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);
881 * Once a preferred node is selected the scheduler balancer will prefer moving
882 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
883 * scans. This will give the process the chance to accumulate more faults on
884 * the preferred node but still allow the scheduler to move the task again if
885 * the nodes CPUs are overloaded.
887 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 3;
889 static inline int task_faults_idx(int nid, int priv)
891 return 2 * nid + priv;
894 static inline unsigned long task_faults(struct task_struct *p, int nid)
899 return p->numa_faults[task_faults_idx(nid, 0)] +
900 p->numa_faults[task_faults_idx(nid, 1)];
903 static unsigned long weighted_cpuload(const int cpu);
907 find_idlest_cpu_node(int this_cpu, int nid)
909 unsigned long load, min_load = ULONG_MAX;
910 int i, idlest_cpu = this_cpu;
912 BUG_ON(cpu_to_node(this_cpu) == nid);
915 for_each_cpu(i, cpumask_of_node(nid)) {
916 load = weighted_cpuload(i);
918 if (load < min_load) {
928 static void task_numa_placement(struct task_struct *p)
930 int seq, nid, max_nid = -1;
931 unsigned long max_faults = 0;
933 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
934 if (p->numa_scan_seq == seq)
936 p->numa_scan_seq = seq;
937 p->numa_migrate_seq++;
938 p->numa_scan_period_max = task_scan_max(p);
940 /* Find the node with the highest number of faults */
941 for_each_online_node(nid) {
942 unsigned long faults;
945 for (priv = 0; priv < 2; priv++) {
946 i = task_faults_idx(nid, priv);
948 /* Decay existing window, copy faults since last scan */
949 p->numa_faults[i] >>= 1;
950 p->numa_faults[i] += p->numa_faults_buffer[i];
951 p->numa_faults_buffer[i] = 0;
954 /* Find maximum private faults */
955 faults = p->numa_faults[task_faults_idx(nid, 1)];
956 if (faults > max_faults) {
963 * Record the preferred node as the node with the most faults,
964 * requeue the task to be running on the idlest CPU on the
965 * preferred node and reset the scanning rate to recheck
966 * the working set placement.
968 if (max_faults && max_nid != p->numa_preferred_nid) {
972 * If the task is not on the preferred node then find the most
973 * idle CPU to migrate to.
975 preferred_cpu = task_cpu(p);
976 if (cpu_to_node(preferred_cpu) != max_nid) {
977 preferred_cpu = find_idlest_cpu_node(preferred_cpu,
981 /* Update the preferred nid and migrate task if possible */
982 p->numa_preferred_nid = max_nid;
983 p->numa_migrate_seq = 0;
984 migrate_task_to(p, preferred_cpu);
989 * Got a PROT_NONE fault for a page on @node.
991 void task_numa_fault(int last_nid, int node, int pages, bool migrated)
993 struct task_struct *p = current;
996 if (!numabalancing_enabled)
999 /* for example, ksmd faulting in a user's mm */
1003 /* For now, do not attempt to detect private/shared accesses */
1006 /* Allocate buffer to track faults on a per-node basis */
1007 if (unlikely(!p->numa_faults)) {
1008 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1010 /* numa_faults and numa_faults_buffer share the allocation */
1011 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1012 if (!p->numa_faults)
1015 BUG_ON(p->numa_faults_buffer);
1016 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1020 * If pages are properly placed (did not migrate) then scan slower.
1021 * This is reset periodically in case of phase changes
1024 /* Initialise if necessary */
1025 if (!p->numa_scan_period_max)
1026 p->numa_scan_period_max = task_scan_max(p);
1028 p->numa_scan_period = min(p->numa_scan_period_max,
1029 p->numa_scan_period + 10);
1032 task_numa_placement(p);
1034 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1037 static void reset_ptenuma_scan(struct task_struct *p)
1039 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1040 p->mm->numa_scan_offset = 0;
1044 * The expensive part of numa migration is done from task_work context.
1045 * Triggered from task_tick_numa().
1047 void task_numa_work(struct callback_head *work)
1049 unsigned long migrate, next_scan, now = jiffies;
1050 struct task_struct *p = current;
1051 struct mm_struct *mm = p->mm;
1052 struct vm_area_struct *vma;
1053 unsigned long start, end;
1054 unsigned long nr_pte_updates = 0;
1057 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1059 work->next = work; /* protect against double add */
1061 * Who cares about NUMA placement when they're dying.
1063 * NOTE: make sure not to dereference p->mm before this check,
1064 * exit_task_work() happens _after_ exit_mm() so we could be called
1065 * without p->mm even though we still had it when we enqueued this
1068 if (p->flags & PF_EXITING)
1071 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1072 mm->numa_next_scan = now +
1073 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1074 mm->numa_next_reset = now +
1075 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1079 * Reset the scan period if enough time has gone by. Objective is that
1080 * scanning will be reduced if pages are properly placed. As tasks
1081 * can enter different phases this needs to be re-examined. Lacking
1082 * proper tracking of reference behaviour, this blunt hammer is used.
1084 migrate = mm->numa_next_reset;
1085 if (time_after(now, migrate)) {
1086 p->numa_scan_period = task_scan_min(p);
1087 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1088 xchg(&mm->numa_next_reset, next_scan);
1092 * Enforce maximal scan/migration frequency..
1094 migrate = mm->numa_next_scan;
1095 if (time_before(now, migrate))
1098 if (p->numa_scan_period == 0) {
1099 p->numa_scan_period_max = task_scan_max(p);
1100 p->numa_scan_period = task_scan_min(p);
1103 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1104 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1108 * Delay this task enough that another task of this mm will likely win
1109 * the next time around.
1111 p->node_stamp += 2 * TICK_NSEC;
1113 start = mm->numa_scan_offset;
1114 pages = sysctl_numa_balancing_scan_size;
1115 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1119 down_read(&mm->mmap_sem);
1120 vma = find_vma(mm, start);
1122 reset_ptenuma_scan(p);
1126 for (; vma; vma = vma->vm_next) {
1127 if (!vma_migratable(vma))
1131 start = max(start, vma->vm_start);
1132 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1133 end = min(end, vma->vm_end);
1134 nr_pte_updates += change_prot_numa(vma, start, end);
1137 * Scan sysctl_numa_balancing_scan_size but ensure that
1138 * at least one PTE is updated so that unused virtual
1139 * address space is quickly skipped.
1142 pages -= (end - start) >> PAGE_SHIFT;
1147 } while (end != vma->vm_end);
1152 * If the whole process was scanned without updates then no NUMA
1153 * hinting faults are being recorded and scan rate should be lower.
1155 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1156 p->numa_scan_period = min(p->numa_scan_period_max,
1157 p->numa_scan_period << 1);
1159 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1160 mm->numa_next_scan = next_scan;
1164 * It is possible to reach the end of the VMA list but the last few
1165 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1166 * would find the !migratable VMA on the next scan but not reset the
1167 * scanner to the start so check it now.
1170 mm->numa_scan_offset = start;
1172 reset_ptenuma_scan(p);
1173 up_read(&mm->mmap_sem);
1177 * Drive the periodic memory faults..
1179 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1181 struct callback_head *work = &curr->numa_work;
1185 * We don't care about NUMA placement if we don't have memory.
1187 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1191 * Using runtime rather than walltime has the dual advantage that
1192 * we (mostly) drive the selection from busy threads and that the
1193 * task needs to have done some actual work before we bother with
1196 now = curr->se.sum_exec_runtime;
1197 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1199 if (now - curr->node_stamp > period) {
1200 if (!curr->node_stamp)
1201 curr->numa_scan_period = task_scan_min(curr);
1202 curr->node_stamp += period;
1204 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1205 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1206 task_work_add(curr, work, true);
1211 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1214 #endif /* CONFIG_NUMA_BALANCING */
1217 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1219 update_load_add(&cfs_rq->load, se->load.weight);
1220 if (!parent_entity(se))
1221 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1223 if (entity_is_task(se))
1224 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1226 cfs_rq->nr_running++;
1230 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1232 update_load_sub(&cfs_rq->load, se->load.weight);
1233 if (!parent_entity(se))
1234 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1235 if (entity_is_task(se))
1236 list_del_init(&se->group_node);
1237 cfs_rq->nr_running--;
1240 #ifdef CONFIG_FAIR_GROUP_SCHED
1242 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1247 * Use this CPU's actual weight instead of the last load_contribution
1248 * to gain a more accurate current total weight. See
1249 * update_cfs_rq_load_contribution().
1251 tg_weight = atomic_long_read(&tg->load_avg);
1252 tg_weight -= cfs_rq->tg_load_contrib;
1253 tg_weight += cfs_rq->load.weight;
1258 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1260 long tg_weight, load, shares;
1262 tg_weight = calc_tg_weight(tg, cfs_rq);
1263 load = cfs_rq->load.weight;
1265 shares = (tg->shares * load);
1267 shares /= tg_weight;
1269 if (shares < MIN_SHARES)
1270 shares = MIN_SHARES;
1271 if (shares > tg->shares)
1272 shares = tg->shares;
1276 # else /* CONFIG_SMP */
1277 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1281 # endif /* CONFIG_SMP */
1282 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1283 unsigned long weight)
1286 /* commit outstanding execution time */
1287 if (cfs_rq->curr == se)
1288 update_curr(cfs_rq);
1289 account_entity_dequeue(cfs_rq, se);
1292 update_load_set(&se->load, weight);
1295 account_entity_enqueue(cfs_rq, se);
1298 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1300 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1302 struct task_group *tg;
1303 struct sched_entity *se;
1307 se = tg->se[cpu_of(rq_of(cfs_rq))];
1308 if (!se || throttled_hierarchy(cfs_rq))
1311 if (likely(se->load.weight == tg->shares))
1314 shares = calc_cfs_shares(cfs_rq, tg);
1316 reweight_entity(cfs_rq_of(se), se, shares);
1318 #else /* CONFIG_FAIR_GROUP_SCHED */
1319 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1322 #endif /* CONFIG_FAIR_GROUP_SCHED */
1326 * We choose a half-life close to 1 scheduling period.
1327 * Note: The tables below are dependent on this value.
1329 #define LOAD_AVG_PERIOD 32
1330 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1331 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1333 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1334 static const u32 runnable_avg_yN_inv[] = {
1335 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1336 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1337 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1338 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1339 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1340 0x85aac367, 0x82cd8698,
1344 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1345 * over-estimates when re-combining.
1347 static const u32 runnable_avg_yN_sum[] = {
1348 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1349 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1350 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1355 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1357 static __always_inline u64 decay_load(u64 val, u64 n)
1359 unsigned int local_n;
1363 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1366 /* after bounds checking we can collapse to 32-bit */
1370 * As y^PERIOD = 1/2, we can combine
1371 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1372 * With a look-up table which covers k^n (n<PERIOD)
1374 * To achieve constant time decay_load.
1376 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1377 val >>= local_n / LOAD_AVG_PERIOD;
1378 local_n %= LOAD_AVG_PERIOD;
1381 val *= runnable_avg_yN_inv[local_n];
1382 /* We don't use SRR here since we always want to round down. */
1387 * For updates fully spanning n periods, the contribution to runnable
1388 * average will be: \Sum 1024*y^n
1390 * We can compute this reasonably efficiently by combining:
1391 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1393 static u32 __compute_runnable_contrib(u64 n)
1397 if (likely(n <= LOAD_AVG_PERIOD))
1398 return runnable_avg_yN_sum[n];
1399 else if (unlikely(n >= LOAD_AVG_MAX_N))
1400 return LOAD_AVG_MAX;
1402 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1404 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1405 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1407 n -= LOAD_AVG_PERIOD;
1408 } while (n > LOAD_AVG_PERIOD);
1410 contrib = decay_load(contrib, n);
1411 return contrib + runnable_avg_yN_sum[n];
1415 * We can represent the historical contribution to runnable average as the
1416 * coefficients of a geometric series. To do this we sub-divide our runnable
1417 * history into segments of approximately 1ms (1024us); label the segment that
1418 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1420 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1422 * (now) (~1ms ago) (~2ms ago)
1424 * Let u_i denote the fraction of p_i that the entity was runnable.
1426 * We then designate the fractions u_i as our co-efficients, yielding the
1427 * following representation of historical load:
1428 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1430 * We choose y based on the with of a reasonably scheduling period, fixing:
1433 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1434 * approximately half as much as the contribution to load within the last ms
1437 * When a period "rolls over" and we have new u_0`, multiplying the previous
1438 * sum again by y is sufficient to update:
1439 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1440 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1442 static __always_inline int __update_entity_runnable_avg(u64 now,
1443 struct sched_avg *sa,
1447 u32 runnable_contrib;
1448 int delta_w, decayed = 0;
1450 delta = now - sa->last_runnable_update;
1452 * This should only happen when time goes backwards, which it
1453 * unfortunately does during sched clock init when we swap over to TSC.
1455 if ((s64)delta < 0) {
1456 sa->last_runnable_update = now;
1461 * Use 1024ns as the unit of measurement since it's a reasonable
1462 * approximation of 1us and fast to compute.
1467 sa->last_runnable_update = now;
1469 /* delta_w is the amount already accumulated against our next period */
1470 delta_w = sa->runnable_avg_period % 1024;
1471 if (delta + delta_w >= 1024) {
1472 /* period roll-over */
1476 * Now that we know we're crossing a period boundary, figure
1477 * out how much from delta we need to complete the current
1478 * period and accrue it.
1480 delta_w = 1024 - delta_w;
1482 sa->runnable_avg_sum += delta_w;
1483 sa->runnable_avg_period += delta_w;
1487 /* Figure out how many additional periods this update spans */
1488 periods = delta / 1024;
1491 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1493 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1496 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1497 runnable_contrib = __compute_runnable_contrib(periods);
1499 sa->runnable_avg_sum += runnable_contrib;
1500 sa->runnable_avg_period += runnable_contrib;
1503 /* Remainder of delta accrued against u_0` */
1505 sa->runnable_avg_sum += delta;
1506 sa->runnable_avg_period += delta;
1511 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1512 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1514 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1515 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1517 decays -= se->avg.decay_count;
1521 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1522 se->avg.decay_count = 0;
1527 #ifdef CONFIG_FAIR_GROUP_SCHED
1528 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1531 struct task_group *tg = cfs_rq->tg;
1534 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1535 tg_contrib -= cfs_rq->tg_load_contrib;
1537 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1538 atomic_long_add(tg_contrib, &tg->load_avg);
1539 cfs_rq->tg_load_contrib += tg_contrib;
1544 * Aggregate cfs_rq runnable averages into an equivalent task_group
1545 * representation for computing load contributions.
1547 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1548 struct cfs_rq *cfs_rq)
1550 struct task_group *tg = cfs_rq->tg;
1553 /* The fraction of a cpu used by this cfs_rq */
1554 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1555 sa->runnable_avg_period + 1);
1556 contrib -= cfs_rq->tg_runnable_contrib;
1558 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1559 atomic_add(contrib, &tg->runnable_avg);
1560 cfs_rq->tg_runnable_contrib += contrib;
1564 static inline void __update_group_entity_contrib(struct sched_entity *se)
1566 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1567 struct task_group *tg = cfs_rq->tg;
1572 contrib = cfs_rq->tg_load_contrib * tg->shares;
1573 se->avg.load_avg_contrib = div_u64(contrib,
1574 atomic_long_read(&tg->load_avg) + 1);
1577 * For group entities we need to compute a correction term in the case
1578 * that they are consuming <1 cpu so that we would contribute the same
1579 * load as a task of equal weight.
1581 * Explicitly co-ordinating this measurement would be expensive, but
1582 * fortunately the sum of each cpus contribution forms a usable
1583 * lower-bound on the true value.
1585 * Consider the aggregate of 2 contributions. Either they are disjoint
1586 * (and the sum represents true value) or they are disjoint and we are
1587 * understating by the aggregate of their overlap.
1589 * Extending this to N cpus, for a given overlap, the maximum amount we
1590 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1591 * cpus that overlap for this interval and w_i is the interval width.
