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 trace_sched_task_load_contrib(task_of(se), se->avg.load_avg_contrib);
1432 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1433 contrib /= (se->avg.runnable_avg_period + 1);
1434 se->avg.load_avg_ratio = scale_load(contrib);
1435 trace_sched_task_runnable_ratio(task_of(se), se->avg.load_avg_ratio);
1438 /* Compute the current contribution to load_avg by se, return any delta */
1439 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1441 long old_contrib = se->avg.load_avg_contrib;
1443 if (entity_is_task(se)) {
1444 __update_task_entity_contrib(se);
1446 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1447 __update_group_entity_contrib(se);
1450 return se->avg.load_avg_contrib - old_contrib;
1453 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1456 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1457 cfs_rq->blocked_load_avg -= load_contrib;
1459 cfs_rq->blocked_load_avg = 0;
1462 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1464 /* Update a sched_entity's runnable average */
1465 static inline void update_entity_load_avg(struct sched_entity *se,
1468 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1473 * For a group entity we need to use their owned cfs_rq_clock_task() in
1474 * case they are the parent of a throttled hierarchy.
1476 if (entity_is_task(se))
1477 now = cfs_rq_clock_task(cfs_rq);
1479 now = cfs_rq_clock_task(group_cfs_rq(se));
1481 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1482 cfs_rq->curr == se))
1485 contrib_delta = __update_entity_load_avg_contrib(se);
1491 cfs_rq->runnable_load_avg += contrib_delta;
1493 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1497 * Decay the load contributed by all blocked children and account this so that
1498 * their contribution may appropriately discounted when they wake up.
1500 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1502 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1505 decays = now - cfs_rq->last_decay;
1506 if (!decays && !force_update)
1509 if (atomic64_read(&cfs_rq->removed_load)) {
1510 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1511 subtract_blocked_load_contrib(cfs_rq, removed_load);
1515 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1517 atomic64_add(decays, &cfs_rq->decay_counter);
1518 cfs_rq->last_decay = now;
1521 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1524 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1527 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1529 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1530 contrib = rq->avg.runnable_avg_sum * scale_load_down(1024);
1531 contrib /= (rq->avg.runnable_avg_period + 1);
1532 trace_sched_rq_runnable_ratio(cpu_of(rq), scale_load(contrib));
1533 trace_sched_rq_runnable_load(cpu_of(rq), rq->cfs.runnable_load_avg);
1536 /* Add the load generated by se into cfs_rq's child load-average */
1537 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1538 struct sched_entity *se,
1542 * We track migrations using entity decay_count <= 0, on a wake-up
1543 * migration we use a negative decay count to track the remote decays
1544 * accumulated while sleeping.
1546 if (unlikely(se->avg.decay_count <= 0)) {
1547 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1548 if (se->avg.decay_count) {
1550 * In a wake-up migration we have to approximate the
1551 * time sleeping. This is because we can't synchronize
1552 * clock_task between the two cpus, and it is not
1553 * guaranteed to be read-safe. Instead, we can
1554 * approximate this using our carried decays, which are
1555 * explicitly atomically readable.
1557 se->avg.last_runnable_update -= (-se->avg.decay_count)
1559 update_entity_load_avg(se, 0);
1560 /* Indicate that we're now synchronized and on-rq */
1561 se->avg.decay_count = 0;
1565 __synchronize_entity_decay(se);
1568 /* migrated tasks did not contribute to our blocked load */
1570 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1571 update_entity_load_avg(se, 0);
1574 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1575 /* we force update consideration on load-balancer moves */
1576 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1580 * Remove se's load from this cfs_rq child load-average, if the entity is
1581 * transitioning to a blocked state we track its projected decay using
1584 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1585 struct sched_entity *se,
1588 update_entity_load_avg(se, 1);
1589 /* we force update consideration on load-balancer moves */
1590 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1592 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1594 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1595 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1596 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1600 * Update the rq's load with the elapsed running time before entering
1601 * idle. if the last scheduled task is not a CFS task, idle_enter will
1602 * be the only way to update the runnable statistic.
1604 void idle_enter_fair(struct rq *this_rq)
1606 update_rq_runnable_avg(this_rq, 1);
1610 * Update the rq's load with the elapsed idle time before a task is
1611 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1612 * be the only way to update the runnable statistic.
1614 void idle_exit_fair(struct rq *this_rq)
1616 update_rq_runnable_avg(this_rq, 0);
1620 static inline void update_entity_load_avg(struct sched_entity *se,
1621 int update_cfs_rq) {}
1622 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1623 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1624 struct sched_entity *se,
1626 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1627 struct sched_entity *se,
1629 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1630 int force_update) {}
1633 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1635 #ifdef CONFIG_SCHEDSTATS
1636 struct task_struct *tsk = NULL;
1638 if (entity_is_task(se))
1641 if (se->statistics.sleep_start) {
1642 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1647 if (unlikely(delta > se->statistics.sleep_max))
1648 se->statistics.sleep_max = delta;
1650 se->statistics.sleep_start = 0;
1651 se->statistics.sum_sleep_runtime += delta;
1654 account_scheduler_latency(tsk, delta >> 10, 1);
1655 trace_sched_stat_sleep(tsk, delta);
1658 if (se->statistics.block_start) {
1659 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1664 if (unlikely(delta > se->statistics.block_max))
1665 se->statistics.block_max = delta;
1667 se->statistics.block_start = 0;
1668 se->statistics.sum_sleep_runtime += delta;
1671 if (tsk->in_iowait) {
1672 se->statistics.iowait_sum += delta;
1673 se->statistics.iowait_count++;
1674 trace_sched_stat_iowait(tsk, delta);
1677 trace_sched_stat_blocked(tsk, delta);
1680 * Blocking time is in units of nanosecs, so shift by
1681 * 20 to get a milliseconds-range estimation of the
1682 * amount of time that the task spent sleeping:
1684 if (unlikely(prof_on == SLEEP_PROFILING)) {
1685 profile_hits(SLEEP_PROFILING,
1686 (void *)get_wchan(tsk),
1689 account_scheduler_latency(tsk, delta >> 10, 0);
1695 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1697 #ifdef CONFIG_SCHED_DEBUG
1698 s64 d = se->vruntime - cfs_rq->min_vruntime;
1703 if (d > 3*sysctl_sched_latency)
1704 schedstat_inc(cfs_rq, nr_spread_over);
1709 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1711 u64 vruntime = cfs_rq->min_vruntime;
1714 * The 'current' period is already promised to the current tasks,
1715 * however the extra weight of the new task will slow them down a
1716 * little, place the new task so that it fits in the slot that
1717 * stays open at the end.
1719 if (initial && sched_feat(START_DEBIT))
1720 vruntime += sched_vslice(cfs_rq, se);
1722 /* sleeps up to a single latency don't count. */
1724 unsigned long thresh = sysctl_sched_latency;
1727 * Halve their sleep time's effect, to allow
1728 * for a gentler effect of sleepers:
1730 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1736 /* ensure we never gain time by being placed backwards. */
1737 se->vruntime = max_vruntime(se->vruntime, vruntime);
1740 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1743 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1746 * Update the normalized vruntime before updating min_vruntime
1747 * through callig update_curr().
1749 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1750 se->vruntime += cfs_rq->min_vruntime;
1753 * Update run-time statistics of the 'current'.
1755 update_curr(cfs_rq);
1756 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1757 account_entity_enqueue(cfs_rq, se);
1758 update_cfs_shares(cfs_rq);
1760 if (flags & ENQUEUE_WAKEUP) {
1761 place_entity(cfs_rq, se, 0);
1762 enqueue_sleeper(cfs_rq, se);
1765 update_stats_enqueue(cfs_rq, se);
1766 check_spread(cfs_rq, se);
1767 if (se != cfs_rq->curr)
1768 __enqueue_entity(cfs_rq, se);
1771 if (cfs_rq->nr_running == 1) {
1772 list_add_leaf_cfs_rq(cfs_rq);
1773 check_enqueue_throttle(cfs_rq);
1777 static void __clear_buddies_last(struct sched_entity *se)
1779 for_each_sched_entity(se) {
1780 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1781 if (cfs_rq->last == se)
1782 cfs_rq->last = NULL;
1788 static void __clear_buddies_next(struct sched_entity *se)
1790 for_each_sched_entity(se) {
1791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1792 if (cfs_rq->next == se)
1793 cfs_rq->next = NULL;
1799 static void __clear_buddies_skip(struct sched_entity *se)
1801 for_each_sched_entity(se) {
1802 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1803 if (cfs_rq->skip == se)
1804 cfs_rq->skip = NULL;
1810 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1812 if (cfs_rq->last == se)
1813 __clear_buddies_last(se);
1815 if (cfs_rq->next == se)
1816 __clear_buddies_next(se);
1818 if (cfs_rq->skip == se)
1819 __clear_buddies_skip(se);
1822 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1825 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1828 * Update run-time statistics of the 'current'.
1830 update_curr(cfs_rq);
1831 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1833 update_stats_dequeue(cfs_rq, se);
1834 if (flags & DEQUEUE_SLEEP) {
1835 #ifdef CONFIG_SCHEDSTATS
1836 if (entity_is_task(se)) {
1837 struct task_struct *tsk = task_of(se);
1839 if (tsk->state & TASK_INTERRUPTIBLE)
1840 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1841 if (tsk->state & TASK_UNINTERRUPTIBLE)
1842 se->statistics.block_start = rq_of(cfs_rq)->clock;
1847 clear_buddies(cfs_rq, se);
1849 if (se != cfs_rq->curr)
1850 __dequeue_entity(cfs_rq, se);
1852 account_entity_dequeue(cfs_rq, se);
1855 * Normalize the entity after updating the min_vruntime because the
1856 * update can refer to the ->curr item and we need to reflect this
1857 * movement in our normalized position.
1859 if (!(flags & DEQUEUE_SLEEP))
1860 se->vruntime -= cfs_rq->min_vruntime;
1862 /* return excess runtime on last dequeue */
1863 return_cfs_rq_runtime(cfs_rq);
1865 update_min_vruntime(cfs_rq);
1866 update_cfs_shares(cfs_rq);
1870 * Preempt the current task with a newly woken task if needed:
1873 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1875 unsigned long ideal_runtime, delta_exec;
1876 struct sched_entity *se;
1879 ideal_runtime = sched_slice(cfs_rq, curr);
1880 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1881 if (delta_exec > ideal_runtime) {
1882 resched_task(rq_of(cfs_rq)->curr);
1884 * The current task ran long enough, ensure it doesn't get
1885 * re-elected due to buddy favours.
1887 clear_buddies(cfs_rq, curr);
1892 * Ensure that a task that missed wakeup preemption by a
1893 * narrow margin doesn't have to wait for a full slice.
1894 * This also mitigates buddy induced latencies under load.
1896 if (delta_exec < sysctl_sched_min_granularity)
1899 se = __pick_first_entity(cfs_rq);
1900 delta = curr->vruntime - se->vruntime;
1905 if (delta > ideal_runtime)
1906 resched_task(rq_of(cfs_rq)->curr);
1910 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1912 /* 'current' is not kept within the tree. */
1915 * Any task has to be enqueued before it get to execute on
1916 * a CPU. So account for the time it spent waiting on the
1919 update_stats_wait_end(cfs_rq, se);
1920 __dequeue_entity(cfs_rq, se);
1921 update_entity_load_avg(se, 1);
1924 update_stats_curr_start(cfs_rq, se);
1926 #ifdef CONFIG_SCHEDSTATS
1928 * Track our maximum slice length, if the CPU's load is at
1929 * least twice that of our own weight (i.e. dont track it
1930 * when there are only lesser-weight tasks around):
1932 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1933 se->statistics.slice_max = max(se->statistics.slice_max,
1934 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1937 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1941 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1944 * Pick the next process, keeping these things in mind, in this order:
1945 * 1) keep things fair between processes/task groups
1946 * 2) pick the "next" process, since someone really wants that to run
1947 * 3) pick the "last" process, for cache locality
1948 * 4) do not run the "skip" process, if something else is available
1950 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1952 struct sched_entity *se = __pick_first_entity(cfs_rq);
1953 struct sched_entity *left = se;
1956 * Avoid running the skip buddy, if running something else can
1957 * be done without getting too unfair.
1959 if (cfs_rq->skip == se) {
1960 struct sched_entity *second = __pick_next_entity(se);
1961 if (second && wakeup_preempt_entity(second, left) < 1)
1966 * Prefer last buddy, try to return the CPU to a preempted task.
1968 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1972 * Someone really wants this to run. If it's not unfair, run it.
1974 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1977 clear_buddies(cfs_rq, se);
1982 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1984 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1987 * If still on the runqueue then deactivate_task()
1988 * was not called and update_curr() has to be done:
1991 update_curr(cfs_rq);
1993 /* throttle cfs_rqs exceeding runtime */
1994 check_cfs_rq_runtime(cfs_rq);
1996 check_spread(cfs_rq, prev);
1998 update_stats_wait_start(cfs_rq, prev);
1999 /* Put 'current' back into the tree. */
2000 __enqueue_entity(cfs_rq, prev);
2001 /* in !on_rq case, update occurred at dequeue */
2002 update_entity_load_avg(prev, 1);
2004 cfs_rq->curr = NULL;
2008 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2011 * Update run-time statistics of the 'current'.
2013 update_curr(cfs_rq);
2016 * Ensure that runnable average is periodically updated.
2018 update_entity_load_avg(curr, 1);
2019 update_cfs_rq_blocked_load(cfs_rq, 1);
2021 #ifdef CONFIG_SCHED_HRTICK
2023 * queued ticks are scheduled to match the slice, so don't bother
2024 * validating it and just reschedule.
2027 resched_task(rq_of(cfs_rq)->curr);
2031 * don't let the period tick interfere with the hrtick preemption
2033 if (!sched_feat(DOUBLE_TICK) &&
2034 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2038 if (cfs_rq->nr_running > 1)
2039 check_preempt_tick(cfs_rq, curr);
2043 /**************************************************
2044 * CFS bandwidth control machinery
2047 #ifdef CONFIG_CFS_BANDWIDTH
2049 #ifdef HAVE_JUMP_LABEL
2050 static struct static_key __cfs_bandwidth_used;
2052 static inline bool cfs_bandwidth_used(void)
2054 return static_key_false(&__cfs_bandwidth_used);
2057 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2059 /* only need to count groups transitioning between enabled/!enabled */
2060 if (enabled && !was_enabled)
2061 static_key_slow_inc(&__cfs_bandwidth_used);
2062 else if (!enabled && was_enabled)
2063 static_key_slow_dec(&__cfs_bandwidth_used);
2065 #else /* HAVE_JUMP_LABEL */
2066 static bool cfs_bandwidth_used(void)
2071 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2072 #endif /* HAVE_JUMP_LABEL */
2075 * default period for cfs group bandwidth.
2076 * default: 0.1s, units: nanoseconds
2078 static inline u64 default_cfs_period(void)
2080 return 100000000ULL;
2083 static inline u64 sched_cfs_bandwidth_slice(void)
2085 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2089 * Replenish runtime according to assigned quota and update expiration time.
