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>
34 #ifdef CONFIG_HMP_VARIABLE_SCALE
35 #include <linux/sysfs.h>
36 #include <linux/vmalloc.h>
42 * Targeted preemption latency for CPU-bound tasks:
43 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
57 * The initial- and re-scaling of tunables is configurable
58 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
61 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
62 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
63 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
65 enum sched_tunable_scaling sysctl_sched_tunable_scaling
66 = SCHED_TUNABLESCALING_LOG;
69 * Minimal preemption granularity for CPU-bound tasks:
70 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
72 unsigned int sysctl_sched_min_granularity = 750000ULL;
73 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
76 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
78 static unsigned int sched_nr_latency = 8;
81 * After fork, child runs first. If set to 0 (default) then
82 * parent will (try to) run first.
84 unsigned int sysctl_sched_child_runs_first __read_mostly;
87 * SCHED_OTHER wake-up granularity.
88 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
90 * This option delays the preemption effects of decoupled workloads
91 * and reduces their over-scheduling. Synchronous workloads will still
92 * have immediate wakeup/sleep latencies.
94 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
95 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
97 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
100 * The exponential sliding window over which load is averaged for shares
104 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
106 #ifdef CONFIG_CFS_BANDWIDTH
108 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
109 * each time a cfs_rq requests quota.
111 * Note: in the case that the slice exceeds the runtime remaining (either due
112 * to consumption or the quota being specified to be smaller than the slice)
113 * we will always only issue the remaining available time.
115 * default: 5 msec, units: microseconds
117 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
121 * Increase the granularity value when there are more CPUs,
122 * because with more CPUs the 'effective latency' as visible
123 * to users decreases. But the relationship is not linear,
124 * so pick a second-best guess by going with the log2 of the
127 * This idea comes from the SD scheduler of Con Kolivas:
129 static int get_update_sysctl_factor(void)
131 unsigned int cpus = min_t(int, num_online_cpus(), 8);
134 switch (sysctl_sched_tunable_scaling) {
135 case SCHED_TUNABLESCALING_NONE:
138 case SCHED_TUNABLESCALING_LINEAR:
141 case SCHED_TUNABLESCALING_LOG:
143 factor = 1 + ilog2(cpus);
150 static void update_sysctl(void)
152 unsigned int factor = get_update_sysctl_factor();
154 #define SET_SYSCTL(name) \
155 (sysctl_##name = (factor) * normalized_sysctl_##name)
156 SET_SYSCTL(sched_min_granularity);
157 SET_SYSCTL(sched_latency);
158 SET_SYSCTL(sched_wakeup_granularity);
162 void sched_init_granularity(void)
167 #if BITS_PER_LONG == 32
168 # define WMULT_CONST (~0UL)
170 # define WMULT_CONST (1UL << 32)
173 #define WMULT_SHIFT 32
176 * Shift right and round:
178 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
181 * delta *= weight / lw
184 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
185 struct load_weight *lw)
190 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
191 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
192 * 2^SCHED_LOAD_RESOLUTION.
194 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
195 tmp = (u64)delta_exec * scale_load_down(weight);
197 tmp = (u64)delta_exec;
199 if (!lw->inv_weight) {
200 unsigned long w = scale_load_down(lw->weight);
202 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
204 else if (unlikely(!w))
205 lw->inv_weight = WMULT_CONST;
207 lw->inv_weight = WMULT_CONST / w;
211 * Check whether we'd overflow the 64-bit multiplication:
213 if (unlikely(tmp > WMULT_CONST))
214 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
217 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
219 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
223 const struct sched_class fair_sched_class;
225 /**************************************************************
226 * CFS operations on generic schedulable entities:
229 #ifdef CONFIG_FAIR_GROUP_SCHED
231 /* cpu runqueue to which this cfs_rq is attached */
232 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
237 /* An entity is a task if it doesn't "own" a runqueue */
238 #define entity_is_task(se) (!se->my_q)
240 static inline struct task_struct *task_of(struct sched_entity *se)
242 #ifdef CONFIG_SCHED_DEBUG
243 WARN_ON_ONCE(!entity_is_task(se));
245 return container_of(se, struct task_struct, se);
248 /* Walk up scheduling entities hierarchy */
249 #define for_each_sched_entity(se) \
250 for (; se; se = se->parent)
252 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
257 /* runqueue on which this entity is (to be) queued */
258 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
263 /* runqueue "owned" by this group */
264 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
269 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
272 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
274 if (!cfs_rq->on_list) {
276 * Ensure we either appear before our parent (if already
277 * enqueued) or force our parent to appear after us when it is
278 * enqueued. The fact that we always enqueue bottom-up
279 * reduces this to two cases.
281 if (cfs_rq->tg->parent &&
282 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
283 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
284 &rq_of(cfs_rq)->leaf_cfs_rq_list);
286 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
287 &rq_of(cfs_rq)->leaf_cfs_rq_list);
291 /* We should have no load, but we need to update last_decay. */
292 update_cfs_rq_blocked_load(cfs_rq, 0);
296 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
298 if (cfs_rq->on_list) {
299 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
304 /* Iterate thr' all leaf cfs_rq's on a runqueue */
305 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
306 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
308 /* Do the two (enqueued) entities belong to the same group ? */
310 is_same_group(struct sched_entity *se, struct sched_entity *pse)
312 if (se->cfs_rq == pse->cfs_rq)
318 static inline struct sched_entity *parent_entity(struct sched_entity *se)
323 /* return depth at which a sched entity is present in the hierarchy */
324 static inline int depth_se(struct sched_entity *se)
328 for_each_sched_entity(se)
335 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
337 int se_depth, pse_depth;
340 * preemption test can be made between sibling entities who are in the
341 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
342 * both tasks until we find their ancestors who are siblings of common
346 /* First walk up until both entities are at same depth */
347 se_depth = depth_se(*se);
348 pse_depth = depth_se(*pse);
350 while (se_depth > pse_depth) {
352 *se = parent_entity(*se);
355 while (pse_depth > se_depth) {
357 *pse = parent_entity(*pse);
360 while (!is_same_group(*se, *pse)) {
361 *se = parent_entity(*se);
362 *pse = parent_entity(*pse);
366 #else /* !CONFIG_FAIR_GROUP_SCHED */
368 static inline struct task_struct *task_of(struct sched_entity *se)
370 return container_of(se, struct task_struct, se);
373 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
375 return container_of(cfs_rq, struct rq, cfs);
378 #define entity_is_task(se) 1
380 #define for_each_sched_entity(se) \
381 for (; se; se = NULL)
383 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
385 return &task_rq(p)->cfs;
388 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
390 struct task_struct *p = task_of(se);
391 struct rq *rq = task_rq(p);
396 /* runqueue "owned" by this group */
397 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
402 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
406 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
410 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
411 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 is_same_group(struct sched_entity *se, struct sched_entity *pse)
419 static inline struct sched_entity *parent_entity(struct sched_entity *se)
425 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
429 #endif /* CONFIG_FAIR_GROUP_SCHED */
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
434 /**************************************************************
435 * Scheduling class tree data structure manipulation methods:
438 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
440 s64 delta = (s64)(vruntime - max_vruntime);
442 max_vruntime = vruntime;
447 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
449 s64 delta = (s64)(vruntime - min_vruntime);
451 min_vruntime = vruntime;
456 static inline int entity_before(struct sched_entity *a,
457 struct sched_entity *b)
459 return (s64)(a->vruntime - b->vruntime) < 0;
462 static void update_min_vruntime(struct cfs_rq *cfs_rq)
464 u64 vruntime = cfs_rq->min_vruntime;
467 vruntime = cfs_rq->curr->vruntime;
469 if (cfs_rq->rb_leftmost) {
470 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
475 vruntime = se->vruntime;
477 vruntime = min_vruntime(vruntime, se->vruntime);
480 /* ensure we never gain time by being placed backwards. */
481 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
484 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
489 * Enqueue an entity into the rb-tree:
491 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
493 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
494 struct rb_node *parent = NULL;
495 struct sched_entity *entry;
499 * Find the right place in the rbtree:
503 entry = rb_entry(parent, struct sched_entity, run_node);
505 * We dont care about collisions. Nodes with
506 * the same key stay together.
508 if (entity_before(se, entry)) {
509 link = &parent->rb_left;
511 link = &parent->rb_right;
517 * Maintain a cache of leftmost tree entries (it is frequently
521 cfs_rq->rb_leftmost = &se->run_node;
523 rb_link_node(&se->run_node, parent, link);
524 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
527 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 if (cfs_rq->rb_leftmost == &se->run_node) {
530 struct rb_node *next_node;
532 next_node = rb_next(&se->run_node);
533 cfs_rq->rb_leftmost = next_node;
536 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
539 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
541 struct rb_node *left = cfs_rq->rb_leftmost;
546 return rb_entry(left, struct sched_entity, run_node);
549 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
551 struct rb_node *next = rb_next(&se->run_node);
556 return rb_entry(next, struct sched_entity, run_node);
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
562 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
567 return rb_entry(last, struct sched_entity, run_node);
570 /**************************************************************
571 * Scheduling class statistics methods:
574 int sched_proc_update_handler(struct ctl_table *table, int write,
575 void __user *buffer, size_t *lenp,
578 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
579 int factor = get_update_sysctl_factor();
584 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
585 sysctl_sched_min_granularity);
587 #define WRT_SYSCTL(name) \
588 (normalized_sysctl_##name = sysctl_##name / (factor))
589 WRT_SYSCTL(sched_min_granularity);
590 WRT_SYSCTL(sched_latency);
591 WRT_SYSCTL(sched_wakeup_granularity);
601 static inline unsigned long
602 calc_delta_fair(unsigned long delta, struct sched_entity *se)
604 if (unlikely(se->load.weight != NICE_0_LOAD))
605 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
611 * The idea is to set a period in which each task runs once.
613 * When there are too many tasks (sched_nr_latency) we have to stretch
614 * this period because otherwise the slices get too small.
616 * p = (nr <= nl) ? l : l*nr/nl
618 static u64 __sched_period(unsigned long nr_running)
620 u64 period = sysctl_sched_latency;
621 unsigned long nr_latency = sched_nr_latency;
623 if (unlikely(nr_running > nr_latency)) {
624 period = sysctl_sched_min_granularity;
625 period *= nr_running;
632 * We calculate the wall-time slice from the period by taking a part
633 * proportional to the weight.
637 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
639 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
641 for_each_sched_entity(se) {
642 struct load_weight *load;
643 struct load_weight lw;
645 cfs_rq = cfs_rq_of(se);
646 load = &cfs_rq->load;
648 if (unlikely(!se->on_rq)) {
651 update_load_add(&lw, se->load.weight);
654 slice = calc_delta_mine(slice, se->load.weight, load);
660 * We calculate the vruntime slice of a to-be-inserted task.
664 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
666 return calc_delta_fair(sched_slice(cfs_rq, se), se);
670 * Update the current task's runtime statistics. Skip current tasks that
671 * are not in our scheduling class.
674 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
675 unsigned long delta_exec)
677 unsigned long delta_exec_weighted;
679 schedstat_set(curr->statistics.exec_max,
680 max((u64)delta_exec, curr->statistics.exec_max));
682 curr->sum_exec_runtime += delta_exec;
683 schedstat_add(cfs_rq, exec_clock, delta_exec);
684 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
686 curr->vruntime += delta_exec_weighted;
687 update_min_vruntime(cfs_rq);
690 static void update_curr(struct cfs_rq *cfs_rq)
692 struct sched_entity *curr = cfs_rq->curr;
693 u64 now = rq_of(cfs_rq)->clock_task;
694 unsigned long delta_exec;
700 * Get the amount of time the current task was running
701 * since the last time we changed load (this cannot
702 * overflow on 32 bits):
704 delta_exec = (unsigned long)(now - curr->exec_start);
708 __update_curr(cfs_rq, curr, delta_exec);
709 curr->exec_start = now;
711 if (entity_is_task(curr)) {
712 struct task_struct *curtask = task_of(curr);
714 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
715 cpuacct_charge(curtask, delta_exec);
716 account_group_exec_runtime(curtask, delta_exec);
719 account_cfs_rq_runtime(cfs_rq, delta_exec);
723 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
725 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
729 * Task is being enqueued - update stats:
731 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
734 * Are we enqueueing a waiting task? (for current tasks
735 * a dequeue/enqueue event is a NOP)
737 if (se != cfs_rq->curr)
738 update_stats_wait_start(cfs_rq, se);
742 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
745 rq_of(cfs_rq)->clock - se->statistics.wait_start));
746 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
747 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
748 rq_of(cfs_rq)->clock - se->statistics.wait_start);
749 #ifdef CONFIG_SCHEDSTATS
750 if (entity_is_task(se)) {
751 trace_sched_stat_wait(task_of(se),
752 rq_of(cfs_rq)->clock - se->statistics.wait_start);
755 schedstat_set(se->statistics.wait_start, 0);
759 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
762 * Mark the end of the wait period if dequeueing a
765 if (se != cfs_rq->curr)
766 update_stats_wait_end(cfs_rq, se);
770 * We are picking a new current task - update its stats:
773 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
776 * We are starting a new run period:
778 se->exec_start = rq_of(cfs_rq)->clock_task;
781 /**************************************************
782 * Scheduling class queueing methods:
785 #ifdef CONFIG_NUMA_BALANCING
787 * numa task sample period in ms
789 unsigned int sysctl_numa_balancing_scan_period_min = 100;
790 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
791 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
793 /* Portion of address space to scan in MB */
794 unsigned int sysctl_numa_balancing_scan_size = 256;
796 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
797 unsigned int sysctl_numa_balancing_scan_delay = 1000;
799 static void task_numa_placement(struct task_struct *p)
803 if (!p->mm) /* for example, ksmd faulting in a user's mm */
805 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
806 if (p->numa_scan_seq == seq)
808 p->numa_scan_seq = seq;
810 /* FIXME: Scheduling placement policy hints go here */
814 * Got a PROT_NONE fault for a page on @node.
816 void task_numa_fault(int node, int pages, bool migrated)
818 struct task_struct *p = current;
820 if (!sched_feat_numa(NUMA))
823 /* FIXME: Allocate task-specific structure for placement policy here */
826 * If pages are properly placed (did not migrate) then scan slower.
827 * This is reset periodically in case of phase changes
830 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
831 p->numa_scan_period + jiffies_to_msecs(10));
833 task_numa_placement(p);
836 static void reset_ptenuma_scan(struct task_struct *p)
838 ACCESS_ONCE(p->mm->numa_scan_seq)++;
839 p->mm->numa_scan_offset = 0;
843 * The expensive part of numa migration is done from task_work context.
844 * Triggered from task_tick_numa().
846 void task_numa_work(struct callback_head *work)
848 unsigned long migrate, next_scan, now = jiffies;
849 struct task_struct *p = current;
850 struct mm_struct *mm = p->mm;
851 struct vm_area_struct *vma;
852 unsigned long start, end;
855 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
857 work->next = work; /* protect against double add */
859 * Who cares about NUMA placement when they're dying.
861 * NOTE: make sure not to dereference p->mm before this check,
862 * exit_task_work() happens _after_ exit_mm() so we could be called
863 * without p->mm even though we still had it when we enqueued this
866 if (p->flags & PF_EXITING)
870 * We do not care about task placement until a task runs on a node
871 * other than the first one used by the address space. This is
872 * largely because migrations are driven by what CPU the task
873 * is running on. If it's never scheduled on another node, it'll
874 * not migrate so why bother trapping the fault.
876 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
877 mm->first_nid = numa_node_id();
878 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
879 /* Are we running on a new node yet? */
880 if (numa_node_id() == mm->first_nid &&
881 !sched_feat_numa(NUMA_FORCE))
884 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
888 * Reset the scan period if enough time has gone by. Objective is that
889 * scanning will be reduced if pages are properly placed. As tasks
890 * can enter different phases this needs to be re-examined. Lacking
891 * proper tracking of reference behaviour, this blunt hammer is used.
893 migrate = mm->numa_next_reset;
894 if (time_after(now, migrate)) {
895 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
896 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
897 xchg(&mm->numa_next_reset, next_scan);
901 * Enforce maximal scan/migration frequency..
903 migrate = mm->numa_next_scan;
904 if (time_before(now, migrate))
907 if (p->numa_scan_period == 0)
908 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
910 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
911 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
915 * Do not set pte_numa if the current running node is rate-limited.
916 * This loses statistics on the fault but if we are unwilling to
917 * migrate to this node, it is less likely we can do useful work
919 if (migrate_ratelimited(numa_node_id()))
922 start = mm->numa_scan_offset;
923 pages = sysctl_numa_balancing_scan_size;
924 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
928 down_read(&mm->mmap_sem);
929 vma = find_vma(mm, start);
931 reset_ptenuma_scan(p);
935 for (; vma; vma = vma->vm_next) {
936 if (!vma_migratable(vma))
939 /* Skip small VMAs. They are not likely to be of relevance */
940 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
944 start = max(start, vma->vm_start);
945 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
946 end = min(end, vma->vm_end);
947 pages -= change_prot_numa(vma, start, end);
952 } while (end != vma->vm_end);
957 * It is possible to reach the end of the VMA list but the last few VMAs are
958 * not guaranteed to the vma_migratable. If they are not, we would find the
959 * !migratable VMA on the next scan but not reset the scanner to the start
963 mm->numa_scan_offset = start;
965 reset_ptenuma_scan(p);
966 up_read(&mm->mmap_sem);
970 * Drive the periodic memory faults..
972 void task_tick_numa(struct rq *rq, struct task_struct *curr)
974 struct callback_head *work = &curr->numa_work;
978 * We don't care about NUMA placement if we don't have memory.
980 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
984 * Using runtime rather than walltime has the dual advantage that
985 * we (mostly) drive the selection from busy threads and that the
986 * task needs to have done some actual work before we bother with
989 now = curr->se.sum_exec_runtime;
990 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
992 if (now - curr->node_stamp > period) {
993 if (!curr->node_stamp)
994 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
995 curr->node_stamp = now;
997 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
998 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
999 task_work_add(curr, work, true);
1004 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1007 #endif /* CONFIG_NUMA_BALANCING */
1010 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1012 update_load_add(&cfs_rq->load, se->load.weight);
1013 if (!parent_entity(se))
1014 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1016 if (entity_is_task(se))
1017 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1019 cfs_rq->nr_running++;
1023 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1025 update_load_sub(&cfs_rq->load, se->load.weight);
1026 if (!parent_entity(se))
1027 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1028 if (entity_is_task(se))
1029 list_del_init(&se->group_node);
1030 cfs_rq->nr_running--;
1033 #ifdef CONFIG_FAIR_GROUP_SCHED
1035 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1040 * Use this CPU's actual weight instead of the last load_contribution
1041 * to gain a more accurate current total weight. See
1042 * update_cfs_rq_load_contribution().
