[Xen-devel] [PATCH 0/2] Handle

[Xen-devel] [PATCH 0/2] Handle

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 1366 bytes --]

There are cases where pages can get freed with free_heap_pages() when
in fact they were never properly initialised in the heap — they may
have been allocated from the boot allocator, simply assigned directly
to dom0 as part of its initrd, etc.

We have plans to make vmap available during early boot, which would
exacerbate this situation a tiny bit more, as a few more page tables
would stand a small chance of being allocated by the boot allocator and
freed later.

Resolve this by introducing a new page state, PGC_state_uninitialised,
expanding the PGC_state to 3 bits (8 possible values) by subsuming the
PGC_broken bit into it and eliminating the redundant possible
combinations of PGC_broken and various states.

Pages which find their way into free_heap_pages() while still in
PGC_state_uninitialised can thus be detected and properly
rehabilitated, basically by passing them through init_heap_pages().


David Woodhouse (2):
      xen/mm: fold PGC_broken into PGC_state bits
      xen/mm: Introduce PGC_state_uninitialised

xen/arch/x86/domctl.c    |   2 +-
 xen/arch/x86/mm.c        |   3 +-
 xen/common/page_alloc.c  | 110 +++++++++++++++++++++++++++++------------------
 xen/include/asm-arm/mm.h |  39 +++++++++++------
 xen/include/asm-x86/mm.h |  37 +++++++++++-----
 5 files changed, 125 insertions(+), 66 deletions(-)

[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

[Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

Only PGC_state_offlining and PGC_state_offlined are valid in conjunction
with PGC_broken. The other two states (free and inuse) were never valid
for a broken page.

By folding PGC_broken in, we can have three bits for PGC_state which
allows up to 8 states, of which 6 are currently used and 2 are available
for new use cases.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 xen/arch/x86/domctl.c    |  2 +-
 xen/common/page_alloc.c  | 66 ++++++++++++++++++++++------------------
 xen/include/asm-arm/mm.h | 38 +++++++++++++++--------
 xen/include/asm-x86/mm.h | 36 ++++++++++++++++------
 4 files changed, 89 insertions(+), 53 deletions(-)

diff --git a/xen/arch/x86/domctl.c b/xen/arch/x86/domctl.c
index ed86762fa6..a411f64afa 100644
--- a/xen/arch/x86/domctl.c
+++ b/xen/arch/x86/domctl.c
@@ -422,7 +422,7 @@ long arch_do_domctl(
                 if ( page->u.inuse.type_info & PGT_pinned )
                     type |= XEN_DOMCTL_PFINFO_LPINTAB;
 
-                if ( page->count_info & PGC_broken )
+                if ( page_is_broken(page) )
                     type = XEN_DOMCTL_PFINFO_BROKEN;
             }
 
diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
index 76d37226df..8d72a64f4e 100644
--- a/xen/common/page_alloc.c
+++ b/xen/common/page_alloc.c
@@ -1093,7 +1093,7 @@ static int reserve_offlined_page(struct page_info *head)
         struct page_info *pg;
         int next_order;
 
-        if ( page_state_is(cur_head, offlined) )
+        if ( page_is_offlined(cur_head) )
         {
             cur_head++;
             if ( first_dirty != INVALID_DIRTY_IDX && first_dirty )
@@ -1113,7 +1113,7 @@ static int reserve_offlined_page(struct page_info *head)
             for ( i = (1 << cur_order), pg = cur_head + (1 << cur_order );
                   i < (1 << next_order);
                   i++, pg++ )
-                if ( page_state_is(pg, offlined) )
+                if ( page_is_offlined(pg) )
                     break;
             if ( i == ( 1 << next_order) )
             {
@@ -1145,16 +1145,20 @@ static int reserve_offlined_page(struct page_info *head)
 
     for ( cur_head = head; cur_head < head + ( 1UL << head_order); cur_head++ )
     {
-        if ( !page_state_is(cur_head, offlined) )
+        struct page_list_head *list;
+
+        if ( page_state_is(cur_head, offlined) )
+            list = &page_offlined_list;
+        else if (page_state_is(cur_head, broken) )
+            list = &page_broken_list;
+        else
             continue;
 
         avail[node][zone]--;
         total_avail_pages--;
         ASSERT(total_avail_pages >= 0);
 
-        page_list_add_tail(cur_head,
-                           test_bit(_PGC_broken, &cur_head->count_info) ?
-                           &page_broken_list : &page_offlined_list);
+        page_list_add_tail(cur_head, list);
 
         count++;
     }
@@ -1404,13 +1408,16 @@ static void free_heap_pages(
         switch ( pg[i].count_info & PGC_state )
         {
         case PGC_state_inuse:
-            BUG_ON(pg[i].count_info & PGC_broken);
             pg[i].count_info = PGC_state_free;
             break;
 
         case PGC_state_offlining:
-            pg[i].count_info = (pg[i].count_info & PGC_broken) |
-                               PGC_state_offlined;
+            pg[i].count_info = PGC_state_offlined;
+            tainted = 1;
+            break;
+
+        case PGC_state_broken_offlining:
+            pg[i].count_info = PGC_state_broken;
             tainted = 1;
             break;
 
@@ -1527,16 +1534,16 @@ static unsigned long mark_page_offline(struct page_info *pg, int broken)
     do {
         nx = x = y;
 
-        if ( ((x & PGC_state) != PGC_state_offlined) &&
-             ((x & PGC_state) != PGC_state_offlining) )
-        {
-            nx &= ~PGC_state;
-            nx |= (((x & PGC_state) == PGC_state_free)
-                   ? PGC_state_offlined : PGC_state_offlining);
-        }
+        nx &= ~PGC_state;
 
-        if ( broken )
-            nx |= PGC_broken;
+        /* If it was already broken, it stays broken */
+        if ( pgc_is_broken(x) )
+            broken = 1;
+
+        if ( pgc_is_offlined(x) || pgc_is(x, free) )
+            nx |= broken ? PGC_state_broken : PGC_state_offlined;
+        else
+            nx |= broken ? PGC_state_broken_offlining : PGC_state_offlining;
 
         if ( x == nx )
             break;
@@ -1609,7 +1616,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
      * need to prevent malicious guest access the broken page again.
      * Under such case, hypervisor shutdown guest, preventing recursive mce.
      */
-    if ( (pg->count_info & PGC_broken) && (owner = page_get_owner(pg)) )
+    if ( page_is_broken(pg) && (owner = page_get_owner(pg)) )
     {
         *status = PG_OFFLINE_AGAIN;
         domain_crash(owner);
@@ -1620,7 +1627,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
 
     old_info = mark_page_offline(pg, broken);
 
-    if ( page_state_is(pg, offlined) )
+    if ( page_is_offlined(pg) )
     {
         reserve_heap_page(pg);
 
@@ -1699,19 +1706,18 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
     do {
         ret = *status = 0;
 
-        if ( y & PGC_broken )
+        if ( pgc_is_broken(y) )
         {
             ret = -EINVAL;
-            *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
+            *status = PG_ONLINE_FAILED | PG_ONLINE_BROKEN;
             break;
         }
-
-        if ( (y & PGC_state) == PGC_state_offlined )
+        else if ( pgc_is(y, offlined) )
         {
             page_list_del(pg, &page_offlined_list);
             *status = PG_ONLINE_ONLINED;
         }
-        else if ( (y & PGC_state) == PGC_state_offlining )
+        else if ( pgc_is(y, offlining) )
         {
             *status = PG_ONLINE_ONLINED;
         }
@@ -1726,7 +1732,7 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
 
     spin_unlock(&heap_lock);
 
-    if ( (y & PGC_state) == PGC_state_offlined )
+    if ( pgc_is(y, offlined) )
         free_heap_pages(pg, 0, false);
 
     return ret;
@@ -1747,11 +1753,11 @@ int query_page_offline(mfn_t mfn, uint32_t *status)
 
     pg = mfn_to_page(mfn);
 
-    if ( page_state_is(pg, offlining) )
+    if ( page_is_offlining(pg) )
         *status |= PG_OFFLINE_STATUS_OFFLINE_PENDING;
-    if ( pg->count_info & PGC_broken )
+    if ( page_is_broken(pg) )
         *status |= PG_OFFLINE_STATUS_BROKEN;
-    if ( page_state_is(pg, offlined) )
+    if ( page_is_offlined(pg) )
         *status |= PG_OFFLINE_STATUS_OFFLINED;
 
     spin_unlock(&heap_lock);
@@ -2519,7 +2525,7 @@ __initcall(pagealloc_keyhandler_init);
 
 void scrub_one_page(struct page_info *pg)
 {
-    if ( unlikely(pg->count_info & PGC_broken) )
+    if ( unlikely(page_is_broken(pg)) )
         return;
 
 #ifndef NDEBUG
diff --git a/xen/include/asm-arm/mm.h b/xen/include/asm-arm/mm.h
index 7df91280bc..a877791d1c 100644
--- a/xen/include/asm-arm/mm.h
+++ b/xen/include/asm-arm/mm.h
@@ -108,21 +108,35 @@ struct page_info
   /* Page is Xen heap? */
 #define _PGC_xen_heap     PG_shift(2)
 #define PGC_xen_heap      PG_mask(1, 2)
-/* ... */
-/* Page is broken? */
-#define _PGC_broken       PG_shift(7)
-#define PGC_broken        PG_mask(1, 7)
- /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
-#define PGC_state         PG_mask(3, 9)
-#define PGC_state_inuse   PG_mask(0, 9)
-#define PGC_state_offlining PG_mask(1, 9)
-#define PGC_state_offlined PG_mask(2, 9)
-#define PGC_state_free    PG_mask(3, 9)
-#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
+ /*
+  * Mutually-exclusive page states:
+  * { inuse, offlining, offlined, free, broken_offlining, broken }
+  */
+#define PGC_state                  PG_mask(7, 9)
+#define PGC_state_inuse            PG_mask(0, 9)
+#define PGC_state_offlining        PG_mask(1, 9)
+#define PGC_state_offlined         PG_mask(2, 9)
+#define PGC_state_free             PG_mask(3, 9)
+#define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
+#define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
+
+#define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
+#define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
+
+#define pgc_is_broken(pgc)         (pgc_is(pgc, broken) || \
+                                    pgc_is(pgc, broken_offlining))
+#define pgc_is_offlined(pgc)       (pgc_is(pgc, offlined) || \
+                                    pgc_is(pgc, broken))
+#define pgc_is_offlining(pgc)      (pgc_is(pgc, offlining) || \
+                                    pgc_is(pgc, broken_offlining))
+
+#define page_is_broken(pg)         (pgc_is_broken((pg)->count_info))
+#define page_is_offlined(pg)       (pgc_is_broken((pg)->count_info))
+#define page_is_offlining(pg)      (pgc_is_broken((pg)->count_info))
+
 /* Page is not reference counted */
 #define _PGC_extra        PG_shift(10)
 #define PGC_extra         PG_mask(1, 10)
-
 /* Count of references to this frame. */
 #define PGC_count_width   PG_shift(10)
 #define PGC_count_mask    ((1UL<<PGC_count_width)-1)
diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
index a06b2fb81f..1203f1b179 100644
--- a/xen/include/asm-x86/mm.h
+++ b/xen/include/asm-x86/mm.h
@@ -67,16 +67,32 @@
  /* 3-bit PAT/PCD/PWT cache-attribute hint. */
 #define PGC_cacheattr_base PG_shift(6)
 #define PGC_cacheattr_mask PG_mask(7, 6)
- /* Page is broken? */
-#define _PGC_broken       PG_shift(7)
-#define PGC_broken        PG_mask(1, 7)
- /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
-#define PGC_state         PG_mask(3, 9)
-#define PGC_state_inuse   PG_mask(0, 9)
-#define PGC_state_offlining PG_mask(1, 9)
-#define PGC_state_offlined PG_mask(2, 9)
-#define PGC_state_free    PG_mask(3, 9)
-#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
+ /*
+  * Mutually-exclusive page states:
+  * { inuse, offlining, offlined, free, broken_offlining, broken }
+  */
+#define PGC_state                  PG_mask(7, 9)
+#define PGC_state_inuse            PG_mask(0, 9)
+#define PGC_state_offlining        PG_mask(1, 9)
+#define PGC_state_offlined         PG_mask(2, 9)
+#define PGC_state_free             PG_mask(3, 9)
+#define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
+#define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
+
+#define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
+#define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
+
+#define pgc_is_broken(pgc)         (pgc_is(pgc, broken) || \
+                                    pgc_is(pgc, broken_offlining))
+#define pgc_is_offlined(pgc)       (pgc_is(pgc, offlined) || \
+                                    pgc_is(pgc, broken))
+#define pgc_is_offlining(pgc)      (pgc_is(pgc, offlining) || \
+                                    pgc_is(pgc, broken_offlining))
+
+#define page_is_broken(pg)         (pgc_is_broken((pg)->count_info))
+#define page_is_offlined(pg)       (pgc_is_broken((pg)->count_info))
+#define page_is_offlining(pg)      (pgc_is_broken((pg)->count_info))
+
 /* Page is not reference counted */
 #define _PGC_extra        PG_shift(10)
 #define PGC_extra         PG_mask(1, 10)
-- 
2.21.0


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

[Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

It is possible for pages to enter general circulation without ever
being process by init_heap_pages().

For example, pages of the multiboot module containing the initramfs may
be assigned via assign_pages() to dom0 as it is created. And some code
including map_pages_to_xen() has checks on 'system_state' to determine
whether to use the boot or the heap allocator, but it seems impossible
to prove that pages allocated by the boot allocator are not subsequently
freed with free_heap_pages().

This actually works fine in the majority of cases; there are only a few
esoteric corner cases which init_heap_pages() handles before handing the
page range off to free_heap_pages():
 • Excluding MFN #0 to avoid inappropriate cross-zone merging.
 • Ensuring that the node information structures exist, when the first
   page(s) of a given node are handled.
 • High order allocations crossing from one node to another.

To handle this case, shift PG_state_inuse from its current value of
zero, to another value. Use zero, which is the initial state of the
entire frame table, as PG_state_uninitialised.

Fix a couple of assertions which were assuming that PG_state_inuse is
zero, and make them cope with the PG_state_uninitialised case too where
appopriate.

Finally, make free_heap_pages() call through to init_heap_pages() when
given a page range which has not been initialised. This cannot keep
recursing because init_heap_pages() will set each page state to
PGC_state_inuse before passing it back to free_heap_pages() for the
second time.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 xen/arch/x86/mm.c        |  3 ++-
 xen/common/page_alloc.c  | 44 +++++++++++++++++++++++++++++-----------
 xen/include/asm-arm/mm.h |  3 ++-
 xen/include/asm-x86/mm.h |  3 ++-
 4 files changed, 38 insertions(+), 15 deletions(-)

diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
index 62507ca651..5f0581c072 100644
--- a/xen/arch/x86/mm.c
+++ b/xen/arch/x86/mm.c
@@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
 
     page_set_owner(page, d);
     smp_wmb(); /* install valid domain ptr before updating refcnt. */
-    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
+    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
+           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);
 
     /* Only add to the allocation list if the domain isn't dying. */
     if ( !d->is_dying )
diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
index 8d72a64f4e..4f7971f2a1 100644
--- a/xen/common/page_alloc.c
+++ b/xen/common/page_alloc.c
@@ -252,6 +252,8 @@ struct bootmem_region {
 static struct bootmem_region __initdata
     bootmem_region_list[PAGE_SIZE / sizeof(struct bootmem_region)];
 static unsigned int __initdata nr_bootmem_regions;
+static void init_heap_pages(struct page_info *pg, unsigned long nr_pages,
+                            bool scrub);
 
 struct scrub_region {
     unsigned long offset;
@@ -1390,6 +1392,17 @@ static void free_heap_pages(
     ASSERT(order <= MAX_ORDER);
     ASSERT(node >= 0);
 
+    if ( page_state_is(pg, uninitialised) )
+    {
+        init_heap_pages(pg, 1 << order, need_scrub);
+        /*
+         * init_heap_pages() will call back into free_heap_pages() for
+         * each page but cannot keep recursing because each page will
+         * be set to PGC_state_inuse first.
+         */
+        return;
+    }
+
     spin_lock(&heap_lock);
 
     for ( i = 0; i < (1 << order); i++ )
@@ -1771,11 +1784,10 @@ int query_page_offline(mfn_t mfn, uint32_t *status)
  * latter is not on a MAX_ORDER boundary, then we reserve the page by
  * not freeing it to the buddy allocator.
  */
-static void init_heap_pages(
-    struct page_info *pg, unsigned long nr_pages)
+static void init_heap_pages(struct page_info *pg, unsigned long nr_pages,
+                            bool scrub)
 {
     unsigned long i;
-    bool idle_scrub = false;
 
     /*
      * Keep MFN 0 away from the buddy allocator to avoid crossing zone
@@ -1800,7 +1812,7 @@ static void init_heap_pages(
     spin_unlock(&heap_lock);
 
     if ( system_state < SYS_STATE_active && opt_bootscrub == BOOTSCRUB_IDLE )
-        idle_scrub = true;
+        scrub = true;
 
     for ( i = 0; i < nr_pages; i++ )
     {
@@ -1828,7 +1840,8 @@ static void init_heap_pages(
             nr_pages -= n;
         }
 
-        free_heap_pages(pg + i, 0, scrub_debug || idle_scrub);
+        pg[i].count_info = PGC_state_inuse;
+        free_heap_pages(pg + i, 0, scrub_debug || scrub);
     }
 }
 
@@ -1864,7 +1877,7 @@ void __init end_boot_allocator(void)
         if ( (r->s < r->e) &&
              (phys_to_nid(pfn_to_paddr(r->s)) == cpu_to_node(0)) )
         {
-            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s);
+            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s, false);
             r->e = r->s;
             break;
         }
@@ -1873,7 +1886,7 @@ void __init end_boot_allocator(void)
     {
         struct bootmem_region *r = &bootmem_region_list[i];
         if ( r->s < r->e )
-            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s);
+            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s, false);
     }
     nr_bootmem_regions = 0;
 
@@ -2142,7 +2155,7 @@ void init_xenheap_pages(paddr_t ps, paddr_t pe)
 
     memguard_guard_range(maddr_to_virt(ps), pe - ps);
 
-    init_heap_pages(maddr_to_page(ps), (pe - ps) >> PAGE_SHIFT);
+    init_heap_pages(maddr_to_page(ps), (pe - ps) >> PAGE_SHIFT, false);
 }
 
 
@@ -2251,7 +2264,7 @@ void init_domheap_pages(paddr_t ps, paddr_t pe)
     if ( mfn_x(emfn) <= mfn_x(smfn) )
         return;
 
-    init_heap_pages(mfn_to_page(smfn), mfn_x(emfn) - mfn_x(smfn));
+    init_heap_pages(mfn_to_page(smfn), mfn_x(emfn) - mfn_x(smfn), false);
 }
 
 
@@ -2280,7 +2293,8 @@ int assign_pages(
 
         for ( i = 0; i < (1ul << order); i++ )
         {
-            ASSERT(!(pg[i].count_info & ~PGC_extra));
+            ASSERT((pg[i].count_info & ~PGC_extra) == PGC_state_inuse ||
+                   (pg[i].count_info & ~PGC_extra) == PGC_state_uninitialised);
             if ( pg[i].count_info & PGC_extra )
                 extra_pages++;
         }
@@ -2316,10 +2330,16 @@ int assign_pages(
     for ( i = 0; i < (1 << order); i++ )
     {
         ASSERT(page_get_owner(&pg[i]) == NULL);
+        /*
+         * Note: Not using page_state_is() here. The ASSERT requires that
+         * all other bits in count_info are zero, in addition to PGC_state
+         * being appropriate.
+         */
+        ASSERT((pg[i].count_info & ~PGC_extra) == PGC_state_inuse ||
+               (pg[i].count_info & ~PGC_extra) == PGC_state_uninitialised);
         page_set_owner(&pg[i], d);
         smp_wmb(); /* Domain pointer must be visible before updating refcnt. */
-        pg[i].count_info =
-            (pg[i].count_info & PGC_extra) | PGC_allocated | 1;
+        pg[i].count_info = (pg[i].count_info & PGC_state) | PGC_allocated | 1;
         page_list_add_tail(&pg[i], &d->page_list);
     }
 
diff --git a/xen/include/asm-arm/mm.h b/xen/include/asm-arm/mm.h
index a877791d1c..49663fa98a 100644
--- a/xen/include/asm-arm/mm.h
+++ b/xen/include/asm-arm/mm.h
@@ -113,12 +113,13 @@ struct page_info
   * { inuse, offlining, offlined, free, broken_offlining, broken }
   */
 #define PGC_state                  PG_mask(7, 9)
-#define PGC_state_inuse            PG_mask(0, 9)
+#define PGC_state_uninitialised    PG_mask(0, 9)
 #define PGC_state_offlining        PG_mask(1, 9)
 #define PGC_state_offlined         PG_mask(2, 9)
 #define PGC_state_free             PG_mask(3, 9)
 #define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
 #define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
+#define PGC_state_inuse            PG_mask(6, 9)
 
