eas-dev February 2017

eas-dev@lists.linaro.org

10 participants
8 discussions

[PATCH V2] Per Sched domain over utilization

by Thara Gopinath

The current implementation of overutilization, aborts energy aware scheduling if any cpu in the system is over-utilized. This patch introduces over utilization flag per sched domain level instead of a single flag system wide. Load balancing is done at the sched domain where any of the cpu is over utilized. If energy aware scheduling is enabled and no cpu in a sched domain is overuttilized, load balancing is skipped for that sched domain and energy aware scheduling continues at that level. The implementation takes advantage of the shared sched_domain structure that is common across all the sched domains at a level. The new flag introduced is placed in this structure so that all the sched domains the same level share the flag. In case of an overutilized cpu, the flag gets set at level1 sched_domain. The flag at the parent sched_domain level gets set in either of the two following scenarios. 1. There is a misfit task in one of the cpu's in this sched_domain. 2. The total utilization of the domain is greater than the domain capacity The flag is cleared if no cpu in a sched domain is overutilized. This implementation still can have corner scenarios with respect to misfit tasks. For example consider a sched group with n cpus and n+1 70%utilized tasks. Ideally this is a case for load balance to happen in a parent sched domain. But neither the total group utilization is high enough for the load balance to be triggered in the parent domain nor there is a cpu with a single overutilized task so that aload balance is triggered in a parent domain. But again this could be a purely academic sceanrio, as during task wake up these tasks will be placed more appropriately. Signed-off-by: Thara Gopinath <thara.gopinath(a)linaro.org> --- V1->V2: - Removed overutilized flag from sched_group structure. - In case of misfit task, it is ensured that a load balance is triggered in a parent sched domain with assymetric cpu capacities. include/linux/sched.h | 1 + kernel/sched/core.c | 7 ++- kernel/sched/fair.c | 138 +++++++++++++++++++++++++++++++++++++++++--------- kernel/sched/sched.h | 3 -- 4 files changed, 117 insertions(+), 32 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 1c5122e..971842a 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1112,6 +1112,7 @@ struct sched_domain_shared { atomic_t ref; atomic_t nr_busy_cpus; int has_idle_cores; + bool overutilized; }; struct sched_domain { diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 31a466f..e0a8758 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -6659,11 +6659,10 @@ sd_init(struct sched_domain_topology_level *tl, * For all levels sharing cache; connect a sched_domain_shared * instance. */ - if (sd->flags & SD_SHARE_PKG_RESOURCES) { - sd->shared = *per_cpu_ptr(sdd->sds, sd_id); - atomic_inc(&sd->shared->ref); + sd->shared = *per_cpu_ptr(sdd->sds, sd_id); + atomic_inc(&sd->shared->ref); + if (sd->flags & SD_SHARE_PKG_RESOURCES) atomic_set(&sd->shared->nr_busy_cpus, sd_weight); - } sd->private = sdd; diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 489f6d3..9d2bb07 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -4735,6 +4735,30 @@ static inline void hrtick_update(struct rq *rq) static bool cpu_overutilized(int cpu); +static bool +is_sd_overutilized(struct sched_domain *sd) +{ + if (sd) + return sd->shared->overutilized; + else + return false; +} + +static void +set_sd_overutilized(struct sched_domain *sd) +{ + if (sd) + sd->shared->overutilized = true; +} + +static void +clear_sd_overutilized(struct sched_domain *sd) +{ + if (sd) + sd->shared->overutilized = false; +} + + /* * The enqueue_task method is called before nr_running is * increased. Here we update the fair scheduling stats and @@ -4744,6 +4768,7 @@ static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) { struct cfs_rq *cfs_rq; + struct sched_domain *sd; struct sched_entity *se = &p->se; int task_new = !(flags & ENQUEUE_WAKEUP); @@ -4787,9 +4812,12 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) if (!se) { add_nr_running(rq, 1); - if (!task_new && !rq->rd->overutilized && - cpu_overutilized(rq->cpu)) - rq->rd->overutilized = true; + rcu_read_lock(); + sd = rcu_dereference(rq->sd); + if (!task_new && !is_sd_overutilized(sd) && + cpu_overutilized(rq->cpu)) + set_sd_overutilized(sd); + rcu_read_unlock(); } hrtick_update(rq); } @@ -6173,8 +6201,7 @@ static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu) unsigned long max_spare = 0; struct sched_domain *sd; - rcu_read_lock(); - + /* The rcu lock is/should be held in the caller function */ sd = rcu_dereference(per_cpu(sd_ea, prev_cpu)); if (!sd) @@ -6212,8 +6239,6 @@ static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu) } unlock: - rcu_read_unlock(); - if (energy_cpu == prev_cpu && !cpu_overutilized(prev_cpu)) return prev_cpu; @@ -6247,10 +6272,16 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_f && cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); } - if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized)) - return select_energy_cpu_brute(p, prev_cpu); - rcu_read_lock(); + sd = rcu_dereference(cpu_rq(prev_cpu)->sd); + if (energy_aware() && + !is_sd_overutilized(sd)) { + new_cpu = select_energy_cpu_brute(p, prev_cpu); + goto unlock; + } + + sd = NULL; + for_each_domain(cpu, tmp) { if (!