2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
30 #include <trace/events/sched.h>
35 * Targeted preemption latency for CPU-bound tasks:
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 * NOTE: this latency value is not the same as the concept of
39 * 'timeslice length' - timeslices in CFS are of variable length
40 * and have no persistent notion like in traditional, time-slice
41 * based scheduling concepts.
43 * (to see the precise effective timeslice length of your workload,
44 * run vmstat and monitor the context-switches (cs) field)
46 unsigned int sysctl_sched_latency = 6000000ULL;
47 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
50 * The initial- and re-scaling of tunables is configurable
51 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
55 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
56 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
58 enum sched_tunable_scaling sysctl_sched_tunable_scaling
59 = SCHED_TUNABLESCALING_LOG;
62 * Minimal preemption granularity for CPU-bound tasks:
63 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
65 unsigned int sysctl_sched_min_granularity = 750000ULL;
66 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
69 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
71 static unsigned int sched_nr_latency = 8;
74 * After fork, child runs first. If set to 0 (default) then
75 * parent will (try to) run first.
77 unsigned int sysctl_sched_child_runs_first __read_mostly;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 #ifdef CONFIG_CFS_BANDWIDTH
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
108 * default: 5 msec, units: microseconds
110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 * Increase the granularity value when there are more CPUs,
115 * because with more CPUs the 'effective latency' as visible
116 * to users decreases. But the relationship is not linear,
117 * so pick a second-best guess by going with the log2 of the
120 * This idea comes from the SD scheduler of Con Kolivas:
122 static int get_update_sysctl_factor(void)
124 unsigned int cpus = min_t(int, num_online_cpus(), 8);
127 switch (sysctl_sched_tunable_scaling) {
128 case SCHED_TUNABLESCALING_NONE:
131 case SCHED_TUNABLESCALING_LINEAR:
134 case SCHED_TUNABLESCALING_LOG:
136 factor = 1 + ilog2(cpus);
143 static void update_sysctl(void)
145 unsigned int factor = get_update_sysctl_factor();
147 #define SET_SYSCTL(name) \
148 (sysctl_##name = (factor) * normalized_sysctl_##name)
149 SET_SYSCTL(sched_min_granularity);
150 SET_SYSCTL(sched_latency);
151 SET_SYSCTL(sched_wakeup_granularity);
155 void sched_init_granularity(void)
160 #if BITS_PER_LONG == 32
161 # define WMULT_CONST (~0UL)
163 # define WMULT_CONST (1UL << 32)
166 #define WMULT_SHIFT 32
169 * Shift right and round:
171 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
174 * delta *= weight / lw
177 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
178 struct load_weight *lw)
183 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
184 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
185 * 2^SCHED_LOAD_RESOLUTION.
187 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
188 tmp = (u64)delta_exec * scale_load_down(weight);
190 tmp = (u64)delta_exec;
192 if (!lw->inv_weight) {
193 unsigned long w = scale_load_down(lw->weight);
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
200 lw->inv_weight = WMULT_CONST / w;
204 * Check whether we'd overflow the 64-bit multiplication:
206 if (unlikely(tmp > WMULT_CONST))
207 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
210 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
212 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
216 const struct sched_class fair_sched_class;
218 /**************************************************************
219 * CFS operations on generic schedulable entities:
222 #ifdef CONFIG_FAIR_GROUP_SCHED
224 /* cpu runqueue to which this cfs_rq is attached */
225 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
230 /* An entity is a task if it doesn't "own" a runqueue */
231 #define entity_is_task(se) (!se->my_q)
233 static inline struct task_struct *task_of(struct sched_entity *se)
235 #ifdef CONFIG_SCHED_DEBUG
236 WARN_ON_ONCE(!entity_is_task(se));
238 return container_of(se, struct task_struct, se);
241 /* Walk up scheduling entities hierarchy */
242 #define for_each_sched_entity(se) \
243 for (; se; se = se->parent)
245 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
250 /* runqueue on which this entity is (to be) queued */
251 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
256 /* runqueue "owned" by this group */
257 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
262 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
264 if (!cfs_rq->on_list) {
266 * Ensure we either appear before our parent (if already
267 * enqueued) or force our parent to appear after us when it is
268 * enqueued. The fact that we always enqueue bottom-up
269 * reduces this to two cases.
271 if (cfs_rq->tg->parent &&
272 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
273 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
274 &rq_of(cfs_rq)->leaf_cfs_rq_list);
276 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
277 &rq_of(cfs_rq)->leaf_cfs_rq_list);
284 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
286 if (cfs_rq->on_list) {
287 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
292 /* Iterate thr' all leaf cfs_rq's on a runqueue */
293 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
294 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
296 /* Do the two (enqueued) entities belong to the same group ? */
298 is_same_group(struct sched_entity *se, struct sched_entity *pse)
300 if (se->cfs_rq == pse->cfs_rq)
306 static inline struct sched_entity *parent_entity(struct sched_entity *se)
311 /* return depth at which a sched entity is present in the hierarchy */
312 static inline int depth_se(struct sched_entity *se)
316 for_each_sched_entity(se)
323 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
325 int se_depth, pse_depth;
328 * preemption test can be made between sibling entities who are in the
329 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
330 * both tasks until we find their ancestors who are siblings of common
334 /* First walk up until both entities are at same depth */
335 se_depth = depth_se(*se);
336 pse_depth = depth_se(*pse);
338 while (se_depth > pse_depth) {
340 *se = parent_entity(*se);
343 while (pse_depth > se_depth) {
345 *pse = parent_entity(*pse);
348 while (!is_same_group(*se, *pse)) {
349 *se = parent_entity(*se);
350 *pse = parent_entity(*pse);
354 #else /* !CONFIG_FAIR_GROUP_SCHED */
356 static inline struct task_struct *task_of(struct sched_entity *se)
358 return container_of(se, struct task_struct, se);
361 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
363 return container_of(cfs_rq, struct rq, cfs);
366 #define entity_is_task(se) 1
368 #define for_each_sched_entity(se) \
369 for (; se; se = NULL)
371 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
373 return &task_rq(p)->cfs;
376 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
378 struct task_struct *p = task_of(se);
379 struct rq *rq = task_rq(p);
384 /* runqueue "owned" by this group */
385 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
390 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
394 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
398 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
399 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
402 is_same_group(struct sched_entity *se, struct sched_entity *pse)
407 static inline struct sched_entity *parent_entity(struct sched_entity *se)
413 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
417 #endif /* CONFIG_FAIR_GROUP_SCHED */
419 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
420 unsigned long delta_exec);
422 /**************************************************************
423 * Scheduling class tree data structure manipulation methods:
426 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
428 s64 delta = (s64)(vruntime - min_vruntime);
430 min_vruntime = vruntime;
435 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
437 s64 delta = (s64)(vruntime - min_vruntime);
439 min_vruntime = vruntime;
444 static inline int entity_before(struct sched_entity *a,
445 struct sched_entity *b)
447 return (s64)(a->vruntime - b->vruntime) < 0;
450 static void update_min_vruntime(struct cfs_rq *cfs_rq)
452 u64 vruntime = cfs_rq->min_vruntime;
455 vruntime = cfs_rq->curr->vruntime;
457 if (cfs_rq->rb_leftmost) {
458 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
463 vruntime = se->vruntime;
465 vruntime = min_vruntime(vruntime, se->vruntime);
468 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
471 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
476 * Enqueue an entity into the rb-tree:
478 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
480 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
481 struct rb_node *parent = NULL;
482 struct sched_entity *entry;
486 * Find the right place in the rbtree:
490 entry = rb_entry(parent, struct sched_entity, run_node);
492 * We dont care about collisions. Nodes with
493 * the same key stay together.
495 if (entity_before(se, entry)) {
496 link = &parent->rb_left;
498 link = &parent->rb_right;
504 * Maintain a cache of leftmost tree entries (it is frequently
508 cfs_rq->rb_leftmost = &se->run_node;
510 rb_link_node(&se->run_node, parent, link);
511 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
514 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 if (cfs_rq->rb_leftmost == &se->run_node) {
517 struct rb_node *next_node;
519 next_node = rb_next(&se->run_node);
520 cfs_rq->rb_leftmost = next_node;
523 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
526 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
528 struct rb_node *left = cfs_rq->rb_leftmost;
533 return rb_entry(left, struct sched_entity, run_node);
536 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
538 struct rb_node *next = rb_next(&se->run_node);
543 return rb_entry(next, struct sched_entity, run_node);
546 #ifdef CONFIG_SCHED_DEBUG
547 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
549 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
554 return rb_entry(last, struct sched_entity, run_node);
557 /**************************************************************
558 * Scheduling class statistics methods:
561 int sched_proc_update_handler(struct ctl_table *table, int write,
562 void __user *buffer, size_t *lenp,
565 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
566 int factor = get_update_sysctl_factor();
571 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
572 sysctl_sched_min_granularity);
574 #define WRT_SYSCTL(name) \
575 (normalized_sysctl_##name = sysctl_##name / (factor))
576 WRT_SYSCTL(sched_min_granularity);
577 WRT_SYSCTL(sched_latency);
578 WRT_SYSCTL(sched_wakeup_granularity);
588 static inline unsigned long
589 calc_delta_fair(unsigned long delta, struct sched_entity *se)
591 if (unlikely(se->load.weight != NICE_0_LOAD))
592 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
598 * The idea is to set a period in which each task runs once.
600 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
601 * this period because otherwise the slices get too small.
603 * p = (nr <= nl) ? l : l*nr/nl
605 static u64 __sched_period(unsigned long nr_running)
607 u64 period = sysctl_sched_latency;
608 unsigned long nr_latency = sched_nr_latency;
610 if (unlikely(nr_running > nr_latency)) {
611 period = sysctl_sched_min_granularity;
612 period *= nr_running;
619 * We calculate the wall-time slice from the period by taking a part
620 * proportional to the weight.
624 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
626 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
628 for_each_sched_entity(se) {
629 struct load_weight *load;
630 struct load_weight lw;
632 cfs_rq = cfs_rq_of(se);
633 load = &cfs_rq->load;
635 if (unlikely(!se->on_rq)) {
638 update_load_add(&lw, se->load.weight);
641 slice = calc_delta_mine(slice, se->load.weight, load);
647 * We calculate the vruntime slice of a to be inserted task
651 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 return calc_delta_fair(sched_slice(cfs_rq, se), se);
656 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
657 static void update_cfs_shares(struct cfs_rq *cfs_rq);
660 * Update the current task's runtime statistics. Skip current tasks that
661 * are not in our scheduling class.
664 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
665 unsigned long delta_exec)
667 unsigned long delta_exec_weighted;
669 schedstat_set(curr->statistics.exec_max,
670 max((u64)delta_exec, curr->statistics.exec_max));
672 curr->sum_exec_runtime += delta_exec;
673 schedstat_add(cfs_rq, exec_clock, delta_exec);
674 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
676 curr->vruntime += delta_exec_weighted;
677 update_min_vruntime(cfs_rq);
679 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
680 cfs_rq->load_unacc_exec_time += delta_exec;
684 static void update_curr(struct cfs_rq *cfs_rq)
686 struct sched_entity *curr = cfs_rq->curr;
687 u64 now = rq_of(cfs_rq)->clock_task;
688 unsigned long delta_exec;
694 * Get the amount of time the current task was running
695 * since the last time we changed load (this cannot
696 * overflow on 32 bits):
698 delta_exec = (unsigned long)(now - curr->exec_start);
702 __update_curr(cfs_rq, curr, delta_exec);
703 curr->exec_start = now;
705 if (entity_is_task(curr)) {
706 struct task_struct *curtask = task_of(curr);
708 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
709 cpuacct_charge(curtask, delta_exec);
710 account_group_exec_runtime(curtask, delta_exec);
713 account_cfs_rq_runtime(cfs_rq, delta_exec);
717 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
719 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
723 * Task is being enqueued - update stats:
725 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
728 * Are we enqueueing a waiting task? (for current tasks
729 * a dequeue/enqueue event is a NOP)
731 if (se != cfs_rq->curr)
732 update_stats_wait_start(cfs_rq, se);
736 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
738 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
739 rq_of(cfs_rq)->clock - se->statistics.wait_start));
740 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
741 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
742 rq_of(cfs_rq)->clock - se->statistics.wait_start);
743 #ifdef CONFIG_SCHEDSTATS
744 if (entity_is_task(se)) {
745 trace_sched_stat_wait(task_of(se),
746 rq_of(cfs_rq)->clock - se->statistics.wait_start);
749 schedstat_set(se->statistics.wait_start, 0);
753 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
756 * Mark the end of the wait period if dequeueing a
759 if (se != cfs_rq->curr)
760 update_stats_wait_end(cfs_rq, se);
764 * We are picking a new current task - update its stats:
767 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * We are starting a new run period:
772 se->exec_start = rq_of(cfs_rq)->clock_task;
775 /**************************************************
776 * Scheduling class queueing methods:
779 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
781 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
783 cfs_rq->task_weight += weight;
787 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
793 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
795 update_load_add(&cfs_rq->load, se->load.weight);
796 if (!parent_entity(se))
797 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
798 if (entity_is_task(se)) {
799 add_cfs_task_weight(cfs_rq, se->load.weight);
800 list_add(&se->group_node, &cfs_rq->tasks);
802 cfs_rq->nr_running++;
806 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
808 update_load_sub(&cfs_rq->load, se->load.weight);
809 if (!parent_entity(se))
810 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
811 if (entity_is_task(se)) {
812 add_cfs_task_weight(cfs_rq, -se->load.weight);
813 list_del_init(&se->group_node);
815 cfs_rq->nr_running--;
818 #ifdef CONFIG_FAIR_GROUP_SCHED
819 /* we need this in update_cfs_load and load-balance functions below */
820 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
822 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
825 struct task_group *tg = cfs_rq->tg;
828 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
829 load_avg -= cfs_rq->load_contribution;
831 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
832 atomic_add(load_avg, &tg->load_weight);
833 cfs_rq->load_contribution += load_avg;
837 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
839 u64 period = sysctl_sched_shares_window;
841 unsigned long load = cfs_rq->load.weight;
843 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
846 now = rq_of(cfs_rq)->clock_task;
847 delta = now - cfs_rq->load_stamp;
849 /* truncate load history at 4 idle periods */
850 if (cfs_rq->load_stamp > cfs_rq->load_last &&
851 now - cfs_rq->load_last > 4 * period) {
852 cfs_rq->load_period = 0;
853 cfs_rq->load_avg = 0;
857 cfs_rq->load_stamp = now;
858 cfs_rq->load_unacc_exec_time = 0;
859 cfs_rq->load_period += delta;
861 cfs_rq->load_last = now;
862 cfs_rq->load_avg += delta * load;
865 /* consider updating load contribution on each fold or truncate */
866 if (global_update || cfs_rq->load_period > period
867 || !cfs_rq->load_period)
868 update_cfs_rq_load_contribution(cfs_rq, global_update);
870 while (cfs_rq->load_period > period) {
872 * Inline assembly required to prevent the compiler
873 * optimising this loop into a divmod call.
874 * See __iter_div_u64_rem() for another example of this.
