1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
26 * Targeted preemption latency for CPU-bound tasks:
28 * NOTE: this latency value is not the same as the concept of
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
31 * based scheduling concepts.
33 * (to see the precise effective timeslice length of your workload,
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
59 unsigned int sysctl_sched_min_granularity = 750000ULL;
60 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
65 static unsigned int sched_nr_latency = 8;
68 * After fork, child runs first. If set to 0 (default) then
69 * parent will (try to) run first.
71 unsigned int sysctl_sched_child_runs_first __read_mostly;
74 * SCHED_OTHER wake-up granularity.
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
83 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
85 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
87 int sched_thermal_decay_shift;
88 static int __init setup_sched_thermal_decay_shift(char *str)
92 if (kstrtoint(str, 0, &_shift))
93 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
95 sched_thermal_decay_shift = clamp(_shift, 0, 10);
98 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
102 * For asym packing, by default the lower numbered CPU has higher priority.
104 int __weak arch_asym_cpu_priority(int cpu)
110 * The margin used when comparing utilization with CPU capacity.
114 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
117 * The margin used when comparing CPU capacities.
118 * is 'cap1' noticeably greater than 'cap2'
122 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
125 #ifdef CONFIG_CFS_BANDWIDTH
127 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
128 * each time a cfs_rq requests quota.
130 * Note: in the case that the slice exceeds the runtime remaining (either due
131 * to consumption or the quota being specified to be smaller than the slice)
132 * we will always only issue the remaining available time.
134 * (default: 5 msec, units: microseconds)
136 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
139 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
145 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
151 static inline void update_load_set(struct load_weight *lw, unsigned long w)
158 * Increase the granularity value when there are more CPUs,
159 * because with more CPUs the 'effective latency' as visible
160 * to users decreases. But the relationship is not linear,
161 * so pick a second-best guess by going with the log2 of the
164 * This idea comes from the SD scheduler of Con Kolivas:
166 static unsigned int get_update_sysctl_factor(void)
168 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
171 switch (sysctl_sched_tunable_scaling) {
172 case SCHED_TUNABLESCALING_NONE:
175 case SCHED_TUNABLESCALING_LINEAR:
178 case SCHED_TUNABLESCALING_LOG:
180 factor = 1 + ilog2(cpus);
187 static void update_sysctl(void)
189 unsigned int factor = get_update_sysctl_factor();
191 #define SET_SYSCTL(name) \
192 (sysctl_##name = (factor) * normalized_sysctl_##name)
193 SET_SYSCTL(sched_min_granularity);
194 SET_SYSCTL(sched_latency);
195 SET_SYSCTL(sched_wakeup_granularity);
199 void __init sched_init_granularity(void)
204 #define WMULT_CONST (~0U)
205 #define WMULT_SHIFT 32
207 static void __update_inv_weight(struct load_weight *lw)
211 if (likely(lw->inv_weight))
214 w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * delta_exec * weight / lw.weight
227 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
229 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
230 * we're guaranteed shift stays positive because inv_weight is guaranteed to
231 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
233 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
234 * weight/lw.weight <= 1, and therefore our shift will also be positive.
236 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
238 u64 fact = scale_load_down(weight);
239 u32 fact_hi = (u32)(fact >> 32);
240 int shift = WMULT_SHIFT;
243 __update_inv_weight(lw);
245 if (unlikely(fact_hi)) {
251 fact = mul_u32_u32(fact, lw->inv_weight);
253 fact_hi = (u32)(fact >> 32);
260 return mul_u64_u32_shr(delta_exec, fact, shift);
264 const struct sched_class fair_sched_class;
266 /**************************************************************
267 * CFS operations on generic schedulable entities:
270 #ifdef CONFIG_FAIR_GROUP_SCHED
272 /* Walk up scheduling entities hierarchy */
273 #define for_each_sched_entity(se) \
274 for (; se; se = se->parent)
276 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
281 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
282 autogroup_path(cfs_rq->tg, path, len);
283 else if (cfs_rq && cfs_rq->tg->css.cgroup)
284 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
286 strlcpy(path, "(null)", len);
289 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
291 struct rq *rq = rq_of(cfs_rq);
292 int cpu = cpu_of(rq);
295 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
300 * Ensure we either appear before our parent (if already
301 * enqueued) or force our parent to appear after us when it is
302 * enqueued. The fact that we always enqueue bottom-up
303 * reduces this to two cases and a special case for the root
304 * cfs_rq. Furthermore, it also means that we will always reset
305 * tmp_alone_branch either when the branch is connected
306 * to a tree or when we reach the top of the tree
308 if (cfs_rq->tg->parent &&
309 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
311 * If parent is already on the list, we add the child
312 * just before. Thanks to circular linked property of
313 * the list, this means to put the child at the tail
314 * of the list that starts by parent.
316 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
317 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
319 * The branch is now connected to its tree so we can
320 * reset tmp_alone_branch to the beginning of the
323 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
327 if (!cfs_rq->tg->parent) {
329 * cfs rq without parent should be put
330 * at the tail of the list.
332 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
333 &rq->leaf_cfs_rq_list);
335 * We have reach the top of a tree so we can reset
336 * tmp_alone_branch to the beginning of the list.
338 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
343 * The parent has not already been added so we want to
344 * make sure that it will be put after us.
345 * tmp_alone_branch points to the begin of the branch
346 * where we will add parent.
348 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
350 * update tmp_alone_branch to points to the new begin
353 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
357 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
359 if (cfs_rq->on_list) {
360 struct rq *rq = rq_of(cfs_rq);
363 * With cfs_rq being unthrottled/throttled during an enqueue,
364 * it can happen the tmp_alone_branch points the a leaf that
365 * we finally want to del. In this case, tmp_alone_branch moves
366 * to the prev element but it will point to rq->leaf_cfs_rq_list
367 * at the end of the enqueue.
369 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
370 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
372 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
377 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
379 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
382 /* Iterate thr' all leaf cfs_rq's on a runqueue */
383 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
384 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
387 /* Do the two (enqueued) entities belong to the same group ? */
388 static inline struct cfs_rq *
389 is_same_group(struct sched_entity *se, struct sched_entity *pse)
391 if (se->cfs_rq == pse->cfs_rq)
397 static inline struct sched_entity *parent_entity(struct sched_entity *se)
403 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
405 int se_depth, pse_depth;
408 * preemption test can be made between sibling entities who are in the
409 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
410 * both tasks until we find their ancestors who are siblings of common
414 /* First walk up until both entities are at same depth */
415 se_depth = (*se)->depth;
416 pse_depth = (*pse)->depth;
418 while (se_depth > pse_depth) {
420 *se = parent_entity(*se);
423 while (pse_depth > se_depth) {
425 *pse = parent_entity(*pse);
428 while (!is_same_group(*se, *pse)) {
429 *se = parent_entity(*se);
430 *pse = parent_entity(*pse);
434 #else /* !CONFIG_FAIR_GROUP_SCHED */
436 #define for_each_sched_entity(se) \
437 for (; se; se = NULL)
439 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
442 strlcpy(path, "(null)", len);
445 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
450 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
454 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
458 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
459 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
461 static inline struct sched_entity *parent_entity(struct sched_entity *se)
467 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
471 #endif /* CONFIG_FAIR_GROUP_SCHED */
473 static __always_inline
474 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
476 /**************************************************************
477 * Scheduling class tree data structure manipulation methods:
480 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
482 s64 delta = (s64)(vruntime - max_vruntime);
484 max_vruntime = vruntime;
489 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
491 s64 delta = (s64)(vruntime - min_vruntime);
493 min_vruntime = vruntime;
498 static inline bool entity_before(struct sched_entity *a,
499 struct sched_entity *b)
501 return (s64)(a->vruntime - b->vruntime) < 0;
504 #define __node_2_se(node) \
505 rb_entry((node), struct sched_entity, run_node)
507 static void update_min_vruntime(struct cfs_rq *cfs_rq)
509 struct sched_entity *curr = cfs_rq->curr;
510 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
512 u64 vruntime = cfs_rq->min_vruntime;
516 vruntime = curr->vruntime;
521 if (leftmost) { /* non-empty tree */
522 struct sched_entity *se = __node_2_se(leftmost);
525 vruntime = se->vruntime;
527 vruntime = min_vruntime(vruntime, se->vruntime);
530 /* ensure we never gain time by being placed backwards. */
531 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
534 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
538 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
540 return entity_before(__node_2_se(a), __node_2_se(b));
544 * Enqueue an entity into the rb-tree:
546 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
548 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
551 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
553 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
556 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
563 return __node_2_se(left);
566 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
568 struct rb_node *next = rb_next(&se->run_node);
573 return __node_2_se(next);
576 #ifdef CONFIG_SCHED_DEBUG
577 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
579 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
584 return __node_2_se(last);
587 /**************************************************************
588 * Scheduling class statistics methods:
591 int sched_update_scaling(void)
593 unsigned int factor = get_update_sysctl_factor();
595 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
596 sysctl_sched_min_granularity);
598 #define WRT_SYSCTL(name) \
599 (normalized_sysctl_##name = sysctl_##name / (factor))
600 WRT_SYSCTL(sched_min_granularity);
601 WRT_SYSCTL(sched_latency);
602 WRT_SYSCTL(sched_wakeup_granularity);
612 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
614 if (unlikely(se->load.weight != NICE_0_LOAD))
615 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
621 * The idea is to set a period in which each task runs once.
623 * When there are too many tasks (sched_nr_latency) we have to stretch
624 * this period because otherwise the slices get too small.
626 * p = (nr <= nl) ? l : l*nr/nl
628 static u64 __sched_period(unsigned long nr_running)
630 if (unlikely(nr_running > sched_nr_latency))
631 return nr_running * sysctl_sched_min_granularity;
633 return sysctl_sched_latency;
637 * We calculate the wall-time slice from the period by taking a part
638 * proportional to the weight.
642 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
644 unsigned int nr_running = cfs_rq->nr_running;
647 if (sched_feat(ALT_PERIOD))
648 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
650 slice = __sched_period(nr_running + !se->on_rq);
652 for_each_sched_entity(se) {
653 struct load_weight *load;
654 struct load_weight lw;
656 cfs_rq = cfs_rq_of(se);
657 load = &cfs_rq->load;
659 if (unlikely(!se->on_rq)) {
662 update_load_add(&lw, se->load.weight);
665 slice = __calc_delta(slice, se->load.weight, load);
668 if (sched_feat(BASE_SLICE))
669 slice = max(slice, (u64)sysctl_sched_min_granularity);
675 * We calculate the vruntime slice of a to-be-inserted task.
679 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
681 return calc_delta_fair(sched_slice(cfs_rq, se), se);
687 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
688 static unsigned long task_h_load(struct task_struct *p);
689 static unsigned long capacity_of(int cpu);
691 /* Give new sched_entity start runnable values to heavy its load in infant time */
692 void init_entity_runnable_average(struct sched_entity *se)
694 struct sched_avg *sa = &se->avg;
696 memset(sa, 0, sizeof(*sa));
699 * Tasks are initialized with full load to be seen as heavy tasks until
700 * they get a chance to stabilize to their real load level.
701 * Group entities are initialized with zero load to reflect the fact that
702 * nothing has been attached to the task group yet.
704 if (entity_is_task(se))
705 sa->load_avg = scale_load_down(se->load.weight);
707 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
710 static void attach_entity_cfs_rq(struct sched_entity *se);
713 * With new tasks being created, their initial util_avgs are extrapolated
714 * based on the cfs_rq's current util_avg:
716 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
718 * However, in many cases, the above util_avg does not give a desired
719 * value. Moreover, the sum of the util_avgs may be divergent, such
720 * as when the series is a harmonic series.
722 * To solve this problem, we also cap the util_avg of successive tasks to
723 * only 1/2 of the left utilization budget:
725 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
727 * where n denotes the nth task and cpu_scale the CPU capacity.
729 * For example, for a CPU with 1024 of capacity, a simplest series from
730 * the beginning would be like:
732 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
733 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
735 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
736 * if util_avg > util_avg_cap.
738 void post_init_entity_util_avg(struct task_struct *p)
740 struct sched_entity *se = &p->se;
741 struct cfs_rq *cfs_rq = cfs_rq_of(se);
742 struct sched_avg *sa = &se->avg;
743 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
744 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
747 if (cfs_rq->avg.util_avg != 0) {
748 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
749 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
751 if (sa->util_avg > cap)
758 sa->runnable_avg = sa->util_avg;
760 if (p->sched_class != &fair_sched_class) {
762 * For !fair tasks do:
764 update_cfs_rq_load_avg(now, cfs_rq);
765 attach_entity_load_avg(cfs_rq, se);
766 switched_from_fair(rq, p);
768 * such that the next switched_to_fair() has the
771 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
775 attach_entity_cfs_rq(se);
778 #else /* !CONFIG_SMP */
779 void init_entity_runnable_average(struct sched_entity *se)
782 void post_init_entity_util_avg(struct task_struct *p)
785 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
788 #endif /* CONFIG_SMP */
791 * Update the current task's runtime statistics.
793 static void update_curr(struct cfs_rq *cfs_rq)
795 struct sched_entity *curr = cfs_rq->curr;
796 u64 now = rq_clock_task(rq_of(cfs_rq));
802 delta_exec = now - curr->exec_start;
803 if (unlikely((s64)delta_exec <= 0))
806 curr->exec_start = now;
808 schedstat_set(curr->statistics.exec_max,
809 max(delta_exec, curr->statistics.exec_max));
811 curr->sum_exec_runtime += delta_exec;
812 schedstat_add(cfs_rq->exec_clock, delta_exec);
814 curr->vruntime += calc_delta_fair(delta_exec, curr);
815 update_min_vruntime(cfs_rq);
817 if (entity_is_task(curr)) {
818 struct task_struct *curtask = task_of(curr);
820 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
821 cgroup_account_cputime(curtask, delta_exec);
822 account_group_exec_runtime(curtask, delta_exec);
825 account_cfs_rq_runtime(cfs_rq, delta_exec);
828 static void update_curr_fair(struct rq *rq)
830 update_curr(cfs_rq_of(&rq->curr->se));
834 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
836 u64 wait_start, prev_wait_start;
838 if (!schedstat_enabled())
841 wait_start = rq_clock(rq_of(cfs_rq));
842 prev_wait_start = schedstat_val(se->statistics.wait_start);
844 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
845 likely(wait_start > prev_wait_start))
846 wait_start -= prev_wait_start;
848 __schedstat_set(se->statistics.wait_start, wait_start);
852 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
854 struct task_struct *p;
857 if (!schedstat_enabled())
861 * When the sched_schedstat changes from 0 to 1, some sched se
862 * maybe already in the runqueue, the se->statistics.wait_start
863 * will be 0.So it will let the delta wrong. We need to avoid this
866 if (unlikely(!schedstat_val(se->statistics.wait_start)))
869 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
871 if (entity_is_task(se)) {
873 if (task_on_rq_migrating(p)) {
875 * Preserve migrating task's wait time so wait_start
876 * time stamp can be adjusted to accumulate wait time
877 * prior to migration.
879 __schedstat_set(se->statistics.wait_start, delta);
882 trace_sched_stat_wait(p, delta);
885 __schedstat_set(se->statistics.wait_max,
886 max(schedstat_val(se->statistics.wait_max), delta));
887 __schedstat_inc(se->statistics.wait_count);
888 __schedstat_add(se->statistics.wait_sum, delta);
889 __schedstat_set(se->statistics.wait_start, 0);
893 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 struct task_struct *tsk = NULL;
896 u64 sleep_start, block_start;
898 if (!schedstat_enabled())
901 sleep_start = schedstat_val(se->statistics.sleep_start);
902 block_start = schedstat_val(se->statistics.block_start);
904 if (entity_is_task(se))
908 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
913 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
914 __schedstat_set(se->statistics.sleep_max, delta);
916 __schedstat_set(se->statistics.sleep_start, 0);
917 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
920 account_scheduler_latency(tsk, delta >> 10, 1);
921 trace_sched_stat_sleep(tsk, delta);
925 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
930 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
931 __schedstat_set(se->statistics.block_max, delta);
933 __schedstat_set(se->statistics.block_start, 0);
934 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
937 if (tsk->in_iowait) {
938 __schedstat_add(se->statistics.iowait_sum, delta);
939 __schedstat_inc(se->statistics.iowait_count);
940 trace_sched_stat_iowait(tsk, delta);
943 trace_sched_stat_blocked(tsk, delta);
946 * Blocking time is in units of nanosecs, so shift by
947 * 20 to get a milliseconds-range estimation of the
948 * amount of time that the task spent sleeping:
950 if (unlikely(prof_on == SLEEP_PROFILING)) {
951 profile_hits(SLEEP_PROFILING,
952 (void *)get_wchan(tsk),
955 account_scheduler_latency(tsk, delta >> 10, 0);
961 * Task is being enqueued - update stats:
964 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
966 if (!schedstat_enabled())
970 * Are we enqueueing a waiting task? (for current tasks
971 * a dequeue/enqueue event is a NOP)
973 if (se != cfs_rq->curr)
974 update_stats_wait_start(cfs_rq, se);
976 if (flags & ENQUEUE_WAKEUP)
977 update_stats_enqueue_sleeper(cfs_rq, se);
981 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
984 if (!schedstat_enabled())
988 * Mark the end of the wait period if dequeueing a
991 if (se != cfs_rq->curr)
992 update_stats_wait_end(cfs_rq, se);
994 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
995 struct task_struct *tsk = task_of(se);
998 /* XXX racy against TTWU */
999 state = READ_ONCE(tsk->__state);
1000 if (state & TASK_INTERRUPTIBLE)
1001 __schedstat_set(se->statistics.sleep_start,
1002 rq_clock(rq_of(cfs_rq)));
1003 if (state & TASK_UNINTERRUPTIBLE)
1004 __schedstat_set(se->statistics.block_start,
1005 rq_clock(rq_of(cfs_rq)));
1010 * We are picking a new current task - update its stats:
1013 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1016 * We are starting a new run period:
1018 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1021 /**************************************************
1022 * Scheduling class queueing methods:
1025 #ifdef CONFIG_NUMA_BALANCING
1027 * Approximate time to scan a full NUMA task in ms. The task scan period is
1028 * calculated based on the tasks virtual memory size and
1029 * numa_balancing_scan_size.
1031 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1032 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1034 /* Portion of address space to scan in MB */
1035 unsigned int sysctl_numa_balancing_scan_size = 256;
1037 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1038 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1041 refcount_t refcount;
1043 spinlock_t lock; /* nr_tasks, tasks */
1048 struct rcu_head rcu;
1049 unsigned long total_faults;
1050 unsigned long max_faults_cpu;
1052 * Faults_cpu is used to decide whether memory should move
1053 * towards the CPU. As a consequence, these stats are weighted
1054 * more by CPU use than by memory faults.
1056 unsigned long *faults_cpu;
1057 unsigned long faults[];
1061 * For functions that can be called in multiple contexts that permit reading
1062 * ->numa_group (see struct task_struct for locking rules).
1064 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1066 return rcu_dereference_check(p->numa_group, p == current ||
1067 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1070 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1072 return rcu_dereference_protected(p->numa_group, p == current);
1075 static inline unsigned long group_faults_priv(struct numa_group *ng);
1076 static inline unsigned long group_faults_shared(struct numa_group *ng);
1078 static unsigned int task_nr_scan_windows(struct task_struct *p)
1080 unsigned long rss = 0;
1081 unsigned long nr_scan_pages;
1084 * Calculations based on RSS as non-present and empty pages are skipped
1085 * by the PTE scanner and NUMA hinting faults should be trapped based
1088 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1089 rss = get_mm_rss(p->mm);
1091 rss = nr_scan_pages;
1093 rss = round_up(rss, nr_scan_pages);
1094 return rss / nr_scan_pages;
1097 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1098 #define MAX_SCAN_WINDOW 2560
1100 static unsigned int task_scan_min(struct task_struct *p)
1102 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1103 unsigned int scan, floor;
1104 unsigned int windows = 1;
1106 if (scan_size < MAX_SCAN_WINDOW)
1107 windows = MAX_SCAN_WINDOW / scan_size;
1108 floor = 1000 / windows;
1110 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1111 return max_t(unsigned int, floor, scan);
1114 static unsigned int task_scan_start(struct task_struct *p)
1116 unsigned long smin = task_scan_min(p);
1117 unsigned long period = smin;
1118 struct numa_group *ng;
1120 /* Scale the maximum scan period with the amount of shared memory. */
1122 ng = rcu_dereference(p->numa_group);
1124 unsigned long shared = group_faults_shared(ng);
1125 unsigned long private = group_faults_priv(ng);
1127 period *= refcount_read(&ng->refcount);
1128 period *= shared + 1;
1129 period /= private + shared + 1;
1133 return max(smin, period);
1136 static unsigned int task_scan_max(struct task_struct *p)
1138 unsigned long smin = task_scan_min(p);
1140 struct numa_group *ng;
1142 /* Watch for min being lower than max due to floor calculations */
1143 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1145 /* Scale the maximum scan period with the amount of shared memory. */
1146 ng = deref_curr_numa_group(p);
1148 unsigned long shared = group_faults_shared(ng);
1149 unsigned long private = group_faults_priv(ng);
1150 unsigned long period = smax;
1152 period *= refcount_read(&ng->refcount);
1153 period *= shared + 1;
1154 period /= private + shared + 1;
1156 smax = max(smax, period);
1159 return max(smin, smax);
1162 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1164 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1165 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1168 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1170 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1171 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1174 /* Shared or private faults. */
1175 #define NR_NUMA_HINT_FAULT_TYPES 2
1177 /* Memory and CPU locality */
1178 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1180 /* Averaged statistics, and temporary buffers. */
1181 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1183 pid_t task_numa_group_id(struct task_struct *p)
1185 struct numa_group *ng;
1189 ng = rcu_dereference(p->numa_group);
1198 * The averaged statistics, shared & private, memory & CPU,
1199 * occupy the first half of the array. The second half of the
1200 * array is for current counters, which are averaged into the
1201 * first set by task_numa_placement.
1203 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1205 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1208 static inline unsigned long task_faults(struct task_struct *p, int nid)
1210 if (!p->numa_faults)
1213 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1214 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1217 static inline unsigned long group_faults(struct task_struct *p, int nid)
1219 struct numa_group *ng = deref_task_numa_group(p);
1224 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1225 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1228 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1230 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1231 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1234 static inline unsigned long group_faults_priv(struct numa_group *ng)
1236 unsigned long faults = 0;
1239 for_each_online_node(node) {
1240 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1246 static inline unsigned long group_faults_shared(struct numa_group *ng)
1248 unsigned long faults = 0;
1251 for_each_online_node(node) {
1252 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1259 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1260 * considered part of a numa group's pseudo-interleaving set. Migrations
1261 * between these nodes are slowed down, to allow things to settle down.
1263 #define ACTIVE_NODE_FRACTION 3
1265 static bool numa_is_active_node(int nid, struct numa_group *ng)
1267 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1270 /* Handle placement on systems where not all nodes are directly connected. */
1271 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1272 int maxdist, bool task)
1274 unsigned long score = 0;
1278 * All nodes are directly connected, and the same distance
1279 * from each other. No need for fancy placement algorithms.
1281 if (sched_numa_topology_type == NUMA_DIRECT)
1285 * This code is called for each node, introducing N^2 complexity,
1286 * which should be ok given the number of nodes rarely exceeds 8.
1288 for_each_online_node(node) {
1289 unsigned long faults;
1290 int dist = node_distance(nid, node);
1293 * The furthest away nodes in the system are not interesting
1294 * for placement; nid was already counted.
1296 if (dist == sched_max_numa_distance || node == nid)
1300 * On systems with a backplane NUMA topology, compare groups
1301 * of nodes, and move tasks towards the group with the most
1302 * memory accesses. When comparing two nodes at distance
1303 * "hoplimit", only nodes closer by than "hoplimit" are part
1304 * of each group. Skip other nodes.
1306 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1310 /* Add up the faults from nearby nodes. */
1312 faults = task_faults(p, node);
1314 faults = group_faults(p, node);
1317 * On systems with a glueless mesh NUMA topology, there are
1318 * no fixed "groups of nodes". Instead, nodes that are not
1319 * directly connected bounce traffic through intermediate
1320 * nodes; a numa_group can occupy any set of nodes.
1321 * The further away a node is, the less the faults count.
1322 * This seems to result in good task placement.
1324 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1325 faults *= (sched_max_numa_distance - dist);
1326 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1336 * These return the fraction of accesses done by a particular task, or
1337 * task group, on a particular numa node. The group weight is given a
1338 * larger multiplier, in order to group tasks together that are almost
1339 * evenly spread out between numa nodes.
1341 static inline unsigned long task_weight(struct task_struct *p, int nid,
1344 unsigned long faults, total_faults;
1346 if (!p->numa_faults)
1349 total_faults = p->total_numa_faults;
1354 faults = task_faults(p, nid);
1355 faults += score_nearby_nodes(p, nid, dist, true);
1357 return 1000 * faults / total_faults;
1360 static inline unsigned long group_weight(struct task_struct *p, int nid,
1363 struct numa_group *ng = deref_task_numa_group(p);
1364 unsigned long faults, total_faults;
1369 total_faults = ng->total_faults;
1374 faults = group_faults(p, nid);
1375 faults += score_nearby_nodes(p, nid, dist, false);
1377 return 1000 * faults / total_faults;
1380 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1381 int src_nid, int dst_cpu)
1383 struct numa_group *ng = deref_curr_numa_group(p);
1384 int dst_nid = cpu_to_node(dst_cpu);
1385 int last_cpupid, this_cpupid;
1387 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1388 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1391 * Allow first faults or private faults to migrate immediately early in
1392 * the lifetime of a task. The magic number 4 is based on waiting for
1393 * two full passes of the "multi-stage node selection" test that is
1396 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1397 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1401 * Multi-stage node selection is used in conjunction with a periodic
1402 * migration fault to build a temporal task<->page relation. By using
1403 * a two-stage filter we remove short/unlikely relations.
1405 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1406 * a task's usage of a particular page (n_p) per total usage of this
1407 * page (n_t) (in a given time-span) to a probability.
1409 * Our periodic faults will sample this probability and getting the
1410 * same result twice in a row, given these samples are fully
1411 * independent, is then given by P(n)^2, provided our sample period
1412 * is sufficiently short compared to the usage pattern.
1414 * This quadric squishes small probabilities, making it less likely we
1415 * act on an unlikely task<->page relation.
1417 if (!cpupid_pid_unset(last_cpupid) &&
1418 cpupid_to_nid(last_cpupid) != dst_nid)
1421 /* Always allow migrate on private faults */
1422 if (cpupid_match_pid(p, last_cpupid))
1425 /* A shared fault, but p->numa_group has not been set up yet. */
1430 * Destination node is much more heavily used than the source
1431 * node? Allow migration.
1433 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1434 ACTIVE_NODE_FRACTION)
1438 * Distribute memory according to CPU & memory use on each node,
1439 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1441 * faults_cpu(dst) 3 faults_cpu(src)
1442 * --------------- * - > ---------------
1443 * faults_mem(dst) 4 faults_mem(src)
1445 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1446 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1450 * 'numa_type' describes the node at the moment of load balancing.
1453 /* The node has spare capacity that can be used to run more tasks. */
1456 * The node is fully used and the tasks don't compete for more CPU
1457 * cycles. Nevertheless, some tasks might wait before running.
1461 * The node is overloaded and can't provide expected CPU cycles to all
1467 /* Cached statistics for all CPUs within a node */
1470 unsigned long runnable;
1472 /* Total compute capacity of CPUs on a node */
1473 unsigned long compute_capacity;
1474 unsigned int nr_running;
1475 unsigned int weight;
1476 enum numa_type node_type;
1480 static inline bool is_core_idle(int cpu)
1482 #ifdef CONFIG_SCHED_SMT
1485 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1497 struct task_numa_env {
1498 struct task_struct *p;
1500 int src_cpu, src_nid;
1501 int dst_cpu, dst_nid;
1503 struct numa_stats src_stats, dst_stats;
1508 struct task_struct *best_task;
1513 static unsigned long cpu_load(struct rq *rq);
1514 static unsigned long cpu_runnable(struct rq *rq);
1515 static unsigned long cpu_util(int cpu);
1516 static inline long adjust_numa_imbalance(int imbalance,
1517 int dst_running, int dst_weight);
1520 numa_type numa_classify(unsigned int imbalance_pct,
1521 struct numa_stats *ns)
1523 if ((ns->nr_running > ns->weight) &&
1524 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1525 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1526 return node_overloaded;
1528 if ((ns->nr_running < ns->weight) ||
1529 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1530 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1531 return node_has_spare;
1533 return node_fully_busy;
1536 #ifdef CONFIG_SCHED_SMT
1537 /* Forward declarations of select_idle_sibling helpers */
1538 static inline bool test_idle_cores(int cpu, bool def);
1539 static inline int numa_idle_core(int idle_core, int cpu)
1541 if (!static_branch_likely(&sched_smt_present) ||
1542 idle_core >= 0 || !test_idle_cores(cpu, false))
1546 * Prefer cores instead of packing HT siblings
1547 * and triggering future load balancing.
1549 if (is_core_idle(cpu))
1555 static inline int numa_idle_core(int idle_core, int cpu)
1562 * Gather all necessary information to make NUMA balancing placement
1563 * decisions that are compatible with standard load balancer. This
1564 * borrows code and logic from update_sg_lb_stats but sharing a
1565 * common implementation is impractical.
1567 static void update_numa_stats(struct task_numa_env *env,
1568 struct numa_stats *ns, int nid,
1571 int cpu, idle_core = -1;
1573 memset(ns, 0, sizeof(*ns));
1577 for_each_cpu(cpu, cpumask_of_node(nid)) {
1578 struct rq *rq = cpu_rq(cpu);
1580 ns->load += cpu_load(rq);
1581 ns->runnable += cpu_runnable(rq);
1582 ns->util += cpu_util(cpu);
1583 ns->nr_running += rq->cfs.h_nr_running;
1584 ns->compute_capacity += capacity_of(cpu);
1586 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1587 if (READ_ONCE(rq->numa_migrate_on) ||
1588 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1591 if (ns->idle_cpu == -1)
1594 idle_core = numa_idle_core(idle_core, cpu);
1599 ns->weight = cpumask_weight(cpumask_of_node(nid));
1601 ns->node_type = numa_classify(env->imbalance_pct, ns);
1604 ns->idle_cpu = idle_core;
1607 static void task_numa_assign(struct task_numa_env *env,
1608 struct task_struct *p, long imp)
1610 struct rq *rq = cpu_rq(env->dst_cpu);
1612 /* Check if run-queue part of active NUMA balance. */
1613 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1615 int start = env->dst_cpu;
1617 /* Find alternative idle CPU. */
1618 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1619 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1620 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1625 rq = cpu_rq(env->dst_cpu);
1626 if (!xchg(&rq->numa_migrate_on, 1))
1630 /* Failed to find an alternative idle CPU */
1636 * Clear previous best_cpu/rq numa-migrate flag, since task now
1637 * found a better CPU to move/swap.
1639 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1640 rq = cpu_rq(env->best_cpu);
1641 WRITE_ONCE(rq->numa_migrate_on, 0);
1645 put_task_struct(env->best_task);
1650 env->best_imp = imp;
1651 env->best_cpu = env->dst_cpu;
1654 static bool load_too_imbalanced(long src_load, long dst_load,
1655 struct task_numa_env *env)
1658 long orig_src_load, orig_dst_load;
1659 long src_capacity, dst_capacity;
1662 * The load is corrected for the CPU capacity available on each node.
