothing changed, we don't hear anything.

I would agree, and would do almost anything to help/assist/enable any humans 
interested. We need some expertise on when to run certain tests, to avoid
data overload. 
I've noticed that when developer's submit test results with a patch, it sometimes
helps in the decision on patch acceptance. Is there a way to promote this sort of
behaviour?
cliffw
OSDL
> 
> It's a developer role, not a testing role.   All testing is, really.
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>
> 


-- 
The church is near, but the road is icy.
The bar is far, but i will walk carefully. - Russian proverb

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[cliff white]

On Thu, 02 Dec 2004 02:31:35 -0500
Jeff Garzik <jgarzik@pobox.com> wrote:

> Martin J. Bligh wrote:
> > Yeah, probably. Though the stress tests catch a lot more than the 
> > functionality ones. The big pain in the ass is drivers, because I don't
> > have a hope in hell of testing more than 1% of them.
> 
> My dream is that hardware vendors rotate their current machines through 
> a test shop :)  It would be nice to make sure that the popular drivers 
> get daily test coverage.
> 
> 	Jeff, dreaming on

OSDL has recently re-done the donation policy, and we're much better positioned
to support that sort of thing now - Contact Tom Hanrahan at OSDL if you 
are a vendor, or know a vendor. ( Or you can become a vendor ) 

cliffw

> 
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>
> 


-- 
The church is near, but the road is icy.
The bar is far, but i will walk carefully. - Russian proverb

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Gerrit Huizenga]

On Thu, 02 Dec 2004 10:10:29 PST, cliff white wrote:
> On Thu, 02 Dec 2004 02:31:35 -0500
> Jeff Garzik <jgarzik@pobox.com> wrote:
> 
> > Martin J. Bligh wrote:
> > > Yeah, probably. Though the stress tests catch a lot more than the 
> > > functionality ones. The big pain in the ass is drivers, because I don't
> > > have a hope in hell of testing more than 1% of them.
> > 
> > My dream is that hardware vendors rotate their current machines through 
> > a test shop :)  It would be nice to make sure that the popular drivers 
> > get daily test coverage.
> > 
> > 	Jeff, dreaming on
> 
> OSDL has recently re-done the donation policy, and we're much better positioned
> to support that sort of thing now - Contact Tom Hanrahan at OSDL if you 
> are a vendor, or know a vendor. ( Or you can become a vendor ) 

Specifically Tom Hanrahan == hanrahat@osdl.org

gerrit

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Grant Grundler]

On Wed, Dec 01, 2004 at 10:34:41PM -0800, Andrew Morton wrote:
> Of course, nobody will test -rc3 and a zillion people will test final
> 2.6.10, which is when we get lots of useful bug reports.  If this keeps on
> happening then we'll need to get more serious about the 2.6.10.n process.
> 
> Or start alternating between stable and flakey releases, so 2.6.11 will be
> a feature release with a 2-month development period and 2.6.12 will be a
> bugfix-only release, with perhaps a 2-week development period, so people
> know that the even-numbered releases are better stabilised.

No matter what scheme you adopt, I (and others) will adapt as well.
When working on a new feature or bug fix, I don't chase -bk releases
since I don't want to find new, unrelated issues that interfere with
the issue I was originally chasing. I roll to a new release when
the issue I care about is "cooked". Anything that takes longer than
a month or so is just hopeless since I fall too far behind.

(e.g. IRQ handling in parisc-linux needs to be completely rewritten
to pickup irq_affinity support - I just don't have enough time to get
it done in < 2 monthes. We started on this last year and gave up.)

I see "2.6.10.n process" as the right way to handle bug fix only releases.
I'm happy to work on 2.6.10.0 and understand the initial release was a
"best effort".

2.6.odd/.even release described above is a variant of 2.6.10.n releases
where n = {0, 1}. The question is how many parallel releases do people
(you and linus) want us keep "alive" at the same time?
odd/even implies only one vs several if 2.6.X.n scheme is continued
beyond 2.6.8.1.

Also need to think about how well any scheme align's with what distro's
need to support releases. Like the "Adopt-a-Highway" program in
California to pickup trash along highways, I'm wondering if distros
would be willing/interested in adopting a particular release
and maintain it in bk.  e.g. SuSE clearly has interest in some sort
of 2.6.5.n series for SLES9. ditto for RHEL4 (but for 2.6.9.n).
The question of *who* (at respective distro) would be the release
maintainer is a titanic sized rathole. But there is a release manager
today at each distro and perhaps it's easier if s/he remains invisible
to us.

hth,
grant

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Grant Grundler <iod00d@hp.com> wrote:
>
> 2.6.odd/.even release described above is a variant of 2.6.10.n releases
>  where n = {0, 1}. The question is how many parallel releases do people
>  (you and linus) want us keep "alive" at the same time?

2.6.odd/.even is actually a significantly different process.  a) because
there's only one tree, linearly growing.  That's considerably simpler than
maintaining a branch.  And b) because everyone knows that there won't be a
new development tree opened until we've all knuckled down and fixed the
bugs which we put into the previous one, dammit.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Christoph Hellwig]

On Thu, Dec 02, 2004 at 10:27:16AM -0800, Grant Grundler wrote:
> Also need to think about how well any scheme align's with what distro's
> need to support releases. Like the "Adopt-a-Highway" program in
> California to pickup trash along highways, I'm wondering if distros
> would be willing/interested in adopting a particular release
> and maintain it in bk.  e.g. SuSE clearly has interest in some sort
> of 2.6.5.n series for SLES9. ditto for RHEL4 (but for 2.6.9.n).

Unfortunately the SLES9 kernels don't really look anything like 2.6.5
except from the version number.  There's far too much trash from Business
Partners in there.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[cliff white]

On Wed, 01 Dec 2004 23:26:59 -0800
"Martin J. Bligh" <mbligh@aracnet.com> wrote:

> --Andrew Morton <akpm@osdl.org> wrote (on Wednesday, December 01, 2004 23:02:17 -0800):
> 
> > Jeff Garzik <jgarzik@pobox.com> wrote:
> >> 
> >> Andrew Morton wrote:
> >> > We need to be be achieving higher-quality major releases than we did in
> >> > 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
> >> > stabilisation periods.
> >> 
> >> 
> >> I'm still hoping that distros (like my employer) and orgs like OSDL will 
> >> step up, and hook 2.6.x BK snapshots into daily test harnesses.
> > 
> > I believe that both IBM and OSDL are doing this, or are getting geared up
> > to do this.  With both Linus bk and -mm.
> 
> I already run a bunch of tests on a variety of machines for every new 
> kernel ... but don't have an automated way to compare the results as yet, 
> so don't actually look at them much ;-(. Sometime soon (quite possibly over 
> Christmas) things will calm down enough I'll get a couple of days to write 
> the appropriate perl script, and start publishing stuff.

We've had the most success when one person has an itch to scratch, and works
with us to scratch it. We (OSDL) worked with Sebastien at Bull, and we're very 
glad he had the time to do such excellent work. We worked with Con Kolivas, likewise.

We've done tools to automate LTP comparisons ( bryce@osdl.org  has posted results )
and reaim, we've been able to post some regression to lkml, and tied in with developers
to get bugs fixed. But OSDL has been limited by manpower.
 
One of the issues with the performance tests is the amount of data produced - 
 for example, the deep IO tests produce ton's o'  numbers, but the developer community wants
a single "+/- 5%" type response-  we need some opinions and help on how to do the data reduction 
necessary. 

What would be really kewl is some test/analysis code that could be re-used, so the Martin's of the future
have a good starting place. 
cliffw
OSDL




> 
> > However I have my doubts about how useful it will end up being.  These test
> > suites don't seem to pick up many regressions.  I've challenged Gerrit to
> > go back through a release cycle's bugfixes and work out how many of those
> > bugs would have been detected by the test suite.
> > 
> > My suspicion is that the answer will be "a very small proportion", and that
> > really is the bottom line.
> 
> Yeah, probably. Though the stress tests catch a lot more than the 
> functionality ones. The big pain in the ass is drivers, because I don't
> have a hope in hell of testing more than 1% of them.
> 
> M.
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>
> 


-- 
The church is near, but the road is icy.
The bar is far, but i will walk carefully. - Russian proverb

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Diego Calleja]

El Thu, 02 Dec 2004 01:48:25 -0500 Jeff Garzik <jgarzik@pobox.com>
escribió:


> I'm still hoping that distros (like my employer) and orgs like OSDL will 
> step up, and hook 2.6.x BK snapshots into daily test harnesses.

Automated .deb's and .rpm's for the -bk snapshots (and yum/apt repositories)
would be nice for all those people who run unsupported distros.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Diego Calleja wrote:
> El Thu, 02 Dec 2004 01:48:25 -0500 Jeff Garzik <jgarzik@pobox.com>
> escribió:
> 
> 
> 
>>I'm still hoping that distros (like my employer) and orgs like OSDL will 
>>step up, and hook 2.6.x BK snapshots into daily test harnesses.
> 
> 
> Automated .deb's and .rpm's for the -bk snapshots (and yum/apt repositories)
> would be nice for all those people who run unsupported distros.

Now, that's a darned good idea...

Should be simple for rpm at least, given the "make rpm" target.  I 
wonder if we have, or could add, a 'make deb' target.

	Jeff

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[linux-os]

On Thu, 2 Dec 2004, cliff white wrote:

> On Thu, 02 Dec 2004 02:31:35 -0500
> Jeff Garzik <jgarzik@pobox.com> wrote:
>
>> Martin J. Bligh wrote:
>>> Yeah, probably. Though the stress tests catch a lot more than the
>>> functionality ones. The big pain in the ass is drivers, because I don't
>>> have a hope in hell of testing more than 1% of them.
>>
>> My dream is that hardware vendors rotate their current machines through
>> a test shop :)  It would be nice to make sure that the popular drivers
>> get daily test coverage.
>>
>> 	Jeff, dreaming on
>

It isn't going to happen until the time when the vendors
call somebody a liar, try to get them fired, and then
that somebody takes them to court and they lose 100
million dollars or so.

Until that happens, vendors will continue to make junk
and they will continue to lie about the performance of
that junk. It doesn't help that Software Engineering has
become a "hardware junk fixing" job.

Basically many vendors in the PC and PC peripheral
business are, for lack of a better word, liars who
are in the business of perpetrating fraud upon the
unsuspecting PC user.

We have vendors who convincingly change mega-bits
to mega-bytes, improving performance 8-fold without
any expense at all. We have vendors reducing the
size of a kilobyte and a megabyte, then getting
the new lies entered into dictionaries, etc. The
scheme goes on.

In the meantime, if you try to perform DMA
across a PCI/Bus at or near the specified rates,
you will learn that the specifications are
for "this chip" or "that chip", and have nothing
to do with the performance when these chips
get connected together. You will find that real
performance is about 20 percent of the specification.

Occasionally you find a vendor that doesn't lie and
the same chip-set magically performs close to
the published specifications. This is becoming
rare because it costs money to build motherboards
that work. This might require two or more
prototypes to get the timing just right so the
artificial delays and re-clocking, used to make
junk work, isn't required.

Once the PC (and not just the desk-top PC) became
a commodity, everything points to the bottom-line.
You get into the business by making something that
looks and smells new. Then you sell it by writing
specifications that are better than the most
expensive on the market. Your sales-price is
set below average market so you can unload this
junk as rapidly as possible.

Then, you do this over again, claiming that your
equipment is "state-of-the-art"! And if anybody
ever tests the junk and claims that it doesn't
work as specified, you contact the president of
his company and try to kill the messenger.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.9 on an i686 machine (5537.79 BogoMips).
  Notice : All mail here is now cached for review by John Ashcroft.
                  98.36% of all statistics are fiction.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Diego Calleja]

El Thu, 02 Dec 2004 15:12:22 -0500 Jeff Garzik <jgarzik@pobox.com>
escribió:

> > Automated .deb's and .rpm's for the -bk snapshots (and yum/apt
> > repositories) would be nice for all those people who run unsupported
> > distros.
> 
> Now, that's a darned good idea...
> 
> Should be simple for rpm at least, given the "make rpm" target.  I 
> wonder if we have, or could add, a 'make deb' target.


There was a patch for that long time ago before 2.6 was out IIRC? I don't
know where it went (CC'ing Sam who should know ;)

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Wichert Akkerman]

Previously Jeff Garzik wrote:
> Should be simple for rpm at least, given the "make rpm" target.  I 
> wonder if we have, or could add, a 'make deb' target.

make deb-pkg has been there for a while.

Wichert.

-- 
Wichert Akkerman <wichert@wiggy.net>    It is simple to make things.
http://www.wiggy.net/                   It is hard to make things simple.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Francois Romieu]

Jeff Garzik <jgarzik@pobox.com> :
[...]
> Should be simple for rpm at least, given the "make rpm" target.  I 
> wonder if we have, or could add, a 'make deb' target.

http://www.wiggy.net/files/kerneldeb-1.2.ptc ?

--
Ueimor

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Marcelo Tosatti]

On Thu, Dec 02, 2004 at 10:43:30AM -0800, cliff white wrote:
> On Wed, 01 Dec 2004 23:26:59 -0800
> "Martin J. Bligh" <mbligh@aracnet.com> wrote:
> 
> > --Andrew Morton <akpm@osdl.org> wrote (on Wednesday, December 01, 2004 23:02:17 -0800):
> > 
> > > Jeff Garzik <jgarzik@pobox.com> wrote:
> > >> 
> > >> Andrew Morton wrote:
> > >> > We need to be be achieving higher-quality major releases than we did in
> > >> > 2.6.8 and 2.6.9.  Really the only tool we have to ensure this is longer
> > >> > stabilisation periods.
> > >> 
> > >> 
> > >> I'm still hoping that distros (like my employer) and orgs like OSDL will 
> > >> step up, and hook 2.6.x BK snapshots into daily test harnesses.
> > > 
> > > I believe that both IBM and OSDL are doing this, or are getting geared up
> > > to do this.  With both Linus bk and -mm.
> > 
> > I already run a bunch of tests on a variety of machines for every new 
> > kernel ... but don't have an automated way to compare the results as yet, 
> > so don't actually look at them much ;-(. Sometime soon (quite possibly over 
> > Christmas) things will calm down enough I'll get a couple of days to write 
> > the appropriate perl script, and start publishing stuff.
> 
> We've had the most success when one person has an itch to scratch, and works
> with us to scratch it. We (OSDL) worked with Sebastien at Bull, and we're very 
> glad he had the time to do such excellent work. We worked with Con Kolivas, likewise.
> 
> We've done tools to automate LTP comparisons ( bryce@osdl.org  has posted results )
> and reaim, we've been able to post some regression to lkml, and tied in with developers
> to get bugs fixed. But OSDL has been limited by manpower.
>  
> One of the issues with the performance tests is the amount of data produced - 
>  for example, the deep IO tests produce ton's o'  numbers, but the developer community wants
> a single "+/- 5%" type response-  we need some opinions and help on how to do the data reduction 
> necessary. 

Yep, reaim produces a single "global throughput" result in MB/s, which is wonderful 
for readability.

Now iozone on the other extreme produces output for each kind of operation
(read, write, rw, sync version of those) for each client IIRC. tiobench also
has detailed output for each operation.

We ought to reduce all benchmark results to "read", "write" and "global" (read+write/2) 
numbers. 

I'm willing to work on the data reduction and graphic generation scripts
for STP results. I think I can do that.

> 
> What would be really kewl is some test/analysis code that could be re-used, so the Martin's of the future
> have a good starting place. 

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Pavel Machek]

Hi!

> Or start alternating between stable and flakey releases, so 2.6.11 will be
> a feature release with a 2-month development period and 2.6.12 will be a
> bugfix-only release, with perhaps a 2-week development period, so people
> know that the even-numbered releases are better stabilised.

If you expect "feature 2.6.11", you might as well call it 2.7.0, 
followed by 2.8.0.

								Pavel
-- 
People were complaining that M$ turns users into beta-testers...
...jr ghea gurz vagb qrirybcref, naq gurl frrz gb yvxr vg gung jnl!

Anticipatory prefaulting in the page fault handler V1

[Christoph Lameter]

The page fault handler for anonymous pages can generate significant overhead
apart from its essential function which is to clear and setup a new page
table entry for a never accessed memory location. This overhead increases
significantly in an SMP environment.

In the page table scalability patches, we addressed the issue by changing
the locking scheme so that multiple fault handlers are able to be processed
concurrently on multiple cpus. This patch attempts to aggregate multiple
page faults into a single one. It does that by noting
anonymous page faults generated in sequence by an application.

If a fault occurred for page x and is then followed by page x+1 then it may
be reasonable to expect another page fault at x+2 in the future. If page
table entries for x+1 and x+2 would be prepared in the fault handling for
page x+1 then the overhead of taking a fault for x+2 is avoided. However
page x+2 may never be used and thus we may have increased the rss
of an application unnecessarily. The swapper will take care of removing
that page if memory should get tight.

The following patch makes the anonymous fault handler anticipate future
faults. For each fault a prediction is made where the fault would occur
(assuming linear acccess by the application). If the prediction turns out to
be right (next fault is where expected) then a number of pages is
preallocated in order to avoid a series of future faults. The order of the
preallocation increases by the power of two for each success in sequence.

The first successful prediction leads to an additional page being allocated.
Second successful prediction leads to 2 additional pages being allocated.
Third to 4 pages and so on. The max order is 3 by default. In a large
continous allocation the number of faults is reduced by a factor of 8.

The patch may be combined with the page fault scalability patch (another
edition of the patch is needed which will be forthcoming after the
page fault scalability patch has been included). The combined patches
will triple the possible page fault rate from ~1 mio faults sec to 3 mio
faults sec.

Standard Kernel on a 512 Cpu machine allocating 32GB with an increasing
number of threads (and thus increasing parallellism of page faults):

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32   3    1    1.416s    138.165s 139.050s 45073.831  45097.498
 32   3    2    1.397s    148.523s  78.044s 41965.149  80201.646
 32   3    4    1.390s    152.618s  44.044s 40851.258 141545.239
 32   3    8    1.500s    374.008s  53.001s 16754.519 118671.950
 32   3   16    1.415s   1051.759s  73.094s  5973.803  85087.358
 32   3   32    1.867s   3400.417s 117.003s  1849.186  53754.928
 32   3   64    5.361s  11633.040s 197.034s   540.577  31881.112
 32   3  128   23.387s  39386.390s 332.055s   159.642  18918.599
 32   3  256   15.409s  20031.450s 168.095s   313.837  37237.918
 32   3  512   18.720s  10338.511s  86.047s   607.446  72752.686

Patched kernel:

Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32   3    1    1.098s    138.544s 139.063s 45053.657  45057.920
 32   3    2    1.022s    127.770s  67.086s 48849.350  92707.085
 32   3    4    0.995s    119.666s  37.045s 52141.503 167955.292
 32   3    8    0.928s     87.400s  18.034s 71227.407 342934.242
 32   3   16    1.067s     72.943s  11.035s 85007.293 553989.377
 32   3   32    1.248s    133.753s  10.038s 46602.680 606062.151
 32   3   64    5.557s    438.634s  13.093s 14163.802 451418.617
 32   3  128   17.860s   1496.797s  19.048s  4153.714 322808.509
 32   3  256   13.382s    766.063s  10.016s  8071.695 618816.838
 32   3  512   17.067s    369.106s   5.041s 16291.764 1161285.521

These number are roughly equal to what can be accomplished with the
page fault scalability patches.

Kernel patches with both the page fault scalability patches and
prefaulting:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32  10    1    4.103s    456.384s 460.046s 45541.992  45544.369
 32  10    2    4.005s    415.119s 221.095s 50036.407  94484.174
 32  10    4    3.855s    371.317s 111.076s 55898.259 187635.724
 32  10    8    3.902s    308.673s  67.094s 67092.476 308634.397
 32  10   16    4.011s    224.213s  37.016s 91889.781 564241.062
 32  10   32    5.483s    209.391s  27.046s 97598.647 763495.417
 32  10   64   19.166s    219.925s  26.030s 87713.212 797286.395
 32  10  128   53.482s    342.342s  27.024s 52981.744 769687.791
 32  10  256   67.334s    180.321s  15.036s 84679.911 1364614.334
 32  10  512   66.516s     93.098s   9.015s131387.893 2291548.865

The fault rate doubles when both patches are applied.

And on the high end (512 processors allocating 256G) (No numbers
for regular kernels because they are extremely slow, also no
number for a low number of threads. Also very slow)

With prefaulting:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
256   3    4    8.241s   1414.348s 449.016s 35380.301 112056.239
256   3    8    8.306s   1300.982s 247.025s 38441.977 203559.271
256   3   16    8.368s   1223.853s 154.089s 40846.272 324940.924
256   3   32    8.536s   1284.041s 110.097s 38938.970 453556.624
256   3   64   13.025s   3587.203s 110.010s 13980.123 457131.492
256   3  128   25.025s  11460.700s 145.071s  4382.104 345404.909
256   3  256   26.150s   6061.649s  75.086s  8267.625 663414.482
256   3  512   20.637s   3037.097s  38.062s 16460.435 1302993.019

Page fault scalability patch and prefaulting. Max prefault order
increased to 5 (max preallocation of 32 pages):

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
256  10    8   33.571s   4516.293s 863.021s 36874.099 194356.930
256  10   16   33.103s   3737.688s 461.028s 44492.553 363704.484
256  10   32   35.094s   3436.561s 321.080s 48326.262 521352.840
256  10   64   46.675s   2899.997s 245.020s 56936.124 684214.256
256  10  128   85.493s   2890.198s 203.008s 56380.890 826122.524
256  10  256   74.299s   1374.973s  99.088s115762.963 1679630.272
256  10  512   62.760s    706.559s  53.027s218078.311 3149273.714

We are getting into an almost linear scalability in the high end with
both patches and end up with a fault rate > 3 mio faults per second.

The one thing that takes up a lot of time is still be the zeroing
of pages in the page fault handler. There is a another
set of patches that I am working on which will prezero pages
and led to another an increase in performance by a factor of 2-4
(if prezeroed pages are available which may not always be the case).
Maybe we can reach 10 mio fault /sec that way.

Patch against 2.6.10-rc3-bk3:

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-12-01 10:37:31.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-12-01 10:38:15.000000000 -0800
@@ -537,6 +537,8 @@
 #endif

 	struct list_head tasks;
+	unsigned long anon_fault_next_addr;	/* Predicted sequential fault address */
+	int anon_fault_order;			/* Last order of allocation on fault */
 	/*
 	 * ptrace_list/ptrace_children forms the list of my children
 	 * that were stolen by a ptracer.
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-01 10:38:11.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-01 10:45:01.000000000 -0800
@@ -55,6 +55,7 @@

 #include <linux/swapops.h>
 #include <linux/elf.h>
+#include <linux/pagevec.h>

 #ifndef CONFIG_DISCONTIGMEM
 /* use the per-pgdat data instead for discontigmem - mbligh */
@@ -1432,8 +1433,106 @@
 		unsigned long addr)
 {
 	pte_t entry;
-	struct page * page = ZERO_PAGE(addr);
+	struct page * page;
+
+	addr &= PAGE_MASK;
+
+ 	if (current->anon_fault_next_addr == addr) {
+ 		unsigned long end_addr;
+ 		int order = current->anon_fault_order;
+
+		/* Sequence of page faults detected. Perform preallocation of pages */

+		/* The order of preallocations increases with each successful prediction */
+ 		order++;
+
+		if ((1 << order) < PAGEVEC_SIZE)
+			end_addr = addr + (1 << (order + PAGE_SHIFT));
+		else
+			end_addr = addr + PAGEVEC_SIZE * PAGE_SIZE;
+
+		if (end_addr > vma->vm_end)
+			end_addr = vma->vm_end;
+		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
+			end_addr &= PMD_MASK;
+
+		current->anon_fault_next_addr = end_addr;
+	 	current->anon_fault_order = order;
+
+		if (write_access) {
+
+			struct pagevec pv;
+			unsigned long a;
+			struct page **p;
+
+			pte_unmap(page_table);
+			spin_unlock(&mm->page_table_lock);
+
+			pagevec_init(&pv, 0);
+
+			if (unlikely(anon_vma_prepare(vma)))
+				return VM_FAULT_OOM;
+
+			/* Allocate the necessary pages */
+			for(a = addr;a < end_addr ; a += PAGE_SIZE) {
+				struct page *p = alloc_page_vma(GFP_HIGHUSER, vma, a);
+
+				if (p) {
+					clear_user_highpage(p, a);
+					pagevec_add(&pv,p);
+				} else
+					break;
+			}
+			end_addr = a;
+
+			spin_lock(&mm->page_table_lock);
+
+ 			for(p = pv.pages; addr < end_addr; addr += PAGE_SIZE, p++) {
+
+				page_table = pte_offset_map(pmd, addr);
+				if (!pte_none(*page_table)) {
+					/* Someone else got there first */
+					page_cache_release(*p);
+					pte_unmap(page_table);
+					continue;
+				}
+
+ 				entry = maybe_mkwrite(pte_mkdirty(mk_pte(*p,
+ 							 vma->vm_page_prot)),
+ 						      vma);
+
+				mm->rss++;
+				lru_cache_add_active(*p);
+				mark_page_accessed(*p);
+				page_add_anon_rmap(*p, vma, addr);
+
+				set_pte(page_table, entry);
+				pte_unmap(page_table);
+
+ 				/* No need to invalidate - it was non-present before */
+ 				update_mmu_cache(vma, addr, entry);
+			}
+ 		} else {
+ 			/* Read */
+ 			for(;addr < end_addr; addr += PAGE_SIZE) {
+				page_table = pte_offset_map(pmd, addr);
+ 				entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+				set_pte(page_table, entry);
+				pte_unmap(page_table);
+
+ 				/* No need to invalidate - it was non-present before */
+				update_mmu_cache(vma, addr, entry);
+
+			};
+		}
+		spin_unlock(&mm->page_table_lock);
+		return VM_FAULT_MINOR;
+	}
+
+	current->anon_fault_next_addr = addr + PAGE_SIZE;
+	current->anon_fault_order = 0;
+
+	page = ZERO_PAGE(addr);
 	/* Read-only mapping of ZERO_PAGE. */
 	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));

Re: Anticipatory prefaulting in the page fault handler V1

[Jesse Barnes]

On Wednesday, December 8, 2004 9:24 am, Christoph Lameter wrote:
> Page fault scalability patch and prefaulting. Max prefault order
> increased to 5 (max preallocation of 32 pages):
>
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
> 256  10    8   33.571s   4516.293s 863.021s 36874.099 194356.930
> 256  10   16   33.103s   3737.688s 461.028s 44492.553 363704.484
> 256  10   32   35.094s   3436.561s 321.080s 48326.262 521352.840
> 256  10   64   46.675s   2899.997s 245.020s 56936.124 684214.256
> 256  10  128   85.493s   2890.198s 203.008s 56380.890 826122.524
> 256  10  256   74.299s   1374.973s  99.088s115762.963 1679630.272
> 256  10  512   62.760s    706.559s  53.027s218078.311 3149273.714
>
> We are getting into an almost linear scalability in the high end with
> both patches and end up with a fault rate > 3 mio faults per second.

Nice results!  Any idea how many applications benefit from this sort of 
anticipatory faulting?  It has implications for NUMA allocation.  Imagine an 
app that allocates a large virtual address space and then tries to fault in 
pages near each CPU in turn.  With this patch applied, CPU 2 would be 
referencing pages near CPU 1, and CPU 3 would then fault in 4 pages, which 
would then be used by CPUs 4-6.  Unless I'm missing something...

And again, I'm not sure how important that is, maybe this approach will work 
well in the majority of cases (obviously it's a big win in faults/sec for 
your benchmark, but I wonder about subsequent references from other CPUs to 
those pages).  You can look at /sys/devices/platform/nodeN/meminfo to see 
where the pages are coming from.

Jesse

Re: Anticipatory prefaulting in the page fault handler V1

[Dave Hansen]

On Wed, 2004-12-08 at 09:24, Christoph Lameter wrote:
> The page fault handler for anonymous pages can generate significant overhead
> apart from its essential function which is to clear and setup a new page
> table entry for a never accessed memory location. This overhead increases
> significantly in an SMP environment.

do_anonymous_page() is a relatively compact function at this point. 
This would probably be a lot more readable if it was broken out into at
least another function or two that do_anonymous_page() calls into.  That
way, you also get a much cleaner separation if anyone needs to turn it
off in the future.  

Speaking of that, have you seen this impair performance on any other
workloads?  

-- Dave

Re: Anticipatory prefaulting in the page fault handler V1

[Christoph Lameter]

On Wed, 8 Dec 2004, Jesse Barnes wrote:

> Nice results!  Any idea how many applications benefit from this sort of
> anticipatory faulting?  It has implications for NUMA allocation.  Imagine an
> app that allocates a large virtual address space and then tries to fault in
> pages near each CPU in turn.  With this patch applied, CPU 2 would be
> referencing pages near CPU 1, and CPU 3 would then fault in 4 pages, which
> would then be used by CPUs 4-6.  Unless I'm missing something...

Faults are predicted for each thread executing on a different processor.
So each processor does its own predictions which will not generate
preallocations on a different processor (unless the thread is moved to
another processor but that is a very special situation).

> And again, I'm not sure how important that is, maybe this approach will work
> well in the majority of cases (obviously it's a big win in faults/sec for
> your benchmark, but I wonder about subsequent references from other CPUs to
> those pages).  You can look at /sys/devices/platform/nodeN/meminfo to see
> where the pages are coming from.

The origin of the pages has not changed and the existing locality
constraints are observed.

A patch like this is important for applications that allocate and preset
large amounts of memory on startup. It will drastically reduce the startup
times.

Re: Anticipatory prefaulting in the page fault handler V1

[Jesse Barnes]

On Wednesday, December 8, 2004 9:56 am, Christoph Lameter wrote:
> > And again, I'm not sure how important that is, maybe this approach will
> > work well in the majority of cases (obviously it's a big win in
> > faults/sec for your benchmark, but I wonder about subsequent references
> > from other CPUs to those pages).  You can look at
> > /sys/devices/platform/nodeN/meminfo to see where the pages are coming
> > from.
>
> The origin of the pages has not changed and the existing locality
> constraints are observed.
>
> A patch like this is important for applications that allocate and preset
> large amounts of memory on startup. It will drastically reduce the startup
> times.

Ok, that sounds good.  My case was probably a bit contrived, but I'm glad to 
see that you had already thought of it anyway.

Jesse

Re: Anticipatory prefaulting in the page fault handler V1

[Martin J. Bligh]

> The page fault handler for anonymous pages can generate significant overhead
> apart from its essential function which is to clear and setup a new page
> table entry for a never accessed memory location. This overhead increases
> significantly in an SMP environment.
> 
> In the page table scalability patches, we addressed the issue by changing
> the locking scheme so that multiple fault handlers are able to be processed
> concurrently on multiple cpus. This patch attempts to aggregate multiple
> page faults into a single one. It does that by noting
> anonymous page faults generated in sequence by an application.
> 
> If a fault occurred for page x and is then followed by page x+1 then it may
> be reasonable to expect another page fault at x+2 in the future. If page
> table entries for x+1 and x+2 would be prepared in the fault handling for
> page x+1 then the overhead of taking a fault for x+2 is avoided. However
> page x+2 may never be used and thus we may have increased the rss
> of an application unnecessarily. The swapper will take care of removing
> that page if memory should get tight.
> 
> The following patch makes the anonymous fault handler anticipate future
> faults. For each fault a prediction is made where the fault would occur
> (assuming linear acccess by the application). If the prediction turns out to
> be right (next fault is where expected) then a number of pages is
> preallocated in order to avoid a series of future faults. The order of the
> preallocation increases by the power of two for each success in sequence.
> 
> The first successful prediction leads to an additional page being allocated.
> Second successful prediction leads to 2 additional pages being allocated.
> Third to 4 pages and so on. The max order is 3 by default. In a large
> continous allocation the number of faults is reduced by a factor of 8.
> 
> The patch may be combined with the page fault scalability patch (another
> edition of the patch is needed which will be forthcoming after the
> page fault scalability patch has been included). The combined patches
> will triple the possible page fault rate from ~1 mio faults sec to 3 mio
> faults sec.
> 
> Standard Kernel on a 512 Cpu machine allocating 32GB with an increasing
> number of threads (and thus increasing parallellism of page faults):

Mmmm ... we tried doing this before for filebacked pages by sniffing the
pagecache, but it crippled forky workloads (like kernel compile) with the 
extra cost in zap_pte_range, etc. 

Perhaps the locality is better for the anon stuff, but the cost is also
higher. Exactly what benchmark were you running on this? If you just run
a microbenchmark that allocates memory, then it will definitely be faster.
On other things, I suspect not ...

M.

Re: Anticipatory prefaulting in the page fault handler V1

[David S. Miller]

On Wed, 8 Dec 2004 09:56:00 -0800 (PST)
Christoph Lameter <clameter@sgi.com> wrote:

> A patch like this is important for applications that allocate and preset
> large amounts of memory on startup. It will drastically reduce the startup
> times.

I see.  Yet I noticed that while the patch makes system time decrease,
for some reason the wall time is increasing with the patch applied.
Why is that, or am I misreading your tables?

Re: Anticipatory prefaulting in the page fault handler V1

[Linus Torvalds]

On Wed, 8 Dec 2004, David S. Miller wrote:
> 
> I see.  Yet I noticed that while the patch makes system time decrease,
> for some reason the wall time is increasing with the patch applied.
> Why is that, or am I misreading your tables?

I assume that you're looking at the final "both patches applied" case.

It has ten repetitions, while the other two tables only have three. That 
would explain the discrepancy.

		Linus

Re: Anticipatory prefaulting in the page fault handler V1

[Martin J. Bligh]

> The page fault handler for anonymous pages can generate significant overhead
> apart from its essential function which is to clear and setup a new page
> table entry for a never accessed memory location. This overhead increases
> significantly in an SMP environment.
> 
> In the page table scalability patches, we addressed the issue by changing
> the locking scheme so that multiple fault handlers are able to be processed
> concurrently on multiple cpus. This patch attempts to aggregate multiple
> page faults into a single one. It does that by noting
> anonymous page faults generated in sequence by an application.
> 
> If a fault occurred for page x and is then followed by page x+1 then it may
> be reasonable to expect another page fault at x+2 in the future. If page
> table entries for x+1 and x+2 would be prepared in the fault handling for
> page x+1 then the overhead of taking a fault for x+2 is avoided. However
> page x+2 may never be used and thus we may have increased the rss
> of an application unnecessarily. The swapper will take care of removing
> that page if memory should get tight.

I tried benchmarking it ... but processes just segfault all the time. 
Any chance you could try it out on SMP ia32 system?

M.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Christoph Lameter wrote:
> Changes from V11->V12 of this patch:
> - dump sloppy_rss in favor of list_rss (Linus' proposal)
> - keep up against current Linus tree (patch is based on 2.6.10-rc2-bk14)
> 

[snip]

> For more than 8 cpus the page fault rate increases by orders
> of magnitude. For more than 64 cpus the improvement in performace
> is 10 times better.

Those numbers are pretty impressive. I thought you'd said with earlier
patches that performance was about doubled from 8 to 512 CPUS. Did I
remember correctly? If so, where is the improvement coming from? The
per-thread RSS I guess?


On another note, these patches are basically only helpful to new
anonymous page faults. I guess this is the main thing you are concerned
about at the moment, but I wonder if you would see improvements with
my patch to remove the ptl from the other types of faults as well?

The downside of my patch - well the main downsides - compared to yours
are its intrusiveness, and the extra cost involved in copy_page_range
which yours appears not to require.

As I've said earlier though, I wouldn't mind your patches going in. At
least they should probably get into -mm soon, when Andrew has time (and
after the 4level patches are sorted out). That wouldn't stop my patch
(possibly) being merged some time after that if and when it was found
worthy...

Re: Anticipatory prefaulting in the page fault handler V1

[Pavel Machek]

Hi!

> Standard Kernel on a 512 Cpu machine allocating 32GB with an increasing
> number of threads (and thus increasing parallellism of page faults):
> 
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>  32   3    1    1.416s    138.165s 139.050s 45073.831  45097.498
...
> Patched kernel:
> 
> Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>  32   3    1    1.098s    138.544s 139.063s 45053.657  45057.920
...
> These number are roughly equal to what can be accomplished with the
> page fault scalability patches.
> 
> Kernel patches with both the page fault scalability patches and
> prefaulting:
> 
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>  32  10    1    4.103s    456.384s 460.046s 45541.992  45544.369
...
> 
> The fault rate doubles when both patches are applied.
...
> We are getting into an almost linear scalability in the high end with
> both patches and end up with a fault rate > 3 mio faults per second.

Well, with both patches you also slow single-threaded case more than
twice. What are the effects of this patch on UP system?
								Pavel

-- 
People were complaining that M$ turns users into beta-testers...
...jr ghea gurz vagb qrirybcref, naq gurl frrz gb yvxr vg gung jnl!

Re: Anticipatory prefaulting in the page fault handler V1

[Nick Piggin]

Pavel Machek wrote:
> Hi!
> 
> 
>>Standard Kernel on a 512 Cpu machine allocating 32GB with an increasing
>>number of threads (and thus increasing parallellism of page faults):
>>
>> Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>> 32   3    1    1.416s    138.165s 139.050s 45073.831  45097.498
> 
> ...
> 
>>Patched kernel:
>>
>>Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>> 32   3    1    1.098s    138.544s 139.063s 45053.657  45057.920
> 
> ...
> 
>>These number are roughly equal to what can be accomplished with the
>>page fault scalability patches.
>>
>>Kernel patches with both the page fault scalability patches and
>>prefaulting:
>>
>> Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>> 32  10    1    4.103s    456.384s 460.046s 45541.992  45544.369
> 
> ...
> 
>>The fault rate doubles when both patches are applied.
> 
> ...
> 
>>We are getting into an almost linear scalability in the high end with
>>both patches and end up with a fault rate > 3 mio faults per second.
> 
> 
> Well, with both patches you also slow single-threaded case more than
> twice. What are the effects of this patch on UP system?

fault/wsec is the important number.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Christoph Lameter]

On Thu, 9 Dec 2004, Nick Piggin wrote:

> > For more than 8 cpus the page fault rate increases by orders
> > of magnitude. For more than 64 cpus the improvement in performace
> > is 10 times better.
>
> Those numbers are pretty impressive. I thought you'd said with earlier
> patches that performance was about doubled from 8 to 512 CPUS. Did I
> remember correctly? If so, where is the improvement coming from? The
> per-thread RSS I guess?

Right. The per-thread RSS seems to have made a big difference for high CPU
counts. Also I was conservative in the estimates in earlier post since I
did not have the numbers for the very high cpu counts.

> On another note, these patches are basically only helpful to new
> anonymous page faults. I guess this is the main thing you are concerned
> about at the moment, but I wonder if you would see improvements with
> my patch to remove the ptl from the other types of faults as well?

I can try that but I am frankly a bit sceptical since the ptl protects
many other variables. It may be more efficient to have the ptl in these
cases than doing the atomic ops all over the place. Do you have any number
you could post? I believe I send you a copy of the code that I use for
performance tests last week or so,

> The downside of my patch - well the main downsides - compared to yours
> are its intrusiveness, and the extra cost involved in copy_page_range
> which yours appears not to require.

Is the patch known to be okay for ia64? I can try to see how it
does.

> As I've said earlier though, I wouldn't mind your patches going in. At
> least they should probably get into -mm soon, when Andrew has time (and
> after the 4level patches are sorted out). That wouldn't stop my patch
> (possibly) being merged some time after that if and when it was found
> worthy...

I'd certainly be willing to poke around and see how beneficial this is. If
it turns out to accellerate other functionality of the vm then you
have my full support.

Re: Anticipatory prefaulting in the page fault handler V1

[Christoph Lameter]

On Thu, 9 Dec 2004, Pavel Machek wrote:

> Hi!
>
> > Standard Kernel on a 512 Cpu machine allocating 32GB with an increasing
> > number of threads (and thus increasing parallellism of page faults):
> >
> >  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
> >  32   3    1    1.416s    138.165s 139.050s 45073.831  45097.498
> ...
> > Patched kernel:
> >
> > Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
> >  32   3    1    1.098s    138.544s 139.063s 45053.657  45057.920
> ...
> > These number are roughly equal to what can be accomplished with the
> > page fault scalability patches.
> >
> > Kernel patches with both the page fault scalability patches and
> > prefaulting:
> >
> >  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
> >  32  10    1    4.103s    456.384s 460.046s 45541.992  45544.369
> ...
> >
> > The fault rate doubles when both patches are applied.
> ...
> > We are getting into an almost linear scalability in the high end with
> > both patches and end up with a fault rate > 3 mio faults per second.
>
> Well, with both patches you also slow single-threaded case more than
> twice. What are the effects of this patch on UP system?

The faults per second are slightly increased, so its faster. The last
numbers are 10  repetitions and not 3. Do not look at the wall time.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 6655 bytes --]

On Wed, 1 Dec 2004, Christoph Lameter wrote:
> 
> Changes from V11->V12 of this patch:
> - dump sloppy_rss in favor of list_rss (Linus' proposal)
> - keep up against current Linus tree (patch is based on 2.6.10-rc2-bk14)
> 
> This is a series of patches that increases the scalability of
> the page fault handler for SMP. Here are some performance results
> on a machine with 512 processors allocating 32 GB with an increasing
> number of threads (that are assigned a processor each).

Your V12 patches would apply well to 2.6.10-rc3, except that (as noted
before) your mailer or whatever is eating trailing whitespace: trivial
patch attached to apply before yours, removing that whitespace so yours
apply.  But what your patches need to apply to would be 2.6.10-mm.

Your i386 HIGHMEM64G 3level ptep_cmpxchg forgets to use cmpxchg8b, would
have tested out okay up to 4GB but not above: trivial patch attached.

Your scalability figures show a superb improvement.  But they are (I
presume) for the best case: intense initial faulting of distinct areas
of anonymous memory by parallel cpus running a multithreaded process.
This is not a common case: how much do what real-world apps benefit?

Since you also avoid taking the page_table_lock in handle_pte_fault,
there should be some scalability benefit to all kinds of page fault:
do you have any results to show how much (perhaps hard to quantify,
since even tmpfs file faults introduce other scalability issues)?

How do the scalability figures compare if you omit patch 7/7 i.e. revert
the per-task rss complications you added in for Linus?  I remain a fan
of sloppy rss, which you earlier showed to be accurate enough (I'd say),
though I guess should be checked on other architectures than your ia64.
I can't see the point of all that added ugliness for numbers which don't
need to be precise - but perhaps there's no way of rearranging fields,
and the point at which mm->(anon_)rss is updated (near up of mmap_sem?),
to avoid destructive cacheline bounce.  What I'm asking is, do you have
numbers to support 7/7?  Perhaps it's the fact you showed up to 512 cpus
this time, but only up to 32 with sloppy rss?  The ratios do look better
with the latest, but the numbers are altogether lower so we don't know.

The split rss patch, if it stays, needs some work.  For example,
task_statm uses "get_shared" to total up rss-anon_rss from the tasks,
but assumes mm->rss is already accurate.  Scrap the separate get_rss,
get_anon_rss, get_shared functions: just one get_rss to make a single
pass through the tasks adding up both rss and anon_rss at the same time.

I am bothered that every read of /proc/<pid>/status or /proc/<pid>/statm
is going to reread through all of that task_list each time; yet in that
massively parallel case that concerns you, there should be little change
to rss after startup.  Perhaps a later optimization would be to avoid
task_list completely for singly threaded processes.  I'd like get_rss to
update mm->rss and mm->anon_rss and flag it uptodate to avoid subsequent
task_list iterations, but the locking might defeat your whole purpose.

Updating current->rss in do_anonymous_page, current->anon_rss in
page_add_anon_rmap, is not always correct: ptrace's access_process_vm
uses get_user_pages on another task.  You need check that current->mm ==
mm (or vma->vm_mm) before incrementing current->rss or current->anon_rss,
fall back to mm (or vma->vm_mm) in rare case not (taking page_table_lock
for that).  You'll also need to check !(current->flags & PF_BORROWED_MM),
to guard against use_mm.  Or... just go back to sloppy rss.

Moving to the main patch, 1/7, the major issue I see there is the way
do_anonymous_page does update_mmu_cache after setting the pte, without
any page_table_lock to bracket them together.  Obviously no problem on
architectures where update_mmu_cache is a no-op!  But although there's
been plenty of discussion, particularly with Ben and Nick, I've not
noticed anything to guarantee that as safe on all architectures.  I do
think it's fine for you to post your patches before completing hooks in
all the arches, but isn't this a significant issue which needs to be
sorted before your patches go into -mm?  You hazily refer to such issues
in 0/7, but now you need to work with arch maintainers to settle them
and show the patches.

A lesser issue with the reordering in do_anonymous_page: don't you need
to move the lru_cache_add_active after the page_add_anon_rmap, to avoid
the very slight chance that vmscan will pick the page off the LRU and
unmap it before you've counted it in, hitting page_remove_rmap's
BUG_ON(page_mapcount(page) < 0)?

(I do wonder why do_anonymous_page calls mark_page_accessed as well as
lru_cache_add_active.  The other instances of lru_cache_add_active for
an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
why here?  But that's nothing new with your patch, and although you've
reordered the calls, the final page state is the same as before.)

Where handle_pte_fault does "entry = *pte" without page_table_lock:
you're quite right to passing down precisely that entry to the fault
handlers below, but there's still a problem on the 32bit architectures
supporting 64bit ptes (i386, mips, ppc), that the upper and lower ints
of entry may be out of synch.  Not a problem for do_anonymous_page, or
anything else relying on ptep_cmpxchg to check; but a problem for
do_wp_page (which could find !pfn_valid and kill the process) and
probably others (harder to think through).  Your 4/7 patch for i386 has
an unused atomic get_64bit function from Nick, I think you'll have to
define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.

Hmm, that will only work if you're using atomic set_64bit rather than
relying on page_table_lock in the complementary places which matter.
Which I believe you are indeed doing in your 3level set_pte.  Shouldn't
__set_64bit be using LOCK_PREFIX like __get_64bit, instead of lock?

But by making every set_pte use set_64bit, you are significantly slowing
down many operations which do not need that atomicity.  This is quite
visible in the fork/exec/shell results from lmbench on i386 PAE (and is
the only interesting difference, for good or bad, that I noticed with
your patches in lmbench on 2*HT*P4), which run 5-20% slower.  There are
no faults on dst mm (nor on src mm) while copy_page_range is copying,
so its set_ptes don't need to be atomic; likewise during zap_pte_range
(either mmap_sem is held exclusively, or it's in the final exit_mmap).
Probably revert set_pte and set_pte_atomic to what they were, and use
set_pte_atomic where it's needed.

Hugh

[-- Attachment #2: Remove trailing whitespace before C.L. patches --]
[-- Type: TEXT/PLAIN, Size: 1736 bytes --]

--- 2.6.10-rc3/include/asm-i386/system.h	2004-11-15 16:21:12.000000000 +0000
+++ linux/include/asm-i386/system.h	2004-11-22 14:44:30.761904592 +0000
@@ -273,9 +273,9 @@ static inline unsigned long __cmpxchg(vo
 #define cmpxchg(ptr,o,n)\
 	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
 					(unsigned long)(n),sizeof(*(ptr))))
-    
+
 #ifdef __KERNEL__
-struct alt_instr { 
+struct alt_instr {
 	__u8 *instr; 		/* original instruction */
 	__u8 *replacement;
 	__u8  cpuid;		/* cpuid bit set for replacement */
--- 2.6.10-rc3/include/asm-s390/pgalloc.h	2004-05-10 03:33:39.000000000 +0100
+++ linux/include/asm-s390/pgalloc.h	2004-11-22 14:54:43.704723120 +0000
@@ -99,7 +99,7 @@ static inline void pgd_populate(struct m
 
 #endif /* __s390x__ */
 
-static inline void 
+static inline void
 pmd_populate_kernel(struct mm_struct *mm, pmd_t *pmd, pte_t *pte)
 {
 #ifndef __s390x__
--- 2.6.10-rc3/mm/memory.c	2004-11-18 17:56:11.000000000 +0000
+++ linux/mm/memory.c	2004-11-22 14:39:33.924030808 +0000
@@ -1424,7 +1424,7 @@ out:
 /*
  * We are called with the MM semaphore and page_table_lock
  * spinlock held to protect against concurrent faults in
- * multithreaded programs. 
+ * multithreaded programs.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
@@ -1615,7 +1615,7 @@ static int do_file_page(struct mm_struct
 	 * Fall back to the linear mapping if the fs does not support
 	 * ->populate:
 	 */
-	if (!vma->vm_ops || !vma->vm_ops->populate || 
+	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
 		return do_no_page(mm, vma, address, write_access, pte, pmd);


[-- Attachment #3: 3level ptep_cmpxchg use cmpxchg8b --]
[-- Type: TEXT/PLAIN, Size: 570 bytes --]

--- 2.6.10-rc3-cl/include/asm-i386/pgtable-3level.h	2004-12-05 14:01:11.000000000 +0000
+++ linux/include/asm-i386/pgtable-3level.h	2004-12-09 13:17:44.000000000 +0000
@@ -147,7 +147,7 @@ static inline pmd_t pfn_pmd(unsigned lon
 
 static inline int ptep_cmpxchg(struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
 {
-	return cmpxchg((unsigned int *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+	return cmpxchg8b((unsigned long long *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
 }
 
 

Re: Anticipatory prefaulting in the page fault handler V1

[Christoph Lameter]

On Wed, 8 Dec 2004, Martin J. Bligh wrote:

> I tried benchmarking it ... but processes just segfault all the time.
> Any chance you could try it out on SMP ia32 system?

