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 = 4;
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);
904 static unsigned long source_load(int cpu, int type);
905 static unsigned long target_load(int cpu, int type);
906 static unsigned long power_of(int cpu);
907 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
912 unsigned long faults;
915 struct task_numa_env {
916 struct task_struct *p;
918 int src_cpu, src_nid;
919 int dst_cpu, dst_nid;
921 struct numa_stats src_stats, dst_stats;
923 unsigned long best_load;
927 static int task_numa_migrate(struct task_struct *p)
929 int node_cpu = cpumask_first(cpumask_of_node(p->numa_preferred_nid));
930 struct task_numa_env env = {
932 .src_cpu = task_cpu(p),
933 .src_nid = cpu_to_node(task_cpu(p)),
935 .dst_nid = p->numa_preferred_nid,
936 .best_load = ULONG_MAX,
937 .best_cpu = task_cpu(p),
939 struct sched_domain *sd;
941 struct task_group *tg = task_group(p);
942 unsigned long weight;
944 int imbalance_pct, idx = -1;
947 * Find the lowest common scheduling domain covering the nodes of both
948 * the CPU the task is currently running on and the target NUMA node.
951 for_each_domain(env.src_cpu, sd) {
952 if (cpumask_test_cpu(node_cpu, sched_domain_span(sd))) {
954 * busy_idx is used for the load decision as it is the
955 * same index used by the regular load balancer for an
959 imbalance_pct = sd->imbalance_pct;
965 if (WARN_ON_ONCE(idx == -1))
969 * XXX the below is mostly nicked from wake_affine(); we should
970 * see about sharing a bit if at all possible; also it might want
971 * some per entity weight love.
973 weight = p->se.load.weight;
974 env.src_stats.load = source_load(env.src_cpu, idx);
975 env.src_stats.eff_load = 100 + (imbalance_pct - 100) / 2;
976 env.src_stats.eff_load *= power_of(env.src_cpu);
977 env.src_stats.eff_load *= env.src_stats.load + effective_load(tg, env.src_cpu, -weight, -weight);
979 for_each_cpu(cpu, cpumask_of_node(env.dst_nid)) {
981 env.dst_stats.load = target_load(cpu, idx);
983 /* If the CPU is idle, use it */
984 if (!env.dst_stats.load) {
989 /* Otherwise check the target CPU load */
990 env.dst_stats.eff_load = 100;
991 env.dst_stats.eff_load *= power_of(cpu);
992 env.dst_stats.eff_load *= env.dst_stats.load + effective_load(tg, cpu, weight, weight);
995 * Destination is considered balanced if the destination CPU is
996 * less loaded than the source CPU. Unfortunately there is a
997 * risk that a task running on a lightly loaded CPU will not
998 * migrate to its preferred node due to load imbalances.
1000 balanced = (env.dst_stats.eff_load <= env.src_stats.eff_load);
1004 if (env.dst_stats.eff_load < env.best_load) {
1005 env.best_load = env.dst_stats.eff_load;
1011 return migrate_task_to(p, env.best_cpu);
1014 /* Attempt to migrate a task to a CPU on the preferred node. */
1015 static void numa_migrate_preferred(struct task_struct *p)
1017 /* Success if task is already running on preferred CPU */
1018 p->numa_migrate_retry = 0;
1019 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid)
1022 /* This task has no NUMA fault statistics yet */
1023 if (unlikely(p->numa_preferred_nid == -1))
1026 /* Otherwise, try migrate to a CPU on the preferred node */
1027 if (task_numa_migrate(p) != 0)
1028 p->numa_migrate_retry = jiffies + HZ*5;
1031 static void task_numa_placement(struct task_struct *p)
1033 int seq, nid, max_nid = -1;
1034 unsigned long max_faults = 0;
1036 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1037 if (p->numa_scan_seq == seq)
1039 p->numa_scan_seq = seq;
1040 p->numa_migrate_seq++;
1041 p->numa_scan_period_max = task_scan_max(p);
1043 /* Find the node with the highest number of faults */
1044 for_each_online_node(nid) {
1045 unsigned long faults;
1048 for (priv = 0; priv < 2; priv++) {
1049 i = task_faults_idx(nid, priv);
1051 /* Decay existing window, copy faults since last scan */
1052 p->numa_faults[i] >>= 1;
1053 p->numa_faults[i] += p->numa_faults_buffer[i];
1054 p->numa_faults_buffer[i] = 0;
1057 /* Find maximum private faults */
1058 faults = p->numa_faults[task_faults_idx(nid, 1)];
1059 if (faults > max_faults) {
1060 max_faults = faults;
1065 /* Preferred node as the node with the most faults */
1066 if (max_faults && max_nid != p->numa_preferred_nid) {
1067 /* Update the preferred nid and migrate task if possible */
1068 p->numa_preferred_nid = max_nid;
1069 p->numa_migrate_seq = 1;
1070 numa_migrate_preferred(p);
1075 * Got a PROT_NONE fault for a page on @node.
1077 void task_numa_fault(int last_nidpid, int node, int pages, bool migrated)
1079 struct task_struct *p = current;
1082 if (!numabalancing_enabled)
1085 /* for example, ksmd faulting in a user's mm */
1090 * First accesses are treated as private, otherwise consider accesses
1091 * to be private if the accessing pid has not changed
1093 if (!nidpid_pid_unset(last_nidpid))
1094 priv = ((p->pid & LAST__PID_MASK) == nidpid_to_pid(last_nidpid));
1098 /* Allocate buffer to track faults on a per-node basis */
1099 if (unlikely(!p->numa_faults)) {
1100 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1102 /* numa_faults and numa_faults_buffer share the allocation */
1103 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1104 if (!p->numa_faults)
1107 BUG_ON(p->numa_faults_buffer);
1108 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1112 * If pages are properly placed (did not migrate) then scan slower.
1113 * This is reset periodically in case of phase changes
1116 /* Initialise if necessary */
1117 if (!p->numa_scan_period_max)
1118 p->numa_scan_period_max = task_scan_max(p);
1120 p->numa_scan_period = min(p->numa_scan_period_max,
1121 p->numa_scan_period + 10);
1124 task_numa_placement(p);
1126 /* Retry task to preferred node migration if it previously failed */
1127 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1128 numa_migrate_preferred(p);
1130 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1133 static void reset_ptenuma_scan(struct task_struct *p)
1135 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1136 p->mm->numa_scan_offset = 0;
1140 * The expensive part of numa migration is done from task_work context.
1141 * Triggered from task_tick_numa().
1143 void task_numa_work(struct callback_head *work)
1145 unsigned long migrate, next_scan, now = jiffies;
1146 struct task_struct *p = current;
1147 struct mm_struct *mm = p->mm;
1148 struct vm_area_struct *vma;
1149 unsigned long start, end;
1150 unsigned long nr_pte_updates = 0;
1153 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1155 work->next = work; /* protect against double add */
1157 * Who cares about NUMA placement when they're dying.
1159 * NOTE: make sure not to dereference p->mm before this check,
1160 * exit_task_work() happens _after_ exit_mm() so we could be called
1161 * without p->mm even though we still had it when we enqueued this
1164 if (p->flags & PF_EXITING)
1167 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1168 mm->numa_next_scan = now +
1169 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1170 mm->numa_next_reset = now +
1171 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1175 * Reset the scan period if enough time has gone by. Objective is that
1176 * scanning will be reduced if pages are properly placed. As tasks
1177 * can enter different phases this needs to be re-examined. Lacking
1178 * proper tracking of reference behaviour, this blunt hammer is used.
1180 migrate = mm->numa_next_reset;
1181 if (time_after(now, migrate)) {
1182 p->numa_scan_period = task_scan_min(p);
1183 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1184 xchg(&mm->numa_next_reset, next_scan);
1188 * Enforce maximal scan/migration frequency..
1190 migrate = mm->numa_next_scan;
1191 if (time_before(now, migrate))
1194 if (p->numa_scan_period == 0) {
1195 p->numa_scan_period_max = task_scan_max(p);
1196 p->numa_scan_period = task_scan_min(p);
1199 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1200 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1204 * Delay this task enough that another task of this mm will likely win
1205 * the next time around.
1207 p->node_stamp += 2 * TICK_NSEC;
1209 start = mm->numa_scan_offset;
1210 pages = sysctl_numa_balancing_scan_size;
1211 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1215 down_read(&mm->mmap_sem);
1216 vma = find_vma(mm, start);
1218 reset_ptenuma_scan(p);
1222 for (; vma; vma = vma->vm_next) {
1223 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1227 start = max(start, vma->vm_start);
1228 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1229 end = min(end, vma->vm_end);
1230 nr_pte_updates += change_prot_numa(vma, start, end);
1233 * Scan sysctl_numa_balancing_scan_size but ensure that
1234 * at least one PTE is updated so that unused virtual
1235 * address space is quickly skipped.
1238 pages -= (end - start) >> PAGE_SHIFT;
1243 } while (end != vma->vm_end);
1248 * If the whole process was scanned without updates then no NUMA
1249 * hinting faults are being recorded and scan rate should be lower.
1251 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1252 p->numa_scan_period = min(p->numa_scan_period_max,
1253 p->numa_scan_period << 1);
1255 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1256 mm->numa_next_scan = next_scan;
1260 * It is possible to reach the end of the VMA list but the last few
1261 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1262 * would find the !migratable VMA on the next scan but not reset the
1263 * scanner to the start so check it now.
1266 mm->numa_scan_offset = start;
1268 reset_ptenuma_scan(p);
1269 up_read(&mm->mmap_sem);
1273 * Drive the periodic memory faults..
1275 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1277 struct callback_head *work = &curr->numa_work;
1281 * We don't care about NUMA placement if we don't have memory.
1283 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1287 * Using runtime rather than walltime has the dual advantage that
1288 * we (mostly) drive the selection from busy threads and that the
1289 * task needs to have done some actual work before we bother with
1292 now = curr->se.sum_exec_runtime;
1293 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1295 if (now - curr->node_stamp > period) {
1296 if (!curr->node_stamp)
1297 curr->numa_scan_period = task_scan_min(curr);
1298 curr->node_stamp += period;
1300 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1301 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1302 task_work_add(curr, work, true);
1307 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1310 #endif /* CONFIG_NUMA_BALANCING */
1313 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1315 update_load_add(&cfs_rq->load, se->load.weight);
1316 if (!parent_entity(se))
1317 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1319 if (entity_is_task(se))
1320 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1322 cfs_rq->nr_running++;
1326 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1328 update_load_sub(&cfs_rq->load, se->load.weight);
1329 if (!parent_entity(se))
1330 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1331 if (entity_is_task(se))
1332 list_del_init(&se->group_node);
1333 cfs_rq->nr_running--;
1336 #ifdef CONFIG_FAIR_GROUP_SCHED
1338 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1343 * Use this CPU's actual weight instead of the last load_contribution
1344 * to gain a more accurate current total weight. See
1345 * update_cfs_rq_load_contribution().
1347 tg_weight = atomic_long_read(&tg->load_avg);
1348 tg_weight -= cfs_rq->tg_load_contrib;
1349 tg_weight += cfs_rq->load.weight;
1354 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1356 long tg_weight, load, shares;
1358 tg_weight = calc_tg_weight(tg, cfs_rq);
1359 load = cfs_rq->load.weight;
1361 shares = (tg->shares * load);
1363 shares /= tg_weight;
1365 if (shares < MIN_SHARES)
1366 shares = MIN_SHARES;
1367 if (shares > tg->shares)
1368 shares = tg->shares;
1372 # else /* CONFIG_SMP */
1373 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1377 # endif /* CONFIG_SMP */
1378 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1379 unsigned long weight)
1382 /* commit outstanding execution time */
1383 if (cfs_rq->curr == se)
1384 update_curr(cfs_rq);
1385 account_entity_dequeue(cfs_rq, se);
1388 update_load_set(&se->load, weight);
1391 account_entity_enqueue(cfs_rq, se);
1394 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1396 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1398 struct task_group *tg;
1399 struct sched_entity *se;
1403 se = tg->se[cpu_of(rq_of(cfs_rq))];
1404 if (!se || throttled_hierarchy(cfs_rq))
1407 if (likely(se->load.weight == tg->shares))
1410 shares = calc_cfs_shares(cfs_rq, tg);
1412 reweight_entity(cfs_rq_of(se), se, shares);
1414 #else /* CONFIG_FAIR_GROUP_SCHED */
1415 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1418 #endif /* CONFIG_FAIR_GROUP_SCHED */
1422 * We choose a half-life close to 1 scheduling period.
1423 * Note: The tables below are dependent on this value.
1425 #define LOAD_AVG_PERIOD 32
1426 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1427 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1429 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1430 static const u32 runnable_avg_yN_inv[] = {
1431 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1432 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1433 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1434 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1435 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1436 0x85aac367, 0x82cd8698,
1440 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1441 * over-estimates when re-combining.
1443 static const u32 runnable_avg_yN_sum[] = {
1444 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1445 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1446 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1451 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1453 static __always_inline u64 decay_load(u64 val, u64 n)
1455 unsigned int local_n;
1459 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1462 /* after bounds checking we can collapse to 32-bit */
1466 * As y^PERIOD = 1/2, we can combine
1467 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1468 * With a look-up table which covers k^n (n<PERIOD)
1470 * To achieve constant time decay_load.
1472 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1473 val >>= local_n / LOAD_AVG_PERIOD;
1474 local_n %= LOAD_AVG_PERIOD;
1477 val *= runnable_avg_yN_inv[local_n];
1478 /* We don't use SRR here since we always want to round down. */
1483 * For updates fully spanning n periods, the contribution to runnable
1484 * average will be: \Sum 1024*y^n
1486 * We can compute this reasonably efficiently by combining:
1487 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1489 static u32 __compute_runnable_contrib(u64 n)
1493 if (likely(n <= LOAD_AVG_PERIOD))
1494 return runnable_avg_yN_sum[n];
1495 else if (unlikely(n >= LOAD_AVG_MAX_N))
1496 return LOAD_AVG_MAX;
1498 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1500 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1501 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1503 n -= LOAD_AVG_PERIOD;
1504 } while (n > LOAD_AVG_PERIOD);
1506 contrib = decay_load(contrib, n);
1507 return contrib + runnable_avg_yN_sum[n];
1511 * We can represent the historical contribution to runnable average as the
1512 * coefficients of a geometric series. To do this we sub-divide our runnable
1513 * history into segments of approximately 1ms (1024us); label the segment that
1514 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1516 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1518 * (now) (~1ms ago) (~2ms ago)
1520 * Let u_i denote the fraction of p_i that the entity was runnable.
1522 * We then designate the fractions u_i as our co-efficients, yielding the
1523 * following representation of historical load:
1524 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1526 * We choose y based on the with of a reasonably scheduling period, fixing:
1529 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1530 * approximately half as much as the contribution to load within the last ms
1533 * When a period "rolls over" and we have new u_0`, multiplying the previous
1534 * sum again by y is sufficient to update:
1535 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1536 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1538 static __always_inline int __update_entity_runnable_avg(u64 now,
1539 struct sched_avg *sa,
1543 u32 runnable_contrib;
1544 int delta_w, decayed = 0;
1546 delta = now - sa->last_runnable_update;
1548 * This should only happen when time goes backwards, which it
1549 * unfortunately does during sched clock init when we swap over to TSC.
1551 if ((s64)delta < 0) {
1552 sa->last_runnable_update = now;
1557 * Use 1024ns as the unit of measurement since it's a reasonable
1558 * approximation of 1us and fast to compute.
1563 sa->last_runnable_update = now;
1565 /* delta_w is the amount already accumulated against our next period */
1566 delta_w = sa->runnable_avg_period % 1024;
1567 if (delta + delta_w >= 1024) {
1568 /* period roll-over */
1572 * Now that we know we're crossing a period boundary, figure
1573 * out how much from delta we need to complete the current
1574 * period and accrue it.