1665 * ------------ vs ---------
1666 * src_capacity dst_capacity
1668 src_capacity = env->src_stats.compute_capacity;
1669 dst_capacity = env->dst_stats.compute_capacity;
1671 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1673 orig_src_load = env->src_stats.load;
1674 orig_dst_load = env->dst_stats.load;
1676 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1678 /* Would this change make things worse? */
1679 return (imb > old_imb);
1683 * Maximum NUMA importance can be 1998 (2*999);
1684 * SMALLIMP @ 30 would be close to 1998/64.
1685 * Used to deter task migration.
1690 * This checks if the overall compute and NUMA accesses of the system would
1691 * be improved if the source tasks was migrated to the target dst_cpu taking
1692 * into account that it might be best if task running on the dst_cpu should
1693 * be exchanged with the source task
1695 static bool task_numa_compare(struct task_numa_env *env,
1696 long taskimp, long groupimp, bool maymove)
1698 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1699 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1700 long imp = p_ng ? groupimp : taskimp;
1701 struct task_struct *cur;
1702 long src_load, dst_load;
1703 int dist = env->dist;
1706 bool stopsearch = false;
1708 if (READ_ONCE(dst_rq->numa_migrate_on))
1712 cur = rcu_dereference(dst_rq->curr);
1713 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1717 * Because we have preemption enabled we can get migrated around and
1718 * end try selecting ourselves (current == env->p) as a swap candidate.
1720 if (cur == env->p) {
1726 if (maymove && moveimp >= env->best_imp)
1732 /* Skip this swap candidate if cannot move to the source cpu. */
1733 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1737 * Skip this swap candidate if it is not moving to its preferred
1738 * node and the best task is.
1740 if (env->best_task &&
1741 env->best_task->numa_preferred_nid == env->src_nid &&
1742 cur->numa_preferred_nid != env->src_nid) {
1747 * "imp" is the fault differential for the source task between the
1748 * source and destination node. Calculate the total differential for
1749 * the source task and potential destination task. The more negative
1750 * the value is, the more remote accesses that would be expected to
1751 * be incurred if the tasks were swapped.
1753 * If dst and source tasks are in the same NUMA group, or not
1754 * in any group then look only at task weights.
1756 cur_ng = rcu_dereference(cur->numa_group);
1757 if (cur_ng == p_ng) {
1758 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1759 task_weight(cur, env->dst_nid, dist);
1761 * Add some hysteresis to prevent swapping the
1762 * tasks within a group over tiny differences.
1768 * Compare the group weights. If a task is all by itself
1769 * (not part of a group), use the task weight instead.
1772 imp += group_weight(cur, env->src_nid, dist) -
1773 group_weight(cur, env->dst_nid, dist);
1775 imp += task_weight(cur, env->src_nid, dist) -
1776 task_weight(cur, env->dst_nid, dist);
1779 /* Discourage picking a task already on its preferred node */
1780 if (cur->numa_preferred_nid == env->dst_nid)
1784 * Encourage picking a task that moves to its preferred node.
1785 * This potentially makes imp larger than it's maximum of
1786 * 1998 (see SMALLIMP and task_weight for why) but in this
1787 * case, it does not matter.
1789 if (cur->numa_preferred_nid == env->src_nid)
1792 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1799 * Prefer swapping with a task moving to its preferred node over a
1802 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1803 env->best_task->numa_preferred_nid != env->src_nid) {
1808 * If the NUMA importance is less than SMALLIMP,
1809 * task migration might only result in ping pong
1810 * of tasks and also hurt performance due to cache
1813 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1817 * In the overloaded case, try and keep the load balanced.
1819 load = task_h_load(env->p) - task_h_load(cur);
1823 dst_load = env->dst_stats.load + load;
1824 src_load = env->src_stats.load - load;
1826 if (load_too_imbalanced(src_load, dst_load, env))
1830 /* Evaluate an idle CPU for a task numa move. */
1832 int cpu = env->dst_stats.idle_cpu;
1834 /* Nothing cached so current CPU went idle since the search. */
1839 * If the CPU is no longer truly idle and the previous best CPU
1840 * is, keep using it.
1842 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1843 idle_cpu(env->best_cpu)) {
1844 cpu = env->best_cpu;
1850 task_numa_assign(env, cur, imp);
1853 * If a move to idle is allowed because there is capacity or load
1854 * balance improves then stop the search. While a better swap
1855 * candidate may exist, a search is not free.
1857 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1861 * If a swap candidate must be identified and the current best task
1862 * moves its preferred node then stop the search.
1864 if (!maymove && env->best_task &&
1865 env->best_task->numa_preferred_nid == env->src_nid) {
1874 static void task_numa_find_cpu(struct task_numa_env *env,
1875 long taskimp, long groupimp)
1877 bool maymove = false;
1881 * If dst node has spare capacity, then check if there is an
1882 * imbalance that would be overruled by the load balancer.
1884 if (env->dst_stats.node_type == node_has_spare) {
1885 unsigned int imbalance;
1886 int src_running, dst_running;
1889 * Would movement cause an imbalance? Note that if src has
1890 * more running tasks that the imbalance is ignored as the
1891 * move improves the imbalance from the perspective of the
1892 * CPU load balancer.
1894 src_running = env->src_stats.nr_running - 1;
1895 dst_running = env->dst_stats.nr_running + 1;
1896 imbalance = max(0, dst_running - src_running);
1897 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1898 env->dst_stats.weight);
1900 /* Use idle CPU if there is no imbalance */
1903 if (env->dst_stats.idle_cpu >= 0) {
1904 env->dst_cpu = env->dst_stats.idle_cpu;
1905 task_numa_assign(env, NULL, 0);
1910 long src_load, dst_load, load;
1912 * If the improvement from just moving env->p direction is better
1913 * than swapping tasks around, check if a move is possible.
1915 load = task_h_load(env->p);
1916 dst_load = env->dst_stats.load + load;
1917 src_load = env->src_stats.load - load;
1918 maymove = !load_too_imbalanced(src_load, dst_load, env);
1921 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1922 /* Skip this CPU if the source task cannot migrate */
1923 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1927 if (task_numa_compare(env, taskimp, groupimp, maymove))
1932 static int task_numa_migrate(struct task_struct *p)
1934 struct task_numa_env env = {
1937 .src_cpu = task_cpu(p),
1938 .src_nid = task_node(p),
1940 .imbalance_pct = 112,
1946 unsigned long taskweight, groupweight;
1947 struct sched_domain *sd;
1948 long taskimp, groupimp;
1949 struct numa_group *ng;
1954 * Pick the lowest SD_NUMA domain, as that would have the smallest
1955 * imbalance and would be the first to start moving tasks about.
1957 * And we want to avoid any moving of tasks about, as that would create
1958 * random movement of tasks -- counter the numa conditions we're trying
1962 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1964 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1968 * Cpusets can break the scheduler domain tree into smaller
1969 * balance domains, some of which do not cross NUMA boundaries.
1970 * Tasks that are "trapped" in such domains cannot be migrated
1971 * elsewhere, so there is no point in (re)trying.
1973 if (unlikely(!sd)) {
1974 sched_setnuma(p, task_node(p));
1978 env.dst_nid = p->numa_preferred_nid;
1979 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1980 taskweight = task_weight(p, env.src_nid, dist);
1981 groupweight = group_weight(p, env.src_nid, dist);
1982 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
1983 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1984 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1985 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
1987 /* Try to find a spot on the preferred nid. */
1988 task_numa_find_cpu(&env, taskimp, groupimp);
1991 * Look at other nodes in these cases:
1992 * - there is no space available on the preferred_nid
1993 * - the task is part of a numa_group that is interleaved across
1994 * multiple NUMA nodes; in order to better consolidate the group,
1995 * we need to check other locations.
1997 ng = deref_curr_numa_group(p);
1998 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1999 for_each_online_node(nid) {
2000 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2003 dist = node_distance(env.src_nid, env.dst_nid);
2004 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2006 taskweight = task_weight(p, env.src_nid, dist);
2007 groupweight = group_weight(p, env.src_nid, dist);
2010 /* Only consider nodes where both task and groups benefit */
2011 taskimp = task_weight(p, nid, dist) - taskweight;
2012 groupimp = group_weight(p, nid, dist) - groupweight;
2013 if (taskimp < 0 && groupimp < 0)
2018 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2019 task_numa_find_cpu(&env, taskimp, groupimp);
2024 * If the task is part of a workload that spans multiple NUMA nodes,
2025 * and is migrating into one of the workload's active nodes, remember
2026 * this node as the task's preferred numa node, so the workload can
2028 * A task that migrated to a second choice node will be better off
2029 * trying for a better one later. Do not set the preferred node here.
2032 if (env.best_cpu == -1)
2035 nid = cpu_to_node(env.best_cpu);
2037 if (nid != p->numa_preferred_nid)
2038 sched_setnuma(p, nid);
2041 /* No better CPU than the current one was found. */
2042 if (env.best_cpu == -1) {
2043 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2047 best_rq = cpu_rq(env.best_cpu);
2048 if (env.best_task == NULL) {
2049 ret = migrate_task_to(p, env.best_cpu);
2050 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2052 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2056 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2057 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2060 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2061 put_task_struct(env.best_task);
2065 /* Attempt to migrate a task to a CPU on the preferred node. */
2066 static void numa_migrate_preferred(struct task_struct *p)
2068 unsigned long interval = HZ;
2070 /* This task has no NUMA fault statistics yet */
2071 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2074 /* Periodically retry migrating the task to the preferred node */
2075 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2076 p->numa_migrate_retry = jiffies + interval;
2078 /* Success if task is already running on preferred CPU */
2079 if (task_node(p) == p->numa_preferred_nid)
2082 /* Otherwise, try migrate to a CPU on the preferred node */
2083 task_numa_migrate(p);
2087 * Find out how many nodes on the workload is actively running on. Do this by
2088 * tracking the nodes from which NUMA hinting faults are triggered. This can
2089 * be different from the set of nodes where the workload's memory is currently
2092 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2094 unsigned long faults, max_faults = 0;
2095 int nid, active_nodes = 0;
2097 for_each_online_node(nid) {
2098 faults = group_faults_cpu(numa_group, nid);
2099 if (faults > max_faults)
2100 max_faults = faults;
2103 for_each_online_node(nid) {
2104 faults = group_faults_cpu(numa_group, nid);
2105 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2109 numa_group->max_faults_cpu = max_faults;
2110 numa_group->active_nodes = active_nodes;
2114 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2115 * increments. The more local the fault statistics are, the higher the scan
2116 * period will be for the next scan window. If local/(local+remote) ratio is
2117 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2118 * the scan period will decrease. Aim for 70% local accesses.
2120 #define NUMA_PERIOD_SLOTS 10
2121 #define NUMA_PERIOD_THRESHOLD 7
2124 * Increase the scan period (slow down scanning) if the majority of
2125 * our memory is already on our local node, or if the majority of
2126 * the page accesses are shared with other processes.
2127 * Otherwise, decrease the scan period.
2129 static void update_task_scan_period(struct task_struct *p,
2130 unsigned long shared, unsigned long private)
2132 unsigned int period_slot;
2133 int lr_ratio, ps_ratio;
2136 unsigned long remote = p->numa_faults_locality[0];
2137 unsigned long local = p->numa_faults_locality[1];
2140 * If there were no record hinting faults then either the task is
2141 * completely idle or all activity is areas that are not of interest
2142 * to automatic numa balancing. Related to that, if there were failed
2143 * migration then it implies we are migrating too quickly or the local
2144 * node is overloaded. In either case, scan slower
2146 if (local + shared == 0 || p->numa_faults_locality[2]) {
2147 p->numa_scan_period = min(p->numa_scan_period_max,
2148 p->numa_scan_period << 1);
2150 p->mm->numa_next_scan = jiffies +
2151 msecs_to_jiffies(p->numa_scan_period);
2157 * Prepare to scale scan period relative to the current period.
2158 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2159 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2160 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2162 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2163 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2164 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2166 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2168 * Most memory accesses are local. There is no need to
2169 * do fast NUMA scanning, since memory is already local.
2171 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2174 diff = slot * period_slot;
2175 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2177 * Most memory accesses are shared with other tasks.
2178 * There is no point in continuing fast NUMA scanning,
2179 * since other tasks may just move the memory elsewhere.
2181 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2184 diff = slot * period_slot;
2187 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2188 * yet they are not on the local NUMA node. Speed up
2189 * NUMA scanning to get the memory moved over.
2191 int ratio = max(lr_ratio, ps_ratio);
2192 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2195 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2196 task_scan_min(p), task_scan_max(p));
2197 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2201 * Get the fraction of time the task has been running since the last
2202 * NUMA placement cycle. The scheduler keeps similar statistics, but
2203 * decays those on a 32ms period, which is orders of magnitude off
2204 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2205 * stats only if the task is so new there are no NUMA statistics yet.
2207 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2209 u64 runtime, delta, now;
2210 /* Use the start of this time slice to avoid calculations. */
2211 now = p->se.exec_start;
2212 runtime = p->se.sum_exec_runtime;
2214 if (p->last_task_numa_placement) {
2215 delta = runtime - p->last_sum_exec_runtime;
2216 *period = now - p->last_task_numa_placement;
2218 /* Avoid time going backwards, prevent potential divide error: */
2219 if (unlikely((s64)*period < 0))
2222 delta = p->se.avg.load_sum;
2223 *period = LOAD_AVG_MAX;
2226 p->last_sum_exec_runtime = runtime;
2227 p->last_task_numa_placement = now;
2233 * Determine the preferred nid for a task in a numa_group. This needs to
2234 * be done in a way that produces consistent results with group_weight,
2235 * otherwise workloads might not converge.
2237 static int preferred_group_nid(struct task_struct *p, int nid)
2242 /* Direct connections between all NUMA nodes. */
2243 if (sched_numa_topology_type == NUMA_DIRECT)
2247 * On a system with glueless mesh NUMA topology, group_weight
2248 * scores nodes according to the number of NUMA hinting faults on
2249 * both the node itself, and on nearby nodes.
2251 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2252 unsigned long score, max_score = 0;
2253 int node, max_node = nid;
2255 dist = sched_max_numa_distance;
2257 for_each_online_node(node) {
2258 score = group_weight(p, node, dist);
2259 if (score > max_score) {
2268 * Finding the preferred nid in a system with NUMA backplane
2269 * interconnect topology is more involved. The goal is to locate
2270 * tasks from numa_groups near each other in the system, and
2271 * untangle workloads from different sides of the system. This requires
2272 * searching down the hierarchy of node groups, recursively searching
2273 * inside the highest scoring group of nodes. The nodemask tricks
2274 * keep the complexity of the search down.
2276 nodes = node_online_map;
2277 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2278 unsigned long max_faults = 0;
2279 nodemask_t max_group = NODE_MASK_NONE;
2282 /* Are there nodes at this distance from each other? */
2283 if (!find_numa_distance(dist))
2286 for_each_node_mask(a, nodes) {
2287 unsigned long faults = 0;
2288 nodemask_t this_group;
2289 nodes_clear(this_group);
2291 /* Sum group's NUMA faults; includes a==b case. */
2292 for_each_node_mask(b, nodes) {
2293 if (node_distance(a, b) < dist) {
2294 faults += group_faults(p, b);
2295 node_set(b, this_group);
2296 node_clear(b, nodes);
2300 /* Remember the top group. */
2301 if (faults > max_faults) {
2302 max_faults = faults;
2303 max_group = this_group;
2305 * subtle: at the smallest distance there is
2306 * just one node left in each "group", the
2307 * winner is the preferred nid.
2312 /* Next round, evaluate the nodes within max_group. */
2320 static void task_numa_placement(struct task_struct *p)
2322 int seq, nid, max_nid = NUMA_NO_NODE;
2323 unsigned long max_faults = 0;
2324 unsigned long fault_types[2] = { 0, 0 };
2325 unsigned long total_faults;
2326 u64 runtime, period;
2327 spinlock_t *group_lock = NULL;
2328 struct numa_group *ng;
2331 * The p->mm->numa_scan_seq field gets updated without
2332 * exclusive access. Use READ_ONCE() here to ensure
2333 * that the field is read in a single access:
2335 seq = READ_ONCE(p->mm->numa_scan_seq);
2336 if (p->numa_scan_seq == seq)
2338 p->numa_scan_seq = seq;
2339 p->numa_scan_period_max = task_scan_max(p);
2341 total_faults = p->numa_faults_locality[0] +
2342 p->numa_faults_locality[1];
2343 runtime = numa_get_avg_runtime(p, &period);
2345 /* If the task is part of a group prevent parallel updates to group stats */
2346 ng = deref_curr_numa_group(p);
2348 group_lock = &ng->lock;
2349 spin_lock_irq(group_lock);
2352 /* Find the node with the highest number of faults */
2353 for_each_online_node(nid) {
2354 /* Keep track of the offsets in numa_faults array */
2355 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2356 unsigned long faults = 0, group_faults = 0;
2359 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2360 long diff, f_diff, f_weight;
2362 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2363 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2364 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2365 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2367 /* Decay existing window, copy faults since last scan */
2368 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2369 fault_types[priv] += p->numa_faults[membuf_idx];
2370 p->numa_faults[membuf_idx] = 0;
2373 * Normalize the faults_from, so all tasks in a group
2374 * count according to CPU use, instead of by the raw
2375 * number of faults. Tasks with little runtime have
2376 * little over-all impact on throughput, and thus their
2377 * faults are less important.
2379 f_weight = div64_u64(runtime << 16, period + 1);
2380 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2382 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2383 p->numa_faults[cpubuf_idx] = 0;
2385 p->numa_faults[mem_idx] += diff;
2386 p->numa_faults[cpu_idx] += f_diff;
2387 faults += p->numa_faults[mem_idx];
2388 p->total_numa_faults += diff;
2391 * safe because we can only change our own group
2393 * mem_idx represents the offset for a given
2394 * nid and priv in a specific region because it
2395 * is at the beginning of the numa_faults array.
2397 ng->faults[mem_idx] += diff;
2398 ng->faults_cpu[mem_idx] += f_diff;
2399 ng->total_faults += diff;
2400 group_faults += ng->faults[mem_idx];
2405 if (faults > max_faults) {
2406 max_faults = faults;
2409 } else if (group_faults > max_faults) {
2410 max_faults = group_faults;
2416 numa_group_count_active_nodes(ng);
2417 spin_unlock_irq(group_lock);
2418 max_nid = preferred_group_nid(p, max_nid);
2422 /* Set the new preferred node */
2423 if (max_nid != p->numa_preferred_nid)
2424 sched_setnuma(p, max_nid);
2427 update_task_scan_period(p, fault_types[0], fault_types[1]);
2430 static inline int get_numa_group(struct numa_group *grp)
2432 return refcount_inc_not_zero(&grp->refcount);
2435 static inline void put_numa_group(struct numa_group *grp)
2437 if (refcount_dec_and_test(&grp->refcount))
2438 kfree_rcu(grp, rcu);
2441 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2444 struct numa_group *grp, *my_grp;
2445 struct task_struct *tsk;
2447 int cpu = cpupid_to_cpu(cpupid);
2450 if (unlikely(!deref_curr_numa_group(p))) {
2451 unsigned int size = sizeof(struct numa_group) +
2452 4*nr_node_ids*sizeof(unsigned long);
2454 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2458 refcount_set(&grp->refcount, 1);
2459 grp->active_nodes = 1;
2460 grp->max_faults_cpu = 0;
2461 spin_lock_init(&grp->lock);
2463 /* Second half of the array tracks nids where faults happen */
2464 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2467 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2468 grp->faults[i] = p->numa_faults[i];
2470 grp->total_faults = p->total_numa_faults;
2473 rcu_assign_pointer(p->numa_group, grp);
2477 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2479 if (!cpupid_match_pid(tsk, cpupid))
2482 grp = rcu_dereference(tsk->numa_group);
2486 my_grp = deref_curr_numa_group(p);
2491 * Only join the other group if its bigger; if we're the bigger group,
2492 * the other task will join us.
2494 if (my_grp->nr_tasks > grp->nr_tasks)
2498 * Tie-break on the grp address.
2500 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2503 /* Always join threads in the same process. */
2504 if (tsk->mm == current->mm)
2507 /* Simple filter to avoid false positives due to PID collisions */
2508 if (flags & TNF_SHARED)
2511 /* Update priv based on whether false sharing was detected */
2514 if (join && !get_numa_group(grp))
2522 BUG_ON(irqs_disabled());
2523 double_lock_irq(&my_grp->lock, &grp->lock);
2525 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2526 my_grp->faults[i] -= p->numa_faults[i];
2527 grp->faults[i] += p->numa_faults[i];
2529 my_grp->total_faults -= p->total_numa_faults;
2530 grp->total_faults += p->total_numa_faults;
2535 spin_unlock(&my_grp->lock);
2536 spin_unlock_irq(&grp->lock);
2538 rcu_assign_pointer(p->numa_group, grp);
2540 put_numa_group(my_grp);
2549 * Get rid of NUMA statistics associated with a task (either current or dead).
2550 * If @final is set, the task is dead and has reached refcount zero, so we can
2551 * safely free all relevant data structures. Otherwise, there might be
2552 * concurrent reads from places like load balancing and procfs, and we should
2553 * reset the data back to default state without freeing ->numa_faults.
2555 void task_numa_free(struct task_struct *p, bool final)
2557 /* safe: p either is current or is being freed by current */
2558 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2559 unsigned long *numa_faults = p->numa_faults;
2560 unsigned long flags;
2567 spin_lock_irqsave(&grp->lock, flags);
2568 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2569 grp->faults[i] -= p->numa_faults[i];
2570 grp->total_faults -= p->total_numa_faults;
2573 spin_unlock_irqrestore(&grp->lock, flags);
2574 RCU_INIT_POINTER(p->numa_group, NULL);
2575 put_numa_group(grp);
2579 p->numa_faults = NULL;
2582 p->total_numa_faults = 0;
2583 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2589 * Got a PROT_NONE fault for a page on @node.
2591 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2593 struct task_struct *p = current;
2594 bool migrated = flags & TNF_MIGRATED;
2595 int cpu_node = task_node(current);
2596 int local = !!(flags & TNF_FAULT_LOCAL);
2597 struct numa_group *ng;
2600 if (!static_branch_likely(&sched_numa_balancing))
2603 /* for example, ksmd faulting in a user's mm */
2607 /* Allocate buffer to track faults on a per-node basis */
2608 if (unlikely(!p->numa_faults)) {
2609 int size = sizeof(*p->numa_faults) *
2610 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2612 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2613 if (!p->numa_faults)
2616 p->total_numa_faults = 0;
2617 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2621 * First accesses are treated as private, otherwise consider accesses
2622 * to be private if the accessing pid has not changed
2624 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2627 priv = cpupid_match_pid(p, last_cpupid);
2628 if (!priv && !(flags & TNF_NO_GROUP))
2629 task_numa_group(p, last_cpupid, flags, &priv);
2633 * If a workload spans multiple NUMA nodes, a shared fault that
2634 * occurs wholly within the set of nodes that the workload is
2635 * actively using should be counted as local. This allows the
2636 * scan rate to slow down when a workload has settled down.
2638 ng = deref_curr_numa_group(p);
2639 if (!priv && !local && ng && ng->active_nodes > 1 &&
2640 numa_is_active_node(cpu_node, ng) &&
2641 numa_is_active_node(mem_node, ng))
2645 * Retry to migrate task to preferred node periodically, in case it
2646 * previously failed, or the scheduler moved us.
2648 if (time_after(jiffies, p->numa_migrate_retry)) {
2649 task_numa_placement(p);
2650 numa_migrate_preferred(p);
2654 p->numa_pages_migrated += pages;
2655 if (flags & TNF_MIGRATE_FAIL)
2656 p->numa_faults_locality[2] += pages;
2658 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2659 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2660 p->numa_faults_locality[local] += pages;
2663 static void reset_ptenuma_scan(struct task_struct *p)
2666 * We only did a read acquisition of the mmap sem, so
2667 * p->mm->numa_scan_seq is written to without exclusive access
2668 * and the update is not guaranteed to be atomic. That's not
2669 * much of an issue though, since this is just used for
2670 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2671 * expensive, to avoid any form of compiler optimizations:
2673 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2674 p->mm->numa_scan_offset = 0;
2678 * The expensive part of numa migration is done from task_work context.
2679 * Triggered from task_tick_numa().
2681 static void task_numa_work(struct callback_head *work)
2683 unsigned long migrate, next_scan, now = jiffies;
2684 struct task_struct *p = current;
2685 struct mm_struct *mm = p->mm;
2686 u64 runtime = p->se.sum_exec_runtime;
2687 struct vm_area_struct *vma;
2688 unsigned long start, end;
2689 unsigned long nr_pte_updates = 0;
2690 long pages, virtpages;
2692 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2696 * Who cares about NUMA placement when they're dying.
2698 * NOTE: make sure not to dereference p->mm before this check,
2699 * exit_task_work() happens _after_ exit_mm() so we could be called
2700 * without p->mm even though we still had it when we enqueued this
2703 if (p->flags & PF_EXITING)
2706 if (!mm->numa_next_scan) {
2707 mm->numa_next_scan = now +
2708 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2712 * Enforce maximal scan/migration frequency..
2714 migrate = mm->numa_next_scan;
2715 if (time_before(now, migrate))
2718 if (p->numa_scan_period == 0) {
2719 p->numa_scan_period_max = task_scan_max(p);
2720 p->numa_scan_period = task_scan_start(p);
2723 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2724 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2728 * Delay this task enough that another task of this mm will likely win
2729 * the next time around.
2731 p->node_stamp += 2 * TICK_NSEC;
2733 start = mm->numa_scan_offset;
2734 pages = sysctl_numa_balancing_scan_size;
2735 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2736 virtpages = pages * 8; /* Scan up to this much virtual space */
2741 if (!mmap_read_trylock(mm))
2743 vma = find_vma(mm, start);
2745 reset_ptenuma_scan(p);
2749 for (; vma; vma = vma->vm_next) {
2750 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2751 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2756 * Shared library pages mapped by multiple processes are not
2757 * migrated as it is expected they are cache replicated. Avoid
2758 * hinting faults in read-only file-backed mappings or the vdso
2759 * as migrating the pages will be of marginal benefit.
2762 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2766 * Skip inaccessible VMAs to avoid any confusion between
2767 * PROT_NONE and NUMA hinting ptes
2769 if (!vma_is_accessible(vma))
2773 start = max(start, vma->vm_start);
2774 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2775 end = min(end, vma->vm_end);
2776 nr_pte_updates = change_prot_numa(vma, start, end);
2779 * Try to scan sysctl_numa_balancing_size worth of
2780 * hpages that have at least one present PTE that
2781 * is not already pte-numa. If the VMA contains
2782 * areas that are unused or already full of prot_numa
2783 * PTEs, scan up to virtpages, to skip through those
2787 pages -= (end - start) >> PAGE_SHIFT;
2788 virtpages -= (end - start) >> PAGE_SHIFT;
2791 if (pages <= 0 || virtpages <= 0)
2795 } while (end != vma->vm_end);
2800 * It is possible to reach the end of the VMA list but the last few
2801 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2802 * would find the !migratable VMA on the next scan but not reset the
2803 * scanner to the start so check it now.
2806 mm->numa_scan_offset = start;
2808 reset_ptenuma_scan(p);
2809 mmap_read_unlock(mm);
2812 * Make sure tasks use at least 32x as much time to run other code
2813 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2814 * Usually update_task_scan_period slows down scanning enough; on an
2815 * overloaded system we need to limit overhead on a per task basis.
2817 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2818 u64 diff = p->se.sum_exec_runtime - runtime;
2819 p->node_stamp += 32 * diff;
2823 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2826 struct mm_struct *mm = p->mm;
2829 mm_users = atomic_read(&mm->mm_users);
2830 if (mm_users == 1) {
2831 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2832 mm->numa_scan_seq = 0;
2836 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2837 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2838 /* Protect against double add, see task_tick_numa and task_numa_work */
2839 p->numa_work.next = &p->numa_work;
2840 p->numa_faults = NULL;
2841 RCU_INIT_POINTER(p->numa_group, NULL);
2842 p->last_task_numa_placement = 0;
2843 p->last_sum_exec_runtime = 0;
2845 init_task_work(&p->numa_work, task_numa_work);
2847 /* New address space, reset the preferred nid */
2848 if (!(clone_flags & CLONE_VM)) {
2849 p->numa_preferred_nid = NUMA_NO_NODE;
2854 * New thread, keep existing numa_preferred_nid which should be copied
2855 * already by arch_dup_task_struct but stagger when scans start.
2860 delay = min_t(unsigned int, task_scan_max(current),
2861 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2862 delay += 2 * TICK_NSEC;
2863 p->node_stamp = delay;
2868 * Drive the periodic memory faults..
2870 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2872 struct callback_head *work = &curr->numa_work;
2876 * We don't care about NUMA placement if we don't have memory.
2878 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2882 * Using runtime rather than walltime has the dual advantage that
2883 * we (mostly) drive the selection from busy threads and that the
2884 * task needs to have done some actual work before we bother with
2887 now = curr->se.sum_exec_runtime;
2888 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2890 if (now > curr->node_stamp + period) {
2891 if (!curr->node_stamp)
2892 curr->numa_scan_period = task_scan_start(curr);
2893 curr->node_stamp += period;
2895 if (!time_before(jiffies, curr->mm->numa_next_scan))
2896 task_work_add(curr, work, TWA_RESUME);
2900 static void update_scan_period(struct task_struct *p, int new_cpu)
2902 int src_nid = cpu_to_node(task_cpu(p));
2903 int dst_nid = cpu_to_node(new_cpu);
2905 if (!static_branch_likely(&sched_numa_balancing))
2908 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2911 if (src_nid == dst_nid)
2915 * Allow resets if faults have been trapped before one scan
2916 * has completed. This is most likely due to a new task that
2917 * is pulled cross-node due to wakeups or load balancing.
2919 if (p->numa_scan_seq) {
2921 * Avoid scan adjustments if moving to the preferred
2922 * node or if the task was not previously running on
2923 * the preferred node.
2925 if (dst_nid == p->numa_preferred_nid ||
2926 (p->numa_preferred_nid != NUMA_NO_NODE &&
2927 src_nid != p->numa_preferred_nid))
2931 p->numa_scan_period = task_scan_start(p);
2935 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2939 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2943 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2947 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2951 #endif /* CONFIG_NUMA_BALANCING */
2954 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2956 update_load_add(&cfs_rq->load, se->load.weight);
2958 if (entity_is_task(se)) {
2959 struct rq *rq = rq_of(cfs_rq);
2961 account_numa_enqueue(rq, task_of(se));
2962 list_add(&se->group_node, &rq->cfs_tasks);
2965 cfs_rq->nr_running++;
2969 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2971 update_load_sub(&cfs_rq->load, se->load.weight);
2973 if (entity_is_task(se)) {
2974 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2975 list_del_init(&se->group_node);
2978 cfs_rq->nr_running--;
2982 * Signed add and clamp on underflow.
2984 * Explicitly do a load-store to ensure the intermediate value never hits
2985 * memory. This allows lockless observations without ever seeing the negative
2988 #define add_positive(_ptr, _val) do { \
2989 typeof(_ptr) ptr = (_ptr); \
2990 typeof(_val) val = (_val); \
2991 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2995 if (val < 0 && res > var) \
2998 WRITE_ONCE(*ptr, res); \
3002 * Unsigned subtract and clamp on underflow.
3004 * Explicitly do a load-store to ensure the intermediate value never hits
3005 * memory. This allows lockless observations without ever seeing the negative
3008 #define sub_positive(_ptr, _val) do { \
3009 typeof(_ptr) ptr = (_ptr); \
3010 typeof(*ptr) val = (_val); \
3011 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3015 WRITE_ONCE(*ptr, res); \
3019 * Remove and clamp on negative, from a local variable.
