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;
55 * The initial- and re-scaling of tunables is configurable
56 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
59 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
60 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
61 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
63 enum sched_tunable_scaling sysctl_sched_tunable_scaling
64 = SCHED_TUNABLESCALING_LOG;
67 * Minimal preemption granularity for CPU-bound tasks:
68 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
70 unsigned int sysctl_sched_min_granularity = 750000ULL;
71 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
74 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
76 static unsigned int sched_nr_latency = 8;
79 * After fork, child runs first. If set to 0 (default) then
80 * parent will (try to) run first.
82 unsigned int sysctl_sched_child_runs_first __read_mostly;
85 * SCHED_OTHER wake-up granularity.
86 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
88 * This option delays the preemption effects of decoupled workloads
89 * and reduces their over-scheduling. Synchronous workloads will still
90 * have immediate wakeup/sleep latencies.
92 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
93 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
95 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
98 * The exponential sliding window over which load is averaged for shares
102 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
104 #ifdef CONFIG_CFS_BANDWIDTH
106 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
107 * each time a cfs_rq requests quota.
109 * Note: in the case that the slice exceeds the runtime remaining (either due
110 * to consumption or the quota being specified to be smaller than the slice)
111 * we will always only issue the remaining available time.
113 * default: 5 msec, units: microseconds
115 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
118 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
124 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
130 static inline void update_load_set(struct load_weight *lw, unsigned long w)
137 * Increase the granularity value when there are more CPUs,
138 * because with more CPUs the 'effective latency' as visible
139 * to users decreases. But the relationship is not linear,
140 * so pick a second-best guess by going with the log2 of the
143 * This idea comes from the SD scheduler of Con Kolivas:
145 static unsigned int get_update_sysctl_factor(void)
147 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
150 switch (sysctl_sched_tunable_scaling) {
151 case SCHED_TUNABLESCALING_NONE:
154 case SCHED_TUNABLESCALING_LINEAR:
157 case SCHED_TUNABLESCALING_LOG:
159 factor = 1 + ilog2(cpus);
166 static void update_sysctl(void)
168 unsigned int factor = get_update_sysctl_factor();
170 #define SET_SYSCTL(name) \
171 (sysctl_##name = (factor) * normalized_sysctl_##name)
172 SET_SYSCTL(sched_min_granularity);
173 SET_SYSCTL(sched_latency);
174 SET_SYSCTL(sched_wakeup_granularity);
178 void sched_init_granularity(void)
183 #define WMULT_CONST (~0U)
184 #define WMULT_SHIFT 32
186 static void __update_inv_weight(struct load_weight *lw)
190 if (likely(lw->inv_weight))
193 w = scale_load_down(lw->weight);
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
200 lw->inv_weight = WMULT_CONST / w;
204 * delta_exec * weight / lw.weight
206 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
208 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
209 * we're guaranteed shift stays positive because inv_weight is guaranteed to
210 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
212 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
213 * weight/lw.weight <= 1, and therefore our shift will also be positive.
215 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
217 u64 fact = scale_load_down(weight);
218 int shift = WMULT_SHIFT;
220 __update_inv_weight(lw);
222 if (unlikely(fact >> 32)) {
229 /* hint to use a 32x32->64 mul */
230 fact = (u64)(u32)fact * lw->inv_weight;
237 return mul_u64_u32_shr(delta_exec, fact, shift);
241 const struct sched_class fair_sched_class;
243 /**************************************************************
244 * CFS operations on generic schedulable entities:
247 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* cpu runqueue to which this cfs_rq is attached */
250 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
255 /* An entity is a task if it doesn't "own" a runqueue */
256 #define entity_is_task(se) (!se->my_q)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 #ifdef CONFIG_SCHED_DEBUG
261 WARN_ON_ONCE(!entity_is_task(se));
263 return container_of(se, struct task_struct, se);
266 /* Walk up scheduling entities hierarchy */
267 #define for_each_sched_entity(se) \
268 for (; se; se = se->parent)
270 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
275 /* runqueue on which this entity is (to be) queued */
276 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
281 /* runqueue "owned" by this group */
282 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
287 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 if (!cfs_rq->on_list) {
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases.
296 if (cfs_rq->tg->parent &&
297 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
298 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
299 &rq_of(cfs_rq)->leaf_cfs_rq_list);
301 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
302 &rq_of(cfs_rq)->leaf_cfs_rq_list);
309 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
311 if (cfs_rq->on_list) {
312 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
317 /* Iterate thr' all leaf cfs_rq's on a runqueue */
318 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
319 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
321 /* Do the two (enqueued) entities belong to the same group ? */
322 static inline struct cfs_rq *
323 is_same_group(struct sched_entity *se, struct sched_entity *pse)
325 if (se->cfs_rq == pse->cfs_rq)
331 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
339 int se_depth, pse_depth;
342 * preemption test can be made between sibling entities who are in the
343 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
344 * both tasks until we find their ancestors who are siblings of common
348 /* First walk up until both entities are at same depth */
349 se_depth = (*se)->depth;
350 pse_depth = (*pse)->depth;
352 while (se_depth > pse_depth) {
354 *se = parent_entity(*se);
357 while (pse_depth > se_depth) {
359 *pse = parent_entity(*pse);
362 while (!is_same_group(*se, *pse)) {
363 *se = parent_entity(*se);
364 *pse = parent_entity(*pse);
368 #else /* !CONFIG_FAIR_GROUP_SCHED */
370 static inline struct task_struct *task_of(struct sched_entity *se)
372 return container_of(se, struct task_struct, se);
375 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
377 return container_of(cfs_rq, struct rq, cfs);
380 #define entity_is_task(se) 1
382 #define for_each_sched_entity(se) \
383 for (; se; se = NULL)
385 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
387 return &task_rq(p)->cfs;
390 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
392 struct task_struct *p = task_of(se);
393 struct rq *rq = task_rq(p);
398 /* runqueue "owned" by this group */
399 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
404 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
413 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
415 static inline struct sched_entity *parent_entity(struct sched_entity *se)
421 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 #endif /* CONFIG_FAIR_GROUP_SCHED */
427 static __always_inline
428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
430 /**************************************************************
431 * Scheduling class tree data structure manipulation methods:
434 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - max_vruntime);
438 max_vruntime = vruntime;
443 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - min_vruntime);
447 min_vruntime = vruntime;
452 static inline int entity_before(struct sched_entity *a,
453 struct sched_entity *b)
455 return (s64)(a->vruntime - b->vruntime) < 0;
458 static void update_min_vruntime(struct cfs_rq *cfs_rq)
460 u64 vruntime = cfs_rq->min_vruntime;
463 vruntime = cfs_rq->curr->vruntime;
465 if (cfs_rq->rb_leftmost) {
466 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 vruntime = se->vruntime;
473 vruntime = min_vruntime(vruntime, se->vruntime);
476 /* ensure we never gain time by being placed backwards. */
477 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 * Enqueue an entity into the rb-tree:
487 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
490 struct rb_node *parent = NULL;
491 struct sched_entity *entry;
495 * Find the right place in the rbtree:
499 entry = rb_entry(parent, struct sched_entity, run_node);
501 * We dont care about collisions. Nodes with
502 * the same key stay together.
504 if (entity_before(se, entry)) {
505 link = &parent->rb_left;
507 link = &parent->rb_right;
513 * Maintain a cache of leftmost tree entries (it is frequently
517 cfs_rq->rb_leftmost = &se->run_node;
519 rb_link_node(&se->run_node, parent, link);
520 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
523 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 if (cfs_rq->rb_leftmost == &se->run_node) {
526 struct rb_node *next_node;
528 next_node = rb_next(&se->run_node);
529 cfs_rq->rb_leftmost = next_node;
532 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
535 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
537 struct rb_node *left = cfs_rq->rb_leftmost;
542 return rb_entry(left, struct sched_entity, run_node);
545 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
547 struct rb_node *next = rb_next(&se->run_node);
552 return rb_entry(next, struct sched_entity, run_node);
555 #ifdef CONFIG_SCHED_DEBUG
556 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 return rb_entry(last, struct sched_entity, run_node);
566 /**************************************************************
567 * Scheduling class statistics methods:
570 int sched_proc_update_handler(struct ctl_table *table, int write,
571 void __user *buffer, size_t *lenp,
574 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
575 unsigned int factor = get_update_sysctl_factor();
580 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
581 sysctl_sched_min_granularity);
583 #define WRT_SYSCTL(name) \
584 (normalized_sysctl_##name = sysctl_##name / (factor))
585 WRT_SYSCTL(sched_min_granularity);
586 WRT_SYSCTL(sched_latency);
587 WRT_SYSCTL(sched_wakeup_granularity);
597 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
599 if (unlikely(se->load.weight != NICE_0_LOAD))
600 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
606 * The idea is to set a period in which each task runs once.
608 * When there are too many tasks (sched_nr_latency) we have to stretch
609 * this period because otherwise the slices get too small.
611 * p = (nr <= nl) ? l : l*nr/nl
613 static u64 __sched_period(unsigned long nr_running)
615 if (unlikely(nr_running > sched_nr_latency))
616 return nr_running * sysctl_sched_min_granularity;
618 return sysctl_sched_latency;
622 * We calculate the wall-time slice from the period by taking a part
623 * proportional to the weight.
627 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
631 for_each_sched_entity(se) {
632 struct load_weight *load;
633 struct load_weight lw;
635 cfs_rq = cfs_rq_of(se);
636 load = &cfs_rq->load;
638 if (unlikely(!se->on_rq)) {
641 update_load_add(&lw, se->load.weight);
644 slice = __calc_delta(slice, se->load.weight, load);
650 * We calculate the vruntime slice of a to-be-inserted task.
654 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 return calc_delta_fair(sched_slice(cfs_rq, se), se);
660 static int select_idle_sibling(struct task_struct *p, int cpu);
661 static unsigned long task_h_load(struct task_struct *p);
664 * We choose a half-life close to 1 scheduling period.
665 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
666 * dependent on this value.
668 #define LOAD_AVG_PERIOD 32
669 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
670 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
672 /* Give new sched_entity start runnable values to heavy its load in infant time */
673 void init_entity_runnable_average(struct sched_entity *se)
675 struct sched_avg *sa = &se->avg;
677 sa->last_update_time = 0;
679 * sched_avg's period_contrib should be strictly less then 1024, so
680 * we give it 1023 to make sure it is almost a period (1024us), and
681 * will definitely be update (after enqueue).
683 sa->period_contrib = 1023;
684 sa->load_avg = scale_load_down(se->load.weight);
685 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
687 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
688 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
692 void init_entity_runnable_average(struct sched_entity *se)
698 * Update the current task's runtime statistics.
700 static void update_curr(struct cfs_rq *cfs_rq)
702 struct sched_entity *curr = cfs_rq->curr;
703 u64 now = rq_clock_task(rq_of(cfs_rq));
709 delta_exec = now - curr->exec_start;
710 if (unlikely((s64)delta_exec <= 0))
713 curr->exec_start = now;
715 schedstat_set(curr->statistics.exec_max,
716 max(delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
721 curr->vruntime += calc_delta_fair(delta_exec, curr);
722 update_min_vruntime(cfs_rq);
724 if (entity_is_task(curr)) {
725 struct task_struct *curtask = task_of(curr);
727 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
728 cpuacct_charge(curtask, delta_exec);
729 account_group_exec_runtime(curtask, delta_exec);
732 account_cfs_rq_runtime(cfs_rq, delta_exec);
735 static void update_curr_fair(struct rq *rq)
737 update_curr(cfs_rq_of(&rq->curr->se));
741 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
743 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
747 * Task is being enqueued - update stats:
749 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
752 * Are we enqueueing a waiting task? (for current tasks
753 * a dequeue/enqueue event is a NOP)
755 if (se != cfs_rq->curr)
756 update_stats_wait_start(cfs_rq, se);
760 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
762 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
763 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
764 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
765 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
766 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
767 #ifdef CONFIG_SCHEDSTATS
768 if (entity_is_task(se)) {
769 trace_sched_stat_wait(task_of(se),
770 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
773 schedstat_set(se->statistics.wait_start, 0);
777 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 * Mark the end of the wait period if dequeueing a
783 if (se != cfs_rq->curr)
784 update_stats_wait_end(cfs_rq, se);
788 * We are picking a new current task - update its stats:
791 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
794 * We are starting a new run period:
796 se->exec_start = rq_clock_task(rq_of(cfs_rq));
799 /**************************************************
800 * Scheduling class queueing methods:
803 #ifdef CONFIG_NUMA_BALANCING
805 * Approximate time to scan a full NUMA task in ms. The task scan period is
806 * calculated based on the tasks virtual memory size and
807 * numa_balancing_scan_size.
809 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
810 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
812 /* Portion of address space to scan in MB */
813 unsigned int sysctl_numa_balancing_scan_size = 256;
815 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
816 unsigned int sysctl_numa_balancing_scan_delay = 1000;
818 static unsigned int task_nr_scan_windows(struct task_struct *p)
820 unsigned long rss = 0;
821 unsigned long nr_scan_pages;
824 * Calculations based on RSS as non-present and empty pages are skipped
825 * by the PTE scanner and NUMA hinting faults should be trapped based
828 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
829 rss = get_mm_rss(p->mm);
833 rss = round_up(rss, nr_scan_pages);
834 return rss / nr_scan_pages;
837 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
838 #define MAX_SCAN_WINDOW 2560
840 static unsigned int task_scan_min(struct task_struct *p)
842 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
843 unsigned int scan, floor;
844 unsigned int windows = 1;
846 if (scan_size < MAX_SCAN_WINDOW)
847 windows = MAX_SCAN_WINDOW / scan_size;
848 floor = 1000 / windows;
850 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
851 return max_t(unsigned int, floor, scan);
854 static unsigned int task_scan_max(struct task_struct *p)
856 unsigned int smin = task_scan_min(p);
859 /* Watch for min being lower than max due to floor calculations */
860 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
861 return max(smin, smax);
864 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
866 rq->nr_numa_running += (p->numa_preferred_nid != -1);
867 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
870 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
872 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
873 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
879 spinlock_t lock; /* nr_tasks, tasks */
884 nodemask_t active_nodes;
885 unsigned long total_faults;
887 * Faults_cpu is used to decide whether memory should move
888 * towards the CPU. As a consequence, these stats are weighted
889 * more by CPU use than by memory faults.
891 unsigned long *faults_cpu;
892 unsigned long faults[0];
895 /* Shared or private faults. */
896 #define NR_NUMA_HINT_FAULT_TYPES 2
898 /* Memory and CPU locality */
899 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
901 /* Averaged statistics, and temporary buffers. */
902 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
904 pid_t task_numa_group_id(struct task_struct *p)
906 return p->numa_group ? p->numa_group->gid : 0;
910 * The averaged statistics, shared & private, memory & cpu,
911 * occupy the first half of the array. The second half of the
912 * array is for current counters, which are averaged into the
913 * first set by task_numa_placement.
915 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
917 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
920 static inline unsigned long task_faults(struct task_struct *p, int nid)
925 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
926 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
929 static inline unsigned long group_faults(struct task_struct *p, int nid)
934 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
935 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
938 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
940 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
941 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
944 /* Handle placement on systems where not all nodes are directly connected. */
945 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
946 int maxdist, bool task)
948 unsigned long score = 0;
952 * All nodes are directly connected, and the same distance
953 * from each other. No need for fancy placement algorithms.
955 if (sched_numa_topology_type == NUMA_DIRECT)
959 * This code is called for each node, introducing N^2 complexity,
960 * which should be ok given the number of nodes rarely exceeds 8.
962 for_each_online_node(node) {
963 unsigned long faults;
964 int dist = node_distance(nid, node);
967 * The furthest away nodes in the system are not interesting
968 * for placement; nid was already counted.
970 if (dist == sched_max_numa_distance || node == nid)
974 * On systems with a backplane NUMA topology, compare groups
975 * of nodes, and move tasks towards the group with the most
976 * memory accesses. When comparing two nodes at distance
977 * "hoplimit", only nodes closer by than "hoplimit" are part
978 * of each group. Skip other nodes.
980 if (sched_numa_topology_type == NUMA_BACKPLANE &&
984 /* Add up the faults from nearby nodes. */
986 faults = task_faults(p, node);
988 faults = group_faults(p, node);
991 * On systems with a glueless mesh NUMA topology, there are
992 * no fixed "groups of nodes". Instead, nodes that are not
993 * directly connected bounce traffic through intermediate
994 * nodes; a numa_group can occupy any set of nodes.
995 * The further away a node is, the less the faults count.
996 * This seems to result in good task placement.
998 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
999 faults *= (sched_max_numa_distance - dist);
1000 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1010 * These return the fraction of accesses done by a particular task, or
1011 * task group, on a particular numa node. The group weight is given a
1012 * larger multiplier, in order to group tasks together that are almost
1013 * evenly spread out between numa nodes.
1015 static inline unsigned long task_weight(struct task_struct *p, int nid,
1018 unsigned long faults, total_faults;
1020 if (!p->numa_faults)
1023 total_faults = p->total_numa_faults;
1028 faults = task_faults(p, nid);
1029 faults += score_nearby_nodes(p, nid, dist, true);
1031 return 1000 * faults / total_faults;
1034 static inline unsigned long group_weight(struct task_struct *p, int nid,
1037 unsigned long faults, total_faults;
1042 total_faults = p->numa_group->total_faults;
1047 faults = group_faults(p, nid);
1048 faults += score_nearby_nodes(p, nid, dist, false);
1050 return 1000 * faults / total_faults;
1053 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1054 int src_nid, int dst_cpu)
1056 struct numa_group *ng = p->numa_group;
1057 int dst_nid = cpu_to_node(dst_cpu);
1058 int last_cpupid, this_cpupid;
1060 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1063 * Multi-stage node selection is used in conjunction with a periodic
1064 * migration fault to build a temporal task<->page relation. By using
1065 * a two-stage filter we remove short/unlikely relations.
1067 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1068 * a task's usage of a particular page (n_p) per total usage of this
1069 * page (n_t) (in a given time-span) to a probability.
1071 * Our periodic faults will sample this probability and getting the
1072 * same result twice in a row, given these samples are fully
1073 * independent, is then given by P(n)^2, provided our sample period
1074 * is sufficiently short compared to the usage pattern.
1076 * This quadric squishes small probabilities, making it less likely we
1077 * act on an unlikely task<->page relation.
1079 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1080 if (!cpupid_pid_unset(last_cpupid) &&
1081 cpupid_to_nid(last_cpupid) != dst_nid)
1084 /* Always allow migrate on private faults */
1085 if (cpupid_match_pid(p, last_cpupid))
1088 /* A shared fault, but p->numa_group has not been set up yet. */
1093 * Do not migrate if the destination is not a node that
1094 * is actively used by this numa group.
1096 if (!node_isset(dst_nid, ng->active_nodes))
1100 * Source is a node that is not actively used by this
1101 * numa group, while the destination is. Migrate.
1103 if (!node_isset(src_nid, ng->active_nodes))
1107 * Both source and destination are nodes in active
1108 * use by this numa group. Maximize memory bandwidth
1109 * by migrating from more heavily used groups, to less
1110 * heavily used ones, spreading the load around.
1111 * Use a 1/4 hysteresis to avoid spurious page movement.
1113 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1116 static unsigned long weighted_cpuload(const int cpu);
1117 static unsigned long source_load(int cpu, int type);
1118 static unsigned long target_load(int cpu, int type);
1119 static unsigned long capacity_of(int cpu);
1120 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1122 /* Cached statistics for all CPUs within a node */
1124 unsigned long nr_running;
1127 /* Total compute capacity of CPUs on a node */
1128 unsigned long compute_capacity;
1130 /* Approximate capacity in terms of runnable tasks on a node */
1131 unsigned long task_capacity;
1132 int has_free_capacity;
1136 * XXX borrowed from update_sg_lb_stats
1138 static void update_numa_stats(struct numa_stats *ns, int nid)
1140 int smt, cpu, cpus = 0;
1141 unsigned long capacity;
1143 memset(ns, 0, sizeof(*ns));
1144 for_each_cpu(cpu, cpumask_of_node(nid)) {
1145 struct rq *rq = cpu_rq(cpu);
1147 ns->nr_running += rq->nr_running;
1148 ns->load += weighted_cpuload(cpu);
1149 ns->compute_capacity += capacity_of(cpu);
1155 * If we raced with hotplug and there are no CPUs left in our mask
1156 * the @ns structure is NULL'ed and task_numa_compare() will
1157 * not find this node attractive.
1159 * We'll either bail at !has_free_capacity, or we'll detect a huge
1160 * imbalance and bail there.
1165 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1166 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1167 capacity = cpus / smt; /* cores */
1169 ns->task_capacity = min_t(unsigned, capacity,
1170 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1171 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1174 struct task_numa_env {
1175 struct task_struct *p;
1177 int src_cpu, src_nid;
1178 int dst_cpu, dst_nid;
1180 struct numa_stats src_stats, dst_stats;
1185 struct task_struct *best_task;
1190 static void task_numa_assign(struct task_numa_env *env,
1191 struct task_struct *p, long imp)
1194 put_task_struct(env->best_task);
1199 env->best_imp = imp;
1200 env->best_cpu = env->dst_cpu;
1203 static bool load_too_imbalanced(long src_load, long dst_load,
1204 struct task_numa_env *env)
1207 long orig_src_load, orig_dst_load;
1208 long src_capacity, dst_capacity;
1211 * The load is corrected for the CPU capacity available on each node.
1214 * ------------ vs ---------
1215 * src_capacity dst_capacity
1217 src_capacity = env->src_stats.compute_capacity;
1218 dst_capacity = env->dst_stats.compute_capacity;
1220 /* We care about the slope of the imbalance, not the direction. */
1221 if (dst_load < src_load)
1222 swap(dst_load, src_load);
1224 /* Is the difference below the threshold? */
1225 imb = dst_load * src_capacity * 100 -
1226 src_load * dst_capacity * env->imbalance_pct;
1231 * The imbalance is above the allowed threshold.
1232 * Compare it with the old imbalance.
1234 orig_src_load = env->src_stats.load;
1235 orig_dst_load = env->dst_stats.load;
1237 if (orig_dst_load < orig_src_load)
1238 swap(orig_dst_load, orig_src_load);
1240 old_imb = orig_dst_load * src_capacity * 100 -
1241 orig_src_load * dst_capacity * env->imbalance_pct;
1243 /* Would this change make things worse? */
1244 return (imb > old_imb);
1248 * This checks if the overall compute and NUMA accesses of the system would
1249 * be improved if the source tasks was migrated to the target dst_cpu taking
1250 * into account that it might be best if task running on the dst_cpu should
1251 * be exchanged with the source task
1253 static void task_numa_compare(struct task_numa_env *env,
1254 long taskimp, long groupimp)
1256 struct rq *src_rq = cpu_rq(env->src_cpu);
1257 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1258 struct task_struct *cur;
1259 long src_load, dst_load;
1261 long imp = env->p->numa_group ? groupimp : taskimp;
1263 int dist = env->dist;
1267 raw_spin_lock_irq(&dst_rq->lock);
1270 * No need to move the exiting task, and this ensures that ->curr
1271 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1272 * is safe under RCU read lock.
1273 * Note that rcu_read_lock() itself can't protect from the final
1274 * put_task_struct() after the last schedule().
1276 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1278 raw_spin_unlock_irq(&dst_rq->lock);
1281 * Because we have preemption enabled we can get migrated around and
1282 * end try selecting ourselves (current == env->p) as a swap candidate.
1288 * "imp" is the fault differential for the source task between the
1289 * source and destination node. Calculate the total differential for
1290 * the source task and potential destination task. The more negative
1291 * the value is, the more rmeote accesses that would be expected to
1292 * be incurred if the tasks were swapped.
1295 /* Skip this swap candidate if cannot move to the source cpu */
1296 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1300 * If dst and source tasks are in the same NUMA group, or not
1301 * in any group then look only at task weights.
1303 if (cur->numa_group == env->p->numa_group) {
1304 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1305 task_weight(cur, env->dst_nid, dist);
1307 * Add some hysteresis to prevent swapping the
1308 * tasks within a group over tiny differences.
1310 if (cur->numa_group)
1314 * Compare the group weights. If a task is all by
1315 * itself (not part of a group), use the task weight
1318 if (cur->numa_group)
1319 imp += group_weight(cur, env->src_nid, dist) -
1320 group_weight(cur, env->dst_nid, dist);
1322 imp += task_weight(cur, env->src_nid, dist) -
1323 task_weight(cur, env->dst_nid, dist);
1327 if (imp <= env->best_imp && moveimp <= env->best_imp)
1331 /* Is there capacity at our destination? */
1332 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1333 !env->dst_stats.has_free_capacity)
1339 /* Balance doesn't matter much if we're running a task per cpu */
1340 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1341 dst_rq->nr_running == 1)
1345 * In the overloaded case, try and keep the load balanced.
1348 load = task_h_load(env->p);
1349 dst_load = env->dst_stats.load + load;
1350 src_load = env->src_stats.load - load;
1352 if (moveimp > imp && moveimp > env->best_imp) {
1354 * If the improvement from just moving env->p direction is
1355 * better than swapping tasks around, check if a move is
1356 * possible. Store a slightly smaller score than moveimp,
1357 * so an actually idle CPU will win.