1593 * On a small machine; the first term is well-bounded which bounds the
1594 * total error since w_i is a subset of the period. Whereas on a
1595 * larger machine, while this first term can be larger, if w_i is the
1596 * of consequential size guaranteed to see n_i*w_i quickly converge to
1597 * our upper bound of 1-cpu.
1599 runnable_avg = atomic_read(&tg->runnable_avg);
1600 if (runnable_avg < NICE_0_LOAD) {
1601 se->avg.load_avg_contrib *= runnable_avg;
1602 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1606 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1607 int force_update) {}
1608 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1609 struct cfs_rq *cfs_rq) {}
1610 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1613 static inline void __update_task_entity_contrib(struct sched_entity *se)
1617 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1618 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1619 contrib /= (se->avg.runnable_avg_period + 1);
1620 se->avg.load_avg_contrib = scale_load(contrib);
1623 /* Compute the current contribution to load_avg by se, return any delta */
1624 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1626 long old_contrib = se->avg.load_avg_contrib;
1628 if (entity_is_task(se)) {
1629 __update_task_entity_contrib(se);
1631 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1632 __update_group_entity_contrib(se);
1635 return se->avg.load_avg_contrib - old_contrib;
1638 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1641 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1642 cfs_rq->blocked_load_avg -= load_contrib;
1644 cfs_rq->blocked_load_avg = 0;
1647 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1649 /* Update a sched_entity's runnable average */
1650 static inline void update_entity_load_avg(struct sched_entity *se,
1653 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1658 * For a group entity we need to use their owned cfs_rq_clock_task() in
1659 * case they are the parent of a throttled hierarchy.
1661 if (entity_is_task(se))
1662 now = cfs_rq_clock_task(cfs_rq);
1664 now = cfs_rq_clock_task(group_cfs_rq(se));
1666 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1669 contrib_delta = __update_entity_load_avg_contrib(se);
1675 cfs_rq->runnable_load_avg += contrib_delta;
1677 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1681 * Decay the load contributed by all blocked children and account this so that
1682 * their contribution may appropriately discounted when they wake up.
1684 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1686 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1689 decays = now - cfs_rq->last_decay;
1690 if (!decays && !force_update)
1693 if (atomic_long_read(&cfs_rq->removed_load)) {
1694 unsigned long removed_load;
1695 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1696 subtract_blocked_load_contrib(cfs_rq, removed_load);
1700 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1702 atomic64_add(decays, &cfs_rq->decay_counter);
1703 cfs_rq->last_decay = now;
1706 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1709 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1711 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1712 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1715 /* Add the load generated by se into cfs_rq's child load-average */
1716 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1717 struct sched_entity *se,
1721 * We track migrations using entity decay_count <= 0, on a wake-up
1722 * migration we use a negative decay count to track the remote decays
1723 * accumulated while sleeping.
1725 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1726 * are seen by enqueue_entity_load_avg() as a migration with an already
1727 * constructed load_avg_contrib.
1729 if (unlikely(se->avg.decay_count <= 0)) {
1730 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1731 if (se->avg.decay_count) {
1733 * In a wake-up migration we have to approximate the
1734 * time sleeping. This is because we can't synchronize
1735 * clock_task between the two cpus, and it is not
1736 * guaranteed to be read-safe. Instead, we can
1737 * approximate this using our carried decays, which are
1738 * explicitly atomically readable.
1740 se->avg.last_runnable_update -= (-se->avg.decay_count)
1742 update_entity_load_avg(se, 0);
1743 /* Indicate that we're now synchronized and on-rq */
1744 se->avg.decay_count = 0;
1749 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1750 * would have made count negative); we must be careful to avoid
1751 * double-accounting blocked time after synchronizing decays.
1753 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1757 /* migrated tasks did not contribute to our blocked load */
1759 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1760 update_entity_load_avg(se, 0);
1763 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1764 /* we force update consideration on load-balancer moves */
1765 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1769 * Remove se's load from this cfs_rq child load-average, if the entity is
1770 * transitioning to a blocked state we track its projected decay using
1773 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1774 struct sched_entity *se,
1777 update_entity_load_avg(se, 1);
1778 /* we force update consideration on load-balancer moves */
1779 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1781 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1783 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1784 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1785 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1789 * Update the rq's load with the elapsed running time before entering
1790 * idle. if the last scheduled task is not a CFS task, idle_enter will
1791 * be the only way to update the runnable statistic.
1793 void idle_enter_fair(struct rq *this_rq)
1795 update_rq_runnable_avg(this_rq, 1);
1799 * Update the rq's load with the elapsed idle time before a task is
1800 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1801 * be the only way to update the runnable statistic.
1803 void idle_exit_fair(struct rq *this_rq)
1805 update_rq_runnable_avg(this_rq, 0);
1809 static inline void update_entity_load_avg(struct sched_entity *se,
1810 int update_cfs_rq) {}
1811 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1812 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1813 struct sched_entity *se,
1815 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1816 struct sched_entity *se,
1818 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1819 int force_update) {}
1822 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1824 #ifdef CONFIG_SCHEDSTATS
1825 struct task_struct *tsk = NULL;
1827 if (entity_is_task(se))
1830 if (se->statistics.sleep_start) {
1831 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1836 if (unlikely(delta > se->statistics.sleep_max))
1837 se->statistics.sleep_max = delta;
1839 se->statistics.sleep_start = 0;
1840 se->statistics.sum_sleep_runtime += delta;
1843 account_scheduler_latency(tsk, delta >> 10, 1);
1844 trace_sched_stat_sleep(tsk, delta);
1847 if (se->statistics.block_start) {
1848 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1853 if (unlikely(delta > se->statistics.block_max))
1854 se->statistics.block_max = delta;
1856 se->statistics.block_start = 0;
1857 se->statistics.sum_sleep_runtime += delta;
1860 if (tsk->in_iowait) {
1861 se->statistics.iowait_sum += delta;
1862 se->statistics.iowait_count++;
1863 trace_sched_stat_iowait(tsk, delta);
1866 trace_sched_stat_blocked(tsk, delta);
1869 * Blocking time is in units of nanosecs, so shift by
1870 * 20 to get a milliseconds-range estimation of the
1871 * amount of time that the task spent sleeping:
1873 if (unlikely(prof_on == SLEEP_PROFILING)) {
1874 profile_hits(SLEEP_PROFILING,
1875 (void *)get_wchan(tsk),
1878 account_scheduler_latency(tsk, delta >> 10, 0);
1884 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1886 #ifdef CONFIG_SCHED_DEBUG
1887 s64 d = se->vruntime - cfs_rq->min_vruntime;
1892 if (d > 3*sysctl_sched_latency)
1893 schedstat_inc(cfs_rq, nr_spread_over);
1898 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1900 u64 vruntime = cfs_rq->min_vruntime;
1903 * The 'current' period is already promised to the current tasks,
1904 * however the extra weight of the new task will slow them down a
1905 * little, place the new task so that it fits in the slot that
1906 * stays open at the end.
1908 if (initial && sched_feat(START_DEBIT))
1909 vruntime += sched_vslice(cfs_rq, se);
1911 /* sleeps up to a single latency don't count. */
1913 unsigned long thresh = sysctl_sched_latency;
1916 * Halve their sleep time's effect, to allow
1917 * for a gentler effect of sleepers:
1919 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1925 /* ensure we never gain time by being placed backwards. */
1926 se->vruntime = max_vruntime(se->vruntime, vruntime);
1929 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1932 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1935 * Update the normalized vruntime before updating min_vruntime
1936 * through calling update_curr().
1938 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1939 se->vruntime += cfs_rq->min_vruntime;
1942 * Update run-time statistics of the 'current'.
1944 update_curr(cfs_rq);
1945 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1946 account_entity_enqueue(cfs_rq, se);
1947 update_cfs_shares(cfs_rq);
1949 if (flags & ENQUEUE_WAKEUP) {
1950 place_entity(cfs_rq, se, 0);
1951 enqueue_sleeper(cfs_rq, se);
1954 update_stats_enqueue(cfs_rq, se);
1955 check_spread(cfs_rq, se);
1956 if (se != cfs_rq->curr)
1957 __enqueue_entity(cfs_rq, se);
1960 if (cfs_rq->nr_running == 1) {
1961 list_add_leaf_cfs_rq(cfs_rq);
1962 check_enqueue_throttle(cfs_rq);
1966 static void __clear_buddies_last(struct sched_entity *se)
1968 for_each_sched_entity(se) {
1969 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1970 if (cfs_rq->last == se)
1971 cfs_rq->last = NULL;
1977 static void __clear_buddies_next(struct sched_entity *se)
1979 for_each_sched_entity(se) {
1980 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1981 if (cfs_rq->next == se)
1982 cfs_rq->next = NULL;
1988 static void __clear_buddies_skip(struct sched_entity *se)
1990 for_each_sched_entity(se) {
1991 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1992 if (cfs_rq->skip == se)
1993 cfs_rq->skip = NULL;
1999 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2001 if (cfs_rq->last == se)
2002 __clear_buddies_last(se);
2004 if (cfs_rq->next == se)
2005 __clear_buddies_next(se);
2007 if (cfs_rq->skip == se)
2008 __clear_buddies_skip(se);
2011 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2014 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2017 * Update run-time statistics of the 'current'.
2019 update_curr(cfs_rq);
2020 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2022 update_stats_dequeue(cfs_rq, se);
2023 if (flags & DEQUEUE_SLEEP) {
2024 #ifdef CONFIG_SCHEDSTATS
2025 if (entity_is_task(se)) {
2026 struct task_struct *tsk = task_of(se);
2028 if (tsk->state & TASK_INTERRUPTIBLE)
2029 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2030 if (tsk->state & TASK_UNINTERRUPTIBLE)
2031 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2036 clear_buddies(cfs_rq, se);
2038 if (se != cfs_rq->curr)
2039 __dequeue_entity(cfs_rq, se);
2041 account_entity_dequeue(cfs_rq, se);
2044 * Normalize the entity after updating the min_vruntime because the
2045 * update can refer to the ->curr item and we need to reflect this
2046 * movement in our normalized position.
2048 if (!(flags & DEQUEUE_SLEEP))
2049 se->vruntime -= cfs_rq->min_vruntime;
2051 /* return excess runtime on last dequeue */
2052 return_cfs_rq_runtime(cfs_rq);
2054 update_min_vruntime(cfs_rq);
2055 update_cfs_shares(cfs_rq);
2059 * Preempt the current task with a newly woken task if needed:
2062 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2064 unsigned long ideal_runtime, delta_exec;
2065 struct sched_entity *se;
2068 ideal_runtime = sched_slice(cfs_rq, curr);
2069 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2070 if (delta_exec > ideal_runtime) {
2071 resched_task(rq_of(cfs_rq)->curr);
2073 * The current task ran long enough, ensure it doesn't get
2074 * re-elected due to buddy favours.
2076 clear_buddies(cfs_rq, curr);
2081 * Ensure that a task that missed wakeup preemption by a
2082 * narrow margin doesn't have to wait for a full slice.
2083 * This also mitigates buddy induced latencies under load.
2085 if (delta_exec < sysctl_sched_min_granularity)
2088 se = __pick_first_entity(cfs_rq);
2089 delta = curr->vruntime - se->vruntime;
2094 if (delta > ideal_runtime)
2095 resched_task(rq_of(cfs_rq)->curr);
2099 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2101 /* 'current' is not kept within the tree. */
2104 * Any task has to be enqueued before it get to execute on
2105 * a CPU. So account for the time it spent waiting on the
2108 update_stats_wait_end(cfs_rq, se);
2109 __dequeue_entity(cfs_rq, se);
2112 update_stats_curr_start(cfs_rq, se);
2114 #ifdef CONFIG_SCHEDSTATS
2116 * Track our maximum slice length, if the CPU's load is at
2117 * least twice that of our own weight (i.e. dont track it
2118 * when there are only lesser-weight tasks around):
2120 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2121 se->statistics.slice_max = max(se->statistics.slice_max,
2122 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2125 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2129 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2132 * Pick the next process, keeping these things in mind, in this order:
2133 * 1) keep things fair between processes/task groups
2134 * 2) pick the "next" process, since someone really wants that to run
2135 * 3) pick the "last" process, for cache locality
2136 * 4) do not run the "skip" process, if something else is available
2138 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2140 struct sched_entity *se = __pick_first_entity(cfs_rq);
2141 struct sched_entity *left = se;
2144 * Avoid running the skip buddy, if running something else can
2145 * be done without getting too unfair.
2147 if (cfs_rq->skip == se) {
2148 struct sched_entity *second = __pick_next_entity(se);
2149 if (second && wakeup_preempt_entity(second, left) < 1)
2154 * Prefer last buddy, try to return the CPU to a preempted task.
2156 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2160 * Someone really wants this to run. If it's not unfair, run it.
2162 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2165 clear_buddies(cfs_rq, se);
2170 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2172 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2175 * If still on the runqueue then deactivate_task()
2176 * was not called and update_curr() has to be done:
2179 update_curr(cfs_rq);
2181 /* throttle cfs_rqs exceeding runtime */
2182 check_cfs_rq_runtime(cfs_rq);
2184 check_spread(cfs_rq, prev);
2186 update_stats_wait_start(cfs_rq, prev);
2187 /* Put 'current' back into the tree. */
2188 __enqueue_entity(cfs_rq, prev);
2189 /* in !on_rq case, update occurred at dequeue */
2190 update_entity_load_avg(prev, 1);
2192 cfs_rq->curr = NULL;
2196 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2199 * Update run-time statistics of the 'current'.
2201 update_curr(cfs_rq);
2204 * Ensure that runnable average is periodically updated.
2206 update_entity_load_avg(curr, 1);
2207 update_cfs_rq_blocked_load(cfs_rq, 1);
2208 update_cfs_shares(cfs_rq);
2210 #ifdef CONFIG_SCHED_HRTICK
2212 * queued ticks are scheduled to match the slice, so don't bother
2213 * validating it and just reschedule.
2216 resched_task(rq_of(cfs_rq)->curr);
2220 * don't let the period tick interfere with the hrtick preemption
2222 if (!sched_feat(DOUBLE_TICK) &&
2223 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2227 if (cfs_rq->nr_running > 1)
2228 check_preempt_tick(cfs_rq, curr);
2232 /**************************************************
2233 * CFS bandwidth control machinery
2236 #ifdef CONFIG_CFS_BANDWIDTH
2238 #ifdef HAVE_JUMP_LABEL
2239 static struct static_key __cfs_bandwidth_used;
2241 static inline bool cfs_bandwidth_used(void)
2243 return static_key_false(&__cfs_bandwidth_used);
2246 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2248 /* only need to count groups transitioning between enabled/!enabled */
2249 if (enabled && !was_enabled)
2250 static_key_slow_inc(&__cfs_bandwidth_used);
2251 else if (!enabled && was_enabled)
2252 static_key_slow_dec(&__cfs_bandwidth_used);
2254 #else /* HAVE_JUMP_LABEL */
2255 static bool cfs_bandwidth_used(void)
2260 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2261 #endif /* HAVE_JUMP_LABEL */
2264 * default period for cfs group bandwidth.
2265 * default: 0.1s, units: nanoseconds
2267 static inline u64 default_cfs_period(void)
2269 return 100000000ULL;
2272 static inline u64 sched_cfs_bandwidth_slice(void)
2274 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2278 * Replenish runtime according to assigned quota and update expiration time.
2279 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2280 * additional synchronization around rq->lock.
2282 * requires cfs_b->lock
2284 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2288 if (cfs_b->quota == RUNTIME_INF)
2291 now = sched_clock_cpu(smp_processor_id());
2292 cfs_b->runtime = cfs_b->quota;
2293 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2296 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2298 return &tg->cfs_bandwidth;
2301 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2302 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2304 if (unlikely(cfs_rq->throttle_count))
2305 return cfs_rq->throttled_clock_task;
2307 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2310 /* returns 0 on failure to allocate runtime */
2311 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2313 struct task_group *tg = cfs_rq->tg;
2314 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2315 u64 amount = 0, min_amount, expires;
2317 /* note: this is a positive sum as runtime_remaining <= 0 */
2318 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2320 raw_spin_lock(&cfs_b->lock);
2321 if (cfs_b->quota == RUNTIME_INF)
2322 amount = min_amount;
2325 * If the bandwidth pool has become inactive, then at least one
2326 * period must have elapsed since the last consumption.