1044 tg_weight = atomic64_read(&tg->load_avg);
1045 tg_weight -= cfs_rq->tg_load_contrib;
1046 tg_weight += cfs_rq->load.weight;
1051 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1053 long tg_weight, load, shares;
1055 tg_weight = calc_tg_weight(tg, cfs_rq);
1056 load = cfs_rq->load.weight;
1058 shares = (tg->shares * load);
1060 shares /= tg_weight;
1062 if (shares < MIN_SHARES)
1063 shares = MIN_SHARES;
1064 if (shares > tg->shares)
1065 shares = tg->shares;
1069 # else /* CONFIG_SMP */
1070 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1074 # endif /* CONFIG_SMP */
1075 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1076 unsigned long weight)
1079 /* commit outstanding execution time */
1080 if (cfs_rq->curr == se)
1081 update_curr(cfs_rq);
1082 account_entity_dequeue(cfs_rq, se);
1085 update_load_set(&se->load, weight);
1088 account_entity_enqueue(cfs_rq, se);
1091 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1093 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1095 struct task_group *tg;
1096 struct sched_entity *se;
1100 se = tg->se[cpu_of(rq_of(cfs_rq))];
1101 if (!se || throttled_hierarchy(cfs_rq))
1104 if (likely(se->load.weight == tg->shares))
1107 shares = calc_cfs_shares(cfs_rq, tg);
1109 reweight_entity(cfs_rq_of(se), se, shares);
1111 #else /* CONFIG_FAIR_GROUP_SCHED */
1112 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1115 #endif /* CONFIG_FAIR_GROUP_SCHED */
1117 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1118 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1120 * We choose a half-life close to 1 scheduling period.
1121 * Note: The tables below are dependent on this value.
1123 #define LOAD_AVG_PERIOD 32
1124 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1125 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1127 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1128 static const u32 runnable_avg_yN_inv[] = {
1129 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1130 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1131 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1132 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1133 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1134 0x85aac367, 0x82cd8698,
1138 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1139 * over-estimates when re-combining.
1141 static const u32 runnable_avg_yN_sum[] = {
1142 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1143 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1144 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1149 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1151 static __always_inline u64 decay_load(u64 val, u64 n)
1153 unsigned int local_n;
1157 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1160 /* after bounds checking we can collapse to 32-bit */
1164 * As y^PERIOD = 1/2, we can combine
1165 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1166 * With a look-up table which covers k^n (n<PERIOD)
1168 * To achieve constant time decay_load.
1170 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1171 val >>= local_n / LOAD_AVG_PERIOD;
1172 local_n %= LOAD_AVG_PERIOD;
1175 val *= runnable_avg_yN_inv[local_n];
1176 /* We don't use SRR here since we always want to round down. */
1181 * For updates fully spanning n periods, the contribution to runnable
1182 * average will be: \Sum 1024*y^n
1184 * We can compute this reasonably efficiently by combining:
1185 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1187 static u32 __compute_runnable_contrib(u64 n)
1191 if (likely(n <= LOAD_AVG_PERIOD))
1192 return runnable_avg_yN_sum[n];
1193 else if (unlikely(n >= LOAD_AVG_MAX_N))
1194 return LOAD_AVG_MAX;
1196 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1198 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1199 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1201 n -= LOAD_AVG_PERIOD;
1202 } while (n > LOAD_AVG_PERIOD);
1204 contrib = decay_load(contrib, n);
1205 return contrib + runnable_avg_yN_sum[n];
1208 #ifdef CONFIG_HMP_VARIABLE_SCALE
1209 static u64 hmp_variable_scale_convert(u64 delta);
1211 /* We can represent the historical contribution to runnable average as the
1212 * coefficients of a geometric series. To do this we sub-divide our runnable
1213 * history into segments of approximately 1ms (1024us); label the segment that
1214 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1216 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1218 * (now) (~1ms ago) (~2ms ago)
1220 * Let u_i denote the fraction of p_i that the entity was runnable.
1222 * We then designate the fractions u_i as our co-efficients, yielding the
1223 * following representation of historical load:
1224 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1226 * We choose y based on the with of a reasonably scheduling period, fixing:
1229 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1230 * approximately half as much as the contribution to load within the last ms
1233 * When a period "rolls over" and we have new u_0`, multiplying the previous
1234 * sum again by y is sufficient to update:
1235 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1236 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1238 static __always_inline int __update_entity_runnable_avg(u64 now,
1239 struct sched_avg *sa,
1244 u32 runnable_contrib;
1245 int delta_w, decayed = 0;
1247 delta = now - sa->last_runnable_update;
1248 #ifdef CONFIG_HMP_VARIABLE_SCALE
1249 delta = hmp_variable_scale_convert(delta);
1252 * This should only happen when time goes backwards, which it
1253 * unfortunately does during sched clock init when we swap over to TSC.
1255 if ((s64)delta < 0) {
1256 sa->last_runnable_update = now;
1261 * Use 1024ns as the unit of measurement since it's a reasonable
1262 * approximation of 1us and fast to compute.
1267 sa->last_runnable_update = now;
1269 /* delta_w is the amount already accumulated against our next period */
1270 delta_w = sa->runnable_avg_period % 1024;
1271 if (delta + delta_w >= 1024) {
1272 /* period roll-over */
1276 * Now that we know we're crossing a period boundary, figure
1277 * out how much from delta we need to complete the current
1278 * period and accrue it.
1280 delta_w = 1024 - delta_w;
1282 sa->runnable_avg_sum += delta_w;
1284 sa->usage_avg_sum += delta_w;
1285 sa->runnable_avg_period += delta_w;
1289 /* Figure out how many additional periods this update spans */
1290 periods = delta / 1024;
1293 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1295 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1297 sa->usage_avg_sum = decay_load(sa->usage_avg_sum, periods + 1);
1299 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1300 runnable_contrib = __compute_runnable_contrib(periods);
1302 sa->runnable_avg_sum += runnable_contrib;
1304 sa->usage_avg_sum += runnable_contrib;
1305 sa->runnable_avg_period += runnable_contrib;
1308 /* Remainder of delta accrued against u_0` */
1310 sa->runnable_avg_sum += delta;
1312 sa->usage_avg_sum += delta;
1313 sa->runnable_avg_period += delta;
1318 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1319 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1321 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1322 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1324 decays -= se->avg.decay_count;
1328 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1329 se->avg.decay_count = 0;
1334 #ifdef CONFIG_FAIR_GROUP_SCHED
1335 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1338 struct task_group *tg = cfs_rq->tg;
1341 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1342 tg_contrib -= cfs_rq->tg_load_contrib;
1344 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1345 atomic64_add(tg_contrib, &tg->load_avg);
1346 cfs_rq->tg_load_contrib += tg_contrib;
1351 * Aggregate cfs_rq runnable averages into an equivalent task_group
1352 * representation for computing load contributions.
1354 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1355 struct cfs_rq *cfs_rq)
1357 struct task_group *tg = cfs_rq->tg;
1358 long contrib, usage_contrib;
1360 /* The fraction of a cpu used by this cfs_rq */
1361 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1362 sa->runnable_avg_period + 1);
1363 contrib -= cfs_rq->tg_runnable_contrib;
1365 usage_contrib = div_u64(sa->usage_avg_sum << NICE_0_SHIFT,
1366 sa->runnable_avg_period + 1);
1367 usage_contrib -= cfs_rq->tg_usage_contrib;
1370 * contrib/usage at this point represent deltas, only update if they
1373 if ((abs(contrib) > cfs_rq->tg_runnable_contrib / 64) ||
1374 (abs(usage_contrib) > cfs_rq->tg_usage_contrib / 64)) {
1375 atomic_add(contrib, &tg->runnable_avg);
1376 cfs_rq->tg_runnable_contrib += contrib;
1378 atomic_add(usage_contrib, &tg->usage_avg);
1379 cfs_rq->tg_usage_contrib += usage_contrib;
1383 static inline void __update_group_entity_contrib(struct sched_entity *se)
1385 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1386 struct task_group *tg = cfs_rq->tg;
1391 contrib = cfs_rq->tg_load_contrib * tg->shares;
1392 se->avg.load_avg_contrib = div64_u64(contrib,
1393 atomic64_read(&tg->load_avg) + 1);
1396 * For group entities we need to compute a correction term in the case
1397 * that they are consuming <1 cpu so that we would contribute the same
1398 * load as a task of equal weight.
1400 * Explicitly co-ordinating this measurement would be expensive, but
1401 * fortunately the sum of each cpus contribution forms a usable
1402 * lower-bound on the true value.
1404 * Consider the aggregate of 2 contributions. Either they are disjoint
1405 * (and the sum represents true value) or they are disjoint and we are
1406 * understating by the aggregate of their overlap.
1408 * Extending this to N cpus, for a given overlap, the maximum amount we
1409 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1410 * cpus that overlap for this interval and w_i is the interval width.
1412 * On a small machine; the first term is well-bounded which bounds the
1413 * total error since w_i is a subset of the period. Whereas on a
1414 * larger machine, while this first term can be larger, if w_i is the
1415 * of consequential size guaranteed to see n_i*w_i quickly converge to
1416 * our upper bound of 1-cpu.
1418 runnable_avg = atomic_read(&tg->runnable_avg);
1419 if (runnable_avg < NICE_0_LOAD) {
1420 se->avg.load_avg_contrib *= runnable_avg;
1421 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1425 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1426 int force_update) {}
1427 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1428 struct cfs_rq *cfs_rq) {}
1429 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1432 static inline void __update_task_entity_contrib(struct sched_entity *se)
1436 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1437 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1438 contrib /= (se->avg.runnable_avg_period + 1);
1439 se->avg.load_avg_contrib = scale_load(contrib);
1440 trace_sched_task_load_contrib(task_of(se), se->avg.load_avg_contrib);
1441 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1442 contrib /= (se->avg.runnable_avg_period + 1);
1443 se->avg.load_avg_ratio = scale_load(contrib);
1444 trace_sched_task_runnable_ratio(task_of(se), se->avg.load_avg_ratio);
1447 /* Compute the current contribution to load_avg by se, return any delta */
1448 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1450 long old_contrib = se->avg.load_avg_contrib;
1452 if (entity_is_task(se)) {
1453 __update_task_entity_contrib(se);
1455 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1456 __update_group_entity_contrib(se);
1459 return se->avg.load_avg_contrib - old_contrib;
1462 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1465 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1466 cfs_rq->blocked_load_avg -= load_contrib;
1468 cfs_rq->blocked_load_avg = 0;
1471 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1473 /* Update a sched_entity's runnable average */
1474 static inline void update_entity_load_avg(struct sched_entity *se,
1477 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1482 * For a group entity we need to use their owned cfs_rq_clock_task() in
1483 * case they are the parent of a throttled hierarchy.
1485 if (entity_is_task(se))
1486 now = cfs_rq_clock_task(cfs_rq);
1488 now = cfs_rq_clock_task(group_cfs_rq(se));
1490 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1491 cfs_rq->curr == se))
1494 contrib_delta = __update_entity_load_avg_contrib(se);
1500 cfs_rq->runnable_load_avg += contrib_delta;
1502 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1506 * Decay the load contributed by all blocked children and account this so that
1507 * their contribution may appropriately discounted when they wake up.
1509 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1511 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1514 decays = now - cfs_rq->last_decay;
1515 if (!decays && !force_update)
1518 if (atomic64_read(&cfs_rq->removed_load)) {
1519 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1520 subtract_blocked_load_contrib(cfs_rq, removed_load);
1524 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1526 atomic64_add(decays, &cfs_rq->decay_counter);
1527 cfs_rq->last_decay = now;
1530 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1533 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1536 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1538 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1539 contrib = rq->avg.runnable_avg_sum * scale_load_down(1024);
1540 contrib /= (rq->avg.runnable_avg_period + 1);
1541 trace_sched_rq_runnable_ratio(cpu_of(rq), scale_load(contrib));
1542 trace_sched_rq_runnable_load(cpu_of(rq), rq->cfs.runnable_load_avg);
1545 /* Add the load generated by se into cfs_rq's child load-average */
1546 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1547 struct sched_entity *se,
1551 * We track migrations using entity decay_count <= 0, on a wake-up
1552 * migration we use a negative decay count to track the remote decays
1553 * accumulated while sleeping.
1555 if (unlikely(se->avg.decay_count <= 0)) {
1556 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1557 if (se->avg.decay_count) {
1559 * In a wake-up migration we have to approximate the
1560 * time sleeping. This is because we can't synchronize
1561 * clock_task between the two cpus, and it is not
1562 * guaranteed to be read-safe. Instead, we can
1563 * approximate this using our carried decays, which are
1564 * explicitly atomically readable.
1566 se->avg.last_runnable_update -= (-se->avg.decay_count)
1568 update_entity_load_avg(se, 0);
1569 /* Indicate that we're now synchronized and on-rq */
1570 se->avg.decay_count = 0;
1574 __synchronize_entity_decay(se);
1577 /* migrated tasks did not contribute to our blocked load */
1579 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1580 update_entity_load_avg(se, 0);
1583 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1584 /* we force update consideration on load-balancer moves */
1585 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1589 * Remove se's load from this cfs_rq child load-average, if the entity is
1590 * transitioning to a blocked state we track its projected decay using
1593 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1594 struct sched_entity *se,
1597 update_entity_load_avg(se, 1);
1598 /* we force update consideration on load-balancer moves */
1599 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1601 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1603 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1604 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1605 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1609 * Update the rq's load with the elapsed running time before entering
1610 * idle. if the last scheduled task is not a CFS task, idle_enter will
1611 * be the only way to update the runnable statistic.
1613 void idle_enter_fair(struct rq *this_rq)
1615 update_rq_runnable_avg(this_rq, 1);
1619 * Update the rq's load with the elapsed idle time before a task is
1620 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1621 * be the only way to update the runnable statistic.
1623 void idle_exit_fair(struct rq *this_rq)
1625 update_rq_runnable_avg(this_rq, 0);
1629 static inline void update_entity_load_avg(struct sched_entity *se,
1630 int update_cfs_rq) {}
1631 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1632 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1633 struct sched_entity *se,
1635 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1636 struct sched_entity *se,
1638 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1639 int force_update) {}
1642 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1644 #ifdef CONFIG_SCHEDSTATS
1645 struct task_struct *tsk = NULL;
1647 if (entity_is_task(se))
1650 if (se->statistics.sleep_start) {
1651 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1656 if (unlikely(delta > se->statistics.sleep_max))
1657 se->statistics.sleep_max = delta;
1659 se->statistics.sleep_start = 0;
1660 se->statistics.sum_sleep_runtime += delta;
1663 account_scheduler_latency(tsk, delta >> 10, 1);
1664 trace_sched_stat_sleep(tsk, delta);
1667 if (se->statistics.block_start) {
1668 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1673 if (unlikely(delta > se->statistics.block_max))
1674 se->statistics.block_max = delta;
1676 se->statistics.block_start = 0;
1677 se->statistics.sum_sleep_runtime += delta;
1680 if (tsk->in_iowait) {
1681 se->statistics.iowait_sum += delta;
1682 se->statistics.iowait_count++;
1683 trace_sched_stat_iowait(tsk, delta);
1686 trace_sched_stat_blocked(tsk, delta);
1689 * Blocking time is in units of nanosecs, so shift by
1690 * 20 to get a milliseconds-range estimation of the
1691 * amount of time that the task spent sleeping:
1693 if (unlikely(prof_on == SLEEP_PROFILING)) {
1694 profile_hits(SLEEP_PROFILING,
1695 (void *)get_wchan(tsk),
1698 account_scheduler_latency(tsk, delta >> 10, 0);
1704 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1706 #ifdef CONFIG_SCHED_DEBUG
1707 s64 d = se->vruntime - cfs_rq->min_vruntime;
1712 if (d > 3*sysctl_sched_latency)
1713 schedstat_inc(cfs_rq, nr_spread_over);
1718 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1720 u64 vruntime = cfs_rq->min_vruntime;
1723 * The 'current' period is already promised to the current tasks,
1724 * however the extra weight of the new task will slow them down a
1725 * little, place the new task so that it fits in the slot that
1726 * stays open at the end.
1728 if (initial && sched_feat(START_DEBIT))
1729 vruntime += sched_vslice(cfs_rq, se);
1731 /* sleeps up to a single latency don't count. */
1733 unsigned long thresh = sysctl_sched_latency;
1736 * Halve their sleep time's effect, to allow
1737 * for a gentler effect of sleepers:
1739 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1745 /* ensure we never gain time by being placed backwards. */
1746 se->vruntime = max_vruntime(se->vruntime, vruntime);
1749 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1752 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1755 * Update the normalized vruntime before updating min_vruntime
1756 * through callig update_curr().
1758 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1759 se->vruntime += cfs_rq->min_vruntime;
1762 * Update run-time statistics of the 'current'.
1764 update_curr(cfs_rq);
1765 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1766 account_entity_enqueue(cfs_rq, se);
1767 update_cfs_shares(cfs_rq);
1769 if (flags & ENQUEUE_WAKEUP) {
1770 place_entity(cfs_rq, se, 0);
1771 enqueue_sleeper(cfs_rq, se);
1774 update_stats_enqueue(cfs_rq, se);
1775 check_spread(cfs_rq, se);
1776 if (se != cfs_rq->curr)
1777 __enqueue_entity(cfs_rq, se);
1780 if (cfs_rq->nr_running == 1) {
1781 list_add_leaf_cfs_rq(cfs_rq);
1782 check_enqueue_throttle(cfs_rq);
1786 static void __clear_buddies_last(struct sched_entity *se)
1788 for_each_sched_entity(se) {
1789 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1790 if (cfs_rq->last == se)
1791 cfs_rq->last = NULL;
1797 static void __clear_buddies_next(struct sched_entity *se)
1799 for_each_sched_entity(se) {
1800 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1801 if (cfs_rq->next == se)
1802 cfs_rq->next = NULL;
1808 static void __clear_buddies_skip(struct sched_entity *se)
1810 for_each_sched_entity(se) {
1811 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1812 if (cfs_rq->skip == se)
1813 cfs_rq->skip = NULL;
1819 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1821 if (cfs_rq->last == se)
1822 __clear_buddies_last(se);
1824 if (cfs_rq->next == se)
1825 __clear_buddies_next(se);
1827 if (cfs_rq->skip == se)
1828 __clear_buddies_skip(se);
1831 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1834 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1837 * Update run-time statistics of the 'current'.
1839 update_curr(cfs_rq);
1840 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1842 update_stats_dequeue(cfs_rq, se);
1843 if (flags & DEQUEUE_SLEEP) {
1844 #ifdef CONFIG_SCHEDSTATS
1845 if (entity_is_task(se)) {
1846 struct task_struct *tsk = task_of(se);
1848 if (tsk->state & TASK_INTERRUPTIBLE)
1849 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1850 if (tsk->state & TASK_UNINTERRUPTIBLE)
1851 se->statistics.block_start = rq_of(cfs_rq)->clock;
1856 clear_buddies(cfs_rq, se);
1858 if (se != cfs_rq->curr)
1859 __dequeue_entity(cfs_rq, se);
1861 account_entity_dequeue(cfs_rq, se);
1864 * Normalize the entity after updating the min_vruntime because the
1865 * update can refer to the ->curr item and we need to reflect this
1866 * movement in our normalized position.
1868 if (!(flags & DEQUEUE_SLEEP))
1869 se->vruntime -= cfs_rq->min_vruntime;
1871 /* return excess runtime on last dequeue */
1872 return_cfs_rq_runtime(cfs_rq);
1874 update_min_vruntime(cfs_rq);
1875 update_cfs_shares(cfs_rq);
1879 * Preempt the current task with a newly woken task if needed:
1882 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1884 unsigned long ideal_runtime, delta_exec;
1885 struct sched_entity *se;
1888 ideal_runtime = sched_slice(cfs_rq, curr);
1889 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1890 if (delta_exec > ideal_runtime) {
1891 resched_task(rq_of(cfs_rq)->curr);
1893 * The current task ran long enough, ensure it doesn't get
1894 * re-elected due to buddy favours.