(tmp->flags & SD_LOAD_BALANCE)) break; @@ -6315,6 +6346,8 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_f } /* while loop will break here if sd == NULL */ } + +unlock: rcu_read_unlock(); return new_cpu; @@ -7366,6 +7399,7 @@ struct sd_lb_stats { struct sched_group *local; /* Local group in this sd */ unsigned long total_load; /* Total load of all groups in sd */ unsigned long total_capacity; /* Total capacity of all groups in sd */ + unsigned long total_util; /* Total util of all groups in sd */ unsigned long avg_load; /* Average load across all groups in sd */ struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ @@ -7385,6 +7419,7 @@ static inline void init_sd_lb_stats(struct sd_lb_stats *sds) .local = NULL, .total_load = 0UL, .total_capacity = 0UL, + .total_util = 0UL, .busiest_stat = { .avg_load = 0UL, .sum_nr_running = 0, @@ -7664,7 +7699,7 @@ group_type group_classify(struct sched_group *group, static inline void update_sg_lb_stats(struct lb_env *env, struct sched_group *group, int load_idx, int local_group, struct sg_lb_stats *sgs, - bool *overload, bool *overutilized) + bool *overload, bool *overutilized, bool *misfit_task) { unsigned long load; int i, nr_running; @@ -7699,8 +7734,16 @@ static inline void update_sg_lb_stats(struct lb_env *env, if (!nr_running && idle_cpu(i)) sgs->idle_cpus++; - if (cpu_overutilized(i)) + if (cpu_overutilized(i)) { *overutilized = true; + /* + * If the cpu is overutilized and if there is only one + * current task in cfs runqueue, it is potentially a misfit + * task. + */ + if (rq->cfs.h_nr_running == 1) + *misfit_task = true; + } } /* Adjust by relative CPU capacity of the group */ @@ -7825,11 +7868,11 @@ static inline enum fbq_type fbq_classify_rq(struct rq *rq) */ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) { - struct sched_domain *child = env->sd->child; + struct sched_domain *child = env->sd->child, *sd; struct sched_group *sg = env->sd->groups; struct sg_lb_stats tmp_sgs; int load_idx, prefer_sibling = 0; - bool overload = false, overutilized = false; + bool overload = false, overutilized = false, misfit_task = false; if (child && child->flags & SD_PREFER_SIBLING) prefer_sibling = 1; @@ -7851,7 +7894,8 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd } update_sg_lb_stats(env, sg, load_idx, local_group, sgs, - &overload, &overutilized); + &overload, &overutilized, + &misfit_task); if (local_group) goto next_group; @@ -7882,6 +7926,7 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd /* Now, start updating sd_lb_stats */ sds->total_load += sgs->group_load; sds->total_capacity += sgs->group_capacity; + sds->total_util += sgs->group_util; sg = sg->next; } while (sg != env->sd->groups); @@ -7895,14 +7940,45 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd /* update overload indicator if we are at root domain */ if (env->dst_rq->rd->overload != overload) env->dst_rq->rd->overload = overload; + } - /* Update over-utilization (tipping point, U >= 0) indicator */ - if (env->dst_rq->rd->overutilized != overutilized) - env->dst_rq->rd->overutilized = overutilized; - } else { - if (!env->dst_rq->rd->overutilized && overutilized) - env->dst_rq->rd->overutilized = true; + if (overutilized) + set_sd_overutilized(env->sd); + else + clear_sd_overutilized(env->sd); + + /* + * If there is a misfit task in one cpu in this sched_domain + * it is likely that the imbalance cannot be sorted out among + * the cpu's in this sched_domain. In this case set the + * overutilized flag at the parent sched_domain. + */ + if (misfit_task) { + + sd = env->sd->parent; + + /* + * In case of a misfit task, load balance at the parent + * sched domain level will make sense only if the the cpus + * have a different capacity. If cpus at a domain level have + * the same capacity, the misfit task cannot be well + * accomodated in any of the cpus and there in no point in + * trying a load balance at this level + */ + while (sd) { + if (sd->flags & SD_ASYM_CPUCAPACITY) { + set_sd_overutilized(sd); + break; + } + sd = sd->parent; + } } + + /* If the domain util is greater that domain capacity, load balancing + * needs to be done at the next sched domain level as well + */ + if (sds->total_capacity * 1024 < sds->total_util * capacity_margin) + set_sd_overutilized(env->sd->parent); } /** @@ -8122,8 +8198,10 @@ static struct sched_group *find_busiest_group(struct lb_env *env) */ update_sd_lb_stats(env, &sds); - if (energy_aware() && !env->dst_rq->rd->overutilized) - goto out_balanced; + if (energy_aware()) { + if (!is_sd_overutilized(env->sd)) + goto out_balanced; + } local = &sds.local_stat; busiest = &sds.busiest_stat; @@ -8981,6 +9059,11 @@ static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) rcu_read_lock(); for_each_domain(cpu, sd) { + if (energy_aware()) { + if (!is_sd_overutilized(sd)) + continue; + } + /* * Decay the newidle max times here because this is a regular * visit to all the domains. Decay ~1% per second. @@ -9280,6 +9363,7 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) { struct cfs_rq *cfs_rq; struct sched_entity *se = &curr->se; + struct sched_domain *sd; for_each_sched_entity(se) { cfs_rq = cfs_rq_of(se); @@ -9289,8 +9373,12 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) if (static_branch_unlikely(&sched_numa_balancing)) task_tick_numa(rq, curr); - if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) - rq->rd->overutilized = true; + rcu_read_lock(); + sd = rcu_dereference(rq->sd); + if (!is_sd_overutilized(sd) && + cpu_overutilized(task_cpu(curr))) + set_sd_overutilized(sd); + rcu_read_unlock(); } /* diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h index fa98ab3..b24cefa 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -563,9 +563,6 @@ struct root_domain { /* Indicate more than one runnable task for any CPU */ bool overload; - /* Indicate one or more cpus over-utilized (tipping point) */ - bool overutilized; - /* * The bit corresponding to a CPU gets set here if such CPU has more * than one runnable -deadline task (as it is below for RT tasks). -- 2.1.4

