2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * Approximate time to scan a full NUMA task in ms. The task scan period is
822 * calculated based on the tasks virtual memory size and
823 * numa_balancing_scan_size.
825 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
826 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
829 /* Portion of address space to scan in MB */
830 unsigned int sysctl_numa_balancing_scan_size = 256;
832 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
833 unsigned int sysctl_numa_balancing_scan_delay = 1000;
835 static unsigned int task_nr_scan_windows(struct task_struct *p)
837 unsigned long rss = 0;
838 unsigned long nr_scan_pages;
841 * Calculations based on RSS as non-present and empty pages are skipped
842 * by the PTE scanner and NUMA hinting faults should be trapped based
845 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
846 rss = get_mm_rss(p->mm);
850 rss = round_up(rss, nr_scan_pages);
851 return rss / nr_scan_pages;
854 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
855 #define MAX_SCAN_WINDOW 2560
857 static unsigned int task_scan_min(struct task_struct *p)
859 unsigned int scan, floor;
860 unsigned int windows = 1;
862 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
863 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
864 floor = 1000 / windows;
866 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
867 return max_t(unsigned int, floor, scan);
870 static unsigned int task_scan_max(struct task_struct *p)
872 unsigned int smin = task_scan_min(p);
875 /* Watch for min being lower than max due to floor calculations */
876 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
877 return max(smin, smax);
881 * Once a preferred node is selected the scheduler balancer will prefer moving
882 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
883 * scans. This will give the process the chance to accumulate more faults on
884 * the preferred node but still allow the scheduler to move the task again if
885 * the nodes CPUs are overloaded.
887 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 3;
889 static inline int task_faults_idx(int nid, int priv)
891 return 2 * nid + priv;
894 static inline unsigned long task_faults(struct task_struct *p, int nid)
899 return p->numa_faults[task_faults_idx(nid, 0)] +
900 p->numa_faults[task_faults_idx(nid, 1)];
903 static unsigned long weighted_cpuload(const int cpu);
907 find_idlest_cpu_node(int this_cpu, int nid)
909 unsigned long load, min_load = ULONG_MAX;
910 int i, idlest_cpu = this_cpu;
912 BUG_ON(cpu_to_node(this_cpu) == nid);
915 for_each_cpu(i, cpumask_of_node(nid)) {
916 load = weighted_cpuload(i);
918 if (load < min_load) {
928 static void task_numa_placement(struct task_struct *p)
930 int seq, nid, max_nid = -1;
931 unsigned long max_faults = 0;
933 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
934 if (p->numa_scan_seq == seq)
936 p->numa_scan_seq = seq;
937 p->numa_migrate_seq++;
938 p->numa_scan_period_max = task_scan_max(p);
940 /* Find the node with the highest number of faults */
941 for_each_online_node(nid) {
942 unsigned long faults;
945 for (priv = 0; priv < 2; priv++) {
946 i = task_faults_idx(nid, priv);
948 /* Decay existing window, copy faults since last scan */
949 p->numa_faults[i] >>= 1;
950 p->numa_faults[i] += p->numa_faults_buffer[i];
951 p->numa_faults_buffer[i] = 0;
954 /* Find maximum private faults */
955 faults = p->numa_faults[task_faults_idx(nid, 1)];
956 if (faults > max_faults) {
963 * Record the preferred node as the node with the most faults,
964 * requeue the task to be running on the idlest CPU on the
965 * preferred node and reset the scanning rate to recheck
966 * the working set placement.
968 if (max_faults && max_nid != p->numa_preferred_nid) {
972 * If the task is not on the preferred node then find the most
973 * idle CPU to migrate to.
975 preferred_cpu = task_cpu(p);
976 if (cpu_to_node(preferred_cpu) != max_nid) {
977 preferred_cpu = find_idlest_cpu_node(preferred_cpu,
981 /* Update the preferred nid and migrate task if possible */
982 p->numa_preferred_nid = max_nid;
983 p->numa_migrate_seq = 0;
984 migrate_task_to(p, preferred_cpu);
989 * Got a PROT_NONE fault for a page on @node.
991 void task_numa_fault(int last_nid, int node, int pages, bool migrated)
993 struct task_struct *p = current;
996 if (!numabalancing_enabled)
999 /* for example, ksmd faulting in a user's mm */
1003 /* For now, do not attempt to detect private/shared accesses */
1006 /* Allocate buffer to track faults on a per-node basis */
1007 if (unlikely(!p->numa_faults)) {
1008 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1010 /* numa_faults and numa_faults_buffer share the allocation */
1011 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1012 if (!p->numa_faults)
1015 BUG_ON(p->numa_faults_buffer);
1016 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1020 * If pages are properly placed (did not migrate) then scan slower.
1021 * This is reset periodically in case of phase changes
1024 /* Initialise if necessary */
1025 if (!p->numa_scan_period_max)
1026 p->numa_scan_period_max = task_scan_max(p);
1028 p->numa_scan_period = min(p->numa_scan_period_max,
1029 p->numa_scan_period + 10);
1032 task_numa_placement(p);
1034 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1037 static void reset_ptenuma_scan(struct task_struct *p)
1039 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1040 p->mm->numa_scan_offset = 0;
1044 * The expensive part of numa migration is done from task_work context.
1045 * Triggered from task_tick_numa().
1047 void task_numa_work(struct callback_head *work)
1049 unsigned long migrate, next_scan, now = jiffies;
1050 struct task_struct *p = current;
1051 struct mm_struct *mm = p->mm;
1052 struct vm_area_struct *vma;
1053 unsigned long start, end;
1054 unsigned long nr_pte_updates = 0;
1057 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1059 work->next = work; /* protect against double add */
1061 * Who cares about NUMA placement when they're dying.
1063 * NOTE: make sure not to dereference p->mm before this check,
1064 * exit_task_work() happens _after_ exit_mm() so we could be called
1065 * without p->mm even though we still had it when we enqueued this
1068 if (p->flags & PF_EXITING)
1071 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1072 mm->numa_next_scan = now +
1073 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1074 mm->numa_next_reset = now +
1075 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1079 * Reset the scan period if enough time has gone by. Objective is that
1080 * scanning will be reduced if pages are properly placed. As tasks
1081 * can enter different phases this needs to be re-examined. Lacking
1082 * proper tracking of reference behaviour, this blunt hammer is used.
1084 migrate = mm->numa_next_reset;
1085 if (time_after(now, migrate)) {
1086 p->numa_scan_period = task_scan_min(p);
1087 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1088 xchg(&mm->numa_next_reset, next_scan);
1092 * Enforce maximal scan/migration frequency..
1094 migrate = mm->numa_next_scan;
1095 if (time_before(now, migrate))
1098 if (p->numa_scan_period == 0) {
1099 p->numa_scan_period_max = task_scan_max(p);
1100 p->numa_scan_period = task_scan_min(p);
1103 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1104 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1108 * Delay this task enough that another task of this mm will likely win
1109 * the next time around.
1111 p->node_stamp += 2 * TICK_NSEC;
1113 start = mm->numa_scan_offset;
1114 pages = sysctl_numa_balancing_scan_size;
1115 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1119 down_read(&mm->mmap_sem);
1120 vma = find_vma(mm, start);
1122 reset_ptenuma_scan(p);
1126 for (; vma; vma = vma->vm_next) {
1127 if (!vma_migratable(vma))
1130 /* Skip small VMAs. They are not likely to be of relevance */
1131 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1135 start = max(start, vma->vm_start);
1136 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1137 end = min(end, vma->vm_end);
1138 nr_pte_updates += change_prot_numa(vma, start, end);
1141 * Scan sysctl_numa_balancing_scan_size but ensure that
1142 * at least one PTE is updated so that unused virtual
1143 * address space is quickly skipped.
1146 pages -= (end - start) >> PAGE_SHIFT;
1151 } while (end != vma->vm_end);
1156 * If the whole process was scanned without updates then no NUMA
1157 * hinting faults are being recorded and scan rate should be lower.
1159 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1160 p->numa_scan_period = min(p->numa_scan_period_max,
1161 p->numa_scan_period << 1);
1163 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1164 mm->numa_next_scan = next_scan;
1168 * It is possible to reach the end of the VMA list but the last few
1169 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1170 * would find the !migratable VMA on the next scan but not reset the
1171 * scanner to the start so check it now.
1174 mm->numa_scan_offset = start;
1176 reset_ptenuma_scan(p);
1177 up_read(&mm->mmap_sem);
1181 * Drive the periodic memory faults..
1183 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1185 struct callback_head *work = &curr->numa_work;
1189 * We don't care about NUMA placement if we don't have memory.
1191 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1195 * Using runtime rather than walltime has the dual advantage that
1196 * we (mostly) drive the selection from busy threads and that the
1197 * task needs to have done some actual work before we bother with
1200 now = curr->se.sum_exec_runtime;
1201 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1203 if (now - curr->node_stamp > period) {
1204 if (!curr->node_stamp)
1205 curr->numa_scan_period = task_scan_min(curr);
1206 curr->node_stamp += period;
1208 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1209 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1210 task_work_add(curr, work, true);
1215 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1218 #endif /* CONFIG_NUMA_BALANCING */
1221 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1223 update_load_add(&cfs_rq->load, se->load.weight);
1224 if (!parent_entity(se))
1225 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1227 if (entity_is_task(se))
1228 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1230 cfs_rq->nr_running++;
1234 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1236 update_load_sub(&cfs_rq->load, se->load.weight);
1237 if (!parent_entity(se))
1238 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1239 if (entity_is_task(se))
1240 list_del_init(&se->group_node);
1241 cfs_rq->nr_running--;
1244 #ifdef CONFIG_FAIR_GROUP_SCHED
1246 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1251 * Use this CPU's actual weight instead of the last load_contribution
1252 * to gain a more accurate current total weight. See
1253 * update_cfs_rq_load_contribution().
1255 tg_weight = atomic_long_read(&tg->load_avg);
1256 tg_weight -= cfs_rq->tg_load_contrib;
1257 tg_weight += cfs_rq->load.weight;
1262 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1264 long tg_weight, load, shares;
1266 tg_weight = calc_tg_weight(tg, cfs_rq);
1267 load = cfs_rq->load.weight;
1269 shares = (tg->shares * load);
1271 shares /= tg_weight;
1273 if (shares < MIN_SHARES)
1274 shares = MIN_SHARES;
1275 if (shares > tg->shares)
1276 shares = tg->shares;
1280 # else /* CONFIG_SMP */
1281 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1285 # endif /* CONFIG_SMP */
1286 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1287 unsigned long weight)
1290 /* commit outstanding execution time */
1291 if (cfs_rq->curr == se)
1292 update_curr(cfs_rq);
1293 account_entity_dequeue(cfs_rq, se);
1296 update_load_set(&se->load, weight);
1299 account_entity_enqueue(cfs_rq, se);
1302 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1304 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1306 struct task_group *tg;
1307 struct sched_entity *se;
1311 se = tg->se[cpu_of(rq_of(cfs_rq))];
1312 if (!se || throttled_hierarchy(cfs_rq))
1315 if (likely(se->load.weight == tg->shares))
1318 shares = calc_cfs_shares(cfs_rq, tg);
1320 reweight_entity(cfs_rq_of(se), se, shares);
1322 #else /* CONFIG_FAIR_GROUP_SCHED */
1323 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1326 #endif /* CONFIG_FAIR_GROUP_SCHED */
1330 * We choose a half-life close to 1 scheduling period.
1331 * Note: The tables below are dependent on this value.
1333 #define LOAD_AVG_PERIOD 32
1334 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1335 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1337 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1338 static const u32 runnable_avg_yN_inv[] = {
1339 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1340 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1341 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1342 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1343 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1344 0x85aac367, 0x82cd8698,
1348 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1349 * over-estimates when re-combining.
1351 static const u32 runnable_avg_yN_sum[] = {
1352 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1353 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1354 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1359 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1361 static __always_inline u64 decay_load(u64 val, u64 n)
1363 unsigned int local_n;
1367 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1370 /* after bounds checking we can collapse to 32-bit */
1374 * As y^PERIOD = 1/2, we can combine
1375 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1376 * With a look-up table which covers k^n (n<PERIOD)
1378 * To achieve constant time decay_load.
1380 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1381 val >>= local_n / LOAD_AVG_PERIOD;
1382 local_n %= LOAD_AVG_PERIOD;
1385 val *= runnable_avg_yN_inv[local_n];
1386 /* We don't use SRR here since we always want to round down. */
1391 * For updates fully spanning n periods, the contribution to runnable
1392 * average will be: \Sum 1024*y^n
1394 * We can compute this reasonably efficiently by combining:
1395 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1397 static u32 __compute_runnable_contrib(u64 n)
1401 if (likely(n <= LOAD_AVG_PERIOD))
1402 return runnable_avg_yN_sum[n];
1403 else if (unlikely(n >= LOAD_AVG_MAX_N))
1404 return LOAD_AVG_MAX;
1406 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1408 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1409 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1411 n -= LOAD_AVG_PERIOD;
1412 } while (n > LOAD_AVG_PERIOD);
1414 contrib = decay_load(contrib, n);
1415 return contrib + runnable_avg_yN_sum[n];
1419 * We can represent the historical contribution to runnable average as the
1420 * coefficients of a geometric series. To do this we sub-divide our runnable
1421 * history into segments of approximately 1ms (1024us); label the segment that
1422 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1424 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1426 * (now) (~1ms ago) (~2ms ago)
1428 * Let u_i denote the fraction of p_i that the entity was runnable.
1430 * We then designate the fractions u_i as our co-efficients, yielding the
1431 * following representation of historical load:
1432 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1434 * We choose y based on the with of a reasonably scheduling period, fixing:
1437 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1438 * approximately half as much as the contribution to load within the last ms
1441 * When a period "rolls over" and we have new u_0`, multiplying the previous
1442 * sum again by y is sufficient to update:
1443 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1444 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1446 static __always_inline int __update_entity_runnable_avg(u64 now,
1447 struct sched_avg *sa,
1451 u32 runnable_contrib;
1452 int delta_w, decayed = 0;
1454 delta = now - sa->last_runnable_update;
1456 * This should only happen when time goes backwards, which it
1457 * unfortunately does during sched clock init when we swap over to TSC.
1459 if ((s64)delta < 0) {
1460 sa->last_runnable_update = now;
1465 * Use 1024ns as the unit of measurement since it's a reasonable
1466 * approximation of 1us and fast to compute.
1471 sa->last_runnable_update = now;
1473 /* delta_w is the amount already accumulated against our next period */
1474 delta_w = sa->runnable_avg_period % 1024;
1475 if (delta + delta_w >= 1024) {
1476 /* period roll-over */
1480 * Now that we know we're crossing a period boundary, figure
1481 * out how much from delta we need to complete the current
1482 * period and accrue it.
1484 delta_w = 1024 - delta_w;
1486 sa->runnable_avg_sum += delta_w;
1487 sa->runnable_avg_period += delta_w;
1491 /* Figure out how many additional periods this update spans */
1492 periods = delta / 1024;
1495 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1497 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1500 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1501 runnable_contrib = __compute_runnable_contrib(periods);
1503 sa->runnable_avg_sum += runnable_contrib;
1504 sa->runnable_avg_period += runnable_contrib;
1507 /* Remainder of delta accrued against u_0` */
1509 sa->runnable_avg_sum += delta;
1510 sa->runnable_avg_period += delta;
1515 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1516 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1518 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1519 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1521 decays -= se->avg.decay_count;
1525 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1526 se->avg.decay_count = 0;
1531 #ifdef CONFIG_FAIR_GROUP_SCHED
1532 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1535 struct task_group *tg = cfs_rq->tg;
1538 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1539 tg_contrib -= cfs_rq->tg_load_contrib;
1541 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1542 atomic_long_add(tg_contrib, &tg->load_avg);
1543 cfs_rq->tg_load_contrib += tg_contrib;
1548 * Aggregate cfs_rq runnable averages into an equivalent task_group
1549 * representation for computing load contributions.
