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);
907 find_idlest_cpu_node(int this_cpu, int nid)
909 unsigned long load, min_load = ULONG_MAX;
910 int i, idlest_cpu = this_cpu;
912 BUG_ON(cpu_to_node(this_cpu) == nid);
915 for_each_cpu(i, cpumask_of_node(nid)) {
916 load = weighted_cpuload(i);
918 if (load < min_load) {
928 static void task_numa_placement(struct task_struct *p)
930 int seq, nid, max_nid = -1;
931 unsigned long max_faults = 0;
933 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
934 if (p->numa_scan_seq == seq)
936 p->numa_scan_seq = seq;
937 p->numa_migrate_seq++;
938 p->numa_scan_period_max = task_scan_max(p);
940 /* Find the node with the highest number of faults */
941 for_each_online_node(nid) {
942 unsigned long faults;
945 for (priv = 0; priv < 2; priv++) {
946 i = task_faults_idx(nid, priv);
948 /* Decay existing window, copy faults since last scan */
949 p->numa_faults[i] >>= 1;
950 p->numa_faults[i] += p->numa_faults_buffer[i];
951 p->numa_faults_buffer[i] = 0;
954 /* Find maximum private faults */
955 faults = p->numa_faults[task_faults_idx(nid, 1)];
956 if (faults > max_faults) {
963 * Record the preferred node as the node with the most faults,
964 * requeue the task to be running on the idlest CPU on the
965 * preferred node and reset the scanning rate to recheck
966 * the working set placement.
968 if (max_faults && max_nid != p->numa_preferred_nid) {
972 * If the task is not on the preferred node then find the most
973 * idle CPU to migrate to.
975 preferred_cpu = task_cpu(p);
976 if (cpu_to_node(preferred_cpu) != max_nid) {
977 preferred_cpu = find_idlest_cpu_node(preferred_cpu,
981 /* Update the preferred nid and migrate task if possible */
982 p->numa_preferred_nid = max_nid;
983 p->numa_migrate_seq = 1;
984 migrate_task_to(p, preferred_cpu);
989 * Got a PROT_NONE fault for a page on @node.
991 void task_numa_fault(int last_nidpid, int node, int pages, bool migrated)
993 struct task_struct *p = current;
996 if (!numabalancing_enabled)
999 /* for example, ksmd faulting in a user's mm */
1004 * First accesses are treated as private, otherwise consider accesses
1005 * to be private if the accessing pid has not changed
1007 if (!nidpid_pid_unset(last_nidpid))
1008 priv = ((p->pid & LAST__PID_MASK) == nidpid_to_pid(last_nidpid));
1012 /* Allocate buffer to track faults on a per-node basis */
1013 if (unlikely(!p->numa_faults)) {
1014 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1016 /* numa_faults and numa_faults_buffer share the allocation */
1017 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1018 if (!p->numa_faults)
1021 BUG_ON(p->numa_faults_buffer);
1022 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1026 * If pages are properly placed (did not migrate) then scan slower.
1027 * This is reset periodically in case of phase changes
1030 /* Initialise if necessary */
1031 if (!p->numa_scan_period_max)
1032 p->numa_scan_period_max = task_scan_max(p);
1034 p->numa_scan_period = min(p->numa_scan_period_max,
1035 p->numa_scan_period + 10);
1038 task_numa_placement(p);
1040 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1043 static void reset_ptenuma_scan(struct task_struct *p)
1045 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1046 p->mm->numa_scan_offset = 0;
1050 * The expensive part of numa migration is done from task_work context.
1051 * Triggered from task_tick_numa().
1053 void task_numa_work(struct callback_head *work)
1055 unsigned long migrate, next_scan, now = jiffies;
1056 struct task_struct *p = current;
1057 struct mm_struct *mm = p->mm;
1058 struct vm_area_struct *vma;
1059 unsigned long start, end;
1060 unsigned long nr_pte_updates = 0;
1063 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1065 work->next = work; /* protect against double add */
1067 * Who cares about NUMA placement when they're dying.
1069 * NOTE: make sure not to dereference p->mm before this check,
1070 * exit_task_work() happens _after_ exit_mm() so we could be called
1071 * without p->mm even though we still had it when we enqueued this
1074 if (p->flags & PF_EXITING)
1077 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1078 mm->numa_next_scan = now +
1079 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1080 mm->numa_next_reset = now +
1081 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1085 * Reset the scan period if enough time has gone by. Objective is that
1086 * scanning will be reduced if pages are properly placed. As tasks
1087 * can enter different phases this needs to be re-examined. Lacking
1088 * proper tracking of reference behaviour, this blunt hammer is used.
1090 migrate = mm->numa_next_reset;
1091 if (time_after(now, migrate)) {
1092 p->numa_scan_period = task_scan_min(p);
1093 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1094 xchg(&mm->numa_next_reset, next_scan);
1098 * Enforce maximal scan/migration frequency..
1100 migrate = mm->numa_next_scan;
1101 if (time_before(now, migrate))
1104 if (p->numa_scan_period == 0) {
1105 p->numa_scan_period_max = task_scan_max(p);
1106 p->numa_scan_period = task_scan_min(p);
1109 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1110 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1114 * Delay this task enough that another task of this mm will likely win
1115 * the next time around.
1117 p->node_stamp += 2 * TICK_NSEC;
1119 start = mm->numa_scan_offset;
1120 pages = sysctl_numa_balancing_scan_size;
1121 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1125 down_read(&mm->mmap_sem);
1126 vma = find_vma(mm, start);
1128 reset_ptenuma_scan(p);
1132 for (; vma; vma = vma->vm_next) {
1133 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1137 start = max(start, vma->vm_start);
1138 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1139 end = min(end, vma->vm_end);
1140 nr_pte_updates += change_prot_numa(vma, start, end);
1143 * Scan sysctl_numa_balancing_scan_size but ensure that
1144 * at least one PTE is updated so that unused virtual
1145 * address space is quickly skipped.
1148 pages -= (end - start) >> PAGE_SHIFT;
1153 } while (end != vma->vm_end);
1158 * If the whole process was scanned without updates then no NUMA
1159 * hinting faults are being recorded and scan rate should be lower.
1161 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1162 p->numa_scan_period = min(p->numa_scan_period_max,
1163 p->numa_scan_period << 1);
1165 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1166 mm->numa_next_scan = next_scan;
1170 * It is possible to reach the end of the VMA list but the last few
1171 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1172 * would find the !migratable VMA on the next scan but not reset the
1173 * scanner to the start so check it now.
1176 mm->numa_scan_offset = start;
1178 reset_ptenuma_scan(p);
1179 up_read(&mm->mmap_sem);
1183 * Drive the periodic memory faults..
1185 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1187 struct callback_head *work = &curr->numa_work;
1191 * We don't care about NUMA placement if we don't have memory.
1193 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1197 * Using runtime rather than walltime has the dual advantage that
1198 * we (mostly) drive the selection from busy threads and that the
1199 * task needs to have done some actual work before we bother with
1202 now = curr->se.sum_exec_runtime;
1203 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1205 if (now - curr->node_stamp > period) {
1206 if (!curr->node_stamp)
1207 curr->numa_scan_period = task_scan_min(curr);
1208 curr->node_stamp += period;
1210 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1211 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1212 task_work_add(curr, work, true);
1217 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1220 #endif /* CONFIG_NUMA_BALANCING */
1223 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1225 update_load_add(&cfs_rq->load, se->load.weight);
1226 if (!parent_entity(se))
1227 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1229 if (entity_is_task(se))
1230 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1232 cfs_rq->nr_running++;
1236 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1238 update_load_sub(&cfs_rq->load, se->load.weight);
1239 if (!parent_entity(se))
1240 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1241 if (entity_is_task(se))
1242 list_del_init(&se->group_node);
1243 cfs_rq->nr_running--;
1246 #ifdef CONFIG_FAIR_GROUP_SCHED
1248 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1253 * Use this CPU's actual weight instead of the last load_contribution
1254 * to gain a more accurate current total weight. See
1255 * update_cfs_rq_load_contribution().
1257 tg_weight = atomic_long_read(&tg->load_avg);
1258 tg_weight -= cfs_rq->tg_load_contrib;
1259 tg_weight += cfs_rq->load.weight;
1264 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1266 long tg_weight, load, shares;
1268 tg_weight = calc_tg_weight(tg, cfs_rq);
1269 load = cfs_rq->load.weight;
1271 shares = (tg->shares * load);
1273 shares /= tg_weight;
1275 if (shares < MIN_SHARES)
1276 shares = MIN_SHARES;
1277 if (shares > tg->shares)
1278 shares = tg->shares;
1282 # else /* CONFIG_SMP */
1283 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1287 # endif /* CONFIG_SMP */
1288 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1289 unsigned long weight)
1292 /* commit outstanding execution time */
1293 if (cfs_rq->curr == se)
1294 update_curr(cfs_rq);
1295 account_entity_dequeue(cfs_rq, se);
1298 update_load_set(&se->load, weight);
1301 account_entity_enqueue(cfs_rq, se);
1304 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1306 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1308 struct task_group *tg;
1309 struct sched_entity *se;
1313 se = tg->se[cpu_of(rq_of(cfs_rq))];
1314 if (!se || throttled_hierarchy(cfs_rq))
1317 if (likely(se->load.weight == tg->shares))
1320 shares = calc_cfs_shares(cfs_rq, tg);
1322 reweight_entity(cfs_rq_of(se), se, shares);
1324 #else /* CONFIG_FAIR_GROUP_SCHED */
1325 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1328 #endif /* CONFIG_FAIR_GROUP_SCHED */
1332 * We choose a half-life close to 1 scheduling period.
1333 * Note: The tables below are dependent on this value.
1335 #define LOAD_AVG_PERIOD 32
1336 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1337 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1339 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1340 static const u32 runnable_avg_yN_inv[] = {
1341 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1342 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1343 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1344 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1345 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1346 0x85aac367, 0x82cd8698,
1350 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1351 * over-estimates when re-combining.
1353 static const u32 runnable_avg_yN_sum[] = {
1354 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1355 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1356 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1361 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1363 static __always_inline u64 decay_load(u64 val, u64 n)
1365 unsigned int local_n;
1369 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1372 /* after bounds checking we can collapse to 32-bit */
1376 * As y^PERIOD = 1/2, we can combine
1377 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1378 * With a look-up table which covers k^n (n<PERIOD)
1380 * To achieve constant time decay_load.
1382 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1383 val >>= local_n / LOAD_AVG_PERIOD;
1384 local_n %= LOAD_AVG_PERIOD;
1387 val *= runnable_avg_yN_inv[local_n];
1388 /* We don't use SRR here since we always want to round down. */
1393 * For updates fully spanning n periods, the contribution to runnable
1394 * average will be: \Sum 1024*y^n
1396 * We can compute this reasonably efficiently by combining:
1397 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1399 static u32 __compute_runnable_contrib(u64 n)
1403 if (likely(n <= LOAD_AVG_PERIOD))
1404 return runnable_avg_yN_sum[n];
1405 else if (unlikely(n >= LOAD_AVG_MAX_N))
1406 return LOAD_AVG_MAX;
1408 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1410 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1411 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1413 n -= LOAD_AVG_PERIOD;
1414 } while (n > LOAD_AVG_PERIOD);
1416 contrib = decay_load(contrib, n);
1417 return contrib + runnable_avg_yN_sum[n];
1421 * We can represent the historical contribution to runnable average as the
1422 * coefficients of a geometric series. To do this we sub-divide our runnable
1423 * history into segments of approximately 1ms (1024us); label the segment that
1424 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1426 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1428 * (now) (~1ms ago) (~2ms ago)
1430 * Let u_i denote the fraction of p_i that the entity was runnable.
1432 * We then designate the fractions u_i as our co-efficients, yielding the
1433 * following representation of historical load:
1434 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1436 * We choose y based on the with of a reasonably scheduling period, fixing:
1439 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1440 * approximately half as much as the contribution to load within the last ms
1443 * When a period "rolls over" and we have new u_0`, multiplying the previous
1444 * sum again by y is sufficient to update:
1445 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1446 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1448 static __always_inline int __update_entity_runnable_avg(u64 now,
1449 struct sched_avg *sa,
1453 u32 runnable_contrib;
1454 int delta_w, decayed = 0;
1456 delta = now - sa->last_runnable_update;
1458 * This should only happen when time goes backwards, which it
1459 * unfortunately does during sched clock init when we swap over to TSC.
1461 if ((s64)delta < 0) {
1462 sa->last_runnable_update = now;
1467 * Use 1024ns as the unit of measurement since it's a reasonable
1468 * approximation of 1us and fast to compute.
1473 sa->last_runnable_update = now;
1475 /* delta_w is the amount already accumulated against our next period */
1476 delta_w = sa->runnable_avg_period % 1024;
1477 if (delta + delta_w >= 1024) {
1478 /* period roll-over */
1482 * Now that we know we're crossing a period boundary, figure
1483 * out how much from delta we need to complete the current
1484 * period and accrue it.
1486 delta_w = 1024 - delta_w;
1488 sa->runnable_avg_sum += delta_w;
1489 sa->runnable_avg_period += delta_w;
1493 /* Figure out how many additional periods this update spans */
1494 periods = delta / 1024;
1497 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1499 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1502 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1503 runnable_contrib = __compute_runnable_contrib(periods);
1505 sa->runnable_avg_sum += runnable_contrib;
1506 sa->runnable_avg_period += runnable_contrib;
1509 /* Remainder of delta accrued against u_0` */
1511 sa->runnable_avg_sum += delta;
1512 sa->runnable_avg_period += delta;
1517 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1518 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1520 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1521 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1523 decays -= se->avg.decay_count;
1527 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1528 se->avg.decay_count = 0;
1533 #ifdef CONFIG_FAIR_GROUP_SCHED
1534 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1537 struct task_group *tg = cfs_rq->tg;
1540 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1541 tg_contrib -= cfs_rq->tg_load_contrib;
1543 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1544 atomic_long_add(tg_contrib, &tg->load_avg);
1545 cfs_rq->tg_load_contrib += tg_contrib;
1550 * Aggregate cfs_rq runnable averages into an equivalent task_group
1551 * representation for computing load contributions.
1553 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1554 struct cfs_rq *cfs_rq)
1556 struct task_group *tg = cfs_rq->tg;
1559 /* The fraction of a cpu used by this cfs_rq */
1560 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1561 sa->runnable_avg_period + 1);
1562 contrib -= cfs_rq->tg_runnable_contrib;
1564 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1565 atomic_add(contrib, &tg->runnable_avg);
1566 cfs_rq->tg_runnable_contrib += contrib;
1570 static inline void __update_group_entity_contrib(struct sched_entity *se)
1572 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1573 struct task_group *tg = cfs_rq->tg;
1578 contrib = cfs_rq->tg_load_contrib * tg->shares;
1579 se->avg.load_avg_contrib = div_u64(contrib,
1580 atomic_long_read(&tg->load_avg) + 1);
1583 * For group entities we need to compute a correction term in the case
1584 * that they are consuming <1 cpu so that we would contribute the same
1585 * load as a task of equal weight.
1587 * Explicitly co-ordinating this measurement would be expensive, but
1588 * fortunately the sum of each cpus contribution forms a usable
1589 * lower-bound on the true value.
1591 * Consider the aggregate of 2 contributions. Either they are disjoint
1592 * (and the sum represents true value) or they are disjoint and we are
1593 * understating by the aggregate of their overlap.
1595 * Extending this to N cpus, for a given overlap, the maximum amount we
1596 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1597 * cpus that overlap for this interval and w_i is the interval width.
1599 * On a small machine; the first term is well-bounded which bounds the
1600 * total error since w_i is a subset of the period. Whereas on a
1601 * larger machine, while this first term can be larger, if w_i is the
1602 * of consequential size guaranteed to see n_i*w_i quickly converge to
1603 * our upper bound of 1-cpu.
