Re: [PATCH v2 5/5] sched/fair: Fix use of find_idlest_group when local group is idlest.

[Brendan Jackman]

On Mon, Aug 28 2017 at 08:56, Vincent Guittot wrote:
> On 25 August 2017 at 17:51, Brendan Jackman <brendan.jackman@xxxxxxx> wrote:
>>
>> On Fri, Aug 25 2017 at 13:38, Vincent Guittot wrote:
>>> On 25 August 2017 at 12:16, Brendan Jackman <brendan.jackman@xxxxxxx> wrote:
>>>> find_idlest_group currently returns NULL when the local group is
>>>> idlest. The caller then continues the find_idlest_group search at a
>>>> lower level of the current CPU's sched_domain hierarchy. However
>>>> find_idlest_group_cpu is not consulted and, crucially, @new_cpu is
>>>> not updated. This means the search is pointless and we return
>>>> @prev_cpu from select_task_rq_fair.
>>>>
>>>> This patch makes find_idlest_group return the idlest group, and never
>>>> NULL, and then has the caller unconditionally continue to consult
>>>> find_idlest_group_cpu and update @new_cpu.
>>>>
>>>>  * * *
>>>>
>>>> An alternative, simpler, fix for this would be to initialize @new_cpu
>>>> to @cpu instead of @prev_cpu, at the beginning of the new
>>>> find_idlest_cpu. However this would not fix the fact that
>>>
>>> Yes this is simpler
>>>
>>>> find_idlest_group_cpu goes unused, and we miss out on the bias toward
>>>> a shallower-idle CPU, unless find_idlest_group picks a non-local
>>>> group.
>>>
>>> I'm not sure to catch why it's not enough.
>>> If no cpu of sched_domain span is allowed, we continue to return prev_cpu
>>> But if at least 1 cpu is allowed, the default new_cpu is cpu in case
>>> it is really the idlest and no other group will be returned by
>>> find_idlest_group
>>> Otherwise, find_idlest_group will finally return another group than
>>> the local one when running with sd->child
>>
>> This patch isn't about affinity but just about which group is idlest.
>
> So this patch changes 2 things:
> - don't return @prev if @cpu is the idlest one
> - always call find_idlest_group_cpu even for local group
>
> The 1st point is a bug but the 2nd one is more a matter of policy.

I didn't see it this way before but yes I suppose you're right.

> Current implementation favors locality and load/capacity over current
> idle state of CPU which is only used at a last resort to select a cpu.
> What you are proposing, it's to increase the weight of the current
> idle state of cpus in the selection of the cpu in order to become
> sometime higher than the locality of even the load/capacity

Well, not just increase the weight, more give it a weight that is
consistent whether or not find_idlest_group ever returns a non-local
group.

OK, so I'll spin a v3 that just fixes the bug and doesn't change the
policy wrt. using idle states.

Thanks a lot for this review.