1551 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1552 struct cfs_rq *cfs_rq)
1554 struct task_group *tg = cfs_rq->tg;
1557 /* The fraction of a cpu used by this cfs_rq */
1558 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1559 sa->runnable_avg_period + 1);
1560 contrib -= cfs_rq->tg_runnable_contrib;
1562 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1563 atomic_add(contrib, &tg->runnable_avg);
1564 cfs_rq->tg_runnable_contrib += contrib;
1568 static inline void __update_group_entity_contrib(struct sched_entity *se)
1570 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1571 struct task_group *tg = cfs_rq->tg;
1576 contrib = cfs_rq->tg_load_contrib * tg->shares;
1577 se->avg.load_avg_contrib = div_u64(contrib,
1578 atomic_long_read(&tg->load_avg) + 1);
1581 * For group entities we need to compute a correction term in the case
1582 * that they are consuming <1 cpu so that we would contribute the same
1583 * load as a task of equal weight.
1585 * Explicitly co-ordinating this measurement would be expensive, but
1586 * fortunately the sum of each cpus contribution forms a usable
1587 * lower-bound on the true value.
1589 * Consider the aggregate of 2 contributions. Either they are disjoint
1590 * (and the sum represents true value) or they are disjoint and we are
1591 * understating by the aggregate of their overlap.
1593 * Extending this to N cpus, for a given overlap, the maximum amount we
1594 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1595 * cpus that overlap for this interval and w_i is the interval width.
1597 * On a small machine; the first term is well-bounded which bounds the
1598 * total error since w_i is a subset of the period. Whereas on a
1599 * larger machine, while this first term can be larger, if w_i is the
1600 * of consequential size guaranteed to see n_i*w_i quickly converge to
1601 * our upper bound of 1-cpu.
1603 runnable_avg = atomic_read(&tg->runnable_avg);
1604 if (runnable_avg < NICE_0_LOAD) {
1605 se->avg.load_avg_contrib *= runnable_avg;
1606 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1610 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1611 int force_update) {}
1612 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1613 struct cfs_rq *cfs_rq) {}
1614 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1617 static inline void __update_task_entity_contrib(struct sched_entity *se)
1621 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1622 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1623 contrib /= (se->avg.runnable_avg_period + 1);
1624 se->avg.load_avg_contrib = scale_load(contrib);
1627 /* Compute the current contribution to load_avg by se, return any delta */
1628 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1630 long old_contrib = se->avg.load_avg_contrib;
1632 if (entity_is_task(se)) {
1633 __update_task_entity_contrib(se);
1635 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1636 __update_group_entity_contrib(se);
1639 return se->avg.load_avg_contrib - old_contrib;
1642 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1645 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1646 cfs_rq->blocked_load_avg -= load_contrib;
1648 cfs_rq->blocked_load_avg = 0;
1651 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1653 /* Update a sched_entity's runnable average */
1654 static inline void update_entity_load_avg(struct sched_entity *se,
1657 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1662 * For a group entity we need to use their owned cfs_rq_clock_task() in
1663 * case they are the parent of a throttled hierarchy.
1665 if (entity_is_task(se))
1666 now = cfs_rq_clock_task(cfs_rq);
1668 now = cfs_rq_clock_task(group_cfs_rq(se));
1670 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1673 contrib_delta = __update_entity_load_avg_contrib(se);
1679 cfs_rq->runnable_load_avg += contrib_delta;
1681 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1685 * Decay the load contributed by all blocked children and account this so that
1686 * their contribution may appropriately discounted when they wake up.
1688 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1690 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1693 decays = now - cfs_rq->last_decay;
1694 if (!decays && !force_update)
1697 if (atomic_long_read(&cfs_rq->removed_load)) {
1698 unsigned long removed_load;
1699 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1700 subtract_blocked_load_contrib(cfs_rq, removed_load);
1704 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1706 atomic64_add(decays, &cfs_rq->decay_counter);
1707 cfs_rq->last_decay = now;
1710 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1713 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1715 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1716 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1719 /* Add the load generated by se into cfs_rq's child load-average */
1720 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1721 struct sched_entity *se,
1725 * We track migrations using entity decay_count <= 0, on a wake-up
1726 * migration we use a negative decay count to track the remote decays
1727 * accumulated while sleeping.
1729 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1730 * are seen by enqueue_entity_load_avg() as a migration with an already
1731 * constructed load_avg_contrib.
1733 if (unlikely(se->avg.decay_count <= 0)) {
1734 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1735 if (se->avg.decay_count) {
1737 * In a wake-up migration we have to approximate the
1738 * time sleeping. This is because we can't synchronize
1739 * clock_task between the two cpus, and it is not
1740 * guaranteed to be read-safe. Instead, we can
1741 * approximate this using our carried decays, which are
1742 * explicitly atomically readable.
1744 se->avg.last_runnable_update -= (-se->avg.decay_count)
1746 update_entity_load_avg(se, 0);
1747 /* Indicate that we're now synchronized and on-rq */
1748 se->avg.decay_count = 0;
1753 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1754 * would have made count negative); we must be careful to avoid
1755 * double-accounting blocked time after synchronizing decays.
1757 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1761 /* migrated tasks did not contribute to our blocked load */
1763 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1764 update_entity_load_avg(se, 0);
1767 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1768 /* we force update consideration on load-balancer moves */
1769 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1773 * Remove se's load from this cfs_rq child load-average, if the entity is
1774 * transitioning to a blocked state we track its projected decay using
1777 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1778 struct sched_entity *se,
1781 update_entity_load_avg(se, 1);
1782 /* we force update consideration on load-balancer moves */
1783 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1785 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1787 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1788 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1789 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1793 * Update the rq's load with the elapsed running time before entering
1794 * idle. if the last scheduled task is not a CFS task, idle_enter will
1795 * be the only way to update the runnable statistic.
1797 void idle_enter_fair(struct rq *this_rq)
1799 update_rq_runnable_avg(this_rq, 1);
1803 * Update the rq's load with the elapsed idle time before a task is
1804 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1805 * be the only way to update the runnable statistic.
1807 void idle_exit_fair(struct rq *this_rq)
1809 update_rq_runnable_avg(this_rq, 0);
1813 static inline void update_entity_load_avg(struct sched_entity *se,
1814 int update_cfs_rq) {}
1815 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1816 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1817 struct sched_entity *se,
1819 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1820 struct sched_entity *se,
1822 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1823 int force_update) {}
1826 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1828 #ifdef CONFIG_SCHEDSTATS
1829 struct task_struct *tsk = NULL;
1831 if (entity_is_task(se))
1834 if (se->statistics.sleep_start) {
1835 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1840 if (unlikely(delta > se->statistics.sleep_max))
1841 se->statistics.sleep_max = delta;
1843 se->statistics.sleep_start = 0;
1844 se->statistics.sum_sleep_runtime += delta;
1847 account_scheduler_latency(tsk, delta >> 10, 1);
1848 trace_sched_stat_sleep(tsk, delta);
1851 if (se->statistics.block_start) {
1852 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1857 if (unlikely(delta > se->statistics.block_max))
1858 se->statistics.block_max = delta;
1860 se->statistics.block_start = 0;
1861 se->statistics.sum_sleep_runtime += delta;
1864 if (tsk->in_iowait) {
1865 se->statistics.iowait_sum += delta;
1866 se->statistics.iowait_count++;
1867 trace_sched_stat_iowait(tsk, delta);
1870 trace_sched_stat_blocked(tsk, delta);
1873 * Blocking time is in units of nanosecs, so shift by
1874 * 20 to get a milliseconds-range estimation of the
1875 * amount of time that the task spent sleeping:
1877 if (unlikely(prof_on == SLEEP_PROFILING)) {
1878 profile_hits(SLEEP_PROFILING,
1879 (void *)get_wchan(tsk),
1882 account_scheduler_latency(tsk, delta >> 10, 0);
1888 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1890 #ifdef CONFIG_SCHED_DEBUG
1891 s64 d = se->vruntime - cfs_rq->min_vruntime;
1896 if (d > 3*sysctl_sched_latency)
1897 schedstat_inc(cfs_rq, nr_spread_over);
1902 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1904 u64 vruntime = cfs_rq->min_vruntime;
1907 * The 'current' period is already promised to the current tasks,
1908 * however the extra weight of the new task will slow them down a
1909 * little, place the new task so that it fits in the slot that
1910 * stays open at the end.
1912 if (initial && sched_feat(START_DEBIT))
1913 vruntime += sched_vslice(cfs_rq, se);
1915 /* sleeps up to a single latency don't count. */
1917 unsigned long thresh = sysctl_sched_latency;
1920 * Halve their sleep time's effect, to allow
1921 * for a gentler effect of sleepers:
1923 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1929 /* ensure we never gain time by being placed backwards. */
1930 se->vruntime = max_vruntime(se->vruntime, vruntime);
1933 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1936 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1939 * Update the normalized vruntime before updating min_vruntime
1940 * through calling update_curr().
1942 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1943 se->vruntime += cfs_rq->min_vruntime;
1946 * Update run-time statistics of the 'current'.
1948 update_curr(cfs_rq);
1949 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1950 account_entity_enqueue(cfs_rq, se);
1951 update_cfs_shares(cfs_rq);
1953 if (flags & ENQUEUE_WAKEUP) {
1954 place_entity(cfs_rq, se, 0);
1955 enqueue_sleeper(cfs_rq, se);
1958 update_stats_enqueue(cfs_rq, se);
1959 check_spread(cfs_rq, se);
1960 if (se != cfs_rq->curr)
1961 __enqueue_entity(cfs_rq, se);
1964 if (cfs_rq->nr_running == 1) {
1965 list_add_leaf_cfs_rq(cfs_rq);
1966 check_enqueue_throttle(cfs_rq);
1970 static void __clear_buddies_last(struct sched_entity *se)
1972 for_each_sched_entity(se) {
1973 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1974 if (cfs_rq->last == se)
1975 cfs_rq->last = NULL;
1981 static void __clear_buddies_next(struct sched_entity *se)
1983 for_each_sched_entity(se) {
1984 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1985 if (cfs_rq->next == se)
1986 cfs_rq->next = NULL;
1992 static void __clear_buddies_skip(struct sched_entity *se)
1994 for_each_sched_entity(se) {
1995 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1996 if (cfs_rq->skip == se)
1997 cfs_rq->skip = NULL;
2003 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2005 if (cfs_rq->last == se)
2006 __clear_buddies_last(se);
2008 if (cfs_rq->next == se)
2009 __clear_buddies_next(se);
2011 if (cfs_rq->skip == se)
2012 __clear_buddies_skip(se);
2015 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2018 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2021 * Update run-time statistics of the 'current'.
2023 update_curr(cfs_rq);
2024 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2026 update_stats_dequeue(cfs_rq, se);
2027 if (flags & DEQUEUE_SLEEP) {
2028 #ifdef CONFIG_SCHEDSTATS
2029 if (entity_is_task(se)) {
2030 struct task_struct *tsk = task_of(se);
2032 if (tsk->state & TASK_INTERRUPTIBLE)
2033 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2034 if (tsk->state & TASK_UNINTERRUPTIBLE)
2035 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2040 clear_buddies(cfs_rq, se);
2042 if (se != cfs_rq->curr)
2043 __dequeue_entity(cfs_rq, se);
2045 account_entity_dequeue(cfs_rq, se);
2048 * Normalize the entity after updating the min_vruntime because the
2049 * update can refer to the ->curr item and we need to reflect this
2050 * movement in our normalized position.
2052 if (!(flags & DEQUEUE_SLEEP))
2053 se->vruntime -= cfs_rq->min_vruntime;
2055 /* return excess runtime on last dequeue */
2056 return_cfs_rq_runtime(cfs_rq);
2058 update_min_vruntime(cfs_rq);
2059 update_cfs_shares(cfs_rq);
2063 * Preempt the current task with a newly woken task if needed:
2066 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2068 unsigned long ideal_runtime, delta_exec;
2069 struct sched_entity *se;
2072 ideal_runtime = sched_slice(cfs_rq, curr);
2073 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2074 if (delta_exec > ideal_runtime) {
2075 resched_task(rq_of(cfs_rq)->curr);
2077 * The current task ran long enough, ensure it doesn't get
2078 * re-elected due to buddy favours.
2080 clear_buddies(cfs_rq, curr);
2085 * Ensure that a task that missed wakeup preemption by a
2086 * narrow margin doesn't have to wait for a full slice.
2087 * This also mitigates buddy induced latencies under load.
2089 if (delta_exec < sysctl_sched_min_granularity)
2092 se = __pick_first_entity(cfs_rq);
2093 delta = curr->vruntime - se->vruntime;
2098 if (delta > ideal_runtime)
2099 resched_task(rq_of(cfs_rq)->curr);
2103 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2105 /* 'current' is not kept within the tree. */
2108 * Any task has to be enqueued before it get to execute on
2109 * a CPU. So account for the time it spent waiting on the
2112 update_stats_wait_end(cfs_rq, se);
2113 __dequeue_entity(cfs_rq, se);
2116 update_stats_curr_start(cfs_rq, se);
2118 #ifdef CONFIG_SCHEDSTATS
2120 * Track our maximum slice length, if the CPU's load is at
2121 * least twice that of our own weight (i.e. dont track it
2122 * when there are only lesser-weight tasks around):
2124 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2125 se->statistics.slice_max = max(se->statistics.slice_max,
2126 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2129 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2133 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2136 * Pick the next process, keeping these things in mind, in this order:
2137 * 1) keep things fair between processes/task groups
2138 * 2) pick the "next" process, since someone really wants that to run
2139 * 3) pick the "last" process, for cache locality
2140 * 4) do not run the "skip" process, if something else is available
2142 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2144 struct sched_entity *se = __pick_first_entity(cfs_rq);
2145 struct sched_entity *left = se;
2148 * Avoid running the skip buddy, if running something else can
2149 * be done without getting too unfair.
2151 if (cfs_rq->skip == se) {
2152 struct sched_entity *second = __pick_next_entity(se);
2153 if (second && wakeup_preempt_entity(second, left) < 1)
2158 * Prefer last buddy, try to return the CPU to a preempted task.
2160 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2164 * Someone really wants this to run. If it's not unfair, run it.
2166 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2169 clear_buddies(cfs_rq, se);
2174 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2176 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2179 * If still on the runqueue then deactivate_task()
2180 * was not called and update_curr() has to be done:
2183 update_curr(cfs_rq);
2185 /* throttle cfs_rqs exceeding runtime */
2186 check_cfs_rq_runtime(cfs_rq);
2188 check_spread(cfs_rq, prev);
2190 update_stats_wait_start(cfs_rq, prev);
2191 /* Put 'current' back into the tree. */
2192 __enqueue_entity(cfs_rq, prev);
2193 /* in !on_rq case, update occurred at dequeue */
2194 update_entity_load_avg(prev, 1);
2196 cfs_rq->curr = NULL;
2200 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2203 * Update run-time statistics of the 'current'.
2205 update_curr(cfs_rq);
2208 * Ensure that runnable average is periodically updated.
2210 update_entity_load_avg(curr, 1);
2211 update_cfs_rq_blocked_load(cfs_rq, 1);
2212 update_cfs_shares(cfs_rq);
2214 #ifdef CONFIG_SCHED_HRTICK
2216 * queued ticks are scheduled to match the slice, so don't bother
2217 * validating it and just reschedule.
2220 resched_task(rq_of(cfs_rq)->curr);
2224 * don't let the period tick interfere with the hrtick preemption
2226 if (!sched_feat(DOUBLE_TICK) &&
2227 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2231 if (cfs_rq->nr_running > 1)
2232 check_preempt_tick(cfs_rq, curr);
2236 /**************************************************
2237 * CFS bandwidth control machinery
2240 #ifdef CONFIG_CFS_BANDWIDTH
2242 #ifdef HAVE_JUMP_LABEL
2243 static struct static_key __cfs_bandwidth_used;
2245 static inline bool cfs_bandwidth_used(void)
2247 return static_key_false(&__cfs_bandwidth_used);
2250 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2252 /* only need to count groups transitioning between enabled/!enabled */
2253 if (enabled && !was_enabled)
2254 static_key_slow_inc(&__cfs_bandwidth_used);
2255 else if (!enabled && was_enabled)
2256 static_key_slow_dec(&__cfs_bandwidth_used);
2258 #else /* HAVE_JUMP_LABEL */
2259 static bool cfs_bandwidth_used(void)
2264 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2265 #endif /* HAVE_JUMP_LABEL */
2268 * default period for cfs group bandwidth.