8 years, 11 months

[PATCH V2] cpufreq: schedutil: Redefine the rate_limit_us tunable

by Viresh Kumar

The rate_limit_us tunable is intended to reduce the possible overhead from running the schedutil governor. However, that overhead can be divided into two separate parts: the governor computations and the invocation of the scaling driver to set the CPU frequency. The latter is where the real overhead comes from. The former is much less expensive in terms of execution time and running it every time the governor callback is invoked by the scheduler, after rate_limit_us interval has passed since the last frequency update, would not be a problem. For this reason, redefine the rate_limit_us tunable so that it means the minimum time that has to pass between two consecutive invocations of the scaling driver by the schedutil governor (to set the CPU frequency). Signed-off-by: Viresh Kumar <viresh.kumar(a)linaro.org> --- V1->V2: Update $subject and commit log (Rafael) kernel/sched/cpufreq_schedutil.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c index fd4659313640..306d97e7b57c 100644 --- a/kernel/sched/cpufreq_schedutil.c +++ b/kernel/sched/cpufreq_schedutil.c @@ -92,14 +92,13 @@ static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time, { struct cpufreq_policy *policy = sg_policy->policy; - sg_policy->last_freq_update_time = time; - if (policy->fast_switch_enabled) { if (sg_policy->next_freq == next_freq) { trace_cpu_frequency(policy->cur, smp_processor_id()); return; } sg_policy->next_freq = next_freq; + sg_policy->last_freq_update_time = time; next_freq = cpufreq_driver_fast_switch(policy, next_freq); if (next_freq == CPUFREQ_ENTRY_INVALID) return; @@ -108,6 +107,7 @@ static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time, trace_cpu_frequency(next_freq, smp_processor_id()); } else if (sg_policy->next_freq != next_freq) { sg_policy->next_freq = next_freq; + sg_policy->last_freq_update_time = time; sg_policy->work_in_progress = true; irq_work_queue(&sg_policy->irq_work); } -- 2.7.1.410.g6faf27b

8 years, 11 months

[viresh.kumar@linaro.org: [PATCH] cpufreq: schedutil: govern how frequently we change frequency with rate_limit]

by Viresh Kumar

Sorry that I forgot to cc eas-dev list for this patch. ----- Forwarded message from Viresh Kumar <viresh.kumar(a)linaro.org> ----- Date: Wed, 15 Feb 2017 22:45:47 +0530 From: Viresh Kumar <viresh.kumar(a)linaro.org> To: Rafael Wysocki <rjw(a)rjwysocki.net>, Ingo Molnar <mingo(a)redhat.com>, Peter Zijlstra <peterz(a)infradead.org> Cc: linaro-kernel(a)lists.linaro.org, linux-pm(a)vger.kernel.org, linux-kernel(a)vger.kernel.org, Vincent Guittot <vincent.guittot(a)linaro.org>, Viresh Kumar <viresh.kumar(a)linaro.org> Subject: [PATCH] cpufreq: schedutil: govern how frequently we change frequency with rate_limit X-Mailer: git-send-email 2.7.1.410.g6faf27b For an ideal system (where frequency change doesn't incur any penalty) we would like to change the frequency as soon as the load changes for a CPU. But the systems we have to work with are far from ideal and it takes time to change the frequency of a CPU. For many ARM platforms specially, it is at least 1 ms. In order to not spend too much time changing frequency, we have earlier introduced a sysfs controlled tunable for the schedutil governor: rate_limit_us. Currently, rate_limit_us controls how frequently we reevaluate frequency for a set of CPUs controlled by a cpufreq policy. But that may not be the ideal behavior we want. Consider for example the following scenario. The rate_limit_us tunable is set to 10 ms. The CPU has a constant load X and that requires the frequency to be set to Y. The schedutil governor changes the frequency to Y, updates last_freq_update_time and we wait for 10 ms to reevaluate the frequency again. After 10 ms, the schedutil governor reevaluates the load and finds it to be the same. And so it doesn't update the frequency, but updates last_freq_update_time before returning. Right after this point, the scheduler puts more load on the CPU and the CPU needs to go to a higher frequency Z. Because last_freq_update_time was updated just now, the schedutil governor waits for additional 10ms before reevaluating the load again. Normally, the time it takes to reevaluate the frequency is negligible compared to the time it takes to change the frequency. And considering that in the above scenario, as we haven't updated the frequency for over 10ms, we should have changed the frequency as soon as the load changed. This patch changes the way rate_limit_us is used, i.e. It now governs "How frequently we change the frequency" instead of "How frequently we reevaluate the frequency". One may think that this change may have increased the number of times we reevaluate the frequency after a period of rate_limit_us has expired since the last change, if the load isn't changing. But that is protected by the scheduler as normally it doesn't call into the schedutil governor before 1 ms (Hint: "decayed" in update_cfs_rq_load_avg()) since the last call. Tests were performed with this patch on a Dual cluster (same frequency domain), octa-core ARM64 platform (Hikey). Hackbench (Debian) and Vellamo/Galleryfling (Android) didn't had much difference in performance w/ or w/o this patch. Its difficult to create a test case (tried rt-app as well) where this patch will show a lot of improvements as the target of this patch is a real corner case. I.e. Current load is X (resulting in freq change), load after rate_limit_us is also X, but right after that load becomes Y. Undoubtedly this patch would improve the responsiveness in such cases. Signed-off-by: Viresh Kumar <viresh.kumar(a)linaro.org> --- kernel/sched/cpufreq_schedutil.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c index fd4659313640..306d97e7b57c 100644 --- a/kernel/sched/cpufreq_schedutil.c +++ b/kernel/sched/cpufreq_schedutil.c @@ -92,14 +92,13 @@ static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time, { struct cpufreq_policy *policy = sg_policy->policy; - sg_policy->last_freq_update_time = time; - if (policy->fast_switch_enabled) { if (sg_policy->next_freq == next_freq) { trace_cpu_frequency(policy->cur, smp_processor_id()); return; } sg_policy->next_freq = next_freq; + sg_policy->last_freq_update_time = time; next_freq = cpufreq_driver_fast_switch(policy, next_freq); if (next_freq == CPUFREQ_ENTRY_INVALID) return; @@ -108,6 +107,7 @@ static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time, trace_cpu_frequency(next_freq, smp_processor_id()); } else if (sg_policy->next_freq != next_freq) { sg_policy->next_freq = next_freq; + sg_policy->last_freq_update_time = time; sg_policy->work_in_progress = true; irq_work_queue(&sg_policy->irq_work); } -- 2.7.1.410.g6faf27b ----- End forwarded message ----- -- viresh

