2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * Approximate time to scan a full NUMA task in ms. The task scan period is
822 * calculated based on the tasks virtual memory size and
823 * numa_balancing_scan_size.
825 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
826 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
829 /* Portion of address space to scan in MB */
830 unsigned int sysctl_numa_balancing_scan_size = 256;
832 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
833 unsigned int sysctl_numa_balancing_scan_delay = 1000;
835 static unsigned int task_nr_scan_windows(struct task_struct *p)
837 unsigned long rss = 0;
838 unsigned long nr_scan_pages;
841 * Calculations based on RSS as non-present and empty pages are skipped
842 * by the PTE scanner and NUMA hinting faults should be trapped based
845 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
846 rss = get_mm_rss(p->mm);
850 rss = round_up(rss, nr_scan_pages);
851 return rss / nr_scan_pages;
854 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
855 #define MAX_SCAN_WINDOW 2560
857 static unsigned int task_scan_min(struct task_struct *p)
859 unsigned int scan, floor;
860 unsigned int windows = 1;
862 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
863 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
864 floor = 1000 / windows;
866 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
867 return max_t(unsigned int, floor, scan);
870 static unsigned int task_scan_max(struct task_struct *p)
872 unsigned int smin = task_scan_min(p);
875 /* Watch for min being lower than max due to floor calculations */
876 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
877 return max(smin, smax);
881 * Once a preferred node is selected the scheduler balancer will prefer moving
882 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
883 * scans. This will give the process the chance to accumulate more faults on
884 * the preferred node but still allow the scheduler to move the task again if
885 * the nodes CPUs are overloaded.
887 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
889 static inline int task_faults_idx(int nid, int priv)
891 return 2 * nid + priv;
894 static inline unsigned long task_faults(struct task_struct *p, int nid)
899 return p->numa_faults[task_faults_idx(nid, 0)] +
900 p->numa_faults[task_faults_idx(nid, 1)];
903 static unsigned long weighted_cpuload(const int cpu);
904 static unsigned long source_load(int cpu, int type);
905 static unsigned long target_load(int cpu, int type);
906 static unsigned long power_of(int cpu);
907 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
912 unsigned long faults;
915 struct task_numa_env {
916 struct task_struct *p;
918 int src_cpu, src_nid;
919 int dst_cpu, dst_nid;
921 struct numa_stats src_stats, dst_stats;
923 unsigned long best_load;
927 static int task_numa_migrate(struct task_struct *p)
929 int node_cpu = cpumask_first(cpumask_of_node(p->numa_preferred_nid));
930 struct task_numa_env env = {
932 .src_cpu = task_cpu(p),
933 .src_nid = cpu_to_node(task_cpu(p)),
935 .dst_nid = p->numa_preferred_nid,
936 .best_load = ULONG_MAX,
937 .best_cpu = task_cpu(p),
939 struct sched_domain *sd;
941 struct task_group *tg = task_group(p);
942 unsigned long weight;
944 int imbalance_pct, idx = -1;
947 * Find the lowest common scheduling domain covering the nodes of both
948 * the CPU the task is currently running on and the target NUMA node.
951 for_each_domain(env.src_cpu, sd) {
952 if (cpumask_test_cpu(node_cpu, sched_domain_span(sd))) {
954 * busy_idx is used for the load decision as it is the
955 * same index used by the regular load balancer for an
959 imbalance_pct = sd->imbalance_pct;
965 if (WARN_ON_ONCE(idx == -1))
969 * XXX the below is mostly nicked from wake_affine(); we should
970 * see about sharing a bit if at all possible; also it might want
971 * some per entity weight love.
973 weight = p->se.load.weight;
974 env.src_stats.load = source_load(env.src_cpu, idx);
975 env.src_stats.eff_load = 100 + (imbalance_pct - 100) / 2;
976 env.src_stats.eff_load *= power_of(env.src_cpu);
977 env.src_stats.eff_load *= env.src_stats.load + effective_load(tg, env.src_cpu, -weight, -weight);
979 for_each_cpu(cpu, cpumask_of_node(env.dst_nid)) {
981 env.dst_stats.load = target_load(cpu, idx);
983 /* If the CPU is idle, use it */
984 if (!env.dst_stats.load) {
989 /* Otherwise check the target CPU load */
990 env.dst_stats.eff_load = 100;
991 env.dst_stats.eff_load *= power_of(cpu);
992 env.dst_stats.eff_load *= env.dst_stats.load + effective_load(tg, cpu, weight, weight);
995 * Destination is considered balanced if the destination CPU is
996 * less loaded than the source CPU. Unfortunately there is a
997 * risk that a task running on a lightly loaded CPU will not
998 * migrate to its preferred node due to load imbalances.
1000 balanced = (env.dst_stats.eff_load <= env.src_stats.eff_load);
1004 if (env.dst_stats.eff_load < env.best_load) {
1005 env.best_load = env.dst_stats.eff_load;
1011 return migrate_task_to(p, env.best_cpu);
1014 /* Attempt to migrate a task to a CPU on the preferred node. */
1015 static void numa_migrate_preferred(struct task_struct *p)
1017 /* Success if task is already running on preferred CPU */
1018 p->numa_migrate_retry = 0;
1019 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1021 * If migration is temporarily disabled due to a task migration
1022 * then re-enable it now as the task is running on its
1023 * preferred node and memory should migrate locally
1025 if (!p->numa_migrate_seq)
1026 p->numa_migrate_seq++;
1030 /* This task has no NUMA fault statistics yet */
1031 if (unlikely(p->numa_preferred_nid == -1))
1034 /* Otherwise, try migrate to a CPU on the preferred node */
1035 if (task_numa_migrate(p) != 0)
1036 p->numa_migrate_retry = jiffies + HZ*5;
1039 static void task_numa_placement(struct task_struct *p)
1041 int seq, nid, max_nid = -1;
1042 unsigned long max_faults = 0;
1044 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1045 if (p->numa_scan_seq == seq)
1047 p->numa_scan_seq = seq;
1048 p->numa_migrate_seq++;
1049 p->numa_scan_period_max = task_scan_max(p);
1051 /* Find the node with the highest number of faults */
1052 for_each_online_node(nid) {
1053 unsigned long faults;
1056 for (priv = 0; priv < 2; priv++) {
1057 i = task_faults_idx(nid, priv);
1059 /* Decay existing window, copy faults since last scan */
1060 p->numa_faults[i] >>= 1;
1061 p->numa_faults[i] += p->numa_faults_buffer[i];
1062 p->numa_faults_buffer[i] = 0;
1065 /* Find maximum private faults */
1066 faults = p->numa_faults[task_faults_idx(nid, 1)];
1067 if (faults > max_faults) {
1068 max_faults = faults;
1073 /* Preferred node as the node with the most faults */
1074 if (max_faults && max_nid != p->numa_preferred_nid) {
1075 /* Update the preferred nid and migrate task if possible */
1076 p->numa_preferred_nid = max_nid;
1077 p->numa_migrate_seq = 1;
1078 numa_migrate_preferred(p);
1083 * Got a PROT_NONE fault for a page on @node.
1085 void task_numa_fault(int last_nidpid, int node, int pages, bool migrated)
1087 struct task_struct *p = current;
1090 if (!numabalancing_enabled)
1093 /* for example, ksmd faulting in a user's mm */
1098 * First accesses are treated as private, otherwise consider accesses
1099 * to be private if the accessing pid has not changed
1101 if (!nidpid_pid_unset(last_nidpid))
1102 priv = ((p->pid & LAST__PID_MASK) == nidpid_to_pid(last_nidpid));
1106 /* Allocate buffer to track faults on a per-node basis */
1107 if (unlikely(!p->numa_faults)) {
1108 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1110 /* numa_faults and numa_faults_buffer share the allocation */
1111 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1112 if (!p->numa_faults)
1115 BUG_ON(p->numa_faults_buffer);
1116 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1120 * If pages are properly placed (did not migrate) then scan slower.
1121 * This is reset periodically in case of phase changes
1124 /* Initialise if necessary */
1125 if (!p->numa_scan_period_max)
1126 p->numa_scan_period_max = task_scan_max(p);
1128 p->numa_scan_period = min(p->numa_scan_period_max,
1129 p->numa_scan_period + 10);
1132 task_numa_placement(p);
1134 /* Retry task to preferred node migration if it previously failed */
1135 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1136 numa_migrate_preferred(p);
1138 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1141 static void reset_ptenuma_scan(struct task_struct *p)
1143 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1144 p->mm->numa_scan_offset = 0;
1148 * The expensive part of numa migration is done from task_work context.
1149 * Triggered from task_tick_numa().
1151 void task_numa_work(struct callback_head *work)
1153 unsigned long migrate, next_scan, now = jiffies;
1154 struct task_struct *p = current;
1155 struct mm_struct *mm = p->mm;
1156 struct vm_area_struct *vma;
1157 unsigned long start, end;
1158 unsigned long nr_pte_updates = 0;
1161 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1163 work->next = work; /* protect against double add */
1165 * Who cares about NUMA placement when they're dying.
1167 * NOTE: make sure not to dereference p->mm before this check,
1168 * exit_task_work() happens _after_ exit_mm() so we could be called
1169 * without p->mm even though we still had it when we enqueued this
1172 if (p->flags & PF_EXITING)
1175 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1176 mm->numa_next_scan = now +
1177 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1178 mm->numa_next_reset = now +
1179 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1183 * Reset the scan period if enough time has gone by. Objective is that
1184 * scanning will be reduced if pages are properly placed. As tasks
1185 * can enter different phases this needs to be re-examined. Lacking
1186 * proper tracking of reference behaviour, this blunt hammer is used.
1188 migrate = mm->numa_next_reset;
1189 if (time_after(now, migrate)) {
1190 p->numa_scan_period = task_scan_min(p);
1191 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1192 xchg(&mm->numa_next_reset, next_scan);
1196 * Enforce maximal scan/migration frequency..
1198 migrate = mm->numa_next_scan;
1199 if (time_before(now, migrate))
1202 if (p->numa_scan_period == 0) {
1203 p->numa_scan_period_max = task_scan_max(p);
1204 p->numa_scan_period = task_scan_min(p);
1207 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1208 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1212 * Delay this task enough that another task of this mm will likely win
1213 * the next time around.
1215 p->node_stamp += 2 * TICK_NSEC;
1217 start = mm->numa_scan_offset;
1218 pages = sysctl_numa_balancing_scan_size;
1219 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1223 down_read(&mm->mmap_sem);
1224 vma = find_vma(mm, start);
1226 reset_ptenuma_scan(p);
1230 for (; vma; vma = vma->vm_next) {
1231 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1235 start = max(start, vma->vm_start);
1236 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1237 end = min(end, vma->vm_end);
1238 nr_pte_updates += change_prot_numa(vma, start, end);
1241 * Scan sysctl_numa_balancing_scan_size but ensure that
1242 * at least one PTE is updated so that unused virtual
1243 * address space is quickly skipped.
1246 pages -= (end - start) >> PAGE_SHIFT;
1251 } while (end != vma->vm_end);
1256 * If the whole process was scanned without updates then no NUMA
1257 * hinting faults are being recorded and scan rate should be lower.
1259 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1260 p->numa_scan_period = min(p->numa_scan_period_max,
1261 p->numa_scan_period << 1);
1263 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1264 mm->numa_next_scan = next_scan;
1268 * It is possible to reach the end of the VMA list but the last few
1269 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1270 * would find the !migratable VMA on the next scan but not reset the
1271 * scanner to the start so check it now.
1274 mm->numa_scan_offset = start;
1276 reset_ptenuma_scan(p);
1277 up_read(&mm->mmap_sem);
1281 * Drive the periodic memory faults..
1283 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1285 struct callback_head *work = &curr->numa_work;
1289 * We don't care about NUMA placement if we don't have memory.
1291 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1295 * Using runtime rather than walltime has the dual advantage that
1296 * we (mostly) drive the selection from busy threads and that the
1297 * task needs to have done some actual work before we bother with
1300 now = curr->se.sum_exec_runtime;
1301 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1303 if (now - curr->node_stamp > period) {
1304 if (!curr->node_stamp)
1305 curr->numa_scan_period = task_scan_min(curr);
1306 curr->node_stamp += period;
1308 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1309 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1310 task_work_add(curr, work, true);
1315 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1318 #endif /* CONFIG_NUMA_BALANCING */
1321 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1323 update_load_add(&cfs_rq->load, se->load.weight);
1324 if (!parent_entity(se))
1325 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1327 if (entity_is_task(se))
1328 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1330 cfs_rq->nr_running++;
1334 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1336 update_load_sub(&cfs_rq->load, se->load.weight);
1337 if (!parent_entity(se))
1338 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1339 if (entity_is_task(se))
1340 list_del_init(&se->group_node);
1341 cfs_rq->nr_running--;
1344 #ifdef CONFIG_FAIR_GROUP_SCHED
1346 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1351 * Use this CPU's actual weight instead of the last load_contribution
1352 * to gain a more accurate current total weight. See
1353 * update_cfs_rq_load_contribution().
1355 tg_weight = atomic_long_read(&tg->load_avg);
1356 tg_weight -= cfs_rq->tg_load_contrib;
1357 tg_weight += cfs_rq->load.weight;
1362 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1364 long tg_weight, load, shares;
1366 tg_weight = calc_tg_weight(tg, cfs_rq);
1367 load = cfs_rq->load.weight;
1369 shares = (tg->shares * load);
1371 shares /= tg_weight;
1373 if (shares < MIN_SHARES)
1374 shares = MIN_SHARES;
1375 if (shares > tg->shares)
1376 shares = tg->shares;
1380 # else /* CONFIG_SMP */
1381 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1385 # endif /* CONFIG_SMP */
1386 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1387 unsigned long weight)
1390 /* commit outstanding execution time */
1391 if (cfs_rq->curr == se)
1392 update_curr(cfs_rq);
1393 account_entity_dequeue(cfs_rq, se);
1396 update_load_set(&se->load, weight);
1399 account_entity_enqueue(cfs_rq, se);
1402 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1404 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1406 struct task_group *tg;
1407 struct sched_entity *se;
1411 se = tg->se[cpu_of(rq_of(cfs_rq))];
1412 if (!se || throttled_hierarchy(cfs_rq))
1415 if (likely(se->load.weight == tg->shares))
1418 shares = calc_cfs_shares(cfs_rq, tg);
1420 reweight_entity(cfs_rq_of(se), se, shares);
1422 #else /* CONFIG_FAIR_GROUP_SCHED */
1423 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1426 #endif /* CONFIG_FAIR_GROUP_SCHED */
1430 * We choose a half-life close to 1 scheduling period.
1431 * Note: The tables below are dependent on this value.
1433 #define LOAD_AVG_PERIOD 32
1434 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1435 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1437 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1438 static const u32 runnable_avg_yN_inv[] = {
1439 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1440 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1441 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1442 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1443 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1444 0x85aac367, 0x82cd8698,
1448 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1449 * over-estimates when re-combining.
1451 static const u32 runnable_avg_yN_sum[] = {
1452 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1453 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1454 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1459 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1461 static __always_inline u64 decay_load(u64 val, u64 n)
1463 unsigned int local_n;
1467 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1470 /* after bounds checking we can collapse to 32-bit */
1474 * As y^PERIOD = 1/2, we can combine
1475 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1476 * With a look-up table which covers k^n (n<PERIOD)
1478 * To achieve constant time decay_load.
1480 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1481 val >>= local_n / LOAD_AVG_PERIOD;
1482 local_n %= LOAD_AVG_PERIOD;
1485 val *= runnable_avg_yN_inv[local_n];
1486 /* We don't use SRR here since we always want to round down. */
1491 * For updates fully spanning n periods, the contribution to runnable
1492 * average will be: \Sum 1024*y^n
1494 * We can compute this reasonably efficiently by combining:
1495 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1497 static u32 __compute_runnable_contrib(u64 n)
1501 if (likely(n <= LOAD_AVG_PERIOD))
1502 return runnable_avg_yN_sum[n];
1503 else if (unlikely(n >= LOAD_AVG_MAX_N))
1504 return LOAD_AVG_MAX;
1506 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1508 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1509 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1511 n -= LOAD_AVG_PERIOD;
1512 } while (n > LOAD_AVG_PERIOD);
1514 contrib = decay_load(contrib, n);
1515 return contrib + runnable_avg_yN_sum[n];
1519 * We can represent the historical contribution to runnable average as the
1520 * coefficients of a geometric series. To do this we sub-divide our runnable
1521 * history into segments of approximately 1ms (1024us); label the segment that
1522 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1524 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1526 * (now) (~1ms ago) (~2ms ago)
1528 * Let u_i denote the fraction of p_i that the entity was runnable.
