2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 * Increase the granularity value when there are more CPUs,
118 * because with more CPUs the 'effective latency' as visible
119 * to users decreases. But the relationship is not linear,
120 * so pick a second-best guess by going with the log2 of the
123 * This idea comes from the SD scheduler of Con Kolivas:
125 static int get_update_sysctl_factor(void)
127 unsigned int cpus = min_t(int, num_online_cpus(), 8);
130 switch (sysctl_sched_tunable_scaling) {
131 case SCHED_TUNABLESCALING_NONE:
134 case SCHED_TUNABLESCALING_LINEAR:
137 case SCHED_TUNABLESCALING_LOG:
139 factor = 1 + ilog2(cpus);
146 static void update_sysctl(void)
148 unsigned int factor = get_update_sysctl_factor();
150 #define SET_SYSCTL(name) \
151 (sysctl_##name = (factor) * normalized_sysctl_##name)
152 SET_SYSCTL(sched_min_granularity);
153 SET_SYSCTL(sched_latency);
154 SET_SYSCTL(sched_wakeup_granularity);
158 void sched_init_granularity(void)
163 #if BITS_PER_LONG == 32
164 # define WMULT_CONST (~0UL)
166 # define WMULT_CONST (1UL << 32)
169 #define WMULT_SHIFT 32
172 * Shift right and round:
174 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
177 * delta *= weight / lw
180 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
181 struct load_weight *lw)
186 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
187 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
188 * 2^SCHED_LOAD_RESOLUTION.
190 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
191 tmp = (u64)delta_exec * scale_load_down(weight);
193 tmp = (u64)delta_exec;
195 if (!lw->inv_weight) {
196 unsigned long w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * Check whether we'd overflow the 64-bit multiplication:
209 if (unlikely(tmp > WMULT_CONST))
210 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
213 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
215 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
219 const struct sched_class fair_sched_class;
221 /**************************************************************
222 * CFS operations on generic schedulable entities:
225 #ifdef CONFIG_FAIR_GROUP_SCHED
227 /* cpu runqueue to which this cfs_rq is attached */
228 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
233 /* An entity is a task if it doesn't "own" a runqueue */
234 #define entity_is_task(se) (!se->my_q)
236 static inline struct task_struct *task_of(struct sched_entity *se)
238 #ifdef CONFIG_SCHED_DEBUG
239 WARN_ON_ONCE(!entity_is_task(se));
241 return container_of(se, struct task_struct, se);
244 /* Walk up scheduling entities hierarchy */
245 #define for_each_sched_entity(se) \
246 for (; se; se = se->parent)
248 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
253 /* runqueue on which this entity is (to be) queued */
254 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
259 /* runqueue "owned" by this group */
260 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
265 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
268 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
270 if (!cfs_rq->on_list) {
272 * Ensure we either appear before our parent (if already
273 * enqueued) or force our parent to appear after us when it is
274 * enqueued. The fact that we always enqueue bottom-up
275 * reduces this to two cases.
277 if (cfs_rq->tg->parent &&
278 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
279 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
280 &rq_of(cfs_rq)->leaf_cfs_rq_list);
282 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
283 &rq_of(cfs_rq)->leaf_cfs_rq_list);
287 /* We should have no load, but we need to update last_decay. */
288 update_cfs_rq_blocked_load(cfs_rq, 0);
292 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 if (cfs_rq->on_list) {
295 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
300 /* Iterate thr' all leaf cfs_rq's on a runqueue */
301 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
302 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
304 /* Do the two (enqueued) entities belong to the same group ? */
306 is_same_group(struct sched_entity *se, struct sched_entity *pse)
308 if (se->cfs_rq == pse->cfs_rq)
314 static inline struct sched_entity *parent_entity(struct sched_entity *se)
319 /* return depth at which a sched entity is present in the hierarchy */
320 static inline int depth_se(struct sched_entity *se)
324 for_each_sched_entity(se)
331 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
333 int se_depth, pse_depth;
336 * preemption test can be made between sibling entities who are in the
337 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
338 * both tasks until we find their ancestors who are siblings of common
342 /* First walk up until both entities are at same depth */
343 se_depth = depth_se(*se);
344 pse_depth = depth_se(*pse);
346 while (se_depth > pse_depth) {
348 *se = parent_entity(*se);
351 while (pse_depth > se_depth) {
353 *pse = parent_entity(*pse);
356 while (!is_same_group(*se, *pse)) {
357 *se = parent_entity(*se);
358 *pse = parent_entity(*pse);
362 #else /* !CONFIG_FAIR_GROUP_SCHED */
364 static inline struct task_struct *task_of(struct sched_entity *se)
366 return container_of(se, struct task_struct, se);
369 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
371 return container_of(cfs_rq, struct rq, cfs);
374 #define entity_is_task(se) 1
376 #define for_each_sched_entity(se) \
377 for (; se; se = NULL)
379 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
381 return &task_rq(p)->cfs;
384 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
386 struct task_struct *p = task_of(se);
387 struct rq *rq = task_rq(p);
392 /* runqueue "owned" by this group */
393 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
398 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
402 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
406 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
407 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
410 is_same_group(struct sched_entity *se, struct sched_entity *pse)
415 static inline struct sched_entity *parent_entity(struct sched_entity *se)
421 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 #endif /* CONFIG_FAIR_GROUP_SCHED */
427 static __always_inline
428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
430 /**************************************************************
431 * Scheduling class tree data structure manipulation methods:
434 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - max_vruntime);
438 max_vruntime = vruntime;
443 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - min_vruntime);
447 min_vruntime = vruntime;
452 static inline int entity_before(struct sched_entity *a,
453 struct sched_entity *b)
455 return (s64)(a->vruntime - b->vruntime) < 0;
458 static void update_min_vruntime(struct cfs_rq *cfs_rq)
460 u64 vruntime = cfs_rq->min_vruntime;
463 vruntime = cfs_rq->curr->vruntime;
465 if (cfs_rq->rb_leftmost) {
466 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 vruntime = se->vruntime;
473 vruntime = min_vruntime(vruntime, se->vruntime);
476 /* ensure we never gain time by being placed backwards. */
477 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 * Enqueue an entity into the rb-tree:
487 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
490 struct rb_node *parent = NULL;
491 struct sched_entity *entry;
495 * Find the right place in the rbtree:
499 entry = rb_entry(parent, struct sched_entity, run_node);
501 * We dont care about collisions. Nodes with
502 * the same key stay together.
504 if (entity_before(se, entry)) {
505 link = &parent->rb_left;
507 link = &parent->rb_right;
513 * Maintain a cache of leftmost tree entries (it is frequently
517 cfs_rq->rb_leftmost = &se->run_node;
519 rb_link_node(&se->run_node, parent, link);
520 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
523 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 if (cfs_rq->rb_leftmost == &se->run_node) {
526 struct rb_node *next_node;
528 next_node = rb_next(&se->run_node);
529 cfs_rq->rb_leftmost = next_node;
532 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
535 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
537 struct rb_node *left = cfs_rq->rb_leftmost;
542 return rb_entry(left, struct sched_entity, run_node);
545 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
547 struct rb_node *next = rb_next(&se->run_node);
552 return rb_entry(next, struct sched_entity, run_node);
555 #ifdef CONFIG_SCHED_DEBUG
556 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 return rb_entry(last, struct sched_entity, run_node);
566 /**************************************************************
567 * Scheduling class statistics methods:
570 int sched_proc_update_handler(struct ctl_table *table, int write,
571 void __user *buffer, size_t *lenp,
574 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
575 int factor = get_update_sysctl_factor();
580 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
581 sysctl_sched_min_granularity);
583 #define WRT_SYSCTL(name) \
584 (normalized_sysctl_##name = sysctl_##name / (factor))
585 WRT_SYSCTL(sched_min_granularity);
586 WRT_SYSCTL(sched_latency);
587 WRT_SYSCTL(sched_wakeup_granularity);
597 static inline unsigned long
598 calc_delta_fair(unsigned long delta, struct sched_entity *se)
600 if (unlikely(se->load.weight != NICE_0_LOAD))
601 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
607 * The idea is to set a period in which each task runs once.
609 * When there are too many tasks (sched_nr_latency) we have to stretch
610 * this period because otherwise the slices get too small.
612 * p = (nr <= nl) ? l : l*nr/nl
614 static u64 __sched_period(unsigned long nr_running)
616 u64 period = sysctl_sched_latency;
617 unsigned long nr_latency = sched_nr_latency;
619 if (unlikely(nr_running > nr_latency)) {
620 period = sysctl_sched_min_granularity;
621 period *= nr_running;
628 * We calculate the wall-time slice from the period by taking a part
629 * proportional to the weight.
633 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
635 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
637 for_each_sched_entity(se) {
638 struct load_weight *load;
639 struct load_weight lw;
641 cfs_rq = cfs_rq_of(se);
642 load = &cfs_rq->load;
644 if (unlikely(!se->on_rq)) {
647 update_load_add(&lw, se->load.weight);
650 slice = calc_delta_mine(slice, se->load.weight, load);
656 * We calculate the vruntime slice of a to-be-inserted task.
660 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
662 return calc_delta_fair(sched_slice(cfs_rq, se), se);
666 * Update the current task's runtime statistics. Skip current tasks that
667 * are not in our scheduling class.
670 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
671 unsigned long delta_exec)
673 unsigned long delta_exec_weighted;
675 schedstat_set(curr->statistics.exec_max,
676 max((u64)delta_exec, curr->statistics.exec_max));
678 curr->sum_exec_runtime += delta_exec;
679 schedstat_add(cfs_rq, exec_clock, delta_exec);
680 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
682 curr->vruntime += delta_exec_weighted;
683 update_min_vruntime(cfs_rq);
686 static void update_curr(struct cfs_rq *cfs_rq)
688 struct sched_entity *curr = cfs_rq->curr;
689 u64 now = rq_of(cfs_rq)->clock_task;
690 unsigned long delta_exec;
696 * Get the amount of time the current task was running
697 * since the last time we changed load (this cannot
698 * overflow on 32 bits):
700 delta_exec = (unsigned long)(now - curr->exec_start);
704 __update_curr(cfs_rq, curr, delta_exec);
705 curr->exec_start = now;
707 if (entity_is_task(curr)) {
708 struct task_struct *curtask = task_of(curr);
710 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
711 cpuacct_charge(curtask, delta_exec);
712 account_group_exec_runtime(curtask, delta_exec);
715 account_cfs_rq_runtime(cfs_rq, delta_exec);
719 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
721 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
725 * Task is being enqueued - update stats:
727 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 * Are we enqueueing a waiting task? (for current tasks
731 * a dequeue/enqueue event is a NOP)
733 if (se != cfs_rq->curr)
734 update_stats_wait_start(cfs_rq, se);
738 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
740 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
741 rq_of(cfs_rq)->clock - se->statistics.wait_start));
742 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
743 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
744 rq_of(cfs_rq)->clock - se->statistics.wait_start);
745 #ifdef CONFIG_SCHEDSTATS
746 if (entity_is_task(se)) {
747 trace_sched_stat_wait(task_of(se),
748 rq_of(cfs_rq)->clock - se->statistics.wait_start);
751 schedstat_set(se->statistics.wait_start, 0);
755 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
758 * Mark the end of the wait period if dequeueing a
761 if (se != cfs_rq->curr)
762 update_stats_wait_end(cfs_rq, se);
766 * We are picking a new current task - update its stats:
769 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
772 * We are starting a new run period:
774 se->exec_start = rq_of(cfs_rq)->clock_task;
777 /**************************************************
778 * Scheduling class queueing methods:
781 #ifdef CONFIG_NUMA_BALANCING
783 * numa task sample period in ms
785 unsigned int sysctl_numa_balancing_scan_period_min = 100;
786 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
787 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
789 /* Portion of address space to scan in MB */
790 unsigned int sysctl_numa_balancing_scan_size = 256;
792 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
793 unsigned int sysctl_numa_balancing_scan_delay = 1000;
795 static void task_numa_placement(struct task_struct *p)
799 if (!p->mm) /* for example, ksmd faulting in a user's mm */
801 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
802 if (p->numa_scan_seq == seq)
804 p->numa_scan_seq = seq;
806 /* FIXME: Scheduling placement policy hints go here */
810 * Got a PROT_NONE fault for a page on @node.
812 void task_numa_fault(int node, int pages, bool migrated)
814 struct task_struct *p = current;
816 if (!sched_feat_numa(NUMA))
819 /* FIXME: Allocate task-specific structure for placement policy here */
822 * If pages are properly placed (did not migrate) then scan slower.
823 * This is reset periodically in case of phase changes
826 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
827 p->numa_scan_period + jiffies_to_msecs(10));
829 task_numa_placement(p);
832 static void reset_ptenuma_scan(struct task_struct *p)
834 ACCESS_ONCE(p->mm->numa_scan_seq)++;
835 p->mm->numa_scan_offset = 0;
839 * The expensive part of numa migration is done from task_work context.
840 * Triggered from task_tick_numa().
842 void task_numa_work(struct callback_head *work)
844 unsigned long migrate, next_scan, now = jiffies;
845 struct task_struct *p = current;
846 struct mm_struct *mm = p->mm;
847 struct vm_area_struct *vma;
848 unsigned long start, end;
851 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
853 work->next = work; /* protect against double add */
855 * Who cares about NUMA placement when they're dying.
857 * NOTE: make sure not to dereference p->mm before this check,
858 * exit_task_work() happens _after_ exit_mm() so we could be called
859 * without p->mm even though we still had it when we enqueued this
862 if (p->flags & PF_EXITING)
866 * We do not care about task placement until a task runs on a node
867 * other than the first one used by the address space. This is
868 * largely because migrations are driven by what CPU the task
869 * is running on. If it's never scheduled on another node, it'll
870 * not migrate so why bother trapping the fault.
872 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
873 mm->first_nid = numa_node_id();
874 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
875 /* Are we running on a new node yet? */
876 if (numa_node_id() == mm->first_nid &&
877 !sched_feat_numa(NUMA_FORCE))
880 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
884 * Reset the scan period if enough time has gone by. Objective is that
885 * scanning will be reduced if pages are properly placed. As tasks
886 * can enter different phases this needs to be re-examined. Lacking
887 * proper tracking of reference behaviour, this blunt hammer is used.
889 migrate = mm->numa_next_reset;
890 if (time_after(now, migrate)) {
891 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
892 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
893 xchg(&mm->numa_next_reset, next_scan);
897 * Enforce maximal scan/migration frequency..
899 migrate = mm->numa_next_scan;
900 if (time_before(now, migrate))
903 if (p->numa_scan_period == 0)
904 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
906 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
907 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
911 * Do not set pte_numa if the current running node is rate-limited.
912 * This loses statistics on the fault but if we are unwilling to
913 * migrate to this node, it is less likely we can do useful work
915 if (migrate_ratelimited(numa_node_id()))
918 start = mm->numa_scan_offset;
919 pages = sysctl_numa_balancing_scan_size;
920 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
924 down_read(&mm->mmap_sem);
925 vma = find_vma(mm, start);
927 reset_ptenuma_scan(p);
931 for (; vma; vma = vma->vm_next) {
932 if (!vma_migratable(vma))
935 /* Skip small VMAs. They are not likely to be of relevance */
936 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
940 start = max(start, vma->vm_start);
941 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
942 end = min(end, vma->vm_end);
943 pages -= change_prot_numa(vma, start, end);
948 } while (end != vma->vm_end);
953 * It is possible to reach the end of the VMA list but the last few VMAs are
954 * not guaranteed to the vma_migratable. If they are not, we would find the
955 * !migratable VMA on the next scan but not reset the scanner to the start
959 mm->numa_scan_offset = start;
961 reset_ptenuma_scan(p);
962 up_read(&mm->mmap_sem);
966 * Drive the periodic memory faults..
968 void task_tick_numa(struct rq *rq, struct task_struct *curr)
970 struct callback_head *work = &curr->numa_work;
974 * We don't care about NUMA placement if we don't have memory.
976 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
980 * Using runtime rather than walltime has the dual advantage that
981 * we (mostly) drive the selection from busy threads and that the
982 * task needs to have done some actual work before we bother with
985 now = curr->se.sum_exec_runtime;
986 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
988 if (now - curr->node_stamp > period) {
989 if (!curr->node_stamp)
990 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
991 curr->node_stamp = now;
993 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
994 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
995 task_work_add(curr, work, true);
1000 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1003 #endif /* CONFIG_NUMA_BALANCING */
1006 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1008 update_load_add(&cfs_rq->load, se->load.weight);
1009 if (!parent_entity(se))
1010 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1012 if (entity_is_task(se))
1013 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1015 cfs_rq->nr_running++;
1019 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1021 update_load_sub(&cfs_rq->load, se->load.weight);
1022 if (!parent_entity(se))
1023 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1024 if (entity_is_task(se))
1025 list_del_init(&se->group_node);
1026 cfs_rq->nr_running--;
1029 #ifdef CONFIG_FAIR_GROUP_SCHED
1031 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1036 * Use this CPU's actual weight instead of the last load_contribution
1037 * to gain a more accurate current total weight. See
1038 * update_cfs_rq_load_contribution().