3021 * A variant of sub_positive(), which does not use explicit load-store
3022 * and is thus optimized for local variable updates.
3024 #define lsub_positive(_ptr, _val) do { \
3025 typeof(_ptr) ptr = (_ptr); \
3026 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3031 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3033 cfs_rq->avg.load_avg += se->avg.load_avg;
3034 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3038 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3040 u32 divider = get_pelt_divider(&se->avg);
3041 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3042 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider;
3046 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3048 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3051 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3052 unsigned long weight)
3055 /* commit outstanding execution time */
3056 if (cfs_rq->curr == se)
3057 update_curr(cfs_rq);
3058 update_load_sub(&cfs_rq->load, se->load.weight);
3060 dequeue_load_avg(cfs_rq, se);
3062 update_load_set(&se->load, weight);
3066 u32 divider = get_pelt_divider(&se->avg);
3068 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3072 enqueue_load_avg(cfs_rq, se);
3074 update_load_add(&cfs_rq->load, se->load.weight);
3078 void reweight_task(struct task_struct *p, int prio)
3080 struct sched_entity *se = &p->se;
3081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3082 struct load_weight *load = &se->load;
3083 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3085 reweight_entity(cfs_rq, se, weight);
3086 load->inv_weight = sched_prio_to_wmult[prio];
3089 #ifdef CONFIG_FAIR_GROUP_SCHED
3092 * All this does is approximate the hierarchical proportion which includes that
3093 * global sum we all love to hate.
3095 * That is, the weight of a group entity, is the proportional share of the
3096 * group weight based on the group runqueue weights. That is:
3098 * tg->weight * grq->load.weight
3099 * ge->load.weight = ----------------------------- (1)
3100 * \Sum grq->load.weight
3102 * Now, because computing that sum is prohibitively expensive to compute (been
3103 * there, done that) we approximate it with this average stuff. The average
3104 * moves slower and therefore the approximation is cheaper and more stable.
3106 * So instead of the above, we substitute:
3108 * grq->load.weight -> grq->avg.load_avg (2)
3110 * which yields the following:
3112 * tg->weight * grq->avg.load_avg
3113 * ge->load.weight = ------------------------------ (3)
3116 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3118 * That is shares_avg, and it is right (given the approximation (2)).
3120 * The problem with it is that because the average is slow -- it was designed
3121 * to be exactly that of course -- this leads to transients in boundary
3122 * conditions. In specific, the case where the group was idle and we start the
3123 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3124 * yielding bad latency etc..
3126 * Now, in that special case (1) reduces to:
3128 * tg->weight * grq->load.weight
3129 * ge->load.weight = ----------------------------- = tg->weight (4)
3132 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3134 * So what we do is modify our approximation (3) to approach (4) in the (near)
3139 * tg->weight * grq->load.weight
3140 * --------------------------------------------------- (5)
3141 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3143 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3144 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3147 * tg->weight * grq->load.weight
3148 * ge->load.weight = ----------------------------- (6)
3153 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3154 * max(grq->load.weight, grq->avg.load_avg)
3156 * And that is shares_weight and is icky. In the (near) UP case it approaches
3157 * (4) while in the normal case it approaches (3). It consistently
3158 * overestimates the ge->load.weight and therefore:
3160 * \Sum ge->load.weight >= tg->weight
3164 static long calc_group_shares(struct cfs_rq *cfs_rq)
3166 long tg_weight, tg_shares, load, shares;
3167 struct task_group *tg = cfs_rq->tg;
3169 tg_shares = READ_ONCE(tg->shares);
3171 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3173 tg_weight = atomic_long_read(&tg->load_avg);
3175 /* Ensure tg_weight >= load */
3176 tg_weight -= cfs_rq->tg_load_avg_contrib;
3179 shares = (tg_shares * load);
3181 shares /= tg_weight;
3184 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3185 * of a group with small tg->shares value. It is a floor value which is
3186 * assigned as a minimum load.weight to the sched_entity representing
3187 * the group on a CPU.
3189 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3190 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3191 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3192 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3195 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3197 #endif /* CONFIG_SMP */
3199 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3202 * Recomputes the group entity based on the current state of its group
3205 static void update_cfs_group(struct sched_entity *se)
3207 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3213 if (throttled_hierarchy(gcfs_rq))
3217 shares = READ_ONCE(gcfs_rq->tg->shares);
3219 if (likely(se->load.weight == shares))
3222 shares = calc_group_shares(gcfs_rq);
3225 reweight_entity(cfs_rq_of(se), se, shares);
3228 #else /* CONFIG_FAIR_GROUP_SCHED */
3229 static inline void update_cfs_group(struct sched_entity *se)
3232 #endif /* CONFIG_FAIR_GROUP_SCHED */
3234 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3236 struct rq *rq = rq_of(cfs_rq);
3238 if (&rq->cfs == cfs_rq) {
3240 * There are a few boundary cases this might miss but it should
3241 * get called often enough that that should (hopefully) not be
3244 * It will not get called when we go idle, because the idle
3245 * thread is a different class (!fair), nor will the utilization
3246 * number include things like RT tasks.
3248 * As is, the util number is not freq-invariant (we'd have to
3249 * implement arch_scale_freq_capacity() for that).
3253 cpufreq_update_util(rq, flags);
3258 #ifdef CONFIG_FAIR_GROUP_SCHED
3260 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3262 if (cfs_rq->load.weight)
3265 if (cfs_rq->avg.load_sum)
3268 if (cfs_rq->avg.util_sum)
3271 if (cfs_rq->avg.runnable_sum)
3275 * _avg must be null when _sum are null because _avg = _sum / divider
3276 * Make sure that rounding and/or propagation of PELT values never
3279 SCHED_WARN_ON(cfs_rq->avg.load_avg ||
3280 cfs_rq->avg.util_avg ||
3281 cfs_rq->avg.runnable_avg);
3287 * update_tg_load_avg - update the tg's load avg
3288 * @cfs_rq: the cfs_rq whose avg changed
3290 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3291 * However, because tg->load_avg is a global value there are performance
3294 * In order to avoid having to look at the other cfs_rq's, we use a
3295 * differential update where we store the last value we propagated. This in
3296 * turn allows skipping updates if the differential is 'small'.
3298 * Updating tg's load_avg is necessary before update_cfs_share().
3300 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3302 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3305 * No need to update load_avg for root_task_group as it is not used.
3307 if (cfs_rq->tg == &root_task_group)
3310 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3311 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3312 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3317 * Called within set_task_rq() right before setting a task's CPU. The
3318 * caller only guarantees p->pi_lock is held; no other assumptions,
3319 * including the state of rq->lock, should be made.
3321 void set_task_rq_fair(struct sched_entity *se,
3322 struct cfs_rq *prev, struct cfs_rq *next)
3324 u64 p_last_update_time;
3325 u64 n_last_update_time;
3327 if (!sched_feat(ATTACH_AGE_LOAD))
3331 * We are supposed to update the task to "current" time, then its up to
3332 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3333 * getting what current time is, so simply throw away the out-of-date
3334 * time. This will result in the wakee task is less decayed, but giving
3335 * the wakee more load sounds not bad.
3337 if (!(se->avg.last_update_time && prev))
3340 #ifndef CONFIG_64BIT
3342 u64 p_last_update_time_copy;
3343 u64 n_last_update_time_copy;
3346 p_last_update_time_copy = prev->load_last_update_time_copy;
3347 n_last_update_time_copy = next->load_last_update_time_copy;
3351 p_last_update_time = prev->avg.last_update_time;
3352 n_last_update_time = next->avg.last_update_time;
3354 } while (p_last_update_time != p_last_update_time_copy ||
3355 n_last_update_time != n_last_update_time_copy);
3358 p_last_update_time = prev->avg.last_update_time;
3359 n_last_update_time = next->avg.last_update_time;
3361 __update_load_avg_blocked_se(p_last_update_time, se);
3362 se->avg.last_update_time = n_last_update_time;
3367 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3368 * propagate its contribution. The key to this propagation is the invariant
3369 * that for each group:
3371 * ge->avg == grq->avg (1)
3373 * _IFF_ we look at the pure running and runnable sums. Because they
3374 * represent the very same entity, just at different points in the hierarchy.
3376 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3377 * and simply copies the running/runnable sum over (but still wrong, because
3378 * the group entity and group rq do not have their PELT windows aligned).
3380 * However, update_tg_cfs_load() is more complex. So we have:
3382 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3384 * And since, like util, the runnable part should be directly transferable,
3385 * the following would _appear_ to be the straight forward approach:
3387 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3389 * And per (1) we have:
3391 * ge->avg.runnable_avg == grq->avg.runnable_avg
3395 * ge->load.weight * grq->avg.load_avg
3396 * ge->avg.load_avg = ----------------------------------- (4)
3399 * Except that is wrong!
3401 * Because while for entities historical weight is not important and we
3402 * really only care about our future and therefore can consider a pure
3403 * runnable sum, runqueues can NOT do this.
3405 * We specifically want runqueues to have a load_avg that includes
3406 * historical weights. Those represent the blocked load, the load we expect
3407 * to (shortly) return to us. This only works by keeping the weights as
3408 * integral part of the sum. We therefore cannot decompose as per (3).
3410 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3411 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3412 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3413 * runnable section of these tasks overlap (or not). If they were to perfectly
3414 * align the rq as a whole would be runnable 2/3 of the time. If however we
3415 * always have at least 1 runnable task, the rq as a whole is always runnable.
3417 * So we'll have to approximate.. :/
3419 * Given the constraint:
3421 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3423 * We can construct a rule that adds runnable to a rq by assuming minimal
3426 * On removal, we'll assume each task is equally runnable; which yields:
3428 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3430 * XXX: only do this for the part of runnable > running ?
3435 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3437 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3440 /* Nothing to update */
3445 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3446 * See ___update_load_avg() for details.
3448 divider = get_pelt_divider(&cfs_rq->avg);
3450 /* Set new sched_entity's utilization */
3451 se->avg.util_avg = gcfs_rq->avg.util_avg;
3452 se->avg.util_sum = se->avg.util_avg * divider;
3454 /* Update parent cfs_rq utilization */
3455 add_positive(&cfs_rq->avg.util_avg, delta);
3456 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3460 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3462 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3465 /* Nothing to update */
3470 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3471 * See ___update_load_avg() for details.
3473 divider = get_pelt_divider(&cfs_rq->avg);
3475 /* Set new sched_entity's runnable */
3476 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3477 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3479 /* Update parent cfs_rq runnable */
3480 add_positive(&cfs_rq->avg.runnable_avg, delta);
3481 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3485 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3487 long delta, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3488 unsigned long load_avg;
3495 gcfs_rq->prop_runnable_sum = 0;
3498 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3499 * See ___update_load_avg() for details.
3501 divider = get_pelt_divider(&cfs_rq->avg);
3503 if (runnable_sum >= 0) {
3505 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3506 * the CPU is saturated running == runnable.
3508 runnable_sum += se->avg.load_sum;
3509 runnable_sum = min_t(long, runnable_sum, divider);
3512 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3513 * assuming all tasks are equally runnable.
3515 if (scale_load_down(gcfs_rq->load.weight)) {
3516 load_sum = div_s64(gcfs_rq->avg.load_sum,
3517 scale_load_down(gcfs_rq->load.weight));
3520 /* But make sure to not inflate se's runnable */
3521 runnable_sum = min(se->avg.load_sum, load_sum);
3525 * runnable_sum can't be lower than running_sum
3526 * Rescale running sum to be in the same range as runnable sum
3527 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3528 * runnable_sum is in [0 : LOAD_AVG_MAX]
3530 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3531 runnable_sum = max(runnable_sum, running_sum);
3533 load_sum = (s64)se_weight(se) * runnable_sum;
3534 load_avg = div_s64(load_sum, divider);
3536 se->avg.load_sum = runnable_sum;
3538 delta = load_avg - se->avg.load_avg;
3542 se->avg.load_avg = load_avg;
3544 add_positive(&cfs_rq->avg.load_avg, delta);
3545 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider;
3548 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3550 cfs_rq->propagate = 1;
3551 cfs_rq->prop_runnable_sum += runnable_sum;
3554 /* Update task and its cfs_rq load average */
3555 static inline int propagate_entity_load_avg(struct sched_entity *se)
3557 struct cfs_rq *cfs_rq, *gcfs_rq;
3559 if (entity_is_task(se))
3562 gcfs_rq = group_cfs_rq(se);
3563 if (!gcfs_rq->propagate)
3566 gcfs_rq->propagate = 0;
3568 cfs_rq = cfs_rq_of(se);
3570 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3572 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3573 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3574 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3576 trace_pelt_cfs_tp(cfs_rq);
3577 trace_pelt_se_tp(se);
3583 * Check if we need to update the load and the utilization of a blocked
3586 static inline bool skip_blocked_update(struct sched_entity *se)
3588 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3591 * If sched_entity still have not zero load or utilization, we have to
3594 if (se->avg.load_avg || se->avg.util_avg)
3598 * If there is a pending propagation, we have to update the load and
3599 * the utilization of the sched_entity:
3601 if (gcfs_rq->propagate)
3605 * Otherwise, the load and the utilization of the sched_entity is
3606 * already zero and there is no pending propagation, so it will be a
3607 * waste of time to try to decay it:
3612 #else /* CONFIG_FAIR_GROUP_SCHED */
3614 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3616 static inline int propagate_entity_load_avg(struct sched_entity *se)
3621 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3623 #endif /* CONFIG_FAIR_GROUP_SCHED */
3626 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3627 * @now: current time, as per cfs_rq_clock_pelt()
3628 * @cfs_rq: cfs_rq to update
3630 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3631 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3632 * post_init_entity_util_avg().
3634 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3636 * Returns true if the load decayed or we removed load.
3638 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3639 * call update_tg_load_avg() when this function returns true.
3642 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3644 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3645 struct sched_avg *sa = &cfs_rq->avg;
3648 if (cfs_rq->removed.nr) {
3650 u32 divider = get_pelt_divider(&cfs_rq->avg);
3652 raw_spin_lock(&cfs_rq->removed.lock);
3653 swap(cfs_rq->removed.util_avg, removed_util);
3654 swap(cfs_rq->removed.load_avg, removed_load);
3655 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3656 cfs_rq->removed.nr = 0;
3657 raw_spin_unlock(&cfs_rq->removed.lock);
3660 sub_positive(&sa->load_avg, r);
3661 sa->load_sum = sa->load_avg * divider;
3664 sub_positive(&sa->util_avg, r);
3665 sa->util_sum = sa->util_avg * divider;
3667 r = removed_runnable;
3668 sub_positive(&sa->runnable_avg, r);
3669 sa->runnable_sum = sa->runnable_avg * divider;
3672 * removed_runnable is the unweighted version of removed_load so we
3673 * can use it to estimate removed_load_sum.
3675 add_tg_cfs_propagate(cfs_rq,
3676 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3681 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3683 #ifndef CONFIG_64BIT
3685 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3692 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3693 * @cfs_rq: cfs_rq to attach to
3694 * @se: sched_entity to attach
3696 * Must call update_cfs_rq_load_avg() before this, since we rely on
3697 * cfs_rq->avg.last_update_time being current.
3699 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3702 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3703 * See ___update_load_avg() for details.
3705 u32 divider = get_pelt_divider(&cfs_rq->avg);
3708 * When we attach the @se to the @cfs_rq, we must align the decay
3709 * window because without that, really weird and wonderful things can
3714 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3715 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3718 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3719 * period_contrib. This isn't strictly correct, but since we're
3720 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3723 se->avg.util_sum = se->avg.util_avg * divider;
3725 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3727 se->avg.load_sum = divider;
3728 if (se_weight(se)) {
3730 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3733 enqueue_load_avg(cfs_rq, se);
3734 cfs_rq->avg.util_avg += se->avg.util_avg;
3735 cfs_rq->avg.util_sum += se->avg.util_sum;
3736 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3737 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3739 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3741 cfs_rq_util_change(cfs_rq, 0);
3743 trace_pelt_cfs_tp(cfs_rq);
3747 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3748 * @cfs_rq: cfs_rq to detach from
3749 * @se: sched_entity to detach
3751 * Must call update_cfs_rq_load_avg() before this, since we rely on
3752 * cfs_rq->avg.last_update_time being current.
3754 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3757 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3758 * See ___update_load_avg() for details.
3760 u32 divider = get_pelt_divider(&cfs_rq->avg);
3762 dequeue_load_avg(cfs_rq, se);
3763 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3764 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3765 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3766 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3768 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3770 cfs_rq_util_change(cfs_rq, 0);
3772 trace_pelt_cfs_tp(cfs_rq);
3776 * Optional action to be done while updating the load average
3778 #define UPDATE_TG 0x1
3779 #define SKIP_AGE_LOAD 0x2
3780 #define DO_ATTACH 0x4
3782 /* Update task and its cfs_rq load average */
3783 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3785 u64 now = cfs_rq_clock_pelt(cfs_rq);
3789 * Track task load average for carrying it to new CPU after migrated, and
3790 * track group sched_entity load average for task_h_load calc in migration
3792 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3793 __update_load_avg_se(now, cfs_rq, se);
3795 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3796 decayed |= propagate_entity_load_avg(se);
3798 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3801 * DO_ATTACH means we're here from enqueue_entity().
3802 * !last_update_time means we've passed through
3803 * migrate_task_rq_fair() indicating we migrated.
3805 * IOW we're enqueueing a task on a new CPU.
3807 attach_entity_load_avg(cfs_rq, se);
3808 update_tg_load_avg(cfs_rq);
3810 } else if (decayed) {
3811 cfs_rq_util_change(cfs_rq, 0);
3813 if (flags & UPDATE_TG)
3814 update_tg_load_avg(cfs_rq);
3818 #ifndef CONFIG_64BIT
3819 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3821 u64 last_update_time_copy;
3822 u64 last_update_time;
3825 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3827 last_update_time = cfs_rq->avg.last_update_time;
3828 } while (last_update_time != last_update_time_copy);
3830 return last_update_time;
3833 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3835 return cfs_rq->avg.last_update_time;
3840 * Synchronize entity load avg of dequeued entity without locking
3843 static void sync_entity_load_avg(struct sched_entity *se)
3845 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3846 u64 last_update_time;
3848 last_update_time = cfs_rq_last_update_time(cfs_rq);
3849 __update_load_avg_blocked_se(last_update_time, se);
3853 * Task first catches up with cfs_rq, and then subtract
3854 * itself from the cfs_rq (task must be off the queue now).
3856 static void remove_entity_load_avg(struct sched_entity *se)
3858 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3859 unsigned long flags;
3862 * tasks cannot exit without having gone through wake_up_new_task() ->
3863 * post_init_entity_util_avg() which will have added things to the
3864 * cfs_rq, so we can remove unconditionally.
3867 sync_entity_load_avg(se);
3869 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3870 ++cfs_rq->removed.nr;
3871 cfs_rq->removed.util_avg += se->avg.util_avg;
3872 cfs_rq->removed.load_avg += se->avg.load_avg;
3873 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3874 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3877 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3879 return cfs_rq->avg.runnable_avg;
3882 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3884 return cfs_rq->avg.load_avg;
3887 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
3889 static inline unsigned long task_util(struct task_struct *p)
3891 return READ_ONCE(p->se.avg.util_avg);
3894 static inline unsigned long _task_util_est(struct task_struct *p)
3896 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3898 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
3901 static inline unsigned long task_util_est(struct task_struct *p)
3903 return max(task_util(p), _task_util_est(p));
3906 #ifdef CONFIG_UCLAMP_TASK
3907 static inline unsigned long uclamp_task_util(struct task_struct *p)
3909 return clamp(task_util_est(p),
3910 uclamp_eff_value(p, UCLAMP_MIN),
3911 uclamp_eff_value(p, UCLAMP_MAX));
3914 static inline unsigned long uclamp_task_util(struct task_struct *p)
3916 return task_util_est(p);
3920 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3921 struct task_struct *p)
3923 unsigned int enqueued;
3925 if (!sched_feat(UTIL_EST))
3928 /* Update root cfs_rq's estimated utilization */
3929 enqueued = cfs_rq->avg.util_est.enqueued;
3930 enqueued += _task_util_est(p);
3931 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3933 trace_sched_util_est_cfs_tp(cfs_rq);
3936 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
3937 struct task_struct *p)
3939 unsigned int enqueued;
3941 if (!sched_feat(UTIL_EST))
3944 /* Update root cfs_rq's estimated utilization */
3945 enqueued = cfs_rq->avg.util_est.enqueued;
3946 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
3947 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3949 trace_sched_util_est_cfs_tp(cfs_rq);
3952 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3955 * Check if a (signed) value is within a specified (unsigned) margin,
3956 * based on the observation that:
3958 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3960 * NOTE: this only works when value + margin < INT_MAX.
3962 static inline bool within_margin(int value, int margin)
3964 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3967 static inline void util_est_update(struct cfs_rq *cfs_rq,
3968 struct task_struct *p,
3971 long last_ewma_diff, last_enqueued_diff;
3974 if (!sched_feat(UTIL_EST))
3978 * Skip update of task's estimated utilization when the task has not
3979 * yet completed an activation, e.g. being migrated.
3985 * If the PELT values haven't changed since enqueue time,
3986 * skip the util_est update.
3988 ue = p->se.avg.util_est;
3989 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3992 last_enqueued_diff = ue.enqueued;
3995 * Reset EWMA on utilization increases, the moving average is used only
3996 * to smooth utilization decreases.
3998 ue.enqueued = task_util(p);
3999 if (sched_feat(UTIL_EST_FASTUP)) {
4000 if (ue.ewma < ue.enqueued) {
4001 ue.ewma = ue.enqueued;
4007 * Skip update of task's estimated utilization when its members are
4008 * already ~1% close to its last activation value.
4010 last_ewma_diff = ue.enqueued - ue.ewma;
4011 last_enqueued_diff -= ue.enqueued;
4012 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4013 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4020 * To avoid overestimation of actual task utilization, skip updates if
4021 * we cannot grant there is idle time in this CPU.
4023 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4027 * Update Task's estimated utilization
4029 * When *p completes an activation we can consolidate another sample
4030 * of the task size. This is done by storing the current PELT value
4031 * as ue.enqueued and by using this value to update the Exponential
4032 * Weighted Moving Average (EWMA):
4034 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4035 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4036 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4037 * = w * ( last_ewma_diff ) + ewma(t-1)
4038 * = w * (last_ewma_diff + ewma(t-1) / w)
4040 * Where 'w' is the weight of new samples, which is configured to be
4041 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4043 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4044 ue.ewma += last_ewma_diff;
4045 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4047 ue.enqueued |= UTIL_AVG_UNCHANGED;
4048 WRITE_ONCE(p->se.avg.util_est, ue);
4050 trace_sched_util_est_se_tp(&p->se);
4053 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4055 return fits_capacity(uclamp_task_util(p), capacity);
4058 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4060 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4063 if (!p || p->nr_cpus_allowed == 1) {
4064 rq->misfit_task_load = 0;
4068 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4069 rq->misfit_task_load = 0;
4074 * Make sure that misfit_task_load will not be null even if
4075 * task_h_load() returns 0.
4077 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4080 #else /* CONFIG_SMP */
4082 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4087 #define UPDATE_TG 0x0
4088 #define SKIP_AGE_LOAD 0x0
4089 #define DO_ATTACH 0x0
4091 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4093 cfs_rq_util_change(cfs_rq, 0);
4096 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4099 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4101 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4103 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4109 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4112 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4115 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4117 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4119 #endif /* CONFIG_SMP */
4121 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4123 #ifdef CONFIG_SCHED_DEBUG
4124 s64 d = se->vruntime - cfs_rq->min_vruntime;
4129 if (d > 3*sysctl_sched_latency)
4130 schedstat_inc(cfs_rq->nr_spread_over);
4135 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4137 u64 vruntime = cfs_rq->min_vruntime;
4140 * The 'current' period is already promised to the current tasks,
4141 * however the extra weight of the new task will slow them down a
4142 * little, place the new task so that it fits in the slot that
4143 * stays open at the end.
4145 if (initial && sched_feat(START_DEBIT))
4146 vruntime += sched_vslice(cfs_rq, se);
4148 /* sleeps up to a single latency don't count. */
4150 unsigned long thresh = sysctl_sched_latency;
4153 * Halve their sleep time's effect, to allow
4154 * for a gentler effect of sleepers:
4156 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4162 /* ensure we never gain time by being placed backwards. */
4163 se->vruntime = max_vruntime(se->vruntime, vruntime);
4166 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4168 static inline void check_schedstat_required(void)
4170 #ifdef CONFIG_SCHEDSTATS
4171 if (schedstat_enabled())
4174 /* Force schedstat enabled if a dependent tracepoint is active */
4175 if (trace_sched_stat_wait_enabled() ||
4176 trace_sched_stat_sleep_enabled() ||
4177 trace_sched_stat_iowait_enabled() ||
4178 trace_sched_stat_blocked_enabled() ||
4179 trace_sched_stat_runtime_enabled()) {
4180 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4181 "stat_blocked and stat_runtime require the "
4182 "kernel parameter schedstats=enable or "
4183 "kernel.sched_schedstats=1\n");
4188 static inline bool cfs_bandwidth_used(void);
4195 * update_min_vruntime()
4196 * vruntime -= min_vruntime
4200 * update_min_vruntime()
4201 * vruntime += min_vruntime
4203 * this way the vruntime transition between RQs is done when both
4204 * min_vruntime are up-to-date.
4208 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4209 * vruntime -= min_vruntime
4213 * update_min_vruntime()
4214 * vruntime += min_vruntime
4216 * this way we don't have the most up-to-date min_vruntime on the originating
4217 * CPU and an up-to-date min_vruntime on the destination CPU.
4221 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4223 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4224 bool curr = cfs_rq->curr == se;
4227 * If we're the current task, we must renormalise before calling
4231 se->vruntime += cfs_rq->min_vruntime;
4233 update_curr(cfs_rq);
4236 * Otherwise, renormalise after, such that we're placed at the current
4237 * moment in time, instead of some random moment in the past. Being
4238 * placed in the past could significantly boost this task to the
4239 * fairness detriment of existing tasks.
4241 if (renorm && !curr)
4242 se->vruntime += cfs_rq->min_vruntime;
4245 * When enqueuing a sched_entity, we must:
4246 * - Update loads to have both entity and cfs_rq synced with now.
4247 * - Add its load to cfs_rq->runnable_avg
4248 * - For group_entity, update its weight to reflect the new share of
4250 * - Add its new weight to cfs_rq->load.weight
4252 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4253 se_update_runnable(se);
4254 update_cfs_group(se);
4255 account_entity_enqueue(cfs_rq, se);
4257 if (flags & ENQUEUE_WAKEUP)
4258 place_entity(cfs_rq, se, 0);
4260 check_schedstat_required();
4261 update_stats_enqueue(cfs_rq, se, flags);
4262 check_spread(cfs_rq, se);
4264 __enqueue_entity(cfs_rq, se);
4268 * When bandwidth control is enabled, cfs might have been removed
4269 * because of a parent been throttled but cfs->nr_running > 1. Try to
4270 * add it unconditionally.
4272 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4273 list_add_leaf_cfs_rq(cfs_rq);
4275 if (cfs_rq->nr_running == 1)
4276 check_enqueue_throttle(cfs_rq);
4279 static void __clear_buddies_last(struct sched_entity *se)
4281 for_each_sched_entity(se) {
4282 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4283 if (cfs_rq->last != se)
4286 cfs_rq->last = NULL;
4290 static void __clear_buddies_next(struct sched_entity *se)
4292 for_each_sched_entity(se) {
4293 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4294 if (cfs_rq->next != se)
4297 cfs_rq->next = NULL;
4301 static void __clear_buddies_skip(struct sched_entity *se)
4303 for_each_sched_entity(se) {
4304 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4305 if (cfs_rq->skip != se)
4308 cfs_rq->skip = NULL;
4312 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4314 if (cfs_rq->last == se)
4315 __clear_buddies_last(se);
4317 if (cfs_rq->next == se)
4318 __clear_buddies_next(se);
4320 if (cfs_rq->skip == se)
4321 __clear_buddies_skip(se);
4324 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4327 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4330 * Update run-time statistics of the 'current'.
4332 update_curr(cfs_rq);
4335 * When dequeuing a sched_entity, we must:
4336 * - Update loads to have both entity and cfs_rq synced with now.
4337 * - Subtract its load from the cfs_rq->runnable_avg.
4338 * - Subtract its previous weight from cfs_rq->load.weight.
4339 * - For group entity, update its weight to reflect the new share
4340 * of its group cfs_rq.
4342 update_load_avg(cfs_rq, se, UPDATE_TG);
4343 se_update_runnable(se);
4345 update_stats_dequeue(cfs_rq, se, flags);
4347 clear_buddies(cfs_rq, se);
4349 if (se != cfs_rq->curr)
4350 __dequeue_entity(cfs_rq, se);
4352 account_entity_dequeue(cfs_rq, se);
4355 * Normalize after update_curr(); which will also have moved
4356 * min_vruntime if @se is the one holding it back. But before doing
4357 * update_min_vruntime() again, which will discount @se's position and
4358 * can move min_vruntime forward still more.
4360 if (!(flags & DEQUEUE_SLEEP))
4361 se->vruntime -= cfs_rq->min_vruntime;
4363 /* return excess runtime on last dequeue */
4364 return_cfs_rq_runtime(cfs_rq);
4366 update_cfs_group(se);
4369 * Now advance min_vruntime if @se was the entity holding it back,
4370 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4371 * put back on, and if we advance min_vruntime, we'll be placed back
4372 * further than we started -- ie. we'll be penalized.
4374 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4375 update_min_vruntime(cfs_rq);
4379 * Preempt the current task with a newly woken task if needed:
4382 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4384 unsigned long ideal_runtime, delta_exec;
4385 struct sched_entity *se;
4388 ideal_runtime = sched_slice(cfs_rq, curr);
4389 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4390 if (delta_exec > ideal_runtime) {
4391 resched_curr(rq_of(cfs_rq));
4393 * The current task ran long enough, ensure it doesn't get
4394 * re-elected due to buddy favours.
4396 clear_buddies(cfs_rq, curr);
4401 * Ensure that a task that missed wakeup preemption by a
4402 * narrow margin doesn't have to wait for a full slice.
4403 * This also mitigates buddy induced latencies under load.
4405 if (delta_exec < sysctl_sched_min_granularity)
4408 se = __pick_first_entity(cfs_rq);
4409 delta = curr->vruntime - se->vruntime;
4414 if (delta > ideal_runtime)
4415 resched_curr(rq_of(cfs_rq));
4419 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4421 clear_buddies(cfs_rq, se);
4423 /* 'current' is not kept within the tree. */
4426 * Any task has to be enqueued before it get to execute on
4427 * a CPU. So account for the time it spent waiting on the
4430 update_stats_wait_end(cfs_rq, se);
4431 __dequeue_entity(cfs_rq, se);
4432 update_load_avg(cfs_rq, se, UPDATE_TG);
4435 update_stats_curr_start(cfs_rq, se);
4439 * Track our maximum slice length, if the CPU's load is at
4440 * least twice that of our own weight (i.e. dont track it
4441 * when there are only lesser-weight tasks around):
4443 if (schedstat_enabled() &&
4444 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4445 schedstat_set(se->statistics.slice_max,
4446 max((u64)schedstat_val(se->statistics.slice_max),
4447 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4450 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4454 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4457 * Pick the next process, keeping these things in mind, in this order:
4458 * 1) keep things fair between processes/task groups
4459 * 2) pick the "next" process, since someone really wants that to run
4460 * 3) pick the "last" process, for cache locality
4461 * 4) do not run the "skip" process, if something else is available
4463 static struct sched_entity *
4464 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4466 struct sched_entity *left = __pick_first_entity(cfs_rq);
4467 struct sched_entity *se;
4470 * If curr is set we have to see if its left of the leftmost entity
4471 * still in the tree, provided there was anything in the tree at all.
