loop: fix min directio size detection for nested loop devices
Subsystems (2)
PATCH 1/1 loop: fix min directio size detection for nested loop devices
From: Darrick J. Wong <[email protected]>
fstest generic/563 sets up a loop device in front of the SCRATCH_DEV block device to test cgroup I/O accounting, because SCRATCH_DEV could be a partition and cgroup accounting apparently only works on full block devices (e.g. sda, not sda1).
If however SCRATCH_DEV is itself a loop device that we've used to simulate 4k LBA block devices, the minimum directio size discovery introduced in commit f4774e92aab85d is wrong -- we should query the logical block size of the underlying block device because file->f_path points whatever filesystem /dev is.
Otherwise, you get a weird losetup config:
$ losetup -f /dev/sda $ losetup --sector-size 4096 /dev/loop0 $ losetup -f /dev/loop0 $ losetup --raw NAME SIZELIMIT OFFSET AUTOCLEAR RO BACK-FILE DIO LOG-SEC /dev/loop0 0 0 0 0 /dev/sda 1 4096 /dev/loop1 0 0 0 0 /dev/loop0 1 512
(Note loop1 can try to send 512b writes to loop0 which has a sector size of 4k)
and mkfs failures like this:
error reading existing superblock: Invalid argument mkfs.xfs: pwrite failed: Invalid argumentlibxfs_bwrite: write failed on (unknown) bno 0x42a3ef8/0x100, err=22 mkfs.xfs: Releasing dirty buffer to free list!
found dirty buffer (bulk) on free list!
mkfs.xfs: pwrite failed: Invalid argumentlibxfs_bwrite: write failed on (unknown) bno 0x0/0x100, err=22
--- drivers/block/loop.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/block/loop.c +9 -1
@@ -443,9 +443,17 @@ static void loop_reread_partitions(struct loop_device *lo)static unsigned int loop_query_min_dio_size (struct loop_device *lo) {struct file *file = lo->lo_backing_file; -struct block_device *sb_bdev = file->f_mapping->host->i_sb->s_bdev; +struct inode *inode = file->f_mapping->host; +struct block_device *sb_bdev = inode->i_sb->s_bdev;struct kstat st; +/* +* If the backing device is a block device, don't send directios +* smaller than its LBA size. +*/ +if (S_ISBLK (inode->i_mode)) +return bdev_logical_block_size (I_BDEV (inode)); +/* * Use the minimal dio alignment of the file system if provided. */
Quoted 5 lines · Darrick J. Wong
If however
SCRATCH_DEVis itself a loop device that we've used to simulate 4k LBA block devices, the minimum directio size discovery introduced in commitf4774e92aab85dis wrong -- we should query the logical block size of the underlying block device becausefile->f_pathpoints whatever filesystem /dev is.
No, the problem is that special handling of block devices in stat sits in vfs_statx_path and not the exported vfs_getattr helper, and thus we don't call into bdev_statx which would return the right value. I'll send a separate VFS-level fix for this.