 #define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
 #define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
index 1203f1b179..5fbbca5f05 100644
--- a/xen/include/asm-x86/mm.h
+++ b/xen/include/asm-x86/mm.h
@@ -72,12 +72,13 @@
   * { inuse, offlining, offlined, free, broken_offlining, broken }
   */
 #define PGC_state                  PG_mask(7, 9)
-#define PGC_state_inuse            PG_mask(0, 9)
+#define PGC_state_uninitialised    PG_mask(0, 9)
 #define PGC_state_offlining        PG_mask(1, 9)
 #define PGC_state_offlined         PG_mask(2, 9)
 #define PGC_state_free             PG_mask(3, 9)
 #define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
 #define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
+#define PGC_state_inuse            PG_mask(6, 9)
 
 #define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
 #define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
-- 
2.21.0


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Paul Durrant]

> -----Original Message-----
> From: Xen-devel <xen-devel-bounces@lists.xenproject.org> On Behalf Of David Woodhouse
> Sent: 19 March 2020 21:22
> To: xen-devel@lists.xenproject.org
> Cc: Stefano Stabellini <sstabellini@kernel.org>; Julien Grall <julien@xen.org>; Wei Liu <wl@xen.org>;
> Andrew Cooper <andrew.cooper3@citrix.com>; Ian Jackson <ian.jackson@eu.citrix.com>; George Dunlap
> <george.dunlap@citrix.com>; hongyxia@amazon.com; Jan Beulich <jbeulich@suse.com>; Volodymyr Babchuk
> <Volodymyr_Babchuk@epam.com>; Roger Pau Monné <roger.pau@citrix.com>
> Subject: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised
> 
> From: David Woodhouse <dwmw@amazon.co.uk>
> 
> It is possible for pages to enter general circulation without ever
> being process by init_heap_pages().
> 
> For example, pages of the multiboot module containing the initramfs may
> be assigned via assign_pages() to dom0 as it is created. And some code
> including map_pages_to_xen() has checks on 'system_state' to determine
> whether to use the boot or the heap allocator, but it seems impossible
> to prove that pages allocated by the boot allocator are not subsequently
> freed with free_heap_pages().
> 
> This actually works fine in the majority of cases; there are only a few
> esoteric corner cases which init_heap_pages() handles before handing the
> page range off to free_heap_pages():
>  • Excluding MFN #0 to avoid inappropriate cross-zone merging.
>  • Ensuring that the node information structures exist, when the first
>    page(s) of a given node are handled.
>  • High order allocations crossing from one node to another.
> 
> To handle this case, shift PG_state_inuse from its current value of
> zero, to another value. Use zero, which is the initial state of the
> entire frame table, as PG_state_uninitialised.
> 
> Fix a couple of assertions which were assuming that PG_state_inuse is
> zero, and make them cope with the PG_state_uninitialised case too where
> appopriate.
> 
> Finally, make free_heap_pages() call through to init_heap_pages() when
> given a page range which has not been initialised. This cannot keep
> recursing because init_heap_pages() will set each page state to
> PGC_state_inuse before passing it back to free_heap_pages() for the
> second time.
> 
> Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> ---
>  xen/arch/x86/mm.c        |  3 ++-
>  xen/common/page_alloc.c  | 44 +++++++++++++++++++++++++++++-----------
>  xen/include/asm-arm/mm.h |  3 ++-
>  xen/include/asm-x86/mm.h |  3 ++-
>  4 files changed, 38 insertions(+), 15 deletions(-)
> 
> diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
> index 62507ca651..5f0581c072 100644
> --- a/xen/arch/x86/mm.c
> +++ b/xen/arch/x86/mm.c
> @@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
> 
>      page_set_owner(page, d);
>      smp_wmb(); /* install valid domain ptr before updating refcnt. */
> -    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
> +    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
> +           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);

Could the page state perhaps be bumped to inuse in this case? It seems odd to leave state uninitialized yet succeed in sharing with a guest.

> 
>      /* Only add to the allocation list if the domain isn't dying. */
>      if ( !d->is_dying )
> diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
> index 8d72a64f4e..4f7971f2a1 100644
> --- a/xen/common/page_alloc.c
> +++ b/xen/common/page_alloc.c
> @@ -252,6 +252,8 @@ struct bootmem_region {
>  static struct bootmem_region __initdata
>      bootmem_region_list[PAGE_SIZE / sizeof(struct bootmem_region)];
>  static unsigned int __initdata nr_bootmem_regions;
> +static void init_heap_pages(struct page_info *pg, unsigned long nr_pages,
> +                            bool scrub);
> 
>  struct scrub_region {
>      unsigned long offset;
> @@ -1390,6 +1392,17 @@ static void free_heap_pages(
>      ASSERT(order <= MAX_ORDER);
>      ASSERT(node >= 0);
> 
> +    if ( page_state_is(pg, uninitialised) )
> +    {
> +        init_heap_pages(pg, 1 << order, need_scrub);
> +        /*
> +         * init_heap_pages() will call back into free_heap_pages() for
> +         * each page but cannot keep recursing because each page will
> +         * be set to PGC_state_inuse first.
> +         */
> +        return;
> +    }
> +
>      spin_lock(&heap_lock);
> 
>      for ( i = 0; i < (1 << order); i++ )
> @@ -1771,11 +1784,10 @@ int query_page_offline(mfn_t mfn, uint32_t *status)
>   * latter is not on a MAX_ORDER boundary, then we reserve the page by
>   * not freeing it to the buddy allocator.
>   */
> -static void init_heap_pages(
> -    struct page_info *pg, unsigned long nr_pages)
> +static void init_heap_pages(struct page_info *pg, unsigned long nr_pages,
> +                            bool scrub)
>  {
>      unsigned long i;
> -    bool idle_scrub = false;
> 
>      /*
>       * Keep MFN 0 away from the buddy allocator to avoid crossing zone
> @@ -1800,7 +1812,7 @@ static void init_heap_pages(
>      spin_unlock(&heap_lock);
> 
>      if ( system_state < SYS_STATE_active && opt_bootscrub == BOOTSCRUB_IDLE )
> -        idle_scrub = true;
> +        scrub = true;
> 
>      for ( i = 0; i < nr_pages; i++ )
>      {
> @@ -1828,7 +1840,8 @@ static void init_heap_pages(
>              nr_pages -= n;
>          }
> 
> -        free_heap_pages(pg + i, 0, scrub_debug || idle_scrub);

Would it be worth an ASSERT(!pg[i].count_info) here in case something is added which erroneously modifies the page count info before this is done?

> +        pg[i].count_info = PGC_state_inuse;
> +        free_heap_pages(pg + i, 0, scrub_debug || scrub);
>      }
>  }
> 
> @@ -1864,7 +1877,7 @@ void __init end_boot_allocator(void)
>          if ( (r->s < r->e) &&
>               (phys_to_nid(pfn_to_paddr(r->s)) == cpu_to_node(0)) )
>          {
> -            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s);
> +            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s, false);
>              r->e = r->s;
>              break;
>          }
> @@ -1873,7 +1886,7 @@ void __init end_boot_allocator(void)
>      {
>          struct bootmem_region *r = &bootmem_region_list[i];
>          if ( r->s < r->e )
> -            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s);
> +            init_heap_pages(mfn_to_page(_mfn(r->s)), r->e - r->s, false);
>      }
>      nr_bootmem_regions = 0;
> 
> @@ -2142,7 +2155,7 @@ void init_xenheap_pages(paddr_t ps, paddr_t pe)
> 
>      memguard_guard_range(maddr_to_virt(ps), pe - ps);
> 
> -    init_heap_pages(maddr_to_page(ps), (pe - ps) >> PAGE_SHIFT);
> +    init_heap_pages(maddr_to_page(ps), (pe - ps) >> PAGE_SHIFT, false);
>  }
> 
> 
> @@ -2251,7 +2264,7 @@ void init_domheap_pages(paddr_t ps, paddr_t pe)
>      if ( mfn_x(emfn) <= mfn_x(smfn) )
>          return;
> 
> -    init_heap_pages(mfn_to_page(smfn), mfn_x(emfn) - mfn_x(smfn));
> +    init_heap_pages(mfn_to_page(smfn), mfn_x(emfn) - mfn_x(smfn), false);
>  }
> 
> 
> @@ -2280,7 +2293,8 @@ int assign_pages(
> 
>          for ( i = 0; i < (1ul << order); i++ )
>          {
> -            ASSERT(!(pg[i].count_info & ~PGC_extra));
> +            ASSERT((pg[i].count_info & ~PGC_extra) == PGC_state_inuse ||
> +                   (pg[i].count_info & ~PGC_extra) == PGC_state_uninitialised);

Again, perhaps bump the state to inuse if it is uninitialized...

>              if ( pg[i].count_info & PGC_extra )
>                  extra_pages++;
>          }
> @@ -2316,10 +2330,16 @@ int assign_pages(
>      for ( i = 0; i < (1 << order); i++ )
>      {
>          ASSERT(page_get_owner(&pg[i]) == NULL);
> +        /*
> +         * Note: Not using page_state_is() here. The ASSERT requires that
> +         * all other bits in count_info are zero, in addition to PGC_state
> +         * being appropriate.
> +         */
> +        ASSERT((pg[i].count_info & ~PGC_extra) == PGC_state_inuse ||
> +               (pg[i].count_info & ~PGC_extra) == PGC_state_uninitialised);

...then this ASSERT can be tightened.

>          page_set_owner(&pg[i], d);
>          smp_wmb(); /* Domain pointer must be visible before updating refcnt. */
> -        pg[i].count_info =
> -            (pg[i].count_info & PGC_extra) | PGC_allocated | 1;
> +        pg[i].count_info = (pg[i].count_info & PGC_state) | PGC_allocated | 1;

The PGC_extra seems to have vapourized here.

  Paul

>          page_list_add_tail(&pg[i], &d->page_list);
>      }
> 
> diff --git a/xen/include/asm-arm/mm.h b/xen/include/asm-arm/mm.h
> index a877791d1c..49663fa98a 100644
> --- a/xen/include/asm-arm/mm.h
> +++ b/xen/include/asm-arm/mm.h
> @@ -113,12 +113,13 @@ struct page_info
>    * { inuse, offlining, offlined, free, broken_offlining, broken }
>    */
>  #define PGC_state                  PG_mask(7, 9)
> -#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_uninitialised    PG_mask(0, 9)
>  #define PGC_state_offlining        PG_mask(1, 9)
>  #define PGC_state_offlined         PG_mask(2, 9)
>  #define PGC_state_free             PG_mask(3, 9)
>  #define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
>  #define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
> +#define PGC_state_inuse            PG_mask(6, 9)
> 
>  #define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
>  #define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
> diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
> index 1203f1b179..5fbbca5f05 100644
> --- a/xen/include/asm-x86/mm.h
> +++ b/xen/include/asm-x86/mm.h
> @@ -72,12 +72,13 @@
>    * { inuse, offlining, offlined, free, broken_offlining, broken }
>    */
>  #define PGC_state                  PG_mask(7, 9)
> -#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_uninitialised    PG_mask(0, 9)
>  #define PGC_state_offlining        PG_mask(1, 9)
>  #define PGC_state_offlined         PG_mask(2, 9)
>  #define PGC_state_free             PG_mask(3, 9)
>  #define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
>  #define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
> +#define PGC_state_inuse            PG_mask(6, 9)
> 
>  #define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
>  #define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
> --
> 2.21.0
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xenproject.org
> https://lists.xenproject.org/mailman/listinfo/xen-devel


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Jan Beulich]

On 20.03.2020 14:33, Paul Durrant wrote:
>> -----Original Message-----
>> From: Xen-devel <xen-devel-bounces@lists.xenproject.org> On Behalf Of David Woodhouse
>> Sent: 19 March 2020 21:22
>>
>> --- a/xen/arch/x86/mm.c
>> +++ b/xen/arch/x86/mm.c
>> @@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
>>
>>      page_set_owner(page, d);
>>      smp_wmb(); /* install valid domain ptr before updating refcnt. */
>> -    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
>> +    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
>> +           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);
> 
> Could the page state perhaps be bumped to inuse in this case? It
> seems odd to leave state uninitialized yet succeed in sharing with a guest.

This would be quite nice indeed, if of course it doesn't cause
new complications.

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 5010 bytes --]

On Fri, 2020-03-20 at 13:33 +0000, Paul Durrant wrote:
> > -----Original Message-----
> > From: Xen-devel <xen-devel-bounces@lists.xenproject.org> On Behalf Of David Woodhouse
> > Sent: 19 March 2020 21:22
> > To: xen-devel@lists.xenproject.org
> > Cc: Stefano Stabellini <sstabellini@kernel.org>; Julien Grall <julien@xen.org>; Wei Liu <wl@xen.org>;
> > Andrew Cooper <andrew.cooper3@citrix.com>; Ian Jackson <ian.jackson@eu.citrix.com>; George Dunlap
> > <george.dunlap@citrix.com>; hongyxia@amazon.com; Jan Beulich <jbeulich@suse.com>; Volodymyr Babchuk
> > <Volodymyr_Babchuk@epam.com>; Roger Pau Monné <roger.pau@citrix.com>
> > Subject: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised
> > 
> > From: David Woodhouse <dwmw@amazon.co.uk>
> > 
> > It is possible for pages to enter general circulation without ever
> > being process by init_heap_pages().
> > 
> > For example, pages of the multiboot module containing the initramfs may
> > be assigned via assign_pages() to dom0 as it is created. And some code
> > including map_pages_to_xen() has checks on 'system_state' to determine
> > whether to use the boot or the heap allocator, but it seems impossible
> > to prove that pages allocated by the boot allocator are not subsequently
> > freed with free_heap_pages().
> > 
> > This actually works fine in the majority of cases; there are only a few
> > esoteric corner cases which init_heap_pages() handles before handing the
> > page range off to free_heap_pages():
> >  • Excluding MFN #0 to avoid inappropriate cross-zone merging.
> >  • Ensuring that the node information structures exist, when the first
> >    page(s) of a given node are handled.
> >  • High order allocations crossing from one node to another.
> > 
> > To handle this case, shift PG_state_inuse from its current value of
> > zero, to another value. Use zero, which is the initial state of the
> > entire frame table, as PG_state_uninitialised.
> > 
> > Fix a couple of assertions which were assuming that PG_state_inuse is
> > zero, and make them cope with the PG_state_uninitialised case too where
> > appopriate.
> > 
> > Finally, make free_heap_pages() call through to init_heap_pages() when
> > given a page range which has not been initialised. This cannot keep
> > recursing because init_heap_pages() will set each page state to
> > PGC_state_inuse before passing it back to free_heap_pages() for the
> > second time.
> > 
> > Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> > ---
> >  xen/arch/x86/mm.c        |  3 ++-
> >  xen/common/page_alloc.c  | 44 +++++++++++++++++++++++++++++-----------
> >  xen/include/asm-arm/mm.h |  3 ++-
> >  xen/include/asm-x86/mm.h |  3 ++-
> >  4 files changed, 38 insertions(+), 15 deletions(-)
> > 
> > diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
> > index 62507ca651..5f0581c072 100644
> > --- a/xen/arch/x86/mm.c
> > +++ b/xen/arch/x86/mm.c
> > @@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
> > 
> >      page_set_owner(page, d);
> >      smp_wmb(); /* install valid domain ptr before updating refcnt. */
> > -    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
> > +    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
> > +           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);
> 
> Could the page state perhaps be bumped to inuse in this case? It
> seems odd to leave state uninitialized yet succeed in sharing with a
> guest.

No, that doesn't really work.

You can't just *declare* that the page was properly initialised,
because it isn't true. There's a pathological case where the zone
hasn't been initialised at all, because *all* the pages in that zone
got handed out by the boot allocator or used for initrd etc. 

The first pages 'freed' in that zone end up being used (in
init_heap_pages) to create the zone structures.

Likewise, it could include a page which init_heap_pages() doesn't
actually *put* into the buddy allocator, to work around the cross-zone
merge problem. It's fine to use that page and share it with a guest,
but it can't ever be freed into the buddy allocator.

> > @@ -1828,7 +1840,8 @@ static void init_heap_pages(
> >              nr_pages -= n;
> >          }
> > 
> > -        free_heap_pages(pg + i, 0, scrub_debug || idle_scrub);
> 
> Would it be worth an ASSERT(!pg[i].count_info) here in case something
> is added which erroneously modifies the page count info before this
> is done?

That seems valid, I think. Will test it.

> > 
> >          page_set_owner(&pg[i], d);
> >          smp_wmb(); /* Domain pointer must be visible before updating refcnt. */
> > -        pg[i].count_info =
> > -            (pg[i].count_info & PGC_extra) | PGC_allocated | 1;
> > +        pg[i].count_info = (pg[i].count_info & PGC_state) | PGC_allocated | 1;
> 
> The PGC_extra seems to have vapourized here.

Oops. Will fix; thanks.


[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Paul Durrant]

> -----Original Message-----
> > > diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
> > > index 62507ca651..5f0581c072 100644
> > > --- a/xen/arch/x86/mm.c
> > > +++ b/xen/arch/x86/mm.c
> > > @@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
> > >
> > >      page_set_owner(page, d);
> > >      smp_wmb(); /* install valid domain ptr before updating refcnt. */
> > > -    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
> > > +    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
> > > +           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);
> >
> > Could the page state perhaps be bumped to inuse in this case? It
> > seems odd to leave state uninitialized yet succeed in sharing with a
> > guest.
> 
> No, that doesn't really work.
> 
> You can't just *declare* that the page was properly initialised,
> because it isn't true. There's a pathological case where the zone
> hasn't been initialised at all, because *all* the pages in that zone
> got handed out by the boot allocator or used for initrd etc.
> 
> The first pages 'freed' in that zone end up being used (in
> init_heap_pages) to create the zone structures.
> 
> Likewise, it could include a page which init_heap_pages() doesn't
> actually *put* into the buddy allocator, to work around the cross-zone
> merge problem. It's fine to use that page and share it with a guest,
> but it can't ever be freed into the buddy allocator.
> 

Ok, so deferring the call to free_heap_pages() (and consequently init_heap_pages()) is safe to defer until the guest is torn down? (Or is this only safe if the page is being assigned to the initial domain?)

  Paul


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 763 bytes --]

On Mon, 2020-03-23 at 08:49 +0000, Paul Durrant wrote:
> Ok, so deferring the call to free_heap_pages() (and consequently
> init_heap_pages()) is safe to defer until the guest is torn down? (Or
> is this only safe if the page is being assigned to the initial
> domain?)

It's intended to be safe in all cases, including pages which are
allocated from the boot allocator to be used as page tables (cf. the
early-vmap patches), etc.

We kind of have to assume that it's safe to use the page for whatever
purpose it was allocated for, for the lifetime of that usage. If *that*
isn't true, we have bigger problems.

The PGC_state_uninitialised thing is only about recycling it into the
heap *later*, once the lifetime of that initial usage has ended.

