Re: [PATCH RESEND v6 1/8] blk-throttle: fix that io throttle can only work for single bio

[Yu Kuai]

在 2022/07/30 2:04, Tejun Heo 写道:
> Hello,
>
> On Fri, Jul 29, 2022 at 02:32:36PM +0800, Yu Kuai wrote:
> > We need to make sure following conditions is always hold:
> >
> > 1) If a bio is splited, iops limits should count multiple times, while
> > bps limits should only count once.
> > 2) If a bio is issued while some bios are already throttled, bps limits
> > should not be ignored.
> >
> > commit 9f5ede3c01f9 ("block: throttle split bio in case of iops limit")
> > fixes that 1) is not hold, while it breaks 2). Root cause is that such
> > bio will be flaged in __blk_throtl_bio(), and later
> > tg_with_in_bps_limit() will skip flaged bio.
> >
> > In order to fix this problem, at first, I change that flaged bio won't
> > be skipped in tg_with_in_bps_limit():
> >
> > -	if (!bio_flagged(bio, BIO_THROTTLED)) {
> > -		tg->bytes_disp[rw] += bio_size;
> > -		tg->last_bytes_disp[rw] += bio_size;
> > -	}
> > -
> > +	tg->bytes_disp[rw] += bio_size;
> > +	tg->last_bytes_disp[rw] += bio_size;
> >
> > However, this will break that bps limits should only count once. Thus I
> > try to restore the overaccounting in __blk_throtl_bio() in such case:
> >
> > +		if (bio_flagged(bio, BIO_THROTTLED)) {
> > +			unsigned int bio_size = throtl_bio_data_size(bio);
> > +
> > +			if (tg->bytes_disp[rw] >= bio_size)
> > +				tg->bytes_disp[rw] -= bio_size;
> > +			if (tg->last_bytes_disp[rw] >= bio_size)
> > +				tg->last_bytes_disp[rw] -= bio_size;
> > +		}
> >
> > If new slice is not started, then the decrement should make sure this
> > bio won't be counted again. However, if new slice is started and the
> > condition 'bytes_disp >= bio_size' doesn't hold, this bio will end up
> > accounting twice.
> >
> > Pleas let me know if you think this suituation is problematic, I'll try
> > to figure out a new way...
>
> While a bit tricky, I think it's fine but please add comments in the code
> explaining what's going on and why. Also, can you please explain why
> __blk_throtl_bio() being skipped when iops limit is not set doesn't skew the
> result?

Because bps limit is already counted the first time __blk_throtl_bio()
is called for the orignal bio. When splited bio is reentered, we only
need to throttle it again if iops limit is set.

By the way, I found that this way is better after patch 4:

in __blk_throtl_bio():

if (bio_flagged(bio, BIO_THROTTLED)) {
	tg->bytes_skipped[rw] += bio_size;
	if (tg->last_bytes_disp[rw] >= bio_size)
		tg->last_bytes_disp[rw] -= bio_size;
}

The overaccounting can be restored even if new slice is started.

Thanks,
Kuai