summaryrefslogtreecommitdiff
path: root/fs/xfs/xfs_aops.c
diff options
context:
space:
mode:
authorDave Chinner <dchinner@redhat.com>2013-05-20 09:51:08 +1000
committerBen Myers <bpm@sgi.com>2013-05-24 16:26:51 -0500
commit480d7467e4aaa3dc38088baf56bc3eb3599f5d26 (patch)
tree2ef613c4c8b1b4603135f0a777ee5d6dac7d0a4f /fs/xfs/xfs_aops.c
parentf722406faae2d073cc1d01063d1123c35425939e (diff)
downloadkernel-common-480d7467e4aaa3dc38088baf56bc3eb3599f5d26.tar.gz
kernel-common-480d7467e4aaa3dc38088baf56bc3eb3599f5d26.tar.bz2
kernel-common-480d7467e4aaa3dc38088baf56bc3eb3599f5d26.zip
xfs: fix sub-page blocksize data integrity writes
FSX on 512 byte block size filesystems has been failing for some time with corrupted data. The fault dates back to the change in the writeback data integrity algorithm that uses a mark-and-sweep approach to avoid data writeback livelocks. Unfortunately, a side effect of this mark-and-sweep approach is that each page will only be written once for a data integrity sync, and there is a condition in writeback in XFS where a page may require two writeback attempts to be fully written. As a result of the high level change, we now only get a partial page writeback during the integrity sync because the first pass through writeback clears the mark left on the page index to tell writeback that the page needs writeback.... The cause is writing a partial page in the clustering code. This can happen when a mapping boundary falls in the middle of a page - we end up writing back the first part of the page that the mapping covers, but then never revisit the page to have the remainder mapped and written. The fix is simple - if the mapping boundary falls inside a page, then simple abort clustering without touching the page. This means that the next ->writepage entry that write_cache_pages() will make is the page we aborted on, and xfs_vm_writepage() will map all sections of the page correctly. This behaviour is also optimal for non-data integrity writes, as it results in contiguous sequential writeback of the file rather than missing small holes and having to write them a "random" writes in a future pass. With this fix, all the fsx tests in xfstests now pass on a 512 byte block size filesystem on a 4k page machine. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Ben Myers <bpm@sgi.com> (cherry picked from commit 49b137cbbcc836ef231866c137d24f42c42bb483)
Diffstat (limited to 'fs/xfs/xfs_aops.c')
-rw-r--r--fs/xfs/xfs_aops.c19
1 files changed, 19 insertions, 0 deletions
diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 2b2691b73428..41a695048be7 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -725,6 +725,25 @@ xfs_convert_page(
(xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
i_size_read(inode));
+ /*
+ * If the current map does not span the entire page we are about to try
+ * to write, then give up. The only way we can write a page that spans
+ * multiple mappings in a single writeback iteration is via the
+ * xfs_vm_writepage() function. Data integrity writeback requires the
+ * entire page to be written in a single attempt, otherwise the part of
+ * the page we don't write here doesn't get written as part of the data
+ * integrity sync.
+ *
+ * For normal writeback, we also don't attempt to write partial pages
+ * here as it simply means that write_cache_pages() will see it under
+ * writeback and ignore the page until some point in the future, at
+ * which time this will be the only page in the file that needs
+ * writeback. Hence for more optimal IO patterns, we should always
+ * avoid partial page writeback due to multiple mappings on a page here.
+ */
+ if (!xfs_imap_valid(inode, imap, end_offset))
+ goto fail_unlock_page;
+
len = 1 << inode->i_blkbits;
p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
PAGE_CACHE_SIZE);