by Zhang Qiao

[permalink] [raw]

Subject: Re: [PATCH v3] sched/fair: sanitize vruntime of entity being placed

在 2023/3/2 22:55, Vincent Guittot 写道:
> On Thu, 2 Mar 2023 at 15:29, Zhang Qiao <[email protected]> wrote:
>>
>>
>>
>> 在 2023/3/2 21:34, Vincent Guittot 写道:
>>> On Thu, 2 Mar 2023 at 10:36, Zhang Qiao <[email protected]> wrote:
>>>>
>>>>
>>>>
>>>> 在 2023/2/27 22:37, Vincent Guittot 写道:
>>>>> On Mon, 27 Feb 2023 at 09:43, Roman Kagan <[email protected]> wrote:
>>>>>>
>>>>>> On Tue, Feb 21, 2023 at 06:26:11PM +0100, Vincent Guittot wrote:
>>>>>>> On Tue, 21 Feb 2023 at 17:57, Roman Kagan <[email protected]> wrote:
>>>>>>>> What scares me, though, is that I've got a message from the test robot
>>>>>>>> that this commit drammatically affected hackbench results, see the quote
>>>>>>>> below. I expected the commit not to affect any benchmarks.
>>>>>>>>
>>>>>>>> Any idea what could have caused this change?
>>>>>>>
>>>>>>> Hmm, It's most probably because se->exec_start is reset after a
>>>>>>> migration and the condition becomes true for newly migrated task
>>>>>>> whereas its vruntime should be after min_vruntime.
>>>>>>>
>>>>>>> We have missed this condition
>>>>>>
>>>>>> Makes sense to me.
>>>>>>
>>>>>> But what would then be the reliable way to detect a sched_entity which
>>>>>> has slept for long and risks overflowing in .vruntime comparison?
>>>>>
>>>>> For now I don't have a better idea than adding the same check in
>>>>> migrate_task_rq_fair()
>>>>
>>>> Hi, Vincent，
>>>> I fixed this condition as you said, and the test results are as follows.
>>>>
>>>> testcase: hackbench -g 44 -f 20 --process --pipe -l 60000 -s 100
>>>> version1: v6.2
>>>> version2: v6.2 + commit 829c1651e9c4
>>>> version3: v6.2 + commit 829c1651e9c4 + this patch
>>>>
>>>> -------------------------------------------------
>>>> version1 version2 version3
>>>> test1 81.0 118.1 82.1
>>>> test2 82.1 116.9 80.3
>>>> test3 83.2 103.9 83.3
>>>> avg(s) 82.1 113.0 81.9
>>>>
>>>> -------------------------------------------------
>>>> After deal with the task migration case, the hackbench result has restored.
>>>>
>>>> The patch as follow, how does this look?
>>>>
>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>> index ff4dbbae3b10..3a88d20fd29e 100644
>>>> --- a/kernel/sched/fair.c
>>>> +++ b/kernel/sched/fair.c
>>>> @@ -4648,6 +4648,26 @@ static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
>>>> #endif
>>>> }
>>>>
>>>> +static inline u64 sched_sleeper_credit(struct sched_entity *se)
>>>> +{
>>>> +
>>>> + unsigned long thresh;
>>>> +
>>>> + if (se_is_idle(se))
>>>> + thresh = sysctl_sched_min_granularity;
>>>> + else
>>>> + thresh = sysctl_sched_latency;
>>>> +
>>>> + /*
>>>> + * Halve their sleep time's effect, to allow
>>>> + * for a gentler effect of sleepers:
>>>> + */
>>>> + if (sched_feat(GENTLE_FAIR_SLEEPERS))
>>>> + thresh >>= 1;
>>>> +
>>>> + return thresh;
>>>> +}
>>>> +
>>>> static void
>>>> place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
>>>> {
>>>> @@ -4664,23 +4684,8 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
>>>> vruntime += sched_vslice(cfs_rq, se);
>>>>
>>>> /* sleeps up to a single latency don't count. */
>>>> - if (!initial) {
>>>> - unsigned long thresh;
>>>> -
>>>> - if (se_is_idle(se))
>>>> - thresh = sysctl_sched_min_granularity;
>>>> - else
>>>> - thresh = sysctl_sched_latency;
>>>> -
>>>> - /*
>>>> - * Halve their sleep time's effect, to allow
>>>> - * for a gentler effect of sleepers:
>>>> - */
>>>> - if (sched_feat(GENTLE_FAIR_SLEEPERS))
>>>> - thresh >>= 1;
>>>> -
>>>> - vruntime -= thresh;
>>>> - }
>>>> + if (!initial)
>>>> + vruntime -= sched_sleeper_credit(se);
>>>>
>>>> /*
>>>> * Pull vruntime of the entity being placed to the base level of
>>>> @@ -4690,7 +4695,7 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
>>>> * inversed due to s64 overflow.
>>>> */
>>>> sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
>>>> - if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
>>>> + if (se->exec_start != 0 && (s64)sleep_time > 60LL * NSEC_PER_SEC)
>>>> se->vruntime = vruntime;
>>>> else
>>>> se->vruntime = max_vruntime(se->vruntime, vruntime);
>>>> @@ -7634,8 +7639,12 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
>>>> */
>>>> if (READ_ONCE(p->__state) == TASK_WAKING) {
>>>> struct cfs_rq *cfs_rq = cfs_rq_of(se);
>>>> + u64 sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
>>>>
>>>> - se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
>>>> + if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
>>>
>>> You also need to test (se->exec_start !=0) here because the task might
>>
>> Hi,
>>
>> I don't understand when the another migration happend. Could you tell me in more detail?
>
> se->exec_start is update when the task becomes current.
>
> You can have the sequence:
>
> task TA runs on CPU0
> TA's se->exec_start = xxxx
> TA is put back into the rb tree waiting for next slice while another
> task is running
> CPU1 pulls TA which migrates on CPU1
> migrate_task_rq_fair() w/ TA's se->exec_start == xxxx
> TA's se->exec_start = 0
> TA is put into the rb tree of CPU1 waiting to run on CPU1
> CPU2 pulls TA which migrates on CPU2
> migrate_task_rq_fair() w/ TA's se->exec_start == 0
> TA's se->exec_start = 0
Hi, Vincent,