1530 * We then designate the fractions u_i as our co-efficients, yielding the
1531 * following representation of historical load:
1532 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1534 * We choose y based on the with of a reasonably scheduling period, fixing:
1537 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1538 * approximately half as much as the contribution to load within the last ms
1541 * When a period "rolls over" and we have new u_0`, multiplying the previous
1542 * sum again by y is sufficient to update:
1543 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1544 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1546 static __always_inline int __update_entity_runnable_avg(u64 now,
1547 struct sched_avg *sa,
1551 u32 runnable_contrib;
1552 int delta_w, decayed = 0;
1554 delta = now - sa->last_runnable_update;
1556 * This should only happen when time goes backwards, which it
1557 * unfortunately does during sched clock init when we swap over to TSC.
1559 if ((s64)delta < 0) {
1560 sa->last_runnable_update = now;
1565 * Use 1024ns as the unit of measurement since it's a reasonable
1566 * approximation of 1us and fast to compute.
1571 sa->last_runnable_update = now;
1573 /* delta_w is the amount already accumulated against our next period */
1574 delta_w = sa->runnable_avg_period % 1024;
1575 if (delta + delta_w >= 1024) {
1576 /* period roll-over */
1580 * Now that we know we're crossing a period boundary, figure
1581 * out how much from delta we need to complete the current
1582 * period and accrue it.
1584 delta_w = 1024 - delta_w;
1586 sa->runnable_avg_sum += delta_w;
1587 sa->runnable_avg_period += delta_w;
1591 /* Figure out how many additional periods this update spans */
1592 periods = delta / 1024;
1595 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1597 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1600 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1601 runnable_contrib = __compute_runnable_contrib(periods);
1603 sa->runnable_avg_sum += runnable_contrib;
1604 sa->runnable_avg_period += runnable_contrib;
1607 /* Remainder of delta accrued against u_0` */
1609 sa->runnable_avg_sum += delta;
1610 sa->runnable_avg_period += delta;
1615 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1616 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1618 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1619 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1621 decays -= se->avg.decay_count;
1625 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1626 se->avg.decay_count = 0;
1631 #ifdef CONFIG_FAIR_GROUP_SCHED
1632 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1635 struct task_group *tg = cfs_rq->tg;
1638 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1639 tg_contrib -= cfs_rq->tg_load_contrib;
1641 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1642 atomic_long_add(tg_contrib, &tg->load_avg);
1643 cfs_rq->tg_load_contrib += tg_contrib;
1648 * Aggregate cfs_rq runnable averages into an equivalent task_group
1649 * representation for computing load contributions.
1651 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1652 struct cfs_rq *cfs_rq)
1654 struct task_group *tg = cfs_rq->tg;
1657 /* The fraction of a cpu used by this cfs_rq */
1658 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1659 sa->runnable_avg_period + 1);
1660 contrib -= cfs_rq->tg_runnable_contrib;
1662 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1663 atomic_add(contrib, &tg->runnable_avg);
1664 cfs_rq->tg_runnable_contrib += contrib;
1668 static inline void __update_group_entity_contrib(struct sched_entity *se)
1670 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1671 struct task_group *tg = cfs_rq->tg;
1676 contrib = cfs_rq->tg_load_contrib * tg->shares;
1677 se->avg.load_avg_contrib = div_u64(contrib,
1678 atomic_long_read(&tg->load_avg) + 1);
1681 * For group entities we need to compute a correction term in the case
1682 * that they are consuming <1 cpu so that we would contribute the same
1683 * load as a task of equal weight.
1685 * Explicitly co-ordinating this measurement would be expensive, but
1686 * fortunately the sum of each cpus contribution forms a usable
1687 * lower-bound on the true value.
1689 * Consider the aggregate of 2 contributions. Either they are disjoint
1690 * (and the sum represents true value) or they are disjoint and we are
1691 * understating by the aggregate of their overlap.
1693 * Extending this to N cpus, for a given overlap, the maximum amount we
1694 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1695 * cpus that overlap for this interval and w_i is the interval width.
1697 * On a small machine; the first term is well-bounded which bounds the
1698 * total error since w_i is a subset of the period. Whereas on a
1699 * larger machine, while this first term can be larger, if w_i is the
1700 * of consequential size guaranteed to see n_i*w_i quickly converge to
1701 * our upper bound of 1-cpu.
1703 runnable_avg = atomic_read(&tg->runnable_avg);
1704 if (runnable_avg < NICE_0_LOAD) {
1705 se->avg.load_avg_contrib *= runnable_avg;
1706 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1710 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1711 int force_update) {}
1712 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1713 struct cfs_rq *cfs_rq) {}
1714 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1717 static inline void __update_task_entity_contrib(struct sched_entity *se)
1721 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1722 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1723 contrib /= (se->avg.runnable_avg_period + 1);
1724 se->avg.load_avg_contrib = scale_load(contrib);
1727 /* Compute the current contribution to load_avg by se, return any delta */
1728 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1730 long old_contrib = se->avg.load_avg_contrib;
1732 if (entity_is_task(se)) {
1733 __update_task_entity_contrib(se);
1735 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1736 __update_group_entity_contrib(se);
1739 return se->avg.load_avg_contrib - old_contrib;
1742 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1745 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1746 cfs_rq->blocked_load_avg -= load_contrib;
1748 cfs_rq->blocked_load_avg = 0;
1751 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1753 /* Update a sched_entity's runnable average */
1754 static inline void update_entity_load_avg(struct sched_entity *se,
1757 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1762 * For a group entity we need to use their owned cfs_rq_clock_task() in
1763 * case they are the parent of a throttled hierarchy.
1765 if (entity_is_task(se))
1766 now = cfs_rq_clock_task(cfs_rq);
1768 now = cfs_rq_clock_task(group_cfs_rq(se));
1770 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1773 contrib_delta = __update_entity_load_avg_contrib(se);
1779 cfs_rq->runnable_load_avg += contrib_delta;
1781 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1785 * Decay the load contributed by all blocked children and account this so that
1786 * their contribution may appropriately discounted when they wake up.
1788 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1790 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1793 decays = now - cfs_rq->last_decay;
1794 if (!decays && !force_update)
1797 if (atomic_long_read(&cfs_rq->removed_load)) {
1798 unsigned long removed_load;
1799 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1800 subtract_blocked_load_contrib(cfs_rq, removed_load);
1804 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1806 atomic64_add(decays, &cfs_rq->decay_counter);
1807 cfs_rq->last_decay = now;
1810 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1813 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1815 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1816 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1819 /* Add the load generated by se into cfs_rq's child load-average */
1820 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1821 struct sched_entity *se,
1825 * We track migrations using entity decay_count <= 0, on a wake-up
1826 * migration we use a negative decay count to track the remote decays
1827 * accumulated while sleeping.
1829 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1830 * are seen by enqueue_entity_load_avg() as a migration with an already
1831 * constructed load_avg_contrib.
1833 if (unlikely(se->avg.decay_count <= 0)) {
1834 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1835 if (se->avg.decay_count) {
1837 * In a wake-up migration we have to approximate the
1838 * time sleeping. This is because we can't synchronize
1839 * clock_task between the two cpus, and it is not
1840 * guaranteed to be read-safe. Instead, we can
1841 * approximate this using our carried decays, which are
1842 * explicitly atomically readable.
1844 se->avg.last_runnable_update -= (-se->avg.decay_count)
1846 update_entity_load_avg(se, 0);
1847 /* Indicate that we're now synchronized and on-rq */
1848 se->avg.decay_count = 0;
1853 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1854 * would have made count negative); we must be careful to avoid
1855 * double-accounting blocked time after synchronizing decays.
1857 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1861 /* migrated tasks did not contribute to our blocked load */
1863 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1864 update_entity_load_avg(se, 0);
1867 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1868 /* we force update consideration on load-balancer moves */
1869 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1873 * Remove se's load from this cfs_rq child load-average, if the entity is
1874 * transitioning to a blocked state we track its projected decay using
1877 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1878 struct sched_entity *se,
1881 update_entity_load_avg(se, 1);
1882 /* we force update consideration on load-balancer moves */
1883 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1885 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1887 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1888 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1889 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1893 * Update the rq's load with the elapsed running time before entering
1894 * idle. if the last scheduled task is not a CFS task, idle_enter will
1895 * be the only way to update the runnable statistic.
1897 void idle_enter_fair(struct rq *this_rq)
1899 update_rq_runnable_avg(this_rq, 1);
1903 * Update the rq's load with the elapsed idle time before a task is
1904 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1905 * be the only way to update the runnable statistic.
1907 void idle_exit_fair(struct rq *this_rq)
1909 update_rq_runnable_avg(this_rq, 0);
1913 static inline void update_entity_load_avg(struct sched_entity *se,
1914 int update_cfs_rq) {}
1915 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1916 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1917 struct sched_entity *se,
1919 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1920 struct sched_entity *se,
1922 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1923 int force_update) {}
1926 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1928 #ifdef CONFIG_SCHEDSTATS
1929 struct task_struct *tsk = NULL;
1931 if (entity_is_task(se))
1934 if (se->statistics.sleep_start) {
1935 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1940 if (unlikely(delta > se->statistics.sleep_max))
1941 se->statistics.sleep_max = delta;
1943 se->statistics.sleep_start = 0;
1944 se->statistics.sum_sleep_runtime += delta;
1947 account_scheduler_latency(tsk, delta >> 10, 1);
1948 trace_sched_stat_sleep(tsk, delta);
1951 if (se->statistics.block_start) {
1952 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1957 if (unlikely(delta > se->statistics.block_max))
1958 se->statistics.block_max = delta;
1960 se->statistics.block_start = 0;
1961 se->statistics.sum_sleep_runtime += delta;
1964 if (tsk->in_iowait) {
1965 se->statistics.iowait_sum += delta;
1966 se->statistics.iowait_count++;
1967 trace_sched_stat_iowait(tsk, delta);
1970 trace_sched_stat_blocked(tsk, delta);
1973 * Blocking time is in units of nanosecs, so shift by
1974 * 20 to get a milliseconds-range estimation of the
1975 * amount of time that the task spent sleeping:
1977 if (unlikely(prof_on == SLEEP_PROFILING)) {
1978 profile_hits(SLEEP_PROFILING,
1979 (void *)get_wchan(tsk),
1982 account_scheduler_latency(tsk, delta >> 10, 0);
1988 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1990 #ifdef CONFIG_SCHED_DEBUG
1991 s64 d = se->vruntime - cfs_rq->min_vruntime;
1996 if (d > 3*sysctl_sched_latency)
1997 schedstat_inc(cfs_rq, nr_spread_over);
2002 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2004 u64 vruntime = cfs_rq->min_vruntime;
2007 * The 'current' period is already promised to the current tasks,
2008 * however the extra weight of the new task will slow them down a
2009 * little, place the new task so that it fits in the slot that
2010 * stays open at the end.
2012 if (initial && sched_feat(START_DEBIT))
2013 vruntime += sched_vslice(cfs_rq, se);
2015 /* sleeps up to a single latency don't count. */
2017 unsigned long thresh = sysctl_sched_latency;
2020 * Halve their sleep time's effect, to allow
2021 * for a gentler effect of sleepers:
2023 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2029 /* ensure we never gain time by being placed backwards. */
2030 se->vruntime = max_vruntime(se->vruntime, vruntime);
2033 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2036 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2039 * Update the normalized vruntime before updating min_vruntime
2040 * through calling update_curr().
2042 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2043 se->vruntime += cfs_rq->min_vruntime;
2046 * Update run-time statistics of the 'current'.
2048 update_curr(cfs_rq);
2049 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2050 account_entity_enqueue(cfs_rq, se);
2051 update_cfs_shares(cfs_rq);
2053 if (flags & ENQUEUE_WAKEUP) {
2054 place_entity(cfs_rq, se, 0);
2055 enqueue_sleeper(cfs_rq, se);
2058 update_stats_enqueue(cfs_rq, se);
2059 check_spread(cfs_rq, se);
2060 if (se != cfs_rq->curr)
2061 __enqueue_entity(cfs_rq, se);
2064 if (cfs_rq->nr_running == 1) {
2065 list_add_leaf_cfs_rq(cfs_rq);
2066 check_enqueue_throttle(cfs_rq);
2070 static void __clear_buddies_last(struct sched_entity *se)
2072 for_each_sched_entity(se) {
2073 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2074 if (cfs_rq->last == se)
2075 cfs_rq->last = NULL;
2081 static void __clear_buddies_next(struct sched_entity *se)
2083 for_each_sched_entity(se) {
2084 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2085 if (cfs_rq->next == se)
2086 cfs_rq->next = NULL;
2092 static void __clear_buddies_skip(struct sched_entity *se)
2094 for_each_sched_entity(se) {
2095 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2096 if (cfs_rq->skip == se)
2097 cfs_rq->skip = NULL;
2103 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2105 if (cfs_rq->last == se)
2106 __clear_buddies_last(se);
2108 if (cfs_rq->next == se)
2109 __clear_buddies_next(se);
2111 if (cfs_rq->skip == se)
2112 __clear_buddies_skip(se);
2115 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2118 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2121 * Update run-time statistics of the 'current'.
2123 update_curr(cfs_rq);
2124 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2126 update_stats_dequeue(cfs_rq, se);
2127 if (flags & DEQUEUE_SLEEP) {
2128 #ifdef CONFIG_SCHEDSTATS
2129 if (entity_is_task(se)) {
2130 struct task_struct *tsk = task_of(se);
2132 if (tsk->state & TASK_INTERRUPTIBLE)
2133 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2134 if (tsk->state & TASK_UNINTERRUPTIBLE)
2135 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2140 clear_buddies(cfs_rq, se);
2142 if (se != cfs_rq->curr)
2143 __dequeue_entity(cfs_rq, se);
2145 account_entity_dequeue(cfs_rq, se);
2148 * Normalize the entity after updating the min_vruntime because the
2149 * update can refer to the ->curr item and we need to reflect this
2150 * movement in our normalized position.
2152 if (!(flags & DEQUEUE_SLEEP))
2153 se->vruntime -= cfs_rq->min_vruntime;
2155 /* return excess runtime on last dequeue */
2156 return_cfs_rq_runtime(cfs_rq);
2158 update_min_vruntime(cfs_rq);
2159 update_cfs_shares(cfs_rq);
2163 * Preempt the current task with a newly woken task if needed:
2166 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2168 unsigned long ideal_runtime, delta_exec;
2169 struct sched_entity *se;
2172 ideal_runtime = sched_slice(cfs_rq, curr);
2173 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2174 if (delta_exec > ideal_runtime) {
2175 resched_task(rq_of(cfs_rq)->curr);
2177 * The current task ran long enough, ensure it doesn't get
2178 * re-elected due to buddy favours.
2180 clear_buddies(cfs_rq, curr);
2185 * Ensure that a task that missed wakeup preemption by a
2186 * narrow margin doesn't have to wait for a full slice.
2187 * This also mitigates buddy induced latencies under load.
2189 if (delta_exec < sysctl_sched_min_granularity)
2192 se = __pick_first_entity(cfs_rq);
2193 delta = curr->vruntime - se->vruntime;
2198 if (delta > ideal_runtime)
2199 resched_task(rq_of(cfs_rq)->curr);
2203 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2205 /* 'current' is not kept within the tree. */
2208 * Any task has to be enqueued before it get to execute on
2209 * a CPU. So account for the time it spent waiting on the
2212 update_stats_wait_end(cfs_rq, se);
2213 __dequeue_entity(cfs_rq, se);
2216 update_stats_curr_start(cfs_rq, se);
2218 #ifdef CONFIG_SCHEDSTATS
2220 * Track our maximum slice length, if the CPU's load is at
2221 * least twice that of our own weight (i.e. dont track it
2222 * when there are only lesser-weight tasks around):
2224 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2225 se->statistics.slice_max = max(se->statistics.slice_max,
2226 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2229 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2233 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2236 * Pick the next process, keeping these things in mind, in this order:
2237 * 1) keep things fair between processes/task groups
2238 * 2) pick the "next" process, since someone really wants that to run
2239 * 3) pick the "last" process, for cache locality
2240 * 4) do not run the "skip" process, if something else is available
2242 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2244 struct sched_entity *se = __pick_first_entity(cfs_rq);
2245 struct sched_entity *left = se;
2248 * Avoid running the skip buddy, if running something else can
2249 * be done without getting too unfair.
2251 if (cfs_rq->skip == se) {
2252 struct sched_entity *second = __pick_next_entity(se);
2253 if (second && wakeup_preempt_entity(second, left) < 1)
2258 * Prefer last buddy, try to return the CPU to a preempted task.
2260 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2264 * Someone really wants this to run. If it's not unfair, run it.