2090 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2091 * additional synchronization around rq->lock.
2093 * requires cfs_b->lock
2095 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2099 if (cfs_b->quota == RUNTIME_INF)
2102 now = sched_clock_cpu(smp_processor_id());
2103 cfs_b->runtime = cfs_b->quota;
2104 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2107 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2109 return &tg->cfs_bandwidth;
2112 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2113 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2115 if (unlikely(cfs_rq->throttle_count))
2116 return cfs_rq->throttled_clock_task;
2118 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2121 /* returns 0 on failure to allocate runtime */
2122 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2124 struct task_group *tg = cfs_rq->tg;
2125 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2126 u64 amount = 0, min_amount, expires;
2128 /* note: this is a positive sum as runtime_remaining <= 0 */
2129 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2131 raw_spin_lock(&cfs_b->lock);
2132 if (cfs_b->quota == RUNTIME_INF)
2133 amount = min_amount;
2136 * If the bandwidth pool has become inactive, then at least one
2137 * period must have elapsed since the last consumption.
2138 * Refresh the global state and ensure bandwidth timer becomes
2141 if (!cfs_b->timer_active) {
2142 __refill_cfs_bandwidth_runtime(cfs_b);
2143 __start_cfs_bandwidth(cfs_b);
2146 if (cfs_b->runtime > 0) {
2147 amount = min(cfs_b->runtime, min_amount);
2148 cfs_b->runtime -= amount;
2152 expires = cfs_b->runtime_expires;
2153 raw_spin_unlock(&cfs_b->lock);
2155 cfs_rq->runtime_remaining += amount;
2157 * we may have advanced our local expiration to account for allowed
2158 * spread between our sched_clock and the one on which runtime was
2161 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2162 cfs_rq->runtime_expires = expires;
2164 return cfs_rq->runtime_remaining > 0;
2168 * Note: This depends on the synchronization provided by sched_clock and the
2169 * fact that rq->clock snapshots this value.
2171 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2173 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2174 struct rq *rq = rq_of(cfs_rq);
2176 /* if the deadline is ahead of our clock, nothing to do */
2177 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2180 if (cfs_rq->runtime_remaining < 0)
2184 * If the local deadline has passed we have to consider the
2185 * possibility that our sched_clock is 'fast' and the global deadline
2186 * has not truly expired.
2188 * Fortunately we can check determine whether this the case by checking
2189 * whether the global deadline has advanced.
2192 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2193 /* extend local deadline, drift is bounded above by 2 ticks */
2194 cfs_rq->runtime_expires += TICK_NSEC;
2196 /* global deadline is ahead, expiration has passed */
2197 cfs_rq->runtime_remaining = 0;
2201 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2202 unsigned long delta_exec)
2204 /* dock delta_exec before expiring quota (as it could span periods) */
2205 cfs_rq->runtime_remaining -= delta_exec;
2206 expire_cfs_rq_runtime(cfs_rq);
2208 if (likely(cfs_rq->runtime_remaining > 0))
2212 * if we're unable to extend our runtime we resched so that the active
2213 * hierarchy can be throttled
2215 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2216 resched_task(rq_of(cfs_rq)->curr);
2219 static __always_inline
2220 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2222 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2225 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2228 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2230 return cfs_bandwidth_used() && cfs_rq->throttled;
2233 /* check whether cfs_rq, or any parent, is throttled */
2234 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2236 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2240 * Ensure that neither of the group entities corresponding to src_cpu or
2241 * dest_cpu are members of a throttled hierarchy when performing group
2242 * load-balance operations.
2244 static inline int throttled_lb_pair(struct task_group *tg,
2245 int src_cpu, int dest_cpu)
2247 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2249 src_cfs_rq = tg->cfs_rq[src_cpu];
2250 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2252 return throttled_hierarchy(src_cfs_rq) ||
2253 throttled_hierarchy(dest_cfs_rq);
2256 /* updated child weight may affect parent so we have to do this bottom up */
2257 static int tg_unthrottle_up(struct task_group *tg, void *data)
2259 struct rq *rq = data;
2260 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2262 cfs_rq->throttle_count--;
2264 if (!cfs_rq->throttle_count) {
2265 /* adjust cfs_rq_clock_task() */
2266 cfs_rq->throttled_clock_task_time += rq->clock_task -
2267 cfs_rq->throttled_clock_task;
2274 static int tg_throttle_down(struct task_group *tg, void *data)
2276 struct rq *rq = data;
2277 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2279 /* group is entering throttled state, stop time */
2280 if (!cfs_rq->throttle_count)
2281 cfs_rq->throttled_clock_task = rq->clock_task;
2282 cfs_rq->throttle_count++;
2287 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2289 struct rq *rq = rq_of(cfs_rq);
2290 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2291 struct sched_entity *se;
2292 long task_delta, dequeue = 1;
2294 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2296 /* freeze hierarchy runnable averages while throttled */
2298 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2301 task_delta = cfs_rq->h_nr_running;
2302 for_each_sched_entity(se) {
2303 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2304 /* throttled entity or throttle-on-deactivate */
2309 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2310 qcfs_rq->h_nr_running -= task_delta;
2312 if (qcfs_rq->load.weight)
2317 rq->nr_running -= task_delta;
2319 cfs_rq->throttled = 1;
2320 cfs_rq->throttled_clock = rq->clock;
2321 raw_spin_lock(&cfs_b->lock);
2322 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2323 raw_spin_unlock(&cfs_b->lock);
2326 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2328 struct rq *rq = rq_of(cfs_rq);
2329 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2330 struct sched_entity *se;
2334 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2336 cfs_rq->throttled = 0;
2337 raw_spin_lock(&cfs_b->lock);
2338 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2339 list_del_rcu(&cfs_rq->throttled_list);
2340 raw_spin_unlock(&cfs_b->lock);
2342 update_rq_clock(rq);
2343 /* update hierarchical throttle state */
2344 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2346 if (!cfs_rq->load.weight)
2349 task_delta = cfs_rq->h_nr_running;
2350 for_each_sched_entity(se) {
2354 cfs_rq = cfs_rq_of(se);
2356 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2357 cfs_rq->h_nr_running += task_delta;
2359 if (cfs_rq_throttled(cfs_rq))
2364 rq->nr_running += task_delta;
2366 /* determine whether we need to wake up potentially idle cpu */
2367 if (rq->curr == rq->idle && rq->cfs.nr_running)
2368 resched_task(rq->curr);
2371 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2372 u64 remaining, u64 expires)
2374 struct cfs_rq *cfs_rq;
2375 u64 runtime = remaining;
2378 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2380 struct rq *rq = rq_of(cfs_rq);
2382 raw_spin_lock(&rq->lock);
2383 if (!cfs_rq_throttled(cfs_rq))
2386 runtime = -cfs_rq->runtime_remaining + 1;
2387 if (runtime > remaining)
2388 runtime = remaining;
2389 remaining -= runtime;
2391 cfs_rq->runtime_remaining += runtime;
2392 cfs_rq->runtime_expires = expires;
2394 /* we check whether we're throttled above */
2395 if (cfs_rq->runtime_remaining > 0)
2396 unthrottle_cfs_rq(cfs_rq);
2399 raw_spin_unlock(&rq->lock);
2410 * Responsible for refilling a task_group's bandwidth and unthrottling its
2411 * cfs_rqs as appropriate. If there has been no activity within the last
2412 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2413 * used to track this state.
2415 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2417 u64 runtime, runtime_expires;
2418 int idle = 1, throttled;
2420 raw_spin_lock(&cfs_b->lock);
2421 /* no need to continue the timer with no bandwidth constraint */
2422 if (cfs_b->quota == RUNTIME_INF)
2425 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2426 /* idle depends on !throttled (for the case of a large deficit) */
2427 idle = cfs_b->idle && !throttled;
2428 cfs_b->nr_periods += overrun;
2430 /* if we're going inactive then everything else can be deferred */
2434 __refill_cfs_bandwidth_runtime(cfs_b);
2437 /* mark as potentially idle for the upcoming period */
2442 /* account preceding periods in which throttling occurred */
2443 cfs_b->nr_throttled += overrun;
2446 * There are throttled entities so we must first use the new bandwidth
2447 * to unthrottle them before making it generally available. This
2448 * ensures that all existing debts will be paid before a new cfs_rq is
2451 runtime = cfs_b->runtime;
2452 runtime_expires = cfs_b->runtime_expires;
2456 * This check is repeated as we are holding onto the new bandwidth
2457 * while we unthrottle. This can potentially race with an unthrottled
2458 * group trying to acquire new bandwidth from the global pool.
2460 while (throttled && runtime > 0) {
2461 raw_spin_unlock(&cfs_b->lock);
2462 /* we can't nest cfs_b->lock while distributing bandwidth */
2463 runtime = distribute_cfs_runtime(cfs_b, runtime,
2465 raw_spin_lock(&cfs_b->lock);
2467 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2470 /* return (any) remaining runtime */
2471 cfs_b->runtime = runtime;
2473 * While we are ensured activity in the period following an
2474 * unthrottle, this also covers the case in which the new bandwidth is
2475 * insufficient to cover the existing bandwidth deficit. (Forcing the
2476 * timer to remain active while there are any throttled entities.)
2481 cfs_b->timer_active = 0;
2482 raw_spin_unlock(&cfs_b->lock);
2487 /* a cfs_rq won't donate quota below this amount */
2488 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2489 /* minimum remaining period time to redistribute slack quota */
2490 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2491 /* how long we wait to gather additional slack before distributing */
2492 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2494 /* are we near the end of the current quota period? */
2495 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2497 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2500 /* if the call-back is running a quota refresh is already occurring */
2501 if (hrtimer_callback_running(refresh_timer))
2504 /* is a quota refresh about to occur? */
2505 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2506 if (remaining < min_expire)
2512 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2514 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2516 /* if there's a quota refresh soon don't bother with slack */
2517 if (runtime_refresh_within(cfs_b, min_left))
2520 start_bandwidth_timer(&cfs_b->slack_timer,
2521 ns_to_ktime(cfs_bandwidth_slack_period));
2524 /* we know any runtime found here is valid as update_curr() precedes return */
2525 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2527 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2528 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2530 if (slack_runtime <= 0)
2533 raw_spin_lock(&cfs_b->lock);
2534 if (cfs_b->quota != RUNTIME_INF &&
2535 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2536 cfs_b->runtime += slack_runtime;
2538 /* we are under rq->lock, defer unthrottling using a timer */
2539 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2540 !list_empty(&cfs_b->throttled_cfs_rq))
2541 start_cfs_slack_bandwidth(cfs_b);
2543 raw_spin_unlock(&cfs_b->lock);
2545 /* even if it's not valid for return we don't want to try again */
2546 cfs_rq->runtime_remaining -= slack_runtime;
2549 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2551 if (!cfs_bandwidth_used())
2554 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2557 __return_cfs_rq_runtime(cfs_rq);
2561 * This is done with a timer (instead of inline with bandwidth return) since
2562 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2564 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2566 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2569 /* confirm we're still not at a refresh boundary */
2570 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2573 raw_spin_lock(&cfs_b->lock);
2574 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2575 runtime = cfs_b->runtime;
2578 expires = cfs_b->runtime_expires;
2579 raw_spin_unlock(&cfs_b->lock);
2584 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2586 raw_spin_lock(&cfs_b->lock);
2587 if (expires == cfs_b->runtime_expires)
2588 cfs_b->runtime = runtime;
2589 raw_spin_unlock(&cfs_b->lock);
2593 * When a group wakes up we want to make sure that its quota is not already
2594 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2595 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2597 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2599 if (!cfs_bandwidth_used())
2602 /* an active group must be handled by the update_curr()->put() path */
2603 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2606 /* ensure the group is not already throttled */
2607 if (cfs_rq_throttled(cfs_rq))
2610 /* update runtime allocation */
2611 account_cfs_rq_runtime(cfs_rq, 0);
2612 if (cfs_rq->runtime_remaining <= 0)
2613 throttle_cfs_rq(cfs_rq);
2616 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2617 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2619 if (!cfs_bandwidth_used())
2622 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2626 * it's possible for a throttled entity to be forced into a running
2627 * state (e.g. set_curr_task), in this case we're finished.
2629 if (cfs_rq_throttled(cfs_rq))
2632 throttle_cfs_rq(cfs_rq);
2635 static inline u64 default_cfs_period(void);
2636 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2637 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2639 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2641 struct cfs_bandwidth *cfs_b =
2642 container_of(timer, struct cfs_bandwidth, slack_timer);
2643 do_sched_cfs_slack_timer(cfs_b);
2645 return HRTIMER_NORESTART;
2648 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2650 struct cfs_bandwidth *cfs_b =
2651 container_of(timer, struct cfs_bandwidth, period_timer);
2657 now = hrtimer_cb_get_time(timer);
2658 overrun = hrtimer_forward(timer, now, cfs_b->period);
2663 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2666 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2669 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2671 raw_spin_lock_init(&cfs_b->lock);
2673 cfs_b->quota = RUNTIME_INF;
2674 cfs_b->period = ns_to_ktime(default_cfs_period());
2676 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2677 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2678 cfs_b->period_timer.function = sched_cfs_period_timer;
2679 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2680 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2683 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2685 cfs_rq->runtime_enabled = 0;
2686 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2689 /* requires cfs_b->lock, may release to reprogram timer */
2690 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2693 * The timer may be active because we're trying to set a new bandwidth
2694 * period or because we're racing with the tear-down path
2695 * (timer_active==0 becomes visible before the hrtimer call-back
2696 * terminates). In either case we ensure that it's re-programmed
2698 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2699 raw_spin_unlock(&cfs_b->lock);
2700 /* ensure cfs_b->lock is available while we wait */
2701 hrtimer_cancel(&cfs_b->period_timer);
2703 raw_spin_lock(&cfs_b->lock);
2704 /* if someone else restarted the timer then we're done */
2705 if (cfs_b->timer_active)
2709 cfs_b->timer_active = 1;
2710 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2713 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2715 hrtimer_cancel(&cfs_b->period_timer);
2716 hrtimer_cancel(&cfs_b->slack_timer);
2719 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2721 struct cfs_rq *cfs_rq;
2723 for_each_leaf_cfs_rq(rq, cfs_rq) {
2724 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2726 if (!cfs_rq->runtime_enabled)
2730 * clock_task is not advancing so we just need to make sure
2731 * there's some valid quota amount
2733 cfs_rq->runtime_remaining = cfs_b->quota;
2734 if (cfs_rq_throttled(cfs_rq))
2735 unthrottle_cfs_rq(cfs_rq);
2739 #else /* CONFIG_CFS_BANDWIDTH */
2740 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2742 return rq_of(cfs_rq)->clock_task;
2745 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2746 unsigned long delta_exec) {}
2747 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2748 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2749 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2751 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2756 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2761 static inline int throttled_lb_pair(struct task_group *tg,
2762 int src_cpu, int dest_cpu)
2767 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2769 #ifdef CONFIG_FAIR_GROUP_SCHED
2770 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2773 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2777 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2778 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2780 #endif /* CONFIG_CFS_BANDWIDTH */
2782 /**************************************************
2783 * CFS operations on tasks:
2786 #ifdef CONFIG_SCHED_HRTICK
2787 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2789 struct sched_entity *se = &p->se;
2790 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2792 WARN_ON(task_rq(p) != rq);
2794 if (cfs_rq->nr_running > 1) {
2795 u64 slice = sched_slice(cfs_rq, se);
2796 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2797 s64 delta = slice - ran;
2806 * Don't schedule slices shorter than 10000ns, that just
2807 * doesn't make sense. Rely on vruntime for fairness.
