Btrfs: fix direct IO requests not reporting IO error to user space
PATCH 1/1 Btrfs: fix direct IO requests not reporting IO error to user space
From: Filipe Manana <[email protected]>
If a bio for a direct IO request fails, we were not setting the error in the parent bio (the main DIO bio), making us not return the error to user space in btrfs_direct_IO(), that is, it made __blockdev_direct_IO() return the number of bytes issued for IO and not the error a bio created and submitted by btrfs_submit_direct() got from the block layer. This essentially happens because when we call:
dio_end_io(dio_bio, bio->bi_error);
It does not set dio_bio->bi_error to the value of the second argument. So just add this missing assignment in endio callbacks, just as we do in the error path at btrfs_submit_direct() when we fail to clone the dio bio or allocate its private object.
This was detected by the new generic test cases in xfstests: 271, 272, 276 and 278. Which essentially setup a dm error target, then load the error table, do a direct IO write and unload the error table. They expect the write to fail with -EIO, which was not getting reported when testing against btrfs.
--- fs/btrfs/inode.c | 2 ++ 1 file changed, 2 insertions(+)
fs/btrfs/inode.c +2 -0
@@ -7985,6 +7985,7 @@ static void btrfs_endio_direct_read(struct bio *bio)kfree (dip); + dio_bio->bi_error = bio->bi_error;dio_end_io (dio_bio, bio->bi_error);if (io_bio->end_io)
@@ -8039,6 +8040,7 @@ static void btrfs_endio_direct_write(struct bio *bio)kfree (dip); + dio_bio->bi_error = bio->bi_error;dio_end_io (dio_bio, bio->bi_error);bio_put (bio); }
Signature · 1 line
2.7.0.rc3
PATCH v2 1/1 Btrfs: fix direct IO requests not reporting IO error to user space
From: Filipe Manana <[email protected]>
If a bio for a direct IO request fails, we were not setting the error in the parent bio (the main DIO bio), making us not return the error to user space in btrfs_direct_IO(), that is, it made __blockdev_direct_IO() return the number of bytes issued for IO and not the error a bio created and submitted by btrfs_submit_direct() got from the block layer. This essentially happens because when we call:
dio_end_io(dio_bio, bio->bi_error);
It does not set dio_bio->bi_error to the value of the second argument. So just add this missing assignment in endio callbacks, just as we do in the error path at btrfs_submit_direct() when we fail to clone the dio bio or allocate its private object. This follows the convention of what is done with other similar APIs such as bio_endio() where the caller is responsible for setting the bi_error field in the bio it passes as an argument to bio_endio().
This was detected by the new generic test cases in xfstests: 271, 272, 276 and 278. Which essentially setup a dm error target, then load the error table, do a direct IO write and unload the error table. They expect the write to fail with -EIO, which was not getting reported when testing against btrfs.
---
V2: Updated commit message to reflect affected kernel versions.
fs/btrfs/inode.c | 2 ++ 1 file changed, 2 insertions(+)
fs/btrfs/inode.c +2 -0
@@ -7985,6 +7985,7 @@ static void btrfs_endio_direct_read(struct bio *bio)kfree (dip); + dio_bio->bi_error = bio->bi_error;dio_end_io (dio_bio, bio->bi_error);if (io_bio->end_io)
@@ -8039,6 +8040,7 @@ static void btrfs_endio_direct_write(struct bio *bio)kfree (dip); + dio_bio->bi_error = bio->bi_error;dio_end_io (dio_bio, bio->bi_error);bio_put (bio); }
Signature · 1 line
2.7.0.rc3