2266 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2269 clear_buddies(cfs_rq, se);
2274 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2276 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2279 * If still on the runqueue then deactivate_task()
2280 * was not called and update_curr() has to be done:
2283 update_curr(cfs_rq);
2285 /* throttle cfs_rqs exceeding runtime */
2286 check_cfs_rq_runtime(cfs_rq);
2288 check_spread(cfs_rq, prev);
2290 update_stats_wait_start(cfs_rq, prev);
2291 /* Put 'current' back into the tree. */
2292 __enqueue_entity(cfs_rq, prev);
2293 /* in !on_rq case, update occurred at dequeue */
2294 update_entity_load_avg(prev, 1);
2296 cfs_rq->curr = NULL;
2300 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2303 * Update run-time statistics of the 'current'.
2305 update_curr(cfs_rq);
2308 * Ensure that runnable average is periodically updated.
2310 update_entity_load_avg(curr, 1);
2311 update_cfs_rq_blocked_load(cfs_rq, 1);
2312 update_cfs_shares(cfs_rq);
2314 #ifdef CONFIG_SCHED_HRTICK
2316 * queued ticks are scheduled to match the slice, so don't bother
2317 * validating it and just reschedule.
2320 resched_task(rq_of(cfs_rq)->curr);
2324 * don't let the period tick interfere with the hrtick preemption
2326 if (!sched_feat(DOUBLE_TICK) &&
2327 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2331 if (cfs_rq->nr_running > 1)
2332 check_preempt_tick(cfs_rq, curr);
2336 /**************************************************
2337 * CFS bandwidth control machinery
2340 #ifdef CONFIG_CFS_BANDWIDTH
2342 #ifdef HAVE_JUMP_LABEL
2343 static struct static_key __cfs_bandwidth_used;
2345 static inline bool cfs_bandwidth_used(void)
2347 return static_key_false(&__cfs_bandwidth_used);
2350 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2352 /* only need to count groups transitioning between enabled/!enabled */
2353 if (enabled && !was_enabled)
2354 static_key_slow_inc(&__cfs_bandwidth_used);
2355 else if (!enabled && was_enabled)
2356 static_key_slow_dec(&__cfs_bandwidth_used);
2358 #else /* HAVE_JUMP_LABEL */
2359 static bool cfs_bandwidth_used(void)
2364 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2365 #endif /* HAVE_JUMP_LABEL */
2368 * default period for cfs group bandwidth.
2369 * default: 0.1s, units: nanoseconds
2371 static inline u64 default_cfs_period(void)
2373 return 100000000ULL;
2376 static inline u64 sched_cfs_bandwidth_slice(void)
2378 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2382 * Replenish runtime according to assigned quota and update expiration time.
2383 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2384 * additional synchronization around rq->lock.
2386 * requires cfs_b->lock
2388 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2392 if (cfs_b->quota == RUNTIME_INF)
2395 now = sched_clock_cpu(smp_processor_id());
2396 cfs_b->runtime = cfs_b->quota;
2397 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2400 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2402 return &tg->cfs_bandwidth;
2405 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2406 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2408 if (unlikely(cfs_rq->throttle_count))
2409 return cfs_rq->throttled_clock_task;
2411 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2414 /* returns 0 on failure to allocate runtime */
2415 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2417 struct task_group *tg = cfs_rq->tg;
2418 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2419 u64 amount = 0, min_amount, expires;
2421 /* note: this is a positive sum as runtime_remaining <= 0 */
2422 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2424 raw_spin_lock(&cfs_b->lock);
2425 if (cfs_b->quota == RUNTIME_INF)
2426 amount = min_amount;
2429 * If the bandwidth pool has become inactive, then at least one
2430 * period must have elapsed since the last consumption.
2431 * Refresh the global state and ensure bandwidth timer becomes
2434 if (!cfs_b->timer_active) {
2435 __refill_cfs_bandwidth_runtime(cfs_b);
2436 __start_cfs_bandwidth(cfs_b);
2439 if (cfs_b->runtime > 0) {
2440 amount = min(cfs_b->runtime, min_amount);
2441 cfs_b->runtime -= amount;
2445 expires = cfs_b->runtime_expires;
2446 raw_spin_unlock(&cfs_b->lock);
2448 cfs_rq->runtime_remaining += amount;
2450 * we may have advanced our local expiration to account for allowed
2451 * spread between our sched_clock and the one on which runtime was
2454 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2455 cfs_rq->runtime_expires = expires;
2457 return cfs_rq->runtime_remaining > 0;
2461 * Note: This depends on the synchronization provided by sched_clock and the
2462 * fact that rq->clock snapshots this value.
2464 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2466 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2468 /* if the deadline is ahead of our clock, nothing to do */
2469 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2472 if (cfs_rq->runtime_remaining < 0)
2476 * If the local deadline has passed we have to consider the
2477 * possibility that our sched_clock is 'fast' and the global deadline
2478 * has not truly expired.
2480 * Fortunately we can check determine whether this the case by checking
2481 * whether the global deadline has advanced.
2484 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2485 /* extend local deadline, drift is bounded above by 2 ticks */
2486 cfs_rq->runtime_expires += TICK_NSEC;
2488 /* global deadline is ahead, expiration has passed */
2489 cfs_rq->runtime_remaining = 0;
2493 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2494 unsigned long delta_exec)
2496 /* dock delta_exec before expiring quota (as it could span periods) */
2497 cfs_rq->runtime_remaining -= delta_exec;
2498 expire_cfs_rq_runtime(cfs_rq);
2500 if (likely(cfs_rq->runtime_remaining > 0))
2504 * if we're unable to extend our runtime we resched so that the active
2505 * hierarchy can be throttled
2507 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2508 resched_task(rq_of(cfs_rq)->curr);
2511 static __always_inline
2512 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2514 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2517 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2520 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2522 return cfs_bandwidth_used() && cfs_rq->throttled;
2525 /* check whether cfs_rq, or any parent, is throttled */
2526 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2528 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2532 * Ensure that neither of the group entities corresponding to src_cpu or
2533 * dest_cpu are members of a throttled hierarchy when performing group
2534 * load-balance operations.
2536 static inline int throttled_lb_pair(struct task_group *tg,
2537 int src_cpu, int dest_cpu)
2539 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2541 src_cfs_rq = tg->cfs_rq[src_cpu];
2542 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2544 return throttled_hierarchy(src_cfs_rq) ||
2545 throttled_hierarchy(dest_cfs_rq);
2548 /* updated child weight may affect parent so we have to do this bottom up */
2549 static int tg_unthrottle_up(struct task_group *tg, void *data)
2551 struct rq *rq = data;
2552 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2554 cfs_rq->throttle_count--;
2556 if (!cfs_rq->throttle_count) {
2557 /* adjust cfs_rq_clock_task() */
2558 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2559 cfs_rq->throttled_clock_task;
2566 static int tg_throttle_down(struct task_group *tg, void *data)
2568 struct rq *rq = data;
2569 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2571 /* group is entering throttled state, stop time */
2572 if (!cfs_rq->throttle_count)
2573 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2574 cfs_rq->throttle_count++;
2579 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2581 struct rq *rq = rq_of(cfs_rq);
2582 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2583 struct sched_entity *se;
2584 long task_delta, dequeue = 1;
2586 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2588 /* freeze hierarchy runnable averages while throttled */
2590 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2593 task_delta = cfs_rq->h_nr_running;
2594 for_each_sched_entity(se) {
2595 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2596 /* throttled entity or throttle-on-deactivate */
2601 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2602 qcfs_rq->h_nr_running -= task_delta;
2604 if (qcfs_rq->load.weight)
2609 rq->nr_running -= task_delta;
2611 cfs_rq->throttled = 1;
2612 cfs_rq->throttled_clock = rq_clock(rq);
2613 raw_spin_lock(&cfs_b->lock);
2614 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2615 raw_spin_unlock(&cfs_b->lock);
2618 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2620 struct rq *rq = rq_of(cfs_rq);
2621 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2622 struct sched_entity *se;
2626 se = cfs_rq->tg->se[cpu_of(rq)];
2628 cfs_rq->throttled = 0;
2630 update_rq_clock(rq);
2632 raw_spin_lock(&cfs_b->lock);
2633 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2634 list_del_rcu(&cfs_rq->throttled_list);
2635 raw_spin_unlock(&cfs_b->lock);
2637 /* update hierarchical throttle state */
2638 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2640 if (!cfs_rq->load.weight)
2643 task_delta = cfs_rq->h_nr_running;
2644 for_each_sched_entity(se) {
2648 cfs_rq = cfs_rq_of(se);
2650 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2651 cfs_rq->h_nr_running += task_delta;
2653 if (cfs_rq_throttled(cfs_rq))
2658 rq->nr_running += task_delta;
2660 /* determine whether we need to wake up potentially idle cpu */
2661 if (rq->curr == rq->idle && rq->cfs.nr_running)
2662 resched_task(rq->curr);
2665 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2666 u64 remaining, u64 expires)
2668 struct cfs_rq *cfs_rq;
2669 u64 runtime = remaining;
2672 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2674 struct rq *rq = rq_of(cfs_rq);
2676 raw_spin_lock(&rq->lock);
2677 if (!cfs_rq_throttled(cfs_rq))
2680 runtime = -cfs_rq->runtime_remaining + 1;
2681 if (runtime > remaining)
2682 runtime = remaining;
2683 remaining -= runtime;
2685 cfs_rq->runtime_remaining += runtime;
2686 cfs_rq->runtime_expires = expires;
2688 /* we check whether we're throttled above */
2689 if (cfs_rq->runtime_remaining > 0)
2690 unthrottle_cfs_rq(cfs_rq);
2693 raw_spin_unlock(&rq->lock);
2704 * Responsible for refilling a task_group's bandwidth and unthrottling its
2705 * cfs_rqs as appropriate. If there has been no activity within the last
2706 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2707 * used to track this state.
2709 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2711 u64 runtime, runtime_expires;
2712 int idle = 1, throttled;
2714 raw_spin_lock(&cfs_b->lock);
2715 /* no need to continue the timer with no bandwidth constraint */
2716 if (cfs_b->quota == RUNTIME_INF)
2719 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2720 /* idle depends on !throttled (for the case of a large deficit) */
2721 idle = cfs_b->idle && !throttled;
2722 cfs_b->nr_periods += overrun;
2724 /* if we're going inactive then everything else can be deferred */
2728 __refill_cfs_bandwidth_runtime(cfs_b);
2731 /* mark as potentially idle for the upcoming period */
2736 /* account preceding periods in which throttling occurred */
2737 cfs_b->nr_throttled += overrun;
2740 * There are throttled entities so we must first use the new bandwidth
2741 * to unthrottle them before making it generally available. This
2742 * ensures that all existing debts will be paid before a new cfs_rq is
2745 runtime = cfs_b->runtime;
2746 runtime_expires = cfs_b->runtime_expires;
2750 * This check is repeated as we are holding onto the new bandwidth
2751 * while we unthrottle. This can potentially race with an unthrottled
2752 * group trying to acquire new bandwidth from the global pool.
2754 while (throttled && runtime > 0) {
2755 raw_spin_unlock(&cfs_b->lock);
2756 /* we can't nest cfs_b->lock while distributing bandwidth */
2757 runtime = distribute_cfs_runtime(cfs_b, runtime,
2759 raw_spin_lock(&cfs_b->lock);
2761 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2764 /* return (any) remaining runtime */
2765 cfs_b->runtime = runtime;
2767 * While we are ensured activity in the period following an
2768 * unthrottle, this also covers the case in which the new bandwidth is
2769 * insufficient to cover the existing bandwidth deficit. (Forcing the
2770 * timer to remain active while there are any throttled entities.)
2775 cfs_b->timer_active = 0;
2776 raw_spin_unlock(&cfs_b->lock);
2781 /* a cfs_rq won't donate quota below this amount */
2782 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2783 /* minimum remaining period time to redistribute slack quota */
2784 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2785 /* how long we wait to gather additional slack before distributing */
2786 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2788 /* are we near the end of the current quota period? */
2789 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2791 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2794 /* if the call-back is running a quota refresh is already occurring */
2795 if (hrtimer_callback_running(refresh_timer))
2798 /* is a quota refresh about to occur? */
2799 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2800 if (remaining < min_expire)
2806 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2808 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2810 /* if there's a quota refresh soon don't bother with slack */
2811 if (runtime_refresh_within(cfs_b, min_left))
2814 start_bandwidth_timer(&cfs_b->slack_timer,
2815 ns_to_ktime(cfs_bandwidth_slack_period));
2818 /* we know any runtime found here is valid as update_curr() precedes return */
2819 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2821 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2822 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2824 if (slack_runtime <= 0)
2827 raw_spin_lock(&cfs_b->lock);
2828 if (cfs_b->quota != RUNTIME_INF &&
2829 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2830 cfs_b->runtime += slack_runtime;
2832 /* we are under rq->lock, defer unthrottling using a timer */
2833 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2834 !list_empty(&cfs_b->throttled_cfs_rq))
2835 start_cfs_slack_bandwidth(cfs_b);
2837 raw_spin_unlock(&cfs_b->lock);
2839 /* even if it's not valid for return we don't want to try again */
2840 cfs_rq->runtime_remaining -= slack_runtime;
2843 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2845 if (!cfs_bandwidth_used())
2848 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2851 __return_cfs_rq_runtime(cfs_rq);
2855 * This is done with a timer (instead of inline with bandwidth return) since
2856 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2858 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2860 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2863 /* confirm we're still not at a refresh boundary */
2864 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2867 raw_spin_lock(&cfs_b->lock);
2868 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2869 runtime = cfs_b->runtime;
2872 expires = cfs_b->runtime_expires;
2873 raw_spin_unlock(&cfs_b->lock);
2878 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2880 raw_spin_lock(&cfs_b->lock);
2881 if (expires == cfs_b->runtime_expires)
2882 cfs_b->runtime = runtime;
2883 raw_spin_unlock(&cfs_b->lock);
2887 * When a group wakes up we want to make sure that its quota is not already
2888 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2889 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2891 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2893 if (!cfs_bandwidth_used())
2896 /* an active group must be handled by the update_curr()->put() path */
2897 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2900 /* ensure the group is not already throttled */
2901 if (cfs_rq_throttled(cfs_rq))
2904 /* update runtime allocation */
2905 account_cfs_rq_runtime(cfs_rq, 0);
2906 if (cfs_rq->runtime_remaining <= 0)
2907 throttle_cfs_rq(cfs_rq);
2910 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2911 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2913 if (!cfs_bandwidth_used())
2916 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2920 * it's possible for a throttled entity to be forced into a running
2921 * state (e.g. set_curr_task), in this case we're finished.
2923 if (cfs_rq_throttled(cfs_rq))
2926 throttle_cfs_rq(cfs_rq);
2929 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2931 struct cfs_bandwidth *cfs_b =
2932 container_of(timer, struct cfs_bandwidth, slack_timer);
2933 do_sched_cfs_slack_timer(cfs_b);
2935 return HRTIMER_NORESTART;
2938 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2940 struct cfs_bandwidth *cfs_b =
2941 container_of(timer, struct cfs_bandwidth, period_timer);
2947 now = hrtimer_cb_get_time(timer);
2948 overrun = hrtimer_forward(timer, now, cfs_b->period);
2953 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2956 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2959 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2961 raw_spin_lock_init(&cfs_b->lock);
2963 cfs_b->quota = RUNTIME_INF;
2964 cfs_b->period = ns_to_ktime(default_cfs_period());
2966 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2967 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2968 cfs_b->period_timer.function = sched_cfs_period_timer;
2969 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2970 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2973 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2975 cfs_rq->runtime_enabled = 0;
2976 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2979 /* requires cfs_b->lock, may release to reprogram timer */
2980 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2983 * The timer may be active because we're trying to set a new bandwidth
2984 * period or because we're racing with the tear-down path
2985 * (timer_active==0 becomes visible before the hrtimer call-back
2986 * terminates). In either case we ensure that it's re-programmed
2988 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2989 raw_spin_unlock(&cfs_b->lock);
2990 /* ensure cfs_b->lock is available while we wait */
2991 hrtimer_cancel(&cfs_b->period_timer);
2993 raw_spin_lock(&cfs_b->lock);
2994 /* if someone else restarted the timer then we're done */
2995 if (cfs_b->timer_active)
2999 cfs_b->timer_active = 1;
3000 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3003 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3005 hrtimer_cancel(&cfs_b->period_timer);
3006 hrtimer_cancel(&cfs_b->slack_timer);
3009 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3011 struct cfs_rq *cfs_rq;
3013 for_each_leaf_cfs_rq(rq, cfs_rq) {
3014 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3016 if (!cfs_rq->runtime_enabled)
3020 * clock_task is not advancing so we just need to make sure
3021 * there's some valid quota amount
3023 cfs_rq->runtime_remaining = cfs_b->quota;
3024 if (cfs_rq_throttled(cfs_rq))
3025 unthrottle_cfs_rq(cfs_rq);
3029 #else /* CONFIG_CFS_BANDWIDTH */
3030 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3032 return rq_clock_task(rq_of(cfs_rq));
3035 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3036 unsigned long delta_exec) {}
3037 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3038 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3039 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3041 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3046 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3051 static inline int throttled_lb_pair(struct task_group *tg,
3052 int src_cpu, int dest_cpu)
3057 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3059 #ifdef CONFIG_FAIR_GROUP_SCHED
3060 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3063 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3067 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3068 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3070 #endif /* CONFIG_CFS_BANDWIDTH */
3072 /**************************************************
3073 * CFS operations on tasks:
3076 #ifdef CONFIG_SCHED_HRTICK
3077 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3079 struct sched_entity *se = &p->se;
3080 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3082 WARN_ON(task_rq(p) != rq);
3084 if (cfs_rq->nr_running > 1) {
3085 u64 slice = sched_slice(cfs_rq, se);
3086 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3087 s64 delta = slice - ran;
3096 * Don't schedule slices shorter than 10000ns, that just
3097 * doesn't make sense. Rely on vruntime for fairness.