2327 * Refresh the global state and ensure bandwidth timer becomes
2330 if (!cfs_b->timer_active) {
2331 __refill_cfs_bandwidth_runtime(cfs_b);
2332 __start_cfs_bandwidth(cfs_b);
2335 if (cfs_b->runtime > 0) {
2336 amount = min(cfs_b->runtime, min_amount);
2337 cfs_b->runtime -= amount;
2341 expires = cfs_b->runtime_expires;
2342 raw_spin_unlock(&cfs_b->lock);
2344 cfs_rq->runtime_remaining += amount;
2346 * we may have advanced our local expiration to account for allowed
2347 * spread between our sched_clock and the one on which runtime was
2350 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2351 cfs_rq->runtime_expires = expires;
2353 return cfs_rq->runtime_remaining > 0;
2357 * Note: This depends on the synchronization provided by sched_clock and the
2358 * fact that rq->clock snapshots this value.
2360 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2362 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2364 /* if the deadline is ahead of our clock, nothing to do */
2365 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2368 if (cfs_rq->runtime_remaining < 0)
2372 * If the local deadline has passed we have to consider the
2373 * possibility that our sched_clock is 'fast' and the global deadline
2374 * has not truly expired.
2376 * Fortunately we can check determine whether this the case by checking
2377 * whether the global deadline has advanced.
2380 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2381 /* extend local deadline, drift is bounded above by 2 ticks */
2382 cfs_rq->runtime_expires += TICK_NSEC;
2384 /* global deadline is ahead, expiration has passed */
2385 cfs_rq->runtime_remaining = 0;
2389 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2390 unsigned long delta_exec)
2392 /* dock delta_exec before expiring quota (as it could span periods) */
2393 cfs_rq->runtime_remaining -= delta_exec;
2394 expire_cfs_rq_runtime(cfs_rq);
2396 if (likely(cfs_rq->runtime_remaining > 0))
2400 * if we're unable to extend our runtime we resched so that the active
2401 * hierarchy can be throttled
2403 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2404 resched_task(rq_of(cfs_rq)->curr);
2407 static __always_inline
2408 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2410 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2413 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2416 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2418 return cfs_bandwidth_used() && cfs_rq->throttled;
2421 /* check whether cfs_rq, or any parent, is throttled */
2422 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2424 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2428 * Ensure that neither of the group entities corresponding to src_cpu or
2429 * dest_cpu are members of a throttled hierarchy when performing group
2430 * load-balance operations.
2432 static inline int throttled_lb_pair(struct task_group *tg,
2433 int src_cpu, int dest_cpu)
2435 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2437 src_cfs_rq = tg->cfs_rq[src_cpu];
2438 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2440 return throttled_hierarchy(src_cfs_rq) ||
2441 throttled_hierarchy(dest_cfs_rq);
2444 /* updated child weight may affect parent so we have to do this bottom up */
2445 static int tg_unthrottle_up(struct task_group *tg, void *data)
2447 struct rq *rq = data;
2448 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2450 cfs_rq->throttle_count--;
2452 if (!cfs_rq->throttle_count) {
2453 /* adjust cfs_rq_clock_task() */
2454 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2455 cfs_rq->throttled_clock_task;
2462 static int tg_throttle_down(struct task_group *tg, void *data)
2464 struct rq *rq = data;
2465 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2467 /* group is entering throttled state, stop time */
2468 if (!cfs_rq->throttle_count)
2469 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2470 cfs_rq->throttle_count++;
2475 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2477 struct rq *rq = rq_of(cfs_rq);
2478 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2479 struct sched_entity *se;
2480 long task_delta, dequeue = 1;
2482 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2484 /* freeze hierarchy runnable averages while throttled */
2486 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2489 task_delta = cfs_rq->h_nr_running;
2490 for_each_sched_entity(se) {
2491 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2492 /* throttled entity or throttle-on-deactivate */
2497 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2498 qcfs_rq->h_nr_running -= task_delta;
2500 if (qcfs_rq->load.weight)
2505 rq->nr_running -= task_delta;
2507 cfs_rq->throttled = 1;
2508 cfs_rq->throttled_clock = rq_clock(rq);
2509 raw_spin_lock(&cfs_b->lock);
2510 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2511 raw_spin_unlock(&cfs_b->lock);
2514 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2516 struct rq *rq = rq_of(cfs_rq);
2517 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2518 struct sched_entity *se;
2522 se = cfs_rq->tg->se[cpu_of(rq)];
2524 cfs_rq->throttled = 0;
2526 update_rq_clock(rq);
2528 raw_spin_lock(&cfs_b->lock);
2529 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2530 list_del_rcu(&cfs_rq->throttled_list);
2531 raw_spin_unlock(&cfs_b->lock);
2533 /* update hierarchical throttle state */
2534 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2536 if (!cfs_rq->load.weight)
2539 task_delta = cfs_rq->h_nr_running;
2540 for_each_sched_entity(se) {
2544 cfs_rq = cfs_rq_of(se);
2546 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2547 cfs_rq->h_nr_running += task_delta;
2549 if (cfs_rq_throttled(cfs_rq))
2554 rq->nr_running += task_delta;
2556 /* determine whether we need to wake up potentially idle cpu */
2557 if (rq->curr == rq->idle && rq->cfs.nr_running)
2558 resched_task(rq->curr);
2561 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2562 u64 remaining, u64 expires)
2564 struct cfs_rq *cfs_rq;
2565 u64 runtime = remaining;
2568 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2570 struct rq *rq = rq_of(cfs_rq);
2572 raw_spin_lock(&rq->lock);
2573 if (!cfs_rq_throttled(cfs_rq))
2576 runtime = -cfs_rq->runtime_remaining + 1;
2577 if (runtime > remaining)
2578 runtime = remaining;
2579 remaining -= runtime;
2581 cfs_rq->runtime_remaining += runtime;
2582 cfs_rq->runtime_expires = expires;
2584 /* we check whether we're throttled above */
2585 if (cfs_rq->runtime_remaining > 0)
2586 unthrottle_cfs_rq(cfs_rq);
2589 raw_spin_unlock(&rq->lock);
2600 * Responsible for refilling a task_group's bandwidth and unthrottling its
2601 * cfs_rqs as appropriate. If there has been no activity within the last
2602 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2603 * used to track this state.
2605 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2607 u64 runtime, runtime_expires;
2608 int idle = 1, throttled;
2610 raw_spin_lock(&cfs_b->lock);
2611 /* no need to continue the timer with no bandwidth constraint */
2612 if (cfs_b->quota == RUNTIME_INF)
2615 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2616 /* idle depends on !throttled (for the case of a large deficit) */
2617 idle = cfs_b->idle && !throttled;
2618 cfs_b->nr_periods += overrun;
2620 /* if we're going inactive then everything else can be deferred */
2624 __refill_cfs_bandwidth_runtime(cfs_b);
2627 /* mark as potentially idle for the upcoming period */
2632 /* account preceding periods in which throttling occurred */
2633 cfs_b->nr_throttled += overrun;
2636 * There are throttled entities so we must first use the new bandwidth
2637 * to unthrottle them before making it generally available. This
2638 * ensures that all existing debts will be paid before a new cfs_rq is
2641 runtime = cfs_b->runtime;
2642 runtime_expires = cfs_b->runtime_expires;
2646 * This check is repeated as we are holding onto the new bandwidth
2647 * while we unthrottle. This can potentially race with an unthrottled
2648 * group trying to acquire new bandwidth from the global pool.
2650 while (throttled && runtime > 0) {
2651 raw_spin_unlock(&cfs_b->lock);
2652 /* we can't nest cfs_b->lock while distributing bandwidth */
2653 runtime = distribute_cfs_runtime(cfs_b, runtime,
2655 raw_spin_lock(&cfs_b->lock);
2657 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2660 /* return (any) remaining runtime */
2661 cfs_b->runtime = runtime;
2663 * While we are ensured activity in the period following an
2664 * unthrottle, this also covers the case in which the new bandwidth is
2665 * insufficient to cover the existing bandwidth deficit. (Forcing the
2666 * timer to remain active while there are any throttled entities.)
2671 cfs_b->timer_active = 0;
2672 raw_spin_unlock(&cfs_b->lock);
2677 /* a cfs_rq won't donate quota below this amount */
2678 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2679 /* minimum remaining period time to redistribute slack quota */
2680 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2681 /* how long we wait to gather additional slack before distributing */
2682 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2684 /* are we near the end of the current quota period? */
2685 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2687 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2690 /* if the call-back is running a quota refresh is already occurring */
2691 if (hrtimer_callback_running(refresh_timer))
2694 /* is a quota refresh about to occur? */
2695 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2696 if (remaining < min_expire)
2702 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2704 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2706 /* if there's a quota refresh soon don't bother with slack */
2707 if (runtime_refresh_within(cfs_b, min_left))
2710 start_bandwidth_timer(&cfs_b->slack_timer,
2711 ns_to_ktime(cfs_bandwidth_slack_period));
2714 /* we know any runtime found here is valid as update_curr() precedes return */
2715 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2717 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2718 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2720 if (slack_runtime <= 0)
2723 raw_spin_lock(&cfs_b->lock);
2724 if (cfs_b->quota != RUNTIME_INF &&
2725 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2726 cfs_b->runtime += slack_runtime;
2728 /* we are under rq->lock, defer unthrottling using a timer */
2729 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2730 !list_empty(&cfs_b->throttled_cfs_rq))
2731 start_cfs_slack_bandwidth(cfs_b);
2733 raw_spin_unlock(&cfs_b->lock);
2735 /* even if it's not valid for return we don't want to try again */
2736 cfs_rq->runtime_remaining -= slack_runtime;
2739 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2741 if (!cfs_bandwidth_used())
2744 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2747 __return_cfs_rq_runtime(cfs_rq);
2751 * This is done with a timer (instead of inline with bandwidth return) since
2752 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2754 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2756 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2759 /* confirm we're still not at a refresh boundary */
2760 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2763 raw_spin_lock(&cfs_b->lock);
2764 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2765 runtime = cfs_b->runtime;
2768 expires = cfs_b->runtime_expires;
2769 raw_spin_unlock(&cfs_b->lock);
2774 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2776 raw_spin_lock(&cfs_b->lock);
2777 if (expires == cfs_b->runtime_expires)
2778 cfs_b->runtime = runtime;
2779 raw_spin_unlock(&cfs_b->lock);
2783 * When a group wakes up we want to make sure that its quota is not already
2784 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2785 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2787 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2789 if (!cfs_bandwidth_used())
2792 /* an active group must be handled by the update_curr()->put() path */
2793 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2796 /* ensure the group is not already throttled */
2797 if (cfs_rq_throttled(cfs_rq))
2800 /* update runtime allocation */
2801 account_cfs_rq_runtime(cfs_rq, 0);
2802 if (cfs_rq->runtime_remaining <= 0)
2803 throttle_cfs_rq(cfs_rq);
2806 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2807 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2809 if (!cfs_bandwidth_used())
2812 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2816 * it's possible for a throttled entity to be forced into a running
2817 * state (e.g. set_curr_task), in this case we're finished.
2819 if (cfs_rq_throttled(cfs_rq))
2822 throttle_cfs_rq(cfs_rq);
2825 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2827 struct cfs_bandwidth *cfs_b =
2828 container_of(timer, struct cfs_bandwidth, slack_timer);
2829 do_sched_cfs_slack_timer(cfs_b);
2831 return HRTIMER_NORESTART;
2834 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2836 struct cfs_bandwidth *cfs_b =
2837 container_of(timer, struct cfs_bandwidth, period_timer);
2843 now = hrtimer_cb_get_time(timer);
2844 overrun = hrtimer_forward(timer, now, cfs_b->period);
2849 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2852 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2855 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2857 raw_spin_lock_init(&cfs_b->lock);
2859 cfs_b->quota = RUNTIME_INF;
2860 cfs_b->period = ns_to_ktime(default_cfs_period());
2862 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2863 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2864 cfs_b->period_timer.function = sched_cfs_period_timer;
2865 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2866 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2869 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2871 cfs_rq->runtime_enabled = 0;
2872 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2875 /* requires cfs_b->lock, may release to reprogram timer */
2876 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2879 * The timer may be active because we're trying to set a new bandwidth
2880 * period or because we're racing with the tear-down path
2881 * (timer_active==0 becomes visible before the hrtimer call-back
2882 * terminates). In either case we ensure that it's re-programmed
2884 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2885 raw_spin_unlock(&cfs_b->lock);
2886 /* ensure cfs_b->lock is available while we wait */
2887 hrtimer_cancel(&cfs_b->period_timer);
2889 raw_spin_lock(&cfs_b->lock);
2890 /* if someone else restarted the timer then we're done */
2891 if (cfs_b->timer_active)
2895 cfs_b->timer_active = 1;
2896 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2899 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2901 hrtimer_cancel(&cfs_b->period_timer);
2902 hrtimer_cancel(&cfs_b->slack_timer);
2905 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2907 struct cfs_rq *cfs_rq;
2909 for_each_leaf_cfs_rq(rq, cfs_rq) {
2910 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2912 if (!cfs_rq->runtime_enabled)
2916 * clock_task is not advancing so we just need to make sure
2917 * there's some valid quota amount
2919 cfs_rq->runtime_remaining = cfs_b->quota;
2920 if (cfs_rq_throttled(cfs_rq))
2921 unthrottle_cfs_rq(cfs_rq);
2925 #else /* CONFIG_CFS_BANDWIDTH */
2926 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2928 return rq_clock_task(rq_of(cfs_rq));
2931 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2932 unsigned long delta_exec) {}
2933 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2934 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2935 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2937 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2942 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2947 static inline int throttled_lb_pair(struct task_group *tg,
2948 int src_cpu, int dest_cpu)
2953 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2955 #ifdef CONFIG_FAIR_GROUP_SCHED
2956 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2959 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2963 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2964 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2966 #endif /* CONFIG_CFS_BANDWIDTH */
2968 /**************************************************
2969 * CFS operations on tasks:
2972 #ifdef CONFIG_SCHED_HRTICK
2973 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2975 struct sched_entity *se = &p->se;
2976 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2978 WARN_ON(task_rq(p) != rq);
2980 if (cfs_rq->nr_running > 1) {
2981 u64 slice = sched_slice(cfs_rq, se);
2982 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2983 s64 delta = slice - ran;
2992 * Don't schedule slices shorter than 10000ns, that just
2993 * doesn't make sense. Rely on vruntime for fairness.
2996 delta = max_t(s64, 10000LL, delta);
2998 hrtick_start(rq, delta);
3003 * called from enqueue/dequeue and updates the hrtick when the
3004 * current task is from our class and nr_running is low enough
3007 static void hrtick_update(struct rq *rq)
3009 struct task_struct *curr = rq->curr;
3011 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3014 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3015 hrtick_start_fair(rq, curr);
3017 #else /* !CONFIG_SCHED_HRTICK */
3019 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3023 static inline void hrtick_update(struct rq *rq)
3029 * The enqueue_task method is called before nr_running is
3030 * increased. Here we update the fair scheduling stats and
3031 * then put the task into the rbtree:
3034 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3036 struct cfs_rq *cfs_rq;
3037 struct sched_entity *se = &p->se;
3039 for_each_sched_entity(se) {
3042 cfs_rq = cfs_rq_of(se);
3043 enqueue_entity(cfs_rq, se, flags);
3046 * end evaluation on encountering a throttled cfs_rq
3048 * note: in the case of encountering a throttled cfs_rq we will
3049 * post the final h_nr_running increment below.