2810 delta = max_t(s64, 10000LL, delta);
2812 hrtick_start(rq, delta);
2817 * called from enqueue/dequeue and updates the hrtick when the
2818 * current task is from our class and nr_running is low enough
2821 static void hrtick_update(struct rq *rq)
2823 struct task_struct *curr = rq->curr;
2825 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2828 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2829 hrtick_start_fair(rq, curr);
2831 #else /* !CONFIG_SCHED_HRTICK */
2833 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2837 static inline void hrtick_update(struct rq *rq)
2843 * The enqueue_task method is called before nr_running is
2844 * increased. Here we update the fair scheduling stats and
2845 * then put the task into the rbtree:
2848 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2850 struct cfs_rq *cfs_rq;
2851 struct sched_entity *se = &p->se;
2853 for_each_sched_entity(se) {
2856 cfs_rq = cfs_rq_of(se);
2857 enqueue_entity(cfs_rq, se, flags);
2860 * end evaluation on encountering a throttled cfs_rq
2862 * note: in the case of encountering a throttled cfs_rq we will
2863 * post the final h_nr_running increment below.
2865 if (cfs_rq_throttled(cfs_rq))
2867 cfs_rq->h_nr_running++;
2869 flags = ENQUEUE_WAKEUP;
2872 for_each_sched_entity(se) {
2873 cfs_rq = cfs_rq_of(se);
2874 cfs_rq->h_nr_running++;
2876 if (cfs_rq_throttled(cfs_rq))
2879 update_cfs_shares(cfs_rq);
2880 update_entity_load_avg(se, 1);
2884 update_rq_runnable_avg(rq, rq->nr_running);
2890 static void set_next_buddy(struct sched_entity *se);
2893 * The dequeue_task method is called before nr_running is
2894 * decreased. We remove the task from the rbtree and
2895 * update the fair scheduling stats:
2897 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2899 struct cfs_rq *cfs_rq;
2900 struct sched_entity *se = &p->se;
2901 int task_sleep = flags & DEQUEUE_SLEEP;
2903 for_each_sched_entity(se) {
2904 cfs_rq = cfs_rq_of(se);
2905 dequeue_entity(cfs_rq, se, flags);
2908 * end evaluation on encountering a throttled cfs_rq
2910 * note: in the case of encountering a throttled cfs_rq we will
2911 * post the final h_nr_running decrement below.
2913 if (cfs_rq_throttled(cfs_rq))
2915 cfs_rq->h_nr_running--;
2917 /* Don't dequeue parent if it has other entities besides us */
2918 if (cfs_rq->load.weight) {
2920 * Bias pick_next to pick a task from this cfs_rq, as
2921 * p is sleeping when it is within its sched_slice.
2923 if (task_sleep && parent_entity(se))
2924 set_next_buddy(parent_entity(se));
2926 /* avoid re-evaluating load for this entity */
2927 se = parent_entity(se);
2930 flags |= DEQUEUE_SLEEP;
2933 for_each_sched_entity(se) {
2934 cfs_rq = cfs_rq_of(se);
2935 cfs_rq->h_nr_running--;
2937 if (cfs_rq_throttled(cfs_rq))
2940 update_cfs_shares(cfs_rq);
2941 update_entity_load_avg(se, 1);
2946 update_rq_runnable_avg(rq, 1);
2952 /* Used instead of source_load when we know the type == 0 */
2953 static unsigned long weighted_cpuload(const int cpu)
2955 return cpu_rq(cpu)->load.weight;
2959 * Return a low guess at the load of a migration-source cpu weighted
2960 * according to the scheduling class and "nice" value.
2962 * We want to under-estimate the load of migration sources, to
2963 * balance conservatively.
2965 static unsigned long source_load(int cpu, int type)
2967 struct rq *rq = cpu_rq(cpu);
2968 unsigned long total = weighted_cpuload(cpu);
2970 if (type == 0 || !sched_feat(LB_BIAS))
2973 return min(rq->cpu_load[type-1], total);
2977 * Return a high guess at the load of a migration-target cpu weighted
2978 * according to the scheduling class and "nice" value.
2980 static unsigned long target_load(int cpu, int type)
2982 struct rq *rq = cpu_rq(cpu);
2983 unsigned long total = weighted_cpuload(cpu);
2985 if (type == 0 || !sched_feat(LB_BIAS))
2988 return max(rq->cpu_load[type-1], total);
2991 static unsigned long power_of(int cpu)
2993 return cpu_rq(cpu)->cpu_power;
2996 static unsigned long cpu_avg_load_per_task(int cpu)
2998 struct rq *rq = cpu_rq(cpu);
2999 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3002 return rq->load.weight / nr_running;
3008 static void task_waking_fair(struct task_struct *p)
3010 struct sched_entity *se = &p->se;
3011 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3014 #ifndef CONFIG_64BIT
3015 u64 min_vruntime_copy;
3018 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3020 min_vruntime = cfs_rq->min_vruntime;
3021 } while (min_vruntime != min_vruntime_copy);
3023 min_vruntime = cfs_rq->min_vruntime;
3026 se->vruntime -= min_vruntime;
3029 #ifdef CONFIG_FAIR_GROUP_SCHED
3031 * effective_load() calculates the load change as seen from the root_task_group
3033 * Adding load to a group doesn't make a group heavier, but can cause movement
3034 * of group shares between cpus. Assuming the shares were perfectly aligned one
3035 * can calculate the shift in shares.
3037 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3038 * on this @cpu and results in a total addition (subtraction) of @wg to the
3039 * total group weight.
3041 * Given a runqueue weight distribution (rw_i) we can compute a shares
3042 * distribution (s_i) using:
3044 * s_i = rw_i / \Sum rw_j (1)
3046 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3047 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3048 * shares distribution (s_i):
3050 * rw_i = { 2, 4, 1, 0 }
3051 * s_i = { 2/7, 4/7, 1/7, 0 }
3053 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3054 * task used to run on and the CPU the waker is running on), we need to
3055 * compute the effect of waking a task on either CPU and, in case of a sync
3056 * wakeup, compute the effect of the current task going to sleep.
3058 * So for a change of @wl to the local @cpu with an overall group weight change
3059 * of @wl we can compute the new shares distribution (s'_i) using:
3061 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3063 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3064 * differences in waking a task to CPU 0. The additional task changes the
3065 * weight and shares distributions like:
3067 * rw'_i = { 3, 4, 1, 0 }
3068 * s'_i = { 3/8, 4/8, 1/8, 0 }
3070 * We can then compute the difference in effective weight by using:
3072 * dw_i = S * (s'_i - s_i) (3)
3074 * Where 'S' is the group weight as seen by its parent.
3076 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3077 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3078 * 4/7) times the weight of the group.
3080 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3082 struct sched_entity *se = tg->se[cpu];
3084 if (!tg->parent) /* the trivial, non-cgroup case */
3087 for_each_sched_entity(se) {
3093 * W = @wg + \Sum rw_j
3095 W = wg + calc_tg_weight(tg, se->my_q);
3100 w = se->my_q->load.weight + wl;
3103 * wl = S * s'_i; see (2)
3106 wl = (w * tg->shares) / W;
3111 * Per the above, wl is the new se->load.weight value; since
3112 * those are clipped to [MIN_SHARES, ...) do so now. See
3113 * calc_cfs_shares().
3115 if (wl < MIN_SHARES)
3119 * wl = dw_i = S * (s'_i - s_i); see (3)
3121 wl -= se->load.weight;
3124 * Recursively apply this logic to all parent groups to compute
3125 * the final effective load change on the root group. Since
3126 * only the @tg group gets extra weight, all parent groups can
3127 * only redistribute existing shares. @wl is the shift in shares
3128 * resulting from this level per the above.
3137 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3138 unsigned long wl, unsigned long wg)
3145 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3147 s64 this_load, load;
3148 int idx, this_cpu, prev_cpu;
3149 unsigned long tl_per_task;
3150 struct task_group *tg;
3151 unsigned long weight;
3155 this_cpu = smp_processor_id();
3156 prev_cpu = task_cpu(p);
3157 load = source_load(prev_cpu, idx);
3158 this_load = target_load(this_cpu, idx);
3161 * If sync wakeup then subtract the (maximum possible)
3162 * effect of the currently running task from the load
3163 * of the current CPU:
3166 tg = task_group(current);
3167 weight = current->se.load.weight;
3169 this_load += effective_load(tg, this_cpu, -weight, -weight);
3170 load += effective_load(tg, prev_cpu, 0, -weight);
3174 weight = p->se.load.weight;
3177 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3178 * due to the sync cause above having dropped this_load to 0, we'll
3179 * always have an imbalance, but there's really nothing you can do
3180 * about that, so that's good too.
3182 * Otherwise check if either cpus are near enough in load to allow this
3183 * task to be woken on this_cpu.
3185 if (this_load > 0) {
3186 s64 this_eff_load, prev_eff_load;
3188 this_eff_load = 100;
3189 this_eff_load *= power_of(prev_cpu);
3190 this_eff_load *= this_load +
3191 effective_load(tg, this_cpu, weight, weight);
3193 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3194 prev_eff_load *= power_of(this_cpu);
3195 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3197 balanced = this_eff_load <= prev_eff_load;
3202 * If the currently running task will sleep within
3203 * a reasonable amount of time then attract this newly
3206 if (sync && balanced)
3209 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3210 tl_per_task = cpu_avg_load_per_task(this_cpu);
3213 (this_load <= load &&
3214 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3216 * This domain has SD_WAKE_AFFINE and
3217 * p is cache cold in this domain, and
3218 * there is no bad imbalance.
3220 schedstat_inc(sd, ttwu_move_affine);
3221 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3229 * find_idlest_group finds and returns the least busy CPU group within the
3232 static struct sched_group *
3233 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3234 int this_cpu, int load_idx)
3236 struct sched_group *idlest = NULL, *group = sd->groups;
3237 unsigned long min_load = ULONG_MAX, this_load = 0;
3238 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3241 unsigned long load, avg_load;
3245 /* Skip over this group if it has no CPUs allowed */
3246 if (!cpumask_intersects(sched_group_cpus(group),
3247 tsk_cpus_allowed(p)))
3250 local_group = cpumask_test_cpu(this_cpu,
3251 sched_group_cpus(group));
3253 /* Tally up the load of all CPUs in the group */
3256 for_each_cpu(i, sched_group_cpus(group)) {
3257 /* Bias balancing toward cpus of our domain */
3259 load = source_load(i, load_idx);
3261 load = target_load(i, load_idx);
3266 /* Adjust by relative CPU power of the group */
3267 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3270 this_load = avg_load;
3271 } else if (avg_load < min_load) {
3272 min_load = avg_load;
3275 } while (group = group->next, group != sd->groups);
3277 if (!idlest || 100*this_load < imbalance*min_load)
3283 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3286 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3288 unsigned long load, min_load = ULONG_MAX;
3292 /* Traverse only the allowed CPUs */
3293 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3294 load = weighted_cpuload(i);
3296 if (load < min_load || (load == min_load && i == this_cpu)) {
3306 * Try and locate an idle CPU in the sched_domain.
3308 static int select_idle_sibling(struct task_struct *p, int target)
3310 struct sched_domain *sd;
3311 struct sched_group *sg;
3312 int i = task_cpu(p);
3314 if (idle_cpu(target))
3318 * If the prevous cpu is cache affine and idle, don't be stupid.
3320 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3324 * Otherwise, iterate the domains and find an elegible idle cpu.
3326 sd = rcu_dereference(per_cpu(sd_llc, target));
3327 for_each_lower_domain(sd) {
3330 if (!cpumask_intersects(sched_group_cpus(sg),
3331 tsk_cpus_allowed(p)))
3334 for_each_cpu(i, sched_group_cpus(sg)) {
3335 if (i == target || !idle_cpu(i))
3339 target = cpumask_first_and(sched_group_cpus(sg),
3340 tsk_cpus_allowed(p));
3344 } while (sg != sd->groups);
3350 #ifdef CONFIG_SCHED_HMP
3352 * Heterogenous multiprocessor (HMP) optimizations
3354 * The cpu types are distinguished using a list of hmp_domains
3355 * which each represent one cpu type using a cpumask.
3356 * The list is assumed ordered by compute capacity with the
3357 * fastest domain first.
3359 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3361 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3363 /* Setup hmp_domains */
3364 static int __init hmp_cpu_mask_setup(void)
3367 struct hmp_domain *domain;
3368 struct list_head *pos;
3371 pr_debug("Initializing HMP scheduler:\n");
3373 /* Initialize hmp_domains using platform code */
3374 arch_get_hmp_domains(&hmp_domains);
3375 if (list_empty(&hmp_domains)) {
3376 pr_debug("HMP domain list is empty!\n");
3380 /* Print hmp_domains */
3382 list_for_each(pos, &hmp_domains) {
3383 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3384 cpulist_scnprintf(buf, 64, &domain->possible_cpus);
3385 pr_debug(" HMP domain %d: %s\n", dc, buf);
3387 for_each_cpu_mask(cpu, domain->possible_cpus) {
3388 per_cpu(hmp_cpu_domain, cpu) = domain;
3396 static struct hmp_domain *hmp_get_hmp_domain_for_cpu(int cpu)
3398 struct hmp_domain *domain;
3399 struct list_head *pos;
3401 list_for_each(pos, &hmp_domains) {
3402 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3403 if(cpumask_test_cpu(cpu, &domain->possible_cpus))
3409 static void hmp_online_cpu(int cpu)
3411 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3414 cpumask_set_cpu(cpu, &domain->cpus);
3417 static void hmp_offline_cpu(int cpu)
3419 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3422 cpumask_clear_cpu(cpu, &domain->cpus);
3426 * Migration thresholds should be in the range [0..1023]
3427 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3428 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3429 * The default values (512, 256) offer good responsiveness, but may need
3430 * tweaking suit particular needs.
