Re: [PATCH 5/9] sched/fair: Take into account latency priority at wakeup

From: Joel Fernandes
Date: Tue Nov 29 2022 - 10:45:59 EST


Hi Vincent,
Thank you so much for your quick reply.

> On Nov 29, 2022, at 3:58 AM, Vincent Guittot <vincent.guittot@xxxxxxxxxx> wrote:
>
> On Tue, 29 Nov 2022 at 05:25, Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>>
>>> On Tue, Nov 15, 2022 at 12:19 PM Vincent Guittot
>>> <vincent.guittot@xxxxxxxxxx> wrote:
>>>
>>> Take into account the latency priority of a thread when deciding to
>>> preempt the current running thread. We don't want to provide more CPU
>>> bandwidth to a thread but reorder the scheduling to run latency sensitive
>>> task first whenever possible.
>>>
>>> As long as a thread didn't use its bandwidth, it will be able to preempt
>>> the current thread.
>>>
>>> At the opposite, a thread with a low latency priority will preempt current
>>> thread at wakeup only to keep fair CPU bandwidth sharing. Otherwise it will
>>> wait for the tick to get its sched slice.
>>>
>>> curr vruntime
>>> |
>>> sysctl_sched_wakeup_granularity
>>> <-->
>>> ----------------------------------|----|-----------------------|---------------
>>> | |<--------------------->
>>> | . sysctl_sched_latency
>>> | .
>>> default/current latency entity | .
>>> | .
>>> 1111111111111111111111111111111111|0000|-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-
>>> se preempts curr at wakeup ------>|<- se doesn't preempt curr -----------------
>>> | .
>>> | .
>>> | .
>>> low latency entity | .
>>> ---------------------->|
>>> % of sysctl_sched_latency |
>>> 1111111111111111111111111111111111111111111111111111111111|0000|-1-1-1-1-1-1-1-
>>> preempt ------------------------------------------------->|<- do not preempt --
>>> | .
>>> | .
>>> | .
>>> high latency entity | .
>>> |<-----------------------|----.
>>> | % of sysctl_sched_latency .
>>> 111111111|0000|-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1-1
>>> preempt->|<- se doesn't preempt curr ------------------------------------------
>>> Tests results of nice latency impact on heavy load like hackbench:
>>>
>>> hackbench -l (2560 / group) -g group
>>> group latency 0 latency 19
>>> 1 1.378(+/- 1%) 1.337(+/- 1%) + 3%
>>> 4 1.393(+/- 3%) 1.312(+/- 3%) + 6%
>>> 8 1.308(+/- 2%) 1.279(+/- 1%) + 2%
>>> 16 1.347(+/- 1%) 1.317(+/- 1%) + 2%
>>>
>>> hackbench -p -l (2560 / group) -g group
>>> group
>>> 1 1.836(+/- 17%) 1.148(+/- 5%) +37%
>>> 4 1.586(+/- 6%) 1.109(+/- 8%) +30%
>>> 8 1.209(+/- 4%) 0.780(+/- 4%) +35%
>>> 16 0.805(+/- 5%) 0.728(+/- 4%) +10%
>>>
>>> By deacreasing the latency prio, we reduce the number of preemption at
>>> wakeup and help hackbench making progress.
>>>
>>> Test results of nice latency impact on short live load like cyclictest
>>> while competing with heavy load like hackbench:
>>>
>>> hackbench -l 10000 -g $group &
>>> cyclictest --policy other -D 5 -q -n
>>> latency 0 latency -20
>>> group min avg max min avg max
>>> 0 16 19 29 17 18 29
>>> 1 43 299 7359 63 84 3422
>>> 4 56 449 14806 45 83 284
>>> 8 63 820 51123 63 83 283
>>> 16 64 1326 70684 41 157 26852
>>>
>>> group = 0 means that hackbench is not running.
