2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * Approximate time to scan a full NUMA task in ms. The task scan period is
822 * calculated based on the tasks virtual memory size and
823 * numa_balancing_scan_size.
825 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
826 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
829 /* Portion of address space to scan in MB */
830 unsigned int sysctl_numa_balancing_scan_size = 256;
832 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
833 unsigned int sysctl_numa_balancing_scan_delay = 1000;
835 static unsigned int task_nr_scan_windows(struct task_struct *p)
837 unsigned long rss = 0;
838 unsigned long nr_scan_pages;
841 * Calculations based on RSS as non-present and empty pages are skipped
842 * by the PTE scanner and NUMA hinting faults should be trapped based
845 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
846 rss = get_mm_rss(p->mm);
850 rss = round_up(rss, nr_scan_pages);
851 return rss / nr_scan_pages;
854 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
855 #define MAX_SCAN_WINDOW 2560
857 static unsigned int task_scan_min(struct task_struct *p)
859 unsigned int scan, floor;
860 unsigned int windows = 1;
862 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
863 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
864 floor = 1000 / windows;
866 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
867 return max_t(unsigned int, floor, scan);
870 static unsigned int task_scan_max(struct task_struct *p)
872 unsigned int smin = task_scan_min(p);
875 /* Watch for min being lower than max due to floor calculations */
876 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
877 return max(smin, smax);
881 * Once a preferred node is selected the scheduler balancer will prefer moving
882 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
883 * scans. This will give the process the chance to accumulate more faults on
884 * the preferred node but still allow the scheduler to move the task again if
885 * the nodes CPUs are overloaded.
887 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 3;
889 static inline int task_faults_idx(int nid, int priv)
891 return 2 * nid + priv;
894 static inline unsigned long task_faults(struct task_struct *p, int nid)
899 return p->numa_faults[task_faults_idx(nid, 0)] +
900 p->numa_faults[task_faults_idx(nid, 1)];
903 static unsigned long weighted_cpuload(const int cpu);
907 find_idlest_cpu_node(int this_cpu, int nid)
909 unsigned long load, min_load = ULONG_MAX;
910 int i, idlest_cpu = this_cpu;
912 BUG_ON(cpu_to_node(this_cpu) == nid);
915 for_each_cpu(i, cpumask_of_node(nid)) {
916 load = weighted_cpuload(i);
918 if (load < min_load) {
928 static void task_numa_placement(struct task_struct *p)
930 int seq, nid, max_nid = -1;
931 unsigned long max_faults = 0;
933 if (!p->mm) /* for example, ksmd faulting in a user's mm */
935 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
936 if (p->numa_scan_seq == seq)
938 p->numa_scan_seq = seq;
939 p->numa_migrate_seq++;
940 p->numa_scan_period_max = task_scan_max(p);
942 /* Find the node with the highest number of faults */
943 for_each_online_node(nid) {
944 unsigned long faults;
947 for (priv = 0; priv < 2; priv++) {
948 i = task_faults_idx(nid, priv);
950 /* Decay existing window, copy faults since last scan */
951 p->numa_faults[i] >>= 1;
952 p->numa_faults[i] += p->numa_faults_buffer[i];
953 p->numa_faults_buffer[i] = 0;
956 /* Find maximum private faults */
957 faults = p->numa_faults[task_faults_idx(nid, 1)];
958 if (faults > max_faults) {
965 * Record the preferred node as the node with the most faults,
966 * requeue the task to be running on the idlest CPU on the
967 * preferred node and reset the scanning rate to recheck
968 * the working set placement.
970 if (max_faults && max_nid != p->numa_preferred_nid) {
974 * If the task is not on the preferred node then find the most
975 * idle CPU to migrate to.
977 preferred_cpu = task_cpu(p);
978 if (cpu_to_node(preferred_cpu) != max_nid) {
979 preferred_cpu = find_idlest_cpu_node(preferred_cpu,
983 /* Update the preferred nid and migrate task if possible */
984 p->numa_preferred_nid = max_nid;
985 p->numa_migrate_seq = 0;
986 migrate_task_to(p, preferred_cpu);
991 * Got a PROT_NONE fault for a page on @node.
993 void task_numa_fault(int last_nid, int node, int pages, bool migrated)
995 struct task_struct *p = current;
998 if (!numabalancing_enabled)
1001 /* For now, do not attempt to detect private/shared accesses */
1004 /* Allocate buffer to track faults on a per-node basis */
1005 if (unlikely(!p->numa_faults)) {
1006 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1008 /* numa_faults and numa_faults_buffer share the allocation */
1009 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1010 if (!p->numa_faults)
1013 BUG_ON(p->numa_faults_buffer);
1014 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1018 * If pages are properly placed (did not migrate) then scan slower.
1019 * This is reset periodically in case of phase changes
1022 /* Initialise if necessary */
1023 if (!p->numa_scan_period_max)
1024 p->numa_scan_period_max = task_scan_max(p);
1026 p->numa_scan_period = min(p->numa_scan_period_max,
1027 p->numa_scan_period + 10);
1030 task_numa_placement(p);
1032 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1035 static void reset_ptenuma_scan(struct task_struct *p)
1037 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1038 p->mm->numa_scan_offset = 0;
1042 * The expensive part of numa migration is done from task_work context.
1043 * Triggered from task_tick_numa().
1045 void task_numa_work(struct callback_head *work)
1047 unsigned long migrate, next_scan, now = jiffies;
1048 struct task_struct *p = current;
1049 struct mm_struct *mm = p->mm;
1050 struct vm_area_struct *vma;
1051 unsigned long start, end;
1052 unsigned long nr_pte_updates = 0;
1055 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1057 work->next = work; /* protect against double add */
1059 * Who cares about NUMA placement when they're dying.
1061 * NOTE: make sure not to dereference p->mm before this check,
1062 * exit_task_work() happens _after_ exit_mm() so we could be called
1063 * without p->mm even though we still had it when we enqueued this
1066 if (p->flags & PF_EXITING)
1069 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1070 mm->numa_next_scan = now +
1071 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1072 mm->numa_next_reset = now +
1073 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1077 * Reset the scan period if enough time has gone by. Objective is that
1078 * scanning will be reduced if pages are properly placed. As tasks
1079 * can enter different phases this needs to be re-examined. Lacking
1080 * proper tracking of reference behaviour, this blunt hammer is used.
1082 migrate = mm->numa_next_reset;
1083 if (time_after(now, migrate)) {
1084 p->numa_scan_period = task_scan_min(p);
1085 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1086 xchg(&mm->numa_next_reset, next_scan);
1090 * Enforce maximal scan/migration frequency..
1092 migrate = mm->numa_next_scan;
1093 if (time_before(now, migrate))
1096 if (p->numa_scan_period == 0) {
1097 p->numa_scan_period_max = task_scan_max(p);
1098 p->numa_scan_period = task_scan_min(p);
1101 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1102 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1106 * Delay this task enough that another task of this mm will likely win
1107 * the next time around.
1109 p->node_stamp += 2 * TICK_NSEC;
1111 start = mm->numa_scan_offset;
1112 pages = sysctl_numa_balancing_scan_size;
1113 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1117 down_read(&mm->mmap_sem);
1118 vma = find_vma(mm, start);
1120 reset_ptenuma_scan(p);
1124 for (; vma; vma = vma->vm_next) {
1125 if (!vma_migratable(vma))
1128 /* Skip small VMAs. They are not likely to be of relevance */
1129 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1133 start = max(start, vma->vm_start);
1134 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1135 end = min(end, vma->vm_end);
1136 nr_pte_updates += change_prot_numa(vma, start, end);
1139 * Scan sysctl_numa_balancing_scan_size but ensure that
1140 * at least one PTE is updated so that unused virtual
1141 * address space is quickly skipped.
1144 pages -= (end - start) >> PAGE_SHIFT;
1149 } while (end != vma->vm_end);
1154 * If the whole process was scanned without updates then no NUMA
1155 * hinting faults are being recorded and scan rate should be lower.
1157 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1158 p->numa_scan_period = min(p->numa_scan_period_max,
1159 p->numa_scan_period << 1);
1161 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1162 mm->numa_next_scan = next_scan;
1166 * It is possible to reach the end of the VMA list but the last few
1167 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1168 * would find the !migratable VMA on the next scan but not reset the
1169 * scanner to the start so check it now.
1172 mm->numa_scan_offset = start;
1174 reset_ptenuma_scan(p);
1175 up_read(&mm->mmap_sem);
1179 * Drive the periodic memory faults..
1181 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1183 struct callback_head *work = &curr->numa_work;
1187 * We don't care about NUMA placement if we don't have memory.
1189 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1193 * Using runtime rather than walltime has the dual advantage that
1194 * we (mostly) drive the selection from busy threads and that the
1195 * task needs to have done some actual work before we bother with
1198 now = curr->se.sum_exec_runtime;
1199 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1201 if (now - curr->node_stamp > period) {
1202 if (!curr->node_stamp)
1203 curr->numa_scan_period = task_scan_min(curr);
1204 curr->node_stamp += period;
1206 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1207 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1208 task_work_add(curr, work, true);
1213 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1216 #endif /* CONFIG_NUMA_BALANCING */
1219 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1221 update_load_add(&cfs_rq->load, se->load.weight);
1222 if (!parent_entity(se))
1223 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1225 if (entity_is_task(se))
1226 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1228 cfs_rq->nr_running++;
1232 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1234 update_load_sub(&cfs_rq->load, se->load.weight);
1235 if (!parent_entity(se))
1236 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1237 if (entity_is_task(se))
1238 list_del_init(&se->group_node);
1239 cfs_rq->nr_running--;
1242 #ifdef CONFIG_FAIR_GROUP_SCHED
1244 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1249 * Use this CPU's actual weight instead of the last load_contribution
1250 * to gain a more accurate current total weight. See
1251 * update_cfs_rq_load_contribution().
1253 tg_weight = atomic_long_read(&tg->load_avg);
1254 tg_weight -= cfs_rq->tg_load_contrib;
1255 tg_weight += cfs_rq->load.weight;
1260 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1262 long tg_weight, load, shares;
1264 tg_weight = calc_tg_weight(tg, cfs_rq);
1265 load = cfs_rq->load.weight;
1267 shares = (tg->shares * load);
1269 shares /= tg_weight;
1271 if (shares < MIN_SHARES)
1272 shares = MIN_SHARES;
1273 if (shares > tg->shares)
1274 shares = tg->shares;
1278 # else /* CONFIG_SMP */
1279 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1283 # endif /* CONFIG_SMP */
1284 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1285 unsigned long weight)
1288 /* commit outstanding execution time */
1289 if (cfs_rq->curr == se)
1290 update_curr(cfs_rq);
1291 account_entity_dequeue(cfs_rq, se);
1294 update_load_set(&se->load, weight);
1297 account_entity_enqueue(cfs_rq, se);
1300 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1302 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1304 struct task_group *tg;
1305 struct sched_entity *se;
1309 se = tg->se[cpu_of(rq_of(cfs_rq))];
1310 if (!se || throttled_hierarchy(cfs_rq))
1313 if (likely(se->load.weight == tg->shares))
1316 shares = calc_cfs_shares(cfs_rq, tg);
1318 reweight_entity(cfs_rq_of(se), se, shares);
1320 #else /* CONFIG_FAIR_GROUP_SCHED */
1321 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1324 #endif /* CONFIG_FAIR_GROUP_SCHED */
1328 * We choose a half-life close to 1 scheduling period.
1329 * Note: The tables below are dependent on this value.
1331 #define LOAD_AVG_PERIOD 32
1332 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1333 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1335 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1336 static const u32 runnable_avg_yN_inv[] = {
1337 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1338 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1339 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1340 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1341 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1342 0x85aac367, 0x82cd8698,
1346 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1347 * over-estimates when re-combining.
1349 static const u32 runnable_avg_yN_sum[] = {
1350 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1351 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1352 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1357 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1359 static __always_inline u64 decay_load(u64 val, u64 n)
1361 unsigned int local_n;
1365 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1368 /* after bounds checking we can collapse to 32-bit */
1372 * As y^PERIOD = 1/2, we can combine
1373 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1374 * With a look-up table which covers k^n (n<PERIOD)
1376 * To achieve constant time decay_load.
1378 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1379 val >>= local_n / LOAD_AVG_PERIOD;
1380 local_n %= LOAD_AVG_PERIOD;
1383 val *= runnable_avg_yN_inv[local_n];
1384 /* We don't use SRR here since we always want to round down. */
1389 * For updates fully spanning n periods, the contribution to runnable
1390 * average will be: \Sum 1024*y^n
1392 * We can compute this reasonably efficiently by combining:
1393 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1395 static u32 __compute_runnable_contrib(u64 n)
1399 if (likely(n <= LOAD_AVG_PERIOD))
1400 return runnable_avg_yN_sum[n];
1401 else if (unlikely(n >= LOAD_AVG_MAX_N))
1402 return LOAD_AVG_MAX;
1404 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1406 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1407 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1409 n -= LOAD_AVG_PERIOD;
1410 } while (n > LOAD_AVG_PERIOD);
1412 contrib = decay_load(contrib, n);
1413 return contrib + runnable_avg_yN_sum[n];
1417 * We can represent the historical contribution to runnable average as the
1418 * coefficients of a geometric series. To do this we sub-divide our runnable
1419 * history into segments of approximately 1ms (1024us); label the segment that
1420 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1422 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1424 * (now) (~1ms ago) (~2ms ago)
1426 * Let u_i denote the fraction of p_i that the entity was runnable.
1428 * We then designate the fractions u_i as our co-efficients, yielding the
1429 * following representation of historical load:
1430 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1432 * We choose y based on the with of a reasonably scheduling period, fixing:
1435 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1436 * approximately half as much as the contribution to load within the last ms
1439 * When a period "rolls over" and we have new u_0`, multiplying the previous
1440 * sum again by y is sufficient to update:
1441 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1442 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1444 static __always_inline int __update_entity_runnable_avg(u64 now,
1445 struct sched_avg *sa,
1449 u32 runnable_contrib;
1450 int delta_w, decayed = 0;
1452 delta = now - sa->last_runnable_update;
1454 * This should only happen when time goes backwards, which it
1455 * unfortunately does during sched clock init when we swap over to TSC.
1457 if ((s64)delta < 0) {
1458 sa->last_runnable_update = now;
1463 * Use 1024ns as the unit of measurement since it's a reasonable
1464 * approximation of 1us and fast to compute.
1469 sa->last_runnable_update = now;
1471 /* delta_w is the amount already accumulated against our next period */
1472 delta_w = sa->runnable_avg_period % 1024;
1473 if (delta + delta_w >= 1024) {
1474 /* period roll-over */
1478 * Now that we know we're crossing a period boundary, figure
1479 * out how much from delta we need to complete the current
1480 * period and accrue it.
1482 delta_w = 1024 - delta_w;
1484 sa->runnable_avg_sum += delta_w;
1485 sa->runnable_avg_period += delta_w;
1489 /* Figure out how many additional periods this update spans */
1490 periods = delta / 1024;
1493 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1495 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1498 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1499 runnable_contrib = __compute_runnable_contrib(periods);
1501 sa->runnable_avg_sum += runnable_contrib;
1502 sa->runnable_avg_period += runnable_contrib;
1505 /* Remainder of delta accrued against u_0` */
1507 sa->runnable_avg_sum += delta;
1508 sa->runnable_avg_period += delta;
1513 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1514 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1516 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1517 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1519 decays -= se->avg.decay_count;
1523 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1524 se->avg.decay_count = 0;
1529 #ifdef CONFIG_FAIR_GROUP_SCHED
1530 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1533 struct task_group *tg = cfs_rq->tg;
1536 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1537 tg_contrib -= cfs_rq->tg_load_contrib;
1539 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1540 atomic_long_add(tg_contrib, &tg->load_avg);
1541 cfs_rq->tg_load_contrib += tg_contrib;
1546 * Aggregate cfs_rq runnable averages into an equivalent task_group
1547 * representation for computing load contributions.