3432 * hmp_up_prio: Only up migrate task with high priority (<hmp_up_prio)
3433 * hmp_next_up_threshold: Delay before next up migration (1024 ~= 1 ms)
3434 * hmp_next_down_threshold: Delay before next down migration (1024 ~= 1 ms)
3436 unsigned int hmp_up_threshold = 512;
3437 unsigned int hmp_down_threshold = 256;
3438 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
3439 unsigned int hmp_up_prio = NICE_TO_PRIO(CONFIG_SCHED_HMP_PRIO_FILTER_VAL);
3441 unsigned int hmp_next_up_threshold = 4096;
3442 unsigned int hmp_next_down_threshold = 4096;
3444 static unsigned int hmp_up_migration(int cpu, struct sched_entity *se);
3445 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3447 /* Check if cpu is in fastest hmp_domain */
3448 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3450 struct list_head *pos;
3452 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3453 return pos == hmp_domains.next;
3456 /* Check if cpu is in slowest hmp_domain */
3457 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3459 struct list_head *pos;
3461 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3462 return list_is_last(pos, &hmp_domains);
3465 /* Next (slower) hmp_domain relative to cpu */
3466 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3468 struct list_head *pos;
3470 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3471 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3474 /* Previous (faster) hmp_domain relative to cpu */
3475 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3477 struct list_head *pos;
3479 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3480 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3484 * Selects a cpu in previous (faster) hmp_domain
3485 * Note that cpumask_any_and() returns the first cpu in the cpumask
3487 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3490 return cpumask_any_and(&hmp_faster_domain(cpu)->cpus,
3491 tsk_cpus_allowed(tsk));
3495 * Selects a cpu in next (slower) hmp_domain
3496 * Note that cpumask_any_and() returns the first cpu in the cpumask
3498 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3501 return cpumask_any_and(&hmp_slower_domain(cpu)->cpus,
3502 tsk_cpus_allowed(tsk));
3505 static inline void hmp_next_up_delay(struct sched_entity *se, int cpu)
3507 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
3509 se->avg.hmp_last_up_migration = cfs_rq_clock_task(cfs_rq);
3510 se->avg.hmp_last_down_migration = 0;
3513 static inline void hmp_next_down_delay(struct sched_entity *se, int cpu)
3515 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
3517 se->avg.hmp_last_down_migration = cfs_rq_clock_task(cfs_rq);
3518 se->avg.hmp_last_up_migration = 0;
3520 #endif /* CONFIG_SCHED_HMP */
3523 * sched_balance_self: balance the current task (running on cpu) in domains
3524 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3527 * Balance, ie. select the least loaded group.
3529 * Returns the target CPU number, or the same CPU if no balancing is needed.
3531 * preempt must be disabled.
3534 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3536 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3537 int cpu = smp_processor_id();
3538 int prev_cpu = task_cpu(p);
3540 int want_affine = 0;
3541 int sync = wake_flags & WF_SYNC;
3543 if (p->nr_cpus_allowed == 1)
3546 if (sd_flag & SD_BALANCE_WAKE) {
3547 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3553 for_each_domain(cpu, tmp) {
3554 if (!(tmp->flags & SD_LOAD_BALANCE))
3558 * If both cpu and prev_cpu are part of this domain,
3559 * cpu is a valid SD_WAKE_AFFINE target.
3561 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3562 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3567 if (tmp->flags & sd_flag)
3572 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3575 new_cpu = select_idle_sibling(p, prev_cpu);
3580 int load_idx = sd->forkexec_idx;
3581 struct sched_group *group;
3584 if (!(sd->flags & sd_flag)) {
3589 if (sd_flag & SD_BALANCE_WAKE)
3590 load_idx = sd->wake_idx;
3592 group = find_idlest_group(sd, p, cpu, load_idx);
3598 new_cpu = find_idlest_cpu(group, p, cpu);
3599 if (new_cpu == -1 || new_cpu == cpu) {
3600 /* Now try balancing at a lower domain level of cpu */
3605 /* Now try balancing at a lower domain level of new_cpu */
3607 weight = sd->span_weight;
3609 for_each_domain(cpu, tmp) {
3610 if (weight <= tmp->span_weight)
3612 if (tmp->flags & sd_flag)
3615 /* while loop will break here if sd == NULL */
3620 #ifdef CONFIG_SCHED_HMP
3621 if (hmp_up_migration(prev_cpu, &p->se)) {
3622 new_cpu = hmp_select_faster_cpu(p, prev_cpu);
3623 hmp_next_up_delay(&p->se, new_cpu);
3624 trace_sched_hmp_migrate(p, new_cpu, 0);
3627 if (hmp_down_migration(prev_cpu, &p->se)) {
3628 new_cpu = hmp_select_slower_cpu(p, prev_cpu);
3629 hmp_next_down_delay(&p->se, new_cpu);
3630 trace_sched_hmp_migrate(p, new_cpu, 0);
3633 /* Make sure that the task stays in its previous hmp domain */
3634 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
3642 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3643 * removed when useful for applications beyond shares distribution (e.g.
3646 #ifdef CONFIG_FAIR_GROUP_SCHED
3648 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3649 * cfs_rq_of(p) references at time of call are still valid and identify the
3650 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3651 * other assumptions, including the state of rq->lock, should be made.
3654 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3656 struct sched_entity *se = &p->se;
3657 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3660 * Load tracking: accumulate removed load so that it can be processed
3661 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3662 * to blocked load iff they have a positive decay-count. It can never
3663 * be negative here since on-rq tasks have decay-count == 0.
3665 if (se->avg.decay_count) {
3666 se->avg.decay_count = -__synchronize_entity_decay(se);
3667 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3671 #endif /* CONFIG_SMP */
3673 static unsigned long
3674 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3676 unsigned long gran = sysctl_sched_wakeup_granularity;
3679 * Since its curr running now, convert the gran from real-time
3680 * to virtual-time in his units.
3682 * By using 'se' instead of 'curr' we penalize light tasks, so
3683 * they get preempted easier. That is, if 'se' < 'curr' then
3684 * the resulting gran will be larger, therefore penalizing the
3685 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3686 * be smaller, again penalizing the lighter task.
3688 * This is especially important for buddies when the leftmost
3689 * task is higher priority than the buddy.
3691 return calc_delta_fair(gran, se);
3695 * Should 'se' preempt 'curr'.
3709 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3711 s64 gran, vdiff = curr->vruntime - se->vruntime;
3716 gran = wakeup_gran(curr, se);
3723 static void set_last_buddy(struct sched_entity *se)
3725 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3728 for_each_sched_entity(se)
3729 cfs_rq_of(se)->last = se;
3732 static void set_next_buddy(struct sched_entity *se)
3734 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3737 for_each_sched_entity(se)
3738 cfs_rq_of(se)->next = se;
3741 static void set_skip_buddy(struct sched_entity *se)
3743 for_each_sched_entity(se)
3744 cfs_rq_of(se)->skip = se;
3748 * Preempt the current task with a newly woken task if needed:
3750 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3752 struct task_struct *curr = rq->curr;
3753 struct sched_entity *se = &curr->se, *pse = &p->se;
3754 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3755 int scale = cfs_rq->nr_running >= sched_nr_latency;
3756 int next_buddy_marked = 0;
3758 if (unlikely(se == pse))
3762 * This is possible from callers such as move_task(), in which we
3763 * unconditionally check_prempt_curr() after an enqueue (which may have
3764 * lead to a throttle). This both saves work and prevents false
3765 * next-buddy nomination below.
3767 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3770 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3771 set_next_buddy(pse);
3772 next_buddy_marked = 1;
3776 * We can come here with TIF_NEED_RESCHED already set from new task
3779 * Note: this also catches the edge-case of curr being in a throttled
3780 * group (e.g. via set_curr_task), since update_curr() (in the
3781 * enqueue of curr) will have resulted in resched being set. This
3782 * prevents us from potentially nominating it as a false LAST_BUDDY
3785 if (test_tsk_need_resched(curr))
3788 /* Idle tasks are by definition preempted by non-idle tasks. */
3789 if (unlikely(curr->policy == SCHED_IDLE) &&
3790 likely(p->policy != SCHED_IDLE))
3794 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3795 * is driven by the tick):
3797 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3800 find_matching_se(&se, &pse);
3801 update_curr(cfs_rq_of(se));
3803 if (wakeup_preempt_entity(se, pse) == 1) {
3805 * Bias pick_next to pick the sched entity that is
3806 * triggering this preemption.
3808 if (!next_buddy_marked)
3809 set_next_buddy(pse);
3818 * Only set the backward buddy when the current task is still
3819 * on the rq. This can happen when a wakeup gets interleaved
3820 * with schedule on the ->pre_schedule() or idle_balance()
3821 * point, either of which can * drop the rq lock.
3823 * Also, during early boot the idle thread is in the fair class,
3824 * for obvious reasons its a bad idea to schedule back to it.
3826 if (unlikely(!se->on_rq || curr == rq->idle))
3829 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3833 static struct task_struct *pick_next_task_fair(struct rq *rq)
3835 struct task_struct *p;
3836 struct cfs_rq *cfs_rq = &rq->cfs;
3837 struct sched_entity *se;
3839 if (!cfs_rq->nr_running)
3843 se = pick_next_entity(cfs_rq);
3844 set_next_entity(cfs_rq, se);
3845 cfs_rq = group_cfs_rq(se);
3849 if (hrtick_enabled(rq))
3850 hrtick_start_fair(rq, p);
3856 * Account for a descheduled task:
3858 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3860 struct sched_entity *se = &prev->se;
3861 struct cfs_rq *cfs_rq;
3863 for_each_sched_entity(se) {
3864 cfs_rq = cfs_rq_of(se);
3865 put_prev_entity(cfs_rq, se);
3870 * sched_yield() is very simple
3872 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3874 static void yield_task_fair(struct rq *rq)
3876 struct task_struct *curr = rq->curr;
3877 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3878 struct sched_entity *se = &curr->se;
3881 * Are we the only task in the tree?
3883 if (unlikely(rq->nr_running == 1))
3886 clear_buddies(cfs_rq, se);
3888 if (curr->policy != SCHED_BATCH) {
3889 update_rq_clock(rq);
3891 * Update run-time statistics of the 'current'.
3893 update_curr(cfs_rq);
3895 * Tell update_rq_clock() that we've just updated,
3896 * so we don't do microscopic update in schedule()
3897 * and double the fastpath cost.
3899 rq->skip_clock_update = 1;
3905 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3907 struct sched_entity *se = &p->se;
3909 /* throttled hierarchies are not runnable */
3910 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3913 /* Tell the scheduler that we'd really like pse to run next. */
3916 yield_task_fair(rq);
3922 /**************************************************
3923 * Fair scheduling class load-balancing methods.
3927 * The purpose of load-balancing is to achieve the same basic fairness the
3928 * per-cpu scheduler provides, namely provide a proportional amount of compute
3929 * time to each task. This is expressed in the following equation:
3931 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3933 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3934 * W_i,0 is defined as:
3936 * W_i,0 = \Sum_j w_i,j (2)
3938 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3939 * is derived from the nice value as per prio_to_weight[].
3941 * The weight average is an exponential decay average of the instantaneous
3944 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3946 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3947 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3948 * can also include other factors [XXX].
3950 * To achieve this balance we define a measure of imbalance which follows
3951 * directly from (1):
3953 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3955 * We them move tasks around to minimize the imbalance. In the continuous
3956 * function space it is obvious this converges, in the discrete case we get
3957 * a few fun cases generally called infeasible weight scenarios.
3960 * - infeasible weights;
3961 * - local vs global optima in the discrete case. ]
3966 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3967 * for all i,j solution, we create a tree of cpus that follows the hardware
3968 * topology where each level pairs two lower groups (or better). This results
3969 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3970 * tree to only the first of the previous level and we decrease the frequency
3971 * of load-balance at each level inv. proportional to the number of cpus in
3977 * \Sum { --- * --- * 2^i } = O(n) (5)
3979 * `- size of each group
3980 * | | `- number of cpus doing load-balance
3982 * `- sum over all levels
3984 * Coupled with a limit on how many tasks we can migrate every balance pass,
3985 * this makes (5) the runtime complexity of the balancer.
3987 * An important property here is that each CPU is still (indirectly) connected
3988 * to every other cpu in at most O(log n) steps:
3990 * The adjacency matrix of the resulting graph is given by:
3993 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3996 * And you'll find that:
3998 * A^(log_2 n)_i,j != 0 for all i,j (7)
4000 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4001 * The task movement gives a factor of O(m), giving a convergence complexity
4004 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4009 * In order to avoid CPUs going idle while there's still work to do, new idle
4010 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4011 * tree itself instead of relying on other CPUs to bring it work.
4013 * This adds some complexity to both (5) and (8) but it reduces the total idle
4021 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4024 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4029 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4031 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4033 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4036 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4037 * rewrite all of this once again.]
4040 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4042 #define LBF_ALL_PINNED 0x01
4043 #define LBF_NEED_BREAK 0x02
4044 #define LBF_SOME_PINNED 0x04
4047 struct sched_domain *sd;
4055 struct cpumask *dst_grpmask;
4057 enum cpu_idle_type idle;
4059 /* The set of CPUs under consideration for load-balancing */
4060 struct cpumask *cpus;
4065 unsigned int loop_break;
4066 unsigned int loop_max;
4070 * move_task - move a task from one runqueue to another runqueue.
4071 * Both runqueues must be locked.
4073 static void move_task(struct task_struct *p, struct lb_env *env)
4075 deactivate_task(env->src_rq, p, 0);
4076 set_task_cpu(p, env->dst_cpu);
4077 activate_task(env->dst_rq, p, 0);
4078 check_preempt_curr(env->dst_rq, p, 0);
4082 * Is this task likely cache-hot:
4085 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4089 if (p->sched_class != &fair_sched_class)
4092 if (unlikely(p->policy == SCHED_IDLE))
4096 * Buddy candidates are cache hot:
4098 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4099 (&p->se == cfs_rq_of(&p->se)->next ||
4100 &p->se == cfs_rq_of(&p->se)->last))
4103 if (sysctl_sched_migration_cost == -1)
4105 if (sysctl_sched_migration_cost == 0)
4108 delta = now - p->se.exec_start;
4110 return delta < (s64)sysctl_sched_migration_cost;
4114 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4117 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4119 int tsk_cache_hot = 0;
4121 * We do not migrate tasks that are:
4122 * 1) throttled_lb_pair, or
4123 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4124 * 3) running (obviously), or
4125 * 4) are cache-hot on their current CPU.
4127 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4130 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4133 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4136 * Remember if this task can be migrated to any other cpu in
4137 * our sched_group. We may want to revisit it if we couldn't
4138 * meet load balance goals by pulling other tasks on src_cpu.
4140 * Also avoid computing new_dst_cpu if we have already computed
4141 * one in current iteration.
4143 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
4146 /* Prevent to re-select dst_cpu via env's cpus */
4147 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4148 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4149 env->flags |= LBF_SOME_PINNED;
4150 env->new_dst_cpu = cpu;
4158 /* Record that we found atleast one task that could run on dst_cpu */
4159 env->flags &= ~LBF_ALL_PINNED;
4161 if (task_running(env->src_rq, p)) {
4162 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4167 * Aggressive migration if:
4168 * 1) task is cache cold, or
4169 * 2) too many balance attempts have failed.
4171 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
4172 if (!tsk_cache_hot ||
4173 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4175 if (tsk_cache_hot) {
4176 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4177 schedstat_inc(p, se.statistics.nr_forced_migrations);
4183 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4188 * move_one_task tries to move exactly one task from busiest to this_rq, as
4189 * part of active balancing operations within "domain".