876 asm("" : "+rm" (cfs_rq->load_period));
877 cfs_rq->load_period /= 2;
878 cfs_rq->load_avg /= 2;
881 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
882 list_del_leaf_cfs_rq(cfs_rq);
885 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
890 * Use this CPU's actual weight instead of the last load_contribution
891 * to gain a more accurate current total weight. See
892 * update_cfs_rq_load_contribution().
894 tg_weight = atomic_read(&tg->load_weight);
895 tg_weight -= cfs_rq->load_contribution;
896 tg_weight += cfs_rq->load.weight;
901 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
903 long tg_weight, load, shares;
905 tg_weight = calc_tg_weight(tg, cfs_rq);
906 load = cfs_rq->load.weight;
908 shares = (tg->shares * load);
912 if (shares < MIN_SHARES)
914 if (shares > tg->shares)
920 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
922 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
923 update_cfs_load(cfs_rq, 0);
924 update_cfs_shares(cfs_rq);
927 # else /* CONFIG_SMP */
928 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
932 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
937 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
940 # endif /* CONFIG_SMP */
941 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
942 unsigned long weight)
945 /* commit outstanding execution time */
946 if (cfs_rq->curr == se)
948 account_entity_dequeue(cfs_rq, se);
951 update_load_set(&se->load, weight);
954 account_entity_enqueue(cfs_rq, se);
957 static void update_cfs_shares(struct cfs_rq *cfs_rq)
959 struct task_group *tg;
960 struct sched_entity *se;
964 se = tg->se[cpu_of(rq_of(cfs_rq))];
965 if (!se || throttled_hierarchy(cfs_rq))
968 if (likely(se->load.weight == tg->shares))
971 shares = calc_cfs_shares(cfs_rq, tg);
973 reweight_entity(cfs_rq_of(se), se, shares);
975 #else /* CONFIG_FAIR_GROUP_SCHED */
976 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
980 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
984 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
987 #endif /* CONFIG_FAIR_GROUP_SCHED */
989 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
991 #ifdef CONFIG_SCHEDSTATS
992 struct task_struct *tsk = NULL;
994 if (entity_is_task(se))
997 if (se->statistics.sleep_start) {
998 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1003 if (unlikely(delta > se->statistics.sleep_max))
1004 se->statistics.sleep_max = delta;
1006 se->statistics.sleep_start = 0;
1007 se->statistics.sum_sleep_runtime += delta;
1010 account_scheduler_latency(tsk, delta >> 10, 1);
1011 trace_sched_stat_sleep(tsk, delta);
1014 if (se->statistics.block_start) {
1015 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1020 if (unlikely(delta > se->statistics.block_max))
1021 se->statistics.block_max = delta;
1023 se->statistics.block_start = 0;
1024 se->statistics.sum_sleep_runtime += delta;
1027 if (tsk->in_iowait) {
1028 se->statistics.iowait_sum += delta;
1029 se->statistics.iowait_count++;
1030 trace_sched_stat_iowait(tsk, delta);
1033 trace_sched_stat_blocked(tsk, delta);
1036 * Blocking time is in units of nanosecs, so shift by
1037 * 20 to get a milliseconds-range estimation of the
1038 * amount of time that the task spent sleeping:
1040 if (unlikely(prof_on == SLEEP_PROFILING)) {
1041 profile_hits(SLEEP_PROFILING,
1042 (void *)get_wchan(tsk),
1045 account_scheduler_latency(tsk, delta >> 10, 0);
1051 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1053 #ifdef CONFIG_SCHED_DEBUG
1054 s64 d = se->vruntime - cfs_rq->min_vruntime;
1059 if (d > 3*sysctl_sched_latency)
1060 schedstat_inc(cfs_rq, nr_spread_over);
1065 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1067 u64 vruntime = cfs_rq->min_vruntime;
1070 * The 'current' period is already promised to the current tasks,
1071 * however the extra weight of the new task will slow them down a
1072 * little, place the new task so that it fits in the slot that
1073 * stays open at the end.
1075 if (initial && sched_feat(START_DEBIT))
1076 vruntime += sched_vslice(cfs_rq, se);
1078 /* sleeps up to a single latency don't count. */
1080 unsigned long thresh = sysctl_sched_latency;
1083 * Halve their sleep time's effect, to allow
1084 * for a gentler effect of sleepers:
1086 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1092 /* ensure we never gain time by being placed backwards. */
1093 vruntime = max_vruntime(se->vruntime, vruntime);
1095 se->vruntime = vruntime;
1098 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1101 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1104 * Update the normalized vruntime before updating min_vruntime
1105 * through callig update_curr().
1107 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1108 se->vruntime += cfs_rq->min_vruntime;
1111 * Update run-time statistics of the 'current'.
1113 update_curr(cfs_rq);
1114 update_cfs_load(cfs_rq, 0);
1115 account_entity_enqueue(cfs_rq, se);
1116 update_cfs_shares(cfs_rq);
1118 if (flags & ENQUEUE_WAKEUP) {
1119 place_entity(cfs_rq, se, 0);
1120 enqueue_sleeper(cfs_rq, se);
1123 update_stats_enqueue(cfs_rq, se);
1124 check_spread(cfs_rq, se);
1125 if (se != cfs_rq->curr)
1126 __enqueue_entity(cfs_rq, se);
1129 if (cfs_rq->nr_running == 1) {
1130 list_add_leaf_cfs_rq(cfs_rq);
1131 check_enqueue_throttle(cfs_rq);
1135 static void __clear_buddies_last(struct sched_entity *se)
1137 for_each_sched_entity(se) {
1138 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1139 if (cfs_rq->last == se)
1140 cfs_rq->last = NULL;
1146 static void __clear_buddies_next(struct sched_entity *se)
1148 for_each_sched_entity(se) {
1149 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1150 if (cfs_rq->next == se)
1151 cfs_rq->next = NULL;
1157 static void __clear_buddies_skip(struct sched_entity *se)
1159 for_each_sched_entity(se) {
1160 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1161 if (cfs_rq->skip == se)
1162 cfs_rq->skip = NULL;
1168 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1170 if (cfs_rq->last == se)
1171 __clear_buddies_last(se);
1173 if (cfs_rq->next == se)
1174 __clear_buddies_next(se);
1176 if (cfs_rq->skip == se)
1177 __clear_buddies_skip(se);
1180 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1183 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1186 * Update run-time statistics of the 'current'.
1188 update_curr(cfs_rq);
1190 update_stats_dequeue(cfs_rq, se);
1191 if (flags & DEQUEUE_SLEEP) {
1192 #ifdef CONFIG_SCHEDSTATS
1193 if (entity_is_task(se)) {
1194 struct task_struct *tsk = task_of(se);
1196 if (tsk->state & TASK_INTERRUPTIBLE)
1197 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1198 if (tsk->state & TASK_UNINTERRUPTIBLE)
1199 se->statistics.block_start = rq_of(cfs_rq)->clock;
1204 clear_buddies(cfs_rq, se);
1206 if (se != cfs_rq->curr)
1207 __dequeue_entity(cfs_rq, se);
1209 update_cfs_load(cfs_rq, 0);
1210 account_entity_dequeue(cfs_rq, se);
1213 * Normalize the entity after updating the min_vruntime because the
1214 * update can refer to the ->curr item and we need to reflect this
1215 * movement in our normalized position.
1217 if (!(flags & DEQUEUE_SLEEP))
1218 se->vruntime -= cfs_rq->min_vruntime;
1220 /* return excess runtime on last dequeue */
1221 return_cfs_rq_runtime(cfs_rq);
1223 update_min_vruntime(cfs_rq);
1224 update_cfs_shares(cfs_rq);
1228 * Preempt the current task with a newly woken task if needed:
1231 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1233 unsigned long ideal_runtime, delta_exec;
1234 struct sched_entity *se;
1237 ideal_runtime = sched_slice(cfs_rq, curr);
1238 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1239 if (delta_exec > ideal_runtime) {
1240 resched_task(rq_of(cfs_rq)->curr);
1242 * The current task ran long enough, ensure it doesn't get
1243 * re-elected due to buddy favours.
1245 clear_buddies(cfs_rq, curr);
1250 * Ensure that a task that missed wakeup preemption by a
1251 * narrow margin doesn't have to wait for a full slice.
1252 * This also mitigates buddy induced latencies under load.
1254 if (delta_exec < sysctl_sched_min_granularity)
1257 se = __pick_first_entity(cfs_rq);
1258 delta = curr->vruntime - se->vruntime;
1263 if (delta > ideal_runtime)
1264 resched_task(rq_of(cfs_rq)->curr);
1268 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1270 /* 'current' is not kept within the tree. */
1273 * Any task has to be enqueued before it get to execute on
1274 * a CPU. So account for the time it spent waiting on the
1277 update_stats_wait_end(cfs_rq, se);
1278 __dequeue_entity(cfs_rq, se);
1281 update_stats_curr_start(cfs_rq, se);
1283 #ifdef CONFIG_SCHEDSTATS
1285 * Track our maximum slice length, if the CPU's load is at
1286 * least twice that of our own weight (i.e. dont track it
1287 * when there are only lesser-weight tasks around):
1289 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1290 se->statistics.slice_max = max(se->statistics.slice_max,
1291 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1294 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1298 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1301 * Pick the next process, keeping these things in mind, in this order:
1302 * 1) keep things fair between processes/task groups
1303 * 2) pick the "next" process, since someone really wants that to run
1304 * 3) pick the "last" process, for cache locality
1305 * 4) do not run the "skip" process, if something else is available
1307 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1309 struct sched_entity *se = __pick_first_entity(cfs_rq);
1310 struct sched_entity *left = se;
1313 * Avoid running the skip buddy, if running something else can
1314 * be done without getting too unfair.
1316 if (cfs_rq->skip == se) {
1317 struct sched_entity *second = __pick_next_entity(se);
1318 if (second && wakeup_preempt_entity(second, left) < 1)
1323 * Prefer last buddy, try to return the CPU to a preempted task.
1325 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1329 * Someone really wants this to run. If it's not unfair, run it.
1331 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1334 clear_buddies(cfs_rq, se);
1339 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1341 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1344 * If still on the runqueue then deactivate_task()
1345 * was not called and update_curr() has to be done:
1348 update_curr(cfs_rq);
1350 /* throttle cfs_rqs exceeding runtime */
1351 check_cfs_rq_runtime(cfs_rq);
1353 check_spread(cfs_rq, prev);
1355 update_stats_wait_start(cfs_rq, prev);
1356 /* Put 'current' back into the tree. */
1357 __enqueue_entity(cfs_rq, prev);
1359 cfs_rq->curr = NULL;
1363 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1366 * Update run-time statistics of the 'current'.
1368 update_curr(cfs_rq);
1371 * Update share accounting for long-running entities.
1373 update_entity_shares_tick(cfs_rq);
1375 #ifdef CONFIG_SCHED_HRTICK
1377 * queued ticks are scheduled to match the slice, so don't bother
1378 * validating it and just reschedule.
1381 resched_task(rq_of(cfs_rq)->curr);
1385 * don't let the period tick interfere with the hrtick preemption
1387 if (!sched_feat(DOUBLE_TICK) &&
1388 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1392 if (cfs_rq->nr_running > 1)
1393 check_preempt_tick(cfs_rq, curr);
1397 /**************************************************
1398 * CFS bandwidth control machinery
1401 #ifdef CONFIG_CFS_BANDWIDTH
1403 #ifdef HAVE_JUMP_LABEL
1404 static struct jump_label_key __cfs_bandwidth_used;
1406 static inline bool cfs_bandwidth_used(void)
1408 return static_branch(&__cfs_bandwidth_used);
1411 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1413 /* only need to count groups transitioning between enabled/!enabled */
1414 if (enabled && !was_enabled)
1415 jump_label_inc(&__cfs_bandwidth_used);
1416 else if (!enabled && was_enabled)
1417 jump_label_dec(&__cfs_bandwidth_used);
1419 #else /* HAVE_JUMP_LABEL */
1420 static bool cfs_bandwidth_used(void)
1425 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1426 #endif /* HAVE_JUMP_LABEL */
1429 * default period for cfs group bandwidth.
1430 * default: 0.1s, units: nanoseconds
1432 static inline u64 default_cfs_period(void)
1434 return 100000000ULL;
1437 static inline u64 sched_cfs_bandwidth_slice(void)
1439 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1443 * Replenish runtime according to assigned quota and update expiration time.
1444 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1445 * additional synchronization around rq->lock.
1447 * requires cfs_b->lock
1449 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1453 if (cfs_b->quota == RUNTIME_INF)
1456 now = sched_clock_cpu(smp_processor_id());
1457 cfs_b->runtime = cfs_b->quota;
1458 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1461 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1463 return &tg->cfs_bandwidth;
1466 /* returns 0 on failure to allocate runtime */
1467 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1469 struct task_group *tg = cfs_rq->tg;
1470 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1471 u64 amount = 0, min_amount, expires;
1473 /* note: this is a positive sum as runtime_remaining <= 0 */
1474 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1476 raw_spin_lock(&cfs_b->lock);
1477 if (cfs_b->quota == RUNTIME_INF)
1478 amount = min_amount;
1481 * If the bandwidth pool has become inactive, then at least one
1482 * period must have elapsed since the last consumption.
1483 * Refresh the global state and ensure bandwidth timer becomes
1486 if (!cfs_b->timer_active) {
1487 __refill_cfs_bandwidth_runtime(cfs_b);
1488 __start_cfs_bandwidth(cfs_b);
1491 if (cfs_b->runtime > 0) {
1492 amount = min(cfs_b->runtime, min_amount);
1493 cfs_b->runtime -= amount;
1497 expires = cfs_b->runtime_expires;
1498 raw_spin_unlock(&cfs_b->lock);
1500 cfs_rq->runtime_remaining += amount;
1502 * we may have advanced our local expiration to account for allowed
1503 * spread between our sched_clock and the one on which runtime was
1506 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1507 cfs_rq->runtime_expires = expires;
1509 return cfs_rq->runtime_remaining > 0;
1513 * Note: This depends on the synchronization provided by sched_clock and the
1514 * fact that rq->clock snapshots this value.
1516 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1518 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1519 struct rq *rq = rq_of(cfs_rq);
1521 /* if the deadline is ahead of our clock, nothing to do */
1522 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1525 if (cfs_rq->runtime_remaining < 0)
1529 * If the local deadline has passed we have to consider the
1530 * possibility that our sched_clock is 'fast' and the global deadline
1531 * has not truly expired.
1533 * Fortunately we can check determine whether this the case by checking
1534 * whether the global deadline has advanced.
1537 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1538 /* extend local deadline, drift is bounded above by 2 ticks */
1539 cfs_rq->runtime_expires += TICK_NSEC;
1541 /* global deadline is ahead, expiration has passed */
1542 cfs_rq->runtime_remaining = 0;
1546 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1547 unsigned long delta_exec)
1549 /* dock delta_exec before expiring quota (as it could span periods) */
1550 cfs_rq->runtime_remaining -= delta_exec;
1551 expire_cfs_rq_runtime(cfs_rq);
1553 if (likely(cfs_rq->runtime_remaining > 0))
1557 * if we're unable to extend our runtime we resched so that the active
1558 * hierarchy can be throttled
1560 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1561 resched_task(rq_of(cfs_rq)->curr);
1564 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1565 unsigned long delta_exec)
1567 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1570 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1573 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1575 return cfs_bandwidth_used() && cfs_rq->throttled;
1578 /* check whether cfs_rq, or any parent, is throttled */
1579 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1581 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1585 * Ensure that neither of the group entities corresponding to src_cpu or
1586 * dest_cpu are members of a throttled hierarchy when performing group
1587 * load-balance operations.
