[PATCH v2 1/2] mm: make PageReadahead more strict

[PATCH v2 1/2] mm: make PageReadahead more strict

[Minchan Kim]

Recently, I got some bugreports major page fault takes several seconds
sometime. When I review drop mmap_sem logic, I found several bugs.

   CPU 1                                                        CPU 2
mm_populate
 for ()
   ..
   ret = populate_vma_page_range
     __get_user_pages
       faultin_page
         handle_mm_fault
           filemap_fault
             do_async_mmap_readahead
                                                        shrink_page_list
                                                          pageout
                                                            SetPageReclaim(=SetPageReadahead)
                                                              writepage
                                                                SetPageWriteback
               if (PageReadahead(page))
                 maybe_unlock_mmap_for_io
                   up_read(mmap_sem)
                 page_cache_async_readahead()
                   if (PageWriteback(page))
                     return;

Here, since ret from populate_vma_page_range is zero, the loop continue
to run with same address with previous iteration. It will repeat the
loop until the page's writeout is done(ie, PG_writeback or PG_reclaim
is clear).

We could fix the above specific case via adding PageWriteback

   ret = populate_vma_page_range
           ...
           ...
           filemap_fault
             do_async_mmap_readahead
               if (!PageWriteback(page) && PageReadahead(page))
                 maybe_unlock_mmap_for_io
                   up_read(mmap_sem)
                 page_cache_async_readahead()
                   if (PageWriteback(page))
                     return;

Furthermore, to prevent potential issues caused by sharing PG_readahead
with PG_reclaim, let's make page flag wrapper for PageReadahead
with description. With that, we could remove PageWriteback check
in page_cache_async_readahead, which is more clear for maintenance/
readability.

Fixes: 6b4c9f446981 ("filemap: drop the mmap_sem for all blocking operations")
Signed-off-by: Minchan Kim <minchan@kernel.org>
---
 include/linux/page-flags.h | 28 ++++++++++++++++++++++++++--
 mm/readahead.c             |  6 ------
 2 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 1bf83c8fcaa7..f91a9b2a49bd 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -363,8 +363,32 @@ PAGEFLAG(MappedToDisk, mappedtodisk, PF_NO_TAIL)
 /* PG_readahead is only used for reads; PG_reclaim is only for writes */
 PAGEFLAG(Reclaim, reclaim, PF_NO_TAIL)
 	TESTCLEARFLAG(Reclaim, reclaim, PF_NO_TAIL)
-PAGEFLAG(Readahead, reclaim, PF_NO_COMPOUND)
-	TESTCLEARFLAG(Readahead, reclaim, PF_NO_COMPOUND)
+
+SETPAGEFLAG(Readahead, reclaim, PF_NO_COMPOUND)
+CLEARPAGEFLAG(Readahead, reclaim, PF_NO_COMPOUND)
+
+/*
+ * Since PG_readahead is shared with PG_reclaim of the page flags,
+ * PageReadahead should double check whether it's readahead marker
+ * or PG_reclaim. It could be done by PageWriteback check because
+ * PG_reclaim is always with PG_writeback.
+ */
+static inline int PageReadahead(struct page *page)
+{
+	VM_BUG_ON_PGFLAGS(PageCompound(page), page);
+
+	return (page->flags & (1UL << PG_reclaim | 1UL << PG_writeback)) ==
+		(1UL << PG_reclaim);
+}
+
+/* Clear PG_readahead only if it's PG_readahead, not PG_reclaim */
+static inline int TestClearPageReadahead(struct page *page)
+{
+	VM_BUG_ON_PGFLAGS(PageCompound(page), page);
+
+	return !PageWriteback(page) ||
+			test_and_clear_bit(PG_reclaim, &page->flags);
+}
 
 #ifdef CONFIG_HIGHMEM
 /*
diff --git a/mm/readahead.c b/mm/readahead.c
index 2fe72cd29b47..85b15e5a1d7b 100644
--- a/mm/readahead.c
+++ b/mm/readahead.c
@@ -553,12 +553,6 @@ page_cache_async_readahead(struct address_space *mapping,
 	if (!ra->ra_pages)
 		return;
 
-	/*
-	 * Same bit is used for PG_readahead and PG_reclaim.
-	 */
-	if (PageWriteback(page))
-		return;
-
 	ClearPageReadahead(page);
 
 	/*
-- 
2.25.0.265.gbab2e86ba0-goog

[PATCH v2 2/2] mm: fix long time stall from mm_populate

[Minchan Kim]

Basically, fault handler releases mmap_sem before requesting readahead
and then it is supposed to retry lookup the page from page cache with
FAULT_FLAG_TRIED so that it avoids the live lock of infinite retry.