4473 if (!left || (curr && entity_before(curr, left)))
4476 se = left; /* ideally we run the leftmost entity */
4479 * Avoid running the skip buddy, if running something else can
4480 * be done without getting too unfair.
4482 if (cfs_rq->skip && cfs_rq->skip == se) {
4483 struct sched_entity *second;
4486 second = __pick_first_entity(cfs_rq);
4488 second = __pick_next_entity(se);
4489 if (!second || (curr && entity_before(curr, second)))
4493 if (second && wakeup_preempt_entity(second, left) < 1)
4497 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4499 * Someone really wants this to run. If it's not unfair, run it.
4502 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4504 * Prefer last buddy, try to return the CPU to a preempted task.
4512 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4514 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4517 * If still on the runqueue then deactivate_task()
4518 * was not called and update_curr() has to be done:
4521 update_curr(cfs_rq);
4523 /* throttle cfs_rqs exceeding runtime */
4524 check_cfs_rq_runtime(cfs_rq);
4526 check_spread(cfs_rq, prev);
4529 update_stats_wait_start(cfs_rq, prev);
4530 /* Put 'current' back into the tree. */
4531 __enqueue_entity(cfs_rq, prev);
4532 /* in !on_rq case, update occurred at dequeue */
4533 update_load_avg(cfs_rq, prev, 0);
4535 cfs_rq->curr = NULL;
4539 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4542 * Update run-time statistics of the 'current'.
4544 update_curr(cfs_rq);
4547 * Ensure that runnable average is periodically updated.
4549 update_load_avg(cfs_rq, curr, UPDATE_TG);
4550 update_cfs_group(curr);
4552 #ifdef CONFIG_SCHED_HRTICK
4554 * queued ticks are scheduled to match the slice, so don't bother
4555 * validating it and just reschedule.
4558 resched_curr(rq_of(cfs_rq));
4562 * don't let the period tick interfere with the hrtick preemption
4564 if (!sched_feat(DOUBLE_TICK) &&
4565 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4569 if (cfs_rq->nr_running > 1)
4570 check_preempt_tick(cfs_rq, curr);
4574 /**************************************************
4575 * CFS bandwidth control machinery
4578 #ifdef CONFIG_CFS_BANDWIDTH
4580 #ifdef CONFIG_JUMP_LABEL
4581 static struct static_key __cfs_bandwidth_used;
4583 static inline bool cfs_bandwidth_used(void)
4585 return static_key_false(&__cfs_bandwidth_used);
4588 void cfs_bandwidth_usage_inc(void)
4590 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4593 void cfs_bandwidth_usage_dec(void)
4595 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4597 #else /* CONFIG_JUMP_LABEL */
4598 static bool cfs_bandwidth_used(void)
4603 void cfs_bandwidth_usage_inc(void) {}
4604 void cfs_bandwidth_usage_dec(void) {}
4605 #endif /* CONFIG_JUMP_LABEL */
4608 * default period for cfs group bandwidth.
4609 * default: 0.1s, units: nanoseconds
4611 static inline u64 default_cfs_period(void)
4613 return 100000000ULL;
4616 static inline u64 sched_cfs_bandwidth_slice(void)
4618 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4622 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4623 * directly instead of rq->clock to avoid adding additional synchronization
4626 * requires cfs_b->lock
4628 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4630 if (unlikely(cfs_b->quota == RUNTIME_INF))
4633 cfs_b->runtime += cfs_b->quota;
4634 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
4637 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4639 return &tg->cfs_bandwidth;
4642 /* returns 0 on failure to allocate runtime */
4643 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4644 struct cfs_rq *cfs_rq, u64 target_runtime)
4646 u64 min_amount, amount = 0;
4648 lockdep_assert_held(&cfs_b->lock);
4650 /* note: this is a positive sum as runtime_remaining <= 0 */
4651 min_amount = target_runtime - cfs_rq->runtime_remaining;
4653 if (cfs_b->quota == RUNTIME_INF)
4654 amount = min_amount;
4656 start_cfs_bandwidth(cfs_b);
4658 if (cfs_b->runtime > 0) {
4659 amount = min(cfs_b->runtime, min_amount);
4660 cfs_b->runtime -= amount;
4665 cfs_rq->runtime_remaining += amount;
4667 return cfs_rq->runtime_remaining > 0;
4670 /* returns 0 on failure to allocate runtime */
4671 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4673 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4676 raw_spin_lock(&cfs_b->lock);
4677 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4678 raw_spin_unlock(&cfs_b->lock);
4683 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4685 /* dock delta_exec before expiring quota (as it could span periods) */
4686 cfs_rq->runtime_remaining -= delta_exec;
4688 if (likely(cfs_rq->runtime_remaining > 0))
4691 if (cfs_rq->throttled)
4694 * if we're unable to extend our runtime we resched so that the active
4695 * hierarchy can be throttled
4697 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4698 resched_curr(rq_of(cfs_rq));
4701 static __always_inline
4702 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4704 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4707 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4710 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4712 return cfs_bandwidth_used() && cfs_rq->throttled;
4715 /* check whether cfs_rq, or any parent, is throttled */
4716 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4718 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4722 * Ensure that neither of the group entities corresponding to src_cpu or
4723 * dest_cpu are members of a throttled hierarchy when performing group
4724 * load-balance operations.
4726 static inline int throttled_lb_pair(struct task_group *tg,
4727 int src_cpu, int dest_cpu)
4729 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4731 src_cfs_rq = tg->cfs_rq[src_cpu];
4732 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4734 return throttled_hierarchy(src_cfs_rq) ||
4735 throttled_hierarchy(dest_cfs_rq);
4738 static int tg_unthrottle_up(struct task_group *tg, void *data)
4740 struct rq *rq = data;
4741 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4743 cfs_rq->throttle_count--;
4744 if (!cfs_rq->throttle_count) {
4745 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4746 cfs_rq->throttled_clock_task;
4748 /* Add cfs_rq with load or one or more already running entities to the list */
4749 if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running)
4750 list_add_leaf_cfs_rq(cfs_rq);
4756 static int tg_throttle_down(struct task_group *tg, void *data)
4758 struct rq *rq = data;
4759 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4761 /* group is entering throttled state, stop time */
4762 if (!cfs_rq->throttle_count) {
4763 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4764 list_del_leaf_cfs_rq(cfs_rq);
4766 cfs_rq->throttle_count++;
4771 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4773 struct rq *rq = rq_of(cfs_rq);
4774 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4775 struct sched_entity *se;
4776 long task_delta, idle_task_delta, dequeue = 1;
4778 raw_spin_lock(&cfs_b->lock);
4779 /* This will start the period timer if necessary */
4780 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4782 * We have raced with bandwidth becoming available, and if we
4783 * actually throttled the timer might not unthrottle us for an
4784 * entire period. We additionally needed to make sure that any
4785 * subsequent check_cfs_rq_runtime calls agree not to throttle
4786 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4787 * for 1ns of runtime rather than just check cfs_b.
4791 list_add_tail_rcu(&cfs_rq->throttled_list,
4792 &cfs_b->throttled_cfs_rq);
4794 raw_spin_unlock(&cfs_b->lock);
4797 return false; /* Throttle no longer required. */
4799 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4801 /* freeze hierarchy runnable averages while throttled */
4803 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4806 task_delta = cfs_rq->h_nr_running;
4807 idle_task_delta = cfs_rq->idle_h_nr_running;
4808 for_each_sched_entity(se) {
4809 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4810 /* throttled entity or throttle-on-deactivate */
4814 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4816 qcfs_rq->h_nr_running -= task_delta;
4817 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4819 if (qcfs_rq->load.weight) {
4820 /* Avoid re-evaluating load for this entity: */
4821 se = parent_entity(se);
4826 for_each_sched_entity(se) {
4827 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4828 /* throttled entity or throttle-on-deactivate */
4832 update_load_avg(qcfs_rq, se, 0);
4833 se_update_runnable(se);
4835 qcfs_rq->h_nr_running -= task_delta;
4836 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4839 /* At this point se is NULL and we are at root level*/
4840 sub_nr_running(rq, task_delta);
4844 * Note: distribution will already see us throttled via the
4845 * throttled-list. rq->lock protects completion.
4847 cfs_rq->throttled = 1;
4848 cfs_rq->throttled_clock = rq_clock(rq);
4852 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4854 struct rq *rq = rq_of(cfs_rq);
4855 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4856 struct sched_entity *se;
4857 long task_delta, idle_task_delta;
4859 se = cfs_rq->tg->se[cpu_of(rq)];
4861 cfs_rq->throttled = 0;
4863 update_rq_clock(rq);
4865 raw_spin_lock(&cfs_b->lock);
4866 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4867 list_del_rcu(&cfs_rq->throttled_list);
4868 raw_spin_unlock(&cfs_b->lock);
4870 /* update hierarchical throttle state */
4871 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4873 if (!cfs_rq->load.weight)
4876 task_delta = cfs_rq->h_nr_running;
4877 idle_task_delta = cfs_rq->idle_h_nr_running;
4878 for_each_sched_entity(se) {
4881 cfs_rq = cfs_rq_of(se);
4882 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4884 cfs_rq->h_nr_running += task_delta;
4885 cfs_rq->idle_h_nr_running += idle_task_delta;
4887 /* end evaluation on encountering a throttled cfs_rq */
4888 if (cfs_rq_throttled(cfs_rq))
4889 goto unthrottle_throttle;
4892 for_each_sched_entity(se) {
4893 cfs_rq = cfs_rq_of(se);
4895 update_load_avg(cfs_rq, se, UPDATE_TG);
4896 se_update_runnable(se);
4898 cfs_rq->h_nr_running += task_delta;
4899 cfs_rq->idle_h_nr_running += idle_task_delta;
4902 /* end evaluation on encountering a throttled cfs_rq */
4903 if (cfs_rq_throttled(cfs_rq))
4904 goto unthrottle_throttle;
4907 * One parent has been throttled and cfs_rq removed from the
4908 * list. Add it back to not break the leaf list.
4910 if (throttled_hierarchy(cfs_rq))
4911 list_add_leaf_cfs_rq(cfs_rq);
4914 /* At this point se is NULL and we are at root level*/
4915 add_nr_running(rq, task_delta);
4917 unthrottle_throttle:
4919 * The cfs_rq_throttled() breaks in the above iteration can result in
4920 * incomplete leaf list maintenance, resulting in triggering the
4923 for_each_sched_entity(se) {
4924 cfs_rq = cfs_rq_of(se);
4926 if (list_add_leaf_cfs_rq(cfs_rq))
4930 assert_list_leaf_cfs_rq(rq);
4932 /* Determine whether we need to wake up potentially idle CPU: */
4933 if (rq->curr == rq->idle && rq->cfs.nr_running)
4937 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
4939 struct cfs_rq *cfs_rq;
4940 u64 runtime, remaining = 1;
4943 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4945 struct rq *rq = rq_of(cfs_rq);
4948 rq_lock_irqsave(rq, &rf);
4949 if (!cfs_rq_throttled(cfs_rq))
4952 /* By the above check, this should never be true */
4953 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4955 raw_spin_lock(&cfs_b->lock);
4956 runtime = -cfs_rq->runtime_remaining + 1;
4957 if (runtime > cfs_b->runtime)
4958 runtime = cfs_b->runtime;
4959 cfs_b->runtime -= runtime;
4960 remaining = cfs_b->runtime;
4961 raw_spin_unlock(&cfs_b->lock);
4963 cfs_rq->runtime_remaining += runtime;
4965 /* we check whether we're throttled above */
4966 if (cfs_rq->runtime_remaining > 0)
4967 unthrottle_cfs_rq(cfs_rq);
4970 rq_unlock_irqrestore(rq, &rf);
4979 * Responsible for refilling a task_group's bandwidth and unthrottling its
4980 * cfs_rqs as appropriate. If there has been no activity within the last
4981 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4982 * used to track this state.
4984 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4988 /* no need to continue the timer with no bandwidth constraint */
4989 if (cfs_b->quota == RUNTIME_INF)
4990 goto out_deactivate;
4992 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4993 cfs_b->nr_periods += overrun;
4995 /* Refill extra burst quota even if cfs_b->idle */
4996 __refill_cfs_bandwidth_runtime(cfs_b);
4999 * idle depends on !throttled (for the case of a large deficit), and if
5000 * we're going inactive then everything else can be deferred
5002 if (cfs_b->idle && !throttled)
5003 goto out_deactivate;
5006 /* mark as potentially idle for the upcoming period */
5011 /* account preceding periods in which throttling occurred */
5012 cfs_b->nr_throttled += overrun;
5015 * This check is repeated as we release cfs_b->lock while we unthrottle.
5017 while (throttled && cfs_b->runtime > 0) {
5018 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5019 /* we can't nest cfs_b->lock while distributing bandwidth */
5020 distribute_cfs_runtime(cfs_b);
5021 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5023 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5027 * While we are ensured activity in the period following an
5028 * unthrottle, this also covers the case in which the new bandwidth is
5029 * insufficient to cover the existing bandwidth deficit. (Forcing the
5030 * timer to remain active while there are any throttled entities.)
5040 /* a cfs_rq won't donate quota below this amount */
5041 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5042 /* minimum remaining period time to redistribute slack quota */
5043 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5044 /* how long we wait to gather additional slack before distributing */
5045 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5048 * Are we near the end of the current quota period?
5050 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5051 * hrtimer base being cleared by hrtimer_start. In the case of
5052 * migrate_hrtimers, base is never cleared, so we are fine.
5054 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5056 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5059 /* if the call-back is running a quota refresh is already occurring */
5060 if (hrtimer_callback_running(refresh_timer))
5063 /* is a quota refresh about to occur? */
5064 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5065 if (remaining < (s64)min_expire)
5071 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5073 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5075 /* if there's a quota refresh soon don't bother with slack */
5076 if (runtime_refresh_within(cfs_b, min_left))
5079 /* don't push forwards an existing deferred unthrottle */
5080 if (cfs_b->slack_started)
5082 cfs_b->slack_started = true;
5084 hrtimer_start(&cfs_b->slack_timer,
5085 ns_to_ktime(cfs_bandwidth_slack_period),
5089 /* we know any runtime found here is valid as update_curr() precedes return */
5090 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5092 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5093 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5095 if (slack_runtime <= 0)
5098 raw_spin_lock(&cfs_b->lock);
5099 if (cfs_b->quota != RUNTIME_INF) {
5100 cfs_b->runtime += slack_runtime;
5102 /* we are under rq->lock, defer unthrottling using a timer */
5103 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5104 !list_empty(&cfs_b->throttled_cfs_rq))
5105 start_cfs_slack_bandwidth(cfs_b);
5107 raw_spin_unlock(&cfs_b->lock);
5109 /* even if it's not valid for return we don't want to try again */
5110 cfs_rq->runtime_remaining -= slack_runtime;
5113 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5115 if (!cfs_bandwidth_used())
5118 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5121 __return_cfs_rq_runtime(cfs_rq);
5125 * This is done with a timer (instead of inline with bandwidth return) since
5126 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5128 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5130 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5131 unsigned long flags;
5133 /* confirm we're still not at a refresh boundary */
5134 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5135 cfs_b->slack_started = false;
5137 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5138 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5142 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5143 runtime = cfs_b->runtime;
5145 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5150 distribute_cfs_runtime(cfs_b);
5154 * When a group wakes up we want to make sure that its quota is not already
5155 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5156 * runtime as update_curr() throttling can not trigger until it's on-rq.
5158 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5160 if (!cfs_bandwidth_used())
5163 /* an active group must be handled by the update_curr()->put() path */
5164 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5167 /* ensure the group is not already throttled */
5168 if (cfs_rq_throttled(cfs_rq))
5171 /* update runtime allocation */
5172 account_cfs_rq_runtime(cfs_rq, 0);
5173 if (cfs_rq->runtime_remaining <= 0)
5174 throttle_cfs_rq(cfs_rq);
5177 static void sync_throttle(struct task_group *tg, int cpu)
5179 struct cfs_rq *pcfs_rq, *cfs_rq;
5181 if (!cfs_bandwidth_used())
5187 cfs_rq = tg->cfs_rq[cpu];
5188 pcfs_rq = tg->parent->cfs_rq[cpu];
5190 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5191 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5194 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5195 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5197 if (!cfs_bandwidth_used())
5200 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5204 * it's possible for a throttled entity to be forced into a running
5205 * state (e.g. set_curr_task), in this case we're finished.
5207 if (cfs_rq_throttled(cfs_rq))
5210 return throttle_cfs_rq(cfs_rq);
5213 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5215 struct cfs_bandwidth *cfs_b =
5216 container_of(timer, struct cfs_bandwidth, slack_timer);
5218 do_sched_cfs_slack_timer(cfs_b);
5220 return HRTIMER_NORESTART;
5223 extern const u64 max_cfs_quota_period;
5225 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5227 struct cfs_bandwidth *cfs_b =
5228 container_of(timer, struct cfs_bandwidth, period_timer);
5229 unsigned long flags;
5234 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5236 overrun = hrtimer_forward_now(timer, cfs_b->period);
5240 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5243 u64 new, old = ktime_to_ns(cfs_b->period);
5246 * Grow period by a factor of 2 to avoid losing precision.
5247 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5251 if (new < max_cfs_quota_period) {
5252 cfs_b->period = ns_to_ktime(new);
5256 pr_warn_ratelimited(
5257 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5259 div_u64(new, NSEC_PER_USEC),
5260 div_u64(cfs_b->quota, NSEC_PER_USEC));
5262 pr_warn_ratelimited(
5263 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5265 div_u64(old, NSEC_PER_USEC),
5266 div_u64(cfs_b->quota, NSEC_PER_USEC));
5269 /* reset count so we don't come right back in here */
5274 cfs_b->period_active = 0;
5275 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5277 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5280 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5282 raw_spin_lock_init(&cfs_b->lock);
5284 cfs_b->quota = RUNTIME_INF;
5285 cfs_b->period = ns_to_ktime(default_cfs_period());
5288 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5289 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5290 cfs_b->period_timer.function = sched_cfs_period_timer;
5291 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5292 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5293 cfs_b->slack_started = false;
5296 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5298 cfs_rq->runtime_enabled = 0;
5299 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5302 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5304 lockdep_assert_held(&cfs_b->lock);
5306 if (cfs_b->period_active)
5309 cfs_b->period_active = 1;
5310 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5311 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5314 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5316 /* init_cfs_bandwidth() was not called */
5317 if (!cfs_b->throttled_cfs_rq.next)
5320 hrtimer_cancel(&cfs_b->period_timer);
5321 hrtimer_cancel(&cfs_b->slack_timer);
5325 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5327 * The race is harmless, since modifying bandwidth settings of unhooked group
5328 * bits doesn't do much.
5331 /* cpu online callback */
5332 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5334 struct task_group *tg;
5336 lockdep_assert_rq_held(rq);
5339 list_for_each_entry_rcu(tg, &task_groups, list) {
5340 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5341 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5343 raw_spin_lock(&cfs_b->lock);
5344 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5345 raw_spin_unlock(&cfs_b->lock);
5350 /* cpu offline callback */
5351 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5353 struct task_group *tg;
5355 lockdep_assert_rq_held(rq);
5358 list_for_each_entry_rcu(tg, &task_groups, list) {
5359 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5361 if (!cfs_rq->runtime_enabled)
5365 * clock_task is not advancing so we just need to make sure
5366 * there's some valid quota amount
5368 cfs_rq->runtime_remaining = 1;
5370 * Offline rq is schedulable till CPU is completely disabled
5371 * in take_cpu_down(), so we prevent new cfs throttling here.
5373 cfs_rq->runtime_enabled = 0;
5375 if (cfs_rq_throttled(cfs_rq))
5376 unthrottle_cfs_rq(cfs_rq);
5381 #else /* CONFIG_CFS_BANDWIDTH */
5383 static inline bool cfs_bandwidth_used(void)
5388 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5389 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5390 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5391 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5392 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5394 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5399 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5404 static inline int throttled_lb_pair(struct task_group *tg,
5405 int src_cpu, int dest_cpu)
5410 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5412 #ifdef CONFIG_FAIR_GROUP_SCHED
5413 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5416 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5420 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5421 static inline void update_runtime_enabled(struct rq *rq) {}
5422 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5424 #endif /* CONFIG_CFS_BANDWIDTH */
5426 /**************************************************
5427 * CFS operations on tasks:
5430 #ifdef CONFIG_SCHED_HRTICK
5431 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5433 struct sched_entity *se = &p->se;
5434 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5436 SCHED_WARN_ON(task_rq(p) != rq);
5438 if (rq->cfs.h_nr_running > 1) {
5439 u64 slice = sched_slice(cfs_rq, se);
5440 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5441 s64 delta = slice - ran;
5444 if (task_current(rq, p))
5448 hrtick_start(rq, delta);
5453 * called from enqueue/dequeue and updates the hrtick when the
5454 * current task is from our class and nr_running is low enough
5457 static void hrtick_update(struct rq *rq)
5459 struct task_struct *curr = rq->curr;
5461 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5464 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5465 hrtick_start_fair(rq, curr);
5467 #else /* !CONFIG_SCHED_HRTICK */
5469 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5473 static inline void hrtick_update(struct rq *rq)
5479 static inline unsigned long cpu_util(int cpu);
5481 static inline bool cpu_overutilized(int cpu)
5483 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5486 static inline void update_overutilized_status(struct rq *rq)
5488 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5489 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5490 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5494 static inline void update_overutilized_status(struct rq *rq) { }
5497 /* Runqueue only has SCHED_IDLE tasks enqueued */
5498 static int sched_idle_rq(struct rq *rq)
5500 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5505 static int sched_idle_cpu(int cpu)
5507 return sched_idle_rq(cpu_rq(cpu));
5512 * The enqueue_task method is called before nr_running is
5513 * increased. Here we update the fair scheduling stats and
5514 * then put the task into the rbtree:
5517 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5519 struct cfs_rq *cfs_rq;
5520 struct sched_entity *se = &p->se;
5521 int idle_h_nr_running = task_has_idle_policy(p);
5522 int task_new = !(flags & ENQUEUE_WAKEUP);
5525 * The code below (indirectly) updates schedutil which looks at
5526 * the cfs_rq utilization to select a frequency.
5527 * Let's add the task's estimated utilization to the cfs_rq's
5528 * estimated utilization, before we update schedutil.
5530 util_est_enqueue(&rq->cfs, p);
5533 * If in_iowait is set, the code below may not trigger any cpufreq
5534 * utilization updates, so do it here explicitly with the IOWAIT flag
5538 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5540 for_each_sched_entity(se) {
5543 cfs_rq = cfs_rq_of(se);
5544 enqueue_entity(cfs_rq, se, flags);
5546 cfs_rq->h_nr_running++;
5547 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5549 /* end evaluation on encountering a throttled cfs_rq */
5550 if (cfs_rq_throttled(cfs_rq))
5551 goto enqueue_throttle;
5553 flags = ENQUEUE_WAKEUP;
5556 for_each_sched_entity(se) {
5557 cfs_rq = cfs_rq_of(se);
5559 update_load_avg(cfs_rq, se, UPDATE_TG);
5560 se_update_runnable(se);
5561 update_cfs_group(se);
5563 cfs_rq->h_nr_running++;
5564 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5566 /* end evaluation on encountering a throttled cfs_rq */
5567 if (cfs_rq_throttled(cfs_rq))
5568 goto enqueue_throttle;
5571 * One parent has been throttled and cfs_rq removed from the
5572 * list. Add it back to not break the leaf list.
5574 if (throttled_hierarchy(cfs_rq))
5575 list_add_leaf_cfs_rq(cfs_rq);
5578 /* At this point se is NULL and we are at root level*/
5579 add_nr_running(rq, 1);
5582 * Since new tasks are assigned an initial util_avg equal to
5583 * half of the spare capacity of their CPU, tiny tasks have the
5584 * ability to cross the overutilized threshold, which will
5585 * result in the load balancer ruining all the task placement
5586 * done by EAS. As a way to mitigate that effect, do not account
5587 * for the first enqueue operation of new tasks during the
5588 * overutilized flag detection.
5590 * A better way of solving this problem would be to wait for
5591 * the PELT signals of tasks to converge before taking them
5592 * into account, but that is not straightforward to implement,
5593 * and the following generally works well enough in practice.
5596 update_overutilized_status(rq);
5599 if (cfs_bandwidth_used()) {
5601 * When bandwidth control is enabled; the cfs_rq_throttled()
5602 * breaks in the above iteration can result in incomplete
5603 * leaf list maintenance, resulting in triggering the assertion
5606 for_each_sched_entity(se) {
5607 cfs_rq = cfs_rq_of(se);
5609 if (list_add_leaf_cfs_rq(cfs_rq))
5614 assert_list_leaf_cfs_rq(rq);
5619 static void set_next_buddy(struct sched_entity *se);
5622 * The dequeue_task method is called before nr_running is
5623 * decreased. We remove the task from the rbtree and
5624 * update the fair scheduling stats:
5626 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5628 struct cfs_rq *cfs_rq;
5629 struct sched_entity *se = &p->se;
5630 int task_sleep = flags & DEQUEUE_SLEEP;
5631 int idle_h_nr_running = task_has_idle_policy(p);
5632 bool was_sched_idle = sched_idle_rq(rq);
5634 util_est_dequeue(&rq->cfs, p);
5636 for_each_sched_entity(se) {
5637 cfs_rq = cfs_rq_of(se);
5638 dequeue_entity(cfs_rq, se, flags);
5640 cfs_rq->h_nr_running--;
5641 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5643 /* end evaluation on encountering a throttled cfs_rq */
5644 if (cfs_rq_throttled(cfs_rq))
5645 goto dequeue_throttle;
5647 /* Don't dequeue parent if it has other entities besides us */
5648 if (cfs_rq->load.weight) {
5649 /* Avoid re-evaluating load for this entity: */
5650 se = parent_entity(se);
5652 * Bias pick_next to pick a task from this cfs_rq, as
5653 * p is sleeping when it is within its sched_slice.
5655 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5659 flags |= DEQUEUE_SLEEP;
5662 for_each_sched_entity(se) {
5663 cfs_rq = cfs_rq_of(se);
5665 update_load_avg(cfs_rq, se, UPDATE_TG);
5666 se_update_runnable(se);
5667 update_cfs_group(se);
5669 cfs_rq->h_nr_running--;
5670 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5672 /* end evaluation on encountering a throttled cfs_rq */
5673 if (cfs_rq_throttled(cfs_rq))
5674 goto dequeue_throttle;
5678 /* At this point se is NULL and we are at root level*/
5679 sub_nr_running(rq, 1);
5681 /* balance early to pull high priority tasks */
5682 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5683 rq->next_balance = jiffies;
5686 util_est_update(&rq->cfs, p, task_sleep);
5692 /* Working cpumask for: load_balance, load_balance_newidle. */
5693 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5694 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5696 #ifdef CONFIG_NO_HZ_COMMON
5699 cpumask_var_t idle_cpus_mask;
5701 int has_blocked; /* Idle CPUS has blocked load */
5702 unsigned long next_balance; /* in jiffy units */
5703 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5704 } nohz ____cacheline_aligned;
5706 #endif /* CONFIG_NO_HZ_COMMON */
5708 static unsigned long cpu_load(struct rq *rq)
5710 return cfs_rq_load_avg(&rq->cfs);
5714 * cpu_load_without - compute CPU load without any contributions from *p
5715 * @cpu: the CPU which load is requested
5716 * @p: the task which load should be discounted
5718 * The load of a CPU is defined by the load of tasks currently enqueued on that
5719 * CPU as well as tasks which are currently sleeping after an execution on that
5722 * This method returns the load of the specified CPU by discounting the load of
5723 * the specified task, whenever the task is currently contributing to the CPU
5726 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5728 struct cfs_rq *cfs_rq;
5731 /* Task has no contribution or is new */
5732 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5733 return cpu_load(rq);
5736 load = READ_ONCE(cfs_rq->avg.load_avg);
5738 /* Discount task's util from CPU's util */
5739 lsub_positive(&load, task_h_load(p));
5744 static unsigned long cpu_runnable(struct rq *rq)
5746 return cfs_rq_runnable_avg(&rq->cfs);
5749 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5751 struct cfs_rq *cfs_rq;
5752 unsigned int runnable;
5754 /* Task has no contribution or is new */
5755 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5756 return cpu_runnable(rq);
5759 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5761 /* Discount task's runnable from CPU's runnable */
5762 lsub_positive(&runnable, p->se.avg.runnable_avg);
5767 static unsigned long capacity_of(int cpu)
5769 return cpu_rq(cpu)->cpu_capacity;
5772 static void record_wakee(struct task_struct *p)
5775 * Only decay a single time; tasks that have less then 1 wakeup per
5776 * jiffy will not have built up many flips.
5778 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5779 current->wakee_flips >>= 1;
5780 current->wakee_flip_decay_ts = jiffies;
5783 if (current->last_wakee != p) {
5784 current->last_wakee = p;
5785 current->wakee_flips++;
5790 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5792 * A waker of many should wake a different task than the one last awakened
5793 * at a frequency roughly N times higher than one of its wakees.
5795 * In order to determine whether we should let the load spread vs consolidating
5796 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5797 * partner, and a factor of lls_size higher frequency in the other.
5799 * With both conditions met, we can be relatively sure that the relationship is
5800 * non-monogamous, with partner count exceeding socket size.
5802 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5803 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5806 static int wake_wide(struct task_struct *p)
5808 unsigned int master = current->wakee_flips;
5809 unsigned int slave = p->wakee_flips;
5810 int factor = __this_cpu_read(sd_llc_size);
5813 swap(master, slave);
5814 if (slave < factor || master < slave * factor)
5820 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5821 * soonest. For the purpose of speed we only consider the waking and previous
5824 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5825 * cache-affine and is (or will be) idle.
5827 * wake_affine_weight() - considers the weight to reflect the average
5828 * scheduling latency of the CPUs. This seems to work
5829 * for the overloaded case.
5832 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5835 * If this_cpu is idle, it implies the wakeup is from interrupt
5836 * context. Only allow the move if cache is shared. Otherwise an
5837 * interrupt intensive workload could force all tasks onto one
5838 * node depending on the IO topology or IRQ affinity settings.
5840 * If the prev_cpu is idle and cache affine then avoid a migration.
