On Sat, Jan 19, 2019 at 01:05:04PM -0500, Mike Snitzer wrote:
> The risk of redundant IO accounting was not taken into consideration
> when commit 18a25da84354 ("dm: ensure bio submission follows a
> depth-first tree walk") introduced IO splitting in terms of recursion
> via generic_make_request().
>
> Fix this by subtracting the split bio's payload from the IO stats that
> were already accounted for by start_io_acct() upon dm_make_request()
> entry. This repeat oscillation of the IO accounting, up then down,
> isn't ideal but refactoring DM core's IO splitting to pre-split bios
> _before_ they are accounted turned out to be an excessive amount of
> change that will need a full development cycle to refine and verify.
>
> Before this fix:
>
> /dev/mapper/stripe_dev is a 4-way stripe using a 32k chunksize, so
> bios are split on 32k boundaries.
>
> # fio --name=16M --filename=/dev/mapper/stripe_dev --rw=write --bs=64k
> --size=16M \
> --iodepth=1 --ioengine=libaio --direct=1 --refill_buffers
>
> with debugging added:
> [103898.310264] device-mapper: core: start_io_acct: dm-2 WRITE
> bio->bi_iter.bi_sector=0 len=128
> [103898.318704] device-mapper: core: __split_and_process_bio: recursing for
> following split bio:
> [103898.329136] device-mapper: core: start_io_acct: dm-2 WRITE
> bio->bi_iter.bi_sector=64 len=64
> ...
>
> 16M written yet 136M (278528 * 512b) accounted:
> # cat /sys/block/dm-2/stat | awk '{ print $7 }'
> 278528
>
> After this fix:
>
> 16M written and 16M (32768 * 512b) accounted:
> # cat /sys/block/dm-2/stat | awk '{ print $7 }'
> 32768
>
> Fixes: 18a25da84354 ("dm: ensure bio submission follows a depth-first tree
> walk")
> Cc: [email protected] # 4.16+
> Reported-by: Bryan Gurney <[email protected]>
> Signed-off-by: Mike Snitzer <[email protected]>
> ---
> drivers/md/dm.c | 16 ++++++++++++++++
> 1 file changed, 16 insertions(+)
>
> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> index fcb97b0a5743..fbadda68e23b 100644
> --- a/drivers/md/dm.c
> +++ b/drivers/md/dm.c
> @@ -1584,6 +1584,9 @@ static void init_clone_info(struct clone_info *ci,
> struct mapped_device *md,
> ci->sector = bio->bi_iter.bi_sector;
> }
>
> +#define __dm_part_stat_sub(part, field, subnd) \
> + (part_stat_get(part, field) -= (subnd))
> +
> /*
> * Entry point to split a bio into clones and submit them to the targets.
> */
> @@ -1638,6 +1641,19 @@ static blk_qc_t __split_and_process_bio(struct
> mapped_device *md,
> struct bio *b = bio_split(bio, bio_sectors(bio)
> - ci.sector_count,
> GFP_NOIO,
> &md->queue->bio_split);
> ci.io->orig_bio = b;
> +
> + /*
> + * Adjust IO stats for each split, otherwise
> upon queue
> + * reentry there will be redundant IO
> accounting.
> + * NOTE: this is a stop-gap fix, a proper fix
> involves
> + * significant refactoring of DM core's bio
> splitting
> + * (by eliminating DM's splitting and just
> using bio_split)
> + */
> + part_stat_lock();
> + __dm_part_stat_sub(&dm_disk(md)->part0,
> +
> sectors[op_stat_group(bio_op(bio))], ci.sector_count);
> + part_stat_unlock();
> +
> bio_chain(b, bio);
> ret = generic_make_request(bio);
> break;
This ways is a bit ugly, but looks it works and it is simple, especially
DM target may accept partial bio, so:
Reviewed-by: Ming Lei <[email protected]>
Thanks,
Ming