1040 tg_weight = atomic64_read(&tg->load_avg);
1041 tg_weight -= cfs_rq->tg_load_contrib;
1042 tg_weight += cfs_rq->load.weight;
1047 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1049 long tg_weight, load, shares;
1051 tg_weight = calc_tg_weight(tg, cfs_rq);
1052 load = cfs_rq->load.weight;
1054 shares = (tg->shares * load);
1056 shares /= tg_weight;
1058 if (shares < MIN_SHARES)
1059 shares = MIN_SHARES;
1060 if (shares > tg->shares)
1061 shares = tg->shares;
1065 # else /* CONFIG_SMP */
1066 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1070 # endif /* CONFIG_SMP */
1071 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1072 unsigned long weight)
1075 /* commit outstanding execution time */
1076 if (cfs_rq->curr == se)
1077 update_curr(cfs_rq);
1078 account_entity_dequeue(cfs_rq, se);
1081 update_load_set(&se->load, weight);
1084 account_entity_enqueue(cfs_rq, se);
1087 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1089 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1091 struct task_group *tg;
1092 struct sched_entity *se;
1096 se = tg->se[cpu_of(rq_of(cfs_rq))];
1097 if (!se || throttled_hierarchy(cfs_rq))
1100 if (likely(se->load.weight == tg->shares))
1103 shares = calc_cfs_shares(cfs_rq, tg);
1105 reweight_entity(cfs_rq_of(se), se, shares);
1107 #else /* CONFIG_FAIR_GROUP_SCHED */
1108 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1111 #endif /* CONFIG_FAIR_GROUP_SCHED */
1113 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1114 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1116 * We choose a half-life close to 1 scheduling period.
1117 * Note: The tables below are dependent on this value.
1119 #define LOAD_AVG_PERIOD 32
1120 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1121 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1123 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1124 static const u32 runnable_avg_yN_inv[] = {
1125 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1126 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1127 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1128 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1129 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1130 0x85aac367, 0x82cd8698,
1134 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1135 * over-estimates when re-combining.
1137 static const u32 runnable_avg_yN_sum[] = {
1138 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1139 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1140 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1145 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1147 static __always_inline u64 decay_load(u64 val, u64 n)
1149 unsigned int local_n;
1153 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1156 /* after bounds checking we can collapse to 32-bit */
1160 * As y^PERIOD = 1/2, we can combine
1161 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1162 * With a look-up table which covers k^n (n<PERIOD)
1164 * To achieve constant time decay_load.
1166 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1167 val >>= local_n / LOAD_AVG_PERIOD;
1168 local_n %= LOAD_AVG_PERIOD;
1171 val *= runnable_avg_yN_inv[local_n];
1172 /* We don't use SRR here since we always want to round down. */
1177 * For updates fully spanning n periods, the contribution to runnable
1178 * average will be: \Sum 1024*y^n
1180 * We can compute this reasonably efficiently by combining:
1181 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1183 static u32 __compute_runnable_contrib(u64 n)
1187 if (likely(n <= LOAD_AVG_PERIOD))
1188 return runnable_avg_yN_sum[n];
1189 else if (unlikely(n >= LOAD_AVG_MAX_N))
1190 return LOAD_AVG_MAX;
1192 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1194 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1195 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1197 n -= LOAD_AVG_PERIOD;
1198 } while (n > LOAD_AVG_PERIOD);
1200 contrib = decay_load(contrib, n);
1201 return contrib + runnable_avg_yN_sum[n];
1205 * We can represent the historical contribution to runnable average as the
1206 * coefficients of a geometric series. To do this we sub-divide our runnable
1207 * history into segments of approximately 1ms (1024us); label the segment that
1208 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1210 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1212 * (now) (~1ms ago) (~2ms ago)
1214 * Let u_i denote the fraction of p_i that the entity was runnable.
1216 * We then designate the fractions u_i as our co-efficients, yielding the
1217 * following representation of historical load:
1218 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1220 * We choose y based on the with of a reasonably scheduling period, fixing:
1223 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1224 * approximately half as much as the contribution to load within the last ms
1227 * When a period "rolls over" and we have new u_0`, multiplying the previous
1228 * sum again by y is sufficient to update:
1229 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1230 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1232 static __always_inline int __update_entity_runnable_avg(u64 now,
1233 struct sched_avg *sa,
1238 u32 runnable_contrib;
1239 int delta_w, decayed = 0;
1241 delta = now - sa->last_runnable_update;
1243 * This should only happen when time goes backwards, which it
1244 * unfortunately does during sched clock init when we swap over to TSC.
1246 if ((s64)delta < 0) {
1247 sa->last_runnable_update = now;
1252 * Use 1024ns as the unit of measurement since it's a reasonable
1253 * approximation of 1us and fast to compute.
1258 sa->last_runnable_update = now;
1260 /* delta_w is the amount already accumulated against our next period */
1261 delta_w = sa->runnable_avg_period % 1024;
1262 if (delta + delta_w >= 1024) {
1263 /* period roll-over */
1267 * Now that we know we're crossing a period boundary, figure
1268 * out how much from delta we need to complete the current
1269 * period and accrue it.
1271 delta_w = 1024 - delta_w;
1273 sa->runnable_avg_sum += delta_w;
1275 sa->usage_avg_sum += delta_w;
1276 sa->runnable_avg_period += delta_w;
1280 /* Figure out how many additional periods this update spans */
1281 periods = delta / 1024;
1284 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1286 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1288 sa->usage_avg_sum = decay_load(sa->usage_avg_sum, periods + 1);
1290 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1291 runnable_contrib = __compute_runnable_contrib(periods);
1293 sa->runnable_avg_sum += runnable_contrib;
1295 sa->usage_avg_sum += runnable_contrib;
1296 sa->runnable_avg_period += runnable_contrib;
1299 /* Remainder of delta accrued against u_0` */
1301 sa->runnable_avg_sum += delta;
1303 sa->usage_avg_sum += delta;
1304 sa->runnable_avg_period += delta;
1309 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1310 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1312 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1313 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1315 decays -= se->avg.decay_count;
1319 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1320 se->avg.decay_count = 0;
1325 #ifdef CONFIG_FAIR_GROUP_SCHED
1326 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1329 struct task_group *tg = cfs_rq->tg;
1332 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1333 tg_contrib -= cfs_rq->tg_load_contrib;
1335 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1336 atomic64_add(tg_contrib, &tg->load_avg);
1337 cfs_rq->tg_load_contrib += tg_contrib;
1342 * Aggregate cfs_rq runnable averages into an equivalent task_group
1343 * representation for computing load contributions.
1345 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1346 struct cfs_rq *cfs_rq)
1348 struct task_group *tg = cfs_rq->tg;
1349 long contrib, usage_contrib;
1351 /* The fraction of a cpu used by this cfs_rq */
1352 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1353 sa->runnable_avg_period + 1);
1354 contrib -= cfs_rq->tg_runnable_contrib;
1356 usage_contrib = div_u64(sa->usage_avg_sum << NICE_0_SHIFT,
1357 sa->runnable_avg_period + 1);
1358 usage_contrib -= cfs_rq->tg_usage_contrib;
1361 * contrib/usage at this point represent deltas, only update if they
1364 if ((abs(contrib) > cfs_rq->tg_runnable_contrib / 64) ||
1365 (abs(usage_contrib) > cfs_rq->tg_usage_contrib / 64)) {
1366 atomic_add(contrib, &tg->runnable_avg);
1367 cfs_rq->tg_runnable_contrib += contrib;
1369 atomic_add(usage_contrib, &tg->usage_avg);
1370 cfs_rq->tg_usage_contrib += usage_contrib;
1374 static inline void __update_group_entity_contrib(struct sched_entity *se)
1376 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1377 struct task_group *tg = cfs_rq->tg;
1382 contrib = cfs_rq->tg_load_contrib * tg->shares;
1383 se->avg.load_avg_contrib = div64_u64(contrib,
1384 atomic64_read(&tg->load_avg) + 1);
1387 * For group entities we need to compute a correction term in the case
1388 * that they are consuming <1 cpu so that we would contribute the same
1389 * load as a task of equal weight.
1391 * Explicitly co-ordinating this measurement would be expensive, but
1392 * fortunately the sum of each cpus contribution forms a usable
1393 * lower-bound on the true value.
1395 * Consider the aggregate of 2 contributions. Either they are disjoint
1396 * (and the sum represents true value) or they are disjoint and we are
1397 * understating by the aggregate of their overlap.
1399 * Extending this to N cpus, for a given overlap, the maximum amount we
1400 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1401 * cpus that overlap for this interval and w_i is the interval width.
1403 * On a small machine; the first term is well-bounded which bounds the
1404 * total error since w_i is a subset of the period. Whereas on a
1405 * larger machine, while this first term can be larger, if w_i is the
1406 * of consequential size guaranteed to see n_i*w_i quickly converge to
1407 * our upper bound of 1-cpu.
1409 runnable_avg = atomic_read(&tg->runnable_avg);
1410 if (runnable_avg < NICE_0_LOAD) {
1411 se->avg.load_avg_contrib *= runnable_avg;
1412 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1416 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1417 int force_update) {}
1418 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1419 struct cfs_rq *cfs_rq) {}
1420 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1423 static inline void __update_task_entity_contrib(struct sched_entity *se)
1427 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1428 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1429 contrib /= (se->avg.runnable_avg_period + 1);
1430 se->avg.load_avg_contrib = scale_load(contrib);
1431 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1432 contrib /= (se->avg.runnable_avg_period + 1);
1433 se->avg.load_avg_ratio = scale_load(contrib);
1436 /* Compute the current contribution to load_avg by se, return any delta */
1437 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1439 long old_contrib = se->avg.load_avg_contrib;
1441 if (entity_is_task(se)) {
1442 __update_task_entity_contrib(se);
1444 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1445 __update_group_entity_contrib(se);
1448 return se->avg.load_avg_contrib - old_contrib;
1451 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1454 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1455 cfs_rq->blocked_load_avg -= load_contrib;
1457 cfs_rq->blocked_load_avg = 0;
1460 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1462 /* Update a sched_entity's runnable average */
1463 static inline void update_entity_load_avg(struct sched_entity *se,
1466 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1471 * For a group entity we need to use their owned cfs_rq_clock_task() in
1472 * case they are the parent of a throttled hierarchy.
1474 if (entity_is_task(se))
1475 now = cfs_rq_clock_task(cfs_rq);
1477 now = cfs_rq_clock_task(group_cfs_rq(se));
1479 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1480 cfs_rq->curr == se))
1483 contrib_delta = __update_entity_load_avg_contrib(se);
1489 cfs_rq->runnable_load_avg += contrib_delta;
1491 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1495 * Decay the load contributed by all blocked children and account this so that
1496 * their contribution may appropriately discounted when they wake up.
1498 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1500 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1503 decays = now - cfs_rq->last_decay;
1504 if (!decays && !force_update)
1507 if (atomic64_read(&cfs_rq->removed_load)) {
1508 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1509 subtract_blocked_load_contrib(cfs_rq, removed_load);
1513 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1515 atomic64_add(decays, &cfs_rq->decay_counter);
1516 cfs_rq->last_decay = now;
1519 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1522 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1524 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1526 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1529 /* Add the load generated by se into cfs_rq's child load-average */
1530 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1531 struct sched_entity *se,
1535 * We track migrations using entity decay_count <= 0, on a wake-up
1536 * migration we use a negative decay count to track the remote decays
1537 * accumulated while sleeping.
1539 if (unlikely(se->avg.decay_count <= 0)) {
1540 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1541 if (se->avg.decay_count) {
1543 * In a wake-up migration we have to approximate the
1544 * time sleeping. This is because we can't synchronize
1545 * clock_task between the two cpus, and it is not
1546 * guaranteed to be read-safe. Instead, we can
1547 * approximate this using our carried decays, which are
1548 * explicitly atomically readable.
1550 se->avg.last_runnable_update -= (-se->avg.decay_count)
1552 update_entity_load_avg(se, 0);
1553 /* Indicate that we're now synchronized and on-rq */
1554 se->avg.decay_count = 0;
1558 __synchronize_entity_decay(se);
1561 /* migrated tasks did not contribute to our blocked load */
1563 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1564 update_entity_load_avg(se, 0);
1567 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1568 /* we force update consideration on load-balancer moves */
1569 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1573 * Remove se's load from this cfs_rq child load-average, if the entity is
1574 * transitioning to a blocked state we track its projected decay using
1577 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1578 struct sched_entity *se,
1581 update_entity_load_avg(se, 1);
1582 /* we force update consideration on load-balancer moves */
1583 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1585 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1587 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1588 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1589 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1593 * Update the rq's load with the elapsed running time before entering
1594 * idle. if the last scheduled task is not a CFS task, idle_enter will
1595 * be the only way to update the runnable statistic.
1597 void idle_enter_fair(struct rq *this_rq)
1599 update_rq_runnable_avg(this_rq, 1);
1603 * Update the rq's load with the elapsed idle time before a task is
1604 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1605 * be the only way to update the runnable statistic.
1607 void idle_exit_fair(struct rq *this_rq)
1609 update_rq_runnable_avg(this_rq, 0);
1613 static inline void update_entity_load_avg(struct sched_entity *se,
1614 int update_cfs_rq) {}
1615 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1616 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1617 struct sched_entity *se,
1619 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1620 struct sched_entity *se,
1622 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1623 int force_update) {}
1626 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1628 #ifdef CONFIG_SCHEDSTATS
1629 struct task_struct *tsk = NULL;
1631 if (entity_is_task(se))
1634 if (se->statistics.sleep_start) {
1635 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1640 if (unlikely(delta > se->statistics.sleep_max))
1641 se->statistics.sleep_max = delta;
1643 se->statistics.sleep_start = 0;
1644 se->statistics.sum_sleep_runtime += delta;
1647 account_scheduler_latency(tsk, delta >> 10, 1);
1648 trace_sched_stat_sleep(tsk, delta);
1651 if (se->statistics.block_start) {
1652 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1657 if (unlikely(delta > se->statistics.block_max))
1658 se->statistics.block_max = delta;
1660 se->statistics.block_start = 0;
1661 se->statistics.sum_sleep_runtime += delta;
1664 if (tsk->in_iowait) {
1665 se->statistics.iowait_sum += delta;
1666 se->statistics.iowait_count++;
1667 trace_sched_stat_iowait(tsk, delta);
1670 trace_sched_stat_blocked(tsk, delta);
1673 * Blocking time is in units of nanosecs, so shift by
1674 * 20 to get a milliseconds-range estimation of the
1675 * amount of time that the task spent sleeping:
1677 if (unlikely(prof_on == SLEEP_PROFILING)) {
1678 profile_hits(SLEEP_PROFILING,
1679 (void *)get_wchan(tsk),
1682 account_scheduler_latency(tsk, delta >> 10, 0);
1688 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1690 #ifdef CONFIG_SCHED_DEBUG
1691 s64 d = se->vruntime - cfs_rq->min_vruntime;
1696 if (d > 3*sysctl_sched_latency)
1697 schedstat_inc(cfs_rq, nr_spread_over);
1702 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1704 u64 vruntime = cfs_rq->min_vruntime;
1707 * The 'current' period is already promised to the current tasks,
1708 * however the extra weight of the new task will slow them down a
1709 * little, place the new task so that it fits in the slot that
1710 * stays open at the end.
1712 if (initial && sched_feat(START_DEBIT))
1713 vruntime += sched_vslice(cfs_rq, se);
1715 /* sleeps up to a single latency don't count. */
1717 unsigned long thresh = sysctl_sched_latency;
1720 * Halve their sleep time's effect, to allow
1721 * for a gentler effect of sleepers:
1723 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1729 /* ensure we never gain time by being placed backwards. */
1730 se->vruntime = max_vruntime(se->vruntime, vruntime);
1733 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1736 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1739 * Update the normalized vruntime before updating min_vruntime
1740 * through callig update_curr().
1742 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1743 se->vruntime += cfs_rq->min_vruntime;
1746 * Update run-time statistics of the 'current'.
1748 update_curr(cfs_rq);
1749 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1750 account_entity_enqueue(cfs_rq, se);
1751 update_cfs_shares(cfs_rq);
1753 if (flags & ENQUEUE_WAKEUP) {
1754 place_entity(cfs_rq, se, 0);
1755 enqueue_sleeper(cfs_rq, se);
1758 update_stats_enqueue(cfs_rq, se);
1759 check_spread(cfs_rq, se);
1760 if (se != cfs_rq->curr)
1761 __enqueue_entity(cfs_rq, se);
1764 if (cfs_rq->nr_running == 1) {
1765 list_add_leaf_cfs_rq(cfs_rq);
1766 check_enqueue_throttle(cfs_rq);
1770 static void __clear_buddies_last(struct sched_entity *se)
1772 for_each_sched_entity(se) {
1773 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1774 if (cfs_rq->last == se)
1775 cfs_rq->last = NULL;
1781 static void __clear_buddies_next(struct sched_entity *se)
1783 for_each_sched_entity(se) {
1784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1785 if (cfs_rq->next == se)
1786 cfs_rq->next = NULL;
1792 static void __clear_buddies_skip(struct sched_entity *se)
1794 for_each_sched_entity(se) {
1795 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1796 if (cfs_rq->skip == se)
1797 cfs_rq->skip = NULL;
1803 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1805 if (cfs_rq->last == se)
1806 __clear_buddies_last(se);
1808 if (cfs_rq->next == se)
1809 __clear_buddies_next(se);
1811 if (cfs_rq->skip == se)
1812 __clear_buddies_skip(se);
1815 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1818 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1821 * Update run-time statistics of the 'current'.
1823 update_curr(cfs_rq);
1824 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1826 update_stats_dequeue(cfs_rq, se);
1827 if (flags & DEQUEUE_SLEEP) {
1828 #ifdef CONFIG_SCHEDSTATS
1829 if (entity_is_task(se)) {
1830 struct task_struct *tsk = task_of(se);
1832 if (tsk->state & TASK_INTERRUPTIBLE)
1833 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1834 if (tsk->state & TASK_UNINTERRUPTIBLE)
1835 se->statistics.block_start = rq_of(cfs_rq)->clock;
1840 clear_buddies(cfs_rq, se);
1842 if (se != cfs_rq->curr)
1843 __dequeue_entity(cfs_rq, se);
1845 account_entity_dequeue(cfs_rq, se);
1848 * Normalize the entity after updating the min_vruntime because the
1849 * update can refer to the ->curr item and we need to reflect this
1850 * movement in our normalized position.
