2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * Approximate time to scan a full NUMA task in ms. The task scan period is
822 * calculated based on the tasks virtual memory size and
823 * numa_balancing_scan_size.
825 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
826 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
829 /* Portion of address space to scan in MB */
830 unsigned int sysctl_numa_balancing_scan_size = 256;
832 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
833 unsigned int sysctl_numa_balancing_scan_delay = 1000;
835 static unsigned int task_nr_scan_windows(struct task_struct *p)
837 unsigned long rss = 0;
838 unsigned long nr_scan_pages;
841 * Calculations based on RSS as non-present and empty pages are skipped
842 * by the PTE scanner and NUMA hinting faults should be trapped based
845 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
846 rss = get_mm_rss(p->mm);
850 rss = round_up(rss, nr_scan_pages);
851 return rss / nr_scan_pages;
854 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
855 #define MAX_SCAN_WINDOW 2560
857 static unsigned int task_scan_min(struct task_struct *p)
859 unsigned int scan, floor;
860 unsigned int windows = 1;
862 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
863 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
864 floor = 1000 / windows;
866 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
867 return max_t(unsigned int, floor, scan);
870 static unsigned int task_scan_max(struct task_struct *p)
872 unsigned int smin = task_scan_min(p);
875 /* Watch for min being lower than max due to floor calculations */
876 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
877 return max(smin, smax);
880 static void task_numa_placement(struct task_struct *p)
882 int seq, nid, max_nid = -1;
883 unsigned long max_faults = 0;
885 if (!p->mm) /* for example, ksmd faulting in a user's mm */
887 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
888 if (p->numa_scan_seq == seq)
890 p->numa_scan_seq = seq;
891 p->numa_scan_period_max = task_scan_max(p);
893 /* Find the node with the highest number of faults */
894 for_each_online_node(nid) {
895 unsigned long faults;
897 /* Decay existing window and copy faults since last scan */
898 p->numa_faults[nid] >>= 1;
899 p->numa_faults[nid] += p->numa_faults_buffer[nid];
900 p->numa_faults_buffer[nid] = 0;
902 faults = p->numa_faults[nid];
903 if (faults > max_faults) {
909 /* Update the tasks preferred node if necessary */
910 if (max_faults && max_nid != p->numa_preferred_nid)
911 p->numa_preferred_nid = max_nid;
915 * Got a PROT_NONE fault for a page on @node.
917 void task_numa_fault(int node, int pages, bool migrated)
919 struct task_struct *p = current;
921 if (!numabalancing_enabled)
924 /* Allocate buffer to track faults on a per-node basis */
925 if (unlikely(!p->numa_faults)) {
926 int size = sizeof(*p->numa_faults) * nr_node_ids;
928 /* numa_faults and numa_faults_buffer share the allocation */
929 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
933 BUG_ON(p->numa_faults_buffer);
934 p->numa_faults_buffer = p->numa_faults + nr_node_ids;
938 * If pages are properly placed (did not migrate) then scan slower.
939 * This is reset periodically in case of phase changes
942 /* Initialise if necessary */
943 if (!p->numa_scan_period_max)
944 p->numa_scan_period_max = task_scan_max(p);
946 p->numa_scan_period = min(p->numa_scan_period_max,
947 p->numa_scan_period + 10);
950 task_numa_placement(p);
952 p->numa_faults_buffer[node] += pages;
955 static void reset_ptenuma_scan(struct task_struct *p)
957 ACCESS_ONCE(p->mm->numa_scan_seq)++;
958 p->mm->numa_scan_offset = 0;
962 * The expensive part of numa migration is done from task_work context.
963 * Triggered from task_tick_numa().
965 void task_numa_work(struct callback_head *work)
967 unsigned long migrate, next_scan, now = jiffies;
968 struct task_struct *p = current;
969 struct mm_struct *mm = p->mm;
970 struct vm_area_struct *vma;
971 unsigned long start, end;
972 unsigned long nr_pte_updates = 0;
975 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
977 work->next = work; /* protect against double add */
979 * Who cares about NUMA placement when they're dying.
981 * NOTE: make sure not to dereference p->mm before this check,
982 * exit_task_work() happens _after_ exit_mm() so we could be called
983 * without p->mm even though we still had it when we enqueued this
986 if (p->flags & PF_EXITING)
989 if (!mm->numa_next_reset || !mm->numa_next_scan) {
990 mm->numa_next_scan = now +
991 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
992 mm->numa_next_reset = now +
993 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
997 * Reset the scan period if enough time has gone by. Objective is that
998 * scanning will be reduced if pages are properly placed. As tasks
999 * can enter different phases this needs to be re-examined. Lacking
1000 * proper tracking of reference behaviour, this blunt hammer is used.
1002 migrate = mm->numa_next_reset;
1003 if (time_after(now, migrate)) {
1004 p->numa_scan_period = task_scan_min(p);
1005 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1006 xchg(&mm->numa_next_reset, next_scan);
1010 * Enforce maximal scan/migration frequency..
1012 migrate = mm->numa_next_scan;
1013 if (time_before(now, migrate))
1016 if (p->numa_scan_period == 0) {
1017 p->numa_scan_period_max = task_scan_max(p);
1018 p->numa_scan_period = task_scan_min(p);
1021 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1022 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1026 * Delay this task enough that another task of this mm will likely win
1027 * the next time around.
1029 p->node_stamp += 2 * TICK_NSEC;
1031 start = mm->numa_scan_offset;
1032 pages = sysctl_numa_balancing_scan_size;
1033 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1037 down_read(&mm->mmap_sem);
1038 vma = find_vma(mm, start);
1040 reset_ptenuma_scan(p);
1044 for (; vma; vma = vma->vm_next) {
1045 if (!vma_migratable(vma))
1048 /* Skip small VMAs. They are not likely to be of relevance */
1049 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1053 start = max(start, vma->vm_start);
1054 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1055 end = min(end, vma->vm_end);
1056 nr_pte_updates += change_prot_numa(vma, start, end);
1059 * Scan sysctl_numa_balancing_scan_size but ensure that
1060 * at least one PTE is updated so that unused virtual
1061 * address space is quickly skipped.
1064 pages -= (end - start) >> PAGE_SHIFT;
1069 } while (end != vma->vm_end);
1074 * If the whole process was scanned without updates then no NUMA
1075 * hinting faults are being recorded and scan rate should be lower.
1077 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1078 p->numa_scan_period = min(p->numa_scan_period_max,
1079 p->numa_scan_period << 1);
1081 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1082 mm->numa_next_scan = next_scan;
1086 * It is possible to reach the end of the VMA list but the last few
1087 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1088 * would find the !migratable VMA on the next scan but not reset the
1089 * scanner to the start so check it now.
1092 mm->numa_scan_offset = start;
1094 reset_ptenuma_scan(p);
1095 up_read(&mm->mmap_sem);
1099 * Drive the periodic memory faults..
1101 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1103 struct callback_head *work = &curr->numa_work;
1107 * We don't care about NUMA placement if we don't have memory.
1109 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1113 * Using runtime rather than walltime has the dual advantage that
1114 * we (mostly) drive the selection from busy threads and that the
1115 * task needs to have done some actual work before we bother with
1118 now = curr->se.sum_exec_runtime;
1119 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1121 if (now - curr->node_stamp > period) {
1122 if (!curr->node_stamp)
1123 curr->numa_scan_period = task_scan_min(curr);
1124 curr->node_stamp += period;
1126 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1127 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1128 task_work_add(curr, work, true);
1133 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1136 #endif /* CONFIG_NUMA_BALANCING */
1139 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1141 update_load_add(&cfs_rq->load, se->load.weight);
1142 if (!parent_entity(se))
1143 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1145 if (entity_is_task(se))
1146 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1148 cfs_rq->nr_running++;
1152 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1154 update_load_sub(&cfs_rq->load, se->load.weight);
1155 if (!parent_entity(se))
1156 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1157 if (entity_is_task(se))
1158 list_del_init(&se->group_node);
1159 cfs_rq->nr_running--;
1162 #ifdef CONFIG_FAIR_GROUP_SCHED
1164 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1169 * Use this CPU's actual weight instead of the last load_contribution
1170 * to gain a more accurate current total weight. See
1171 * update_cfs_rq_load_contribution().
1173 tg_weight = atomic_long_read(&tg->load_avg);
1174 tg_weight -= cfs_rq->tg_load_contrib;
1175 tg_weight += cfs_rq->load.weight;
1180 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1182 long tg_weight, load, shares;
1184 tg_weight = calc_tg_weight(tg, cfs_rq);
1185 load = cfs_rq->load.weight;
1187 shares = (tg->shares * load);
1189 shares /= tg_weight;
1191 if (shares < MIN_SHARES)
1192 shares = MIN_SHARES;
1193 if (shares > tg->shares)
1194 shares = tg->shares;
1198 # else /* CONFIG_SMP */
1199 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1203 # endif /* CONFIG_SMP */
1204 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1205 unsigned long weight)
1208 /* commit outstanding execution time */
1209 if (cfs_rq->curr == se)
1210 update_curr(cfs_rq);
1211 account_entity_dequeue(cfs_rq, se);
1214 update_load_set(&se->load, weight);
1217 account_entity_enqueue(cfs_rq, se);
1220 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1222 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1224 struct task_group *tg;
1225 struct sched_entity *se;
1229 se = tg->se[cpu_of(rq_of(cfs_rq))];
1230 if (!se || throttled_hierarchy(cfs_rq))
1233 if (likely(se->load.weight == tg->shares))
1236 shares = calc_cfs_shares(cfs_rq, tg);
1238 reweight_entity(cfs_rq_of(se), se, shares);
1240 #else /* CONFIG_FAIR_GROUP_SCHED */
1241 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1244 #endif /* CONFIG_FAIR_GROUP_SCHED */
1248 * We choose a half-life close to 1 scheduling period.
1249 * Note: The tables below are dependent on this value.
1251 #define LOAD_AVG_PERIOD 32
1252 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1253 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1255 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1256 static const u32 runnable_avg_yN_inv[] = {
1257 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1258 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1259 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1260 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1261 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1262 0x85aac367, 0x82cd8698,
1266 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1267 * over-estimates when re-combining.
1269 static const u32 runnable_avg_yN_sum[] = {
1270 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1271 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1272 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1277 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1279 static __always_inline u64 decay_load(u64 val, u64 n)
1281 unsigned int local_n;
1285 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1288 /* after bounds checking we can collapse to 32-bit */
1292 * As y^PERIOD = 1/2, we can combine
1293 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1294 * With a look-up table which covers k^n (n<PERIOD)
1296 * To achieve constant time decay_load.
1298 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1299 val >>= local_n / LOAD_AVG_PERIOD;
1300 local_n %= LOAD_AVG_PERIOD;
1303 val *= runnable_avg_yN_inv[local_n];
1304 /* We don't use SRR here since we always want to round down. */
1309 * For updates fully spanning n periods, the contribution to runnable
1310 * average will be: \Sum 1024*y^n
1312 * We can compute this reasonably efficiently by combining:
1313 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1315 static u32 __compute_runnable_contrib(u64 n)
1319 if (likely(n <= LOAD_AVG_PERIOD))
1320 return runnable_avg_yN_sum[n];
1321 else if (unlikely(n >= LOAD_AVG_MAX_N))
1322 return LOAD_AVG_MAX;
1324 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1326 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1327 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1329 n -= LOAD_AVG_PERIOD;
1330 } while (n > LOAD_AVG_PERIOD);
1332 contrib = decay_load(contrib, n);
1333 return contrib + runnable_avg_yN_sum[n];
1337 * We can represent the historical contribution to runnable average as the
1338 * coefficients of a geometric series. To do this we sub-divide our runnable
1339 * history into segments of approximately 1ms (1024us); label the segment that
1340 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1342 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1344 * (now) (~1ms ago) (~2ms ago)
1346 * Let u_i denote the fraction of p_i that the entity was runnable.
1348 * We then designate the fractions u_i as our co-efficients, yielding the
1349 * following representation of historical load:
1350 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1352 * We choose y based on the with of a reasonably scheduling period, fixing:
1355 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1356 * approximately half as much as the contribution to load within the last ms
1359 * When a period "rolls over" and we have new u_0`, multiplying the previous
1360 * sum again by y is sufficient to update:
1361 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1362 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1364 static __always_inline int __update_entity_runnable_avg(u64 now,
1365 struct sched_avg *sa,
1369 u32 runnable_contrib;
1370 int delta_w, decayed = 0;
1372 delta = now - sa->last_runnable_update;
1374 * This should only happen when time goes backwards, which it
1375 * unfortunately does during sched clock init when we swap over to TSC.
1377 if ((s64)delta < 0) {
1378 sa->last_runnable_update = now;
1383 * Use 1024ns as the unit of measurement since it's a reasonable
1384 * approximation of 1us and fast to compute.
1389 sa->last_runnable_update = now;
1391 /* delta_w is the amount already accumulated against our next period */
1392 delta_w = sa->runnable_avg_period % 1024;
1393 if (delta + delta_w >= 1024) {
1394 /* period roll-over */
1398 * Now that we know we're crossing a period boundary, figure
1399 * out how much from delta we need to complete the current
1400 * period and accrue it.
1402 delta_w = 1024 - delta_w;
1404 sa->runnable_avg_sum += delta_w;
1405 sa->runnable_avg_period += delta_w;
1409 /* Figure out how many additional periods this update spans */
1410 periods = delta / 1024;
1413 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1415 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1418 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1419 runnable_contrib = __compute_runnable_contrib(periods);
1421 sa->runnable_avg_sum += runnable_contrib;
1422 sa->runnable_avg_period += runnable_contrib;
1425 /* Remainder of delta accrued against u_0` */
1427 sa->runnable_avg_sum += delta;
1428 sa->runnable_avg_period += delta;
1433 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1434 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1436 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1437 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1439 decays -= se->avg.decay_count;
1443 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1444 se->avg.decay_count = 0;
1449 #ifdef CONFIG_FAIR_GROUP_SCHED
1450 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1453 struct task_group *tg = cfs_rq->tg;
1456 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1457 tg_contrib -= cfs_rq->tg_load_contrib;
1459 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1460 atomic_long_add(tg_contrib, &tg->load_avg);
1461 cfs_rq->tg_load_contrib += tg_contrib;
1466 * Aggregate cfs_rq runnable averages into an equivalent task_group
1467 * representation for computing load contributions.
1469 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1470 struct cfs_rq *cfs_rq)
1472 struct task_group *tg = cfs_rq->tg;
1475 /* The fraction of a cpu used by this cfs_rq */
1476 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1477 sa->runnable_avg_period + 1);
1478 contrib -= cfs_rq->tg_runnable_contrib;
1480 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1481 atomic_add(contrib, &tg->runnable_avg);
1482 cfs_rq->tg_runnable_contrib += contrib;
1486 static inline void __update_group_entity_contrib(struct sched_entity *se)
1488 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1489 struct task_group *tg = cfs_rq->tg;
1494 contrib = cfs_rq->tg_load_contrib * tg->shares;
1495 se->avg.load_avg_contrib = div_u64(contrib,
1496 atomic_long_read(&tg->load_avg) + 1);
1499 * For group entities we need to compute a correction term in the case
1500 * that they are consuming <1 cpu so that we would contribute the same
1501 * load as a task of equal weight.
1503 * Explicitly co-ordinating this measurement would be expensive, but
1504 * fortunately the sum of each cpus contribution forms a usable
1505 * lower-bound on the true value.
1507 * Consider the aggregate of 2 contributions. Either they are disjoint
1508 * (and the sum represents true value) or they are disjoint and we are
1509 * understating by the aggregate of their overlap.
1511 * Extending this to N cpus, for a given overlap, the maximum amount we
1512 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1513 * cpus that overlap for this interval and w_i is the interval width.
