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 * Shared library pages mapped by multiple processes are not
1236 * migrated as it is expected they are cache replicated. Avoid
1237 * hinting faults in read-only file-backed mappings or the vdso
1238 * as migrating the pages will be of marginal benefit.
1241 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1245 start = max(start, vma->vm_start);
1246 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1247 end = min(end, vma->vm_end);
1248 nr_pte_updates += change_prot_numa(vma, start, end);
1251 * Scan sysctl_numa_balancing_scan_size but ensure that
1252 * at least one PTE is updated so that unused virtual
1253 * address space is quickly skipped.
1256 pages -= (end - start) >> PAGE_SHIFT;
1261 } while (end != vma->vm_end);
1266 * If the whole process was scanned without updates then no NUMA
1267 * hinting faults are being recorded and scan rate should be lower.
1269 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1270 p->numa_scan_period = min(p->numa_scan_period_max,
1271 p->numa_scan_period << 1);
1273 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1274 mm->numa_next_scan = next_scan;
1278 * It is possible to reach the end of the VMA list but the last few
1279 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1280 * would find the !migratable VMA on the next scan but not reset the
1281 * scanner to the start so check it now.
1284 mm->numa_scan_offset = start;
1286 reset_ptenuma_scan(p);
1287 up_read(&mm->mmap_sem);
1291 * Drive the periodic memory faults..
1293 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1295 struct callback_head *work = &curr->numa_work;
1299 * We don't care about NUMA placement if we don't have memory.
1301 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1305 * Using runtime rather than walltime has the dual advantage that
1306 * we (mostly) drive the selection from busy threads and that the
1307 * task needs to have done some actual work before we bother with
1310 now = curr->se.sum_exec_runtime;
1311 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1313 if (now - curr->node_stamp > period) {
1314 if (!curr->node_stamp)
1315 curr->numa_scan_period = task_scan_min(curr);
1316 curr->node_stamp += period;
1318 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1319 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1320 task_work_add(curr, work, true);
1325 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1328 #endif /* CONFIG_NUMA_BALANCING */
1331 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1333 update_load_add(&cfs_rq->load, se->load.weight);
1334 if (!parent_entity(se))
1335 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1337 if (entity_is_task(se))
1338 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1340 cfs_rq->nr_running++;
1344 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1346 update_load_sub(&cfs_rq->load, se->load.weight);
1347 if (!parent_entity(se))
1348 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1349 if (entity_is_task(se))
1350 list_del_init(&se->group_node);
1351 cfs_rq->nr_running--;
1354 #ifdef CONFIG_FAIR_GROUP_SCHED
1356 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1361 * Use this CPU's actual weight instead of the last load_contribution
1362 * to gain a more accurate current total weight. See
1363 * update_cfs_rq_load_contribution().
1365 tg_weight = atomic_long_read(&tg->load_avg);
1366 tg_weight -= cfs_rq->tg_load_contrib;
1367 tg_weight += cfs_rq->load.weight;
1372 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1374 long tg_weight, load, shares;
1376 tg_weight = calc_tg_weight(tg, cfs_rq);
1377 load = cfs_rq->load.weight;
1379 shares = (tg->shares * load);
1381 shares /= tg_weight;
1383 if (shares < MIN_SHARES)
1384 shares = MIN_SHARES;
1385 if (shares > tg->shares)
1386 shares = tg->shares;
1390 # else /* CONFIG_SMP */
1391 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1395 # endif /* CONFIG_SMP */
1396 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1397 unsigned long weight)
1400 /* commit outstanding execution time */
1401 if (cfs_rq->curr == se)
1402 update_curr(cfs_rq);
1403 account_entity_dequeue(cfs_rq, se);
1406 update_load_set(&se->load, weight);
1409 account_entity_enqueue(cfs_rq, se);
1412 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1414 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1416 struct task_group *tg;
1417 struct sched_entity *se;
1421 se = tg->se[cpu_of(rq_of(cfs_rq))];
1422 if (!se || throttled_hierarchy(cfs_rq))
1425 if (likely(se->load.weight == tg->shares))
1428 shares = calc_cfs_shares(cfs_rq, tg);
1430 reweight_entity(cfs_rq_of(se), se, shares);
1432 #else /* CONFIG_FAIR_GROUP_SCHED */
1433 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1436 #endif /* CONFIG_FAIR_GROUP_SCHED */
1440 * We choose a half-life close to 1 scheduling period.
1441 * Note: The tables below are dependent on this value.
1443 #define LOAD_AVG_PERIOD 32
1444 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1445 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1447 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1448 static const u32 runnable_avg_yN_inv[] = {
1449 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1450 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1451 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1452 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1453 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1454 0x85aac367, 0x82cd8698,
1458 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1459 * over-estimates when re-combining.
1461 static const u32 runnable_avg_yN_sum[] = {
1462 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1463 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1464 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1469 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1471 static __always_inline u64 decay_load(u64 val, u64 n)
1473 unsigned int local_n;
1477 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1480 /* after bounds checking we can collapse to 32-bit */
1484 * As y^PERIOD = 1/2, we can combine
1485 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1486 * With a look-up table which covers k^n (n<PERIOD)
1488 * To achieve constant time decay_load.
1490 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1491 val >>= local_n / LOAD_AVG_PERIOD;
1492 local_n %= LOAD_AVG_PERIOD;
1495 val *= runnable_avg_yN_inv[local_n];
1496 /* We don't use SRR here since we always want to round down. */
1501 * For updates fully spanning n periods, the contribution to runnable
1502 * average will be: \Sum 1024*y^n
1504 * We can compute this reasonably efficiently by combining:
1505 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1507 static u32 __compute_runnable_contrib(u64 n)
1511 if (likely(n <= LOAD_AVG_PERIOD))
1512 return runnable_avg_yN_sum[n];
1513 else if (unlikely(n >= LOAD_AVG_MAX_N))
1514 return LOAD_AVG_MAX;
1516 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1518 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1519 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1521 n -= LOAD_AVG_PERIOD;
1522 } while (n > LOAD_AVG_PERIOD);
1524 contrib = decay_load(contrib, n);
1525 return contrib + runnable_avg_yN_sum[n];
1529 * We can represent the historical contribution to runnable average as the
1530 * coefficients of a geometric series. To do this we sub-divide our runnable
1531 * history into segments of approximately 1ms (1024us); label the segment that
1532 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1534 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1536 * (now) (~1ms ago) (~2ms ago)
1538 * Let u_i denote the fraction of p_i that the entity was runnable.
1540 * We then designate the fractions u_i as our co-efficients, yielding the
1541 * following representation of historical load:
1542 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1544 * We choose y based on the with of a reasonably scheduling period, fixing:
1547 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1548 * approximately half as much as the contribution to load within the last ms
1551 * When a period "rolls over" and we have new u_0`, multiplying the previous
1552 * sum again by y is sufficient to update:
1553 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1554 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1556 static __always_inline int __update_entity_runnable_avg(u64 now,
1557 struct sched_avg *sa,
1561 u32 runnable_contrib;
1562 int delta_w, decayed = 0;
1564 delta = now - sa->last_runnable_update;
1566 * This should only happen when time goes backwards, which it
1567 * unfortunately does during sched clock init when we swap over to TSC.
1569 if ((s64)delta < 0) {
1570 sa->last_runnable_update = now;
1575 * Use 1024ns as the unit of measurement since it's a reasonable
1576 * approximation of 1us and fast to compute.
1581 sa->last_runnable_update = now;
1583 /* delta_w is the amount already accumulated against our next period */
1584 delta_w = sa->runnable_avg_period % 1024;
1585 if (delta + delta_w >= 1024) {
1586 /* period roll-over */
1590 * Now that we know we're crossing a period boundary, figure
1591 * out how much from delta we need to complete the current
1592 * period and accrue it.
1594 delta_w = 1024 - delta_w;
1596 sa->runnable_avg_sum += delta_w;
1597 sa->runnable_avg_period += delta_w;
1601 /* Figure out how many additional periods this update spans */
1602 periods = delta / 1024;
1605 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1607 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1610 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1611 runnable_contrib = __compute_runnable_contrib(periods);
1613 sa->runnable_avg_sum += runnable_contrib;
1614 sa->runnable_avg_period += runnable_contrib;
1617 /* Remainder of delta accrued against u_0` */
1619 sa->runnable_avg_sum += delta;
1620 sa->runnable_avg_period += delta;
1625 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1626 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1628 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1629 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1631 decays -= se->avg.decay_count;
1635 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1636 se->avg.decay_count = 0;
1641 #ifdef CONFIG_FAIR_GROUP_SCHED
1642 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1645 struct task_group *tg = cfs_rq->tg;
1648 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1649 tg_contrib -= cfs_rq->tg_load_contrib;
1651 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1652 atomic_long_add(tg_contrib, &tg->load_avg);
1653 cfs_rq->tg_load_contrib += tg_contrib;
1658 * Aggregate cfs_rq runnable averages into an equivalent task_group
1659 * representation for computing load contributions.
1661 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1662 struct cfs_rq *cfs_rq)
1664 struct task_group *tg = cfs_rq->tg;
1667 /* The fraction of a cpu used by this cfs_rq */
1668 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1669 sa->runnable_avg_period + 1);
1670 contrib -= cfs_rq->tg_runnable_contrib;
1672 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1673 atomic_add(contrib, &tg->runnable_avg);
1674 cfs_rq->tg_runnable_contrib += contrib;
1678 static inline void __update_group_entity_contrib(struct sched_entity *se)
1680 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1681 struct task_group *tg = cfs_rq->tg;
1686 contrib = cfs_rq->tg_load_contrib * tg->shares;
1687 se->avg.load_avg_contrib = div_u64(contrib,
1688 atomic_long_read(&tg->load_avg) + 1);
1691 * For group entities we need to compute a correction term in the case
1692 * that they are consuming <1 cpu so that we would contribute the same
1693 * load as a task of equal weight.
1695 * Explicitly co-ordinating this measurement would be expensive, but
1696 * fortunately the sum of each cpus contribution forms a usable
1697 * lower-bound on the true value.
1699 * Consider the aggregate of 2 contributions. Either they are disjoint
1700 * (and the sum represents true value) or they are disjoint and we are
1701 * understating by the aggregate of their overlap.
1703 * Extending this to N cpus, for a given overlap, the maximum amount we
1704 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1705 * cpus that overlap for this interval and w_i is the interval width.
1707 * On a small machine; the first term is well-bounded which bounds the
1708 * total error since w_i is a subset of the period. Whereas on a
1709 * larger machine, while this first term can be larger, if w_i is the
1710 * of consequential size guaranteed to see n_i*w_i quickly converge to
1711 * our upper bound of 1-cpu.
1713 runnable_avg = atomic_read(&tg->runnable_avg);
1714 if (runnable_avg < NICE_0_LOAD) {
1715 se->avg.load_avg_contrib *= runnable_avg;
1716 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1720 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1721 int force_update) {}
1722 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1723 struct cfs_rq *cfs_rq) {}
1724 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1727 static inline void __update_task_entity_contrib(struct sched_entity *se)
1731 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1732 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1733 contrib /= (se->avg.runnable_avg_period + 1);
1734 se->avg.load_avg_contrib = scale_load(contrib);
1737 /* Compute the current contribution to load_avg by se, return any delta */
1738 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1740 long old_contrib = se->avg.load_avg_contrib;
1742 if (entity_is_task(se)) {
1743 __update_task_entity_contrib(se);
1745 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1746 __update_group_entity_contrib(se);
1749 return se->avg.load_avg_contrib - old_contrib;
1752 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1755 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1756 cfs_rq->blocked_load_avg -= load_contrib;
1758 cfs_rq->blocked_load_avg = 0;
1761 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1763 /* Update a sched_entity's runnable average */
1764 static inline void update_entity_load_avg(struct sched_entity *se,
1767 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1772 * For a group entity we need to use their owned cfs_rq_clock_task() in
1773 * case they are the parent of a throttled hierarchy.
1775 if (entity_is_task(se))
1776 now = cfs_rq_clock_task(cfs_rq);
1778 now = cfs_rq_clock_task(group_cfs_rq(se));
1780 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1783 contrib_delta = __update_entity_load_avg_contrib(se);
1789 cfs_rq->runnable_load_avg += contrib_delta;
1791 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1795 * Decay the load contributed by all blocked children and account this so that
1796 * their contribution may appropriately discounted when they wake up.
1798 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1800 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1803 decays = now - cfs_rq->last_decay;
1804 if (!decays && !force_update)
1807 if (atomic_long_read(&cfs_rq->removed_load)) {
1808 unsigned long removed_load;
1809 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1810 subtract_blocked_load_contrib(cfs_rq, removed_load);
1814 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1816 atomic64_add(decays, &cfs_rq->decay_counter);
1817 cfs_rq->last_decay = now;
1820 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1823 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1825 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1826 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1829 /* Add the load generated by se into cfs_rq's child load-average */
1830 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1831 struct sched_entity *se,
1835 * We track migrations using entity decay_count <= 0, on a wake-up
1836 * migration we use a negative decay count to track the remote decays
1837 * accumulated while sleeping.
1839 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1840 * are seen by enqueue_entity_load_avg() as a migration with an already
1841 * constructed load_avg_contrib.
1843 if (unlikely(se->avg.decay_count <= 0)) {
1844 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1845 if (se->avg.decay_count) {
1847 * In a wake-up migration we have to approximate the
1848 * time sleeping. This is because we can't synchronize
1849 * clock_task between the two cpus, and it is not
1850 * guaranteed to be read-safe. Instead, we can
1851 * approximate this using our carried decays, which are
1852 * explicitly atomically readable.
1854 se->avg.last_runnable_update -= (-se->avg.decay_count)
1856 update_entity_load_avg(se, 0);
1857 /* Indicate that we're now synchronized and on-rq */
1858 se->avg.decay_count = 0;
1863 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1864 * would have made count negative); we must be careful to avoid
1865 * double-accounting blocked time after synchronizing decays.
1867 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1871 /* migrated tasks did not contribute to our blocked load */
1873 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1874 update_entity_load_avg(se, 0);
1877 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1878 /* we force update consideration on load-balancer moves */
1879 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1883 * Remove se's load from this cfs_rq child load-average, if the entity is
1884 * transitioning to a blocked state we track its projected decay using
1887 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1888 struct sched_entity *se,
1891 update_entity_load_avg(se, 1);
1892 /* we force update consideration on load-balancer moves */
1893 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1895 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1897 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1898 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1899 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1903 * Update the rq's load with the elapsed running time before entering
1904 * idle. if the last scheduled task is not a CFS task, idle_enter will
1905 * be the only way to update the runnable statistic.
1907 void idle_enter_fair(struct rq *this_rq)
1909 update_rq_runnable_avg(this_rq, 1);
1913 * Update the rq's load with the elapsed idle time before a task is
1914 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1915 * be the only way to update the runnable statistic.