8 years, 11 months

[PATCH] Per Sched domain over utilization

by Thara Gopinath

The current implementation of overutilization, aborts energy aware scheduling if any cpu in the system is over-utilized. This patch introduces over utilization flag per sched domain level instead of a single flag system wide. Load balancing is done at the sched domain where any of the cpu is over utilized. If energy aware scheduling is enabled and no cpu in a sched domain is overuttilized, load balancing is skipped for that sched domain and energy aware scheduling continues at that level. The implementation takes advantage of the shared sched_domain structure that is common across all the sched domains at a level. The new flag introduced is placed in this structure so that all the sched domains the same level share the flag. In case of an overutilized cpu, the flag gets set at level1 sched_domain. The flag at the parent sched_domain level gets set in either of the two following scenarios. 1. There is a misfit task in one of the cpu's in this sched_domain. 2. The total utilization of the domain is greater than the domain capacity The flag is cleared if no cpu in a sched domain is overutilized. Signed-off-by: Thara Gopinath <thara.gopinath(a)linaro.org> --- include/linux/sched.h | 1 + kernel/sched/core.c | 7 ++- kernel/sched/fair.c | 120 +++++++++++++++++++++++++++++++++++++++----------- 3 files changed, 99 insertions(+), 29 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 1c5122e..971842a 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1112,6 +1112,7 @@ struct sched_domain_shared { atomic_t ref; atomic_t nr_busy_cpus; int has_idle_cores; + bool overutilized; }; struct sched_domain { diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 31a466f..e0a8758 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -6659,11 +6659,10 @@ sd_init(struct sched_domain_topology_level *tl, * For all levels sharing cache; connect a sched_domain_shared * instance. */ - if (sd->flags & SD_SHARE_PKG_RESOURCES) { - sd->shared = *per_cpu_ptr(sdd->sds, sd_id); - atomic_inc(&sd->shared->ref); + sd->shared = *per_cpu_ptr(sdd->sds, sd_id); + atomic_inc(&sd->shared->ref); + if (sd->flags & SD_SHARE_PKG_RESOURCES) atomic_set(&sd->shared->nr_busy_cpus, sd_weight); - } sd->private = sdd; diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 489f6d3..485f597 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -4735,6 +4735,30 @@ static inline void hrtick_update(struct rq *rq) static bool cpu_overutilized(int cpu); +static bool +is_sd_overutilized(struct sched_domain *sd) +{ + if (sd) + return sd->shared->overutilized; + else + return false; +} + +static void +set_sd_overutilized(struct sched_domain *sd) +{ + if (sd) + sd->shared->overutilized = true; +} + +static void +clear_sd_overutilized(struct sched_domain *sd) +{ + if (sd) + sd->shared->overutilized = false; +} + + /* * The enqueue_task method is called before nr_running is * increased. Here we update the fair scheduling stats and @@ -4744,6 +4768,7 @@ static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) { struct cfs_rq *cfs_rq; + struct sched_domain *sd; struct sched_entity *se = &p->se; int task_new = !(flags & ENQUEUE_WAKEUP); @@ -4787,9 +4812,12 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) if (!se) { add_nr_running(rq, 1); - if (!task_new && !rq->rd->overutilized && - cpu_overutilized(rq->cpu)) - rq->rd->overutilized = true; + rcu_read_lock(); + sd = rcu_dereference(rq->sd); + if (!task_new && !is_sd_overutilized(sd) && + cpu_overutilized(rq->cpu)) + set_sd_overutilized(sd); + rcu_read_unlock(); } hrtick_update(rq); } @@ -6173,8 +6201,7 @@ static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu) unsigned long max_spare = 0; struct sched_domain *sd; - rcu_read_lock(); - + /* The rcu lock is/should be held in the caller function */ sd = rcu_dereference(per_cpu(sd_ea, prev_cpu)); if (!sd) @@ -6212,8 +6239,6 @@ static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu) } unlock: - rcu_read_unlock(); - if (energy_cpu == prev_cpu && !cpu_overutilized(prev_cpu)) return prev_cpu; @@ -6247,10 +6272,16 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_f && cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); } - if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized)) - return select_energy_cpu_brute(p, prev_cpu); - rcu_read_lock(); + sd = rcu_dereference(cpu_rq(prev_cpu)->sd); + if (energy_aware() && + !is_sd_overutilized(sd)) { + new_cpu = select_energy_cpu_brute(p, prev_cpu); + goto unlock; + } + + sd = NULL; + for_each_domain(cpu, tmp) { if (!