2269 * default: 0.1s, units: nanoseconds
2271 static inline u64 default_cfs_period(void)
2273 return 100000000ULL;
2276 static inline u64 sched_cfs_bandwidth_slice(void)
2278 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2282 * Replenish runtime according to assigned quota and update expiration time.
2283 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2284 * additional synchronization around rq->lock.
2286 * requires cfs_b->lock
2288 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2292 if (cfs_b->quota == RUNTIME_INF)
2295 now = sched_clock_cpu(smp_processor_id());
2296 cfs_b->runtime = cfs_b->quota;
2297 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2300 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2302 return &tg->cfs_bandwidth;
2305 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2306 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2308 if (unlikely(cfs_rq->throttle_count))
2309 return cfs_rq->throttled_clock_task;
2311 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2314 /* returns 0 on failure to allocate runtime */
2315 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2317 struct task_group *tg = cfs_rq->tg;
2318 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2319 u64 amount = 0, min_amount, expires;
2321 /* note: this is a positive sum as runtime_remaining <= 0 */
2322 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2324 raw_spin_lock(&cfs_b->lock);
2325 if (cfs_b->quota == RUNTIME_INF)
2326 amount = min_amount;
2329 * If the bandwidth pool has become inactive, then at least one
2330 * period must have elapsed since the last consumption.
2331 * Refresh the global state and ensure bandwidth timer becomes
2334 if (!cfs_b->timer_active) {
2335 __refill_cfs_bandwidth_runtime(cfs_b);
2336 __start_cfs_bandwidth(cfs_b);
2339 if (cfs_b->runtime > 0) {
2340 amount = min(cfs_b->runtime, min_amount);
2341 cfs_b->runtime -= amount;
2345 expires = cfs_b->runtime_expires;
2346 raw_spin_unlock(&cfs_b->lock);
2348 cfs_rq->runtime_remaining += amount;
2350 * we may have advanced our local expiration to account for allowed
2351 * spread between our sched_clock and the one on which runtime was
2354 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2355 cfs_rq->runtime_expires = expires;
2357 return cfs_rq->runtime_remaining > 0;
2361 * Note: This depends on the synchronization provided by sched_clock and the
2362 * fact that rq->clock snapshots this value.
2364 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2366 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2368 /* if the deadline is ahead of our clock, nothing to do */
2369 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2372 if (cfs_rq->runtime_remaining < 0)
2376 * If the local deadline has passed we have to consider the
2377 * possibility that our sched_clock is 'fast' and the global deadline
2378 * has not truly expired.
2380 * Fortunately we can check determine whether this the case by checking
2381 * whether the global deadline has advanced.
2384 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2385 /* extend local deadline, drift is bounded above by 2 ticks */
2386 cfs_rq->runtime_expires += TICK_NSEC;
2388 /* global deadline is ahead, expiration has passed */
2389 cfs_rq->runtime_remaining = 0;
2393 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2394 unsigned long delta_exec)
2396 /* dock delta_exec before expiring quota (as it could span periods) */
2397 cfs_rq->runtime_remaining -= delta_exec;
2398 expire_cfs_rq_runtime(cfs_rq);
2400 if (likely(cfs_rq->runtime_remaining > 0))
2404 * if we're unable to extend our runtime we resched so that the active
2405 * hierarchy can be throttled
2407 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2408 resched_task(rq_of(cfs_rq)->curr);
2411 static __always_inline
2412 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2414 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2417 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2420 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2422 return cfs_bandwidth_used() && cfs_rq->throttled;
2425 /* check whether cfs_rq, or any parent, is throttled */
2426 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2428 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2432 * Ensure that neither of the group entities corresponding to src_cpu or
2433 * dest_cpu are members of a throttled hierarchy when performing group
2434 * load-balance operations.
2436 static inline int throttled_lb_pair(struct task_group *tg,
2437 int src_cpu, int dest_cpu)
2439 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2441 src_cfs_rq = tg->cfs_rq[src_cpu];
2442 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2444 return throttled_hierarchy(src_cfs_rq) ||
2445 throttled_hierarchy(dest_cfs_rq);
2448 /* updated child weight may affect parent so we have to do this bottom up */
2449 static int tg_unthrottle_up(struct task_group *tg, void *data)
2451 struct rq *rq = data;
2452 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2454 cfs_rq->throttle_count--;
2456 if (!cfs_rq->throttle_count) {
2457 /* adjust cfs_rq_clock_task() */
2458 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2459 cfs_rq->throttled_clock_task;
2466 static int tg_throttle_down(struct task_group *tg, void *data)
2468 struct rq *rq = data;
2469 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2471 /* group is entering throttled state, stop time */
2472 if (!cfs_rq->throttle_count)
2473 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2474 cfs_rq->throttle_count++;
2479 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2481 struct rq *rq = rq_of(cfs_rq);
2482 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2483 struct sched_entity *se;
2484 long task_delta, dequeue = 1;
2486 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2488 /* freeze hierarchy runnable averages while throttled */
2490 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2493 task_delta = cfs_rq->h_nr_running;
2494 for_each_sched_entity(se) {
2495 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2496 /* throttled entity or throttle-on-deactivate */
2501 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2502 qcfs_rq->h_nr_running -= task_delta;
2504 if (qcfs_rq->load.weight)
2509 rq->nr_running -= task_delta;
2511 cfs_rq->throttled = 1;
2512 cfs_rq->throttled_clock = rq_clock(rq);
2513 raw_spin_lock(&cfs_b->lock);
2514 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2515 raw_spin_unlock(&cfs_b->lock);
2518 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2520 struct rq *rq = rq_of(cfs_rq);
2521 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2522 struct sched_entity *se;
2526 se = cfs_rq->tg->se[cpu_of(rq)];
2528 cfs_rq->throttled = 0;
2530 update_rq_clock(rq);
2532 raw_spin_lock(&cfs_b->lock);
2533 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2534 list_del_rcu(&cfs_rq->throttled_list);
2535 raw_spin_unlock(&cfs_b->lock);
2537 /* update hierarchical throttle state */
2538 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2540 if (!cfs_rq->load.weight)
2543 task_delta = cfs_rq->h_nr_running;
2544 for_each_sched_entity(se) {
2548 cfs_rq = cfs_rq_of(se);
2550 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2551 cfs_rq->h_nr_running += task_delta;
2553 if (cfs_rq_throttled(cfs_rq))
2558 rq->nr_running += task_delta;
2560 /* determine whether we need to wake up potentially idle cpu */
2561 if (rq->curr == rq->idle && rq->cfs.nr_running)
2562 resched_task(rq->curr);
2565 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2566 u64 remaining, u64 expires)
2568 struct cfs_rq *cfs_rq;
2569 u64 runtime = remaining;
2572 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2574 struct rq *rq = rq_of(cfs_rq);
2576 raw_spin_lock(&rq->lock);
2577 if (!cfs_rq_throttled(cfs_rq))
2580 runtime = -cfs_rq->runtime_remaining + 1;
2581 if (runtime > remaining)
2582 runtime = remaining;
2583 remaining -= runtime;
2585 cfs_rq->runtime_remaining += runtime;
2586 cfs_rq->runtime_expires = expires;
2588 /* we check whether we're throttled above */
2589 if (cfs_rq->runtime_remaining > 0)
2590 unthrottle_cfs_rq(cfs_rq);
2593 raw_spin_unlock(&rq->lock);
2604 * Responsible for refilling a task_group's bandwidth and unthrottling its
2605 * cfs_rqs as appropriate. If there has been no activity within the last
2606 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2607 * used to track this state.
2609 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2611 u64 runtime, runtime_expires;
2612 int idle = 1, throttled;
2614 raw_spin_lock(&cfs_b->lock);
2615 /* no need to continue the timer with no bandwidth constraint */
2616 if (cfs_b->quota == RUNTIME_INF)
2619 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2620 /* idle depends on !throttled (for the case of a large deficit) */
2621 idle = cfs_b->idle && !throttled;
2622 cfs_b->nr_periods += overrun;
2624 /* if we're going inactive then everything else can be deferred */
2628 __refill_cfs_bandwidth_runtime(cfs_b);
2631 /* mark as potentially idle for the upcoming period */
2636 /* account preceding periods in which throttling occurred */
2637 cfs_b->nr_throttled += overrun;
2640 * There are throttled entities so we must first use the new bandwidth
2641 * to unthrottle them before making it generally available. This
2642 * ensures that all existing debts will be paid before a new cfs_rq is
2645 runtime = cfs_b->runtime;
2646 runtime_expires = cfs_b->runtime_expires;
2650 * This check is repeated as we are holding onto the new bandwidth
2651 * while we unthrottle. This can potentially race with an unthrottled
2652 * group trying to acquire new bandwidth from the global pool.
2654 while (throttled && runtime > 0) {
2655 raw_spin_unlock(&cfs_b->lock);
2656 /* we can't nest cfs_b->lock while distributing bandwidth */
2657 runtime = distribute_cfs_runtime(cfs_b, runtime,
2659 raw_spin_lock(&cfs_b->lock);
2661 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2664 /* return (any) remaining runtime */
2665 cfs_b->runtime = runtime;
2667 * While we are ensured activity in the period following an
2668 * unthrottle, this also covers the case in which the new bandwidth is
2669 * insufficient to cover the existing bandwidth deficit. (Forcing the
2670 * timer to remain active while there are any throttled entities.)
2675 cfs_b->timer_active = 0;
2676 raw_spin_unlock(&cfs_b->lock);
2681 /* a cfs_rq won't donate quota below this amount */
2682 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2683 /* minimum remaining period time to redistribute slack quota */
2684 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2685 /* how long we wait to gather additional slack before distributing */
2686 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2688 /* are we near the end of the current quota period? */
2689 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2691 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2694 /* if the call-back is running a quota refresh is already occurring */
2695 if (hrtimer_callback_running(refresh_timer))
2698 /* is a quota refresh about to occur? */
2699 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2700 if (remaining < min_expire)
2706 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2708 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2710 /* if there's a quota refresh soon don't bother with slack */
2711 if (runtime_refresh_within(cfs_b, min_left))
2714 start_bandwidth_timer(&cfs_b->slack_timer,
2715 ns_to_ktime(cfs_bandwidth_slack_period));
2718 /* we know any runtime found here is valid as update_curr() precedes return */
2719 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2721 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2722 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2724 if (slack_runtime <= 0)
2727 raw_spin_lock(&cfs_b->lock);
2728 if (cfs_b->quota != RUNTIME_INF &&
2729 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2730 cfs_b->runtime += slack_runtime;
2732 /* we are under rq->lock, defer unthrottling using a timer */
2733 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2734 !list_empty(&cfs_b->throttled_cfs_rq))
2735 start_cfs_slack_bandwidth(cfs_b);
2737 raw_spin_unlock(&cfs_b->lock);
2739 /* even if it's not valid for return we don't want to try again */
2740 cfs_rq->runtime_remaining -= slack_runtime;
2743 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2745 if (!cfs_bandwidth_used())
2748 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2751 __return_cfs_rq_runtime(cfs_rq);
2755 * This is done with a timer (instead of inline with bandwidth return) since
2756 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2758 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2760 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2763 /* confirm we're still not at a refresh boundary */
2764 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2767 raw_spin_lock(&cfs_b->lock);
2768 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2769 runtime = cfs_b->runtime;
2772 expires = cfs_b->runtime_expires;
2773 raw_spin_unlock(&cfs_b->lock);
2778 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2780 raw_spin_lock(&cfs_b->lock);
2781 if (expires == cfs_b->runtime_expires)
2782 cfs_b->runtime = runtime;
2783 raw_spin_unlock(&cfs_b->lock);
2787 * When a group wakes up we want to make sure that its quota is not already
2788 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2789 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2791 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2793 if (!cfs_bandwidth_used())
2796 /* an active group must be handled by the update_curr()->put() path */
2797 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2800 /* ensure the group is not already throttled */
2801 if (cfs_rq_throttled(cfs_rq))
2804 /* update runtime allocation */
2805 account_cfs_rq_runtime(cfs_rq, 0);
2806 if (cfs_rq->runtime_remaining <= 0)
2807 throttle_cfs_rq(cfs_rq);
2810 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2811 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2813 if (!cfs_bandwidth_used())
2816 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2820 * it's possible for a throttled entity to be forced into a running
2821 * state (e.g. set_curr_task), in this case we're finished.
2823 if (cfs_rq_throttled(cfs_rq))
2826 throttle_cfs_rq(cfs_rq);
2829 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2831 struct cfs_bandwidth *cfs_b =
2832 container_of(timer, struct cfs_bandwidth, slack_timer);
2833 do_sched_cfs_slack_timer(cfs_b);
2835 return HRTIMER_NORESTART;
2838 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2840 struct cfs_bandwidth *cfs_b =
2841 container_of(timer, struct cfs_bandwidth, period_timer);
2847 now = hrtimer_cb_get_time(timer);
2848 overrun = hrtimer_forward(timer, now, cfs_b->period);
2853 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2856 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2859 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2861 raw_spin_lock_init(&cfs_b->lock);
2863 cfs_b->quota = RUNTIME_INF;
2864 cfs_b->period = ns_to_ktime(default_cfs_period());
2866 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2867 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2868 cfs_b->period_timer.function = sched_cfs_period_timer;
2869 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2870 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2873 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2875 cfs_rq->runtime_enabled = 0;
2876 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2879 /* requires cfs_b->lock, may release to reprogram timer */
2880 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2883 * The timer may be active because we're trying to set a new bandwidth
2884 * period or because we're racing with the tear-down path
2885 * (timer_active==0 becomes visible before the hrtimer call-back
2886 * terminates). In either case we ensure that it's re-programmed
2888 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2889 raw_spin_unlock(&cfs_b->lock);
2890 /* ensure cfs_b->lock is available while we wait */
2891 hrtimer_cancel(&cfs_b->period_timer);
2893 raw_spin_lock(&cfs_b->lock);
2894 /* if someone else restarted the timer then we're done */
2895 if (cfs_b->timer_active)
2899 cfs_b->timer_active = 1;
2900 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2903 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2905 hrtimer_cancel(&cfs_b->period_timer);
2906 hrtimer_cancel(&cfs_b->slack_timer);
2909 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2911 struct cfs_rq *cfs_rq;
2913 for_each_leaf_cfs_rq(rq, cfs_rq) {
2914 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2916 if (!cfs_rq->runtime_enabled)
2920 * clock_task is not advancing so we just need to make sure
2921 * there's some valid quota amount
2923 cfs_rq->runtime_remaining = cfs_b->quota;
2924 if (cfs_rq_throttled(cfs_rq))
2925 unthrottle_cfs_rq(cfs_rq);
2929 #else /* CONFIG_CFS_BANDWIDTH */
2930 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2932 return rq_clock_task(rq_of(cfs_rq));
2935 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2936 unsigned long delta_exec) {}
2937 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2938 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2939 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2941 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2946 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2951 static inline int throttled_lb_pair(struct task_group *tg,
2952 int src_cpu, int dest_cpu)
2957 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2959 #ifdef CONFIG_FAIR_GROUP_SCHED
2960 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2963 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2967 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2968 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2970 #endif /* CONFIG_CFS_BANDWIDTH */
2972 /**************************************************
2973 * CFS operations on tasks:
2976 #ifdef CONFIG_SCHED_HRTICK
2977 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2979 struct sched_entity *se = &p->se;
2980 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2982 WARN_ON(task_rq(p) != rq);
2984 if (cfs_rq->nr_running > 1) {
2985 u64 slice = sched_slice(cfs_rq, se);
2986 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2987 s64 delta = slice - ran;
2996 * Don't schedule slices shorter than 10000ns, that just
2997 * doesn't make sense. Rely on vruntime for fairness.
3000 delta = max_t(s64, 10000LL, delta);
3002 hrtick_start(rq, delta);
3007 * called from enqueue/dequeue and updates the hrtick when the
3008 * current task is from our class and nr_running is low enough
3011 static void hrtick_update(struct rq *rq)
3013 struct task_struct *curr = rq->curr;
3015 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3018 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3019 hrtick_start_fair(rq, curr);
3021 #else /* !CONFIG_SCHED_HRTICK */
3023 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3027 static inline void hrtick_update(struct rq *rq)
3033 * The enqueue_task method is called before nr_running is
3034 * increased. Here we update the fair scheduling stats and
3035 * then put the task into the rbtree:
3038 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3040 struct cfs_rq *cfs_rq;
3041 struct sched_entity *se = &p->se;
3043 for_each_sched_entity(se) {
3046 cfs_rq = cfs_rq_of(se);
3047 enqueue_entity(cfs_rq, se, flags);
3050 * end evaluation on encountering a throttled cfs_rq
3052 * note: in the case of encountering a throttled cfs_rq we will
3053 * post the final h_nr_running increment below.