1852 if (!(flags & DEQUEUE_SLEEP))
1853 se->vruntime -= cfs_rq->min_vruntime;
1855 /* return excess runtime on last dequeue */
1856 return_cfs_rq_runtime(cfs_rq);
1858 update_min_vruntime(cfs_rq);
1859 update_cfs_shares(cfs_rq);
1863 * Preempt the current task with a newly woken task if needed:
1866 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1868 unsigned long ideal_runtime, delta_exec;
1869 struct sched_entity *se;
1872 ideal_runtime = sched_slice(cfs_rq, curr);
1873 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1874 if (delta_exec > ideal_runtime) {
1875 resched_task(rq_of(cfs_rq)->curr);
1877 * The current task ran long enough, ensure it doesn't get
1878 * re-elected due to buddy favours.
1880 clear_buddies(cfs_rq, curr);
1885 * Ensure that a task that missed wakeup preemption by a
1886 * narrow margin doesn't have to wait for a full slice.
1887 * This also mitigates buddy induced latencies under load.
1889 if (delta_exec < sysctl_sched_min_granularity)
1892 se = __pick_first_entity(cfs_rq);
1893 delta = curr->vruntime - se->vruntime;
1898 if (delta > ideal_runtime)
1899 resched_task(rq_of(cfs_rq)->curr);
1903 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1905 /* 'current' is not kept within the tree. */
1908 * Any task has to be enqueued before it get to execute on
1909 * a CPU. So account for the time it spent waiting on the
1912 update_stats_wait_end(cfs_rq, se);
1913 __dequeue_entity(cfs_rq, se);
1914 update_entity_load_avg(se, 1);
1917 update_stats_curr_start(cfs_rq, se);
1919 #ifdef CONFIG_SCHEDSTATS
1921 * Track our maximum slice length, if the CPU's load is at
1922 * least twice that of our own weight (i.e. dont track it
1923 * when there are only lesser-weight tasks around):
1925 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1926 se->statistics.slice_max = max(se->statistics.slice_max,
1927 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1930 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1934 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1937 * Pick the next process, keeping these things in mind, in this order:
1938 * 1) keep things fair between processes/task groups
1939 * 2) pick the "next" process, since someone really wants that to run
1940 * 3) pick the "last" process, for cache locality
1941 * 4) do not run the "skip" process, if something else is available
1943 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1945 struct sched_entity *se = __pick_first_entity(cfs_rq);
1946 struct sched_entity *left = se;
1949 * Avoid running the skip buddy, if running something else can
1950 * be done without getting too unfair.
1952 if (cfs_rq->skip == se) {
1953 struct sched_entity *second = __pick_next_entity(se);
1954 if (second && wakeup_preempt_entity(second, left) < 1)
1959 * Prefer last buddy, try to return the CPU to a preempted task.
1961 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1965 * Someone really wants this to run. If it's not unfair, run it.
1967 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1970 clear_buddies(cfs_rq, se);
1975 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1977 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1980 * If still on the runqueue then deactivate_task()
1981 * was not called and update_curr() has to be done:
1984 update_curr(cfs_rq);
1986 /* throttle cfs_rqs exceeding runtime */
1987 check_cfs_rq_runtime(cfs_rq);
1989 check_spread(cfs_rq, prev);
1991 update_stats_wait_start(cfs_rq, prev);
1992 /* Put 'current' back into the tree. */
1993 __enqueue_entity(cfs_rq, prev);
1994 /* in !on_rq case, update occurred at dequeue */
1995 update_entity_load_avg(prev, 1);
1997 cfs_rq->curr = NULL;
2001 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2004 * Update run-time statistics of the 'current'.
2006 update_curr(cfs_rq);
2009 * Ensure that runnable average is periodically updated.
2011 update_entity_load_avg(curr, 1);
2012 update_cfs_rq_blocked_load(cfs_rq, 1);
2014 #ifdef CONFIG_SCHED_HRTICK
2016 * queued ticks are scheduled to match the slice, so don't bother
2017 * validating it and just reschedule.
2020 resched_task(rq_of(cfs_rq)->curr);
2024 * don't let the period tick interfere with the hrtick preemption
2026 if (!sched_feat(DOUBLE_TICK) &&
2027 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2031 if (cfs_rq->nr_running > 1)
2032 check_preempt_tick(cfs_rq, curr);
2036 /**************************************************
2037 * CFS bandwidth control machinery
2040 #ifdef CONFIG_CFS_BANDWIDTH
2042 #ifdef HAVE_JUMP_LABEL
2043 static struct static_key __cfs_bandwidth_used;
2045 static inline bool cfs_bandwidth_used(void)
2047 return static_key_false(&__cfs_bandwidth_used);
2050 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2052 /* only need to count groups transitioning between enabled/!enabled */
2053 if (enabled && !was_enabled)
2054 static_key_slow_inc(&__cfs_bandwidth_used);
2055 else if (!enabled && was_enabled)
2056 static_key_slow_dec(&__cfs_bandwidth_used);
2058 #else /* HAVE_JUMP_LABEL */
2059 static bool cfs_bandwidth_used(void)
2064 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2065 #endif /* HAVE_JUMP_LABEL */
2068 * default period for cfs group bandwidth.
2069 * default: 0.1s, units: nanoseconds
2071 static inline u64 default_cfs_period(void)
2073 return 100000000ULL;
2076 static inline u64 sched_cfs_bandwidth_slice(void)
2078 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2082 * Replenish runtime according to assigned quota and update expiration time.
2083 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2084 * additional synchronization around rq->lock.
2086 * requires cfs_b->lock
2088 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2092 if (cfs_b->quota == RUNTIME_INF)
2095 now = sched_clock_cpu(smp_processor_id());
2096 cfs_b->runtime = cfs_b->quota;
2097 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2100 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2102 return &tg->cfs_bandwidth;
2105 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2106 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2108 if (unlikely(cfs_rq->throttle_count))
2109 return cfs_rq->throttled_clock_task;
2111 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2114 /* returns 0 on failure to allocate runtime */
2115 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2117 struct task_group *tg = cfs_rq->tg;
2118 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2119 u64 amount = 0, min_amount, expires;
2121 /* note: this is a positive sum as runtime_remaining <= 0 */
2122 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2124 raw_spin_lock(&cfs_b->lock);
2125 if (cfs_b->quota == RUNTIME_INF)
2126 amount = min_amount;
2129 * If the bandwidth pool has become inactive, then at least one
2130 * period must have elapsed since the last consumption.
2131 * Refresh the global state and ensure bandwidth timer becomes
2134 if (!cfs_b->timer_active) {
2135 __refill_cfs_bandwidth_runtime(cfs_b);
2136 __start_cfs_bandwidth(cfs_b);
2139 if (cfs_b->runtime > 0) {
2140 amount = min(cfs_b->runtime, min_amount);
2141 cfs_b->runtime -= amount;
2145 expires = cfs_b->runtime_expires;
2146 raw_spin_unlock(&cfs_b->lock);
2148 cfs_rq->runtime_remaining += amount;
2150 * we may have advanced our local expiration to account for allowed
2151 * spread between our sched_clock and the one on which runtime was
2154 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2155 cfs_rq->runtime_expires = expires;
2157 return cfs_rq->runtime_remaining > 0;
2161 * Note: This depends on the synchronization provided by sched_clock and the
2162 * fact that rq->clock snapshots this value.
2164 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2166 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2167 struct rq *rq = rq_of(cfs_rq);
2169 /* if the deadline is ahead of our clock, nothing to do */
2170 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2173 if (cfs_rq->runtime_remaining < 0)
2177 * If the local deadline has passed we have to consider the
2178 * possibility that our sched_clock is 'fast' and the global deadline
2179 * has not truly expired.
2181 * Fortunately we can check determine whether this the case by checking
2182 * whether the global deadline has advanced.
2185 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2186 /* extend local deadline, drift is bounded above by 2 ticks */
2187 cfs_rq->runtime_expires += TICK_NSEC;
2189 /* global deadline is ahead, expiration has passed */
2190 cfs_rq->runtime_remaining = 0;
2194 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2195 unsigned long delta_exec)
2197 /* dock delta_exec before expiring quota (as it could span periods) */
2198 cfs_rq->runtime_remaining -= delta_exec;
2199 expire_cfs_rq_runtime(cfs_rq);
2201 if (likely(cfs_rq->runtime_remaining > 0))
2205 * if we're unable to extend our runtime we resched so that the active
2206 * hierarchy can be throttled
2208 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2209 resched_task(rq_of(cfs_rq)->curr);
2212 static __always_inline
2213 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2215 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2218 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2221 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2223 return cfs_bandwidth_used() && cfs_rq->throttled;
2226 /* check whether cfs_rq, or any parent, is throttled */
2227 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2229 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2233 * Ensure that neither of the group entities corresponding to src_cpu or
2234 * dest_cpu are members of a throttled hierarchy when performing group
2235 * load-balance operations.
2237 static inline int throttled_lb_pair(struct task_group *tg,
2238 int src_cpu, int dest_cpu)
2240 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2242 src_cfs_rq = tg->cfs_rq[src_cpu];
2243 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2245 return throttled_hierarchy(src_cfs_rq) ||
2246 throttled_hierarchy(dest_cfs_rq);
2249 /* updated child weight may affect parent so we have to do this bottom up */
2250 static int tg_unthrottle_up(struct task_group *tg, void *data)
2252 struct rq *rq = data;
2253 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2255 cfs_rq->throttle_count--;
2257 if (!cfs_rq->throttle_count) {
2258 /* adjust cfs_rq_clock_task() */
2259 cfs_rq->throttled_clock_task_time += rq->clock_task -
2260 cfs_rq->throttled_clock_task;
2267 static int tg_throttle_down(struct task_group *tg, void *data)
2269 struct rq *rq = data;
2270 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2272 /* group is entering throttled state, stop time */
2273 if (!cfs_rq->throttle_count)
2274 cfs_rq->throttled_clock_task = rq->clock_task;
2275 cfs_rq->throttle_count++;
2280 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2282 struct rq *rq = rq_of(cfs_rq);
2283 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2284 struct sched_entity *se;
2285 long task_delta, dequeue = 1;
2287 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2289 /* freeze hierarchy runnable averages while throttled */
2291 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2294 task_delta = cfs_rq->h_nr_running;
2295 for_each_sched_entity(se) {
2296 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2297 /* throttled entity or throttle-on-deactivate */
2302 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2303 qcfs_rq->h_nr_running -= task_delta;
2305 if (qcfs_rq->load.weight)
2310 rq->nr_running -= task_delta;
2312 cfs_rq->throttled = 1;
2313 cfs_rq->throttled_clock = rq->clock;
2314 raw_spin_lock(&cfs_b->lock);
2315 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2316 raw_spin_unlock(&cfs_b->lock);
2319 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2321 struct rq *rq = rq_of(cfs_rq);
2322 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2323 struct sched_entity *se;
2327 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2329 cfs_rq->throttled = 0;
2330 raw_spin_lock(&cfs_b->lock);
2331 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2332 list_del_rcu(&cfs_rq->throttled_list);
2333 raw_spin_unlock(&cfs_b->lock);
2335 update_rq_clock(rq);
2336 /* update hierarchical throttle state */
2337 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2339 if (!cfs_rq->load.weight)
2342 task_delta = cfs_rq->h_nr_running;
2343 for_each_sched_entity(se) {
2347 cfs_rq = cfs_rq_of(se);
2349 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2350 cfs_rq->h_nr_running += task_delta;
2352 if (cfs_rq_throttled(cfs_rq))
2357 rq->nr_running += task_delta;
2359 /* determine whether we need to wake up potentially idle cpu */
2360 if (rq->curr == rq->idle && rq->cfs.nr_running)
2361 resched_task(rq->curr);
2364 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2365 u64 remaining, u64 expires)
2367 struct cfs_rq *cfs_rq;
2368 u64 runtime = remaining;
2371 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2373 struct rq *rq = rq_of(cfs_rq);
2375 raw_spin_lock(&rq->lock);
2376 if (!cfs_rq_throttled(cfs_rq))
2379 runtime = -cfs_rq->runtime_remaining + 1;
2380 if (runtime > remaining)
2381 runtime = remaining;
2382 remaining -= runtime;
2384 cfs_rq->runtime_remaining += runtime;
2385 cfs_rq->runtime_expires = expires;
2387 /* we check whether we're throttled above */
2388 if (cfs_rq->runtime_remaining > 0)
2389 unthrottle_cfs_rq(cfs_rq);
2392 raw_spin_unlock(&rq->lock);
2403 * Responsible for refilling a task_group's bandwidth and unthrottling its
2404 * cfs_rqs as appropriate. If there has been no activity within the last
2405 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2406 * used to track this state.
2408 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2410 u64 runtime, runtime_expires;
2411 int idle = 1, throttled;
2413 raw_spin_lock(&cfs_b->lock);
2414 /* no need to continue the timer with no bandwidth constraint */
2415 if (cfs_b->quota == RUNTIME_INF)
2418 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2419 /* idle depends on !throttled (for the case of a large deficit) */
2420 idle = cfs_b->idle && !throttled;
2421 cfs_b->nr_periods += overrun;
2423 /* if we're going inactive then everything else can be deferred */
2427 __refill_cfs_bandwidth_runtime(cfs_b);
2430 /* mark as potentially idle for the upcoming period */
2435 /* account preceding periods in which throttling occurred */
2436 cfs_b->nr_throttled += overrun;
2439 * There are throttled entities so we must first use the new bandwidth
2440 * to unthrottle them before making it generally available. This
2441 * ensures that all existing debts will be paid before a new cfs_rq is
2444 runtime = cfs_b->runtime;
2445 runtime_expires = cfs_b->runtime_expires;
2449 * This check is repeated as we are holding onto the new bandwidth
2450 * while we unthrottle. This can potentially race with an unthrottled
2451 * group trying to acquire new bandwidth from the global pool.
2453 while (throttled && runtime > 0) {
2454 raw_spin_unlock(&cfs_b->lock);
2455 /* we can't nest cfs_b->lock while distributing bandwidth */
2456 runtime = distribute_cfs_runtime(cfs_b, runtime,
2458 raw_spin_lock(&cfs_b->lock);
2460 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2463 /* return (any) remaining runtime */
2464 cfs_b->runtime = runtime;
2466 * While we are ensured activity in the period following an
2467 * unthrottle, this also covers the case in which the new bandwidth is
2468 * insufficient to cover the existing bandwidth deficit. (Forcing the
2469 * timer to remain active while there are any throttled entities.)
2474 cfs_b->timer_active = 0;
2475 raw_spin_unlock(&cfs_b->lock);
2480 /* a cfs_rq won't donate quota below this amount */
2481 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2482 /* minimum remaining period time to redistribute slack quota */
2483 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2484 /* how long we wait to gather additional slack before distributing */
2485 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2487 /* are we near the end of the current quota period? */
2488 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2490 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2493 /* if the call-back is running a quota refresh is already occurring */
2494 if (hrtimer_callback_running(refresh_timer))
2497 /* is a quota refresh about to occur? */
2498 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2499 if (remaining < min_expire)
2505 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2507 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2509 /* if there's a quota refresh soon don't bother with slack */
2510 if (runtime_refresh_within(cfs_b, min_left))
2513 start_bandwidth_timer(&cfs_b->slack_timer,
2514 ns_to_ktime(cfs_bandwidth_slack_period));
2517 /* we know any runtime found here is valid as update_curr() precedes return */
2518 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2520 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2521 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2523 if (slack_runtime <= 0)
2526 raw_spin_lock(&cfs_b->lock);
2527 if (cfs_b->quota != RUNTIME_INF &&
2528 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2529 cfs_b->runtime += slack_runtime;
2531 /* we are under rq->lock, defer unthrottling using a timer */
2532 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2533 !list_empty(&cfs_b->throttled_cfs_rq))
2534 start_cfs_slack_bandwidth(cfs_b);
2536 raw_spin_unlock(&cfs_b->lock);
2538 /* even if it's not valid for return we don't want to try again */
2539 cfs_rq->runtime_remaining -= slack_runtime;
2542 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2544 if (!cfs_bandwidth_used())
2547 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2550 __return_cfs_rq_runtime(cfs_rq);
2554 * This is done with a timer (instead of inline with bandwidth return) since
2555 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2557 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2559 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2562 /* confirm we're still not at a refresh boundary */
2563 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2566 raw_spin_lock(&cfs_b->lock);
2567 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2568 runtime = cfs_b->runtime;
2571 expires = cfs_b->runtime_expires;
2572 raw_spin_unlock(&cfs_b->lock);
2577 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2579 raw_spin_lock(&cfs_b->lock);
2580 if (expires == cfs_b->runtime_expires)
2581 cfs_b->runtime = runtime;
2582 raw_spin_unlock(&cfs_b->lock);
2586 * When a group wakes up we want to make sure that its quota is not already
2587 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2588 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2590 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2592 if (!cfs_bandwidth_used())
2595 /* an active group must be handled by the update_curr()->put() path */
2596 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2599 /* ensure the group is not already throttled */
2600 if (cfs_rq_throttled(cfs_rq))
2603 /* update runtime allocation */
2604 account_cfs_rq_runtime(cfs_rq, 0);
2605 if (cfs_rq->runtime_remaining <= 0)
2606 throttle_cfs_rq(cfs_rq);
2609 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2610 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2612 if (!cfs_bandwidth_used())
2615 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2619 * it's possible for a throttled entity to be forced into a running
2620 * state (e.g. set_curr_task), in this case we're finished.