(tmp->flags & SD_LOAD_BALANCE)) break; @@ -6315,6 +6346,8 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_f } /* while loop will break here if sd == NULL */ } + +unlock: rcu_read_unlock(); return new_cpu; @@ -7366,6 +7399,7 @@ struct sd_lb_stats { struct sched_group *local; /* Local group in this sd */ unsigned long total_load; /* Total load of all groups in sd */ unsigned long total_capacity; /* Total capacity of all groups in sd */ + unsigned long total_util; /* Total util of all groups in sd */ unsigned long avg_load; /* Average load across all groups in sd */ struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ @@ -7385,6 +7419,7 @@ static inline void init_sd_lb_stats(struct sd_lb_stats *sds) .local = NULL, .total_load = 0UL, .total_capacity = 0UL, + .total_util = 0UL, .busiest_stat = { .avg_load = 0UL, .sum_nr_running = 0, @@ -7664,7 +7699,7 @@ group_type group_classify(struct sched_group *group, static inline void update_sg_lb_stats(struct lb_env *env, struct sched_group *group, int load_idx, int local_group, struct sg_lb_stats *sgs, - bool *overload, bool *overutilized) + bool *overload, bool *overutilized, bool *misfit_task) { unsigned long load; int i, nr_running; @@ -7699,8 +7734,16 @@ static inline void update_sg_lb_stats(struct lb_env *env, if (!nr_running && idle_cpu(i)) sgs->idle_cpus++; - if (cpu_overutilized(i)) + if (cpu_overutilized(i)) { *overutilized = true; + /* + * If the cpu is overutilized and if there is only one + * current task in cfs runqueue, it is potentially a misfit + * task. + */ + if (rq->cfs.h_nr_running == 1) + *misfit_task = true; + } } /* Adjust by relative CPU capacity of the group */ @@ -7829,7 +7872,7 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd struct sched_group *sg = env->sd->groups; struct sg_lb_stats tmp_sgs; int load_idx, prefer_sibling = 0; - bool overload = false, overutilized = false; + bool overload = false, overutilized = false, misfit_task = false; if (child && child->flags & SD_PREFER_SIBLING) prefer_sibling = 1; @@ -7851,7 +7894,8 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd } update_sg_lb_stats(env, sg, load_idx, local_group, sgs, - &overload, &overutilized); + &overload, &overutilized, + &misfit_task); if (local_group) goto next_group; @@ -7882,6 +7926,7 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd /* Now, start updating sd_lb_stats */ sds->total_load += sgs->group_load; sds->total_capacity += sgs->group_capacity; + sds->total_util += sgs->group_util; sg = sg->next; } while (sg != env->sd->groups); @@ -7895,14 +7940,27 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd /* update overload indicator if we are at root domain */ if (env->dst_rq->rd->overload != overload) env->dst_rq->rd->overload = overload; - - /* Update over-utilization (tipping point, U >= 0) indicator */ - if (env->dst_rq->rd->overutilized != overutilized) - env->dst_rq->rd->overutilized = overutilized; - } else { - if (!env->dst_rq->rd->overutilized && overutilized) - env->dst_rq->rd->overutilized = true; } + + if (overutilized) + set_sd_overutilized(env->sd); + else + clear_sd_overutilized(env->sd); + + /* + * If there is a misfit task in one cpu in this sched_domain + * it is likely that the imbalance cannot be sorted out among + * the cpu's in this sched_domain. In this case set the + * overutilized flag at the parent sched_domain. + */ + if (misfit_task) + set_sd_overutilized(env->sd->parent); + + /* If the domain util is greater that domain capacity, load balancing + * needs to be done at the next sched domain level as well + */ + if (sds->total_capacity * 1024 < sds->total_util * capacity_margin) + set_sd_overutilized(env->sd->parent); } /** @@ -8122,8 +8180,10 @@ static struct sched_group *find_busiest_group(struct lb_env *env) */ update_sd_lb_stats(env, &sds); - if (energy_aware() && !env->dst_rq->rd->overutilized) - goto out_balanced; + if (energy_aware()) { + if (!is_sd_overutilized(env->sd)) + goto out_balanced; + } local = &sds.local_stat; busiest = &sds.busiest_stat; @@ -8981,6 +9041,11 @@ static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) rcu_read_lock(); for_each_domain(cpu, sd) { + if (energy_aware()) { + if (!is_sd_overutilized(sd)) + continue; + } + /* * Decay the newidle max times here because this is a regular * visit to all the domains. Decay ~1% per second. @@ -9280,6 +9345,7 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) { struct cfs_rq *cfs_rq; struct sched_entity *se = &curr->se; + struct sched_domain *sd; for_each_sched_entity(se) { cfs_rq = cfs_rq_of(se); @@ -9289,8 +9355,12 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) if (static_branch_unlikely(&sched_numa_balancing)) task_tick_numa(rq, curr); - if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) - rq->rd->overutilized = true; + rcu_read_lock(); + sd = rcu_dereference(rq->sd); + if (!is_sd_overutilized(sd) && + cpu_overutilized(task_cpu(curr))) + set_sd_overutilized(sd); + rcu_read_unlock(); } /* -- 2.1.4

