This patch fixes a race between the shutdown path and bio completion
handling. In the ext4 direct io path with async io, after submitting a
bio to the block layer, if journal starting fails,
ext4_direct_IO_write() would bail out pretending that the IO
failed. The caller would have had no way of knowing whether or not the
IO was successfully submitted. So instead, we return -EIOCBQUEUED in
this case. Now, the caller knows that the IO was submitted. The bio
completion handler takes care of the error.
Tested: Ran the shutdown xfstest test 461 in loop for over 2 hours across
4 machines resulting in over 400 runs. Verified that the race didn't
occur. Usually the race was seen in about 20-30 iterations.
Signed-off-by: Harshad Shirwadkar <[email protected]>
---
fs/ext4/inode.c | 16 ++++++++++++----
1 file changed, 12 insertions(+), 4 deletions(-)
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 534a9130f625..4c2f8b57bdc7 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -3767,10 +3767,18 @@ static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
/* Credits for sb + inode write */
handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
if (IS_ERR(handle)) {
- /* This is really bad luck. We've written the data
- * but cannot extend i_size. Bail out and pretend
- * the write failed... */
- ret = PTR_ERR(handle);
+ /*
+ * We wrote the data but cannot extend
+ * i_size. Bail out. In async io case, we do
+ * not return error here because we have
+ * already submmitted the corresponding
+ * bio. Returning error here makes the caller
+ * think that this IO is done and failed
+ * resulting in race with bio's completion
+ * handler.
+ */
+ if (!ret)
+ ret = PTR_ERR(handle);
if (inode->i_nlink)
ext4_orphan_del(NULL, inode);
--
2.16.0.rc1.238.g530d649a79-goog
On Tue, Jan 09, 2018 at 02:52:57PM -0800, harshads wrote:
> This patch fixes a race between the shutdown path and bio completion
> handling. In the ext4 direct io path with async io, after submitting a
> bio to the block layer, if journal starting fails,
> ext4_direct_IO_write() would bail out pretending that the IO
> failed. The caller would have had no way of knowing whether or not the
> IO was successfully submitted. So instead, we return -EIOCBQUEUED in
> this case. Now, the caller knows that the IO was submitted. The bio
> completion handler takes care of the error.
>
> Tested: Ran the shutdown xfstest test 461 in loop for over 2 hours across
> 4 machines resulting in over 400 runs. Verified that the race didn't
> occur. Usually the race was seen in about 20-30 iterations.
>
> Signed-off-by: Harshad Shirwadkar <[email protected]>
Thanks, applied.
- Ted