summaryrefslogtreecommitdiffstats
path: root/fs/buffer.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2008-04-04 14:38:17 -0700
committerLinus Torvalds <torvalds@linux-foundation.org>2008-04-04 14:38:17 -0700
commit1be62dc190ebaca331038962c873e7967de6cc4b (patch)
treea8eb0a106bf362819d2bb0fc602b0e52df6a4198 /fs/buffer.c
parent4ed919014eb2b591eb8fdd4dd00226a65faddef4 (diff)
downloadlinux-stable-1be62dc190ebaca331038962c873e7967de6cc4b.tar.gz
linux-stable-1be62dc190ebaca331038962c873e7967de6cc4b.tar.bz2
linux-stable-1be62dc190ebaca331038962c873e7967de6cc4b.zip
Be more careful about marking buffers dirty
Mikulas Patocka noted that the optimization where we check if a buffer was already dirty (and we avoid re-dirtying it) was not really SMP-safe. Since the read of the old status was not synchronized with anything, an aggressive CPU re-ordering of memory accesses might have moved that read up to before the data was even written to the buffer, and another CPU that cleaned it again, causing the newly dirty state to never actually hit the disk. Admittedly this would probably never trigger in practice, but it's still wrong. Mikulas sent a patch that fixed the problem, but I dislike the subtlety of the whole optimization, so this is an alternate fix that is more explicit about the particular SMP ordering for the optimization, and separates out the speculative reads of the buffer state into its own conditional (and makes the memory barrier only happen if we are likely to actually hit the optimized case in the first place). I considered removing the optimization entirely, but Andrew argued for it's continued existence. I'm a push-over. Cc: Mikulas Patocka <mikulas@artax.karlin.mff.cuni.cz> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'fs/buffer.c')
-rw-r--r--fs/buffer.c15
1 files changed, 14 insertions, 1 deletions
diff --git a/fs/buffer.c b/fs/buffer.c
index 98196327ddf0..39ff14403d13 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -1181,7 +1181,20 @@ __getblk_slow(struct block_device *bdev, sector_t block, int size)
void mark_buffer_dirty(struct buffer_head *bh)
{
WARN_ON_ONCE(!buffer_uptodate(bh));
- if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
+
+ /*
+ * Very *carefully* optimize the it-is-already-dirty case.
+ *
+ * Don't let the final "is it dirty" escape to before we
+ * perhaps modified the buffer.
+ */
+ if (buffer_dirty(bh)) {
+ smp_mb();
+ if (buffer_dirty(bh))
+ return;
+ }
+
+ if (!test_set_buffer_dirty(bh))
__set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
}