On Tue, 2014-03-18 at 16:55 +0000, Jon Medhurst (Tixy) wrote:

On Tue, 2014-03-18 at 15:32 +0000, Chris Redpath wrote:

...

The HMP active migration code is functionally identical to the CFS active migration code apart from one flag check. Share the code and make the flag check optional.

Two wrapper functions allow the flag check to be present or not.

Signed-off-by: Chris Redpath chris.redpath@arm.com

This gets build errors if SCHED_HMP is not configured.

The changes below fix the build errors. Chris, are you OK for me to squash that into your patch?

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 4e3686b..3431533 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6211,10 +6211,17 @@ out_one_pinned: out: return ld_moved; } + #ifdef CONFIG_SCHED_HMP static unsigned int hmp_idle_pull(int this_cpu); -#endif static int move_specific_task(struct lb_env *env, struct task_struct *pm); +#else +inline int move_specific_task(struct lb_env *env, struct task_struct *pm) +{ + return 0; +} +#endif + /* * idle_balance is called by schedule() if this_cpu is about to become * idle. Attempts to pull tasks from other CPUs. @@ -6281,11 +6288,12 @@ static int __do_active_load_balance_cpu_stop(void *data, bool check_sd_lb_flag) int target_cpu = busiest_rq->push_cpu; struct rq *target_rq = cpu_rq(target_cpu); struct sched_domain *sd; - struct task_struct *p; + struct task_struct *p = NULL;

raw_spin_lock_irq(&busiest_rq->lock); +#ifdef CONFIG_SCHED_HMP p = busiest_rq->migrate_task; - +#endif /* make sure the requested cpu hasn't gone down in the meantime */ if (unlikely(busiest_cpu != smp_processor_id() || !busiest_rq->active_balance))

In anticipation of modifying the up_threshold handling, make all instances use the same utility fn to check if a task is eligible for up-migration. This also removes the previous difference in threshold comparison where up-migration used '!<threshold' and idle pull used '>threshold' to decide up-migration eligibility. Make them both use '!<threshold' instead for consistency, although this is unlikely to change any results.

Signed-off-by: Chris Redpath chris.redpath@arm.com --- kernel/sched/fair.c | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 62a8808..29e2c74 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6734,6 +6734,14 @@ static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } #endif

#ifdef CONFIG_SCHED_HMP +static unsigned int hmp_task_eligible_for_up_migration(struct sched_entity *se) +{ + /* below hmp_up_threshold, never eligible */ + if (se->avg.load_avg_ratio < hmp_up_threshold) + return 0; + return 1; +} + /* Check if task should migrate to a faster cpu */ static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se) { @@ -6749,7 +6757,7 @@ static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_enti if (p->prio >= hmp_up_prio) return 0; #endif - if (se->avg.load_avg_ratio < hmp_up_threshold) + if (!hmp_task_eligible_for_up_migration(se)) return 0;

/* Let the task load settle before doing another up migration */ @@ -7237,7 +7245,10 @@ static unsigned int hmp_idle_pull(int this_cpu) } orig = curr; curr = hmp_get_heaviest_task(curr, 1); - if (curr->avg.load_avg_ratio > hmp_up_threshold && + /* check if heaviest eligible task on this + * CPU is heavier than previous task + */ + if (hmp_task_eligible_for_up_migration(curr) && curr->avg.load_avg_ratio > ratio) { p = task_of(curr); target = rq;

When a normal forced up-migration takes place we stop the task to be migrated while the target CPU becomes available. This delay can range from 80us to 1500us on TC2 if the target CPU is in a deep idle state.

Instead, interrupt the target CPU and ask it to pull a task. This lets the current eligible task continue executing on the original CPU while the target CPU wakes. Use a pinned timer to prevent the pulling CPU going back into power-down with pending up-migrations.

If we trigger for a nohz kick, it doesn't matter about triggering for an idle pull since the idle_pull flag will be set when we execute the softirq and we'll still do the idle pull.

If the target CPU is busy, we will not pull any tasks.