3051 if (cfs_rq_throttled(cfs_rq))
3053 cfs_rq->h_nr_running++;
3055 flags = ENQUEUE_WAKEUP;
3058 for_each_sched_entity(se) {
3059 cfs_rq = cfs_rq_of(se);
3060 cfs_rq->h_nr_running++;
3062 if (cfs_rq_throttled(cfs_rq))
3065 update_cfs_shares(cfs_rq);
3066 update_entity_load_avg(se, 1);
3070 update_rq_runnable_avg(rq, rq->nr_running);
3076 static void set_next_buddy(struct sched_entity *se);
3079 * The dequeue_task method is called before nr_running is
3080 * decreased. We remove the task from the rbtree and
3081 * update the fair scheduling stats:
3083 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3085 struct cfs_rq *cfs_rq;
3086 struct sched_entity *se = &p->se;
3087 int task_sleep = flags & DEQUEUE_SLEEP;
3089 for_each_sched_entity(se) {
3090 cfs_rq = cfs_rq_of(se);
3091 dequeue_entity(cfs_rq, se, flags);
3094 * end evaluation on encountering a throttled cfs_rq
3096 * note: in the case of encountering a throttled cfs_rq we will
3097 * post the final h_nr_running decrement below.
3099 if (cfs_rq_throttled(cfs_rq))
3101 cfs_rq->h_nr_running--;
3103 /* Don't dequeue parent if it has other entities besides us */
3104 if (cfs_rq->load.weight) {
3106 * Bias pick_next to pick a task from this cfs_rq, as
3107 * p is sleeping when it is within its sched_slice.
3109 if (task_sleep && parent_entity(se))
3110 set_next_buddy(parent_entity(se));
3112 /* avoid re-evaluating load for this entity */
3113 se = parent_entity(se);
3116 flags |= DEQUEUE_SLEEP;
3119 for_each_sched_entity(se) {
3120 cfs_rq = cfs_rq_of(se);
3121 cfs_rq->h_nr_running--;
3123 if (cfs_rq_throttled(cfs_rq))
3126 update_cfs_shares(cfs_rq);
3127 update_entity_load_avg(se, 1);
3132 update_rq_runnable_avg(rq, 1);
3138 /* Used instead of source_load when we know the type == 0 */
3139 static unsigned long weighted_cpuload(const int cpu)
3141 return cpu_rq(cpu)->cfs.runnable_load_avg;
3145 * Return a low guess at the load of a migration-source cpu weighted
3146 * according to the scheduling class and "nice" value.
3148 * We want to under-estimate the load of migration sources, to
3149 * balance conservatively.
3151 static unsigned long source_load(int cpu, int type)
3153 struct rq *rq = cpu_rq(cpu);
3154 unsigned long total = weighted_cpuload(cpu);
3156 if (type == 0 || !sched_feat(LB_BIAS))
3159 return min(rq->cpu_load[type-1], total);
3163 * Return a high guess at the load of a migration-target cpu weighted
3164 * according to the scheduling class and "nice" value.
3166 static unsigned long target_load(int cpu, int type)
3168 struct rq *rq = cpu_rq(cpu);
3169 unsigned long total = weighted_cpuload(cpu);
3171 if (type == 0 || !sched_feat(LB_BIAS))
3174 return max(rq->cpu_load[type-1], total);
3177 static unsigned long power_of(int cpu)
3179 return cpu_rq(cpu)->cpu_power;
3182 static unsigned long cpu_avg_load_per_task(int cpu)
3184 struct rq *rq = cpu_rq(cpu);
3185 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3186 unsigned long load_avg = rq->cfs.runnable_load_avg;
3189 return load_avg / nr_running;
3194 static void record_wakee(struct task_struct *p)
3197 * Rough decay (wiping) for cost saving, don't worry
3198 * about the boundary, really active task won't care
3201 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3202 current->wakee_flips = 0;
3203 current->wakee_flip_decay_ts = jiffies;
3206 if (current->last_wakee != p) {
3207 current->last_wakee = p;
3208 current->wakee_flips++;
3212 static void task_waking_fair(struct task_struct *p)
3214 struct sched_entity *se = &p->se;
3215 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3218 #ifndef CONFIG_64BIT
3219 u64 min_vruntime_copy;
3222 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3224 min_vruntime = cfs_rq->min_vruntime;
3225 } while (min_vruntime != min_vruntime_copy);
3227 min_vruntime = cfs_rq->min_vruntime;
3230 se->vruntime -= min_vruntime;
3234 #ifdef CONFIG_FAIR_GROUP_SCHED
3236 * effective_load() calculates the load change as seen from the root_task_group
3238 * Adding load to a group doesn't make a group heavier, but can cause movement
3239 * of group shares between cpus. Assuming the shares were perfectly aligned one
3240 * can calculate the shift in shares.
3242 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3243 * on this @cpu and results in a total addition (subtraction) of @wg to the
3244 * total group weight.
3246 * Given a runqueue weight distribution (rw_i) we can compute a shares
3247 * distribution (s_i) using:
3249 * s_i = rw_i / \Sum rw_j (1)
3251 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3252 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3253 * shares distribution (s_i):
3255 * rw_i = { 2, 4, 1, 0 }
3256 * s_i = { 2/7, 4/7, 1/7, 0 }
3258 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3259 * task used to run on and the CPU the waker is running on), we need to
3260 * compute the effect of waking a task on either CPU and, in case of a sync
3261 * wakeup, compute the effect of the current task going to sleep.
3263 * So for a change of @wl to the local @cpu with an overall group weight change
3264 * of @wl we can compute the new shares distribution (s'_i) using:
3266 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3268 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3269 * differences in waking a task to CPU 0. The additional task changes the
3270 * weight and shares distributions like:
3272 * rw'_i = { 3, 4, 1, 0 }
3273 * s'_i = { 3/8, 4/8, 1/8, 0 }
3275 * We can then compute the difference in effective weight by using:
3277 * dw_i = S * (s'_i - s_i) (3)
3279 * Where 'S' is the group weight as seen by its parent.
3281 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3282 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3283 * 4/7) times the weight of the group.
3285 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3287 struct sched_entity *se = tg->se[cpu];
3289 if (!tg->parent) /* the trivial, non-cgroup case */
3292 for_each_sched_entity(se) {
3298 * W = @wg + \Sum rw_j
3300 W = wg + calc_tg_weight(tg, se->my_q);
3305 w = se->my_q->load.weight + wl;
3308 * wl = S * s'_i; see (2)
3311 wl = (w * tg->shares) / W;
3316 * Per the above, wl is the new se->load.weight value; since
3317 * those are clipped to [MIN_SHARES, ...) do so now. See
3318 * calc_cfs_shares().
3320 if (wl < MIN_SHARES)
3324 * wl = dw_i = S * (s'_i - s_i); see (3)
3326 wl -= se->load.weight;
3329 * Recursively apply this logic to all parent groups to compute
3330 * the final effective load change on the root group. Since
3331 * only the @tg group gets extra weight, all parent groups can
3332 * only redistribute existing shares. @wl is the shift in shares
3333 * resulting from this level per the above.
3342 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3343 unsigned long wl, unsigned long wg)
3350 static int wake_wide(struct task_struct *p)
3352 int factor = this_cpu_read(sd_llc_size);
3355 * Yeah, it's the switching-frequency, could means many wakee or
3356 * rapidly switch, use factor here will just help to automatically
3357 * adjust the loose-degree, so bigger node will lead to more pull.
3359 if (p->wakee_flips > factor) {
3361 * wakee is somewhat hot, it needs certain amount of cpu
3362 * resource, so if waker is far more hot, prefer to leave
3365 if (current->wakee_flips > (factor * p->wakee_flips))
3372 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3374 s64 this_load, load;
3375 int idx, this_cpu, prev_cpu;
3376 unsigned long tl_per_task;
3377 struct task_group *tg;
3378 unsigned long weight;
3382 * If we wake multiple tasks be careful to not bounce
3383 * ourselves around too much.
3389 this_cpu = smp_processor_id();
3390 prev_cpu = task_cpu(p);
3391 load = source_load(prev_cpu, idx);
3392 this_load = target_load(this_cpu, idx);
3395 * If sync wakeup then subtract the (maximum possible)
3396 * effect of the currently running task from the load
3397 * of the current CPU:
3400 tg = task_group(current);
3401 weight = current->se.load.weight;
3403 this_load += effective_load(tg, this_cpu, -weight, -weight);
3404 load += effective_load(tg, prev_cpu, 0, -weight);
3408 weight = p->se.load.weight;
3411 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3412 * due to the sync cause above having dropped this_load to 0, we'll
3413 * always have an imbalance, but there's really nothing you can do
3414 * about that, so that's good too.
3416 * Otherwise check if either cpus are near enough in load to allow this
3417 * task to be woken on this_cpu.
3419 if (this_load > 0) {
3420 s64 this_eff_load, prev_eff_load;
3422 this_eff_load = 100;
3423 this_eff_load *= power_of(prev_cpu);
3424 this_eff_load *= this_load +
3425 effective_load(tg, this_cpu, weight, weight);
3427 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3428 prev_eff_load *= power_of(this_cpu);
3429 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3431 balanced = this_eff_load <= prev_eff_load;
3436 * If the currently running task will sleep within
3437 * a reasonable amount of time then attract this newly
3440 if (sync && balanced)
3443 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3444 tl_per_task = cpu_avg_load_per_task(this_cpu);
3447 (this_load <= load &&
3448 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3450 * This domain has SD_WAKE_AFFINE and
3451 * p is cache cold in this domain, and
3452 * there is no bad imbalance.
3454 schedstat_inc(sd, ttwu_move_affine);
3455 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3463 * find_idlest_group finds and returns the least busy CPU group within the
3466 static struct sched_group *
3467 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3468 int this_cpu, int load_idx)
3470 struct sched_group *idlest = NULL, *group = sd->groups;
3471 unsigned long min_load = ULONG_MAX, this_load = 0;
3472 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3475 unsigned long load, avg_load;
3479 /* Skip over this group if it has no CPUs allowed */
3480 if (!cpumask_intersects(sched_group_cpus(group),
3481 tsk_cpus_allowed(p)))
3484 local_group = cpumask_test_cpu(this_cpu,
3485 sched_group_cpus(group));
3487 /* Tally up the load of all CPUs in the group */
3490 for_each_cpu(i, sched_group_cpus(group)) {
3491 /* Bias balancing toward cpus of our domain */
3493 load = source_load(i, load_idx);
3495 load = target_load(i, load_idx);
3500 /* Adjust by relative CPU power of the group */
3501 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3504 this_load = avg_load;
3505 } else if (avg_load < min_load) {
3506 min_load = avg_load;
3509 } while (group = group->next, group != sd->groups);
3511 if (!idlest || 100*this_load < imbalance*min_load)
3517 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3520 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3522 unsigned long load, min_load = ULONG_MAX;
3526 /* Traverse only the allowed CPUs */
3527 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3528 load = weighted_cpuload(i);
3530 if (load < min_load || (load == min_load && i == this_cpu)) {
3540 * Try and locate an idle CPU in the sched_domain.
3542 static int select_idle_sibling(struct task_struct *p, int target)
3544 struct sched_domain *sd;
3545 struct sched_group *sg;
3546 int i = task_cpu(p);
3548 if (idle_cpu(target))
3552 * If the prevous cpu is cache affine and idle, don't be stupid.
3554 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3558 * Otherwise, iterate the domains and find an elegible idle cpu.
3560 sd = rcu_dereference(per_cpu(sd_llc, target));
3561 for_each_lower_domain(sd) {
3564 if (!cpumask_intersects(sched_group_cpus(sg),
3565 tsk_cpus_allowed(p)))
3568 for_each_cpu(i, sched_group_cpus(sg)) {
3569 if (i == target || !idle_cpu(i))
3573 target = cpumask_first_and(sched_group_cpus(sg),
3574 tsk_cpus_allowed(p));
3578 } while (sg != sd->groups);
3585 * sched_balance_self: balance the current task (running on cpu) in domains
3586 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3589 * Balance, ie. select the least loaded group.
3591 * Returns the target CPU number, or the same CPU if no balancing is needed.
3593 * preempt must be disabled.
3596 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3598 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3599 int cpu = smp_processor_id();
3600 int prev_cpu = task_cpu(p);
3602 int want_affine = 0;
3603 int sync = wake_flags & WF_SYNC;
3605 if (p->nr_cpus_allowed == 1)
3608 if (sd_flag & SD_BALANCE_WAKE) {
3609 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3615 for_each_domain(cpu, tmp) {
3616 if (!(tmp->flags & SD_LOAD_BALANCE))
3620 * If both cpu and prev_cpu are part of this domain,
3621 * cpu is a valid SD_WAKE_AFFINE target.
3623 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3624 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3629 if (tmp->flags & sd_flag)
3634 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3637 new_cpu = select_idle_sibling(p, prev_cpu);
3642 int load_idx = sd->forkexec_idx;
3643 struct sched_group *group;
3646 if (!(sd->flags & sd_flag)) {
3651 if (sd_flag & SD_BALANCE_WAKE)
3652 load_idx = sd->wake_idx;
3654 group = find_idlest_group(sd, p, cpu, load_idx);
3660 new_cpu = find_idlest_cpu(group, p, cpu);
3661 if (new_cpu == -1 || new_cpu == cpu) {
3662 /* Now try balancing at a lower domain level of cpu */
3667 /* Now try balancing at a lower domain level of new_cpu */
3669 weight = sd->span_weight;
3671 for_each_domain(cpu, tmp) {
3672 if (weight <= tmp->span_weight)
3674 if (tmp->flags & sd_flag)
3677 /* while loop will break here if sd == NULL */
3686 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3687 * cfs_rq_of(p) references at time of call are still valid and identify the
3688 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3689 * other assumptions, including the state of rq->lock, should be made.
3692 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3694 struct sched_entity *se = &p->se;
3695 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3698 * Load tracking: accumulate removed load so that it can be processed
3699 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3700 * to blocked load iff they have a positive decay-count. It can never
3701 * be negative here since on-rq tasks have decay-count == 0.
3703 if (se->avg.decay_count) {
3704 se->avg.decay_count = -__synchronize_entity_decay(se);
3705 atomic_long_add(se->avg.load_avg_contrib,
3706 &cfs_rq->removed_load);
3709 #endif /* CONFIG_SMP */
3711 static unsigned long
3712 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3714 unsigned long gran = sysctl_sched_wakeup_granularity;
3717 * Since its curr running now, convert the gran from real-time
3718 * to virtual-time in his units.
3720 * By using 'se' instead of 'curr' we penalize light tasks, so
3721 * they get preempted easier. That is, if 'se' < 'curr' then
3722 * the resulting gran will be larger, therefore penalizing the
3723 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3724 * be smaller, again penalizing the lighter task.
3726 * This is especially important for buddies when the leftmost
3727 * task is higher priority than the buddy.
3729 return calc_delta_fair(gran, se);
3733 * Should 'se' preempt 'curr'.
3747 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3749 s64 gran, vdiff = curr->vruntime - se->vruntime;
3754 gran = wakeup_gran(curr, se);
3761 static void set_last_buddy(struct sched_entity *se)
3763 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3766 for_each_sched_entity(se)
3767 cfs_rq_of(se)->last = se;
3770 static void set_next_buddy(struct sched_entity *se)
3772 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3775 for_each_sched_entity(se)
3776 cfs_rq_of(se)->next = se;
3779 static void set_skip_buddy(struct sched_entity *se)
3781 for_each_sched_entity(se)
3782 cfs_rq_of(se)->skip = se;
3786 * Preempt the current task with a newly woken task if needed:
3788 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3790 struct task_struct *curr = rq->curr;
3791 struct sched_entity *se = &curr->se, *pse = &p->se;
3792 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3793 int scale = cfs_rq->nr_running >= sched_nr_latency;
3794 int next_buddy_marked = 0;
3796 if (unlikely(se == pse))
3800 * This is possible from callers such as move_task(), in which we
3801 * unconditionally check_prempt_curr() after an enqueue (which may have
3802 * lead to a throttle). This both saves work and prevents false
3803 * next-buddy nomination below.
