Re: [PATCH v2] sched/fair: sanitize vruntime of entity being migrated

[Vincent Guittot]

On Thu, 9 Mar 2023 at 09:37, Zhang Qiao <zhangqiao22@xxxxxxxxxx> wrote:
>
>
>
> 在 2023/3/8 20:55, Vincent Guittot 写道:
> > Le mercredi 08 mars 2023 à 09:01:05 (+0100), Vincent Guittot a écrit :
> >> On Tue, 7 Mar 2023 at 14:41, Zhang Qiao <zhangqiao22@xxxxxxxxxx> wrote:
> >>>
> >>>
> >>>
> >>> 在 2023/3/7 18:26, Vincent Guittot 写道:
> >>>> On Mon, 6 Mar 2023 at 14:53, Vincent Guittot <vincent.guittot@xxxxxxxxxx> wrote:
> >>>>>
> >>>>> On Mon, 6 Mar 2023 at 13:57, Zhang Qiao <zhangqiao22@xxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> Commit 829c1651e9c4 ("sched/fair: sanitize vruntime of
> >>>>>> entity being placed") fix an overflowing bug, but ignore
> >>>>>> a case that se->exec_start is reset after a migration.
> >>>>>>
> >>>>>> For fixing this case, we reset the vruntime of a long
> >>>>>> sleeping task in migrate_task_rq_fair().
> >>>>>>
> >>>>>> Fixes: 829c1651e9c4 ("sched/fair: sanitize vruntime of entity being placed")
> >>>>>> Suggested-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> >>>>>> Signed-off-by: Zhang Qiao <zhangqiao22@xxxxxxxxxx>
> >>>>>
> >>>>> Reviewed-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> >>>>>
> >>>>>> ---
> >>>>>>
> >>>>>> v1 -> v2:
> >>>>>> - fix some typos and update comments
> >>>>>> - reformat the patch
> >>>>>>
> >>>>>> ---
> >>>>>>  kernel/sched/fair.c | 76 ++++++++++++++++++++++++++++++++-------------
> >>>>>>  1 file changed, 55 insertions(+), 21 deletions(-)
> >>>>>>
> >>>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>>>>> index 7a1b1f855b96..74c9918ffe76 100644
> >>>>>> --- a/kernel/sched/fair.c
> >>>>>> +++ b/kernel/sched/fair.c
> >>>>>> @@ -4648,11 +4648,45 @@ static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
> >>>>>>  #endif
> >>>>>>  }
> >>>>>>
> >>>>>> +static inline bool entity_is_long_sleep(struct sched_entity *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;
> >>>>>> +}
> >>>>>> +
> >>>>>> +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)
> >>>>>>  {
> >>>>>>         u64 vruntime = cfs_rq->min_vruntime;
> >>>>>> -       u64 sleep_time;
> >>>>>>
> >>>>>>         /*
> >>>>>>          * The 'current' period is already promised to the current tasks,
> >>>>>> @@ -4664,23 +4698,8 @@ 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
> >>>>>> @@ -4689,8 +4708,7 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
> >>>>>>          * the base as it may be too far off and the comparison may get
> >>>>>>          * 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 (entity_is_long_sleep(se))
> >>>>>>                 se->vruntime = vruntime;
> >>>>>>         else
> >>>>>>                 se->vruntime = max_vruntime(se->vruntime, vruntime);
> >>>>>> @@ -7635,7 +7653,23 @@ 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);
> >>>>>>
> >>>>>> -               se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
> >>>>>> +               /*
> >>>>>> +                * We determine whether a task sleeps for long by checking
> >>>>>> +                * se->exec_start, and if it is, we sanitize its vruntime at
> >>>>>> +                * place_entity(). However, after a migration, this detection
> >>>>>> +                * method fails due to se->exec_start being reset.
> >>>>>> +                *
> >>>>>> +                * For fixing this case, we add the same check here. For a task
> >>>>>> +                * which has slept for a long time, its vruntime should be reset
> >>>>>> +                * to cfs_rq->min_vruntime with a sleep credit. Because waking
> >>>>>> +                * task's vruntime will be added to cfs_rq->min_vruntime when
> >>>>>> +                * enqueue, we only need to reset the se->vruntime of waking task
> >>>>>> +                * to a credit here.
> >>>>>> +                */
> >>>>>> +               if (entity_is_long_sleep(se))
> >>>>
> >>>> I completely overlooked that we can't use rq_clock_task here. Need to
> >>>> think a bit more on this
> >>>
> >>> Hi,Vincent,
> >>>
> >>> How about using exec_start of the parent sched_entity instant of rq_clock_task()?
> >>
> >> How do we handle sched_entity without a parent sched_entity ?
> >
> > The change below should do the stuff. Not that we can't use it in place entity because
> > pelt is not enabled in UP.
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 74c9918ffe76..b8b381b0ff20 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -7635,6 +7635,32 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> >         return new_cpu;
> >  }
> >
> > +static inline bool migrate_long_sleeper(struct sched_entity *se)
> > +{
> > +       struct cfs_rq *cfs_rq;
> > +       u64 sleep_time;
> > +
> > +       if (se->exec_start == 0)
>
> How about use `se->avg.last_update_time == 0` here?
>
> > +               return false;
> > +
> > +       cfs_rq = cfs_rq_of(se);
> > +       /*
> > +        * If the entity slept for a long time, don't even try to normalize its
> > +        * vruntime with the base as it may be too far off and might generate
> > +        * wrong decision because of s64 overflow.
> > +        * We estimate its sleep duration with the last update of se's pelt.
> > +        * The last update happened before sleeping. The cfs' pelt is not
> > +        * always updated when cfs is idle but this is not a problem because
> > +        * its min_vruntime is not updated too, so the situation can't get
> > +        * worse.
> > +        */
> > +       sleep_time = cfs_rq_last_update_time(cfs_rq) - se->avg.last_update_time;
> > +       if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
>
> I tested with this patch and found it make hackbench slower.

Compared to which version ?
- v6.2 + revert commit 829c1651e9c4 ?
- v6.2 ?

v6.2 has a bug because newly migrated task gets a runtime credit
whatever its previous vruntime and hackbench generates a lot of
migration



>
>
> > +               return true;
> > +
> > +       return false;
> > +}
> > +
> >  /*
> >   * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
> >   * cfs_rq_of(p) references at time of call are still valid and identify the
> > @@ -7666,7 +7692,7 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
> >                  * enqueue, we only need to reset the se->vruntime of waking task
> >                  * to a credit here.
> >                  */
> > -               if (entity_is_long_sleep(se))
> > +               if (migrate_long_sleeper(se))
> >                         se->vruntime = -sched_sleeper_credit(se);
> >                 else
> >                         se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
> >
> >
> >
> >>
> >>
> >>
> >>>>
> >>>>>> +                       se->vruntime = -sched_sleeper_credit(se);
> >>>>>> +               else
> >>>>>> +                       se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
> >>>>>>         }
> >>>>>>
> >>>>>>         if (!task_on_rq_migrating(p)) {
> >>>>>> --
> >>>>>> 2.17.1
> >>>>>>
> >>>> .
> >>>>
> > .
> >