However, what happens if the fault handler find a page from page
cache and the page has readahead marker but are waiting under
writeback? Plus one more condition, it happens under mm_populate
which repeats faulting unless it encounters error. So let's assemble
conditions below.

       CPU 1                                                        CPU 2

- first loop
    mm_populate
     for ()
       ..
       ret = populate_vma_page_range
         __get_user_pages
           faultin_page
             handle_mm_fault
               filemap_fault
                 do_async_mmap_readahead
                   if (PageReadahead(pageA))
                     maybe_unlock_mmap_for_io
                       up_read(mmap_sem)
					                    shrink_page_list
                                                              pageout
                                                                SetPageReclaim(=SetPageReadahead)(pageA)
                                                                writepage
                                                                  SetPageWriteback(pageA)

                     page_cache_async_readahead()
		       ClearPageReadahead(pageA)
                 do_async_mmap_readahead
		 lock_page_maybe_drop_mmap
		   goto out_retry

					                    the pageA is reclaimed
							    and new pageB is populated to the file offset
							    and finally has become PG_readahead

- second loop

	  __get_user_pages
           faultin_page
             handle_mm_fault
               filemap_fault
                 do_async_mmap_readahead
                   if (PageReadahead(pageB))
                     maybe_unlock_mmap_for_io
                       up_read(mmap_sem)
					                    shrink_page_list
                                                              pageout
                                                                SetPageReclaim(=SetPageReadahead)(pageB)
                                                                writepage
                                                                  SetPageWriteback(pageB)

                     page_cache_async_readahead()
		       ClearPageReadahead(pageB)
                 do_async_mmap_readahead
		 lock_page_maybe_drop_mmap
		   goto out_retry

It could be repeated forever so it's livelock. without involving reclaim,
it could happens if ra_pages become zero by fadvise/other threads who
have same fd one doing randome while the other one is sequential
because page_cache_async_readahead has following condition check like
PageWriteback and ra_pages are never synchrnized with fadvise and
shrink_readahead_size_eio from other threads.

void page_cache_async_readahead(struct address_space *mapping,
                           unsigned long req_size)
{
        /* no read-ahead */
        if (!ra->ra_pages)
                return;

Thus, we need to limit fault retry from mm_populate like page
fault handler.

Fixes: 6b4c9f446981 ("filemap: drop the mmap_sem for all blocking operations")
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Minchan Kim <minchan@kernel.org>
---
 mm/gup.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/mm/gup.c b/mm/gup.c
index 1b521e0ac1de..6f6548c63ad5 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -1133,7 +1133,7 @@ static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
  *
  * This takes care of mlocking the pages too if VM_LOCKED is set.
  *
- * return 0 on success, negative error code on error.
+ * return number of pages pinned on success, negative error code on error.
  *
  * vma->vm_mm->mmap_sem must be held.
  *
@@ -1196,6 +1196,7 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 	struct vm_area_struct *vma = NULL;
 	int locked = 0;
 	long ret = 0;
+	bool tried = false;
 
 	end = start + len;
 
@@ -1226,14 +1227,18 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 		 * double checks the vma flags, so that it won't mlock pages
 		 * if the vma was already munlocked.
 		 */
-		ret = populate_vma_page_range(vma, nstart, nend, &locked);
+		ret = populate_vma_page_range(vma, nstart, nend,
+						tried ? NULL : &locked);
 		if (ret < 0) {
 			if (ignore_errors) {
 				ret = 0;
 				continue;	/* continue at next VMA */
 			}
 			break;
-		}
+		} else if (ret == 0)
+			tried = true;
+		else
+			tried = false;
 		nend = nstart + ret * PAGE_SIZE;
 		ret = 0;
 	}
-- 
2.25.0.265.gbab2e86ba0-goog

Re: [PATCH v2 2/2] mm: fix long time stall from mm_populate

[Minchan Kim]

Bumping up.