yes, you're right, such sequence does exist. But at this point, p->__state != TASK_WAKING.

I have a question, Whether there is case that is "p->se.exec_start == 0 && p->__state == TASK_WAKING" ?
I analyzed the code and concluded that this case isn't existed, is it right?

Thanks.
ZhangQiao.

>
>>
>> I think the next migration will happend after the wakee task enqueued, but at this time
>> the p->__state isn't TASK_WAKING, p->__state already be changed to TASK_RUNNING at ttwu_do_wakeup().
>>
>> If such a migration exists, Previous code "se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);" maybe
>> perform multiple times，wouldn't it go wrong in this way?
>
> the vruntime have been updated when enqueued but not exec_start
>
>>
>>> migrate another time before being scheduled. You should create a
>>> helper function like below and use it in both place
>>
>> Ok, I will update at next version.
>>
>>
>> Thanks,
>> ZhangQiao.
>>
>>>
>>> static inline bool entity_long_sleep(se)
>>> {
>>> struct cfs_rq *cfs_rq;
>>> u64 sleep_time;
>>>
>>> if (se->exec_start == 0)
>>> return false;
>>>
>>> cfs_rq = cfs_rq_of(se);
>>> sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
>>> if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
>>> return true;
>>>
>>> return false;
>>> }
>>>
>>>
>>>> + se->vruntime = -sched_sleeper_credit(se);
>>>> + else
>>>> + se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
>>>> }
>>>>
>>>> if (!task_on_rq_migrating(p)) {
>>>>
>>>>
>>>>
>>>> Thanks.
>>>> Zhang Qiao.
>>>>
>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>> Roman.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Amazon Development Center Germany GmbH
>>>>>> Krausenstr. 38
>>>>>> 10117 Berlin
>>>>>> Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss
>>>>>> Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
>>>>>> Sitz: Berlin
>>>>>> Ust-ID: DE 289 237 879
>>>>>>
>>>>>>
>>>>>>
>>>>> .
>>>>>
>>> .
>>>
> .
>

2023-03-03 08:33:10

by Vincent Guittot

[permalink] [raw]