1549 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1550 struct cfs_rq *cfs_rq)
1552 struct task_group *tg = cfs_rq->tg;
1555 /* The fraction of a cpu used by this cfs_rq */
1556 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1557 sa->runnable_avg_period + 1);
1558 contrib -= cfs_rq->tg_runnable_contrib;
1560 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1561 atomic_add(contrib, &tg->runnable_avg);
1562 cfs_rq->tg_runnable_contrib += contrib;
1566 static inline void __update_group_entity_contrib(struct sched_entity *se)
1568 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1569 struct task_group *tg = cfs_rq->tg;
1574 contrib = cfs_rq->tg_load_contrib * tg->shares;
1575 se->avg.load_avg_contrib = div_u64(contrib,
1576 atomic_long_read(&tg->load_avg) + 1);
1579 * For group entities we need to compute a correction term in the case
1580 * that they are consuming <1 cpu so that we would contribute the same
1581 * load as a task of equal weight.
1583 * Explicitly co-ordinating this measurement would be expensive, but
1584 * fortunately the sum of each cpus contribution forms a usable
1585 * lower-bound on the true value.
1587 * Consider the aggregate of 2 contributions. Either they are disjoint
1588 * (and the sum represents true value) or they are disjoint and we are
1589 * understating by the aggregate of their overlap.
1591 * Extending this to N cpus, for a given overlap, the maximum amount we
1592 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1593 * cpus that overlap for this interval and w_i is the interval width.
1595 * On a small machine; the first term is well-bounded which bounds the
1596 * total error since w_i is a subset of the period. Whereas on a
1597 * larger machine, while this first term can be larger, if w_i is the
1598 * of consequential size guaranteed to see n_i*w_i quickly converge to
1599 * our upper bound of 1-cpu.
1601 runnable_avg = atomic_read(&tg->runnable_avg);
1602 if (runnable_avg < NICE_0_LOAD) {
1603 se->avg.load_avg_contrib *= runnable_avg;
1604 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1608 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1609 int force_update) {}
1610 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1611 struct cfs_rq *cfs_rq) {}
1612 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1615 static inline void __update_task_entity_contrib(struct sched_entity *se)
1619 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1620 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1621 contrib /= (se->avg.runnable_avg_period + 1);
1622 se->avg.load_avg_contrib = scale_load(contrib);
1625 /* Compute the current contribution to load_avg by se, return any delta */
1626 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1628 long old_contrib = se->avg.load_avg_contrib;
1630 if (entity_is_task(se)) {
1631 __update_task_entity_contrib(se);
1633 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1634 __update_group_entity_contrib(se);
1637 return se->avg.load_avg_contrib - old_contrib;
1640 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1643 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1644 cfs_rq->blocked_load_avg -= load_contrib;
1646 cfs_rq->blocked_load_avg = 0;
1649 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1651 /* Update a sched_entity's runnable average */
1652 static inline void update_entity_load_avg(struct sched_entity *se,
1655 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1660 * For a group entity we need to use their owned cfs_rq_clock_task() in
1661 * case they are the parent of a throttled hierarchy.
1663 if (entity_is_task(se))
1664 now = cfs_rq_clock_task(cfs_rq);
1666 now = cfs_rq_clock_task(group_cfs_rq(se));
1668 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1671 contrib_delta = __update_entity_load_avg_contrib(se);
1677 cfs_rq->runnable_load_avg += contrib_delta;
1679 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1683 * Decay the load contributed by all blocked children and account this so that
1684 * their contribution may appropriately discounted when they wake up.
1686 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1688 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1691 decays = now - cfs_rq->last_decay;
1692 if (!decays && !force_update)
1695 if (atomic_long_read(&cfs_rq->removed_load)) {
1696 unsigned long removed_load;
1697 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1698 subtract_blocked_load_contrib(cfs_rq, removed_load);
1702 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1704 atomic64_add(decays, &cfs_rq->decay_counter);
1705 cfs_rq->last_decay = now;
1708 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1711 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1713 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1714 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1717 /* Add the load generated by se into cfs_rq's child load-average */
1718 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1719 struct sched_entity *se,
1723 * We track migrations using entity decay_count <= 0, on a wake-up
1724 * migration we use a negative decay count to track the remote decays
1725 * accumulated while sleeping.
1727 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1728 * are seen by enqueue_entity_load_avg() as a migration with an already
1729 * constructed load_avg_contrib.
1731 if (unlikely(se->avg.decay_count <= 0)) {
1732 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1733 if (se->avg.decay_count) {
1735 * In a wake-up migration we have to approximate the
1736 * time sleeping. This is because we can't synchronize
1737 * clock_task between the two cpus, and it is not
1738 * guaranteed to be read-safe. Instead, we can
1739 * approximate this using our carried decays, which are
1740 * explicitly atomically readable.
1742 se->avg.last_runnable_update -= (-se->avg.decay_count)
1744 update_entity_load_avg(se, 0);
1745 /* Indicate that we're now synchronized and on-rq */
1746 se->avg.decay_count = 0;
1751 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1752 * would have made count negative); we must be careful to avoid
1753 * double-accounting blocked time after synchronizing decays.
1755 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1759 /* migrated tasks did not contribute to our blocked load */
1761 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1762 update_entity_load_avg(se, 0);
1765 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1766 /* we force update consideration on load-balancer moves */
1767 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1771 * Remove se's load from this cfs_rq child load-average, if the entity is
1772 * transitioning to a blocked state we track its projected decay using
1775 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1776 struct sched_entity *se,
1779 update_entity_load_avg(se, 1);
1780 /* we force update consideration on load-balancer moves */
1781 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1783 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1785 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1786 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1787 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1791 * Update the rq's load with the elapsed running time before entering
1792 * idle. if the last scheduled task is not a CFS task, idle_enter will
1793 * be the only way to update the runnable statistic.
1795 void idle_enter_fair(struct rq *this_rq)
1797 update_rq_runnable_avg(this_rq, 1);
1801 * Update the rq's load with the elapsed idle time before a task is
1802 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1803 * be the only way to update the runnable statistic.
1805 void idle_exit_fair(struct rq *this_rq)
1807 update_rq_runnable_avg(this_rq, 0);
1811 static inline void update_entity_load_avg(struct sched_entity *se,
1812 int update_cfs_rq) {}
1813 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1814 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1815 struct sched_entity *se,
1817 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1818 struct sched_entity *se,
1820 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1821 int force_update) {}
1824 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1826 #ifdef CONFIG_SCHEDSTATS
1827 struct task_struct *tsk = NULL;
1829 if (entity_is_task(se))
1832 if (se->statistics.sleep_start) {
1833 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1838 if (unlikely(delta > se->statistics.sleep_max))
1839 se->statistics.sleep_max = delta;
1841 se->statistics.sleep_start = 0;
1842 se->statistics.sum_sleep_runtime += delta;
1845 account_scheduler_latency(tsk, delta >> 10, 1);
1846 trace_sched_stat_sleep(tsk, delta);
1849 if (se->statistics.block_start) {
1850 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1855 if (unlikely(delta > se->statistics.block_max))
1856 se->statistics.block_max = delta;
1858 se->statistics.block_start = 0;
1859 se->statistics.sum_sleep_runtime += delta;
1862 if (tsk->in_iowait) {
1863 se->statistics.iowait_sum += delta;
1864 se->statistics.iowait_count++;
1865 trace_sched_stat_iowait(tsk, delta);
1868 trace_sched_stat_blocked(tsk, delta);
1871 * Blocking time is in units of nanosecs, so shift by
1872 * 20 to get a milliseconds-range estimation of the
1873 * amount of time that the task spent sleeping:
1875 if (unlikely(prof_on == SLEEP_PROFILING)) {
1876 profile_hits(SLEEP_PROFILING,
1877 (void *)get_wchan(tsk),
1880 account_scheduler_latency(tsk, delta >> 10, 0);
1886 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1888 #ifdef CONFIG_SCHED_DEBUG
1889 s64 d = se->vruntime - cfs_rq->min_vruntime;
1894 if (d > 3*sysctl_sched_latency)
1895 schedstat_inc(cfs_rq, nr_spread_over);
1900 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1902 u64 vruntime = cfs_rq->min_vruntime;
1905 * The 'current' period is already promised to the current tasks,
1906 * however the extra weight of the new task will slow them down a
1907 * little, place the new task so that it fits in the slot that
1908 * stays open at the end.
1910 if (initial && sched_feat(START_DEBIT))
1911 vruntime += sched_vslice(cfs_rq, se);
1913 /* sleeps up to a single latency don't count. */
1915 unsigned long thresh = sysctl_sched_latency;
1918 * Halve their sleep time's effect, to allow
1919 * for a gentler effect of sleepers:
1921 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1927 /* ensure we never gain time by being placed backwards. */
1928 se->vruntime = max_vruntime(se->vruntime, vruntime);
1931 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1934 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1937 * Update the normalized vruntime before updating min_vruntime
1938 * through calling update_curr().
1940 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1941 se->vruntime += cfs_rq->min_vruntime;
1944 * Update run-time statistics of the 'current'.
1946 update_curr(cfs_rq);
1947 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1948 account_entity_enqueue(cfs_rq, se);
1949 update_cfs_shares(cfs_rq);
1951 if (flags & ENQUEUE_WAKEUP) {
1952 place_entity(cfs_rq, se, 0);
1953 enqueue_sleeper(cfs_rq, se);
1956 update_stats_enqueue(cfs_rq, se);
1957 check_spread(cfs_rq, se);
1958 if (se != cfs_rq->curr)
1959 __enqueue_entity(cfs_rq, se);
1962 if (cfs_rq->nr_running == 1) {
1963 list_add_leaf_cfs_rq(cfs_rq);
1964 check_enqueue_throttle(cfs_rq);
1968 static void __clear_buddies_last(struct sched_entity *se)
1970 for_each_sched_entity(se) {
1971 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1972 if (cfs_rq->last == se)
1973 cfs_rq->last = NULL;
1979 static void __clear_buddies_next(struct sched_entity *se)
1981 for_each_sched_entity(se) {
1982 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1983 if (cfs_rq->next == se)
1984 cfs_rq->next = NULL;
1990 static void __clear_buddies_skip(struct sched_entity *se)
1992 for_each_sched_entity(se) {
1993 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1994 if (cfs_rq->skip == se)
1995 cfs_rq->skip = NULL;
2001 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2003 if (cfs_rq->last == se)
2004 __clear_buddies_last(se);
2006 if (cfs_rq->next == se)
2007 __clear_buddies_next(se);
2009 if (cfs_rq->skip == se)
2010 __clear_buddies_skip(se);
2013 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2016 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2019 * Update run-time statistics of the 'current'.
2021 update_curr(cfs_rq);
2022 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2024 update_stats_dequeue(cfs_rq, se);
2025 if (flags & DEQUEUE_SLEEP) {
2026 #ifdef CONFIG_SCHEDSTATS
2027 if (entity_is_task(se)) {
2028 struct task_struct *tsk = task_of(se);
2030 if (tsk->state & TASK_INTERRUPTIBLE)
2031 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2032 if (tsk->state & TASK_UNINTERRUPTIBLE)
2033 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2038 clear_buddies(cfs_rq, se);
2040 if (se != cfs_rq->curr)
2041 __dequeue_entity(cfs_rq, se);
2043 account_entity_dequeue(cfs_rq, se);
2046 * Normalize the entity after updating the min_vruntime because the
2047 * update can refer to the ->curr item and we need to reflect this
2048 * movement in our normalized position.
2050 if (!(flags & DEQUEUE_SLEEP))
2051 se->vruntime -= cfs_rq->min_vruntime;
2053 /* return excess runtime on last dequeue */
2054 return_cfs_rq_runtime(cfs_rq);
2056 update_min_vruntime(cfs_rq);
2057 update_cfs_shares(cfs_rq);
2061 * Preempt the current task with a newly woken task if needed:
2064 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2066 unsigned long ideal_runtime, delta_exec;
2067 struct sched_entity *se;
2070 ideal_runtime = sched_slice(cfs_rq, curr);
2071 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2072 if (delta_exec > ideal_runtime) {
2073 resched_task(rq_of(cfs_rq)->curr);
2075 * The current task ran long enough, ensure it doesn't get
2076 * re-elected due to buddy favours.
2078 clear_buddies(cfs_rq, curr);
2083 * Ensure that a task that missed wakeup preemption by a
2084 * narrow margin doesn't have to wait for a full slice.
2085 * This also mitigates buddy induced latencies under load.
2087 if (delta_exec < sysctl_sched_min_granularity)
2090 se = __pick_first_entity(cfs_rq);
2091 delta = curr->vruntime - se->vruntime;
2096 if (delta > ideal_runtime)
2097 resched_task(rq_of(cfs_rq)->curr);
2101 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2103 /* 'current' is not kept within the tree. */
2106 * Any task has to be enqueued before it get to execute on
2107 * a CPU. So account for the time it spent waiting on the
2110 update_stats_wait_end(cfs_rq, se);
2111 __dequeue_entity(cfs_rq, se);
2114 update_stats_curr_start(cfs_rq, se);
2116 #ifdef CONFIG_SCHEDSTATS
2118 * Track our maximum slice length, if the CPU's load is at
2119 * least twice that of our own weight (i.e. dont track it
2120 * when there are only lesser-weight tasks around):
2122 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2123 se->statistics.slice_max = max(se->statistics.slice_max,
2124 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2127 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2131 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2134 * Pick the next process, keeping these things in mind, in this order:
2135 * 1) keep things fair between processes/task groups
2136 * 2) pick the "next" process, since someone really wants that to run
2137 * 3) pick the "last" process, for cache locality
2138 * 4) do not run the "skip" process, if something else is available
2140 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2142 struct sched_entity *se = __pick_first_entity(cfs_rq);
2143 struct sched_entity *left = se;
2146 * Avoid running the skip buddy, if running something else can
2147 * be done without getting too unfair.
2149 if (cfs_rq->skip == se) {
2150 struct sched_entity *second = __pick_next_entity(se);
2151 if (second && wakeup_preempt_entity(second, left) < 1)
2156 * Prefer last buddy, try to return the CPU to a preempted task.
2158 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2162 * Someone really wants this to run. If it's not unfair, run it.
2164 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2167 clear_buddies(cfs_rq, se);
2172 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2174 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2177 * If still on the runqueue then deactivate_task()
2178 * was not called and update_curr() has to be done:
2181 update_curr(cfs_rq);
2183 /* throttle cfs_rqs exceeding runtime */
2184 check_cfs_rq_runtime(cfs_rq);
2186 check_spread(cfs_rq, prev);
2188 update_stats_wait_start(cfs_rq, prev);
2189 /* Put 'current' back into the tree. */
2190 __enqueue_entity(cfs_rq, prev);
2191 /* in !on_rq case, update occurred at dequeue */
2192 update_entity_load_avg(prev, 1);
2194 cfs_rq->curr = NULL;
2198 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2201 * Update run-time statistics of the 'current'.
2203 update_curr(cfs_rq);
2206 * Ensure that runnable average is periodically updated.
2208 update_entity_load_avg(curr, 1);
2209 update_cfs_rq_blocked_load(cfs_rq, 1);
2210 update_cfs_shares(cfs_rq);
2212 #ifdef CONFIG_SCHED_HRTICK
2214 * queued ticks are scheduled to match the slice, so don't bother
2215 * validating it and just reschedule.
2218 resched_task(rq_of(cfs_rq)->curr);
2222 * don't let the period tick interfere with the hrtick preemption
2224 if (!sched_feat(DOUBLE_TICK) &&
2225 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2229 if (cfs_rq->nr_running > 1)
2230 check_preempt_tick(cfs_rq, curr);
2234 /**************************************************
2235 * CFS bandwidth control machinery
2238 #ifdef CONFIG_CFS_BANDWIDTH
2240 #ifdef HAVE_JUMP_LABEL
2241 static struct static_key __cfs_bandwidth_used;
2243 static inline bool cfs_bandwidth_used(void)
2245 return static_key_false(&__cfs_bandwidth_used);
2248 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2250 /* only need to count groups transitioning between enabled/!enabled */
2251 if (enabled && !was_enabled)
2252 static_key_slow_inc(&__cfs_bandwidth_used);
2253 else if (!enabled && was_enabled)
2254 static_key_slow_dec(&__cfs_bandwidth_used);
2256 #else /* HAVE_JUMP_LABEL */
2257 static bool cfs_bandwidth_used(void)
2262 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2263 #endif /* HAVE_JUMP_LABEL */
2266 * default period for cfs group bandwidth.