1896 clear_buddies(cfs_rq, curr);
1901 * Ensure that a task that missed wakeup preemption by a
1902 * narrow margin doesn't have to wait for a full slice.
1903 * This also mitigates buddy induced latencies under load.
1905 if (delta_exec < sysctl_sched_min_granularity)
1908 se = __pick_first_entity(cfs_rq);
1909 delta = curr->vruntime - se->vruntime;
1914 if (delta > ideal_runtime)
1915 resched_task(rq_of(cfs_rq)->curr);
1919 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1921 /* 'current' is not kept within the tree. */
1924 * Any task has to be enqueued before it get to execute on
1925 * a CPU. So account for the time it spent waiting on the
1928 update_stats_wait_end(cfs_rq, se);
1929 __dequeue_entity(cfs_rq, se);
1930 update_entity_load_avg(se, 1);
1933 update_stats_curr_start(cfs_rq, se);
1935 #ifdef CONFIG_SCHEDSTATS
1937 * Track our maximum slice length, if the CPU's load is at
1938 * least twice that of our own weight (i.e. dont track it
1939 * when there are only lesser-weight tasks around):
1941 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1942 se->statistics.slice_max = max(se->statistics.slice_max,
1943 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1946 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1950 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1953 * Pick the next process, keeping these things in mind, in this order:
1954 * 1) keep things fair between processes/task groups
1955 * 2) pick the "next" process, since someone really wants that to run
1956 * 3) pick the "last" process, for cache locality
1957 * 4) do not run the "skip" process, if something else is available
1959 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1961 struct sched_entity *se = __pick_first_entity(cfs_rq);
1962 struct sched_entity *left = se;
1965 * Avoid running the skip buddy, if running something else can
1966 * be done without getting too unfair.
1968 if (cfs_rq->skip == se) {
1969 struct sched_entity *second = __pick_next_entity(se);
1970 if (second && wakeup_preempt_entity(second, left) < 1)
1975 * Prefer last buddy, try to return the CPU to a preempted task.
1977 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1981 * Someone really wants this to run. If it's not unfair, run it.
1983 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1986 clear_buddies(cfs_rq, se);
1991 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1993 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1996 * If still on the runqueue then deactivate_task()
1997 * was not called and update_curr() has to be done:
2000 update_curr(cfs_rq);
2002 /* throttle cfs_rqs exceeding runtime */
2003 check_cfs_rq_runtime(cfs_rq);
2005 check_spread(cfs_rq, prev);
2007 update_stats_wait_start(cfs_rq, prev);
2008 /* Put 'current' back into the tree. */
2009 __enqueue_entity(cfs_rq, prev);
2010 /* in !on_rq case, update occurred at dequeue */
2011 update_entity_load_avg(prev, 1);
2013 cfs_rq->curr = NULL;
2017 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2020 * Update run-time statistics of the 'current'.
2022 update_curr(cfs_rq);
2025 * Ensure that runnable average is periodically updated.
2027 update_entity_load_avg(curr, 1);
2028 update_cfs_rq_blocked_load(cfs_rq, 1);
2030 #ifdef CONFIG_SCHED_HRTICK
2032 * queued ticks are scheduled to match the slice, so don't bother
2033 * validating it and just reschedule.
2036 resched_task(rq_of(cfs_rq)->curr);
2040 * don't let the period tick interfere with the hrtick preemption
2042 if (!sched_feat(DOUBLE_TICK) &&
2043 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2047 if (cfs_rq->nr_running > 1)
2048 check_preempt_tick(cfs_rq, curr);
2052 /**************************************************
2053 * CFS bandwidth control machinery
2056 #ifdef CONFIG_CFS_BANDWIDTH
2058 #ifdef HAVE_JUMP_LABEL
2059 static struct static_key __cfs_bandwidth_used;
2061 static inline bool cfs_bandwidth_used(void)
2063 return static_key_false(&__cfs_bandwidth_used);
2066 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2068 /* only need to count groups transitioning between enabled/!enabled */
2069 if (enabled && !was_enabled)
2070 static_key_slow_inc(&__cfs_bandwidth_used);
2071 else if (!enabled && was_enabled)
2072 static_key_slow_dec(&__cfs_bandwidth_used);
2074 #else /* HAVE_JUMP_LABEL */
2075 static bool cfs_bandwidth_used(void)
2080 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2081 #endif /* HAVE_JUMP_LABEL */
2084 * default period for cfs group bandwidth.
2085 * default: 0.1s, units: nanoseconds
2087 static inline u64 default_cfs_period(void)
2089 return 100000000ULL;
2092 static inline u64 sched_cfs_bandwidth_slice(void)
2094 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2098 * Replenish runtime according to assigned quota and update expiration time.
2099 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2100 * additional synchronization around rq->lock.
2102 * requires cfs_b->lock
2104 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2108 if (cfs_b->quota == RUNTIME_INF)
2111 now = sched_clock_cpu(smp_processor_id());
2112 cfs_b->runtime = cfs_b->quota;
2113 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2116 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2118 return &tg->cfs_bandwidth;
2121 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2122 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2124 if (unlikely(cfs_rq->throttle_count))
2125 return cfs_rq->throttled_clock_task;
2127 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2130 /* returns 0 on failure to allocate runtime */
2131 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2133 struct task_group *tg = cfs_rq->tg;
2134 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2135 u64 amount = 0, min_amount, expires;
2137 /* note: this is a positive sum as runtime_remaining <= 0 */
2138 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2140 raw_spin_lock(&cfs_b->lock);
2141 if (cfs_b->quota == RUNTIME_INF)
2142 amount = min_amount;
2145 * If the bandwidth pool has become inactive, then at least one
2146 * period must have elapsed since the last consumption.
2147 * Refresh the global state and ensure bandwidth timer becomes
2150 if (!cfs_b->timer_active) {
2151 __refill_cfs_bandwidth_runtime(cfs_b);
2152 __start_cfs_bandwidth(cfs_b);
2155 if (cfs_b->runtime > 0) {
2156 amount = min(cfs_b->runtime, min_amount);
2157 cfs_b->runtime -= amount;
2161 expires = cfs_b->runtime_expires;
2162 raw_spin_unlock(&cfs_b->lock);
2164 cfs_rq->runtime_remaining += amount;
2166 * we may have advanced our local expiration to account for allowed
2167 * spread between our sched_clock and the one on which runtime was
2170 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2171 cfs_rq->runtime_expires = expires;
2173 return cfs_rq->runtime_remaining > 0;
2177 * Note: This depends on the synchronization provided by sched_clock and the
2178 * fact that rq->clock snapshots this value.
2180 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2182 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2183 struct rq *rq = rq_of(cfs_rq);
2185 /* if the deadline is ahead of our clock, nothing to do */
2186 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2189 if (cfs_rq->runtime_remaining < 0)
2193 * If the local deadline has passed we have to consider the
2194 * possibility that our sched_clock is 'fast' and the global deadline
2195 * has not truly expired.
2197 * Fortunately we can check determine whether this the case by checking
2198 * whether the global deadline has advanced.
2201 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2202 /* extend local deadline, drift is bounded above by 2 ticks */
2203 cfs_rq->runtime_expires += TICK_NSEC;
2205 /* global deadline is ahead, expiration has passed */
2206 cfs_rq->runtime_remaining = 0;
2210 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2211 unsigned long delta_exec)
2213 /* dock delta_exec before expiring quota (as it could span periods) */
2214 cfs_rq->runtime_remaining -= delta_exec;
2215 expire_cfs_rq_runtime(cfs_rq);
2217 if (likely(cfs_rq->runtime_remaining > 0))
2221 * if we're unable to extend our runtime we resched so that the active
2222 * hierarchy can be throttled
2224 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2225 resched_task(rq_of(cfs_rq)->curr);
2228 static __always_inline
2229 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2231 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2234 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2237 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2239 return cfs_bandwidth_used() && cfs_rq->throttled;
2242 /* check whether cfs_rq, or any parent, is throttled */
2243 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2245 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2249 * Ensure that neither of the group entities corresponding to src_cpu or
2250 * dest_cpu are members of a throttled hierarchy when performing group
2251 * load-balance operations.
2253 static inline int throttled_lb_pair(struct task_group *tg,
2254 int src_cpu, int dest_cpu)
2256 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2258 src_cfs_rq = tg->cfs_rq[src_cpu];
2259 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2261 return throttled_hierarchy(src_cfs_rq) ||
2262 throttled_hierarchy(dest_cfs_rq);
2265 /* updated child weight may affect parent so we have to do this bottom up */
2266 static int tg_unthrottle_up(struct task_group *tg, void *data)
2268 struct rq *rq = data;
2269 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2271 cfs_rq->throttle_count--;
2273 if (!cfs_rq->throttle_count) {
2274 /* adjust cfs_rq_clock_task() */
2275 cfs_rq->throttled_clock_task_time += rq->clock_task -
2276 cfs_rq->throttled_clock_task;
2283 static int tg_throttle_down(struct task_group *tg, void *data)
2285 struct rq *rq = data;
2286 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2288 /* group is entering throttled state, stop time */
2289 if (!cfs_rq->throttle_count)
2290 cfs_rq->throttled_clock_task = rq->clock_task;
2291 cfs_rq->throttle_count++;
2296 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2298 struct rq *rq = rq_of(cfs_rq);
2299 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2300 struct sched_entity *se;
2301 long task_delta, dequeue = 1;
2303 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2305 /* freeze hierarchy runnable averages while throttled */
2307 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2310 task_delta = cfs_rq->h_nr_running;
2311 for_each_sched_entity(se) {
2312 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2313 /* throttled entity or throttle-on-deactivate */
2318 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2319 qcfs_rq->h_nr_running -= task_delta;
2321 if (qcfs_rq->load.weight)
2326 rq->nr_running -= task_delta;
2328 cfs_rq->throttled = 1;
2329 cfs_rq->throttled_clock = rq->clock;
2330 raw_spin_lock(&cfs_b->lock);
2331 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2332 raw_spin_unlock(&cfs_b->lock);
2335 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2337 struct rq *rq = rq_of(cfs_rq);
2338 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2339 struct sched_entity *se;
2343 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2345 cfs_rq->throttled = 0;
2346 raw_spin_lock(&cfs_b->lock);
2347 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2348 list_del_rcu(&cfs_rq->throttled_list);
2349 raw_spin_unlock(&cfs_b->lock);
2351 update_rq_clock(rq);
2352 /* update hierarchical throttle state */
2353 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2355 if (!cfs_rq->load.weight)
2358 task_delta = cfs_rq->h_nr_running;
2359 for_each_sched_entity(se) {
2363 cfs_rq = cfs_rq_of(se);
2365 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2366 cfs_rq->h_nr_running += task_delta;
2368 if (cfs_rq_throttled(cfs_rq))
2373 rq->nr_running += task_delta;
2375 /* determine whether we need to wake up potentially idle cpu */
2376 if (rq->curr == rq->idle && rq->cfs.nr_running)
2377 resched_task(rq->curr);
2380 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2381 u64 remaining, u64 expires)
2383 struct cfs_rq *cfs_rq;
2384 u64 runtime = remaining;
2387 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2389 struct rq *rq = rq_of(cfs_rq);
2391 raw_spin_lock(&rq->lock);
2392 if (!cfs_rq_throttled(cfs_rq))
2395 runtime = -cfs_rq->runtime_remaining + 1;
2396 if (runtime > remaining)
2397 runtime = remaining;
2398 remaining -= runtime;
2400 cfs_rq->runtime_remaining += runtime;
2401 cfs_rq->runtime_expires = expires;
2403 /* we check whether we're throttled above */
2404 if (cfs_rq->runtime_remaining > 0)
2405 unthrottle_cfs_rq(cfs_rq);
2408 raw_spin_unlock(&rq->lock);
2419 * Responsible for refilling a task_group's bandwidth and unthrottling its
2420 * cfs_rqs as appropriate. If there has been no activity within the last
2421 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2422 * used to track this state.
2424 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2426 u64 runtime, runtime_expires;
2427 int idle = 1, throttled;
2429 raw_spin_lock(&cfs_b->lock);
2430 /* no need to continue the timer with no bandwidth constraint */
2431 if (cfs_b->quota == RUNTIME_INF)
2434 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2435 /* idle depends on !throttled (for the case of a large deficit) */
2436 idle = cfs_b->idle && !throttled;
2437 cfs_b->nr_periods += overrun;
2439 /* if we're going inactive then everything else can be deferred */
2443 __refill_cfs_bandwidth_runtime(cfs_b);
2446 /* mark as potentially idle for the upcoming period */
2451 /* account preceding periods in which throttling occurred */
2452 cfs_b->nr_throttled += overrun;
2455 * There are throttled entities so we must first use the new bandwidth
2456 * to unthrottle them before making it generally available. This
2457 * ensures that all existing debts will be paid before a new cfs_rq is
2460 runtime = cfs_b->runtime;
2461 runtime_expires = cfs_b->runtime_expires;
2465 * This check is repeated as we are holding onto the new bandwidth
2466 * while we unthrottle. This can potentially race with an unthrottled
2467 * group trying to acquire new bandwidth from the global pool.
2469 while (throttled && runtime > 0) {
2470 raw_spin_unlock(&cfs_b->lock);
2471 /* we can't nest cfs_b->lock while distributing bandwidth */
2472 runtime = distribute_cfs_runtime(cfs_b, runtime,
2474 raw_spin_lock(&cfs_b->lock);
2476 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2479 /* return (any) remaining runtime */
2480 cfs_b->runtime = runtime;
2482 * While we are ensured activity in the period following an
2483 * unthrottle, this also covers the case in which the new bandwidth is
2484 * insufficient to cover the existing bandwidth deficit. (Forcing the
2485 * timer to remain active while there are any throttled entities.)
2490 cfs_b->timer_active = 0;
2491 raw_spin_unlock(&cfs_b->lock);
2496 /* a cfs_rq won't donate quota below this amount */
2497 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2498 /* minimum remaining period time to redistribute slack quota */
2499 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2500 /* how long we wait to gather additional slack before distributing */
2501 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2503 /* are we near the end of the current quota period? */
2504 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2506 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2509 /* if the call-back is running a quota refresh is already occurring */
2510 if (hrtimer_callback_running(refresh_timer))
2513 /* is a quota refresh about to occur? */
2514 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2515 if (remaining < min_expire)
2521 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2523 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2525 /* if there's a quota refresh soon don't bother with slack */
2526 if (runtime_refresh_within(cfs_b, min_left))
2529 start_bandwidth_timer(&cfs_b->slack_timer,
2530 ns_to_ktime(cfs_bandwidth_slack_period));
2533 /* we know any runtime found here is valid as update_curr() precedes return */
2534 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2536 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2537 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2539 if (slack_runtime <= 0)
2542 raw_spin_lock(&cfs_b->lock);
2543 if (cfs_b->quota != RUNTIME_INF &&
2544 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2545 cfs_b->runtime += slack_runtime;
2547 /* we are under rq->lock, defer unthrottling using a timer */
2548 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2549 !list_empty(&cfs_b->throttled_cfs_rq))
2550 start_cfs_slack_bandwidth(cfs_b);
2552 raw_spin_unlock(&cfs_b->lock);
2554 /* even if it's not valid for return we don't want to try again */
2555 cfs_rq->runtime_remaining -= slack_runtime;
2558 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2560 if (!cfs_bandwidth_used())
2563 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2566 __return_cfs_rq_runtime(cfs_rq);
2570 * This is done with a timer (instead of inline with bandwidth return) since
2571 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2573 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2575 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2578 /* confirm we're still not at a refresh boundary */
2579 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2582 raw_spin_lock(&cfs_b->lock);
2583 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2584 runtime = cfs_b->runtime;
2587 expires = cfs_b->runtime_expires;
2588 raw_spin_unlock(&cfs_b->lock);
2593 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2595 raw_spin_lock(&cfs_b->lock);
2596 if (expires == cfs_b->runtime_expires)
2597 cfs_b->runtime = runtime;
2598 raw_spin_unlock(&cfs_b->lock);
2602 * When a group wakes up we want to make sure that its quota is not already
2603 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2604 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2606 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2608 if (!cfs_bandwidth_used())
2611 /* an active group must be handled by the update_curr()->put() path */
2612 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2615 /* ensure the group is not already throttled */
2616 if (cfs_rq_throttled(cfs_rq))
2619 /* update runtime allocation */
2620 account_cfs_rq_runtime(cfs_rq, 0);
2621 if (cfs_rq->runtime_remaining <= 0)
2622 throttle_cfs_rq(cfs_rq);
2625 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2626 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2628 if (!cfs_bandwidth_used())
2631 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2635 * it's possible for a throttled entity to be forced into a running
2636 * state (e.g. set_curr_task), in this case we're finished.
2638 if (cfs_rq_throttled(cfs_rq))
2641 throttle_cfs_rq(cfs_rq);
2644 static inline u64 default_cfs_period(void);
2645 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2646 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2648 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2650 struct cfs_bandwidth *cfs_b =
2651 container_of(timer, struct cfs_bandwidth, slack_timer);
2652 do_sched_cfs_slack_timer(cfs_b);
2654 return HRTIMER_NORESTART;
2657 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2659 struct cfs_bandwidth *cfs_b =
2660 container_of(timer, struct cfs_bandwidth, period_timer);
2666 now = hrtimer_cb_get_time(timer);
2667 overrun = hrtimer_forward(timer, now, cfs_b->period);
2672 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2675 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2678 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2680 raw_spin_lock_init(&cfs_b->lock);
2682 cfs_b->quota = RUNTIME_INF;
2683 cfs_b->period = ns_to_ktime(default_cfs_period());
2685 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2686 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2687 cfs_b->period_timer.function = sched_cfs_period_timer;
2688 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2689 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2692 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2694 cfs_rq->runtime_enabled = 0;
2695 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2698 /* requires cfs_b->lock, may release to reprogram timer */
2699 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2702 * The timer may be active because we're trying to set a new bandwidth
2703 * period or because we're racing with the tear-down path
2704 * (timer_active==0 becomes visible before the hrtimer call-back
2705 * terminates). In either case we ensure that it's re-programmed
2707 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2708 raw_spin_unlock(&cfs_b->lock);
2709 /* ensure cfs_b->lock is available while we wait */
2710 hrtimer_cancel(&cfs_b->period_timer);
2712 raw_spin_lock(&cfs_b->lock);
2713 /* if someone else restarted the timer then we're done */
2714 if (cfs_b->timer_active)
2718 cfs_b->timer_active = 1;
2719 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2722 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2724 hrtimer_cancel(&cfs_b->period_timer);
2725 hrtimer_cancel(&cfs_b->slack_timer);
2728 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2730 struct cfs_rq *cfs_rq;
2732 for_each_leaf_cfs_rq(rq, cfs_rq) {
2733 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2735 if (!cfs_rq->runtime_enabled)
2739 * clock_task is not advancing so we just need to make sure
2740 * there's some valid quota amount
2742 cfs_rq->runtime_remaining = cfs_b->quota;
2743 if (cfs_rq_throttled(cfs_rq))
2744 unthrottle_cfs_rq(cfs_rq);
2748 #else /* CONFIG_CFS_BANDWIDTH */
2749 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2751 return rq_of(cfs_rq)->clock_task;
2754 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2755 unsigned long delta_exec) {}
2756 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2757 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2758 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2760 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2765 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2770 static inline int throttled_lb_pair(struct task_group *tg,
2771 int src_cpu, int dest_cpu)
2776 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2778 #ifdef CONFIG_FAIR_GROUP_SCHED
2779 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2782 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2786 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2787 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2789 #endif /* CONFIG_CFS_BANDWIDTH */
2791 /**************************************************
2792 * CFS operations on tasks:
2795 #ifdef CONFIG_SCHED_HRTICK
2796 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2798 struct sched_entity *se = &p->se;
2799 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2801 WARN_ON(task_rq(p) != rq);
2803 if (cfs_rq->nr_running > 1) {
2804 u64 slice = sched_slice(cfs_rq, se);
2805 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2806 s64 delta = slice - ran;
2815 * Don't schedule slices shorter than 10000ns, that just
2816 * doesn't make sense. Rely on vruntime for fairness.
