aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched.c
diff options
context:
space:
mode:
authorPeter Zijlstra <a.p.zijlstra@chello.nl>2009-09-01 04:34:37 -0400
committerIngo Molnar <mingo@elte.hu>2009-09-04 04:09:55 -0400
commite9e9250bc78e7f6342517214c0178a529807964b (patch)
tree9466a437ccfa93b200f4ee434fe807dd6b5fc050 /kernel/sched.c
parentab29230e673c646292c90c8b9d378b9562145af0 (diff)
sched: Scale down cpu_power due to RT tasks
Keep an average on the amount of time spend on RT tasks and use that fraction to scale down the cpu_power for regular tasks. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Tested-by: Andreas Herrmann <andreas.herrmann3@amd.com> Acked-by: Andreas Herrmann <andreas.herrmann3@amd.com> Acked-by: Gautham R Shenoy <ego@in.ibm.com> Cc: Balbir Singh <balbir@in.ibm.com> LKML-Reference: <20090901083826.287778431@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'kernel/sched.c')
-rw-r--r--kernel/sched.c64
1 files changed, 61 insertions, 3 deletions
diff --git a/kernel/sched.c b/kernel/sched.c
index 036600fd70bb..ab532b5de40e 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -627,6 +627,9 @@ struct rq {
627 627
628 struct task_struct *migration_thread; 628 struct task_struct *migration_thread;
629 struct list_head migration_queue; 629 struct list_head migration_queue;
630
631 u64 rt_avg;
632 u64 age_stamp;
630#endif 633#endif
631 634
632 /* calc_load related fields */ 635 /* calc_load related fields */
@@ -863,6 +866,14 @@ unsigned int sysctl_sched_shares_ratelimit = 250000;
863unsigned int sysctl_sched_shares_thresh = 4; 866unsigned int sysctl_sched_shares_thresh = 4;
864 867
865/* 868/*
869 * period over which we average the RT time consumption, measured
870 * in ms.
871 *
872 * default: 1s
873 */
874const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
875
876/*
866 * period over which we measure -rt task cpu usage in us. 877 * period over which we measure -rt task cpu usage in us.
867 * default: 1s 878 * default: 1s
868 */ 879 */
@@ -1280,12 +1291,37 @@ void wake_up_idle_cpu(int cpu)
1280} 1291}
1281#endif /* CONFIG_NO_HZ */ 1292#endif /* CONFIG_NO_HZ */
1282 1293
1294static u64 sched_avg_period(void)
1295{
1296 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1297}
1298
1299static void sched_avg_update(struct rq *rq)
1300{
1301 s64 period = sched_avg_period();
1302
1303 while ((s64)(rq->clock - rq->age_stamp) > period) {
1304 rq->age_stamp += period;
1305 rq->rt_avg /= 2;
1306 }
1307}
1308
1309static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1310{
1311 rq->rt_avg += rt_delta;
1312 sched_avg_update(rq);
1313}
1314
1283#else /* !CONFIG_SMP */ 1315#else /* !CONFIG_SMP */
1284static void resched_task(struct task_struct *p) 1316static void resched_task(struct task_struct *p)
1285{ 1317{
1286 assert_spin_locked(&task_rq(p)->lock); 1318 assert_spin_locked(&task_rq(p)->lock);
1287 set_tsk_need_resched(p); 1319 set_tsk_need_resched(p);
1288} 1320}
1321
1322static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1323{
1324}
1289#endif /* CONFIG_SMP */ 1325#endif /* CONFIG_SMP */
1290 1326
1291#if BITS_PER_LONG == 32 1327#if BITS_PER_LONG == 32
@@ -3699,7 +3735,7 @@ static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3699} 3735}
3700#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 3736#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3701 3737
3702unsigned long __weak arch_smt_gain(struct sched_domain *sd, int cpu) 3738unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3703{ 3739{
3704 unsigned long weight = cpumask_weight(sched_domain_span(sd)); 3740 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3705 unsigned long smt_gain = sd->smt_gain; 3741 unsigned long smt_gain = sd->smt_gain;
@@ -3709,6 +3745,24 @@ unsigned long __weak arch_smt_gain(struct sched_domain *sd, int cpu)
3709 return smt_gain; 3745 return smt_gain;
3710} 3746}
3711 3747
3748unsigned long scale_rt_power(int cpu)
3749{
3750 struct rq *rq = cpu_rq(cpu);
3751 u64 total, available;
3752
3753 sched_avg_update(rq);
3754
3755 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3756 available = total - rq->rt_avg;
3757
3758 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3759 total = SCHED_LOAD_SCALE;
3760
3761 total >>= SCHED_LOAD_SHIFT;
3762
3763 return div_u64(available, total);
3764}
3765
3712static void update_cpu_power(struct sched_domain *sd, int cpu) 3766static void update_cpu_power(struct sched_domain *sd, int cpu)
3713{ 3767{
3714 unsigned long weight = cpumask_weight(sched_domain_span(sd)); 3768 unsigned long weight = cpumask_weight(sched_domain_span(sd));
@@ -3719,11 +3773,15 @@ static void update_cpu_power(struct sched_domain *sd, int cpu)
3719 /* here we could scale based on cpufreq */ 3773 /* here we could scale based on cpufreq */
3720 3774
3721 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { 3775 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3722 power *= arch_smt_gain(sd, cpu); 3776 power *= arch_scale_smt_power(sd, cpu);
3723 power >>= SCHED_LOAD_SHIFT; 3777 power >>= SCHED_LOAD_SHIFT;
3724 } 3778 }
3725 3779
3726 /* here we could scale based on RT time */ 3780 power *= scale_rt_power(cpu);
3781 power >>= SCHED_LOAD_SHIFT;
3782
3783 if (!power)
3784 power = 1;
3727 3785
3728 if (power != old) { 3786 if (power != old) {
3729 sdg->__cpu_power = power; 3787 sdg->__cpu_power = power;