GitLab

We noticed that we were having regular CG OOM kills in cases where there
was still enough dirty pages to avoid OOM'ing.  It turned out there's
this corner case in btrfs's handling of range_cyclic where files that
were being redirtied were not getting fully written out because of how
we do range_cyclic writeback.

We unconditionally were setting scanned = 1; the first time we found any
pages in the inode.  This isn't actually what we want, we want it to be
set if we've scanned the entire file.  For range_cyclic we could be
starting in the middle or towards the end of the file, so we could write
one page and then not write any of the other dirty pages in the file
because we set scanned = 1.

Fix this by not setting scanned = 1 if we find pages.  The rules for
setting scanned should be

1) !range_cyclic.  In this case we have a specified range to write out.
2) range_cyclic && index == 0.  In this case we've started at the
   beginning and there is no need to loop around a second time.
3) range_cyclic && we started at index > 0 and we've reached the end of
   the file without satisfying our nr_to_write.

This patch fixes both of our writepages implementations to make sure
these rules hold true.  This fixed our over zealous CG OOMs in
production.

Fixes: d1310b2e

 ("Btrfs: Split the extent_map code into two parts")
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add comment ]
Signed-off-by: David Sterba <dsterba@suse.com>