2267 * default: 0.1s, units: nanoseconds
2269 static inline u64 default_cfs_period(void)
2271 return 100000000ULL;
2274 static inline u64 sched_cfs_bandwidth_slice(void)
2276 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2280 * Replenish runtime according to assigned quota and update expiration time.
2281 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2282 * additional synchronization around rq->lock.
2284 * requires cfs_b->lock
2286 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2290 if (cfs_b->quota == RUNTIME_INF)
2293 now = sched_clock_cpu(smp_processor_id());
2294 cfs_b->runtime = cfs_b->quota;
2295 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2298 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2300 return &tg->cfs_bandwidth;
2303 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2304 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2306 if (unlikely(cfs_rq->throttle_count))
2307 return cfs_rq->throttled_clock_task;
2309 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2312 /* returns 0 on failure to allocate runtime */
2313 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2315 struct task_group *tg = cfs_rq->tg;
2316 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2317 u64 amount = 0, min_amount, expires;
2319 /* note: this is a positive sum as runtime_remaining <= 0 */
2320 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2322 raw_spin_lock(&cfs_b->lock);
2323 if (cfs_b->quota == RUNTIME_INF)
2324 amount = min_amount;
2327 * If the bandwidth pool has become inactive, then at least one
2328 * period must have elapsed since the last consumption.
2329 * Refresh the global state and ensure bandwidth timer becomes
2332 if (!cfs_b->timer_active) {
2333 __refill_cfs_bandwidth_runtime(cfs_b);
2334 __start_cfs_bandwidth(cfs_b);
2337 if (cfs_b->runtime > 0) {
2338 amount = min(cfs_b->runtime, min_amount);
2339 cfs_b->runtime -= amount;
2343 expires = cfs_b->runtime_expires;
2344 raw_spin_unlock(&cfs_b->lock);
2346 cfs_rq->runtime_remaining += amount;
2348 * we may have advanced our local expiration to account for allowed
2349 * spread between our sched_clock and the one on which runtime was
2352 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2353 cfs_rq->runtime_expires = expires;
2355 return cfs_rq->runtime_remaining > 0;
2359 * Note: This depends on the synchronization provided by sched_clock and the
2360 * fact that rq->clock snapshots this value.
2362 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2364 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2366 /* if the deadline is ahead of our clock, nothing to do */
2367 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2370 if (cfs_rq->runtime_remaining < 0)
2374 * If the local deadline has passed we have to consider the
2375 * possibility that our sched_clock is 'fast' and the global deadline
2376 * has not truly expired.
2378 * Fortunately we can check determine whether this the case by checking
2379 * whether the global deadline has advanced.
2382 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2383 /* extend local deadline, drift is bounded above by 2 ticks */
2384 cfs_rq->runtime_expires += TICK_NSEC;
2386 /* global deadline is ahead, expiration has passed */
2387 cfs_rq->runtime_remaining = 0;
2391 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2392 unsigned long delta_exec)
2394 /* dock delta_exec before expiring quota (as it could span periods) */
2395 cfs_rq->runtime_remaining -= delta_exec;
2396 expire_cfs_rq_runtime(cfs_rq);
2398 if (likely(cfs_rq->runtime_remaining > 0))
2402 * if we're unable to extend our runtime we resched so that the active
2403 * hierarchy can be throttled
2405 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2406 resched_task(rq_of(cfs_rq)->curr);
2409 static __always_inline
2410 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2412 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2415 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2418 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2420 return cfs_bandwidth_used() && cfs_rq->throttled;
2423 /* check whether cfs_rq, or any parent, is throttled */
2424 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2426 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2430 * Ensure that neither of the group entities corresponding to src_cpu or
2431 * dest_cpu are members of a throttled hierarchy when performing group
2432 * load-balance operations.
2434 static inline int throttled_lb_pair(struct task_group *tg,
2435 int src_cpu, int dest_cpu)
2437 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2439 src_cfs_rq = tg->cfs_rq[src_cpu];
2440 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2442 return throttled_hierarchy(src_cfs_rq) ||
2443 throttled_hierarchy(dest_cfs_rq);
2446 /* updated child weight may affect parent so we have to do this bottom up */
2447 static int tg_unthrottle_up(struct task_group *tg, void *data)
2449 struct rq *rq = data;
2450 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2452 cfs_rq->throttle_count--;
2454 if (!cfs_rq->throttle_count) {
2455 /* adjust cfs_rq_clock_task() */
2456 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2457 cfs_rq->throttled_clock_task;
2464 static int tg_throttle_down(struct task_group *tg, void *data)
2466 struct rq *rq = data;
2467 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2469 /* group is entering throttled state, stop time */
2470 if (!cfs_rq->throttle_count)
2471 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2472 cfs_rq->throttle_count++;
2477 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2479 struct rq *rq = rq_of(cfs_rq);
2480 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2481 struct sched_entity *se;
2482 long task_delta, dequeue = 1;
2484 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2486 /* freeze hierarchy runnable averages while throttled */
2488 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2491 task_delta = cfs_rq->h_nr_running;
2492 for_each_sched_entity(se) {
2493 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2494 /* throttled entity or throttle-on-deactivate */
2499 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2500 qcfs_rq->h_nr_running -= task_delta;
2502 if (qcfs_rq->load.weight)
2507 rq->nr_running -= task_delta;
2509 cfs_rq->throttled = 1;
2510 cfs_rq->throttled_clock = rq_clock(rq);
2511 raw_spin_lock(&cfs_b->lock);
2512 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2513 raw_spin_unlock(&cfs_b->lock);
2516 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2518 struct rq *rq = rq_of(cfs_rq);
2519 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2520 struct sched_entity *se;
2524 se = cfs_rq->tg->se[cpu_of(rq)];
2526 cfs_rq->throttled = 0;
2528 update_rq_clock(rq);
2530 raw_spin_lock(&cfs_b->lock);
2531 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2532 list_del_rcu(&cfs_rq->throttled_list);
2533 raw_spin_unlock(&cfs_b->lock);
2535 /* update hierarchical throttle state */
2536 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2538 if (!cfs_rq->load.weight)
2541 task_delta = cfs_rq->h_nr_running;
2542 for_each_sched_entity(se) {
2546 cfs_rq = cfs_rq_of(se);
2548 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2549 cfs_rq->h_nr_running += task_delta;
2551 if (cfs_rq_throttled(cfs_rq))
2556 rq->nr_running += task_delta;
2558 /* determine whether we need to wake up potentially idle cpu */
2559 if (rq->curr == rq->idle && rq->cfs.nr_running)
2560 resched_task(rq->curr);
2563 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2564 u64 remaining, u64 expires)
2566 struct cfs_rq *cfs_rq;
2567 u64 runtime = remaining;
2570 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2572 struct rq *rq = rq_of(cfs_rq);
2574 raw_spin_lock(&rq->lock);
2575 if (!cfs_rq_throttled(cfs_rq))
2578 runtime = -cfs_rq->runtime_remaining + 1;
2579 if (runtime > remaining)
2580 runtime = remaining;
2581 remaining -= runtime;
2583 cfs_rq->runtime_remaining += runtime;
2584 cfs_rq->runtime_expires = expires;
2586 /* we check whether we're throttled above */
2587 if (cfs_rq->runtime_remaining > 0)
2588 unthrottle_cfs_rq(cfs_rq);
2591 raw_spin_unlock(&rq->lock);
2602 * Responsible for refilling a task_group's bandwidth and unthrottling its
2603 * cfs_rqs as appropriate. If there has been no activity within the last
2604 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2605 * used to track this state.
2607 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2609 u64 runtime, runtime_expires;
2610 int idle = 1, throttled;
2612 raw_spin_lock(&cfs_b->lock);
2613 /* no need to continue the timer with no bandwidth constraint */
2614 if (cfs_b->quota == RUNTIME_INF)
2617 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2618 /* idle depends on !throttled (for the case of a large deficit) */
2619 idle = cfs_b->idle && !throttled;
2620 cfs_b->nr_periods += overrun;
2622 /* if we're going inactive then everything else can be deferred */
2626 __refill_cfs_bandwidth_runtime(cfs_b);
2629 /* mark as potentially idle for the upcoming period */
2634 /* account preceding periods in which throttling occurred */
2635 cfs_b->nr_throttled += overrun;
2638 * There are throttled entities so we must first use the new bandwidth
2639 * to unthrottle them before making it generally available. This
2640 * ensures that all existing debts will be paid before a new cfs_rq is
2643 runtime = cfs_b->runtime;
2644 runtime_expires = cfs_b->runtime_expires;
2648 * This check is repeated as we are holding onto the new bandwidth
2649 * while we unthrottle. This can potentially race with an unthrottled
2650 * group trying to acquire new bandwidth from the global pool.
2652 while (throttled && runtime > 0) {
2653 raw_spin_unlock(&cfs_b->lock);
2654 /* we can't nest cfs_b->lock while distributing bandwidth */
2655 runtime = distribute_cfs_runtime(cfs_b, runtime,
2657 raw_spin_lock(&cfs_b->lock);
2659 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2662 /* return (any) remaining runtime */
2663 cfs_b->runtime = runtime;
2665 * While we are ensured activity in the period following an
2666 * unthrottle, this also covers the case in which the new bandwidth is
2667 * insufficient to cover the existing bandwidth deficit. (Forcing the
2668 * timer to remain active while there are any throttled entities.)
2673 cfs_b->timer_active = 0;
2674 raw_spin_unlock(&cfs_b->lock);
2679 /* a cfs_rq won't donate quota below this amount */
2680 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2681 /* minimum remaining period time to redistribute slack quota */
2682 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2683 /* how long we wait to gather additional slack before distributing */
2684 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2686 /* are we near the end of the current quota period? */
2687 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2689 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2692 /* if the call-back is running a quota refresh is already occurring */
2693 if (hrtimer_callback_running(refresh_timer))
2696 /* is a quota refresh about to occur? */
2697 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2698 if (remaining < min_expire)
2704 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2706 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2708 /* if there's a quota refresh soon don't bother with slack */
2709 if (runtime_refresh_within(cfs_b, min_left))
2712 start_bandwidth_timer(&cfs_b->slack_timer,
2713 ns_to_ktime(cfs_bandwidth_slack_period));
2716 /* we know any runtime found here is valid as update_curr() precedes return */
2717 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2719 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2720 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2722 if (slack_runtime <= 0)
2725 raw_spin_lock(&cfs_b->lock);
2726 if (cfs_b->quota != RUNTIME_INF &&
2727 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2728 cfs_b->runtime += slack_runtime;
2730 /* we are under rq->lock, defer unthrottling using a timer */
2731 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2732 !list_empty(&cfs_b->throttled_cfs_rq))
2733 start_cfs_slack_bandwidth(cfs_b);
2735 raw_spin_unlock(&cfs_b->lock);
2737 /* even if it's not valid for return we don't want to try again */
2738 cfs_rq->runtime_remaining -= slack_runtime;
2741 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2743 if (!cfs_bandwidth_used())
2746 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2749 __return_cfs_rq_runtime(cfs_rq);
2753 * This is done with a timer (instead of inline with bandwidth return) since
2754 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2756 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2758 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2761 /* confirm we're still not at a refresh boundary */
2762 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2765 raw_spin_lock(&cfs_b->lock);
2766 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2767 runtime = cfs_b->runtime;
2770 expires = cfs_b->runtime_expires;
2771 raw_spin_unlock(&cfs_b->lock);
2776 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2778 raw_spin_lock(&cfs_b->lock);
2779 if (expires == cfs_b->runtime_expires)
2780 cfs_b->runtime = runtime;
2781 raw_spin_unlock(&cfs_b->lock);
2785 * When a group wakes up we want to make sure that its quota is not already
2786 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2787 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2789 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2791 if (!cfs_bandwidth_used())
2794 /* an active group must be handled by the update_curr()->put() path */
2795 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2798 /* ensure the group is not already throttled */
2799 if (cfs_rq_throttled(cfs_rq))
2802 /* update runtime allocation */
2803 account_cfs_rq_runtime(cfs_rq, 0);
2804 if (cfs_rq->runtime_remaining <= 0)
2805 throttle_cfs_rq(cfs_rq);
2808 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2809 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2811 if (!cfs_bandwidth_used())
2814 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2818 * it's possible for a throttled entity to be forced into a running
2819 * state (e.g. set_curr_task), in this case we're finished.
2821 if (cfs_rq_throttled(cfs_rq))
2824 throttle_cfs_rq(cfs_rq);
2827 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2829 struct cfs_bandwidth *cfs_b =
2830 container_of(timer, struct cfs_bandwidth, slack_timer);
2831 do_sched_cfs_slack_timer(cfs_b);
2833 return HRTIMER_NORESTART;
2836 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2838 struct cfs_bandwidth *cfs_b =
2839 container_of(timer, struct cfs_bandwidth, period_timer);
2845 now = hrtimer_cb_get_time(timer);
2846 overrun = hrtimer_forward(timer, now, cfs_b->period);
2851 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2854 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2857 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2859 raw_spin_lock_init(&cfs_b->lock);
2861 cfs_b->quota = RUNTIME_INF;
2862 cfs_b->period = ns_to_ktime(default_cfs_period());
2864 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2865 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2866 cfs_b->period_timer.function = sched_cfs_period_timer;
2867 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2868 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2871 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2873 cfs_rq->runtime_enabled = 0;
2874 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2877 /* requires cfs_b->lock, may release to reprogram timer */
2878 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2881 * The timer may be active because we're trying to set a new bandwidth
2882 * period or because we're racing with the tear-down path
2883 * (timer_active==0 becomes visible before the hrtimer call-back
2884 * terminates). In either case we ensure that it's re-programmed
2886 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2887 raw_spin_unlock(&cfs_b->lock);
2888 /* ensure cfs_b->lock is available while we wait */
2889 hrtimer_cancel(&cfs_b->period_timer);
2891 raw_spin_lock(&cfs_b->lock);
2892 /* if someone else restarted the timer then we're done */
2893 if (cfs_b->timer_active)
2897 cfs_b->timer_active = 1;
2898 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2901 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2903 hrtimer_cancel(&cfs_b->period_timer);
2904 hrtimer_cancel(&cfs_b->slack_timer);
2907 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2909 struct cfs_rq *cfs_rq;
2911 for_each_leaf_cfs_rq(rq, cfs_rq) {
2912 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2914 if (!cfs_rq->runtime_enabled)
2918 * clock_task is not advancing so we just need to make sure
2919 * there's some valid quota amount
2921 cfs_rq->runtime_remaining = cfs_b->quota;
2922 if (cfs_rq_throttled(cfs_rq))
2923 unthrottle_cfs_rq(cfs_rq);
2927 #else /* CONFIG_CFS_BANDWIDTH */
2928 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2930 return rq_clock_task(rq_of(cfs_rq));
2933 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2934 unsigned long delta_exec) {}
2935 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2936 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2937 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2939 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2944 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2949 static inline int throttled_lb_pair(struct task_group *tg,
2950 int src_cpu, int dest_cpu)
2955 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2957 #ifdef CONFIG_FAIR_GROUP_SCHED
2958 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2961 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2965 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2966 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2968 #endif /* CONFIG_CFS_BANDWIDTH */
2970 /**************************************************
2971 * CFS operations on tasks:
2974 #ifdef CONFIG_SCHED_HRTICK
2975 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2977 struct sched_entity *se = &p->se;
2978 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2980 WARN_ON(task_rq(p) != rq);
2982 if (cfs_rq->nr_running > 1) {
2983 u64 slice = sched_slice(cfs_rq, se);
2984 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2985 s64 delta = slice - ran;
2994 * Don't schedule slices shorter than 10000ns, that just
2995 * doesn't make sense. Rely on vruntime for fairness.
2998 delta = max_t(s64, 10000LL, delta);
3000 hrtick_start(rq, delta);
3005 * called from enqueue/dequeue and updates the hrtick when the
3006 * current task is from our class and nr_running is low enough
3009 static void hrtick_update(struct rq *rq)
3011 struct task_struct *curr = rq->curr;
3013 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3016 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3017 hrtick_start_fair(rq, curr);
3019 #else /* !CONFIG_SCHED_HRTICK */
3021 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3025 static inline void hrtick_update(struct rq *rq)
3031 * The enqueue_task method is called before nr_running is
3032 * increased. Here we update the fair scheduling stats and
3033 * then put the task into the rbtree:
3036 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3038 struct cfs_rq *cfs_rq;
3039 struct sched_entity *se = &p->se;
3041 for_each_sched_entity(se) {
3044 cfs_rq = cfs_rq_of(se);
3045 enqueue_entity(cfs_rq, se, flags);
3048 * end evaluation on encountering a throttled cfs_rq
3050 * note: in the case of encountering a throttled cfs_rq we will
3051 * post the final h_nr_running increment below.