3100 delta = max_t(s64, 10000LL, delta);
3102 hrtick_start(rq, delta);
3107 * called from enqueue/dequeue and updates the hrtick when the
3108 * current task is from our class and nr_running is low enough
3111 static void hrtick_update(struct rq *rq)
3113 struct task_struct *curr = rq->curr;
3115 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3118 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3119 hrtick_start_fair(rq, curr);
3121 #else /* !CONFIG_SCHED_HRTICK */
3123 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3127 static inline void hrtick_update(struct rq *rq)
3133 * The enqueue_task method is called before nr_running is
3134 * increased. Here we update the fair scheduling stats and
3135 * then put the task into the rbtree:
3138 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3140 struct cfs_rq *cfs_rq;
3141 struct sched_entity *se = &p->se;
3143 for_each_sched_entity(se) {
3146 cfs_rq = cfs_rq_of(se);
3147 enqueue_entity(cfs_rq, se, flags);
3150 * end evaluation on encountering a throttled cfs_rq
3152 * note: in the case of encountering a throttled cfs_rq we will
3153 * post the final h_nr_running increment below.
3155 if (cfs_rq_throttled(cfs_rq))
3157 cfs_rq->h_nr_running++;
3159 flags = ENQUEUE_WAKEUP;
3162 for_each_sched_entity(se) {
3163 cfs_rq = cfs_rq_of(se);
3164 cfs_rq->h_nr_running++;
3166 if (cfs_rq_throttled(cfs_rq))
3169 update_cfs_shares(cfs_rq);
3170 update_entity_load_avg(se, 1);
3174 update_rq_runnable_avg(rq, rq->nr_running);
3180 static void set_next_buddy(struct sched_entity *se);
3183 * The dequeue_task method is called before nr_running is
3184 * decreased. We remove the task from the rbtree and
3185 * update the fair scheduling stats:
3187 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3189 struct cfs_rq *cfs_rq;
3190 struct sched_entity *se = &p->se;
3191 int task_sleep = flags & DEQUEUE_SLEEP;
3193 for_each_sched_entity(se) {
3194 cfs_rq = cfs_rq_of(se);
3195 dequeue_entity(cfs_rq, se, flags);
3198 * end evaluation on encountering a throttled cfs_rq
3200 * note: in the case of encountering a throttled cfs_rq we will
3201 * post the final h_nr_running decrement below.
3203 if (cfs_rq_throttled(cfs_rq))
3205 cfs_rq->h_nr_running--;
3207 /* Don't dequeue parent if it has other entities besides us */
3208 if (cfs_rq->load.weight) {
3210 * Bias pick_next to pick a task from this cfs_rq, as
3211 * p is sleeping when it is within its sched_slice.
3213 if (task_sleep && parent_entity(se))
3214 set_next_buddy(parent_entity(se));
3216 /* avoid re-evaluating load for this entity */
3217 se = parent_entity(se);
3220 flags |= DEQUEUE_SLEEP;
3223 for_each_sched_entity(se) {
3224 cfs_rq = cfs_rq_of(se);
3225 cfs_rq->h_nr_running--;
3227 if (cfs_rq_throttled(cfs_rq))
3230 update_cfs_shares(cfs_rq);
3231 update_entity_load_avg(se, 1);
3236 update_rq_runnable_avg(rq, 1);
3242 /* Used instead of source_load when we know the type == 0 */
3243 static unsigned long weighted_cpuload(const int cpu)
3245 return cpu_rq(cpu)->cfs.runnable_load_avg;
3249 * Return a low guess at the load of a migration-source cpu weighted
3250 * according to the scheduling class and "nice" value.
3252 * We want to under-estimate the load of migration sources, to
3253 * balance conservatively.
3255 static unsigned long source_load(int cpu, int type)
3257 struct rq *rq = cpu_rq(cpu);
3258 unsigned long total = weighted_cpuload(cpu);
3260 if (type == 0 || !sched_feat(LB_BIAS))
3263 return min(rq->cpu_load[type-1], total);
3267 * Return a high guess at the load of a migration-target cpu weighted
3268 * according to the scheduling class and "nice" value.
3270 static unsigned long target_load(int cpu, int type)
3272 struct rq *rq = cpu_rq(cpu);
3273 unsigned long total = weighted_cpuload(cpu);
3275 if (type == 0 || !sched_feat(LB_BIAS))
3278 return max(rq->cpu_load[type-1], total);
3281 static unsigned long power_of(int cpu)
3283 return cpu_rq(cpu)->cpu_power;
3286 static unsigned long cpu_avg_load_per_task(int cpu)
3288 struct rq *rq = cpu_rq(cpu);
3289 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3290 unsigned long load_avg = rq->cfs.runnable_load_avg;
3293 return load_avg / nr_running;
3298 static void record_wakee(struct task_struct *p)
3301 * Rough decay (wiping) for cost saving, don't worry
3302 * about the boundary, really active task won't care
3305 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3306 current->wakee_flips = 0;
3307 current->wakee_flip_decay_ts = jiffies;
3310 if (current->last_wakee != p) {
3311 current->last_wakee = p;
3312 current->wakee_flips++;
3316 static void task_waking_fair(struct task_struct *p)
3318 struct sched_entity *se = &p->se;
3319 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3322 #ifndef CONFIG_64BIT
3323 u64 min_vruntime_copy;
3326 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3328 min_vruntime = cfs_rq->min_vruntime;
3329 } while (min_vruntime != min_vruntime_copy);
3331 min_vruntime = cfs_rq->min_vruntime;
3334 se->vruntime -= min_vruntime;
3338 #ifdef CONFIG_FAIR_GROUP_SCHED
3340 * effective_load() calculates the load change as seen from the root_task_group
3342 * Adding load to a group doesn't make a group heavier, but can cause movement
3343 * of group shares between cpus. Assuming the shares were perfectly aligned one
3344 * can calculate the shift in shares.
3346 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3347 * on this @cpu and results in a total addition (subtraction) of @wg to the
3348 * total group weight.
3350 * Given a runqueue weight distribution (rw_i) we can compute a shares
3351 * distribution (s_i) using:
3353 * s_i = rw_i / \Sum rw_j (1)
3355 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3356 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3357 * shares distribution (s_i):
3359 * rw_i = { 2, 4, 1, 0 }
3360 * s_i = { 2/7, 4/7, 1/7, 0 }
3362 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3363 * task used to run on and the CPU the waker is running on), we need to
3364 * compute the effect of waking a task on either CPU and, in case of a sync
3365 * wakeup, compute the effect of the current task going to sleep.
3367 * So for a change of @wl to the local @cpu with an overall group weight change
3368 * of @wl we can compute the new shares distribution (s'_i) using:
3370 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3372 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3373 * differences in waking a task to CPU 0. The additional task changes the
3374 * weight and shares distributions like:
3376 * rw'_i = { 3, 4, 1, 0 }
3377 * s'_i = { 3/8, 4/8, 1/8, 0 }
3379 * We can then compute the difference in effective weight by using:
3381 * dw_i = S * (s'_i - s_i) (3)
3383 * Where 'S' is the group weight as seen by its parent.
3385 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3386 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3387 * 4/7) times the weight of the group.
3389 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3391 struct sched_entity *se = tg->se[cpu];
3393 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3396 for_each_sched_entity(se) {
3402 * W = @wg + \Sum rw_j
3404 W = wg + calc_tg_weight(tg, se->my_q);
3409 w = se->my_q->load.weight + wl;
3412 * wl = S * s'_i; see (2)
3415 wl = (w * tg->shares) / W;
3420 * Per the above, wl is the new se->load.weight value; since
3421 * those are clipped to [MIN_SHARES, ...) do so now. See
3422 * calc_cfs_shares().
3424 if (wl < MIN_SHARES)
3428 * wl = dw_i = S * (s'_i - s_i); see (3)
3430 wl -= se->load.weight;
3433 * Recursively apply this logic to all parent groups to compute
3434 * the final effective load change on the root group. Since
3435 * only the @tg group gets extra weight, all parent groups can
3436 * only redistribute existing shares. @wl is the shift in shares
3437 * resulting from this level per the above.
3446 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3453 static int wake_wide(struct task_struct *p)
3455 int factor = this_cpu_read(sd_llc_size);
3458 * Yeah, it's the switching-frequency, could means many wakee or
3459 * rapidly switch, use factor here will just help to automatically
3460 * adjust the loose-degree, so bigger node will lead to more pull.
3462 if (p->wakee_flips > factor) {
3464 * wakee is somewhat hot, it needs certain amount of cpu
3465 * resource, so if waker is far more hot, prefer to leave
3468 if (current->wakee_flips > (factor * p->wakee_flips))
3475 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3477 s64 this_load, load;
3478 int idx, this_cpu, prev_cpu;
3479 unsigned long tl_per_task;
3480 struct task_group *tg;
3481 unsigned long weight;
3485 * If we wake multiple tasks be careful to not bounce
3486 * ourselves around too much.
3492 this_cpu = smp_processor_id();
3493 prev_cpu = task_cpu(p);
3494 load = source_load(prev_cpu, idx);
3495 this_load = target_load(this_cpu, idx);
3498 * If sync wakeup then subtract the (maximum possible)
3499 * effect of the currently running task from the load
3500 * of the current CPU:
3503 tg = task_group(current);
3504 weight = current->se.load.weight;
3506 this_load += effective_load(tg, this_cpu, -weight, -weight);
3507 load += effective_load(tg, prev_cpu, 0, -weight);
3511 weight = p->se.load.weight;
3514 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3515 * due to the sync cause above having dropped this_load to 0, we'll
3516 * always have an imbalance, but there's really nothing you can do
3517 * about that, so that's good too.
3519 * Otherwise check if either cpus are near enough in load to allow this
3520 * task to be woken on this_cpu.
3522 if (this_load > 0) {
3523 s64 this_eff_load, prev_eff_load;
3525 this_eff_load = 100;
3526 this_eff_load *= power_of(prev_cpu);
3527 this_eff_load *= this_load +
3528 effective_load(tg, this_cpu, weight, weight);
3530 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3531 prev_eff_load *= power_of(this_cpu);
3532 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3534 balanced = this_eff_load <= prev_eff_load;
3539 * If the currently running task will sleep within
3540 * a reasonable amount of time then attract this newly
3543 if (sync && balanced)
3546 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3547 tl_per_task = cpu_avg_load_per_task(this_cpu);
3550 (this_load <= load &&
3551 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3553 * This domain has SD_WAKE_AFFINE and
3554 * p is cache cold in this domain, and
3555 * there is no bad imbalance.
3557 schedstat_inc(sd, ttwu_move_affine);
3558 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3566 * find_idlest_group finds and returns the least busy CPU group within the
3569 static struct sched_group *
3570 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3571 int this_cpu, int load_idx)
3573 struct sched_group *idlest = NULL, *group = sd->groups;
3574 unsigned long min_load = ULONG_MAX, this_load = 0;
3575 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3578 unsigned long load, avg_load;
3582 /* Skip over this group if it has no CPUs allowed */
3583 if (!cpumask_intersects(sched_group_cpus(group),
3584 tsk_cpus_allowed(p)))
3587 local_group = cpumask_test_cpu(this_cpu,
3588 sched_group_cpus(group));
3590 /* Tally up the load of all CPUs in the group */
3593 for_each_cpu(i, sched_group_cpus(group)) {
3594 /* Bias balancing toward cpus of our domain */
3596 load = source_load(i, load_idx);
3598 load = target_load(i, load_idx);
3603 /* Adjust by relative CPU power of the group */
3604 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3607 this_load = avg_load;
3608 } else if (avg_load < min_load) {
3609 min_load = avg_load;
3612 } while (group = group->next, group != sd->groups);
3614 if (!idlest || 100*this_load < imbalance*min_load)
3620 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3623 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3625 unsigned long load, min_load = ULONG_MAX;
3629 /* Traverse only the allowed CPUs */
3630 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3631 load = weighted_cpuload(i);
3633 if (load < min_load || (load == min_load && i == this_cpu)) {
3643 * Try and locate an idle CPU in the sched_domain.
3645 static int select_idle_sibling(struct task_struct *p, int target)
3647 struct sched_domain *sd;
3648 struct sched_group *sg;
3649 int i = task_cpu(p);
3651 if (idle_cpu(target))
3655 * If the prevous cpu is cache affine and idle, don't be stupid.
3657 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3661 * Otherwise, iterate the domains and find an elegible idle cpu.
3663 sd = rcu_dereference(per_cpu(sd_llc, target));
3664 for_each_lower_domain(sd) {
3667 if (!cpumask_intersects(sched_group_cpus(sg),
3668 tsk_cpus_allowed(p)))
3671 for_each_cpu(i, sched_group_cpus(sg)) {
3672 if (i == target || !idle_cpu(i))
3676 target = cpumask_first_and(sched_group_cpus(sg),
3677 tsk_cpus_allowed(p));
3681 } while (sg != sd->groups);
3688 * sched_balance_self: balance the current task (running on cpu) in domains
3689 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3692 * Balance, ie. select the least loaded group.
3694 * Returns the target CPU number, or the same CPU if no balancing is needed.
3696 * preempt must be disabled.
3699 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3701 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3702 int cpu = smp_processor_id();
3703 int prev_cpu = task_cpu(p);
3705 int want_affine = 0;
3706 int sync = wake_flags & WF_SYNC;
3708 if (p->nr_cpus_allowed == 1)
3711 if (sd_flag & SD_BALANCE_WAKE) {
3712 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3718 for_each_domain(cpu, tmp) {
3719 if (!(tmp->flags & SD_LOAD_BALANCE))
3723 * If both cpu and prev_cpu are part of this domain,
3724 * cpu is a valid SD_WAKE_AFFINE target.
3726 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3727 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3732 if (tmp->flags & sd_flag)
3737 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3740 new_cpu = select_idle_sibling(p, prev_cpu);
3745 int load_idx = sd->forkexec_idx;
3746 struct sched_group *group;
3749 if (!(sd->flags & sd_flag)) {
3754 if (sd_flag & SD_BALANCE_WAKE)
3755 load_idx = sd->wake_idx;
3757 group = find_idlest_group(sd, p, cpu, load_idx);
3763 new_cpu = find_idlest_cpu(group, p, cpu);
3764 if (new_cpu == -1 || new_cpu == cpu) {
3765 /* Now try balancing at a lower domain level of cpu */
3770 /* Now try balancing at a lower domain level of new_cpu */
3772 weight = sd->span_weight;
3774 for_each_domain(cpu, tmp) {
3775 if (weight <= tmp->span_weight)
3777 if (tmp->flags & sd_flag)
3780 /* while loop will break here if sd == NULL */
3789 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3790 * cfs_rq_of(p) references at time of call are still valid and identify the
3791 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3792 * other assumptions, including the state of rq->lock, should be made.
3795 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3797 struct sched_entity *se = &p->se;
3798 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3801 * Load tracking: accumulate removed load so that it can be processed
3802 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3803 * to blocked load iff they have a positive decay-count. It can never
3804 * be negative here since on-rq tasks have decay-count == 0.
3806 if (se->avg.decay_count) {
3807 se->avg.decay_count = -__synchronize_entity_decay(se);
3808 atomic_long_add(se->avg.load_avg_contrib,
3809 &cfs_rq->removed_load);
3812 #endif /* CONFIG_SMP */
3814 static unsigned long
3815 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3817 unsigned long gran = sysctl_sched_wakeup_granularity;
3820 * Since its curr running now, convert the gran from real-time
3821 * to virtual-time in his units.