4190 * Returns 1 if successful and 0 otherwise.
4192 * Called with both runqueues locked.
4194 static int move_one_task(struct lb_env *env)
4196 struct task_struct *p, *n;
4198 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4199 if (!can_migrate_task(p, env))
4204 * Right now, this is only the second place move_task()
4205 * is called, so we can safely collect move_task()
4206 * stats here rather than inside move_task().
4208 schedstat_inc(env->sd, lb_gained[env->idle]);
4214 static unsigned long task_h_load(struct task_struct *p);
4216 static const unsigned int sched_nr_migrate_break = 32;
4219 * move_tasks tries to move up to imbalance weighted load from busiest to
4220 * this_rq, as part of a balancing operation within domain "sd".
4221 * Returns 1 if successful and 0 otherwise.
4223 * Called with both runqueues locked.
4225 static int move_tasks(struct lb_env *env)
4227 struct list_head *tasks = &env->src_rq->cfs_tasks;
4228 struct task_struct *p;
4232 if (env->imbalance <= 0)
4235 while (!list_empty(tasks)) {
4236 p = list_first_entry(tasks, struct task_struct, se.group_node);
4239 /* We've more or less seen every task there is, call it quits */
4240 if (env->loop > env->loop_max)
4243 /* take a breather every nr_migrate tasks */
4244 if (env->loop > env->loop_break) {
4245 env->loop_break += sched_nr_migrate_break;
4246 env->flags |= LBF_NEED_BREAK;
4250 if (!can_migrate_task(p, env))
4253 load = task_h_load(p);
4255 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4258 if ((load / 2) > env->imbalance)
4263 env->imbalance -= load;
4265 #ifdef CONFIG_PREEMPT
4267 * NEWIDLE balancing is a source of latency, so preemptible
4268 * kernels will stop after the first task is pulled to minimize
4269 * the critical section.
4271 if (env->idle == CPU_NEWLY_IDLE)
4276 * We only want to steal up to the prescribed amount of
4279 if (env->imbalance <= 0)
4284 list_move_tail(&p->se.group_node, tasks);
4288 * Right now, this is one of only two places move_task() is called,
4289 * so we can safely collect move_task() stats here rather than
4290 * inside move_task().
4292 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4297 #ifdef CONFIG_FAIR_GROUP_SCHED
4299 * update tg->load_weight by folding this cpu's load_avg
4301 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4303 struct sched_entity *se = tg->se[cpu];
4304 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4306 /* throttled entities do not contribute to load */
4307 if (throttled_hierarchy(cfs_rq))
4310 update_cfs_rq_blocked_load(cfs_rq, 1);
4313 update_entity_load_avg(se, 1);
4315 * We pivot on our runnable average having decayed to zero for
4316 * list removal. This generally implies that all our children
4317 * have also been removed (modulo rounding error or bandwidth
4318 * control); however, such cases are rare and we can fix these
4321 * TODO: fix up out-of-order children on enqueue.
4323 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4324 list_del_leaf_cfs_rq(cfs_rq);
4326 struct rq *rq = rq_of(cfs_rq);
4327 update_rq_runnable_avg(rq, rq->nr_running);
4331 static void update_blocked_averages(int cpu)
4333 struct rq *rq = cpu_rq(cpu);
4334 struct cfs_rq *cfs_rq;
4335 unsigned long flags;
4337 raw_spin_lock_irqsave(&rq->lock, flags);
4338 update_rq_clock(rq);
4340 * Iterates the task_group tree in a bottom up fashion, see
4341 * list_add_leaf_cfs_rq() for details.
4343 for_each_leaf_cfs_rq(rq, cfs_rq) {
4345 * Note: We may want to consider periodically releasing
4346 * rq->lock about these updates so that creating many task
4347 * groups does not result in continually extending hold time.
4349 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4352 raw_spin_unlock_irqrestore(&rq->lock, flags);
4356 * Compute the cpu's hierarchical load factor for each task group.
4357 * This needs to be done in a top-down fashion because the load of a child
4358 * group is a fraction of its parents load.
4360 static int tg_load_down(struct task_group *tg, void *data)
4363 long cpu = (long)data;
4366 load = cpu_rq(cpu)->load.weight;
4368 load = tg->parent->cfs_rq[cpu]->h_load;
4369 load *= tg->se[cpu]->load.weight;
4370 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
4373 tg->cfs_rq[cpu]->h_load = load;
4378 static void update_h_load(long cpu)
4380 struct rq *rq = cpu_rq(cpu);
4381 unsigned long now = jiffies;
4383 if (rq->h_load_throttle == now)
4386 rq->h_load_throttle = now;
4389 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4393 static unsigned long task_h_load(struct task_struct *p)
4395 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4398 load = p->se.load.weight;
4399 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
4404 static inline void update_blocked_averages(int cpu)
4408 static inline void update_h_load(long cpu)
4412 static unsigned long task_h_load(struct task_struct *p)
4414 return p->se.load.weight;
4418 /********** Helpers for find_busiest_group ************************/
4420 * sd_lb_stats - Structure to store the statistics of a sched_domain
4421 * during load balancing.
4423 struct sd_lb_stats {
4424 struct sched_group *busiest; /* Busiest group in this sd */
4425 struct sched_group *this; /* Local group in this sd */
4426 unsigned long total_load; /* Total load of all groups in sd */
4427 unsigned long total_pwr; /* Total power of all groups in sd */
4428 unsigned long avg_load; /* Average load across all groups in sd */
4430 /** Statistics of this group */
4431 unsigned long this_load;
4432 unsigned long this_load_per_task;
4433 unsigned long this_nr_running;
4434 unsigned long this_has_capacity;
4435 unsigned int this_idle_cpus;
4437 /* Statistics of the busiest group */
4438 unsigned int busiest_idle_cpus;
4439 unsigned long max_load;
4440 unsigned long busiest_load_per_task;
4441 unsigned long busiest_nr_running;
4442 unsigned long busiest_group_capacity;
4443 unsigned long busiest_has_capacity;
4444 unsigned int busiest_group_weight;
4446 int group_imb; /* Is there imbalance in this sd */
4450 * sg_lb_stats - stats of a sched_group required for load_balancing
4452 struct sg_lb_stats {
4453 unsigned long avg_load; /*Avg load across the CPUs of the group */
4454 unsigned long group_load; /* Total load over the CPUs of the group */
4455 unsigned long sum_nr_running; /* Nr tasks running in the group */
4456 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4457 unsigned long group_capacity;
4458 unsigned long idle_cpus;
4459 unsigned long group_weight;
4460 int group_imb; /* Is there an imbalance in the group ? */
4461 int group_has_capacity; /* Is there extra capacity in the group? */
4465 * get_sd_load_idx - Obtain the load index for a given sched domain.
4466 * @sd: The sched_domain whose load_idx is to be obtained.
4467 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4469 static inline int get_sd_load_idx(struct sched_domain *sd,
4470 enum cpu_idle_type idle)
4476 load_idx = sd->busy_idx;
4479 case CPU_NEWLY_IDLE:
4480 load_idx = sd->newidle_idx;
4483 load_idx = sd->idle_idx;
4490 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4492 return SCHED_POWER_SCALE;
4495 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4497 return default_scale_freq_power(sd, cpu);
4500 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4502 unsigned long weight = sd->span_weight;
4503 unsigned long smt_gain = sd->smt_gain;
4510 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4512 return default_scale_smt_power(sd, cpu);
4515 static unsigned long scale_rt_power(int cpu)
4517 struct rq *rq = cpu_rq(cpu);
4518 u64 total, available, age_stamp, avg;
4521 * Since we're reading these variables without serialization make sure
4522 * we read them once before doing sanity checks on them.
4524 age_stamp = ACCESS_ONCE(rq->age_stamp);
4525 avg = ACCESS_ONCE(rq->rt_avg);
4527 total = sched_avg_period() + (rq->clock - age_stamp);
4529 if (unlikely(total < avg)) {
4530 /* Ensures that power won't end up being negative */
4533 available = total - avg;
4536 if (unlikely((s64)total < SCHED_POWER_SCALE))
4537 total = SCHED_POWER_SCALE;
4539 total >>= SCHED_POWER_SHIFT;
4541 return div_u64(available, total);
4544 static void update_cpu_power(struct sched_domain *sd, int cpu)
4546 unsigned long weight = sd->span_weight;
4547 unsigned long power = SCHED_POWER_SCALE;
4548 struct sched_group *sdg = sd->groups;
4550 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4551 if (sched_feat(ARCH_POWER))
4552 power *= arch_scale_smt_power(sd, cpu);
4554 power *= default_scale_smt_power(sd, cpu);
4556 power >>= SCHED_POWER_SHIFT;
4559 sdg->sgp->power_orig = power;
4561 if (sched_feat(ARCH_POWER))
4562 power *= arch_scale_freq_power(sd, cpu);
4564 power *= default_scale_freq_power(sd, cpu);
4566 power >>= SCHED_POWER_SHIFT;
4568 power *= scale_rt_power(cpu);
4569 power >>= SCHED_POWER_SHIFT;
4574 cpu_rq(cpu)->cpu_power = power;
4575 sdg->sgp->power = power;
4578 void update_group_power(struct sched_domain *sd, int cpu)
4580 struct sched_domain *child = sd->child;
4581 struct sched_group *group, *sdg = sd->groups;
4582 unsigned long power;
4583 unsigned long interval;
4585 interval = msecs_to_jiffies(sd->balance_interval);
4586 interval = clamp(interval, 1UL, max_load_balance_interval);
4587 sdg->sgp->next_update = jiffies + interval;
4590 update_cpu_power(sd, cpu);
4596 if (child->flags & SD_OVERLAP) {
4598 * SD_OVERLAP domains cannot assume that child groups
4599 * span the current group.
4602 for_each_cpu(cpu, sched_group_cpus(sdg))
4603 power += power_of(cpu);
4606 * !SD_OVERLAP domains can assume that child groups
4607 * span the current group.
4610 group = child->groups;
4612 power += group->sgp->power;
4613 group = group->next;
4614 } while (group != child->groups);
4617 sdg->sgp->power_orig = sdg->sgp->power = power;
4621 * Try and fix up capacity for tiny siblings, this is needed when
4622 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4623 * which on its own isn't powerful enough.
4625 * See update_sd_pick_busiest() and check_asym_packing().
4628 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4631 * Only siblings can have significantly less than SCHED_POWER_SCALE
4633 if (!(sd->flags & SD_SHARE_CPUPOWER))
4637 * If ~90% of the cpu_power is still there, we're good.
4639 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4646 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4647 * @env: The load balancing environment.
4648 * @group: sched_group whose statistics are to be updated.
4649 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4650 * @local_group: Does group contain this_cpu.
4651 * @balance: Should we balance.
4652 * @sgs: variable to hold the statistics for this group.
4654 static inline void update_sg_lb_stats(struct lb_env *env,
4655 struct sched_group *group, int load_idx,
4656 int local_group, int *balance, struct sg_lb_stats *sgs)
4658 unsigned long nr_running, max_nr_running, min_nr_running;
4659 unsigned long load, max_cpu_load, min_cpu_load;
4660 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4661 unsigned long avg_load_per_task = 0;
4665 balance_cpu = group_balance_cpu(group);
4667 /* Tally up the load of all CPUs in the group */
4669 min_cpu_load = ~0UL;
4671 min_nr_running = ~0UL;
4673 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4674 struct rq *rq = cpu_rq(i);
4676 nr_running = rq->nr_running;
4678 /* Bias balancing toward cpus of our domain */
4680 if (idle_cpu(i) && !first_idle_cpu &&
4681 cpumask_test_cpu(i, sched_group_mask(group))) {
4686 load = target_load(i, load_idx);
4688 load = source_load(i, load_idx);
4689 if (load > max_cpu_load)
4690 max_cpu_load = load;
4691 if (min_cpu_load > load)
4692 min_cpu_load = load;
4694 if (nr_running > max_nr_running)
4695 max_nr_running = nr_running;
4696 if (min_nr_running > nr_running)
4697 min_nr_running = nr_running;
4700 sgs->group_load += load;
4701 sgs->sum_nr_running += nr_running;
4702 sgs->sum_weighted_load += weighted_cpuload(i);
4708 * First idle cpu or the first cpu(busiest) in this sched group
4709 * is eligible for doing load balancing at this and above
4710 * domains. In the newly idle case, we will allow all the cpu's
4711 * to do the newly idle load balance.
4714 if (env->idle != CPU_NEWLY_IDLE) {
4715 if (balance_cpu != env->dst_cpu) {
4719 update_group_power(env->sd, env->dst_cpu);
4720 } else if (time_after_eq(jiffies, group->sgp->next_update))
4721 update_group_power(env->sd, env->dst_cpu);
4724 /* Adjust by relative CPU power of the group */
4725 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4728 * Consider the group unbalanced when the imbalance is larger
4729 * than the average weight of a task.
4731 * APZ: with cgroup the avg task weight can vary wildly and
4732 * might not be a suitable number - should we keep a
4733 * normalized nr_running number somewhere that negates
4736 if (sgs->sum_nr_running)
4737 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4739 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4740 (max_nr_running - min_nr_running) > 1)
4743 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4745 if (!sgs->group_capacity)
4746 sgs->group_capacity = fix_small_capacity(env->sd, group);
4747 sgs->group_weight = group->group_weight;
4749 if (sgs->group_capacity > sgs->sum_nr_running)
4750 sgs->group_has_capacity = 1;
4754 * update_sd_pick_busiest - return 1 on busiest group
4755 * @env: The load balancing environment.
4756 * @sds: sched_domain statistics
4757 * @sg: sched_group candidate to be checked for being the busiest
4758 * @sgs: sched_group statistics
4760 * Determine if @sg is a busier group than the previously selected
4763 static bool update_sd_pick_busiest(struct lb_env *env,
4764 struct sd_lb_stats *sds,
4765 struct sched_group *sg,
4766 struct sg_lb_stats *sgs)
4768 if (sgs->avg_load <= sds->max_load)
4771 if (sgs->sum_nr_running > sgs->group_capacity)
4778 * ASYM_PACKING needs to move all the work to the lowest
4779 * numbered CPUs in the group, therefore mark all groups
4780 * higher than ourself as busy.
4782 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4783 env->dst_cpu < group_first_cpu(sg)) {
4787 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4795 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4796 * @env: The load balancing environment.
4797 * @balance: Should we balance.
4798 * @sds: variable to hold the statistics for this sched_domain.