3053 if (cfs_rq_throttled(cfs_rq))
3055 cfs_rq->h_nr_running++;
3057 flags = ENQUEUE_WAKEUP;
3060 for_each_sched_entity(se) {
3061 cfs_rq = cfs_rq_of(se);
3062 cfs_rq->h_nr_running++;
3064 if (cfs_rq_throttled(cfs_rq))
3067 update_cfs_shares(cfs_rq);
3068 update_entity_load_avg(se, 1);
3072 update_rq_runnable_avg(rq, rq->nr_running);
3078 static void set_next_buddy(struct sched_entity *se);
3081 * The dequeue_task method is called before nr_running is
3082 * decreased. We remove the task from the rbtree and
3083 * update the fair scheduling stats:
3085 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3087 struct cfs_rq *cfs_rq;
3088 struct sched_entity *se = &p->se;
3089 int task_sleep = flags & DEQUEUE_SLEEP;
3091 for_each_sched_entity(se) {
3092 cfs_rq = cfs_rq_of(se);
3093 dequeue_entity(cfs_rq, se, flags);
3096 * end evaluation on encountering a throttled cfs_rq
3098 * note: in the case of encountering a throttled cfs_rq we will
3099 * post the final h_nr_running decrement below.
3101 if (cfs_rq_throttled(cfs_rq))
3103 cfs_rq->h_nr_running--;
3105 /* Don't dequeue parent if it has other entities besides us */
3106 if (cfs_rq->load.weight) {
3108 * Bias pick_next to pick a task from this cfs_rq, as
3109 * p is sleeping when it is within its sched_slice.
3111 if (task_sleep && parent_entity(se))
3112 set_next_buddy(parent_entity(se));
3114 /* avoid re-evaluating load for this entity */
3115 se = parent_entity(se);
3118 flags |= DEQUEUE_SLEEP;
3121 for_each_sched_entity(se) {
3122 cfs_rq = cfs_rq_of(se);
3123 cfs_rq->h_nr_running--;
3125 if (cfs_rq_throttled(cfs_rq))
3128 update_cfs_shares(cfs_rq);
3129 update_entity_load_avg(se, 1);
3134 update_rq_runnable_avg(rq, 1);
3140 /* Used instead of source_load when we know the type == 0 */
3141 static unsigned long weighted_cpuload(const int cpu)
3143 return cpu_rq(cpu)->cfs.runnable_load_avg;
3147 * Return a low guess at the load of a migration-source cpu weighted
3148 * according to the scheduling class and "nice" value.
3150 * We want to under-estimate the load of migration sources, to
3151 * balance conservatively.
3153 static unsigned long source_load(int cpu, int type)
3155 struct rq *rq = cpu_rq(cpu);
3156 unsigned long total = weighted_cpuload(cpu);
3158 if (type == 0 || !sched_feat(LB_BIAS))
3161 return min(rq->cpu_load[type-1], total);
3165 * Return a high guess at the load of a migration-target cpu weighted
3166 * according to the scheduling class and "nice" value.
3168 static unsigned long target_load(int cpu, int type)
3170 struct rq *rq = cpu_rq(cpu);
3171 unsigned long total = weighted_cpuload(cpu);
3173 if (type == 0 || !sched_feat(LB_BIAS))
3176 return max(rq->cpu_load[type-1], total);
3179 static unsigned long power_of(int cpu)
3181 return cpu_rq(cpu)->cpu_power;
3184 static unsigned long cpu_avg_load_per_task(int cpu)
3186 struct rq *rq = cpu_rq(cpu);
3187 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3188 unsigned long load_avg = rq->cfs.runnable_load_avg;
3191 return load_avg / nr_running;
3196 static void record_wakee(struct task_struct *p)
3199 * Rough decay (wiping) for cost saving, don't worry
3200 * about the boundary, really active task won't care
3203 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3204 current->wakee_flips = 0;
3205 current->wakee_flip_decay_ts = jiffies;
3208 if (current->last_wakee != p) {
3209 current->last_wakee = p;
3210 current->wakee_flips++;
3214 static void task_waking_fair(struct task_struct *p)
3216 struct sched_entity *se = &p->se;
3217 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3220 #ifndef CONFIG_64BIT
3221 u64 min_vruntime_copy;
3224 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3226 min_vruntime = cfs_rq->min_vruntime;
3227 } while (min_vruntime != min_vruntime_copy);
3229 min_vruntime = cfs_rq->min_vruntime;
3232 se->vruntime -= min_vruntime;
3236 #ifdef CONFIG_FAIR_GROUP_SCHED
3238 * effective_load() calculates the load change as seen from the root_task_group
3240 * Adding load to a group doesn't make a group heavier, but can cause movement
3241 * of group shares between cpus. Assuming the shares were perfectly aligned one
3242 * can calculate the shift in shares.
3244 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3245 * on this @cpu and results in a total addition (subtraction) of @wg to the
3246 * total group weight.
3248 * Given a runqueue weight distribution (rw_i) we can compute a shares
3249 * distribution (s_i) using:
3251 * s_i = rw_i / \Sum rw_j (1)
3253 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3254 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3255 * shares distribution (s_i):
3257 * rw_i = { 2, 4, 1, 0 }
3258 * s_i = { 2/7, 4/7, 1/7, 0 }
3260 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3261 * task used to run on and the CPU the waker is running on), we need to
3262 * compute the effect of waking a task on either CPU and, in case of a sync
3263 * wakeup, compute the effect of the current task going to sleep.
3265 * So for a change of @wl to the local @cpu with an overall group weight change
3266 * of @wl we can compute the new shares distribution (s'_i) using:
3268 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3270 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3271 * differences in waking a task to CPU 0. The additional task changes the
3272 * weight and shares distributions like:
3274 * rw'_i = { 3, 4, 1, 0 }
3275 * s'_i = { 3/8, 4/8, 1/8, 0 }
3277 * We can then compute the difference in effective weight by using:
3279 * dw_i = S * (s'_i - s_i) (3)
3281 * Where 'S' is the group weight as seen by its parent.
3283 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3284 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3285 * 4/7) times the weight of the group.
3287 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3289 struct sched_entity *se = tg->se[cpu];
3291 if (!tg->parent) /* the trivial, non-cgroup case */
3294 for_each_sched_entity(se) {
3300 * W = @wg + \Sum rw_j
3302 W = wg + calc_tg_weight(tg, se->my_q);
3307 w = se->my_q->load.weight + wl;
3310 * wl = S * s'_i; see (2)
3313 wl = (w * tg->shares) / W;
3318 * Per the above, wl is the new se->load.weight value; since
3319 * those are clipped to [MIN_SHARES, ...) do so now. See
3320 * calc_cfs_shares().
3322 if (wl < MIN_SHARES)
3326 * wl = dw_i = S * (s'_i - s_i); see (3)
3328 wl -= se->load.weight;
3331 * Recursively apply this logic to all parent groups to compute
3332 * the final effective load change on the root group. Since
3333 * only the @tg group gets extra weight, all parent groups can
3334 * only redistribute existing shares. @wl is the shift in shares
3335 * resulting from this level per the above.
3344 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3345 unsigned long wl, unsigned long wg)
3352 static int wake_wide(struct task_struct *p)
3354 int factor = this_cpu_read(sd_llc_size);
3357 * Yeah, it's the switching-frequency, could means many wakee or
3358 * rapidly switch, use factor here will just help to automatically
3359 * adjust the loose-degree, so bigger node will lead to more pull.
3361 if (p->wakee_flips > factor) {
3363 * wakee is somewhat hot, it needs certain amount of cpu
3364 * resource, so if waker is far more hot, prefer to leave
3367 if (current->wakee_flips > (factor * p->wakee_flips))
3374 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3376 s64 this_load, load;
3377 int idx, this_cpu, prev_cpu;
3378 unsigned long tl_per_task;
3379 struct task_group *tg;
3380 unsigned long weight;
3384 * If we wake multiple tasks be careful to not bounce
3385 * ourselves around too much.
3391 this_cpu = smp_processor_id();
3392 prev_cpu = task_cpu(p);
3393 load = source_load(prev_cpu, idx);
3394 this_load = target_load(this_cpu, idx);
3397 * If sync wakeup then subtract the (maximum possible)
3398 * effect of the currently running task from the load
3399 * of the current CPU:
3402 tg = task_group(current);
3403 weight = current->se.load.weight;
3405 this_load += effective_load(tg, this_cpu, -weight, -weight);
3406 load += effective_load(tg, prev_cpu, 0, -weight);
3410 weight = p->se.load.weight;
3413 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3414 * due to the sync cause above having dropped this_load to 0, we'll
3415 * always have an imbalance, but there's really nothing you can do
3416 * about that, so that's good too.
3418 * Otherwise check if either cpus are near enough in load to allow this
3419 * task to be woken on this_cpu.
3421 if (this_load > 0) {
3422 s64 this_eff_load, prev_eff_load;
3424 this_eff_load = 100;
3425 this_eff_load *= power_of(prev_cpu);
3426 this_eff_load *= this_load +
3427 effective_load(tg, this_cpu, weight, weight);
3429 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3430 prev_eff_load *= power_of(this_cpu);
3431 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3433 balanced = this_eff_load <= prev_eff_load;
3438 * If the currently running task will sleep within
3439 * a reasonable amount of time then attract this newly
3442 if (sync && balanced)
3445 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3446 tl_per_task = cpu_avg_load_per_task(this_cpu);
3449 (this_load <= load &&
3450 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3452 * This domain has SD_WAKE_AFFINE and
3453 * p is cache cold in this domain, and
3454 * there is no bad imbalance.
3456 schedstat_inc(sd, ttwu_move_affine);
3457 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3465 * find_idlest_group finds and returns the least busy CPU group within the
3468 static struct sched_group *
3469 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3470 int this_cpu, int load_idx)
3472 struct sched_group *idlest = NULL, *group = sd->groups;
3473 unsigned long min_load = ULONG_MAX, this_load = 0;
3474 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3477 unsigned long load, avg_load;
3481 /* Skip over this group if it has no CPUs allowed */
3482 if (!cpumask_intersects(sched_group_cpus(group),
3483 tsk_cpus_allowed(p)))
3486 local_group = cpumask_test_cpu(this_cpu,
3487 sched_group_cpus(group));
3489 /* Tally up the load of all CPUs in the group */
3492 for_each_cpu(i, sched_group_cpus(group)) {
3493 /* Bias balancing toward cpus of our domain */
3495 load = source_load(i, load_idx);
3497 load = target_load(i, load_idx);
3502 /* Adjust by relative CPU power of the group */
3503 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3506 this_load = avg_load;
3507 } else if (avg_load < min_load) {
3508 min_load = avg_load;
3511 } while (group = group->next, group != sd->groups);
3513 if (!idlest || 100*this_load < imbalance*min_load)
3519 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3522 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3524 unsigned long load, min_load = ULONG_MAX;
3528 /* Traverse only the allowed CPUs */
3529 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3530 load = weighted_cpuload(i);
3532 if (load < min_load || (load == min_load && i == this_cpu)) {
3542 * Try and locate an idle CPU in the sched_domain.
3544 static int select_idle_sibling(struct task_struct *p, int target)
3546 struct sched_domain *sd;
3547 struct sched_group *sg;
3548 int i = task_cpu(p);
3550 if (idle_cpu(target))
3554 * If the prevous cpu is cache affine and idle, don't be stupid.
3556 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3560 * Otherwise, iterate the domains and find an elegible idle cpu.
3562 sd = rcu_dereference(per_cpu(sd_llc, target));
3563 for_each_lower_domain(sd) {
3566 if (!cpumask_intersects(sched_group_cpus(sg),
3567 tsk_cpus_allowed(p)))
3570 for_each_cpu(i, sched_group_cpus(sg)) {
3571 if (i == target || !idle_cpu(i))
3575 target = cpumask_first_and(sched_group_cpus(sg),
3576 tsk_cpus_allowed(p));
3580 } while (sg != sd->groups);
3587 * sched_balance_self: balance the current task (running on cpu) in domains
3588 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3591 * Balance, ie. select the least loaded group.
3593 * Returns the target CPU number, or the same CPU if no balancing is needed.
3595 * preempt must be disabled.
3598 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3600 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3601 int cpu = smp_processor_id();
3602 int prev_cpu = task_cpu(p);
3604 int want_affine = 0;
3605 int sync = wake_flags & WF_SYNC;
3607 if (p->nr_cpus_allowed == 1)
3610 if (sd_flag & SD_BALANCE_WAKE) {
3611 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3617 for_each_domain(cpu, tmp) {
3618 if (!(tmp->flags & SD_LOAD_BALANCE))
3622 * If both cpu and prev_cpu are part of this domain,
3623 * cpu is a valid SD_WAKE_AFFINE target.
3625 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3626 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3631 if (tmp->flags & sd_flag)
3636 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3639 new_cpu = select_idle_sibling(p, prev_cpu);
3644 int load_idx = sd->forkexec_idx;
3645 struct sched_group *group;
3648 if (!(sd->flags & sd_flag)) {
3653 if (sd_flag & SD_BALANCE_WAKE)
3654 load_idx = sd->wake_idx;
3656 group = find_idlest_group(sd, p, cpu, load_idx);
3662 new_cpu = find_idlest_cpu(group, p, cpu);
3663 if (new_cpu == -1 || new_cpu == cpu) {
3664 /* Now try balancing at a lower domain level of cpu */
3669 /* Now try balancing at a lower domain level of new_cpu */
3671 weight = sd->span_weight;
3673 for_each_domain(cpu, tmp) {
3674 if (weight <= tmp->span_weight)
3676 if (tmp->flags & sd_flag)
3679 /* while loop will break here if sd == NULL */
3688 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3689 * cfs_rq_of(p) references at time of call are still valid and identify the
3690 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3691 * other assumptions, including the state of rq->lock, should be made.
3694 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3696 struct sched_entity *se = &p->se;
3697 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3700 * Load tracking: accumulate removed load so that it can be processed
3701 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3702 * to blocked load iff they have a positive decay-count. It can never
3703 * be negative here since on-rq tasks have decay-count == 0.
3705 if (se->avg.decay_count) {
3706 se->avg.decay_count = -__synchronize_entity_decay(se);
3707 atomic_long_add(se->avg.load_avg_contrib,
3708 &cfs_rq->removed_load);
3711 #endif /* CONFIG_SMP */
3713 static unsigned long
3714 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3716 unsigned long gran = sysctl_sched_wakeup_granularity;
3719 * Since its curr running now, convert the gran from real-time
3720 * to virtual-time in his units.
3722 * By using 'se' instead of 'curr' we penalize light tasks, so
3723 * they get preempted easier. That is, if 'se' < 'curr' then
3724 * the resulting gran will be larger, therefore penalizing the
3725 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3726 * be smaller, again penalizing the lighter task.
3728 * This is especially important for buddies when the leftmost
3729 * task is higher priority than the buddy.
3731 return calc_delta_fair(gran, se);
3735 * Should 'se' preempt 'curr'.
3749 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3751 s64 gran, vdiff = curr->vruntime - se->vruntime;
3756 gran = wakeup_gran(curr, se);
3763 static void set_last_buddy(struct sched_entity *se)
3765 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3768 for_each_sched_entity(se)
3769 cfs_rq_of(se)->last = se;
3772 static void set_next_buddy(struct sched_entity *se)
3774 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3777 for_each_sched_entity(se)
3778 cfs_rq_of(se)->next = se;
3781 static void set_skip_buddy(struct sched_entity *se)
3783 for_each_sched_entity(se)
3784 cfs_rq_of(se)->skip = se;
3788 * Preempt the current task with a newly woken task if needed:
3790 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3792 struct task_struct *curr = rq->curr;
3793 struct sched_entity *se = &curr->se, *pse = &p->se;
3794 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3795 int scale = cfs_rq->nr_running >= sched_nr_latency;
3796 int next_buddy_marked = 0;
3798 if (unlikely(se == pse))
3802 * This is possible from callers such as move_task(), in which we
3803 * unconditionally check_prempt_curr() after an enqueue (which may have
3804 * lead to a throttle). This both saves work and prevents false
3805 * next-buddy nomination below.