[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Julien Grall]

Hi David,

On 20/03/2020 15:17, David Woodhouse wrote:
> On Fri, 2020-03-20 at 13:33 +0000, Paul Durrant wrote:
>>> -----Original Message-----
>>> From: Xen-devel <xen-devel-bounces@lists.xenproject.org> On Behalf Of David Woodhouse
>>> Sent: 19 March 2020 21:22
>>> To: xen-devel@lists.xenproject.org
>>> Cc: Stefano Stabellini <sstabellini@kernel.org>; Julien Grall <julien@xen.org>; Wei Liu <wl@xen.org>;
>>> Andrew Cooper <andrew.cooper3@citrix.com>; Ian Jackson <ian.jackson@eu.citrix.com>; George Dunlap
>>> <george.dunlap@citrix.com>; hongyxia@amazon.com; Jan Beulich <jbeulich@suse.com>; Volodymyr Babchuk
>>> <Volodymyr_Babchuk@epam.com>; Roger Pau Monné <roger.pau@citrix.com>
>>> Subject: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised
>>>
>>> From: David Woodhouse <dwmw@amazon.co.uk>
>>>
>>> It is possible for pages to enter general circulation without ever
>>> being process by init_heap_pages().
>>>
>>> For example, pages of the multiboot module containing the initramfs may
>>> be assigned via assign_pages() to dom0 as it is created. And some code
>>> including map_pages_to_xen() has checks on 'system_state' to determine
>>> whether to use the boot or the heap allocator, but it seems impossible
>>> to prove that pages allocated by the boot allocator are not subsequently
>>> freed with free_heap_pages().
>>>
>>> This actually works fine in the majority of cases; there are only a few
>>> esoteric corner cases which init_heap_pages() handles before handing the
>>> page range off to free_heap_pages():
>>>   • Excluding MFN #0 to avoid inappropriate cross-zone merging.
>>>   • Ensuring that the node information structures exist, when the first
>>>     page(s) of a given node are handled.
>>>   • High order allocations crossing from one node to another.
>>>
>>> To handle this case, shift PG_state_inuse from its current value of
>>> zero, to another value. Use zero, which is the initial state of the
>>> entire frame table, as PG_state_uninitialised.
>>>
>>> Fix a couple of assertions which were assuming that PG_state_inuse is
>>> zero, and make them cope with the PG_state_uninitialised case too where
>>> appopriate.
>>>
>>> Finally, make free_heap_pages() call through to init_heap_pages() when
>>> given a page range which has not been initialised. This cannot keep
>>> recursing because init_heap_pages() will set each page state to
>>> PGC_state_inuse before passing it back to free_heap_pages() for the
>>> second time.
>>>
>>> Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
>>> ---
>>>   xen/arch/x86/mm.c        |  3 ++-
>>>   xen/common/page_alloc.c  | 44 +++++++++++++++++++++++++++++-----------
>>>   xen/include/asm-arm/mm.h |  3 ++-
>>>   xen/include/asm-x86/mm.h |  3 ++-
>>>   4 files changed, 38 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
>>> index 62507ca651..5f0581c072 100644
>>> --- a/xen/arch/x86/mm.c
>>> +++ b/xen/arch/x86/mm.c
>>> @@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
>>>
>>>       page_set_owner(page, d);
>>>       smp_wmb(); /* install valid domain ptr before updating refcnt. */
>>> -    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
>>> +    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
>>> +           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);
>>
>> Could the page state perhaps be bumped to inuse in this case? It
>> seems odd to leave state uninitialized yet succeed in sharing with a
>> guest.
> 
> No, that doesn't really work.
> 
> You can't just *declare* that the page was properly initialised,
> because it isn't true. There's a pathological case where the zone
> hasn't been initialised at all, because *all* the pages in that zone
> got handed out by the boot allocator or used for initrd etc.
> 
> The first pages 'freed' in that zone end up being used (in
> init_heap_pages) to create the zone structures.
> 
> Likewise, it could include a page which init_heap_pages() doesn't
> actually *put* into the buddy allocator, to work around the cross-zone
> merge problem. It's fine to use that page and share it with a guest,
> but it can't ever be freed into the buddy allocator.

For liveupdate, we will need a way to initialize a page but mark it as 
already inuse (i.e in the same state as they would be if allocated 
normally).

It feels to me, this is also what we want in this case. The page would 
be first initialize and then we can use it normally including freeing 
later on.

Would it make sense to introduce an helper for this purpose?

Cheers,

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 1726 bytes --]

On Mon, 2020-03-23 at 09:34 +0000, Julien Grall wrote:
> For liveupdate, we will need a way to initialize a page but mark it as 
> already inuse (i.e in the same state as they would be if allocated 
> normally).

I am unconvinced of the veracity of this claim.

We don't want to turn specific details of the current Xen buddy
allocator part into of the implicit ABI of live update. That goes for
the power-of-two zone boundaries, amongst other things.

What if Xen receives LU state in which *all* pages in a given zone are
marked as already in use? That's one of the cases in which we *really*
want to pass through init_heap_pages() instead of just
free_heap_pages(), in order to allocate the zone data structures for
the first pages that get freed into that zone.

What if Xen starts to exclude more pages, like the exclusion at zero?

What if new Xen wants to exclude an additional page due to a hardware
erratum? It can't take it away from existing domains (especially if
there are assigned PCI devices) but it could be part of the vetting in
init_heap_pages(), for example.

My intent for PGC_state_uninitialised was to mark pages that haven't
been through init_heap_pages(), whatever init_heap_pages() does in the
current version of Xen.

The pages which are "already in use" because they're inherited through
LU state should be in PGC_state_uninitialised, shouldn't they?

Perhaps if there's a need for a helper, it could be a companion
function to init_heap_pages() which would return a boolean saying,
"nah, I didn't want to do anything to this page anyway", which could
short-circuit it into the PGC_state_inuse state. But I'm not sure I see
the need for such an optimisation. 


[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Julien Grall]

Hi David,

On 23/03/2020 10:55, David Woodhouse wrote:
> On Mon, 2020-03-23 at 09:34 +0000, Julien Grall wrote:
>> For liveupdate, we will need a way to initialize a page but mark it as
>> already inuse (i.e in the same state as they would be if allocated
>> normally).
> 
> I am unconvinced of the veracity of this claim.
> 
> We don't want to turn specific details of the current Xen buddy
> allocator part into of the implicit ABI of live update. That goes for
> the power-of-two zone boundaries, amongst other things.

Why would you to do that? Marking the page as already used is no 
different to "PGC_state_unitialized" except the "struct page_info" and 
the internal of the buddy allocator would be properly setup for start 
rather than at free.

> 
> What if Xen receives LU state in which *all* pages in a given zone are
> marked as already in use? That's one of the cases in which we *really*
> want to pass through init_heap_pages() instead of just
> free_heap_pages(), in order to allocate the zone data structures for
> the first pages that get freed into that zone.
> 
> What if Xen starts to exclude more pages, like the exclusion at zero?
> 
> What if new Xen wants to exclude an additional page due to a hardware
> erratum? It can't take it away from existing domains (especially if
> there are assigned PCI devices) but it could be part of the vetting in
> init_heap_pages(), for example.

I don't think it would be safe to continue to run a guest using pages 
that were excluded for an HW erratum. It would be safer to not restart 
the domain (or replace the page) in the target Xen if that's hapenning.

> 
> My intent for PGC_state_uninitialised was to mark pages that haven't
> been through init_heap_pages(), whatever init_heap_pages() does in the
> current version of Xen.
> 
> The pages which are "already in use" because they're inherited through
> LU state should be in PGC_state_uninitialised, shouldn't they?

I think using "PGC_state_unitialised" for preserved page is an abuse. I 
understand this is existing in other part of Xen (particularly on x86), 
but I would rather not try to add more.

The PGC_state_unitialised may work for the current allocator because 
most of the fields are not going to be used after allocation. But it may 
not hold for any new allocator (I know the embedded folks are working on 
a new one).

> 
> Perhaps if there's a need for a helper, it could be a companion
> function to init_heap_pages() which would return a boolean saying,
> "nah, I didn't want to do anything to this page anyway", which could
> short-circuit it into the PGC_state_inuse state. But I'm not sure I see
> the need for such an optimisation.

I don't view it as an optimisation but as a way to avoid spreading the 
current misbehavior.

Cheers,

-- 
Julien Grall

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[David Woodhouse]

[-- Attachment #1: Type: text/plain, Size: 2038 bytes --]

On Tue, 2020-03-24 at 10:08 +0000, Julien Grall wrote:
> Hi David,
> 
> On 23/03/2020 10:55, David Woodhouse wrote:
> > On Mon, 2020-03-23 at 09:34 +0000, Julien Grall wrote:
> > > For liveupdate, we will need a way to initialize a page but mark it as
> > > already inuse (i.e in the same state as they would be if allocated
> > > normally).
> > 
> > I am unconvinced of the veracity of this claim.
> > 
> > We don't want to turn specific details of the current Xen buddy
> > allocator part into of the implicit ABI of live update. That goes for
> > the power-of-two zone boundaries, amongst other things.
> 
> Why would you to do that? Marking the page as already used is no 
> different to "PGC_state_unitialized" except the "struct page_info" and 
> the internal of the buddy allocator would be properly setup for start 
> rather than at free.

The internals of the buddy allocator *cannot* be set up properly for a
page which it would not actually give out — like the zero page.

We *could* do some tricks to allocate the zone structures for zones
which *do* exist but contain only "pre-allocated" pages so the buddy
allocator has never seen those zones... yet.


> I think using "PGC_state_unitialised" for preserved page is an abuse. I 
> understand this is existing in other part of Xen (particularly on x86), 
> but I would rather not try to add more.


I am perfectly happy to try avoiding PGC_state_uninitialised for pages
which are "in use" at boot time due to live update.

All I insist on is that you explicitly describe the ABI constraints
that it imposes, if any.

For example, that hack which stops the buddy allocator from giving out
page zero: Could we have live updated from a Xen without that hack, to
a Xen which has it? With page zero already given out to a domain?

With PGC_state_initialised and passing the page through
init_heap_pages() if/when it ever gets freed, my answer would be 'yes'.

What's yours? How would we cope with a situation like that, over LU?


[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

Re: [Xen-devel] [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Julien Grall]

Hi David,

On 24/03/2020 17:55, David Woodhouse wrote:
> On Tue, 2020-03-24 at 10:08 +0000, Julien Grall wrote:
>> Hi David,
>>
>> On 23/03/2020 10:55, David Woodhouse wrote:
>>> On Mon, 2020-03-23 at 09:34 +0000, Julien Grall wrote:
>>>> For liveupdate, we will need a way to initialize a page but mark it as
>>>> already inuse (i.e in the same state as they would be if allocated
>>>> normally).
>>>
>>> I am unconvinced of the veracity of this claim.
>>>
>>> We don't want to turn specific details of the current Xen buddy
>>> allocator part into of the implicit ABI of live update. That goes for
>>> the power-of-two zone boundaries, amongst other things.
>>
>> Why would you to do that? Marking the page as already used is no
>> different to "PGC_state_unitialized" except the "struct page_info" and
>> the internal of the buddy allocator would be properly setup for start
>> rather than at free.
> 
> The internals of the buddy allocator *cannot* be set up properly for a
> page which it would not actually give out — like the zero page.
> 
> We *could* do some tricks to allocate the zone structures for zones
> which *do* exist but contain only "pre-allocated" pages so the buddy
> allocator has never seen those zones... yet.
> 
> 
>> I think using "PGC_state_unitialised" for preserved page is an abuse. I
>> understand this is existing in other part of Xen (particularly on x86),
>> but I would rather not try to add more.
> 
> 
> I am perfectly happy to try avoiding PGC_state_uninitialised for pages
> which are "in use" at boot time due to live update.
> 
> All I insist on is that you explicitly describe the ABI constraints
> that it imposes, if any.

Agreed.

> 
> For example, that hack which stops the buddy allocator from giving out
> page zero: Could we have live updated from a Xen without that hack, to
> a Xen which has it? With page zero already given out to a domain?

The buddy allocator could never have given out page 0 on x86 because it 
is part of the first MB of the RAM (see arch_init_memory() in 
arch/x86/mm.c). AFAIK, the first MB cannot be freed..

The change in the buddy allocator was to address the Arm side and also 
make clear this was a problem this is a weakness of the allocator.

> What's yours? How would we cope with a situation like that, over LU?

When do you expect the pages to be carved out from the buddy allocator?

I can see only two situations:
	1) Workaround a bug in the allocator.
         2) A CPU errata requiring to not use memory.

In the case of 1), it is still safe for a domain to run with that page. 
However, we don't want to give it back to the page allocator. A solution 
is to mark them as "offlining/broken". Alternatively, you could remove 
the swap the page (see more below).

In the case of 2), it is not safe for a domain to run with that page. So 
it is probably best to swap the pages with a new one. For HVM domain 
(including the P2M), it should be easy. For PV domain, I am not entirely 
sure if that's feasible.

Cheers,

-- 
Julien Grall

Re: [PATCH 2/2] xen/mm: Introduce PGC_state_uninitialised

[Jan Beulich]

On 19.03.2020 22:21, David Woodhouse wrote:
> --- a/xen/arch/x86/mm.c
> +++ b/xen/arch/x86/mm.c
> @@ -491,7 +491,8 @@ void share_xen_page_with_guest(struct page_info *page, struct domain *d,
>  
>      page_set_owner(page, d);
>      smp_wmb(); /* install valid domain ptr before updating refcnt. */
> -    ASSERT((page->count_info & ~PGC_xen_heap) == 0);
> +    ASSERT((page->count_info & ~PGC_xen_heap) == PGC_state_inuse ||
> +           (page->count_info & ~PGC_xen_heap) == PGC_state_uninitialised);

Like for patch 1, there's a risk of the page transitioning from
uninitialised to inuse between these two comparisons, making the
ASSERT() trigger when it shouldn't. As you've got two more
similar constructs further down in the patch, maybe this also
warrants a helper function/macro?

> --- a/xen/common/page_alloc.c
> +++ b/xen/common/page_alloc.c
> @@ -252,6 +252,8 @@ struct bootmem_region {
>  static struct bootmem_region __initdata
>      bootmem_region_list[PAGE_SIZE / sizeof(struct bootmem_region)];
>  static unsigned int __initdata nr_bootmem_regions;
> +static void init_heap_pages(struct page_info *pg, unsigned long nr_pages,
> +                            bool scrub);
>  
>  struct scrub_region {
>      unsigned long offset;
> @@ -1390,6 +1392,17 @@ static void free_heap_pages(
>      ASSERT(order <= MAX_ORDER);
>      ASSERT(node >= 0);
>  
> +    if ( page_state_is(pg, uninitialised) )
> +    {
> +        init_heap_pages(pg, 1 << order, need_scrub);

While, at least for now, it shouldn't matter in practice, and
while code overall is very inconsistent in this regard, with
the respective parameter type being "unsigned long" may I
suggest to use 1UL here?

Jan

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Paul Durrant]

> -----Original Message-----
> From: Xen-devel <xen-devel-bounces@lists.xenproject.org> On Behalf Of David Woodhouse
> Sent: 19 March 2020 21:22
> To: xen-devel@lists.xenproject.org
> Cc: Stefano Stabellini <sstabellini@kernel.org>; Julien Grall <julien@xen.org>; Wei Liu <wl@xen.org>;
> Andrew Cooper <andrew.cooper3@citrix.com>; Ian Jackson <ian.jackson@eu.citrix.com>; George Dunlap
> <george.dunlap@citrix.com>; hongyxia@amazon.com; Jan Beulich <jbeulich@suse.com>; Volodymyr Babchuk
> <Volodymyr_Babchuk@epam.com>; Roger Pau Monné <roger.pau@citrix.com>
> Subject: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits
> 
> From: David Woodhouse <dwmw@amazon.co.uk>
> 
> Only PGC_state_offlining and PGC_state_offlined are valid in conjunction
> with PGC_broken. The other two states (free and inuse) were never valid
> for a broken page.
> 
> By folding PGC_broken in, we can have three bits for PGC_state which
> allows up to 8 states, of which 6 are currently used and 2 are available
> for new use cases.
> 
> Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> ---
>  xen/arch/x86/domctl.c    |  2 +-
>  xen/common/page_alloc.c  | 66 ++++++++++++++++++++++------------------
>  xen/include/asm-arm/mm.h | 38 +++++++++++++++--------
>  xen/include/asm-x86/mm.h | 36 ++++++++++++++++------
>  4 files changed, 89 insertions(+), 53 deletions(-)
> 
> diff --git a/xen/arch/x86/domctl.c b/xen/arch/x86/domctl.c
> index ed86762fa6..a411f64afa 100644
> --- a/xen/arch/x86/domctl.c
> +++ b/xen/arch/x86/domctl.c
> @@ -422,7 +422,7 @@ long arch_do_domctl(
>                  if ( page->u.inuse.type_info & PGT_pinned )
>                      type |= XEN_DOMCTL_PFINFO_LPINTAB;
> 
> -                if ( page->count_info & PGC_broken )
> +                if ( page_is_broken(page) )
>                      type = XEN_DOMCTL_PFINFO_BROKEN;
>              }
> 
> diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
> index 76d37226df..8d72a64f4e 100644
> --- a/xen/common/page_alloc.c
> +++ b/xen/common/page_alloc.c
> @@ -1093,7 +1093,7 @@ static int reserve_offlined_page(struct page_info *head)
>          struct page_info *pg;
>          int next_order;
> 
> -        if ( page_state_is(cur_head, offlined) )
> +        if ( page_is_offlined(cur_head) )
>          {
>              cur_head++;
>              if ( first_dirty != INVALID_DIRTY_IDX && first_dirty )
> @@ -1113,7 +1113,7 @@ static int reserve_offlined_page(struct page_info *head)
>              for ( i = (1 << cur_order), pg = cur_head + (1 << cur_order );
>                    i < (1 << next_order);
>                    i++, pg++ )
> -                if ( page_state_is(pg, offlined) )
> +                if ( page_is_offlined(pg) )
>                      break;
>              if ( i == ( 1 << next_order) )
>              {
> @@ -1145,16 +1145,20 @@ static int reserve_offlined_page(struct page_info *head)
> 
>      for ( cur_head = head; cur_head < head + ( 1UL << head_order); cur_head++ )
>      {
> -        if ( !page_state_is(cur_head, offlined) )
> +        struct page_list_head *list;
> +
> +        if ( page_state_is(cur_head, offlined) )
> +            list = &page_offlined_list;
> +        else if (page_state_is(cur_head, broken) )
> +            list = &page_broken_list;
> +        else
>              continue;
> 
>          avail[node][zone]--;
>          total_avail_pages--;
>          ASSERT(total_avail_pages >= 0);
> 
> -        page_list_add_tail(cur_head,
> -                           test_bit(_PGC_broken, &cur_head->count_info) ?
> -                           &page_broken_list : &page_offlined_list);
> +        page_list_add_tail(cur_head, list);
> 
>          count++;
>      }
> @@ -1404,13 +1408,16 @@ static void free_heap_pages(
>          switch ( pg[i].count_info & PGC_state )
>          {
>          case PGC_state_inuse:
> -            BUG_ON(pg[i].count_info & PGC_broken);
>              pg[i].count_info = PGC_state_free;
>              break;
> 
>          case PGC_state_offlining:
> -            pg[i].count_info = (pg[i].count_info & PGC_broken) |
> -                               PGC_state_offlined;
> +            pg[i].count_info = PGC_state_offlined;
> +            tainted = 1;
> +            break;
> +
> +        case PGC_state_broken_offlining:
> +            pg[i].count_info = PGC_state_broken;
>              tainted = 1;
>              break;
> 
> @@ -1527,16 +1534,16 @@ static unsigned long mark_page_offline(struct page_info *pg, int broken)
>      do {
>          nx = x = y;
> 
> -        if ( ((x & PGC_state) != PGC_state_offlined) &&
> -             ((x & PGC_state) != PGC_state_offlining) )
> -        {
> -            nx &= ~PGC_state;
> -            nx |= (((x & PGC_state) == PGC_state_free)
> -                   ? PGC_state_offlined : PGC_state_offlining);
> -        }
> +        nx &= ~PGC_state;
> 
> -        if ( broken )
> -            nx |= PGC_broken;
> +        /* If it was already broken, it stays broken */
> +        if ( pgc_is_broken(x) )
> +            broken = 1;
> +
> +        if ( pgc_is_offlined(x) || pgc_is(x, free) )
> +            nx |= broken ? PGC_state_broken : PGC_state_offlined;
> +        else
> +            nx |= broken ? PGC_state_broken_offlining : PGC_state_offlining;
> 
>          if ( x == nx )
>              break;
> @@ -1609,7 +1616,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
>       * need to prevent malicious guest access the broken page again.
>       * Under such case, hypervisor shutdown guest, preventing recursive mce.
>       */
> -    if ( (pg->count_info & PGC_broken) && (owner = page_get_owner(pg)) )
> +    if ( page_is_broken(pg) && (owner = page_get_owner(pg)) )
>      {
>          *status = PG_OFFLINE_AGAIN;
>          domain_crash(owner);
> @@ -1620,7 +1627,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
> 
>      old_info = mark_page_offline(pg, broken);
> 
> -    if ( page_state_is(pg, offlined) )
> +    if ( page_is_offlined(pg) )
>      {
>          reserve_heap_page(pg);
> 
> @@ -1699,19 +1706,18 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
>      do {
>          ret = *status = 0;
> 
> -        if ( y & PGC_broken )
> +        if ( pgc_is_broken(y) )
>          {
>              ret = -EINVAL;
> -            *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
> +            *status = PG_ONLINE_FAILED | PG_ONLINE_BROKEN;

Whitespace fix. Ought to be called out in the commit comment.