1576 delta_w = 1024 - delta_w;
1578 sa->runnable_avg_sum += delta_w;
1579 sa->runnable_avg_period += delta_w;
1583 /* Figure out how many additional periods this update spans */
1584 periods = delta / 1024;
1587 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1589 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1592 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1593 runnable_contrib = __compute_runnable_contrib(periods);
1595 sa->runnable_avg_sum += runnable_contrib;
1596 sa->runnable_avg_period += runnable_contrib;
1599 /* Remainder of delta accrued against u_0` */
1601 sa->runnable_avg_sum += delta;
1602 sa->runnable_avg_period += delta;
1607 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1608 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1610 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1611 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1613 decays -= se->avg.decay_count;
1617 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1618 se->avg.decay_count = 0;
1623 #ifdef CONFIG_FAIR_GROUP_SCHED
1624 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1627 struct task_group *tg = cfs_rq->tg;
1630 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1631 tg_contrib -= cfs_rq->tg_load_contrib;
1633 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1634 atomic_long_add(tg_contrib, &tg->load_avg);
1635 cfs_rq->tg_load_contrib += tg_contrib;
1640 * Aggregate cfs_rq runnable averages into an equivalent task_group
1641 * representation for computing load contributions.
1643 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1644 struct cfs_rq *cfs_rq)
1646 struct task_group *tg = cfs_rq->tg;
1649 /* The fraction of a cpu used by this cfs_rq */
1650 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1651 sa->runnable_avg_period + 1);
1652 contrib -= cfs_rq->tg_runnable_contrib;
1654 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1655 atomic_add(contrib, &tg->runnable_avg);
1656 cfs_rq->tg_runnable_contrib += contrib;
1660 static inline void __update_group_entity_contrib(struct sched_entity *se)
1662 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1663 struct task_group *tg = cfs_rq->tg;
1668 contrib = cfs_rq->tg_load_contrib * tg->shares;
1669 se->avg.load_avg_contrib = div_u64(contrib,
1670 atomic_long_read(&tg->load_avg) + 1);
1673 * For group entities we need to compute a correction term in the case
1674 * that they are consuming <1 cpu so that we would contribute the same
1675 * load as a task of equal weight.
1677 * Explicitly co-ordinating this measurement would be expensive, but
1678 * fortunately the sum of each cpus contribution forms a usable
1679 * lower-bound on the true value.
1681 * Consider the aggregate of 2 contributions. Either they are disjoint
1682 * (and the sum represents true value) or they are disjoint and we are
1683 * understating by the aggregate of their overlap.
1685 * Extending this to N cpus, for a given overlap, the maximum amount we
1686 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1687 * cpus that overlap for this interval and w_i is the interval width.
1689 * On a small machine; the first term is well-bounded which bounds the
1690 * total error since w_i is a subset of the period. Whereas on a
1691 * larger machine, while this first term can be larger, if w_i is the
1692 * of consequential size guaranteed to see n_i*w_i quickly converge to
1693 * our upper bound of 1-cpu.
1695 runnable_avg = atomic_read(&tg->runnable_avg);
1696 if (runnable_avg < NICE_0_LOAD) {
1697 se->avg.load_avg_contrib *= runnable_avg;
1698 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1702 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1703 int force_update) {}
1704 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1705 struct cfs_rq *cfs_rq) {}
1706 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1709 static inline void __update_task_entity_contrib(struct sched_entity *se)
1713 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1714 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1715 contrib /= (se->avg.runnable_avg_period + 1);
1716 se->avg.load_avg_contrib = scale_load(contrib);
1719 /* Compute the current contribution to load_avg by se, return any delta */
1720 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1722 long old_contrib = se->avg.load_avg_contrib;
1724 if (entity_is_task(se)) {
1725 __update_task_entity_contrib(se);
1727 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1728 __update_group_entity_contrib(se);
1731 return se->avg.load_avg_contrib - old_contrib;
1734 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1737 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1738 cfs_rq->blocked_load_avg -= load_contrib;
1740 cfs_rq->blocked_load_avg = 0;
1743 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1745 /* Update a sched_entity's runnable average */
1746 static inline void update_entity_load_avg(struct sched_entity *se,
1749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1754 * For a group entity we need to use their owned cfs_rq_clock_task() in
1755 * case they are the parent of a throttled hierarchy.
1757 if (entity_is_task(se))
1758 now = cfs_rq_clock_task(cfs_rq);
1760 now = cfs_rq_clock_task(group_cfs_rq(se));
1762 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1765 contrib_delta = __update_entity_load_avg_contrib(se);
1771 cfs_rq->runnable_load_avg += contrib_delta;
1773 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1777 * Decay the load contributed by all blocked children and account this so that
1778 * their contribution may appropriately discounted when they wake up.
1780 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1782 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1785 decays = now - cfs_rq->last_decay;
1786 if (!decays && !force_update)
1789 if (atomic_long_read(&cfs_rq->removed_load)) {
1790 unsigned long removed_load;
1791 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1792 subtract_blocked_load_contrib(cfs_rq, removed_load);
1796 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1798 atomic64_add(decays, &cfs_rq->decay_counter);
1799 cfs_rq->last_decay = now;
1802 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1805 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1807 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1808 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1811 /* Add the load generated by se into cfs_rq's child load-average */
1812 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1813 struct sched_entity *se,
1817 * We track migrations using entity decay_count <= 0, on a wake-up
1818 * migration we use a negative decay count to track the remote decays
1819 * accumulated while sleeping.
1821 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1822 * are seen by enqueue_entity_load_avg() as a migration with an already
1823 * constructed load_avg_contrib.
1825 if (unlikely(se->avg.decay_count <= 0)) {
1826 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1827 if (se->avg.decay_count) {
1829 * In a wake-up migration we have to approximate the
1830 * time sleeping. This is because we can't synchronize
1831 * clock_task between the two cpus, and it is not
1832 * guaranteed to be read-safe. Instead, we can
1833 * approximate this using our carried decays, which are
1834 * explicitly atomically readable.
1836 se->avg.last_runnable_update -= (-se->avg.decay_count)
1838 update_entity_load_avg(se, 0);
1839 /* Indicate that we're now synchronized and on-rq */
1840 se->avg.decay_count = 0;
1845 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1846 * would have made count negative); we must be careful to avoid
1847 * double-accounting blocked time after synchronizing decays.
1849 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1853 /* migrated tasks did not contribute to our blocked load */
1855 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1856 update_entity_load_avg(se, 0);
1859 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1860 /* we force update consideration on load-balancer moves */
1861 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1865 * Remove se's load from this cfs_rq child load-average, if the entity is
1866 * transitioning to a blocked state we track its projected decay using
1869 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1870 struct sched_entity *se,
1873 update_entity_load_avg(se, 1);
1874 /* we force update consideration on load-balancer moves */
1875 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1877 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1879 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1880 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1881 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1885 * Update the rq's load with the elapsed running time before entering
1886 * idle. if the last scheduled task is not a CFS task, idle_enter will
1887 * be the only way to update the runnable statistic.
1889 void idle_enter_fair(struct rq *this_rq)
1891 update_rq_runnable_avg(this_rq, 1);
1895 * Update the rq's load with the elapsed idle time before a task is
1896 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1897 * be the only way to update the runnable statistic.
1899 void idle_exit_fair(struct rq *this_rq)
1901 update_rq_runnable_avg(this_rq, 0);
1905 static inline void update_entity_load_avg(struct sched_entity *se,
1906 int update_cfs_rq) {}
1907 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1908 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1909 struct sched_entity *se,
1911 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1912 struct sched_entity *se,
1914 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1915 int force_update) {}
1918 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1920 #ifdef CONFIG_SCHEDSTATS
1921 struct task_struct *tsk = NULL;
1923 if (entity_is_task(se))
1926 if (se->statistics.sleep_start) {
1927 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1932 if (unlikely(delta > se->statistics.sleep_max))
1933 se->statistics.sleep_max = delta;
1935 se->statistics.sleep_start = 0;
1936 se->statistics.sum_sleep_runtime += delta;
1939 account_scheduler_latency(tsk, delta >> 10, 1);
1940 trace_sched_stat_sleep(tsk, delta);
1943 if (se->statistics.block_start) {
1944 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1949 if (unlikely(delta > se->statistics.block_max))
1950 se->statistics.block_max = delta;
1952 se->statistics.block_start = 0;
1953 se->statistics.sum_sleep_runtime += delta;
1956 if (tsk->in_iowait) {
1957 se->statistics.iowait_sum += delta;
1958 se->statistics.iowait_count++;
1959 trace_sched_stat_iowait(tsk, delta);
1962 trace_sched_stat_blocked(tsk, delta);
1965 * Blocking time is in units of nanosecs, so shift by
1966 * 20 to get a milliseconds-range estimation of the
1967 * amount of time that the task spent sleeping:
1969 if (unlikely(prof_on == SLEEP_PROFILING)) {
1970 profile_hits(SLEEP_PROFILING,
1971 (void *)get_wchan(tsk),
1974 account_scheduler_latency(tsk, delta >> 10, 0);
1980 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1982 #ifdef CONFIG_SCHED_DEBUG
1983 s64 d = se->vruntime - cfs_rq->min_vruntime;
1988 if (d > 3*sysctl_sched_latency)
1989 schedstat_inc(cfs_rq, nr_spread_over);
1994 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1996 u64 vruntime = cfs_rq->min_vruntime;
1999 * The 'current' period is already promised to the current tasks,
2000 * however the extra weight of the new task will slow them down a
2001 * little, place the new task so that it fits in the slot that
2002 * stays open at the end.
2004 if (initial && sched_feat(START_DEBIT))
2005 vruntime += sched_vslice(cfs_rq, se);
2007 /* sleeps up to a single latency don't count. */
2009 unsigned long thresh = sysctl_sched_latency;
2012 * Halve their sleep time's effect, to allow
2013 * for a gentler effect of sleepers:
2015 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2021 /* ensure we never gain time by being placed backwards. */
2022 se->vruntime = max_vruntime(se->vruntime, vruntime);
2025 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2028 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2031 * Update the normalized vruntime before updating min_vruntime
2032 * through calling update_curr().
2034 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2035 se->vruntime += cfs_rq->min_vruntime;
2038 * Update run-time statistics of the 'current'.
2040 update_curr(cfs_rq);
2041 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2042 account_entity_enqueue(cfs_rq, se);
2043 update_cfs_shares(cfs_rq);
2045 if (flags & ENQUEUE_WAKEUP) {
2046 place_entity(cfs_rq, se, 0);
2047 enqueue_sleeper(cfs_rq, se);
2050 update_stats_enqueue(cfs_rq, se);
2051 check_spread(cfs_rq, se);
2052 if (se != cfs_rq->curr)
2053 __enqueue_entity(cfs_rq, se);
2056 if (cfs_rq->nr_running == 1) {
2057 list_add_leaf_cfs_rq(cfs_rq);
2058 check_enqueue_throttle(cfs_rq);
2062 static void __clear_buddies_last(struct sched_entity *se)
2064 for_each_sched_entity(se) {
2065 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2066 if (cfs_rq->last == se)
2067 cfs_rq->last = NULL;
2073 static void __clear_buddies_next(struct sched_entity *se)
2075 for_each_sched_entity(se) {
2076 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2077 if (cfs_rq->next == se)
2078 cfs_rq->next = NULL;
2084 static void __clear_buddies_skip(struct sched_entity *se)
2086 for_each_sched_entity(se) {
2087 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2088 if (cfs_rq->skip == se)
2089 cfs_rq->skip = NULL;
2095 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2097 if (cfs_rq->last == se)
2098 __clear_buddies_last(se);
2100 if (cfs_rq->next == se)
2101 __clear_buddies_next(se);
2103 if (cfs_rq->skip == se)
2104 __clear_buddies_skip(se);
2107 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2110 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2113 * Update run-time statistics of the 'current'.
2115 update_curr(cfs_rq);
2116 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2118 update_stats_dequeue(cfs_rq, se);
2119 if (flags & DEQUEUE_SLEEP) {
2120 #ifdef CONFIG_SCHEDSTATS
2121 if (entity_is_task(se)) {
2122 struct task_struct *tsk = task_of(se);
2124 if (tsk->state & TASK_INTERRUPTIBLE)
2125 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2126 if (tsk->state & TASK_UNINTERRUPTIBLE)
2127 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2132 clear_buddies(cfs_rq, se);
2134 if (se != cfs_rq->curr)
2135 __dequeue_entity(cfs_rq, se);
2137 account_entity_dequeue(cfs_rq, se);
2140 * Normalize the entity after updating the min_vruntime because the
2141 * update can refer to the ->curr item and we need to reflect this
2142 * movement in our normalized position.
2144 if (!(flags & DEQUEUE_SLEEP))
2145 se->vruntime -= cfs_rq->min_vruntime;
2147 /* return excess runtime on last dequeue */
2148 return_cfs_rq_runtime(cfs_rq);
2150 update_min_vruntime(cfs_rq);
2151 update_cfs_shares(cfs_rq);
2155 * Preempt the current task with a newly woken task if needed:
2158 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2160 unsigned long ideal_runtime, delta_exec;
2161 struct sched_entity *se;
2164 ideal_runtime = sched_slice(cfs_rq, curr);
2165 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2166 if (delta_exec > ideal_runtime) {
2167 resched_task(rq_of(cfs_rq)->curr);
2169 * The current task ran long enough, ensure it doesn't get
2170 * re-elected due to buddy favours.
2172 clear_buddies(cfs_rq, curr);
2177 * Ensure that a task that missed wakeup preemption by a
2178 * narrow margin doesn't have to wait for a full slice.
2179 * This also mitigates buddy induced latencies under load.
2181 if (delta_exec < sysctl_sched_min_granularity)
2184 se = __pick_first_entity(cfs_rq);
2185 delta = curr->vruntime - se->vruntime;
2190 if (delta > ideal_runtime)
2191 resched_task(rq_of(cfs_rq)->curr);
2195 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2197 /* 'current' is not kept within the tree. */
2200 * Any task has to be enqueued before it get to execute on
2201 * a CPU. So account for the time it spent waiting on the
2204 update_stats_wait_end(cfs_rq, se);
2205 __dequeue_entity(cfs_rq, se);
2208 update_stats_curr_start(cfs_rq, se);
2210 #ifdef CONFIG_SCHEDSTATS
2212 * Track our maximum slice length, if the CPU's load is at
2213 * least twice that of our own weight (i.e. dont track it
2214 * when there are only lesser-weight tasks around):
2216 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2217 se->statistics.slice_max = max(se->statistics.slice_max,
2218 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2221 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2225 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2228 * Pick the next process, keeping these things in mind, in this order:
2229 * 1) keep things fair between processes/task groups
2230 * 2) pick the "next" process, since someone really wants that to run
2231 * 3) pick the "last" process, for cache locality
2232 * 4) do not run the "skip" process, if something else is available
2234 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2236 struct sched_entity *se = __pick_first_entity(cfs_rq);
2237 struct sched_entity *left = se;
2240 * Avoid running the skip buddy, if running something else can
2241 * be done without getting too unfair.
2243 if (cfs_rq->skip == se) {
2244 struct sched_entity *second = __pick_next_entity(se);
2245 if (second && wakeup_preempt_entity(second, left) < 1)
2250 * Prefer last buddy, try to return the CPU to a preempted task.
2252 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2256 * Someone really wants this to run. If it's not unfair, run it.
2258 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2261 clear_buddies(cfs_rq, se);
2266 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2268 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2271 * If still on the runqueue then deactivate_task()
2272 * was not called and update_curr() has to be done:
2275 update_curr(cfs_rq);
2277 /* throttle cfs_rqs exceeding runtime */
2278 check_cfs_rq_runtime(cfs_rq);
2280 check_spread(cfs_rq, prev);
2282 update_stats_wait_start(cfs_rq, prev);
2283 /* Put 'current' back into the tree. */
2284 __enqueue_entity(cfs_rq, prev);
2285 /* in !on_rq case, update occurred at dequeue */
2286 update_entity_load_avg(prev, 1);
2288 cfs_rq->curr = NULL;
2292 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2295 * Update run-time statistics of the 'current'.
2297 update_curr(cfs_rq);
2300 * Ensure that runnable average is periodically updated.
2302 update_entity_load_avg(curr, 1);
2303 update_cfs_rq_blocked_load(cfs_rq, 1);
2304 update_cfs_shares(cfs_rq);
2306 #ifdef CONFIG_SCHED_HRTICK
2308 * queued ticks are scheduled to match the slice, so don't bother
2309 * validating it and just reschedule.