1589 static inline int throttled_lb_pair(struct task_group *tg,
1590 int src_cpu, int dest_cpu)
1592 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1594 src_cfs_rq = tg->cfs_rq[src_cpu];
1595 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1597 return throttled_hierarchy(src_cfs_rq) ||
1598 throttled_hierarchy(dest_cfs_rq);
1601 /* updated child weight may affect parent so we have to do this bottom up */
1602 static int tg_unthrottle_up(struct task_group *tg, void *data)
1604 struct rq *rq = data;
1605 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1607 cfs_rq->throttle_count--;
1609 if (!cfs_rq->throttle_count) {
1610 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1612 /* leaving throttled state, advance shares averaging windows */
1613 cfs_rq->load_stamp += delta;
1614 cfs_rq->load_last += delta;
1616 /* update entity weight now that we are on_rq again */
1617 update_cfs_shares(cfs_rq);
1624 static int tg_throttle_down(struct task_group *tg, void *data)
1626 struct rq *rq = data;
1627 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1629 /* group is entering throttled state, record last load */
1630 if (!cfs_rq->throttle_count)
1631 update_cfs_load(cfs_rq, 0);
1632 cfs_rq->throttle_count++;
1637 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1639 struct rq *rq = rq_of(cfs_rq);
1640 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1641 struct sched_entity *se;
1642 long task_delta, dequeue = 1;
1644 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1646 /* account load preceding throttle */
1648 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1651 task_delta = cfs_rq->h_nr_running;
1652 for_each_sched_entity(se) {
1653 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1654 /* throttled entity or throttle-on-deactivate */
1659 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1660 qcfs_rq->h_nr_running -= task_delta;
1662 if (qcfs_rq->load.weight)
1667 rq->nr_running -= task_delta;
1669 cfs_rq->throttled = 1;
1670 cfs_rq->throttled_timestamp = rq->clock;
1671 raw_spin_lock(&cfs_b->lock);
1672 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1673 raw_spin_unlock(&cfs_b->lock);
1676 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1678 struct rq *rq = rq_of(cfs_rq);
1679 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1680 struct sched_entity *se;
1684 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1686 cfs_rq->throttled = 0;
1687 raw_spin_lock(&cfs_b->lock);
1688 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1689 list_del_rcu(&cfs_rq->throttled_list);
1690 raw_spin_unlock(&cfs_b->lock);
1691 cfs_rq->throttled_timestamp = 0;
1693 update_rq_clock(rq);
1694 /* update hierarchical throttle state */
1695 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1697 if (!cfs_rq->load.weight)
1700 task_delta = cfs_rq->h_nr_running;
1701 for_each_sched_entity(se) {
1705 cfs_rq = cfs_rq_of(se);
1707 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1708 cfs_rq->h_nr_running += task_delta;
1710 if (cfs_rq_throttled(cfs_rq))
1715 rq->nr_running += task_delta;
1717 /* determine whether we need to wake up potentially idle cpu */
1718 if (rq->curr == rq->idle && rq->cfs.nr_running)
1719 resched_task(rq->curr);
1722 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1723 u64 remaining, u64 expires)
1725 struct cfs_rq *cfs_rq;
1726 u64 runtime = remaining;
1729 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1731 struct rq *rq = rq_of(cfs_rq);
1733 raw_spin_lock(&rq->lock);
1734 if (!cfs_rq_throttled(cfs_rq))
1737 runtime = -cfs_rq->runtime_remaining + 1;
1738 if (runtime > remaining)
1739 runtime = remaining;
1740 remaining -= runtime;
1742 cfs_rq->runtime_remaining += runtime;
1743 cfs_rq->runtime_expires = expires;
1745 /* we check whether we're throttled above */
1746 if (cfs_rq->runtime_remaining > 0)
1747 unthrottle_cfs_rq(cfs_rq);
1750 raw_spin_unlock(&rq->lock);
1761 * Responsible for refilling a task_group's bandwidth and unthrottling its
1762 * cfs_rqs as appropriate. If there has been no activity within the last
1763 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1764 * used to track this state.
1766 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1768 u64 runtime, runtime_expires;
1769 int idle = 1, throttled;
1771 raw_spin_lock(&cfs_b->lock);
1772 /* no need to continue the timer with no bandwidth constraint */
1773 if (cfs_b->quota == RUNTIME_INF)
1776 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1777 /* idle depends on !throttled (for the case of a large deficit) */
1778 idle = cfs_b->idle && !throttled;
1779 cfs_b->nr_periods += overrun;
1781 /* if we're going inactive then everything else can be deferred */
1785 __refill_cfs_bandwidth_runtime(cfs_b);
1788 /* mark as potentially idle for the upcoming period */
1793 /* account preceding periods in which throttling occurred */
1794 cfs_b->nr_throttled += overrun;
1797 * There are throttled entities so we must first use the new bandwidth
1798 * to unthrottle them before making it generally available. This
1799 * ensures that all existing debts will be paid before a new cfs_rq is
1802 runtime = cfs_b->runtime;
1803 runtime_expires = cfs_b->runtime_expires;
1807 * This check is repeated as we are holding onto the new bandwidth
1808 * while we unthrottle. This can potentially race with an unthrottled
1809 * group trying to acquire new bandwidth from the global pool.
1811 while (throttled && runtime > 0) {
1812 raw_spin_unlock(&cfs_b->lock);
1813 /* we can't nest cfs_b->lock while distributing bandwidth */
1814 runtime = distribute_cfs_runtime(cfs_b, runtime,
1816 raw_spin_lock(&cfs_b->lock);
1818 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1821 /* return (any) remaining runtime */
1822 cfs_b->runtime = runtime;
1824 * While we are ensured activity in the period following an
1825 * unthrottle, this also covers the case in which the new bandwidth is
1826 * insufficient to cover the existing bandwidth deficit. (Forcing the
1827 * timer to remain active while there are any throttled entities.)
1832 cfs_b->timer_active = 0;
1833 raw_spin_unlock(&cfs_b->lock);
1838 /* a cfs_rq won't donate quota below this amount */
1839 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1840 /* minimum remaining period time to redistribute slack quota */
1841 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1842 /* how long we wait to gather additional slack before distributing */
1843 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1845 /* are we near the end of the current quota period? */
1846 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1848 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1851 /* if the call-back is running a quota refresh is already occurring */
1852 if (hrtimer_callback_running(refresh_timer))
1855 /* is a quota refresh about to occur? */
1856 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1857 if (remaining < min_expire)
1863 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1865 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1867 /* if there's a quota refresh soon don't bother with slack */
1868 if (runtime_refresh_within(cfs_b, min_left))
1871 start_bandwidth_timer(&cfs_b->slack_timer,
1872 ns_to_ktime(cfs_bandwidth_slack_period));
1875 /* we know any runtime found here is valid as update_curr() precedes return */
1876 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1878 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1879 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1881 if (slack_runtime <= 0)
1884 raw_spin_lock(&cfs_b->lock);
1885 if (cfs_b->quota != RUNTIME_INF &&
1886 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1887 cfs_b->runtime += slack_runtime;
1889 /* we are under rq->lock, defer unthrottling using a timer */
1890 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1891 !list_empty(&cfs_b->throttled_cfs_rq))
1892 start_cfs_slack_bandwidth(cfs_b);
1894 raw_spin_unlock(&cfs_b->lock);
1896 /* even if it's not valid for return we don't want to try again */
1897 cfs_rq->runtime_remaining -= slack_runtime;
1900 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1902 if (!cfs_bandwidth_used())
1905 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
1908 __return_cfs_rq_runtime(cfs_rq);
1912 * This is done with a timer (instead of inline with bandwidth return) since
1913 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1915 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
1917 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
1920 /* confirm we're still not at a refresh boundary */
1921 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
1924 raw_spin_lock(&cfs_b->lock);
1925 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
1926 runtime = cfs_b->runtime;
1929 expires = cfs_b->runtime_expires;
1930 raw_spin_unlock(&cfs_b->lock);
1935 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
1937 raw_spin_lock(&cfs_b->lock);
1938 if (expires == cfs_b->runtime_expires)
1939 cfs_b->runtime = runtime;
1940 raw_spin_unlock(&cfs_b->lock);
1944 * When a group wakes up we want to make sure that its quota is not already
1945 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1946 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1948 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1950 if (!cfs_bandwidth_used())
1953 /* an active group must be handled by the update_curr()->put() path */
1954 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1957 /* ensure the group is not already throttled */
1958 if (cfs_rq_throttled(cfs_rq))
1961 /* update runtime allocation */
1962 account_cfs_rq_runtime(cfs_rq, 0);
1963 if (cfs_rq->runtime_remaining <= 0)
1964 throttle_cfs_rq(cfs_rq);
1967 /* conditionally throttle active cfs_rq's from put_prev_entity() */
1968 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1970 if (!cfs_bandwidth_used())
1973 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1977 * it's possible for a throttled entity to be forced into a running
1978 * state (e.g. set_curr_task), in this case we're finished.
1980 if (cfs_rq_throttled(cfs_rq))
1983 throttle_cfs_rq(cfs_rq);
1986 static inline u64 default_cfs_period(void);
1987 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
1988 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
1990 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
1992 struct cfs_bandwidth *cfs_b =
1993 container_of(timer, struct cfs_bandwidth, slack_timer);
1994 do_sched_cfs_slack_timer(cfs_b);
1996 return HRTIMER_NORESTART;
1999 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2001 struct cfs_bandwidth *cfs_b =
2002 container_of(timer, struct cfs_bandwidth, period_timer);
2008 now = hrtimer_cb_get_time(timer);
2009 overrun = hrtimer_forward(timer, now, cfs_b->period);
2014 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2017 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2020 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2022 raw_spin_lock_init(&cfs_b->lock);
2024 cfs_b->quota = RUNTIME_INF;
2025 cfs_b->period = ns_to_ktime(default_cfs_period());
2027 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2028 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2029 cfs_b->period_timer.function = sched_cfs_period_timer;
2030 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2031 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2034 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2036 cfs_rq->runtime_enabled = 0;
2037 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2040 /* requires cfs_b->lock, may release to reprogram timer */
2041 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2044 * The timer may be active because we're trying to set a new bandwidth
2045 * period or because we're racing with the tear-down path
2046 * (timer_active==0 becomes visible before the hrtimer call-back
2047 * terminates). In either case we ensure that it's re-programmed
2049 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2050 raw_spin_unlock(&cfs_b->lock);
2051 /* ensure cfs_b->lock is available while we wait */
2052 hrtimer_cancel(&cfs_b->period_timer);
2054 raw_spin_lock(&cfs_b->lock);
2055 /* if someone else restarted the timer then we're done */
2056 if (cfs_b->timer_active)
2060 cfs_b->timer_active = 1;
2061 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2064 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2066 hrtimer_cancel(&cfs_b->period_timer);
2067 hrtimer_cancel(&cfs_b->slack_timer);
2070 void unthrottle_offline_cfs_rqs(struct rq *rq)
2072 struct cfs_rq *cfs_rq;
2074 for_each_leaf_cfs_rq(rq, cfs_rq) {
2075 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2077 if (!cfs_rq->runtime_enabled)
2081 * clock_task is not advancing so we just need to make sure
2082 * there's some valid quota amount
2084 cfs_rq->runtime_remaining = cfs_b->quota;
2085 if (cfs_rq_throttled(cfs_rq))
2086 unthrottle_cfs_rq(cfs_rq);
2090 #else /* CONFIG_CFS_BANDWIDTH */
2091 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2092 unsigned long delta_exec) {}
2093 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2094 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2095 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2097 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2102 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2107 static inline int throttled_lb_pair(struct task_group *tg,
2108 int src_cpu, int dest_cpu)
2113 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2115 #ifdef CONFIG_FAIR_GROUP_SCHED
2116 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2119 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2123 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2124 void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2126 #endif /* CONFIG_CFS_BANDWIDTH */
2128 /**************************************************
2129 * CFS operations on tasks:
2132 #ifdef CONFIG_SCHED_HRTICK
2133 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2135 struct sched_entity *se = &p->se;
2136 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2138 WARN_ON(task_rq(p) != rq);
2140 if (cfs_rq->nr_running > 1) {
2141 u64 slice = sched_slice(cfs_rq, se);
2142 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2143 s64 delta = slice - ran;
2152 * Don't schedule slices shorter than 10000ns, that just
2153 * doesn't make sense. Rely on vruntime for fairness.
2156 delta = max_t(s64, 10000LL, delta);
2158 hrtick_start(rq, delta);
2163 * called from enqueue/dequeue and updates the hrtick when the
2164 * current task is from our class and nr_running is low enough
2167 static void hrtick_update(struct rq *rq)
2169 struct task_struct *curr = rq->curr;
2171 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2174 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2175 hrtick_start_fair(rq, curr);
2177 #else /* !CONFIG_SCHED_HRTICK */
2179 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2183 static inline void hrtick_update(struct rq *rq)
2189 * The enqueue_task method is called before nr_running is
2190 * increased. Here we update the fair scheduling stats and
2191 * then put the task into the rbtree:
2194 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2196 struct cfs_rq *cfs_rq;
2197 struct sched_entity *se = &p->se;
2199 for_each_sched_entity(se) {
2202 cfs_rq = cfs_rq_of(se);
2203 enqueue_entity(cfs_rq, se, flags);
2206 * end evaluation on encountering a throttled cfs_rq
2208 * note: in the case of encountering a throttled cfs_rq we will
2209 * post the final h_nr_running increment below.
2211 if (cfs_rq_throttled(cfs_rq))
2213 cfs_rq->h_nr_running++;
2215 flags = ENQUEUE_WAKEUP;
2218 for_each_sched_entity(se) {
2219 cfs_rq = cfs_rq_of(se);
2220 cfs_rq->h_nr_running++;
2222 if (cfs_rq_throttled(cfs_rq))
2225 update_cfs_load(cfs_rq, 0);
2226 update_cfs_shares(cfs_rq);
2234 static void set_next_buddy(struct sched_entity *se);
2237 * The dequeue_task method is called before nr_running is
2238 * decreased. We remove the task from the rbtree and
2239 * update the fair scheduling stats:
2241 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2243 struct cfs_rq *cfs_rq;
2244 struct sched_entity *se = &p->se;
2245 int task_sleep = flags & DEQUEUE_SLEEP;
2247 for_each_sched_entity(se) {
2248 cfs_rq = cfs_rq_of(se);
2249 dequeue_entity(cfs_rq, se, flags);
2252 * end evaluation on encountering a throttled cfs_rq
2254 * note: in the case of encountering a throttled cfs_rq we will
2255 * post the final h_nr_running decrement below.