>              break;
>          }
> -
> -        if ( (y & PGC_state) == PGC_state_offlined )
> +        else if ( pgc_is(y, offlined) )
>          {
>              page_list_del(pg, &page_offlined_list);
>              *status = PG_ONLINE_ONLINED;
>          }
> -        else if ( (y & PGC_state) == PGC_state_offlining )
> +        else if ( pgc_is(y, offlining) )
>          {
>              *status = PG_ONLINE_ONLINED;
>          }
> @@ -1726,7 +1732,7 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
> 
>      spin_unlock(&heap_lock);
> 
> -    if ( (y & PGC_state) == PGC_state_offlined )
> +    if ( pgc_is(y, offlined) )
>          free_heap_pages(pg, 0, false);
> 
>      return ret;
> @@ -1747,11 +1753,11 @@ int query_page_offline(mfn_t mfn, uint32_t *status)
> 
>      pg = mfn_to_page(mfn);
> 
> -    if ( page_state_is(pg, offlining) )
> +    if ( page_is_offlining(pg) )
>          *status |= PG_OFFLINE_STATUS_OFFLINE_PENDING;
> -    if ( pg->count_info & PGC_broken )
> +    if ( page_is_broken(pg) )
>          *status |= PG_OFFLINE_STATUS_BROKEN;
> -    if ( page_state_is(pg, offlined) )
> +    if ( page_is_offlined(pg) )
>          *status |= PG_OFFLINE_STATUS_OFFLINED;
> 
>      spin_unlock(&heap_lock);
> @@ -2519,7 +2525,7 @@ __initcall(pagealloc_keyhandler_init);
> 
>  void scrub_one_page(struct page_info *pg)
>  {
> -    if ( unlikely(pg->count_info & PGC_broken) )
> +    if ( unlikely(page_is_broken(pg)) )
>          return;
> 
>  #ifndef NDEBUG
> diff --git a/xen/include/asm-arm/mm.h b/xen/include/asm-arm/mm.h
> index 7df91280bc..a877791d1c 100644
> --- a/xen/include/asm-arm/mm.h
> +++ b/xen/include/asm-arm/mm.h
> @@ -108,21 +108,35 @@ struct page_info
>    /* Page is Xen heap? */
>  #define _PGC_xen_heap     PG_shift(2)
>  #define PGC_xen_heap      PG_mask(1, 2)
> -/* ... */
> -/* Page is broken? */
> -#define _PGC_broken       PG_shift(7)
> -#define PGC_broken        PG_mask(1, 7)
> - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> -#define PGC_state         PG_mask(3, 9)
> -#define PGC_state_inuse   PG_mask(0, 9)
> -#define PGC_state_offlining PG_mask(1, 9)
> -#define PGC_state_offlined PG_mask(2, 9)
> -#define PGC_state_free    PG_mask(3, 9)
> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> + /*
> +  * Mutually-exclusive page states:
> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> +  */
> +#define PGC_state                  PG_mask(7, 9)
> +#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_offlining        PG_mask(1, 9)
> +#define PGC_state_offlined         PG_mask(2, 9)
> +#define PGC_state_free             PG_mask(3, 9)
> +#define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
> +#define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
> +
> +#define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
> +#define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
> +
> +#define pgc_is_broken(pgc)         (pgc_is(pgc, broken) || \
> +                                    pgc_is(pgc, broken_offlining))
> +#define pgc_is_offlined(pgc)       (pgc_is(pgc, offlined) || \
> +                                    pgc_is(pgc, broken))
> +#define pgc_is_offlining(pgc)      (pgc_is(pgc, offlining) || \
> +                                    pgc_is(pgc, broken_offlining))
> +
> +#define page_is_broken(pg)         (pgc_is_broken((pg)->count_info))
> +#define page_is_offlined(pg)       (pgc_is_broken((pg)->count_info))
> +#define page_is_offlining(pg)      (pgc_is_broken((pg)->count_info))
> +
>  /* Page is not reference counted */
>  #define _PGC_extra        PG_shift(10)
>  #define PGC_extra         PG_mask(1, 10)
> -

Extraneous whitespace change.

>  /* Count of references to this frame. */
>  #define PGC_count_width   PG_shift(10)
>  #define PGC_count_mask    ((1UL<<PGC_count_width)-1)
> diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
> index a06b2fb81f..1203f1b179 100644
> --- a/xen/include/asm-x86/mm.h
> +++ b/xen/include/asm-x86/mm.h
> @@ -67,16 +67,32 @@
>   /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>  #define PGC_cacheattr_base PG_shift(6)
>  #define PGC_cacheattr_mask PG_mask(7, 6)
> - /* Page is broken? */
> -#define _PGC_broken       PG_shift(7)
> -#define PGC_broken        PG_mask(1, 7)
> - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> -#define PGC_state         PG_mask(3, 9)
> -#define PGC_state_inuse   PG_mask(0, 9)
> -#define PGC_state_offlining PG_mask(1, 9)
> -#define PGC_state_offlined PG_mask(2, 9)
> -#define PGC_state_free    PG_mask(3, 9)
> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> + /*
> +  * Mutually-exclusive page states:
> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> +  */
> +#define PGC_state                  PG_mask(7, 9)
> +#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_offlining        PG_mask(1, 9)
> +#define PGC_state_offlined         PG_mask(2, 9)
> +#define PGC_state_free             PG_mask(3, 9)
> +#define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
> +#define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
> +
> +#define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)

Maybe pgc_state_is() for consistency? Might also draw attention to the difference between e.g.:

pgc_is(pgc, offlined) and pgc_is_offlined(pgc) 

> +#define page_state_is(pg, st)       pgc_is((pg)->count_info, st)

Indentation looks wrong.

^^ Same for the arm code.

  Paul

> +
> +#define pgc_is_broken(pgc)         (pgc_is(pgc, broken) || \
> +                                    pgc_is(pgc, broken_offlining))
> +#define pgc_is_offlined(pgc)       (pgc_is(pgc, offlined) || \
> +                                    pgc_is(pgc, broken))
> +#define pgc_is_offlining(pgc)      (pgc_is(pgc, offlining) || \
> +                                    pgc_is(pgc, broken_offlining))
> +
> +#define page_is_broken(pg)         (pgc_is_broken((pg)->count_info))
> +#define page_is_offlined(pg)       (pgc_is_broken((pg)->count_info))
> +#define page_is_offlining(pg)      (pgc_is_broken((pg)->count_info))
> +
>  /* Page is not reference counted */
>  #define _PGC_extra        PG_shift(10)
>  #define PGC_extra         PG_mask(1, 10)
> --
> 2.21.0
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xenproject.org
> https://lists.xenproject.org/mailman/listinfo/xen-devel


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 19.03.2020 22:21, David Woodhouse wrote:
> --- a/xen/arch/x86/domctl.c
> +++ b/xen/arch/x86/domctl.c
> @@ -422,7 +422,7 @@ long arch_do_domctl(
>                  if ( page->u.inuse.type_info & PGT_pinned )
>                      type |= XEN_DOMCTL_PFINFO_LPINTAB;
>  
> -                if ( page->count_info & PGC_broken )
> +                if ( page_is_broken(page) )
>                      type = XEN_DOMCTL_PFINFO_BROKEN;

Coming back to an earlier request of mine: There are no locks being
held here afaics, so with

#define page_is_broken(pg)         (pgc_is_broken((pg)->count_info))

and

#define pgc_is_broken(pgc)         (pgc_is(pgc, broken) || \
                                    pgc_is(pgc, broken_offlining))

there's a chance that the page gets transitioned from
broken_offlining to broken (by another CPU) between these two
checks, resulting in wrong returned state. Either the latter macro
uses an intermediate variable and ACCESS_ONCE() or, as suggested
before, enumerators get arranged such that the check can be done
(e.g. using binary masking operations) with a single evaluation of
pgc.

This may or may not also be an issue for the other two pgc_is_*(),
but I think at least for symmetry they should then follow suit. At
the very least all three macros need to be immune to uses like
page_is_offlined(pg++) or similar argument expressions with side
effects.

> @@ -1699,19 +1706,18 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
>      do {
>          ret = *status = 0;
>  
> -        if ( y & PGC_broken )
> +        if ( pgc_is_broken(y) )
>          {
>              ret = -EINVAL;
> -            *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
> +            *status = PG_ONLINE_FAILED | PG_ONLINE_BROKEN;
>              break;
>          }
> -
> -        if ( (y & PGC_state) == PGC_state_offlined )
> +        else if ( pgc_is(y, offlined) )

At the risk of getting flamed again: Even if it was a matter of
taste in new code whether to use "else" in a case like this one,
this isn't new code, and it is in no way necessary to change what
we have for the purpose of this patch. I.e. without even having
to resort to the question of whether personal taste decisions are
to be accepted, this simply falls under "no unrelated /
unnecessary changes please". (FAOD this includes the deletion of
the blank line then as well.)

> --- a/xen/include/asm-x86/mm.h
> +++ b/xen/include/asm-x86/mm.h
> @@ -67,16 +67,32 @@
>   /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>  #define PGC_cacheattr_base PG_shift(6)
>  #define PGC_cacheattr_mask PG_mask(7, 6)
> - /* Page is broken? */
> -#define _PGC_broken       PG_shift(7)
> -#define PGC_broken        PG_mask(1, 7)
> - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> -#define PGC_state         PG_mask(3, 9)
> -#define PGC_state_inuse   PG_mask(0, 9)
> -#define PGC_state_offlining PG_mask(1, 9)
> -#define PGC_state_offlined PG_mask(2, 9)
> -#define PGC_state_free    PG_mask(3, 9)
> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> + /*
> +  * Mutually-exclusive page states:
> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> +  */
> +#define PGC_state                  PG_mask(7, 9)
> +#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_offlining        PG_mask(1, 9)
> +#define PGC_state_offlined         PG_mask(2, 9)
> +#define PGC_state_free             PG_mask(3, 9)
> +#define PGC_state_broken_offlining PG_mask(4, 9) /* Broken and offlining */
> +#define PGC_state_broken           PG_mask(5, 9) /* Broken and offlined */
> +
> +#define pgc_is(pgc, st)            (((pgc) & PGC_state) == PGC_state_##st)
> +#define page_state_is(pg, st)       pgc_is((pg)->count_info, st)
> +
> +#define pgc_is_broken(pgc)         (pgc_is(pgc, broken) || \
> +                                    pgc_is(pgc, broken_offlining))
> +#define pgc_is_offlined(pgc)       (pgc_is(pgc, offlined) || \
> +                                    pgc_is(pgc, broken))
> +#define pgc_is_offlining(pgc)      (pgc_is(pgc, offlining) || \
> +                                    pgc_is(pgc, broken_offlining))
> +
> +#define page_is_broken(pg)         (pgc_is_broken((pg)->count_info))
> +#define page_is_offlined(pg)       (pgc_is_broken((pg)->count_info))
> +#define page_is_offlining(pg)      (pgc_is_broken((pg)->count_info))

Copy-and-paste mistake (rhs is the same for all three; same for Arm)?
Also there's no need here for the outer pairs of parentheses.

Also, for the next version, may I ask that you number versions in
the subject's tag and that you provide a brief description of
changes from the previous version (if any, but there ought to be
some in a series for there to be a point to send out)? Thanks.

Jan

Re: [Xen-devel] [XEN PATCH v2 1/2] Check zone before merging adjacent blocks in heap

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 1328 bytes --]

On Wed, 2020-02-05 at 14:12 +0000, David Woodhouse wrote:
> I think we have a viable path to fixing that, by folding PGC_broken in to
> the state bits so that we can disambiguate. Will experiment.

Here, it looks something like this. First we fold PGC_broken into the
state bits giving us 8 possible states of which only 6 are currently in
use.

Then in the second patch we can move PGC_state_inuse from zero (the
default contents of the frame table at startup), and make a new state
PGC_state_uninitialised with the value zero.

We can make free_xenheap_pages() and free_domheap_pages() call
init_heap_pages() instead of free_heap_pages() if they see a page range
which is still in PGC_state_uninitialised.

We need to fix up a couple of asserts not only to cope with
PGC_state_inuse not being zero (as Julien has already looked at) but
also to take PGC_state_uninitialised pages.

In assign_pages() that's because we map the multiboot module containing
the initramfs to dom0. That might actually cross node boundaries,
contain MFN#0, etc. — so if/when that gets released by dom0 we'd want
those pages to be passed to init_heap_pages() just the same as boot-
allocated memory.

In share_xen_page_with_guest() it happens because we share all non-RAM
page frames with dom0.

(2 patches follow...)

[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

[Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

From: David Woodhouse <dwmw@amazon.co.uk>

Only PGC_state_offlining and PGC_state_offlined are valid in conjunction
with PGC_broken. The other two states (free and inuse) were never valid
for a broken page.

By folding PGC_broken in, we can have three bits for PGC_state which
allows up to 8 states, of which 6 are currently used and 2 are available
for new use cases.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---
 xen/arch/x86/domctl.c    |  2 +-
 xen/common/page_alloc.c  | 67 ++++++++++++++++++++++++----------------
 xen/include/asm-arm/mm.h | 26 +++++++++++-----
 xen/include/asm-x86/mm.h | 33 +++++++++++++-------
 4 files changed, 82 insertions(+), 46 deletions(-)

diff --git a/xen/arch/x86/domctl.c b/xen/arch/x86/domctl.c
index 4fa9c91140..17a318e16d 100644
--- a/xen/arch/x86/domctl.c
+++ b/xen/arch/x86/domctl.c
@@ -415,7 +415,7 @@ long arch_do_domctl(
                 if ( page->u.inuse.type_info & PGT_pinned )
                     type |= XEN_DOMCTL_PFINFO_LPINTAB;
 
-                if ( page->count_info & PGC_broken )
+                if ( page_is_broken(page) )
                     type = XEN_DOMCTL_PFINFO_BROKEN;
             }
 
diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
index 97902d42c1..4084503554 100644
--- a/xen/common/page_alloc.c
+++ b/xen/common/page_alloc.c
@@ -1093,7 +1093,7 @@ static int reserve_offlined_page(struct page_info *head)
         struct page_info *pg;
         int next_order;
 
-        if ( page_state_is(cur_head, offlined) )
+        if ( page_is_offlined(cur_head) )
         {
             cur_head++;
             if ( first_dirty != INVALID_DIRTY_IDX && first_dirty )
@@ -1113,7 +1113,7 @@ static int reserve_offlined_page(struct page_info *head)
             for ( i = (1 << cur_order), pg = cur_head + (1 << cur_order );
                   i < (1 << next_order);
                   i++, pg++ )
-                if ( page_state_is(pg, offlined) )
+                if ( page_is_offlined(pg) )
                     break;
             if ( i == ( 1 << next_order) )
             {
@@ -1145,16 +1145,19 @@ static int reserve_offlined_page(struct page_info *head)
 
     for ( cur_head = head; cur_head < head + ( 1UL << head_order); cur_head++ )
     {
-        if ( !page_state_is(cur_head, offlined) )
+        struct page_list_head *list;
+        if ( page_state_is(cur_head, offlined) )
+            list = &page_offlined_list;
+        else if (page_state_is(cur_head, broken) )
+            list = &page_broken_list;
+        else
             continue;
 
         avail[node][zone]--;
         total_avail_pages--;
         ASSERT(total_avail_pages >= 0);
 
-        page_list_add_tail(cur_head,
-                           test_bit(_PGC_broken, &cur_head->count_info) ?
-                           &page_broken_list : &page_offlined_list);
+        page_list_add_tail(cur_head, list);
 
         count++;
     }
@@ -1404,13 +1407,16 @@ static void free_heap_pages(
         switch ( pg[i].count_info & PGC_state )
         {
         case PGC_state_inuse:
-            BUG_ON(pg[i].count_info & PGC_broken);
             pg[i].count_info = PGC_state_free;
             break;
 
         case PGC_state_offlining:
-            pg[i].count_info = (pg[i].count_info & PGC_broken) |
-                               PGC_state_offlined;
+            pg[i].count_info = PGC_state_offlined;
+            tainted = 1;
+            break;
+
+        case PGC_state_broken_offlining:
+            pg[i].count_info = PGC_state_broken;
             tainted = 1;
             break;
 
@@ -1527,16 +1533,25 @@ static unsigned long mark_page_offline(struct page_info *pg, int broken)
     do {
         nx = x = y;
 
-        if ( ((x & PGC_state) != PGC_state_offlined) &&
-             ((x & PGC_state) != PGC_state_offlining) )
+        nx &= ~PGC_state;
+
+        switch ( x & PGC_state )
         {
-            nx &= ~PGC_state;
-            nx |= (((x & PGC_state) == PGC_state_free)
-                   ? PGC_state_offlined : PGC_state_offlining);
-        }
+        case PGC_state_inuse:
+        case PGC_state_offlining:
+            nx |= broken ? PGC_state_offlining : PGC_state_broken_offlining;
+            break;
+
+        case PGC_state_free:
+            nx |= broken ? PGC_state_broken : PGC_state_offlined;
 
-        if ( broken )
-            nx |= PGC_broken;
+        case PGC_state_broken_offlining:
+            nx |= PGC_state_broken_offlining;
+
+        case PGC_state_offlined:
+        case PGC_state_broken:
+            nx |= PGC_state_broken;
+        }
 
         if ( x == nx )
             break;
@@ -1609,7 +1624,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
      * need to prevent malicious guest access the broken page again.
      * Under such case, hypervisor shutdown guest, preventing recursive mce.
      */
-    if ( (pg->count_info & PGC_broken) && (owner = page_get_owner(pg)) )
+    if ( page_is_broken(pg) && (owner = page_get_owner(pg)) )
     {
         *status = PG_OFFLINE_AGAIN;
         domain_crash(owner);
@@ -1620,7 +1635,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
 
     old_info = mark_page_offline(pg, broken);
 
-    if ( page_state_is(pg, offlined) )
+    if ( page_is_offlined(pg) )
     {
         reserve_heap_page(pg);
 
@@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
     do {
         ret = *status = 0;
 
-        if ( y & PGC_broken )
+        if ( (y & PGC_state) == PGC_state_broken ||
+             (y & PGC_state) == PGC_state_broken_offlining )
         {
             ret = -EINVAL;
             *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
             break;
         }
-
-        if ( (y & PGC_state) == PGC_state_offlined )
+        else if ( (y & PGC_state) == PGC_state_offlined )
         {
             page_list_del(pg, &page_offlined_list);
             *status = PG_ONLINE_ONLINED;
@@ -1747,11 +1762,11 @@ int query_page_offline(mfn_t mfn, uint32_t *status)
 
     pg = mfn_to_page(mfn);
 
-    if ( page_state_is(pg, offlining) )
+    if ( page_is_offlining(pg) )
         *status |= PG_OFFLINE_STATUS_OFFLINE_PENDING;
-    if ( pg->count_info & PGC_broken )
+    if ( page_is_broken(pg) )
         *status |= PG_OFFLINE_STATUS_BROKEN;
-    if ( page_state_is(pg, offlined) )
+    if ( page_is_offlined(pg) )
         *status |= PG_OFFLINE_STATUS_OFFLINED;
 
     spin_unlock(&heap_lock);
@@ -2483,7 +2498,7 @@ __initcall(pagealloc_keyhandler_init);
 
 void scrub_one_page(struct page_info *pg)
 {
-    if ( unlikely(pg->count_info & PGC_broken) )
+    if ( unlikely(page_is_broken(pg)) )
         return;
 
 #ifndef NDEBUG
diff --git a/xen/include/asm-arm/mm.h b/xen/include/asm-arm/mm.h
index 333efd3a60..c9466c8ca0 100644
--- a/xen/include/asm-arm/mm.h
+++ b/xen/include/asm-arm/mm.h
@@ -112,13 +112,25 @@ struct page_info
 /* Page is broken? */
 #define _PGC_broken       PG_shift(7)
 #define PGC_broken        PG_mask(1, 7)
- /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
-#define PGC_state         PG_mask(3, 9)
-#define PGC_state_inuse   PG_mask(0, 9)
-#define PGC_state_offlining PG_mask(1, 9)
-#define PGC_state_offlined PG_mask(2, 9)
-#define PGC_state_free    PG_mask(3, 9)
-#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
+ /*
+  * Mutually-exclusive page states:
+  * { inuse, offlining, offlined, free, broken_offlining, broken }
+  */
+#define PGC_state                  PG_mask(7, 9)
+#define PGC_state_inuse            PG_mask(0, 9)
+#define PGC_state_offlining        PG_mask(1, 9)
+#define PGC_state_offlined         PG_mask(2, 9)
+#define PGC_state_free             PG_mask(3, 9)
+#define PGC_state_broken_offlining PG_mask(4, 9)
+#define PGC_state_broken           PG_mask(5, 9)
+
+#define page_state_is(pg, st)      (((pg)->count_info&PGC_state) == PGC_state_##st)
+#define page_is_broken(pg)         (page_state_is((pg), broken_offlining) ||  \
+                                    page_state_is((pg), broken))
+#define page_is_offlined(pg)       (page_state_is((pg), broken) ||    \
+                                    page_state_is((pg), offlined))
+#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
+                                    page_state_is((pg), offlining))
 
 /* Count of references to this frame. */
 #define PGC_count_width   PG_shift(9)
diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
index 2ca8882ad0..3edadf7a7c 100644
--- a/xen/include/asm-x86/mm.h
+++ b/xen/include/asm-x86/mm.h
@@ -67,18 +67,27 @@
  /* 3-bit PAT/PCD/PWT cache-attribute hint. */
 #define PGC_cacheattr_base PG_shift(6)
 #define PGC_cacheattr_mask PG_mask(7, 6)
- /* Page is broken? */
-#define _PGC_broken       PG_shift(7)
-#define PGC_broken        PG_mask(1, 7)
- /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
-#define PGC_state         PG_mask(3, 9)
-#define PGC_state_inuse   PG_mask(0, 9)
-#define PGC_state_offlining PG_mask(1, 9)
-#define PGC_state_offlined PG_mask(2, 9)
-#define PGC_state_free    PG_mask(3, 9)
-#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
-
- /* Count of references to this frame. */
+ /*
+  * Mutually-exclusive page states:
+  * { inuse, offlining, offlined, free, broken_offlining, broken }
+  */
+#define PGC_state                  PG_mask(7, 9)
+#define PGC_state_inuse            PG_mask(0, 9)
+#define PGC_state_offlining        PG_mask(1, 9)
+#define PGC_state_offlined         PG_mask(2, 9)
+#define PGC_state_free             PG_mask(3, 9)
+#define PGC_state_broken_offlining PG_mask(4, 9)
+#define PGC_state_broken           PG_mask(5, 9)
+
+#define page_state_is(pg, st)      (((pg)->count_info&PGC_state) == PGC_state_##st)
+#define page_is_broken(pg)         (page_state_is((pg), broken_offlining) ||  \
+                                    page_state_is((pg), broken))
+#define page_is_offlined(pg)       (page_state_is((pg), broken) ||    \
+                                    page_state_is((pg), offlined))
+#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
+                                    page_state_is((pg), offlining))
+
+/* Count of references to this frame. */
 #define PGC_count_width   PG_shift(9)
 #define PGC_count_mask    ((1UL<<PGC_count_width)-1)
 
-- 
2.21.0


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Julien Grall]

Hi David,

Could you please send the series in a separate thread? So we don't mix 
the two discussions (i.e merge and free on boot allocated page) together.