3055 if (cfs_rq_throttled(cfs_rq))
3057 cfs_rq->h_nr_running++;
3059 flags = ENQUEUE_WAKEUP;
3062 for_each_sched_entity(se) {
3063 cfs_rq = cfs_rq_of(se);
3064 cfs_rq->h_nr_running++;
3066 if (cfs_rq_throttled(cfs_rq))
3069 update_cfs_shares(cfs_rq);
3070 update_entity_load_avg(se, 1);
3074 update_rq_runnable_avg(rq, rq->nr_running);
3080 static void set_next_buddy(struct sched_entity *se);
3083 * The dequeue_task method is called before nr_running is
3084 * decreased. We remove the task from the rbtree and
3085 * update the fair scheduling stats:
3087 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3089 struct cfs_rq *cfs_rq;
3090 struct sched_entity *se = &p->se;
3091 int task_sleep = flags & DEQUEUE_SLEEP;
3093 for_each_sched_entity(se) {
3094 cfs_rq = cfs_rq_of(se);
3095 dequeue_entity(cfs_rq, se, flags);
3098 * end evaluation on encountering a throttled cfs_rq
3100 * note: in the case of encountering a throttled cfs_rq we will
3101 * post the final h_nr_running decrement below.
3103 if (cfs_rq_throttled(cfs_rq))
3105 cfs_rq->h_nr_running--;
3107 /* Don't dequeue parent if it has other entities besides us */
3108 if (cfs_rq->load.weight) {
3110 * Bias pick_next to pick a task from this cfs_rq, as
3111 * p is sleeping when it is within its sched_slice.
3113 if (task_sleep && parent_entity(se))
3114 set_next_buddy(parent_entity(se));
3116 /* avoid re-evaluating load for this entity */
3117 se = parent_entity(se);
3120 flags |= DEQUEUE_SLEEP;
3123 for_each_sched_entity(se) {
3124 cfs_rq = cfs_rq_of(se);
3125 cfs_rq->h_nr_running--;
3127 if (cfs_rq_throttled(cfs_rq))
3130 update_cfs_shares(cfs_rq);
3131 update_entity_load_avg(se, 1);
3136 update_rq_runnable_avg(rq, 1);
3142 /* Used instead of source_load when we know the type == 0 */
3143 static unsigned long weighted_cpuload(const int cpu)
3145 return cpu_rq(cpu)->cfs.runnable_load_avg;
3149 * Return a low guess at the load of a migration-source cpu weighted
3150 * according to the scheduling class and "nice" value.
3152 * We want to under-estimate the load of migration sources, to
3153 * balance conservatively.
3155 static unsigned long source_load(int cpu, int type)
3157 struct rq *rq = cpu_rq(cpu);
3158 unsigned long total = weighted_cpuload(cpu);
3160 if (type == 0 || !sched_feat(LB_BIAS))
3163 return min(rq->cpu_load[type-1], total);
3167 * Return a high guess at the load of a migration-target cpu weighted
3168 * according to the scheduling class and "nice" value.
3170 static unsigned long target_load(int cpu, int type)
3172 struct rq *rq = cpu_rq(cpu);
3173 unsigned long total = weighted_cpuload(cpu);
3175 if (type == 0 || !sched_feat(LB_BIAS))
3178 return max(rq->cpu_load[type-1], total);
3181 static unsigned long power_of(int cpu)
3183 return cpu_rq(cpu)->cpu_power;
3186 static unsigned long cpu_avg_load_per_task(int cpu)
3188 struct rq *rq = cpu_rq(cpu);
3189 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3190 unsigned long load_avg = rq->cfs.runnable_load_avg;
3193 return load_avg / nr_running;
3198 static void record_wakee(struct task_struct *p)
3201 * Rough decay (wiping) for cost saving, don't worry
3202 * about the boundary, really active task won't care
3205 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3206 current->wakee_flips = 0;
3207 current->wakee_flip_decay_ts = jiffies;
3210 if (current->last_wakee != p) {
3211 current->last_wakee = p;
3212 current->wakee_flips++;
3216 static void task_waking_fair(struct task_struct *p)
3218 struct sched_entity *se = &p->se;
3219 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3222 #ifndef CONFIG_64BIT
3223 u64 min_vruntime_copy;
3226 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3228 min_vruntime = cfs_rq->min_vruntime;
3229 } while (min_vruntime != min_vruntime_copy);
3231 min_vruntime = cfs_rq->min_vruntime;
3234 se->vruntime -= min_vruntime;
3238 #ifdef CONFIG_FAIR_GROUP_SCHED
3240 * effective_load() calculates the load change as seen from the root_task_group
3242 * Adding load to a group doesn't make a group heavier, but can cause movement
3243 * of group shares between cpus. Assuming the shares were perfectly aligned one
3244 * can calculate the shift in shares.
3246 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3247 * on this @cpu and results in a total addition (subtraction) of @wg to the
3248 * total group weight.
3250 * Given a runqueue weight distribution (rw_i) we can compute a shares
3251 * distribution (s_i) using:
3253 * s_i = rw_i / \Sum rw_j (1)
3255 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3256 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3257 * shares distribution (s_i):
3259 * rw_i = { 2, 4, 1, 0 }
3260 * s_i = { 2/7, 4/7, 1/7, 0 }
3262 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3263 * task used to run on and the CPU the waker is running on), we need to
3264 * compute the effect of waking a task on either CPU and, in case of a sync
3265 * wakeup, compute the effect of the current task going to sleep.
3267 * So for a change of @wl to the local @cpu with an overall group weight change
3268 * of @wl we can compute the new shares distribution (s'_i) using:
3270 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3272 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3273 * differences in waking a task to CPU 0. The additional task changes the
3274 * weight and shares distributions like:
3276 * rw'_i = { 3, 4, 1, 0 }
3277 * s'_i = { 3/8, 4/8, 1/8, 0 }
3279 * We can then compute the difference in effective weight by using:
3281 * dw_i = S * (s'_i - s_i) (3)
3283 * Where 'S' is the group weight as seen by its parent.
3285 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3286 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3287 * 4/7) times the weight of the group.
3289 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3291 struct sched_entity *se = tg->se[cpu];
3293 if (!tg->parent) /* the trivial, non-cgroup case */
3296 for_each_sched_entity(se) {
3302 * W = @wg + \Sum rw_j
3304 W = wg + calc_tg_weight(tg, se->my_q);
3309 w = se->my_q->load.weight + wl;
3312 * wl = S * s'_i; see (2)
3315 wl = (w * tg->shares) / W;
3320 * Per the above, wl is the new se->load.weight value; since
3321 * those are clipped to [MIN_SHARES, ...) do so now. See
3322 * calc_cfs_shares().
3324 if (wl < MIN_SHARES)
3328 * wl = dw_i = S * (s'_i - s_i); see (3)
3330 wl -= se->load.weight;
3333 * Recursively apply this logic to all parent groups to compute
3334 * the final effective load change on the root group. Since
3335 * only the @tg group gets extra weight, all parent groups can
3336 * only redistribute existing shares. @wl is the shift in shares
3337 * resulting from this level per the above.
3346 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3347 unsigned long wl, unsigned long wg)
3354 static int wake_wide(struct task_struct *p)
3356 int factor = this_cpu_read(sd_llc_size);
3359 * Yeah, it's the switching-frequency, could means many wakee or
3360 * rapidly switch, use factor here will just help to automatically
3361 * adjust the loose-degree, so bigger node will lead to more pull.
3363 if (p->wakee_flips > factor) {
3365 * wakee is somewhat hot, it needs certain amount of cpu
3366 * resource, so if waker is far more hot, prefer to leave
3369 if (current->wakee_flips > (factor * p->wakee_flips))
3376 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3378 s64 this_load, load;
3379 int idx, this_cpu, prev_cpu;
3380 unsigned long tl_per_task;
3381 struct task_group *tg;
3382 unsigned long weight;
3386 * If we wake multiple tasks be careful to not bounce
3387 * ourselves around too much.
3393 this_cpu = smp_processor_id();
3394 prev_cpu = task_cpu(p);
3395 load = source_load(prev_cpu, idx);
3396 this_load = target_load(this_cpu, idx);
3399 * If sync wakeup then subtract the (maximum possible)
3400 * effect of the currently running task from the load
3401 * of the current CPU:
3404 tg = task_group(current);
3405 weight = current->se.load.weight;
3407 this_load += effective_load(tg, this_cpu, -weight, -weight);
3408 load += effective_load(tg, prev_cpu, 0, -weight);
3412 weight = p->se.load.weight;
3415 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3416 * due to the sync cause above having dropped this_load to 0, we'll
3417 * always have an imbalance, but there's really nothing you can do
3418 * about that, so that's good too.
3420 * Otherwise check if either cpus are near enough in load to allow this
3421 * task to be woken on this_cpu.
3423 if (this_load > 0) {
3424 s64 this_eff_load, prev_eff_load;
3426 this_eff_load = 100;
3427 this_eff_load *= power_of(prev_cpu);
3428 this_eff_load *= this_load +
3429 effective_load(tg, this_cpu, weight, weight);
3431 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3432 prev_eff_load *= power_of(this_cpu);
3433 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3435 balanced = this_eff_load <= prev_eff_load;
3440 * If the currently running task will sleep within
3441 * a reasonable amount of time then attract this newly
3444 if (sync && balanced)
3447 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3448 tl_per_task = cpu_avg_load_per_task(this_cpu);
3451 (this_load <= load &&
3452 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3454 * This domain has SD_WAKE_AFFINE and
3455 * p is cache cold in this domain, and
3456 * there is no bad imbalance.
3458 schedstat_inc(sd, ttwu_move_affine);
3459 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3467 * find_idlest_group finds and returns the least busy CPU group within the
3470 static struct sched_group *
3471 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3472 int this_cpu, int load_idx)
3474 struct sched_group *idlest = NULL, *group = sd->groups;
3475 unsigned long min_load = ULONG_MAX, this_load = 0;
3476 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3479 unsigned long load, avg_load;
3483 /* Skip over this group if it has no CPUs allowed */
3484 if (!cpumask_intersects(sched_group_cpus(group),
3485 tsk_cpus_allowed(p)))
3488 local_group = cpumask_test_cpu(this_cpu,
3489 sched_group_cpus(group));
3491 /* Tally up the load of all CPUs in the group */
3494 for_each_cpu(i, sched_group_cpus(group)) {
3495 /* Bias balancing toward cpus of our domain */
3497 load = source_load(i, load_idx);
3499 load = target_load(i, load_idx);
3504 /* Adjust by relative CPU power of the group */
3505 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3508 this_load = avg_load;
3509 } else if (avg_load < min_load) {
3510 min_load = avg_load;
3513 } while (group = group->next, group != sd->groups);
3515 if (!idlest || 100*this_load < imbalance*min_load)
3521 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3524 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3526 unsigned long load, min_load = ULONG_MAX;
3530 /* Traverse only the allowed CPUs */
3531 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3532 load = weighted_cpuload(i);
3534 if (load < min_load || (load == min_load && i == this_cpu)) {
3544 * Try and locate an idle CPU in the sched_domain.
3546 static int select_idle_sibling(struct task_struct *p, int target)
3548 struct sched_domain *sd;
3549 struct sched_group *sg;
3550 int i = task_cpu(p);
3552 if (idle_cpu(target))
3556 * If the prevous cpu is cache affine and idle, don't be stupid.
3558 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3562 * Otherwise, iterate the domains and find an elegible idle cpu.
3564 sd = rcu_dereference(per_cpu(sd_llc, target));
3565 for_each_lower_domain(sd) {
3568 if (!cpumask_intersects(sched_group_cpus(sg),
3569 tsk_cpus_allowed(p)))
3572 for_each_cpu(i, sched_group_cpus(sg)) {
3573 if (i == target || !idle_cpu(i))
3577 target = cpumask_first_and(sched_group_cpus(sg),
3578 tsk_cpus_allowed(p));
3582 } while (sg != sd->groups);
3589 * sched_balance_self: balance the current task (running on cpu) in domains
3590 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3593 * Balance, ie. select the least loaded group.
3595 * Returns the target CPU number, or the same CPU if no balancing is needed.
3597 * preempt must be disabled.
3600 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3602 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3603 int cpu = smp_processor_id();
3604 int prev_cpu = task_cpu(p);
3606 int want_affine = 0;
3607 int sync = wake_flags & WF_SYNC;
3609 if (p->nr_cpus_allowed == 1)
3612 if (sd_flag & SD_BALANCE_WAKE) {
3613 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3619 for_each_domain(cpu, tmp) {
3620 if (!(tmp->flags & SD_LOAD_BALANCE))
3624 * If both cpu and prev_cpu are part of this domain,
3625 * cpu is a valid SD_WAKE_AFFINE target.
3627 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3628 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3633 if (tmp->flags & sd_flag)
3638 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3641 new_cpu = select_idle_sibling(p, prev_cpu);
3646 int load_idx = sd->forkexec_idx;
3647 struct sched_group *group;
3650 if (!(sd->flags & sd_flag)) {
3655 if (sd_flag & SD_BALANCE_WAKE)
3656 load_idx = sd->wake_idx;
3658 group = find_idlest_group(sd, p, cpu, load_idx);
3664 new_cpu = find_idlest_cpu(group, p, cpu);
3665 if (new_cpu == -1 || new_cpu == cpu) {
3666 /* Now try balancing at a lower domain level of cpu */
3671 /* Now try balancing at a lower domain level of new_cpu */
3673 weight = sd->span_weight;
3675 for_each_domain(cpu, tmp) {
3676 if (weight <= tmp->span_weight)
3678 if (tmp->flags & sd_flag)
3681 /* while loop will break here if sd == NULL */
3690 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3691 * cfs_rq_of(p) references at time of call are still valid and identify the
3692 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3693 * other assumptions, including the state of rq->lock, should be made.
3696 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3698 struct sched_entity *se = &p->se;
3699 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3702 * Load tracking: accumulate removed load so that it can be processed
3703 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3704 * to blocked load iff they have a positive decay-count. It can never
3705 * be negative here since on-rq tasks have decay-count == 0.
3707 if (se->avg.decay_count) {
3708 se->avg.decay_count = -__synchronize_entity_decay(se);
3709 atomic_long_add(se->avg.load_avg_contrib,
3710 &cfs_rq->removed_load);
3713 #endif /* CONFIG_SMP */
3715 static unsigned long
3716 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3718 unsigned long gran = sysctl_sched_wakeup_granularity;
3721 * Since its curr running now, convert the gran from real-time
3722 * to virtual-time in his units.
3724 * By using 'se' instead of 'curr' we penalize light tasks, so
3725 * they get preempted easier. That is, if 'se' < 'curr' then
3726 * the resulting gran will be larger, therefore penalizing the
3727 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3728 * be smaller, again penalizing the lighter task.
3730 * This is especially important for buddies when the leftmost
3731 * task is higher priority than the buddy.
3733 return calc_delta_fair(gran, se);
3737 * Should 'se' preempt 'curr'.
3751 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3753 s64 gran, vdiff = curr->vruntime - se->vruntime;
3758 gran = wakeup_gran(curr, se);
3765 static void set_last_buddy(struct sched_entity *se)
3767 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3770 for_each_sched_entity(se)
3771 cfs_rq_of(se)->last = se;
3774 static void set_next_buddy(struct sched_entity *se)
3776 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3779 for_each_sched_entity(se)
3780 cfs_rq_of(se)->next = se;
3783 static void set_skip_buddy(struct sched_entity *se)
3785 for_each_sched_entity(se)
3786 cfs_rq_of(se)->skip = se;
3790 * Preempt the current task with a newly woken task if needed:
3792 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3794 struct task_struct *curr = rq->curr;
3795 struct sched_entity *se = &curr->se, *pse = &p->se;
3796 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3797 int scale = cfs_rq->nr_running >= sched_nr_latency;
3798 int next_buddy_marked = 0;
3800 if (unlikely(se == pse))
3804 * This is possible from callers such as move_task(), in which we
3805 * unconditionally check_prempt_curr() after an enqueue (which may have
3806 * lead to a throttle). This both saves work and prevents false
3807 * next-buddy nomination below.