2622 if (cfs_rq_throttled(cfs_rq))
2625 throttle_cfs_rq(cfs_rq);
2628 static inline u64 default_cfs_period(void);
2629 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2630 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2632 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2634 struct cfs_bandwidth *cfs_b =
2635 container_of(timer, struct cfs_bandwidth, slack_timer);
2636 do_sched_cfs_slack_timer(cfs_b);
2638 return HRTIMER_NORESTART;
2641 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2643 struct cfs_bandwidth *cfs_b =
2644 container_of(timer, struct cfs_bandwidth, period_timer);
2650 now = hrtimer_cb_get_time(timer);
2651 overrun = hrtimer_forward(timer, now, cfs_b->period);
2656 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2659 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2662 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2664 raw_spin_lock_init(&cfs_b->lock);
2666 cfs_b->quota = RUNTIME_INF;
2667 cfs_b->period = ns_to_ktime(default_cfs_period());
2669 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2670 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2671 cfs_b->period_timer.function = sched_cfs_period_timer;
2672 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2673 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2676 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2678 cfs_rq->runtime_enabled = 0;
2679 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2682 /* requires cfs_b->lock, may release to reprogram timer */
2683 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2686 * The timer may be active because we're trying to set a new bandwidth
2687 * period or because we're racing with the tear-down path
2688 * (timer_active==0 becomes visible before the hrtimer call-back
2689 * terminates). In either case we ensure that it's re-programmed
2691 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2692 raw_spin_unlock(&cfs_b->lock);
2693 /* ensure cfs_b->lock is available while we wait */
2694 hrtimer_cancel(&cfs_b->period_timer);
2696 raw_spin_lock(&cfs_b->lock);
2697 /* if someone else restarted the timer then we're done */
2698 if (cfs_b->timer_active)
2702 cfs_b->timer_active = 1;
2703 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2706 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2708 hrtimer_cancel(&cfs_b->period_timer);
2709 hrtimer_cancel(&cfs_b->slack_timer);
2712 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2714 struct cfs_rq *cfs_rq;
2716 for_each_leaf_cfs_rq(rq, cfs_rq) {
2717 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2719 if (!cfs_rq->runtime_enabled)
2723 * clock_task is not advancing so we just need to make sure
2724 * there's some valid quota amount
2726 cfs_rq->runtime_remaining = cfs_b->quota;
2727 if (cfs_rq_throttled(cfs_rq))
2728 unthrottle_cfs_rq(cfs_rq);
2732 #else /* CONFIG_CFS_BANDWIDTH */
2733 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2735 return rq_of(cfs_rq)->clock_task;
2738 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2739 unsigned long delta_exec) {}
2740 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2741 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2742 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2744 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2749 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2754 static inline int throttled_lb_pair(struct task_group *tg,
2755 int src_cpu, int dest_cpu)
2760 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2762 #ifdef CONFIG_FAIR_GROUP_SCHED
2763 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2766 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2770 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2771 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2773 #endif /* CONFIG_CFS_BANDWIDTH */
2775 /**************************************************
2776 * CFS operations on tasks:
2779 #ifdef CONFIG_SCHED_HRTICK
2780 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2782 struct sched_entity *se = &p->se;
2783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2785 WARN_ON(task_rq(p) != rq);
2787 if (cfs_rq->nr_running > 1) {
2788 u64 slice = sched_slice(cfs_rq, se);
2789 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2790 s64 delta = slice - ran;
2799 * Don't schedule slices shorter than 10000ns, that just
2800 * doesn't make sense. Rely on vruntime for fairness.
2803 delta = max_t(s64, 10000LL, delta);
2805 hrtick_start(rq, delta);
2810 * called from enqueue/dequeue and updates the hrtick when the
2811 * current task is from our class and nr_running is low enough
2814 static void hrtick_update(struct rq *rq)
2816 struct task_struct *curr = rq->curr;
2818 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2821 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2822 hrtick_start_fair(rq, curr);
2824 #else /* !CONFIG_SCHED_HRTICK */
2826 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2830 static inline void hrtick_update(struct rq *rq)
2836 * The enqueue_task method is called before nr_running is
2837 * increased. Here we update the fair scheduling stats and
2838 * then put the task into the rbtree:
2841 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2843 struct cfs_rq *cfs_rq;
2844 struct sched_entity *se = &p->se;
2846 for_each_sched_entity(se) {
2849 cfs_rq = cfs_rq_of(se);
2850 enqueue_entity(cfs_rq, se, flags);
2853 * end evaluation on encountering a throttled cfs_rq
2855 * note: in the case of encountering a throttled cfs_rq we will
2856 * post the final h_nr_running increment below.
2858 if (cfs_rq_throttled(cfs_rq))
2860 cfs_rq->h_nr_running++;
2862 flags = ENQUEUE_WAKEUP;
2865 for_each_sched_entity(se) {
2866 cfs_rq = cfs_rq_of(se);
2867 cfs_rq->h_nr_running++;
2869 if (cfs_rq_throttled(cfs_rq))
2872 update_cfs_shares(cfs_rq);
2873 update_entity_load_avg(se, 1);
2877 update_rq_runnable_avg(rq, rq->nr_running);
2883 static void set_next_buddy(struct sched_entity *se);
2886 * The dequeue_task method is called before nr_running is
2887 * decreased. We remove the task from the rbtree and
2888 * update the fair scheduling stats:
2890 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2892 struct cfs_rq *cfs_rq;
2893 struct sched_entity *se = &p->se;
2894 int task_sleep = flags & DEQUEUE_SLEEP;
2896 for_each_sched_entity(se) {
2897 cfs_rq = cfs_rq_of(se);
2898 dequeue_entity(cfs_rq, se, flags);
2901 * end evaluation on encountering a throttled cfs_rq
2903 * note: in the case of encountering a throttled cfs_rq we will
2904 * post the final h_nr_running decrement below.
2906 if (cfs_rq_throttled(cfs_rq))
2908 cfs_rq->h_nr_running--;
2910 /* Don't dequeue parent if it has other entities besides us */
2911 if (cfs_rq->load.weight) {
2913 * Bias pick_next to pick a task from this cfs_rq, as
2914 * p is sleeping when it is within its sched_slice.
2916 if (task_sleep && parent_entity(se))
2917 set_next_buddy(parent_entity(se));
2919 /* avoid re-evaluating load for this entity */
2920 se = parent_entity(se);
2923 flags |= DEQUEUE_SLEEP;
2926 for_each_sched_entity(se) {
2927 cfs_rq = cfs_rq_of(se);
2928 cfs_rq->h_nr_running--;
2930 if (cfs_rq_throttled(cfs_rq))
2933 update_cfs_shares(cfs_rq);
2934 update_entity_load_avg(se, 1);
2939 update_rq_runnable_avg(rq, 1);
2945 /* Used instead of source_load when we know the type == 0 */
2946 static unsigned long weighted_cpuload(const int cpu)
2948 return cpu_rq(cpu)->load.weight;
2952 * Return a low guess at the load of a migration-source cpu weighted
2953 * according to the scheduling class and "nice" value.
2955 * We want to under-estimate the load of migration sources, to
2956 * balance conservatively.
2958 static unsigned long source_load(int cpu, int type)
2960 struct rq *rq = cpu_rq(cpu);
2961 unsigned long total = weighted_cpuload(cpu);
2963 if (type == 0 || !sched_feat(LB_BIAS))
2966 return min(rq->cpu_load[type-1], total);
2970 * Return a high guess at the load of a migration-target cpu weighted
2971 * according to the scheduling class and "nice" value.
2973 static unsigned long target_load(int cpu, int type)
2975 struct rq *rq = cpu_rq(cpu);
2976 unsigned long total = weighted_cpuload(cpu);
2978 if (type == 0 || !sched_feat(LB_BIAS))
2981 return max(rq->cpu_load[type-1], total);
2984 static unsigned long power_of(int cpu)
2986 return cpu_rq(cpu)->cpu_power;
2989 static unsigned long cpu_avg_load_per_task(int cpu)
2991 struct rq *rq = cpu_rq(cpu);
2992 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2995 return rq->load.weight / nr_running;
3001 static void task_waking_fair(struct task_struct *p)
3003 struct sched_entity *se = &p->se;
3004 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3007 #ifndef CONFIG_64BIT
3008 u64 min_vruntime_copy;
3011 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3013 min_vruntime = cfs_rq->min_vruntime;
3014 } while (min_vruntime != min_vruntime_copy);
3016 min_vruntime = cfs_rq->min_vruntime;
3019 se->vruntime -= min_vruntime;
3022 #ifdef CONFIG_FAIR_GROUP_SCHED
3024 * effective_load() calculates the load change as seen from the root_task_group
3026 * Adding load to a group doesn't make a group heavier, but can cause movement
3027 * of group shares between cpus. Assuming the shares were perfectly aligned one
3028 * can calculate the shift in shares.
3030 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3031 * on this @cpu and results in a total addition (subtraction) of @wg to the
3032 * total group weight.
3034 * Given a runqueue weight distribution (rw_i) we can compute a shares
3035 * distribution (s_i) using:
3037 * s_i = rw_i / \Sum rw_j (1)
3039 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3040 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3041 * shares distribution (s_i):
3043 * rw_i = { 2, 4, 1, 0 }
3044 * s_i = { 2/7, 4/7, 1/7, 0 }
3046 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3047 * task used to run on and the CPU the waker is running on), we need to
3048 * compute the effect of waking a task on either CPU and, in case of a sync
3049 * wakeup, compute the effect of the current task going to sleep.
3051 * So for a change of @wl to the local @cpu with an overall group weight change
3052 * of @wl we can compute the new shares distribution (s'_i) using:
3054 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3056 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3057 * differences in waking a task to CPU 0. The additional task changes the
3058 * weight and shares distributions like:
3060 * rw'_i = { 3, 4, 1, 0 }
3061 * s'_i = { 3/8, 4/8, 1/8, 0 }
3063 * We can then compute the difference in effective weight by using:
3065 * dw_i = S * (s'_i - s_i) (3)
3067 * Where 'S' is the group weight as seen by its parent.
3069 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3070 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3071 * 4/7) times the weight of the group.
3073 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3075 struct sched_entity *se = tg->se[cpu];
3077 if (!tg->parent) /* the trivial, non-cgroup case */
3080 for_each_sched_entity(se) {
3086 * W = @wg + \Sum rw_j
3088 W = wg + calc_tg_weight(tg, se->my_q);
3093 w = se->my_q->load.weight + wl;
3096 * wl = S * s'_i; see (2)
3099 wl = (w * tg->shares) / W;
3104 * Per the above, wl is the new se->load.weight value; since
3105 * those are clipped to [MIN_SHARES, ...) do so now. See
3106 * calc_cfs_shares().
3108 if (wl < MIN_SHARES)
3112 * wl = dw_i = S * (s'_i - s_i); see (3)
3114 wl -= se->load.weight;
3117 * Recursively apply this logic to all parent groups to compute
3118 * the final effective load change on the root group. Since
3119 * only the @tg group gets extra weight, all parent groups can
3120 * only redistribute existing shares. @wl is the shift in shares
3121 * resulting from this level per the above.
3130 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3131 unsigned long wl, unsigned long wg)
3138 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3140 s64 this_load, load;
3141 int idx, this_cpu, prev_cpu;
3142 unsigned long tl_per_task;
3143 struct task_group *tg;
3144 unsigned long weight;
3148 this_cpu = smp_processor_id();
3149 prev_cpu = task_cpu(p);
3150 load = source_load(prev_cpu, idx);
3151 this_load = target_load(this_cpu, idx);
3154 * If sync wakeup then subtract the (maximum possible)
3155 * effect of the currently running task from the load
3156 * of the current CPU:
3159 tg = task_group(current);
3160 weight = current->se.load.weight;
3162 this_load += effective_load(tg, this_cpu, -weight, -weight);
3163 load += effective_load(tg, prev_cpu, 0, -weight);
3167 weight = p->se.load.weight;
3170 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3171 * due to the sync cause above having dropped this_load to 0, we'll
3172 * always have an imbalance, but there's really nothing you can do
3173 * about that, so that's good too.
3175 * Otherwise check if either cpus are near enough in load to allow this
3176 * task to be woken on this_cpu.
3178 if (this_load > 0) {
3179 s64 this_eff_load, prev_eff_load;
3181 this_eff_load = 100;
3182 this_eff_load *= power_of(prev_cpu);
3183 this_eff_load *= this_load +
3184 effective_load(tg, this_cpu, weight, weight);
3186 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3187 prev_eff_load *= power_of(this_cpu);
3188 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3190 balanced = this_eff_load <= prev_eff_load;
3195 * If the currently running task will sleep within
3196 * a reasonable amount of time then attract this newly
3199 if (sync && balanced)
3202 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3203 tl_per_task = cpu_avg_load_per_task(this_cpu);
3206 (this_load <= load &&
3207 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3209 * This domain has SD_WAKE_AFFINE and
3210 * p is cache cold in this domain, and
3211 * there is no bad imbalance.
3213 schedstat_inc(sd, ttwu_move_affine);
3214 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3222 * find_idlest_group finds and returns the least busy CPU group within the
3225 static struct sched_group *
3226 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3227 int this_cpu, int load_idx)
3229 struct sched_group *idlest = NULL, *group = sd->groups;
3230 unsigned long min_load = ULONG_MAX, this_load = 0;
3231 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3234 unsigned long load, avg_load;
3238 /* Skip over this group if it has no CPUs allowed */
3239 if (!cpumask_intersects(sched_group_cpus(group),
3240 tsk_cpus_allowed(p)))
3243 local_group = cpumask_test_cpu(this_cpu,
3244 sched_group_cpus(group));
3246 /* Tally up the load of all CPUs in the group */
3249 for_each_cpu(i, sched_group_cpus(group)) {
3250 /* Bias balancing toward cpus of our domain */
3252 load = source_load(i, load_idx);
3254 load = target_load(i, load_idx);
3259 /* Adjust by relative CPU power of the group */
3260 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3263 this_load = avg_load;
3264 } else if (avg_load < min_load) {
3265 min_load = avg_load;
3268 } while (group = group->next, group != sd->groups);
3270 if (!idlest || 100*this_load < imbalance*min_load)
3276 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3279 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3281 unsigned long load, min_load = ULONG_MAX;
3285 /* Traverse only the allowed CPUs */
3286 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3287 load = weighted_cpuload(i);
3289 if (load < min_load || (load == min_load && i == this_cpu)) {
3299 * Try and locate an idle CPU in the sched_domain.
3301 static int select_idle_sibling(struct task_struct *p, int target)
3303 struct sched_domain *sd;
3304 struct sched_group *sg;
3305 int i = task_cpu(p);
3307 if (idle_cpu(target))
3311 * If the prevous cpu is cache affine and idle, don't be stupid.
3313 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3317 * Otherwise, iterate the domains and find an elegible idle cpu.
3319 sd = rcu_dereference(per_cpu(sd_llc, target));
3320 for_each_lower_domain(sd) {
3323 if (!cpumask_intersects(sched_group_cpus(sg),
3324 tsk_cpus_allowed(p)))
3327 for_each_cpu(i, sched_group_cpus(sg)) {
3328 if (i == target || !idle_cpu(i))
3332 target = cpumask_first_and(sched_group_cpus(sg),
3333 tsk_cpus_allowed(p));
3337 } while (sg != sd->groups);
3343 #ifdef CONFIG_SCHED_HMP
3345 * Heterogenous multiprocessor (HMP) optimizations
3347 * The cpu types are distinguished using a list of hmp_domains
3348 * which each represent one cpu type using a cpumask.
3349 * The list is assumed ordered by compute capacity with the
3350 * fastest domain first.
3352 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3354 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3356 /* Setup hmp_domains */
3357 static int __init hmp_cpu_mask_setup(void)
3360 struct hmp_domain *domain;
3361 struct list_head *pos;
3364 pr_debug("Initializing HMP scheduler:\n");
3366 /* Initialize hmp_domains using platform code */
3367 arch_get_hmp_domains(&hmp_domains);
3368 if (list_empty(&hmp_domains)) {
3369 pr_debug("HMP domain list is empty!\n");
3373 /* Print hmp_domains */
3375 list_for_each(pos, &hmp_domains) {
3376 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3377 cpulist_scnprintf(buf, 64, &domain->cpus);
3378 pr_debug(" HMP domain %d: %s\n", dc, buf);
3380 for_each_cpu_mask(cpu, domain->cpus) {
3381 per_cpu(hmp_cpu_domain, cpu) = domain;
3390 * Migration thresholds should be in the range [0..1023]
3391 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3392 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3393 * The default values (512, 256) offer good responsiveness, but may need
3394 * tweaking suit particular needs.
3396 unsigned int hmp_up_threshold = 512;
3397 unsigned int hmp_down_threshold = 256;
3399 static unsigned int hmp_up_migration(int cpu, struct sched_entity *se);
3400 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3402 /* Check if cpu is in fastest hmp_domain */
3403 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3405 struct list_head *pos;
3407 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3408 return pos == hmp_domains.next;
3411 /* Check if cpu is in slowest hmp_domain */
3412 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3414 struct list_head *pos;
3416 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3417 return list_is_last(pos, &hmp_domains);
3420 /* Next (slower) hmp_domain relative to cpu */
3421 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3423 struct list_head *pos;
3425 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3426 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3429 /* Previous (faster) hmp_domain relative to cpu */
3430 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3432 struct list_head *pos;
3434 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3435 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3439 * Selects a cpu in previous (faster) hmp_domain
3440 * Note that cpumask_any_and() returns the first cpu in the cpumask
3442 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3445 return cpumask_any_and(&hmp_faster_domain(cpu)->cpus,
3446 tsk_cpus_allowed(tsk));
3450 * Selects a cpu in next (slower) hmp_domain
3451 * Note that cpumask_any_and() returns the first cpu in the cpumask
3453 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3456 return cpumask_any_and(&hmp_slower_domain(cpu)->cpus,
3457 tsk_cpus_allowed(tsk));
3460 #endif /* CONFIG_SCHED_HMP */
3463 * sched_balance_self: balance the current task (running on cpu) in domains
3464 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3467 * Balance, ie. select the least loaded group.