3807 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3810 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3811 set_next_buddy(pse);
3812 next_buddy_marked = 1;
3816 * We can come here with TIF_NEED_RESCHED already set from new task
3819 * Note: this also catches the edge-case of curr being in a throttled
3820 * group (e.g. via set_curr_task), since update_curr() (in the
3821 * enqueue of curr) will have resulted in resched being set. This
3822 * prevents us from potentially nominating it as a false LAST_BUDDY
3825 if (test_tsk_need_resched(curr))
3828 /* Idle tasks are by definition preempted by non-idle tasks. */
3829 if (unlikely(curr->policy == SCHED_IDLE) &&
3830 likely(p->policy != SCHED_IDLE))
3834 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3835 * is driven by the tick):
3837 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3840 find_matching_se(&se, &pse);
3841 update_curr(cfs_rq_of(se));
3843 if (wakeup_preempt_entity(se, pse) == 1) {
3845 * Bias pick_next to pick the sched entity that is
3846 * triggering this preemption.
3848 if (!next_buddy_marked)
3849 set_next_buddy(pse);
3858 * Only set the backward buddy when the current task is still
3859 * on the rq. This can happen when a wakeup gets interleaved
3860 * with schedule on the ->pre_schedule() or idle_balance()
3861 * point, either of which can * drop the rq lock.
3863 * Also, during early boot the idle thread is in the fair class,
3864 * for obvious reasons its a bad idea to schedule back to it.
3866 if (unlikely(!se->on_rq || curr == rq->idle))
3869 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3873 static struct task_struct *pick_next_task_fair(struct rq *rq)
3875 struct task_struct *p;
3876 struct cfs_rq *cfs_rq = &rq->cfs;
3877 struct sched_entity *se;
3879 if (!cfs_rq->nr_running)
3883 se = pick_next_entity(cfs_rq);
3884 set_next_entity(cfs_rq, se);
3885 cfs_rq = group_cfs_rq(se);
3889 if (hrtick_enabled(rq))
3890 hrtick_start_fair(rq, p);
3896 * Account for a descheduled task:
3898 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3900 struct sched_entity *se = &prev->se;
3901 struct cfs_rq *cfs_rq;
3903 for_each_sched_entity(se) {
3904 cfs_rq = cfs_rq_of(se);
3905 put_prev_entity(cfs_rq, se);
3910 * sched_yield() is very simple
3912 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3914 static void yield_task_fair(struct rq *rq)
3916 struct task_struct *curr = rq->curr;
3917 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3918 struct sched_entity *se = &curr->se;
3921 * Are we the only task in the tree?
3923 if (unlikely(rq->nr_running == 1))
3926 clear_buddies(cfs_rq, se);
3928 if (curr->policy != SCHED_BATCH) {
3929 update_rq_clock(rq);
3931 * Update run-time statistics of the 'current'.
3933 update_curr(cfs_rq);
3935 * Tell update_rq_clock() that we've just updated,
3936 * so we don't do microscopic update in schedule()
3937 * and double the fastpath cost.
3939 rq->skip_clock_update = 1;
3945 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3947 struct sched_entity *se = &p->se;
3949 /* throttled hierarchies are not runnable */
3950 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3953 /* Tell the scheduler that we'd really like pse to run next. */
3956 yield_task_fair(rq);
3962 /**************************************************
3963 * Fair scheduling class load-balancing methods.
3967 * The purpose of load-balancing is to achieve the same basic fairness the
3968 * per-cpu scheduler provides, namely provide a proportional amount of compute
3969 * time to each task. This is expressed in the following equation:
3971 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3973 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3974 * W_i,0 is defined as:
3976 * W_i,0 = \Sum_j w_i,j (2)
3978 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3979 * is derived from the nice value as per prio_to_weight[].
3981 * The weight average is an exponential decay average of the instantaneous
3984 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3986 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3987 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3988 * can also include other factors [XXX].
3990 * To achieve this balance we define a measure of imbalance which follows
3991 * directly from (1):
3993 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3995 * We them move tasks around to minimize the imbalance. In the continuous
3996 * function space it is obvious this converges, in the discrete case we get
3997 * a few fun cases generally called infeasible weight scenarios.
4000 * - infeasible weights;
4001 * - local vs global optima in the discrete case. ]
4006 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4007 * for all i,j solution, we create a tree of cpus that follows the hardware
4008 * topology where each level pairs two lower groups (or better). This results
4009 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4010 * tree to only the first of the previous level and we decrease the frequency
4011 * of load-balance at each level inv. proportional to the number of cpus in
4017 * \Sum { --- * --- * 2^i } = O(n) (5)
4019 * `- size of each group
4020 * | | `- number of cpus doing load-balance
4022 * `- sum over all levels
4024 * Coupled with a limit on how many tasks we can migrate every balance pass,
4025 * this makes (5) the runtime complexity of the balancer.
4027 * An important property here is that each CPU is still (indirectly) connected
4028 * to every other cpu in at most O(log n) steps:
4030 * The adjacency matrix of the resulting graph is given by:
4033 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4036 * And you'll find that:
4038 * A^(log_2 n)_i,j != 0 for all i,j (7)
4040 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4041 * The task movement gives a factor of O(m), giving a convergence complexity
4044 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4049 * In order to avoid CPUs going idle while there's still work to do, new idle
4050 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4051 * tree itself instead of relying on other CPUs to bring it work.
4053 * This adds some complexity to both (5) and (8) but it reduces the total idle
4061 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4064 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4069 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4071 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4073 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4076 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4077 * rewrite all of this once again.]
4080 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4082 #define LBF_ALL_PINNED 0x01
4083 #define LBF_NEED_BREAK 0x02
4084 #define LBF_DST_PINNED 0x04
4085 #define LBF_SOME_PINNED 0x08
4088 struct sched_domain *sd;
4096 struct cpumask *dst_grpmask;
4098 enum cpu_idle_type idle;
4100 /* The set of CPUs under consideration for load-balancing */
4101 struct cpumask *cpus;
4106 unsigned int loop_break;
4107 unsigned int loop_max;
4111 * move_task - move a task from one runqueue to another runqueue.
4112 * Both runqueues must be locked.
4114 static void move_task(struct task_struct *p, struct lb_env *env)
4116 deactivate_task(env->src_rq, p, 0);
4117 set_task_cpu(p, env->dst_cpu);
4118 activate_task(env->dst_rq, p, 0);
4119 check_preempt_curr(env->dst_rq, p, 0);
4123 * Is this task likely cache-hot:
4126 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4130 if (p->sched_class != &fair_sched_class)
4133 if (unlikely(p->policy == SCHED_IDLE))
4137 * Buddy candidates are cache hot:
4139 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4140 (&p->se == cfs_rq_of(&p->se)->next ||
4141 &p->se == cfs_rq_of(&p->se)->last))
4144 if (sysctl_sched_migration_cost == -1)
4146 if (sysctl_sched_migration_cost == 0)
4149 delta = now - p->se.exec_start;
4151 return delta < (s64)sysctl_sched_migration_cost;
4154 #ifdef CONFIG_NUMA_BALANCING
4155 /* Returns true if the destination node has incurred more faults */
4156 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4158 int src_nid, dst_nid;
4160 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4161 !(env->sd->flags & SD_NUMA)) {
4165 src_nid = cpu_to_node(env->src_cpu);
4166 dst_nid = cpu_to_node(env->dst_cpu);
4168 if (src_nid == dst_nid ||
4169 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4172 if (dst_nid == p->numa_preferred_nid ||
4173 task_faults(p, dst_nid) > task_faults(p, src_nid))
4180 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4182 int src_nid, dst_nid;
4184 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4187 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4190 src_nid = cpu_to_node(env->src_cpu);
4191 dst_nid = cpu_to_node(env->dst_cpu);
4193 if (src_nid == dst_nid ||
4194 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4197 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4204 static inline bool migrate_improves_locality(struct task_struct *p,
4210 static inline bool migrate_degrades_locality(struct task_struct *p,
4218 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4221 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4223 int tsk_cache_hot = 0;
4225 * We do not migrate tasks that are:
4226 * 1) throttled_lb_pair, or
4227 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4228 * 3) running (obviously), or
4229 * 4) are cache-hot on their current CPU.
4231 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4234 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4237 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4239 env->flags |= LBF_SOME_PINNED;
4242 * Remember if this task can be migrated to any other cpu in
4243 * our sched_group. We may want to revisit it if we couldn't
4244 * meet load balance goals by pulling other tasks on src_cpu.
4246 * Also avoid computing new_dst_cpu if we have already computed
4247 * one in current iteration.
4249 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4252 /* Prevent to re-select dst_cpu via env's cpus */
4253 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4254 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4255 env->flags |= LBF_DST_PINNED;
4256 env->new_dst_cpu = cpu;
4264 /* Record that we found atleast one task that could run on dst_cpu */
4265 env->flags &= ~LBF_ALL_PINNED;
4267 if (task_running(env->src_rq, p)) {
4268 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4273 * Aggressive migration if:
4274 * 1) destination numa is preferred
4275 * 2) task is cache cold, or
4276 * 3) too many balance attempts have failed.
4278 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4280 tsk_cache_hot = migrate_degrades_locality(p, env);
4282 if (migrate_improves_locality(p, env)) {
4283 #ifdef CONFIG_SCHEDSTATS
4284 if (tsk_cache_hot) {
4285 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4286 schedstat_inc(p, se.statistics.nr_forced_migrations);
4292 if (!tsk_cache_hot ||
4293 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4295 if (tsk_cache_hot) {
4296 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4297 schedstat_inc(p, se.statistics.nr_forced_migrations);
4303 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4308 * move_one_task tries to move exactly one task from busiest to this_rq, as
4309 * part of active balancing operations within "domain".
4310 * Returns 1 if successful and 0 otherwise.
4312 * Called with both runqueues locked.
4314 static int move_one_task(struct lb_env *env)
4316 struct task_struct *p, *n;
4318 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4319 if (!can_migrate_task(p, env))
4324 * Right now, this is only the second place move_task()
4325 * is called, so we can safely collect move_task()
4326 * stats here rather than inside move_task().
4328 schedstat_inc(env->sd, lb_gained[env->idle]);
4334 static unsigned long task_h_load(struct task_struct *p);
4336 static const unsigned int sched_nr_migrate_break = 32;
4339 * move_tasks tries to move up to imbalance weighted load from busiest to
4340 * this_rq, as part of a balancing operation within domain "sd".
4341 * Returns 1 if successful and 0 otherwise.
4343 * Called with both runqueues locked.
4345 static int move_tasks(struct lb_env *env)
4347 struct list_head *tasks = &env->src_rq->cfs_tasks;
4348 struct task_struct *p;
4352 if (env->imbalance <= 0)
4355 while (!list_empty(tasks)) {
4356 p = list_first_entry(tasks, struct task_struct, se.group_node);
4359 /* We've more or less seen every task there is, call it quits */
4360 if (env->loop > env->loop_max)
4363 /* take a breather every nr_migrate tasks */
4364 if (env->loop > env->loop_break) {
4365 env->loop_break += sched_nr_migrate_break;
4366 env->flags |= LBF_NEED_BREAK;
4370 if (!can_migrate_task(p, env))
4373 load = task_h_load(p);
4375 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4378 if ((load / 2) > env->imbalance)
4383 env->imbalance -= load;
4385 #ifdef CONFIG_PREEMPT
4387 * NEWIDLE balancing is a source of latency, so preemptible
4388 * kernels will stop after the first task is pulled to minimize
4389 * the critical section.
4391 if (env->idle == CPU_NEWLY_IDLE)
4396 * We only want to steal up to the prescribed amount of
4399 if (env->imbalance <= 0)
4404 list_move_tail(&p->se.group_node, tasks);
4408 * Right now, this is one of only two places move_task() is called,
4409 * so we can safely collect move_task() stats here rather than
4410 * inside move_task().
4412 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4417 #ifdef CONFIG_FAIR_GROUP_SCHED
4419 * update tg->load_weight by folding this cpu's load_avg
4421 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4423 struct sched_entity *se = tg->se[cpu];
4424 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4426 /* throttled entities do not contribute to load */
4427 if (throttled_hierarchy(cfs_rq))
4430 update_cfs_rq_blocked_load(cfs_rq, 1);
4433 update_entity_load_avg(se, 1);
4435 * We pivot on our runnable average having decayed to zero for
4436 * list removal. This generally implies that all our children
4437 * have also been removed (modulo rounding error or bandwidth
4438 * control); however, such cases are rare and we can fix these
4441 * TODO: fix up out-of-order children on enqueue.
4443 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4444 list_del_leaf_cfs_rq(cfs_rq);
4446 struct rq *rq = rq_of(cfs_rq);
4447 update_rq_runnable_avg(rq, rq->nr_running);
4451 static void update_blocked_averages(int cpu)
4453 struct rq *rq = cpu_rq(cpu);
4454 struct cfs_rq *cfs_rq;
4455 unsigned long flags;
4457 raw_spin_lock_irqsave(&rq->lock, flags);
4458 update_rq_clock(rq);
4460 * Iterates the task_group tree in a bottom up fashion, see
4461 * list_add_leaf_cfs_rq() for details.
4463 for_each_leaf_cfs_rq(rq, cfs_rq) {
4465 * Note: We may want to consider periodically releasing
4466 * rq->lock about these updates so that creating many task
4467 * groups does not result in continually extending hold time.
4469 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4472 raw_spin_unlock_irqrestore(&rq->lock, flags);
4476 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4477 * This needs to be done in a top-down fashion because the load of a child
4478 * group is a fraction of its parents load.
4480 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4482 struct rq *rq = rq_of(cfs_rq);
4483 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4484 unsigned long now = jiffies;
4487 if (cfs_rq->last_h_load_update == now)
4490 cfs_rq->h_load_next = NULL;
4491 for_each_sched_entity(se) {
4492 cfs_rq = cfs_rq_of(se);
4493 cfs_rq->h_load_next = se;
4494 if (cfs_rq->last_h_load_update == now)
4499 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4500 cfs_rq->last_h_load_update = now;
4503 while ((se = cfs_rq->h_load_next) != NULL) {
4504 load = cfs_rq->h_load;
4505 load = div64_ul(load * se->avg.load_avg_contrib,
4506 cfs_rq->runnable_load_avg + 1);
4507 cfs_rq = group_cfs_rq(se);
4508 cfs_rq->h_load = load;
4509 cfs_rq->last_h_load_update = now;
4513 static unsigned long task_h_load(struct task_struct *p)
4515 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4517 update_cfs_rq_h_load(cfs_rq);
4518 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4519 cfs_rq->runnable_load_avg + 1);
4522 static inline void update_blocked_averages(int cpu)
4526 static unsigned long task_h_load(struct task_struct *p)
4528 return p->se.avg.load_avg_contrib;
4532 /********** Helpers for find_busiest_group ************************/
4534 * sg_lb_stats - stats of a sched_group required for load_balancing
4536 struct sg_lb_stats {
4537 unsigned long avg_load; /*Avg load across the CPUs of the group */
4538 unsigned long group_load; /* Total load over the CPUs of the group */
4539 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4540 unsigned long load_per_task;
4541 unsigned long group_power;
4542 unsigned int sum_nr_running; /* Nr tasks running in the group */
4543 unsigned int group_capacity;
4544 unsigned int idle_cpus;
4545 unsigned int group_weight;
4546 int group_imb; /* Is there an imbalance in the group ? */
4547 int group_has_capacity; /* Is there extra capacity in the group? */
4551 * sd_lb_stats - Structure to store the statistics of a sched_domain
4552 * during load balancing.
4554 struct sd_lb_stats {
4555 struct sched_group *busiest; /* Busiest group in this sd */
4556 struct sched_group *local; /* Local group in this sd */
4557 unsigned long total_load; /* Total load of all groups in sd */
4558 unsigned long total_pwr; /* Total power of all groups in sd */
4559 unsigned long avg_load; /* Average load across all groups in sd */
4561 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4562 struct sg_lb_stats local_stat; /* Statistics of the local group */
4565 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4568 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4569 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4570 * We must however clear busiest_stat::avg_load because
4571 * update_sd_pick_busiest() reads this before assignment.