1605 runnable_avg = atomic_read(&tg->runnable_avg);
1606 if (runnable_avg < NICE_0_LOAD) {
1607 se->avg.load_avg_contrib *= runnable_avg;
1608 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1612 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1613 int force_update) {}
1614 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1615 struct cfs_rq *cfs_rq) {}
1616 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1619 static inline void __update_task_entity_contrib(struct sched_entity *se)
1623 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1624 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1625 contrib /= (se->avg.runnable_avg_period + 1);
1626 se->avg.load_avg_contrib = scale_load(contrib);
1629 /* Compute the current contribution to load_avg by se, return any delta */
1630 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1632 long old_contrib = se->avg.load_avg_contrib;
1634 if (entity_is_task(se)) {
1635 __update_task_entity_contrib(se);
1637 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1638 __update_group_entity_contrib(se);
1641 return se->avg.load_avg_contrib - old_contrib;
1644 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1647 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1648 cfs_rq->blocked_load_avg -= load_contrib;
1650 cfs_rq->blocked_load_avg = 0;
1653 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1655 /* Update a sched_entity's runnable average */
1656 static inline void update_entity_load_avg(struct sched_entity *se,
1659 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1664 * For a group entity we need to use their owned cfs_rq_clock_task() in
1665 * case they are the parent of a throttled hierarchy.
1667 if (entity_is_task(se))
1668 now = cfs_rq_clock_task(cfs_rq);
1670 now = cfs_rq_clock_task(group_cfs_rq(se));
1672 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1675 contrib_delta = __update_entity_load_avg_contrib(se);
1681 cfs_rq->runnable_load_avg += contrib_delta;
1683 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1687 * Decay the load contributed by all blocked children and account this so that
1688 * their contribution may appropriately discounted when they wake up.
1690 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1692 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1695 decays = now - cfs_rq->last_decay;
1696 if (!decays && !force_update)
1699 if (atomic_long_read(&cfs_rq->removed_load)) {
1700 unsigned long removed_load;
1701 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1702 subtract_blocked_load_contrib(cfs_rq, removed_load);
1706 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1708 atomic64_add(decays, &cfs_rq->decay_counter);
1709 cfs_rq->last_decay = now;
1712 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1715 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1717 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1718 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1721 /* Add the load generated by se into cfs_rq's child load-average */
1722 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1723 struct sched_entity *se,
1727 * We track migrations using entity decay_count <= 0, on a wake-up
1728 * migration we use a negative decay count to track the remote decays
1729 * accumulated while sleeping.
1731 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1732 * are seen by enqueue_entity_load_avg() as a migration with an already
1733 * constructed load_avg_contrib.
1735 if (unlikely(se->avg.decay_count <= 0)) {
1736 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1737 if (se->avg.decay_count) {
1739 * In a wake-up migration we have to approximate the
1740 * time sleeping. This is because we can't synchronize
1741 * clock_task between the two cpus, and it is not
1742 * guaranteed to be read-safe. Instead, we can
1743 * approximate this using our carried decays, which are
1744 * explicitly atomically readable.
1746 se->avg.last_runnable_update -= (-se->avg.decay_count)
1748 update_entity_load_avg(se, 0);
1749 /* Indicate that we're now synchronized and on-rq */
1750 se->avg.decay_count = 0;
1755 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1756 * would have made count negative); we must be careful to avoid
1757 * double-accounting blocked time after synchronizing decays.
1759 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1763 /* migrated tasks did not contribute to our blocked load */
1765 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1766 update_entity_load_avg(se, 0);
1769 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1770 /* we force update consideration on load-balancer moves */
1771 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1775 * Remove se's load from this cfs_rq child load-average, if the entity is
1776 * transitioning to a blocked state we track its projected decay using
1779 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1780 struct sched_entity *se,
1783 update_entity_load_avg(se, 1);
1784 /* we force update consideration on load-balancer moves */
1785 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1787 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1789 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1790 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1791 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1795 * Update the rq's load with the elapsed running time before entering
1796 * idle. if the last scheduled task is not a CFS task, idle_enter will
1797 * be the only way to update the runnable statistic.
1799 void idle_enter_fair(struct rq *this_rq)
1801 update_rq_runnable_avg(this_rq, 1);
1805 * Update the rq's load with the elapsed idle time before a task is
1806 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1807 * be the only way to update the runnable statistic.
1809 void idle_exit_fair(struct rq *this_rq)
1811 update_rq_runnable_avg(this_rq, 0);
1815 static inline void update_entity_load_avg(struct sched_entity *se,
1816 int update_cfs_rq) {}
1817 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1818 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1819 struct sched_entity *se,
1821 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1822 struct sched_entity *se,
1824 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1825 int force_update) {}
1828 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1830 #ifdef CONFIG_SCHEDSTATS
1831 struct task_struct *tsk = NULL;
1833 if (entity_is_task(se))
1836 if (se->statistics.sleep_start) {
1837 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1842 if (unlikely(delta > se->statistics.sleep_max))
1843 se->statistics.sleep_max = delta;
1845 se->statistics.sleep_start = 0;
1846 se->statistics.sum_sleep_runtime += delta;
1849 account_scheduler_latency(tsk, delta >> 10, 1);
1850 trace_sched_stat_sleep(tsk, delta);
1853 if (se->statistics.block_start) {
1854 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1859 if (unlikely(delta > se->statistics.block_max))
1860 se->statistics.block_max = delta;
1862 se->statistics.block_start = 0;
1863 se->statistics.sum_sleep_runtime += delta;
1866 if (tsk->in_iowait) {
1867 se->statistics.iowait_sum += delta;
1868 se->statistics.iowait_count++;
1869 trace_sched_stat_iowait(tsk, delta);
1872 trace_sched_stat_blocked(tsk, delta);
1875 * Blocking time is in units of nanosecs, so shift by
1876 * 20 to get a milliseconds-range estimation of the
1877 * amount of time that the task spent sleeping:
1879 if (unlikely(prof_on == SLEEP_PROFILING)) {
1880 profile_hits(SLEEP_PROFILING,
1881 (void *)get_wchan(tsk),
1884 account_scheduler_latency(tsk, delta >> 10, 0);
1890 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1892 #ifdef CONFIG_SCHED_DEBUG
1893 s64 d = se->vruntime - cfs_rq->min_vruntime;
1898 if (d > 3*sysctl_sched_latency)
1899 schedstat_inc(cfs_rq, nr_spread_over);
1904 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1906 u64 vruntime = cfs_rq->min_vruntime;
1909 * The 'current' period is already promised to the current tasks,
1910 * however the extra weight of the new task will slow them down a
1911 * little, place the new task so that it fits in the slot that
1912 * stays open at the end.
1914 if (initial && sched_feat(START_DEBIT))
1915 vruntime += sched_vslice(cfs_rq, se);
1917 /* sleeps up to a single latency don't count. */
1919 unsigned long thresh = sysctl_sched_latency;
1922 * Halve their sleep time's effect, to allow
1923 * for a gentler effect of sleepers:
1925 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1931 /* ensure we never gain time by being placed backwards. */
1932 se->vruntime = max_vruntime(se->vruntime, vruntime);
1935 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1938 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1941 * Update the normalized vruntime before updating min_vruntime
1942 * through calling update_curr().
1944 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1945 se->vruntime += cfs_rq->min_vruntime;
1948 * Update run-time statistics of the 'current'.
1950 update_curr(cfs_rq);
1951 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1952 account_entity_enqueue(cfs_rq, se);
1953 update_cfs_shares(cfs_rq);
1955 if (flags & ENQUEUE_WAKEUP) {
1956 place_entity(cfs_rq, se, 0);
1957 enqueue_sleeper(cfs_rq, se);
1960 update_stats_enqueue(cfs_rq, se);
1961 check_spread(cfs_rq, se);
1962 if (se != cfs_rq->curr)
1963 __enqueue_entity(cfs_rq, se);
1966 if (cfs_rq->nr_running == 1) {
1967 list_add_leaf_cfs_rq(cfs_rq);
1968 check_enqueue_throttle(cfs_rq);
1972 static void __clear_buddies_last(struct sched_entity *se)
1974 for_each_sched_entity(se) {
1975 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1976 if (cfs_rq->last == se)
1977 cfs_rq->last = NULL;
1983 static void __clear_buddies_next(struct sched_entity *se)
1985 for_each_sched_entity(se) {
1986 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1987 if (cfs_rq->next == se)
1988 cfs_rq->next = NULL;
1994 static void __clear_buddies_skip(struct sched_entity *se)
1996 for_each_sched_entity(se) {
1997 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1998 if (cfs_rq->skip == se)
1999 cfs_rq->skip = NULL;
2005 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2007 if (cfs_rq->last == se)
2008 __clear_buddies_last(se);
2010 if (cfs_rq->next == se)
2011 __clear_buddies_next(se);
2013 if (cfs_rq->skip == se)
2014 __clear_buddies_skip(se);
2017 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2020 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2023 * Update run-time statistics of the 'current'.
2025 update_curr(cfs_rq);
2026 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2028 update_stats_dequeue(cfs_rq, se);
2029 if (flags & DEQUEUE_SLEEP) {
2030 #ifdef CONFIG_SCHEDSTATS
2031 if (entity_is_task(se)) {
2032 struct task_struct *tsk = task_of(se);
2034 if (tsk->state & TASK_INTERRUPTIBLE)
2035 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2036 if (tsk->state & TASK_UNINTERRUPTIBLE)
2037 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2042 clear_buddies(cfs_rq, se);
2044 if (se != cfs_rq->curr)
2045 __dequeue_entity(cfs_rq, se);
2047 account_entity_dequeue(cfs_rq, se);
2050 * Normalize the entity after updating the min_vruntime because the
2051 * update can refer to the ->curr item and we need to reflect this
2052 * movement in our normalized position.
2054 if (!(flags & DEQUEUE_SLEEP))
2055 se->vruntime -= cfs_rq->min_vruntime;
2057 /* return excess runtime on last dequeue */
2058 return_cfs_rq_runtime(cfs_rq);
2060 update_min_vruntime(cfs_rq);
2061 update_cfs_shares(cfs_rq);
2065 * Preempt the current task with a newly woken task if needed:
2068 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2070 unsigned long ideal_runtime, delta_exec;
2071 struct sched_entity *se;
2074 ideal_runtime = sched_slice(cfs_rq, curr);
2075 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2076 if (delta_exec > ideal_runtime) {
2077 resched_task(rq_of(cfs_rq)->curr);
2079 * The current task ran long enough, ensure it doesn't get
2080 * re-elected due to buddy favours.
2082 clear_buddies(cfs_rq, curr);
2087 * Ensure that a task that missed wakeup preemption by a
2088 * narrow margin doesn't have to wait for a full slice.
2089 * This also mitigates buddy induced latencies under load.
2091 if (delta_exec < sysctl_sched_min_granularity)
2094 se = __pick_first_entity(cfs_rq);
2095 delta = curr->vruntime - se->vruntime;
2100 if (delta > ideal_runtime)
2101 resched_task(rq_of(cfs_rq)->curr);
2105 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2107 /* 'current' is not kept within the tree. */
2110 * Any task has to be enqueued before it get to execute on
2111 * a CPU. So account for the time it spent waiting on the
2114 update_stats_wait_end(cfs_rq, se);
2115 __dequeue_entity(cfs_rq, se);
2118 update_stats_curr_start(cfs_rq, se);
2120 #ifdef CONFIG_SCHEDSTATS
2122 * Track our maximum slice length, if the CPU's load is at
2123 * least twice that of our own weight (i.e. dont track it
2124 * when there are only lesser-weight tasks around):
2126 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2127 se->statistics.slice_max = max(se->statistics.slice_max,
2128 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2131 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2135 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2138 * Pick the next process, keeping these things in mind, in this order:
2139 * 1) keep things fair between processes/task groups
2140 * 2) pick the "next" process, since someone really wants that to run
2141 * 3) pick the "last" process, for cache locality
2142 * 4) do not run the "skip" process, if something else is available
2144 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2146 struct sched_entity *se = __pick_first_entity(cfs_rq);
2147 struct sched_entity *left = se;
2150 * Avoid running the skip buddy, if running something else can
2151 * be done without getting too unfair.
2153 if (cfs_rq->skip == se) {
2154 struct sched_entity *second = __pick_next_entity(se);
2155 if (second && wakeup_preempt_entity(second, left) < 1)
2160 * Prefer last buddy, try to return the CPU to a preempted task.
2162 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2166 * Someone really wants this to run. If it's not unfair, run it.
2168 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2171 clear_buddies(cfs_rq, se);
2176 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2178 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2181 * If still on the runqueue then deactivate_task()
2182 * was not called and update_curr() has to be done:
2185 update_curr(cfs_rq);
2187 /* throttle cfs_rqs exceeding runtime */
2188 check_cfs_rq_runtime(cfs_rq);
2190 check_spread(cfs_rq, prev);
2192 update_stats_wait_start(cfs_rq, prev);
2193 /* Put 'current' back into the tree. */
2194 __enqueue_entity(cfs_rq, prev);
2195 /* in !on_rq case, update occurred at dequeue */
2196 update_entity_load_avg(prev, 1);
2198 cfs_rq->curr = NULL;
2202 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2205 * Update run-time statistics of the 'current'.
2207 update_curr(cfs_rq);
2210 * Ensure that runnable average is periodically updated.
2212 update_entity_load_avg(curr, 1);
2213 update_cfs_rq_blocked_load(cfs_rq, 1);
2214 update_cfs_shares(cfs_rq);
2216 #ifdef CONFIG_SCHED_HRTICK
2218 * queued ticks are scheduled to match the slice, so don't bother
2219 * validating it and just reschedule.
2222 resched_task(rq_of(cfs_rq)->curr);
2226 * don't let the period tick interfere with the hrtick preemption
2228 if (!sched_feat(DOUBLE_TICK) &&
2229 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2233 if (cfs_rq->nr_running > 1)
2234 check_preempt_tick(cfs_rq, curr);
2238 /**************************************************
2239 * CFS bandwidth control machinery
2242 #ifdef CONFIG_CFS_BANDWIDTH
2244 #ifdef HAVE_JUMP_LABEL
2245 static struct static_key __cfs_bandwidth_used;
2247 static inline bool cfs_bandwidth_used(void)
2249 return static_key_false(&__cfs_bandwidth_used);
2252 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2254 /* only need to count groups transitioning between enabled/!enabled */
2255 if (enabled && !was_enabled)
2256 static_key_slow_inc(&__cfs_bandwidth_used);
2257 else if (!enabled && was_enabled)
2258 static_key_slow_dec(&__cfs_bandwidth_used);
2260 #else /* HAVE_JUMP_LABEL */
2261 static bool cfs_bandwidth_used(void)
2266 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2267 #endif /* HAVE_JUMP_LABEL */
2270 * default period for cfs group bandwidth.
2271 * default: 0.1s, units: nanoseconds
2273 static inline u64 default_cfs_period(void)
2275 return 100000000ULL;
2278 static inline u64 sched_cfs_bandwidth_slice(void)
2280 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2284 * Replenish runtime according to assigned quota and update expiration time.
2285 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2286 * additional synchronization around rq->lock.
2288 * requires cfs_b->lock
2290 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2294 if (cfs_b->quota == RUNTIME_INF)
2297 now = sched_clock_cpu(smp_processor_id());
2298 cfs_b->runtime = cfs_b->quota;
2299 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2302 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2304 return &tg->cfs_bandwidth;
2307 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2308 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2310 if (unlikely(cfs_rq->throttle_count))
2311 return cfs_rq->throttled_clock_task;
2313 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2316 /* returns 0 on failure to allocate runtime */
2317 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2319 struct task_group *tg = cfs_rq->tg;
2320 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2321 u64 amount = 0, min_amount, expires;
2323 /* note: this is a positive sum as runtime_remaining <= 0 */
2324 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2326 raw_spin_lock(&cfs_b->lock);
2327 if (cfs_b->quota == RUNTIME_INF)
2328 amount = min_amount;
2331 * If the bandwidth pool has become inactive, then at least one
2332 * period must have elapsed since the last consumption.