3469 * Returns the target CPU number, or the same CPU if no balancing is needed.
3471 * preempt must be disabled.
3474 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3476 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3477 int cpu = smp_processor_id();
3478 int prev_cpu = task_cpu(p);
3480 int want_affine = 0;
3481 int sync = wake_flags & WF_SYNC;
3483 if (p->nr_cpus_allowed == 1)
3486 if (sd_flag & SD_BALANCE_WAKE) {
3487 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3493 for_each_domain(cpu, tmp) {
3494 if (!(tmp->flags & SD_LOAD_BALANCE))
3498 * If both cpu and prev_cpu are part of this domain,
3499 * cpu is a valid SD_WAKE_AFFINE target.
3501 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3502 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3507 if (tmp->flags & sd_flag)
3512 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3515 new_cpu = select_idle_sibling(p, prev_cpu);
3520 int load_idx = sd->forkexec_idx;
3521 struct sched_group *group;
3524 if (!(sd->flags & sd_flag)) {
3529 if (sd_flag & SD_BALANCE_WAKE)
3530 load_idx = sd->wake_idx;
3532 group = find_idlest_group(sd, p, cpu, load_idx);
3538 new_cpu = find_idlest_cpu(group, p, cpu);
3539 if (new_cpu == -1 || new_cpu == cpu) {
3540 /* Now try balancing at a lower domain level of cpu */
3545 /* Now try balancing at a lower domain level of new_cpu */
3547 weight = sd->span_weight;
3549 for_each_domain(cpu, tmp) {
3550 if (weight <= tmp->span_weight)
3552 if (tmp->flags & sd_flag)
3555 /* while loop will break here if sd == NULL */
3560 #ifdef CONFIG_SCHED_HMP
3561 if (hmp_up_migration(prev_cpu, &p->se))
3562 return hmp_select_faster_cpu(p, prev_cpu);
3563 if (hmp_down_migration(prev_cpu, &p->se))
3564 return hmp_select_slower_cpu(p, prev_cpu);
3565 /* Make sure that the task stays in its previous hmp domain */
3566 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
3574 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3575 * removed when useful for applications beyond shares distribution (e.g.
3578 #ifdef CONFIG_FAIR_GROUP_SCHED
3580 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3581 * cfs_rq_of(p) references at time of call are still valid and identify the
3582 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3583 * other assumptions, including the state of rq->lock, should be made.
3586 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3588 struct sched_entity *se = &p->se;
3589 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3592 * Load tracking: accumulate removed load so that it can be processed
3593 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3594 * to blocked load iff they have a positive decay-count. It can never
3595 * be negative here since on-rq tasks have decay-count == 0.
3597 if (se->avg.decay_count) {
3598 se->avg.decay_count = -__synchronize_entity_decay(se);
3599 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3603 #endif /* CONFIG_SMP */
3605 static unsigned long
3606 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3608 unsigned long gran = sysctl_sched_wakeup_granularity;
3611 * Since its curr running now, convert the gran from real-time
3612 * to virtual-time in his units.
3614 * By using 'se' instead of 'curr' we penalize light tasks, so
3615 * they get preempted easier. That is, if 'se' < 'curr' then
3616 * the resulting gran will be larger, therefore penalizing the
3617 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3618 * be smaller, again penalizing the lighter task.
3620 * This is especially important for buddies when the leftmost
3621 * task is higher priority than the buddy.
3623 return calc_delta_fair(gran, se);
3627 * Should 'se' preempt 'curr'.
3641 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3643 s64 gran, vdiff = curr->vruntime - se->vruntime;
3648 gran = wakeup_gran(curr, se);
3655 static void set_last_buddy(struct sched_entity *se)
3657 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3660 for_each_sched_entity(se)
3661 cfs_rq_of(se)->last = se;
3664 static void set_next_buddy(struct sched_entity *se)
3666 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3669 for_each_sched_entity(se)
3670 cfs_rq_of(se)->next = se;
3673 static void set_skip_buddy(struct sched_entity *se)
3675 for_each_sched_entity(se)
3676 cfs_rq_of(se)->skip = se;
3680 * Preempt the current task with a newly woken task if needed:
3682 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3684 struct task_struct *curr = rq->curr;
3685 struct sched_entity *se = &curr->se, *pse = &p->se;
3686 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3687 int scale = cfs_rq->nr_running >= sched_nr_latency;
3688 int next_buddy_marked = 0;
3690 if (unlikely(se == pse))
3694 * This is possible from callers such as move_task(), in which we
3695 * unconditionally check_prempt_curr() after an enqueue (which may have
3696 * lead to a throttle). This both saves work and prevents false
3697 * next-buddy nomination below.
3699 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3702 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3703 set_next_buddy(pse);
3704 next_buddy_marked = 1;
3708 * We can come here with TIF_NEED_RESCHED already set from new task
3711 * Note: this also catches the edge-case of curr being in a throttled
3712 * group (e.g. via set_curr_task), since update_curr() (in the
3713 * enqueue of curr) will have resulted in resched being set. This
3714 * prevents us from potentially nominating it as a false LAST_BUDDY
3717 if (test_tsk_need_resched(curr))
3720 /* Idle tasks are by definition preempted by non-idle tasks. */
3721 if (unlikely(curr->policy == SCHED_IDLE) &&
3722 likely(p->policy != SCHED_IDLE))
3726 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3727 * is driven by the tick):
3729 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3732 find_matching_se(&se, &pse);
3733 update_curr(cfs_rq_of(se));
3735 if (wakeup_preempt_entity(se, pse) == 1) {
3737 * Bias pick_next to pick the sched entity that is
3738 * triggering this preemption.
3740 if (!next_buddy_marked)
3741 set_next_buddy(pse);
3750 * Only set the backward buddy when the current task is still
3751 * on the rq. This can happen when a wakeup gets interleaved
3752 * with schedule on the ->pre_schedule() or idle_balance()
3753 * point, either of which can * drop the rq lock.
3755 * Also, during early boot the idle thread is in the fair class,
3756 * for obvious reasons its a bad idea to schedule back to it.
3758 if (unlikely(!se->on_rq || curr == rq->idle))
3761 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3765 static struct task_struct *pick_next_task_fair(struct rq *rq)
3767 struct task_struct *p;
3768 struct cfs_rq *cfs_rq = &rq->cfs;
3769 struct sched_entity *se;
3771 if (!cfs_rq->nr_running)
3775 se = pick_next_entity(cfs_rq);
3776 set_next_entity(cfs_rq, se);
3777 cfs_rq = group_cfs_rq(se);
3781 if (hrtick_enabled(rq))
3782 hrtick_start_fair(rq, p);
3788 * Account for a descheduled task:
3790 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3792 struct sched_entity *se = &prev->se;
3793 struct cfs_rq *cfs_rq;
3795 for_each_sched_entity(se) {
3796 cfs_rq = cfs_rq_of(se);
3797 put_prev_entity(cfs_rq, se);
3802 * sched_yield() is very simple
3804 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3806 static void yield_task_fair(struct rq *rq)
3808 struct task_struct *curr = rq->curr;
3809 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3810 struct sched_entity *se = &curr->se;
3813 * Are we the only task in the tree?
3815 if (unlikely(rq->nr_running == 1))
3818 clear_buddies(cfs_rq, se);
3820 if (curr->policy != SCHED_BATCH) {
3821 update_rq_clock(rq);
3823 * Update run-time statistics of the 'current'.
3825 update_curr(cfs_rq);
3827 * Tell update_rq_clock() that we've just updated,
3828 * so we don't do microscopic update in schedule()
3829 * and double the fastpath cost.
3831 rq->skip_clock_update = 1;
3837 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3839 struct sched_entity *se = &p->se;
3841 /* throttled hierarchies are not runnable */
3842 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3845 /* Tell the scheduler that we'd really like pse to run next. */
3848 yield_task_fair(rq);
3854 /**************************************************
3855 * Fair scheduling class load-balancing methods.
3859 * The purpose of load-balancing is to achieve the same basic fairness the
3860 * per-cpu scheduler provides, namely provide a proportional amount of compute
3861 * time to each task. This is expressed in the following equation:
3863 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3865 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3866 * W_i,0 is defined as:
3868 * W_i,0 = \Sum_j w_i,j (2)
3870 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3871 * is derived from the nice value as per prio_to_weight[].
3873 * The weight average is an exponential decay average of the instantaneous
3876 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3878 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3879 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3880 * can also include other factors [XXX].
3882 * To achieve this balance we define a measure of imbalance which follows
3883 * directly from (1):
3885 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3887 * We them move tasks around to minimize the imbalance. In the continuous
3888 * function space it is obvious this converges, in the discrete case we get
3889 * a few fun cases generally called infeasible weight scenarios.
3892 * - infeasible weights;
3893 * - local vs global optima in the discrete case. ]
3898 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3899 * for all i,j solution, we create a tree of cpus that follows the hardware
3900 * topology where each level pairs two lower groups (or better). This results
3901 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3902 * tree to only the first of the previous level and we decrease the frequency
3903 * of load-balance at each level inv. proportional to the number of cpus in
3909 * \Sum { --- * --- * 2^i } = O(n) (5)
3911 * `- size of each group
3912 * | | `- number of cpus doing load-balance
3914 * `- sum over all levels
3916 * Coupled with a limit on how many tasks we can migrate every balance pass,
3917 * this makes (5) the runtime complexity of the balancer.
3919 * An important property here is that each CPU is still (indirectly) connected
3920 * to every other cpu in at most O(log n) steps:
3922 * The adjacency matrix of the resulting graph is given by:
3925 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3928 * And you'll find that:
3930 * A^(log_2 n)_i,j != 0 for all i,j (7)
3932 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3933 * The task movement gives a factor of O(m), giving a convergence complexity
3936 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3941 * In order to avoid CPUs going idle while there's still work to do, new idle
3942 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3943 * tree itself instead of relying on other CPUs to bring it work.
3945 * This adds some complexity to both (5) and (8) but it reduces the total idle
3953 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3956 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3961 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3963 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3965 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3968 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3969 * rewrite all of this once again.]
3972 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3974 #define LBF_ALL_PINNED 0x01
3975 #define LBF_NEED_BREAK 0x02
3976 #define LBF_SOME_PINNED 0x04
3979 struct sched_domain *sd;
3987 struct cpumask *dst_grpmask;
3989 enum cpu_idle_type idle;
3991 /* The set of CPUs under consideration for load-balancing */
3992 struct cpumask *cpus;
3997 unsigned int loop_break;
3998 unsigned int loop_max;
4002 * move_task - move a task from one runqueue to another runqueue.
4003 * Both runqueues must be locked.
4005 static void move_task(struct task_struct *p, struct lb_env *env)
4007 deactivate_task(env->src_rq, p, 0);
4008 set_task_cpu(p, env->dst_cpu);
4009 activate_task(env->dst_rq, p, 0);
4010 check_preempt_curr(env->dst_rq, p, 0);
4014 * Is this task likely cache-hot:
4017 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4021 if (p->sched_class != &fair_sched_class)
4024 if (unlikely(p->policy == SCHED_IDLE))
4028 * Buddy candidates are cache hot:
4030 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4031 (&p->se == cfs_rq_of(&p->se)->next ||
4032 &p->se == cfs_rq_of(&p->se)->last))
4035 if (sysctl_sched_migration_cost == -1)
4037 if (sysctl_sched_migration_cost == 0)
4040 delta = now - p->se.exec_start;
4042 return delta < (s64)sysctl_sched_migration_cost;
4046 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4049 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4051 int tsk_cache_hot = 0;
4053 * We do not migrate tasks that are:
4054 * 1) throttled_lb_pair, or
4055 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4056 * 3) running (obviously), or
4057 * 4) are cache-hot on their current CPU.
4059 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4062 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4065 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4068 * Remember if this task can be migrated to any other cpu in
4069 * our sched_group. We may want to revisit it if we couldn't
4070 * meet load balance goals by pulling other tasks on src_cpu.
4072 * Also avoid computing new_dst_cpu if we have already computed
4073 * one in current iteration.
4075 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
4078 /* Prevent to re-select dst_cpu via env's cpus */
4079 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4080 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4081 env->flags |= LBF_SOME_PINNED;
4082 env->new_dst_cpu = cpu;
4090 /* Record that we found atleast one task that could run on dst_cpu */
4091 env->flags &= ~LBF_ALL_PINNED;
4093 if (task_running(env->src_rq, p)) {
4094 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4099 * Aggressive migration if:
4100 * 1) task is cache cold, or
4101 * 2) too many balance attempts have failed.
4103 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
4104 if (!tsk_cache_hot ||
4105 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4107 if (tsk_cache_hot) {
4108 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4109 schedstat_inc(p, se.statistics.nr_forced_migrations);
4115 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4120 * move_one_task tries to move exactly one task from busiest to this_rq, as
4121 * part of active balancing operations within "domain".
4122 * Returns 1 if successful and 0 otherwise.
4124 * Called with both runqueues locked.
4126 static int move_one_task(struct lb_env *env)
4128 struct task_struct *p, *n;
4130 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4131 if (!can_migrate_task(p, env))
4136 * Right now, this is only the second place move_task()
4137 * is called, so we can safely collect move_task()
4138 * stats here rather than inside move_task().
4140 schedstat_inc(env->sd, lb_gained[env->idle]);
4146 static unsigned long task_h_load(struct task_struct *p);
4148 static const unsigned int sched_nr_migrate_break = 32;
4151 * move_tasks tries to move up to imbalance weighted load from busiest to
4152 * this_rq, as part of a balancing operation within domain "sd".
4153 * Returns 1 if successful and 0 otherwise.
4155 * Called with both runqueues locked.
4157 static int move_tasks(struct lb_env *env)
4159 struct list_head *tasks = &env->src_rq->cfs_tasks;
4160 struct task_struct *p;
4164 if (env->imbalance <= 0)
4167 while (!list_empty(tasks)) {
4168 p = list_first_entry(tasks, struct task_struct, se.group_node);
4171 /* We've more or less seen every task there is, call it quits */
4172 if (env->loop > env->loop_max)
4175 /* take a breather every nr_migrate tasks */
4176 if (env->loop > env->loop_break) {
4177 env->loop_break += sched_nr_migrate_break;
4178 env->flags |= LBF_NEED_BREAK;
4182 if (!can_migrate_task(p, env))
4185 load = task_h_load(p);
4187 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4190 if ((load / 2) > env->imbalance)
4195 env->imbalance -= load;
4197 #ifdef CONFIG_PREEMPT
4199 * NEWIDLE balancing is a source of latency, so preemptible
4200 * kernels will stop after the first task is pulled to minimize
4201 * the critical section.
4203 if (env->idle == CPU_NEWLY_IDLE)
4208 * We only want to steal up to the prescribed amount of
4211 if (env->imbalance <= 0)
4216 list_move_tail(&p->se.group_node, tasks);
4220 * Right now, this is one of only two places move_task() is called,
4221 * so we can safely collect move_task() stats here rather than
4222 * inside move_task().
4224 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4229 #ifdef CONFIG_FAIR_GROUP_SCHED
4231 * update tg->load_weight by folding this cpu's load_avg
4233 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4235 struct sched_entity *se = tg->se[cpu];
4236 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4238 /* throttled entities do not contribute to load */
4239 if (throttled_hierarchy(cfs_rq))
4242 update_cfs_rq_blocked_load(cfs_rq, 1);
4245 update_entity_load_avg(se, 1);
4247 * We pivot on our runnable average having decayed to zero for
4248 * list removal. This generally implies that all our children
4249 * have also been removed (modulo rounding error or bandwidth
4250 * control); however, such cases are rare and we can fix these
4253 * TODO: fix up out-of-order children on enqueue.
4255 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4256 list_del_leaf_cfs_rq(cfs_rq);
4258 struct rq *rq = rq_of(cfs_rq);
4259 update_rq_runnable_avg(rq, rq->nr_running);
4263 static void update_blocked_averages(int cpu)
4265 struct rq *rq = cpu_rq(cpu);
4266 struct cfs_rq *cfs_rq;
4267 unsigned long flags;
4269 raw_spin_lock_irqsave(&rq->lock, flags);
4270 update_rq_clock(rq);
4272 * Iterates the task_group tree in a bottom up fashion, see
4273 * list_add_leaf_cfs_rq() for details.
4275 for_each_leaf_cfs_rq(rq, cfs_rq) {
4277 * Note: We may want to consider periodically releasing
4278 * rq->lock about these updates so that creating many task
4279 * groups does not result in continually extending hold time.
4281 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4284 raw_spin_unlock_irqrestore(&rq->lock, flags);
4288 * Compute the cpu's hierarchical load factor for each task group.
4289 * This needs to be done in a top-down fashion because the load of a child
4290 * group is a fraction of its parents load.
4292 static int tg_load_down(struct task_group *tg, void *data)
4295 long cpu = (long)data;
4298 load = cpu_rq(cpu)->load.weight;
4300 load = tg->parent->cfs_rq[cpu]->h_load;
4301 load *= tg->se[cpu]->load.weight;
4302 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
4305 tg->cfs_rq[cpu]->h_load = load;
4310 static void update_h_load(long cpu)
4312 struct rq *rq = cpu_rq(cpu);
4313 unsigned long now = jiffies;
4315 if (rq->h_load_throttle == now)
4318 rq->h_load_throttle = now;
4321 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4325 static unsigned long task_h_load(struct task_struct *p)
4327 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4330 load = p->se.load.weight;
4331 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
4336 static inline void update_blocked_averages(int cpu)
4340 static inline void update_h_load(long cpu)
4344 static unsigned long task_h_load(struct task_struct *p)
4346 return p->se.load.weight;
4350 /********** Helpers for find_busiest_group ************************/
4352 * sd_lb_stats - Structure to store the statistics of a sched_domain
4353 * during load balancing.