1359 if (!load_too_imbalanced(src_load, dst_load, env)) {
1366 if (imp <= env->best_imp)
1370 load = task_h_load(cur);
1375 if (load_too_imbalanced(src_load, dst_load, env))
1379 * One idle CPU per node is evaluated for a task numa move.
1380 * Call select_idle_sibling to maybe find a better one.
1383 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1386 task_numa_assign(env, cur, imp);
1391 static void task_numa_find_cpu(struct task_numa_env *env,
1392 long taskimp, long groupimp)
1396 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1397 /* Skip this CPU if the source task cannot migrate */
1398 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1402 task_numa_compare(env, taskimp, groupimp);
1406 /* Only move tasks to a NUMA node less busy than the current node. */
1407 static bool numa_has_capacity(struct task_numa_env *env)
1409 struct numa_stats *src = &env->src_stats;
1410 struct numa_stats *dst = &env->dst_stats;
1412 if (src->has_free_capacity && !dst->has_free_capacity)
1416 * Only consider a task move if the source has a higher load
1417 * than the destination, corrected for CPU capacity on each node.
1419 * src->load dst->load
1420 * --------------------- vs ---------------------
1421 * src->compute_capacity dst->compute_capacity
1423 if (src->load * dst->compute_capacity * env->imbalance_pct >
1425 dst->load * src->compute_capacity * 100)
1431 static int task_numa_migrate(struct task_struct *p)
1433 struct task_numa_env env = {
1436 .src_cpu = task_cpu(p),
1437 .src_nid = task_node(p),
1439 .imbalance_pct = 112,
1445 struct sched_domain *sd;
1446 unsigned long taskweight, groupweight;
1448 long taskimp, groupimp;
1451 * Pick the lowest SD_NUMA domain, as that would have the smallest
1452 * imbalance and would be the first to start moving tasks about.
1454 * And we want to avoid any moving of tasks about, as that would create
1455 * random movement of tasks -- counter the numa conditions we're trying
1459 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1461 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1465 * Cpusets can break the scheduler domain tree into smaller
1466 * balance domains, some of which do not cross NUMA boundaries.
1467 * Tasks that are "trapped" in such domains cannot be migrated
1468 * elsewhere, so there is no point in (re)trying.
1470 if (unlikely(!sd)) {
1471 p->numa_preferred_nid = task_node(p);
1475 env.dst_nid = p->numa_preferred_nid;
1476 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1477 taskweight = task_weight(p, env.src_nid, dist);
1478 groupweight = group_weight(p, env.src_nid, dist);
1479 update_numa_stats(&env.src_stats, env.src_nid);
1480 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1481 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1482 update_numa_stats(&env.dst_stats, env.dst_nid);
1484 /* Try to find a spot on the preferred nid. */
1485 if (numa_has_capacity(&env))
1486 task_numa_find_cpu(&env, taskimp, groupimp);
1489 * Look at other nodes in these cases:
1490 * - there is no space available on the preferred_nid
1491 * - the task is part of a numa_group that is interleaved across
1492 * multiple NUMA nodes; in order to better consolidate the group,
1493 * we need to check other locations.
1495 if (env.best_cpu == -1 || (p->numa_group &&
1496 nodes_weight(p->numa_group->active_nodes) > 1)) {
1497 for_each_online_node(nid) {
1498 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1501 dist = node_distance(env.src_nid, env.dst_nid);
1502 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1504 taskweight = task_weight(p, env.src_nid, dist);
1505 groupweight = group_weight(p, env.src_nid, dist);
1508 /* Only consider nodes where both task and groups benefit */
1509 taskimp = task_weight(p, nid, dist) - taskweight;
1510 groupimp = group_weight(p, nid, dist) - groupweight;
1511 if (taskimp < 0 && groupimp < 0)
1516 update_numa_stats(&env.dst_stats, env.dst_nid);
1517 if (numa_has_capacity(&env))
1518 task_numa_find_cpu(&env, taskimp, groupimp);
1523 * If the task is part of a workload that spans multiple NUMA nodes,
1524 * and is migrating into one of the workload's active nodes, remember
1525 * this node as the task's preferred numa node, so the workload can
1527 * A task that migrated to a second choice node will be better off
1528 * trying for a better one later. Do not set the preferred node here.
1530 if (p->numa_group) {
1531 if (env.best_cpu == -1)
1536 if (node_isset(nid, p->numa_group->active_nodes))
1537 sched_setnuma(p, env.dst_nid);
1540 /* No better CPU than the current one was found. */
1541 if (env.best_cpu == -1)
1545 * Reset the scan period if the task is being rescheduled on an
1546 * alternative node to recheck if the tasks is now properly placed.
1548 p->numa_scan_period = task_scan_min(p);
1550 if (env.best_task == NULL) {
1551 ret = migrate_task_to(p, env.best_cpu);
1553 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1557 ret = migrate_swap(p, env.best_task);
1559 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1560 put_task_struct(env.best_task);
1564 /* Attempt to migrate a task to a CPU on the preferred node. */
1565 static void numa_migrate_preferred(struct task_struct *p)
1567 unsigned long interval = HZ;
1569 /* This task has no NUMA fault statistics yet */
1570 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1573 /* Periodically retry migrating the task to the preferred node */
1574 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1575 p->numa_migrate_retry = jiffies + interval;
1577 /* Success if task is already running on preferred CPU */
1578 if (task_node(p) == p->numa_preferred_nid)
1581 /* Otherwise, try migrate to a CPU on the preferred node */
1582 task_numa_migrate(p);
1586 * Find the nodes on which the workload is actively running. We do this by
1587 * tracking the nodes from which NUMA hinting faults are triggered. This can
1588 * be different from the set of nodes where the workload's memory is currently
1591 * The bitmask is used to make smarter decisions on when to do NUMA page
1592 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1593 * are added when they cause over 6/16 of the maximum number of faults, but
1594 * only removed when they drop below 3/16.
1596 static void update_numa_active_node_mask(struct numa_group *numa_group)
1598 unsigned long faults, max_faults = 0;
1601 for_each_online_node(nid) {
1602 faults = group_faults_cpu(numa_group, nid);
1603 if (faults > max_faults)
1604 max_faults = faults;
1607 for_each_online_node(nid) {
1608 faults = group_faults_cpu(numa_group, nid);
1609 if (!node_isset(nid, numa_group->active_nodes)) {
1610 if (faults > max_faults * 6 / 16)
1611 node_set(nid, numa_group->active_nodes);
1612 } else if (faults < max_faults * 3 / 16)
1613 node_clear(nid, numa_group->active_nodes);
1618 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1619 * increments. The more local the fault statistics are, the higher the scan
1620 * period will be for the next scan window. If local/(local+remote) ratio is
1621 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1622 * the scan period will decrease. Aim for 70% local accesses.
1624 #define NUMA_PERIOD_SLOTS 10
1625 #define NUMA_PERIOD_THRESHOLD 7
1628 * Increase the scan period (slow down scanning) if the majority of
1629 * our memory is already on our local node, or if the majority of
1630 * the page accesses are shared with other processes.
1631 * Otherwise, decrease the scan period.
1633 static void update_task_scan_period(struct task_struct *p,
1634 unsigned long shared, unsigned long private)
1636 unsigned int period_slot;
1640 unsigned long remote = p->numa_faults_locality[0];
1641 unsigned long local = p->numa_faults_locality[1];
1644 * If there were no record hinting faults then either the task is
1645 * completely idle or all activity is areas that are not of interest
1646 * to automatic numa balancing. Related to that, if there were failed
1647 * migration then it implies we are migrating too quickly or the local
1648 * node is overloaded. In either case, scan slower
1650 if (local + shared == 0 || p->numa_faults_locality[2]) {
1651 p->numa_scan_period = min(p->numa_scan_period_max,
1652 p->numa_scan_period << 1);
1654 p->mm->numa_next_scan = jiffies +
1655 msecs_to_jiffies(p->numa_scan_period);
1661 * Prepare to scale scan period relative to the current period.
1662 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1663 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1664 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1666 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1667 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1668 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1669 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1672 diff = slot * period_slot;
1674 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1677 * Scale scan rate increases based on sharing. There is an
1678 * inverse relationship between the degree of sharing and
1679 * the adjustment made to the scanning period. Broadly
1680 * speaking the intent is that there is little point
1681 * scanning faster if shared accesses dominate as it may
1682 * simply bounce migrations uselessly
1684 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1685 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1688 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1689 task_scan_min(p), task_scan_max(p));
1690 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1694 * Get the fraction of time the task has been running since the last
1695 * NUMA placement cycle. The scheduler keeps similar statistics, but
1696 * decays those on a 32ms period, which is orders of magnitude off
1697 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1698 * stats only if the task is so new there are no NUMA statistics yet.
1700 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1702 u64 runtime, delta, now;
1703 /* Use the start of this time slice to avoid calculations. */
1704 now = p->se.exec_start;
1705 runtime = p->se.sum_exec_runtime;
1707 if (p->last_task_numa_placement) {
1708 delta = runtime - p->last_sum_exec_runtime;
1709 *period = now - p->last_task_numa_placement;
1711 delta = p->se.avg.load_sum / p->se.load.weight;
1712 *period = LOAD_AVG_MAX;
1715 p->last_sum_exec_runtime = runtime;
1716 p->last_task_numa_placement = now;
1722 * Determine the preferred nid for a task in a numa_group. This needs to
1723 * be done in a way that produces consistent results with group_weight,
1724 * otherwise workloads might not converge.
1726 static int preferred_group_nid(struct task_struct *p, int nid)
1731 /* Direct connections between all NUMA nodes. */
1732 if (sched_numa_topology_type == NUMA_DIRECT)
1736 * On a system with glueless mesh NUMA topology, group_weight
1737 * scores nodes according to the number of NUMA hinting faults on
1738 * both the node itself, and on nearby nodes.
1740 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1741 unsigned long score, max_score = 0;
1742 int node, max_node = nid;
1744 dist = sched_max_numa_distance;
1746 for_each_online_node(node) {
1747 score = group_weight(p, node, dist);
1748 if (score > max_score) {
1757 * Finding the preferred nid in a system with NUMA backplane
1758 * interconnect topology is more involved. The goal is to locate
1759 * tasks from numa_groups near each other in the system, and
1760 * untangle workloads from different sides of the system. This requires
1761 * searching down the hierarchy of node groups, recursively searching
1762 * inside the highest scoring group of nodes. The nodemask tricks
1763 * keep the complexity of the search down.
1765 nodes = node_online_map;
1766 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1767 unsigned long max_faults = 0;
1768 nodemask_t max_group = NODE_MASK_NONE;
1771 /* Are there nodes at this distance from each other? */
1772 if (!find_numa_distance(dist))
1775 for_each_node_mask(a, nodes) {
1776 unsigned long faults = 0;
1777 nodemask_t this_group;
1778 nodes_clear(this_group);
1780 /* Sum group's NUMA faults; includes a==b case. */
1781 for_each_node_mask(b, nodes) {
1782 if (node_distance(a, b) < dist) {
1783 faults += group_faults(p, b);
1784 node_set(b, this_group);
1785 node_clear(b, nodes);
1789 /* Remember the top group. */
1790 if (faults > max_faults) {
1791 max_faults = faults;
1792 max_group = this_group;
1794 * subtle: at the smallest distance there is
1795 * just one node left in each "group", the
1796 * winner is the preferred nid.
1801 /* Next round, evaluate the nodes within max_group. */
1809 static void task_numa_placement(struct task_struct *p)
1811 int seq, nid, max_nid = -1, max_group_nid = -1;
1812 unsigned long max_faults = 0, max_group_faults = 0;
1813 unsigned long fault_types[2] = { 0, 0 };
1814 unsigned long total_faults;
1815 u64 runtime, period;
1816 spinlock_t *group_lock = NULL;
1819 * The p->mm->numa_scan_seq field gets updated without
1820 * exclusive access. Use READ_ONCE() here to ensure
1821 * that the field is read in a single access:
1823 seq = READ_ONCE(p->mm->numa_scan_seq);
1824 if (p->numa_scan_seq == seq)
1826 p->numa_scan_seq = seq;
1827 p->numa_scan_period_max = task_scan_max(p);
1829 total_faults = p->numa_faults_locality[0] +
1830 p->numa_faults_locality[1];
1831 runtime = numa_get_avg_runtime(p, &period);
1833 /* If the task is part of a group prevent parallel updates to group stats */
1834 if (p->numa_group) {
1835 group_lock = &p->numa_group->lock;
1836 spin_lock_irq(group_lock);
1839 /* Find the node with the highest number of faults */
1840 for_each_online_node(nid) {
1841 /* Keep track of the offsets in numa_faults array */
1842 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1843 unsigned long faults = 0, group_faults = 0;
1846 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1847 long diff, f_diff, f_weight;
1849 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1850 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1851 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1852 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1854 /* Decay existing window, copy faults since last scan */
1855 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1856 fault_types[priv] += p->numa_faults[membuf_idx];
1857 p->numa_faults[membuf_idx] = 0;
1860 * Normalize the faults_from, so all tasks in a group
1861 * count according to CPU use, instead of by the raw
1862 * number of faults. Tasks with little runtime have
1863 * little over-all impact on throughput, and thus their
1864 * faults are less important.
1866 f_weight = div64_u64(runtime << 16, period + 1);
1867 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1869 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1870 p->numa_faults[cpubuf_idx] = 0;
1872 p->numa_faults[mem_idx] += diff;
1873 p->numa_faults[cpu_idx] += f_diff;
1874 faults += p->numa_faults[mem_idx];
1875 p->total_numa_faults += diff;
1876 if (p->numa_group) {
1878 * safe because we can only change our own group
1880 * mem_idx represents the offset for a given
1881 * nid and priv in a specific region because it
1882 * is at the beginning of the numa_faults array.
1884 p->numa_group->faults[mem_idx] += diff;
1885 p->numa_group->faults_cpu[mem_idx] += f_diff;
1886 p->numa_group->total_faults += diff;
1887 group_faults += p->numa_group->faults[mem_idx];
1891 if (faults > max_faults) {
1892 max_faults = faults;
1896 if (group_faults > max_group_faults) {
1897 max_group_faults = group_faults;
1898 max_group_nid = nid;
1902 update_task_scan_period(p, fault_types[0], fault_types[1]);
1904 if (p->numa_group) {
1905 update_numa_active_node_mask(p->numa_group);
1906 spin_unlock_irq(group_lock);
1907 max_nid = preferred_group_nid(p, max_group_nid);
1911 /* Set the new preferred node */
1912 if (max_nid != p->numa_preferred_nid)
1913 sched_setnuma(p, max_nid);
1915 if (task_node(p) != p->numa_preferred_nid)
1916 numa_migrate_preferred(p);
1920 static inline int get_numa_group(struct numa_group *grp)
1922 return atomic_inc_not_zero(&grp->refcount);
1925 static inline void put_numa_group(struct numa_group *grp)
1927 if (atomic_dec_and_test(&grp->refcount))
1928 kfree_rcu(grp, rcu);
1931 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1934 struct numa_group *grp, *my_grp;
1935 struct task_struct *tsk;
1937 int cpu = cpupid_to_cpu(cpupid);
1940 if (unlikely(!p->numa_group)) {
1941 unsigned int size = sizeof(struct numa_group) +
1942 4*nr_node_ids*sizeof(unsigned long);
1944 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1948 atomic_set(&grp->refcount, 1);
1949 spin_lock_init(&grp->lock);
1951 /* Second half of the array tracks nids where faults happen */
1952 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1955 node_set(task_node(current), grp->active_nodes);
1957 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1958 grp->faults[i] = p->numa_faults[i];
1960 grp->total_faults = p->total_numa_faults;
1963 rcu_assign_pointer(p->numa_group, grp);
1967 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1969 if (!cpupid_match_pid(tsk, cpupid))
1972 grp = rcu_dereference(tsk->numa_group);
1976 my_grp = p->numa_group;
1981 * Only join the other group if its bigger; if we're the bigger group,
1982 * the other task will join us.
1984 if (my_grp->nr_tasks > grp->nr_tasks)
1988 * Tie-break on the grp address.
1990 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1993 /* Always join threads in the same process. */
1994 if (tsk->mm == current->mm)
1997 /* Simple filter to avoid false positives due to PID collisions */
1998 if (flags & TNF_SHARED)
2001 /* Update priv based on whether false sharing was detected */
2004 if (join && !get_numa_group(grp))
2012 BUG_ON(irqs_disabled());
2013 double_lock_irq(&my_grp->lock, &grp->lock);
2015 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2016 my_grp->faults[i] -= p->numa_faults[i];
2017 grp->faults[i] += p->numa_faults[i];
2019 my_grp->total_faults -= p->total_numa_faults;
2020 grp->total_faults += p->total_numa_faults;
2025 spin_unlock(&my_grp->lock);
2026 spin_unlock_irq(&grp->lock);
2028 rcu_assign_pointer(p->numa_group, grp);
2030 put_numa_group(my_grp);
2038 void task_numa_free(struct task_struct *p)
2040 struct numa_group *grp = p->numa_group;
2041 void *numa_faults = p->numa_faults;
2042 unsigned long flags;
2046 spin_lock_irqsave(&grp->lock, flags);
2047 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2048 grp->faults[i] -= p->numa_faults[i];
2049 grp->total_faults -= p->total_numa_faults;
2052 spin_unlock_irqrestore(&grp->lock, flags);
2053 RCU_INIT_POINTER(p->numa_group, NULL);
2054 put_numa_group(grp);
2057 p->numa_faults = NULL;
2062 * Got a PROT_NONE fault for a page on @node.
2064 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2066 struct task_struct *p = current;
2067 bool migrated = flags & TNF_MIGRATED;
2068 int cpu_node = task_node(current);
2069 int local = !!(flags & TNF_FAULT_LOCAL);
2072 if (!static_branch_likely(&sched_numa_balancing))
2075 /* for example, ksmd faulting in a user's mm */
2079 /* Allocate buffer to track faults on a per-node basis */
2080 if (unlikely(!p->numa_faults)) {
2081 int size = sizeof(*p->numa_faults) *
2082 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2084 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2085 if (!p->numa_faults)
2088 p->total_numa_faults = 0;
2089 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2093 * First accesses are treated as private, otherwise consider accesses
2094 * to be private if the accessing pid has not changed
2096 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2099 priv = cpupid_match_pid(p, last_cpupid);
2100 if (!priv && !(flags & TNF_NO_GROUP))
2101 task_numa_group(p, last_cpupid, flags, &priv);
2105 * If a workload spans multiple NUMA nodes, a shared fault that
2106 * occurs wholly within the set of nodes that the workload is
2107 * actively using should be counted as local. This allows the
2108 * scan rate to slow down when a workload has settled down.
2110 if (!priv && !local && p->numa_group &&
2111 node_isset(cpu_node, p->numa_group->active_nodes) &&
2112 node_isset(mem_node, p->numa_group->active_nodes))
2115 task_numa_placement(p);
2118 * Retry task to preferred node migration periodically, in case it
2119 * case it previously failed, or the scheduler moved us.
2121 if (time_after(jiffies, p->numa_migrate_retry))
2122 numa_migrate_preferred(p);
2125 p->numa_pages_migrated += pages;
2126 if (flags & TNF_MIGRATE_FAIL)
2127 p->numa_faults_locality[2] += pages;
2129 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2130 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2131 p->numa_faults_locality[local] += pages;
2134 static void reset_ptenuma_scan(struct task_struct *p)
2137 * We only did a read acquisition of the mmap sem, so
2138 * p->mm->numa_scan_seq is written to without exclusive access
2139 * and the update is not guaranteed to be atomic. That's not
2140 * much of an issue though, since this is just used for
2141 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2142 * expensive, to avoid any form of compiler optimizations:
2144 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2145 p->mm->numa_scan_offset = 0;
2149 * The expensive part of numa migration is done from task_work context.
2150 * Triggered from task_tick_numa().
2152 void task_numa_work(struct callback_head *work)
2154 unsigned long migrate, next_scan, now = jiffies;
2155 struct task_struct *p = current;
2156 struct mm_struct *mm = p->mm;
2157 struct vm_area_struct *vma;
2158 unsigned long start, end;
2159 unsigned long nr_pte_updates = 0;
2160 long pages, virtpages;
2162 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2164 work->next = work; /* protect against double add */
2166 * Who cares about NUMA placement when they're dying.
2168 * NOTE: make sure not to dereference p->mm before this check,
2169 * exit_task_work() happens _after_ exit_mm() so we could be called
2170 * without p->mm even though we still had it when we enqueued this
2173 if (p->flags & PF_EXITING)
2176 if (!mm->numa_next_scan) {
2177 mm->numa_next_scan = now +
2178 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2182 * Enforce maximal scan/migration frequency..
2184 migrate = mm->numa_next_scan;
2185 if (time_before(now, migrate))
2188 if (p->numa_scan_period == 0) {
2189 p->numa_scan_period_max = task_scan_max(p);
2190 p->numa_scan_period = task_scan_min(p);
2193 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2194 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2198 * Delay this task enough that another task of this mm will likely win
2199 * the next time around.
2201 p->node_stamp += 2 * TICK_NSEC;
2203 start = mm->numa_scan_offset;
2204 pages = sysctl_numa_balancing_scan_size;
2205 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2206 virtpages = pages * 8; /* Scan up to this much virtual space */
2211 down_read(&mm->mmap_sem);
2212 vma = find_vma(mm, start);
2214 reset_ptenuma_scan(p);
2218 for (; vma; vma = vma->vm_next) {
2219 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2220 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2225 * Shared library pages mapped by multiple processes are not
2226 * migrated as it is expected they are cache replicated. Avoid
2227 * hinting faults in read-only file-backed mappings or the vdso
2228 * as migrating the pages will be of marginal benefit.
2231 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2235 * Skip inaccessible VMAs to avoid any confusion between
2236 * PROT_NONE and NUMA hinting ptes
2238 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2242 start = max(start, vma->vm_start);
2243 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2244 end = min(end, vma->vm_end);
2245 nr_pte_updates = change_prot_numa(vma, start, end);
2248 * Try to scan sysctl_numa_balancing_size worth of
2249 * hpages that have at least one present PTE that
2250 * is not already pte-numa. If the VMA contains
2251 * areas that are unused or already full of prot_numa
2252 * PTEs, scan up to virtpages, to skip through those
2256 pages -= (end - start) >> PAGE_SHIFT;
2257 virtpages -= (end - start) >> PAGE_SHIFT;
2260 if (pages <= 0 || virtpages <= 0)
2264 } while (end != vma->vm_end);
2269 * It is possible to reach the end of the VMA list but the last few
2270 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2271 * would find the !migratable VMA on the next scan but not reset the
2272 * scanner to the start so check it now.
2275 mm->numa_scan_offset = start;
2277 reset_ptenuma_scan(p);
2278 up_read(&mm->mmap_sem);
2282 * Drive the periodic memory faults..
2284 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2286 struct callback_head *work = &curr->numa_work;
2290 * We don't care about NUMA placement if we don't have memory.
2292 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2296 * Using runtime rather than walltime has the dual advantage that
2297 * we (mostly) drive the selection from busy threads and that the
2298 * task needs to have done some actual work before we bother with
2301 now = curr->se.sum_exec_runtime;
2302 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2304 if (now > curr->node_stamp + period) {
2305 if (!curr->node_stamp)
2306 curr->numa_scan_period = task_scan_min(curr);
2307 curr->node_stamp += period;
2309 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2310 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2311 task_work_add(curr, work, true);
2316 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2320 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2324 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2327 #endif /* CONFIG_NUMA_BALANCING */
2330 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2332 update_load_add(&cfs_rq->load, se->load.weight);
2333 if (!parent_entity(se))
2334 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2336 if (entity_is_task(se)) {
2337 struct rq *rq = rq_of(cfs_rq);
2339 account_numa_enqueue(rq, task_of(se));
2340 list_add(&se->group_node, &rq->cfs_tasks);
2343 cfs_rq->nr_running++;
2347 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2349 update_load_sub(&cfs_rq->load, se->load.weight);
2350 if (!parent_entity(se))
2351 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2352 if (entity_is_task(se)) {
2353 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2354 list_del_init(&se->group_node);
2356 cfs_rq->nr_running--;
2359 #ifdef CONFIG_FAIR_GROUP_SCHED
2361 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2366 * Use this CPU's real-time load instead of the last load contribution
2367 * as the updating of the contribution is delayed, and we will use the
2368 * the real-time load to calc the share. See update_tg_load_avg().
2370 tg_weight = atomic_long_read(&tg->load_avg);
2371 tg_weight -= cfs_rq->tg_load_avg_contrib;
2372 tg_weight += cfs_rq->load.weight;
2377 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2379 long tg_weight, load, shares;
2381 tg_weight = calc_tg_weight(tg, cfs_rq);
2382 load = cfs_rq->load.weight;
2384 shares = (tg->shares * load);
2386 shares /= tg_weight;
2388 if (shares < MIN_SHARES)
2389 shares = MIN_SHARES;
2390 if (shares > tg->shares)
2391 shares = tg->shares;
2395 # else /* CONFIG_SMP */
2396 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2400 # endif /* CONFIG_SMP */
2401 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2402 unsigned long weight)
2405 /* commit outstanding execution time */
2406 if (cfs_rq->curr == se)
2407 update_curr(cfs_rq);
2408 account_entity_dequeue(cfs_rq, se);
2411 update_load_set(&se->load, weight);
2414 account_entity_enqueue(cfs_rq, se);
2417 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2419 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2421 struct task_group *tg;
2422 struct sched_entity *se;
2426 se = tg->se[cpu_of(rq_of(cfs_rq))];
2427 if (!se || throttled_hierarchy(cfs_rq))
2430 if (likely(se->load.weight == tg->shares))
2433 shares = calc_cfs_shares(cfs_rq, tg);
2435 reweight_entity(cfs_rq_of(se), se, shares);
2437 #else /* CONFIG_FAIR_GROUP_SCHED */
2438 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2441 #endif /* CONFIG_FAIR_GROUP_SCHED */
2444 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2445 static const u32 runnable_avg_yN_inv[] = {
2446 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2447 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2448 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2449 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2450 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2451 0x85aac367, 0x82cd8698,
2455 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2456 * over-estimates when re-combining.
2458 static const u32 runnable_avg_yN_sum[] = {
2459 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2460 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2461 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2466 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2468 static __always_inline u64 decay_load(u64 val, u64 n)
2470 unsigned int local_n;
2474 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2477 /* after bounds checking we can collapse to 32-bit */
2481 * As y^PERIOD = 1/2, we can combine
2482 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2483 * With a look-up table which covers y^n (n<PERIOD)
2485 * To achieve constant time decay_load.
2487 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2488 val >>= local_n / LOAD_AVG_PERIOD;
2489 local_n %= LOAD_AVG_PERIOD;
2492 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2497 * For updates fully spanning n periods, the contribution to runnable
2498 * average will be: \Sum 1024*y^n
2500 * We can compute this reasonably efficiently by combining:
2501 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2503 static u32 __compute_runnable_contrib(u64 n)
2507 if (likely(n <= LOAD_AVG_PERIOD))
2508 return runnable_avg_yN_sum[n];
2509 else if (unlikely(n >= LOAD_AVG_MAX_N))
2510 return LOAD_AVG_MAX;
2512 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2514 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2515 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2517 n -= LOAD_AVG_PERIOD;
2518 } while (n > LOAD_AVG_PERIOD);
2520 contrib = decay_load(contrib, n);
2521 return contrib + runnable_avg_yN_sum[n];
2524 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2525 #error "load tracking assumes 2^10 as unit"
2528 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2531 * We can represent the historical contribution to runnable average as the
2532 * coefficients of a geometric series. To do this we sub-divide our runnable
2533 * history into segments of approximately 1ms (1024us); label the segment that
2534 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2536 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2538 * (now) (~1ms ago) (~2ms ago)
2540 * Let u_i denote the fraction of p_i that the entity was runnable.
2542 * We then designate the fractions u_i as our co-efficients, yielding the
2543 * following representation of historical load:
2544 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2546 * We choose y based on the with of a reasonably scheduling period, fixing:
2549 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2550 * approximately half as much as the contribution to load within the last ms
2553 * When a period "rolls over" and we have new u_0`, multiplying the previous
2554 * sum again by y is sufficient to update:
2555 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2556 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2558 static __always_inline int
2559 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2560 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2562 u64 delta, scaled_delta, periods;
2564 unsigned int delta_w, scaled_delta_w, decayed = 0;
2565 unsigned long scale_freq, scale_cpu;
2567 delta = now - sa->last_update_time;
2569 * This should only happen when time goes backwards, which it
2570 * unfortunately does during sched clock init when we swap over to TSC.
2572 if ((s64)delta < 0) {
2573 sa->last_update_time = now;
2578 * Use 1024ns as the unit of measurement since it's a reasonable
2579 * approximation of 1us and fast to compute.
2584 sa->last_update_time = now;
2586 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2587 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2589 /* delta_w is the amount already accumulated against our next period */
2590 delta_w = sa->period_contrib;
2591 if (delta + delta_w >= 1024) {
2594 /* how much left for next period will start over, we don't know yet */
2595 sa->period_contrib = 0;
2598 * Now that we know we're crossing a period boundary, figure
2599 * out how much from delta we need to complete the current
2600 * period and accrue it.
2602 delta_w = 1024 - delta_w;
2603 scaled_delta_w = cap_scale(delta_w, scale_freq);
2605 sa->load_sum += weight * scaled_delta_w;
2607 cfs_rq->runnable_load_sum +=
2608 weight * scaled_delta_w;
2612 sa->util_sum += scaled_delta_w * scale_cpu;
2616 /* Figure out how many additional periods this update spans */
2617 periods = delta / 1024;
2620 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2622 cfs_rq->runnable_load_sum =
2623 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2625 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2627 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2628 contrib = __compute_runnable_contrib(periods);
2629 contrib = cap_scale(contrib, scale_freq);
2631 sa->load_sum += weight * contrib;
2633 cfs_rq->runnable_load_sum += weight * contrib;
2636 sa->util_sum += contrib * scale_cpu;
2639 /* Remainder of delta accrued against u_0` */
2640 scaled_delta = cap_scale(delta, scale_freq);
2642 sa->load_sum += weight * scaled_delta;
2644 cfs_rq->runnable_load_sum += weight * scaled_delta;
2647 sa->util_sum += scaled_delta * scale_cpu;
2649 sa->period_contrib += delta;
2652 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2654 cfs_rq->runnable_load_avg =
2655 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2657 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2663 #ifdef CONFIG_FAIR_GROUP_SCHED
2665 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2666 * and effective_load (which is not done because it is too costly).
2668 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2670 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2672 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2673 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2674 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2678 #else /* CONFIG_FAIR_GROUP_SCHED */
2679 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2680 #endif /* CONFIG_FAIR_GROUP_SCHED */
2682 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2685 * Unsigned subtract and clamp on underflow.
2687 * Explicitly do a load-store to ensure the intermediate value never hits
2688 * memory. This allows lockless observations without ever seeing the negative
2691 #define sub_positive(_ptr, _val) do { \
2692 typeof(_ptr) ptr = (_ptr); \
2693 typeof(*ptr) val = (_val); \
2694 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2698 WRITE_ONCE(*ptr, res); \
2701 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2702 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2704 struct sched_avg *sa = &cfs_rq->avg;
2705 int decayed, removed = 0;
2707 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2708 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2709 sub_positive(&sa->load_avg, r);
2710 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2714 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2715 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2716 sub_positive(&sa->util_avg, r);
2717 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2720 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2721 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2723 #ifndef CONFIG_64BIT
2725 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2728 return decayed || removed;
2731 /* Update task and its cfs_rq load average */
2732 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2734 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2735 u64 now = cfs_rq_clock_task(cfs_rq);
2736 int cpu = cpu_of(rq_of(cfs_rq));
2739 * Track task load average for carrying it to new CPU after migrated, and
2740 * track group sched_entity load average for task_h_load calc in migration
2742 __update_load_avg(now, cpu, &se->avg,
2743 se->on_rq * scale_load_down(se->load.weight),
2744 cfs_rq->curr == se, NULL);
2746 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2747 update_tg_load_avg(cfs_rq, 0);
2750 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2752 if (!sched_feat(ATTACH_AGE_LOAD))
2756 * If we got migrated (either between CPUs or between cgroups) we'll
2757 * have aged the average right before clearing @last_update_time.
2759 if (se->avg.last_update_time) {
2760 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2761 &se->avg, 0, 0, NULL);
2764 * XXX: we could have just aged the entire load away if we've been
2765 * absent from the fair class for too long.
2770 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2771 cfs_rq->avg.load_avg += se->avg.load_avg;
2772 cfs_rq->avg.load_sum += se->avg.load_sum;
2773 cfs_rq->avg.util_avg += se->avg.util_avg;
2774 cfs_rq->avg.util_sum += se->avg.util_sum;
2777 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2779 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2780 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2781 cfs_rq->curr == se, NULL);
2783 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2784 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2785 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2786 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2789 /* Add the load generated by se into cfs_rq's load average */
2791 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2793 struct sched_avg *sa = &se->avg;
2794 u64 now = cfs_rq_clock_task(cfs_rq);
2795 int migrated, decayed;
2797 migrated = !sa->last_update_time;
2799 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2800 se->on_rq * scale_load_down(se->load.weight),
2801 cfs_rq->curr == se, NULL);
2804 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2806 cfs_rq->runnable_load_avg += sa->load_avg;
2807 cfs_rq->runnable_load_sum += sa->load_sum;
2810 attach_entity_load_avg(cfs_rq, se);
2812 if (decayed || migrated)
2813 update_tg_load_avg(cfs_rq, 0);
2816 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2818 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2820 update_load_avg(se, 1);
2822 cfs_rq->runnable_load_avg =
2823 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2824 cfs_rq->runnable_load_sum =
2825 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2828 #ifndef CONFIG_64BIT
2829 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2831 u64 last_update_time_copy;
2832 u64 last_update_time;
2835 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2837 last_update_time = cfs_rq->avg.last_update_time;
2838 } while (last_update_time != last_update_time_copy);
2840 return last_update_time;
2843 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2845 return cfs_rq->avg.last_update_time;
2850 * Task first catches up with cfs_rq, and then subtract
2851 * itself from the cfs_rq (task must be off the queue now).
2853 void remove_entity_load_avg(struct sched_entity *se)
2855 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2856 u64 last_update_time;
2859 * Newly created task or never used group entity should not be removed
2860 * from its (source) cfs_rq
2862 if (se->avg.last_update_time == 0)
2865 last_update_time = cfs_rq_last_update_time(cfs_rq);
2867 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2868 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2869 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2873 * Update the rq's load with the elapsed running time before entering
2874 * idle. if the last scheduled task is not a CFS task, idle_enter will
2875 * be the only way to update the runnable statistic.
2877 void idle_enter_fair(struct rq *this_rq)
2882 * Update the rq's load with the elapsed idle time before a task is
2883 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2884 * be the only way to update the runnable statistic.
2886 void idle_exit_fair(struct rq *this_rq)
2890 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2892 return cfs_rq->runnable_load_avg;
2895 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2897 return cfs_rq->avg.load_avg;
2900 static int idle_balance(struct rq *this_rq);
2902 #else /* CONFIG_SMP */
2904 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2906 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2908 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2909 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2912 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2914 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2916 static inline int idle_balance(struct rq *rq)
2921 #endif /* CONFIG_SMP */
2923 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2925 #ifdef CONFIG_SCHEDSTATS
2926 struct task_struct *tsk = NULL;
2928 if (entity_is_task(se))
2931 if (se->statistics.sleep_start) {
2932 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2937 if (unlikely(delta > se->statistics.sleep_max))
2938 se->statistics.sleep_max = delta;
2940 se->statistics.sleep_start = 0;
2941 se->statistics.sum_sleep_runtime += delta;
2944 account_scheduler_latency(tsk, delta >> 10, 1);
2945 trace_sched_stat_sleep(tsk, delta);
2948 if (se->statistics.block_start) {
2949 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2954 if (unlikely(delta > se->statistics.block_max))
2955 se->statistics.block_max = delta;
2957 se->statistics.block_start = 0;
2958 se->statistics.sum_sleep_runtime += delta;
2961 if (tsk->in_iowait) {
2962 se->statistics.iowait_sum += delta;
2963 se->statistics.iowait_count++;
2964 trace_sched_stat_iowait(tsk, delta);
2967 trace_sched_stat_blocked(tsk, delta);
2968 trace_sched_blocked_reason(tsk);
2971 * Blocking time is in units of nanosecs, so shift by
2972 * 20 to get a milliseconds-range estimation of the
2973 * amount of time that the task spent sleeping:
2975 if (unlikely(prof_on == SLEEP_PROFILING)) {
2976 profile_hits(SLEEP_PROFILING,
2977 (void *)get_wchan(tsk),
2980 account_scheduler_latency(tsk, delta >> 10, 0);
2986 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2988 #ifdef CONFIG_SCHED_DEBUG
2989 s64 d = se->vruntime - cfs_rq->min_vruntime;
2994 if (d > 3*sysctl_sched_latency)
2995 schedstat_inc(cfs_rq, nr_spread_over);
3000 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3002 u64 vruntime = cfs_rq->min_vruntime;
3005 * The 'current' period is already promised to the current tasks,
3006 * however the extra weight of the new task will slow them down a
3007 * little, place the new task so that it fits in the slot that
3008 * stays open at the end.
3010 if (initial && sched_feat(START_DEBIT))
3011 vruntime += sched_vslice(cfs_rq, se);
3013 /* sleeps up to a single latency don't count. */
3015 unsigned long thresh = sysctl_sched_latency;
3018 * Halve their sleep time's effect, to allow
3019 * for a gentler effect of sleepers:
3021 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3027 /* ensure we never gain time by being placed backwards. */
3028 se->vruntime = max_vruntime(se->vruntime, vruntime);
3031 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3034 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3037 * Update the normalized vruntime before updating min_vruntime
3038 * through calling update_curr().
3040 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3041 se->vruntime += cfs_rq->min_vruntime;
3044 * Update run-time statistics of the 'current'.
3046 update_curr(cfs_rq);
3047 enqueue_entity_load_avg(cfs_rq, se);
3048 account_entity_enqueue(cfs_rq, se);
3049 update_cfs_shares(cfs_rq);
3051 if (flags & ENQUEUE_WAKEUP) {
3052 place_entity(cfs_rq, se, 0);
3053 enqueue_sleeper(cfs_rq, se);
3056 update_stats_enqueue(cfs_rq, se);
3057 check_spread(cfs_rq, se);
3058 if (se != cfs_rq->curr)
3059 __enqueue_entity(cfs_rq, se);
3062 if (cfs_rq->nr_running == 1) {
3063 list_add_leaf_cfs_rq(cfs_rq);
3064 check_enqueue_throttle(cfs_rq);
3068 static void __clear_buddies_last(struct sched_entity *se)
3070 for_each_sched_entity(se) {
3071 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3072 if (cfs_rq->last != se)
3075 cfs_rq->last = NULL;
3079 static void __clear_buddies_next(struct sched_entity *se)
3081 for_each_sched_entity(se) {
3082 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3083 if (cfs_rq->next != se)
3086 cfs_rq->next = NULL;
3090 static void __clear_buddies_skip(struct sched_entity *se)
3092 for_each_sched_entity(se) {
3093 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3094 if (cfs_rq->skip != se)
3097 cfs_rq->skip = NULL;
3101 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3103 if (cfs_rq->last == se)
3104 __clear_buddies_last(se);
3106 if (cfs_rq->next == se)
3107 __clear_buddies_next(se);
3109 if (cfs_rq->skip == se)
3110 __clear_buddies_skip(se);
3113 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3116 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3119 * Update run-time statistics of the 'current'.
3121 update_curr(cfs_rq);
3122 dequeue_entity_load_avg(cfs_rq, se);
3124 update_stats_dequeue(cfs_rq, se);
3125 if (flags & DEQUEUE_SLEEP) {
3126 #ifdef CONFIG_SCHEDSTATS
3127 if (entity_is_task(se)) {
3128 struct task_struct *tsk = task_of(se);
3130 if (tsk->state & TASK_INTERRUPTIBLE)
3131 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3132 if (tsk->state & TASK_UNINTERRUPTIBLE)
3133 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3138 clear_buddies(cfs_rq, se);
3140 if (se != cfs_rq->curr)
3141 __dequeue_entity(cfs_rq, se);
3143 account_entity_dequeue(cfs_rq, se);
3146 * Normalize the entity after updating the min_vruntime because the
3147 * update can refer to the ->curr item and we need to reflect this
3148 * movement in our normalized position.
3150 if (!(flags & DEQUEUE_SLEEP))
3151 se->vruntime -= cfs_rq->min_vruntime;
3153 /* return excess runtime on last dequeue */
3154 return_cfs_rq_runtime(cfs_rq);
3156 update_min_vruntime(cfs_rq);
3157 update_cfs_shares(cfs_rq);
3161 * Preempt the current task with a newly woken task if needed:
3164 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3166 unsigned long ideal_runtime, delta_exec;
3167 struct sched_entity *se;
3170 ideal_runtime = sched_slice(cfs_rq, curr);
3171 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3172 if (delta_exec > ideal_runtime) {
3173 resched_curr(rq_of(cfs_rq));
3175 * The current task ran long enough, ensure it doesn't get
3176 * re-elected due to buddy favours.
3178 clear_buddies(cfs_rq, curr);
3183 * Ensure that a task that missed wakeup preemption by a
3184 * narrow margin doesn't have to wait for a full slice.
3185 * This also mitigates buddy induced latencies under load.
3187 if (delta_exec < sysctl_sched_min_granularity)
3190 se = __pick_first_entity(cfs_rq);
3191 delta = curr->vruntime - se->vruntime;
3196 if (delta > ideal_runtime)
3197 resched_curr(rq_of(cfs_rq));
3201 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3203 /* 'current' is not kept within the tree. */
3206 * Any task has to be enqueued before it get to execute on
3207 * a CPU. So account for the time it spent waiting on the
3210 update_stats_wait_end(cfs_rq, se);
3211 __dequeue_entity(cfs_rq, se);
3212 update_load_avg(se, 1);
3215 update_stats_curr_start(cfs_rq, se);
3217 #ifdef CONFIG_SCHEDSTATS
3219 * Track our maximum slice length, if the CPU's load is at
3220 * least twice that of our own weight (i.e. dont track it
3221 * when there are only lesser-weight tasks around):
3223 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3224 se->statistics.slice_max = max(se->statistics.slice_max,
3225 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3228 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3232 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3235 * Pick the next process, keeping these things in mind, in this order:
3236 * 1) keep things fair between processes/task groups
3237 * 2) pick the "next" process, since someone really wants that to run
3238 * 3) pick the "last" process, for cache locality
3239 * 4) do not run the "skip" process, if something else is available
3241 static struct sched_entity *
3242 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3244 struct sched_entity *left = __pick_first_entity(cfs_rq);
3245 struct sched_entity *se;
3248 * If curr is set we have to see if its left of the leftmost entity
3249 * still in the tree, provided there was anything in the tree at all.
3251 if (!left || (curr && entity_before(curr, left)))
3254 se = left; /* ideally we run the leftmost entity */
3257 * Avoid running the skip buddy, if running something else can
3258 * be done without getting too unfair.
3260 if (cfs_rq->skip == se) {
3261 struct sched_entity *second;
3264 second = __pick_first_entity(cfs_rq);
3266 second = __pick_next_entity(se);
3267 if (!second || (curr && entity_before(curr, second)))
3271 if (second && wakeup_preempt_entity(second, left) < 1)
3276 * Prefer last buddy, try to return the CPU to a preempted task.
3278 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3282 * Someone really wants this to run. If it's not unfair, run it.
3284 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3287 clear_buddies(cfs_rq, se);
3292 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3294 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3297 * If still on the runqueue then deactivate_task()
3298 * was not called and update_curr() has to be done:
3301 update_curr(cfs_rq);
3303 /* throttle cfs_rqs exceeding runtime */
3304 check_cfs_rq_runtime(cfs_rq);
3306 check_spread(cfs_rq, prev);
3308 update_stats_wait_start(cfs_rq, prev);
3309 /* Put 'current' back into the tree. */
3310 __enqueue_entity(cfs_rq, prev);
3311 /* in !on_rq case, update occurred at dequeue */
3312 update_load_avg(prev, 0);
3314 cfs_rq->curr = NULL;
3318 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3321 * Update run-time statistics of the 'current'.
3323 update_curr(cfs_rq);
3326 * Ensure that runnable average is periodically updated.
3328 update_load_avg(curr, 1);
3329 update_cfs_shares(cfs_rq);
3331 #ifdef CONFIG_SCHED_HRTICK
3333 * queued ticks are scheduled to match the slice, so don't bother
3334 * validating it and just reschedule.
3337 resched_curr(rq_of(cfs_rq));
3341 * don't let the period tick interfere with the hrtick preemption
3343 if (!sched_feat(DOUBLE_TICK) &&
3344 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3348 if (cfs_rq->nr_running > 1)
3349 check_preempt_tick(cfs_rq, curr);
3353 /**************************************************
3354 * CFS bandwidth control machinery
3357 #ifdef CONFIG_CFS_BANDWIDTH
3359 #ifdef HAVE_JUMP_LABEL
3360 static struct static_key __cfs_bandwidth_used;
3362 static inline bool cfs_bandwidth_used(void)
3364 return static_key_false(&__cfs_bandwidth_used);
3367 void cfs_bandwidth_usage_inc(void)
3369 static_key_slow_inc(&__cfs_bandwidth_used);
3372 void cfs_bandwidth_usage_dec(void)
3374 static_key_slow_dec(&__cfs_bandwidth_used);
3376 #else /* HAVE_JUMP_LABEL */
3377 static bool cfs_bandwidth_used(void)
3382 void cfs_bandwidth_usage_inc(void) {}
3383 void cfs_bandwidth_usage_dec(void) {}
3384 #endif /* HAVE_JUMP_LABEL */
3387 * default period for cfs group bandwidth.
3388 * default: 0.1s, units: nanoseconds
3390 static inline u64 default_cfs_period(void)
3392 return 100000000ULL;
3395 static inline u64 sched_cfs_bandwidth_slice(void)
3397 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3401 * Replenish runtime according to assigned quota and update expiration time.
3402 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3403 * additional synchronization around rq->lock.
3405 * requires cfs_b->lock
3407 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3411 if (cfs_b->quota == RUNTIME_INF)
3414 now = sched_clock_cpu(smp_processor_id());
3415 cfs_b->runtime = cfs_b->quota;
3416 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3419 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3421 return &tg->cfs_bandwidth;
3424 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3425 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3427 if (unlikely(cfs_rq->throttle_count))
3428 return cfs_rq->throttled_clock_task;
3430 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3433 /* returns 0 on failure to allocate runtime */
3434 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3436 struct task_group *tg = cfs_rq->tg;
3437 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3438 u64 amount = 0, min_amount, expires;
3440 /* note: this is a positive sum as runtime_remaining <= 0 */
3441 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3443 raw_spin_lock(&cfs_b->lock);
3444 if (cfs_b->quota == RUNTIME_INF)
3445 amount = min_amount;
3447 start_cfs_bandwidth(cfs_b);
3449 if (cfs_b->runtime > 0) {
3450 amount = min(cfs_b->runtime, min_amount);
3451 cfs_b->runtime -= amount;
3455 expires = cfs_b->runtime_expires;
3456 raw_spin_unlock(&cfs_b->lock);
3458 cfs_rq->runtime_remaining += amount;
3460 * we may have advanced our local expiration to account for allowed
3461 * spread between our sched_clock and the one on which runtime was
3464 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3465 cfs_rq->runtime_expires = expires;
3467 return cfs_rq->runtime_remaining > 0;
3471 * Note: This depends on the synchronization provided by sched_clock and the
3472 * fact that rq->clock snapshots this value.
3474 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3476 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3478 /* if the deadline is ahead of our clock, nothing to do */
3479 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3482 if (cfs_rq->runtime_remaining < 0)
3486 * If the local deadline has passed we have to consider the
3487 * possibility that our sched_clock is 'fast' and the global deadline
3488 * has not truly expired.
3490 * Fortunately we can check determine whether this the case by checking
3491 * whether the global deadline has advanced. It is valid to compare
3492 * cfs_b->runtime_expires without any locks since we only care about
3493 * exact equality, so a partial write will still work.
3496 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3497 /* extend local deadline, drift is bounded above by 2 ticks */
3498 cfs_rq->runtime_expires += TICK_NSEC;
3500 /* global deadline is ahead, expiration has passed */
3501 cfs_rq->runtime_remaining = 0;
3505 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3507 /* dock delta_exec before expiring quota (as it could span periods) */
3508 cfs_rq->runtime_remaining -= delta_exec;
3509 expire_cfs_rq_runtime(cfs_rq);
3511 if (likely(cfs_rq->runtime_remaining > 0))
3515 * if we're unable to extend our runtime we resched so that the active
3516 * hierarchy can be throttled
3518 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3519 resched_curr(rq_of(cfs_rq));
3522 static __always_inline
3523 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3525 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3528 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3531 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3533 return cfs_bandwidth_used() && cfs_rq->throttled;
3536 /* check whether cfs_rq, or any parent, is throttled */
3537 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3539 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3543 * Ensure that neither of the group entities corresponding to src_cpu or
3544 * dest_cpu are members of a throttled hierarchy when performing group
3545 * load-balance operations.
3547 static inline int throttled_lb_pair(struct task_group *tg,
3548 int src_cpu, int dest_cpu)
3550 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3552 src_cfs_rq = tg->cfs_rq[src_cpu];
3553 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3555 return throttled_hierarchy(src_cfs_rq) ||
3556 throttled_hierarchy(dest_cfs_rq);
3559 /* updated child weight may affect parent so we have to do this bottom up */
3560 static int tg_unthrottle_up(struct task_group *tg, void *data)
3562 struct rq *rq = data;
3563 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3565 cfs_rq->throttle_count--;
3567 if (!cfs_rq->throttle_count) {
3568 /* adjust cfs_rq_clock_task() */
3569 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3570 cfs_rq->throttled_clock_task;
3577 static int tg_throttle_down(struct task_group *tg, void *data)
3579 struct rq *rq = data;
3580 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3582 /* group is entering throttled state, stop time */
3583 if (!cfs_rq->throttle_count)
3584 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3585 cfs_rq->throttle_count++;
3590 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3592 struct rq *rq = rq_of(cfs_rq);
3593 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3594 struct sched_entity *se;
3595 long task_delta, dequeue = 1;
3598 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3600 /* freeze hierarchy runnable averages while throttled */
3602 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3605 task_delta = cfs_rq->h_nr_running;
3606 for_each_sched_entity(se) {
3607 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3608 /* throttled entity or throttle-on-deactivate */
3613 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3614 qcfs_rq->h_nr_running -= task_delta;
3616 if (qcfs_rq->load.weight)
3621 sub_nr_running(rq, task_delta);
3623 cfs_rq->throttled = 1;
3624 cfs_rq->throttled_clock = rq_clock(rq);
3625 raw_spin_lock(&cfs_b->lock);
3626 empty = list_empty(&cfs_b->throttled_cfs_rq);
3629 * Add to the _head_ of the list, so that an already-started
3630 * distribute_cfs_runtime will not see us
3632 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3635 * If we're the first throttled task, make sure the bandwidth
3639 start_cfs_bandwidth(cfs_b);
3641 raw_spin_unlock(&cfs_b->lock);
3644 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3646 struct rq *rq = rq_of(cfs_rq);
3647 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3648 struct sched_entity *se;
3652 se = cfs_rq->tg->se[cpu_of(rq)];
3654 cfs_rq->throttled = 0;
3656 update_rq_clock(rq);
3658 raw_spin_lock(&cfs_b->lock);
3659 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3660 list_del_rcu(&cfs_rq->throttled_list);
3661 raw_spin_unlock(&cfs_b->lock);
3663 /* update hierarchical throttle state */
3664 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3666 if (!cfs_rq->load.weight)
3669 task_delta = cfs_rq->h_nr_running;
3670 for_each_sched_entity(se) {
3674 cfs_rq = cfs_rq_of(se);
3676 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3677 cfs_rq->h_nr_running += task_delta;
3679 if (cfs_rq_throttled(cfs_rq))
3684 add_nr_running(rq, task_delta);
3686 /* determine whether we need to wake up potentially idle cpu */
3687 if (rq->curr == rq->idle && rq->cfs.nr_running)
3691 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3692 u64 remaining, u64 expires)
3694 struct cfs_rq *cfs_rq;
3696 u64 starting_runtime = remaining;
3699 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3701 struct rq *rq = rq_of(cfs_rq);
3703 raw_spin_lock(&rq->lock);
3704 if (!cfs_rq_throttled(cfs_rq))
3707 runtime = -cfs_rq->runtime_remaining + 1;
3708 if (runtime > remaining)
3709 runtime = remaining;
3710 remaining -= runtime;
3712 cfs_rq->runtime_remaining += runtime;
3713 cfs_rq->runtime_expires = expires;
3715 /* we check whether we're throttled above */
3716 if (cfs_rq->runtime_remaining > 0)
3717 unthrottle_cfs_rq(cfs_rq);
3720 raw_spin_unlock(&rq->lock);
3727 return starting_runtime - remaining;
3731 * Responsible for refilling a task_group's bandwidth and unthrottling its
3732 * cfs_rqs as appropriate. If there has been no activity within the last
3733 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3734 * used to track this state.
3736 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3738 u64 runtime, runtime_expires;
3741 /* no need to continue the timer with no bandwidth constraint */
3742 if (cfs_b->quota == RUNTIME_INF)
3743 goto out_deactivate;
3745 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3746 cfs_b->nr_periods += overrun;
3749 * idle depends on !throttled (for the case of a large deficit), and if
3750 * we're going inactive then everything else can be deferred
3752 if (cfs_b->idle && !throttled)
3753 goto out_deactivate;
3755 __refill_cfs_bandwidth_runtime(cfs_b);
3758 /* mark as potentially idle for the upcoming period */
3763 /* account preceding periods in which throttling occurred */
3764 cfs_b->nr_throttled += overrun;
3766 runtime_expires = cfs_b->runtime_expires;
3769 * This check is repeated as we are holding onto the new bandwidth while
3770 * we unthrottle. This can potentially race with an unthrottled group
3771 * trying to acquire new bandwidth from the global pool. This can result
3772 * in us over-using our runtime if it is all used during this loop, but
3773 * only by limited amounts in that extreme case.
3775 while (throttled && cfs_b->runtime > 0) {
3776 runtime = cfs_b->runtime;
3777 raw_spin_unlock(&cfs_b->lock);
3778 /* we can't nest cfs_b->lock while distributing bandwidth */
3779 runtime = distribute_cfs_runtime(cfs_b, runtime,
3781 raw_spin_lock(&cfs_b->lock);
3783 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3785 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3789 * While we are ensured activity in the period following an
3790 * unthrottle, this also covers the case in which the new bandwidth is
3791 * insufficient to cover the existing bandwidth deficit. (Forcing the
3792 * timer to remain active while there are any throttled entities.)
3802 /* a cfs_rq won't donate quota below this amount */
3803 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3804 /* minimum remaining period time to redistribute slack quota */
3805 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3806 /* how long we wait to gather additional slack before distributing */
3807 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3810 * Are we near the end of the current quota period?
3812 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3813 * hrtimer base being cleared by hrtimer_start. In the case of
3814 * migrate_hrtimers, base is never cleared, so we are fine.
3816 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3818 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3821 /* if the call-back is running a quota refresh is already occurring */
3822 if (hrtimer_callback_running(refresh_timer))
3825 /* is a quota refresh about to occur? */
3826 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3827 if (remaining < min_expire)
3833 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3835 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3837 /* if there's a quota refresh soon don't bother with slack */
3838 if (runtime_refresh_within(cfs_b, min_left))
3841 hrtimer_start(&cfs_b->slack_timer,
3842 ns_to_ktime(cfs_bandwidth_slack_period),
3846 /* we know any runtime found here is valid as update_curr() precedes return */
3847 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3849 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3850 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3852 if (slack_runtime <= 0)
3855 raw_spin_lock(&cfs_b->lock);
3856 if (cfs_b->quota != RUNTIME_INF &&
3857 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3858 cfs_b->runtime += slack_runtime;
3860 /* we are under rq->lock, defer unthrottling using a timer */
3861 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3862 !list_empty(&cfs_b->throttled_cfs_rq))
3863 start_cfs_slack_bandwidth(cfs_b);
3865 raw_spin_unlock(&cfs_b->lock);
3867 /* even if it's not valid for return we don't want to try again */
3868 cfs_rq->runtime_remaining -= slack_runtime;
3871 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3873 if (!cfs_bandwidth_used())
3876 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3879 __return_cfs_rq_runtime(cfs_rq);
3883 * This is done with a timer (instead of inline with bandwidth return) since
3884 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3886 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3888 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3891 /* confirm we're still not at a refresh boundary */
3892 raw_spin_lock(&cfs_b->lock);
3893 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3894 raw_spin_unlock(&cfs_b->lock);
3898 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3899 runtime = cfs_b->runtime;
3901 expires = cfs_b->runtime_expires;
3902 raw_spin_unlock(&cfs_b->lock);
3907 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3909 raw_spin_lock(&cfs_b->lock);
3910 if (expires == cfs_b->runtime_expires)
3911 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3912 raw_spin_unlock(&cfs_b->lock);
3916 * When a group wakes up we want to make sure that its quota is not already
3917 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3918 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3920 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3922 if (!cfs_bandwidth_used())
3925 /* an active group must be handled by the update_curr()->put() path */
3926 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3929 /* ensure the group is not already throttled */
3930 if (cfs_rq_throttled(cfs_rq))
3933 /* update runtime allocation */
3934 account_cfs_rq_runtime(cfs_rq, 0);
3935 if (cfs_rq->runtime_remaining <= 0)
3936 throttle_cfs_rq(cfs_rq);
3939 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3940 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3942 if (!cfs_bandwidth_used())
3945 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3949 * it's possible for a throttled entity to be forced into a running
3950 * state (e.g. set_curr_task), in this case we're finished.
3952 if (cfs_rq_throttled(cfs_rq))
3955 throttle_cfs_rq(cfs_rq);
3959 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3961 struct cfs_bandwidth *cfs_b =
3962 container_of(timer, struct cfs_bandwidth, slack_timer);
3964 do_sched_cfs_slack_timer(cfs_b);
3966 return HRTIMER_NORESTART;
3969 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3971 struct cfs_bandwidth *cfs_b =
3972 container_of(timer, struct cfs_bandwidth, period_timer);
3976 raw_spin_lock(&cfs_b->lock);
3978 overrun = hrtimer_forward_now(timer, cfs_b->period);
3982 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3985 cfs_b->period_active = 0;
3986 raw_spin_unlock(&cfs_b->lock);
3988 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3991 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3993 raw_spin_lock_init(&cfs_b->lock);
3995 cfs_b->quota = RUNTIME_INF;
3996 cfs_b->period = ns_to_ktime(default_cfs_period());
3998 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3999 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4000 cfs_b->period_timer.function = sched_cfs_period_timer;
4001 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4002 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4005 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4007 cfs_rq->runtime_enabled = 0;
4008 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4011 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4013 lockdep_assert_held(&cfs_b->lock);
4015 if (!cfs_b->period_active) {
4016 cfs_b->period_active = 1;
4017 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4018 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4022 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4024 /* init_cfs_bandwidth() was not called */
4025 if (!cfs_b->throttled_cfs_rq.next)
4028 hrtimer_cancel(&cfs_b->period_timer);
4029 hrtimer_cancel(&cfs_b->slack_timer);
4032 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4034 struct cfs_rq *cfs_rq;
4036 for_each_leaf_cfs_rq(rq, cfs_rq) {
4037 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4039 raw_spin_lock(&cfs_b->lock);
4040 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4041 raw_spin_unlock(&cfs_b->lock);
4045 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4047 struct cfs_rq *cfs_rq;
4049 for_each_leaf_cfs_rq(rq, cfs_rq) {
4050 if (!cfs_rq->runtime_enabled)
4054 * clock_task is not advancing so we just need to make sure
4055 * there's some valid quota amount
4057 cfs_rq->runtime_remaining = 1;
4059 * Offline rq is schedulable till cpu is completely disabled
4060 * in take_cpu_down(), so we prevent new cfs throttling here.
4062 cfs_rq->runtime_enabled = 0;
4064 if (cfs_rq_throttled(cfs_rq))
4065 unthrottle_cfs_rq(cfs_rq);
4069 #else /* CONFIG_CFS_BANDWIDTH */
4070 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4072 return rq_clock_task(rq_of(cfs_rq));
4075 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4076 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4077 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4078 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4080 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4085 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4090 static inline int throttled_lb_pair(struct task_group *tg,
4091 int src_cpu, int dest_cpu)
4096 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4098 #ifdef CONFIG_FAIR_GROUP_SCHED
4099 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4102 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4106 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4107 static inline void update_runtime_enabled(struct rq *rq) {}
4108 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4110 #endif /* CONFIG_CFS_BANDWIDTH */
4112 /**************************************************
4113 * CFS operations on tasks:
4116 #ifdef CONFIG_SCHED_HRTICK
4117 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4119 struct sched_entity *se = &p->se;
4120 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4122 WARN_ON(task_rq(p) != rq);
4124 if (cfs_rq->nr_running > 1) {
4125 u64 slice = sched_slice(cfs_rq, se);
4126 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4127 s64 delta = slice - ran;
4134 hrtick_start(rq, delta);
4139 * called from enqueue/dequeue and updates the hrtick when the
4140 * current task is from our class and nr_running is low enough
4143 static void hrtick_update(struct rq *rq)
4145 struct task_struct *curr = rq->curr;
4147 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4150 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4151 hrtick_start_fair(rq, curr);
4153 #else /* !CONFIG_SCHED_HRTICK */
4155 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4159 static inline void hrtick_update(struct rq *rq)
4164 static inline unsigned long boosted_cpu_util(int cpu);
4166 static void update_capacity_of(int cpu)
4168 unsigned long req_cap;
4173 /* Convert scale-invariant capacity to cpu. */
4174 req_cap = boosted_cpu_util(cpu);
4175 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4176 set_cfs_cpu_capacity(cpu, true, req_cap);
4179 static bool cpu_overutilized(int cpu);
4182 * The enqueue_task method is called before nr_running is
4183 * increased. Here we update the fair scheduling stats and
4184 * then put the task into the rbtree:
4187 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4189 struct cfs_rq *cfs_rq;
4190 struct sched_entity *se = &p->se;
4191 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4192 int task_wakeup = flags & ENQUEUE_WAKEUP;
4194 for_each_sched_entity(se) {
4197 cfs_rq = cfs_rq_of(se);
4198 enqueue_entity(cfs_rq, se, flags);
4201 * end evaluation on encountering a throttled cfs_rq
4203 * note: in the case of encountering a throttled cfs_rq we will
4204 * post the final h_nr_running increment below.
4206 if (cfs_rq_throttled(cfs_rq))
4208 cfs_rq->h_nr_running++;
4210 flags = ENQUEUE_WAKEUP;
4213 for_each_sched_entity(se) {
4214 cfs_rq = cfs_rq_of(se);
4215 cfs_rq->h_nr_running++;
4217 if (cfs_rq_throttled(cfs_rq))
4220 update_load_avg(se, 1);
4221 update_cfs_shares(cfs_rq);
4225 add_nr_running(rq, 1);
4226 if (!task_new && !rq->rd->overutilized &&
4227 cpu_overutilized(rq->cpu))
4228 rq->rd->overutilized = true;
4230 schedtune_enqueue_task(p, cpu_of(rq));
4233 * We want to potentially trigger a freq switch
4234 * request only for tasks that are waking up; this is
4235 * because we get here also during load balancing, but
4236 * in these cases it seems wise to trigger as single
4237 * request after load balancing is done.
4239 if (task_new || task_wakeup)
4240 update_capacity_of(cpu_of(rq));
4245 static void set_next_buddy(struct sched_entity *se);
4248 * The dequeue_task method is called before nr_running is
4249 * decreased. We remove the task from the rbtree and
4250 * update the fair scheduling stats:
4252 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4254 struct cfs_rq *cfs_rq;
4255 struct sched_entity *se = &p->se;
4256 int task_sleep = flags & DEQUEUE_SLEEP;
4258 for_each_sched_entity(se) {
4259 cfs_rq = cfs_rq_of(se);
4260 dequeue_entity(cfs_rq, se, flags);
4263 * end evaluation on encountering a throttled cfs_rq
4265 * note: in the case of encountering a throttled cfs_rq we will
4266 * post the final h_nr_running decrement below.
4268 if (cfs_rq_throttled(cfs_rq))
4270 cfs_rq->h_nr_running--;
4272 /* Don't dequeue parent if it has other entities besides us */
4273 if (cfs_rq->load.weight) {
4275 * Bias pick_next to pick a task from this cfs_rq, as
4276 * p is sleeping when it is within its sched_slice.
4278 if (task_sleep && parent_entity(se))
4279 set_next_buddy(parent_entity(se));
4281 /* avoid re-evaluating load for this entity */
4282 se = parent_entity(se);
4285 flags |= DEQUEUE_SLEEP;
4288 for_each_sched_entity(se) {
4289 cfs_rq = cfs_rq_of(se);
4290 cfs_rq->h_nr_running--;
4292 if (cfs_rq_throttled(cfs_rq))
4295 update_load_avg(se, 1);
4296 update_cfs_shares(cfs_rq);
4300 sub_nr_running(rq, 1);
4301 schedtune_dequeue_task(p, cpu_of(rq));
4304 * We want to potentially trigger a freq switch
4305 * request only for tasks that are going to sleep;
4306 * this is because we get here also during load
4307 * balancing, but in these cases it seems wise to
4308 * trigger as single request after load balancing is
4312 if (rq->cfs.nr_running)
4313 update_capacity_of(cpu_of(rq));
4314 else if (sched_freq())
4315 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4324 * per rq 'load' arrray crap; XXX kill this.
4328 * The exact cpuload at various idx values, calculated at every tick would be
4329 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4331 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4332 * on nth tick when cpu may be busy, then we have:
4333 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4334 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4336 * decay_load_missed() below does efficient calculation of
4337 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4338 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4340 * The calculation is approximated on a 128 point scale.
4341 * degrade_zero_ticks is the number of ticks after which load at any
4342 * particular idx is approximated to be zero.
4343 * degrade_factor is a precomputed table, a row for each load idx.
4344 * Each column corresponds to degradation factor for a power of two ticks,
4345 * based on 128 point scale.
4347 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4348 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4350 * With this power of 2 load factors, we can degrade the load n times
4351 * by looking at 1 bits in n and doing as many mult/shift instead of
4352 * n mult/shifts needed by the exact degradation.
4354 #define DEGRADE_SHIFT 7
4355 static const unsigned char
4356 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4357 static const unsigned char
4358 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4359 {0, 0, 0, 0, 0, 0, 0, 0},
4360 {64, 32, 8, 0, 0, 0, 0, 0},
4361 {96, 72, 40, 12, 1, 0, 0},
4362 {112, 98, 75, 43, 15, 1, 0},
4363 {120, 112, 98, 76, 45, 16, 2} };
4366 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4367 * would be when CPU is idle and so we just decay the old load without
4368 * adding any new load.
4370 static unsigned long
4371 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4375 if (!missed_updates)
4378 if (missed_updates >= degrade_zero_ticks[idx])
4382 return load >> missed_updates;
4384 while (missed_updates) {
4385 if (missed_updates % 2)
4386 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4388 missed_updates >>= 1;
4395 * Update rq->cpu_load[] statistics. This function is usually called every
4396 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4397 * every tick. We fix it up based on jiffies.
4399 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4400 unsigned long pending_updates)
4404 this_rq->nr_load_updates++;
4406 /* Update our load: */
4407 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4408 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4409 unsigned long old_load, new_load;
4411 /* scale is effectively 1 << i now, and >> i divides by scale */
4413 old_load = this_rq->cpu_load[i];
4414 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4415 new_load = this_load;
4417 * Round up the averaging division if load is increasing. This
4418 * prevents us from getting stuck on 9 if the load is 10, for
4421 if (new_load > old_load)
4422 new_load += scale - 1;
4424 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4427 sched_avg_update(this_rq);
4430 /* Used instead of source_load when we know the type == 0 */
4431 static unsigned long weighted_cpuload(const int cpu)
4433 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4436 #ifdef CONFIG_NO_HZ_COMMON
4438 * There is no sane way to deal with nohz on smp when using jiffies because the
4439 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4440 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4442 * Therefore we cannot use the delta approach from the regular tick since that
4443 * would seriously skew the load calculation. However we'll make do for those
4444 * updates happening while idle (nohz_idle_balance) or coming out of idle
4445 * (tick_nohz_idle_exit).
4447 * This means we might still be one tick off for nohz periods.
4451 * Called from nohz_idle_balance() to update the load ratings before doing the
4454 static void update_idle_cpu_load(struct rq *this_rq)
4456 unsigned long curr_jiffies = READ_ONCE(jiffies);
4457 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4458 unsigned long pending_updates;
4461 * bail if there's load or we're actually up-to-date.
4463 if (load || curr_jiffies == this_rq->last_load_update_tick)
4466 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4467 this_rq->last_load_update_tick = curr_jiffies;
4469 __update_cpu_load(this_rq, load, pending_updates);
4473 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4475 void update_cpu_load_nohz(void)
4477 struct rq *this_rq = this_rq();
4478 unsigned long curr_jiffies = READ_ONCE(jiffies);
4479 unsigned long pending_updates;
4481 if (curr_jiffies == this_rq->last_load_update_tick)
4484 raw_spin_lock(&this_rq->lock);
4485 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4486 if (pending_updates) {
4487 this_rq->last_load_update_tick = curr_jiffies;
4489 * We were idle, this means load 0, the current load might be
4490 * !0 due to remote wakeups and the sort.
4492 __update_cpu_load(this_rq, 0, pending_updates);
4494 raw_spin_unlock(&this_rq->lock);
4496 #endif /* CONFIG_NO_HZ */
4499 * Called from scheduler_tick()
4501 void update_cpu_load_active(struct rq *this_rq)
4503 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4505 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4507 this_rq->last_load_update_tick = jiffies;
4508 __update_cpu_load(this_rq, load, 1);
4512 * Return a low guess at the load of a migration-source cpu weighted
4513 * according to the scheduling class and "nice" value.
4515 * We want to under-estimate the load of migration sources, to
4516 * balance conservatively.
4518 static unsigned long source_load(int cpu, int type)
4520 struct rq *rq = cpu_rq(cpu);
4521 unsigned long total = weighted_cpuload(cpu);
4523 if (type == 0 || !sched_feat(LB_BIAS))
4526 return min(rq->cpu_load[type-1], total);
4530 * Return a high guess at the load of a migration-target cpu weighted
4531 * according to the scheduling class and "nice" value.
4533 static unsigned long target_load(int cpu, int type)
4535 struct rq *rq = cpu_rq(cpu);
4536 unsigned long total = weighted_cpuload(cpu);
4538 if (type == 0 || !sched_feat(LB_BIAS))
4541 return max(rq->cpu_load[type-1], total);
4545 static unsigned long cpu_avg_load_per_task(int cpu)
4547 struct rq *rq = cpu_rq(cpu);
4548 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4549 unsigned long load_avg = weighted_cpuload(cpu);
4552 return load_avg / nr_running;
4557 static void record_wakee(struct task_struct *p)
4560 * Rough decay (wiping) for cost saving, don't worry
4561 * about the boundary, really active task won't care
4564 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4565 current->wakee_flips >>= 1;
4566 current->wakee_flip_decay_ts = jiffies;
4569 if (current->last_wakee != p) {
4570 current->last_wakee = p;
4571 current->wakee_flips++;
4575 static void task_waking_fair(struct task_struct *p)
4577 struct sched_entity *se = &p->se;
4578 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4581 #ifndef CONFIG_64BIT
4582 u64 min_vruntime_copy;
4585 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4587 min_vruntime = cfs_rq->min_vruntime;
4588 } while (min_vruntime != min_vruntime_copy);
4590 min_vruntime = cfs_rq->min_vruntime;
4593 se->vruntime -= min_vruntime;
4597 #ifdef CONFIG_FAIR_GROUP_SCHED
4599 * effective_load() calculates the load change as seen from the root_task_group
4601 * Adding load to a group doesn't make a group heavier, but can cause movement
4602 * of group shares between cpus. Assuming the shares were perfectly aligned one
4603 * can calculate the shift in shares.
4605 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4606 * on this @cpu and results in a total addition (subtraction) of @wg to the
4607 * total group weight.
4609 * Given a runqueue weight distribution (rw_i) we can compute a shares
4610 * distribution (s_i) using:
4612 * s_i = rw_i / \Sum rw_j (1)
4614 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4615 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4616 * shares distribution (s_i):
4618 * rw_i = { 2, 4, 1, 0 }
4619 * s_i = { 2/7, 4/7, 1/7, 0 }
4621 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4622 * task used to run on and the CPU the waker is running on), we need to
4623 * compute the effect of waking a task on either CPU and, in case of a sync
4624 * wakeup, compute the effect of the current task going to sleep.
4626 * So for a change of @wl to the local @cpu with an overall group weight change
4627 * of @wl we can compute the new shares distribution (s'_i) using:
4629 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4631 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4632 * differences in waking a task to CPU 0. The additional task changes the
4633 * weight and shares distributions like:
4635 * rw'_i = { 3, 4, 1, 0 }
4636 * s'_i = { 3/8, 4/8, 1/8, 0 }
4638 * We can then compute the difference in effective weight by using:
4640 * dw_i = S * (s'_i - s_i) (3)
4642 * Where 'S' is the group weight as seen by its parent.
4644 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4645 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4646 * 4/7) times the weight of the group.
4648 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4650 struct sched_entity *se = tg->se[cpu];
4652 if (!tg->parent) /* the trivial, non-cgroup case */
4655 for_each_sched_entity(se) {
4656 struct cfs_rq *cfs_rq = se->my_q;
4657 long W, w = cfs_rq_load_avg(cfs_rq);
4662 * W = @wg + \Sum rw_j
4664 W = wg + atomic_long_read(&tg->load_avg);
4666 /* Ensure \Sum rw_j >= rw_i */
4667 W -= cfs_rq->tg_load_avg_contrib;
4676 * wl = S * s'_i; see (2)
4679 wl = (w * (long)tg->shares) / W;
4684 * Per the above, wl is the new se->load.weight value; since
4685 * those are clipped to [MIN_SHARES, ...) do so now. See
4686 * calc_cfs_shares().
4688 if (wl < MIN_SHARES)
4692 * wl = dw_i = S * (s'_i - s_i); see (3)
4694 wl -= se->avg.load_avg;
4697 * Recursively apply this logic to all parent groups to compute
4698 * the final effective load change on the root group. Since
4699 * only the @tg group gets extra weight, all parent groups can
4700 * only redistribute existing shares. @wl is the shift in shares
4701 * resulting from this level per the above.
4710 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4717 static inline bool energy_aware(void)
4719 return sched_feat(ENERGY_AWARE);
4723 struct sched_group *sg_top;
4724 struct sched_group *sg_cap;
4733 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4734 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4735 * energy calculations. Using the scale-invariant util returned by
4736 * cpu_util() and approximating scale-invariant util by:
4738 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4740 * the normalized util can be found using the specific capacity.
4742 * capacity = capacity_orig * curr_freq/max_freq
4744 * norm_util = running_time/time ~ util/capacity
4746 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4748 int util = __cpu_util(cpu, delta);
4750 if (util >= capacity)
4751 return SCHED_CAPACITY_SCALE;
4753 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4756 static int calc_util_delta(struct energy_env *eenv, int cpu)
4758 if (cpu == eenv->src_cpu)
4759 return -eenv->util_delta;
4760 if (cpu == eenv->dst_cpu)
4761 return eenv->util_delta;
4766 unsigned long group_max_util(struct energy_env *eenv)
4769 unsigned long max_util = 0;
4771 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4772 delta = calc_util_delta(eenv, i);
4773 max_util = max(max_util, __cpu_util(i, delta));
4780 * group_norm_util() returns the approximated group util relative to it's
4781 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4782 * energy calculations. Since task executions may or may not overlap in time in
4783 * the group the true normalized util is between max(cpu_norm_util(i)) and
4784 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4785 * latter is used as the estimate as it leads to a more pessimistic energy
4786 * estimate (more busy).
4789 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4792 unsigned long util_sum = 0;
4793 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4795 for_each_cpu(i, sched_group_cpus(sg)) {
4796 delta = calc_util_delta(eenv, i);
4797 util_sum += __cpu_norm_util(i, capacity, delta);
4800 if (util_sum > SCHED_CAPACITY_SCALE)
4801 return SCHED_CAPACITY_SCALE;
4805 static int find_new_capacity(struct energy_env *eenv,
4806 const struct sched_group_energy const *sge)
4809 unsigned long util = group_max_util(eenv);
4811 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4812 if (sge->cap_states[idx].cap >= util)
4816 eenv->cap_idx = idx;
4821 static int group_idle_state(struct sched_group *sg)
4823 int i, state = INT_MAX;
4825 /* Find the shallowest idle state in the sched group. */
4826 for_each_cpu(i, sched_group_cpus(sg))
4827 state = min(state, idle_get_state_idx(cpu_rq(i)));
4829 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4836 * sched_group_energy(): Computes the absolute energy consumption of cpus
4837 * belonging to the sched_group including shared resources shared only by
4838 * members of the group. Iterates over all cpus in the hierarchy below the
4839 * sched_group starting from the bottom working it's way up before going to
4840 * the next cpu until all cpus are covered at all levels. The current
4841 * implementation is likely to gather the same util statistics multiple times.
4842 * This can probably be done in a faster but more complex way.
4843 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4845 static int sched_group_energy(struct energy_env *eenv)
4847 struct sched_domain *sd;
4848 int cpu, total_energy = 0;
4849 struct cpumask visit_cpus;
4850 struct sched_group *sg;
4852 WARN_ON(!eenv->sg_top->sge);
4854 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4856 while (!cpumask_empty(&visit_cpus)) {
4857 struct sched_group *sg_shared_cap = NULL;
4859 cpu = cpumask_first(&visit_cpus);
4862 * Is the group utilization affected by cpus outside this
4865 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4869 * We most probably raced with hotplug; returning a
4870 * wrong energy estimation is better than entering an
4876 sg_shared_cap = sd->parent->groups;
4878 for_each_domain(cpu, sd) {
4881 /* Has this sched_domain already been visited? */
4882 if (sd->child && group_first_cpu(sg) != cpu)
4886 unsigned long group_util;
4887 int sg_busy_energy, sg_idle_energy;
4888 int cap_idx, idle_idx;
4890 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4891 eenv->sg_cap = sg_shared_cap;
4895 cap_idx = find_new_capacity(eenv, sg->sge);
4896 idle_idx = group_idle_state(sg);
4897 group_util = group_norm_util(eenv, sg);
4898 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4899 >> SCHED_CAPACITY_SHIFT;
4900 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4901 * sg->sge->idle_states[idle_idx].power)
4902 >> SCHED_CAPACITY_SHIFT;
4904 total_energy += sg_busy_energy + sg_idle_energy;
4907 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4909 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4912 } while (sg = sg->next, sg != sd->groups);
4918 eenv->energy = total_energy;
4922 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4924 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4928 * energy_diff(): Estimate the energy impact of changing the utilization
4929 * distribution. eenv specifies the change: utilisation amount, source, and
4930 * destination cpu. Source or destination cpu may be -1 in which case the
4931 * utilization is removed from or added to the system (e.g. task wake-up). If
4932 * both are specified, the utilization is migrated.
4934 static int energy_diff(struct energy_env *eenv)
4936 struct sched_domain *sd;
4937 struct sched_group *sg;
4938 int sd_cpu = -1, energy_before = 0, energy_after = 0;
4940 struct energy_env eenv_before = {
4942 .src_cpu = eenv->src_cpu,
4943 .dst_cpu = eenv->dst_cpu,
4946 if (eenv->src_cpu == eenv->dst_cpu)
4949 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
4950 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
4953 return 0; /* Error */
4958 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
4959 eenv_before.sg_top = eenv->sg_top = sg;
4961 if (sched_group_energy(&eenv_before))
4962 return 0; /* Invalid result abort */
4963 energy_before += eenv_before.energy;
4965 if (sched_group_energy(eenv))
4966 return 0; /* Invalid result abort */
4967 energy_after += eenv->energy;
4969 } while (sg = sg->next, sg != sd->groups);
4971 return energy_after-energy_before;
4975 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4976 * A waker of many should wake a different task than the one last awakened
4977 * at a frequency roughly N times higher than one of its wakees. In order
4978 * to determine whether we should let the load spread vs consolodating to
4979 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4980 * partner, and a factor of lls_size higher frequency in the other. With
4981 * both conditions met, we can be relatively sure that the relationship is
4982 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4983 * being client/server, worker/dispatcher, interrupt source or whatever is
4984 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4986 static int wake_wide(struct task_struct *p)
4988 unsigned int master = current->wakee_flips;
4989 unsigned int slave = p->wakee_flips;
4990 int factor = this_cpu_read(sd_llc_size);
4993 swap(master, slave);
4994 if (slave < factor || master < slave * factor)
4999 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5001 s64 this_load, load;
5002 s64 this_eff_load, prev_eff_load;
5003 int idx, this_cpu, prev_cpu;
5004 struct task_group *tg;
5005 unsigned long weight;
5009 this_cpu = smp_processor_id();
5010 prev_cpu = task_cpu(p);
5011 load = source_load(prev_cpu, idx);
5012 this_load = target_load(this_cpu, idx);
5015 * If sync wakeup then subtract the (maximum possible)
5016 * effect of the currently running task from the load
5017 * of the current CPU:
5020 tg = task_group(current);
5021 weight = current->se.avg.load_avg;
5023 this_load += effective_load(tg, this_cpu, -weight, -weight);
5024 load += effective_load(tg, prev_cpu, 0, -weight);
5028 weight = p->se.avg.load_avg;
5031 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5032 * due to the sync cause above having dropped this_load to 0, we'll
5033 * always have an imbalance, but there's really nothing you can do
5034 * about that, so that's good too.
5036 * Otherwise check if either cpus are near enough in load to allow this
5037 * task to be woken on this_cpu.
5039 this_eff_load = 100;
5040 this_eff_load *= capacity_of(prev_cpu);
5042 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5043 prev_eff_load *= capacity_of(this_cpu);
5045 if (this_load > 0) {
5046 this_eff_load *= this_load +
5047 effective_load(tg, this_cpu, weight, weight);
5049 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5052 balanced = this_eff_load <= prev_eff_load;
5054 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5059 schedstat_inc(sd, ttwu_move_affine);
5060 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5065 static inline unsigned long task_util(struct task_struct *p)
5067 return p->se.avg.util_avg;
5070 unsigned int capacity_margin = 1280; /* ~20% margin */
5072 static inline unsigned long boosted_task_util(struct task_struct *task);
5074 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5076 unsigned long capacity = capacity_of(cpu);
5078 util += boosted_task_util(p);
5080 return (capacity * 1024) > (util * capacity_margin);
5083 static inline bool task_fits_max(struct task_struct *p, int cpu)
5085 unsigned long capacity = capacity_of(cpu);
5086 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5088 if (capacity == max_capacity)
5091 if (capacity * capacity_margin > max_capacity * 1024)
5094 return __task_fits(p, cpu, 0);
5097 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5099 return __task_fits(p, cpu, cpu_util(cpu));
5102 static bool cpu_overutilized(int cpu)
5104 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5107 #ifdef CONFIG_SCHED_TUNE
5109 static unsigned long
5110 schedtune_margin(unsigned long signal, unsigned long boost)
5112 unsigned long long margin = 0;
5115 * Signal proportional compensation (SPC)
5117 * The Boost (B) value is used to compute a Margin (M) which is
5118 * proportional to the complement of the original Signal (S):
5119 * M = B * (SCHED_LOAD_SCALE - S)
5120 * The obtained M could be used by the caller to "boost" S.
5122 margin = SCHED_LOAD_SCALE - signal;
5126 * Fast integer division by constant:
5127 * Constant : (C) = 100
5128 * Precision : 0.1% (P) = 0.1
5129 * Reference : C * 100 / P (R) = 100000
5132 * Shift bits : ceil(log(R,2)) (S) = 17
5133 * Mult const : round(2^S/C) (M) = 1311
5143 static inline unsigned int
5144 schedtune_cpu_margin(unsigned long util, int cpu)
5148 #ifdef CONFIG_CGROUP_SCHEDTUNE
5149 boost = schedtune_cpu_boost(cpu);
5151 boost = get_sysctl_sched_cfs_boost();
5156 return schedtune_margin(util, boost);
5159 static inline unsigned long
5160 schedtune_task_margin(struct task_struct *task)
5164 unsigned long margin;
5166 #ifdef CONFIG_CGROUP_SCHEDTUNE
5167 boost = schedtune_task_boost(task);
5169 boost = get_sysctl_sched_cfs_boost();
5174 util = task_util(task);
5175 margin = schedtune_margin(util, boost);
5180 #else /* CONFIG_SCHED_TUNE */
5182 static inline unsigned int
5183 schedtune_cpu_margin(unsigned long util, int cpu)
5188 static inline unsigned int
5189 schedtune_task_margin(struct task_struct *task)
5194 #endif /* CONFIG_SCHED_TUNE */
5196 static inline unsigned long
5197 boosted_cpu_util(int cpu)
5199 unsigned long util = cpu_util(cpu);
5200 unsigned long margin = schedtune_cpu_margin(util, cpu);
5202 return util + margin;
5205 static inline unsigned long
5206 boosted_task_util(struct task_struct *task)
5208 unsigned long util = task_util(task);
5209 unsigned long margin = schedtune_task_margin(task);
5211 return util + margin;
5215 * find_idlest_group finds and returns the least busy CPU group within the
5218 static struct sched_group *
5219 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5220 int this_cpu, int sd_flag)
5222 struct sched_group *idlest = NULL, *group = sd->groups;
5223 struct sched_group *fit_group = NULL, *spare_group = NULL;
5224 unsigned long min_load = ULONG_MAX, this_load = 0;
5225 unsigned long fit_capacity = ULONG_MAX;
5226 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5227 int load_idx = sd->forkexec_idx;
5228 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5230 if (sd_flag & SD_BALANCE_WAKE)
5231 load_idx = sd->wake_idx;
5234 unsigned long load, avg_load, spare_capacity;
5238 /* Skip over this group if it has no CPUs allowed */
5239 if (!cpumask_intersects(sched_group_cpus(group),
5240 tsk_cpus_allowed(p)))
5243 local_group = cpumask_test_cpu(this_cpu,
5244 sched_group_cpus(group));
5246 /* Tally up the load of all CPUs in the group */
5249 for_each_cpu(i, sched_group_cpus(group)) {
5250 /* Bias balancing toward cpus of our domain */
5252 load = source_load(i, load_idx);
5254 load = target_load(i, load_idx);
5259 * Look for most energy-efficient group that can fit
5260 * that can fit the task.
5262 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5263 fit_capacity = capacity_of(i);
5268 * Look for group which has most spare capacity on a
5271 spare_capacity = capacity_of(i) - cpu_util(i);
5272 if (spare_capacity > max_spare_capacity) {
5273 max_spare_capacity = spare_capacity;
5274 spare_group = group;
5278 /* Adjust by relative CPU capacity of the group */
5279 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5282 this_load = avg_load;
5283 } else if (avg_load < min_load) {
5284 min_load = avg_load;
5287 } while (group = group->next, group != sd->groups);
5295 if (!idlest || 100*this_load < imbalance*min_load)
5301 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5304 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5306 unsigned long load, min_load = ULONG_MAX;
5307 unsigned int min_exit_latency = UINT_MAX;
5308 u64 latest_idle_timestamp = 0;
5309 int least_loaded_cpu = this_cpu;
5310 int shallowest_idle_cpu = -1;
5313 /* Traverse only the allowed CPUs */
5314 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5315 if (task_fits_spare(p, i)) {
5316 struct rq *rq = cpu_rq(i);
5317 struct cpuidle_state *idle = idle_get_state(rq);
5318 if (idle && idle->exit_latency < min_exit_latency) {
5320 * We give priority to a CPU whose idle state
5321 * has the smallest exit latency irrespective
5322 * of any idle timestamp.
5324 min_exit_latency = idle->exit_latency;
5325 latest_idle_timestamp = rq->idle_stamp;
5326 shallowest_idle_cpu = i;
5327 } else if (idle_cpu(i) &&
5328 (!idle || idle->exit_latency == min_exit_latency) &&
5329 rq->idle_stamp > latest_idle_timestamp) {
5331 * If equal or no active idle state, then
5332 * the most recently idled CPU might have
5335 latest_idle_timestamp = rq->idle_stamp;
5336 shallowest_idle_cpu = i;
5337 } else if (shallowest_idle_cpu == -1) {
5339 * If we haven't found an idle CPU yet
5340 * pick a non-idle one that can fit the task as
5343 shallowest_idle_cpu = i;
5345 } else if (shallowest_idle_cpu == -1) {
5346 load = weighted_cpuload(i);
5347 if (load < min_load || (load == min_load && i == this_cpu)) {
5349 least_loaded_cpu = i;
5354 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5358 * Try and locate an idle CPU in the sched_domain.
5360 static int select_idle_sibling(struct task_struct *p, int target)
5362 struct sched_domain *sd;
5363 struct sched_group *sg;
5364 int i = task_cpu(p);
5366 if (idle_cpu(target))
5370 * If the prevous cpu is cache affine and idle, don't be stupid.
5372 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5376 * Otherwise, iterate the domains and find an elegible idle cpu.
5378 sd = rcu_dereference(per_cpu(sd_llc, target));
5379 for_each_lower_domain(sd) {
5382 if (!cpumask_intersects(sched_group_cpus(sg),
5383 tsk_cpus_allowed(p)))
5386 for_each_cpu(i, sched_group_cpus(sg)) {
5387 if (i == target || !idle_cpu(i))
5391 target = cpumask_first_and(sched_group_cpus(sg),
5392 tsk_cpus_allowed(p));
5396 } while (sg != sd->groups);
5402 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5404 struct sched_domain *sd;
5405 struct sched_group *sg, *sg_target;
5406 int target_max_cap = INT_MAX;
5407 int target_cpu = task_cpu(p);
5410 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5419 * Find group with sufficient capacity. We only get here if no cpu is
5420 * overutilized. We may end up overutilizing a cpu by adding the task,
5421 * but that should not be any worse than select_idle_sibling().
5422 * load_balance() should sort it out later as we get above the tipping
5426 /* Assuming all cpus are the same in group */
5427 int max_cap_cpu = group_first_cpu(sg);
5430 * Assume smaller max capacity means more energy-efficient.
5431 * Ideally we should query the energy model for the right
5432 * answer but it easily ends up in an exhaustive search.
5434 if (capacity_of(max_cap_cpu) < target_max_cap &&
5435 task_fits_max(p, max_cap_cpu)) {
5437 target_max_cap = capacity_of(max_cap_cpu);
5439 } while (sg = sg->next, sg != sd->groups);
5441 /* Find cpu with sufficient capacity */
5442 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5444 * p's blocked utilization is still accounted for on prev_cpu
5445 * so prev_cpu will receive a negative bias due to the double
5446 * accounting. However, the blocked utilization may be zero.
5448 int new_util = cpu_util(i) + boosted_task_util(p);
5450 if (new_util > capacity_orig_of(i))
5453 if (new_util < capacity_curr_of(i)) {
5455 if (cpu_rq(i)->nr_running)
5459 /* cpu has capacity at higher OPP, keep it as fallback */
5460 if (target_cpu == task_cpu(p))
5464 if (target_cpu != task_cpu(p)) {
5465 struct energy_env eenv = {
5466 .util_delta = task_util(p),
5467 .src_cpu = task_cpu(p),
5468 .dst_cpu = target_cpu,
5471 /* Not enough spare capacity on previous cpu */
5472 if (cpu_overutilized(task_cpu(p)))
5475 if (energy_diff(&eenv) >= 0)
5483 * select_task_rq_fair: Select target runqueue for the waking task in domains
5484 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5485 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5487 * Balances load by selecting the idlest cpu in the idlest group, or under
5488 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5490 * Returns the target cpu number.
5492 * preempt must be disabled.
5495 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5497 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5498 int cpu = smp_processor_id();
5499 int new_cpu = prev_cpu;
5500 int want_affine = 0;
5501 int sync = wake_flags & WF_SYNC;
5503 if (sd_flag & SD_BALANCE_WAKE)
5504 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5505 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5509 for_each_domain(cpu, tmp) {
5510 if (!(tmp->flags & SD_LOAD_BALANCE))
5514 * If both cpu and prev_cpu are part of this domain,
5515 * cpu is a valid SD_WAKE_AFFINE target.
5517 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5518 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5523 if (tmp->flags & sd_flag)
5525 else if (!want_affine)
5530 sd = NULL; /* Prefer wake_affine over balance flags */
5531 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5536 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5537 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5538 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5539 new_cpu = select_idle_sibling(p, new_cpu);
5542 struct sched_group *group;
5545 if (!(sd->flags & sd_flag)) {
5550 group = find_idlest_group(sd, p, cpu, sd_flag);
5556 new_cpu = find_idlest_cpu(group, p, cpu);
5557 if (new_cpu == -1 || new_cpu == cpu) {
5558 /* Now try balancing at a lower domain level of cpu */
5563 /* Now try balancing at a lower domain level of new_cpu */
5565 weight = sd->span_weight;
5567 for_each_domain(cpu, tmp) {
5568 if (weight <= tmp->span_weight)
5570 if (tmp->flags & sd_flag)
5573 /* while loop will break here if sd == NULL */
5581 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5582 * cfs_rq_of(p) references at time of call are still valid and identify the
5583 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5584 * other assumptions, including the state of rq->lock, should be made.
5586 static void migrate_task_rq_fair(struct task_struct *p)
5589 * We are supposed to update the task to "current" time, then its up to date
5590 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5591 * what current time is, so simply throw away the out-of-date time. This
5592 * will result in the wakee task is less decayed, but giving the wakee more
5593 * load sounds not bad.
5595 remove_entity_load_avg(&p->se);
5597 /* Tell new CPU we are migrated */
5598 p->se.avg.last_update_time = 0;
5600 /* We have migrated, no longer consider this task hot */
5601 p->se.exec_start = 0;
5604 static void task_dead_fair(struct task_struct *p)
5606 remove_entity_load_avg(&p->se);
5608 #endif /* CONFIG_SMP */
5610 static unsigned long
5611 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5613 unsigned long gran = sysctl_sched_wakeup_granularity;
5616 * Since its curr running now, convert the gran from real-time
5617 * to virtual-time in his units.
5619 * By using 'se' instead of 'curr' we penalize light tasks, so
5620 * they get preempted easier. That is, if 'se' < 'curr' then
5621 * the resulting gran will be larger, therefore penalizing the
5622 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5623 * be smaller, again penalizing the lighter task.
5625 * This is especially important for buddies when the leftmost
5626 * task is higher priority than the buddy.
5628 return calc_delta_fair(gran, se);
5632 * Should 'se' preempt 'curr'.
5646 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5648 s64 gran, vdiff = curr->vruntime - se->vruntime;
5653 gran = wakeup_gran(curr, se);
5660 static void set_last_buddy(struct sched_entity *se)
5662 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5665 for_each_sched_entity(se)
5666 cfs_rq_of(se)->last = se;
5669 static void set_next_buddy(struct sched_entity *se)
5671 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5674 for_each_sched_entity(se)
5675 cfs_rq_of(se)->next = se;
5678 static void set_skip_buddy(struct sched_entity *se)
5680 for_each_sched_entity(se)
5681 cfs_rq_of(se)->skip = se;
5685 * Preempt the current task with a newly woken task if needed:
5687 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5689 struct task_struct *curr = rq->curr;
5690 struct sched_entity *se = &curr->se, *pse = &p->se;
5691 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5692 int scale = cfs_rq->nr_running >= sched_nr_latency;
5693 int next_buddy_marked = 0;
5695 if (unlikely(se == pse))
5699 * This is possible from callers such as attach_tasks(), in which we
5700 * unconditionally check_prempt_curr() after an enqueue (which may have
5701 * lead to a throttle). This both saves work and prevents false
5702 * next-buddy nomination below.
5704 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5707 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5708 set_next_buddy(pse);
5709 next_buddy_marked = 1;
5713 * We can come here with TIF_NEED_RESCHED already set from new task
5716 * Note: this also catches the edge-case of curr being in a throttled
5717 * group (e.g. via set_curr_task), since update_curr() (in the
5718 * enqueue of curr) will have resulted in resched being set. This
5719 * prevents us from potentially nominating it as a false LAST_BUDDY
5722 if (test_tsk_need_resched(curr))
5725 /* Idle tasks are by definition preempted by non-idle tasks. */
5726 if (unlikely(curr->policy == SCHED_IDLE) &&
5727 likely(p->policy != SCHED_IDLE))
5731 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5732 * is driven by the tick):
5734 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5737 find_matching_se(&se, &pse);
5738 update_curr(cfs_rq_of(se));
5740 if (wakeup_preempt_entity(se, pse) == 1) {
5742 * Bias pick_next to pick the sched entity that is
5743 * triggering this preemption.
5745 if (!next_buddy_marked)
5746 set_next_buddy(pse);
5755 * Only set the backward buddy when the current task is still
5756 * on the rq. This can happen when a wakeup gets interleaved
5757 * with schedule on the ->pre_schedule() or idle_balance()
5758 * point, either of which can * drop the rq lock.
5760 * Also, during early boot the idle thread is in the fair class,
5761 * for obvious reasons its a bad idea to schedule back to it.
5763 if (unlikely(!se->on_rq || curr == rq->idle))
5766 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5770 static struct task_struct *
5771 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5773 struct cfs_rq *cfs_rq = &rq->cfs;
5774 struct sched_entity *se;
5775 struct task_struct *p;
5779 #ifdef CONFIG_FAIR_GROUP_SCHED
5780 if (!cfs_rq->nr_running)
5783 if (prev->sched_class != &fair_sched_class)
5787 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5788 * likely that a next task is from the same cgroup as the current.
5790 * Therefore attempt to avoid putting and setting the entire cgroup
5791 * hierarchy, only change the part that actually changes.
5795 struct sched_entity *curr = cfs_rq->curr;
5798 * Since we got here without doing put_prev_entity() we also
5799 * have to consider cfs_rq->curr. If it is still a runnable
5800 * entity, update_curr() will update its vruntime, otherwise
5801 * forget we've ever seen it.
5805 update_curr(cfs_rq);
5810 * This call to check_cfs_rq_runtime() will do the
5811 * throttle and dequeue its entity in the parent(s).
5812 * Therefore the 'simple' nr_running test will indeed
5815 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5819 se = pick_next_entity(cfs_rq, curr);
5820 cfs_rq = group_cfs_rq(se);
5826 * Since we haven't yet done put_prev_entity and if the selected task
5827 * is a different task than we started out with, try and touch the
5828 * least amount of cfs_rqs.
5831 struct sched_entity *pse = &prev->se;
5833 while (!(cfs_rq = is_same_group(se, pse))) {
5834 int se_depth = se->depth;
5835 int pse_depth = pse->depth;
5837 if (se_depth <= pse_depth) {
5838 put_prev_entity(cfs_rq_of(pse), pse);
5839 pse = parent_entity(pse);
5841 if (se_depth >= pse_depth) {
5842 set_next_entity(cfs_rq_of(se), se);
5843 se = parent_entity(se);
5847 put_prev_entity(cfs_rq, pse);
5848 set_next_entity(cfs_rq, se);
5851 if (hrtick_enabled(rq))
5852 hrtick_start_fair(rq, p);
5854 rq->misfit_task = !task_fits_max(p, rq->cpu);
5861 if (!cfs_rq->nr_running)
5864 put_prev_task(rq, prev);
5867 se = pick_next_entity(cfs_rq, NULL);
5868 set_next_entity(cfs_rq, se);
5869 cfs_rq = group_cfs_rq(se);
5874 if (hrtick_enabled(rq))
5875 hrtick_start_fair(rq, p);
5877 rq->misfit_task = !task_fits_max(p, rq->cpu);
5882 rq->misfit_task = 0;
5884 * This is OK, because current is on_cpu, which avoids it being picked
5885 * for load-balance and preemption/IRQs are still disabled avoiding
5886 * further scheduler activity on it and we're being very careful to
5887 * re-start the picking loop.
5889 lockdep_unpin_lock(&rq->lock);
5890 new_tasks = idle_balance(rq);
5891 lockdep_pin_lock(&rq->lock);
5893 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5894 * possible for any higher priority task to appear. In that case we
5895 * must re-start the pick_next_entity() loop.
5907 * Account for a descheduled task:
5909 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5911 struct sched_entity *se = &prev->se;
5912 struct cfs_rq *cfs_rq;
5914 for_each_sched_entity(se) {
5915 cfs_rq = cfs_rq_of(se);
5916 put_prev_entity(cfs_rq, se);
5921 * sched_yield() is very simple
5923 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5925 static void yield_task_fair(struct rq *rq)
5927 struct task_struct *curr = rq->curr;
5928 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5929 struct sched_entity *se = &curr->se;
5932 * Are we the only task in the tree?
5934 if (unlikely(rq->nr_running == 1))
5937 clear_buddies(cfs_rq, se);
5939 if (curr->policy != SCHED_BATCH) {
5940 update_rq_clock(rq);
5942 * Update run-time statistics of the 'current'.
5944 update_curr(cfs_rq);
5946 * Tell update_rq_clock() that we've just updated,
5947 * so we don't do microscopic update in schedule()
5948 * and double the fastpath cost.
5950 rq_clock_skip_update(rq, true);
5956 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5958 struct sched_entity *se = &p->se;
5960 /* throttled hierarchies are not runnable */
5961 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5964 /* Tell the scheduler that we'd really like pse to run next. */
5967 yield_task_fair(rq);
5973 /**************************************************
5974 * Fair scheduling class load-balancing methods.
5978 * The purpose of load-balancing is to achieve the same basic fairness the
5979 * per-cpu scheduler provides, namely provide a proportional amount of compute
5980 * time to each task. This is expressed in the following equation:
5982 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5984 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5985 * W_i,0 is defined as:
5987 * W_i,0 = \Sum_j w_i,j (2)
5989 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5990 * is derived from the nice value as per prio_to_weight[].
5992 * The weight average is an exponential decay average of the instantaneous
5995 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5997 * C_i is the compute capacity of cpu i, typically it is the
5998 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5999 * can also include other factors [XXX].
6001 * To achieve this balance we define a measure of imbalance which follows
6002 * directly from (1):
6004 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6006 * We them move tasks around to minimize the imbalance. In the continuous
6007 * function space it is obvious this converges, in the discrete case we get
6008 * a few fun cases generally called infeasible weight scenarios.
6011 * - infeasible weights;
6012 * - local vs global optima in the discrete case. ]
6017 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6018 * for all i,j solution, we create a tree of cpus that follows the hardware
6019 * topology where each level pairs two lower groups (or better). This results
6020 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6021 * tree to only the first of the previous level and we decrease the frequency
6022 * of load-balance at each level inv. proportional to the number of cpus in
6028 * \Sum { --- * --- * 2^i } = O(n) (5)
6030 * `- size of each group
6031 * | | `- number of cpus doing load-balance
6033 * `- sum over all levels
6035 * Coupled with a limit on how many tasks we can migrate every balance pass,
6036 * this makes (5) the runtime complexity of the balancer.
6038 * An important property here is that each CPU is still (indirectly) connected
6039 * to every other cpu in at most O(log n) steps:
6041 * The adjacency matrix of the resulting graph is given by:
6044 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6047 * And you'll find that:
6049 * A^(log_2 n)_i,j != 0 for all i,j (7)
6051 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6052 * The task movement gives a factor of O(m), giving a convergence complexity
6055 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6060 * In order to avoid CPUs going idle while there's still work to do, new idle
6061 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6062 * tree itself instead of relying on other CPUs to bring it work.
6064 * This adds some complexity to both (5) and (8) but it reduces the total idle
6072 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6075 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6080 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6082 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6084 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6087 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6088 * rewrite all of this once again.]
6091 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6093 enum fbq_type { regular, remote, all };
6102 #define LBF_ALL_PINNED 0x01
6103 #define LBF_NEED_BREAK 0x02
6104 #define LBF_DST_PINNED 0x04
6105 #define LBF_SOME_PINNED 0x08
6108 struct sched_domain *sd;
6116 struct cpumask *dst_grpmask;
6118 enum cpu_idle_type idle;
6120 unsigned int src_grp_nr_running;
6121 /* The set of CPUs under consideration for load-balancing */
6122 struct cpumask *cpus;
6127 unsigned int loop_break;
6128 unsigned int loop_max;
6130 enum fbq_type fbq_type;
6131 enum group_type busiest_group_type;
6132 struct list_head tasks;
6136 * Is this task likely cache-hot:
6138 static int task_hot(struct task_struct *p, struct lb_env *env)
6142 lockdep_assert_held(&env->src_rq->lock);
6144 if (p->sched_class != &fair_sched_class)
6147 if (unlikely(p->policy == SCHED_IDLE))
6151 * Buddy candidates are cache hot:
6153 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6154 (&p->se == cfs_rq_of(&p->se)->next ||
6155 &p->se == cfs_rq_of(&p->se)->last))
6158 if (sysctl_sched_migration_cost == -1)
6160 if (sysctl_sched_migration_cost == 0)
6163 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6165 return delta < (s64)sysctl_sched_migration_cost;
6168 #ifdef CONFIG_NUMA_BALANCING
6170 * Returns 1, if task migration degrades locality
6171 * Returns 0, if task migration improves locality i.e migration preferred.
6172 * Returns -1, if task migration is not affected by locality.
6174 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6176 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6177 unsigned long src_faults, dst_faults;
6178 int src_nid, dst_nid;
6180 if (!static_branch_likely(&sched_numa_balancing))
6183 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6186 src_nid = cpu_to_node(env->src_cpu);
6187 dst_nid = cpu_to_node(env->dst_cpu);
6189 if (src_nid == dst_nid)
6192 /* Migrating away from the preferred node is always bad. */
6193 if (src_nid == p->numa_preferred_nid) {
6194 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6200 /* Encourage migration to the preferred node. */
6201 if (dst_nid == p->numa_preferred_nid)
6205 src_faults = group_faults(p, src_nid);
6206 dst_faults = group_faults(p, dst_nid);
6208 src_faults = task_faults(p, src_nid);
6209 dst_faults = task_faults(p, dst_nid);
6212 return dst_faults < src_faults;
6216 static inline int migrate_degrades_locality(struct task_struct *p,
6224 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6227 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6231 lockdep_assert_held(&env->src_rq->lock);
6234 * We do not migrate tasks that are:
6235 * 1) throttled_lb_pair, or
6236 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6237 * 3) running (obviously), or
6238 * 4) are cache-hot on their current CPU.
6240 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6243 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6246 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6248 env->flags |= LBF_SOME_PINNED;
6251 * Remember if this task can be migrated to any other cpu in
6252 * our sched_group. We may want to revisit it if we couldn't
6253 * meet load balance goals by pulling other tasks on src_cpu.
6255 * Also avoid computing new_dst_cpu if we have already computed
6256 * one in current iteration.
6258 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6261 /* Prevent to re-select dst_cpu via env's cpus */
6262 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6263 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6264 env->flags |= LBF_DST_PINNED;
6265 env->new_dst_cpu = cpu;
6273 /* Record that we found atleast one task that could run on dst_cpu */
6274 env->flags &= ~LBF_ALL_PINNED;
6276 if (task_running(env->src_rq, p)) {
6277 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6282 * Aggressive migration if:
6283 * 1) destination numa is preferred
6284 * 2) task is cache cold, or
6285 * 3) too many balance attempts have failed.
6287 tsk_cache_hot = migrate_degrades_locality(p, env);
6288 if (tsk_cache_hot == -1)
6289 tsk_cache_hot = task_hot(p, env);
6291 if (tsk_cache_hot <= 0 ||
6292 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6293 if (tsk_cache_hot == 1) {
6294 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6295 schedstat_inc(p, se.statistics.nr_forced_migrations);
6300 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6305 * detach_task() -- detach the task for the migration specified in env
6307 static void detach_task(struct task_struct *p, struct lb_env *env)
6309 lockdep_assert_held(&env->src_rq->lock);
6311 deactivate_task(env->src_rq, p, 0);
6312 p->on_rq = TASK_ON_RQ_MIGRATING;
6313 set_task_cpu(p, env->dst_cpu);
6317 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6318 * part of active balancing operations within "domain".
6320 * Returns a task if successful and NULL otherwise.
6322 static struct task_struct *detach_one_task(struct lb_env *env)
6324 struct task_struct *p, *n;
6326 lockdep_assert_held(&env->src_rq->lock);
6328 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6329 if (!can_migrate_task(p, env))
6332 detach_task(p, env);
6335 * Right now, this is only the second place where
6336 * lb_gained[env->idle] is updated (other is detach_tasks)
6337 * so we can safely collect stats here rather than
6338 * inside detach_tasks().
6340 schedstat_inc(env->sd, lb_gained[env->idle]);
6346 static const unsigned int sched_nr_migrate_break = 32;
6349 * detach_tasks() -- tries to detach up to imbalance weighted load from
6350 * busiest_rq, as part of a balancing operation within domain "sd".
6352 * Returns number of detached tasks if successful and 0 otherwise.
6354 static int detach_tasks(struct lb_env *env)
6356 struct list_head *tasks = &env->src_rq->cfs_tasks;
6357 struct task_struct *p;
6361 lockdep_assert_held(&env->src_rq->lock);
6363 if (env->imbalance <= 0)
6366 while (!list_empty(tasks)) {
6368 * We don't want to steal all, otherwise we may be treated likewise,
6369 * which could at worst lead to a livelock crash.
6371 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6374 p = list_first_entry(tasks, struct task_struct, se.group_node);
6377 /* We've more or less seen every task there is, call it quits */
6378 if (env->loop > env->loop_max)
6381 /* take a breather every nr_migrate tasks */
6382 if (env->loop > env->loop_break) {
6383 env->loop_break += sched_nr_migrate_break;
6384 env->flags |= LBF_NEED_BREAK;
6388 if (!can_migrate_task(p, env))
6391 load = task_h_load(p);
6393 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6396 if ((load / 2) > env->imbalance)
6399 detach_task(p, env);
6400 list_add(&p->se.group_node, &env->tasks);
6403 env->imbalance -= load;
6405 #ifdef CONFIG_PREEMPT
6407 * NEWIDLE balancing is a source of latency, so preemptible
6408 * kernels will stop after the first task is detached to minimize
6409 * the critical section.
6411 if (env->idle == CPU_NEWLY_IDLE)
6416 * We only want to steal up to the prescribed amount of
6419 if (env->imbalance <= 0)
6424 list_move_tail(&p->se.group_node, tasks);
6428 * Right now, this is one of only two places we collect this stat
6429 * so we can safely collect detach_one_task() stats here rather
6430 * than inside detach_one_task().
6432 schedstat_add(env->sd, lb_gained[env->idle], detached);
6438 * attach_task() -- attach the task detached by detach_task() to its new rq.
6440 static void attach_task(struct rq *rq, struct task_struct *p)
6442 lockdep_assert_held(&rq->lock);
6444 BUG_ON(task_rq(p) != rq);
6445 p->on_rq = TASK_ON_RQ_QUEUED;
6446 activate_task(rq, p, 0);
6447 check_preempt_curr(rq, p, 0);
6451 * attach_one_task() -- attaches the task returned from detach_one_task() to
6454 static void attach_one_task(struct rq *rq, struct task_struct *p)
6456 raw_spin_lock(&rq->lock);
6459 * We want to potentially raise target_cpu's OPP.
6461 update_capacity_of(cpu_of(rq));
6462 raw_spin_unlock(&rq->lock);
6466 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6469 static void attach_tasks(struct lb_env *env)
6471 struct list_head *tasks = &env->tasks;
6472 struct task_struct *p;
6474 raw_spin_lock(&env->dst_rq->lock);
6476 while (!list_empty(tasks)) {
6477 p = list_first_entry(tasks, struct task_struct, se.group_node);
6478 list_del_init(&p->se.group_node);
6480 attach_task(env->dst_rq, p);
6484 * We want to potentially raise env.dst_cpu's OPP.
6486 update_capacity_of(env->dst_cpu);
6488 raw_spin_unlock(&env->dst_rq->lock);
6491 #ifdef CONFIG_FAIR_GROUP_SCHED
6492 static void update_blocked_averages(int cpu)
6494 struct rq *rq = cpu_rq(cpu);
6495 struct cfs_rq *cfs_rq;
6496 unsigned long flags;
6498 raw_spin_lock_irqsave(&rq->lock, flags);
6499 update_rq_clock(rq);
6502 * Iterates the task_group tree in a bottom up fashion, see
6503 * list_add_leaf_cfs_rq() for details.
6505 for_each_leaf_cfs_rq(rq, cfs_rq) {
6506 /* throttled entities do not contribute to load */
6507 if (throttled_hierarchy(cfs_rq))
6510 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6511 update_tg_load_avg(cfs_rq, 0);
6513 raw_spin_unlock_irqrestore(&rq->lock, flags);
6517 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6518 * This needs to be done in a top-down fashion because the load of a child
6519 * group is a fraction of its parents load.
6521 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6523 struct rq *rq = rq_of(cfs_rq);
6524 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6525 unsigned long now = jiffies;
6528 if (cfs_rq->last_h_load_update == now)
6531 cfs_rq->h_load_next = NULL;
6532 for_each_sched_entity(se) {
6533 cfs_rq = cfs_rq_of(se);
6534 cfs_rq->h_load_next = se;
6535 if (cfs_rq->last_h_load_update == now)
6540 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6541 cfs_rq->last_h_load_update = now;
6544 while ((se = cfs_rq->h_load_next) != NULL) {
6545 load = cfs_rq->h_load;
6546 load = div64_ul(load * se->avg.load_avg,
6547 cfs_rq_load_avg(cfs_rq) + 1);
6548 cfs_rq = group_cfs_rq(se);
6549 cfs_rq->h_load = load;
6550 cfs_rq->last_h_load_update = now;
6554 static unsigned long task_h_load(struct task_struct *p)
6556 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6558 update_cfs_rq_h_load(cfs_rq);
6559 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6560 cfs_rq_load_avg(cfs_rq) + 1);
6563 static inline void update_blocked_averages(int cpu)
6565 struct rq *rq = cpu_rq(cpu);
6566 struct cfs_rq *cfs_rq = &rq->cfs;
6567 unsigned long flags;
6569 raw_spin_lock_irqsave(&rq->lock, flags);
6570 update_rq_clock(rq);
6571 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6572 raw_spin_unlock_irqrestore(&rq->lock, flags);
6575 static unsigned long task_h_load(struct task_struct *p)
6577 return p->se.avg.load_avg;
6581 /********** Helpers for find_busiest_group ************************/
6584 * sg_lb_stats - stats of a sched_group required for load_balancing
6586 struct sg_lb_stats {
6587 unsigned long avg_load; /*Avg load across the CPUs of the group */
6588 unsigned long group_load; /* Total load over the CPUs of the group */
6589 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6590 unsigned long load_per_task;
6591 unsigned long group_capacity;
6592 unsigned long group_util; /* Total utilization of the group */
6593 unsigned int sum_nr_running; /* Nr tasks running in the group */
6594 unsigned int idle_cpus;
6595 unsigned int group_weight;
6596 enum group_type group_type;
6597 int group_no_capacity;
6598 int group_misfit_task; /* A cpu has a task too big for its capacity */
6599 #ifdef CONFIG_NUMA_BALANCING
6600 unsigned int nr_numa_running;
6601 unsigned int nr_preferred_running;
6606 * sd_lb_stats - Structure to store the statistics of a sched_domain
6607 * during load balancing.
6609 struct sd_lb_stats {
6610 struct sched_group *busiest; /* Busiest group in this sd */
6611 struct sched_group *local; /* Local group in this sd */
6612 unsigned long total_load; /* Total load of all groups in sd */
6613 unsigned long total_capacity; /* Total capacity of all groups in sd */
6614 unsigned long avg_load; /* Average load across all groups in sd */
6616 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6617 struct sg_lb_stats local_stat; /* Statistics of the local group */
6620 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6623 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6624 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6625 * We must however clear busiest_stat::avg_load because
6626 * update_sd_pick_busiest() reads this before assignment.
6628 *sds = (struct sd_lb_stats){
6632 .total_capacity = 0UL,
6635 .sum_nr_running = 0,
6636 .group_type = group_other,
6642 * get_sd_load_idx - Obtain the load index for a given sched domain.
6643 * @sd: The sched_domain whose load_idx is to be obtained.
6644 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6646 * Return: The load index.
6648 static inline int get_sd_load_idx(struct sched_domain *sd,
6649 enum cpu_idle_type idle)
6655 load_idx = sd->busy_idx;
6658 case CPU_NEWLY_IDLE:
6659 load_idx = sd->newidle_idx;
6662 load_idx = sd->idle_idx;
6669 static unsigned long scale_rt_capacity(int cpu)
6671 struct rq *rq = cpu_rq(cpu);
6672 u64 total, used, age_stamp, avg;
6676 * Since we're reading these variables without serialization make sure
6677 * we read them once before doing sanity checks on them.
6679 age_stamp = READ_ONCE(rq->age_stamp);
6680 avg = READ_ONCE(rq->rt_avg);
6681 delta = __rq_clock_broken(rq) - age_stamp;
6683 if (unlikely(delta < 0))
6686 total = sched_avg_period() + delta;
6688 used = div_u64(avg, total);
6691 * deadline bandwidth is defined at system level so we must
6692 * weight this bandwidth with the max capacity of the system.
6693 * As a reminder, avg_bw is 20bits width and
6694 * scale_cpu_capacity is 10 bits width
6696 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6698 if (likely(used < SCHED_CAPACITY_SCALE))
6699 return SCHED_CAPACITY_SCALE - used;
6704 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6706 raw_spin_lock_init(&mcc->lock);
6711 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6713 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6714 struct sched_group *sdg = sd->groups;
6715 struct max_cpu_capacity *mcc;
6716 unsigned long max_capacity;
6718 unsigned long flags;
6720 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6722 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6724 raw_spin_lock_irqsave(&mcc->lock, flags);
6725 max_capacity = mcc->val;
6726 max_cap_cpu = mcc->cpu;
6728 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6729 (max_capacity < capacity)) {
6730 mcc->val = capacity;
6732 #ifdef CONFIG_SCHED_DEBUG
6733 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6734 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6738 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6740 skip_unlock: __attribute__ ((unused));
6741 capacity *= scale_rt_capacity(cpu);
6742 capacity >>= SCHED_CAPACITY_SHIFT;
6747 cpu_rq(cpu)->cpu_capacity = capacity;
6748 sdg->sgc->capacity = capacity;
6749 sdg->sgc->max_capacity = capacity;
6752 void update_group_capacity(struct sched_domain *sd, int cpu)
6754 struct sched_domain *child = sd->child;
6755 struct sched_group *group, *sdg = sd->groups;
6756 unsigned long capacity, max_capacity;
6757 unsigned long interval;
6759 interval = msecs_to_jiffies(sd->balance_interval);
6760 interval = clamp(interval, 1UL, max_load_balance_interval);
6761 sdg->sgc->next_update = jiffies + interval;
6764 update_cpu_capacity(sd, cpu);
6771 if (child->flags & SD_OVERLAP) {
6773 * SD_OVERLAP domains cannot assume that child groups
6774 * span the current group.
6777 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6778 struct sched_group_capacity *sgc;
6779 struct rq *rq = cpu_rq(cpu);
6782 * build_sched_domains() -> init_sched_groups_capacity()
6783 * gets here before we've attached the domains to the
6786 * Use capacity_of(), which is set irrespective of domains
6787 * in update_cpu_capacity().
6789 * This avoids capacity from being 0 and
6790 * causing divide-by-zero issues on boot.
6792 if (unlikely(!rq->sd)) {
6793 capacity += capacity_of(cpu);
6795 sgc = rq->sd->groups->sgc;
6796 capacity += sgc->capacity;
6799 max_capacity = max(capacity, max_capacity);
6803 * !SD_OVERLAP domains can assume that child groups
6804 * span the current group.
6807 group = child->groups;
6809 struct sched_group_capacity *sgc = group->sgc;
6811 capacity += sgc->capacity;
6812 max_capacity = max(sgc->max_capacity, max_capacity);
6813 group = group->next;
6814 } while (group != child->groups);
6817 sdg->sgc->capacity = capacity;
6818 sdg->sgc->max_capacity = max_capacity;
6822 * Check whether the capacity of the rq has been noticeably reduced by side
6823 * activity. The imbalance_pct is used for the threshold.
6824 * Return true is the capacity is reduced
6827 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6829 return ((rq->cpu_capacity * sd->imbalance_pct) <
6830 (rq->cpu_capacity_orig * 100));
6834 * Group imbalance indicates (and tries to solve) the problem where balancing
6835 * groups is inadequate due to tsk_cpus_allowed() constraints.
6837 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6838 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6841 * { 0 1 2 3 } { 4 5 6 7 }
6844 * If we were to balance group-wise we'd place two tasks in the first group and
6845 * two tasks in the second group. Clearly this is undesired as it will overload
6846 * cpu 3 and leave one of the cpus in the second group unused.
6848 * The current solution to this issue is detecting the skew in the first group
6849 * by noticing the lower domain failed to reach balance and had difficulty
6850 * moving tasks due to affinity constraints.
6852 * When this is so detected; this group becomes a candidate for busiest; see
6853 * update_sd_pick_busiest(). And calculate_imbalance() and
6854 * find_busiest_group() avoid some of the usual balance conditions to allow it
6855 * to create an effective group imbalance.
6857 * This is a somewhat tricky proposition since the next run might not find the
6858 * group imbalance and decide the groups need to be balanced again. A most
6859 * subtle and fragile situation.
6862 static inline int sg_imbalanced(struct sched_group *group)
6864 return group->sgc->imbalance;
6868 * group_has_capacity returns true if the group has spare capacity that could
6869 * be used by some tasks.
6870 * We consider that a group has spare capacity if the * number of task is
6871 * smaller than the number of CPUs or if the utilization is lower than the
6872 * available capacity for CFS tasks.
6873 * For the latter, we use a threshold to stabilize the state, to take into
6874 * account the variance of the tasks' load and to return true if the available
6875 * capacity in meaningful for the load balancer.
6876 * As an example, an available capacity of 1% can appear but it doesn't make
6877 * any benefit for the load balance.
6880 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6882 if (sgs->sum_nr_running < sgs->group_weight)
6885 if ((sgs->group_capacity * 100) >
6886 (sgs->group_util * env->sd->imbalance_pct))
6893 * group_is_overloaded returns true if the group has more tasks than it can
6895 * group_is_overloaded is not equals to !group_has_capacity because a group
6896 * with the exact right number of tasks, has no more spare capacity but is not
6897 * overloaded so both group_has_capacity and group_is_overloaded return
6901 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6903 if (sgs->sum_nr_running <= sgs->group_weight)
6906 if ((sgs->group_capacity * 100) <
6907 (sgs->group_util * env->sd->imbalance_pct))
6915 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
6916 * per-cpu capacity than sched_group ref.
6919 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
6921 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
6922 ref->sgc->max_capacity;
6926 group_type group_classify(struct sched_group *group,
6927 struct sg_lb_stats *sgs)
6929 if (sgs->group_no_capacity)
6930 return group_overloaded;
6932 if (sg_imbalanced(group))
6933 return group_imbalanced;
6935 if (sgs->group_misfit_task)
6936 return group_misfit_task;
6942 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6943 * @env: The load balancing environment.
6944 * @group: sched_group whose statistics are to be updated.
6945 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6946 * @local_group: Does group contain this_cpu.
6947 * @sgs: variable to hold the statistics for this group.
6948 * @overload: Indicate more than one runnable task for any CPU.
6949 * @overutilized: Indicate overutilization for any CPU.
6951 static inline void update_sg_lb_stats(struct lb_env *env,
6952 struct sched_group *group, int load_idx,
6953 int local_group, struct sg_lb_stats *sgs,
6954 bool *overload, bool *overutilized)
6959 memset(sgs, 0, sizeof(*sgs));
6961 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6962 struct rq *rq = cpu_rq(i);
6964 /* Bias balancing toward cpus of our domain */
6966 load = target_load(i, load_idx);
6968 load = source_load(i, load_idx);
6970 sgs->group_load += load;
6971 sgs->group_util += cpu_util(i);
6972 sgs->sum_nr_running += rq->cfs.h_nr_running;
6974 if (rq->nr_running > 1)
6977 #ifdef CONFIG_NUMA_BALANCING
6978 sgs->nr_numa_running += rq->nr_numa_running;
6979 sgs->nr_preferred_running += rq->nr_preferred_running;
6981 sgs->sum_weighted_load += weighted_cpuload(i);
6985 if (cpu_overutilized(i)) {
6986 *overutilized = true;
6987 if (!sgs->group_misfit_task && rq->misfit_task)
6988 sgs->group_misfit_task = capacity_of(i);
6992 /* Adjust by relative CPU capacity of the group */
6993 sgs->group_capacity = group->sgc->capacity;
6994 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6996 if (sgs->sum_nr_running)
6997 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6999 sgs->group_weight = group->group_weight;
7001 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7002 sgs->group_type = group_classify(group, sgs);
7006 * update_sd_pick_busiest - return 1 on busiest group
7007 * @env: The load balancing environment.
7008 * @sds: sched_domain statistics
7009 * @sg: sched_group candidate to be checked for being the busiest
7010 * @sgs: sched_group statistics
7012 * Determine if @sg is a busier group than the previously selected
7015 * Return: %true if @sg is a busier group than the previously selected
7016 * busiest group. %false otherwise.
7018 static bool update_sd_pick_busiest(struct lb_env *env,
7019 struct sd_lb_stats *sds,
7020 struct sched_group *sg,
7021 struct sg_lb_stats *sgs)
7023 struct sg_lb_stats *busiest = &sds->busiest_stat;
7025 if (sgs->group_type > busiest->group_type)
7028 if (sgs->group_type < busiest->group_type)
7032 * Candidate sg doesn't face any serious load-balance problems
7033 * so don't pick it if the local sg is already filled up.
7035 if (sgs->group_type == group_other &&
7036 !group_has_capacity(env, &sds->local_stat))
7039 if (sgs->avg_load <= busiest->avg_load)
7043 * Candiate sg has no more than one task per cpu and has higher
7044 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7046 if (sgs->sum_nr_running <= sgs->group_weight &&
7047 group_smaller_cpu_capacity(sds->local, sg))
7050 /* This is the busiest node in its class. */
7051 if (!(env->sd->flags & SD_ASYM_PACKING))
7055 * ASYM_PACKING needs to move all the work to the lowest
7056 * numbered CPUs in the group, therefore mark all groups
7057 * higher than ourself as busy.
7059 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7063 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7070 #ifdef CONFIG_NUMA_BALANCING
7071 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7073 if (sgs->sum_nr_running > sgs->nr_numa_running)
7075 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7080 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7082 if (rq->nr_running > rq->nr_numa_running)
7084 if (rq->nr_running > rq->nr_preferred_running)
7089 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7094 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7098 #endif /* CONFIG_NUMA_BALANCING */
7101 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7102 * @env: The load balancing environment.
7103 * @sds: variable to hold the statistics for this sched_domain.
7105 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7107 struct sched_domain *child = env->sd->child;
7108 struct sched_group *sg = env->sd->groups;
7109 struct sg_lb_stats tmp_sgs;
7110 int load_idx, prefer_sibling = 0;
7111 bool overload = false, overutilized = false;
7113 if (child && child->flags & SD_PREFER_SIBLING)
7116 load_idx = get_sd_load_idx(env->sd, env->idle);
7119 struct sg_lb_stats *sgs = &tmp_sgs;
7122 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7125 sgs = &sds->local_stat;
7127 if (env->idle != CPU_NEWLY_IDLE ||
7128 time_after_eq(jiffies, sg->sgc->next_update))
7129 update_group_capacity(env->sd, env->dst_cpu);
7132 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7133 &overload, &overutilized);
7139 * In case the child domain prefers tasks go to siblings
7140 * first, lower the sg capacity so that we'll try
7141 * and move all the excess tasks away. We lower the capacity
7142 * of a group only if the local group has the capacity to fit
7143 * these excess tasks. The extra check prevents the case where
7144 * you always pull from the heaviest group when it is already
7145 * under-utilized (possible with a large weight task outweighs
7146 * the tasks on the system).
7148 if (prefer_sibling && sds->local &&
7149 group_has_capacity(env, &sds->local_stat) &&
7150 (sgs->sum_nr_running > 1)) {
7151 sgs->group_no_capacity = 1;
7152 sgs->group_type = group_classify(sg, sgs);
7156 * Ignore task groups with misfit tasks if local group has no
7157 * capacity or if per-cpu capacity isn't higher.
7159 if (sgs->group_type == group_misfit_task &&
7160 (!group_has_capacity(env, &sds->local_stat) ||
7161 !group_smaller_cpu_capacity(sg, sds->local)))
7162 sgs->group_type = group_other;
7164 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7166 sds->busiest_stat = *sgs;
7170 /* Now, start updating sd_lb_stats */
7171 sds->total_load += sgs->group_load;
7172 sds->total_capacity += sgs->group_capacity;
7175 } while (sg != env->sd->groups);
7177 if (env->sd->flags & SD_NUMA)
7178 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7180 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7182 if (!env->sd->parent) {
7183 /* update overload indicator if we are at root domain */
7184 if (env->dst_rq->rd->overload != overload)
7185 env->dst_rq->rd->overload = overload;
7187 /* Update over-utilization (tipping point, U >= 0) indicator */
7188 if (env->dst_rq->rd->overutilized != overutilized)
7189 env->dst_rq->rd->overutilized = overutilized;
7191 if (!env->dst_rq->rd->overutilized && overutilized)
7192 env->dst_rq->rd->overutilized = true;
7197 * check_asym_packing - Check to see if the group is packed into the
7200 * This is primarily intended to used at the sibling level. Some
7201 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7202 * case of POWER7, it can move to lower SMT modes only when higher
7203 * threads are idle. When in lower SMT modes, the threads will
7204 * perform better since they share less core resources. Hence when we
7205 * have idle threads, we want them to be the higher ones.
7207 * This packing function is run on idle threads. It checks to see if
7208 * the busiest CPU in this domain (core in the P7 case) has a higher
7209 * CPU number than the packing function is being run on. Here we are
7210 * assuming lower CPU number will be equivalent to lower a SMT thread
7213 * Return: 1 when packing is required and a task should be moved to
7214 * this CPU. The amount of the imbalance is returned in *imbalance.
7216 * @env: The load balancing environment.
7217 * @sds: Statistics of the sched_domain which is to be packed
7219 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7223 if (!(env->sd->flags & SD_ASYM_PACKING))
7229 busiest_cpu = group_first_cpu(sds->busiest);
7230 if (env->dst_cpu > busiest_cpu)
7233 env->imbalance = DIV_ROUND_CLOSEST(
7234 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7235 SCHED_CAPACITY_SCALE);
7241 * fix_small_imbalance - Calculate the minor imbalance that exists
7242 * amongst the groups of a sched_domain, during
7244 * @env: The load balancing environment.
7245 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7248 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7250 unsigned long tmp, capa_now = 0, capa_move = 0;
7251 unsigned int imbn = 2;
7252 unsigned long scaled_busy_load_per_task;
7253 struct sg_lb_stats *local, *busiest;
7255 local = &sds->local_stat;
7256 busiest = &sds->busiest_stat;
7258 if (!local->sum_nr_running)
7259 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7260 else if (busiest->load_per_task > local->load_per_task)
7263 scaled_busy_load_per_task =
7264 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7265 busiest->group_capacity;
7267 if (busiest->avg_load + scaled_busy_load_per_task >=
7268 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7269 env->imbalance = busiest->load_per_task;
7274 * OK, we don't have enough imbalance to justify moving tasks,
7275 * however we may be able to increase total CPU capacity used by
7279 capa_now += busiest->group_capacity *
7280 min(busiest->load_per_task, busiest->avg_load);
7281 capa_now += local->group_capacity *
7282 min(local->load_per_task, local->avg_load);
7283 capa_now /= SCHED_CAPACITY_SCALE;
7285 /* Amount of load we'd subtract */
7286 if (busiest->avg_load > scaled_busy_load_per_task) {
7287 capa_move += busiest->group_capacity *
7288 min(busiest->load_per_task,
7289 busiest->avg_load - scaled_busy_load_per_task);
7292 /* Amount of load we'd add */
7293 if (busiest->avg_load * busiest->group_capacity <
7294 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7295 tmp = (busiest->avg_load * busiest->group_capacity) /
7296 local->group_capacity;
7298 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7299 local->group_capacity;
7301 capa_move += local->group_capacity *
7302 min(local->load_per_task, local->avg_load + tmp);
7303 capa_move /= SCHED_CAPACITY_SCALE;
7305 /* Move if we gain throughput */
7306 if (capa_move > capa_now)
7307 env->imbalance = busiest->load_per_task;
7311 * calculate_imbalance - Calculate the amount of imbalance present within the
7312 * groups of a given sched_domain during load balance.
7313 * @env: load balance environment
7314 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7316 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7318 unsigned long max_pull, load_above_capacity = ~0UL;
7319 struct sg_lb_stats *local, *busiest;
7321 local = &sds->local_stat;
7322 busiest = &sds->busiest_stat;
7324 if (busiest->group_type == group_imbalanced) {
7326 * In the group_imb case we cannot rely on group-wide averages
7327 * to ensure cpu-load equilibrium, look at wider averages. XXX
7329 busiest->load_per_task =
7330 min(busiest->load_per_task, sds->avg_load);
7334 * In the presence of smp nice balancing, certain scenarios can have
7335 * max load less than avg load(as we skip the groups at or below
7336 * its cpu_capacity, while calculating max_load..)
7338 if (busiest->avg_load <= sds->avg_load ||
7339 local->avg_load >= sds->avg_load) {
7340 /* Misfitting tasks should be migrated in any case */
7341 if (busiest->group_type == group_misfit_task) {
7342 env->imbalance = busiest->group_misfit_task;
7347 * Busiest group is overloaded, local is not, use the spare
7348 * cycles to maximize throughput
7350 if (busiest->group_type == group_overloaded &&
7351 local->group_type <= group_misfit_task) {
7352 env->imbalance = busiest->load_per_task;
7357 return fix_small_imbalance(env, sds);
7361 * If there aren't any idle cpus, avoid creating some.
7363 if (busiest->group_type == group_overloaded &&
7364 local->group_type == group_overloaded) {
7365 load_above_capacity = busiest->sum_nr_running *
7367 if (load_above_capacity > busiest->group_capacity)
7368 load_above_capacity -= busiest->group_capacity;
7370 load_above_capacity = ~0UL;
7374 * We're trying to get all the cpus to the average_load, so we don't
7375 * want to push ourselves above the average load, nor do we wish to
7376 * reduce the max loaded cpu below the average load. At the same time,
7377 * we also don't want to reduce the group load below the group capacity
7378 * (so that we can implement power-savings policies etc). Thus we look
7379 * for the minimum possible imbalance.
7381 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7383 /* How much load to actually move to equalise the imbalance */
7384 env->imbalance = min(
7385 max_pull * busiest->group_capacity,
7386 (sds->avg_load - local->avg_load) * local->group_capacity
7387 ) / SCHED_CAPACITY_SCALE;
7389 /* Boost imbalance to allow misfit task to be balanced. */
7390 if (busiest->group_type == group_misfit_task)
7391 env->imbalance = max_t(long, env->imbalance,
7392 busiest->group_misfit_task);
7395 * if *imbalance is less than the average load per runnable task
7396 * there is no guarantee that any tasks will be moved so we'll have
7397 * a think about bumping its value to force at least one task to be
7400 if (env->imbalance < busiest->load_per_task)
7401 return fix_small_imbalance(env, sds);
7404 /******* find_busiest_group() helpers end here *********************/
7407 * find_busiest_group - Returns the busiest group within the sched_domain
7408 * if there is an imbalance. If there isn't an imbalance, and
7409 * the user has opted for power-savings, it returns a group whose
7410 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7411 * such a group exists.
7413 * Also calculates the amount of weighted load which should be moved
7414 * to restore balance.
7416 * @env: The load balancing environment.
7418 * Return: - The busiest group if imbalance exists.
7419 * - If no imbalance and user has opted for power-savings balance,
7420 * return the least loaded group whose CPUs can be
7421 * put to idle by rebalancing its tasks onto our group.
7423 static struct sched_group *find_busiest_group(struct lb_env *env)
7425 struct sg_lb_stats *local, *busiest;
7426 struct sd_lb_stats sds;
7428 init_sd_lb_stats(&sds);
7431 * Compute the various statistics relavent for load balancing at
7434 update_sd_lb_stats(env, &sds);
7436 if (energy_aware() && !env->dst_rq->rd->overutilized)
7439 local = &sds.local_stat;
7440 busiest = &sds.busiest_stat;
7442 /* ASYM feature bypasses nice load balance check */
7443 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7444 check_asym_packing(env, &sds))
7447 /* There is no busy sibling group to pull tasks from */
7448 if (!sds.busiest || busiest->sum_nr_running == 0)
7451 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7452 / sds.total_capacity;
7455 * If the busiest group is imbalanced the below checks don't
7456 * work because they assume all things are equal, which typically
7457 * isn't true due to cpus_allowed constraints and the like.
7459 if (busiest->group_type == group_imbalanced)
7462 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7463 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7464 busiest->group_no_capacity)
7467 /* Misfitting tasks should be dealt with regardless of the avg load */
7468 if (busiest->group_type == group_misfit_task) {
7473 * If the local group is busier than the selected busiest group
7474 * don't try and pull any tasks.
7476 if (local->avg_load >= busiest->avg_load)
7480 * Don't pull any tasks if this group is already above the domain
7483 if (local->avg_load >= sds.avg_load)
7486 if (env->idle == CPU_IDLE) {
7488 * This cpu is idle. If the busiest group is not overloaded
7489 * and there is no imbalance between this and busiest group
7490 * wrt idle cpus, it is balanced. The imbalance becomes
7491 * significant if the diff is greater than 1 otherwise we
7492 * might end up to just move the imbalance on another group
7494 if ((busiest->group_type != group_overloaded) &&
7495 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7496 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7500 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7501 * imbalance_pct to be conservative.
7503 if (100 * busiest->avg_load <=
7504 env->sd->imbalance_pct * local->avg_load)
7509 env->busiest_group_type = busiest->group_type;
7510 /* Looks like there is an imbalance. Compute it */
7511 calculate_imbalance(env, &sds);
7520 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7522 static struct rq *find_busiest_queue(struct lb_env *env,
7523 struct sched_group *group)
7525 struct rq *busiest = NULL, *rq;
7526 unsigned long busiest_load = 0, busiest_capacity = 1;
7529 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7530 unsigned long capacity, wl;
7534 rt = fbq_classify_rq(rq);
7537 * We classify groups/runqueues into three groups:
7538 * - regular: there are !numa tasks
7539 * - remote: there are numa tasks that run on the 'wrong' node
7540 * - all: there is no distinction
7542 * In order to avoid migrating ideally placed numa tasks,
7543 * ignore those when there's better options.
7545 * If we ignore the actual busiest queue to migrate another
7546 * task, the next balance pass can still reduce the busiest
7547 * queue by moving tasks around inside the node.
7549 * If we cannot move enough load due to this classification
7550 * the next pass will adjust the group classification and
7551 * allow migration of more tasks.
7553 * Both cases only affect the total convergence complexity.
7555 if (rt > env->fbq_type)
7558 capacity = capacity_of(i);
7560 wl = weighted_cpuload(i);
7563 * When comparing with imbalance, use weighted_cpuload()
7564 * which is not scaled with the cpu capacity.
7567 if (rq->nr_running == 1 && wl > env->imbalance &&
7568 !check_cpu_capacity(rq, env->sd) &&
7569 env->busiest_group_type != group_misfit_task)
7573 * For the load comparisons with the other cpu's, consider
7574 * the weighted_cpuload() scaled with the cpu capacity, so
7575 * that the load can be moved away from the cpu that is
7576 * potentially running at a lower capacity.
7578 * Thus we're looking for max(wl_i / capacity_i), crosswise
7579 * multiplication to rid ourselves of the division works out
7580 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7581 * our previous maximum.
7583 if (wl * busiest_capacity > busiest_load * capacity) {
7585 busiest_capacity = capacity;
7594 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7595 * so long as it is large enough.
7597 #define MAX_PINNED_INTERVAL 512
7599 /* Working cpumask for load_balance and load_balance_newidle. */
7600 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7602 static int need_active_balance(struct lb_env *env)
7604 struct sched_domain *sd = env->sd;
7606 if (env->idle == CPU_NEWLY_IDLE) {
7609 * ASYM_PACKING needs to force migrate tasks from busy but
7610 * higher numbered CPUs in order to pack all tasks in the
7611 * lowest numbered CPUs.
7613 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7618 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7619 * It's worth migrating the task if the src_cpu's capacity is reduced
7620 * because of other sched_class or IRQs if more capacity stays
7621 * available on dst_cpu.
7623 if ((env->idle != CPU_NOT_IDLE) &&
7624 (env->src_rq->cfs.h_nr_running == 1)) {
7625 if ((check_cpu_capacity(env->src_rq, sd)) &&
7626 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7630 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7631 env->src_rq->cfs.h_nr_running == 1 &&
7632 cpu_overutilized(env->src_cpu) &&
7633 !cpu_overutilized(env->dst_cpu)) {
7637 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7640 static int active_load_balance_cpu_stop(void *data);
7642 static int should_we_balance(struct lb_env *env)
7644 struct sched_group *sg = env->sd->groups;
7645 struct cpumask *sg_cpus, *sg_mask;
7646 int cpu, balance_cpu = -1;
7649 * In the newly idle case, we will allow all the cpu's
7650 * to do the newly idle load balance.
7652 if (env->idle == CPU_NEWLY_IDLE)
7655 sg_cpus = sched_group_cpus(sg);
7656 sg_mask = sched_group_mask(sg);
7657 /* Try to find first idle cpu */
7658 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7659 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7666 if (balance_cpu == -1)
7667 balance_cpu = group_balance_cpu(sg);
7670 * First idle cpu or the first cpu(busiest) in this sched group
7671 * is eligible for doing load balancing at this and above domains.
7673 return balance_cpu == env->dst_cpu;
7677 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7678 * tasks if there is an imbalance.
7680 static int load_balance(int this_cpu, struct rq *this_rq,
7681 struct sched_domain *sd, enum cpu_idle_type idle,
7682 int *continue_balancing)
7684 int ld_moved, cur_ld_moved, active_balance = 0;
7685 struct sched_domain *sd_parent = sd->parent;
7686 struct sched_group *group;
7688 unsigned long flags;
7689 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7691 struct lb_env env = {
7693 .dst_cpu = this_cpu,
7695 .dst_grpmask = sched_group_cpus(sd->groups),
7697 .loop_break = sched_nr_migrate_break,
7700 .tasks = LIST_HEAD_INIT(env.tasks),
7704 * For NEWLY_IDLE load_balancing, we don't need to consider
7705 * other cpus in our group
7707 if (idle == CPU_NEWLY_IDLE)
7708 env.dst_grpmask = NULL;
7710 cpumask_copy(cpus, cpu_active_mask);
7712 schedstat_inc(sd, lb_count[idle]);
7715 if (!should_we_balance(&env)) {
7716 *continue_balancing = 0;
7720 group = find_busiest_group(&env);
7722 schedstat_inc(sd, lb_nobusyg[idle]);
7726 busiest = find_busiest_queue(&env, group);
7728 schedstat_inc(sd, lb_nobusyq[idle]);
7732 BUG_ON(busiest == env.dst_rq);
7734 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7736 env.src_cpu = busiest->cpu;
7737 env.src_rq = busiest;
7740 if (busiest->nr_running > 1) {
7742 * Attempt to move tasks. If find_busiest_group has found
7743 * an imbalance but busiest->nr_running <= 1, the group is
7744 * still unbalanced. ld_moved simply stays zero, so it is
7745 * correctly treated as an imbalance.
7747 env.flags |= LBF_ALL_PINNED;
7748 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7751 raw_spin_lock_irqsave(&busiest->lock, flags);
7754 * cur_ld_moved - load moved in current iteration
7755 * ld_moved - cumulative load moved across iterations
7757 cur_ld_moved = detach_tasks(&env);
7759 * We want to potentially lower env.src_cpu's OPP.
7762 update_capacity_of(env.src_cpu);
7765 * We've detached some tasks from busiest_rq. Every
7766 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7767 * unlock busiest->lock, and we are able to be sure
7768 * that nobody can manipulate the tasks in parallel.
7769 * See task_rq_lock() family for the details.
7772 raw_spin_unlock(&busiest->lock);
7776 ld_moved += cur_ld_moved;
7779 local_irq_restore(flags);
7781 if (env.flags & LBF_NEED_BREAK) {
7782 env.flags &= ~LBF_NEED_BREAK;
7787 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7788 * us and move them to an alternate dst_cpu in our sched_group
7789 * where they can run. The upper limit on how many times we
7790 * iterate on same src_cpu is dependent on number of cpus in our
7793 * This changes load balance semantics a bit on who can move
7794 * load to a given_cpu. In addition to the given_cpu itself
7795 * (or a ilb_cpu acting on its behalf where given_cpu is
7796 * nohz-idle), we now have balance_cpu in a position to move
7797 * load to given_cpu. In rare situations, this may cause
7798 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7799 * _independently_ and at _same_ time to move some load to
7800 * given_cpu) causing exceess load to be moved to given_cpu.
7801 * This however should not happen so much in practice and
7802 * moreover subsequent load balance cycles should correct the
7803 * excess load moved.
7805 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7807 /* Prevent to re-select dst_cpu via env's cpus */
7808 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7810 env.dst_rq = cpu_rq(env.new_dst_cpu);
7811 env.dst_cpu = env.new_dst_cpu;
7812 env.flags &= ~LBF_DST_PINNED;
7814 env.loop_break = sched_nr_migrate_break;
7817 * Go back to "more_balance" rather than "redo" since we
7818 * need to continue with same src_cpu.
7824 * We failed to reach balance because of affinity.
7827 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7829 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7830 *group_imbalance = 1;
7833 /* All tasks on this runqueue were pinned by CPU affinity */
7834 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7835 cpumask_clear_cpu(cpu_of(busiest), cpus);
7836 if (!cpumask_empty(cpus)) {
7838 env.loop_break = sched_nr_migrate_break;
7841 goto out_all_pinned;
7846 schedstat_inc(sd, lb_failed[idle]);
7848 * Increment the failure counter only on periodic balance.
7849 * We do not want newidle balance, which can be very
7850 * frequent, pollute the failure counter causing
7851 * excessive cache_hot migrations and active balances.
7853 if (idle != CPU_NEWLY_IDLE)
7854 if (env.src_grp_nr_running > 1)
7855 sd->nr_balance_failed++;
7857 if (need_active_balance(&env)) {
7858 raw_spin_lock_irqsave(&busiest->lock, flags);
7860 /* don't kick the active_load_balance_cpu_stop,
7861 * if the curr task on busiest cpu can't be
7864 if (!cpumask_test_cpu(this_cpu,
7865 tsk_cpus_allowed(busiest->curr))) {
7866 raw_spin_unlock_irqrestore(&busiest->lock,
7868 env.flags |= LBF_ALL_PINNED;
7869 goto out_one_pinned;
7873 * ->active_balance synchronizes accesses to
7874 * ->active_balance_work. Once set, it's cleared
7875 * only after active load balance is finished.
7877 if (!busiest->active_balance) {
7878 busiest->active_balance = 1;
7879 busiest->push_cpu = this_cpu;
7882 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7884 if (active_balance) {
7885 stop_one_cpu_nowait(cpu_of(busiest),
7886 active_load_balance_cpu_stop, busiest,
7887 &busiest->active_balance_work);
7891 * We've kicked active balancing, reset the failure
7894 sd->nr_balance_failed = sd->cache_nice_tries+1;
7897 sd->nr_balance_failed = 0;
7899 if (likely(!active_balance)) {
7900 /* We were unbalanced, so reset the balancing interval */
7901 sd->balance_interval = sd->min_interval;
7904 * If we've begun active balancing, start to back off. This
7905 * case may not be covered by the all_pinned logic if there
7906 * is only 1 task on the busy runqueue (because we don't call
7909 if (sd->balance_interval < sd->max_interval)
7910 sd->balance_interval *= 2;
7917 * We reach balance although we may have faced some affinity
7918 * constraints. Clear the imbalance flag if it was set.
7921 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7923 if (*group_imbalance)
7924 *group_imbalance = 0;
7929 * We reach balance because all tasks are pinned at this level so
7930 * we can't migrate them. Let the imbalance flag set so parent level
7931 * can try to migrate them.
7933 schedstat_inc(sd, lb_balanced[idle]);
7935 sd->nr_balance_failed = 0;
7938 /* tune up the balancing interval */
7939 if (((env.flags & LBF_ALL_PINNED) &&
7940 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7941 (sd->balance_interval < sd->max_interval))
7942 sd->balance_interval *= 2;
7949 static inline unsigned long
7950 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7952 unsigned long interval = sd->balance_interval;
7955 interval *= sd->busy_factor;
7957 /* scale ms to jiffies */
7958 interval = msecs_to_jiffies(interval);
7959 interval = clamp(interval, 1UL, max_load_balance_interval);
7965 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7967 unsigned long interval, next;
7969 interval = get_sd_balance_interval(sd, cpu_busy);
7970 next = sd->last_balance + interval;
7972 if (time_after(*next_balance, next))
7973 *next_balance = next;
7977 * idle_balance is called by schedule() if this_cpu is about to become
7978 * idle. Attempts to pull tasks from other CPUs.
7980 static int idle_balance(struct rq *this_rq)
7982 unsigned long next_balance = jiffies + HZ;
7983 int this_cpu = this_rq->cpu;
7984 struct sched_domain *sd;
7985 int pulled_task = 0;
7988 idle_enter_fair(this_rq);
7991 * We must set idle_stamp _before_ calling idle_balance(), such that we
7992 * measure the duration of idle_balance() as idle time.
7994 this_rq->idle_stamp = rq_clock(this_rq);
7996 if (!energy_aware() &&
7997 (this_rq->avg_idle < sysctl_sched_migration_cost ||
7998 !this_rq->rd->overload)) {
8000 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8002 update_next_balance(sd, 0, &next_balance);
8008 raw_spin_unlock(&this_rq->lock);
8010 update_blocked_averages(this_cpu);
8012 for_each_domain(this_cpu, sd) {
8013 int continue_balancing = 1;
8014 u64 t0, domain_cost;
8016 if (!(sd->flags & SD_LOAD_BALANCE))
8019 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8020 update_next_balance(sd, 0, &next_balance);
8024 if (sd->flags & SD_BALANCE_NEWIDLE) {
8025 t0 = sched_clock_cpu(this_cpu);
8027 pulled_task = load_balance(this_cpu, this_rq,
8029 &continue_balancing);
8031 domain_cost = sched_clock_cpu(this_cpu) - t0;
8032 if (domain_cost > sd->max_newidle_lb_cost)
8033 sd->max_newidle_lb_cost = domain_cost;
8035 curr_cost += domain_cost;
8038 update_next_balance(sd, 0, &next_balance);
8041 * Stop searching for tasks to pull if there are
8042 * now runnable tasks on this rq.
8044 if (pulled_task || this_rq->nr_running > 0)
8049 raw_spin_lock(&this_rq->lock);
8051 if (curr_cost > this_rq->max_idle_balance_cost)
8052 this_rq->max_idle_balance_cost = curr_cost;
8055 * While browsing the domains, we released the rq lock, a task could
8056 * have been enqueued in the meantime. Since we're not going idle,
8057 * pretend we pulled a task.
8059 if (this_rq->cfs.h_nr_running && !pulled_task)
8063 /* Move the next balance forward */
8064 if (time_after(this_rq->next_balance, next_balance))
8065 this_rq->next_balance = next_balance;
8067 /* Is there a task of a high priority class? */
8068 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8072 idle_exit_fair(this_rq);
8073 this_rq->idle_stamp = 0;
8080 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8081 * running tasks off the busiest CPU onto idle CPUs. It requires at
8082 * least 1 task to be running on each physical CPU where possible, and
8083 * avoids physical / logical imbalances.
8085 static int active_load_balance_cpu_stop(void *data)
8087 struct rq *busiest_rq = data;
8088 int busiest_cpu = cpu_of(busiest_rq);
8089 int target_cpu = busiest_rq->push_cpu;
8090 struct rq *target_rq = cpu_rq(target_cpu);
8091 struct sched_domain *sd;
8092 struct task_struct *p = NULL;
8094 raw_spin_lock_irq(&busiest_rq->lock);
8096 /* make sure the requested cpu hasn't gone down in the meantime */
8097 if (unlikely(busiest_cpu != smp_processor_id() ||
8098 !busiest_rq->active_balance))
8101 /* Is there any task to move? */
8102 if (busiest_rq->nr_running <= 1)
8106 * This condition is "impossible", if it occurs
8107 * we need to fix it. Originally reported by
8108 * Bjorn Helgaas on a 128-cpu setup.
8110 BUG_ON(busiest_rq == target_rq);
8112 /* Search for an sd spanning us and the target CPU. */
8114 for_each_domain(target_cpu, sd) {
8115 if ((sd->flags & SD_LOAD_BALANCE) &&
8116 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8121 struct lb_env env = {
8123 .dst_cpu = target_cpu,
8124 .dst_rq = target_rq,
8125 .src_cpu = busiest_rq->cpu,
8126 .src_rq = busiest_rq,
8130 schedstat_inc(sd, alb_count);
8132 p = detach_one_task(&env);
8134 schedstat_inc(sd, alb_pushed);
8136 * We want to potentially lower env.src_cpu's OPP.
8138 update_capacity_of(env.src_cpu);
8141 schedstat_inc(sd, alb_failed);
8145 busiest_rq->active_balance = 0;
8146 raw_spin_unlock(&busiest_rq->lock);
8149 attach_one_task(target_rq, p);
8156 static inline int on_null_domain(struct rq *rq)
8158 return unlikely(!rcu_dereference_sched(rq->sd));
8161 #ifdef CONFIG_NO_HZ_COMMON
8163 * idle load balancing details
8164 * - When one of the busy CPUs notice that there may be an idle rebalancing
8165 * needed, they will kick the idle load balancer, which then does idle
8166 * load balancing for all the idle CPUs.
8169 cpumask_var_t idle_cpus_mask;
8171 unsigned long next_balance; /* in jiffy units */
8172 } nohz ____cacheline_aligned;
8174 static inline int find_new_ilb(void)
8176 int ilb = cpumask_first(nohz.idle_cpus_mask);
8178 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8185 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8186 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8187 * CPU (if there is one).
8189 static void nohz_balancer_kick(void)
8193 nohz.next_balance++;
8195 ilb_cpu = find_new_ilb();
8197 if (ilb_cpu >= nr_cpu_ids)
8200 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8203 * Use smp_send_reschedule() instead of resched_cpu().
8204 * This way we generate a sched IPI on the target cpu which
8205 * is idle. And the softirq performing nohz idle load balance
8206 * will be run before returning from the IPI.
8208 smp_send_reschedule(ilb_cpu);
8212 static inline void nohz_balance_exit_idle(int cpu)
8214 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8216 * Completely isolated CPUs don't ever set, so we must test.
8218 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8219 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8220 atomic_dec(&nohz.nr_cpus);
8222 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8226 static inline void set_cpu_sd_state_busy(void)
8228 struct sched_domain *sd;
8229 int cpu = smp_processor_id();
8232 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8234 if (!sd || !sd->nohz_idle)
8238 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8243 void set_cpu_sd_state_idle(void)
8245 struct sched_domain *sd;
8246 int cpu = smp_processor_id();
8249 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8251 if (!sd || sd->nohz_idle)
8255 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8261 * This routine will record that the cpu is going idle with tick stopped.
8262 * This info will be used in performing idle load balancing in the future.
8264 void nohz_balance_enter_idle(int cpu)
8267 * If this cpu is going down, then nothing needs to be done.
8269 if (!cpu_active(cpu))
8272 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8276 * If we're a completely isolated CPU, we don't play.
8278 if (on_null_domain(cpu_rq(cpu)))
8281 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8282 atomic_inc(&nohz.nr_cpus);
8283 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8286 static int sched_ilb_notifier(struct notifier_block *nfb,
8287 unsigned long action, void *hcpu)
8289 switch (action & ~CPU_TASKS_FROZEN) {
8291 nohz_balance_exit_idle(smp_processor_id());
8299 static DEFINE_SPINLOCK(balancing);
8302 * Scale the max load_balance interval with the number of CPUs in the system.
8303 * This trades load-balance latency on larger machines for less cross talk.
8305 void update_max_interval(void)
8307 max_load_balance_interval = HZ*num_online_cpus()/10;
8311 * It checks each scheduling domain to see if it is due to be balanced,
8312 * and initiates a balancing operation if so.
8314 * Balancing parameters are set up in init_sched_domains.
8316 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8318 int continue_balancing = 1;
8320 unsigned long interval;
8321 struct sched_domain *sd;
8322 /* Earliest time when we have to do rebalance again */
8323 unsigned long next_balance = jiffies + 60*HZ;
8324 int update_next_balance = 0;
8325 int need_serialize, need_decay = 0;
8328 update_blocked_averages(cpu);
8331 for_each_domain(cpu, sd) {
8333 * Decay the newidle max times here because this is a regular
8334 * visit to all the domains. Decay ~1% per second.
8336 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8337 sd->max_newidle_lb_cost =
8338 (sd->max_newidle_lb_cost * 253) / 256;
8339 sd->next_decay_max_lb_cost = jiffies + HZ;
8342 max_cost += sd->max_newidle_lb_cost;
8344 if (!(sd->flags & SD_LOAD_BALANCE))
8348 * Stop the load balance at this level. There is another
8349 * CPU in our sched group which is doing load balancing more
8352 if (!continue_balancing) {
8358 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8360 need_serialize = sd->flags & SD_SERIALIZE;
8361 if (need_serialize) {
8362 if (!spin_trylock(&balancing))
8366 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8367 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8369 * The LBF_DST_PINNED logic could have changed
8370 * env->dst_cpu, so we can't know our idle
8371 * state even if we migrated tasks. Update it.
8373 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8375 sd->last_balance = jiffies;
8376 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8379 spin_unlock(&balancing);
8381 if (time_after(next_balance, sd->last_balance + interval)) {
8382 next_balance = sd->last_balance + interval;
8383 update_next_balance = 1;
8388 * Ensure the rq-wide value also decays but keep it at a
8389 * reasonable floor to avoid funnies with rq->avg_idle.
8391 rq->max_idle_balance_cost =
8392 max((u64)sysctl_sched_migration_cost, max_cost);
8397 * next_balance will be updated only when there is a need.
8398 * When the cpu is attached to null domain for ex, it will not be
8401 if (likely(update_next_balance)) {
8402 rq->next_balance = next_balance;
8404 #ifdef CONFIG_NO_HZ_COMMON
8406 * If this CPU has been elected to perform the nohz idle
8407 * balance. Other idle CPUs have already rebalanced with
8408 * nohz_idle_balance() and nohz.next_balance has been
8409 * updated accordingly. This CPU is now running the idle load
8410 * balance for itself and we need to update the
8411 * nohz.next_balance accordingly.
8413 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8414 nohz.next_balance = rq->next_balance;
8419 #ifdef CONFIG_NO_HZ_COMMON
8421 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8422 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8424 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8426 int this_cpu = this_rq->cpu;
8429 /* Earliest time when we have to do rebalance again */
8430 unsigned long next_balance = jiffies + 60*HZ;
8431 int update_next_balance = 0;
8433 if (idle != CPU_IDLE ||
8434 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8437 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8438 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8442 * If this cpu gets work to do, stop the load balancing
8443 * work being done for other cpus. Next load
8444 * balancing owner will pick it up.
8449 rq = cpu_rq(balance_cpu);
8452 * If time for next balance is due,
8455 if (time_after_eq(jiffies, rq->next_balance)) {
8456 raw_spin_lock_irq(&rq->lock);
8457 update_rq_clock(rq);
8458 update_idle_cpu_load(rq);
8459 raw_spin_unlock_irq(&rq->lock);
8460 rebalance_domains(rq, CPU_IDLE);
8463 if (time_after(next_balance, rq->next_balance)) {
8464 next_balance = rq->next_balance;
8465 update_next_balance = 1;
8470 * next_balance will be updated only when there is a need.
8471 * When the CPU is attached to null domain for ex, it will not be
8474 if (likely(update_next_balance))
8475 nohz.next_balance = next_balance;
8477 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8481 * Current heuristic for kicking the idle load balancer in the presence
8482 * of an idle cpu in the system.
8483 * - This rq has more than one task.
8484 * - This rq has at least one CFS task and the capacity of the CPU is
8485 * significantly reduced because of RT tasks or IRQs.
8486 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8487 * multiple busy cpu.
8488 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8489 * domain span are idle.
8491 static inline bool nohz_kick_needed(struct rq *rq)
8493 unsigned long now = jiffies;
8494 struct sched_domain *sd;
8495 struct sched_group_capacity *sgc;
8496 int nr_busy, cpu = rq->cpu;
8499 if (unlikely(rq->idle_balance))
8503 * We may be recently in ticked or tickless idle mode. At the first
8504 * busy tick after returning from idle, we will update the busy stats.
8506 set_cpu_sd_state_busy();
8507 nohz_balance_exit_idle(cpu);
8510 * None are in tickless mode and hence no need for NOHZ idle load
8513 if (likely(!atomic_read(&nohz.nr_cpus)))
8516 if (time_before(now, nohz.next_balance))
8519 if (rq->nr_running >= 2 &&
8520 (!energy_aware() || cpu_overutilized(cpu)))
8524 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8525 if (sd && !energy_aware()) {
8526 sgc = sd->groups->sgc;
8527 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8536 sd = rcu_dereference(rq->sd);
8538 if ((rq->cfs.h_nr_running >= 1) &&
8539 check_cpu_capacity(rq, sd)) {
8545 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8546 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8547 sched_domain_span(sd)) < cpu)) {
8557 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8561 * run_rebalance_domains is triggered when needed from the scheduler tick.
8562 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8564 static void run_rebalance_domains(struct softirq_action *h)
8566 struct rq *this_rq = this_rq();
8567 enum cpu_idle_type idle = this_rq->idle_balance ?
8568 CPU_IDLE : CPU_NOT_IDLE;
8571 * If this cpu has a pending nohz_balance_kick, then do the
8572 * balancing on behalf of the other idle cpus whose ticks are
8573 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8574 * give the idle cpus a chance to load balance. Else we may
8575 * load balance only within the local sched_domain hierarchy
8576 * and abort nohz_idle_balance altogether if we pull some load.
8578 nohz_idle_balance(this_rq, idle);
8579 rebalance_domains(this_rq, idle);
8583 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8585 void trigger_load_balance(struct rq *rq)
8587 /* Don't need to rebalance while attached to NULL domain */
8588 if (unlikely(on_null_domain(rq)))
8591 if (time_after_eq(jiffies, rq->next_balance))
8592 raise_softirq(SCHED_SOFTIRQ);
8593 #ifdef CONFIG_NO_HZ_COMMON
8594 if (nohz_kick_needed(rq))
8595 nohz_balancer_kick();
8599 static void rq_online_fair(struct rq *rq)
8603 update_runtime_enabled(rq);
8606 static void rq_offline_fair(struct rq *rq)
8610 /* Ensure any throttled groups are reachable by pick_next_task */
8611 unthrottle_offline_cfs_rqs(rq);
8614 #endif /* CONFIG_SMP */
8617 * scheduler tick hitting a task of our scheduling class:
8619 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8621 struct cfs_rq *cfs_rq;
8622 struct sched_entity *se = &curr->se;
8624 for_each_sched_entity(se) {
8625 cfs_rq = cfs_rq_of(se);
8626 entity_tick(cfs_rq, se, queued);
8629 if (static_branch_unlikely(&sched_numa_balancing))
8630 task_tick_numa(rq, curr);
8632 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8633 rq->rd->overutilized = true;
8635 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8639 * called on fork with the child task as argument from the parent's context
8640 * - child not yet on the tasklist
8641 * - preemption disabled
8643 static void task_fork_fair(struct task_struct *p)
8645 struct cfs_rq *cfs_rq;
8646 struct sched_entity *se = &p->se, *curr;
8647 int this_cpu = smp_processor_id();
8648 struct rq *rq = this_rq();
8649 unsigned long flags;
8651 raw_spin_lock_irqsave(&rq->lock, flags);
8653 update_rq_clock(rq);
8655 cfs_rq = task_cfs_rq(current);
8656 curr = cfs_rq->curr;
8659 * Not only the cpu but also the task_group of the parent might have
8660 * been changed after parent->se.parent,cfs_rq were copied to
8661 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8662 * of child point to valid ones.
8665 __set_task_cpu(p, this_cpu);
8668 update_curr(cfs_rq);
8671 se->vruntime = curr->vruntime;
8672 place_entity(cfs_rq, se, 1);
8674 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8676 * Upon rescheduling, sched_class::put_prev_task() will place
8677 * 'current' within the tree based on its new key value.
8679 swap(curr->vruntime, se->vruntime);
8683 se->vruntime -= cfs_rq->min_vruntime;
8685 raw_spin_unlock_irqrestore(&rq->lock, flags);
8689 * Priority of the task has changed. Check to see if we preempt
8693 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8695 if (!task_on_rq_queued(p))
8699 * Reschedule if we are currently running on this runqueue and
8700 * our priority decreased, or if we are not currently running on
8701 * this runqueue and our priority is higher than the current's
8703 if (rq->curr == p) {
8704 if (p->prio > oldprio)
8707 check_preempt_curr(rq, p, 0);
8710 static inline bool vruntime_normalized(struct task_struct *p)
8712 struct sched_entity *se = &p->se;
8715 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8716 * the dequeue_entity(.flags=0) will already have normalized the
8723 * When !on_rq, vruntime of the task has usually NOT been normalized.
8724 * But there are some cases where it has already been normalized:
8726 * - A forked child which is waiting for being woken up by
8727 * wake_up_new_task().
8728 * - A task which has been woken up by try_to_wake_up() and
8729 * waiting for actually being woken up by sched_ttwu_pending().
8731 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8737 static void detach_task_cfs_rq(struct task_struct *p)
8739 struct sched_entity *se = &p->se;
8740 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8742 if (!vruntime_normalized(p)) {
8744 * Fix up our vruntime so that the current sleep doesn't
8745 * cause 'unlimited' sleep bonus.
8747 place_entity(cfs_rq, se, 0);
8748 se->vruntime -= cfs_rq->min_vruntime;
8751 /* Catch up with the cfs_rq and remove our load when we leave */
8752 detach_entity_load_avg(cfs_rq, se);
8755 static void attach_task_cfs_rq(struct task_struct *p)
8757 struct sched_entity *se = &p->se;
8758 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8760 #ifdef CONFIG_FAIR_GROUP_SCHED
8762 * Since the real-depth could have been changed (only FAIR
8763 * class maintain depth value), reset depth properly.
8765 se->depth = se->parent ? se->parent->depth + 1 : 0;
8768 /* Synchronize task with its cfs_rq */
8769 attach_entity_load_avg(cfs_rq, se);
8771 if (!vruntime_normalized(p))
8772 se->vruntime += cfs_rq->min_vruntime;
8775 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8777 detach_task_cfs_rq(p);
8780 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8782 attach_task_cfs_rq(p);
8784 if (task_on_rq_queued(p)) {
8786 * We were most likely switched from sched_rt, so
8787 * kick off the schedule if running, otherwise just see
8788 * if we can still preempt the current task.
8793 check_preempt_curr(rq, p, 0);
8797 /* Account for a task changing its policy or group.
8799 * This routine is mostly called to set cfs_rq->curr field when a task
8800 * migrates between groups/classes.
8802 static void set_curr_task_fair(struct rq *rq)
8804 struct sched_entity *se = &rq->curr->se;
8806 for_each_sched_entity(se) {
8807 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8809 set_next_entity(cfs_rq, se);
8810 /* ensure bandwidth has been allocated on our new cfs_rq */
8811 account_cfs_rq_runtime(cfs_rq, 0);
8815 void init_cfs_rq(struct cfs_rq *cfs_rq)
8817 cfs_rq->tasks_timeline = RB_ROOT;
8818 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8819 #ifndef CONFIG_64BIT
8820 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8823 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8824 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8828 #ifdef CONFIG_FAIR_GROUP_SCHED
8829 static void task_move_group_fair(struct task_struct *p)
8831 detach_task_cfs_rq(p);
8832 set_task_rq(p, task_cpu(p));
8835 /* Tell se's cfs_rq has been changed -- migrated */
8836 p->se.avg.last_update_time = 0;
8838 attach_task_cfs_rq(p);
8841 void free_fair_sched_group(struct task_group *tg)
8845 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8847 for_each_possible_cpu(i) {
8849 kfree(tg->cfs_rq[i]);
8852 remove_entity_load_avg(tg->se[i]);
8861 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8863 struct cfs_rq *cfs_rq;
8864 struct sched_entity *se;
8867 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8870 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8874 tg->shares = NICE_0_LOAD;
8876 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8878 for_each_possible_cpu(i) {
8879 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8880 GFP_KERNEL, cpu_to_node(i));
8884 se = kzalloc_node(sizeof(struct sched_entity),
8885 GFP_KERNEL, cpu_to_node(i));
8889 init_cfs_rq(cfs_rq);
8890 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8891 init_entity_runnable_average(se);
8902 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8904 struct rq *rq = cpu_rq(cpu);
8905 unsigned long flags;
8908 * Only empty task groups can be destroyed; so we can speculatively
8909 * check on_list without danger of it being re-added.
8911 if (!tg->cfs_rq[cpu]->on_list)
8914 raw_spin_lock_irqsave(&rq->lock, flags);
8915 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8916 raw_spin_unlock_irqrestore(&rq->lock, flags);
8919 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8920 struct sched_entity *se, int cpu,
8921 struct sched_entity *parent)
8923 struct rq *rq = cpu_rq(cpu);
8927 init_cfs_rq_runtime(cfs_rq);
8929 tg->cfs_rq[cpu] = cfs_rq;
8932 /* se could be NULL for root_task_group */
8937 se->cfs_rq = &rq->cfs;
8940 se->cfs_rq = parent->my_q;
8941 se->depth = parent->depth + 1;
8945 /* guarantee group entities always have weight */
8946 update_load_set(&se->load, NICE_0_LOAD);
8947 se->parent = parent;
8950 static DEFINE_MUTEX(shares_mutex);
8952 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8955 unsigned long flags;
8958 * We can't change the weight of the root cgroup.
8963 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8965 mutex_lock(&shares_mutex);
8966 if (tg->shares == shares)
8969 tg->shares = shares;
8970 for_each_possible_cpu(i) {
8971 struct rq *rq = cpu_rq(i);
8972 struct sched_entity *se;
8975 /* Propagate contribution to hierarchy */
8976 raw_spin_lock_irqsave(&rq->lock, flags);
8978 /* Possible calls to update_curr() need rq clock */
8979 update_rq_clock(rq);
8980 for_each_sched_entity(se)
8981 update_cfs_shares(group_cfs_rq(se));
8982 raw_spin_unlock_irqrestore(&rq->lock, flags);
8986 mutex_unlock(&shares_mutex);
8989 #else /* CONFIG_FAIR_GROUP_SCHED */
8991 void free_fair_sched_group(struct task_group *tg) { }
8993 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8998 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9000 #endif /* CONFIG_FAIR_GROUP_SCHED */
9003 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9005 struct sched_entity *se = &task->se;
9006 unsigned int rr_interval = 0;
9009 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9012 if (rq->cfs.load.weight)
9013 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9019 * All the scheduling class methods:
9021 const struct sched_class fair_sched_class = {
9022 .next = &idle_sched_class,
9023 .enqueue_task = enqueue_task_fair,
9024 .dequeue_task = dequeue_task_fair,
9025 .yield_task = yield_task_fair,
9026 .yield_to_task = yield_to_task_fair,
9028 .check_preempt_curr = check_preempt_wakeup,
9030 .pick_next_task = pick_next_task_fair,
9031 .put_prev_task = put_prev_task_fair,
9034 .select_task_rq = select_task_rq_fair,
9035 .migrate_task_rq = migrate_task_rq_fair,
9037 .rq_online = rq_online_fair,
9038 .rq_offline = rq_offline_fair,
9040 .task_waking = task_waking_fair,
9041 .task_dead = task_dead_fair,
9042 .set_cpus_allowed = set_cpus_allowed_common,
9045 .set_curr_task = set_curr_task_fair,
9046 .task_tick = task_tick_fair,
9047 .task_fork = task_fork_fair,
9049 .prio_changed = prio_changed_fair,
9050 .switched_from = switched_from_fair,
9051 .switched_to = switched_to_fair,
9053 .get_rr_interval = get_rr_interval_fair,
9055 .update_curr = update_curr_fair,
9057 #ifdef CONFIG_FAIR_GROUP_SCHED
9058 .task_move_group = task_move_group_fair,
9062 #ifdef CONFIG_SCHED_DEBUG
9063 void print_cfs_stats(struct seq_file *m, int cpu)
9065 struct cfs_rq *cfs_rq;
9068 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9069 print_cfs_rq(m, cpu, cfs_rq);
9073 #ifdef CONFIG_NUMA_BALANCING
9074 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9077 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9079 for_each_online_node(node) {
9080 if (p->numa_faults) {
9081 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9082 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9084 if (p->numa_group) {
9085 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9086 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9088 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9091 #endif /* CONFIG_NUMA_BALANCING */
9092 #endif /* CONFIG_SCHED_DEBUG */
9094 __init void init_sched_fair_class(void)
9097 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9099 #ifdef CONFIG_NO_HZ_COMMON
9100 nohz.next_balance = jiffies;
9101 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9102 cpu_notifier(sched_ilb_notifier, 0);