2819 delta = max_t(s64, 10000LL, delta);
2821 hrtick_start(rq, delta);
2826 * called from enqueue/dequeue and updates the hrtick when the
2827 * current task is from our class and nr_running is low enough
2830 static void hrtick_update(struct rq *rq)
2832 struct task_struct *curr = rq->curr;
2834 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2837 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2838 hrtick_start_fair(rq, curr);
2840 #else /* !CONFIG_SCHED_HRTICK */
2842 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2846 static inline void hrtick_update(struct rq *rq)
2852 * The enqueue_task method is called before nr_running is
2853 * increased. Here we update the fair scheduling stats and
2854 * then put the task into the rbtree:
2857 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2859 struct cfs_rq *cfs_rq;
2860 struct sched_entity *se = &p->se;
2862 for_each_sched_entity(se) {
2865 cfs_rq = cfs_rq_of(se);
2866 enqueue_entity(cfs_rq, se, flags);
2869 * end evaluation on encountering a throttled cfs_rq
2871 * note: in the case of encountering a throttled cfs_rq we will
2872 * post the final h_nr_running increment below.
2874 if (cfs_rq_throttled(cfs_rq))
2876 cfs_rq->h_nr_running++;
2878 flags = ENQUEUE_WAKEUP;
2881 for_each_sched_entity(se) {
2882 cfs_rq = cfs_rq_of(se);
2883 cfs_rq->h_nr_running++;
2885 if (cfs_rq_throttled(cfs_rq))
2888 update_cfs_shares(cfs_rq);
2889 update_entity_load_avg(se, 1);
2893 update_rq_runnable_avg(rq, rq->nr_running);
2899 static void set_next_buddy(struct sched_entity *se);
2902 * The dequeue_task method is called before nr_running is
2903 * decreased. We remove the task from the rbtree and
2904 * update the fair scheduling stats:
2906 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2908 struct cfs_rq *cfs_rq;
2909 struct sched_entity *se = &p->se;
2910 int task_sleep = flags & DEQUEUE_SLEEP;
2912 for_each_sched_entity(se) {
2913 cfs_rq = cfs_rq_of(se);
2914 dequeue_entity(cfs_rq, se, flags);
2917 * end evaluation on encountering a throttled cfs_rq
2919 * note: in the case of encountering a throttled cfs_rq we will
2920 * post the final h_nr_running decrement below.
2922 if (cfs_rq_throttled(cfs_rq))
2924 cfs_rq->h_nr_running--;
2926 /* Don't dequeue parent if it has other entities besides us */
2927 if (cfs_rq->load.weight) {
2929 * Bias pick_next to pick a task from this cfs_rq, as
2930 * p is sleeping when it is within its sched_slice.
2932 if (task_sleep && parent_entity(se))
2933 set_next_buddy(parent_entity(se));
2935 /* avoid re-evaluating load for this entity */
2936 se = parent_entity(se);
2939 flags |= DEQUEUE_SLEEP;
2942 for_each_sched_entity(se) {
2943 cfs_rq = cfs_rq_of(se);
2944 cfs_rq->h_nr_running--;
2946 if (cfs_rq_throttled(cfs_rq))
2949 update_cfs_shares(cfs_rq);
2950 update_entity_load_avg(se, 1);
2955 update_rq_runnable_avg(rq, 1);
2961 /* Used instead of source_load when we know the type == 0 */
2962 static unsigned long weighted_cpuload(const int cpu)
2964 return cpu_rq(cpu)->load.weight;
2968 * Return a low guess at the load of a migration-source cpu weighted
2969 * according to the scheduling class and "nice" value.
2971 * We want to under-estimate the load of migration sources, to
2972 * balance conservatively.
2974 static unsigned long source_load(int cpu, int type)
2976 struct rq *rq = cpu_rq(cpu);
2977 unsigned long total = weighted_cpuload(cpu);
2979 if (type == 0 || !sched_feat(LB_BIAS))
2982 return min(rq->cpu_load[type-1], total);
2986 * Return a high guess at the load of a migration-target cpu weighted
2987 * according to the scheduling class and "nice" value.
2989 static unsigned long target_load(int cpu, int type)
2991 struct rq *rq = cpu_rq(cpu);
2992 unsigned long total = weighted_cpuload(cpu);
2994 if (type == 0 || !sched_feat(LB_BIAS))
2997 return max(rq->cpu_load[type-1], total);
3000 static unsigned long power_of(int cpu)
3002 return cpu_rq(cpu)->cpu_power;
3005 static unsigned long cpu_avg_load_per_task(int cpu)
3007 struct rq *rq = cpu_rq(cpu);
3008 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3011 return rq->load.weight / nr_running;
3017 static void task_waking_fair(struct task_struct *p)
3019 struct sched_entity *se = &p->se;
3020 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3023 #ifndef CONFIG_64BIT
3024 u64 min_vruntime_copy;
3027 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3029 min_vruntime = cfs_rq->min_vruntime;
3030 } while (min_vruntime != min_vruntime_copy);
3032 min_vruntime = cfs_rq->min_vruntime;
3035 se->vruntime -= min_vruntime;
3038 #ifdef CONFIG_FAIR_GROUP_SCHED
3040 * effective_load() calculates the load change as seen from the root_task_group
3042 * Adding load to a group doesn't make a group heavier, but can cause movement
3043 * of group shares between cpus. Assuming the shares were perfectly aligned one
3044 * can calculate the shift in shares.
3046 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3047 * on this @cpu and results in a total addition (subtraction) of @wg to the
3048 * total group weight.
3050 * Given a runqueue weight distribution (rw_i) we can compute a shares
3051 * distribution (s_i) using:
3053 * s_i = rw_i / \Sum rw_j (1)
3055 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3056 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3057 * shares distribution (s_i):
3059 * rw_i = { 2, 4, 1, 0 }
3060 * s_i = { 2/7, 4/7, 1/7, 0 }
3062 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3063 * task used to run on and the CPU the waker is running on), we need to
3064 * compute the effect of waking a task on either CPU and, in case of a sync
3065 * wakeup, compute the effect of the current task going to sleep.
3067 * So for a change of @wl to the local @cpu with an overall group weight change
3068 * of @wl we can compute the new shares distribution (s'_i) using:
3070 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3072 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3073 * differences in waking a task to CPU 0. The additional task changes the
3074 * weight and shares distributions like:
3076 * rw'_i = { 3, 4, 1, 0 }
3077 * s'_i = { 3/8, 4/8, 1/8, 0 }
3079 * We can then compute the difference in effective weight by using:
3081 * dw_i = S * (s'_i - s_i) (3)
3083 * Where 'S' is the group weight as seen by its parent.
3085 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3086 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3087 * 4/7) times the weight of the group.
3089 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3091 struct sched_entity *se = tg->se[cpu];
3093 if (!tg->parent) /* the trivial, non-cgroup case */
3096 for_each_sched_entity(se) {
3102 * W = @wg + \Sum rw_j
3104 W = wg + calc_tg_weight(tg, se->my_q);
3109 w = se->my_q->load.weight + wl;
3112 * wl = S * s'_i; see (2)
3115 wl = (w * tg->shares) / W;
3120 * Per the above, wl is the new se->load.weight value; since
3121 * those are clipped to [MIN_SHARES, ...) do so now. See
3122 * calc_cfs_shares().
3124 if (wl < MIN_SHARES)
3128 * wl = dw_i = S * (s'_i - s_i); see (3)
3130 wl -= se->load.weight;
3133 * Recursively apply this logic to all parent groups to compute
3134 * the final effective load change on the root group. Since
3135 * only the @tg group gets extra weight, all parent groups can
3136 * only redistribute existing shares. @wl is the shift in shares
3137 * resulting from this level per the above.
3146 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3147 unsigned long wl, unsigned long wg)
3154 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3156 s64 this_load, load;
3157 int idx, this_cpu, prev_cpu;
3158 unsigned long tl_per_task;
3159 struct task_group *tg;
3160 unsigned long weight;
3164 this_cpu = smp_processor_id();
3165 prev_cpu = task_cpu(p);
3166 load = source_load(prev_cpu, idx);
3167 this_load = target_load(this_cpu, idx);
3170 * If sync wakeup then subtract the (maximum possible)
3171 * effect of the currently running task from the load
3172 * of the current CPU:
3175 tg = task_group(current);
3176 weight = current->se.load.weight;
3178 this_load += effective_load(tg, this_cpu, -weight, -weight);
3179 load += effective_load(tg, prev_cpu, 0, -weight);
3183 weight = p->se.load.weight;
3186 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3187 * due to the sync cause above having dropped this_load to 0, we'll
3188 * always have an imbalance, but there's really nothing you can do
3189 * about that, so that's good too.
3191 * Otherwise check if either cpus are near enough in load to allow this
3192 * task to be woken on this_cpu.
3194 if (this_load > 0) {
3195 s64 this_eff_load, prev_eff_load;
3197 this_eff_load = 100;
3198 this_eff_load *= power_of(prev_cpu);
3199 this_eff_load *= this_load +
3200 effective_load(tg, this_cpu, weight, weight);
3202 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3203 prev_eff_load *= power_of(this_cpu);
3204 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3206 balanced = this_eff_load <= prev_eff_load;
3211 * If the currently running task will sleep within
3212 * a reasonable amount of time then attract this newly
3215 if (sync && balanced)
3218 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3219 tl_per_task = cpu_avg_load_per_task(this_cpu);
3222 (this_load <= load &&
3223 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3225 * This domain has SD_WAKE_AFFINE and
3226 * p is cache cold in this domain, and
3227 * there is no bad imbalance.
3229 schedstat_inc(sd, ttwu_move_affine);
3230 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3238 * find_idlest_group finds and returns the least busy CPU group within the
3241 static struct sched_group *
3242 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3243 int this_cpu, int load_idx)
3245 struct sched_group *idlest = NULL, *group = sd->groups;
3246 unsigned long min_load = ULONG_MAX, this_load = 0;
3247 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3250 unsigned long load, avg_load;
3254 /* Skip over this group if it has no CPUs allowed */
3255 if (!cpumask_intersects(sched_group_cpus(group),
3256 tsk_cpus_allowed(p)))
3259 local_group = cpumask_test_cpu(this_cpu,
3260 sched_group_cpus(group));
3262 /* Tally up the load of all CPUs in the group */
3265 for_each_cpu(i, sched_group_cpus(group)) {
3266 /* Bias balancing toward cpus of our domain */
3268 load = source_load(i, load_idx);
3270 load = target_load(i, load_idx);
3275 /* Adjust by relative CPU power of the group */
3276 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3279 this_load = avg_load;
3280 } else if (avg_load < min_load) {
3281 min_load = avg_load;
3284 } while (group = group->next, group != sd->groups);
3286 if (!idlest || 100*this_load < imbalance*min_load)
3292 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3295 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3297 unsigned long load, min_load = ULONG_MAX;
3301 /* Traverse only the allowed CPUs */
3302 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3303 load = weighted_cpuload(i);
3305 if (load < min_load || (load == min_load && i == this_cpu)) {
3315 * Try and locate an idle CPU in the sched_domain.
3317 static int select_idle_sibling(struct task_struct *p, int target)
3319 struct sched_domain *sd;
3320 struct sched_group *sg;
3321 int i = task_cpu(p);
3323 if (idle_cpu(target))
3327 * If the prevous cpu is cache affine and idle, don't be stupid.
3329 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3333 * Otherwise, iterate the domains and find an elegible idle cpu.
3335 sd = rcu_dereference(per_cpu(sd_llc, target));
3336 for_each_lower_domain(sd) {
3339 if (!cpumask_intersects(sched_group_cpus(sg),
3340 tsk_cpus_allowed(p)))
3343 for_each_cpu(i, sched_group_cpus(sg)) {
3344 if (i == target || !idle_cpu(i))
3348 target = cpumask_first_and(sched_group_cpus(sg),
3349 tsk_cpus_allowed(p));
3353 } while (sg != sd->groups);
3359 #ifdef CONFIG_SCHED_HMP
3361 * Heterogenous multiprocessor (HMP) optimizations
3363 * The cpu types are distinguished using a list of hmp_domains
3364 * which each represent one cpu type using a cpumask.
3365 * The list is assumed ordered by compute capacity with the
3366 * fastest domain first.
3368 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3370 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3372 /* Setup hmp_domains */
3373 static int __init hmp_cpu_mask_setup(void)
3376 struct hmp_domain *domain;
3377 struct list_head *pos;
3380 pr_debug("Initializing HMP scheduler:\n");
3382 /* Initialize hmp_domains using platform code */
3383 arch_get_hmp_domains(&hmp_domains);
3384 if (list_empty(&hmp_domains)) {
3385 pr_debug("HMP domain list is empty!\n");
3389 /* Print hmp_domains */
3391 list_for_each(pos, &hmp_domains) {
3392 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3393 cpulist_scnprintf(buf, 64, &domain->possible_cpus);
3394 pr_debug(" HMP domain %d: %s\n", dc, buf);
3396 for_each_cpu_mask(cpu, domain->possible_cpus) {
3397 per_cpu(hmp_cpu_domain, cpu) = domain;
3405 static struct hmp_domain *hmp_get_hmp_domain_for_cpu(int cpu)
3407 struct hmp_domain *domain;
3408 struct list_head *pos;
3410 list_for_each(pos, &hmp_domains) {
3411 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3412 if(cpumask_test_cpu(cpu, &domain->possible_cpus))
3418 static void hmp_online_cpu(int cpu)
3420 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3423 cpumask_set_cpu(cpu, &domain->cpus);
3426 static void hmp_offline_cpu(int cpu)
3428 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3431 cpumask_clear_cpu(cpu, &domain->cpus);
3435 * Migration thresholds should be in the range [0..1023]
3436 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3437 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3438 * The default values (512, 256) offer good responsiveness, but may need
3439 * tweaking suit particular needs.
3441 * hmp_up_prio: Only up migrate task with high priority (<hmp_up_prio)
3442 * hmp_next_up_threshold: Delay before next up migration (1024 ~= 1 ms)
3443 * hmp_next_down_threshold: Delay before next down migration (1024 ~= 1 ms)
3445 unsigned int hmp_up_threshold = 512;
3446 unsigned int hmp_down_threshold = 256;
3447 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
3448 unsigned int hmp_up_prio = NICE_TO_PRIO(CONFIG_SCHED_HMP_PRIO_FILTER_VAL);
3450 unsigned int hmp_next_up_threshold = 4096;
3451 unsigned int hmp_next_down_threshold = 4096;
3453 static unsigned int hmp_up_migration(int cpu, struct sched_entity *se);
3454 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3456 /* Check if cpu is in fastest hmp_domain */
3457 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3459 struct list_head *pos;
3461 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3462 return pos == hmp_domains.next;
3465 /* Check if cpu is in slowest hmp_domain */
3466 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3468 struct list_head *pos;
3470 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3471 return list_is_last(pos, &hmp_domains);
3474 /* Next (slower) hmp_domain relative to cpu */
3475 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3477 struct list_head *pos;
3479 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3480 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3483 /* Previous (faster) hmp_domain relative to cpu */
3484 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3486 struct list_head *pos;
3488 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3489 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3493 * Selects a cpu in previous (faster) hmp_domain
3494 * Note that cpumask_any_and() returns the first cpu in the cpumask
3496 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3499 return cpumask_any_and(&hmp_faster_domain(cpu)->cpus,
3500 tsk_cpus_allowed(tsk));
3504 * Selects a cpu in next (slower) hmp_domain
3505 * Note that cpumask_any_and() returns the first cpu in the cpumask
3507 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3510 return cpumask_any_and(&hmp_slower_domain(cpu)->cpus,
3511 tsk_cpus_allowed(tsk));
3514 static inline void hmp_next_up_delay(struct sched_entity *se, int cpu)
3516 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
3518 se->avg.hmp_last_up_migration = cfs_rq_clock_task(cfs_rq);
3519 se->avg.hmp_last_down_migration = 0;
3522 static inline void hmp_next_down_delay(struct sched_entity *se, int cpu)
3524 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
3526 se->avg.hmp_last_down_migration = cfs_rq_clock_task(cfs_rq);
3527 se->avg.hmp_last_up_migration = 0;
3530 #ifdef CONFIG_HMP_VARIABLE_SCALE
3532 * Heterogenous multiprocessor (HMP) optimizations
3534 * These functions allow to change the growing speed of the load_avg_ratio
3535 * by default it goes from 0 to 0.5 in LOAD_AVG_PERIOD = 32ms
3536 * This can now be changed with /sys/kernel/hmp/load_avg_period_ms.
3538 * These functions also allow to change the up and down threshold of HMP
3539 * using /sys/kernel/hmp/{up,down}_threshold.
3540 * Both must be between 0 and 1023. The threshold that is compared
3541 * to the load_avg_ratio is up_threshold/1024 and down_threshold/1024.
3543 * For instance, if load_avg_period = 64 and up_threshold = 512, an idle
3544 * task with a load of 0 will reach the threshold after 64ms of busy loop.
3546 * Changing load_avg_periods_ms has the same effect than changing the
3547 * default scaling factor Y=1002/1024 in the load_avg_ratio computation to
3548 * (1002/1024.0)^(LOAD_AVG_PERIOD/load_avg_period_ms), but the last one
3549 * could trigger overflows.
3550 * For instance, with Y = 1023/1024 in __update_task_entity_contrib()
3551 * "contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);"
3552 * could be overflowed for a weight > 2^12 even is the load_avg_contrib
3553 * should still be a 32bits result. This would not happen by multiplicating
3554 * delta time by 1/22 and setting load_avg_period_ms = 706.