1515 * On a small machine; the first term is well-bounded which bounds the
1516 * total error since w_i is a subset of the period. Whereas on a
1517 * larger machine, while this first term can be larger, if w_i is the
1518 * of consequential size guaranteed to see n_i*w_i quickly converge to
1519 * our upper bound of 1-cpu.
1521 runnable_avg = atomic_read(&tg->runnable_avg);
1522 if (runnable_avg < NICE_0_LOAD) {
1523 se->avg.load_avg_contrib *= runnable_avg;
1524 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1528 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1529 int force_update) {}
1530 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1531 struct cfs_rq *cfs_rq) {}
1532 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1535 static inline void __update_task_entity_contrib(struct sched_entity *se)
1539 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1540 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1541 contrib /= (se->avg.runnable_avg_period + 1);
1542 se->avg.load_avg_contrib = scale_load(contrib);
1545 /* Compute the current contribution to load_avg by se, return any delta */
1546 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1548 long old_contrib = se->avg.load_avg_contrib;
1550 if (entity_is_task(se)) {
1551 __update_task_entity_contrib(se);
1553 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1554 __update_group_entity_contrib(se);
1557 return se->avg.load_avg_contrib - old_contrib;
1560 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1563 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1564 cfs_rq->blocked_load_avg -= load_contrib;
1566 cfs_rq->blocked_load_avg = 0;
1569 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1571 /* Update a sched_entity's runnable average */
1572 static inline void update_entity_load_avg(struct sched_entity *se,
1575 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1580 * For a group entity we need to use their owned cfs_rq_clock_task() in
1581 * case they are the parent of a throttled hierarchy.
1583 if (entity_is_task(se))
1584 now = cfs_rq_clock_task(cfs_rq);
1586 now = cfs_rq_clock_task(group_cfs_rq(se));
1588 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1591 contrib_delta = __update_entity_load_avg_contrib(se);
1597 cfs_rq->runnable_load_avg += contrib_delta;
1599 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1603 * Decay the load contributed by all blocked children and account this so that
1604 * their contribution may appropriately discounted when they wake up.
1606 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1608 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1611 decays = now - cfs_rq->last_decay;
1612 if (!decays && !force_update)
1615 if (atomic_long_read(&cfs_rq->removed_load)) {
1616 unsigned long removed_load;
1617 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1618 subtract_blocked_load_contrib(cfs_rq, removed_load);
1622 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1624 atomic64_add(decays, &cfs_rq->decay_counter);
1625 cfs_rq->last_decay = now;
1628 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1631 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1633 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1634 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1637 /* Add the load generated by se into cfs_rq's child load-average */
1638 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1639 struct sched_entity *se,
1643 * We track migrations using entity decay_count <= 0, on a wake-up
1644 * migration we use a negative decay count to track the remote decays
1645 * accumulated while sleeping.
1647 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1648 * are seen by enqueue_entity_load_avg() as a migration with an already
1649 * constructed load_avg_contrib.
1651 if (unlikely(se->avg.decay_count <= 0)) {
1652 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1653 if (se->avg.decay_count) {
1655 * In a wake-up migration we have to approximate the
1656 * time sleeping. This is because we can't synchronize
1657 * clock_task between the two cpus, and it is not
1658 * guaranteed to be read-safe. Instead, we can
1659 * approximate this using our carried decays, which are
1660 * explicitly atomically readable.
1662 se->avg.last_runnable_update -= (-se->avg.decay_count)
1664 update_entity_load_avg(se, 0);
1665 /* Indicate that we're now synchronized and on-rq */
1666 se->avg.decay_count = 0;
1671 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1672 * would have made count negative); we must be careful to avoid
1673 * double-accounting blocked time after synchronizing decays.
1675 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1679 /* migrated tasks did not contribute to our blocked load */
1681 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1682 update_entity_load_avg(se, 0);
1685 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1686 /* we force update consideration on load-balancer moves */
1687 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1691 * Remove se's load from this cfs_rq child load-average, if the entity is
1692 * transitioning to a blocked state we track its projected decay using
1695 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1696 struct sched_entity *se,
1699 update_entity_load_avg(se, 1);
1700 /* we force update consideration on load-balancer moves */
1701 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1703 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1705 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1706 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1707 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1711 * Update the rq's load with the elapsed running time before entering
1712 * idle. if the last scheduled task is not a CFS task, idle_enter will
1713 * be the only way to update the runnable statistic.
1715 void idle_enter_fair(struct rq *this_rq)
1717 update_rq_runnable_avg(this_rq, 1);
1721 * Update the rq's load with the elapsed idle time before a task is
1722 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1723 * be the only way to update the runnable statistic.
1725 void idle_exit_fair(struct rq *this_rq)
1727 update_rq_runnable_avg(this_rq, 0);
1731 static inline void update_entity_load_avg(struct sched_entity *se,
1732 int update_cfs_rq) {}
1733 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1734 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1735 struct sched_entity *se,
1737 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1738 struct sched_entity *se,
1740 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1741 int force_update) {}
1744 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1746 #ifdef CONFIG_SCHEDSTATS
1747 struct task_struct *tsk = NULL;
1749 if (entity_is_task(se))
1752 if (se->statistics.sleep_start) {
1753 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1758 if (unlikely(delta > se->statistics.sleep_max))
1759 se->statistics.sleep_max = delta;
1761 se->statistics.sleep_start = 0;
1762 se->statistics.sum_sleep_runtime += delta;
1765 account_scheduler_latency(tsk, delta >> 10, 1);
1766 trace_sched_stat_sleep(tsk, delta);
1769 if (se->statistics.block_start) {
1770 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1775 if (unlikely(delta > se->statistics.block_max))
1776 se->statistics.block_max = delta;
1778 se->statistics.block_start = 0;
1779 se->statistics.sum_sleep_runtime += delta;
1782 if (tsk->in_iowait) {
1783 se->statistics.iowait_sum += delta;
1784 se->statistics.iowait_count++;
1785 trace_sched_stat_iowait(tsk, delta);
1788 trace_sched_stat_blocked(tsk, delta);
1791 * Blocking time is in units of nanosecs, so shift by
1792 * 20 to get a milliseconds-range estimation of the
1793 * amount of time that the task spent sleeping:
1795 if (unlikely(prof_on == SLEEP_PROFILING)) {
1796 profile_hits(SLEEP_PROFILING,
1797 (void *)get_wchan(tsk),
1800 account_scheduler_latency(tsk, delta >> 10, 0);
1806 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1808 #ifdef CONFIG_SCHED_DEBUG
1809 s64 d = se->vruntime - cfs_rq->min_vruntime;
1814 if (d > 3*sysctl_sched_latency)
1815 schedstat_inc(cfs_rq, nr_spread_over);
1820 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1822 u64 vruntime = cfs_rq->min_vruntime;
1825 * The 'current' period is already promised to the current tasks,
1826 * however the extra weight of the new task will slow them down a
1827 * little, place the new task so that it fits in the slot that
1828 * stays open at the end.
1830 if (initial && sched_feat(START_DEBIT))
1831 vruntime += sched_vslice(cfs_rq, se);
1833 /* sleeps up to a single latency don't count. */
1835 unsigned long thresh = sysctl_sched_latency;
1838 * Halve their sleep time's effect, to allow
1839 * for a gentler effect of sleepers:
1841 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1847 /* ensure we never gain time by being placed backwards. */
1848 se->vruntime = max_vruntime(se->vruntime, vruntime);
1851 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1854 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1857 * Update the normalized vruntime before updating min_vruntime
1858 * through calling update_curr().
1860 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1861 se->vruntime += cfs_rq->min_vruntime;
1864 * Update run-time statistics of the 'current'.
1866 update_curr(cfs_rq);
1867 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1868 account_entity_enqueue(cfs_rq, se);
1869 update_cfs_shares(cfs_rq);
1871 if (flags & ENQUEUE_WAKEUP) {
1872 place_entity(cfs_rq, se, 0);
1873 enqueue_sleeper(cfs_rq, se);
1876 update_stats_enqueue(cfs_rq, se);
1877 check_spread(cfs_rq, se);
1878 if (se != cfs_rq->curr)
1879 __enqueue_entity(cfs_rq, se);
1882 if (cfs_rq->nr_running == 1) {
1883 list_add_leaf_cfs_rq(cfs_rq);
1884 check_enqueue_throttle(cfs_rq);
1888 static void __clear_buddies_last(struct sched_entity *se)
1890 for_each_sched_entity(se) {
1891 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1892 if (cfs_rq->last == se)
1893 cfs_rq->last = NULL;
1899 static void __clear_buddies_next(struct sched_entity *se)
1901 for_each_sched_entity(se) {
1902 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1903 if (cfs_rq->next == se)
1904 cfs_rq->next = NULL;
1910 static void __clear_buddies_skip(struct sched_entity *se)
1912 for_each_sched_entity(se) {
1913 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1914 if (cfs_rq->skip == se)
1915 cfs_rq->skip = NULL;
1921 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1923 if (cfs_rq->last == se)
1924 __clear_buddies_last(se);
1926 if (cfs_rq->next == se)
1927 __clear_buddies_next(se);
1929 if (cfs_rq->skip == se)
1930 __clear_buddies_skip(se);
1933 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1936 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1939 * Update run-time statistics of the 'current'.
1941 update_curr(cfs_rq);
1942 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1944 update_stats_dequeue(cfs_rq, se);
1945 if (flags & DEQUEUE_SLEEP) {
1946 #ifdef CONFIG_SCHEDSTATS
1947 if (entity_is_task(se)) {
1948 struct task_struct *tsk = task_of(se);
1950 if (tsk->state & TASK_INTERRUPTIBLE)
1951 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1952 if (tsk->state & TASK_UNINTERRUPTIBLE)
1953 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1958 clear_buddies(cfs_rq, se);
1960 if (se != cfs_rq->curr)
1961 __dequeue_entity(cfs_rq, se);
1963 account_entity_dequeue(cfs_rq, se);
1966 * Normalize the entity after updating the min_vruntime because the
1967 * update can refer to the ->curr item and we need to reflect this
1968 * movement in our normalized position.
1970 if (!(flags & DEQUEUE_SLEEP))
1971 se->vruntime -= cfs_rq->min_vruntime;
1973 /* return excess runtime on last dequeue */
1974 return_cfs_rq_runtime(cfs_rq);
1976 update_min_vruntime(cfs_rq);
1977 update_cfs_shares(cfs_rq);
1981 * Preempt the current task with a newly woken task if needed:
1984 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1986 unsigned long ideal_runtime, delta_exec;
1987 struct sched_entity *se;
1990 ideal_runtime = sched_slice(cfs_rq, curr);
1991 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1992 if (delta_exec > ideal_runtime) {
1993 resched_task(rq_of(cfs_rq)->curr);
1995 * The current task ran long enough, ensure it doesn't get
1996 * re-elected due to buddy favours.
1998 clear_buddies(cfs_rq, curr);
2003 * Ensure that a task that missed wakeup preemption by a
2004 * narrow margin doesn't have to wait for a full slice.
2005 * This also mitigates buddy induced latencies under load.
2007 if (delta_exec < sysctl_sched_min_granularity)
2010 se = __pick_first_entity(cfs_rq);
2011 delta = curr->vruntime - se->vruntime;
2016 if (delta > ideal_runtime)
2017 resched_task(rq_of(cfs_rq)->curr);
2021 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2023 /* 'current' is not kept within the tree. */
2026 * Any task has to be enqueued before it get to execute on
2027 * a CPU. So account for the time it spent waiting on the
2030 update_stats_wait_end(cfs_rq, se);
2031 __dequeue_entity(cfs_rq, se);
2034 update_stats_curr_start(cfs_rq, se);
2036 #ifdef CONFIG_SCHEDSTATS
2038 * Track our maximum slice length, if the CPU's load is at
2039 * least twice that of our own weight (i.e. dont track it
2040 * when there are only lesser-weight tasks around):
2042 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2043 se->statistics.slice_max = max(se->statistics.slice_max,
2044 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2047 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2051 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2054 * Pick the next process, keeping these things in mind, in this order:
2055 * 1) keep things fair between processes/task groups
2056 * 2) pick the "next" process, since someone really wants that to run
2057 * 3) pick the "last" process, for cache locality
2058 * 4) do not run the "skip" process, if something else is available
2060 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2062 struct sched_entity *se = __pick_first_entity(cfs_rq);
2063 struct sched_entity *left = se;
2066 * Avoid running the skip buddy, if running something else can
2067 * be done without getting too unfair.
2069 if (cfs_rq->skip == se) {
2070 struct sched_entity *second = __pick_next_entity(se);
2071 if (second && wakeup_preempt_entity(second, left) < 1)
2076 * Prefer last buddy, try to return the CPU to a preempted task.
2078 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2082 * Someone really wants this to run. If it's not unfair, run it.
2084 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2087 clear_buddies(cfs_rq, se);
2092 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2094 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2097 * If still on the runqueue then deactivate_task()
2098 * was not called and update_curr() has to be done:
2101 update_curr(cfs_rq);
2103 /* throttle cfs_rqs exceeding runtime */
2104 check_cfs_rq_runtime(cfs_rq);
2106 check_spread(cfs_rq, prev);
2108 update_stats_wait_start(cfs_rq, prev);
2109 /* Put 'current' back into the tree. */
2110 __enqueue_entity(cfs_rq, prev);
2111 /* in !on_rq case, update occurred at dequeue */
2112 update_entity_load_avg(prev, 1);
2114 cfs_rq->curr = NULL;
2118 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2121 * Update run-time statistics of the 'current'.
2123 update_curr(cfs_rq);
2126 * Ensure that runnable average is periodically updated.
2128 update_entity_load_avg(curr, 1);
2129 update_cfs_rq_blocked_load(cfs_rq, 1);
2130 update_cfs_shares(cfs_rq);
2132 #ifdef CONFIG_SCHED_HRTICK
2134 * queued ticks are scheduled to match the slice, so don't bother
2135 * validating it and just reschedule.
2138 resched_task(rq_of(cfs_rq)->curr);
2142 * don't let the period tick interfere with the hrtick preemption
2144 if (!sched_feat(DOUBLE_TICK) &&
2145 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2149 if (cfs_rq->nr_running > 1)
2150 check_preempt_tick(cfs_rq, curr);
2154 /**************************************************
2155 * CFS bandwidth control machinery
2158 #ifdef CONFIG_CFS_BANDWIDTH
2160 #ifdef HAVE_JUMP_LABEL
2161 static struct static_key __cfs_bandwidth_used;
2163 static inline bool cfs_bandwidth_used(void)
2165 return static_key_false(&__cfs_bandwidth_used);
2168 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2170 /* only need to count groups transitioning between enabled/!enabled */
2171 if (enabled && !was_enabled)
2172 static_key_slow_inc(&__cfs_bandwidth_used);
2173 else if (!enabled && was_enabled)
2174 static_key_slow_dec(&__cfs_bandwidth_used);
2176 #else /* HAVE_JUMP_LABEL */
2177 static bool cfs_bandwidth_used(void)
2182 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2183 #endif /* HAVE_JUMP_LABEL */
2186 * default period for cfs group bandwidth.
2187 * default: 0.1s, units: nanoseconds
2189 static inline u64 default_cfs_period(void)
2191 return 100000000ULL;
2194 static inline u64 sched_cfs_bandwidth_slice(void)
2196 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2200 * Replenish runtime according to assigned quota and update expiration time.
2201 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2202 * additional synchronization around rq->lock.
2204 * requires cfs_b->lock
2206 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2210 if (cfs_b->quota == RUNTIME_INF)
2213 now = sched_clock_cpu(smp_processor_id());
2214 cfs_b->runtime = cfs_b->quota;
2215 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2218 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2220 return &tg->cfs_bandwidth;
2223 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2224 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2226 if (unlikely(cfs_rq->throttle_count))
2227 return cfs_rq->throttled_clock_task;
2229 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2232 /* returns 0 on failure to allocate runtime */
2233 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2235 struct task_group *tg = cfs_rq->tg;
2236 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2237 u64 amount = 0, min_amount, expires;
2239 /* note: this is a positive sum as runtime_remaining <= 0 */
2240 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2242 raw_spin_lock(&cfs_b->lock);
2243 if (cfs_b->quota == RUNTIME_INF)
2244 amount = min_amount;
2247 * If the bandwidth pool has become inactive, then at least one
2248 * period must have elapsed since the last consumption.