1917 void idle_exit_fair(struct rq *this_rq)
1919 update_rq_runnable_avg(this_rq, 0);
1923 static inline void update_entity_load_avg(struct sched_entity *se,
1924 int update_cfs_rq) {}
1925 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1926 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1927 struct sched_entity *se,
1929 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1930 struct sched_entity *se,
1932 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1933 int force_update) {}
1936 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1938 #ifdef CONFIG_SCHEDSTATS
1939 struct task_struct *tsk = NULL;
1941 if (entity_is_task(se))
1944 if (se->statistics.sleep_start) {
1945 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1950 if (unlikely(delta > se->statistics.sleep_max))
1951 se->statistics.sleep_max = delta;
1953 se->statistics.sleep_start = 0;
1954 se->statistics.sum_sleep_runtime += delta;
1957 account_scheduler_latency(tsk, delta >> 10, 1);
1958 trace_sched_stat_sleep(tsk, delta);
1961 if (se->statistics.block_start) {
1962 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1967 if (unlikely(delta > se->statistics.block_max))
1968 se->statistics.block_max = delta;
1970 se->statistics.block_start = 0;
1971 se->statistics.sum_sleep_runtime += delta;
1974 if (tsk->in_iowait) {
1975 se->statistics.iowait_sum += delta;
1976 se->statistics.iowait_count++;
1977 trace_sched_stat_iowait(tsk, delta);
1980 trace_sched_stat_blocked(tsk, delta);
1983 * Blocking time is in units of nanosecs, so shift by
1984 * 20 to get a milliseconds-range estimation of the
1985 * amount of time that the task spent sleeping:
1987 if (unlikely(prof_on == SLEEP_PROFILING)) {
1988 profile_hits(SLEEP_PROFILING,
1989 (void *)get_wchan(tsk),
1992 account_scheduler_latency(tsk, delta >> 10, 0);
1998 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2000 #ifdef CONFIG_SCHED_DEBUG
2001 s64 d = se->vruntime - cfs_rq->min_vruntime;
2006 if (d > 3*sysctl_sched_latency)
2007 schedstat_inc(cfs_rq, nr_spread_over);
2012 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2014 u64 vruntime = cfs_rq->min_vruntime;
2017 * The 'current' period is already promised to the current tasks,
2018 * however the extra weight of the new task will slow them down a
2019 * little, place the new task so that it fits in the slot that
2020 * stays open at the end.
2022 if (initial && sched_feat(START_DEBIT))
2023 vruntime += sched_vslice(cfs_rq, se);
2025 /* sleeps up to a single latency don't count. */
2027 unsigned long thresh = sysctl_sched_latency;
2030 * Halve their sleep time's effect, to allow
2031 * for a gentler effect of sleepers:
2033 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2039 /* ensure we never gain time by being placed backwards. */
2040 se->vruntime = max_vruntime(se->vruntime, vruntime);
2043 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2046 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2049 * Update the normalized vruntime before updating min_vruntime
2050 * through calling update_curr().
2052 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2053 se->vruntime += cfs_rq->min_vruntime;
2056 * Update run-time statistics of the 'current'.
2058 update_curr(cfs_rq);
2059 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2060 account_entity_enqueue(cfs_rq, se);
2061 update_cfs_shares(cfs_rq);
2063 if (flags & ENQUEUE_WAKEUP) {
2064 place_entity(cfs_rq, se, 0);
2065 enqueue_sleeper(cfs_rq, se);
2068 update_stats_enqueue(cfs_rq, se);
2069 check_spread(cfs_rq, se);
2070 if (se != cfs_rq->curr)
2071 __enqueue_entity(cfs_rq, se);
2074 if (cfs_rq->nr_running == 1) {
2075 list_add_leaf_cfs_rq(cfs_rq);
2076 check_enqueue_throttle(cfs_rq);
2080 static void __clear_buddies_last(struct sched_entity *se)
2082 for_each_sched_entity(se) {
2083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2084 if (cfs_rq->last == se)
2085 cfs_rq->last = NULL;
2091 static void __clear_buddies_next(struct sched_entity *se)
2093 for_each_sched_entity(se) {
2094 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2095 if (cfs_rq->next == se)
2096 cfs_rq->next = NULL;
2102 static void __clear_buddies_skip(struct sched_entity *se)
2104 for_each_sched_entity(se) {
2105 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2106 if (cfs_rq->skip == se)
2107 cfs_rq->skip = NULL;
2113 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2115 if (cfs_rq->last == se)
2116 __clear_buddies_last(se);
2118 if (cfs_rq->next == se)
2119 __clear_buddies_next(se);
2121 if (cfs_rq->skip == se)
2122 __clear_buddies_skip(se);
2125 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2128 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2131 * Update run-time statistics of the 'current'.
2133 update_curr(cfs_rq);
2134 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2136 update_stats_dequeue(cfs_rq, se);
2137 if (flags & DEQUEUE_SLEEP) {
2138 #ifdef CONFIG_SCHEDSTATS
2139 if (entity_is_task(se)) {
2140 struct task_struct *tsk = task_of(se);
2142 if (tsk->state & TASK_INTERRUPTIBLE)
2143 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2144 if (tsk->state & TASK_UNINTERRUPTIBLE)
2145 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2150 clear_buddies(cfs_rq, se);
2152 if (se != cfs_rq->curr)
2153 __dequeue_entity(cfs_rq, se);
2155 account_entity_dequeue(cfs_rq, se);
2158 * Normalize the entity after updating the min_vruntime because the
2159 * update can refer to the ->curr item and we need to reflect this
2160 * movement in our normalized position.
2162 if (!(flags & DEQUEUE_SLEEP))
2163 se->vruntime -= cfs_rq->min_vruntime;
2165 /* return excess runtime on last dequeue */
2166 return_cfs_rq_runtime(cfs_rq);
2168 update_min_vruntime(cfs_rq);
2169 update_cfs_shares(cfs_rq);
2173 * Preempt the current task with a newly woken task if needed:
2176 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2178 unsigned long ideal_runtime, delta_exec;
2179 struct sched_entity *se;
2182 ideal_runtime = sched_slice(cfs_rq, curr);
2183 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2184 if (delta_exec > ideal_runtime) {
2185 resched_task(rq_of(cfs_rq)->curr);
2187 * The current task ran long enough, ensure it doesn't get
2188 * re-elected due to buddy favours.
2190 clear_buddies(cfs_rq, curr);
2195 * Ensure that a task that missed wakeup preemption by a
2196 * narrow margin doesn't have to wait for a full slice.
2197 * This also mitigates buddy induced latencies under load.
2199 if (delta_exec < sysctl_sched_min_granularity)
2202 se = __pick_first_entity(cfs_rq);
2203 delta = curr->vruntime - se->vruntime;
2208 if (delta > ideal_runtime)
2209 resched_task(rq_of(cfs_rq)->curr);
2213 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2215 /* 'current' is not kept within the tree. */
2218 * Any task has to be enqueued before it get to execute on
2219 * a CPU. So account for the time it spent waiting on the
2222 update_stats_wait_end(cfs_rq, se);
2223 __dequeue_entity(cfs_rq, se);
2226 update_stats_curr_start(cfs_rq, se);
2228 #ifdef CONFIG_SCHEDSTATS
2230 * Track our maximum slice length, if the CPU's load is at
2231 * least twice that of our own weight (i.e. dont track it
2232 * when there are only lesser-weight tasks around):
2234 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2235 se->statistics.slice_max = max(se->statistics.slice_max,
2236 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2239 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2243 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2246 * Pick the next process, keeping these things in mind, in this order:
2247 * 1) keep things fair between processes/task groups
2248 * 2) pick the "next" process, since someone really wants that to run
2249 * 3) pick the "last" process, for cache locality
2250 * 4) do not run the "skip" process, if something else is available
2252 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2254 struct sched_entity *se = __pick_first_entity(cfs_rq);
2255 struct sched_entity *left = se;
2258 * Avoid running the skip buddy, if running something else can
2259 * be done without getting too unfair.
2261 if (cfs_rq->skip == se) {
2262 struct sched_entity *second = __pick_next_entity(se);
2263 if (second && wakeup_preempt_entity(second, left) < 1)
2268 * Prefer last buddy, try to return the CPU to a preempted task.
2270 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2274 * Someone really wants this to run. If it's not unfair, run it.
2276 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2279 clear_buddies(cfs_rq, se);
2284 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2286 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2289 * If still on the runqueue then deactivate_task()
2290 * was not called and update_curr() has to be done:
2293 update_curr(cfs_rq);
2295 /* throttle cfs_rqs exceeding runtime */
2296 check_cfs_rq_runtime(cfs_rq);
2298 check_spread(cfs_rq, prev);
2300 update_stats_wait_start(cfs_rq, prev);
2301 /* Put 'current' back into the tree. */
2302 __enqueue_entity(cfs_rq, prev);
2303 /* in !on_rq case, update occurred at dequeue */
2304 update_entity_load_avg(prev, 1);
2306 cfs_rq->curr = NULL;
2310 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2313 * Update run-time statistics of the 'current'.
2315 update_curr(cfs_rq);
2318 * Ensure that runnable average is periodically updated.
2320 update_entity_load_avg(curr, 1);
2321 update_cfs_rq_blocked_load(cfs_rq, 1);
2322 update_cfs_shares(cfs_rq);
2324 #ifdef CONFIG_SCHED_HRTICK
2326 * queued ticks are scheduled to match the slice, so don't bother
2327 * validating it and just reschedule.
2330 resched_task(rq_of(cfs_rq)->curr);
2334 * don't let the period tick interfere with the hrtick preemption
2336 if (!sched_feat(DOUBLE_TICK) &&
2337 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2341 if (cfs_rq->nr_running > 1)
2342 check_preempt_tick(cfs_rq, curr);
2346 /**************************************************
2347 * CFS bandwidth control machinery
2350 #ifdef CONFIG_CFS_BANDWIDTH
2352 #ifdef HAVE_JUMP_LABEL
2353 static struct static_key __cfs_bandwidth_used;
2355 static inline bool cfs_bandwidth_used(void)
2357 return static_key_false(&__cfs_bandwidth_used);
2360 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2362 /* only need to count groups transitioning between enabled/!enabled */
2363 if (enabled && !was_enabled)
2364 static_key_slow_inc(&__cfs_bandwidth_used);
2365 else if (!enabled && was_enabled)
2366 static_key_slow_dec(&__cfs_bandwidth_used);
2368 #else /* HAVE_JUMP_LABEL */
2369 static bool cfs_bandwidth_used(void)
2374 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2375 #endif /* HAVE_JUMP_LABEL */
2378 * default period for cfs group bandwidth.
2379 * default: 0.1s, units: nanoseconds
2381 static inline u64 default_cfs_period(void)
2383 return 100000000ULL;
2386 static inline u64 sched_cfs_bandwidth_slice(void)
2388 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2392 * Replenish runtime according to assigned quota and update expiration time.
2393 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2394 * additional synchronization around rq->lock.
2396 * requires cfs_b->lock
2398 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2402 if (cfs_b->quota == RUNTIME_INF)
2405 now = sched_clock_cpu(smp_processor_id());
2406 cfs_b->runtime = cfs_b->quota;
2407 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2410 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2412 return &tg->cfs_bandwidth;
2415 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2416 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2418 if (unlikely(cfs_rq->throttle_count))
2419 return cfs_rq->throttled_clock_task;
2421 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2424 /* returns 0 on failure to allocate runtime */
2425 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2427 struct task_group *tg = cfs_rq->tg;
2428 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2429 u64 amount = 0, min_amount, expires;
2431 /* note: this is a positive sum as runtime_remaining <= 0 */
2432 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2434 raw_spin_lock(&cfs_b->lock);
2435 if (cfs_b->quota == RUNTIME_INF)
2436 amount = min_amount;
2439 * If the bandwidth pool has become inactive, then at least one
2440 * period must have elapsed since the last consumption.
2441 * Refresh the global state and ensure bandwidth timer becomes
2444 if (!cfs_b->timer_active) {
2445 __refill_cfs_bandwidth_runtime(cfs_b);
2446 __start_cfs_bandwidth(cfs_b);
2449 if (cfs_b->runtime > 0) {
2450 amount = min(cfs_b->runtime, min_amount);
2451 cfs_b->runtime -= amount;
2455 expires = cfs_b->runtime_expires;
2456 raw_spin_unlock(&cfs_b->lock);
2458 cfs_rq->runtime_remaining += amount;
2460 * we may have advanced our local expiration to account for allowed
2461 * spread between our sched_clock and the one on which runtime was
2464 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2465 cfs_rq->runtime_expires = expires;
2467 return cfs_rq->runtime_remaining > 0;
2471 * Note: This depends on the synchronization provided by sched_clock and the
2472 * fact that rq->clock snapshots this value.
2474 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2476 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2478 /* if the deadline is ahead of our clock, nothing to do */
2479 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2482 if (cfs_rq->runtime_remaining < 0)
2486 * If the local deadline has passed we have to consider the
2487 * possibility that our sched_clock is 'fast' and the global deadline
2488 * has not truly expired.
2490 * Fortunately we can check determine whether this the case by checking
2491 * whether the global deadline has advanced.
2494 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2495 /* extend local deadline, drift is bounded above by 2 ticks */
2496 cfs_rq->runtime_expires += TICK_NSEC;
2498 /* global deadline is ahead, expiration has passed */
2499 cfs_rq->runtime_remaining = 0;
2503 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2504 unsigned long delta_exec)
2506 /* dock delta_exec before expiring quota (as it could span periods) */
2507 cfs_rq->runtime_remaining -= delta_exec;
2508 expire_cfs_rq_runtime(cfs_rq);
2510 if (likely(cfs_rq->runtime_remaining > 0))
2514 * if we're unable to extend our runtime we resched so that the active
2515 * hierarchy can be throttled
2517 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2518 resched_task(rq_of(cfs_rq)->curr);
2521 static __always_inline
2522 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2524 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2527 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2530 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2532 return cfs_bandwidth_used() && cfs_rq->throttled;
2535 /* check whether cfs_rq, or any parent, is throttled */
2536 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2538 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2542 * Ensure that neither of the group entities corresponding to src_cpu or
2543 * dest_cpu are members of a throttled hierarchy when performing group
2544 * load-balance operations.