2333 * Refresh the global state and ensure bandwidth timer becomes
2336 if (!cfs_b->timer_active) {
2337 __refill_cfs_bandwidth_runtime(cfs_b);
2338 __start_cfs_bandwidth(cfs_b);
2341 if (cfs_b->runtime > 0) {
2342 amount = min(cfs_b->runtime, min_amount);
2343 cfs_b->runtime -= amount;
2347 expires = cfs_b->runtime_expires;
2348 raw_spin_unlock(&cfs_b->lock);
2350 cfs_rq->runtime_remaining += amount;
2352 * we may have advanced our local expiration to account for allowed
2353 * spread between our sched_clock and the one on which runtime was
2356 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2357 cfs_rq->runtime_expires = expires;
2359 return cfs_rq->runtime_remaining > 0;
2363 * Note: This depends on the synchronization provided by sched_clock and the
2364 * fact that rq->clock snapshots this value.
2366 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2368 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2370 /* if the deadline is ahead of our clock, nothing to do */
2371 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2374 if (cfs_rq->runtime_remaining < 0)
2378 * If the local deadline has passed we have to consider the
2379 * possibility that our sched_clock is 'fast' and the global deadline
2380 * has not truly expired.
2382 * Fortunately we can check determine whether this the case by checking
2383 * whether the global deadline has advanced.
2386 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2387 /* extend local deadline, drift is bounded above by 2 ticks */
2388 cfs_rq->runtime_expires += TICK_NSEC;
2390 /* global deadline is ahead, expiration has passed */
2391 cfs_rq->runtime_remaining = 0;
2395 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2396 unsigned long delta_exec)
2398 /* dock delta_exec before expiring quota (as it could span periods) */
2399 cfs_rq->runtime_remaining -= delta_exec;
2400 expire_cfs_rq_runtime(cfs_rq);
2402 if (likely(cfs_rq->runtime_remaining > 0))
2406 * if we're unable to extend our runtime we resched so that the active
2407 * hierarchy can be throttled
2409 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2410 resched_task(rq_of(cfs_rq)->curr);
2413 static __always_inline
2414 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2416 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2419 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2422 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2424 return cfs_bandwidth_used() && cfs_rq->throttled;
2427 /* check whether cfs_rq, or any parent, is throttled */
2428 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2430 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2434 * Ensure that neither of the group entities corresponding to src_cpu or
2435 * dest_cpu are members of a throttled hierarchy when performing group
2436 * load-balance operations.
2438 static inline int throttled_lb_pair(struct task_group *tg,
2439 int src_cpu, int dest_cpu)
2441 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2443 src_cfs_rq = tg->cfs_rq[src_cpu];
2444 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2446 return throttled_hierarchy(src_cfs_rq) ||
2447 throttled_hierarchy(dest_cfs_rq);
2450 /* updated child weight may affect parent so we have to do this bottom up */
2451 static int tg_unthrottle_up(struct task_group *tg, void *data)
2453 struct rq *rq = data;
2454 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2456 cfs_rq->throttle_count--;
2458 if (!cfs_rq->throttle_count) {
2459 /* adjust cfs_rq_clock_task() */
2460 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2461 cfs_rq->throttled_clock_task;
2468 static int tg_throttle_down(struct task_group *tg, void *data)
2470 struct rq *rq = data;
2471 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2473 /* group is entering throttled state, stop time */
2474 if (!cfs_rq->throttle_count)
2475 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2476 cfs_rq->throttle_count++;
2481 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2483 struct rq *rq = rq_of(cfs_rq);
2484 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2485 struct sched_entity *se;
2486 long task_delta, dequeue = 1;
2488 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2490 /* freeze hierarchy runnable averages while throttled */
2492 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2495 task_delta = cfs_rq->h_nr_running;
2496 for_each_sched_entity(se) {
2497 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2498 /* throttled entity or throttle-on-deactivate */
2503 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2504 qcfs_rq->h_nr_running -= task_delta;
2506 if (qcfs_rq->load.weight)
2511 rq->nr_running -= task_delta;
2513 cfs_rq->throttled = 1;
2514 cfs_rq->throttled_clock = rq_clock(rq);
2515 raw_spin_lock(&cfs_b->lock);
2516 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2517 raw_spin_unlock(&cfs_b->lock);
2520 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2522 struct rq *rq = rq_of(cfs_rq);
2523 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2524 struct sched_entity *se;
2528 se = cfs_rq->tg->se[cpu_of(rq)];
2530 cfs_rq->throttled = 0;
2532 update_rq_clock(rq);
2534 raw_spin_lock(&cfs_b->lock);
2535 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2536 list_del_rcu(&cfs_rq->throttled_list);
2537 raw_spin_unlock(&cfs_b->lock);
2539 /* update hierarchical throttle state */
2540 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2542 if (!cfs_rq->load.weight)
2545 task_delta = cfs_rq->h_nr_running;
2546 for_each_sched_entity(se) {
2550 cfs_rq = cfs_rq_of(se);
2552 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2553 cfs_rq->h_nr_running += task_delta;
2555 if (cfs_rq_throttled(cfs_rq))
2560 rq->nr_running += task_delta;
2562 /* determine whether we need to wake up potentially idle cpu */
2563 if (rq->curr == rq->idle && rq->cfs.nr_running)
2564 resched_task(rq->curr);
2567 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2568 u64 remaining, u64 expires)
2570 struct cfs_rq *cfs_rq;
2571 u64 runtime = remaining;
2574 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2576 struct rq *rq = rq_of(cfs_rq);
2578 raw_spin_lock(&rq->lock);
2579 if (!cfs_rq_throttled(cfs_rq))
2582 runtime = -cfs_rq->runtime_remaining + 1;
2583 if (runtime > remaining)
2584 runtime = remaining;
2585 remaining -= runtime;
2587 cfs_rq->runtime_remaining += runtime;
2588 cfs_rq->runtime_expires = expires;
2590 /* we check whether we're throttled above */
2591 if (cfs_rq->runtime_remaining > 0)
2592 unthrottle_cfs_rq(cfs_rq);
2595 raw_spin_unlock(&rq->lock);
2606 * Responsible for refilling a task_group's bandwidth and unthrottling its
2607 * cfs_rqs as appropriate. If there has been no activity within the last
2608 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2609 * used to track this state.
2611 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2613 u64 runtime, runtime_expires;
2614 int idle = 1, throttled;
2616 raw_spin_lock(&cfs_b->lock);
2617 /* no need to continue the timer with no bandwidth constraint */
2618 if (cfs_b->quota == RUNTIME_INF)
2621 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2622 /* idle depends on !throttled (for the case of a large deficit) */
2623 idle = cfs_b->idle && !throttled;
2624 cfs_b->nr_periods += overrun;
2626 /* if we're going inactive then everything else can be deferred */
2630 __refill_cfs_bandwidth_runtime(cfs_b);
2633 /* mark as potentially idle for the upcoming period */
2638 /* account preceding periods in which throttling occurred */
2639 cfs_b->nr_throttled += overrun;
2642 * There are throttled entities so we must first use the new bandwidth
2643 * to unthrottle them before making it generally available. This
2644 * ensures that all existing debts will be paid before a new cfs_rq is
2647 runtime = cfs_b->runtime;
2648 runtime_expires = cfs_b->runtime_expires;
2652 * This check is repeated as we are holding onto the new bandwidth
2653 * while we unthrottle. This can potentially race with an unthrottled
2654 * group trying to acquire new bandwidth from the global pool.
2656 while (throttled && runtime > 0) {
2657 raw_spin_unlock(&cfs_b->lock);
2658 /* we can't nest cfs_b->lock while distributing bandwidth */
2659 runtime = distribute_cfs_runtime(cfs_b, runtime,
2661 raw_spin_lock(&cfs_b->lock);
2663 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2666 /* return (any) remaining runtime */
2667 cfs_b->runtime = runtime;
2669 * While we are ensured activity in the period following an
2670 * unthrottle, this also covers the case in which the new bandwidth is
2671 * insufficient to cover the existing bandwidth deficit. (Forcing the
2672 * timer to remain active while there are any throttled entities.)
2677 cfs_b->timer_active = 0;
2678 raw_spin_unlock(&cfs_b->lock);
2683 /* a cfs_rq won't donate quota below this amount */
2684 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2685 /* minimum remaining period time to redistribute slack quota */
2686 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2687 /* how long we wait to gather additional slack before distributing */
2688 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2690 /* are we near the end of the current quota period? */
2691 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2693 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2696 /* if the call-back is running a quota refresh is already occurring */
2697 if (hrtimer_callback_running(refresh_timer))
2700 /* is a quota refresh about to occur? */
2701 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2702 if (remaining < min_expire)
2708 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2710 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2712 /* if there's a quota refresh soon don't bother with slack */
2713 if (runtime_refresh_within(cfs_b, min_left))
2716 start_bandwidth_timer(&cfs_b->slack_timer,
2717 ns_to_ktime(cfs_bandwidth_slack_period));
2720 /* we know any runtime found here is valid as update_curr() precedes return */
2721 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2723 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2724 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2726 if (slack_runtime <= 0)
2729 raw_spin_lock(&cfs_b->lock);
2730 if (cfs_b->quota != RUNTIME_INF &&
2731 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2732 cfs_b->runtime += slack_runtime;
2734 /* we are under rq->lock, defer unthrottling using a timer */
2735 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2736 !list_empty(&cfs_b->throttled_cfs_rq))
2737 start_cfs_slack_bandwidth(cfs_b);
2739 raw_spin_unlock(&cfs_b->lock);
2741 /* even if it's not valid for return we don't want to try again */
2742 cfs_rq->runtime_remaining -= slack_runtime;
2745 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2747 if (!cfs_bandwidth_used())
2750 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2753 __return_cfs_rq_runtime(cfs_rq);
2757 * This is done with a timer (instead of inline with bandwidth return) since
2758 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2760 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2762 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2765 /* confirm we're still not at a refresh boundary */
2766 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2769 raw_spin_lock(&cfs_b->lock);
2770 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2771 runtime = cfs_b->runtime;
2774 expires = cfs_b->runtime_expires;
2775 raw_spin_unlock(&cfs_b->lock);
2780 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2782 raw_spin_lock(&cfs_b->lock);
2783 if (expires == cfs_b->runtime_expires)
2784 cfs_b->runtime = runtime;
2785 raw_spin_unlock(&cfs_b->lock);
2789 * When a group wakes up we want to make sure that its quota is not already
2790 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2791 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2793 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2795 if (!cfs_bandwidth_used())
2798 /* an active group must be handled by the update_curr()->put() path */
2799 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2802 /* ensure the group is not already throttled */
2803 if (cfs_rq_throttled(cfs_rq))
2806 /* update runtime allocation */
2807 account_cfs_rq_runtime(cfs_rq, 0);
2808 if (cfs_rq->runtime_remaining <= 0)
2809 throttle_cfs_rq(cfs_rq);
2812 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2813 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2815 if (!cfs_bandwidth_used())
2818 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2822 * it's possible for a throttled entity to be forced into a running
2823 * state (e.g. set_curr_task), in this case we're finished.
2825 if (cfs_rq_throttled(cfs_rq))
2828 throttle_cfs_rq(cfs_rq);
2831 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2833 struct cfs_bandwidth *cfs_b =
2834 container_of(timer, struct cfs_bandwidth, slack_timer);
2835 do_sched_cfs_slack_timer(cfs_b);
2837 return HRTIMER_NORESTART;
2840 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2842 struct cfs_bandwidth *cfs_b =
2843 container_of(timer, struct cfs_bandwidth, period_timer);
2849 now = hrtimer_cb_get_time(timer);
2850 overrun = hrtimer_forward(timer, now, cfs_b->period);
2855 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2858 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2861 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2863 raw_spin_lock_init(&cfs_b->lock);
2865 cfs_b->quota = RUNTIME_INF;
2866 cfs_b->period = ns_to_ktime(default_cfs_period());
2868 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2869 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2870 cfs_b->period_timer.function = sched_cfs_period_timer;
2871 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2872 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2875 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2877 cfs_rq->runtime_enabled = 0;
2878 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2881 /* requires cfs_b->lock, may release to reprogram timer */
2882 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2885 * The timer may be active because we're trying to set a new bandwidth
2886 * period or because we're racing with the tear-down path
2887 * (timer_active==0 becomes visible before the hrtimer call-back
2888 * terminates). In either case we ensure that it's re-programmed
2890 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2891 raw_spin_unlock(&cfs_b->lock);
2892 /* ensure cfs_b->lock is available while we wait */
2893 hrtimer_cancel(&cfs_b->period_timer);
2895 raw_spin_lock(&cfs_b->lock);
2896 /* if someone else restarted the timer then we're done */
2897 if (cfs_b->timer_active)
2901 cfs_b->timer_active = 1;
2902 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2905 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2907 hrtimer_cancel(&cfs_b->period_timer);
2908 hrtimer_cancel(&cfs_b->slack_timer);
2911 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2913 struct cfs_rq *cfs_rq;
2915 for_each_leaf_cfs_rq(rq, cfs_rq) {
2916 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2918 if (!cfs_rq->runtime_enabled)
2922 * clock_task is not advancing so we just need to make sure
2923 * there's some valid quota amount
2925 cfs_rq->runtime_remaining = cfs_b->quota;
2926 if (cfs_rq_throttled(cfs_rq))
2927 unthrottle_cfs_rq(cfs_rq);
2931 #else /* CONFIG_CFS_BANDWIDTH */
2932 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2934 return rq_clock_task(rq_of(cfs_rq));
2937 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2938 unsigned long delta_exec) {}
2939 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2940 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2941 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2943 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2948 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2953 static inline int throttled_lb_pair(struct task_group *tg,
2954 int src_cpu, int dest_cpu)
2959 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2961 #ifdef CONFIG_FAIR_GROUP_SCHED
2962 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2965 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2969 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2970 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2972 #endif /* CONFIG_CFS_BANDWIDTH */
2974 /**************************************************
2975 * CFS operations on tasks:
2978 #ifdef CONFIG_SCHED_HRTICK
2979 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2981 struct sched_entity *se = &p->se;
2982 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2984 WARN_ON(task_rq(p) != rq);
2986 if (cfs_rq->nr_running > 1) {
2987 u64 slice = sched_slice(cfs_rq, se);
2988 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2989 s64 delta = slice - ran;
2998 * Don't schedule slices shorter than 10000ns, that just
2999 * doesn't make sense. Rely on vruntime for fairness.
3002 delta = max_t(s64, 10000LL, delta);
3004 hrtick_start(rq, delta);
3009 * called from enqueue/dequeue and updates the hrtick when the
3010 * current task is from our class and nr_running is low enough
3013 static void hrtick_update(struct rq *rq)
3015 struct task_struct *curr = rq->curr;
3017 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3020 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3021 hrtick_start_fair(rq, curr);
3023 #else /* !CONFIG_SCHED_HRTICK */
3025 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3029 static inline void hrtick_update(struct rq *rq)
3035 * The enqueue_task method is called before nr_running is
3036 * increased. Here we update the fair scheduling stats and
3037 * then put the task into the rbtree:
3040 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3042 struct cfs_rq *cfs_rq;
3043 struct sched_entity *se = &p->se;
3045 for_each_sched_entity(se) {
3048 cfs_rq = cfs_rq_of(se);
3049 enqueue_entity(cfs_rq, se, flags);
3052 * end evaluation on encountering a throttled cfs_rq
3054 * note: in the case of encountering a throttled cfs_rq we will
3055 * post the final h_nr_running increment below.
3057 if (cfs_rq_throttled(cfs_rq))
3059 cfs_rq->h_nr_running++;
3061 flags = ENQUEUE_WAKEUP;
3064 for_each_sched_entity(se) {
3065 cfs_rq = cfs_rq_of(se);
3066 cfs_rq->h_nr_running++;
3068 if (cfs_rq_throttled(cfs_rq))
3071 update_cfs_shares(cfs_rq);
3072 update_entity_load_avg(se, 1);
3076 update_rq_runnable_avg(rq, rq->nr_running);
3082 static void set_next_buddy(struct sched_entity *se);
3085 * The dequeue_task method is called before nr_running is
3086 * decreased. We remove the task from the rbtree and
3087 * update the fair scheduling stats:
3089 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3091 struct cfs_rq *cfs_rq;
3092 struct sched_entity *se = &p->se;
3093 int task_sleep = flags & DEQUEUE_SLEEP;
3095 for_each_sched_entity(se) {
3096 cfs_rq = cfs_rq_of(se);
3097 dequeue_entity(cfs_rq, se, flags);
3100 * end evaluation on encountering a throttled cfs_rq
3102 * note: in the case of encountering a throttled cfs_rq we will
3103 * post the final h_nr_running decrement below.