Signed-off-by: Chris Redpath chris.redpath@arm.com --- kernel/sched/core.c | 11 +++- kernel/sched/fair.c | 142 ++++++++++++++++++++++++++++++++++++++++++++++---- kernel/sched/sched.h | 1 + 3 files changed, 142 insertions(+), 12 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c index de9d360..f583951 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -1407,7 +1407,11 @@ void scheduler_ipi(void) { if (llist_empty(&this_rq()->wake_list) && !tick_nohz_full_cpu(smp_processor_id()) - && !got_nohz_idle_kick()) + && !got_nohz_idle_kick() +#ifdef CONFIG_SCHED_HMP + && !this_rq()->wake_for_idle_pull +#endif + ) return;

/* @@ -1434,6 +1438,11 @@ void scheduler_ipi(void) this_rq()->idle_balance = 1; raise_softirq_irqoff(SCHED_SOFTIRQ); } +#ifdef CONFIG_SCHED_HMP + else if (unlikely(this_rq()->wake_for_idle_pull)) + raise_softirq_irqoff(SCHED_SOFTIRQ); +#endif + irq_exit(); }

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index af82120..1929e32 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -39,6 +39,9 @@ */ #include <linux/cpufreq.h> #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */ +#ifdef CONFIG_SCHED_HMP +#include <linux/cpuidle.h> +#endif

#include "sched.h"

@@ -3541,6 +3544,110 @@ static const int hmp_max_tasks = 5;

extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);

+#ifdef CONFIG_CPU_IDLE +/* + * hmp_idle_pull: + * + * In this version we have stopped using forced up migrations when we + * detect that a task running on a little CPU should be moved to a bigger + * CPU. In most cases, the bigger CPU is in a deep sleep state and a forced + * migration means we stop the task immediately but need to wait for the + * target CPU to wake up before we can restart the task which is being + * moved. Instead, we now wake a big CPU with an IPI and ask it to pull + * a task when ready. This allows the task to continue executing on its + * current CPU, reducing the amount of time that the task is stalled for. + * + * keepalive timers: + * + * The keepalive timer is used as a way to keep a CPU engaged in an + * idle pull operation out of idle while waiting for the source + * CPU to stop and move the task. Ideally this would not be necessary + * and we could impose a temporary zero-latency requirement on the + * current CPU, but in the current QoS framework this will result in + * all CPUs in the system being unable to enter idle states which is + * not desirable. The timer does not perform any work when it expires. + */ +struct hmp_keepalive { + bool init; + ktime_t delay; /* if zero, no need for timer */ + struct hrtimer timer; +}; +DEFINE_PER_CPU(struct hmp_keepalive, hmp_cpu_keepalive); + +/* setup per-cpu keepalive timers */ +static enum hrtimer_restart hmp_cpu_keepalive_notify(struct hrtimer *hrtimer) +{ + return HRTIMER_NORESTART; +} + +/* + * Work out if any of the idle states have an exit latency too high for us. + * ns_delay is passed in containing the max we are willing to tolerate. + * If there are none, set ns_delay to zero. + * If there are any, set ns_delay to + * ('target_residency of state with shortest too-big latency' - 1) * 1000. + */ +static void hmp_keepalive_delay(unsigned int *ns_delay) +{ + struct cpuidle_driver *drv; + drv = cpuidle_driver_ref(); + if (drv) { + unsigned int us_delay = UINT_MAX; + unsigned int us_max_delay = *ns_delay / 1000; + int idx; + /* if cpuidle states are guaranteed to be sorted we + * could stop at the first match. + */ + for (idx = 0; idx < drv->state_count; idx++) { + if (drv->states[idx].exit_latency > us_max_delay && + drv->states[idx].target_residency < us_delay) { + us_delay = drv->states[idx].target_residency; + } + } + if (us_delay == UINT_MAX) + *ns_delay = 0; /* no timer required */ + else + *ns_delay = 1000 * (us_delay - 1); + } + cpuidle_driver_unref(); +} + +static void hmp_cpu_keepalive_trigger(void) +{ + int cpu = smp_processor_id(); + struct hmp_keepalive *keepalive = &per_cpu(hmp_cpu_keepalive, cpu); + if (!keepalive->init) { + unsigned int ns_delay = 100000; /* tolerate 100usec delay */ + + hrtimer_init(&keepalive->timer, + CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); + keepalive->timer.function = hmp_cpu_keepalive_notify; + + hmp_keepalive_delay(&ns_delay); + keepalive->delay = ns_to_ktime(ns_delay); + keepalive->init = true; + } + if (ktime_to_ns(keepalive->delay)) + hrtimer_start(&keepalive->timer, + keepalive->delay, HRTIMER_MODE_REL_PINNED); +} + +static void hmp_cpu_keepalive_cancel(int cpu) +{ + struct hmp_keepalive *keepalive = &per_cpu(hmp_cpu_keepalive, cpu); + if (keepalive->init) + hrtimer_cancel(&keepalive->timer); +} +#else /* !CONFIG_CPU_IDLE */ +static void hmp_cpu_keepalive_trigger(void) +{ +} + +static void hmp_cpu_keepalive_cancel(int cpu) +{ +} +#endif + /* Setup hmp_domains */ static int __init hmp_cpu_mask_setup(void) { @@ -3601,6 +3708,8 @@ static void hmp_offline_cpu(int cpu)