3823 * By using 'se' instead of 'curr' we penalize light tasks, so
3824 * they get preempted easier. That is, if 'se' < 'curr' then
3825 * the resulting gran will be larger, therefore penalizing the
3826 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3827 * be smaller, again penalizing the lighter task.
3829 * This is especially important for buddies when the leftmost
3830 * task is higher priority than the buddy.
3832 return calc_delta_fair(gran, se);
3836 * Should 'se' preempt 'curr'.
3850 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3852 s64 gran, vdiff = curr->vruntime - se->vruntime;
3857 gran = wakeup_gran(curr, se);
3864 static void set_last_buddy(struct sched_entity *se)
3866 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3869 for_each_sched_entity(se)
3870 cfs_rq_of(se)->last = se;
3873 static void set_next_buddy(struct sched_entity *se)
3875 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3878 for_each_sched_entity(se)
3879 cfs_rq_of(se)->next = se;
3882 static void set_skip_buddy(struct sched_entity *se)
3884 for_each_sched_entity(se)
3885 cfs_rq_of(se)->skip = se;
3889 * Preempt the current task with a newly woken task if needed:
3891 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3893 struct task_struct *curr = rq->curr;
3894 struct sched_entity *se = &curr->se, *pse = &p->se;
3895 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3896 int scale = cfs_rq->nr_running >= sched_nr_latency;
3897 int next_buddy_marked = 0;
3899 if (unlikely(se == pse))
3903 * This is possible from callers such as move_task(), in which we
3904 * unconditionally check_prempt_curr() after an enqueue (which may have
3905 * lead to a throttle). This both saves work and prevents false
3906 * next-buddy nomination below.
3908 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3911 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3912 set_next_buddy(pse);
3913 next_buddy_marked = 1;
3917 * We can come here with TIF_NEED_RESCHED already set from new task
3920 * Note: this also catches the edge-case of curr being in a throttled
3921 * group (e.g. via set_curr_task), since update_curr() (in the
3922 * enqueue of curr) will have resulted in resched being set. This
3923 * prevents us from potentially nominating it as a false LAST_BUDDY
3926 if (test_tsk_need_resched(curr))
3929 /* Idle tasks are by definition preempted by non-idle tasks. */
3930 if (unlikely(curr->policy == SCHED_IDLE) &&
3931 likely(p->policy != SCHED_IDLE))
3935 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3936 * is driven by the tick):
3938 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3941 find_matching_se(&se, &pse);
3942 update_curr(cfs_rq_of(se));
3944 if (wakeup_preempt_entity(se, pse) == 1) {
3946 * Bias pick_next to pick the sched entity that is
3947 * triggering this preemption.
3949 if (!next_buddy_marked)
3950 set_next_buddy(pse);
3959 * Only set the backward buddy when the current task is still
3960 * on the rq. This can happen when a wakeup gets interleaved
3961 * with schedule on the ->pre_schedule() or idle_balance()
3962 * point, either of which can * drop the rq lock.
3964 * Also, during early boot the idle thread is in the fair class,
3965 * for obvious reasons its a bad idea to schedule back to it.
3967 if (unlikely(!se->on_rq || curr == rq->idle))
3970 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3974 static struct task_struct *pick_next_task_fair(struct rq *rq)
3976 struct task_struct *p;
3977 struct cfs_rq *cfs_rq = &rq->cfs;
3978 struct sched_entity *se;
3980 if (!cfs_rq->nr_running)
3984 se = pick_next_entity(cfs_rq);
3985 set_next_entity(cfs_rq, se);
3986 cfs_rq = group_cfs_rq(se);
3990 if (hrtick_enabled(rq))
3991 hrtick_start_fair(rq, p);
3997 * Account for a descheduled task:
3999 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4001 struct sched_entity *se = &prev->se;
4002 struct cfs_rq *cfs_rq;
4004 for_each_sched_entity(se) {
4005 cfs_rq = cfs_rq_of(se);
4006 put_prev_entity(cfs_rq, se);
4011 * sched_yield() is very simple
4013 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4015 static void yield_task_fair(struct rq *rq)
4017 struct task_struct *curr = rq->curr;
4018 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4019 struct sched_entity *se = &curr->se;
4022 * Are we the only task in the tree?
4024 if (unlikely(rq->nr_running == 1))
4027 clear_buddies(cfs_rq, se);
4029 if (curr->policy != SCHED_BATCH) {
4030 update_rq_clock(rq);
4032 * Update run-time statistics of the 'current'.
4034 update_curr(cfs_rq);
4036 * Tell update_rq_clock() that we've just updated,
4037 * so we don't do microscopic update in schedule()
4038 * and double the fastpath cost.
4040 rq->skip_clock_update = 1;
4046 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4048 struct sched_entity *se = &p->se;
4050 /* throttled hierarchies are not runnable */
4051 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4054 /* Tell the scheduler that we'd really like pse to run next. */
4057 yield_task_fair(rq);
4063 /**************************************************
4064 * Fair scheduling class load-balancing methods.
4068 * The purpose of load-balancing is to achieve the same basic fairness the
4069 * per-cpu scheduler provides, namely provide a proportional amount of compute
4070 * time to each task. This is expressed in the following equation:
4072 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4074 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4075 * W_i,0 is defined as:
4077 * W_i,0 = \Sum_j w_i,j (2)
4079 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4080 * is derived from the nice value as per prio_to_weight[].
4082 * The weight average is an exponential decay average of the instantaneous
4085 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4087 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4088 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4089 * can also include other factors [XXX].
4091 * To achieve this balance we define a measure of imbalance which follows
4092 * directly from (1):
4094 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4096 * We them move tasks around to minimize the imbalance. In the continuous
4097 * function space it is obvious this converges, in the discrete case we get
4098 * a few fun cases generally called infeasible weight scenarios.
4101 * - infeasible weights;
4102 * - local vs global optima in the discrete case. ]
4107 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4108 * for all i,j solution, we create a tree of cpus that follows the hardware
4109 * topology where each level pairs two lower groups (or better). This results
4110 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4111 * tree to only the first of the previous level and we decrease the frequency
4112 * of load-balance at each level inv. proportional to the number of cpus in
4118 * \Sum { --- * --- * 2^i } = O(n) (5)
4120 * `- size of each group
4121 * | | `- number of cpus doing load-balance
4123 * `- sum over all levels
4125 * Coupled with a limit on how many tasks we can migrate every balance pass,
4126 * this makes (5) the runtime complexity of the balancer.
4128 * An important property here is that each CPU is still (indirectly) connected
4129 * to every other cpu in at most O(log n) steps:
4131 * The adjacency matrix of the resulting graph is given by:
4134 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4137 * And you'll find that:
4139 * A^(log_2 n)_i,j != 0 for all i,j (7)
4141 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4142 * The task movement gives a factor of O(m), giving a convergence complexity
4145 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4150 * In order to avoid CPUs going idle while there's still work to do, new idle
4151 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4152 * tree itself instead of relying on other CPUs to bring it work.
4154 * This adds some complexity to both (5) and (8) but it reduces the total idle
4162 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4165 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4170 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4172 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4174 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4177 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4178 * rewrite all of this once again.]
4181 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4183 #define LBF_ALL_PINNED 0x01
4184 #define LBF_NEED_BREAK 0x02
4185 #define LBF_DST_PINNED 0x04
4186 #define LBF_SOME_PINNED 0x08
4189 struct sched_domain *sd;
4197 struct cpumask *dst_grpmask;
4199 enum cpu_idle_type idle;
4201 /* The set of CPUs under consideration for load-balancing */
4202 struct cpumask *cpus;
4207 unsigned int loop_break;
4208 unsigned int loop_max;
4212 * move_task - move a task from one runqueue to another runqueue.
4213 * Both runqueues must be locked.
4215 static void move_task(struct task_struct *p, struct lb_env *env)
4217 deactivate_task(env->src_rq, p, 0);
4218 set_task_cpu(p, env->dst_cpu);
4219 activate_task(env->dst_rq, p, 0);
4220 check_preempt_curr(env->dst_rq, p, 0);
4221 #ifdef CONFIG_NUMA_BALANCING
4222 if (p->numa_preferred_nid != -1) {
4223 int src_nid = cpu_to_node(env->src_cpu);
4224 int dst_nid = cpu_to_node(env->dst_cpu);
4227 * If the load balancer has moved the task then limit
4228 * migrations from taking place in the short term in
4229 * case this is a short-lived migration.
4231 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4232 p->numa_migrate_seq = 0;
4238 * Is this task likely cache-hot:
4241 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4245 if (p->sched_class != &fair_sched_class)
4248 if (unlikely(p->policy == SCHED_IDLE))
4252 * Buddy candidates are cache hot:
4254 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4255 (&p->se == cfs_rq_of(&p->se)->next ||
4256 &p->se == cfs_rq_of(&p->se)->last))
4259 if (sysctl_sched_migration_cost == -1)
4261 if (sysctl_sched_migration_cost == 0)
4264 delta = now - p->se.exec_start;
4266 return delta < (s64)sysctl_sched_migration_cost;
4269 #ifdef CONFIG_NUMA_BALANCING
4270 /* Returns true if the destination node has incurred more faults */
4271 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4273 int src_nid, dst_nid;
4275 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4276 !(env->sd->flags & SD_NUMA)) {
4280 src_nid = cpu_to_node(env->src_cpu);
4281 dst_nid = cpu_to_node(env->dst_cpu);
4283 if (src_nid == dst_nid ||
4284 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4287 if (dst_nid == p->numa_preferred_nid ||
4288 task_faults(p, dst_nid) > task_faults(p, src_nid))
4295 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4297 int src_nid, dst_nid;
4299 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4302 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4305 src_nid = cpu_to_node(env->src_cpu);
4306 dst_nid = cpu_to_node(env->dst_cpu);
4308 if (src_nid == dst_nid ||
4309 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4312 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4319 static inline bool migrate_improves_locality(struct task_struct *p,
4325 static inline bool migrate_degrades_locality(struct task_struct *p,
4333 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4336 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4338 int tsk_cache_hot = 0;
4340 * We do not migrate tasks that are:
4341 * 1) throttled_lb_pair, or
4342 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4343 * 3) running (obviously), or
4344 * 4) are cache-hot on their current CPU.
4346 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4349 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4352 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4354 env->flags |= LBF_SOME_PINNED;
4357 * Remember if this task can be migrated to any other cpu in
4358 * our sched_group. We may want to revisit it if we couldn't
4359 * meet load balance goals by pulling other tasks on src_cpu.
4361 * Also avoid computing new_dst_cpu if we have already computed
4362 * one in current iteration.
4364 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4367 /* Prevent to re-select dst_cpu via env's cpus */
4368 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4369 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4370 env->flags |= LBF_DST_PINNED;
4371 env->new_dst_cpu = cpu;
4379 /* Record that we found atleast one task that could run on dst_cpu */
4380 env->flags &= ~LBF_ALL_PINNED;
4382 if (task_running(env->src_rq, p)) {
4383 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4388 * Aggressive migration if:
4389 * 1) destination numa is preferred
4390 * 2) task is cache cold, or
4391 * 3) too many balance attempts have failed.
4393 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4395 tsk_cache_hot = migrate_degrades_locality(p, env);
4397 if (migrate_improves_locality(p, env)) {
4398 #ifdef CONFIG_SCHEDSTATS
4399 if (tsk_cache_hot) {
4400 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4401 schedstat_inc(p, se.statistics.nr_forced_migrations);
4407 if (!tsk_cache_hot ||
4408 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4410 if (tsk_cache_hot) {
4411 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4412 schedstat_inc(p, se.statistics.nr_forced_migrations);
4418 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4423 * move_one_task tries to move exactly one task from busiest to this_rq, as
4424 * part of active balancing operations within "domain".
4425 * Returns 1 if successful and 0 otherwise.
4427 * Called with both runqueues locked.
4429 static int move_one_task(struct lb_env *env)
4431 struct task_struct *p, *n;
4433 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4434 if (!can_migrate_task(p, env))
4439 * Right now, this is only the second place move_task()
4440 * is called, so we can safely collect move_task()
4441 * stats here rather than inside move_task().
4443 schedstat_inc(env->sd, lb_gained[env->idle]);
4449 static unsigned long task_h_load(struct task_struct *p);
4451 static const unsigned int sched_nr_migrate_break = 32;
4454 * move_tasks tries to move up to imbalance weighted load from busiest to
4455 * this_rq, as part of a balancing operation within domain "sd".
4456 * Returns 1 if successful and 0 otherwise.
4458 * Called with both runqueues locked.
4460 static int move_tasks(struct lb_env *env)
4462 struct list_head *tasks = &env->src_rq->cfs_tasks;
4463 struct task_struct *p;
4467 if (env->imbalance <= 0)
4470 while (!list_empty(tasks)) {
4471 p = list_first_entry(tasks, struct task_struct, se.group_node);
4474 /* We've more or less seen every task there is, call it quits */
4475 if (env->loop > env->loop_max)
4478 /* take a breather every nr_migrate tasks */
4479 if (env->loop > env->loop_break) {
4480 env->loop_break += sched_nr_migrate_break;
4481 env->flags |= LBF_NEED_BREAK;
4485 if (!can_migrate_task(p, env))
4488 load = task_h_load(p);
4490 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4493 if ((load / 2) > env->imbalance)
4498 env->imbalance -= load;
4500 #ifdef CONFIG_PREEMPT
4502 * NEWIDLE balancing is a source of latency, so preemptible
4503 * kernels will stop after the first task is pulled to minimize
4504 * the critical section.
4506 if (env->idle == CPU_NEWLY_IDLE)
4511 * We only want to steal up to the prescribed amount of
4514 if (env->imbalance <= 0)
4519 list_move_tail(&p->se.group_node, tasks);
4523 * Right now, this is one of only two places move_task() is called,
4524 * so we can safely collect move_task() stats here rather than
4525 * inside move_task().
4527 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4532 #ifdef CONFIG_FAIR_GROUP_SCHED
4534 * update tg->load_weight by folding this cpu's load_avg
4536 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4538 struct sched_entity *se = tg->se[cpu];
4539 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4541 /* throttled entities do not contribute to load */
4542 if (throttled_hierarchy(cfs_rq))
4545 update_cfs_rq_blocked_load(cfs_rq, 1);
4548 update_entity_load_avg(se, 1);
4550 * We pivot on our runnable average having decayed to zero for
4551 * list removal. This generally implies that all our children
4552 * have also been removed (modulo rounding error or bandwidth
4553 * control); however, such cases are rare and we can fix these
4556 * TODO: fix up out-of-order children on enqueue.
4558 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4559 list_del_leaf_cfs_rq(cfs_rq);
4561 struct rq *rq = rq_of(cfs_rq);
4562 update_rq_runnable_avg(rq, rq->nr_running);
4566 static void update_blocked_averages(int cpu)
4568 struct rq *rq = cpu_rq(cpu);
4569 struct cfs_rq *cfs_rq;
4570 unsigned long flags;
4572 raw_spin_lock_irqsave(&rq->lock, flags);
4573 update_rq_clock(rq);
4575 * Iterates the task_group tree in a bottom up fashion, see
4576 * list_add_leaf_cfs_rq() for details.
4578 for_each_leaf_cfs_rq(rq, cfs_rq) {
4580 * Note: We may want to consider periodically releasing
4581 * rq->lock about these updates so that creating many task
4582 * groups does not result in continually extending hold time.
4584 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4587 raw_spin_unlock_irqrestore(&rq->lock, flags);
4591 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4592 * This needs to be done in a top-down fashion because the load of a child
4593 * group is a fraction of its parents load.
4595 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4597 struct rq *rq = rq_of(cfs_rq);
4598 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4599 unsigned long now = jiffies;
4602 if (cfs_rq->last_h_load_update == now)
4605 cfs_rq->h_load_next = NULL;
4606 for_each_sched_entity(se) {
4607 cfs_rq = cfs_rq_of(se);
4608 cfs_rq->h_load_next = se;
4609 if (cfs_rq->last_h_load_update == now)
4614 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4615 cfs_rq->last_h_load_update = now;
4618 while ((se = cfs_rq->h_load_next) != NULL) {
4619 load = cfs_rq->h_load;
4620 load = div64_ul(load * se->avg.load_avg_contrib,
4621 cfs_rq->runnable_load_avg + 1);
4622 cfs_rq = group_cfs_rq(se);
4623 cfs_rq->h_load = load;
4624 cfs_rq->last_h_load_update = now;
4628 static unsigned long task_h_load(struct task_struct *p)
4630 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4632 update_cfs_rq_h_load(cfs_rq);
4633 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4634 cfs_rq->runnable_load_avg + 1);
4637 static inline void update_blocked_averages(int cpu)
4641 static unsigned long task_h_load(struct task_struct *p)
4643 return p->se.avg.load_avg_contrib;
4647 /********** Helpers for find_busiest_group ************************/
4649 * sg_lb_stats - stats of a sched_group required for load_balancing
4651 struct sg_lb_stats {
4652 unsigned long avg_load; /*Avg load across the CPUs of the group */
4653 unsigned long group_load; /* Total load over the CPUs of the group */
4654 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4655 unsigned long load_per_task;
4656 unsigned long group_power;
4657 unsigned int sum_nr_running; /* Nr tasks running in the group */
4658 unsigned int group_capacity;
4659 unsigned int idle_cpus;
4660 unsigned int group_weight;
4661 int group_imb; /* Is there an imbalance in the group ? */
4662 int group_has_capacity; /* Is there extra capacity in the group? */
4666 * sd_lb_stats - Structure to store the statistics of a sched_domain
4667 * during load balancing.