3105 if (cfs_rq_throttled(cfs_rq))
3107 cfs_rq->h_nr_running--;
3109 /* Don't dequeue parent if it has other entities besides us */
3110 if (cfs_rq->load.weight) {
3112 * Bias pick_next to pick a task from this cfs_rq, as
3113 * p is sleeping when it is within its sched_slice.
3115 if (task_sleep && parent_entity(se))
3116 set_next_buddy(parent_entity(se));
3118 /* avoid re-evaluating load for this entity */
3119 se = parent_entity(se);
3122 flags |= DEQUEUE_SLEEP;
3125 for_each_sched_entity(se) {
3126 cfs_rq = cfs_rq_of(se);
3127 cfs_rq->h_nr_running--;
3129 if (cfs_rq_throttled(cfs_rq))
3132 update_cfs_shares(cfs_rq);
3133 update_entity_load_avg(se, 1);
3138 update_rq_runnable_avg(rq, 1);
3144 /* Used instead of source_load when we know the type == 0 */
3145 static unsigned long weighted_cpuload(const int cpu)
3147 return cpu_rq(cpu)->cfs.runnable_load_avg;
3151 * Return a low guess at the load of a migration-source cpu weighted
3152 * according to the scheduling class and "nice" value.
3154 * We want to under-estimate the load of migration sources, to
3155 * balance conservatively.
3157 static unsigned long source_load(int cpu, int type)
3159 struct rq *rq = cpu_rq(cpu);
3160 unsigned long total = weighted_cpuload(cpu);
3162 if (type == 0 || !sched_feat(LB_BIAS))
3165 return min(rq->cpu_load[type-1], total);
3169 * Return a high guess at the load of a migration-target cpu weighted
3170 * according to the scheduling class and "nice" value.
3172 static unsigned long target_load(int cpu, int type)
3174 struct rq *rq = cpu_rq(cpu);
3175 unsigned long total = weighted_cpuload(cpu);
3177 if (type == 0 || !sched_feat(LB_BIAS))
3180 return max(rq->cpu_load[type-1], total);
3183 static unsigned long power_of(int cpu)
3185 return cpu_rq(cpu)->cpu_power;
3188 static unsigned long cpu_avg_load_per_task(int cpu)
3190 struct rq *rq = cpu_rq(cpu);
3191 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3192 unsigned long load_avg = rq->cfs.runnable_load_avg;
3195 return load_avg / nr_running;
3200 static void record_wakee(struct task_struct *p)
3203 * Rough decay (wiping) for cost saving, don't worry
3204 * about the boundary, really active task won't care
3207 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3208 current->wakee_flips = 0;
3209 current->wakee_flip_decay_ts = jiffies;
3212 if (current->last_wakee != p) {
3213 current->last_wakee = p;
3214 current->wakee_flips++;
3218 static void task_waking_fair(struct task_struct *p)
3220 struct sched_entity *se = &p->se;
3221 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3224 #ifndef CONFIG_64BIT
3225 u64 min_vruntime_copy;
3228 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3230 min_vruntime = cfs_rq->min_vruntime;
3231 } while (min_vruntime != min_vruntime_copy);
3233 min_vruntime = cfs_rq->min_vruntime;
3236 se->vruntime -= min_vruntime;
3240 #ifdef CONFIG_FAIR_GROUP_SCHED
3242 * effective_load() calculates the load change as seen from the root_task_group
3244 * Adding load to a group doesn't make a group heavier, but can cause movement
3245 * of group shares between cpus. Assuming the shares were perfectly aligned one
3246 * can calculate the shift in shares.
3248 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3249 * on this @cpu and results in a total addition (subtraction) of @wg to the
3250 * total group weight.
3252 * Given a runqueue weight distribution (rw_i) we can compute a shares
3253 * distribution (s_i) using:
3255 * s_i = rw_i / \Sum rw_j (1)
3257 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3258 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3259 * shares distribution (s_i):
3261 * rw_i = { 2, 4, 1, 0 }
3262 * s_i = { 2/7, 4/7, 1/7, 0 }
3264 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3265 * task used to run on and the CPU the waker is running on), we need to
3266 * compute the effect of waking a task on either CPU and, in case of a sync
3267 * wakeup, compute the effect of the current task going to sleep.
3269 * So for a change of @wl to the local @cpu with an overall group weight change
3270 * of @wl we can compute the new shares distribution (s'_i) using:
3272 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3274 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3275 * differences in waking a task to CPU 0. The additional task changes the
3276 * weight and shares distributions like:
3278 * rw'_i = { 3, 4, 1, 0 }
3279 * s'_i = { 3/8, 4/8, 1/8, 0 }
3281 * We can then compute the difference in effective weight by using:
3283 * dw_i = S * (s'_i - s_i) (3)
3285 * Where 'S' is the group weight as seen by its parent.
3287 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3288 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3289 * 4/7) times the weight of the group.
3291 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3293 struct sched_entity *se = tg->se[cpu];
3295 if (!tg->parent) /* the trivial, non-cgroup case */
3298 for_each_sched_entity(se) {
3304 * W = @wg + \Sum rw_j
3306 W = wg + calc_tg_weight(tg, se->my_q);
3311 w = se->my_q->load.weight + wl;
3314 * wl = S * s'_i; see (2)
3317 wl = (w * tg->shares) / W;
3322 * Per the above, wl is the new se->load.weight value; since
3323 * those are clipped to [MIN_SHARES, ...) do so now. See
3324 * calc_cfs_shares().
3326 if (wl < MIN_SHARES)
3330 * wl = dw_i = S * (s'_i - s_i); see (3)
3332 wl -= se->load.weight;
3335 * Recursively apply this logic to all parent groups to compute
3336 * the final effective load change on the root group. Since
3337 * only the @tg group gets extra weight, all parent groups can
3338 * only redistribute existing shares. @wl is the shift in shares
3339 * resulting from this level per the above.
3348 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3349 unsigned long wl, unsigned long wg)
3356 static int wake_wide(struct task_struct *p)
3358 int factor = this_cpu_read(sd_llc_size);
3361 * Yeah, it's the switching-frequency, could means many wakee or
3362 * rapidly switch, use factor here will just help to automatically
3363 * adjust the loose-degree, so bigger node will lead to more pull.
3365 if (p->wakee_flips > factor) {
3367 * wakee is somewhat hot, it needs certain amount of cpu
3368 * resource, so if waker is far more hot, prefer to leave
3371 if (current->wakee_flips > (factor * p->wakee_flips))
3378 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3380 s64 this_load, load;
3381 int idx, this_cpu, prev_cpu;
3382 unsigned long tl_per_task;
3383 struct task_group *tg;
3384 unsigned long weight;
3388 * If we wake multiple tasks be careful to not bounce
3389 * ourselves around too much.
3395 this_cpu = smp_processor_id();
3396 prev_cpu = task_cpu(p);
3397 load = source_load(prev_cpu, idx);
3398 this_load = target_load(this_cpu, idx);
3401 * If sync wakeup then subtract the (maximum possible)
3402 * effect of the currently running task from the load
3403 * of the current CPU:
3406 tg = task_group(current);
3407 weight = current->se.load.weight;
3409 this_load += effective_load(tg, this_cpu, -weight, -weight);
3410 load += effective_load(tg, prev_cpu, 0, -weight);
3414 weight = p->se.load.weight;
3417 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3418 * due to the sync cause above having dropped this_load to 0, we'll
3419 * always have an imbalance, but there's really nothing you can do
3420 * about that, so that's good too.
3422 * Otherwise check if either cpus are near enough in load to allow this
3423 * task to be woken on this_cpu.
3425 if (this_load > 0) {
3426 s64 this_eff_load, prev_eff_load;
3428 this_eff_load = 100;
3429 this_eff_load *= power_of(prev_cpu);
3430 this_eff_load *= this_load +
3431 effective_load(tg, this_cpu, weight, weight);
3433 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3434 prev_eff_load *= power_of(this_cpu);
3435 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3437 balanced = this_eff_load <= prev_eff_load;
3442 * If the currently running task will sleep within
3443 * a reasonable amount of time then attract this newly
3446 if (sync && balanced)
3449 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3450 tl_per_task = cpu_avg_load_per_task(this_cpu);
3453 (this_load <= load &&
3454 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3456 * This domain has SD_WAKE_AFFINE and
3457 * p is cache cold in this domain, and
3458 * there is no bad imbalance.
3460 schedstat_inc(sd, ttwu_move_affine);
3461 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3469 * find_idlest_group finds and returns the least busy CPU group within the
3472 static struct sched_group *
3473 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3474 int this_cpu, int load_idx)
3476 struct sched_group *idlest = NULL, *group = sd->groups;
3477 unsigned long min_load = ULONG_MAX, this_load = 0;
3478 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3481 unsigned long load, avg_load;
3485 /* Skip over this group if it has no CPUs allowed */
3486 if (!cpumask_intersects(sched_group_cpus(group),
3487 tsk_cpus_allowed(p)))
3490 local_group = cpumask_test_cpu(this_cpu,
3491 sched_group_cpus(group));
3493 /* Tally up the load of all CPUs in the group */
3496 for_each_cpu(i, sched_group_cpus(group)) {
3497 /* Bias balancing toward cpus of our domain */
3499 load = source_load(i, load_idx);
3501 load = target_load(i, load_idx);
3506 /* Adjust by relative CPU power of the group */
3507 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3510 this_load = avg_load;
3511 } else if (avg_load < min_load) {
3512 min_load = avg_load;
3515 } while (group = group->next, group != sd->groups);
3517 if (!idlest || 100*this_load < imbalance*min_load)
3523 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3526 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3528 unsigned long load, min_load = ULONG_MAX;
3532 /* Traverse only the allowed CPUs */
3533 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3534 load = weighted_cpuload(i);
3536 if (load < min_load || (load == min_load && i == this_cpu)) {
3546 * Try and locate an idle CPU in the sched_domain.
3548 static int select_idle_sibling(struct task_struct *p, int target)
3550 struct sched_domain *sd;
3551 struct sched_group *sg;
3552 int i = task_cpu(p);
3554 if (idle_cpu(target))
3558 * If the prevous cpu is cache affine and idle, don't be stupid.
3560 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3564 * Otherwise, iterate the domains and find an elegible idle cpu.
3566 sd = rcu_dereference(per_cpu(sd_llc, target));
3567 for_each_lower_domain(sd) {
3570 if (!cpumask_intersects(sched_group_cpus(sg),
3571 tsk_cpus_allowed(p)))
3574 for_each_cpu(i, sched_group_cpus(sg)) {
3575 if (i == target || !idle_cpu(i))
3579 target = cpumask_first_and(sched_group_cpus(sg),
3580 tsk_cpus_allowed(p));
3584 } while (sg != sd->groups);
3591 * sched_balance_self: balance the current task (running on cpu) in domains
3592 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3595 * Balance, ie. select the least loaded group.
3597 * Returns the target CPU number, or the same CPU if no balancing is needed.
3599 * preempt must be disabled.
3602 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3604 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3605 int cpu = smp_processor_id();
3606 int prev_cpu = task_cpu(p);
3608 int want_affine = 0;
3609 int sync = wake_flags & WF_SYNC;
3611 if (p->nr_cpus_allowed == 1)
3614 if (sd_flag & SD_BALANCE_WAKE) {
3615 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3621 for_each_domain(cpu, tmp) {
3622 if (!(tmp->flags & SD_LOAD_BALANCE))
3626 * If both cpu and prev_cpu are part of this domain,
3627 * cpu is a valid SD_WAKE_AFFINE target.
3629 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3630 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3635 if (tmp->flags & sd_flag)
3640 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3643 new_cpu = select_idle_sibling(p, prev_cpu);
3648 int load_idx = sd->forkexec_idx;
3649 struct sched_group *group;
3652 if (!(sd->flags & sd_flag)) {
3657 if (sd_flag & SD_BALANCE_WAKE)
3658 load_idx = sd->wake_idx;
3660 group = find_idlest_group(sd, p, cpu, load_idx);
3666 new_cpu = find_idlest_cpu(group, p, cpu);
3667 if (new_cpu == -1 || new_cpu == cpu) {
3668 /* Now try balancing at a lower domain level of cpu */
3673 /* Now try balancing at a lower domain level of new_cpu */
3675 weight = sd->span_weight;
3677 for_each_domain(cpu, tmp) {
3678 if (weight <= tmp->span_weight)
3680 if (tmp->flags & sd_flag)
3683 /* while loop will break here if sd == NULL */
3692 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3693 * cfs_rq_of(p) references at time of call are still valid and identify the
3694 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3695 * other assumptions, including the state of rq->lock, should be made.
3698 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3700 struct sched_entity *se = &p->se;
3701 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3704 * Load tracking: accumulate removed load so that it can be processed
3705 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3706 * to blocked load iff they have a positive decay-count. It can never
3707 * be negative here since on-rq tasks have decay-count == 0.
3709 if (se->avg.decay_count) {
3710 se->avg.decay_count = -__synchronize_entity_decay(se);
3711 atomic_long_add(se->avg.load_avg_contrib,
3712 &cfs_rq->removed_load);
3715 #endif /* CONFIG_SMP */
3717 static unsigned long
3718 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3720 unsigned long gran = sysctl_sched_wakeup_granularity;
3723 * Since its curr running now, convert the gran from real-time
3724 * to virtual-time in his units.
3726 * By using 'se' instead of 'curr' we penalize light tasks, so
3727 * they get preempted easier. That is, if 'se' < 'curr' then
3728 * the resulting gran will be larger, therefore penalizing the
3729 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3730 * be smaller, again penalizing the lighter task.
3732 * This is especially important for buddies when the leftmost
3733 * task is higher priority than the buddy.
3735 return calc_delta_fair(gran, se);
3739 * Should 'se' preempt 'curr'.
3753 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3755 s64 gran, vdiff = curr->vruntime - se->vruntime;
3760 gran = wakeup_gran(curr, se);
3767 static void set_last_buddy(struct sched_entity *se)
3769 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3772 for_each_sched_entity(se)
3773 cfs_rq_of(se)->last = se;
3776 static void set_next_buddy(struct sched_entity *se)
3778 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3781 for_each_sched_entity(se)
3782 cfs_rq_of(se)->next = se;
3785 static void set_skip_buddy(struct sched_entity *se)
3787 for_each_sched_entity(se)
3788 cfs_rq_of(se)->skip = se;
3792 * Preempt the current task with a newly woken task if needed:
3794 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3796 struct task_struct *curr = rq->curr;
3797 struct sched_entity *se = &curr->se, *pse = &p->se;
3798 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3799 int scale = cfs_rq->nr_running >= sched_nr_latency;
3800 int next_buddy_marked = 0;
3802 if (unlikely(se == pse))
3806 * This is possible from callers such as move_task(), in which we
3807 * unconditionally check_prempt_curr() after an enqueue (which may have
3808 * lead to a throttle). This both saves work and prevents false
3809 * next-buddy nomination below.
3811 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3814 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3815 set_next_buddy(pse);
3816 next_buddy_marked = 1;
3820 * We can come here with TIF_NEED_RESCHED already set from new task
3823 * Note: this also catches the edge-case of curr being in a throttled
3824 * group (e.g. via set_curr_task), since update_curr() (in the
3825 * enqueue of curr) will have resulted in resched being set. This
3826 * prevents us from potentially nominating it as a false LAST_BUDDY
3829 if (test_tsk_need_resched(curr))
3832 /* Idle tasks are by definition preempted by non-idle tasks. */
3833 if (unlikely(curr->policy == SCHED_IDLE) &&
3834 likely(p->policy != SCHED_IDLE))
3838 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3839 * is driven by the tick):
3841 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3844 find_matching_se(&se, &pse);
3845 update_curr(cfs_rq_of(se));
3847 if (wakeup_preempt_entity(se, pse) == 1) {
3849 * Bias pick_next to pick the sched entity that is
3850 * triggering this preemption.