On 07/02/2020 15:57, David Woodhouse wrote:
> From: David Woodhouse <dwmw@amazon.co.uk>
> 
> Only PGC_state_offlining and PGC_state_offlined are valid in conjunction
> with PGC_broken. The other two states (free and inuse) were never valid
> for a broken page.
> 
> By folding PGC_broken in, we can have three bits for PGC_state which
> allows up to 8 states, of which 6 are currently used and 2 are available
> for new use cases.

The idea looks good to me. I have a few  mostly nitpicking comment below.

> 
> Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> ---
>   xen/arch/x86/domctl.c    |  2 +-
>   xen/common/page_alloc.c  | 67 ++++++++++++++++++++++++----------------
>   xen/include/asm-arm/mm.h | 26 +++++++++++-----
>   xen/include/asm-x86/mm.h | 33 +++++++++++++-------
>   4 files changed, 82 insertions(+), 46 deletions(-)
> 
> diff --git a/xen/arch/x86/domctl.c b/xen/arch/x86/domctl.c
> index 4fa9c91140..17a318e16d 100644
> --- a/xen/arch/x86/domctl.c
> +++ b/xen/arch/x86/domctl.c
> @@ -415,7 +415,7 @@ long arch_do_domctl(
>                   if ( page->u.inuse.type_info & PGT_pinned )
>                       type |= XEN_DOMCTL_PFINFO_LPINTAB;
>   
> -                if ( page->count_info & PGC_broken )
> +                if ( page_is_broken(page) )
>                       type = XEN_DOMCTL_PFINFO_BROKEN;
>               }
>   
> diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
> index 97902d42c1..4084503554 100644
> --- a/xen/common/page_alloc.c
> +++ b/xen/common/page_alloc.c
> @@ -1093,7 +1093,7 @@ static int reserve_offlined_page(struct page_info *head)
>           struct page_info *pg;
>           int next_order;
>   
> -        if ( page_state_is(cur_head, offlined) )
> +        if ( page_is_offlined(cur_head) )
>           {
>               cur_head++;
>               if ( first_dirty != INVALID_DIRTY_IDX && first_dirty )
> @@ -1113,7 +1113,7 @@ static int reserve_offlined_page(struct page_info *head)
>               for ( i = (1 << cur_order), pg = cur_head + (1 << cur_order );
>                     i < (1 << next_order);
>                     i++, pg++ )
> -                if ( page_state_is(pg, offlined) )
> +                if ( page_is_offlined(pg) )
>                       break;
>               if ( i == ( 1 << next_order) )
>               {
> @@ -1145,16 +1145,19 @@ static int reserve_offlined_page(struct page_info *head)
>   
>       for ( cur_head = head; cur_head < head + ( 1UL << head_order); cur_head++ )
>       {
> -        if ( !page_state_is(cur_head, offlined) )
> +        struct page_list_head *list;

We tend to add a newline after a series of declaration.

> +        if ( page_state_is(cur_head, offlined) )
> +            list = &page_offlined_list;
> +        else if (page_state_is(cur_head, broken) )
> +            list = &page_broken_list;
> +        else
>               continue;
>   
>           avail[node][zone]--;
>           total_avail_pages--;
>           ASSERT(total_avail_pages >= 0);
>   
> -        page_list_add_tail(cur_head,
> -                           test_bit(_PGC_broken, &cur_head->count_info) ?
> -                           &page_broken_list : &page_offlined_list);
> +        page_list_add_tail(cur_head, list);
>   
>           count++;
>       }
> @@ -1404,13 +1407,16 @@ static void free_heap_pages(
>           switch ( pg[i].count_info & PGC_state )
>           {
>           case PGC_state_inuse:
> -            BUG_ON(pg[i].count_info & PGC_broken);
>               pg[i].count_info = PGC_state_free;
>               break;
>   
>           case PGC_state_offlining:
> -            pg[i].count_info = (pg[i].count_info & PGC_broken) |
> -                               PGC_state_offlined;
> +            pg[i].count_info = PGC_state_offlined;
> +            tainted = 1;
> +            break;
> +
> +        case PGC_state_broken_offlining:
> +            pg[i].count_info = PGC_state_broken;
>               tainted = 1;
>               break;
>   
> @@ -1527,16 +1533,25 @@ static unsigned long mark_page_offline(struct page_info *pg, int broken)
>       do {
>           nx = x = y;
>   
> -        if ( ((x & PGC_state) != PGC_state_offlined) &&
> -             ((x & PGC_state) != PGC_state_offlining) )
> +        nx &= ~PGC_state;
> +
> +        switch ( x & PGC_state )
>           {
> -            nx &= ~PGC_state;
> -            nx |= (((x & PGC_state) == PGC_state_free)
> -                   ? PGC_state_offlined : PGC_state_offlining);
> -        }
> +        case PGC_state_inuse:
> +        case PGC_state_offlining:
> +            nx |= broken ? PGC_state_offlining : PGC_state_broken_offlining;
> +            break;
> +
> +        case PGC_state_free:
> +            nx |= broken ? PGC_state_broken : PGC_state_offlined;
>   
> -        if ( broken )
> -            nx |= PGC_broken;
> +        case PGC_state_broken_offlining:
> +            nx |= PGC_state_broken_offlining;
> +
> +        case PGC_state_offlined:
> +        case PGC_state_broken:
> +            nx |= PGC_state_broken;

Shouldn't this be:

case PGC_state_offlined:
     nx |= broken ? PGC_state_offlined : PGC_state_broken;

case PGC_state_broken:
     nx |= PGC_state_broken;

There are also quite a difference with the default case now. Without 
this patch, if you were to add a new state but not handling it here, you 
would transition to PGC_state_offlining. With this patch, we will 
transtion to 0 (i.e PGC_state_inuse for now).

PGC_state_* are not an enum, the compiler can't help to catch new state 
that doesn't have a corresponding case. So I would suggest to add a 
default matching the current behavior and adding an 
ASSERT_UNREACHABLE(). Note that I am open to a different kind of 
handling here.

> +        }
>   
>           if ( x == nx )
>               break;
> @@ -1609,7 +1624,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
>        * need to prevent malicious guest access the broken page again.
>        * Under such case, hypervisor shutdown guest, preventing recursive mce.
>        */
> -    if ( (pg->count_info & PGC_broken) && (owner = page_get_owner(pg)) )
> +    if ( page_is_broken(pg) && (owner = page_get_owner(pg)) )
>       {
>           *status = PG_OFFLINE_AGAIN;
>           domain_crash(owner);
> @@ -1620,7 +1635,7 @@ int offline_page(mfn_t mfn, int broken, uint32_t *status)
>   
>       old_info = mark_page_offline(pg, broken);
>   
> -    if ( page_state_is(pg, offlined) )
> +    if ( page_is_offlined(pg) )
>       {
>           reserve_heap_page(pg);
>   
> @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
>       do {
>           ret = *status = 0;
>   
> -        if ( y & PGC_broken )
> +        if ( (y & PGC_state) == PGC_state_broken ||
> +             (y & PGC_state) == PGC_state_broken_offlining )

This is a bit a shame we can't use page_is_broken. Would it make sense 
to introduce an helper that just take a count_info?

>           {
>               ret = -EINVAL;
>               *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
>               break;
>           }
> -
> -        if ( (y & PGC_state) == PGC_state_offlined )
> +        else if ( (y & PGC_state) == PGC_state_offlined )
>           {
>               page_list_del(pg, &page_offlined_list);
>               *status = PG_ONLINE_ONLINED;
> @@ -1747,11 +1762,11 @@ int query_page_offline(mfn_t mfn, uint32_t *status)
>   
>       pg = mfn_to_page(mfn);
>   
> -    if ( page_state_is(pg, offlining) )
> +    if ( page_is_offlining(pg) )
>           *status |= PG_OFFLINE_STATUS_OFFLINE_PENDING;
> -    if ( pg->count_info & PGC_broken )
> +    if ( page_is_broken(pg) )
>           *status |= PG_OFFLINE_STATUS_BROKEN;
> -    if ( page_state_is(pg, offlined) )
> +    if ( page_is_offlined(pg) )
>           *status |= PG_OFFLINE_STATUS_OFFLINED;
>   
>       spin_unlock(&heap_lock);
> @@ -2483,7 +2498,7 @@ __initcall(pagealloc_keyhandler_init);
>   
>   void scrub_one_page(struct page_info *pg)
>   {
> -    if ( unlikely(pg->count_info & PGC_broken) )
> +    if ( unlikely(page_is_broken(pg)) )
>           return;
>   
>   #ifndef NDEBUG
> diff --git a/xen/include/asm-arm/mm.h b/xen/include/asm-arm/mm.h
> index 333efd3a60..c9466c8ca0 100644
> --- a/xen/include/asm-arm/mm.h
> +++ b/xen/include/asm-arm/mm.h
> @@ -112,13 +112,25 @@ struct page_info
>   /* Page is broken? */
>   #define _PGC_broken       PG_shift(7)
>   #define PGC_broken        PG_mask(1, 7)

Shouldn't this be dropped now?

> - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> -#define PGC_state         PG_mask(3, 9)
> -#define PGC_state_inuse   PG_mask(0, 9)
> -#define PGC_state_offlining PG_mask(1, 9)
> -#define PGC_state_offlined PG_mask(2, 9)
> -#define PGC_state_free    PG_mask(3, 9)
> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> + /*
> +  * Mutually-exclusive page states:
> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> +  */
> +#define PGC_state                  PG_mask(7, 9)
> +#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_offlining        PG_mask(1, 9)
> +#define PGC_state_offlined         PG_mask(2, 9)
> +#define PGC_state_free             PG_mask(3, 9)
> +#define PGC_state_broken_offlining PG_mask(4, 9)
> +#define PGC_state_broken           PG_mask(5, 9)

I agree that making all the value aligned make it nicer to read, but 
this is increasing number of "unrelated" changes and makes the review 
more difficult.

I would prefer if we leave the indentation alone for the current 
#define. But I am not going to push for it :).

> +
> +#define page_state_is(pg, st)      (((pg)->count_info&PGC_state) == PGC_state_##st)
> +#define page_is_broken(pg)         (page_state_is((pg), broken_offlining) ||  \
> +                                    page_state_is((pg), broken))
> +#define page_is_offlined(pg)       (page_state_is((pg), broken) ||    \
> +                                    page_state_is((pg), offlined))
> +#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
> +                                    page_state_is((pg), offlining))
>   
>   /* Count of references to this frame. */
>   #define PGC_count_width   PG_shift(9)
> diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
> index 2ca8882ad0..3edadf7a7c 100644
> --- a/xen/include/asm-x86/mm.h
> +++ b/xen/include/asm-x86/mm.h
> @@ -67,18 +67,27 @@
>    /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>   #define PGC_cacheattr_base PG_shift(6)
>   #define PGC_cacheattr_mask PG_mask(7, 6)
> - /* Page is broken? */
> -#define _PGC_broken       PG_shift(7)
> -#define PGC_broken        PG_mask(1, 7)
> - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> -#define PGC_state         PG_mask(3, 9)
> -#define PGC_state_inuse   PG_mask(0, 9)
> -#define PGC_state_offlining PG_mask(1, 9)
> -#define PGC_state_offlined PG_mask(2, 9)
> -#define PGC_state_free    PG_mask(3, 9)
> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> -
> - /* Count of references to this frame. */
> + /*
> +  * Mutually-exclusive page states:
> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> +  */
> +#define PGC_state                  PG_mask(7, 9)
> +#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_offlining        PG_mask(1, 9)
> +#define PGC_state_offlined         PG_mask(2, 9)
> +#define PGC_state_free             PG_mask(3, 9)
> +#define PGC_state_broken_offlining PG_mask(4, 9)
> +#define PGC_state_broken           PG_mask(5, 9)
> +
> +#define page_state_is(pg, st)      (((pg)->count_info&PGC_state) == PGC_state_##st)
> +#define page_is_broken(pg)         (page_state_is((pg), broken_offlining) ||  \
> +                                    page_state_is((pg), broken))
> +#define page_is_offlined(pg)       (page_state_is((pg), broken) ||    \
> +                                    page_state_is((pg), offlined))
> +#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
> +                                    page_state_is((pg), offlining))
> +
> +/* Count of references to this frame. */
>   #define PGC_count_width   PG_shift(9)
>   #define PGC_count_mask    ((1UL<<PGC_count_width)-1)
>   
> 

Cheers,

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 1657 bytes --]

On Fri, 2020-02-07 at 20:27 +0000, Julien Grall wrote:
> Could you please send the series in a separate thread? So we don't mix 
> the two discussions (i.e merge and free on boot allocated page) together.

Well, they're all part of the same mess, really.

There are cases where pages end up in free_heap_pages() which were
never vetted by init_heap_pages(). While that normally works just fine
— to the extent that we've never really cared — the hack excluding MFN0
is one of the things that gets 'missed' for such pages.

I was only looking at this because the early vmap support makes it a
tiny bit more likely that some pages will be freed that way after being
given out by the boot allocator, but there were plenty of reasons it
might happen already.

These patches basically legitimise that — we make it OK for
free_heap_pages() to be given a range which was never in the heap in
the first place. And in so doing, we fix the MFN0/zone merge trick even
when (for example) MFN0 was part of the dom0 initramfs and assigned
directly, but gets freed up later.

But sure, having failed to elicit any screams of "don't do it like
that", I'll fix up the things you pointed out and I'll repost it as
part of the series containing the early vmap() support, since that's
why I'm working on it.

Although at this point I'm half tempted to declare that there are *so*
many ways this can happen already, that the tiny little bit that it's
made more likely by the early vmap() support is basically in the noise.

In that case we can separate these patches out as an independent fix
rather than a dependency of early vmap.




[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Julien Grall]

On 09/02/2020 13:22, David Woodhouse wrote:
> On Fri, 2020-02-07 at 20:27 +0000, Julien Grall wrote:
>> Could you please send the series in a separate thread? So we don't mix
>> the two discussions (i.e merge and free on boot allocated page) together.
> 
> Well, they're all part of the same mess, really.

Sending a series in the middle of another series is always more 
difficult to track :). The more if they are handled by two different 
person...

> 
> There are cases where pages end up in free_heap_pages() which were
> never vetted by init_heap_pages(). While that normally works just fine
> — to the extent that we've never really cared — the hack excluding MFN0
> is one of the things that gets 'missed' for such pages.
> 
> I was only looking at this because the early vmap support makes it a
> tiny bit more likely that some pages will be freed that way after being
> given out by the boot allocator, but there were plenty of reasons it
> might happen already.
> 
> These patches basically legitimise that — we make it OK for
> free_heap_pages() to be given a range which was never in the heap in
> the first place. And in so doing, we fix the MFN0/zone merge trick even
> when (for example) MFN0 was part of the dom0 initramfs and assigned
> directly, but gets freed up later.
> 
> But sure, having failed to elicit any screams of "don't do it like
> that", I'll fix up the things you pointed out and I'll repost it as
> part of the series containing the early vmap() support, since that's
> why I'm working on it.
> 
> Although at this point I'm half tempted to declare that there are *so*
> many ways this can happen already, that the tiny little bit that it's
> made more likely by the early vmap() support is basically in the noise.
> 
> In that case we can separate these patches out as an independent fix
> rather than a dependency of early vmap.

I hadn't realize how messy it was because I had Arm in mind and wasn't 
expected x86 to abuse so much the interface.

For x86, I agree that this is noise as they are abusing the interface 
pretty much everywhere.

However, on Arm there is only one place that is abusing the interface. 
It is in the ACPI code, although I think it will be just a leak given 
the implementation of acpi_os_free_memory(). As we don't free 
page-tables yet on Arm, the introduction of the early vmap would not 
introduce any more abuse on Arm.

It would obviously nice to fix it, but as you said this is noise on x86. 
So that's really up to the x86 folks (Andrew, George, Jan) to see 
whether yet another abuse is ok for them :).

Cheers,

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 3528 bytes --]

On Fri, 2020-02-07 at 20:27 +0000, Julien Grall wrote:

> > +        switch ( x & PGC_state )
> >           {
> > -            nx &= ~PGC_state;
> > -            nx |= (((x & PGC_state) == PGC_state_free)
> > -                   ? PGC_state_offlined : PGC_state_offlining);
> > -        }
> > +        case PGC_state_inuse:
> > +        case PGC_state_offlining:
> > +            nx |= broken ? PGC_state_offlining : PGC_state_broken_offlining;
> > +            break;
> > +
> > +        case PGC_state_free:
> > +            nx |= broken ? PGC_state_broken : PGC_state_offlined;
> >   
> > -        if ( broken )
> > -            nx |= PGC_broken;
> > +        case PGC_state_broken_offlining:
> > +            nx |= PGC_state_broken_offlining;
> > +
> > +        case PGC_state_offlined:
> > +        case PGC_state_broken:
> > +            nx |= PGC_state_broken;
> 
> Shouldn't this be:
> 
> case PGC_state_offlined:
>      nx |= broken ? PGC_state_offlined : PGC_state_broken;
> 
> case PGC_state_broken:
>      nx |= PGC_state_broken;
> 
> There are also quite a difference with the default case now. Without 
> this patch, if you were to add a new state but not handling it here, you 
> would transition to PGC_state_offlining. With this patch, we will 
> transtion to 0 (i.e PGC_state_inuse for now).
> 
> PGC_state_* are not an enum, the compiler can't help to catch new state 
> that doesn't have a corresponding case. So I would suggest to add a 
> default matching the current behavior and adding an 
> ASSERT_UNREACHABLE(). Note that I am open to a different kind of 
> handling here.

I revamped this, taking into account your later suggestion of a helper
that works on the count_info. Looks more like this:


        /* If it was already broken, it stays broken */
        if ( pgc_is_broken(x) )
            broken = 1;

        if ( pgc_is_offlined(x) || pgc_is(x, free) )
            nx |= broken ? PGC_state_broken : PGC_state_offlined;
        else
            nx |= broken ? PGC_state_broken_offlining : PGC_state_offlining;


> > - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> > -#define PGC_state         PG_mask(3, 9)
> > -#define PGC_state_inuse   PG_mask(0, 9)
> > -#define PGC_state_offlining PG_mask(1, 9)
> > -#define PGC_state_offlined PG_mask(2, 9)
> > -#define PGC_state_free    PG_mask(3, 9)
> > -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> > + /*
> > +  * Mutually-exclusive page states:
> > +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> > +  */
> > +#define PGC_state                  PG_mask(7, 9)
> > +#define PGC_state_inuse            PG_mask(0, 9)
> > +#define PGC_state_offlining        PG_mask(1, 9)
> > +#define PGC_state_offlined         PG_mask(2, 9)
> > +#define PGC_state_free             PG_mask(3, 9)
> > +#define PGC_state_broken_offlining PG_mask(4, 9)
> > +#define PGC_state_broken           PG_mask(5, 9)
> 
> I agree that making all the value aligned make it nicer to read, but 
> this is increasing number of "unrelated" changes and makes the review 
> more difficult.
> 
> I would prefer if we leave the indentation alone for the current 
> #define. But I am not going to push for it :).

I'm generally sympathetic to that point of view but at this point, all
those page states are kind of being redefined and it makes sense to
think about them all; having them all change doesn't hurt.