>>>
>>> The avg is significantly improved with nice latency -20 especially with
>>> large number of groups but min and max remain quite similar. If we add the
>>> histogram parameter to get details of latency, we have :
>>>
>>> hackbench -l 10000 -g 16 &
>>> cyclictest --policy other -D 5 -q -n -H 20000 --histfile data.txt
>>> latency 0 latency -20
>>> Min Latencies: 64 62
>>> Avg Latencies: 1170 107
>>> Max Latencies: 88069 10417
>>> 50% latencies: 122 86
>>> 75% latencies: 614 91
>>> 85% latencies: 961 94
>>> 90% latencies: 1225 97
>>> 95% latencies: 6120 102
>>> 99% latencies: 18328 159
>>>
>>> With percentile details, we see the benefit of nice latency -20 as
>>> only 1% of the latencies are above 159us whereas the default latency
>>> has got 15% around ~1ms or above and 5% over the 6ms.
>>>
>>> Signed-off-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
>>> ---
>>> include/linux/sched.h | 4 ++-
>>> include/linux/sched/prio.h | 9 ++++++
>>> init/init_task.c | 2 +-
>>> kernel/sched/core.c | 38 +++++++++++++++++++---
>>> kernel/sched/debug.c | 2 +-
>>> kernel/sched/fair.c | 66 ++++++++++++++++++++++++++++++++++----
>>> kernel/sched/sched.h | 6 ++++
>>> 7 files changed, 112 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>>> index 856240573300..2f33326adb8d 100644
>>> --- a/include/linux/sched.h
>>> +++ b/include/linux/sched.h
>>> @@ -568,6 +568,8 @@ struct sched_entity {
>>> /* cached value of my_q->h_nr_running */
>>> unsigned long runnable_weight;
>>> #endif
>>> + /* preemption offset in ns */
>>> + long latency_offset;
>>>
>>> #ifdef CONFIG_SMP
>>> /*
>>> @@ -784,7 +786,7 @@ struct task_struct {
>>> int static_prio;
>>> int normal_prio;
>>> unsigned int rt_priority;
>>> - int latency_nice;
>>> + int latency_prio;
>>>
>>> struct sched_entity se;
>>> struct sched_rt_entity rt;
>>> diff --git a/include/linux/sched/prio.h b/include/linux/sched/prio.h
>>> index bfcd7f1d1e11..be79503d86af 100644
>>> --- a/include/linux/sched/prio.h
>>> +++ b/include/linux/sched/prio.h
>>> @@ -59,5 +59,14 @@ static inline long rlimit_to_nice(long prio)
>>> * Default tasks should be treated as a task with latency_nice = 0.
>>> */
>>> #define DEFAULT_LATENCY_NICE 0
>>> +#define DEFAULT_LATENCY_PRIO (DEFAULT_LATENCY_NICE + LATENCY_NICE_WIDTH/2)
>>> +
>>> +/*
>>> + * Convert user-nice values [ -20 ... 0 ... 19 ]
>>> + * to static latency [ 0..39 ],
>>> + * and back.
>>> + */
>>> +#define NICE_TO_LATENCY(nice) ((nice) + DEFAULT_LATENCY_PRIO)
>>> +#define LATENCY_TO_NICE(prio) ((prio) - DEFAULT_LATENCY_PRIO)
>>>
>>> #endif /* _LINUX_SCHED_PRIO_H */
>>> diff --git a/init/init_task.c b/init/init_task.c
>>> index 7dd71dd2d261..071deff8dbd1 100644
>>> --- a/init/init_task.c
>>> +++ b/init/init_task.c
>>> @@ -78,7 +78,7 @@ struct task_struct init_task
>>> .prio = MAX_PRIO - 20,
>>> .static_prio = MAX_PRIO - 20,
>>> .normal_prio = MAX_PRIO - 20,
>>> - .latency_nice = DEFAULT_LATENCY_NICE,
>>> + .latency_prio = DEFAULT_LATENCY_PRIO,
>>> .policy = SCHED_NORMAL,
>>> .cpus_ptr = &init_task.cpus_mask,
>>> .user_cpus_ptr = NULL,
>>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>>> index 18c31a68eb18..b2b8cb6c08cd 100644