Subject: Re: [PATCH v3] sched/fair: sanitize vruntime of entity being placed

On Fri, 3 Mar 2023 at 07:51, Zhang Qiao <[email protected]> wrote:
>
>
>
> 在 2023/3/2 22:55, Vincent Guittot 写道:
> > On Thu, 2 Mar 2023 at 15:29, Zhang Qiao <[email protected]> wrote:
> >>
> >>
> >>
> >> 在 2023/3/2 21:34, Vincent Guittot 写道:
> >>> On Thu, 2 Mar 2023 at 10:36, Zhang Qiao <[email protected]> wrote:
> >>>>
> >>>>
> >>>>
> >>>> 在 2023/2/27 22:37, Vincent Guittot 写道:
> >>>>> On Mon, 27 Feb 2023 at 09:43, Roman Kagan <[email protected]> wrote:
> >>>>>>
> >>>>>> On Tue, Feb 21, 2023 at 06:26:11PM +0100, Vincent Guittot wrote:
> >>>>>>> On Tue, 21 Feb 2023 at 17:57, Roman Kagan <[email protected]> wrote:
> >>>>>>>> What scares me, though, is that I've got a message from the test robot
> >>>>>>>> that this commit drammatically affected hackbench results, see the quote
> >>>>>>>> below. I expected the commit not to affect any benchmarks.
> >>>>>>>>
> >>>>>>>> Any idea what could have caused this change?
> >>>>>>>
> >>>>>>> Hmm, It's most probably because se->exec_start is reset after a
> >>>>>>> migration and the condition becomes true for newly migrated task
> >>>>>>> whereas its vruntime should be after min_vruntime.
> >>>>>>>
> >>>>>>> We have missed this condition
> >>>>>>
> >>>>>> Makes sense to me.
> >>>>>>
> >>>>>> But what would then be the reliable way to detect a sched_entity which
> >>>>>> has slept for long and risks overflowing in .vruntime comparison?
> >>>>>
> >>>>> For now I don't have a better idea than adding the same check in
> >>>>> migrate_task_rq_fair()
> >>>>
> >>>> Hi, Vincent，
> >>>> I fixed this condition as you said, and the test results are as follows.
> >>>>
> >>>> testcase: hackbench -g 44 -f 20 --process --pipe -l 60000 -s 100
> >>>> version1: v6.2
> >>>> version2: v6.2 + commit 829c1651e9c4
> >>>> version3: v6.2 + commit 829c1651e9c4 + this patch
> >>>>
> >>>> -------------------------------------------------
> >>>> version1 version2 version3
> >>>> test1 81.0 118.1 82.1
> >>>> test2 82.1 116.9 80.3
> >>>> test3 83.2 103.9 83.3
> >>>> avg(s) 82.1 113.0 81.9
> >>>>
> >>>> -------------------------------------------------
> >>>> After deal with the task migration case, the hackbench result has restored.
> >>>>
> >>>> The patch as follow, how does this look?
> >>>>
> >>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>>> index ff4dbbae3b10..3a88d20fd29e 100644
> >>>> --- a/kernel/sched/fair.c
> >>>> +++ b/kernel/sched/fair.c
> >>>> @@ -4648,6 +4648,26 @@ static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
> >>>> #endif
> >>>> }
> >>>>
> >>>> +static inline u64 sched_sleeper_credit(struct sched_entity *se)
> >>>> +{
> >>>> +
> >>>> + unsigned long thresh;
> >>>> +
> >>>> + if (se_is_idle(se))
> >>>> + thresh = sysctl_sched_min_granularity;
> >>>> + else
> >>>> + thresh = sysctl_sched_latency;
> >>>> +
> >>>> + /*
> >>>> + * Halve their sleep time's effect, to allow
> >>>> + * for a gentler effect of sleepers:
> >>>> + */
> >>>> + if (sched_feat(GENTLE_FAIR_SLEEPERS))
> >>>> + thresh >>= 1;
> >>>> +
> >>>> + return thresh;
> >>>> +}
> >>>> +
> >>>> static void
> >>>> place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
> >>>> {
> >>>> @@ -4664,23 +4684,8 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
> >>>> vruntime += sched_vslice(cfs_rq, se);
> >>>>
> >>>> /* sleeps up to a single latency don't count. */
> >>>> - if (!initial) {
> >>>> - unsigned long thresh;
> >>>> -
> >>>> - if (se_is_idle(se))
> >>>> - thresh = sysctl_sched_min_granularity;
> >>>> - else
> >>>> - thresh = sysctl_sched_latency;
> >>>> -
> >>>> - /*
> >>>> - * Halve their sleep time's effect, to allow
> >>>> - * for a gentler effect of sleepers:
> >>>> - */
> >>>> - if (sched_feat(GENTLE_FAIR_SLEEPERS))
> >>>> - thresh >>= 1;
> >>>> -
> >>>> - vruntime -= thresh;
> >>>> - }
> >>>> + if (!initial)
> >>>> + vruntime -= sched_sleeper_credit(se);
> >>>>
> >>>> /*
> >>>> * Pull vruntime of the entity being placed to the base level of
> >>>> @@ -4690,7 +4695,7 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
> >>>> * inversed due to s64 overflow.
> >>>> */
> >>>> sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
> >>>> - if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
> >>>> + if (se->exec_start != 0 && (s64)sleep_time > 60LL * NSEC_PER_SEC)
> >>>> se->vruntime = vruntime;
> >>>> else
> >>>> se->vruntime = max_vruntime(se->vruntime, vruntime);
> >>>> @@ -7634,8 +7639,12 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
> >>>> */
> >>>> if (READ_ONCE(p->__state) == TASK_WAKING) {
> >>>> struct cfs_rq *cfs_rq = cfs_rq_of(se);
> >>>> + u64 sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
> >>>>
> >>>> - se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
> >>>> + if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
> >>>
> >>> You also need to test (se->exec_start !=0) here because the task might
> >>
> >> Hi,
> >>
> >> I don't understand when the another migration happend. Could you tell me in more detail?
> >
> > se->exec_start is update when the task becomes current.
> >
> > You can have the sequence:
> >
> > task TA runs on CPU0
> > TA's se->exec_start = xxxx
> > TA is put back into the rb tree waiting for next slice while another
> > task is running
> > CPU1 pulls TA which migrates on CPU1
> > migrate_task_rq_fair() w/ TA's se->exec_start == xxxx
> > TA's se->exec_start = 0
> > TA is put into the rb tree of CPU1 waiting to run on CPU1
> > CPU2 pulls TA which migrates on CPU2
> > migrate_task_rq_fair() w/ TA's se->exec_start == 0
> > TA's se->exec_start = 0
> Hi, Vincent,
>
> yes, you're right, such sequence does exist. But at this point, p->__state != TASK_WAKING.
>
> I have a question, Whether there is case that is "p->se.exec_start == 0 && p->__state == TASK_WAKING" ?
> I analyzed the code and concluded that this case isn't existed, is it right?