2312 resched_task(rq_of(cfs_rq)->curr);
2316 * don't let the period tick interfere with the hrtick preemption
2318 if (!sched_feat(DOUBLE_TICK) &&
2319 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2323 if (cfs_rq->nr_running > 1)
2324 check_preempt_tick(cfs_rq, curr);
2328 /**************************************************
2329 * CFS bandwidth control machinery
2332 #ifdef CONFIG_CFS_BANDWIDTH
2334 #ifdef HAVE_JUMP_LABEL
2335 static struct static_key __cfs_bandwidth_used;
2337 static inline bool cfs_bandwidth_used(void)
2339 return static_key_false(&__cfs_bandwidth_used);
2342 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2344 /* only need to count groups transitioning between enabled/!enabled */
2345 if (enabled && !was_enabled)
2346 static_key_slow_inc(&__cfs_bandwidth_used);
2347 else if (!enabled && was_enabled)
2348 static_key_slow_dec(&__cfs_bandwidth_used);
2350 #else /* HAVE_JUMP_LABEL */
2351 static bool cfs_bandwidth_used(void)
2356 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2357 #endif /* HAVE_JUMP_LABEL */
2360 * default period for cfs group bandwidth.
2361 * default: 0.1s, units: nanoseconds
2363 static inline u64 default_cfs_period(void)
2365 return 100000000ULL;
2368 static inline u64 sched_cfs_bandwidth_slice(void)
2370 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2374 * Replenish runtime according to assigned quota and update expiration time.
2375 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2376 * additional synchronization around rq->lock.
2378 * requires cfs_b->lock
2380 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2384 if (cfs_b->quota == RUNTIME_INF)
2387 now = sched_clock_cpu(smp_processor_id());
2388 cfs_b->runtime = cfs_b->quota;
2389 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2392 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2394 return &tg->cfs_bandwidth;
2397 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2398 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2400 if (unlikely(cfs_rq->throttle_count))
2401 return cfs_rq->throttled_clock_task;
2403 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2406 /* returns 0 on failure to allocate runtime */
2407 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2409 struct task_group *tg = cfs_rq->tg;
2410 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2411 u64 amount = 0, min_amount, expires;
2413 /* note: this is a positive sum as runtime_remaining <= 0 */
2414 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2416 raw_spin_lock(&cfs_b->lock);
2417 if (cfs_b->quota == RUNTIME_INF)
2418 amount = min_amount;
2421 * If the bandwidth pool has become inactive, then at least one
2422 * period must have elapsed since the last consumption.
2423 * Refresh the global state and ensure bandwidth timer becomes
2426 if (!cfs_b->timer_active) {
2427 __refill_cfs_bandwidth_runtime(cfs_b);
2428 __start_cfs_bandwidth(cfs_b);
2431 if (cfs_b->runtime > 0) {
2432 amount = min(cfs_b->runtime, min_amount);
2433 cfs_b->runtime -= amount;
2437 expires = cfs_b->runtime_expires;
2438 raw_spin_unlock(&cfs_b->lock);
2440 cfs_rq->runtime_remaining += amount;
2442 * we may have advanced our local expiration to account for allowed
2443 * spread between our sched_clock and the one on which runtime was
2446 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2447 cfs_rq->runtime_expires = expires;
2449 return cfs_rq->runtime_remaining > 0;
2453 * Note: This depends on the synchronization provided by sched_clock and the
2454 * fact that rq->clock snapshots this value.
2456 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2458 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2460 /* if the deadline is ahead of our clock, nothing to do */
2461 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2464 if (cfs_rq->runtime_remaining < 0)
2468 * If the local deadline has passed we have to consider the
2469 * possibility that our sched_clock is 'fast' and the global deadline
2470 * has not truly expired.
2472 * Fortunately we can check determine whether this the case by checking
2473 * whether the global deadline has advanced.
2476 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2477 /* extend local deadline, drift is bounded above by 2 ticks */
2478 cfs_rq->runtime_expires += TICK_NSEC;
2480 /* global deadline is ahead, expiration has passed */
2481 cfs_rq->runtime_remaining = 0;
2485 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2486 unsigned long delta_exec)
2488 /* dock delta_exec before expiring quota (as it could span periods) */
2489 cfs_rq->runtime_remaining -= delta_exec;
2490 expire_cfs_rq_runtime(cfs_rq);
2492 if (likely(cfs_rq->runtime_remaining > 0))
2496 * if we're unable to extend our runtime we resched so that the active
2497 * hierarchy can be throttled
2499 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2500 resched_task(rq_of(cfs_rq)->curr);
2503 static __always_inline
2504 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2506 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2509 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2512 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2514 return cfs_bandwidth_used() && cfs_rq->throttled;
2517 /* check whether cfs_rq, or any parent, is throttled */
2518 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2520 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2524 * Ensure that neither of the group entities corresponding to src_cpu or
2525 * dest_cpu are members of a throttled hierarchy when performing group
2526 * load-balance operations.
2528 static inline int throttled_lb_pair(struct task_group *tg,
2529 int src_cpu, int dest_cpu)
2531 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2533 src_cfs_rq = tg->cfs_rq[src_cpu];
2534 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2536 return throttled_hierarchy(src_cfs_rq) ||
2537 throttled_hierarchy(dest_cfs_rq);
2540 /* updated child weight may affect parent so we have to do this bottom up */
2541 static int tg_unthrottle_up(struct task_group *tg, void *data)
2543 struct rq *rq = data;
2544 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2546 cfs_rq->throttle_count--;
2548 if (!cfs_rq->throttle_count) {
2549 /* adjust cfs_rq_clock_task() */
2550 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2551 cfs_rq->throttled_clock_task;
2558 static int tg_throttle_down(struct task_group *tg, void *data)
2560 struct rq *rq = data;
2561 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2563 /* group is entering throttled state, stop time */
2564 if (!cfs_rq->throttle_count)
2565 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2566 cfs_rq->throttle_count++;
2571 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2573 struct rq *rq = rq_of(cfs_rq);
2574 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2575 struct sched_entity *se;
2576 long task_delta, dequeue = 1;
2578 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2580 /* freeze hierarchy runnable averages while throttled */
2582 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2585 task_delta = cfs_rq->h_nr_running;
2586 for_each_sched_entity(se) {
2587 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2588 /* throttled entity or throttle-on-deactivate */
2593 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2594 qcfs_rq->h_nr_running -= task_delta;
2596 if (qcfs_rq->load.weight)
2601 rq->nr_running -= task_delta;
2603 cfs_rq->throttled = 1;
2604 cfs_rq->throttled_clock = rq_clock(rq);
2605 raw_spin_lock(&cfs_b->lock);
2606 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2607 raw_spin_unlock(&cfs_b->lock);
2610 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2612 struct rq *rq = rq_of(cfs_rq);
2613 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2614 struct sched_entity *se;
2618 se = cfs_rq->tg->se[cpu_of(rq)];
2620 cfs_rq->throttled = 0;
2622 update_rq_clock(rq);
2624 raw_spin_lock(&cfs_b->lock);
2625 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2626 list_del_rcu(&cfs_rq->throttled_list);
2627 raw_spin_unlock(&cfs_b->lock);
2629 /* update hierarchical throttle state */
2630 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2632 if (!cfs_rq->load.weight)
2635 task_delta = cfs_rq->h_nr_running;
2636 for_each_sched_entity(se) {
2640 cfs_rq = cfs_rq_of(se);
2642 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2643 cfs_rq->h_nr_running += task_delta;
2645 if (cfs_rq_throttled(cfs_rq))
2650 rq->nr_running += task_delta;
2652 /* determine whether we need to wake up potentially idle cpu */
2653 if (rq->curr == rq->idle && rq->cfs.nr_running)
2654 resched_task(rq->curr);
2657 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2658 u64 remaining, u64 expires)
2660 struct cfs_rq *cfs_rq;
2661 u64 runtime = remaining;
2664 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2666 struct rq *rq = rq_of(cfs_rq);
2668 raw_spin_lock(&rq->lock);
2669 if (!cfs_rq_throttled(cfs_rq))
2672 runtime = -cfs_rq->runtime_remaining + 1;
2673 if (runtime > remaining)
2674 runtime = remaining;
2675 remaining -= runtime;
2677 cfs_rq->runtime_remaining += runtime;
2678 cfs_rq->runtime_expires = expires;
2680 /* we check whether we're throttled above */
2681 if (cfs_rq->runtime_remaining > 0)
2682 unthrottle_cfs_rq(cfs_rq);
2685 raw_spin_unlock(&rq->lock);
2696 * Responsible for refilling a task_group's bandwidth and unthrottling its
2697 * cfs_rqs as appropriate. If there has been no activity within the last
2698 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2699 * used to track this state.
2701 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2703 u64 runtime, runtime_expires;
2704 int idle = 1, throttled;
2706 raw_spin_lock(&cfs_b->lock);
2707 /* no need to continue the timer with no bandwidth constraint */
2708 if (cfs_b->quota == RUNTIME_INF)
2711 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2712 /* idle depends on !throttled (for the case of a large deficit) */
2713 idle = cfs_b->idle && !throttled;
2714 cfs_b->nr_periods += overrun;
2716 /* if we're going inactive then everything else can be deferred */
2720 __refill_cfs_bandwidth_runtime(cfs_b);
2723 /* mark as potentially idle for the upcoming period */
2728 /* account preceding periods in which throttling occurred */
2729 cfs_b->nr_throttled += overrun;
2732 * There are throttled entities so we must first use the new bandwidth
2733 * to unthrottle them before making it generally available. This
2734 * ensures that all existing debts will be paid before a new cfs_rq is
2737 runtime = cfs_b->runtime;
2738 runtime_expires = cfs_b->runtime_expires;
2742 * This check is repeated as we are holding onto the new bandwidth
2743 * while we unthrottle. This can potentially race with an unthrottled
2744 * group trying to acquire new bandwidth from the global pool.
2746 while (throttled && runtime > 0) {
2747 raw_spin_unlock(&cfs_b->lock);
2748 /* we can't nest cfs_b->lock while distributing bandwidth */
2749 runtime = distribute_cfs_runtime(cfs_b, runtime,
2751 raw_spin_lock(&cfs_b->lock);
2753 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2756 /* return (any) remaining runtime */
2757 cfs_b->runtime = runtime;
2759 * While we are ensured activity in the period following an
2760 * unthrottle, this also covers the case in which the new bandwidth is
2761 * insufficient to cover the existing bandwidth deficit. (Forcing the
2762 * timer to remain active while there are any throttled entities.)
2767 cfs_b->timer_active = 0;
2768 raw_spin_unlock(&cfs_b->lock);
2773 /* a cfs_rq won't donate quota below this amount */
2774 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2775 /* minimum remaining period time to redistribute slack quota */
2776 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2777 /* how long we wait to gather additional slack before distributing */
2778 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2780 /* are we near the end of the current quota period? */
2781 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2783 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2786 /* if the call-back is running a quota refresh is already occurring */
2787 if (hrtimer_callback_running(refresh_timer))
2790 /* is a quota refresh about to occur? */
2791 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2792 if (remaining < min_expire)
2798 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2800 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2802 /* if there's a quota refresh soon don't bother with slack */
2803 if (runtime_refresh_within(cfs_b, min_left))
2806 start_bandwidth_timer(&cfs_b->slack_timer,
2807 ns_to_ktime(cfs_bandwidth_slack_period));
2810 /* we know any runtime found here is valid as update_curr() precedes return */
2811 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2813 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2814 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2816 if (slack_runtime <= 0)
2819 raw_spin_lock(&cfs_b->lock);
2820 if (cfs_b->quota != RUNTIME_INF &&
2821 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2822 cfs_b->runtime += slack_runtime;
2824 /* we are under rq->lock, defer unthrottling using a timer */
2825 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2826 !list_empty(&cfs_b->throttled_cfs_rq))
2827 start_cfs_slack_bandwidth(cfs_b);
2829 raw_spin_unlock(&cfs_b->lock);
2831 /* even if it's not valid for return we don't want to try again */
2832 cfs_rq->runtime_remaining -= slack_runtime;
2835 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2837 if (!cfs_bandwidth_used())
2840 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2843 __return_cfs_rq_runtime(cfs_rq);
2847 * This is done with a timer (instead of inline with bandwidth return) since
2848 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2850 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2852 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2855 /* confirm we're still not at a refresh boundary */
2856 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2859 raw_spin_lock(&cfs_b->lock);
2860 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2861 runtime = cfs_b->runtime;
2864 expires = cfs_b->runtime_expires;
2865 raw_spin_unlock(&cfs_b->lock);
2870 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2872 raw_spin_lock(&cfs_b->lock);
2873 if (expires == cfs_b->runtime_expires)
2874 cfs_b->runtime = runtime;
2875 raw_spin_unlock(&cfs_b->lock);
2879 * When a group wakes up we want to make sure that its quota is not already
2880 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2881 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2883 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2885 if (!cfs_bandwidth_used())
2888 /* an active group must be handled by the update_curr()->put() path */
2889 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2892 /* ensure the group is not already throttled */
2893 if (cfs_rq_throttled(cfs_rq))
2896 /* update runtime allocation */
2897 account_cfs_rq_runtime(cfs_rq, 0);
2898 if (cfs_rq->runtime_remaining <= 0)
2899 throttle_cfs_rq(cfs_rq);
2902 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2903 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2905 if (!cfs_bandwidth_used())
2908 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2912 * it's possible for a throttled entity to be forced into a running
2913 * state (e.g. set_curr_task), in this case we're finished.
2915 if (cfs_rq_throttled(cfs_rq))
2918 throttle_cfs_rq(cfs_rq);
2921 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2923 struct cfs_bandwidth *cfs_b =
2924 container_of(timer, struct cfs_bandwidth, slack_timer);
2925 do_sched_cfs_slack_timer(cfs_b);
2927 return HRTIMER_NORESTART;
2930 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2932 struct cfs_bandwidth *cfs_b =
2933 container_of(timer, struct cfs_bandwidth, period_timer);
2939 now = hrtimer_cb_get_time(timer);
2940 overrun = hrtimer_forward(timer, now, cfs_b->period);
2945 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2948 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2951 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2953 raw_spin_lock_init(&cfs_b->lock);
2955 cfs_b->quota = RUNTIME_INF;
2956 cfs_b->period = ns_to_ktime(default_cfs_period());
2958 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2959 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2960 cfs_b->period_timer.function = sched_cfs_period_timer;
2961 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2962 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2965 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2967 cfs_rq->runtime_enabled = 0;
2968 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2971 /* requires cfs_b->lock, may release to reprogram timer */
2972 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2975 * The timer may be active because we're trying to set a new bandwidth
2976 * period or because we're racing with the tear-down path
2977 * (timer_active==0 becomes visible before the hrtimer call-back
2978 * terminates). In either case we ensure that it's re-programmed
2980 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2981 raw_spin_unlock(&cfs_b->lock);
2982 /* ensure cfs_b->lock is available while we wait */
2983 hrtimer_cancel(&cfs_b->period_timer);
2985 raw_spin_lock(&cfs_b->lock);
2986 /* if someone else restarted the timer then we're done */
2987 if (cfs_b->timer_active)
2991 cfs_b->timer_active = 1;
2992 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2995 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2997 hrtimer_cancel(&cfs_b->period_timer);
2998 hrtimer_cancel(&cfs_b->slack_timer);
3001 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3003 struct cfs_rq *cfs_rq;
3005 for_each_leaf_cfs_rq(rq, cfs_rq) {
3006 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3008 if (!cfs_rq->runtime_enabled)
3012 * clock_task is not advancing so we just need to make sure
3013 * there's some valid quota amount
3015 cfs_rq->runtime_remaining = cfs_b->quota;
3016 if (cfs_rq_throttled(cfs_rq))
3017 unthrottle_cfs_rq(cfs_rq);
3021 #else /* CONFIG_CFS_BANDWIDTH */
3022 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3024 return rq_clock_task(rq_of(cfs_rq));
3027 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3028 unsigned long delta_exec) {}
3029 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3030 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3031 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3033 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3038 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3043 static inline int throttled_lb_pair(struct task_group *tg,
3044 int src_cpu, int dest_cpu)
3049 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3051 #ifdef CONFIG_FAIR_GROUP_SCHED
3052 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3055 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3059 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3060 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3062 #endif /* CONFIG_CFS_BANDWIDTH */
3064 /**************************************************
3065 * CFS operations on tasks:
3068 #ifdef CONFIG_SCHED_HRTICK
3069 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3071 struct sched_entity *se = &p->se;
3072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3074 WARN_ON(task_rq(p) != rq);
3076 if (cfs_rq->nr_running > 1) {
3077 u64 slice = sched_slice(cfs_rq, se);
3078 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3079 s64 delta = slice - ran;
3088 * Don't schedule slices shorter than 10000ns, that just
3089 * doesn't make sense. Rely on vruntime for fairness.
3092 delta = max_t(s64, 10000LL, delta);
3094 hrtick_start(rq, delta);
3099 * called from enqueue/dequeue and updates the hrtick when the
3100 * current task is from our class and nr_running is low enough
3103 static void hrtick_update(struct rq *rq)
3105 struct task_struct *curr = rq->curr;
3107 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3110 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3111 hrtick_start_fair(rq, curr);
3113 #else /* !CONFIG_SCHED_HRTICK */
3115 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3119 static inline void hrtick_update(struct rq *rq)
3125 * The enqueue_task method is called before nr_running is
3126 * increased. Here we update the fair scheduling stats and
3127 * then put the task into the rbtree:
3130 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3132 struct cfs_rq *cfs_rq;
3133 struct sched_entity *se = &p->se;
3135 for_each_sched_entity(se) {
3138 cfs_rq = cfs_rq_of(se);
3139 enqueue_entity(cfs_rq, se, flags);
3142 * end evaluation on encountering a throttled cfs_rq
3144 * note: in the case of encountering a throttled cfs_rq we will
3145 * post the final h_nr_running increment below.
3147 if (cfs_rq_throttled(cfs_rq))
3149 cfs_rq->h_nr_running++;
3151 flags = ENQUEUE_WAKEUP;
3154 for_each_sched_entity(se) {
3155 cfs_rq = cfs_rq_of(se);
3156 cfs_rq->h_nr_running++;
3158 if (cfs_rq_throttled(cfs_rq))
3161 update_cfs_shares(cfs_rq);
3162 update_entity_load_avg(se, 1);
3166 update_rq_runnable_avg(rq, rq->nr_running);
3172 static void set_next_buddy(struct sched_entity *se);
3175 * The dequeue_task method is called before nr_running is
3176 * decreased. We remove the task from the rbtree and
3177 * update the fair scheduling stats:
3179 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3181 struct cfs_rq *cfs_rq;
3182 struct sched_entity *se = &p->se;
3183 int task_sleep = flags & DEQUEUE_SLEEP;
3185 for_each_sched_entity(se) {
3186 cfs_rq = cfs_rq_of(se);
3187 dequeue_entity(cfs_rq, se, flags);
3190 * end evaluation on encountering a throttled cfs_rq
3192 * note: in the case of encountering a throttled cfs_rq we will
3193 * post the final h_nr_running decrement below.
3195 if (cfs_rq_throttled(cfs_rq))
3197 cfs_rq->h_nr_running--;
3199 /* Don't dequeue parent if it has other entities besides us */
3200 if (cfs_rq->load.weight) {
3202 * Bias pick_next to pick a task from this cfs_rq, as
3203 * p is sleeping when it is within its sched_slice.
3205 if (task_sleep && parent_entity(se))
3206 set_next_buddy(parent_entity(se));
3208 /* avoid re-evaluating load for this entity */
3209 se = parent_entity(se);
3212 flags |= DEQUEUE_SLEEP;
3215 for_each_sched_entity(se) {
3216 cfs_rq = cfs_rq_of(se);
3217 cfs_rq->h_nr_running--;
3219 if (cfs_rq_throttled(cfs_rq))
3222 update_cfs_shares(cfs_rq);
3223 update_entity_load_avg(se, 1);
3228 update_rq_runnable_avg(rq, 1);
3234 /* Used instead of source_load when we know the type == 0 */
3235 static unsigned long weighted_cpuload(const int cpu)
3237 return cpu_rq(cpu)->cfs.runnable_load_avg;
3241 * Return a low guess at the load of a migration-source cpu weighted
3242 * according to the scheduling class and "nice" value.
3244 * We want to under-estimate the load of migration sources, to
3245 * balance conservatively.
3247 static unsigned long source_load(int cpu, int type)
3249 struct rq *rq = cpu_rq(cpu);
3250 unsigned long total = weighted_cpuload(cpu);
3252 if (type == 0 || !sched_feat(LB_BIAS))
3255 return min(rq->cpu_load[type-1], total);
3259 * Return a high guess at the load of a migration-target cpu weighted
3260 * according to the scheduling class and "nice" value.
3262 static unsigned long target_load(int cpu, int type)
3264 struct rq *rq = cpu_rq(cpu);
3265 unsigned long total = weighted_cpuload(cpu);
3267 if (type == 0 || !sched_feat(LB_BIAS))
3270 return max(rq->cpu_load[type-1], total);
3273 static unsigned long power_of(int cpu)
3275 return cpu_rq(cpu)->cpu_power;
3278 static unsigned long cpu_avg_load_per_task(int cpu)
3280 struct rq *rq = cpu_rq(cpu);
3281 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3282 unsigned long load_avg = rq->cfs.runnable_load_avg;
3285 return load_avg / nr_running;
3290 static void record_wakee(struct task_struct *p)
3293 * Rough decay (wiping) for cost saving, don't worry
3294 * about the boundary, really active task won't care
3297 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3298 current->wakee_flips = 0;
3299 current->wakee_flip_decay_ts = jiffies;
3302 if (current->last_wakee != p) {
3303 current->last_wakee = p;
3304 current->wakee_flips++;
3308 static void task_waking_fair(struct task_struct *p)
3310 struct sched_entity *se = &p->se;
3311 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3314 #ifndef CONFIG_64BIT
3315 u64 min_vruntime_copy;
3318 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3320 min_vruntime = cfs_rq->min_vruntime;
3321 } while (min_vruntime != min_vruntime_copy);
3323 min_vruntime = cfs_rq->min_vruntime;
3326 se->vruntime -= min_vruntime;
3330 #ifdef CONFIG_FAIR_GROUP_SCHED
3332 * effective_load() calculates the load change as seen from the root_task_group
3334 * Adding load to a group doesn't make a group heavier, but can cause movement
3335 * of group shares between cpus. Assuming the shares were perfectly aligned one
3336 * can calculate the shift in shares.
3338 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3339 * on this @cpu and results in a total addition (subtraction) of @wg to the
3340 * total group weight.
3342 * Given a runqueue weight distribution (rw_i) we can compute a shares
3343 * distribution (s_i) using:
3345 * s_i = rw_i / \Sum rw_j (1)
3347 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3348 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3349 * shares distribution (s_i):
3351 * rw_i = { 2, 4, 1, 0 }
3352 * s_i = { 2/7, 4/7, 1/7, 0 }
3354 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3355 * task used to run on and the CPU the waker is running on), we need to
3356 * compute the effect of waking a task on either CPU and, in case of a sync
3357 * wakeup, compute the effect of the current task going to sleep.
3359 * So for a change of @wl to the local @cpu with an overall group weight change
3360 * of @wl we can compute the new shares distribution (s'_i) using:
3362 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3364 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3365 * differences in waking a task to CPU 0. The additional task changes the
3366 * weight and shares distributions like:
3368 * rw'_i = { 3, 4, 1, 0 }
3369 * s'_i = { 3/8, 4/8, 1/8, 0 }
3371 * We can then compute the difference in effective weight by using:
3373 * dw_i = S * (s'_i - s_i) (3)
3375 * Where 'S' is the group weight as seen by its parent.
3377 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3378 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3379 * 4/7) times the weight of the group.
3381 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3383 struct sched_entity *se = tg->se[cpu];
3385 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3388 for_each_sched_entity(se) {
3394 * W = @wg + \Sum rw_j
3396 W = wg + calc_tg_weight(tg, se->my_q);
3401 w = se->my_q->load.weight + wl;
3404 * wl = S * s'_i; see (2)
3407 wl = (w * tg->shares) / W;
3412 * Per the above, wl is the new se->load.weight value; since
3413 * those are clipped to [MIN_SHARES, ...) do so now. See
3414 * calc_cfs_shares().
3416 if (wl < MIN_SHARES)
3420 * wl = dw_i = S * (s'_i - s_i); see (3)
3422 wl -= se->load.weight;
3425 * Recursively apply this logic to all parent groups to compute
3426 * the final effective load change on the root group. Since
3427 * only the @tg group gets extra weight, all parent groups can
3428 * only redistribute existing shares. @wl is the shift in shares
3429 * resulting from this level per the above.
3438 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3445 static int wake_wide(struct task_struct *p)
3447 int factor = this_cpu_read(sd_llc_size);
3450 * Yeah, it's the switching-frequency, could means many wakee or
3451 * rapidly switch, use factor here will just help to automatically
3452 * adjust the loose-degree, so bigger node will lead to more pull.
3454 if (p->wakee_flips > factor) {
3456 * wakee is somewhat hot, it needs certain amount of cpu
3457 * resource, so if waker is far more hot, prefer to leave
3460 if (current->wakee_flips > (factor * p->wakee_flips))
3467 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3469 s64 this_load, load;
3470 int idx, this_cpu, prev_cpu;
3471 unsigned long tl_per_task;
3472 struct task_group *tg;
3473 unsigned long weight;
3477 * If we wake multiple tasks be careful to not bounce
3478 * ourselves around too much.
3484 this_cpu = smp_processor_id();
3485 prev_cpu = task_cpu(p);
3486 load = source_load(prev_cpu, idx);
3487 this_load = target_load(this_cpu, idx);
3490 * If sync wakeup then subtract the (maximum possible)
3491 * effect of the currently running task from the load
3492 * of the current CPU:
3495 tg = task_group(current);
3496 weight = current->se.load.weight;
3498 this_load += effective_load(tg, this_cpu, -weight, -weight);
3499 load += effective_load(tg, prev_cpu, 0, -weight);
3503 weight = p->se.load.weight;
3506 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3507 * due to the sync cause above having dropped this_load to 0, we'll
3508 * always have an imbalance, but there's really nothing you can do
3509 * about that, so that's good too.
3511 * Otherwise check if either cpus are near enough in load to allow this
3512 * task to be woken on this_cpu.
3514 if (this_load > 0) {
3515 s64 this_eff_load, prev_eff_load;
3517 this_eff_load = 100;
3518 this_eff_load *= power_of(prev_cpu);
3519 this_eff_load *= this_load +
3520 effective_load(tg, this_cpu, weight, weight);
3522 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3523 prev_eff_load *= power_of(this_cpu);
3524 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3526 balanced = this_eff_load <= prev_eff_load;
3531 * If the currently running task will sleep within
3532 * a reasonable amount of time then attract this newly
3535 if (sync && balanced)
3538 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3539 tl_per_task = cpu_avg_load_per_task(this_cpu);
3542 (this_load <= load &&
3543 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3545 * This domain has SD_WAKE_AFFINE and
3546 * p is cache cold in this domain, and
3547 * there is no bad imbalance.
3549 schedstat_inc(sd, ttwu_move_affine);
3550 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3558 * find_idlest_group finds and returns the least busy CPU group within the
3561 static struct sched_group *
3562 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3563 int this_cpu, int load_idx)
3565 struct sched_group *idlest = NULL, *group = sd->groups;
3566 unsigned long min_load = ULONG_MAX, this_load = 0;
3567 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3570 unsigned long load, avg_load;
3574 /* Skip over this group if it has no CPUs allowed */
3575 if (!cpumask_intersects(sched_group_cpus(group),
3576 tsk_cpus_allowed(p)))
3579 local_group = cpumask_test_cpu(this_cpu,
3580 sched_group_cpus(group));
3582 /* Tally up the load of all CPUs in the group */
3585 for_each_cpu(i, sched_group_cpus(group)) {
3586 /* Bias balancing toward cpus of our domain */
3588 load = source_load(i, load_idx);
3590 load = target_load(i, load_idx);
3595 /* Adjust by relative CPU power of the group */
3596 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3599 this_load = avg_load;
3600 } else if (avg_load < min_load) {
3601 min_load = avg_load;
3604 } while (group = group->next, group != sd->groups);
3606 if (!idlest || 100*this_load < imbalance*min_load)
3612 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3615 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3617 unsigned long load, min_load = ULONG_MAX;
3621 /* Traverse only the allowed CPUs */
3622 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3623 load = weighted_cpuload(i);
3625 if (load < min_load || (load == min_load && i == this_cpu)) {
3635 * Try and locate an idle CPU in the sched_domain.
3637 static int select_idle_sibling(struct task_struct *p, int target)
3639 struct sched_domain *sd;
3640 struct sched_group *sg;
3641 int i = task_cpu(p);
3643 if (idle_cpu(target))
3647 * If the prevous cpu is cache affine and idle, don't be stupid.
3649 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3653 * Otherwise, iterate the domains and find an elegible idle cpu.
3655 sd = rcu_dereference(per_cpu(sd_llc, target));
3656 for_each_lower_domain(sd) {
3659 if (!cpumask_intersects(sched_group_cpus(sg),
3660 tsk_cpus_allowed(p)))
3663 for_each_cpu(i, sched_group_cpus(sg)) {
3664 if (i == target || !idle_cpu(i))
3668 target = cpumask_first_and(sched_group_cpus(sg),
3669 tsk_cpus_allowed(p));
3673 } while (sg != sd->groups);
3680 * sched_balance_self: balance the current task (running on cpu) in domains
3681 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3684 * Balance, ie. select the least loaded group.
3686 * Returns the target CPU number, or the same CPU if no balancing is needed.
3688 * preempt must be disabled.
3691 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3693 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3694 int cpu = smp_processor_id();
3695 int prev_cpu = task_cpu(p);
3697 int want_affine = 0;
3698 int sync = wake_flags & WF_SYNC;
3700 if (p->nr_cpus_allowed == 1)
3703 if (sd_flag & SD_BALANCE_WAKE) {
3704 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3710 for_each_domain(cpu, tmp) {
3711 if (!(tmp->flags & SD_LOAD_BALANCE))
3715 * If both cpu and prev_cpu are part of this domain,
3716 * cpu is a valid SD_WAKE_AFFINE target.
3718 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3719 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3724 if (tmp->flags & sd_flag)
3729 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3732 new_cpu = select_idle_sibling(p, prev_cpu);
3737 int load_idx = sd->forkexec_idx;
3738 struct sched_group *group;
3741 if (!(sd->flags & sd_flag)) {
3746 if (sd_flag & SD_BALANCE_WAKE)
3747 load_idx = sd->wake_idx;
3749 group = find_idlest_group(sd, p, cpu, load_idx);
3755 new_cpu = find_idlest_cpu(group, p, cpu);
3756 if (new_cpu == -1 || new_cpu == cpu) {
3757 /* Now try balancing at a lower domain level of cpu */
3762 /* Now try balancing at a lower domain level of new_cpu */
3764 weight = sd->span_weight;
3766 for_each_domain(cpu, tmp) {
3767 if (weight <= tmp->span_weight)
3769 if (tmp->flags & sd_flag)
3772 /* while loop will break here if sd == NULL */
3781 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3782 * cfs_rq_of(p) references at time of call are still valid and identify the
3783 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3784 * other assumptions, including the state of rq->lock, should be made.
3787 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3789 struct sched_entity *se = &p->se;
3790 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3793 * Load tracking: accumulate removed load so that it can be processed
3794 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3795 * to blocked load iff they have a positive decay-count. It can never
3796 * be negative here since on-rq tasks have decay-count == 0.
3798 if (se->avg.decay_count) {
3799 se->avg.decay_count = -__synchronize_entity_decay(se);
3800 atomic_long_add(se->avg.load_avg_contrib,
3801 &cfs_rq->removed_load);
3804 #endif /* CONFIG_SMP */
3806 static unsigned long
3807 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3809 unsigned long gran = sysctl_sched_wakeup_granularity;
3812 * Since its curr running now, convert the gran from real-time
3813 * to virtual-time in his units.
3815 * By using 'se' instead of 'curr' we penalize light tasks, so
3816 * they get preempted easier. That is, if 'se' < 'curr' then
3817 * the resulting gran will be larger, therefore penalizing the
3818 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3819 * be smaller, again penalizing the lighter task.
3821 * This is especially important for buddies when the leftmost
3822 * task is higher priority than the buddy.
3824 return calc_delta_fair(gran, se);
3828 * Should 'se' preempt 'curr'.
3842 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3844 s64 gran, vdiff = curr->vruntime - se->vruntime;
3849 gran = wakeup_gran(curr, se);
3856 static void set_last_buddy(struct sched_entity *se)
3858 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3861 for_each_sched_entity(se)
3862 cfs_rq_of(se)->last = se;
3865 static void set_next_buddy(struct sched_entity *se)
3867 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3870 for_each_sched_entity(se)
3871 cfs_rq_of(se)->next = se;
3874 static void set_skip_buddy(struct sched_entity *se)
3876 for_each_sched_entity(se)
3877 cfs_rq_of(se)->skip = se;
3881 * Preempt the current task with a newly woken task if needed:
3883 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3885 struct task_struct *curr = rq->curr;
3886 struct sched_entity *se = &curr->se, *pse = &p->se;
3887 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3888 int scale = cfs_rq->nr_running >= sched_nr_latency;
3889 int next_buddy_marked = 0;
3891 if (unlikely(se == pse))
3895 * This is possible from callers such as move_task(), in which we
3896 * unconditionally check_prempt_curr() after an enqueue (which may have
3897 * lead to a throttle). This both saves work and prevents false
3898 * next-buddy nomination below.
3900 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3903 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3904 set_next_buddy(pse);
3905 next_buddy_marked = 1;
3909 * We can come here with TIF_NEED_RESCHED already set from new task
3912 * Note: this also catches the edge-case of curr being in a throttled
3913 * group (e.g. via set_curr_task), since update_curr() (in the
3914 * enqueue of curr) will have resulted in resched being set. This
3915 * prevents us from potentially nominating it as a false LAST_BUDDY
3918 if (test_tsk_need_resched(curr))
3921 /* Idle tasks are by definition preempted by non-idle tasks. */
3922 if (unlikely(curr->policy == SCHED_IDLE) &&
3923 likely(p->policy != SCHED_IDLE))
3927 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3928 * is driven by the tick):
3930 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3933 find_matching_se(&se, &pse);
3934 update_curr(cfs_rq_of(se));
3936 if (wakeup_preempt_entity(se, pse) == 1) {
3938 * Bias pick_next to pick the sched entity that is
3939 * triggering this preemption.
3941 if (!next_buddy_marked)
3942 set_next_buddy(pse);
3951 * Only set the backward buddy when the current task is still
3952 * on the rq. This can happen when a wakeup gets interleaved
3953 * with schedule on the ->pre_schedule() or idle_balance()
3954 * point, either of which can * drop the rq lock.
3956 * Also, during early boot the idle thread is in the fair class,
3957 * for obvious reasons its a bad idea to schedule back to it.
3959 if (unlikely(!se->on_rq || curr == rq->idle))
3962 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3966 static struct task_struct *pick_next_task_fair(struct rq *rq)
3968 struct task_struct *p;
3969 struct cfs_rq *cfs_rq = &rq->cfs;
3970 struct sched_entity *se;
3972 if (!cfs_rq->nr_running)
3976 se = pick_next_entity(cfs_rq);
3977 set_next_entity(cfs_rq, se);
3978 cfs_rq = group_cfs_rq(se);
3982 if (hrtick_enabled(rq))
3983 hrtick_start_fair(rq, p);
3989 * Account for a descheduled task:
3991 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3993 struct sched_entity *se = &prev->se;
3994 struct cfs_rq *cfs_rq;
3996 for_each_sched_entity(se) {
3997 cfs_rq = cfs_rq_of(se);
3998 put_prev_entity(cfs_rq, se);
4003 * sched_yield() is very simple
4005 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4007 static void yield_task_fair(struct rq *rq)
4009 struct task_struct *curr = rq->curr;
4010 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4011 struct sched_entity *se = &curr->se;
4014 * Are we the only task in the tree?
4016 if (unlikely(rq->nr_running == 1))
4019 clear_buddies(cfs_rq, se);
4021 if (curr->policy != SCHED_BATCH) {
4022 update_rq_clock(rq);
4024 * Update run-time statistics of the 'current'.
4026 update_curr(cfs_rq);
4028 * Tell update_rq_clock() that we've just updated,
4029 * so we don't do microscopic update in schedule()
4030 * and double the fastpath cost.
4032 rq->skip_clock_update = 1;
4038 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4040 struct sched_entity *se = &p->se;
4042 /* throttled hierarchies are not runnable */
4043 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4046 /* Tell the scheduler that we'd really like pse to run next. */
4049 yield_task_fair(rq);
4055 /**************************************************
4056 * Fair scheduling class load-balancing methods.
4060 * The purpose of load-balancing is to achieve the same basic fairness the
4061 * per-cpu scheduler provides, namely provide a proportional amount of compute
4062 * time to each task. This is expressed in the following equation:
4064 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4066 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4067 * W_i,0 is defined as:
4069 * W_i,0 = \Sum_j w_i,j (2)
4071 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4072 * is derived from the nice value as per prio_to_weight[].
4074 * The weight average is an exponential decay average of the instantaneous
4077 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4079 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4080 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4081 * can also include other factors [XXX].
4083 * To achieve this balance we define a measure of imbalance which follows
4084 * directly from (1):
4086 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4088 * We them move tasks around to minimize the imbalance. In the continuous
4089 * function space it is obvious this converges, in the discrete case we get
4090 * a few fun cases generally called infeasible weight scenarios.
4093 * - infeasible weights;
4094 * - local vs global optima in the discrete case. ]
4099 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4100 * for all i,j solution, we create a tree of cpus that follows the hardware
4101 * topology where each level pairs two lower groups (or better). This results
4102 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4103 * tree to only the first of the previous level and we decrease the frequency
4104 * of load-balance at each level inv. proportional to the number of cpus in
4110 * \Sum { --- * --- * 2^i } = O(n) (5)
4112 * `- size of each group
4113 * | | `- number of cpus doing load-balance
4115 * `- sum over all levels
4117 * Coupled with a limit on how many tasks we can migrate every balance pass,
4118 * this makes (5) the runtime complexity of the balancer.
4120 * An important property here is that each CPU is still (indirectly) connected
4121 * to every other cpu in at most O(log n) steps:
4123 * The adjacency matrix of the resulting graph is given by:
4126 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4129 * And you'll find that:
4131 * A^(log_2 n)_i,j != 0 for all i,j (7)
4133 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4134 * The task movement gives a factor of O(m), giving a convergence complexity
4137 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4142 * In order to avoid CPUs going idle while there's still work to do, new idle
4143 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4144 * tree itself instead of relying on other CPUs to bring it work.
4146 * This adds some complexity to both (5) and (8) but it reduces the total idle
4154 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4157 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4162 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4164 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4166 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4169 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4170 * rewrite all of this once again.]
4173 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4175 #define LBF_ALL_PINNED 0x01
4176 #define LBF_NEED_BREAK 0x02
4177 #define LBF_DST_PINNED 0x04
4178 #define LBF_SOME_PINNED 0x08
4181 struct sched_domain *sd;
4189 struct cpumask *dst_grpmask;
4191 enum cpu_idle_type idle;
4193 /* The set of CPUs under consideration for load-balancing */
4194 struct cpumask *cpus;
4199 unsigned int loop_break;
4200 unsigned int loop_max;
4204 * move_task - move a task from one runqueue to another runqueue.
4205 * Both runqueues must be locked.
4207 static void move_task(struct task_struct *p, struct lb_env *env)
4209 deactivate_task(env->src_rq, p, 0);
4210 set_task_cpu(p, env->dst_cpu);
4211 activate_task(env->dst_rq, p, 0);
4212 check_preempt_curr(env->dst_rq, p, 0);
4213 #ifdef CONFIG_NUMA_BALANCING
4214 if (p->numa_preferred_nid != -1) {
4215 int src_nid = cpu_to_node(env->src_cpu);
4216 int dst_nid = cpu_to_node(env->dst_cpu);
4219 * If the load balancer has moved the task then limit
4220 * migrations from taking place in the short term in
4221 * case this is a short-lived migration.
4223 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4224 p->numa_migrate_seq = 0;
4230 * Is this task likely cache-hot:
4233 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4237 if (p->sched_class != &fair_sched_class)
4240 if (unlikely(p->policy == SCHED_IDLE))
4244 * Buddy candidates are cache hot:
4246 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4247 (&p->se == cfs_rq_of(&p->se)->next ||
4248 &p->se == cfs_rq_of(&p->se)->last))
4251 if (sysctl_sched_migration_cost == -1)
4253 if (sysctl_sched_migration_cost == 0)
4256 delta = now - p->se.exec_start;
4258 return delta < (s64)sysctl_sched_migration_cost;
4261 #ifdef CONFIG_NUMA_BALANCING
4262 /* Returns true if the destination node has incurred more faults */
4263 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4265 int src_nid, dst_nid;
4267 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4268 !(env->sd->flags & SD_NUMA)) {
4272 src_nid = cpu_to_node(env->src_cpu);
4273 dst_nid = cpu_to_node(env->dst_cpu);
4275 if (src_nid == dst_nid ||
4276 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4279 if (dst_nid == p->numa_preferred_nid ||
4280 task_faults(p, dst_nid) > task_faults(p, src_nid))
4287 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4289 int src_nid, dst_nid;
4291 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4294 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4297 src_nid = cpu_to_node(env->src_cpu);
4298 dst_nid = cpu_to_node(env->dst_cpu);
4300 if (src_nid == dst_nid ||
4301 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4304 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4311 static inline bool migrate_improves_locality(struct task_struct *p,
4317 static inline bool migrate_degrades_locality(struct task_struct *p,
4325 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4328 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4330 int tsk_cache_hot = 0;
4332 * We do not migrate tasks that are:
4333 * 1) throttled_lb_pair, or
4334 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4335 * 3) running (obviously), or
4336 * 4) are cache-hot on their current CPU.
4338 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4341 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4344 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4346 env->flags |= LBF_SOME_PINNED;
4349 * Remember if this task can be migrated to any other cpu in
4350 * our sched_group. We may want to revisit it if we couldn't
4351 * meet load balance goals by pulling other tasks on src_cpu.
4353 * Also avoid computing new_dst_cpu if we have already computed
4354 * one in current iteration.
4356 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4359 /* Prevent to re-select dst_cpu via env's cpus */
4360 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4361 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4362 env->flags |= LBF_DST_PINNED;
4363 env->new_dst_cpu = cpu;
4371 /* Record that we found atleast one task that could run on dst_cpu */
4372 env->flags &= ~LBF_ALL_PINNED;
4374 if (task_running(env->src_rq, p)) {
4375 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4380 * Aggressive migration if:
4381 * 1) destination numa is preferred
4382 * 2) task is cache cold, or
4383 * 3) too many balance attempts have failed.
4385 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4387 tsk_cache_hot = migrate_degrades_locality(p, env);
4389 if (migrate_improves_locality(p, env)) {
4390 #ifdef CONFIG_SCHEDSTATS
4391 if (tsk_cache_hot) {
4392 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4393 schedstat_inc(p, se.statistics.nr_forced_migrations);
4399 if (!tsk_cache_hot ||
4400 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4402 if (tsk_cache_hot) {
4403 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4404 schedstat_inc(p, se.statistics.nr_forced_migrations);
4410 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4415 * move_one_task tries to move exactly one task from busiest to this_rq, as
4416 * part of active balancing operations within "domain".
4417 * Returns 1 if successful and 0 otherwise.
4419 * Called with both runqueues locked.
4421 static int move_one_task(struct lb_env *env)
4423 struct task_struct *p, *n;
4425 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4426 if (!can_migrate_task(p, env))
4431 * Right now, this is only the second place move_task()
4432 * is called, so we can safely collect move_task()
4433 * stats here rather than inside move_task().
4435 schedstat_inc(env->sd, lb_gained[env->idle]);
4441 static unsigned long task_h_load(struct task_struct *p);
4443 static const unsigned int sched_nr_migrate_break = 32;
4446 * move_tasks tries to move up to imbalance weighted load from busiest to
4447 * this_rq, as part of a balancing operation within domain "sd".
4448 * Returns 1 if successful and 0 otherwise.
4450 * Called with both runqueues locked.
4452 static int move_tasks(struct lb_env *env)
4454 struct list_head *tasks = &env->src_rq->cfs_tasks;
4455 struct task_struct *p;
4459 if (env->imbalance <= 0)
4462 while (!list_empty(tasks)) {
4463 p = list_first_entry(tasks, struct task_struct, se.group_node);
4466 /* We've more or less seen every task there is, call it quits */
4467 if (env->loop > env->loop_max)
4470 /* take a breather every nr_migrate tasks */
4471 if (env->loop > env->loop_break) {
4472 env->loop_break += sched_nr_migrate_break;
4473 env->flags |= LBF_NEED_BREAK;
4477 if (!can_migrate_task(p, env))
4480 load = task_h_load(p);
4482 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4485 if ((load / 2) > env->imbalance)
4490 env->imbalance -= load;
4492 #ifdef CONFIG_PREEMPT
4494 * NEWIDLE balancing is a source of latency, so preemptible
4495 * kernels will stop after the first task is pulled to minimize
4496 * the critical section.
4498 if (env->idle == CPU_NEWLY_IDLE)
4503 * We only want to steal up to the prescribed amount of
4506 if (env->imbalance <= 0)
4511 list_move_tail(&p->se.group_node, tasks);
4515 * Right now, this is one of only two places move_task() is called,
4516 * so we can safely collect move_task() stats here rather than
4517 * inside move_task().
4519 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4524 #ifdef CONFIG_FAIR_GROUP_SCHED
4526 * update tg->load_weight by folding this cpu's load_avg
4528 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4530 struct sched_entity *se = tg->se[cpu];
4531 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4533 /* throttled entities do not contribute to load */
4534 if (throttled_hierarchy(cfs_rq))
4537 update_cfs_rq_blocked_load(cfs_rq, 1);
4540 update_entity_load_avg(se, 1);
4542 * We pivot on our runnable average having decayed to zero for
4543 * list removal. This generally implies that all our children
4544 * have also been removed (modulo rounding error or bandwidth
4545 * control); however, such cases are rare and we can fix these
4548 * TODO: fix up out-of-order children on enqueue.
4550 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4551 list_del_leaf_cfs_rq(cfs_rq);
4553 struct rq *rq = rq_of(cfs_rq);
4554 update_rq_runnable_avg(rq, rq->nr_running);
4558 static void update_blocked_averages(int cpu)
4560 struct rq *rq = cpu_rq(cpu);
4561 struct cfs_rq *cfs_rq;
4562 unsigned long flags;
4564 raw_spin_lock_irqsave(&rq->lock, flags);
4565 update_rq_clock(rq);
4567 * Iterates the task_group tree in a bottom up fashion, see
4568 * list_add_leaf_cfs_rq() for details.
4570 for_each_leaf_cfs_rq(rq, cfs_rq) {
4572 * Note: We may want to consider periodically releasing
4573 * rq->lock about these updates so that creating many task
4574 * groups does not result in continually extending hold time.
4576 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4579 raw_spin_unlock_irqrestore(&rq->lock, flags);
4583 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4584 * This needs to be done in a top-down fashion because the load of a child
4585 * group is a fraction of its parents load.
4587 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4589 struct rq *rq = rq_of(cfs_rq);
4590 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4591 unsigned long now = jiffies;
4594 if (cfs_rq->last_h_load_update == now)
4597 cfs_rq->h_load_next = NULL;
4598 for_each_sched_entity(se) {
4599 cfs_rq = cfs_rq_of(se);
4600 cfs_rq->h_load_next = se;
4601 if (cfs_rq->last_h_load_update == now)
4606 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4607 cfs_rq->last_h_load_update = now;
4610 while ((se = cfs_rq->h_load_next) != NULL) {
4611 load = cfs_rq->h_load;
4612 load = div64_ul(load * se->avg.load_avg_contrib,
4613 cfs_rq->runnable_load_avg + 1);
4614 cfs_rq = group_cfs_rq(se);
4615 cfs_rq->h_load = load;
4616 cfs_rq->last_h_load_update = now;
4620 static unsigned long task_h_load(struct task_struct *p)
4622 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4624 update_cfs_rq_h_load(cfs_rq);
4625 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4626 cfs_rq->runnable_load_avg + 1);
4629 static inline void update_blocked_averages(int cpu)
4633 static unsigned long task_h_load(struct task_struct *p)
4635 return p->se.avg.load_avg_contrib;
4639 /********** Helpers for find_busiest_group ************************/
4641 * sg_lb_stats - stats of a sched_group required for load_balancing
4643 struct sg_lb_stats {
4644 unsigned long avg_load; /*Avg load across the CPUs of the group */
4645 unsigned long group_load; /* Total load over the CPUs of the group */
4646 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4647 unsigned long load_per_task;
4648 unsigned long group_power;
4649 unsigned int sum_nr_running; /* Nr tasks running in the group */
4650 unsigned int group_capacity;
4651 unsigned int idle_cpus;
4652 unsigned int group_weight;
4653 int group_imb; /* Is there an imbalance in the group ? */
4654 int group_has_capacity; /* Is there extra capacity in the group? */
4658 * sd_lb_stats - Structure to store the statistics of a sched_domain
4659 * during load balancing.