3809 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3812 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3813 set_next_buddy(pse);
3814 next_buddy_marked = 1;
3818 * We can come here with TIF_NEED_RESCHED already set from new task
3821 * Note: this also catches the edge-case of curr being in a throttled
3822 * group (e.g. via set_curr_task), since update_curr() (in the
3823 * enqueue of curr) will have resulted in resched being set. This
3824 * prevents us from potentially nominating it as a false LAST_BUDDY
3827 if (test_tsk_need_resched(curr))
3830 /* Idle tasks are by definition preempted by non-idle tasks. */
3831 if (unlikely(curr->policy == SCHED_IDLE) &&
3832 likely(p->policy != SCHED_IDLE))
3836 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3837 * is driven by the tick):
3839 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3842 find_matching_se(&se, &pse);
3843 update_curr(cfs_rq_of(se));
3845 if (wakeup_preempt_entity(se, pse) == 1) {
3847 * Bias pick_next to pick the sched entity that is
3848 * triggering this preemption.
3850 if (!next_buddy_marked)
3851 set_next_buddy(pse);
3860 * Only set the backward buddy when the current task is still
3861 * on the rq. This can happen when a wakeup gets interleaved
3862 * with schedule on the ->pre_schedule() or idle_balance()
3863 * point, either of which can * drop the rq lock.
3865 * Also, during early boot the idle thread is in the fair class,
3866 * for obvious reasons its a bad idea to schedule back to it.
3868 if (unlikely(!se->on_rq || curr == rq->idle))
3871 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3875 static struct task_struct *pick_next_task_fair(struct rq *rq)
3877 struct task_struct *p;
3878 struct cfs_rq *cfs_rq = &rq->cfs;
3879 struct sched_entity *se;
3881 if (!cfs_rq->nr_running)
3885 se = pick_next_entity(cfs_rq);
3886 set_next_entity(cfs_rq, se);
3887 cfs_rq = group_cfs_rq(se);
3891 if (hrtick_enabled(rq))
3892 hrtick_start_fair(rq, p);
3898 * Account for a descheduled task:
3900 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3902 struct sched_entity *se = &prev->se;
3903 struct cfs_rq *cfs_rq;
3905 for_each_sched_entity(se) {
3906 cfs_rq = cfs_rq_of(se);
3907 put_prev_entity(cfs_rq, se);
3912 * sched_yield() is very simple
3914 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3916 static void yield_task_fair(struct rq *rq)
3918 struct task_struct *curr = rq->curr;
3919 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3920 struct sched_entity *se = &curr->se;
3923 * Are we the only task in the tree?
3925 if (unlikely(rq->nr_running == 1))
3928 clear_buddies(cfs_rq, se);
3930 if (curr->policy != SCHED_BATCH) {
3931 update_rq_clock(rq);
3933 * Update run-time statistics of the 'current'.
3935 update_curr(cfs_rq);
3937 * Tell update_rq_clock() that we've just updated,
3938 * so we don't do microscopic update in schedule()
3939 * and double the fastpath cost.
3941 rq->skip_clock_update = 1;
3947 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3949 struct sched_entity *se = &p->se;
3951 /* throttled hierarchies are not runnable */
3952 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3955 /* Tell the scheduler that we'd really like pse to run next. */
3958 yield_task_fair(rq);
3964 /**************************************************
3965 * Fair scheduling class load-balancing methods.
3969 * The purpose of load-balancing is to achieve the same basic fairness the
3970 * per-cpu scheduler provides, namely provide a proportional amount of compute
3971 * time to each task. This is expressed in the following equation:
3973 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3975 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3976 * W_i,0 is defined as:
3978 * W_i,0 = \Sum_j w_i,j (2)
3980 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3981 * is derived from the nice value as per prio_to_weight[].
3983 * The weight average is an exponential decay average of the instantaneous
3986 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3988 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3989 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3990 * can also include other factors [XXX].
3992 * To achieve this balance we define a measure of imbalance which follows
3993 * directly from (1):
3995 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3997 * We them move tasks around to minimize the imbalance. In the continuous
3998 * function space it is obvious this converges, in the discrete case we get
3999 * a few fun cases generally called infeasible weight scenarios.
4002 * - infeasible weights;
4003 * - local vs global optima in the discrete case. ]
4008 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4009 * for all i,j solution, we create a tree of cpus that follows the hardware
4010 * topology where each level pairs two lower groups (or better). This results
4011 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4012 * tree to only the first of the previous level and we decrease the frequency
4013 * of load-balance at each level inv. proportional to the number of cpus in
4019 * \Sum { --- * --- * 2^i } = O(n) (5)
4021 * `- size of each group
4022 * | | `- number of cpus doing load-balance
4024 * `- sum over all levels
4026 * Coupled with a limit on how many tasks we can migrate every balance pass,
4027 * this makes (5) the runtime complexity of the balancer.
4029 * An important property here is that each CPU is still (indirectly) connected
4030 * to every other cpu in at most O(log n) steps:
4032 * The adjacency matrix of the resulting graph is given by:
4035 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4038 * And you'll find that:
4040 * A^(log_2 n)_i,j != 0 for all i,j (7)
4042 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4043 * The task movement gives a factor of O(m), giving a convergence complexity
4046 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4051 * In order to avoid CPUs going idle while there's still work to do, new idle
4052 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4053 * tree itself instead of relying on other CPUs to bring it work.
4055 * This adds some complexity to both (5) and (8) but it reduces the total idle
4063 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4066 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4071 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4073 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4075 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4078 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4079 * rewrite all of this once again.]
4082 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4084 #define LBF_ALL_PINNED 0x01
4085 #define LBF_NEED_BREAK 0x02
4086 #define LBF_DST_PINNED 0x04
4087 #define LBF_SOME_PINNED 0x08
4090 struct sched_domain *sd;
4098 struct cpumask *dst_grpmask;
4100 enum cpu_idle_type idle;
4102 /* The set of CPUs under consideration for load-balancing */
4103 struct cpumask *cpus;
4108 unsigned int loop_break;
4109 unsigned int loop_max;
4113 * move_task - move a task from one runqueue to another runqueue.
4114 * Both runqueues must be locked.
4116 static void move_task(struct task_struct *p, struct lb_env *env)
4118 deactivate_task(env->src_rq, p, 0);
4119 set_task_cpu(p, env->dst_cpu);
4120 activate_task(env->dst_rq, p, 0);
4121 check_preempt_curr(env->dst_rq, p, 0);
4125 * Is this task likely cache-hot:
4128 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4132 if (p->sched_class != &fair_sched_class)
4135 if (unlikely(p->policy == SCHED_IDLE))
4139 * Buddy candidates are cache hot:
4141 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4142 (&p->se == cfs_rq_of(&p->se)->next ||
4143 &p->se == cfs_rq_of(&p->se)->last))
4146 if (sysctl_sched_migration_cost == -1)
4148 if (sysctl_sched_migration_cost == 0)
4151 delta = now - p->se.exec_start;
4153 return delta < (s64)sysctl_sched_migration_cost;
4156 #ifdef CONFIG_NUMA_BALANCING
4157 /* Returns true if the destination node has incurred more faults */
4158 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4160 int src_nid, dst_nid;
4162 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4163 !(env->sd->flags & SD_NUMA)) {
4167 src_nid = cpu_to_node(env->src_cpu);
4168 dst_nid = cpu_to_node(env->dst_cpu);
4170 if (src_nid == dst_nid ||
4171 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4174 if (dst_nid == p->numa_preferred_nid ||
4175 task_faults(p, dst_nid) > task_faults(p, src_nid))
4182 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4184 int src_nid, dst_nid;
4186 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4189 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4192 src_nid = cpu_to_node(env->src_cpu);
4193 dst_nid = cpu_to_node(env->dst_cpu);
4195 if (src_nid == dst_nid ||
4196 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4199 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4206 static inline bool migrate_improves_locality(struct task_struct *p,
4212 static inline bool migrate_degrades_locality(struct task_struct *p,
4220 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4223 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4225 int tsk_cache_hot = 0;
4227 * We do not migrate tasks that are:
4228 * 1) throttled_lb_pair, or
4229 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4230 * 3) running (obviously), or
4231 * 4) are cache-hot on their current CPU.
4233 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4236 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4239 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4241 env->flags |= LBF_SOME_PINNED;
4244 * Remember if this task can be migrated to any other cpu in
4245 * our sched_group. We may want to revisit it if we couldn't
4246 * meet load balance goals by pulling other tasks on src_cpu.
4248 * Also avoid computing new_dst_cpu if we have already computed
4249 * one in current iteration.
4251 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4254 /* Prevent to re-select dst_cpu via env's cpus */
4255 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4256 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4257 env->flags |= LBF_DST_PINNED;
4258 env->new_dst_cpu = cpu;
4266 /* Record that we found atleast one task that could run on dst_cpu */
4267 env->flags &= ~LBF_ALL_PINNED;
4269 if (task_running(env->src_rq, p)) {
4270 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4275 * Aggressive migration if:
4276 * 1) destination numa is preferred
4277 * 2) task is cache cold, or
4278 * 3) too many balance attempts have failed.
4280 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4282 tsk_cache_hot = migrate_degrades_locality(p, env);
4284 if (migrate_improves_locality(p, env)) {
4285 #ifdef CONFIG_SCHEDSTATS
4286 if (tsk_cache_hot) {
4287 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4288 schedstat_inc(p, se.statistics.nr_forced_migrations);
4294 if (!tsk_cache_hot ||
4295 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4297 if (tsk_cache_hot) {
4298 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4299 schedstat_inc(p, se.statistics.nr_forced_migrations);
4305 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4310 * move_one_task tries to move exactly one task from busiest to this_rq, as
4311 * part of active balancing operations within "domain".
4312 * Returns 1 if successful and 0 otherwise.
4314 * Called with both runqueues locked.
4316 static int move_one_task(struct lb_env *env)
4318 struct task_struct *p, *n;
4320 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4321 if (!can_migrate_task(p, env))
4326 * Right now, this is only the second place move_task()
4327 * is called, so we can safely collect move_task()
4328 * stats here rather than inside move_task().
4330 schedstat_inc(env->sd, lb_gained[env->idle]);
4336 static unsigned long task_h_load(struct task_struct *p);
4338 static const unsigned int sched_nr_migrate_break = 32;
4341 * move_tasks tries to move up to imbalance weighted load from busiest to
4342 * this_rq, as part of a balancing operation within domain "sd".
4343 * Returns 1 if successful and 0 otherwise.
4345 * Called with both runqueues locked.
4347 static int move_tasks(struct lb_env *env)
4349 struct list_head *tasks = &env->src_rq->cfs_tasks;
4350 struct task_struct *p;
4354 if (env->imbalance <= 0)
4357 while (!list_empty(tasks)) {
4358 p = list_first_entry(tasks, struct task_struct, se.group_node);
4361 /* We've more or less seen every task there is, call it quits */
4362 if (env->loop > env->loop_max)
4365 /* take a breather every nr_migrate tasks */
4366 if (env->loop > env->loop_break) {
4367 env->loop_break += sched_nr_migrate_break;
4368 env->flags |= LBF_NEED_BREAK;
4372 if (!can_migrate_task(p, env))
4375 load = task_h_load(p);
4377 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4380 if ((load / 2) > env->imbalance)
4385 env->imbalance -= load;
4387 #ifdef CONFIG_PREEMPT
4389 * NEWIDLE balancing is a source of latency, so preemptible
4390 * kernels will stop after the first task is pulled to minimize
4391 * the critical section.
4393 if (env->idle == CPU_NEWLY_IDLE)
4398 * We only want to steal up to the prescribed amount of
4401 if (env->imbalance <= 0)
4406 list_move_tail(&p->se.group_node, tasks);
4410 * Right now, this is one of only two places move_task() is called,
4411 * so we can safely collect move_task() stats here rather than
4412 * inside move_task().
4414 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4419 #ifdef CONFIG_FAIR_GROUP_SCHED
4421 * update tg->load_weight by folding this cpu's load_avg
4423 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4425 struct sched_entity *se = tg->se[cpu];
4426 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4428 /* throttled entities do not contribute to load */
4429 if (throttled_hierarchy(cfs_rq))
4432 update_cfs_rq_blocked_load(cfs_rq, 1);
4435 update_entity_load_avg(se, 1);
4437 * We pivot on our runnable average having decayed to zero for
4438 * list removal. This generally implies that all our children
4439 * have also been removed (modulo rounding error or bandwidth
4440 * control); however, such cases are rare and we can fix these
4443 * TODO: fix up out-of-order children on enqueue.
4445 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4446 list_del_leaf_cfs_rq(cfs_rq);
4448 struct rq *rq = rq_of(cfs_rq);
4449 update_rq_runnable_avg(rq, rq->nr_running);
4453 static void update_blocked_averages(int cpu)
4455 struct rq *rq = cpu_rq(cpu);
4456 struct cfs_rq *cfs_rq;
4457 unsigned long flags;
4459 raw_spin_lock_irqsave(&rq->lock, flags);
4460 update_rq_clock(rq);
4462 * Iterates the task_group tree in a bottom up fashion, see
4463 * list_add_leaf_cfs_rq() for details.
4465 for_each_leaf_cfs_rq(rq, cfs_rq) {
4467 * Note: We may want to consider periodically releasing
4468 * rq->lock about these updates so that creating many task
4469 * groups does not result in continually extending hold time.
4471 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4474 raw_spin_unlock_irqrestore(&rq->lock, flags);
4478 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4479 * This needs to be done in a top-down fashion because the load of a child
4480 * group is a fraction of its parents load.
4482 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4484 struct rq *rq = rq_of(cfs_rq);
4485 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4486 unsigned long now = jiffies;
4489 if (cfs_rq->last_h_load_update == now)
4492 cfs_rq->h_load_next = NULL;
4493 for_each_sched_entity(se) {
4494 cfs_rq = cfs_rq_of(se);
4495 cfs_rq->h_load_next = se;
4496 if (cfs_rq->last_h_load_update == now)
4501 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4502 cfs_rq->last_h_load_update = now;
4505 while ((se = cfs_rq->h_load_next) != NULL) {
4506 load = cfs_rq->h_load;
4507 load = div64_ul(load * se->avg.load_avg_contrib,
4508 cfs_rq->runnable_load_avg + 1);
4509 cfs_rq = group_cfs_rq(se);
4510 cfs_rq->h_load = load;
4511 cfs_rq->last_h_load_update = now;
4515 static unsigned long task_h_load(struct task_struct *p)
4517 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4519 update_cfs_rq_h_load(cfs_rq);
4520 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4521 cfs_rq->runnable_load_avg + 1);
4524 static inline void update_blocked_averages(int cpu)
4528 static unsigned long task_h_load(struct task_struct *p)
4530 return p->se.avg.load_avg_contrib;
4534 /********** Helpers for find_busiest_group ************************/
4536 * sg_lb_stats - stats of a sched_group required for load_balancing
4538 struct sg_lb_stats {
4539 unsigned long avg_load; /*Avg load across the CPUs of the group */
4540 unsigned long group_load; /* Total load over the CPUs of the group */
4541 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4542 unsigned long load_per_task;
4543 unsigned long group_power;
4544 unsigned int sum_nr_running; /* Nr tasks running in the group */
4545 unsigned int group_capacity;
4546 unsigned int idle_cpus;
4547 unsigned int group_weight;
4548 int group_imb; /* Is there an imbalance in the group ? */
4549 int group_has_capacity; /* Is there extra capacity in the group? */
4553 * sd_lb_stats - Structure to store the statistics of a sched_domain
4554 * during load balancing.
4556 struct sd_lb_stats {
4557 struct sched_group *busiest; /* Busiest group in this sd */
4558 struct sched_group *local; /* Local group in this sd */
4559 unsigned long total_load; /* Total load of all groups in sd */
4560 unsigned long total_pwr; /* Total power of all groups in sd */
4561 unsigned long avg_load; /* Average load across all groups in sd */
4563 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4564 struct sg_lb_stats local_stat; /* Statistics of the local group */
4567 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4570 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4571 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4572 * We must however clear busiest_stat::avg_load because
4573 * update_sd_pick_busiest() reads this before assignment.