3805 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3808 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3809 set_next_buddy(pse);
3810 next_buddy_marked = 1;
3814 * We can come here with TIF_NEED_RESCHED already set from new task
3817 * Note: this also catches the edge-case of curr being in a throttled
3818 * group (e.g. via set_curr_task), since update_curr() (in the
3819 * enqueue of curr) will have resulted in resched being set. This
3820 * prevents us from potentially nominating it as a false LAST_BUDDY
3823 if (test_tsk_need_resched(curr))
3826 /* Idle tasks are by definition preempted by non-idle tasks. */
3827 if (unlikely(curr->policy == SCHED_IDLE) &&
3828 likely(p->policy != SCHED_IDLE))
3832 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3833 * is driven by the tick):
3835 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3838 find_matching_se(&se, &pse);
3839 update_curr(cfs_rq_of(se));
3841 if (wakeup_preempt_entity(se, pse) == 1) {
3843 * Bias pick_next to pick the sched entity that is
3844 * triggering this preemption.
3846 if (!next_buddy_marked)
3847 set_next_buddy(pse);
3856 * Only set the backward buddy when the current task is still
3857 * on the rq. This can happen when a wakeup gets interleaved
3858 * with schedule on the ->pre_schedule() or idle_balance()
3859 * point, either of which can * drop the rq lock.
3861 * Also, during early boot the idle thread is in the fair class,
3862 * for obvious reasons its a bad idea to schedule back to it.
3864 if (unlikely(!se->on_rq || curr == rq->idle))
3867 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3871 static struct task_struct *pick_next_task_fair(struct rq *rq)
3873 struct task_struct *p;
3874 struct cfs_rq *cfs_rq = &rq->cfs;
3875 struct sched_entity *se;
3877 if (!cfs_rq->nr_running)
3881 se = pick_next_entity(cfs_rq);
3882 set_next_entity(cfs_rq, se);
3883 cfs_rq = group_cfs_rq(se);
3887 if (hrtick_enabled(rq))
3888 hrtick_start_fair(rq, p);
3894 * Account for a descheduled task:
3896 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3898 struct sched_entity *se = &prev->se;
3899 struct cfs_rq *cfs_rq;
3901 for_each_sched_entity(se) {
3902 cfs_rq = cfs_rq_of(se);
3903 put_prev_entity(cfs_rq, se);
3908 * sched_yield() is very simple
3910 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3912 static void yield_task_fair(struct rq *rq)
3914 struct task_struct *curr = rq->curr;
3915 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3916 struct sched_entity *se = &curr->se;
3919 * Are we the only task in the tree?
3921 if (unlikely(rq->nr_running == 1))
3924 clear_buddies(cfs_rq, se);
3926 if (curr->policy != SCHED_BATCH) {
3927 update_rq_clock(rq);
3929 * Update run-time statistics of the 'current'.
3931 update_curr(cfs_rq);
3933 * Tell update_rq_clock() that we've just updated,
3934 * so we don't do microscopic update in schedule()
3935 * and double the fastpath cost.
3937 rq->skip_clock_update = 1;
3943 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3945 struct sched_entity *se = &p->se;
3947 /* throttled hierarchies are not runnable */
3948 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3951 /* Tell the scheduler that we'd really like pse to run next. */
3954 yield_task_fair(rq);
3960 /**************************************************
3961 * Fair scheduling class load-balancing methods.
3965 * The purpose of load-balancing is to achieve the same basic fairness the
3966 * per-cpu scheduler provides, namely provide a proportional amount of compute
3967 * time to each task. This is expressed in the following equation:
3969 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3971 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3972 * W_i,0 is defined as:
3974 * W_i,0 = \Sum_j w_i,j (2)
3976 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3977 * is derived from the nice value as per prio_to_weight[].
3979 * The weight average is an exponential decay average of the instantaneous
3982 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3984 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3985 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3986 * can also include other factors [XXX].
3988 * To achieve this balance we define a measure of imbalance which follows
3989 * directly from (1):
3991 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3993 * We them move tasks around to minimize the imbalance. In the continuous
3994 * function space it is obvious this converges, in the discrete case we get
3995 * a few fun cases generally called infeasible weight scenarios.
3998 * - infeasible weights;
3999 * - local vs global optima in the discrete case. ]
4004 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4005 * for all i,j solution, we create a tree of cpus that follows the hardware
4006 * topology where each level pairs two lower groups (or better). This results
4007 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4008 * tree to only the first of the previous level and we decrease the frequency
4009 * of load-balance at each level inv. proportional to the number of cpus in
4015 * \Sum { --- * --- * 2^i } = O(n) (5)
4017 * `- size of each group
4018 * | | `- number of cpus doing load-balance
4020 * `- sum over all levels
4022 * Coupled with a limit on how many tasks we can migrate every balance pass,
4023 * this makes (5) the runtime complexity of the balancer.
4025 * An important property here is that each CPU is still (indirectly) connected
4026 * to every other cpu in at most O(log n) steps:
4028 * The adjacency matrix of the resulting graph is given by:
4031 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4034 * And you'll find that:
4036 * A^(log_2 n)_i,j != 0 for all i,j (7)
4038 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4039 * The task movement gives a factor of O(m), giving a convergence complexity
4042 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4047 * In order to avoid CPUs going idle while there's still work to do, new idle
4048 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4049 * tree itself instead of relying on other CPUs to bring it work.
4051 * This adds some complexity to both (5) and (8) but it reduces the total idle
4059 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4062 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4067 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4069 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4071 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4074 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4075 * rewrite all of this once again.]
4078 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4080 #define LBF_ALL_PINNED 0x01
4081 #define LBF_NEED_BREAK 0x02
4082 #define LBF_DST_PINNED 0x04
4083 #define LBF_SOME_PINNED 0x08
4086 struct sched_domain *sd;
4094 struct cpumask *dst_grpmask;
4096 enum cpu_idle_type idle;
4098 /* The set of CPUs under consideration for load-balancing */
4099 struct cpumask *cpus;
4104 unsigned int loop_break;
4105 unsigned int loop_max;
4109 * move_task - move a task from one runqueue to another runqueue.
4110 * Both runqueues must be locked.
4112 static void move_task(struct task_struct *p, struct lb_env *env)
4114 deactivate_task(env->src_rq, p, 0);
4115 set_task_cpu(p, env->dst_cpu);
4116 activate_task(env->dst_rq, p, 0);
4117 check_preempt_curr(env->dst_rq, p, 0);
4121 * Is this task likely cache-hot:
4124 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4128 if (p->sched_class != &fair_sched_class)
4131 if (unlikely(p->policy == SCHED_IDLE))
4135 * Buddy candidates are cache hot:
4137 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4138 (&p->se == cfs_rq_of(&p->se)->next ||
4139 &p->se == cfs_rq_of(&p->se)->last))
4142 if (sysctl_sched_migration_cost == -1)
4144 if (sysctl_sched_migration_cost == 0)
4147 delta = now - p->se.exec_start;
4149 return delta < (s64)sysctl_sched_migration_cost;
4152 #ifdef CONFIG_NUMA_BALANCING
4153 /* Returns true if the destination node has incurred more faults */
4154 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4156 int src_nid, dst_nid;
4158 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4159 !(env->sd->flags & SD_NUMA)) {
4163 src_nid = cpu_to_node(env->src_cpu);
4164 dst_nid = cpu_to_node(env->dst_cpu);
4166 if (src_nid == dst_nid ||
4167 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4170 if (dst_nid == p->numa_preferred_nid ||
4171 task_faults(p, dst_nid) > task_faults(p, src_nid))
4178 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4180 int src_nid, dst_nid;
4182 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4185 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4188 src_nid = cpu_to_node(env->src_cpu);
4189 dst_nid = cpu_to_node(env->dst_cpu);
4191 if (src_nid == dst_nid ||
4192 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4195 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4202 static inline bool migrate_improves_locality(struct task_struct *p,
4208 static inline bool migrate_degrades_locality(struct task_struct *p,
4216 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4219 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4221 int tsk_cache_hot = 0;
4223 * We do not migrate tasks that are:
4224 * 1) throttled_lb_pair, or
4225 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4226 * 3) running (obviously), or
4227 * 4) are cache-hot on their current CPU.
4229 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4232 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4235 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4237 env->flags |= LBF_SOME_PINNED;
4240 * Remember if this task can be migrated to any other cpu in
4241 * our sched_group. We may want to revisit it if we couldn't
4242 * meet load balance goals by pulling other tasks on src_cpu.
4244 * Also avoid computing new_dst_cpu if we have already computed
4245 * one in current iteration.
4247 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4250 /* Prevent to re-select dst_cpu via env's cpus */
4251 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4252 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4253 env->flags |= LBF_DST_PINNED;
4254 env->new_dst_cpu = cpu;
4262 /* Record that we found atleast one task that could run on dst_cpu */
4263 env->flags &= ~LBF_ALL_PINNED;
4265 if (task_running(env->src_rq, p)) {
4266 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4271 * Aggressive migration if:
4272 * 1) destination numa is preferred
4273 * 2) task is cache cold, or
4274 * 3) too many balance attempts have failed.
4276 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4278 tsk_cache_hot = migrate_degrades_locality(p, env);
4280 if (migrate_improves_locality(p, env)) {
4281 #ifdef CONFIG_SCHEDSTATS
4282 if (tsk_cache_hot) {
4283 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4284 schedstat_inc(p, se.statistics.nr_forced_migrations);
4290 if (!tsk_cache_hot ||
4291 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4293 if (tsk_cache_hot) {
4294 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4295 schedstat_inc(p, se.statistics.nr_forced_migrations);
4301 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4306 * move_one_task tries to move exactly one task from busiest to this_rq, as
4307 * part of active balancing operations within "domain".
4308 * Returns 1 if successful and 0 otherwise.
4310 * Called with both runqueues locked.
4312 static int move_one_task(struct lb_env *env)
4314 struct task_struct *p, *n;
4316 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4317 if (!can_migrate_task(p, env))
4322 * Right now, this is only the second place move_task()
4323 * is called, so we can safely collect move_task()
4324 * stats here rather than inside move_task().
4326 schedstat_inc(env->sd, lb_gained[env->idle]);
4332 static unsigned long task_h_load(struct task_struct *p);
4334 static const unsigned int sched_nr_migrate_break = 32;
4337 * move_tasks tries to move up to imbalance weighted load from busiest to
4338 * this_rq, as part of a balancing operation within domain "sd".
4339 * Returns 1 if successful and 0 otherwise.
4341 * Called with both runqueues locked.
4343 static int move_tasks(struct lb_env *env)
4345 struct list_head *tasks = &env->src_rq->cfs_tasks;
4346 struct task_struct *p;
4350 if (env->imbalance <= 0)
4353 while (!list_empty(tasks)) {
4354 p = list_first_entry(tasks, struct task_struct, se.group_node);
4357 /* We've more or less seen every task there is, call it quits */
4358 if (env->loop > env->loop_max)
4361 /* take a breather every nr_migrate tasks */
4362 if (env->loop > env->loop_break) {
4363 env->loop_break += sched_nr_migrate_break;
4364 env->flags |= LBF_NEED_BREAK;
4368 if (!can_migrate_task(p, env))
4371 load = task_h_load(p);
4373 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4376 if ((load / 2) > env->imbalance)
4381 env->imbalance -= load;
4383 #ifdef CONFIG_PREEMPT
4385 * NEWIDLE balancing is a source of latency, so preemptible
4386 * kernels will stop after the first task is pulled to minimize
4387 * the critical section.
4389 if (env->idle == CPU_NEWLY_IDLE)
4394 * We only want to steal up to the prescribed amount of
4397 if (env->imbalance <= 0)
4402 list_move_tail(&p->se.group_node, tasks);
4406 * Right now, this is one of only two places move_task() is called,
4407 * so we can safely collect move_task() stats here rather than
4408 * inside move_task().
4410 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4415 #ifdef CONFIG_FAIR_GROUP_SCHED
4417 * update tg->load_weight by folding this cpu's load_avg
4419 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4421 struct sched_entity *se = tg->se[cpu];
4422 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4424 /* throttled entities do not contribute to load */
4425 if (throttled_hierarchy(cfs_rq))
4428 update_cfs_rq_blocked_load(cfs_rq, 1);
4431 update_entity_load_avg(se, 1);
4433 * We pivot on our runnable average having decayed to zero for
4434 * list removal. This generally implies that all our children
4435 * have also been removed (modulo rounding error or bandwidth
4436 * control); however, such cases are rare and we can fix these
4439 * TODO: fix up out-of-order children on enqueue.
4441 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4442 list_del_leaf_cfs_rq(cfs_rq);
4444 struct rq *rq = rq_of(cfs_rq);
4445 update_rq_runnable_avg(rq, rq->nr_running);
4449 static void update_blocked_averages(int cpu)
4451 struct rq *rq = cpu_rq(cpu);
4452 struct cfs_rq *cfs_rq;
4453 unsigned long flags;
4455 raw_spin_lock_irqsave(&rq->lock, flags);
4456 update_rq_clock(rq);
4458 * Iterates the task_group tree in a bottom up fashion, see
4459 * list_add_leaf_cfs_rq() for details.
4461 for_each_leaf_cfs_rq(rq, cfs_rq) {
4463 * Note: We may want to consider periodically releasing
4464 * rq->lock about these updates so that creating many task
4465 * groups does not result in continually extending hold time.
4467 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4470 raw_spin_unlock_irqrestore(&rq->lock, flags);
4474 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4475 * This needs to be done in a top-down fashion because the load of a child
4476 * group is a fraction of its parents load.
4478 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4480 struct rq *rq = rq_of(cfs_rq);
4481 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4482 unsigned long now = jiffies;
4485 if (cfs_rq->last_h_load_update == now)
4488 cfs_rq->h_load_next = NULL;
4489 for_each_sched_entity(se) {
4490 cfs_rq = cfs_rq_of(se);
4491 cfs_rq->h_load_next = se;
4492 if (cfs_rq->last_h_load_update == now)
4497 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4498 cfs_rq->last_h_load_update = now;
4501 while ((se = cfs_rq->h_load_next) != NULL) {
4502 load = cfs_rq->h_load;
4503 load = div64_ul(load * se->avg.load_avg_contrib,
4504 cfs_rq->runnable_load_avg + 1);
4505 cfs_rq = group_cfs_rq(se);
4506 cfs_rq->h_load = load;
4507 cfs_rq->last_h_load_update = now;
4511 static unsigned long task_h_load(struct task_struct *p)
4513 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4515 update_cfs_rq_h_load(cfs_rq);
4516 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4517 cfs_rq->runnable_load_avg + 1);
4520 static inline void update_blocked_averages(int cpu)
4524 static unsigned long task_h_load(struct task_struct *p)
4526 return p->se.avg.load_avg_contrib;
4530 /********** Helpers for find_busiest_group ************************/
4532 * sg_lb_stats - stats of a sched_group required for load_balancing
4534 struct sg_lb_stats {
4535 unsigned long avg_load; /*Avg load across the CPUs of the group */
4536 unsigned long group_load; /* Total load over the CPUs of the group */
4537 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4538 unsigned long load_per_task;
4539 unsigned long group_power;
4540 unsigned int sum_nr_running; /* Nr tasks running in the group */
4541 unsigned int group_capacity;
4542 unsigned int idle_cpus;
4543 unsigned int group_weight;
4544 int group_imb; /* Is there an imbalance in the group ? */
4545 int group_has_capacity; /* Is there extra capacity in the group? */
4549 * sd_lb_stats - Structure to store the statistics of a sched_domain
4550 * during load balancing.
4552 struct sd_lb_stats {
4553 struct sched_group *busiest; /* Busiest group in this sd */
4554 struct sched_group *local; /* Local group in this sd */
4555 unsigned long total_load; /* Total load of all groups in sd */
4556 unsigned long total_pwr; /* Total power of all groups in sd */
4557 unsigned long avg_load; /* Average load across all groups in sd */
4559 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4560 struct sg_lb_stats local_stat; /* Statistics of the local group */
4563 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4566 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4567 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4568 * We must however clear busiest_stat::avg_load because
4569 * update_sd_pick_busiest() reads this before assignment.
4571 *sds = (struct sd_lb_stats){
4583 * get_sd_load_idx - Obtain the load index for a given sched domain.
4584 * @sd: The sched_domain whose load_idx is to be obtained.
4585 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4587 * Return: The load index.
