On Thu 12-05-16 12:45:22, Ross Zwisler wrote:

On Wed, May 11, 2016 at 11:58:49AM +0200, Jan Kara wrote:

quoted

Currently ext2 zeroes any data blocks allocated for DAX inode however it
still returns them as BH_New. Thus DAX code zeroes them again in
dax_insert_mapping() which can possibly overwrite the data that has been
already stored to those blocks by a racing dax_io(). Avoid marking
pre-zeroed buffers as new.

Reviewed-by: Ross Zwisler <redacted>
Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/ext2/inode.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/ext2/inode.c b/fs/ext2/inode.c
index 6bd58e6ff038..1f07b758b968 100644
--- a/fs/ext2/inode.c
+++ b/fs/ext2/inode.c

@@ -745,11 +745,11 @@ static int ext2_get_blocks(struct inode *inode,
 			mutex_unlock(&ei->truncate_mutex);
 			goto cleanup;
 		}
-	}
+	} else
+		set_buffer_new(bh_result);
 
 	ext2_splice_branch(inode, iblock, partial, indirect_blks, count);
 	mutex_unlock(&ei->truncate_mutex);
-	set_buffer_new(bh_result);
 got_it:
 	map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
 	if (count > blocks_to_boundary)

-- 
2.6.6

Interestingly this change is causing a bunch of xfstests regressions for me
with ext2 + DAX.  All of these tests pass without this one change.

Good catch. Attached patch fixes this issue for me. Preferably it should be
merged before the above ext2 change.

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR