blk-throttle: Fix io statistics for cgroup v1

Commit Message

hanjinke March 28, 2023, 2:23 p.m. UTC

From: Jinke Han <[e-mail]>

Now the io statistics of cgroup v1 are no longer accurate. Although
in the long run it's best that rstat is a good implementation of
cgroup v1 io statistics. But before that, we'd better fix this issue.

Signed-off-by: Jinke Han <[e-mail]>
---
 block/blk-cgroup.c   | 3 +++
 block/blk-throttle.c | 6 ------
 block/blk-throttle.h | 9 +++++++++
 3 files changed, 12 insertions(+), 6 deletions(-)

Comments

Tejun Heo March 29, 2023, 6:54 p.m. UTC | #1

On Tue, Mar 28, 2023 at 10:23:09PM +0800, Jinke Han wrote:
> From: Jinke Han <[e-mail]>
> 
> Now the io statistics of cgroup v1 are no longer accurate. Although
> in the long run it's best that rstat is a good implementation of
> cgroup v1 io statistics. But before that, we'd better fix this issue.

Can you please expand on how the stats are wrong on v1 and how the patch
fixes it?

Thanks.

hanjinke March 30, 2023, 3:44 a.m. UTC | #2

在 2023/3/30 上午2:54, Tejun Heo 写道:
> On Tue, Mar 28, 2023 at 10:23:09PM +0800, Jinke Han wrote:
>> From: Jinke Han <[e-mail]>
>>
>> Now the io statistics of cgroup v1 are no longer accurate. Although
>> in the long run it's best that rstat is a good implementation of
>> cgroup v1 io statistics. But before that, we'd better fix this issue.
> 
> Can you please expand on how the stats are wrong on v1 and how the patch
> fixes it?
> 
> Thanks.
> 
Now blkio.throttle.io_serviced and blkio.throttle.io_serviced become the 
only stable io stats interface of cgroup v1, and these statistics are 
done in the blk-throttle code. But the current code only counts the bios 
that are actually throttled. When the user does not add the throttle 
limit, the io stats for cgroup v1 has nothing. I fix it according to the 
statistical method of v2, and made it count all ios accurately.

Thanks.

Tejun Heo March 31, 2023, 1:44 a.m. UTC | #3

Hello,

On Thu, Mar 30, 2023 at 11:44:04AM +0800, hanjinke wrote:
> 在 2023/3/30 上午2:54, Tejun Heo 写道:
> > On Tue, Mar 28, 2023 at 10:23:09PM +0800, Jinke Han wrote:
> > > From: Jinke Han <[e-mail]>
> > > 
> > > Now the io statistics of cgroup v1 are no longer accurate. Although
> > > in the long run it's best that rstat is a good implementation of
> > > cgroup v1 io statistics. But before that, we'd better fix this issue.
> > 
> > Can you please expand on how the stats are wrong on v1 and how the patch
> > fixes it?
> > 
> > Thanks.
> > 
> Now blkio.throttle.io_serviced and blkio.throttle.io_serviced become the

"now" might be a bit too vague. Can you point to the commit which made the
change?

> only stable io stats interface of cgroup v1, and these statistics are done
> in the blk-throttle code. But the current code only counts the bios that are

Ah, okay, so the stats are now updated by blk-throtl itself but

> actually throttled. When the user does not add the throttle limit, the io
> stats for cgroup v1 has nothing. I fix it according to the statistical
> method of v2, and made it count all ios accurately.

updated only when limits are configured which can be confusing. Makes sense
to me. Can you please update the patch description accordingly?

Also, the following change:

@@ -2033,6 +2033,9 @@ void blk_cgroup_bio_start(struct bio *bio)
        struct blkg_iostat_set *bis;
        unsigned long flags;

+       if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
+               return;
+
        /* Root-level stats are sourced from system-wide IO stats */
        if (!cgroup_parent(blkcg->css.cgroup))
                return;

seems incomplete as there's an additional
cgroup_subsys_on_dfl(io_cgrp_subsys) test in the function. We probably wanna
remove that?