2249 * Refresh the global state and ensure bandwidth timer becomes
2252 if (!cfs_b->timer_active) {
2253 __refill_cfs_bandwidth_runtime(cfs_b);
2254 __start_cfs_bandwidth(cfs_b);
2257 if (cfs_b->runtime > 0) {
2258 amount = min(cfs_b->runtime, min_amount);
2259 cfs_b->runtime -= amount;
2263 expires = cfs_b->runtime_expires;
2264 raw_spin_unlock(&cfs_b->lock);
2266 cfs_rq->runtime_remaining += amount;
2268 * we may have advanced our local expiration to account for allowed
2269 * spread between our sched_clock and the one on which runtime was
2272 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2273 cfs_rq->runtime_expires = expires;
2275 return cfs_rq->runtime_remaining > 0;
2279 * Note: This depends on the synchronization provided by sched_clock and the
2280 * fact that rq->clock snapshots this value.
2282 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2284 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2286 /* if the deadline is ahead of our clock, nothing to do */
2287 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2290 if (cfs_rq->runtime_remaining < 0)
2294 * If the local deadline has passed we have to consider the
2295 * possibility that our sched_clock is 'fast' and the global deadline
2296 * has not truly expired.
2298 * Fortunately we can check determine whether this the case by checking
2299 * whether the global deadline has advanced.
2302 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2303 /* extend local deadline, drift is bounded above by 2 ticks */
2304 cfs_rq->runtime_expires += TICK_NSEC;
2306 /* global deadline is ahead, expiration has passed */
2307 cfs_rq->runtime_remaining = 0;
2311 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2312 unsigned long delta_exec)
2314 /* dock delta_exec before expiring quota (as it could span periods) */
2315 cfs_rq->runtime_remaining -= delta_exec;
2316 expire_cfs_rq_runtime(cfs_rq);
2318 if (likely(cfs_rq->runtime_remaining > 0))
2322 * if we're unable to extend our runtime we resched so that the active
2323 * hierarchy can be throttled
2325 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2326 resched_task(rq_of(cfs_rq)->curr);
2329 static __always_inline
2330 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2332 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2335 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2338 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2340 return cfs_bandwidth_used() && cfs_rq->throttled;
2343 /* check whether cfs_rq, or any parent, is throttled */
2344 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2346 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2350 * Ensure that neither of the group entities corresponding to src_cpu or
2351 * dest_cpu are members of a throttled hierarchy when performing group
2352 * load-balance operations.
2354 static inline int throttled_lb_pair(struct task_group *tg,
2355 int src_cpu, int dest_cpu)
2357 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2359 src_cfs_rq = tg->cfs_rq[src_cpu];
2360 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2362 return throttled_hierarchy(src_cfs_rq) ||
2363 throttled_hierarchy(dest_cfs_rq);
2366 /* updated child weight may affect parent so we have to do this bottom up */
2367 static int tg_unthrottle_up(struct task_group *tg, void *data)
2369 struct rq *rq = data;
2370 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2372 cfs_rq->throttle_count--;
2374 if (!cfs_rq->throttle_count) {
2375 /* adjust cfs_rq_clock_task() */
2376 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2377 cfs_rq->throttled_clock_task;
2384 static int tg_throttle_down(struct task_group *tg, void *data)
2386 struct rq *rq = data;
2387 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2389 /* group is entering throttled state, stop time */
2390 if (!cfs_rq->throttle_count)
2391 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2392 cfs_rq->throttle_count++;
2397 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2399 struct rq *rq = rq_of(cfs_rq);
2400 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2401 struct sched_entity *se;
2402 long task_delta, dequeue = 1;
2404 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2406 /* freeze hierarchy runnable averages while throttled */
2408 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2411 task_delta = cfs_rq->h_nr_running;
2412 for_each_sched_entity(se) {
2413 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2414 /* throttled entity or throttle-on-deactivate */
2419 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2420 qcfs_rq->h_nr_running -= task_delta;
2422 if (qcfs_rq->load.weight)
2427 rq->nr_running -= task_delta;
2429 cfs_rq->throttled = 1;
2430 cfs_rq->throttled_clock = rq_clock(rq);
2431 raw_spin_lock(&cfs_b->lock);
2432 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2433 raw_spin_unlock(&cfs_b->lock);
2436 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2438 struct rq *rq = rq_of(cfs_rq);
2439 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2440 struct sched_entity *se;
2444 se = cfs_rq->tg->se[cpu_of(rq)];
2446 cfs_rq->throttled = 0;
2448 update_rq_clock(rq);
2450 raw_spin_lock(&cfs_b->lock);
2451 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2452 list_del_rcu(&cfs_rq->throttled_list);
2453 raw_spin_unlock(&cfs_b->lock);
2455 /* update hierarchical throttle state */
2456 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2458 if (!cfs_rq->load.weight)
2461 task_delta = cfs_rq->h_nr_running;
2462 for_each_sched_entity(se) {
2466 cfs_rq = cfs_rq_of(se);
2468 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2469 cfs_rq->h_nr_running += task_delta;
2471 if (cfs_rq_throttled(cfs_rq))
2476 rq->nr_running += task_delta;
2478 /* determine whether we need to wake up potentially idle cpu */
2479 if (rq->curr == rq->idle && rq->cfs.nr_running)
2480 resched_task(rq->curr);
2483 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2484 u64 remaining, u64 expires)
2486 struct cfs_rq *cfs_rq;
2487 u64 runtime = remaining;
2490 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2492 struct rq *rq = rq_of(cfs_rq);
2494 raw_spin_lock(&rq->lock);
2495 if (!cfs_rq_throttled(cfs_rq))
2498 runtime = -cfs_rq->runtime_remaining + 1;
2499 if (runtime > remaining)
2500 runtime = remaining;
2501 remaining -= runtime;
2503 cfs_rq->runtime_remaining += runtime;
2504 cfs_rq->runtime_expires = expires;
2506 /* we check whether we're throttled above */
2507 if (cfs_rq->runtime_remaining > 0)
2508 unthrottle_cfs_rq(cfs_rq);
2511 raw_spin_unlock(&rq->lock);
2522 * Responsible for refilling a task_group's bandwidth and unthrottling its
2523 * cfs_rqs as appropriate. If there has been no activity within the last
2524 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2525 * used to track this state.
2527 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2529 u64 runtime, runtime_expires;
2530 int idle = 1, throttled;
2532 raw_spin_lock(&cfs_b->lock);
2533 /* no need to continue the timer with no bandwidth constraint */
2534 if (cfs_b->quota == RUNTIME_INF)
2537 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2538 /* idle depends on !throttled (for the case of a large deficit) */
2539 idle = cfs_b->idle && !throttled;
2540 cfs_b->nr_periods += overrun;
2542 /* if we're going inactive then everything else can be deferred */
2546 __refill_cfs_bandwidth_runtime(cfs_b);
2549 /* mark as potentially idle for the upcoming period */
2554 /* account preceding periods in which throttling occurred */
2555 cfs_b->nr_throttled += overrun;
2558 * There are throttled entities so we must first use the new bandwidth
2559 * to unthrottle them before making it generally available. This
2560 * ensures that all existing debts will be paid before a new cfs_rq is
2563 runtime = cfs_b->runtime;
2564 runtime_expires = cfs_b->runtime_expires;
2568 * This check is repeated as we are holding onto the new bandwidth
2569 * while we unthrottle. This can potentially race with an unthrottled
2570 * group trying to acquire new bandwidth from the global pool.
2572 while (throttled && runtime > 0) {
2573 raw_spin_unlock(&cfs_b->lock);
2574 /* we can't nest cfs_b->lock while distributing bandwidth */
2575 runtime = distribute_cfs_runtime(cfs_b, runtime,
2577 raw_spin_lock(&cfs_b->lock);
2579 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2582 /* return (any) remaining runtime */
2583 cfs_b->runtime = runtime;
2585 * While we are ensured activity in the period following an
2586 * unthrottle, this also covers the case in which the new bandwidth is
2587 * insufficient to cover the existing bandwidth deficit. (Forcing the
2588 * timer to remain active while there are any throttled entities.)
2593 cfs_b->timer_active = 0;
2594 raw_spin_unlock(&cfs_b->lock);
2599 /* a cfs_rq won't donate quota below this amount */
2600 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2601 /* minimum remaining period time to redistribute slack quota */
2602 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2603 /* how long we wait to gather additional slack before distributing */
2604 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2606 /* are we near the end of the current quota period? */
2607 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2609 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2612 /* if the call-back is running a quota refresh is already occurring */
2613 if (hrtimer_callback_running(refresh_timer))
2616 /* is a quota refresh about to occur? */
2617 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2618 if (remaining < min_expire)
2624 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2626 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2628 /* if there's a quota refresh soon don't bother with slack */
2629 if (runtime_refresh_within(cfs_b, min_left))
2632 start_bandwidth_timer(&cfs_b->slack_timer,
2633 ns_to_ktime(cfs_bandwidth_slack_period));
2636 /* we know any runtime found here is valid as update_curr() precedes return */
2637 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2639 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2640 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2642 if (slack_runtime <= 0)
2645 raw_spin_lock(&cfs_b->lock);
2646 if (cfs_b->quota != RUNTIME_INF &&
2647 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2648 cfs_b->runtime += slack_runtime;
2650 /* we are under rq->lock, defer unthrottling using a timer */
2651 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2652 !list_empty(&cfs_b->throttled_cfs_rq))
2653 start_cfs_slack_bandwidth(cfs_b);
2655 raw_spin_unlock(&cfs_b->lock);
2657 /* even if it's not valid for return we don't want to try again */
2658 cfs_rq->runtime_remaining -= slack_runtime;
2661 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2663 if (!cfs_bandwidth_used())
2666 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2669 __return_cfs_rq_runtime(cfs_rq);
2673 * This is done with a timer (instead of inline with bandwidth return) since
2674 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2676 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2678 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2681 /* confirm we're still not at a refresh boundary */
2682 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2685 raw_spin_lock(&cfs_b->lock);
2686 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2687 runtime = cfs_b->runtime;
2690 expires = cfs_b->runtime_expires;
2691 raw_spin_unlock(&cfs_b->lock);
2696 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2698 raw_spin_lock(&cfs_b->lock);
2699 if (expires == cfs_b->runtime_expires)
2700 cfs_b->runtime = runtime;
2701 raw_spin_unlock(&cfs_b->lock);
2705 * When a group wakes up we want to make sure that its quota is not already
2706 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2707 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2709 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2711 if (!cfs_bandwidth_used())
2714 /* an active group must be handled by the update_curr()->put() path */
2715 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2718 /* ensure the group is not already throttled */
2719 if (cfs_rq_throttled(cfs_rq))
2722 /* update runtime allocation */
2723 account_cfs_rq_runtime(cfs_rq, 0);
2724 if (cfs_rq->runtime_remaining <= 0)
2725 throttle_cfs_rq(cfs_rq);
2728 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2729 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2731 if (!cfs_bandwidth_used())
2734 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2738 * it's possible for a throttled entity to be forced into a running
2739 * state (e.g. set_curr_task), in this case we're finished.
2741 if (cfs_rq_throttled(cfs_rq))
2744 throttle_cfs_rq(cfs_rq);
2747 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2749 struct cfs_bandwidth *cfs_b =
2750 container_of(timer, struct cfs_bandwidth, slack_timer);
2751 do_sched_cfs_slack_timer(cfs_b);
2753 return HRTIMER_NORESTART;
2756 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2758 struct cfs_bandwidth *cfs_b =
2759 container_of(timer, struct cfs_bandwidth, period_timer);
2765 now = hrtimer_cb_get_time(timer);
2766 overrun = hrtimer_forward(timer, now, cfs_b->period);
2771 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2774 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2777 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2779 raw_spin_lock_init(&cfs_b->lock);
2781 cfs_b->quota = RUNTIME_INF;
2782 cfs_b->period = ns_to_ktime(default_cfs_period());
2784 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2785 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2786 cfs_b->period_timer.function = sched_cfs_period_timer;
2787 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2788 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2791 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2793 cfs_rq->runtime_enabled = 0;
2794 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2797 /* requires cfs_b->lock, may release to reprogram timer */
2798 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2801 * The timer may be active because we're trying to set a new bandwidth
2802 * period or because we're racing with the tear-down path
2803 * (timer_active==0 becomes visible before the hrtimer call-back
2804 * terminates). In either case we ensure that it's re-programmed
2806 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2807 raw_spin_unlock(&cfs_b->lock);
2808 /* ensure cfs_b->lock is available while we wait */
2809 hrtimer_cancel(&cfs_b->period_timer);
2811 raw_spin_lock(&cfs_b->lock);
2812 /* if someone else restarted the timer then we're done */
2813 if (cfs_b->timer_active)
2817 cfs_b->timer_active = 1;
2818 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2821 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2823 hrtimer_cancel(&cfs_b->period_timer);
2824 hrtimer_cancel(&cfs_b->slack_timer);
2827 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2829 struct cfs_rq *cfs_rq;
2831 for_each_leaf_cfs_rq(rq, cfs_rq) {
2832 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2834 if (!cfs_rq->runtime_enabled)
2838 * clock_task is not advancing so we just need to make sure
2839 * there's some valid quota amount
2841 cfs_rq->runtime_remaining = cfs_b->quota;
2842 if (cfs_rq_throttled(cfs_rq))
2843 unthrottle_cfs_rq(cfs_rq);
2847 #else /* CONFIG_CFS_BANDWIDTH */
2848 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2850 return rq_clock_task(rq_of(cfs_rq));
2853 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2854 unsigned long delta_exec) {}
2855 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2856 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2857 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2859 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2864 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2869 static inline int throttled_lb_pair(struct task_group *tg,
2870 int src_cpu, int dest_cpu)
2875 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2877 #ifdef CONFIG_FAIR_GROUP_SCHED
2878 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2881 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2885 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2886 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2888 #endif /* CONFIG_CFS_BANDWIDTH */
2890 /**************************************************
2891 * CFS operations on tasks:
2894 #ifdef CONFIG_SCHED_HRTICK
2895 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2897 struct sched_entity *se = &p->se;
2898 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2900 WARN_ON(task_rq(p) != rq);
2902 if (cfs_rq->nr_running > 1) {
2903 u64 slice = sched_slice(cfs_rq, se);
2904 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2905 s64 delta = slice - ran;
2914 * Don't schedule slices shorter than 10000ns, that just
2915 * doesn't make sense. Rely on vruntime for fairness.
2918 delta = max_t(s64, 10000LL, delta);
2920 hrtick_start(rq, delta);
2925 * called from enqueue/dequeue and updates the hrtick when the
2926 * current task is from our class and nr_running is low enough
2929 static void hrtick_update(struct rq *rq)
2931 struct task_struct *curr = rq->curr;
2933 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2936 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2937 hrtick_start_fair(rq, curr);
2939 #else /* !CONFIG_SCHED_HRTICK */
2941 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2945 static inline void hrtick_update(struct rq *rq)
2951 * The enqueue_task method is called before nr_running is
2952 * increased. Here we update the fair scheduling stats and
2953 * then put the task into the rbtree:
2956 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2958 struct cfs_rq *cfs_rq;
2959 struct sched_entity *se = &p->se;
2961 for_each_sched_entity(se) {
2964 cfs_rq = cfs_rq_of(se);
2965 enqueue_entity(cfs_rq, se, flags);
2968 * end evaluation on encountering a throttled cfs_rq
2970 * note: in the case of encountering a throttled cfs_rq we will
2971 * post the final h_nr_running increment below.