3557 #define HMP_VARIABLE_SCALE_SHIFT 16ULL
3558 struct hmp_global_attr {
3559 struct attribute attr;
3560 ssize_t (*show)(struct kobject *kobj,
3561 struct attribute *attr, char *buf);
3562 ssize_t (*store)(struct kobject *a, struct attribute *b,
3563 const char *c, size_t count);
3565 int (*to_sysfs)(int);
3566 int (*from_sysfs)(int);
3569 #define HMP_DATA_SYSFS_MAX 3
3571 struct hmp_data_struct {
3572 int multiplier; /* used to scale the time delta */
3573 struct attribute_group attr_group;
3574 struct attribute *attributes[HMP_DATA_SYSFS_MAX + 1];
3575 struct hmp_global_attr attr[HMP_DATA_SYSFS_MAX];
3579 * By scaling the delta time it end-up increasing or decrease the
3580 * growing speed of the per entity load_avg_ratio
3581 * The scale factor hmp_data.multiplier is a fixed point
3582 * number: (32-HMP_VARIABLE_SCALE_SHIFT).HMP_VARIABLE_SCALE_SHIFT
3584 static u64 hmp_variable_scale_convert(u64 delta)
3586 u64 high = delta >> 32ULL;
3587 u64 low = delta & 0xffffffffULL;
3588 low *= hmp_data.multiplier;
3589 high *= hmp_data.multiplier;
3590 return (low >> HMP_VARIABLE_SCALE_SHIFT)
3591 + (high << (32ULL - HMP_VARIABLE_SCALE_SHIFT));
3594 static ssize_t hmp_show(struct kobject *kobj,
3595 struct attribute *attr, char *buf)
3598 struct hmp_global_attr *hmp_attr =
3599 container_of(attr, struct hmp_global_attr, attr);
3600 int temp = *(hmp_attr->value);
3601 if (hmp_attr->to_sysfs != NULL)
3602 temp = hmp_attr->to_sysfs(temp);
3603 ret = sprintf(buf, "%d\n", temp);
3607 static ssize_t hmp_store(struct kobject *a, struct attribute *attr,
3608 const char *buf, size_t count)
3611 ssize_t ret = count;
3612 struct hmp_global_attr *hmp_attr =
3613 container_of(attr, struct hmp_global_attr, attr);
3614 char *str = vmalloc(count + 1);
3617 memcpy(str, buf, count);
3619 if (sscanf(str, "%d", &temp) < 1)
3622 if (hmp_attr->from_sysfs != NULL)
3623 temp = hmp_attr->from_sysfs(temp);
3627 *(hmp_attr->value) = temp;
3633 static int hmp_period_tofrom_sysfs(int value)
3635 return (LOAD_AVG_PERIOD << HMP_VARIABLE_SCALE_SHIFT) / value;
3638 /* max value for threshold is 1024 */
3639 static int hmp_theshold_from_sysfs(int value)
3646 static void hmp_attr_add(
3649 int (*to_sysfs)(int),
3650 int (*from_sysfs)(int))
3653 while (hmp_data.attributes[i] != NULL) {
3655 if (i >= HMP_DATA_SYSFS_MAX)
3658 hmp_data.attr[i].attr.mode = 0644;
3659 hmp_data.attr[i].show = hmp_show;
3660 hmp_data.attr[i].store = hmp_store;
3661 hmp_data.attr[i].attr.name = name;
3662 hmp_data.attr[i].value = value;
3663 hmp_data.attr[i].to_sysfs = to_sysfs;
3664 hmp_data.attr[i].from_sysfs = from_sysfs;
3665 hmp_data.attributes[i] = &hmp_data.attr[i].attr;
3666 hmp_data.attributes[i + 1] = NULL;
3669 static int hmp_attr_init(void)
3672 memset(&hmp_data, sizeof(hmp_data), 0);
3673 /* by default load_avg_period_ms == LOAD_AVG_PERIOD
3676 hmp_data.multiplier = hmp_period_tofrom_sysfs(LOAD_AVG_PERIOD);
3678 hmp_attr_add("load_avg_period_ms",
3679 &hmp_data.multiplier,
3680 hmp_period_tofrom_sysfs,
3681 hmp_period_tofrom_sysfs);
3682 hmp_attr_add("up_threshold",
3685 hmp_theshold_from_sysfs);
3686 hmp_attr_add("down_threshold",
3687 &hmp_down_threshold,
3689 hmp_theshold_from_sysfs);
3691 hmp_data.attr_group.name = "hmp";
3692 hmp_data.attr_group.attrs = hmp_data.attributes;
3693 ret = sysfs_create_group(kernel_kobj,
3694 &hmp_data.attr_group);
3697 late_initcall(hmp_attr_init);
3698 #endif /* CONFIG_HMP_VARIABLE_SCALE */
3699 #endif /* CONFIG_SCHED_HMP */
3702 * sched_balance_self: balance the current task (running on cpu) in domains
3703 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3706 * Balance, ie. select the least loaded group.
3708 * Returns the target CPU number, or the same CPU if no balancing is needed.
3710 * preempt must be disabled.
3713 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3715 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3716 int cpu = smp_processor_id();
3717 int prev_cpu = task_cpu(p);
3719 int want_affine = 0;
3720 int sync = wake_flags & WF_SYNC;
3722 if (p->nr_cpus_allowed == 1)
3725 if (sd_flag & SD_BALANCE_WAKE) {
3726 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3732 for_each_domain(cpu, tmp) {
3733 if (!(tmp->flags & SD_LOAD_BALANCE))
3737 * If both cpu and prev_cpu are part of this domain,
3738 * cpu is a valid SD_WAKE_AFFINE target.
3740 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3741 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3746 if (tmp->flags & sd_flag)
3751 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3754 new_cpu = select_idle_sibling(p, prev_cpu);
3759 int load_idx = sd->forkexec_idx;
3760 struct sched_group *group;
3763 if (!(sd->flags & sd_flag)) {
3768 if (sd_flag & SD_BALANCE_WAKE)
3769 load_idx = sd->wake_idx;
3771 group = find_idlest_group(sd, p, cpu, load_idx);
3777 new_cpu = find_idlest_cpu(group, p, cpu);
3778 if (new_cpu == -1 || new_cpu == cpu) {
3779 /* Now try balancing at a lower domain level of cpu */
3784 /* Now try balancing at a lower domain level of new_cpu */
3786 weight = sd->span_weight;
3788 for_each_domain(cpu, tmp) {
3789 if (weight <= tmp->span_weight)
3791 if (tmp->flags & sd_flag)
3794 /* while loop will break here if sd == NULL */
3799 #ifdef CONFIG_SCHED_HMP
3800 if (hmp_up_migration(prev_cpu, &p->se)) {
3801 new_cpu = hmp_select_faster_cpu(p, prev_cpu);
3802 hmp_next_up_delay(&p->se, new_cpu);
3803 trace_sched_hmp_migrate(p, new_cpu, 0);
3806 if (hmp_down_migration(prev_cpu, &p->se)) {
3807 new_cpu = hmp_select_slower_cpu(p, prev_cpu);
3808 hmp_next_down_delay(&p->se, new_cpu);
3809 trace_sched_hmp_migrate(p, new_cpu, 0);
3812 /* Make sure that the task stays in its previous hmp domain */
3813 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
3821 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3822 * removed when useful for applications beyond shares distribution (e.g.
3825 #ifdef CONFIG_FAIR_GROUP_SCHED
3827 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3828 * cfs_rq_of(p) references at time of call are still valid and identify the
3829 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3830 * other assumptions, including the state of rq->lock, should be made.
3833 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3835 struct sched_entity *se = &p->se;
3836 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3839 * Load tracking: accumulate removed load so that it can be processed
3840 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3841 * to blocked load iff they have a positive decay-count. It can never
3842 * be negative here since on-rq tasks have decay-count == 0.
3844 if (se->avg.decay_count) {
3845 se->avg.decay_count = -__synchronize_entity_decay(se);
3846 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3850 #endif /* CONFIG_SMP */
3852 static unsigned long
3853 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3855 unsigned long gran = sysctl_sched_wakeup_granularity;
3858 * Since its curr running now, convert the gran from real-time
3859 * to virtual-time in his units.
3861 * By using 'se' instead of 'curr' we penalize light tasks, so
3862 * they get preempted easier. That is, if 'se' < 'curr' then
3863 * the resulting gran will be larger, therefore penalizing the
3864 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3865 * be smaller, again penalizing the lighter task.
3867 * This is especially important for buddies when the leftmost
3868 * task is higher priority than the buddy.
3870 return calc_delta_fair(gran, se);
3874 * Should 'se' preempt 'curr'.
3888 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3890 s64 gran, vdiff = curr->vruntime - se->vruntime;
3895 gran = wakeup_gran(curr, se);
3902 static void set_last_buddy(struct sched_entity *se)
3904 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3907 for_each_sched_entity(se)
3908 cfs_rq_of(se)->last = se;
3911 static void set_next_buddy(struct sched_entity *se)
3913 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3916 for_each_sched_entity(se)
3917 cfs_rq_of(se)->next = se;
3920 static void set_skip_buddy(struct sched_entity *se)
3922 for_each_sched_entity(se)
3923 cfs_rq_of(se)->skip = se;
3927 * Preempt the current task with a newly woken task if needed:
3929 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3931 struct task_struct *curr = rq->curr;
3932 struct sched_entity *se = &curr->se, *pse = &p->se;
3933 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3934 int scale = cfs_rq->nr_running >= sched_nr_latency;
3935 int next_buddy_marked = 0;
3937 if (unlikely(se == pse))
3941 * This is possible from callers such as move_task(), in which we
3942 * unconditionally check_prempt_curr() after an enqueue (which may have
3943 * lead to a throttle). This both saves work and prevents false
3944 * next-buddy nomination below.
3946 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3949 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3950 set_next_buddy(pse);
3951 next_buddy_marked = 1;
3955 * We can come here with TIF_NEED_RESCHED already set from new task
3958 * Note: this also catches the edge-case of curr being in a throttled
3959 * group (e.g. via set_curr_task), since update_curr() (in the
3960 * enqueue of curr) will have resulted in resched being set. This
3961 * prevents us from potentially nominating it as a false LAST_BUDDY
3964 if (test_tsk_need_resched(curr))
3967 /* Idle tasks are by definition preempted by non-idle tasks. */
3968 if (unlikely(curr->policy == SCHED_IDLE) &&
3969 likely(p->policy != SCHED_IDLE))
3973 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3974 * is driven by the tick):
3976 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3979 find_matching_se(&se, &pse);
3980 update_curr(cfs_rq_of(se));
3982 if (wakeup_preempt_entity(se, pse) == 1) {
3984 * Bias pick_next to pick the sched entity that is
3985 * triggering this preemption.
3987 if (!next_buddy_marked)
3988 set_next_buddy(pse);
3997 * Only set the backward buddy when the current task is still
3998 * on the rq. This can happen when a wakeup gets interleaved
3999 * with schedule on the ->pre_schedule() or idle_balance()
4000 * point, either of which can * drop the rq lock.
4002 * Also, during early boot the idle thread is in the fair class,
4003 * for obvious reasons its a bad idea to schedule back to it.
4005 if (unlikely(!se->on_rq || curr == rq->idle))
4008 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4012 static struct task_struct *pick_next_task_fair(struct rq *rq)
4014 struct task_struct *p;
4015 struct cfs_rq *cfs_rq = &rq->cfs;
4016 struct sched_entity *se;
4018 if (!cfs_rq->nr_running)
4022 se = pick_next_entity(cfs_rq);
4023 set_next_entity(cfs_rq, se);
4024 cfs_rq = group_cfs_rq(se);
4028 if (hrtick_enabled(rq))
4029 hrtick_start_fair(rq, p);
4035 * Account for a descheduled task:
4037 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4039 struct sched_entity *se = &prev->se;
4040 struct cfs_rq *cfs_rq;
4042 for_each_sched_entity(se) {
4043 cfs_rq = cfs_rq_of(se);
4044 put_prev_entity(cfs_rq, se);
4049 * sched_yield() is very simple
4051 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4053 static void yield_task_fair(struct rq *rq)
4055 struct task_struct *curr = rq->curr;
4056 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4057 struct sched_entity *se = &curr->se;
4060 * Are we the only task in the tree?
4062 if (unlikely(rq->nr_running == 1))
4065 clear_buddies(cfs_rq, se);
4067 if (curr->policy != SCHED_BATCH) {
4068 update_rq_clock(rq);
4070 * Update run-time statistics of the 'current'.
4072 update_curr(cfs_rq);
4074 * Tell update_rq_clock() that we've just updated,
4075 * so we don't do microscopic update in schedule()
4076 * and double the fastpath cost.
4078 rq->skip_clock_update = 1;
4084 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4086 struct sched_entity *se = &p->se;
4088 /* throttled hierarchies are not runnable */
4089 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4092 /* Tell the scheduler that we'd really like pse to run next. */
4095 yield_task_fair(rq);
4101 /**************************************************
4102 * Fair scheduling class load-balancing methods.
4106 * The purpose of load-balancing is to achieve the same basic fairness the
4107 * per-cpu scheduler provides, namely provide a proportional amount of compute
4108 * time to each task. This is expressed in the following equation:
4110 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4112 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4113 * W_i,0 is defined as:
4115 * W_i,0 = \Sum_j w_i,j (2)
4117 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4118 * is derived from the nice value as per prio_to_weight[].
4120 * The weight average is an exponential decay average of the instantaneous
4123 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4125 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4126 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4127 * can also include other factors [XXX].
4129 * To achieve this balance we define a measure of imbalance which follows
4130 * directly from (1):
4132 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4134 * We them move tasks around to minimize the imbalance. In the continuous
4135 * function space it is obvious this converges, in the discrete case we get
4136 * a few fun cases generally called infeasible weight scenarios.
4139 * - infeasible weights;
4140 * - local vs global optima in the discrete case. ]
4145 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4146 * for all i,j solution, we create a tree of cpus that follows the hardware
4147 * topology where each level pairs two lower groups (or better). This results
4148 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4149 * tree to only the first of the previous level and we decrease the frequency
4150 * of load-balance at each level inv. proportional to the number of cpus in
4156 * \Sum { --- * --- * 2^i } = O(n) (5)
4158 * `- size of each group
4159 * | | `- number of cpus doing load-balance
4161 * `- sum over all levels
4163 * Coupled with a limit on how many tasks we can migrate every balance pass,
4164 * this makes (5) the runtime complexity of the balancer.
4166 * An important property here is that each CPU is still (indirectly) connected
4167 * to every other cpu in at most O(log n) steps:
4169 * The adjacency matrix of the resulting graph is given by:
4172 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4175 * And you'll find that:
4177 * A^(log_2 n)_i,j != 0 for all i,j (7)
4179 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4180 * The task movement gives a factor of O(m), giving a convergence complexity
4183 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4188 * In order to avoid CPUs going idle while there's still work to do, new idle
4189 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4190 * tree itself instead of relying on other CPUs to bring it work.
4192 * This adds some complexity to both (5) and (8) but it reduces the total idle
4200 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4203 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4208 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4210 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4212 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4215 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4216 * rewrite all of this once again.]
4219 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4221 #define LBF_ALL_PINNED 0x01
4222 #define LBF_NEED_BREAK 0x02
4223 #define LBF_SOME_PINNED 0x04
4226 struct sched_domain *sd;
4234 struct cpumask *dst_grpmask;
4236 enum cpu_idle_type idle;
4238 /* The set of CPUs under consideration for load-balancing */
4239 struct cpumask *cpus;
4244 unsigned int loop_break;
4245 unsigned int loop_max;
4249 * move_task - move a task from one runqueue to another runqueue.
4250 * Both runqueues must be locked.
4252 static void move_task(struct task_struct *p, struct lb_env *env)
4254 deactivate_task(env->src_rq, p, 0);
4255 set_task_cpu(p, env->dst_cpu);
4256 activate_task(env->dst_rq, p, 0);
4257 check_preempt_curr(env->dst_rq, p, 0);
4261 * Is this task likely cache-hot:
4264 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4268 if (p->sched_class != &fair_sched_class)
4271 if (unlikely(p->policy == SCHED_IDLE))
4275 * Buddy candidates are cache hot:
4277 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4278 (&p->se == cfs_rq_of(&p->se)->next ||
4279 &p->se == cfs_rq_of(&p->se)->last))
4282 if (sysctl_sched_migration_cost == -1)
4284 if (sysctl_sched_migration_cost == 0)
4287 delta = now - p->se.exec_start;
4289 return delta < (s64)sysctl_sched_migration_cost;
4293 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4296 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4298 int tsk_cache_hot = 0;
4300 * We do not migrate tasks that are:
4301 * 1) throttled_lb_pair, or
4302 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4303 * 3) running (obviously), or
4304 * 4) are cache-hot on their current CPU.
4306 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4309 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4312 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4315 * Remember if this task can be migrated to any other cpu in
4316 * our sched_group. We may want to revisit it if we couldn't
4317 * meet load balance goals by pulling other tasks on src_cpu.
4319 * Also avoid computing new_dst_cpu if we have already computed
4320 * one in current iteration.
4322 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
4325 /* Prevent to re-select dst_cpu via env's cpus */
4326 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4327 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4328 env->flags |= LBF_SOME_PINNED;
4329 env->new_dst_cpu = cpu;
4337 /* Record that we found atleast one task that could run on dst_cpu */
4338 env->flags &= ~LBF_ALL_PINNED;
4340 if (task_running(env->src_rq, p)) {
4341 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4346 * Aggressive migration if:
4347 * 1) task is cache cold, or
4348 * 2) too many balance attempts have failed.
4350 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
4351 if (!tsk_cache_hot ||
4352 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4354 if (tsk_cache_hot) {
4355 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4356 schedstat_inc(p, se.statistics.nr_forced_migrations);
4362 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4367 * move_one_task tries to move exactly one task from busiest to this_rq, as
4368 * part of active balancing operations within "domain".
4369 * Returns 1 if successful and 0 otherwise.
4371 * Called with both runqueues locked.
4373 static int move_one_task(struct lb_env *env)
4375 struct task_struct *p, *n;
4377 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4378 if (!can_migrate_task(p, env))
4383 * Right now, this is only the second place move_task()
4384 * is called, so we can safely collect move_task()
4385 * stats here rather than inside move_task().
4387 schedstat_inc(env->sd, lb_gained[env->idle]);
4393 static unsigned long task_h_load(struct task_struct *p);
4395 static const unsigned int sched_nr_migrate_break = 32;
4398 * move_tasks tries to move up to imbalance weighted load from busiest to
4399 * this_rq, as part of a balancing operation within domain "sd".
4400 * Returns 1 if successful and 0 otherwise.
4402 * Called with both runqueues locked.
4404 static int move_tasks(struct lb_env *env)
4406 struct list_head *tasks = &env->src_rq->cfs_tasks;
4407 struct task_struct *p;
4411 if (env->imbalance <= 0)
4414 while (!list_empty(tasks)) {
4415 p = list_first_entry(tasks, struct task_struct, se.group_node);
4418 /* We've more or less seen every task there is, call it quits */
4419 if (env->loop > env->loop_max)
4422 /* take a breather every nr_migrate tasks */
4423 if (env->loop > env->loop_break) {
4424 env->loop_break += sched_nr_migrate_break;
4425 env->flags |= LBF_NEED_BREAK;
4429 if (!can_migrate_task(p, env))
4432 load = task_h_load(p);
4434 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4437 if ((load / 2) > env->imbalance)
4442 env->imbalance -= load;
4444 #ifdef CONFIG_PREEMPT
4446 * NEWIDLE balancing is a source of latency, so preemptible
4447 * kernels will stop after the first task is pulled to minimize
4448 * the critical section.
4450 if (env->idle == CPU_NEWLY_IDLE)
4455 * We only want to steal up to the prescribed amount of
4458 if (env->imbalance <= 0)
4463 list_move_tail(&p->se.group_node, tasks);
4467 * Right now, this is one of only two places move_task() is called,
4468 * so we can safely collect move_task() stats here rather than
4469 * inside move_task().
