On 04/02/21 11:49, Dietmar Eggemann wrote:
> On 03/02/2021 16:15, Qais Yousef wrote:
>> On 01/28/21 18:31, Valentin Schneider wrote:
>
> [...]
>
>>> @@ -10238,7 +10236,7 @@ static void nohz_balancer_kick(struct rq *rq)
>>> * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
>>> * to run the misfit task on.
>>> */
>>> - if (check_misfit_status(rq, sd)) {
>>> + if (check_misfit_status(rq)) {
>
> Since check_misfit_status() doesn't need sd anymore it looks like that
> rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)) could be replaced by
> static_branch_unlikely(&sched_asym_cpucapacity)) in nohz_balancer_kick().
>
> But as you mentioned in an earlier conversation we do need to check sd
> because of asymmetric CPU capacity systems w/ exclusive cpusets which
> could create symmetric islands (unique capacity_orig among CPUs).
>
> Maybe worth putting a comment here (similar to the one in sis()) so
> people don't try to optimize?
How about:
--->8---
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index c2351b87824f..4b71f4d1d324 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6322,15 +6322,8 @@ static int select_idle_sibling(struct task_struct *p,
int prev, int target)
* sd_asym_cpucapacity rather than sd_llc.
*/
if (static_branch_unlikely(&sched_asym_cpucapacity)) {
+ /* See sd_has_asym_cpucapacity() */
sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
- /*
- * On an asymmetric CPU capacity system where an exclusive
- * cpuset defines a symmetric island (i.e. one unique
- * capacity_orig value through the cpuset), the key will be set
- * but the CPUs within that cpuset will not have a domain with
- * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
- * capacity path.
- */
if (sd) {
i = select_idle_capacity(p, sd, target);
return ((unsigned)i < nr_cpumask_bits) ? i : target;
@@ -10274,6 +10267,10 @@ static void nohz_balancer_kick(struct rq *rq)
}
}
+ /*
+ * Below checks don't actually use the sd, but they still hinge on its
+ * presence. See sd_has_asym_cpucapacity().
+ */
sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
if (sd) {
/*
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 21bd71f58c06..ea7f0155e268 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1482,6 +1482,33 @@ DECLARE_PER_CPU(struct sched_domain __rcu *,
sd_asym_packing);
DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
extern struct static_key_false sched_asym_cpucapacity;
+/*
+ * Note that the static key is system-wide, but the visibility of
+ * SD_ASYM_CPUCAPACITY isn't. Thus the static key being enabled does not
+ * imply all CPUs can see asymmetry.
+ *
+ * Consider an asymmetric CPU capacity system such as:
+ *
+ * MC [ ]
+ * 0 1 2 3 4 5
+ * L L L L B B
+ *
+ * w/ arch_scale_cpu_capacity(L) < arch_scale_cpu_capacity(B)
+ *
+ * By default, booting this system will enable the sched_asym_cpucapacity
+ * static key, and all CPUs will see SD_ASYM_CPUCAPACITY set at their MC
+ * sched_domain.
+ *
+ * Further consider exclusive cpusets creating a "symmetric island":
+ *
+ * MC [ ][ ]
+ * 0 1 2 3 4 5
+ * L L L L B B
+ *
+ * Again, booting this will enable the static key, but CPUs 0-1 will *not* have
+ * SD_ASYM_CPUCAPACITY set in any of their sched_domain. This is the intending
+ * behaviour, as CPUs 0-1 should be treated as a regular, isolated SMP system.
+ */
static inline bool sd_has_asym_cpucapacity(struct sched_domain *sd)
{
return static_branch_unlikely(&sched_asym_cpucapacity) &&