2973 if (cfs_rq_throttled(cfs_rq))
2975 cfs_rq->h_nr_running++;
2977 flags = ENQUEUE_WAKEUP;
2980 for_each_sched_entity(se) {
2981 cfs_rq = cfs_rq_of(se);
2982 cfs_rq->h_nr_running++;
2984 if (cfs_rq_throttled(cfs_rq))
2987 update_cfs_shares(cfs_rq);
2988 update_entity_load_avg(se, 1);
2992 update_rq_runnable_avg(rq, rq->nr_running);
2998 static void set_next_buddy(struct sched_entity *se);
3001 * The dequeue_task method is called before nr_running is
3002 * decreased. We remove the task from the rbtree and
3003 * update the fair scheduling stats:
3005 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3007 struct cfs_rq *cfs_rq;
3008 struct sched_entity *se = &p->se;
3009 int task_sleep = flags & DEQUEUE_SLEEP;
3011 for_each_sched_entity(se) {
3012 cfs_rq = cfs_rq_of(se);
3013 dequeue_entity(cfs_rq, se, flags);
3016 * end evaluation on encountering a throttled cfs_rq
3018 * note: in the case of encountering a throttled cfs_rq we will
3019 * post the final h_nr_running decrement below.
3021 if (cfs_rq_throttled(cfs_rq))
3023 cfs_rq->h_nr_running--;
3025 /* Don't dequeue parent if it has other entities besides us */
3026 if (cfs_rq->load.weight) {
3028 * Bias pick_next to pick a task from this cfs_rq, as
3029 * p is sleeping when it is within its sched_slice.
3031 if (task_sleep && parent_entity(se))
3032 set_next_buddy(parent_entity(se));
3034 /* avoid re-evaluating load for this entity */
3035 se = parent_entity(se);
3038 flags |= DEQUEUE_SLEEP;
3041 for_each_sched_entity(se) {
3042 cfs_rq = cfs_rq_of(se);
3043 cfs_rq->h_nr_running--;
3045 if (cfs_rq_throttled(cfs_rq))
3048 update_cfs_shares(cfs_rq);
3049 update_entity_load_avg(se, 1);
3054 update_rq_runnable_avg(rq, 1);
3060 /* Used instead of source_load when we know the type == 0 */
3061 static unsigned long weighted_cpuload(const int cpu)
3063 return cpu_rq(cpu)->cfs.runnable_load_avg;
3067 * Return a low guess at the load of a migration-source cpu weighted
3068 * according to the scheduling class and "nice" value.
3070 * We want to under-estimate the load of migration sources, to
3071 * balance conservatively.
3073 static unsigned long source_load(int cpu, int type)
3075 struct rq *rq = cpu_rq(cpu);
3076 unsigned long total = weighted_cpuload(cpu);
3078 if (type == 0 || !sched_feat(LB_BIAS))
3081 return min(rq->cpu_load[type-1], total);
3085 * Return a high guess at the load of a migration-target cpu weighted
3086 * according to the scheduling class and "nice" value.
3088 static unsigned long target_load(int cpu, int type)
3090 struct rq *rq = cpu_rq(cpu);
3091 unsigned long total = weighted_cpuload(cpu);
3093 if (type == 0 || !sched_feat(LB_BIAS))
3096 return max(rq->cpu_load[type-1], total);
3099 static unsigned long power_of(int cpu)
3101 return cpu_rq(cpu)->cpu_power;
3104 static unsigned long cpu_avg_load_per_task(int cpu)
3106 struct rq *rq = cpu_rq(cpu);
3107 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3108 unsigned long load_avg = rq->cfs.runnable_load_avg;
3111 return load_avg / nr_running;
3116 static void record_wakee(struct task_struct *p)
3119 * Rough decay (wiping) for cost saving, don't worry
3120 * about the boundary, really active task won't care
3123 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3124 current->wakee_flips = 0;
3125 current->wakee_flip_decay_ts = jiffies;
3128 if (current->last_wakee != p) {
3129 current->last_wakee = p;
3130 current->wakee_flips++;
3134 static void task_waking_fair(struct task_struct *p)
3136 struct sched_entity *se = &p->se;
3137 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3140 #ifndef CONFIG_64BIT
3141 u64 min_vruntime_copy;
3144 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3146 min_vruntime = cfs_rq->min_vruntime;
3147 } while (min_vruntime != min_vruntime_copy);
3149 min_vruntime = cfs_rq->min_vruntime;
3152 se->vruntime -= min_vruntime;
3156 #ifdef CONFIG_FAIR_GROUP_SCHED
3158 * effective_load() calculates the load change as seen from the root_task_group
3160 * Adding load to a group doesn't make a group heavier, but can cause movement
3161 * of group shares between cpus. Assuming the shares were perfectly aligned one
3162 * can calculate the shift in shares.
3164 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3165 * on this @cpu and results in a total addition (subtraction) of @wg to the
3166 * total group weight.
3168 * Given a runqueue weight distribution (rw_i) we can compute a shares
3169 * distribution (s_i) using:
3171 * s_i = rw_i / \Sum rw_j (1)
3173 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3174 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3175 * shares distribution (s_i):
3177 * rw_i = { 2, 4, 1, 0 }
3178 * s_i = { 2/7, 4/7, 1/7, 0 }
3180 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3181 * task used to run on and the CPU the waker is running on), we need to
3182 * compute the effect of waking a task on either CPU and, in case of a sync
3183 * wakeup, compute the effect of the current task going to sleep.
3185 * So for a change of @wl to the local @cpu with an overall group weight change
3186 * of @wl we can compute the new shares distribution (s'_i) using:
3188 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3190 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3191 * differences in waking a task to CPU 0. The additional task changes the
3192 * weight and shares distributions like:
3194 * rw'_i = { 3, 4, 1, 0 }
3195 * s'_i = { 3/8, 4/8, 1/8, 0 }
3197 * We can then compute the difference in effective weight by using:
3199 * dw_i = S * (s'_i - s_i) (3)
3201 * Where 'S' is the group weight as seen by its parent.
3203 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3204 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3205 * 4/7) times the weight of the group.
3207 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3209 struct sched_entity *se = tg->se[cpu];
3211 if (!tg->parent) /* the trivial, non-cgroup case */
3214 for_each_sched_entity(se) {
3220 * W = @wg + \Sum rw_j
3222 W = wg + calc_tg_weight(tg, se->my_q);
3227 w = se->my_q->load.weight + wl;
3230 * wl = S * s'_i; see (2)
3233 wl = (w * tg->shares) / W;
3238 * Per the above, wl is the new se->load.weight value; since
3239 * those are clipped to [MIN_SHARES, ...) do so now. See
3240 * calc_cfs_shares().
3242 if (wl < MIN_SHARES)
3246 * wl = dw_i = S * (s'_i - s_i); see (3)
3248 wl -= se->load.weight;
3251 * Recursively apply this logic to all parent groups to compute
3252 * the final effective load change on the root group. Since
3253 * only the @tg group gets extra weight, all parent groups can
3254 * only redistribute existing shares. @wl is the shift in shares
3255 * resulting from this level per the above.
3264 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3265 unsigned long wl, unsigned long wg)
3272 static int wake_wide(struct task_struct *p)
3274 int factor = this_cpu_read(sd_llc_size);
3277 * Yeah, it's the switching-frequency, could means many wakee or
3278 * rapidly switch, use factor here will just help to automatically
3279 * adjust the loose-degree, so bigger node will lead to more pull.
3281 if (p->wakee_flips > factor) {
3283 * wakee is somewhat hot, it needs certain amount of cpu
3284 * resource, so if waker is far more hot, prefer to leave
3287 if (current->wakee_flips > (factor * p->wakee_flips))
3294 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3296 s64 this_load, load;
3297 int idx, this_cpu, prev_cpu;
3298 unsigned long tl_per_task;
3299 struct task_group *tg;
3300 unsigned long weight;
3304 * If we wake multiple tasks be careful to not bounce
3305 * ourselves around too much.
3311 this_cpu = smp_processor_id();
3312 prev_cpu = task_cpu(p);
3313 load = source_load(prev_cpu, idx);
3314 this_load = target_load(this_cpu, idx);
3317 * If sync wakeup then subtract the (maximum possible)
3318 * effect of the currently running task from the load
3319 * of the current CPU:
3322 tg = task_group(current);
3323 weight = current->se.load.weight;
3325 this_load += effective_load(tg, this_cpu, -weight, -weight);
3326 load += effective_load(tg, prev_cpu, 0, -weight);
3330 weight = p->se.load.weight;
3333 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3334 * due to the sync cause above having dropped this_load to 0, we'll
3335 * always have an imbalance, but there's really nothing you can do
3336 * about that, so that's good too.
3338 * Otherwise check if either cpus are near enough in load to allow this
3339 * task to be woken on this_cpu.
3341 if (this_load > 0) {
3342 s64 this_eff_load, prev_eff_load;
3344 this_eff_load = 100;
3345 this_eff_load *= power_of(prev_cpu);
3346 this_eff_load *= this_load +
3347 effective_load(tg, this_cpu, weight, weight);
3349 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3350 prev_eff_load *= power_of(this_cpu);
3351 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3353 balanced = this_eff_load <= prev_eff_load;
3358 * If the currently running task will sleep within
3359 * a reasonable amount of time then attract this newly
3362 if (sync && balanced)
3365 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3366 tl_per_task = cpu_avg_load_per_task(this_cpu);
3369 (this_load <= load &&
3370 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3372 * This domain has SD_WAKE_AFFINE and
3373 * p is cache cold in this domain, and
3374 * there is no bad imbalance.
3376 schedstat_inc(sd, ttwu_move_affine);
3377 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3385 * find_idlest_group finds and returns the least busy CPU group within the
3388 static struct sched_group *
3389 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3390 int this_cpu, int load_idx)
3392 struct sched_group *idlest = NULL, *group = sd->groups;
3393 unsigned long min_load = ULONG_MAX, this_load = 0;
3394 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3397 unsigned long load, avg_load;
3401 /* Skip over this group if it has no CPUs allowed */
3402 if (!cpumask_intersects(sched_group_cpus(group),
3403 tsk_cpus_allowed(p)))
3406 local_group = cpumask_test_cpu(this_cpu,
3407 sched_group_cpus(group));
3409 /* Tally up the load of all CPUs in the group */
3412 for_each_cpu(i, sched_group_cpus(group)) {
3413 /* Bias balancing toward cpus of our domain */
3415 load = source_load(i, load_idx);
3417 load = target_load(i, load_idx);
3422 /* Adjust by relative CPU power of the group */
3423 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3426 this_load = avg_load;
3427 } else if (avg_load < min_load) {
3428 min_load = avg_load;
3431 } while (group = group->next, group != sd->groups);
3433 if (!idlest || 100*this_load < imbalance*min_load)
3439 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3442 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3444 unsigned long load, min_load = ULONG_MAX;
3448 /* Traverse only the allowed CPUs */
3449 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3450 load = weighted_cpuload(i);
3452 if (load < min_load || (load == min_load && i == this_cpu)) {
3462 * Try and locate an idle CPU in the sched_domain.
3464 static int select_idle_sibling(struct task_struct *p, int target)
3466 struct sched_domain *sd;
3467 struct sched_group *sg;
3468 int i = task_cpu(p);
3470 if (idle_cpu(target))
3474 * If the prevous cpu is cache affine and idle, don't be stupid.
3476 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3480 * Otherwise, iterate the domains and find an elegible idle cpu.
3482 sd = rcu_dereference(per_cpu(sd_llc, target));
3483 for_each_lower_domain(sd) {
3486 if (!cpumask_intersects(sched_group_cpus(sg),
3487 tsk_cpus_allowed(p)))
3490 for_each_cpu(i, sched_group_cpus(sg)) {
3491 if (i == target || !idle_cpu(i))
3495 target = cpumask_first_and(sched_group_cpus(sg),
3496 tsk_cpus_allowed(p));
3500 } while (sg != sd->groups);
3507 * sched_balance_self: balance the current task (running on cpu) in domains
3508 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3511 * Balance, ie. select the least loaded group.
3513 * Returns the target CPU number, or the same CPU if no balancing is needed.
3515 * preempt must be disabled.
3518 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3520 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3521 int cpu = smp_processor_id();
3522 int prev_cpu = task_cpu(p);
3524 int want_affine = 0;
3525 int sync = wake_flags & WF_SYNC;
3527 if (p->nr_cpus_allowed == 1)
3530 if (sd_flag & SD_BALANCE_WAKE) {
3531 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3537 for_each_domain(cpu, tmp) {
3538 if (!(tmp->flags & SD_LOAD_BALANCE))
3542 * If both cpu and prev_cpu are part of this domain,
3543 * cpu is a valid SD_WAKE_AFFINE target.
3545 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3546 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3551 if (tmp->flags & sd_flag)
3556 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3559 new_cpu = select_idle_sibling(p, prev_cpu);
3564 int load_idx = sd->forkexec_idx;
3565 struct sched_group *group;
3568 if (!(sd->flags & sd_flag)) {
3573 if (sd_flag & SD_BALANCE_WAKE)
3574 load_idx = sd->wake_idx;
3576 group = find_idlest_group(sd, p, cpu, load_idx);
3582 new_cpu = find_idlest_cpu(group, p, cpu);
3583 if (new_cpu == -1 || new_cpu == cpu) {
3584 /* Now try balancing at a lower domain level of cpu */
3589 /* Now try balancing at a lower domain level of new_cpu */
3591 weight = sd->span_weight;
3593 for_each_domain(cpu, tmp) {
3594 if (weight <= tmp->span_weight)
3596 if (tmp->flags & sd_flag)
3599 /* while loop will break here if sd == NULL */
3608 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3609 * cfs_rq_of(p) references at time of call are still valid and identify the
3610 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3611 * other assumptions, including the state of rq->lock, should be made.
3614 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3616 struct sched_entity *se = &p->se;
3617 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3620 * Load tracking: accumulate removed load so that it can be processed
3621 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3622 * to blocked load iff they have a positive decay-count. It can never
3623 * be negative here since on-rq tasks have decay-count == 0.
3625 if (se->avg.decay_count) {
3626 se->avg.decay_count = -__synchronize_entity_decay(se);
3627 atomic_long_add(se->avg.load_avg_contrib,
3628 &cfs_rq->removed_load);
3631 #endif /* CONFIG_SMP */
3633 static unsigned long
3634 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3636 unsigned long gran = sysctl_sched_wakeup_granularity;
3639 * Since its curr running now, convert the gran from real-time
3640 * to virtual-time in his units.
3642 * By using 'se' instead of 'curr' we penalize light tasks, so
3643 * they get preempted easier. That is, if 'se' < 'curr' then
3644 * the resulting gran will be larger, therefore penalizing the
3645 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3646 * be smaller, again penalizing the lighter task.
3648 * This is especially important for buddies when the leftmost
3649 * task is higher priority than the buddy.
3651 return calc_delta_fair(gran, se);
3655 * Should 'se' preempt 'curr'.
3669 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3671 s64 gran, vdiff = curr->vruntime - se->vruntime;
3676 gran = wakeup_gran(curr, se);
3683 static void set_last_buddy(struct sched_entity *se)
3685 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3688 for_each_sched_entity(se)
3689 cfs_rq_of(se)->last = se;
3692 static void set_next_buddy(struct sched_entity *se)
3694 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3697 for_each_sched_entity(se)
3698 cfs_rq_of(se)->next = se;
3701 static void set_skip_buddy(struct sched_entity *se)
3703 for_each_sched_entity(se)
3704 cfs_rq_of(se)->skip = se;
3708 * Preempt the current task with a newly woken task if needed:
3710 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3712 struct task_struct *curr = rq->curr;
3713 struct sched_entity *se = &curr->se, *pse = &p->se;
3714 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3715 int scale = cfs_rq->nr_running >= sched_nr_latency;
3716 int next_buddy_marked = 0;
3718 if (unlikely(se == pse))
3722 * This is possible from callers such as move_task(), in which we
3723 * unconditionally check_prempt_curr() after an enqueue (which may have
3724 * lead to a throttle). This both saves work and prevents false
3725 * next-buddy nomination below.