8 years, 11 months

[ANNOUNCE][CFP] Power Management and Scheduling in the Linux Kernel (OSPM-summit)

by Juri Lelli

Power Management and Scheduling in the Linux Kernel (OSPM-summit) April 3-4, 2017 Scuola Superiore Sant'Anna (SSSA) Pisa, Italy http://retis.sssup.it/ospm-summit/ --- .:: FOCUS Power management and scheduling techniques to reduce energy consumption while meeting performance and latency requirements are receiving considerable attention from the Linux Kernel development community. The Power Management and Scheduling in the Linux Kernel (OSPM-summit) summit aims at fostering further interest and discussion to happen. .:: FORMAT The summit is organized to cover two days of discussions and talks. First day is mainly focused on discussion and hacking sessions about topics/patches that are already under review in the Linux kernel mailing lists and to debate and plan development tasks for more forward looking work items centred around power management in the Linux kernel. The list of topics includes (but it is not limited to): * Energy Aware Scheduling: next steps and energy model expression; * SCHED_DEADLINE reclaiming of unused bandwidth, coupling with schedutil cpufreq governor and group scheduling support; * fix the load metric exposed to cpuidle; * IRQ prediction; * ACPI power management: kernel/firmware bindings and development model; Second day instead welcomes presentations from both end users and developers on topics about Power management and scheduling in Linux covering, but not limited to: * Power management techniques * Real-time and non real-time scheduling techniques * Energy awareness * Mobile/Server power management real-world use cases (successes and failures) * Power management and scheduling tooling (tracing, configuration, integration testing, etc.) Presentations can cover recently developed technologies, ongoing work and new ideas. Please understand that this workshop is not intended for presenting sales and marketing pitches. .:: ATTENDING Attending the OSPM-summit is free of charge, but registration to the event is mandatory. The event can allow a maximum of 50 people (so, be sure to register early!). To register send an email to ospm-registration(a)retis.sssup.it. While it is not strictly required to submit a topic/presentation, registrations with a topic/presentation proposal will take precedence. .:: VENUE The workshop will take place at ReTiS Lab*, Scuola Superiore Sant'Anna, Pisa, Italy. Pisa is a small town, walking distance from the city center to the venue is 20 minutes, walking distance from the airport to the city center is 30 minutes. More details are available from the summit web page: http://retis.sssup.it/ospm-summit/ * https://goo.gl/maps/2pPXG2v7Lfp .:: SUBMIT A TOPIC/PRESENTATION To submit a topic/presentation send an email to ospm-registration(a)retis.sssup.it specifying: subject - [TOPIC] or [PRESENTATION] - short title body - first name, family name - abstract/topic of interest - affiliation (if any) - short biography - expected duration (only for topics, presentations get 30min slots) Deadline for submitting topics/presentations is 26th of February 2017. Notifications for accepted topics/presentations will be sent out on 5th of March 2017. .:: ORGANIZERS (in alphabetical order) Luca Abeni (SSSA) Patrick Bellasi (ARM) Tommaso Cucinotta (SSSA) Dietmar Eggemann (ARM) Sudeep Holla (ARM) Juri Lelli (ARM) Lorenzo Pieralisi (ARM) Morten Rasmussen (ARM)

9 years

[PATCH 0/6] Improvement load balance for misfit task

by Leo Yan

This patch series is to improve load balance with more proper behaviour for misfit task. Current code introduces type 'group_misfit_task' to indicate one schedule group has misfit task, but before the misfit task can be really migrated onto higher capacity CPU there still have some barriers we need clear up. The first patch is to correct task_fits_max() so it can properly filter out misfit task on low capacity CPU. If without this patch, in system it's possible this function can always return true so the 'misfit' task mechanism will totally not be triggered. The second patch is to fix function group_smaller_cpu_capacity(), so we can make sure the schedule group with type 'group_misfit_task' will not wrongly be roll back to type 'group_other'. This will let all misfit related info be abondoned. The third patch is to fix nr_running accounting, if without this patch the scheduler will wronly consider the destination CPU has running task and skip migrate task on it. This patch is to give correct info like the destination CPU has no running task on it when the CPU is going into idle state, so should migrate misfit task by utilizing this time balance. The forth patch is a temperary patch if we have not backported Vincent's patches "sched: reflect sched_entity move into task_group's load" [1], If without this patch series, it's possible that the CPU is not overutilized but the CPU has one misfit task has been enqueued on it. So we set sgs->group_misfit_task by checking rq->misfit_task but not rely on cpu is overutilized or not. The fifth patch is to select busiest rq if the rq has misfit task, we let this kind rq has higher priority than the rq with highest weighted load. This criteria is only enabled for energy aware scheduling. The sixth patch is to aggressively kick active load balance for misfit task, so it has quite high chance for higher capacity CPU to immediately pull misfit task on it. [1] https://lkml.org/lkml/2016/10/17/223 Leo Yan (6): sched/fair: correct task_fits_max() for misfit task sched/fair: fix for group_smaller_cpu_capacity() sched/fair: fix nr_running accounting for new idle CPU sched/fair: fix to set sgs->group_misfit_task sched/fair: select busiest rq with misfit task sched/fair: kick active load balance for misfit task kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++++++++++++------------ 1 file changed, 46 insertions(+), 13 deletions(-) -- 2.7.4

9 years

eas on non-arm platforms

by Darren Greene

Hello all, I have an x86 based platform which is running android. I wanted to play around with the EAS patches to see if it would improve power numbers on it. I had a few basic questions regarding this: 1) Can EAS be used with x86 based platforms ? I see some arm/arm64 energy model related patches in the eas integration tree (git://linux-arm.org/linux-power.git). However, there aren't any x86 specific changes present. Is that because no x86 specific changes are required or just that it is untested there ? 2) Is it expected that EAS would show significant power savings on SMP systems or just on HMP systems ? 3) Would there be any cpufreq/cpuidle integration be required for x86 specifically ? If so, would I need to base them on the arm stuff or is there any other reference code. 4) Are there other in-flight patches that need to be applied over the patches in the eas integration tree for best results ? If indeed the EAS patches can be used on x86. then I would be interested in integrating and providing results on my platform. Please guide. Regards, Darren