Yes, you're right. Your proposal is enough

Thanks

>
> Thanks.
> ZhangQiao.
>
> >
> >>
> >> I think the next migration will happend after the wakee task enqueued, but at this time
> >> the p->__state isn't TASK_WAKING, p->__state already be changed to TASK_RUNNING at ttwu_do_wakeup().
> >>
> >> If such a migration exists, Previous code "se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);" maybe
> >> perform multiple times，wouldn't it go wrong in this way?
> >
> > the vruntime have been updated when enqueued but not exec_start
> >
> >>
> >>> migrate another time before being scheduled. You should create a
> >>> helper function like below and use it in both place
> >>
> >> Ok, I will update at next version.
> >>
> >>
> >> Thanks,
> >> ZhangQiao.
> >>
> >>>
> >>> static inline bool entity_long_sleep(se)
> >>> {
> >>> struct cfs_rq *cfs_rq;
> >>> u64 sleep_time;
> >>>
> >>> if (se->exec_start == 0)
> >>> return false;
> >>>
> >>> cfs_rq = cfs_rq_of(se);
> >>> sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
> >>> if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
> >>> return true;
> >>>
> >>> return false;
> >>> }
> >>>
> >>>
> >>>> + se->vruntime = -sched_sleeper_credit(se);
> >>>> + else
> >>>> + se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
> >>>> }
> >>>>
> >>>> if (!task_on_rq_migrating(p)) {
> >>>>
> >>>>
> >>>>
> >>>> Thanks.
> >>>> Zhang Qiao.
> >>>>
> >>>>>
> >>>>>>
> >>>>>> Thanks,
> >>>>>> Roman.
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>>> Amazon Development Center Germany GmbH
> >>>>>> Krausenstr. 38
> >>>>>> 10117 Berlin
> >>>>>> Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss
> >>>>>> Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
> >>>>>> Sitz: Berlin
> >>>>>> Ust-ID: DE 289 237 879
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>> .
> >>>>>
> >>> .
> >>>
> > .
> >