2546 static inline int throttled_lb_pair(struct task_group *tg,
2547 int src_cpu, int dest_cpu)
2549 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2551 src_cfs_rq = tg->cfs_rq[src_cpu];
2552 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2554 return throttled_hierarchy(src_cfs_rq) ||
2555 throttled_hierarchy(dest_cfs_rq);
2558 /* updated child weight may affect parent so we have to do this bottom up */
2559 static int tg_unthrottle_up(struct task_group *tg, void *data)
2561 struct rq *rq = data;
2562 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2564 cfs_rq->throttle_count--;
2566 if (!cfs_rq->throttle_count) {
2567 /* adjust cfs_rq_clock_task() */
2568 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2569 cfs_rq->throttled_clock_task;
2576 static int tg_throttle_down(struct task_group *tg, void *data)
2578 struct rq *rq = data;
2579 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2581 /* group is entering throttled state, stop time */
2582 if (!cfs_rq->throttle_count)
2583 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2584 cfs_rq->throttle_count++;
2589 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2591 struct rq *rq = rq_of(cfs_rq);
2592 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2593 struct sched_entity *se;
2594 long task_delta, dequeue = 1;
2596 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2598 /* freeze hierarchy runnable averages while throttled */
2600 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2603 task_delta = cfs_rq->h_nr_running;
2604 for_each_sched_entity(se) {
2605 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2606 /* throttled entity or throttle-on-deactivate */
2611 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2612 qcfs_rq->h_nr_running -= task_delta;
2614 if (qcfs_rq->load.weight)
2619 rq->nr_running -= task_delta;
2621 cfs_rq->throttled = 1;
2622 cfs_rq->throttled_clock = rq_clock(rq);
2623 raw_spin_lock(&cfs_b->lock);
2624 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2625 raw_spin_unlock(&cfs_b->lock);
2628 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2630 struct rq *rq = rq_of(cfs_rq);
2631 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2632 struct sched_entity *se;
2636 se = cfs_rq->tg->se[cpu_of(rq)];
2638 cfs_rq->throttled = 0;
2640 update_rq_clock(rq);
2642 raw_spin_lock(&cfs_b->lock);
2643 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2644 list_del_rcu(&cfs_rq->throttled_list);
2645 raw_spin_unlock(&cfs_b->lock);
2647 /* update hierarchical throttle state */
2648 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2650 if (!cfs_rq->load.weight)
2653 task_delta = cfs_rq->h_nr_running;
2654 for_each_sched_entity(se) {
2658 cfs_rq = cfs_rq_of(se);
2660 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2661 cfs_rq->h_nr_running += task_delta;
2663 if (cfs_rq_throttled(cfs_rq))
2668 rq->nr_running += task_delta;
2670 /* determine whether we need to wake up potentially idle cpu */
2671 if (rq->curr == rq->idle && rq->cfs.nr_running)
2672 resched_task(rq->curr);
2675 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2676 u64 remaining, u64 expires)
2678 struct cfs_rq *cfs_rq;
2679 u64 runtime = remaining;
2682 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2684 struct rq *rq = rq_of(cfs_rq);
2686 raw_spin_lock(&rq->lock);
2687 if (!cfs_rq_throttled(cfs_rq))
2690 runtime = -cfs_rq->runtime_remaining + 1;
2691 if (runtime > remaining)
2692 runtime = remaining;
2693 remaining -= runtime;
2695 cfs_rq->runtime_remaining += runtime;
2696 cfs_rq->runtime_expires = expires;
2698 /* we check whether we're throttled above */
2699 if (cfs_rq->runtime_remaining > 0)
2700 unthrottle_cfs_rq(cfs_rq);
2703 raw_spin_unlock(&rq->lock);
2714 * Responsible for refilling a task_group's bandwidth and unthrottling its
2715 * cfs_rqs as appropriate. If there has been no activity within the last
2716 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2717 * used to track this state.
2719 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2721 u64 runtime, runtime_expires;
2722 int idle = 1, throttled;
2724 raw_spin_lock(&cfs_b->lock);
2725 /* no need to continue the timer with no bandwidth constraint */
2726 if (cfs_b->quota == RUNTIME_INF)
2729 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2730 /* idle depends on !throttled (for the case of a large deficit) */
2731 idle = cfs_b->idle && !throttled;
2732 cfs_b->nr_periods += overrun;
2734 /* if we're going inactive then everything else can be deferred */
2738 __refill_cfs_bandwidth_runtime(cfs_b);
2741 /* mark as potentially idle for the upcoming period */
2746 /* account preceding periods in which throttling occurred */
2747 cfs_b->nr_throttled += overrun;
2750 * There are throttled entities so we must first use the new bandwidth
2751 * to unthrottle them before making it generally available. This
2752 * ensures that all existing debts will be paid before a new cfs_rq is
2755 runtime = cfs_b->runtime;
2756 runtime_expires = cfs_b->runtime_expires;
2760 * This check is repeated as we are holding onto the new bandwidth
2761 * while we unthrottle. This can potentially race with an unthrottled
2762 * group trying to acquire new bandwidth from the global pool.
2764 while (throttled && runtime > 0) {
2765 raw_spin_unlock(&cfs_b->lock);
2766 /* we can't nest cfs_b->lock while distributing bandwidth */
2767 runtime = distribute_cfs_runtime(cfs_b, runtime,
2769 raw_spin_lock(&cfs_b->lock);
2771 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2774 /* return (any) remaining runtime */
2775 cfs_b->runtime = runtime;
2777 * While we are ensured activity in the period following an
2778 * unthrottle, this also covers the case in which the new bandwidth is
2779 * insufficient to cover the existing bandwidth deficit. (Forcing the
2780 * timer to remain active while there are any throttled entities.)
2785 cfs_b->timer_active = 0;
2786 raw_spin_unlock(&cfs_b->lock);
2791 /* a cfs_rq won't donate quota below this amount */
2792 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2793 /* minimum remaining period time to redistribute slack quota */
2794 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2795 /* how long we wait to gather additional slack before distributing */
2796 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2798 /* are we near the end of the current quota period? */
2799 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2801 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2804 /* if the call-back is running a quota refresh is already occurring */
2805 if (hrtimer_callback_running(refresh_timer))
2808 /* is a quota refresh about to occur? */
2809 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2810 if (remaining < min_expire)
2816 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2818 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2820 /* if there's a quota refresh soon don't bother with slack */
2821 if (runtime_refresh_within(cfs_b, min_left))
2824 start_bandwidth_timer(&cfs_b->slack_timer,
2825 ns_to_ktime(cfs_bandwidth_slack_period));
2828 /* we know any runtime found here is valid as update_curr() precedes return */
2829 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2831 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2832 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2834 if (slack_runtime <= 0)
2837 raw_spin_lock(&cfs_b->lock);
2838 if (cfs_b->quota != RUNTIME_INF &&
2839 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2840 cfs_b->runtime += slack_runtime;
2842 /* we are under rq->lock, defer unthrottling using a timer */
2843 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2844 !list_empty(&cfs_b->throttled_cfs_rq))
2845 start_cfs_slack_bandwidth(cfs_b);
2847 raw_spin_unlock(&cfs_b->lock);
2849 /* even if it's not valid for return we don't want to try again */
2850 cfs_rq->runtime_remaining -= slack_runtime;
2853 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2855 if (!cfs_bandwidth_used())
2858 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2861 __return_cfs_rq_runtime(cfs_rq);
2865 * This is done with a timer (instead of inline with bandwidth return) since
2866 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2868 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2870 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2873 /* confirm we're still not at a refresh boundary */
2874 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2877 raw_spin_lock(&cfs_b->lock);
2878 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2879 runtime = cfs_b->runtime;
2882 expires = cfs_b->runtime_expires;
2883 raw_spin_unlock(&cfs_b->lock);
2888 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2890 raw_spin_lock(&cfs_b->lock);
2891 if (expires == cfs_b->runtime_expires)
2892 cfs_b->runtime = runtime;
2893 raw_spin_unlock(&cfs_b->lock);
2897 * When a group wakes up we want to make sure that its quota is not already
2898 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2899 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2901 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2903 if (!cfs_bandwidth_used())
2906 /* an active group must be handled by the update_curr()->put() path */
2907 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2910 /* ensure the group is not already throttled */
2911 if (cfs_rq_throttled(cfs_rq))
2914 /* update runtime allocation */
2915 account_cfs_rq_runtime(cfs_rq, 0);
2916 if (cfs_rq->runtime_remaining <= 0)
2917 throttle_cfs_rq(cfs_rq);
2920 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2921 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2923 if (!cfs_bandwidth_used())
2926 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2930 * it's possible for a throttled entity to be forced into a running
2931 * state (e.g. set_curr_task), in this case we're finished.
2933 if (cfs_rq_throttled(cfs_rq))
2936 throttle_cfs_rq(cfs_rq);
2939 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2941 struct cfs_bandwidth *cfs_b =
2942 container_of(timer, struct cfs_bandwidth, slack_timer);
2943 do_sched_cfs_slack_timer(cfs_b);
2945 return HRTIMER_NORESTART;
2948 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2950 struct cfs_bandwidth *cfs_b =
2951 container_of(timer, struct cfs_bandwidth, period_timer);
2957 now = hrtimer_cb_get_time(timer);
2958 overrun = hrtimer_forward(timer, now, cfs_b->period);
2963 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2966 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2969 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2971 raw_spin_lock_init(&cfs_b->lock);
2973 cfs_b->quota = RUNTIME_INF;
2974 cfs_b->period = ns_to_ktime(default_cfs_period());
2976 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2977 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2978 cfs_b->period_timer.function = sched_cfs_period_timer;
2979 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2980 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2983 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2985 cfs_rq->runtime_enabled = 0;
2986 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2989 /* requires cfs_b->lock, may release to reprogram timer */
2990 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2993 * The timer may be active because we're trying to set a new bandwidth
2994 * period or because we're racing with the tear-down path
2995 * (timer_active==0 becomes visible before the hrtimer call-back
2996 * terminates). In either case we ensure that it's re-programmed
2998 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2999 raw_spin_unlock(&cfs_b->lock);
3000 /* ensure cfs_b->lock is available while we wait */
3001 hrtimer_cancel(&cfs_b->period_timer);
3003 raw_spin_lock(&cfs_b->lock);
3004 /* if someone else restarted the timer then we're done */
3005 if (cfs_b->timer_active)
3009 cfs_b->timer_active = 1;
3010 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3013 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3015 hrtimer_cancel(&cfs_b->period_timer);
3016 hrtimer_cancel(&cfs_b->slack_timer);
3019 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3021 struct cfs_rq *cfs_rq;
3023 for_each_leaf_cfs_rq(rq, cfs_rq) {
3024 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3026 if (!cfs_rq->runtime_enabled)
3030 * clock_task is not advancing so we just need to make sure
3031 * there's some valid quota amount
3033 cfs_rq->runtime_remaining = cfs_b->quota;
3034 if (cfs_rq_throttled(cfs_rq))
3035 unthrottle_cfs_rq(cfs_rq);
3039 #else /* CONFIG_CFS_BANDWIDTH */
3040 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3042 return rq_clock_task(rq_of(cfs_rq));
3045 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3046 unsigned long delta_exec) {}
3047 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3048 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3049 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3051 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3056 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3061 static inline int throttled_lb_pair(struct task_group *tg,
3062 int src_cpu, int dest_cpu)
3067 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3069 #ifdef CONFIG_FAIR_GROUP_SCHED
3070 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3073 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3077 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3078 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3080 #endif /* CONFIG_CFS_BANDWIDTH */
3082 /**************************************************
3083 * CFS operations on tasks:
3086 #ifdef CONFIG_SCHED_HRTICK
3087 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3089 struct sched_entity *se = &p->se;
3090 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3092 WARN_ON(task_rq(p) != rq);
3094 if (cfs_rq->nr_running > 1) {
3095 u64 slice = sched_slice(cfs_rq, se);
3096 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3097 s64 delta = slice - ran;
3106 * Don't schedule slices shorter than 10000ns, that just
3107 * doesn't make sense. Rely on vruntime for fairness.
3110 delta = max_t(s64, 10000LL, delta);
3112 hrtick_start(rq, delta);
3117 * called from enqueue/dequeue and updates the hrtick when the
3118 * current task is from our class and nr_running is low enough
3121 static void hrtick_update(struct rq *rq)
3123 struct task_struct *curr = rq->curr;
3125 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3128 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3129 hrtick_start_fair(rq, curr);
3131 #else /* !CONFIG_SCHED_HRTICK */
3133 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3137 static inline void hrtick_update(struct rq *rq)
3143 * The enqueue_task method is called before nr_running is
3144 * increased. Here we update the fair scheduling stats and
3145 * then put the task into the rbtree:
3148 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3150 struct cfs_rq *cfs_rq;
3151 struct sched_entity *se = &p->se;
3153 for_each_sched_entity(se) {
3156 cfs_rq = cfs_rq_of(se);
3157 enqueue_entity(cfs_rq, se, flags);
3160 * end evaluation on encountering a throttled cfs_rq
3162 * note: in the case of encountering a throttled cfs_rq we will
3163 * post the final h_nr_running increment below.
3165 if (cfs_rq_throttled(cfs_rq))
3167 cfs_rq->h_nr_running++;
3169 flags = ENQUEUE_WAKEUP;
3172 for_each_sched_entity(se) {
3173 cfs_rq = cfs_rq_of(se);
3174 cfs_rq->h_nr_running++;
3176 if (cfs_rq_throttled(cfs_rq))
3179 update_cfs_shares(cfs_rq);
3180 update_entity_load_avg(se, 1);
3184 update_rq_runnable_avg(rq, rq->nr_running);
3190 static void set_next_buddy(struct sched_entity *se);
3193 * The dequeue_task method is called before nr_running is
3194 * decreased. We remove the task from the rbtree and
3195 * update the fair scheduling stats:
3197 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3199 struct cfs_rq *cfs_rq;
3200 struct sched_entity *se = &p->se;
3201 int task_sleep = flags & DEQUEUE_SLEEP;
3203 for_each_sched_entity(se) {
3204 cfs_rq = cfs_rq_of(se);
3205 dequeue_entity(cfs_rq, se, flags);
3208 * end evaluation on encountering a throttled cfs_rq
3210 * note: in the case of encountering a throttled cfs_rq we will
3211 * post the final h_nr_running decrement below.
3213 if (cfs_rq_throttled(cfs_rq))
3215 cfs_rq->h_nr_running--;
3217 /* Don't dequeue parent if it has other entities besides us */
3218 if (cfs_rq->load.weight) {
3220 * Bias pick_next to pick a task from this cfs_rq, as
3221 * p is sleeping when it is within its sched_slice.
3223 if (task_sleep && parent_entity(se))
3224 set_next_buddy(parent_entity(se));
3226 /* avoid re-evaluating load for this entity */
3227 se = parent_entity(se);
3230 flags |= DEQUEUE_SLEEP;
3233 for_each_sched_entity(se) {
3234 cfs_rq = cfs_rq_of(se);
3235 cfs_rq->h_nr_running--;
3237 if (cfs_rq_throttled(cfs_rq))
3240 update_cfs_shares(cfs_rq);
3241 update_entity_load_avg(se, 1);
3246 update_rq_runnable_avg(rq, 1);
3252 /* Used instead of source_load when we know the type == 0 */
3253 static unsigned long weighted_cpuload(const int cpu)
3255 return cpu_rq(cpu)->cfs.runnable_load_avg;
3259 * Return a low guess at the load of a migration-source cpu weighted
3260 * according to the scheduling class and "nice" value.
3262 * We want to under-estimate the load of migration sources, to
3263 * balance conservatively.
3265 static unsigned long source_load(int cpu, int type)
3267 struct rq *rq = cpu_rq(cpu);
3268 unsigned long total = weighted_cpuload(cpu);
3270 if (type == 0 || !sched_feat(LB_BIAS))
3273 return min(rq->cpu_load[type-1], total);
3277 * Return a high guess at the load of a migration-target cpu weighted
3278 * according to the scheduling class and "nice" value.
3280 static unsigned long target_load(int cpu, int type)
3282 struct rq *rq = cpu_rq(cpu);
3283 unsigned long total = weighted_cpuload(cpu);
3285 if (type == 0 || !sched_feat(LB_BIAS))
3288 return max(rq->cpu_load[type-1], total);
3291 static unsigned long power_of(int cpu)
3293 return cpu_rq(cpu)->cpu_power;
3296 static unsigned long cpu_avg_load_per_task(int cpu)
3298 struct rq *rq = cpu_rq(cpu);
3299 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3300 unsigned long load_avg = rq->cfs.runnable_load_avg;
3303 return load_avg / nr_running;
3308 static void record_wakee(struct task_struct *p)
3311 * Rough decay (wiping) for cost saving, don't worry
3312 * about the boundary, really active task won't care
3315 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3316 current->wakee_flips = 0;
3317 current->wakee_flip_decay_ts = jiffies;
3320 if (current->last_wakee != p) {
3321 current->last_wakee = p;
3322 current->wakee_flips++;
3326 static void task_waking_fair(struct task_struct *p)
3328 struct sched_entity *se = &p->se;
3329 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3332 #ifndef CONFIG_64BIT
3333 u64 min_vruntime_copy;
3336 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3338 min_vruntime = cfs_rq->min_vruntime;
3339 } while (min_vruntime != min_vruntime_copy);
3341 min_vruntime = cfs_rq->min_vruntime;
3344 se->vruntime -= min_vruntime;
3348 #ifdef CONFIG_FAIR_GROUP_SCHED
3350 * effective_load() calculates the load change as seen from the root_task_group
3352 * Adding load to a group doesn't make a group heavier, but can cause movement
3353 * of group shares between cpus. Assuming the shares were perfectly aligned one
3354 * can calculate the shift in shares.