[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 07.02.2020 16:57, David Woodhouse wrote:
> @@ -1145,16 +1145,19 @@ static int reserve_offlined_page(struct page_info *head)
>  
>      for ( cur_head = head; cur_head < head + ( 1UL << head_order); cur_head++ )
>      {
> -        if ( !page_state_is(cur_head, offlined) )
> +        struct page_list_head *list;
> +        if ( page_state_is(cur_head, offlined) )
> +            list = &page_offlined_list;
> +        else if (page_state_is(cur_head, broken) )
> +            list = &page_broken_list;
> +        else
>              continue;
>  
>          avail[node][zone]--;
>          total_avail_pages--;
>          ASSERT(total_avail_pages >= 0);
>  
> -        page_list_add_tail(cur_head,
> -                           test_bit(_PGC_broken, &cur_head->count_info) ?
> -                           &page_broken_list : &page_offlined_list);
> +        page_list_add_tail(cur_head, list);

While I realize it's fewer comparisons this way, I still wonder
whether for the reader's sake it wouldn't better be
page_is_offlined() first and then page_is_broken() down here.

> @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn, uint32_t *status)
>      do {
>          ret = *status = 0;
>  
> -        if ( y & PGC_broken )
> +        if ( (y & PGC_state) == PGC_state_broken ||
> +             (y & PGC_state) == PGC_state_broken_offlining )
>          {
>              ret = -EINVAL;
>              *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
>              break;
>          }
> -
> -        if ( (y & PGC_state) == PGC_state_offlined )
> +        else if ( (y & PGC_state) == PGC_state_offlined )

I don't see a need for adding "else" here.

> --- a/xen/include/asm-x86/mm.h
> +++ b/xen/include/asm-x86/mm.h
> @@ -67,18 +67,27 @@
>   /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>  #define PGC_cacheattr_base PG_shift(6)
>  #define PGC_cacheattr_mask PG_mask(7, 6)
> - /* Page is broken? */
> -#define _PGC_broken       PG_shift(7)
> -#define PGC_broken        PG_mask(1, 7)
> - /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
> -#define PGC_state         PG_mask(3, 9)
> -#define PGC_state_inuse   PG_mask(0, 9)
> -#define PGC_state_offlining PG_mask(1, 9)
> -#define PGC_state_offlined PG_mask(2, 9)
> -#define PGC_state_free    PG_mask(3, 9)
> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) == PGC_state_##st)
> -
> - /* Count of references to this frame. */
> + /*
> +  * Mutually-exclusive page states:
> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> +  */
> +#define PGC_state                  PG_mask(7, 9)
> +#define PGC_state_inuse            PG_mask(0, 9)
> +#define PGC_state_offlining        PG_mask(1, 9)
> +#define PGC_state_offlined         PG_mask(2, 9)
> +#define PGC_state_free             PG_mask(3, 9)
> +#define PGC_state_broken_offlining PG_mask(4, 9)

TBH I'd prefer PGC_state_offlining_broken, as it's not the
offlining which is broken, but a broken page is being
offlined.

> +#define PGC_state_broken           PG_mask(5, 9)
> +
> +#define page_state_is(pg, st)      (((pg)->count_info&PGC_state) == PGC_state_##st)

Blanks around & please.

> +#define page_is_broken(pg)         (page_state_is((pg), broken_offlining) ||  \
> +                                    page_state_is((pg), broken))
> +#define page_is_offlined(pg)       (page_state_is((pg), broken) ||    \
> +                                    page_state_is((pg), offlined))

The inclusion of "broken" here would seem to deserve a (brief)
comment, either here or next to PGC_state_broken's #define.

> +#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
> +                                    page_state_is((pg), offlining))

Overall I wonder whether the PGC_state_* ordering couldn't be
adjusted such that at least some of these three won't need
two comparisons (by masking off a bit before comparing).

Also for all three - no need for extra parentheses around pg
(or in general macro arguments which get handed on without
being part of an expression).

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 5400 bytes --]

On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
> On 07.02.2020 16:57, David Woodhouse wrote:
> > @@ -1145,16 +1145,19 @@ static int reserve_offlined_page(struct
> > page_info *head)
> >  
> >      for ( cur_head = head; cur_head < head + ( 1UL << head_order);
> > cur_head++ )
> >      {
> > -        if ( !page_state_is(cur_head, offlined) )
> > +        struct page_list_head *list;
> > +        if ( page_state_is(cur_head, offlined) )
> > +            list = &page_offlined_list;
> > +        else if (page_state_is(cur_head, broken) )
> > +            list = &page_broken_list;
> > +        else
> >              continue;
> >  
> >          avail[node][zone]--;
> >          total_avail_pages--;
> >          ASSERT(total_avail_pages >= 0);
> >  
> > -        page_list_add_tail(cur_head,
> > -                           test_bit(_PGC_broken, &cur_head->count_info) ?
> > -                           &page_broken_list : &page_offlined_list);
> > +        page_list_add_tail(cur_head, list);
> 
> While I realize it's fewer comparisons this way, I still wonder
> whether for the reader's sake it wouldn't better be
> page_is_offlined() first and then page_is_broken() down here.

Nah, that would be worse. This way there are two cases which are
explicitly handled and the list to use for each of them is explicitly
set. The 'if (a||b) …    some_function(a ? thing_for_a : thing_for_b)'
construct is much less comprehensible.



> > @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn,
> > uint32_t *status)
> >      do {
> >          ret = *status = 0;
> >  
> > -        if ( y & PGC_broken )
> > +        if ( (y & PGC_state) == PGC_state_broken ||
> > +             (y & PGC_state) == PGC_state_broken_offlining )
> >          {
> >              ret = -EINVAL;
> >              *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
> >              break;
> >          }
> > -
> > -        if ( (y & PGC_state) == PGC_state_offlined )
> > +        else if ( (y & PGC_state) == PGC_state_offlined )
> 
> I don't see a need for adding "else" here.

They are mutually exclusive cases. It makes things a whole lot clearer
to the reader to put the 'else' there, and sometimes helps a naïve
compiler along the way too.


> > --- a/xen/include/asm-x86/mm.h
> > +++ b/xen/include/asm-x86/mm.h
> > @@ -67,18 +67,27 @@
> >   /* 3-bit PAT/PCD/PWT cache-attribute hint. */
> >  #define PGC_cacheattr_base PG_shift(6)
> >  #define PGC_cacheattr_mask PG_mask(7, 6)
> > - /* Page is broken? */
> > -#define _PGC_broken       PG_shift(7)
> > -#define PGC_broken        PG_mask(1, 7)
> > - /* Mutually-exclusive page states: { inuse, offlining, offlined,
> > free }. */
> > -#define PGC_state         PG_mask(3, 9)
> > -#define PGC_state_inuse   PG_mask(0, 9)
> > -#define PGC_state_offlining PG_mask(1, 9)
> > -#define PGC_state_offlined PG_mask(2, 9)
> > -#define PGC_state_free    PG_mask(3, 9)
> > -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) ==
> > PGC_state_##st)
> > -
> > - /* Count of references to this frame. */
> > + /*
> > +  * Mutually-exclusive page states:
> > +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> > +  */
> > +#define PGC_state                  PG_mask(7, 9)
> > +#define PGC_state_inuse            PG_mask(0, 9)
> > +#define PGC_state_offlining        PG_mask(1, 9)
> > +#define PGC_state_offlined         PG_mask(2, 9)
> > +#define PGC_state_free             PG_mask(3, 9)
> > +#define PGC_state_broken_offlining PG_mask(4, 9)
> 
> TBH I'd prefer PGC_state_offlining_broken, as it's not the
> offlining which is broken, but a broken page is being
> offlined.

It is the page which is both broken and offlining.
Or indeed it is the page which is both offlining and broken.


> > +#define PGC_state_broken           PG_mask(5, 9)
> > +
> > +#define page_state_is(pg, st)      (((pg)->count_info&PGC_state)
> > == PGC_state_##st)
> 
> Blanks around & please.

That part I hadn't touched but sure, I'll add those while I'm touching
it. I'd already ignored Julien's request *not* to make whitespace
cleanups while I'm here, after all :)

> > +#define page_is_broken(pg)         (page_state_is((pg),
> > broken_offlining) ||  \
> > +                                    page_state_is((pg), broken))
> > +#define page_is_offlined(pg)       (page_state_is((pg), broken)
> > ||    \
> > +                                    page_state_is((pg), offlined))
> 
> The inclusion of "broken" here would seem to deserve a (brief)
> comment, either here or next to PGC_state_broken's #define.

Done, in the version which will be sent shortly.

> > +#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
> > +                                    page_state_is((pg), offlining))
> 
> Overall I wonder whether the PGC_state_* ordering couldn't be
> adjusted such that at least some of these three won't need
> two comparisons (by masking off a bit before comparing).

The whole point in this exercise is that there isn't a whole bit for
these; they are each *two* states out of the possible 8.

> Also for all three - no need for extra parentheses around pg
> (or in general macro arguments which get handed on without
> being part of an expression).

Yeah, I'll remove some of those.




[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 17.03.2020 22:52, David Woodhouse wrote:
> On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
>> On 07.02.2020 16:57, David Woodhouse wrote:
>>> @@ -1145,16 +1145,19 @@ static int reserve_offlined_page(struct
>>> page_info *head)
>>>   
>>>       for ( cur_head = head; cur_head < head + ( 1UL << head_order);
>>> cur_head++ )
>>>       {
>>> -        if ( !page_state_is(cur_head, offlined) )
>>> +        struct page_list_head *list;
>>> +        if ( page_state_is(cur_head, offlined) )
>>> +            list = &page_offlined_list;
>>> +        else if (page_state_is(cur_head, broken) )
>>> +            list = &page_broken_list;
>>> +        else
>>>               continue;
>>>   
>>>           avail[node][zone]--;
>>>           total_avail_pages--;
>>>           ASSERT(total_avail_pages >= 0);
>>>   
>>> -        page_list_add_tail(cur_head,
>>> -                           test_bit(_PGC_broken, &cur_head->count_info) ?
>>> -                           &page_broken_list : &page_offlined_list);
>>> +        page_list_add_tail(cur_head, list);
>>
>> While I realize it's fewer comparisons this way, I still wonder
>> whether for the reader's sake it wouldn't better be
>> page_is_offlined() first and then page_is_broken() down here.
> 
> Nah, that would be worse. This way there are two cases which are
> explicitly handled and the list to use for each of them is explicitly
> set. The 'if (a||b) …    some_function(a ? thing_for_a : thing_for_b)'
> construct is much less comprehensible.

It's a matter of taste, I agree, and in such a case I generally advise
to see about limiting code churn. For code you then still introduce
anew, yes, taste decisions may typically be to the authors judgement
(there are exceptions, though).

>>> @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn,
>>> uint32_t *status)
>>>       do {
>>>           ret = *status = 0;
>>>   
>>> -        if ( y & PGC_broken )
>>> +        if ( (y & PGC_state) == PGC_state_broken ||
>>> +             (y & PGC_state) == PGC_state_broken_offlining )
>>>           {
>>>               ret = -EINVAL;
>>>               *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
>>>               break;
>>>           }
>>> -
>>> -        if ( (y & PGC_state) == PGC_state_offlined )
>>> +        else if ( (y & PGC_state) == PGC_state_offlined )
>>
>> I don't see a need for adding "else" here.
> 
> They are mutually exclusive cases. It makes things a whole lot clearer
> to the reader to put the 'else' there, and sometimes helps a naïve
> compiler along the way too.

Well, I'm afraid I'm going to be pretty strict about this: It's again
a matter of taste, yes, but we generally try to avoid pointless else.
What you consider "a whole lot clearer to the reader" is the opposite
from my pov.

>>> --- a/xen/include/asm-x86/mm.h
>>> +++ b/xen/include/asm-x86/mm.h
>>> @@ -67,18 +67,27 @@
>>>    /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>>>   #define PGC_cacheattr_base PG_shift(6)
>>>   #define PGC_cacheattr_mask PG_mask(7, 6)
>>> - /* Page is broken? */
>>> -#define _PGC_broken       PG_shift(7)
>>> -#define PGC_broken        PG_mask(1, 7)
>>> - /* Mutually-exclusive page states: { inuse, offlining, offlined,
>>> free }. */
>>> -#define PGC_state         PG_mask(3, 9)
>>> -#define PGC_state_inuse   PG_mask(0, 9)
>>> -#define PGC_state_offlining PG_mask(1, 9)
>>> -#define PGC_state_offlined PG_mask(2, 9)
>>> -#define PGC_state_free    PG_mask(3, 9)
>>> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) ==
>>> PGC_state_##st)
>>> -
>>> - /* Count of references to this frame. */
>>> + /*
>>> +  * Mutually-exclusive page states:
>>> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
>>> +  */
>>> +#define PGC_state                  PG_mask(7, 9)
>>> +#define PGC_state_inuse            PG_mask(0, 9)
>>> +#define PGC_state_offlining        PG_mask(1, 9)
>>> +#define PGC_state_offlined         PG_mask(2, 9)
>>> +#define PGC_state_free             PG_mask(3, 9)
>>> +#define PGC_state_broken_offlining PG_mask(4, 9)
>>
>> TBH I'd prefer PGC_state_offlining_broken, as it's not the
>> offlining which is broken, but a broken page is being
>> offlined.
> 
> It is the page which is both broken and offlining.
> Or indeed it is the page which is both offlining and broken.

I.e. you agree with flipping the two parts around?

>>> +#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
>>> +                                    page_state_is((pg), offlining))
>>
>> Overall I wonder whether the PGC_state_* ordering couldn't be
>> adjusted such that at least some of these three won't need
>> two comparisons (by masking off a bit before comparing).
> 
> The whole point in this exercise is that there isn't a whole bit for
> these; they are each *two* states out of the possible 8.

Sure. But just consider the more general case: Instead of writing

	if ( i == 6 || i == 7 )

you can as well write

	if ( (i | 1) == 7 )

Similar for multiple == vs a single <= or >=.

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Julien Grall]

Hi Jan,

On 18/03/2020 09:56, Jan Beulich wrote:
> On 17.03.2020 22:52, David Woodhouse wrote:
>> On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
>>> On 07.02.2020 16:57, David Woodhouse wrote:
>>>> @@ -1145,16 +1145,19 @@ static int reserve_offlined_page(struct
>>>> page_info *head)
>>>>       for ( cur_head = head; cur_head < head + ( 1UL << head_order);
>>>> cur_head++ )
>>>>       {
>>>> -        if ( !page_state_is(cur_head, offlined) )
>>>> +        struct page_list_head *list;
>>>> +        if ( page_state_is(cur_head, offlined) )
>>>> +            list = &page_offlined_list;
>>>> +        else if (page_state_is(cur_head, broken) )
>>>> +            list = &page_broken_list;
>>>> +        else
>>>>               continue;
>>>>           avail[node][zone]--;
>>>>           total_avail_pages--;
>>>>           ASSERT(total_avail_pages >= 0);
>>>> -        page_list_add_tail(cur_head,
>>>> -                           test_bit(_PGC_broken, 
>>>> &cur_head->count_info) ?
>>>> -                           &page_broken_list : &page_offlined_list);
>>>> +        page_list_add_tail(cur_head, list);
>>>
>>> While I realize it's fewer comparisons this way, I still wonder
>>> whether for the reader's sake it wouldn't better be
>>> page_is_offlined() first and then page_is_broken() down here.
>>
>> Nah, that would be worse. This way there are two cases which are
>> explicitly handled and the list to use for each of them is explicitly
>> set. The 'if (a||b) …    some_function(a ? thing_for_a : thing_for_b)'
>> construct is much less comprehensible.
> 
> It's a matter of taste, I agree, and in such a case I generally advise
> to see about limiting code churn. For code you then still introduce
> anew, yes, taste decisions may typically be to the authors judgement
> (there are exceptions, though).
> 
>>>> @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn,
>>>> uint32_t *status)
>>>>       do {
>>>>           ret = *status = 0;
>>>> -        if ( y & PGC_broken )
>>>> +        if ( (y & PGC_state) == PGC_state_broken ||
>>>> +             (y & PGC_state) == PGC_state_broken_offlining )
>>>>           {
>>>>               ret = -EINVAL;
>>>>               *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
>>>>               break;
>>>>           }
>>>> -
>>>> -        if ( (y & PGC_state) == PGC_state_offlined )
>>>> +        else if ( (y & PGC_state) == PGC_state_offlined )
>>>
>>> I don't see a need for adding "else" here.
>>
>> They are mutually exclusive cases. It makes things a whole lot clearer
>> to the reader to put the 'else' there, and sometimes helps a naïve
>> compiler along the way too.
> 
> Well, I'm afraid I'm going to be pretty strict about this: It's again
> a matter of taste, yes, but we generally try to avoid pointless else.
> What you consider "a whole lot clearer to the reader" is the opposite
> from my pov.

While I agree the 'else' may be pointless, I don't think it is worth an 
argument. As the author of the patch, it is his choice to write the code 
like that.

> 
>>>> --- a/xen/include/asm-x86/mm.h
>>>> +++ b/xen/include/asm-x86/mm.h
>>>> @@ -67,18 +67,27 @@
>>>>    /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>>>>   #define PGC_cacheattr_base PG_shift(6)
>>>>   #define PGC_cacheattr_mask PG_mask(7, 6)
>>>> - /* Page is broken? */
>>>> -#define _PGC_broken       PG_shift(7)
>>>> -#define PGC_broken        PG_mask(1, 7)
>>>> - /* Mutually-exclusive page states: { inuse, offlining, offlined,
>>>> free }. */
>>>> -#define PGC_state         PG_mask(3, 9)
>>>> -#define PGC_state_inuse   PG_mask(0, 9)
>>>> -#define PGC_state_offlining PG_mask(1, 9)
>>>> -#define PGC_state_offlined PG_mask(2, 9)
>>>> -#define PGC_state_free    PG_mask(3, 9)
>>>> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) ==
>>>> PGC_state_##st)
>>>> -
>>>> - /* Count of references to this frame. */
>>>> + /*
>>>> +  * Mutually-exclusive page states:
>>>> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
>>>> +  */
>>>> +#define PGC_state                  PG_mask(7, 9)
>>>> +#define PGC_state_inuse            PG_mask(0, 9)
>>>> +#define PGC_state_offlining        PG_mask(1, 9)
>>>> +#define PGC_state_offlined         PG_mask(2, 9)
>>>> +#define PGC_state_free             PG_mask(3, 9)
>>>> +#define PGC_state_broken_offlining PG_mask(4, 9)
>>>
>>> TBH I'd prefer PGC_state_offlining_broken, as it's not the
>>> offlining which is broken, but a broken page is being
>>> offlined.
>>
>> It is the page which is both broken and offlining.
>> Or indeed it is the page which is both offlining and broken.
> 
> I.e. you agree with flipping the two parts around?
> 
>>>> +#define page_is_offlining(pg)      (page_state_is((pg), 
>>>> broken_offlining) || \
>>>> +                                    page_state_is((pg), offlining))
>>>
>>> Overall I wonder whether the PGC_state_* ordering couldn't be
>>> adjusted such that at least some of these three won't need
>>> two comparisons (by masking off a bit before comparing).
>>
>> The whole point in this exercise is that there isn't a whole bit for
>> these; they are each *two* states out of the possible 8.
> 
> Sure. But just consider the more general case: Instead of writing
> 
>      if ( i == 6 || i == 7 )
> 
> you can as well write
> 
>      if ( (i | 1) == 7 )

I stumbled accross a few of those recently and this is not the obvious 
things to read. Yes, your code may be faster. But is it really worth it 
compare to the cost of readability and futureproofness?

Cheers,

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 18.03.2020 13:31, Julien Grall wrote:
> On 18/03/2020 09:56, Jan Beulich wrote:
>> On 17.03.2020 22:52, David Woodhouse wrote:
>>> On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
>>>> On 07.02.2020 16:57, David Woodhouse wrote:
>>>>> +#define page_is_offlining(pg)      (page_state_is((pg), broken_offlining) || \
>>>>> +                                    page_state_is((pg), offlining))
>>>>
>>>> Overall I wonder whether the PGC_state_* ordering couldn't be
>>>> adjusted such that at least some of these three won't need
>>>> two comparisons (by masking off a bit before comparing).
>>>
>>> The whole point in this exercise is that there isn't a whole bit for
>>> these; they are each *two* states out of the possible 8.
>>
>> Sure. But just consider the more general case: Instead of writing
>>
>>      if ( i == 6 || i == 7 )
>>
>> you can as well write
>>
>>      if ( (i | 1) == 7 )
> 
> I stumbled accross a few of those recently and this is not the
> obvious things to read.

Depends on the reader, I guess.

> Yes, your code may be faster. But is it really worth it compare
> to the cost of readability and futureproofness?

Not sure how being future proof comes into play here. I'm not
suggesting to use literal numbers. I'd also be happy to be proven
wrong in assuming that the compiler still can't do such
transformations itself; it couldn't when I check a while back.
Reducing the number of conditional branches is, imo, worth at
least some effort at the source level.

