tmpfs: fix spurious ENOSPC when racing with unswap

Testing the shmem_swaplist replacements for igrab() revealed another bug:
writes to /dev/loop0 on a tmpfs file which fills its filesystem were
sometimes failing with "Buffer I/O error"s.

These came from ENOSPC failures of shmem_getpage(), when racing with
swapoff: the same could happen when racing with another shmem_getpage(),
pulling the page in from swap in between our find_lock_page() and our
taking the info->lock (though not in the single-threaded loop case).

This is unacceptable, and surprising that I've not noticed it before:
it dates back many years, but (presumably) was made a lot easier to
reproduce in 2.6.36, which sited a page preallocation in the race window.

Fix it by rechecking the page cache before settling on an ENOSPC error.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 22 insertions, 10 deletions

diff --git a/mm/shmem.c b/mm/shmem.c
index dc17551..9e755c1 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1407,20 +1407,14 @@ repeat:
 		if (sbinfo->max_blocks) {
 			if (percpu_counter_compare(&sbinfo->used_blocks,
 						sbinfo->max_blocks) >= 0 ||
-			    shmem_acct_block(info->flags)) {
-				spin_unlock(&info->lock);
-				error = -ENOSPC;
-				goto failed;
-			}
+			    shmem_acct_block(info->flags))
+				goto nospace;
 			percpu_counter_inc(&sbinfo->used_blocks);
 			spin_lock(&inode->i_lock);
 			inode->i_blocks += BLOCKS_PER_PAGE;
 			spin_unlock(&inode->i_lock);
-		} else if (shmem_acct_block(info->flags)) {
-			spin_unlock(&info->lock);
-			error = -ENOSPC;
-			goto failed;
-		}
+		} else if (shmem_acct_block(info->flags))
+			goto nospace;
 
 		if (!filepage) {
 			int ret;
@@ -1500,6 +1494,24 @@ done:
 	error = 0;
 	goto out;
 
+nospace:
+	/*
+	 * Perhaps the page was brought in from swap between find_lock_page
+	 * and taking info->lock?  We allow for that at add_to_page_cache_lru,
+	 * but must also avoid reporting a spurious ENOSPC while working on a
+	 * full tmpfs.  (When filepage has been passed in to shmem_getpage, it
+	 * is already in page cache, which prevents this race from occurring.)
+	 */
+	if (!filepage) {
+		struct page *page = find_get_page(mapping, idx);
+		if (page) {
+			spin_unlock(&info->lock);
+			page_cache_release(page);
+			goto repeat;
+		}
+	}
+	spin_unlock(&info->lock);
+	error = -ENOSPC;
 failed:
 	if (*pagep != filepage) {
 		unlock_page(filepage);