Thanks.

hanjinke March 31, 2023, 3 a.m. UTC | #4

在 2023/3/31 上午9:44, Tejun Heo 写道:
> Hello,
> 
> On Thu, Mar 30, 2023 at 11:44:04AM +0800, hanjinke wrote:
>> 在 2023/3/30 上午2:54, Tejun Heo 写道:
>>> On Tue, Mar 28, 2023 at 10:23:09PM +0800, Jinke Han wrote:
>>>> From: Jinke Han <[e-mail]>
>>>>
>>>> Now the io statistics of cgroup v1 are no longer accurate. Although
>>>> in the long run it's best that rstat is a good implementation of
>>>> cgroup v1 io statistics. But before that, we'd better fix this issue.
>>>
>>> Can you please expand on how the stats are wrong on v1 and how the patch
>>> fixes it?
>>>
>>> Thanks.
>>>
>> Now blkio.throttle.io_serviced and blkio.throttle.io_serviced become the
> 
> "now" might be a bit too vague. Can you point to the commit which made the
> change?
> 
>> only stable io stats interface of cgroup v1, and these statistics are done
>> in the blk-throttle code. But the current code only counts the bios that are
> 
> Ah, okay, so the stats are now updated by blk-throtl itself but
> 
>> actually throttled. When the user does not add the throttle limit, the io
>> stats for cgroup v1 has nothing. I fix it according to the statistical
>> method of v2, and made it count all ios accurately.
> 
> updated only when limits are configured which can be confusing. Makes sense
> to me. Can you please update the patch description accordingly?
> 
> Also, the following change:
> 
> @@ -2033,6 +2033,9 @@ void blk_cgroup_bio_start(struct bio *bio)
>          struct blkg_iostat_set *bis;
>          unsigned long flags;
> 
> +       if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
> +               return;
> +
>          /* Root-level stats are sourced from system-wide IO stats */
>          if (!cgroup_parent(blkcg->css.cgroup))
>                  return;
> 
> seems incomplete as there's an additional
> cgroup_subsys_on_dfl(io_cgrp_subsys) test in the function. We probably wanna
> remove that?
> 
> Thanks.
> 

okay, according to your suggestion, I will send a v2.

Thanks

Patch

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index bd50b55bdb61..677e4473e45e 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -2033,6 +2033,9 @@  void blk_cgroup_bio_start(struct bio *bio)
 	struct blkg_iostat_set *bis;
 	unsigned long flags;
 
+	if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
+		return;
+
 	/* Root-level stats are sourced from system-wide IO stats */
 	if (!cgroup_parent(blkcg->css.cgroup))
 		return;
diff --git a/block/blk-throttle.c b/block/blk-throttle.c
index 47e9d8be68f3..2be66e9430f7 100644
--- a/block/blk-throttle.c
+++ b/block/blk-throttle.c
@@ -2174,12 +2174,6 @@  bool __blk_throtl_bio(struct bio *bio)
 
 	rcu_read_lock();
 
-	if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
-		blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
-				bio->bi_iter.bi_size);
-		blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
-	}
-
 	spin_lock_irq(&q->queue_lock);
 
 	throtl_update_latency_buckets(td);
diff --git a/block/blk-throttle.h b/block/blk-throttle.h
index ef4b7a4de987..d1ccbfe9f797 100644
--- a/block/blk-throttle.h
+++ b/block/blk-throttle.h
@@ -185,6 +185,15 @@  static inline bool blk_should_throtl(struct bio *bio)
 	struct throtl_grp *tg = blkg_to_tg(bio->bi_blkg);
 	int rw = bio_data_dir(bio);
 
+	if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
+		if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
+			bio_set_flag(bio, BIO_CGROUP_ACCT);
+			blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
+					bio->bi_iter.bi_size);
+		}
+		blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
+	}
+
 	/* iops limit is always counted */
 	if (tg->has_rules_iops[rw])
 		return true;