if(domain) cpumask_clear_cpu(cpu, &domain->cpus); + + hmp_cpu_keepalive_cancel(cpu); } /* * Needed to determine heaviest tasks etc. @@ -7030,7 +7139,7 @@ static void hmp_force_up_migration(int this_cpu) target = cpu_rq(cpu); raw_spin_lock_irqsave(&target->lock, flags); curr = target->cfs.curr; - if (!curr) { + if (!curr || target->active_balance) { raw_spin_unlock_irqrestore(&target->lock, flags); continue; } @@ -7047,16 +7156,13 @@ static void hmp_force_up_migration(int this_cpu) curr = hmp_get_heaviest_task(curr, 1); p = task_of(curr); if (hmp_up_migration(cpu, &target_cpu, curr)) { - if (!target->active_balance) { - get_task_struct(p); - target->push_cpu = target_cpu; - target->migrate_task = p; - got_target = 1; - trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_FORCE); - hmp_next_up_delay(&p->se, target->push_cpu); - } + cpu_rq(target_cpu)->wake_for_idle_pull = 1; + raw_spin_unlock_irqrestore(&target->lock, flags); + spin_unlock(&hmp_force_migration); + smp_send_reschedule(target_cpu); + return; } - if (!got_target && !target->active_balance) { + if (!got_target) { /* * For now we just check the currently running task. * Selecting the lightest task for offloading will @@ -7078,7 +7184,7 @@ static void hmp_force_up_migration(int this_cpu) * is not currently running move it, otherwise let the * CPU stopper take care of it. */ - if (got_target && !target->active_balance) { + if (got_target) { if (!task_running(target, p)) { trace_sched_hmp_migrate_force_running(p, 0); hmp_migrate_runnable_task(target); @@ -7184,6 +7290,8 @@ static unsigned int hmp_idle_pull(int this_cpu) raw_spin_unlock_irqrestore(&target->lock, flags);

if (force) { + /* start timer to keep us awake */ + hmp_cpu_keepalive_trigger(); stop_one_cpu_nowait(cpu_of(target), hmp_active_task_migration_cpu_stop, target, &target->active_balance_work); @@ -7207,6 +7315,18 @@ static void run_rebalance_domains(struct softirq_action *h) enum cpu_idle_type idle = this_rq->idle_balance ? CPU_IDLE : CPU_NOT_IDLE;