3852 if (!next_buddy_marked)
3853 set_next_buddy(pse);
3862 * Only set the backward buddy when the current task is still
3863 * on the rq. This can happen when a wakeup gets interleaved
3864 * with schedule on the ->pre_schedule() or idle_balance()
3865 * point, either of which can * drop the rq lock.
3867 * Also, during early boot the idle thread is in the fair class,
3868 * for obvious reasons its a bad idea to schedule back to it.
3870 if (unlikely(!se->on_rq || curr == rq->idle))
3873 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3877 static struct task_struct *pick_next_task_fair(struct rq *rq)
3879 struct task_struct *p;
3880 struct cfs_rq *cfs_rq = &rq->cfs;
3881 struct sched_entity *se;
3883 if (!cfs_rq->nr_running)
3887 se = pick_next_entity(cfs_rq);
3888 set_next_entity(cfs_rq, se);
3889 cfs_rq = group_cfs_rq(se);
3893 if (hrtick_enabled(rq))
3894 hrtick_start_fair(rq, p);
3900 * Account for a descheduled task:
3902 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3904 struct sched_entity *se = &prev->se;
3905 struct cfs_rq *cfs_rq;
3907 for_each_sched_entity(se) {
3908 cfs_rq = cfs_rq_of(se);
3909 put_prev_entity(cfs_rq, se);
3914 * sched_yield() is very simple
3916 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3918 static void yield_task_fair(struct rq *rq)
3920 struct task_struct *curr = rq->curr;
3921 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3922 struct sched_entity *se = &curr->se;
3925 * Are we the only task in the tree?
3927 if (unlikely(rq->nr_running == 1))
3930 clear_buddies(cfs_rq, se);
3932 if (curr->policy != SCHED_BATCH) {
3933 update_rq_clock(rq);
3935 * Update run-time statistics of the 'current'.
3937 update_curr(cfs_rq);
3939 * Tell update_rq_clock() that we've just updated,
3940 * so we don't do microscopic update in schedule()
3941 * and double the fastpath cost.
3943 rq->skip_clock_update = 1;
3949 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3951 struct sched_entity *se = &p->se;
3953 /* throttled hierarchies are not runnable */
3954 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3957 /* Tell the scheduler that we'd really like pse to run next. */
3960 yield_task_fair(rq);
3966 /**************************************************
3967 * Fair scheduling class load-balancing methods.
3971 * The purpose of load-balancing is to achieve the same basic fairness the
3972 * per-cpu scheduler provides, namely provide a proportional amount of compute
3973 * time to each task. This is expressed in the following equation:
3975 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3977 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3978 * W_i,0 is defined as:
3980 * W_i,0 = \Sum_j w_i,j (2)
3982 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3983 * is derived from the nice value as per prio_to_weight[].
3985 * The weight average is an exponential decay average of the instantaneous
3988 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3990 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3991 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3992 * can also include other factors [XXX].
3994 * To achieve this balance we define a measure of imbalance which follows
3995 * directly from (1):
3997 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3999 * We them move tasks around to minimize the imbalance. In the continuous
4000 * function space it is obvious this converges, in the discrete case we get
4001 * a few fun cases generally called infeasible weight scenarios.
4004 * - infeasible weights;
4005 * - local vs global optima in the discrete case. ]
4010 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4011 * for all i,j solution, we create a tree of cpus that follows the hardware
4012 * topology where each level pairs two lower groups (or better). This results
4013 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4014 * tree to only the first of the previous level and we decrease the frequency
4015 * of load-balance at each level inv. proportional to the number of cpus in
4021 * \Sum { --- * --- * 2^i } = O(n) (5)
4023 * `- size of each group
4024 * | | `- number of cpus doing load-balance
4026 * `- sum over all levels
4028 * Coupled with a limit on how many tasks we can migrate every balance pass,
4029 * this makes (5) the runtime complexity of the balancer.
4031 * An important property here is that each CPU is still (indirectly) connected
4032 * to every other cpu in at most O(log n) steps:
4034 * The adjacency matrix of the resulting graph is given by:
4037 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4040 * And you'll find that:
4042 * A^(log_2 n)_i,j != 0 for all i,j (7)
4044 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4045 * The task movement gives a factor of O(m), giving a convergence complexity
4048 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4053 * In order to avoid CPUs going idle while there's still work to do, new idle
4054 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4055 * tree itself instead of relying on other CPUs to bring it work.
4057 * This adds some complexity to both (5) and (8) but it reduces the total idle
4065 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4068 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4073 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4075 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4077 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4080 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4081 * rewrite all of this once again.]
4084 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4086 #define LBF_ALL_PINNED 0x01
4087 #define LBF_NEED_BREAK 0x02
4088 #define LBF_DST_PINNED 0x04
4089 #define LBF_SOME_PINNED 0x08
4092 struct sched_domain *sd;
4100 struct cpumask *dst_grpmask;
4102 enum cpu_idle_type idle;
4104 /* The set of CPUs under consideration for load-balancing */
4105 struct cpumask *cpus;
4110 unsigned int loop_break;
4111 unsigned int loop_max;
4115 * move_task - move a task from one runqueue to another runqueue.
4116 * Both runqueues must be locked.
4118 static void move_task(struct task_struct *p, struct lb_env *env)
4120 deactivate_task(env->src_rq, p, 0);
4121 set_task_cpu(p, env->dst_cpu);
4122 activate_task(env->dst_rq, p, 0);
4123 check_preempt_curr(env->dst_rq, p, 0);
4124 #ifdef CONFIG_NUMA_BALANCING
4125 if (p->numa_preferred_nid != -1) {
4126 int src_nid = cpu_to_node(env->src_cpu);
4127 int dst_nid = cpu_to_node(env->dst_cpu);
4130 * If the load balancer has moved the task then limit
4131 * migrations from taking place in the short term in
4132 * case this is a short-lived migration.
4134 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4135 p->numa_migrate_seq = 0;
4141 * Is this task likely cache-hot:
4144 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4148 if (p->sched_class != &fair_sched_class)
4151 if (unlikely(p->policy == SCHED_IDLE))
4155 * Buddy candidates are cache hot:
4157 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4158 (&p->se == cfs_rq_of(&p->se)->next ||
4159 &p->se == cfs_rq_of(&p->se)->last))
4162 if (sysctl_sched_migration_cost == -1)
4164 if (sysctl_sched_migration_cost == 0)
4167 delta = now - p->se.exec_start;
4169 return delta < (s64)sysctl_sched_migration_cost;
4172 #ifdef CONFIG_NUMA_BALANCING
4173 /* Returns true if the destination node has incurred more faults */
4174 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4176 int src_nid, dst_nid;
4178 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4179 !(env->sd->flags & SD_NUMA)) {
4183 src_nid = cpu_to_node(env->src_cpu);
4184 dst_nid = cpu_to_node(env->dst_cpu);
4186 if (src_nid == dst_nid ||
4187 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4190 if (dst_nid == p->numa_preferred_nid ||
4191 task_faults(p, dst_nid) > task_faults(p, src_nid))
4198 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4200 int src_nid, dst_nid;
4202 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4205 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4208 src_nid = cpu_to_node(env->src_cpu);
4209 dst_nid = cpu_to_node(env->dst_cpu);
4211 if (src_nid == dst_nid ||
4212 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4215 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4222 static inline bool migrate_improves_locality(struct task_struct *p,
4228 static inline bool migrate_degrades_locality(struct task_struct *p,
4236 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4239 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4241 int tsk_cache_hot = 0;
4243 * We do not migrate tasks that are:
4244 * 1) throttled_lb_pair, or
4245 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4246 * 3) running (obviously), or
4247 * 4) are cache-hot on their current CPU.
4249 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4252 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4255 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4257 env->flags |= LBF_SOME_PINNED;
4260 * Remember if this task can be migrated to any other cpu in
4261 * our sched_group. We may want to revisit it if we couldn't
4262 * meet load balance goals by pulling other tasks on src_cpu.
4264 * Also avoid computing new_dst_cpu if we have already computed
4265 * one in current iteration.
4267 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4270 /* Prevent to re-select dst_cpu via env's cpus */
4271 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4272 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4273 env->flags |= LBF_DST_PINNED;
4274 env->new_dst_cpu = cpu;
4282 /* Record that we found atleast one task that could run on dst_cpu */
4283 env->flags &= ~LBF_ALL_PINNED;
4285 if (task_running(env->src_rq, p)) {
4286 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4291 * Aggressive migration if:
4292 * 1) destination numa is preferred
4293 * 2) task is cache cold, or
4294 * 3) too many balance attempts have failed.
4296 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4298 tsk_cache_hot = migrate_degrades_locality(p, env);
4300 if (migrate_improves_locality(p, env)) {
4301 #ifdef CONFIG_SCHEDSTATS
4302 if (tsk_cache_hot) {
4303 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4304 schedstat_inc(p, se.statistics.nr_forced_migrations);
4310 if (!tsk_cache_hot ||
4311 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4313 if (tsk_cache_hot) {
4314 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4315 schedstat_inc(p, se.statistics.nr_forced_migrations);
4321 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4326 * move_one_task tries to move exactly one task from busiest to this_rq, as
4327 * part of active balancing operations within "domain".
4328 * Returns 1 if successful and 0 otherwise.
4330 * Called with both runqueues locked.
4332 static int move_one_task(struct lb_env *env)
4334 struct task_struct *p, *n;
4336 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4337 if (!can_migrate_task(p, env))
4342 * Right now, this is only the second place move_task()
4343 * is called, so we can safely collect move_task()
4344 * stats here rather than inside move_task().
4346 schedstat_inc(env->sd, lb_gained[env->idle]);
4352 static unsigned long task_h_load(struct task_struct *p);
4354 static const unsigned int sched_nr_migrate_break = 32;
4357 * move_tasks tries to move up to imbalance weighted load from busiest to
4358 * this_rq, as part of a balancing operation within domain "sd".
4359 * Returns 1 if successful and 0 otherwise.
4361 * Called with both runqueues locked.
4363 static int move_tasks(struct lb_env *env)
4365 struct list_head *tasks = &env->src_rq->cfs_tasks;
4366 struct task_struct *p;
4370 if (env->imbalance <= 0)
4373 while (!list_empty(tasks)) {
4374 p = list_first_entry(tasks, struct task_struct, se.group_node);
4377 /* We've more or less seen every task there is, call it quits */
4378 if (env->loop > env->loop_max)
4381 /* take a breather every nr_migrate tasks */
4382 if (env->loop > env->loop_break) {
4383 env->loop_break += sched_nr_migrate_break;
4384 env->flags |= LBF_NEED_BREAK;
4388 if (!can_migrate_task(p, env))
4391 load = task_h_load(p);
4393 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4396 if ((load / 2) > env->imbalance)
4401 env->imbalance -= load;
4403 #ifdef CONFIG_PREEMPT
4405 * NEWIDLE balancing is a source of latency, so preemptible
4406 * kernels will stop after the first task is pulled to minimize
4407 * the critical section.
4409 if (env->idle == CPU_NEWLY_IDLE)
4414 * We only want to steal up to the prescribed amount of
4417 if (env->imbalance <= 0)
4422 list_move_tail(&p->se.group_node, tasks);
4426 * Right now, this is one of only two places move_task() is called,
4427 * so we can safely collect move_task() stats here rather than
4428 * inside move_task().
4430 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4435 #ifdef CONFIG_FAIR_GROUP_SCHED
4437 * update tg->load_weight by folding this cpu's load_avg
4439 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4441 struct sched_entity *se = tg->se[cpu];
4442 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4444 /* throttled entities do not contribute to load */
4445 if (throttled_hierarchy(cfs_rq))
4448 update_cfs_rq_blocked_load(cfs_rq, 1);
4451 update_entity_load_avg(se, 1);
4453 * We pivot on our runnable average having decayed to zero for
4454 * list removal. This generally implies that all our children
4455 * have also been removed (modulo rounding error or bandwidth
4456 * control); however, such cases are rare and we can fix these
4459 * TODO: fix up out-of-order children on enqueue.
4461 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4462 list_del_leaf_cfs_rq(cfs_rq);
4464 struct rq *rq = rq_of(cfs_rq);
4465 update_rq_runnable_avg(rq, rq->nr_running);
4469 static void update_blocked_averages(int cpu)
4471 struct rq *rq = cpu_rq(cpu);
4472 struct cfs_rq *cfs_rq;
4473 unsigned long flags;
4475 raw_spin_lock_irqsave(&rq->lock, flags);
4476 update_rq_clock(rq);
4478 * Iterates the task_group tree in a bottom up fashion, see
4479 * list_add_leaf_cfs_rq() for details.
4481 for_each_leaf_cfs_rq(rq, cfs_rq) {
4483 * Note: We may want to consider periodically releasing
4484 * rq->lock about these updates so that creating many task
4485 * groups does not result in continually extending hold time.
4487 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4490 raw_spin_unlock_irqrestore(&rq->lock, flags);
4494 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4495 * This needs to be done in a top-down fashion because the load of a child
4496 * group is a fraction of its parents load.
4498 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4500 struct rq *rq = rq_of(cfs_rq);
4501 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4502 unsigned long now = jiffies;
4505 if (cfs_rq->last_h_load_update == now)
4508 cfs_rq->h_load_next = NULL;
4509 for_each_sched_entity(se) {
4510 cfs_rq = cfs_rq_of(se);
4511 cfs_rq->h_load_next = se;
4512 if (cfs_rq->last_h_load_update == now)
4517 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4518 cfs_rq->last_h_load_update = now;
4521 while ((se = cfs_rq->h_load_next) != NULL) {
4522 load = cfs_rq->h_load;
4523 load = div64_ul(load * se->avg.load_avg_contrib,
4524 cfs_rq->runnable_load_avg + 1);
4525 cfs_rq = group_cfs_rq(se);
4526 cfs_rq->h_load = load;
4527 cfs_rq->last_h_load_update = now;
4531 static unsigned long task_h_load(struct task_struct *p)
4533 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4535 update_cfs_rq_h_load(cfs_rq);
4536 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4537 cfs_rq->runnable_load_avg + 1);
4540 static inline void update_blocked_averages(int cpu)
4544 static unsigned long task_h_load(struct task_struct *p)
4546 return p->se.avg.load_avg_contrib;
4550 /********** Helpers for find_busiest_group ************************/
4552 * sg_lb_stats - stats of a sched_group required for load_balancing
4554 struct sg_lb_stats {
4555 unsigned long avg_load; /*Avg load across the CPUs of the group */
4556 unsigned long group_load; /* Total load over the CPUs of the group */
4557 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4558 unsigned long load_per_task;
4559 unsigned long group_power;
4560 unsigned int sum_nr_running; /* Nr tasks running in the group */
4561 unsigned int group_capacity;
4562 unsigned int idle_cpus;
4563 unsigned int group_weight;
4564 int group_imb; /* Is there an imbalance in the group ? */
4565 int group_has_capacity; /* Is there extra capacity in the group? */
4569 * sd_lb_stats - Structure to store the statistics of a sched_domain
4570 * during load balancing.
4572 struct sd_lb_stats {
4573 struct sched_group *busiest; /* Busiest group in this sd */
4574 struct sched_group *local; /* Local group in this sd */
4575 unsigned long total_load; /* Total load of all groups in sd */
4576 unsigned long total_pwr; /* Total power of all groups in sd */
4577 unsigned long avg_load; /* Average load across all groups in sd */
4579 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4580 struct sg_lb_stats local_stat; /* Statistics of the local group */
4583 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4586 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4587 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4588 * We must however clear busiest_stat::avg_load because
4589 * update_sd_pick_busiest() reads this before assignment.
4591 *sds = (struct sd_lb_stats){
4603 * get_sd_load_idx - Obtain the load index for a given sched domain.
4604 * @sd: The sched_domain whose load_idx is to be obtained.
4605 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4607 * Return: The load index.