2257 if (cfs_rq_throttled(cfs_rq))
2259 cfs_rq->h_nr_running--;
2261 /* Don't dequeue parent if it has other entities besides us */
2262 if (cfs_rq->load.weight) {
2264 * Bias pick_next to pick a task from this cfs_rq, as
2265 * p is sleeping when it is within its sched_slice.
2267 if (task_sleep && parent_entity(se))
2268 set_next_buddy(parent_entity(se));
2270 /* avoid re-evaluating load for this entity */
2271 se = parent_entity(se);
2274 flags |= DEQUEUE_SLEEP;
2277 for_each_sched_entity(se) {
2278 cfs_rq = cfs_rq_of(se);
2279 cfs_rq->h_nr_running--;
2281 if (cfs_rq_throttled(cfs_rq))
2284 update_cfs_load(cfs_rq, 0);
2285 update_cfs_shares(cfs_rq);
2294 /* Used instead of source_load when we know the type == 0 */
2295 static unsigned long weighted_cpuload(const int cpu)
2297 return cpu_rq(cpu)->load.weight;
2301 * Return a low guess at the load of a migration-source cpu weighted
2302 * according to the scheduling class and "nice" value.
2304 * We want to under-estimate the load of migration sources, to
2305 * balance conservatively.
2307 static unsigned long source_load(int cpu, int type)
2309 struct rq *rq = cpu_rq(cpu);
2310 unsigned long total = weighted_cpuload(cpu);
2312 if (type == 0 || !sched_feat(LB_BIAS))
2315 return min(rq->cpu_load[type-1], total);
2319 * Return a high guess at the load of a migration-target cpu weighted
2320 * according to the scheduling class and "nice" value.
2322 static unsigned long target_load(int cpu, int type)
2324 struct rq *rq = cpu_rq(cpu);
2325 unsigned long total = weighted_cpuload(cpu);
2327 if (type == 0 || !sched_feat(LB_BIAS))
2330 return max(rq->cpu_load[type-1], total);
2333 static unsigned long power_of(int cpu)
2335 return cpu_rq(cpu)->cpu_power;
2338 static unsigned long cpu_avg_load_per_task(int cpu)
2340 struct rq *rq = cpu_rq(cpu);
2341 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2344 return rq->load.weight / nr_running;
2350 static void task_waking_fair(struct task_struct *p)
2352 struct sched_entity *se = &p->se;
2353 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2356 #ifndef CONFIG_64BIT
2357 u64 min_vruntime_copy;
2360 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2362 min_vruntime = cfs_rq->min_vruntime;
2363 } while (min_vruntime != min_vruntime_copy);
2365 min_vruntime = cfs_rq->min_vruntime;
2368 se->vruntime -= min_vruntime;
2371 #ifdef CONFIG_FAIR_GROUP_SCHED
2373 * effective_load() calculates the load change as seen from the root_task_group
2375 * Adding load to a group doesn't make a group heavier, but can cause movement
2376 * of group shares between cpus. Assuming the shares were perfectly aligned one
2377 * can calculate the shift in shares.
2379 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2380 * on this @cpu and results in a total addition (subtraction) of @wg to the
2381 * total group weight.
2383 * Given a runqueue weight distribution (rw_i) we can compute a shares
2384 * distribution (s_i) using:
2386 * s_i = rw_i / \Sum rw_j (1)
2388 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2389 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2390 * shares distribution (s_i):
2392 * rw_i = { 2, 4, 1, 0 }
2393 * s_i = { 2/7, 4/7, 1/7, 0 }
2395 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2396 * task used to run on and the CPU the waker is running on), we need to
2397 * compute the effect of waking a task on either CPU and, in case of a sync
2398 * wakeup, compute the effect of the current task going to sleep.
2400 * So for a change of @wl to the local @cpu with an overall group weight change
2401 * of @wl we can compute the new shares distribution (s'_i) using:
2403 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2405 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2406 * differences in waking a task to CPU 0. The additional task changes the
2407 * weight and shares distributions like:
2409 * rw'_i = { 3, 4, 1, 0 }
2410 * s'_i = { 3/8, 4/8, 1/8, 0 }
2412 * We can then compute the difference in effective weight by using:
2414 * dw_i = S * (s'_i - s_i) (3)
2416 * Where 'S' is the group weight as seen by its parent.
2418 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2419 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2420 * 4/7) times the weight of the group.
2422 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2424 struct sched_entity *se = tg->se[cpu];
2426 if (!tg->parent) /* the trivial, non-cgroup case */
2429 for_each_sched_entity(se) {
2435 * W = @wg + \Sum rw_j
2437 W = wg + calc_tg_weight(tg, se->my_q);
2442 w = se->my_q->load.weight + wl;
2445 * wl = S * s'_i; see (2)
2448 wl = (w * tg->shares) / W;
2453 * Per the above, wl is the new se->load.weight value; since
2454 * those are clipped to [MIN_SHARES, ...) do so now. See
2455 * calc_cfs_shares().
2457 if (wl < MIN_SHARES)
2461 * wl = dw_i = S * (s'_i - s_i); see (3)
2463 wl -= se->load.weight;
2466 * Recursively apply this logic to all parent groups to compute
2467 * the final effective load change on the root group. Since
2468 * only the @tg group gets extra weight, all parent groups can
2469 * only redistribute existing shares. @wl is the shift in shares
2470 * resulting from this level per the above.
2479 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2480 unsigned long wl, unsigned long wg)
2487 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2489 s64 this_load, load;
2490 int idx, this_cpu, prev_cpu;
2491 unsigned long tl_per_task;
2492 struct task_group *tg;
2493 unsigned long weight;
2497 this_cpu = smp_processor_id();
2498 prev_cpu = task_cpu(p);
2499 load = source_load(prev_cpu, idx);
2500 this_load = target_load(this_cpu, idx);
2503 * If sync wakeup then subtract the (maximum possible)
2504 * effect of the currently running task from the load
2505 * of the current CPU:
2508 tg = task_group(current);
2509 weight = current->se.load.weight;
2511 this_load += effective_load(tg, this_cpu, -weight, -weight);
2512 load += effective_load(tg, prev_cpu, 0, -weight);
2516 weight = p->se.load.weight;
2519 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2520 * due to the sync cause above having dropped this_load to 0, we'll
2521 * always have an imbalance, but there's really nothing you can do
2522 * about that, so that's good too.
2524 * Otherwise check if either cpus are near enough in load to allow this
2525 * task to be woken on this_cpu.
2527 if (this_load > 0) {
2528 s64 this_eff_load, prev_eff_load;
2530 this_eff_load = 100;
2531 this_eff_load *= power_of(prev_cpu);
2532 this_eff_load *= this_load +
2533 effective_load(tg, this_cpu, weight, weight);
2535 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2536 prev_eff_load *= power_of(this_cpu);
2537 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2539 balanced = this_eff_load <= prev_eff_load;
2544 * If the currently running task will sleep within
2545 * a reasonable amount of time then attract this newly
2548 if (sync && balanced)
2551 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2552 tl_per_task = cpu_avg_load_per_task(this_cpu);
2555 (this_load <= load &&
2556 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2558 * This domain has SD_WAKE_AFFINE and
2559 * p is cache cold in this domain, and
2560 * there is no bad imbalance.
2562 schedstat_inc(sd, ttwu_move_affine);
2563 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2571 * find_idlest_group finds and returns the least busy CPU group within the
2574 static struct sched_group *
2575 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2576 int this_cpu, int load_idx)
2578 struct sched_group *idlest = NULL, *group = sd->groups;
2579 unsigned long min_load = ULONG_MAX, this_load = 0;
2580 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2583 unsigned long load, avg_load;
2587 /* Skip over this group if it has no CPUs allowed */
2588 if (!cpumask_intersects(sched_group_cpus(group),
2589 tsk_cpus_allowed(p)))
2592 local_group = cpumask_test_cpu(this_cpu,
2593 sched_group_cpus(group));
2595 /* Tally up the load of all CPUs in the group */
2598 for_each_cpu(i, sched_group_cpus(group)) {
2599 /* Bias balancing toward cpus of our domain */
2601 load = source_load(i, load_idx);
2603 load = target_load(i, load_idx);
2608 /* Adjust by relative CPU power of the group */
2609 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2612 this_load = avg_load;
2613 } else if (avg_load < min_load) {
2614 min_load = avg_load;
2617 } while (group = group->next, group != sd->groups);
2619 if (!idlest || 100*this_load < imbalance*min_load)
2625 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2628 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2630 unsigned long load, min_load = ULONG_MAX;
2634 /* Traverse only the allowed CPUs */
2635 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2636 load = weighted_cpuload(i);
2638 if (load < min_load || (load == min_load && i == this_cpu)) {
2648 * highest_flag_domain - Return highest sched_domain containing flag.
2649 * @cpu: The cpu whose highest level of sched domain is to
2651 * @flag: The flag to check for the highest sched_domain
2652 * for the given cpu.
2654 * Returns the highest sched_domain of a cpu which contains the given flag.
2656 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
2658 struct sched_domain *sd, *hsd = NULL;
2660 for_each_domain(cpu, sd) {
2661 if (!(sd->flags & flag))
2670 * Try and locate an idle CPU in the sched_domain.
2672 static int select_idle_sibling(struct task_struct *p, int target)
2674 int cpu = smp_processor_id();
2675 int prev_cpu = task_cpu(p);
2676 struct sched_domain *sd;
2677 struct sched_group *sg;
2681 * If the task is going to be woken-up on this cpu and if it is
2682 * already idle, then it is the right target.
2684 if (target == cpu && idle_cpu(cpu))
2688 * If the task is going to be woken-up on the cpu where it previously
2689 * ran and if it is currently idle, then it the right target.
2691 if (target == prev_cpu && idle_cpu(prev_cpu))
2695 * Otherwise, iterate the domains and find an elegible idle cpu.
2699 sd = highest_flag_domain(target, SD_SHARE_PKG_RESOURCES);
2700 for_each_lower_domain(sd) {
2703 if (!cpumask_intersects(sched_group_cpus(sg),
2704 tsk_cpus_allowed(p)))
2707 for_each_cpu(i, sched_group_cpus(sg)) {
2712 target = cpumask_first_and(sched_group_cpus(sg),
2713 tsk_cpus_allowed(p));
2717 } while (sg != sd->groups);
2726 * sched_balance_self: balance the current task (running on cpu) in domains
2727 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2730 * Balance, ie. select the least loaded group.
2732 * Returns the target CPU number, or the same CPU if no balancing is needed.
2734 * preempt must be disabled.
2737 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2739 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2740 int cpu = smp_processor_id();
2741 int prev_cpu = task_cpu(p);
2743 int want_affine = 0;
2745 int sync = wake_flags & WF_SYNC;
2747 if (p->rt.nr_cpus_allowed == 1)
2750 if (sd_flag & SD_BALANCE_WAKE) {
2751 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2757 for_each_domain(cpu, tmp) {
2758 if (!(tmp->flags & SD_LOAD_BALANCE))
2762 * If power savings logic is enabled for a domain, see if we
2763 * are not overloaded, if so, don't balance wider.
2765 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2766 unsigned long power = 0;
2767 unsigned long nr_running = 0;
2768 unsigned long capacity;
2771 for_each_cpu(i, sched_domain_span(tmp)) {
2772 power += power_of(i);
2773 nr_running += cpu_rq(i)->cfs.nr_running;
2776 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2778 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2781 if (nr_running < capacity)
2786 * If both cpu and prev_cpu are part of this domain,
2787 * cpu is a valid SD_WAKE_AFFINE target.
2789 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2790 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2795 if (!want_sd && !want_affine)
2798 if (!(tmp->flags & sd_flag))
2806 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2809 new_cpu = select_idle_sibling(p, prev_cpu);
2814 int load_idx = sd->forkexec_idx;
2815 struct sched_group *group;
2818 if (!(sd->flags & sd_flag)) {
2823 if (sd_flag & SD_BALANCE_WAKE)
2824 load_idx = sd->wake_idx;
2826 group = find_idlest_group(sd, p, cpu, load_idx);
2832 new_cpu = find_idlest_cpu(group, p, cpu);
2833 if (new_cpu == -1 || new_cpu == cpu) {
2834 /* Now try balancing at a lower domain level of cpu */
2839 /* Now try balancing at a lower domain level of new_cpu */
2841 weight = sd->span_weight;
2843 for_each_domain(cpu, tmp) {
2844 if (weight <= tmp->span_weight)
2846 if (tmp->flags & sd_flag)
2849 /* while loop will break here if sd == NULL */
2856 #endif /* CONFIG_SMP */
2858 static unsigned long
2859 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2861 unsigned long gran = sysctl_sched_wakeup_granularity;
2864 * Since its curr running now, convert the gran from real-time
2865 * to virtual-time in his units.
2867 * By using 'se' instead of 'curr' we penalize light tasks, so
2868 * they get preempted easier. That is, if 'se' < 'curr' then
2869 * the resulting gran will be larger, therefore penalizing the
2870 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2871 * be smaller, again penalizing the lighter task.
2873 * This is especially important for buddies when the leftmost
2874 * task is higher priority than the buddy.
2876 return calc_delta_fair(gran, se);
2880 * Should 'se' preempt 'curr'.
2894 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2896 s64 gran, vdiff = curr->vruntime - se->vruntime;
2901 gran = wakeup_gran(curr, se);
2908 static void set_last_buddy(struct sched_entity *se)
2910 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2913 for_each_sched_entity(se)
2914 cfs_rq_of(se)->last = se;
2917 static void set_next_buddy(struct sched_entity *se)
2919 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2922 for_each_sched_entity(se)
2923 cfs_rq_of(se)->next = se;
2926 static void set_skip_buddy(struct sched_entity *se)
2928 for_each_sched_entity(se)
2929 cfs_rq_of(se)->skip = se;
2933 * Preempt the current task with a newly woken task if needed:
2935 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2937 struct task_struct *curr = rq->curr;
2938 struct sched_entity *se = &curr->se, *pse = &p->se;
2939 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2940 int scale = cfs_rq->nr_running >= sched_nr_latency;
2941 int next_buddy_marked = 0;
2943 if (unlikely(se == pse))
2947 * This is possible from callers such as pull_task(), in which we
2948 * unconditionally check_prempt_curr() after an enqueue (which may have
2949 * lead to a throttle). This both saves work and prevents false
2950 * next-buddy nomination below.
2952 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2955 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2956 set_next_buddy(pse);
2957 next_buddy_marked = 1;
2961 * We can come here with TIF_NEED_RESCHED already set from new task
2964 * Note: this also catches the edge-case of curr being in a throttled
2965 * group (e.g. via set_curr_task), since update_curr() (in the
2966 * enqueue of curr) will have resulted in resched being set. This
2967 * prevents us from potentially nominating it as a false LAST_BUDDY
2970 if (test_tsk_need_resched(curr))
2973 /* Idle tasks are by definition preempted by non-idle tasks. */
2974 if (unlikely(curr->policy == SCHED_IDLE) &&
2975 likely(p->policy != SCHED_IDLE))
2979 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2980 * is driven by the tick):
2982 if (unlikely(p->policy != SCHED_NORMAL))
2985 find_matching_se(&se, &pse);
2986 update_curr(cfs_rq_of(se));
2988 if (wakeup_preempt_entity(se, pse) == 1) {
2990 * Bias pick_next to pick the sched entity that is
2991 * triggering this preemption.