I tried it on my i386 system and it works fine. Sorry about the puny
memory sizes (the system is a PIII-450 with 384k memory)

clameter@schroedinger:~/pfault/code$ ./pft -t -b256000 -r3 -f1
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  0   3    1    0.000s      0.004s   0.000s 37407.481  29200.500
  0   3    2    0.002s      0.002s   0.000s 31177.059  27227.723

clameter@schroedinger:~/pfault/code$ uname -a
Linux schroedinger 2.6.10-rc3-bk3-prezero #8 SMP Wed Dec 8 15:22:28 PST
2004 i686 GNU/Linux

Could you send me your .config?

page fault scalability patch V12: rss tasklist vs sloppy rss

[Christoph Lameter]

On Thu, 9 Dec 2004, Hugh Dickins wrote:

> How do the scalability figures compare if you omit patch 7/7 i.e. revert
> the per-task rss complications you added in for Linus?  I remain a fan
> of sloppy rss, which you earlier showed to be accurate enough (I'd say),
> though I guess should be checked on other architectures than your ia64.
> I can't see the point of all that added ugliness for numbers which don't
> need to be precise - but perhaps there's no way of rearranging fields,
> and the point at which mm->(anon_)rss is updated (near up of mmap_sem?),
> to avoid destructive cacheline bounce.  What I'm asking is, do you have
> numbers to support 7/7?  Perhaps it's the fact you showed up to 512 cpus
> this time, but only up to 32 with sloppy rss?  The ratios do look better
> with the latest, but the numbers are altogether lower so we don't know.

Here is a full set of numbers for sloppy and tasklist. The sloppy version
is 2.6.9-rc2-bk14 with the prefault patch also applied and the tasklist
version is 2.6.9-rc2-bk12 w/o prefault (you can get the numbers of
2.6.9-rc2-bk12 w prefault in the post titled "anticipatory prefaulting
in the page fault handler")). Even with this handicap
tasklist is still slightly better! I would expect tasklist to increase in
importance for combination patches which increase the fault rate even
more. The tasklist is likely to be unavoidable once I get the prezeroing
patch debugged and integrated which should at least give us a peak pulse
performance for page faults > 5 mio faults /sec.

I was not also able to get the high numbers of > 3 mio faults with atomic
rss + prefaulting but was able to get that with tasklist + prefault. The
atomic version shares the locality problems with the sloppy approach.

sloppy (2.6.10-bk14-rss-sloppy-prefault):
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  1  10    1    0.040s      6.505s   6.054s100117.616 100072.760
  1  10    2    0.041s      7.394s   4.005s 88138.739 161535.358
  1  10    4    0.049s      7.863s   2.049s 82819.743 262839.190
  1  10    8    0.093s      8.657s   1.077s 74889.898 369606.184
  1  10   16    0.621s     13.278s   1.076s 47150.165 371506.561
  1  10   32    3.154s     35.337s   2.029s 17025.784 285469.956
  1  10   64   11.602s     77.548s   2.086s  7351.089 228908.831
  1  10  128   41.999s    217.106s   4.030s  2529.316 152087.458
  1  10  256   40.482s    106.627s   3.022s  4454.885 203363.548
  1  10  512   63.673s     61.361s   3.040s  5241.403 192528.941
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  4  10    1    0.176s     41.276s  41.045s 63238.628  63237.008
  4  10    2    0.154s     31.074s  16.095s 83943.753 154606.489
  4  10    4    0.193s     31.886s   9.096s 81715.471 263190.941
  4  10    8    0.210s     33.577s   6.061s 77584.707 396402.083
  4  10   16    0.473s     52.997s   6.036s 49025.701 411640.587
  4  10   32    3.331s    142.296s   7.093s 18000.934 330197.326
  4  10   64   10.820s    318.485s   8.088s  7960.503 295042.520
  4  10  128   56.012s    928.004s  12.037s  2664.019 211812.600
  4  10  256   46.197s    464.579s   7.026s  5132.263 360940.189
  4  10  512   57.396s    225.876s   4.081s  9254.125 544185.485
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 16  10    1    0.948s    221.167s 222.009s 47208.624  47212.786
 16  10    2    0.824s    205.021s 110.022s 50939.876  95134.456
 16  10    4    0.689s    168.670s  53.055s 61914.226 195802.740
 16  10    8    0.683s    137.278s  27.034s 76004.706 383471.968
 16  10   16    0.969s    216.288s  24.031s 48264.109 431329.422
 16  10   32    3.932s    587.987s  30.002s 17714.820 349219.905
 16  10   64   13.542s   1253.834s  32.051s  8273.588 322528.516
 16  10  128   54.197s   3161.896s  38.064s  3260.403 271357.849
 16  10  256   57.610s   1668.913s  21.038s  6073.335 490410.386
 16  10  512   36.721s    833.691s  11.069s 12046.872 896970.623
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32  10    1    2.080s    470.722s 472.075s 44355.728  44360.409
 32  10    2    1.836s    456.343s 242.088s 45771.267  86344.100
 32  10    4    1.671s    432.569s 131.065s 48294.609 159291.360
 32  10    8    1.457s    354.825s  71.027s 58862.070 294242.410
 32  10   16    1.660s    431.057s  48.038s 48464.636 433466.055
 32  10   32    3.639s   1190.388s  59.040s 17563.676 353012.708
 32  10   64   14.623s   2490.393s  63.040s  8371.808 330750.309
 32  10  128   68.481s   6415.265s  76.053s  3234.476 274023.655
 32  10  256   63.428s   3216.337s  39.044s  6394.212 531665.931
 32  10  512   50.693s   1644.307s  21.035s 12372.572 982183.559
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 64  10    1    4.457s   1021.948s1026.030s 40863.994  40868.119
 64  10    2    3.929s    994.825s 525.030s 41995.308  79844.658
 64  10    4    3.661s    931.523s 269.014s 44849.990 155838.443
 64  10    8    3.355s    858.565s 153.098s 48662.260 272381.402
 64  10   16    3.130s    904.485s 101.090s 46212.285 411581.778
 64  10   32    5.007s   2366.494s 116.079s 17686.275 359107.203
 64  10   64   17.472s   5195.222s 126.012s  8046.325 332545.646
 64  10  128   65.249s  12515.845s 147.053s  3333.815 284290.928
 64  10  256   61.328s   6706.566s  78.061s  6197.354 533523.711
 64  10  512   60.656s   3201.068s  39.095s 12859.162 1049637.054
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
128  10    8    7.481s   1875.297s 318.049s 44554.389 263386.340
128  10   16    7.128s   2048.919s 230.060s 40799.672 363757.736
128  10   32    9.584s   4758.868s 241.094s 17591.883 346711.571
128  10   64   17.955s  10135.674s 249.025s  8261.684 336547.279
128  10  128   66.939s  25006.914s 287.019s  3345.560 292086.404
128  10  256   62.454s  12892.242s 149.035s  6475.341 561653.696
128  10  512   59.082s   6456.965s  77.002s 12873.768 1089026.647
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
256  10    8   17.201s   4672.781s 860.094s 35772.446 194870.225
256  10   16   16.641s   5071.433s 588.076s 32973.603 284954.772
256  10   32   17.745s   9193.335s 478.005s 18214.166 350950.045
256  10   64   25.474s  20440.137s 510.037s  8197.759 328725.189
256  10  128   65.451s  50015.195s 572.044s  3350.040 293079.914
256  10  256   61.296s  25191.675s 290.084s  6643.660 576852.282
256  10  512   58.911s  12589.530s 149.012s 13264.255 1125015.367

tasklist (2.6.10-rc2-bk12-rss-tasklist):

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  1   3    1    0.045s      2.042s   2.009s 94121.837  94039.902
  1   3    2    0.053s      2.217s   1.022s 86554.869 160093.661
  1   3    4    0.036s      2.325s   0.074s 83261.622 265213.249
  1   3    8    0.065s      2.507s   0.053s 76404.784 370587.422
  1   3   16    0.168s      4.727s   0.057s 40152.877 341385.368
  1   3   32    0.829s     11.408s   0.070s 16066.277 280690.973
  1   3   64    4.324s     25.591s   0.093s  6571.995 209956.473
  1   3  128   19.370s     81.568s   1.055s  1947.799 126774.712
  1   3  256   13.042s     46.608s   1.009s  3295.950 179708.774
  1   3  512   19.410s     28.085s   0.092s  4139.454 211823.959
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  4   3    1    0.161s     12.698s  12.086s 61156.292  61149.853
  4   3    2    0.152s     10.469s   5.073s 74037.518 137039.041
  4   3    4    0.179s      9.401s   2.098s 82081.949 263750.289
  4   3    8    0.156s     10.194s   1.098s 75979.430 395361.526
  4   3   16    0.407s     18.084s   2.010s 42527.778 373673.111
  4   3   32    0.824s     44.316s   2.031s 17421.815 339975.566
  4   3   64    4.706s     96.587s   2.066s  7763.856 295588.217
  4   3  128   17.453s    259.672s   3.053s  2837.813 222395.530
  4   3  256   17.090s    136.816s   2.017s  5109.777 361440.098
  4   3  512   13.466s     78.242s   1.043s  8575.295 548859.306
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 16   3    1    0.678s     61.548s  62.023s 50551.998  50544.748
 16   3    2    0.691s     63.381s  34.027s 49095.790  91791.474
 16   3    4    0.663s     52.083s  16.086s 59639.041 186542.124
 16   3    8    0.585s     43.339s   9.031s 71614.583 337721.897
 16   3   16    0.744s     75.174s   8.003s 41435.328 391278.035
 16   3   32    1.713s    171.942s   8.086s 18114.674 354760.887
 16   3   64    4.720s    366.803s   9.055s  8467.079 329273.168
 16   3  128   22.637s    849.059s  10.093s  3608.741 287572.764
 16   3  256   15.849s    472.565s   6.009s  6440.683 515916.601
 16   3  512   15.479s    245.305s   3.046s 12062.521 909147.611
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 32   3    1    1.451s    140.151s 141.060s 44430.367  44428.115
 32   3    2    1.399s    136.349s  73.041s 45673.303  85699.793
 32   3    4    1.321s    129.760s  39.027s 47996.303 160197.217
 32   3    8    1.279s    100.648s  20.039s 61724.641 308454.557
 32   3   16    1.414s    153.975s  15.090s 40488.236 395681.716
 32   3   32    2.534s    337.021s  17.016s 18528.487 366445.400
 32   3   64    4.271s    709.872s  18.057s  8809.787 338656.440
 32   3  128   18.734s   1805.094s  21.084s  3449.586 288005.644
 32   3  256   14.698s    963.787s  11.078s  6429.787 534077.540
 32   3  512   15.299s    453.990s   5.098s 13406.321 1050416.414
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
 64   3    1    3.018s    301.014s 304.004s 41386.617  41384.901
 64   3    2    2.941s    296.780s 157.005s 41981.967  80116.179
 64   3    4    2.810s    280.803s  82.047s 44366.266 152575.551
 64   3    8    2.763s    268.745s  48.099s 46344.377 256813.576
 64   3   16    2.764s    332.029s  34.030s 37584.030 366744.317
 64   3   32    3.337s    704.321s  34.074s 17781.025 362195.710
 64   3   64    7.395s   1475.497s  36.078s  8485.379 342026.888
 64   3  128   22.227s   3188.934s  40.044s  3918.492 311115.971
 64   3  256   18.004s   1834.246s  21.093s  6793.308 573753.797
 64   3  512   19.367s    861.324s  10.099s 14287.531 1144168.224

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
128   3    4    5.857s    626.055s 189.010s 39824.798 133076.331
128   3    8    5.837s    592.587s 107.080s 42053.423 233443.791
128   3   16    5.852s    666.252s  71.008s 37443.301 354011.649
128   3   32    6.305s   1365.184s  69.075s 18349.259 360755.364
128   3   64    8.450s   2914.730s  72.046s  8609.057 347288.474
128   3  128   21.188s   6719.590s  79.078s  3733.370 315402.750
128   3  256   18.263s   3672.379s  43.049s  6818.817 578587.427
128   3  512   17.625s   1901.969s  22.082s 13109.967 1102629.479

128   3  256   24.035s   3392.117s  40.074s  7366.714 617628.607
128   3  512   17.000s   1820.242s  21.072s 13697.601 1158632.106

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
256   3    4   11.976s   1660.924s 514.023s 30086.443  97877.018
256   3    8   11.618s   1301.448s 223.063s 38331.361 225057.902
256   3   16   11.696s   1409.158s 148.074s 35423.488 338379.838
256   3   32   12.678s   2668.417s 140.042s 18772.788 358421.926
256   3   64   15.933s   5833.804s 145.068s  8604.085 345487.685
256   3  128   32.640s  13437.080s 159.079s  3736.651 314981.569
256   3  256   23.875s   6835.241s  81.007s  7337.919 620777.397
256   3  512   17.566s   3392.148s  41.003s 14761.249 1226507.319

256   3  256   21.314s   6648.629s  79.085s  7546.038 630270.726
256   3  512   15.994s   3400.378s  40.087s 14732.481 1231399.906

Re: page fault scalability patch V12: rss tasklist vs sloppy rss

[Andrew Morton]

Christoph Lameter <clameter@sgi.com> wrote:
>
> On Thu, 9 Dec 2004, Hugh Dickins wrote:
> 
> > How do the scalability figures compare if you omit patch 7/7 i.e. revert
> > the per-task rss complications you added in for Linus?  I remain a fan
> > of sloppy rss, which you earlier showed to be accurate enough (I'd say),
> > though I guess should be checked on other architectures than your ia64.
> > I can't see the point of all that added ugliness for numbers which don't
> > need to be precise - but perhaps there's no way of rearranging fields,
> > and the point at which mm->(anon_)rss is updated (near up of mmap_sem?),
> > to avoid destructive cacheline bounce.  What I'm asking is, do you have
> > numbers to support 7/7?  Perhaps it's the fact you showed up to 512 cpus
> > this time, but only up to 32 with sloppy rss?  The ratios do look better
> > with the latest, but the numbers are altogether lower so we don't know.
> 
> Here is a full set of numbers for sloppy and tasklist.

Yes, but that only tests the thing-which-you're-trying-to-improve.  We also
need to work out the impact of that tasklist walk on other people's worst
cases.

> sloppy (2.6.10-bk14-rss-sloppy-prefault):

It would be helpful if you could generate a breif summary of benchmarking
results as well as dumping the raw numbers, please.

Re: page fault scalability patch V12: rss tasklist vs sloppy rss

[William Lee Irwin III]

On Thu, Dec 09, 2004 at 02:02:37PM -0800, Christoph Lameter wrote:
> I was not also able to get the high numbers of > 3 mio faults with atomic
> rss + prefaulting but was able to get that with tasklist + prefault. The
> atomic version shares the locality problems with the sloppy approach.

The implementation of the atomic version at least improperly places
the counter's cacheline, so the results for that are gibberish.

Unless the algorithms being compared are properly implemented, they're
straw men, not valid comparisons.


-- wli

Re: page fault scalability patch V12: rss tasklist vs sloppy rss

[Christoph Lameter]

On Thu, 9 Dec 2004, William Lee Irwin III wrote:
> Unless the algorithms being compared are properly implemented, they're
> straw men, not valid comparisons.

Sloppy rss left the rss in the section of mm that contained the counters.
So that has a separate cacheline. The idea of putting the atomic ops in a
group was to only have one exclusive cacheline for mmap_sem and the rss.
Which could lead to more bouncing of a single cache line rather than
bouncing multiple cache lines less. But it seems to me that the problem
essentially remains the same if the rss counter is not split.

Re: page fault scalability patch V12: rss tasklist vs sloppy rss

[William Lee Irwin III]

On Thu, Dec 09, 2004 at 03:07:13PM -0800, Christoph Lameter wrote:
> Sloppy rss left the rss in the section of mm that contained the counters.
> So that has a separate cacheline. The idea of putting the atomic ops in a
> group was to only have one exclusive cacheline for mmap_sem and the rss.
> Which could lead to more bouncing of a single cache line rather than
> bouncing multiple cache lines less. But it seems to me that the problem
> essentially remains the same if the rss counter is not split.

The prior results Robin Holt cited were that the counter needed to be
in a different cacheline from the ->mmap_sem and ->page_table_lock.
We shouldn't need to evaluate splitting for the atomic RSS algorithm.

A faithful implementation would just move the atomic counters away from
the ->mmap_sem and ->page_table_lock (just shuffle some mm fields).
Obviously a complete set of results won't be needed unless it's very
surprisingly competitive with the stronger algorithms. Things should be
fine just making sure that behaves similarly to the one with the shared
cacheline with ->mmap_sem in the sense of having a curve of similar shape
on smaller systems. The absolute difference probably doesn't matter,
but there is something to prove, and the largest risk of not doing so
is exaggerating the low-end performance benefits of stronger algorithms.

-- wli

Re: page fault scalability patch V12: rss tasklist vs sloppy rss

[Christoph Lameter]

On Thu, 9 Dec 2004, William Lee Irwin III wrote:

> On Thu, Dec 09, 2004 at 03:07:13PM -0800, Christoph Lameter wrote:
> > Sloppy rss left the rss in the section of mm that contained the counters.
> > So that has a separate cacheline. The idea of putting the atomic ops in a
> > group was to only have one exclusive cacheline for mmap_sem and the rss.
> > Which could lead to more bouncing of a single cache line rather than
> > bouncing multiple cache lines less. But it seems to me that the problem
> > essentially remains the same if the rss counter is not split.
>
> The prior results Robin Holt cited were that the counter needed to be
> in a different cacheline from the ->mmap_sem and ->page_table_lock.
> We shouldn't need to evaluate splitting for the atomic RSS algorithm.

Ok. Then we would need rss and rss_anon on two additional cache lines?
Both rss and anon_rss on one line? mmap_sem and the page_table_lock also
each on different cache lines?

> A faithful implementation would just move the atomic counters away from
> the ->mmap_sem and ->page_table_lock (just shuffle some mm fields).
> Obviously a complete set of results won't be needed unless it's very
> surprisingly competitive with the stronger algorithms. Things should be
> fine just making sure that behaves similarly to the one with the shared
> cacheline with ->mmap_sem in the sense of having a curve of similar shape
> on smaller systems. The absolute difference probably doesn't matter,
> but there is something to prove, and the largest risk of not doing so
> is exaggerating the low-end performance benefits of stronger algorithms.

The advantage in the split rss solution is that it can be placed on the
same cacheline as other stuff from task that is already needed. So there
is minimal overhead involved. But I can certainly give it a spin and see
what the results are.

[OT:HUMOR] Re: Anticipatory prefaulting in the page fault handler V1

[Adam Heath]

On Thu, 9 Dec 2004, Christoph Lameter wrote:

> On Wed, 8 Dec 2004, Martin J. Bligh wrote:
>
> > I tried benchmarking it ... but processes just segfault all the time.
> > Any chance you could try it out on SMP ia32 system?
>
> I tried it on my i386 system and it works fine. Sorry about the puny
> memory sizes (the system is a PIII-450 with 384k memory)

You probably get such fast numbers, because I mean, how many pages can 384k of
memory have?  I only figure 96 total pages.

> clameter@schroedinger:~/pfault/code$ ./pft -t -b256000 -r3 -f1
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>   0   3    1    0.000s      0.004s   0.000s 37407.481  29200.500
>   0   3    2    0.002s      0.002s   0.000s 31177.059  27227.723
>
> clameter@schroedinger:~/pfault/code$ uname -a
> Linux schroedinger 2.6.10-rc3-bk3-prezero #8 SMP Wed Dec 8 15:22:28 PST
> 2004 i686 GNU/Linux

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Hugh Dickins wrote:
> On Wed, 1 Dec 2004, Christoph Lameter wrote:
> 
>>Changes from V11->V12 of this patch:
>>- dump sloppy_rss in favor of list_rss (Linus' proposal)
>>- keep up against current Linus tree (patch is based on 2.6.10-rc2-bk14)
>>
>>This is a series of patches that increases the scalability of
>>the page fault handler for SMP. Here are some performance results
>>on a machine with 512 processors allocating 32 GB with an increasing
>>number of threads (that are assigned a processor each).
> 
> 
> Your V12 patches would apply well to 2.6.10-rc3, except that (as noted
> before) your mailer or whatever is eating trailing whitespace: trivial
> patch attached to apply before yours, removing that whitespace so yours
> apply.  But what your patches need to apply to would be 2.6.10-mm.
> 
> Your i386 HIGHMEM64G 3level ptep_cmpxchg forgets to use cmpxchg8b, would
> have tested out okay up to 4GB but not above: trivial patch attached.
> 

That looks obviously correct. Probably the reason why Martin was
getting crashes.

[snip]

> Moving to the main patch, 1/7, the major issue I see there is the way
> do_anonymous_page does update_mmu_cache after setting the pte, without
> any page_table_lock to bracket them together.  Obviously no problem on
> architectures where update_mmu_cache is a no-op!  But although there's
> been plenty of discussion, particularly with Ben and Nick, I've not
> noticed anything to guarantee that as safe on all architectures.  I do
> think it's fine for you to post your patches before completing hooks in
> all the arches, but isn't this a significant issue which needs to be
> sorted before your patches go into -mm?  You hazily refer to such issues
> in 0/7, but now you need to work with arch maintainers to settle them
> and show the patches.
> 

Yep, the update_mmu_cache issue is real. There is a parallel problem
that is update_mmu_cache can be called on a pte who's page has since
been evicted and reused. Again, that looks safe on IA64, but maybe
not on other architectures.

It can be solved by moving lru_cache_add to after update_mmu_cache in
all cases but the "update accessed bit" type fault. I solved that by
simply defining that out for architectures that don't need it - a raced
fault will simply get repeated if need be.

> A lesser issue with the reordering in do_anonymous_page: don't you need
> to move the lru_cache_add_active after the page_add_anon_rmap, to avoid
> the very slight chance that vmscan will pick the page off the LRU and
> unmap it before you've counted it in, hitting page_remove_rmap's
> BUG_ON(page_mapcount(page) < 0)?
> 

That's what I had been doing too. Seems to be the right way to go.

> (I do wonder why do_anonymous_page calls mark_page_accessed as well as
> lru_cache_add_active.  The other instances of lru_cache_add_active for
> an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
> why here?  But that's nothing new with your patch, and although you've
> reordered the calls, the final page state is the same as before.)
> 
> Where handle_pte_fault does "entry = *pte" without page_table_lock:
> you're quite right to passing down precisely that entry to the fault
> handlers below, but there's still a problem on the 32bit architectures
> supporting 64bit ptes (i386, mips, ppc), that the upper and lower ints
> of entry may be out of synch.  Not a problem for do_anonymous_page, or
> anything else relying on ptep_cmpxchg to check; but a problem for
> do_wp_page (which could find !pfn_valid and kill the process) and
> probably others (harder to think through).  Your 4/7 patch for i386 has
> an unused atomic get_64bit function from Nick, I think you'll have to
> define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.
> 

Indeed. This was a real problem for my patch, definitely.

> Hmm, that will only work if you're using atomic set_64bit rather than
> relying on page_table_lock in the complementary places which matter.
> Which I believe you are indeed doing in your 3level set_pte.  Shouldn't
> __set_64bit be using LOCK_PREFIX like __get_64bit, instead of lock?
> 

That's what I was wondering. It could be that the actual 64-bit store is
still atomic without the lock prefix (just not the entire rmw), which I
think would be sufficient.

In that case, get_64bit may be able to drop the lock prefix as well.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Christoph Lameter wrote:
> On Thu, 9 Dec 2004, Nick Piggin wrote:
> 
> 
>>>For more than 8 cpus the page fault rate increases by orders
>>>of magnitude. For more than 64 cpus the improvement in performace
>>>is 10 times better.
>>
>>Those numbers are pretty impressive. I thought you'd said with earlier
>>patches that performance was about doubled from 8 to 512 CPUS. Did I
>>remember correctly? If so, where is the improvement coming from? The
>>per-thread RSS I guess?
> 
> 
> Right. The per-thread RSS seems to have made a big difference for high CPU
> counts. Also I was conservative in the estimates in earlier post since I
> did not have the numbers for the very high cpu counts.
> 

Ah OK.

> 
>>On another note, these patches are basically only helpful to new
>>anonymous page faults. I guess this is the main thing you are concerned
>>about at the moment, but I wonder if you would see improvements with
>>my patch to remove the ptl from the other types of faults as well?
> 
> 
> I can try that but I am frankly a bit sceptical since the ptl protects
> many other variables. It may be more efficient to have the ptl in these
> cases than doing the atomic ops all over the place. Do you have any number
> you could post? I believe I send you a copy of the code that I use for
> performance tests last week or so,
> 

Yep I have your test program. No real numbers because the biggest thing
I have to test on is a 4-way - there is improvement, but it is not so
impressive as your 512 way tests! :)

> 
>>The downside of my patch - well the main downsides - compared to yours
>>are its intrusiveness, and the extra cost involved in copy_page_range
>>which yours appears not to require.
> 
> 
> Is the patch known to be okay for ia64? I can try to see how it
> does.
> 

I think it just needs one small fix to the swapping code, and it should
be pretty stable. So in fact it would probably work for you as is (if you
don't swap), but I'd rather have something more stable before I ask you
to test. I'll try to find time to do that in the next few days.

> 
>>As I've said earlier though, I wouldn't mind your patches going in. At
>>least they should probably get into -mm soon, when Andrew has time (and
>>after the 4level patches are sorted out). That wouldn't stop my patch
>>(possibly) being merged some time after that if and when it was found
>>worthy...
> 
> 
> I'd certainly be willing to poke around and see how beneficial this is. If
> it turns out to accellerate other functionality of the vm then you
> have my full support.
> 

Great, thanks.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Nick Piggin wrote:

> Yep, the update_mmu_cache issue is real. There is a parallel problem
> that is update_mmu_cache can be called on a pte who's page has since
> been evicted and reused. Again, that looks safe on IA64, but maybe
> not on other architectures.
> 
> It can be solved by moving lru_cache_add to after update_mmu_cache in
> all cases but the "update accessed bit" type fault. I solved that by
> simply defining that out for architectures that don't need it - a raced
> fault will simply get repeated if need be.
> 

The page-freed-before-update_mmu_cache issue can be solved in that way,
not the set_pte and update_mmu_cache not performed under the same ptl
section issue that you raised.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Benjamin Herrenschmidt]

On Fri, 2004-12-10 at 15:54 +1100, Nick Piggin wrote:
> Nick Piggin wrote:
> 
> > Yep, the update_mmu_cache issue is real. There is a parallel problem
> > that is update_mmu_cache can be called on a pte who's page has since
> > been evicted and reused. Again, that looks safe on IA64, but maybe
> > not on other architectures.
> > 
> > It can be solved by moving lru_cache_add to after update_mmu_cache in
> > all cases but the "update accessed bit" type fault. I solved that by
> > simply defining that out for architectures that don't need it - a raced
> > fault will simply get repeated if need be.
> > 
> 
> The page-freed-before-update_mmu_cache issue can be solved in that way,
> not the set_pte and update_mmu_cache not performed under the same ptl
> section issue that you raised.

What is the problem with update_mmu_cache ? It doesn't need to be done
in the same lock section since it's approx. equivalent to a HW fault,
which doesn't take the ptl...

Ben.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Benjamin Herrenschmidt wrote:
> On Fri, 2004-12-10 at 15:54 +1100, Nick Piggin wrote:
> 
>>Nick Piggin wrote:
>>
>>The page-freed-before-update_mmu_cache issue can be solved in that way,
>>not the set_pte and update_mmu_cache not performed under the same ptl
>>section issue that you raised.
> 
> 
> What is the problem with update_mmu_cache ? It doesn't need to be done
> in the same lock section since it's approx. equivalent to a HW fault,
> which doesn't take the ptl...
> 

I don't think a problem has been observed, I think Hugh was just raising
it as a general issue.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

On Fri, 10 Dec 2004, Nick Piggin wrote:
> Benjamin Herrenschmidt wrote:
> > On Fri, 2004-12-10 at 15:54 +1100, Nick Piggin wrote:
> >>
> >>The page-freed-before-update_mmu_cache issue can be solved in that way,
> >>not the set_pte and update_mmu_cache not performed under the same ptl
> >>section issue that you raised.
> > 
> > What is the problem with update_mmu_cache ? It doesn't need to be done
> > in the same lock section since it's approx. equivalent to a HW fault,
> > which doesn't take the ptl...
> 
> I don't think a problem has been observed, I think Hugh was just raising
> it as a general issue.

That's right, I know little of the arches on which update_mmu_cache does
something, so cannot say that separation is a problem.  And I did see mail
from Ben a month ago in which he arrived at the conclusion that it's not a
problem - but assumed he was speaking for ppc and ppc64.  (He was also
writing in the context of your patches rather than Christoph's.)

Perhaps Ben has in mind a logical argument that if update_mmu_cache does
just what its name implies, then doing it under a separate acquisition
of page_table_lock cannot introduce incorrectness on any architecture.
Maybe, but I'd still rather we heard that from an expert in each of the
affected architectures.

As it stands in Christoph's patches, update_mmu_cache is sometimes
called inside page_table_lock and sometimes outside: I'd be surprised
if that doesn't require adjustment for some architecture.

Your idea to raise do_anonymous_page's update_mmu_cache before the
lru_cache_add_active sounds just right; perhaps it should then even be
subsumed into the architectural ptep_cmpxchg.  But once we get this far,
I do wonder again whether it's right to be changing the rules in
do_anonymous_page alone (Christoph's patches) rather than all the
other faults together (your patches).

But there's no doubt that the do_anonymous_page case is easier,
or more obviously easy, to deal with - it helps a lot to know
that the page cannot yet be exposed to vmscan.c and rmap.c.

Hugh

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Christoph Lameter]

Thank you for the thorough review of my patches. Comments below

On Thu, 9 Dec 2004, Hugh Dickins wrote:

> Your V12 patches would apply well to 2.6.10-rc3, except that (as noted
> before) your mailer or whatever is eating trailing whitespace: trivial
> patch attached to apply before yours, removing that whitespace so yours
> apply.  But what your patches need to apply to would be 2.6.10-mm.

I am still mystified as to why this is an issue at all. The patches apply
just fine to the kernel sources as is. I have patched kernels numerous
times with this patchset and never ran into any issue. quilt removes trailing
whitespace from patches when they are generated as far as I can tell.

Patches will be made against mm after Nick's modifications to the 4 level
patches are in.

> Your i386 HIGHMEM64G 3level ptep_cmpxchg forgets to use cmpxchg8b, would
> have tested out okay up to 4GB but not above: trivial patch attached.

Thanks for the patch.

> Your scalability figures show a superb improvement.  But they are (I
> presume) for the best case: intense initial faulting of distinct areas
> of anonymous memory by parallel cpus running a multithreaded process.
> This is not a common case: how much do what real-world apps benefit?

This is common during the startup of distributed applications on our large
machines. They seem to freeze for minutes on bootup. I am not sure how
much real-world apps benefit. The numbers show that the benefit would
mostly be for SMP applications. UP has only very minor improvements.

> Since you also avoid taking the page_table_lock in handle_pte_fault,
> there should be some scalability benefit to all kinds of page fault:
> do you have any results to show how much (perhaps hard to quantify,
> since even tmpfs file faults introduce other scalability issues)?

I have not done such tests (yet).

> The split rss patch, if it stays, needs some work.  For example,
> task_statm uses "get_shared" to total up rss-anon_rss from the tasks,
> but assumes mm->rss is already accurate.  Scrap the separate get_rss,
> get_anon_rss, get_shared functions: just one get_rss to make a single
> pass through the tasks adding up both rss and anon_rss at the same time.

Next rev will have that.

> Updating current->rss in do_anonymous_page, current->anon_rss in
> page_add_anon_rmap, is not always correct: ptrace's access_process_vm
> uses get_user_pages on another task.  You need check that current->mm ==
> mm (or vma->vm_mm) before incrementing current->rss or current->anon_rss,
> fall back to mm (or vma->vm_mm) in rare case not (taking page_table_lock
> for that).  You'll also need to check !(current->flags & PF_BORROWED_MM),
> to guard against use_mm.  Or... just go back to sloppy rss.

I will look into this issue.

> Moving to the main patch, 1/7, the major issue I see there is the way
> do_anonymous_page does update_mmu_cache after setting the pte, without
> any page_table_lock to bracket them together.  Obviously no problem on
> architectures where update_mmu_cache is a no-op!  But although there's
> been plenty of discussion, particularly with Ben and Nick, I've not
> noticed anything to guarantee that as safe on all architectures.  I do
> think it's fine for you to post your patches before completing hooks in
> all the arches, but isn't this a significant issue which needs to be
> sorted before your patches go into -mm?  You hazily refer to such issues
> in 0/7, but now you need to work with arch maintainers to settle them
> and show the patches.

I have worked with a couple of arches and received feedback that was
integrated. I certainly welcome more feedback. A vague idea if there is
more trouble on that front: One could take the ptl in the cmpxchg
emulation and then unlock on update_mmu cache.

> A lesser issue with the reordering in do_anonymous_page: don't you need
> to move the lru_cache_add_active after the page_add_anon_rmap, to avoid
> the very slight chance that vmscan will pick the page off the LRU and
> unmap it before you've counted it in, hitting page_remove_rmap's
> BUG_ON(page_mapcount(page) < 0)?

Changed.

> (I do wonder why do_anonymous_page calls mark_page_accessed as well as
> lru_cache_add_active.  The other instances of lru_cache_add_active for
> an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
> why here?  But that's nothing new with your patch, and although you've
> reordered the calls, the final page state is the same as before.)

The mark_page_accessed is likely there avoid a future fault just to set
the accessed bit.

> Where handle_pte_fault does "entry = *pte" without page_table_lock:
> you're quite right to passing down precisely that entry to the fault
> handlers below, but there's still a problem on the 32bit architectures
> supporting 64bit ptes (i386, mips, ppc), that the upper and lower ints
> of entry may be out of synch.  Not a problem for do_anonymous_page, or
> anything else relying on ptep_cmpxchg to check; but a problem for
> do_wp_page (which could find !pfn_valid and kill the process) and
> probably others (harder to think through).  Your 4/7 patch for i386 has
> an unused atomic get_64bit function from Nick, I think you'll have to
> define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.

That would be a performance issue.

> Hmm, that will only work if you're using atomic set_64bit rather than
> relying on page_table_lock in the complementary places which matter.
> Which I believe you are indeed doing in your 3level set_pte.  Shouldn't
> __set_64bit be using LOCK_PREFIX like __get_64bit, instead of lock?

> But by making every set_pte use set_64bit, you are significantly slowing
> down many operations which do not need that atomicity.  This is quite
> visible in the fork/exec/shell results from lmbench on i386 PAE (and is
> the only interesting difference, for good or bad, that I noticed with
> your patches in lmbench on 2*HT*P4), which run 5-20% slower.  There are
> no faults on dst mm (nor on src mm) while copy_page_range is copying,
> so its set_ptes don't need to be atomic; likewise during zap_pte_range
> (either mmap_sem is held exclusively, or it's in the final exit_mmap).
> Probably revert set_pte and set_pte_atomic to what they were, and use
> set_pte_atomic where it's needed.

Good suggestions. Will see what I can do but I will need some assistence
my main platform is ia64 and the hardware and opportunities for testing on
i386 are limited.

Again thanks for the detailed review.

Re: pfault V12 : correction to tasklist rss

[Christoph Lameter]

On Thu, 9 Dec 2004, Hugh Dickins wrote:

> Updating current->rss in do_anonymous_page, current->anon_rss in
> page_add_anon_rmap, is not always correct: ptrace's access_process_vm
> uses get_user_pages on another task.  You need check that current->mm ==
> mm (or vma->vm_mm) before incrementing current->rss or current->anon_rss,
> fall back to mm (or vma->vm_mm) in rare case not (taking page_table_lock
> for that).  You'll also need to check !(current->flags & PF_BORROWED_MM),
> to guard against use_mm.  Or... just go back to sloppy rss.

Use_mm can simply attach the kernel thread to the mm via mm_add_thread
and will then update mm->rss when being detached again.

The issue with ptrace and get_user_pages is a bit thorny. I did the check
for mm = current->mm in the following patch. If mm != current->mm then
do the sloppy thing and increment mm->rss without the page table lock.
This should be a very special rare case.

One could also set current to the target task in get_user_pages but then
faults for the actual current task may increment the wrong counters. Could
we live with that?

Or simply leave as is. The pages are after all allocated by the ptrace
process and it should be held responsible for it.

My favorite rss solution is still just getting rid of rss and
anon_rss and do the long loops in procfs. Whichever process wants to
know better be willing to pay the price in cpu time and the code for
incrementing rss can be removed from the page fault handler.

We have no  real way of establishing the ownership of shared pages
anyways. Its counted when allocated. But the page may live on afterwards
in another process and then not be accounted for although its only user is
the new process. IMHO vm scans may be the only way of really getting an
accurate count.

But here is the improved list_rss patch:

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-12-06 17:23:55.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-12-10 11:39:00.000000000 -0800
@@ -30,6 +30,7 @@
 #include <linux/pid.h>
 #include <linux/percpu.h>
 #include <linux/topology.h>
+#include <linux/rcupdate.h>

 struct exec_domain;

@@ -217,6 +218,7 @@
 	int map_count;				/* number of VMAs */
 	struct rw_semaphore mmap_sem;
 	spinlock_t page_table_lock;		/* Protects page tables, mm->rss, mm->anon_rss */
+	long rss, anon_rss;

 	struct list_head mmlist;		/* List of maybe swapped mm's.  These are globally strung
 						 * together off init_mm.mmlist, and are protected
@@ -226,7 +228,7 @@
 	unsigned long start_code, end_code, start_data, end_data;
 	unsigned long start_brk, brk, start_stack;
 	unsigned long arg_start, arg_end, env_start, env_end;
-	unsigned long rss, anon_rss, total_vm, locked_vm, shared_vm;
+	unsigned long total_vm, locked_vm, shared_vm;
 	unsigned long exec_vm, stack_vm, reserved_vm, def_flags, nr_ptes;

 	unsigned long saved_auxv[42]; /* for /proc/PID/auxv */
@@ -236,6 +238,8 @@

 	/* Architecture-specific MM context */
 	mm_context_t context;
+	struct list_head task_list;		/* Tasks using this mm */
+	struct rcu_head rcu_head;		/* For freeing mm via rcu */

 	/* Token based thrashing protection. */
 	unsigned long swap_token_time;
@@ -545,6 +549,9 @@
 	struct list_head ptrace_list;

 	struct mm_struct *mm, *active_mm;
+	/* Split counters from mm */
+	long rss;
+	long anon_rss;

 /* task state */
 	struct linux_binfmt *binfmt;
@@ -578,6 +585,9 @@
 	struct completion *vfork_done;		/* for vfork() */
 	int __user *set_child_tid;		/* CLONE_CHILD_SETTID */
 	int __user *clear_child_tid;		/* CLONE_CHILD_CLEARTID */
+
+	/* List of other tasks using the same mm */
+	struct list_head mm_tasks;

 	unsigned long rt_priority;
 	unsigned long it_real_value, it_prof_value, it_virt_value;
@@ -1124,6 +1134,12 @@

 #endif

+void get_rss(struct mm_struct *mm, unsigned long *rss, unsigned long *anon_rss);
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk);
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk);
+
 #endif /* __KERNEL__ */

 #endif
+
Index: linux-2.6.9/fs/proc/task_mmu.c
===================================================================
--- linux-2.6.9.orig/fs/proc/task_mmu.c	2004-12-06 17:23:54.000000000 -0800
+++ linux-2.6.9/fs/proc/task_mmu.c	2004-12-10 11:39:00.000000000 -0800
@@ -6,8 +6,9 @@

 char *task_mem(struct mm_struct *mm, char *buffer)
 {
-	unsigned long data, text, lib;
+	unsigned long data, text, lib, rss, anon_rss;

+	get_rss(mm, &rss, &anon_rss);
 	data = mm->total_vm - mm->shared_vm - mm->stack_vm;
 	text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK)) >> 10;
 	lib = (mm->exec_vm << (PAGE_SHIFT-10)) - text;
@@ -22,7 +23,7 @@
 		"VmPTE:\t%8lu kB\n",
 		(mm->total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),
-		mm->rss << (PAGE_SHIFT-10),
+		rss << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
@@ -37,11 +38,14 @@
 int task_statm(struct mm_struct *mm, int *shared, int *text,
 	       int *data, int *resident)
 {
-	*shared = mm->rss - mm->anon_rss;
+	unsigned long rss, anon_rss;
+
+	get_rss(mm, &rss, &anon_rss);
+	*shared = rss - anon_rss;
 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
 								>> PAGE_SHIFT;
 	*data = mm->total_vm - mm->shared_vm;
-	*resident = mm->rss;
+	*resident = rss;
 	return mm->total_vm;
 }

Index: linux-2.6.9/fs/proc/array.c
===================================================================
--- linux-2.6.9.orig/fs/proc/array.c	2004-12-06 17:23:54.000000000 -0800
+++ linux-2.6.9/fs/proc/array.c	2004-12-10 11:39:00.000000000 -0800
@@ -302,7 +302,7 @@

 static int do_task_stat(struct task_struct *task, char * buffer, int whole)
 {
-	unsigned long vsize, eip, esp, wchan = ~0UL;
+	unsigned long rss, anon_rss, vsize, eip, esp, wchan = ~0UL;
 	long priority, nice;
 	int tty_pgrp = -1, tty_nr = 0;
 	sigset_t sigign, sigcatch;
@@ -325,6 +325,7 @@
 		vsize = task_vsize(mm);
 		eip = KSTK_EIP(task);
 		esp = KSTK_ESP(task);
+		get_rss(mm, &rss, &anon_rss);
 	}

 	get_task_comm(tcomm, task);
@@ -420,7 +421,7 @@
 		jiffies_to_clock_t(task->it_real_value),
 		start_time,
 		vsize,
-		mm ? mm->rss : 0, /* you might want to shift this left 3 */
+		mm ? rss : 0, /* you might want to shift this left 3 */
 	        rsslim,
 		mm ? mm->start_code : 0,
 		mm ? mm->end_code : 0,
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-12-10 11:11:26.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-12-10 11:46:07.000000000 -0800
@@ -263,8 +263,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -438,7 +436,10 @@
 	BUG_ON(PageReserved(page));
 	BUG_ON(!anon_vma);

-	vma->vm_mm->anon_rss++;
+	if (current->mm == vma->vm_mm)
+		current->anon_rss++;
+	else
+		vma->vm_mm->anon_rss++;

 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 	index = (address - vma->vm_start) >> PAGE_SHIFT;
@@ -510,8 +511,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -799,8 +798,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;
Index: linux-2.6.9/kernel/fork.c
===================================================================
--- linux-2.6.9.orig/kernel/fork.c	2004-12-06 17:23:55.000000000 -0800
+++ linux-2.6.9/kernel/fork.c	2004-12-10 11:39:00.000000000 -0800
@@ -151,6 +151,7 @@
 	*tsk = *orig;
 	tsk->thread_info = ti;
 	ti->task = tsk;
+	tsk->rss = 0;

 	/* One for us, one for whoever does the "release_task()" (usually parent) */
 	atomic_set(&tsk->usage,2);
@@ -292,6 +293,7 @@
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);
 	INIT_LIST_HEAD(&mm->mmlist);
+	INIT_LIST_HEAD(&mm->task_list);
 	mm->core_waiters = 0;
 	mm->nr_ptes = 0;
 	spin_lock_init(&mm->page_table_lock);
@@ -323,6 +325,13 @@
 	return mm;
 }

+static void rcu_free_mm(struct rcu_head *head)
+{
+	struct mm_struct *mm = container_of(head ,struct mm_struct, rcu_head);
+
+	free_mm(mm);
+}
+
 /*
  * Called when the last reference to the mm
  * is dropped: either by a lazy thread or by
@@ -333,7 +342,7 @@
 	BUG_ON(mm == &init_mm);
 	mm_free_pgd(mm);
 	destroy_context(mm);
-	free_mm(mm);
+	call_rcu(&mm->rcu_head, rcu_free_mm);
 }

 /*
@@ -400,6 +409,8 @@

 	/* Get rid of any cached register state */
 	deactivate_mm(tsk, mm);
+	if (mm)
+		mm_remove_thread(mm, tsk);

 	/* notify parent sleeping on vfork() */
 	if (vfork_done) {
@@ -447,8 +458,8 @@
 		 * new threads start up in user mode using an mm, which
 		 * allows optimizing out ipis; the tlb_gather_mmu code
 		 * is an example.
+		 * (mm_add_thread does use the ptl .... )
 		 */
-		spin_unlock_wait(&oldmm->page_table_lock);
 		goto good_mm;
 	}

@@ -470,6 +481,7 @@
 		goto free_pt;

 good_mm:
+	mm_add_thread(mm, tsk);
 	tsk->mm = mm;
 	tsk->active_mm = mm;
 	return 0;
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-10 11:12:44.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-10 11:45:00.000000000 -0800
@@ -1467,8 +1467,10 @@
 		 */
 		page_add_anon_rmap(page, vma, addr);
 		lru_cache_add_active(page);
-		mm->rss++;
-
+		if (current->mm == mm)
+			current->rss++;
+		else
+			mm->rss++;
 	}
 	pte_unmap(page_table);

@@ -1859,3 +1861,49 @@
 }

 #endif
+
+void get_rss(struct mm_struct *mm, unsigned long *rss, unsigned long *anon_rss)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss_sum, anon_rss_sum;
+
+	rcu_read_lock();
+	rss_sum = mm->rss;
+	anon_rss_sum = mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss_sum += t->rss;
+		anon_rss_sum += t->anon_rss;
+	}
+	if (rss_sum < 0)
+		rss_sum = 0;
+	if (anon_rss_sum < 0)
+		anon_rss_sum = 0;
+	rcu_read_unlock();
+	*rss = rss_sum;
+	*anon_rss = anon_rss_sum;
+}
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	if (!mm)
+		return;
+
+	spin_lock(&mm->page_table_lock);
+	mm->rss += tsk->rss;
+	mm->anon_rss += tsk->anon_rss;
+	list_del_rcu(&tsk->mm_tasks);
+	spin_unlock(&mm->page_table_lock);
+}
+
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	spin_lock(&mm->page_table_lock);
+	tsk->rss = 0;
+	tsk->anon_rss = 0;
+	list_add_rcu(&tsk->mm_tasks, &mm->task_list);
+	spin_unlock(&mm->page_table_lock);
+}
+
+
Index: linux-2.6.9/include/linux/init_task.h
===================================================================
--- linux-2.6.9.orig/include/linux/init_task.h	2004-12-06 17:23:55.000000000 -0800
+++ linux-2.6.9/include/linux/init_task.h	2004-12-10 11:39:00.000000000 -0800
@@ -42,6 +42,7 @@
 	.mmlist		= LIST_HEAD_INIT(name.mmlist),		\
 	.cpu_vm_mask	= CPU_MASK_ALL,				\
 	.default_kioctx = INIT_KIOCTX(name.default_kioctx, name),	\
+	.task_list	= LIST_HEAD_INIT(name.task_list),	\
 }

 #define INIT_SIGNALS(sig) {	\
@@ -112,6 +113,7 @@
 	.proc_lock	= SPIN_LOCK_UNLOCKED,				\
 	.switch_lock	= SPIN_LOCK_UNLOCKED,				\
 	.journal_info	= NULL,						\
+	.mm_tasks	= LIST_HEAD_INIT(tsk.mm_tasks),			\
 }


Index: linux-2.6.9/fs/exec.c
===================================================================
--- linux-2.6.9.orig/fs/exec.c	2004-12-06 17:23:54.000000000 -0800
+++ linux-2.6.9/fs/exec.c	2004-12-10 11:39:00.000000000 -0800
@@ -543,6 +543,7 @@
 	active_mm = tsk->active_mm;
 	tsk->mm = mm;
 	tsk->active_mm = mm;
+	mm_add_thread(mm, current);
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
 	arch_pick_mmap_layout(mm);
Index: linux-2.6.9/fs/aio.c
===================================================================
--- linux-2.6.9.orig/fs/aio.c	2004-12-06 17:23:54.000000000 -0800
+++ linux-2.6.9/fs/aio.c	2004-12-10 11:39:00.000000000 -0800
@@ -575,6 +575,7 @@
 	atomic_inc(&mm->mm_count);
 	tsk->mm = mm;
 	tsk->active_mm = mm;
+	mm_add_thread(mm, tsk);
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);

@@ -597,6 +598,7 @@
 	struct task_struct *tsk = current;

 	task_lock(tsk);
+	mm_remove_thread(mm,tsk);
 	tsk->flags &= ~PF_BORROWED_MM;
 	tsk->mm = NULL;
 	/* active_mm is still 'mm' */

Re: pfault V12 : correction to tasklist rss

[Hugh Dickins]

On Fri, 10 Dec 2004, Christoph Lameter wrote:
> On Thu, 9 Dec 2004, Hugh Dickins wrote:
> 
> > Updating current->rss in do_anonymous_page, current->anon_rss in
> > page_add_anon_rmap, is not always correct: ptrace's access_process_vm
> > uses get_user_pages on another task.  You need check that current->mm ==
> > mm (or vma->vm_mm) before incrementing current->rss or current->anon_rss,
> > fall back to mm (or vma->vm_mm) in rare case not (taking page_table_lock
> > for that).  You'll also need to check !(current->flags & PF_BORROWED_MM),
> > to guard against use_mm.  Or... just go back to sloppy rss.
> 
> Use_mm can simply attach the kernel thread to the mm via mm_add_thread
> and will then update mm->rss when being detached again.