4800 static inline void update_sd_lb_stats(struct lb_env *env,
4801 int *balance, struct sd_lb_stats *sds)
4803 struct sched_domain *child = env->sd->child;
4804 struct sched_group *sg = env->sd->groups;
4805 struct sg_lb_stats sgs;
4806 int load_idx, prefer_sibling = 0;
4808 if (child && child->flags & SD_PREFER_SIBLING)
4811 load_idx = get_sd_load_idx(env->sd, env->idle);
4816 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4817 memset(&sgs, 0, sizeof(sgs));
4818 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4820 if (local_group && !(*balance))
4823 sds->total_load += sgs.group_load;
4824 sds->total_pwr += sg->sgp->power;
4827 * In case the child domain prefers tasks go to siblings
4828 * first, lower the sg capacity to one so that we'll try
4829 * and move all the excess tasks away. We lower the capacity
4830 * of a group only if the local group has the capacity to fit
4831 * these excess tasks, i.e. nr_running < group_capacity. The
4832 * extra check prevents the case where you always pull from the
4833 * heaviest group when it is already under-utilized (possible
4834 * with a large weight task outweighs the tasks on the system).
4836 if (prefer_sibling && !local_group && sds->this_has_capacity)
4837 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4840 sds->this_load = sgs.avg_load;
4842 sds->this_nr_running = sgs.sum_nr_running;
4843 sds->this_load_per_task = sgs.sum_weighted_load;
4844 sds->this_has_capacity = sgs.group_has_capacity;
4845 sds->this_idle_cpus = sgs.idle_cpus;
4846 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4847 sds->max_load = sgs.avg_load;
4849 sds->busiest_nr_running = sgs.sum_nr_running;
4850 sds->busiest_idle_cpus = sgs.idle_cpus;
4851 sds->busiest_group_capacity = sgs.group_capacity;
4852 sds->busiest_load_per_task = sgs.sum_weighted_load;
4853 sds->busiest_has_capacity = sgs.group_has_capacity;
4854 sds->busiest_group_weight = sgs.group_weight;
4855 sds->group_imb = sgs.group_imb;
4859 } while (sg != env->sd->groups);
4863 * check_asym_packing - Check to see if the group is packed into the
4866 * This is primarily intended to used at the sibling level. Some
4867 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4868 * case of POWER7, it can move to lower SMT modes only when higher
4869 * threads are idle. When in lower SMT modes, the threads will
4870 * perform better since they share less core resources. Hence when we
4871 * have idle threads, we want them to be the higher ones.
4873 * This packing function is run on idle threads. It checks to see if
4874 * the busiest CPU in this domain (core in the P7 case) has a higher
4875 * CPU number than the packing function is being run on. Here we are
4876 * assuming lower CPU number will be equivalent to lower a SMT thread
4879 * Returns 1 when packing is required and a task should be moved to
4880 * this CPU. The amount of the imbalance is returned in *imbalance.
4882 * @env: The load balancing environment.
4883 * @sds: Statistics of the sched_domain which is to be packed
4885 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4889 if (!(env->sd->flags & SD_ASYM_PACKING))
4895 busiest_cpu = group_first_cpu(sds->busiest);
4896 if (env->dst_cpu > busiest_cpu)
4899 env->imbalance = DIV_ROUND_CLOSEST(
4900 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4906 * fix_small_imbalance - Calculate the minor imbalance that exists
4907 * amongst the groups of a sched_domain, during
4909 * @env: The load balancing environment.
4910 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4913 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4915 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4916 unsigned int imbn = 2;
4917 unsigned long scaled_busy_load_per_task;
4919 if (sds->this_nr_running) {
4920 sds->this_load_per_task /= sds->this_nr_running;
4921 if (sds->busiest_load_per_task >
4922 sds->this_load_per_task)
4925 sds->this_load_per_task =
4926 cpu_avg_load_per_task(env->dst_cpu);
4929 scaled_busy_load_per_task = sds->busiest_load_per_task
4930 * SCHED_POWER_SCALE;
4931 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4933 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4934 (scaled_busy_load_per_task * imbn)) {
4935 env->imbalance = sds->busiest_load_per_task;
4940 * OK, we don't have enough imbalance to justify moving tasks,
4941 * however we may be able to increase total CPU power used by
4945 pwr_now += sds->busiest->sgp->power *
4946 min(sds->busiest_load_per_task, sds->max_load);
4947 pwr_now += sds->this->sgp->power *
4948 min(sds->this_load_per_task, sds->this_load);
4949 pwr_now /= SCHED_POWER_SCALE;
4951 /* Amount of load we'd subtract */
4952 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4953 sds->busiest->sgp->power;
4954 if (sds->max_load > tmp)
4955 pwr_move += sds->busiest->sgp->power *
4956 min(sds->busiest_load_per_task, sds->max_load - tmp);
4958 /* Amount of load we'd add */
4959 if (sds->max_load * sds->busiest->sgp->power <
4960 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4961 tmp = (sds->max_load * sds->busiest->sgp->power) /
4962 sds->this->sgp->power;
4964 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4965 sds->this->sgp->power;
4966 pwr_move += sds->this->sgp->power *
4967 min(sds->this_load_per_task, sds->this_load + tmp);
4968 pwr_move /= SCHED_POWER_SCALE;
4970 /* Move if we gain throughput */
4971 if (pwr_move > pwr_now)
4972 env->imbalance = sds->busiest_load_per_task;
4976 * calculate_imbalance - Calculate the amount of imbalance present within the
4977 * groups of a given sched_domain during load balance.
4978 * @env: load balance environment
4979 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4981 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4983 unsigned long max_pull, load_above_capacity = ~0UL;
4985 sds->busiest_load_per_task /= sds->busiest_nr_running;
4986 if (sds->group_imb) {
4987 sds->busiest_load_per_task =
4988 min(sds->busiest_load_per_task, sds->avg_load);
4992 * In the presence of smp nice balancing, certain scenarios can have
4993 * max load less than avg load(as we skip the groups at or below
4994 * its cpu_power, while calculating max_load..)
4996 if (sds->max_load < sds->avg_load) {
4998 return fix_small_imbalance(env, sds);
5001 if (!sds->group_imb) {
5003 * Don't want to pull so many tasks that a group would go idle.
5005 load_above_capacity = (sds->busiest_nr_running -
5006 sds->busiest_group_capacity);
5008 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5010 load_above_capacity /= sds->busiest->sgp->power;
5014 * We're trying to get all the cpus to the average_load, so we don't
5015 * want to push ourselves above the average load, nor do we wish to
5016 * reduce the max loaded cpu below the average load. At the same time,
5017 * we also don't want to reduce the group load below the group capacity
5018 * (so that we can implement power-savings policies etc). Thus we look
5019 * for the minimum possible imbalance.
5020 * Be careful of negative numbers as they'll appear as very large values
5021 * with unsigned longs.
5023 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
5025 /* How much load to actually move to equalise the imbalance */
5026 env->imbalance = min(max_pull * sds->busiest->sgp->power,
5027 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
5028 / SCHED_POWER_SCALE;
5031 * if *imbalance is less than the average load per runnable task
5032 * there is no guarantee that any tasks will be moved so we'll have
5033 * a think about bumping its value to force at least one task to be
5036 if (env->imbalance < sds->busiest_load_per_task)
5037 return fix_small_imbalance(env, sds);
5041 /******* find_busiest_group() helpers end here *********************/
5044 * find_busiest_group - Returns the busiest group within the sched_domain
5045 * if there is an imbalance. If there isn't an imbalance, and
5046 * the user has opted for power-savings, it returns a group whose
5047 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5048 * such a group exists.
5050 * Also calculates the amount of weighted load which should be moved
5051 * to restore balance.
5053 * @env: The load balancing environment.
5054 * @balance: Pointer to a variable indicating if this_cpu
5055 * is the appropriate cpu to perform load balancing at this_level.
5057 * Returns: - the busiest group if imbalance exists.
5058 * - If no imbalance and user has opted for power-savings balance,
5059 * return the least loaded group whose CPUs can be
5060 * put to idle by rebalancing its tasks onto our group.
5062 static struct sched_group *
5063 find_busiest_group(struct lb_env *env, int *balance)
5065 struct sd_lb_stats sds;
5067 memset(&sds, 0, sizeof(sds));
5070 * Compute the various statistics relavent for load balancing at
5073 update_sd_lb_stats(env, balance, &sds);
5076 * this_cpu is not the appropriate cpu to perform load balancing at
5082 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5083 check_asym_packing(env, &sds))
5086 /* There is no busy sibling group to pull tasks from */
5087 if (!sds.busiest || sds.busiest_nr_running == 0)
5090 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5093 * If the busiest group is imbalanced the below checks don't
5094 * work because they assumes all things are equal, which typically
5095 * isn't true due to cpus_allowed constraints and the like.
5100 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5101 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5102 !sds.busiest_has_capacity)
5106 * If the local group is more busy than the selected busiest group
5107 * don't try and pull any tasks.
5109 if (sds.this_load >= sds.max_load)
5113 * Don't pull any tasks if this group is already above the domain
5116 if (sds.this_load >= sds.avg_load)
5119 if (env->idle == CPU_IDLE) {
5121 * This cpu is idle. If the busiest group load doesn't
5122 * have more tasks than the number of available cpu's and
5123 * there is no imbalance between this and busiest group
5124 * wrt to idle cpu's, it is balanced.
5126 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
5127 sds.busiest_nr_running <= sds.busiest_group_weight)
5131 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5132 * imbalance_pct to be conservative.
5134 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
5139 /* Looks like there is an imbalance. Compute it */
5140 calculate_imbalance(env, &sds);
5150 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5152 static struct rq *find_busiest_queue(struct lb_env *env,
5153 struct sched_group *group)
5155 struct rq *busiest = NULL, *rq;
5156 unsigned long max_load = 0;
5159 for_each_cpu(i, sched_group_cpus(group)) {
5160 unsigned long power = power_of(i);
5161 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5166 capacity = fix_small_capacity(env->sd, group);
5168 if (!cpumask_test_cpu(i, env->cpus))
5172 wl = weighted_cpuload(i);
5175 * When comparing with imbalance, use weighted_cpuload()
5176 * which is not scaled with the cpu power.
5178 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5182 * For the load comparisons with the other cpu's, consider
5183 * the weighted_cpuload() scaled with the cpu power, so that
5184 * the load can be moved away from the cpu that is potentially
5185 * running at a lower capacity.
5187 wl = (wl * SCHED_POWER_SCALE) / power;
5189 if (wl > max_load) {
5199 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5200 * so long as it is large enough.
5202 #define MAX_PINNED_INTERVAL 512
5204 /* Working cpumask for load_balance and load_balance_newidle. */
5205 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5207 static int need_active_balance(struct lb_env *env)
5209 struct sched_domain *sd = env->sd;
5211 if (env->idle == CPU_NEWLY_IDLE) {
5214 * ASYM_PACKING needs to force migrate tasks from busy but
5215 * higher numbered CPUs in order to pack all tasks in the
5216 * lowest numbered CPUs.
5218 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5222 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5225 static int active_load_balance_cpu_stop(void *data);
5228 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5229 * tasks if there is an imbalance.
5231 static int load_balance(int this_cpu, struct rq *this_rq,
5232 struct sched_domain *sd, enum cpu_idle_type idle,
5235 int ld_moved, cur_ld_moved, active_balance = 0;
5236 struct sched_group *group;
5238 unsigned long flags;
5239 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5241 struct lb_env env = {
5243 .dst_cpu = this_cpu,
5245 .dst_grpmask = sched_group_cpus(sd->groups),
5247 .loop_break = sched_nr_migrate_break,
5252 * For NEWLY_IDLE load_balancing, we don't need to consider
5253 * other cpus in our group
5255 if (idle == CPU_NEWLY_IDLE)
5256 env.dst_grpmask = NULL;
5258 cpumask_copy(cpus, cpu_active_mask);
5260 schedstat_inc(sd, lb_count[idle]);
5263 group = find_busiest_group(&env, balance);
5269 schedstat_inc(sd, lb_nobusyg[idle]);
5273 busiest = find_busiest_queue(&env, group);
5275 schedstat_inc(sd, lb_nobusyq[idle]);
5279 BUG_ON(busiest == env.dst_rq);
5281 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5284 if (busiest->nr_running > 1) {
5286 * Attempt to move tasks. If find_busiest_group has found
5287 * an imbalance but busiest->nr_running <= 1, the group is
5288 * still unbalanced. ld_moved simply stays zero, so it is
5289 * correctly treated as an imbalance.
5291 env.flags |= LBF_ALL_PINNED;
5292 env.src_cpu = busiest->cpu;
5293 env.src_rq = busiest;
5294 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5296 update_h_load(env.src_cpu);
5298 local_irq_save(flags);
5299 double_rq_lock(env.dst_rq, busiest);
5302 * cur_ld_moved - load moved in current iteration
5303 * ld_moved - cumulative load moved across iterations
5305 cur_ld_moved = move_tasks(&env);
5306 ld_moved += cur_ld_moved;
5307 double_rq_unlock(env.dst_rq, busiest);
5308 local_irq_restore(flags);
5311 * some other cpu did the load balance for us.
5313 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5314 resched_cpu(env.dst_cpu);
5316 if (env.flags & LBF_NEED_BREAK) {
5317 env.flags &= ~LBF_NEED_BREAK;
5322 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5323 * us and move them to an alternate dst_cpu in our sched_group
5324 * where they can run. The upper limit on how many times we
5325 * iterate on same src_cpu is dependent on number of cpus in our
5328 * This changes load balance semantics a bit on who can move
5329 * load to a given_cpu. In addition to the given_cpu itself
5330 * (or a ilb_cpu acting on its behalf where given_cpu is
5331 * nohz-idle), we now have balance_cpu in a position to move
5332 * load to given_cpu. In rare situations, this may cause
5333 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5334 * _independently_ and at _same_ time to move some load to
5335 * given_cpu) causing exceess load to be moved to given_cpu.
5336 * This however should not happen so much in practice and
5337 * moreover subsequent load balance cycles should correct the
5338 * excess load moved.
5340 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5342 env.dst_rq = cpu_rq(env.new_dst_cpu);
5343 env.dst_cpu = env.new_dst_cpu;
5344 env.flags &= ~LBF_SOME_PINNED;
5346 env.loop_break = sched_nr_migrate_break;
5348 /* Prevent to re-select dst_cpu via env's cpus */
5349 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5352 * Go back to "more_balance" rather than "redo" since we
5353 * need to continue with same src_cpu.
5358 /* All tasks on this runqueue were pinned by CPU affinity */
5359 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5360 cpumask_clear_cpu(cpu_of(busiest), cpus);
5361 if (!cpumask_empty(cpus)) {
5363 env.loop_break = sched_nr_migrate_break;
5371 schedstat_inc(sd, lb_failed[idle]);
5373 * Increment the failure counter only on periodic balance.
5374 * We do not want newidle balance, which can be very
5375 * frequent, pollute the failure counter causing
5376 * excessive cache_hot migrations and active balances.
5378 if (idle != CPU_NEWLY_IDLE)
5379 sd->nr_balance_failed++;
5381 if (need_active_balance(&env)) {
5382 raw_spin_lock_irqsave(&busiest->lock, flags);
5384 /* don't kick the active_load_balance_cpu_stop,
5385 * if the curr task on busiest cpu can't be
5388 if (!cpumask_test_cpu(this_cpu,
5389 tsk_cpus_allowed(busiest->curr))) {
5390 raw_spin_unlock_irqrestore(&busiest->lock,
5392 env.flags |= LBF_ALL_PINNED;
5393 goto out_one_pinned;
5397 * ->active_balance synchronizes accesses to
5398 * ->active_balance_work. Once set, it's cleared
5399 * only after active load balance is finished.