4589 static inline int get_sd_load_idx(struct sched_domain *sd,
4590 enum cpu_idle_type idle)
4596 load_idx = sd->busy_idx;
4599 case CPU_NEWLY_IDLE:
4600 load_idx = sd->newidle_idx;
4603 load_idx = sd->idle_idx;
4610 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4612 return SCHED_POWER_SCALE;
4615 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4617 return default_scale_freq_power(sd, cpu);
4620 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4622 unsigned long weight = sd->span_weight;
4623 unsigned long smt_gain = sd->smt_gain;
4630 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4632 return default_scale_smt_power(sd, cpu);
4635 static unsigned long scale_rt_power(int cpu)
4637 struct rq *rq = cpu_rq(cpu);
4638 u64 total, available, age_stamp, avg;
4641 * Since we're reading these variables without serialization make sure
4642 * we read them once before doing sanity checks on them.
4644 age_stamp = ACCESS_ONCE(rq->age_stamp);
4645 avg = ACCESS_ONCE(rq->rt_avg);
4647 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4649 if (unlikely(total < avg)) {
4650 /* Ensures that power won't end up being negative */
4653 available = total - avg;
4656 if (unlikely((s64)total < SCHED_POWER_SCALE))
4657 total = SCHED_POWER_SCALE;
4659 total >>= SCHED_POWER_SHIFT;
4661 return div_u64(available, total);
4664 static void update_cpu_power(struct sched_domain *sd, int cpu)
4666 unsigned long weight = sd->span_weight;
4667 unsigned long power = SCHED_POWER_SCALE;
4668 struct sched_group *sdg = sd->groups;
4670 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4671 if (sched_feat(ARCH_POWER))
4672 power *= arch_scale_smt_power(sd, cpu);
4674 power *= default_scale_smt_power(sd, cpu);
4676 power >>= SCHED_POWER_SHIFT;
4679 sdg->sgp->power_orig = power;
4681 if (sched_feat(ARCH_POWER))
4682 power *= arch_scale_freq_power(sd, cpu);
4684 power *= default_scale_freq_power(sd, cpu);
4686 power >>= SCHED_POWER_SHIFT;
4688 power *= scale_rt_power(cpu);
4689 power >>= SCHED_POWER_SHIFT;
4694 cpu_rq(cpu)->cpu_power = power;
4695 sdg->sgp->power = power;
4698 void update_group_power(struct sched_domain *sd, int cpu)
4700 struct sched_domain *child = sd->child;
4701 struct sched_group *group, *sdg = sd->groups;
4702 unsigned long power, power_orig;
4703 unsigned long interval;
4705 interval = msecs_to_jiffies(sd->balance_interval);
4706 interval = clamp(interval, 1UL, max_load_balance_interval);
4707 sdg->sgp->next_update = jiffies + interval;
4710 update_cpu_power(sd, cpu);
4714 power_orig = power = 0;
4716 if (child->flags & SD_OVERLAP) {
4718 * SD_OVERLAP domains cannot assume that child groups
4719 * span the current group.
4722 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4723 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4725 power_orig += sg->sgp->power_orig;
4726 power += sg->sgp->power;
4730 * !SD_OVERLAP domains can assume that child groups
4731 * span the current group.
4734 group = child->groups;
4736 power_orig += group->sgp->power_orig;
4737 power += group->sgp->power;
4738 group = group->next;
4739 } while (group != child->groups);
4742 sdg->sgp->power_orig = power_orig;
4743 sdg->sgp->power = power;
4747 * Try and fix up capacity for tiny siblings, this is needed when
4748 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4749 * which on its own isn't powerful enough.
4751 * See update_sd_pick_busiest() and check_asym_packing().
4754 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4757 * Only siblings can have significantly less than SCHED_POWER_SCALE
4759 if (!(sd->flags & SD_SHARE_CPUPOWER))
4763 * If ~90% of the cpu_power is still there, we're good.
4765 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4772 * Group imbalance indicates (and tries to solve) the problem where balancing
4773 * groups is inadequate due to tsk_cpus_allowed() constraints.
4775 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4776 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4779 * { 0 1 2 3 } { 4 5 6 7 }
4782 * If we were to balance group-wise we'd place two tasks in the first group and
4783 * two tasks in the second group. Clearly this is undesired as it will overload
4784 * cpu 3 and leave one of the cpus in the second group unused.
4786 * The current solution to this issue is detecting the skew in the first group
4787 * by noticing the lower domain failed to reach balance and had difficulty
4788 * moving tasks due to affinity constraints.
4790 * When this is so detected; this group becomes a candidate for busiest; see
4791 * update_sd_pick_busiest(). And calculcate_imbalance() and
4792 * find_busiest_group() avoid some of the usual balance conditions to allow it
4793 * to create an effective group imbalance.
4795 * This is a somewhat tricky proposition since the next run might not find the
4796 * group imbalance and decide the groups need to be balanced again. A most
4797 * subtle and fragile situation.
4800 static inline int sg_imbalanced(struct sched_group *group)
4802 return group->sgp->imbalance;
4806 * Compute the group capacity.
4808 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4809 * first dividing out the smt factor and computing the actual number of cores
4810 * and limit power unit capacity with that.
4812 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4814 unsigned int capacity, smt, cpus;
4815 unsigned int power, power_orig;
4817 power = group->sgp->power;
4818 power_orig = group->sgp->power_orig;
4819 cpus = group->group_weight;
4821 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4822 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4823 capacity = cpus / smt; /* cores */
4825 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4827 capacity = fix_small_capacity(env->sd, group);
4833 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4834 * @env: The load balancing environment.
4835 * @group: sched_group whose statistics are to be updated.
4836 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4837 * @local_group: Does group contain this_cpu.
4838 * @sgs: variable to hold the statistics for this group.
4840 static inline void update_sg_lb_stats(struct lb_env *env,
4841 struct sched_group *group, int load_idx,
4842 int local_group, struct sg_lb_stats *sgs)
4844 unsigned long nr_running;
4848 memset(sgs, 0, sizeof(*sgs));
4850 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4851 struct rq *rq = cpu_rq(i);
4853 nr_running = rq->nr_running;
4855 /* Bias balancing toward cpus of our domain */
4857 load = target_load(i, load_idx);
4859 load = source_load(i, load_idx);
4861 sgs->group_load += load;
4862 sgs->sum_nr_running += nr_running;
4863 sgs->sum_weighted_load += weighted_cpuload(i);
4868 /* Adjust by relative CPU power of the group */
4869 sgs->group_power = group->sgp->power;
4870 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4872 if (sgs->sum_nr_running)
4873 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4875 sgs->group_weight = group->group_weight;
4877 sgs->group_imb = sg_imbalanced(group);
4878 sgs->group_capacity = sg_capacity(env, group);
4880 if (sgs->group_capacity > sgs->sum_nr_running)
4881 sgs->group_has_capacity = 1;
4885 * update_sd_pick_busiest - return 1 on busiest group
4886 * @env: The load balancing environment.
4887 * @sds: sched_domain statistics
4888 * @sg: sched_group candidate to be checked for being the busiest
4889 * @sgs: sched_group statistics
4891 * Determine if @sg is a busier group than the previously selected
4894 * Return: %true if @sg is a busier group than the previously selected
4895 * busiest group. %false otherwise.
4897 static bool update_sd_pick_busiest(struct lb_env *env,
4898 struct sd_lb_stats *sds,
4899 struct sched_group *sg,
4900 struct sg_lb_stats *sgs)
4902 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4905 if (sgs->sum_nr_running > sgs->group_capacity)
4912 * ASYM_PACKING needs to move all the work to the lowest
4913 * numbered CPUs in the group, therefore mark all groups
4914 * higher than ourself as busy.
4916 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4917 env->dst_cpu < group_first_cpu(sg)) {
4921 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4929 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4930 * @env: The load balancing environment.
4931 * @balance: Should we balance.
4932 * @sds: variable to hold the statistics for this sched_domain.
4934 static inline void update_sd_lb_stats(struct lb_env *env,
4935 struct sd_lb_stats *sds)
4937 struct sched_domain *child = env->sd->child;
4938 struct sched_group *sg = env->sd->groups;
4939 struct sg_lb_stats tmp_sgs;
4940 int load_idx, prefer_sibling = 0;
4942 if (child && child->flags & SD_PREFER_SIBLING)
4945 load_idx = get_sd_load_idx(env->sd, env->idle);
4948 struct sg_lb_stats *sgs = &tmp_sgs;
4951 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4954 sgs = &sds->local_stat;
4956 if (env->idle != CPU_NEWLY_IDLE ||
4957 time_after_eq(jiffies, sg->sgp->next_update))
4958 update_group_power(env->sd, env->dst_cpu);
4961 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4967 * In case the child domain prefers tasks go to siblings
4968 * first, lower the sg capacity to one so that we'll try
4969 * and move all the excess tasks away. We lower the capacity
4970 * of a group only if the local group has the capacity to fit
4971 * these excess tasks, i.e. nr_running < group_capacity. The
4972 * extra check prevents the case where you always pull from the
4973 * heaviest group when it is already under-utilized (possible
4974 * with a large weight task outweighs the tasks on the system).
4976 if (prefer_sibling && sds->local &&
4977 sds->local_stat.group_has_capacity)
4978 sgs->group_capacity = min(sgs->group_capacity, 1U);
4980 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4982 sds->busiest_stat = *sgs;
4986 /* Now, start updating sd_lb_stats */
4987 sds->total_load += sgs->group_load;
4988 sds->total_pwr += sgs->group_power;
4991 } while (sg != env->sd->groups);
4995 * check_asym_packing - Check to see if the group is packed into the
4998 * This is primarily intended to used at the sibling level. Some
4999 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5000 * case of POWER7, it can move to lower SMT modes only when higher
5001 * threads are idle. When in lower SMT modes, the threads will
5002 * perform better since they share less core resources. Hence when we
5003 * have idle threads, we want them to be the higher ones.
5005 * This packing function is run on idle threads. It checks to see if
5006 * the busiest CPU in this domain (core in the P7 case) has a higher
5007 * CPU number than the packing function is being run on. Here we are
5008 * assuming lower CPU number will be equivalent to lower a SMT thread
5011 * Return: 1 when packing is required and a task should be moved to
5012 * this CPU. The amount of the imbalance is returned in *imbalance.
5014 * @env: The load balancing environment.
5015 * @sds: Statistics of the sched_domain which is to be packed
5017 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5021 if (!(env->sd->flags & SD_ASYM_PACKING))
5027 busiest_cpu = group_first_cpu(sds->busiest);
5028 if (env->dst_cpu > busiest_cpu)
5031 env->imbalance = DIV_ROUND_CLOSEST(
5032 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5039 * fix_small_imbalance - Calculate the minor imbalance that exists
5040 * amongst the groups of a sched_domain, during
5042 * @env: The load balancing environment.
5043 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5046 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5048 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5049 unsigned int imbn = 2;
5050 unsigned long scaled_busy_load_per_task;
5051 struct sg_lb_stats *local, *busiest;
5053 local = &sds->local_stat;
5054 busiest = &sds->busiest_stat;
5056 if (!local->sum_nr_running)
5057 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5058 else if (busiest->load_per_task > local->load_per_task)
5061 scaled_busy_load_per_task =
5062 (busiest->load_per_task * SCHED_POWER_SCALE) /
5063 busiest->group_power;
5065 if (busiest->avg_load + scaled_busy_load_per_task >=
5066 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5067 env->imbalance = busiest->load_per_task;
5072 * OK, we don't have enough imbalance to justify moving tasks,
5073 * however we may be able to increase total CPU power used by
5077 pwr_now += busiest->group_power *
5078 min(busiest->load_per_task, busiest->avg_load);
5079 pwr_now += local->group_power *
5080 min(local->load_per_task, local->avg_load);
5081 pwr_now /= SCHED_POWER_SCALE;
5083 /* Amount of load we'd subtract */
5084 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5085 busiest->group_power;
5086 if (busiest->avg_load > tmp) {
5087 pwr_move += busiest->group_power *
5088 min(busiest->load_per_task,
5089 busiest->avg_load - tmp);
5092 /* Amount of load we'd add */
5093 if (busiest->avg_load * busiest->group_power <
5094 busiest->load_per_task * SCHED_POWER_SCALE) {
5095 tmp = (busiest->avg_load * busiest->group_power) /
5098 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5101 pwr_move += local->group_power *
5102 min(local->load_per_task, local->avg_load + tmp);
5103 pwr_move /= SCHED_POWER_SCALE;
5105 /* Move if we gain throughput */
5106 if (pwr_move > pwr_now)
5107 env->imbalance = busiest->load_per_task;
5111 * calculate_imbalance - Calculate the amount of imbalance present within the
5112 * groups of a given sched_domain during load balance.
5113 * @env: load balance environment
5114 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5116 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5118 unsigned long max_pull, load_above_capacity = ~0UL;
5119 struct sg_lb_stats *local, *busiest;
5121 local = &sds->local_stat;
5122 busiest = &sds->busiest_stat;
5124 if (busiest->group_imb) {
5126 * In the group_imb case we cannot rely on group-wide averages
5127 * to ensure cpu-load equilibrium, look at wider averages. XXX
5129 busiest->load_per_task =
5130 min(busiest->load_per_task, sds->avg_load);
5134 * In the presence of smp nice balancing, certain scenarios can have
5135 * max load less than avg load(as we skip the groups at or below
5136 * its cpu_power, while calculating max_load..)
5138 if (busiest->avg_load <= sds->avg_load ||
5139 local->avg_load >= sds->avg_load) {
5141 return fix_small_imbalance(env, sds);
5144 if (!busiest->group_imb) {
5146 * Don't want to pull so many tasks that a group would go idle.
5147 * Except of course for the group_imb case, since then we might
5148 * have to drop below capacity to reach cpu-load equilibrium.
5150 load_above_capacity =
5151 (busiest->sum_nr_running - busiest->group_capacity);
5153 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5154 load_above_capacity /= busiest->group_power;
5158 * We're trying to get all the cpus to the average_load, so we don't
5159 * want to push ourselves above the average load, nor do we wish to
5160 * reduce the max loaded cpu below the average load. At the same time,
5161 * we also don't want to reduce the group load below the group capacity
5162 * (so that we can implement power-savings policies etc). Thus we look
5163 * for the minimum possible imbalance.
5165 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5167 /* How much load to actually move to equalise the imbalance */
5168 env->imbalance = min(
5169 max_pull * busiest->group_power,
5170 (sds->avg_load - local->avg_load) * local->group_power
5171 ) / SCHED_POWER_SCALE;
5174 * if *imbalance is less than the average load per runnable task
5175 * there is no guarantee that any tasks will be moved so we'll have
5176 * a think about bumping its value to force at least one task to be
5179 if (env->imbalance < busiest->load_per_task)
5180 return fix_small_imbalance(env, sds);
5183 /******* find_busiest_group() helpers end here *********************/
5186 * find_busiest_group - Returns the busiest group within the sched_domain
5187 * if there is an imbalance. If there isn't an imbalance, and
5188 * the user has opted for power-savings, it returns a group whose
5189 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5190 * such a group exists.
5192 * Also calculates the amount of weighted load which should be moved
5193 * to restore balance.
5195 * @env: The load balancing environment.
5197 * Return: - The busiest group if imbalance exists.
5198 * - If no imbalance and user has opted for power-savings balance,
5199 * return the least loaded group whose CPUs can be
5200 * put to idle by rebalancing its tasks onto our group.
5202 static struct sched_group *find_busiest_group(struct lb_env *env)
5204 struct sg_lb_stats *local, *busiest;
5205 struct sd_lb_stats sds;
5207 init_sd_lb_stats(&sds);
5210 * Compute the various statistics relavent for load balancing at
5213 update_sd_lb_stats(env, &sds);
5214 local = &sds.local_stat;
5215 busiest = &sds.busiest_stat;
5217 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5218 check_asym_packing(env, &sds))
5221 /* There is no busy sibling group to pull tasks from */
5222 if (!sds.busiest || busiest->sum_nr_running == 0)
5225 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5228 * If the busiest group is imbalanced the below checks don't
5229 * work because they assume all things are equal, which typically
5230 * isn't true due to cpus_allowed constraints and the like.