3356 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3357 * on this @cpu and results in a total addition (subtraction) of @wg to the
3358 * total group weight.
3360 * Given a runqueue weight distribution (rw_i) we can compute a shares
3361 * distribution (s_i) using:
3363 * s_i = rw_i / \Sum rw_j (1)
3365 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3366 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3367 * shares distribution (s_i):
3369 * rw_i = { 2, 4, 1, 0 }
3370 * s_i = { 2/7, 4/7, 1/7, 0 }
3372 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3373 * task used to run on and the CPU the waker is running on), we need to
3374 * compute the effect of waking a task on either CPU and, in case of a sync
3375 * wakeup, compute the effect of the current task going to sleep.
3377 * So for a change of @wl to the local @cpu with an overall group weight change
3378 * of @wl we can compute the new shares distribution (s'_i) using:
3380 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3382 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3383 * differences in waking a task to CPU 0. The additional task changes the
3384 * weight and shares distributions like:
3386 * rw'_i = { 3, 4, 1, 0 }
3387 * s'_i = { 3/8, 4/8, 1/8, 0 }
3389 * We can then compute the difference in effective weight by using:
3391 * dw_i = S * (s'_i - s_i) (3)
3393 * Where 'S' is the group weight as seen by its parent.
3395 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3396 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3397 * 4/7) times the weight of the group.
3399 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3401 struct sched_entity *se = tg->se[cpu];
3403 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3406 for_each_sched_entity(se) {
3412 * W = @wg + \Sum rw_j
3414 W = wg + calc_tg_weight(tg, se->my_q);
3419 w = se->my_q->load.weight + wl;
3422 * wl = S * s'_i; see (2)
3425 wl = (w * tg->shares) / W;
3430 * Per the above, wl is the new se->load.weight value; since
3431 * those are clipped to [MIN_SHARES, ...) do so now. See
3432 * calc_cfs_shares().
3434 if (wl < MIN_SHARES)
3438 * wl = dw_i = S * (s'_i - s_i); see (3)
3440 wl -= se->load.weight;
3443 * Recursively apply this logic to all parent groups to compute
3444 * the final effective load change on the root group. Since
3445 * only the @tg group gets extra weight, all parent groups can
3446 * only redistribute existing shares. @wl is the shift in shares
3447 * resulting from this level per the above.
3456 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3463 static int wake_wide(struct task_struct *p)
3465 int factor = this_cpu_read(sd_llc_size);
3468 * Yeah, it's the switching-frequency, could means many wakee or
3469 * rapidly switch, use factor here will just help to automatically
3470 * adjust the loose-degree, so bigger node will lead to more pull.
3472 if (p->wakee_flips > factor) {
3474 * wakee is somewhat hot, it needs certain amount of cpu
3475 * resource, so if waker is far more hot, prefer to leave
3478 if (current->wakee_flips > (factor * p->wakee_flips))
3485 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3487 s64 this_load, load;
3488 int idx, this_cpu, prev_cpu;
3489 unsigned long tl_per_task;
3490 struct task_group *tg;
3491 unsigned long weight;
3495 * If we wake multiple tasks be careful to not bounce
3496 * ourselves around too much.
3502 this_cpu = smp_processor_id();
3503 prev_cpu = task_cpu(p);
3504 load = source_load(prev_cpu, idx);
3505 this_load = target_load(this_cpu, idx);
3508 * If sync wakeup then subtract the (maximum possible)
3509 * effect of the currently running task from the load
3510 * of the current CPU:
3513 tg = task_group(current);
3514 weight = current->se.load.weight;
3516 this_load += effective_load(tg, this_cpu, -weight, -weight);
3517 load += effective_load(tg, prev_cpu, 0, -weight);
3521 weight = p->se.load.weight;
3524 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3525 * due to the sync cause above having dropped this_load to 0, we'll
3526 * always have an imbalance, but there's really nothing you can do
3527 * about that, so that's good too.
3529 * Otherwise check if either cpus are near enough in load to allow this
3530 * task to be woken on this_cpu.
3532 if (this_load > 0) {
3533 s64 this_eff_load, prev_eff_load;
3535 this_eff_load = 100;
3536 this_eff_load *= power_of(prev_cpu);
3537 this_eff_load *= this_load +
3538 effective_load(tg, this_cpu, weight, weight);
3540 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3541 prev_eff_load *= power_of(this_cpu);
3542 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3544 balanced = this_eff_load <= prev_eff_load;
3549 * If the currently running task will sleep within
3550 * a reasonable amount of time then attract this newly
3553 if (sync && balanced)
3556 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3557 tl_per_task = cpu_avg_load_per_task(this_cpu);
3560 (this_load <= load &&
3561 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3563 * This domain has SD_WAKE_AFFINE and
3564 * p is cache cold in this domain, and
3565 * there is no bad imbalance.
3567 schedstat_inc(sd, ttwu_move_affine);
3568 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3576 * find_idlest_group finds and returns the least busy CPU group within the
3579 static struct sched_group *
3580 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3581 int this_cpu, int load_idx)
3583 struct sched_group *idlest = NULL, *group = sd->groups;
3584 unsigned long min_load = ULONG_MAX, this_load = 0;
3585 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3588 unsigned long load, avg_load;
3592 /* Skip over this group if it has no CPUs allowed */
3593 if (!cpumask_intersects(sched_group_cpus(group),
3594 tsk_cpus_allowed(p)))
3597 local_group = cpumask_test_cpu(this_cpu,
3598 sched_group_cpus(group));
3600 /* Tally up the load of all CPUs in the group */
3603 for_each_cpu(i, sched_group_cpus(group)) {
3604 /* Bias balancing toward cpus of our domain */
3606 load = source_load(i, load_idx);
3608 load = target_load(i, load_idx);
3613 /* Adjust by relative CPU power of the group */
3614 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3617 this_load = avg_load;
3618 } else if (avg_load < min_load) {
3619 min_load = avg_load;
3622 } while (group = group->next, group != sd->groups);
3624 if (!idlest || 100*this_load < imbalance*min_load)
3630 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3633 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3635 unsigned long load, min_load = ULONG_MAX;
3639 /* Traverse only the allowed CPUs */
3640 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3641 load = weighted_cpuload(i);
3643 if (load < min_load || (load == min_load && i == this_cpu)) {
3653 * Try and locate an idle CPU in the sched_domain.
3655 static int select_idle_sibling(struct task_struct *p, int target)
3657 struct sched_domain *sd;
3658 struct sched_group *sg;
3659 int i = task_cpu(p);
3661 if (idle_cpu(target))
3665 * If the prevous cpu is cache affine and idle, don't be stupid.
3667 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3671 * Otherwise, iterate the domains and find an elegible idle cpu.
3673 sd = rcu_dereference(per_cpu(sd_llc, target));
3674 for_each_lower_domain(sd) {
3677 if (!cpumask_intersects(sched_group_cpus(sg),
3678 tsk_cpus_allowed(p)))
3681 for_each_cpu(i, sched_group_cpus(sg)) {
3682 if (i == target || !idle_cpu(i))
3686 target = cpumask_first_and(sched_group_cpus(sg),
3687 tsk_cpus_allowed(p));
3691 } while (sg != sd->groups);
3698 * sched_balance_self: balance the current task (running on cpu) in domains
3699 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3702 * Balance, ie. select the least loaded group.
3704 * Returns the target CPU number, or the same CPU if no balancing is needed.
3706 * preempt must be disabled.
3709 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
3711 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3712 int cpu = smp_processor_id();
3714 int want_affine = 0;
3715 int sync = wake_flags & WF_SYNC;
3717 if (p->nr_cpus_allowed == 1)
3720 if (sd_flag & SD_BALANCE_WAKE) {
3721 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3727 for_each_domain(cpu, tmp) {
3728 if (!(tmp->flags & SD_LOAD_BALANCE))
3732 * If both cpu and prev_cpu are part of this domain,
3733 * cpu is a valid SD_WAKE_AFFINE target.
3735 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3736 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3741 if (tmp->flags & sd_flag)
3746 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3749 new_cpu = select_idle_sibling(p, prev_cpu);
3754 int load_idx = sd->forkexec_idx;
3755 struct sched_group *group;
3758 if (!(sd->flags & sd_flag)) {
3763 if (sd_flag & SD_BALANCE_WAKE)
3764 load_idx = sd->wake_idx;
3766 group = find_idlest_group(sd, p, cpu, load_idx);
3772 new_cpu = find_idlest_cpu(group, p, cpu);
3773 if (new_cpu == -1 || new_cpu == cpu) {
3774 /* Now try balancing at a lower domain level of cpu */
3779 /* Now try balancing at a lower domain level of new_cpu */
3781 weight = sd->span_weight;
3783 for_each_domain(cpu, tmp) {
3784 if (weight <= tmp->span_weight)
3786 if (tmp->flags & sd_flag)
3789 /* while loop will break here if sd == NULL */
3798 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3799 * cfs_rq_of(p) references at time of call are still valid and identify the
3800 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3801 * other assumptions, including the state of rq->lock, should be made.
3804 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3806 struct sched_entity *se = &p->se;
3807 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3810 * Load tracking: accumulate removed load so that it can be processed
3811 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3812 * to blocked load iff they have a positive decay-count. It can never
3813 * be negative here since on-rq tasks have decay-count == 0.
3815 if (se->avg.decay_count) {
3816 se->avg.decay_count = -__synchronize_entity_decay(se);
3817 atomic_long_add(se->avg.load_avg_contrib,
3818 &cfs_rq->removed_load);
3821 #endif /* CONFIG_SMP */
3823 static unsigned long
3824 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3826 unsigned long gran = sysctl_sched_wakeup_granularity;
3829 * Since its curr running now, convert the gran from real-time
3830 * to virtual-time in his units.
3832 * By using 'se' instead of 'curr' we penalize light tasks, so
3833 * they get preempted easier. That is, if 'se' < 'curr' then
3834 * the resulting gran will be larger, therefore penalizing the
3835 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3836 * be smaller, again penalizing the lighter task.
3838 * This is especially important for buddies when the leftmost
3839 * task is higher priority than the buddy.
3841 return calc_delta_fair(gran, se);
3845 * Should 'se' preempt 'curr'.
3859 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3861 s64 gran, vdiff = curr->vruntime - se->vruntime;
3866 gran = wakeup_gran(curr, se);
3873 static void set_last_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)->last = se;
3882 static void set_next_buddy(struct sched_entity *se)
3884 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3887 for_each_sched_entity(se)
3888 cfs_rq_of(se)->next = se;
3891 static void set_skip_buddy(struct sched_entity *se)
3893 for_each_sched_entity(se)
3894 cfs_rq_of(se)->skip = se;
3898 * Preempt the current task with a newly woken task if needed:
3900 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3902 struct task_struct *curr = rq->curr;
3903 struct sched_entity *se = &curr->se, *pse = &p->se;
3904 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3905 int scale = cfs_rq->nr_running >= sched_nr_latency;
3906 int next_buddy_marked = 0;
3908 if (unlikely(se == pse))
3912 * This is possible from callers such as move_task(), in which we
3913 * unconditionally check_prempt_curr() after an enqueue (which may have
3914 * lead to a throttle). This both saves work and prevents false
3915 * next-buddy nomination below.
3917 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3920 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3921 set_next_buddy(pse);
3922 next_buddy_marked = 1;
3926 * We can come here with TIF_NEED_RESCHED already set from new task
3929 * Note: this also catches the edge-case of curr being in a throttled
3930 * group (e.g. via set_curr_task), since update_curr() (in the
3931 * enqueue of curr) will have resulted in resched being set. This
3932 * prevents us from potentially nominating it as a false LAST_BUDDY
3935 if (test_tsk_need_resched(curr))
3938 /* Idle tasks are by definition preempted by non-idle tasks. */
3939 if (unlikely(curr->policy == SCHED_IDLE) &&
3940 likely(p->policy != SCHED_IDLE))
3944 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3945 * is driven by the tick):
3947 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3950 find_matching_se(&se, &pse);
3951 update_curr(cfs_rq_of(se));
3953 if (wakeup_preempt_entity(se, pse) == 1) {
3955 * Bias pick_next to pick the sched entity that is
3956 * triggering this preemption.
3958 if (!next_buddy_marked)
3959 set_next_buddy(pse);
3968 * Only set the backward buddy when the current task is still
3969 * on the rq. This can happen when a wakeup gets interleaved
3970 * with schedule on the ->pre_schedule() or idle_balance()
3971 * point, either of which can * drop the rq lock.
3973 * Also, during early boot the idle thread is in the fair class,
3974 * for obvious reasons its a bad idea to schedule back to it.
3976 if (unlikely(!se->on_rq || curr == rq->idle))
3979 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3983 static struct task_struct *pick_next_task_fair(struct rq *rq)
3985 struct task_struct *p;
3986 struct cfs_rq *cfs_rq = &rq->cfs;
3987 struct sched_entity *se;
3989 if (!cfs_rq->nr_running)
3993 se = pick_next_entity(cfs_rq);
3994 set_next_entity(cfs_rq, se);
3995 cfs_rq = group_cfs_rq(se);
3999 if (hrtick_enabled(rq))
4000 hrtick_start_fair(rq, p);
4006 * Account for a descheduled task:
4008 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4010 struct sched_entity *se = &prev->se;
4011 struct cfs_rq *cfs_rq;
4013 for_each_sched_entity(se) {
4014 cfs_rq = cfs_rq_of(se);
4015 put_prev_entity(cfs_rq, se);
4020 * sched_yield() is very simple
4022 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4024 static void yield_task_fair(struct rq *rq)
4026 struct task_struct *curr = rq->curr;
4027 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4028 struct sched_entity *se = &curr->se;
4031 * Are we the only task in the tree?
4033 if (unlikely(rq->nr_running == 1))
4036 clear_buddies(cfs_rq, se);
4038 if (curr->policy != SCHED_BATCH) {
4039 update_rq_clock(rq);
4041 * Update run-time statistics of the 'current'.
4043 update_curr(cfs_rq);
4045 * Tell update_rq_clock() that we've just updated,
4046 * so we don't do microscopic update in schedule()
4047 * and double the fastpath cost.
4049 rq->skip_clock_update = 1;
4055 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4057 struct sched_entity *se = &p->se;
4059 /* throttled hierarchies are not runnable */
4060 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4063 /* Tell the scheduler that we'd really like pse to run next. */
4066 yield_task_fair(rq);
4072 /**************************************************
4073 * Fair scheduling class load-balancing methods.
4077 * The purpose of load-balancing is to achieve the same basic fairness the
4078 * per-cpu scheduler provides, namely provide a proportional amount of compute
4079 * time to each task. This is expressed in the following equation:
4081 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4083 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4084 * W_i,0 is defined as:
4086 * W_i,0 = \Sum_j w_i,j (2)
4088 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4089 * is derived from the nice value as per prio_to_weight[].