4609 static inline int get_sd_load_idx(struct sched_domain *sd,
4610 enum cpu_idle_type idle)
4616 load_idx = sd->busy_idx;
4619 case CPU_NEWLY_IDLE:
4620 load_idx = sd->newidle_idx;
4623 load_idx = sd->idle_idx;
4630 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4632 return SCHED_POWER_SCALE;
4635 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4637 return default_scale_freq_power(sd, cpu);
4640 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4642 unsigned long weight = sd->span_weight;
4643 unsigned long smt_gain = sd->smt_gain;
4650 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4652 return default_scale_smt_power(sd, cpu);
4655 static unsigned long scale_rt_power(int cpu)
4657 struct rq *rq = cpu_rq(cpu);
4658 u64 total, available, age_stamp, avg;
4661 * Since we're reading these variables without serialization make sure
4662 * we read them once before doing sanity checks on them.
4664 age_stamp = ACCESS_ONCE(rq->age_stamp);
4665 avg = ACCESS_ONCE(rq->rt_avg);
4667 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4669 if (unlikely(total < avg)) {
4670 /* Ensures that power won't end up being negative */
4673 available = total - avg;
4676 if (unlikely((s64)total < SCHED_POWER_SCALE))
4677 total = SCHED_POWER_SCALE;
4679 total >>= SCHED_POWER_SHIFT;
4681 return div_u64(available, total);
4684 static void update_cpu_power(struct sched_domain *sd, int cpu)
4686 unsigned long weight = sd->span_weight;
4687 unsigned long power = SCHED_POWER_SCALE;
4688 struct sched_group *sdg = sd->groups;
4690 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4691 if (sched_feat(ARCH_POWER))
4692 power *= arch_scale_smt_power(sd, cpu);
4694 power *= default_scale_smt_power(sd, cpu);
4696 power >>= SCHED_POWER_SHIFT;
4699 sdg->sgp->power_orig = power;
4701 if (sched_feat(ARCH_POWER))
4702 power *= arch_scale_freq_power(sd, cpu);
4704 power *= default_scale_freq_power(sd, cpu);
4706 power >>= SCHED_POWER_SHIFT;
4708 power *= scale_rt_power(cpu);
4709 power >>= SCHED_POWER_SHIFT;
4714 cpu_rq(cpu)->cpu_power = power;
4715 sdg->sgp->power = power;
4718 void update_group_power(struct sched_domain *sd, int cpu)
4720 struct sched_domain *child = sd->child;
4721 struct sched_group *group, *sdg = sd->groups;
4722 unsigned long power, power_orig;
4723 unsigned long interval;
4725 interval = msecs_to_jiffies(sd->balance_interval);
4726 interval = clamp(interval, 1UL, max_load_balance_interval);
4727 sdg->sgp->next_update = jiffies + interval;
4730 update_cpu_power(sd, cpu);
4734 power_orig = power = 0;
4736 if (child->flags & SD_OVERLAP) {
4738 * SD_OVERLAP domains cannot assume that child groups
4739 * span the current group.
4742 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4743 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4745 power_orig += sg->sgp->power_orig;
4746 power += sg->sgp->power;
4750 * !SD_OVERLAP domains can assume that child groups
4751 * span the current group.
4754 group = child->groups;
4756 power_orig += group->sgp->power_orig;
4757 power += group->sgp->power;
4758 group = group->next;
4759 } while (group != child->groups);
4762 sdg->sgp->power_orig = power_orig;
4763 sdg->sgp->power = power;
4767 * Try and fix up capacity for tiny siblings, this is needed when
4768 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4769 * which on its own isn't powerful enough.
4771 * See update_sd_pick_busiest() and check_asym_packing().
4774 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4777 * Only siblings can have significantly less than SCHED_POWER_SCALE
4779 if (!(sd->flags & SD_SHARE_CPUPOWER))
4783 * If ~90% of the cpu_power is still there, we're good.
4785 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4792 * Group imbalance indicates (and tries to solve) the problem where balancing
4793 * groups is inadequate due to tsk_cpus_allowed() constraints.
4795 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4796 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4799 * { 0 1 2 3 } { 4 5 6 7 }
4802 * If we were to balance group-wise we'd place two tasks in the first group and
4803 * two tasks in the second group. Clearly this is undesired as it will overload
4804 * cpu 3 and leave one of the cpus in the second group unused.
4806 * The current solution to this issue is detecting the skew in the first group
4807 * by noticing the lower domain failed to reach balance and had difficulty
4808 * moving tasks due to affinity constraints.
4810 * When this is so detected; this group becomes a candidate for busiest; see
4811 * update_sd_pick_busiest(). And calculcate_imbalance() and
4812 * find_busiest_group() avoid some of the usual balance conditions to allow it
4813 * to create an effective group imbalance.
4815 * This is a somewhat tricky proposition since the next run might not find the
4816 * group imbalance and decide the groups need to be balanced again. A most
4817 * subtle and fragile situation.
4820 static inline int sg_imbalanced(struct sched_group *group)
4822 return group->sgp->imbalance;
4826 * Compute the group capacity.
4828 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4829 * first dividing out the smt factor and computing the actual number of cores
4830 * and limit power unit capacity with that.
4832 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4834 unsigned int capacity, smt, cpus;
4835 unsigned int power, power_orig;
4837 power = group->sgp->power;
4838 power_orig = group->sgp->power_orig;
4839 cpus = group->group_weight;
4841 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4842 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4843 capacity = cpus / smt; /* cores */
4845 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4847 capacity = fix_small_capacity(env->sd, group);
4853 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4854 * @env: The load balancing environment.
4855 * @group: sched_group whose statistics are to be updated.
4856 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4857 * @local_group: Does group contain this_cpu.
4858 * @sgs: variable to hold the statistics for this group.
4860 static inline void update_sg_lb_stats(struct lb_env *env,
4861 struct sched_group *group, int load_idx,
4862 int local_group, struct sg_lb_stats *sgs)
4864 unsigned long nr_running;
4868 memset(sgs, 0, sizeof(*sgs));
4870 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4871 struct rq *rq = cpu_rq(i);
4873 nr_running = rq->nr_running;
4875 /* Bias balancing toward cpus of our domain */
4877 load = target_load(i, load_idx);
4879 load = source_load(i, load_idx);
4881 sgs->group_load += load;
4882 sgs->sum_nr_running += nr_running;
4883 sgs->sum_weighted_load += weighted_cpuload(i);
4888 /* Adjust by relative CPU power of the group */
4889 sgs->group_power = group->sgp->power;
4890 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4892 if (sgs->sum_nr_running)
4893 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4895 sgs->group_weight = group->group_weight;
4897 sgs->group_imb = sg_imbalanced(group);
4898 sgs->group_capacity = sg_capacity(env, group);
4900 if (sgs->group_capacity > sgs->sum_nr_running)
4901 sgs->group_has_capacity = 1;
4905 * update_sd_pick_busiest - return 1 on busiest group
4906 * @env: The load balancing environment.
4907 * @sds: sched_domain statistics
4908 * @sg: sched_group candidate to be checked for being the busiest
4909 * @sgs: sched_group statistics
4911 * Determine if @sg is a busier group than the previously selected
4914 * Return: %true if @sg is a busier group than the previously selected
4915 * busiest group. %false otherwise.
4917 static bool update_sd_pick_busiest(struct lb_env *env,
4918 struct sd_lb_stats *sds,
4919 struct sched_group *sg,
4920 struct sg_lb_stats *sgs)
4922 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4925 if (sgs->sum_nr_running > sgs->group_capacity)
4932 * ASYM_PACKING needs to move all the work to the lowest
4933 * numbered CPUs in the group, therefore mark all groups
4934 * higher than ourself as busy.
4936 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4937 env->dst_cpu < group_first_cpu(sg)) {
4941 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4949 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4950 * @env: The load balancing environment.
4951 * @balance: Should we balance.
4952 * @sds: variable to hold the statistics for this sched_domain.
4954 static inline void update_sd_lb_stats(struct lb_env *env,
4955 struct sd_lb_stats *sds)
4957 struct sched_domain *child = env->sd->child;
4958 struct sched_group *sg = env->sd->groups;
4959 struct sg_lb_stats tmp_sgs;
4960 int load_idx, prefer_sibling = 0;
4962 if (child && child->flags & SD_PREFER_SIBLING)
4965 load_idx = get_sd_load_idx(env->sd, env->idle);
4968 struct sg_lb_stats *sgs = &tmp_sgs;
4971 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4974 sgs = &sds->local_stat;
4976 if (env->idle != CPU_NEWLY_IDLE ||
4977 time_after_eq(jiffies, sg->sgp->next_update))
4978 update_group_power(env->sd, env->dst_cpu);
4981 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4987 * In case the child domain prefers tasks go to siblings
4988 * first, lower the sg capacity to one so that we'll try
4989 * and move all the excess tasks away. We lower the capacity
4990 * of a group only if the local group has the capacity to fit
4991 * these excess tasks, i.e. nr_running < group_capacity. The
4992 * extra check prevents the case where you always pull from the
4993 * heaviest group when it is already under-utilized (possible
4994 * with a large weight task outweighs the tasks on the system).
4996 if (prefer_sibling && sds->local &&
4997 sds->local_stat.group_has_capacity)
4998 sgs->group_capacity = min(sgs->group_capacity, 1U);
5000 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5002 sds->busiest_stat = *sgs;
5006 /* Now, start updating sd_lb_stats */
5007 sds->total_load += sgs->group_load;
5008 sds->total_pwr += sgs->group_power;
5011 } while (sg != env->sd->groups);
5015 * check_asym_packing - Check to see if the group is packed into the
5018 * This is primarily intended to used at the sibling level. Some
5019 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5020 * case of POWER7, it can move to lower SMT modes only when higher
5021 * threads are idle. When in lower SMT modes, the threads will
5022 * perform better since they share less core resources. Hence when we
5023 * have idle threads, we want them to be the higher ones.
5025 * This packing function is run on idle threads. It checks to see if
5026 * the busiest CPU in this domain (core in the P7 case) has a higher
5027 * CPU number than the packing function is being run on. Here we are
5028 * assuming lower CPU number will be equivalent to lower a SMT thread
5031 * Return: 1 when packing is required and a task should be moved to
5032 * this CPU. The amount of the imbalance is returned in *imbalance.
5034 * @env: The load balancing environment.
5035 * @sds: Statistics of the sched_domain which is to be packed
5037 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5041 if (!(env->sd->flags & SD_ASYM_PACKING))
5047 busiest_cpu = group_first_cpu(sds->busiest);
5048 if (env->dst_cpu > busiest_cpu)
5051 env->imbalance = DIV_ROUND_CLOSEST(
5052 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5059 * fix_small_imbalance - Calculate the minor imbalance that exists
5060 * amongst the groups of a sched_domain, during
5062 * @env: The load balancing environment.
5063 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5066 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5068 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5069 unsigned int imbn = 2;
5070 unsigned long scaled_busy_load_per_task;
5071 struct sg_lb_stats *local, *busiest;
5073 local = &sds->local_stat;
5074 busiest = &sds->busiest_stat;
5076 if (!local->sum_nr_running)
5077 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5078 else if (busiest->load_per_task > local->load_per_task)
5081 scaled_busy_load_per_task =
5082 (busiest->load_per_task * SCHED_POWER_SCALE) /
5083 busiest->group_power;
5085 if (busiest->avg_load + scaled_busy_load_per_task >=
5086 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5087 env->imbalance = busiest->load_per_task;
5092 * OK, we don't have enough imbalance to justify moving tasks,
5093 * however we may be able to increase total CPU power used by
5097 pwr_now += busiest->group_power *
5098 min(busiest->load_per_task, busiest->avg_load);
5099 pwr_now += local->group_power *
5100 min(local->load_per_task, local->avg_load);
5101 pwr_now /= SCHED_POWER_SCALE;
5103 /* Amount of load we'd subtract */
5104 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5105 busiest->group_power;
5106 if (busiest->avg_load > tmp) {
5107 pwr_move += busiest->group_power *
5108 min(busiest->load_per_task,
5109 busiest->avg_load - tmp);
5112 /* Amount of load we'd add */
5113 if (busiest->avg_load * busiest->group_power <
5114 busiest->load_per_task * SCHED_POWER_SCALE) {
5115 tmp = (busiest->avg_load * busiest->group_power) /
5118 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5121 pwr_move += local->group_power *
5122 min(local->load_per_task, local->avg_load + tmp);
5123 pwr_move /= SCHED_POWER_SCALE;
5125 /* Move if we gain throughput */
5126 if (pwr_move > pwr_now)
5127 env->imbalance = busiest->load_per_task;
5131 * calculate_imbalance - Calculate the amount of imbalance present within the
5132 * groups of a given sched_domain during load balance.
5133 * @env: load balance environment
5134 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5136 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5138 unsigned long max_pull, load_above_capacity = ~0UL;
5139 struct sg_lb_stats *local, *busiest;
5141 local = &sds->local_stat;
5142 busiest = &sds->busiest_stat;
5144 if (busiest->group_imb) {
5146 * In the group_imb case we cannot rely on group-wide averages
5147 * to ensure cpu-load equilibrium, look at wider averages. XXX
5149 busiest->load_per_task =
5150 min(busiest->load_per_task, sds->avg_load);
5154 * In the presence of smp nice balancing, certain scenarios can have
5155 * max load less than avg load(as we skip the groups at or below
5156 * its cpu_power, while calculating max_load..)
5158 if (busiest->avg_load <= sds->avg_load ||
5159 local->avg_load >= sds->avg_load) {
5161 return fix_small_imbalance(env, sds);
5164 if (!busiest->group_imb) {
5166 * Don't want to pull so many tasks that a group would go idle.
5167 * Except of course for the group_imb case, since then we might
5168 * have to drop below capacity to reach cpu-load equilibrium.
5170 load_above_capacity =
5171 (busiest->sum_nr_running - busiest->group_capacity);
5173 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5174 load_above_capacity /= busiest->group_power;
5178 * We're trying to get all the cpus to the average_load, so we don't
5179 * want to push ourselves above the average load, nor do we wish to
5180 * reduce the max loaded cpu below the average load. At the same time,
5181 * we also don't want to reduce the group load below the group capacity
5182 * (so that we can implement power-savings policies etc). Thus we look
5183 * for the minimum possible imbalance.
5185 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5187 /* How much load to actually move to equalise the imbalance */
5188 env->imbalance = min(
5189 max_pull * busiest->group_power,
5190 (sds->avg_load - local->avg_load) * local->group_power
5191 ) / SCHED_POWER_SCALE;
5194 * if *imbalance is less than the average load per runnable task
5195 * there is no guarantee that any tasks will be moved so we'll have
5196 * a think about bumping its value to force at least one task to be
5199 if (env->imbalance < busiest->load_per_task)
5200 return fix_small_imbalance(env, sds);
5203 /******* find_busiest_group() helpers end here *********************/
5206 * find_busiest_group - Returns the busiest group within the sched_domain
5207 * if there is an imbalance. If there isn't an imbalance, and
5208 * the user has opted for power-savings, it returns a group whose
5209 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5210 * such a group exists.
5212 * Also calculates the amount of weighted load which should be moved
5213 * to restore balance.
5215 * @env: The load balancing environment.
5217 * Return: - The busiest group if imbalance exists.
5218 * - If no imbalance and user has opted for power-savings balance,
5219 * return the least loaded group whose CPUs can be
5220 * put to idle by rebalancing its tasks onto our group.
5222 static struct sched_group *find_busiest_group(struct lb_env *env)
5224 struct sg_lb_stats *local, *busiest;
5225 struct sd_lb_stats sds;
5227 init_sd_lb_stats(&sds);
5230 * Compute the various statistics relavent for load balancing at
5233 update_sd_lb_stats(env, &sds);
5234 local = &sds.local_stat;
5235 busiest = &sds.busiest_stat;
5237 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5238 check_asym_packing(env, &sds))
5241 /* There is no busy sibling group to pull tasks from */
5242 if (!sds.busiest || busiest->sum_nr_running == 0)
5245 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5248 * If the busiest group is imbalanced the below checks don't
5249 * work because they assume all things are equal, which typically
5250 * isn't true due to cpus_allowed constraints and the like.