True.  But please add and remove mm outside of the task_lock,
there's no need to nest page_table_lock within it, is there?

> The issue with ptrace and get_user_pages is a bit thorny. I did the check
> for mm = current->mm in the following patch. If mm != current->mm then
> do the sloppy thing and increment mm->rss without the page table lock.
> This should be a very special rare case.

I don't understand why you want to avoid taking mm->page_table_lock
in that special rare case.  I do prefer the sloppy rss approach, but if
you're trying to be exact then it's regrettable to leave sloppy corners.

Oh, is it because page_add_anon_rmap is usually called with page_table_lock,
but without in your do_anonymous_page case?  You'll have to move the
anon_rss incrementation out of page_add_anon_rmap to its callsites
(I was being a little bit lazy when I sited it in that one place,
it's probably better to do it near mm->rss anyway.)

> One could also set current to the target task in get_user_pages but then
> faults for the actual current task may increment the wrong counters. Could
> we live with that?

No, "current" is not nearly so easy to play with as that.
See i386.  Even if it were, you might get burnt for heresy.

> Or simply leave as is. The pages are after all allocated by the ptrace
> process and it should be held responsible for it.

No.

> My favorite rss solution is still just getting rid of rss and
> anon_rss and do the long loops in procfs. Whichever process wants to
> know better be willing to pay the price in cpu time and the code for
> incrementing rss can be removed from the page fault handler.

We all seem to have different favourites.  Your favourite makes
quite a few people very angry.  We've been there, we've done that,
we've no wish to return.  It'd be fine if just the process which
wants to know paid the price; but it's every other that has to pay.

> We have no  real way of establishing the ownership of shared pages
> anyways. Its counted when allocated. But the page may live on afterwards
> in another process and then not be accounted for although its only user is
> the new process.

I didn't understand that bit.

> IMHO vm scans may be the only way of really getting an accurate count.
> 
> But here is the improved list_rss patch:

Not studied in depth, but... am I going mad, or is your impressive
RCUing the wrong way round?  While we're scanning the list of tasks
sharing the mm, there's no danger of the mm vanishing, but there is
a danger of the task vanishing.  Isn't it therefore the task which
needs to be freed via RCU, not the mm?

Hugh

Re: pfault V12 : correction to tasklist rss

[Andrew Morton]

Hugh Dickins <hugh@veritas.com> wrote:
>
> > We have no  real way of establishing the ownership of shared pages
>  > anyways. Its counted when allocated. But the page may live on afterwards
>  > in another process and then not be accounted for although its only user is
>  > the new process.
> 
>  I didn't understand that bit.

We did lose some accounting accuracy when the pagetable walk and the big
tasklist walks were removed.  Bill would probably have more details.  Given
that the code as it stood was a complete showstopper, the tradeoff seemed
reasonable.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

On Fri, 10 Dec 2004, Christoph Lameter wrote:
> On Thu, 9 Dec 2004, Hugh Dickins wrote:
> 
> > Your V12 patches would apply well to 2.6.10-rc3, except that (as noted
> > before) your mailer or whatever is eating trailing whitespace: trivial
> > patch attached to apply before yours, removing that whitespace so yours
> > apply.  But what your patches need to apply to would be 2.6.10-mm.
> 
> I am still mystified as to why this is an issue at all. The patches apply
> just fine to the kernel sources as is. I have patched kernels numerous
> times with this patchset and never ran into any issue. quilt removes trailing
> whitespace from patches when they are generated as far as I can tell.

Perhaps you've only tried applying your original patches, not the ones
as received through the mail.  It discourages people from trying them
when "patch -p1" fails with rejects, however trivial.  Or am I alone
in seeing this?  never had such a problem with other patches before.

> > Your scalability figures show a superb improvement.  But they are (I
> > presume) for the best case: intense initial faulting of distinct areas
> > of anonymous memory by parallel cpus running a multithreaded process.
> > This is not a common case: how much do what real-world apps benefit?
> 
> This is common during the startup of distributed applications on our large
> machines. They seem to freeze for minutes on bootup. I am not sure how
> much real-world apps benefit. The numbers show that the benefit would
> mostly be for SMP applications. UP has only very minor improvements.

How much do your patches speed the startup of these applications?
Can you name them?

> I have worked with a couple of arches and received feedback that was
> integrated. I certainly welcome more feedback. A vague idea if there is
> more trouble on that front: One could take the ptl in the cmpxchg
> emulation and then unlock on update_mmu cache.

Or move the update_mmu_cache into the ptep_cmpxchg emulation perhaps.

> > (I do wonder why do_anonymous_page calls mark_page_accessed as well as
> > lru_cache_add_active.  The other instances of lru_cache_add_active for
> > an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
> > why here?  But that's nothing new with your patch, and although you've
> > reordered the calls, the final page state is the same as before.)
> 
> The mark_page_accessed is likely there avoid a future fault just to set
> the accessed bit.

No, mark_page_accessed is an operation on the struct page
(and the accessed bit of the pte is preset too anyway).

> > Where handle_pte_fault does "entry = *pte" without page_table_lock:
> > you're quite right to passing down precisely that entry to the fault
> > handlers below, but there's still a problem on the 32bit architectures
> > supporting 64bit ptes (i386, mips, ppc), that the upper and lower ints
> > of entry may be out of synch.  Not a problem for do_anonymous_page, or
> > anything else relying on ptep_cmpxchg to check; but a problem for
> > do_wp_page (which could find !pfn_valid and kill the process) and
> > probably others (harder to think through).  Your 4/7 patch for i386 has
> > an unused atomic get_64bit function from Nick, I think you'll have to
> > define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.
> 
> That would be a performance issue.

Sadly, yes, but correctness must take precedence over performance.
It may be possible to avoid it in most cases, doing the atomic
later when in doubt: but would need careful thought.

> Good suggestions. Will see what I can do but I will need some assistence
> my main platform is ia64 and the hardware and opportunities for testing on
> i386 are limited.

There's plenty of us can be trying i386.  It's other arches worrying me.

Hugh

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Hugh Dickins <hugh@veritas.com> wrote:
>
> > > (I do wonder why do_anonymous_page calls mark_page_accessed as well as
> > > lru_cache_add_active.  The other instances of lru_cache_add_active for
> > > an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
> > > why here?  But that's nothing new with your patch, and although you've
> > > reordered the calls, the final page state is the same as before.)
> > 
> > The mark_page_accessed is likely there avoid a future fault just to set
> > the accessed bit.
> 
> No, mark_page_accessed is an operation on the struct page
> (and the accessed bit of the pte is preset too anyway).

The point is a good one - I guess that code is a holdover from earlier
implementations.

This is equivalent, no?

--- 25/mm/memory.c~do_anonymous_page-use-setpagereferenced	Fri Dec 10 14:11:32 2004
+++ 25-akpm/mm/memory.c	Fri Dec 10 14:11:42 2004
@@ -1464,7 +1464,7 @@ do_anonymous_page(struct mm_struct *mm, 
 							 vma->vm_page_prot)),
 				      vma);
 		lru_cache_add_active(page);
-		mark_page_accessed(page);
+		SetPageReferenced(page);
 		page_add_anon_rmap(page, vma, addr);
 	}
 
_

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

On Fri, 10 Dec 2004, Andrew Morton wrote:
> Hugh Dickins <hugh@veritas.com> wrote:
> >
> > > > (I do wonder why do_anonymous_page calls mark_page_accessed as well as
> > > > lru_cache_add_active.  The other instances of lru_cache_add_active for
> > > > an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
> > > > why here?  But that's nothing new with your patch, and although you've
> > > > reordered the calls, the final page state is the same as before.)
> 
> The point is a good one - I guess that code is a holdover from earlier
> implementations.
> 
> This is equivalent, no?

Yes, it is equivalent to use SetPageReferenced(page) there instead.
But why is do_anonymous_page adding anything to lru_cache_add_active,
when its other callers leave it at that?  What's special about the
do_anonymous_page case?

Hugh

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Hugh Dickins <hugh@veritas.com> wrote:
>
> On Fri, 10 Dec 2004, Andrew Morton wrote:
> > Hugh Dickins <hugh@veritas.com> wrote:
> > >
> > > > > (I do wonder why do_anonymous_page calls mark_page_accessed as well as
> > > > > lru_cache_add_active.  The other instances of lru_cache_add_active for
> > > > > an anonymous page don't mark_page_accessed i.e. SetPageReferenced too,
> > > > > why here?  But that's nothing new with your patch, and although you've
> > > > > reordered the calls, the final page state is the same as before.)
> > 
> > The point is a good one - I guess that code is a holdover from earlier
> > implementations.
> > 
> > This is equivalent, no?
> 
> Yes, it is equivalent to use SetPageReferenced(page) there instead.
> But why is do_anonymous_page adding anything to lru_cache_add_active,
> when its other callers leave it at that?  What's special about the
> do_anonymous_page case?

do_swap_page() is effectively doing the same as do_anonymous_page(). 
do_wp_page() and do_no_page() appear to be errant.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

On Fri, 10 Dec 2004, Andrew Morton wrote:
> Hugh Dickins <hugh@veritas.com> wrote:
> > But why is do_anonymous_page adding anything to lru_cache_add_active,
> > when its other callers leave it at that?  What's special about the
> > do_anonymous_page case?
> 
> do_swap_page() is effectively doing the same as do_anonymous_page(). 
> do_wp_page() and do_no_page() appear to be errant.

Demur.  do_swap_page has to mark_page_accessed because the page from
the swap cache is already on the LRU, and for who knows how long.
The others (and count in fs/exec.c's install_arg_page) are dealing
with a freshly allocated page they are putting onto the active LRU.

My inclination would be simply to remove the mark_page_accessed
from do_anonymous_page; but I have no numbers to back that hunch.

Hugh

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Andrew Morton]

Hugh Dickins <hugh@veritas.com> wrote:
>
> On Fri, 10 Dec 2004, Andrew Morton wrote:
> > Hugh Dickins <hugh@veritas.com> wrote:
> > > But why is do_anonymous_page adding anything to lru_cache_add_active,
> > > when its other callers leave it at that?  What's special about the
> > > do_anonymous_page case?
> > 
> > do_swap_page() is effectively doing the same as do_anonymous_page(). 
> > do_wp_page() and do_no_page() appear to be errant.
> 
> Demur.  do_swap_page has to mark_page_accessed because the page from
> the swap cache is already on the LRU, and for who knows how long.

Well.  Some of the time.  If the page was just read from swap, it's known
to be on the active list.

> The others (and count in fs/exec.c's install_arg_page) are dealing
> with a freshly allocated page they are putting onto the active LRU.
> 
> My inclination would be simply to remove the mark_page_accessed
> from do_anonymous_page; but I have no numbers to back that hunch.
> 

With the current implementation of page_referenced() the
software-referenced bit doesn't matter anyway, as long as the pte's
referenced bit got set.  So as long as the thing is on the active list, we
can simply remove the mark_page_accessed() call.

Except one day the VM might get smarter about pages which are both
software-referenced and pte-referenced.

Re: pfault V12 : correction to tasklist rss

[William Lee Irwin III]

Hugh Dickins <hugh@veritas.com> wrote:
>>> We have no  real way of establishing the ownership of shared pages
>>> anyways. Its counted when allocated. But the page may live on afterwards
>>> in another process and then not be accounted for although its only user is
>>> the new process.

On Fri, Dec 10, 2004 at 01:38:59PM -0800, Andrew Morton wrote:
> We did lose some accounting accuracy when the pagetable walk and the big
> tasklist walks were removed.  Bill would probably have more details.  Given
> that the code as it stood was a complete showstopper, the tradeoff seemed
> reasonable.

There are several issues, not listed in order of importance here:
(1) Workload monitoring with high multiprogramming levels was infeasible.
(2) The long address space walks interfered with mmap() and page
	faults in the monitored processes, disturbing cluster membership
	and exceeding maximum response times in monitored workloads.
(3) There's a general long-running ongoing effort to take on various
	places tasklist_lock is abused one-by-one to incrementally
	resolve or otherwise mitigate the rwlock starvation issues.


-- wli

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

On Fri, 10 Dec 2004, Andrew Morton wrote:
> Hugh Dickins <hugh@veritas.com> wrote:
> > 
> > My inclination would be simply to remove the mark_page_accessed
> > from do_anonymous_page; but I have no numbers to back that hunch.
> 
> With the current implementation of page_referenced() the
> software-referenced bit doesn't matter anyway, as long as the pte's
> referenced bit got set.  So as long as the thing is on the active list, we
> can simply remove the mark_page_accessed() call.

Yes, you're right.  So we don't need numbers, can just delete that line.

> Except one day the VM might get smarter about pages which are both
> software-referenced and pte-referenced.

And on that day, we'd be making other changes, which might well
involve restoring the mark_page_accessed to do_anonymous_page
and adding it in the similar places which currently lack it.

But for now...

--- 2.6.10-rc3/mm/memory.c	2004-12-05 12:56:12.000000000 +0000
+++ linux/mm/memory.c	2004-12-11 09:18:39.000000000 +0000
@@ -1464,7 +1464,6 @@ do_anonymous_page(struct mm_struct *mm, 
 							 vma->vm_page_prot)),
 				      vma);
 		lru_cache_add_active(page);
-		mark_page_accessed(page);
 		page_add_anon_rmap(page, vma, addr);
 	}
 

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Christoph Lameter wrote:
> Thank you for the thorough review of my patches. Comments below
> 
> On Thu, 9 Dec 2004, Hugh Dickins wrote:
> 
> 
>>Your V12 patches would apply well to 2.6.10-rc3, except that (as noted
>>before) your mailer or whatever is eating trailing whitespace: trivial
>>patch attached to apply before yours, removing that whitespace so yours
>>apply.  But what your patches need to apply to would be 2.6.10-mm.
> 
> 
> I am still mystified as to why this is an issue at all. The patches apply
> just fine to the kernel sources as is. I have patched kernels numerous
> times with this patchset and never ran into any issue. quilt removes trailing
> whitespace from patches when they are generated as far as I can tell.
> 
> Patches will be made against mm after Nick's modifications to the 4 level
> patches are in.
> 

I've been a bit slow with them, sorry.... but there hasn't been a hard
decision to go one way or the other with the 4level patches yet.
Fortunately, it looks like 2.6.10 is having a longish drying out period,
so I should have something before it is released.

I would just sit on them for a while, and submit them to -mm when the
4level patches get merged / ready to merge into 2.6. It shouldn't slow
down the progress of your patch too much - they'll may have to wait until
after 2.6.11 anyway I'd say (probably depends on the progress of other
changes going in).


>>probably others (harder to think through).  Your 4/7 patch for i386 has
>>an unused atomic get_64bit function from Nick, I think you'll have to
>>define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.
> 
> 
> That would be a performance issue.
> 
> 

Problems were pretty trivial to reproduce here with non atomic 64-bit
loads being cut in half by atomic 64 bit stores. I don't see a way
around them, unfortunately.

Test case is to run with CONFIG_HIGHMEM (you needn't have > 4 GB of
memory in the system, of course), and run 2-4 threads on a dual CPU
system, doing parallel faulting of the *same* anonymous pages.

What happens is that the load (`entry = *pte`) in handle_pte_fault
gets cut in half, and handle_pte_fault drops down to do_swap_page,
and you get an infinite loop trying to read in a non existant swap
entry IIRC.

>>Hmm, that will only work if you're using atomic set_64bit rather than
>>relying on page_table_lock in the complementary places which matter.
>>Which I believe you are indeed doing in your 3level set_pte.  Shouldn't
>>__set_64bit be using LOCK_PREFIX like __get_64bit, instead of lock?
> 
> 
>>But by making every set_pte use set_64bit, you are significantly slowing
>>down many operations which do not need that atomicity.  This is quite
>>visible in the fork/exec/shell results from lmbench on i386 PAE (and is
>>the only interesting difference, for good or bad, that I noticed with
>>your patches in lmbench on 2*HT*P4), which run 5-20% slower.  There are
>>no faults on dst mm (nor on src mm) while copy_page_range is copying,
>>so its set_ptes don't need to be atomic; likewise during zap_pte_range
>>(either mmap_sem is held exclusively, or it's in the final exit_mmap).
>>Probably revert set_pte and set_pte_atomic to what they were, and use
>>set_pte_atomic where it's needed.
> 
> 
> Good suggestions. Will see what I can do but I will need some assistence
> my main platform is ia64 and the hardware and opportunities for testing on
> i386 are limited.
> 

I think I (and/or others) should be able to help with i386 if you are
having trouble :)

Nick

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Hugh Dickins]

On Sun, 12 Dec 2004, Nick Piggin wrote:
> Christoph Lameter wrote:
> > On Thu, 9 Dec 2004, Hugh Dickins wrote:
> 
> >>probably others (harder to think through).  Your 4/7 patch for i386 has
> >>an unused atomic get_64bit function from Nick, I think you'll have to
> >>define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.
> > 
> > That would be a performance issue.
> 
> Problems were pretty trivial to reproduce here with non atomic 64-bit
> loads being cut in half by atomic 64 bit stores. I don't see a way
> around them, unfortunately.

Of course, it'll only be a performance issue in the 64-on-32 cases:
the 64-on-64 and 32-on-32 macro should reduce to exactly the present
"entry = *pte".

I've had the impression that Christoph and SGI have to care a great
deal more about ia64 than the others; and as x86_64 advances, so
i386 PAE grows less important.  Just so long as a get_64bit there
isn't a serious degradation from present behaviour, it's okay.

Oh, hold on, isn't handle_mm_fault's pmd without page_table_lock
similarly racy, in both the 64-on-32 cases, and on architectures
which have a more complex pmd_t (sparc, m68k, h8300)?  Sigh.

Hugh

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Nick Piggin]

Hugh Dickins wrote:
> On Sun, 12 Dec 2004, Nick Piggin wrote:
> 
>>Christoph Lameter wrote:
>>
>>>On Thu, 9 Dec 2004, Hugh Dickins wrote:
>>
>>>>probably others (harder to think through).  Your 4/7 patch for i386 has
>>>>an unused atomic get_64bit function from Nick, I think you'll have to
>>>>define a get_pte_atomic macro and use get_64bit in its 64-on-32 cases.
>>>
>>>That would be a performance issue.
>>
>>Problems were pretty trivial to reproduce here with non atomic 64-bit
>>loads being cut in half by atomic 64 bit stores. I don't see a way
>>around them, unfortunately.
> 
> 
> Of course, it'll only be a performance issue in the 64-on-32 cases:
> the 64-on-64 and 32-on-32 macro should reduce to exactly the present
> "entry = *pte".
> 

That's right, yep. There is no ordering requirement, only that
the actual store and load be atomic.

> I've had the impression that Christoph and SGI have to care a great
> deal more about ia64 than the others; and as x86_64 advances, so
> i386 PAE grows less important.  Just so long as a get_64bit there
> isn't a serious degradation from present behaviour, it's okay.
> 

I don't think it was particularly serious for PAE. Probably not
worth holding off until 2.7. We'll see.

> Oh, hold on, isn't handle_mm_fault's pmd without page_table_lock
> similarly racy, in both the 64-on-32 cases, and on architectures
> which have a more complex pmd_t (sparc, m68k, h8300)?  Sigh.
> 

Can't comment on a specific architecture... some may have problems.
I think i386 prepopulates pmds, so it is no problem; but generally:

I think you can get away with it if you write the "unimportant"
word(s) first, do a wmb(), then write the word containing the
present bit. I guess this has to be done this way otherwise the
hardware walker will blow up...

Of course, the hardware walker would be doing either atomic or
correctly ordered reads, while a plain dereference doesn't
guarantee anything.

I'm not sure of the history behind the code, but I would be in
favour of making _all_ pagetable access go through accessor
functions, even if nobody quite needs them yet.

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[William Lee Irwin III]

On Sun, Dec 12, 2004 at 09:33:11AM +0000, Hugh Dickins wrote:
> Oh, hold on, isn't handle_mm_fault's pmd without page_table_lock
> similarly racy, in both the 64-on-32 cases, and on architectures
> which have a more complex pmd_t (sparc, m68k, h8300)?  Sigh.

yes.


-- wli

Re: Anticipatory prefaulting in the page fault handler V1

[Akinobu Mita]

On Friday 10 December 2004 04:32, Christoph Lameter wrote:
> On Wed, 8 Dec 2004, Martin J. Bligh wrote:
> > I tried benchmarking it ... but processes just segfault all the time.
> > Any chance you could try it out on SMP ia32 system?
>
> I tried it on my i386 system and it works fine. Sorry about the puny
> memory sizes (the system is a PIII-450 with 384k memory)
>

I also encountered processes segfault.
Below patch fix several problems.

1) if no pages could allocated, returns VM_FAULT_OOM
2) fix duplicated pte_offset_map() call
3) don't set_pte() for the entry which already have been set

Acutually, 3) fixes my segfault problem.

--- 2.6-rc/mm/memory.c.orig	2004-12-13 22:17:04.000000000 +0900
+++ 2.6-rc/mm/memory.c	2004-12-13 22:22:14.000000000 +0900
@@ -1483,6 +1483,8 @@ do_anonymous_page(struct mm_struct *mm, 
 				} else
 					break;
 			}
+			if (a == addr)
+				goto no_mem;
 			end_addr = a;
 
 			spin_lock(&mm->page_table_lock);
@@ -1514,8 +1516,17 @@ do_anonymous_page(struct mm_struct *mm, 
 			}
  		} else {
  			/* Read */
+			int first = 1;
+
  			for(;addr < end_addr; addr += PAGE_SIZE) {
-				page_table = pte_offset_map(pmd, addr);
+				if (!first)
+					page_table = pte_offset_map(pmd, addr);
+				first = 0;
+				if (!pte_none(*page_table)) {
+					/* Someone else got there first */
+					pte_unmap(page_table);
+					continue;
+				}
  				entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
 				set_pte(page_table, entry);
 				pte_unmap(page_table);

Re: Anticipatory prefaulting in the page fault handler V1

[Christoph Lameter]

On Mon, 13 Dec 2004, Akinobu Mita wrote:

> I also encountered processes segfault.
> Below patch fix several problems.
>
> 1) if no pages could allocated, returns VM_FAULT_OOM
> 2) fix duplicated pte_offset_map() call

I also saw these two issues and I think I dealt with them in a forthcoming
patch.

> 3) don't set_pte() for the entry which already have been set

Not sure how this could have happened in the patch.

Could you try my updated version:

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-12-08 15:01:48.801457702 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-12-08 15:02:04.286479345 -0800
@@ -537,6 +537,8 @@
 #endif

 	struct list_head tasks;
+	unsigned long anon_fault_next_addr;	/* Predicted sequential fault address */
+	int anon_fault_order;			/* Last order of allocation on fault */
 	/*
 	 * ptrace_list/ptrace_children forms the list of my children
 	 * that were stolen by a ptracer.
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-08 15:01:50.668339751 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-09 14:21:17.090061608 -0800
@@ -55,6 +55,7 @@

 #include <linux/swapops.h>
 #include <linux/elf.h>
+#include <linux/pagevec.h>

 #ifndef CONFIG_DISCONTIGMEM
 /* use the per-pgdat data instead for discontigmem - mbligh */
@@ -1432,52 +1433,99 @@
 		unsigned long addr)
 {
 	pte_t entry;
-	struct page * page = ZERO_PAGE(addr);
-
-	/* Read-only mapping of ZERO_PAGE. */
-	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+ 	unsigned long end_addr;
+
+	addr &= PAGE_MASK;
+
+ 	if (likely((vma->vm_flags & VM_RAND_READ) || current->anon_fault_next_addr != addr)) {
+		/* Single page */
+		current->anon_fault_order = 0;
+		end_addr = addr + PAGE_SIZE;
+	} else {
+		/* Sequence of faults detect. Perform preallocation */
+ 		int order = ++current->anon_fault_order;
+
+		if ((1 << order) < PAGEVEC_SIZE)
+			end_addr = addr + (PAGE_SIZE << order);
+		else
+			end_addr = addr + PAGEVEC_SIZE * PAGE_SIZE;

-	/* ..except if it's a write access */
+		if (end_addr > vma->vm_end)
+			end_addr = vma->vm_end;
+		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
+			end_addr &= PMD_MASK;
+	}
 	if (write_access) {
-		/* Allocate our own private page. */
+
+		unsigned long a;
+		struct page **p;
+		struct pagevec pv;
+
 		pte_unmap(page_table);
 		spin_unlock(&mm->page_table_lock);

+		pagevec_init(&pv, 0);
+
 		if (unlikely(anon_vma_prepare(vma)))
-			goto no_mem;
-		page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
-		if (!page)
-			goto no_mem;
-		clear_user_highpage(page, addr);
+			return VM_FAULT_OOM;
+
+		/* Allocate the necessary pages */
+		for(a = addr; a < end_addr ; a += PAGE_SIZE) {
+			struct page *p = alloc_page_vma(GFP_HIGHUSER, vma, a);
+
+			if (likely(p)) {
+				clear_user_highpage(p, a);
+				pagevec_add(&pv, p);
+			} else {
+				if (a == addr)
+					return VM_FAULT_OOM;
+				break;
+			}
+		}

 		spin_lock(&mm->page_table_lock);
-		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
+		for(p = pv.pages; addr < a; addr += PAGE_SIZE, p++) {
+
+			page_table = pte_offset_map(pmd, addr);
+			if (unlikely(!pte_none(*page_table))) {
+				/* Someone else got there first */
+				pte_unmap(page_table);
+				page_cache_release(*p);
+				continue;
+			}
+
+ 			entry = maybe_mkwrite(pte_mkdirty(mk_pte(*p,
+ 						 vma->vm_page_prot)),
+ 					      vma);
+
+			mm->rss++;
+			lru_cache_add_active(*p);
+			mark_page_accessed(*p);
+			page_add_anon_rmap(*p, vma, addr);
+
+			set_pte(page_table, entry);
 			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
+
+ 			/* No need to invalidate - it was non-present before */
+ 			update_mmu_cache(vma, addr, entry);
+		}
+ 	} else {
+ 		/* Read */
+		entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+nextread:
+		set_pte(page_table, entry);
+		pte_unmap(page_table);
+		update_mmu_cache(vma, addr, entry);
+		addr += PAGE_SIZE;
+		if (unlikely(addr < end_addr)) {
+			pte_offset_map(pmd, addr);
+			goto nextread;
 		}
-		mm->rss++;
-		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-							 vma->vm_page_prot)),
-				      vma);
-		lru_cache_add_active(page);
-		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}
-
-	set_pte(page_table, entry);
-	pte_unmap(page_table);
-
-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
+	current->anon_fault_next_addr = addr;
 	spin_unlock(&mm->page_table_lock);
-out:
 	return VM_FAULT_MINOR;
-no_mem:
-	return VM_FAULT_OOM;
 }

 /*

Re: Anticipatory prefaulting in the page fault handler V1

[Martin J. Bligh]

>> I also encountered processes segfault.
>> Below patch fix several problems.
>> 
>> 1) if no pages could allocated, returns VM_FAULT_OOM
>> 2) fix duplicated pte_offset_map() call
> 
> I also saw these two issues and I think I dealt with them in a forthcoming
> patch.
> 
>> 3) don't set_pte() for the entry which already have been set
> 
> Not sure how this could have happened in the patch.
> 
> Could you try my updated version:

Urgle. There was a fix from Hugh too ... any chance you could just stick
a whole new patch somewhere? I'm too idle/stupid to work it out ;-)

M.

Anticipatory prefaulting in the page fault handler V2

[Christoph Lameter]

Changes from V1 to V2:
- Eliminate duplicate code and reorganize things
- Use SetReferenced instead of mark_accessed (Hugh Dickins)
- Fix the problem of the preallocation order increasing out of bounds
(leading to memory being overwritten with pointers to struct page)
- Return VM_FAULT_OOM if not able to allocate a single page
- Tested on i386 and ia64
- New performance test for low cpu counts (up to 8 so that this does not
seem to be too exotic)

The page fault handler for anonymous pages can generate significant overhead
apart from its essential function which is to clear and setup a new page
table entry for a never accessed memory location. This overhead increases
significantly in an SMP environment.

In the page table scalability patches, we addressed the issue by changing
the locking scheme so that multiple fault handlers are able to be processed
concurrently on multiple cpus. This patch attempts to aggregate multiple
page faults into a single one. It does that by noting
anonymous page faults generated in sequence by an application.

If a fault occurred for page x and is then followed by page x+1 then it may
be reasonable to expect another page fault at x+2 in the future. If page
table entries for x+1 and x+2 would be prepared in the fault handling for
page x+1 then the overhead of taking a fault for x+2 is avoided. However
page x+2 may never be used and thus we may have increased the rss
of an application unnecessarily. The swapper will take care of removing
that page if memory should get tight.

The following patch makes the anonymous fault handler anticipate future
faults. For each fault a prediction is made where the fault would occur
(assuming linear acccess by the application). If the prediction turns out to
be right (next fault is where expected) then a number of pages is
preallocated in order to avoid a series of future faults. The order of the
preallocation increases by the power of two for each success in sequence.

The first successful prediction leads to an additional page being allocated.
Second successful prediction leads to 2 additional pages being allocated.
Third to 4 pages and so on. The max order is 3 by default. In a large
continous allocation the number of faults is reduced by a factor of 8.

Standard Kernel on a 8 Cpu machine allocating 1 and 4GB with an increasing
number of threads (and thus increasing parallellism of page faults):

ia64 2.6.10-rc3-bk7

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  1   3    1    0.047s      2.163s   2.021s 88925.153  88859.030
  1   3    2    0.040s      3.215s   1.069s 60385.889 115677.685
  1   3    4    0.041s      3.509s   1.023s 55370.338 158971.609
  1   3    8    0.047s      4.130s   1.014s 47049.904 172405.990

Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  4   3    1    0.155s     11.277s  11.043s 68788.420  68747.223
  4   3    2    0.161s     16.459s   8.061s 47315.277  91322.962
  4   3    4    0.170s     14.708s   4.079s 52852.007 164043.773
  4   3    8    0.171s     23.257s   4.028s 33565.604 183348.574

ia64 Patched kernel:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  1   3    1    0.008s      2.080s   2.008s 94121.792  94101.359
  1   3    2    0.015s      3.128s   1.064s 62523.771 119563.496
  1   3    4    0.008s      2.714s   1.012s 72185.910 175020.971
  1   3    8    0.016s      2.963s   0.087s 65965.457 223921.949

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  4   3    1    0.034s     10.861s  10.089s 72179.444  72181.353
  4   3    2    0.050s     14.303s   7.072s 54786.447 101738.901
  4   3    4    0.038s     13.478s   4.044s 58182.649 176913.840
  4   3    8    0.063s     13.584s   3.007s 57620.638 256109.927

i386 2.6.10-rc3-bk3 256M allocation 2x Pentium III 500 Mhz

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  0   3    1    0.020s      1.566s   1.058s123827.513 123842.098
  0   3    2    0.017s      2.439s   1.043s 79999.154 136931.671

i386 2.6.10-rc3-bk3 patches

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  0   3    1    0.020s      1.527s   1.039s126945.181 140930.664
  0   3    2    0.016s      2.417s   1.026s 80754.809 155162.903

Patch against 2.6.10-rc3-bk7:

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-12-13 15:14:40.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-12-13 15:15:55.000000000 -0800
@@ -537,6 +537,8 @@
 #endif

 	struct list_head tasks;
+	unsigned long anon_fault_next_addr;	/* Predicted sequential fault address */
+	int anon_fault_order;			/* Last order of allocation on fault */
 	/*
 	 * ptrace_list/ptrace_children forms the list of my children
 	 * that were stolen by a ptracer.
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-13 15:14:40.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-13 16:49:31.000000000 -0800
@@ -55,6 +55,7 @@

 #include <linux/swapops.h>
 #include <linux/elf.h>
+#include <linux/pagevec.h>

 #ifndef CONFIG_DISCONTIGMEM
 /* use the per-pgdat data instead for discontigmem - mbligh */
@@ -1432,52 +1433,102 @@
 		unsigned long addr)
 {
 	pte_t entry;
-	struct page * page = ZERO_PAGE(addr);
-
-	/* Read-only mapping of ZERO_PAGE. */
-	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+ 	unsigned long end_addr;
+
+	addr &= PAGE_MASK;
+
+ 	if (likely((vma->vm_flags & VM_RAND_READ) || current->anon_fault_next_addr != addr)) {
+		/* Single page */
+		current->anon_fault_order = 0;
+		end_addr = addr + PAGE_SIZE;
+	} else {
+		/* Sequence of faults detect. Perform preallocation */
+ 		int order = ++current->anon_fault_order;
+
+		if ((1 << order) < PAGEVEC_SIZE)
+			end_addr = addr + (PAGE_SIZE << order);
+		else {
+			end_addr = addr + PAGEVEC_SIZE * PAGE_SIZE;
+			current->anon_fault_order = 3;
+		}

-	/* ..except if it's a write access */
+		if (end_addr > vma->vm_end)
+			end_addr = vma->vm_end;
+		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
+			end_addr &= PMD_MASK;
+	}
 	if (write_access) {
-		/* Allocate our own private page. */
+
+		unsigned long a;
+		int i;
+		struct pagevec pv;
+
 		pte_unmap(page_table);
 		spin_unlock(&mm->page_table_lock);

+		pagevec_init(&pv, 0);
+
 		if (unlikely(anon_vma_prepare(vma)))
-			goto no_mem;
-		page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
-		if (!page)
-			goto no_mem;
-		clear_user_highpage(page, addr);
+			return VM_FAULT_OOM;
+
+		/* Allocate the necessary pages */
+		for(a = addr; a < end_addr ; a += PAGE_SIZE) {
+			struct page *p = alloc_page_vma(GFP_HIGHUSER, vma, a);
+
+			if (likely(p)) {
+				clear_user_highpage(p, a);
+				pagevec_add(&pv, p);
+			} else {
+				if (a == addr)
+					return VM_FAULT_OOM;
+				break;
+			}
+		}

 		spin_lock(&mm->page_table_lock);
-		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
+		for(i = 0; addr < a; addr += PAGE_SIZE, i++) {
+			struct page *p = pv.pages[i];
+
+			page_table = pte_offset_map(pmd, addr);
+			if (unlikely(!pte_none(*page_table))) {
+				/* Someone else got there first */
+				pte_unmap(page_table);
+				page_cache_release(p);
+				continue;
+			}
+
+ 			entry = maybe_mkwrite(pte_mkdirty(mk_pte(p,
+ 						 vma->vm_page_prot)),
+ 					      vma);
+
+			mm->rss++;
+			lru_cache_add_active(p);
+			SetPageReferenced(p);
+			page_add_anon_rmap(p, vma, addr);
+
+			set_pte(page_table, entry);
 			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
+
+ 			/* No need to invalidate - it was non-present before */
+ 			update_mmu_cache(vma, addr, entry);
+		}
+ 	} else {
+ 		/* Read */
+		entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+nextread:
+		set_pte(page_table, entry);
+		pte_unmap(page_table);
+		update_mmu_cache(vma, addr, entry);
+		addr += PAGE_SIZE;
+		if (unlikely(addr < end_addr)) {
+			pte_offset_map(pmd, addr);
+			goto nextread;
 		}
-		mm->rss++;
-		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-							 vma->vm_page_prot)),
-				      vma);
-		lru_cache_add_active(page);
-		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}
-
-	set_pte(page_table, entry);
-	pte_unmap(page_table);
-
-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
+	current->anon_fault_next_addr = addr;
 	spin_unlock(&mm->page_table_lock);
-out:
 	return VM_FAULT_MINOR;
-no_mem:
-	return VM_FAULT_OOM;
 }

 /*

Re: Anticipatory prefaulting in the page fault handler V1

[Akinobu Mita]

On Tuesday 14 December 2004 02:10, Christoph Lameter wrote:
> On Mon, 13 Dec 2004, Akinobu Mita wrote:

> > 3) don't set_pte() for the entry which already have been set
>
> Not sure how this could have happened in the patch.

This is why I inserted pte_none() for each page_table in case of
read fault too.

If read access fault occured for the address "addr".
It is completely unnecessary to check by pte_none() to the page_table
for "addr". Because page_table_lock has never been released until
do_anonymous_page returns (in case of read access fault)

But there is not any guarantee that the page_tables for addr+PAGE_SIZE,
addr+2*PAGE_SIZE, ...  have not been mapped yet.

Anyway, I will try your V2 patch.

Re: Anticipatory prefaulting in the page fault handler V1

[Akinobu Mita]

On Tuesday 14 December 2004 21:24, Akinobu Mita wrote:

> But there is not any guarantee that the page_tables for addr+PAGE_SIZE,
> addr+2*PAGE_SIZE, ...  have not been mapped yet.
>
> Anyway, I will try your V2 patch.
>

Below patch fixes V2 patch, and adds debug printk. 
The output coincides with segfaulted processes.

# dmesg | grep ^comm:

comm: xscreensaver, addr_orig: ccdc40, addr: cce000, pid: 2995
comm: rhn-applet-gui, addr_orig: b6fd8020, addr: b6fd9000, pid: 3029
comm: rhn-applet-gui, addr_orig: b6e95020, addr: b6e96000, pid: 3029
comm: rhn-applet-gui, addr_orig: b6fd8020, addr: b6fd9000, pid: 3029
comm: rhn-applet-gui, addr_orig: b6e95020, addr: b6e96000, pid: 3029
comm: rhn-applet-gui, addr_orig: b6fd8020, addr: b6fd9000, pid: 3029
comm: X, addr_orig: 87e8000, addr: 87e9000, pid: 2874
comm: X, addr_orig: 87ea000, addr: 87eb000, pid: 2874

---
The read access prefaulting may override the page_table which has been
already mapped. this patch fixes it. and it shows which process might
suffer this problem.


--- 2.6-rc/mm/memory.c.orig	2004-12-14 22:06:08.000000000 +0900
+++ 2.6-rc/mm/memory.c	2004-12-14 23:42:34.000000000 +0900
@@ -1434,6 +1434,7 @@ do_anonymous_page(struct mm_struct *mm, 
 {
 	pte_t entry;
  	unsigned long end_addr;
+ 	unsigned long addr_orig = addr;
 
 	addr &= PAGE_MASK;
 
@@ -1517,9 +1518,15 @@ do_anonymous_page(struct mm_struct *mm, 
  		/* Read */
 		entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
 nextread:
-		set_pte(page_table, entry);
-		pte_unmap(page_table);
-		update_mmu_cache(vma, addr, entry);
+		if (!pte_none(*page_table)) {
+			printk("comm: %s, addr_orig: %lx, addr: %lx, pid: %d\n",
+				current->comm, addr_orig, addr, current->pid);
+			pte_unmap(page_table);
+		} else {
+			set_pte(page_table, entry);
+			pte_unmap(page_table);
+			update_mmu_cache(vma, addr, entry);
+		}
 		addr += PAGE_SIZE;
 		if (unlikely(addr < end_addr)) {
 			pte_offset_map(pmd, addr);

Re: Anticipatory prefaulting in the page fault handler V1

[Adam Litke]

What benchmark are you using to generate the following results?  I'd
like to run this on some of my hardware and see how the results compare.

On Wed, 2004-12-08 at 11:24, Christoph Lameter wrote:
> Standard Kernel on a 512 Cpu machine allocating 32GB with an increasing
> number of threads (and thus increasing parallellism of page faults):
> 
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>  32   3    1    1.416s    138.165s 139.050s 45073.831  45097.498
>  32   3    2    1.397s    148.523s  78.044s 41965.149  80201.646
>  32   3    4    1.390s    152.618s  44.044s 40851.258 141545.239
>  32   3    8    1.500s    374.008s  53.001s 16754.519 118671.950
>  32   3   16    1.415s   1051.759s  73.094s  5973.803  85087.358
>  32   3   32    1.867s   3400.417s 117.003s  1849.186  53754.928
>  32   3   64    5.361s  11633.040s 197.034s   540.577  31881.112
>  32   3  128   23.387s  39386.390s 332.055s   159.642  18918.599
>  32   3  256   15.409s  20031.450s 168.095s   313.837  37237.918
>  32   3  512   18.720s  10338.511s  86.047s   607.446  72752.686
> 
> Patched kernel:
> 
> Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>  32   3    1    1.098s    138.544s 139.063s 45053.657  45057.920
>  32   3    2    1.022s    127.770s  67.086s 48849.350  92707.085
>  32   3    4    0.995s    119.666s  37.045s 52141.503 167955.292
>  32   3    8    0.928s     87.400s  18.034s 71227.407 342934.242
>  32   3   16    1.067s     72.943s  11.035s 85007.293 553989.377
>  32   3   32    1.248s    133.753s  10.038s 46602.680 606062.151
>  32   3   64    5.557s    438.634s  13.093s 14163.802 451418.617
>  32   3  128   17.860s   1496.797s  19.048s  4153.714 322808.509
>  32   3  256   13.382s    766.063s  10.016s  8071.695 618816.838
>  32   3  512   17.067s    369.106s   5.041s 16291.764 1161285.521
> 
> These number are roughly equal to what can be accomplished with the
> page fault scalability patches.
> 
> Kernel patches with both the page fault scalability patches and
> prefaulting:
> 
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>  32  10    1    4.103s    456.384s 460.046s 45541.992  45544.369
>  32  10    2    4.005s    415.119s 221.095s 50036.407  94484.174
>  32  10    4    3.855s    371.317s 111.076s 55898.259 187635.724
>  32  10    8    3.902s    308.673s  67.094s 67092.476 308634.397
>  32  10   16    4.011s    224.213s  37.016s 91889.781 564241.062
>  32  10   32    5.483s    209.391s  27.046s 97598.647 763495.417
>  32  10   64   19.166s    219.925s  26.030s 87713.212 797286.395
>  32  10  128   53.482s    342.342s  27.024s 52981.744 769687.791
>  32  10  256   67.334s    180.321s  15.036s 84679.911 1364614.334
>  32  10  512   66.516s     93.098s   9.015s131387.893 2291548.865
> 
> The fault rate doubles when both patches are applied.
> 
> And on the high end (512 processors allocating 256G) (No numbers
> for regular kernels because they are extremely slow, also no
> number for a low number of threads. Also very slow)
> 
> With prefaulting:
> 
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
> 256   3    4    8.241s   1414.348s 449.016s 35380.301 112056.239
> 256   3    8    8.306s   1300.982s 247.025s 38441.977 203559.271
> 256   3   16    8.368s   1223.853s 154.089s 40846.272 324940.924
> 256   3   32    8.536s   1284.041s 110.097s 38938.970 453556.624
> 256   3   64   13.025s   3587.203s 110.010s 13980.123 457131.492
> 256   3  128   25.025s  11460.700s 145.071s  4382.104 345404.909
> 256   3  256   26.150s   6061.649s  75.086s  8267.625 663414.482
> 256   3  512   20.637s   3037.097s  38.062s 16460.435 1302993.019
> 
> Page fault scalability patch and prefaulting. Max prefault order
> increased to 5 (max preallocation of 32 pages):
> 
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
> 256  10    8   33.571s   4516.293s 863.021s 36874.099 194356.930
> 256  10   16   33.103s   3737.688s 461.028s 44492.553 363704.484
> 256  10   32   35.094s   3436.561s 321.080s 48326.262 521352.840
> 256  10   64   46.675s   2899.997s 245.020s 56936.124 684214.256
> 256  10  128   85.493s   2890.198s 203.008s 56380.890 826122.524
> 256  10  256   74.299s   1374.973s  99.088s115762.963 1679630.272
> 256  10  512   62.760s    706.559s  53.027s218078.311 3149273.714
> 
> We are getting into an almost linear scalability in the high end with
> both patches and end up with a fault rate > 3 mio faults per second.
> 
> The one thing that takes up a lot of time is still be the zeroing
> of pages in the page fault handler. There is a another
> set of patches that I am working on which will prezero pages
> and led to another an increase in performance by a factor of 2-4
> (if prezeroed pages are available which may not always be the case).
> Maybe we can reach 10 mio fault /sec that way.
> 
> Patch against 2.6.10-rc3-bk3:
> 
> Index: linux-2.6.9/include/linux/sched.h
> ===================================================================
> --- linux-2.6.9.orig/include/linux/sched.h	2004-12-01 10:37:31.000000000 -0800
> +++ linux-2.6.9/include/linux/sched.h	2004-12-01 10:38:15.000000000 -0800
> @@ -537,6 +537,8 @@
>  #endif
> 
>  	struct list_head tasks;
> +	unsigned long anon_fault_next_addr;	/* Predicted sequential fault address */
> +	int anon_fault_order;			/* Last order of allocation on fault */
>  	/*
>  	 * ptrace_list/ptrace_children forms the list of my children
>  	 * that were stolen by a ptracer.
> Index: linux-2.6.9/mm/memory.c
> ===================================================================
> --- linux-2.6.9.orig/mm/memory.c	2004-12-01 10:38:11.000000000 -0800
> +++ linux-2.6.9/mm/memory.c	2004-12-01 10:45:01.000000000 -0800
> @@ -55,6 +55,7 @@
> 
>  #include <linux/swapops.h>
>  #include <linux/elf.h>
> +#include <linux/pagevec.h>
> 
>  #ifndef CONFIG_DISCONTIGMEM
>  /* use the per-pgdat data instead for discontigmem - mbligh */
> @@ -1432,8 +1433,106 @@
>  		unsigned long addr)
>  {
>  	pte_t entry;
> -	struct page * page = ZERO_PAGE(addr);
> +	struct page * page;
> +
> +	addr &= PAGE_MASK;
> +
> + 	if (current->anon_fault_next_addr == addr) {
> + 		unsigned long end_addr;
> + 		int order = current->anon_fault_order;
> +
> +		/* Sequence of page faults detected. Perform preallocation of pages */
> 
> +		/* The order of preallocations increases with each successful prediction */
> + 		order++;
> +
> +		if ((1 << order) < PAGEVEC_SIZE)
> +			end_addr = addr + (1 << (order + PAGE_SHIFT));
> +		else
> +			end_addr = addr + PAGEVEC_SIZE * PAGE_SIZE;
> +
> +		if (end_addr > vma->vm_end)
> +			end_addr = vma->vm_end;
> +		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
> +			end_addr &= PMD_MASK;
> +
> +		current->anon_fault_next_addr = end_addr;
> +	 	current->anon_fault_order = order;
> +
> +		if (write_access) {
> +
> +			struct pagevec pv;
> +			unsigned long a;
> +			struct page **p;
> +
> +			pte_unmap(page_table);
> +			spin_unlock(&mm->page_table_lock);
> +
> +			pagevec_init(&pv, 0);
> +
> +			if (unlikely(anon_vma_prepare(vma)))
> +				return VM_FAULT_OOM;
> +
> +			/* Allocate the necessary pages */
> +			for(a = addr;a < end_addr ; a += PAGE_SIZE) {
> +				struct page *p = alloc_page_vma(GFP_HIGHUSER, vma, a);
> +
> +				if (p) {
> +					clear_user_highpage(p, a);
> +					pagevec_add(&pv,p);
> +				} else
> +					break;
> +			}
> +			end_addr = a;
> +
> +			spin_lock(&mm->page_table_lock);
> +
> + 			for(p = pv.pages; addr < end_addr; addr += PAGE_SIZE, p++) {
> +
> +				page_table = pte_offset_map(pmd, addr);
> +				if (!pte_none(*page_table)) {
> +					/* Someone else got there first */
> +					page_cache_release(*p);
> +					pte_unmap(page_table);
> +					continue;
> +				}
> +
> + 				entry = maybe_mkwrite(pte_mkdirty(mk_pte(*p,
> + 							 vma->vm_page_prot)),
> + 						      vma);
> +
> +				mm->rss++;
> +				lru_cache_add_active(*p);
> +				mark_page_accessed(*p);
> +				page_add_anon_rmap(*p, vma, addr);
> +
> +				set_pte(page_table, entry);
> +				pte_unmap(page_table);
> +
> + 				/* No need to invalidate - it was non-present before */
> + 				update_mmu_cache(vma, addr, entry);
> +			}
> + 		} else {
> + 			/* Read */
> + 			for(;addr < end_addr; addr += PAGE_SIZE) {
> +				page_table = pte_offset_map(pmd, addr);
> + 				entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
> +				set_pte(page_table, entry);
> +				pte_unmap(page_table);
> +
> + 				/* No need to invalidate - it was non-present before */
> +				update_mmu_cache(vma, addr, entry);
> +
> +			};
> +		}
> +		spin_unlock(&mm->page_table_lock);
> +		return VM_FAULT_MINOR;
> +	}
> +
> +	current->anon_fault_next_addr = addr + PAGE_SIZE;
> +	current->anon_fault_order = 0;
> +
> +	page = ZERO_PAGE(addr);
>  	/* Read-only mapping of ZERO_PAGE. */
>  	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
> 
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
-- 
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center

Re: Anticipatory prefaulting in the page fault handler V2

[Adam Litke]

Just to add another data point:  This works on my 4-way ppc64 (Power4)
box.  I am seeing no degradation when running this on kernbench (which
is expected).  For the curious, here are the results:

Kernbench results with anon-prefault:
349.86user 49.64system 1:57.85elapsed 338%CPU (0avgtext+0avgdata 0maxresident)k
349.65user 49.81system 1:58.31elapsed 337%CPU (0avgtext+0avgdata 0maxresident)k
349.48user 50.00system 1:53.70elapsed 351%CPU (0avgtext+0avgdata 0maxresident)k
349.73user 49.69system 1:57.67elapsed 339%CPU (0avgtext+0avgdata 0maxresident)k
349.75user 49.85system 1:52.71elapsed 354%CPU (0avgtext+0avgdata 0maxresident)k
Elapsed: 116.048s User: 349.694s System: 49.798s CPU: 343.8%

Kernbench results without anon-prefault:
350.86user 52.54system 1:53.45elapsed 355%CPU (0avgtext+0avgdata 0maxresident)k
350.99user 52.36system 1:52.05elapsed 359%CPU (0avgtext+0avgdata 0maxresident)k
350.92user 52.68system 1:54.14elapsed 353%CPU (0avgtext+0avgdata 0maxresident)k
350.98user 52.38system 1:56.17elapsed 347%CPU (0avgtext+0avgdata 0maxresident)k
351.16user 52.31system 1:53.90elapsed 354%CPU (0avgtext+0avgdata 0maxresident)k
Elapsed: 113.942s User: 350.982s System: 52.454s CPU: 353.6%

On Mon, 2004-12-13 at 19:32, Christoph Lameter wrote:
> Changes from V1 to V2:
> - Eliminate duplicate code and reorganize things
> - Use SetReferenced instead of mark_accessed (Hugh Dickins)
> - Fix the problem of the preallocation order increasing out of bounds
> (leading to memory being overwritten with pointers to struct page)
> - Return VM_FAULT_OOM if not able to allocate a single page
> - Tested on i386 and ia64
> - New performance test for low cpu counts (up to 8 so that this does not
> seem to be too exotic)
-- 
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center

Re: Anticipatory prefaulting in the page fault handler V1

[Christoph Lameter]

On Tue, 14 Dec 2004, Akinobu Mita wrote:

> This is why I inserted pte_none() for each page_table in case of
> read fault too.
>
> If read access fault occured for the address "addr".
> It is completely unnecessary to check by pte_none() to the page_table
> for "addr". Because page_table_lock has never been released until
> do_anonymous_page returns (in case of read access fault)
>
> But there is not any guarantee that the page_tables for addr+PAGE_SIZE,
> addr+2*PAGE_SIZE, ...  have not been mapped yet.