4091 * The weight average is an exponential decay average of the instantaneous
4094 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4096 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4097 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4098 * can also include other factors [XXX].
4100 * To achieve this balance we define a measure of imbalance which follows
4101 * directly from (1):
4103 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4105 * We them move tasks around to minimize the imbalance. In the continuous
4106 * function space it is obvious this converges, in the discrete case we get
4107 * a few fun cases generally called infeasible weight scenarios.
4110 * - infeasible weights;
4111 * - local vs global optima in the discrete case. ]
4116 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4117 * for all i,j solution, we create a tree of cpus that follows the hardware
4118 * topology where each level pairs two lower groups (or better). This results
4119 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4120 * tree to only the first of the previous level and we decrease the frequency
4121 * of load-balance at each level inv. proportional to the number of cpus in
4127 * \Sum { --- * --- * 2^i } = O(n) (5)
4129 * `- size of each group
4130 * | | `- number of cpus doing load-balance
4132 * `- sum over all levels
4134 * Coupled with a limit on how many tasks we can migrate every balance pass,
4135 * this makes (5) the runtime complexity of the balancer.
4137 * An important property here is that each CPU is still (indirectly) connected
4138 * to every other cpu in at most O(log n) steps:
4140 * The adjacency matrix of the resulting graph is given by:
4143 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4146 * And you'll find that:
4148 * A^(log_2 n)_i,j != 0 for all i,j (7)
4150 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4151 * The task movement gives a factor of O(m), giving a convergence complexity
4154 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4159 * In order to avoid CPUs going idle while there's still work to do, new idle
4160 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4161 * tree itself instead of relying on other CPUs to bring it work.
4163 * This adds some complexity to both (5) and (8) but it reduces the total idle
4171 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4174 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4179 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4181 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4183 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4186 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4187 * rewrite all of this once again.]
4190 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4192 #define LBF_ALL_PINNED 0x01
4193 #define LBF_NEED_BREAK 0x02
4194 #define LBF_DST_PINNED 0x04
4195 #define LBF_SOME_PINNED 0x08
4198 struct sched_domain *sd;
4206 struct cpumask *dst_grpmask;
4208 enum cpu_idle_type idle;
4210 /* The set of CPUs under consideration for load-balancing */
4211 struct cpumask *cpus;
4216 unsigned int loop_break;
4217 unsigned int loop_max;
4221 * move_task - move a task from one runqueue to another runqueue.
4222 * Both runqueues must be locked.
4224 static void move_task(struct task_struct *p, struct lb_env *env)
4226 deactivate_task(env->src_rq, p, 0);
4227 set_task_cpu(p, env->dst_cpu);
4228 activate_task(env->dst_rq, p, 0);
4229 check_preempt_curr(env->dst_rq, p, 0);
4230 #ifdef CONFIG_NUMA_BALANCING
4231 if (p->numa_preferred_nid != -1) {
4232 int src_nid = cpu_to_node(env->src_cpu);
4233 int dst_nid = cpu_to_node(env->dst_cpu);
4236 * If the load balancer has moved the task then limit
4237 * migrations from taking place in the short term in
4238 * case this is a short-lived migration.
4240 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4241 p->numa_migrate_seq = 0;
4247 * Is this task likely cache-hot:
4250 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4254 if (p->sched_class != &fair_sched_class)
4257 if (unlikely(p->policy == SCHED_IDLE))
4261 * Buddy candidates are cache hot:
4263 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4264 (&p->se == cfs_rq_of(&p->se)->next ||
4265 &p->se == cfs_rq_of(&p->se)->last))
4268 if (sysctl_sched_migration_cost == -1)
4270 if (sysctl_sched_migration_cost == 0)
4273 delta = now - p->se.exec_start;
4275 return delta < (s64)sysctl_sched_migration_cost;
4278 #ifdef CONFIG_NUMA_BALANCING
4279 /* Returns true if the destination node has incurred more faults */
4280 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4282 int src_nid, dst_nid;
4284 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4285 !(env->sd->flags & SD_NUMA)) {
4289 src_nid = cpu_to_node(env->src_cpu);
4290 dst_nid = cpu_to_node(env->dst_cpu);
4292 if (src_nid == dst_nid ||
4293 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4296 if (dst_nid == p->numa_preferred_nid ||
4297 task_faults(p, dst_nid) > task_faults(p, src_nid))
4304 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4306 int src_nid, dst_nid;
4308 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4311 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4314 src_nid = cpu_to_node(env->src_cpu);
4315 dst_nid = cpu_to_node(env->dst_cpu);
4317 if (src_nid == dst_nid ||
4318 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4321 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4328 static inline bool migrate_improves_locality(struct task_struct *p,
4334 static inline bool migrate_degrades_locality(struct task_struct *p,
4342 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4345 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4347 int tsk_cache_hot = 0;
4349 * We do not migrate tasks that are:
4350 * 1) throttled_lb_pair, or
4351 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4352 * 3) running (obviously), or
4353 * 4) are cache-hot on their current CPU.
4355 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4358 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4361 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4363 env->flags |= LBF_SOME_PINNED;
4366 * Remember if this task can be migrated to any other cpu in
4367 * our sched_group. We may want to revisit it if we couldn't
4368 * meet load balance goals by pulling other tasks on src_cpu.
4370 * Also avoid computing new_dst_cpu if we have already computed
4371 * one in current iteration.
4373 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4376 /* Prevent to re-select dst_cpu via env's cpus */
4377 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4378 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4379 env->flags |= LBF_DST_PINNED;
4380 env->new_dst_cpu = cpu;
4388 /* Record that we found atleast one task that could run on dst_cpu */
4389 env->flags &= ~LBF_ALL_PINNED;
4391 if (task_running(env->src_rq, p)) {
4392 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4397 * Aggressive migration if:
4398 * 1) destination numa is preferred
4399 * 2) task is cache cold, or
4400 * 3) too many balance attempts have failed.
4402 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4404 tsk_cache_hot = migrate_degrades_locality(p, env);
4406 if (migrate_improves_locality(p, env)) {
4407 #ifdef CONFIG_SCHEDSTATS
4408 if (tsk_cache_hot) {
4409 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4410 schedstat_inc(p, se.statistics.nr_forced_migrations);
4416 if (!tsk_cache_hot ||
4417 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4419 if (tsk_cache_hot) {
4420 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4421 schedstat_inc(p, se.statistics.nr_forced_migrations);
4427 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4432 * move_one_task tries to move exactly one task from busiest to this_rq, as
4433 * part of active balancing operations within "domain".
4434 * Returns 1 if successful and 0 otherwise.
4436 * Called with both runqueues locked.
4438 static int move_one_task(struct lb_env *env)
4440 struct task_struct *p, *n;
4442 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4443 if (!can_migrate_task(p, env))
4448 * Right now, this is only the second place move_task()
4449 * is called, so we can safely collect move_task()
4450 * stats here rather than inside move_task().
4452 schedstat_inc(env->sd, lb_gained[env->idle]);
4458 static unsigned long task_h_load(struct task_struct *p);
4460 static const unsigned int sched_nr_migrate_break = 32;
4463 * move_tasks tries to move up to imbalance weighted load from busiest to
4464 * this_rq, as part of a balancing operation within domain "sd".
4465 * Returns 1 if successful and 0 otherwise.
4467 * Called with both runqueues locked.
4469 static int move_tasks(struct lb_env *env)
4471 struct list_head *tasks = &env->src_rq->cfs_tasks;
4472 struct task_struct *p;
4476 if (env->imbalance <= 0)
4479 while (!list_empty(tasks)) {
4480 p = list_first_entry(tasks, struct task_struct, se.group_node);
4483 /* We've more or less seen every task there is, call it quits */
4484 if (env->loop > env->loop_max)
4487 /* take a breather every nr_migrate tasks */
4488 if (env->loop > env->loop_break) {
4489 env->loop_break += sched_nr_migrate_break;
4490 env->flags |= LBF_NEED_BREAK;
4494 if (!can_migrate_task(p, env))
4497 load = task_h_load(p);
4499 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4502 if ((load / 2) > env->imbalance)
4507 env->imbalance -= load;
4509 #ifdef CONFIG_PREEMPT
4511 * NEWIDLE balancing is a source of latency, so preemptible
4512 * kernels will stop after the first task is pulled to minimize
4513 * the critical section.
4515 if (env->idle == CPU_NEWLY_IDLE)
4520 * We only want to steal up to the prescribed amount of
4523 if (env->imbalance <= 0)
4528 list_move_tail(&p->se.group_node, tasks);
4532 * Right now, this is one of only two places move_task() is called,
4533 * so we can safely collect move_task() stats here rather than
4534 * inside move_task().
4536 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4541 #ifdef CONFIG_FAIR_GROUP_SCHED
4543 * update tg->load_weight by folding this cpu's load_avg
4545 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4547 struct sched_entity *se = tg->se[cpu];
4548 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4550 /* throttled entities do not contribute to load */
4551 if (throttled_hierarchy(cfs_rq))
4554 update_cfs_rq_blocked_load(cfs_rq, 1);
4557 update_entity_load_avg(se, 1);
4559 * We pivot on our runnable average having decayed to zero for
4560 * list removal. This generally implies that all our children
4561 * have also been removed (modulo rounding error or bandwidth
4562 * control); however, such cases are rare and we can fix these
4565 * TODO: fix up out-of-order children on enqueue.
4567 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4568 list_del_leaf_cfs_rq(cfs_rq);
4570 struct rq *rq = rq_of(cfs_rq);
4571 update_rq_runnable_avg(rq, rq->nr_running);
4575 static void update_blocked_averages(int cpu)
4577 struct rq *rq = cpu_rq(cpu);
4578 struct cfs_rq *cfs_rq;
4579 unsigned long flags;
4581 raw_spin_lock_irqsave(&rq->lock, flags);
4582 update_rq_clock(rq);
4584 * Iterates the task_group tree in a bottom up fashion, see
4585 * list_add_leaf_cfs_rq() for details.
4587 for_each_leaf_cfs_rq(rq, cfs_rq) {
4589 * Note: We may want to consider periodically releasing
4590 * rq->lock about these updates so that creating many task
4591 * groups does not result in continually extending hold time.
4593 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4596 raw_spin_unlock_irqrestore(&rq->lock, flags);
4600 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4601 * This needs to be done in a top-down fashion because the load of a child
4602 * group is a fraction of its parents load.
4604 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4606 struct rq *rq = rq_of(cfs_rq);
4607 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4608 unsigned long now = jiffies;
4611 if (cfs_rq->last_h_load_update == now)
4614 cfs_rq->h_load_next = NULL;
4615 for_each_sched_entity(se) {
4616 cfs_rq = cfs_rq_of(se);
4617 cfs_rq->h_load_next = se;
4618 if (cfs_rq->last_h_load_update == now)
4623 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4624 cfs_rq->last_h_load_update = now;
4627 while ((se = cfs_rq->h_load_next) != NULL) {
4628 load = cfs_rq->h_load;
4629 load = div64_ul(load * se->avg.load_avg_contrib,
4630 cfs_rq->runnable_load_avg + 1);
4631 cfs_rq = group_cfs_rq(se);
4632 cfs_rq->h_load = load;
4633 cfs_rq->last_h_load_update = now;
4637 static unsigned long task_h_load(struct task_struct *p)
4639 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4641 update_cfs_rq_h_load(cfs_rq);
4642 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4643 cfs_rq->runnable_load_avg + 1);
4646 static inline void update_blocked_averages(int cpu)
4650 static unsigned long task_h_load(struct task_struct *p)
4652 return p->se.avg.load_avg_contrib;
4656 /********** Helpers for find_busiest_group ************************/
4658 * sg_lb_stats - stats of a sched_group required for load_balancing
4660 struct sg_lb_stats {
4661 unsigned long avg_load; /*Avg load across the CPUs of the group */
4662 unsigned long group_load; /* Total load over the CPUs of the group */
4663 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4664 unsigned long load_per_task;
4665 unsigned long group_power;
4666 unsigned int sum_nr_running; /* Nr tasks running in the group */
4667 unsigned int group_capacity;
4668 unsigned int idle_cpus;
4669 unsigned int group_weight;
4670 int group_imb; /* Is there an imbalance in the group ? */
4671 int group_has_capacity; /* Is there extra capacity in the group? */
4675 * sd_lb_stats - Structure to store the statistics of a sched_domain
4676 * during load balancing.
4678 struct sd_lb_stats {
4679 struct sched_group *busiest; /* Busiest group in this sd */
4680 struct sched_group *local; /* Local group in this sd */
4681 unsigned long total_load; /* Total load of all groups in sd */
4682 unsigned long total_pwr; /* Total power of all groups in sd */
4683 unsigned long avg_load; /* Average load across all groups in sd */
4685 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4686 struct sg_lb_stats local_stat; /* Statistics of the local group */
4689 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4692 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4693 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4694 * We must however clear busiest_stat::avg_load because
4695 * update_sd_pick_busiest() reads this before assignment.
4697 *sds = (struct sd_lb_stats){
4709 * get_sd_load_idx - Obtain the load index for a given sched domain.
4710 * @sd: The sched_domain whose load_idx is to be obtained.
4711 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4713 * Return: The load index.
4715 static inline int get_sd_load_idx(struct sched_domain *sd,
4716 enum cpu_idle_type idle)
4722 load_idx = sd->busy_idx;
4725 case CPU_NEWLY_IDLE:
4726 load_idx = sd->newidle_idx;
4729 load_idx = sd->idle_idx;
4736 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4738 return SCHED_POWER_SCALE;
4741 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4743 return default_scale_freq_power(sd, cpu);
4746 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4748 unsigned long weight = sd->span_weight;
4749 unsigned long smt_gain = sd->smt_gain;
4756 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4758 return default_scale_smt_power(sd, cpu);
4761 static unsigned long scale_rt_power(int cpu)
4763 struct rq *rq = cpu_rq(cpu);
4764 u64 total, available, age_stamp, avg;
4767 * Since we're reading these variables without serialization make sure
4768 * we read them once before doing sanity checks on them.
4770 age_stamp = ACCESS_ONCE(rq->age_stamp);
4771 avg = ACCESS_ONCE(rq->rt_avg);
4773 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4775 if (unlikely(total < avg)) {
4776 /* Ensures that power won't end up being negative */
4779 available = total - avg;
4782 if (unlikely((s64)total < SCHED_POWER_SCALE))
4783 total = SCHED_POWER_SCALE;
4785 total >>= SCHED_POWER_SHIFT;
4787 return div_u64(available, total);
4790 static void update_cpu_power(struct sched_domain *sd, int cpu)
4792 unsigned long weight = sd->span_weight;
4793 unsigned long power = SCHED_POWER_SCALE;
4794 struct sched_group *sdg = sd->groups;
4796 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4797 if (sched_feat(ARCH_POWER))
4798 power *= arch_scale_smt_power(sd, cpu);
4800 power *= default_scale_smt_power(sd, cpu);
4802 power >>= SCHED_POWER_SHIFT;
4805 sdg->sgp->power_orig = power;
4807 if (sched_feat(ARCH_POWER))
4808 power *= arch_scale_freq_power(sd, cpu);
4810 power *= default_scale_freq_power(sd, cpu);
4812 power >>= SCHED_POWER_SHIFT;
4814 power *= scale_rt_power(cpu);
4815 power >>= SCHED_POWER_SHIFT;
4820 cpu_rq(cpu)->cpu_power = power;
4821 sdg->sgp->power = power;
4824 void update_group_power(struct sched_domain *sd, int cpu)
4826 struct sched_domain *child = sd->child;
4827 struct sched_group *group, *sdg = sd->groups;
4828 unsigned long power, power_orig;
4829 unsigned long interval;
4831 interval = msecs_to_jiffies(sd->balance_interval);
4832 interval = clamp(interval, 1UL, max_load_balance_interval);
4833 sdg->sgp->next_update = jiffies + interval;
4836 update_cpu_power(sd, cpu);
4840 power_orig = power = 0;
4842 if (child->flags & SD_OVERLAP) {
4844 * SD_OVERLAP domains cannot assume that child groups
4845 * span the current group.