>>> --- a/kernel/sched/core.c
>>> +++ b/kernel/sched/core.c
>>> @@ -1283,6 +1283,16 @@ static void set_load_weight(struct task_struct *p, bool update_load)
>>> }
>>> }
>>>
>>> +static void set_latency_offset(struct task_struct *p)
>>> +{
>>> + long weight = sched_latency_to_weight[p->latency_prio];
>>> + s64 offset;
>>> +
>>> + offset = weight * get_sleep_latency(false);
>>> + offset = div_s64(offset, NICE_LATENCY_WEIGHT_MAX);
>>> + p->se.latency_offset = (long)offset;
>>> +}
>>> +
>>> #ifdef CONFIG_UCLAMP_TASK
>>> /*
>>> * Serializes updates of utilization clamp values
>>> @@ -4592,7 +4602,9 @@ int sched_fork(unsigned long clone_flags, struct task_struct *p)
>>> p->prio = p->normal_prio = p->static_prio;
>>> set_load_weight(p, false);
>>>
>>> - p->latency_nice = DEFAULT_LATENCY_NICE;
>>> + p->latency_prio = NICE_TO_LATENCY(0);
>>> + set_latency_offset(p);
>>> +
>>> /*
>>> * We don't need the reset flag anymore after the fork. It has
>>> * fulfilled its duty:
>>> @@ -7358,8 +7370,10 @@ static void __setscheduler_params(struct task_struct *p,
>>> static void __setscheduler_latency(struct task_struct *p,
>>> const struct sched_attr *attr)
>>> {
>>> - if (attr->sched_flags & SCHED_FLAG_LATENCY_NICE)
>>> - p->latency_nice = attr->sched_latency_nice;
>>> + if (attr->sched_flags & SCHED_FLAG_LATENCY_NICE) {
>>> + p->latency_prio = NICE_TO_LATENCY(attr->sched_latency_nice);
>>> + set_latency_offset(p);
>>> + }
>>> }
>>>
>>> /*
>>> @@ -7544,7 +7558,7 @@ static int __sched_setscheduler(struct task_struct *p,
>>> if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
>>> goto change;
>>> if (attr->sched_flags & SCHED_FLAG_LATENCY_NICE &&
>>> - attr->sched_latency_nice != p->latency_nice)
>>> + attr->sched_latency_nice != LATENCY_TO_NICE(p->latency_prio))
>>> goto change;
>>>
>>> p->sched_reset_on_fork = reset_on_fork;
>>> @@ -8085,7 +8099,7 @@ SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
>>> get_params(p, &kattr);
>>> kattr.sched_flags &= SCHED_FLAG_ALL;
>>>
>>> - kattr.sched_latency_nice = p->latency_nice;
>>> + kattr.sched_latency_nice = LATENCY_TO_NICE(p->latency_prio);
>>>
>>> #ifdef CONFIG_UCLAMP_TASK
>>> /*
>>> @@ -11294,6 +11308,20 @@ const u32 sched_prio_to_wmult[40] = {
>>> /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
>>> };
>>>
>>> +/*
>>> + * latency weight for wakeup preemption
>>> + */
>>> +const int sched_latency_to_weight[40] = {
>>> + /* -20 */ -1024, -973, -922, -870, -819,
>>> + /* -15 */ -768, -717, -666, -614, -563,
>>> + /* -10 */ -512, -461, -410, -358, -307,
>>> + /* -5 */ -256, -205, -154, -102, -51,
>>> + /* 0 */ 0, 51, 102, 154, 205,
>>> + /* 5 */ 256, 307, 358, 410, 461,
>>> + /* 10 */ 512, 563, 614, 666, 717,
>>> + /* 15 */ 768, 819, 870, 922, 973,
>>> +};
>>> +
>>
>> The table is linear. You could approximate this as: weight = nice * 51
>> since it is a linear scale and do the conversion in place.
>>
>> Or, since the only place you are using the latency_to_weight is in
>> set_latency_offset(), can we drop the sched_latency_to_weight array
>> and simplify as follows?
>
> It's also used in cgroup patch and keeps a coherency between
> nice/weight an latency_nice/offset so I prefer