Right. Thanks for pointing that out.

Anticipatory prefaulting in the page fault handler V3

[Christoph Lameter]

Changes from V2 to V3:
- check for empty pte before setting additional pte's in aggregate
  read

The page fault handler for anonymous pages can generate significant overhead
apart from its essential function which is to clear and setup a new page
table entry for a never accessed memory location. This overhead increases
significantly in an SMP environment.

In the page table scalability patches, we addressed the issue by changing
the locking scheme so that multiple fault handlers are able to be processed
concurrently on multiple cpus. This patch attempts to aggregate multiple
page faults into a single one. It does that by noting
anonymous page faults generated in sequence by an application.

If a fault occurred for page x and is then followed by page x+1 then it may
be reasonable to expect another page fault at x+2 in the future. If page
table entries for x+1 and x+2 would be prepared in the fault handling for
page x+1 then the overhead of taking a fault for x+2 is avoided. However
page x+2 may never be used and thus we may have increased the rss
of an application unnecessarily. The swapper will take care of removing
that page if memory should get tight.

The following patch makes the anonymous fault handler anticipate future
faults. For each fault a prediction is made where the fault would occur
(assuming linear acccess by the application). If the prediction turns out to
be right (next fault is where expected) then a number of pages is
preallocated in order to avoid a series of future faults. The order of the
preallocation increases by the power of two for each success in sequence.

The first successful prediction leads to an additional page being allocated.
Second successful prediction leads to 2 additional pages being allocated.
Third to 4 pages and so on. The max order is 3 by default. In a large
continous allocation the number of faults is reduced by a factor of 8.

Patch against 2.6.10-rc3-bk7:

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-12-13 15:14:40.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-12-14 12:21:26.000000000 -0800
@@ -537,6 +537,8 @@
 #endif

 	struct list_head tasks;
+	unsigned long anon_fault_next_addr;	/* Predicted sequential fault address */
+	int anon_fault_order;			/* Last order of allocation on fault */
 	/*
 	 * ptrace_list/ptrace_children forms the list of my children
 	 * that were stolen by a ptracer.
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-13 15:14:40.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-14 12:23:36.000000000 -0800
@@ -55,6 +55,7 @@

 #include <linux/swapops.h>
 #include <linux/elf.h>
+#include <linux/pagevec.h>

 #ifndef CONFIG_DISCONTIGMEM
 /* use the per-pgdat data instead for discontigmem - mbligh */
@@ -1432,52 +1433,103 @@
 		unsigned long addr)
 {
 	pte_t entry;
-	struct page * page = ZERO_PAGE(addr);
-
-	/* Read-only mapping of ZERO_PAGE. */
-	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+ 	unsigned long end_addr;
+
+	addr &= PAGE_MASK;
+
+ 	if (likely((vma->vm_flags & VM_RAND_READ) || current->anon_fault_next_addr != addr)) {
+		/* Single page */
+		current->anon_fault_order = 0;
+		end_addr = addr + PAGE_SIZE;
+	} else {
+		/* Sequence of faults detect. Perform preallocation */
+ 		int order = ++current->anon_fault_order;
+
+		if ((1 << order) < PAGEVEC_SIZE)
+			end_addr = addr + (PAGE_SIZE << order);
+		else {
+			end_addr = addr + PAGEVEC_SIZE * PAGE_SIZE;
+			current->anon_fault_order = 3;
+		}

-	/* ..except if it's a write access */
+		if (end_addr > vma->vm_end)
+			end_addr = vma->vm_end;
+		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
+			end_addr &= PMD_MASK;
+	}
 	if (write_access) {
-		/* Allocate our own private page. */
+
+		unsigned long a;
+		int i;
+		struct pagevec pv;
+
 		pte_unmap(page_table);
 		spin_unlock(&mm->page_table_lock);

+		pagevec_init(&pv, 0);
+
 		if (unlikely(anon_vma_prepare(vma)))
-			goto no_mem;
-		page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
-		if (!page)
-			goto no_mem;
-		clear_user_highpage(page, addr);
+			return VM_FAULT_OOM;
+
+		/* Allocate the necessary pages */
+		for(a = addr; a < end_addr ; a += PAGE_SIZE) {
+			struct page *p = alloc_page_vma(GFP_HIGHUSER, vma, a);
+
+			if (likely(p)) {
+				clear_user_highpage(p, a);
+				pagevec_add(&pv, p);
+			} else {
+				if (a == addr)
+					return VM_FAULT_OOM;
+				break;
+			}
+		}

 		spin_lock(&mm->page_table_lock);
-		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
+		for(i = 0; addr < a; addr += PAGE_SIZE, i++) {
+			struct page *p = pv.pages[i];
+
+			page_table = pte_offset_map(pmd, addr);
+			if (unlikely(!pte_none(*page_table))) {
+				/* Someone else got there first */
+				pte_unmap(page_table);
+				page_cache_release(p);
+				continue;
+			}
+
+ 			entry = maybe_mkwrite(pte_mkdirty(mk_pte(p,
+ 						 vma->vm_page_prot)),
+ 					      vma);
+
+			mm->rss++;
+			lru_cache_add_active(p);
+			SetPageReferenced(p);
+			page_add_anon_rmap(p, vma, addr);
+
+			set_pte(page_table, entry);
 			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
+
+ 			/* No need to invalidate - it was non-present before */
+ 			update_mmu_cache(vma, addr, entry);
+		}
+ 	} else {
+ 		/* Read */
+		entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+nextread:
+		set_pte(page_table, entry);
+		pte_unmap(page_table);
+		update_mmu_cache(vma, addr, entry);
+		addr += PAGE_SIZE;
+		if (unlikely(addr < end_addr)) {
+			page_table = pte_offset_map(pmd, addr);
+			if (likely(pte_none(*page_table)))
+				goto nextread;
 		}
-		mm->rss++;
-		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-							 vma->vm_page_prot)),
-				      vma);
-		lru_cache_add_active(page);
-		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}
-
-	set_pte(page_table, entry);
-	pte_unmap(page_table);
-
-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
+	current->anon_fault_next_addr = addr;
 	spin_unlock(&mm->page_table_lock);
-out:
 	return VM_FAULT_MINOR;
-no_mem:
-	return VM_FAULT_OOM;
 }

 /*

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Christoph Lameter]

On Sun, 12 Dec 2004, William Lee Irwin III wrote:

> On Sun, Dec 12, 2004 at 09:33:11AM +0000, Hugh Dickins wrote:
> > Oh, hold on, isn't handle_mm_fault's pmd without page_table_lock
> > similarly racy, in both the 64-on-32 cases, and on architectures
> > which have a more complex pmd_t (sparc, m68k, h8300)?  Sigh.
>
> yes.

Those may fall back to use the page_table_lock for individual operations
that cannot be realized in an atomic way.

page fault scalability patch V13 [0/8]: Overview

[Christoph Lameter]

Changes from V12->V13 of this patch:
- list_rss: Fix issues Hugh dickins pointed out.
- i386: Hugh Dickins patch. Fall back to not use get_64 if ptl is used to
  restore performance on PAE..
- introduce get_pte_atomic for non ptl access to pte.
- i386 PAE: get_pte_atomic uses get_64bit
- ptep cmpxchg must now include update_mmu_cache functionality. All
  arches updated.
- add optional prefault patch [8/8] which is controllable via
  /proc/sys/vm/max_pre_alloc
- This patch series tested by me on i386 PAE and ia64

Potential issues:
- I avoided to check mm == current->mm for incrementing current->rss in
  the anonymous fault handler. Instead I added some code
  around handle_mm_fault in memory.c to deal with the situation
  if mm != current->mm.

Potential unintended benefits from mm containing a tasklist:
- Threads are now freed via rcu thus the tasklist may be traversed
  without acquiring the tasklist lock if the other list operations
  for task_struct are also switched to be of rcu type.
- The list of threads of a mm is determined in two locations by a
  walk of all the tasks. The code could be change to use the
  tasklist in mm.

This is a series of patches that increases the scalability of
the page fault handler for SMP. The performance increase is
accomplished by avoiding the use of the
page_table_lock spinlock (but not mm->mmap_sem) through new atomic
operations on pte's (ptep_xchg, ptep_cmpxchg) and on pmd and pgd's
(pgd_test_and_populate, pmd_test_and_populate).

The page table lock can be avoided in the following situations:

1. An empty pte or pmd entry is populated

This is safe since the swapper may only depopulate them and the
swapper code has been changed to never set a pte to be empty until the
page has been evicted. The population of an empty pte is frequent
if a process touches newly allocated memory.

2. Modifications of flags in a pte entry (write/accessed).

These modifications are done by the CPU or by low level handlers
on various platforms also bypassing the page_table_lock. So this
seems to be safe too.

One essential change in the VM is the use of pte_cmpxchg (or its
generic emulation) on page table entries before doing an
update_mmu_change without holding the page table lock. However, we do
similar things now with other atomic pte operations such as
ptep_get_and_clear and ptep_test_and_clear_dirty. These operations
clear a pte *after* doing an operation on it. The ptep_cmpxchg as used
in this patch operates on an *cleared* pte and replaces it with a pte
pointing to valid memory. The effect of this change on various
architectures has to be thought through. Local definitions of
ptep_cmpxchg and ptep_xchg may be necessary.

For IA64 an icache coherency issue may arise that potentially requires
the flushing of the icache (as done via update_mmu_cache on IA64) prior
to the use of ptep_cmpxchg. Similar issues may arise on other platforms.

The patch introduces a split counter for rss handling to avoid atomic
operations and locks currently necessary for rss modifications. In
addition to mm->rss, tsk->rss is introduced. tsk->rss is defined to be
in the same cache line as tsk->mm (which is already used by the fault
handler) and thus tsk->rss can be incremented without locks
in a fast way. The cache line does not need to be shared between
processors for the page table handler.

A tasklist is generated for each mm (rcu based). Values in that list
are added up to calculate rss or anon_rss values.

The patchset is composed of 8 patches (and was tested against 2.6.10-rc3-bk10):

1/8: Avoid page_table_lock in handle_mm_fault

   This patch defers the acquisition of the page_table_lock as much as
   possible and uses atomic operations for allocating anonymous memory.
   These atomic operations are simulated by acquiring the page_table_lock
   for very small time frames if an architecture does not define
   __HAVE_ARCH_ATOMIC_TABLE_OPS. It also changes the swapper so that a
   pte will not be set to empty if a page is in transition to swap.

   If only the first two patches are applied then the time that the
   page_table_lock is held is simply reduced. The lock may then be
   acquired multiple times during a page fault.

2/8: Atomic pte operations for ia64

3/8: Make cmpxchg generally available on i386

   The atomic operations on the page table rely heavily on cmpxchg
   instructions. This patch adds emulations for cmpxchg and cmpxchg8b
   for old 80386 and 80486 cpus. The emulations are only included if a
   kernel is build for these old cpus and are skipped for the real
   cmpxchg instructions if the kernel that is build for 386 or 486 is
   then run on a more recent cpu.

   This patch may be used independently of the other patches.

4/8: Atomic pte operations for i386

   A generally available cmpxchg (last patch) must be available for
   this patch to preserve the ability to build kernels for 386 and 486.

5/8: Atomic pte operation for x86_64

6/8: Atomic pte operations for s390

7/8: Split counter implementation for rss
  Add tsk->rss and tsk->anon_rss. Add tasklist. Add logic
  to calculate rss from tasklist.

8/8: Prefaulting for the page table scalability patchset.

  Note that this patch is significantly different from the patches
  posted under the title "Anticipatory prefaulting" because the handling
  of the pte's differs significantly. This prefault patch can reach
  higher orders during prefaulting since no pagevec is needed to store the
  preallocated pages.

  The maximum order of preallocation can be controlled via
  /proc/sys/vm/max_prealloc_order and is set to 3 by default. Setting
  max_prealloc_order to zero switches off preallocation altogether.

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Re: page fault scalability patch V13 [1/8]: Reduce the use of the page_table_lock

[Christoph Lameter]

Changelog
        * Increase parallelism in SMP configurations by deferring
          the acquisition of page_table_lock in handle_mm_fault
        * Anonymous memory page faults bypass the page_table_lock
          through the use of atomic page table operations
        * Swapper does not set pte to empty in transition to swap
        * Simulate atomic page table operations using the
          page_table_lock if an arch does not define
          __HAVE_ARCH_ATOMIC_TABLE_OPS. This still provides
          a performance benefit since the page_table_lock
          is held for shorter periods of time.

Signed-off-by: Christoph Lameter <clameter@sgi.com

Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-15 15:00:22.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-16 10:03:27.000000000 -0800
@@ -1330,8 +1330,7 @@
 }

 /*
- * We hold the mm semaphore and the page_table_lock on entry and
- * should release the pagetable lock on exit..
+ * We hold the mm semaphore
  */
 static int do_swap_page(struct mm_struct * mm,
 	struct vm_area_struct * vma, unsigned long address,
@@ -1343,15 +1342,13 @@
 	int ret = VM_FAULT_MINOR;

 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);
 	page = lookup_swap_cache(entry);
 	if (!page) {
  		swapin_readahead(entry, address, vma);
  		page = read_swap_cache_async(entry, vma, address);
 		if (!page) {
 			/*
-			 * Back out if somebody else faulted in this pte while
-			 * we released the page table lock.
+			 * Back out if somebody else faulted in this pte
 			 */
 			spin_lock(&mm->page_table_lock);
 			page_table = pte_offset_map(pmd, address);
@@ -1374,8 +1371,7 @@
 	lock_page(page);

 	/*
-	 * Back out if somebody else faulted in this pte while we
-	 * released the page table lock.
+	 * Back out if somebody else faulted in this pte
 	 */
 	spin_lock(&mm->page_table_lock);
 	page_table = pte_offset_map(pmd, address);
@@ -1422,14 +1418,12 @@
 }

 /*
- * We are called with the MM semaphore and page_table_lock
- * spinlock held to protect against concurrent faults in
- * multithreaded programs.
+ * We are called with the MM semaphore held.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 		pte_t *page_table, pmd_t *pmd, int write_access,
-		unsigned long addr)
+		unsigned long addr, pte_t orig_entry)
 {
 	pte_t entry;
 	struct page * page = ZERO_PAGE(addr);
@@ -1441,7 +1435,6 @@
 	if (write_access) {
 		/* Allocate our own private page. */
 		pte_unmap(page_table);
-		spin_unlock(&mm->page_table_lock);

 		if (unlikely(anon_vma_prepare(vma)))
 			goto no_mem;
@@ -1450,30 +1443,34 @@
 			goto no_mem;
 		clear_user_highpage(page, addr);

-		spin_lock(&mm->page_table_lock);
 		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
-			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
-		}
-		mm->rss++;
 		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
 							 vma->vm_page_prot)),
 				      vma);
-		lru_cache_add_active(page);
-		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}

-	set_pte(page_table, entry);
+	/* update the entry */
+	if (!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)) {
+		if (write_access) {
+			pte_unmap(page_table);
+			page_cache_release(page);
+		}
+		goto out;
+	}
+	if (write_access) {
+		/*
+		 * These two functions must come after the cmpxchg
+		 * because if the page is on the LRU then try_to_unmap may come
+		 * in and unmap the pte.
+		 */
+		page_add_anon_rmap(page, vma, addr);
+		lru_cache_add_active(page);
+		mm->rss++;
+
+	}
 	pte_unmap(page_table);

-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
-	spin_unlock(&mm->page_table_lock);
 out:
 	return VM_FAULT_MINOR;
 no_mem:
@@ -1489,12 +1486,12 @@
  * As this is called only for pages that do not currently exist, we
  * do not need to flush old virtual caches or the TLB.
  *
- * This is called with the MM semaphore held and the page table
- * spinlock held. Exit with the spinlock released.
+ * This is called with the MM semaphore held.
  */
 static int
 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-	unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *page_table,
+        pmd_t *pmd, pte_t orig_entry)
 {
 	struct page * new_page;
 	struct address_space *mapping = NULL;
@@ -1505,9 +1502,8 @@

 	if (!vma->vm_ops || !vma->vm_ops->nopage)
 		return do_anonymous_page(mm, vma, page_table,
-					pmd, write_access, address);
+					pmd, write_access, address, orig_entry);
 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);

 	if (vma->vm_file) {
 		mapping = vma->vm_file->f_mapping;
@@ -1605,7 +1601,7 @@
  * nonlinear vmas.
  */
 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
-	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd, pte_t entry)
 {
 	unsigned long pgoff;
 	int err;
@@ -1618,13 +1614,12 @@
 	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
-		return do_no_page(mm, vma, address, write_access, pte, pmd);
+		return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 	}

-	pgoff = pte_to_pgoff(*pte);
+	pgoff = pte_to_pgoff(entry);

 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);

 	err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
 	if (err == -ENOMEM)
@@ -1643,49 +1638,46 @@
  * with external mmu caches can use to update those (ie the Sparc or
  * PowerPC hashed page tables that act as extended TLBs).
  *
- * Note the "page_table_lock". It is to protect against kswapd removing
- * pages from under us. Note that kswapd only ever _removes_ pages, never
- * adds them. As such, once we have noticed that the page is not present,
- * we can drop the lock early.
- *
- * The adding of pages is protected by the MM semaphore (which we hold),
- * so we don't need to worry about a page being suddenly been added into
- * our VM.
- *
- * We enter with the pagetable spinlock held, we are supposed to
- * release it when done.
+ * Note that kswapd only ever _removes_ pages, never adds them.
+ * We need to insure to handle that case properly.
  */
 static inline int handle_pte_fault(struct mm_struct *mm,
 	struct vm_area_struct * vma, unsigned long address,
 	int write_access, pte_t *pte, pmd_t *pmd)
 {
 	pte_t entry;
+	pte_t new_entry;

-	entry = *pte;
+	/*
+	 * This must be a atomic operation since the page_table_lock is
+	 * not held. If a pte_t larger than the word size is used an
+	 * incorrect value could be read because another processor is
+	 * concurrently updating the multi-word pte. The i386 PAE mode
+	 * is raising its ugly head here.
+	 */
+	entry = get_pte_atomic(pte);
 	if (!pte_present(entry)) {
-		/*
-		 * If it truly wasn't present, we know that kswapd
-		 * and the PTE updates will not touch it later. So
-		 * drop the lock.
-		 */
 		if (pte_none(entry))
-			return do_no_page(mm, vma, address, write_access, pte, pmd);
+			return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 		if (pte_file(entry))
-			return do_file_page(mm, vma, address, write_access, pte, pmd);
+			return do_file_page(mm, vma, address, write_access, pte, pmd, entry);
 		return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
 	}

+	/*
+	 * This is the case in which we only update some bits in the pte.
+	 */
+	new_entry = pte_mkyoung(entry);
 	if (write_access) {
-		if (!pte_write(entry))
+		if (!pte_write(entry)) {
+			/* do_wp_page expects us to hold the page_table_lock */
+			spin_lock(&mm->page_table_lock);
 			return do_wp_page(mm, vma, address, pte, pmd, entry);
-
-		entry = pte_mkdirty(entry);
+		}
+		new_entry = pte_mkdirty(new_entry);
 	}
-	entry = pte_mkyoung(entry);
-	ptep_set_access_flags(vma, address, pte, entry, write_access);
-	update_mmu_cache(vma, address, entry);
+	ptep_cmpxchg(vma, address, pte, entry, new_entry);
 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_MINOR;
 }

@@ -1703,22 +1695,45 @@

 	inc_page_state(pgfault);

-	if (is_vm_hugetlb_page(vma))
+	if (unlikely(is_vm_hugetlb_page(vma)))
 		return VM_FAULT_SIGBUS;	/* mapping truncation does this. */

 	/*
-	 * We need the page table lock to synchronize with kswapd
-	 * and the SMP-safe atomic PTE updates.
+	 * We rely on the mmap_sem and the SMP-safe atomic PTE updates.
+	 * to synchronize with kswapd
 	 */
-	spin_lock(&mm->page_table_lock);
-	pmd = pmd_alloc(mm, pgd, address);
+	if (unlikely(pgd_none(*pgd))) {
+		pmd_t *new = pmd_alloc_one(mm, address);
+		if (!new)
+			return VM_FAULT_OOM;
+
+		/* Insure that the update is done in an atomic way */
+		if (!pgd_test_and_populate(mm, pgd, new))
+			pmd_free(new);
+	}
+
+	pmd = pmd_offset(pgd, address);
+
+	if (likely(pmd)) {
+		pte_t *pte;

-	if (pmd) {
-		pte_t * pte = pte_alloc_map(mm, pmd, address);
-		if (pte)
+		if (!pmd_present(*pmd)) {
+			struct page *new;
+
+			new = pte_alloc_one(mm, address);
+			if (!new)
+				return VM_FAULT_OOM;
+
+			if (!pmd_test_and_populate(mm, pmd, new))
+				pte_free(new);
+			else
+				inc_page_state(nr_page_table_pages);
+		}
+
+		pte = pte_offset_map(pmd, address);
+		if (likely(pte))
 			return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
 	}
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_OOM;
 }

Index: linux-2.6.9/include/asm-generic/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-generic/pgtable.h	2004-10-18 14:53:46.000000000 -0700
+++ linux-2.6.9/include/asm-generic/pgtable.h	2004-12-16 09:59:58.000000000 -0800
@@ -28,6 +28,11 @@
 #endif /* __HAVE_ARCH_SET_PTE_ATOMIC */
 #endif

+/* Get a pte entry without the page table lock */
+#ifndef __HAVE_ARCH_GET_PTE_ATOMIC
+#define get_pte_atomic(__x)	*(__x)
+#endif
+
 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Largely same as above, but only sets the access flags (dirty,
@@ -134,4 +139,61 @@
 #define pgd_offset_gate(mm, addr)	pgd_offset(mm, addr)
 #endif

+#ifndef __HAVE_ARCH_ATOMIC_TABLE_OPS
+/*
+ * If atomic page table operations are not available then use
+ * the page_table_lock to insure some form of locking.
+ * Note thought that low level operations as well as the
+ * page_table_handling of the cpu may bypass all locking.
+ */
+
+#ifndef __HAVE_ARCH_PTEP_CMPXCHG
+#define ptep_cmpxchg(__vma, __addr, __ptep, __oldval, __newval)		\
+({									\
+	int __rc;							\
+	spin_lock(&__vma->vm_mm->page_table_lock);			\
+	__rc = pte_same(*(__ptep), __oldval);				\
+	if (__rc) { set_pte(__ptep, __newval);				\
+		update_mmu_cache(__vma, __addr, __newval); }		\
+	spin_unlock(&__vma->vm_mm->page_table_lock);			\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_ARCH_PGP_TEST_AND_POPULATE
+#define pgd_test_and_populate(__mm, __pgd, __pmd)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pgd_present(*(__pgd));					\
+	if (__rc) pgd_populate(__mm, __pgd, __pmd);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_PMD_TEST_AND_POPULATE
+#define pmd_test_and_populate(__mm, __pmd, __page)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pmd_present(*(__pmd));					\
+	if (__rc) pmd_populate(__mm, __pmd, __page);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#endif
+
+#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH
+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)		\
+({									\
+	pte_t __p = __pte(xchg(&pte_val(*(__ptep)), pte_val(__pteval)));\
+	flush_tlb_page(__vma, __address);				\
+	__p;								\
+})
+
+#endif
+
 #endif /* _ASM_GENERIC_PGTABLE_H */
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-12-15 15:00:22.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-12-16 09:59:58.000000000 -0800
@@ -424,7 +424,10 @@
  * @vma:	the vm area in which the mapping is added
  * @address:	the user virtual address mapped
  *
- * The caller needs to hold the mm->page_table_lock.
+ * The caller needs to hold the mm->page_table_lock if page
+ * is pointing to something that is known by the vm.
+ * The lock does not need to be held if page is pointing
+ * to a newly allocated page.
  */
 void page_add_anon_rmap(struct page *page,
 	struct vm_area_struct *vma, unsigned long address)
@@ -568,11 +571,6 @@

 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address);
-	pteval = ptep_clear_flush(vma, address, pte);
-
-	/* Move the dirty bit to the physical page now the pte is gone. */
-	if (pte_dirty(pteval))
-		set_page_dirty(page);

 	if (PageAnon(page)) {
 		swp_entry_t entry = { .val = page->private };
@@ -587,11 +585,15 @@
 			list_add(&mm->mmlist, &init_mm.mmlist);
 			spin_unlock(&mmlist_lock);
 		}
-		set_pte(pte, swp_entry_to_pte(entry));
+		pteval = ptep_xchg_flush(vma, address, pte, swp_entry_to_pte(entry));
 		BUG_ON(pte_file(*pte));
 		mm->anon_rss--;
-	}
+	} else
+		pteval = ptep_clear_flush(vma, address, pte);

+	/* Move the dirty bit to the physical page now the pte is gone. */
+	if (pte_dirty(pteval))
+		set_page_dirty(page);
 	mm->rss--;
 	page_remove_rmap(page);
 	page_cache_release(page);
@@ -678,15 +680,21 @@
 		if (ptep_clear_flush_young(vma, address, pte))
 			continue;

-		/* Nuke the page table entry. */
 		flush_cache_page(vma, address);
-		pteval = ptep_clear_flush(vma, address, pte);
+		/*
+		 * There would be a race here with handle_mm_fault and do_anonymous_page
+		 * which  bypasses the page_table_lock if we would zap the pte before
+		 * putting something into it. On the other hand we need to
+		 * have the dirty flag setting at the time we replaced the value.
+		 */

 		/* If nonlinear, store the file page offset in the pte. */
 		if (page->index != linear_page_index(vma, address))
-			set_pte(pte, pgoff_to_pte(page->index));
+			pteval = ptep_xchg_flush(vma, address, pte, pgoff_to_pte(page->index));
+		else
+			pteval = ptep_get_and_clear(pte);

-		/* Move the dirty bit to the physical page now the pte is gone. */
+		/* Move the dirty bit to the physical page now that the pte is gone. */
 		if (pte_dirty(pteval))
 			set_page_dirty(page);

page fault scalability patch V13 [2/8]: ia64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for ia64
        * Enhanced parallelism in page fault handler if applied together
          with the generic patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-ia64/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-ia64/pgalloc.h	2004-10-18 14:53:06.000000000 -0700
+++ linux-2.6.9/include/asm-ia64/pgalloc.h	2004-11-19 07:54:19.000000000 -0800
@@ -34,6 +34,10 @@
 #define pmd_quicklist		(local_cpu_data->pmd_quick)
 #define pgtable_cache_size	(local_cpu_data->pgtable_cache_sz)

+/* Empty entries of PMD and PGD */
+#define PMD_NONE       0
+#define PGD_NONE       0
+
 static inline pgd_t*
 pgd_alloc_one_fast (struct mm_struct *mm)
 {
@@ -78,12 +82,19 @@
 	preempt_enable();
 }

+
 static inline void
 pgd_populate (struct mm_struct *mm, pgd_t *pgd_entry, pmd_t *pmd)
 {
 	pgd_val(*pgd_entry) = __pa(pmd);
 }

+/* Atomic populate */
+static inline int
+pgd_test_and_populate (struct mm_struct *mm, pgd_t *pgd_entry, pmd_t *pmd)
+{
+	return ia64_cmpxchg8_acq(pgd_entry,__pa(pmd), PGD_NONE) == PGD_NONE;
+}

 static inline pmd_t*
 pmd_alloc_one_fast (struct mm_struct *mm, unsigned long addr)
@@ -132,6 +143,13 @@
 	pmd_val(*pmd_entry) = page_to_phys(pte);
 }

+/* Atomic populate */
+static inline int
+pmd_test_and_populate (struct mm_struct *mm, pmd_t *pmd_entry, struct page *pte)
+{
+	return ia64_cmpxchg8_acq(pmd_entry, page_to_phys(pte), PMD_NONE) == PMD_NONE;
+}
+
 static inline void
 pmd_populate_kernel (struct mm_struct *mm, pmd_t *pmd_entry, pte_t *pte)
 {
Index: linux-2.6.9/include/asm-ia64/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-ia64/pgtable.h	2004-11-15 11:13:38.000000000 -0800
+++ linux-2.6.9/include/asm-ia64/pgtable.h	2004-11-19 07:55:35.000000000 -0800
@@ -30,6 +30,8 @@
 #define _PAGE_P_BIT		0
 #define _PAGE_A_BIT		5
 #define _PAGE_D_BIT		6
+#define _PAGE_IG_BITS          53
+#define _PAGE_LOCK_BIT         (_PAGE_IG_BITS+3)       /* bit 56. Aligned to 8 bits */

 #define _PAGE_P			(1 << _PAGE_P_BIT)	/* page present bit */
 #define _PAGE_MA_WB		(0x0 <<  2)	/* write back memory attribute */
@@ -58,6 +60,7 @@
 #define _PAGE_PPN_MASK		(((__IA64_UL(1) << IA64_MAX_PHYS_BITS) - 1) & ~0xfffUL)
 #define _PAGE_ED		(__IA64_UL(1) << 52)	/* exception deferral */
 #define _PAGE_PROTNONE		(__IA64_UL(1) << 63)
+#define _PAGE_LOCK		(__IA64_UL(1) << _PAGE_LOCK_BIT)

 /* Valid only for a PTE with the present bit cleared: */
 #define _PAGE_FILE		(1 << 1)		/* see swap & file pte remarks below */
@@ -270,6 +273,8 @@
 #define pte_dirty(pte)		((pte_val(pte) & _PAGE_D) != 0)
 #define pte_young(pte)		((pte_val(pte) & _PAGE_A) != 0)
 #define pte_file(pte)		((pte_val(pte) & _PAGE_FILE) != 0)
+#define pte_locked(pte)		((pte_val(pte) & _PAGE_LOCK)!=0)
+
 /*
  * Note: we convert AR_RWX to AR_RX and AR_RW to AR_R by clearing the 2nd bit in the
  * access rights:
@@ -281,8 +286,15 @@
 #define pte_mkyoung(pte)	(__pte(pte_val(pte) | _PAGE_A))
 #define pte_mkclean(pte)	(__pte(pte_val(pte) & ~_PAGE_D))
 #define pte_mkdirty(pte)	(__pte(pte_val(pte) | _PAGE_D))
+#define pte_mkunlocked(pte)	(__pte(pte_val(pte) & ~_PAGE_LOCK))

 /*
+ * Lock functions for pte's
+ */
+#define ptep_lock(ptep)		test_and_set_bit(_PAGE_LOCK_BIT, ptep)
+#define ptep_unlock(ptep)	{ clear_bit(_PAGE_LOCK_BIT,ptep); smp_mb__after_clear_bit(); }
+#define ptep_unlock_set(ptep, val) set_pte(ptep, pte_mkunlocked(val))
+/*
  * Macro to a page protection value as "uncacheable".  Note that "protection" is really a
  * misnomer here as the protection value contains the memory attribute bits, dirty bits,
  * and various other bits as well.
@@ -342,7 +354,6 @@
 #define pte_unmap_nested(pte)		do { } while (0)

 /* atomic versions of the some PTE manipulations: */
-
 static inline int
 ptep_test_and_clear_young (pte_t *ptep)
 {
@@ -414,6 +425,26 @@
 #endif
 }

+/*
+ * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
+ * information.  However, we use this routine to take care of any (delayed) i-cache
+ * flushing that may be necessary.
+ */
+extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
+
+static inline int
+ptep_cmpxchg (struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	/*
+	 * IA64 defers icache flushes. If the new pte is executable we may
+	 * have to flush the icache to insure cache coherency immediately
+	 * after the cmpxchg.
+	 */
+	if (pte_exec(newval))
+		update_mmu_cache(vma, addr, newval);
+	return ia64_cmpxchg8_acq(&ptep->pte, newval.pte, oldval.pte) == oldval.pte;
+}
+
 static inline int
 pte_same (pte_t a, pte_t b)
 {
@@ -476,13 +507,6 @@
 	struct vm_area_struct * prev, unsigned long start, unsigned long end);
 #endif

-/*
- * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
- * information.  However, we use this routine to take care of any (delayed) i-cache
- * flushing that may be necessary.
- */
-extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
-
 #define __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Update PTEP with ENTRY, which is guaranteed to be a less
@@ -560,6 +584,8 @@
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
 #define __HAVE_ARCH_PGD_OFFSET_GATE
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+#define __HAVE_ARCH_LOCK_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _ASM_IA64_PGTABLE_H */

page fault scalability patch V13 [3/8]: universal cmpxchg for i386

[Christoph Lameter]

Changelog
        * Make cmpxchg and cmpxchg8b generally available on the i386
	  platform.
        * Provide emulation of cmpxchg suitable for uniprocessor if
	  build and run on 386.
        * Provide emulation of cmpxchg8b suitable for uniprocessor
	  systems if build and run on 386 or 486.
	* Provide an inline function to atomically get a 64 bit value
	  via cmpxchg8b in an SMP system (courtesy of Nick Piggin)
	  (important for i386 PAE mode and other places where atomic
	  64 bit operations are useful)

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/arch/i386/Kconfig
===================================================================
--- linux-2.6.9.orig/arch/i386/Kconfig	2004-12-10 09:58:03.000000000 -0800
+++ linux-2.6.9/arch/i386/Kconfig	2004-12-10 09:59:27.000000000 -0800
@@ -351,6 +351,11 @@
 	depends on !M386
 	default y

+config X86_CMPXCHG8B
+	bool
+	depends on !M386 && !M486
+	default y
+
 config X86_XADD
 	bool
 	depends on !M386
Index: linux-2.6.9/arch/i386/kernel/cpu/intel.c
===================================================================
--- linux-2.6.9.orig/arch/i386/kernel/cpu/intel.c	2004-12-06 17:23:49.000000000 -0800
+++ linux-2.6.9/arch/i386/kernel/cpu/intel.c	2004-12-10 09:59:27.000000000 -0800
@@ -6,6 +6,7 @@
 #include <linux/bitops.h>
 #include <linux/smp.h>
 #include <linux/thread_info.h>
+#include <linux/module.h>

 #include <asm/processor.h>
 #include <asm/msr.h>
@@ -287,5 +288,103 @@
 	return 0;
 }

+#ifndef CONFIG_X86_CMPXCHG
+unsigned long cmpxchg_386_u8(volatile void *ptr, u8 old, u8 new)
+{
+	u8 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u8));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u8 *)ptr;
+	if (prev == old)
+		*(u8 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u8);
+
+unsigned long cmpxchg_386_u16(volatile void *ptr, u16 old, u16 new)
+{
+	u16 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u16));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u16 *)ptr;
+	if (prev == old)
+		*(u16 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u16);
+
+unsigned long cmpxchg_386_u32(volatile void *ptr, u32 old, u32 new)
+{
+	u32 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u32));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u32 *)ptr;
+	if (prev == old)
+		*(u32 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u32);
+#endif
+
+#ifndef CONFIG_X86_CMPXCHG8B
+unsigned long long cmpxchg8b_486(volatile unsigned long long *ptr,
+	       unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	unsigned long flags;
+
+	/*
+	 * Check if the kernel was compiled for an old cpu but
+	 * we are running really on a cpu capable of cmpxchg8b
+	 */
+
+	if (cpu_has(cpu_data, X86_FEATURE_CX8))
+		return __cmpxchg8b(ptr, old, newv);
+
+	/* Poor mans cmpxchg8b for 386 and 486. Not suitable for SMP */
+	local_irq_save(flags);
+	prev = *ptr;
+	if (prev == old)
+		*ptr = newv;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg8b_486);
+#endif
+
 // arch_initcall(intel_cpu_init);

Index: linux-2.6.9/include/asm-i386/system.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/system.h	2004-12-06 17:23:55.000000000 -0800
+++ linux-2.6.9/include/asm-i386/system.h	2004-12-10 10:00:49.000000000 -0800
@@ -149,6 +149,9 @@
 #define __xg(x) ((struct __xchg_dummy *)(x))


+#define ll_low(x)	*(((unsigned int*)&(x))+0)
+#define ll_high(x)	*(((unsigned int*)&(x))+1)
+
 /*
  * The semantics of XCHGCMP8B are a bit strange, this is why
  * there is a loop and the loading of %%eax and %%edx has to
@@ -184,8 +187,6 @@
 {
 	__set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
 }
-#define ll_low(x)	*(((unsigned int*)&(x))+0)
-#define ll_high(x)	*(((unsigned int*)&(x))+1)

 static inline void __set_64bit_var (unsigned long long *ptr,
 			 unsigned long long value)
@@ -203,6 +204,26 @@
  __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
  __set_64bit(ptr, ll_low(value), ll_high(value)) )

+static inline unsigned long long __get_64bit(unsigned long long * ptr)
+{
+	unsigned long long ret;
+	__asm__ __volatile__ (
+		"\n1:\t"
+		"movl (%1), %%eax\n\t"
+		"movl 4(%1), %%edx\n\t"
+		"movl %%eax, %%ebx\n\t"
+		"movl %%edx, %%ecx\n\t"
+		LOCK_PREFIX "cmpxchg8b (%1)\n\t"
+		"jnz 1b"
+		:	"=A"(ret)
+		:	"D"(ptr)
+		:	"ebx", "ecx", "memory");
+	return ret;
+}
+
+#define get_64bit(ptr) __get_64bit(ptr)
+
+
 /*
  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
  * Note 2: xchg has side effect, so that attribute volatile is necessary,
@@ -240,7 +261,41 @@
  */

 #ifdef CONFIG_X86_CMPXCHG
+
 #define __HAVE_ARCH_CMPXCHG 1
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))__cmpxchg((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
+
+#else
+
+/*
+ * Building a kernel capable running on 80386. It may be necessary to
+ * simulate the cmpxchg on the 80386 CPU. For that purpose we define
+ * a function for each of the sizes we support.
+ */
+
+extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
+extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
+extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
+
+static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
+				      unsigned long new, int size)
+{
+	switch (size) {
+	case 1:
+		return cmpxchg_386_u8(ptr, old, new);
+	case 2:
+		return cmpxchg_386_u16(ptr, old, new);
+	case 4:
+		return cmpxchg_386_u32(ptr, old, new);
+	}
+	return old;
+}
+
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))cmpxchg_386((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
 #endif

 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
@@ -270,12 +325,34 @@
 	return old;
 }

-#define cmpxchg(ptr,o,n)\
-	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
-					(unsigned long)(n),sizeof(*(ptr))))
-
+static inline unsigned long long __cmpxchg8b(volatile unsigned long long *ptr,
+		unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	__asm__ __volatile__(
+	LOCK_PREFIX "cmpxchg8b (%4)"
+		: "=A" (prev)
+		: "0" (old), "c" ((unsigned long)(newv >> 32)),
+		  "b" ((unsigned long)(newv & 0xffffffffULL)), "D" (ptr)
+		: "memory");
+	return prev;
+}
+
+#ifdef CONFIG_X86_CMPXCHG8B
+#define cmpxchg8b __cmpxchg8b
+#else
+/*
+ * Building a kernel capable of running on 80486 and 80386. Both
+ * do not support cmpxchg8b. Call a function that emulates the
+ * instruction if necessary.
+ */
+extern unsigned long long cmpxchg8b_486(volatile unsigned long long *,
+		unsigned long long, unsigned long long);
+#define cmpxchg8b cmpxchg8b_486
+#endif
+
 #ifdef __KERNEL__
-struct alt_instr {
+struct alt_instr {
 	__u8 *instr; 		/* original instruction */
 	__u8 *replacement;
 	__u8  cpuid;		/* cpuid bit set for replacement */

page fault scalability patch V13 [4/8]: atomic pte operations for i386

[Christoph Lameter]

Changelog
	* Atomic pte operations for i386 in regular and PAE modes

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-i386/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable.h	2004-12-15 15:00:20.000000000 -0800
+++ linux-2.6.9/include/asm-i386/pgtable.h	2004-12-16 10:08:38.000000000 -0800
@@ -413,6 +413,7 @@
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _I386_PGTABLE_H */
Index: linux-2.6.9/include/asm-i386/pgtable-3level.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable-3level.h	2004-10-18 14:54:55.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgtable-3level.h	2004-12-16 10:13:11.000000000 -0800
@@ -6,7 +6,8 @@
  * tables on PPro+ CPUs.
  *
  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
- */
+ * August 26, 2004 added ptep_cmpxchg <christoph@lameter.com>
+*/

 #define pte_ERROR(e) \
 	printk("%s:%d: bad pte %p(%08lx%08lx).\n", __FILE__, __LINE__, &(e), (e).pte_high, (e).pte_low)
@@ -42,21 +43,11 @@
 	return pte_x(pte);
 }

-/* Rules for using set_pte: the pte being assigned *must* be
- * either not present or in a state where the hardware will
- * not attempt to update the pte.  In places where this is
- * not possible, use pte_get_and_clear to obtain the old pte
- * value and then use set_pte to update it.  -ben
- */
-static inline void set_pte(pte_t *ptep, pte_t pte)
-{
-	ptep->pte_high = pte.pte_high;
-	smp_wmb();
-	ptep->pte_low = pte.pte_low;
-}
 #define __HAVE_ARCH_SET_PTE_ATOMIC
 #define set_pte_atomic(pteptr,pteval) \
 		set_64bit((unsigned long long *)(pteptr),pte_val(pteval))
+#define set_pte(pteptr,pteval) \
+		*(unsigned long long *)(pteptr) = pte_val(pteval)
 #define set_pmd(pmdptr,pmdval) \
 		set_64bit((unsigned long long *)(pmdptr),pmd_val(pmdval))
 #define set_pgd(pgdptr,pgdval) \
@@ -142,4 +133,25 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t){ (pte).pte_high })
 #define __swp_entry_to_pte(x)		((pte_t){ 0, (x).val })

+/* Atomic PTE operations */
+#define ptep_xchg_flush(__vma, __addr, __ptep, __newval) \
+({	pte_t __r;							\
+	/* xchg acts as a barrier before the setting of the high bits. */\
+	__r.pte_low = xchg(&(__ptep)->pte_low, (__newval).pte_low);	\
+	__r.pte_high = (__ptep)->pte_high;				\
+	(__ptep)->pte_high = (__newval).pte_high;			\
+	flush_tlb_page(__vma, __addr);					\
+	(__r);								\
+})
+
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
+
+static inline int ptep_cmpxchg(struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg8b((unsigned long long *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
+#define __HAVE_ARCH_GET_PTE_ATOMIC
+#define get_pte_atomic(__ptep) __pte(get_64bit((unsigned long long *)(__ptep)))
+
 #endif /* _I386_PGTABLE_3LEVEL_H */
Index: linux-2.6.9/include/asm-i386/pgtable-2level.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgtable-2level.h	2004-10-18 14:54:31.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgtable-2level.h	2004-12-16 10:08:38.000000000 -0800
@@ -82,4 +82,7 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t) { (pte).pte_low })
 #define __swp_entry_to_pte(x)		((pte_t) { (x).val })

+/* Atomic PTE operations */
+#define ptep_cmpxchg(__vma,__a,__xp,__oldpte,__newpte) (cmpxchg(&(__xp)->pte_low, (__oldpte).pte_low, (__newpte).pte_low)==(__oldpte).pte_low)
+
 #endif /* _I386_PGTABLE_2LEVEL_H */
Index: linux-2.6.9/include/asm-i386/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/pgalloc.h	2004-10-18 14:53:10.000000000 -0700
+++ linux-2.6.9/include/asm-i386/pgalloc.h	2004-12-16 10:08:38.000000000 -0800
@@ -4,9 +4,12 @@
 #include <linux/config.h>
 #include <asm/processor.h>
 #include <asm/fixmap.h>
+#include <asm/system.h>
 #include <linux/threads.h>
 #include <linux/mm.h>		/* for struct page */

+#define PMD_NONE 0L
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte)))

@@ -16,6 +19,19 @@
 		((unsigned long long)page_to_pfn(pte) <<
 			(unsigned long long) PAGE_SHIFT)));
 }
+
+/* Atomic version */
+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+#ifdef CONFIG_X86_PAE
+	return cmpxchg8b( ((unsigned long long *)pmd), PMD_NONE, _PAGE_TABLE +
+		((unsigned long long)page_to_pfn(pte) <<
+			(unsigned long long) PAGE_SHIFT) ) == PMD_NONE;
+#else
+	return cmpxchg( (unsigned long *)pmd, PMD_NONE, _PAGE_TABLE + (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+#endif
+}
+
 /*
  * Allocate and free page tables.
  */
@@ -49,6 +65,7 @@
 #define pmd_free(x)			do { } while (0)
 #define __pmd_free_tlb(tlb,x)		do { } while (0)
 #define pgd_populate(mm, pmd, pte)	BUG()
+#define pgd_test_and_populate(mm, pmd, pte)	({ BUG(); 1; })

 #define check_pgt_cache()	do { } while (0)

page fault scalability patch V13 [5/8]: atomic pte operations for AMD64

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for x86_64

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.9.orig/include/asm-x86_64/pgalloc.h	2004-10-18 14:54:30.000000000 -0700
+++ linux-2.6.9/include/asm-x86_64/pgalloc.h	2004-11-23 10:59:01.000000000 -0800
@@ -7,16 +7,26 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pgd_populate(mm, pgd, pmd) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pmd)))
+#define pgd_test_and_populate(mm, pgd, pmd) \
+		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pmd)) == PGD_NONE)

 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
 	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
 }

+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+}
+
 extern __inline__ pmd_t *get_pmd(void)
 {
 	return (pmd_t *)get_zeroed_page(GFP_KERNEL);
Index: linux-2.6.9/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.9.orig/include/asm-x86_64/pgtable.h	2004-11-22 15:08:43.000000000 -0800
+++ linux-2.6.9/include/asm-x86_64/pgtable.h	2004-11-23 10:59:01.000000000 -0800
@@ -437,6 +437,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

page fault scalability patch V13 [7/8]: Split RSS

[Christoph Lameter]

Changelog
	* Split rss counter into the task structure
	* remove 3 checks of rss in mm/rmap.c
	* increment current->rss instead of mm->rss in the page fault handler
	* move incrementing of anon_rss out of page_add_anon_rmap to group
	  the increments more tightly and allow a better cache utilization

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-11-30 20:33:31.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-11-30 20:33:50.000000000 -0800
@@ -30,6 +30,7 @@
 #include <linux/pid.h>
 #include <linux/percpu.h>
 #include <linux/topology.h>
+#include <linux/rcupdate.h>

 struct exec_domain;

@@ -217,6 +218,7 @@
 	int map_count;				/* number of VMAs */
 	struct rw_semaphore mmap_sem;
 	spinlock_t page_table_lock;		/* Protects page tables, mm->rss, mm->anon_rss */
+	long rss, anon_rss;

 	struct list_head mmlist;		/* List of maybe swapped mm's.  These are globally strung
 						 * together off init_mm.mmlist, and are protected
@@ -226,7 +228,7 @@
 	unsigned long start_code, end_code, start_data, end_data;
 	unsigned long start_brk, brk, start_stack;
 	unsigned long arg_start, arg_end, env_start, env_end;
-	unsigned long rss, anon_rss, total_vm, locked_vm, shared_vm;
+	unsigned long total_vm, locked_vm, shared_vm;
 	unsigned long exec_vm, stack_vm, reserved_vm, def_flags, nr_ptes;

 	unsigned long saved_auxv[42]; /* for /proc/PID/auxv */
@@ -236,6 +238,8 @@

 	/* Architecture-specific MM context */
 	mm_context_t context;
+	struct list_head task_list;		/* Tasks using this mm */
+	struct rcu_head rcu_head;		/* For freeing mm via rcu */