4848 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4849 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4851 power_orig += sg->sgp->power_orig;
4852 power += sg->sgp->power;
4856 * !SD_OVERLAP domains can assume that child groups
4857 * span the current group.
4860 group = child->groups;
4862 power_orig += group->sgp->power_orig;
4863 power += group->sgp->power;
4864 group = group->next;
4865 } while (group != child->groups);
4868 sdg->sgp->power_orig = power_orig;
4869 sdg->sgp->power = power;
4873 * Try and fix up capacity for tiny siblings, this is needed when
4874 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4875 * which on its own isn't powerful enough.
4877 * See update_sd_pick_busiest() and check_asym_packing().
4880 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4883 * Only siblings can have significantly less than SCHED_POWER_SCALE
4885 if (!(sd->flags & SD_SHARE_CPUPOWER))
4889 * If ~90% of the cpu_power is still there, we're good.
4891 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4898 * Group imbalance indicates (and tries to solve) the problem where balancing
4899 * groups is inadequate due to tsk_cpus_allowed() constraints.
4901 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4902 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4905 * { 0 1 2 3 } { 4 5 6 7 }
4908 * If we were to balance group-wise we'd place two tasks in the first group and
4909 * two tasks in the second group. Clearly this is undesired as it will overload
4910 * cpu 3 and leave one of the cpus in the second group unused.
4912 * The current solution to this issue is detecting the skew in the first group
4913 * by noticing the lower domain failed to reach balance and had difficulty
4914 * moving tasks due to affinity constraints.
4916 * When this is so detected; this group becomes a candidate for busiest; see
4917 * update_sd_pick_busiest(). And calculcate_imbalance() and
4918 * find_busiest_group() avoid some of the usual balance conditions to allow it
4919 * to create an effective group imbalance.
4921 * This is a somewhat tricky proposition since the next run might not find the
4922 * group imbalance and decide the groups need to be balanced again. A most
4923 * subtle and fragile situation.
4926 static inline int sg_imbalanced(struct sched_group *group)
4928 return group->sgp->imbalance;
4932 * Compute the group capacity.
4934 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4935 * first dividing out the smt factor and computing the actual number of cores
4936 * and limit power unit capacity with that.
4938 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4940 unsigned int capacity, smt, cpus;
4941 unsigned int power, power_orig;
4943 power = group->sgp->power;
4944 power_orig = group->sgp->power_orig;
4945 cpus = group->group_weight;
4947 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4948 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4949 capacity = cpus / smt; /* cores */
4951 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4953 capacity = fix_small_capacity(env->sd, group);
4959 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4960 * @env: The load balancing environment.
4961 * @group: sched_group whose statistics are to be updated.
4962 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4963 * @local_group: Does group contain this_cpu.
4964 * @sgs: variable to hold the statistics for this group.
4966 static inline void update_sg_lb_stats(struct lb_env *env,
4967 struct sched_group *group, int load_idx,
4968 int local_group, struct sg_lb_stats *sgs)
4970 unsigned long nr_running;
4974 memset(sgs, 0, sizeof(*sgs));
4976 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4977 struct rq *rq = cpu_rq(i);
4979 nr_running = rq->nr_running;
4981 /* Bias balancing toward cpus of our domain */
4983 load = target_load(i, load_idx);
4985 load = source_load(i, load_idx);
4987 sgs->group_load += load;
4988 sgs->sum_nr_running += nr_running;
4989 sgs->sum_weighted_load += weighted_cpuload(i);
4994 /* Adjust by relative CPU power of the group */
4995 sgs->group_power = group->sgp->power;
4996 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4998 if (sgs->sum_nr_running)
4999 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5001 sgs->group_weight = group->group_weight;
5003 sgs->group_imb = sg_imbalanced(group);
5004 sgs->group_capacity = sg_capacity(env, group);
5006 if (sgs->group_capacity > sgs->sum_nr_running)
5007 sgs->group_has_capacity = 1;
5011 * update_sd_pick_busiest - return 1 on busiest group
5012 * @env: The load balancing environment.
5013 * @sds: sched_domain statistics
5014 * @sg: sched_group candidate to be checked for being the busiest
5015 * @sgs: sched_group statistics
5017 * Determine if @sg is a busier group than the previously selected
5020 * Return: %true if @sg is a busier group than the previously selected
5021 * busiest group. %false otherwise.
5023 static bool update_sd_pick_busiest(struct lb_env *env,
5024 struct sd_lb_stats *sds,
5025 struct sched_group *sg,
5026 struct sg_lb_stats *sgs)
5028 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5031 if (sgs->sum_nr_running > sgs->group_capacity)
5038 * ASYM_PACKING needs to move all the work to the lowest
5039 * numbered CPUs in the group, therefore mark all groups
5040 * higher than ourself as busy.
5042 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5043 env->dst_cpu < group_first_cpu(sg)) {
5047 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5055 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5056 * @env: The load balancing environment.
5057 * @balance: Should we balance.
5058 * @sds: variable to hold the statistics for this sched_domain.
5060 static inline void update_sd_lb_stats(struct lb_env *env,
5061 struct sd_lb_stats *sds)
5063 struct sched_domain *child = env->sd->child;
5064 struct sched_group *sg = env->sd->groups;
5065 struct sg_lb_stats tmp_sgs;
5066 int load_idx, prefer_sibling = 0;
5068 if (child && child->flags & SD_PREFER_SIBLING)
5071 load_idx = get_sd_load_idx(env->sd, env->idle);
5074 struct sg_lb_stats *sgs = &tmp_sgs;
5077 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5080 sgs = &sds->local_stat;
5082 if (env->idle != CPU_NEWLY_IDLE ||
5083 time_after_eq(jiffies, sg->sgp->next_update))
5084 update_group_power(env->sd, env->dst_cpu);
5087 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5093 * In case the child domain prefers tasks go to siblings
5094 * first, lower the sg capacity to one so that we'll try
5095 * and move all the excess tasks away. We lower the capacity
5096 * of a group only if the local group has the capacity to fit
5097 * these excess tasks, i.e. nr_running < group_capacity. The
5098 * extra check prevents the case where you always pull from the
5099 * heaviest group when it is already under-utilized (possible
5100 * with a large weight task outweighs the tasks on the system).
5102 if (prefer_sibling && sds->local &&
5103 sds->local_stat.group_has_capacity)
5104 sgs->group_capacity = min(sgs->group_capacity, 1U);
5106 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5108 sds->busiest_stat = *sgs;
5112 /* Now, start updating sd_lb_stats */
5113 sds->total_load += sgs->group_load;
5114 sds->total_pwr += sgs->group_power;
5117 } while (sg != env->sd->groups);
5121 * check_asym_packing - Check to see if the group is packed into the
5124 * This is primarily intended to used at the sibling level. Some
5125 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5126 * case of POWER7, it can move to lower SMT modes only when higher
5127 * threads are idle. When in lower SMT modes, the threads will
5128 * perform better since they share less core resources. Hence when we
5129 * have idle threads, we want them to be the higher ones.
5131 * This packing function is run on idle threads. It checks to see if
5132 * the busiest CPU in this domain (core in the P7 case) has a higher
5133 * CPU number than the packing function is being run on. Here we are
5134 * assuming lower CPU number will be equivalent to lower a SMT thread
5137 * Return: 1 when packing is required and a task should be moved to
5138 * this CPU. The amount of the imbalance is returned in *imbalance.
5140 * @env: The load balancing environment.
5141 * @sds: Statistics of the sched_domain which is to be packed
5143 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5147 if (!(env->sd->flags & SD_ASYM_PACKING))
5153 busiest_cpu = group_first_cpu(sds->busiest);
5154 if (env->dst_cpu > busiest_cpu)
5157 env->imbalance = DIV_ROUND_CLOSEST(
5158 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5165 * fix_small_imbalance - Calculate the minor imbalance that exists
5166 * amongst the groups of a sched_domain, during
5168 * @env: The load balancing environment.
5169 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5172 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5174 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5175 unsigned int imbn = 2;
5176 unsigned long scaled_busy_load_per_task;
5177 struct sg_lb_stats *local, *busiest;
5179 local = &sds->local_stat;
5180 busiest = &sds->busiest_stat;
5182 if (!local->sum_nr_running)
5183 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5184 else if (busiest->load_per_task > local->load_per_task)
5187 scaled_busy_load_per_task =
5188 (busiest->load_per_task * SCHED_POWER_SCALE) /
5189 busiest->group_power;
5191 if (busiest->avg_load + scaled_busy_load_per_task >=
5192 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5193 env->imbalance = busiest->load_per_task;
5198 * OK, we don't have enough imbalance to justify moving tasks,
5199 * however we may be able to increase total CPU power used by
5203 pwr_now += busiest->group_power *
5204 min(busiest->load_per_task, busiest->avg_load);
5205 pwr_now += local->group_power *
5206 min(local->load_per_task, local->avg_load);
5207 pwr_now /= SCHED_POWER_SCALE;
5209 /* Amount of load we'd subtract */
5210 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5211 busiest->group_power;
5212 if (busiest->avg_load > tmp) {
5213 pwr_move += busiest->group_power *
5214 min(busiest->load_per_task,
5215 busiest->avg_load - tmp);
5218 /* Amount of load we'd add */
5219 if (busiest->avg_load * busiest->group_power <
5220 busiest->load_per_task * SCHED_POWER_SCALE) {
5221 tmp = (busiest->avg_load * busiest->group_power) /
5224 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5227 pwr_move += local->group_power *
5228 min(local->load_per_task, local->avg_load + tmp);
5229 pwr_move /= SCHED_POWER_SCALE;
5231 /* Move if we gain throughput */
5232 if (pwr_move > pwr_now)
5233 env->imbalance = busiest->load_per_task;
5237 * calculate_imbalance - Calculate the amount of imbalance present within the
5238 * groups of a given sched_domain during load balance.
5239 * @env: load balance environment
5240 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5242 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5244 unsigned long max_pull, load_above_capacity = ~0UL;
5245 struct sg_lb_stats *local, *busiest;
5247 local = &sds->local_stat;
5248 busiest = &sds->busiest_stat;
5250 if (busiest->group_imb) {
5252 * In the group_imb case we cannot rely on group-wide averages
5253 * to ensure cpu-load equilibrium, look at wider averages. XXX
5255 busiest->load_per_task =
5256 min(busiest->load_per_task, sds->avg_load);
5260 * In the presence of smp nice balancing, certain scenarios can have
5261 * max load less than avg load(as we skip the groups at or below
5262 * its cpu_power, while calculating max_load..)
5264 if (busiest->avg_load <= sds->avg_load ||
5265 local->avg_load >= sds->avg_load) {
5267 return fix_small_imbalance(env, sds);
5270 if (!busiest->group_imb) {
5272 * Don't want to pull so many tasks that a group would go idle.
5273 * Except of course for the group_imb case, since then we might
5274 * have to drop below capacity to reach cpu-load equilibrium.
5276 load_above_capacity =
5277 (busiest->sum_nr_running - busiest->group_capacity);
5279 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5280 load_above_capacity /= busiest->group_power;
5284 * We're trying to get all the cpus to the average_load, so we don't
5285 * want to push ourselves above the average load, nor do we wish to
5286 * reduce the max loaded cpu below the average load. At the same time,
5287 * we also don't want to reduce the group load below the group capacity
5288 * (so that we can implement power-savings policies etc). Thus we look
5289 * for the minimum possible imbalance.
5291 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5293 /* How much load to actually move to equalise the imbalance */
5294 env->imbalance = min(
5295 max_pull * busiest->group_power,
5296 (sds->avg_load - local->avg_load) * local->group_power
5297 ) / SCHED_POWER_SCALE;
5300 * if *imbalance is less than the average load per runnable task
5301 * there is no guarantee that any tasks will be moved so we'll have
5302 * a think about bumping its value to force at least one task to be
5305 if (env->imbalance < busiest->load_per_task)
5306 return fix_small_imbalance(env, sds);
5309 /******* find_busiest_group() helpers end here *********************/
5312 * find_busiest_group - Returns the busiest group within the sched_domain
5313 * if there is an imbalance. If there isn't an imbalance, and
5314 * the user has opted for power-savings, it returns a group whose
5315 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5316 * such a group exists.
5318 * Also calculates the amount of weighted load which should be moved
5319 * to restore balance.
5321 * @env: The load balancing environment.
5323 * Return: - The busiest group if imbalance exists.
5324 * - If no imbalance and user has opted for power-savings balance,
5325 * return the least loaded group whose CPUs can be
5326 * put to idle by rebalancing its tasks onto our group.
5328 static struct sched_group *find_busiest_group(struct lb_env *env)
5330 struct sg_lb_stats *local, *busiest;
5331 struct sd_lb_stats sds;
5333 init_sd_lb_stats(&sds);
5336 * Compute the various statistics relavent for load balancing at
5339 update_sd_lb_stats(env, &sds);
5340 local = &sds.local_stat;
5341 busiest = &sds.busiest_stat;
5343 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5344 check_asym_packing(env, &sds))
5347 /* There is no busy sibling group to pull tasks from */
5348 if (!sds.busiest || busiest->sum_nr_running == 0)
5351 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5354 * If the busiest group is imbalanced the below checks don't
5355 * work because they assume all things are equal, which typically
5356 * isn't true due to cpus_allowed constraints and the like.
5358 if (busiest->group_imb)
5361 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5362 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5363 !busiest->group_has_capacity)
5367 * If the local group is more busy than the selected busiest group
5368 * don't try and pull any tasks.
5370 if (local->avg_load >= busiest->avg_load)
5374 * Don't pull any tasks if this group is already above the domain
5377 if (local->avg_load >= sds.avg_load)
5380 if (env->idle == CPU_IDLE) {
5382 * This cpu is idle. If the busiest group load doesn't
5383 * have more tasks than the number of available cpu's and
5384 * there is no imbalance between this and busiest group
5385 * wrt to idle cpu's, it is balanced.
5387 if ((local->idle_cpus < busiest->idle_cpus) &&
5388 busiest->sum_nr_running <= busiest->group_weight)
5392 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5393 * imbalance_pct to be conservative.
5395 if (100 * busiest->avg_load <=
5396 env->sd->imbalance_pct * local->avg_load)
5401 /* Looks like there is an imbalance. Compute it */
5402 calculate_imbalance(env, &sds);
5411 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5413 static struct rq *find_busiest_queue(struct lb_env *env,
5414 struct sched_group *group)
5416 struct rq *busiest = NULL, *rq;
5417 unsigned long busiest_load = 0, busiest_power = 1;
5420 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5421 unsigned long power = power_of(i);
5422 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5427 capacity = fix_small_capacity(env->sd, group);
5430 wl = weighted_cpuload(i);
5433 * When comparing with imbalance, use weighted_cpuload()
5434 * which is not scaled with the cpu power.