5841 * There is no guarantee that the cache hot data from an interrupt
5842 * is more important than cache hot data on the prev_cpu and from
5843 * a cpufreq perspective, it's better to have higher utilisation
5846 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5847 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5849 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5852 if (available_idle_cpu(prev_cpu))
5855 return nr_cpumask_bits;
5859 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5860 int this_cpu, int prev_cpu, int sync)
5862 s64 this_eff_load, prev_eff_load;
5863 unsigned long task_load;
5865 this_eff_load = cpu_load(cpu_rq(this_cpu));
5868 unsigned long current_load = task_h_load(current);
5870 if (current_load > this_eff_load)
5873 this_eff_load -= current_load;
5876 task_load = task_h_load(p);
5878 this_eff_load += task_load;
5879 if (sched_feat(WA_BIAS))
5880 this_eff_load *= 100;
5881 this_eff_load *= capacity_of(prev_cpu);
5883 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5884 prev_eff_load -= task_load;
5885 if (sched_feat(WA_BIAS))
5886 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5887 prev_eff_load *= capacity_of(this_cpu);
5890 * If sync, adjust the weight of prev_eff_load such that if
5891 * prev_eff == this_eff that select_idle_sibling() will consider
5892 * stacking the wakee on top of the waker if no other CPU is
5898 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5901 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5902 int this_cpu, int prev_cpu, int sync)
5904 int target = nr_cpumask_bits;
5906 if (sched_feat(WA_IDLE))
5907 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5909 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5910 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5912 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5913 if (target == nr_cpumask_bits)
5916 schedstat_inc(sd->ttwu_move_affine);
5917 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5921 static struct sched_group *
5922 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
5925 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5928 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5930 unsigned long load, min_load = ULONG_MAX;
5931 unsigned int min_exit_latency = UINT_MAX;
5932 u64 latest_idle_timestamp = 0;
5933 int least_loaded_cpu = this_cpu;
5934 int shallowest_idle_cpu = -1;
5937 /* Check if we have any choice: */
5938 if (group->group_weight == 1)
5939 return cpumask_first(sched_group_span(group));
5941 /* Traverse only the allowed CPUs */
5942 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5943 struct rq *rq = cpu_rq(i);
5945 if (!sched_core_cookie_match(rq, p))
5948 if (sched_idle_cpu(i))
5951 if (available_idle_cpu(i)) {
5952 struct cpuidle_state *idle = idle_get_state(rq);
5953 if (idle && idle->exit_latency < min_exit_latency) {
5955 * We give priority to a CPU whose idle state
5956 * has the smallest exit latency irrespective
5957 * of any idle timestamp.
5959 min_exit_latency = idle->exit_latency;
5960 latest_idle_timestamp = rq->idle_stamp;
5961 shallowest_idle_cpu = i;
5962 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5963 rq->idle_stamp > latest_idle_timestamp) {
5965 * If equal or no active idle state, then
5966 * the most recently idled CPU might have
5969 latest_idle_timestamp = rq->idle_stamp;
5970 shallowest_idle_cpu = i;
5972 } else if (shallowest_idle_cpu == -1) {
5973 load = cpu_load(cpu_rq(i));
5974 if (load < min_load) {
5976 least_loaded_cpu = i;
5981 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5984 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5985 int cpu, int prev_cpu, int sd_flag)
5989 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5993 * We need task's util for cpu_util_without, sync it up to
5994 * prev_cpu's last_update_time.
5996 if (!(sd_flag & SD_BALANCE_FORK))
5997 sync_entity_load_avg(&p->se);
6000 struct sched_group *group;
6001 struct sched_domain *tmp;
6004 if (!(sd->flags & sd_flag)) {
6009 group = find_idlest_group(sd, p, cpu);
6015 new_cpu = find_idlest_group_cpu(group, p, cpu);
6016 if (new_cpu == cpu) {
6017 /* Now try balancing at a lower domain level of 'cpu': */
6022 /* Now try balancing at a lower domain level of 'new_cpu': */
6024 weight = sd->span_weight;
6026 for_each_domain(cpu, tmp) {
6027 if (weight <= tmp->span_weight)
6029 if (tmp->flags & sd_flag)
6037 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6039 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6040 sched_cpu_cookie_match(cpu_rq(cpu), p))
6046 #ifdef CONFIG_SCHED_SMT
6047 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6048 EXPORT_SYMBOL_GPL(sched_smt_present);
6050 static inline void set_idle_cores(int cpu, int val)
6052 struct sched_domain_shared *sds;
6054 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6056 WRITE_ONCE(sds->has_idle_cores, val);
6059 static inline bool test_idle_cores(int cpu, bool def)
6061 struct sched_domain_shared *sds;
6063 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6065 return READ_ONCE(sds->has_idle_cores);
6071 * Scans the local SMT mask to see if the entire core is idle, and records this
6072 * information in sd_llc_shared->has_idle_cores.
6074 * Since SMT siblings share all cache levels, inspecting this limited remote
6075 * state should be fairly cheap.
6077 void __update_idle_core(struct rq *rq)
6079 int core = cpu_of(rq);
6083 if (test_idle_cores(core, true))
6086 for_each_cpu(cpu, cpu_smt_mask(core)) {
6090 if (!available_idle_cpu(cpu))
6094 set_idle_cores(core, 1);
6100 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6101 * there are no idle cores left in the system; tracked through
6102 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6104 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6109 if (!static_branch_likely(&sched_smt_present))
6110 return __select_idle_cpu(core, p);
6112 for_each_cpu(cpu, cpu_smt_mask(core)) {
6113 if (!available_idle_cpu(cpu)) {
6115 if (*idle_cpu == -1) {
6116 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6124 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6131 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6136 * Scan the local SMT mask for idle CPUs.
6138 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6142 for_each_cpu(cpu, cpu_smt_mask(target)) {
6143 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6144 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6146 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6153 #else /* CONFIG_SCHED_SMT */
6155 static inline void set_idle_cores(int cpu, int val)
6159 static inline bool test_idle_cores(int cpu, bool def)
6164 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6166 return __select_idle_cpu(core, p);
6169 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6174 #endif /* CONFIG_SCHED_SMT */
6177 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6178 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6179 * average idle time for this rq (as found in rq->avg_idle).
6181 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6183 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6184 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6185 struct rq *this_rq = this_rq();
6186 int this = smp_processor_id();
6187 struct sched_domain *this_sd;
6190 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6194 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6196 if (sched_feat(SIS_PROP) && !has_idle_core) {
6197 u64 avg_cost, avg_idle, span_avg;
6198 unsigned long now = jiffies;
6201 * If we're busy, the assumption that the last idle period
6202 * predicts the future is flawed; age away the remaining
6203 * predicted idle time.
6205 if (unlikely(this_rq->wake_stamp < now)) {
6206 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6207 this_rq->wake_stamp++;
6208 this_rq->wake_avg_idle >>= 1;
6212 avg_idle = this_rq->wake_avg_idle;
6213 avg_cost = this_sd->avg_scan_cost + 1;
6215 span_avg = sd->span_weight * avg_idle;
6216 if (span_avg > 4*avg_cost)
6217 nr = div_u64(span_avg, avg_cost);
6221 time = cpu_clock(this);
6224 for_each_cpu_wrap(cpu, cpus, target) {
6225 if (has_idle_core) {
6226 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6227 if ((unsigned int)i < nr_cpumask_bits)
6233 idle_cpu = __select_idle_cpu(cpu, p);
6234 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6240 set_idle_cores(target, false);
6242 if (sched_feat(SIS_PROP) && !has_idle_core) {
6243 time = cpu_clock(this) - time;
6246 * Account for the scan cost of wakeups against the average
6249 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6251 update_avg(&this_sd->avg_scan_cost, time);
6258 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6259 * the task fits. If no CPU is big enough, but there are idle ones, try to
6260 * maximize capacity.
6263 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6265 unsigned long task_util, best_cap = 0;
6266 int cpu, best_cpu = -1;
6267 struct cpumask *cpus;
6269 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6270 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6272 task_util = uclamp_task_util(p);
6274 for_each_cpu_wrap(cpu, cpus, target) {
6275 unsigned long cpu_cap = capacity_of(cpu);
6277 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6279 if (fits_capacity(task_util, cpu_cap))
6282 if (cpu_cap > best_cap) {
6291 static inline bool asym_fits_capacity(int task_util, int cpu)
6293 if (static_branch_unlikely(&sched_asym_cpucapacity))
6294 return fits_capacity(task_util, capacity_of(cpu));
6300 * Try and locate an idle core/thread in the LLC cache domain.
6302 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6304 bool has_idle_core = false;
6305 struct sched_domain *sd;
6306 unsigned long task_util;
6307 int i, recent_used_cpu;
6310 * On asymmetric system, update task utilization because we will check
6311 * that the task fits with cpu's capacity.
6313 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6314 sync_entity_load_avg(&p->se);
6315 task_util = uclamp_task_util(p);
6319 * per-cpu select_idle_mask usage
6321 lockdep_assert_irqs_disabled();
6323 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6324 asym_fits_capacity(task_util, target))
6328 * If the previous CPU is cache affine and idle, don't be stupid:
6330 if (prev != target && cpus_share_cache(prev, target) &&
6331 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6332 asym_fits_capacity(task_util, prev))
6336 * Allow a per-cpu kthread to stack with the wakee if the
6337 * kworker thread and the tasks previous CPUs are the same.
6338 * The assumption is that the wakee queued work for the
6339 * per-cpu kthread that is now complete and the wakeup is
6340 * essentially a sync wakeup. An obvious example of this
6341 * pattern is IO completions.
6343 if (is_per_cpu_kthread(current) &&
6344 prev == smp_processor_id() &&
6345 this_rq()->nr_running <= 1) {
6349 /* Check a recently used CPU as a potential idle candidate: */
6350 recent_used_cpu = p->recent_used_cpu;
6351 if (recent_used_cpu != prev &&
6352 recent_used_cpu != target &&
6353 cpus_share_cache(recent_used_cpu, target) &&
6354 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6355 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6356 asym_fits_capacity(task_util, recent_used_cpu)) {
6358 * Replace recent_used_cpu with prev as it is a potential
6359 * candidate for the next wake:
6361 p->recent_used_cpu = prev;
6362 return recent_used_cpu;
6366 * For asymmetric CPU capacity systems, our domain of interest is
6367 * sd_asym_cpucapacity rather than sd_llc.
6369 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6370 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6372 * On an asymmetric CPU capacity system where an exclusive
6373 * cpuset defines a symmetric island (i.e. one unique
6374 * capacity_orig value through the cpuset), the key will be set
6375 * but the CPUs within that cpuset will not have a domain with
6376 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6380 i = select_idle_capacity(p, sd, target);
6381 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6385 sd = rcu_dereference(per_cpu(sd_llc, target));
6389 if (sched_smt_active()) {
6390 has_idle_core = test_idle_cores(target, false);
6392 if (!has_idle_core && cpus_share_cache(prev, target)) {
6393 i = select_idle_smt(p, sd, prev);
6394 if ((unsigned int)i < nr_cpumask_bits)
6399 i = select_idle_cpu(p, sd, has_idle_core, target);
6400 if ((unsigned)i < nr_cpumask_bits)
6407 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6408 * @cpu: the CPU to get the utilization of
6410 * The unit of the return value must be the one of capacity so we can compare
6411 * the utilization with the capacity of the CPU that is available for CFS task
6412 * (ie cpu_capacity).
6414 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6415 * recent utilization of currently non-runnable tasks on a CPU. It represents
6416 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6417 * capacity_orig is the cpu_capacity available at the highest frequency
6418 * (arch_scale_freq_capacity()).
6419 * The utilization of a CPU converges towards a sum equal to or less than the
6420 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6421 * the running time on this CPU scaled by capacity_curr.
6423 * The estimated utilization of a CPU is defined to be the maximum between its
6424 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6425 * currently RUNNABLE on that CPU.
6426 * This allows to properly represent the expected utilization of a CPU which
6427 * has just got a big task running since a long sleep period. At the same time
6428 * however it preserves the benefits of the "blocked utilization" in
6429 * describing the potential for other tasks waking up on the same CPU.
6431 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6432 * higher than capacity_orig because of unfortunate rounding in
6433 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6434 * the average stabilizes with the new running time. We need to check that the
6435 * utilization stays within the range of [0..capacity_orig] and cap it if
6436 * necessary. Without utilization capping, a group could be seen as overloaded
6437 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6438 * available capacity. We allow utilization to overshoot capacity_curr (but not
6439 * capacity_orig) as it useful for predicting the capacity required after task
6440 * migrations (scheduler-driven DVFS).
6442 * Return: the (estimated) utilization for the specified CPU
6444 static inline unsigned long cpu_util(int cpu)
6446 struct cfs_rq *cfs_rq;
6449 cfs_rq = &cpu_rq(cpu)->cfs;
6450 util = READ_ONCE(cfs_rq->avg.util_avg);
6452 if (sched_feat(UTIL_EST))
6453 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6455 return min_t(unsigned long, util, capacity_orig_of(cpu));
6459 * cpu_util_without: compute cpu utilization without any contributions from *p
6460 * @cpu: the CPU which utilization is requested
6461 * @p: the task which utilization should be discounted
6463 * The utilization of a CPU is defined by the utilization of tasks currently
6464 * enqueued on that CPU as well as tasks which are currently sleeping after an
6465 * execution on that CPU.
6467 * This method returns the utilization of the specified CPU by discounting the
6468 * utilization of the specified task, whenever the task is currently
6469 * contributing to the CPU utilization.
6471 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6473 struct cfs_rq *cfs_rq;
6476 /* Task has no contribution or is new */
6477 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6478 return cpu_util(cpu);
6480 cfs_rq = &cpu_rq(cpu)->cfs;
6481 util = READ_ONCE(cfs_rq->avg.util_avg);
6483 /* Discount task's util from CPU's util */
6484 lsub_positive(&util, task_util(p));
6489 * a) if *p is the only task sleeping on this CPU, then:
6490 * cpu_util (== task_util) > util_est (== 0)
6491 * and thus we return:
6492 * cpu_util_without = (cpu_util - task_util) = 0
6494 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6496 * cpu_util >= task_util
6497 * cpu_util > util_est (== 0)
6498 * and thus we discount *p's blocked utilization to return:
6499 * cpu_util_without = (cpu_util - task_util) >= 0
6501 * c) if other tasks are RUNNABLE on that CPU and
6502 * util_est > cpu_util
6503 * then we use util_est since it returns a more restrictive
6504 * estimation of the spare capacity on that CPU, by just
6505 * considering the expected utilization of tasks already
6506 * runnable on that CPU.
6508 * Cases a) and b) are covered by the above code, while case c) is
6509 * covered by the following code when estimated utilization is
6512 if (sched_feat(UTIL_EST)) {
6513 unsigned int estimated =
6514 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6517 * Despite the following checks we still have a small window
6518 * for a possible race, when an execl's select_task_rq_fair()
6519 * races with LB's detach_task():
6522 * p->on_rq = TASK_ON_RQ_MIGRATING;
6523 * ---------------------------------- A
6524 * deactivate_task() \
6525 * dequeue_task() + RaceTime
6526 * util_est_dequeue() /
6527 * ---------------------------------- B
6529 * The additional check on "current == p" it's required to
6530 * properly fix the execl regression and it helps in further
6531 * reducing the chances for the above race.
6533 if (unlikely(task_on_rq_queued(p) || current == p))
6534 lsub_positive(&estimated, _task_util_est(p));
6536 util = max(util, estimated);
6540 * Utilization (estimated) can exceed the CPU capacity, thus let's
6541 * clamp to the maximum CPU capacity to ensure consistency with
6542 * the cpu_util call.
6544 return min_t(unsigned long, util, capacity_orig_of(cpu));
6548 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6551 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6553 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6554 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6557 * If @p migrates from @cpu to another, remove its contribution. Or,
6558 * if @p migrates from another CPU to @cpu, add its contribution. In
6559 * the other cases, @cpu is not impacted by the migration, so the
6560 * util_avg should already be correct.
6562 if (task_cpu(p) == cpu && dst_cpu != cpu)
6563 lsub_positive(&util, task_util(p));
6564 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6565 util += task_util(p);
6567 if (sched_feat(UTIL_EST)) {
6568 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6571 * During wake-up, the task isn't enqueued yet and doesn't
6572 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6573 * so just add it (if needed) to "simulate" what will be
6574 * cpu_util() after the task has been enqueued.
6577 util_est += _task_util_est(p);
6579 util = max(util, util_est);
6582 return min(util, capacity_orig_of(cpu));
6586 * compute_energy(): Estimates the energy that @pd would consume if @p was
6587 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6588 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6589 * to compute what would be the energy if we decided to actually migrate that
6593 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6595 struct cpumask *pd_mask = perf_domain_span(pd);
6596 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6597 unsigned long max_util = 0, sum_util = 0;
6598 unsigned long _cpu_cap = cpu_cap;
6601 _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask));
6604 * The capacity state of CPUs of the current rd can be driven by CPUs
6605 * of another rd if they belong to the same pd. So, account for the
6606 * utilization of these CPUs too by masking pd with cpu_online_mask
6607 * instead of the rd span.
6609 * If an entire pd is outside of the current rd, it will not appear in
6610 * its pd list and will not be accounted by compute_energy().
6612 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6613 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu);
6614 unsigned long cpu_util, util_running = util_freq;
6615 struct task_struct *tsk = NULL;
6618 * When @p is placed on @cpu:
6620 * util_running = max(cpu_util, cpu_util_est) +
6621 * max(task_util, _task_util_est)
6623 * while cpu_util_next is: max(cpu_util + task_util,
6624 * cpu_util_est + _task_util_est)
6626 if (cpu == dst_cpu) {
6629 cpu_util_next(cpu, p, -1) + task_util_est(p);
6633 * Busy time computation: utilization clamping is not
6634 * required since the ratio (sum_util / cpu_capacity)
6635 * is already enough to scale the EM reported power
6636 * consumption at the (eventually clamped) cpu_capacity.
6638 cpu_util = effective_cpu_util(cpu, util_running, cpu_cap,
6641 sum_util += min(cpu_util, _cpu_cap);
6644 * Performance domain frequency: utilization clamping
6645 * must be considered since it affects the selection
6646 * of the performance domain frequency.
6647 * NOTE: in case RT tasks are running, by default the
6648 * FREQUENCY_UTIL's utilization can be max OPP.
6650 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap,
6651 FREQUENCY_UTIL, tsk);
6652 max_util = max(max_util, min(cpu_util, _cpu_cap));
6655 return em_cpu_energy(pd->em_pd, max_util, sum_util, _cpu_cap);
6659 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6660 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6661 * spare capacity in each performance domain and uses it as a potential
6662 * candidate to execute the task. Then, it uses the Energy Model to figure
6663 * out which of the CPU candidates is the most energy-efficient.
6665 * The rationale for this heuristic is as follows. In a performance domain,
6666 * all the most energy efficient CPU candidates (according to the Energy
6667 * Model) are those for which we'll request a low frequency. When there are
6668 * several CPUs for which the frequency request will be the same, we don't
6669 * have enough data to break the tie between them, because the Energy Model
6670 * only includes active power costs. With this model, if we assume that
6671 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6672 * the maximum spare capacity in a performance domain is guaranteed to be among
6673 * the best candidates of the performance domain.
6675 * In practice, it could be preferable from an energy standpoint to pack
6676 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6677 * but that could also hurt our chances to go cluster idle, and we have no
6678 * ways to tell with the current Energy Model if this is actually a good
6679 * idea or not. So, find_energy_efficient_cpu() basically favors
6680 * cluster-packing, and spreading inside a cluster. That should at least be
6681 * a good thing for latency, and this is consistent with the idea that most
6682 * of the energy savings of EAS come from the asymmetry of the system, and
6683 * not so much from breaking the tie between identical CPUs. That's also the
6684 * reason why EAS is enabled in the topology code only for systems where
6685 * SD_ASYM_CPUCAPACITY is set.
6687 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6688 * they don't have any useful utilization data yet and it's not possible to
6689 * forecast their impact on energy consumption. Consequently, they will be
6690 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6691 * to be energy-inefficient in some use-cases. The alternative would be to
6692 * bias new tasks towards specific types of CPUs first, or to try to infer
6693 * their util_avg from the parent task, but those heuristics could hurt
6694 * other use-cases too. So, until someone finds a better way to solve this,
6695 * let's keep things simple by re-using the existing slow path.
6697 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6699 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6700 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6701 int cpu, best_energy_cpu = prev_cpu, target = -1;
6702 unsigned long cpu_cap, util, base_energy = 0;
6703 struct sched_domain *sd;
6704 struct perf_domain *pd;
6707 pd = rcu_dereference(rd->pd);
6708 if (!pd || READ_ONCE(rd->overutilized))
6712 * Energy-aware wake-up happens on the lowest sched_domain starting
6713 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6715 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6716 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6723 sync_entity_load_avg(&p->se);
6724 if (!task_util_est(p))
6727 for (; pd; pd = pd->next) {
6728 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6729 bool compute_prev_delta = false;
6730 unsigned long base_energy_pd;
6731 int max_spare_cap_cpu = -1;
6733 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6734 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6737 util = cpu_util_next(cpu, p, cpu);
6738 cpu_cap = capacity_of(cpu);
6739 spare_cap = cpu_cap;
6740 lsub_positive(&spare_cap, util);
6743 * Skip CPUs that cannot satisfy the capacity request.
6744 * IOW, placing the task there would make the CPU
6745 * overutilized. Take uclamp into account to see how
6746 * much capacity we can get out of the CPU; this is
6747 * aligned with sched_cpu_util().
6749 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6750 if (!fits_capacity(util, cpu_cap))
6753 if (cpu == prev_cpu) {
6754 /* Always use prev_cpu as a candidate. */
6755 compute_prev_delta = true;
6756 } else if (spare_cap > max_spare_cap) {
6758 * Find the CPU with the maximum spare capacity
6759 * in the performance domain.
6761 max_spare_cap = spare_cap;
6762 max_spare_cap_cpu = cpu;
6766 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6769 /* Compute the 'base' energy of the pd, without @p */
6770 base_energy_pd = compute_energy(p, -1, pd);
6771 base_energy += base_energy_pd;
6773 /* Evaluate the energy impact of using prev_cpu. */
6774 if (compute_prev_delta) {
6775 prev_delta = compute_energy(p, prev_cpu, pd);
6776 if (prev_delta < base_energy_pd)
6778 prev_delta -= base_energy_pd;
6779 best_delta = min(best_delta, prev_delta);
6782 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6783 if (max_spare_cap_cpu >= 0) {
6784 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6785 if (cur_delta < base_energy_pd)
6787 cur_delta -= base_energy_pd;
6788 if (cur_delta < best_delta) {
6789 best_delta = cur_delta;
6790 best_energy_cpu = max_spare_cap_cpu;
6797 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6798 * least 6% of the energy used by prev_cpu.
6800 if ((prev_delta == ULONG_MAX) ||
6801 (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6802 target = best_energy_cpu;
6813 * select_task_rq_fair: Select target runqueue for the waking task in domains
6814 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6815 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6817 * Balances load by selecting the idlest CPU in the idlest group, or under
6818 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6820 * Returns the target CPU number.
6823 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
6825 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6826 struct sched_domain *tmp, *sd = NULL;
6827 int cpu = smp_processor_id();
6828 int new_cpu = prev_cpu;
6829 int want_affine = 0;
6830 /* SD_flags and WF_flags share the first nibble */
6831 int sd_flag = wake_flags & 0xF;
6834 * required for stable ->cpus_allowed
6836 lockdep_assert_held(&p->pi_lock);
6837 if (wake_flags & WF_TTWU) {
6840 if (sched_energy_enabled()) {
6841 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6847 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6851 for_each_domain(cpu, tmp) {
6853 * If both 'cpu' and 'prev_cpu' are part of this domain,
6854 * cpu is a valid SD_WAKE_AFFINE target.
6856 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6857 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6858 if (cpu != prev_cpu)
6859 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6861 sd = NULL; /* Prefer wake_affine over balance flags */
6865 if (tmp->flags & sd_flag)
6867 else if (!want_affine)
6873 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6874 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
6876 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6879 current->recent_used_cpu = cpu;
6886 static void detach_entity_cfs_rq(struct sched_entity *se);
6889 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6890 * cfs_rq_of(p) references at time of call are still valid and identify the
6891 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6893 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6896 * As blocked tasks retain absolute vruntime the migration needs to
6897 * deal with this by subtracting the old and adding the new
6898 * min_vruntime -- the latter is done by enqueue_entity() when placing
6899 * the task on the new runqueue.
6901 if (READ_ONCE(p->__state) == TASK_WAKING) {
6902 struct sched_entity *se = &p->se;
6903 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6906 #ifndef CONFIG_64BIT
6907 u64 min_vruntime_copy;
6910 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6912 min_vruntime = cfs_rq->min_vruntime;
6913 } while (min_vruntime != min_vruntime_copy);
6915 min_vruntime = cfs_rq->min_vruntime;
6918 se->vruntime -= min_vruntime;
6921 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6923 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6924 * rq->lock and can modify state directly.
6926 lockdep_assert_rq_held(task_rq(p));
6927 detach_entity_cfs_rq(&p->se);
6931 * We are supposed to update the task to "current" time, then
6932 * its up to date and ready to go to new CPU/cfs_rq. But we
6933 * have difficulty in getting what current time is, so simply
6934 * throw away the out-of-date time. This will result in the
6935 * wakee task is less decayed, but giving the wakee more load
6938 remove_entity_load_avg(&p->se);
6941 /* Tell new CPU we are migrated */
6942 p->se.avg.last_update_time = 0;
6944 /* We have migrated, no longer consider this task hot */
6945 p->se.exec_start = 0;
6947 update_scan_period(p, new_cpu);
6950 static void task_dead_fair(struct task_struct *p)
6952 remove_entity_load_avg(&p->se);
6956 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6961 return newidle_balance(rq, rf) != 0;
6963 #endif /* CONFIG_SMP */
6965 static unsigned long wakeup_gran(struct sched_entity *se)
6967 unsigned long gran = sysctl_sched_wakeup_granularity;
6970 * Since its curr running now, convert the gran from real-time
6971 * to virtual-time in his units.
6973 * By using 'se' instead of 'curr' we penalize light tasks, so
6974 * they get preempted easier. That is, if 'se' < 'curr' then
6975 * the resulting gran will be larger, therefore penalizing the
6976 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6977 * be smaller, again penalizing the lighter task.
6979 * This is especially important for buddies when the leftmost
6980 * task is higher priority than the buddy.
6982 return calc_delta_fair(gran, se);
6986 * Should 'se' preempt 'curr'.
7000 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7002 s64 gran, vdiff = curr->vruntime - se->vruntime;
7007 gran = wakeup_gran(se);
7014 static void set_last_buddy(struct sched_entity *se)
7016 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
7019 for_each_sched_entity(se) {
7020 if (SCHED_WARN_ON(!se->on_rq))
7022 cfs_rq_of(se)->last = se;
7026 static void set_next_buddy(struct sched_entity *se)
7028 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
7031 for_each_sched_entity(se) {
7032 if (SCHED_WARN_ON(!se->on_rq))
7034 cfs_rq_of(se)->next = se;
7038 static void set_skip_buddy(struct sched_entity *se)
7040 for_each_sched_entity(se)
7041 cfs_rq_of(se)->skip = se;
7045 * Preempt the current task with a newly woken task if needed:
7047 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7049 struct task_struct *curr = rq->curr;
7050 struct sched_entity *se = &curr->se, *pse = &p->se;
7051 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7052 int scale = cfs_rq->nr_running >= sched_nr_latency;
7053 int next_buddy_marked = 0;
7055 if (unlikely(se == pse))
7059 * This is possible from callers such as attach_tasks(), in which we
7060 * unconditionally check_preempt_curr() after an enqueue (which may have
7061 * lead to a throttle). This both saves work and prevents false
7062 * next-buddy nomination below.
7064 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7067 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7068 set_next_buddy(pse);
7069 next_buddy_marked = 1;
7073 * We can come here with TIF_NEED_RESCHED already set from new task
7076 * Note: this also catches the edge-case of curr being in a throttled
7077 * group (e.g. via set_curr_task), since update_curr() (in the
7078 * enqueue of curr) will have resulted in resched being set. This
7079 * prevents us from potentially nominating it as a false LAST_BUDDY
7082 if (test_tsk_need_resched(curr))
7085 /* Idle tasks are by definition preempted by non-idle tasks. */
7086 if (unlikely(task_has_idle_policy(curr)) &&
7087 likely(!task_has_idle_policy(p)))
7091 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7092 * is driven by the tick):
7094 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7097 find_matching_se(&se, &pse);
7098 update_curr(cfs_rq_of(se));
7100 if (wakeup_preempt_entity(se, pse) == 1) {
7102 * Bias pick_next to pick the sched entity that is
7103 * triggering this preemption.
7105 if (!next_buddy_marked)
7106 set_next_buddy(pse);
7115 * Only set the backward buddy when the current task is still
7116 * on the rq. This can happen when a wakeup gets interleaved
7117 * with schedule on the ->pre_schedule() or idle_balance()
7118 * point, either of which can * drop the rq lock.
7120 * Also, during early boot the idle thread is in the fair class,
7121 * for obvious reasons its a bad idea to schedule back to it.
7123 if (unlikely(!se->on_rq || curr == rq->idle))
7126 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7131 static struct task_struct *pick_task_fair(struct rq *rq)
7133 struct sched_entity *se;
7134 struct cfs_rq *cfs_rq;
7138 if (!cfs_rq->nr_running)
7142 struct sched_entity *curr = cfs_rq->curr;
7144 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7147 update_curr(cfs_rq);
7151 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7155 se = pick_next_entity(cfs_rq, curr);
7156 cfs_rq = group_cfs_rq(se);
7163 struct task_struct *
7164 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7166 struct cfs_rq *cfs_rq = &rq->cfs;
7167 struct sched_entity *se;
7168 struct task_struct *p;
7172 if (!sched_fair_runnable(rq))
7175 #ifdef CONFIG_FAIR_GROUP_SCHED
7176 if (!prev || prev->sched_class != &fair_sched_class)
7180 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7181 * likely that a next task is from the same cgroup as the current.
7183 * Therefore attempt to avoid putting and setting the entire cgroup
7184 * hierarchy, only change the part that actually changes.
7188 struct sched_entity *curr = cfs_rq->curr;
7191 * Since we got here without doing put_prev_entity() we also
7192 * have to consider cfs_rq->curr. If it is still a runnable
7193 * entity, update_curr() will update its vruntime, otherwise
7194 * forget we've ever seen it.
7198 update_curr(cfs_rq);
7203 * This call to check_cfs_rq_runtime() will do the
7204 * throttle and dequeue its entity in the parent(s).
7205 * Therefore the nr_running test will indeed
7208 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7211 if (!cfs_rq->nr_running)
7218 se = pick_next_entity(cfs_rq, curr);
7219 cfs_rq = group_cfs_rq(se);
7225 * Since we haven't yet done put_prev_entity and if the selected task
7226 * is a different task than we started out with, try and touch the
7227 * least amount of cfs_rqs.
7230 struct sched_entity *pse = &prev->se;
7232 while (!(cfs_rq = is_same_group(se, pse))) {
7233 int se_depth = se->depth;
7234 int pse_depth = pse->depth;
7236 if (se_depth <= pse_depth) {
7237 put_prev_entity(cfs_rq_of(pse), pse);
7238 pse = parent_entity(pse);
7240 if (se_depth >= pse_depth) {
7241 set_next_entity(cfs_rq_of(se), se);
7242 se = parent_entity(se);
7246 put_prev_entity(cfs_rq, pse);
7247 set_next_entity(cfs_rq, se);
7254 put_prev_task(rq, prev);
7257 se = pick_next_entity(cfs_rq, NULL);
7258 set_next_entity(cfs_rq, se);
7259 cfs_rq = group_cfs_rq(se);
7264 done: __maybe_unused;
7267 * Move the next running task to the front of
7268 * the list, so our cfs_tasks list becomes MRU
7271 list_move(&p->se.group_node, &rq->cfs_tasks);
7274 if (hrtick_enabled_fair(rq))
7275 hrtick_start_fair(rq, p);
7277 update_misfit_status(p, rq);
7285 new_tasks = newidle_balance(rq, rf);
7288 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7289 * possible for any higher priority task to appear. In that case we
7290 * must re-start the pick_next_entity() loop.