4355 struct sd_lb_stats {
4356 struct sched_group *busiest; /* Busiest group in this sd */
4357 struct sched_group *this; /* Local group in this sd */
4358 unsigned long total_load; /* Total load of all groups in sd */
4359 unsigned long total_pwr; /* Total power of all groups in sd */
4360 unsigned long avg_load; /* Average load across all groups in sd */
4362 /** Statistics of this group */
4363 unsigned long this_load;
4364 unsigned long this_load_per_task;
4365 unsigned long this_nr_running;
4366 unsigned long this_has_capacity;
4367 unsigned int this_idle_cpus;
4369 /* Statistics of the busiest group */
4370 unsigned int busiest_idle_cpus;
4371 unsigned long max_load;
4372 unsigned long busiest_load_per_task;
4373 unsigned long busiest_nr_running;
4374 unsigned long busiest_group_capacity;
4375 unsigned long busiest_has_capacity;
4376 unsigned int busiest_group_weight;
4378 int group_imb; /* Is there imbalance in this sd */
4382 * sg_lb_stats - stats of a sched_group required for load_balancing
4384 struct sg_lb_stats {
4385 unsigned long avg_load; /*Avg load across the CPUs of the group */
4386 unsigned long group_load; /* Total load over the CPUs of the group */
4387 unsigned long sum_nr_running; /* Nr tasks running in the group */
4388 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4389 unsigned long group_capacity;
4390 unsigned long idle_cpus;
4391 unsigned long group_weight;
4392 int group_imb; /* Is there an imbalance in the group ? */
4393 int group_has_capacity; /* Is there extra capacity in the group? */
4397 * get_sd_load_idx - Obtain the load index for a given sched domain.
4398 * @sd: The sched_domain whose load_idx is to be obtained.
4399 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4401 static inline int get_sd_load_idx(struct sched_domain *sd,
4402 enum cpu_idle_type idle)
4408 load_idx = sd->busy_idx;
4411 case CPU_NEWLY_IDLE:
4412 load_idx = sd->newidle_idx;
4415 load_idx = sd->idle_idx;
4422 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4424 return SCHED_POWER_SCALE;
4427 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4429 return default_scale_freq_power(sd, cpu);
4432 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4434 unsigned long weight = sd->span_weight;
4435 unsigned long smt_gain = sd->smt_gain;
4442 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4444 return default_scale_smt_power(sd, cpu);
4447 static unsigned long scale_rt_power(int cpu)
4449 struct rq *rq = cpu_rq(cpu);
4450 u64 total, available, age_stamp, avg;
4453 * Since we're reading these variables without serialization make sure
4454 * we read them once before doing sanity checks on them.
4456 age_stamp = ACCESS_ONCE(rq->age_stamp);
4457 avg = ACCESS_ONCE(rq->rt_avg);
4459 total = sched_avg_period() + (rq->clock - age_stamp);
4461 if (unlikely(total < avg)) {
4462 /* Ensures that power won't end up being negative */
4465 available = total - avg;
4468 if (unlikely((s64)total < SCHED_POWER_SCALE))
4469 total = SCHED_POWER_SCALE;
4471 total >>= SCHED_POWER_SHIFT;
4473 return div_u64(available, total);
4476 static void update_cpu_power(struct sched_domain *sd, int cpu)
4478 unsigned long weight = sd->span_weight;
4479 unsigned long power = SCHED_POWER_SCALE;
4480 struct sched_group *sdg = sd->groups;
4482 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4483 if (sched_feat(ARCH_POWER))
4484 power *= arch_scale_smt_power(sd, cpu);
4486 power *= default_scale_smt_power(sd, cpu);
4488 power >>= SCHED_POWER_SHIFT;
4491 sdg->sgp->power_orig = power;
4493 if (sched_feat(ARCH_POWER))
4494 power *= arch_scale_freq_power(sd, cpu);
4496 power *= default_scale_freq_power(sd, cpu);
4498 power >>= SCHED_POWER_SHIFT;
4500 power *= scale_rt_power(cpu);
4501 power >>= SCHED_POWER_SHIFT;
4506 cpu_rq(cpu)->cpu_power = power;
4507 sdg->sgp->power = power;
4510 void update_group_power(struct sched_domain *sd, int cpu)
4512 struct sched_domain *child = sd->child;
4513 struct sched_group *group, *sdg = sd->groups;
4514 unsigned long power;
4515 unsigned long interval;
4517 interval = msecs_to_jiffies(sd->balance_interval);
4518 interval = clamp(interval, 1UL, max_load_balance_interval);
4519 sdg->sgp->next_update = jiffies + interval;
4522 update_cpu_power(sd, cpu);
4528 if (child->flags & SD_OVERLAP) {
4530 * SD_OVERLAP domains cannot assume that child groups
4531 * span the current group.
4534 for_each_cpu(cpu, sched_group_cpus(sdg))
4535 power += power_of(cpu);
4538 * !SD_OVERLAP domains can assume that child groups
4539 * span the current group.
4542 group = child->groups;
4544 power += group->sgp->power;
4545 group = group->next;
4546 } while (group != child->groups);
4549 sdg->sgp->power_orig = sdg->sgp->power = power;
4553 * Try and fix up capacity for tiny siblings, this is needed when
4554 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4555 * which on its own isn't powerful enough.
4557 * See update_sd_pick_busiest() and check_asym_packing().
4560 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4563 * Only siblings can have significantly less than SCHED_POWER_SCALE
4565 if (!(sd->flags & SD_SHARE_CPUPOWER))
4569 * If ~90% of the cpu_power is still there, we're good.
4571 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4578 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4579 * @env: The load balancing environment.
4580 * @group: sched_group whose statistics are to be updated.
4581 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4582 * @local_group: Does group contain this_cpu.
4583 * @balance: Should we balance.
4584 * @sgs: variable to hold the statistics for this group.
4586 static inline void update_sg_lb_stats(struct lb_env *env,
4587 struct sched_group *group, int load_idx,
4588 int local_group, int *balance, struct sg_lb_stats *sgs)
4590 unsigned long nr_running, max_nr_running, min_nr_running;
4591 unsigned long load, max_cpu_load, min_cpu_load;
4592 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4593 unsigned long avg_load_per_task = 0;
4597 balance_cpu = group_balance_cpu(group);
4599 /* Tally up the load of all CPUs in the group */
4601 min_cpu_load = ~0UL;
4603 min_nr_running = ~0UL;
4605 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4606 struct rq *rq = cpu_rq(i);
4608 nr_running = rq->nr_running;
4610 /* Bias balancing toward cpus of our domain */
4612 if (idle_cpu(i) && !first_idle_cpu &&
4613 cpumask_test_cpu(i, sched_group_mask(group))) {
4618 load = target_load(i, load_idx);
4620 load = source_load(i, load_idx);
4621 if (load > max_cpu_load)
4622 max_cpu_load = load;
4623 if (min_cpu_load > load)
4624 min_cpu_load = load;
4626 if (nr_running > max_nr_running)
4627 max_nr_running = nr_running;
4628 if (min_nr_running > nr_running)
4629 min_nr_running = nr_running;
4632 sgs->group_load += load;
4633 sgs->sum_nr_running += nr_running;
4634 sgs->sum_weighted_load += weighted_cpuload(i);
4640 * First idle cpu or the first cpu(busiest) in this sched group
4641 * is eligible for doing load balancing at this and above
4642 * domains. In the newly idle case, we will allow all the cpu's
4643 * to do the newly idle load balance.
4646 if (env->idle != CPU_NEWLY_IDLE) {
4647 if (balance_cpu != env->dst_cpu) {
4651 update_group_power(env->sd, env->dst_cpu);
4652 } else if (time_after_eq(jiffies, group->sgp->next_update))
4653 update_group_power(env->sd, env->dst_cpu);
4656 /* Adjust by relative CPU power of the group */
4657 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4660 * Consider the group unbalanced when the imbalance is larger
4661 * than the average weight of a task.
4663 * APZ: with cgroup the avg task weight can vary wildly and
4664 * might not be a suitable number - should we keep a
4665 * normalized nr_running number somewhere that negates
4668 if (sgs->sum_nr_running)
4669 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4671 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4672 (max_nr_running - min_nr_running) > 1)
4675 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4677 if (!sgs->group_capacity)
4678 sgs->group_capacity = fix_small_capacity(env->sd, group);
4679 sgs->group_weight = group->group_weight;
4681 if (sgs->group_capacity > sgs->sum_nr_running)
4682 sgs->group_has_capacity = 1;
4686 * update_sd_pick_busiest - return 1 on busiest group
4687 * @env: The load balancing environment.
4688 * @sds: sched_domain statistics
4689 * @sg: sched_group candidate to be checked for being the busiest
4690 * @sgs: sched_group statistics
4692 * Determine if @sg is a busier group than the previously selected
4695 static bool update_sd_pick_busiest(struct lb_env *env,
4696 struct sd_lb_stats *sds,
4697 struct sched_group *sg,
4698 struct sg_lb_stats *sgs)
4700 if (sgs->avg_load <= sds->max_load)
4703 if (sgs->sum_nr_running > sgs->group_capacity)
4710 * ASYM_PACKING needs to move all the work to the lowest
4711 * numbered CPUs in the group, therefore mark all groups
4712 * higher than ourself as busy.
4714 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4715 env->dst_cpu < group_first_cpu(sg)) {
4719 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4727 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4728 * @env: The load balancing environment.
4729 * @balance: Should we balance.
4730 * @sds: variable to hold the statistics for this sched_domain.
4732 static inline void update_sd_lb_stats(struct lb_env *env,
4733 int *balance, struct sd_lb_stats *sds)
4735 struct sched_domain *child = env->sd->child;
4736 struct sched_group *sg = env->sd->groups;
4737 struct sg_lb_stats sgs;
4738 int load_idx, prefer_sibling = 0;
4740 if (child && child->flags & SD_PREFER_SIBLING)
4743 load_idx = get_sd_load_idx(env->sd, env->idle);
4748 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4749 memset(&sgs, 0, sizeof(sgs));
4750 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4752 if (local_group && !(*balance))
4755 sds->total_load += sgs.group_load;
4756 sds->total_pwr += sg->sgp->power;
4759 * In case the child domain prefers tasks go to siblings
4760 * first, lower the sg capacity to one so that we'll try
4761 * and move all the excess tasks away. We lower the capacity
4762 * of a group only if the local group has the capacity to fit
4763 * these excess tasks, i.e. nr_running < group_capacity. The
4764 * extra check prevents the case where you always pull from the
4765 * heaviest group when it is already under-utilized (possible
4766 * with a large weight task outweighs the tasks on the system).
4768 if (prefer_sibling && !local_group && sds->this_has_capacity)
4769 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4772 sds->this_load = sgs.avg_load;
4774 sds->this_nr_running = sgs.sum_nr_running;
4775 sds->this_load_per_task = sgs.sum_weighted_load;
4776 sds->this_has_capacity = sgs.group_has_capacity;
4777 sds->this_idle_cpus = sgs.idle_cpus;
4778 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4779 sds->max_load = sgs.avg_load;
4781 sds->busiest_nr_running = sgs.sum_nr_running;
4782 sds->busiest_idle_cpus = sgs.idle_cpus;
4783 sds->busiest_group_capacity = sgs.group_capacity;
4784 sds->busiest_load_per_task = sgs.sum_weighted_load;
4785 sds->busiest_has_capacity = sgs.group_has_capacity;
4786 sds->busiest_group_weight = sgs.group_weight;
4787 sds->group_imb = sgs.group_imb;
4791 } while (sg != env->sd->groups);
4795 * check_asym_packing - Check to see if the group is packed into the
4798 * This is primarily intended to used at the sibling level. Some
4799 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4800 * case of POWER7, it can move to lower SMT modes only when higher
4801 * threads are idle. When in lower SMT modes, the threads will
4802 * perform better since they share less core resources. Hence when we
4803 * have idle threads, we want them to be the higher ones.
4805 * This packing function is run on idle threads. It checks to see if
4806 * the busiest CPU in this domain (core in the P7 case) has a higher
4807 * CPU number than the packing function is being run on. Here we are
4808 * assuming lower CPU number will be equivalent to lower a SMT thread
4811 * Returns 1 when packing is required and a task should be moved to
4812 * this CPU. The amount of the imbalance is returned in *imbalance.
4814 * @env: The load balancing environment.
4815 * @sds: Statistics of the sched_domain which is to be packed
4817 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4821 if (!(env->sd->flags & SD_ASYM_PACKING))
4827 busiest_cpu = group_first_cpu(sds->busiest);
4828 if (env->dst_cpu > busiest_cpu)
4831 env->imbalance = DIV_ROUND_CLOSEST(
4832 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4838 * fix_small_imbalance - Calculate the minor imbalance that exists
4839 * amongst the groups of a sched_domain, during
4841 * @env: The load balancing environment.
4842 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4845 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4847 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4848 unsigned int imbn = 2;
4849 unsigned long scaled_busy_load_per_task;
4851 if (sds->this_nr_running) {
4852 sds->this_load_per_task /= sds->this_nr_running;
4853 if (sds->busiest_load_per_task >
4854 sds->this_load_per_task)
4857 sds->this_load_per_task =
4858 cpu_avg_load_per_task(env->dst_cpu);
4861 scaled_busy_load_per_task = sds->busiest_load_per_task
4862 * SCHED_POWER_SCALE;
4863 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4865 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4866 (scaled_busy_load_per_task * imbn)) {
4867 env->imbalance = sds->busiest_load_per_task;
4872 * OK, we don't have enough imbalance to justify moving tasks,
4873 * however we may be able to increase total CPU power used by
4877 pwr_now += sds->busiest->sgp->power *
4878 min(sds->busiest_load_per_task, sds->max_load);
4879 pwr_now += sds->this->sgp->power *
4880 min(sds->this_load_per_task, sds->this_load);
4881 pwr_now /= SCHED_POWER_SCALE;
4883 /* Amount of load we'd subtract */
4884 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4885 sds->busiest->sgp->power;
4886 if (sds->max_load > tmp)
4887 pwr_move += sds->busiest->sgp->power *
4888 min(sds->busiest_load_per_task, sds->max_load - tmp);
4890 /* Amount of load we'd add */
4891 if (sds->max_load * sds->busiest->sgp->power <
4892 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4893 tmp = (sds->max_load * sds->busiest->sgp->power) /
4894 sds->this->sgp->power;
4896 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4897 sds->this->sgp->power;
4898 pwr_move += sds->this->sgp->power *
4899 min(sds->this_load_per_task, sds->this_load + tmp);
4900 pwr_move /= SCHED_POWER_SCALE;
4902 /* Move if we gain throughput */
4903 if (pwr_move > pwr_now)
4904 env->imbalance = sds->busiest_load_per_task;
4908 * calculate_imbalance - Calculate the amount of imbalance present within the
4909 * groups of a given sched_domain during load balance.
4910 * @env: load balance environment
4911 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4913 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4915 unsigned long max_pull, load_above_capacity = ~0UL;
4917 sds->busiest_load_per_task /= sds->busiest_nr_running;
4918 if (sds->group_imb) {
4919 sds->busiest_load_per_task =
4920 min(sds->busiest_load_per_task, sds->avg_load);
4924 * In the presence of smp nice balancing, certain scenarios can have
4925 * max load less than avg load(as we skip the groups at or below
4926 * its cpu_power, while calculating max_load..)
4928 if (sds->max_load < sds->avg_load) {
4930 return fix_small_imbalance(env, sds);
4933 if (!sds->group_imb) {
4935 * Don't want to pull so many tasks that a group would go idle.
4937 load_above_capacity = (sds->busiest_nr_running -
4938 sds->busiest_group_capacity);
4940 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4942 load_above_capacity /= sds->busiest->sgp->power;
4946 * We're trying to get all the cpus to the average_load, so we don't
4947 * want to push ourselves above the average load, nor do we wish to
4948 * reduce the max loaded cpu below the average load. At the same time,
4949 * we also don't want to reduce the group load below the group capacity
4950 * (so that we can implement power-savings policies etc). Thus we look
4951 * for the minimum possible imbalance.
4952 * Be careful of negative numbers as they'll appear as very large values
4953 * with unsigned longs.
4955 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4957 /* How much load to actually move to equalise the imbalance */
4958 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4959 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4960 / SCHED_POWER_SCALE;
4963 * if *imbalance is less than the average load per runnable task
4964 * there is no guarantee that any tasks will be moved so we'll have
4965 * a think about bumping its value to force at least one task to be
4968 if (env->imbalance < sds->busiest_load_per_task)
4969 return fix_small_imbalance(env, sds);
4973 /******* find_busiest_group() helpers end here *********************/
4976 * find_busiest_group - Returns the busiest group within the sched_domain
4977 * if there is an imbalance. If there isn't an imbalance, and
4978 * the user has opted for power-savings, it returns a group whose
4979 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4980 * such a group exists.
4982 * Also calculates the amount of weighted load which should be moved
4983 * to restore balance.
4985 * @env: The load balancing environment.
4986 * @balance: Pointer to a variable indicating if this_cpu
4987 * is the appropriate cpu to perform load balancing at this_level.
4989 * Returns: - the busiest group if imbalance exists.
4990 * - If no imbalance and user has opted for power-savings balance,
4991 * return the least loaded group whose CPUs can be
4992 * put to idle by rebalancing its tasks onto our group.
4994 static struct sched_group *
4995 find_busiest_group(struct lb_env *env, int *balance)
4997 struct sd_lb_stats sds;
4999 memset(&sds, 0, sizeof(sds));
5002 * Compute the various statistics relavent for load balancing at
5005 update_sd_lb_stats(env, balance, &sds);
5008 * this_cpu is not the appropriate cpu to perform load balancing at
5014 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5015 check_asym_packing(env, &sds))
5018 /* There is no busy sibling group to pull tasks from */
5019 if (!sds.busiest || sds.busiest_nr_running == 0)
5022 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5025 * If the busiest group is imbalanced the below checks don't
5026 * work because they assumes all things are equal, which typically
5027 * isn't true due to cpus_allowed constraints and the like.