4573 *sds = (struct sd_lb_stats){
4585 * get_sd_load_idx - Obtain the load index for a given sched domain.
4586 * @sd: The sched_domain whose load_idx is to be obtained.
4587 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4589 * Return: The load index.
4591 static inline int get_sd_load_idx(struct sched_domain *sd,
4592 enum cpu_idle_type idle)
4598 load_idx = sd->busy_idx;
4601 case CPU_NEWLY_IDLE:
4602 load_idx = sd->newidle_idx;
4605 load_idx = sd->idle_idx;
4612 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4614 return SCHED_POWER_SCALE;
4617 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4619 return default_scale_freq_power(sd, cpu);
4622 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4624 unsigned long weight = sd->span_weight;
4625 unsigned long smt_gain = sd->smt_gain;
4632 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4634 return default_scale_smt_power(sd, cpu);
4637 static unsigned long scale_rt_power(int cpu)
4639 struct rq *rq = cpu_rq(cpu);
4640 u64 total, available, age_stamp, avg;
4643 * Since we're reading these variables without serialization make sure
4644 * we read them once before doing sanity checks on them.
4646 age_stamp = ACCESS_ONCE(rq->age_stamp);
4647 avg = ACCESS_ONCE(rq->rt_avg);
4649 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4651 if (unlikely(total < avg)) {
4652 /* Ensures that power won't end up being negative */
4655 available = total - avg;
4658 if (unlikely((s64)total < SCHED_POWER_SCALE))
4659 total = SCHED_POWER_SCALE;
4661 total >>= SCHED_POWER_SHIFT;
4663 return div_u64(available, total);
4666 static void update_cpu_power(struct sched_domain *sd, int cpu)
4668 unsigned long weight = sd->span_weight;
4669 unsigned long power = SCHED_POWER_SCALE;
4670 struct sched_group *sdg = sd->groups;
4672 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4673 if (sched_feat(ARCH_POWER))
4674 power *= arch_scale_smt_power(sd, cpu);
4676 power *= default_scale_smt_power(sd, cpu);
4678 power >>= SCHED_POWER_SHIFT;
4681 sdg->sgp->power_orig = power;
4683 if (sched_feat(ARCH_POWER))
4684 power *= arch_scale_freq_power(sd, cpu);
4686 power *= default_scale_freq_power(sd, cpu);
4688 power >>= SCHED_POWER_SHIFT;
4690 power *= scale_rt_power(cpu);
4691 power >>= SCHED_POWER_SHIFT;
4696 cpu_rq(cpu)->cpu_power = power;
4697 sdg->sgp->power = power;
4700 void update_group_power(struct sched_domain *sd, int cpu)
4702 struct sched_domain *child = sd->child;
4703 struct sched_group *group, *sdg = sd->groups;
4704 unsigned long power, power_orig;
4705 unsigned long interval;
4707 interval = msecs_to_jiffies(sd->balance_interval);
4708 interval = clamp(interval, 1UL, max_load_balance_interval);
4709 sdg->sgp->next_update = jiffies + interval;
4712 update_cpu_power(sd, cpu);
4716 power_orig = power = 0;
4718 if (child->flags & SD_OVERLAP) {
4720 * SD_OVERLAP domains cannot assume that child groups
4721 * span the current group.
4724 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4725 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4727 power_orig += sg->sgp->power_orig;
4728 power += sg->sgp->power;
4732 * !SD_OVERLAP domains can assume that child groups
4733 * span the current group.
4736 group = child->groups;
4738 power_orig += group->sgp->power_orig;
4739 power += group->sgp->power;
4740 group = group->next;
4741 } while (group != child->groups);
4744 sdg->sgp->power_orig = power_orig;
4745 sdg->sgp->power = power;
4749 * Try and fix up capacity for tiny siblings, this is needed when
4750 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4751 * which on its own isn't powerful enough.
4753 * See update_sd_pick_busiest() and check_asym_packing().
4756 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4759 * Only siblings can have significantly less than SCHED_POWER_SCALE
4761 if (!(sd->flags & SD_SHARE_CPUPOWER))
4765 * If ~90% of the cpu_power is still there, we're good.
4767 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4774 * Group imbalance indicates (and tries to solve) the problem where balancing
4775 * groups is inadequate due to tsk_cpus_allowed() constraints.
4777 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4778 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4781 * { 0 1 2 3 } { 4 5 6 7 }
4784 * If we were to balance group-wise we'd place two tasks in the first group and
4785 * two tasks in the second group. Clearly this is undesired as it will overload
4786 * cpu 3 and leave one of the cpus in the second group unused.
4788 * The current solution to this issue is detecting the skew in the first group
4789 * by noticing the lower domain failed to reach balance and had difficulty
4790 * moving tasks due to affinity constraints.
4792 * When this is so detected; this group becomes a candidate for busiest; see
4793 * update_sd_pick_busiest(). And calculcate_imbalance() and
4794 * find_busiest_group() avoid some of the usual balance conditions to allow it
4795 * to create an effective group imbalance.
4797 * This is a somewhat tricky proposition since the next run might not find the
4798 * group imbalance and decide the groups need to be balanced again. A most
4799 * subtle and fragile situation.
4802 static inline int sg_imbalanced(struct sched_group *group)
4804 return group->sgp->imbalance;
4808 * Compute the group capacity.
4810 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4811 * first dividing out the smt factor and computing the actual number of cores
4812 * and limit power unit capacity with that.
4814 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4816 unsigned int capacity, smt, cpus;
4817 unsigned int power, power_orig;
4819 power = group->sgp->power;
4820 power_orig = group->sgp->power_orig;
4821 cpus = group->group_weight;
4823 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4824 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4825 capacity = cpus / smt; /* cores */
4827 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4829 capacity = fix_small_capacity(env->sd, group);
4835 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4836 * @env: The load balancing environment.
4837 * @group: sched_group whose statistics are to be updated.
4838 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4839 * @local_group: Does group contain this_cpu.
4840 * @sgs: variable to hold the statistics for this group.
4842 static inline void update_sg_lb_stats(struct lb_env *env,
4843 struct sched_group *group, int load_idx,
4844 int local_group, struct sg_lb_stats *sgs)
4846 unsigned long nr_running;
4850 memset(sgs, 0, sizeof(*sgs));
4852 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4853 struct rq *rq = cpu_rq(i);
4855 nr_running = rq->nr_running;
4857 /* Bias balancing toward cpus of our domain */
4859 load = target_load(i, load_idx);
4861 load = source_load(i, load_idx);
4863 sgs->group_load += load;
4864 sgs->sum_nr_running += nr_running;
4865 sgs->sum_weighted_load += weighted_cpuload(i);
4870 /* Adjust by relative CPU power of the group */
4871 sgs->group_power = group->sgp->power;
4872 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4874 if (sgs->sum_nr_running)
4875 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4877 sgs->group_weight = group->group_weight;
4879 sgs->group_imb = sg_imbalanced(group);
4880 sgs->group_capacity = sg_capacity(env, group);
4882 if (sgs->group_capacity > sgs->sum_nr_running)
4883 sgs->group_has_capacity = 1;
4887 * update_sd_pick_busiest - return 1 on busiest group
4888 * @env: The load balancing environment.
4889 * @sds: sched_domain statistics
4890 * @sg: sched_group candidate to be checked for being the busiest
4891 * @sgs: sched_group statistics
4893 * Determine if @sg is a busier group than the previously selected
4896 * Return: %true if @sg is a busier group than the previously selected
4897 * busiest group. %false otherwise.
4899 static bool update_sd_pick_busiest(struct lb_env *env,
4900 struct sd_lb_stats *sds,
4901 struct sched_group *sg,
4902 struct sg_lb_stats *sgs)
4904 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4907 if (sgs->sum_nr_running > sgs->group_capacity)
4914 * ASYM_PACKING needs to move all the work to the lowest
4915 * numbered CPUs in the group, therefore mark all groups
4916 * higher than ourself as busy.
4918 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4919 env->dst_cpu < group_first_cpu(sg)) {
4923 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4931 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4932 * @env: The load balancing environment.
4933 * @balance: Should we balance.
4934 * @sds: variable to hold the statistics for this sched_domain.
4936 static inline void update_sd_lb_stats(struct lb_env *env,
4937 struct sd_lb_stats *sds)
4939 struct sched_domain *child = env->sd->child;
4940 struct sched_group *sg = env->sd->groups;
4941 struct sg_lb_stats tmp_sgs;
4942 int load_idx, prefer_sibling = 0;
4944 if (child && child->flags & SD_PREFER_SIBLING)
4947 load_idx = get_sd_load_idx(env->sd, env->idle);
4950 struct sg_lb_stats *sgs = &tmp_sgs;
4953 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4956 sgs = &sds->local_stat;
4958 if (env->idle != CPU_NEWLY_IDLE ||
4959 time_after_eq(jiffies, sg->sgp->next_update))
4960 update_group_power(env->sd, env->dst_cpu);
4963 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4969 * In case the child domain prefers tasks go to siblings
4970 * first, lower the sg capacity to one so that we'll try
4971 * and move all the excess tasks away. We lower the capacity
4972 * of a group only if the local group has the capacity to fit
4973 * these excess tasks, i.e. nr_running < group_capacity. The
4974 * extra check prevents the case where you always pull from the
4975 * heaviest group when it is already under-utilized (possible
4976 * with a large weight task outweighs the tasks on the system).
4978 if (prefer_sibling && sds->local &&
4979 sds->local_stat.group_has_capacity)
4980 sgs->group_capacity = min(sgs->group_capacity, 1U);
4982 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4984 sds->busiest_stat = *sgs;
4988 /* Now, start updating sd_lb_stats */
4989 sds->total_load += sgs->group_load;
4990 sds->total_pwr += sgs->group_power;
4993 } while (sg != env->sd->groups);
4997 * check_asym_packing - Check to see if the group is packed into the
5000 * This is primarily intended to used at the sibling level. Some
5001 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5002 * case of POWER7, it can move to lower SMT modes only when higher
5003 * threads are idle. When in lower SMT modes, the threads will
5004 * perform better since they share less core resources. Hence when we
5005 * have idle threads, we want them to be the higher ones.
5007 * This packing function is run on idle threads. It checks to see if
5008 * the busiest CPU in this domain (core in the P7 case) has a higher
5009 * CPU number than the packing function is being run on. Here we are
5010 * assuming lower CPU number will be equivalent to lower a SMT thread
5013 * Return: 1 when packing is required and a task should be moved to
5014 * this CPU. The amount of the imbalance is returned in *imbalance.
5016 * @env: The load balancing environment.
5017 * @sds: Statistics of the sched_domain which is to be packed
5019 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5023 if (!(env->sd->flags & SD_ASYM_PACKING))
5029 busiest_cpu = group_first_cpu(sds->busiest);
5030 if (env->dst_cpu > busiest_cpu)
5033 env->imbalance = DIV_ROUND_CLOSEST(
5034 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5041 * fix_small_imbalance - Calculate the minor imbalance that exists
5042 * amongst the groups of a sched_domain, during
5044 * @env: The load balancing environment.
5045 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5048 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5050 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5051 unsigned int imbn = 2;
5052 unsigned long scaled_busy_load_per_task;
5053 struct sg_lb_stats *local, *busiest;
5055 local = &sds->local_stat;
5056 busiest = &sds->busiest_stat;
5058 if (!local->sum_nr_running)
5059 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5060 else if (busiest->load_per_task > local->load_per_task)
5063 scaled_busy_load_per_task =
5064 (busiest->load_per_task * SCHED_POWER_SCALE) /
5065 busiest->group_power;
5067 if (busiest->avg_load + scaled_busy_load_per_task >=
5068 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5069 env->imbalance = busiest->load_per_task;
5074 * OK, we don't have enough imbalance to justify moving tasks,
5075 * however we may be able to increase total CPU power used by
5079 pwr_now += busiest->group_power *
5080 min(busiest->load_per_task, busiest->avg_load);
5081 pwr_now += local->group_power *
5082 min(local->load_per_task, local->avg_load);
5083 pwr_now /= SCHED_POWER_SCALE;
5085 /* Amount of load we'd subtract */
5086 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5087 busiest->group_power;
5088 if (busiest->avg_load > tmp) {
5089 pwr_move += busiest->group_power *
5090 min(busiest->load_per_task,
5091 busiest->avg_load - tmp);
5094 /* Amount of load we'd add */
5095 if (busiest->avg_load * busiest->group_power <
5096 busiest->load_per_task * SCHED_POWER_SCALE) {
5097 tmp = (busiest->avg_load * busiest->group_power) /
5100 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5103 pwr_move += local->group_power *
5104 min(local->load_per_task, local->avg_load + tmp);
5105 pwr_move /= SCHED_POWER_SCALE;
5107 /* Move if we gain throughput */
5108 if (pwr_move > pwr_now)
5109 env->imbalance = busiest->load_per_task;
5113 * calculate_imbalance - Calculate the amount of imbalance present within the
5114 * groups of a given sched_domain during load balance.
5115 * @env: load balance environment
5116 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5118 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5120 unsigned long max_pull, load_above_capacity = ~0UL;
5121 struct sg_lb_stats *local, *busiest;
5123 local = &sds->local_stat;
5124 busiest = &sds->busiest_stat;
5126 if (busiest->group_imb) {
5128 * In the group_imb case we cannot rely on group-wide averages
5129 * to ensure cpu-load equilibrium, look at wider averages. XXX
5131 busiest->load_per_task =
5132 min(busiest->load_per_task, sds->avg_load);
5136 * In the presence of smp nice balancing, certain scenarios can have
5137 * max load less than avg load(as we skip the groups at or below
5138 * its cpu_power, while calculating max_load..)
5140 if (busiest->avg_load <= sds->avg_load ||
5141 local->avg_load >= sds->avg_load) {
5143 return fix_small_imbalance(env, sds);
5146 if (!busiest->group_imb) {
5148 * Don't want to pull so many tasks that a group would go idle.
5149 * Except of course for the group_imb case, since then we might
5150 * have to drop below capacity to reach cpu-load equilibrium.
5152 load_above_capacity =
5153 (busiest->sum_nr_running - busiest->group_capacity);
5155 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5156 load_above_capacity /= busiest->group_power;
5160 * We're trying to get all the cpus to the average_load, so we don't
5161 * want to push ourselves above the average load, nor do we wish to
5162 * reduce the max loaded cpu below the average load. At the same time,
5163 * we also don't want to reduce the group load below the group capacity
5164 * (so that we can implement power-savings policies etc). Thus we look
5165 * for the minimum possible imbalance.
5167 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5169 /* How much load to actually move to equalise the imbalance */
5170 env->imbalance = min(
5171 max_pull * busiest->group_power,
5172 (sds->avg_load - local->avg_load) * local->group_power
5173 ) / SCHED_POWER_SCALE;
5176 * if *imbalance is less than the average load per runnable task
5177 * there is no guarantee that any tasks will be moved so we'll have
5178 * a think about bumping its value to force at least one task to be
5181 if (env->imbalance < busiest->load_per_task)
5182 return fix_small_imbalance(env, sds);
5185 /******* find_busiest_group() helpers end here *********************/
5188 * find_busiest_group - Returns the busiest group within the sched_domain
5189 * if there is an imbalance. If there isn't an imbalance, and
5190 * the user has opted for power-savings, it returns a group whose
5191 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5192 * such a group exists.
5194 * Also calculates the amount of weighted load which should be moved
5195 * to restore balance.
5197 * @env: The load balancing environment.
5199 * Return: - The busiest group if imbalance exists.
5200 * - If no imbalance and user has opted for power-savings balance,
5201 * return the least loaded group whose CPUs can be
5202 * put to idle by rebalancing its tasks onto our group.
5204 static struct sched_group *find_busiest_group(struct lb_env *env)
5206 struct sg_lb_stats *local, *busiest;
5207 struct sd_lb_stats sds;
5209 init_sd_lb_stats(&sds);
5212 * Compute the various statistics relavent for load balancing at
5215 update_sd_lb_stats(env, &sds);
5216 local = &sds.local_stat;
5217 busiest = &sds.busiest_stat;
5219 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5220 check_asym_packing(env, &sds))
5223 /* There is no busy sibling group to pull tasks from */
5224 if (!sds.busiest || busiest->sum_nr_running == 0)
5227 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5230 * If the busiest group is imbalanced the below checks don't
5231 * work because they assume all things are equal, which typically
5232 * isn't true due to cpus_allowed constraints and the like.