5232 if (busiest->group_imb)
5235 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5236 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5237 !busiest->group_has_capacity)
5241 * If the local group is more busy than the selected busiest group
5242 * don't try and pull any tasks.
5244 if (local->avg_load >= busiest->avg_load)
5248 * Don't pull any tasks if this group is already above the domain
5251 if (local->avg_load >= sds.avg_load)
5254 if (env->idle == CPU_IDLE) {
5256 * This cpu is idle. If the busiest group load doesn't
5257 * have more tasks than the number of available cpu's and
5258 * there is no imbalance between this and busiest group
5259 * wrt to idle cpu's, it is balanced.
5261 if ((local->idle_cpus < busiest->idle_cpus) &&
5262 busiest->sum_nr_running <= busiest->group_weight)
5266 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5267 * imbalance_pct to be conservative.
5269 if (100 * busiest->avg_load <=
5270 env->sd->imbalance_pct * local->avg_load)
5275 /* Looks like there is an imbalance. Compute it */
5276 calculate_imbalance(env, &sds);
5285 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5287 static struct rq *find_busiest_queue(struct lb_env *env,
5288 struct sched_group *group)
5290 struct rq *busiest = NULL, *rq;
5291 unsigned long busiest_load = 0, busiest_power = 1;
5294 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5295 unsigned long power = power_of(i);
5296 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5301 capacity = fix_small_capacity(env->sd, group);
5304 wl = weighted_cpuload(i);
5307 * When comparing with imbalance, use weighted_cpuload()
5308 * which is not scaled with the cpu power.
5310 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5314 * For the load comparisons with the other cpu's, consider
5315 * the weighted_cpuload() scaled with the cpu power, so that
5316 * the load can be moved away from the cpu that is potentially
5317 * running at a lower capacity.
5319 * Thus we're looking for max(wl_i / power_i), crosswise
5320 * multiplication to rid ourselves of the division works out
5321 * to: wl_i * power_j > wl_j * power_i; where j is our
5324 if (wl * busiest_power > busiest_load * power) {
5326 busiest_power = power;
5335 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5336 * so long as it is large enough.
5338 #define MAX_PINNED_INTERVAL 512
5340 /* Working cpumask for load_balance and load_balance_newidle. */
5341 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5343 static int need_active_balance(struct lb_env *env)
5345 struct sched_domain *sd = env->sd;
5347 if (env->idle == CPU_NEWLY_IDLE) {
5350 * ASYM_PACKING needs to force migrate tasks from busy but
5351 * higher numbered CPUs in order to pack all tasks in the
5352 * lowest numbered CPUs.
5354 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5358 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5361 static int active_load_balance_cpu_stop(void *data);
5363 static int should_we_balance(struct lb_env *env)
5365 struct sched_group *sg = env->sd->groups;
5366 struct cpumask *sg_cpus, *sg_mask;
5367 int cpu, balance_cpu = -1;
5370 * In the newly idle case, we will allow all the cpu's
5371 * to do the newly idle load balance.
5373 if (env->idle == CPU_NEWLY_IDLE)
5376 sg_cpus = sched_group_cpus(sg);
5377 sg_mask = sched_group_mask(sg);
5378 /* Try to find first idle cpu */
5379 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5380 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5387 if (balance_cpu == -1)
5388 balance_cpu = group_balance_cpu(sg);
5391 * First idle cpu or the first cpu(busiest) in this sched group
5392 * is eligible for doing load balancing at this and above domains.
5394 return balance_cpu == env->dst_cpu;
5398 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5399 * tasks if there is an imbalance.
5401 static int load_balance(int this_cpu, struct rq *this_rq,
5402 struct sched_domain *sd, enum cpu_idle_type idle,
5403 int *continue_balancing)
5405 int ld_moved, cur_ld_moved, active_balance = 0;
5406 struct sched_domain *sd_parent = sd->parent;
5407 struct sched_group *group;
5409 unsigned long flags;
5410 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5412 struct lb_env env = {
5414 .dst_cpu = this_cpu,
5416 .dst_grpmask = sched_group_cpus(sd->groups),
5418 .loop_break = sched_nr_migrate_break,
5423 * For NEWLY_IDLE load_balancing, we don't need to consider
5424 * other cpus in our group
5426 if (idle == CPU_NEWLY_IDLE)
5427 env.dst_grpmask = NULL;
5429 cpumask_copy(cpus, cpu_active_mask);
5431 schedstat_inc(sd, lb_count[idle]);
5434 if (!should_we_balance(&env)) {
5435 *continue_balancing = 0;
5439 group = find_busiest_group(&env);
5441 schedstat_inc(sd, lb_nobusyg[idle]);
5445 busiest = find_busiest_queue(&env, group);
5447 schedstat_inc(sd, lb_nobusyq[idle]);
5451 BUG_ON(busiest == env.dst_rq);
5453 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5456 if (busiest->nr_running > 1) {
5458 * Attempt to move tasks. If find_busiest_group has found
5459 * an imbalance but busiest->nr_running <= 1, the group is
5460 * still unbalanced. ld_moved simply stays zero, so it is
5461 * correctly treated as an imbalance.
5463 env.flags |= LBF_ALL_PINNED;
5464 env.src_cpu = busiest->cpu;
5465 env.src_rq = busiest;
5466 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5469 local_irq_save(flags);
5470 double_rq_lock(env.dst_rq, busiest);
5473 * cur_ld_moved - load moved in current iteration
5474 * ld_moved - cumulative load moved across iterations
5476 cur_ld_moved = move_tasks(&env);
5477 ld_moved += cur_ld_moved;
5478 double_rq_unlock(env.dst_rq, busiest);
5479 local_irq_restore(flags);
5482 * some other cpu did the load balance for us.
5484 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5485 resched_cpu(env.dst_cpu);
5487 if (env.flags & LBF_NEED_BREAK) {
5488 env.flags &= ~LBF_NEED_BREAK;
5493 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5494 * us and move them to an alternate dst_cpu in our sched_group
5495 * where they can run. The upper limit on how many times we
5496 * iterate on same src_cpu is dependent on number of cpus in our
5499 * This changes load balance semantics a bit on who can move
5500 * load to a given_cpu. In addition to the given_cpu itself
5501 * (or a ilb_cpu acting on its behalf where given_cpu is
5502 * nohz-idle), we now have balance_cpu in a position to move
5503 * load to given_cpu. In rare situations, this may cause
5504 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5505 * _independently_ and at _same_ time to move some load to
5506 * given_cpu) causing exceess load to be moved to given_cpu.
5507 * This however should not happen so much in practice and
5508 * moreover subsequent load balance cycles should correct the
5509 * excess load moved.
5511 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5513 /* Prevent to re-select dst_cpu via env's cpus */
5514 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5516 env.dst_rq = cpu_rq(env.new_dst_cpu);
5517 env.dst_cpu = env.new_dst_cpu;
5518 env.flags &= ~LBF_DST_PINNED;
5520 env.loop_break = sched_nr_migrate_break;
5523 * Go back to "more_balance" rather than "redo" since we
5524 * need to continue with same src_cpu.
5530 * We failed to reach balance because of affinity.
5533 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5535 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5536 *group_imbalance = 1;
5537 } else if (*group_imbalance)
5538 *group_imbalance = 0;
5541 /* All tasks on this runqueue were pinned by CPU affinity */
5542 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5543 cpumask_clear_cpu(cpu_of(busiest), cpus);
5544 if (!cpumask_empty(cpus)) {
5546 env.loop_break = sched_nr_migrate_break;
5554 schedstat_inc(sd, lb_failed[idle]);
5556 * Increment the failure counter only on periodic balance.
5557 * We do not want newidle balance, which can be very
5558 * frequent, pollute the failure counter causing
5559 * excessive cache_hot migrations and active balances.
5561 if (idle != CPU_NEWLY_IDLE)
5562 sd->nr_balance_failed++;
5564 if (need_active_balance(&env)) {
5565 raw_spin_lock_irqsave(&busiest->lock, flags);
5567 /* don't kick the active_load_balance_cpu_stop,
5568 * if the curr task on busiest cpu can't be
5571 if (!cpumask_test_cpu(this_cpu,
5572 tsk_cpus_allowed(busiest->curr))) {
5573 raw_spin_unlock_irqrestore(&busiest->lock,
5575 env.flags |= LBF_ALL_PINNED;
5576 goto out_one_pinned;
5580 * ->active_balance synchronizes accesses to
5581 * ->active_balance_work. Once set, it's cleared
5582 * only after active load balance is finished.
5584 if (!busiest->active_balance) {
5585 busiest->active_balance = 1;
5586 busiest->push_cpu = this_cpu;
5589 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5591 if (active_balance) {
5592 stop_one_cpu_nowait(cpu_of(busiest),
5593 active_load_balance_cpu_stop, busiest,
5594 &busiest->active_balance_work);
5598 * We've kicked active balancing, reset the failure
5601 sd->nr_balance_failed = sd->cache_nice_tries+1;
5604 sd->nr_balance_failed = 0;
5606 if (likely(!active_balance)) {
5607 /* We were unbalanced, so reset the balancing interval */
5608 sd->balance_interval = sd->min_interval;
5611 * If we've begun active balancing, start to back off. This
5612 * case may not be covered by the all_pinned logic if there
5613 * is only 1 task on the busy runqueue (because we don't call
5616 if (sd->balance_interval < sd->max_interval)
5617 sd->balance_interval *= 2;
5623 schedstat_inc(sd, lb_balanced[idle]);
5625 sd->nr_balance_failed = 0;
5628 /* tune up the balancing interval */
5629 if (((env.flags & LBF_ALL_PINNED) &&
5630 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5631 (sd->balance_interval < sd->max_interval))
5632 sd->balance_interval *= 2;
5640 * idle_balance is called by schedule() if this_cpu is about to become
5641 * idle. Attempts to pull tasks from other CPUs.
5643 void idle_balance(int this_cpu, struct rq *this_rq)
5645 struct sched_domain *sd;
5646 int pulled_task = 0;
5647 unsigned long next_balance = jiffies + HZ;
5650 this_rq->idle_stamp = rq_clock(this_rq);
5652 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5656 * Drop the rq->lock, but keep IRQ/preempt disabled.
5658 raw_spin_unlock(&this_rq->lock);
5660 update_blocked_averages(this_cpu);
5662 for_each_domain(this_cpu, sd) {
5663 unsigned long interval;
5664 int continue_balancing = 1;
5665 u64 t0, domain_cost;
5667 if (!(sd->flags & SD_LOAD_BALANCE))
5670 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5673 if (sd->flags & SD_BALANCE_NEWIDLE) {
5674 t0 = sched_clock_cpu(this_cpu);
5676 /* If we've pulled tasks over stop searching: */
5677 pulled_task = load_balance(this_cpu, this_rq,
5679 &continue_balancing);
5681 domain_cost = sched_clock_cpu(this_cpu) - t0;
5682 if (domain_cost > sd->max_newidle_lb_cost)
5683 sd->max_newidle_lb_cost = domain_cost;
5685 curr_cost += domain_cost;
5688 interval = msecs_to_jiffies(sd->balance_interval);
5689 if (time_after(next_balance, sd->last_balance + interval))
5690 next_balance = sd->last_balance + interval;
5692 this_rq->idle_stamp = 0;
5698 raw_spin_lock(&this_rq->lock);
5700 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5702 * We are going idle. next_balance may be set based on
5703 * a busy processor. So reset next_balance.
5705 this_rq->next_balance = next_balance;
5708 if (curr_cost > this_rq->max_idle_balance_cost)
5709 this_rq->max_idle_balance_cost = curr_cost;
5713 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5714 * running tasks off the busiest CPU onto idle CPUs. It requires at
5715 * least 1 task to be running on each physical CPU where possible, and
5716 * avoids physical / logical imbalances.
5718 static int active_load_balance_cpu_stop(void *data)
5720 struct rq *busiest_rq = data;
5721 int busiest_cpu = cpu_of(busiest_rq);
5722 int target_cpu = busiest_rq->push_cpu;
5723 struct rq *target_rq = cpu_rq(target_cpu);
5724 struct sched_domain *sd;
5726 raw_spin_lock_irq(&busiest_rq->lock);
5728 /* make sure the requested cpu hasn't gone down in the meantime */
5729 if (unlikely(busiest_cpu != smp_processor_id() ||
5730 !busiest_rq->active_balance))
5733 /* Is there any task to move? */
5734 if (busiest_rq->nr_running <= 1)
5738 * This condition is "impossible", if it occurs
5739 * we need to fix it. Originally reported by
5740 * Bjorn Helgaas on a 128-cpu setup.
5742 BUG_ON(busiest_rq == target_rq);
5744 /* move a task from busiest_rq to target_rq */
5745 double_lock_balance(busiest_rq, target_rq);
5747 /* Search for an sd spanning us and the target CPU. */
5749 for_each_domain(target_cpu, sd) {
5750 if ((sd->flags & SD_LOAD_BALANCE) &&
5751 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5756 struct lb_env env = {
5758 .dst_cpu = target_cpu,
5759 .dst_rq = target_rq,
5760 .src_cpu = busiest_rq->cpu,
5761 .src_rq = busiest_rq,
5765 schedstat_inc(sd, alb_count);
5767 if (move_one_task(&env))
5768 schedstat_inc(sd, alb_pushed);
5770 schedstat_inc(sd, alb_failed);
5773 double_unlock_balance(busiest_rq, target_rq);
5775 busiest_rq->active_balance = 0;
5776 raw_spin_unlock_irq(&busiest_rq->lock);
5780 #ifdef CONFIG_NO_HZ_COMMON
5782 * idle load balancing details
5783 * - When one of the busy CPUs notice that there may be an idle rebalancing
5784 * needed, they will kick the idle load balancer, which then does idle
5785 * load balancing for all the idle CPUs.
5788 cpumask_var_t idle_cpus_mask;
5790 unsigned long next_balance; /* in jiffy units */
5791 } nohz ____cacheline_aligned;
5793 static inline int find_new_ilb(int call_cpu)
5795 int ilb = cpumask_first(nohz.idle_cpus_mask);
5797 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5804 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5805 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5806 * CPU (if there is one).
5808 static void nohz_balancer_kick(int cpu)
5812 nohz.next_balance++;
5814 ilb_cpu = find_new_ilb(cpu);
5816 if (ilb_cpu >= nr_cpu_ids)
5819 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5822 * Use smp_send_reschedule() instead of resched_cpu().
5823 * This way we generate a sched IPI on the target cpu which
5824 * is idle. And the softirq performing nohz idle load balance
5825 * will be run before returning from the IPI.
5827 smp_send_reschedule(ilb_cpu);
5831 static inline void nohz_balance_exit_idle(int cpu)
5833 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5834 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5835 atomic_dec(&nohz.nr_cpus);
5836 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5840 static inline void set_cpu_sd_state_busy(void)
5842 struct sched_domain *sd;
5845 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5847 if (!sd || !sd->nohz_idle)
5851 for (; sd; sd = sd->parent)
5852 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5857 void set_cpu_sd_state_idle(void)
5859 struct sched_domain *sd;
5862 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5864 if (!sd || sd->nohz_idle)
5868 for (; sd; sd = sd->parent)
5869 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5875 * This routine will record that the cpu is going idle with tick stopped.
5876 * This info will be used in performing idle load balancing in the future.
5878 void nohz_balance_enter_idle(int cpu)
5881 * If this cpu is going down, then nothing needs to be done.
5883 if (!cpu_active(cpu))
5886 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5889 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5890 atomic_inc(&nohz.nr_cpus);
5891 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5894 static int sched_ilb_notifier(struct notifier_block *nfb,
5895 unsigned long action, void *hcpu)
5897 switch (action & ~CPU_TASKS_FROZEN) {
5899 nohz_balance_exit_idle(smp_processor_id());
5907 static DEFINE_SPINLOCK(balancing);
5910 * Scale the max load_balance interval with the number of CPUs in the system.
5911 * This trades load-balance latency on larger machines for less cross talk.
5913 void update_max_interval(void)
5915 max_load_balance_interval = HZ*num_online_cpus()/10;
5919 * It checks each scheduling domain to see if it is due to be balanced,
5920 * and initiates a balancing operation if so.