4661 struct sd_lb_stats {
4662 struct sched_group *busiest; /* Busiest group in this sd */
4663 struct sched_group *local; /* Local group in this sd */
4664 unsigned long total_load; /* Total load of all groups in sd */
4665 unsigned long total_pwr; /* Total power of all groups in sd */
4666 unsigned long avg_load; /* Average load across all groups in sd */
4668 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4669 struct sg_lb_stats local_stat; /* Statistics of the local group */
4672 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4675 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4676 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4677 * We must however clear busiest_stat::avg_load because
4678 * update_sd_pick_busiest() reads this before assignment.
4680 *sds = (struct sd_lb_stats){
4692 * get_sd_load_idx - Obtain the load index for a given sched domain.
4693 * @sd: The sched_domain whose load_idx is to be obtained.
4694 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4696 * Return: The load index.
4698 static inline int get_sd_load_idx(struct sched_domain *sd,
4699 enum cpu_idle_type idle)
4705 load_idx = sd->busy_idx;
4708 case CPU_NEWLY_IDLE:
4709 load_idx = sd->newidle_idx;
4712 load_idx = sd->idle_idx;
4719 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4721 return SCHED_POWER_SCALE;
4724 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4726 return default_scale_freq_power(sd, cpu);
4729 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4731 unsigned long weight = sd->span_weight;
4732 unsigned long smt_gain = sd->smt_gain;
4739 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4741 return default_scale_smt_power(sd, cpu);
4744 static unsigned long scale_rt_power(int cpu)
4746 struct rq *rq = cpu_rq(cpu);
4747 u64 total, available, age_stamp, avg;
4750 * Since we're reading these variables without serialization make sure
4751 * we read them once before doing sanity checks on them.
4753 age_stamp = ACCESS_ONCE(rq->age_stamp);
4754 avg = ACCESS_ONCE(rq->rt_avg);
4756 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4758 if (unlikely(total < avg)) {
4759 /* Ensures that power won't end up being negative */
4762 available = total - avg;
4765 if (unlikely((s64)total < SCHED_POWER_SCALE))
4766 total = SCHED_POWER_SCALE;
4768 total >>= SCHED_POWER_SHIFT;
4770 return div_u64(available, total);
4773 static void update_cpu_power(struct sched_domain *sd, int cpu)
4775 unsigned long weight = sd->span_weight;
4776 unsigned long power = SCHED_POWER_SCALE;
4777 struct sched_group *sdg = sd->groups;
4779 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4780 if (sched_feat(ARCH_POWER))
4781 power *= arch_scale_smt_power(sd, cpu);
4783 power *= default_scale_smt_power(sd, cpu);
4785 power >>= SCHED_POWER_SHIFT;
4788 sdg->sgp->power_orig = power;
4790 if (sched_feat(ARCH_POWER))
4791 power *= arch_scale_freq_power(sd, cpu);
4793 power *= default_scale_freq_power(sd, cpu);
4795 power >>= SCHED_POWER_SHIFT;
4797 power *= scale_rt_power(cpu);
4798 power >>= SCHED_POWER_SHIFT;
4803 cpu_rq(cpu)->cpu_power = power;
4804 sdg->sgp->power = power;
4807 void update_group_power(struct sched_domain *sd, int cpu)
4809 struct sched_domain *child = sd->child;
4810 struct sched_group *group, *sdg = sd->groups;
4811 unsigned long power, power_orig;
4812 unsigned long interval;
4814 interval = msecs_to_jiffies(sd->balance_interval);
4815 interval = clamp(interval, 1UL, max_load_balance_interval);
4816 sdg->sgp->next_update = jiffies + interval;
4819 update_cpu_power(sd, cpu);
4823 power_orig = power = 0;
4825 if (child->flags & SD_OVERLAP) {
4827 * SD_OVERLAP domains cannot assume that child groups
4828 * span the current group.
4831 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4832 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4834 power_orig += sg->sgp->power_orig;
4835 power += sg->sgp->power;
4839 * !SD_OVERLAP domains can assume that child groups
4840 * span the current group.
4843 group = child->groups;
4845 power_orig += group->sgp->power_orig;
4846 power += group->sgp->power;
4847 group = group->next;
4848 } while (group != child->groups);
4851 sdg->sgp->power_orig = power_orig;
4852 sdg->sgp->power = power;
4856 * Try and fix up capacity for tiny siblings, this is needed when
4857 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4858 * which on its own isn't powerful enough.
4860 * See update_sd_pick_busiest() and check_asym_packing().
4863 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4866 * Only siblings can have significantly less than SCHED_POWER_SCALE
4868 if (!(sd->flags & SD_SHARE_CPUPOWER))
4872 * If ~90% of the cpu_power is still there, we're good.
4874 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4881 * Group imbalance indicates (and tries to solve) the problem where balancing
4882 * groups is inadequate due to tsk_cpus_allowed() constraints.
4884 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4885 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4888 * { 0 1 2 3 } { 4 5 6 7 }
4891 * If we were to balance group-wise we'd place two tasks in the first group and
4892 * two tasks in the second group. Clearly this is undesired as it will overload
4893 * cpu 3 and leave one of the cpus in the second group unused.
4895 * The current solution to this issue is detecting the skew in the first group
4896 * by noticing the lower domain failed to reach balance and had difficulty
4897 * moving tasks due to affinity constraints.
4899 * When this is so detected; this group becomes a candidate for busiest; see
4900 * update_sd_pick_busiest(). And calculcate_imbalance() and
4901 * find_busiest_group() avoid some of the usual balance conditions to allow it
4902 * to create an effective group imbalance.
4904 * This is a somewhat tricky proposition since the next run might not find the
4905 * group imbalance and decide the groups need to be balanced again. A most
4906 * subtle and fragile situation.
4909 static inline int sg_imbalanced(struct sched_group *group)
4911 return group->sgp->imbalance;
4915 * Compute the group capacity.
4917 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4918 * first dividing out the smt factor and computing the actual number of cores
4919 * and limit power unit capacity with that.
4921 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4923 unsigned int capacity, smt, cpus;
4924 unsigned int power, power_orig;
4926 power = group->sgp->power;
4927 power_orig = group->sgp->power_orig;
4928 cpus = group->group_weight;
4930 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4931 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4932 capacity = cpus / smt; /* cores */
4934 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4936 capacity = fix_small_capacity(env->sd, group);
4942 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4943 * @env: The load balancing environment.
4944 * @group: sched_group whose statistics are to be updated.
4945 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4946 * @local_group: Does group contain this_cpu.
4947 * @sgs: variable to hold the statistics for this group.
4949 static inline void update_sg_lb_stats(struct lb_env *env,
4950 struct sched_group *group, int load_idx,
4951 int local_group, struct sg_lb_stats *sgs)
4953 unsigned long nr_running;
4957 memset(sgs, 0, sizeof(*sgs));
4959 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4960 struct rq *rq = cpu_rq(i);
4962 nr_running = rq->nr_running;
4964 /* Bias balancing toward cpus of our domain */
4966 load = target_load(i, load_idx);
4968 load = source_load(i, load_idx);
4970 sgs->group_load += load;
4971 sgs->sum_nr_running += nr_running;
4972 sgs->sum_weighted_load += weighted_cpuload(i);
4977 /* Adjust by relative CPU power of the group */
4978 sgs->group_power = group->sgp->power;
4979 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4981 if (sgs->sum_nr_running)
4982 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4984 sgs->group_weight = group->group_weight;
4986 sgs->group_imb = sg_imbalanced(group);
4987 sgs->group_capacity = sg_capacity(env, group);
4989 if (sgs->group_capacity > sgs->sum_nr_running)
4990 sgs->group_has_capacity = 1;
4994 * update_sd_pick_busiest - return 1 on busiest group
4995 * @env: The load balancing environment.
4996 * @sds: sched_domain statistics
4997 * @sg: sched_group candidate to be checked for being the busiest
4998 * @sgs: sched_group statistics
5000 * Determine if @sg is a busier group than the previously selected
5003 * Return: %true if @sg is a busier group than the previously selected
5004 * busiest group. %false otherwise.
5006 static bool update_sd_pick_busiest(struct lb_env *env,
5007 struct sd_lb_stats *sds,
5008 struct sched_group *sg,
5009 struct sg_lb_stats *sgs)
5011 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5014 if (sgs->sum_nr_running > sgs->group_capacity)
5021 * ASYM_PACKING needs to move all the work to the lowest
5022 * numbered CPUs in the group, therefore mark all groups
5023 * higher than ourself as busy.
5025 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5026 env->dst_cpu < group_first_cpu(sg)) {
5030 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5038 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5039 * @env: The load balancing environment.
5040 * @balance: Should we balance.
5041 * @sds: variable to hold the statistics for this sched_domain.
5043 static inline void update_sd_lb_stats(struct lb_env *env,
5044 struct sd_lb_stats *sds)
5046 struct sched_domain *child = env->sd->child;
5047 struct sched_group *sg = env->sd->groups;
5048 struct sg_lb_stats tmp_sgs;
5049 int load_idx, prefer_sibling = 0;
5051 if (child && child->flags & SD_PREFER_SIBLING)
5054 load_idx = get_sd_load_idx(env->sd, env->idle);
5057 struct sg_lb_stats *sgs = &tmp_sgs;
5060 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5063 sgs = &sds->local_stat;
5065 if (env->idle != CPU_NEWLY_IDLE ||
5066 time_after_eq(jiffies, sg->sgp->next_update))
5067 update_group_power(env->sd, env->dst_cpu);
5070 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5076 * In case the child domain prefers tasks go to siblings
5077 * first, lower the sg capacity to one so that we'll try
5078 * and move all the excess tasks away. We lower the capacity
5079 * of a group only if the local group has the capacity to fit
5080 * these excess tasks, i.e. nr_running < group_capacity. The
5081 * extra check prevents the case where you always pull from the
5082 * heaviest group when it is already under-utilized (possible
5083 * with a large weight task outweighs the tasks on the system).
5085 if (prefer_sibling && sds->local &&
5086 sds->local_stat.group_has_capacity)
5087 sgs->group_capacity = min(sgs->group_capacity, 1U);
5089 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5091 sds->busiest_stat = *sgs;
5095 /* Now, start updating sd_lb_stats */
5096 sds->total_load += sgs->group_load;
5097 sds->total_pwr += sgs->group_power;
5100 } while (sg != env->sd->groups);
5104 * check_asym_packing - Check to see if the group is packed into the
5107 * This is primarily intended to used at the sibling level. Some
5108 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5109 * case of POWER7, it can move to lower SMT modes only when higher
5110 * threads are idle. When in lower SMT modes, the threads will
5111 * perform better since they share less core resources. Hence when we
5112 * have idle threads, we want them to be the higher ones.
5114 * This packing function is run on idle threads. It checks to see if
5115 * the busiest CPU in this domain (core in the P7 case) has a higher
5116 * CPU number than the packing function is being run on. Here we are
5117 * assuming lower CPU number will be equivalent to lower a SMT thread
5120 * Return: 1 when packing is required and a task should be moved to
5121 * this CPU. The amount of the imbalance is returned in *imbalance.
5123 * @env: The load balancing environment.
5124 * @sds: Statistics of the sched_domain which is to be packed
5126 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5130 if (!(env->sd->flags & SD_ASYM_PACKING))
5136 busiest_cpu = group_first_cpu(sds->busiest);
5137 if (env->dst_cpu > busiest_cpu)
5140 env->imbalance = DIV_ROUND_CLOSEST(
5141 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5148 * fix_small_imbalance - Calculate the minor imbalance that exists
5149 * amongst the groups of a sched_domain, during
5151 * @env: The load balancing environment.
5152 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5155 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5157 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5158 unsigned int imbn = 2;
5159 unsigned long scaled_busy_load_per_task;
5160 struct sg_lb_stats *local, *busiest;
5162 local = &sds->local_stat;
5163 busiest = &sds->busiest_stat;
5165 if (!local->sum_nr_running)
5166 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5167 else if (busiest->load_per_task > local->load_per_task)
5170 scaled_busy_load_per_task =
5171 (busiest->load_per_task * SCHED_POWER_SCALE) /
5172 busiest->group_power;
5174 if (busiest->avg_load + scaled_busy_load_per_task >=
5175 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5176 env->imbalance = busiest->load_per_task;
5181 * OK, we don't have enough imbalance to justify moving tasks,
5182 * however we may be able to increase total CPU power used by
5186 pwr_now += busiest->group_power *
5187 min(busiest->load_per_task, busiest->avg_load);
5188 pwr_now += local->group_power *
5189 min(local->load_per_task, local->avg_load);
5190 pwr_now /= SCHED_POWER_SCALE;
5192 /* Amount of load we'd subtract */
5193 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5194 busiest->group_power;
5195 if (busiest->avg_load > tmp) {
5196 pwr_move += busiest->group_power *
5197 min(busiest->load_per_task,
5198 busiest->avg_load - tmp);
5201 /* Amount of load we'd add */
5202 if (busiest->avg_load * busiest->group_power <
5203 busiest->load_per_task * SCHED_POWER_SCALE) {
5204 tmp = (busiest->avg_load * busiest->group_power) /
5207 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5210 pwr_move += local->group_power *
5211 min(local->load_per_task, local->avg_load + tmp);
5212 pwr_move /= SCHED_POWER_SCALE;
5214 /* Move if we gain throughput */
5215 if (pwr_move > pwr_now)
5216 env->imbalance = busiest->load_per_task;
5220 * calculate_imbalance - Calculate the amount of imbalance present within the
5221 * groups of a given sched_domain during load balance.
5222 * @env: load balance environment
5223 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5225 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5227 unsigned long max_pull, load_above_capacity = ~0UL;
5228 struct sg_lb_stats *local, *busiest;
5230 local = &sds->local_stat;
5231 busiest = &sds->busiest_stat;
5233 if (busiest->group_imb) {
5235 * In the group_imb case we cannot rely on group-wide averages
5236 * to ensure cpu-load equilibrium, look at wider averages. XXX
5238 busiest->load_per_task =
5239 min(busiest->load_per_task, sds->avg_load);
5243 * In the presence of smp nice balancing, certain scenarios can have
5244 * max load less than avg load(as we skip the groups at or below
5245 * its cpu_power, while calculating max_load..)
5247 if (busiest->avg_load <= sds->avg_load ||
5248 local->avg_load >= sds->avg_load) {
5250 return fix_small_imbalance(env, sds);
5253 if (!busiest->group_imb) {
5255 * Don't want to pull so many tasks that a group would go idle.
5256 * Except of course for the group_imb case, since then we might
5257 * have to drop below capacity to reach cpu-load equilibrium.
5259 load_above_capacity =
5260 (busiest->sum_nr_running - busiest->group_capacity);
5262 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5263 load_above_capacity /= busiest->group_power;
5267 * We're trying to get all the cpus to the average_load, so we don't
5268 * want to push ourselves above the average load, nor do we wish to
5269 * reduce the max loaded cpu below the average load. At the same time,
5270 * we also don't want to reduce the group load below the group capacity
5271 * (so that we can implement power-savings policies etc). Thus we look
5272 * for the minimum possible imbalance.
5274 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5276 /* How much load to actually move to equalise the imbalance */
5277 env->imbalance = min(
5278 max_pull * busiest->group_power,
5279 (sds->avg_load - local->avg_load) * local->group_power
5280 ) / SCHED_POWER_SCALE;
5283 * if *imbalance is less than the average load per runnable task
5284 * there is no guarantee that any tasks will be moved so we'll have
5285 * a think about bumping its value to force at least one task to be
5288 if (env->imbalance < busiest->load_per_task)
5289 return fix_small_imbalance(env, sds);
5292 /******* find_busiest_group() helpers end here *********************/
5295 * find_busiest_group - Returns the busiest group within the sched_domain
5296 * if there is an imbalance. If there isn't an imbalance, and
5297 * the user has opted for power-savings, it returns a group whose
5298 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5299 * such a group exists.
5301 * Also calculates the amount of weighted load which should be moved
5302 * to restore balance.
5304 * @env: The load balancing environment.
5306 * Return: - The busiest group if imbalance exists.
5307 * - If no imbalance and user has opted for power-savings balance,
5308 * return the least loaded group whose CPUs can be
5309 * put to idle by rebalancing its tasks onto our group.