7299 * rq is about to be idle, check if we need to update the
7300 * lost_idle_time of clock_pelt
7302 update_idle_rq_clock_pelt(rq);
7307 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7309 return pick_next_task_fair(rq, NULL, NULL);
7313 * Account for a descheduled task:
7315 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7317 struct sched_entity *se = &prev->se;
7318 struct cfs_rq *cfs_rq;
7320 for_each_sched_entity(se) {
7321 cfs_rq = cfs_rq_of(se);
7322 put_prev_entity(cfs_rq, se);
7327 * sched_yield() is very simple
7329 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7331 static void yield_task_fair(struct rq *rq)
7333 struct task_struct *curr = rq->curr;
7334 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7335 struct sched_entity *se = &curr->se;
7338 * Are we the only task in the tree?
7340 if (unlikely(rq->nr_running == 1))
7343 clear_buddies(cfs_rq, se);
7345 if (curr->policy != SCHED_BATCH) {
7346 update_rq_clock(rq);
7348 * Update run-time statistics of the 'current'.
7350 update_curr(cfs_rq);
7352 * Tell update_rq_clock() that we've just updated,
7353 * so we don't do microscopic update in schedule()
7354 * and double the fastpath cost.
7356 rq_clock_skip_update(rq);
7362 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7364 struct sched_entity *se = &p->se;
7366 /* throttled hierarchies are not runnable */
7367 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7370 /* Tell the scheduler that we'd really like pse to run next. */
7373 yield_task_fair(rq);
7379 /**************************************************
7380 * Fair scheduling class load-balancing methods.
7384 * The purpose of load-balancing is to achieve the same basic fairness the
7385 * per-CPU scheduler provides, namely provide a proportional amount of compute
7386 * time to each task. This is expressed in the following equation:
7388 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7390 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7391 * W_i,0 is defined as:
7393 * W_i,0 = \Sum_j w_i,j (2)
7395 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7396 * is derived from the nice value as per sched_prio_to_weight[].
7398 * The weight average is an exponential decay average of the instantaneous
7401 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7403 * C_i is the compute capacity of CPU i, typically it is the
7404 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7405 * can also include other factors [XXX].
7407 * To achieve this balance we define a measure of imbalance which follows
7408 * directly from (1):
7410 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7412 * We them move tasks around to minimize the imbalance. In the continuous
7413 * function space it is obvious this converges, in the discrete case we get
7414 * a few fun cases generally called infeasible weight scenarios.
7417 * - infeasible weights;
7418 * - local vs global optima in the discrete case. ]
7423 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7424 * for all i,j solution, we create a tree of CPUs that follows the hardware
7425 * topology where each level pairs two lower groups (or better). This results
7426 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7427 * tree to only the first of the previous level and we decrease the frequency
7428 * of load-balance at each level inv. proportional to the number of CPUs in
7434 * \Sum { --- * --- * 2^i } = O(n) (5)
7436 * `- size of each group
7437 * | | `- number of CPUs doing load-balance
7439 * `- sum over all levels
7441 * Coupled with a limit on how many tasks we can migrate every balance pass,
7442 * this makes (5) the runtime complexity of the balancer.
7444 * An important property here is that each CPU is still (indirectly) connected
7445 * to every other CPU in at most O(log n) steps:
7447 * The adjacency matrix of the resulting graph is given by:
7450 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7453 * And you'll find that:
7455 * A^(log_2 n)_i,j != 0 for all i,j (7)
7457 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7458 * The task movement gives a factor of O(m), giving a convergence complexity
7461 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7466 * In order to avoid CPUs going idle while there's still work to do, new idle
7467 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7468 * tree itself instead of relying on other CPUs to bring it work.
7470 * This adds some complexity to both (5) and (8) but it reduces the total idle
7478 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7481 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7486 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7488 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7490 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7493 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7494 * rewrite all of this once again.]
7497 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7499 enum fbq_type { regular, remote, all };
7502 * 'group_type' describes the group of CPUs at the moment of load balancing.
7504 * The enum is ordered by pulling priority, with the group with lowest priority
7505 * first so the group_type can simply be compared when selecting the busiest
7506 * group. See update_sd_pick_busiest().
7509 /* The group has spare capacity that can be used to run more tasks. */
7510 group_has_spare = 0,
7512 * The group is fully used and the tasks don't compete for more CPU
7513 * cycles. Nevertheless, some tasks might wait before running.
7517 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7518 * and must be migrated to a more powerful CPU.
7522 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7523 * and the task should be migrated to it instead of running on the
7528 * The tasks' affinity constraints previously prevented the scheduler
7529 * from balancing the load across the system.
7533 * The CPU is overloaded and can't provide expected CPU cycles to all
7539 enum migration_type {
7546 #define LBF_ALL_PINNED 0x01
7547 #define LBF_NEED_BREAK 0x02
7548 #define LBF_DST_PINNED 0x04
7549 #define LBF_SOME_PINNED 0x08
7550 #define LBF_ACTIVE_LB 0x10
7553 struct sched_domain *sd;
7561 struct cpumask *dst_grpmask;
7563 enum cpu_idle_type idle;
7565 /* The set of CPUs under consideration for load-balancing */
7566 struct cpumask *cpus;
7571 unsigned int loop_break;
7572 unsigned int loop_max;
7574 enum fbq_type fbq_type;
7575 enum migration_type migration_type;
7576 struct list_head tasks;
7580 * Is this task likely cache-hot:
7582 static int task_hot(struct task_struct *p, struct lb_env *env)
7586 lockdep_assert_rq_held(env->src_rq);
7588 if (p->sched_class != &fair_sched_class)
7591 if (unlikely(task_has_idle_policy(p)))
7594 /* SMT siblings share cache */
7595 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7599 * Buddy candidates are cache hot:
7601 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7602 (&p->se == cfs_rq_of(&p->se)->next ||
7603 &p->se == cfs_rq_of(&p->se)->last))
7606 if (sysctl_sched_migration_cost == -1)
7610 * Don't migrate task if the task's cookie does not match
7611 * with the destination CPU's core cookie.
7613 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7616 if (sysctl_sched_migration_cost == 0)
7619 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7621 return delta < (s64)sysctl_sched_migration_cost;
7624 #ifdef CONFIG_NUMA_BALANCING
7626 * Returns 1, if task migration degrades locality
7627 * Returns 0, if task migration improves locality i.e migration preferred.
7628 * Returns -1, if task migration is not affected by locality.
7630 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7632 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7633 unsigned long src_weight, dst_weight;
7634 int src_nid, dst_nid, dist;
7636 if (!static_branch_likely(&sched_numa_balancing))
7639 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7642 src_nid = cpu_to_node(env->src_cpu);
7643 dst_nid = cpu_to_node(env->dst_cpu);
7645 if (src_nid == dst_nid)
7648 /* Migrating away from the preferred node is always bad. */
7649 if (src_nid == p->numa_preferred_nid) {
7650 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7656 /* Encourage migration to the preferred node. */
7657 if (dst_nid == p->numa_preferred_nid)
7660 /* Leaving a core idle is often worse than degrading locality. */
7661 if (env->idle == CPU_IDLE)
7664 dist = node_distance(src_nid, dst_nid);
7666 src_weight = group_weight(p, src_nid, dist);
7667 dst_weight = group_weight(p, dst_nid, dist);
7669 src_weight = task_weight(p, src_nid, dist);
7670 dst_weight = task_weight(p, dst_nid, dist);
7673 return dst_weight < src_weight;
7677 static inline int migrate_degrades_locality(struct task_struct *p,
7685 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7688 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7692 lockdep_assert_rq_held(env->src_rq);
7695 * We do not migrate tasks that are:
7696 * 1) throttled_lb_pair, or
7697 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7698 * 3) running (obviously), or
7699 * 4) are cache-hot on their current CPU.
7701 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7704 /* Disregard pcpu kthreads; they are where they need to be. */
7705 if (kthread_is_per_cpu(p))
7708 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7711 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7713 env->flags |= LBF_SOME_PINNED;
7716 * Remember if this task can be migrated to any other CPU in
7717 * our sched_group. We may want to revisit it if we couldn't
7718 * meet load balance goals by pulling other tasks on src_cpu.
7720 * Avoid computing new_dst_cpu
7722 * - if we have already computed one in current iteration
7723 * - if it's an active balance
7725 if (env->idle == CPU_NEWLY_IDLE ||
7726 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7729 /* Prevent to re-select dst_cpu via env's CPUs: */
7730 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7731 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7732 env->flags |= LBF_DST_PINNED;
7733 env->new_dst_cpu = cpu;
7741 /* Record that we found at least one task that could run on dst_cpu */
7742 env->flags &= ~LBF_ALL_PINNED;
7744 if (task_running(env->src_rq, p)) {
7745 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7750 * Aggressive migration if:
7752 * 2) destination numa is preferred
7753 * 3) task is cache cold, or
7754 * 4) too many balance attempts have failed.
7756 if (env->flags & LBF_ACTIVE_LB)
7759 tsk_cache_hot = migrate_degrades_locality(p, env);
7760 if (tsk_cache_hot == -1)
7761 tsk_cache_hot = task_hot(p, env);
7763 if (tsk_cache_hot <= 0 ||
7764 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7765 if (tsk_cache_hot == 1) {
7766 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7767 schedstat_inc(p->se.statistics.nr_forced_migrations);
7772 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7777 * detach_task() -- detach the task for the migration specified in env
7779 static void detach_task(struct task_struct *p, struct lb_env *env)
7781 lockdep_assert_rq_held(env->src_rq);
7783 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7784 set_task_cpu(p, env->dst_cpu);
7788 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7789 * part of active balancing operations within "domain".
7791 * Returns a task if successful and NULL otherwise.
7793 static struct task_struct *detach_one_task(struct lb_env *env)
7795 struct task_struct *p;
7797 lockdep_assert_rq_held(env->src_rq);
7799 list_for_each_entry_reverse(p,
7800 &env->src_rq->cfs_tasks, se.group_node) {
7801 if (!can_migrate_task(p, env))
7804 detach_task(p, env);
7807 * Right now, this is only the second place where
7808 * lb_gained[env->idle] is updated (other is detach_tasks)
7809 * so we can safely collect stats here rather than
7810 * inside detach_tasks().
7812 schedstat_inc(env->sd->lb_gained[env->idle]);
7818 static const unsigned int sched_nr_migrate_break = 32;
7821 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7822 * busiest_rq, as part of a balancing operation within domain "sd".
7824 * Returns number of detached tasks if successful and 0 otherwise.
7826 static int detach_tasks(struct lb_env *env)
7828 struct list_head *tasks = &env->src_rq->cfs_tasks;
7829 unsigned long util, load;
7830 struct task_struct *p;
7833 lockdep_assert_rq_held(env->src_rq);
7836 * Source run queue has been emptied by another CPU, clear
7837 * LBF_ALL_PINNED flag as we will not test any task.
7839 if (env->src_rq->nr_running <= 1) {
7840 env->flags &= ~LBF_ALL_PINNED;
7844 if (env->imbalance <= 0)
7847 while (!list_empty(tasks)) {
7849 * We don't want to steal all, otherwise we may be treated likewise,
7850 * which could at worst lead to a livelock crash.
7852 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7855 p = list_last_entry(tasks, struct task_struct, se.group_node);
7858 /* We've more or less seen every task there is, call it quits */
7859 if (env->loop > env->loop_max)
7862 /* take a breather every nr_migrate tasks */
7863 if (env->loop > env->loop_break) {
7864 env->loop_break += sched_nr_migrate_break;
7865 env->flags |= LBF_NEED_BREAK;
7869 if (!can_migrate_task(p, env))
7872 switch (env->migration_type) {
7875 * Depending of the number of CPUs and tasks and the
7876 * cgroup hierarchy, task_h_load() can return a null
7877 * value. Make sure that env->imbalance decreases
7878 * otherwise detach_tasks() will stop only after
7879 * detaching up to loop_max tasks.
7881 load = max_t(unsigned long, task_h_load(p), 1);
7883 if (sched_feat(LB_MIN) &&
7884 load < 16 && !env->sd->nr_balance_failed)
7888 * Make sure that we don't migrate too much load.
7889 * Nevertheless, let relax the constraint if
7890 * scheduler fails to find a good waiting task to
7893 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
7896 env->imbalance -= load;
7900 util = task_util_est(p);
7902 if (util > env->imbalance)
7905 env->imbalance -= util;
7912 case migrate_misfit:
7913 /* This is not a misfit task */
7914 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7921 detach_task(p, env);
7922 list_add(&p->se.group_node, &env->tasks);
7926 #ifdef CONFIG_PREEMPTION
7928 * NEWIDLE balancing is a source of latency, so preemptible
7929 * kernels will stop after the first task is detached to minimize
7930 * the critical section.
7932 if (env->idle == CPU_NEWLY_IDLE)
7937 * We only want to steal up to the prescribed amount of
7940 if (env->imbalance <= 0)
7945 list_move(&p->se.group_node, tasks);
7949 * Right now, this is one of only two places we collect this stat
7950 * so we can safely collect detach_one_task() stats here rather
7951 * than inside detach_one_task().
7953 schedstat_add(env->sd->lb_gained[env->idle], detached);
7959 * attach_task() -- attach the task detached by detach_task() to its new rq.
7961 static void attach_task(struct rq *rq, struct task_struct *p)
7963 lockdep_assert_rq_held(rq);
7965 BUG_ON(task_rq(p) != rq);
7966 activate_task(rq, p, ENQUEUE_NOCLOCK);
7967 check_preempt_curr(rq, p, 0);
7971 * attach_one_task() -- attaches the task returned from detach_one_task() to
7974 static void attach_one_task(struct rq *rq, struct task_struct *p)
7979 update_rq_clock(rq);
7985 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7988 static void attach_tasks(struct lb_env *env)
7990 struct list_head *tasks = &env->tasks;
7991 struct task_struct *p;
7994 rq_lock(env->dst_rq, &rf);
7995 update_rq_clock(env->dst_rq);
7997 while (!list_empty(tasks)) {
7998 p = list_first_entry(tasks, struct task_struct, se.group_node);
7999 list_del_init(&p->se.group_node);
8001 attach_task(env->dst_rq, p);
8004 rq_unlock(env->dst_rq, &rf);
8007 #ifdef CONFIG_NO_HZ_COMMON
8008 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8010 if (cfs_rq->avg.load_avg)
8013 if (cfs_rq->avg.util_avg)
8019 static inline bool others_have_blocked(struct rq *rq)
8021 if (READ_ONCE(rq->avg_rt.util_avg))
8024 if (READ_ONCE(rq->avg_dl.util_avg))
8027 if (thermal_load_avg(rq))
8030 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8031 if (READ_ONCE(rq->avg_irq.util_avg))
8038 static inline void update_blocked_load_tick(struct rq *rq)
8040 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8043 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8046 rq->has_blocked_load = 0;
8049 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8050 static inline bool others_have_blocked(struct rq *rq) { return false; }
8051 static inline void update_blocked_load_tick(struct rq *rq) {}
8052 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8055 static bool __update_blocked_others(struct rq *rq, bool *done)
8057 const struct sched_class *curr_class;
8058 u64 now = rq_clock_pelt(rq);
8059 unsigned long thermal_pressure;
8063 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8064 * DL and IRQ signals have been updated before updating CFS.
8066 curr_class = rq->curr->sched_class;
8068 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8070 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8071 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8072 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8073 update_irq_load_avg(rq, 0);
8075 if (others_have_blocked(rq))
8081 #ifdef CONFIG_FAIR_GROUP_SCHED
8083 static bool __update_blocked_fair(struct rq *rq, bool *done)
8085 struct cfs_rq *cfs_rq, *pos;
8086 bool decayed = false;
8087 int cpu = cpu_of(rq);
8090 * Iterates the task_group tree in a bottom up fashion, see
8091 * list_add_leaf_cfs_rq() for details.
8093 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8094 struct sched_entity *se;
8096 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8097 update_tg_load_avg(cfs_rq);
8099 if (cfs_rq == &rq->cfs)
8103 /* Propagate pending load changes to the parent, if any: */
8104 se = cfs_rq->tg->se[cpu];
8105 if (se && !skip_blocked_update(se))
8106 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8109 * There can be a lot of idle CPU cgroups. Don't let fully
8110 * decayed cfs_rqs linger on the list.
8112 if (cfs_rq_is_decayed(cfs_rq))
8113 list_del_leaf_cfs_rq(cfs_rq);
8115 /* Don't need periodic decay once load/util_avg are null */
8116 if (cfs_rq_has_blocked(cfs_rq))
8124 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8125 * This needs to be done in a top-down fashion because the load of a child
8126 * group is a fraction of its parents load.
8128 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8130 struct rq *rq = rq_of(cfs_rq);
8131 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8132 unsigned long now = jiffies;
8135 if (cfs_rq->last_h_load_update == now)
8138 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8139 for_each_sched_entity(se) {
8140 cfs_rq = cfs_rq_of(se);
8141 WRITE_ONCE(cfs_rq->h_load_next, se);
8142 if (cfs_rq->last_h_load_update == now)
8147 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8148 cfs_rq->last_h_load_update = now;
8151 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8152 load = cfs_rq->h_load;
8153 load = div64_ul(load * se->avg.load_avg,
8154 cfs_rq_load_avg(cfs_rq) + 1);
8155 cfs_rq = group_cfs_rq(se);
8156 cfs_rq->h_load = load;
8157 cfs_rq->last_h_load_update = now;
8161 static unsigned long task_h_load(struct task_struct *p)
8163 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8165 update_cfs_rq_h_load(cfs_rq);
8166 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8167 cfs_rq_load_avg(cfs_rq) + 1);
8170 static bool __update_blocked_fair(struct rq *rq, bool *done)
8172 struct cfs_rq *cfs_rq = &rq->cfs;
8175 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8176 if (cfs_rq_has_blocked(cfs_rq))
8182 static unsigned long task_h_load(struct task_struct *p)
8184 return p->se.avg.load_avg;
8188 static void update_blocked_averages(int cpu)
8190 bool decayed = false, done = true;
8191 struct rq *rq = cpu_rq(cpu);
8194 rq_lock_irqsave(rq, &rf);
8195 update_blocked_load_tick(rq);
8196 update_rq_clock(rq);
8198 decayed |= __update_blocked_others(rq, &done);
8199 decayed |= __update_blocked_fair(rq, &done);
8201 update_blocked_load_status(rq, !done);
8203 cpufreq_update_util(rq, 0);
8204 rq_unlock_irqrestore(rq, &rf);
8207 /********** Helpers for find_busiest_group ************************/
8210 * sg_lb_stats - stats of a sched_group required for load_balancing
8212 struct sg_lb_stats {
8213 unsigned long avg_load; /*Avg load across the CPUs of the group */
8214 unsigned long group_load; /* Total load over the CPUs of the group */
8215 unsigned long group_capacity;
8216 unsigned long group_util; /* Total utilization over the CPUs of the group */
8217 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8218 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8219 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8220 unsigned int idle_cpus;
8221 unsigned int group_weight;
8222 enum group_type group_type;
8223 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8224 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8225 #ifdef CONFIG_NUMA_BALANCING
8226 unsigned int nr_numa_running;
8227 unsigned int nr_preferred_running;
8232 * sd_lb_stats - Structure to store the statistics of a sched_domain
8233 * during load balancing.
8235 struct sd_lb_stats {
8236 struct sched_group *busiest; /* Busiest group in this sd */
8237 struct sched_group *local; /* Local group in this sd */
8238 unsigned long total_load; /* Total load of all groups in sd */
8239 unsigned long total_capacity; /* Total capacity of all groups in sd */
8240 unsigned long avg_load; /* Average load across all groups in sd */
8241 unsigned int prefer_sibling; /* tasks should go to sibling first */
8243 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8244 struct sg_lb_stats local_stat; /* Statistics of the local group */
8247 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8250 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8251 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8252 * We must however set busiest_stat::group_type and
8253 * busiest_stat::idle_cpus to the worst busiest group because
8254 * update_sd_pick_busiest() reads these before assignment.
8256 *sds = (struct sd_lb_stats){
8260 .total_capacity = 0UL,
8262 .idle_cpus = UINT_MAX,
8263 .group_type = group_has_spare,
8268 static unsigned long scale_rt_capacity(int cpu)
8270 struct rq *rq = cpu_rq(cpu);
8271 unsigned long max = arch_scale_cpu_capacity(cpu);
8272 unsigned long used, free;
8275 irq = cpu_util_irq(rq);
8277 if (unlikely(irq >= max))
8281 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8282 * (running and not running) with weights 0 and 1024 respectively.
8283 * avg_thermal.load_avg tracks thermal pressure and the weighted
8284 * average uses the actual delta max capacity(load).
8286 used = READ_ONCE(rq->avg_rt.util_avg);
8287 used += READ_ONCE(rq->avg_dl.util_avg);
8288 used += thermal_load_avg(rq);
8290 if (unlikely(used >= max))
8295 return scale_irq_capacity(free, irq, max);
8298 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8300 unsigned long capacity = scale_rt_capacity(cpu);
8301 struct sched_group *sdg = sd->groups;
8303 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8308 cpu_rq(cpu)->cpu_capacity = capacity;
8309 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8311 sdg->sgc->capacity = capacity;
8312 sdg->sgc->min_capacity = capacity;
8313 sdg->sgc->max_capacity = capacity;
8316 void update_group_capacity(struct sched_domain *sd, int cpu)
8318 struct sched_domain *child = sd->child;
8319 struct sched_group *group, *sdg = sd->groups;
8320 unsigned long capacity, min_capacity, max_capacity;
8321 unsigned long interval;
8323 interval = msecs_to_jiffies(sd->balance_interval);
8324 interval = clamp(interval, 1UL, max_load_balance_interval);
8325 sdg->sgc->next_update = jiffies + interval;
8328 update_cpu_capacity(sd, cpu);
8333 min_capacity = ULONG_MAX;
8336 if (child->flags & SD_OVERLAP) {
8338 * SD_OVERLAP domains cannot assume that child groups
8339 * span the current group.
8342 for_each_cpu(cpu, sched_group_span(sdg)) {
8343 unsigned long cpu_cap = capacity_of(cpu);
8345 capacity += cpu_cap;
8346 min_capacity = min(cpu_cap, min_capacity);
8347 max_capacity = max(cpu_cap, max_capacity);
8351 * !SD_OVERLAP domains can assume that child groups
8352 * span the current group.
8355 group = child->groups;
8357 struct sched_group_capacity *sgc = group->sgc;
8359 capacity += sgc->capacity;
8360 min_capacity = min(sgc->min_capacity, min_capacity);
8361 max_capacity = max(sgc->max_capacity, max_capacity);
8362 group = group->next;
8363 } while (group != child->groups);
8366 sdg->sgc->capacity = capacity;
8367 sdg->sgc->min_capacity = min_capacity;
8368 sdg->sgc->max_capacity = max_capacity;
8372 * Check whether the capacity of the rq has been noticeably reduced by side
8373 * activity. The imbalance_pct is used for the threshold.
8374 * Return true is the capacity is reduced
8377 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8379 return ((rq->cpu_capacity * sd->imbalance_pct) <
8380 (rq->cpu_capacity_orig * 100));
8384 * Check whether a rq has a misfit task and if it looks like we can actually
8385 * help that task: we can migrate the task to a CPU of higher capacity, or
8386 * the task's current CPU is heavily pressured.
8388 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8390 return rq->misfit_task_load &&
8391 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8392 check_cpu_capacity(rq, sd));
8396 * Group imbalance indicates (and tries to solve) the problem where balancing
8397 * groups is inadequate due to ->cpus_ptr constraints.
8399 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8400 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8403 * { 0 1 2 3 } { 4 5 6 7 }
8406 * If we were to balance group-wise we'd place two tasks in the first group and
8407 * two tasks in the second group. Clearly this is undesired as it will overload
8408 * cpu 3 and leave one of the CPUs in the second group unused.
8410 * The current solution to this issue is detecting the skew in the first group
8411 * by noticing the lower domain failed to reach balance and had difficulty
8412 * moving tasks due to affinity constraints.
8414 * When this is so detected; this group becomes a candidate for busiest; see
8415 * update_sd_pick_busiest(). And calculate_imbalance() and
8416 * find_busiest_group() avoid some of the usual balance conditions to allow it
8417 * to create an effective group imbalance.
8419 * This is a somewhat tricky proposition since the next run might not find the
8420 * group imbalance and decide the groups need to be balanced again. A most
8421 * subtle and fragile situation.
8424 static inline int sg_imbalanced(struct sched_group *group)
8426 return group->sgc->imbalance;
8430 * group_has_capacity returns true if the group has spare capacity that could
8431 * be used by some tasks.
8432 * We consider that a group has spare capacity if the * number of task is
8433 * smaller than the number of CPUs or if the utilization is lower than the
8434 * available capacity for CFS tasks.
8435 * For the latter, we use a threshold to stabilize the state, to take into
8436 * account the variance of the tasks' load and to return true if the available
8437 * capacity in meaningful for the load balancer.
8438 * As an example, an available capacity of 1% can appear but it doesn't make
8439 * any benefit for the load balance.
8442 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8444 if (sgs->sum_nr_running < sgs->group_weight)
8447 if ((sgs->group_capacity * imbalance_pct) <
8448 (sgs->group_runnable * 100))
8451 if ((sgs->group_capacity * 100) >
8452 (sgs->group_util * imbalance_pct))
8459 * group_is_overloaded returns true if the group has more tasks than it can
8461 * group_is_overloaded is not equals to !group_has_capacity because a group
8462 * with the exact right number of tasks, has no more spare capacity but is not
8463 * overloaded so both group_has_capacity and group_is_overloaded return
8467 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8469 if (sgs->sum_nr_running <= sgs->group_weight)
8472 if ((sgs->group_capacity * 100) <
8473 (sgs->group_util * imbalance_pct))
8476 if ((sgs->group_capacity * imbalance_pct) <
8477 (sgs->group_runnable * 100))
8484 group_type group_classify(unsigned int imbalance_pct,
8485 struct sched_group *group,
8486 struct sg_lb_stats *sgs)
8488 if (group_is_overloaded(imbalance_pct, sgs))
8489 return group_overloaded;
8491 if (sg_imbalanced(group))
8492 return group_imbalanced;
8494 if (sgs->group_asym_packing)
8495 return group_asym_packing;
8497 if (sgs->group_misfit_task_load)
8498 return group_misfit_task;
8500 if (!group_has_capacity(imbalance_pct, sgs))
8501 return group_fully_busy;
8503 return group_has_spare;
8507 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8508 * @env: The load balancing environment.
8509 * @group: sched_group whose statistics are to be updated.
8510 * @sgs: variable to hold the statistics for this group.
8511 * @sg_status: Holds flag indicating the status of the sched_group
8513 static inline void update_sg_lb_stats(struct lb_env *env,
8514 struct sched_group *group,
8515 struct sg_lb_stats *sgs,
8518 int i, nr_running, local_group;
8520 memset(sgs, 0, sizeof(*sgs));
8522 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8524 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8525 struct rq *rq = cpu_rq(i);
8527 sgs->group_load += cpu_load(rq);
8528 sgs->group_util += cpu_util(i);
8529 sgs->group_runnable += cpu_runnable(rq);
8530 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8532 nr_running = rq->nr_running;
8533 sgs->sum_nr_running += nr_running;
8536 *sg_status |= SG_OVERLOAD;
8538 if (cpu_overutilized(i))
8539 *sg_status |= SG_OVERUTILIZED;
8541 #ifdef CONFIG_NUMA_BALANCING
8542 sgs->nr_numa_running += rq->nr_numa_running;
8543 sgs->nr_preferred_running += rq->nr_preferred_running;
8546 * No need to call idle_cpu() if nr_running is not 0
8548 if (!nr_running && idle_cpu(i)) {
8550 /* Idle cpu can't have misfit task */
8557 /* Check for a misfit task on the cpu */
8558 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8559 sgs->group_misfit_task_load < rq->misfit_task_load) {
8560 sgs->group_misfit_task_load = rq->misfit_task_load;
8561 *sg_status |= SG_OVERLOAD;
8565 /* Check if dst CPU is idle and preferred to this group */
8566 if (env->sd->flags & SD_ASYM_PACKING &&
8567 env->idle != CPU_NOT_IDLE &&
8568 sgs->sum_h_nr_running &&
8569 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8570 sgs->group_asym_packing = 1;
8573 sgs->group_capacity = group->sgc->capacity;
8575 sgs->group_weight = group->group_weight;
8577 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8579 /* Computing avg_load makes sense only when group is overloaded */
8580 if (sgs->group_type == group_overloaded)
8581 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8582 sgs->group_capacity;
8586 * update_sd_pick_busiest - return 1 on busiest group
8587 * @env: The load balancing environment.
8588 * @sds: sched_domain statistics
8589 * @sg: sched_group candidate to be checked for being the busiest
8590 * @sgs: sched_group statistics
8592 * Determine if @sg is a busier group than the previously selected
8595 * Return: %true if @sg is a busier group than the previously selected
8596 * busiest group. %false otherwise.
8598 static bool update_sd_pick_busiest(struct lb_env *env,
8599 struct sd_lb_stats *sds,
8600 struct sched_group *sg,
8601 struct sg_lb_stats *sgs)
8603 struct sg_lb_stats *busiest = &sds->busiest_stat;
8605 /* Make sure that there is at least one task to pull */
8606 if (!sgs->sum_h_nr_running)
8610 * Don't try to pull misfit tasks we can't help.
8611 * We can use max_capacity here as reduction in capacity on some
8612 * CPUs in the group should either be possible to resolve
8613 * internally or be covered by avg_load imbalance (eventually).
8615 if (sgs->group_type == group_misfit_task &&
8616 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8617 sds->local_stat.group_type != group_has_spare))
8620 if (sgs->group_type > busiest->group_type)
8623 if (sgs->group_type < busiest->group_type)
8627 * The candidate and the current busiest group are the same type of
8628 * group. Let check which one is the busiest according to the type.
8631 switch (sgs->group_type) {
8632 case group_overloaded:
8633 /* Select the overloaded group with highest avg_load. */
8634 if (sgs->avg_load <= busiest->avg_load)
8638 case group_imbalanced:
8640 * Select the 1st imbalanced group as we don't have any way to
8641 * choose one more than another.
8645 case group_asym_packing:
8646 /* Prefer to move from lowest priority CPU's work */
8647 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8651 case group_misfit_task:
8653 * If we have more than one misfit sg go with the biggest
8656 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8660 case group_fully_busy:
8662 * Select the fully busy group with highest avg_load. In
8663 * theory, there is no need to pull task from such kind of
8664 * group because tasks have all compute capacity that they need
8665 * but we can still improve the overall throughput by reducing
8666 * contention when accessing shared HW resources.
8668 * XXX for now avg_load is not computed and always 0 so we
8669 * select the 1st one.
8671 if (sgs->avg_load <= busiest->avg_load)
8675 case group_has_spare:
8677 * Select not overloaded group with lowest number of idle cpus
8678 * and highest number of running tasks. We could also compare
8679 * the spare capacity which is more stable but it can end up
8680 * that the group has less spare capacity but finally more idle
8681 * CPUs which means less opportunity to pull tasks.
8683 if (sgs->idle_cpus > busiest->idle_cpus)
8685 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8686 (sgs->sum_nr_running <= busiest->sum_nr_running))
8693 * Candidate sg has no more than one task per CPU and has higher
8694 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8695 * throughput. Maximize throughput, power/energy consequences are not
8698 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8699 (sgs->group_type <= group_fully_busy) &&
8700 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
8706 #ifdef CONFIG_NUMA_BALANCING
8707 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8709 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8711 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8716 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8718 if (rq->nr_running > rq->nr_numa_running)
8720 if (rq->nr_running > rq->nr_preferred_running)
8725 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8730 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8734 #endif /* CONFIG_NUMA_BALANCING */
8740 * task_running_on_cpu - return 1 if @p is running on @cpu.
8743 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8745 /* Task has no contribution or is new */
8746 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8749 if (task_on_rq_queued(p))
8756 * idle_cpu_without - would a given CPU be idle without p ?