4471 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4476 #ifdef CONFIG_FAIR_GROUP_SCHED
4478 * update tg->load_weight by folding this cpu's load_avg
4480 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4482 struct sched_entity *se = tg->se[cpu];
4483 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4485 /* throttled entities do not contribute to load */
4486 if (throttled_hierarchy(cfs_rq))
4489 update_cfs_rq_blocked_load(cfs_rq, 1);
4492 update_entity_load_avg(se, 1);
4494 * We pivot on our runnable average having decayed to zero for
4495 * list removal. This generally implies that all our children
4496 * have also been removed (modulo rounding error or bandwidth
4497 * control); however, such cases are rare and we can fix these
4500 * TODO: fix up out-of-order children on enqueue.
4502 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4503 list_del_leaf_cfs_rq(cfs_rq);
4505 struct rq *rq = rq_of(cfs_rq);
4506 update_rq_runnable_avg(rq, rq->nr_running);
4510 static void update_blocked_averages(int cpu)
4512 struct rq *rq = cpu_rq(cpu);
4513 struct cfs_rq *cfs_rq;
4514 unsigned long flags;
4516 raw_spin_lock_irqsave(&rq->lock, flags);
4517 update_rq_clock(rq);
4519 * Iterates the task_group tree in a bottom up fashion, see
4520 * list_add_leaf_cfs_rq() for details.
4522 for_each_leaf_cfs_rq(rq, cfs_rq) {
4524 * Note: We may want to consider periodically releasing
4525 * rq->lock about these updates so that creating many task
4526 * groups does not result in continually extending hold time.
4528 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4531 raw_spin_unlock_irqrestore(&rq->lock, flags);
4535 * Compute the cpu's hierarchical load factor for each task group.
4536 * This needs to be done in a top-down fashion because the load of a child
4537 * group is a fraction of its parents load.
4539 static int tg_load_down(struct task_group *tg, void *data)
4542 long cpu = (long)data;
4545 load = cpu_rq(cpu)->load.weight;
4547 load = tg->parent->cfs_rq[cpu]->h_load;
4548 load *= tg->se[cpu]->load.weight;
4549 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
4552 tg->cfs_rq[cpu]->h_load = load;
4557 static void update_h_load(long cpu)
4559 struct rq *rq = cpu_rq(cpu);
4560 unsigned long now = jiffies;
4562 if (rq->h_load_throttle == now)
4565 rq->h_load_throttle = now;
4568 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4572 static unsigned long task_h_load(struct task_struct *p)
4574 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4577 load = p->se.load.weight;
4578 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
4583 static inline void update_blocked_averages(int cpu)
4587 static inline void update_h_load(long cpu)
4591 static unsigned long task_h_load(struct task_struct *p)
4593 return p->se.load.weight;
4597 /********** Helpers for find_busiest_group ************************/
4599 * sd_lb_stats - Structure to store the statistics of a sched_domain
4600 * during load balancing.
4602 struct sd_lb_stats {
4603 struct sched_group *busiest; /* Busiest group in this sd */
4604 struct sched_group *this; /* Local group in this sd */
4605 unsigned long total_load; /* Total load of all groups in sd */
4606 unsigned long total_pwr; /* Total power of all groups in sd */
4607 unsigned long avg_load; /* Average load across all groups in sd */
4609 /** Statistics of this group */
4610 unsigned long this_load;
4611 unsigned long this_load_per_task;
4612 unsigned long this_nr_running;
4613 unsigned long this_has_capacity;
4614 unsigned int this_idle_cpus;
4616 /* Statistics of the busiest group */
4617 unsigned int busiest_idle_cpus;
4618 unsigned long max_load;
4619 unsigned long busiest_load_per_task;
4620 unsigned long busiest_nr_running;
4621 unsigned long busiest_group_capacity;
4622 unsigned long busiest_has_capacity;
4623 unsigned int busiest_group_weight;
4625 int group_imb; /* Is there imbalance in this sd */
4629 * sg_lb_stats - stats of a sched_group required for load_balancing
4631 struct sg_lb_stats {
4632 unsigned long avg_load; /*Avg load across the CPUs of the group */
4633 unsigned long group_load; /* Total load over the CPUs of the group */
4634 unsigned long sum_nr_running; /* Nr tasks running in the group */
4635 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4636 unsigned long group_capacity;
4637 unsigned long idle_cpus;
4638 unsigned long group_weight;
4639 int group_imb; /* Is there an imbalance in the group ? */
4640 int group_has_capacity; /* Is there extra capacity in the group? */
4644 * get_sd_load_idx - Obtain the load index for a given sched domain.
4645 * @sd: The sched_domain whose load_idx is to be obtained.
4646 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4648 static inline int get_sd_load_idx(struct sched_domain *sd,
4649 enum cpu_idle_type idle)
4655 load_idx = sd->busy_idx;
4658 case CPU_NEWLY_IDLE:
4659 load_idx = sd->newidle_idx;
4662 load_idx = sd->idle_idx;
4669 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4671 return SCHED_POWER_SCALE;
4674 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4676 return default_scale_freq_power(sd, cpu);
4679 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4681 unsigned long weight = sd->span_weight;
4682 unsigned long smt_gain = sd->smt_gain;
4689 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4691 return default_scale_smt_power(sd, cpu);
4694 static unsigned long scale_rt_power(int cpu)
4696 struct rq *rq = cpu_rq(cpu);
4697 u64 total, available, age_stamp, avg;
4700 * Since we're reading these variables without serialization make sure
4701 * we read them once before doing sanity checks on them.
4703 age_stamp = ACCESS_ONCE(rq->age_stamp);
4704 avg = ACCESS_ONCE(rq->rt_avg);
4706 total = sched_avg_period() + (rq->clock - age_stamp);
4708 if (unlikely(total < avg)) {
4709 /* Ensures that power won't end up being negative */
4712 available = total - avg;
4715 if (unlikely((s64)total < SCHED_POWER_SCALE))
4716 total = SCHED_POWER_SCALE;
4718 total >>= SCHED_POWER_SHIFT;
4720 return div_u64(available, total);
4723 static void update_cpu_power(struct sched_domain *sd, int cpu)
4725 unsigned long weight = sd->span_weight;
4726 unsigned long power = SCHED_POWER_SCALE;
4727 struct sched_group *sdg = sd->groups;
4729 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4730 if (sched_feat(ARCH_POWER))
4731 power *= arch_scale_smt_power(sd, cpu);
4733 power *= default_scale_smt_power(sd, cpu);
4735 power >>= SCHED_POWER_SHIFT;
4738 sdg->sgp->power_orig = power;
4740 if (sched_feat(ARCH_POWER))
4741 power *= arch_scale_freq_power(sd, cpu);
4743 power *= default_scale_freq_power(sd, cpu);
4745 power >>= SCHED_POWER_SHIFT;
4747 power *= scale_rt_power(cpu);
4748 power >>= SCHED_POWER_SHIFT;
4753 cpu_rq(cpu)->cpu_power = power;
4754 sdg->sgp->power = power;
4757 void update_group_power(struct sched_domain *sd, int cpu)
4759 struct sched_domain *child = sd->child;
4760 struct sched_group *group, *sdg = sd->groups;
4761 unsigned long power;
4762 unsigned long interval;
4764 interval = msecs_to_jiffies(sd->balance_interval);
4765 interval = clamp(interval, 1UL, max_load_balance_interval);
4766 sdg->sgp->next_update = jiffies + interval;
4769 update_cpu_power(sd, cpu);
4775 if (child->flags & SD_OVERLAP) {
4777 * SD_OVERLAP domains cannot assume that child groups
4778 * span the current group.
4781 for_each_cpu(cpu, sched_group_cpus(sdg))
4782 power += power_of(cpu);
4785 * !SD_OVERLAP domains can assume that child groups
4786 * span the current group.
4789 group = child->groups;
4791 power += group->sgp->power;
4792 group = group->next;
4793 } while (group != child->groups);
4796 sdg->sgp->power_orig = sdg->sgp->power = power;
4800 * Try and fix up capacity for tiny siblings, this is needed when
4801 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4802 * which on its own isn't powerful enough.
4804 * See update_sd_pick_busiest() and check_asym_packing().
4807 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4810 * Only siblings can have significantly less than SCHED_POWER_SCALE
4812 if (!(sd->flags & SD_SHARE_CPUPOWER))
4816 * If ~90% of the cpu_power is still there, we're good.
4818 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4825 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4826 * @env: The load balancing environment.
4827 * @group: sched_group whose statistics are to be updated.
4828 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4829 * @local_group: Does group contain this_cpu.
4830 * @balance: Should we balance.
4831 * @sgs: variable to hold the statistics for this group.
4833 static inline void update_sg_lb_stats(struct lb_env *env,
4834 struct sched_group *group, int load_idx,
4835 int local_group, int *balance, struct sg_lb_stats *sgs)
4837 unsigned long nr_running, max_nr_running, min_nr_running;
4838 unsigned long load, max_cpu_load, min_cpu_load;
4839 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4840 unsigned long avg_load_per_task = 0;
4844 balance_cpu = group_balance_cpu(group);
4846 /* Tally up the load of all CPUs in the group */
4848 min_cpu_load = ~0UL;
4850 min_nr_running = ~0UL;
4852 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4853 struct rq *rq = cpu_rq(i);
4855 nr_running = rq->nr_running;
4857 /* Bias balancing toward cpus of our domain */
4859 if (idle_cpu(i) && !first_idle_cpu &&
4860 cpumask_test_cpu(i, sched_group_mask(group))) {
4865 load = target_load(i, load_idx);
4867 load = source_load(i, load_idx);
4868 if (load > max_cpu_load)
4869 max_cpu_load = load;
4870 if (min_cpu_load > load)
4871 min_cpu_load = load;
4873 if (nr_running > max_nr_running)
4874 max_nr_running = nr_running;
4875 if (min_nr_running > nr_running)
4876 min_nr_running = nr_running;
4879 sgs->group_load += load;
4880 sgs->sum_nr_running += nr_running;
4881 sgs->sum_weighted_load += weighted_cpuload(i);
4887 * First idle cpu or the first cpu(busiest) in this sched group
4888 * is eligible for doing load balancing at this and above
4889 * domains. In the newly idle case, we will allow all the cpu's
4890 * to do the newly idle load balance.
4893 if (env->idle != CPU_NEWLY_IDLE) {
4894 if (balance_cpu != env->dst_cpu) {
4898 update_group_power(env->sd, env->dst_cpu);
4899 } else if (time_after_eq(jiffies, group->sgp->next_update))
4900 update_group_power(env->sd, env->dst_cpu);
4903 /* Adjust by relative CPU power of the group */
4904 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4907 * Consider the group unbalanced when the imbalance is larger
4908 * than the average weight of a task.
4910 * APZ: with cgroup the avg task weight can vary wildly and
4911 * might not be a suitable number - should we keep a
4912 * normalized nr_running number somewhere that negates
4915 if (sgs->sum_nr_running)
4916 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4918 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4919 (max_nr_running - min_nr_running) > 1)
4922 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4924 if (!sgs->group_capacity)
4925 sgs->group_capacity = fix_small_capacity(env->sd, group);
4926 sgs->group_weight = group->group_weight;
4928 if (sgs->group_capacity > sgs->sum_nr_running)
4929 sgs->group_has_capacity = 1;
4933 * update_sd_pick_busiest - return 1 on busiest group
4934 * @env: The load balancing environment.
4935 * @sds: sched_domain statistics
4936 * @sg: sched_group candidate to be checked for being the busiest
4937 * @sgs: sched_group statistics
4939 * Determine if @sg is a busier group than the previously selected
4942 static bool update_sd_pick_busiest(struct lb_env *env,
4943 struct sd_lb_stats *sds,
4944 struct sched_group *sg,
4945 struct sg_lb_stats *sgs)
4947 if (sgs->avg_load <= sds->max_load)
4950 if (sgs->sum_nr_running > sgs->group_capacity)
4957 * ASYM_PACKING needs to move all the work to the lowest
4958 * numbered CPUs in the group, therefore mark all groups
4959 * higher than ourself as busy.
4961 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4962 env->dst_cpu < group_first_cpu(sg)) {
4966 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4974 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4975 * @env: The load balancing environment.
4976 * @balance: Should we balance.
4977 * @sds: variable to hold the statistics for this sched_domain.
4979 static inline void update_sd_lb_stats(struct lb_env *env,
4980 int *balance, struct sd_lb_stats *sds)
4982 struct sched_domain *child = env->sd->child;
4983 struct sched_group *sg = env->sd->groups;
4984 struct sg_lb_stats sgs;
4985 int load_idx, prefer_sibling = 0;
4987 if (child && child->flags & SD_PREFER_SIBLING)
4990 load_idx = get_sd_load_idx(env->sd, env->idle);
4995 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4996 memset(&sgs, 0, sizeof(sgs));
4997 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4999 if (local_group && !(*balance))
5002 sds->total_load += sgs.group_load;
5003 sds->total_pwr += sg->sgp->power;
5006 * In case the child domain prefers tasks go to siblings
5007 * first, lower the sg capacity to one so that we'll try
5008 * and move all the excess tasks away. We lower the capacity
5009 * of a group only if the local group has the capacity to fit
5010 * these excess tasks, i.e. nr_running < group_capacity. The
5011 * extra check prevents the case where you always pull from the
5012 * heaviest group when it is already under-utilized (possible
5013 * with a large weight task outweighs the tasks on the system).
5015 if (prefer_sibling && !local_group && sds->this_has_capacity)
5016 sgs.group_capacity = min(sgs.group_capacity, 1UL);
5019 sds->this_load = sgs.avg_load;
5021 sds->this_nr_running = sgs.sum_nr_running;
5022 sds->this_load_per_task = sgs.sum_weighted_load;
5023 sds->this_has_capacity = sgs.group_has_capacity;
5024 sds->this_idle_cpus = sgs.idle_cpus;
5025 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
5026 sds->max_load = sgs.avg_load;
5028 sds->busiest_nr_running = sgs.sum_nr_running;
5029 sds->busiest_idle_cpus = sgs.idle_cpus;
5030 sds->busiest_group_capacity = sgs.group_capacity;
5031 sds->busiest_load_per_task = sgs.sum_weighted_load;
5032 sds->busiest_has_capacity = sgs.group_has_capacity;
5033 sds->busiest_group_weight = sgs.group_weight;
5034 sds->group_imb = sgs.group_imb;
5038 } while (sg != env->sd->groups);
5042 * check_asym_packing - Check to see if the group is packed into the
5045 * This is primarily intended to used at the sibling level. Some
5046 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5047 * case of POWER7, it can move to lower SMT modes only when higher
5048 * threads are idle. When in lower SMT modes, the threads will
5049 * perform better since they share less core resources. Hence when we
5050 * have idle threads, we want them to be the higher ones.
5052 * This packing function is run on idle threads. It checks to see if
5053 * the busiest CPU in this domain (core in the P7 case) has a higher
5054 * CPU number than the packing function is being run on. Here we are
5055 * assuming lower CPU number will be equivalent to lower a SMT thread
5058 * Returns 1 when packing is required and a task should be moved to
5059 * this CPU. The amount of the imbalance is returned in *imbalance.
5061 * @env: The load balancing environment.
5062 * @sds: Statistics of the sched_domain which is to be packed
5064 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5068 if (!(env->sd->flags & SD_ASYM_PACKING))
5074 busiest_cpu = group_first_cpu(sds->busiest);
5075 if (env->dst_cpu > busiest_cpu)
5078 env->imbalance = DIV_ROUND_CLOSEST(
5079 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
5085 * fix_small_imbalance - Calculate the minor imbalance that exists
5086 * amongst the groups of a sched_domain, during
5088 * @env: The load balancing environment.
5089 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5092 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5094 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5095 unsigned int imbn = 2;
5096 unsigned long scaled_busy_load_per_task;
5098 if (sds->this_nr_running) {
5099 sds->this_load_per_task /= sds->this_nr_running;
5100 if (sds->busiest_load_per_task >
5101 sds->this_load_per_task)
5104 sds->this_load_per_task =
5105 cpu_avg_load_per_task(env->dst_cpu);
5108 scaled_busy_load_per_task = sds->busiest_load_per_task
5109 * SCHED_POWER_SCALE;
5110 scaled_busy_load_per_task /= sds->busiest->sgp->power;
5112 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
5113 (scaled_busy_load_per_task * imbn)) {
5114 env->imbalance = sds->busiest_load_per_task;
5119 * OK, we don't have enough imbalance to justify moving tasks,
5120 * however we may be able to increase total CPU power used by
5124 pwr_now += sds->busiest->sgp->power *
5125 min(sds->busiest_load_per_task, sds->max_load);
5126 pwr_now += sds->this->sgp->power *
5127 min(sds->this_load_per_task, sds->this_load);
5128 pwr_now /= SCHED_POWER_SCALE;
5130 /* Amount of load we'd subtract */
5131 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5132 sds->busiest->sgp->power;
5133 if (sds->max_load > tmp)
5134 pwr_move += sds->busiest->sgp->power *
5135 min(sds->busiest_load_per_task, sds->max_load - tmp);
5137 /* Amount of load we'd add */
5138 if (sds->max_load * sds->busiest->sgp->power <
5139 sds->busiest_load_per_task * SCHED_POWER_SCALE)
5140 tmp = (sds->max_load * sds->busiest->sgp->power) /
5141 sds->this->sgp->power;
5143 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5144 sds->this->sgp->power;
5145 pwr_move += sds->this->sgp->power *
5146 min(sds->this_load_per_task, sds->this_load + tmp);
5147 pwr_move /= SCHED_POWER_SCALE;
5149 /* Move if we gain throughput */
5150 if (pwr_move > pwr_now)
5151 env->imbalance = sds->busiest_load_per_task;
5155 * calculate_imbalance - Calculate the amount of imbalance present within the
5156 * groups of a given sched_domain during load balance.
5157 * @env: load balance environment
5158 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5160 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5162 unsigned long max_pull, load_above_capacity = ~0UL;
5164 sds->busiest_load_per_task /= sds->busiest_nr_running;
5165 if (sds->group_imb) {
5166 sds->busiest_load_per_task =
5167 min(sds->busiest_load_per_task, sds->avg_load);
5171 * In the presence of smp nice balancing, certain scenarios can have
5172 * max load less than avg load(as we skip the groups at or below
5173 * its cpu_power, while calculating max_load..)
5175 if (sds->max_load < sds->avg_load) {
5177 return fix_small_imbalance(env, sds);
5180 if (!sds->group_imb) {
5182 * Don't want to pull so many tasks that a group would go idle.
5184 load_above_capacity = (sds->busiest_nr_running -
5185 sds->busiest_group_capacity);
5187 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5189 load_above_capacity /= sds->busiest->sgp->power;
5193 * We're trying to get all the cpus to the average_load, so we don't
5194 * want to push ourselves above the average load, nor do we wish to
5195 * reduce the max loaded cpu below the average load. At the same time,
5196 * we also don't want to reduce the group load below the group capacity
5197 * (so that we can implement power-savings policies etc). Thus we look
5198 * for the minimum possible imbalance.
5199 * Be careful of negative numbers as they'll appear as very large values
5200 * with unsigned longs.