2993 if (!next_buddy_marked)
2994 set_next_buddy(pse);
3003 * Only set the backward buddy when the current task is still
3004 * on the rq. This can happen when a wakeup gets interleaved
3005 * with schedule on the ->pre_schedule() or idle_balance()
3006 * point, either of which can * drop the rq lock.
3008 * Also, during early boot the idle thread is in the fair class,
3009 * for obvious reasons its a bad idea to schedule back to it.
3011 if (unlikely(!se->on_rq || curr == rq->idle))
3014 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3018 static struct task_struct *pick_next_task_fair(struct rq *rq)
3020 struct task_struct *p;
3021 struct cfs_rq *cfs_rq = &rq->cfs;
3022 struct sched_entity *se;
3024 if (!cfs_rq->nr_running)
3028 se = pick_next_entity(cfs_rq);
3029 set_next_entity(cfs_rq, se);
3030 cfs_rq = group_cfs_rq(se);
3034 if (hrtick_enabled(rq))
3035 hrtick_start_fair(rq, p);
3041 * Account for a descheduled task:
3043 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3045 struct sched_entity *se = &prev->se;
3046 struct cfs_rq *cfs_rq;
3048 for_each_sched_entity(se) {
3049 cfs_rq = cfs_rq_of(se);
3050 put_prev_entity(cfs_rq, se);
3055 * sched_yield() is very simple
3057 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3059 static void yield_task_fair(struct rq *rq)
3061 struct task_struct *curr = rq->curr;
3062 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3063 struct sched_entity *se = &curr->se;
3066 * Are we the only task in the tree?
3068 if (unlikely(rq->nr_running == 1))
3071 clear_buddies(cfs_rq, se);
3073 if (curr->policy != SCHED_BATCH) {
3074 update_rq_clock(rq);
3076 * Update run-time statistics of the 'current'.
3078 update_curr(cfs_rq);
3080 * Tell update_rq_clock() that we've just updated,
3081 * so we don't do microscopic update in schedule()
3082 * and double the fastpath cost.
3084 rq->skip_clock_update = 1;
3090 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3092 struct sched_entity *se = &p->se;
3094 /* throttled hierarchies are not runnable */
3095 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3098 /* Tell the scheduler that we'd really like pse to run next. */
3101 yield_task_fair(rq);
3107 /**************************************************
3108 * Fair scheduling class load-balancing methods:
3112 * pull_task - move a task from a remote runqueue to the local runqueue.
3113 * Both runqueues must be locked.
3115 static void pull_task(struct rq *src_rq, struct task_struct *p,
3116 struct rq *this_rq, int this_cpu)
3118 deactivate_task(src_rq, p, 0);
3119 set_task_cpu(p, this_cpu);
3120 activate_task(this_rq, p, 0);
3121 check_preempt_curr(this_rq, p, 0);
3125 * Is this task likely cache-hot:
3128 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3132 if (p->sched_class != &fair_sched_class)
3135 if (unlikely(p->policy == SCHED_IDLE))
3139 * Buddy candidates are cache hot:
3141 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3142 (&p->se == cfs_rq_of(&p->se)->next ||
3143 &p->se == cfs_rq_of(&p->se)->last))
3146 if (sysctl_sched_migration_cost == -1)
3148 if (sysctl_sched_migration_cost == 0)
3151 delta = now - p->se.exec_start;
3153 return delta < (s64)sysctl_sched_migration_cost;
3157 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3160 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3161 struct sched_domain *sd, enum cpu_idle_type idle,
3164 int tsk_cache_hot = 0;
3166 * We do not migrate tasks that are:
3167 * 1) running (obviously), or
3168 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3169 * 3) are cache-hot on their current CPU.
3171 if (!cpumask_test_cpu(this_cpu, tsk_cpus_allowed(p))) {
3172 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3177 if (task_running(rq, p)) {
3178 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3183 * Aggressive migration if:
3184 * 1) task is cache cold, or
3185 * 2) too many balance attempts have failed.
3188 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
3189 if (!tsk_cache_hot ||
3190 sd->nr_balance_failed > sd->cache_nice_tries) {
3191 #ifdef CONFIG_SCHEDSTATS
3192 if (tsk_cache_hot) {
3193 schedstat_inc(sd, lb_hot_gained[idle]);
3194 schedstat_inc(p, se.statistics.nr_forced_migrations);
3200 if (tsk_cache_hot) {
3201 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3208 * move_one_task tries to move exactly one task from busiest to this_rq, as
3209 * part of active balancing operations within "domain".
3210 * Returns 1 if successful and 0 otherwise.
3212 * Called with both runqueues locked.
3215 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3216 struct sched_domain *sd, enum cpu_idle_type idle)
3218 struct task_struct *p, *n;
3219 struct cfs_rq *cfs_rq;
3222 for_each_leaf_cfs_rq(busiest, cfs_rq) {
3223 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
3224 if (throttled_lb_pair(task_group(p),
3225 busiest->cpu, this_cpu))
3228 if (!can_migrate_task(p, busiest, this_cpu,
3232 pull_task(busiest, p, this_rq, this_cpu);
3234 * Right now, this is only the second place pull_task()
3235 * is called, so we can safely collect pull_task()
3236 * stats here rather than inside pull_task().
3238 schedstat_inc(sd, lb_gained[idle]);
3246 static unsigned long
3247 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3248 unsigned long max_load_move, struct sched_domain *sd,
3249 enum cpu_idle_type idle, int *all_pinned,
3250 struct cfs_rq *busiest_cfs_rq)
3252 int loops = 0, pulled = 0;
3253 long rem_load_move = max_load_move;
3254 struct task_struct *p, *n;
3256 if (max_load_move == 0)
3259 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
3260 if (loops++ > sysctl_sched_nr_migrate)
3263 if ((p->se.load.weight >> 1) > rem_load_move ||
3264 !can_migrate_task(p, busiest, this_cpu, sd, idle,
3268 pull_task(busiest, p, this_rq, this_cpu);
3270 rem_load_move -= p->se.load.weight;
3272 #ifdef CONFIG_PREEMPT
3274 * NEWIDLE balancing is a source of latency, so preemptible
3275 * kernels will stop after the first task is pulled to minimize
3276 * the critical section.
3278 if (idle == CPU_NEWLY_IDLE)
3283 * We only want to steal up to the prescribed amount of
3286 if (rem_load_move <= 0)
3291 * Right now, this is one of only two places pull_task() is called,
3292 * so we can safely collect pull_task() stats here rather than
3293 * inside pull_task().
3295 schedstat_add(sd, lb_gained[idle], pulled);
3297 return max_load_move - rem_load_move;
3300 #ifdef CONFIG_FAIR_GROUP_SCHED
3302 * update tg->load_weight by folding this cpu's load_avg
3304 static int update_shares_cpu(struct task_group *tg, int cpu)
3306 struct cfs_rq *cfs_rq;
3307 unsigned long flags;
3314 cfs_rq = tg->cfs_rq[cpu];
3316 raw_spin_lock_irqsave(&rq->lock, flags);
3318 update_rq_clock(rq);
3319 update_cfs_load(cfs_rq, 1);
3322 * We need to update shares after updating tg->load_weight in
3323 * order to adjust the weight of groups with long running tasks.
3325 update_cfs_shares(cfs_rq);
3327 raw_spin_unlock_irqrestore(&rq->lock, flags);
3332 static void update_shares(int cpu)
3334 struct cfs_rq *cfs_rq;
3335 struct rq *rq = cpu_rq(cpu);
3339 * Iterates the task_group tree in a bottom up fashion, see
3340 * list_add_leaf_cfs_rq() for details.
3342 for_each_leaf_cfs_rq(rq, cfs_rq) {
3343 /* throttled entities do not contribute to load */
3344 if (throttled_hierarchy(cfs_rq))
3347 update_shares_cpu(cfs_rq->tg, cpu);
3353 * Compute the cpu's hierarchical load factor for each task group.
3354 * This needs to be done in a top-down fashion because the load of a child
3355 * group is a fraction of its parents load.
3357 static int tg_load_down(struct task_group *tg, void *data)
3360 long cpu = (long)data;
3363 load = cpu_rq(cpu)->load.weight;
3365 load = tg->parent->cfs_rq[cpu]->h_load;
3366 load *= tg->se[cpu]->load.weight;
3367 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3370 tg->cfs_rq[cpu]->h_load = load;
3375 static void update_h_load(long cpu)
3377 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3380 static unsigned long
3381 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3382 unsigned long max_load_move,
3383 struct sched_domain *sd, enum cpu_idle_type idle,
3386 long rem_load_move = max_load_move;
3387 struct cfs_rq *busiest_cfs_rq;
3390 update_h_load(cpu_of(busiest));
3392 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
3393 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
3394 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
3395 u64 rem_load, moved_load;
3398 * empty group or part of a throttled hierarchy
3400 if (!busiest_cfs_rq->task_weight ||
3401 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
3404 rem_load = (u64)rem_load_move * busiest_weight;
3405 rem_load = div_u64(rem_load, busiest_h_load + 1);
3407 moved_load = balance_tasks(this_rq, this_cpu, busiest,
3408 rem_load, sd, idle, all_pinned,
3414 moved_load *= busiest_h_load;
3415 moved_load = div_u64(moved_load, busiest_weight + 1);
3417 rem_load_move -= moved_load;
3418 if (rem_load_move < 0)
3423 return max_load_move - rem_load_move;
3426 static inline void update_shares(int cpu)
3430 static unsigned long
3431 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3432 unsigned long max_load_move,
3433 struct sched_domain *sd, enum cpu_idle_type idle,
3436 return balance_tasks(this_rq, this_cpu, busiest,
3437 max_load_move, sd, idle, all_pinned,
3443 * move_tasks tries to move up to max_load_move weighted load from busiest to
3444 * this_rq, as part of a balancing operation within domain "sd".
3445 * Returns 1 if successful and 0 otherwise.
3447 * Called with both runqueues locked.
3449 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3450 unsigned long max_load_move,
3451 struct sched_domain *sd, enum cpu_idle_type idle,
3454 unsigned long total_load_moved = 0, load_moved;
3457 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
3458 max_load_move - total_load_moved,
3459 sd, idle, all_pinned);
3461 total_load_moved += load_moved;
3463 #ifdef CONFIG_PREEMPT
3465 * NEWIDLE balancing is a source of latency, so preemptible
3466 * kernels will stop after the first task is pulled to minimize
3467 * the critical section.
3469 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3472 if (raw_spin_is_contended(&this_rq->lock) ||
3473 raw_spin_is_contended(&busiest->lock))
3476 } while (load_moved && max_load_move > total_load_moved);
3478 return total_load_moved > 0;
3481 /********** Helpers for find_busiest_group ************************/
3483 * sd_lb_stats - Structure to store the statistics of a sched_domain
3484 * during load balancing.
3486 struct sd_lb_stats {
3487 struct sched_group *busiest; /* Busiest group in this sd */
3488 struct sched_group *this; /* Local group in this sd */
3489 unsigned long total_load; /* Total load of all groups in sd */
3490 unsigned long total_pwr; /* Total power of all groups in sd */
3491 unsigned long avg_load; /* Average load across all groups in sd */
3493 /** Statistics of this group */
3494 unsigned long this_load;
3495 unsigned long this_load_per_task;
3496 unsigned long this_nr_running;
3497 unsigned long this_has_capacity;
3498 unsigned int this_idle_cpus;
3500 /* Statistics of the busiest group */
3501 unsigned int busiest_idle_cpus;
3502 unsigned long max_load;
3503 unsigned long busiest_load_per_task;
3504 unsigned long busiest_nr_running;
3505 unsigned long busiest_group_capacity;
3506 unsigned long busiest_has_capacity;
3507 unsigned int busiest_group_weight;
3509 int group_imb; /* Is there imbalance in this sd */
3510 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3511 int power_savings_balance; /* Is powersave balance needed for this sd */
3512 struct sched_group *group_min; /* Least loaded group in sd */
3513 struct sched_group *group_leader; /* Group which relieves group_min */
3514 unsigned long min_load_per_task; /* load_per_task in group_min */
3515 unsigned long leader_nr_running; /* Nr running of group_leader */
3516 unsigned long min_nr_running; /* Nr running of group_min */
3521 * sg_lb_stats - stats of a sched_group required for load_balancing
3523 struct sg_lb_stats {
3524 unsigned long avg_load; /*Avg load across the CPUs of the group */
3525 unsigned long group_load; /* Total load over the CPUs of the group */
3526 unsigned long sum_nr_running; /* Nr tasks running in the group */
3527 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3528 unsigned long group_capacity;
3529 unsigned long idle_cpus;
3530 unsigned long group_weight;
3531 int group_imb; /* Is there an imbalance in the group ? */
3532 int group_has_capacity; /* Is there extra capacity in the group? */
3536 * get_sd_load_idx - Obtain the load index for a given sched domain.
3537 * @sd: The sched_domain whose load_idx is to be obtained.
3538 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3540 static inline int get_sd_load_idx(struct sched_domain *sd,
3541 enum cpu_idle_type idle)
3547 load_idx = sd->busy_idx;
3550 case CPU_NEWLY_IDLE:
3551 load_idx = sd->newidle_idx;
3554 load_idx = sd->idle_idx;
3562 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3564 * init_sd_power_savings_stats - Initialize power savings statistics for
3565 * the given sched_domain, during load balancing.
3567 * @sd: Sched domain whose power-savings statistics are to be initialized.
3568 * @sds: Variable containing the statistics for sd.
3569 * @idle: Idle status of the CPU at which we're performing load-balancing.
3571 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3572 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3575 * Busy processors will not participate in power savings
3578 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3579 sds->power_savings_balance = 0;
3581 sds->power_savings_balance = 1;
3582 sds->min_nr_running = ULONG_MAX;
3583 sds->leader_nr_running = 0;
3588 * update_sd_power_savings_stats - Update the power saving stats for a
3589 * sched_domain while performing load balancing.
3591 * @group: sched_group belonging to the sched_domain under consideration.
3592 * @sds: Variable containing the statistics of the sched_domain
3593 * @local_group: Does group contain the CPU for which we're performing
3595 * @sgs: Variable containing the statistics of the group.
3597 static inline void update_sd_power_savings_stats(struct sched_group *group,
3598 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3601 if (!sds->power_savings_balance)
3605 * If the local group is idle or completely loaded
3606 * no need to do power savings balance at this domain
3608 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3609 !sds->this_nr_running))
3610 sds->power_savings_balance = 0;
3613 * If a group is already running at full capacity or idle,
3614 * don't include that group in power savings calculations
3616 if (!sds->power_savings_balance ||
3617 sgs->sum_nr_running >= sgs->group_capacity ||
3618 !sgs->sum_nr_running)
3622 * Calculate the group which has the least non-idle load.