5401 if (!busiest->active_balance) {
5402 busiest->active_balance = 1;
5403 busiest->push_cpu = this_cpu;
5406 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5408 if (active_balance) {
5409 stop_one_cpu_nowait(cpu_of(busiest),
5410 active_load_balance_cpu_stop, busiest,
5411 &busiest->active_balance_work);
5415 * We've kicked active balancing, reset the failure
5418 sd->nr_balance_failed = sd->cache_nice_tries+1;
5421 sd->nr_balance_failed = 0;
5423 if (likely(!active_balance)) {
5424 /* We were unbalanced, so reset the balancing interval */
5425 sd->balance_interval = sd->min_interval;
5428 * If we've begun active balancing, start to back off. This
5429 * case may not be covered by the all_pinned logic if there
5430 * is only 1 task on the busy runqueue (because we don't call
5433 if (sd->balance_interval < sd->max_interval)
5434 sd->balance_interval *= 2;
5440 schedstat_inc(sd, lb_balanced[idle]);
5442 sd->nr_balance_failed = 0;
5445 /* tune up the balancing interval */
5446 if (((env.flags & LBF_ALL_PINNED) &&
5447 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5448 (sd->balance_interval < sd->max_interval))
5449 sd->balance_interval *= 2;
5457 * idle_balance is called by schedule() if this_cpu is about to become
5458 * idle. Attempts to pull tasks from other CPUs.
5460 void idle_balance(int this_cpu, struct rq *this_rq)
5462 struct sched_domain *sd;
5463 int pulled_task = 0;
5464 unsigned long next_balance = jiffies + HZ;
5466 this_rq->idle_stamp = this_rq->clock;
5468 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5472 * Drop the rq->lock, but keep IRQ/preempt disabled.
5474 raw_spin_unlock(&this_rq->lock);
5476 update_blocked_averages(this_cpu);
5478 for_each_domain(this_cpu, sd) {
5479 unsigned long interval;
5482 if (!(sd->flags & SD_LOAD_BALANCE))
5485 if (sd->flags & SD_BALANCE_NEWIDLE) {
5486 /* If we've pulled tasks over stop searching: */
5487 pulled_task = load_balance(this_cpu, this_rq,
5488 sd, CPU_NEWLY_IDLE, &balance);
5491 interval = msecs_to_jiffies(sd->balance_interval);
5492 if (time_after(next_balance, sd->last_balance + interval))
5493 next_balance = sd->last_balance + interval;
5495 this_rq->idle_stamp = 0;
5501 raw_spin_lock(&this_rq->lock);
5503 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5505 * We are going idle. next_balance may be set based on
5506 * a busy processor. So reset next_balance.
5508 this_rq->next_balance = next_balance;
5513 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5514 * running tasks off the busiest CPU onto idle CPUs. It requires at
5515 * least 1 task to be running on each physical CPU where possible, and
5516 * avoids physical / logical imbalances.
5518 static int active_load_balance_cpu_stop(void *data)
5520 struct rq *busiest_rq = data;
5521 int busiest_cpu = cpu_of(busiest_rq);
5522 int target_cpu = busiest_rq->push_cpu;
5523 struct rq *target_rq = cpu_rq(target_cpu);
5524 struct sched_domain *sd;
5526 raw_spin_lock_irq(&busiest_rq->lock);
5528 /* make sure the requested cpu hasn't gone down in the meantime */
5529 if (unlikely(busiest_cpu != smp_processor_id() ||
5530 !busiest_rq->active_balance))
5533 /* Is there any task to move? */
5534 if (busiest_rq->nr_running <= 1)
5538 * This condition is "impossible", if it occurs
5539 * we need to fix it. Originally reported by
5540 * Bjorn Helgaas on a 128-cpu setup.
5542 BUG_ON(busiest_rq == target_rq);
5544 /* move a task from busiest_rq to target_rq */
5545 double_lock_balance(busiest_rq, target_rq);
5547 /* Search for an sd spanning us and the target CPU. */
5549 for_each_domain(target_cpu, sd) {
5550 if ((sd->flags & SD_LOAD_BALANCE) &&
5551 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5556 struct lb_env env = {
5558 .dst_cpu = target_cpu,
5559 .dst_rq = target_rq,
5560 .src_cpu = busiest_rq->cpu,
5561 .src_rq = busiest_rq,
5565 schedstat_inc(sd, alb_count);
5567 if (move_one_task(&env))
5568 schedstat_inc(sd, alb_pushed);
5570 schedstat_inc(sd, alb_failed);
5573 double_unlock_balance(busiest_rq, target_rq);
5575 busiest_rq->active_balance = 0;
5576 raw_spin_unlock_irq(&busiest_rq->lock);
5580 #ifdef CONFIG_NO_HZ_COMMON
5582 * idle load balancing details
5583 * - When one of the busy CPUs notice that there may be an idle rebalancing
5584 * needed, they will kick the idle load balancer, which then does idle
5585 * load balancing for all the idle CPUs.
5588 cpumask_var_t idle_cpus_mask;
5590 unsigned long next_balance; /* in jiffy units */
5591 } nohz ____cacheline_aligned;
5593 static inline int find_new_ilb(int call_cpu)
5595 int ilb = cpumask_first(nohz.idle_cpus_mask);
5597 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5604 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5605 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5606 * CPU (if there is one).
5608 static void nohz_balancer_kick(int cpu)
5612 nohz.next_balance++;
5614 ilb_cpu = find_new_ilb(cpu);
5616 if (ilb_cpu >= nr_cpu_ids)
5619 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5622 * Use smp_send_reschedule() instead of resched_cpu().
5623 * This way we generate a sched IPI on the target cpu which
5624 * is idle. And the softirq performing nohz idle load balance
5625 * will be run before returning from the IPI.
5627 smp_send_reschedule(ilb_cpu);
5631 static inline void nohz_balance_exit_idle(int cpu)
5633 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5634 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5635 atomic_dec(&nohz.nr_cpus);
5636 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5640 static inline void set_cpu_sd_state_busy(void)
5642 struct sched_domain *sd;
5643 int cpu = smp_processor_id();
5646 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5648 if (!sd || !sd->nohz_idle)
5652 for (; sd; sd = sd->parent)
5653 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5658 void set_cpu_sd_state_idle(void)
5660 struct sched_domain *sd;
5661 int cpu = smp_processor_id();
5664 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5666 if (!sd || sd->nohz_idle)
5670 for (; sd; sd = sd->parent)
5671 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5677 * This routine will record that the cpu is going idle with tick stopped.
5678 * This info will be used in performing idle load balancing in the future.
5680 void nohz_balance_enter_idle(int cpu)
5683 * If this cpu is going down, then nothing needs to be done.
5685 if (!cpu_active(cpu))
5688 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5691 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5692 atomic_inc(&nohz.nr_cpus);
5693 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5696 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5697 unsigned long action, void *hcpu)
5699 switch (action & ~CPU_TASKS_FROZEN) {
5701 nohz_balance_exit_idle(smp_processor_id());
5709 static DEFINE_SPINLOCK(balancing);
5712 * Scale the max load_balance interval with the number of CPUs in the system.
5713 * This trades load-balance latency on larger machines for less cross talk.
5715 void update_max_interval(void)
5717 max_load_balance_interval = HZ*num_online_cpus()/10;
5721 * It checks each scheduling domain to see if it is due to be balanced,
5722 * and initiates a balancing operation if so.
5724 * Balancing parameters are set up in init_sched_domains.
5726 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5729 struct rq *rq = cpu_rq(cpu);
5730 unsigned long interval;
5731 struct sched_domain *sd;
5732 /* Earliest time when we have to do rebalance again */
5733 unsigned long next_balance = jiffies + 60*HZ;
5734 int update_next_balance = 0;
5737 update_blocked_averages(cpu);
5740 for_each_domain(cpu, sd) {
5741 if (!(sd->flags & SD_LOAD_BALANCE))
5744 interval = sd->balance_interval;
5745 if (idle != CPU_IDLE)
5746 interval *= sd->busy_factor;
5748 /* scale ms to jiffies */
5749 interval = msecs_to_jiffies(interval);
5750 interval = clamp(interval, 1UL, max_load_balance_interval);
5752 need_serialize = sd->flags & SD_SERIALIZE;
5754 if (need_serialize) {
5755 if (!spin_trylock(&balancing))
5759 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5760 if (load_balance(cpu, rq, sd, idle, &balance)) {
5762 * The LBF_SOME_PINNED logic could have changed
5763 * env->dst_cpu, so we can't know our idle
5764 * state even if we migrated tasks. Update it.
5766 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5768 sd->last_balance = jiffies;
5771 spin_unlock(&balancing);
5773 if (time_after(next_balance, sd->last_balance + interval)) {
5774 next_balance = sd->last_balance + interval;
5775 update_next_balance = 1;
5779 * Stop the load balance at this level. There is another
5780 * CPU in our sched group which is doing load balancing more
5789 * next_balance will be updated only when there is a need.
5790 * When the cpu is attached to null domain for ex, it will not be
5793 if (likely(update_next_balance))
5794 rq->next_balance = next_balance;
5797 #ifdef CONFIG_NO_HZ_COMMON
5799 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5800 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5802 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5804 struct rq *this_rq = cpu_rq(this_cpu);
5808 if (idle != CPU_IDLE ||
5809 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5812 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5813 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5817 * If this cpu gets work to do, stop the load balancing
5818 * work being done for other cpus. Next load
5819 * balancing owner will pick it up.
5824 rq = cpu_rq(balance_cpu);
5826 raw_spin_lock_irq(&rq->lock);
5827 update_rq_clock(rq);
5828 update_idle_cpu_load(rq);
5829 raw_spin_unlock_irq(&rq->lock);
5831 rebalance_domains(balance_cpu, CPU_IDLE);
5833 if (time_after(this_rq->next_balance, rq->next_balance))
5834 this_rq->next_balance = rq->next_balance;
5836 nohz.next_balance = this_rq->next_balance;
5838 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5842 * Current heuristic for kicking the idle load balancer in the presence
5843 * of an idle cpu is the system.
5844 * - This rq has more than one task.
5845 * - At any scheduler domain level, this cpu's scheduler group has multiple
5846 * busy cpu's exceeding the group's power.
5847 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5848 * domain span are idle.
5850 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5852 unsigned long now = jiffies;
5853 struct sched_domain *sd;
5855 if (unlikely(idle_cpu(cpu)))
5859 * We may be recently in ticked or tickless idle mode. At the first
5860 * busy tick after returning from idle, we will update the busy stats.
5862 set_cpu_sd_state_busy();
5863 nohz_balance_exit_idle(cpu);
5866 * None are in tickless mode and hence no need for NOHZ idle load
5869 if (likely(!atomic_read(&nohz.nr_cpus)))
5872 if (time_before(now, nohz.next_balance))
5875 if (rq->nr_running >= 2)
5879 for_each_domain(cpu, sd) {
5880 struct sched_group *sg = sd->groups;
5881 struct sched_group_power *sgp = sg->sgp;
5882 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5884 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5885 goto need_kick_unlock;
5887 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5888 && (cpumask_first_and(nohz.idle_cpus_mask,
5889 sched_domain_span(sd)) < cpu))
5890 goto need_kick_unlock;
5892 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5904 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5907 #ifdef CONFIG_SCHED_HMP
5908 /* Check if task should migrate to a faster cpu */
5909 static unsigned int hmp_up_migration(int cpu, struct sched_entity *se)
5911 struct task_struct *p = task_of(se);
5912 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
5915 if (hmp_cpu_is_fastest(cpu))
5918 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
5919 /* Filter by task priority */
5920 if (p->prio >= hmp_up_prio)
5924 /* Let the task load settle before doing another up migration */
5925 now = cfs_rq_clock_task(cfs_rq);
5926 if (((now - se->avg.hmp_last_up_migration) >> 10)
5927 < hmp_next_up_threshold)
5930 if (cpumask_intersects(&hmp_faster_domain(cpu)->cpus,
5931 tsk_cpus_allowed(p))
5932 && se->avg.load_avg_ratio > hmp_up_threshold) {
5938 /* Check if task should migrate to a slower cpu */
5939 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
5941 struct task_struct *p = task_of(se);
5942 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
5945 if (hmp_cpu_is_slowest(cpu))
5948 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
5949 /* Filter by task priority */
5950 if ((p->prio >= hmp_up_prio) &&
5951 cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
5952 tsk_cpus_allowed(p))) {
5957 /* Let the task load settle before doing another down migration */
5958 now = cfs_rq_clock_task(cfs_rq);
5959 if (((now - se->avg.hmp_last_down_migration) >> 10)
5960 < hmp_next_down_threshold)
5963 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
5964 tsk_cpus_allowed(p))
5965 && se->avg.load_avg_ratio < hmp_down_threshold) {
5972 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5973 * Ideally this function should be merged with can_migrate_task() to avoid
5976 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
5978 int tsk_cache_hot = 0;
5981 * We do not migrate tasks that are:
5982 * 1) running (obviously), or
5983 * 2) cannot be migrated to this CPU due to cpus_allowed
5985 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5986 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5989 env->flags &= ~LBF_ALL_PINNED;
5991 if (task_running(env->src_rq, p)) {
5992 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5997 * Aggressive migration if:
5998 * 1) task is cache cold, or
5999 * 2) too many balance attempts have failed.
6002 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
6003 if (!tsk_cache_hot ||
6004 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6005 #ifdef CONFIG_SCHEDSTATS
6006 if (tsk_cache_hot) {
6007 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6008 schedstat_inc(p, se.statistics.nr_forced_migrations);
6018 * move_specific_task tries to move a specific task.
6019 * Returns 1 if successful and 0 otherwise.
6020 * Called with both runqueues locked.
6022 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
6024 struct task_struct *p, *n;
6026 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6027 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
6031 if (!hmp_can_migrate_task(p, env))
6033 /* Check if we found the right task */
6039 * Right now, this is only the third place move_task()
6040 * is called, so we can safely collect move_task()
6041 * stats here rather than inside move_task().
6043 schedstat_inc(env->sd, lb_gained[env->idle]);
6050 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
6051 * migrate a specific task from one runqueue to another.
6052 * hmp_force_up_migration uses this to push a currently running task
6054 * Based on active_load_balance_stop_cpu and can potentially be merged.