4669 struct sd_lb_stats {
4670 struct sched_group *busiest; /* Busiest group in this sd */
4671 struct sched_group *local; /* Local group in this sd */
4672 unsigned long total_load; /* Total load of all groups in sd */
4673 unsigned long total_pwr; /* Total power of all groups in sd */
4674 unsigned long avg_load; /* Average load across all groups in sd */
4676 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4677 struct sg_lb_stats local_stat; /* Statistics of the local group */
4680 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4683 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4684 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4685 * We must however clear busiest_stat::avg_load because
4686 * update_sd_pick_busiest() reads this before assignment.
4688 *sds = (struct sd_lb_stats){
4700 * get_sd_load_idx - Obtain the load index for a given sched domain.
4701 * @sd: The sched_domain whose load_idx is to be obtained.
4702 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4704 * Return: The load index.
4706 static inline int get_sd_load_idx(struct sched_domain *sd,
4707 enum cpu_idle_type idle)
4713 load_idx = sd->busy_idx;
4716 case CPU_NEWLY_IDLE:
4717 load_idx = sd->newidle_idx;
4720 load_idx = sd->idle_idx;
4727 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4729 return SCHED_POWER_SCALE;
4732 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4734 return default_scale_freq_power(sd, cpu);
4737 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4739 unsigned long weight = sd->span_weight;
4740 unsigned long smt_gain = sd->smt_gain;
4747 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4749 return default_scale_smt_power(sd, cpu);
4752 static unsigned long scale_rt_power(int cpu)
4754 struct rq *rq = cpu_rq(cpu);
4755 u64 total, available, age_stamp, avg;
4758 * Since we're reading these variables without serialization make sure
4759 * we read them once before doing sanity checks on them.
4761 age_stamp = ACCESS_ONCE(rq->age_stamp);
4762 avg = ACCESS_ONCE(rq->rt_avg);
4764 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4766 if (unlikely(total < avg)) {
4767 /* Ensures that power won't end up being negative */
4770 available = total - avg;
4773 if (unlikely((s64)total < SCHED_POWER_SCALE))
4774 total = SCHED_POWER_SCALE;
4776 total >>= SCHED_POWER_SHIFT;
4778 return div_u64(available, total);
4781 static void update_cpu_power(struct sched_domain *sd, int cpu)
4783 unsigned long weight = sd->span_weight;
4784 unsigned long power = SCHED_POWER_SCALE;
4785 struct sched_group *sdg = sd->groups;
4787 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4788 if (sched_feat(ARCH_POWER))
4789 power *= arch_scale_smt_power(sd, cpu);
4791 power *= default_scale_smt_power(sd, cpu);
4793 power >>= SCHED_POWER_SHIFT;
4796 sdg->sgp->power_orig = power;
4798 if (sched_feat(ARCH_POWER))
4799 power *= arch_scale_freq_power(sd, cpu);
4801 power *= default_scale_freq_power(sd, cpu);
4803 power >>= SCHED_POWER_SHIFT;
4805 power *= scale_rt_power(cpu);
4806 power >>= SCHED_POWER_SHIFT;
4811 cpu_rq(cpu)->cpu_power = power;
4812 sdg->sgp->power = power;
4815 void update_group_power(struct sched_domain *sd, int cpu)
4817 struct sched_domain *child = sd->child;
4818 struct sched_group *group, *sdg = sd->groups;
4819 unsigned long power, power_orig;
4820 unsigned long interval;
4822 interval = msecs_to_jiffies(sd->balance_interval);
4823 interval = clamp(interval, 1UL, max_load_balance_interval);
4824 sdg->sgp->next_update = jiffies + interval;
4827 update_cpu_power(sd, cpu);
4831 power_orig = power = 0;
4833 if (child->flags & SD_OVERLAP) {
4835 * SD_OVERLAP domains cannot assume that child groups
4836 * span the current group.
4839 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4840 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4842 power_orig += sg->sgp->power_orig;
4843 power += sg->sgp->power;
4847 * !SD_OVERLAP domains can assume that child groups
4848 * span the current group.
4851 group = child->groups;
4853 power_orig += group->sgp->power_orig;
4854 power += group->sgp->power;
4855 group = group->next;
4856 } while (group != child->groups);
4859 sdg->sgp->power_orig = power_orig;
4860 sdg->sgp->power = power;
4864 * Try and fix up capacity for tiny siblings, this is needed when
4865 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4866 * which on its own isn't powerful enough.
4868 * See update_sd_pick_busiest() and check_asym_packing().
4871 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4874 * Only siblings can have significantly less than SCHED_POWER_SCALE
4876 if (!(sd->flags & SD_SHARE_CPUPOWER))
4880 * If ~90% of the cpu_power is still there, we're good.
4882 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4889 * Group imbalance indicates (and tries to solve) the problem where balancing
4890 * groups is inadequate due to tsk_cpus_allowed() constraints.
4892 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4893 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4896 * { 0 1 2 3 } { 4 5 6 7 }
4899 * If we were to balance group-wise we'd place two tasks in the first group and
4900 * two tasks in the second group. Clearly this is undesired as it will overload
4901 * cpu 3 and leave one of the cpus in the second group unused.
4903 * The current solution to this issue is detecting the skew in the first group
4904 * by noticing the lower domain failed to reach balance and had difficulty
4905 * moving tasks due to affinity constraints.
4907 * When this is so detected; this group becomes a candidate for busiest; see
4908 * update_sd_pick_busiest(). And calculcate_imbalance() and
4909 * find_busiest_group() avoid some of the usual balance conditions to allow it
4910 * to create an effective group imbalance.
4912 * This is a somewhat tricky proposition since the next run might not find the
4913 * group imbalance and decide the groups need to be balanced again. A most
4914 * subtle and fragile situation.
4917 static inline int sg_imbalanced(struct sched_group *group)
4919 return group->sgp->imbalance;
4923 * Compute the group capacity.
4925 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4926 * first dividing out the smt factor and computing the actual number of cores
4927 * and limit power unit capacity with that.
4929 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4931 unsigned int capacity, smt, cpus;
4932 unsigned int power, power_orig;
4934 power = group->sgp->power;
4935 power_orig = group->sgp->power_orig;
4936 cpus = group->group_weight;
4938 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4939 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4940 capacity = cpus / smt; /* cores */
4942 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4944 capacity = fix_small_capacity(env->sd, group);
4950 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4951 * @env: The load balancing environment.
4952 * @group: sched_group whose statistics are to be updated.
4953 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4954 * @local_group: Does group contain this_cpu.
4955 * @sgs: variable to hold the statistics for this group.
4957 static inline void update_sg_lb_stats(struct lb_env *env,
4958 struct sched_group *group, int load_idx,
4959 int local_group, struct sg_lb_stats *sgs)
4961 unsigned long nr_running;
4965 memset(sgs, 0, sizeof(*sgs));
4967 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4968 struct rq *rq = cpu_rq(i);
4970 nr_running = rq->nr_running;
4972 /* Bias balancing toward cpus of our domain */
4974 load = target_load(i, load_idx);
4976 load = source_load(i, load_idx);
4978 sgs->group_load += load;
4979 sgs->sum_nr_running += nr_running;
4980 sgs->sum_weighted_load += weighted_cpuload(i);
4985 /* Adjust by relative CPU power of the group */
4986 sgs->group_power = group->sgp->power;
4987 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4989 if (sgs->sum_nr_running)
4990 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4992 sgs->group_weight = group->group_weight;
4994 sgs->group_imb = sg_imbalanced(group);
4995 sgs->group_capacity = sg_capacity(env, group);
4997 if (sgs->group_capacity > sgs->sum_nr_running)
4998 sgs->group_has_capacity = 1;
5002 * update_sd_pick_busiest - return 1 on busiest group
5003 * @env: The load balancing environment.
5004 * @sds: sched_domain statistics
5005 * @sg: sched_group candidate to be checked for being the busiest
5006 * @sgs: sched_group statistics
5008 * Determine if @sg is a busier group than the previously selected
5011 * Return: %true if @sg is a busier group than the previously selected
5012 * busiest group. %false otherwise.
5014 static bool update_sd_pick_busiest(struct lb_env *env,
5015 struct sd_lb_stats *sds,
5016 struct sched_group *sg,
5017 struct sg_lb_stats *sgs)
5019 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5022 if (sgs->sum_nr_running > sgs->group_capacity)
5029 * ASYM_PACKING needs to move all the work to the lowest
5030 * numbered CPUs in the group, therefore mark all groups
5031 * higher than ourself as busy.
5033 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5034 env->dst_cpu < group_first_cpu(sg)) {
5038 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5046 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5047 * @env: The load balancing environment.
5048 * @balance: Should we balance.
5049 * @sds: variable to hold the statistics for this sched_domain.
5051 static inline void update_sd_lb_stats(struct lb_env *env,
5052 struct sd_lb_stats *sds)
5054 struct sched_domain *child = env->sd->child;
5055 struct sched_group *sg = env->sd->groups;
5056 struct sg_lb_stats tmp_sgs;
5057 int load_idx, prefer_sibling = 0;
5059 if (child && child->flags & SD_PREFER_SIBLING)
5062 load_idx = get_sd_load_idx(env->sd, env->idle);
5065 struct sg_lb_stats *sgs = &tmp_sgs;
5068 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5071 sgs = &sds->local_stat;
5073 if (env->idle != CPU_NEWLY_IDLE ||
5074 time_after_eq(jiffies, sg->sgp->next_update))
5075 update_group_power(env->sd, env->dst_cpu);
5078 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5084 * In case the child domain prefers tasks go to siblings
5085 * first, lower the sg capacity to one so that we'll try
5086 * and move all the excess tasks away. We lower the capacity
5087 * of a group only if the local group has the capacity to fit
5088 * these excess tasks, i.e. nr_running < group_capacity. The
5089 * extra check prevents the case where you always pull from the
5090 * heaviest group when it is already under-utilized (possible
5091 * with a large weight task outweighs the tasks on the system).
5093 if (prefer_sibling && sds->local &&
5094 sds->local_stat.group_has_capacity)
5095 sgs->group_capacity = min(sgs->group_capacity, 1U);
5097 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5099 sds->busiest_stat = *sgs;
5103 /* Now, start updating sd_lb_stats */
5104 sds->total_load += sgs->group_load;
5105 sds->total_pwr += sgs->group_power;
5108 } while (sg != env->sd->groups);
5112 * check_asym_packing - Check to see if the group is packed into the
5115 * This is primarily intended to used at the sibling level. Some
5116 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5117 * case of POWER7, it can move to lower SMT modes only when higher
5118 * threads are idle. When in lower SMT modes, the threads will
5119 * perform better since they share less core resources. Hence when we
5120 * have idle threads, we want them to be the higher ones.
5122 * This packing function is run on idle threads. It checks to see if
5123 * the busiest CPU in this domain (core in the P7 case) has a higher
5124 * CPU number than the packing function is being run on. Here we are
5125 * assuming lower CPU number will be equivalent to lower a SMT thread
5128 * Return: 1 when packing is required and a task should be moved to
5129 * this CPU. The amount of the imbalance is returned in *imbalance.
5131 * @env: The load balancing environment.
5132 * @sds: Statistics of the sched_domain which is to be packed
5134 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5138 if (!(env->sd->flags & SD_ASYM_PACKING))
5144 busiest_cpu = group_first_cpu(sds->busiest);
5145 if (env->dst_cpu > busiest_cpu)
5148 env->imbalance = DIV_ROUND_CLOSEST(
5149 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5156 * fix_small_imbalance - Calculate the minor imbalance that exists
5157 * amongst the groups of a sched_domain, during
5159 * @env: The load balancing environment.
5160 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5163 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5165 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5166 unsigned int imbn = 2;
5167 unsigned long scaled_busy_load_per_task;
5168 struct sg_lb_stats *local, *busiest;
5170 local = &sds->local_stat;
5171 busiest = &sds->busiest_stat;
5173 if (!local->sum_nr_running)
5174 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5175 else if (busiest->load_per_task > local->load_per_task)
5178 scaled_busy_load_per_task =
5179 (busiest->load_per_task * SCHED_POWER_SCALE) /
5180 busiest->group_power;
5182 if (busiest->avg_load + scaled_busy_load_per_task >=
5183 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5184 env->imbalance = busiest->load_per_task;
5189 * OK, we don't have enough imbalance to justify moving tasks,
5190 * however we may be able to increase total CPU power used by
5194 pwr_now += busiest->group_power *
5195 min(busiest->load_per_task, busiest->avg_load);
5196 pwr_now += local->group_power *
5197 min(local->load_per_task, local->avg_load);
5198 pwr_now /= SCHED_POWER_SCALE;
5200 /* Amount of load we'd subtract */
5201 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5202 busiest->group_power;
5203 if (busiest->avg_load > tmp) {
5204 pwr_move += busiest->group_power *
5205 min(busiest->load_per_task,
5206 busiest->avg_load - tmp);
5209 /* Amount of load we'd add */
5210 if (busiest->avg_load * busiest->group_power <
5211 busiest->load_per_task * SCHED_POWER_SCALE) {
5212 tmp = (busiest->avg_load * busiest->group_power) /
5215 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5218 pwr_move += local->group_power *
5219 min(local->load_per_task, local->avg_load + tmp);
5220 pwr_move /= SCHED_POWER_SCALE;
5222 /* Move if we gain throughput */
5223 if (pwr_move > pwr_now)
5224 env->imbalance = busiest->load_per_task;
5228 * calculate_imbalance - Calculate the amount of imbalance present within the
5229 * groups of a given sched_domain during load balance.
5230 * @env: load balance environment
5231 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5233 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5235 unsigned long max_pull, load_above_capacity = ~0UL;
5236 struct sg_lb_stats *local, *busiest;
5238 local = &sds->local_stat;
5239 busiest = &sds->busiest_stat;
5241 if (busiest->group_imb) {
5243 * In the group_imb case we cannot rely on group-wide averages
5244 * to ensure cpu-load equilibrium, look at wider averages. XXX
5246 busiest->load_per_task =
5247 min(busiest->load_per_task, sds->avg_load);
5251 * In the presence of smp nice balancing, certain scenarios can have
5252 * max load less than avg load(as we skip the groups at or below
5253 * its cpu_power, while calculating max_load..)
5255 if (busiest->avg_load <= sds->avg_load ||
5256 local->avg_load >= sds->avg_load) {
5258 return fix_small_imbalance(env, sds);
5261 if (!busiest->group_imb) {
5263 * Don't want to pull so many tasks that a group would go idle.
5264 * Except of course for the group_imb case, since then we might
5265 * have to drop below capacity to reach cpu-load equilibrium.
5267 load_above_capacity =
5268 (busiest->sum_nr_running - busiest->group_capacity);
5270 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5271 load_above_capacity /= busiest->group_power;
5275 * We're trying to get all the cpus to the average_load, so we don't
5276 * want to push ourselves above the average load, nor do we wish to
5277 * reduce the max loaded cpu below the average load. At the same time,
5278 * we also don't want to reduce the group load below the group capacity
5279 * (so that we can implement power-savings policies etc). Thus we look
5280 * for the minimum possible imbalance.
5282 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5284 /* How much load to actually move to equalise the imbalance */
5285 env->imbalance = min(
5286 max_pull * busiest->group_power,
5287 (sds->avg_load - local->avg_load) * local->group_power
5288 ) / SCHED_POWER_SCALE;
5291 * if *imbalance is less than the average load per runnable task
5292 * there is no guarantee that any tasks will be moved so we'll have
5293 * a think about bumping its value to force at least one task to be
5296 if (env->imbalance < busiest->load_per_task)
5297 return fix_small_imbalance(env, sds);
5300 /******* find_busiest_group() helpers end here *********************/
5303 * find_busiest_group - Returns the busiest group within the sched_domain
5304 * if there is an imbalance. If there isn't an imbalance, and
5305 * the user has opted for power-savings, it returns a group whose
5306 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5307 * such a group exists.