4575 *sds = (struct sd_lb_stats){
4587 * get_sd_load_idx - Obtain the load index for a given sched domain.
4588 * @sd: The sched_domain whose load_idx is to be obtained.
4589 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4591 * Return: The load index.
4593 static inline int get_sd_load_idx(struct sched_domain *sd,
4594 enum cpu_idle_type idle)
4600 load_idx = sd->busy_idx;
4603 case CPU_NEWLY_IDLE:
4604 load_idx = sd->newidle_idx;
4607 load_idx = sd->idle_idx;
4614 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4616 return SCHED_POWER_SCALE;
4619 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4621 return default_scale_freq_power(sd, cpu);
4624 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4626 unsigned long weight = sd->span_weight;
4627 unsigned long smt_gain = sd->smt_gain;
4634 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4636 return default_scale_smt_power(sd, cpu);
4639 static unsigned long scale_rt_power(int cpu)
4641 struct rq *rq = cpu_rq(cpu);
4642 u64 total, available, age_stamp, avg;
4645 * Since we're reading these variables without serialization make sure
4646 * we read them once before doing sanity checks on them.
4648 age_stamp = ACCESS_ONCE(rq->age_stamp);
4649 avg = ACCESS_ONCE(rq->rt_avg);
4651 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4653 if (unlikely(total < avg)) {
4654 /* Ensures that power won't end up being negative */
4657 available = total - avg;
4660 if (unlikely((s64)total < SCHED_POWER_SCALE))
4661 total = SCHED_POWER_SCALE;
4663 total >>= SCHED_POWER_SHIFT;
4665 return div_u64(available, total);
4668 static void update_cpu_power(struct sched_domain *sd, int cpu)
4670 unsigned long weight = sd->span_weight;
4671 unsigned long power = SCHED_POWER_SCALE;
4672 struct sched_group *sdg = sd->groups;
4674 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4675 if (sched_feat(ARCH_POWER))
4676 power *= arch_scale_smt_power(sd, cpu);
4678 power *= default_scale_smt_power(sd, cpu);
4680 power >>= SCHED_POWER_SHIFT;
4683 sdg->sgp->power_orig = power;
4685 if (sched_feat(ARCH_POWER))
4686 power *= arch_scale_freq_power(sd, cpu);
4688 power *= default_scale_freq_power(sd, cpu);
4690 power >>= SCHED_POWER_SHIFT;
4692 power *= scale_rt_power(cpu);
4693 power >>= SCHED_POWER_SHIFT;
4698 cpu_rq(cpu)->cpu_power = power;
4699 sdg->sgp->power = power;
4702 void update_group_power(struct sched_domain *sd, int cpu)
4704 struct sched_domain *child = sd->child;
4705 struct sched_group *group, *sdg = sd->groups;
4706 unsigned long power, power_orig;
4707 unsigned long interval;
4709 interval = msecs_to_jiffies(sd->balance_interval);
4710 interval = clamp(interval, 1UL, max_load_balance_interval);
4711 sdg->sgp->next_update = jiffies + interval;
4714 update_cpu_power(sd, cpu);
4718 power_orig = power = 0;
4720 if (child->flags & SD_OVERLAP) {
4722 * SD_OVERLAP domains cannot assume that child groups
4723 * span the current group.
4726 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4727 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4729 power_orig += sg->sgp->power_orig;
4730 power += sg->sgp->power;
4734 * !SD_OVERLAP domains can assume that child groups
4735 * span the current group.
4738 group = child->groups;
4740 power_orig += group->sgp->power_orig;
4741 power += group->sgp->power;
4742 group = group->next;
4743 } while (group != child->groups);
4746 sdg->sgp->power_orig = power_orig;
4747 sdg->sgp->power = power;
4751 * Try and fix up capacity for tiny siblings, this is needed when
4752 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4753 * which on its own isn't powerful enough.
4755 * See update_sd_pick_busiest() and check_asym_packing().
4758 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4761 * Only siblings can have significantly less than SCHED_POWER_SCALE
4763 if (!(sd->flags & SD_SHARE_CPUPOWER))
4767 * If ~90% of the cpu_power is still there, we're good.
4769 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4776 * Group imbalance indicates (and tries to solve) the problem where balancing
4777 * groups is inadequate due to tsk_cpus_allowed() constraints.
4779 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4780 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4783 * { 0 1 2 3 } { 4 5 6 7 }
4786 * If we were to balance group-wise we'd place two tasks in the first group and
4787 * two tasks in the second group. Clearly this is undesired as it will overload
4788 * cpu 3 and leave one of the cpus in the second group unused.
4790 * The current solution to this issue is detecting the skew in the first group
4791 * by noticing the lower domain failed to reach balance and had difficulty
4792 * moving tasks due to affinity constraints.
4794 * When this is so detected; this group becomes a candidate for busiest; see
4795 * update_sd_pick_busiest(). And calculcate_imbalance() and
4796 * find_busiest_group() avoid some of the usual balance conditions to allow it
4797 * to create an effective group imbalance.
4799 * This is a somewhat tricky proposition since the next run might not find the
4800 * group imbalance and decide the groups need to be balanced again. A most
4801 * subtle and fragile situation.
4804 static inline int sg_imbalanced(struct sched_group *group)
4806 return group->sgp->imbalance;
4810 * Compute the group capacity.
4812 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4813 * first dividing out the smt factor and computing the actual number of cores
4814 * and limit power unit capacity with that.
4816 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4818 unsigned int capacity, smt, cpus;
4819 unsigned int power, power_orig;
4821 power = group->sgp->power;
4822 power_orig = group->sgp->power_orig;
4823 cpus = group->group_weight;
4825 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4826 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4827 capacity = cpus / smt; /* cores */
4829 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4831 capacity = fix_small_capacity(env->sd, group);
4837 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4838 * @env: The load balancing environment.
4839 * @group: sched_group whose statistics are to be updated.
4840 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4841 * @local_group: Does group contain this_cpu.
4842 * @sgs: variable to hold the statistics for this group.
4844 static inline void update_sg_lb_stats(struct lb_env *env,
4845 struct sched_group *group, int load_idx,
4846 int local_group, struct sg_lb_stats *sgs)
4848 unsigned long nr_running;
4852 memset(sgs, 0, sizeof(*sgs));
4854 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4855 struct rq *rq = cpu_rq(i);
4857 nr_running = rq->nr_running;
4859 /* Bias balancing toward cpus of our domain */
4861 load = target_load(i, load_idx);
4863 load = source_load(i, load_idx);
4865 sgs->group_load += load;
4866 sgs->sum_nr_running += nr_running;
4867 sgs->sum_weighted_load += weighted_cpuload(i);
4872 /* Adjust by relative CPU power of the group */
4873 sgs->group_power = group->sgp->power;
4874 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4876 if (sgs->sum_nr_running)
4877 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4879 sgs->group_weight = group->group_weight;
4881 sgs->group_imb = sg_imbalanced(group);
4882 sgs->group_capacity = sg_capacity(env, group);
4884 if (sgs->group_capacity > sgs->sum_nr_running)
4885 sgs->group_has_capacity = 1;
4889 * update_sd_pick_busiest - return 1 on busiest group
4890 * @env: The load balancing environment.
4891 * @sds: sched_domain statistics
4892 * @sg: sched_group candidate to be checked for being the busiest
4893 * @sgs: sched_group statistics
4895 * Determine if @sg is a busier group than the previously selected
4898 * Return: %true if @sg is a busier group than the previously selected
4899 * busiest group. %false otherwise.
4901 static bool update_sd_pick_busiest(struct lb_env *env,
4902 struct sd_lb_stats *sds,
4903 struct sched_group *sg,
4904 struct sg_lb_stats *sgs)
4906 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4909 if (sgs->sum_nr_running > sgs->group_capacity)
4916 * ASYM_PACKING needs to move all the work to the lowest
4917 * numbered CPUs in the group, therefore mark all groups
4918 * higher than ourself as busy.
4920 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4921 env->dst_cpu < group_first_cpu(sg)) {
4925 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4933 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4934 * @env: The load balancing environment.
4935 * @balance: Should we balance.
4936 * @sds: variable to hold the statistics for this sched_domain.
4938 static inline void update_sd_lb_stats(struct lb_env *env,
4939 struct sd_lb_stats *sds)
4941 struct sched_domain *child = env->sd->child;
4942 struct sched_group *sg = env->sd->groups;
4943 struct sg_lb_stats tmp_sgs;
4944 int load_idx, prefer_sibling = 0;
4946 if (child && child->flags & SD_PREFER_SIBLING)
4949 load_idx = get_sd_load_idx(env->sd, env->idle);
4952 struct sg_lb_stats *sgs = &tmp_sgs;
4955 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4958 sgs = &sds->local_stat;
4960 if (env->idle != CPU_NEWLY_IDLE ||
4961 time_after_eq(jiffies, sg->sgp->next_update))
4962 update_group_power(env->sd, env->dst_cpu);
4965 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4971 * In case the child domain prefers tasks go to siblings
4972 * first, lower the sg capacity to one so that we'll try
4973 * and move all the excess tasks away. We lower the capacity
4974 * of a group only if the local group has the capacity to fit
4975 * these excess tasks, i.e. nr_running < group_capacity. The
4976 * extra check prevents the case where you always pull from the
4977 * heaviest group when it is already under-utilized (possible
4978 * with a large weight task outweighs the tasks on the system).
4980 if (prefer_sibling && sds->local &&
4981 sds->local_stat.group_has_capacity)
4982 sgs->group_capacity = min(sgs->group_capacity, 1U);
4984 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4986 sds->busiest_stat = *sgs;
4990 /* Now, start updating sd_lb_stats */
4991 sds->total_load += sgs->group_load;
4992 sds->total_pwr += sgs->group_power;
4995 } while (sg != env->sd->groups);
4999 * check_asym_packing - Check to see if the group is packed into the
5002 * This is primarily intended to used at the sibling level. Some
5003 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5004 * case of POWER7, it can move to lower SMT modes only when higher
5005 * threads are idle. When in lower SMT modes, the threads will
5006 * perform better since they share less core resources. Hence when we
5007 * have idle threads, we want them to be the higher ones.
5009 * This packing function is run on idle threads. It checks to see if
5010 * the busiest CPU in this domain (core in the P7 case) has a higher
5011 * CPU number than the packing function is being run on. Here we are
5012 * assuming lower CPU number will be equivalent to lower a SMT thread
5015 * Return: 1 when packing is required and a task should be moved to
5016 * this CPU. The amount of the imbalance is returned in *imbalance.
5018 * @env: The load balancing environment.
5019 * @sds: Statistics of the sched_domain which is to be packed
5021 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5025 if (!(env->sd->flags & SD_ASYM_PACKING))
5031 busiest_cpu = group_first_cpu(sds->busiest);
5032 if (env->dst_cpu > busiest_cpu)
5035 env->imbalance = DIV_ROUND_CLOSEST(
5036 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5043 * fix_small_imbalance - Calculate the minor imbalance that exists
5044 * amongst the groups of a sched_domain, during
5046 * @env: The load balancing environment.
5047 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5050 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5052 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5053 unsigned int imbn = 2;
5054 unsigned long scaled_busy_load_per_task;
5055 struct sg_lb_stats *local, *busiest;
5057 local = &sds->local_stat;
5058 busiest = &sds->busiest_stat;
5060 if (!local->sum_nr_running)
5061 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5062 else if (busiest->load_per_task > local->load_per_task)
5065 scaled_busy_load_per_task =
5066 (busiest->load_per_task * SCHED_POWER_SCALE) /
5067 busiest->group_power;
5069 if (busiest->avg_load + scaled_busy_load_per_task >=
5070 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5071 env->imbalance = busiest->load_per_task;
5076 * OK, we don't have enough imbalance to justify moving tasks,
5077 * however we may be able to increase total CPU power used by
5081 pwr_now += busiest->group_power *
5082 min(busiest->load_per_task, busiest->avg_load);
5083 pwr_now += local->group_power *
5084 min(local->load_per_task, local->avg_load);
5085 pwr_now /= SCHED_POWER_SCALE;
5087 /* Amount of load we'd subtract */
5088 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5089 busiest->group_power;
5090 if (busiest->avg_load > tmp) {
5091 pwr_move += busiest->group_power *
5092 min(busiest->load_per_task,
5093 busiest->avg_load - tmp);
5096 /* Amount of load we'd add */
5097 if (busiest->avg_load * busiest->group_power <
5098 busiest->load_per_task * SCHED_POWER_SCALE) {
5099 tmp = (busiest->avg_load * busiest->group_power) /
5102 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5105 pwr_move += local->group_power *
5106 min(local->load_per_task, local->avg_load + tmp);
5107 pwr_move /= SCHED_POWER_SCALE;
5109 /* Move if we gain throughput */
5110 if (pwr_move > pwr_now)
5111 env->imbalance = busiest->load_per_task;
5115 * calculate_imbalance - Calculate the amount of imbalance present within the
5116 * groups of a given sched_domain during load balance.
5117 * @env: load balance environment
5118 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5120 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5122 unsigned long max_pull, load_above_capacity = ~0UL;
5123 struct sg_lb_stats *local, *busiest;
5125 local = &sds->local_stat;
5126 busiest = &sds->busiest_stat;
5128 if (busiest->group_imb) {
5130 * In the group_imb case we cannot rely on group-wide averages
5131 * to ensure cpu-load equilibrium, look at wider averages. XXX
5133 busiest->load_per_task =
5134 min(busiest->load_per_task, sds->avg_load);
5138 * In the presence of smp nice balancing, certain scenarios can have
5139 * max load less than avg load(as we skip the groups at or below
5140 * its cpu_power, while calculating max_load..)
5142 if (busiest->avg_load <= sds->avg_load ||
5143 local->avg_load >= sds->avg_load) {
5145 return fix_small_imbalance(env, sds);
5148 if (!busiest->group_imb) {
5150 * Don't want to pull so many tasks that a group would go idle.
5151 * Except of course for the group_imb case, since then we might
5152 * have to drop below capacity to reach cpu-load equilibrium.
5154 load_above_capacity =
5155 (busiest->sum_nr_running - busiest->group_capacity);
5157 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5158 load_above_capacity /= busiest->group_power;
5162 * We're trying to get all the cpus to the average_load, so we don't
5163 * want to push ourselves above the average load, nor do we wish to
5164 * reduce the max loaded cpu below the average load. At the same time,
5165 * we also don't want to reduce the group load below the group capacity
5166 * (so that we can implement power-savings policies etc). Thus we look
5167 * for the minimum possible imbalance.
5169 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5171 /* How much load to actually move to equalise the imbalance */
5172 env->imbalance = min(
5173 max_pull * busiest->group_power,
5174 (sds->avg_load - local->avg_load) * local->group_power
5175 ) / SCHED_POWER_SCALE;
5178 * if *imbalance is less than the average load per runnable task
5179 * there is no guarantee that any tasks will be moved so we'll have
5180 * a think about bumping its value to force at least one task to be
5183 if (env->imbalance < busiest->load_per_task)
5184 return fix_small_imbalance(env, sds);
5187 /******* find_busiest_group() helpers end here *********************/
5190 * find_busiest_group - Returns the busiest group within the sched_domain
5191 * if there is an imbalance. If there isn't an imbalance, and
5192 * the user has opted for power-savings, it returns a group whose
5193 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5194 * such a group exists.
5196 * Also calculates the amount of weighted load which should be moved
5197 * to restore balance.
5199 * @env: The load balancing environment.
5201 * Return: - The busiest group if imbalance exists.
5202 * - If no imbalance and user has opted for power-savings balance,
5203 * return the least loaded group whose CPUs can be
5204 * put to idle by rebalancing its tasks onto our group.
5206 static struct sched_group *find_busiest_group(struct lb_env *env)
5208 struct sg_lb_stats *local, *busiest;
5209 struct sd_lb_stats sds;
5211 init_sd_lb_stats(&sds);
5214 * Compute the various statistics relavent for load balancing at
5217 update_sd_lb_stats(env, &sds);
5218 local = &sds.local_stat;
5219 busiest = &sds.busiest_stat;
5221 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5222 check_asym_packing(env, &sds))
5225 /* There is no busy sibling group to pull tasks from */
5226 if (!sds.busiest || busiest->sum_nr_running == 0)
5229 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5232 * If the busiest group is imbalanced the below checks don't
5233 * work because they assume all things are equal, which typically
5234 * isn't true due to cpus_allowed constraints and the like.
