From mboxrd@z Thu Jan 1 00:00:00 1970 From: "yukuai (C)" Subject: Re: [PATCH -next] blk-throttle: delay the setting of 'BIO_THROTTLED' to when throttle is done Date: Wed, 18 May 2022 14:30:07 +0800 Message-ID: <73e8bdef-0f80-d255-e4f2-170813a12f5f@huawei.com> References: <20220517134909.2910251-1-yukuai3@huawei.com> Mime-Version: 1.0 Content-Transfer-Encoding: quoted-printable Return-path: In-Reply-To: List-ID: Content-Type: text/plain; charset="iso-8859-1"; format="flowed" To: Ming Lei Cc: tj@kernel.org, axboe@kernel.dk, cgroups@vger.kernel.org, linux-block@vger.kernel.org, linux-kernel@vger.kernel.org, yi.zhang@huawei.com =D4=DA 2022/05/18 12:07, Ming Lei =D0=B4=B5=C0: > On Tue, May 17, 2022 at 09:49:09PM +0800, Yu Kuai wrote: >> commit 9f5ede3c01f9 ("block: throttle split bio in case of iops limit") >> introduce a new problem, for example: >> >> [root@localhost ~]# echo "8:0 1024" > /sys/fs/cgroup/blkio/blkio.throttl= e.write_bps_device >> [root@localhost ~]# echo $$ > /sys/fs/cgroup/blkio/cgroup.procs >> [root@localhost ~]# dd if=3D/dev/zero of=3D/dev/sda bs=3D10k count=3D1 o= flag=3Ddirect & >> [1] 620 >> [root@localhost ~]# dd if=3D/dev/zero of=3D/dev/sda bs=3D10k count=3D1 o= flag=3Ddirect & >> [2] 626 >> [root@localhost ~]# 1+0 records in >> 1+0 records out >> 10240 bytes (10 kB, 10 KiB) copied, 10.0038 s, 1.0 kB/s1+0 records in >> 1+0 records out >> >> 10240 bytes (10 kB, 10 KiB) copied, 9.23076 s, 1.1 kB/s >> -> the second bio is issued after 10s instead of 20s. >> >> This is because if some bios are already queued, current bio is queued >> directly and the flag 'BIO_THROTTLED' is set. And later, when former >> bios are dispatched, this bio will be dispatched without waiting at all, >> this is due to tg_with_in_bps_limit() will return 0 if the flag is set. >> >> Instead of setting the flag when bio starts throttle, delay to when >> throttle is done to fix the problem. >> >> Fixes: 9f5ede3c01f9 ("block: throttle split bio in case of iops limit") >> Signed-off-by: Yu Kuai >> --- >> block/blk-throttle.c | 10 ++++++---- >> 1 file changed, 6 insertions(+), 4 deletions(-) >> >> diff --git a/block/blk-throttle.c b/block/blk-throttle.c >> index 447e1b8722f7..f952f2d942ff 100644 >> --- a/block/blk-throttle.c >> +++ b/block/blk-throttle.c >> @@ -811,7 +811,7 @@ static bool tg_with_in_bps_limit(struct throtl_grp *= tg, struct bio *bio, >> unsigned int bio_size =3D throtl_bio_data_size(bio); >> =20 >> /* no need to throttle if this bio's bytes have been accounted */ >> - if (bps_limit =3D=3D U64_MAX || bio_flagged(bio, BIO_THROTTLED)) { >> + if (bps_limit =3D=3D U64_MAX) { >=20 > This way may double account bio size for re-entered split bio. Hi, Ming Yes, you are right, I forgot that... >=20 >=20 >> if (wait) >> *wait =3D 0; >> return true; >> @@ -1226,8 +1226,10 @@ static void blk_throtl_dispatch_work_fn(struct wo= rk_struct *work) >> =20 >> spin_lock_irq(&q->queue_lock); >> for (rw =3D READ; rw <=3D WRITE; rw++) >> - while ((bio =3D throtl_pop_queued(&td_sq->queued[rw], NULL))) >> + while ((bio =3D throtl_pop_queued(&td_sq->queued[rw], NULL))) { >> + bio_set_flag(bio, BIO_THROTTLED); >> bio_list_add(&bio_list_on_stack, bio); >> + } >> spin_unlock_irq(&q->queue_lock); >> =20 >> if (!bio_list_empty(&bio_list_on_stack)) { >> @@ -2134,7 +2136,8 @@ bool __blk_throtl_bio(struct bio *bio) >> } >> break; >> } >> - >> + /* this bio will be issued directly */ >> + bio_set_flag(bio, BIO_THROTTLED); >> /* within limits, let's charge and dispatch directly */ >> throtl_charge_bio(tg, bio); >=20 > Marking BIO_THROTTLED before throtle_charge_bio() causes the bio > bytes not be charged. Yes, thanks for spotting this. >=20 > Another simple way is to compensate for previous extra bytes accounting, > something like the following patch: >=20 >=20 > diff --git a/block/blk-throttle.c b/block/blk-throttle.c > index 139b2d7a99e2..44773d2ba257 100644 > --- a/block/blk-throttle.c > +++ b/block/blk-throttle.c > @@ -810,8 +810,7 @@ static bool tg_with_in_bps_limit(struct throtl_grp *t= g, struct bio *bio, > unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; > unsigned int bio_size =3D throtl_bio_data_size(bio); > =20 > - /* no need to throttle if this bio's bytes have been accounted */ > - if (bps_limit =3D=3D U64_MAX || bio_flagged(bio, BIO_THROTTLED)) { > + if (bps_limit =3D=3D U64_MAX) { > if (wait) > *wait =3D 0; > return true; > @@ -921,10 +920,8 @@ static void throtl_charge_bio(struct throtl_grp *tg,= struct bio *bio) > unsigned int bio_size =3D throtl_bio_data_size(bio); > =20 > /* Charge the bio to the group */ > - if (!bio_flagged(bio, BIO_THROTTLED)) { > - tg->bytes_disp[rw] +=3D bio_size; > - tg->last_bytes_disp[rw] +=3D bio_size; > - } > + tg->bytes_disp[rw] +=3D bio_size; > + tg->last_bytes_disp[rw] +=3D bio_size; > =20 > tg->io_disp[rw]++; > tg->last_io_disp[rw]++; > @@ -2125,6 +2122,20 @@ bool __blk_throtl_bio(struct bio *bio) > if (sq->nr_queued[rw]) > break; > =20 > + /* > + * re-entered bio has accounted bytes already, so try to > + * compensate previous over-accounting. However, if new > + * slice is started, just forget it > + */ > + if (bio_flagged(bio, BIO_THROTTLED)) { > + unsigned int bio_size =3D throtl_bio_data_size(bio); > + > + if (tg->bytes_disp[rw] >=3D bio_size) > + tg->bytes_disp[rw] -=3D bio_size; > + if (tg->last_bytes_disp[rw] - bio_size) > + tg->last_bytes_disp[rw] -=3D bio_size; > + } > + If new slice is not started, then this should fix the problem. From what I can see, other than tg_conf_updated, new silce can start if there are still bio throttled: tg_may_dispatch if (!(tg->service_queue.nr_queued[rw])) throtl_start_new_slice Thus I think the change is ok. For the case in tg_conf_updated, I'll remove the throtl_start_new_slice() to fix a hung problem. I'll add this patch with this one in next version. Thanks, Kuai > /* if above limits, break to queue */ > if (!tg_may_dispatch(tg, bio, NULL)) { > tg->last_low_overflow_time[rw] =3D jiffies; >=20 > Thanks, > Ming >=20 > . >=20