+#ifdef CONFIG_SCHED_HMP + /* shortcut for hmp idle pull wakeups */ + if (unlikely(this_rq->wake_for_idle_pull)) { + this_rq->wake_for_idle_pull = 0; + if (hmp_idle_pull(this_cpu)) { + /* break out unless running nohz idle as well */ + if (idle != CPU_IDLE) + return; + } + } +#endif + hmp_force_up_migration(this_cpu);

rebalance_domains(this_cpu, idle); diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h index 989c5ae..0d19ede 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -466,6 +466,7 @@ struct rq { struct cpu_stop_work active_balance_work; #ifdef CONFIG_SCHED_HMP struct task_struct *migrate_task; + int wake_for_idle_pull; #endif /* cpu of this runqueue: */ int cpu;

Hi Chris,

Any comments are appreciated! :)

On 04/04/2014 10:49 PM, Alex Shi wrote:

Hi Chris, I very appreciate for your so detailed explanations! And looks like your patch do a very excellent improvement on this issue.

I am just wondering if we could use a bit simple way to resolve this problem like the following patch. let the task moving destination cpu do active load balance instead of source cpu. So it could give the time to source cpu when destination waking. and don't need let destination keepalive. What's your opinions on this?

iff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 7984458..f30e598 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6275,7 +6275,7 @@ more_balance: raw_spin_unlock_irqrestore(&busiest->lock, flags); if (active_balance) {

•                           stop_one_cpu_nowait(cpu_of(busiest),
  

•                           stop_one_cpu_nowait(busiest->push_cpu,
                                    active_load_balance_cpu_stop, busiest,
                                    &busiest->active_balance_work);
                    }
  

@@ -7198,7 +7198,7 @@ static void hmp_force_up_migration(int this_cpu) raw_spin_unlock_irqrestore(&target->lock, flags); if (force)

•                   stop_one_cpu_nowait(cpu_of(target),
  

•                   stop_one_cpu_nowait(target->push_cpu,
                            hmp_active_task_migration_cpu_stop,
                            target, &target->active_balance_work);
    }
  

@@ -7295,7 +7295,7 @@ static unsigned int hmp_idle_pull(int this_cpu) if (force) { /* start timer to keep us awake */ hmp_cpu_keepalive_trigger();

•           stop_one_cpu_nowait(cpu_of(target),
  

•           stop_one_cpu_nowait(target->push_cpu,
                    hmp_active_task_migration_cpu_stop,
                    target, &target->active_balance_work);
    }
  
  

...

...

...

...

So do you have data show the trade off is worthy? like, the res interrupt cost, cpu alive cost VS go to idle and be wakeup cost. or benchmark data to show we get benefit from performance/power.

I have traces which show the resulting improvement but it's so small that it is lost in the noise in all the benchmarks we have. Most of the benchmarks do not actually involve that much migration between clusters

• typically the 'benchmark' app tasks start heavy processing and

continue until complete, and with the HMP thresholds we use our lighter workloads generally migrate once or twice per operation.

Is your statistic too sensitive to share in linaro-kernel ml? If not, I will be very glad to see your data. :)

When looking for a task to be idle-pulled, don't consider tasks where the affinity does not allow that task to be placed on the target CPU. Also ensure that tasks with restricted affinity do not block selecting other unrestricted busy tasks.

Use the knowledge of target CPU more effectively in idle pull by passing to hmp_get_heaviest_task when we know it, otherwise only checking for general affinity matches with any of the CPUs in the bigger HMP domain.

We still need to explicitly check affinity is allowed in idle pull since if we find no match in hmp_get_heaviest_task we will return the current one, which may not be affine to the new CPU despite having high enough load. In this case, there is nothing to move.