5032 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5033 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5034 !sds.busiest_has_capacity)
5038 * If the local group is more busy than the selected busiest group
5039 * don't try and pull any tasks.
5041 if (sds.this_load >= sds.max_load)
5045 * Don't pull any tasks if this group is already above the domain
5048 if (sds.this_load >= sds.avg_load)
5051 if (env->idle == CPU_IDLE) {
5053 * This cpu is idle. If the busiest group load doesn't
5054 * have more tasks than the number of available cpu's and
5055 * there is no imbalance between this and busiest group
5056 * wrt to idle cpu's, it is balanced.
5058 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
5059 sds.busiest_nr_running <= sds.busiest_group_weight)
5063 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5064 * imbalance_pct to be conservative.
5066 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
5071 /* Looks like there is an imbalance. Compute it */
5072 calculate_imbalance(env, &sds);
5082 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5084 static struct rq *find_busiest_queue(struct lb_env *env,
5085 struct sched_group *group)
5087 struct rq *busiest = NULL, *rq;
5088 unsigned long max_load = 0;
5091 for_each_cpu(i, sched_group_cpus(group)) {
5092 unsigned long power = power_of(i);
5093 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5098 capacity = fix_small_capacity(env->sd, group);
5100 if (!cpumask_test_cpu(i, env->cpus))
5104 wl = weighted_cpuload(i);
5107 * When comparing with imbalance, use weighted_cpuload()
5108 * which is not scaled with the cpu power.
5110 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5114 * For the load comparisons with the other cpu's, consider
5115 * the weighted_cpuload() scaled with the cpu power, so that
5116 * the load can be moved away from the cpu that is potentially
5117 * running at a lower capacity.
5119 wl = (wl * SCHED_POWER_SCALE) / power;
5121 if (wl > max_load) {
5131 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5132 * so long as it is large enough.
5134 #define MAX_PINNED_INTERVAL 512
5136 /* Working cpumask for load_balance and load_balance_newidle. */
5137 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5139 static int need_active_balance(struct lb_env *env)
5141 struct sched_domain *sd = env->sd;
5143 if (env->idle == CPU_NEWLY_IDLE) {
5146 * ASYM_PACKING needs to force migrate tasks from busy but
5147 * higher numbered CPUs in order to pack all tasks in the
5148 * lowest numbered CPUs.
5150 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5154 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5157 static int active_load_balance_cpu_stop(void *data);
5160 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5161 * tasks if there is an imbalance.
5163 static int load_balance(int this_cpu, struct rq *this_rq,
5164 struct sched_domain *sd, enum cpu_idle_type idle,
5167 int ld_moved, cur_ld_moved, active_balance = 0;
5168 struct sched_group *group;
5170 unsigned long flags;
5171 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5173 struct lb_env env = {
5175 .dst_cpu = this_cpu,
5177 .dst_grpmask = sched_group_cpus(sd->groups),
5179 .loop_break = sched_nr_migrate_break,
5184 * For NEWLY_IDLE load_balancing, we don't need to consider
5185 * other cpus in our group
5187 if (idle == CPU_NEWLY_IDLE)
5188 env.dst_grpmask = NULL;
5190 cpumask_copy(cpus, cpu_active_mask);
5192 schedstat_inc(sd, lb_count[idle]);
5195 group = find_busiest_group(&env, balance);
5201 schedstat_inc(sd, lb_nobusyg[idle]);
5205 busiest = find_busiest_queue(&env, group);
5207 schedstat_inc(sd, lb_nobusyq[idle]);
5211 BUG_ON(busiest == env.dst_rq);
5213 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5216 if (busiest->nr_running > 1) {
5218 * Attempt to move tasks. If find_busiest_group has found
5219 * an imbalance but busiest->nr_running <= 1, the group is
5220 * still unbalanced. ld_moved simply stays zero, so it is
5221 * correctly treated as an imbalance.
5223 env.flags |= LBF_ALL_PINNED;
5224 env.src_cpu = busiest->cpu;
5225 env.src_rq = busiest;
5226 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5228 update_h_load(env.src_cpu);
5230 local_irq_save(flags);
5231 double_rq_lock(env.dst_rq, busiest);
5234 * cur_ld_moved - load moved in current iteration
5235 * ld_moved - cumulative load moved across iterations
5237 cur_ld_moved = move_tasks(&env);
5238 ld_moved += cur_ld_moved;
5239 double_rq_unlock(env.dst_rq, busiest);
5240 local_irq_restore(flags);
5243 * some other cpu did the load balance for us.
5245 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5246 resched_cpu(env.dst_cpu);
5248 if (env.flags & LBF_NEED_BREAK) {
5249 env.flags &= ~LBF_NEED_BREAK;
5254 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5255 * us and move them to an alternate dst_cpu in our sched_group
5256 * where they can run. The upper limit on how many times we
5257 * iterate on same src_cpu is dependent on number of cpus in our
5260 * This changes load balance semantics a bit on who can move
5261 * load to a given_cpu. In addition to the given_cpu itself
5262 * (or a ilb_cpu acting on its behalf where given_cpu is
5263 * nohz-idle), we now have balance_cpu in a position to move
5264 * load to given_cpu. In rare situations, this may cause
5265 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5266 * _independently_ and at _same_ time to move some load to
5267 * given_cpu) causing exceess load to be moved to given_cpu.
5268 * This however should not happen so much in practice and
5269 * moreover subsequent load balance cycles should correct the
5270 * excess load moved.
5272 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5274 env.dst_rq = cpu_rq(env.new_dst_cpu);
5275 env.dst_cpu = env.new_dst_cpu;
5276 env.flags &= ~LBF_SOME_PINNED;
5278 env.loop_break = sched_nr_migrate_break;
5280 /* Prevent to re-select dst_cpu via env's cpus */
5281 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5284 * Go back to "more_balance" rather than "redo" since we
5285 * need to continue with same src_cpu.
5290 /* All tasks on this runqueue were pinned by CPU affinity */
5291 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5292 cpumask_clear_cpu(cpu_of(busiest), cpus);
5293 if (!cpumask_empty(cpus)) {
5295 env.loop_break = sched_nr_migrate_break;
5303 schedstat_inc(sd, lb_failed[idle]);
5305 * Increment the failure counter only on periodic balance.
5306 * We do not want newidle balance, which can be very
5307 * frequent, pollute the failure counter causing
5308 * excessive cache_hot migrations and active balances.
5310 if (idle != CPU_NEWLY_IDLE)
5311 sd->nr_balance_failed++;
5313 if (need_active_balance(&env)) {
5314 raw_spin_lock_irqsave(&busiest->lock, flags);
5316 /* don't kick the active_load_balance_cpu_stop,
5317 * if the curr task on busiest cpu can't be
5320 if (!cpumask_test_cpu(this_cpu,
5321 tsk_cpus_allowed(busiest->curr))) {
5322 raw_spin_unlock_irqrestore(&busiest->lock,
5324 env.flags |= LBF_ALL_PINNED;
5325 goto out_one_pinned;
5329 * ->active_balance synchronizes accesses to
5330 * ->active_balance_work. Once set, it's cleared
5331 * only after active load balance is finished.
5333 if (!busiest->active_balance) {
5334 busiest->active_balance = 1;
5335 busiest->push_cpu = this_cpu;
5338 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5340 if (active_balance) {
5341 stop_one_cpu_nowait(cpu_of(busiest),
5342 active_load_balance_cpu_stop, busiest,
5343 &busiest->active_balance_work);
5347 * We've kicked active balancing, reset the failure
5350 sd->nr_balance_failed = sd->cache_nice_tries+1;
5353 sd->nr_balance_failed = 0;
5355 if (likely(!active_balance)) {
5356 /* We were unbalanced, so reset the balancing interval */
5357 sd->balance_interval = sd->min_interval;
5360 * If we've begun active balancing, start to back off. This
5361 * case may not be covered by the all_pinned logic if there
5362 * is only 1 task on the busy runqueue (because we don't call
5365 if (sd->balance_interval < sd->max_interval)
5366 sd->balance_interval *= 2;
5372 schedstat_inc(sd, lb_balanced[idle]);
5374 sd->nr_balance_failed = 0;
5377 /* tune up the balancing interval */
5378 if (((env.flags & LBF_ALL_PINNED) &&
5379 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5380 (sd->balance_interval < sd->max_interval))
5381 sd->balance_interval *= 2;
5389 * idle_balance is called by schedule() if this_cpu is about to become
5390 * idle. Attempts to pull tasks from other CPUs.
5392 void idle_balance(int this_cpu, struct rq *this_rq)
5394 struct sched_domain *sd;
5395 int pulled_task = 0;
5396 unsigned long next_balance = jiffies + HZ;
5398 this_rq->idle_stamp = this_rq->clock;
5400 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5404 * Drop the rq->lock, but keep IRQ/preempt disabled.
5406 raw_spin_unlock(&this_rq->lock);
5408 update_blocked_averages(this_cpu);
5410 for_each_domain(this_cpu, sd) {
5411 unsigned long interval;
5414 if (!(sd->flags & SD_LOAD_BALANCE))
5417 if (sd->flags & SD_BALANCE_NEWIDLE) {
5418 /* If we've pulled tasks over stop searching: */
5419 pulled_task = load_balance(this_cpu, this_rq,
5420 sd, CPU_NEWLY_IDLE, &balance);
5423 interval = msecs_to_jiffies(sd->balance_interval);
5424 if (time_after(next_balance, sd->last_balance + interval))
5425 next_balance = sd->last_balance + interval;
5427 this_rq->idle_stamp = 0;
5433 raw_spin_lock(&this_rq->lock);
5435 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5437 * We are going idle. next_balance may be set based on
5438 * a busy processor. So reset next_balance.
5440 this_rq->next_balance = next_balance;
5445 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5446 * running tasks off the busiest CPU onto idle CPUs. It requires at
5447 * least 1 task to be running on each physical CPU where possible, and
5448 * avoids physical / logical imbalances.
5450 static int active_load_balance_cpu_stop(void *data)
5452 struct rq *busiest_rq = data;
5453 int busiest_cpu = cpu_of(busiest_rq);
5454 int target_cpu = busiest_rq->push_cpu;
5455 struct rq *target_rq = cpu_rq(target_cpu);
5456 struct sched_domain *sd;
5458 raw_spin_lock_irq(&busiest_rq->lock);
5460 /* make sure the requested cpu hasn't gone down in the meantime */
5461 if (unlikely(busiest_cpu != smp_processor_id() ||
5462 !busiest_rq->active_balance))
5465 /* Is there any task to move? */
5466 if (busiest_rq->nr_running <= 1)
5470 * This condition is "impossible", if it occurs
5471 * we need to fix it. Originally reported by
5472 * Bjorn Helgaas on a 128-cpu setup.
5474 BUG_ON(busiest_rq == target_rq);
5476 /* move a task from busiest_rq to target_rq */
5477 double_lock_balance(busiest_rq, target_rq);
5479 /* Search for an sd spanning us and the target CPU. */
5481 for_each_domain(target_cpu, sd) {
5482 if ((sd->flags & SD_LOAD_BALANCE) &&
5483 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5488 struct lb_env env = {
5490 .dst_cpu = target_cpu,
5491 .dst_rq = target_rq,
5492 .src_cpu = busiest_rq->cpu,
5493 .src_rq = busiest_rq,
5497 schedstat_inc(sd, alb_count);
5499 if (move_one_task(&env))
5500 schedstat_inc(sd, alb_pushed);
5502 schedstat_inc(sd, alb_failed);
5505 double_unlock_balance(busiest_rq, target_rq);
5507 busiest_rq->active_balance = 0;
5508 raw_spin_unlock_irq(&busiest_rq->lock);
5512 #ifdef CONFIG_NO_HZ_COMMON
5514 * idle load balancing details
5515 * - When one of the busy CPUs notice that there may be an idle rebalancing
5516 * needed, they will kick the idle load balancer, which then does idle
5517 * load balancing for all the idle CPUs.
5520 cpumask_var_t idle_cpus_mask;
5522 unsigned long next_balance; /* in jiffy units */
5523 } nohz ____cacheline_aligned;
5525 static inline int find_new_ilb(int call_cpu)
5527 int ilb = cpumask_first(nohz.idle_cpus_mask);
5529 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5536 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5537 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5538 * CPU (if there is one).
5540 static void nohz_balancer_kick(int cpu)
5544 nohz.next_balance++;
5546 ilb_cpu = find_new_ilb(cpu);
5548 if (ilb_cpu >= nr_cpu_ids)
5551 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5554 * Use smp_send_reschedule() instead of resched_cpu().
5555 * This way we generate a sched IPI on the target cpu which
5556 * is idle. And the softirq performing nohz idle load balance
5557 * will be run before returning from the IPI.
5559 smp_send_reschedule(ilb_cpu);
5563 static inline void nohz_balance_exit_idle(int cpu)
5565 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5566 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5567 atomic_dec(&nohz.nr_cpus);
5568 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5572 static inline void set_cpu_sd_state_busy(void)
5574 struct sched_domain *sd;
5575 int cpu = smp_processor_id();
5578 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5580 if (!sd || !sd->nohz_idle)
5584 for (; sd; sd = sd->parent)
5585 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5590 void set_cpu_sd_state_idle(void)
5592 struct sched_domain *sd;
5593 int cpu = smp_processor_id();
5596 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5598 if (!sd || sd->nohz_idle)
5602 for (; sd; sd = sd->parent)
5603 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5609 * This routine will record that the cpu is going idle with tick stopped.
5610 * This info will be used in performing idle load balancing in the future.
5612 void nohz_balance_enter_idle(int cpu)
5615 * If this cpu is going down, then nothing needs to be done.
5617 if (!cpu_active(cpu))
5620 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5623 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5624 atomic_inc(&nohz.nr_cpus);
5625 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5628 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5629 unsigned long action, void *hcpu)
5631 switch (action & ~CPU_TASKS_FROZEN) {
5633 nohz_balance_exit_idle(smp_processor_id());
5641 static DEFINE_SPINLOCK(balancing);
5644 * Scale the max load_balance interval with the number of CPUs in the system.
5645 * This trades load-balance latency on larger machines for less cross talk.
5647 void update_max_interval(void)
5649 max_load_balance_interval = HZ*num_online_cpus()/10;
5653 * It checks each scheduling domain to see if it is due to be balanced,
5654 * and initiates a balancing operation if so.
5656 * Balancing parameters are set up in init_sched_domains.
5658 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5661 struct rq *rq = cpu_rq(cpu);
5662 unsigned long interval;
5663 struct sched_domain *sd;
5664 /* Earliest time when we have to do rebalance again */
5665 unsigned long next_balance = jiffies + 60*HZ;
5666 int update_next_balance = 0;
5669 update_blocked_averages(cpu);
5672 for_each_domain(cpu, sd) {
5673 if (!(sd->flags & SD_LOAD_BALANCE))
5676 interval = sd->balance_interval;
5677 if (idle != CPU_IDLE)
5678 interval *= sd->busy_factor;
5680 /* scale ms to jiffies */
5681 interval = msecs_to_jiffies(interval);
5682 interval = clamp(interval, 1UL, max_load_balance_interval);
5684 need_serialize = sd->flags & SD_SERIALIZE;
5686 if (need_serialize) {
5687 if (!spin_trylock(&balancing))
5691 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5692 if (load_balance(cpu, rq, sd, idle, &balance)) {
5694 * The LBF_SOME_PINNED logic could have changed
5695 * env->dst_cpu, so we can't know our idle
5696 * state even if we migrated tasks. Update it.
5698 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5700 sd->last_balance = jiffies;
5703 spin_unlock(&balancing);
5705 if (time_after(next_balance, sd->last_balance + interval)) {
5706 next_balance = sd->last_balance + interval;
5707 update_next_balance = 1;
5711 * Stop the load balance at this level. There is another
5712 * CPU in our sched group which is doing load balancing more
5721 * next_balance will be updated only when there is a need.
5722 * When the cpu is attached to null domain for ex, it will not be
5725 if (likely(update_next_balance))
5726 rq->next_balance = next_balance;
5729 #ifdef CONFIG_NO_HZ_COMMON
5731 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5732 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5734 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5736 struct rq *this_rq = cpu_rq(this_cpu);
5740 if (idle != CPU_IDLE ||
5741 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5744 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5745 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5749 * If this cpu gets work to do, stop the load balancing
5750 * work being done for other cpus. Next load
5751 * balancing owner will pick it up.
5756 rq = cpu_rq(balance_cpu);
5758 raw_spin_lock_irq(&rq->lock);
5759 update_rq_clock(rq);
5760 update_idle_cpu_load(rq);
5761 raw_spin_unlock_irq(&rq->lock);
5763 rebalance_domains(balance_cpu, CPU_IDLE);
5765 if (time_after(this_rq->next_balance, rq->next_balance))
5766 this_rq->next_balance = rq->next_balance;
5768 nohz.next_balance = this_rq->next_balance;
5770 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5774 * Current heuristic for kicking the idle load balancer in the presence
5775 * of an idle cpu is the system.
5776 * - This rq has more than one task.
5777 * - At any scheduler domain level, this cpu's scheduler group has multiple
5778 * busy cpu's exceeding the group's power.
5779 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5780 * domain span are idle.
5782 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5784 unsigned long now = jiffies;
5785 struct sched_domain *sd;
5787 if (unlikely(idle_cpu(cpu)))
5791 * We may be recently in ticked or tickless idle mode. At the first
5792 * busy tick after returning from idle, we will update the busy stats.