>
>>
>> Consider a 4 CPU system with 2 shared caches where CPU0's sched_domains
>> look like:
>>
>> DIE: [0  1][2  3]
>>  MC: [0][1]
>>
>> Suppose CPUs 0 and 1 are equally idle wrt load and util, but CPU1 is in
>> a shallower idle state. Suppose all CPUs are allowed.
>>
>> Assuming @cpu is 0, on the first iteration, find_idlest_group (fig) will
>> currently return NULL meaning [0 1], then on the second iteration it
>> will return NULL meaning [0].
>>
>> With this patch, on the first iteration fig will return [0 1], then
>> find_idlest_group_cpu (figc) will return 1. Then on the second
>> iteration, fig should return [1] (because it is biased towards the local
>> group i.e. [@cpu] i.e. [1]). This is better because CPU1 is in a
>> shallower idle state.
>>
>> So basically we're currently missing out on the biasing in figc, unless
>> we at some point return a non-local group.
>>
>> Does that make sense?
>
> With your previous example if the load/capacity of CPU0 is
> significantly lower than CPU1, we will also select CPU1 instead of
> CPU0:
>
> @cpu is 0.
> fig returns local group [0 1]
> figc returns 1
> @new_cpu != @cpu so we don't go to child
>
> cpu = new cpu
> sd = NULL
> weight = 4
> sd stays null  because weight == tmp->span_weight (DIE)
> we don't loop at lower level
> fic returns CPU1
>
>>
>>>
>>> Is there sched_domain topology where the sched_group of lowest level
>>> is not with only 1 cpu ?
>>>
>>>>
>>>> If that alternative solution was preferred, then some code could be
>>>> removed in recognition of the fact that when find_idlest_group_cpu
>>>> was called, @cpu would never be a member of @group (because @group
>>>> would be a non-local group, and since @sd is derived from @cpu, @cpu
>>>> would always be in @sd's local group).
>>>>
>>>> Signed-off-by: Brendan Jackman <brendan.jackman@xxxxxxx>
>>>> Cc: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
>>>> Cc: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
>>>> Cc: Josef Bacik <josef@xxxxxxxxxxxxxx>
>>>> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
>>>> Cc: Morten Rasmussen <morten.rasmussen@xxxxxxx>
>>>> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>>>> ---
>>>>  kernel/sched/fair.c | 29 ++++++++++-------------------
>>>>  1 file changed, 10 insertions(+), 19 deletions(-)
>>>>
>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>> index 26080917ff8d..93e2746d3c30 100644
>>>> --- a/kernel/sched/fair.c
>>>> +++ b/kernel/sched/fair.c
>>>> @@ -5384,11 +5384,10 @@ static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
>>>>   * Assumes p is allowed on at least one CPU in sd.
>>>>   */
>>>>  static struct sched_group *
>>>> -find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>>>> -                 int this_cpu, int sd_flag)
>>>> +find_idlest_group(struct sched_domain *sd, struct task_struct *p, int sd_flag)
>>>>  {
>>>> -       struct sched_group *idlest = NULL, *group = sd->groups;
>>>> -       struct sched_group *most_spare_sg = NULL;
>>>> +       struct sched_group *group = sd->groups, *local_group = sd->groups;
>>>> +       struct sched_group *idlest = NULL, *most_spare_sg = NULL;
>>>>         unsigned long min_runnable_load = ULONG_MAX;
>>>>         unsigned long this_runnable_load = ULONG_MAX;
>>>>         unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
>>>> @@ -5404,7 +5403,6 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>>>>         do {
>>>>                 unsigned long load, avg_load, runnable_load;
>>>>                 unsigned long spare_cap, max_spare_cap;
>>>> -               int local_group;
>>>>                 int i;
>>>>
>>>>                 /* Skip over this group if it has no CPUs allowed */
>>>> @@ -5412,9 +5410,6 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>>>>                                         &p->cpus_allowed))
>>>>                         continue;
>>>>
>>>> -               local_group = cpumask_test_cpu(this_cpu,
>>>> -                                              sched_group_span(group));
>>>> -
>>>>                 /*
>>>>                  * Tally up the load of all CPUs in the group and find
>>>>                  * the group containing the CPU with most spare capacity.
>>>> @@ -5425,7 +5420,7 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>>>>
>>>>                 for_each_cpu(i, sched_group_span(group)) {
>>>>                         /* Bias balancing toward cpus of our domain */
>>>> -                       if (local_group)
>>>> +                       if (group == local_group)
>>>>                                 load = source_load(i, load_idx);
>>>>                         else
>>>>                                 load = target_load(i, load_idx);
>>>> @@ -5446,7 +5441,7 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>>>>                 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
>>>>                                         group->sgc->capacity;
>>>>
>>>> -               if (local_group) {
>>>> +               if (group == local_group) {
>>>>                         this_runnable_load = runnable_load;
>>>>                         this_avg_load = avg_load;
>>>>                         this_spare = max_spare_cap;
>>>> @@ -5492,21 +5487,21 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>>>>
>>>>         if (this_spare > task_util(p) / 2 &&
>>>>             imbalance_scale*this_spare > 100*most_spare)
>>>> -               return NULL;
>>>> +               return local_group;
>>>>
>>>>         if (most_spare > task_util(p) / 2)
>>>>                 return most_spare_sg;
>>>>
>>>>  skip_spare:
>>>>         if (!idlest)
>>>> -               return NULL;
>>>> +               return local_group;
>>>>
>>>>         if (min_runnable_load > (this_runnable_load + imbalance))
>>>> -               return NULL;
>>>> +               return local_group;
>>>>
>>>>         if ((this_runnable_load < (min_runnable_load + imbalance)) &&
>>>>              (100*this_avg_load < imbalance_scale*min_avg_load))
>>>> -               return NULL;
>>>> +               return local_group;
>>>>
>>>>         return idlest;
>>>>  }
>>>> @@ -5582,11 +5577,7 @@ static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p
>>>>                         continue;
>>>>                 }
>>>>
>>>> -               group = find_idlest_group(sd, p, cpu, sd_flag);
>>>> -               if (!group) {
>>>> -                       sd = sd->child;
>>>> -                       continue;
>>>> -               }
>>>> +               group = find_idlest_group(sd, p, sd_flag);
>>>>
>>>>                 new_cpu = find_idlest_group_cpu(group, p, cpu);
>>>>                 if (new_cpu == cpu) {
>>>> --
>>>> 2.14.1
>>>>
>>