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 8790 bytes --]

On Wed, 2020-03-18 at 12:31 +0000, Julien Grall wrote:
> On 18/03/2020 09:56, Jan Beulich wrote:
> > On 17.03.2020 22:52, David Woodhouse wrote:
> > > On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
> > > > > @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn,
> > > > > uint32_t *status)
> > > > >       do {
> > > > >           ret = *status = 0;
> > > > > -        if ( y & PGC_broken )
> > > > > +        if ( (y & PGC_state) == PGC_state_broken ||
> > > > > +             (y & PGC_state) == PGC_state_broken_offlining )
> > > > >           {
> > > > >               ret = -EINVAL;
> > > > >               *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
> > > > >               break;
> > > > >           }
> > > > > -
> > > > > -        if ( (y & PGC_state) == PGC_state_offlined )
> > > > > +        else if ( (y & PGC_state) == PGC_state_offlined )
> > > > 
> > > > I don't see a need for adding "else" here.
> > > 
> > > They are mutually exclusive cases. It makes things a whole lot clearer
> > > to the reader to put the 'else' there, and sometimes helps a naïve
> > > compiler along the way too.
> > 
> > Well, I'm afraid I'm going to be pretty strict about this: It's again
> > a matter of taste, yes, but we generally try to avoid pointless else.
> > What you consider "a whole lot clearer to the reader" is the opposite
> > from my pov.
> 
> While I agree the 'else' may be pointless, I don't think it is worth an 
> argument. As the author of the patch, it is his choice to write the code 
> like that.

Indeed. While I appreciate your insight, Jan, and your detailed reviews
are undoubtedly helpful — especially to me as I poke around the Xen
code base without knowing where the bodies are buried — I do sometimes
find that it degenerates into what appears to be gratuitous
bikeshedding.

Like *some* others, I'm perfectly capable of responding "I understand
you would have done it differently, but I prefer it this way".

But even for those like me who have the self-confidence (or arrogance?)
to respond in such a way, the end result is often the same — a patch
series which the maintainer doesn't apply because it has "unresolved
issues".

Perfect is the enemy of good. Especially when perfection is so
subjective.

This definitely isn't the kind of welcoming community that I enjoy
trying to get my junior engineers to contribute to. And they aren't
snowflakes; they cope with the Linux community just fine, for the most
part.


Earlier today, I found myself adjusting a patch in order to tweak the
behaviour of a "can never happen" situation, when it was far from clear
that the *existing* behaviour in that situation would have been correct
anyway.

There is a lot of value in your reviews, and they are appreciated. But
the overall effect is seen by some as making the Xen community somewhat
dysfunctional. 

The -MP makefile patch I posted yesterday... I almost didn't bother.
And when I allowed myself to be talked into it, I was entirely
unsurprised when a review came in basically asking me to prove a
negative before the patch could proceed. So as far as I can tell, it'll
fall by the wayside and the build will remain broken any time anyone
removes or renames a header file. Because life's too short to invest
the energy to make improvements like that.

One of these days, I may attempt to revive my series cleaning up the
16-bit and 32-bit boot code. Which was a clear improvement and
simplification, and again you gave extremely valid feedback which
helped to improve it — but again it was interspersed with more
subjective and less helpful comments which basically derailed it. 

Having carefully threaded my way through the existing byzantine code
and made incremental bisectable changes, I ended up with feedback that
basically would have required me to start again from scratch, in order
to satisfy what appeared to be fairly arbitrary and subjective demands.

As is often the case in creating a bisectable series out of complex
changes, I had sometimes moved/refactored code, only to move/refactor
it again in a subsequent patch. Sometimes the ordering of such inter-
related changes can be fairly arbitrary, and I had made my choices as
I'd progressed; the real focus being the end result. At one point you
were picking on intermediate details of how I'd made my overall series
bisectable, and seemed to be demanding that I go back and refactor (the
intermediate stages for no better reason than because you would have
done it differently. 

Again, your attention to detail and your expertise are massively
appreciated. But please let's remember that "perfect is the enemy of
good", and strike a balance which allows forward progress without
blocking improvements.

Sometimes I wonder if you truly realise how much you derail the
progress of a patch series just by raising well-intentioned "queries"
around it.


> > > > > --- a/xen/include/asm-x86/mm.h
> > > > > +++ b/xen/include/asm-x86/mm.h
> > > > > @@ -67,18 +67,27 @@
> > > > >    /* 3-bit PAT/PCD/PWT cache-attribute hint. */
> > > > >   #define PGC_cacheattr_base PG_shift(6)
> > > > >   #define PGC_cacheattr_mask PG_mask(7, 6)
> > > > > - /* Page is broken? */
> > > > > -#define _PGC_broken       PG_shift(7)
> > > > > -#define PGC_broken        PG_mask(1, 7)
> > > > > - /* Mutually-exclusive page states: { inuse, offlining, offlined,
> > > > > free }. */
> > > > > -#define PGC_state         PG_mask(3, 9)
> > > > > -#define PGC_state_inuse   PG_mask(0, 9)
> > > > > -#define PGC_state_offlining PG_mask(1, 9)
> > > > > -#define PGC_state_offlined PG_mask(2, 9)
> > > > > -#define PGC_state_free    PG_mask(3, 9)
> > > > > -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) ==
> > > > > PGC_state_##st)
> > > > > -
> > > > > - /* Count of references to this frame. */
> > > > > + /*
> > > > > +  * Mutually-exclusive page states:
> > > > > +  * { inuse, offlining, offlined, free, broken_offlining, broken }
> > > > > +  */
> > > > > +#define PGC_state                  PG_mask(7, 9)
> > > > > +#define PGC_state_inuse            PG_mask(0, 9)
> > > > > +#define PGC_state_offlining        PG_mask(1, 9)
> > > > > +#define PGC_state_offlined         PG_mask(2, 9)
> > > > > +#define PGC_state_free             PG_mask(3, 9)
> > > > > +#define PGC_state_broken_offlining PG_mask(4, 9)
> > > > 
> > > > TBH I'd prefer PGC_state_offlining_broken, as it's not the
> > > > offlining which is broken, but a broken page is being
> > > > offlined.
> > > 
> > > It is the page which is both broken and offlining.
> > > Or indeed it is the page which is both offlining and broken.
> > 
> > I.e. you agree with flipping the two parts around?

I hope I have respectfully made it clear that no, I'm really not happy
with the very concept of such a request.

Perhaps it would be easier for me to acquiesce, in the short term.

But on the whole I think it's better to put my foot down and say 'no',
and focus on real work and things that matter.

> > > > > +#define page_is_offlining(pg)      (page_state_is((pg), 
> > > > > broken_offlining) || \
> > > > > +                                    page_state_is((pg), offlining))
> > > > 
> > > > Overall I wonder whether the PGC_state_* ordering couldn't be
> > > > adjusted such that at least some of these three won't need
> > > > two comparisons (by masking off a bit before comparing).
> > > 
> > > The whole point in this exercise is that there isn't a whole bit for
> > > these; they are each *two* states out of the possible 8.
> > 
> > Sure. But just consider the more general case: Instead of writing
> > 
> >      if ( i == 6 || i == 7 )
> > 
> > you can as well write
> > 
> >      if ( (i | 1) == 7 )
> 
> I stumbled accross a few of those recently and this is not the obvious 
> things to read. Yes, your code may be faster. But is it really worth it 
> compare to the cost of readability and futureproofness?

No. Just no.

If that kind of change is really a worthwhile win, it'll depend on the
CPU. File a GCC PR with a test case as a missed optimisation. Don't
make the source code gratuitously harder to read.

Honestly, this, right here, is the *epitome* of why I, and others,
sometimes feel that submitting a patch to Xen can be more effort than
it's worth.

This email is not intended as a personal attack; I hope you don't feel
that it is. For about the fifth time: your careful reviews and your
attention to detail are *massively* appreciated. Just a little over the
time sometimes, and I'd like to ask you to take care to be aware of the
overall effect, and that you are not blocking progress.

Thanks.



[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 18.03.2020 18:13, David Woodhouse wrote:
> On Wed, 2020-03-18 at 12:31 +0000, Julien Grall wrote:
>> On 18/03/2020 09:56, Jan Beulich wrote:
>>> On 17.03.2020 22:52, David Woodhouse wrote:
>>>> On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
>>>>>> @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn,
>>>>>> uint32_t *status)
>>>>>>       do {
>>>>>>           ret = *status = 0;
>>>>>> -        if ( y & PGC_broken )
>>>>>> +        if ( (y & PGC_state) == PGC_state_broken ||
>>>>>> +             (y & PGC_state) == PGC_state_broken_offlining )
>>>>>>           {
>>>>>>               ret = -EINVAL;
>>>>>>               *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
>>>>>>               break;
>>>>>>           }
>>>>>> -
>>>>>> -        if ( (y & PGC_state) == PGC_state_offlined )
>>>>>> +        else if ( (y & PGC_state) == PGC_state_offlined )
>>>>>
>>>>> I don't see a need for adding "else" here.
>>>>
>>>> They are mutually exclusive cases. It makes things a whole lot clearer
>>>> to the reader to put the 'else' there, and sometimes helps a naïve
>>>> compiler along the way too.
>>>
>>> Well, I'm afraid I'm going to be pretty strict about this: It's again
>>> a matter of taste, yes, but we generally try to avoid pointless else.
>>> What you consider "a whole lot clearer to the reader" is the opposite
>>> from my pov.
>>
>> While I agree the 'else' may be pointless, I don't think it is worth an 
>> argument. As the author of the patch, it is his choice to write the code 
>> like that.
> 
> Indeed. While I appreciate your insight, Jan, and your detailed reviews
> are undoubtedly helpful — especially to me as I poke around the Xen
> code base without knowing where the bodies are buried — I do sometimes
> find that it degenerates into what appears to be gratuitous
> bikeshedding.
> 
> Like *some* others, I'm perfectly capable of responding "I understand
> you would have done it differently, but I prefer it this way".
> 
> But even for those like me who have the self-confidence (or arrogance?)
> to respond in such a way, the end result is often the same — a patch
> series which the maintainer doesn't apply because it has "unresolved
> issues".
> 
> Perfect is the enemy of good. Especially when perfection is so
> subjective.
> 
> This definitely isn't the kind of welcoming community that I enjoy
> trying to get my junior engineers to contribute to. And they aren't
> snowflakes; they cope with the Linux community just fine, for the most
> part.

I appreciate your open an honest feedback, and having had similar
comments in the past I can assure you that I've already tried to
adjust where I find this acceptable. I take it you realize that
there are two limitations in this - trying doesn't mean succeeding,
and the boundaries of what I'd consider acceptable to let go with
no comments.

Of course there are always two sides of the medal.

As a maintainer of some piece of code, I view it as my
responsibility to look after not only the technical correctness of
that code, but also after its style (in the broadest sense of the
word). Looking at some very bad examples in our tree, many of
which I'm afraid have a Linux origin, I'm in particular of the
opinion that consistent style is a significant aid in code
readability and maintainability. And I hope you agree that _style_
adjustments are pretty easy to make, so I don't view asking for
such as placing a significant burden on the submitter. The
alternative of letting it go uncommented and then take the time
myself to clean up seems quite a bit worse to me, not the least
because of this scaling even less well than the amount of code
review that needs doing.

The mentioned Linux origin of some of the particularly bad
examples in our tree is why I view your "they cope with the Linux
community just fine" as not really applicable. This is despite
our subsequent changes to those files often having made the
situation worse rather than better.

To some degree the same goes for bigger than necessary code churn,
albeit I agree that in a number of cases it is far less objective
to judge than the aim for consistent style. Extra code churn
instead is often making review harder, irrespective of the often
good intentions behind doing so.

> There is a lot of value in your reviews, and they are appreciated. But
> the overall effect is seen by some as making the Xen community somewhat
> dysfunctional. 

In which case I ought to consider, of course after first checking
with my management, to step back as a maintainer. I'd very much
regret doing so, but if it's in the interest of the community ...

(As an aside, likely being among those doing the largest part of
code reviews, helping with that part of the overall workload the
project generates would reduce the number of reviews I'd have to
do, and hence the chances of me giving comments viewed as
unhelpful or worse by submitters. Or, to put it in different,
frank, but hopefully not offending words - I'd like to see you do
a fair amount of code review, including looking after merely
cosmetic aspects in the spirit of our written and unwritten rules,
before you actually comment on me going too far with some of my
feedback. And without me wanting to put too much emphasis on this:
It is my opinion that maintainer views generally have somewhat
higher weight than non-maintainer ones. I'm not going to claim
though there aren't cases where I might go too far and hence abuse
rather than use this, but as per above I can only try to avoid
doing so, I can't promise to succeed. And of course I, like others,
can be convinced to be wrong.)

> The -MP makefile patch I posted yesterday... I almost didn't bother.
> And when I allowed myself to be talked into it, I was entirely
> unsurprised when a review came in basically asking me to prove a
> negative before the patch could proceed. So as far as I can tell, it'll
> fall by the wayside and the build will remain broken any time anyone
> removes or renames a header file. Because life's too short to invest
> the energy to make improvements like that.

So are you saying that as a maintainer I should let go uncommented a
change which I'm unconvinced doesn't have negative side effects,
besides its positive intended behavioral change? The more that here
the workaround is rather trivial? As you may imagine, I've run into
the situation myself a number of times, without considering this a
reason to make any adjustments to the build machinery.

>>>>>> --- a/xen/include/asm-x86/mm.h
>>>>>> +++ b/xen/include/asm-x86/mm.h
>>>>>> @@ -67,18 +67,27 @@
>>>>>>    /* 3-bit PAT/PCD/PWT cache-attribute hint. */
>>>>>>   #define PGC_cacheattr_base PG_shift(6)
>>>>>>   #define PGC_cacheattr_mask PG_mask(7, 6)
>>>>>> - /* Page is broken? */
>>>>>> -#define _PGC_broken       PG_shift(7)
>>>>>> -#define PGC_broken        PG_mask(1, 7)
>>>>>> - /* Mutually-exclusive page states: { inuse, offlining, offlined,
>>>>>> free }. */
>>>>>> -#define PGC_state         PG_mask(3, 9)
>>>>>> -#define PGC_state_inuse   PG_mask(0, 9)
>>>>>> -#define PGC_state_offlining PG_mask(1, 9)
>>>>>> -#define PGC_state_offlined PG_mask(2, 9)
>>>>>> -#define PGC_state_free    PG_mask(3, 9)
>>>>>> -#define page_state_is(pg, st) (((pg)->count_info&PGC_state) ==
>>>>>> PGC_state_##st)
>>>>>> -
>>>>>> - /* Count of references to this frame. */
>>>>>> + /*
>>>>>> +  * Mutually-exclusive page states:
>>>>>> +  * { inuse, offlining, offlined, free, broken_offlining, broken }
>>>>>> +  */
>>>>>> +#define PGC_state                  PG_mask(7, 9)
>>>>>> +#define PGC_state_inuse            PG_mask(0, 9)
>>>>>> +#define PGC_state_offlining        PG_mask(1, 9)
>>>>>> +#define PGC_state_offlined         PG_mask(2, 9)
>>>>>> +#define PGC_state_free             PG_mask(3, 9)
>>>>>> +#define PGC_state_broken_offlining PG_mask(4, 9)
>>>>>
>>>>> TBH I'd prefer PGC_state_offlining_broken, as it's not the
>>>>> offlining which is broken, but a broken page is being
>>>>> offlined.
>>>>
>>>> It is the page which is both broken and offlining.
>>>> Or indeed it is the page which is both offlining and broken.
>>>
>>> I.e. you agree with flipping the two parts around?
> 
> I hope I have respectfully made it clear that no, I'm really not happy
> with the very concept of such a request.
> 
> Perhaps it would be easier for me to acquiesce, in the short term.
> 
> But on the whole I think it's better to put my foot down and say 'no',
> and focus on real work and things that matter.

Well, in the specific case here I've meanwhile realized that my
alternative naming suggested in in no way less ambiguous. So
stick to what you've chosen, albeit I continue to dislike the
name ambiguously also suggesting that the offlining operation
might be broken (e.g. as in "can't be offlined"), rather than the
page itself. I'm not going to exclude though that this is just
because of not being a native English speaker.

>>>>>> +#define page_is_offlining(pg)      (page_state_is((pg), 
>>>>>> broken_offlining) || \
>>>>>> +                                    page_state_is((pg), offlining))
>>>>>
>>>>> Overall I wonder whether the PGC_state_* ordering couldn't be
>>>>> adjusted such that at least some of these three won't need
>>>>> two comparisons (by masking off a bit before comparing).
>>>>
>>>> The whole point in this exercise is that there isn't a whole bit for
>>>> these; they are each *two* states out of the possible 8.
>>>
>>> Sure. But just consider the more general case: Instead of writing
>>>
>>>      if ( i == 6 || i == 7 )
>>>
>>> you can as well write
>>>
>>>      if ( (i | 1) == 7 )
>>
>> I stumbled accross a few of those recently and this is not the obvious 
>> things to read. Yes, your code may be faster. But is it really worth it 
>> compare to the cost of readability and futureproofness?
> 
> No. Just no.
> 
> If that kind of change is really a worthwhile win, it'll depend on the
> CPU. File a GCC PR with a test case as a missed optimisation.

Your experience may be different, but I hardly ever see any effect from
reporting bugs (not just against gcc) unless they're really bad or really
easy to address. That's why I generally prefer to fix such issues myself,
provided of course that I can find the time.

> Don't make the source code gratuitously harder to read.

That's a very subjective aspect again. Personally I find two comparisons
of the same variable against different constants harder to read.

> Honestly, this, right here, is the *epitome* of why I, and others,
> sometimes feel that submitting a patch to Xen can be more effort than
> it's worth.

Note how I said "I wonder", not "please make".

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 16193 bytes --]

On Thu, 2020-03-19 at 09:49 +0100, Jan Beulich wrote:
> On 18.03.2020 18:13, David Woodhouse wrote:
> > On Wed, 2020-03-18 at 12:31 +0000, Julien Grall wrote:
> > > On 18/03/2020 09:56, Jan Beulich wrote:
> > > > On 17.03.2020 22:52, David Woodhouse wrote:
> > > > > On Thu, 2020-02-20 at 12:10 +0100, Jan Beulich wrote:
> > > > > > > @@ -1699,14 +1714,14 @@ unsigned int online_page(mfn_t mfn,
> > > > > > > uint32_t *status)
> > > > > > >       do {
> > > > > > >           ret = *status = 0;
> > > > > > > -        if ( y & PGC_broken )
> > > > > > > +        if ( (y & PGC_state) == PGC_state_broken ||
> > > > > > > +             (y & PGC_state) == PGC_state_broken_offlining )
> > > > > > >           {
> > > > > > >               ret = -EINVAL;
> > > > > > >               *status = PG_ONLINE_FAILED |PG_ONLINE_BROKEN;
> > > > > > >               break;
> > > > > > >           }
> > > > > > > -
> > > > > > > -        if ( (y & PGC_state) == PGC_state_offlined )
> > > > > > > +        else if ( (y & PGC_state) == PGC_state_offlined )
> > > > > > 
> > > > > > I don't see a need for adding "else" here.
> > > > > 
> > > > > They are mutually exclusive cases. It makes things a whole lot clearer
> > > > > to the reader to put the 'else' there, and sometimes helps a naïve
> > > > > compiler along the way too.
> > > > 
> > > > Well, I'm afraid I'm going to be pretty strict about this: It's again
> > > > a matter of taste, yes, but we generally try to avoid pointless else.
> > > > What you consider "a whole lot clearer to the reader" is the opposite
> > > > from my pov.
> > > 
> > > While I agree the 'else' may be pointless, I don't think it is worth an 
> > > argument. As the author of the patch, it is his choice to write the code 
> > > like that.
> > 
> > Indeed. While I appreciate your insight, Jan, and your detailed reviews
> > are undoubtedly helpful — especially to me as I poke around the Xen
> > code base without knowing where the bodies are buried — I do sometimes
> > find that it degenerates into what appears to be gratuitous
> > bikeshedding.
> > 
> > Like *some* others, I'm perfectly capable of responding "I understand
> > you would have done it differently, but I prefer it this way".
> > 
> > But even for those like me who have the self-confidence (or arrogance?)
> > to respond in such a way, the end result is often the same — a patch
> > series which the maintainer doesn't apply because it has "unresolved
> > issues".
> > 
> > Perfect is the enemy of good. Especially when perfection is so
> > subjective.
> > 
> > This definitely isn't the kind of welcoming community that I enjoy
> > trying to get my junior engineers to contribute to. And they aren't
> > snowflakes; they cope with the Linux community just fine, for the most
> > part.
> 
> I appreciate your open an honest feedback, and having had similar
> comments in the past I can assure you that I've already tried to
> adjust where I find this acceptable. I take it you realize that
> there are two limitations in this - trying doesn't mean succeeding,
> and the boundaries of what I'd consider acceptable to let go with
> no comments.
> 
> Of course there are always two sides of the medal.
> 
> As a maintainer of some piece of code, I view it as my
> responsibility to look after not only the technical correctness of
> that code, but also after its style (in the broadest sense of the
> word). Looking at some very bad examples in our tree, many of
> which I'm afraid have a Linux origin, I'm in particular of the
> opinion that consistent style is a significant aid in code
> readability and maintainability. And I hope you agree that _style_
> adjustments are pretty easy to make, so I don't view asking for
> such as placing a significant burden on the submitter. The
> alternative of letting it go uncommented and then take the time
> myself to clean up seems quite a bit worse to me, not the least
> because of this scaling even less well than the amount of code
> review that needs doing.