I dont think it’s a valid comparison as nice/weight conversion are non linear and over there a table makes sense: weight = 1024 / 1.25 ^ nice

> keeping current
> implementation

I could be missing something, but, since its a linear scale, why does cgroup need weight at all? Just store nice directly. Why would that not work?

In the end the TG and SE has the latency offset in the struct, that is all you care about. All the conversion back and forth is unnecessary, as it is a linear scale and just increases LOC and takes more memory to store linear arrays.

Again I could be missing something and I will try to play with your series and see if I can show you what I mean (or convince myself it’s needed).

>> static void set_latency_offset(struct task_struct *p)
>> {
>> s64 offset = p->latency_prio * get_sleep_latency(false);
>> p->latency_prio = (long)div_s64(offset, 40);
>> }
>>
>>> void call_trace_sched_update_nr_running(struct rq *rq, int count)
>>> {
>>> trace_sched_update_nr_running_tp(rq, count);
>>> diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c
>>> index 68be7a3e42a3..b3922184af91 100644
>>> --- a/kernel/sched/debug.c
>>> +++ b/kernel/sched/debug.c
>>> @@ -1043,7 +1043,7 @@ void proc_sched_show_task(struct task_struct *p, struct pid_namespace *ns,
>>> #endif
>>> P(policy);
>>> P(prio);
>>> - P(latency_nice);
>>> + P(latency_prio);
>>> if (task_has_dl_policy(p)) {
>>> P(dl.runtime);
>>> P(dl.deadline);
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index c8a697f8db88..0e80e65113bd 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -4858,6 +4858,8 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
>>> update_idle_cfs_rq_clock_pelt(cfs_rq);
>>> }
>>>
>>> +static long wakeup_latency_gran(struct sched_entity *curr, struct sched_entity *se);
>>> +
>>> /*
>>> * Preempt the current task with a newly woken task if needed:
>>> */
>>> @@ -4866,7 +4868,7 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>>> {
>>> unsigned long ideal_runtime, delta_exec;
>>> struct sched_entity *se;
>>> - s64 delta;
>>> + s64 delta, offset;
>>>
>>> ideal_runtime = sched_slice(cfs_rq, curr);
>>> delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
>>> @@ -4891,10 +4893,12 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>>> se = __pick_first_entity(cfs_rq);
>>> delta = curr->vruntime - se->vruntime;
>>>
>>> - if (delta < 0)
>>> + offset = wakeup_latency_gran(curr, se);
>>> + if (delta < offset)
>>> return;
>>
>> Agreed.
>>
>>> - if (delta > ideal_runtime)
>>> + if ((delta > ideal_runtime) ||
>>> + (delta > get_latency_max()))
>>> resched_curr(rq_of(cfs_rq));
>>> }
>>>
>>> @@ -6019,6 +6023,35 @@ static int sched_idle_cpu(int cpu)
>>> }
>>> #endif
>>>
>>> +static void set_next_buddy(struct sched_entity *se);
>>> +
>>> +static void check_preempt_from_others(struct cfs_rq *cfs, struct sched_entity *se)
>>> +{
>>> + struct sched_entity *next;
>>> +
>>> + if (se->latency_offset >= 0)
>>> + return;
>>> +
>>> + if (cfs->nr_running <= 1)
>>> + return;
>>> + /*
>>> + * When waking from another class, we don't need to check to preempt at
>>> + * wakeup and don't set next buddy as a candidate for being picked in
>>> + * priority.
>>> + * In case of simultaneous wakeup when current is another class, the
>>> + * latency sensitive tasks lost opportunity to preempt non sensitive
>>> + * tasks which woke up simultaneously.
>>> + */
>>> +
>>> + if (cfs->next)
>>> + next = cfs->next;
>>> + else
>>> + next = __pick_first_entity(cfs);
>>> +
>>> + if (next && wakeup_preempt_entity(next, se) == 1)
>>> + set_next_buddy(se);
>>> +}
>>> +
>>> /*
>>> * The enqueue_task method is called before nr_running is
>>> * increased. Here we update the fair scheduling stats and
>>> @@ -6105,14 +6138,15 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
>>> if (!task_new)
>>> update_overutilized_status(rq);
>>>
>>> + if (rq->curr->sched_class != &fair_sched_class)
>>> + check_preempt_from_others(cfs_rq_of(&p->se), &p->se);
>>> +
>>> enqueue_throttle:
>>> assert_list_leaf_cfs_rq(rq);
>>>
>>> hrtick_update(rq);
>>> }
>>>
>>> -static void set_next_buddy(struct sched_entity *se);
>>> -
>>> /*
>>> * The dequeue_task method is called before nr_running is
>>> * decreased. We remove the task from the rbtree and
>>> @@ -7461,6 +7495,23 @@ balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
>>> }
>>> #endif /* CONFIG_SMP */
>>>
>>> +static long wakeup_latency_gran(struct sched_entity *curr, struct sched_entity *se)
>>> +{
>>> + long latency_offset = se->latency_offset;
>>> +
>>> + /*
>>> + * A negative latency offset means that the sched_entity has latency
>>> + * requirement that needs to be evaluated versus other entity.
>>> + * Otherwise, use the latency weight to evaluate how much scheduling
>>> + * delay is acceptable by se.
>>> + */
>>> + if ((latency_offset < 0) || (curr->latency_offset < 0))
>>> + latency_offset -= curr->latency_offset;
>>> + latency_offset = min_t(long, latency_offset, get_latency_max());
>>
>> Over here can we make positive latency offsets also be evaluated
>> "versus other entity"?
>>
>> It feels strange to have different rules for positive latency_offset
>> when comparing curr and se. IMO we should also factor in latency
>> requirements by comparing 2 positive nice values. It should be
>> relative even for positive values, just like regular nice IMO and not
>> have hidden meaning. If there is hidden meaning, it confuses the user
>> and requires documentation that most users will not read. Especially
>> because latency_nice shares the word "nice" with regular nice values.
>
> This has already been discussed in the previous revisions.