3727 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3730 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3731 set_next_buddy(pse);
3732 next_buddy_marked = 1;
3736 * We can come here with TIF_NEED_RESCHED already set from new task
3739 * Note: this also catches the edge-case of curr being in a throttled
3740 * group (e.g. via set_curr_task), since update_curr() (in the
3741 * enqueue of curr) will have resulted in resched being set. This
3742 * prevents us from potentially nominating it as a false LAST_BUDDY
3745 if (test_tsk_need_resched(curr))
3748 /* Idle tasks are by definition preempted by non-idle tasks. */
3749 if (unlikely(curr->policy == SCHED_IDLE) &&
3750 likely(p->policy != SCHED_IDLE))
3754 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3755 * is driven by the tick):
3757 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3760 find_matching_se(&se, &pse);
3761 update_curr(cfs_rq_of(se));
3763 if (wakeup_preempt_entity(se, pse) == 1) {
3765 * Bias pick_next to pick the sched entity that is
3766 * triggering this preemption.
3768 if (!next_buddy_marked)
3769 set_next_buddy(pse);
3778 * Only set the backward buddy when the current task is still
3779 * on the rq. This can happen when a wakeup gets interleaved
3780 * with schedule on the ->pre_schedule() or idle_balance()
3781 * point, either of which can * drop the rq lock.
3783 * Also, during early boot the idle thread is in the fair class,
3784 * for obvious reasons its a bad idea to schedule back to it.
3786 if (unlikely(!se->on_rq || curr == rq->idle))
3789 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3793 static struct task_struct *pick_next_task_fair(struct rq *rq)
3795 struct task_struct *p;
3796 struct cfs_rq *cfs_rq = &rq->cfs;
3797 struct sched_entity *se;
3799 if (!cfs_rq->nr_running)
3803 se = pick_next_entity(cfs_rq);
3804 set_next_entity(cfs_rq, se);
3805 cfs_rq = group_cfs_rq(se);
3809 if (hrtick_enabled(rq))
3810 hrtick_start_fair(rq, p);
3816 * Account for a descheduled task:
3818 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3820 struct sched_entity *se = &prev->se;
3821 struct cfs_rq *cfs_rq;
3823 for_each_sched_entity(se) {
3824 cfs_rq = cfs_rq_of(se);
3825 put_prev_entity(cfs_rq, se);
3830 * sched_yield() is very simple
3832 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3834 static void yield_task_fair(struct rq *rq)
3836 struct task_struct *curr = rq->curr;
3837 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3838 struct sched_entity *se = &curr->se;
3841 * Are we the only task in the tree?
3843 if (unlikely(rq->nr_running == 1))
3846 clear_buddies(cfs_rq, se);
3848 if (curr->policy != SCHED_BATCH) {
3849 update_rq_clock(rq);
3851 * Update run-time statistics of the 'current'.
3853 update_curr(cfs_rq);
3855 * Tell update_rq_clock() that we've just updated,
3856 * so we don't do microscopic update in schedule()
3857 * and double the fastpath cost.
3859 rq->skip_clock_update = 1;
3865 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3867 struct sched_entity *se = &p->se;
3869 /* throttled hierarchies are not runnable */
3870 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3873 /* Tell the scheduler that we'd really like pse to run next. */
3876 yield_task_fair(rq);
3882 /**************************************************
3883 * Fair scheduling class load-balancing methods.
3887 * The purpose of load-balancing is to achieve the same basic fairness the
3888 * per-cpu scheduler provides, namely provide a proportional amount of compute
3889 * time to each task. This is expressed in the following equation:
3891 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3893 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3894 * W_i,0 is defined as:
3896 * W_i,0 = \Sum_j w_i,j (2)
3898 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3899 * is derived from the nice value as per prio_to_weight[].
3901 * The weight average is an exponential decay average of the instantaneous
3904 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3906 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3907 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3908 * can also include other factors [XXX].
3910 * To achieve this balance we define a measure of imbalance which follows
3911 * directly from (1):
3913 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3915 * We them move tasks around to minimize the imbalance. In the continuous
3916 * function space it is obvious this converges, in the discrete case we get
3917 * a few fun cases generally called infeasible weight scenarios.
3920 * - infeasible weights;
3921 * - local vs global optima in the discrete case. ]
3926 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3927 * for all i,j solution, we create a tree of cpus that follows the hardware
3928 * topology where each level pairs two lower groups (or better). This results
3929 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3930 * tree to only the first of the previous level and we decrease the frequency
3931 * of load-balance at each level inv. proportional to the number of cpus in
3937 * \Sum { --- * --- * 2^i } = O(n) (5)
3939 * `- size of each group
3940 * | | `- number of cpus doing load-balance
3942 * `- sum over all levels
3944 * Coupled with a limit on how many tasks we can migrate every balance pass,
3945 * this makes (5) the runtime complexity of the balancer.
3947 * An important property here is that each CPU is still (indirectly) connected
3948 * to every other cpu in at most O(log n) steps:
3950 * The adjacency matrix of the resulting graph is given by:
3953 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3956 * And you'll find that:
3958 * A^(log_2 n)_i,j != 0 for all i,j (7)
3960 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3961 * The task movement gives a factor of O(m), giving a convergence complexity
3964 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3969 * In order to avoid CPUs going idle while there's still work to do, new idle
3970 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3971 * tree itself instead of relying on other CPUs to bring it work.
3973 * This adds some complexity to both (5) and (8) but it reduces the total idle
3981 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3984 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3989 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3991 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3993 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3996 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3997 * rewrite all of this once again.]
4000 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4002 #define LBF_ALL_PINNED 0x01
4003 #define LBF_NEED_BREAK 0x02
4004 #define LBF_DST_PINNED 0x04
4005 #define LBF_SOME_PINNED 0x08
4008 struct sched_domain *sd;
4016 struct cpumask *dst_grpmask;
4018 enum cpu_idle_type idle;
4020 /* The set of CPUs under consideration for load-balancing */
4021 struct cpumask *cpus;
4026 unsigned int loop_break;
4027 unsigned int loop_max;
4031 * move_task - move a task from one runqueue to another runqueue.
4032 * Both runqueues must be locked.
4034 static void move_task(struct task_struct *p, struct lb_env *env)
4036 deactivate_task(env->src_rq, p, 0);
4037 set_task_cpu(p, env->dst_cpu);
4038 activate_task(env->dst_rq, p, 0);
4039 check_preempt_curr(env->dst_rq, p, 0);
4043 * Is this task likely cache-hot:
4046 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4050 if (p->sched_class != &fair_sched_class)
4053 if (unlikely(p->policy == SCHED_IDLE))
4057 * Buddy candidates are cache hot:
4059 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4060 (&p->se == cfs_rq_of(&p->se)->next ||
4061 &p->se == cfs_rq_of(&p->se)->last))
4064 if (sysctl_sched_migration_cost == -1)
4066 if (sysctl_sched_migration_cost == 0)
4069 delta = now - p->se.exec_start;
4071 return delta < (s64)sysctl_sched_migration_cost;
4075 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4078 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4080 int tsk_cache_hot = 0;
4082 * We do not migrate tasks that are:
4083 * 1) throttled_lb_pair, or
4084 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4085 * 3) running (obviously), or
4086 * 4) are cache-hot on their current CPU.
4088 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4091 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4094 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4096 env->flags |= LBF_SOME_PINNED;
4099 * Remember if this task can be migrated to any other cpu in
4100 * our sched_group. We may want to revisit it if we couldn't
4101 * meet load balance goals by pulling other tasks on src_cpu.
4103 * Also avoid computing new_dst_cpu if we have already computed
4104 * one in current iteration.
4106 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4109 /* Prevent to re-select dst_cpu via env's cpus */
4110 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4111 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4112 env->flags |= LBF_DST_PINNED;
4113 env->new_dst_cpu = cpu;
4121 /* Record that we found atleast one task that could run on dst_cpu */
4122 env->flags &= ~LBF_ALL_PINNED;
4124 if (task_running(env->src_rq, p)) {
4125 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4130 * Aggressive migration if:
4131 * 1) task is cache cold, or
4132 * 2) too many balance attempts have failed.
4135 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4136 if (!tsk_cache_hot ||
4137 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4139 if (tsk_cache_hot) {
4140 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4141 schedstat_inc(p, se.statistics.nr_forced_migrations);
4147 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4152 * move_one_task tries to move exactly one task from busiest to this_rq, as
4153 * part of active balancing operations within "domain".
4154 * Returns 1 if successful and 0 otherwise.
4156 * Called with both runqueues locked.
4158 static int move_one_task(struct lb_env *env)
4160 struct task_struct *p, *n;
4162 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4163 if (!can_migrate_task(p, env))
4168 * Right now, this is only the second place move_task()
4169 * is called, so we can safely collect move_task()
4170 * stats here rather than inside move_task().
4172 schedstat_inc(env->sd, lb_gained[env->idle]);
4178 static unsigned long task_h_load(struct task_struct *p);
4180 static const unsigned int sched_nr_migrate_break = 32;
4183 * move_tasks tries to move up to imbalance weighted load from busiest to
4184 * this_rq, as part of a balancing operation within domain "sd".
4185 * Returns 1 if successful and 0 otherwise.
4187 * Called with both runqueues locked.
4189 static int move_tasks(struct lb_env *env)
4191 struct list_head *tasks = &env->src_rq->cfs_tasks;
4192 struct task_struct *p;
4196 if (env->imbalance <= 0)
4199 while (!list_empty(tasks)) {
4200 p = list_first_entry(tasks, struct task_struct, se.group_node);
4203 /* We've more or less seen every task there is, call it quits */
4204 if (env->loop > env->loop_max)
4207 /* take a breather every nr_migrate tasks */
4208 if (env->loop > env->loop_break) {
4209 env->loop_break += sched_nr_migrate_break;
4210 env->flags |= LBF_NEED_BREAK;
4214 if (!can_migrate_task(p, env))
4217 load = task_h_load(p);
4219 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4222 if ((load / 2) > env->imbalance)
4227 env->imbalance -= load;
4229 #ifdef CONFIG_PREEMPT
4231 * NEWIDLE balancing is a source of latency, so preemptible
4232 * kernels will stop after the first task is pulled to minimize
4233 * the critical section.
4235 if (env->idle == CPU_NEWLY_IDLE)
4240 * We only want to steal up to the prescribed amount of
4243 if (env->imbalance <= 0)
4248 list_move_tail(&p->se.group_node, tasks);
4252 * Right now, this is one of only two places move_task() is called,
4253 * so we can safely collect move_task() stats here rather than
4254 * inside move_task().
4256 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4261 #ifdef CONFIG_FAIR_GROUP_SCHED
4263 * update tg->load_weight by folding this cpu's load_avg
4265 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4267 struct sched_entity *se = tg->se[cpu];
4268 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4270 /* throttled entities do not contribute to load */
4271 if (throttled_hierarchy(cfs_rq))
4274 update_cfs_rq_blocked_load(cfs_rq, 1);
4277 update_entity_load_avg(se, 1);
4279 * We pivot on our runnable average having decayed to zero for
4280 * list removal. This generally implies that all our children
4281 * have also been removed (modulo rounding error or bandwidth
4282 * control); however, such cases are rare and we can fix these
4285 * TODO: fix up out-of-order children on enqueue.
4287 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4288 list_del_leaf_cfs_rq(cfs_rq);
4290 struct rq *rq = rq_of(cfs_rq);
4291 update_rq_runnable_avg(rq, rq->nr_running);
4295 static void update_blocked_averages(int cpu)
4297 struct rq *rq = cpu_rq(cpu);
4298 struct cfs_rq *cfs_rq;
4299 unsigned long flags;
4301 raw_spin_lock_irqsave(&rq->lock, flags);
4302 update_rq_clock(rq);
4304 * Iterates the task_group tree in a bottom up fashion, see
4305 * list_add_leaf_cfs_rq() for details.
4307 for_each_leaf_cfs_rq(rq, cfs_rq) {
4309 * Note: We may want to consider periodically releasing
4310 * rq->lock about these updates so that creating many task
4311 * groups does not result in continually extending hold time.
4313 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4316 raw_spin_unlock_irqrestore(&rq->lock, flags);
4320 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4321 * This needs to be done in a top-down fashion because the load of a child
4322 * group is a fraction of its parents load.
4324 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4326 struct rq *rq = rq_of(cfs_rq);
4327 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4328 unsigned long now = jiffies;
4331 if (cfs_rq->last_h_load_update == now)
4334 cfs_rq->h_load_next = NULL;
4335 for_each_sched_entity(se) {
4336 cfs_rq = cfs_rq_of(se);
4337 cfs_rq->h_load_next = se;
4338 if (cfs_rq->last_h_load_update == now)
4343 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4344 cfs_rq->last_h_load_update = now;
4347 while ((se = cfs_rq->h_load_next) != NULL) {
4348 load = cfs_rq->h_load;
4349 load = div64_ul(load * se->avg.load_avg_contrib,
4350 cfs_rq->runnable_load_avg + 1);
4351 cfs_rq = group_cfs_rq(se);
4352 cfs_rq->h_load = load;
4353 cfs_rq->last_h_load_update = now;
4357 static unsigned long task_h_load(struct task_struct *p)
4359 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4361 update_cfs_rq_h_load(cfs_rq);
4362 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4363 cfs_rq->runnable_load_avg + 1);
4366 static inline void update_blocked_averages(int cpu)
4370 static unsigned long task_h_load(struct task_struct *p)
4372 return p->se.avg.load_avg_contrib;
4376 /********** Helpers for find_busiest_group ************************/
4378 * sg_lb_stats - stats of a sched_group required for load_balancing
4380 struct sg_lb_stats {
4381 unsigned long avg_load; /*Avg load across the CPUs of the group */
4382 unsigned long group_load; /* Total load over the CPUs of the group */
4383 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4384 unsigned long load_per_task;
4385 unsigned long group_power;
4386 unsigned int sum_nr_running; /* Nr tasks running in the group */
4387 unsigned int group_capacity;
4388 unsigned int idle_cpus;
4389 unsigned int group_weight;
4390 int group_imb; /* Is there an imbalance in the group ? */
4391 int group_has_capacity; /* Is there extra capacity in the group? */
4395 * sd_lb_stats - Structure to store the statistics of a sched_domain
4396 * during load balancing.
4398 struct sd_lb_stats {
4399 struct sched_group *busiest; /* Busiest group in this sd */
4400 struct sched_group *local; /* Local group in this sd */
4401 unsigned long total_load; /* Total load of all groups in sd */
4402 unsigned long total_pwr; /* Total power of all groups in sd */
4403 unsigned long avg_load; /* Average load across all groups in sd */
4405 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4406 struct sg_lb_stats local_stat; /* Statistics of the local group */
4409 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4412 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4413 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4414 * We must however clear busiest_stat::avg_load because
4415 * update_sd_pick_busiest() reads this before assignment.
4417 *sds = (struct sd_lb_stats){
4429 * get_sd_load_idx - Obtain the load index for a given sched domain.
4430 * @sd: The sched_domain whose load_idx is to be obtained.
4431 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4433 * Return: The load index.
4435 static inline int get_sd_load_idx(struct sched_domain *sd,
4436 enum cpu_idle_type idle)
4442 load_idx = sd->busy_idx;
4445 case CPU_NEWLY_IDLE:
4446 load_idx = sd->newidle_idx;
4449 load_idx = sd->idle_idx;
4456 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4458 return SCHED_POWER_SCALE;
4461 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4463 return default_scale_freq_power(sd, cpu);
4466 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4468 unsigned long weight = sd->span_weight;
4469 unsigned long smt_gain = sd->smt_gain;
4476 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4478 return default_scale_smt_power(sd, cpu);
4481 static unsigned long scale_rt_power(int cpu)
4483 struct rq *rq = cpu_rq(cpu);
4484 u64 total, available, age_stamp, avg;
4487 * Since we're reading these variables without serialization make sure
4488 * we read them once before doing sanity checks on them.