6056 static int hmp_active_task_migration_cpu_stop(void *data)
6058 struct rq *busiest_rq = data;
6059 struct task_struct *p = busiest_rq->migrate_task;
6060 int busiest_cpu = cpu_of(busiest_rq);
6061 int target_cpu = busiest_rq->push_cpu;
6062 struct rq *target_rq = cpu_rq(target_cpu);
6063 struct sched_domain *sd;
6065 raw_spin_lock_irq(&busiest_rq->lock);
6066 /* make sure the requested cpu hasn't gone down in the meantime */
6067 if (unlikely(busiest_cpu != smp_processor_id() ||
6068 !busiest_rq->active_balance)) {
6071 /* Is there any task to move? */
6072 if (busiest_rq->nr_running <= 1)
6074 /* Task has migrated meanwhile, abort forced migration */
6075 if (task_rq(p) != busiest_rq)
6078 * This condition is "impossible", if it occurs
6079 * we need to fix it. Originally reported by
6080 * Bjorn Helgaas on a 128-cpu setup.
6082 BUG_ON(busiest_rq == target_rq);
6084 /* move a task from busiest_rq to target_rq */
6085 double_lock_balance(busiest_rq, target_rq);
6087 /* Search for an sd spanning us and the target CPU. */
6089 for_each_domain(target_cpu, sd) {
6090 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6095 struct lb_env env = {
6097 .dst_cpu = target_cpu,
6098 .dst_rq = target_rq,
6099 .src_cpu = busiest_rq->cpu,
6100 .src_rq = busiest_rq,
6104 schedstat_inc(sd, alb_count);
6106 if (move_specific_task(&env, p))
6107 schedstat_inc(sd, alb_pushed);
6109 schedstat_inc(sd, alb_failed);
6112 double_unlock_balance(busiest_rq, target_rq);
6114 busiest_rq->active_balance = 0;
6115 raw_spin_unlock_irq(&busiest_rq->lock);
6119 static DEFINE_SPINLOCK(hmp_force_migration);
6122 * hmp_force_up_migration checks runqueues for tasks that need to
6123 * be actively migrated to a faster cpu.
6125 static void hmp_force_up_migration(int this_cpu)
6128 struct sched_entity *curr;
6130 unsigned long flags;
6132 struct task_struct *p;
6134 if (!spin_trylock(&hmp_force_migration))
6136 for_each_online_cpu(cpu) {
6138 target = cpu_rq(cpu);
6139 raw_spin_lock_irqsave(&target->lock, flags);
6140 curr = target->cfs.curr;
6142 raw_spin_unlock_irqrestore(&target->lock, flags);
6145 if (!entity_is_task(curr)) {
6146 struct cfs_rq *cfs_rq;
6148 cfs_rq = group_cfs_rq(curr);
6150 curr = cfs_rq->curr;
6151 cfs_rq = group_cfs_rq(curr);
6155 if (hmp_up_migration(cpu, curr)) {
6156 if (!target->active_balance) {
6157 target->active_balance = 1;
6158 target->push_cpu = hmp_select_faster_cpu(p, cpu);
6159 target->migrate_task = p;
6161 trace_sched_hmp_migrate(p, target->push_cpu, 1);
6162 hmp_next_up_delay(&p->se, target->push_cpu);
6165 raw_spin_unlock_irqrestore(&target->lock, flags);
6167 stop_one_cpu_nowait(cpu_of(target),
6168 hmp_active_task_migration_cpu_stop,
6169 target, &target->active_balance_work);
6171 spin_unlock(&hmp_force_migration);
6174 static void hmp_force_up_migration(int this_cpu) { }
6175 #endif /* CONFIG_SCHED_HMP */
6178 * run_rebalance_domains is triggered when needed from the scheduler tick.
6179 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6181 static void run_rebalance_domains(struct softirq_action *h)
6183 int this_cpu = smp_processor_id();
6184 struct rq *this_rq = cpu_rq(this_cpu);
6185 enum cpu_idle_type idle = this_rq->idle_balance ?
6186 CPU_IDLE : CPU_NOT_IDLE;
6188 hmp_force_up_migration(this_cpu);
6190 rebalance_domains(this_cpu, idle);
6193 * If this cpu has a pending nohz_balance_kick, then do the
6194 * balancing on behalf of the other idle cpus whose ticks are
6197 nohz_idle_balance(this_cpu, idle);
6200 static inline int on_null_domain(int cpu)
6202 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6206 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6208 void trigger_load_balance(struct rq *rq, int cpu)
6210 /* Don't need to rebalance while attached to NULL domain */
6211 if (time_after_eq(jiffies, rq->next_balance) &&
6212 likely(!on_null_domain(cpu)))
6213 raise_softirq(SCHED_SOFTIRQ);
6214 #ifdef CONFIG_NO_HZ_COMMON
6215 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6216 nohz_balancer_kick(cpu);
6220 static void rq_online_fair(struct rq *rq)
6222 #ifdef CONFIG_SCHED_HMP
6223 hmp_online_cpu(rq->cpu);
6228 static void rq_offline_fair(struct rq *rq)
6230 #ifdef CONFIG_SCHED_HMP
6231 hmp_offline_cpu(rq->cpu);
6235 /* Ensure any throttled groups are reachable by pick_next_task */
6236 unthrottle_offline_cfs_rqs(rq);
6239 #endif /* CONFIG_SMP */
6242 * scheduler tick hitting a task of our scheduling class:
6244 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6246 struct cfs_rq *cfs_rq;
6247 struct sched_entity *se = &curr->se;
6249 for_each_sched_entity(se) {
6250 cfs_rq = cfs_rq_of(se);
6251 entity_tick(cfs_rq, se, queued);
6254 if (sched_feat_numa(NUMA))
6255 task_tick_numa(rq, curr);
6257 update_rq_runnable_avg(rq, 1);
6261 * called on fork with the child task as argument from the parent's context
6262 * - child not yet on the tasklist
6263 * - preemption disabled
6265 static void task_fork_fair(struct task_struct *p)
6267 struct cfs_rq *cfs_rq;
6268 struct sched_entity *se = &p->se, *curr;
6269 int this_cpu = smp_processor_id();
6270 struct rq *rq = this_rq();
6271 unsigned long flags;
6273 raw_spin_lock_irqsave(&rq->lock, flags);
6275 update_rq_clock(rq);
6277 cfs_rq = task_cfs_rq(current);
6278 curr = cfs_rq->curr;
6280 if (unlikely(task_cpu(p) != this_cpu)) {
6282 __set_task_cpu(p, this_cpu);
6286 update_curr(cfs_rq);
6289 se->vruntime = curr->vruntime;
6290 place_entity(cfs_rq, se, 1);
6292 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6294 * Upon rescheduling, sched_class::put_prev_task() will place
6295 * 'current' within the tree based on its new key value.
6297 swap(curr->vruntime, se->vruntime);
6298 resched_task(rq->curr);
6301 se->vruntime -= cfs_rq->min_vruntime;
6303 raw_spin_unlock_irqrestore(&rq->lock, flags);
6307 * Priority of the task has changed. Check to see if we preempt
6311 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6317 * Reschedule if we are currently running on this runqueue and
6318 * our priority decreased, or if we are not currently running on
6319 * this runqueue and our priority is higher than the current's
6321 if (rq->curr == p) {
6322 if (p->prio > oldprio)
6323 resched_task(rq->curr);
6325 check_preempt_curr(rq, p, 0);
6328 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6330 struct sched_entity *se = &p->se;
6331 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6334 * Ensure the task's vruntime is normalized, so that when its
6335 * switched back to the fair class the enqueue_entity(.flags=0) will
6336 * do the right thing.
6338 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6339 * have normalized the vruntime, if it was !on_rq, then only when
6340 * the task is sleeping will it still have non-normalized vruntime.
6342 if (!se->on_rq && p->state != TASK_RUNNING) {
6344 * Fix up our vruntime so that the current sleep doesn't
6345 * cause 'unlimited' sleep bonus.
6347 place_entity(cfs_rq, se, 0);
6348 se->vruntime -= cfs_rq->min_vruntime;
6351 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
6353 * Remove our load from contribution when we leave sched_fair
6354 * and ensure we don't carry in an old decay_count if we
6357 if (p->se.avg.decay_count) {
6358 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
6359 __synchronize_entity_decay(&p->se);
6360 subtract_blocked_load_contrib(cfs_rq,
6361 p->se.avg.load_avg_contrib);
6367 * We switched to the sched_fair class.
6369 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6375 * We were most likely switched from sched_rt, so
6376 * kick off the schedule if running, otherwise just see
6377 * if we can still preempt the current task.
6380 resched_task(rq->curr);
6382 check_preempt_curr(rq, p, 0);
6385 /* Account for a task changing its policy or group.
6387 * This routine is mostly called to set cfs_rq->curr field when a task
6388 * migrates between groups/classes.
6390 static void set_curr_task_fair(struct rq *rq)
6392 struct sched_entity *se = &rq->curr->se;
6394 for_each_sched_entity(se) {
6395 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6397 set_next_entity(cfs_rq, se);
6398 /* ensure bandwidth has been allocated on our new cfs_rq */
6399 account_cfs_rq_runtime(cfs_rq, 0);
6403 void init_cfs_rq(struct cfs_rq *cfs_rq)
6405 cfs_rq->tasks_timeline = RB_ROOT;
6406 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6407 #ifndef CONFIG_64BIT
6408 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6410 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
6411 atomic64_set(&cfs_rq->decay_counter, 1);
6412 atomic64_set(&cfs_rq->removed_load, 0);
6416 #ifdef CONFIG_FAIR_GROUP_SCHED
6417 static void task_move_group_fair(struct task_struct *p, int on_rq)
6419 struct cfs_rq *cfs_rq;
6421 * If the task was not on the rq at the time of this cgroup movement
6422 * it must have been asleep, sleeping tasks keep their ->vruntime
6423 * absolute on their old rq until wakeup (needed for the fair sleeper
6424 * bonus in place_entity()).
6426 * If it was on the rq, we've just 'preempted' it, which does convert
6427 * ->vruntime to a relative base.
6429 * Make sure both cases convert their relative position when migrating
6430 * to another cgroup's rq. This does somewhat interfere with the
6431 * fair sleeper stuff for the first placement, but who cares.
6434 * When !on_rq, vruntime of the task has usually NOT been normalized.
6435 * But there are some cases where it has already been normalized:
6437 * - Moving a forked child which is waiting for being woken up by
6438 * wake_up_new_task().
6439 * - Moving a task which has been woken up by try_to_wake_up() and
6440 * waiting for actually being woken up by sched_ttwu_pending().
6442 * To prevent boost or penalty in the new cfs_rq caused by delta
6443 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6445 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6449 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6450 set_task_rq(p, task_cpu(p));
6452 cfs_rq = cfs_rq_of(&p->se);
6453 p->se.vruntime += cfs_rq->min_vruntime;
6456 * migrate_task_rq_fair() will have removed our previous
6457 * contribution, but we must synchronize for ongoing future
6460 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6461 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6466 void free_fair_sched_group(struct task_group *tg)
6470 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6472 for_each_possible_cpu(i) {
6474 kfree(tg->cfs_rq[i]);
6483 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6485 struct cfs_rq *cfs_rq;
6486 struct sched_entity *se;
6489 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6492 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6496 tg->shares = NICE_0_LOAD;
6498 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6500 for_each_possible_cpu(i) {
6501 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6502 GFP_KERNEL, cpu_to_node(i));
6506 se = kzalloc_node(sizeof(struct sched_entity),
6507 GFP_KERNEL, cpu_to_node(i));
6511 init_cfs_rq(cfs_rq);
6512 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6523 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6525 struct rq *rq = cpu_rq(cpu);
6526 unsigned long flags;
6529 * Only empty task groups can be destroyed; so we can speculatively
6530 * check on_list without danger of it being re-added.
6532 if (!tg->cfs_rq[cpu]->on_list)
6535 raw_spin_lock_irqsave(&rq->lock, flags);
6536 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6537 raw_spin_unlock_irqrestore(&rq->lock, flags);
6540 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6541 struct sched_entity *se, int cpu,
6542 struct sched_entity *parent)
6544 struct rq *rq = cpu_rq(cpu);
6548 init_cfs_rq_runtime(cfs_rq);
6550 tg->cfs_rq[cpu] = cfs_rq;
6553 /* se could be NULL for root_task_group */
6558 se->cfs_rq = &rq->cfs;
6560 se->cfs_rq = parent->my_q;
6563 update_load_set(&se->load, 0);
6564 se->parent = parent;
6567 static DEFINE_MUTEX(shares_mutex);
6569 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6572 unsigned long flags;
6575 * We can't change the weight of the root cgroup.
6580 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6582 mutex_lock(&shares_mutex);
6583 if (tg->shares == shares)
6586 tg->shares = shares;
6587 for_each_possible_cpu(i) {
6588 struct rq *rq = cpu_rq(i);
6589 struct sched_entity *se;
6592 /* Propagate contribution to hierarchy */
6593 raw_spin_lock_irqsave(&rq->lock, flags);
6594 for_each_sched_entity(se)
6595 update_cfs_shares(group_cfs_rq(se));
6596 raw_spin_unlock_irqrestore(&rq->lock, flags);
6600 mutex_unlock(&shares_mutex);
6603 #else /* CONFIG_FAIR_GROUP_SCHED */
6605 void free_fair_sched_group(struct task_group *tg) { }
6607 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6612 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6614 #endif /* CONFIG_FAIR_GROUP_SCHED */
6617 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6619 struct sched_entity *se = &task->se;
6620 unsigned int rr_interval = 0;
6623 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6626 if (rq->cfs.load.weight)
6627 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6633 * All the scheduling class methods:
6635 const struct sched_class fair_sched_class = {
6636 .next = &idle_sched_class,
6637 .enqueue_task = enqueue_task_fair,
6638 .dequeue_task = dequeue_task_fair,
6639 .yield_task = yield_task_fair,
6640 .yield_to_task = yield_to_task_fair,
6642 .check_preempt_curr = check_preempt_wakeup,
6644 .pick_next_task = pick_next_task_fair,
6645 .put_prev_task = put_prev_task_fair,
6648 .select_task_rq = select_task_rq_fair,
6649 #ifdef CONFIG_FAIR_GROUP_SCHED
6650 .migrate_task_rq = migrate_task_rq_fair,
6652 .rq_online = rq_online_fair,
6653 .rq_offline = rq_offline_fair,
6655 .task_waking = task_waking_fair,
6658 .set_curr_task = set_curr_task_fair,
6659 .task_tick = task_tick_fair,
6660 .task_fork = task_fork_fair,
6662 .prio_changed = prio_changed_fair,
6663 .switched_from = switched_from_fair,
6664 .switched_to = switched_to_fair,
6666 .get_rr_interval = get_rr_interval_fair,
6668 #ifdef CONFIG_FAIR_GROUP_SCHED
6669 .task_move_group = task_move_group_fair,
6673 #ifdef CONFIG_SCHED_DEBUG
6674 void print_cfs_stats(struct seq_file *m, int cpu)
6676 struct cfs_rq *cfs_rq;
6679 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6680 print_cfs_rq(m, cpu, cfs_rq);
6685 __init void init_sched_fair_class(void)
6688 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6690 #ifdef CONFIG_NO_HZ_COMMON
6691 nohz.next_balance = jiffies;
6692 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6693 cpu_notifier(sched_ilb_notifier, 0);
6696 #ifdef CONFIG_SCHED_HMP
6697 hmp_cpu_mask_setup();