9 years

EAS 1.2 Pre-alpha Experimental Branch now available for Pixel/Pixel XL

by Chris Redpath

Hello, I'm pleased to announce that we have pushed a very early version of some of the key features we intend to make available as EAS 1.2 this year to Google's msm repository ( https://android.googlesource.com/kernel/msm.git/ ) as android-msm-marlin-3.18-nougat-mr1-eas-experimental. EAS 1.2 is intended to be the next iteration of EAS for AOSP, including improvements to the wakeup path to better support big.LITTLE and trialling other upstream scheduler enhancements such as schedutil along with some important load/util tracking enhancements to PELT. Although EAS 1.2 will be primarily focused on a 4.4-based kernel, we are making this experimental branch available on the 3.18-based Pixel kernel (marlin_defconfig) in order that we have a readily-available real platform with an optimised userspace for experimentation. There are some differences in the scheduler task wake-up path between this release and that shipping in the Pixel kernel which should be taken into account when using this kernel. The most visible change in the wake-up path is the removal of the is_big_little sysctl. Wake-up now uses a single cpu selection algorithm (the same one used previously for !isBigLittle) but modified to remove the assumption that the highest capacity cpus have the highest logical cpu number. We now allow cpu topology independent selection of max capacity cpus for tasks which belong to a schedtune group which has some boost applied irrespective of the cpu numbering. This changes the iteration order of cpus when looking for a place to run these tasks from [3,2],[1,0] to [2,3], [0,1]. This has an impact on runtime configuration. Not making a change to this configuration is likely to have a small impact for lightly-loaded systems where there will usually be two idle high-capacity cpus, but we should anyway match cpuset configuration to the selection ordering to restore the expectations used when tuning. In Pixel, cpusets are arranged such that one of the highest capacity cpus is available only to tasks belonging to the ‘top-app’ cpuset. In combination with the cpu iteration order used for schedtune boosted tasks, we hope to find an empty cpu more often for these tasks to wake on. As a result of the changed iteration order, the top-app should now be set to the lowest numbered high capacity cpu (in this case #2 for Pixel). The impact of this is likely to be small for most light use cases if not changed. This is done in the initrc: The usual group setup for Pixel is in init.sailfish.rc - the part which configures the CPUSets for the tuning groups is normally as follows: on property:sys.boot_completed=1 write /proc/sys/kernel/sched_boost 0 # update cpusets now that boot is complete and we want better load balancing write /dev/cpuset/top-app/cpus 0-3 write /dev/cpuset/foreground/boost/cpus 0-2 write /dev/cpuset/foreground/cpus 0-2 write /dev/cpuset/background/cpus 0 write /dev/cpuset/system-background/cpus 0-2 As we wish to make cpu 2 the one which is only available for tasks in the top-app group, we should exclude cpu 2 from the other groups. on property:sys.boot_completed=1 write /proc/sys/kernel/sched_boost 0 # update cpusets now that boot is complete and we want better load balancing write /dev/cpuset/top-app/cpus 0-3 write /dev/cpuset/foreground/boost/cpus 0-1,3 write /dev/cpuset/foreground/cpus 0-1,3 write /dev/cpuset/background/cpus 0 write /dev/cpuset/system-background/cpus 0-1,3 We normally do this at run time in a root shell rather than modifying the init scripts. The schedutil governor is present but not selected as the default cpufreq governor. It is important to note that there is a slight difference in the meaning of the up & down frequency select throttling for the 'sched' governor (sched-dvfs) and 'schedutil'. The 'sched' governor considers time to be measured since the last *frequency change* whilst the 'schedutil' governor considers the time to be measured since the last *utilisation request*. This means that we need to shorten the throttle periods used for schedutil when comparing it to sched-dvfs to avoid staying at the maximum frequency for long periods in UI-driven workloads. We have been experimenting with up_rate_limit_usec set to 500 and down_rate_limit_usec set to 2000 or 5000 which appears to give results comparable with those of the 'sched' governor. The branch is based upon the mr1 kernel release, and contains the patches shown at the end of this mail. They are comprised of 6 main areas of functionality. * ec114ba...d2238c2 and 8646350...35ea67a patches to reduce the delta between the msm kernel and the common kernel * b055eba...d2e2970 introduce a backport of the upstream schedutil governor (but it is not the default governor in marlin_defconfig) * 7f7e79e...14531d4e bring the energy-aware-scheduling calculations into line with our mainline-focused implementation and backport capacity-based-scheduling to 3.18 * b75b728...407d2a7 integrate the current EAS 1.1 wakeup path with the mainline-focused wakeup path and introduce a way to provide a common algorithm implementing the alternate CPU search algorithm for schedtune boosted tasks * f966249...1ad6d08 Backport some important upstream CFS fixes to 3.18. This fixes some critical group accounting issues which had a negative impact on the suitability of PELT utilisation signals for Android * 6ae4707 Allows EAS to continue to calculate energy for systems which end up with a single CPU in a sched domain Best Regards, Chris Amit Pundir (3): sched/walt: use do_div instead of division operator ANDROID: sched/walt: fix build failure if FAIR_GROUP_SCHED=n Revert "cgroup: Fix issues in allow_attach callback" Brendan Jackman (2): DEBUG: sched/fair: Fix missing sched_load_avg_cpu events DEBUG: sched/fair: Fix sched_load_avg_cpu events for task_groups Chris Redpath (17): Revert "WIP: UTIL_EST: use estimated utilization on load balancing paths" Revert "WIP: UTIL_EST: use estimated utilization on energy aware wakeup path" Revert "WIP: UTIL_EST: sched/fair: use estimated utilization to drive CPUFreq" Revert "WIP: UTIL_EST: switch to usage of tasks's estimated utilization" sched: revert UTIL_EST usage from commit 6bf72ca7f1 Revert "WIP: UTIL_EST: sched/{core,fair}: add support to use estimated utilization" Revert "WIP: UTIL_EST: sched/fair: add support for estimated utilization" sched/fair: missing parts of 'optimize idle cpu selection for boosted tasks' sched/fair: Fix uninitialised variable in idle_balance Revert: UTIL_EST code from 'fix set_cfs_cpu_capacity when WALT is in use" Unify whitespace layout with android-3.18 schedtune: Guarding against compile errors sched/walt: Drop arch-specific timer access Revert "DEBUG: UTIL_EST: sched: update tracepoint to report estimated CPU utilzation" sched: This kernel expects sched_cfs_boost to be signed schedutil: Fix linkage of schedutil and walt config: Update marlin_defconfig to include schedutil governor Dietmar Eggemann (20): Revert "WIP: sched: Consider spare cpu capacity at task wake-up" Partial Revert: "WIP: sched: Add cpu capacity awareness to wakeup balancing" Experimental! arm64: Set SD_SHARE_CAP_STATES sched_domain flag on DIE level Experimental!: sched/fair: Do not force want_affine eq. true if EAS is enabled Experimental!: sched/fair: Decommission energy_aware_wake_cpu() Fixup!: sched/fair.c: Set SchedTune specific struct energy_env.task Experimental!: EAS: sched/fair: Re-integrate 'honor sync wakeups' into wakeup path Experimental!: sched/fair: Code !is_big_little path into select_energy_cpu_brute() Experimental!: sched: Remove sysctl_sched_is_big_little sched/core: Remove remnants of commit fd5c98da1a42 Experimental!: sched/core: Add first cpu w/ max/min orig capacity to root domain Experimental!: sched/fair: Change cpu iteration order in find_best_target() sched/fair: Simplify backup_capacity handling in find_best_target() Fixup!: sched/fair: Simplify target_util handling in find_best_target() Fixup!: sched/fair: Simplify idle_idx handling in find_best_target() Fixup!: sched/fair: Refactor min_util, new_util in find_best_target() Fixup!: sched/fair: Simplify idle_idx handling in select_idle_sibling() Fixup!: Return first idle cpu for prefer_idle task immediately Fixup!: sched/fair: No need to 'and' current cpu w/ online mask in wakeup sched: EAS & 'single cpu per cluster'/cpu hotplug interoperability Dmitry Shmidt (1): sched: Fix sysctl_sched_cfs_boost type to be int Juri Lelli (3): sched/cpufreq: make schedutil use WALT signal trace/sched: add rq utilization signal for WALT sched/walt: kill {min,max}_capacity Ke Wang (1): sched: tune: Fix lacking spinlock initialization Morten Rasmussen (15): sched/core: Fix power to capacity renaming in comment sched/fair: Make the use of prev_cpu consistent in the wakeup path sched/fair: Optimize find_idlest_cpu() when there is no choice sched/core: Remove unnecessary NULL-pointer check sched/core: Introduce SD_ASYM_CPUCAPACITY sched_domain topology flag sched/core: Pass child domain into sd_init() sched/core: Enable SD_BALANCE_WAKE for asymmetric capacity systems sched/fair: Let asymmetric CPU configurations balance at wake-up sched/fair: Compute task/cpu utilization at wake-up correctly sched/fair: Consider spare capacity in find_idlest_group() sched/fair: Add per-CPU min capacity to sched_group_capacity sched/fair: Avoid pulling tasks from non-overloaded higher capacity groups sched/fair: Fix incorrect comment for capacity_margin Experimental!: sched/fair: Add energy_diff dead-zone margin Experimental!: sched/fair: Energy-aware wake-up task placement Patrick Bellasi (3): FIXUP: sched/tune: update accouting before CPU capacity FIX: sched/tune: move schedtune_nornalize_energy into fair.c sched/tune: backport 'fix accounting for runnable tasks' Peter Zijlstra (Intel) (3): sched/fair: Apply more PELT fixes sched/fair: Improve PELT stuff some more sched/fair: Fix effective_load() to consistently use smoothed load Petr Mladek (1): kthread: allow to cancel kthread work Srinath Sridharan (1): eas/sched/fair: Fixing comments in find_best_target. Steve Muckle (5): sched/cpufreq: fix tunables for schedfreq governor sched: backport cpufreq hooks from 4.9-rc4 sched: backport schedutil governor from 4.9-rc4 sched: cpufreq: use rt_avg as estimate of required RT CPU capacity cpufreq: schedutil: add up/down frequency transition rate limits Vincent Guittot (6): sched: factorize attach entity sched: factorize PELT update sched: fix hierarchical order in rq->leaf_cfs_rq_list sched: propagate load during synchronous attach/detach sched: propagate asynchrous detach sched: Multiple upstream load tracking changes Viresh Kumar (1): cpufreq: schedutil: move slow path from workqueue to SCHED_FIFO task Yuyang Du (1): sched/fair: Initiate a new task's util avg to a bounded value kbuild test robot (2): ANDROID: sched/tune: __pcpu_scope_cpu_boost_groups can be static ANDROID: sched/tune: schedtune_allow_attach() can be static arch/arm64/configs/marlin_defconfig | 2 +- arch/arm64/kernel/topology.c | 7 +- drivers/cpufreq/Kconfig | 27 + drivers/cpufreq/Makefile | 2 +- drivers/cpufreq/cpufreq.c | 32 + drivers/cpufreq/cpufreq_governor_attr_set.c | 84 ++ include/linux/cgroup.h | 2 +- include/linux/cpufreq.h | 49 ++ include/linux/kthread.h | 4 + include/linux/sched.h | 20 +- include/linux/sched/sysctl.h | 7 +- include/trace/events/sched.h | 22 +- init/Kconfig | 1 + kernel/kthread.c | 96 +- kernel/sched/Makefile | 2 + kernel/sched/core.c | 84 +- kernel/sched/cpufreq.c | 63 ++ kernel/sched/cpufreq_sched.c | 220 ++--- kernel/sched/cpufreq_schedutil.c | 762 ++++++++++++++++ kernel/sched/deadline.c | 3 + kernel/sched/debug.c | 4 - kernel/sched/fair.c | 1254 ++++++++++++++++++--------- kernel/sched/features.h | 5 - kernel/sched/rt.c | 3 + kernel/sched/sched.h | 84 +- kernel/sched/tune.c | 5 +- kernel/sched/tune.h | 3 + kernel/sched/walt.c | 52 +- kernel/sysctl.c | 7 - 29 files changed, 2261 insertions(+), 645 deletions(-) create mode 100644 drivers/cpufreq/cpufreq_governor_attr_set.c create mode 100644 kernel/sched/cpufreq.c create mode 100644 kernel/sched/cpufreq_schedutil.c -- 1.9.1 IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

9 years

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

eas-dev February 2017