On Fri, Feb 14, 2020 at 11:29:51AM -0800, Minchan Kim wrote:
> Basically, fault handler releases mmap_sem before requesting readahead
> and then it is supposed to retry lookup the page from page cache with
> FAULT_FLAG_TRIED so that it avoids the live lock of infinite retry.
> 
> However, what happens if the fault handler find a page from page
> cache and the page has readahead marker but are waiting under
> writeback? Plus one more condition, it happens under mm_populate
> which repeats faulting unless it encounters error. So let's assemble
> conditions below.
> 
>        CPU 1                                                        CPU 2
> 
> - first loop
>     mm_populate
>      for ()
>        ..
>        ret = populate_vma_page_range
>          __get_user_pages
>            faultin_page
>              handle_mm_fault
>                filemap_fault
>                  do_async_mmap_readahead
>                    if (PageReadahead(pageA))
>                      maybe_unlock_mmap_for_io
>                        up_read(mmap_sem)
> 					                    shrink_page_list
>                                                               pageout
>                                                                 SetPageReclaim(=SetPageReadahead)(pageA)
>                                                                 writepage
>                                                                   SetPageWriteback(pageA)
> 
>                      page_cache_async_readahead()
> 		       ClearPageReadahead(pageA)
>                  do_async_mmap_readahead
> 		 lock_page_maybe_drop_mmap
> 		   goto out_retry
> 
> 					                    the pageA is reclaimed
> 							    and new pageB is populated to the file offset
> 							    and finally has become PG_readahead
> 
> - second loop
> 
> 	  __get_user_pages
>            faultin_page
>              handle_mm_fault
>                filemap_fault
>                  do_async_mmap_readahead
>                    if (PageReadahead(pageB))
>                      maybe_unlock_mmap_for_io
>                        up_read(mmap_sem)
> 					                    shrink_page_list
>                                                               pageout
>                                                                 SetPageReclaim(=SetPageReadahead)(pageB)
>                                                                 writepage
>                                                                   SetPageWriteback(pageB)
> 
>                      page_cache_async_readahead()
> 		       ClearPageReadahead(pageB)
>                  do_async_mmap_readahead
> 		 lock_page_maybe_drop_mmap
> 		   goto out_retry
> 
> It could be repeated forever so it's livelock. without involving reclaim,
> it could happens if ra_pages become zero by fadvise/other threads who
> have same fd one doing randome while the other one is sequential
> because page_cache_async_readahead has following condition check like
> PageWriteback and ra_pages are never synchrnized with fadvise and
> shrink_readahead_size_eio from other threads.
> 
> void page_cache_async_readahead(struct address_space *mapping,
>                            unsigned long req_size)
> {
>         /* no read-ahead */
>         if (!ra->ra_pages)
>                 return;
> 
> Thus, we need to limit fault retry from mm_populate like page
> fault handler.
> 
> Fixes: 6b4c9f446981 ("filemap: drop the mmap_sem for all blocking operations")
> Reviewed-by: Jan Kara <jack@suse.cz>
> Signed-off-by: Minchan Kim <minchan@kernel.org>
> ---
>  mm/gup.c | 11 ++++++++---
>  1 file changed, 8 insertions(+), 3 deletions(-)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 1b521e0ac1de..6f6548c63ad5 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1133,7 +1133,7 @@ static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
>   *
>   * This takes care of mlocking the pages too if VM_LOCKED is set.
>   *
> - * return 0 on success, negative error code on error.
> + * return number of pages pinned on success, negative error code on error.
>   *
>   * vma->vm_mm->mmap_sem must be held.
>   *
> @@ -1196,6 +1196,7 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
>  	struct vm_area_struct *vma = NULL;
>  	int locked = 0;
>  	long ret = 0;
> +	bool tried = false;
>  
>  	end = start + len;
>  
> @@ -1226,14 +1227,18 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
>  		 * double checks the vma flags, so that it won't mlock pages
>  		 * if the vma was already munlocked.
>  		 */
> -		ret = populate_vma_page_range(vma, nstart, nend, &locked);
> +		ret = populate_vma_page_range(vma, nstart, nend,
> +						tried ? NULL : &locked);
>  		if (ret < 0) {
>  			if (ignore_errors) {
>  				ret = 0;
>  				continue;	/* continue at next VMA */
>  			}
>  			break;
> -		}
> +		} else if (ret == 0)
> +			tried = true;
> +		else
> +			tried = false;
>  		nend = nstart + ret * PAGE_SIZE;
>  		ret = 0;
>  	}
> -- 
> 2.25.0.265.gbab2e86ba0-goog
>