5309 * Also calculates the amount of weighted load which should be moved
5310 * to restore balance.
5312 * @env: The load balancing environment.
5314 * Return: - The busiest group if imbalance exists.
5315 * - If no imbalance and user has opted for power-savings balance,
5316 * return the least loaded group whose CPUs can be
5317 * put to idle by rebalancing its tasks onto our group.
5319 static struct sched_group *find_busiest_group(struct lb_env *env)
5321 struct sg_lb_stats *local, *busiest;
5322 struct sd_lb_stats sds;
5324 init_sd_lb_stats(&sds);
5327 * Compute the various statistics relavent for load balancing at
5330 update_sd_lb_stats(env, &sds);
5331 local = &sds.local_stat;
5332 busiest = &sds.busiest_stat;
5334 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5335 check_asym_packing(env, &sds))
5338 /* There is no busy sibling group to pull tasks from */
5339 if (!sds.busiest || busiest->sum_nr_running == 0)
5342 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5345 * If the busiest group is imbalanced the below checks don't
5346 * work because they assume all things are equal, which typically
5347 * isn't true due to cpus_allowed constraints and the like.
5349 if (busiest->group_imb)
5352 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5353 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5354 !busiest->group_has_capacity)
5358 * If the local group is more busy than the selected busiest group
5359 * don't try and pull any tasks.
5361 if (local->avg_load >= busiest->avg_load)
5365 * Don't pull any tasks if this group is already above the domain
5368 if (local->avg_load >= sds.avg_load)
5371 if (env->idle == CPU_IDLE) {
5373 * This cpu is idle. If the busiest group load doesn't
5374 * have more tasks than the number of available cpu's and
5375 * there is no imbalance between this and busiest group
5376 * wrt to idle cpu's, it is balanced.
5378 if ((local->idle_cpus < busiest->idle_cpus) &&
5379 busiest->sum_nr_running <= busiest->group_weight)
5383 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5384 * imbalance_pct to be conservative.
5386 if (100 * busiest->avg_load <=
5387 env->sd->imbalance_pct * local->avg_load)
5392 /* Looks like there is an imbalance. Compute it */
5393 calculate_imbalance(env, &sds);
5402 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5404 static struct rq *find_busiest_queue(struct lb_env *env,
5405 struct sched_group *group)
5407 struct rq *busiest = NULL, *rq;
5408 unsigned long busiest_load = 0, busiest_power = 1;
5411 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5412 unsigned long power = power_of(i);
5413 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5418 capacity = fix_small_capacity(env->sd, group);
5421 wl = weighted_cpuload(i);
5424 * When comparing with imbalance, use weighted_cpuload()
5425 * which is not scaled with the cpu power.
5427 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5431 * For the load comparisons with the other cpu's, consider
5432 * the weighted_cpuload() scaled with the cpu power, so that
5433 * the load can be moved away from the cpu that is potentially
5434 * running at a lower capacity.
5436 * Thus we're looking for max(wl_i / power_i), crosswise
5437 * multiplication to rid ourselves of the division works out
5438 * to: wl_i * power_j > wl_j * power_i; where j is our
5441 if (wl * busiest_power > busiest_load * power) {
5443 busiest_power = power;
5452 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5453 * so long as it is large enough.
5455 #define MAX_PINNED_INTERVAL 512
5457 /* Working cpumask for load_balance and load_balance_newidle. */
5458 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5460 static int need_active_balance(struct lb_env *env)
5462 struct sched_domain *sd = env->sd;
5464 if (env->idle == CPU_NEWLY_IDLE) {
5467 * ASYM_PACKING needs to force migrate tasks from busy but
5468 * higher numbered CPUs in order to pack all tasks in the
5469 * lowest numbered CPUs.
5471 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5475 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5478 static int active_load_balance_cpu_stop(void *data);
5480 static int should_we_balance(struct lb_env *env)
5482 struct sched_group *sg = env->sd->groups;
5483 struct cpumask *sg_cpus, *sg_mask;
5484 int cpu, balance_cpu = -1;
5487 * In the newly idle case, we will allow all the cpu's
5488 * to do the newly idle load balance.
5490 if (env->idle == CPU_NEWLY_IDLE)
5493 sg_cpus = sched_group_cpus(sg);
5494 sg_mask = sched_group_mask(sg);
5495 /* Try to find first idle cpu */
5496 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5497 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5504 if (balance_cpu == -1)
5505 balance_cpu = group_balance_cpu(sg);
5508 * First idle cpu or the first cpu(busiest) in this sched group
5509 * is eligible for doing load balancing at this and above domains.
5511 return balance_cpu == env->dst_cpu;
5515 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5516 * tasks if there is an imbalance.
5518 static int load_balance(int this_cpu, struct rq *this_rq,
5519 struct sched_domain *sd, enum cpu_idle_type idle,
5520 int *continue_balancing)
5522 int ld_moved, cur_ld_moved, active_balance = 0;
5523 struct sched_domain *sd_parent = sd->parent;
5524 struct sched_group *group;
5526 unsigned long flags;
5527 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5529 struct lb_env env = {
5531 .dst_cpu = this_cpu,
5533 .dst_grpmask = sched_group_cpus(sd->groups),
5535 .loop_break = sched_nr_migrate_break,
5540 * For NEWLY_IDLE load_balancing, we don't need to consider
5541 * other cpus in our group
5543 if (idle == CPU_NEWLY_IDLE)
5544 env.dst_grpmask = NULL;
5546 cpumask_copy(cpus, cpu_active_mask);
5548 schedstat_inc(sd, lb_count[idle]);
5551 if (!should_we_balance(&env)) {
5552 *continue_balancing = 0;
5556 group = find_busiest_group(&env);
5558 schedstat_inc(sd, lb_nobusyg[idle]);
5562 busiest = find_busiest_queue(&env, group);
5564 schedstat_inc(sd, lb_nobusyq[idle]);
5568 BUG_ON(busiest == env.dst_rq);
5570 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5573 if (busiest->nr_running > 1) {
5575 * Attempt to move tasks. If find_busiest_group has found
5576 * an imbalance but busiest->nr_running <= 1, the group is
5577 * still unbalanced. ld_moved simply stays zero, so it is
5578 * correctly treated as an imbalance.
5580 env.flags |= LBF_ALL_PINNED;
5581 env.src_cpu = busiest->cpu;
5582 env.src_rq = busiest;
5583 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5586 local_irq_save(flags);
5587 double_rq_lock(env.dst_rq, busiest);
5590 * cur_ld_moved - load moved in current iteration
5591 * ld_moved - cumulative load moved across iterations
5593 cur_ld_moved = move_tasks(&env);
5594 ld_moved += cur_ld_moved;
5595 double_rq_unlock(env.dst_rq, busiest);
5596 local_irq_restore(flags);
5599 * some other cpu did the load balance for us.
5601 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5602 resched_cpu(env.dst_cpu);
5604 if (env.flags & LBF_NEED_BREAK) {
5605 env.flags &= ~LBF_NEED_BREAK;
5610 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5611 * us and move them to an alternate dst_cpu in our sched_group
5612 * where they can run. The upper limit on how many times we
5613 * iterate on same src_cpu is dependent on number of cpus in our
5616 * This changes load balance semantics a bit on who can move
5617 * load to a given_cpu. In addition to the given_cpu itself
5618 * (or a ilb_cpu acting on its behalf where given_cpu is
5619 * nohz-idle), we now have balance_cpu in a position to move
5620 * load to given_cpu. In rare situations, this may cause
5621 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5622 * _independently_ and at _same_ time to move some load to
5623 * given_cpu) causing exceess load to be moved to given_cpu.
5624 * This however should not happen so much in practice and
5625 * moreover subsequent load balance cycles should correct the
5626 * excess load moved.
5628 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5630 /* Prevent to re-select dst_cpu via env's cpus */
5631 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5633 env.dst_rq = cpu_rq(env.new_dst_cpu);
5634 env.dst_cpu = env.new_dst_cpu;
5635 env.flags &= ~LBF_DST_PINNED;
5637 env.loop_break = sched_nr_migrate_break;
5640 * Go back to "more_balance" rather than "redo" since we
5641 * need to continue with same src_cpu.
5647 * We failed to reach balance because of affinity.
5650 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5652 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5653 *group_imbalance = 1;
5654 } else if (*group_imbalance)
5655 *group_imbalance = 0;
5658 /* All tasks on this runqueue were pinned by CPU affinity */
5659 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5660 cpumask_clear_cpu(cpu_of(busiest), cpus);
5661 if (!cpumask_empty(cpus)) {
5663 env.loop_break = sched_nr_migrate_break;
5671 schedstat_inc(sd, lb_failed[idle]);
5673 * Increment the failure counter only on periodic balance.
5674 * We do not want newidle balance, which can be very
5675 * frequent, pollute the failure counter causing
5676 * excessive cache_hot migrations and active balances.
5678 if (idle != CPU_NEWLY_IDLE)
5679 sd->nr_balance_failed++;
5681 if (need_active_balance(&env)) {
5682 raw_spin_lock_irqsave(&busiest->lock, flags);
5684 /* don't kick the active_load_balance_cpu_stop,
5685 * if the curr task on busiest cpu can't be
5688 if (!cpumask_test_cpu(this_cpu,
5689 tsk_cpus_allowed(busiest->curr))) {
5690 raw_spin_unlock_irqrestore(&busiest->lock,
5692 env.flags |= LBF_ALL_PINNED;
5693 goto out_one_pinned;
5697 * ->active_balance synchronizes accesses to
5698 * ->active_balance_work. Once set, it's cleared
5699 * only after active load balance is finished.
5701 if (!busiest->active_balance) {
5702 busiest->active_balance = 1;
5703 busiest->push_cpu = this_cpu;
5706 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5708 if (active_balance) {
5709 stop_one_cpu_nowait(cpu_of(busiest),
5710 active_load_balance_cpu_stop, busiest,
5711 &busiest->active_balance_work);
5715 * We've kicked active balancing, reset the failure
5718 sd->nr_balance_failed = sd->cache_nice_tries+1;
5721 sd->nr_balance_failed = 0;
5723 if (likely(!active_balance)) {
5724 /* We were unbalanced, so reset the balancing interval */
5725 sd->balance_interval = sd->min_interval;
5728 * If we've begun active balancing, start to back off. This
5729 * case may not be covered by the all_pinned logic if there
5730 * is only 1 task on the busy runqueue (because we don't call
5733 if (sd->balance_interval < sd->max_interval)
5734 sd->balance_interval *= 2;
5740 schedstat_inc(sd, lb_balanced[idle]);
5742 sd->nr_balance_failed = 0;
5745 /* tune up the balancing interval */
5746 if (((env.flags & LBF_ALL_PINNED) &&
5747 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5748 (sd->balance_interval < sd->max_interval))
5749 sd->balance_interval *= 2;
5757 * idle_balance is called by schedule() if this_cpu is about to become
5758 * idle. Attempts to pull tasks from other CPUs.
5760 void idle_balance(int this_cpu, struct rq *this_rq)
5762 struct sched_domain *sd;
5763 int pulled_task = 0;
5764 unsigned long next_balance = jiffies + HZ;
5767 this_rq->idle_stamp = rq_clock(this_rq);
5769 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5773 * Drop the rq->lock, but keep IRQ/preempt disabled.
5775 raw_spin_unlock(&this_rq->lock);
5777 update_blocked_averages(this_cpu);
5779 for_each_domain(this_cpu, sd) {
5780 unsigned long interval;
5781 int continue_balancing = 1;
5782 u64 t0, domain_cost;
5784 if (!(sd->flags & SD_LOAD_BALANCE))
5787 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5790 if (sd->flags & SD_BALANCE_NEWIDLE) {
5791 t0 = sched_clock_cpu(this_cpu);
5793 /* If we've pulled tasks over stop searching: */
5794 pulled_task = load_balance(this_cpu, this_rq,
5796 &continue_balancing);
5798 domain_cost = sched_clock_cpu(this_cpu) - t0;
5799 if (domain_cost > sd->max_newidle_lb_cost)
5800 sd->max_newidle_lb_cost = domain_cost;
5802 curr_cost += domain_cost;
5805 interval = msecs_to_jiffies(sd->balance_interval);
5806 if (time_after(next_balance, sd->last_balance + interval))
5807 next_balance = sd->last_balance + interval;
5809 this_rq->idle_stamp = 0;
5815 raw_spin_lock(&this_rq->lock);
5817 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5819 * We are going idle. next_balance may be set based on
5820 * a busy processor. So reset next_balance.
5822 this_rq->next_balance = next_balance;
5825 if (curr_cost > this_rq->max_idle_balance_cost)
5826 this_rq->max_idle_balance_cost = curr_cost;
5830 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5831 * running tasks off the busiest CPU onto idle CPUs. It requires at
5832 * least 1 task to be running on each physical CPU where possible, and
5833 * avoids physical / logical imbalances.
5835 static int active_load_balance_cpu_stop(void *data)
5837 struct rq *busiest_rq = data;
5838 int busiest_cpu = cpu_of(busiest_rq);
5839 int target_cpu = busiest_rq->push_cpu;
5840 struct rq *target_rq = cpu_rq(target_cpu);
5841 struct sched_domain *sd;
5843 raw_spin_lock_irq(&busiest_rq->lock);
5845 /* make sure the requested cpu hasn't gone down in the meantime */
5846 if (unlikely(busiest_cpu != smp_processor_id() ||
5847 !busiest_rq->active_balance))
5850 /* Is there any task to move? */
5851 if (busiest_rq->nr_running <= 1)
5855 * This condition is "impossible", if it occurs
5856 * we need to fix it. Originally reported by
5857 * Bjorn Helgaas on a 128-cpu setup.
5859 BUG_ON(busiest_rq == target_rq);
5861 /* move a task from busiest_rq to target_rq */
5862 double_lock_balance(busiest_rq, target_rq);
5864 /* Search for an sd spanning us and the target CPU. */
5866 for_each_domain(target_cpu, sd) {
5867 if ((sd->flags & SD_LOAD_BALANCE) &&
5868 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5873 struct lb_env env = {
5875 .dst_cpu = target_cpu,
5876 .dst_rq = target_rq,
5877 .src_cpu = busiest_rq->cpu,
5878 .src_rq = busiest_rq,
5882 schedstat_inc(sd, alb_count);
5884 if (move_one_task(&env))
5885 schedstat_inc(sd, alb_pushed);
5887 schedstat_inc(sd, alb_failed);
5890 double_unlock_balance(busiest_rq, target_rq);
5892 busiest_rq->active_balance = 0;
5893 raw_spin_unlock_irq(&busiest_rq->lock);
5897 #ifdef CONFIG_NO_HZ_COMMON
5899 * idle load balancing details
5900 * - When one of the busy CPUs notice that there may be an idle rebalancing
5901 * needed, they will kick the idle load balancer, which then does idle
5902 * load balancing for all the idle CPUs.
5905 cpumask_var_t idle_cpus_mask;
5907 unsigned long next_balance; /* in jiffy units */
5908 } nohz ____cacheline_aligned;
5910 static inline int find_new_ilb(int call_cpu)
5912 int ilb = cpumask_first(nohz.idle_cpus_mask);
5914 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5921 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5922 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5923 * CPU (if there is one).
5925 static void nohz_balancer_kick(int cpu)
5929 nohz.next_balance++;
5931 ilb_cpu = find_new_ilb(cpu);
5933 if (ilb_cpu >= nr_cpu_ids)
5936 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5939 * Use smp_send_reschedule() instead of resched_cpu().
5940 * This way we generate a sched IPI on the target cpu which
5941 * is idle. And the softirq performing nohz idle load balance
5942 * will be run before returning from the IPI.
5944 smp_send_reschedule(ilb_cpu);
5948 static inline void nohz_balance_exit_idle(int cpu)
5950 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5951 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5952 atomic_dec(&nohz.nr_cpus);
5953 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5957 static inline void set_cpu_sd_state_busy(void)
5959 struct sched_domain *sd;
5962 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5964 if (!sd || !sd->nohz_idle)
5968 for (; sd; sd = sd->parent)
5969 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5974 void set_cpu_sd_state_idle(void)
5976 struct sched_domain *sd;
5979 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5981 if (!sd || sd->nohz_idle)
5985 for (; sd; sd = sd->parent)
5986 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5992 * This routine will record that the cpu is going idle with tick stopped.
5993 * This info will be used in performing idle load balancing in the future.
5995 void nohz_balance_enter_idle(int cpu)
5998 * If this cpu is going down, then nothing needs to be done.