5794 set_cpu_sd_state_busy();
5795 nohz_balance_exit_idle(cpu);
5798 * None are in tickless mode and hence no need for NOHZ idle load
5801 if (likely(!atomic_read(&nohz.nr_cpus)))
5804 if (time_before(now, nohz.next_balance))
5807 if (rq->nr_running >= 2)
5811 for_each_domain(cpu, sd) {
5812 struct sched_group *sg = sd->groups;
5813 struct sched_group_power *sgp = sg->sgp;
5814 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5816 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5817 goto need_kick_unlock;
5819 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5820 && (cpumask_first_and(nohz.idle_cpus_mask,
5821 sched_domain_span(sd)) < cpu))
5822 goto need_kick_unlock;
5824 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5836 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5839 #ifdef CONFIG_SCHED_HMP
5840 /* Check if task should migrate to a faster cpu */
5841 static unsigned int hmp_up_migration(int cpu, struct sched_entity *se)
5843 struct task_struct *p = task_of(se);
5845 if (hmp_cpu_is_fastest(cpu))
5848 if (cpumask_intersects(&hmp_faster_domain(cpu)->cpus,
5849 tsk_cpus_allowed(p))
5850 && se->avg.load_avg_ratio > hmp_up_threshold) {
5856 /* Check if task should migrate to a slower cpu */
5857 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
5859 struct task_struct *p = task_of(se);
5861 if (hmp_cpu_is_slowest(cpu))
5864 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
5865 tsk_cpus_allowed(p))
5866 && se->avg.load_avg_ratio < hmp_down_threshold) {
5873 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5874 * Ideally this function should be merged with can_migrate_task() to avoid
5877 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
5879 int tsk_cache_hot = 0;
5882 * We do not migrate tasks that are:
5883 * 1) running (obviously), or
5884 * 2) cannot be migrated to this CPU due to cpus_allowed
5886 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5887 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5890 env->flags &= ~LBF_ALL_PINNED;
5892 if (task_running(env->src_rq, p)) {
5893 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5898 * Aggressive migration if:
5899 * 1) task is cache cold, or
5900 * 2) too many balance attempts have failed.
5903 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
5904 if (!tsk_cache_hot ||
5905 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5906 #ifdef CONFIG_SCHEDSTATS
5907 if (tsk_cache_hot) {
5908 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5909 schedstat_inc(p, se.statistics.nr_forced_migrations);
5919 * move_specific_task tries to move a specific task.
5920 * Returns 1 if successful and 0 otherwise.
5921 * Called with both runqueues locked.
5923 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
5925 struct task_struct *p, *n;
5927 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5928 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
5932 if (!hmp_can_migrate_task(p, env))
5934 /* Check if we found the right task */
5940 * Right now, this is only the third place move_task()
5941 * is called, so we can safely collect move_task()
5942 * stats here rather than inside move_task().
5944 schedstat_inc(env->sd, lb_gained[env->idle]);
5951 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
5952 * migrate a specific task from one runqueue to another.
5953 * hmp_force_up_migration uses this to push a currently running task
5955 * Based on active_load_balance_stop_cpu and can potentially be merged.
5957 static int hmp_active_task_migration_cpu_stop(void *data)
5959 struct rq *busiest_rq = data;
5960 struct task_struct *p = busiest_rq->migrate_task;
5961 int busiest_cpu = cpu_of(busiest_rq);
5962 int target_cpu = busiest_rq->push_cpu;
5963 struct rq *target_rq = cpu_rq(target_cpu);
5964 struct sched_domain *sd;
5966 raw_spin_lock_irq(&busiest_rq->lock);
5967 /* make sure the requested cpu hasn't gone down in the meantime */
5968 if (unlikely(busiest_cpu != smp_processor_id() ||
5969 !busiest_rq->active_balance)) {
5972 /* Is there any task to move? */
5973 if (busiest_rq->nr_running <= 1)
5975 /* Task has migrated meanwhile, abort forced migration */
5976 if (task_rq(p) != busiest_rq)
5979 * This condition is "impossible", if it occurs
5980 * we need to fix it. Originally reported by
5981 * Bjorn Helgaas on a 128-cpu setup.
5983 BUG_ON(busiest_rq == target_rq);
5985 /* move a task from busiest_rq to target_rq */
5986 double_lock_balance(busiest_rq, target_rq);
5988 /* Search for an sd spanning us and the target CPU. */
5990 for_each_domain(target_cpu, sd) {
5991 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5996 struct lb_env env = {
5998 .dst_cpu = target_cpu,
5999 .dst_rq = target_rq,
6000 .src_cpu = busiest_rq->cpu,
6001 .src_rq = busiest_rq,
6005 schedstat_inc(sd, alb_count);
6007 if (move_specific_task(&env, p))
6008 schedstat_inc(sd, alb_pushed);
6010 schedstat_inc(sd, alb_failed);
6013 double_unlock_balance(busiest_rq, target_rq);
6015 busiest_rq->active_balance = 0;
6016 raw_spin_unlock_irq(&busiest_rq->lock);
6020 static DEFINE_SPINLOCK(hmp_force_migration);
6023 * hmp_force_up_migration checks runqueues for tasks that need to
6024 * be actively migrated to a faster cpu.
6026 static void hmp_force_up_migration(int this_cpu)
6029 struct sched_entity *curr;
6031 unsigned long flags;
6033 struct task_struct *p;
6035 if (!spin_trylock(&hmp_force_migration))
6037 for_each_online_cpu(cpu) {
6039 target = cpu_rq(cpu);
6040 raw_spin_lock_irqsave(&target->lock, flags);
6041 curr = target->cfs.curr;
6042 if (!curr || !entity_is_task(curr)) {
6043 raw_spin_unlock_irqrestore(&target->lock, flags);
6047 if (hmp_up_migration(cpu, curr)) {
6048 if (!target->active_balance) {
6049 target->active_balance = 1;
6050 target->push_cpu = hmp_select_faster_cpu(p, cpu);
6051 target->migrate_task = p;
6055 raw_spin_unlock_irqrestore(&target->lock, flags);
6057 stop_one_cpu_nowait(cpu_of(target),
6058 hmp_active_task_migration_cpu_stop,
6059 target, &target->active_balance_work);
6061 spin_unlock(&hmp_force_migration);
6064 static void hmp_force_up_migration(int this_cpu) { }
6065 #endif /* CONFIG_SCHED_HMP */
6068 * run_rebalance_domains is triggered when needed from the scheduler tick.
6069 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6071 static void run_rebalance_domains(struct softirq_action *h)
6073 int this_cpu = smp_processor_id();
6074 struct rq *this_rq = cpu_rq(this_cpu);
6075 enum cpu_idle_type idle = this_rq->idle_balance ?
6076 CPU_IDLE : CPU_NOT_IDLE;
6078 hmp_force_up_migration(this_cpu);
6080 rebalance_domains(this_cpu, idle);
6083 * If this cpu has a pending nohz_balance_kick, then do the
6084 * balancing on behalf of the other idle cpus whose ticks are
6087 nohz_idle_balance(this_cpu, idle);
6090 static inline int on_null_domain(int cpu)
6092 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6096 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6098 void trigger_load_balance(struct rq *rq, int cpu)
6100 /* Don't need to rebalance while attached to NULL domain */
6101 if (time_after_eq(jiffies, rq->next_balance) &&
6102 likely(!on_null_domain(cpu)))
6103 raise_softirq(SCHED_SOFTIRQ);
6104 #ifdef CONFIG_NO_HZ_COMMON
6105 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6106 nohz_balancer_kick(cpu);
6110 static void rq_online_fair(struct rq *rq)
6115 static void rq_offline_fair(struct rq *rq)
6119 /* Ensure any throttled groups are reachable by pick_next_task */
6120 unthrottle_offline_cfs_rqs(rq);
6123 #endif /* CONFIG_SMP */
6126 * scheduler tick hitting a task of our scheduling class:
6128 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6130 struct cfs_rq *cfs_rq;
6131 struct sched_entity *se = &curr->se;
6133 for_each_sched_entity(se) {
6134 cfs_rq = cfs_rq_of(se);
6135 entity_tick(cfs_rq, se, queued);
6138 if (sched_feat_numa(NUMA))
6139 task_tick_numa(rq, curr);
6141 update_rq_runnable_avg(rq, 1);
6145 * called on fork with the child task as argument from the parent's context
6146 * - child not yet on the tasklist
6147 * - preemption disabled
6149 static void task_fork_fair(struct task_struct *p)
6151 struct cfs_rq *cfs_rq;
6152 struct sched_entity *se = &p->se, *curr;
6153 int this_cpu = smp_processor_id();
6154 struct rq *rq = this_rq();
6155 unsigned long flags;
6157 raw_spin_lock_irqsave(&rq->lock, flags);
6159 update_rq_clock(rq);
6161 cfs_rq = task_cfs_rq(current);
6162 curr = cfs_rq->curr;
6164 if (unlikely(task_cpu(p) != this_cpu)) {
6166 __set_task_cpu(p, this_cpu);
6170 update_curr(cfs_rq);
6173 se->vruntime = curr->vruntime;
6174 place_entity(cfs_rq, se, 1);
6176 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6178 * Upon rescheduling, sched_class::put_prev_task() will place
6179 * 'current' within the tree based on its new key value.
6181 swap(curr->vruntime, se->vruntime);
6182 resched_task(rq->curr);
6185 se->vruntime -= cfs_rq->min_vruntime;
6187 raw_spin_unlock_irqrestore(&rq->lock, flags);
6191 * Priority of the task has changed. Check to see if we preempt
6195 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6201 * Reschedule if we are currently running on this runqueue and
6202 * our priority decreased, or if we are not currently running on
6203 * this runqueue and our priority is higher than the current's
6205 if (rq->curr == p) {
6206 if (p->prio > oldprio)
6207 resched_task(rq->curr);
6209 check_preempt_curr(rq, p, 0);
6212 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6214 struct sched_entity *se = &p->se;
6215 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6218 * Ensure the task's vruntime is normalized, so that when its
6219 * switched back to the fair class the enqueue_entity(.flags=0) will
6220 * do the right thing.
6222 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6223 * have normalized the vruntime, if it was !on_rq, then only when
6224 * the task is sleeping will it still have non-normalized vruntime.
6226 if (!se->on_rq && p->state != TASK_RUNNING) {
6228 * Fix up our vruntime so that the current sleep doesn't
6229 * cause 'unlimited' sleep bonus.
6231 place_entity(cfs_rq, se, 0);
6232 se->vruntime -= cfs_rq->min_vruntime;
6235 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
6237 * Remove our load from contribution when we leave sched_fair
6238 * and ensure we don't carry in an old decay_count if we
6241 if (p->se.avg.decay_count) {
6242 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
6243 __synchronize_entity_decay(&p->se);
6244 subtract_blocked_load_contrib(cfs_rq,
6245 p->se.avg.load_avg_contrib);
6251 * We switched to the sched_fair class.
6253 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6259 * We were most likely switched from sched_rt, so
6260 * kick off the schedule if running, otherwise just see
6261 * if we can still preempt the current task.
6264 resched_task(rq->curr);
6266 check_preempt_curr(rq, p, 0);
6269 /* Account for a task changing its policy or group.
6271 * This routine is mostly called to set cfs_rq->curr field when a task
6272 * migrates between groups/classes.
6274 static void set_curr_task_fair(struct rq *rq)
6276 struct sched_entity *se = &rq->curr->se;
6278 for_each_sched_entity(se) {
6279 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6281 set_next_entity(cfs_rq, se);
6282 /* ensure bandwidth has been allocated on our new cfs_rq */
6283 account_cfs_rq_runtime(cfs_rq, 0);
6287 void init_cfs_rq(struct cfs_rq *cfs_rq)
6289 cfs_rq->tasks_timeline = RB_ROOT;
6290 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6291 #ifndef CONFIG_64BIT
6292 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6294 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
6295 atomic64_set(&cfs_rq->decay_counter, 1);
6296 atomic64_set(&cfs_rq->removed_load, 0);
6300 #ifdef CONFIG_FAIR_GROUP_SCHED
6301 static void task_move_group_fair(struct task_struct *p, int on_rq)
6303 struct cfs_rq *cfs_rq;
6305 * If the task was not on the rq at the time of this cgroup movement
6306 * it must have been asleep, sleeping tasks keep their ->vruntime
6307 * absolute on their old rq until wakeup (needed for the fair sleeper
6308 * bonus in place_entity()).
6310 * If it was on the rq, we've just 'preempted' it, which does convert
6311 * ->vruntime to a relative base.
6313 * Make sure both cases convert their relative position when migrating
6314 * to another cgroup's rq. This does somewhat interfere with the
6315 * fair sleeper stuff for the first placement, but who cares.
6318 * When !on_rq, vruntime of the task has usually NOT been normalized.
6319 * But there are some cases where it has already been normalized:
6321 * - Moving a forked child which is waiting for being woken up by
6322 * wake_up_new_task().
6323 * - Moving a task which has been woken up by try_to_wake_up() and
6324 * waiting for actually being woken up by sched_ttwu_pending().
6326 * To prevent boost or penalty in the new cfs_rq caused by delta
6327 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6329 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6333 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6334 set_task_rq(p, task_cpu(p));
6336 cfs_rq = cfs_rq_of(&p->se);
6337 p->se.vruntime += cfs_rq->min_vruntime;
6340 * migrate_task_rq_fair() will have removed our previous
6341 * contribution, but we must synchronize for ongoing future
6344 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6345 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6350 void free_fair_sched_group(struct task_group *tg)
6354 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6356 for_each_possible_cpu(i) {
6358 kfree(tg->cfs_rq[i]);
6367 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6369 struct cfs_rq *cfs_rq;
6370 struct sched_entity *se;
6373 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6376 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6380 tg->shares = NICE_0_LOAD;
6382 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6384 for_each_possible_cpu(i) {
6385 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6386 GFP_KERNEL, cpu_to_node(i));
6390 se = kzalloc_node(sizeof(struct sched_entity),
6391 GFP_KERNEL, cpu_to_node(i));
6395 init_cfs_rq(cfs_rq);
6396 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6407 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6409 struct rq *rq = cpu_rq(cpu);
6410 unsigned long flags;
6413 * Only empty task groups can be destroyed; so we can speculatively
6414 * check on_list without danger of it being re-added.
6416 if (!tg->cfs_rq[cpu]->on_list)
6419 raw_spin_lock_irqsave(&rq->lock, flags);
6420 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6421 raw_spin_unlock_irqrestore(&rq->lock, flags);
6424 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6425 struct sched_entity *se, int cpu,
6426 struct sched_entity *parent)
6428 struct rq *rq = cpu_rq(cpu);
6432 init_cfs_rq_runtime(cfs_rq);
6434 tg->cfs_rq[cpu] = cfs_rq;
6437 /* se could be NULL for root_task_group */
6442 se->cfs_rq = &rq->cfs;
6444 se->cfs_rq = parent->my_q;
6447 update_load_set(&se->load, 0);
6448 se->parent = parent;
6451 static DEFINE_MUTEX(shares_mutex);
6453 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6456 unsigned long flags;
6459 * We can't change the weight of the root cgroup.
6464 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6466 mutex_lock(&shares_mutex);
6467 if (tg->shares == shares)
6470 tg->shares = shares;
6471 for_each_possible_cpu(i) {
6472 struct rq *rq = cpu_rq(i);
6473 struct sched_entity *se;
6476 /* Propagate contribution to hierarchy */
6477 raw_spin_lock_irqsave(&rq->lock, flags);
6478 for_each_sched_entity(se)
6479 update_cfs_shares(group_cfs_rq(se));
6480 raw_spin_unlock_irqrestore(&rq->lock, flags);
6484 mutex_unlock(&shares_mutex);
6487 #else /* CONFIG_FAIR_GROUP_SCHED */
6489 void free_fair_sched_group(struct task_group *tg) { }
6491 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6496 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6498 #endif /* CONFIG_FAIR_GROUP_SCHED */
6501 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6503 struct sched_entity *se = &task->se;
6504 unsigned int rr_interval = 0;
6507 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6510 if (rq->cfs.load.weight)
6511 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6517 * All the scheduling class methods:
6519 const struct sched_class fair_sched_class = {
6520 .next = &idle_sched_class,
6521 .enqueue_task = enqueue_task_fair,
6522 .dequeue_task = dequeue_task_fair,
6523 .yield_task = yield_task_fair,
6524 .yield_to_task = yield_to_task_fair,
6526 .check_preempt_curr = check_preempt_wakeup,
6528 .pick_next_task = pick_next_task_fair,
6529 .put_prev_task = put_prev_task_fair,
6532 .select_task_rq = select_task_rq_fair,
6533 #ifdef CONFIG_FAIR_GROUP_SCHED
6534 .migrate_task_rq = migrate_task_rq_fair,
6536 .rq_online = rq_online_fair,
6537 .rq_offline = rq_offline_fair,
6539 .task_waking = task_waking_fair,
6542 .set_curr_task = set_curr_task_fair,
6543 .task_tick = task_tick_fair,
6544 .task_fork = task_fork_fair,
6546 .prio_changed = prio_changed_fair,
6547 .switched_from = switched_from_fair,
6548 .switched_to = switched_to_fair,
6550 .get_rr_interval = get_rr_interval_fair,
6552 #ifdef CONFIG_FAIR_GROUP_SCHED
6553 .task_move_group = task_move_group_fair,
6557 #ifdef CONFIG_SCHED_DEBUG
6558 void print_cfs_stats(struct seq_file *m, int cpu)
6560 struct cfs_rq *cfs_rq;
6563 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6564 print_cfs_rq(m, cpu, cfs_rq);
6569 __init void init_sched_fair_class(void)
6572 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6574 #ifdef CONFIG_NO_HZ_COMMON
6575 nohz.next_balance = jiffies;
6576 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6577 cpu_notifier(sched_ilb_notifier, 0);
6580 #ifdef CONFIG_SCHED_HMP
6581 hmp_cpu_mask_setup();