 	/* Token based thrashing protection. */
 	unsigned long swap_token_time;
@@ -545,6 +549,9 @@
 	struct list_head ptrace_list;

 	struct mm_struct *mm, *active_mm;
+	/* Split counters from mm */
+	long rss;
+	long anon_rss;

 /* task state */
 	struct linux_binfmt *binfmt;
@@ -578,6 +585,9 @@
 	struct completion *vfork_done;		/* for vfork() */
 	int __user *set_child_tid;		/* CLONE_CHILD_SETTID */
 	int __user *clear_child_tid;		/* CLONE_CHILD_CLEARTID */
+
+	/* List of other tasks using the same mm */
+	struct list_head mm_tasks;

 	unsigned long rt_priority;
 	unsigned long it_real_value, it_prof_value, it_virt_value;
@@ -1111,6 +1121,14 @@

 #endif

+unsigned long get_rss(struct mm_struct *mm);
+unsigned long get_anon_rss(struct mm_struct *mm);
+unsigned long get_shared(struct mm_struct *mm);
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk);
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk);
+
 #endif /* __KERNEL__ */

 #endif
+
Index: linux-2.6.9/fs/proc/task_mmu.c
===================================================================
--- linux-2.6.9.orig/fs/proc/task_mmu.c	2004-11-30 20:33:26.000000000 -0800
+++ linux-2.6.9/fs/proc/task_mmu.c	2004-11-30 20:33:50.000000000 -0800
@@ -22,7 +22,7 @@
 		"VmPTE:\t%8lu kB\n",
 		(mm->total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),
-		mm->rss << (PAGE_SHIFT-10),
+		get_rss(mm) << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
@@ -37,7 +37,7 @@
 int task_statm(struct mm_struct *mm, int *shared, int *text,
 	       int *data, int *resident)
 {
-	*shared = mm->rss - mm->anon_rss;
+	*shared = get_shared(mm);
 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
 								>> PAGE_SHIFT;
 	*data = mm->total_vm - mm->shared_vm;
Index: linux-2.6.9/fs/proc/array.c
===================================================================
--- linux-2.6.9.orig/fs/proc/array.c	2004-11-30 20:33:26.000000000 -0800
+++ linux-2.6.9/fs/proc/array.c	2004-11-30 20:33:50.000000000 -0800
@@ -420,7 +420,7 @@
 		jiffies_to_clock_t(task->it_real_value),
 		start_time,
 		vsize,
-		mm ? mm->rss : 0, /* you might want to shift this left 3 */
+		mm ? get_rss(mm) : 0, /* you might want to shift this left 3 */
 	        rsslim,
 		mm ? mm->start_code : 0,
 		mm ? mm->end_code : 0,
Index: linux-2.6.9/mm/rmap.c
===================================================================
--- linux-2.6.9.orig/mm/rmap.c	2004-11-30 20:33:46.000000000 -0800
+++ linux-2.6.9/mm/rmap.c	2004-11-30 20:33:50.000000000 -0800
@@ -263,8 +263,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -438,7 +436,7 @@
 	BUG_ON(PageReserved(page));
 	BUG_ON(!anon_vma);

-	vma->vm_mm->anon_rss++;
+	current->anon_rss++;

 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 	index = (address - vma->vm_start) >> PAGE_SHIFT;
@@ -510,8 +508,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -799,8 +795,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;
Index: linux-2.6.9/kernel/fork.c
===================================================================
--- linux-2.6.9.orig/kernel/fork.c	2004-11-30 20:33:42.000000000 -0800
+++ linux-2.6.9/kernel/fork.c	2004-11-30 20:33:50.000000000 -0800
@@ -151,6 +151,7 @@
 	*tsk = *orig;
 	tsk->thread_info = ti;
 	ti->task = tsk;
+	tsk->rss = 0;

 	/* One for us, one for whoever does the "release_task()" (usually parent) */
 	atomic_set(&tsk->usage,2);
@@ -292,6 +293,7 @@
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);
 	INIT_LIST_HEAD(&mm->mmlist);
+	INIT_LIST_HEAD(&mm->task_list);
 	mm->core_waiters = 0;
 	mm->nr_ptes = 0;
 	spin_lock_init(&mm->page_table_lock);
@@ -323,6 +325,13 @@
 	return mm;
 }

+static void rcu_free_mm(struct rcu_head *head)
+{
+	struct mm_struct *mm = container_of(head ,struct mm_struct, rcu_head);
+
+	free_mm(mm);
+}
+
 /*
  * Called when the last reference to the mm
  * is dropped: either by a lazy thread or by
@@ -333,7 +342,7 @@
 	BUG_ON(mm == &init_mm);
 	mm_free_pgd(mm);
 	destroy_context(mm);
-	free_mm(mm);
+	call_rcu(&mm->rcu_head, rcu_free_mm);
 }

 /*
@@ -400,6 +409,8 @@

 	/* Get rid of any cached register state */
 	deactivate_mm(tsk, mm);
+	if (mm)
+		mm_remove_thread(mm, tsk);

 	/* notify parent sleeping on vfork() */
 	if (vfork_done) {
@@ -447,8 +458,8 @@
 		 * new threads start up in user mode using an mm, which
 		 * allows optimizing out ipis; the tlb_gather_mmu code
 		 * is an example.
+		 * (mm_add_thread does use the ptl .... )
 		 */
-		spin_unlock_wait(&oldmm->page_table_lock);
 		goto good_mm;
 	}

@@ -470,6 +481,7 @@
 		goto free_pt;

 good_mm:
+	mm_add_thread(mm, tsk);
 	tsk->mm = mm;
 	tsk->active_mm = mm;
 	return 0;
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-11-30 20:33:46.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-11-30 20:33:50.000000000 -0800
@@ -1467,7 +1467,7 @@
 		 */
 		lru_cache_add_active(page);
 		page_add_anon_rmap(page, vma, addr);
-		mm->rss++;
+		current->rss++;

 	}
 	pte_unmap(page_table);
@@ -1859,3 +1859,87 @@
 }

 #endif
+
+unsigned long get_rss(struct mm_struct *mm)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss;
+
+	if (!mm)
+		return 0;
+
+	rcu_read_lock();
+	rss = mm->rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss += t->rss;
+	}
+	if (rss < 0)
+		rss = 0;
+	rcu_read_unlock();
+	return rss;
+}
+
+unsigned long get_anon_rss(struct mm_struct *mm)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss;
+
+	if (!mm)
+		return 0;
+
+	rcu_read_lock();
+	rss = mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss += t->anon_rss;
+	}
+	if (rss < 0)
+		rss = 0;
+	rcu_read_unlock();
+	return rss;
+}
+
+unsigned long get_shared(struct mm_struct *mm)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss;
+
+	if (!mm)
+		return 0;
+
+	rcu_read_lock();
+	rss = mm->rss - mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss += t->rss - t->anon_rss;
+	}
+	if (rss < 0)
+		rss = 0;
+	rcu_read_unlock();
+	return rss;
+}
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	if (!mm)
+		return;
+
+	spin_lock(&mm->page_table_lock);
+	mm->rss += tsk->rss;
+	mm->anon_rss += tsk->anon_rss;
+	list_del_rcu(&tsk->mm_tasks);
+	spin_unlock(&mm->page_table_lock);
+}
+
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	spin_lock(&mm->page_table_lock);
+	list_add_rcu(&tsk->mm_tasks, &mm->task_list);
+	spin_unlock(&mm->page_table_lock);
+}
+
+
Index: linux-2.6.9/include/linux/init_task.h
===================================================================
--- linux-2.6.9.orig/include/linux/init_task.h	2004-11-30 20:33:30.000000000 -0800
+++ linux-2.6.9/include/linux/init_task.h	2004-11-30 20:33:50.000000000 -0800
@@ -42,6 +42,7 @@
 	.mmlist		= LIST_HEAD_INIT(name.mmlist),		\
 	.cpu_vm_mask	= CPU_MASK_ALL,				\
 	.default_kioctx = INIT_KIOCTX(name.default_kioctx, name),	\
+	.task_list	= LIST_HEAD_INIT(name.task_list),	\
 }

 #define INIT_SIGNALS(sig) {	\
@@ -112,6 +113,7 @@
 	.proc_lock	= SPIN_LOCK_UNLOCKED,				\
 	.switch_lock	= SPIN_LOCK_UNLOCKED,				\
 	.journal_info	= NULL,						\
+	.mm_tasks	= LIST_HEAD_INIT(tsk.mm_tasks),			\
 }


Index: linux-2.6.9/fs/exec.c
===================================================================
--- linux-2.6.9.orig/fs/exec.c	2004-11-30 20:33:41.000000000 -0800
+++ linux-2.6.9/fs/exec.c	2004-11-30 20:33:50.000000000 -0800
@@ -543,6 +543,7 @@
 	active_mm = tsk->active_mm;
 	tsk->mm = mm;
 	tsk->active_mm = mm;
+	mm_add_thread(mm, current);
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
 	arch_pick_mmap_layout(mm);

page fault scalability patch V13 [8/8]: Prefaulting using ptep_cmpxchg

[Christoph Lameter]

The page fault handler for anonymous pages can generate significant overhead
apart from its essential function which is to clear and setup a new page
table entry for a never accessed memory location. This overhead increases
significantly in an SMP environment.

If a fault occurred for page x and is then followed by page x+1 then it may
be reasonable to expect another page fault at x+2 in the future. If page
table entries for x+1 and x+2 would be prepared in the fault handling for
page x+1 then the overhead of taking a fault for x+2 is avoided. However
page x+2 may never be used and thus we may have increased the rss
of an application unnecessarily. The swapper will take care of removing
that page if memory should get tight.

The following patch makes the anonymous fault handler anticipate future
faults. For each fault a prediction is made where the fault would occur
(assuming linear acccess by the application). If the prediction turns out to
be right (next fault is where expected) then a number of pages is
preallocated in order to avoid a series of future faults. The order of the
preallocation increases by the power of two for each success in sequence.

The first successful prediction leads to an additional page being allocated.
Second successful prediction leads to 2 additional pages being allocated.
Third to 4 pages and so on. The max order is 3 by default. In a large
continous allocation the number of faults is reduced by a factor of 8.

The order of preallocation may be controlled through setting the maximum order
in /proc/sys/vm/max_prealloc_order. Setting it to zero will disable
preallocations.

Signed_off_by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/include/linux/sched.h
===================================================================
--- linux-2.6.9.orig/include/linux/sched.h	2004-12-16 10:59:26.000000000 -0800
+++ linux-2.6.9/include/linux/sched.h	2004-12-16 11:06:37.000000000 -0800
@@ -548,6 +548,9 @@
 	struct list_head ptrace_list;

 	struct mm_struct *mm, *active_mm;
+	/* Prefaulting */
+	unsigned long anon_fault_next_addr;
+	int anon_fault_order;
 	/* Split counters from mm */
 	long rss;
 	long anon_rss;
Index: linux-2.6.9/mm/memory.c
===================================================================
--- linux-2.6.9.orig/mm/memory.c	2004-12-16 10:59:26.000000000 -0800
+++ linux-2.6.9/mm/memory.c	2004-12-16 11:06:37.000000000 -0800
@@ -1437,6 +1437,8 @@
 	return ret;
 }

+int sysctl_max_prealloc_order = 3;
+
 /*
  * We are called with the MM semaphore held.
  */
@@ -1445,57 +1447,103 @@
 		pte_t *page_table, pmd_t *pmd, int write_access,
 		unsigned long addr, pte_t orig_entry)
 {
-	pte_t entry;
-	struct page * page = ZERO_PAGE(addr);
+	unsigned long end_addr;
+
+	addr &= PAGE_MASK;
+
+	/* Check if there is a sequential allocation of pages */
+	if (likely((vma->vm_flags & VM_RAND_READ) || current->anon_fault_next_addr != addr)) {
+
+		/* Single page */
+		current->anon_fault_order = 0;
+		end_addr = addr + PAGE_SIZE;
+
+	} else {
+		int order = ++current->anon_fault_order;
+
+		/*
+		 * Calculate the number of pages to preallocate. The order of preallocations
+		 * increases with each successful prediction
+		 */
+		if (unlikely(order > sysctl_max_prealloc_order))
+			order = current->anon_fault_order = sysctl_max_prealloc_order;

-	/* Read-only mapping of ZERO_PAGE. */
-	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+		end_addr = addr + (PAGE_SIZE << order);
+
+		/* Do not prefault beyond vm limits */
+		if (end_addr > vma->vm_end)
+			end_addr = vma->vm_end;
+
+		/* Stay in pmd */
+		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
+		end_addr &= PMD_MASK;
+	}

-	/* ..except if it's a write access */
 	if (write_access) {
-		/* Allocate our own private page. */
-		pte_unmap(page_table);
+		int count = 0;

 		if (unlikely(anon_vma_prepare(vma)))
-			goto no_mem;
-		page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
-		if (!page)
-			goto no_mem;
-		clear_user_highpage(page, addr);
+			return VM_FAULT_OOM;

-		page_table = pte_offset_map(pmd, addr);
+		do {
+			pte_t entry;
+			struct page *page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
+
+			if (unlikely(!page)) {
+				if (!count)
+					return VM_FAULT_OOM;
+				else
+					break;
+			}
+
+			clear_user_highpage(page, addr);
+
+			entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
+							vma->vm_page_prot)),
+						vma);
+
+			/* update the entry */
+			if (unlikely(!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry))) {
+				pte_unmap(page_table);
+				page_cache_release(page);
+				break;
+			}
+
+			page_add_anon_rmap(page, vma, addr);
+			lru_cache_add_active(page);
+			count++;

-		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-							 vma->vm_page_prot)),
-				      vma);
-	}
+			pte_unmap(page_table);
+			addr += PAGE_SIZE;
+			if (addr >= end_addr)
+				break;
+			page_table = pte_offset_map(pmd, addr);
+			orig_entry = *page_table;
+
+		} while (pte_none(orig_entry));
+
+		current->rss += count;
+		current->anon_rss += count;
+
+	} else {
+		pte_t entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+		/* Read */
+		do {
+			if (unlikely(!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)))
+			break;

-	/* update the entry */
-	if (!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)) {
-		if (write_access) {
 			pte_unmap(page_table);
-			page_cache_release(page);
-		}
-		goto out;
-	}
-	if (write_access) {
-		/*
-		 * These two functions must come after the cmpxchg
-		 * because if the page is on the LRU then try_to_unmap may come
-		 * in and unmap the pte.
-		 */
-		page_add_anon_rmap(page, vma, addr);
-		lru_cache_add_active(page);
-		mm->rss++;
-		mm->anon_rss++;
-
+			addr += PAGE_SIZE;
+
+			if (addr >= end_addr)
+				break;
+			page_table = pte_offset_map(pmd, addr);
+			orig_entry = *page_table;
+		} while (pte_none(orig_entry));
 	}
-	pte_unmap(page_table);

-out:
-	return VM_FAULT_MINOR;
-no_mem:
-	return VM_FAULT_OOM;
+	current->anon_fault_next_addr = addr;
+        return VM_FAULT_MINOR;
 }

 /*
Index: linux-2.6.9/kernel/sysctl.c
===================================================================
--- linux-2.6.9.orig/kernel/sysctl.c	2004-12-15 15:00:22.000000000 -0800
+++ linux-2.6.9/kernel/sysctl.c	2004-12-16 11:06:37.000000000 -0800
@@ -56,6 +56,7 @@
 extern int C_A_D;
 extern int sysctl_overcommit_memory;
 extern int sysctl_overcommit_ratio;
+extern int sysctl_max_prealloc_order;
 extern int max_threads;
 extern int sysrq_enabled;
 extern int core_uses_pid;
@@ -816,6 +817,16 @@
 		.strategy	= &sysctl_jiffies,
 	},
 #endif
+	{
+		.ctl_name	= VM_MAX_PREFAULT_ORDER,
+		.procname	= "max_prealloc_order",
+		.data		= &sysctl_max_prealloc_order,
+		.maxlen		= sizeof(sysctl_max_prealloc_order),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &zero,
+	},
 	{ .ctl_name = 0 }
 };

Index: linux-2.6.9/include/linux/sysctl.h
===================================================================
--- linux-2.6.9.orig/include/linux/sysctl.h	2004-12-15 15:00:22.000000000 -0800
+++ linux-2.6.9/include/linux/sysctl.h	2004-12-16 11:06:37.000000000 -0800
@@ -168,6 +168,7 @@
 	VM_VFS_CACHE_PRESSURE=26, /* dcache/icache reclaim pressure */
 	VM_LEGACY_VA_LAYOUT=27, /* legacy/compatibility virtual address space layout */
 	VM_SWAP_TOKEN_TIMEOUT=28, /* default time for token time out */
+	VM_MAX_PREFAULT_ORDER=29, /* max prefault order during anonymous page faults */
 };

Re: page fault scalability patch V13 [0/8]: Overview

[Christoph Lameter]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 149 bytes --]

Its confirmed that my MUA (pine) eats blanks and reformats patches.
Archive of all the patches attached. Use the reformatted stuff for comments only.

[-- Attachment #2: page fault scalability patches v13 --]
[-- Type: APPLICATION/x-gtar, Size: 15736 bytes --]

page fault scalability patch V14 [0/7]: Overview

[Christoph Lameter]

Changes from V13->V14 of this patch:
- 4level page support
- Tested on ia64, i386 and i386 in PAE mode

This is a series of patches that increases the scalability of
the page fault handler for SMP. The performance increase is
accomplished by avoiding the use of the
page_table_lock spinlock (but not mm->mmap_sem) through new atomic
operations on pte's (ptep_xchg, ptep_cmpxchg) and on pmd and pgd's
(pgd_test_and_populate, pmd_test_and_populate).

The page table lock can be avoided in the following situations:

1. An empty pte or pmd entry is populated

This is safe since the swapper may only depopulate them and the
swapper code has been changed to never set a pte to be empty until the
page has been evicted. The population of an empty pte is frequent
if a process touches newly allocated memory.

2. Modifications of flags in a pte entry (write/accessed).

These modifications are done by the CPU or by low level handlers
on various platforms also bypassing the page_table_lock. So this
seems to be safe too.

One essential change in the VM is the use of pte_cmpxchg (or its
generic emulation) on page table entries before doing an
update_mmu_change without holding the page table lock. However, we do
similar things now with other atomic pte operations such as
ptep_get_and_clear and ptep_test_and_clear_dirty. These operations
clear a pte *after* doing an operation on it. The ptep_cmpxchg as used
in this patch operates on an *cleared* pte and replaces it with a pte
pointing to valid memory. The effect of this change on various
architectures has to be thought through. Local definitions of
ptep_cmpxchg and ptep_xchg may be necessary.

For IA64 an icache coherency issue may arise that potentially requires
the flushing of the icache (as done via update_mmu_cache on IA64) prior
to the use of ptep_cmpxchg. Similar issues may arise on other platforms.

The patch introduces a split counter for rss handling to avoid atomic
operations and locks currently necessary for rss modifications. In
addition to mm->rss, tsk->rss is introduced. tsk->rss is defined to be
in the same cache line as tsk->mm (which is already used by the fault
handler) and thus tsk->rss can be incremented without locks
in a fast way. The cache line does not need to be shared between
processors for the page table handler.

A tasklist is generated for each mm (rcu based). Values in that list
are added up to calculate rss or anon_rss values.

The patchset is composed of 7 patches (and was tested against 2.6.10-bk6):

1/7: Avoid page_table_lock in handle_mm_fault

   This patch defers the acquisition of the page_table_lock as much as
   possible and uses atomic operations for allocating anonymous memory.
   These atomic operations are simulated by acquiring the page_table_lock
   for very small time frames if an architecture does not define
   __HAVE_ARCH_ATOMIC_TABLE_OPS. It also changes kswapd so that a
   pte will not be set to empty if a page is in transition to swap.

   If only the first two patches are applied then the time that the
   page_table_lock is held is simply reduced. The lock may then be
   acquired multiple times during a page fault.

2/7: Atomic pte operations for ia64

3/7: Make cmpxchg generally available on i386

   The atomic operations on the page table rely heavily on cmpxchg
   instructions. This patch adds emulations for cmpxchg and cmpxchg8b
   for old 80386 and 80486 cpus. The emulations are only included if a
   kernel is build for these old cpus and are skipped for the real
   cmpxchg instructions if the kernel that is build for 386 or 486 is
   then run on a more recent cpu.

   This patch may be used independently of the other patches.

4/7: Atomic pte operations for i386

   A generally available cmpxchg (last patch) must be available for
   this patch to preserve the ability to build kernels for 386 and 486.

5/7: Atomic pte operation for x86_64

6/7: Atomic pte operations for s390

7/7: Split counter implementation for rss
  Add tsk->rss and tsk->anon_rss. Add tasklist. Add logic
  to calculate rss from tasklist.

Signed-off-by: Christoph Lameter <clameter@sgi.com>

page fault scalability patch V14 [1/7]: Avoid taking page_table_lock

[Christoph Lameter]

Changelog
        * Increase parallelism in SMP configurations by deferring
          the acquisition of page_table_lock in handle_mm_fault
        * Anonymous memory page faults bypass the page_table_lock
          through the use of atomic page table operations
        * Swapper does not set pte to empty in transition to swap
        * Simulate atomic page table operations using the
          page_table_lock if an arch does not define
          __HAVE_ARCH_ATOMIC_TABLE_OPS. This still provides
          a performance benefit since the page_table_lock
          is held for shorter periods of time.

Signed-off-by: Christoph Lameter <clameter@sgi.com

Index: linux-2.6.10/mm/memory.c
===================================================================
--- linux-2.6.10.orig/mm/memory.c	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/mm/memory.c	2005-01-03 15:48:34.000000000 -0800
@@ -1537,8 +1537,7 @@
 }

 /*
- * We hold the mm semaphore and the page_table_lock on entry and
- * should release the pagetable lock on exit..
+ * We hold the mm semaphore
  */
 static int do_swap_page(struct mm_struct * mm,
 	struct vm_area_struct * vma, unsigned long address,
@@ -1550,15 +1549,13 @@
 	int ret = VM_FAULT_MINOR;

 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);
 	page = lookup_swap_cache(entry);
 	if (!page) {
  		swapin_readahead(entry, address, vma);
  		page = read_swap_cache_async(entry, vma, address);
 		if (!page) {
 			/*
-			 * Back out if somebody else faulted in this pte while
-			 * we released the page table lock.
+			 * Back out if somebody else faulted in this pte
 			 */
 			spin_lock(&mm->page_table_lock);
 			page_table = pte_offset_map(pmd, address);
@@ -1581,8 +1578,7 @@
 	lock_page(page);

 	/*
-	 * Back out if somebody else faulted in this pte while we
-	 * released the page table lock.
+	 * Back out if somebody else faulted in this pte
 	 */
 	spin_lock(&mm->page_table_lock);
 	page_table = pte_offset_map(pmd, address);
@@ -1629,14 +1625,12 @@
 }

 /*
- * We are called with the MM semaphore and page_table_lock
- * spinlock held to protect against concurrent faults in
- * multithreaded programs.
+ * We are called with the MM semaphore held.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 		pte_t *page_table, pmd_t *pmd, int write_access,
-		unsigned long addr)
+		unsigned long addr, pte_t orig_entry)
 {
 	pte_t entry;
 	struct page * page = ZERO_PAGE(addr);
@@ -1648,7 +1642,6 @@
 	if (write_access) {
 		/* Allocate our own private page. */
 		pte_unmap(page_table);
-		spin_unlock(&mm->page_table_lock);

 		if (unlikely(anon_vma_prepare(vma)))
 			goto no_mem;
@@ -1657,30 +1650,34 @@
 			goto no_mem;
 		clear_user_highpage(page, addr);

-		spin_lock(&mm->page_table_lock);
 		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
-			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
-		}
-		mm->rss++;
 		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
 							 vma->vm_page_prot)),
 				      vma);
-		lru_cache_add_active(page);
-		mark_page_accessed(page);
-		page_add_anon_rmap(page, vma, addr);
 	}

-	set_pte(page_table, entry);
+	/* update the entry */
+	if (!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)) {
+		if (write_access) {
+			pte_unmap(page_table);
+			page_cache_release(page);
+		}
+		goto out;
+	}
+	if (write_access) {
+		/*
+		 * These two functions must come after the cmpxchg
+		 * because if the page is on the LRU then try_to_unmap may come
+		 * in and unmap the pte.
+		 */
+		page_add_anon_rmap(page, vma, addr);
+		lru_cache_add_active(page);
+		mm->rss++;
+
+	}
 	pte_unmap(page_table);

-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
-	spin_unlock(&mm->page_table_lock);
 out:
 	return VM_FAULT_MINOR;
 no_mem:
@@ -1696,12 +1693,12 @@
  * As this is called only for pages that do not currently exist, we
  * do not need to flush old virtual caches or the TLB.
  *
- * This is called with the MM semaphore held and the page table
- * spinlock held. Exit with the spinlock released.
+ * This is called with the MM semaphore held.
  */
 static int
 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-	unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *page_table,
+        pmd_t *pmd, pte_t orig_entry)
 {
 	struct page * new_page;
 	struct address_space *mapping = NULL;
@@ -1712,9 +1709,8 @@

 	if (!vma->vm_ops || !vma->vm_ops->nopage)
 		return do_anonymous_page(mm, vma, page_table,
-					pmd, write_access, address);
+					pmd, write_access, address, orig_entry);
 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);

 	if (vma->vm_file) {
 		mapping = vma->vm_file->f_mapping;
@@ -1812,7 +1808,7 @@
  * nonlinear vmas.
  */
 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
-	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd, pte_t entry)
 {
 	unsigned long pgoff;
 	int err;
@@ -1825,13 +1821,12 @@
 	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
-		return do_no_page(mm, vma, address, write_access, pte, pmd);
+		return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 	}

-	pgoff = pte_to_pgoff(*pte);
+	pgoff = pte_to_pgoff(entry);

 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);

 	err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
 	if (err == -ENOMEM)
@@ -1850,49 +1845,46 @@
  * with external mmu caches can use to update those (ie the Sparc or
  * PowerPC hashed page tables that act as extended TLBs).
  *
- * Note the "page_table_lock". It is to protect against kswapd removing
- * pages from under us. Note that kswapd only ever _removes_ pages, never
- * adds them. As such, once we have noticed that the page is not present,
- * we can drop the lock early.
- *
- * The adding of pages is protected by the MM semaphore (which we hold),
- * so we don't need to worry about a page being suddenly been added into
- * our VM.
- *
- * We enter with the pagetable spinlock held, we are supposed to
- * release it when done.
+ * Note that kswapd only ever _removes_ pages, never adds them.
+ * We need to insure to handle that case properly.
  */
 static inline int handle_pte_fault(struct mm_struct *mm,
 	struct vm_area_struct * vma, unsigned long address,
 	int write_access, pte_t *pte, pmd_t *pmd)
 {
 	pte_t entry;
+	pte_t new_entry;

-	entry = *pte;
+	/*
+	 * This must be a atomic operation since the page_table_lock is
+	 * not held. If a pte_t larger than the word size is used an
+	 * incorrect value could be read because another processor is
+	 * concurrently updating the multi-word pte. The i386 PAE mode
+	 * is raising its ugly head here.
+	 */
+	entry = get_pte_atomic(pte);
 	if (!pte_present(entry)) {
-		/*
-		 * If it truly wasn't present, we know that kswapd
-		 * and the PTE updates will not touch it later. So
-		 * drop the lock.
-		 */
 		if (pte_none(entry))
-			return do_no_page(mm, vma, address, write_access, pte, pmd);
+			return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 		if (pte_file(entry))
-			return do_file_page(mm, vma, address, write_access, pte, pmd);
+			return do_file_page(mm, vma, address, write_access, pte, pmd, entry);
 		return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
 	}

+	/*
+	 * This is the case in which we only update some bits in the pte.
+	 */
+	new_entry = pte_mkyoung(entry);
 	if (write_access) {
-		if (!pte_write(entry))
+		if (!pte_write(entry)) {
+			/* do_wp_page expects us to hold the page_table_lock */
+			spin_lock(&mm->page_table_lock);
 			return do_wp_page(mm, vma, address, pte, pmd, entry);
-
-		entry = pte_mkdirty(entry);
+		}
+		new_entry = pte_mkdirty(new_entry);
 	}
-	entry = pte_mkyoung(entry);
-	ptep_set_access_flags(vma, address, pte, entry, write_access);
-	update_mmu_cache(vma, address, entry);
+	ptep_cmpxchg(vma, address, pte, entry, new_entry);
 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_MINOR;
 }

@@ -1911,33 +1903,54 @@

 	inc_page_state(pgfault);

-	if (is_vm_hugetlb_page(vma))
+	if (unlikely(is_vm_hugetlb_page(vma)))
 		return VM_FAULT_SIGBUS;	/* mapping truncation does this. */

 	/*
-	 * We need the page table lock to synchronize with kswapd
-	 * and the SMP-safe atomic PTE updates.
+	 * We rely on the mmap_sem and the SMP-safe atomic PTE updates.
+	 * to synchronize with kswapd. We can avoid the overhead
+	 * of the p??_alloc functions through atomic operations so
+	 * we duplicate the functionality of pmd_alloc, pud_alloc and
+	 * pte_alloc_map here.
 	 */
 	pgd = pgd_offset(mm, address);
-	spin_lock(&mm->page_table_lock);
+	if (unlikely(pgd_none(*pgd))) {
+		pud_t *new = pud_alloc_one(mm, address);
+
+		if (!new)
+			return VM_FAULT_OOM;
+		if (!pgd_test_and_populate(mm, pgd, new));
+			pud_free(new);
+	}

-	pud = pud_alloc(mm, pgd, address);
-	if (!pud)
-		goto oom;
-
-	pmd = pmd_alloc(mm, pud, address);
-	if (!pmd)
-		goto oom;
-
-	pte = pte_alloc_map(mm, pmd, address);
-	if (!pte)
-		goto oom;
+	pud = pud_offset(pgd, address);
+	if (unlikely(pud_none(*pud))) {
+		pmd_t *new = pmd_alloc_one(mm, address);
+
+		if (!new)
+			return VM_FAULT_OOM;
+
+		if (!pud_test_and_populate(mm, pud, new))
+			pmd_free(new);
+	}
+
+	pmd = pmd_offset(pud, address);
+	if (unlikely(!pmd_present(*pmd))) {
+		struct page *new = pte_alloc_one(mm, address);

-	return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
+		if (!new)
+			return VM_FAULT_OOM;

- oom:
-	spin_unlock(&mm->page_table_lock);
-	return VM_FAULT_OOM;
+		if (!pmd_test_and_populate(mm, pmd, new))
+			pte_free(new);
+		else {
+			inc_page_state(nr_page_table_pages);
+			mm->nr_ptes++;
+		}
+	}
+
+	pte = pte_offset_map(pmd, address);
+	return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
 }

 #ifndef __ARCH_HAS_4LEVEL_HACK
Index: linux-2.6.10/include/asm-generic/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-generic/pgtable.h	2004-12-24 13:34:30.000000000 -0800
+++ linux-2.6.10/include/asm-generic/pgtable.h	2005-01-03 15:48:34.000000000 -0800
@@ -28,6 +28,11 @@
 #endif /* __HAVE_ARCH_SET_PTE_ATOMIC */
 #endif

+/* Get a pte entry without the page table lock */
+#ifndef __HAVE_ARCH_GET_PTE_ATOMIC
+#define get_pte_atomic(__x)	*(__x)
+#endif
+
 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Largely same as above, but only sets the access flags (dirty,
@@ -134,4 +139,61 @@
 #define pgd_offset_gate(mm, addr)	pgd_offset(mm, addr)
 #endif

+#ifndef __HAVE_ARCH_ATOMIC_TABLE_OPS
+/*
+ * If atomic page table operations are not available then use
+ * the page_table_lock to insure some form of locking.
+ * Note thought that low level operations as well as the
+ * page_table_handling of the cpu may bypass all locking.
+ */
+
+#ifndef __HAVE_ARCH_PTEP_CMPXCHG
+#define ptep_cmpxchg(__vma, __addr, __ptep, __oldval, __newval)		\
+({									\
+	int __rc;							\
+	spin_lock(&__vma->vm_mm->page_table_lock);			\
+	__rc = pte_same(*(__ptep), __oldval);				\
+	if (__rc) { set_pte(__ptep, __newval);				\
+		update_mmu_cache(__vma, __addr, __newval); }		\
+	spin_unlock(&__vma->vm_mm->page_table_lock);			\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_ARCH_PGP_TEST_AND_POPULATE
+#define pgd_test_and_populate(__mm, __pgd, __pmd)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pgd_present(*(__pgd));					\
+	if (__rc) pgd_populate(__mm, __pgd, __pmd);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_PMD_TEST_AND_POPULATE
+#define pmd_test_and_populate(__mm, __pmd, __page)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pmd_present(*(__pmd));					\
+	if (__rc) pmd_populate(__mm, __pmd, __page);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#endif
+
+#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH
+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)		\
+({									\
+	pte_t __p = __pte(xchg(&pte_val(*(__ptep)), pte_val(__pteval)));\
+	flush_tlb_page(__vma, __address);				\
+	__p;								\
+})
+
+#endif
+
 #endif /* _ASM_GENERIC_PGTABLE_H */
Index: linux-2.6.10/mm/rmap.c
===================================================================
--- linux-2.6.10.orig/mm/rmap.c	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/mm/rmap.c	2005-01-03 15:48:34.000000000 -0800
@@ -432,7 +432,10 @@
  * @vma:	the vm area in which the mapping is added
  * @address:	the user virtual address mapped
  *
- * The caller needs to hold the mm->page_table_lock.
+ * The caller needs to hold the mm->page_table_lock if page
+ * is pointing to something that is known by the vm.
+ * The lock does not need to be held if page is pointing
+ * to a newly allocated page.
  */
 void page_add_anon_rmap(struct page *page,
 	struct vm_area_struct *vma, unsigned long address)
@@ -581,11 +584,6 @@

 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address);
-	pteval = ptep_clear_flush(vma, address, pte);
-
-	/* Move the dirty bit to the physical page now the pte is gone. */
-	if (pte_dirty(pteval))
-		set_page_dirty(page);

 	if (PageAnon(page)) {
 		swp_entry_t entry = { .val = page->private };
@@ -600,11 +598,15 @@
 			list_add(&mm->mmlist, &init_mm.mmlist);
 			spin_unlock(&mmlist_lock);
 		}
-		set_pte(pte, swp_entry_to_pte(entry));
+		pteval = ptep_xchg_flush(vma, address, pte, swp_entry_to_pte(entry));
 		BUG_ON(pte_file(*pte));
 		mm->anon_rss--;
-	}
+	} else
+		pteval = ptep_clear_flush(vma, address, pte);

+	/* Move the dirty bit to the physical page now the pte is gone. */
+	if (pte_dirty(pteval))
+		set_page_dirty(page);
 	mm->rss--;
 	page_remove_rmap(page);
 	page_cache_release(page);
@@ -696,15 +698,21 @@
 		if (ptep_clear_flush_young(vma, address, pte))
 			continue;

-		/* Nuke the page table entry. */
 		flush_cache_page(vma, address);
-		pteval = ptep_clear_flush(vma, address, pte);
+		/*
+		 * There would be a race here with handle_mm_fault and do_anonymous_page
+		 * which  bypasses the page_table_lock if we would zap the pte before
+		 * putting something into it. On the other hand we need to
+		 * have the dirty flag setting at the time we replaced the value.
+		 */

 		/* If nonlinear, store the file page offset in the pte. */
 		if (page->index != linear_page_index(vma, address))
-			set_pte(pte, pgoff_to_pte(page->index));
+			pteval = ptep_xchg_flush(vma, address, pte, pgoff_to_pte(page->index));
+		else
+			pteval = ptep_get_and_clear(pte);

-		/* Move the dirty bit to the physical page now the pte is gone. */
+		/* Move the dirty bit to the physical page now that the pte is gone. */
 		if (pte_dirty(pteval))
 			set_page_dirty(page);

Index: linux-2.6.10/include/asm-generic/pgtable-nopud.h
===================================================================
--- linux-2.6.10.orig/include/asm-generic/pgtable-nopud.h	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/include/asm-generic/pgtable-nopud.h	2005-01-03 15:48:34.000000000 -0800
@@ -25,8 +25,9 @@
 static inline int pgd_present(pgd_t pgd)	{ return 1; }
 static inline void pgd_clear(pgd_t *pgd)	{ }
 #define pud_ERROR(pud)				(pgd_ERROR((pud).pgd))
-
 #define pgd_populate(mm, pgd, pud)		do { } while (0)
+static inline int pgd_test_and_populate(struct mm_struct *mm, pgd_t *pgd, pud_t *pud)	{ return 1; }
+
 /*
  * (puds are folded into pgds so this doesn't get actually called,
  * but the define is needed for a generic inline function.)
Index: linux-2.6.10/include/asm-generic/pgtable-nopmd.h
===================================================================
--- linux-2.6.10.orig/include/asm-generic/pgtable-nopmd.h	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/include/asm-generic/pgtable-nopmd.h	2005-01-03 15:49:12.000000000 -0800
@@ -29,6 +29,7 @@
 #define pmd_ERROR(pmd)				(pud_ERROR((pmd).pud))

 #define pud_populate(mm, pmd, pte)		do { } while (0)
+static inline int pud_test_and_populate(struct mm_struct *mm, pud_t *pud, pmd_t *pmd)	{ return 1; }

 /*
  * (pmds are folded into puds so this doesn't get actually called,

page fault scalability patch V14 [2/7]: ia64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for ia64
        * Enhanced parallelism in page fault handler if applied together
          with the generic patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-ia64/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-ia64/pgalloc.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-ia64/pgalloc.h	2005-01-03 12:37:44.000000000 -0800
@@ -34,6 +34,10 @@
 #define pmd_quicklist		(local_cpu_data->pmd_quick)
 #define pgtable_cache_size	(local_cpu_data->pgtable_cache_sz)

+/* Empty entries of PMD and PGD */
+#define PMD_NONE       0
+#define PUD_NONE       0
+
 static inline pgd_t*
 pgd_alloc_one_fast (struct mm_struct *mm)
 {
@@ -84,6 +88,13 @@
 	pud_val(*pud_entry) = __pa(pmd);
 }

+/* Atomic populate */
+static inline int
+pud_test_and_populate (struct mm_struct *mm, pud_t *pud_entry, pmd_t *pmd)
+{
+	return ia64_cmpxchg8_acq(pud_entry,__pa(pmd), PUD_NONE) == PUD_NONE;
+}
+
 static inline pmd_t*
 pmd_alloc_one_fast (struct mm_struct *mm, unsigned long addr)
 {
@@ -131,6 +142,13 @@
 	pmd_val(*pmd_entry) = page_to_phys(pte);
 }

+/* Atomic populate */
+static inline int
+pmd_test_and_populate (struct mm_struct *mm, pmd_t *pmd_entry, struct page *pte)
+{
+	return ia64_cmpxchg8_acq(pmd_entry, page_to_phys(pte), PMD_NONE) == PMD_NONE;
+}
+
 static inline void
 pmd_populate_kernel (struct mm_struct *mm, pmd_t *pmd_entry, pte_t *pte)
 {
Index: linux-2.6.10/include/asm-ia64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-ia64/pgtable.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-ia64/pgtable.h	2005-01-03 12:37:44.000000000 -0800
@@ -30,6 +30,8 @@
 #define _PAGE_P_BIT		0
 #define _PAGE_A_BIT		5
 #define _PAGE_D_BIT		6
+#define _PAGE_IG_BITS          53
+#define _PAGE_LOCK_BIT         (_PAGE_IG_BITS+3)       /* bit 56. Aligned to 8 bits */

 #define _PAGE_P			(1 << _PAGE_P_BIT)	/* page present bit */
 #define _PAGE_MA_WB		(0x0 <<  2)	/* write back memory attribute */
@@ -58,6 +60,7 @@
 #define _PAGE_PPN_MASK		(((__IA64_UL(1) << IA64_MAX_PHYS_BITS) - 1) & ~0xfffUL)
 #define _PAGE_ED		(__IA64_UL(1) << 52)	/* exception deferral */
 #define _PAGE_PROTNONE		(__IA64_UL(1) << 63)
+#define _PAGE_LOCK		(__IA64_UL(1) << _PAGE_LOCK_BIT)

 /* Valid only for a PTE with the present bit cleared: */
 #define _PAGE_FILE		(1 << 1)		/* see swap & file pte remarks below */
@@ -271,6 +274,8 @@
 #define pte_dirty(pte)		((pte_val(pte) & _PAGE_D) != 0)
 #define pte_young(pte)		((pte_val(pte) & _PAGE_A) != 0)
 #define pte_file(pte)		((pte_val(pte) & _PAGE_FILE) != 0)
+#define pte_locked(pte)		((pte_val(pte) & _PAGE_LOCK)!=0)
+
 /*
  * Note: we convert AR_RWX to AR_RX and AR_RW to AR_R by clearing the 2nd bit in the
  * access rights:
@@ -282,8 +287,15 @@
 #define pte_mkyoung(pte)	(__pte(pte_val(pte) | _PAGE_A))
 #define pte_mkclean(pte)	(__pte(pte_val(pte) & ~_PAGE_D))
 #define pte_mkdirty(pte)	(__pte(pte_val(pte) | _PAGE_D))
+#define pte_mkunlocked(pte)	(__pte(pte_val(pte) & ~_PAGE_LOCK))

 /*
+ * Lock functions for pte's
+ */
+#define ptep_lock(ptep)		test_and_set_bit(_PAGE_LOCK_BIT, ptep)
+#define ptep_unlock(ptep)	{ clear_bit(_PAGE_LOCK_BIT,ptep); smp_mb__after_clear_bit(); }
+#define ptep_unlock_set(ptep, val) set_pte(ptep, pte_mkunlocked(val))
+/*
  * Macro to a page protection value as "uncacheable".  Note that "protection" is really a
  * misnomer here as the protection value contains the memory attribute bits, dirty bits,
  * and various other bits as well.
@@ -343,7 +355,6 @@
 #define pte_unmap_nested(pte)		do { } while (0)

 /* atomic versions of the some PTE manipulations: */
-
 static inline int
 ptep_test_and_clear_young (pte_t *ptep)
 {
@@ -415,6 +426,26 @@
 #endif
 }

+/*
+ * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
+ * information.  However, we use this routine to take care of any (delayed) i-cache
+ * flushing that may be necessary.
+ */
+extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
+
+static inline int
+ptep_cmpxchg (struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	/*
+	 * IA64 defers icache flushes. If the new pte is executable we may
+	 * have to flush the icache to insure cache coherency immediately
+	 * after the cmpxchg.
+	 */
+	if (pte_exec(newval))
+		update_mmu_cache(vma, addr, newval);
+	return ia64_cmpxchg8_acq(&ptep->pte, newval.pte, oldval.pte) == oldval.pte;
+}
+
 static inline int
 pte_same (pte_t a, pte_t b)
 {
@@ -477,13 +508,6 @@
 	struct vm_area_struct * prev, unsigned long start, unsigned long end);
 #endif

-/*
- * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
- * information.  However, we use this routine to take care of any (delayed) i-cache
- * flushing that may be necessary.
- */
-extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
-
 #define __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Update PTEP with ENTRY, which is guaranteed to be a less
@@ -561,6 +585,8 @@
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
 #define __HAVE_ARCH_PGD_OFFSET_GATE
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+#define __HAVE_ARCH_LOCK_TABLE_OPS
 #include <asm-generic/pgtable.h>
 #include <asm-generic/pgtable-nopud.h>

page fault scalability patch V14 [3/7]: i386 universal cmpxchg

[Christoph Lameter]

Changelog
        * Make cmpxchg and cmpxchg8b generally available on the i386
	  platform.
        * Provide emulation of cmpxchg suitable for uniprocessor if
	  build and run on 386.
        * Provide emulation of cmpxchg8b suitable for uniprocessor
	  systems if build and run on 386 or 486.
	* Provide an inline function to atomically get a 64 bit value
	  via cmpxchg8b in an SMP system (courtesy of Nick Piggin)
	  (important for i386 PAE mode and other places where atomic
	  64 bit operations are useful)

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/arch/i386/Kconfig
===================================================================
--- linux-2.6.9.orig/arch/i386/Kconfig	2004-12-10 09:58:03.000000000 -0800
+++ linux-2.6.9/arch/i386/Kconfig	2004-12-10 09:59:27.000000000 -0800
@@ -351,6 +351,11 @@
 	depends on !M386
 	default y

+config X86_CMPXCHG8B
+	bool
+	depends on !M386 && !M486
+	default y
+
 config X86_XADD
 	bool
 	depends on !M386
Index: linux-2.6.9/arch/i386/kernel/cpu/intel.c
===================================================================
--- linux-2.6.9.orig/arch/i386/kernel/cpu/intel.c	2004-12-06 17:23:49.000000000 -0800
+++ linux-2.6.9/arch/i386/kernel/cpu/intel.c	2004-12-10 09:59:27.000000000 -0800
@@ -6,6 +6,7 @@
 #include <linux/bitops.h>
 #include <linux/smp.h>
 #include <linux/thread_info.h>
+#include <linux/module.h>

 #include <asm/processor.h>
 #include <asm/msr.h>
@@ -287,5 +288,103 @@
 	return 0;
 }

+#ifndef CONFIG_X86_CMPXCHG
+unsigned long cmpxchg_386_u8(volatile void *ptr, u8 old, u8 new)
+{
+	u8 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u8));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u8 *)ptr;
+	if (prev == old)
+		*(u8 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u8);
+
+unsigned long cmpxchg_386_u16(volatile void *ptr, u16 old, u16 new)
+{
+	u16 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u16));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u16 *)ptr;
+	if (prev == old)
+		*(u16 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u16);
+
+unsigned long cmpxchg_386_u32(volatile void *ptr, u32 old, u32 new)
+{
+	u32 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u32));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u32 *)ptr;
+	if (prev == old)
+		*(u32 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u32);
+#endif
+
+#ifndef CONFIG_X86_CMPXCHG8B
+unsigned long long cmpxchg8b_486(volatile unsigned long long *ptr,
+	       unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	unsigned long flags;
+
+	/*
+	 * Check if the kernel was compiled for an old cpu but
+	 * we are running really on a cpu capable of cmpxchg8b
+	 */
+
+	if (cpu_has(cpu_data, X86_FEATURE_CX8))
+		return __cmpxchg8b(ptr, old, newv);
+
+	/* Poor mans cmpxchg8b for 386 and 486. Not suitable for SMP */
+	local_irq_save(flags);
+	prev = *ptr;
+	if (prev == old)
+		*ptr = newv;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg8b_486);
+#endif
+
 // arch_initcall(intel_cpu_init);

Index: linux-2.6.9/include/asm-i386/system.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/system.h	2004-12-06 17:23:55.000000000 -0800
+++ linux-2.6.9/include/asm-i386/system.h	2004-12-10 10:00:49.000000000 -0800
@@ -149,6 +149,9 @@
 #define __xg(x) ((struct __xchg_dummy *)(x))


+#define ll_low(x)	*(((unsigned int*)&(x))+0)
+#define ll_high(x)	*(((unsigned int*)&(x))+1)
+
 /*
  * The semantics of XCHGCMP8B are a bit strange, this is why
  * there is a loop and the loading of %%eax and %%edx has to
@@ -184,8 +187,6 @@
 {
 	__set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
 }
-#define ll_low(x)	*(((unsigned int*)&(x))+0)
-#define ll_high(x)	*(((unsigned int*)&(x))+1)

 static inline void __set_64bit_var (unsigned long long *ptr,
 			 unsigned long long value)
@@ -203,6 +204,26 @@
  __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
  __set_64bit(ptr, ll_low(value), ll_high(value)) )

+static inline unsigned long long __get_64bit(unsigned long long * ptr)
+{
+	unsigned long long ret;
+	__asm__ __volatile__ (
+		"\n1:\t"
+		"movl (%1), %%eax\n\t"
+		"movl 4(%1), %%edx\n\t"
+		"movl %%eax, %%ebx\n\t"
+		"movl %%edx, %%ecx\n\t"
+		LOCK_PREFIX "cmpxchg8b (%1)\n\t"
+		"jnz 1b"
+		:	"=A"(ret)
+		:	"D"(ptr)
+		:	"ebx", "ecx", "memory");
+	return ret;
+}
+
+#define get_64bit(ptr) __get_64bit(ptr)
+
+
 /*
  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
  * Note 2: xchg has side effect, so that attribute volatile is necessary,
@@ -240,7 +261,41 @@
  */

 #ifdef CONFIG_X86_CMPXCHG
+
 #define __HAVE_ARCH_CMPXCHG 1
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))__cmpxchg((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
+
+#else
+
+/*
+ * Building a kernel capable running on 80386. It may be necessary to
+ * simulate the cmpxchg on the 80386 CPU. For that purpose we define
+ * a function for each of the sizes we support.
+ */
+
+extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
+extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
+extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
+
+static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
+				      unsigned long new, int size)
+{
+	switch (size) {
+	case 1:
+		return cmpxchg_386_u8(ptr, old, new);
+	case 2:
+		return cmpxchg_386_u16(ptr, old, new);
+	case 4:
+		return cmpxchg_386_u32(ptr, old, new);
+	}
+	return old;
+}
+
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))cmpxchg_386((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
 #endif