5252 if (busiest->group_imb)
5255 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5256 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5257 !busiest->group_has_capacity)
5261 * If the local group is more busy than the selected busiest group
5262 * don't try and pull any tasks.
5264 if (local->avg_load >= busiest->avg_load)
5268 * Don't pull any tasks if this group is already above the domain
5271 if (local->avg_load >= sds.avg_load)
5274 if (env->idle == CPU_IDLE) {
5276 * This cpu is idle. If the busiest group load doesn't
5277 * have more tasks than the number of available cpu's and
5278 * there is no imbalance between this and busiest group
5279 * wrt to idle cpu's, it is balanced.
5281 if ((local->idle_cpus < busiest->idle_cpus) &&
5282 busiest->sum_nr_running <= busiest->group_weight)
5286 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5287 * imbalance_pct to be conservative.
5289 if (100 * busiest->avg_load <=
5290 env->sd->imbalance_pct * local->avg_load)
5295 /* Looks like there is an imbalance. Compute it */
5296 calculate_imbalance(env, &sds);
5305 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5307 static struct rq *find_busiest_queue(struct lb_env *env,
5308 struct sched_group *group)
5310 struct rq *busiest = NULL, *rq;
5311 unsigned long busiest_load = 0, busiest_power = 1;
5314 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5315 unsigned long power = power_of(i);
5316 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5321 capacity = fix_small_capacity(env->sd, group);
5324 wl = weighted_cpuload(i);
5327 * When comparing with imbalance, use weighted_cpuload()
5328 * which is not scaled with the cpu power.
5330 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5334 * For the load comparisons with the other cpu's, consider
5335 * the weighted_cpuload() scaled with the cpu power, so that
5336 * the load can be moved away from the cpu that is potentially
5337 * running at a lower capacity.
5339 * Thus we're looking for max(wl_i / power_i), crosswise
5340 * multiplication to rid ourselves of the division works out
5341 * to: wl_i * power_j > wl_j * power_i; where j is our
5344 if (wl * busiest_power > busiest_load * power) {
5346 busiest_power = power;
5355 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5356 * so long as it is large enough.
5358 #define MAX_PINNED_INTERVAL 512
5360 /* Working cpumask for load_balance and load_balance_newidle. */
5361 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5363 static int need_active_balance(struct lb_env *env)
5365 struct sched_domain *sd = env->sd;
5367 if (env->idle == CPU_NEWLY_IDLE) {
5370 * ASYM_PACKING needs to force migrate tasks from busy but
5371 * higher numbered CPUs in order to pack all tasks in the
5372 * lowest numbered CPUs.
5374 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5378 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5381 static int active_load_balance_cpu_stop(void *data);
5383 static int should_we_balance(struct lb_env *env)
5385 struct sched_group *sg = env->sd->groups;
5386 struct cpumask *sg_cpus, *sg_mask;
5387 int cpu, balance_cpu = -1;
5390 * In the newly idle case, we will allow all the cpu's
5391 * to do the newly idle load balance.
5393 if (env->idle == CPU_NEWLY_IDLE)
5396 sg_cpus = sched_group_cpus(sg);
5397 sg_mask = sched_group_mask(sg);
5398 /* Try to find first idle cpu */
5399 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5400 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5407 if (balance_cpu == -1)
5408 balance_cpu = group_balance_cpu(sg);
5411 * First idle cpu or the first cpu(busiest) in this sched group
5412 * is eligible for doing load balancing at this and above domains.
5414 return balance_cpu == env->dst_cpu;
5418 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5419 * tasks if there is an imbalance.
5421 static int load_balance(int this_cpu, struct rq *this_rq,
5422 struct sched_domain *sd, enum cpu_idle_type idle,
5423 int *continue_balancing)
5425 int ld_moved, cur_ld_moved, active_balance = 0;
5426 struct sched_domain *sd_parent = sd->parent;
5427 struct sched_group *group;
5429 unsigned long flags;
5430 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5432 struct lb_env env = {
5434 .dst_cpu = this_cpu,
5436 .dst_grpmask = sched_group_cpus(sd->groups),
5438 .loop_break = sched_nr_migrate_break,
5443 * For NEWLY_IDLE load_balancing, we don't need to consider
5444 * other cpus in our group
5446 if (idle == CPU_NEWLY_IDLE)
5447 env.dst_grpmask = NULL;
5449 cpumask_copy(cpus, cpu_active_mask);
5451 schedstat_inc(sd, lb_count[idle]);
5454 if (!should_we_balance(&env)) {
5455 *continue_balancing = 0;
5459 group = find_busiest_group(&env);
5461 schedstat_inc(sd, lb_nobusyg[idle]);
5465 busiest = find_busiest_queue(&env, group);
5467 schedstat_inc(sd, lb_nobusyq[idle]);
5471 BUG_ON(busiest == env.dst_rq);
5473 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5476 if (busiest->nr_running > 1) {
5478 * Attempt to move tasks. If find_busiest_group has found
5479 * an imbalance but busiest->nr_running <= 1, the group is
5480 * still unbalanced. ld_moved simply stays zero, so it is
5481 * correctly treated as an imbalance.
5483 env.flags |= LBF_ALL_PINNED;
5484 env.src_cpu = busiest->cpu;
5485 env.src_rq = busiest;
5486 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5489 local_irq_save(flags);
5490 double_rq_lock(env.dst_rq, busiest);
5493 * cur_ld_moved - load moved in current iteration
5494 * ld_moved - cumulative load moved across iterations
5496 cur_ld_moved = move_tasks(&env);
5497 ld_moved += cur_ld_moved;
5498 double_rq_unlock(env.dst_rq, busiest);
5499 local_irq_restore(flags);
5502 * some other cpu did the load balance for us.
5504 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5505 resched_cpu(env.dst_cpu);
5507 if (env.flags & LBF_NEED_BREAK) {
5508 env.flags &= ~LBF_NEED_BREAK;
5513 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5514 * us and move them to an alternate dst_cpu in our sched_group
5515 * where they can run. The upper limit on how many times we
5516 * iterate on same src_cpu is dependent on number of cpus in our
5519 * This changes load balance semantics a bit on who can move
5520 * load to a given_cpu. In addition to the given_cpu itself
5521 * (or a ilb_cpu acting on its behalf where given_cpu is
5522 * nohz-idle), we now have balance_cpu in a position to move
5523 * load to given_cpu. In rare situations, this may cause
5524 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5525 * _independently_ and at _same_ time to move some load to
5526 * given_cpu) causing exceess load to be moved to given_cpu.
5527 * This however should not happen so much in practice and
5528 * moreover subsequent load balance cycles should correct the
5529 * excess load moved.
5531 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5533 /* Prevent to re-select dst_cpu via env's cpus */
5534 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5536 env.dst_rq = cpu_rq(env.new_dst_cpu);
5537 env.dst_cpu = env.new_dst_cpu;
5538 env.flags &= ~LBF_DST_PINNED;
5540 env.loop_break = sched_nr_migrate_break;
5543 * Go back to "more_balance" rather than "redo" since we
5544 * need to continue with same src_cpu.
5550 * We failed to reach balance because of affinity.
5553 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5555 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5556 *group_imbalance = 1;
5557 } else if (*group_imbalance)
5558 *group_imbalance = 0;
5561 /* All tasks on this runqueue were pinned by CPU affinity */
5562 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5563 cpumask_clear_cpu(cpu_of(busiest), cpus);
5564 if (!cpumask_empty(cpus)) {
5566 env.loop_break = sched_nr_migrate_break;
5574 schedstat_inc(sd, lb_failed[idle]);
5576 * Increment the failure counter only on periodic balance.
5577 * We do not want newidle balance, which can be very
5578 * frequent, pollute the failure counter causing
5579 * excessive cache_hot migrations and active balances.
5581 if (idle != CPU_NEWLY_IDLE)
5582 sd->nr_balance_failed++;
5584 if (need_active_balance(&env)) {
5585 raw_spin_lock_irqsave(&busiest->lock, flags);
5587 /* don't kick the active_load_balance_cpu_stop,
5588 * if the curr task on busiest cpu can't be
5591 if (!cpumask_test_cpu(this_cpu,
5592 tsk_cpus_allowed(busiest->curr))) {
5593 raw_spin_unlock_irqrestore(&busiest->lock,
5595 env.flags |= LBF_ALL_PINNED;
5596 goto out_one_pinned;
5600 * ->active_balance synchronizes accesses to
5601 * ->active_balance_work. Once set, it's cleared
5602 * only after active load balance is finished.
5604 if (!busiest->active_balance) {
5605 busiest->active_balance = 1;
5606 busiest->push_cpu = this_cpu;
5609 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5611 if (active_balance) {
5612 stop_one_cpu_nowait(cpu_of(busiest),
5613 active_load_balance_cpu_stop, busiest,
5614 &busiest->active_balance_work);
5618 * We've kicked active balancing, reset the failure
5621 sd->nr_balance_failed = sd->cache_nice_tries+1;
5624 sd->nr_balance_failed = 0;
5626 if (likely(!active_balance)) {
5627 /* We were unbalanced, so reset the balancing interval */
5628 sd->balance_interval = sd->min_interval;
5631 * If we've begun active balancing, start to back off. This
5632 * case may not be covered by the all_pinned logic if there
5633 * is only 1 task on the busy runqueue (because we don't call
5636 if (sd->balance_interval < sd->max_interval)
5637 sd->balance_interval *= 2;
5643 schedstat_inc(sd, lb_balanced[idle]);
5645 sd->nr_balance_failed = 0;
5648 /* tune up the balancing interval */
5649 if (((env.flags & LBF_ALL_PINNED) &&
5650 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5651 (sd->balance_interval < sd->max_interval))
5652 sd->balance_interval *= 2;
5660 * idle_balance is called by schedule() if this_cpu is about to become
5661 * idle. Attempts to pull tasks from other CPUs.
5663 void idle_balance(int this_cpu, struct rq *this_rq)
5665 struct sched_domain *sd;
5666 int pulled_task = 0;
5667 unsigned long next_balance = jiffies + HZ;
5670 this_rq->idle_stamp = rq_clock(this_rq);
5672 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5676 * Drop the rq->lock, but keep IRQ/preempt disabled.
5678 raw_spin_unlock(&this_rq->lock);
5680 update_blocked_averages(this_cpu);
5682 for_each_domain(this_cpu, sd) {
5683 unsigned long interval;
5684 int continue_balancing = 1;
5685 u64 t0, domain_cost;
5687 if (!(sd->flags & SD_LOAD_BALANCE))
5690 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5693 if (sd->flags & SD_BALANCE_NEWIDLE) {
5694 t0 = sched_clock_cpu(this_cpu);
5696 /* If we've pulled tasks over stop searching: */
5697 pulled_task = load_balance(this_cpu, this_rq,
5699 &continue_balancing);
5701 domain_cost = sched_clock_cpu(this_cpu) - t0;
5702 if (domain_cost > sd->max_newidle_lb_cost)
5703 sd->max_newidle_lb_cost = domain_cost;
5705 curr_cost += domain_cost;
5708 interval = msecs_to_jiffies(sd->balance_interval);
5709 if (time_after(next_balance, sd->last_balance + interval))
5710 next_balance = sd->last_balance + interval;
5712 this_rq->idle_stamp = 0;
5718 raw_spin_lock(&this_rq->lock);
5720 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5722 * We are going idle. next_balance may be set based on
5723 * a busy processor. So reset next_balance.
5725 this_rq->next_balance = next_balance;
5728 if (curr_cost > this_rq->max_idle_balance_cost)
5729 this_rq->max_idle_balance_cost = curr_cost;
5733 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5734 * running tasks off the busiest CPU onto idle CPUs. It requires at
5735 * least 1 task to be running on each physical CPU where possible, and
5736 * avoids physical / logical imbalances.
5738 static int active_load_balance_cpu_stop(void *data)
5740 struct rq *busiest_rq = data;
5741 int busiest_cpu = cpu_of(busiest_rq);
5742 int target_cpu = busiest_rq->push_cpu;
5743 struct rq *target_rq = cpu_rq(target_cpu);
5744 struct sched_domain *sd;
5746 raw_spin_lock_irq(&busiest_rq->lock);
5748 /* make sure the requested cpu hasn't gone down in the meantime */
5749 if (unlikely(busiest_cpu != smp_processor_id() ||
5750 !busiest_rq->active_balance))
5753 /* Is there any task to move? */
5754 if (busiest_rq->nr_running <= 1)
5758 * This condition is "impossible", if it occurs
5759 * we need to fix it. Originally reported by
5760 * Bjorn Helgaas on a 128-cpu setup.
5762 BUG_ON(busiest_rq == target_rq);
5764 /* move a task from busiest_rq to target_rq */
5765 double_lock_balance(busiest_rq, target_rq);
5767 /* Search for an sd spanning us and the target CPU. */
5769 for_each_domain(target_cpu, sd) {
5770 if ((sd->flags & SD_LOAD_BALANCE) &&
5771 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5776 struct lb_env env = {
5778 .dst_cpu = target_cpu,
5779 .dst_rq = target_rq,
5780 .src_cpu = busiest_rq->cpu,
5781 .src_rq = busiest_rq,
5785 schedstat_inc(sd, alb_count);
5787 if (move_one_task(&env))
5788 schedstat_inc(sd, alb_pushed);
5790 schedstat_inc(sd, alb_failed);
5793 double_unlock_balance(busiest_rq, target_rq);
5795 busiest_rq->active_balance = 0;
5796 raw_spin_unlock_irq(&busiest_rq->lock);
5800 #ifdef CONFIG_NO_HZ_COMMON
5802 * idle load balancing details
5803 * - When one of the busy CPUs notice that there may be an idle rebalancing
5804 * needed, they will kick the idle load balancer, which then does idle
5805 * load balancing for all the idle CPUs.
5808 cpumask_var_t idle_cpus_mask;
5810 unsigned long next_balance; /* in jiffy units */
5811 } nohz ____cacheline_aligned;
5813 static inline int find_new_ilb(int call_cpu)
5815 int ilb = cpumask_first(nohz.idle_cpus_mask);
5817 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5824 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5825 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5826 * CPU (if there is one).
5828 static void nohz_balancer_kick(int cpu)
5832 nohz.next_balance++;
5834 ilb_cpu = find_new_ilb(cpu);
5836 if (ilb_cpu >= nr_cpu_ids)
5839 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5842 * Use smp_send_reschedule() instead of resched_cpu().
5843 * This way we generate a sched IPI on the target cpu which
5844 * is idle. And the softirq performing nohz idle load balance
5845 * will be run before returning from the IPI.
5847 smp_send_reschedule(ilb_cpu);
5851 static inline void nohz_balance_exit_idle(int cpu)
5853 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5854 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5855 atomic_dec(&nohz.nr_cpus);
5856 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5860 static inline void set_cpu_sd_state_busy(void)
5862 struct sched_domain *sd;
5865 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5867 if (!sd || !sd->nohz_idle)
5871 for (; sd; sd = sd->parent)
5872 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5877 void set_cpu_sd_state_idle(void)
5879 struct sched_domain *sd;
5882 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5884 if (!sd || sd->nohz_idle)
5888 for (; sd; sd = sd->parent)
5889 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5895 * This routine will record that the cpu is going idle with tick stopped.
5896 * This info will be used in performing idle load balancing in the future.
5898 void nohz_balance_enter_idle(int cpu)
5901 * If this cpu is going down, then nothing needs to be done.
5903 if (!cpu_active(cpu))
5906 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5909 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5910 atomic_inc(&nohz.nr_cpus);
5911 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5914 static int sched_ilb_notifier(struct notifier_block *nfb,
5915 unsigned long action, void *hcpu)
5917 switch (action & ~CPU_TASKS_FROZEN) {
5919 nohz_balance_exit_idle(smp_processor_id());
5927 static DEFINE_SPINLOCK(balancing);
5930 * Scale the max load_balance interval with the number of CPUs in the system.
5931 * This trades load-balance latency on larger machines for less cross talk.
5933 void update_max_interval(void)
5935 max_load_balance_interval = HZ*num_online_cpus()/10;
5939 * It checks each scheduling domain to see if it is due to be balanced,
5940 * and initiates a balancing operation if so.