Sorry to be late to the party.

> This is not
> a hidden behavior but the normal behavior.
>
> A negative latency nice, means that the task are not tolerant to
> scheduling delay and it want to preempt current and run now. Or, if
> the task is current, it doesn't want to be preempted and finish its
> slice. In this case, we compare current and wake up task in case there
> is 2 latency sensitive tasks that are fighting to run 1st.
>
> Whereas a positive latency nice means that the task is tolerant to
> scheduling delay and you don't care preempting current as long as it's
> in an acceptable vruntime range. Why would the latency nice of the
> current task make the wakeup task less tolerant to the scheduling
> delay ? As an example, If current is latency_nice 19 and the wakeup
> task is latency nice 19 too, both are tolerant to scheduling delay and
> the waking up task should preempt current only if there is an
> unfairness problem. By comparing their positive latency nice values,
> you are back to the normal behavior which defeats the purpose of the
> feature.

I see it as, if 2 tasks are latency tolerant, then they will have higher latency with respect to a third tasks that is latency in tolerant. But I am ok with your definition as well…

Thanks!

- Joel

>
> Thanks
> Vincent
>
>>
>> Thanks,
>>
>> - Joel
>> .
>>
>>> +
>>> + return latency_offset;
>>> +}
>>> +
>>> static unsigned long wakeup_gran(struct sched_entity *se)
>>> {
>>> unsigned long gran = sysctl_sched_wakeup_granularity;
>>> @@ -7499,11 +7550,12 @@ static int
>>> wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
>>> {
>>> s64 gran, vdiff = curr->vruntime - se->vruntime;
>>> + s64 offset = wakeup_latency_gran(curr, se);
>>>
>>> - if (vdiff <= 0)
>>> + if (vdiff < offset)
>>> return -1;
>>>
>>> - gran = wakeup_gran(se);
>>> + gran = offset + wakeup_gran(se);
>>>
>>> /*
>>> * At wake up, the vruntime of a task is capped to not be older than
>>> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
>>> index 842ce0094d9c..7292652731d0 100644
>>> --- a/kernel/sched/sched.h
>>> +++ b/kernel/sched/sched.h
>>> @@ -125,6 +125,11 @@ extern int sched_rr_timeslice;
>>> */
>>> #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
>>>
>>> +/* Maximum nice latency weight used to scale the latency_offset */
>>> +
>>> +#define NICE_LATENCY_SHIFT (SCHED_FIXEDPOINT_SHIFT)
>>> +#define NICE_LATENCY_WEIGHT_MAX (1L << NICE_LATENCY_SHIFT)
>>> +
>>> /*
>>> * Increase resolution of nice-level calculations for 64-bit architectures.
>>> * The extra resolution improves shares distribution and load balancing of
>>> @@ -2115,6 +2120,7 @@ static_assert(WF_TTWU == SD_BALANCE_WAKE);
>>>
>>> extern const int sched_prio_to_weight[40];
>>> extern const u32 sched_prio_to_wmult[40];
>>> +extern const int sched_latency_to_weight[40];
>>>
>>> /*
>>> * {de,en}queue flags:
>>> --
>>> 2.17.1
>>>