4490 age_stamp = ACCESS_ONCE(rq->age_stamp);
4491 avg = ACCESS_ONCE(rq->rt_avg);
4493 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4495 if (unlikely(total < avg)) {
4496 /* Ensures that power won't end up being negative */
4499 available = total - avg;
4502 if (unlikely((s64)total < SCHED_POWER_SCALE))
4503 total = SCHED_POWER_SCALE;
4505 total >>= SCHED_POWER_SHIFT;
4507 return div_u64(available, total);
4510 static void update_cpu_power(struct sched_domain *sd, int cpu)
4512 unsigned long weight = sd->span_weight;
4513 unsigned long power = SCHED_POWER_SCALE;
4514 struct sched_group *sdg = sd->groups;
4516 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4517 if (sched_feat(ARCH_POWER))
4518 power *= arch_scale_smt_power(sd, cpu);
4520 power *= default_scale_smt_power(sd, cpu);
4522 power >>= SCHED_POWER_SHIFT;
4525 sdg->sgp->power_orig = power;
4527 if (sched_feat(ARCH_POWER))
4528 power *= arch_scale_freq_power(sd, cpu);
4530 power *= default_scale_freq_power(sd, cpu);
4532 power >>= SCHED_POWER_SHIFT;
4534 power *= scale_rt_power(cpu);
4535 power >>= SCHED_POWER_SHIFT;
4540 cpu_rq(cpu)->cpu_power = power;
4541 sdg->sgp->power = power;
4544 void update_group_power(struct sched_domain *sd, int cpu)
4546 struct sched_domain *child = sd->child;
4547 struct sched_group *group, *sdg = sd->groups;
4548 unsigned long power, power_orig;
4549 unsigned long interval;
4551 interval = msecs_to_jiffies(sd->balance_interval);
4552 interval = clamp(interval, 1UL, max_load_balance_interval);
4553 sdg->sgp->next_update = jiffies + interval;
4556 update_cpu_power(sd, cpu);
4560 power_orig = power = 0;
4562 if (child->flags & SD_OVERLAP) {
4564 * SD_OVERLAP domains cannot assume that child groups
4565 * span the current group.
4568 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4569 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4571 power_orig += sg->sgp->power_orig;
4572 power += sg->sgp->power;
4576 * !SD_OVERLAP domains can assume that child groups
4577 * span the current group.
4580 group = child->groups;
4582 power_orig += group->sgp->power_orig;
4583 power += group->sgp->power;
4584 group = group->next;
4585 } while (group != child->groups);
4588 sdg->sgp->power_orig = power_orig;
4589 sdg->sgp->power = power;
4593 * Try and fix up capacity for tiny siblings, this is needed when
4594 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4595 * which on its own isn't powerful enough.
4597 * See update_sd_pick_busiest() and check_asym_packing().
4600 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4603 * Only siblings can have significantly less than SCHED_POWER_SCALE
4605 if (!(sd->flags & SD_SHARE_CPUPOWER))
4609 * If ~90% of the cpu_power is still there, we're good.
4611 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4618 * Group imbalance indicates (and tries to solve) the problem where balancing
4619 * groups is inadequate due to tsk_cpus_allowed() constraints.
4621 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4622 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4625 * { 0 1 2 3 } { 4 5 6 7 }
4628 * If we were to balance group-wise we'd place two tasks in the first group and
4629 * two tasks in the second group. Clearly this is undesired as it will overload
4630 * cpu 3 and leave one of the cpus in the second group unused.
4632 * The current solution to this issue is detecting the skew in the first group
4633 * by noticing the lower domain failed to reach balance and had difficulty
4634 * moving tasks due to affinity constraints.
4636 * When this is so detected; this group becomes a candidate for busiest; see
4637 * update_sd_pick_busiest(). And calculcate_imbalance() and
4638 * find_busiest_group() avoid some of the usual balance conditions to allow it
4639 * to create an effective group imbalance.
4641 * This is a somewhat tricky proposition since the next run might not find the
4642 * group imbalance and decide the groups need to be balanced again. A most
4643 * subtle and fragile situation.
4646 static inline int sg_imbalanced(struct sched_group *group)
4648 return group->sgp->imbalance;
4652 * Compute the group capacity.
4654 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4655 * first dividing out the smt factor and computing the actual number of cores
4656 * and limit power unit capacity with that.
4658 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4660 unsigned int capacity, smt, cpus;
4661 unsigned int power, power_orig;
4663 power = group->sgp->power;
4664 power_orig = group->sgp->power_orig;
4665 cpus = group->group_weight;
4667 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4668 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4669 capacity = cpus / smt; /* cores */
4671 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4673 capacity = fix_small_capacity(env->sd, group);
4679 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4680 * @env: The load balancing environment.
4681 * @group: sched_group whose statistics are to be updated.
4682 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4683 * @local_group: Does group contain this_cpu.
4684 * @sgs: variable to hold the statistics for this group.
4686 static inline void update_sg_lb_stats(struct lb_env *env,
4687 struct sched_group *group, int load_idx,
4688 int local_group, struct sg_lb_stats *sgs)
4690 unsigned long nr_running;
4694 memset(sgs, 0, sizeof(*sgs));
4696 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4697 struct rq *rq = cpu_rq(i);
4699 nr_running = rq->nr_running;
4701 /* Bias balancing toward cpus of our domain */
4703 load = target_load(i, load_idx);
4705 load = source_load(i, load_idx);
4707 sgs->group_load += load;
4708 sgs->sum_nr_running += nr_running;
4709 sgs->sum_weighted_load += weighted_cpuload(i);
4714 /* Adjust by relative CPU power of the group */
4715 sgs->group_power = group->sgp->power;
4716 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4718 if (sgs->sum_nr_running)
4719 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4721 sgs->group_weight = group->group_weight;
4723 sgs->group_imb = sg_imbalanced(group);
4724 sgs->group_capacity = sg_capacity(env, group);
4726 if (sgs->group_capacity > sgs->sum_nr_running)
4727 sgs->group_has_capacity = 1;
4731 * update_sd_pick_busiest - return 1 on busiest group
4732 * @env: The load balancing environment.
4733 * @sds: sched_domain statistics
4734 * @sg: sched_group candidate to be checked for being the busiest
4735 * @sgs: sched_group statistics
4737 * Determine if @sg is a busier group than the previously selected
4740 * Return: %true if @sg is a busier group than the previously selected
4741 * busiest group. %false otherwise.
4743 static bool update_sd_pick_busiest(struct lb_env *env,
4744 struct sd_lb_stats *sds,
4745 struct sched_group *sg,
4746 struct sg_lb_stats *sgs)
4748 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4751 if (sgs->sum_nr_running > sgs->group_capacity)
4758 * ASYM_PACKING needs to move all the work to the lowest
4759 * numbered CPUs in the group, therefore mark all groups
4760 * higher than ourself as busy.
4762 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4763 env->dst_cpu < group_first_cpu(sg)) {
4767 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4775 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4776 * @env: The load balancing environment.
4777 * @balance: Should we balance.
4778 * @sds: variable to hold the statistics for this sched_domain.
4780 static inline void update_sd_lb_stats(struct lb_env *env,
4781 struct sd_lb_stats *sds)
4783 struct sched_domain *child = env->sd->child;
4784 struct sched_group *sg = env->sd->groups;
4785 struct sg_lb_stats tmp_sgs;
4786 int load_idx, prefer_sibling = 0;
4788 if (child && child->flags & SD_PREFER_SIBLING)
4791 load_idx = get_sd_load_idx(env->sd, env->idle);
4794 struct sg_lb_stats *sgs = &tmp_sgs;
4797 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4800 sgs = &sds->local_stat;
4802 if (env->idle != CPU_NEWLY_IDLE ||
4803 time_after_eq(jiffies, sg->sgp->next_update))
4804 update_group_power(env->sd, env->dst_cpu);
4807 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4813 * In case the child domain prefers tasks go to siblings
4814 * first, lower the sg capacity to one so that we'll try
4815 * and move all the excess tasks away. We lower the capacity
4816 * of a group only if the local group has the capacity to fit
4817 * these excess tasks, i.e. nr_running < group_capacity. The
4818 * extra check prevents the case where you always pull from the
4819 * heaviest group when it is already under-utilized (possible
4820 * with a large weight task outweighs the tasks on the system).
4822 if (prefer_sibling && sds->local &&
4823 sds->local_stat.group_has_capacity)
4824 sgs->group_capacity = min(sgs->group_capacity, 1U);
4826 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4828 sds->busiest_stat = *sgs;
4832 /* Now, start updating sd_lb_stats */
4833 sds->total_load += sgs->group_load;
4834 sds->total_pwr += sgs->group_power;
4837 } while (sg != env->sd->groups);
4841 * check_asym_packing - Check to see if the group is packed into the
4844 * This is primarily intended to used at the sibling level. Some
4845 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4846 * case of POWER7, it can move to lower SMT modes only when higher
4847 * threads are idle. When in lower SMT modes, the threads will
4848 * perform better since they share less core resources. Hence when we
4849 * have idle threads, we want them to be the higher ones.
4851 * This packing function is run on idle threads. It checks to see if
4852 * the busiest CPU in this domain (core in the P7 case) has a higher
4853 * CPU number than the packing function is being run on. Here we are
4854 * assuming lower CPU number will be equivalent to lower a SMT thread
4857 * Return: 1 when packing is required and a task should be moved to
4858 * this CPU. The amount of the imbalance is returned in *imbalance.
4860 * @env: The load balancing environment.
4861 * @sds: Statistics of the sched_domain which is to be packed
4863 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4867 if (!(env->sd->flags & SD_ASYM_PACKING))
4873 busiest_cpu = group_first_cpu(sds->busiest);
4874 if (env->dst_cpu > busiest_cpu)
4877 env->imbalance = DIV_ROUND_CLOSEST(
4878 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4885 * fix_small_imbalance - Calculate the minor imbalance that exists
4886 * amongst the groups of a sched_domain, during
4888 * @env: The load balancing environment.
4889 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4892 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4894 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4895 unsigned int imbn = 2;
4896 unsigned long scaled_busy_load_per_task;
4897 struct sg_lb_stats *local, *busiest;
4899 local = &sds->local_stat;
4900 busiest = &sds->busiest_stat;
4902 if (!local->sum_nr_running)
4903 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4904 else if (busiest->load_per_task > local->load_per_task)
4907 scaled_busy_load_per_task =
4908 (busiest->load_per_task * SCHED_POWER_SCALE) /
4909 busiest->group_power;
4911 if (busiest->avg_load + scaled_busy_load_per_task >=
4912 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4913 env->imbalance = busiest->load_per_task;
4918 * OK, we don't have enough imbalance to justify moving tasks,
4919 * however we may be able to increase total CPU power used by
4923 pwr_now += busiest->group_power *
4924 min(busiest->load_per_task, busiest->avg_load);
4925 pwr_now += local->group_power *
4926 min(local->load_per_task, local->avg_load);
4927 pwr_now /= SCHED_POWER_SCALE;
4929 /* Amount of load we'd subtract */
4930 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4931 busiest->group_power;
4932 if (busiest->avg_load > tmp) {
4933 pwr_move += busiest->group_power *
4934 min(busiest->load_per_task,
4935 busiest->avg_load - tmp);
4938 /* Amount of load we'd add */
4939 if (busiest->avg_load * busiest->group_power <
4940 busiest->load_per_task * SCHED_POWER_SCALE) {
4941 tmp = (busiest->avg_load * busiest->group_power) /
4944 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4947 pwr_move += local->group_power *
4948 min(local->load_per_task, local->avg_load + tmp);
4949 pwr_move /= SCHED_POWER_SCALE;
4951 /* Move if we gain throughput */
4952 if (pwr_move > pwr_now)
4953 env->imbalance = busiest->load_per_task;
4957 * calculate_imbalance - Calculate the amount of imbalance present within the
4958 * groups of a given sched_domain during load balance.
4959 * @env: load balance environment
4960 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4962 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4964 unsigned long max_pull, load_above_capacity = ~0UL;
4965 struct sg_lb_stats *local, *busiest;
4967 local = &sds->local_stat;
4968 busiest = &sds->busiest_stat;
4970 if (busiest->group_imb) {
4972 * In the group_imb case we cannot rely on group-wide averages
4973 * to ensure cpu-load equilibrium, look at wider averages. XXX
4975 busiest->load_per_task =
4976 min(busiest->load_per_task, sds->avg_load);
4980 * In the presence of smp nice balancing, certain scenarios can have
4981 * max load less than avg load(as we skip the groups at or below
4982 * its cpu_power, while calculating max_load..)
4984 if (busiest->avg_load <= sds->avg_load ||
4985 local->avg_load >= sds->avg_load) {
4987 return fix_small_imbalance(env, sds);
4990 if (!busiest->group_imb) {
4992 * Don't want to pull so many tasks that a group would go idle.
4993 * Except of course for the group_imb case, since then we might
4994 * have to drop below capacity to reach cpu-load equilibrium.
4996 load_above_capacity =
4997 (busiest->sum_nr_running - busiest->group_capacity);
4999 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5000 load_above_capacity /= busiest->group_power;
5004 * We're trying to get all the cpus to the average_load, so we don't
5005 * want to push ourselves above the average load, nor do we wish to
5006 * reduce the max loaded cpu below the average load. At the same time,
5007 * we also don't want to reduce the group load below the group capacity
5008 * (so that we can implement power-savings policies etc). Thus we look
5009 * for the minimum possible imbalance.
5011 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5013 /* How much load to actually move to equalise the imbalance */
5014 env->imbalance = min(
5015 max_pull * busiest->group_power,
5016 (sds->avg_load - local->avg_load) * local->group_power
5017 ) / SCHED_POWER_SCALE;
5020 * if *imbalance is less than the average load per runnable task
5021 * there is no guarantee that any tasks will be moved so we'll have
5022 * a think about bumping its value to force at least one task to be
5025 if (env->imbalance < busiest->load_per_task)
5026 return fix_small_imbalance(env, sds);
5029 /******* find_busiest_group() helpers end here *********************/
5032 * find_busiest_group - Returns the busiest group within the sched_domain
5033 * if there is an imbalance. If there isn't an imbalance, and
5034 * the user has opted for power-savings, it returns a group whose
5035 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5036 * such a group exists.
5038 * Also calculates the amount of weighted load which should be moved
5039 * to restore balance.
5041 * @env: The load balancing environment.
5043 * Return: - The busiest group if imbalance exists.
5044 * - If no imbalance and user has opted for power-savings balance,
5045 * return the least loaded group whose CPUs can be
5046 * put to idle by rebalancing its tasks onto our group.
5048 static struct sched_group *find_busiest_group(struct lb_env *env)
5050 struct sg_lb_stats *local, *busiest;
5051 struct sd_lb_stats sds;
5053 init_sd_lb_stats(&sds);
5056 * Compute the various statistics relavent for load balancing at
5059 update_sd_lb_stats(env, &sds);
5060 local = &sds.local_stat;
5061 busiest = &sds.busiest_stat;
5063 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5064 check_asym_packing(env, &sds))
5067 /* There is no busy sibling group to pull tasks from */
5068 if (!sds.busiest || busiest->sum_nr_running == 0)
5071 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5074 * If the busiest group is imbalanced the below checks don't
5075 * work because they assume all things are equal, which typically
5076 * isn't true due to cpus_allowed constraints and the like.
5078 if (busiest->group_imb)
5081 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5082 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5083 !busiest->group_has_capacity)
5087 * If the local group is more busy than the selected busiest group
5088 * don't try and pull any tasks.
5090 if (local->avg_load >= busiest->avg_load)
5094 * Don't pull any tasks if this group is already above the domain
5097 if (local->avg_load >= sds.avg_load)
5100 if (env->idle == CPU_IDLE) {
5102 * This cpu is idle. If the busiest group load doesn't
5103 * have more tasks than the number of available cpu's and
5104 * there is no imbalance between this and busiest group
5105 * wrt to idle cpu's, it is balanced.