5236 if (busiest->group_imb)
5239 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5240 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5241 !busiest->group_has_capacity)
5245 * If the local group is more busy than the selected busiest group
5246 * don't try and pull any tasks.
5248 if (local->avg_load >= busiest->avg_load)
5252 * Don't pull any tasks if this group is already above the domain
5255 if (local->avg_load >= sds.avg_load)
5258 if (env->idle == CPU_IDLE) {
5260 * This cpu is idle. If the busiest group load doesn't
5261 * have more tasks than the number of available cpu's and
5262 * there is no imbalance between this and busiest group
5263 * wrt to idle cpu's, it is balanced.
5265 if ((local->idle_cpus < busiest->idle_cpus) &&
5266 busiest->sum_nr_running <= busiest->group_weight)
5270 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5271 * imbalance_pct to be conservative.
5273 if (100 * busiest->avg_load <=
5274 env->sd->imbalance_pct * local->avg_load)
5279 /* Looks like there is an imbalance. Compute it */
5280 calculate_imbalance(env, &sds);
5289 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5291 static struct rq *find_busiest_queue(struct lb_env *env,
5292 struct sched_group *group)
5294 struct rq *busiest = NULL, *rq;
5295 unsigned long busiest_load = 0, busiest_power = 1;
5298 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5299 unsigned long power = power_of(i);
5300 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5305 capacity = fix_small_capacity(env->sd, group);
5308 wl = weighted_cpuload(i);
5311 * When comparing with imbalance, use weighted_cpuload()
5312 * which is not scaled with the cpu power.
5314 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5318 * For the load comparisons with the other cpu's, consider
5319 * the weighted_cpuload() scaled with the cpu power, so that
5320 * the load can be moved away from the cpu that is potentially
5321 * running at a lower capacity.
5323 * Thus we're looking for max(wl_i / power_i), crosswise
5324 * multiplication to rid ourselves of the division works out
5325 * to: wl_i * power_j > wl_j * power_i; where j is our
5328 if (wl * busiest_power > busiest_load * power) {
5330 busiest_power = power;
5339 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5340 * so long as it is large enough.
5342 #define MAX_PINNED_INTERVAL 512
5344 /* Working cpumask for load_balance and load_balance_newidle. */
5345 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5347 static int need_active_balance(struct lb_env *env)
5349 struct sched_domain *sd = env->sd;
5351 if (env->idle == CPU_NEWLY_IDLE) {
5354 * ASYM_PACKING needs to force migrate tasks from busy but
5355 * higher numbered CPUs in order to pack all tasks in the
5356 * lowest numbered CPUs.
5358 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5362 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5365 static int active_load_balance_cpu_stop(void *data);
5367 static int should_we_balance(struct lb_env *env)
5369 struct sched_group *sg = env->sd->groups;
5370 struct cpumask *sg_cpus, *sg_mask;
5371 int cpu, balance_cpu = -1;
5374 * In the newly idle case, we will allow all the cpu's
5375 * to do the newly idle load balance.
5377 if (env->idle == CPU_NEWLY_IDLE)
5380 sg_cpus = sched_group_cpus(sg);
5381 sg_mask = sched_group_mask(sg);
5382 /* Try to find first idle cpu */
5383 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5384 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5391 if (balance_cpu == -1)
5392 balance_cpu = group_balance_cpu(sg);
5395 * First idle cpu or the first cpu(busiest) in this sched group
5396 * is eligible for doing load balancing at this and above domains.
5398 return balance_cpu == env->dst_cpu;
5402 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5403 * tasks if there is an imbalance.
5405 static int load_balance(int this_cpu, struct rq *this_rq,
5406 struct sched_domain *sd, enum cpu_idle_type idle,
5407 int *continue_balancing)
5409 int ld_moved, cur_ld_moved, active_balance = 0;
5410 struct sched_domain *sd_parent = sd->parent;
5411 struct sched_group *group;
5413 unsigned long flags;
5414 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5416 struct lb_env env = {
5418 .dst_cpu = this_cpu,
5420 .dst_grpmask = sched_group_cpus(sd->groups),
5422 .loop_break = sched_nr_migrate_break,
5427 * For NEWLY_IDLE load_balancing, we don't need to consider
5428 * other cpus in our group
5430 if (idle == CPU_NEWLY_IDLE)
5431 env.dst_grpmask = NULL;
5433 cpumask_copy(cpus, cpu_active_mask);
5435 schedstat_inc(sd, lb_count[idle]);
5438 if (!should_we_balance(&env)) {
5439 *continue_balancing = 0;
5443 group = find_busiest_group(&env);
5445 schedstat_inc(sd, lb_nobusyg[idle]);
5449 busiest = find_busiest_queue(&env, group);
5451 schedstat_inc(sd, lb_nobusyq[idle]);
5455 BUG_ON(busiest == env.dst_rq);
5457 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5460 if (busiest->nr_running > 1) {
5462 * Attempt to move tasks. If find_busiest_group has found
5463 * an imbalance but busiest->nr_running <= 1, the group is
5464 * still unbalanced. ld_moved simply stays zero, so it is
5465 * correctly treated as an imbalance.
5467 env.flags |= LBF_ALL_PINNED;
5468 env.src_cpu = busiest->cpu;
5469 env.src_rq = busiest;
5470 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5473 local_irq_save(flags);
5474 double_rq_lock(env.dst_rq, busiest);
5477 * cur_ld_moved - load moved in current iteration
5478 * ld_moved - cumulative load moved across iterations
5480 cur_ld_moved = move_tasks(&env);
5481 ld_moved += cur_ld_moved;
5482 double_rq_unlock(env.dst_rq, busiest);
5483 local_irq_restore(flags);
5486 * some other cpu did the load balance for us.
5488 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5489 resched_cpu(env.dst_cpu);
5491 if (env.flags & LBF_NEED_BREAK) {
5492 env.flags &= ~LBF_NEED_BREAK;
5497 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5498 * us and move them to an alternate dst_cpu in our sched_group
5499 * where they can run. The upper limit on how many times we
5500 * iterate on same src_cpu is dependent on number of cpus in our
5503 * This changes load balance semantics a bit on who can move
5504 * load to a given_cpu. In addition to the given_cpu itself
5505 * (or a ilb_cpu acting on its behalf where given_cpu is
5506 * nohz-idle), we now have balance_cpu in a position to move
5507 * load to given_cpu. In rare situations, this may cause
5508 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5509 * _independently_ and at _same_ time to move some load to
5510 * given_cpu) causing exceess load to be moved to given_cpu.
5511 * This however should not happen so much in practice and
5512 * moreover subsequent load balance cycles should correct the
5513 * excess load moved.
5515 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5517 /* Prevent to re-select dst_cpu via env's cpus */
5518 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5520 env.dst_rq = cpu_rq(env.new_dst_cpu);
5521 env.dst_cpu = env.new_dst_cpu;
5522 env.flags &= ~LBF_DST_PINNED;
5524 env.loop_break = sched_nr_migrate_break;
5527 * Go back to "more_balance" rather than "redo" since we
5528 * need to continue with same src_cpu.
5534 * We failed to reach balance because of affinity.
5537 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5539 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5540 *group_imbalance = 1;
5541 } else if (*group_imbalance)
5542 *group_imbalance = 0;
5545 /* All tasks on this runqueue were pinned by CPU affinity */
5546 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5547 cpumask_clear_cpu(cpu_of(busiest), cpus);
5548 if (!cpumask_empty(cpus)) {
5550 env.loop_break = sched_nr_migrate_break;
5558 schedstat_inc(sd, lb_failed[idle]);
5560 * Increment the failure counter only on periodic balance.
5561 * We do not want newidle balance, which can be very
5562 * frequent, pollute the failure counter causing
5563 * excessive cache_hot migrations and active balances.
5565 if (idle != CPU_NEWLY_IDLE)
5566 sd->nr_balance_failed++;
5568 if (need_active_balance(&env)) {
5569 raw_spin_lock_irqsave(&busiest->lock, flags);
5571 /* don't kick the active_load_balance_cpu_stop,
5572 * if the curr task on busiest cpu can't be
5575 if (!cpumask_test_cpu(this_cpu,
5576 tsk_cpus_allowed(busiest->curr))) {
5577 raw_spin_unlock_irqrestore(&busiest->lock,
5579 env.flags |= LBF_ALL_PINNED;
5580 goto out_one_pinned;
5584 * ->active_balance synchronizes accesses to
5585 * ->active_balance_work. Once set, it's cleared
5586 * only after active load balance is finished.
5588 if (!busiest->active_balance) {
5589 busiest->active_balance = 1;
5590 busiest->push_cpu = this_cpu;
5593 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5595 if (active_balance) {
5596 stop_one_cpu_nowait(cpu_of(busiest),
5597 active_load_balance_cpu_stop, busiest,
5598 &busiest->active_balance_work);
5602 * We've kicked active balancing, reset the failure
5605 sd->nr_balance_failed = sd->cache_nice_tries+1;
5608 sd->nr_balance_failed = 0;
5610 if (likely(!active_balance)) {
5611 /* We were unbalanced, so reset the balancing interval */
5612 sd->balance_interval = sd->min_interval;
5615 * If we've begun active balancing, start to back off. This
5616 * case may not be covered by the all_pinned logic if there
5617 * is only 1 task on the busy runqueue (because we don't call
5620 if (sd->balance_interval < sd->max_interval)
5621 sd->balance_interval *= 2;
5627 schedstat_inc(sd, lb_balanced[idle]);
5629 sd->nr_balance_failed = 0;
5632 /* tune up the balancing interval */
5633 if (((env.flags & LBF_ALL_PINNED) &&
5634 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5635 (sd->balance_interval < sd->max_interval))
5636 sd->balance_interval *= 2;
5644 * idle_balance is called by schedule() if this_cpu is about to become
5645 * idle. Attempts to pull tasks from other CPUs.
5647 void idle_balance(int this_cpu, struct rq *this_rq)
5649 struct sched_domain *sd;
5650 int pulled_task = 0;
5651 unsigned long next_balance = jiffies + HZ;
5654 this_rq->idle_stamp = rq_clock(this_rq);
5656 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5660 * Drop the rq->lock, but keep IRQ/preempt disabled.
5662 raw_spin_unlock(&this_rq->lock);
5664 update_blocked_averages(this_cpu);
5666 for_each_domain(this_cpu, sd) {
5667 unsigned long interval;
5668 int continue_balancing = 1;
5669 u64 t0, domain_cost;
5671 if (!(sd->flags & SD_LOAD_BALANCE))
5674 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5677 if (sd->flags & SD_BALANCE_NEWIDLE) {
5678 t0 = sched_clock_cpu(this_cpu);
5680 /* If we've pulled tasks over stop searching: */
5681 pulled_task = load_balance(this_cpu, this_rq,
5683 &continue_balancing);
5685 domain_cost = sched_clock_cpu(this_cpu) - t0;
5686 if (domain_cost > sd->max_newidle_lb_cost)
5687 sd->max_newidle_lb_cost = domain_cost;
5689 curr_cost += domain_cost;
5692 interval = msecs_to_jiffies(sd->balance_interval);
5693 if (time_after(next_balance, sd->last_balance + interval))
5694 next_balance = sd->last_balance + interval;
5696 this_rq->idle_stamp = 0;
5702 raw_spin_lock(&this_rq->lock);
5704 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5706 * We are going idle. next_balance may be set based on
5707 * a busy processor. So reset next_balance.
5709 this_rq->next_balance = next_balance;
5712 if (curr_cost > this_rq->max_idle_balance_cost)
5713 this_rq->max_idle_balance_cost = curr_cost;
5717 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5718 * running tasks off the busiest CPU onto idle CPUs. It requires at
5719 * least 1 task to be running on each physical CPU where possible, and
5720 * avoids physical / logical imbalances.
5722 static int active_load_balance_cpu_stop(void *data)
5724 struct rq *busiest_rq = data;
5725 int busiest_cpu = cpu_of(busiest_rq);
5726 int target_cpu = busiest_rq->push_cpu;
5727 struct rq *target_rq = cpu_rq(target_cpu);
5728 struct sched_domain *sd;
5730 raw_spin_lock_irq(&busiest_rq->lock);
5732 /* make sure the requested cpu hasn't gone down in the meantime */
5733 if (unlikely(busiest_cpu != smp_processor_id() ||
5734 !busiest_rq->active_balance))
5737 /* Is there any task to move? */
5738 if (busiest_rq->nr_running <= 1)
5742 * This condition is "impossible", if it occurs
5743 * we need to fix it. Originally reported by
5744 * Bjorn Helgaas on a 128-cpu setup.
5746 BUG_ON(busiest_rq == target_rq);
5748 /* move a task from busiest_rq to target_rq */
5749 double_lock_balance(busiest_rq, target_rq);
5751 /* Search for an sd spanning us and the target CPU. */
5753 for_each_domain(target_cpu, sd) {
5754 if ((sd->flags & SD_LOAD_BALANCE) &&
5755 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5760 struct lb_env env = {
5762 .dst_cpu = target_cpu,
5763 .dst_rq = target_rq,
5764 .src_cpu = busiest_rq->cpu,
5765 .src_rq = busiest_rq,
5769 schedstat_inc(sd, alb_count);
5771 if (move_one_task(&env))
5772 schedstat_inc(sd, alb_pushed);
5774 schedstat_inc(sd, alb_failed);
5777 double_unlock_balance(busiest_rq, target_rq);
5779 busiest_rq->active_balance = 0;
5780 raw_spin_unlock_irq(&busiest_rq->lock);
5784 #ifdef CONFIG_NO_HZ_COMMON
5786 * idle load balancing details
5787 * - When one of the busy CPUs notice that there may be an idle rebalancing
5788 * needed, they will kick the idle load balancer, which then does idle
5789 * load balancing for all the idle CPUs.
5792 cpumask_var_t idle_cpus_mask;
5794 unsigned long next_balance; /* in jiffy units */
5795 } nohz ____cacheline_aligned;
5797 static inline int find_new_ilb(int call_cpu)
5799 int ilb = cpumask_first(nohz.idle_cpus_mask);
5801 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5808 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5809 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5810 * CPU (if there is one).
5812 static void nohz_balancer_kick(int cpu)
5816 nohz.next_balance++;
5818 ilb_cpu = find_new_ilb(cpu);
5820 if (ilb_cpu >= nr_cpu_ids)
5823 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5826 * Use smp_send_reschedule() instead of resched_cpu().
5827 * This way we generate a sched IPI on the target cpu which
5828 * is idle. And the softirq performing nohz idle load balance
5829 * will be run before returning from the IPI.
5831 smp_send_reschedule(ilb_cpu);
5835 static inline void nohz_balance_exit_idle(int cpu)
5837 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5838 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5839 atomic_dec(&nohz.nr_cpus);
5840 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5844 static inline void set_cpu_sd_state_busy(void)
5846 struct sched_domain *sd;
5849 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5851 if (!sd || !sd->nohz_idle)
5855 for (; sd; sd = sd->parent)
5856 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5861 void set_cpu_sd_state_idle(void)
5863 struct sched_domain *sd;
5866 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5868 if (!sd || sd->nohz_idle)
5872 for (; sd; sd = sd->parent)
5873 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5879 * This routine will record that the cpu is going idle with tick stopped.
5880 * This info will be used in performing idle load balancing in the future.
5882 void nohz_balance_enter_idle(int cpu)
5885 * If this cpu is going down, then nothing needs to be done.
5887 if (!cpu_active(cpu))
5890 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5893 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5894 atomic_inc(&nohz.nr_cpus);
5895 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5898 static int sched_ilb_notifier(struct notifier_block *nfb,
5899 unsigned long action, void *hcpu)
5901 switch (action & ~CPU_TASKS_FROZEN) {
5903 nohz_balance_exit_idle(smp_processor_id());
5911 static DEFINE_SPINLOCK(balancing);
5914 * Scale the max load_balance interval with the number of CPUs in the system.
5915 * This trades load-balance latency on larger machines for less cross talk.