5234 if (busiest->group_imb)
5237 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5238 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5239 !busiest->group_has_capacity)
5243 * If the local group is more busy than the selected busiest group
5244 * don't try and pull any tasks.
5246 if (local->avg_load >= busiest->avg_load)
5250 * Don't pull any tasks if this group is already above the domain
5253 if (local->avg_load >= sds.avg_load)
5256 if (env->idle == CPU_IDLE) {
5258 * This cpu is idle. If the busiest group load doesn't
5259 * have more tasks than the number of available cpu's and
5260 * there is no imbalance between this and busiest group
5261 * wrt to idle cpu's, it is balanced.
5263 if ((local->idle_cpus < busiest->idle_cpus) &&
5264 busiest->sum_nr_running <= busiest->group_weight)
5268 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5269 * imbalance_pct to be conservative.
5271 if (100 * busiest->avg_load <=
5272 env->sd->imbalance_pct * local->avg_load)
5277 /* Looks like there is an imbalance. Compute it */
5278 calculate_imbalance(env, &sds);
5287 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5289 static struct rq *find_busiest_queue(struct lb_env *env,
5290 struct sched_group *group)
5292 struct rq *busiest = NULL, *rq;
5293 unsigned long busiest_load = 0, busiest_power = 1;
5296 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5297 unsigned long power = power_of(i);
5298 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5303 capacity = fix_small_capacity(env->sd, group);
5306 wl = weighted_cpuload(i);
5309 * When comparing with imbalance, use weighted_cpuload()
5310 * which is not scaled with the cpu power.
5312 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5316 * For the load comparisons with the other cpu's, consider
5317 * the weighted_cpuload() scaled with the cpu power, so that
5318 * the load can be moved away from the cpu that is potentially
5319 * running at a lower capacity.
5321 * Thus we're looking for max(wl_i / power_i), crosswise
5322 * multiplication to rid ourselves of the division works out
5323 * to: wl_i * power_j > wl_j * power_i; where j is our
5326 if (wl * busiest_power > busiest_load * power) {
5328 busiest_power = power;
5337 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5338 * so long as it is large enough.
5340 #define MAX_PINNED_INTERVAL 512
5342 /* Working cpumask for load_balance and load_balance_newidle. */
5343 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5345 static int need_active_balance(struct lb_env *env)
5347 struct sched_domain *sd = env->sd;
5349 if (env->idle == CPU_NEWLY_IDLE) {
5352 * ASYM_PACKING needs to force migrate tasks from busy but
5353 * higher numbered CPUs in order to pack all tasks in the
5354 * lowest numbered CPUs.
5356 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5360 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5363 static int active_load_balance_cpu_stop(void *data);
5365 static int should_we_balance(struct lb_env *env)
5367 struct sched_group *sg = env->sd->groups;
5368 struct cpumask *sg_cpus, *sg_mask;
5369 int cpu, balance_cpu = -1;
5372 * In the newly idle case, we will allow all the cpu's
5373 * to do the newly idle load balance.
5375 if (env->idle == CPU_NEWLY_IDLE)
5378 sg_cpus = sched_group_cpus(sg);
5379 sg_mask = sched_group_mask(sg);
5380 /* Try to find first idle cpu */
5381 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5382 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5389 if (balance_cpu == -1)
5390 balance_cpu = group_balance_cpu(sg);
5393 * First idle cpu or the first cpu(busiest) in this sched group
5394 * is eligible for doing load balancing at this and above domains.
5396 return balance_cpu == env->dst_cpu;
5400 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5401 * tasks if there is an imbalance.
5403 static int load_balance(int this_cpu, struct rq *this_rq,
5404 struct sched_domain *sd, enum cpu_idle_type idle,
5405 int *continue_balancing)
5407 int ld_moved, cur_ld_moved, active_balance = 0;
5408 struct sched_domain *sd_parent = sd->parent;
5409 struct sched_group *group;
5411 unsigned long flags;
5412 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5414 struct lb_env env = {
5416 .dst_cpu = this_cpu,
5418 .dst_grpmask = sched_group_cpus(sd->groups),
5420 .loop_break = sched_nr_migrate_break,
5425 * For NEWLY_IDLE load_balancing, we don't need to consider
5426 * other cpus in our group
5428 if (idle == CPU_NEWLY_IDLE)
5429 env.dst_grpmask = NULL;
5431 cpumask_copy(cpus, cpu_active_mask);
5433 schedstat_inc(sd, lb_count[idle]);
5436 if (!should_we_balance(&env)) {
5437 *continue_balancing = 0;
5441 group = find_busiest_group(&env);
5443 schedstat_inc(sd, lb_nobusyg[idle]);
5447 busiest = find_busiest_queue(&env, group);
5449 schedstat_inc(sd, lb_nobusyq[idle]);
5453 BUG_ON(busiest == env.dst_rq);
5455 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5458 if (busiest->nr_running > 1) {
5460 * Attempt to move tasks. If find_busiest_group has found
5461 * an imbalance but busiest->nr_running <= 1, the group is
5462 * still unbalanced. ld_moved simply stays zero, so it is
5463 * correctly treated as an imbalance.
5465 env.flags |= LBF_ALL_PINNED;
5466 env.src_cpu = busiest->cpu;
5467 env.src_rq = busiest;
5468 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5471 local_irq_save(flags);
5472 double_rq_lock(env.dst_rq, busiest);
5475 * cur_ld_moved - load moved in current iteration
5476 * ld_moved - cumulative load moved across iterations
5478 cur_ld_moved = move_tasks(&env);
5479 ld_moved += cur_ld_moved;
5480 double_rq_unlock(env.dst_rq, busiest);
5481 local_irq_restore(flags);
5484 * some other cpu did the load balance for us.
5486 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5487 resched_cpu(env.dst_cpu);
5489 if (env.flags & LBF_NEED_BREAK) {
5490 env.flags &= ~LBF_NEED_BREAK;
5495 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5496 * us and move them to an alternate dst_cpu in our sched_group
5497 * where they can run. The upper limit on how many times we
5498 * iterate on same src_cpu is dependent on number of cpus in our
5501 * This changes load balance semantics a bit on who can move
5502 * load to a given_cpu. In addition to the given_cpu itself
5503 * (or a ilb_cpu acting on its behalf where given_cpu is
5504 * nohz-idle), we now have balance_cpu in a position to move
5505 * load to given_cpu. In rare situations, this may cause
5506 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5507 * _independently_ and at _same_ time to move some load to
5508 * given_cpu) causing exceess load to be moved to given_cpu.
5509 * This however should not happen so much in practice and
5510 * moreover subsequent load balance cycles should correct the
5511 * excess load moved.
5513 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5515 /* Prevent to re-select dst_cpu via env's cpus */
5516 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5518 env.dst_rq = cpu_rq(env.new_dst_cpu);
5519 env.dst_cpu = env.new_dst_cpu;
5520 env.flags &= ~LBF_DST_PINNED;
5522 env.loop_break = sched_nr_migrate_break;
5525 * Go back to "more_balance" rather than "redo" since we
5526 * need to continue with same src_cpu.
5532 * We failed to reach balance because of affinity.
5535 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5537 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5538 *group_imbalance = 1;
5539 } else if (*group_imbalance)
5540 *group_imbalance = 0;
5543 /* All tasks on this runqueue were pinned by CPU affinity */
5544 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5545 cpumask_clear_cpu(cpu_of(busiest), cpus);
5546 if (!cpumask_empty(cpus)) {
5548 env.loop_break = sched_nr_migrate_break;
5556 schedstat_inc(sd, lb_failed[idle]);
5558 * Increment the failure counter only on periodic balance.
5559 * We do not want newidle balance, which can be very
5560 * frequent, pollute the failure counter causing
5561 * excessive cache_hot migrations and active balances.
5563 if (idle != CPU_NEWLY_IDLE)
5564 sd->nr_balance_failed++;
5566 if (need_active_balance(&env)) {
5567 raw_spin_lock_irqsave(&busiest->lock, flags);
5569 /* don't kick the active_load_balance_cpu_stop,
5570 * if the curr task on busiest cpu can't be
5573 if (!cpumask_test_cpu(this_cpu,
5574 tsk_cpus_allowed(busiest->curr))) {
5575 raw_spin_unlock_irqrestore(&busiest->lock,
5577 env.flags |= LBF_ALL_PINNED;
5578 goto out_one_pinned;
5582 * ->active_balance synchronizes accesses to
5583 * ->active_balance_work. Once set, it's cleared
5584 * only after active load balance is finished.
5586 if (!busiest->active_balance) {
5587 busiest->active_balance = 1;
5588 busiest->push_cpu = this_cpu;
5591 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5593 if (active_balance) {
5594 stop_one_cpu_nowait(cpu_of(busiest),
5595 active_load_balance_cpu_stop, busiest,
5596 &busiest->active_balance_work);
5600 * We've kicked active balancing, reset the failure
5603 sd->nr_balance_failed = sd->cache_nice_tries+1;
5606 sd->nr_balance_failed = 0;
5608 if (likely(!active_balance)) {
5609 /* We were unbalanced, so reset the balancing interval */
5610 sd->balance_interval = sd->min_interval;
5613 * If we've begun active balancing, start to back off. This
5614 * case may not be covered by the all_pinned logic if there
5615 * is only 1 task on the busy runqueue (because we don't call
5618 if (sd->balance_interval < sd->max_interval)
5619 sd->balance_interval *= 2;
5625 schedstat_inc(sd, lb_balanced[idle]);
5627 sd->nr_balance_failed = 0;
5630 /* tune up the balancing interval */
5631 if (((env.flags & LBF_ALL_PINNED) &&
5632 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5633 (sd->balance_interval < sd->max_interval))
5634 sd->balance_interval *= 2;
5642 * idle_balance is called by schedule() if this_cpu is about to become
5643 * idle. Attempts to pull tasks from other CPUs.
5645 void idle_balance(int this_cpu, struct rq *this_rq)
5647 struct sched_domain *sd;
5648 int pulled_task = 0;
5649 unsigned long next_balance = jiffies + HZ;
5652 this_rq->idle_stamp = rq_clock(this_rq);
5654 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5658 * Drop the rq->lock, but keep IRQ/preempt disabled.
5660 raw_spin_unlock(&this_rq->lock);
5662 update_blocked_averages(this_cpu);
5664 for_each_domain(this_cpu, sd) {
5665 unsigned long interval;
5666 int continue_balancing = 1;
5667 u64 t0, domain_cost;
5669 if (!(sd->flags & SD_LOAD_BALANCE))
5672 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5675 if (sd->flags & SD_BALANCE_NEWIDLE) {
5676 t0 = sched_clock_cpu(this_cpu);
5678 /* If we've pulled tasks over stop searching: */
5679 pulled_task = load_balance(this_cpu, this_rq,
5681 &continue_balancing);
5683 domain_cost = sched_clock_cpu(this_cpu) - t0;
5684 if (domain_cost > sd->max_newidle_lb_cost)
5685 sd->max_newidle_lb_cost = domain_cost;
5687 curr_cost += domain_cost;
5690 interval = msecs_to_jiffies(sd->balance_interval);
5691 if (time_after(next_balance, sd->last_balance + interval))
5692 next_balance = sd->last_balance + interval;
5694 this_rq->idle_stamp = 0;
5700 raw_spin_lock(&this_rq->lock);
5702 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5704 * We are going idle. next_balance may be set based on
5705 * a busy processor. So reset next_balance.
5707 this_rq->next_balance = next_balance;
5710 if (curr_cost > this_rq->max_idle_balance_cost)
5711 this_rq->max_idle_balance_cost = curr_cost;
5715 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5716 * running tasks off the busiest CPU onto idle CPUs. It requires at
5717 * least 1 task to be running on each physical CPU where possible, and
5718 * avoids physical / logical imbalances.
5720 static int active_load_balance_cpu_stop(void *data)
5722 struct rq *busiest_rq = data;
5723 int busiest_cpu = cpu_of(busiest_rq);
5724 int target_cpu = busiest_rq->push_cpu;
5725 struct rq *target_rq = cpu_rq(target_cpu);
5726 struct sched_domain *sd;
5728 raw_spin_lock_irq(&busiest_rq->lock);
5730 /* make sure the requested cpu hasn't gone down in the meantime */
5731 if (unlikely(busiest_cpu != smp_processor_id() ||
5732 !busiest_rq->active_balance))
5735 /* Is there any task to move? */
5736 if (busiest_rq->nr_running <= 1)
5740 * This condition is "impossible", if it occurs
5741 * we need to fix it. Originally reported by
5742 * Bjorn Helgaas on a 128-cpu setup.
5744 BUG_ON(busiest_rq == target_rq);
5746 /* move a task from busiest_rq to target_rq */
5747 double_lock_balance(busiest_rq, target_rq);
5749 /* Search for an sd spanning us and the target CPU. */
5751 for_each_domain(target_cpu, sd) {
5752 if ((sd->flags & SD_LOAD_BALANCE) &&
5753 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5758 struct lb_env env = {
5760 .dst_cpu = target_cpu,
5761 .dst_rq = target_rq,
5762 .src_cpu = busiest_rq->cpu,
5763 .src_rq = busiest_rq,
5767 schedstat_inc(sd, alb_count);
5769 if (move_one_task(&env))
5770 schedstat_inc(sd, alb_pushed);
5772 schedstat_inc(sd, alb_failed);
5775 double_unlock_balance(busiest_rq, target_rq);
5777 busiest_rq->active_balance = 0;
5778 raw_spin_unlock_irq(&busiest_rq->lock);
5782 #ifdef CONFIG_NO_HZ_COMMON
5784 * idle load balancing details
5785 * - When one of the busy CPUs notice that there may be an idle rebalancing
5786 * needed, they will kick the idle load balancer, which then does idle
5787 * load balancing for all the idle CPUs.
5790 cpumask_var_t idle_cpus_mask;
5792 unsigned long next_balance; /* in jiffy units */
5793 } nohz ____cacheline_aligned;
5795 static inline int find_new_ilb(int call_cpu)
5797 int ilb = cpumask_first(nohz.idle_cpus_mask);
5799 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5806 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5807 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5808 * CPU (if there is one).
5810 static void nohz_balancer_kick(int cpu)
5814 nohz.next_balance++;
5816 ilb_cpu = find_new_ilb(cpu);
5818 if (ilb_cpu >= nr_cpu_ids)
5821 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5824 * Use smp_send_reschedule() instead of resched_cpu().
5825 * This way we generate a sched IPI on the target cpu which
5826 * is idle. And the softirq performing nohz idle load balance
5827 * will be run before returning from the IPI.
5829 smp_send_reschedule(ilb_cpu);
5833 static inline void nohz_balance_exit_idle(int cpu)
5835 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5836 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5837 atomic_dec(&nohz.nr_cpus);
5838 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5842 static inline void set_cpu_sd_state_busy(void)
5844 struct sched_domain *sd;
5847 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5849 if (!sd || !sd->nohz_idle)
5853 for (; sd; sd = sd->parent)
5854 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5859 void set_cpu_sd_state_idle(void)
5861 struct sched_domain *sd;
5864 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5866 if (!sd || sd->nohz_idle)
5870 for (; sd; sd = sd->parent)
5871 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5877 * This routine will record that the cpu is going idle with tick stopped.
5878 * This info will be used in performing idle load balancing in the future.
5880 void nohz_balance_enter_idle(int cpu)
5883 * If this cpu is going down, then nothing needs to be done.
5885 if (!cpu_active(cpu))
5888 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5891 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5892 atomic_inc(&nohz.nr_cpus);
5893 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5896 static int sched_ilb_notifier(struct notifier_block *nfb,
5897 unsigned long action, void *hcpu)
5899 switch (action & ~CPU_TASKS_FROZEN) {
5901 nohz_balance_exit_idle(smp_processor_id());
5909 static DEFINE_SPINLOCK(balancing);
5912 * Scale the max load_balance interval with the number of CPUs in the system.
5913 * This trades load-balance latency on larger machines for less cross talk.