 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
@@ -270,12 +325,34 @@
 	return old;
 }

-#define cmpxchg(ptr,o,n)\
-	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
-					(unsigned long)(n),sizeof(*(ptr))))
-
+static inline unsigned long long __cmpxchg8b(volatile unsigned long long *ptr,
+		unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	__asm__ __volatile__(
+	LOCK_PREFIX "cmpxchg8b (%4)"
+		: "=A" (prev)
+		: "0" (old), "c" ((unsigned long)(newv >> 32)),
+		  "b" ((unsigned long)(newv & 0xffffffffULL)), "D" (ptr)
+		: "memory");
+	return prev;
+}
+
+#ifdef CONFIG_X86_CMPXCHG8B
+#define cmpxchg8b __cmpxchg8b
+#else
+/*
+ * Building a kernel capable of running on 80486 and 80386. Both
+ * do not support cmpxchg8b. Call a function that emulates the
+ * instruction if necessary.
+ */
+extern unsigned long long cmpxchg8b_486(volatile unsigned long long *,
+		unsigned long long, unsigned long long);
+#define cmpxchg8b cmpxchg8b_486
+#endif
+
 #ifdef __KERNEL__
-struct alt_instr {
+struct alt_instr {
 	__u8 *instr; 		/* original instruction */
 	__u8 *replacement;
 	__u8  cpuid;		/* cpuid bit set for replacement */

page fault scalability patch V14 [4/7]: i386 atomic pte operations

[Christoph Lameter]

Changelog
	* Atomic pte operations for i386 in regular and PAE modes

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-i386/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgtable.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgtable.h	2005-01-03 12:08:35.000000000 -0800
@@ -407,6 +407,7 @@
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _I386_PGTABLE_H */
Index: linux-2.6.10/include/asm-i386/pgtable-3level.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgtable-3level.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgtable-3level.h	2005-01-03 12:11:59.000000000 -0800
@@ -8,7 +8,8 @@
  * tables on PPro+ CPUs.
  *
  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
- */
+ * August 26, 2004 added ptep_cmpxchg <christoph@lameter.com>
+*/

 #define pte_ERROR(e) \
 	printk("%s:%d: bad pte %p(%08lx%08lx).\n", __FILE__, __LINE__, &(e), (e).pte_high, (e).pte_low)
@@ -44,21 +45,11 @@
 	return pte_x(pte);
 }

-/* Rules for using set_pte: the pte being assigned *must* be
- * either not present or in a state where the hardware will
- * not attempt to update the pte.  In places where this is
- * not possible, use pte_get_and_clear to obtain the old pte
- * value and then use set_pte to update it.  -ben
- */
-static inline void set_pte(pte_t *ptep, pte_t pte)
-{
-	ptep->pte_high = pte.pte_high;
-	smp_wmb();
-	ptep->pte_low = pte.pte_low;
-}
 #define __HAVE_ARCH_SET_PTE_ATOMIC
 #define set_pte_atomic(pteptr,pteval) \
 		set_64bit((unsigned long long *)(pteptr),pte_val(pteval))
+#define set_pte(pteptr,pteval) \
+		*(unsigned long long *)(pteptr) = pte_val(pteval)
 #define set_pmd(pmdptr,pmdval) \
 		set_64bit((unsigned long long *)(pmdptr),pmd_val(pmdval))
 #define set_pud(pudptr,pudval) \
@@ -155,4 +146,25 @@

 #define __pmd_free_tlb(tlb, x)		do { } while (0)

+/* Atomic PTE operations */
+#define ptep_xchg_flush(__vma, __addr, __ptep, __newval) \
+({	pte_t __r;							\
+	/* xchg acts as a barrier before the setting of the high bits. */\
+	__r.pte_low = xchg(&(__ptep)->pte_low, (__newval).pte_low);	\
+	__r.pte_high = (__ptep)->pte_high;				\
+	(__ptep)->pte_high = (__newval).pte_high;			\
+	flush_tlb_page(__vma, __addr);					\
+	(__r);								\
+})
+
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
+
+static inline int ptep_cmpxchg(struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg8b((unsigned long long *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
+#define __HAVE_ARCH_GET_PTE_ATOMIC
+#define get_pte_atomic(__ptep) __pte(get_64bit((unsigned long long *)(__ptep)))
+
 #endif /* _I386_PGTABLE_3LEVEL_H */
Index: linux-2.6.10/include/asm-i386/pgtable-2level.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgtable-2level.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgtable-2level.h	2005-01-03 12:08:35.000000000 -0800
@@ -65,4 +65,7 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t) { (pte).pte_low })
 #define __swp_entry_to_pte(x)		((pte_t) { (x).val })

+/* Atomic PTE operations */
+#define ptep_cmpxchg(__vma,__a,__xp,__oldpte,__newpte) (cmpxchg(&(__xp)->pte_low, (__oldpte).pte_low, (__newpte).pte_low)==(__oldpte).pte_low)
+
 #endif /* _I386_PGTABLE_2LEVEL_H */
Index: linux-2.6.10/include/asm-i386/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgalloc.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgalloc.h	2005-01-03 12:11:23.000000000 -0800
@@ -4,9 +4,12 @@
 #include <linux/config.h>
 #include <asm/processor.h>
 #include <asm/fixmap.h>
+#include <asm/system.h>
 #include <linux/threads.h>
 #include <linux/mm.h>		/* for struct page */

+#define PMD_NONE 0L
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte)))

@@ -14,6 +17,18 @@
 	set_pmd(pmd, __pmd(_PAGE_TABLE +			\
 		((unsigned long long)page_to_pfn(pte) <<	\
 			(unsigned long long) PAGE_SHIFT)))
+/* Atomic version */
+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+#ifdef CONFIG_X86_PAE
+	return cmpxchg8b( ((unsigned long long *)pmd), PMD_NONE, _PAGE_TABLE +
+		((unsigned long long)page_to_pfn(pte) <<
+			(unsigned long long) PAGE_SHIFT) ) == PMD_NONE;
+#else
+	return cmpxchg( (unsigned long *)pmd, PMD_NONE, _PAGE_TABLE + (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+#endif
+}
+
 /*
  * Allocate and free page tables.
  */
@@ -44,6 +59,7 @@
 #define pmd_free(x)			do { } while (0)
 #define __pmd_free_tlb(tlb,x)		do { } while (0)
 #define pud_populate(mm, pmd, pte)	BUG()
+#define pud_test_and_populate(mm, pmd, pte) 	({ BUG(); 1; })
 #endif

 #define check_pgt_cache()	do { } while (0)

page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for x86_64

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgalloc.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgalloc.h	2005-01-03 12:21:28.000000000 -0800
@@ -7,6 +7,10 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PUD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pud_populate(mm, pud, pmd) \
@@ -14,11 +18,24 @@
 #define pgd_populate(mm, pgd, pud) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pud)))

+#define pmd_test_and_populate(mm, pmd, pte) \
+		(cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
+#define pud_test_and_populate(mm, pud, pmd) \
+		(cmpxchg((int *)pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
+#define pgd_test_and_populate(mm, pgd, pud) \
+		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
+
+
 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
 	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
 }

+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+}
+
 extern __inline__ pmd_t *get_pmd(void)
 {
 	return (pmd_t *)get_zeroed_page(GFP_KERNEL);
Index: linux-2.6.10/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgtable.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgtable.h	2005-01-03 12:13:17.000000000 -0800
@@ -413,6 +413,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

page fault scalability patch V14 [6/7]: s390 atomic pte operationsw

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for s390

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-s390/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-s390/pgtable.h	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/include/asm-s390/pgtable.h	2005-01-03 12:12:03.000000000 -0800
@@ -569,6 +569,15 @@
 	return pte;
 }

+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)            \
+({                                                                     \
+	struct mm_struct *__mm = __vma->vm_mm;                          \
+	pte_t __pte;                                                    \
+	__pte = ptep_clear_flush(__vma, __address, __ptep);             \
+	set_pte(__ptep, __pteval);                                      \
+	__pte;                                                          \
+})
+
 static inline void ptep_set_wrprotect(pte_t *ptep)
 {
 	pte_t old_pte = *ptep;
@@ -780,6 +789,14 @@

 #define kern_addr_valid(addr)   (1)

+/* Atomic PTE operations */
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
+static inline int ptep_cmpxchg (struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg(ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
 /*
  * No page table caches to initialise
  */
@@ -793,6 +810,7 @@
 #define __HAVE_ARCH_PTEP_CLEAR_DIRTY_FLUSH
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR
 #define __HAVE_ARCH_PTEP_CLEAR_FLUSH
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
Index: linux-2.6.10/include/asm-s390/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-s390/pgalloc.h	2004-12-24 13:35:00.000000000 -0800
+++ linux-2.6.10/include/asm-s390/pgalloc.h	2005-01-03 12:12:03.000000000 -0800
@@ -97,6 +97,10 @@
 	pgd_val(*pgd) = _PGD_ENTRY | __pa(pmd);
 }

+static inline int pgd_test_and_populate(struct mm_struct *mm, pdg_t *pgd, pmd_t *pmd)
+{
+	return cmpxchg(pgd, _PAGE_TABLE_INV, _PGD_ENTRY | __pa(pmd)) == _PAGE_TABLE_INV;
+}
 #endif /* __s390x__ */

 static inline void
@@ -119,6 +123,18 @@
 	pmd_populate_kernel(mm, pmd, (pte_t *)((page-mem_map) << PAGE_SHIFT));
 }

+static inline int
+pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *page)
+{
+	int rc;
+	spin_lock(&mm->page_table_lock);
+
+	rc=pte_same(*pmd, _PAGE_INVALID_EMPTY);
+	if (rc) pmd_populate(mm, pmd, page);
+	spin_unlock(&mm->page_table_lock);
+	return rc;
+}
+
 /*
  * page table entry allocation/free routines.
  */

page fault scalability patch V14 [7/7]: Split RSS counters

[Christoph Lameter]

Changelog
	* Split rss counter into the task structure
	* remove 3 checks of rss in mm/rmap.c
	* increment current->rss instead of mm->rss in the page fault handler
	* move incrementing of anon_rss out of page_add_anon_rmap to group
	  the increments more tightly and allow a better cache utilization

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/linux/sched.h
===================================================================
--- linux-2.6.10.orig/include/linux/sched.h	2004-12-24 13:33:59.000000000 -0800
+++ linux-2.6.10/include/linux/sched.h	2005-01-03 12:21:32.000000000 -0800
@@ -30,6 +30,7 @@
 #include <linux/pid.h>
 #include <linux/percpu.h>
 #include <linux/topology.h>
+#include <linux/rcupdate.h>

 struct exec_domain;

@@ -217,6 +218,7 @@
 	int map_count;				/* number of VMAs */
 	struct rw_semaphore mmap_sem;
 	spinlock_t page_table_lock;		/* Protects page tables, mm->rss, mm->anon_rss */
+	long rss, anon_rss;

 	struct list_head mmlist;		/* List of maybe swapped mm's.  These are globally strung
 						 * together off init_mm.mmlist, and are protected
@@ -226,7 +228,7 @@
 	unsigned long start_code, end_code, start_data, end_data;
 	unsigned long start_brk, brk, start_stack;
 	unsigned long arg_start, arg_end, env_start, env_end;
-	unsigned long rss, anon_rss, total_vm, locked_vm, shared_vm;
+	unsigned long total_vm, locked_vm, shared_vm;
 	unsigned long exec_vm, stack_vm, reserved_vm, def_flags, nr_ptes;

 	unsigned long saved_auxv[42]; /* for /proc/PID/auxv */
@@ -236,6 +238,7 @@

 	/* Architecture-specific MM context */
 	mm_context_t context;
+	struct list_head task_list;             /* Tasks using this mm */

 	/* Token based thrashing protection. */
 	unsigned long swap_token_time;
@@ -545,6 +548,9 @@
 	struct list_head ptrace_list;

 	struct mm_struct *mm, *active_mm;
+	/* Split counters from mm */
+	long rss;
+	long anon_rss;

 /* task state */
 	struct linux_binfmt *binfmt;
@@ -578,6 +584,10 @@
 	struct completion *vfork_done;		/* for vfork() */
 	int __user *set_child_tid;		/* CLONE_CHILD_SETTID */
 	int __user *clear_child_tid;		/* CLONE_CHILD_CLEARTID */
+
+	/* List of other tasks using the same mm */
+	struct list_head mm_tasks;
+	struct rcu_head rcu_head;               /* For freeing the task via rcu */

 	unsigned long rt_priority;
 	unsigned long it_real_value, it_prof_value, it_virt_value;
@@ -1124,6 +1134,12 @@

 #endif

+void get_rss(struct mm_struct *mm, unsigned long *rss, unsigned long *anon_rss);
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk);
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk);
+
 #endif /* __KERNEL__ */

 #endif
+
Index: linux-2.6.10/fs/proc/task_mmu.c
===================================================================
--- linux-2.6.10.orig/fs/proc/task_mmu.c	2004-12-24 13:34:01.000000000 -0800
+++ linux-2.6.10/fs/proc/task_mmu.c	2005-01-03 12:21:32.000000000 -0800
@@ -6,8 +6,9 @@

 char *task_mem(struct mm_struct *mm, char *buffer)
 {
-	unsigned long data, text, lib;
+	unsigned long data, text, lib, rss, anon_rss;

+	get_rss(mm, &rss, &anon_rss);
 	data = mm->total_vm - mm->shared_vm - mm->stack_vm;
 	text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK)) >> 10;
 	lib = (mm->exec_vm << (PAGE_SHIFT-10)) - text;
@@ -22,7 +23,7 @@
 		"VmPTE:\t%8lu kB\n",
 		(mm->total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),
-		mm->rss << (PAGE_SHIFT-10),
+		rss << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
@@ -37,11 +38,14 @@
 int task_statm(struct mm_struct *mm, int *shared, int *text,
 	       int *data, int *resident)
 {
-	*shared = mm->rss - mm->anon_rss;
+	unsigned long rss, anon_rss;
+
+	get_rss(mm, &rss, &anon_rss);
+	*shared = rss - anon_rss;
 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
 								>> PAGE_SHIFT;
 	*data = mm->total_vm - mm->shared_vm;
-	*resident = mm->rss;
+	*resident = rss;
 	return mm->total_vm;
 }

Index: linux-2.6.10/fs/proc/array.c
===================================================================
--- linux-2.6.10.orig/fs/proc/array.c	2004-12-24 13:35:00.000000000 -0800
+++ linux-2.6.10/fs/proc/array.c	2005-01-03 12:21:32.000000000 -0800
@@ -302,7 +302,7 @@

 static int do_task_stat(struct task_struct *task, char * buffer, int whole)
 {
-	unsigned long vsize, eip, esp, wchan = ~0UL;
+	unsigned long rss, anon_rss, vsize, eip, esp, wchan = ~0UL;
 	long priority, nice;
 	int tty_pgrp = -1, tty_nr = 0;
 	sigset_t sigign, sigcatch;
@@ -325,6 +325,7 @@
 		vsize = task_vsize(mm);
 		eip = KSTK_EIP(task);
 		esp = KSTK_ESP(task);
+		get_rss(mm, &rss, &anon_rss);
 	}

 	get_task_comm(tcomm, task);
@@ -420,7 +421,7 @@
 		jiffies_to_clock_t(task->it_real_value),
 		start_time,
 		vsize,
-		mm ? mm->rss : 0, /* you might want to shift this left 3 */
+		mm ? rss : 0, /* you might want to shift this left 3 */
 	        rsslim,
 		mm ? mm->start_code : 0,
 		mm ? mm->end_code : 0,
Index: linux-2.6.10/mm/rmap.c
===================================================================
--- linux-2.6.10.orig/mm/rmap.c	2005-01-03 10:31:41.000000000 -0800
+++ linux-2.6.10/mm/rmap.c	2005-01-03 12:21:32.000000000 -0800
@@ -264,8 +264,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -446,8 +444,6 @@
 	BUG_ON(PageReserved(page));
 	BUG_ON(!anon_vma);

-	vma->vm_mm->anon_rss++;
-
 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 	index = (address - vma->vm_start) >> PAGE_SHIFT;
 	index += vma->vm_pgoff;
@@ -519,8 +515,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -817,8 +811,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;
Index: linux-2.6.10/kernel/fork.c
===================================================================
--- linux-2.6.10.orig/kernel/fork.c	2004-12-24 13:33:59.000000000 -0800
+++ linux-2.6.10/kernel/fork.c	2005-01-03 12:21:32.000000000 -0800
@@ -78,10 +78,16 @@
 static kmem_cache_t *task_struct_cachep;
 #endif

+static void rcu_free_task(struct rcu_head *head)
+{
+	struct task_struct *tsk = container_of(head ,struct task_struct, rcu_head);
+	free_task_struct(tsk);
+}
+
 void free_task(struct task_struct *tsk)
 {
 	free_thread_info(tsk->thread_info);
-	free_task_struct(tsk);
+	call_rcu(&tsk->rcu_head, rcu_free_task);
 }
 EXPORT_SYMBOL(free_task);

@@ -98,7 +104,7 @@
 	put_group_info(tsk->group_info);

 	if (!profile_handoff_task(tsk))
-		free_task(tsk);
+		call_rcu(&tsk->rcu_head, rcu_free_task);
 }

 void __init fork_init(unsigned long mempages)
@@ -151,6 +157,7 @@
 	*tsk = *orig;
 	tsk->thread_info = ti;
 	ti->task = tsk;
+	tsk->rss = 0;

 	/* One for us, one for whoever does the "release_task()" (usually parent) */
 	atomic_set(&tsk->usage,2);
@@ -292,6 +299,7 @@
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);
 	INIT_LIST_HEAD(&mm->mmlist);
+	INIT_LIST_HEAD(&mm->task_list);
 	mm->core_waiters = 0;
 	mm->nr_ptes = 0;
 	spin_lock_init(&mm->page_table_lock);
@@ -400,6 +408,8 @@

 	/* Get rid of any cached register state */
 	deactivate_mm(tsk, mm);
+	if (mm)
+		mm_remove_thread(mm, tsk);

 	/* notify parent sleeping on vfork() */
 	if (vfork_done) {
@@ -447,8 +457,8 @@
 		 * new threads start up in user mode using an mm, which
 		 * allows optimizing out ipis; the tlb_gather_mmu code
 		 * is an example.
+		 * (mm_add_thread does use the ptl .... )
 		 */
-		spin_unlock_wait(&oldmm->page_table_lock);
 		goto good_mm;
 	}

@@ -470,6 +480,7 @@
 		goto free_pt;

 good_mm:
+	mm_add_thread(mm, tsk);
 	tsk->mm = mm;
 	tsk->active_mm = mm;
 	return 0;
@@ -1063,7 +1074,7 @@
 	atomic_dec(&p->user->processes);
 	free_uid(p->user);
 bad_fork_free:
-	free_task(p);
+	call_rcu(&p->rcu_head, rcu_free_task);
 	goto fork_out;
 }

Index: linux-2.6.10/mm/memory.c
===================================================================
--- linux-2.6.10.orig/mm/memory.c	2005-01-03 11:15:55.000000000 -0800
+++ linux-2.6.10/mm/memory.c	2005-01-03 12:21:32.000000000 -0800
@@ -909,6 +909,7 @@
 			struct page *map;
 			int lookup_write = write;
 			while (!(map = follow_page(mm, start, lookup_write))) {
+				unsigned long rss, anon_rss;
 				/*
 				 * Shortcut for anonymous pages. We don't want
 				 * to force the creation of pages tables for
@@ -921,6 +922,17 @@
 					map = ZERO_PAGE(start);
 					break;
 				}
+				if (mm != current->mm) {
+					/*
+					 * handle_mm_fault uses the current pointer
+					 * for a split rss counter. The current pointer
+					 * is not correct if we are using a different mm
+					 */
+					rss = current->rss;
+					anon_rss = current->anon_rss;
+					current->rss = 0;
+					current->anon_rss = 0;
+				}
 				spin_unlock(&mm->page_table_lock);
 				switch (handle_mm_fault(mm,vma,start,write)) {
 				case VM_FAULT_MINOR:
@@ -945,6 +957,12 @@
 				 */
 				lookup_write = write && !force;
 				spin_lock(&mm->page_table_lock);
+				if (mm != current->mm) {
+					mm->rss += current->rss;
+					mm->anon_rss += current->anon_rss;
+					current->rss = rss;
+					current->anon_rss = anon_rss;
+				}
 			}
 			if (pages) {
 				pages[i] = get_page_map(map);
@@ -1325,6 +1343,7 @@
 		break_cow(vma, new_page, address, page_table);
 		lru_cache_add_active(new_page);
 		page_add_anon_rmap(new_page, vma, address);
+		mm->anon_rss++;

 		/* Free the old page.. */
 		new_page = old_page;
@@ -1608,6 +1627,7 @@
 	flush_icache_page(vma, page);
 	set_pte(page_table, pte);
 	page_add_anon_rmap(page, vma, address);
+	mm->anon_rss++;

 	if (write_access) {
 		if (do_wp_page(mm, vma, address,
@@ -1674,6 +1694,7 @@
 		page_add_anon_rmap(page, vma, addr);
 		lru_cache_add_active(page);
 		mm->rss++;
+		mm->anon_rss++;

 	}
 	pte_unmap(page_table);
@@ -1780,6 +1801,7 @@
 		if (anon) {
 			lru_cache_add_active(new_page);
 			page_add_anon_rmap(new_page, vma, address);
+			mm->anon_rss++;
 		} else
 			page_add_file_rmap(new_page);
 		pte_unmap(page_table);
@@ -2143,3 +2165,49 @@
 }

 #endif
+
+void get_rss(struct mm_struct *mm, unsigned long *rss, unsigned long *anon_rss)
+{
+	struct list_head *y;
+	struct task_struct *t;
+        long rss_sum, anon_rss_sum;
+
+	rcu_read_lock();
+	rss_sum = mm->rss;
+	anon_rss_sum = mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss_sum += t->rss;
+		anon_rss_sum += t->anon_rss;
+	}
+	if (rss_sum < 0)
+		rss_sum = 0;
+	if (anon_rss_sum < 0)
+		anon_rss_sum = 0;
+	rcu_read_unlock();
+	*rss = rss_sum;
+	*anon_rss = anon_rss_sum;
+}
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	if (!mm)
+		return;
+
+	spin_lock(&mm->page_table_lock);
+	mm->rss += tsk->rss;
+	mm->anon_rss += tsk->anon_rss;
+	list_del_rcu(&tsk->mm_tasks);
+	spin_unlock(&mm->page_table_lock);
+}
+
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	spin_lock(&mm->page_table_lock);
+	tsk->rss = 0;
+	tsk->anon_rss = 0;
+	list_add_rcu(&tsk->mm_tasks, &mm->task_list);
+	spin_unlock(&mm->page_table_lock);
+}
+
+
Index: linux-2.6.10/include/linux/init_task.h
===================================================================
--- linux-2.6.10.orig/include/linux/init_task.h	2004-12-24 13:33:52.000000000 -0800
+++ linux-2.6.10/include/linux/init_task.h	2005-01-03 12:21:32.000000000 -0800
@@ -42,6 +42,7 @@
 	.mmlist		= LIST_HEAD_INIT(name.mmlist),		\
 	.cpu_vm_mask	= CPU_MASK_ALL,				\
 	.default_kioctx = INIT_KIOCTX(name.default_kioctx, name),	\
+	.task_list	= LIST_HEAD_INIT(name.task_list),	\
 }

 #define INIT_SIGNALS(sig) {	\
@@ -112,6 +113,7 @@
 	.proc_lock	= SPIN_LOCK_UNLOCKED,				\
 	.switch_lock	= SPIN_LOCK_UNLOCKED,				\
 	.journal_info	= NULL,						\
+	.mm_tasks	= LIST_HEAD_INIT(tsk.mm_tasks),			\
 }


Index: linux-2.6.10/fs/exec.c
===================================================================
--- linux-2.6.10.orig/fs/exec.c	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/fs/exec.c	2005-01-03 12:21:32.000000000 -0800
@@ -549,6 +549,7 @@
 	tsk->active_mm = mm;
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
+	mm_add_thread(mm, current);
 	arch_pick_mmap_layout(mm);
 	if (old_mm) {
 		if (active_mm != old_mm) BUG();
Index: linux-2.6.10/fs/aio.c
===================================================================
--- linux-2.6.10.orig/fs/aio.c	2004-12-24 13:34:44.000000000 -0800
+++ linux-2.6.10/fs/aio.c	2005-01-03 12:21:32.000000000 -0800
@@ -577,6 +577,7 @@
 	tsk->active_mm = mm;
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
+	mm_add_thread(mm, tsk);

 	mmdrop(active_mm);
 }
@@ -596,6 +597,7 @@
 {
 	struct task_struct *tsk = current;

+	mm_remove_thread(mm,tsk);
 	task_lock(tsk);
 	tsk->flags &= ~PF_BORROWED_MM;
 	tsk->mm = NULL;
Index: linux-2.6.10/mm/swapfile.c
===================================================================
--- linux-2.6.10.orig/mm/swapfile.c	2005-01-03 10:31:31.000000000 -0800
+++ linux-2.6.10/mm/swapfile.c	2005-01-03 12:21:32.000000000 -0800
@@ -432,6 +432,7 @@
 	swp_entry_t entry, struct page *page)
 {
 	vma->vm_mm->rss++;
+	vma->vm_mm->anon_rss++;
 	get_page(page);
 	set_pte(dir, pte_mkold(mk_pte(page, vma->vm_page_prot)));
 	page_add_anon_rmap(page, vma, address);

Re: page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Andi Kleen]

Christoph Lameter <clameter@sgi.com> writes:

I bet this has been never tested.

>  #define pmd_populate_kernel(mm, pmd, pte) \
>  		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
>  #define pud_populate(mm, pud, pmd) \
> @@ -14,11 +18,24 @@
>  #define pgd_populate(mm, pgd, pud) \
>  		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pud)))
>
> +#define pmd_test_and_populate(mm, pmd, pte) \
> +		(cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
> +#define pud_test_and_populate(mm, pud, pmd) \
> +		(cmpxchg((int *)pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
> +#define pgd_test_and_populate(mm, pgd, pud) \
> +		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
> +

Shouldn't this all be (long *)pmd ? page table entries on x86-64 are 64bit.
Also why do you cast at all? i think the macro should handle an arbitary
pointer.

> +
>  static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
>  {
>  	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
>  }
>
> +static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
> +{
> +	return cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;

Same.

-Andi

Re: page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Christoph Lameter]

On Tue, 4 Jan 2005, Andi Kleen wrote:

> Christoph Lameter <clameter@sgi.com> writes:
>
> I bet this has been never tested.

I tested this back in October and it worked fine. Would you be able to
test your proposed modifications and send me a patch?

> > +#define pmd_test_and_populate(mm, pmd, pte) \
> > +		(cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
> > +#define pud_test_and_populate(mm, pud, pmd) \
> > +		(cmpxchg((int *)pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
> > +#define pgd_test_and_populate(mm, pgd, pud) \
> > +		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
> > +
>
> Shouldn't this all be (long *)pmd ? page table entries on x86-64 are 64bit.
> Also why do you cast at all? i think the macro should handle an arbitary
> pointer.

The macro checks for the size of the pointer and then generates the
appropriate cmpxchg instruction. pgd_t is a struct which may be
problematic for the cmpxchg macros.

Re: page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Andi Kleen]

On Tue, Jan 04, 2005 at 11:58:13AM -0800, Christoph Lameter wrote:
> On Tue, 4 Jan 2005, Andi Kleen wrote:
> 
> > Christoph Lameter <clameter@sgi.com> writes:
> >
> > I bet this has been never tested.
> 
> I tested this back in October and it worked fine. Would you be able to
> test your proposed modifications and send me a patch?

Hmm, I don't think it could have worked this way, except
if you only tested page faults < 4GB. 

> 
> > > +#define pmd_test_and_populate(mm, pmd, pte) \
> > > +		(cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
> > > +#define pud_test_and_populate(mm, pud, pmd) \
> > > +		(cmpxchg((int *)pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
> > > +#define pgd_test_and_populate(mm, pgd, pud) \
> > > +		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
> > > +
> >
> > Shouldn't this all be (long *)pmd ? page table entries on x86-64 are 64bit.
> > Also why do you cast at all? i think the macro should handle an arbitary
> > pointer.
> 
> The macro checks for the size of the pointer and then generates the
> appropriate cmpxchg instruction. pgd_t is a struct which may be
> problematic for the cmpxchg macros.

It just checks sizeof, that should be fine.

-Andi

Re: page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Christoph Lameter]

On Tue, 4 Jan 2005, Andi Kleen wrote:

> > The macro checks for the size of the pointer and then generates the
> > appropriate cmpxchg instruction. pgd_t is a struct which may be
> > problematic for the cmpxchg macros.
>
> It just checks sizeof, that should be fine.

Index: linux-2.6.10/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgalloc.h	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgalloc.h	2005-01-04 12:31:14.000000000 -0800
@@ -7,6 +7,10 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PUD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pud_populate(mm, pud, pmd) \
@@ -14,11 +18,24 @@
 #define pgd_populate(mm, pgd, pud) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pud)))

+#define pmd_test_and_populate(mm, pmd, pte) \
+		(cmpxchg(pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
+#define pud_test_and_populate(mm, pud, pmd) \
+		(cmpxchg(pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
+#define pgd_test_and_populate(mm, pgd, pud) \
+		(cmpxchg(pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
+
+
 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
 	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
 }

+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg(pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+}
+
 extern __inline__ pmd_t *get_pmd(void)
 {
 	return (pmd_t *)get_zeroed_page(GFP_KERNEL);
Index: linux-2.6.10/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgtable.h	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgtable.h	2005-01-04 12:29:25.000000000 -0800
@@ -413,6 +413,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

Re: page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Brian Gerst]

Christoph Lameter wrote:
> Changelog
>         * Provide atomic pte operations for x86_64
> 
> Signed-off-by: Christoph Lameter <clameter@sgi.com>
> 
> Index: linux-2.6.10/include/asm-x86_64/pgalloc.h
> ===================================================================
> --- linux-2.6.10.orig/include/asm-x86_64/pgalloc.h	2005-01-03 10:31:31.000000000 -0800
> +++ linux-2.6.10/include/asm-x86_64/pgalloc.h	2005-01-03 12:21:28.000000000 -0800
> @@ -7,6 +7,10 @@
>  #include <linux/threads.h>
>  #include <linux/mm.h>
> 
> +#define PMD_NONE 0
> +#define PUD_NONE 0
> +#define PGD_NONE 0
> +
>  #define pmd_populate_kernel(mm, pmd, pte) \
>  		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
>  #define pud_populate(mm, pud, pmd) \
> @@ -14,11 +18,24 @@
>  #define pgd_populate(mm, pgd, pud) \
>  		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pud)))
> 
> +#define pmd_test_and_populate(mm, pmd, pte) \
> +		(cmpxchg((int *)pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
> +#define pud_test_and_populate(mm, pud, pmd) \
> +		(cmpxchg((int *)pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
                                 ^^^
Shouldn't this be pud?

> +#define pgd_test_and_populate(mm, pgd, pud) \
> +		(cmpxchg((int *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
> +
> +

--
				Brian Gerst

Re: page fault scalability patch V14 [5/7]: x86_64 atomic pte operations

[Christoph Lameter]

On Tue, 4 Jan 2005, Brian Gerst wrote:

> > +#define pud_test_and_populate(mm, pud, pmd) \
> > +		(cmpxchg((int *)pgd, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
>                                  ^^^
> Shouldn't this be pud?

Corrrect. Sigh. Could someone test this on x86_64?

Index: linux-2.6.10/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgalloc.h	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgalloc.h	2005-01-04 12:31:14.000000000 -0800
@@ -7,6 +7,10 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PUD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pud_populate(mm, pud, pmd) \
@@ -14,11 +18,24 @@
 #define pgd_populate(mm, pgd, pud) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pud)))

+#define pmd_test_and_populate(mm, pmd, pte) \
+		(cmpxchg(pmd, PMD_NONE, _PAGE_TABLE | __pa(pte)) == PMD_NONE)
+#define pud_test_and_populate(mm, pud, pmd) \
+		(cmpxchg(pud, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
+#define pgd_test_and_populate(mm, pgd, pud) \
+		(cmpxchg(pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
+
+
 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
 	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
 }

+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg(pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+}
+
 extern __inline__ pmd_t *get_pmd(void)
 {
 	return (pmd_t *)get_zeroed_page(GFP_KERNEL);
Index: linux-2.6.10/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgtable.h	2005-01-03 15:02:01.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgtable.h	2005-01-04 12:29:25.000000000 -0800
@@ -413,6 +413,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

Anticipatory prefaulting in the page fault handler V4

[Christoph Lameter]

Changes from V3 to V4:
- Add /proc/sys/vm/max_prealloc_order to limit preallocations
- Tested against 2.6.10-bk7

(This version of the patch is not depending on atomic pte operations and will
conflict with the page fault scalabilty patchset. I have another patch
that works with atomic pte operations)

The page fault handler for anonymous pages can generate significant overhead
apart from its essential function which is to clear and setup a new page
table entry for a never accessed memory location. This overhead increases
significantly in an SMP environment.

In the page table scalability patches, we addressed the issue by changing
the locking scheme so that multiple fault handlers are able to be processed
concurrently on multiple cpus. This patch attempts to aggregate multiple
page faults into a single one. It does that by noting
anonymous page faults generated in sequence by an application.

If a fault occurred for page x and is then followed by page x+1 then it may
be reasonable to expect another page fault at x+2 in the future. If page
table entries for x+1 and x+2 would be prepared in the fault handling for
page x+1 then the overhead of taking a fault for x+2 is avoided. However
page x+2 may never be used and thus we may have increased the rss
of an application unnecessarily. The swapper will take care of removing
that page if memory should get tight.

The following patch makes the anonymous fault handler anticipate future
faults. For each fault a prediction is made where the fault would occur
(assuming linear acccess by the application). If the prediction turns out to
be right (next fault is where expected) then a number of pages is
preallocated in order to avoid a series of future faults. The order of the
preallocation increases by the power of two for each success in sequence.

The first successful prediction leads to an additional page being allocated.
Second successful prediction leads to 2 additional pages being allocated.
Third to 4 pages and so on. The max order is 3 by default. In a large
continous allocation the number of faults is reduced by a factor of 8.

Patch against 2.6.10-bk7:

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/linux/sched.h
===================================================================
--- linux-2.6.10.orig/include/linux/sched.h	2005-01-04 13:55:00.000000000 -0800
+++ linux-2.6.10/include/linux/sched.h	2005-01-04 14:00:27.000000000 -0800
@@ -537,6 +537,8 @@
 #endif

 	struct list_head tasks;
+	unsigned long anon_fault_next_addr;	/* Predicted sequential fault address */
+	int anon_fault_order;			/* Last order of allocation on fault */
 	/*
 	 * ptrace_list/ptrace_children forms the list of my children
 	 * that were stolen by a ptracer.
Index: linux-2.6.10/mm/memory.c
===================================================================
--- linux-2.6.10.orig/mm/memory.c	2005-01-04 13:55:00.000000000 -0800
+++ linux-2.6.10/mm/memory.c	2005-01-04 14:00:27.000000000 -0800
@@ -57,6 +57,7 @@

 #include <linux/swapops.h>
 #include <linux/elf.h>
+#include <linux/pagevec.h>

 #ifndef CONFIG_DISCONTIGMEM
 /* use the per-pgdat data instead for discontigmem - mbligh */
@@ -1626,6 +1627,8 @@
 	return ret;
 }

+int sysctl_max_prealloc_order = 4;
+
 /*
  * We are called with the MM semaphore and page_table_lock
  * spinlock held to protect against concurrent faults in
@@ -1637,52 +1640,105 @@
 		unsigned long addr)
 {
 	pte_t entry;
-	struct page * page = ZERO_PAGE(addr);
+ 	unsigned long end_addr;
+
+	addr &= PAGE_MASK;

-	/* Read-only mapping of ZERO_PAGE. */
-	entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+ 	if (likely((vma->vm_flags & VM_RAND_READ)
+		|| current->anon_fault_next_addr != addr)
+		|| current->anon_fault_order >= sysctl_max_prealloc_order) {
+		/* Single page */
+		current->anon_fault_order = 0;
+		end_addr = addr + PAGE_SIZE;
+	} else {
+		/* Sequence of faults detect. Perform preallocation */
+ 		int order = ++current->anon_fault_order;
+
+		if ((1 << order) < PAGEVEC_SIZE)
+			end_addr = addr + (PAGE_SIZE << order);
+		else {
+			end_addr = addr + PAGEVEC_SIZE * PAGE_SIZE;
+			current->anon_fault_order = 3;
+		}

-	/* ..except if it's a write access */
+		if (end_addr > vma->vm_end)
+			end_addr = vma->vm_end;
+		if ((addr & PMD_MASK) != (end_addr & PMD_MASK))
+			end_addr &= PMD_MASK;
+	}
 	if (write_access) {
-		/* Allocate our own private page. */
+
+		unsigned long a;
+		int i;
+		struct pagevec pv;
+
 		pte_unmap(page_table);
 		spin_unlock(&mm->page_table_lock);

+		pagevec_init(&pv, 0);
+
 		if (unlikely(anon_vma_prepare(vma)))
-			goto no_mem;
-		page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
-		if (!page)
-			goto no_mem;
-		clear_user_highpage(page, addr);
+			return VM_FAULT_OOM;
+
+		/* Allocate the necessary pages */
+		for(a = addr; a < end_addr ; a += PAGE_SIZE) {
+			struct page *p = alloc_page_vma(GFP_HIGHUSER, vma, a);
+
+			if (likely(p)) {
+				clear_user_highpage(p, a);
+				pagevec_add(&pv, p);
+			} else {
+				if (a == addr)
+					return VM_FAULT_OOM;
+				break;
+			}
+		}

 		spin_lock(&mm->page_table_lock);
-		page_table = pte_offset_map(pmd, addr);

-		if (!pte_none(*page_table)) {
+		for(i = 0; addr < a; addr += PAGE_SIZE, i++) {
+			struct page *p = pv.pages[i];
+
+			page_table = pte_offset_map(pmd, addr);
+			if (unlikely(!pte_none(*page_table))) {
+				/* Someone else got there first */
+				pte_unmap(page_table);
+				page_cache_release(p);
+				continue;
+			}
+
+ 			entry = maybe_mkwrite(pte_mkdirty(mk_pte(p,
+ 						 vma->vm_page_prot)),
+ 					      vma);
+
+			mm->rss++;
+			lru_cache_add_active(p);
+			SetPageReferenced(p);
+			page_add_anon_rmap(p, vma, addr);
+
+			set_pte(page_table, entry);
 			pte_unmap(page_table);
-			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
+
+ 			/* No need to invalidate - it was non-present before */
+ 			update_mmu_cache(vma, addr, entry);
+		}
+ 	} else {
+ 		/* Read */
+		entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
+nextread:
+		set_pte(page_table, entry);
+		pte_unmap(page_table);
+		update_mmu_cache(vma, addr, entry);
+		addr += PAGE_SIZE;
+		if (unlikely(addr < end_addr)) {
+			page_table = pte_offset_map(pmd, addr);
+			if (likely(pte_none(*page_table)))
+				goto nextread;
 		}
-		mm->rss++;
-		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-							 vma->vm_page_prot)),
-				      vma);
-		lru_cache_add_active(page);
-		SetPageReferenced(page);
-		page_add_anon_rmap(page, vma, addr);
 	}
-
-	set_pte(page_table, entry);
-	pte_unmap(page_table);
-
-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
+	current->anon_fault_next_addr = addr;
 	spin_unlock(&mm->page_table_lock);
-out:
 	return VM_FAULT_MINOR;
-no_mem:
-	return VM_FAULT_OOM;
 }

 /*
Index: linux-2.6.10/kernel/sysctl.c
===================================================================
--- linux-2.6.10.orig/kernel/sysctl.c	2005-01-04 13:55:00.000000000 -0800
+++ linux-2.6.10/kernel/sysctl.c	2005-01-04 14:00:27.000000000 -0800
@@ -56,6 +56,7 @@
 extern int C_A_D;
 extern int sysctl_overcommit_memory;
 extern int sysctl_overcommit_ratio;
+extern int sysctl_max_prealloc_order;
 extern int max_threads;
 extern int sysrq_enabled;
 extern int core_uses_pid;
@@ -826,6 +827,16 @@
 		.strategy	= &sysctl_jiffies,
 	},
 #endif
+	{
+		.ctl_name	= VM_MAX_PREFAULT_ORDER,
+		.procname	= "max_prealloc_order",
+		.data		= &sysctl_max_prealloc_order,
+		.maxlen		= sizeof(sysctl_max_prealloc_order),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &zero,
+	},
 	{ .ctl_name = 0 }
 };

Index: linux-2.6.10/include/linux/sysctl.h
===================================================================
--- linux-2.6.10.orig/include/linux/sysctl.h	2005-01-04 13:55:00.000000000 -0800
+++ linux-2.6.10/include/linux/sysctl.h	2005-01-04 14:00:27.000000000 -0800
@@ -169,6 +169,7 @@
 	VM_VFS_CACHE_PRESSURE=26, /* dcache/icache reclaim pressure */
 	VM_LEGACY_VA_LAYOUT=27, /* legacy/compatibility virtual address space layout */
 	VM_SWAP_TOKEN_TIMEOUT=28, /* default time for token time out */
+	VM_MAX_PREFAULT_ORDER=29, /* max prefault order during anonymous page faults */
 };

Re: page fault scalability patch V14 [3/7]: i386 universal cmpxchg

[Roman Zippel]

Hi,

On Tuesday 04 January 2005 20:37, Christoph Lameter wrote:

>         * Provide emulation of cmpxchg suitable for uniprocessor if
>    build and run on 386.
>         * Provide emulation of cmpxchg8b suitable for uniprocessor
>    systems if build and run on 386 or 486.

I'm not sure that's such a good idea. This emulation is more expensive as it 
has to disable interrupts and you already have emulation functions using 
spinlocks anyway, so why not use them? This way your patch would not just 
scale up, but also still scale down.

bye, Roman

Re: page fault scalability patch V12 [0/7]: Overview and performance tests

[Jeff Garzik]

Andy Warner wrote:
> Jeff Garzik wrote:
> 
>>[...]
>>I am currently chasing a 2.6.8->2.6.9 SATA regression, which causes 
>>ata_piix (Intel ICH5/6/7) to not-find some SATA devices on x86-64 SMP, 
>>but works on UP.  Potentially related to >=4GB of RAM.
>>
>>
>>
>>Details, in case anyone is interested:
>>Unless my code is screwed up (certainly possible), PIO data-in [using 
>>the insw() call] seems to return all zeroes on a true-blue SMP machine, 
>>for the identify-device command.  When this happens, libata (correctly) 
>>detects a bad id page and bails.  (problem doesn't show up on single CPU 
>>w/ HT)
> 
> 
> Ah, I might have been here recently, with the pass-thru stuff.
> 
> What I saw was that in an SMP machine:
> 
> 1. queue_work() can result in the work running (on another
>    CPU) instantly.
> 
> 2. Having one CPU beat on PIO registers reading data from one port
>    would significantly alter the timing of the CMD->BSY->DRQ sequence
>    used in PIO. This behaviour was far worse for competing ports
>    within one chip, which I put down to arbitration problems.
> 
> 3. CPU utilisation would go through the roof. Effectively the
>    entire pio_task state machine reduced to a busy spin loop.
> 
> 4. The state machine needed some tweaks, especially in error
>    handling cases.
> 
> I made some changes, which effectively solved the problem for promise
> TX4-150 cards, and was going to test the results on other chipsets
> next week before speaking up. Specifically, I have seen some
> issues with SiI 3114 cards.
> 
> I was trying to explore using interrupts instead of polling state
> but for some reason, I was not getting them for PIO data operations,
> or I misunderstand the spec, after removing ata_qc_set_polling() - again
> I saw a difference in behaviour between the Promise & SiI cards
> here.
> 
> I'm about to go offline for 3 days, and hadn't prepared for this
> yet. The best I can do is provide a patch (attached) that applies
> against 2.6.9. It also seems to apply against libata-2.6, but
> barfs a bit against libata-dev-2.6.
> 
> The changes boil down to these:
> 
> 1. Minor changes in how status/error regs are read.
>    Including attempts to use altstatus, while I was
>    exploring interrupts.
> 
> 2. State machine logic changes.
> 
> 3. Replace calls to queue_work() with queue_delayed_work()
>    to stop SMP machines going crazy.
> 
> With these changes, on a platform consisting of 2.6.9 and
> Promise TX4-150 cards, I can move terabytes of parallel
> PIO data, without error.
> 
> My gut says that the PIO mechanism should be overhauled, I
> composed a "how much should we pay for this muffler" email
> to linux-ide at least twice while working on this, but never
> sent it - wanting to send a solution in rather than just
> making more comments from the peanut gallery.
> 
> I'll pick up the thread on this next week, when I'm back online.
> I hope this helps.

Please let me know if you still have problems?

The PIO SMP problems seem to be fixed here.

	Jeff

page table lock patch V15 [0/7]: overview

[Christoph Lameter]

Changes from V14->V15 of this patch:
- Remove misplaced semicolon in handle_mm_fault (caused x86_64 troubles)
- Fixed up and tested x86_64 arch specific patch
- Redone against 2.6.10-bk14

This is a series of patches that increases the scalability of
the page fault handler for SMP. The performance increase is
accomplished by avoiding the use of the page_table_lock spinlock
(but not mm->mmap_sem) through new atomic operations on pte's
(ptep_xchg, ptep_cmpxchg) and on pmd, pud and
pgd's (pgd_test_and_populate, pud_test_and_populate,
pmd_test_and_populate).

The page table lock can be avoided in the following situations:

1. An empty pte or pmd entry is populated

This is safe since the swapper may only depopulate them and the
swapper code has been changed to never set a pte to be empty until the
page has been evicted. The population of an empty pte is frequent
if a process touches newly allocated memory.

2. Modifications of flags in a pte entry (write/accessed).

These modifications are done by the CPU or by low level handlers
on various platforms also bypassing the page_table_lock. So this
seems to be safe too.

One essential change in the VM is the use of pte_cmpxchg (or its
generic emulation) on page table entries before doing an
update_mmu_change without holding the page table lock. However, we do
similar things now with other atomic pte operations such as
ptep_get_and_clear and ptep_test_and_clear_dirty. These operations
clear a pte *after* doing an operation on it. The ptep_cmpxchg as used
in this patch operates on an *cleared* pte and replaces it with a pte
pointing to valid memory. The effect of this change on various
architectures has to be thought through. Local definitions of
ptep_cmpxchg and ptep_xchg may be necessary.

For ia64 an icache coherency issue may arise that potentially requires
the flushing of the icache (as done via update_mmu_cache on ia64) prior
to the use of ptep_cmpxchg. Similar issues may arise on other platforms.

The patch introduces a split counter for rss handling to avoid atomic
operations and locks currently necessary for rss modifications. In
addition to mm->rss, tsk->rss is introduced. tsk->rss is defined to be
in the same cache line as tsk->mm (which is already used by the fault
handler) and thus tsk->rss can be incremented without locks
in a fast way. The cache line does not need to be shared between
processors for the page table handler.

A tasklist is generated for each mm (rcu based). Values in that list
are added up to calculate rss or anon_rss values.

The patchset is composed of 7 patches (and was tested against 2.6.10-bk6):

1/7: Avoid page_table_lock in handle_mm_fault

   This patch defers the acquisition of the page_table_lock as much as
   possible and uses atomic operations for allocating anonymous memory.
   These atomic operations are simulated by acquiring the page_table_lock
   for very small time frames if an architecture does not define
   __HAVE_ARCH_ATOMIC_TABLE_OPS. It also changes kswapd so that a
   pte will not be set to empty if a page is in transition to swap.

   If only the first two patches are applied then the time that the
   page_table_lock is held is simply reduced. The lock may then be
   acquired multiple times during a page fault.

2/7: Atomic pte operations for ia64

3/7: Make cmpxchg generally available on i386

   The atomic operations on the page table rely heavily on cmpxchg
   instructions. This patch adds emulations for cmpxchg and cmpxchg8b
   for old 80386 and 80486 cpus. The emulations are only included if a
   kernel is build for these old cpus and are skipped for the real
   cmpxchg instructions if the kernel that is build for 386 or 486 is
   then run on a more recent cpu.

   This patch may be used independently of the other patches.

4/7: Atomic pte operations for i386

   A generally available cmpxchg (last patch) must be available for
   this patch to preserve the ability to build kernels for 386 and 486.

5/7: Atomic pte operation for x86_64

6/7: Atomic pte operations for s390

7/7: Split counter implementation for rss
  Add tsk->rss and tsk->anon_rss. Add tasklist. Add logic
  to calculate rss from tasklist.

Signed-off-by: Christoph Lameter <clameter@sgi.com>

page table lock patch V15 [1/7]: Reduce use of page table lock

[Christoph Lameter]

Changelog
        * Increase parallelism in SMP configurations by deferring
          the acquisition of page_table_lock in handle_mm_fault
        * Anonymous memory page faults bypass the page_table_lock
          through the use of atomic page table operations
        * Swapper does not set pte to empty in transition to swap
        * Simulate atomic page table operations using the
          page_table_lock if an arch does not define
          __HAVE_ARCH_ATOMIC_TABLE_OPS. This still provides
          a performance benefit since the page_table_lock
          is held for shorter periods of time.