5436 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5440 * For the load comparisons with the other cpu's, consider
5441 * the weighted_cpuload() scaled with the cpu power, so that
5442 * the load can be moved away from the cpu that is potentially
5443 * running at a lower capacity.
5445 * Thus we're looking for max(wl_i / power_i), crosswise
5446 * multiplication to rid ourselves of the division works out
5447 * to: wl_i * power_j > wl_j * power_i; where j is our
5450 if (wl * busiest_power > busiest_load * power) {
5452 busiest_power = power;
5461 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5462 * so long as it is large enough.
5464 #define MAX_PINNED_INTERVAL 512
5466 /* Working cpumask for load_balance and load_balance_newidle. */
5467 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5469 static int need_active_balance(struct lb_env *env)
5471 struct sched_domain *sd = env->sd;
5473 if (env->idle == CPU_NEWLY_IDLE) {
5476 * ASYM_PACKING needs to force migrate tasks from busy but
5477 * higher numbered CPUs in order to pack all tasks in the
5478 * lowest numbered CPUs.
5480 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5484 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5487 static int active_load_balance_cpu_stop(void *data);
5489 static int should_we_balance(struct lb_env *env)
5491 struct sched_group *sg = env->sd->groups;
5492 struct cpumask *sg_cpus, *sg_mask;
5493 int cpu, balance_cpu = -1;
5496 * In the newly idle case, we will allow all the cpu's
5497 * to do the newly idle load balance.
5499 if (env->idle == CPU_NEWLY_IDLE)
5502 sg_cpus = sched_group_cpus(sg);
5503 sg_mask = sched_group_mask(sg);
5504 /* Try to find first idle cpu */
5505 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5506 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5513 if (balance_cpu == -1)
5514 balance_cpu = group_balance_cpu(sg);
5517 * First idle cpu or the first cpu(busiest) in this sched group
5518 * is eligible for doing load balancing at this and above domains.
5520 return balance_cpu == env->dst_cpu;
5524 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5525 * tasks if there is an imbalance.
5527 static int load_balance(int this_cpu, struct rq *this_rq,
5528 struct sched_domain *sd, enum cpu_idle_type idle,
5529 int *continue_balancing)
5531 int ld_moved, cur_ld_moved, active_balance = 0;
5532 struct sched_domain *sd_parent = sd->parent;
5533 struct sched_group *group;
5535 unsigned long flags;
5536 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5538 struct lb_env env = {
5540 .dst_cpu = this_cpu,
5542 .dst_grpmask = sched_group_cpus(sd->groups),
5544 .loop_break = sched_nr_migrate_break,
5549 * For NEWLY_IDLE load_balancing, we don't need to consider
5550 * other cpus in our group
5552 if (idle == CPU_NEWLY_IDLE)
5553 env.dst_grpmask = NULL;
5555 cpumask_copy(cpus, cpu_active_mask);
5557 schedstat_inc(sd, lb_count[idle]);
5560 if (!should_we_balance(&env)) {
5561 *continue_balancing = 0;
5565 group = find_busiest_group(&env);
5567 schedstat_inc(sd, lb_nobusyg[idle]);
5571 busiest = find_busiest_queue(&env, group);
5573 schedstat_inc(sd, lb_nobusyq[idle]);
5577 BUG_ON(busiest == env.dst_rq);
5579 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5582 if (busiest->nr_running > 1) {
5584 * Attempt to move tasks. If find_busiest_group has found
5585 * an imbalance but busiest->nr_running <= 1, the group is
5586 * still unbalanced. ld_moved simply stays zero, so it is
5587 * correctly treated as an imbalance.
5589 env.flags |= LBF_ALL_PINNED;
5590 env.src_cpu = busiest->cpu;
5591 env.src_rq = busiest;
5592 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5595 local_irq_save(flags);
5596 double_rq_lock(env.dst_rq, busiest);
5599 * cur_ld_moved - load moved in current iteration
5600 * ld_moved - cumulative load moved across iterations
5602 cur_ld_moved = move_tasks(&env);
5603 ld_moved += cur_ld_moved;
5604 double_rq_unlock(env.dst_rq, busiest);
5605 local_irq_restore(flags);
5608 * some other cpu did the load balance for us.
5610 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5611 resched_cpu(env.dst_cpu);
5613 if (env.flags & LBF_NEED_BREAK) {
5614 env.flags &= ~LBF_NEED_BREAK;
5619 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5620 * us and move them to an alternate dst_cpu in our sched_group
5621 * where they can run. The upper limit on how many times we
5622 * iterate on same src_cpu is dependent on number of cpus in our
5625 * This changes load balance semantics a bit on who can move
5626 * load to a given_cpu. In addition to the given_cpu itself
5627 * (or a ilb_cpu acting on its behalf where given_cpu is
5628 * nohz-idle), we now have balance_cpu in a position to move
5629 * load to given_cpu. In rare situations, this may cause
5630 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5631 * _independently_ and at _same_ time to move some load to
5632 * given_cpu) causing exceess load to be moved to given_cpu.
5633 * This however should not happen so much in practice and
5634 * moreover subsequent load balance cycles should correct the
5635 * excess load moved.
5637 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5639 /* Prevent to re-select dst_cpu via env's cpus */
5640 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5642 env.dst_rq = cpu_rq(env.new_dst_cpu);
5643 env.dst_cpu = env.new_dst_cpu;
5644 env.flags &= ~LBF_DST_PINNED;
5646 env.loop_break = sched_nr_migrate_break;
5649 * Go back to "more_balance" rather than "redo" since we
5650 * need to continue with same src_cpu.
5656 * We failed to reach balance because of affinity.
5659 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5661 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5662 *group_imbalance = 1;
5663 } else if (*group_imbalance)
5664 *group_imbalance = 0;
5667 /* All tasks on this runqueue were pinned by CPU affinity */
5668 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5669 cpumask_clear_cpu(cpu_of(busiest), cpus);
5670 if (!cpumask_empty(cpus)) {
5672 env.loop_break = sched_nr_migrate_break;
5680 schedstat_inc(sd, lb_failed[idle]);
5682 * Increment the failure counter only on periodic balance.
5683 * We do not want newidle balance, which can be very
5684 * frequent, pollute the failure counter causing
5685 * excessive cache_hot migrations and active balances.
5687 if (idle != CPU_NEWLY_IDLE)
5688 sd->nr_balance_failed++;
5690 if (need_active_balance(&env)) {
5691 raw_spin_lock_irqsave(&busiest->lock, flags);
5693 /* don't kick the active_load_balance_cpu_stop,
5694 * if the curr task on busiest cpu can't be
5697 if (!cpumask_test_cpu(this_cpu,
5698 tsk_cpus_allowed(busiest->curr))) {
5699 raw_spin_unlock_irqrestore(&busiest->lock,
5701 env.flags |= LBF_ALL_PINNED;
5702 goto out_one_pinned;
5706 * ->active_balance synchronizes accesses to
5707 * ->active_balance_work. Once set, it's cleared
5708 * only after active load balance is finished.
5710 if (!busiest->active_balance) {
5711 busiest->active_balance = 1;
5712 busiest->push_cpu = this_cpu;
5715 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5717 if (active_balance) {
5718 stop_one_cpu_nowait(cpu_of(busiest),
5719 active_load_balance_cpu_stop, busiest,
5720 &busiest->active_balance_work);
5724 * We've kicked active balancing, reset the failure
5727 sd->nr_balance_failed = sd->cache_nice_tries+1;
5730 sd->nr_balance_failed = 0;
5732 if (likely(!active_balance)) {
5733 /* We were unbalanced, so reset the balancing interval */
5734 sd->balance_interval = sd->min_interval;
5737 * If we've begun active balancing, start to back off. This
5738 * case may not be covered by the all_pinned logic if there
5739 * is only 1 task on the busy runqueue (because we don't call
5742 if (sd->balance_interval < sd->max_interval)
5743 sd->balance_interval *= 2;
5749 schedstat_inc(sd, lb_balanced[idle]);
5751 sd->nr_balance_failed = 0;
5754 /* tune up the balancing interval */
5755 if (((env.flags & LBF_ALL_PINNED) &&
5756 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5757 (sd->balance_interval < sd->max_interval))
5758 sd->balance_interval *= 2;
5766 * idle_balance is called by schedule() if this_cpu is about to become
5767 * idle. Attempts to pull tasks from other CPUs.
5769 void idle_balance(int this_cpu, struct rq *this_rq)
5771 struct sched_domain *sd;
5772 int pulled_task = 0;
5773 unsigned long next_balance = jiffies + HZ;
5776 this_rq->idle_stamp = rq_clock(this_rq);
5778 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5782 * Drop the rq->lock, but keep IRQ/preempt disabled.
5784 raw_spin_unlock(&this_rq->lock);
5786 update_blocked_averages(this_cpu);
5788 for_each_domain(this_cpu, sd) {
5789 unsigned long interval;
5790 int continue_balancing = 1;
5791 u64 t0, domain_cost;
5793 if (!(sd->flags & SD_LOAD_BALANCE))
5796 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5799 if (sd->flags & SD_BALANCE_NEWIDLE) {
5800 t0 = sched_clock_cpu(this_cpu);
5802 /* If we've pulled tasks over stop searching: */
5803 pulled_task = load_balance(this_cpu, this_rq,
5805 &continue_balancing);
5807 domain_cost = sched_clock_cpu(this_cpu) - t0;
5808 if (domain_cost > sd->max_newidle_lb_cost)
5809 sd->max_newidle_lb_cost = domain_cost;
5811 curr_cost += domain_cost;
5814 interval = msecs_to_jiffies(sd->balance_interval);
5815 if (time_after(next_balance, sd->last_balance + interval))
5816 next_balance = sd->last_balance + interval;
5818 this_rq->idle_stamp = 0;
5824 raw_spin_lock(&this_rq->lock);
5826 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5828 * We are going idle. next_balance may be set based on
5829 * a busy processor. So reset next_balance.
5831 this_rq->next_balance = next_balance;
5834 if (curr_cost > this_rq->max_idle_balance_cost)
5835 this_rq->max_idle_balance_cost = curr_cost;
5839 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5840 * running tasks off the busiest CPU onto idle CPUs. It requires at
5841 * least 1 task to be running on each physical CPU where possible, and
5842 * avoids physical / logical imbalances.
5844 static int active_load_balance_cpu_stop(void *data)
5846 struct rq *busiest_rq = data;
5847 int busiest_cpu = cpu_of(busiest_rq);
5848 int target_cpu = busiest_rq->push_cpu;
5849 struct rq *target_rq = cpu_rq(target_cpu);
5850 struct sched_domain *sd;
5852 raw_spin_lock_irq(&busiest_rq->lock);
5854 /* make sure the requested cpu hasn't gone down in the meantime */
5855 if (unlikely(busiest_cpu != smp_processor_id() ||
5856 !busiest_rq->active_balance))
5859 /* Is there any task to move? */
5860 if (busiest_rq->nr_running <= 1)
5864 * This condition is "impossible", if it occurs
5865 * we need to fix it. Originally reported by
5866 * Bjorn Helgaas on a 128-cpu setup.
5868 BUG_ON(busiest_rq == target_rq);
5870 /* move a task from busiest_rq to target_rq */
5871 double_lock_balance(busiest_rq, target_rq);
5873 /* Search for an sd spanning us and the target CPU. */
5875 for_each_domain(target_cpu, sd) {
5876 if ((sd->flags & SD_LOAD_BALANCE) &&
5877 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5882 struct lb_env env = {
5884 .dst_cpu = target_cpu,
5885 .dst_rq = target_rq,
5886 .src_cpu = busiest_rq->cpu,
5887 .src_rq = busiest_rq,
5891 schedstat_inc(sd, alb_count);
5893 if (move_one_task(&env))
5894 schedstat_inc(sd, alb_pushed);
5896 schedstat_inc(sd, alb_failed);
5899 double_unlock_balance(busiest_rq, target_rq);
5901 busiest_rq->active_balance = 0;
5902 raw_spin_unlock_irq(&busiest_rq->lock);
5906 #ifdef CONFIG_NO_HZ_COMMON
5908 * idle load balancing details
5909 * - When one of the busy CPUs notice that there may be an idle rebalancing
5910 * needed, they will kick the idle load balancer, which then does idle
5911 * load balancing for all the idle CPUs.
5914 cpumask_var_t idle_cpus_mask;
5916 unsigned long next_balance; /* in jiffy units */
5917 } nohz ____cacheline_aligned;
5919 static inline int find_new_ilb(int call_cpu)
5921 int ilb = cpumask_first(nohz.idle_cpus_mask);
5923 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5930 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5931 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5932 * CPU (if there is one).
5934 static void nohz_balancer_kick(int cpu)
5938 nohz.next_balance++;
5940 ilb_cpu = find_new_ilb(cpu);
5942 if (ilb_cpu >= nr_cpu_ids)
5945 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5948 * Use smp_send_reschedule() instead of resched_cpu().
5949 * This way we generate a sched IPI on the target cpu which
5950 * is idle. And the softirq performing nohz idle load balance
5951 * will be run before returning from the IPI.
5953 smp_send_reschedule(ilb_cpu);
5957 static inline void nohz_balance_exit_idle(int cpu)
5959 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5960 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5961 atomic_dec(&nohz.nr_cpus);
5962 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5966 static inline void set_cpu_sd_state_busy(void)
5968 struct sched_domain *sd;
5971 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5973 if (!sd || !sd->nohz_idle)
5977 for (; sd; sd = sd->parent)
5978 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5983 void set_cpu_sd_state_idle(void)
5985 struct sched_domain *sd;
5988 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5990 if (!sd || sd->nohz_idle)
5994 for (; sd; sd = sd->parent)
5995 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6001 * This routine will record that the cpu is going idle with tick stopped.
6002 * This info will be used in performing idle load balancing in the future.
6004 void nohz_balance_enter_idle(int cpu)
6007 * If this cpu is going down, then nothing needs to be done.
6009 if (!cpu_active(cpu))
6012 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6015 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6016 atomic_inc(&nohz.nr_cpus);
6017 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6020 static int sched_ilb_notifier(struct notifier_block *nfb,
6021 unsigned long action, void *hcpu)
6023 switch (action & ~CPU_TASKS_FROZEN) {
6025 nohz_balance_exit_idle(smp_processor_id());
6033 static DEFINE_SPINLOCK(balancing);
6036 * Scale the max load_balance interval with the number of CPUs in the system.
6037 * This trades load-balance latency on larger machines for less cross talk.
6039 void update_max_interval(void)
6041 max_load_balance_interval = HZ*num_online_cpus()/10;
6045 * It checks each scheduling domain to see if it is due to be balanced,
6046 * and initiates a balancing operation if so.
6048 * Balancing parameters are set up in init_sched_domains.
6050 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6052 int continue_balancing = 1;
6053 struct rq *rq = cpu_rq(cpu);
6054 unsigned long interval;
6055 struct sched_domain *sd;
6056 /* Earliest time when we have to do rebalance again */
6057 unsigned long next_balance = jiffies + 60*HZ;
6058 int update_next_balance = 0;
6059 int need_serialize, need_decay = 0;
6062 update_blocked_averages(cpu);
6065 for_each_domain(cpu, sd) {
6067 * Decay the newidle max times here because this is a regular
6068 * visit to all the domains. Decay ~1% per second.