5311 static struct sched_group *find_busiest_group(struct lb_env *env)
5313 struct sg_lb_stats *local, *busiest;
5314 struct sd_lb_stats sds;
5316 init_sd_lb_stats(&sds);
5319 * Compute the various statistics relavent for load balancing at
5322 update_sd_lb_stats(env, &sds);
5323 local = &sds.local_stat;
5324 busiest = &sds.busiest_stat;
5326 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5327 check_asym_packing(env, &sds))
5330 /* There is no busy sibling group to pull tasks from */
5331 if (!sds.busiest || busiest->sum_nr_running == 0)
5334 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5337 * If the busiest group is imbalanced the below checks don't
5338 * work because they assume all things are equal, which typically
5339 * isn't true due to cpus_allowed constraints and the like.
5341 if (busiest->group_imb)
5344 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5345 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5346 !busiest->group_has_capacity)
5350 * If the local group is more busy than the selected busiest group
5351 * don't try and pull any tasks.
5353 if (local->avg_load >= busiest->avg_load)
5357 * Don't pull any tasks if this group is already above the domain
5360 if (local->avg_load >= sds.avg_load)
5363 if (env->idle == CPU_IDLE) {
5365 * This cpu is idle. If the busiest group load doesn't
5366 * have more tasks than the number of available cpu's and
5367 * there is no imbalance between this and busiest group
5368 * wrt to idle cpu's, it is balanced.
5370 if ((local->idle_cpus < busiest->idle_cpus) &&
5371 busiest->sum_nr_running <= busiest->group_weight)
5375 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5376 * imbalance_pct to be conservative.
5378 if (100 * busiest->avg_load <=
5379 env->sd->imbalance_pct * local->avg_load)
5384 /* Looks like there is an imbalance. Compute it */
5385 calculate_imbalance(env, &sds);
5394 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5396 static struct rq *find_busiest_queue(struct lb_env *env,
5397 struct sched_group *group)
5399 struct rq *busiest = NULL, *rq;
5400 unsigned long busiest_load = 0, busiest_power = 1;
5403 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5404 unsigned long power = power_of(i);
5405 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5410 capacity = fix_small_capacity(env->sd, group);
5413 wl = weighted_cpuload(i);
5416 * When comparing with imbalance, use weighted_cpuload()
5417 * which is not scaled with the cpu power.
5419 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5423 * For the load comparisons with the other cpu's, consider
5424 * the weighted_cpuload() scaled with the cpu power, so that
5425 * the load can be moved away from the cpu that is potentially
5426 * running at a lower capacity.
5428 * Thus we're looking for max(wl_i / power_i), crosswise
5429 * multiplication to rid ourselves of the division works out
5430 * to: wl_i * power_j > wl_j * power_i; where j is our
5433 if (wl * busiest_power > busiest_load * power) {
5435 busiest_power = power;
5444 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5445 * so long as it is large enough.
5447 #define MAX_PINNED_INTERVAL 512
5449 /* Working cpumask for load_balance and load_balance_newidle. */
5450 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5452 static int need_active_balance(struct lb_env *env)
5454 struct sched_domain *sd = env->sd;
5456 if (env->idle == CPU_NEWLY_IDLE) {
5459 * ASYM_PACKING needs to force migrate tasks from busy but
5460 * higher numbered CPUs in order to pack all tasks in the
5461 * lowest numbered CPUs.
5463 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5467 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5470 static int active_load_balance_cpu_stop(void *data);
5472 static int should_we_balance(struct lb_env *env)
5474 struct sched_group *sg = env->sd->groups;
5475 struct cpumask *sg_cpus, *sg_mask;
5476 int cpu, balance_cpu = -1;
5479 * In the newly idle case, we will allow all the cpu's
5480 * to do the newly idle load balance.
5482 if (env->idle == CPU_NEWLY_IDLE)
5485 sg_cpus = sched_group_cpus(sg);
5486 sg_mask = sched_group_mask(sg);
5487 /* Try to find first idle cpu */
5488 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5489 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5496 if (balance_cpu == -1)
5497 balance_cpu = group_balance_cpu(sg);
5500 * First idle cpu or the first cpu(busiest) in this sched group
5501 * is eligible for doing load balancing at this and above domains.
5503 return balance_cpu == env->dst_cpu;
5507 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5508 * tasks if there is an imbalance.
5510 static int load_balance(int this_cpu, struct rq *this_rq,
5511 struct sched_domain *sd, enum cpu_idle_type idle,
5512 int *continue_balancing)
5514 int ld_moved, cur_ld_moved, active_balance = 0;
5515 struct sched_domain *sd_parent = sd->parent;
5516 struct sched_group *group;
5518 unsigned long flags;
5519 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5521 struct lb_env env = {
5523 .dst_cpu = this_cpu,
5525 .dst_grpmask = sched_group_cpus(sd->groups),
5527 .loop_break = sched_nr_migrate_break,
5532 * For NEWLY_IDLE load_balancing, we don't need to consider
5533 * other cpus in our group
5535 if (idle == CPU_NEWLY_IDLE)
5536 env.dst_grpmask = NULL;
5538 cpumask_copy(cpus, cpu_active_mask);
5540 schedstat_inc(sd, lb_count[idle]);
5543 if (!should_we_balance(&env)) {
5544 *continue_balancing = 0;
5548 group = find_busiest_group(&env);
5550 schedstat_inc(sd, lb_nobusyg[idle]);
5554 busiest = find_busiest_queue(&env, group);
5556 schedstat_inc(sd, lb_nobusyq[idle]);
5560 BUG_ON(busiest == env.dst_rq);
5562 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5565 if (busiest->nr_running > 1) {
5567 * Attempt to move tasks. If find_busiest_group has found
5568 * an imbalance but busiest->nr_running <= 1, the group is
5569 * still unbalanced. ld_moved simply stays zero, so it is
5570 * correctly treated as an imbalance.
5572 env.flags |= LBF_ALL_PINNED;
5573 env.src_cpu = busiest->cpu;
5574 env.src_rq = busiest;
5575 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5578 local_irq_save(flags);
5579 double_rq_lock(env.dst_rq, busiest);
5582 * cur_ld_moved - load moved in current iteration
5583 * ld_moved - cumulative load moved across iterations
5585 cur_ld_moved = move_tasks(&env);
5586 ld_moved += cur_ld_moved;
5587 double_rq_unlock(env.dst_rq, busiest);
5588 local_irq_restore(flags);
5591 * some other cpu did the load balance for us.
5593 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5594 resched_cpu(env.dst_cpu);
5596 if (env.flags & LBF_NEED_BREAK) {
5597 env.flags &= ~LBF_NEED_BREAK;
5602 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5603 * us and move them to an alternate dst_cpu in our sched_group
5604 * where they can run. The upper limit on how many times we
5605 * iterate on same src_cpu is dependent on number of cpus in our
5608 * This changes load balance semantics a bit on who can move
5609 * load to a given_cpu. In addition to the given_cpu itself
5610 * (or a ilb_cpu acting on its behalf where given_cpu is
5611 * nohz-idle), we now have balance_cpu in a position to move
5612 * load to given_cpu. In rare situations, this may cause
5613 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5614 * _independently_ and at _same_ time to move some load to
5615 * given_cpu) causing exceess load to be moved to given_cpu.
5616 * This however should not happen so much in practice and
5617 * moreover subsequent load balance cycles should correct the
5618 * excess load moved.
5620 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5622 /* Prevent to re-select dst_cpu via env's cpus */
5623 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5625 env.dst_rq = cpu_rq(env.new_dst_cpu);
5626 env.dst_cpu = env.new_dst_cpu;
5627 env.flags &= ~LBF_DST_PINNED;
5629 env.loop_break = sched_nr_migrate_break;
5632 * Go back to "more_balance" rather than "redo" since we
5633 * need to continue with same src_cpu.
5639 * We failed to reach balance because of affinity.
5642 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5644 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5645 *group_imbalance = 1;
5646 } else if (*group_imbalance)
5647 *group_imbalance = 0;
5650 /* All tasks on this runqueue were pinned by CPU affinity */
5651 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5652 cpumask_clear_cpu(cpu_of(busiest), cpus);
5653 if (!cpumask_empty(cpus)) {
5655 env.loop_break = sched_nr_migrate_break;
5663 schedstat_inc(sd, lb_failed[idle]);
5665 * Increment the failure counter only on periodic balance.
5666 * We do not want newidle balance, which can be very
5667 * frequent, pollute the failure counter causing
5668 * excessive cache_hot migrations and active balances.
5670 if (idle != CPU_NEWLY_IDLE)
5671 sd->nr_balance_failed++;
5673 if (need_active_balance(&env)) {
5674 raw_spin_lock_irqsave(&busiest->lock, flags);
5676 /* don't kick the active_load_balance_cpu_stop,
5677 * if the curr task on busiest cpu can't be
5680 if (!cpumask_test_cpu(this_cpu,
5681 tsk_cpus_allowed(busiest->curr))) {
5682 raw_spin_unlock_irqrestore(&busiest->lock,
5684 env.flags |= LBF_ALL_PINNED;
5685 goto out_one_pinned;
5689 * ->active_balance synchronizes accesses to
5690 * ->active_balance_work. Once set, it's cleared
5691 * only after active load balance is finished.
5693 if (!busiest->active_balance) {
5694 busiest->active_balance = 1;
5695 busiest->push_cpu = this_cpu;
5698 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5700 if (active_balance) {
5701 stop_one_cpu_nowait(cpu_of(busiest),
5702 active_load_balance_cpu_stop, busiest,
5703 &busiest->active_balance_work);
5707 * We've kicked active balancing, reset the failure
5710 sd->nr_balance_failed = sd->cache_nice_tries+1;
5713 sd->nr_balance_failed = 0;
5715 if (likely(!active_balance)) {
5716 /* We were unbalanced, so reset the balancing interval */
5717 sd->balance_interval = sd->min_interval;
5720 * If we've begun active balancing, start to back off. This
5721 * case may not be covered by the all_pinned logic if there
5722 * is only 1 task on the busy runqueue (because we don't call
5725 if (sd->balance_interval < sd->max_interval)
5726 sd->balance_interval *= 2;
5732 schedstat_inc(sd, lb_balanced[idle]);
5734 sd->nr_balance_failed = 0;
5737 /* tune up the balancing interval */
5738 if (((env.flags & LBF_ALL_PINNED) &&
5739 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5740 (sd->balance_interval < sd->max_interval))
5741 sd->balance_interval *= 2;
5749 * idle_balance is called by schedule() if this_cpu is about to become
5750 * idle. Attempts to pull tasks from other CPUs.
5752 void idle_balance(int this_cpu, struct rq *this_rq)
5754 struct sched_domain *sd;
5755 int pulled_task = 0;
5756 unsigned long next_balance = jiffies + HZ;
5759 this_rq->idle_stamp = rq_clock(this_rq);
5761 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5765 * Drop the rq->lock, but keep IRQ/preempt disabled.
5767 raw_spin_unlock(&this_rq->lock);
5769 update_blocked_averages(this_cpu);
5771 for_each_domain(this_cpu, sd) {
5772 unsigned long interval;
5773 int continue_balancing = 1;
5774 u64 t0, domain_cost;
5776 if (!(sd->flags & SD_LOAD_BALANCE))
5779 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5782 if (sd->flags & SD_BALANCE_NEWIDLE) {
5783 t0 = sched_clock_cpu(this_cpu);
5785 /* If we've pulled tasks over stop searching: */
5786 pulled_task = load_balance(this_cpu, this_rq,
5788 &continue_balancing);
5790 domain_cost = sched_clock_cpu(this_cpu) - t0;
5791 if (domain_cost > sd->max_newidle_lb_cost)
5792 sd->max_newidle_lb_cost = domain_cost;
5794 curr_cost += domain_cost;
5797 interval = msecs_to_jiffies(sd->balance_interval);
5798 if (time_after(next_balance, sd->last_balance + interval))
5799 next_balance = sd->last_balance + interval;
5801 this_rq->idle_stamp = 0;
5807 raw_spin_lock(&this_rq->lock);
5809 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5811 * We are going idle. next_balance may be set based on
5812 * a busy processor. So reset next_balance.
5814 this_rq->next_balance = next_balance;
5817 if (curr_cost > this_rq->max_idle_balance_cost)
5818 this_rq->max_idle_balance_cost = curr_cost;
5822 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5823 * running tasks off the busiest CPU onto idle CPUs. It requires at
5824 * least 1 task to be running on each physical CPU where possible, and
5825 * avoids physical / logical imbalances.
5827 static int active_load_balance_cpu_stop(void *data)
5829 struct rq *busiest_rq = data;
5830 int busiest_cpu = cpu_of(busiest_rq);
5831 int target_cpu = busiest_rq->push_cpu;
5832 struct rq *target_rq = cpu_rq(target_cpu);
5833 struct sched_domain *sd;
5835 raw_spin_lock_irq(&busiest_rq->lock);
5837 /* make sure the requested cpu hasn't gone down in the meantime */
5838 if (unlikely(busiest_cpu != smp_processor_id() ||
5839 !busiest_rq->active_balance))
5842 /* Is there any task to move? */
5843 if (busiest_rq->nr_running <= 1)
5847 * This condition is "impossible", if it occurs
5848 * we need to fix it. Originally reported by
5849 * Bjorn Helgaas on a 128-cpu setup.
5851 BUG_ON(busiest_rq == target_rq);
5853 /* move a task from busiest_rq to target_rq */
5854 double_lock_balance(busiest_rq, target_rq);
5856 /* Search for an sd spanning us and the target CPU. */
5858 for_each_domain(target_cpu, sd) {
5859 if ((sd->flags & SD_LOAD_BALANCE) &&
5860 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5865 struct lb_env env = {
5867 .dst_cpu = target_cpu,
5868 .dst_rq = target_rq,
5869 .src_cpu = busiest_rq->cpu,
5870 .src_rq = busiest_rq,
5874 schedstat_inc(sd, alb_count);
5876 if (move_one_task(&env))
5877 schedstat_inc(sd, alb_pushed);
5879 schedstat_inc(sd, alb_failed);
5882 double_unlock_balance(busiest_rq, target_rq);
5884 busiest_rq->active_balance = 0;
5885 raw_spin_unlock_irq(&busiest_rq->lock);
5889 #ifdef CONFIG_NO_HZ_COMMON
5891 * idle load balancing details
5892 * - When one of the busy CPUs notice that there may be an idle rebalancing
5893 * needed, they will kick the idle load balancer, which then does idle
5894 * load balancing for all the idle CPUs.
5897 cpumask_var_t idle_cpus_mask;
5899 unsigned long next_balance; /* in jiffy units */
5900 } nohz ____cacheline_aligned;
5902 static inline int find_new_ilb(int call_cpu)
5904 int ilb = cpumask_first(nohz.idle_cpus_mask);
5906 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5913 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5914 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5915 * CPU (if there is one).
5917 static void nohz_balancer_kick(int cpu)
5921 nohz.next_balance++;
5923 ilb_cpu = find_new_ilb(cpu);
5925 if (ilb_cpu >= nr_cpu_ids)
5928 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5931 * Use smp_send_reschedule() instead of resched_cpu().
5932 * This way we generate a sched IPI on the target cpu which
5933 * is idle. And the softirq performing nohz idle load balance
5934 * will be run before returning from the IPI.
5936 smp_send_reschedule(ilb_cpu);
5940 static inline void nohz_balance_exit_idle(int cpu)
5942 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5943 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5944 atomic_dec(&nohz.nr_cpus);
5945 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5949 static inline void set_cpu_sd_state_busy(void)
5951 struct sched_domain *sd;
5954 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5956 if (!sd || !sd->nohz_idle)
5960 for (; sd; sd = sd->parent)
5961 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5966 void set_cpu_sd_state_idle(void)
5968 struct sched_domain *sd;
5971 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5973 if (!sd || sd->nohz_idle)
5977 for (; sd; sd = sd->parent)
5978 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5984 * This routine will record that the cpu is going idle with tick stopped.
5985 * This info will be used in performing idle load balancing in the future.
5987 void nohz_balance_enter_idle(int cpu)
5990 * If this cpu is going down, then nothing needs to be done.
5992 if (!cpu_active(cpu))
5995 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5998 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5999 atomic_inc(&nohz.nr_cpus);
6000 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6003 static int sched_ilb_notifier(struct notifier_block *nfb,
6004 unsigned long action, void *hcpu)
6006 switch (action & ~CPU_TASKS_FROZEN) {
6008 nohz_balance_exit_idle(smp_processor_id());
6016 static DEFINE_SPINLOCK(balancing);
6019 * Scale the max load_balance interval with the number of CPUs in the system.
