2019-06-07 13:14:48

by Naohiro Aota

Subject: [PATCH 13/19] btrfs: avoid sync IO prioritization on checksum in HMZONED mode

Btrfs prioritize sync I/Os to be handled by async checksum worker earlier.
As a result, checksumming sync I/Os to larger logical extent address can
finish faster than checksumming non-sync I/Os to smaller logical extent
address.

Since we have upper limit of number of checksum worker, it is possible that
sync I/Os to wait forever for non-starting checksum of I/Os for smaller
address.

This situation can be reproduced by e.g. fstests btrfs/073.

To avoid such disordering, disable sync IO prioritization for now. Note
that sync I/Os anyway must wait for I/Os to smaller address to finish. So,
actually prioritization have no benefit in HMZONED mode.

Signed-off-by: Naohiro Aota <[email protected]>
---
fs/btrfs/disk-io.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 56a416902ce7..6651986da470 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -838,7 +838,7 @@ blk_status_t btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct bio *bio,

async->status = 0;

- if (op_is_sync(bio->bi_opf))
+ if (op_is_sync(bio->bi_opf) && !btrfs_fs_incompat(fs_info, HMZONED))
btrfs_set_work_high_priority(&async->work);

btrfs_queue_work(fs_info->workers, &async->work);
--
2.21.0

2019-06-13 15:07:16

by Josef Bacik

Subject: Re: [PATCH 13/19] btrfs: avoid sync IO prioritization on checksum in HMZONED mode

On Fri, Jun 07, 2019 at 10:10:19PM +0900, Naohiro Aota wrote:
> Btrfs prioritize sync I/Os to be handled by async checksum worker earlier.
> As a result, checksumming sync I/Os to larger logical extent address can
> finish faster than checksumming non-sync I/Os to smaller logical extent
> address.
>
> Since we have upper limit of number of checksum worker, it is possible that
> sync I/Os to wait forever for non-starting checksum of I/Os for smaller
> address.
>
> This situation can be reproduced by e.g. fstests btrfs/073.
>
> To avoid such disordering, disable sync IO prioritization for now. Note
> that sync I/Os anyway must wait for I/Os to smaller address to finish. So,
> actually prioritization have no benefit in HMZONED mode.
>

This stuff is going away once we finish the io.weight work anyway, I wouldn't
worry about this. Thanks,

Josef