6070 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6071 sd->max_newidle_lb_cost =
6072 (sd->max_newidle_lb_cost * 253) / 256;
6073 sd->next_decay_max_lb_cost = jiffies + HZ;
6076 max_cost += sd->max_newidle_lb_cost;
6078 if (!(sd->flags & SD_LOAD_BALANCE))
6082 * Stop the load balance at this level. There is another
6083 * CPU in our sched group which is doing load balancing more
6086 if (!continue_balancing) {
6092 interval = sd->balance_interval;
6093 if (idle != CPU_IDLE)
6094 interval *= sd->busy_factor;
6096 /* scale ms to jiffies */
6097 interval = msecs_to_jiffies(interval);
6098 interval = clamp(interval, 1UL, max_load_balance_interval);
6100 need_serialize = sd->flags & SD_SERIALIZE;
6102 if (need_serialize) {
6103 if (!spin_trylock(&balancing))
6107 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6108 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6110 * The LBF_DST_PINNED logic could have changed
6111 * env->dst_cpu, so we can't know our idle
6112 * state even if we migrated tasks. Update it.
6114 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6116 sd->last_balance = jiffies;
6119 spin_unlock(&balancing);
6121 if (time_after(next_balance, sd->last_balance + interval)) {
6122 next_balance = sd->last_balance + interval;
6123 update_next_balance = 1;
6128 * Ensure the rq-wide value also decays but keep it at a
6129 * reasonable floor to avoid funnies with rq->avg_idle.
6131 rq->max_idle_balance_cost =
6132 max((u64)sysctl_sched_migration_cost, max_cost);
6137 * next_balance will be updated only when there is a need.
6138 * When the cpu is attached to null domain for ex, it will not be
6141 if (likely(update_next_balance))
6142 rq->next_balance = next_balance;
6145 #ifdef CONFIG_NO_HZ_COMMON
6147 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6148 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6150 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6152 struct rq *this_rq = cpu_rq(this_cpu);
6156 if (idle != CPU_IDLE ||
6157 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6160 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6161 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6165 * If this cpu gets work to do, stop the load balancing
6166 * work being done for other cpus. Next load
6167 * balancing owner will pick it up.
6172 rq = cpu_rq(balance_cpu);
6174 raw_spin_lock_irq(&rq->lock);
6175 update_rq_clock(rq);
6176 update_idle_cpu_load(rq);
6177 raw_spin_unlock_irq(&rq->lock);
6179 rebalance_domains(balance_cpu, CPU_IDLE);
6181 if (time_after(this_rq->next_balance, rq->next_balance))
6182 this_rq->next_balance = rq->next_balance;
6184 nohz.next_balance = this_rq->next_balance;
6186 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6190 * Current heuristic for kicking the idle load balancer in the presence
6191 * of an idle cpu is the system.
6192 * - This rq has more than one task.
6193 * - At any scheduler domain level, this cpu's scheduler group has multiple
6194 * busy cpu's exceeding the group's power.
6195 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6196 * domain span are idle.
6198 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6200 unsigned long now = jiffies;
6201 struct sched_domain *sd;
6203 if (unlikely(idle_cpu(cpu)))
6207 * We may be recently in ticked or tickless idle mode. At the first
6208 * busy tick after returning from idle, we will update the busy stats.
6210 set_cpu_sd_state_busy();
6211 nohz_balance_exit_idle(cpu);
6214 * None are in tickless mode and hence no need for NOHZ idle load
6217 if (likely(!atomic_read(&nohz.nr_cpus)))
6220 if (time_before(now, nohz.next_balance))
6223 if (rq->nr_running >= 2)
6227 for_each_domain(cpu, sd) {
6228 struct sched_group *sg = sd->groups;
6229 struct sched_group_power *sgp = sg->sgp;
6230 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6232 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6233 goto need_kick_unlock;
6235 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6236 && (cpumask_first_and(nohz.idle_cpus_mask,
6237 sched_domain_span(sd)) < cpu))
6238 goto need_kick_unlock;
6240 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6252 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6256 * run_rebalance_domains is triggered when needed from the scheduler tick.
6257 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6259 static void run_rebalance_domains(struct softirq_action *h)
6261 int this_cpu = smp_processor_id();
6262 struct rq *this_rq = cpu_rq(this_cpu);
6263 enum cpu_idle_type idle = this_rq->idle_balance ?
6264 CPU_IDLE : CPU_NOT_IDLE;
6266 rebalance_domains(this_cpu, idle);
6269 * If this cpu has a pending nohz_balance_kick, then do the
6270 * balancing on behalf of the other idle cpus whose ticks are
6273 nohz_idle_balance(this_cpu, idle);
6276 static inline int on_null_domain(int cpu)
6278 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6282 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6284 void trigger_load_balance(struct rq *rq, int cpu)
6286 /* Don't need to rebalance while attached to NULL domain */
6287 if (time_after_eq(jiffies, rq->next_balance) &&
6288 likely(!on_null_domain(cpu)))
6289 raise_softirq(SCHED_SOFTIRQ);
6290 #ifdef CONFIG_NO_HZ_COMMON
6291 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6292 nohz_balancer_kick(cpu);
6296 static void rq_online_fair(struct rq *rq)
6301 static void rq_offline_fair(struct rq *rq)
6305 /* Ensure any throttled groups are reachable by pick_next_task */
6306 unthrottle_offline_cfs_rqs(rq);
6309 #endif /* CONFIG_SMP */
6312 * scheduler tick hitting a task of our scheduling class:
6314 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6316 struct cfs_rq *cfs_rq;
6317 struct sched_entity *se = &curr->se;
6319 for_each_sched_entity(se) {
6320 cfs_rq = cfs_rq_of(se);
6321 entity_tick(cfs_rq, se, queued);
6324 if (numabalancing_enabled)
6325 task_tick_numa(rq, curr);
6327 update_rq_runnable_avg(rq, 1);
6331 * called on fork with the child task as argument from the parent's context
6332 * - child not yet on the tasklist
6333 * - preemption disabled
6335 static void task_fork_fair(struct task_struct *p)
6337 struct cfs_rq *cfs_rq;
6338 struct sched_entity *se = &p->se, *curr;
6339 int this_cpu = smp_processor_id();
6340 struct rq *rq = this_rq();
6341 unsigned long flags;
6343 raw_spin_lock_irqsave(&rq->lock, flags);
6345 update_rq_clock(rq);
6347 cfs_rq = task_cfs_rq(current);
6348 curr = cfs_rq->curr;
6351 * Not only the cpu but also the task_group of the parent might have
6352 * been changed after parent->se.parent,cfs_rq were copied to
6353 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6354 * of child point to valid ones.
6357 __set_task_cpu(p, this_cpu);
6360 update_curr(cfs_rq);
6363 se->vruntime = curr->vruntime;
6364 place_entity(cfs_rq, se, 1);
6366 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6368 * Upon rescheduling, sched_class::put_prev_task() will place
6369 * 'current' within the tree based on its new key value.
6371 swap(curr->vruntime, se->vruntime);
6372 resched_task(rq->curr);
6375 se->vruntime -= cfs_rq->min_vruntime;
6377 raw_spin_unlock_irqrestore(&rq->lock, flags);
6381 * Priority of the task has changed. Check to see if we preempt
6385 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6391 * Reschedule if we are currently running on this runqueue and
6392 * our priority decreased, or if we are not currently running on
6393 * this runqueue and our priority is higher than the current's
6395 if (rq->curr == p) {
6396 if (p->prio > oldprio)
6397 resched_task(rq->curr);
6399 check_preempt_curr(rq, p, 0);
6402 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6404 struct sched_entity *se = &p->se;
6405 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6408 * Ensure the task's vruntime is normalized, so that when its
6409 * switched back to the fair class the enqueue_entity(.flags=0) will
6410 * do the right thing.
6412 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6413 * have normalized the vruntime, if it was !on_rq, then only when
6414 * the task is sleeping will it still have non-normalized vruntime.
6416 if (!se->on_rq && p->state != TASK_RUNNING) {
6418 * Fix up our vruntime so that the current sleep doesn't
6419 * cause 'unlimited' sleep bonus.
6421 place_entity(cfs_rq, se, 0);
6422 se->vruntime -= cfs_rq->min_vruntime;
6427 * Remove our load from contribution when we leave sched_fair
6428 * and ensure we don't carry in an old decay_count if we
6431 if (se->avg.decay_count) {
6432 __synchronize_entity_decay(se);
6433 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6439 * We switched to the sched_fair class.
6441 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6447 * We were most likely switched from sched_rt, so
6448 * kick off the schedule if running, otherwise just see
6449 * if we can still preempt the current task.
6452 resched_task(rq->curr);
6454 check_preempt_curr(rq, p, 0);
6457 /* Account for a task changing its policy or group.
6459 * This routine is mostly called to set cfs_rq->curr field when a task
6460 * migrates between groups/classes.
6462 static void set_curr_task_fair(struct rq *rq)
6464 struct sched_entity *se = &rq->curr->se;
6466 for_each_sched_entity(se) {
6467 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6469 set_next_entity(cfs_rq, se);
6470 /* ensure bandwidth has been allocated on our new cfs_rq */
6471 account_cfs_rq_runtime(cfs_rq, 0);
6475 void init_cfs_rq(struct cfs_rq *cfs_rq)
6477 cfs_rq->tasks_timeline = RB_ROOT;
6478 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6479 #ifndef CONFIG_64BIT
6480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6483 atomic64_set(&cfs_rq->decay_counter, 1);
6484 atomic_long_set(&cfs_rq->removed_load, 0);
6488 #ifdef CONFIG_FAIR_GROUP_SCHED
6489 static void task_move_group_fair(struct task_struct *p, int on_rq)
6491 struct cfs_rq *cfs_rq;
6493 * If the task was not on the rq at the time of this cgroup movement
6494 * it must have been asleep, sleeping tasks keep their ->vruntime
6495 * absolute on their old rq until wakeup (needed for the fair sleeper
6496 * bonus in place_entity()).
6498 * If it was on the rq, we've just 'preempted' it, which does convert
6499 * ->vruntime to a relative base.
6501 * Make sure both cases convert their relative position when migrating
6502 * to another cgroup's rq. This does somewhat interfere with the
6503 * fair sleeper stuff for the first placement, but who cares.
6506 * When !on_rq, vruntime of the task has usually NOT been normalized.
6507 * But there are some cases where it has already been normalized:
6509 * - Moving a forked child which is waiting for being woken up by
6510 * wake_up_new_task().
6511 * - Moving a task which has been woken up by try_to_wake_up() and
6512 * waiting for actually being woken up by sched_ttwu_pending().
6514 * To prevent boost or penalty in the new cfs_rq caused by delta
6515 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6517 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6521 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6522 set_task_rq(p, task_cpu(p));
6524 cfs_rq = cfs_rq_of(&p->se);
6525 p->se.vruntime += cfs_rq->min_vruntime;
6528 * migrate_task_rq_fair() will have removed our previous
6529 * contribution, but we must synchronize for ongoing future
6532 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6533 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6538 void free_fair_sched_group(struct task_group *tg)
6542 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6544 for_each_possible_cpu(i) {
6546 kfree(tg->cfs_rq[i]);
6555 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6557 struct cfs_rq *cfs_rq;
6558 struct sched_entity *se;
6561 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6564 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6568 tg->shares = NICE_0_LOAD;
6570 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6572 for_each_possible_cpu(i) {
6573 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6574 GFP_KERNEL, cpu_to_node(i));
6578 se = kzalloc_node(sizeof(struct sched_entity),
6579 GFP_KERNEL, cpu_to_node(i));
6583 init_cfs_rq(cfs_rq);
6584 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6595 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6597 struct rq *rq = cpu_rq(cpu);
6598 unsigned long flags;
6601 * Only empty task groups can be destroyed; so we can speculatively
6602 * check on_list without danger of it being re-added.
6604 if (!tg->cfs_rq[cpu]->on_list)
6607 raw_spin_lock_irqsave(&rq->lock, flags);
6608 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6609 raw_spin_unlock_irqrestore(&rq->lock, flags);
6612 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6613 struct sched_entity *se, int cpu,
6614 struct sched_entity *parent)
6616 struct rq *rq = cpu_rq(cpu);
6620 init_cfs_rq_runtime(cfs_rq);
6622 tg->cfs_rq[cpu] = cfs_rq;
6625 /* se could be NULL for root_task_group */
6630 se->cfs_rq = &rq->cfs;
6632 se->cfs_rq = parent->my_q;
6635 update_load_set(&se->load, 0);
6636 se->parent = parent;
6639 static DEFINE_MUTEX(shares_mutex);
6641 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6644 unsigned long flags;
6647 * We can't change the weight of the root cgroup.
6652 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6654 mutex_lock(&shares_mutex);
6655 if (tg->shares == shares)
6658 tg->shares = shares;
6659 for_each_possible_cpu(i) {
6660 struct rq *rq = cpu_rq(i);
6661 struct sched_entity *se;
6664 /* Propagate contribution to hierarchy */
6665 raw_spin_lock_irqsave(&rq->lock, flags);
6667 /* Possible calls to update_curr() need rq clock */
6668 update_rq_clock(rq);
6669 for_each_sched_entity(se)
6670 update_cfs_shares(group_cfs_rq(se));
6671 raw_spin_unlock_irqrestore(&rq->lock, flags);
6675 mutex_unlock(&shares_mutex);
6678 #else /* CONFIG_FAIR_GROUP_SCHED */
6680 void free_fair_sched_group(struct task_group *tg) { }
6682 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6687 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6689 #endif /* CONFIG_FAIR_GROUP_SCHED */
6692 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6694 struct sched_entity *se = &task->se;
6695 unsigned int rr_interval = 0;
6698 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6701 if (rq->cfs.load.weight)
6702 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6708 * All the scheduling class methods:
6710 const struct sched_class fair_sched_class = {
6711 .next = &idle_sched_class,
6712 .enqueue_task = enqueue_task_fair,
6713 .dequeue_task = dequeue_task_fair,
6714 .yield_task = yield_task_fair,
6715 .yield_to_task = yield_to_task_fair,
6717 .check_preempt_curr = check_preempt_wakeup,
6719 .pick_next_task = pick_next_task_fair,
6720 .put_prev_task = put_prev_task_fair,
6723 .select_task_rq = select_task_rq_fair,
6724 .migrate_task_rq = migrate_task_rq_fair,
6726 .rq_online = rq_online_fair,
6727 .rq_offline = rq_offline_fair,
6729 .task_waking = task_waking_fair,
6732 .set_curr_task = set_curr_task_fair,
6733 .task_tick = task_tick_fair,
6734 .task_fork = task_fork_fair,
6736 .prio_changed = prio_changed_fair,
6737 .switched_from = switched_from_fair,
6738 .switched_to = switched_to_fair,
6740 .get_rr_interval = get_rr_interval_fair,
6742 #ifdef CONFIG_FAIR_GROUP_SCHED
6743 .task_move_group = task_move_group_fair,
6747 #ifdef CONFIG_SCHED_DEBUG
6748 void print_cfs_stats(struct seq_file *m, int cpu)
6750 struct cfs_rq *cfs_rq;
6753 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6754 print_cfs_rq(m, cpu, cfs_rq);
6759 __init void init_sched_fair_class(void)
6762 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6764 #ifdef CONFIG_NO_HZ_COMMON
6765 nohz.next_balance = jiffies;
6766 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6767 cpu_notifier(sched_ilb_notifier, 0);