5922 * Balancing parameters are set up in init_sched_domains.
5924 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5926 int continue_balancing = 1;
5927 struct rq *rq = cpu_rq(cpu);
5928 unsigned long interval;
5929 struct sched_domain *sd;
5930 /* Earliest time when we have to do rebalance again */
5931 unsigned long next_balance = jiffies + 60*HZ;
5932 int update_next_balance = 0;
5933 int need_serialize, need_decay = 0;
5936 update_blocked_averages(cpu);
5939 for_each_domain(cpu, sd) {
5941 * Decay the newidle max times here because this is a regular
5942 * visit to all the domains. Decay ~1% per second.
5944 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5945 sd->max_newidle_lb_cost =
5946 (sd->max_newidle_lb_cost * 253) / 256;
5947 sd->next_decay_max_lb_cost = jiffies + HZ;
5950 max_cost += sd->max_newidle_lb_cost;
5952 if (!(sd->flags & SD_LOAD_BALANCE))
5956 * Stop the load balance at this level. There is another
5957 * CPU in our sched group which is doing load balancing more
5960 if (!continue_balancing) {
5966 interval = sd->balance_interval;
5967 if (idle != CPU_IDLE)
5968 interval *= sd->busy_factor;
5970 /* scale ms to jiffies */
5971 interval = msecs_to_jiffies(interval);
5972 interval = clamp(interval, 1UL, max_load_balance_interval);
5974 need_serialize = sd->flags & SD_SERIALIZE;
5976 if (need_serialize) {
5977 if (!spin_trylock(&balancing))
5981 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5982 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5984 * The LBF_DST_PINNED logic could have changed
5985 * env->dst_cpu, so we can't know our idle
5986 * state even if we migrated tasks. Update it.
5988 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5990 sd->last_balance = jiffies;
5993 spin_unlock(&balancing);
5995 if (time_after(next_balance, sd->last_balance + interval)) {
5996 next_balance = sd->last_balance + interval;
5997 update_next_balance = 1;
6002 * Ensure the rq-wide value also decays but keep it at a
6003 * reasonable floor to avoid funnies with rq->avg_idle.
6005 rq->max_idle_balance_cost =
6006 max((u64)sysctl_sched_migration_cost, max_cost);
6011 * next_balance will be updated only when there is a need.
6012 * When the cpu is attached to null domain for ex, it will not be
6015 if (likely(update_next_balance))
6016 rq->next_balance = next_balance;
6019 #ifdef CONFIG_NO_HZ_COMMON
6021 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6022 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6024 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6026 struct rq *this_rq = cpu_rq(this_cpu);
6030 if (idle != CPU_IDLE ||
6031 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6034 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6035 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6039 * If this cpu gets work to do, stop the load balancing
6040 * work being done for other cpus. Next load
6041 * balancing owner will pick it up.
6046 rq = cpu_rq(balance_cpu);
6048 raw_spin_lock_irq(&rq->lock);
6049 update_rq_clock(rq);
6050 update_idle_cpu_load(rq);
6051 raw_spin_unlock_irq(&rq->lock);
6053 rebalance_domains(balance_cpu, CPU_IDLE);
6055 if (time_after(this_rq->next_balance, rq->next_balance))
6056 this_rq->next_balance = rq->next_balance;
6058 nohz.next_balance = this_rq->next_balance;
6060 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6064 * Current heuristic for kicking the idle load balancer in the presence
6065 * of an idle cpu is the system.
6066 * - This rq has more than one task.
6067 * - At any scheduler domain level, this cpu's scheduler group has multiple
6068 * busy cpu's exceeding the group's power.
6069 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6070 * domain span are idle.
6072 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6074 unsigned long now = jiffies;
6075 struct sched_domain *sd;
6077 if (unlikely(idle_cpu(cpu)))
6081 * We may be recently in ticked or tickless idle mode. At the first
6082 * busy tick after returning from idle, we will update the busy stats.
6084 set_cpu_sd_state_busy();
6085 nohz_balance_exit_idle(cpu);
6088 * None are in tickless mode and hence no need for NOHZ idle load
6091 if (likely(!atomic_read(&nohz.nr_cpus)))
6094 if (time_before(now, nohz.next_balance))
6097 if (rq->nr_running >= 2)
6101 for_each_domain(cpu, sd) {
6102 struct sched_group *sg = sd->groups;
6103 struct sched_group_power *sgp = sg->sgp;
6104 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6106 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6107 goto need_kick_unlock;
6109 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6110 && (cpumask_first_and(nohz.idle_cpus_mask,
6111 sched_domain_span(sd)) < cpu))
6112 goto need_kick_unlock;
6114 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6126 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6130 * run_rebalance_domains is triggered when needed from the scheduler tick.
6131 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6133 static void run_rebalance_domains(struct softirq_action *h)
6135 int this_cpu = smp_processor_id();
6136 struct rq *this_rq = cpu_rq(this_cpu);
6137 enum cpu_idle_type idle = this_rq->idle_balance ?
6138 CPU_IDLE : CPU_NOT_IDLE;
6140 rebalance_domains(this_cpu, idle);
6143 * If this cpu has a pending nohz_balance_kick, then do the
6144 * balancing on behalf of the other idle cpus whose ticks are
6147 nohz_idle_balance(this_cpu, idle);
6150 static inline int on_null_domain(int cpu)
6152 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6156 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6158 void trigger_load_balance(struct rq *rq, int cpu)
6160 /* Don't need to rebalance while attached to NULL domain */
6161 if (time_after_eq(jiffies, rq->next_balance) &&
6162 likely(!on_null_domain(cpu)))
6163 raise_softirq(SCHED_SOFTIRQ);
6164 #ifdef CONFIG_NO_HZ_COMMON
6165 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6166 nohz_balancer_kick(cpu);
6170 static void rq_online_fair(struct rq *rq)
6175 static void rq_offline_fair(struct rq *rq)
6179 /* Ensure any throttled groups are reachable by pick_next_task */
6180 unthrottle_offline_cfs_rqs(rq);
6183 #endif /* CONFIG_SMP */
6186 * scheduler tick hitting a task of our scheduling class:
6188 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6190 struct cfs_rq *cfs_rq;
6191 struct sched_entity *se = &curr->se;
6193 for_each_sched_entity(se) {
6194 cfs_rq = cfs_rq_of(se);
6195 entity_tick(cfs_rq, se, queued);
6198 if (numabalancing_enabled)
6199 task_tick_numa(rq, curr);
6201 update_rq_runnable_avg(rq, 1);
6205 * called on fork with the child task as argument from the parent's context
6206 * - child not yet on the tasklist
6207 * - preemption disabled
6209 static void task_fork_fair(struct task_struct *p)
6211 struct cfs_rq *cfs_rq;
6212 struct sched_entity *se = &p->se, *curr;
6213 int this_cpu = smp_processor_id();
6214 struct rq *rq = this_rq();
6215 unsigned long flags;
6217 raw_spin_lock_irqsave(&rq->lock, flags);
6219 update_rq_clock(rq);
6221 cfs_rq = task_cfs_rq(current);
6222 curr = cfs_rq->curr;
6225 * Not only the cpu but also the task_group of the parent might have
6226 * been changed after parent->se.parent,cfs_rq were copied to
6227 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6228 * of child point to valid ones.
6231 __set_task_cpu(p, this_cpu);
6234 update_curr(cfs_rq);
6237 se->vruntime = curr->vruntime;
6238 place_entity(cfs_rq, se, 1);
6240 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6242 * Upon rescheduling, sched_class::put_prev_task() will place
6243 * 'current' within the tree based on its new key value.
6245 swap(curr->vruntime, se->vruntime);
6246 resched_task(rq->curr);
6249 se->vruntime -= cfs_rq->min_vruntime;
6251 raw_spin_unlock_irqrestore(&rq->lock, flags);
6255 * Priority of the task has changed. Check to see if we preempt
6259 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6265 * Reschedule if we are currently running on this runqueue and
6266 * our priority decreased, or if we are not currently running on
6267 * this runqueue and our priority is higher than the current's
6269 if (rq->curr == p) {
6270 if (p->prio > oldprio)
6271 resched_task(rq->curr);
6273 check_preempt_curr(rq, p, 0);
6276 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6278 struct sched_entity *se = &p->se;
6279 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6282 * Ensure the task's vruntime is normalized, so that when its
6283 * switched back to the fair class the enqueue_entity(.flags=0) will
6284 * do the right thing.
6286 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6287 * have normalized the vruntime, if it was !on_rq, then only when
6288 * the task is sleeping will it still have non-normalized vruntime.
6290 if (!se->on_rq && p->state != TASK_RUNNING) {
6292 * Fix up our vruntime so that the current sleep doesn't
6293 * cause 'unlimited' sleep bonus.
6295 place_entity(cfs_rq, se, 0);
6296 se->vruntime -= cfs_rq->min_vruntime;
6301 * Remove our load from contribution when we leave sched_fair
6302 * and ensure we don't carry in an old decay_count if we
6305 if (se->avg.decay_count) {
6306 __synchronize_entity_decay(se);
6307 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6313 * We switched to the sched_fair class.
6315 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6321 * We were most likely switched from sched_rt, so
6322 * kick off the schedule if running, otherwise just see
6323 * if we can still preempt the current task.
6326 resched_task(rq->curr);
6328 check_preempt_curr(rq, p, 0);
6331 /* Account for a task changing its policy or group.
6333 * This routine is mostly called to set cfs_rq->curr field when a task
6334 * migrates between groups/classes.
6336 static void set_curr_task_fair(struct rq *rq)
6338 struct sched_entity *se = &rq->curr->se;
6340 for_each_sched_entity(se) {
6341 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6343 set_next_entity(cfs_rq, se);
6344 /* ensure bandwidth has been allocated on our new cfs_rq */
6345 account_cfs_rq_runtime(cfs_rq, 0);
6349 void init_cfs_rq(struct cfs_rq *cfs_rq)
6351 cfs_rq->tasks_timeline = RB_ROOT;
6352 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6353 #ifndef CONFIG_64BIT
6354 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6357 atomic64_set(&cfs_rq->decay_counter, 1);
6358 atomic_long_set(&cfs_rq->removed_load, 0);
6362 #ifdef CONFIG_FAIR_GROUP_SCHED
6363 static void task_move_group_fair(struct task_struct *p, int on_rq)
6365 struct cfs_rq *cfs_rq;
6367 * If the task was not on the rq at the time of this cgroup movement
6368 * it must have been asleep, sleeping tasks keep their ->vruntime
6369 * absolute on their old rq until wakeup (needed for the fair sleeper
6370 * bonus in place_entity()).
6372 * If it was on the rq, we've just 'preempted' it, which does convert
6373 * ->vruntime to a relative base.
6375 * Make sure both cases convert their relative position when migrating
6376 * to another cgroup's rq. This does somewhat interfere with the
6377 * fair sleeper stuff for the first placement, but who cares.
6380 * When !on_rq, vruntime of the task has usually NOT been normalized.
6381 * But there are some cases where it has already been normalized:
6383 * - Moving a forked child which is waiting for being woken up by
6384 * wake_up_new_task().
6385 * - Moving a task which has been woken up by try_to_wake_up() and
6386 * waiting for actually being woken up by sched_ttwu_pending().
6388 * To prevent boost or penalty in the new cfs_rq caused by delta
6389 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6391 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6395 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6396 set_task_rq(p, task_cpu(p));
6398 cfs_rq = cfs_rq_of(&p->se);
6399 p->se.vruntime += cfs_rq->min_vruntime;
6402 * migrate_task_rq_fair() will have removed our previous
6403 * contribution, but we must synchronize for ongoing future
6406 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6407 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6412 void free_fair_sched_group(struct task_group *tg)
6416 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6418 for_each_possible_cpu(i) {
6420 kfree(tg->cfs_rq[i]);
6429 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6431 struct cfs_rq *cfs_rq;
6432 struct sched_entity *se;
6435 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6438 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6442 tg->shares = NICE_0_LOAD;
6444 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6446 for_each_possible_cpu(i) {
6447 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6448 GFP_KERNEL, cpu_to_node(i));
6452 se = kzalloc_node(sizeof(struct sched_entity),
6453 GFP_KERNEL, cpu_to_node(i));
6457 init_cfs_rq(cfs_rq);
6458 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6469 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6471 struct rq *rq = cpu_rq(cpu);
6472 unsigned long flags;
6475 * Only empty task groups can be destroyed; so we can speculatively
6476 * check on_list without danger of it being re-added.
6478 if (!tg->cfs_rq[cpu]->on_list)
6481 raw_spin_lock_irqsave(&rq->lock, flags);
6482 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6483 raw_spin_unlock_irqrestore(&rq->lock, flags);
6486 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6487 struct sched_entity *se, int cpu,
6488 struct sched_entity *parent)
6490 struct rq *rq = cpu_rq(cpu);
6494 init_cfs_rq_runtime(cfs_rq);
6496 tg->cfs_rq[cpu] = cfs_rq;
6499 /* se could be NULL for root_task_group */
6504 se->cfs_rq = &rq->cfs;
6506 se->cfs_rq = parent->my_q;
6509 update_load_set(&se->load, 0);
6510 se->parent = parent;
6513 static DEFINE_MUTEX(shares_mutex);
6515 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6518 unsigned long flags;
6521 * We can't change the weight of the root cgroup.
6526 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6528 mutex_lock(&shares_mutex);
6529 if (tg->shares == shares)
6532 tg->shares = shares;
6533 for_each_possible_cpu(i) {
6534 struct rq *rq = cpu_rq(i);
6535 struct sched_entity *se;
6538 /* Propagate contribution to hierarchy */
6539 raw_spin_lock_irqsave(&rq->lock, flags);
6541 /* Possible calls to update_curr() need rq clock */
6542 update_rq_clock(rq);
6543 for_each_sched_entity(se)
6544 update_cfs_shares(group_cfs_rq(se));
6545 raw_spin_unlock_irqrestore(&rq->lock, flags);
6549 mutex_unlock(&shares_mutex);
6552 #else /* CONFIG_FAIR_GROUP_SCHED */
6554 void free_fair_sched_group(struct task_group *tg) { }
6556 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6561 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6563 #endif /* CONFIG_FAIR_GROUP_SCHED */
6566 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6568 struct sched_entity *se = &task->se;
6569 unsigned int rr_interval = 0;
6572 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6575 if (rq->cfs.load.weight)
6576 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6582 * All the scheduling class methods:
6584 const struct sched_class fair_sched_class = {
6585 .next = &idle_sched_class,
6586 .enqueue_task = enqueue_task_fair,
6587 .dequeue_task = dequeue_task_fair,
6588 .yield_task = yield_task_fair,
6589 .yield_to_task = yield_to_task_fair,
6591 .check_preempt_curr = check_preempt_wakeup,
6593 .pick_next_task = pick_next_task_fair,
6594 .put_prev_task = put_prev_task_fair,
6597 .select_task_rq = select_task_rq_fair,
6598 .migrate_task_rq = migrate_task_rq_fair,
6600 .rq_online = rq_online_fair,
6601 .rq_offline = rq_offline_fair,
6603 .task_waking = task_waking_fair,
6606 .set_curr_task = set_curr_task_fair,
6607 .task_tick = task_tick_fair,
6608 .task_fork = task_fork_fair,
6610 .prio_changed = prio_changed_fair,
6611 .switched_from = switched_from_fair,
6612 .switched_to = switched_to_fair,
6614 .get_rr_interval = get_rr_interval_fair,
6616 #ifdef CONFIG_FAIR_GROUP_SCHED
6617 .task_move_group = task_move_group_fair,
6621 #ifdef CONFIG_SCHED_DEBUG
6622 void print_cfs_stats(struct seq_file *m, int cpu)
6624 struct cfs_rq *cfs_rq;
6627 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6628 print_cfs_rq(m, cpu, cfs_rq);
6633 __init void init_sched_fair_class(void)
6636 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6638 #ifdef CONFIG_NO_HZ_COMMON
6639 nohz.next_balance = jiffies;
6640 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6641 cpu_notifier(sched_ilb_notifier, 0);