3623 * This is the group from where we need to pick up the load
3626 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3627 (sgs->sum_nr_running == sds->min_nr_running &&
3628 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3629 sds->group_min = group;
3630 sds->min_nr_running = sgs->sum_nr_running;
3631 sds->min_load_per_task = sgs->sum_weighted_load /
3632 sgs->sum_nr_running;
3636 * Calculate the group which is almost near its
3637 * capacity but still has some space to pick up some load
3638 * from other group and save more power
3640 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3643 if (sgs->sum_nr_running > sds->leader_nr_running ||
3644 (sgs->sum_nr_running == sds->leader_nr_running &&
3645 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3646 sds->group_leader = group;
3647 sds->leader_nr_running = sgs->sum_nr_running;
3652 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3653 * @sds: Variable containing the statistics of the sched_domain
3654 * under consideration.
3655 * @this_cpu: Cpu at which we're currently performing load-balancing.
3656 * @imbalance: Variable to store the imbalance.
3659 * Check if we have potential to perform some power-savings balance.
3660 * If yes, set the busiest group to be the least loaded group in the
3661 * sched_domain, so that it's CPUs can be put to idle.
3663 * Returns 1 if there is potential to perform power-savings balance.
3666 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3667 int this_cpu, unsigned long *imbalance)
3669 if (!sds->power_savings_balance)
3672 if (sds->this != sds->group_leader ||
3673 sds->group_leader == sds->group_min)
3676 *imbalance = sds->min_load_per_task;
3677 sds->busiest = sds->group_min;
3682 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3683 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3684 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3689 static inline void update_sd_power_savings_stats(struct sched_group *group,
3690 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3695 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3696 int this_cpu, unsigned long *imbalance)
3700 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3703 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3705 return SCHED_POWER_SCALE;
3708 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3710 return default_scale_freq_power(sd, cpu);
3713 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3715 unsigned long weight = sd->span_weight;
3716 unsigned long smt_gain = sd->smt_gain;
3723 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3725 return default_scale_smt_power(sd, cpu);
3728 unsigned long scale_rt_power(int cpu)
3730 struct rq *rq = cpu_rq(cpu);
3731 u64 total, available;
3733 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3735 if (unlikely(total < rq->rt_avg)) {
3736 /* Ensures that power won't end up being negative */
3739 available = total - rq->rt_avg;
3742 if (unlikely((s64)total < SCHED_POWER_SCALE))
3743 total = SCHED_POWER_SCALE;
3745 total >>= SCHED_POWER_SHIFT;
3747 return div_u64(available, total);
3750 static void update_cpu_power(struct sched_domain *sd, int cpu)
3752 unsigned long weight = sd->span_weight;
3753 unsigned long power = SCHED_POWER_SCALE;
3754 struct sched_group *sdg = sd->groups;
3756 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3757 if (sched_feat(ARCH_POWER))
3758 power *= arch_scale_smt_power(sd, cpu);
3760 power *= default_scale_smt_power(sd, cpu);
3762 power >>= SCHED_POWER_SHIFT;
3765 sdg->sgp->power_orig = power;
3767 if (sched_feat(ARCH_POWER))
3768 power *= arch_scale_freq_power(sd, cpu);
3770 power *= default_scale_freq_power(sd, cpu);
3772 power >>= SCHED_POWER_SHIFT;
3774 power *= scale_rt_power(cpu);
3775 power >>= SCHED_POWER_SHIFT;
3780 cpu_rq(cpu)->cpu_power = power;
3781 sdg->sgp->power = power;
3784 void update_group_power(struct sched_domain *sd, int cpu)
3786 struct sched_domain *child = sd->child;
3787 struct sched_group *group, *sdg = sd->groups;
3788 unsigned long power;
3791 update_cpu_power(sd, cpu);
3797 group = child->groups;
3799 power += group->sgp->power;
3800 group = group->next;
3801 } while (group != child->groups);
3803 sdg->sgp->power = power;
3807 * Try and fix up capacity for tiny siblings, this is needed when
3808 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3809 * which on its own isn't powerful enough.
3811 * See update_sd_pick_busiest() and check_asym_packing().
3814 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3817 * Only siblings can have significantly less than SCHED_POWER_SCALE
3819 if (!(sd->flags & SD_SHARE_CPUPOWER))
3823 * If ~90% of the cpu_power is still there, we're good.
3825 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3832 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3833 * @sd: The sched_domain whose statistics are to be updated.
3834 * @group: sched_group whose statistics are to be updated.
3835 * @this_cpu: Cpu for which load balance is currently performed.
3836 * @idle: Idle status of this_cpu
3837 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3838 * @local_group: Does group contain this_cpu.
3839 * @cpus: Set of cpus considered for load balancing.
3840 * @balance: Should we balance.
3841 * @sgs: variable to hold the statistics for this group.
3843 static inline void update_sg_lb_stats(struct sched_domain *sd,
3844 struct sched_group *group, int this_cpu,
3845 enum cpu_idle_type idle, int load_idx,
3846 int local_group, const struct cpumask *cpus,
3847 int *balance, struct sg_lb_stats *sgs)
3849 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3851 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3852 unsigned long avg_load_per_task = 0;
3855 balance_cpu = group_first_cpu(group);
3857 /* Tally up the load of all CPUs in the group */
3859 min_cpu_load = ~0UL;
3862 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3863 struct rq *rq = cpu_rq(i);
3865 /* Bias balancing toward cpus of our domain */
3867 if (idle_cpu(i) && !first_idle_cpu) {
3872 load = target_load(i, load_idx);
3874 load = source_load(i, load_idx);
3875 if (load > max_cpu_load) {
3876 max_cpu_load = load;
3877 max_nr_running = rq->nr_running;
3879 if (min_cpu_load > load)
3880 min_cpu_load = load;
3883 sgs->group_load += load;
3884 sgs->sum_nr_running += rq->nr_running;
3885 sgs->sum_weighted_load += weighted_cpuload(i);
3891 * First idle cpu or the first cpu(busiest) in this sched group
3892 * is eligible for doing load balancing at this and above
3893 * domains. In the newly idle case, we will allow all the cpu's
3894 * to do the newly idle load balance.
3896 if (idle != CPU_NEWLY_IDLE && local_group) {
3897 if (balance_cpu != this_cpu) {
3901 update_group_power(sd, this_cpu);
3904 /* Adjust by relative CPU power of the group */
3905 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3908 * Consider the group unbalanced when the imbalance is larger
3909 * than the average weight of a task.
3911 * APZ: with cgroup the avg task weight can vary wildly and
3912 * might not be a suitable number - should we keep a
3913 * normalized nr_running number somewhere that negates
3916 if (sgs->sum_nr_running)
3917 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3919 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3922 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3924 if (!sgs->group_capacity)
3925 sgs->group_capacity = fix_small_capacity(sd, group);
3926 sgs->group_weight = group->group_weight;
3928 if (sgs->group_capacity > sgs->sum_nr_running)
3929 sgs->group_has_capacity = 1;
3933 * update_sd_pick_busiest - return 1 on busiest group
3934 * @sd: sched_domain whose statistics are to be checked
3935 * @sds: sched_domain statistics
3936 * @sg: sched_group candidate to be checked for being the busiest
3937 * @sgs: sched_group statistics
3938 * @this_cpu: the current cpu
3940 * Determine if @sg is a busier group than the previously selected
3943 static bool update_sd_pick_busiest(struct sched_domain *sd,
3944 struct sd_lb_stats *sds,
3945 struct sched_group *sg,
3946 struct sg_lb_stats *sgs,
3949 if (sgs->avg_load <= sds->max_load)
3952 if (sgs->sum_nr_running > sgs->group_capacity)
3959 * ASYM_PACKING needs to move all the work to the lowest
3960 * numbered CPUs in the group, therefore mark all groups
3961 * higher than ourself as busy.
3963 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3964 this_cpu < group_first_cpu(sg)) {
3968 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3976 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3977 * @sd: sched_domain whose statistics are to be updated.
3978 * @this_cpu: Cpu for which load balance is currently performed.
3979 * @idle: Idle status of this_cpu
3980 * @cpus: Set of cpus considered for load balancing.
3981 * @balance: Should we balance.
3982 * @sds: variable to hold the statistics for this sched_domain.
3984 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3985 enum cpu_idle_type idle, const struct cpumask *cpus,
3986 int *balance, struct sd_lb_stats *sds)
3988 struct sched_domain *child = sd->child;
3989 struct sched_group *sg = sd->groups;
3990 struct sg_lb_stats sgs;
3991 int load_idx, prefer_sibling = 0;
3993 if (child && child->flags & SD_PREFER_SIBLING)
3996 init_sd_power_savings_stats(sd, sds, idle);
3997 load_idx = get_sd_load_idx(sd, idle);
4002 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
4003 memset(&sgs, 0, sizeof(sgs));
4004 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
4005 local_group, cpus, balance, &sgs);
4007 if (local_group && !(*balance))
4010 sds->total_load += sgs.group_load;
4011 sds->total_pwr += sg->sgp->power;
4014 * In case the child domain prefers tasks go to siblings
4015 * first, lower the sg capacity to one so that we'll try
4016 * and move all the excess tasks away. We lower the capacity
4017 * of a group only if the local group has the capacity to fit
4018 * these excess tasks, i.e. nr_running < group_capacity. The
4019 * extra check prevents the case where you always pull from the
4020 * heaviest group when it is already under-utilized (possible
4021 * with a large weight task outweighs the tasks on the system).
4023 if (prefer_sibling && !local_group && sds->this_has_capacity)
4024 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4027 sds->this_load = sgs.avg_load;
4029 sds->this_nr_running = sgs.sum_nr_running;
4030 sds->this_load_per_task = sgs.sum_weighted_load;
4031 sds->this_has_capacity = sgs.group_has_capacity;
4032 sds->this_idle_cpus = sgs.idle_cpus;
4033 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
4034 sds->max_load = sgs.avg_load;
4036 sds->busiest_nr_running = sgs.sum_nr_running;
4037 sds->busiest_idle_cpus = sgs.idle_cpus;
4038 sds->busiest_group_capacity = sgs.group_capacity;
4039 sds->busiest_load_per_task = sgs.sum_weighted_load;
4040 sds->busiest_has_capacity = sgs.group_has_capacity;
4041 sds->busiest_group_weight = sgs.group_weight;
4042 sds->group_imb = sgs.group_imb;
4045 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
4047 } while (sg != sd->groups);
4051 * check_asym_packing - Check to see if the group is packed into the
4054 * This is primarily intended to used at the sibling level. Some
4055 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4056 * case of POWER7, it can move to lower SMT modes only when higher
4057 * threads are idle. When in lower SMT modes, the threads will
4058 * perform better since they share less core resources. Hence when we
4059 * have idle threads, we want them to be the higher ones.
4061 * This packing function is run on idle threads. It checks to see if
4062 * the busiest CPU in this domain (core in the P7 case) has a higher
4063 * CPU number than the packing function is being run on. Here we are
4064 * assuming lower CPU number will be equivalent to lower a SMT thread
4067 * Returns 1 when packing is required and a task should be moved to
4068 * this CPU. The amount of the imbalance is returned in *imbalance.
4070 * @sd: The sched_domain whose packing is to be checked.
4071 * @sds: Statistics of the sched_domain which is to be packed
4072 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4073 * @imbalance: returns amount of imbalanced due to packing.
4075 static int check_asym_packing(struct sched_domain *sd,
4076 struct sd_lb_stats *sds,
4077 int this_cpu, unsigned long *imbalance)
4081 if (!(sd->flags & SD_ASYM_PACKING))
4087 busiest_cpu = group_first_cpu(sds->busiest);
4088 if (this_cpu > busiest_cpu)
4091 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
4097 * fix_small_imbalance - Calculate the minor imbalance that exists
4098 * amongst the groups of a sched_domain, during
4100 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4101 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4102 * @imbalance: Variable to store the imbalance.
4104 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
4105 int this_cpu, unsigned long *imbalance)
4107 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4108 unsigned int imbn = 2;
4109 unsigned long scaled_busy_load_per_task;
4111 if (sds->this_nr_running) {
4112 sds->this_load_per_task /= sds->this_nr_running;
4113 if (sds->busiest_load_per_task >
4114 sds->this_load_per_task)
4117 sds->this_load_per_task =
4118 cpu_avg_load_per_task(this_cpu);
4120 scaled_busy_load_per_task = sds->busiest_load_per_task
4121 * SCHED_POWER_SCALE;
4122 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4124 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4125 (scaled_busy_load_per_task * imbn)) {
4126 *imbalance = sds->busiest_load_per_task;
4131 * OK, we don't have enough imbalance to justify moving tasks,
4132 * however we may be able to increase total CPU power used by
4136 pwr_now += sds->busiest->sgp->power *
4137 min(sds->busiest_load_per_task, sds->max_load);
4138 pwr_now += sds->this->sgp->power *
4139 min(sds->this_load_per_task, sds->this_load);
4140 pwr_now /= SCHED_POWER_SCALE;
4142 /* Amount of load we'd subtract */
4143 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4144 sds->busiest->sgp->power;
4145 if (sds->max_load > tmp)
4146 pwr_move += sds->busiest->sgp->power *
4147 min(sds->busiest_load_per_task, sds->max_load - tmp);
4149 /* Amount of load we'd add */
4150 if (sds->max_load * sds->busiest->sgp->power <
4151 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4152 tmp = (sds->max_load * sds->busiest->sgp->power) /
4153 sds->this->sgp->power;
4155 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4156 sds->this->sgp->power;
4157 pwr_move += sds->this->sgp->power *
4158 min(sds->this_load_per_task, sds->this_load + tmp);
4159 pwr_move /= SCHED_POWER_SCALE;
4161 /* Move if we gain throughput */
4162 if (pwr_move > pwr_now)
4163 *imbalance = sds->busiest_load_per_task;
4167 * calculate_imbalance - Calculate the amount of imbalance present within the
4168 * groups of a given sched_domain during load balance.
4169 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4170 * @this_cpu: Cpu for which currently load balance is being performed.
4171 * @imbalance: The variable to store the imbalance.
4173 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4174 unsigned long *imbalance)
4176 unsigned long max_pull, load_above_capacity = ~0UL;
4178 sds->busiest_load_per_task /= sds->busiest_nr_running;
4179 if (sds->group_imb) {
4180 sds->busiest_load_per_task =
4181 min(sds->busiest_load_per_task, sds->avg_load);
4185 * In the presence of smp nice balancing, certain scenarios can have
4186 * max load less than avg load(as we skip the groups at or below
4187 * its cpu_power, while calculating max_load..)
4189 if (sds->max_load < sds->avg_load) {
4191 return fix_small_imbalance(sds, this_cpu, imbalance);
4194 if (!sds->group_imb) {
4196 * Don't want to pull so many tasks that a group would go idle.
4198 load_above_capacity = (sds->busiest_nr_running -
4199 sds->busiest_group_capacity);
4201 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4203 load_above_capacity /= sds->busiest->sgp->power;
4207 * We're trying to get all the cpus to the average_load, so we don't
4208 * want to push ourselves above the average load, nor do we wish to
4209 * reduce the max loaded cpu below the average load. At the same time,
4210 * we also don't want to reduce the group load below the group capacity
4211 * (so that we can implement power-savings policies etc). Thus we look
4212 * for the minimum possible imbalance.
4213 * Be careful of negative numbers as they'll appear as very large values
4214 * with unsigned longs.