8757 * @cpu: the processor on which idleness is tested.
8758 * @p: task which should be ignored.
8760 * Return: 1 if the CPU would be idle. 0 otherwise.
8762 static int idle_cpu_without(int cpu, struct task_struct *p)
8764 struct rq *rq = cpu_rq(cpu);
8766 if (rq->curr != rq->idle && rq->curr != p)
8770 * rq->nr_running can't be used but an updated version without the
8771 * impact of p on cpu must be used instead. The updated nr_running
8772 * be computed and tested before calling idle_cpu_without().
8776 if (rq->ttwu_pending)
8784 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8785 * @sd: The sched_domain level to look for idlest group.
8786 * @group: sched_group whose statistics are to be updated.
8787 * @sgs: variable to hold the statistics for this group.
8788 * @p: The task for which we look for the idlest group/CPU.
8790 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8791 struct sched_group *group,
8792 struct sg_lb_stats *sgs,
8793 struct task_struct *p)
8797 memset(sgs, 0, sizeof(*sgs));
8799 for_each_cpu(i, sched_group_span(group)) {
8800 struct rq *rq = cpu_rq(i);
8803 sgs->group_load += cpu_load_without(rq, p);
8804 sgs->group_util += cpu_util_without(i, p);
8805 sgs->group_runnable += cpu_runnable_without(rq, p);
8806 local = task_running_on_cpu(i, p);
8807 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8809 nr_running = rq->nr_running - local;
8810 sgs->sum_nr_running += nr_running;
8813 * No need to call idle_cpu_without() if nr_running is not 0
8815 if (!nr_running && idle_cpu_without(i, p))
8820 /* Check if task fits in the group */
8821 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8822 !task_fits_capacity(p, group->sgc->max_capacity)) {
8823 sgs->group_misfit_task_load = 1;
8826 sgs->group_capacity = group->sgc->capacity;
8828 sgs->group_weight = group->group_weight;
8830 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8833 * Computing avg_load makes sense only when group is fully busy or
8836 if (sgs->group_type == group_fully_busy ||
8837 sgs->group_type == group_overloaded)
8838 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8839 sgs->group_capacity;
8842 static bool update_pick_idlest(struct sched_group *idlest,
8843 struct sg_lb_stats *idlest_sgs,
8844 struct sched_group *group,
8845 struct sg_lb_stats *sgs)
8847 if (sgs->group_type < idlest_sgs->group_type)
8850 if (sgs->group_type > idlest_sgs->group_type)
8854 * The candidate and the current idlest group are the same type of
8855 * group. Let check which one is the idlest according to the type.
8858 switch (sgs->group_type) {
8859 case group_overloaded:
8860 case group_fully_busy:
8861 /* Select the group with lowest avg_load. */
8862 if (idlest_sgs->avg_load <= sgs->avg_load)
8866 case group_imbalanced:
8867 case group_asym_packing:
8868 /* Those types are not used in the slow wakeup path */
8871 case group_misfit_task:
8872 /* Select group with the highest max capacity */
8873 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8877 case group_has_spare:
8878 /* Select group with most idle CPUs */
8879 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
8882 /* Select group with lowest group_util */
8883 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
8884 idlest_sgs->group_util <= sgs->group_util)
8894 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
8895 * This is an approximation as the number of running tasks may not be
8896 * related to the number of busy CPUs due to sched_setaffinity.
8898 static inline bool allow_numa_imbalance(int dst_running, int dst_weight)
8900 return (dst_running < (dst_weight >> 2));
8904 * find_idlest_group() finds and returns the least busy CPU group within the
8907 * Assumes p is allowed on at least one CPU in sd.
8909 static struct sched_group *
8910 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
8912 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8913 struct sg_lb_stats local_sgs, tmp_sgs;
8914 struct sg_lb_stats *sgs;
8915 unsigned long imbalance;
8916 struct sg_lb_stats idlest_sgs = {
8917 .avg_load = UINT_MAX,
8918 .group_type = group_overloaded,
8924 /* Skip over this group if it has no CPUs allowed */
8925 if (!cpumask_intersects(sched_group_span(group),
8929 /* Skip over this group if no cookie matched */
8930 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
8933 local_group = cpumask_test_cpu(this_cpu,
8934 sched_group_span(group));
8943 update_sg_wakeup_stats(sd, group, sgs, p);
8945 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8950 } while (group = group->next, group != sd->groups);
8953 /* There is no idlest group to push tasks to */
8957 /* The local group has been skipped because of CPU affinity */
8962 * If the local group is idler than the selected idlest group
8963 * don't try and push the task.
8965 if (local_sgs.group_type < idlest_sgs.group_type)
8969 * If the local group is busier than the selected idlest group
8970 * try and push the task.
8972 if (local_sgs.group_type > idlest_sgs.group_type)
8975 switch (local_sgs.group_type) {
8976 case group_overloaded:
8977 case group_fully_busy:
8979 /* Calculate allowed imbalance based on load */
8980 imbalance = scale_load_down(NICE_0_LOAD) *
8981 (sd->imbalance_pct-100) / 100;
8984 * When comparing groups across NUMA domains, it's possible for
8985 * the local domain to be very lightly loaded relative to the
8986 * remote domains but "imbalance" skews the comparison making
8987 * remote CPUs look much more favourable. When considering
8988 * cross-domain, add imbalance to the load on the remote node
8989 * and consider staying local.
8992 if ((sd->flags & SD_NUMA) &&
8993 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8997 * If the local group is less loaded than the selected
8998 * idlest group don't try and push any tasks.
9000 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9003 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9007 case group_imbalanced:
9008 case group_asym_packing:
9009 /* Those type are not used in the slow wakeup path */
9012 case group_misfit_task:
9013 /* Select group with the highest max capacity */
9014 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9018 case group_has_spare:
9019 if (sd->flags & SD_NUMA) {
9020 #ifdef CONFIG_NUMA_BALANCING
9023 * If there is spare capacity at NUMA, try to select
9024 * the preferred node
9026 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9029 idlest_cpu = cpumask_first(sched_group_span(idlest));
9030 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9034 * Otherwise, keep the task on this node to stay close
9035 * its wakeup source and improve locality. If there is
9036 * a real need of migration, periodic load balance will
9039 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight))
9044 * Select group with highest number of idle CPUs. We could also
9045 * compare the utilization which is more stable but it can end
9046 * up that the group has less spare capacity but finally more
9047 * idle CPUs which means more opportunity to run task.
9049 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9058 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9059 * @env: The load balancing environment.
9060 * @sds: variable to hold the statistics for this sched_domain.
9063 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9065 struct sched_domain *child = env->sd->child;
9066 struct sched_group *sg = env->sd->groups;
9067 struct sg_lb_stats *local = &sds->local_stat;
9068 struct sg_lb_stats tmp_sgs;
9072 struct sg_lb_stats *sgs = &tmp_sgs;
9075 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9080 if (env->idle != CPU_NEWLY_IDLE ||
9081 time_after_eq(jiffies, sg->sgc->next_update))
9082 update_group_capacity(env->sd, env->dst_cpu);
9085 update_sg_lb_stats(env, sg, sgs, &sg_status);
9091 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9093 sds->busiest_stat = *sgs;
9097 /* Now, start updating sd_lb_stats */
9098 sds->total_load += sgs->group_load;
9099 sds->total_capacity += sgs->group_capacity;
9102 } while (sg != env->sd->groups);
9104 /* Tag domain that child domain prefers tasks go to siblings first */
9105 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9108 if (env->sd->flags & SD_NUMA)
9109 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9111 if (!env->sd->parent) {
9112 struct root_domain *rd = env->dst_rq->rd;
9114 /* update overload indicator if we are at root domain */
9115 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9117 /* Update over-utilization (tipping point, U >= 0) indicator */
9118 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9119 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9120 } else if (sg_status & SG_OVERUTILIZED) {
9121 struct root_domain *rd = env->dst_rq->rd;
9123 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9124 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9128 #define NUMA_IMBALANCE_MIN 2
9130 static inline long adjust_numa_imbalance(int imbalance,
9131 int dst_running, int dst_weight)
9133 if (!allow_numa_imbalance(dst_running, dst_weight))
9137 * Allow a small imbalance based on a simple pair of communicating
9138 * tasks that remain local when the destination is lightly loaded.
9140 if (imbalance <= NUMA_IMBALANCE_MIN)
9147 * calculate_imbalance - Calculate the amount of imbalance present within the
9148 * groups of a given sched_domain during load balance.
9149 * @env: load balance environment
9150 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9152 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9154 struct sg_lb_stats *local, *busiest;
9156 local = &sds->local_stat;
9157 busiest = &sds->busiest_stat;
9159 if (busiest->group_type == group_misfit_task) {
9160 /* Set imbalance to allow misfit tasks to be balanced. */
9161 env->migration_type = migrate_misfit;
9166 if (busiest->group_type == group_asym_packing) {
9168 * In case of asym capacity, we will try to migrate all load to
9169 * the preferred CPU.
9171 env->migration_type = migrate_task;
9172 env->imbalance = busiest->sum_h_nr_running;
9176 if (busiest->group_type == group_imbalanced) {
9178 * In the group_imb case we cannot rely on group-wide averages
9179 * to ensure CPU-load equilibrium, try to move any task to fix
9180 * the imbalance. The next load balance will take care of
9181 * balancing back the system.
9183 env->migration_type = migrate_task;
9189 * Try to use spare capacity of local group without overloading it or
9192 if (local->group_type == group_has_spare) {
9193 if ((busiest->group_type > group_fully_busy) &&
9194 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9196 * If busiest is overloaded, try to fill spare
9197 * capacity. This might end up creating spare capacity
9198 * in busiest or busiest still being overloaded but
9199 * there is no simple way to directly compute the
9200 * amount of load to migrate in order to balance the
9203 env->migration_type = migrate_util;
9204 env->imbalance = max(local->group_capacity, local->group_util) -
9208 * In some cases, the group's utilization is max or even
9209 * higher than capacity because of migrations but the
9210 * local CPU is (newly) idle. There is at least one
9211 * waiting task in this overloaded busiest group. Let's
9214 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9215 env->migration_type = migrate_task;
9222 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9223 unsigned int nr_diff = busiest->sum_nr_running;
9225 * When prefer sibling, evenly spread running tasks on
9228 env->migration_type = migrate_task;
9229 lsub_positive(&nr_diff, local->sum_nr_running);
9230 env->imbalance = nr_diff >> 1;
9234 * If there is no overload, we just want to even the number of
9237 env->migration_type = migrate_task;
9238 env->imbalance = max_t(long, 0, (local->idle_cpus -
9239 busiest->idle_cpus) >> 1);
9242 /* Consider allowing a small imbalance between NUMA groups */
9243 if (env->sd->flags & SD_NUMA) {
9244 env->imbalance = adjust_numa_imbalance(env->imbalance,
9245 busiest->sum_nr_running, busiest->group_weight);
9252 * Local is fully busy but has to take more load to relieve the
9255 if (local->group_type < group_overloaded) {
9257 * Local will become overloaded so the avg_load metrics are
9261 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9262 local->group_capacity;
9264 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9265 sds->total_capacity;
9267 * If the local group is more loaded than the selected
9268 * busiest group don't try to pull any tasks.
9270 if (local->avg_load >= busiest->avg_load) {
9277 * Both group are or will become overloaded and we're trying to get all
9278 * the CPUs to the average_load, so we don't want to push ourselves
9279 * above the average load, nor do we wish to reduce the max loaded CPU
9280 * below the average load. At the same time, we also don't want to
9281 * reduce the group load below the group capacity. Thus we look for
9282 * the minimum possible imbalance.
9284 env->migration_type = migrate_load;
9285 env->imbalance = min(
9286 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9287 (sds->avg_load - local->avg_load) * local->group_capacity
9288 ) / SCHED_CAPACITY_SCALE;
9291 /******* find_busiest_group() helpers end here *********************/
9294 * Decision matrix according to the local and busiest group type:
9296 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9297 * has_spare nr_idle balanced N/A N/A balanced balanced
9298 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9299 * misfit_task force N/A N/A N/A force force
9300 * asym_packing force force N/A N/A force force
9301 * imbalanced force force N/A N/A force force
9302 * overloaded force force N/A N/A force avg_load
9304 * N/A : Not Applicable because already filtered while updating
9306 * balanced : The system is balanced for these 2 groups.
9307 * force : Calculate the imbalance as load migration is probably needed.
9308 * avg_load : Only if imbalance is significant enough.
9309 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9310 * different in groups.
9314 * find_busiest_group - Returns the busiest group within the sched_domain
9315 * if there is an imbalance.
9317 * Also calculates the amount of runnable load which should be moved
9318 * to restore balance.
9320 * @env: The load balancing environment.
9322 * Return: - The busiest group if imbalance exists.
9324 static struct sched_group *find_busiest_group(struct lb_env *env)
9326 struct sg_lb_stats *local, *busiest;
9327 struct sd_lb_stats sds;
9329 init_sd_lb_stats(&sds);
9332 * Compute the various statistics relevant for load balancing at
9335 update_sd_lb_stats(env, &sds);
9337 if (sched_energy_enabled()) {
9338 struct root_domain *rd = env->dst_rq->rd;
9340 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9344 local = &sds.local_stat;
9345 busiest = &sds.busiest_stat;
9347 /* There is no busy sibling group to pull tasks from */
9351 /* Misfit tasks should be dealt with regardless of the avg load */
9352 if (busiest->group_type == group_misfit_task)
9355 /* ASYM feature bypasses nice load balance check */
9356 if (busiest->group_type == group_asym_packing)
9360 * If the busiest group is imbalanced the below checks don't
9361 * work because they assume all things are equal, which typically
9362 * isn't true due to cpus_ptr constraints and the like.
9364 if (busiest->group_type == group_imbalanced)
9368 * If the local group is busier than the selected busiest group
9369 * don't try and pull any tasks.
9371 if (local->group_type > busiest->group_type)
9375 * When groups are overloaded, use the avg_load to ensure fairness
9378 if (local->group_type == group_overloaded) {
9380 * If the local group is more loaded than the selected
9381 * busiest group don't try to pull any tasks.
9383 if (local->avg_load >= busiest->avg_load)
9386 /* XXX broken for overlapping NUMA groups */
9387 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9391 * Don't pull any tasks if this group is already above the
9392 * domain average load.
9394 if (local->avg_load >= sds.avg_load)
9398 * If the busiest group is more loaded, use imbalance_pct to be
9401 if (100 * busiest->avg_load <=
9402 env->sd->imbalance_pct * local->avg_load)
9406 /* Try to move all excess tasks to child's sibling domain */
9407 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9408 busiest->sum_nr_running > local->sum_nr_running + 1)
9411 if (busiest->group_type != group_overloaded) {
9412 if (env->idle == CPU_NOT_IDLE)
9414 * If the busiest group is not overloaded (and as a
9415 * result the local one too) but this CPU is already
9416 * busy, let another idle CPU try to pull task.
9420 if (busiest->group_weight > 1 &&
9421 local->idle_cpus <= (busiest->idle_cpus + 1))
9423 * If the busiest group is not overloaded
9424 * and there is no imbalance between this and busiest
9425 * group wrt idle CPUs, it is balanced. The imbalance
9426 * becomes significant if the diff is greater than 1
9427 * otherwise we might end up to just move the imbalance
9428 * on another group. Of course this applies only if
9429 * there is more than 1 CPU per group.
9433 if (busiest->sum_h_nr_running == 1)
9435 * busiest doesn't have any tasks waiting to run
9441 /* Looks like there is an imbalance. Compute it */
9442 calculate_imbalance(env, &sds);
9443 return env->imbalance ? sds.busiest : NULL;
9451 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9453 static struct rq *find_busiest_queue(struct lb_env *env,
9454 struct sched_group *group)
9456 struct rq *busiest = NULL, *rq;
9457 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9458 unsigned int busiest_nr = 0;
9461 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9462 unsigned long capacity, load, util;
9463 unsigned int nr_running;
9467 rt = fbq_classify_rq(rq);
9470 * We classify groups/runqueues into three groups:
9471 * - regular: there are !numa tasks
9472 * - remote: there are numa tasks that run on the 'wrong' node
9473 * - all: there is no distinction
9475 * In order to avoid migrating ideally placed numa tasks,
9476 * ignore those when there's better options.
9478 * If we ignore the actual busiest queue to migrate another
9479 * task, the next balance pass can still reduce the busiest
9480 * queue by moving tasks around inside the node.
9482 * If we cannot move enough load due to this classification
9483 * the next pass will adjust the group classification and
9484 * allow migration of more tasks.
9486 * Both cases only affect the total convergence complexity.
9488 if (rt > env->fbq_type)
9491 nr_running = rq->cfs.h_nr_running;
9495 capacity = capacity_of(i);
9498 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9499 * eventually lead to active_balancing high->low capacity.
9500 * Higher per-CPU capacity is considered better than balancing
9503 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9504 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9508 switch (env->migration_type) {
9511 * When comparing with load imbalance, use cpu_load()
9512 * which is not scaled with the CPU capacity.
9514 load = cpu_load(rq);
9516 if (nr_running == 1 && load > env->imbalance &&
9517 !check_cpu_capacity(rq, env->sd))
9521 * For the load comparisons with the other CPUs,
9522 * consider the cpu_load() scaled with the CPU
9523 * capacity, so that the load can be moved away
9524 * from the CPU that is potentially running at a
9527 * Thus we're looking for max(load_i / capacity_i),
9528 * crosswise multiplication to rid ourselves of the
9529 * division works out to:
9530 * load_i * capacity_j > load_j * capacity_i;
9531 * where j is our previous maximum.
9533 if (load * busiest_capacity > busiest_load * capacity) {
9534 busiest_load = load;
9535 busiest_capacity = capacity;
9541 util = cpu_util(cpu_of(rq));
9544 * Don't try to pull utilization from a CPU with one
9545 * running task. Whatever its utilization, we will fail
9548 if (nr_running <= 1)
9551 if (busiest_util < util) {
9552 busiest_util = util;
9558 if (busiest_nr < nr_running) {
9559 busiest_nr = nr_running;
9564 case migrate_misfit:
9566 * For ASYM_CPUCAPACITY domains with misfit tasks we
9567 * simply seek the "biggest" misfit task.
9569 if (rq->misfit_task_load > busiest_load) {
9570 busiest_load = rq->misfit_task_load;
9583 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9584 * so long as it is large enough.
9586 #define MAX_PINNED_INTERVAL 512
9589 asym_active_balance(struct lb_env *env)
9592 * ASYM_PACKING needs to force migrate tasks from busy but
9593 * lower priority CPUs in order to pack all tasks in the
9594 * highest priority CPUs.
9596 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9597 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9601 imbalanced_active_balance(struct lb_env *env)
9603 struct sched_domain *sd = env->sd;
9606 * The imbalanced case includes the case of pinned tasks preventing a fair
9607 * distribution of the load on the system but also the even distribution of the
9608 * threads on a system with spare capacity
9610 if ((env->migration_type == migrate_task) &&
9611 (sd->nr_balance_failed > sd->cache_nice_tries+2))
9617 static int need_active_balance(struct lb_env *env)
9619 struct sched_domain *sd = env->sd;
9621 if (asym_active_balance(env))
9624 if (imbalanced_active_balance(env))
9628 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9629 * It's worth migrating the task if the src_cpu's capacity is reduced
9630 * because of other sched_class or IRQs if more capacity stays
9631 * available on dst_cpu.
9633 if ((env->idle != CPU_NOT_IDLE) &&
9634 (env->src_rq->cfs.h_nr_running == 1)) {
9635 if ((check_cpu_capacity(env->src_rq, sd)) &&
9636 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9640 if (env->migration_type == migrate_misfit)
9646 static int active_load_balance_cpu_stop(void *data);
9648 static int should_we_balance(struct lb_env *env)
9650 struct sched_group *sg = env->sd->groups;
9654 * Ensure the balancing environment is consistent; can happen
9655 * when the softirq triggers 'during' hotplug.
9657 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9661 * In the newly idle case, we will allow all the CPUs
9662 * to do the newly idle load balance.
9664 if (env->idle == CPU_NEWLY_IDLE)
9667 /* Try to find first idle CPU */
9668 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9672 /* Are we the first idle CPU? */
9673 return cpu == env->dst_cpu;
9676 /* Are we the first CPU of this group ? */
9677 return group_balance_cpu(sg) == env->dst_cpu;
9681 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9682 * tasks if there is an imbalance.
9684 static int load_balance(int this_cpu, struct rq *this_rq,
9685 struct sched_domain *sd, enum cpu_idle_type idle,
9686 int *continue_balancing)
9688 int ld_moved, cur_ld_moved, active_balance = 0;
9689 struct sched_domain *sd_parent = sd->parent;
9690 struct sched_group *group;
9693 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9695 struct lb_env env = {
9697 .dst_cpu = this_cpu,
9699 .dst_grpmask = sched_group_span(sd->groups),
9701 .loop_break = sched_nr_migrate_break,
9704 .tasks = LIST_HEAD_INIT(env.tasks),
9707 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9709 schedstat_inc(sd->lb_count[idle]);
9712 if (!should_we_balance(&env)) {
9713 *continue_balancing = 0;
9717 group = find_busiest_group(&env);
9719 schedstat_inc(sd->lb_nobusyg[idle]);
9723 busiest = find_busiest_queue(&env, group);
9725 schedstat_inc(sd->lb_nobusyq[idle]);
9729 BUG_ON(busiest == env.dst_rq);
9731 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9733 env.src_cpu = busiest->cpu;
9734 env.src_rq = busiest;
9737 /* Clear this flag as soon as we find a pullable task */
9738 env.flags |= LBF_ALL_PINNED;
9739 if (busiest->nr_running > 1) {
9741 * Attempt to move tasks. If find_busiest_group has found
9742 * an imbalance but busiest->nr_running <= 1, the group is
9743 * still unbalanced. ld_moved simply stays zero, so it is
9744 * correctly treated as an imbalance.
9746 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9749 rq_lock_irqsave(busiest, &rf);
9750 update_rq_clock(busiest);
9753 * cur_ld_moved - load moved in current iteration
9754 * ld_moved - cumulative load moved across iterations
9756 cur_ld_moved = detach_tasks(&env);
9759 * We've detached some tasks from busiest_rq. Every
9760 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9761 * unlock busiest->lock, and we are able to be sure
9762 * that nobody can manipulate the tasks in parallel.
9763 * See task_rq_lock() family for the details.
9766 rq_unlock(busiest, &rf);
9770 ld_moved += cur_ld_moved;
9773 local_irq_restore(rf.flags);
9775 if (env.flags & LBF_NEED_BREAK) {
9776 env.flags &= ~LBF_NEED_BREAK;
9781 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9782 * us and move them to an alternate dst_cpu in our sched_group
9783 * where they can run. The upper limit on how many times we
9784 * iterate on same src_cpu is dependent on number of CPUs in our
9787 * This changes load balance semantics a bit on who can move
9788 * load to a given_cpu. In addition to the given_cpu itself
9789 * (or a ilb_cpu acting on its behalf where given_cpu is
9790 * nohz-idle), we now have balance_cpu in a position to move
9791 * load to given_cpu. In rare situations, this may cause
9792 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9793 * _independently_ and at _same_ time to move some load to
9794 * given_cpu) causing excess load to be moved to given_cpu.
9795 * This however should not happen so much in practice and
9796 * moreover subsequent load balance cycles should correct the
9797 * excess load moved.
9799 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9801 /* Prevent to re-select dst_cpu via env's CPUs */
9802 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9804 env.dst_rq = cpu_rq(env.new_dst_cpu);
9805 env.dst_cpu = env.new_dst_cpu;
9806 env.flags &= ~LBF_DST_PINNED;
9808 env.loop_break = sched_nr_migrate_break;
9811 * Go back to "more_balance" rather than "redo" since we
9812 * need to continue with same src_cpu.
9818 * We failed to reach balance because of affinity.
9821 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9823 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9824 *group_imbalance = 1;
9827 /* All tasks on this runqueue were pinned by CPU affinity */
9828 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9829 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9831 * Attempting to continue load balancing at the current
9832 * sched_domain level only makes sense if there are
9833 * active CPUs remaining as possible busiest CPUs to
9834 * pull load from which are not contained within the
9835 * destination group that is receiving any migrated
9838 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9840 env.loop_break = sched_nr_migrate_break;
9843 goto out_all_pinned;
9848 schedstat_inc(sd->lb_failed[idle]);
9850 * Increment the failure counter only on periodic balance.
9851 * We do not want newidle balance, which can be very
9852 * frequent, pollute the failure counter causing
9853 * excessive cache_hot migrations and active balances.
9855 if (idle != CPU_NEWLY_IDLE)
9856 sd->nr_balance_failed++;
9858 if (need_active_balance(&env)) {
9859 unsigned long flags;
9861 raw_spin_rq_lock_irqsave(busiest, flags);
9864 * Don't kick the active_load_balance_cpu_stop,
9865 * if the curr task on busiest CPU can't be
9866 * moved to this_cpu:
9868 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9869 raw_spin_rq_unlock_irqrestore(busiest, flags);
9870 goto out_one_pinned;
9873 /* Record that we found at least one task that could run on this_cpu */
9874 env.flags &= ~LBF_ALL_PINNED;
9877 * ->active_balance synchronizes accesses to
9878 * ->active_balance_work. Once set, it's cleared
9879 * only after active load balance is finished.
9881 if (!busiest->active_balance) {
9882 busiest->active_balance = 1;
9883 busiest->push_cpu = this_cpu;
9886 raw_spin_rq_unlock_irqrestore(busiest, flags);
9888 if (active_balance) {
9889 stop_one_cpu_nowait(cpu_of(busiest),
9890 active_load_balance_cpu_stop, busiest,
9891 &busiest->active_balance_work);
9895 sd->nr_balance_failed = 0;
9898 if (likely(!active_balance) || need_active_balance(&env)) {
9899 /* We were unbalanced, so reset the balancing interval */
9900 sd->balance_interval = sd->min_interval;
9907 * We reach balance although we may have faced some affinity
9908 * constraints. Clear the imbalance flag only if other tasks got
9909 * a chance to move and fix the imbalance.
9911 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9912 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9914 if (*group_imbalance)
9915 *group_imbalance = 0;
9920 * We reach balance because all tasks are pinned at this level so
9921 * we can't migrate them. Let the imbalance flag set so parent level
9922 * can try to migrate them.
9924 schedstat_inc(sd->lb_balanced[idle]);
9926 sd->nr_balance_failed = 0;
9932 * newidle_balance() disregards balance intervals, so we could
9933 * repeatedly reach this code, which would lead to balance_interval
9934 * skyrocketing in a short amount of time. Skip the balance_interval
9935 * increase logic to avoid that.
9937 if (env.idle == CPU_NEWLY_IDLE)
9940 /* tune up the balancing interval */
9941 if ((env.flags & LBF_ALL_PINNED &&
9942 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9943 sd->balance_interval < sd->max_interval)
9944 sd->balance_interval *= 2;
9949 static inline unsigned long
9950 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9952 unsigned long interval = sd->balance_interval;
9955 interval *= sd->busy_factor;
9957 /* scale ms to jiffies */
9958 interval = msecs_to_jiffies(interval);
9961 * Reduce likelihood of busy balancing at higher domains racing with
9962 * balancing at lower domains by preventing their balancing periods
9963 * from being multiples of each other.
9968 interval = clamp(interval, 1UL, max_load_balance_interval);
9974 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9976 unsigned long interval, next;
9978 /* used by idle balance, so cpu_busy = 0 */
9979 interval = get_sd_balance_interval(sd, 0);
9980 next = sd->last_balance + interval;
9982 if (time_after(*next_balance, next))
9983 *next_balance = next;
9987 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9988 * running tasks off the busiest CPU onto idle CPUs. It requires at
9989 * least 1 task to be running on each physical CPU where possible, and
9990 * avoids physical / logical imbalances.
9992 static int active_load_balance_cpu_stop(void *data)
9994 struct rq *busiest_rq = data;
9995 int busiest_cpu = cpu_of(busiest_rq);
9996 int target_cpu = busiest_rq->push_cpu;
9997 struct rq *target_rq = cpu_rq(target_cpu);
9998 struct sched_domain *sd;
9999 struct task_struct *p = NULL;
10000 struct rq_flags rf;
10002 rq_lock_irq(busiest_rq, &rf);
10004 * Between queueing the stop-work and running it is a hole in which
10005 * CPUs can become inactive. We should not move tasks from or to
10008 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10011 /* Make sure the requested CPU hasn't gone down in the meantime: */
10012 if (unlikely(busiest_cpu != smp_processor_id() ||
10013 !busiest_rq->active_balance))
10016 /* Is there any task to move? */
10017 if (busiest_rq->nr_running <= 1)
10021 * This condition is "impossible", if it occurs
10022 * we need to fix it. Originally reported by
10023 * Bjorn Helgaas on a 128-CPU setup.
10025 BUG_ON(busiest_rq == target_rq);
10027 /* Search for an sd spanning us and the target CPU. */
10029 for_each_domain(target_cpu, sd) {
10030 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10035 struct lb_env env = {
10037 .dst_cpu = target_cpu,
10038 .dst_rq = target_rq,
10039 .src_cpu = busiest_rq->cpu,
10040 .src_rq = busiest_rq,
10042 .flags = LBF_ACTIVE_LB,
10045 schedstat_inc(sd->alb_count);
10046 update_rq_clock(busiest_rq);
10048 p = detach_one_task(&env);
10050 schedstat_inc(sd->alb_pushed);
10051 /* Active balancing done, reset the failure counter. */
10052 sd->nr_balance_failed = 0;
10054 schedstat_inc(sd->alb_failed);
10059 busiest_rq->active_balance = 0;
10060 rq_unlock(busiest_rq, &rf);
10063 attach_one_task(target_rq, p);
10065 local_irq_enable();
10070 static DEFINE_SPINLOCK(balancing);
10073 * Scale the max load_balance interval with the number of CPUs in the system.
10074 * This trades load-balance latency on larger machines for less cross talk.
10076 void update_max_interval(void)
10078 max_load_balance_interval = HZ*num_online_cpus()/10;
10082 * It checks each scheduling domain to see if it is due to be balanced,
10083 * and initiates a balancing operation if so.
10085 * Balancing parameters are set up in init_sched_domains.
10087 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10089 int continue_balancing = 1;
10091 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10092 unsigned long interval;
10093 struct sched_domain *sd;
10094 /* Earliest time when we have to do rebalance again */
10095 unsigned long next_balance = jiffies + 60*HZ;
10096 int update_next_balance = 0;
10097 int need_serialize, need_decay = 0;
10101 for_each_domain(cpu, sd) {
10103 * Decay the newidle max times here because this is a regular
10104 * visit to all the domains. Decay ~1% per second.
10106 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
10107 sd->max_newidle_lb_cost =
10108 (sd->max_newidle_lb_cost * 253) / 256;
10109 sd->next_decay_max_lb_cost = jiffies + HZ;
10112 max_cost += sd->max_newidle_lb_cost;
10115 * Stop the load balance at this level. There is another
10116 * CPU in our sched group which is doing load balancing more
10119 if (!continue_balancing) {
10125 interval = get_sd_balance_interval(sd, busy);
10127 need_serialize = sd->flags & SD_SERIALIZE;
10128 if (need_serialize) {
10129 if (!spin_trylock(&balancing))
10133 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10134 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10136 * The LBF_DST_PINNED logic could have changed
10137 * env->dst_cpu, so we can't know our idle
10138 * state even if we migrated tasks. Update it.