6020 * This trades load-balance latency on larger machines for less cross talk.
6022 void update_max_interval(void)
6024 max_load_balance_interval = HZ*num_online_cpus()/10;
6028 * It checks each scheduling domain to see if it is due to be balanced,
6029 * and initiates a balancing operation if so.
6031 * Balancing parameters are set up in init_sched_domains.
6033 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6035 int continue_balancing = 1;
6036 struct rq *rq = cpu_rq(cpu);
6037 unsigned long interval;
6038 struct sched_domain *sd;
6039 /* Earliest time when we have to do rebalance again */
6040 unsigned long next_balance = jiffies + 60*HZ;
6041 int update_next_balance = 0;
6042 int need_serialize, need_decay = 0;
6045 update_blocked_averages(cpu);
6048 for_each_domain(cpu, sd) {
6050 * Decay the newidle max times here because this is a regular
6051 * visit to all the domains. Decay ~1% per second.
6053 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6054 sd->max_newidle_lb_cost =
6055 (sd->max_newidle_lb_cost * 253) / 256;
6056 sd->next_decay_max_lb_cost = jiffies + HZ;
6059 max_cost += sd->max_newidle_lb_cost;
6061 if (!(sd->flags & SD_LOAD_BALANCE))
6065 * Stop the load balance at this level. There is another
6066 * CPU in our sched group which is doing load balancing more
6069 if (!continue_balancing) {
6075 interval = sd->balance_interval;
6076 if (idle != CPU_IDLE)
6077 interval *= sd->busy_factor;
6079 /* scale ms to jiffies */
6080 interval = msecs_to_jiffies(interval);
6081 interval = clamp(interval, 1UL, max_load_balance_interval);
6083 need_serialize = sd->flags & SD_SERIALIZE;
6085 if (need_serialize) {
6086 if (!spin_trylock(&balancing))
6090 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6091 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6093 * The LBF_DST_PINNED logic could have changed
6094 * env->dst_cpu, so we can't know our idle
6095 * state even if we migrated tasks. Update it.
6097 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6099 sd->last_balance = jiffies;
6102 spin_unlock(&balancing);
6104 if (time_after(next_balance, sd->last_balance + interval)) {
6105 next_balance = sd->last_balance + interval;
6106 update_next_balance = 1;
6111 * Ensure the rq-wide value also decays but keep it at a
6112 * reasonable floor to avoid funnies with rq->avg_idle.
6114 rq->max_idle_balance_cost =
6115 max((u64)sysctl_sched_migration_cost, max_cost);
6120 * next_balance will be updated only when there is a need.
6121 * When the cpu is attached to null domain for ex, it will not be
6124 if (likely(update_next_balance))
6125 rq->next_balance = next_balance;
6128 #ifdef CONFIG_NO_HZ_COMMON
6130 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6131 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6133 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6135 struct rq *this_rq = cpu_rq(this_cpu);
6139 if (idle != CPU_IDLE ||
6140 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6143 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6144 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6148 * If this cpu gets work to do, stop the load balancing
6149 * work being done for other cpus. Next load
6150 * balancing owner will pick it up.
6155 rq = cpu_rq(balance_cpu);
6157 raw_spin_lock_irq(&rq->lock);
6158 update_rq_clock(rq);
6159 update_idle_cpu_load(rq);
6160 raw_spin_unlock_irq(&rq->lock);
6162 rebalance_domains(balance_cpu, CPU_IDLE);
6164 if (time_after(this_rq->next_balance, rq->next_balance))
6165 this_rq->next_balance = rq->next_balance;
6167 nohz.next_balance = this_rq->next_balance;
6169 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6173 * Current heuristic for kicking the idle load balancer in the presence
6174 * of an idle cpu is the system.
6175 * - This rq has more than one task.
6176 * - At any scheduler domain level, this cpu's scheduler group has multiple
6177 * busy cpu's exceeding the group's power.
6178 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6179 * domain span are idle.
6181 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6183 unsigned long now = jiffies;
6184 struct sched_domain *sd;
6186 if (unlikely(idle_cpu(cpu)))
6190 * We may be recently in ticked or tickless idle mode. At the first
6191 * busy tick after returning from idle, we will update the busy stats.
6193 set_cpu_sd_state_busy();
6194 nohz_balance_exit_idle(cpu);
6197 * None are in tickless mode and hence no need for NOHZ idle load
6200 if (likely(!atomic_read(&nohz.nr_cpus)))
6203 if (time_before(now, nohz.next_balance))
6206 if (rq->nr_running >= 2)
6210 for_each_domain(cpu, sd) {
6211 struct sched_group *sg = sd->groups;
6212 struct sched_group_power *sgp = sg->sgp;
6213 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6215 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6216 goto need_kick_unlock;
6218 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6219 && (cpumask_first_and(nohz.idle_cpus_mask,
6220 sched_domain_span(sd)) < cpu))
6221 goto need_kick_unlock;
6223 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6235 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6239 * run_rebalance_domains is triggered when needed from the scheduler tick.
6240 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6242 static void run_rebalance_domains(struct softirq_action *h)
6244 int this_cpu = smp_processor_id();
6245 struct rq *this_rq = cpu_rq(this_cpu);
6246 enum cpu_idle_type idle = this_rq->idle_balance ?
6247 CPU_IDLE : CPU_NOT_IDLE;
6249 rebalance_domains(this_cpu, idle);
6252 * If this cpu has a pending nohz_balance_kick, then do the
6253 * balancing on behalf of the other idle cpus whose ticks are
6256 nohz_idle_balance(this_cpu, idle);
6259 static inline int on_null_domain(int cpu)
6261 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6265 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6267 void trigger_load_balance(struct rq *rq, int cpu)
6269 /* Don't need to rebalance while attached to NULL domain */
6270 if (time_after_eq(jiffies, rq->next_balance) &&
6271 likely(!on_null_domain(cpu)))
6272 raise_softirq(SCHED_SOFTIRQ);
6273 #ifdef CONFIG_NO_HZ_COMMON
6274 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6275 nohz_balancer_kick(cpu);
6279 static void rq_online_fair(struct rq *rq)
6284 static void rq_offline_fair(struct rq *rq)
6288 /* Ensure any throttled groups are reachable by pick_next_task */
6289 unthrottle_offline_cfs_rqs(rq);
6292 #endif /* CONFIG_SMP */
6295 * scheduler tick hitting a task of our scheduling class:
6297 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6299 struct cfs_rq *cfs_rq;
6300 struct sched_entity *se = &curr->se;
6302 for_each_sched_entity(se) {
6303 cfs_rq = cfs_rq_of(se);
6304 entity_tick(cfs_rq, se, queued);
6307 if (numabalancing_enabled)
6308 task_tick_numa(rq, curr);
6310 update_rq_runnable_avg(rq, 1);
6314 * called on fork with the child task as argument from the parent's context
6315 * - child not yet on the tasklist
6316 * - preemption disabled
6318 static void task_fork_fair(struct task_struct *p)
6320 struct cfs_rq *cfs_rq;
6321 struct sched_entity *se = &p->se, *curr;
6322 int this_cpu = smp_processor_id();
6323 struct rq *rq = this_rq();
6324 unsigned long flags;
6326 raw_spin_lock_irqsave(&rq->lock, flags);
6328 update_rq_clock(rq);
6330 cfs_rq = task_cfs_rq(current);
6331 curr = cfs_rq->curr;
6334 * Not only the cpu but also the task_group of the parent might have
6335 * been changed after parent->se.parent,cfs_rq were copied to
6336 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6337 * of child point to valid ones.
6340 __set_task_cpu(p, this_cpu);
6343 update_curr(cfs_rq);
6346 se->vruntime = curr->vruntime;
6347 place_entity(cfs_rq, se, 1);
6349 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6351 * Upon rescheduling, sched_class::put_prev_task() will place
6352 * 'current' within the tree based on its new key value.
6354 swap(curr->vruntime, se->vruntime);
6355 resched_task(rq->curr);
6358 se->vruntime -= cfs_rq->min_vruntime;
6360 raw_spin_unlock_irqrestore(&rq->lock, flags);
6364 * Priority of the task has changed. Check to see if we preempt
6368 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6374 * Reschedule if we are currently running on this runqueue and
6375 * our priority decreased, or if we are not currently running on
6376 * this runqueue and our priority is higher than the current's
6378 if (rq->curr == p) {
6379 if (p->prio > oldprio)
6380 resched_task(rq->curr);
6382 check_preempt_curr(rq, p, 0);
6385 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6387 struct sched_entity *se = &p->se;
6388 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6391 * Ensure the task's vruntime is normalized, so that when its
6392 * switched back to the fair class the enqueue_entity(.flags=0) will
6393 * do the right thing.
6395 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6396 * have normalized the vruntime, if it was !on_rq, then only when
6397 * the task is sleeping will it still have non-normalized vruntime.
6399 if (!se->on_rq && p->state != TASK_RUNNING) {
6401 * Fix up our vruntime so that the current sleep doesn't
6402 * cause 'unlimited' sleep bonus.
6404 place_entity(cfs_rq, se, 0);
6405 se->vruntime -= cfs_rq->min_vruntime;
6410 * Remove our load from contribution when we leave sched_fair
6411 * and ensure we don't carry in an old decay_count if we
6414 if (se->avg.decay_count) {
6415 __synchronize_entity_decay(se);
6416 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6422 * We switched to the sched_fair class.
6424 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6430 * We were most likely switched from sched_rt, so
6431 * kick off the schedule if running, otherwise just see
6432 * if we can still preempt the current task.
6435 resched_task(rq->curr);
6437 check_preempt_curr(rq, p, 0);
6440 /* Account for a task changing its policy or group.
6442 * This routine is mostly called to set cfs_rq->curr field when a task
6443 * migrates between groups/classes.
6445 static void set_curr_task_fair(struct rq *rq)
6447 struct sched_entity *se = &rq->curr->se;
6449 for_each_sched_entity(se) {
6450 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6452 set_next_entity(cfs_rq, se);
6453 /* ensure bandwidth has been allocated on our new cfs_rq */
6454 account_cfs_rq_runtime(cfs_rq, 0);
6458 void init_cfs_rq(struct cfs_rq *cfs_rq)
6460 cfs_rq->tasks_timeline = RB_ROOT;
6461 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6462 #ifndef CONFIG_64BIT
6463 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6466 atomic64_set(&cfs_rq->decay_counter, 1);
6467 atomic_long_set(&cfs_rq->removed_load, 0);
6471 #ifdef CONFIG_FAIR_GROUP_SCHED
6472 static void task_move_group_fair(struct task_struct *p, int on_rq)
6474 struct cfs_rq *cfs_rq;
6476 * If the task was not on the rq at the time of this cgroup movement
6477 * it must have been asleep, sleeping tasks keep their ->vruntime
6478 * absolute on their old rq until wakeup (needed for the fair sleeper
6479 * bonus in place_entity()).
6481 * If it was on the rq, we've just 'preempted' it, which does convert
6482 * ->vruntime to a relative base.
6484 * Make sure both cases convert their relative position when migrating
6485 * to another cgroup's rq. This does somewhat interfere with the
6486 * fair sleeper stuff for the first placement, but who cares.
6489 * When !on_rq, vruntime of the task has usually NOT been normalized.
6490 * But there are some cases where it has already been normalized:
6492 * - Moving a forked child which is waiting for being woken up by
6493 * wake_up_new_task().
6494 * - Moving a task which has been woken up by try_to_wake_up() and
6495 * waiting for actually being woken up by sched_ttwu_pending().
6497 * To prevent boost or penalty in the new cfs_rq caused by delta
6498 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6500 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6504 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6505 set_task_rq(p, task_cpu(p));
6507 cfs_rq = cfs_rq_of(&p->se);
6508 p->se.vruntime += cfs_rq->min_vruntime;
6511 * migrate_task_rq_fair() will have removed our previous
6512 * contribution, but we must synchronize for ongoing future
6515 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6516 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6521 void free_fair_sched_group(struct task_group *tg)
6525 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6527 for_each_possible_cpu(i) {
6529 kfree(tg->cfs_rq[i]);
6538 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6540 struct cfs_rq *cfs_rq;
6541 struct sched_entity *se;
6544 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6547 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6551 tg->shares = NICE_0_LOAD;
6553 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6555 for_each_possible_cpu(i) {
6556 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6557 GFP_KERNEL, cpu_to_node(i));
6561 se = kzalloc_node(sizeof(struct sched_entity),
6562 GFP_KERNEL, cpu_to_node(i));
6566 init_cfs_rq(cfs_rq);
6567 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6578 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6580 struct rq *rq = cpu_rq(cpu);
6581 unsigned long flags;
6584 * Only empty task groups can be destroyed; so we can speculatively
6585 * check on_list without danger of it being re-added.
6587 if (!tg->cfs_rq[cpu]->on_list)
6590 raw_spin_lock_irqsave(&rq->lock, flags);
6591 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6592 raw_spin_unlock_irqrestore(&rq->lock, flags);
6595 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6596 struct sched_entity *se, int cpu,
6597 struct sched_entity *parent)
6599 struct rq *rq = cpu_rq(cpu);
6603 init_cfs_rq_runtime(cfs_rq);
6605 tg->cfs_rq[cpu] = cfs_rq;
6608 /* se could be NULL for root_task_group */
6613 se->cfs_rq = &rq->cfs;
6615 se->cfs_rq = parent->my_q;
6618 update_load_set(&se->load, 0);
6619 se->parent = parent;
6622 static DEFINE_MUTEX(shares_mutex);
6624 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6627 unsigned long flags;
6630 * We can't change the weight of the root cgroup.
6635 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6637 mutex_lock(&shares_mutex);
6638 if (tg->shares == shares)
6641 tg->shares = shares;
6642 for_each_possible_cpu(i) {
6643 struct rq *rq = cpu_rq(i);
6644 struct sched_entity *se;
6647 /* Propagate contribution to hierarchy */
6648 raw_spin_lock_irqsave(&rq->lock, flags);
6650 /* Possible calls to update_curr() need rq clock */
6651 update_rq_clock(rq);
6652 for_each_sched_entity(se)
6653 update_cfs_shares(group_cfs_rq(se));
6654 raw_spin_unlock_irqrestore(&rq->lock, flags);
6658 mutex_unlock(&shares_mutex);
6661 #else /* CONFIG_FAIR_GROUP_SCHED */
6663 void free_fair_sched_group(struct task_group *tg) { }
6665 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6670 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6672 #endif /* CONFIG_FAIR_GROUP_SCHED */
6675 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6677 struct sched_entity *se = &task->se;
6678 unsigned int rr_interval = 0;
6681 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6684 if (rq->cfs.load.weight)
6685 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6691 * All the scheduling class methods:
6693 const struct sched_class fair_sched_class = {
6694 .next = &idle_sched_class,
6695 .enqueue_task = enqueue_task_fair,
6696 .dequeue_task = dequeue_task_fair,
6697 .yield_task = yield_task_fair,
6698 .yield_to_task = yield_to_task_fair,
6700 .check_preempt_curr = check_preempt_wakeup,
6702 .pick_next_task = pick_next_task_fair,
6703 .put_prev_task = put_prev_task_fair,
6706 .select_task_rq = select_task_rq_fair,
6707 .migrate_task_rq = migrate_task_rq_fair,
6709 .rq_online = rq_online_fair,
6710 .rq_offline = rq_offline_fair,
6712 .task_waking = task_waking_fair,
6715 .set_curr_task = set_curr_task_fair,
6716 .task_tick = task_tick_fair,
6717 .task_fork = task_fork_fair,
6719 .prio_changed = prio_changed_fair,
6720 .switched_from = switched_from_fair,
6721 .switched_to = switched_to_fair,
6723 .get_rr_interval = get_rr_interval_fair,
6725 #ifdef CONFIG_FAIR_GROUP_SCHED
6726 .task_move_group = task_move_group_fair,
6730 #ifdef CONFIG_SCHED_DEBUG
6731 void print_cfs_stats(struct seq_file *m, int cpu)
6733 struct cfs_rq *cfs_rq;
6736 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6737 print_cfs_rq(m, cpu, cfs_rq);
6742 __init void init_sched_fair_class(void)
6745 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6747 #ifdef CONFIG_NO_HZ_COMMON
6748 nohz.next_balance = jiffies;
6749 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6750 cpu_notifier(sched_ilb_notifier, 0);