Signed-off-by: Chris Redpath chris.redpath@arm.com --- kernel/sched/fair.c | 31 +++++++++++++++++-------------- 1 file changed, 17 insertions(+), 14 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 1929e32..681a24e 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3721,30 +3721,31 @@ static inline struct hmp_domain *hmp_faster_domain(int cpu);

/* must hold runqueue lock for queue se is currently on */ static struct sched_entity *hmp_get_heaviest_task( - struct sched_entity *se, int migrate_up) + struct sched_entity *se, int target_cpu) { int num_tasks = hmp_max_tasks; struct sched_entity *max_se = se; unsigned long int max_ratio = se->avg.load_avg_ratio; const struct cpumask *hmp_target_mask = NULL; + struct hmp_domain *hmp;

- if (migrate_up) { - struct hmp_domain *hmp; - if (hmp_cpu_is_fastest(cpu_of(se->cfs_rq->rq))) - return max_se; + if (hmp_cpu_is_fastest(cpu_of(se->cfs_rq->rq))) + return max_se;

- hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq)); - hmp_target_mask = &hmp->cpus; + hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq)); + hmp_target_mask = &hmp->cpus; + if (target_cpu >= 0) { + BUG_ON(!cpumask_test_cpu(target_cpu, hmp_target_mask)); + hmp_target_mask = cpumask_of(target_cpu); } /* The currently running task is not on the runqueue */ se = __pick_first_entity(cfs_rq_of(se));

while (num_tasks && se) { if (entity_is_task(se) && - (se->avg.load_avg_ratio > max_ratio && - hmp_target_mask && - cpumask_intersects(hmp_target_mask, - tsk_cpus_allowed(task_of(se))))) { + se->avg.load_avg_ratio > max_ratio && + cpumask_intersects(hmp_target_mask, + tsk_cpus_allowed(task_of(se)))) { max_se = se; max_ratio = se->avg.load_avg_ratio; } @@ -7153,7 +7154,7 @@ static void hmp_force_up_migration(int this_cpu) } } orig = curr; - curr = hmp_get_heaviest_task(curr, 1); + curr = hmp_get_heaviest_task(curr, -1); p = task_of(curr); if (hmp_up_migration(cpu, &target_cpu, curr)) { cpu_rq(target_cpu)->wake_for_idle_pull = 1; @@ -7250,12 +7251,14 @@ static unsigned int hmp_idle_pull(int this_cpu) } } orig = curr; - curr = hmp_get_heaviest_task(curr, 1); + curr = hmp_get_heaviest_task(curr, this_cpu); /* check if heaviest eligible task on this * CPU is heavier than previous task */ if (hmp_task_eligible_for_up_migration(curr) && - curr->avg.load_avg_ratio > ratio) { + curr->avg.load_avg_ratio > ratio && + cpumask_test_cpu(this_cpu, + tsk_cpus_allowed(task_of(curr)))) { p = task_of(curr); target = rq; ratio = curr->avg.load_avg_ratio;

On Thu, 2014-04-03 at 14:32 +0100, Jon Medhurst (Tixy) wrote: [...]

It looks like updates to hmp_target_mask can lag or lead idle-pull attempts. I don't know if it's worth trying to investigate if there's something that can or should be done to avoid that situation occurring, or just cope with it by adding something like...

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 7984458..38bc70b 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3735,7 +3735,8 @@ static struct sched_entity *hmp_get_heaviest_task( hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq)); hmp_target_mask = &hmp->cpus; if (target_cpu >= 0) {

• BUG_ON(!cpumask_test_cpu(target_cpu, hmp_target_mask));
  

• if (!cpumask_test_cpu(target_cpu, hmp_target_mask))
  

• 	return 0;
  

  hmp_target_mask = cpumask_of(target_cpu); } /* The currently running task is not on the runqueue */

@@ -7255,7 +7256,7 @@ static unsigned int hmp_idle_pull(int this_cpu) /* check if heaviest eligible task on this * CPU is heavier than previous task */

• if (hmp_task_eligible_for_up_migration(curr) &&
  

• if (curr && hmp_task_eligible_for_up_migration(curr) &&
  curr->avg.load_avg_ratio > ratio &&
  cpumask_test_cpu(this_cpu,
  		tsk_cpus_allowed(task_of(curr)))) {
  
  

Assuming I don't hear otherwise soon, I plan on going with the above change. It seems to me that BUG_ON is excessive anyway, for something we can safely and simply deal with.