5915 void update_max_interval(void)
5917 max_load_balance_interval = HZ*num_online_cpus()/10;
5921 * It checks each scheduling domain to see if it is due to be balanced,
5922 * and initiates a balancing operation if so.
5924 * Balancing parameters are set up in init_sched_domains.
5926 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5928 int continue_balancing = 1;
5929 struct rq *rq = cpu_rq(cpu);
5930 unsigned long interval;
5931 struct sched_domain *sd;
5932 /* Earliest time when we have to do rebalance again */
5933 unsigned long next_balance = jiffies + 60*HZ;
5934 int update_next_balance = 0;
5935 int need_serialize, need_decay = 0;
5938 update_blocked_averages(cpu);
5941 for_each_domain(cpu, sd) {
5943 * Decay the newidle max times here because this is a regular
5944 * visit to all the domains. Decay ~1% per second.
5946 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5947 sd->max_newidle_lb_cost =
5948 (sd->max_newidle_lb_cost * 253) / 256;
5949 sd->next_decay_max_lb_cost = jiffies + HZ;
5952 max_cost += sd->max_newidle_lb_cost;
5954 if (!(sd->flags & SD_LOAD_BALANCE))
5958 * Stop the load balance at this level. There is another
5959 * CPU in our sched group which is doing load balancing more
5962 if (!continue_balancing) {
5968 interval = sd->balance_interval;
5969 if (idle != CPU_IDLE)
5970 interval *= sd->busy_factor;
5972 /* scale ms to jiffies */
5973 interval = msecs_to_jiffies(interval);
5974 interval = clamp(interval, 1UL, max_load_balance_interval);
5976 need_serialize = sd->flags & SD_SERIALIZE;
5978 if (need_serialize) {
5979 if (!spin_trylock(&balancing))
5983 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5984 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5986 * The LBF_DST_PINNED logic could have changed
5987 * env->dst_cpu, so we can't know our idle
5988 * state even if we migrated tasks. Update it.
5990 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5992 sd->last_balance = jiffies;
5995 spin_unlock(&balancing);
5997 if (time_after(next_balance, sd->last_balance + interval)) {
5998 next_balance = sd->last_balance + interval;
5999 update_next_balance = 1;
6004 * Ensure the rq-wide value also decays but keep it at a
6005 * reasonable floor to avoid funnies with rq->avg_idle.
6007 rq->max_idle_balance_cost =
6008 max((u64)sysctl_sched_migration_cost, max_cost);
6013 * next_balance will be updated only when there is a need.
6014 * When the cpu is attached to null domain for ex, it will not be
6017 if (likely(update_next_balance))
6018 rq->next_balance = next_balance;
6021 #ifdef CONFIG_NO_HZ_COMMON
6023 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6024 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6026 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6028 struct rq *this_rq = cpu_rq(this_cpu);
6032 if (idle != CPU_IDLE ||
6033 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6036 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6037 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6041 * If this cpu gets work to do, stop the load balancing
6042 * work being done for other cpus. Next load
6043 * balancing owner will pick it up.
6048 rq = cpu_rq(balance_cpu);
6050 raw_spin_lock_irq(&rq->lock);
6051 update_rq_clock(rq);
6052 update_idle_cpu_load(rq);
6053 raw_spin_unlock_irq(&rq->lock);
6055 rebalance_domains(balance_cpu, CPU_IDLE);
6057 if (time_after(this_rq->next_balance, rq->next_balance))
6058 this_rq->next_balance = rq->next_balance;
6060 nohz.next_balance = this_rq->next_balance;
6062 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6066 * Current heuristic for kicking the idle load balancer in the presence
6067 * of an idle cpu is the system.
6068 * - This rq has more than one task.
6069 * - At any scheduler domain level, this cpu's scheduler group has multiple
6070 * busy cpu's exceeding the group's power.
6071 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6072 * domain span are idle.
6074 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6076 unsigned long now = jiffies;
6077 struct sched_domain *sd;
6079 if (unlikely(idle_cpu(cpu)))
6083 * We may be recently in ticked or tickless idle mode. At the first
6084 * busy tick after returning from idle, we will update the busy stats.
6086 set_cpu_sd_state_busy();
6087 nohz_balance_exit_idle(cpu);
6090 * None are in tickless mode and hence no need for NOHZ idle load
6093 if (likely(!atomic_read(&nohz.nr_cpus)))
6096 if (time_before(now, nohz.next_balance))
6099 if (rq->nr_running >= 2)
6103 for_each_domain(cpu, sd) {
6104 struct sched_group *sg = sd->groups;
6105 struct sched_group_power *sgp = sg->sgp;
6106 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6108 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6109 goto need_kick_unlock;
6111 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6112 && (cpumask_first_and(nohz.idle_cpus_mask,
6113 sched_domain_span(sd)) < cpu))
6114 goto need_kick_unlock;
6116 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6128 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6132 * run_rebalance_domains is triggered when needed from the scheduler tick.
6133 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6135 static void run_rebalance_domains(struct softirq_action *h)
6137 int this_cpu = smp_processor_id();
6138 struct rq *this_rq = cpu_rq(this_cpu);
6139 enum cpu_idle_type idle = this_rq->idle_balance ?
6140 CPU_IDLE : CPU_NOT_IDLE;
6142 rebalance_domains(this_cpu, idle);
6145 * If this cpu has a pending nohz_balance_kick, then do the
6146 * balancing on behalf of the other idle cpus whose ticks are
6149 nohz_idle_balance(this_cpu, idle);
6152 static inline int on_null_domain(int cpu)
6154 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6158 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6160 void trigger_load_balance(struct rq *rq, int cpu)
6162 /* Don't need to rebalance while attached to NULL domain */
6163 if (time_after_eq(jiffies, rq->next_balance) &&
6164 likely(!on_null_domain(cpu)))
6165 raise_softirq(SCHED_SOFTIRQ);
6166 #ifdef CONFIG_NO_HZ_COMMON
6167 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6168 nohz_balancer_kick(cpu);
6172 static void rq_online_fair(struct rq *rq)
6177 static void rq_offline_fair(struct rq *rq)
6181 /* Ensure any throttled groups are reachable by pick_next_task */
6182 unthrottle_offline_cfs_rqs(rq);
6185 #endif /* CONFIG_SMP */
6188 * scheduler tick hitting a task of our scheduling class:
6190 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6192 struct cfs_rq *cfs_rq;
6193 struct sched_entity *se = &curr->se;
6195 for_each_sched_entity(se) {
6196 cfs_rq = cfs_rq_of(se);
6197 entity_tick(cfs_rq, se, queued);
6200 if (numabalancing_enabled)
6201 task_tick_numa(rq, curr);
6203 update_rq_runnable_avg(rq, 1);
6207 * called on fork with the child task as argument from the parent's context
6208 * - child not yet on the tasklist
6209 * - preemption disabled
6211 static void task_fork_fair(struct task_struct *p)
6213 struct cfs_rq *cfs_rq;
6214 struct sched_entity *se = &p->se, *curr;
6215 int this_cpu = smp_processor_id();
6216 struct rq *rq = this_rq();
6217 unsigned long flags;
6219 raw_spin_lock_irqsave(&rq->lock, flags);
6221 update_rq_clock(rq);
6223 cfs_rq = task_cfs_rq(current);
6224 curr = cfs_rq->curr;
6227 * Not only the cpu but also the task_group of the parent might have
6228 * been changed after parent->se.parent,cfs_rq were copied to
6229 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6230 * of child point to valid ones.
6233 __set_task_cpu(p, this_cpu);
6236 update_curr(cfs_rq);
6239 se->vruntime = curr->vruntime;
6240 place_entity(cfs_rq, se, 1);
6242 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6244 * Upon rescheduling, sched_class::put_prev_task() will place
6245 * 'current' within the tree based on its new key value.
6247 swap(curr->vruntime, se->vruntime);
6248 resched_task(rq->curr);
6251 se->vruntime -= cfs_rq->min_vruntime;
6253 raw_spin_unlock_irqrestore(&rq->lock, flags);
6257 * Priority of the task has changed. Check to see if we preempt
6261 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6267 * Reschedule if we are currently running on this runqueue and
6268 * our priority decreased, or if we are not currently running on
6269 * this runqueue and our priority is higher than the current's
6271 if (rq->curr == p) {
6272 if (p->prio > oldprio)
6273 resched_task(rq->curr);
6275 check_preempt_curr(rq, p, 0);
6278 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6280 struct sched_entity *se = &p->se;
6281 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6284 * Ensure the task's vruntime is normalized, so that when its
6285 * switched back to the fair class the enqueue_entity(.flags=0) will
6286 * do the right thing.
6288 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6289 * have normalized the vruntime, if it was !on_rq, then only when
6290 * the task is sleeping will it still have non-normalized vruntime.
6292 if (!se->on_rq && p->state != TASK_RUNNING) {
6294 * Fix up our vruntime so that the current sleep doesn't
6295 * cause 'unlimited' sleep bonus.
6297 place_entity(cfs_rq, se, 0);
6298 se->vruntime -= cfs_rq->min_vruntime;
6303 * Remove our load from contribution when we leave sched_fair
6304 * and ensure we don't carry in an old decay_count if we
6307 if (se->avg.decay_count) {
6308 __synchronize_entity_decay(se);
6309 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6315 * We switched to the sched_fair class.
6317 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6323 * We were most likely switched from sched_rt, so
6324 * kick off the schedule if running, otherwise just see
6325 * if we can still preempt the current task.
6328 resched_task(rq->curr);
6330 check_preempt_curr(rq, p, 0);
6333 /* Account for a task changing its policy or group.
6335 * This routine is mostly called to set cfs_rq->curr field when a task
6336 * migrates between groups/classes.
6338 static void set_curr_task_fair(struct rq *rq)
6340 struct sched_entity *se = &rq->curr->se;
6342 for_each_sched_entity(se) {
6343 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6345 set_next_entity(cfs_rq, se);
6346 /* ensure bandwidth has been allocated on our new cfs_rq */
6347 account_cfs_rq_runtime(cfs_rq, 0);
6351 void init_cfs_rq(struct cfs_rq *cfs_rq)
6353 cfs_rq->tasks_timeline = RB_ROOT;
6354 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6355 #ifndef CONFIG_64BIT
6356 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6359 atomic64_set(&cfs_rq->decay_counter, 1);
6360 atomic_long_set(&cfs_rq->removed_load, 0);
6364 #ifdef CONFIG_FAIR_GROUP_SCHED
6365 static void task_move_group_fair(struct task_struct *p, int on_rq)
6367 struct cfs_rq *cfs_rq;
6369 * If the task was not on the rq at the time of this cgroup movement
6370 * it must have been asleep, sleeping tasks keep their ->vruntime
6371 * absolute on their old rq until wakeup (needed for the fair sleeper
6372 * bonus in place_entity()).
6374 * If it was on the rq, we've just 'preempted' it, which does convert
6375 * ->vruntime to a relative base.
6377 * Make sure both cases convert their relative position when migrating
6378 * to another cgroup's rq. This does somewhat interfere with the
6379 * fair sleeper stuff for the first placement, but who cares.
6382 * When !on_rq, vruntime of the task has usually NOT been normalized.
6383 * But there are some cases where it has already been normalized:
6385 * - Moving a forked child which is waiting for being woken up by
6386 * wake_up_new_task().
6387 * - Moving a task which has been woken up by try_to_wake_up() and
6388 * waiting for actually being woken up by sched_ttwu_pending().
6390 * To prevent boost or penalty in the new cfs_rq caused by delta
6391 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6393 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6397 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6398 set_task_rq(p, task_cpu(p));
6400 cfs_rq = cfs_rq_of(&p->se);
6401 p->se.vruntime += cfs_rq->min_vruntime;
6404 * migrate_task_rq_fair() will have removed our previous
6405 * contribution, but we must synchronize for ongoing future
6408 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6409 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6414 void free_fair_sched_group(struct task_group *tg)
6418 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6420 for_each_possible_cpu(i) {
6422 kfree(tg->cfs_rq[i]);
6431 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6433 struct cfs_rq *cfs_rq;
6434 struct sched_entity *se;
6437 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6440 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6444 tg->shares = NICE_0_LOAD;
6446 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6448 for_each_possible_cpu(i) {
6449 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6450 GFP_KERNEL, cpu_to_node(i));
6454 se = kzalloc_node(sizeof(struct sched_entity),
6455 GFP_KERNEL, cpu_to_node(i));
6459 init_cfs_rq(cfs_rq);
6460 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6471 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6473 struct rq *rq = cpu_rq(cpu);
6474 unsigned long flags;
6477 * Only empty task groups can be destroyed; so we can speculatively
6478 * check on_list without danger of it being re-added.
6480 if (!tg->cfs_rq[cpu]->on_list)
6483 raw_spin_lock_irqsave(&rq->lock, flags);
6484 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6485 raw_spin_unlock_irqrestore(&rq->lock, flags);
6488 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6489 struct sched_entity *se, int cpu,
6490 struct sched_entity *parent)
6492 struct rq *rq = cpu_rq(cpu);
6496 init_cfs_rq_runtime(cfs_rq);
6498 tg->cfs_rq[cpu] = cfs_rq;
6501 /* se could be NULL for root_task_group */
6506 se->cfs_rq = &rq->cfs;
6508 se->cfs_rq = parent->my_q;
6511 update_load_set(&se->load, 0);
6512 se->parent = parent;
6515 static DEFINE_MUTEX(shares_mutex);
6517 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6520 unsigned long flags;
6523 * We can't change the weight of the root cgroup.
6528 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6530 mutex_lock(&shares_mutex);
6531 if (tg->shares == shares)
6534 tg->shares = shares;
6535 for_each_possible_cpu(i) {
6536 struct rq *rq = cpu_rq(i);
6537 struct sched_entity *se;
6540 /* Propagate contribution to hierarchy */
6541 raw_spin_lock_irqsave(&rq->lock, flags);
6543 /* Possible calls to update_curr() need rq clock */
6544 update_rq_clock(rq);
6545 for_each_sched_entity(se)
6546 update_cfs_shares(group_cfs_rq(se));
6547 raw_spin_unlock_irqrestore(&rq->lock, flags);
6551 mutex_unlock(&shares_mutex);
6554 #else /* CONFIG_FAIR_GROUP_SCHED */
6556 void free_fair_sched_group(struct task_group *tg) { }
6558 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6563 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6565 #endif /* CONFIG_FAIR_GROUP_SCHED */
6568 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6570 struct sched_entity *se = &task->se;
6571 unsigned int rr_interval = 0;
6574 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6577 if (rq->cfs.load.weight)
6578 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6584 * All the scheduling class methods:
6586 const struct sched_class fair_sched_class = {
6587 .next = &idle_sched_class,
6588 .enqueue_task = enqueue_task_fair,
6589 .dequeue_task = dequeue_task_fair,
6590 .yield_task = yield_task_fair,
6591 .yield_to_task = yield_to_task_fair,
6593 .check_preempt_curr = check_preempt_wakeup,
6595 .pick_next_task = pick_next_task_fair,
6596 .put_prev_task = put_prev_task_fair,
6599 .select_task_rq = select_task_rq_fair,
6600 .migrate_task_rq = migrate_task_rq_fair,
6602 .rq_online = rq_online_fair,
6603 .rq_offline = rq_offline_fair,
6605 .task_waking = task_waking_fair,
6608 .set_curr_task = set_curr_task_fair,
6609 .task_tick = task_tick_fair,
6610 .task_fork = task_fork_fair,
6612 .prio_changed = prio_changed_fair,
6613 .switched_from = switched_from_fair,
6614 .switched_to = switched_to_fair,
6616 .get_rr_interval = get_rr_interval_fair,
6618 #ifdef CONFIG_FAIR_GROUP_SCHED
6619 .task_move_group = task_move_group_fair,
6623 #ifdef CONFIG_SCHED_DEBUG
6624 void print_cfs_stats(struct seq_file *m, int cpu)
6626 struct cfs_rq *cfs_rq;
6629 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6630 print_cfs_rq(m, cpu, cfs_rq);
6635 __init void init_sched_fair_class(void)
6638 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6640 #ifdef CONFIG_NO_HZ_COMMON
6641 nohz.next_balance = jiffies;
6642 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6643 cpu_notifier(sched_ilb_notifier, 0);