10140 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10141 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10143 sd->last_balance = jiffies;
10144 interval = get_sd_balance_interval(sd, busy);
10146 if (need_serialize)
10147 spin_unlock(&balancing);
10149 if (time_after(next_balance, sd->last_balance + interval)) {
10150 next_balance = sd->last_balance + interval;
10151 update_next_balance = 1;
10156 * Ensure the rq-wide value also decays but keep it at a
10157 * reasonable floor to avoid funnies with rq->avg_idle.
10159 rq->max_idle_balance_cost =
10160 max((u64)sysctl_sched_migration_cost, max_cost);
10165 * next_balance will be updated only when there is a need.
10166 * When the cpu is attached to null domain for ex, it will not be
10169 if (likely(update_next_balance))
10170 rq->next_balance = next_balance;
10174 static inline int on_null_domain(struct rq *rq)
10176 return unlikely(!rcu_dereference_sched(rq->sd));
10179 #ifdef CONFIG_NO_HZ_COMMON
10181 * idle load balancing details
10182 * - When one of the busy CPUs notice that there may be an idle rebalancing
10183 * needed, they will kick the idle load balancer, which then does idle
10184 * load balancing for all the idle CPUs.
10185 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10189 static inline int find_new_ilb(void)
10193 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
10194 housekeeping_cpumask(HK_FLAG_MISC)) {
10196 if (ilb == smp_processor_id())
10207 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10208 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10210 static void kick_ilb(unsigned int flags)
10215 * Increase nohz.next_balance only when if full ilb is triggered but
10216 * not if we only update stats.
10218 if (flags & NOHZ_BALANCE_KICK)
10219 nohz.next_balance = jiffies+1;
10221 ilb_cpu = find_new_ilb();
10223 if (ilb_cpu >= nr_cpu_ids)
10227 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10228 * the first flag owns it; cleared by nohz_csd_func().
10230 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10231 if (flags & NOHZ_KICK_MASK)
10235 * This way we generate an IPI on the target CPU which
10236 * is idle. And the softirq performing nohz idle load balance
10237 * will be run before returning from the IPI.
10239 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10243 * Current decision point for kicking the idle load balancer in the presence
10244 * of idle CPUs in the system.
10246 static void nohz_balancer_kick(struct rq *rq)
10248 unsigned long now = jiffies;
10249 struct sched_domain_shared *sds;
10250 struct sched_domain *sd;
10251 int nr_busy, i, cpu = rq->cpu;
10252 unsigned int flags = 0;
10254 if (unlikely(rq->idle_balance))
10258 * We may be recently in ticked or tickless idle mode. At the first
10259 * busy tick after returning from idle, we will update the busy stats.
10261 nohz_balance_exit_idle(rq);
10264 * None are in tickless mode and hence no need for NOHZ idle load
10267 if (likely(!atomic_read(&nohz.nr_cpus)))
10270 if (READ_ONCE(nohz.has_blocked) &&
10271 time_after(now, READ_ONCE(nohz.next_blocked)))
10272 flags = NOHZ_STATS_KICK;
10274 if (time_before(now, nohz.next_balance))
10277 if (rq->nr_running >= 2) {
10278 flags = NOHZ_KICK_MASK;
10284 sd = rcu_dereference(rq->sd);
10287 * If there's a CFS task and the current CPU has reduced
10288 * capacity; kick the ILB to see if there's a better CPU to run
10291 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10292 flags = NOHZ_KICK_MASK;
10297 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10300 * When ASYM_PACKING; see if there's a more preferred CPU
10301 * currently idle; in which case, kick the ILB to move tasks
10304 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10305 if (sched_asym_prefer(i, cpu)) {
10306 flags = NOHZ_KICK_MASK;
10312 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10315 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10316 * to run the misfit task on.
10318 if (check_misfit_status(rq, sd)) {
10319 flags = NOHZ_KICK_MASK;
10324 * For asymmetric systems, we do not want to nicely balance
10325 * cache use, instead we want to embrace asymmetry and only
10326 * ensure tasks have enough CPU capacity.
10328 * Skip the LLC logic because it's not relevant in that case.
10333 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10336 * If there is an imbalance between LLC domains (IOW we could
10337 * increase the overall cache use), we need some less-loaded LLC
10338 * domain to pull some load. Likewise, we may need to spread
10339 * load within the current LLC domain (e.g. packed SMT cores but
10340 * other CPUs are idle). We can't really know from here how busy
10341 * the others are - so just get a nohz balance going if it looks
10342 * like this LLC domain has tasks we could move.
10344 nr_busy = atomic_read(&sds->nr_busy_cpus);
10346 flags = NOHZ_KICK_MASK;
10357 static void set_cpu_sd_state_busy(int cpu)
10359 struct sched_domain *sd;
10362 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10364 if (!sd || !sd->nohz_idle)
10368 atomic_inc(&sd->shared->nr_busy_cpus);
10373 void nohz_balance_exit_idle(struct rq *rq)
10375 SCHED_WARN_ON(rq != this_rq());
10377 if (likely(!rq->nohz_tick_stopped))
10380 rq->nohz_tick_stopped = 0;
10381 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10382 atomic_dec(&nohz.nr_cpus);
10384 set_cpu_sd_state_busy(rq->cpu);
10387 static void set_cpu_sd_state_idle(int cpu)
10389 struct sched_domain *sd;
10392 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10394 if (!sd || sd->nohz_idle)
10398 atomic_dec(&sd->shared->nr_busy_cpus);
10404 * This routine will record that the CPU is going idle with tick stopped.
10405 * This info will be used in performing idle load balancing in the future.
10407 void nohz_balance_enter_idle(int cpu)
10409 struct rq *rq = cpu_rq(cpu);
10411 SCHED_WARN_ON(cpu != smp_processor_id());
10413 /* If this CPU is going down, then nothing needs to be done: */
10414 if (!cpu_active(cpu))
10417 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10418 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10422 * Can be set safely without rq->lock held
10423 * If a clear happens, it will have evaluated last additions because
10424 * rq->lock is held during the check and the clear
10426 rq->has_blocked_load = 1;
10429 * The tick is still stopped but load could have been added in the
10430 * meantime. We set the nohz.has_blocked flag to trig a check of the
10431 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10432 * of nohz.has_blocked can only happen after checking the new load
10434 if (rq->nohz_tick_stopped)
10437 /* If we're a completely isolated CPU, we don't play: */
10438 if (on_null_domain(rq))
10441 rq->nohz_tick_stopped = 1;
10443 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10444 atomic_inc(&nohz.nr_cpus);
10447 * Ensures that if nohz_idle_balance() fails to observe our
10448 * @idle_cpus_mask store, it must observe the @has_blocked
10451 smp_mb__after_atomic();
10453 set_cpu_sd_state_idle(cpu);
10457 * Each time a cpu enter idle, we assume that it has blocked load and
10458 * enable the periodic update of the load of idle cpus
10460 WRITE_ONCE(nohz.has_blocked, 1);
10463 static bool update_nohz_stats(struct rq *rq)
10465 unsigned int cpu = rq->cpu;
10467 if (!rq->has_blocked_load)
10470 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10473 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10476 update_blocked_averages(cpu);
10478 return rq->has_blocked_load;
10482 * Internal function that runs load balance for all idle cpus. The load balance
10483 * can be a simple update of blocked load or a complete load balance with
10484 * tasks movement depending of flags.
10486 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10487 enum cpu_idle_type idle)
10489 /* Earliest time when we have to do rebalance again */
10490 unsigned long now = jiffies;
10491 unsigned long next_balance = now + 60*HZ;
10492 bool has_blocked_load = false;
10493 int update_next_balance = 0;
10494 int this_cpu = this_rq->cpu;
10498 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10501 * We assume there will be no idle load after this update and clear
10502 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10503 * set the has_blocked flag and trig another update of idle load.
10504 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10505 * setting the flag, we are sure to not clear the state and not
10506 * check the load of an idle cpu.
10508 WRITE_ONCE(nohz.has_blocked, 0);
10511 * Ensures that if we miss the CPU, we must see the has_blocked
10512 * store from nohz_balance_enter_idle().
10517 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10518 * chance for other idle cpu to pull load.
10520 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10521 if (!idle_cpu(balance_cpu))
10525 * If this CPU gets work to do, stop the load balancing
10526 * work being done for other CPUs. Next load
10527 * balancing owner will pick it up.
10529 if (need_resched()) {
10530 has_blocked_load = true;
10534 rq = cpu_rq(balance_cpu);
10536 has_blocked_load |= update_nohz_stats(rq);
10539 * If time for next balance is due,
10542 if (time_after_eq(jiffies, rq->next_balance)) {
10543 struct rq_flags rf;
10545 rq_lock_irqsave(rq, &rf);
10546 update_rq_clock(rq);
10547 rq_unlock_irqrestore(rq, &rf);
10549 if (flags & NOHZ_BALANCE_KICK)
10550 rebalance_domains(rq, CPU_IDLE);
10553 if (time_after(next_balance, rq->next_balance)) {
10554 next_balance = rq->next_balance;
10555 update_next_balance = 1;
10560 * next_balance will be updated only when there is a need.
10561 * When the CPU is attached to null domain for ex, it will not be
10564 if (likely(update_next_balance))
10565 nohz.next_balance = next_balance;
10567 WRITE_ONCE(nohz.next_blocked,
10568 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10571 /* There is still blocked load, enable periodic update */
10572 if (has_blocked_load)
10573 WRITE_ONCE(nohz.has_blocked, 1);
10577 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10578 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10580 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10582 unsigned int flags = this_rq->nohz_idle_balance;
10587 this_rq->nohz_idle_balance = 0;
10589 if (idle != CPU_IDLE)
10592 _nohz_idle_balance(this_rq, flags, idle);
10598 * Check if we need to run the ILB for updating blocked load before entering
10601 void nohz_run_idle_balance(int cpu)
10603 unsigned int flags;
10605 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
10608 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10609 * (ie NOHZ_STATS_KICK set) and will do the same.
10611 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
10612 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
10615 static void nohz_newidle_balance(struct rq *this_rq)
10617 int this_cpu = this_rq->cpu;
10620 * This CPU doesn't want to be disturbed by scheduler
10623 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10626 /* Will wake up very soon. No time for doing anything else*/
10627 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10630 /* Don't need to update blocked load of idle CPUs*/
10631 if (!READ_ONCE(nohz.has_blocked) ||
10632 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10636 * Set the need to trigger ILB in order to update blocked load
10637 * before entering idle state.
10639 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
10642 #else /* !CONFIG_NO_HZ_COMMON */
10643 static inline void nohz_balancer_kick(struct rq *rq) { }
10645 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10650 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10651 #endif /* CONFIG_NO_HZ_COMMON */
10654 * newidle_balance is called by schedule() if this_cpu is about to become
10655 * idle. Attempts to pull tasks from other CPUs.
10658 * < 0 - we released the lock and there are !fair tasks present
10659 * 0 - failed, no new tasks
10660 * > 0 - success, new (fair) tasks present
10662 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10664 unsigned long next_balance = jiffies + HZ;
10665 int this_cpu = this_rq->cpu;
10666 struct sched_domain *sd;
10667 int pulled_task = 0;
10670 update_misfit_status(NULL, this_rq);
10673 * There is a task waiting to run. No need to search for one.
10674 * Return 0; the task will be enqueued when switching to idle.
10676 if (this_rq->ttwu_pending)
10680 * We must set idle_stamp _before_ calling idle_balance(), such that we
10681 * measure the duration of idle_balance() as idle time.
10683 this_rq->idle_stamp = rq_clock(this_rq);
10686 * Do not pull tasks towards !active CPUs...
10688 if (!cpu_active(this_cpu))
10692 * This is OK, because current is on_cpu, which avoids it being picked
10693 * for load-balance and preemption/IRQs are still disabled avoiding
10694 * further scheduler activity on it and we're being very careful to
10695 * re-start the picking loop.
10697 rq_unpin_lock(this_rq, rf);
10699 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10700 !READ_ONCE(this_rq->rd->overload)) {
10703 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10705 update_next_balance(sd, &next_balance);
10711 raw_spin_rq_unlock(this_rq);
10713 update_blocked_averages(this_cpu);
10715 for_each_domain(this_cpu, sd) {
10716 int continue_balancing = 1;
10717 u64 t0, domain_cost;
10719 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10720 update_next_balance(sd, &next_balance);
10724 if (sd->flags & SD_BALANCE_NEWIDLE) {
10725 t0 = sched_clock_cpu(this_cpu);
10727 pulled_task = load_balance(this_cpu, this_rq,
10728 sd, CPU_NEWLY_IDLE,
10729 &continue_balancing);
10731 domain_cost = sched_clock_cpu(this_cpu) - t0;
10732 if (domain_cost > sd->max_newidle_lb_cost)
10733 sd->max_newidle_lb_cost = domain_cost;
10735 curr_cost += domain_cost;
10738 update_next_balance(sd, &next_balance);
10741 * Stop searching for tasks to pull if there are
10742 * now runnable tasks on this rq.
10744 if (pulled_task || this_rq->nr_running > 0 ||
10745 this_rq->ttwu_pending)
10750 raw_spin_rq_lock(this_rq);
10752 if (curr_cost > this_rq->max_idle_balance_cost)
10753 this_rq->max_idle_balance_cost = curr_cost;
10756 * While browsing the domains, we released the rq lock, a task could
10757 * have been enqueued in the meantime. Since we're not going idle,
10758 * pretend we pulled a task.
10760 if (this_rq->cfs.h_nr_running && !pulled_task)
10763 /* Is there a task of a high priority class? */
10764 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10768 /* Move the next balance forward */
10769 if (time_after(this_rq->next_balance, next_balance))
10770 this_rq->next_balance = next_balance;
10773 this_rq->idle_stamp = 0;
10775 nohz_newidle_balance(this_rq);
10777 rq_repin_lock(this_rq, rf);
10779 return pulled_task;
10783 * run_rebalance_domains is triggered when needed from the scheduler tick.
10784 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10786 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10788 struct rq *this_rq = this_rq();
10789 enum cpu_idle_type idle = this_rq->idle_balance ?
10790 CPU_IDLE : CPU_NOT_IDLE;
10793 * If this CPU has a pending nohz_balance_kick, then do the
10794 * balancing on behalf of the other idle CPUs whose ticks are
10795 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10796 * give the idle CPUs a chance to load balance. Else we may
10797 * load balance only within the local sched_domain hierarchy
10798 * and abort nohz_idle_balance altogether if we pull some load.
10800 if (nohz_idle_balance(this_rq, idle))
10803 /* normal load balance */
10804 update_blocked_averages(this_rq->cpu);
10805 rebalance_domains(this_rq, idle);
10809 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10811 void trigger_load_balance(struct rq *rq)
10814 * Don't need to rebalance while attached to NULL domain or
10815 * runqueue CPU is not active
10817 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
10820 if (time_after_eq(jiffies, rq->next_balance))
10821 raise_softirq(SCHED_SOFTIRQ);
10823 nohz_balancer_kick(rq);
10826 static void rq_online_fair(struct rq *rq)
10830 update_runtime_enabled(rq);
10833 static void rq_offline_fair(struct rq *rq)
10837 /* Ensure any throttled groups are reachable by pick_next_task */
10838 unthrottle_offline_cfs_rqs(rq);
10841 #endif /* CONFIG_SMP */
10843 #ifdef CONFIG_SCHED_CORE
10845 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
10847 u64 slice = sched_slice(cfs_rq_of(se), se);
10848 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
10850 return (rtime * min_nr_tasks > slice);
10853 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
10854 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
10856 if (!sched_core_enabled(rq))
10860 * If runqueue has only one task which used up its slice and
10861 * if the sibling is forced idle, then trigger schedule to
10862 * give forced idle task a chance.
10864 * sched_slice() considers only this active rq and it gets the
10865 * whole slice. But during force idle, we have siblings acting
10866 * like a single runqueue and hence we need to consider runnable
10867 * tasks on this CPU and the forced idle CPU. Ideally, we should
10868 * go through the forced idle rq, but that would be a perf hit.
10869 * We can assume that the forced idle CPU has at least
10870 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
10871 * if we need to give up the CPU.
10873 if (rq->core->core_forceidle && rq->cfs.nr_running == 1 &&
10874 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
10879 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
10881 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
10883 for_each_sched_entity(se) {
10884 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10887 if (cfs_rq->forceidle_seq == fi_seq)
10889 cfs_rq->forceidle_seq = fi_seq;
10892 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
10896 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
10898 struct sched_entity *se = &p->se;
10900 if (p->sched_class != &fair_sched_class)
10903 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
10906 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
10908 struct rq *rq = task_rq(a);
10909 struct sched_entity *sea = &a->se;
10910 struct sched_entity *seb = &b->se;
10911 struct cfs_rq *cfs_rqa;
10912 struct cfs_rq *cfs_rqb;
10915 SCHED_WARN_ON(task_rq(b)->core != rq->core);
10917 #ifdef CONFIG_FAIR_GROUP_SCHED
10919 * Find an se in the hierarchy for tasks a and b, such that the se's
10920 * are immediate siblings.
10922 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
10923 int sea_depth = sea->depth;
10924 int seb_depth = seb->depth;
10926 if (sea_depth >= seb_depth)
10927 sea = parent_entity(sea);
10928 if (sea_depth <= seb_depth)
10929 seb = parent_entity(seb);
10932 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
10933 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
10935 cfs_rqa = sea->cfs_rq;
10936 cfs_rqb = seb->cfs_rq;
10938 cfs_rqa = &task_rq(a)->cfs;
10939 cfs_rqb = &task_rq(b)->cfs;
10943 * Find delta after normalizing se's vruntime with its cfs_rq's
10944 * min_vruntime_fi, which would have been updated in prior calls
10945 * to se_fi_update().
10947 delta = (s64)(sea->vruntime - seb->vruntime) +
10948 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
10953 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
10957 * scheduler tick hitting a task of our scheduling class.
10959 * NOTE: This function can be called remotely by the tick offload that
10960 * goes along full dynticks. Therefore no local assumption can be made
10961 * and everything must be accessed through the @rq and @curr passed in
10964 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10966 struct cfs_rq *cfs_rq;
10967 struct sched_entity *se = &curr->se;
10969 for_each_sched_entity(se) {
10970 cfs_rq = cfs_rq_of(se);
10971 entity_tick(cfs_rq, se, queued);
10974 if (static_branch_unlikely(&sched_numa_balancing))
10975 task_tick_numa(rq, curr);
10977 update_misfit_status(curr, rq);
10978 update_overutilized_status(task_rq(curr));
10980 task_tick_core(rq, curr);
10984 * called on fork with the child task as argument from the parent's context
10985 * - child not yet on the tasklist
10986 * - preemption disabled
10988 static void task_fork_fair(struct task_struct *p)
10990 struct cfs_rq *cfs_rq;
10991 struct sched_entity *se = &p->se, *curr;
10992 struct rq *rq = this_rq();
10993 struct rq_flags rf;
10996 update_rq_clock(rq);
10998 cfs_rq = task_cfs_rq(current);
10999 curr = cfs_rq->curr;
11001 update_curr(cfs_rq);
11002 se->vruntime = curr->vruntime;
11004 place_entity(cfs_rq, se, 1);
11006 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11008 * Upon rescheduling, sched_class::put_prev_task() will place
11009 * 'current' within the tree based on its new key value.
11011 swap(curr->vruntime, se->vruntime);
11015 se->vruntime -= cfs_rq->min_vruntime;
11016 rq_unlock(rq, &rf);
11020 * Priority of the task has changed. Check to see if we preempt
11021 * the current task.
11024 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11026 if (!task_on_rq_queued(p))
11029 if (rq->cfs.nr_running == 1)
11033 * Reschedule if we are currently running on this runqueue and
11034 * our priority decreased, or if we are not currently running on
11035 * this runqueue and our priority is higher than the current's
11037 if (task_current(rq, p)) {
11038 if (p->prio > oldprio)
11041 check_preempt_curr(rq, p, 0);
11044 static inline bool vruntime_normalized(struct task_struct *p)
11046 struct sched_entity *se = &p->se;
11049 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11050 * the dequeue_entity(.flags=0) will already have normalized the
11057 * When !on_rq, vruntime of the task has usually NOT been normalized.
11058 * But there are some cases where it has already been normalized:
11060 * - A forked child which is waiting for being woken up by
11061 * wake_up_new_task().
11062 * - A task which has been woken up by try_to_wake_up() and
11063 * waiting for actually being woken up by sched_ttwu_pending().
11065 if (!se->sum_exec_runtime ||
11066 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11072 #ifdef CONFIG_FAIR_GROUP_SCHED
11074 * Propagate the changes of the sched_entity across the tg tree to make it
11075 * visible to the root
11077 static void propagate_entity_cfs_rq(struct sched_entity *se)
11079 struct cfs_rq *cfs_rq;
11081 list_add_leaf_cfs_rq(cfs_rq_of(se));
11083 /* Start to propagate at parent */
11086 for_each_sched_entity(se) {
11087 cfs_rq = cfs_rq_of(se);
11089 if (!cfs_rq_throttled(cfs_rq)){
11090 update_load_avg(cfs_rq, se, UPDATE_TG);
11091 list_add_leaf_cfs_rq(cfs_rq);
11095 if (list_add_leaf_cfs_rq(cfs_rq))
11100 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11103 static void detach_entity_cfs_rq(struct sched_entity *se)
11105 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11107 /* Catch up with the cfs_rq and remove our load when we leave */
11108 update_load_avg(cfs_rq, se, 0);
11109 detach_entity_load_avg(cfs_rq, se);
11110 update_tg_load_avg(cfs_rq);
11111 propagate_entity_cfs_rq(se);
11114 static void attach_entity_cfs_rq(struct sched_entity *se)
11116 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11118 #ifdef CONFIG_FAIR_GROUP_SCHED
11120 * Since the real-depth could have been changed (only FAIR
11121 * class maintain depth value), reset depth properly.
11123 se->depth = se->parent ? se->parent->depth + 1 : 0;
11126 /* Synchronize entity with its cfs_rq */
11127 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11128 attach_entity_load_avg(cfs_rq, se);
11129 update_tg_load_avg(cfs_rq);
11130 propagate_entity_cfs_rq(se);
11133 static void detach_task_cfs_rq(struct task_struct *p)
11135 struct sched_entity *se = &p->se;
11136 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11138 if (!vruntime_normalized(p)) {
11140 * Fix up our vruntime so that the current sleep doesn't
11141 * cause 'unlimited' sleep bonus.
11143 place_entity(cfs_rq, se, 0);
11144 se->vruntime -= cfs_rq->min_vruntime;
11147 detach_entity_cfs_rq(se);
11150 static void attach_task_cfs_rq(struct task_struct *p)
11152 struct sched_entity *se = &p->se;
11153 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11155 attach_entity_cfs_rq(se);
11157 if (!vruntime_normalized(p))
11158 se->vruntime += cfs_rq->min_vruntime;
11161 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11163 detach_task_cfs_rq(p);
11166 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11168 attach_task_cfs_rq(p);
11170 if (task_on_rq_queued(p)) {
11172 * We were most likely switched from sched_rt, so
11173 * kick off the schedule if running, otherwise just see
11174 * if we can still preempt the current task.
11176 if (task_current(rq, p))
11179 check_preempt_curr(rq, p, 0);
11183 /* Account for a task changing its policy or group.
11185 * This routine is mostly called to set cfs_rq->curr field when a task
11186 * migrates between groups/classes.
11188 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11190 struct sched_entity *se = &p->se;
11193 if (task_on_rq_queued(p)) {
11195 * Move the next running task to the front of the list, so our
11196 * cfs_tasks list becomes MRU one.
11198 list_move(&se->group_node, &rq->cfs_tasks);
11202 for_each_sched_entity(se) {
11203 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11205 set_next_entity(cfs_rq, se);
11206 /* ensure bandwidth has been allocated on our new cfs_rq */
11207 account_cfs_rq_runtime(cfs_rq, 0);
11211 void init_cfs_rq(struct cfs_rq *cfs_rq)
11213 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11214 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
11215 #ifndef CONFIG_64BIT
11216 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
11219 raw_spin_lock_init(&cfs_rq->removed.lock);
11223 #ifdef CONFIG_FAIR_GROUP_SCHED
11224 static void task_set_group_fair(struct task_struct *p)
11226 struct sched_entity *se = &p->se;
11228 set_task_rq(p, task_cpu(p));
11229 se->depth = se->parent ? se->parent->depth + 1 : 0;
11232 static void task_move_group_fair(struct task_struct *p)
11234 detach_task_cfs_rq(p);
11235 set_task_rq(p, task_cpu(p));
11238 /* Tell se's cfs_rq has been changed -- migrated */
11239 p->se.avg.last_update_time = 0;
11241 attach_task_cfs_rq(p);
11244 static void task_change_group_fair(struct task_struct *p, int type)
11247 case TASK_SET_GROUP:
11248 task_set_group_fair(p);
11251 case TASK_MOVE_GROUP:
11252 task_move_group_fair(p);
11257 void free_fair_sched_group(struct task_group *tg)
11261 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11263 for_each_possible_cpu(i) {
11265 kfree(tg->cfs_rq[i]);
11274 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11276 struct sched_entity *se;
11277 struct cfs_rq *cfs_rq;
11280 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11283 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11287 tg->shares = NICE_0_LOAD;
11289 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11291 for_each_possible_cpu(i) {
11292 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11293 GFP_KERNEL, cpu_to_node(i));
11297 se = kzalloc_node(sizeof(struct sched_entity),
11298 GFP_KERNEL, cpu_to_node(i));
11302 init_cfs_rq(cfs_rq);
11303 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11304 init_entity_runnable_average(se);
11315 void online_fair_sched_group(struct task_group *tg)
11317 struct sched_entity *se;
11318 struct rq_flags rf;
11322 for_each_possible_cpu(i) {
11325 rq_lock_irq(rq, &rf);
11326 update_rq_clock(rq);
11327 attach_entity_cfs_rq(se);
11328 sync_throttle(tg, i);
11329 rq_unlock_irq(rq, &rf);
11333 void unregister_fair_sched_group(struct task_group *tg)
11335 unsigned long flags;
11339 for_each_possible_cpu(cpu) {
11341 remove_entity_load_avg(tg->se[cpu]);
11344 * Only empty task groups can be destroyed; so we can speculatively
11345 * check on_list without danger of it being re-added.
11347 if (!tg->cfs_rq[cpu]->on_list)
11352 raw_spin_rq_lock_irqsave(rq, flags);
11353 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11354 raw_spin_rq_unlock_irqrestore(rq, flags);
11358 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11359 struct sched_entity *se, int cpu,
11360 struct sched_entity *parent)
11362 struct rq *rq = cpu_rq(cpu);
11366 init_cfs_rq_runtime(cfs_rq);
11368 tg->cfs_rq[cpu] = cfs_rq;
11371 /* se could be NULL for root_task_group */
11376 se->cfs_rq = &rq->cfs;
11379 se->cfs_rq = parent->my_q;
11380 se->depth = parent->depth + 1;
11384 /* guarantee group entities always have weight */
11385 update_load_set(&se->load, NICE_0_LOAD);
11386 se->parent = parent;
11389 static DEFINE_MUTEX(shares_mutex);
11391 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11396 * We can't change the weight of the root cgroup.
11401 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11403 mutex_lock(&shares_mutex);
11404 if (tg->shares == shares)
11407 tg->shares = shares;
11408 for_each_possible_cpu(i) {
11409 struct rq *rq = cpu_rq(i);
11410 struct sched_entity *se = tg->se[i];
11411 struct rq_flags rf;
11413 /* Propagate contribution to hierarchy */
11414 rq_lock_irqsave(rq, &rf);
11415 update_rq_clock(rq);
11416 for_each_sched_entity(se) {
11417 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11418 update_cfs_group(se);
11420 rq_unlock_irqrestore(rq, &rf);
11424 mutex_unlock(&shares_mutex);
11427 #else /* CONFIG_FAIR_GROUP_SCHED */
11429 void free_fair_sched_group(struct task_group *tg) { }
11431 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11436 void online_fair_sched_group(struct task_group *tg) { }
11438 void unregister_fair_sched_group(struct task_group *tg) { }
11440 #endif /* CONFIG_FAIR_GROUP_SCHED */
11443 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11445 struct sched_entity *se = &task->se;
11446 unsigned int rr_interval = 0;
11449 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11452 if (rq->cfs.load.weight)
11453 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11455 return rr_interval;
11459 * All the scheduling class methods:
11461 DEFINE_SCHED_CLASS(fair) = {
11463 .enqueue_task = enqueue_task_fair,
11464 .dequeue_task = dequeue_task_fair,
11465 .yield_task = yield_task_fair,
11466 .yield_to_task = yield_to_task_fair,
11468 .check_preempt_curr = check_preempt_wakeup,
11470 .pick_next_task = __pick_next_task_fair,
11471 .put_prev_task = put_prev_task_fair,
11472 .set_next_task = set_next_task_fair,
11475 .balance = balance_fair,
11476 .pick_task = pick_task_fair,
11477 .select_task_rq = select_task_rq_fair,
11478 .migrate_task_rq = migrate_task_rq_fair,
11480 .rq_online = rq_online_fair,
11481 .rq_offline = rq_offline_fair,
11483 .task_dead = task_dead_fair,
11484 .set_cpus_allowed = set_cpus_allowed_common,
11487 .task_tick = task_tick_fair,
11488 .task_fork = task_fork_fair,
11490 .prio_changed = prio_changed_fair,
11491 .switched_from = switched_from_fair,
11492 .switched_to = switched_to_fair,
11494 .get_rr_interval = get_rr_interval_fair,
11496 .update_curr = update_curr_fair,
11498 #ifdef CONFIG_FAIR_GROUP_SCHED
11499 .task_change_group = task_change_group_fair,
11502 #ifdef CONFIG_UCLAMP_TASK
11503 .uclamp_enabled = 1,
11507 #ifdef CONFIG_SCHED_DEBUG
11508 void print_cfs_stats(struct seq_file *m, int cpu)
11510 struct cfs_rq *cfs_rq, *pos;
11513 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11514 print_cfs_rq(m, cpu, cfs_rq);
11518 #ifdef CONFIG_NUMA_BALANCING
11519 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11522 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11523 struct numa_group *ng;
11526 ng = rcu_dereference(p->numa_group);
11527 for_each_online_node(node) {
11528 if (p->numa_faults) {
11529 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11530 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11533 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11534 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11536 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11540 #endif /* CONFIG_NUMA_BALANCING */
11541 #endif /* CONFIG_SCHED_DEBUG */
11543 __init void init_sched_fair_class(void)
11546 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11548 #ifdef CONFIG_NO_HZ_COMMON
11549 nohz.next_balance = jiffies;
11550 nohz.next_blocked = jiffies;
11551 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11558 * Helper functions to facilitate extracting info from tracepoints.
11561 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11564 return cfs_rq ? &cfs_rq->avg : NULL;
11569 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11571 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11575 strlcpy(str, "(null)", len);
11580 cfs_rq_tg_path(cfs_rq, str, len);
11583 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11585 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11587 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11589 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11591 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11594 return rq ? &rq->avg_rt : NULL;
11599 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11601 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11604 return rq ? &rq->avg_dl : NULL;
11609 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11611 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11613 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11614 return rq ? &rq->avg_irq : NULL;
11619 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11621 int sched_trace_rq_cpu(struct rq *rq)
11623 return rq ? cpu_of(rq) : -1;
11625 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11627 int sched_trace_rq_cpu_capacity(struct rq *rq)
11633 SCHED_CAPACITY_SCALE
11637 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
11639 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11642 return rd ? rd->span : NULL;
11647 EXPORT_SYMBOL_GPL(sched_trace_rd_span);
11649 int sched_trace_rq_nr_running(struct rq *rq)
11651 return rq ? rq->nr_running : -1;
11653 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);