Signed-off-by: Christoph Lameter <clameter@sgi.com

Index: linux-2.6.10/mm/memory.c
===================================================================
--- linux-2.6.10.orig/mm/memory.c	2005-01-11 08:46:16.000000000 -0800
+++ linux-2.6.10/mm/memory.c	2005-01-11 09:16:30.000000000 -0800
@@ -36,6 +36,8 @@
  *		(Gerhard.Wichert@pdb.siemens.de)
  *
  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
+ * Jan 2005 	Scalability improvement by reducing the use and the length of time
+ *		the page table lock is held (Christoph Lameter)
  */

 #include <linux/kernel_stat.h>
@@ -1677,8 +1679,7 @@
 }

 /*
- * We hold the mm semaphore and the page_table_lock on entry and
- * should release the pagetable lock on exit..
+ * We hold the mm semaphore
  */
 static int do_swap_page(struct mm_struct * mm,
 	struct vm_area_struct * vma, unsigned long address,
@@ -1690,15 +1691,13 @@
 	int ret = VM_FAULT_MINOR;

 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);
 	page = lookup_swap_cache(entry);
 	if (!page) {
  		swapin_readahead(entry, address, vma);
  		page = read_swap_cache_async(entry, vma, address);
 		if (!page) {
 			/*
-			 * Back out if somebody else faulted in this pte while
-			 * we released the page table lock.
+			 * Back out if somebody else faulted in this pte
 			 */
 			spin_lock(&mm->page_table_lock);
 			page_table = pte_offset_map(pmd, address);
@@ -1721,8 +1720,7 @@
 	lock_page(page);

 	/*
-	 * Back out if somebody else faulted in this pte while we
-	 * released the page table lock.
+	 * Back out if somebody else faulted in this pte
 	 */
 	spin_lock(&mm->page_table_lock);
 	page_table = pte_offset_map(pmd, address);
@@ -1772,14 +1770,12 @@
 }

 /*
- * We are called with the MM semaphore and page_table_lock
- * spinlock held to protect against concurrent faults in
- * multithreaded programs.
+ * We are called with the MM semaphore held.
  */
 static int
 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 		pte_t *page_table, pmd_t *pmd, int write_access,
-		unsigned long addr)
+		unsigned long addr, pte_t orig_entry)
 {
 	pte_t entry;
 	struct page * page = ZERO_PAGE(addr);
@@ -1789,47 +1785,44 @@

 	/* ..except if it's a write access */
 	if (write_access) {
-		/* Allocate our own private page. */
-		pte_unmap(page_table);
-		spin_unlock(&mm->page_table_lock);

+		/* Allocate our own private page. */
 		if (unlikely(anon_vma_prepare(vma)))
-			goto no_mem;
+			return VM_FAULT_OOM;
+
 		page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 		if (!page)
-			goto no_mem;
+			return VM_FAULT_OOM;
 		clear_user_highpage(page, addr);

-		spin_lock(&mm->page_table_lock);
-		page_table = pte_offset_map(pmd, addr);
+		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
+							 vma->vm_page_prot)),
+				      vma);
+	}

-		if (!pte_none(*page_table)) {
+	if (!ptep_cmpxchg(vma, addr, page_table, orig_entry, entry)) {
+		if (write_access) {
 			pte_unmap(page_table);
 			page_cache_release(page);
-			spin_unlock(&mm->page_table_lock);
-			goto out;
 		}
+		return VM_FAULT_MINOR;
+	}
+	if (write_access) {
+		/*
+		 * These two functions must come after the cmpxchg
+		 * because if the page is on the LRU then try_to_unmap may come
+		 * in and unmap the pte.
+		 */
+		page_add_anon_rmap(page, vma, addr);
+		lru_cache_add_active(page);
 		mm->rss++;
 		acct_update_integrals();
 		update_mem_hiwater();
-		entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-							 vma->vm_page_prot)),
-				      vma);
-		lru_cache_add_active(page);
-		SetPageReferenced(page);
-		page_add_anon_rmap(page, vma, addr);
-	}

-	set_pte(page_table, entry);
+	}
 	pte_unmap(page_table);

-	/* No need to invalidate - it was non-present before */
-	update_mmu_cache(vma, addr, entry);
-	spin_unlock(&mm->page_table_lock);
-out:
 	return VM_FAULT_MINOR;
-no_mem:
-	return VM_FAULT_OOM;
 }

 /*
@@ -1841,12 +1834,12 @@
  * As this is called only for pages that do not currently exist, we
  * do not need to flush old virtual caches or the TLB.
  *
- * This is called with the MM semaphore held and the page table
- * spinlock held. Exit with the spinlock released.
+ * This is called with the MM semaphore held.
  */
 static int
 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-	unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *page_table,
+        pmd_t *pmd, pte_t orig_entry)
 {
 	struct page * new_page;
 	struct address_space *mapping = NULL;
@@ -1857,9 +1850,8 @@

 	if (!vma->vm_ops || !vma->vm_ops->nopage)
 		return do_anonymous_page(mm, vma, page_table,
-					pmd, write_access, address);
+					pmd, write_access, address, orig_entry);
 	pte_unmap(page_table);
-	spin_unlock(&mm->page_table_lock);

 	if (vma->vm_file) {
 		mapping = vma->vm_file->f_mapping;
@@ -1959,7 +1951,7 @@
  * nonlinear vmas.
  */
 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
-	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
+	unsigned long address, int write_access, pte_t *pte, pmd_t *pmd, pte_t entry)
 {
 	unsigned long pgoff;
 	int err;
@@ -1972,13 +1964,12 @@
 	if (!vma->vm_ops || !vma->vm_ops->populate ||
 			(write_access && !(vma->vm_flags & VM_SHARED))) {
 		pte_clear(pte);
-		return do_no_page(mm, vma, address, write_access, pte, pmd);
+		return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 	}

-	pgoff = pte_to_pgoff(*pte);
+	pgoff = pte_to_pgoff(entry);

 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);

 	err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
 	if (err == -ENOMEM)
@@ -1997,49 +1988,46 @@
  * with external mmu caches can use to update those (ie the Sparc or
  * PowerPC hashed page tables that act as extended TLBs).
  *
- * Note the "page_table_lock". It is to protect against kswapd removing
- * pages from under us. Note that kswapd only ever _removes_ pages, never
- * adds them. As such, once we have noticed that the page is not present,
- * we can drop the lock early.
- *
- * The adding of pages is protected by the MM semaphore (which we hold),
- * so we don't need to worry about a page being suddenly been added into
- * our VM.
- *
- * We enter with the pagetable spinlock held, we are supposed to
- * release it when done.
+ * Note that kswapd only ever _removes_ pages, never adds them.
+ * We need to insure to handle that case properly.
  */
 static inline int handle_pte_fault(struct mm_struct *mm,
 	struct vm_area_struct * vma, unsigned long address,
 	int write_access, pte_t *pte, pmd_t *pmd)
 {
 	pte_t entry;
+	pte_t new_entry;

-	entry = *pte;
+	/*
+	 * This must be an atomic operation since the page_table_lock is
+	 * not held. If a pte_t larger than the word size is used then an
+	 * incorrect value could be read because another processor is
+	 * concurrently updating the multi-word pte. The i386 PAE mode
+	 * is raising its ugly head here.
+	 */
+	entry = get_pte_atomic(pte);
 	if (!pte_present(entry)) {
-		/*
-		 * If it truly wasn't present, we know that kswapd
-		 * and the PTE updates will not touch it later. So
-		 * drop the lock.
-		 */
 		if (pte_none(entry))
-			return do_no_page(mm, vma, address, write_access, pte, pmd);
+			return do_no_page(mm, vma, address, write_access, pte, pmd, entry);
 		if (pte_file(entry))
-			return do_file_page(mm, vma, address, write_access, pte, pmd);
+			return do_file_page(mm, vma, address, write_access, pte, pmd, entry);
 		return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
 	}

+	/*
+	 * This is the case in which we only update some bits in the pte.
+	 */
+	new_entry = pte_mkyoung(entry);
 	if (write_access) {
-		if (!pte_write(entry))
+		if (!pte_write(entry)) {
+			/* do_wp_page expects us to hold the page_table_lock */
+			spin_lock(&mm->page_table_lock);
 			return do_wp_page(mm, vma, address, pte, pmd, entry);
-
-		entry = pte_mkdirty(entry);
+		}
+		new_entry = pte_mkdirty(new_entry);
 	}
-	entry = pte_mkyoung(entry);
-	ptep_set_access_flags(vma, address, pte, entry, write_access);
-	update_mmu_cache(vma, address, entry);
+	ptep_cmpxchg(vma, address, pte, entry, new_entry);
 	pte_unmap(pte);
-	spin_unlock(&mm->page_table_lock);
 	return VM_FAULT_MINOR;
 }

@@ -2058,33 +2046,55 @@

 	inc_page_state(pgfault);

-	if (is_vm_hugetlb_page(vma))
+	if (unlikely(is_vm_hugetlb_page(vma)))
 		return VM_FAULT_SIGBUS;	/* mapping truncation does this. */

 	/*
-	 * We need the page table lock to synchronize with kswapd
-	 * and the SMP-safe atomic PTE updates.
+	 * We rely on the mmap_sem and the SMP-safe atomic PTE updates.
+	 * to synchronize with kswapd. We can avoid the overhead
+	 * of the p??_alloc functions through atomic operations so
+	 * we duplicate the functionality of pmd_alloc, pud_alloc and
+	 * pte_alloc_map here.
 	 */
 	pgd = pgd_offset(mm, address);
-	spin_lock(&mm->page_table_lock);
+	if (unlikely(pgd_none(*pgd))) {
+		pud_t *new = pud_alloc_one(mm, address);
+
+		if (!new)
+			return VM_FAULT_OOM;

-	pud = pud_alloc(mm, pgd, address);
-	if (!pud)
-		goto oom;
-
-	pmd = pmd_alloc(mm, pud, address);
-	if (!pmd)
-		goto oom;
-
-	pte = pte_alloc_map(mm, pmd, address);
-	if (!pte)
-		goto oom;
+		if (!pgd_test_and_populate(mm, pgd, new))
+			pud_free(new);
+	}
+
+	pud = pud_offset(pgd, address);
+	if (unlikely(pud_none(*pud))) {
+		pmd_t *new = pmd_alloc_one(mm, address);
+
+		if (!new)
+			return VM_FAULT_OOM;
+
+		if (!pud_test_and_populate(mm, pud, new))
+			pmd_free(new);
+	}
+
+	pmd = pmd_offset(pud, address);
+	if (unlikely(!pmd_present(*pmd))) {
+		struct page *new = pte_alloc_one(mm, address);

-	return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
+		if (!new)
+			return VM_FAULT_OOM;

- oom:
-	spin_unlock(&mm->page_table_lock);
-	return VM_FAULT_OOM;
+		if (!pmd_test_and_populate(mm, pmd, new))
+			pte_free(new);
+		else {
+			inc_page_state(nr_page_table_pages);
+			mm->nr_ptes++;
+		}
+	}
+
+	pte = pte_offset_map(pmd, address);
+	return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
 }

 #ifndef __ARCH_HAS_4LEVEL_HACK
Index: linux-2.6.10/include/asm-generic/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-generic/pgtable.h	2004-12-24 13:34:30.000000000 -0800
+++ linux-2.6.10/include/asm-generic/pgtable.h	2005-01-11 08:48:30.000000000 -0800
@@ -28,6 +28,11 @@
 #endif /* __HAVE_ARCH_SET_PTE_ATOMIC */
 #endif

+/* Get a pte entry without the page table lock */
+#ifndef __HAVE_ARCH_GET_PTE_ATOMIC
+#define get_pte_atomic(__x)	*(__x)
+#endif
+
 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Largely same as above, but only sets the access flags (dirty,
@@ -134,4 +139,73 @@
 #define pgd_offset_gate(mm, addr)	pgd_offset(mm, addr)
 #endif

+#ifndef __HAVE_ARCH_ATOMIC_TABLE_OPS
+/*
+ * If atomic page table operations are not available then use
+ * the page_table_lock to insure some form of locking.
+ * Note thought that low level operations as well as the
+ * page_table_handling of the cpu may bypass all locking.
+ */
+
+#ifndef __HAVE_ARCH_PTEP_CMPXCHG
+#define ptep_cmpxchg(__vma, __addr, __ptep, __oldval, __newval)		\
+({									\
+	int __rc;							\
+	spin_lock(&__vma->vm_mm->page_table_lock);			\
+	__rc = pte_same(*(__ptep), __oldval);				\
+	if (__rc) { set_pte(__ptep, __newval);				\
+		update_mmu_cache(__vma, __addr, __newval); }		\
+	spin_unlock(&__vma->vm_mm->page_table_lock);			\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_ARCH_PGD_TEST_AND_POPULATE
+#define pgd_test_and_populate(__mm, __pgd, __pud)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pgd_present(*(__pgd));					\
+	if (__rc) pgd_populate(__mm, __pgd, __pud);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_ARCH_PUD_TEST_AND_POPULATE
+#define pud_test_and_populate(__mm, __pud, __pmd)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pud_present(*(__pud));					\
+	if (__rc) pud_populate(__mm, __pud, __pmd);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#ifndef __HAVE_PMD_TEST_AND_POPULATE
+#define pmd_test_and_populate(__mm, __pmd, __page)			\
+({									\
+	int __rc;							\
+	spin_lock(&__mm->page_table_lock);				\
+	__rc = !pmd_present(*(__pmd));					\
+	if (__rc) pmd_populate(__mm, __pmd, __page);			\
+	spin_unlock(&__mm->page_table_lock);				\
+	__rc;								\
+})
+#endif
+
+#endif
+
+#ifndef __HAVE_ARCH_PTEP_XCHG_FLUSH
+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)		\
+({									\
+	pte_t __p = __pte(xchg(&pte_val(*(__ptep)), pte_val(__pteval)));\
+	flush_tlb_page(__vma, __address);				\
+	__p;								\
+})
+
+#endif
+
 #endif /* _ASM_GENERIC_PGTABLE_H */
Index: linux-2.6.10/mm/rmap.c
===================================================================
--- linux-2.6.10.orig/mm/rmap.c	2005-01-11 08:46:16.000000000 -0800
+++ linux-2.6.10/mm/rmap.c	2005-01-11 08:48:30.000000000 -0800
@@ -426,7 +426,10 @@
  * @vma:	the vm area in which the mapping is added
  * @address:	the user virtual address mapped
  *
- * The caller needs to hold the mm->page_table_lock.
+ * The caller needs to hold the mm->page_table_lock if page
+ * is pointing to something that is known by the vm.
+ * The lock does not need to be held if page is pointing
+ * to a newly allocated page.
  */
 void page_add_anon_rmap(struct page *page,
 	struct vm_area_struct *vma, unsigned long address)
@@ -575,11 +578,6 @@

 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address);
-	pteval = ptep_clear_flush(vma, address, pte);
-
-	/* Move the dirty bit to the physical page now the pte is gone. */
-	if (pte_dirty(pteval))
-		set_page_dirty(page);

 	if (PageAnon(page)) {
 		swp_entry_t entry = { .val = page->private };
@@ -594,11 +592,15 @@
 			list_add(&mm->mmlist, &init_mm.mmlist);
 			spin_unlock(&mmlist_lock);
 		}
-		set_pte(pte, swp_entry_to_pte(entry));
+		pteval = ptep_xchg_flush(vma, address, pte, swp_entry_to_pte(entry));
 		BUG_ON(pte_file(*pte));
 		mm->anon_rss--;
-	}
+	} else
+		pteval = ptep_clear_flush(vma, address, pte);

+	/* Move the dirty bit to the physical page now the pte is gone. */
+	if (pte_dirty(pteval))
+		set_page_dirty(page);
 	mm->rss--;
 	acct_update_integrals();
 	page_remove_rmap(page);
@@ -691,15 +693,21 @@
 		if (ptep_clear_flush_young(vma, address, pte))
 			continue;

-		/* Nuke the page table entry. */
 		flush_cache_page(vma, address);
-		pteval = ptep_clear_flush(vma, address, pte);
+		/*
+		 * There would be a race here with handle_mm_fault and do_anonymous_page
+		 * which  bypasses the page_table_lock if we would zap the pte before
+		 * putting something into it. On the other hand we need to
+		 * have the dirty flag setting at the time we replaced the value.
+		 */

 		/* If nonlinear, store the file page offset in the pte. */
 		if (page->index != linear_page_index(vma, address))
-			set_pte(pte, pgoff_to_pte(page->index));
+			pteval = ptep_xchg_flush(vma, address, pte, pgoff_to_pte(page->index));
+		else
+			pteval = ptep_get_and_clear(pte);

-		/* Move the dirty bit to the physical page now the pte is gone. */
+		/* Move the dirty bit to the physical page now that the pte is gone. */
 		if (pte_dirty(pteval))
 			set_page_dirty(page);

Index: linux-2.6.10/include/asm-generic/pgtable-nopud.h
===================================================================
--- linux-2.6.10.orig/include/asm-generic/pgtable-nopud.h	2005-01-11 08:46:15.000000000 -0800
+++ linux-2.6.10/include/asm-generic/pgtable-nopud.h	2005-01-11 08:48:30.000000000 -0800
@@ -25,8 +25,13 @@
 static inline int pgd_present(pgd_t pgd)	{ return 1; }
 static inline void pgd_clear(pgd_t *pgd)	{ }
 #define pud_ERROR(pud)				(pgd_ERROR((pud).pgd))
-
 #define pgd_populate(mm, pgd, pud)		do { } while (0)
+
+static inline int pgd_test_and_populate(struct mm_struct *mm, pgd_t *pgd, pud_t *pud)
+{
+	return 1;
+}
+
 /*
  * (puds are folded into pgds so this doesn't get actually called,
  * but the define is needed for a generic inline function.)
Index: linux-2.6.10/include/asm-generic/pgtable-nopmd.h
===================================================================
--- linux-2.6.10.orig/include/asm-generic/pgtable-nopmd.h	2005-01-11 08:46:15.000000000 -0800
+++ linux-2.6.10/include/asm-generic/pgtable-nopmd.h	2005-01-11 08:48:30.000000000 -0800
@@ -29,6 +29,7 @@
 #define pmd_ERROR(pmd)				(pud_ERROR((pmd).pud))

 #define pud_populate(mm, pmd, pte)		do { } while (0)
+static inline int pud_test_and_populate(struct mm_struct *mm, pud_t *pud, pmd_t *pmd)	{ return 1; }

 /*
  * (pmds are folded into puds so this doesn't get actually called,
Index: linux-2.6.10/include/asm-ia64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-ia64/pgtable.h	2005-01-11 08:46:15.000000000 -0800
+++ linux-2.6.10/include/asm-ia64/pgtable.h	2005-01-11 08:48:30.000000000 -0800
@@ -561,7 +561,7 @@
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
 #define __HAVE_ARCH_PGD_OFFSET_GATE
-#include <asm-generic/pgtable.h>
 #include <asm-generic/pgtable-nopud.h>
+#include <asm-generic/pgtable.h>

 #endif /* _ASM_IA64_PGTABLE_H */

page table lock patch V15 [2/7]: ia64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for ia64
        * Enhanced parallelism in page fault handler if applied together
          with the generic patch

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-ia64/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-ia64/pgalloc.h	2005-01-10 16:31:56.000000000 -0800
+++ linux-2.6.10/include/asm-ia64/pgalloc.h	2005-01-10 16:41:00.000000000 -0800
@@ -34,6 +34,10 @@
 #define pmd_quicklist		(local_cpu_data->pmd_quick)
 #define pgtable_cache_size	(local_cpu_data->pgtable_cache_sz)

+/* Empty entries of PMD and PGD */
+#define PMD_NONE       0
+#define PUD_NONE       0
+
 static inline pgd_t*
 pgd_alloc_one_fast (struct mm_struct *mm)
 {
@@ -82,6 +86,13 @@
 	pud_val(*pud_entry) = __pa(pmd);
 }

+/* Atomic populate */
+static inline int
+pud_test_and_populate (struct mm_struct *mm, pud_t *pud_entry, pmd_t *pmd)
+{
+	return ia64_cmpxchg8_acq(pud_entry,__pa(pmd), PUD_NONE) == PUD_NONE;
+}
+
 static inline pmd_t*
 pmd_alloc_one_fast (struct mm_struct *mm, unsigned long addr)
 {
@@ -127,6 +138,13 @@
 	pmd_val(*pmd_entry) = page_to_phys(pte);
 }

+/* Atomic populate */
+static inline int
+pmd_test_and_populate (struct mm_struct *mm, pmd_t *pmd_entry, struct page *pte)
+{
+	return ia64_cmpxchg8_acq(pmd_entry, page_to_phys(pte), PMD_NONE) == PMD_NONE;
+}
+
 static inline void
 pmd_populate_kernel (struct mm_struct *mm, pmd_t *pmd_entry, pte_t *pte)
 {
Index: linux-2.6.10/include/asm-ia64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-ia64/pgtable.h	2005-01-10 16:32:35.000000000 -0800
+++ linux-2.6.10/include/asm-ia64/pgtable.h	2005-01-10 16:41:00.000000000 -0800
@@ -30,6 +30,8 @@
 #define _PAGE_P_BIT		0
 #define _PAGE_A_BIT		5
 #define _PAGE_D_BIT		6
+#define _PAGE_IG_BITS          53
+#define _PAGE_LOCK_BIT         (_PAGE_IG_BITS+3)       /* bit 56. Aligned to 8 bits */

 #define _PAGE_P			(1 << _PAGE_P_BIT)	/* page present bit */
 #define _PAGE_MA_WB		(0x0 <<  2)	/* write back memory attribute */
@@ -58,6 +60,7 @@
 #define _PAGE_PPN_MASK		(((__IA64_UL(1) << IA64_MAX_PHYS_BITS) - 1) & ~0xfffUL)
 #define _PAGE_ED		(__IA64_UL(1) << 52)	/* exception deferral */
 #define _PAGE_PROTNONE		(__IA64_UL(1) << 63)
+#define _PAGE_LOCK		(__IA64_UL(1) << _PAGE_LOCK_BIT)

 /* Valid only for a PTE with the present bit cleared: */
 #define _PAGE_FILE		(1 << 1)		/* see swap & file pte remarks below */
@@ -271,6 +274,8 @@
 #define pte_dirty(pte)		((pte_val(pte) & _PAGE_D) != 0)
 #define pte_young(pte)		((pte_val(pte) & _PAGE_A) != 0)
 #define pte_file(pte)		((pte_val(pte) & _PAGE_FILE) != 0)
+#define pte_locked(pte)		((pte_val(pte) & _PAGE_LOCK)!=0)
+
 /*
  * Note: we convert AR_RWX to AR_RX and AR_RW to AR_R by clearing the 2nd bit in the
  * access rights:
@@ -282,8 +287,15 @@
 #define pte_mkyoung(pte)	(__pte(pte_val(pte) | _PAGE_A))
 #define pte_mkclean(pte)	(__pte(pte_val(pte) & ~_PAGE_D))
 #define pte_mkdirty(pte)	(__pte(pte_val(pte) | _PAGE_D))
+#define pte_mkunlocked(pte)	(__pte(pte_val(pte) & ~_PAGE_LOCK))

 /*
+ * Lock functions for pte's
+ */
+#define ptep_lock(ptep)		test_and_set_bit(_PAGE_LOCK_BIT, ptep)
+#define ptep_unlock(ptep)	{ clear_bit(_PAGE_LOCK_BIT,ptep); smp_mb__after_clear_bit(); }
+#define ptep_unlock_set(ptep, val) set_pte(ptep, pte_mkunlocked(val))
+/*
  * Macro to a page protection value as "uncacheable".  Note that "protection" is really a
  * misnomer here as the protection value contains the memory attribute bits, dirty bits,
  * and various other bits as well.
@@ -343,7 +355,6 @@
 #define pte_unmap_nested(pte)		do { } while (0)

 /* atomic versions of the some PTE manipulations: */
-
 static inline int
 ptep_test_and_clear_young (pte_t *ptep)
 {
@@ -415,6 +426,26 @@
 #endif
 }

+/*
+ * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
+ * information.  However, we use this routine to take care of any (delayed) i-cache
+ * flushing that may be necessary.
+ */
+extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
+
+static inline int
+ptep_cmpxchg (struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	/*
+	 * IA64 defers icache flushes. If the new pte is executable we may
+	 * have to flush the icache to insure cache coherency immediately
+	 * after the cmpxchg.
+	 */
+	if (pte_exec(newval))
+		update_mmu_cache(vma, addr, newval);
+	return ia64_cmpxchg8_acq(&ptep->pte, newval.pte, oldval.pte) == oldval.pte;
+}
+
 static inline int
 pte_same (pte_t a, pte_t b)
 {
@@ -477,13 +508,6 @@
 	struct vm_area_struct * prev, unsigned long start, unsigned long end);
 #endif

-/*
- * IA-64 doesn't have any external MMU info: the page tables contain all the necessary
- * information.  However, we use this routine to take care of any (delayed) i-cache
- * flushing that may be necessary.
- */
-extern void update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte);
-
 #define __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 /*
  * Update PTEP with ENTRY, which is guaranteed to be a less
@@ -561,6 +585,8 @@
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
 #define __HAVE_ARCH_PGD_OFFSET_GATE
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+#define __HAVE_ARCH_LOCK_TABLE_OPS
 #include <asm-generic/pgtable-nopud.h>
 #include <asm-generic/pgtable.h>

page table lock patch V15 [3/7]: i386 universal cmpxchg

[Christoph Lameter]

Changelog
        * Make cmpxchg and cmpxchg8b generally available on the i386
	  platform.
        * Provide emulation of cmpxchg suitable for uniprocessor if
	  build and run on 386.
        * Provide emulation of cmpxchg8b suitable for uniprocessor
	  systems if build and run on 386 or 486.
	* Provide an inline function to atomically get a 64 bit value
	  via cmpxchg8b in an SMP system (courtesy of Nick Piggin)
	  (important for i386 PAE mode and other places where atomic
	  64 bit operations are useful)

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.9/arch/i386/Kconfig
===================================================================
--- linux-2.6.9.orig/arch/i386/Kconfig	2004-12-10 09:58:03.000000000 -0800
+++ linux-2.6.9/arch/i386/Kconfig	2004-12-10 09:59:27.000000000 -0800
@@ -351,6 +351,11 @@
 	depends on !M386
 	default y

+config X86_CMPXCHG8B
+	bool
+	depends on !M386 && !M486
+	default y
+
 config X86_XADD
 	bool
 	depends on !M386
Index: linux-2.6.9/arch/i386/kernel/cpu/intel.c
===================================================================
--- linux-2.6.9.orig/arch/i386/kernel/cpu/intel.c	2004-12-06 17:23:49.000000000 -0800
+++ linux-2.6.9/arch/i386/kernel/cpu/intel.c	2004-12-10 09:59:27.000000000 -0800
@@ -6,6 +6,7 @@
 #include <linux/bitops.h>
 #include <linux/smp.h>
 #include <linux/thread_info.h>
+#include <linux/module.h>

 #include <asm/processor.h>
 #include <asm/msr.h>
@@ -287,5 +288,103 @@
 	return 0;
 }

+#ifndef CONFIG_X86_CMPXCHG
+unsigned long cmpxchg_386_u8(volatile void *ptr, u8 old, u8 new)
+{
+	u8 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u8));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u8 *)ptr;
+	if (prev == old)
+		*(u8 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u8);
+
+unsigned long cmpxchg_386_u16(volatile void *ptr, u16 old, u16 new)
+{
+	u16 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u16));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u16 *)ptr;
+	if (prev == old)
+		*(u16 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u16);
+
+unsigned long cmpxchg_386_u32(volatile void *ptr, u32 old, u32 new)
+{
+	u32 prev;
+	unsigned long flags;
+	/*
+	 * Check if the kernel was compiled for an old cpu but the
+	 * currently running cpu can do cmpxchg after all
+	 * All CPUs except 386 support CMPXCHG
+	 */
+	if (cpu_data->x86 > 3)
+		return __cmpxchg(ptr, old, new, sizeof(u32));
+
+	/* Poor man's cmpxchg for 386. Unsuitable for SMP */
+	local_irq_save(flags);
+	prev = *(u32 *)ptr;
+	if (prev == old)
+		*(u32 *)ptr = new;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg_386_u32);
+#endif
+
+#ifndef CONFIG_X86_CMPXCHG8B
+unsigned long long cmpxchg8b_486(volatile unsigned long long *ptr,
+	       unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	unsigned long flags;
+
+	/*
+	 * Check if the kernel was compiled for an old cpu but
+	 * we are running really on a cpu capable of cmpxchg8b
+	 */
+
+	if (cpu_has(cpu_data, X86_FEATURE_CX8))
+		return __cmpxchg8b(ptr, old, newv);
+
+	/* Poor mans cmpxchg8b for 386 and 486. Not suitable for SMP */
+	local_irq_save(flags);
+	prev = *ptr;
+	if (prev == old)
+		*ptr = newv;
+	local_irq_restore(flags);
+	return prev;
+}
+
+EXPORT_SYMBOL(cmpxchg8b_486);
+#endif
+
 // arch_initcall(intel_cpu_init);

Index: linux-2.6.9/include/asm-i386/system.h
===================================================================
--- linux-2.6.9.orig/include/asm-i386/system.h	2004-12-06 17:23:55.000000000 -0800
+++ linux-2.6.9/include/asm-i386/system.h	2004-12-10 10:00:49.000000000 -0800
@@ -149,6 +149,9 @@
 #define __xg(x) ((struct __xchg_dummy *)(x))


+#define ll_low(x)	*(((unsigned int*)&(x))+0)
+#define ll_high(x)	*(((unsigned int*)&(x))+1)
+
 /*
  * The semantics of XCHGCMP8B are a bit strange, this is why
  * there is a loop and the loading of %%eax and %%edx has to
@@ -184,8 +187,6 @@
 {
 	__set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
 }
-#define ll_low(x)	*(((unsigned int*)&(x))+0)
-#define ll_high(x)	*(((unsigned int*)&(x))+1)

 static inline void __set_64bit_var (unsigned long long *ptr,
 			 unsigned long long value)
@@ -203,6 +204,26 @@
  __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
  __set_64bit(ptr, ll_low(value), ll_high(value)) )

+static inline unsigned long long __get_64bit(unsigned long long * ptr)
+{
+	unsigned long long ret;
+	__asm__ __volatile__ (
+		"\n1:\t"
+		"movl (%1), %%eax\n\t"
+		"movl 4(%1), %%edx\n\t"
+		"movl %%eax, %%ebx\n\t"
+		"movl %%edx, %%ecx\n\t"
+		LOCK_PREFIX "cmpxchg8b (%1)\n\t"
+		"jnz 1b"
+		:	"=A"(ret)
+		:	"D"(ptr)
+		:	"ebx", "ecx", "memory");
+	return ret;
+}
+
+#define get_64bit(ptr) __get_64bit(ptr)
+
+
 /*
  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
  * Note 2: xchg has side effect, so that attribute volatile is necessary,
@@ -240,7 +261,41 @@
  */

 #ifdef CONFIG_X86_CMPXCHG
+
 #define __HAVE_ARCH_CMPXCHG 1
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))__cmpxchg((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
+
+#else
+
+/*
+ * Building a kernel capable running on 80386. It may be necessary to
+ * simulate the cmpxchg on the 80386 CPU. For that purpose we define
+ * a function for each of the sizes we support.
+ */
+
+extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
+extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
+extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
+
+static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
+				      unsigned long new, int size)
+{
+	switch (size) {
+	case 1:
+		return cmpxchg_386_u8(ptr, old, new);
+	case 2:
+		return cmpxchg_386_u16(ptr, old, new);
+	case 4:
+		return cmpxchg_386_u32(ptr, old, new);
+	}
+	return old;
+}
+
+#define cmpxchg(ptr,o,n)\
+	((__typeof__(*(ptr)))cmpxchg_386((ptr), (unsigned long)(o), \
+					(unsigned long)(n), sizeof(*(ptr))))
 #endif

 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
@@ -270,12 +325,34 @@
 	return old;
 }

-#define cmpxchg(ptr,o,n)\
-	((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
-					(unsigned long)(n),sizeof(*(ptr))))
-
+static inline unsigned long long __cmpxchg8b(volatile unsigned long long *ptr,
+		unsigned long long old, unsigned long long newv)
+{
+	unsigned long long prev;
+	__asm__ __volatile__(
+	LOCK_PREFIX "cmpxchg8b (%4)"
+		: "=A" (prev)
+		: "0" (old), "c" ((unsigned long)(newv >> 32)),
+		  "b" ((unsigned long)(newv & 0xffffffffULL)), "D" (ptr)
+		: "memory");
+	return prev;
+}
+
+#ifdef CONFIG_X86_CMPXCHG8B
+#define cmpxchg8b __cmpxchg8b
+#else
+/*
+ * Building a kernel capable of running on 80486 and 80386. Both
+ * do not support cmpxchg8b. Call a function that emulates the
+ * instruction if necessary.
+ */
+extern unsigned long long cmpxchg8b_486(volatile unsigned long long *,
+		unsigned long long, unsigned long long);
+#define cmpxchg8b cmpxchg8b_486
+#endif
+
 #ifdef __KERNEL__
-struct alt_instr {
+struct alt_instr {
 	__u8 *instr; 		/* original instruction */
 	__u8 *replacement;
 	__u8  cpuid;		/* cpuid bit set for replacement */

page table lock patch V15 [4/7]: i386 atomic pte operations

[Christoph Lameter]

Changelog
	* Atomic pte operations for i386 in regular and PAE modes

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-i386/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgtable.h	2005-01-07 09:48:57.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgtable.h	2005-01-07 09:51:09.000000000 -0800
@@ -409,6 +409,7 @@
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
 #include <asm-generic/pgtable.h>

 #endif /* _I386_PGTABLE_H */
Index: linux-2.6.10/include/asm-i386/pgtable-3level.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgtable-3level.h	2005-01-07 09:48:57.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgtable-3level.h	2005-01-07 09:51:09.000000000 -0800
@@ -8,7 +8,8 @@
  * tables on PPro+ CPUs.
  *
  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
- */
+ * August 26, 2004 added ptep_cmpxchg <christoph@lameter.com>
+*/

 #define pte_ERROR(e) \
 	printk("%s:%d: bad pte %p(%08lx%08lx).\n", __FILE__, __LINE__, &(e), (e).pte_high, (e).pte_low)
@@ -44,21 +45,11 @@
 	return pte_x(pte);
 }

-/* Rules for using set_pte: the pte being assigned *must* be
- * either not present or in a state where the hardware will
- * not attempt to update the pte.  In places where this is
- * not possible, use pte_get_and_clear to obtain the old pte
- * value and then use set_pte to update it.  -ben
- */
-static inline void set_pte(pte_t *ptep, pte_t pte)
-{
-	ptep->pte_high = pte.pte_high;
-	smp_wmb();
-	ptep->pte_low = pte.pte_low;
-}
 #define __HAVE_ARCH_SET_PTE_ATOMIC
 #define set_pte_atomic(pteptr,pteval) \
 		set_64bit((unsigned long long *)(pteptr),pte_val(pteval))
+#define set_pte(pteptr,pteval) \
+		*(unsigned long long *)(pteptr) = pte_val(pteval)
 #define set_pmd(pmdptr,pmdval) \
 		set_64bit((unsigned long long *)(pmdptr),pmd_val(pmdval))
 #define set_pud(pudptr,pudval) \
@@ -155,4 +146,25 @@

 #define __pmd_free_tlb(tlb, x)		do { } while (0)

+/* Atomic PTE operations */
+#define ptep_xchg_flush(__vma, __addr, __ptep, __newval) \
+({	pte_t __r;							\
+	/* xchg acts as a barrier before the setting of the high bits. */\
+	__r.pte_low = xchg(&(__ptep)->pte_low, (__newval).pte_low);	\
+	__r.pte_high = (__ptep)->pte_high;				\
+	(__ptep)->pte_high = (__newval).pte_high;			\
+	flush_tlb_page(__vma, __addr);					\
+	(__r);								\
+})
+
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
+
+static inline int ptep_cmpxchg(struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg8b((unsigned long long *)ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
+#define __HAVE_ARCH_GET_PTE_ATOMIC
+#define get_pte_atomic(__ptep) __pte(get_64bit((unsigned long long *)(__ptep)))
+
 #endif /* _I386_PGTABLE_3LEVEL_H */
Index: linux-2.6.10/include/asm-i386/pgtable-2level.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgtable-2level.h	2005-01-07 09:48:57.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgtable-2level.h	2005-01-07 09:51:09.000000000 -0800
@@ -65,4 +65,7 @@
 #define __pte_to_swp_entry(pte)		((swp_entry_t) { (pte).pte_low })
 #define __swp_entry_to_pte(x)		((pte_t) { (x).val })

+/* Atomic PTE operations */
+#define ptep_cmpxchg(__vma,__a,__xp,__oldpte,__newpte) (cmpxchg(&(__xp)->pte_low, (__oldpte).pte_low, (__newpte).pte_low)==(__oldpte).pte_low)
+
 #endif /* _I386_PGTABLE_2LEVEL_H */
Index: linux-2.6.10/include/asm-i386/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-i386/pgalloc.h	2005-01-07 09:48:57.000000000 -0800
+++ linux-2.6.10/include/asm-i386/pgalloc.h	2005-01-07 11:15:55.000000000 -0800
@@ -4,9 +4,12 @@
 #include <linux/config.h>
 #include <asm/processor.h>
 #include <asm/fixmap.h>
+#include <asm/system.h>
 #include <linux/threads.h>
 #include <linux/mm.h>		/* for struct page */

+#define PMD_NONE 0L
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte)))

@@ -14,6 +17,18 @@
 	set_pmd(pmd, __pmd(_PAGE_TABLE +			\
 		((unsigned long long)page_to_pfn(pte) <<	\
 			(unsigned long long) PAGE_SHIFT)))
+/* Atomic version */
+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+#ifdef CONFIG_X86_PAE
+	return cmpxchg8b( ((unsigned long long *)pmd), PMD_NONE, _PAGE_TABLE +
+		((unsigned long long)page_to_pfn(pte) <<
+			(unsigned long long) PAGE_SHIFT) ) == PMD_NONE;
+#else
+	return cmpxchg( (unsigned long *)pmd, PMD_NONE, _PAGE_TABLE + (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
+#endif
+}
+
 /*
  * Allocate and free page tables.
  */
@@ -44,6 +59,7 @@
 #define pmd_free(x)			do { } while (0)
 #define __pmd_free_tlb(tlb,x)		do { } while (0)
 #define pud_populate(mm, pmd, pte)	BUG()
+#define pud_test_and_populate(mm, pmd, pte) 	({ BUG(); 1; })
 #endif

 #define check_pgt_cache()	do { } while (0)

page table lock patch V15 [5/7]: x86_64 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for x86_64

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-x86_64/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgalloc.h	2005-01-10 16:31:56.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgalloc.h	2005-01-10 16:41:24.000000000 -0800
@@ -7,6 +7,10 @@
 #include <linux/threads.h>
 #include <linux/mm.h>

+#define PMD_NONE 0
+#define PUD_NONE 0
+#define PGD_NONE 0
+
 #define pmd_populate_kernel(mm, pmd, pte) \
 		set_pmd(pmd, __pmd(_PAGE_TABLE | __pa(pte)))
 #define pud_populate(mm, pud, pmd) \
@@ -14,9 +18,20 @@
 #define pgd_populate(mm, pgd, pud) \
 		set_pgd(pgd, __pgd(_PAGE_TABLE | __pa(pud)))

+#define pud_test_and_populate(mm, pud, pmd) \
+		(cmpxchg((unsigned long *)pud, PUD_NONE, _PAGE_TABLE | __pa(pmd)) == PUD_NONE)
+#define pgd_test_and_populate(mm, pgd, pud) \
+		(cmpxchg((unsigned long *)pgd, PGD_NONE, _PAGE_TABLE | __pa(pud)) == PGD_NONE)
+
+
 static inline void pmd_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
 {
-	set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
+       set_pmd(pmd, __pmd(_PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)));
+}
+
+static inline int pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *pte)
+{
+	return cmpxchg((unsigned long *)pmd, PMD_NONE, _PAGE_TABLE | (page_to_pfn(pte) << PAGE_SHIFT)) == PMD_NONE;
 }

 extern __inline__ pmd_t *get_pmd(void)
Index: linux-2.6.10/include/asm-x86_64/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-x86_64/pgtable.h	2005-01-10 16:31:56.000000000 -0800
+++ linux-2.6.10/include/asm-x86_64/pgtable.h	2005-01-10 16:41:24.000000000 -0800
@@ -414,6 +414,10 @@
 #define	kc_offset_to_vaddr(o) \
    (((o) & (1UL << (__VIRTUAL_MASK_SHIFT-1))) ? ((o) | (~__VIRTUAL_MASK)) : (o))

+
+#define ptep_cmpxchg(__vma,__addr,__xp,__oldval,__newval) (cmpxchg(&(__xp)->pte, pte_val(__oldval), pte_val(__newval)) == pte_val(__oldval))
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_DIRTY
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR

page table lock patch V15 [6/7]: s390 atomic pte operations

[Christoph Lameter]

Changelog
        * Provide atomic pte operations for s390

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/asm-s390/pgtable.h
===================================================================
--- linux-2.6.10.orig/include/asm-s390/pgtable.h	2005-01-10 16:31:56.000000000 -0800
+++ linux-2.6.10/include/asm-s390/pgtable.h	2005-01-10 16:41:07.000000000 -0800
@@ -577,6 +577,15 @@
 	return pte;
 }

+#define ptep_xchg_flush(__vma, __address, __ptep, __pteval)            \
+({                                                                     \
+	struct mm_struct *__mm = __vma->vm_mm;                          \
+	pte_t __pte;                                                    \
+	__pte = ptep_clear_flush(__vma, __address, __ptep);             \
+	set_pte(__ptep, __pteval);                                      \
+	__pte;                                                          \
+})
+
 static inline void ptep_set_wrprotect(pte_t *ptep)
 {
 	pte_t old_pte = *ptep;
@@ -788,6 +797,14 @@

 #define kern_addr_valid(addr)   (1)

+/* Atomic PTE operations */
+#define __HAVE_ARCH_ATOMIC_TABLE_OPS
+
+static inline int ptep_cmpxchg (struct vm_area_struct *vma, unsigned long address, pte_t *ptep, pte_t oldval, pte_t newval)
+{
+	return cmpxchg(ptep, pte_val(oldval), pte_val(newval)) == pte_val(oldval);
+}
+
 /*
  * No page table caches to initialise
  */
@@ -801,6 +818,7 @@
 #define __HAVE_ARCH_PTEP_CLEAR_DIRTY_FLUSH
 #define __HAVE_ARCH_PTEP_GET_AND_CLEAR
 #define __HAVE_ARCH_PTEP_CLEAR_FLUSH
+#define __HAVE_ARCH_PTEP_XCHG_FLUSH
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
 #define __HAVE_ARCH_PTEP_MKDIRTY
 #define __HAVE_ARCH_PTE_SAME
Index: linux-2.6.10/include/asm-s390/pgalloc.h
===================================================================
--- linux-2.6.10.orig/include/asm-s390/pgalloc.h	2004-12-24 13:35:00.000000000 -0800
+++ linux-2.6.10/include/asm-s390/pgalloc.h	2005-01-10 16:41:07.000000000 -0800
@@ -97,6 +97,10 @@
 	pgd_val(*pgd) = _PGD_ENTRY | __pa(pmd);
 }

+static inline int pgd_test_and_populate(struct mm_struct *mm, pdg_t *pgd, pmd_t *pmd)
+{
+	return cmpxchg(pgd, _PAGE_TABLE_INV, _PGD_ENTRY | __pa(pmd)) == _PAGE_TABLE_INV;
+}
 #endif /* __s390x__ */

 static inline void
@@ -119,6 +123,18 @@
 	pmd_populate_kernel(mm, pmd, (pte_t *)((page-mem_map) << PAGE_SHIFT));
 }

+static inline int
+pmd_test_and_populate(struct mm_struct *mm, pmd_t *pmd, struct page *page)
+{
+	int rc;
+	spin_lock(&mm->page_table_lock);
+
+	rc=pte_same(*pmd, _PAGE_INVALID_EMPTY);
+	if (rc) pmd_populate(mm, pmd, page);
+	spin_unlock(&mm->page_table_lock);
+	return rc;
+}
+
 /*
  * page table entry allocation/free routines.
  */

page table lock patch V15 [7/7]: Split RSS counter

[Christoph Lameter]

Changelog
	* Split rss counter into the task structure
	* remove 3 checks of rss in mm/rmap.c
	* increment current->rss instead of mm->rss in the page fault handler
	* move incrementing of anon_rss out of page_add_anon_rmap to group
	  the increments more tightly and allow a better cache utilization

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.10/include/linux/sched.h
===================================================================
--- linux-2.6.10.orig/include/linux/sched.h	2005-01-11 08:46:16.000000000 -0800
+++ linux-2.6.10/include/linux/sched.h	2005-01-11 08:56:45.000000000 -0800
@@ -31,6 +31,7 @@
 #include <linux/pid.h>
 #include <linux/percpu.h>
 #include <linux/topology.h>
+#include <linux/rcupdate.h>

 struct exec_domain;

@@ -216,6 +217,7 @@
 	int map_count;				/* number of VMAs */
 	struct rw_semaphore mmap_sem;
 	spinlock_t page_table_lock;		/* Protects page tables, mm->rss, mm->anon_rss */
+	long rss, anon_rss;

 	struct list_head mmlist;		/* List of maybe swapped mm's.  These are globally strung
 						 * together off init_mm.mmlist, and are protected
@@ -225,7 +227,7 @@
 	unsigned long start_code, end_code, start_data, end_data;
 	unsigned long start_brk, brk, start_stack;
 	unsigned long arg_start, arg_end, env_start, env_end;
-	unsigned long rss, anon_rss, total_vm, locked_vm, shared_vm;
+	unsigned long total_vm, locked_vm, shared_vm;
 	unsigned long exec_vm, stack_vm, reserved_vm, def_flags, nr_ptes;

 	unsigned long saved_auxv[42]; /* for /proc/PID/auxv */
@@ -235,6 +237,7 @@

 	/* Architecture-specific MM context */
 	mm_context_t context;
+	struct list_head task_list;             /* Tasks using this mm */

 	/* Token based thrashing protection. */
 	unsigned long swap_token_time;
@@ -555,6 +558,9 @@
 	struct list_head ptrace_list;

 	struct mm_struct *mm, *active_mm;
+	/* Split counters from mm */
+	long rss;
+	long anon_rss;

 /* task state */
 	struct linux_binfmt *binfmt;
@@ -587,6 +593,10 @@
 	struct completion *vfork_done;		/* for vfork() */
 	int __user *set_child_tid;		/* CLONE_CHILD_SETTID */
 	int __user *clear_child_tid;		/* CLONE_CHILD_CLEARTID */
+
+	/* List of other tasks using the same mm */
+	struct list_head mm_tasks;
+	struct rcu_head rcu_head;               /* For freeing the task via rcu */

 	unsigned long rt_priority;
 	unsigned long it_real_value, it_prof_value, it_virt_value;
@@ -1184,6 +1194,11 @@
 	return 0;
 }
 #endif /* CONFIG_PM */
+
+void get_rss(struct mm_struct *mm, unsigned long *rss, unsigned long *anon_rss);
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk);
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk);
+
 #endif /* __KERNEL__ */

 #endif
Index: linux-2.6.10/fs/proc/task_mmu.c
===================================================================
--- linux-2.6.10.orig/fs/proc/task_mmu.c	2005-01-11 08:46:15.000000000 -0800
+++ linux-2.6.10/fs/proc/task_mmu.c	2005-01-11 08:56:45.000000000 -0800
@@ -8,8 +8,9 @@

 char *task_mem(struct mm_struct *mm, char *buffer)
 {
-	unsigned long data, text, lib;
+	unsigned long data, text, lib, rss, anon_rss;

+	get_rss(mm, &rss, &anon_rss);
 	data = mm->total_vm - mm->shared_vm - mm->stack_vm;
 	text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK)) >> 10;
 	lib = (mm->exec_vm << (PAGE_SHIFT-10)) - text;
@@ -24,7 +25,7 @@
 		"VmPTE:\t%8lu kB\n",
 		(mm->total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),
-		mm->rss << (PAGE_SHIFT-10),
+		rss << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
 		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
@@ -39,11 +40,14 @@
 int task_statm(struct mm_struct *mm, int *shared, int *text,
 	       int *data, int *resident)
 {
-	*shared = mm->rss - mm->anon_rss;
+	unsigned long rss, anon_rss;
+
+	get_rss(mm, &rss, &anon_rss);
+	*shared = rss - anon_rss;
 	*text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK))
 								>> PAGE_SHIFT;
 	*data = mm->total_vm - mm->shared_vm;
-	*resident = mm->rss;
+	*resident = rss;
 	return mm->total_vm;
 }