Yes, 100% agreed. And I'll even concede that for the cases of moving
code around that happens to not conform to the current style, and
asking contributors to fix it up as they go.

I was agreeing with you on that point, while simultaneously telling
Julien "nah, I'll fix it while I'm here" when he suggested that I *not*
realign the PGC_state bit definitions.


> The mentioned Linux origin of some of the particularly bad
> examples in our tree is why I view your "they cope with the Linux
> community just fine" as not really applicable. This is despite
> our subsequent changes to those files often having made the
> situation worse rather than better.

Was more about the community effect than technical matters, but let's
not rathole on that.

> To some degree the same goes for bigger than necessary code churn,
> albeit I agree that in a number of cases it is far less objective
> to judge than the aim for consistent style. Extra code churn
> instead is often making review harder, irrespective of the often
> good intentions behind doing so.

Completely agreed.

> > There is a lot of value in your reviews, and they are appreciated. But
> > the overall effect is seen by some as making the Xen community somewhat
> > dysfunctional. 
> 
> In which case I ought to consider, of course after first checking
> with my management, to step back as a maintainer. I'd very much
> regret doing so, but if it's in the interest of the community ...

I definitely don't think that would be in the interest of the
community. As I think I may have mentioned once or twice in my previous
message, your detailed reviews are massively appreciated and useful.

> (As an aside, likely being among those doing the largest part of
> code reviews, helping with that part of the overall workload the
> project generates would reduce the number of reviews I'd have to
> do, and hence the chances of me giving comments viewed as
> unhelpful or worse by submitters. Or, to put it in different,
> frank, but hopefully not offending words - I'd like to see you do
> a fair amount of code review, including looking after merely
> cosmetic aspects in the spirit of our written and unwritten rules,
> before you actually comment on me going too far with some of my
> feedback. And without me wanting to put too much emphasis on this:
> It is my opinion that maintainer views generally have somewhat
> higher weight than non-maintainer ones. I'm not going to claim
> though there aren't cases where I might go too far and hence abuse
> rather than use this, but as per above I can only try to avoid
> doing so, I can't promise to succeed. And of course I, like others,
> can be convinced to be wrong.)

Understood.

> > The -MP makefile patch I posted yesterday... I almost didn't bother.
> > And when I allowed myself to be talked into it, I was entirely
> > unsurprised when a review came in basically asking me to prove a
> > negative before the patch could proceed. So as far as I can tell, it'll
> > fall by the wayside and the build will remain broken any time anyone
> > removes or renames a header file. Because life's too short to invest
> > the energy to make improvements like that.
> 
> So are you saying that as a maintainer I should let go uncommented a
> change which I'm unconvinced doesn't have negative side effects,
> besides its positive intended behavioral change? The more that here
> the workaround is rather trivial? As you may imagine, I've run into
> the situation myself a number of times, without considering this a
> reason to make any adjustments to the build machinery.

Jan, I would respectfully request that you take another look at your
initial response, but put yourself in the shoes of a patch submitter:
https://lists.xenproject.org/archives/html/xen-devel/2020-03/msg01171.html

You mention a "simple" workaround... but the workaround I've been using
is to manually remove the offending .o.d files, one at a time (or at
least one directory at a time), until the broken build starts working
again. Is that what you meant? And you really didn't ever consider that
it should be fixed?

And the substance of the response is basically saying "this is voodoo
and we can't touch it or unspecified things might break, but I have no
idea where to tell you to look."

Looking back I realise that the concern about phony rules overriding
pattern rules didn't even come from you; your concern was more nebulous
and unaddressable. It looks like I came up with a straw man and shot
*that* down in my later analysis (although that wasn't my intent; I
think the concern about pattern rules really did come from somewhere).

You asked a question about "why isn't this default behaviour", which is
kind of a silly question when asking about an option (-MP) that was
added to GCC almost a decade after the initial -MD behaviour was
established. Of *course* they didn't retroactively change the default.


Read that message again from the point of view of a contributor.
Pretend it isn't even me; pretend it's someone attempting to make their
first, trivial, improvement to the Xen ecosystem.

I hope you'll understand why my initial reaction was just a
monosyllabic 'no'.


> > > > > > > +#define PGC_state_broken_offlining PG_mask(4, 9)
> > > > > > 
> > > > > > TBH I'd prefer PGC_state_offlining_broken, as it's not the
> > > > > > offlining which is broken, but a broken page is being
> > > > > > offlined.
> > > > > 
> > > > > It is the page which is both broken and offlining.
> > > > > Or indeed it is the page which is both offlining and broken.
> > > > 
> > > > I.e. you agree with flipping the two parts around?
> > 
> > I hope I have respectfully made it clear that no, I'm really not happy
> > with the very concept of such a request.
> > 
> > Perhaps it would be easier for me to acquiesce, in the short term.
> > 
> > But on the whole I think it's better to put my foot down and say 'no',
> > and focus on real work and things that matter.
> 
> Well, in the specific case here I've meanwhile realized that my
> alternative naming suggested in in no way less ambiguous. So
> stick to what you've chosen, albeit I continue to dislike the
> name ambiguously also suggesting that the offlining operation
> might be broken (e.g. as in "can't be offlined"), rather than the
> page itself. I'm not going to exclude though that this is just
> because of not being a native English speaker.

As a native English speaker, the naming of these bothered me too.
They're too long and redundant. But subsuming the PGC_broken but into a
3-bit PGC_state makes sense, and we can't abandon that idea purely
because we can't come up with a *name* that fills us with joy.

There wasn't a *good* answer. I vacillated for a while, and picked the
one that offended me least.

And then ended up in a debate about it when it really wasn't important.


> > > > > > > +#define page_is_offlining(pg)      (page_state_is((pg), 
> > > > > > > broken_offlining) || \
> > > > > > > +                                    page_state_is((pg), offlining))
> > > > > > 
> > > > > > Overall I wonder whether the PGC_state_* ordering couldn't be
> > > > > > adjusted such that at least some of these three won't need
> > > > > > two comparisons (by masking off a bit before comparing).
> > > > > 
> > > > > The whole point in this exercise is that there isn't a whole bit for
> > > > > these; they are each *two* states out of the possible 8.
> > > > 
> > > > Sure. But just consider the more general case: Instead of writing
> > > > 
> > > >      if ( i == 6 || i == 7 )
> > > > 
> > > > you can as well write
> > > > 
> > > >      if ( (i | 1) == 7 )
> > > 
> > > I stumbled accross a few of those recently and this is not the obvious 
> > > things to read. Yes, your code may be faster. But is it really worth it 
> > > compare to the cost of readability and futureproofness?
> > 
> > No. Just no.
> > 
> > If that kind of change is really a worthwhile win, it'll depend on the
> > CPU. File a GCC PR with a test case as a missed optimisation.
> 
> Your experience may be different, but I hardly ever see any effect from
> reporting bugs (not just against gcc) unless they're really bad or really
> easy to address. That's why I generally prefer to fix such issues myself,
> provided of course that I can find the time.

Perhaps so. But if I *don't* file it, it *certainly* doesn't get fixed.

And I've learned over the years *not* to second-guess the optimisations
that today's compiler might make, with the wind blowing in this
particular direction.

FWIW 'return (x == 6 || x == 7)' ends up being emitted by GCC on x86 as

	subl	$6, %edi
	xorl	%eax, %eax
	cmpl	$1, %edi
	setbe	%al
	ret

And 'return (x == 5 || x == 7)' gives:

	andl	$-3, %edi
	xorl	%eax, %eax
	cmpl	$5, %edi
	sete	%al
	ret

So it does look like GCC is actually doing its job, on this occasion.

But that's entirely beside the point, which is that I'm having some
pointless discussion about compiler optimisation minutiæ when fixing
PGC_broken was *already* deep into yak-shaving for the improvement I
was *actually* trying to make. It's distracting from real work, raising
barriers to getting fixes merged.

> > Don't make the source code gratuitously harder to read.
> 
> That's a very subjective aspect again. Personally I find two comparisons
> of the same variable against different constants harder to read.
> 
> > Honestly, this, right here, is the *epitome* of why I, and others,
> > sometimes feel that submitting a patch to Xen can be more effort than
> > it's worth.
> 
> Note how I said "I wonder", not "please make".

Perspective again. That distinction really doesn't matter. Perhaps you
underestimate the weight your words carry, as a well-respected
maintainer. You can't negate that effect purely by word tricks like
saying 'I wonder'. 

Because understatement is a very common tool in the English language,
especially in British English — and we've all seen people write "I
wonder if you should..." when what was really meant was "I will set
fire to you if you don't...". 

Understatement like that doesn't work. It still derails the patch
review. It just didn't need to be said, in that context.

Let me repeat — because I've only said it once today, I think, that
your reviews are incredibly useful. I'm only asking that you recognise
the weight that your 'wondering' can have, and recognise when something
you are asking for is *subjective*.

A review is not about "is this code precisely how it would look if I
had written it myself", but it is about "is this code correct and
maintainable".

Sometimes, as in the example with the PGC_state_ naming above, there
isn't a "nice" answer. We pick the solution that offends us least. And
I completely understand as a maintainer, what it's like to be on the
receiving end of such a choice. You think "that could be nicer"... and
have to work through the alternatives yourself before you realise that
actually, it was the best of the choices available.

Each of the responses I've identified from you as 'excessive' has some
merit, we can focus on each of them and you can justify them, to a
certain extent. But as a whole, the effect is of a barrage of nitpicks
of questionable utility which really does hinder forward progress.

Let's try to focus on comments which will genuinely make the code
better. It's not that we should deliberately stop paying attention to
detail, or deliberately allow buggy and broken code into the tree. It's
that we should be aware that "perfect is the enemy of good".

For my part, I'll stop whining at you now. If I end up giving responses
to parts of your code review which seem to be along the lines of
"that's nice, dear, but I didn't think so and I did the typing", please
hark back to this conversation. I'll try to phrase them more
appropriately than that, but no promises :)

Thanks.


[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 19.03.2020 11:26, David Woodhouse wrote:
> On Thu, 2020-03-19 at 09:49 +0100, Jan Beulich wrote:
>> On 18.03.2020 18:13, David Woodhouse wrote:
>>> The -MP makefile patch I posted yesterday... I almost didn't bother.
>>> And when I allowed myself to be talked into it, I was entirely
>>> unsurprised when a review came in basically asking me to prove a
>>> negative before the patch could proceed. So as far as I can tell, it'll
>>> fall by the wayside and the build will remain broken any time anyone
>>> removes or renames a header file. Because life's too short to invest
>>> the energy to make improvements like that.
>>
>> So are you saying that as a maintainer I should let go uncommented a
>> change which I'm unconvinced doesn't have negative side effects,
>> besides its positive intended behavioral change? The more that here
>> the workaround is rather trivial? As you may imagine, I've run into
>> the situation myself a number of times, without considering this a
>> reason to make any adjustments to the build machinery.
> 
> Jan, I would respectfully request that you take another look at your
> initial response, but put yourself in the shoes of a patch submitter:
> https://lists.xenproject.org/archives/html/xen-devel/2020-03/msg01171.html
> 
> You mention a "simple" workaround... but the workaround I've been using
> is to manually remove the offending .o.d files, one at a time (or at
> least one directory at a time), until the broken build starts working
> again. Is that what you meant? And you really didn't ever consider that
> it should be fixed?

No, the workaround I've been using (after initially doing the expensive
one you describe) was to simply put in an empty file (or perhaps one
with an #error directive) in the place of the deleted one, rebuild, and
all .*.o.d files would have been updated. I might do so even before
fully deleting the file.

> And the substance of the response is basically saying "this is voodoo
> and we can't touch it or unspecified things might break, but I have no
> idea where to tell you to look."
> 
> Looking back I realise that the concern about phony rules overriding
> pattern rules didn't even come from you; your concern was more nebulous
> and unaddressable. It looks like I came up with a straw man and shot
> *that* down in my later analysis (although that wasn't my intent; I
> think the concern about pattern rules really did come from somewhere).
> 
> You asked a question about "why isn't this default behaviour", which is
> kind of a silly question when asking about an option (-MP) that was
> added to GCC almost a decade after the initial -MD behaviour was
> established. Of *course* they didn't retroactively change the default.

I don't see at all why this would be "of course" - if there really
was no undue side effect, why couldn't they?

> Read that message again from the point of view of a contributor.
> Pretend it isn't even me; pretend it's someone attempting to make their
> first, trivial, improvement to the Xen ecosystem.
> 
> I hope you'll understand why my initial reaction was just a
> monosyllabic 'no'.

To be honest- no, I don't. I didn't say "no way". Instead I asked
back to see whether there's more background to this. It is a useful
piece of information to know that -MP post-dates -MD by 10 or more
years. It's still speculation of why a new option was added rather
than making this default behavior, but I feel less afraid of the
change this way than by an implied "this not going to do any harm"
without really being certain why there is a separate option in the
first place (and gcc doc also not saying anything to this effect).

I can certainly follow your sentiment, not the least because
especially in my early days I also frequently got back replies I
didn't like, in various projects. Yet in a case like this one I'm
afraid it is not the reviewer's job to point out the unsafety of
a change, but it's the submitter who has to (sufficiently) prove
that a change won't break anything. Yes, in the typical case,
when there's a recognizable bug, the reviewer would point this
out. But there are cases where there's no obvious bug, but
experience (and, as so often, insufficient documentation) tells
one to be wary of changes of, possibly, any kind.

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[David Woodhouse]

[-- Attachment #1.1: Type: text/plain, Size: 3256 bytes --]

On Thu, 2020-03-19 at 12:59 +0100, Jan Beulich wrote:
> > Read that message again from the point of view of a contributor.
> > Pretend it isn't even me; pretend it's someone attempting to make
> > their first, trivial, improvement to the Xen ecosystem.
> > 
> > I hope you'll understand why my initial reaction was just a
> > monosyllabic 'no'.
> 
> To be honest- no, I don't. I didn't say "no way".

Then you have completely missed my point about how subtly understating
your 'objections' makes no difference at all to the outcome.

But OK, I'll come back to that at the end. You have made your intent
clear, more than once now, and we should take it on board.

>  Instead I asked back to see whether there's more background to this.
> It is a useful piece of information to know that -MP post-dates -MD
> by 10 or more years. It's still speculation of why a new option was
> added rather than making this default behavior, but I feel less
> afraid of the change this way than by an implied "this not going to
> do any harm" without really being certain why there is a separate
> option in the first place (and gcc doc also not saying anything to
> this effect).

It is not my job to make you feel less afraid of change.

> I can certainly follow your sentiment, not the least because
> especially in my early days I also frequently got back replies I
> didn't like, in various projects. Yet in a case like this one I'm
> afraid it is not the reviewer's job to point out the unsafety of
> a change, but it's the submitter who has to (sufficiently) prove
> that a change won't break anything. 

I'm sure you didn't mean it as such, Jan, but FYI that response could
be construed as being fairly patronising. If you were to direct it
towards someone who's even remotely inclined to feeling patronised,
that is. :)

> Yes, in the typical case, when there's a recognizable bug, the
> reviewer would point this out. But there are cases where there's no
> obvious bug, but experience (and, as so often, insufficient
> documentation) tells one to be wary of changes of, possibly, any
> kind.

I find this response to be purely obstructive and unhelpful. Your
response to my patch was basically asking me to prove a negative, and I
find myself surprised and disappointed that you cannot acknowledge
that. I didn't think our viewpoints were really that far apart; perhaps
I was wrong.

If there was an actual bug — or even the suspicion of a bug — I could
understand it. But this is just voodoo "we're too scared to change
things because we don't understand".

We are better than that. You can be better than that.

But I will take on board your comments about understatement and the
fact that you hadn't actually said "no". In future I shall consider
merely ignoring such interjections unless you explicitly state that you
are blocking the acceptance of a patch. Or, I suppose, resorting to the
style of monosyllabic answer that I had originally given in this case.

I trust that maintainers will take that on board too, and that open
"questions" from you in a thread will not be considered sufficient
reason not to merge a patch.

That seems to be what you're saying is your intent, yes? 


[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH 1/2] xen/mm: fold PGC_broken into PGC_state bits

[Jan Beulich]

On 19.03.2020 14:54, David Woodhouse wrote:
> On Thu, 2020-03-19 at 12:59 +0100, Jan Beulich wrote:
>>> Read that message again from the point of view of a contributor.
>>> Pretend it isn't even me; pretend it's someone attempting to make
>>> their first, trivial, improvement to the Xen ecosystem.
>>>
>>> I hope you'll understand why my initial reaction was just a
>>> monosyllabic 'no'.
>>
>> To be honest- no, I don't. I didn't say "no way".
> 
> Then you have completely missed my point about how subtly understating
> your 'objections' makes no difference at all to the outcome.
> 
> But OK, I'll come back to that at the end. You have made your intent
> clear, more than once now, and we should take it on board.
> 
>>  Instead I asked back to see whether there's more background to this.
>> It is a useful piece of information to know that -MP post-dates -MD
>> by 10 or more years. It's still speculation of why a new option was
>> added rather than making this default behavior, but I feel less
>> afraid of the change this way than by an implied "this not going to
>> do any harm" without really being certain why there is a separate
>> option in the first place (and gcc doc also not saying anything to
>> this effect).
> 
> It is not my job to make you feel less afraid of change.
> 
>> I can certainly follow your sentiment, not the least because
>> especially in my early days I also frequently got back replies I
>> didn't like, in various projects. Yet in a case like this one I'm
>> afraid it is not the reviewer's job to point out the unsafety of
>> a change, but it's the submitter who has to (sufficiently) prove
>> that a change won't break anything. 
> 
> I'm sure you didn't mean it as such, Jan, but FYI that response could
> be construed as being fairly patronising. If you were to direct it
> towards someone who's even remotely inclined to feeling patronised,
> that is. :)

I certainly didn't mean to, I apologize. (My dictionary gives me
several very different meanings of "patronize", so I'm somewhat
guessing which meaning you infer here.)

>> Yes, in the typical case, when there's a recognizable bug, the
>> reviewer would point this out. But there are cases where there's no
>> obvious bug, but experience (and, as so often, insufficient
>> documentation) tells one to be wary of changes of, possibly, any
>> kind.
> 
> I find this response to be purely obstructive and unhelpful.

I'm sorry if it feels like this to you.

> Your
> response to my patch was basically asking me to prove a negative, and I
> find myself surprised and disappointed that you cannot acknowledge
> that. I didn't think our viewpoints were really that far apart; perhaps
> I was wrong.

I'm certainly willing to acknowledge that I've asked a question
that may be difficult if possible at all to answer in a way that
we'd be fully certain in the end. Yet even after all of the
discussion we've had here I still think the question was
appropriate to ask. It continues to be unobvious to me that non-
default behavior of a tool would imply using this behavior is
going to be free of side effects. The historical aspect you've
dug out afterwards is at least a partial explanation which,
seeing that you've got an unconditional and a conditional ack,
is good enough for me to let the change go in, despite still
not being finally convinced of it being free of side effects.

> If there was an actual bug — or even the suspicion of a bug — I could
> understand it. But this is just voodoo "we're too scared to change
> things because we don't understand".

Not just this, but also because things had been broken in subtle
ways in the past. Until we get a better one, we have to live with
the build system being fragile here and there.

> We are better than that. You can be better than that.
> 
> But I will take on board your comments about understatement and the
> fact that you hadn't actually said "no". In future I shall consider
> merely ignoring such interjections unless you explicitly state that you
> are blocking the acceptance of a patch. Or, I suppose, resorting to the
> style of monosyllabic answer that I had originally given in this case.
> 
> I trust that maintainers will take that on board too, and that open
> "questions" from you in a thread will not be considered sufficient
> reason not to merge a patch.
> 
> That seems to be what you're saying is your intent, yes? 

My intent was to get clarification before the patch would go in.
I didn't mean to block it, but I also didn't see it go in without
such clarification. I'm struggling to see what's bad in asking
whether you/we are certain enough that a change won't have bad
side effects; if there were, we might treat an easy to work around
situation by one hard to recognize and address. Seeing you reply
just "no" seemed a fair answer to me (while I sensed a certain
level of annoyance), albeit not one that would resolve the
question. In anticipation I did include anyone else who might
know right away. Had I known the answer myself, I of course
wouldn't have asked.

Bottom line - when I say "no", I mean "no". When I ask a question
I expect it to be resolved, at least to a reasonable degree. When
I say "I wonder" I indeed mean just that; to me "may I suggest to
consider as an alternative" is simply more words, which may again
be an effect of English not being my native language. And when I
say "ack", I mean "ack". (I also didn't think I made any comments
about understatement; it was you who brought up that [cultural]
aspect.)

I'm afraid as a result of this discussion I'm now more confused
as to finding common grounds than I was before.

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel