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
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
40 * Targeted preemption latency for CPU-bound tasks:
41 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 unsigned int sysctl_sched_latency = 6000000ULL;
52 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 unsigned int sysctl_sched_is_big_little = 0;
55 unsigned int sysctl_sched_sync_hint_enable = 1;
56 unsigned int sysctl_sched_initial_task_util = 0;
57 unsigned int sysctl_sched_cstate_aware = 1;
60 * The initial- and re-scaling of tunables is configurable
61 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
64 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
65 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
66 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
68 enum sched_tunable_scaling sysctl_sched_tunable_scaling
69 = SCHED_TUNABLESCALING_LOG;
72 * Minimal preemption granularity for CPU-bound tasks:
73 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
75 unsigned int sysctl_sched_min_granularity = 750000ULL;
76 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
79 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
81 static unsigned int sched_nr_latency = 8;
84 * After fork, child runs first. If set to 0 (default) then
85 * parent will (try to) run first.
87 unsigned int sysctl_sched_child_runs_first __read_mostly;
90 * SCHED_OTHER wake-up granularity.
91 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
93 * This option delays the preemption effects of decoupled workloads
94 * and reduces their over-scheduling. Synchronous workloads will still
95 * have immediate wakeup/sleep latencies.
97 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
98 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
100 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
103 * The exponential sliding window over which load is averaged for shares
107 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
109 #ifdef CONFIG_CFS_BANDWIDTH
111 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
112 * each time a cfs_rq requests quota.
114 * Note: in the case that the slice exceeds the runtime remaining (either due
115 * to consumption or the quota being specified to be smaller than the slice)
116 * we will always only issue the remaining available time.
118 * default: 5 msec, units: microseconds
120 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
123 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
129 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
135 static inline void update_load_set(struct load_weight *lw, unsigned long w)
142 * Increase the granularity value when there are more CPUs,
143 * because with more CPUs the 'effective latency' as visible
144 * to users decreases. But the relationship is not linear,
145 * so pick a second-best guess by going with the log2 of the
148 * This idea comes from the SD scheduler of Con Kolivas:
150 static unsigned int get_update_sysctl_factor(void)
152 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
155 switch (sysctl_sched_tunable_scaling) {
156 case SCHED_TUNABLESCALING_NONE:
159 case SCHED_TUNABLESCALING_LINEAR:
162 case SCHED_TUNABLESCALING_LOG:
164 factor = 1 + ilog2(cpus);
171 static void update_sysctl(void)
173 unsigned int factor = get_update_sysctl_factor();
175 #define SET_SYSCTL(name) \
176 (sysctl_##name = (factor) * normalized_sysctl_##name)
177 SET_SYSCTL(sched_min_granularity);
178 SET_SYSCTL(sched_latency);
179 SET_SYSCTL(sched_wakeup_granularity);
183 void sched_init_granularity(void)
188 #define WMULT_CONST (~0U)
189 #define WMULT_SHIFT 32
191 static void __update_inv_weight(struct load_weight *lw)
195 if (likely(lw->inv_weight))
198 w = scale_load_down(lw->weight);
200 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
202 else if (unlikely(!w))
203 lw->inv_weight = WMULT_CONST;
205 lw->inv_weight = WMULT_CONST / w;
209 * delta_exec * weight / lw.weight
211 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
213 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
214 * we're guaranteed shift stays positive because inv_weight is guaranteed to
215 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
217 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
218 * weight/lw.weight <= 1, and therefore our shift will also be positive.
220 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
222 u64 fact = scale_load_down(weight);
223 int shift = WMULT_SHIFT;
225 __update_inv_weight(lw);
227 if (unlikely(fact >> 32)) {
234 /* hint to use a 32x32->64 mul */
235 fact = (u64)(u32)fact * lw->inv_weight;
242 return mul_u64_u32_shr(delta_exec, fact, shift);
246 const struct sched_class fair_sched_class;
248 /**************************************************************
249 * CFS operations on generic schedulable entities:
252 #ifdef CONFIG_FAIR_GROUP_SCHED
254 /* cpu runqueue to which this cfs_rq is attached */
255 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
260 /* An entity is a task if it doesn't "own" a runqueue */
261 #define entity_is_task(se) (!se->my_q)
263 static inline struct task_struct *task_of(struct sched_entity *se)
265 #ifdef CONFIG_SCHED_DEBUG
266 WARN_ON_ONCE(!entity_is_task(se));
268 return container_of(se, struct task_struct, se);
271 /* Walk up scheduling entities hierarchy */
272 #define for_each_sched_entity(se) \
273 for (; se; se = se->parent)
275 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
280 /* runqueue on which this entity is (to be) queued */
281 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
286 /* runqueue "owned" by this group */
287 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
292 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 if (!cfs_rq->on_list) {
296 * Ensure we either appear before our parent (if already
297 * enqueued) or force our parent to appear after us when it is
298 * enqueued. The fact that we always enqueue bottom-up
299 * reduces this to two cases.
301 if (cfs_rq->tg->parent &&
302 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
303 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
304 &rq_of(cfs_rq)->leaf_cfs_rq_list);
306 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
307 &rq_of(cfs_rq)->leaf_cfs_rq_list);
314 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
316 if (cfs_rq->on_list) {
317 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
322 /* Iterate thr' all leaf cfs_rq's on a runqueue */
323 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
324 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
326 /* Do the two (enqueued) entities belong to the same group ? */
327 static inline struct cfs_rq *
328 is_same_group(struct sched_entity *se, struct sched_entity *pse)
330 if (se->cfs_rq == pse->cfs_rq)
336 static inline struct sched_entity *parent_entity(struct sched_entity *se)
342 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
344 int se_depth, pse_depth;
347 * preemption test can be made between sibling entities who are in the
348 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
349 * both tasks until we find their ancestors who are siblings of common
353 /* First walk up until both entities are at same depth */
354 se_depth = (*se)->depth;
355 pse_depth = (*pse)->depth;
357 while (se_depth > pse_depth) {
359 *se = parent_entity(*se);
362 while (pse_depth > se_depth) {
364 *pse = parent_entity(*pse);
367 while (!is_same_group(*se, *pse)) {
368 *se = parent_entity(*se);
369 *pse = parent_entity(*pse);
373 #else /* !CONFIG_FAIR_GROUP_SCHED */
375 static inline struct task_struct *task_of(struct sched_entity *se)
377 return container_of(se, struct task_struct, se);
380 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
382 return container_of(cfs_rq, struct rq, cfs);
385 #define entity_is_task(se) 1
387 #define for_each_sched_entity(se) \
388 for (; se; se = NULL)
390 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
392 return &task_rq(p)->cfs;
395 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
397 struct task_struct *p = task_of(se);
398 struct rq *rq = task_rq(p);
403 /* runqueue "owned" by this group */
404 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
409 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
413 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
417 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
418 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
420 static inline struct sched_entity *parent_entity(struct sched_entity *se)
426 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
430 #endif /* CONFIG_FAIR_GROUP_SCHED */
432 static __always_inline
433 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
435 /**************************************************************
436 * Scheduling class tree data structure manipulation methods:
439 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
441 s64 delta = (s64)(vruntime - max_vruntime);
443 max_vruntime = vruntime;
448 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
450 s64 delta = (s64)(vruntime - min_vruntime);
452 min_vruntime = vruntime;
457 static inline int entity_before(struct sched_entity *a,
458 struct sched_entity *b)
460 return (s64)(a->vruntime - b->vruntime) < 0;
463 static void update_min_vruntime(struct cfs_rq *cfs_rq)
465 u64 vruntime = cfs_rq->min_vruntime;
468 vruntime = cfs_rq->curr->vruntime;
470 if (cfs_rq->rb_leftmost) {
471 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
476 vruntime = se->vruntime;
478 vruntime = min_vruntime(vruntime, se->vruntime);
481 /* ensure we never gain time by being placed backwards. */
482 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
485 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
490 * Enqueue an entity into the rb-tree:
492 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
494 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
495 struct rb_node *parent = NULL;
496 struct sched_entity *entry;
500 * Find the right place in the rbtree:
504 entry = rb_entry(parent, struct sched_entity, run_node);
506 * We dont care about collisions. Nodes with
507 * the same key stay together.
509 if (entity_before(se, entry)) {
510 link = &parent->rb_left;
512 link = &parent->rb_right;
518 * Maintain a cache of leftmost tree entries (it is frequently
522 cfs_rq->rb_leftmost = &se->run_node;
524 rb_link_node(&se->run_node, parent, link);
525 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
528 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
530 if (cfs_rq->rb_leftmost == &se->run_node) {
531 struct rb_node *next_node;
533 next_node = rb_next(&se->run_node);
534 cfs_rq->rb_leftmost = next_node;
537 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
540 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
542 struct rb_node *left = cfs_rq->rb_leftmost;
547 return rb_entry(left, struct sched_entity, run_node);
550 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
552 struct rb_node *next = rb_next(&se->run_node);
557 return rb_entry(next, struct sched_entity, run_node);
560 #ifdef CONFIG_SCHED_DEBUG
561 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
563 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
568 return rb_entry(last, struct sched_entity, run_node);
571 /**************************************************************
572 * Scheduling class statistics methods:
575 int sched_proc_update_handler(struct ctl_table *table, int write,
576 void __user *buffer, size_t *lenp,
579 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
580 unsigned int factor = get_update_sysctl_factor();
585 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
586 sysctl_sched_min_granularity);
588 #define WRT_SYSCTL(name) \
589 (normalized_sysctl_##name = sysctl_##name / (factor))
590 WRT_SYSCTL(sched_min_granularity);
591 WRT_SYSCTL(sched_latency);
592 WRT_SYSCTL(sched_wakeup_granularity);
602 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
604 if (unlikely(se->load.weight != NICE_0_LOAD))
605 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
611 * The idea is to set a period in which each task runs once.
613 * When there are too many tasks (sched_nr_latency) we have to stretch
614 * this period because otherwise the slices get too small.
616 * p = (nr <= nl) ? l : l*nr/nl
618 static u64 __sched_period(unsigned long nr_running)
620 if (unlikely(nr_running > sched_nr_latency))
621 return nr_running * sysctl_sched_min_granularity;
623 return sysctl_sched_latency;
627 * We calculate the wall-time slice from the period by taking a part
628 * proportional to the weight.
632 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
634 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
636 for_each_sched_entity(se) {
637 struct load_weight *load;
638 struct load_weight lw;
640 cfs_rq = cfs_rq_of(se);
641 load = &cfs_rq->load;
643 if (unlikely(!se->on_rq)) {
646 update_load_add(&lw, se->load.weight);
649 slice = __calc_delta(slice, se->load.weight, load);
655 * We calculate the vruntime slice of a to-be-inserted task.
659 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
661 return calc_delta_fair(sched_slice(cfs_rq, se), se);
665 static int select_idle_sibling(struct task_struct *p, int cpu);
666 static unsigned long task_h_load(struct task_struct *p);
669 * We choose a half-life close to 1 scheduling period.
670 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
671 * dependent on this value.
673 #define LOAD_AVG_PERIOD 32
674 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
675 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
677 /* Give new sched_entity start runnable values to heavy its load in infant time */
678 void init_entity_runnable_average(struct sched_entity *se)
680 struct sched_avg *sa = &se->avg;
682 sa->last_update_time = 0;
684 * sched_avg's period_contrib should be strictly less then 1024, so
685 * we give it 1023 to make sure it is almost a period (1024us), and
686 * will definitely be update (after enqueue).
688 sa->period_contrib = 1023;
689 sa->load_avg = scale_load_down(se->load.weight);
690 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
691 sa->util_avg = sched_freq() ?
692 sysctl_sched_initial_task_util :
693 scale_load_down(SCHED_LOAD_SCALE);
694 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
695 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
699 void init_entity_runnable_average(struct sched_entity *se)
705 * Update the current task's runtime statistics.
707 static void update_curr(struct cfs_rq *cfs_rq)
709 struct sched_entity *curr = cfs_rq->curr;
710 u64 now = rq_clock_task(rq_of(cfs_rq));
716 delta_exec = now - curr->exec_start;
717 if (unlikely((s64)delta_exec <= 0))
720 curr->exec_start = now;
722 schedstat_set(curr->statistics.exec_max,
723 max(delta_exec, curr->statistics.exec_max));
725 curr->sum_exec_runtime += delta_exec;
726 schedstat_add(cfs_rq, exec_clock, delta_exec);
728 curr->vruntime += calc_delta_fair(delta_exec, curr);
729 update_min_vruntime(cfs_rq);
731 if (entity_is_task(curr)) {
732 struct task_struct *curtask = task_of(curr);
734 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
735 cpuacct_charge(curtask, delta_exec);
736 account_group_exec_runtime(curtask, delta_exec);
739 account_cfs_rq_runtime(cfs_rq, delta_exec);
742 static void update_curr_fair(struct rq *rq)
744 update_curr(cfs_rq_of(&rq->curr->se));
748 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
750 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
754 * Task is being enqueued - update stats:
756 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 * Are we enqueueing a waiting task? (for current tasks
760 * a dequeue/enqueue event is a NOP)
762 if (se != cfs_rq->curr)
763 update_stats_wait_start(cfs_rq, se);
767 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
769 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
770 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
771 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
772 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
773 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
774 #ifdef CONFIG_SCHEDSTATS
775 if (entity_is_task(se)) {
776 trace_sched_stat_wait(task_of(se),
777 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
780 schedstat_set(se->statistics.wait_start, 0);
784 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
787 * Mark the end of the wait period if dequeueing a
790 if (se != cfs_rq->curr)
791 update_stats_wait_end(cfs_rq, se);
795 * We are picking a new current task - update its stats:
798 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
801 * We are starting a new run period:
803 se->exec_start = rq_clock_task(rq_of(cfs_rq));
806 /**************************************************
807 * Scheduling class queueing methods:
810 #ifdef CONFIG_NUMA_BALANCING
812 * Approximate time to scan a full NUMA task in ms. The task scan period is
813 * calculated based on the tasks virtual memory size and
814 * numa_balancing_scan_size.
816 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
817 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
819 /* Portion of address space to scan in MB */
820 unsigned int sysctl_numa_balancing_scan_size = 256;
822 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
823 unsigned int sysctl_numa_balancing_scan_delay = 1000;
825 static unsigned int task_nr_scan_windows(struct task_struct *p)
827 unsigned long rss = 0;
828 unsigned long nr_scan_pages;
831 * Calculations based on RSS as non-present and empty pages are skipped
832 * by the PTE scanner and NUMA hinting faults should be trapped based
835 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
836 rss = get_mm_rss(p->mm);
840 rss = round_up(rss, nr_scan_pages);
841 return rss / nr_scan_pages;
844 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
845 #define MAX_SCAN_WINDOW 2560
847 static unsigned int task_scan_min(struct task_struct *p)
849 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
850 unsigned int scan, floor;
851 unsigned int windows = 1;
853 if (scan_size < MAX_SCAN_WINDOW)
854 windows = MAX_SCAN_WINDOW / scan_size;
855 floor = 1000 / windows;
857 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
858 return max_t(unsigned int, floor, scan);
861 static unsigned int task_scan_max(struct task_struct *p)
863 unsigned int smin = task_scan_min(p);
866 /* Watch for min being lower than max due to floor calculations */
867 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
868 return max(smin, smax);
871 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
873 rq->nr_numa_running += (p->numa_preferred_nid != -1);
874 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
877 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
879 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
880 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
886 spinlock_t lock; /* nr_tasks, tasks */
891 nodemask_t active_nodes;
892 unsigned long total_faults;
894 * Faults_cpu is used to decide whether memory should move
895 * towards the CPU. As a consequence, these stats are weighted
896 * more by CPU use than by memory faults.
898 unsigned long *faults_cpu;
899 unsigned long faults[0];
902 /* Shared or private faults. */
903 #define NR_NUMA_HINT_FAULT_TYPES 2
905 /* Memory and CPU locality */
906 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
908 /* Averaged statistics, and temporary buffers. */
909 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
911 pid_t task_numa_group_id(struct task_struct *p)
913 return p->numa_group ? p->numa_group->gid : 0;
917 * The averaged statistics, shared & private, memory & cpu,
918 * occupy the first half of the array. The second half of the
919 * array is for current counters, which are averaged into the
920 * first set by task_numa_placement.
922 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
924 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
927 static inline unsigned long task_faults(struct task_struct *p, int nid)
932 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
933 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
936 static inline unsigned long group_faults(struct task_struct *p, int nid)
941 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
942 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
945 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
947 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
948 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
951 /* Handle placement on systems where not all nodes are directly connected. */
952 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
953 int maxdist, bool task)
955 unsigned long score = 0;
959 * All nodes are directly connected, and the same distance
960 * from each other. No need for fancy placement algorithms.
962 if (sched_numa_topology_type == NUMA_DIRECT)
966 * This code is called for each node, introducing N^2 complexity,
967 * which should be ok given the number of nodes rarely exceeds 8.
969 for_each_online_node(node) {
970 unsigned long faults;
971 int dist = node_distance(nid, node);
974 * The furthest away nodes in the system are not interesting
975 * for placement; nid was already counted.
977 if (dist == sched_max_numa_distance || node == nid)
981 * On systems with a backplane NUMA topology, compare groups
982 * of nodes, and move tasks towards the group with the most
983 * memory accesses. When comparing two nodes at distance
984 * "hoplimit", only nodes closer by than "hoplimit" are part
985 * of each group. Skip other nodes.
987 if (sched_numa_topology_type == NUMA_BACKPLANE &&
991 /* Add up the faults from nearby nodes. */
993 faults = task_faults(p, node);
995 faults = group_faults(p, node);
998 * On systems with a glueless mesh NUMA topology, there are
999 * no fixed "groups of nodes". Instead, nodes that are not
1000 * directly connected bounce traffic through intermediate
1001 * nodes; a numa_group can occupy any set of nodes.
1002 * The further away a node is, the less the faults count.
1003 * This seems to result in good task placement.
1005 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1006 faults *= (sched_max_numa_distance - dist);
1007 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1017 * These return the fraction of accesses done by a particular task, or
1018 * task group, on a particular numa node. The group weight is given a
1019 * larger multiplier, in order to group tasks together that are almost
1020 * evenly spread out between numa nodes.
1022 static inline unsigned long task_weight(struct task_struct *p, int nid,
1025 unsigned long faults, total_faults;
1027 if (!p->numa_faults)
1030 total_faults = p->total_numa_faults;
1035 faults = task_faults(p, nid);
1036 faults += score_nearby_nodes(p, nid, dist, true);
1038 return 1000 * faults / total_faults;
1041 static inline unsigned long group_weight(struct task_struct *p, int nid,
1044 unsigned long faults, total_faults;
1049 total_faults = p->numa_group->total_faults;
1054 faults = group_faults(p, nid);
1055 faults += score_nearby_nodes(p, nid, dist, false);
1057 return 1000 * faults / total_faults;
1060 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1061 int src_nid, int dst_cpu)
1063 struct numa_group *ng = p->numa_group;
1064 int dst_nid = cpu_to_node(dst_cpu);
1065 int last_cpupid, this_cpupid;
1067 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1070 * Multi-stage node selection is used in conjunction with a periodic
1071 * migration fault to build a temporal task<->page relation. By using
1072 * a two-stage filter we remove short/unlikely relations.
1074 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1075 * a task's usage of a particular page (n_p) per total usage of this
1076 * page (n_t) (in a given time-span) to a probability.
1078 * Our periodic faults will sample this probability and getting the
1079 * same result twice in a row, given these samples are fully
1080 * independent, is then given by P(n)^2, provided our sample period
1081 * is sufficiently short compared to the usage pattern.
1083 * This quadric squishes small probabilities, making it less likely we
1084 * act on an unlikely task<->page relation.
1086 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1087 if (!cpupid_pid_unset(last_cpupid) &&
1088 cpupid_to_nid(last_cpupid) != dst_nid)
1091 /* Always allow migrate on private faults */
1092 if (cpupid_match_pid(p, last_cpupid))
1095 /* A shared fault, but p->numa_group has not been set up yet. */
1100 * Do not migrate if the destination is not a node that
1101 * is actively used by this numa group.
1103 if (!node_isset(dst_nid, ng->active_nodes))
1107 * Source is a node that is not actively used by this
1108 * numa group, while the destination is. Migrate.
1110 if (!node_isset(src_nid, ng->active_nodes))
1114 * Both source and destination are nodes in active
1115 * use by this numa group. Maximize memory bandwidth
1116 * by migrating from more heavily used groups, to less
1117 * heavily used ones, spreading the load around.
1118 * Use a 1/4 hysteresis to avoid spurious page movement.
1120 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1123 static unsigned long weighted_cpuload(const int cpu);
1124 static unsigned long source_load(int cpu, int type);
1125 static unsigned long target_load(int cpu, int type);
1126 static unsigned long capacity_of(int cpu);
1127 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1129 /* Cached statistics for all CPUs within a node */
1131 unsigned long nr_running;
1134 /* Total compute capacity of CPUs on a node */
1135 unsigned long compute_capacity;
1137 /* Approximate capacity in terms of runnable tasks on a node */
1138 unsigned long task_capacity;
1139 int has_free_capacity;
1143 * XXX borrowed from update_sg_lb_stats
1145 static void update_numa_stats(struct numa_stats *ns, int nid)
1147 int smt, cpu, cpus = 0;
1148 unsigned long capacity;
1150 memset(ns, 0, sizeof(*ns));
1151 for_each_cpu(cpu, cpumask_of_node(nid)) {
1152 struct rq *rq = cpu_rq(cpu);
1154 ns->nr_running += rq->nr_running;
1155 ns->load += weighted_cpuload(cpu);
1156 ns->compute_capacity += capacity_of(cpu);
1162 * If we raced with hotplug and there are no CPUs left in our mask
1163 * the @ns structure is NULL'ed and task_numa_compare() will
1164 * not find this node attractive.
1166 * We'll either bail at !has_free_capacity, or we'll detect a huge
1167 * imbalance and bail there.
1172 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1173 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1174 capacity = cpus / smt; /* cores */
1176 ns->task_capacity = min_t(unsigned, capacity,
1177 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1178 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1181 struct task_numa_env {
1182 struct task_struct *p;
1184 int src_cpu, src_nid;
1185 int dst_cpu, dst_nid;
1187 struct numa_stats src_stats, dst_stats;
1192 struct task_struct *best_task;
1197 static void task_numa_assign(struct task_numa_env *env,
1198 struct task_struct *p, long imp)
1201 put_task_struct(env->best_task);
1206 env->best_imp = imp;
1207 env->best_cpu = env->dst_cpu;
1210 static bool load_too_imbalanced(long src_load, long dst_load,
1211 struct task_numa_env *env)
1214 long orig_src_load, orig_dst_load;
1215 long src_capacity, dst_capacity;
1218 * The load is corrected for the CPU capacity available on each node.
1221 * ------------ vs ---------
1222 * src_capacity dst_capacity
1224 src_capacity = env->src_stats.compute_capacity;
1225 dst_capacity = env->dst_stats.compute_capacity;
1227 /* We care about the slope of the imbalance, not the direction. */
1228 if (dst_load < src_load)
1229 swap(dst_load, src_load);
1231 /* Is the difference below the threshold? */
1232 imb = dst_load * src_capacity * 100 -
1233 src_load * dst_capacity * env->imbalance_pct;
1238 * The imbalance is above the allowed threshold.
1239 * Compare it with the old imbalance.
1241 orig_src_load = env->src_stats.load;
1242 orig_dst_load = env->dst_stats.load;
1244 if (orig_dst_load < orig_src_load)
1245 swap(orig_dst_load, orig_src_load);
1247 old_imb = orig_dst_load * src_capacity * 100 -
1248 orig_src_load * dst_capacity * env->imbalance_pct;
1250 /* Would this change make things worse? */
1251 return (imb > old_imb);
1255 * This checks if the overall compute and NUMA accesses of the system would
1256 * be improved if the source tasks was migrated to the target dst_cpu taking
1257 * into account that it might be best if task running on the dst_cpu should
1258 * be exchanged with the source task
1260 static void task_numa_compare(struct task_numa_env *env,
1261 long taskimp, long groupimp)
1263 struct rq *src_rq = cpu_rq(env->src_cpu);
1264 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1265 struct task_struct *cur;
1266 long src_load, dst_load;
1268 long imp = env->p->numa_group ? groupimp : taskimp;
1270 int dist = env->dist;
1274 raw_spin_lock_irq(&dst_rq->lock);
1277 * No need to move the exiting task, and this ensures that ->curr
1278 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1279 * is safe under RCU read lock.
1280 * Note that rcu_read_lock() itself can't protect from the final
1281 * put_task_struct() after the last schedule().
1283 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1285 raw_spin_unlock_irq(&dst_rq->lock);
1288 * Because we have preemption enabled we can get migrated around and
1289 * end try selecting ourselves (current == env->p) as a swap candidate.
1295 * "imp" is the fault differential for the source task between the
1296 * source and destination node. Calculate the total differential for
1297 * the source task and potential destination task. The more negative
1298 * the value is, the more rmeote accesses that would be expected to
1299 * be incurred if the tasks were swapped.
1302 /* Skip this swap candidate if cannot move to the source cpu */
1303 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1307 * If dst and source tasks are in the same NUMA group, or not
1308 * in any group then look only at task weights.
1310 if (cur->numa_group == env->p->numa_group) {
1311 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1312 task_weight(cur, env->dst_nid, dist);
1314 * Add some hysteresis to prevent swapping the
1315 * tasks within a group over tiny differences.
1317 if (cur->numa_group)
1321 * Compare the group weights. If a task is all by
1322 * itself (not part of a group), use the task weight
1325 if (cur->numa_group)
1326 imp += group_weight(cur, env->src_nid, dist) -
1327 group_weight(cur, env->dst_nid, dist);
1329 imp += task_weight(cur, env->src_nid, dist) -
1330 task_weight(cur, env->dst_nid, dist);
1334 if (imp <= env->best_imp && moveimp <= env->best_imp)
1338 /* Is there capacity at our destination? */
1339 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1340 !env->dst_stats.has_free_capacity)
1346 /* Balance doesn't matter much if we're running a task per cpu */
1347 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1348 dst_rq->nr_running == 1)
1352 * In the overloaded case, try and keep the load balanced.
1355 load = task_h_load(env->p);
1356 dst_load = env->dst_stats.load + load;
1357 src_load = env->src_stats.load - load;
1359 if (moveimp > imp && moveimp > env->best_imp) {
1361 * If the improvement from just moving env->p direction is
1362 * better than swapping tasks around, check if a move is
1363 * possible. Store a slightly smaller score than moveimp,
1364 * so an actually idle CPU will win.
1366 if (!load_too_imbalanced(src_load, dst_load, env)) {
1373 if (imp <= env->best_imp)
1377 load = task_h_load(cur);
1382 if (load_too_imbalanced(src_load, dst_load, env))
1386 * One idle CPU per node is evaluated for a task numa move.
1387 * Call select_idle_sibling to maybe find a better one.
1390 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1393 task_numa_assign(env, cur, imp);
1398 static void task_numa_find_cpu(struct task_numa_env *env,
1399 long taskimp, long groupimp)
1403 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1404 /* Skip this CPU if the source task cannot migrate */
1405 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1409 task_numa_compare(env, taskimp, groupimp);
1413 /* Only move tasks to a NUMA node less busy than the current node. */
1414 static bool numa_has_capacity(struct task_numa_env *env)
1416 struct numa_stats *src = &env->src_stats;
1417 struct numa_stats *dst = &env->dst_stats;
1419 if (src->has_free_capacity && !dst->has_free_capacity)
1423 * Only consider a task move if the source has a higher load
1424 * than the destination, corrected for CPU capacity on each node.
1426 * src->load dst->load
1427 * --------------------- vs ---------------------
1428 * src->compute_capacity dst->compute_capacity
1430 if (src->load * dst->compute_capacity * env->imbalance_pct >
1432 dst->load * src->compute_capacity * 100)
1438 static int task_numa_migrate(struct task_struct *p)
1440 struct task_numa_env env = {
1443 .src_cpu = task_cpu(p),
1444 .src_nid = task_node(p),
1446 .imbalance_pct = 112,
1452 struct sched_domain *sd;
1453 unsigned long taskweight, groupweight;
1455 long taskimp, groupimp;
1458 * Pick the lowest SD_NUMA domain, as that would have the smallest
1459 * imbalance and would be the first to start moving tasks about.
1461 * And we want to avoid any moving of tasks about, as that would create
1462 * random movement of tasks -- counter the numa conditions we're trying
1466 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1468 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1472 * Cpusets can break the scheduler domain tree into smaller
1473 * balance domains, some of which do not cross NUMA boundaries.
1474 * Tasks that are "trapped" in such domains cannot be migrated
1475 * elsewhere, so there is no point in (re)trying.
1477 if (unlikely(!sd)) {
1478 p->numa_preferred_nid = task_node(p);
1482 env.dst_nid = p->numa_preferred_nid;
1483 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1484 taskweight = task_weight(p, env.src_nid, dist);
1485 groupweight = group_weight(p, env.src_nid, dist);
1486 update_numa_stats(&env.src_stats, env.src_nid);
1487 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1488 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1489 update_numa_stats(&env.dst_stats, env.dst_nid);
1491 /* Try to find a spot on the preferred nid. */
1492 if (numa_has_capacity(&env))
1493 task_numa_find_cpu(&env, taskimp, groupimp);
1496 * Look at other nodes in these cases:
1497 * - there is no space available on the preferred_nid
1498 * - the task is part of a numa_group that is interleaved across
1499 * multiple NUMA nodes; in order to better consolidate the group,
1500 * we need to check other locations.
1502 if (env.best_cpu == -1 || (p->numa_group &&
1503 nodes_weight(p->numa_group->active_nodes) > 1)) {
1504 for_each_online_node(nid) {
1505 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1508 dist = node_distance(env.src_nid, env.dst_nid);
1509 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1511 taskweight = task_weight(p, env.src_nid, dist);
1512 groupweight = group_weight(p, env.src_nid, dist);
1515 /* Only consider nodes where both task and groups benefit */
1516 taskimp = task_weight(p, nid, dist) - taskweight;
1517 groupimp = group_weight(p, nid, dist) - groupweight;
1518 if (taskimp < 0 && groupimp < 0)
1523 update_numa_stats(&env.dst_stats, env.dst_nid);
1524 if (numa_has_capacity(&env))
1525 task_numa_find_cpu(&env, taskimp, groupimp);
1530 * If the task is part of a workload that spans multiple NUMA nodes,
1531 * and is migrating into one of the workload's active nodes, remember
1532 * this node as the task's preferred numa node, so the workload can
1534 * A task that migrated to a second choice node will be better off
1535 * trying for a better one later. Do not set the preferred node here.
1537 if (p->numa_group) {
1538 if (env.best_cpu == -1)
1543 if (node_isset(nid, p->numa_group->active_nodes))
1544 sched_setnuma(p, env.dst_nid);
1547 /* No better CPU than the current one was found. */
1548 if (env.best_cpu == -1)
1552 * Reset the scan period if the task is being rescheduled on an
1553 * alternative node to recheck if the tasks is now properly placed.
1555 p->numa_scan_period = task_scan_min(p);
1557 if (env.best_task == NULL) {
1558 ret = migrate_task_to(p, env.best_cpu);
1560 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1564 ret = migrate_swap(p, env.best_task);
1566 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1567 put_task_struct(env.best_task);
1571 /* Attempt to migrate a task to a CPU on the preferred node. */
1572 static void numa_migrate_preferred(struct task_struct *p)
1574 unsigned long interval = HZ;
1576 /* This task has no NUMA fault statistics yet */
1577 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1580 /* Periodically retry migrating the task to the preferred node */
1581 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1582 p->numa_migrate_retry = jiffies + interval;
1584 /* Success if task is already running on preferred CPU */
1585 if (task_node(p) == p->numa_preferred_nid)
1588 /* Otherwise, try migrate to a CPU on the preferred node */
1589 task_numa_migrate(p);
1593 * Find the nodes on which the workload is actively running. We do this by
1594 * tracking the nodes from which NUMA hinting faults are triggered. This can
1595 * be different from the set of nodes where the workload's memory is currently
1598 * The bitmask is used to make smarter decisions on when to do NUMA page
1599 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1600 * are added when they cause over 6/16 of the maximum number of faults, but
1601 * only removed when they drop below 3/16.
1603 static void update_numa_active_node_mask(struct numa_group *numa_group)
1605 unsigned long faults, max_faults = 0;
1608 for_each_online_node(nid) {
1609 faults = group_faults_cpu(numa_group, nid);
1610 if (faults > max_faults)
1611 max_faults = faults;
1614 for_each_online_node(nid) {
1615 faults = group_faults_cpu(numa_group, nid);
1616 if (!node_isset(nid, numa_group->active_nodes)) {
1617 if (faults > max_faults * 6 / 16)
1618 node_set(nid, numa_group->active_nodes);
1619 } else if (faults < max_faults * 3 / 16)
1620 node_clear(nid, numa_group->active_nodes);
1625 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1626 * increments. The more local the fault statistics are, the higher the scan
1627 * period will be for the next scan window. If local/(local+remote) ratio is
1628 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1629 * the scan period will decrease. Aim for 70% local accesses.
1631 #define NUMA_PERIOD_SLOTS 10
1632 #define NUMA_PERIOD_THRESHOLD 7
1635 * Increase the scan period (slow down scanning) if the majority of
1636 * our memory is already on our local node, or if the majority of
1637 * the page accesses are shared with other processes.
1638 * Otherwise, decrease the scan period.
1640 static void update_task_scan_period(struct task_struct *p,
1641 unsigned long shared, unsigned long private)
1643 unsigned int period_slot;
1647 unsigned long remote = p->numa_faults_locality[0];
1648 unsigned long local = p->numa_faults_locality[1];
1651 * If there were no record hinting faults then either the task is
1652 * completely idle or all activity is areas that are not of interest
1653 * to automatic numa balancing. Related to that, if there were failed
1654 * migration then it implies we are migrating too quickly or the local
1655 * node is overloaded. In either case, scan slower
1657 if (local + shared == 0 || p->numa_faults_locality[2]) {
1658 p->numa_scan_period = min(p->numa_scan_period_max,
1659 p->numa_scan_period << 1);
1661 p->mm->numa_next_scan = jiffies +
1662 msecs_to_jiffies(p->numa_scan_period);
1668 * Prepare to scale scan period relative to the current period.
1669 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1670 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1671 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1673 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1674 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1675 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1676 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1679 diff = slot * period_slot;
1681 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1684 * Scale scan rate increases based on sharing. There is an
1685 * inverse relationship between the degree of sharing and
1686 * the adjustment made to the scanning period. Broadly
1687 * speaking the intent is that there is little point
1688 * scanning faster if shared accesses dominate as it may
1689 * simply bounce migrations uselessly
1691 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1692 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1695 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1696 task_scan_min(p), task_scan_max(p));
1697 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1701 * Get the fraction of time the task has been running since the last
1702 * NUMA placement cycle. The scheduler keeps similar statistics, but
1703 * decays those on a 32ms period, which is orders of magnitude off
1704 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1705 * stats only if the task is so new there are no NUMA statistics yet.
1707 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1709 u64 runtime, delta, now;
1710 /* Use the start of this time slice to avoid calculations. */
1711 now = p->se.exec_start;
1712 runtime = p->se.sum_exec_runtime;
1714 if (p->last_task_numa_placement) {
1715 delta = runtime - p->last_sum_exec_runtime;
1716 *period = now - p->last_task_numa_placement;
1718 delta = p->se.avg.load_sum / p->se.load.weight;
1719 *period = LOAD_AVG_MAX;
1722 p->last_sum_exec_runtime = runtime;
1723 p->last_task_numa_placement = now;
1729 * Determine the preferred nid for a task in a numa_group. This needs to
1730 * be done in a way that produces consistent results with group_weight,
1731 * otherwise workloads might not converge.
1733 static int preferred_group_nid(struct task_struct *p, int nid)
1738 /* Direct connections between all NUMA nodes. */
1739 if (sched_numa_topology_type == NUMA_DIRECT)
1743 * On a system with glueless mesh NUMA topology, group_weight
1744 * scores nodes according to the number of NUMA hinting faults on
1745 * both the node itself, and on nearby nodes.
1747 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1748 unsigned long score, max_score = 0;
1749 int node, max_node = nid;
1751 dist = sched_max_numa_distance;
1753 for_each_online_node(node) {
1754 score = group_weight(p, node, dist);
1755 if (score > max_score) {
1764 * Finding the preferred nid in a system with NUMA backplane
1765 * interconnect topology is more involved. The goal is to locate
1766 * tasks from numa_groups near each other in the system, and
1767 * untangle workloads from different sides of the system. This requires
1768 * searching down the hierarchy of node groups, recursively searching
1769 * inside the highest scoring group of nodes. The nodemask tricks
1770 * keep the complexity of the search down.
1772 nodes = node_online_map;
1773 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1774 unsigned long max_faults = 0;
1775 nodemask_t max_group = NODE_MASK_NONE;
1778 /* Are there nodes at this distance from each other? */
1779 if (!find_numa_distance(dist))
1782 for_each_node_mask(a, nodes) {
1783 unsigned long faults = 0;
1784 nodemask_t this_group;
1785 nodes_clear(this_group);
1787 /* Sum group's NUMA faults; includes a==b case. */
1788 for_each_node_mask(b, nodes) {
1789 if (node_distance(a, b) < dist) {
1790 faults += group_faults(p, b);
1791 node_set(b, this_group);
1792 node_clear(b, nodes);
1796 /* Remember the top group. */
1797 if (faults > max_faults) {
1798 max_faults = faults;
1799 max_group = this_group;
1801 * subtle: at the smallest distance there is
1802 * just one node left in each "group", the
1803 * winner is the preferred nid.
1808 /* Next round, evaluate the nodes within max_group. */
1816 static void task_numa_placement(struct task_struct *p)
1818 int seq, nid, max_nid = -1, max_group_nid = -1;
1819 unsigned long max_faults = 0, max_group_faults = 0;
1820 unsigned long fault_types[2] = { 0, 0 };
1821 unsigned long total_faults;
1822 u64 runtime, period;
1823 spinlock_t *group_lock = NULL;
1826 * The p->mm->numa_scan_seq field gets updated without
1827 * exclusive access. Use READ_ONCE() here to ensure
1828 * that the field is read in a single access:
1830 seq = READ_ONCE(p->mm->numa_scan_seq);
1831 if (p->numa_scan_seq == seq)
1833 p->numa_scan_seq = seq;
1834 p->numa_scan_period_max = task_scan_max(p);
1836 total_faults = p->numa_faults_locality[0] +
1837 p->numa_faults_locality[1];
1838 runtime = numa_get_avg_runtime(p, &period);
1840 /* If the task is part of a group prevent parallel updates to group stats */
1841 if (p->numa_group) {
1842 group_lock = &p->numa_group->lock;
1843 spin_lock_irq(group_lock);
1846 /* Find the node with the highest number of faults */
1847 for_each_online_node(nid) {
1848 /* Keep track of the offsets in numa_faults array */
1849 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1850 unsigned long faults = 0, group_faults = 0;
1853 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1854 long diff, f_diff, f_weight;
1856 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1857 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1858 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1859 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1861 /* Decay existing window, copy faults since last scan */
1862 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1863 fault_types[priv] += p->numa_faults[membuf_idx];
1864 p->numa_faults[membuf_idx] = 0;
1867 * Normalize the faults_from, so all tasks in a group
1868 * count according to CPU use, instead of by the raw
1869 * number of faults. Tasks with little runtime have
1870 * little over-all impact on throughput, and thus their
1871 * faults are less important.
1873 f_weight = div64_u64(runtime << 16, period + 1);
1874 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1876 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1877 p->numa_faults[cpubuf_idx] = 0;
1879 p->numa_faults[mem_idx] += diff;
1880 p->numa_faults[cpu_idx] += f_diff;
1881 faults += p->numa_faults[mem_idx];
1882 p->total_numa_faults += diff;
1883 if (p->numa_group) {
1885 * safe because we can only change our own group
1887 * mem_idx represents the offset for a given
1888 * nid and priv in a specific region because it
1889 * is at the beginning of the numa_faults array.
1891 p->numa_group->faults[mem_idx] += diff;
1892 p->numa_group->faults_cpu[mem_idx] += f_diff;
1893 p->numa_group->total_faults += diff;
1894 group_faults += p->numa_group->faults[mem_idx];
1898 if (faults > max_faults) {
1899 max_faults = faults;
1903 if (group_faults > max_group_faults) {
1904 max_group_faults = group_faults;
1905 max_group_nid = nid;
1909 update_task_scan_period(p, fault_types[0], fault_types[1]);
1911 if (p->numa_group) {
1912 update_numa_active_node_mask(p->numa_group);
1913 spin_unlock_irq(group_lock);
1914 max_nid = preferred_group_nid(p, max_group_nid);
1918 /* Set the new preferred node */
1919 if (max_nid != p->numa_preferred_nid)
1920 sched_setnuma(p, max_nid);
1922 if (task_node(p) != p->numa_preferred_nid)
1923 numa_migrate_preferred(p);
1927 static inline int get_numa_group(struct numa_group *grp)
1929 return atomic_inc_not_zero(&grp->refcount);
1932 static inline void put_numa_group(struct numa_group *grp)
1934 if (atomic_dec_and_test(&grp->refcount))
1935 kfree_rcu(grp, rcu);
1938 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1941 struct numa_group *grp, *my_grp;
1942 struct task_struct *tsk;
1944 int cpu = cpupid_to_cpu(cpupid);
1947 if (unlikely(!p->numa_group)) {
1948 unsigned int size = sizeof(struct numa_group) +
1949 4*nr_node_ids*sizeof(unsigned long);
1951 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1955 atomic_set(&grp->refcount, 1);
1956 spin_lock_init(&grp->lock);
1958 /* Second half of the array tracks nids where faults happen */
1959 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1962 node_set(task_node(current), grp->active_nodes);
1964 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1965 grp->faults[i] = p->numa_faults[i];
1967 grp->total_faults = p->total_numa_faults;
1970 rcu_assign_pointer(p->numa_group, grp);
1974 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1976 if (!cpupid_match_pid(tsk, cpupid))
1979 grp = rcu_dereference(tsk->numa_group);
1983 my_grp = p->numa_group;
1988 * Only join the other group if its bigger; if we're the bigger group,
1989 * the other task will join us.
1991 if (my_grp->nr_tasks > grp->nr_tasks)
1995 * Tie-break on the grp address.
1997 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2000 /* Always join threads in the same process. */
2001 if (tsk->mm == current->mm)
2004 /* Simple filter to avoid false positives due to PID collisions */
2005 if (flags & TNF_SHARED)
2008 /* Update priv based on whether false sharing was detected */
2011 if (join && !get_numa_group(grp))
2019 BUG_ON(irqs_disabled());
2020 double_lock_irq(&my_grp->lock, &grp->lock);
2022 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2023 my_grp->faults[i] -= p->numa_faults[i];
2024 grp->faults[i] += p->numa_faults[i];
2026 my_grp->total_faults -= p->total_numa_faults;
2027 grp->total_faults += p->total_numa_faults;
2032 spin_unlock(&my_grp->lock);
2033 spin_unlock_irq(&grp->lock);
2035 rcu_assign_pointer(p->numa_group, grp);
2037 put_numa_group(my_grp);
2045 void task_numa_free(struct task_struct *p)
2047 struct numa_group *grp = p->numa_group;
2048 void *numa_faults = p->numa_faults;
2049 unsigned long flags;
2053 spin_lock_irqsave(&grp->lock, flags);
2054 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2055 grp->faults[i] -= p->numa_faults[i];
2056 grp->total_faults -= p->total_numa_faults;
2059 spin_unlock_irqrestore(&grp->lock, flags);
2060 RCU_INIT_POINTER(p->numa_group, NULL);
2061 put_numa_group(grp);
2064 p->numa_faults = NULL;
2069 * Got a PROT_NONE fault for a page on @node.
2071 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2073 struct task_struct *p = current;
2074 bool migrated = flags & TNF_MIGRATED;
2075 int cpu_node = task_node(current);
2076 int local = !!(flags & TNF_FAULT_LOCAL);
2079 if (!static_branch_likely(&sched_numa_balancing))
2082 /* for example, ksmd faulting in a user's mm */
2086 /* Allocate buffer to track faults on a per-node basis */
2087 if (unlikely(!p->numa_faults)) {
2088 int size = sizeof(*p->numa_faults) *
2089 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2091 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2092 if (!p->numa_faults)
2095 p->total_numa_faults = 0;
2096 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2100 * First accesses are treated as private, otherwise consider accesses
2101 * to be private if the accessing pid has not changed
2103 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2106 priv = cpupid_match_pid(p, last_cpupid);
2107 if (!priv && !(flags & TNF_NO_GROUP))
2108 task_numa_group(p, last_cpupid, flags, &priv);
2112 * If a workload spans multiple NUMA nodes, a shared fault that
2113 * occurs wholly within the set of nodes that the workload is
2114 * actively using should be counted as local. This allows the
2115 * scan rate to slow down when a workload has settled down.
2117 if (!priv && !local && p->numa_group &&
2118 node_isset(cpu_node, p->numa_group->active_nodes) &&
2119 node_isset(mem_node, p->numa_group->active_nodes))
2122 task_numa_placement(p);
2125 * Retry task to preferred node migration periodically, in case it
2126 * case it previously failed, or the scheduler moved us.
2128 if (time_after(jiffies, p->numa_migrate_retry))
2129 numa_migrate_preferred(p);
2132 p->numa_pages_migrated += pages;
2133 if (flags & TNF_MIGRATE_FAIL)
2134 p->numa_faults_locality[2] += pages;
2136 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2137 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2138 p->numa_faults_locality[local] += pages;
2141 static void reset_ptenuma_scan(struct task_struct *p)
2144 * We only did a read acquisition of the mmap sem, so
2145 * p->mm->numa_scan_seq is written to without exclusive access
2146 * and the update is not guaranteed to be atomic. That's not
2147 * much of an issue though, since this is just used for
2148 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2149 * expensive, to avoid any form of compiler optimizations:
2151 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2152 p->mm->numa_scan_offset = 0;
2156 * The expensive part of numa migration is done from task_work context.
2157 * Triggered from task_tick_numa().
2159 void task_numa_work(struct callback_head *work)
2161 unsigned long migrate, next_scan, now = jiffies;
2162 struct task_struct *p = current;
2163 struct mm_struct *mm = p->mm;
2164 struct vm_area_struct *vma;
2165 unsigned long start, end;
2166 unsigned long nr_pte_updates = 0;
2167 long pages, virtpages;
2169 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2171 work->next = work; /* protect against double add */
2173 * Who cares about NUMA placement when they're dying.
2175 * NOTE: make sure not to dereference p->mm before this check,
2176 * exit_task_work() happens _after_ exit_mm() so we could be called
2177 * without p->mm even though we still had it when we enqueued this
2180 if (p->flags & PF_EXITING)
2183 if (!mm->numa_next_scan) {
2184 mm->numa_next_scan = now +
2185 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2189 * Enforce maximal scan/migration frequency..
2191 migrate = mm->numa_next_scan;
2192 if (time_before(now, migrate))
2195 if (p->numa_scan_period == 0) {
2196 p->numa_scan_period_max = task_scan_max(p);
2197 p->numa_scan_period = task_scan_min(p);
2200 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2201 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2205 * Delay this task enough that another task of this mm will likely win
2206 * the next time around.
2208 p->node_stamp += 2 * TICK_NSEC;
2210 start = mm->numa_scan_offset;
2211 pages = sysctl_numa_balancing_scan_size;
2212 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2213 virtpages = pages * 8; /* Scan up to this much virtual space */
2218 down_read(&mm->mmap_sem);
2219 vma = find_vma(mm, start);
2221 reset_ptenuma_scan(p);
2225 for (; vma; vma = vma->vm_next) {
2226 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2227 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2232 * Shared library pages mapped by multiple processes are not
2233 * migrated as it is expected they are cache replicated. Avoid
2234 * hinting faults in read-only file-backed mappings or the vdso
2235 * as migrating the pages will be of marginal benefit.
2238 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2242 * Skip inaccessible VMAs to avoid any confusion between
2243 * PROT_NONE and NUMA hinting ptes
2245 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2249 start = max(start, vma->vm_start);
2250 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2251 end = min(end, vma->vm_end);
2252 nr_pte_updates = change_prot_numa(vma, start, end);
2255 * Try to scan sysctl_numa_balancing_size worth of
2256 * hpages that have at least one present PTE that
2257 * is not already pte-numa. If the VMA contains
2258 * areas that are unused or already full of prot_numa
2259 * PTEs, scan up to virtpages, to skip through those
2263 pages -= (end - start) >> PAGE_SHIFT;
2264 virtpages -= (end - start) >> PAGE_SHIFT;
2267 if (pages <= 0 || virtpages <= 0)
2271 } while (end != vma->vm_end);
2276 * It is possible to reach the end of the VMA list but the last few
2277 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2278 * would find the !migratable VMA on the next scan but not reset the
2279 * scanner to the start so check it now.
2282 mm->numa_scan_offset = start;
2284 reset_ptenuma_scan(p);
2285 up_read(&mm->mmap_sem);
2289 * Drive the periodic memory faults..
2291 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2293 struct callback_head *work = &curr->numa_work;
2297 * We don't care about NUMA placement if we don't have memory.
2299 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2303 * Using runtime rather than walltime has the dual advantage that
2304 * we (mostly) drive the selection from busy threads and that the
2305 * task needs to have done some actual work before we bother with
2308 now = curr->se.sum_exec_runtime;
2309 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2311 if (now > curr->node_stamp + period) {
2312 if (!curr->node_stamp)
2313 curr->numa_scan_period = task_scan_min(curr);
2314 curr->node_stamp += period;
2316 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2317 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2318 task_work_add(curr, work, true);
2323 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2327 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2331 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2334 #endif /* CONFIG_NUMA_BALANCING */
2337 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2339 update_load_add(&cfs_rq->load, se->load.weight);
2340 if (!parent_entity(se))
2341 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2343 if (entity_is_task(se)) {
2344 struct rq *rq = rq_of(cfs_rq);
2346 account_numa_enqueue(rq, task_of(se));
2347 list_add(&se->group_node, &rq->cfs_tasks);
2350 cfs_rq->nr_running++;
2354 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2356 update_load_sub(&cfs_rq->load, se->load.weight);
2357 if (!parent_entity(se))
2358 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2359 if (entity_is_task(se)) {
2360 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2361 list_del_init(&se->group_node);
2363 cfs_rq->nr_running--;
2366 #ifdef CONFIG_FAIR_GROUP_SCHED
2368 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2373 * Use this CPU's real-time load instead of the last load contribution
2374 * as the updating of the contribution is delayed, and we will use the
2375 * the real-time load to calc the share. See update_tg_load_avg().
2377 tg_weight = atomic_long_read(&tg->load_avg);
2378 tg_weight -= cfs_rq->tg_load_avg_contrib;
2379 tg_weight += cfs_rq->load.weight;
2384 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2386 long tg_weight, load, shares;
2388 tg_weight = calc_tg_weight(tg, cfs_rq);
2389 load = cfs_rq->load.weight;
2391 shares = (tg->shares * load);
2393 shares /= tg_weight;
2395 if (shares < MIN_SHARES)
2396 shares = MIN_SHARES;
2397 if (shares > tg->shares)
2398 shares = tg->shares;
2402 # else /* CONFIG_SMP */
2403 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2407 # endif /* CONFIG_SMP */
2408 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2409 unsigned long weight)
2412 /* commit outstanding execution time */
2413 if (cfs_rq->curr == se)
2414 update_curr(cfs_rq);
2415 account_entity_dequeue(cfs_rq, se);
2418 update_load_set(&se->load, weight);
2421 account_entity_enqueue(cfs_rq, se);
2424 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2426 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2428 struct task_group *tg;
2429 struct sched_entity *se;
2433 se = tg->se[cpu_of(rq_of(cfs_rq))];
2434 if (!se || throttled_hierarchy(cfs_rq))
2437 if (likely(se->load.weight == tg->shares))
2440 shares = calc_cfs_shares(cfs_rq, tg);
2442 reweight_entity(cfs_rq_of(se), se, shares);
2444 #else /* CONFIG_FAIR_GROUP_SCHED */
2445 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2448 #endif /* CONFIG_FAIR_GROUP_SCHED */
2451 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2452 static const u32 runnable_avg_yN_inv[] = {
2453 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2454 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2455 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2456 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2457 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2458 0x85aac367, 0x82cd8698,
2462 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2463 * over-estimates when re-combining.
2465 static const u32 runnable_avg_yN_sum[] = {
2466 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2467 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2468 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2473 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2475 static __always_inline u64 decay_load(u64 val, u64 n)
2477 unsigned int local_n;
2481 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2484 /* after bounds checking we can collapse to 32-bit */
2488 * As y^PERIOD = 1/2, we can combine
2489 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2490 * With a look-up table which covers y^n (n<PERIOD)
2492 * To achieve constant time decay_load.
2494 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2495 val >>= local_n / LOAD_AVG_PERIOD;
2496 local_n %= LOAD_AVG_PERIOD;
2499 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2504 * For updates fully spanning n periods, the contribution to runnable
2505 * average will be: \Sum 1024*y^n
2507 * We can compute this reasonably efficiently by combining:
2508 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2510 static u32 __compute_runnable_contrib(u64 n)
2514 if (likely(n <= LOAD_AVG_PERIOD))
2515 return runnable_avg_yN_sum[n];
2516 else if (unlikely(n >= LOAD_AVG_MAX_N))
2517 return LOAD_AVG_MAX;
2519 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2521 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2522 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2524 n -= LOAD_AVG_PERIOD;
2525 } while (n > LOAD_AVG_PERIOD);
2527 contrib = decay_load(contrib, n);
2528 return contrib + runnable_avg_yN_sum[n];
2531 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2532 #error "load tracking assumes 2^10 as unit"
2535 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2538 * We can represent the historical contribution to runnable average as the
2539 * coefficients of a geometric series. To do this we sub-divide our runnable
2540 * history into segments of approximately 1ms (1024us); label the segment that
2541 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2543 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2545 * (now) (~1ms ago) (~2ms ago)
2547 * Let u_i denote the fraction of p_i that the entity was runnable.
2549 * We then designate the fractions u_i as our co-efficients, yielding the
2550 * following representation of historical load:
2551 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2553 * We choose y based on the with of a reasonably scheduling period, fixing:
2556 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2557 * approximately half as much as the contribution to load within the last ms
2560 * When a period "rolls over" and we have new u_0`, multiplying the previous
2561 * sum again by y is sufficient to update:
2562 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2563 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2565 static __always_inline int
2566 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2567 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2569 u64 delta, scaled_delta, periods;
2571 unsigned int delta_w, scaled_delta_w, decayed = 0;
2572 unsigned long scale_freq, scale_cpu;
2574 delta = now - sa->last_update_time;
2576 * This should only happen when time goes backwards, which it
2577 * unfortunately does during sched clock init when we swap over to TSC.
2579 if ((s64)delta < 0) {
2580 sa->last_update_time = now;
2585 * Use 1024ns as the unit of measurement since it's a reasonable
2586 * approximation of 1us and fast to compute.
2591 sa->last_update_time = now;
2593 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2594 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2595 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2597 /* delta_w is the amount already accumulated against our next period */
2598 delta_w = sa->period_contrib;
2599 if (delta + delta_w >= 1024) {
2602 /* how much left for next period will start over, we don't know yet */
2603 sa->period_contrib = 0;
2606 * Now that we know we're crossing a period boundary, figure
2607 * out how much from delta we need to complete the current
2608 * period and accrue it.
2610 delta_w = 1024 - delta_w;
2611 scaled_delta_w = cap_scale(delta_w, scale_freq);
2613 sa->load_sum += weight * scaled_delta_w;
2615 cfs_rq->runnable_load_sum +=
2616 weight * scaled_delta_w;
2620 sa->util_sum += scaled_delta_w * scale_cpu;
2624 /* Figure out how many additional periods this update spans */
2625 periods = delta / 1024;
2628 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2630 cfs_rq->runnable_load_sum =
2631 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2633 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2635 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2636 contrib = __compute_runnable_contrib(periods);
2637 contrib = cap_scale(contrib, scale_freq);
2639 sa->load_sum += weight * contrib;
2641 cfs_rq->runnable_load_sum += weight * contrib;
2644 sa->util_sum += contrib * scale_cpu;
2647 /* Remainder of delta accrued against u_0` */
2648 scaled_delta = cap_scale(delta, scale_freq);
2650 sa->load_sum += weight * scaled_delta;
2652 cfs_rq->runnable_load_sum += weight * scaled_delta;
2655 sa->util_sum += scaled_delta * scale_cpu;
2657 sa->period_contrib += delta;
2660 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2662 cfs_rq->runnable_load_avg =
2663 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2665 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2671 #ifdef CONFIG_FAIR_GROUP_SCHED
2673 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2674 * and effective_load (which is not done because it is too costly).
2676 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2678 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2680 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2681 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2682 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2686 #else /* CONFIG_FAIR_GROUP_SCHED */
2687 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2688 #endif /* CONFIG_FAIR_GROUP_SCHED */
2690 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2693 * Unsigned subtract and clamp on underflow.
2695 * Explicitly do a load-store to ensure the intermediate value never hits
2696 * memory. This allows lockless observations without ever seeing the negative
2699 #define sub_positive(_ptr, _val) do { \
2700 typeof(_ptr) ptr = (_ptr); \
2701 typeof(*ptr) val = (_val); \
2702 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2706 WRITE_ONCE(*ptr, res); \
2709 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2710 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2712 struct sched_avg *sa = &cfs_rq->avg;
2713 int decayed, removed = 0;
2715 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2716 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2717 sub_positive(&sa->load_avg, r);
2718 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2722 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2723 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2724 sub_positive(&sa->util_avg, r);
2725 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2728 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2729 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2731 #ifndef CONFIG_64BIT
2733 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2736 return decayed || removed;
2739 /* Update task and its cfs_rq load average */
2740 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2742 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2743 u64 now = cfs_rq_clock_task(cfs_rq);
2744 int cpu = cpu_of(rq_of(cfs_rq));
2747 * Track task load average for carrying it to new CPU after migrated, and
2748 * track group sched_entity load average for task_h_load calc in migration
2750 __update_load_avg(now, cpu, &se->avg,
2751 se->on_rq * scale_load_down(se->load.weight),
2752 cfs_rq->curr == se, NULL);
2754 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2755 update_tg_load_avg(cfs_rq, 0);
2757 if (entity_is_task(se))
2758 trace_sched_load_avg_task(task_of(se), &se->avg);
2759 trace_sched_load_avg_cpu(cpu, cfs_rq);
2762 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2764 if (!sched_feat(ATTACH_AGE_LOAD))
2768 * If we got migrated (either between CPUs or between cgroups) we'll
2769 * have aged the average right before clearing @last_update_time.
2771 if (se->avg.last_update_time) {
2772 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2773 &se->avg, 0, 0, NULL);
2776 * XXX: we could have just aged the entire load away if we've been
2777 * absent from the fair class for too long.
2782 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2783 cfs_rq->avg.load_avg += se->avg.load_avg;
2784 cfs_rq->avg.load_sum += se->avg.load_sum;
2785 cfs_rq->avg.util_avg += se->avg.util_avg;
2786 cfs_rq->avg.util_sum += se->avg.util_sum;
2789 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2791 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2792 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2793 cfs_rq->curr == se, NULL);
2795 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2796 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2797 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2798 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2801 /* Add the load generated by se into cfs_rq's load average */
2803 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2805 struct sched_avg *sa = &se->avg;
2806 u64 now = cfs_rq_clock_task(cfs_rq);
2807 int migrated, decayed;
2809 migrated = !sa->last_update_time;
2811 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2812 se->on_rq * scale_load_down(se->load.weight),
2813 cfs_rq->curr == se, NULL);
2816 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2818 cfs_rq->runnable_load_avg += sa->load_avg;
2819 cfs_rq->runnable_load_sum += sa->load_sum;
2822 attach_entity_load_avg(cfs_rq, se);
2824 if (decayed || migrated)
2825 update_tg_load_avg(cfs_rq, 0);
2828 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2830 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2832 update_load_avg(se, 1);
2834 cfs_rq->runnable_load_avg =
2835 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2836 cfs_rq->runnable_load_sum =
2837 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2840 #ifndef CONFIG_64BIT
2841 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2843 u64 last_update_time_copy;
2844 u64 last_update_time;
2847 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2849 last_update_time = cfs_rq->avg.last_update_time;
2850 } while (last_update_time != last_update_time_copy);
2852 return last_update_time;
2855 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2857 return cfs_rq->avg.last_update_time;
2862 * Task first catches up with cfs_rq, and then subtract
2863 * itself from the cfs_rq (task must be off the queue now).
2865 void remove_entity_load_avg(struct sched_entity *se)
2867 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2868 u64 last_update_time;
2871 * Newly created task or never used group entity should not be removed
2872 * from its (source) cfs_rq
2874 if (se->avg.last_update_time == 0)
2877 last_update_time = cfs_rq_last_update_time(cfs_rq);
2879 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2880 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2881 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2885 * Update the rq's load with the elapsed running time before entering
2886 * idle. if the last scheduled task is not a CFS task, idle_enter will
2887 * be the only way to update the runnable statistic.
2889 void idle_enter_fair(struct rq *this_rq)
2894 * Update the rq's load with the elapsed idle time before a task is
2895 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2896 * be the only way to update the runnable statistic.
2898 void idle_exit_fair(struct rq *this_rq)
2902 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2904 return cfs_rq->runnable_load_avg;
2907 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2909 return cfs_rq->avg.load_avg;
2912 static int idle_balance(struct rq *this_rq);
2914 #else /* CONFIG_SMP */
2916 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2918 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2920 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2921 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2924 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2926 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2928 static inline int idle_balance(struct rq *rq)
2933 #endif /* CONFIG_SMP */
2935 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2937 #ifdef CONFIG_SCHEDSTATS
2938 struct task_struct *tsk = NULL;
2940 if (entity_is_task(se))
2943 if (se->statistics.sleep_start) {
2944 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2949 if (unlikely(delta > se->statistics.sleep_max))
2950 se->statistics.sleep_max = delta;
2952 se->statistics.sleep_start = 0;
2953 se->statistics.sum_sleep_runtime += delta;
2956 account_scheduler_latency(tsk, delta >> 10, 1);
2957 trace_sched_stat_sleep(tsk, delta);
2960 if (se->statistics.block_start) {
2961 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2966 if (unlikely(delta > se->statistics.block_max))
2967 se->statistics.block_max = delta;
2969 se->statistics.block_start = 0;
2970 se->statistics.sum_sleep_runtime += delta;
2973 if (tsk->in_iowait) {
2974 se->statistics.iowait_sum += delta;
2975 se->statistics.iowait_count++;
2976 trace_sched_stat_iowait(tsk, delta);
2979 trace_sched_stat_blocked(tsk, delta);
2980 trace_sched_blocked_reason(tsk);
2983 * Blocking time is in units of nanosecs, so shift by
2984 * 20 to get a milliseconds-range estimation of the
2985 * amount of time that the task spent sleeping:
2987 if (unlikely(prof_on == SLEEP_PROFILING)) {
2988 profile_hits(SLEEP_PROFILING,
2989 (void *)get_wchan(tsk),
2992 account_scheduler_latency(tsk, delta >> 10, 0);
2998 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3000 #ifdef CONFIG_SCHED_DEBUG
3001 s64 d = se->vruntime - cfs_rq->min_vruntime;
3006 if (d > 3*sysctl_sched_latency)
3007 schedstat_inc(cfs_rq, nr_spread_over);
3012 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3014 u64 vruntime = cfs_rq->min_vruntime;
3017 * The 'current' period is already promised to the current tasks,
3018 * however the extra weight of the new task will slow them down a
3019 * little, place the new task so that it fits in the slot that
3020 * stays open at the end.
3022 if (initial && sched_feat(START_DEBIT))
3023 vruntime += sched_vslice(cfs_rq, se);
3025 /* sleeps up to a single latency don't count. */
3027 unsigned long thresh = sysctl_sched_latency;
3030 * Halve their sleep time's effect, to allow
3031 * for a gentler effect of sleepers:
3033 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3039 /* ensure we never gain time by being placed backwards. */
3040 se->vruntime = max_vruntime(se->vruntime, vruntime);
3043 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3046 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3049 * Update the normalized vruntime before updating min_vruntime
3050 * through calling update_curr().
3052 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3053 se->vruntime += cfs_rq->min_vruntime;
3056 * Update run-time statistics of the 'current'.
3058 update_curr(cfs_rq);
3059 enqueue_entity_load_avg(cfs_rq, se);
3060 account_entity_enqueue(cfs_rq, se);
3061 update_cfs_shares(cfs_rq);
3063 if (flags & ENQUEUE_WAKEUP) {
3064 place_entity(cfs_rq, se, 0);
3065 enqueue_sleeper(cfs_rq, se);
3068 update_stats_enqueue(cfs_rq, se);
3069 check_spread(cfs_rq, se);
3070 if (se != cfs_rq->curr)
3071 __enqueue_entity(cfs_rq, se);
3074 if (cfs_rq->nr_running == 1) {
3075 list_add_leaf_cfs_rq(cfs_rq);
3076 check_enqueue_throttle(cfs_rq);
3080 static void __clear_buddies_last(struct sched_entity *se)
3082 for_each_sched_entity(se) {
3083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3084 if (cfs_rq->last != se)
3087 cfs_rq->last = NULL;
3091 static void __clear_buddies_next(struct sched_entity *se)
3093 for_each_sched_entity(se) {
3094 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3095 if (cfs_rq->next != se)
3098 cfs_rq->next = NULL;
3102 static void __clear_buddies_skip(struct sched_entity *se)
3104 for_each_sched_entity(se) {
3105 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3106 if (cfs_rq->skip != se)
3109 cfs_rq->skip = NULL;
3113 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3115 if (cfs_rq->last == se)
3116 __clear_buddies_last(se);
3118 if (cfs_rq->next == se)
3119 __clear_buddies_next(se);
3121 if (cfs_rq->skip == se)
3122 __clear_buddies_skip(se);
3125 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3128 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3131 * Update run-time statistics of the 'current'.
3133 update_curr(cfs_rq);
3134 dequeue_entity_load_avg(cfs_rq, se);
3136 update_stats_dequeue(cfs_rq, se);
3137 if (flags & DEQUEUE_SLEEP) {
3138 #ifdef CONFIG_SCHEDSTATS
3139 if (entity_is_task(se)) {
3140 struct task_struct *tsk = task_of(se);
3142 if (tsk->state & TASK_INTERRUPTIBLE)
3143 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3144 if (tsk->state & TASK_UNINTERRUPTIBLE)
3145 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3150 clear_buddies(cfs_rq, se);
3152 if (se != cfs_rq->curr)
3153 __dequeue_entity(cfs_rq, se);
3155 account_entity_dequeue(cfs_rq, se);
3158 * Normalize the entity after updating the min_vruntime because the
3159 * update can refer to the ->curr item and we need to reflect this
3160 * movement in our normalized position.
3162 if (!(flags & DEQUEUE_SLEEP))
3163 se->vruntime -= cfs_rq->min_vruntime;
3165 /* return excess runtime on last dequeue */
3166 return_cfs_rq_runtime(cfs_rq);
3168 update_min_vruntime(cfs_rq);
3169 update_cfs_shares(cfs_rq);
3173 * Preempt the current task with a newly woken task if needed:
3176 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3178 unsigned long ideal_runtime, delta_exec;
3179 struct sched_entity *se;
3182 ideal_runtime = sched_slice(cfs_rq, curr);
3183 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3184 if (delta_exec > ideal_runtime) {
3185 resched_curr(rq_of(cfs_rq));
3187 * The current task ran long enough, ensure it doesn't get
3188 * re-elected due to buddy favours.
3190 clear_buddies(cfs_rq, curr);
3195 * Ensure that a task that missed wakeup preemption by a
3196 * narrow margin doesn't have to wait for a full slice.
3197 * This also mitigates buddy induced latencies under load.
3199 if (delta_exec < sysctl_sched_min_granularity)
3202 se = __pick_first_entity(cfs_rq);
3203 delta = curr->vruntime - se->vruntime;
3208 if (delta > ideal_runtime)
3209 resched_curr(rq_of(cfs_rq));
3213 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3215 /* 'current' is not kept within the tree. */
3218 * Any task has to be enqueued before it get to execute on
3219 * a CPU. So account for the time it spent waiting on the
3222 update_stats_wait_end(cfs_rq, se);
3223 __dequeue_entity(cfs_rq, se);
3224 update_load_avg(se, 1);
3227 update_stats_curr_start(cfs_rq, se);
3229 #ifdef CONFIG_SCHEDSTATS
3231 * Track our maximum slice length, if the CPU's load is at
3232 * least twice that of our own weight (i.e. dont track it
3233 * when there are only lesser-weight tasks around):
3235 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3236 se->statistics.slice_max = max(se->statistics.slice_max,
3237 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3240 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3244 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3247 * Pick the next process, keeping these things in mind, in this order:
3248 * 1) keep things fair between processes/task groups
3249 * 2) pick the "next" process, since someone really wants that to run
3250 * 3) pick the "last" process, for cache locality
3251 * 4) do not run the "skip" process, if something else is available
3253 static struct sched_entity *
3254 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3256 struct sched_entity *left = __pick_first_entity(cfs_rq);
3257 struct sched_entity *se;
3260 * If curr is set we have to see if its left of the leftmost entity
3261 * still in the tree, provided there was anything in the tree at all.
3263 if (!left || (curr && entity_before(curr, left)))
3266 se = left; /* ideally we run the leftmost entity */
3269 * Avoid running the skip buddy, if running something else can
3270 * be done without getting too unfair.
3272 if (cfs_rq->skip == se) {
3273 struct sched_entity *second;
3276 second = __pick_first_entity(cfs_rq);
3278 second = __pick_next_entity(se);
3279 if (!second || (curr && entity_before(curr, second)))
3283 if (second && wakeup_preempt_entity(second, left) < 1)
3288 * Prefer last buddy, try to return the CPU to a preempted task.
3290 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3294 * Someone really wants this to run. If it's not unfair, run it.
3296 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3299 clear_buddies(cfs_rq, se);
3304 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3306 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3309 * If still on the runqueue then deactivate_task()
3310 * was not called and update_curr() has to be done:
3313 update_curr(cfs_rq);
3315 /* throttle cfs_rqs exceeding runtime */
3316 check_cfs_rq_runtime(cfs_rq);
3318 check_spread(cfs_rq, prev);
3320 update_stats_wait_start(cfs_rq, prev);
3321 /* Put 'current' back into the tree. */
3322 __enqueue_entity(cfs_rq, prev);
3323 /* in !on_rq case, update occurred at dequeue */
3324 update_load_avg(prev, 0);
3326 cfs_rq->curr = NULL;
3330 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3333 * Update run-time statistics of the 'current'.
3335 update_curr(cfs_rq);
3338 * Ensure that runnable average is periodically updated.
3340 update_load_avg(curr, 1);
3341 update_cfs_shares(cfs_rq);
3343 #ifdef CONFIG_SCHED_HRTICK
3345 * queued ticks are scheduled to match the slice, so don't bother
3346 * validating it and just reschedule.
3349 resched_curr(rq_of(cfs_rq));
3353 * don't let the period tick interfere with the hrtick preemption
3355 if (!sched_feat(DOUBLE_TICK) &&
3356 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3360 if (cfs_rq->nr_running > 1)
3361 check_preempt_tick(cfs_rq, curr);
3365 /**************************************************
3366 * CFS bandwidth control machinery
3369 #ifdef CONFIG_CFS_BANDWIDTH
3371 #ifdef HAVE_JUMP_LABEL
3372 static struct static_key __cfs_bandwidth_used;
3374 static inline bool cfs_bandwidth_used(void)
3376 return static_key_false(&__cfs_bandwidth_used);
3379 void cfs_bandwidth_usage_inc(void)
3381 static_key_slow_inc(&__cfs_bandwidth_used);
3384 void cfs_bandwidth_usage_dec(void)
3386 static_key_slow_dec(&__cfs_bandwidth_used);
3388 #else /* HAVE_JUMP_LABEL */
3389 static bool cfs_bandwidth_used(void)
3394 void cfs_bandwidth_usage_inc(void) {}
3395 void cfs_bandwidth_usage_dec(void) {}
3396 #endif /* HAVE_JUMP_LABEL */
3399 * default period for cfs group bandwidth.
3400 * default: 0.1s, units: nanoseconds
3402 static inline u64 default_cfs_period(void)
3404 return 100000000ULL;
3407 static inline u64 sched_cfs_bandwidth_slice(void)
3409 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3413 * Replenish runtime according to assigned quota and update expiration time.
3414 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3415 * additional synchronization around rq->lock.
3417 * requires cfs_b->lock
3419 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3423 if (cfs_b->quota == RUNTIME_INF)
3426 now = sched_clock_cpu(smp_processor_id());
3427 cfs_b->runtime = cfs_b->quota;
3428 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3431 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3433 return &tg->cfs_bandwidth;
3436 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3437 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3439 if (unlikely(cfs_rq->throttle_count))
3440 return cfs_rq->throttled_clock_task;
3442 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3445 /* returns 0 on failure to allocate runtime */
3446 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3448 struct task_group *tg = cfs_rq->tg;
3449 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3450 u64 amount = 0, min_amount, expires;
3452 /* note: this is a positive sum as runtime_remaining <= 0 */
3453 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3455 raw_spin_lock(&cfs_b->lock);
3456 if (cfs_b->quota == RUNTIME_INF)
3457 amount = min_amount;
3459 start_cfs_bandwidth(cfs_b);
3461 if (cfs_b->runtime > 0) {
3462 amount = min(cfs_b->runtime, min_amount);
3463 cfs_b->runtime -= amount;
3467 expires = cfs_b->runtime_expires;
3468 raw_spin_unlock(&cfs_b->lock);
3470 cfs_rq->runtime_remaining += amount;
3472 * we may have advanced our local expiration to account for allowed
3473 * spread between our sched_clock and the one on which runtime was
3476 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3477 cfs_rq->runtime_expires = expires;
3479 return cfs_rq->runtime_remaining > 0;
3483 * Note: This depends on the synchronization provided by sched_clock and the
3484 * fact that rq->clock snapshots this value.
3486 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3488 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3490 /* if the deadline is ahead of our clock, nothing to do */
3491 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3494 if (cfs_rq->runtime_remaining < 0)
3498 * If the local deadline has passed we have to consider the
3499 * possibility that our sched_clock is 'fast' and the global deadline
3500 * has not truly expired.
3502 * Fortunately we can check determine whether this the case by checking
3503 * whether the global deadline has advanced. It is valid to compare
3504 * cfs_b->runtime_expires without any locks since we only care about
3505 * exact equality, so a partial write will still work.
3508 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3509 /* extend local deadline, drift is bounded above by 2 ticks */
3510 cfs_rq->runtime_expires += TICK_NSEC;
3512 /* global deadline is ahead, expiration has passed */
3513 cfs_rq->runtime_remaining = 0;
3517 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3519 /* dock delta_exec before expiring quota (as it could span periods) */
3520 cfs_rq->runtime_remaining -= delta_exec;
3521 expire_cfs_rq_runtime(cfs_rq);
3523 if (likely(cfs_rq->runtime_remaining > 0))
3527 * if we're unable to extend our runtime we resched so that the active
3528 * hierarchy can be throttled
3530 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3531 resched_curr(rq_of(cfs_rq));
3534 static __always_inline
3535 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3537 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3540 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3543 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3545 return cfs_bandwidth_used() && cfs_rq->throttled;
3548 /* check whether cfs_rq, or any parent, is throttled */
3549 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3551 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3555 * Ensure that neither of the group entities corresponding to src_cpu or
3556 * dest_cpu are members of a throttled hierarchy when performing group
3557 * load-balance operations.
3559 static inline int throttled_lb_pair(struct task_group *tg,
3560 int src_cpu, int dest_cpu)
3562 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3564 src_cfs_rq = tg->cfs_rq[src_cpu];
3565 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3567 return throttled_hierarchy(src_cfs_rq) ||
3568 throttled_hierarchy(dest_cfs_rq);
3571 /* updated child weight may affect parent so we have to do this bottom up */
3572 static int tg_unthrottle_up(struct task_group *tg, void *data)
3574 struct rq *rq = data;
3575 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3577 cfs_rq->throttle_count--;
3579 if (!cfs_rq->throttle_count) {
3580 /* adjust cfs_rq_clock_task() */
3581 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3582 cfs_rq->throttled_clock_task;
3589 static int tg_throttle_down(struct task_group *tg, void *data)
3591 struct rq *rq = data;
3592 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3594 /* group is entering throttled state, stop time */
3595 if (!cfs_rq->throttle_count)
3596 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3597 cfs_rq->throttle_count++;
3602 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3604 struct rq *rq = rq_of(cfs_rq);
3605 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3606 struct sched_entity *se;
3607 long task_delta, dequeue = 1;
3610 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3612 /* freeze hierarchy runnable averages while throttled */
3614 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3617 task_delta = cfs_rq->h_nr_running;
3618 for_each_sched_entity(se) {
3619 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3620 /* throttled entity or throttle-on-deactivate */
3625 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3626 qcfs_rq->h_nr_running -= task_delta;
3628 if (qcfs_rq->load.weight)
3633 sub_nr_running(rq, task_delta);
3635 cfs_rq->throttled = 1;
3636 cfs_rq->throttled_clock = rq_clock(rq);
3637 raw_spin_lock(&cfs_b->lock);
3638 empty = list_empty(&cfs_b->throttled_cfs_rq);
3641 * Add to the _head_ of the list, so that an already-started
3642 * distribute_cfs_runtime will not see us
3644 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3647 * If we're the first throttled task, make sure the bandwidth
3651 start_cfs_bandwidth(cfs_b);
3653 raw_spin_unlock(&cfs_b->lock);
3656 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3658 struct rq *rq = rq_of(cfs_rq);
3659 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3660 struct sched_entity *se;
3664 se = cfs_rq->tg->se[cpu_of(rq)];
3666 cfs_rq->throttled = 0;
3668 update_rq_clock(rq);
3670 raw_spin_lock(&cfs_b->lock);
3671 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3672 list_del_rcu(&cfs_rq->throttled_list);
3673 raw_spin_unlock(&cfs_b->lock);
3675 /* update hierarchical throttle state */
3676 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3678 if (!cfs_rq->load.weight)
3681 task_delta = cfs_rq->h_nr_running;
3682 for_each_sched_entity(se) {
3686 cfs_rq = cfs_rq_of(se);
3688 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3689 cfs_rq->h_nr_running += task_delta;
3691 if (cfs_rq_throttled(cfs_rq))
3696 add_nr_running(rq, task_delta);
3698 /* determine whether we need to wake up potentially idle cpu */
3699 if (rq->curr == rq->idle && rq->cfs.nr_running)
3703 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3704 u64 remaining, u64 expires)
3706 struct cfs_rq *cfs_rq;
3708 u64 starting_runtime = remaining;
3711 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3713 struct rq *rq = rq_of(cfs_rq);
3715 raw_spin_lock(&rq->lock);
3716 if (!cfs_rq_throttled(cfs_rq))
3719 runtime = -cfs_rq->runtime_remaining + 1;
3720 if (runtime > remaining)
3721 runtime = remaining;
3722 remaining -= runtime;
3724 cfs_rq->runtime_remaining += runtime;
3725 cfs_rq->runtime_expires = expires;
3727 /* we check whether we're throttled above */
3728 if (cfs_rq->runtime_remaining > 0)
3729 unthrottle_cfs_rq(cfs_rq);
3732 raw_spin_unlock(&rq->lock);
3739 return starting_runtime - remaining;
3743 * Responsible for refilling a task_group's bandwidth and unthrottling its
3744 * cfs_rqs as appropriate. If there has been no activity within the last
3745 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3746 * used to track this state.
3748 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3750 u64 runtime, runtime_expires;
3753 /* no need to continue the timer with no bandwidth constraint */
3754 if (cfs_b->quota == RUNTIME_INF)
3755 goto out_deactivate;
3757 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3758 cfs_b->nr_periods += overrun;
3761 * idle depends on !throttled (for the case of a large deficit), and if
3762 * we're going inactive then everything else can be deferred
3764 if (cfs_b->idle && !throttled)
3765 goto out_deactivate;
3767 __refill_cfs_bandwidth_runtime(cfs_b);
3770 /* mark as potentially idle for the upcoming period */
3775 /* account preceding periods in which throttling occurred */
3776 cfs_b->nr_throttled += overrun;
3778 runtime_expires = cfs_b->runtime_expires;
3781 * This check is repeated as we are holding onto the new bandwidth while
3782 * we unthrottle. This can potentially race with an unthrottled group
3783 * trying to acquire new bandwidth from the global pool. This can result
3784 * in us over-using our runtime if it is all used during this loop, but
3785 * only by limited amounts in that extreme case.
3787 while (throttled && cfs_b->runtime > 0) {
3788 runtime = cfs_b->runtime;
3789 raw_spin_unlock(&cfs_b->lock);
3790 /* we can't nest cfs_b->lock while distributing bandwidth */
3791 runtime = distribute_cfs_runtime(cfs_b, runtime,
3793 raw_spin_lock(&cfs_b->lock);
3795 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3797 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3801 * While we are ensured activity in the period following an
3802 * unthrottle, this also covers the case in which the new bandwidth is
3803 * insufficient to cover the existing bandwidth deficit. (Forcing the
3804 * timer to remain active while there are any throttled entities.)
3814 /* a cfs_rq won't donate quota below this amount */
3815 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3816 /* minimum remaining period time to redistribute slack quota */
3817 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3818 /* how long we wait to gather additional slack before distributing */
3819 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3822 * Are we near the end of the current quota period?
3824 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3825 * hrtimer base being cleared by hrtimer_start. In the case of
3826 * migrate_hrtimers, base is never cleared, so we are fine.
3828 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3830 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3833 /* if the call-back is running a quota refresh is already occurring */
3834 if (hrtimer_callback_running(refresh_timer))
3837 /* is a quota refresh about to occur? */
3838 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3839 if (remaining < min_expire)
3845 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3847 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3849 /* if there's a quota refresh soon don't bother with slack */
3850 if (runtime_refresh_within(cfs_b, min_left))
3853 hrtimer_start(&cfs_b->slack_timer,
3854 ns_to_ktime(cfs_bandwidth_slack_period),
3858 /* we know any runtime found here is valid as update_curr() precedes return */
3859 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3861 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3862 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3864 if (slack_runtime <= 0)
3867 raw_spin_lock(&cfs_b->lock);
3868 if (cfs_b->quota != RUNTIME_INF &&
3869 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3870 cfs_b->runtime += slack_runtime;
3872 /* we are under rq->lock, defer unthrottling using a timer */
3873 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3874 !list_empty(&cfs_b->throttled_cfs_rq))
3875 start_cfs_slack_bandwidth(cfs_b);
3877 raw_spin_unlock(&cfs_b->lock);
3879 /* even if it's not valid for return we don't want to try again */
3880 cfs_rq->runtime_remaining -= slack_runtime;
3883 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3885 if (!cfs_bandwidth_used())
3888 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3891 __return_cfs_rq_runtime(cfs_rq);
3895 * This is done with a timer (instead of inline with bandwidth return) since
3896 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3898 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3900 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3903 /* confirm we're still not at a refresh boundary */
3904 raw_spin_lock(&cfs_b->lock);
3905 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3906 raw_spin_unlock(&cfs_b->lock);
3910 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3911 runtime = cfs_b->runtime;
3913 expires = cfs_b->runtime_expires;
3914 raw_spin_unlock(&cfs_b->lock);
3919 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3921 raw_spin_lock(&cfs_b->lock);
3922 if (expires == cfs_b->runtime_expires)
3923 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3924 raw_spin_unlock(&cfs_b->lock);
3928 * When a group wakes up we want to make sure that its quota is not already
3929 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3930 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3932 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3934 if (!cfs_bandwidth_used())
3937 /* an active group must be handled by the update_curr()->put() path */
3938 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3941 /* ensure the group is not already throttled */
3942 if (cfs_rq_throttled(cfs_rq))
3945 /* update runtime allocation */
3946 account_cfs_rq_runtime(cfs_rq, 0);
3947 if (cfs_rq->runtime_remaining <= 0)
3948 throttle_cfs_rq(cfs_rq);
3951 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3952 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3954 if (!cfs_bandwidth_used())
3957 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3961 * it's possible for a throttled entity to be forced into a running
3962 * state (e.g. set_curr_task), in this case we're finished.
3964 if (cfs_rq_throttled(cfs_rq))
3967 throttle_cfs_rq(cfs_rq);
3971 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3973 struct cfs_bandwidth *cfs_b =
3974 container_of(timer, struct cfs_bandwidth, slack_timer);
3976 do_sched_cfs_slack_timer(cfs_b);
3978 return HRTIMER_NORESTART;
3981 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3983 struct cfs_bandwidth *cfs_b =
3984 container_of(timer, struct cfs_bandwidth, period_timer);
3988 raw_spin_lock(&cfs_b->lock);
3990 overrun = hrtimer_forward_now(timer, cfs_b->period);
3994 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3997 cfs_b->period_active = 0;
3998 raw_spin_unlock(&cfs_b->lock);
4000 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4003 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4005 raw_spin_lock_init(&cfs_b->lock);
4007 cfs_b->quota = RUNTIME_INF;
4008 cfs_b->period = ns_to_ktime(default_cfs_period());
4010 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4011 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4012 cfs_b->period_timer.function = sched_cfs_period_timer;
4013 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4014 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4017 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4019 cfs_rq->runtime_enabled = 0;
4020 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4023 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4025 lockdep_assert_held(&cfs_b->lock);
4027 if (!cfs_b->period_active) {
4028 cfs_b->period_active = 1;
4029 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4030 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4034 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4036 /* init_cfs_bandwidth() was not called */
4037 if (!cfs_b->throttled_cfs_rq.next)
4040 hrtimer_cancel(&cfs_b->period_timer);
4041 hrtimer_cancel(&cfs_b->slack_timer);
4044 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4046 struct cfs_rq *cfs_rq;
4048 for_each_leaf_cfs_rq(rq, cfs_rq) {
4049 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4051 raw_spin_lock(&cfs_b->lock);
4052 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4053 raw_spin_unlock(&cfs_b->lock);
4057 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4059 struct cfs_rq *cfs_rq;
4061 for_each_leaf_cfs_rq(rq, cfs_rq) {
4062 if (!cfs_rq->runtime_enabled)
4066 * clock_task is not advancing so we just need to make sure
4067 * there's some valid quota amount
4069 cfs_rq->runtime_remaining = 1;
4071 * Offline rq is schedulable till cpu is completely disabled
4072 * in take_cpu_down(), so we prevent new cfs throttling here.
4074 cfs_rq->runtime_enabled = 0;
4076 if (cfs_rq_throttled(cfs_rq))
4077 unthrottle_cfs_rq(cfs_rq);
4081 #else /* CONFIG_CFS_BANDWIDTH */
4082 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4084 return rq_clock_task(rq_of(cfs_rq));
4087 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4088 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4089 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4090 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4092 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4097 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4102 static inline int throttled_lb_pair(struct task_group *tg,
4103 int src_cpu, int dest_cpu)
4108 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4110 #ifdef CONFIG_FAIR_GROUP_SCHED
4111 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4114 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4118 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4119 static inline void update_runtime_enabled(struct rq *rq) {}
4120 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4122 #endif /* CONFIG_CFS_BANDWIDTH */
4124 /**************************************************
4125 * CFS operations on tasks:
4128 #ifdef CONFIG_SCHED_HRTICK
4129 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4131 struct sched_entity *se = &p->se;
4132 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4134 WARN_ON(task_rq(p) != rq);
4136 if (cfs_rq->nr_running > 1) {
4137 u64 slice = sched_slice(cfs_rq, se);
4138 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4139 s64 delta = slice - ran;
4146 hrtick_start(rq, delta);
4151 * called from enqueue/dequeue and updates the hrtick when the
4152 * current task is from our class and nr_running is low enough
4155 static void hrtick_update(struct rq *rq)
4157 struct task_struct *curr = rq->curr;
4159 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4162 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4163 hrtick_start_fair(rq, curr);
4165 #else /* !CONFIG_SCHED_HRTICK */
4167 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4171 static inline void hrtick_update(struct rq *rq)
4177 static bool cpu_overutilized(int cpu);
4178 static inline unsigned long boosted_cpu_util(int cpu);
4180 #define boosted_cpu_util(cpu) cpu_util(cpu)
4184 static void update_capacity_of(int cpu)
4186 unsigned long req_cap;
4191 /* Convert scale-invariant capacity to cpu. */
4192 req_cap = boosted_cpu_util(cpu);
4193 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4194 set_cfs_cpu_capacity(cpu, true, req_cap);
4199 * The enqueue_task method is called before nr_running is
4200 * increased. Here we update the fair scheduling stats and
4201 * then put the task into the rbtree:
4204 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4206 struct cfs_rq *cfs_rq;
4207 struct sched_entity *se = &p->se;
4209 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4210 int task_wakeup = flags & ENQUEUE_WAKEUP;
4213 for_each_sched_entity(se) {
4216 cfs_rq = cfs_rq_of(se);
4217 enqueue_entity(cfs_rq, se, flags);
4220 * end evaluation on encountering a throttled cfs_rq
4222 * note: in the case of encountering a throttled cfs_rq we will
4223 * post the final h_nr_running increment below.
4225 if (cfs_rq_throttled(cfs_rq))
4227 cfs_rq->h_nr_running++;
4229 flags = ENQUEUE_WAKEUP;
4232 for_each_sched_entity(se) {
4233 cfs_rq = cfs_rq_of(se);
4234 cfs_rq->h_nr_running++;
4236 if (cfs_rq_throttled(cfs_rq))
4239 update_load_avg(se, 1);
4240 update_cfs_shares(cfs_rq);
4244 add_nr_running(rq, 1);
4249 if (!task_new && !rq->rd->overutilized &&
4250 cpu_overutilized(rq->cpu))
4251 rq->rd->overutilized = true;
4253 schedtune_enqueue_task(p, cpu_of(rq));
4256 * We want to potentially trigger a freq switch
4257 * request only for tasks that are waking up; this is
4258 * because we get here also during load balancing, but
4259 * in these cases it seems wise to trigger as single
4260 * request after load balancing is done.
4262 if (task_new || task_wakeup)
4263 update_capacity_of(cpu_of(rq));
4265 #endif /* CONFIG_SMP */
4270 static void set_next_buddy(struct sched_entity *se);
4273 * The dequeue_task method is called before nr_running is
4274 * decreased. We remove the task from the rbtree and
4275 * update the fair scheduling stats:
4277 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4279 struct cfs_rq *cfs_rq;
4280 struct sched_entity *se = &p->se;
4281 int task_sleep = flags & DEQUEUE_SLEEP;
4283 for_each_sched_entity(se) {
4284 cfs_rq = cfs_rq_of(se);
4285 dequeue_entity(cfs_rq, se, flags);
4288 * end evaluation on encountering a throttled cfs_rq
4290 * note: in the case of encountering a throttled cfs_rq we will
4291 * post the final h_nr_running decrement below.
4293 if (cfs_rq_throttled(cfs_rq))
4295 cfs_rq->h_nr_running--;
4297 /* Don't dequeue parent if it has other entities besides us */
4298 if (cfs_rq->load.weight) {
4300 * Bias pick_next to pick a task from this cfs_rq, as
4301 * p is sleeping when it is within its sched_slice.
4303 if (task_sleep && parent_entity(se))
4304 set_next_buddy(parent_entity(se));
4306 /* avoid re-evaluating load for this entity */
4307 se = parent_entity(se);
4310 flags |= DEQUEUE_SLEEP;
4313 for_each_sched_entity(se) {
4314 cfs_rq = cfs_rq_of(se);
4315 cfs_rq->h_nr_running--;
4317 if (cfs_rq_throttled(cfs_rq))
4320 update_load_avg(se, 1);
4321 update_cfs_shares(cfs_rq);
4325 sub_nr_running(rq, 1);
4330 schedtune_dequeue_task(p, cpu_of(rq));
4333 * We want to potentially trigger a freq switch
4334 * request only for tasks that are going to sleep;
4335 * this is because we get here also during load
4336 * balancing, but in these cases it seems wise to
4337 * trigger as single request after load balancing is
4341 if (rq->cfs.nr_running)
4342 update_capacity_of(cpu_of(rq));
4343 else if (sched_freq())
4344 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4348 #endif /* CONFIG_SMP */
4356 * per rq 'load' arrray crap; XXX kill this.
4360 * The exact cpuload at various idx values, calculated at every tick would be
4361 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4363 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4364 * on nth tick when cpu may be busy, then we have:
4365 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4366 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4368 * decay_load_missed() below does efficient calculation of
4369 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4370 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4372 * The calculation is approximated on a 128 point scale.
4373 * degrade_zero_ticks is the number of ticks after which load at any
4374 * particular idx is approximated to be zero.
4375 * degrade_factor is a precomputed table, a row for each load idx.
4376 * Each column corresponds to degradation factor for a power of two ticks,
4377 * based on 128 point scale.
4379 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4380 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4382 * With this power of 2 load factors, we can degrade the load n times
4383 * by looking at 1 bits in n and doing as many mult/shift instead of
4384 * n mult/shifts needed by the exact degradation.
4386 #define DEGRADE_SHIFT 7
4387 static const unsigned char
4388 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4389 static const unsigned char
4390 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4391 {0, 0, 0, 0, 0, 0, 0, 0},
4392 {64, 32, 8, 0, 0, 0, 0, 0},
4393 {96, 72, 40, 12, 1, 0, 0},
4394 {112, 98, 75, 43, 15, 1, 0},
4395 {120, 112, 98, 76, 45, 16, 2} };
4398 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4399 * would be when CPU is idle and so we just decay the old load without
4400 * adding any new load.
4402 static unsigned long
4403 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4407 if (!missed_updates)
4410 if (missed_updates >= degrade_zero_ticks[idx])
4414 return load >> missed_updates;
4416 while (missed_updates) {
4417 if (missed_updates % 2)
4418 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4420 missed_updates >>= 1;
4427 * Update rq->cpu_load[] statistics. This function is usually called every
4428 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4429 * every tick. We fix it up based on jiffies.
4431 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4432 unsigned long pending_updates)
4436 this_rq->nr_load_updates++;
4438 /* Update our load: */
4439 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4440 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4441 unsigned long old_load, new_load;
4443 /* scale is effectively 1 << i now, and >> i divides by scale */
4445 old_load = this_rq->cpu_load[i];
4446 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4447 new_load = this_load;
4449 * Round up the averaging division if load is increasing. This
4450 * prevents us from getting stuck on 9 if the load is 10, for
4453 if (new_load > old_load)
4454 new_load += scale - 1;
4456 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4459 sched_avg_update(this_rq);
4462 /* Used instead of source_load when we know the type == 0 */
4463 static unsigned long weighted_cpuload(const int cpu)
4465 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4468 #ifdef CONFIG_NO_HZ_COMMON
4470 * There is no sane way to deal with nohz on smp when using jiffies because the
4471 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4472 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4474 * Therefore we cannot use the delta approach from the regular tick since that
4475 * would seriously skew the load calculation. However we'll make do for those
4476 * updates happening while idle (nohz_idle_balance) or coming out of idle
4477 * (tick_nohz_idle_exit).
4479 * This means we might still be one tick off for nohz periods.
4483 * Called from nohz_idle_balance() to update the load ratings before doing the
4486 static void update_idle_cpu_load(struct rq *this_rq)
4488 unsigned long curr_jiffies = READ_ONCE(jiffies);
4489 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4490 unsigned long pending_updates;
4493 * bail if there's load or we're actually up-to-date.
4495 if (load || curr_jiffies == this_rq->last_load_update_tick)
4498 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4499 this_rq->last_load_update_tick = curr_jiffies;
4501 __update_cpu_load(this_rq, load, pending_updates);
4505 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4507 void update_cpu_load_nohz(void)
4509 struct rq *this_rq = this_rq();
4510 unsigned long curr_jiffies = READ_ONCE(jiffies);
4511 unsigned long pending_updates;
4513 if (curr_jiffies == this_rq->last_load_update_tick)
4516 raw_spin_lock(&this_rq->lock);
4517 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4518 if (pending_updates) {
4519 this_rq->last_load_update_tick = curr_jiffies;
4521 * We were idle, this means load 0, the current load might be
4522 * !0 due to remote wakeups and the sort.
4524 __update_cpu_load(this_rq, 0, pending_updates);
4526 raw_spin_unlock(&this_rq->lock);
4528 #endif /* CONFIG_NO_HZ */
4531 * Called from scheduler_tick()
4533 void update_cpu_load_active(struct rq *this_rq)
4535 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4537 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4539 this_rq->last_load_update_tick = jiffies;
4540 __update_cpu_load(this_rq, load, 1);
4544 * Return a low guess at the load of a migration-source cpu weighted
4545 * according to the scheduling class and "nice" value.
4547 * We want to under-estimate the load of migration sources, to
4548 * balance conservatively.
4550 static unsigned long source_load(int cpu, int type)
4552 struct rq *rq = cpu_rq(cpu);
4553 unsigned long total = weighted_cpuload(cpu);
4555 if (type == 0 || !sched_feat(LB_BIAS))
4558 return min(rq->cpu_load[type-1], total);
4562 * Return a high guess at the load of a migration-target cpu weighted
4563 * according to the scheduling class and "nice" value.
4565 static unsigned long target_load(int cpu, int type)
4567 struct rq *rq = cpu_rq(cpu);
4568 unsigned long total = weighted_cpuload(cpu);
4570 if (type == 0 || !sched_feat(LB_BIAS))
4573 return max(rq->cpu_load[type-1], total);
4577 static unsigned long cpu_avg_load_per_task(int cpu)
4579 struct rq *rq = cpu_rq(cpu);
4580 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4581 unsigned long load_avg = weighted_cpuload(cpu);
4584 return load_avg / nr_running;
4589 static void record_wakee(struct task_struct *p)
4592 * Rough decay (wiping) for cost saving, don't worry
4593 * about the boundary, really active task won't care
4596 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4597 current->wakee_flips >>= 1;
4598 current->wakee_flip_decay_ts = jiffies;
4601 if (current->last_wakee != p) {
4602 current->last_wakee = p;
4603 current->wakee_flips++;
4607 static void task_waking_fair(struct task_struct *p)
4609 struct sched_entity *se = &p->se;
4610 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4613 #ifndef CONFIG_64BIT
4614 u64 min_vruntime_copy;
4617 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4619 min_vruntime = cfs_rq->min_vruntime;
4620 } while (min_vruntime != min_vruntime_copy);
4622 min_vruntime = cfs_rq->min_vruntime;
4625 se->vruntime -= min_vruntime;
4629 #ifdef CONFIG_FAIR_GROUP_SCHED
4631 * effective_load() calculates the load change as seen from the root_task_group
4633 * Adding load to a group doesn't make a group heavier, but can cause movement
4634 * of group shares between cpus. Assuming the shares were perfectly aligned one
4635 * can calculate the shift in shares.
4637 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4638 * on this @cpu and results in a total addition (subtraction) of @wg to the
4639 * total group weight.
4641 * Given a runqueue weight distribution (rw_i) we can compute a shares
4642 * distribution (s_i) using:
4644 * s_i = rw_i / \Sum rw_j (1)
4646 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4647 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4648 * shares distribution (s_i):
4650 * rw_i = { 2, 4, 1, 0 }
4651 * s_i = { 2/7, 4/7, 1/7, 0 }
4653 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4654 * task used to run on and the CPU the waker is running on), we need to
4655 * compute the effect of waking a task on either CPU and, in case of a sync
4656 * wakeup, compute the effect of the current task going to sleep.
4658 * So for a change of @wl to the local @cpu with an overall group weight change
4659 * of @wl we can compute the new shares distribution (s'_i) using:
4661 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4663 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4664 * differences in waking a task to CPU 0. The additional task changes the
4665 * weight and shares distributions like:
4667 * rw'_i = { 3, 4, 1, 0 }
4668 * s'_i = { 3/8, 4/8, 1/8, 0 }
4670 * We can then compute the difference in effective weight by using:
4672 * dw_i = S * (s'_i - s_i) (3)
4674 * Where 'S' is the group weight as seen by its parent.
4676 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4677 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4678 * 4/7) times the weight of the group.
4680 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4682 struct sched_entity *se = tg->se[cpu];
4684 if (!tg->parent) /* the trivial, non-cgroup case */
4687 for_each_sched_entity(se) {
4688 struct cfs_rq *cfs_rq = se->my_q;
4689 long W, w = cfs_rq_load_avg(cfs_rq);
4694 * W = @wg + \Sum rw_j
4696 W = wg + atomic_long_read(&tg->load_avg);
4698 /* Ensure \Sum rw_j >= rw_i */
4699 W -= cfs_rq->tg_load_avg_contrib;
4708 * wl = S * s'_i; see (2)
4711 wl = (w * (long)tg->shares) / W;
4716 * Per the above, wl is the new se->load.weight value; since
4717 * those are clipped to [MIN_SHARES, ...) do so now. See
4718 * calc_cfs_shares().
4720 if (wl < MIN_SHARES)
4724 * wl = dw_i = S * (s'_i - s_i); see (3)
4726 wl -= se->avg.load_avg;
4729 * Recursively apply this logic to all parent groups to compute
4730 * the final effective load change on the root group. Since
4731 * only the @tg group gets extra weight, all parent groups can
4732 * only redistribute existing shares. @wl is the shift in shares
4733 * resulting from this level per the above.
4742 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4750 * Returns the current capacity of cpu after applying both
4751 * cpu and freq scaling.
4753 unsigned long capacity_curr_of(int cpu)
4755 return cpu_rq(cpu)->cpu_capacity_orig *
4756 arch_scale_freq_capacity(NULL, cpu)
4757 >> SCHED_CAPACITY_SHIFT;
4760 static inline bool energy_aware(void)
4762 return sched_feat(ENERGY_AWARE);
4766 struct sched_group *sg_top;
4767 struct sched_group *sg_cap;
4774 struct task_struct *task;
4789 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4790 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4791 * energy calculations. Using the scale-invariant util returned by
4792 * cpu_util() and approximating scale-invariant util by:
4794 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4796 * the normalized util can be found using the specific capacity.
4798 * capacity = capacity_orig * curr_freq/max_freq
4800 * norm_util = running_time/time ~ util/capacity
4802 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4804 int util = __cpu_util(cpu, delta);
4806 if (util >= capacity)
4807 return SCHED_CAPACITY_SCALE;
4809 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4812 static int calc_util_delta(struct energy_env *eenv, int cpu)
4814 if (cpu == eenv->src_cpu)
4815 return -eenv->util_delta;
4816 if (cpu == eenv->dst_cpu)
4817 return eenv->util_delta;
4822 unsigned long group_max_util(struct energy_env *eenv)
4825 unsigned long max_util = 0;
4827 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4828 delta = calc_util_delta(eenv, i);
4829 max_util = max(max_util, __cpu_util(i, delta));
4836 * group_norm_util() returns the approximated group util relative to it's
4837 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4838 * energy calculations. Since task executions may or may not overlap in time in
4839 * the group the true normalized util is between max(cpu_norm_util(i)) and
4840 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4841 * latter is used as the estimate as it leads to a more pessimistic energy
4842 * estimate (more busy).
4845 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4848 unsigned long util_sum = 0;
4849 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4851 for_each_cpu(i, sched_group_cpus(sg)) {
4852 delta = calc_util_delta(eenv, i);
4853 util_sum += __cpu_norm_util(i, capacity, delta);
4856 if (util_sum > SCHED_CAPACITY_SCALE)
4857 return SCHED_CAPACITY_SCALE;
4861 static int find_new_capacity(struct energy_env *eenv,
4862 const struct sched_group_energy const *sge)
4865 unsigned long util = group_max_util(eenv);
4867 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4868 if (sge->cap_states[idx].cap >= util)
4872 eenv->cap_idx = idx;
4877 static int group_idle_state(struct sched_group *sg)
4879 int i, state = INT_MAX;
4881 /* Find the shallowest idle state in the sched group. */
4882 for_each_cpu(i, sched_group_cpus(sg))
4883 state = min(state, idle_get_state_idx(cpu_rq(i)));
4885 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4892 * sched_group_energy(): Computes the absolute energy consumption of cpus
4893 * belonging to the sched_group including shared resources shared only by
4894 * members of the group. Iterates over all cpus in the hierarchy below the
4895 * sched_group starting from the bottom working it's way up before going to
4896 * the next cpu until all cpus are covered at all levels. The current
4897 * implementation is likely to gather the same util statistics multiple times.
4898 * This can probably be done in a faster but more complex way.
4899 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4901 static int sched_group_energy(struct energy_env *eenv)
4903 struct sched_domain *sd;
4904 int cpu, total_energy = 0;
4905 struct cpumask visit_cpus;
4906 struct sched_group *sg;
4908 WARN_ON(!eenv->sg_top->sge);
4910 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4912 while (!cpumask_empty(&visit_cpus)) {
4913 struct sched_group *sg_shared_cap = NULL;
4915 cpu = cpumask_first(&visit_cpus);
4918 * Is the group utilization affected by cpus outside this
4921 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4925 * We most probably raced with hotplug; returning a
4926 * wrong energy estimation is better than entering an
4932 sg_shared_cap = sd->parent->groups;
4934 for_each_domain(cpu, sd) {
4937 /* Has this sched_domain already been visited? */
4938 if (sd->child && group_first_cpu(sg) != cpu)
4942 unsigned long group_util;
4943 int sg_busy_energy, sg_idle_energy;
4944 int cap_idx, idle_idx;
4946 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4947 eenv->sg_cap = sg_shared_cap;
4951 cap_idx = find_new_capacity(eenv, sg->sge);
4953 if (sg->group_weight == 1) {
4954 /* Remove capacity of src CPU (before task move) */
4955 if (eenv->util_delta == 0 &&
4956 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4957 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4958 eenv->cap.delta -= eenv->cap.before;
4960 /* Add capacity of dst CPU (after task move) */
4961 if (eenv->util_delta != 0 &&
4962 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4963 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4964 eenv->cap.delta += eenv->cap.after;
4968 idle_idx = group_idle_state(sg);
4969 group_util = group_norm_util(eenv, sg);
4970 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4971 >> SCHED_CAPACITY_SHIFT;
4972 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4973 * sg->sge->idle_states[idle_idx].power)
4974 >> SCHED_CAPACITY_SHIFT;
4976 total_energy += sg_busy_energy + sg_idle_energy;
4979 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4981 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4984 } while (sg = sg->next, sg != sd->groups);
4990 eenv->energy = total_energy;
4994 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4996 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5000 * energy_diff(): Estimate the energy impact of changing the utilization
5001 * distribution. eenv specifies the change: utilisation amount, source, and
5002 * destination cpu. Source or destination cpu may be -1 in which case the
5003 * utilization is removed from or added to the system (e.g. task wake-up). If
5004 * both are specified, the utilization is migrated.
5006 static inline int __energy_diff(struct energy_env *eenv)
5008 struct sched_domain *sd;
5009 struct sched_group *sg;
5010 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5012 struct energy_env eenv_before = {
5014 .src_cpu = eenv->src_cpu,
5015 .dst_cpu = eenv->dst_cpu,
5016 .nrg = { 0, 0, 0, 0},
5020 if (eenv->src_cpu == eenv->dst_cpu)
5023 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5024 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5027 return 0; /* Error */
5032 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5033 eenv_before.sg_top = eenv->sg_top = sg;
5035 if (sched_group_energy(&eenv_before))
5036 return 0; /* Invalid result abort */
5037 energy_before += eenv_before.energy;
5039 /* Keep track of SRC cpu (before) capacity */
5040 eenv->cap.before = eenv_before.cap.before;
5041 eenv->cap.delta = eenv_before.cap.delta;
5043 if (sched_group_energy(eenv))
5044 return 0; /* Invalid result abort */
5045 energy_after += eenv->energy;
5047 } while (sg = sg->next, sg != sd->groups);
5049 eenv->nrg.before = energy_before;
5050 eenv->nrg.after = energy_after;
5051 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5054 trace_sched_energy_diff(eenv->task,
5055 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5056 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5057 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5058 eenv->nrg.delta, eenv->payoff);
5060 return eenv->nrg.diff;
5063 #ifdef CONFIG_SCHED_TUNE
5065 struct target_nrg schedtune_target_nrg;
5068 * System energy normalization
5069 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5070 * corresponding to the specified energy variation.
5073 normalize_energy(int energy_diff)
5076 #ifdef CONFIG_SCHED_DEBUG
5079 /* Check for boundaries */
5080 max_delta = schedtune_target_nrg.max_power;
5081 max_delta -= schedtune_target_nrg.min_power;
5082 WARN_ON(abs(energy_diff) >= max_delta);
5085 /* Do scaling using positive numbers to increase the range */
5086 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5088 /* Scale by energy magnitude */
5089 normalized_nrg <<= SCHED_LOAD_SHIFT;
5091 /* Normalize on max energy for target platform */
5092 normalized_nrg = reciprocal_divide(
5093 normalized_nrg, schedtune_target_nrg.rdiv);
5095 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5099 energy_diff(struct energy_env *eenv)
5104 /* Conpute "absolute" energy diff */
5105 __energy_diff(eenv);
5107 /* Return energy diff when boost margin is 0 */
5108 #ifdef CONFIG_CGROUP_SCHEDTUNE
5109 boost = schedtune_task_boost(eenv->task);
5111 boost = get_sysctl_sched_cfs_boost();
5114 return eenv->nrg.diff;
5116 /* Compute normalized energy diff */
5117 nrg_delta = normalize_energy(eenv->nrg.diff);
5118 eenv->nrg.delta = nrg_delta;
5120 eenv->payoff = schedtune_accept_deltas(
5126 * When SchedTune is enabled, the energy_diff() function will return
5127 * the computed energy payoff value. Since the energy_diff() return
5128 * value is expected to be negative by its callers, this evaluation
5129 * function return a negative value each time the evaluation return a
5130 * positive payoff, which is the condition for the acceptance of
5131 * a scheduling decision
5133 return -eenv->payoff;
5135 #else /* CONFIG_SCHED_TUNE */
5136 #define energy_diff(eenv) __energy_diff(eenv)
5140 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5141 * A waker of many should wake a different task than the one last awakened
5142 * at a frequency roughly N times higher than one of its wakees. In order
5143 * to determine whether we should let the load spread vs consolodating to
5144 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5145 * partner, and a factor of lls_size higher frequency in the other. With
5146 * both conditions met, we can be relatively sure that the relationship is
5147 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5148 * being client/server, worker/dispatcher, interrupt source or whatever is
5149 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5151 static int wake_wide(struct task_struct *p)
5153 unsigned int master = current->wakee_flips;
5154 unsigned int slave = p->wakee_flips;
5155 int factor = this_cpu_read(sd_llc_size);
5158 swap(master, slave);
5159 if (slave < factor || master < slave * factor)
5164 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5166 s64 this_load, load;
5167 s64 this_eff_load, prev_eff_load;
5168 int idx, this_cpu, prev_cpu;
5169 struct task_group *tg;
5170 unsigned long weight;
5174 this_cpu = smp_processor_id();
5175 prev_cpu = task_cpu(p);
5176 load = source_load(prev_cpu, idx);
5177 this_load = target_load(this_cpu, idx);
5180 * If sync wakeup then subtract the (maximum possible)
5181 * effect of the currently running task from the load
5182 * of the current CPU:
5185 tg = task_group(current);
5186 weight = current->se.avg.load_avg;
5188 this_load += effective_load(tg, this_cpu, -weight, -weight);
5189 load += effective_load(tg, prev_cpu, 0, -weight);
5193 weight = p->se.avg.load_avg;
5196 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5197 * due to the sync cause above having dropped this_load to 0, we'll
5198 * always have an imbalance, but there's really nothing you can do
5199 * about that, so that's good too.
5201 * Otherwise check if either cpus are near enough in load to allow this
5202 * task to be woken on this_cpu.
5204 this_eff_load = 100;
5205 this_eff_load *= capacity_of(prev_cpu);
5207 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5208 prev_eff_load *= capacity_of(this_cpu);
5210 if (this_load > 0) {
5211 this_eff_load *= this_load +
5212 effective_load(tg, this_cpu, weight, weight);
5214 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5217 balanced = this_eff_load <= prev_eff_load;
5219 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5224 schedstat_inc(sd, ttwu_move_affine);
5225 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5230 static inline unsigned long task_util(struct task_struct *p)
5232 return p->se.avg.util_avg;
5235 unsigned int capacity_margin = 1280; /* ~20% margin */
5237 static inline unsigned long boosted_task_util(struct task_struct *task);
5239 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5241 unsigned long capacity = capacity_of(cpu);
5243 util += boosted_task_util(p);
5245 return (capacity * 1024) > (util * capacity_margin);
5248 static inline bool task_fits_max(struct task_struct *p, int cpu)
5250 unsigned long capacity = capacity_of(cpu);
5251 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5253 if (capacity == max_capacity)
5256 if (capacity * capacity_margin > max_capacity * 1024)
5259 return __task_fits(p, cpu, 0);
5262 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5264 return __task_fits(p, cpu, cpu_util(cpu));
5267 static bool cpu_overutilized(int cpu)
5269 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5272 #ifdef CONFIG_SCHED_TUNE
5274 static unsigned long
5275 schedtune_margin(unsigned long signal, unsigned long boost)
5277 unsigned long long margin = 0;
5280 * Signal proportional compensation (SPC)
5282 * The Boost (B) value is used to compute a Margin (M) which is
5283 * proportional to the complement of the original Signal (S):
5284 * M = B * (SCHED_LOAD_SCALE - S)
5285 * The obtained M could be used by the caller to "boost" S.
5287 margin = SCHED_LOAD_SCALE - signal;
5291 * Fast integer division by constant:
5292 * Constant : (C) = 100
5293 * Precision : 0.1% (P) = 0.1
5294 * Reference : C * 100 / P (R) = 100000
5297 * Shift bits : ceil(log(R,2)) (S) = 17
5298 * Mult const : round(2^S/C) (M) = 1311
5308 static inline unsigned int
5309 schedtune_cpu_margin(unsigned long util, int cpu)
5313 #ifdef CONFIG_CGROUP_SCHEDTUNE
5314 boost = schedtune_cpu_boost(cpu);
5316 boost = get_sysctl_sched_cfs_boost();
5321 return schedtune_margin(util, boost);
5324 static inline unsigned long
5325 schedtune_task_margin(struct task_struct *task)
5329 unsigned long margin;
5331 #ifdef CONFIG_CGROUP_SCHEDTUNE
5332 boost = schedtune_task_boost(task);
5334 boost = get_sysctl_sched_cfs_boost();
5339 util = task_util(task);
5340 margin = schedtune_margin(util, boost);
5345 #else /* CONFIG_SCHED_TUNE */
5347 static inline unsigned int
5348 schedtune_cpu_margin(unsigned long util, int cpu)
5353 static inline unsigned int
5354 schedtune_task_margin(struct task_struct *task)
5359 #endif /* CONFIG_SCHED_TUNE */
5361 static inline unsigned long
5362 boosted_cpu_util(int cpu)
5364 unsigned long util = cpu_util(cpu);
5365 unsigned long margin = schedtune_cpu_margin(util, cpu);
5367 trace_sched_boost_cpu(cpu, util, margin);
5369 return util + margin;
5372 static inline unsigned long
5373 boosted_task_util(struct task_struct *task)
5375 unsigned long util = task_util(task);
5376 unsigned long margin = schedtune_task_margin(task);
5378 trace_sched_boost_task(task, util, margin);
5380 return util + margin;
5384 * find_idlest_group finds and returns the least busy CPU group within the
5387 static struct sched_group *
5388 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5389 int this_cpu, int sd_flag)
5391 struct sched_group *idlest = NULL, *group = sd->groups;
5392 struct sched_group *fit_group = NULL, *spare_group = NULL;
5393 unsigned long min_load = ULONG_MAX, this_load = 0;
5394 unsigned long fit_capacity = ULONG_MAX;
5395 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5396 int load_idx = sd->forkexec_idx;
5397 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5399 if (sd_flag & SD_BALANCE_WAKE)
5400 load_idx = sd->wake_idx;
5403 unsigned long load, avg_load, spare_capacity;
5407 /* Skip over this group if it has no CPUs allowed */
5408 if (!cpumask_intersects(sched_group_cpus(group),
5409 tsk_cpus_allowed(p)))
5412 local_group = cpumask_test_cpu(this_cpu,
5413 sched_group_cpus(group));
5415 /* Tally up the load of all CPUs in the group */
5418 for_each_cpu(i, sched_group_cpus(group)) {
5419 /* Bias balancing toward cpus of our domain */
5421 load = source_load(i, load_idx);
5423 load = target_load(i, load_idx);
5428 * Look for most energy-efficient group that can fit
5429 * that can fit the task.
5431 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5432 fit_capacity = capacity_of(i);
5437 * Look for group which has most spare capacity on a
5440 spare_capacity = capacity_of(i) - cpu_util(i);
5441 if (spare_capacity > max_spare_capacity) {
5442 max_spare_capacity = spare_capacity;
5443 spare_group = group;
5447 /* Adjust by relative CPU capacity of the group */
5448 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5451 this_load = avg_load;
5452 } else if (avg_load < min_load) {
5453 min_load = avg_load;
5456 } while (group = group->next, group != sd->groups);
5464 if (!idlest || 100*this_load < imbalance*min_load)
5470 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5473 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5475 unsigned long load, min_load = ULONG_MAX;
5476 unsigned int min_exit_latency = UINT_MAX;
5477 u64 latest_idle_timestamp = 0;
5478 int least_loaded_cpu = this_cpu;
5479 int shallowest_idle_cpu = -1;
5482 /* Traverse only the allowed CPUs */
5483 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5484 if (task_fits_spare(p, i)) {
5485 struct rq *rq = cpu_rq(i);
5486 struct cpuidle_state *idle = idle_get_state(rq);
5487 if (idle && idle->exit_latency < min_exit_latency) {
5489 * We give priority to a CPU whose idle state
5490 * has the smallest exit latency irrespective
5491 * of any idle timestamp.
5493 min_exit_latency = idle->exit_latency;
5494 latest_idle_timestamp = rq->idle_stamp;
5495 shallowest_idle_cpu = i;
5496 } else if (idle_cpu(i) &&
5497 (!idle || idle->exit_latency == min_exit_latency) &&
5498 rq->idle_stamp > latest_idle_timestamp) {
5500 * If equal or no active idle state, then
5501 * the most recently idled CPU might have
5504 latest_idle_timestamp = rq->idle_stamp;
5505 shallowest_idle_cpu = i;
5506 } else if (shallowest_idle_cpu == -1) {
5508 * If we haven't found an idle CPU yet
5509 * pick a non-idle one that can fit the task as
5512 shallowest_idle_cpu = i;
5514 } else if (shallowest_idle_cpu == -1) {
5515 load = weighted_cpuload(i);
5516 if (load < min_load || (load == min_load && i == this_cpu)) {
5518 least_loaded_cpu = i;
5523 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5527 * Try and locate an idle CPU in the sched_domain.
5529 static int select_idle_sibling(struct task_struct *p, int target)
5531 struct sched_domain *sd;
5532 struct sched_group *sg;
5533 int i = task_cpu(p);
5535 int best_idle_cstate = -1;
5536 int best_idle_capacity = INT_MAX;
5538 if (!sysctl_sched_cstate_aware) {
5539 if (idle_cpu(target))
5543 * If the prevous cpu is cache affine and idle, don't be stupid.
5545 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5550 * Otherwise, iterate the domains and find an elegible idle cpu.
5552 sd = rcu_dereference(per_cpu(sd_llc, target));
5553 for_each_lower_domain(sd) {
5556 if (!cpumask_intersects(sched_group_cpus(sg),
5557 tsk_cpus_allowed(p)))
5560 if (sysctl_sched_cstate_aware) {
5561 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5562 struct rq *rq = cpu_rq(i);
5563 int idle_idx = idle_get_state_idx(rq);
5564 unsigned long new_usage = boosted_task_util(p);
5565 unsigned long capacity_orig = capacity_orig_of(i);
5566 if (new_usage > capacity_orig || !idle_cpu(i))
5569 if (i == target && new_usage <= capacity_curr_of(target))
5572 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5574 best_idle_cstate = idle_idx;
5575 best_idle_capacity = capacity_orig;
5579 for_each_cpu(i, sched_group_cpus(sg)) {
5580 if (i == target || !idle_cpu(i))
5584 target = cpumask_first_and(sched_group_cpus(sg),
5585 tsk_cpus_allowed(p));
5590 } while (sg != sd->groups);
5599 static inline int find_best_target(struct task_struct *p)
5602 int target_cpu = -1;
5603 int target_capacity = 0;
5604 int backup_capacity = 0;
5606 int best_idle_cstate = INT_MAX;
5607 int backup_cpu = -1;
5608 unsigned long task_util_boosted, new_util;
5611 * Favor 1) busy cpu with most capacity at current OPP
5612 * 2) idle_cpu with capacity at current OPP
5613 * 3) busy cpu with capacity at higher OPP
5615 #ifdef CONFIG_CGROUP_SCHEDTUNE
5616 boosted = schedtune_task_boost(p);
5620 task_util_boosted = boosted_task_util(p);
5621 for_each_cpu(i, tsk_cpus_allowed(p)) {
5622 int cur_capacity = capacity_curr_of(i);
5623 struct rq *rq = cpu_rq(i);
5624 int idle_idx = idle_get_state_idx(rq);
5627 * p's blocked utilization is still accounted for on prev_cpu
5628 * so prev_cpu will receive a negative bias due to the double
5629 * accounting. However, the blocked utilization may be zero.
5631 new_util = cpu_util(i) + task_util_boosted;
5634 * Ensure minimum capacity to grant the required boost.
5635 * The target CPU can be already at a capacity level higher
5636 * than the one required to boost the task.
5639 if (new_util > capacity_orig_of(i))
5643 * For boosted tasks we favor idle cpus unconditionally to
5646 if (idle_idx >= 0 && boosted) {
5648 (sysctl_sched_cstate_aware &&
5649 best_idle_cstate > idle_idx)) {
5650 best_idle_cstate = idle_idx;
5656 if (new_util < cur_capacity) {
5657 if (cpu_rq(i)->nr_running) {
5658 if (target_capacity == 0 ||
5659 target_capacity > cur_capacity) {
5660 /* busy CPU with most capacity at current OPP */
5662 target_capacity = cur_capacity;
5664 } else if (!boosted) {
5666 (sysctl_sched_cstate_aware &&
5667 best_idle_cstate > idle_idx)) {
5668 best_idle_cstate = idle_idx;
5672 } else if (backup_capacity == 0 ||
5673 backup_capacity > cur_capacity) {
5674 /* first busy CPU with capacity at higher OPP */
5675 backup_capacity = cur_capacity;
5680 if (!boosted && target_cpu < 0) {
5681 target_cpu = idle_cpu >= 0 ? idle_cpu : backup_cpu;
5684 if (boosted && idle_cpu >= 0)
5685 target_cpu = idle_cpu;
5689 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5691 struct sched_domain *sd;
5692 struct sched_group *sg, *sg_target;
5693 int target_max_cap = INT_MAX;
5694 int target_cpu = task_cpu(p);
5695 unsigned long task_util_boosted, new_util;
5698 if (sysctl_sched_sync_hint_enable && sync) {
5699 int cpu = smp_processor_id();
5700 cpumask_t search_cpus;
5701 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5702 if (cpumask_test_cpu(cpu, &search_cpus))
5706 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5714 if (sysctl_sched_is_big_little) {
5717 * Find group with sufficient capacity. We only get here if no cpu is
5718 * overutilized. We may end up overutilizing a cpu by adding the task,
5719 * but that should not be any worse than select_idle_sibling().
5720 * load_balance() should sort it out later as we get above the tipping
5724 /* Assuming all cpus are the same in group */
5725 int max_cap_cpu = group_first_cpu(sg);
5728 * Assume smaller max capacity means more energy-efficient.
5729 * Ideally we should query the energy model for the right
5730 * answer but it easily ends up in an exhaustive search.
5732 if (capacity_of(max_cap_cpu) < target_max_cap &&
5733 task_fits_max(p, max_cap_cpu)) {
5735 target_max_cap = capacity_of(max_cap_cpu);
5737 } while (sg = sg->next, sg != sd->groups);
5739 task_util_boosted = boosted_task_util(p);
5740 /* Find cpu with sufficient capacity */
5741 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5743 * p's blocked utilization is still accounted for on prev_cpu
5744 * so prev_cpu will receive a negative bias due to the double
5745 * accounting. However, the blocked utilization may be zero.
5747 new_util = cpu_util(i) + task_util_boosted;
5750 * Ensure minimum capacity to grant the required boost.
5751 * The target CPU can be already at a capacity level higher
5752 * than the one required to boost the task.
5754 if (new_util > capacity_orig_of(i))
5757 if (new_util < capacity_curr_of(i)) {
5759 if (cpu_rq(i)->nr_running)
5763 /* cpu has capacity at higher OPP, keep it as fallback */
5764 if (target_cpu == task_cpu(p))
5769 * Find a cpu with sufficient capacity
5771 int tmp_target = find_best_target(p);
5772 if (tmp_target >= 0)
5773 target_cpu = tmp_target;
5776 if (target_cpu != task_cpu(p)) {
5777 struct energy_env eenv = {
5778 .util_delta = task_util(p),
5779 .src_cpu = task_cpu(p),
5780 .dst_cpu = target_cpu,
5784 /* Not enough spare capacity on previous cpu */
5785 if (cpu_overutilized(task_cpu(p)))
5788 if (energy_diff(&eenv) >= 0)
5796 * select_task_rq_fair: Select target runqueue for the waking task in domains
5797 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5798 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5800 * Balances load by selecting the idlest cpu in the idlest group, or under
5801 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5803 * Returns the target cpu number.
5805 * preempt must be disabled.
5808 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5810 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5811 int cpu = smp_processor_id();
5812 int new_cpu = prev_cpu;
5813 int want_affine = 0;
5814 int sync = wake_flags & WF_SYNC;
5816 if (sd_flag & SD_BALANCE_WAKE)
5817 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5818 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5822 for_each_domain(cpu, tmp) {
5823 if (!(tmp->flags & SD_LOAD_BALANCE))
5827 * If both cpu and prev_cpu are part of this domain,
5828 * cpu is a valid SD_WAKE_AFFINE target.
5830 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5831 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5836 if (tmp->flags & sd_flag)
5838 else if (!want_affine)
5843 sd = NULL; /* Prefer wake_affine over balance flags */
5844 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5849 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5850 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5851 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5852 new_cpu = select_idle_sibling(p, new_cpu);
5855 struct sched_group *group;
5858 if (!(sd->flags & sd_flag)) {
5863 group = find_idlest_group(sd, p, cpu, sd_flag);
5869 new_cpu = find_idlest_cpu(group, p, cpu);
5870 if (new_cpu == -1 || new_cpu == cpu) {
5871 /* Now try balancing at a lower domain level of cpu */
5876 /* Now try balancing at a lower domain level of new_cpu */
5878 weight = sd->span_weight;
5880 for_each_domain(cpu, tmp) {
5881 if (weight <= tmp->span_weight)
5883 if (tmp->flags & sd_flag)
5886 /* while loop will break here if sd == NULL */
5894 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5895 * cfs_rq_of(p) references at time of call are still valid and identify the
5896 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5897 * other assumptions, including the state of rq->lock, should be made.
5899 static void migrate_task_rq_fair(struct task_struct *p)
5902 * We are supposed to update the task to "current" time, then its up to date
5903 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5904 * what current time is, so simply throw away the out-of-date time. This
5905 * will result in the wakee task is less decayed, but giving the wakee more
5906 * load sounds not bad.
5908 remove_entity_load_avg(&p->se);
5910 /* Tell new CPU we are migrated */
5911 p->se.avg.last_update_time = 0;
5913 /* We have migrated, no longer consider this task hot */
5914 p->se.exec_start = 0;
5917 static void task_dead_fair(struct task_struct *p)
5919 remove_entity_load_avg(&p->se);
5922 #define task_fits_max(p, cpu) true
5923 #endif /* CONFIG_SMP */
5925 static unsigned long
5926 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5928 unsigned long gran = sysctl_sched_wakeup_granularity;
5931 * Since its curr running now, convert the gran from real-time
5932 * to virtual-time in his units.
5934 * By using 'se' instead of 'curr' we penalize light tasks, so
5935 * they get preempted easier. That is, if 'se' < 'curr' then
5936 * the resulting gran will be larger, therefore penalizing the
5937 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5938 * be smaller, again penalizing the lighter task.
5940 * This is especially important for buddies when the leftmost
5941 * task is higher priority than the buddy.
5943 return calc_delta_fair(gran, se);
5947 * Should 'se' preempt 'curr'.
5961 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5963 s64 gran, vdiff = curr->vruntime - se->vruntime;
5968 gran = wakeup_gran(curr, se);
5975 static void set_last_buddy(struct sched_entity *se)
5977 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5980 for_each_sched_entity(se)
5981 cfs_rq_of(se)->last = se;
5984 static void set_next_buddy(struct sched_entity *se)
5986 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5989 for_each_sched_entity(se)
5990 cfs_rq_of(se)->next = se;
5993 static void set_skip_buddy(struct sched_entity *se)
5995 for_each_sched_entity(se)
5996 cfs_rq_of(se)->skip = se;
6000 * Preempt the current task with a newly woken task if needed:
6002 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6004 struct task_struct *curr = rq->curr;
6005 struct sched_entity *se = &curr->se, *pse = &p->se;
6006 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6007 int scale = cfs_rq->nr_running >= sched_nr_latency;
6008 int next_buddy_marked = 0;
6010 if (unlikely(se == pse))
6014 * This is possible from callers such as attach_tasks(), in which we
6015 * unconditionally check_prempt_curr() after an enqueue (which may have
6016 * lead to a throttle). This both saves work and prevents false
6017 * next-buddy nomination below.
6019 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6022 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6023 set_next_buddy(pse);
6024 next_buddy_marked = 1;
6028 * We can come here with TIF_NEED_RESCHED already set from new task
6031 * Note: this also catches the edge-case of curr being in a throttled
6032 * group (e.g. via set_curr_task), since update_curr() (in the
6033 * enqueue of curr) will have resulted in resched being set. This
6034 * prevents us from potentially nominating it as a false LAST_BUDDY
6037 if (test_tsk_need_resched(curr))
6040 /* Idle tasks are by definition preempted by non-idle tasks. */
6041 if (unlikely(curr->policy == SCHED_IDLE) &&
6042 likely(p->policy != SCHED_IDLE))
6046 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6047 * is driven by the tick):
6049 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6052 find_matching_se(&se, &pse);
6053 update_curr(cfs_rq_of(se));
6055 if (wakeup_preempt_entity(se, pse) == 1) {
6057 * Bias pick_next to pick the sched entity that is
6058 * triggering this preemption.
6060 if (!next_buddy_marked)
6061 set_next_buddy(pse);
6070 * Only set the backward buddy when the current task is still
6071 * on the rq. This can happen when a wakeup gets interleaved
6072 * with schedule on the ->pre_schedule() or idle_balance()
6073 * point, either of which can * drop the rq lock.
6075 * Also, during early boot the idle thread is in the fair class,
6076 * for obvious reasons its a bad idea to schedule back to it.
6078 if (unlikely(!se->on_rq || curr == rq->idle))
6081 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6085 static struct task_struct *
6086 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6088 struct cfs_rq *cfs_rq = &rq->cfs;
6089 struct sched_entity *se;
6090 struct task_struct *p;
6094 #ifdef CONFIG_FAIR_GROUP_SCHED
6095 if (!cfs_rq->nr_running)
6098 if (prev->sched_class != &fair_sched_class)
6102 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6103 * likely that a next task is from the same cgroup as the current.
6105 * Therefore attempt to avoid putting and setting the entire cgroup
6106 * hierarchy, only change the part that actually changes.
6110 struct sched_entity *curr = cfs_rq->curr;
6113 * Since we got here without doing put_prev_entity() we also
6114 * have to consider cfs_rq->curr. If it is still a runnable
6115 * entity, update_curr() will update its vruntime, otherwise
6116 * forget we've ever seen it.
6120 update_curr(cfs_rq);
6125 * This call to check_cfs_rq_runtime() will do the
6126 * throttle and dequeue its entity in the parent(s).
6127 * Therefore the 'simple' nr_running test will indeed
6130 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6134 se = pick_next_entity(cfs_rq, curr);
6135 cfs_rq = group_cfs_rq(se);
6141 * Since we haven't yet done put_prev_entity and if the selected task
6142 * is a different task than we started out with, try and touch the
6143 * least amount of cfs_rqs.
6146 struct sched_entity *pse = &prev->se;
6148 while (!(cfs_rq = is_same_group(se, pse))) {
6149 int se_depth = se->depth;
6150 int pse_depth = pse->depth;
6152 if (se_depth <= pse_depth) {
6153 put_prev_entity(cfs_rq_of(pse), pse);
6154 pse = parent_entity(pse);
6156 if (se_depth >= pse_depth) {
6157 set_next_entity(cfs_rq_of(se), se);
6158 se = parent_entity(se);
6162 put_prev_entity(cfs_rq, pse);
6163 set_next_entity(cfs_rq, se);
6166 if (hrtick_enabled(rq))
6167 hrtick_start_fair(rq, p);
6169 rq->misfit_task = !task_fits_max(p, rq->cpu);
6176 if (!cfs_rq->nr_running)
6179 put_prev_task(rq, prev);
6182 se = pick_next_entity(cfs_rq, NULL);
6183 set_next_entity(cfs_rq, se);
6184 cfs_rq = group_cfs_rq(se);
6189 if (hrtick_enabled(rq))
6190 hrtick_start_fair(rq, p);
6192 rq->misfit_task = !task_fits_max(p, rq->cpu);
6197 rq->misfit_task = 0;
6199 * This is OK, because current is on_cpu, which avoids it being picked
6200 * for load-balance and preemption/IRQs are still disabled avoiding
6201 * further scheduler activity on it and we're being very careful to
6202 * re-start the picking loop.
6204 lockdep_unpin_lock(&rq->lock);
6205 new_tasks = idle_balance(rq);
6206 lockdep_pin_lock(&rq->lock);
6208 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6209 * possible for any higher priority task to appear. In that case we
6210 * must re-start the pick_next_entity() loop.
6222 * Account for a descheduled task:
6224 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6226 struct sched_entity *se = &prev->se;
6227 struct cfs_rq *cfs_rq;
6229 for_each_sched_entity(se) {
6230 cfs_rq = cfs_rq_of(se);
6231 put_prev_entity(cfs_rq, se);
6236 * sched_yield() is very simple
6238 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6240 static void yield_task_fair(struct rq *rq)
6242 struct task_struct *curr = rq->curr;
6243 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6244 struct sched_entity *se = &curr->se;
6247 * Are we the only task in the tree?
6249 if (unlikely(rq->nr_running == 1))
6252 clear_buddies(cfs_rq, se);
6254 if (curr->policy != SCHED_BATCH) {
6255 update_rq_clock(rq);
6257 * Update run-time statistics of the 'current'.
6259 update_curr(cfs_rq);
6261 * Tell update_rq_clock() that we've just updated,
6262 * so we don't do microscopic update in schedule()
6263 * and double the fastpath cost.
6265 rq_clock_skip_update(rq, true);
6271 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6273 struct sched_entity *se = &p->se;
6275 /* throttled hierarchies are not runnable */
6276 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6279 /* Tell the scheduler that we'd really like pse to run next. */
6282 yield_task_fair(rq);
6288 /**************************************************
6289 * Fair scheduling class load-balancing methods.
6293 * The purpose of load-balancing is to achieve the same basic fairness the
6294 * per-cpu scheduler provides, namely provide a proportional amount of compute
6295 * time to each task. This is expressed in the following equation:
6297 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6299 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6300 * W_i,0 is defined as:
6302 * W_i,0 = \Sum_j w_i,j (2)
6304 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6305 * is derived from the nice value as per prio_to_weight[].
6307 * The weight average is an exponential decay average of the instantaneous
6310 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6312 * C_i is the compute capacity of cpu i, typically it is the
6313 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6314 * can also include other factors [XXX].
6316 * To achieve this balance we define a measure of imbalance which follows
6317 * directly from (1):
6319 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6321 * We them move tasks around to minimize the imbalance. In the continuous
6322 * function space it is obvious this converges, in the discrete case we get
6323 * a few fun cases generally called infeasible weight scenarios.
6326 * - infeasible weights;
6327 * - local vs global optima in the discrete case. ]
6332 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6333 * for all i,j solution, we create a tree of cpus that follows the hardware
6334 * topology where each level pairs two lower groups (or better). This results
6335 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6336 * tree to only the first of the previous level and we decrease the frequency
6337 * of load-balance at each level inv. proportional to the number of cpus in
6343 * \Sum { --- * --- * 2^i } = O(n) (5)
6345 * `- size of each group
6346 * | | `- number of cpus doing load-balance
6348 * `- sum over all levels
6350 * Coupled with a limit on how many tasks we can migrate every balance pass,
6351 * this makes (5) the runtime complexity of the balancer.
6353 * An important property here is that each CPU is still (indirectly) connected
6354 * to every other cpu in at most O(log n) steps:
6356 * The adjacency matrix of the resulting graph is given by:
6359 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6362 * And you'll find that:
6364 * A^(log_2 n)_i,j != 0 for all i,j (7)
6366 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6367 * The task movement gives a factor of O(m), giving a convergence complexity
6370 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6375 * In order to avoid CPUs going idle while there's still work to do, new idle
6376 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6377 * tree itself instead of relying on other CPUs to bring it work.
6379 * This adds some complexity to both (5) and (8) but it reduces the total idle
6387 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6390 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6395 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6397 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6399 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6402 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6403 * rewrite all of this once again.]
6406 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6408 enum fbq_type { regular, remote, all };
6417 #define LBF_ALL_PINNED 0x01
6418 #define LBF_NEED_BREAK 0x02
6419 #define LBF_DST_PINNED 0x04
6420 #define LBF_SOME_PINNED 0x08
6423 struct sched_domain *sd;
6431 struct cpumask *dst_grpmask;
6433 enum cpu_idle_type idle;
6435 unsigned int src_grp_nr_running;
6436 /* The set of CPUs under consideration for load-balancing */
6437 struct cpumask *cpus;
6442 unsigned int loop_break;
6443 unsigned int loop_max;
6445 enum fbq_type fbq_type;
6446 enum group_type busiest_group_type;
6447 struct list_head tasks;
6451 * Is this task likely cache-hot:
6453 static int task_hot(struct task_struct *p, struct lb_env *env)
6457 lockdep_assert_held(&env->src_rq->lock);
6459 if (p->sched_class != &fair_sched_class)
6462 if (unlikely(p->policy == SCHED_IDLE))
6466 * Buddy candidates are cache hot:
6468 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6469 (&p->se == cfs_rq_of(&p->se)->next ||
6470 &p->se == cfs_rq_of(&p->se)->last))
6473 if (sysctl_sched_migration_cost == -1)
6475 if (sysctl_sched_migration_cost == 0)
6478 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6480 return delta < (s64)sysctl_sched_migration_cost;
6483 #ifdef CONFIG_NUMA_BALANCING
6485 * Returns 1, if task migration degrades locality
6486 * Returns 0, if task migration improves locality i.e migration preferred.
6487 * Returns -1, if task migration is not affected by locality.
6489 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6491 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6492 unsigned long src_faults, dst_faults;
6493 int src_nid, dst_nid;
6495 if (!static_branch_likely(&sched_numa_balancing))
6498 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6501 src_nid = cpu_to_node(env->src_cpu);
6502 dst_nid = cpu_to_node(env->dst_cpu);
6504 if (src_nid == dst_nid)
6507 /* Migrating away from the preferred node is always bad. */
6508 if (src_nid == p->numa_preferred_nid) {
6509 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6515 /* Encourage migration to the preferred node. */
6516 if (dst_nid == p->numa_preferred_nid)
6520 src_faults = group_faults(p, src_nid);
6521 dst_faults = group_faults(p, dst_nid);
6523 src_faults = task_faults(p, src_nid);
6524 dst_faults = task_faults(p, dst_nid);
6527 return dst_faults < src_faults;
6531 static inline int migrate_degrades_locality(struct task_struct *p,
6539 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6542 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6546 lockdep_assert_held(&env->src_rq->lock);
6549 * We do not migrate tasks that are:
6550 * 1) throttled_lb_pair, or
6551 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6552 * 3) running (obviously), or
6553 * 4) are cache-hot on their current CPU.
6555 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6558 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6561 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6563 env->flags |= LBF_SOME_PINNED;
6566 * Remember if this task can be migrated to any other cpu in
6567 * our sched_group. We may want to revisit it if we couldn't
6568 * meet load balance goals by pulling other tasks on src_cpu.
6570 * Also avoid computing new_dst_cpu if we have already computed
6571 * one in current iteration.
6573 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6576 /* Prevent to re-select dst_cpu via env's cpus */
6577 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6578 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6579 env->flags |= LBF_DST_PINNED;
6580 env->new_dst_cpu = cpu;
6588 /* Record that we found atleast one task that could run on dst_cpu */
6589 env->flags &= ~LBF_ALL_PINNED;
6591 if (task_running(env->src_rq, p)) {
6592 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6597 * Aggressive migration if:
6598 * 1) destination numa is preferred
6599 * 2) task is cache cold, or
6600 * 3) too many balance attempts have failed.
6602 tsk_cache_hot = migrate_degrades_locality(p, env);
6603 if (tsk_cache_hot == -1)
6604 tsk_cache_hot = task_hot(p, env);
6606 if (tsk_cache_hot <= 0 ||
6607 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6608 if (tsk_cache_hot == 1) {
6609 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6610 schedstat_inc(p, se.statistics.nr_forced_migrations);
6615 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6620 * detach_task() -- detach the task for the migration specified in env
6622 static void detach_task(struct task_struct *p, struct lb_env *env)
6624 lockdep_assert_held(&env->src_rq->lock);
6626 deactivate_task(env->src_rq, p, 0);
6627 p->on_rq = TASK_ON_RQ_MIGRATING;
6628 set_task_cpu(p, env->dst_cpu);
6632 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6633 * part of active balancing operations within "domain".
6635 * Returns a task if successful and NULL otherwise.
6637 static struct task_struct *detach_one_task(struct lb_env *env)
6639 struct task_struct *p, *n;
6641 lockdep_assert_held(&env->src_rq->lock);
6643 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6644 if (!can_migrate_task(p, env))
6647 detach_task(p, env);
6650 * Right now, this is only the second place where
6651 * lb_gained[env->idle] is updated (other is detach_tasks)
6652 * so we can safely collect stats here rather than
6653 * inside detach_tasks().
6655 schedstat_inc(env->sd, lb_gained[env->idle]);
6661 static const unsigned int sched_nr_migrate_break = 32;
6664 * detach_tasks() -- tries to detach up to imbalance weighted load from
6665 * busiest_rq, as part of a balancing operation within domain "sd".
6667 * Returns number of detached tasks if successful and 0 otherwise.
6669 static int detach_tasks(struct lb_env *env)
6671 struct list_head *tasks = &env->src_rq->cfs_tasks;
6672 struct task_struct *p;
6676 lockdep_assert_held(&env->src_rq->lock);
6678 if (env->imbalance <= 0)
6681 while (!list_empty(tasks)) {
6683 * We don't want to steal all, otherwise we may be treated likewise,
6684 * which could at worst lead to a livelock crash.
6686 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6689 p = list_first_entry(tasks, struct task_struct, se.group_node);
6692 /* We've more or less seen every task there is, call it quits */
6693 if (env->loop > env->loop_max)
6696 /* take a breather every nr_migrate tasks */
6697 if (env->loop > env->loop_break) {
6698 env->loop_break += sched_nr_migrate_break;
6699 env->flags |= LBF_NEED_BREAK;
6703 if (!can_migrate_task(p, env))
6706 load = task_h_load(p);
6708 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6711 if ((load / 2) > env->imbalance)
6714 detach_task(p, env);
6715 list_add(&p->se.group_node, &env->tasks);
6718 env->imbalance -= load;
6720 #ifdef CONFIG_PREEMPT
6722 * NEWIDLE balancing is a source of latency, so preemptible
6723 * kernels will stop after the first task is detached to minimize
6724 * the critical section.
6726 if (env->idle == CPU_NEWLY_IDLE)
6731 * We only want to steal up to the prescribed amount of
6734 if (env->imbalance <= 0)
6739 list_move_tail(&p->se.group_node, tasks);
6743 * Right now, this is one of only two places we collect this stat
6744 * so we can safely collect detach_one_task() stats here rather
6745 * than inside detach_one_task().
6747 schedstat_add(env->sd, lb_gained[env->idle], detached);
6753 * attach_task() -- attach the task detached by detach_task() to its new rq.
6755 static void attach_task(struct rq *rq, struct task_struct *p)
6757 lockdep_assert_held(&rq->lock);
6759 BUG_ON(task_rq(p) != rq);
6760 p->on_rq = TASK_ON_RQ_QUEUED;
6761 activate_task(rq, p, 0);
6762 check_preempt_curr(rq, p, 0);
6766 * attach_one_task() -- attaches the task returned from detach_one_task() to
6769 static void attach_one_task(struct rq *rq, struct task_struct *p)
6771 raw_spin_lock(&rq->lock);
6774 * We want to potentially raise target_cpu's OPP.
6776 update_capacity_of(cpu_of(rq));
6777 raw_spin_unlock(&rq->lock);
6781 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6784 static void attach_tasks(struct lb_env *env)
6786 struct list_head *tasks = &env->tasks;
6787 struct task_struct *p;
6789 raw_spin_lock(&env->dst_rq->lock);
6791 while (!list_empty(tasks)) {
6792 p = list_first_entry(tasks, struct task_struct, se.group_node);
6793 list_del_init(&p->se.group_node);
6795 attach_task(env->dst_rq, p);
6799 * We want to potentially raise env.dst_cpu's OPP.
6801 update_capacity_of(env->dst_cpu);
6803 raw_spin_unlock(&env->dst_rq->lock);
6806 #ifdef CONFIG_FAIR_GROUP_SCHED
6807 static void update_blocked_averages(int cpu)
6809 struct rq *rq = cpu_rq(cpu);
6810 struct cfs_rq *cfs_rq;
6811 unsigned long flags;
6813 raw_spin_lock_irqsave(&rq->lock, flags);
6814 update_rq_clock(rq);
6817 * Iterates the task_group tree in a bottom up fashion, see
6818 * list_add_leaf_cfs_rq() for details.
6820 for_each_leaf_cfs_rq(rq, cfs_rq) {
6821 /* throttled entities do not contribute to load */
6822 if (throttled_hierarchy(cfs_rq))
6825 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6826 update_tg_load_avg(cfs_rq, 0);
6828 raw_spin_unlock_irqrestore(&rq->lock, flags);
6832 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6833 * This needs to be done in a top-down fashion because the load of a child
6834 * group is a fraction of its parents load.
6836 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6838 struct rq *rq = rq_of(cfs_rq);
6839 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6840 unsigned long now = jiffies;
6843 if (cfs_rq->last_h_load_update == now)
6846 cfs_rq->h_load_next = NULL;
6847 for_each_sched_entity(se) {
6848 cfs_rq = cfs_rq_of(se);
6849 cfs_rq->h_load_next = se;
6850 if (cfs_rq->last_h_load_update == now)
6855 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6856 cfs_rq->last_h_load_update = now;
6859 while ((se = cfs_rq->h_load_next) != NULL) {
6860 load = cfs_rq->h_load;
6861 load = div64_ul(load * se->avg.load_avg,
6862 cfs_rq_load_avg(cfs_rq) + 1);
6863 cfs_rq = group_cfs_rq(se);
6864 cfs_rq->h_load = load;
6865 cfs_rq->last_h_load_update = now;
6869 static unsigned long task_h_load(struct task_struct *p)
6871 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6873 update_cfs_rq_h_load(cfs_rq);
6874 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6875 cfs_rq_load_avg(cfs_rq) + 1);
6878 static inline void update_blocked_averages(int cpu)
6880 struct rq *rq = cpu_rq(cpu);
6881 struct cfs_rq *cfs_rq = &rq->cfs;
6882 unsigned long flags;
6884 raw_spin_lock_irqsave(&rq->lock, flags);
6885 update_rq_clock(rq);
6886 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6887 raw_spin_unlock_irqrestore(&rq->lock, flags);
6890 static unsigned long task_h_load(struct task_struct *p)
6892 return p->se.avg.load_avg;
6896 /********** Helpers for find_busiest_group ************************/
6899 * sg_lb_stats - stats of a sched_group required for load_balancing
6901 struct sg_lb_stats {
6902 unsigned long avg_load; /*Avg load across the CPUs of the group */
6903 unsigned long group_load; /* Total load over the CPUs of the group */
6904 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6905 unsigned long load_per_task;
6906 unsigned long group_capacity;
6907 unsigned long group_util; /* Total utilization of the group */
6908 unsigned int sum_nr_running; /* Nr tasks running in the group */
6909 unsigned int idle_cpus;
6910 unsigned int group_weight;
6911 enum group_type group_type;
6912 int group_no_capacity;
6913 int group_misfit_task; /* A cpu has a task too big for its capacity */
6914 #ifdef CONFIG_NUMA_BALANCING
6915 unsigned int nr_numa_running;
6916 unsigned int nr_preferred_running;
6921 * sd_lb_stats - Structure to store the statistics of a sched_domain
6922 * during load balancing.
6924 struct sd_lb_stats {
6925 struct sched_group *busiest; /* Busiest group in this sd */
6926 struct sched_group *local; /* Local group in this sd */
6927 unsigned long total_load; /* Total load of all groups in sd */
6928 unsigned long total_capacity; /* Total capacity of all groups in sd */
6929 unsigned long avg_load; /* Average load across all groups in sd */
6931 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6932 struct sg_lb_stats local_stat; /* Statistics of the local group */
6935 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6938 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6939 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6940 * We must however clear busiest_stat::avg_load because
6941 * update_sd_pick_busiest() reads this before assignment.
6943 *sds = (struct sd_lb_stats){
6947 .total_capacity = 0UL,
6950 .sum_nr_running = 0,
6951 .group_type = group_other,
6957 * get_sd_load_idx - Obtain the load index for a given sched domain.
6958 * @sd: The sched_domain whose load_idx is to be obtained.
6959 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6961 * Return: The load index.
6963 static inline int get_sd_load_idx(struct sched_domain *sd,
6964 enum cpu_idle_type idle)
6970 load_idx = sd->busy_idx;
6973 case CPU_NEWLY_IDLE:
6974 load_idx = sd->newidle_idx;
6977 load_idx = sd->idle_idx;
6984 static unsigned long scale_rt_capacity(int cpu)
6986 struct rq *rq = cpu_rq(cpu);
6987 u64 total, used, age_stamp, avg;
6991 * Since we're reading these variables without serialization make sure
6992 * we read them once before doing sanity checks on them.
6994 age_stamp = READ_ONCE(rq->age_stamp);
6995 avg = READ_ONCE(rq->rt_avg);
6996 delta = __rq_clock_broken(rq) - age_stamp;
6998 if (unlikely(delta < 0))
7001 total = sched_avg_period() + delta;
7003 used = div_u64(avg, total);
7006 * deadline bandwidth is defined at system level so we must
7007 * weight this bandwidth with the max capacity of the system.
7008 * As a reminder, avg_bw is 20bits width and
7009 * scale_cpu_capacity is 10 bits width
7011 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7013 if (likely(used < SCHED_CAPACITY_SCALE))
7014 return SCHED_CAPACITY_SCALE - used;
7019 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7021 raw_spin_lock_init(&mcc->lock);
7026 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7028 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7029 struct sched_group *sdg = sd->groups;
7030 struct max_cpu_capacity *mcc;
7031 unsigned long max_capacity;
7033 unsigned long flags;
7035 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7037 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7039 raw_spin_lock_irqsave(&mcc->lock, flags);
7040 max_capacity = mcc->val;
7041 max_cap_cpu = mcc->cpu;
7043 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7044 (max_capacity < capacity)) {
7045 mcc->val = capacity;
7047 #ifdef CONFIG_SCHED_DEBUG
7048 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7049 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
7053 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7055 skip_unlock: __attribute__ ((unused));
7056 capacity *= scale_rt_capacity(cpu);
7057 capacity >>= SCHED_CAPACITY_SHIFT;
7062 cpu_rq(cpu)->cpu_capacity = capacity;
7063 sdg->sgc->capacity = capacity;
7064 sdg->sgc->max_capacity = capacity;
7067 void update_group_capacity(struct sched_domain *sd, int cpu)
7069 struct sched_domain *child = sd->child;
7070 struct sched_group *group, *sdg = sd->groups;
7071 unsigned long capacity, max_capacity;
7072 unsigned long interval;
7074 interval = msecs_to_jiffies(sd->balance_interval);
7075 interval = clamp(interval, 1UL, max_load_balance_interval);
7076 sdg->sgc->next_update = jiffies + interval;
7079 update_cpu_capacity(sd, cpu);
7086 if (child->flags & SD_OVERLAP) {
7088 * SD_OVERLAP domains cannot assume that child groups
7089 * span the current group.
7092 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7093 struct sched_group_capacity *sgc;
7094 struct rq *rq = cpu_rq(cpu);
7097 * build_sched_domains() -> init_sched_groups_capacity()
7098 * gets here before we've attached the domains to the
7101 * Use capacity_of(), which is set irrespective of domains
7102 * in update_cpu_capacity().
7104 * This avoids capacity from being 0 and
7105 * causing divide-by-zero issues on boot.
7107 if (unlikely(!rq->sd)) {
7108 capacity += capacity_of(cpu);
7110 sgc = rq->sd->groups->sgc;
7111 capacity += sgc->capacity;
7114 max_capacity = max(capacity, max_capacity);
7118 * !SD_OVERLAP domains can assume that child groups
7119 * span the current group.
7122 group = child->groups;
7124 struct sched_group_capacity *sgc = group->sgc;
7126 capacity += sgc->capacity;
7127 max_capacity = max(sgc->max_capacity, max_capacity);
7128 group = group->next;
7129 } while (group != child->groups);
7132 sdg->sgc->capacity = capacity;
7133 sdg->sgc->max_capacity = max_capacity;
7137 * Check whether the capacity of the rq has been noticeably reduced by side
7138 * activity. The imbalance_pct is used for the threshold.
7139 * Return true is the capacity is reduced
7142 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7144 return ((rq->cpu_capacity * sd->imbalance_pct) <
7145 (rq->cpu_capacity_orig * 100));
7149 * Group imbalance indicates (and tries to solve) the problem where balancing
7150 * groups is inadequate due to tsk_cpus_allowed() constraints.
7152 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7153 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7156 * { 0 1 2 3 } { 4 5 6 7 }
7159 * If we were to balance group-wise we'd place two tasks in the first group and
7160 * two tasks in the second group. Clearly this is undesired as it will overload
7161 * cpu 3 and leave one of the cpus in the second group unused.
7163 * The current solution to this issue is detecting the skew in the first group
7164 * by noticing the lower domain failed to reach balance and had difficulty
7165 * moving tasks due to affinity constraints.
7167 * When this is so detected; this group becomes a candidate for busiest; see
7168 * update_sd_pick_busiest(). And calculate_imbalance() and
7169 * find_busiest_group() avoid some of the usual balance conditions to allow it
7170 * to create an effective group imbalance.
7172 * This is a somewhat tricky proposition since the next run might not find the
7173 * group imbalance and decide the groups need to be balanced again. A most
7174 * subtle and fragile situation.
7177 static inline int sg_imbalanced(struct sched_group *group)
7179 return group->sgc->imbalance;
7183 * group_has_capacity returns true if the group has spare capacity that could
7184 * be used by some tasks.
7185 * We consider that a group has spare capacity if the * number of task is
7186 * smaller than the number of CPUs or if the utilization is lower than the
7187 * available capacity for CFS tasks.
7188 * For the latter, we use a threshold to stabilize the state, to take into
7189 * account the variance of the tasks' load and to return true if the available
7190 * capacity in meaningful for the load balancer.
7191 * As an example, an available capacity of 1% can appear but it doesn't make
7192 * any benefit for the load balance.
7195 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7197 if (sgs->sum_nr_running < sgs->group_weight)
7200 if ((sgs->group_capacity * 100) >
7201 (sgs->group_util * env->sd->imbalance_pct))
7208 * group_is_overloaded returns true if the group has more tasks than it can
7210 * group_is_overloaded is not equals to !group_has_capacity because a group
7211 * with the exact right number of tasks, has no more spare capacity but is not
7212 * overloaded so both group_has_capacity and group_is_overloaded return
7216 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7218 if (sgs->sum_nr_running <= sgs->group_weight)
7221 if ((sgs->group_capacity * 100) <
7222 (sgs->group_util * env->sd->imbalance_pct))
7230 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7231 * per-cpu capacity than sched_group ref.
7234 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7236 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7237 ref->sgc->max_capacity;
7241 group_type group_classify(struct sched_group *group,
7242 struct sg_lb_stats *sgs)
7244 if (sgs->group_no_capacity)
7245 return group_overloaded;
7247 if (sg_imbalanced(group))
7248 return group_imbalanced;
7250 if (sgs->group_misfit_task)
7251 return group_misfit_task;
7257 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7258 * @env: The load balancing environment.
7259 * @group: sched_group whose statistics are to be updated.
7260 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7261 * @local_group: Does group contain this_cpu.
7262 * @sgs: variable to hold the statistics for this group.
7263 * @overload: Indicate more than one runnable task for any CPU.
7264 * @overutilized: Indicate overutilization for any CPU.
7266 static inline void update_sg_lb_stats(struct lb_env *env,
7267 struct sched_group *group, int load_idx,
7268 int local_group, struct sg_lb_stats *sgs,
7269 bool *overload, bool *overutilized)
7274 memset(sgs, 0, sizeof(*sgs));
7276 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7277 struct rq *rq = cpu_rq(i);
7279 /* Bias balancing toward cpus of our domain */
7281 load = target_load(i, load_idx);
7283 load = source_load(i, load_idx);
7285 sgs->group_load += load;
7286 sgs->group_util += cpu_util(i);
7287 sgs->sum_nr_running += rq->cfs.h_nr_running;
7289 if (rq->nr_running > 1)
7292 #ifdef CONFIG_NUMA_BALANCING
7293 sgs->nr_numa_running += rq->nr_numa_running;
7294 sgs->nr_preferred_running += rq->nr_preferred_running;
7296 sgs->sum_weighted_load += weighted_cpuload(i);
7300 if (cpu_overutilized(i)) {
7301 *overutilized = true;
7302 if (!sgs->group_misfit_task && rq->misfit_task)
7303 sgs->group_misfit_task = capacity_of(i);
7307 /* Adjust by relative CPU capacity of the group */
7308 sgs->group_capacity = group->sgc->capacity;
7309 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7311 if (sgs->sum_nr_running)
7312 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7314 sgs->group_weight = group->group_weight;
7316 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7317 sgs->group_type = group_classify(group, sgs);
7321 * update_sd_pick_busiest - return 1 on busiest group
7322 * @env: The load balancing environment.
7323 * @sds: sched_domain statistics
7324 * @sg: sched_group candidate to be checked for being the busiest
7325 * @sgs: sched_group statistics
7327 * Determine if @sg is a busier group than the previously selected
7330 * Return: %true if @sg is a busier group than the previously selected
7331 * busiest group. %false otherwise.
7333 static bool update_sd_pick_busiest(struct lb_env *env,
7334 struct sd_lb_stats *sds,
7335 struct sched_group *sg,
7336 struct sg_lb_stats *sgs)
7338 struct sg_lb_stats *busiest = &sds->busiest_stat;
7340 if (sgs->group_type > busiest->group_type)
7343 if (sgs->group_type < busiest->group_type)
7347 * Candidate sg doesn't face any serious load-balance problems
7348 * so don't pick it if the local sg is already filled up.
7350 if (sgs->group_type == group_other &&
7351 !group_has_capacity(env, &sds->local_stat))
7354 if (sgs->avg_load <= busiest->avg_load)
7358 * Candiate sg has no more than one task per cpu and has higher
7359 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7361 if (sgs->sum_nr_running <= sgs->group_weight &&
7362 group_smaller_cpu_capacity(sds->local, sg))
7365 /* This is the busiest node in its class. */
7366 if (!(env->sd->flags & SD_ASYM_PACKING))
7370 * ASYM_PACKING needs to move all the work to the lowest
7371 * numbered CPUs in the group, therefore mark all groups
7372 * higher than ourself as busy.
7374 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7378 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7385 #ifdef CONFIG_NUMA_BALANCING
7386 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7388 if (sgs->sum_nr_running > sgs->nr_numa_running)
7390 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7395 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7397 if (rq->nr_running > rq->nr_numa_running)
7399 if (rq->nr_running > rq->nr_preferred_running)
7404 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7409 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7413 #endif /* CONFIG_NUMA_BALANCING */
7416 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7417 * @env: The load balancing environment.
7418 * @sds: variable to hold the statistics for this sched_domain.
7420 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7422 struct sched_domain *child = env->sd->child;
7423 struct sched_group *sg = env->sd->groups;
7424 struct sg_lb_stats tmp_sgs;
7425 int load_idx, prefer_sibling = 0;
7426 bool overload = false, overutilized = false;
7428 if (child && child->flags & SD_PREFER_SIBLING)
7431 load_idx = get_sd_load_idx(env->sd, env->idle);
7434 struct sg_lb_stats *sgs = &tmp_sgs;
7437 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7440 sgs = &sds->local_stat;
7442 if (env->idle != CPU_NEWLY_IDLE ||
7443 time_after_eq(jiffies, sg->sgc->next_update))
7444 update_group_capacity(env->sd, env->dst_cpu);
7447 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7448 &overload, &overutilized);
7454 * In case the child domain prefers tasks go to siblings
7455 * first, lower the sg capacity so that we'll try
7456 * and move all the excess tasks away. We lower the capacity
7457 * of a group only if the local group has the capacity to fit
7458 * these excess tasks. The extra check prevents the case where
7459 * you always pull from the heaviest group when it is already
7460 * under-utilized (possible with a large weight task outweighs
7461 * the tasks on the system).
7463 if (prefer_sibling && sds->local &&
7464 group_has_capacity(env, &sds->local_stat) &&
7465 (sgs->sum_nr_running > 1)) {
7466 sgs->group_no_capacity = 1;
7467 sgs->group_type = group_classify(sg, sgs);
7471 * Ignore task groups with misfit tasks if local group has no
7472 * capacity or if per-cpu capacity isn't higher.
7474 if (sgs->group_type == group_misfit_task &&
7475 (!group_has_capacity(env, &sds->local_stat) ||
7476 !group_smaller_cpu_capacity(sg, sds->local)))
7477 sgs->group_type = group_other;
7479 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7481 sds->busiest_stat = *sgs;
7485 /* Now, start updating sd_lb_stats */
7486 sds->total_load += sgs->group_load;
7487 sds->total_capacity += sgs->group_capacity;
7490 } while (sg != env->sd->groups);
7492 if (env->sd->flags & SD_NUMA)
7493 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7495 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7497 if (!env->sd->parent) {
7498 /* update overload indicator if we are at root domain */
7499 if (env->dst_rq->rd->overload != overload)
7500 env->dst_rq->rd->overload = overload;
7502 /* Update over-utilization (tipping point, U >= 0) indicator */
7503 if (env->dst_rq->rd->overutilized != overutilized)
7504 env->dst_rq->rd->overutilized = overutilized;
7506 if (!env->dst_rq->rd->overutilized && overutilized)
7507 env->dst_rq->rd->overutilized = true;
7512 * check_asym_packing - Check to see if the group is packed into the
7515 * This is primarily intended to used at the sibling level. Some
7516 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7517 * case of POWER7, it can move to lower SMT modes only when higher
7518 * threads are idle. When in lower SMT modes, the threads will
7519 * perform better since they share less core resources. Hence when we
7520 * have idle threads, we want them to be the higher ones.
7522 * This packing function is run on idle threads. It checks to see if
7523 * the busiest CPU in this domain (core in the P7 case) has a higher
7524 * CPU number than the packing function is being run on. Here we are
7525 * assuming lower CPU number will be equivalent to lower a SMT thread
7528 * Return: 1 when packing is required and a task should be moved to
7529 * this CPU. The amount of the imbalance is returned in *imbalance.
7531 * @env: The load balancing environment.
7532 * @sds: Statistics of the sched_domain which is to be packed
7534 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7538 if (!(env->sd->flags & SD_ASYM_PACKING))
7544 busiest_cpu = group_first_cpu(sds->busiest);
7545 if (env->dst_cpu > busiest_cpu)
7548 env->imbalance = DIV_ROUND_CLOSEST(
7549 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7550 SCHED_CAPACITY_SCALE);
7556 * fix_small_imbalance - Calculate the minor imbalance that exists
7557 * amongst the groups of a sched_domain, during
7559 * @env: The load balancing environment.
7560 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7563 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7565 unsigned long tmp, capa_now = 0, capa_move = 0;
7566 unsigned int imbn = 2;
7567 unsigned long scaled_busy_load_per_task;
7568 struct sg_lb_stats *local, *busiest;
7570 local = &sds->local_stat;
7571 busiest = &sds->busiest_stat;
7573 if (!local->sum_nr_running)
7574 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7575 else if (busiest->load_per_task > local->load_per_task)
7578 scaled_busy_load_per_task =
7579 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7580 busiest->group_capacity;
7582 if (busiest->avg_load + scaled_busy_load_per_task >=
7583 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7584 env->imbalance = busiest->load_per_task;
7589 * OK, we don't have enough imbalance to justify moving tasks,
7590 * however we may be able to increase total CPU capacity used by
7594 capa_now += busiest->group_capacity *
7595 min(busiest->load_per_task, busiest->avg_load);
7596 capa_now += local->group_capacity *
7597 min(local->load_per_task, local->avg_load);
7598 capa_now /= SCHED_CAPACITY_SCALE;
7600 /* Amount of load we'd subtract */
7601 if (busiest->avg_load > scaled_busy_load_per_task) {
7602 capa_move += busiest->group_capacity *
7603 min(busiest->load_per_task,
7604 busiest->avg_load - scaled_busy_load_per_task);
7607 /* Amount of load we'd add */
7608 if (busiest->avg_load * busiest->group_capacity <
7609 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7610 tmp = (busiest->avg_load * busiest->group_capacity) /
7611 local->group_capacity;
7613 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7614 local->group_capacity;
7616 capa_move += local->group_capacity *
7617 min(local->load_per_task, local->avg_load + tmp);
7618 capa_move /= SCHED_CAPACITY_SCALE;
7620 /* Move if we gain throughput */
7621 if (capa_move > capa_now)
7622 env->imbalance = busiest->load_per_task;
7626 * calculate_imbalance - Calculate the amount of imbalance present within the
7627 * groups of a given sched_domain during load balance.
7628 * @env: load balance environment
7629 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7631 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7633 unsigned long max_pull, load_above_capacity = ~0UL;
7634 struct sg_lb_stats *local, *busiest;
7636 local = &sds->local_stat;
7637 busiest = &sds->busiest_stat;
7639 if (busiest->group_type == group_imbalanced) {
7641 * In the group_imb case we cannot rely on group-wide averages
7642 * to ensure cpu-load equilibrium, look at wider averages. XXX
7644 busiest->load_per_task =
7645 min(busiest->load_per_task, sds->avg_load);
7649 * In the presence of smp nice balancing, certain scenarios can have
7650 * max load less than avg load(as we skip the groups at or below
7651 * its cpu_capacity, while calculating max_load..)
7653 if (busiest->avg_load <= sds->avg_load ||
7654 local->avg_load >= sds->avg_load) {
7655 /* Misfitting tasks should be migrated in any case */
7656 if (busiest->group_type == group_misfit_task) {
7657 env->imbalance = busiest->group_misfit_task;
7662 * Busiest group is overloaded, local is not, use the spare
7663 * cycles to maximize throughput
7665 if (busiest->group_type == group_overloaded &&
7666 local->group_type <= group_misfit_task) {
7667 env->imbalance = busiest->load_per_task;
7672 return fix_small_imbalance(env, sds);
7676 * If there aren't any idle cpus, avoid creating some.
7678 if (busiest->group_type == group_overloaded &&
7679 local->group_type == group_overloaded) {
7680 load_above_capacity = busiest->sum_nr_running *
7682 if (load_above_capacity > busiest->group_capacity)
7683 load_above_capacity -= busiest->group_capacity;
7685 load_above_capacity = ~0UL;
7689 * We're trying to get all the cpus to the average_load, so we don't
7690 * want to push ourselves above the average load, nor do we wish to
7691 * reduce the max loaded cpu below the average load. At the same time,
7692 * we also don't want to reduce the group load below the group capacity
7693 * (so that we can implement power-savings policies etc). Thus we look
7694 * for the minimum possible imbalance.
7696 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7698 /* How much load to actually move to equalise the imbalance */
7699 env->imbalance = min(
7700 max_pull * busiest->group_capacity,
7701 (sds->avg_load - local->avg_load) * local->group_capacity
7702 ) / SCHED_CAPACITY_SCALE;
7704 /* Boost imbalance to allow misfit task to be balanced. */
7705 if (busiest->group_type == group_misfit_task)
7706 env->imbalance = max_t(long, env->imbalance,
7707 busiest->group_misfit_task);
7710 * if *imbalance is less than the average load per runnable task
7711 * there is no guarantee that any tasks will be moved so we'll have
7712 * a think about bumping its value to force at least one task to be
7715 if (env->imbalance < busiest->load_per_task)
7716 return fix_small_imbalance(env, sds);
7719 /******* find_busiest_group() helpers end here *********************/
7722 * find_busiest_group - Returns the busiest group within the sched_domain
7723 * if there is an imbalance. If there isn't an imbalance, and
7724 * the user has opted for power-savings, it returns a group whose
7725 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7726 * such a group exists.
7728 * Also calculates the amount of weighted load which should be moved
7729 * to restore balance.
7731 * @env: The load balancing environment.
7733 * Return: - The busiest group if imbalance exists.
7734 * - If no imbalance and user has opted for power-savings balance,
7735 * return the least loaded group whose CPUs can be
7736 * put to idle by rebalancing its tasks onto our group.
7738 static struct sched_group *find_busiest_group(struct lb_env *env)
7740 struct sg_lb_stats *local, *busiest;
7741 struct sd_lb_stats sds;
7743 init_sd_lb_stats(&sds);
7746 * Compute the various statistics relavent for load balancing at
7749 update_sd_lb_stats(env, &sds);
7751 if (energy_aware() && !env->dst_rq->rd->overutilized)
7754 local = &sds.local_stat;
7755 busiest = &sds.busiest_stat;
7757 /* ASYM feature bypasses nice load balance check */
7758 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7759 check_asym_packing(env, &sds))
7762 /* There is no busy sibling group to pull tasks from */
7763 if (!sds.busiest || busiest->sum_nr_running == 0)
7766 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7767 / sds.total_capacity;
7770 * If the busiest group is imbalanced the below checks don't
7771 * work because they assume all things are equal, which typically
7772 * isn't true due to cpus_allowed constraints and the like.
7774 if (busiest->group_type == group_imbalanced)
7777 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7778 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7779 busiest->group_no_capacity)
7782 /* Misfitting tasks should be dealt with regardless of the avg load */
7783 if (busiest->group_type == group_misfit_task) {
7788 * If the local group is busier than the selected busiest group
7789 * don't try and pull any tasks.
7791 if (local->avg_load >= busiest->avg_load)
7795 * Don't pull any tasks if this group is already above the domain
7798 if (local->avg_load >= sds.avg_load)
7801 if (env->idle == CPU_IDLE) {
7803 * This cpu is idle. If the busiest group is not overloaded
7804 * and there is no imbalance between this and busiest group
7805 * wrt idle cpus, it is balanced. The imbalance becomes
7806 * significant if the diff is greater than 1 otherwise we
7807 * might end up to just move the imbalance on another group
7809 if ((busiest->group_type != group_overloaded) &&
7810 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7811 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7815 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7816 * imbalance_pct to be conservative.
7818 if (100 * busiest->avg_load <=
7819 env->sd->imbalance_pct * local->avg_load)
7824 env->busiest_group_type = busiest->group_type;
7825 /* Looks like there is an imbalance. Compute it */
7826 calculate_imbalance(env, &sds);
7835 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7837 static struct rq *find_busiest_queue(struct lb_env *env,
7838 struct sched_group *group)
7840 struct rq *busiest = NULL, *rq;
7841 unsigned long busiest_load = 0, busiest_capacity = 1;
7844 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7845 unsigned long capacity, wl;
7849 rt = fbq_classify_rq(rq);
7852 * We classify groups/runqueues into three groups:
7853 * - regular: there are !numa tasks
7854 * - remote: there are numa tasks that run on the 'wrong' node
7855 * - all: there is no distinction
7857 * In order to avoid migrating ideally placed numa tasks,
7858 * ignore those when there's better options.
7860 * If we ignore the actual busiest queue to migrate another
7861 * task, the next balance pass can still reduce the busiest
7862 * queue by moving tasks around inside the node.
7864 * If we cannot move enough load due to this classification
7865 * the next pass will adjust the group classification and
7866 * allow migration of more tasks.
7868 * Both cases only affect the total convergence complexity.
7870 if (rt > env->fbq_type)
7873 capacity = capacity_of(i);
7875 wl = weighted_cpuload(i);
7878 * When comparing with imbalance, use weighted_cpuload()
7879 * which is not scaled with the cpu capacity.
7882 if (rq->nr_running == 1 && wl > env->imbalance &&
7883 !check_cpu_capacity(rq, env->sd) &&
7884 env->busiest_group_type != group_misfit_task)
7888 * For the load comparisons with the other cpu's, consider
7889 * the weighted_cpuload() scaled with the cpu capacity, so
7890 * that the load can be moved away from the cpu that is
7891 * potentially running at a lower capacity.
7893 * Thus we're looking for max(wl_i / capacity_i), crosswise
7894 * multiplication to rid ourselves of the division works out
7895 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7896 * our previous maximum.
7898 if (wl * busiest_capacity > busiest_load * capacity) {
7900 busiest_capacity = capacity;
7909 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7910 * so long as it is large enough.
7912 #define MAX_PINNED_INTERVAL 512
7914 /* Working cpumask for load_balance and load_balance_newidle. */
7915 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7917 static int need_active_balance(struct lb_env *env)
7919 struct sched_domain *sd = env->sd;
7921 if (env->idle == CPU_NEWLY_IDLE) {
7924 * ASYM_PACKING needs to force migrate tasks from busy but
7925 * higher numbered CPUs in order to pack all tasks in the
7926 * lowest numbered CPUs.
7928 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7933 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7934 * It's worth migrating the task if the src_cpu's capacity is reduced
7935 * because of other sched_class or IRQs if more capacity stays
7936 * available on dst_cpu.
7938 if ((env->idle != CPU_NOT_IDLE) &&
7939 (env->src_rq->cfs.h_nr_running == 1)) {
7940 if ((check_cpu_capacity(env->src_rq, sd)) &&
7941 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7945 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7946 env->src_rq->cfs.h_nr_running == 1 &&
7947 cpu_overutilized(env->src_cpu) &&
7948 !cpu_overutilized(env->dst_cpu)) {
7952 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7955 static int active_load_balance_cpu_stop(void *data);
7957 static int should_we_balance(struct lb_env *env)
7959 struct sched_group *sg = env->sd->groups;
7960 struct cpumask *sg_cpus, *sg_mask;
7961 int cpu, balance_cpu = -1;
7964 * In the newly idle case, we will allow all the cpu's
7965 * to do the newly idle load balance.
7967 if (env->idle == CPU_NEWLY_IDLE)
7970 sg_cpus = sched_group_cpus(sg);
7971 sg_mask = sched_group_mask(sg);
7972 /* Try to find first idle cpu */
7973 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7974 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7981 if (balance_cpu == -1)
7982 balance_cpu = group_balance_cpu(sg);
7985 * First idle cpu or the first cpu(busiest) in this sched group
7986 * is eligible for doing load balancing at this and above domains.
7988 return balance_cpu == env->dst_cpu;
7992 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7993 * tasks if there is an imbalance.
7995 static int load_balance(int this_cpu, struct rq *this_rq,
7996 struct sched_domain *sd, enum cpu_idle_type idle,
7997 int *continue_balancing)
7999 int ld_moved, cur_ld_moved, active_balance = 0;
8000 struct sched_domain *sd_parent = sd->parent;
8001 struct sched_group *group;
8003 unsigned long flags;
8004 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8006 struct lb_env env = {
8008 .dst_cpu = this_cpu,
8010 .dst_grpmask = sched_group_cpus(sd->groups),
8012 .loop_break = sched_nr_migrate_break,
8015 .tasks = LIST_HEAD_INIT(env.tasks),
8019 * For NEWLY_IDLE load_balancing, we don't need to consider
8020 * other cpus in our group
8022 if (idle == CPU_NEWLY_IDLE)
8023 env.dst_grpmask = NULL;
8025 cpumask_copy(cpus, cpu_active_mask);
8027 schedstat_inc(sd, lb_count[idle]);
8030 if (!should_we_balance(&env)) {
8031 *continue_balancing = 0;
8035 group = find_busiest_group(&env);
8037 schedstat_inc(sd, lb_nobusyg[idle]);
8041 busiest = find_busiest_queue(&env, group);
8043 schedstat_inc(sd, lb_nobusyq[idle]);
8047 BUG_ON(busiest == env.dst_rq);
8049 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8051 env.src_cpu = busiest->cpu;
8052 env.src_rq = busiest;
8055 if (busiest->nr_running > 1) {
8057 * Attempt to move tasks. If find_busiest_group has found
8058 * an imbalance but busiest->nr_running <= 1, the group is
8059 * still unbalanced. ld_moved simply stays zero, so it is
8060 * correctly treated as an imbalance.
8062 env.flags |= LBF_ALL_PINNED;
8063 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8066 raw_spin_lock_irqsave(&busiest->lock, flags);
8069 * cur_ld_moved - load moved in current iteration
8070 * ld_moved - cumulative load moved across iterations
8072 cur_ld_moved = detach_tasks(&env);
8074 * We want to potentially lower env.src_cpu's OPP.
8077 update_capacity_of(env.src_cpu);
8080 * We've detached some tasks from busiest_rq. Every
8081 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8082 * unlock busiest->lock, and we are able to be sure
8083 * that nobody can manipulate the tasks in parallel.
8084 * See task_rq_lock() family for the details.
8087 raw_spin_unlock(&busiest->lock);
8091 ld_moved += cur_ld_moved;
8094 local_irq_restore(flags);
8096 if (env.flags & LBF_NEED_BREAK) {
8097 env.flags &= ~LBF_NEED_BREAK;
8102 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8103 * us and move them to an alternate dst_cpu in our sched_group
8104 * where they can run. The upper limit on how many times we
8105 * iterate on same src_cpu is dependent on number of cpus in our
8108 * This changes load balance semantics a bit on who can move
8109 * load to a given_cpu. In addition to the given_cpu itself
8110 * (or a ilb_cpu acting on its behalf where given_cpu is
8111 * nohz-idle), we now have balance_cpu in a position to move
8112 * load to given_cpu. In rare situations, this may cause
8113 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8114 * _independently_ and at _same_ time to move some load to
8115 * given_cpu) causing exceess load to be moved to given_cpu.
8116 * This however should not happen so much in practice and
8117 * moreover subsequent load balance cycles should correct the
8118 * excess load moved.
8120 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8122 /* Prevent to re-select dst_cpu via env's cpus */
8123 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8125 env.dst_rq = cpu_rq(env.new_dst_cpu);
8126 env.dst_cpu = env.new_dst_cpu;
8127 env.flags &= ~LBF_DST_PINNED;
8129 env.loop_break = sched_nr_migrate_break;
8132 * Go back to "more_balance" rather than "redo" since we
8133 * need to continue with same src_cpu.
8139 * We failed to reach balance because of affinity.
8142 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8144 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8145 *group_imbalance = 1;
8148 /* All tasks on this runqueue were pinned by CPU affinity */
8149 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8150 cpumask_clear_cpu(cpu_of(busiest), cpus);
8151 if (!cpumask_empty(cpus)) {
8153 env.loop_break = sched_nr_migrate_break;
8156 goto out_all_pinned;
8161 schedstat_inc(sd, lb_failed[idle]);
8163 * Increment the failure counter only on periodic balance.
8164 * We do not want newidle balance, which can be very
8165 * frequent, pollute the failure counter causing
8166 * excessive cache_hot migrations and active balances.
8168 if (idle != CPU_NEWLY_IDLE)
8169 if (env.src_grp_nr_running > 1)
8170 sd->nr_balance_failed++;
8172 if (need_active_balance(&env)) {
8173 raw_spin_lock_irqsave(&busiest->lock, flags);
8175 /* don't kick the active_load_balance_cpu_stop,
8176 * if the curr task on busiest cpu can't be
8179 if (!cpumask_test_cpu(this_cpu,
8180 tsk_cpus_allowed(busiest->curr))) {
8181 raw_spin_unlock_irqrestore(&busiest->lock,
8183 env.flags |= LBF_ALL_PINNED;
8184 goto out_one_pinned;
8188 * ->active_balance synchronizes accesses to
8189 * ->active_balance_work. Once set, it's cleared
8190 * only after active load balance is finished.
8192 if (!busiest->active_balance) {
8193 busiest->active_balance = 1;
8194 busiest->push_cpu = this_cpu;
8197 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8199 if (active_balance) {
8200 stop_one_cpu_nowait(cpu_of(busiest),
8201 active_load_balance_cpu_stop, busiest,
8202 &busiest->active_balance_work);
8206 * We've kicked active balancing, reset the failure
8209 sd->nr_balance_failed = sd->cache_nice_tries+1;
8212 sd->nr_balance_failed = 0;
8214 if (likely(!active_balance)) {
8215 /* We were unbalanced, so reset the balancing interval */
8216 sd->balance_interval = sd->min_interval;
8219 * If we've begun active balancing, start to back off. This
8220 * case may not be covered by the all_pinned logic if there
8221 * is only 1 task on the busy runqueue (because we don't call
8224 if (sd->balance_interval < sd->max_interval)
8225 sd->balance_interval *= 2;
8232 * We reach balance although we may have faced some affinity
8233 * constraints. Clear the imbalance flag if it was set.
8236 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8238 if (*group_imbalance)
8239 *group_imbalance = 0;
8244 * We reach balance because all tasks are pinned at this level so
8245 * we can't migrate them. Let the imbalance flag set so parent level
8246 * can try to migrate them.
8248 schedstat_inc(sd, lb_balanced[idle]);
8250 sd->nr_balance_failed = 0;
8253 /* tune up the balancing interval */
8254 if (((env.flags & LBF_ALL_PINNED) &&
8255 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8256 (sd->balance_interval < sd->max_interval))
8257 sd->balance_interval *= 2;
8264 static inline unsigned long
8265 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8267 unsigned long interval = sd->balance_interval;
8270 interval *= sd->busy_factor;
8272 /* scale ms to jiffies */
8273 interval = msecs_to_jiffies(interval);
8274 interval = clamp(interval, 1UL, max_load_balance_interval);
8280 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8282 unsigned long interval, next;
8284 interval = get_sd_balance_interval(sd, cpu_busy);
8285 next = sd->last_balance + interval;
8287 if (time_after(*next_balance, next))
8288 *next_balance = next;
8292 * idle_balance is called by schedule() if this_cpu is about to become
8293 * idle. Attempts to pull tasks from other CPUs.
8295 static int idle_balance(struct rq *this_rq)
8297 unsigned long next_balance = jiffies + HZ;
8298 int this_cpu = this_rq->cpu;
8299 struct sched_domain *sd;
8300 int pulled_task = 0;
8303 idle_enter_fair(this_rq);
8306 * We must set idle_stamp _before_ calling idle_balance(), such that we
8307 * measure the duration of idle_balance() as idle time.
8309 this_rq->idle_stamp = rq_clock(this_rq);
8311 if (!energy_aware() &&
8312 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8313 !this_rq->rd->overload)) {
8315 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8317 update_next_balance(sd, 0, &next_balance);
8323 raw_spin_unlock(&this_rq->lock);
8325 update_blocked_averages(this_cpu);
8327 for_each_domain(this_cpu, sd) {
8328 int continue_balancing = 1;
8329 u64 t0, domain_cost;
8331 if (!(sd->flags & SD_LOAD_BALANCE))
8334 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8335 update_next_balance(sd, 0, &next_balance);
8339 if (sd->flags & SD_BALANCE_NEWIDLE) {
8340 t0 = sched_clock_cpu(this_cpu);
8342 pulled_task = load_balance(this_cpu, this_rq,
8344 &continue_balancing);
8346 domain_cost = sched_clock_cpu(this_cpu) - t0;
8347 if (domain_cost > sd->max_newidle_lb_cost)
8348 sd->max_newidle_lb_cost = domain_cost;
8350 curr_cost += domain_cost;
8353 update_next_balance(sd, 0, &next_balance);
8356 * Stop searching for tasks to pull if there are
8357 * now runnable tasks on this rq.
8359 if (pulled_task || this_rq->nr_running > 0)
8364 raw_spin_lock(&this_rq->lock);
8366 if (curr_cost > this_rq->max_idle_balance_cost)
8367 this_rq->max_idle_balance_cost = curr_cost;
8370 * While browsing the domains, we released the rq lock, a task could
8371 * have been enqueued in the meantime. Since we're not going idle,
8372 * pretend we pulled a task.
8374 if (this_rq->cfs.h_nr_running && !pulled_task)
8378 /* Move the next balance forward */
8379 if (time_after(this_rq->next_balance, next_balance))
8380 this_rq->next_balance = next_balance;
8382 /* Is there a task of a high priority class? */
8383 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8387 idle_exit_fair(this_rq);
8388 this_rq->idle_stamp = 0;
8395 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8396 * running tasks off the busiest CPU onto idle CPUs. It requires at
8397 * least 1 task to be running on each physical CPU where possible, and
8398 * avoids physical / logical imbalances.
8400 static int active_load_balance_cpu_stop(void *data)
8402 struct rq *busiest_rq = data;
8403 int busiest_cpu = cpu_of(busiest_rq);
8404 int target_cpu = busiest_rq->push_cpu;
8405 struct rq *target_rq = cpu_rq(target_cpu);
8406 struct sched_domain *sd;
8407 struct task_struct *p = NULL;
8409 raw_spin_lock_irq(&busiest_rq->lock);
8411 /* make sure the requested cpu hasn't gone down in the meantime */
8412 if (unlikely(busiest_cpu != smp_processor_id() ||
8413 !busiest_rq->active_balance))
8416 /* Is there any task to move? */
8417 if (busiest_rq->nr_running <= 1)
8421 * This condition is "impossible", if it occurs
8422 * we need to fix it. Originally reported by
8423 * Bjorn Helgaas on a 128-cpu setup.
8425 BUG_ON(busiest_rq == target_rq);
8427 /* Search for an sd spanning us and the target CPU. */
8429 for_each_domain(target_cpu, sd) {
8430 if ((sd->flags & SD_LOAD_BALANCE) &&
8431 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8436 struct lb_env env = {
8438 .dst_cpu = target_cpu,
8439 .dst_rq = target_rq,
8440 .src_cpu = busiest_rq->cpu,
8441 .src_rq = busiest_rq,
8445 schedstat_inc(sd, alb_count);
8447 p = detach_one_task(&env);
8449 schedstat_inc(sd, alb_pushed);
8451 * We want to potentially lower env.src_cpu's OPP.
8453 update_capacity_of(env.src_cpu);
8456 schedstat_inc(sd, alb_failed);
8460 busiest_rq->active_balance = 0;
8461 raw_spin_unlock(&busiest_rq->lock);
8464 attach_one_task(target_rq, p);
8471 static inline int on_null_domain(struct rq *rq)
8473 return unlikely(!rcu_dereference_sched(rq->sd));
8476 #ifdef CONFIG_NO_HZ_COMMON
8478 * idle load balancing details
8479 * - When one of the busy CPUs notice that there may be an idle rebalancing
8480 * needed, they will kick the idle load balancer, which then does idle
8481 * load balancing for all the idle CPUs.
8484 cpumask_var_t idle_cpus_mask;
8486 unsigned long next_balance; /* in jiffy units */
8487 } nohz ____cacheline_aligned;
8489 static inline int find_new_ilb(void)
8491 int ilb = cpumask_first(nohz.idle_cpus_mask);
8493 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8500 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8501 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8502 * CPU (if there is one).
8504 static void nohz_balancer_kick(void)
8508 nohz.next_balance++;
8510 ilb_cpu = find_new_ilb();
8512 if (ilb_cpu >= nr_cpu_ids)
8515 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8518 * Use smp_send_reschedule() instead of resched_cpu().
8519 * This way we generate a sched IPI on the target cpu which
8520 * is idle. And the softirq performing nohz idle load balance
8521 * will be run before returning from the IPI.
8523 smp_send_reschedule(ilb_cpu);
8527 static inline void nohz_balance_exit_idle(int cpu)
8529 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8531 * Completely isolated CPUs don't ever set, so we must test.
8533 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8534 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8535 atomic_dec(&nohz.nr_cpus);
8537 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8541 static inline void set_cpu_sd_state_busy(void)
8543 struct sched_domain *sd;
8544 int cpu = smp_processor_id();
8547 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8549 if (!sd || !sd->nohz_idle)
8553 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8558 void set_cpu_sd_state_idle(void)
8560 struct sched_domain *sd;
8561 int cpu = smp_processor_id();
8564 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8566 if (!sd || sd->nohz_idle)
8570 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8576 * This routine will record that the cpu is going idle with tick stopped.
8577 * This info will be used in performing idle load balancing in the future.
8579 void nohz_balance_enter_idle(int cpu)
8582 * If this cpu is going down, then nothing needs to be done.
8584 if (!cpu_active(cpu))
8587 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8591 * If we're a completely isolated CPU, we don't play.
8593 if (on_null_domain(cpu_rq(cpu)))
8596 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8597 atomic_inc(&nohz.nr_cpus);
8598 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8601 static int sched_ilb_notifier(struct notifier_block *nfb,
8602 unsigned long action, void *hcpu)
8604 switch (action & ~CPU_TASKS_FROZEN) {
8606 nohz_balance_exit_idle(smp_processor_id());
8614 static DEFINE_SPINLOCK(balancing);
8617 * Scale the max load_balance interval with the number of CPUs in the system.
8618 * This trades load-balance latency on larger machines for less cross talk.
8620 void update_max_interval(void)
8622 max_load_balance_interval = HZ*num_online_cpus()/10;
8626 * It checks each scheduling domain to see if it is due to be balanced,
8627 * and initiates a balancing operation if so.
8629 * Balancing parameters are set up in init_sched_domains.
8631 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8633 int continue_balancing = 1;
8635 unsigned long interval;
8636 struct sched_domain *sd;
8637 /* Earliest time when we have to do rebalance again */
8638 unsigned long next_balance = jiffies + 60*HZ;
8639 int update_next_balance = 0;
8640 int need_serialize, need_decay = 0;
8643 update_blocked_averages(cpu);
8646 for_each_domain(cpu, sd) {
8648 * Decay the newidle max times here because this is a regular
8649 * visit to all the domains. Decay ~1% per second.
8651 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8652 sd->max_newidle_lb_cost =
8653 (sd->max_newidle_lb_cost * 253) / 256;
8654 sd->next_decay_max_lb_cost = jiffies + HZ;
8657 max_cost += sd->max_newidle_lb_cost;
8659 if (!(sd->flags & SD_LOAD_BALANCE))
8663 * Stop the load balance at this level. There is another
8664 * CPU in our sched group which is doing load balancing more
8667 if (!continue_balancing) {
8673 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8675 need_serialize = sd->flags & SD_SERIALIZE;
8676 if (need_serialize) {
8677 if (!spin_trylock(&balancing))
8681 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8682 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8684 * The LBF_DST_PINNED logic could have changed
8685 * env->dst_cpu, so we can't know our idle
8686 * state even if we migrated tasks. Update it.
8688 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8690 sd->last_balance = jiffies;
8691 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8694 spin_unlock(&balancing);
8696 if (time_after(next_balance, sd->last_balance + interval)) {
8697 next_balance = sd->last_balance + interval;
8698 update_next_balance = 1;
8703 * Ensure the rq-wide value also decays but keep it at a
8704 * reasonable floor to avoid funnies with rq->avg_idle.
8706 rq->max_idle_balance_cost =
8707 max((u64)sysctl_sched_migration_cost, max_cost);
8712 * next_balance will be updated only when there is a need.
8713 * When the cpu is attached to null domain for ex, it will not be
8716 if (likely(update_next_balance)) {
8717 rq->next_balance = next_balance;
8719 #ifdef CONFIG_NO_HZ_COMMON
8721 * If this CPU has been elected to perform the nohz idle
8722 * balance. Other idle CPUs have already rebalanced with
8723 * nohz_idle_balance() and nohz.next_balance has been
8724 * updated accordingly. This CPU is now running the idle load
8725 * balance for itself and we need to update the
8726 * nohz.next_balance accordingly.
8728 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8729 nohz.next_balance = rq->next_balance;
8734 #ifdef CONFIG_NO_HZ_COMMON
8736 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8737 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8739 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8741 int this_cpu = this_rq->cpu;
8744 /* Earliest time when we have to do rebalance again */
8745 unsigned long next_balance = jiffies + 60*HZ;
8746 int update_next_balance = 0;
8748 if (idle != CPU_IDLE ||
8749 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8752 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8753 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8757 * If this cpu gets work to do, stop the load balancing
8758 * work being done for other cpus. Next load
8759 * balancing owner will pick it up.
8764 rq = cpu_rq(balance_cpu);
8767 * If time for next balance is due,
8770 if (time_after_eq(jiffies, rq->next_balance)) {
8771 raw_spin_lock_irq(&rq->lock);
8772 update_rq_clock(rq);
8773 update_idle_cpu_load(rq);
8774 raw_spin_unlock_irq(&rq->lock);
8775 rebalance_domains(rq, CPU_IDLE);
8778 if (time_after(next_balance, rq->next_balance)) {
8779 next_balance = rq->next_balance;
8780 update_next_balance = 1;
8785 * next_balance will be updated only when there is a need.
8786 * When the CPU is attached to null domain for ex, it will not be
8789 if (likely(update_next_balance))
8790 nohz.next_balance = next_balance;
8792 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8796 * Current heuristic for kicking the idle load balancer in the presence
8797 * of an idle cpu in the system.
8798 * - This rq has more than one task.
8799 * - This rq has at least one CFS task and the capacity of the CPU is
8800 * significantly reduced because of RT tasks or IRQs.
8801 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8802 * multiple busy cpu.
8803 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8804 * domain span are idle.
8806 static inline bool nohz_kick_needed(struct rq *rq)
8808 unsigned long now = jiffies;
8809 struct sched_domain *sd;
8810 struct sched_group_capacity *sgc;
8811 int nr_busy, cpu = rq->cpu;
8814 if (unlikely(rq->idle_balance))
8818 * We may be recently in ticked or tickless idle mode. At the first
8819 * busy tick after returning from idle, we will update the busy stats.
8821 set_cpu_sd_state_busy();
8822 nohz_balance_exit_idle(cpu);
8825 * None are in tickless mode and hence no need for NOHZ idle load
8828 if (likely(!atomic_read(&nohz.nr_cpus)))
8831 if (time_before(now, nohz.next_balance))
8834 if (rq->nr_running >= 2 &&
8835 (!energy_aware() || cpu_overutilized(cpu)))
8839 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8840 if (sd && !energy_aware()) {
8841 sgc = sd->groups->sgc;
8842 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8851 sd = rcu_dereference(rq->sd);
8853 if ((rq->cfs.h_nr_running >= 1) &&
8854 check_cpu_capacity(rq, sd)) {
8860 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8861 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8862 sched_domain_span(sd)) < cpu)) {
8872 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8876 * run_rebalance_domains is triggered when needed from the scheduler tick.
8877 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8879 static void run_rebalance_domains(struct softirq_action *h)
8881 struct rq *this_rq = this_rq();
8882 enum cpu_idle_type idle = this_rq->idle_balance ?
8883 CPU_IDLE : CPU_NOT_IDLE;
8886 * If this cpu has a pending nohz_balance_kick, then do the
8887 * balancing on behalf of the other idle cpus whose ticks are
8888 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8889 * give the idle cpus a chance to load balance. Else we may
8890 * load balance only within the local sched_domain hierarchy
8891 * and abort nohz_idle_balance altogether if we pull some load.
8893 nohz_idle_balance(this_rq, idle);
8894 rebalance_domains(this_rq, idle);
8898 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8900 void trigger_load_balance(struct rq *rq)
8902 /* Don't need to rebalance while attached to NULL domain */
8903 if (unlikely(on_null_domain(rq)))
8906 if (time_after_eq(jiffies, rq->next_balance))
8907 raise_softirq(SCHED_SOFTIRQ);
8908 #ifdef CONFIG_NO_HZ_COMMON
8909 if (nohz_kick_needed(rq))
8910 nohz_balancer_kick();
8914 static void rq_online_fair(struct rq *rq)
8918 update_runtime_enabled(rq);
8921 static void rq_offline_fair(struct rq *rq)
8925 /* Ensure any throttled groups are reachable by pick_next_task */
8926 unthrottle_offline_cfs_rqs(rq);
8929 #endif /* CONFIG_SMP */
8932 * scheduler tick hitting a task of our scheduling class:
8934 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8936 struct cfs_rq *cfs_rq;
8937 struct sched_entity *se = &curr->se;
8939 for_each_sched_entity(se) {
8940 cfs_rq = cfs_rq_of(se);
8941 entity_tick(cfs_rq, se, queued);
8944 if (static_branch_unlikely(&sched_numa_balancing))
8945 task_tick_numa(rq, curr);
8948 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8949 rq->rd->overutilized = true;
8951 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8957 * called on fork with the child task as argument from the parent's context
8958 * - child not yet on the tasklist
8959 * - preemption disabled
8961 static void task_fork_fair(struct task_struct *p)
8963 struct cfs_rq *cfs_rq;
8964 struct sched_entity *se = &p->se, *curr;
8965 int this_cpu = smp_processor_id();
8966 struct rq *rq = this_rq();
8967 unsigned long flags;
8969 raw_spin_lock_irqsave(&rq->lock, flags);
8971 update_rq_clock(rq);
8973 cfs_rq = task_cfs_rq(current);
8974 curr = cfs_rq->curr;
8977 * Not only the cpu but also the task_group of the parent might have
8978 * been changed after parent->se.parent,cfs_rq were copied to
8979 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8980 * of child point to valid ones.
8983 __set_task_cpu(p, this_cpu);
8986 update_curr(cfs_rq);
8989 se->vruntime = curr->vruntime;
8990 place_entity(cfs_rq, se, 1);
8992 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8994 * Upon rescheduling, sched_class::put_prev_task() will place
8995 * 'current' within the tree based on its new key value.
8997 swap(curr->vruntime, se->vruntime);
9001 se->vruntime -= cfs_rq->min_vruntime;
9003 raw_spin_unlock_irqrestore(&rq->lock, flags);
9007 * Priority of the task has changed. Check to see if we preempt
9011 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9013 if (!task_on_rq_queued(p))
9017 * Reschedule if we are currently running on this runqueue and
9018 * our priority decreased, or if we are not currently running on
9019 * this runqueue and our priority is higher than the current's
9021 if (rq->curr == p) {
9022 if (p->prio > oldprio)
9025 check_preempt_curr(rq, p, 0);
9028 static inline bool vruntime_normalized(struct task_struct *p)
9030 struct sched_entity *se = &p->se;
9033 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9034 * the dequeue_entity(.flags=0) will already have normalized the
9041 * When !on_rq, vruntime of the task has usually NOT been normalized.
9042 * But there are some cases where it has already been normalized:
9044 * - A forked child which is waiting for being woken up by
9045 * wake_up_new_task().
9046 * - A task which has been woken up by try_to_wake_up() and
9047 * waiting for actually being woken up by sched_ttwu_pending().
9049 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9055 static void detach_task_cfs_rq(struct task_struct *p)
9057 struct sched_entity *se = &p->se;
9058 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9060 if (!vruntime_normalized(p)) {
9062 * Fix up our vruntime so that the current sleep doesn't
9063 * cause 'unlimited' sleep bonus.
9065 place_entity(cfs_rq, se, 0);
9066 se->vruntime -= cfs_rq->min_vruntime;
9069 /* Catch up with the cfs_rq and remove our load when we leave */
9070 detach_entity_load_avg(cfs_rq, se);
9073 static void attach_task_cfs_rq(struct task_struct *p)
9075 struct sched_entity *se = &p->se;
9076 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9078 #ifdef CONFIG_FAIR_GROUP_SCHED
9080 * Since the real-depth could have been changed (only FAIR
9081 * class maintain depth value), reset depth properly.
9083 se->depth = se->parent ? se->parent->depth + 1 : 0;
9086 /* Synchronize task with its cfs_rq */
9087 attach_entity_load_avg(cfs_rq, se);
9089 if (!vruntime_normalized(p))
9090 se->vruntime += cfs_rq->min_vruntime;
9093 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9095 detach_task_cfs_rq(p);
9098 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9100 attach_task_cfs_rq(p);
9102 if (task_on_rq_queued(p)) {
9104 * We were most likely switched from sched_rt, so
9105 * kick off the schedule if running, otherwise just see
9106 * if we can still preempt the current task.
9111 check_preempt_curr(rq, p, 0);
9115 /* Account for a task changing its policy or group.
9117 * This routine is mostly called to set cfs_rq->curr field when a task
9118 * migrates between groups/classes.
9120 static void set_curr_task_fair(struct rq *rq)
9122 struct sched_entity *se = &rq->curr->se;
9124 for_each_sched_entity(se) {
9125 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9127 set_next_entity(cfs_rq, se);
9128 /* ensure bandwidth has been allocated on our new cfs_rq */
9129 account_cfs_rq_runtime(cfs_rq, 0);
9133 void init_cfs_rq(struct cfs_rq *cfs_rq)
9135 cfs_rq->tasks_timeline = RB_ROOT;
9136 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9137 #ifndef CONFIG_64BIT
9138 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9141 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9142 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9146 #ifdef CONFIG_FAIR_GROUP_SCHED
9147 static void task_move_group_fair(struct task_struct *p)
9149 detach_task_cfs_rq(p);
9150 set_task_rq(p, task_cpu(p));
9153 /* Tell se's cfs_rq has been changed -- migrated */
9154 p->se.avg.last_update_time = 0;
9156 attach_task_cfs_rq(p);
9159 void free_fair_sched_group(struct task_group *tg)
9163 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9165 for_each_possible_cpu(i) {
9167 kfree(tg->cfs_rq[i]);
9170 remove_entity_load_avg(tg->se[i]);
9179 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9181 struct cfs_rq *cfs_rq;
9182 struct sched_entity *se;
9185 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9188 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9192 tg->shares = NICE_0_LOAD;
9194 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9196 for_each_possible_cpu(i) {
9197 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9198 GFP_KERNEL, cpu_to_node(i));
9202 se = kzalloc_node(sizeof(struct sched_entity),
9203 GFP_KERNEL, cpu_to_node(i));
9207 init_cfs_rq(cfs_rq);
9208 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9209 init_entity_runnable_average(se);
9220 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9222 struct rq *rq = cpu_rq(cpu);
9223 unsigned long flags;
9226 * Only empty task groups can be destroyed; so we can speculatively
9227 * check on_list without danger of it being re-added.
9229 if (!tg->cfs_rq[cpu]->on_list)
9232 raw_spin_lock_irqsave(&rq->lock, flags);
9233 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9234 raw_spin_unlock_irqrestore(&rq->lock, flags);
9237 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9238 struct sched_entity *se, int cpu,
9239 struct sched_entity *parent)
9241 struct rq *rq = cpu_rq(cpu);
9245 init_cfs_rq_runtime(cfs_rq);
9247 tg->cfs_rq[cpu] = cfs_rq;
9250 /* se could be NULL for root_task_group */
9255 se->cfs_rq = &rq->cfs;
9258 se->cfs_rq = parent->my_q;
9259 se->depth = parent->depth + 1;
9263 /* guarantee group entities always have weight */
9264 update_load_set(&se->load, NICE_0_LOAD);
9265 se->parent = parent;
9268 static DEFINE_MUTEX(shares_mutex);
9270 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9273 unsigned long flags;
9276 * We can't change the weight of the root cgroup.
9281 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9283 mutex_lock(&shares_mutex);
9284 if (tg->shares == shares)
9287 tg->shares = shares;
9288 for_each_possible_cpu(i) {
9289 struct rq *rq = cpu_rq(i);
9290 struct sched_entity *se;
9293 /* Propagate contribution to hierarchy */
9294 raw_spin_lock_irqsave(&rq->lock, flags);
9296 /* Possible calls to update_curr() need rq clock */
9297 update_rq_clock(rq);
9298 for_each_sched_entity(se)
9299 update_cfs_shares(group_cfs_rq(se));
9300 raw_spin_unlock_irqrestore(&rq->lock, flags);
9304 mutex_unlock(&shares_mutex);
9307 #else /* CONFIG_FAIR_GROUP_SCHED */
9309 void free_fair_sched_group(struct task_group *tg) { }
9311 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9316 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9318 #endif /* CONFIG_FAIR_GROUP_SCHED */
9321 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9323 struct sched_entity *se = &task->se;
9324 unsigned int rr_interval = 0;
9327 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9330 if (rq->cfs.load.weight)
9331 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9337 * All the scheduling class methods:
9339 const struct sched_class fair_sched_class = {
9340 .next = &idle_sched_class,
9341 .enqueue_task = enqueue_task_fair,
9342 .dequeue_task = dequeue_task_fair,
9343 .yield_task = yield_task_fair,
9344 .yield_to_task = yield_to_task_fair,
9346 .check_preempt_curr = check_preempt_wakeup,
9348 .pick_next_task = pick_next_task_fair,
9349 .put_prev_task = put_prev_task_fair,
9352 .select_task_rq = select_task_rq_fair,
9353 .migrate_task_rq = migrate_task_rq_fair,
9355 .rq_online = rq_online_fair,
9356 .rq_offline = rq_offline_fair,
9358 .task_waking = task_waking_fair,
9359 .task_dead = task_dead_fair,
9360 .set_cpus_allowed = set_cpus_allowed_common,
9363 .set_curr_task = set_curr_task_fair,
9364 .task_tick = task_tick_fair,
9365 .task_fork = task_fork_fair,
9367 .prio_changed = prio_changed_fair,
9368 .switched_from = switched_from_fair,
9369 .switched_to = switched_to_fair,
9371 .get_rr_interval = get_rr_interval_fair,
9373 .update_curr = update_curr_fair,
9375 #ifdef CONFIG_FAIR_GROUP_SCHED
9376 .task_move_group = task_move_group_fair,
9380 #ifdef CONFIG_SCHED_DEBUG
9381 void print_cfs_stats(struct seq_file *m, int cpu)
9383 struct cfs_rq *cfs_rq;
9386 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9387 print_cfs_rq(m, cpu, cfs_rq);
9391 #ifdef CONFIG_NUMA_BALANCING
9392 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9395 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9397 for_each_online_node(node) {
9398 if (p->numa_faults) {
9399 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9400 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9402 if (p->numa_group) {
9403 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9404 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9406 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9409 #endif /* CONFIG_NUMA_BALANCING */
9410 #endif /* CONFIG_SCHED_DEBUG */
9412 __init void init_sched_fair_class(void)
9415 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9417 #ifdef CONFIG_NO_HZ_COMMON
9418 nohz.next_balance = jiffies;
9419 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9420 cpu_notifier(sched_ilb_notifier, 0);