5107 if ((local->idle_cpus < busiest->idle_cpus) &&
5108 busiest->sum_nr_running <= busiest->group_weight)
5112 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5113 * imbalance_pct to be conservative.
5115 if (100 * busiest->avg_load <=
5116 env->sd->imbalance_pct * local->avg_load)
5121 /* Looks like there is an imbalance. Compute it */
5122 calculate_imbalance(env, &sds);
5131 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5133 static struct rq *find_busiest_queue(struct lb_env *env,
5134 struct sched_group *group)
5136 struct rq *busiest = NULL, *rq;
5137 unsigned long busiest_load = 0, busiest_power = 1;
5140 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5141 unsigned long power = power_of(i);
5142 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5147 capacity = fix_small_capacity(env->sd, group);
5150 wl = weighted_cpuload(i);
5153 * When comparing with imbalance, use weighted_cpuload()
5154 * which is not scaled with the cpu power.
5156 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5160 * For the load comparisons with the other cpu's, consider
5161 * the weighted_cpuload() scaled with the cpu power, so that
5162 * the load can be moved away from the cpu that is potentially
5163 * running at a lower capacity.
5165 * Thus we're looking for max(wl_i / power_i), crosswise
5166 * multiplication to rid ourselves of the division works out
5167 * to: wl_i * power_j > wl_j * power_i; where j is our
5170 if (wl * busiest_power > busiest_load * power) {
5172 busiest_power = power;
5181 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5182 * so long as it is large enough.
5184 #define MAX_PINNED_INTERVAL 512
5186 /* Working cpumask for load_balance and load_balance_newidle. */
5187 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5189 static int need_active_balance(struct lb_env *env)
5191 struct sched_domain *sd = env->sd;
5193 if (env->idle == CPU_NEWLY_IDLE) {
5196 * ASYM_PACKING needs to force migrate tasks from busy but
5197 * higher numbered CPUs in order to pack all tasks in the
5198 * lowest numbered CPUs.
5200 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5204 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5207 static int active_load_balance_cpu_stop(void *data);
5209 static int should_we_balance(struct lb_env *env)
5211 struct sched_group *sg = env->sd->groups;
5212 struct cpumask *sg_cpus, *sg_mask;
5213 int cpu, balance_cpu = -1;
5216 * In the newly idle case, we will allow all the cpu's
5217 * to do the newly idle load balance.
5219 if (env->idle == CPU_NEWLY_IDLE)
5222 sg_cpus = sched_group_cpus(sg);
5223 sg_mask = sched_group_mask(sg);
5224 /* Try to find first idle cpu */
5225 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5226 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5233 if (balance_cpu == -1)
5234 balance_cpu = group_balance_cpu(sg);
5237 * First idle cpu or the first cpu(busiest) in this sched group
5238 * is eligible for doing load balancing at this and above domains.
5240 return balance_cpu == env->dst_cpu;
5244 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5245 * tasks if there is an imbalance.
5247 static int load_balance(int this_cpu, struct rq *this_rq,
5248 struct sched_domain *sd, enum cpu_idle_type idle,
5249 int *continue_balancing)
5251 int ld_moved, cur_ld_moved, active_balance = 0;
5252 struct sched_domain *sd_parent = sd->parent;
5253 struct sched_group *group;
5255 unsigned long flags;
5256 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5258 struct lb_env env = {
5260 .dst_cpu = this_cpu,
5262 .dst_grpmask = sched_group_cpus(sd->groups),
5264 .loop_break = sched_nr_migrate_break,
5269 * For NEWLY_IDLE load_balancing, we don't need to consider
5270 * other cpus in our group
5272 if (idle == CPU_NEWLY_IDLE)
5273 env.dst_grpmask = NULL;
5275 cpumask_copy(cpus, cpu_active_mask);
5277 schedstat_inc(sd, lb_count[idle]);
5280 if (!should_we_balance(&env)) {
5281 *continue_balancing = 0;
5285 group = find_busiest_group(&env);
5287 schedstat_inc(sd, lb_nobusyg[idle]);
5291 busiest = find_busiest_queue(&env, group);
5293 schedstat_inc(sd, lb_nobusyq[idle]);
5297 BUG_ON(busiest == env.dst_rq);
5299 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5302 if (busiest->nr_running > 1) {
5304 * Attempt to move tasks. If find_busiest_group has found
5305 * an imbalance but busiest->nr_running <= 1, the group is
5306 * still unbalanced. ld_moved simply stays zero, so it is
5307 * correctly treated as an imbalance.
5309 env.flags |= LBF_ALL_PINNED;
5310 env.src_cpu = busiest->cpu;
5311 env.src_rq = busiest;
5312 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5315 local_irq_save(flags);
5316 double_rq_lock(env.dst_rq, busiest);
5319 * cur_ld_moved - load moved in current iteration
5320 * ld_moved - cumulative load moved across iterations
5322 cur_ld_moved = move_tasks(&env);
5323 ld_moved += cur_ld_moved;
5324 double_rq_unlock(env.dst_rq, busiest);
5325 local_irq_restore(flags);
5328 * some other cpu did the load balance for us.
5330 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5331 resched_cpu(env.dst_cpu);
5333 if (env.flags & LBF_NEED_BREAK) {
5334 env.flags &= ~LBF_NEED_BREAK;
5339 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5340 * us and move them to an alternate dst_cpu in our sched_group
5341 * where they can run. The upper limit on how many times we
5342 * iterate on same src_cpu is dependent on number of cpus in our
5345 * This changes load balance semantics a bit on who can move
5346 * load to a given_cpu. In addition to the given_cpu itself
5347 * (or a ilb_cpu acting on its behalf where given_cpu is
5348 * nohz-idle), we now have balance_cpu in a position to move
5349 * load to given_cpu. In rare situations, this may cause
5350 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5351 * _independently_ and at _same_ time to move some load to
5352 * given_cpu) causing exceess load to be moved to given_cpu.
5353 * This however should not happen so much in practice and
5354 * moreover subsequent load balance cycles should correct the
5355 * excess load moved.
5357 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5359 /* Prevent to re-select dst_cpu via env's cpus */
5360 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5362 env.dst_rq = cpu_rq(env.new_dst_cpu);
5363 env.dst_cpu = env.new_dst_cpu;
5364 env.flags &= ~LBF_DST_PINNED;
5366 env.loop_break = sched_nr_migrate_break;
5369 * Go back to "more_balance" rather than "redo" since we
5370 * need to continue with same src_cpu.
5376 * We failed to reach balance because of affinity.
5379 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5381 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5382 *group_imbalance = 1;
5383 } else if (*group_imbalance)
5384 *group_imbalance = 0;
5387 /* All tasks on this runqueue were pinned by CPU affinity */
5388 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5389 cpumask_clear_cpu(cpu_of(busiest), cpus);
5390 if (!cpumask_empty(cpus)) {
5392 env.loop_break = sched_nr_migrate_break;
5400 schedstat_inc(sd, lb_failed[idle]);
5402 * Increment the failure counter only on periodic balance.
5403 * We do not want newidle balance, which can be very
5404 * frequent, pollute the failure counter causing
5405 * excessive cache_hot migrations and active balances.
5407 if (idle != CPU_NEWLY_IDLE)
5408 sd->nr_balance_failed++;
5410 if (need_active_balance(&env)) {
5411 raw_spin_lock_irqsave(&busiest->lock, flags);
5413 /* don't kick the active_load_balance_cpu_stop,
5414 * if the curr task on busiest cpu can't be
5417 if (!cpumask_test_cpu(this_cpu,
5418 tsk_cpus_allowed(busiest->curr))) {
5419 raw_spin_unlock_irqrestore(&busiest->lock,
5421 env.flags |= LBF_ALL_PINNED;
5422 goto out_one_pinned;
5426 * ->active_balance synchronizes accesses to
5427 * ->active_balance_work. Once set, it's cleared
5428 * only after active load balance is finished.
5430 if (!busiest->active_balance) {
5431 busiest->active_balance = 1;
5432 busiest->push_cpu = this_cpu;
5435 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5437 if (active_balance) {
5438 stop_one_cpu_nowait(cpu_of(busiest),
5439 active_load_balance_cpu_stop, busiest,
5440 &busiest->active_balance_work);
5444 * We've kicked active balancing, reset the failure
5447 sd->nr_balance_failed = sd->cache_nice_tries+1;
5450 sd->nr_balance_failed = 0;
5452 if (likely(!active_balance)) {
5453 /* We were unbalanced, so reset the balancing interval */
5454 sd->balance_interval = sd->min_interval;
5457 * If we've begun active balancing, start to back off. This
5458 * case may not be covered by the all_pinned logic if there
5459 * is only 1 task on the busy runqueue (because we don't call
5462 if (sd->balance_interval < sd->max_interval)
5463 sd->balance_interval *= 2;
5469 schedstat_inc(sd, lb_balanced[idle]);
5471 sd->nr_balance_failed = 0;
5474 /* tune up the balancing interval */
5475 if (((env.flags & LBF_ALL_PINNED) &&
5476 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5477 (sd->balance_interval < sd->max_interval))
5478 sd->balance_interval *= 2;
5486 * idle_balance is called by schedule() if this_cpu is about to become
5487 * idle. Attempts to pull tasks from other CPUs.
5489 void idle_balance(int this_cpu, struct rq *this_rq)
5491 struct sched_domain *sd;
5492 int pulled_task = 0;
5493 unsigned long next_balance = jiffies + HZ;
5496 this_rq->idle_stamp = rq_clock(this_rq);
5498 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5502 * Drop the rq->lock, but keep IRQ/preempt disabled.
5504 raw_spin_unlock(&this_rq->lock);
5506 update_blocked_averages(this_cpu);
5508 for_each_domain(this_cpu, sd) {
5509 unsigned long interval;
5510 int continue_balancing = 1;
5511 u64 t0, domain_cost;
5513 if (!(sd->flags & SD_LOAD_BALANCE))
5516 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5519 if (sd->flags & SD_BALANCE_NEWIDLE) {
5520 t0 = sched_clock_cpu(this_cpu);
5522 /* If we've pulled tasks over stop searching: */
5523 pulled_task = load_balance(this_cpu, this_rq,
5525 &continue_balancing);
5527 domain_cost = sched_clock_cpu(this_cpu) - t0;
5528 if (domain_cost > sd->max_newidle_lb_cost)
5529 sd->max_newidle_lb_cost = domain_cost;
5531 curr_cost += domain_cost;
5534 interval = msecs_to_jiffies(sd->balance_interval);
5535 if (time_after(next_balance, sd->last_balance + interval))
5536 next_balance = sd->last_balance + interval;
5538 this_rq->idle_stamp = 0;
5544 raw_spin_lock(&this_rq->lock);
5546 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5548 * We are going idle. next_balance may be set based on
5549 * a busy processor. So reset next_balance.
5551 this_rq->next_balance = next_balance;
5554 if (curr_cost > this_rq->max_idle_balance_cost)
5555 this_rq->max_idle_balance_cost = curr_cost;
5559 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5560 * running tasks off the busiest CPU onto idle CPUs. It requires at
5561 * least 1 task to be running on each physical CPU where possible, and
5562 * avoids physical / logical imbalances.
5564 static int active_load_balance_cpu_stop(void *data)
5566 struct rq *busiest_rq = data;
5567 int busiest_cpu = cpu_of(busiest_rq);
5568 int target_cpu = busiest_rq->push_cpu;
5569 struct rq *target_rq = cpu_rq(target_cpu);
5570 struct sched_domain *sd;
5572 raw_spin_lock_irq(&busiest_rq->lock);
5574 /* make sure the requested cpu hasn't gone down in the meantime */
5575 if (unlikely(busiest_cpu != smp_processor_id() ||
5576 !busiest_rq->active_balance))
5579 /* Is there any task to move? */
5580 if (busiest_rq->nr_running <= 1)
5584 * This condition is "impossible", if it occurs
5585 * we need to fix it. Originally reported by
5586 * Bjorn Helgaas on a 128-cpu setup.
5588 BUG_ON(busiest_rq == target_rq);
5590 /* move a task from busiest_rq to target_rq */
5591 double_lock_balance(busiest_rq, target_rq);
5593 /* Search for an sd spanning us and the target CPU. */
5595 for_each_domain(target_cpu, sd) {
5596 if ((sd->flags & SD_LOAD_BALANCE) &&
5597 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5602 struct lb_env env = {
5604 .dst_cpu = target_cpu,
5605 .dst_rq = target_rq,
5606 .src_cpu = busiest_rq->cpu,
5607 .src_rq = busiest_rq,
5611 schedstat_inc(sd, alb_count);
5613 if (move_one_task(&env))
5614 schedstat_inc(sd, alb_pushed);
5616 schedstat_inc(sd, alb_failed);
5619 double_unlock_balance(busiest_rq, target_rq);
5621 busiest_rq->active_balance = 0;
5622 raw_spin_unlock_irq(&busiest_rq->lock);
5626 #ifdef CONFIG_NO_HZ_COMMON
5628 * idle load balancing details
5629 * - When one of the busy CPUs notice that there may be an idle rebalancing
5630 * needed, they will kick the idle load balancer, which then does idle
5631 * load balancing for all the idle CPUs.
5634 cpumask_var_t idle_cpus_mask;
5636 unsigned long next_balance; /* in jiffy units */
5637 } nohz ____cacheline_aligned;
5639 static inline int find_new_ilb(int call_cpu)
5641 int ilb = cpumask_first(nohz.idle_cpus_mask);
5643 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5650 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5651 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5652 * CPU (if there is one).
5654 static void nohz_balancer_kick(int cpu)
5658 nohz.next_balance++;
5660 ilb_cpu = find_new_ilb(cpu);
5662 if (ilb_cpu >= nr_cpu_ids)
5665 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5668 * Use smp_send_reschedule() instead of resched_cpu().
5669 * This way we generate a sched IPI on the target cpu which
5670 * is idle. And the softirq performing nohz idle load balance
5671 * will be run before returning from the IPI.
5673 smp_send_reschedule(ilb_cpu);
5677 static inline void nohz_balance_exit_idle(int cpu)
5679 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5680 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5681 atomic_dec(&nohz.nr_cpus);
5682 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5686 static inline void set_cpu_sd_state_busy(void)
5688 struct sched_domain *sd;
5691 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5693 if (!sd || !sd->nohz_idle)
5697 for (; sd; sd = sd->parent)
5698 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5703 void set_cpu_sd_state_idle(void)
5705 struct sched_domain *sd;
5708 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5710 if (!sd || sd->nohz_idle)
5714 for (; sd; sd = sd->parent)
5715 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5721 * This routine will record that the cpu is going idle with tick stopped.
5722 * This info will be used in performing idle load balancing in the future.
5724 void nohz_balance_enter_idle(int cpu)
5727 * If this cpu is going down, then nothing needs to be done.
5729 if (!cpu_active(cpu))
5732 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5735 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5736 atomic_inc(&nohz.nr_cpus);
5737 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5740 static int sched_ilb_notifier(struct notifier_block *nfb,
5741 unsigned long action, void *hcpu)
5743 switch (action & ~CPU_TASKS_FROZEN) {
5745 nohz_balance_exit_idle(smp_processor_id());
5753 static DEFINE_SPINLOCK(balancing);
5756 * Scale the max load_balance interval with the number of CPUs in the system.
5757 * This trades load-balance latency on larger machines for less cross talk.
5759 void update_max_interval(void)
5761 max_load_balance_interval = HZ*num_online_cpus()/10;
5765 * It checks each scheduling domain to see if it is due to be balanced,
5766 * and initiates a balancing operation if so.
5768 * Balancing parameters are set up in init_sched_domains.