4216 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4218 /* How much load to actually move to equalise the imbalance */
4219 *imbalance = min(max_pull * sds->busiest->sgp->power,
4220 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4221 / SCHED_POWER_SCALE;
4224 * if *imbalance is less than the average load per runnable task
4225 * there is no guarantee that any tasks will be moved so we'll have
4226 * a think about bumping its value to force at least one task to be
4229 if (*imbalance < sds->busiest_load_per_task)
4230 return fix_small_imbalance(sds, this_cpu, imbalance);
4234 /******* find_busiest_group() helpers end here *********************/
4237 * find_busiest_group - Returns the busiest group within the sched_domain
4238 * if there is an imbalance. If there isn't an imbalance, and
4239 * the user has opted for power-savings, it returns a group whose
4240 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4241 * such a group exists.
4243 * Also calculates the amount of weighted load which should be moved
4244 * to restore balance.
4246 * @sd: The sched_domain whose busiest group is to be returned.
4247 * @this_cpu: The cpu for which load balancing is currently being performed.
4248 * @imbalance: Variable which stores amount of weighted load which should
4249 * be moved to restore balance/put a group to idle.
4250 * @idle: The idle status of this_cpu.
4251 * @cpus: The set of CPUs under consideration for load-balancing.
4252 * @balance: Pointer to a variable indicating if this_cpu
4253 * is the appropriate cpu to perform load balancing at this_level.
4255 * Returns: - the busiest group if imbalance exists.
4256 * - If no imbalance and user has opted for power-savings balance,
4257 * return the least loaded group whose CPUs can be
4258 * put to idle by rebalancing its tasks onto our group.
4260 static struct sched_group *
4261 find_busiest_group(struct sched_domain *sd, int this_cpu,
4262 unsigned long *imbalance, enum cpu_idle_type idle,
4263 const struct cpumask *cpus, int *balance)
4265 struct sd_lb_stats sds;
4267 memset(&sds, 0, sizeof(sds));
4270 * Compute the various statistics relavent for load balancing at
4273 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
4276 * this_cpu is not the appropriate cpu to perform load balancing at
4282 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
4283 check_asym_packing(sd, &sds, this_cpu, imbalance))
4286 /* There is no busy sibling group to pull tasks from */
4287 if (!sds.busiest || sds.busiest_nr_running == 0)
4290 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4293 * If the busiest group is imbalanced the below checks don't
4294 * work because they assumes all things are equal, which typically
4295 * isn't true due to cpus_allowed constraints and the like.
4300 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4301 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4302 !sds.busiest_has_capacity)
4306 * If the local group is more busy than the selected busiest group
4307 * don't try and pull any tasks.
4309 if (sds.this_load >= sds.max_load)
4313 * Don't pull any tasks if this group is already above the domain
4316 if (sds.this_load >= sds.avg_load)
4319 if (idle == CPU_IDLE) {
4321 * This cpu is idle. If the busiest group load doesn't
4322 * have more tasks than the number of available cpu's and
4323 * there is no imbalance between this and busiest group
4324 * wrt to idle cpu's, it is balanced.
4326 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4327 sds.busiest_nr_running <= sds.busiest_group_weight)
4331 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4332 * imbalance_pct to be conservative.
4334 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4339 /* Looks like there is an imbalance. Compute it */
4340 calculate_imbalance(&sds, this_cpu, imbalance);
4345 * There is no obvious imbalance. But check if we can do some balancing
4348 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4356 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4359 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
4360 enum cpu_idle_type idle, unsigned long imbalance,
4361 const struct cpumask *cpus)
4363 struct rq *busiest = NULL, *rq;
4364 unsigned long max_load = 0;
4367 for_each_cpu(i, sched_group_cpus(group)) {
4368 unsigned long power = power_of(i);
4369 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4374 capacity = fix_small_capacity(sd, group);
4376 if (!cpumask_test_cpu(i, cpus))
4380 wl = weighted_cpuload(i);
4383 * When comparing with imbalance, use weighted_cpuload()
4384 * which is not scaled with the cpu power.
4386 if (capacity && rq->nr_running == 1 && wl > imbalance)
4390 * For the load comparisons with the other cpu's, consider
4391 * the weighted_cpuload() scaled with the cpu power, so that
4392 * the load can be moved away from the cpu that is potentially
4393 * running at a lower capacity.
4395 wl = (wl * SCHED_POWER_SCALE) / power;
4397 if (wl > max_load) {
4407 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4408 * so long as it is large enough.
4410 #define MAX_PINNED_INTERVAL 512
4412 /* Working cpumask for load_balance and load_balance_newidle. */
4413 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4415 static int need_active_balance(struct sched_domain *sd, int idle,
4416 int busiest_cpu, int this_cpu)
4418 if (idle == CPU_NEWLY_IDLE) {
4421 * ASYM_PACKING needs to force migrate tasks from busy but
4422 * higher numbered CPUs in order to pack all tasks in the
4423 * lowest numbered CPUs.
4425 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
4429 * The only task running in a non-idle cpu can be moved to this
4430 * cpu in an attempt to completely freeup the other CPU
4433 * The package power saving logic comes from
4434 * find_busiest_group(). If there are no imbalance, then
4435 * f_b_g() will return NULL. However when sched_mc={1,2} then
4436 * f_b_g() will select a group from which a running task may be
4437 * pulled to this cpu in order to make the other package idle.
4438 * If there is no opportunity to make a package idle and if
4439 * there are no imbalance, then f_b_g() will return NULL and no
4440 * action will be taken in load_balance_newidle().
4442 * Under normal task pull operation due to imbalance, there
4443 * will be more than one task in the source run queue and
4444 * move_tasks() will succeed. ld_moved will be true and this
4445 * active balance code will not be triggered.
4447 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4451 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4454 static int active_load_balance_cpu_stop(void *data);
4457 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4458 * tasks if there is an imbalance.
4460 static int load_balance(int this_cpu, struct rq *this_rq,
4461 struct sched_domain *sd, enum cpu_idle_type idle,
4464 int ld_moved, all_pinned = 0, active_balance = 0;
4465 struct sched_group *group;
4466 unsigned long imbalance;
4468 unsigned long flags;
4469 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4471 cpumask_copy(cpus, cpu_active_mask);
4473 schedstat_inc(sd, lb_count[idle]);
4476 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
4483 schedstat_inc(sd, lb_nobusyg[idle]);
4487 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
4489 schedstat_inc(sd, lb_nobusyq[idle]);
4493 BUG_ON(busiest == this_rq);
4495 schedstat_add(sd, lb_imbalance[idle], imbalance);
4498 if (busiest->nr_running > 1) {
4500 * Attempt to move tasks. If find_busiest_group has found
4501 * an imbalance but busiest->nr_running <= 1, the group is
4502 * still unbalanced. ld_moved simply stays zero, so it is
4503 * correctly treated as an imbalance.
4506 local_irq_save(flags);
4507 double_rq_lock(this_rq, busiest);
4508 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4509 imbalance, sd, idle, &all_pinned);
4510 double_rq_unlock(this_rq, busiest);
4511 local_irq_restore(flags);
4514 * some other cpu did the load balance for us.
4516 if (ld_moved && this_cpu != smp_processor_id())
4517 resched_cpu(this_cpu);
4519 /* All tasks on this runqueue were pinned by CPU affinity */
4520 if (unlikely(all_pinned)) {
4521 cpumask_clear_cpu(cpu_of(busiest), cpus);
4522 if (!cpumask_empty(cpus))
4529 schedstat_inc(sd, lb_failed[idle]);
4531 * Increment the failure counter only on periodic balance.
4532 * We do not want newidle balance, which can be very
4533 * frequent, pollute the failure counter causing
4534 * excessive cache_hot migrations and active balances.
4536 if (idle != CPU_NEWLY_IDLE)
4537 sd->nr_balance_failed++;
4539 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
4540 raw_spin_lock_irqsave(&busiest->lock, flags);
4542 /* don't kick the active_load_balance_cpu_stop,
4543 * if the curr task on busiest cpu can't be
4546 if (!cpumask_test_cpu(this_cpu,
4547 tsk_cpus_allowed(busiest->curr))) {
4548 raw_spin_unlock_irqrestore(&busiest->lock,
4551 goto out_one_pinned;
4555 * ->active_balance synchronizes accesses to
4556 * ->active_balance_work. Once set, it's cleared
4557 * only after active load balance is finished.
4559 if (!busiest->active_balance) {
4560 busiest->active_balance = 1;
4561 busiest->push_cpu = this_cpu;
4564 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4567 stop_one_cpu_nowait(cpu_of(busiest),
4568 active_load_balance_cpu_stop, busiest,
4569 &busiest->active_balance_work);
4572 * We've kicked active balancing, reset the failure
4575 sd->nr_balance_failed = sd->cache_nice_tries+1;
4578 sd->nr_balance_failed = 0;
4580 if (likely(!active_balance)) {
4581 /* We were unbalanced, so reset the balancing interval */
4582 sd->balance_interval = sd->min_interval;
4585 * If we've begun active balancing, start to back off. This
4586 * case may not be covered by the all_pinned logic if there
4587 * is only 1 task on the busy runqueue (because we don't call
4590 if (sd->balance_interval < sd->max_interval)
4591 sd->balance_interval *= 2;
4597 schedstat_inc(sd, lb_balanced[idle]);
4599 sd->nr_balance_failed = 0;
4602 /* tune up the balancing interval */
4603 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4604 (sd->balance_interval < sd->max_interval))
4605 sd->balance_interval *= 2;
4613 * idle_balance is called by schedule() if this_cpu is about to become
4614 * idle. Attempts to pull tasks from other CPUs.
4616 void idle_balance(int this_cpu, struct rq *this_rq)
4618 struct sched_domain *sd;
4619 int pulled_task = 0;
4620 unsigned long next_balance = jiffies + HZ;
4622 this_rq->idle_stamp = this_rq->clock;
4624 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4628 * Drop the rq->lock, but keep IRQ/preempt disabled.
4630 raw_spin_unlock(&this_rq->lock);
4632 update_shares(this_cpu);
4634 for_each_domain(this_cpu, sd) {
4635 unsigned long interval;
4638 if (!(sd->flags & SD_LOAD_BALANCE))
4641 if (sd->flags & SD_BALANCE_NEWIDLE) {
4642 /* If we've pulled tasks over stop searching: */
4643 pulled_task = load_balance(this_cpu, this_rq,
4644 sd, CPU_NEWLY_IDLE, &balance);
4647 interval = msecs_to_jiffies(sd->balance_interval);
4648 if (time_after(next_balance, sd->last_balance + interval))
4649 next_balance = sd->last_balance + interval;
4651 this_rq->idle_stamp = 0;
4657 raw_spin_lock(&this_rq->lock);
4659 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4661 * We are going idle. next_balance may be set based on
4662 * a busy processor. So reset next_balance.
4664 this_rq->next_balance = next_balance;
4669 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4670 * running tasks off the busiest CPU onto idle CPUs. It requires at
4671 * least 1 task to be running on each physical CPU where possible, and
4672 * avoids physical / logical imbalances.
4674 static int active_load_balance_cpu_stop(void *data)
4676 struct rq *busiest_rq = data;
4677 int busiest_cpu = cpu_of(busiest_rq);
4678 int target_cpu = busiest_rq->push_cpu;
4679 struct rq *target_rq = cpu_rq(target_cpu);
4680 struct sched_domain *sd;
4682 raw_spin_lock_irq(&busiest_rq->lock);
4684 /* make sure the requested cpu hasn't gone down in the meantime */
4685 if (unlikely(busiest_cpu != smp_processor_id() ||
4686 !busiest_rq->active_balance))
4689 /* Is there any task to move? */
4690 if (busiest_rq->nr_running <= 1)
4694 * This condition is "impossible", if it occurs
4695 * we need to fix it. Originally reported by
4696 * Bjorn Helgaas on a 128-cpu setup.
4698 BUG_ON(busiest_rq == target_rq);
4700 /* move a task from busiest_rq to target_rq */
4701 double_lock_balance(busiest_rq, target_rq);
4703 /* Search for an sd spanning us and the target CPU. */
4705 for_each_domain(target_cpu, sd) {
4706 if ((sd->flags & SD_LOAD_BALANCE) &&
4707 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4712 schedstat_inc(sd, alb_count);
4714 if (move_one_task(target_rq, target_cpu, busiest_rq,
4716 schedstat_inc(sd, alb_pushed);
4718 schedstat_inc(sd, alb_failed);
4721 double_unlock_balance(busiest_rq, target_rq);
4723 busiest_rq->active_balance = 0;
4724 raw_spin_unlock_irq(&busiest_rq->lock);
4730 * idle load balancing details
4731 * - When one of the busy CPUs notice that there may be an idle rebalancing
4732 * needed, they will kick the idle load balancer, which then does idle
4733 * load balancing for all the idle CPUs.
4736 cpumask_var_t idle_cpus_mask;
4738 unsigned long next_balance; /* in jiffy units */
4739 } nohz ____cacheline_aligned;
4741 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4743 * lowest_flag_domain - Return lowest sched_domain containing flag.
4744 * @cpu: The cpu whose lowest level of sched domain is to
4746 * @flag: The flag to check for the lowest sched_domain
4747 * for the given cpu.
4749 * Returns the lowest sched_domain of a cpu which contains the given flag.
4751 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4753 struct sched_domain *sd;
4755 for_each_domain(cpu, sd)
4756 if (sd->flags & flag)
4763 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4764 * @cpu: The cpu whose domains we're iterating over.
4765 * @sd: variable holding the value of the power_savings_sd
4767 * @flag: The flag to filter the sched_domains to be iterated.
4769 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4770 * set, starting from the lowest sched_domain to the highest.
4772 #define for_each_flag_domain(cpu, sd, flag) \
4773 for (sd = lowest_flag_domain(cpu, flag); \
4774 (sd && (sd->flags & flag)); sd = sd->parent)
4777 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4778 * @cpu: The cpu which is nominating a new idle_load_balancer.
4780 * Returns: Returns the id of the idle load balancer if it exists,
4781 * Else, returns >= nr_cpu_ids.
4783 * This algorithm picks the idle load balancer such that it belongs to a
4784 * semi-idle powersavings sched_domain. The idea is to try and avoid
4785 * completely idle packages/cores just for the purpose of idle load balancing
4786 * when there are other idle cpu's which are better suited for that job.
4788 static int find_new_ilb(int cpu)
4790 int ilb = cpumask_first(nohz.idle_cpus_mask);
4791 struct sched_group *ilbg;
4792 struct sched_domain *sd;
4795 * Have idle load balancer selection from semi-idle packages only
4796 * when power-aware load balancing is enabled
4798 if (!(sched_smt_power_savings || sched_mc_power_savings))
4802 * Optimize for the case when we have no idle CPUs or only one
4803 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4805 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4809 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4813 if (ilbg->group_weight !=
4814 atomic_read(&ilbg->sgp->nr_busy_cpus)) {
4815 ilb = cpumask_first_and(nohz.idle_cpus_mask,
4816 sched_group_cpus(ilbg));
4822 } while (ilbg != sd->groups);
4828 if (ilb < nr_cpu_ids && idle_cpu(ilb))
4833 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4834 static inline int find_new_ilb(int call_cpu)
4841 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4842 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4843 * CPU (if there is one).