5202 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
5204 /* How much load to actually move to equalise the imbalance */
5205 env->imbalance = min(max_pull * sds->busiest->sgp->power,
5206 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
5207 / SCHED_POWER_SCALE;
5210 * if *imbalance is less than the average load per runnable task
5211 * there is no guarantee that any tasks will be moved so we'll have
5212 * a think about bumping its value to force at least one task to be
5215 if (env->imbalance < sds->busiest_load_per_task)
5216 return fix_small_imbalance(env, sds);
5220 /******* find_busiest_group() helpers end here *********************/
5223 * find_busiest_group - Returns the busiest group within the sched_domain
5224 * if there is an imbalance. If there isn't an imbalance, and
5225 * the user has opted for power-savings, it returns a group whose
5226 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5227 * such a group exists.
5229 * Also calculates the amount of weighted load which should be moved
5230 * to restore balance.
5232 * @env: The load balancing environment.
5233 * @balance: Pointer to a variable indicating if this_cpu
5234 * is the appropriate cpu to perform load balancing at this_level.
5236 * Returns: - the busiest group if imbalance exists.
5237 * - If no imbalance and user has opted for power-savings balance,
5238 * return the least loaded group whose CPUs can be
5239 * put to idle by rebalancing its tasks onto our group.
5241 static struct sched_group *
5242 find_busiest_group(struct lb_env *env, int *balance)
5244 struct sd_lb_stats sds;
5246 memset(&sds, 0, sizeof(sds));
5249 * Compute the various statistics relavent for load balancing at
5252 update_sd_lb_stats(env, balance, &sds);
5255 * this_cpu is not the appropriate cpu to perform load balancing at
5261 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5262 check_asym_packing(env, &sds))
5265 /* There is no busy sibling group to pull tasks from */
5266 if (!sds.busiest || sds.busiest_nr_running == 0)
5269 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5272 * If the busiest group is imbalanced the below checks don't
5273 * work because they assumes all things are equal, which typically
5274 * isn't true due to cpus_allowed constraints and the like.
5279 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5280 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5281 !sds.busiest_has_capacity)
5285 * If the local group is more busy than the selected busiest group
5286 * don't try and pull any tasks.
5288 if (sds.this_load >= sds.max_load)
5292 * Don't pull any tasks if this group is already above the domain
5295 if (sds.this_load >= sds.avg_load)
5298 if (env->idle == CPU_IDLE) {
5300 * This cpu is idle. If the busiest group load doesn't
5301 * have more tasks than the number of available cpu's and
5302 * there is no imbalance between this and busiest group
5303 * wrt to idle cpu's, it is balanced.
5305 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
5306 sds.busiest_nr_running <= sds.busiest_group_weight)
5310 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5311 * imbalance_pct to be conservative.
5313 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
5318 /* Looks like there is an imbalance. Compute it */
5319 calculate_imbalance(env, &sds);
5329 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5331 static struct rq *find_busiest_queue(struct lb_env *env,
5332 struct sched_group *group)
5334 struct rq *busiest = NULL, *rq;
5335 unsigned long max_load = 0;
5338 for_each_cpu(i, sched_group_cpus(group)) {
5339 unsigned long power = power_of(i);
5340 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5345 capacity = fix_small_capacity(env->sd, group);
5347 if (!cpumask_test_cpu(i, env->cpus))
5351 wl = weighted_cpuload(i);
5354 * When comparing with imbalance, use weighted_cpuload()
5355 * which is not scaled with the cpu power.
5357 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5361 * For the load comparisons with the other cpu's, consider
5362 * the weighted_cpuload() scaled with the cpu power, so that
5363 * the load can be moved away from the cpu that is potentially
5364 * running at a lower capacity.
5366 wl = (wl * SCHED_POWER_SCALE) / power;
5368 if (wl > max_load) {
5378 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5379 * so long as it is large enough.
5381 #define MAX_PINNED_INTERVAL 512
5383 /* Working cpumask for load_balance and load_balance_newidle. */
5384 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5386 static int need_active_balance(struct lb_env *env)
5388 struct sched_domain *sd = env->sd;
5390 if (env->idle == CPU_NEWLY_IDLE) {
5393 * ASYM_PACKING needs to force migrate tasks from busy but
5394 * higher numbered CPUs in order to pack all tasks in the
5395 * lowest numbered CPUs.
5397 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5401 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5404 static int active_load_balance_cpu_stop(void *data);
5407 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5408 * tasks if there is an imbalance.
5410 static int load_balance(int this_cpu, struct rq *this_rq,
5411 struct sched_domain *sd, enum cpu_idle_type idle,
5414 int ld_moved, cur_ld_moved, active_balance = 0;
5415 struct sched_group *group;
5417 unsigned long flags;
5418 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5420 struct lb_env env = {
5422 .dst_cpu = this_cpu,
5424 .dst_grpmask = sched_group_cpus(sd->groups),
5426 .loop_break = sched_nr_migrate_break,
5431 * For NEWLY_IDLE load_balancing, we don't need to consider
5432 * other cpus in our group
5434 if (idle == CPU_NEWLY_IDLE)
5435 env.dst_grpmask = NULL;
5437 cpumask_copy(cpus, cpu_active_mask);
5439 schedstat_inc(sd, lb_count[idle]);
5442 group = find_busiest_group(&env, balance);
5448 schedstat_inc(sd, lb_nobusyg[idle]);
5452 busiest = find_busiest_queue(&env, group);
5454 schedstat_inc(sd, lb_nobusyq[idle]);
5458 BUG_ON(busiest == env.dst_rq);
5460 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5463 if (busiest->nr_running > 1) {
5465 * Attempt to move tasks. If find_busiest_group has found
5466 * an imbalance but busiest->nr_running <= 1, the group is
5467 * still unbalanced. ld_moved simply stays zero, so it is
5468 * correctly treated as an imbalance.
5470 env.flags |= LBF_ALL_PINNED;
5471 env.src_cpu = busiest->cpu;
5472 env.src_rq = busiest;
5473 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5475 update_h_load(env.src_cpu);
5477 local_irq_save(flags);
5478 double_rq_lock(env.dst_rq, busiest);
5481 * cur_ld_moved - load moved in current iteration
5482 * ld_moved - cumulative load moved across iterations
5484 cur_ld_moved = move_tasks(&env);
5485 ld_moved += cur_ld_moved;
5486 double_rq_unlock(env.dst_rq, busiest);
5487 local_irq_restore(flags);
5490 * some other cpu did the load balance for us.
5492 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5493 resched_cpu(env.dst_cpu);
5495 if (env.flags & LBF_NEED_BREAK) {
5496 env.flags &= ~LBF_NEED_BREAK;
5501 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5502 * us and move them to an alternate dst_cpu in our sched_group
5503 * where they can run. The upper limit on how many times we
5504 * iterate on same src_cpu is dependent on number of cpus in our
5507 * This changes load balance semantics a bit on who can move
5508 * load to a given_cpu. In addition to the given_cpu itself
5509 * (or a ilb_cpu acting on its behalf where given_cpu is
5510 * nohz-idle), we now have balance_cpu in a position to move
5511 * load to given_cpu. In rare situations, this may cause
5512 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5513 * _independently_ and at _same_ time to move some load to
5514 * given_cpu) causing exceess load to be moved to given_cpu.
5515 * This however should not happen so much in practice and
5516 * moreover subsequent load balance cycles should correct the
5517 * excess load moved.
5519 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5521 env.dst_rq = cpu_rq(env.new_dst_cpu);
5522 env.dst_cpu = env.new_dst_cpu;
5523 env.flags &= ~LBF_SOME_PINNED;
5525 env.loop_break = sched_nr_migrate_break;
5527 /* Prevent to re-select dst_cpu via env's cpus */
5528 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5531 * Go back to "more_balance" rather than "redo" since we
5532 * need to continue with same src_cpu.
5537 /* All tasks on this runqueue were pinned by CPU affinity */
5538 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5539 cpumask_clear_cpu(cpu_of(busiest), cpus);
5540 if (!cpumask_empty(cpus)) {
5542 env.loop_break = sched_nr_migrate_break;
5550 schedstat_inc(sd, lb_failed[idle]);
5552 * Increment the failure counter only on periodic balance.
5553 * We do not want newidle balance, which can be very
5554 * frequent, pollute the failure counter causing
5555 * excessive cache_hot migrations and active balances.
5557 if (idle != CPU_NEWLY_IDLE)
5558 sd->nr_balance_failed++;
5560 if (need_active_balance(&env)) {
5561 raw_spin_lock_irqsave(&busiest->lock, flags);
5563 /* don't kick the active_load_balance_cpu_stop,
5564 * if the curr task on busiest cpu can't be
5567 if (!cpumask_test_cpu(this_cpu,
5568 tsk_cpus_allowed(busiest->curr))) {
5569 raw_spin_unlock_irqrestore(&busiest->lock,
5571 env.flags |= LBF_ALL_PINNED;
5572 goto out_one_pinned;
5576 * ->active_balance synchronizes accesses to
5577 * ->active_balance_work. Once set, it's cleared
5578 * only after active load balance is finished.
5580 if (!busiest->active_balance) {
5581 busiest->active_balance = 1;
5582 busiest->push_cpu = this_cpu;
5585 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5587 if (active_balance) {
5588 stop_one_cpu_nowait(cpu_of(busiest),
5589 active_load_balance_cpu_stop, busiest,
5590 &busiest->active_balance_work);
5594 * We've kicked active balancing, reset the failure
5597 sd->nr_balance_failed = sd->cache_nice_tries+1;
5600 sd->nr_balance_failed = 0;
5602 if (likely(!active_balance)) {
5603 /* We were unbalanced, so reset the balancing interval */
5604 sd->balance_interval = sd->min_interval;
5607 * If we've begun active balancing, start to back off. This
5608 * case may not be covered by the all_pinned logic if there
5609 * is only 1 task on the busy runqueue (because we don't call
5612 if (sd->balance_interval < sd->max_interval)
5613 sd->balance_interval *= 2;
5619 schedstat_inc(sd, lb_balanced[idle]);
5621 sd->nr_balance_failed = 0;
5624 /* tune up the balancing interval */
5625 if (((env.flags & LBF_ALL_PINNED) &&
5626 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5627 (sd->balance_interval < sd->max_interval))
5628 sd->balance_interval *= 2;
5636 * idle_balance is called by schedule() if this_cpu is about to become
5637 * idle. Attempts to pull tasks from other CPUs.
5639 void idle_balance(int this_cpu, struct rq *this_rq)
5641 struct sched_domain *sd;
5642 int pulled_task = 0;
5643 unsigned long next_balance = jiffies + HZ;
5645 this_rq->idle_stamp = this_rq->clock;
5647 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5651 * Drop the rq->lock, but keep IRQ/preempt disabled.
5653 raw_spin_unlock(&this_rq->lock);
5655 update_blocked_averages(this_cpu);
5657 for_each_domain(this_cpu, sd) {
5658 unsigned long interval;
5661 if (!(sd->flags & SD_LOAD_BALANCE))
5664 if (sd->flags & SD_BALANCE_NEWIDLE) {
5665 /* If we've pulled tasks over stop searching: */
5666 pulled_task = load_balance(this_cpu, this_rq,
5667 sd, CPU_NEWLY_IDLE, &balance);
5670 interval = msecs_to_jiffies(sd->balance_interval);
5671 if (time_after(next_balance, sd->last_balance + interval))
5672 next_balance = sd->last_balance + interval;
5674 this_rq->idle_stamp = 0;
5680 raw_spin_lock(&this_rq->lock);
5682 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5684 * We are going idle. next_balance may be set based on
5685 * a busy processor. So reset next_balance.
5687 this_rq->next_balance = next_balance;
5692 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5693 * running tasks off the busiest CPU onto idle CPUs. It requires at
5694 * least 1 task to be running on each physical CPU where possible, and
5695 * avoids physical / logical imbalances.
5697 static int active_load_balance_cpu_stop(void *data)
5699 struct rq *busiest_rq = data;
5700 int busiest_cpu = cpu_of(busiest_rq);
5701 int target_cpu = busiest_rq->push_cpu;
5702 struct rq *target_rq = cpu_rq(target_cpu);
5703 struct sched_domain *sd;
5705 raw_spin_lock_irq(&busiest_rq->lock);
5707 /* make sure the requested cpu hasn't gone down in the meantime */
5708 if (unlikely(busiest_cpu != smp_processor_id() ||
5709 !busiest_rq->active_balance))
5712 /* Is there any task to move? */
5713 if (busiest_rq->nr_running <= 1)
5717 * This condition is "impossible", if it occurs
5718 * we need to fix it. Originally reported by
5719 * Bjorn Helgaas on a 128-cpu setup.
5721 BUG_ON(busiest_rq == target_rq);
5723 /* move a task from busiest_rq to target_rq */
5724 double_lock_balance(busiest_rq, target_rq);
5726 /* Search for an sd spanning us and the target CPU. */
5728 for_each_domain(target_cpu, sd) {
5729 if ((sd->flags & SD_LOAD_BALANCE) &&
5730 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5735 struct lb_env env = {
5737 .dst_cpu = target_cpu,
5738 .dst_rq = target_rq,
5739 .src_cpu = busiest_rq->cpu,
5740 .src_rq = busiest_rq,
5744 schedstat_inc(sd, alb_count);
5746 if (move_one_task(&env))
5747 schedstat_inc(sd, alb_pushed);
5749 schedstat_inc(sd, alb_failed);
5752 double_unlock_balance(busiest_rq, target_rq);
5754 busiest_rq->active_balance = 0;
5755 raw_spin_unlock_irq(&busiest_rq->lock);
5759 #ifdef CONFIG_NO_HZ_COMMON
5761 * idle load balancing details
5762 * - When one of the busy CPUs notice that there may be an idle rebalancing
5763 * needed, they will kick the idle load balancer, which then does idle
5764 * load balancing for all the idle CPUs.
5767 cpumask_var_t idle_cpus_mask;
5769 unsigned long next_balance; /* in jiffy units */
5770 } nohz ____cacheline_aligned;
5772 static inline int find_new_ilb(int call_cpu)
5774 int ilb = cpumask_first(nohz.idle_cpus_mask);
5776 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5783 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5784 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5785 * CPU (if there is one).
5787 static void nohz_balancer_kick(int cpu)
5791 nohz.next_balance++;
5793 ilb_cpu = find_new_ilb(cpu);
5795 if (ilb_cpu >= nr_cpu_ids)
5798 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5801 * Use smp_send_reschedule() instead of resched_cpu().
5802 * This way we generate a sched IPI on the target cpu which
5803 * is idle. And the softirq performing nohz idle load balance
5804 * will be run before returning from the IPI.
5806 smp_send_reschedule(ilb_cpu);
5810 static inline void nohz_balance_exit_idle(int cpu)
5812 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5813 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5814 atomic_dec(&nohz.nr_cpus);
5815 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5819 static inline void set_cpu_sd_state_busy(void)
5821 struct sched_domain *sd;
5822 int cpu = smp_processor_id();
5825 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5827 if (!sd || !sd->nohz_idle)
5831 for (; sd; sd = sd->parent)
5832 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5837 void set_cpu_sd_state_idle(void)
5839 struct sched_domain *sd;
5840 int cpu = smp_processor_id();
5843 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5845 if (!sd || sd->nohz_idle)
5849 for (; sd; sd = sd->parent)
5850 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5856 * This routine will record that the cpu is going idle with tick stopped.
5857 * This info will be used in performing idle load balancing in the future.
5859 void nohz_balance_enter_idle(int cpu)
5862 * If this cpu is going down, then nothing needs to be done.
5864 if (!cpu_active(cpu))
5867 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5870 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5871 atomic_inc(&nohz.nr_cpus);
5872 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5875 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5876 unsigned long action, void *hcpu)
5878 switch (action & ~CPU_TASKS_FROZEN) {
5880 nohz_balance_exit_idle(smp_processor_id());
5888 static DEFINE_SPINLOCK(balancing);
5891 * Scale the max load_balance interval with the number of CPUs in the system.
5892 * This trades load-balance latency on larger machines for less cross talk.
5894 void update_max_interval(void)
5896 max_load_balance_interval = HZ*num_online_cpus()/10;
5900 * It checks each scheduling domain to see if it is due to be balanced,
5901 * and initiates a balancing operation if so.
5903 * Balancing parameters are set up in init_sched_domains.
5905 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5908 struct rq *rq = cpu_rq(cpu);
5909 unsigned long interval;
5910 struct sched_domain *sd;
5911 /* Earliest time when we have to do rebalance again */
5912 unsigned long next_balance = jiffies + 60*HZ;
5913 int update_next_balance = 0;
5916 update_blocked_averages(cpu);
5919 for_each_domain(cpu, sd) {
5920 if (!(sd->flags & SD_LOAD_BALANCE))
5923 interval = sd->balance_interval;
5924 if (idle != CPU_IDLE)
5925 interval *= sd->busy_factor;
5927 /* scale ms to jiffies */
5928 interval = msecs_to_jiffies(interval);
5929 interval = clamp(interval, 1UL, max_load_balance_interval);
5931 need_serialize = sd->flags & SD_SERIALIZE;
5933 if (need_serialize) {
5934 if (!spin_trylock(&balancing))
5938 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5939 if (load_balance(cpu, rq, sd, idle, &balance)) {
5941 * The LBF_SOME_PINNED logic could have changed
5942 * env->dst_cpu, so we can't know our idle
5943 * state even if we migrated tasks. Update it.
5945 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5947 sd->last_balance = jiffies;
5950 spin_unlock(&balancing);
5952 if (time_after(next_balance, sd->last_balance + interval)) {
5953 next_balance = sd->last_balance + interval;
5954 update_next_balance = 1;
5958 * Stop the load balance at this level. There is another
5959 * CPU in our sched group which is doing load balancing more
5968 * next_balance will be updated only when there is a need.
5969 * When the cpu is attached to null domain for ex, it will not be
5972 if (likely(update_next_balance))
5973 rq->next_balance = next_balance;
5976 #ifdef CONFIG_NO_HZ_COMMON
5978 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5979 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5981 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5983 struct rq *this_rq = cpu_rq(this_cpu);
5987 if (idle != CPU_IDLE ||
5988 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5991 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5992 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5996 * If this cpu gets work to do, stop the load balancing
5997 * work being done for other cpus. Next load
5998 * balancing owner will pick it up.
6003 rq = cpu_rq(balance_cpu);
6005 raw_spin_lock_irq(&rq->lock);
6006 update_rq_clock(rq);
6007 update_idle_cpu_load(rq);
6008 raw_spin_unlock_irq(&rq->lock);
6010 rebalance_domains(balance_cpu, CPU_IDLE);
6012 if (time_after(this_rq->next_balance, rq->next_balance))
6013 this_rq->next_balance = rq->next_balance;
6015 nohz.next_balance = this_rq->next_balance;
6017 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6021 * Current heuristic for kicking the idle load balancer in the presence
6022 * of an idle cpu is the system.
6023 * - This rq has more than one task.
6024 * - At any scheduler domain level, this cpu's scheduler group has multiple
6025 * busy cpu's exceeding the group's power.
6026 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6027 * domain span are idle.
6029 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6031 unsigned long now = jiffies;
6032 struct sched_domain *sd;
6034 if (unlikely(idle_cpu(cpu)))
6038 * We may be recently in ticked or tickless idle mode. At the first
6039 * busy tick after returning from idle, we will update the busy stats.