5770 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5772 int continue_balancing = 1;
5773 struct rq *rq = cpu_rq(cpu);
5774 unsigned long interval;
5775 struct sched_domain *sd;
5776 /* Earliest time when we have to do rebalance again */
5777 unsigned long next_balance = jiffies + 60*HZ;
5778 int update_next_balance = 0;
5779 int need_serialize, need_decay = 0;
5782 update_blocked_averages(cpu);
5785 for_each_domain(cpu, sd) {
5787 * Decay the newidle max times here because this is a regular
5788 * visit to all the domains. Decay ~1% per second.
5790 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5791 sd->max_newidle_lb_cost =
5792 (sd->max_newidle_lb_cost * 253) / 256;
5793 sd->next_decay_max_lb_cost = jiffies + HZ;
5796 max_cost += sd->max_newidle_lb_cost;
5798 if (!(sd->flags & SD_LOAD_BALANCE))
5802 * Stop the load balance at this level. There is another
5803 * CPU in our sched group which is doing load balancing more
5806 if (!continue_balancing) {
5812 interval = sd->balance_interval;
5813 if (idle != CPU_IDLE)
5814 interval *= sd->busy_factor;
5816 /* scale ms to jiffies */
5817 interval = msecs_to_jiffies(interval);
5818 interval = clamp(interval, 1UL, max_load_balance_interval);
5820 need_serialize = sd->flags & SD_SERIALIZE;
5822 if (need_serialize) {
5823 if (!spin_trylock(&balancing))
5827 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5828 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5830 * The LBF_DST_PINNED logic could have changed
5831 * env->dst_cpu, so we can't know our idle
5832 * state even if we migrated tasks. Update it.
5834 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5836 sd->last_balance = jiffies;
5839 spin_unlock(&balancing);
5841 if (time_after(next_balance, sd->last_balance + interval)) {
5842 next_balance = sd->last_balance + interval;
5843 update_next_balance = 1;
5848 * Ensure the rq-wide value also decays but keep it at a
5849 * reasonable floor to avoid funnies with rq->avg_idle.
5851 rq->max_idle_balance_cost =
5852 max((u64)sysctl_sched_migration_cost, max_cost);
5857 * next_balance will be updated only when there is a need.
5858 * When the cpu is attached to null domain for ex, it will not be
5861 if (likely(update_next_balance))
5862 rq->next_balance = next_balance;
5865 #ifdef CONFIG_NO_HZ_COMMON
5867 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5868 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5870 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5872 struct rq *this_rq = cpu_rq(this_cpu);
5876 if (idle != CPU_IDLE ||
5877 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5880 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5881 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5885 * If this cpu gets work to do, stop the load balancing
5886 * work being done for other cpus. Next load
5887 * balancing owner will pick it up.
5892 rq = cpu_rq(balance_cpu);
5894 raw_spin_lock_irq(&rq->lock);
5895 update_rq_clock(rq);
5896 update_idle_cpu_load(rq);
5897 raw_spin_unlock_irq(&rq->lock);
5899 rebalance_domains(balance_cpu, CPU_IDLE);
5901 if (time_after(this_rq->next_balance, rq->next_balance))
5902 this_rq->next_balance = rq->next_balance;
5904 nohz.next_balance = this_rq->next_balance;
5906 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5910 * Current heuristic for kicking the idle load balancer in the presence
5911 * of an idle cpu is the system.
5912 * - This rq has more than one task.
5913 * - At any scheduler domain level, this cpu's scheduler group has multiple
5914 * busy cpu's exceeding the group's power.
5915 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5916 * domain span are idle.
5918 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5920 unsigned long now = jiffies;
5921 struct sched_domain *sd;
5923 if (unlikely(idle_cpu(cpu)))
5927 * We may be recently in ticked or tickless idle mode. At the first
5928 * busy tick after returning from idle, we will update the busy stats.
5930 set_cpu_sd_state_busy();
5931 nohz_balance_exit_idle(cpu);
5934 * None are in tickless mode and hence no need for NOHZ idle load
5937 if (likely(!atomic_read(&nohz.nr_cpus)))
5940 if (time_before(now, nohz.next_balance))
5943 if (rq->nr_running >= 2)
5947 for_each_domain(cpu, sd) {
5948 struct sched_group *sg = sd->groups;
5949 struct sched_group_power *sgp = sg->sgp;
5950 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5952 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5953 goto need_kick_unlock;
5955 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5956 && (cpumask_first_and(nohz.idle_cpus_mask,
5957 sched_domain_span(sd)) < cpu))
5958 goto need_kick_unlock;
5960 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5972 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5976 * run_rebalance_domains is triggered when needed from the scheduler tick.
5977 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5979 static void run_rebalance_domains(struct softirq_action *h)
5981 int this_cpu = smp_processor_id();
5982 struct rq *this_rq = cpu_rq(this_cpu);
5983 enum cpu_idle_type idle = this_rq->idle_balance ?
5984 CPU_IDLE : CPU_NOT_IDLE;
5986 rebalance_domains(this_cpu, idle);
5989 * If this cpu has a pending nohz_balance_kick, then do the
5990 * balancing on behalf of the other idle cpus whose ticks are
5993 nohz_idle_balance(this_cpu, idle);
5996 static inline int on_null_domain(int cpu)
5998 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6002 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6004 void trigger_load_balance(struct rq *rq, int cpu)
6006 /* Don't need to rebalance while attached to NULL domain */
6007 if (time_after_eq(jiffies, rq->next_balance) &&
6008 likely(!on_null_domain(cpu)))
6009 raise_softirq(SCHED_SOFTIRQ);
6010 #ifdef CONFIG_NO_HZ_COMMON
6011 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6012 nohz_balancer_kick(cpu);
6016 static void rq_online_fair(struct rq *rq)
6021 static void rq_offline_fair(struct rq *rq)
6025 /* Ensure any throttled groups are reachable by pick_next_task */
6026 unthrottle_offline_cfs_rqs(rq);
6029 #endif /* CONFIG_SMP */
6032 * scheduler tick hitting a task of our scheduling class:
6034 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6036 struct cfs_rq *cfs_rq;
6037 struct sched_entity *se = &curr->se;
6039 for_each_sched_entity(se) {
6040 cfs_rq = cfs_rq_of(se);
6041 entity_tick(cfs_rq, se, queued);
6044 if (numabalancing_enabled)
6045 task_tick_numa(rq, curr);
6047 update_rq_runnable_avg(rq, 1);
6051 * called on fork with the child task as argument from the parent's context
6052 * - child not yet on the tasklist
6053 * - preemption disabled
6055 static void task_fork_fair(struct task_struct *p)
6057 struct cfs_rq *cfs_rq;
6058 struct sched_entity *se = &p->se, *curr;
6059 int this_cpu = smp_processor_id();
6060 struct rq *rq = this_rq();
6061 unsigned long flags;
6063 raw_spin_lock_irqsave(&rq->lock, flags);
6065 update_rq_clock(rq);
6067 cfs_rq = task_cfs_rq(current);
6068 curr = cfs_rq->curr;
6071 * Not only the cpu but also the task_group of the parent might have
6072 * been changed after parent->se.parent,cfs_rq were copied to
6073 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6074 * of child point to valid ones.
6077 __set_task_cpu(p, this_cpu);
6080 update_curr(cfs_rq);
6083 se->vruntime = curr->vruntime;
6084 place_entity(cfs_rq, se, 1);
6086 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6088 * Upon rescheduling, sched_class::put_prev_task() will place
6089 * 'current' within the tree based on its new key value.
6091 swap(curr->vruntime, se->vruntime);
6092 resched_task(rq->curr);
6095 se->vruntime -= cfs_rq->min_vruntime;
6097 raw_spin_unlock_irqrestore(&rq->lock, flags);
6101 * Priority of the task has changed. Check to see if we preempt
6105 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6111 * Reschedule if we are currently running on this runqueue and
6112 * our priority decreased, or if we are not currently running on
6113 * this runqueue and our priority is higher than the current's
6115 if (rq->curr == p) {
6116 if (p->prio > oldprio)
6117 resched_task(rq->curr);
6119 check_preempt_curr(rq, p, 0);
6122 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6124 struct sched_entity *se = &p->se;
6125 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6128 * Ensure the task's vruntime is normalized, so that when its
6129 * switched back to the fair class the enqueue_entity(.flags=0) will
6130 * do the right thing.
6132 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6133 * have normalized the vruntime, if it was !on_rq, then only when
6134 * the task is sleeping will it still have non-normalized vruntime.
6136 if (!se->on_rq && p->state != TASK_RUNNING) {
6138 * Fix up our vruntime so that the current sleep doesn't
6139 * cause 'unlimited' sleep bonus.
6141 place_entity(cfs_rq, se, 0);
6142 se->vruntime -= cfs_rq->min_vruntime;
6147 * Remove our load from contribution when we leave sched_fair
6148 * and ensure we don't carry in an old decay_count if we
6151 if (se->avg.decay_count) {
6152 __synchronize_entity_decay(se);
6153 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6159 * We switched to the sched_fair class.
6161 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6167 * We were most likely switched from sched_rt, so
6168 * kick off the schedule if running, otherwise just see
6169 * if we can still preempt the current task.
6172 resched_task(rq->curr);
6174 check_preempt_curr(rq, p, 0);
6177 /* Account for a task changing its policy or group.
6179 * This routine is mostly called to set cfs_rq->curr field when a task
6180 * migrates between groups/classes.
6182 static void set_curr_task_fair(struct rq *rq)
6184 struct sched_entity *se = &rq->curr->se;
6186 for_each_sched_entity(se) {
6187 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6189 set_next_entity(cfs_rq, se);
6190 /* ensure bandwidth has been allocated on our new cfs_rq */
6191 account_cfs_rq_runtime(cfs_rq, 0);
6195 void init_cfs_rq(struct cfs_rq *cfs_rq)
6197 cfs_rq->tasks_timeline = RB_ROOT;
6198 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6199 #ifndef CONFIG_64BIT
6200 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6203 atomic64_set(&cfs_rq->decay_counter, 1);
6204 atomic_long_set(&cfs_rq->removed_load, 0);
6208 #ifdef CONFIG_FAIR_GROUP_SCHED
6209 static void task_move_group_fair(struct task_struct *p, int on_rq)
6211 struct cfs_rq *cfs_rq;
6213 * If the task was not on the rq at the time of this cgroup movement
6214 * it must have been asleep, sleeping tasks keep their ->vruntime
6215 * absolute on their old rq until wakeup (needed for the fair sleeper
6216 * bonus in place_entity()).
6218 * If it was on the rq, we've just 'preempted' it, which does convert
6219 * ->vruntime to a relative base.
6221 * Make sure both cases convert their relative position when migrating
6222 * to another cgroup's rq. This does somewhat interfere with the
6223 * fair sleeper stuff for the first placement, but who cares.
6226 * When !on_rq, vruntime of the task has usually NOT been normalized.
6227 * But there are some cases where it has already been normalized:
6229 * - Moving a forked child which is waiting for being woken up by
6230 * wake_up_new_task().
6231 * - Moving a task which has been woken up by try_to_wake_up() and
6232 * waiting for actually being woken up by sched_ttwu_pending().
6234 * To prevent boost or penalty in the new cfs_rq caused by delta
6235 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6237 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6241 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6242 set_task_rq(p, task_cpu(p));
6244 cfs_rq = cfs_rq_of(&p->se);
6245 p->se.vruntime += cfs_rq->min_vruntime;
6248 * migrate_task_rq_fair() will have removed our previous
6249 * contribution, but we must synchronize for ongoing future
6252 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6253 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6258 void free_fair_sched_group(struct task_group *tg)
6262 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6264 for_each_possible_cpu(i) {
6266 kfree(tg->cfs_rq[i]);
6275 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6277 struct cfs_rq *cfs_rq;
6278 struct sched_entity *se;
6281 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6284 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6288 tg->shares = NICE_0_LOAD;
6290 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6292 for_each_possible_cpu(i) {
6293 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6294 GFP_KERNEL, cpu_to_node(i));
6298 se = kzalloc_node(sizeof(struct sched_entity),
6299 GFP_KERNEL, cpu_to_node(i));
6303 init_cfs_rq(cfs_rq);
6304 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6315 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6317 struct rq *rq = cpu_rq(cpu);
6318 unsigned long flags;
6321 * Only empty task groups can be destroyed; so we can speculatively
6322 * check on_list without danger of it being re-added.
6324 if (!tg->cfs_rq[cpu]->on_list)
6327 raw_spin_lock_irqsave(&rq->lock, flags);
6328 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6329 raw_spin_unlock_irqrestore(&rq->lock, flags);
6332 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6333 struct sched_entity *se, int cpu,
6334 struct sched_entity *parent)
6336 struct rq *rq = cpu_rq(cpu);
6340 init_cfs_rq_runtime(cfs_rq);
6342 tg->cfs_rq[cpu] = cfs_rq;
6345 /* se could be NULL for root_task_group */
6350 se->cfs_rq = &rq->cfs;
6352 se->cfs_rq = parent->my_q;
6355 update_load_set(&se->load, 0);
6356 se->parent = parent;
6359 static DEFINE_MUTEX(shares_mutex);
6361 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6364 unsigned long flags;
6367 * We can't change the weight of the root cgroup.
6372 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6374 mutex_lock(&shares_mutex);
6375 if (tg->shares == shares)
6378 tg->shares = shares;
6379 for_each_possible_cpu(i) {
6380 struct rq *rq = cpu_rq(i);
6381 struct sched_entity *se;
6384 /* Propagate contribution to hierarchy */
6385 raw_spin_lock_irqsave(&rq->lock, flags);
6387 /* Possible calls to update_curr() need rq clock */
6388 update_rq_clock(rq);
6389 for_each_sched_entity(se)
6390 update_cfs_shares(group_cfs_rq(se));
6391 raw_spin_unlock_irqrestore(&rq->lock, flags);
6395 mutex_unlock(&shares_mutex);
6398 #else /* CONFIG_FAIR_GROUP_SCHED */
6400 void free_fair_sched_group(struct task_group *tg) { }
6402 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6407 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6409 #endif /* CONFIG_FAIR_GROUP_SCHED */
6412 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6414 struct sched_entity *se = &task->se;
6415 unsigned int rr_interval = 0;
6418 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6421 if (rq->cfs.load.weight)
6422 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6428 * All the scheduling class methods:
6430 const struct sched_class fair_sched_class = {
6431 .next = &idle_sched_class,
6432 .enqueue_task = enqueue_task_fair,
6433 .dequeue_task = dequeue_task_fair,
6434 .yield_task = yield_task_fair,
6435 .yield_to_task = yield_to_task_fair,
6437 .check_preempt_curr = check_preempt_wakeup,
6439 .pick_next_task = pick_next_task_fair,
6440 .put_prev_task = put_prev_task_fair,
6443 .select_task_rq = select_task_rq_fair,
6444 .migrate_task_rq = migrate_task_rq_fair,
6446 .rq_online = rq_online_fair,
6447 .rq_offline = rq_offline_fair,
6449 .task_waking = task_waking_fair,
6452 .set_curr_task = set_curr_task_fair,
6453 .task_tick = task_tick_fair,
6454 .task_fork = task_fork_fair,
6456 .prio_changed = prio_changed_fair,
6457 .switched_from = switched_from_fair,
6458 .switched_to = switched_to_fair,
6460 .get_rr_interval = get_rr_interval_fair,
6462 #ifdef CONFIG_FAIR_GROUP_SCHED
6463 .task_move_group = task_move_group_fair,
6467 #ifdef CONFIG_SCHED_DEBUG
6468 void print_cfs_stats(struct seq_file *m, int cpu)
6470 struct cfs_rq *cfs_rq;
6473 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6474 print_cfs_rq(m, cpu, cfs_rq);
6479 __init void init_sched_fair_class(void)
6482 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6484 #ifdef CONFIG_NO_HZ_COMMON
6485 nohz.next_balance = jiffies;
6486 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6487 cpu_notifier(sched_ilb_notifier, 0);