6000 if (!cpu_active(cpu))
6003 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6006 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6007 atomic_inc(&nohz.nr_cpus);
6008 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6011 static int sched_ilb_notifier(struct notifier_block *nfb,
6012 unsigned long action, void *hcpu)
6014 switch (action & ~CPU_TASKS_FROZEN) {
6016 nohz_balance_exit_idle(smp_processor_id());
6024 static DEFINE_SPINLOCK(balancing);
6027 * Scale the max load_balance interval with the number of CPUs in the system.
6028 * This trades load-balance latency on larger machines for less cross talk.
6030 void update_max_interval(void)
6032 max_load_balance_interval = HZ*num_online_cpus()/10;
6036 * It checks each scheduling domain to see if it is due to be balanced,
6037 * and initiates a balancing operation if so.
6039 * Balancing parameters are set up in init_sched_domains.
6041 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6043 int continue_balancing = 1;
6044 struct rq *rq = cpu_rq(cpu);
6045 unsigned long interval;
6046 struct sched_domain *sd;
6047 /* Earliest time when we have to do rebalance again */
6048 unsigned long next_balance = jiffies + 60*HZ;
6049 int update_next_balance = 0;
6050 int need_serialize, need_decay = 0;
6053 update_blocked_averages(cpu);
6056 for_each_domain(cpu, sd) {
6058 * Decay the newidle max times here because this is a regular
6059 * visit to all the domains. Decay ~1% per second.
6061 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6062 sd->max_newidle_lb_cost =
6063 (sd->max_newidle_lb_cost * 253) / 256;
6064 sd->next_decay_max_lb_cost = jiffies + HZ;
6067 max_cost += sd->max_newidle_lb_cost;
6069 if (!(sd->flags & SD_LOAD_BALANCE))
6073 * Stop the load balance at this level. There is another
6074 * CPU in our sched group which is doing load balancing more
6077 if (!continue_balancing) {
6083 interval = sd->balance_interval;
6084 if (idle != CPU_IDLE)
6085 interval *= sd->busy_factor;
6087 /* scale ms to jiffies */
6088 interval = msecs_to_jiffies(interval);
6089 interval = clamp(interval, 1UL, max_load_balance_interval);
6091 need_serialize = sd->flags & SD_SERIALIZE;
6093 if (need_serialize) {
6094 if (!spin_trylock(&balancing))
6098 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6099 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6101 * The LBF_DST_PINNED logic could have changed
6102 * env->dst_cpu, so we can't know our idle
6103 * state even if we migrated tasks. Update it.
6105 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6107 sd->last_balance = jiffies;
6110 spin_unlock(&balancing);
6112 if (time_after(next_balance, sd->last_balance + interval)) {
6113 next_balance = sd->last_balance + interval;
6114 update_next_balance = 1;
6119 * Ensure the rq-wide value also decays but keep it at a
6120 * reasonable floor to avoid funnies with rq->avg_idle.
6122 rq->max_idle_balance_cost =
6123 max((u64)sysctl_sched_migration_cost, max_cost);
6128 * next_balance will be updated only when there is a need.
6129 * When the cpu is attached to null domain for ex, it will not be
6132 if (likely(update_next_balance))
6133 rq->next_balance = next_balance;
6136 #ifdef CONFIG_NO_HZ_COMMON
6138 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6139 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6141 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6143 struct rq *this_rq = cpu_rq(this_cpu);
6147 if (idle != CPU_IDLE ||
6148 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6151 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6152 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6156 * If this cpu gets work to do, stop the load balancing
6157 * work being done for other cpus. Next load
6158 * balancing owner will pick it up.
6163 rq = cpu_rq(balance_cpu);
6165 raw_spin_lock_irq(&rq->lock);
6166 update_rq_clock(rq);
6167 update_idle_cpu_load(rq);
6168 raw_spin_unlock_irq(&rq->lock);
6170 rebalance_domains(balance_cpu, CPU_IDLE);
6172 if (time_after(this_rq->next_balance, rq->next_balance))
6173 this_rq->next_balance = rq->next_balance;
6175 nohz.next_balance = this_rq->next_balance;
6177 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6181 * Current heuristic for kicking the idle load balancer in the presence
6182 * of an idle cpu is the system.
6183 * - This rq has more than one task.
6184 * - At any scheduler domain level, this cpu's scheduler group has multiple
6185 * busy cpu's exceeding the group's power.
6186 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6187 * domain span are idle.
6189 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6191 unsigned long now = jiffies;
6192 struct sched_domain *sd;
6194 if (unlikely(idle_cpu(cpu)))
6198 * We may be recently in ticked or tickless idle mode. At the first
6199 * busy tick after returning from idle, we will update the busy stats.
6201 set_cpu_sd_state_busy();
6202 nohz_balance_exit_idle(cpu);
6205 * None are in tickless mode and hence no need for NOHZ idle load
6208 if (likely(!atomic_read(&nohz.nr_cpus)))
6211 if (time_before(now, nohz.next_balance))
6214 if (rq->nr_running >= 2)
6218 for_each_domain(cpu, sd) {
6219 struct sched_group *sg = sd->groups;
6220 struct sched_group_power *sgp = sg->sgp;
6221 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6223 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6224 goto need_kick_unlock;
6226 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6227 && (cpumask_first_and(nohz.idle_cpus_mask,
6228 sched_domain_span(sd)) < cpu))
6229 goto need_kick_unlock;
6231 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6243 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6247 * run_rebalance_domains is triggered when needed from the scheduler tick.
6248 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6250 static void run_rebalance_domains(struct softirq_action *h)
6252 int this_cpu = smp_processor_id();
6253 struct rq *this_rq = cpu_rq(this_cpu);
6254 enum cpu_idle_type idle = this_rq->idle_balance ?
6255 CPU_IDLE : CPU_NOT_IDLE;
6257 rebalance_domains(this_cpu, idle);
6260 * If this cpu has a pending nohz_balance_kick, then do the
6261 * balancing on behalf of the other idle cpus whose ticks are
6264 nohz_idle_balance(this_cpu, idle);
6267 static inline int on_null_domain(int cpu)
6269 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6273 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6275 void trigger_load_balance(struct rq *rq, int cpu)
6277 /* Don't need to rebalance while attached to NULL domain */
6278 if (time_after_eq(jiffies, rq->next_balance) &&
6279 likely(!on_null_domain(cpu)))
6280 raise_softirq(SCHED_SOFTIRQ);
6281 #ifdef CONFIG_NO_HZ_COMMON
6282 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6283 nohz_balancer_kick(cpu);
6287 static void rq_online_fair(struct rq *rq)
6292 static void rq_offline_fair(struct rq *rq)
6296 /* Ensure any throttled groups are reachable by pick_next_task */
6297 unthrottle_offline_cfs_rqs(rq);
6300 #endif /* CONFIG_SMP */
6303 * scheduler tick hitting a task of our scheduling class:
6305 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6307 struct cfs_rq *cfs_rq;
6308 struct sched_entity *se = &curr->se;
6310 for_each_sched_entity(se) {
6311 cfs_rq = cfs_rq_of(se);
6312 entity_tick(cfs_rq, se, queued);
6315 if (numabalancing_enabled)
6316 task_tick_numa(rq, curr);
6318 update_rq_runnable_avg(rq, 1);
6322 * called on fork with the child task as argument from the parent's context
6323 * - child not yet on the tasklist
6324 * - preemption disabled
6326 static void task_fork_fair(struct task_struct *p)
6328 struct cfs_rq *cfs_rq;
6329 struct sched_entity *se = &p->se, *curr;
6330 int this_cpu = smp_processor_id();
6331 struct rq *rq = this_rq();
6332 unsigned long flags;
6334 raw_spin_lock_irqsave(&rq->lock, flags);
6336 update_rq_clock(rq);
6338 cfs_rq = task_cfs_rq(current);
6339 curr = cfs_rq->curr;
6342 * Not only the cpu but also the task_group of the parent might have
6343 * been changed after parent->se.parent,cfs_rq were copied to
6344 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6345 * of child point to valid ones.
6348 __set_task_cpu(p, this_cpu);
6351 update_curr(cfs_rq);
6354 se->vruntime = curr->vruntime;
6355 place_entity(cfs_rq, se, 1);
6357 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6359 * Upon rescheduling, sched_class::put_prev_task() will place
6360 * 'current' within the tree based on its new key value.
6362 swap(curr->vruntime, se->vruntime);
6363 resched_task(rq->curr);
6366 se->vruntime -= cfs_rq->min_vruntime;
6368 raw_spin_unlock_irqrestore(&rq->lock, flags);
6372 * Priority of the task has changed. Check to see if we preempt
6376 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6382 * Reschedule if we are currently running on this runqueue and
6383 * our priority decreased, or if we are not currently running on
6384 * this runqueue and our priority is higher than the current's
6386 if (rq->curr == p) {
6387 if (p->prio > oldprio)
6388 resched_task(rq->curr);
6390 check_preempt_curr(rq, p, 0);
6393 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6395 struct sched_entity *se = &p->se;
6396 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6399 * Ensure the task's vruntime is normalized, so that when its
6400 * switched back to the fair class the enqueue_entity(.flags=0) will
6401 * do the right thing.
6403 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6404 * have normalized the vruntime, if it was !on_rq, then only when
6405 * the task is sleeping will it still have non-normalized vruntime.
6407 if (!se->on_rq && p->state != TASK_RUNNING) {
6409 * Fix up our vruntime so that the current sleep doesn't
6410 * cause 'unlimited' sleep bonus.
6412 place_entity(cfs_rq, se, 0);
6413 se->vruntime -= cfs_rq->min_vruntime;
6418 * Remove our load from contribution when we leave sched_fair
6419 * and ensure we don't carry in an old decay_count if we
6422 if (se->avg.decay_count) {
6423 __synchronize_entity_decay(se);
6424 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6430 * We switched to the sched_fair class.
6432 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6438 * We were most likely switched from sched_rt, so
6439 * kick off the schedule if running, otherwise just see
6440 * if we can still preempt the current task.
6443 resched_task(rq->curr);
6445 check_preempt_curr(rq, p, 0);
6448 /* Account for a task changing its policy or group.
6450 * This routine is mostly called to set cfs_rq->curr field when a task
6451 * migrates between groups/classes.
6453 static void set_curr_task_fair(struct rq *rq)
6455 struct sched_entity *se = &rq->curr->se;
6457 for_each_sched_entity(se) {
6458 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6460 set_next_entity(cfs_rq, se);
6461 /* ensure bandwidth has been allocated on our new cfs_rq */
6462 account_cfs_rq_runtime(cfs_rq, 0);
6466 void init_cfs_rq(struct cfs_rq *cfs_rq)
6468 cfs_rq->tasks_timeline = RB_ROOT;
6469 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6470 #ifndef CONFIG_64BIT
6471 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6474 atomic64_set(&cfs_rq->decay_counter, 1);
6475 atomic_long_set(&cfs_rq->removed_load, 0);
6479 #ifdef CONFIG_FAIR_GROUP_SCHED
6480 static void task_move_group_fair(struct task_struct *p, int on_rq)
6482 struct cfs_rq *cfs_rq;
6484 * If the task was not on the rq at the time of this cgroup movement
6485 * it must have been asleep, sleeping tasks keep their ->vruntime
6486 * absolute on their old rq until wakeup (needed for the fair sleeper
6487 * bonus in place_entity()).
6489 * If it was on the rq, we've just 'preempted' it, which does convert
6490 * ->vruntime to a relative base.
6492 * Make sure both cases convert their relative position when migrating
6493 * to another cgroup's rq. This does somewhat interfere with the
6494 * fair sleeper stuff for the first placement, but who cares.
6497 * When !on_rq, vruntime of the task has usually NOT been normalized.
6498 * But there are some cases where it has already been normalized:
6500 * - Moving a forked child which is waiting for being woken up by
6501 * wake_up_new_task().
6502 * - Moving a task which has been woken up by try_to_wake_up() and
6503 * waiting for actually being woken up by sched_ttwu_pending().
6505 * To prevent boost or penalty in the new cfs_rq caused by delta
6506 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6508 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6512 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6513 set_task_rq(p, task_cpu(p));
6515 cfs_rq = cfs_rq_of(&p->se);
6516 p->se.vruntime += cfs_rq->min_vruntime;
6519 * migrate_task_rq_fair() will have removed our previous
6520 * contribution, but we must synchronize for ongoing future
6523 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6524 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6529 void free_fair_sched_group(struct task_group *tg)
6533 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6535 for_each_possible_cpu(i) {
6537 kfree(tg->cfs_rq[i]);
6546 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6548 struct cfs_rq *cfs_rq;
6549 struct sched_entity *se;
6552 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6555 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6559 tg->shares = NICE_0_LOAD;
6561 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6563 for_each_possible_cpu(i) {
6564 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6565 GFP_KERNEL, cpu_to_node(i));
6569 se = kzalloc_node(sizeof(struct sched_entity),
6570 GFP_KERNEL, cpu_to_node(i));
6574 init_cfs_rq(cfs_rq);
6575 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6586 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6588 struct rq *rq = cpu_rq(cpu);
6589 unsigned long flags;
6592 * Only empty task groups can be destroyed; so we can speculatively
6593 * check on_list without danger of it being re-added.
6595 if (!tg->cfs_rq[cpu]->on_list)
6598 raw_spin_lock_irqsave(&rq->lock, flags);
6599 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6600 raw_spin_unlock_irqrestore(&rq->lock, flags);
6603 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6604 struct sched_entity *se, int cpu,
6605 struct sched_entity *parent)
6607 struct rq *rq = cpu_rq(cpu);
6611 init_cfs_rq_runtime(cfs_rq);
6613 tg->cfs_rq[cpu] = cfs_rq;
6616 /* se could be NULL for root_task_group */
6621 se->cfs_rq = &rq->cfs;
6623 se->cfs_rq = parent->my_q;
6626 update_load_set(&se->load, 0);
6627 se->parent = parent;
6630 static DEFINE_MUTEX(shares_mutex);
6632 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6635 unsigned long flags;
6638 * We can't change the weight of the root cgroup.
6643 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6645 mutex_lock(&shares_mutex);
6646 if (tg->shares == shares)
6649 tg->shares = shares;
6650 for_each_possible_cpu(i) {
6651 struct rq *rq = cpu_rq(i);
6652 struct sched_entity *se;
6655 /* Propagate contribution to hierarchy */
6656 raw_spin_lock_irqsave(&rq->lock, flags);
6658 /* Possible calls to update_curr() need rq clock */
6659 update_rq_clock(rq);
6660 for_each_sched_entity(se)
6661 update_cfs_shares(group_cfs_rq(se));
6662 raw_spin_unlock_irqrestore(&rq->lock, flags);
6666 mutex_unlock(&shares_mutex);
6669 #else /* CONFIG_FAIR_GROUP_SCHED */
6671 void free_fair_sched_group(struct task_group *tg) { }
6673 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6678 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6680 #endif /* CONFIG_FAIR_GROUP_SCHED */
6683 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6685 struct sched_entity *se = &task->se;
6686 unsigned int rr_interval = 0;
6689 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6692 if (rq->cfs.load.weight)
6693 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6699 * All the scheduling class methods:
6701 const struct sched_class fair_sched_class = {
6702 .next = &idle_sched_class,
6703 .enqueue_task = enqueue_task_fair,
6704 .dequeue_task = dequeue_task_fair,
6705 .yield_task = yield_task_fair,
6706 .yield_to_task = yield_to_task_fair,
6708 .check_preempt_curr = check_preempt_wakeup,
6710 .pick_next_task = pick_next_task_fair,
6711 .put_prev_task = put_prev_task_fair,
6714 .select_task_rq = select_task_rq_fair,
6715 .migrate_task_rq = migrate_task_rq_fair,
6717 .rq_online = rq_online_fair,
6718 .rq_offline = rq_offline_fair,
6720 .task_waking = task_waking_fair,
6723 .set_curr_task = set_curr_task_fair,
6724 .task_tick = task_tick_fair,
6725 .task_fork = task_fork_fair,
6727 .prio_changed = prio_changed_fair,
6728 .switched_from = switched_from_fair,
6729 .switched_to = switched_to_fair,
6731 .get_rr_interval = get_rr_interval_fair,
6733 #ifdef CONFIG_FAIR_GROUP_SCHED
6734 .task_move_group = task_move_group_fair,
6738 #ifdef CONFIG_SCHED_DEBUG
6739 void print_cfs_stats(struct seq_file *m, int cpu)
6741 struct cfs_rq *cfs_rq;
6744 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6745 print_cfs_rq(m, cpu, cfs_rq);
6750 __init void init_sched_fair_class(void)
6753 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6755 #ifdef CONFIG_NO_HZ_COMMON
6756 nohz.next_balance = jiffies;
6757 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6758 cpu_notifier(sched_ilb_notifier, 0);