6041 set_cpu_sd_state_busy();
6042 nohz_balance_exit_idle(cpu);
6045 * None are in tickless mode and hence no need for NOHZ idle load
6048 if (likely(!atomic_read(&nohz.nr_cpus)))
6051 if (time_before(now, nohz.next_balance))
6054 if (rq->nr_running >= 2)
6058 for_each_domain(cpu, sd) {
6059 struct sched_group *sg = sd->groups;
6060 struct sched_group_power *sgp = sg->sgp;
6061 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6063 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6064 goto need_kick_unlock;
6066 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6067 && (cpumask_first_and(nohz.idle_cpus_mask,
6068 sched_domain_span(sd)) < cpu))
6069 goto need_kick_unlock;
6071 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6083 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6086 #ifdef CONFIG_SCHED_HMP
6087 /* Check if task should migrate to a faster cpu */
6088 static unsigned int hmp_up_migration(int cpu, struct sched_entity *se)
6090 struct task_struct *p = task_of(se);
6091 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6094 if (hmp_cpu_is_fastest(cpu))
6097 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6098 /* Filter by task priority */
6099 if (p->prio >= hmp_up_prio)
6103 /* Let the task load settle before doing another up migration */
6104 now = cfs_rq_clock_task(cfs_rq);
6105 if (((now - se->avg.hmp_last_up_migration) >> 10)
6106 < hmp_next_up_threshold)
6109 if (cpumask_intersects(&hmp_faster_domain(cpu)->cpus,
6110 tsk_cpus_allowed(p))
6111 && se->avg.load_avg_ratio > hmp_up_threshold) {
6117 /* Check if task should migrate to a slower cpu */
6118 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
6120 struct task_struct *p = task_of(se);
6121 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6124 if (hmp_cpu_is_slowest(cpu))
6127 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6128 /* Filter by task priority */
6129 if ((p->prio >= hmp_up_prio) &&
6130 cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6131 tsk_cpus_allowed(p))) {
6136 /* Let the task load settle before doing another down migration */
6137 now = cfs_rq_clock_task(cfs_rq);
6138 if (((now - se->avg.hmp_last_down_migration) >> 10)
6139 < hmp_next_down_threshold)
6142 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6143 tsk_cpus_allowed(p))
6144 && se->avg.load_avg_ratio < hmp_down_threshold) {
6151 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6152 * Ideally this function should be merged with can_migrate_task() to avoid
6155 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
6157 int tsk_cache_hot = 0;
6160 * We do not migrate tasks that are:
6161 * 1) running (obviously), or
6162 * 2) cannot be migrated to this CPU due to cpus_allowed
6164 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6165 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6168 env->flags &= ~LBF_ALL_PINNED;
6170 if (task_running(env->src_rq, p)) {
6171 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6176 * Aggressive migration if:
6177 * 1) task is cache cold, or
6178 * 2) too many balance attempts have failed.
6181 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
6182 if (!tsk_cache_hot ||
6183 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6184 #ifdef CONFIG_SCHEDSTATS
6185 if (tsk_cache_hot) {
6186 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6187 schedstat_inc(p, se.statistics.nr_forced_migrations);
6197 * move_specific_task tries to move a specific task.
6198 * Returns 1 if successful and 0 otherwise.
6199 * Called with both runqueues locked.
6201 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
6203 struct task_struct *p, *n;
6205 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6206 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
6210 if (!hmp_can_migrate_task(p, env))
6212 /* Check if we found the right task */
6218 * Right now, this is only the third place move_task()
6219 * is called, so we can safely collect move_task()
6220 * stats here rather than inside move_task().
6222 schedstat_inc(env->sd, lb_gained[env->idle]);
6229 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
6230 * migrate a specific task from one runqueue to another.
6231 * hmp_force_up_migration uses this to push a currently running task
6233 * Based on active_load_balance_stop_cpu and can potentially be merged.
6235 static int hmp_active_task_migration_cpu_stop(void *data)
6237 struct rq *busiest_rq = data;
6238 struct task_struct *p = busiest_rq->migrate_task;
6239 int busiest_cpu = cpu_of(busiest_rq);
6240 int target_cpu = busiest_rq->push_cpu;
6241 struct rq *target_rq = cpu_rq(target_cpu);
6242 struct sched_domain *sd;
6244 raw_spin_lock_irq(&busiest_rq->lock);
6245 /* make sure the requested cpu hasn't gone down in the meantime */
6246 if (unlikely(busiest_cpu != smp_processor_id() ||
6247 !busiest_rq->active_balance)) {
6250 /* Is there any task to move? */
6251 if (busiest_rq->nr_running <= 1)
6253 /* Task has migrated meanwhile, abort forced migration */
6254 if (task_rq(p) != busiest_rq)
6257 * This condition is "impossible", if it occurs
6258 * we need to fix it. Originally reported by
6259 * Bjorn Helgaas on a 128-cpu setup.
6261 BUG_ON(busiest_rq == target_rq);
6263 /* move a task from busiest_rq to target_rq */
6264 double_lock_balance(busiest_rq, target_rq);
6266 /* Search for an sd spanning us and the target CPU. */
6268 for_each_domain(target_cpu, sd) {
6269 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6274 struct lb_env env = {
6276 .dst_cpu = target_cpu,
6277 .dst_rq = target_rq,
6278 .src_cpu = busiest_rq->cpu,
6279 .src_rq = busiest_rq,
6283 schedstat_inc(sd, alb_count);
6285 if (move_specific_task(&env, p))
6286 schedstat_inc(sd, alb_pushed);
6288 schedstat_inc(sd, alb_failed);
6291 double_unlock_balance(busiest_rq, target_rq);
6293 busiest_rq->active_balance = 0;
6294 raw_spin_unlock_irq(&busiest_rq->lock);
6298 static DEFINE_SPINLOCK(hmp_force_migration);
6301 * hmp_force_up_migration checks runqueues for tasks that need to
6302 * be actively migrated to a faster cpu.
6304 static void hmp_force_up_migration(int this_cpu)
6307 struct sched_entity *curr;
6309 unsigned long flags;
6311 struct task_struct *p;
6313 if (!spin_trylock(&hmp_force_migration))
6315 for_each_online_cpu(cpu) {
6317 target = cpu_rq(cpu);
6318 raw_spin_lock_irqsave(&target->lock, flags);
6319 curr = target->cfs.curr;
6321 raw_spin_unlock_irqrestore(&target->lock, flags);
6324 if (!entity_is_task(curr)) {
6325 struct cfs_rq *cfs_rq;
6327 cfs_rq = group_cfs_rq(curr);
6329 curr = cfs_rq->curr;
6330 cfs_rq = group_cfs_rq(curr);
6334 if (hmp_up_migration(cpu, curr)) {
6335 if (!target->active_balance) {
6336 target->active_balance = 1;
6337 target->push_cpu = hmp_select_faster_cpu(p, cpu);
6338 target->migrate_task = p;
6340 trace_sched_hmp_migrate(p, target->push_cpu, 1);
6341 hmp_next_up_delay(&p->se, target->push_cpu);
6344 raw_spin_unlock_irqrestore(&target->lock, flags);
6346 stop_one_cpu_nowait(cpu_of(target),
6347 hmp_active_task_migration_cpu_stop,
6348 target, &target->active_balance_work);
6350 spin_unlock(&hmp_force_migration);
6353 static void hmp_force_up_migration(int this_cpu) { }
6354 #endif /* CONFIG_SCHED_HMP */
6357 * run_rebalance_domains is triggered when needed from the scheduler tick.
6358 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6360 static void run_rebalance_domains(struct softirq_action *h)
6362 int this_cpu = smp_processor_id();
6363 struct rq *this_rq = cpu_rq(this_cpu);
6364 enum cpu_idle_type idle = this_rq->idle_balance ?
6365 CPU_IDLE : CPU_NOT_IDLE;
6367 hmp_force_up_migration(this_cpu);
6369 rebalance_domains(this_cpu, idle);
6372 * If this cpu has a pending nohz_balance_kick, then do the
6373 * balancing on behalf of the other idle cpus whose ticks are
6376 nohz_idle_balance(this_cpu, idle);
6379 static inline int on_null_domain(int cpu)
6381 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6385 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6387 void trigger_load_balance(struct rq *rq, int cpu)
6389 /* Don't need to rebalance while attached to NULL domain */
6390 if (time_after_eq(jiffies, rq->next_balance) &&
6391 likely(!on_null_domain(cpu)))
6392 raise_softirq(SCHED_SOFTIRQ);
6393 #ifdef CONFIG_NO_HZ_COMMON
6394 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6395 nohz_balancer_kick(cpu);
6399 static void rq_online_fair(struct rq *rq)
6401 #ifdef CONFIG_SCHED_HMP
6402 hmp_online_cpu(rq->cpu);
6407 static void rq_offline_fair(struct rq *rq)
6409 #ifdef CONFIG_SCHED_HMP
6410 hmp_offline_cpu(rq->cpu);
6414 /* Ensure any throttled groups are reachable by pick_next_task */
6415 unthrottle_offline_cfs_rqs(rq);
6418 #endif /* CONFIG_SMP */
6421 * scheduler tick hitting a task of our scheduling class:
6423 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6425 struct cfs_rq *cfs_rq;
6426 struct sched_entity *se = &curr->se;
6428 for_each_sched_entity(se) {
6429 cfs_rq = cfs_rq_of(se);
6430 entity_tick(cfs_rq, se, queued);
6433 if (sched_feat_numa(NUMA))
6434 task_tick_numa(rq, curr);
6436 update_rq_runnable_avg(rq, 1);
6440 * called on fork with the child task as argument from the parent's context
6441 * - child not yet on the tasklist
6442 * - preemption disabled
6444 static void task_fork_fair(struct task_struct *p)
6446 struct cfs_rq *cfs_rq;
6447 struct sched_entity *se = &p->se, *curr;
6448 int this_cpu = smp_processor_id();
6449 struct rq *rq = this_rq();
6450 unsigned long flags;
6452 raw_spin_lock_irqsave(&rq->lock, flags);
6454 update_rq_clock(rq);
6456 cfs_rq = task_cfs_rq(current);
6457 curr = cfs_rq->curr;
6459 if (unlikely(task_cpu(p) != this_cpu)) {
6461 __set_task_cpu(p, this_cpu);
6465 update_curr(cfs_rq);
6468 se->vruntime = curr->vruntime;
6469 place_entity(cfs_rq, se, 1);
6471 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6473 * Upon rescheduling, sched_class::put_prev_task() will place
6474 * 'current' within the tree based on its new key value.
6476 swap(curr->vruntime, se->vruntime);
6477 resched_task(rq->curr);
6480 se->vruntime -= cfs_rq->min_vruntime;
6482 raw_spin_unlock_irqrestore(&rq->lock, flags);
6486 * Priority of the task has changed. Check to see if we preempt
6490 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6496 * Reschedule if we are currently running on this runqueue and
6497 * our priority decreased, or if we are not currently running on
6498 * this runqueue and our priority is higher than the current's
6500 if (rq->curr == p) {
6501 if (p->prio > oldprio)
6502 resched_task(rq->curr);
6504 check_preempt_curr(rq, p, 0);
6507 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6509 struct sched_entity *se = &p->se;
6510 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6513 * Ensure the task's vruntime is normalized, so that when its
6514 * switched back to the fair class the enqueue_entity(.flags=0) will
6515 * do the right thing.
6517 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6518 * have normalized the vruntime, if it was !on_rq, then only when
6519 * the task is sleeping will it still have non-normalized vruntime.
6521 if (!se->on_rq && p->state != TASK_RUNNING) {
6523 * Fix up our vruntime so that the current sleep doesn't
6524 * cause 'unlimited' sleep bonus.
6526 place_entity(cfs_rq, se, 0);
6527 se->vruntime -= cfs_rq->min_vruntime;
6530 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
6532 * Remove our load from contribution when we leave sched_fair
6533 * and ensure we don't carry in an old decay_count if we
6536 if (p->se.avg.decay_count) {
6537 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
6538 __synchronize_entity_decay(&p->se);
6539 subtract_blocked_load_contrib(cfs_rq,
6540 p->se.avg.load_avg_contrib);
6546 * We switched to the sched_fair class.
6548 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6554 * We were most likely switched from sched_rt, so
6555 * kick off the schedule if running, otherwise just see
6556 * if we can still preempt the current task.
6559 resched_task(rq->curr);
6561 check_preempt_curr(rq, p, 0);
6564 /* Account for a task changing its policy or group.
6566 * This routine is mostly called to set cfs_rq->curr field when a task
6567 * migrates between groups/classes.
6569 static void set_curr_task_fair(struct rq *rq)
6571 struct sched_entity *se = &rq->curr->se;
6573 for_each_sched_entity(se) {
6574 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6576 set_next_entity(cfs_rq, se);
6577 /* ensure bandwidth has been allocated on our new cfs_rq */
6578 account_cfs_rq_runtime(cfs_rq, 0);
6582 void init_cfs_rq(struct cfs_rq *cfs_rq)
6584 cfs_rq->tasks_timeline = RB_ROOT;
6585 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6586 #ifndef CONFIG_64BIT
6587 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6589 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
6590 atomic64_set(&cfs_rq->decay_counter, 1);
6591 atomic64_set(&cfs_rq->removed_load, 0);
6595 #ifdef CONFIG_FAIR_GROUP_SCHED
6596 static void task_move_group_fair(struct task_struct *p, int on_rq)
6598 struct cfs_rq *cfs_rq;
6600 * If the task was not on the rq at the time of this cgroup movement
6601 * it must have been asleep, sleeping tasks keep their ->vruntime
6602 * absolute on their old rq until wakeup (needed for the fair sleeper
6603 * bonus in place_entity()).
6605 * If it was on the rq, we've just 'preempted' it, which does convert
6606 * ->vruntime to a relative base.
6608 * Make sure both cases convert their relative position when migrating
6609 * to another cgroup's rq. This does somewhat interfere with the
6610 * fair sleeper stuff for the first placement, but who cares.
6613 * When !on_rq, vruntime of the task has usually NOT been normalized.
6614 * But there are some cases where it has already been normalized:
6616 * - Moving a forked child which is waiting for being woken up by
6617 * wake_up_new_task().
6618 * - Moving a task which has been woken up by try_to_wake_up() and
6619 * waiting for actually being woken up by sched_ttwu_pending().
6621 * To prevent boost or penalty in the new cfs_rq caused by delta
6622 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6624 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6628 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6629 set_task_rq(p, task_cpu(p));
6631 cfs_rq = cfs_rq_of(&p->se);
6632 p->se.vruntime += cfs_rq->min_vruntime;
6635 * migrate_task_rq_fair() will have removed our previous
6636 * contribution, but we must synchronize for ongoing future
6639 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6640 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6645 void free_fair_sched_group(struct task_group *tg)
6649 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6651 for_each_possible_cpu(i) {
6653 kfree(tg->cfs_rq[i]);
6662 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6664 struct cfs_rq *cfs_rq;
6665 struct sched_entity *se;
6668 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6671 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6675 tg->shares = NICE_0_LOAD;
6677 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6679 for_each_possible_cpu(i) {
6680 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6681 GFP_KERNEL, cpu_to_node(i));
6685 se = kzalloc_node(sizeof(struct sched_entity),
6686 GFP_KERNEL, cpu_to_node(i));
6690 init_cfs_rq(cfs_rq);
6691 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6702 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6704 struct rq *rq = cpu_rq(cpu);
6705 unsigned long flags;
6708 * Only empty task groups can be destroyed; so we can speculatively
6709 * check on_list without danger of it being re-added.
6711 if (!tg->cfs_rq[cpu]->on_list)
6714 raw_spin_lock_irqsave(&rq->lock, flags);
6715 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6716 raw_spin_unlock_irqrestore(&rq->lock, flags);
6719 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6720 struct sched_entity *se, int cpu,
6721 struct sched_entity *parent)
6723 struct rq *rq = cpu_rq(cpu);
6727 init_cfs_rq_runtime(cfs_rq);
6729 tg->cfs_rq[cpu] = cfs_rq;
6732 /* se could be NULL for root_task_group */
6737 se->cfs_rq = &rq->cfs;
6739 se->cfs_rq = parent->my_q;
6742 update_load_set(&se->load, 0);
6743 se->parent = parent;
6746 static DEFINE_MUTEX(shares_mutex);
6748 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6751 unsigned long flags;
6754 * We can't change the weight of the root cgroup.
6759 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6761 mutex_lock(&shares_mutex);
6762 if (tg->shares == shares)
6765 tg->shares = shares;
6766 for_each_possible_cpu(i) {
6767 struct rq *rq = cpu_rq(i);
6768 struct sched_entity *se;
6771 /* Propagate contribution to hierarchy */
6772 raw_spin_lock_irqsave(&rq->lock, flags);
6773 for_each_sched_entity(se)
6774 update_cfs_shares(group_cfs_rq(se));
6775 raw_spin_unlock_irqrestore(&rq->lock, flags);
6779 mutex_unlock(&shares_mutex);
6782 #else /* CONFIG_FAIR_GROUP_SCHED */
6784 void free_fair_sched_group(struct task_group *tg) { }
6786 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6791 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6793 #endif /* CONFIG_FAIR_GROUP_SCHED */
6796 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6798 struct sched_entity *se = &task->se;
6799 unsigned int rr_interval = 0;
6802 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6805 if (rq->cfs.load.weight)
6806 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6812 * All the scheduling class methods:
6814 const struct sched_class fair_sched_class = {
6815 .next = &idle_sched_class,
6816 .enqueue_task = enqueue_task_fair,
6817 .dequeue_task = dequeue_task_fair,
6818 .yield_task = yield_task_fair,
6819 .yield_to_task = yield_to_task_fair,
6821 .check_preempt_curr = check_preempt_wakeup,
6823 .pick_next_task = pick_next_task_fair,
6824 .put_prev_task = put_prev_task_fair,
6827 .select_task_rq = select_task_rq_fair,
6828 #ifdef CONFIG_FAIR_GROUP_SCHED
6829 .migrate_task_rq = migrate_task_rq_fair,
6831 .rq_online = rq_online_fair,
6832 .rq_offline = rq_offline_fair,
6834 .task_waking = task_waking_fair,
6837 .set_curr_task = set_curr_task_fair,
6838 .task_tick = task_tick_fair,
6839 .task_fork = task_fork_fair,
6841 .prio_changed = prio_changed_fair,
6842 .switched_from = switched_from_fair,
6843 .switched_to = switched_to_fair,
6845 .get_rr_interval = get_rr_interval_fair,
6847 #ifdef CONFIG_FAIR_GROUP_SCHED
6848 .task_move_group = task_move_group_fair,
6852 #ifdef CONFIG_SCHED_DEBUG
6853 void print_cfs_stats(struct seq_file *m, int cpu)
6855 struct cfs_rq *cfs_rq;
6858 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6859 print_cfs_rq(m, cpu, cfs_rq);
6864 __init void init_sched_fair_class(void)
6867 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6869 #ifdef CONFIG_NO_HZ_COMMON
6870 nohz.next_balance = jiffies;
6871 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6872 cpu_notifier(sched_ilb_notifier, 0);
6875 #ifdef CONFIG_SCHED_HMP
6876 hmp_cpu_mask_setup();