5917 void update_max_interval(void)
5919 max_load_balance_interval = HZ*num_online_cpus()/10;
5923 * It checks each scheduling domain to see if it is due to be balanced,
5924 * and initiates a balancing operation if so.
5926 * Balancing parameters are set up in init_sched_domains.
5928 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5930 int continue_balancing = 1;
5931 struct rq *rq = cpu_rq(cpu);
5932 unsigned long interval;
5933 struct sched_domain *sd;
5934 /* Earliest time when we have to do rebalance again */
5935 unsigned long next_balance = jiffies + 60*HZ;
5936 int update_next_balance = 0;
5937 int need_serialize, need_decay = 0;
5940 update_blocked_averages(cpu);
5943 for_each_domain(cpu, sd) {
5945 * Decay the newidle max times here because this is a regular
5946 * visit to all the domains. Decay ~1% per second.
5948 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5949 sd->max_newidle_lb_cost =
5950 (sd->max_newidle_lb_cost * 253) / 256;
5951 sd->next_decay_max_lb_cost = jiffies + HZ;
5954 max_cost += sd->max_newidle_lb_cost;
5956 if (!(sd->flags & SD_LOAD_BALANCE))
5960 * Stop the load balance at this level. There is another
5961 * CPU in our sched group which is doing load balancing more
5964 if (!continue_balancing) {
5970 interval = sd->balance_interval;
5971 if (idle != CPU_IDLE)
5972 interval *= sd->busy_factor;
5974 /* scale ms to jiffies */
5975 interval = msecs_to_jiffies(interval);
5976 interval = clamp(interval, 1UL, max_load_balance_interval);
5978 need_serialize = sd->flags & SD_SERIALIZE;
5980 if (need_serialize) {
5981 if (!spin_trylock(&balancing))
5985 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5986 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5988 * The LBF_DST_PINNED logic could have changed
5989 * env->dst_cpu, so we can't know our idle
5990 * state even if we migrated tasks. Update it.
5992 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5994 sd->last_balance = jiffies;
5997 spin_unlock(&balancing);
5999 if (time_after(next_balance, sd->last_balance + interval)) {
6000 next_balance = sd->last_balance + interval;
6001 update_next_balance = 1;
6006 * Ensure the rq-wide value also decays but keep it at a
6007 * reasonable floor to avoid funnies with rq->avg_idle.
6009 rq->max_idle_balance_cost =
6010 max((u64)sysctl_sched_migration_cost, max_cost);
6015 * next_balance will be updated only when there is a need.
6016 * When the cpu is attached to null domain for ex, it will not be
6019 if (likely(update_next_balance))
6020 rq->next_balance = next_balance;
6023 #ifdef CONFIG_NO_HZ_COMMON
6025 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6026 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6028 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6030 struct rq *this_rq = cpu_rq(this_cpu);
6034 if (idle != CPU_IDLE ||
6035 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6038 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6039 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6043 * If this cpu gets work to do, stop the load balancing
6044 * work being done for other cpus. Next load
6045 * balancing owner will pick it up.
6050 rq = cpu_rq(balance_cpu);
6052 raw_spin_lock_irq(&rq->lock);
6053 update_rq_clock(rq);
6054 update_idle_cpu_load(rq);
6055 raw_spin_unlock_irq(&rq->lock);
6057 rebalance_domains(balance_cpu, CPU_IDLE);
6059 if (time_after(this_rq->next_balance, rq->next_balance))
6060 this_rq->next_balance = rq->next_balance;
6062 nohz.next_balance = this_rq->next_balance;
6064 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6068 * Current heuristic for kicking the idle load balancer in the presence
6069 * of an idle cpu is the system.
6070 * - This rq has more than one task.
6071 * - At any scheduler domain level, this cpu's scheduler group has multiple
6072 * busy cpu's exceeding the group's power.
6073 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6074 * domain span are idle.
6076 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6078 unsigned long now = jiffies;
6079 struct sched_domain *sd;
6081 if (unlikely(idle_cpu(cpu)))
6085 * We may be recently in ticked or tickless idle mode. At the first
6086 * busy tick after returning from idle, we will update the busy stats.
6088 set_cpu_sd_state_busy();
6089 nohz_balance_exit_idle(cpu);
6092 * None are in tickless mode and hence no need for NOHZ idle load
6095 if (likely(!atomic_read(&nohz.nr_cpus)))
6098 if (time_before(now, nohz.next_balance))
6101 if (rq->nr_running >= 2)
6105 for_each_domain(cpu, sd) {
6106 struct sched_group *sg = sd->groups;
6107 struct sched_group_power *sgp = sg->sgp;
6108 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6110 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6111 goto need_kick_unlock;
6113 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6114 && (cpumask_first_and(nohz.idle_cpus_mask,
6115 sched_domain_span(sd)) < cpu))
6116 goto need_kick_unlock;
6118 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6130 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6134 * run_rebalance_domains is triggered when needed from the scheduler tick.
6135 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6137 static void run_rebalance_domains(struct softirq_action *h)
6139 int this_cpu = smp_processor_id();
6140 struct rq *this_rq = cpu_rq(this_cpu);
6141 enum cpu_idle_type idle = this_rq->idle_balance ?
6142 CPU_IDLE : CPU_NOT_IDLE;
6144 rebalance_domains(this_cpu, idle);
6147 * If this cpu has a pending nohz_balance_kick, then do the
6148 * balancing on behalf of the other idle cpus whose ticks are
6151 nohz_idle_balance(this_cpu, idle);
6154 static inline int on_null_domain(int cpu)
6156 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6160 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6162 void trigger_load_balance(struct rq *rq, int cpu)
6164 /* Don't need to rebalance while attached to NULL domain */
6165 if (time_after_eq(jiffies, rq->next_balance) &&
6166 likely(!on_null_domain(cpu)))
6167 raise_softirq(SCHED_SOFTIRQ);
6168 #ifdef CONFIG_NO_HZ_COMMON
6169 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6170 nohz_balancer_kick(cpu);
6174 static void rq_online_fair(struct rq *rq)
6179 static void rq_offline_fair(struct rq *rq)
6183 /* Ensure any throttled groups are reachable by pick_next_task */
6184 unthrottle_offline_cfs_rqs(rq);
6187 #endif /* CONFIG_SMP */
6190 * scheduler tick hitting a task of our scheduling class:
6192 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6194 struct cfs_rq *cfs_rq;
6195 struct sched_entity *se = &curr->se;
6197 for_each_sched_entity(se) {
6198 cfs_rq = cfs_rq_of(se);
6199 entity_tick(cfs_rq, se, queued);
6202 if (numabalancing_enabled)
6203 task_tick_numa(rq, curr);
6205 update_rq_runnable_avg(rq, 1);
6209 * called on fork with the child task as argument from the parent's context
6210 * - child not yet on the tasklist
6211 * - preemption disabled
6213 static void task_fork_fair(struct task_struct *p)
6215 struct cfs_rq *cfs_rq;
6216 struct sched_entity *se = &p->se, *curr;
6217 int this_cpu = smp_processor_id();
6218 struct rq *rq = this_rq();
6219 unsigned long flags;
6221 raw_spin_lock_irqsave(&rq->lock, flags);
6223 update_rq_clock(rq);
6225 cfs_rq = task_cfs_rq(current);
6226 curr = cfs_rq->curr;
6229 * Not only the cpu but also the task_group of the parent might have
6230 * been changed after parent->se.parent,cfs_rq were copied to
6231 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6232 * of child point to valid ones.
6235 __set_task_cpu(p, this_cpu);
6238 update_curr(cfs_rq);
6241 se->vruntime = curr->vruntime;
6242 place_entity(cfs_rq, se, 1);
6244 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6246 * Upon rescheduling, sched_class::put_prev_task() will place
6247 * 'current' within the tree based on its new key value.
6249 swap(curr->vruntime, se->vruntime);
6250 resched_task(rq->curr);
6253 se->vruntime -= cfs_rq->min_vruntime;
6255 raw_spin_unlock_irqrestore(&rq->lock, flags);
6259 * Priority of the task has changed. Check to see if we preempt
6263 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6269 * Reschedule if we are currently running on this runqueue and
6270 * our priority decreased, or if we are not currently running on
6271 * this runqueue and our priority is higher than the current's
6273 if (rq->curr == p) {
6274 if (p->prio > oldprio)
6275 resched_task(rq->curr);
6277 check_preempt_curr(rq, p, 0);
6280 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6282 struct sched_entity *se = &p->se;
6283 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6286 * Ensure the task's vruntime is normalized, so that when its
6287 * switched back to the fair class the enqueue_entity(.flags=0) will
6288 * do the right thing.
6290 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6291 * have normalized the vruntime, if it was !on_rq, then only when
6292 * the task is sleeping will it still have non-normalized vruntime.
6294 if (!se->on_rq && p->state != TASK_RUNNING) {
6296 * Fix up our vruntime so that the current sleep doesn't
6297 * cause 'unlimited' sleep bonus.
6299 place_entity(cfs_rq, se, 0);
6300 se->vruntime -= cfs_rq->min_vruntime;
6305 * Remove our load from contribution when we leave sched_fair
6306 * and ensure we don't carry in an old decay_count if we
6309 if (se->avg.decay_count) {
6310 __synchronize_entity_decay(se);
6311 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6317 * We switched to the sched_fair class.
6319 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6325 * We were most likely switched from sched_rt, so
6326 * kick off the schedule if running, otherwise just see
6327 * if we can still preempt the current task.
6330 resched_task(rq->curr);
6332 check_preempt_curr(rq, p, 0);
6335 /* Account for a task changing its policy or group.
6337 * This routine is mostly called to set cfs_rq->curr field when a task
6338 * migrates between groups/classes.
6340 static void set_curr_task_fair(struct rq *rq)
6342 struct sched_entity *se = &rq->curr->se;
6344 for_each_sched_entity(se) {
6345 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6347 set_next_entity(cfs_rq, se);
6348 /* ensure bandwidth has been allocated on our new cfs_rq */
6349 account_cfs_rq_runtime(cfs_rq, 0);
6353 void init_cfs_rq(struct cfs_rq *cfs_rq)
6355 cfs_rq->tasks_timeline = RB_ROOT;
6356 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6357 #ifndef CONFIG_64BIT
6358 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6361 atomic64_set(&cfs_rq->decay_counter, 1);
6362 atomic_long_set(&cfs_rq->removed_load, 0);
6366 #ifdef CONFIG_FAIR_GROUP_SCHED
6367 static void task_move_group_fair(struct task_struct *p, int on_rq)
6369 struct cfs_rq *cfs_rq;
6371 * If the task was not on the rq at the time of this cgroup movement
6372 * it must have been asleep, sleeping tasks keep their ->vruntime
6373 * absolute on their old rq until wakeup (needed for the fair sleeper
6374 * bonus in place_entity()).
6376 * If it was on the rq, we've just 'preempted' it, which does convert
6377 * ->vruntime to a relative base.
6379 * Make sure both cases convert their relative position when migrating
6380 * to another cgroup's rq. This does somewhat interfere with the
6381 * fair sleeper stuff for the first placement, but who cares.
6384 * When !on_rq, vruntime of the task has usually NOT been normalized.
6385 * But there are some cases where it has already been normalized:
6387 * - Moving a forked child which is waiting for being woken up by
6388 * wake_up_new_task().
6389 * - Moving a task which has been woken up by try_to_wake_up() and
6390 * waiting for actually being woken up by sched_ttwu_pending().
6392 * To prevent boost or penalty in the new cfs_rq caused by delta
6393 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6395 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6399 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6400 set_task_rq(p, task_cpu(p));
6402 cfs_rq = cfs_rq_of(&p->se);
6403 p->se.vruntime += cfs_rq->min_vruntime;
6406 * migrate_task_rq_fair() will have removed our previous
6407 * contribution, but we must synchronize for ongoing future
6410 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6411 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6416 void free_fair_sched_group(struct task_group *tg)
6420 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6422 for_each_possible_cpu(i) {
6424 kfree(tg->cfs_rq[i]);
6433 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6435 struct cfs_rq *cfs_rq;
6436 struct sched_entity *se;
6439 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6442 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6446 tg->shares = NICE_0_LOAD;
6448 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6450 for_each_possible_cpu(i) {
6451 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6452 GFP_KERNEL, cpu_to_node(i));
6456 se = kzalloc_node(sizeof(struct sched_entity),
6457 GFP_KERNEL, cpu_to_node(i));
6461 init_cfs_rq(cfs_rq);
6462 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6473 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6475 struct rq *rq = cpu_rq(cpu);
6476 unsigned long flags;
6479 * Only empty task groups can be destroyed; so we can speculatively
6480 * check on_list without danger of it being re-added.
6482 if (!tg->cfs_rq[cpu]->on_list)
6485 raw_spin_lock_irqsave(&rq->lock, flags);
6486 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6487 raw_spin_unlock_irqrestore(&rq->lock, flags);
6490 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6491 struct sched_entity *se, int cpu,
6492 struct sched_entity *parent)
6494 struct rq *rq = cpu_rq(cpu);
6498 init_cfs_rq_runtime(cfs_rq);
6500 tg->cfs_rq[cpu] = cfs_rq;
6503 /* se could be NULL for root_task_group */
6508 se->cfs_rq = &rq->cfs;
6510 se->cfs_rq = parent->my_q;
6513 update_load_set(&se->load, 0);
6514 se->parent = parent;
6517 static DEFINE_MUTEX(shares_mutex);
6519 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6522 unsigned long flags;
6525 * We can't change the weight of the root cgroup.
6530 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6532 mutex_lock(&shares_mutex);
6533 if (tg->shares == shares)
6536 tg->shares = shares;
6537 for_each_possible_cpu(i) {
6538 struct rq *rq = cpu_rq(i);
6539 struct sched_entity *se;
6542 /* Propagate contribution to hierarchy */
6543 raw_spin_lock_irqsave(&rq->lock, flags);
6545 /* Possible calls to update_curr() need rq clock */
6546 update_rq_clock(rq);
6547 for_each_sched_entity(se)
6548 update_cfs_shares(group_cfs_rq(se));
6549 raw_spin_unlock_irqrestore(&rq->lock, flags);
6553 mutex_unlock(&shares_mutex);
6556 #else /* CONFIG_FAIR_GROUP_SCHED */
6558 void free_fair_sched_group(struct task_group *tg) { }
6560 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6565 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6567 #endif /* CONFIG_FAIR_GROUP_SCHED */
6570 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6572 struct sched_entity *se = &task->se;
6573 unsigned int rr_interval = 0;
6576 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6579 if (rq->cfs.load.weight)
6580 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6586 * All the scheduling class methods:
6588 const struct sched_class fair_sched_class = {
6589 .next = &idle_sched_class,
6590 .enqueue_task = enqueue_task_fair,
6591 .dequeue_task = dequeue_task_fair,
6592 .yield_task = yield_task_fair,
6593 .yield_to_task = yield_to_task_fair,
6595 .check_preempt_curr = check_preempt_wakeup,
6597 .pick_next_task = pick_next_task_fair,
6598 .put_prev_task = put_prev_task_fair,
6601 .select_task_rq = select_task_rq_fair,
6602 .migrate_task_rq = migrate_task_rq_fair,
6604 .rq_online = rq_online_fair,
6605 .rq_offline = rq_offline_fair,
6607 .task_waking = task_waking_fair,
6610 .set_curr_task = set_curr_task_fair,
6611 .task_tick = task_tick_fair,
6612 .task_fork = task_fork_fair,
6614 .prio_changed = prio_changed_fair,
6615 .switched_from = switched_from_fair,
6616 .switched_to = switched_to_fair,
6618 .get_rr_interval = get_rr_interval_fair,
6620 #ifdef CONFIG_FAIR_GROUP_SCHED
6621 .task_move_group = task_move_group_fair,
6625 #ifdef CONFIG_SCHED_DEBUG
6626 void print_cfs_stats(struct seq_file *m, int cpu)
6628 struct cfs_rq *cfs_rq;
6631 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6632 print_cfs_rq(m, cpu, cfs_rq);
6637 __init void init_sched_fair_class(void)
6640 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6642 #ifdef CONFIG_NO_HZ_COMMON
6643 nohz.next_balance = jiffies;
6644 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6645 cpu_notifier(sched_ilb_notifier, 0);