5942 * Balancing parameters are set up in init_sched_domains.
5944 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5946 int continue_balancing = 1;
5947 struct rq *rq = cpu_rq(cpu);
5948 unsigned long interval;
5949 struct sched_domain *sd;
5950 /* Earliest time when we have to do rebalance again */
5951 unsigned long next_balance = jiffies + 60*HZ;
5952 int update_next_balance = 0;
5953 int need_serialize, need_decay = 0;
5956 update_blocked_averages(cpu);
5959 for_each_domain(cpu, sd) {
5961 * Decay the newidle max times here because this is a regular
5962 * visit to all the domains. Decay ~1% per second.
5964 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5965 sd->max_newidle_lb_cost =
5966 (sd->max_newidle_lb_cost * 253) / 256;
5967 sd->next_decay_max_lb_cost = jiffies + HZ;
5970 max_cost += sd->max_newidle_lb_cost;
5972 if (!(sd->flags & SD_LOAD_BALANCE))
5976 * Stop the load balance at this level. There is another
5977 * CPU in our sched group which is doing load balancing more
5980 if (!continue_balancing) {
5986 interval = sd->balance_interval;
5987 if (idle != CPU_IDLE)
5988 interval *= sd->busy_factor;
5990 /* scale ms to jiffies */
5991 interval = msecs_to_jiffies(interval);
5992 interval = clamp(interval, 1UL, max_load_balance_interval);
5994 need_serialize = sd->flags & SD_SERIALIZE;
5996 if (need_serialize) {
5997 if (!spin_trylock(&balancing))
6001 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6002 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6004 * The LBF_DST_PINNED logic could have changed
6005 * env->dst_cpu, so we can't know our idle
6006 * state even if we migrated tasks. Update it.
6008 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6010 sd->last_balance = jiffies;
6013 spin_unlock(&balancing);
6015 if (time_after(next_balance, sd->last_balance + interval)) {
6016 next_balance = sd->last_balance + interval;
6017 update_next_balance = 1;
6022 * Ensure the rq-wide value also decays but keep it at a
6023 * reasonable floor to avoid funnies with rq->avg_idle.
6025 rq->max_idle_balance_cost =
6026 max((u64)sysctl_sched_migration_cost, max_cost);
6031 * next_balance will be updated only when there is a need.
6032 * When the cpu is attached to null domain for ex, it will not be
6035 if (likely(update_next_balance))
6036 rq->next_balance = next_balance;
6039 #ifdef CONFIG_NO_HZ_COMMON
6041 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6042 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6044 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6046 struct rq *this_rq = cpu_rq(this_cpu);
6050 if (idle != CPU_IDLE ||
6051 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6054 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6055 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6059 * If this cpu gets work to do, stop the load balancing
6060 * work being done for other cpus. Next load
6061 * balancing owner will pick it up.
6066 rq = cpu_rq(balance_cpu);
6068 raw_spin_lock_irq(&rq->lock);
6069 update_rq_clock(rq);
6070 update_idle_cpu_load(rq);
6071 raw_spin_unlock_irq(&rq->lock);
6073 rebalance_domains(balance_cpu, CPU_IDLE);
6075 if (time_after(this_rq->next_balance, rq->next_balance))
6076 this_rq->next_balance = rq->next_balance;
6078 nohz.next_balance = this_rq->next_balance;
6080 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6084 * Current heuristic for kicking the idle load balancer in the presence
6085 * of an idle cpu is the system.
6086 * - This rq has more than one task.
6087 * - At any scheduler domain level, this cpu's scheduler group has multiple
6088 * busy cpu's exceeding the group's power.
6089 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6090 * domain span are idle.
6092 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6094 unsigned long now = jiffies;
6095 struct sched_domain *sd;
6097 if (unlikely(idle_cpu(cpu)))
6101 * We may be recently in ticked or tickless idle mode. At the first
6102 * busy tick after returning from idle, we will update the busy stats.
6104 set_cpu_sd_state_busy();
6105 nohz_balance_exit_idle(cpu);
6108 * None are in tickless mode and hence no need for NOHZ idle load
6111 if (likely(!atomic_read(&nohz.nr_cpus)))
6114 if (time_before(now, nohz.next_balance))
6117 if (rq->nr_running >= 2)
6121 for_each_domain(cpu, sd) {
6122 struct sched_group *sg = sd->groups;
6123 struct sched_group_power *sgp = sg->sgp;
6124 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6126 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6127 goto need_kick_unlock;
6129 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6130 && (cpumask_first_and(nohz.idle_cpus_mask,
6131 sched_domain_span(sd)) < cpu))
6132 goto need_kick_unlock;
6134 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6146 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6150 * run_rebalance_domains is triggered when needed from the scheduler tick.
6151 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6153 static void run_rebalance_domains(struct softirq_action *h)
6155 int this_cpu = smp_processor_id();
6156 struct rq *this_rq = cpu_rq(this_cpu);
6157 enum cpu_idle_type idle = this_rq->idle_balance ?
6158 CPU_IDLE : CPU_NOT_IDLE;
6160 rebalance_domains(this_cpu, idle);
6163 * If this cpu has a pending nohz_balance_kick, then do the
6164 * balancing on behalf of the other idle cpus whose ticks are
6167 nohz_idle_balance(this_cpu, idle);
6170 static inline int on_null_domain(int cpu)
6172 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6176 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6178 void trigger_load_balance(struct rq *rq, int cpu)
6180 /* Don't need to rebalance while attached to NULL domain */
6181 if (time_after_eq(jiffies, rq->next_balance) &&
6182 likely(!on_null_domain(cpu)))
6183 raise_softirq(SCHED_SOFTIRQ);
6184 #ifdef CONFIG_NO_HZ_COMMON
6185 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6186 nohz_balancer_kick(cpu);
6190 static void rq_online_fair(struct rq *rq)
6195 static void rq_offline_fair(struct rq *rq)
6199 /* Ensure any throttled groups are reachable by pick_next_task */
6200 unthrottle_offline_cfs_rqs(rq);
6203 #endif /* CONFIG_SMP */
6206 * scheduler tick hitting a task of our scheduling class:
6208 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6210 struct cfs_rq *cfs_rq;
6211 struct sched_entity *se = &curr->se;
6213 for_each_sched_entity(se) {
6214 cfs_rq = cfs_rq_of(se);
6215 entity_tick(cfs_rq, se, queued);
6218 if (numabalancing_enabled)
6219 task_tick_numa(rq, curr);
6221 update_rq_runnable_avg(rq, 1);
6225 * called on fork with the child task as argument from the parent's context
6226 * - child not yet on the tasklist
6227 * - preemption disabled
6229 static void task_fork_fair(struct task_struct *p)
6231 struct cfs_rq *cfs_rq;
6232 struct sched_entity *se = &p->se, *curr;
6233 int this_cpu = smp_processor_id();
6234 struct rq *rq = this_rq();
6235 unsigned long flags;
6237 raw_spin_lock_irqsave(&rq->lock, flags);
6239 update_rq_clock(rq);
6241 cfs_rq = task_cfs_rq(current);
6242 curr = cfs_rq->curr;
6245 * Not only the cpu but also the task_group of the parent might have
6246 * been changed after parent->se.parent,cfs_rq were copied to
6247 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6248 * of child point to valid ones.
6251 __set_task_cpu(p, this_cpu);
6254 update_curr(cfs_rq);
6257 se->vruntime = curr->vruntime;
6258 place_entity(cfs_rq, se, 1);
6260 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6262 * Upon rescheduling, sched_class::put_prev_task() will place
6263 * 'current' within the tree based on its new key value.
6265 swap(curr->vruntime, se->vruntime);
6266 resched_task(rq->curr);
6269 se->vruntime -= cfs_rq->min_vruntime;
6271 raw_spin_unlock_irqrestore(&rq->lock, flags);
6275 * Priority of the task has changed. Check to see if we preempt
6279 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6285 * Reschedule if we are currently running on this runqueue and
6286 * our priority decreased, or if we are not currently running on
6287 * this runqueue and our priority is higher than the current's
6289 if (rq->curr == p) {
6290 if (p->prio > oldprio)
6291 resched_task(rq->curr);
6293 check_preempt_curr(rq, p, 0);
6296 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6298 struct sched_entity *se = &p->se;
6299 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6302 * Ensure the task's vruntime is normalized, so that when its
6303 * switched back to the fair class the enqueue_entity(.flags=0) will
6304 * do the right thing.
6306 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6307 * have normalized the vruntime, if it was !on_rq, then only when
6308 * the task is sleeping will it still have non-normalized vruntime.
6310 if (!se->on_rq && p->state != TASK_RUNNING) {
6312 * Fix up our vruntime so that the current sleep doesn't
6313 * cause 'unlimited' sleep bonus.
6315 place_entity(cfs_rq, se, 0);
6316 se->vruntime -= cfs_rq->min_vruntime;
6321 * Remove our load from contribution when we leave sched_fair
6322 * and ensure we don't carry in an old decay_count if we
6325 if (se->avg.decay_count) {
6326 __synchronize_entity_decay(se);
6327 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6333 * We switched to the sched_fair class.
6335 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6341 * We were most likely switched from sched_rt, so
6342 * kick off the schedule if running, otherwise just see
6343 * if we can still preempt the current task.
6346 resched_task(rq->curr);
6348 check_preempt_curr(rq, p, 0);
6351 /* Account for a task changing its policy or group.
6353 * This routine is mostly called to set cfs_rq->curr field when a task
6354 * migrates between groups/classes.
6356 static void set_curr_task_fair(struct rq *rq)
6358 struct sched_entity *se = &rq->curr->se;
6360 for_each_sched_entity(se) {
6361 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6363 set_next_entity(cfs_rq, se);
6364 /* ensure bandwidth has been allocated on our new cfs_rq */
6365 account_cfs_rq_runtime(cfs_rq, 0);
6369 void init_cfs_rq(struct cfs_rq *cfs_rq)
6371 cfs_rq->tasks_timeline = RB_ROOT;
6372 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6373 #ifndef CONFIG_64BIT
6374 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6377 atomic64_set(&cfs_rq->decay_counter, 1);
6378 atomic_long_set(&cfs_rq->removed_load, 0);
6382 #ifdef CONFIG_FAIR_GROUP_SCHED
6383 static void task_move_group_fair(struct task_struct *p, int on_rq)
6385 struct cfs_rq *cfs_rq;
6387 * If the task was not on the rq at the time of this cgroup movement
6388 * it must have been asleep, sleeping tasks keep their ->vruntime
6389 * absolute on their old rq until wakeup (needed for the fair sleeper
6390 * bonus in place_entity()).
6392 * If it was on the rq, we've just 'preempted' it, which does convert
6393 * ->vruntime to a relative base.
6395 * Make sure both cases convert their relative position when migrating
6396 * to another cgroup's rq. This does somewhat interfere with the
6397 * fair sleeper stuff for the first placement, but who cares.
6400 * When !on_rq, vruntime of the task has usually NOT been normalized.
6401 * But there are some cases where it has already been normalized:
6403 * - Moving a forked child which is waiting for being woken up by
6404 * wake_up_new_task().
6405 * - Moving a task which has been woken up by try_to_wake_up() and
6406 * waiting for actually being woken up by sched_ttwu_pending().
6408 * To prevent boost or penalty in the new cfs_rq caused by delta
6409 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6411 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6415 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6416 set_task_rq(p, task_cpu(p));
6418 cfs_rq = cfs_rq_of(&p->se);
6419 p->se.vruntime += cfs_rq->min_vruntime;
6422 * migrate_task_rq_fair() will have removed our previous
6423 * contribution, but we must synchronize for ongoing future
6426 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6427 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6432 void free_fair_sched_group(struct task_group *tg)
6436 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6438 for_each_possible_cpu(i) {
6440 kfree(tg->cfs_rq[i]);
6449 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6451 struct cfs_rq *cfs_rq;
6452 struct sched_entity *se;
6455 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6458 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6462 tg->shares = NICE_0_LOAD;
6464 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6466 for_each_possible_cpu(i) {
6467 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6468 GFP_KERNEL, cpu_to_node(i));
6472 se = kzalloc_node(sizeof(struct sched_entity),
6473 GFP_KERNEL, cpu_to_node(i));
6477 init_cfs_rq(cfs_rq);
6478 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6489 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6491 struct rq *rq = cpu_rq(cpu);
6492 unsigned long flags;
6495 * Only empty task groups can be destroyed; so we can speculatively
6496 * check on_list without danger of it being re-added.
6498 if (!tg->cfs_rq[cpu]->on_list)
6501 raw_spin_lock_irqsave(&rq->lock, flags);
6502 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6503 raw_spin_unlock_irqrestore(&rq->lock, flags);
6506 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6507 struct sched_entity *se, int cpu,
6508 struct sched_entity *parent)
6510 struct rq *rq = cpu_rq(cpu);
6514 init_cfs_rq_runtime(cfs_rq);
6516 tg->cfs_rq[cpu] = cfs_rq;
6519 /* se could be NULL for root_task_group */
6524 se->cfs_rq = &rq->cfs;
6526 se->cfs_rq = parent->my_q;
6529 update_load_set(&se->load, 0);
6530 se->parent = parent;
6533 static DEFINE_MUTEX(shares_mutex);
6535 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6538 unsigned long flags;
6541 * We can't change the weight of the root cgroup.
6546 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6548 mutex_lock(&shares_mutex);
6549 if (tg->shares == shares)
6552 tg->shares = shares;
6553 for_each_possible_cpu(i) {
6554 struct rq *rq = cpu_rq(i);
6555 struct sched_entity *se;
6558 /* Propagate contribution to hierarchy */
6559 raw_spin_lock_irqsave(&rq->lock, flags);
6561 /* Possible calls to update_curr() need rq clock */
6562 update_rq_clock(rq);
6563 for_each_sched_entity(se)
6564 update_cfs_shares(group_cfs_rq(se));
6565 raw_spin_unlock_irqrestore(&rq->lock, flags);
6569 mutex_unlock(&shares_mutex);
6572 #else /* CONFIG_FAIR_GROUP_SCHED */
6574 void free_fair_sched_group(struct task_group *tg) { }
6576 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6581 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6583 #endif /* CONFIG_FAIR_GROUP_SCHED */
6586 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6588 struct sched_entity *se = &task->se;
6589 unsigned int rr_interval = 0;
6592 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6595 if (rq->cfs.load.weight)
6596 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6602 * All the scheduling class methods:
6604 const struct sched_class fair_sched_class = {
6605 .next = &idle_sched_class,
6606 .enqueue_task = enqueue_task_fair,
6607 .dequeue_task = dequeue_task_fair,
6608 .yield_task = yield_task_fair,
6609 .yield_to_task = yield_to_task_fair,
6611 .check_preempt_curr = check_preempt_wakeup,
6613 .pick_next_task = pick_next_task_fair,
6614 .put_prev_task = put_prev_task_fair,
6617 .select_task_rq = select_task_rq_fair,
6618 .migrate_task_rq = migrate_task_rq_fair,
6620 .rq_online = rq_online_fair,
6621 .rq_offline = rq_offline_fair,
6623 .task_waking = task_waking_fair,
6626 .set_curr_task = set_curr_task_fair,
6627 .task_tick = task_tick_fair,
6628 .task_fork = task_fork_fair,
6630 .prio_changed = prio_changed_fair,
6631 .switched_from = switched_from_fair,
6632 .switched_to = switched_to_fair,
6634 .get_rr_interval = get_rr_interval_fair,
6636 #ifdef CONFIG_FAIR_GROUP_SCHED
6637 .task_move_group = task_move_group_fair,
6641 #ifdef CONFIG_SCHED_DEBUG
6642 void print_cfs_stats(struct seq_file *m, int cpu)
6644 struct cfs_rq *cfs_rq;
6647 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6648 print_cfs_rq(m, cpu, cfs_rq);
6653 __init void init_sched_fair_class(void)
6656 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6658 #ifdef CONFIG_NO_HZ_COMMON
6659 nohz.next_balance = jiffies;
6660 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6661 cpu_notifier(sched_ilb_notifier, 0);