4845 static void nohz_balancer_kick(int cpu)
4849 nohz.next_balance++;
4851 ilb_cpu = find_new_ilb(cpu);
4853 if (ilb_cpu >= nr_cpu_ids)
4856 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
4859 * Use smp_send_reschedule() instead of resched_cpu().
4860 * This way we generate a sched IPI on the target cpu which
4861 * is idle. And the softirq performing nohz idle load balance
4862 * will be run before returning from the IPI.
4864 smp_send_reschedule(ilb_cpu);
4868 static inline void set_cpu_sd_state_busy(void)
4870 struct sched_domain *sd;
4871 int cpu = smp_processor_id();
4873 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4875 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
4878 for_each_domain(cpu, sd)
4879 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
4883 void set_cpu_sd_state_idle(void)
4885 struct sched_domain *sd;
4886 int cpu = smp_processor_id();
4888 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4890 set_bit(NOHZ_IDLE, nohz_flags(cpu));
4893 for_each_domain(cpu, sd)
4894 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
4899 * This routine will record that this cpu is going idle with tick stopped.
4900 * This info will be used in performing idle load balancing in the future.
4902 void select_nohz_load_balancer(int stop_tick)
4904 int cpu = smp_processor_id();
4907 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
4910 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4911 atomic_inc(&nohz.nr_cpus);
4912 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4918 static DEFINE_SPINLOCK(balancing);
4920 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4923 * Scale the max load_balance interval with the number of CPUs in the system.
4924 * This trades load-balance latency on larger machines for less cross talk.
4926 void update_max_interval(void)
4928 max_load_balance_interval = HZ*num_online_cpus()/10;
4932 * It checks each scheduling domain to see if it is due to be balanced,
4933 * and initiates a balancing operation if so.
4935 * Balancing parameters are set up in arch_init_sched_domains.
4937 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4940 struct rq *rq = cpu_rq(cpu);
4941 unsigned long interval;
4942 struct sched_domain *sd;
4943 /* Earliest time when we have to do rebalance again */
4944 unsigned long next_balance = jiffies + 60*HZ;
4945 int update_next_balance = 0;
4951 for_each_domain(cpu, sd) {
4952 if (!(sd->flags & SD_LOAD_BALANCE))
4955 interval = sd->balance_interval;
4956 if (idle != CPU_IDLE)
4957 interval *= sd->busy_factor;
4959 /* scale ms to jiffies */
4960 interval = msecs_to_jiffies(interval);
4961 interval = clamp(interval, 1UL, max_load_balance_interval);
4963 need_serialize = sd->flags & SD_SERIALIZE;
4965 if (need_serialize) {
4966 if (!spin_trylock(&balancing))
4970 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4971 if (load_balance(cpu, rq, sd, idle, &balance)) {
4973 * We've pulled tasks over so either we're no
4976 idle = CPU_NOT_IDLE;
4978 sd->last_balance = jiffies;
4981 spin_unlock(&balancing);
4983 if (time_after(next_balance, sd->last_balance + interval)) {
4984 next_balance = sd->last_balance + interval;
4985 update_next_balance = 1;
4989 * Stop the load balance at this level. There is another
4990 * CPU in our sched group which is doing load balancing more
4999 * next_balance will be updated only when there is a need.
5000 * When the cpu is attached to null domain for ex, it will not be
5003 if (likely(update_next_balance))
5004 rq->next_balance = next_balance;
5009 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5010 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5012 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5014 struct rq *this_rq = cpu_rq(this_cpu);
5018 if (idle != CPU_IDLE ||
5019 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5022 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5023 if (balance_cpu == this_cpu || !idle_cpu(this_cpu))
5027 * If this cpu gets work to do, stop the load balancing
5028 * work being done for other cpus. Next load
5029 * balancing owner will pick it up.
5034 raw_spin_lock_irq(&this_rq->lock);
5035 update_rq_clock(this_rq);
5036 update_cpu_load(this_rq);
5037 raw_spin_unlock_irq(&this_rq->lock);
5039 rebalance_domains(balance_cpu, CPU_IDLE);
5041 rq = cpu_rq(balance_cpu);
5042 if (time_after(this_rq->next_balance, rq->next_balance))
5043 this_rq->next_balance = rq->next_balance;
5045 nohz.next_balance = this_rq->next_balance;
5047 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5051 * Current heuristic for kicking the idle load balancer in the presence
5052 * of an idle cpu is the system.
5053 * - This rq has more than one task.
5054 * - At any scheduler domain level, this cpu's scheduler group has multiple
5055 * busy cpu's exceeding the group's power.
5056 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5057 * domain span are idle.
5059 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5061 unsigned long now = jiffies;
5062 struct sched_domain *sd;
5064 if (unlikely(idle_cpu(cpu)))
5068 * We may be recently in ticked or tickless idle mode. At the first
5069 * busy tick after returning from idle, we will update the busy stats.
5071 set_cpu_sd_state_busy();
5072 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5073 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5074 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5075 atomic_dec(&nohz.nr_cpus);
5079 * None are in tickless mode and hence no need for NOHZ idle load
5082 if (likely(!atomic_read(&nohz.nr_cpus)))
5085 if (time_before(now, nohz.next_balance))
5088 if (rq->nr_running >= 2)
5091 for_each_domain(cpu, sd) {
5092 struct sched_group *sg = sd->groups;
5093 struct sched_group_power *sgp = sg->sgp;
5094 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5096 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5099 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5100 && (cpumask_first_and(nohz.idle_cpus_mask,
5101 sched_domain_span(sd)) < cpu))
5104 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5112 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5116 * run_rebalance_domains is triggered when needed from the scheduler tick.
5117 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5119 static void run_rebalance_domains(struct softirq_action *h)
5121 int this_cpu = smp_processor_id();
5122 struct rq *this_rq = cpu_rq(this_cpu);
5123 enum cpu_idle_type idle = this_rq->idle_balance ?
5124 CPU_IDLE : CPU_NOT_IDLE;
5126 rebalance_domains(this_cpu, idle);
5129 * If this cpu has a pending nohz_balance_kick, then do the
5130 * balancing on behalf of the other idle cpus whose ticks are
5133 nohz_idle_balance(this_cpu, idle);
5136 static inline int on_null_domain(int cpu)
5138 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5142 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5144 void trigger_load_balance(struct rq *rq, int cpu)
5146 /* Don't need to rebalance while attached to NULL domain */
5147 if (time_after_eq(jiffies, rq->next_balance) &&
5148 likely(!on_null_domain(cpu)))
5149 raise_softirq(SCHED_SOFTIRQ);
5151 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5152 nohz_balancer_kick(cpu);
5156 static void rq_online_fair(struct rq *rq)
5161 static void rq_offline_fair(struct rq *rq)
5166 #endif /* CONFIG_SMP */
5169 * scheduler tick hitting a task of our scheduling class:
5171 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5173 struct cfs_rq *cfs_rq;
5174 struct sched_entity *se = &curr->se;
5176 for_each_sched_entity(se) {
5177 cfs_rq = cfs_rq_of(se);
5178 entity_tick(cfs_rq, se, queued);
5183 * called on fork with the child task as argument from the parent's context
5184 * - child not yet on the tasklist
5185 * - preemption disabled
5187 static void task_fork_fair(struct task_struct *p)
5189 struct cfs_rq *cfs_rq = task_cfs_rq(current);
5190 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
5191 int this_cpu = smp_processor_id();
5192 struct rq *rq = this_rq();
5193 unsigned long flags;
5195 raw_spin_lock_irqsave(&rq->lock, flags);
5197 update_rq_clock(rq);
5199 if (unlikely(task_cpu(p) != this_cpu)) {
5201 __set_task_cpu(p, this_cpu);
5205 update_curr(cfs_rq);
5208 se->vruntime = curr->vruntime;
5209 place_entity(cfs_rq, se, 1);
5211 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5213 * Upon rescheduling, sched_class::put_prev_task() will place
5214 * 'current' within the tree based on its new key value.
5216 swap(curr->vruntime, se->vruntime);
5217 resched_task(rq->curr);
5220 se->vruntime -= cfs_rq->min_vruntime;
5222 raw_spin_unlock_irqrestore(&rq->lock, flags);
5226 * Priority of the task has changed. Check to see if we preempt
5230 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5236 * Reschedule if we are currently running on this runqueue and
5237 * our priority decreased, or if we are not currently running on
5238 * this runqueue and our priority is higher than the current's
5240 if (rq->curr == p) {
5241 if (p->prio > oldprio)
5242 resched_task(rq->curr);
5244 check_preempt_curr(rq, p, 0);
5247 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5249 struct sched_entity *se = &p->se;
5250 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5253 * Ensure the task's vruntime is normalized, so that when its
5254 * switched back to the fair class the enqueue_entity(.flags=0) will
5255 * do the right thing.
5257 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5258 * have normalized the vruntime, if it was !on_rq, then only when
5259 * the task is sleeping will it still have non-normalized vruntime.
5261 if (!se->on_rq && p->state != TASK_RUNNING) {
5263 * Fix up our vruntime so that the current sleep doesn't
5264 * cause 'unlimited' sleep bonus.
5266 place_entity(cfs_rq, se, 0);
5267 se->vruntime -= cfs_rq->min_vruntime;
5272 * We switched to the sched_fair class.
5274 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5280 * We were most likely switched from sched_rt, so
5281 * kick off the schedule if running, otherwise just see
5282 * if we can still preempt the current task.
5285 resched_task(rq->curr);
5287 check_preempt_curr(rq, p, 0);
5290 /* Account for a task changing its policy or group.
5292 * This routine is mostly called to set cfs_rq->curr field when a task
5293 * migrates between groups/classes.
5295 static void set_curr_task_fair(struct rq *rq)
5297 struct sched_entity *se = &rq->curr->se;
5299 for_each_sched_entity(se) {
5300 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5302 set_next_entity(cfs_rq, se);
5303 /* ensure bandwidth has been allocated on our new cfs_rq */
5304 account_cfs_rq_runtime(cfs_rq, 0);
5308 void init_cfs_rq(struct cfs_rq *cfs_rq)
5310 cfs_rq->tasks_timeline = RB_ROOT;
5311 INIT_LIST_HEAD(&cfs_rq->tasks);
5312 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5313 #ifndef CONFIG_64BIT
5314 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5318 #ifdef CONFIG_FAIR_GROUP_SCHED
5319 static void task_move_group_fair(struct task_struct *p, int on_rq)
5322 * If the task was not on the rq at the time of this cgroup movement
5323 * it must have been asleep, sleeping tasks keep their ->vruntime
5324 * absolute on their old rq until wakeup (needed for the fair sleeper
5325 * bonus in place_entity()).
5327 * If it was on the rq, we've just 'preempted' it, which does convert
5328 * ->vruntime to a relative base.
5330 * Make sure both cases convert their relative position when migrating
5331 * to another cgroup's rq. This does somewhat interfere with the
5332 * fair sleeper stuff for the first placement, but who cares.
5335 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5336 set_task_rq(p, task_cpu(p));
5338 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5341 void free_fair_sched_group(struct task_group *tg)
5345 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5347 for_each_possible_cpu(i) {
5349 kfree(tg->cfs_rq[i]);
5358 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5360 struct cfs_rq *cfs_rq;
5361 struct sched_entity *se;
5364 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5367 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5371 tg->shares = NICE_0_LOAD;
5373 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5375 for_each_possible_cpu(i) {
5376 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5377 GFP_KERNEL, cpu_to_node(i));
5381 se = kzalloc_node(sizeof(struct sched_entity),
5382 GFP_KERNEL, cpu_to_node(i));
5386 init_cfs_rq(cfs_rq);
5387 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5398 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5400 struct rq *rq = cpu_rq(cpu);
5401 unsigned long flags;
5404 * Only empty task groups can be destroyed; so we can speculatively
5405 * check on_list without danger of it being re-added.
5407 if (!tg->cfs_rq[cpu]->on_list)
5410 raw_spin_lock_irqsave(&rq->lock, flags);
5411 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5412 raw_spin_unlock_irqrestore(&rq->lock, flags);
5415 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5416 struct sched_entity *se, int cpu,
5417 struct sched_entity *parent)
5419 struct rq *rq = cpu_rq(cpu);
5424 /* allow initial update_cfs_load() to truncate */
5425 cfs_rq->load_stamp = 1;
5427 init_cfs_rq_runtime(cfs_rq);
5429 tg->cfs_rq[cpu] = cfs_rq;
5432 /* se could be NULL for root_task_group */
5437 se->cfs_rq = &rq->cfs;
5439 se->cfs_rq = parent->my_q;
5442 update_load_set(&se->load, 0);
5443 se->parent = parent;
5446 static DEFINE_MUTEX(shares_mutex);
5448 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5451 unsigned long flags;
5454 * We can't change the weight of the root cgroup.
5459 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5461 mutex_lock(&shares_mutex);
5462 if (tg->shares == shares)
5465 tg->shares = shares;
5466 for_each_possible_cpu(i) {
5467 struct rq *rq = cpu_rq(i);
5468 struct sched_entity *se;
5471 /* Propagate contribution to hierarchy */
5472 raw_spin_lock_irqsave(&rq->lock, flags);
5473 for_each_sched_entity(se)
5474 update_cfs_shares(group_cfs_rq(se));
5475 raw_spin_unlock_irqrestore(&rq->lock, flags);
5479 mutex_unlock(&shares_mutex);
5482 #else /* CONFIG_FAIR_GROUP_SCHED */
5484 void free_fair_sched_group(struct task_group *tg) { }
5486 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5491 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5493 #endif /* CONFIG_FAIR_GROUP_SCHED */
5496 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5498 struct sched_entity *se = &task->se;
5499 unsigned int rr_interval = 0;
5502 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5505 if (rq->cfs.load.weight)
5506 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5512 * All the scheduling class methods:
5514 const struct sched_class fair_sched_class = {
5515 .next = &idle_sched_class,
5516 .enqueue_task = enqueue_task_fair,
5517 .dequeue_task = dequeue_task_fair,
5518 .yield_task = yield_task_fair,
5519 .yield_to_task = yield_to_task_fair,
5521 .check_preempt_curr = check_preempt_wakeup,
5523 .pick_next_task = pick_next_task_fair,
5524 .put_prev_task = put_prev_task_fair,
5527 .select_task_rq = select_task_rq_fair,
5529 .rq_online = rq_online_fair,
5530 .rq_offline = rq_offline_fair,
5532 .task_waking = task_waking_fair,
5535 .set_curr_task = set_curr_task_fair,
5536 .task_tick = task_tick_fair,
5537 .task_fork = task_fork_fair,
5539 .prio_changed = prio_changed_fair,
5540 .switched_from = switched_from_fair,
5541 .switched_to = switched_to_fair,
5543 .get_rr_interval = get_rr_interval_fair,
5545 #ifdef CONFIG_FAIR_GROUP_SCHED
5546 .task_move_group = task_move_group_fair,
5550 #ifdef CONFIG_SCHED_DEBUG
5551 void print_cfs_stats(struct seq_file *m, int cpu)
5553 struct cfs_rq *cfs_rq;
5556 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5557 print_cfs_rq(m, cpu, cfs_rq);
5562 __init void init_sched_fair_class(void)
5565 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5568 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);