Index: linux-2.6.10/fs/proc/array.c
===================================================================
--- linux-2.6.10.orig/fs/proc/array.c	2005-01-11 08:46:15.000000000 -0800
+++ linux-2.6.10/fs/proc/array.c	2005-01-11 08:56:45.000000000 -0800
@@ -303,7 +303,7 @@

 static int do_task_stat(struct task_struct *task, char * buffer, int whole)
 {
-	unsigned long vsize, eip, esp, wchan = ~0UL;
+	unsigned long rss, anon_rss, vsize, eip, esp, wchan = ~0UL;
 	long priority, nice;
 	int tty_pgrp = -1, tty_nr = 0;
 	sigset_t sigign, sigcatch;
@@ -326,6 +326,7 @@
 		vsize = task_vsize(mm);
 		eip = KSTK_EIP(task);
 		esp = KSTK_ESP(task);
+		get_rss(mm, &rss, &anon_rss);
 	}

 	get_task_comm(tcomm, task);
@@ -421,7 +422,7 @@
 		jiffies_to_clock_t(task->it_real_value),
 		start_time,
 		vsize,
-		mm ? mm->rss : 0, /* you might want to shift this left 3 */
+		mm ? rss : 0, /* you might want to shift this left 3 */
 	        rsslim,
 		mm ? mm->start_code : 0,
 		mm ? mm->end_code : 0,
Index: linux-2.6.10/mm/rmap.c
===================================================================
--- linux-2.6.10.orig/mm/rmap.c	2005-01-11 08:48:30.000000000 -0800
+++ linux-2.6.10/mm/rmap.c	2005-01-11 08:56:45.000000000 -0800
@@ -258,8 +258,6 @@
 	pte_t *pte;
 	int referenced = 0;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -440,8 +438,6 @@
 	BUG_ON(PageReserved(page));
 	BUG_ON(!anon_vma);

-	vma->vm_mm->anon_rss++;
-
 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 	index = (address - vma->vm_start) >> PAGE_SHIFT;
 	index += vma->vm_pgoff;
@@ -513,8 +509,6 @@
 	pte_t pteval;
 	int ret = SWAP_AGAIN;

-	if (!mm->rss)
-		goto out;
 	address = vma_address(page, vma);
 	if (address == -EFAULT)
 		goto out;
@@ -813,8 +807,7 @@
 			if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
 				continue;
 			cursor = (unsigned long) vma->vm_private_data;
-			while (vma->vm_mm->rss &&
-				cursor < max_nl_cursor &&
+			while (cursor < max_nl_cursor &&
 				cursor < vma->vm_end - vma->vm_start) {
 				try_to_unmap_cluster(cursor, &mapcount, vma);
 				cursor += CLUSTER_SIZE;
Index: linux-2.6.10/kernel/fork.c
===================================================================
--- linux-2.6.10.orig/kernel/fork.c	2005-01-11 08:46:16.000000000 -0800
+++ linux-2.6.10/kernel/fork.c	2005-01-11 08:56:45.000000000 -0800
@@ -79,10 +79,16 @@
 static kmem_cache_t *task_struct_cachep;
 #endif

+static void rcu_free_task(struct rcu_head *head)
+{
+	struct task_struct *tsk = container_of(head ,struct task_struct, rcu_head);
+	free_task_struct(tsk);
+}
+
 void free_task(struct task_struct *tsk)
 {
 	free_thread_info(tsk->thread_info);
-	free_task_struct(tsk);
+	call_rcu(&tsk->rcu_head, rcu_free_task);
 }
 EXPORT_SYMBOL(free_task);

@@ -99,7 +105,7 @@
 	put_group_info(tsk->group_info);

 	if (!profile_handoff_task(tsk))
-		free_task(tsk);
+		call_rcu(&tsk->rcu_head, rcu_free_task);
 }

 void __init fork_init(unsigned long mempages)
@@ -152,6 +158,7 @@
 	*tsk = *orig;
 	tsk->thread_info = ti;
 	ti->task = tsk;
+	tsk->rss = 0;

 	/* One for us, one for whoever does the "release_task()" (usually parent) */
 	atomic_set(&tsk->usage,2);
@@ -294,6 +301,7 @@
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);
 	INIT_LIST_HEAD(&mm->mmlist);
+	INIT_LIST_HEAD(&mm->task_list);
 	mm->core_waiters = 0;
 	mm->nr_ptes = 0;
 	spin_lock_init(&mm->page_table_lock);
@@ -402,6 +410,8 @@

 	/* Get rid of any cached register state */
 	deactivate_mm(tsk, mm);
+	if (mm)
+		mm_remove_thread(mm, tsk);

 	/* notify parent sleeping on vfork() */
 	if (vfork_done) {
@@ -449,8 +459,8 @@
 		 * new threads start up in user mode using an mm, which
 		 * allows optimizing out ipis; the tlb_gather_mmu code
 		 * is an example.
+		 * (mm_add_thread does use the ptl .... )
 		 */
-		spin_unlock_wait(&oldmm->page_table_lock);
 		goto good_mm;
 	}

@@ -475,6 +485,7 @@
 	mm->hiwater_vm = mm->total_vm;

 good_mm:
+	mm_add_thread(mm, tsk);
 	tsk->mm = mm;
 	tsk->active_mm = mm;
 	return 0;
@@ -1079,7 +1090,7 @@
 	atomic_dec(&p->user->processes);
 	free_uid(p->user);
 bad_fork_free:
-	free_task(p);
+	call_rcu(&p->rcu_head, rcu_free_task);
 	goto fork_out;
 }

Index: linux-2.6.10/mm/memory.c
===================================================================
--- linux-2.6.10.orig/mm/memory.c	2005-01-11 08:56:37.000000000 -0800
+++ linux-2.6.10/mm/memory.c	2005-01-11 08:56:45.000000000 -0800
@@ -935,6 +935,7 @@

 			cond_resched_lock(&mm->page_table_lock);
 			while (!(map = follow_page(mm, start, lookup_write))) {
+				unsigned long rss, anon_rss;
 				/*
 				 * Shortcut for anonymous pages. We don't want
 				 * to force the creation of pages tables for
@@ -947,6 +948,17 @@
 					map = ZERO_PAGE(start);
 					break;
 				}
+				if (mm != current->mm) {
+					/*
+					 * handle_mm_fault uses the current pointer
+					 * for a split rss counter. The current pointer
+					 * is not correct if we are using a different mm
+					 */
+					rss = current->rss;
+					anon_rss = current->anon_rss;
+					current->rss = 0;
+					current->anon_rss = 0;
+				}
 				spin_unlock(&mm->page_table_lock);
 				switch (handle_mm_fault(mm,vma,start,write)) {
 				case VM_FAULT_MINOR:
@@ -971,6 +983,12 @@
 				 */
 				lookup_write = write && !force;
 				spin_lock(&mm->page_table_lock);
+				if (mm != current->mm) {
+					mm->rss += current->rss;
+					mm->anon_rss += current->anon_rss;
+					current->rss = rss;
+					current->anon_rss = anon_rss;
+				}
 			}
 			if (pages) {
 				pages[i] = get_page_map(map);
@@ -1353,6 +1371,7 @@
 		break_cow(vma, new_page, address, page_table);
 		lru_cache_add_active(new_page);
 		page_add_anon_rmap(new_page, vma, address);
+		mm->anon_rss++;

 		/* Free the old page.. */
 		new_page = old_page;
@@ -1753,6 +1772,7 @@
 	flush_icache_page(vma, page);
 	set_pte(page_table, pte);
 	page_add_anon_rmap(page, vma, address);
+	mm->anon_rss++;

 	if (write_access) {
 		if (do_wp_page(mm, vma, address,
@@ -1815,6 +1835,7 @@
 		page_add_anon_rmap(page, vma, addr);
 		lru_cache_add_active(page);
 		mm->rss++;
+		mm->anon_rss++;
 		acct_update_integrals();
 		update_mem_hiwater();

@@ -1922,6 +1943,7 @@
 		if (anon) {
 			lru_cache_add_active(new_page);
 			page_add_anon_rmap(new_page, vma, address);
+			mm->anon_rss++;
 		} else
 			page_add_file_rmap(new_page);
 		pte_unmap(page_table);
@@ -2250,6 +2272,49 @@

 EXPORT_SYMBOL(vmalloc_to_pfn);

+void get_rss(struct mm_struct *mm, unsigned long *rss, unsigned long *anon_rss)
+{
+	struct list_head *y;
+	struct task_struct *t;
+	long rss_sum, anon_rss_sum;
+
+	rcu_read_lock();
+	rss_sum = mm->rss;
+	anon_rss_sum = mm->anon_rss;
+	list_for_each_rcu(y, &mm->task_list) {
+		t = list_entry(y, struct task_struct, mm_tasks);
+		rss_sum += t->rss;
+		anon_rss_sum += t->anon_rss;
+	}
+	if (rss_sum < 0)
+		rss_sum = 0;
+	if (anon_rss_sum < 0)
+		anon_rss_sum = 0;
+	rcu_read_unlock();
+	*rss = rss_sum;
+	*anon_rss = anon_rss_sum;
+}
+
+void mm_remove_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	if (!mm)
+		return;
+
+	spin_lock(&mm->page_table_lock);
+	mm->rss += tsk->rss;
+	mm->anon_rss += tsk->anon_rss;
+	list_del_rcu(&tsk->mm_tasks);
+	spin_unlock(&mm->page_table_lock);
+}
+
+void mm_add_thread(struct mm_struct *mm, struct task_struct *tsk)
+{
+	spin_lock(&mm->page_table_lock);
+	tsk->rss = 0;
+	tsk->anon_rss = 0;
+	list_add_rcu(&tsk->mm_tasks, &mm->task_list);
+	spin_unlock(&mm->page_table_lock);
+}
 /*
  * update_mem_hiwater
  *	- update per process rss and vm high water data
Index: linux-2.6.10/include/linux/init_task.h
===================================================================
--- linux-2.6.10.orig/include/linux/init_task.h	2005-01-11 08:46:16.000000000 -0800
+++ linux-2.6.10/include/linux/init_task.h	2005-01-11 08:56:45.000000000 -0800
@@ -42,6 +42,7 @@
 	.mmlist		= LIST_HEAD_INIT(name.mmlist),		\
 	.cpu_vm_mask	= CPU_MASK_ALL,				\
 	.default_kioctx = INIT_KIOCTX(name.default_kioctx, name),	\
+	.task_list	= LIST_HEAD_INIT(name.task_list),	\
 }

 #define INIT_SIGNALS(sig) {	\
@@ -112,6 +113,7 @@
 	.proc_lock	= SPIN_LOCK_UNLOCKED,				\
 	.switch_lock	= SPIN_LOCK_UNLOCKED,				\
 	.journal_info	= NULL,						\
+	.mm_tasks	= LIST_HEAD_INIT(tsk.mm_tasks),			\
 }


Index: linux-2.6.10/fs/exec.c
===================================================================
--- linux-2.6.10.orig/fs/exec.c	2005-01-11 08:46:15.000000000 -0800
+++ linux-2.6.10/fs/exec.c	2005-01-11 08:56:45.000000000 -0800
@@ -556,6 +556,7 @@
 	tsk->active_mm = mm;
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
+	mm_add_thread(mm, current);
 	arch_pick_mmap_layout(mm);
 	if (old_mm) {
 		if (active_mm != old_mm) BUG();
Index: linux-2.6.10/fs/aio.c
===================================================================
--- linux-2.6.10.orig/fs/aio.c	2004-12-24 13:34:44.000000000 -0800
+++ linux-2.6.10/fs/aio.c	2005-01-11 08:56:45.000000000 -0800
@@ -577,6 +577,7 @@
 	tsk->active_mm = mm;
 	activate_mm(active_mm, mm);
 	task_unlock(tsk);
+	mm_add_thread(mm, tsk);

 	mmdrop(active_mm);
 }
@@ -596,6 +597,7 @@
 {
 	struct task_struct *tsk = current;

+	mm_remove_thread(mm,tsk);
 	task_lock(tsk);
 	tsk->flags &= ~PF_BORROWED_MM;
 	tsk->mm = NULL;
Index: linux-2.6.10/mm/swapfile.c
===================================================================
--- linux-2.6.10.orig/mm/swapfile.c	2005-01-11 08:46:16.000000000 -0800
+++ linux-2.6.10/mm/swapfile.c	2005-01-11 08:56:45.000000000 -0800
@@ -433,6 +433,7 @@
 	swp_entry_t entry, struct page *page)
 {
 	vma->vm_mm->rss++;
+	vma->vm_mm->anon_rss++;
 	get_page(page);
 	set_pte(dir, pte_mkold(mk_pte(page, vma->vm_page_prot)));
 	page_add_anon_rmap(page, vma, address);

Re: page table lock patch V15 [0/7]: overview

[Nick Piggin]

Christoph Lameter wrote:
> Changes from V14->V15 of this patch:

Hi,

I wonder what everyone thinks about moving forward with these patches?
Has it been decided that they'll be merged soon? Christoph has been
working fairly hard on them, but there hasn't been a lot of feedback.


And for those few people who have looked at my patches for page table
lock removal, is there is any preference to one implementation or the
other?

It is probably fair to say that my patches are more comprehensive
(in terms of ptl removal, ie. the complete removal**), and can allow
architectures to be more flexible in their page table synchronisation
methods.

However, Christoph's are simpler and probably more widely tested and
reviewed at this stage, and more polished. Christoph's implementation
probably also covers the most pressing performance cases.

On the other hand, my patches *do* allow for the use of a spin-locked
synchronisation implementation, which is probably closer to the
current code than Christoph's spin-locked pte_cmpxchg fallback in
terms of changes to locking semantics.


[** Aside, I didn't see a very significant improvement in mm/rmap.c
functions from ptl removal. Mostly I think due to contention on
mapping->i_mmap_lock (I didn't test anonymous pages, they may have
a better yield)]

Re: page table lock patch V15 [0/7]: overview

[Andrew Morton]

Nick Piggin <nickpiggin@yahoo.com.au> wrote:
>
> Christoph Lameter wrote:
>  > Changes from V14->V15 of this patch:
> 
>  Hi,
> 
>  I wonder what everyone thinks about moving forward with these patches?

I was waiting for them to settle down before paying more attention.

My general take is that these patches address a single workload on
exceedingly rare and expensive machines.  If they adversely affect common
and cheap machines via code complexity, memory footprint or via runtime
impact then it would be pretty hard to justify their inclusion.

Do we have measurements of the negative and/or positive impact on smaller
machines?

Re: page table lock patch V15 [0/7]: overview

[Marcelo Tosatti]

On Wed, Jan 12, 2005 at 01:42:35AM -0800, Andrew Morton wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> >
> > Christoph Lameter wrote:
> >  > Changes from V14->V15 of this patch:
> > 
> >  Hi,
> > 
> >  I wonder what everyone thinks about moving forward with these patches?
> 
> I was waiting for them to settle down before paying more attention.
> 
> My general take is that these patches address a single workload on
> exceedingly rare and expensive machines.  If they adversely affect common
> and cheap machines via code complexity, memory footprint or via runtime
> impact then it would be pretty hard to justify their inclusion.
> 
> Do we have measurements of the negative and/or positive impact on smaller
> machines?

I haven't seen wide performance numbers of this patch yet.

Hint: STP is really easy.

Re: page table lock patch V15 [0/7]: overview

[Hugh Dickins]

On Wed, 12 Jan 2005, Andrew Morton wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> >
> > Christoph Lameter wrote:
> >  > Changes from V14->V15 of this patch:
> > 
> >  I wonder what everyone thinks about moving forward with these patches?
> 
> I was waiting for them to settle down before paying more attention.

They seem to have settled down, without advancing to anything satisfactory.
7/7 is particularly amusing at the moment (added complexity with no payoff).

> My general take is that these patches address a single workload on
> exceedingly rare and expensive machines.

Well put.  Christoph's patches stubbornly remain a _good_ hack for one
very specific initial workload (multi-parallel faulting of anon memory)
on one architecture (ia64, perhaps a few more) important to SGI.
I don't see why the mainline kernel should want them.

> If they adversely affect common
> and cheap machines via code complexity, memory footprint or via runtime
> impact then it would be pretty hard to justify their inclusion.

Aside from 7/7 (and some good asm primitives within headers),
the code itself is not complex; but it is more complex to think about,
and so less obviously correct.

> Do we have measurements of the negative and/or positive impact on smaller
> machines?

I don't think so.  But my main worry remains the detriment to other
architectures, which still remains unaddressed.

Nick's patches (I've not seen for some while) are a different case:
on the minus side, considerably more complex; on the plus side,
more general and more aware of the range of architectures.

I'll write at greater length to support these accusations later on.

Hugh

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Wed, 12 Jan 2005, Andrew Morton wrote:

> My general take is that these patches address a single workload on
> exceedingly rare and expensive machines.  If they adversely affect common
> and cheap machines via code complexity, memory footprint or via runtime
> impact then it would be pretty hard to justify their inclusion.

The future is in higher and higher SMP counts since the chase for the
higher clock frequency has ended. We will increasingly see multi-core
cpus etc. Machines with higher CPU counts are becoming common in business.
Of course SGI uses much higher CPU counts and our supercomputer
applications would benefit most from this patch.

I thought this patch was already approved by Linus?

> Do we have measurements of the negative and/or positive impact on smaller
> machines?

Here is a measurement of 256M allocation on a 2 way SMP machine 2x
PIII-500Mhz:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  0  10    1    0.005s      0.016s   0.002s 54357.280  52261.895
  0  10    2    0.008s      0.019s   0.002s 43112.368  42463.566

With patch:

 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  0  10    1    0.005s      0.016s   0.002s 54357.280  53439.357
  0  10    2    0.008s      0.018s   0.002s 44650.831  44202.412

So only a very minor improvements for old machines (this one from ~ 98).

Re: page table lock patch V15 [0/7]: overview

[Christoph Hellwig]

On Wed, Jan 12, 2005 at 08:39:21AM -0800, Christoph Lameter wrote:
> The future is in higher and higher SMP counts since the chase for the
> higher clock frequency has ended. We will increasingly see multi-core
> cpus etc. Machines with higher CPU counts are becoming common in business.

An they still are absolutely in the minority.  In fact with multicore
cpus it becomes more and more important to be fast for SMP systtems with
a _small_ number of CPUs, while really larget CPUs will remain a small
nische for the forseeable future.

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Wed, 12 Jan 2005, Christoph Hellwig wrote:

> On Wed, Jan 12, 2005 at 08:39:21AM -0800, Christoph Lameter wrote:
> > The future is in higher and higher SMP counts since the chase for the
> > higher clock frequency has ended. We will increasingly see multi-core
> > cpus etc. Machines with higher CPU counts are becoming common in business.
>
> An they still are absolutely in the minority.  In fact with multicore
> cpus it becomes more and more important to be fast for SMP systtems with
> a _small_ number of CPUs, while really larget CPUs will remain a small
> nische for the forseeable future.

The benefits start to be significant pretty fast with even a few cpus
on modern architectures:

Altix  no patch:
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  1  10    1    0.107s      6.444s   6.055s100028.084 100006.622
  1  10    2    0.121s      9.048s   4.082s 71468.414 135904.412
  1  10    4    0.129s     10.185s   3.011s 63531.985 210146.600

w/patch
 Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
  1  10    1    0.094s      6.116s   6.021s105517.039 105517.574
  1  10    2    0.134s      6.998s   3.087s 91879.573 169079.712
  1  10    4    0.095s      7.658s   2.043s 84519.939 268955.165

There is even a small benefit to the single thread case.

Its not the case that this patch only benefits systems with a large number
of CPUs. Of course that is when the benefits results in performance gains
by orders of magnitude.

Re: page table lock patch V15 [0/7]: overview

[Christoph Hellwig]

On Wed, Jan 12, 2005 at 09:37:27AM -0800, Christoph Lameter wrote:
> 
> The benefits start to be significant pretty fast with even a few cpus
> on modern architectures:
> 
> Altix  no patch:
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>   1  10    1    0.107s      6.444s   6.055s100028.084 100006.622
>   1  10    2    0.121s      9.048s   4.082s 71468.414 135904.412
>   1  10    4    0.129s     10.185s   3.011s 63531.985 210146.600
> 
> w/patch
>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>   1  10    1    0.094s      6.116s   6.021s105517.039 105517.574
>   1  10    2    0.134s      6.998s   3.087s 91879.573 169079.712
>   1  10    4    0.095s      7.658s   2.043s 84519.939 268955.165

These smaller systems are more likely x86/x86_64 machines ;-)

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Wed, 12 Jan 2005, Christoph Hellwig wrote:

> These smaller systems are more likely x86/x86_64 machines ;-)

But they will not have been build in 1998 either like the machine I used
for the i386 tests. Could you do some tests on contemporary x86/x86_64
SMP systems with large memory?

Re: page table lock patch V15 [0/7]: overview

[Christoph Hellwig]

On Wed, Jan 12, 2005 at 09:52:53AM -0800, Christoph Lameter wrote:
> On Wed, 12 Jan 2005, Christoph Hellwig wrote:
> 
> > These smaller systems are more likely x86/x86_64 machines ;-)
> 
> But they will not have been build in 1998 either like the machine I used
> for the i386 tests. Could you do some tests on contemporary x86/x86_64
> SMP systems with large memory?

I don't have such systems.

Re: page table lock patch V15 [0/7]: overview

[Andrew Walrond]

On Wednesday 12 January 2005 17:52, Christoph Lameter wrote:
> On Wed, 12 Jan 2005, Christoph Hellwig wrote:
> > These smaller systems are more likely x86/x86_64 machines ;-)
>
> But they will not have been build in 1998 either like the machine I used
> for the i386 tests. Could you do some tests on contemporary x86/x86_64
> SMP systems with large memory?

I have various dual x86_64 systems with 1-4Gb ram. What tests do you want run?

Andrew Walrond

Re: page table lock patch V15 [0/7]: overview

[Andrew Morton]

Christoph Lameter <clameter@sgi.com> wrote:
>
> > Do we have measurements of the negative and/or positive impact on smaller
>  > machines?
> 
>  Here is a measurement of 256M allocation on a 2 way SMP machine 2x
>  PIII-500Mhz:
> 
>   Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>    0  10    1    0.005s      0.016s   0.002s 54357.280  52261.895
>    0  10    2    0.008s      0.019s   0.002s 43112.368  42463.566
> 
>  With patch:
> 
>   Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>    0  10    1    0.005s      0.016s   0.002s 54357.280  53439.357
>    0  10    2    0.008s      0.018s   0.002s 44650.831  44202.412
> 
>  So only a very minor improvements for old machines (this one from ~ 98).

OK.  But have you written a test to demonstrate any performance
regressions?  From, say, the use of atomic ops on ptes?

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Wed, 12 Jan 2005, Andrew Morton wrote:

> >  So only a very minor improvements for old machines (this one from ~ 98).
>
> OK.  But have you written a test to demonstrate any performance
> regressions?  From, say, the use of atomic ops on ptes?

If I knew of any regressions, I would certain try to deal with them. The
test is written to check for concurrent page fault performance and it has
repeatedly helped me to find problems with page faults. I have used it for
a couple of other patchsets too. If the patch would be available in mm
then it certainly would get more exposure and it may become clear that
there are some regressions.

Introduction of the cmpxchg is one atomic operations that replaces the two
spinlock ops typically necessary in an unpatched kernel. Obtaining the
spinlock requires an spinlock (which is an atomic operation) and then the
release involves a barrier. So there is a net win for all SMP cases as far
as I can see.

Re: page table lock patch V15 [0/7]: overview

[Hugh Dickins]

On Wed, 12 Jan 2005, Hugh Dickins wrote:
> On Wed, 12 Jan 2005, Andrew Morton wrote:
> > Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> > > Christoph Lameter wrote:
> > >  > Changes from V14->V15 of this patch:
> > >  I wonder what everyone thinks about moving forward with these patches?
> > I was waiting for them to settle down before paying more attention.
> They seem to have settled down, without advancing to anything satisfactory.

Well, I studied the patches a bit more, and wrote
"That remark looks a bit unfair to me now I've looked closer."
Sorry.  But I do still think it remains unsatisfactory."

Then I studied it a bit more, and I think my hostility melted away
once I thought about the other-arch-defaults: I'd been supposing that
taking and dropping the page_table_lock within each primitive was
adding up to an unhealthy flurry of takes and drops on the non-target
architectures.  But that doesn't look like the case to me now (except
in those rarer paths where a page table has to be allocated: of course,
not a problem).

I owe Christoph an apology.  It's not quite satisfactory yet,
but it does look a lot better than an ia64 hack for one special case.

Might I save face by suggesting that it would be a lot clearer and
better if 1/1 got split into two?  The first entirely concerned with
removing the spin_lock(&mm->page_table_lock) from handle_mm_fault,
and dealing with the consequences of that - moving the locking into
the allocating blocks, atomic getting of pud and pmd and pte,
passing the atomically-gotten orig_pte down to subfunctions
(which no longer expect page_table_lock held on entry) etc.

If there's a slight increase in the number of atomic operations
in each i386 PAE page fault, well, I think the superiority of
x86_64 makes that now an acceptable tradeoff.

That would be quite a decent patch, wouldn't it? that could go into
-mm for a few days and be measured, before any more.  Then one using
something called ptep_cmpxchg to encapsulate the page_table_lock'ed
checking of pte_same and set_pte in do_anonymous page.  Then ones to
implement ptep_cmpxchg per selected arches without page_table_lock.

Dismiss those suggestions if they'd just waste everyone's time.

Christoph has made some strides in correcting for other architectures
e.g. update_mmu_cache within default ptep_cmpxchg's page_table_lock
(probably correct but I can't be sure myself), and get_pte_atomic to
get even i386 PAE pte correctly without page_table_lock; and reverted
the pessimization of set_pte being always atomic on i386 PAE (but now
I've forgotten and can't find the case where it needed to be atomic).

Unless it's just been fixed in this latest version, the well-intentioned
get_pte_atomic doesn't actually work on i386 PAE: once you get swapping,
the swap entries look like pte_nones and all collapses.  Presumably just
#define get_pte_atomic(__ptep) __pte(get_64bit((unsigned long long *)(__ptep)))
doesn't quite do what it's trying to do, and needs a slight adjustment.

But no sign of get_pmd(atomic) or get_pud(atomic) to get the higher level
entries - I thought we'd agreed they were also necessary on some arches?

> 7/7 is particularly amusing at the moment (added complexity with no payoff).

I still dislike 7/7, despite seeing the sense of keeping stats in the
task struct.  It's at the very end anyway, and I'd be glad for it to
be delayed (in the hope that time somehow magically makes it nicer).

In its present state it is absurd: partly because Christoph seems to
have forgotten the point of it, so after all the per-thread infrastructure,
has ended up with do_anonymous_page saying mm->rss++, mm->anon_rss++.

And partly because others at SGI have been working in the opposite
direction, adding mysterious and tasteless acct_update_integrals
and update_mem_hiwater calls.  I say mysterious because there's
nothing in the tree which actually uses the accumulated statistics,
or shows how they might be used (when many threads share the mm),
- so Adrian/Arjan/HCH might remove them any day.  But looking at
December mails suggests there's lse-tech agreement that all kinds
of addons would find them useful.  I say tasteless because they
don't even take "mm" arguments (what happens when ptrace or AIO
daemon faults something? perhaps it's okay but there's no use of
the stats to judge by), and the places where you'd want to update
hiwater_rss are almost entirely disjoint from the places where
you'd want to update hiwater_vm (expand_stack the exception).

If those new stats stay, and the per-task-rss idea stays,
then I suppose those new stats need to be split per task too.

> I'll write at greater length to support these accusations later on.

I rather failed to do so!  And perhaps tomorrow I'll have to be
apologizing to Jay for my uncomprehending attack on hiwater etc.

Hugh

Re: page table lock patch V15 [0/7]: overview

[Nick Piggin]

Andrew Morton wrote:
> Christoph Lameter <clameter@sgi.com> wrote:
> 
>>>Do we have measurements of the negative and/or positive impact on smaller
>>
>> > machines?
>>
>> Here is a measurement of 256M allocation on a 2 way SMP machine 2x
>> PIII-500Mhz:
>>
>>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>>   0  10    1    0.005s      0.016s   0.002s 54357.280  52261.895
>>   0  10    2    0.008s      0.019s   0.002s 43112.368  42463.566
>>
>> With patch:
>>
>>  Gb Rep Threads   User      System     Wall flt/cpu/s fault/wsec
>>   0  10    1    0.005s      0.016s   0.002s 54357.280  53439.357
>>   0  10    2    0.008s      0.018s   0.002s 44650.831  44202.412
>>
>> So only a very minor improvements for old machines (this one from ~ 98).
> 
> 
> OK.  But have you written a test to demonstrate any performance
> regressions?  From, say, the use of atomic ops on ptes?
> 

Performance wise, Christoph's never had as much of a problem as my
patches because it isn't doing extra atomic operations in copy_page_range.

However, it looks like it should be. For the same reason there needs to
be an atomic read in handle_mm_fault. And it probably needs atomic ops
in other places too, I think.

So my patches cost about 7% in lmbench fork benchmark... however, I've
been thinking we could take the mmap_sem for writing before doing the
copy_page_range which could reduce the need for atomic ops.

Re: page table lock patch V15 [0/7]: overview

[Andrew Morton]

Nick Piggin <nickpiggin@yahoo.com.au> wrote:
>
> So my patches cost about 7% in lmbench fork benchmark.

OK, well that's the sort of thing we need to understand fully.  What sort
of CPU was that on?

Look, -7% on a 2-way versus +700% on a many-way might well be a tradeoff we
agree to take.  But we need to fully understand all the costs and benefits.

Re: page table lock patch V15 [0/7]: overview

[Nick Piggin]

Andrew Morton wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> 
>>So my patches cost about 7% in lmbench fork benchmark.
> 
> 
> OK, well that's the sort of thing we need to understand fully.  What sort
> of CPU was that on?
> 

That was on a P4, although I've seen pretty similar results on ia64 and
other x86 CPUs.

Note that this was with my ptl removal patches. I can't see why Christoph's
would have _any_ extra overhead as they are, but it looks to me like they're
lacking in atomic ops. So I'd expect something similar for Christoph's when
they're properly atomic.

> Look, -7% on a 2-way versus +700% on a many-way might well be a tradeoff we
> agree to take.  But we need to fully understand all the costs and benefits.
> 

I think copy_page_range is the one to keep an eye on.

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Wed, 12 Jan 2005, Hugh Dickins wrote:

> Well, I studied the patches a bit more, and wrote
> "That remark looks a bit unfair to me now I've looked closer."
> Sorry.  But I do still think it remains unsatisfactory."

Well then thanks for not ccing me on the initial rant but a whole bunch of
other people instead that you then did not send the following email too.
Is this standard behavior on linux-mm?

> Might I save face by suggesting that it would be a lot clearer and
> better if 1/1 got split into two?  The first entirely concerned with
> removing the spin_lock(&mm->page_table_lock) from handle_mm_fault,
> and dealing with the consequences of that - moving the locking into
> the allocating blocks, atomic getting of pud and pmd and pte,
> passing the atomically-gotten orig_pte down to subfunctions
> (which no longer expect page_table_lock held on entry) etc.

That wont do any good since the pte's are not always updated in an atomic
way. One would have to change set_pte to always be atomic. The reason
that I added get_pte_atomic was that you told me that this would fix the
PAE mode. I did not think too much about this but simply added it
according to your wish and it seemed to run fine. If you have any
complaints, complain to yourself.

> If there's a slight increase in the number of atomic operations
> in each i386 PAE page fault, well, I think the superiority of
> x86_64 makes that now an acceptable tradeoff.

Could we have PAE mode drop back to using the page_table_lock?

> Dismiss those suggestions if they'd just waste everyone's time.

They dont fix the PAE mode issue.

> Christoph has made some strides in correcting for other architectures
> e.g. update_mmu_cache within default ptep_cmpxchg's page_table_lock
> (probably correct but I can't be sure myself), and get_pte_atomic to
> get even i386 PAE pte correctly without page_table_lock; and reverted
> the pessimization of set_pte being always atomic on i386 PAE (but now
> I've forgotten and can't find the case where it needed to be atomic).

Well this was another suggestion of yours that I followed. Turns out that
the set_pte must be atomic for this to work! Look I am no expert on the
i386 PAE mode and I rely on other for this to check up on it. And you were
the expert.

> But no sign of get_pmd(atomic) or get_pud(atomic) to get the higher level
> entries - I thought we'd agreed they were also necessary on some arches?

I did not hear about that. Maybe you also sent that email to other people
instead?

> In its present state it is absurd: partly because Christoph seems to
> have forgotten the point of it, so after all the per-thread infrastructure,
> has ended up with do_anonymous_page saying mm->rss++, mm->anon_rss++.

Sorry that seems to have dropped out of the patch somehow. Here is the
fix:

Index: linux-2.6.10/mm/memory.c
===================================================================
--- linux-2.6.10.orig/mm/memory.c	2005-01-11 09:16:34.000000000 -0800
+++ linux-2.6.10/mm/memory.c	2005-01-12 15:49:45.000000000 -0800
@@ -1835,8 +1835,8 @@ do_anonymous_page(struct mm_struct *mm,
 		 */
 		page_add_anon_rmap(page, vma, addr);
 		lru_cache_add_active(page);
-		mm->rss++;
-		mm->anon_rss++;
+		current->rss++;
+		current->anon_rss++;
 		acct_update_integrals();
 		update_mem_hiwater();



> And partly because others at SGI have been working in the opposite
> direction, adding mysterious and tasteless acct_update_integrals
> and update_mem_hiwater calls.  I say mysterious because there's

Yea. I posted a patch to move that stuff out of the vm. No
good deed gets unpunished.

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Thu, 13 Jan 2005, Nick Piggin wrote:

> Note that this was with my ptl removal patches. I can't see why Christoph's
> would have _any_ extra overhead as they are, but it looks to me like they're
> lacking in atomic ops. So I'd expect something similar for Christoph's when
> they're properly atomic.

Pointer operations and word size operations are atomic. So this is mostly
okay.

The issue arises on architectures that have a large pte size than the
wordsize. This is only on i386 PAE mode and S/390. S/390 falls back to
the page table lock  for these operations. PAE mode should do the same and
not use atomic ops if they cannot be made to work in a reasonable manner.

Re: page table lock patch V15 [0/7]: overview

[Nick Piggin]

Christoph Lameter wrote:
> On Thu, 13 Jan 2005, Nick Piggin wrote:
> 
> 
>>Note that this was with my ptl removal patches. I can't see why Christoph's
>>would have _any_ extra overhead as they are, but it looks to me like they're
>>lacking in atomic ops. So I'd expect something similar for Christoph's when
>>they're properly atomic.
> 
> 
> Pointer operations and word size operations are atomic. So this is mostly
> okay.
> 
> The issue arises on architectures that have a large pte size than the
> wordsize. This is only on i386 PAE mode and S/390. S/390 falls back to
> the page table lock  for these operations. PAE mode should do the same and
> not use atomic ops if they cannot be made to work in a reasonable manner.
> 

Yep well you should be OK then. Your implementation has the advantage
that it only instantiates previously clear ptes... hmm, no I'm wrong,
your ptep_set_access_flags path modifies an existing pte. I think this
can cause subtle races in copy_page_range, and maybe other places,
can't it?

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Thu, 13 Jan 2005, Nick Piggin wrote:

> > Pointer operations and word size operations are atomic. So this is mostly
> > okay.
> >
> > The issue arises on architectures that have a large pte size than the
> > wordsize. This is only on i386 PAE mode and S/390. S/390 falls back to
> > the page table lock  for these operations. PAE mode should do the same and
> > not use atomic ops if they cannot be made to work in a reasonable manner.
> >
>
> Yep well you should be OK then. Your implementation has the advantage
> that it only instantiates previously clear ptes... hmm, no I'm wrong,
> your ptep_set_access_flags path modifies an existing pte. I think this
> can cause subtle races in copy_page_range, and maybe other places,
> can't it?

ptep_set_access_flags is only used after acquiring the page_table_lock and
does not clear a pte. That is safe. The only critical thing is if a pte
would be cleared while holding the page_table_lock. That used to occur in
the swapper code but we modified that.

There is still an issue as Hugh rightly observed. One cannot rely on a
read of a pte/pud/pmd being atomic if the pte is > word size. This occurs
for all higher levels in handle_mm_fault. Thus we would need to either
acuire the page_table_lock for some architectures or provide primitives
get_pgd, get_pud etc that take the page_table_lock on PAE mode. ARGH.

Re: page table lock patch V15 [0/7]: overview

[Nick Piggin]

Christoph Lameter wrote:
> On Thu, 13 Jan 2005, Nick Piggin wrote:
> 
> 
>>>Pointer operations and word size operations are atomic. So this is mostly
>>>okay.
>>>
>>>The issue arises on architectures that have a large pte size than the
>>>wordsize. This is only on i386 PAE mode and S/390. S/390 falls back to
>>>the page table lock  for these operations. PAE mode should do the same and
>>>not use atomic ops if they cannot be made to work in a reasonable manner.
>>>
>>
>>Yep well you should be OK then. Your implementation has the advantage
>>that it only instantiates previously clear ptes... hmm, no I'm wrong,
>>your ptep_set_access_flags path modifies an existing pte. I think this
>>can cause subtle races in copy_page_range, and maybe other places,
>>can't it?
> 
> 
> ptep_set_access_flags is only used after acquiring the page_table_lock and
> does not clear a pte. That is safe. The only critical thing is if a pte
> would be cleared while holding the page_table_lock. That used to occur in
> the swapper code but we modified that.
> 

I mean what used to be the ptep_set_access_flags path. Where you are
now modifying a pte without the ptl. However after a second look, it
seems like that won't be a problem.

> There is still an issue as Hugh rightly observed. One cannot rely on a
> read of a pte/pud/pmd being atomic if the pte is > word size. This occurs
> for all higher levels in handle_mm_fault. Thus we would need to either
> acuire the page_table_lock for some architectures or provide primitives
> get_pgd, get_pud etc that take the page_table_lock on PAE mode. ARGH.
> 

Yes I know. I would say that having arch-definable accessors for the
page tables wouldn't be a bad idea anyway, and the flexibility may
come in handy for other things.

It would be a big, annoying patch though :(

Re: page table lock patch V15 [0/7]: overview

[Hugh Dickins]

On Wed, 12 Jan 2005, Christoph Lameter wrote:
> On Wed, 12 Jan 2005, Hugh Dickins wrote:
> 
> > Well, I studied the patches a bit more, and wrote
> > "That remark looks a bit unfair to me now I've looked closer."
> > Sorry.  But I do still think it remains unsatisfactory."
> 
> Well then thanks for not ccing me on the initial rant but a whole bunch of
> other people instead that you then did not send the following email too.
> Is this standard behavior on linux-mm?

I did cc you.  What whole bunch of other people?  The list of recipients
was the same, except (for obvious reasons) I added Jay the second time
(and having more time, spelt out most names in full).

Perhaps we've a misunderstanding: when I say "and wrote..." above,
I'm not quoting from some mail I sent others not you, I'm referring
to an earlier draft of the mail I'm then sending.

Or perhaps SGI has a spam filter which chose to gobble it up.
I'll try forwarding it to you again.

> > Might I save face by suggesting that it would be a lot clearer and
> > better if 1/1 got split into two?  The first entirely concerned with
> > removing the spin_lock(&mm->page_table_lock) from handle_mm_fault,
> > and dealing with the consequences of that - moving the locking into
> > the allocating blocks, atomic getting of pud and pmd and pte,
> > passing the atomically-gotten orig_pte down to subfunctions
> > (which no longer expect page_table_lock held on entry) etc.
> 
> That wont do any good since the pte's are not always updated in an atomic
> way. One would have to change set_pte to always be atomic.

You did have set_pte always atomic at one point, to the detriment of
(PAE) set_page_range.  You rightly reverted that, but you've reminded
me of what I confessed to forgetting, where you do need set_pte_atomic
in various places, mainly (only?) the fault handlers in mm/memory.c.
And yes, I think you're right, that needs to be in this first patch.

> The reason
> that I added get_pte_atomic was that you told me that this would fix the
> PAE mode. I did not think too much about this but simply added it
> according to your wish and it seemed to run fine.

Please don't leave the thinking to me or anyone else.

> If you have any complaints, complain to yourself.

I'd better omit my response to that.

> > If there's a slight increase in the number of atomic operations
> > in each i386 PAE page fault, well, I think the superiority of
> > x86_64 makes that now an acceptable tradeoff.
> 
> Could we have PAE mode drop back to using the page_table_lock?

That sounds a simple and sensible alternative (to more atomics):
haven't really thought it through, but if the default arch code is
right, and not overhead, then why not use it for the PAE case instead
of cluttering up with cleverness.  Yes, I think that's a good idea:
anyone see why not?

> > Dismiss those suggestions if they'd just waste everyone's time.
> 
> They dont fix the PAE mode issue.
> 
> > Christoph has made some strides in correcting for other architectures
> > e.g. update_mmu_cache within default ptep_cmpxchg's page_table_lock
> > (probably correct but I can't be sure myself), and get_pte_atomic to
> > get even i386 PAE pte correctly without page_table_lock; and reverted
> > the pessimization of set_pte being always atomic on i386 PAE (but now
> > I've forgotten and can't find the case where it needed to be atomic).
> 
> Well this was another suggestion of yours that I followed. Turns out that
> the set_pte must be atomic for this to work!

I didn't say you never needed an atomic set_pte, I said that making
set_pte always atomic (in the PAE case) unnecessarily slowed down
copy_page_range and zap_pte_range.  Probably a misunderstanding.

> Look I am no expert on the
> i386 PAE mode and I rely on other for this to check up on it. And you were
> the expert.

Expert?  I was trying to help, but you seem to resent that.

> > But no sign of get_pmd(atomic) or get_pud(atomic) to get the higher level
> > entries - I thought we'd agreed they were also necessary on some arches?
> 
> I did not hear about that. Maybe you also sent that email to other people
> instead?

No, you were cc'ed on that one too (Sun, 12 Dec to Nick Piggin).
The spam filter again.  Not that I have total recall of every
exchange about these patches either.

Hugh

Re: page table lock patch V15 [0/7]: overview

[Hugh Dickins]

On Thu, 13 Jan 2005, Nick Piggin wrote:
> Andrew Morton wrote:
> 
> Note that this was with my ptl removal patches. I can't see why Christoph's
> would have _any_ extra overhead as they are, but it looks to me like they're
> lacking in atomic ops. So I'd expect something similar for Christoph's when
> they're properly atomic.
> 
> > Look, -7% on a 2-way versus +700% on a many-way might well be a tradeoff we
> > agree to take.  But we need to fully understand all the costs and benefits.
> 
> I think copy_page_range is the one to keep an eye on.

Christoph's currently lack set_pte_atomics in the fault handlers, yes.
But I don't see why they should need set_pte_atomics in copy_page_range
(which is why I persuaded him to drop forcing set_pte to atomic).

dup_mmap has down_write of the src mmap_sem, keeping out any faults on
that.  copy_pte_range has spin_lock of the dst page_table_lock and the
src page_table_lock, keeping swapout away from those.  Why would atomic
set_ptes be needed there?  Probably in yours, but not in Christoph's.

Hugh

Re: page table lock patch V15 [0/7]: overview

[Andi Kleen]

> There is still an issue as Hugh rightly observed. One cannot rely on a
> read of a pte/pud/pmd being atomic if the pte is > word size. This occurs
> for all higher levels in handle_mm_fault. Thus we would need to either
> acuire the page_table_lock for some architectures or provide primitives
> get_pgd, get_pud etc that take the page_table_lock on PAE mode. ARGH.
> 

Alternatively you can use a lazy load, checking for changes. 
(untested) 

pte_t read_pte(volatile pte_t *pte) 
{ 
	pte_t n;
	do { 
		n.pte_low = pte->pte_low;
		rmb();
		n.pte_high = pte->pte_high;
		rmb();
	} while (n.pte_low != pte->pte_low); 
	return pte; 	
} 

No atomic operations, I bet it's actually faster than the cmpxchg8.
There is a small risk for livelock, but not much worse than with an
ordinary spinlock.

Not that I get it what you want it for exactly - the content
of the pte could change any time when you don't hold page_table_lock, right?

-Andi

Re: page table lock patch V15 [0/7]: overview

[Nick Piggin]

Hugh Dickins wrote:
> On Thu, 13 Jan 2005, Nick Piggin wrote:
> 
>>Andrew Morton wrote:
>>
>>Note that this was with my ptl removal patches. I can't see why Christoph's
>>would have _any_ extra overhead as they are, but it looks to me like they're
>>lacking in atomic ops. So I'd expect something similar for Christoph's when
>>they're properly atomic.
>>
>>
>>>Look, -7% on a 2-way versus +700% on a many-way might well be a tradeoff we
>>>agree to take.  But we need to fully understand all the costs and benefits.
>>
>>I think copy_page_range is the one to keep an eye on.
> 
> 
> Christoph's currently lack set_pte_atomics in the fault handlers, yes.
> But I don't see why they should need set_pte_atomics in copy_page_range
> (which is why I persuaded him to drop forcing set_pte to atomic).
> 
> dup_mmap has down_write of the src mmap_sem, keeping out any faults on
> that.  copy_pte_range has spin_lock of the dst page_table_lock and the
> src page_table_lock, keeping swapout away from those.  Why would atomic
> set_ptes be needed there?  Probably in yours, but not in Christoph's.
> 

I was more thinking of atomic pte reads there. I had for some reason
thought that dup_mmap only had a down_read of the mmap_sem. But even if
it did only down_read, a further look showed this wouldn't have been a
problem for Christoph anyway. That dim light-bulb probably changes things
for my patches too; I may be able to do copy_page_range with fewer atomics.

I'm still not too sure that all places read the pte atomically where needed.
But presently this is not a really big concern because it only would
really slow down i386 PAE if anything.

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Thu, 13 Jan 2005, Hugh Dickins wrote:

> I did cc you.  What whole bunch of other people?  The list of recipients
> was the same, except (for obvious reasons) I added Jay the second time
> (and having more time, spelt out most names in full).
>
> Or perhaps SGI has a spam filter which chose to gobble it up.
> I'll try forwarding it to you again.

Yes sorry it was the spam filter. I got my copy from linux-mm just fine on
another account.

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Wed, 13 Jan 2005, Andi Kleen wrote:

> Alternatively you can use a lazy load, checking for changes.
> (untested)
>
> pte_t read_pte(volatile pte_t *pte)
> {
> 	pte_t n;
> 	do {
> 		n.pte_low = pte->pte_low;
> 		rmb();
> 		n.pte_high = pte->pte_high;
> 		rmb();
> 	} while (n.pte_low != pte->pte_low);
> 	return pte;
> }
>
> No atomic operations, I bet it's actually faster than the cmpxchg8.
> There is a small risk for livelock, but not much worse than with an
> ordinary spinlock.

Hmm.... This may replace the get of a 64 bit value. But here could still
be another process that is setting the pte in a non-atomic way.

> Not that I get it what you want it for exactly - the content
> of the pte could change any time when you don't hold page_table_lock, right?

The content of the pte can change anytime the page_table_lock is held and
it may change from cleared to a value through a cmpxchg while the lock is
not held.

Re: page table lock patch V15 [0/7]: overview

[Christoph Lameter]

On Thu, 13 Jan 2005, Nick Piggin wrote:

> I'm still not too sure that all places read the pte atomically where needed.
> But presently this is not a really big concern because it only would
> really slow down i386 PAE if anything.

S/390 is also affected. And I vaguely recall special issues with sparc
too. That is why I dropped the arch support for that a long time ago from
the patchset. Then there was some talk a couple of months back to use
another addressing mode on IA64 that may also require 128 bit ptes. There
are significantly different ways of doing optimal SMP locking for these
scenarios.

Re: page table lock patch V15 [0/7]: overview

[Linus Torvalds]

On Thu, 13 Jan 2005, Christoph Lameter wrote:
>
> On Wed, 13 Jan 2005, Andi Kleen wrote:
> 
> > Alternatively you can use a lazy load, checking for changes.
> > (untested)
> >
> > pte_t read_pte(volatile pte_t *pte)
> > {
> > 	pte_t n;
> > 	do {
> > 		n.pte_low = pte->pte_low;
> > 		rmb();
> > 		n.pte_high = pte->pte_high;
> > 		rmb();
> > 	} while (n.pte_low != pte->pte_low);
> > 	return pte;
> > }
> >
> > No atomic operations, I bet it's actually faster than the cmpxchg8.
> > There is a small risk for livelock, but not much worse than with an
> > ordinary spinlock.
> 
> Hmm.... This may replace the get of a 64 bit value. But here could still
> be another process that is setting the pte in a non-atomic way.

There's a nice standard way of doing that, namely sequence numbers. 

However, most of the time it isn't actually faster than just getting the 
lock. There are two real costs in getting a lock: serialization and cache 
bouncing. The ordering often requires _more_ serialization than a 
lock/unlock sequence, so sequences like the above are often slower than 
the trivial lock is, at least in the absense of lock contention.

So sequence numbers (or multiple reads) only tend make sense where there
is a _lot_ more reads than writes, and where you get lots of lock 
contention. If there are lots of writes, my gut feel (but hey, all locking 
optimization should be backed up by real numbers) is that it's better to 
have a lock close to the data, since you'll get the cacheline bounces 
_anyway_, and locking often has lower serialization costs.

		Linus

Re: page table lock patch V15 [0/7]: overview

[Andi Kleen]

On Thu, Jan 13, 2005 at 09:11:29AM -0800, Christoph Lameter wrote:
> On Wed, 13 Jan 2005, Andi Klee