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);
2588 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2590 /* delta_w is the amount already accumulated against our next period */
2591 delta_w = sa->period_contrib;
2592 if (delta + delta_w >= 1024) {
2595 /* how much left for next period will start over, we don't know yet */
2596 sa->period_contrib = 0;
2599 * Now that we know we're crossing a period boundary, figure
2600 * out how much from delta we need to complete the current
2601 * period and accrue it.
2603 delta_w = 1024 - delta_w;
2604 scaled_delta_w = cap_scale(delta_w, scale_freq);
2606 sa->load_sum += weight * scaled_delta_w;
2608 cfs_rq->runnable_load_sum +=
2609 weight * scaled_delta_w;
2613 sa->util_sum += scaled_delta_w * scale_cpu;
2617 /* Figure out how many additional periods this update spans */
2618 periods = delta / 1024;
2621 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2623 cfs_rq->runnable_load_sum =
2624 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2626 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2628 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2629 contrib = __compute_runnable_contrib(periods);
2630 contrib = cap_scale(contrib, scale_freq);
2632 sa->load_sum += weight * contrib;
2634 cfs_rq->runnable_load_sum += weight * contrib;
2637 sa->util_sum += contrib * scale_cpu;
2640 /* Remainder of delta accrued against u_0` */
2641 scaled_delta = cap_scale(delta, scale_freq);
2643 sa->load_sum += weight * scaled_delta;
2645 cfs_rq->runnable_load_sum += weight * scaled_delta;
2648 sa->util_sum += scaled_delta * scale_cpu;
2650 sa->period_contrib += delta;
2653 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2655 cfs_rq->runnable_load_avg =
2656 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2658 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2664 #ifdef CONFIG_FAIR_GROUP_SCHED
2666 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2667 * and effective_load (which is not done because it is too costly).
2669 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2671 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2673 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2674 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2675 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2679 #else /* CONFIG_FAIR_GROUP_SCHED */
2680 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2681 #endif /* CONFIG_FAIR_GROUP_SCHED */
2683 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2686 * Unsigned subtract and clamp on underflow.
2688 * Explicitly do a load-store to ensure the intermediate value never hits
2689 * memory. This allows lockless observations without ever seeing the negative
2692 #define sub_positive(_ptr, _val) do { \
2693 typeof(_ptr) ptr = (_ptr); \
2694 typeof(*ptr) val = (_val); \
2695 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2699 WRITE_ONCE(*ptr, res); \
2702 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2703 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2705 struct sched_avg *sa = &cfs_rq->avg;
2706 int decayed, removed = 0;
2708 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2709 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2710 sub_positive(&sa->load_avg, r);
2711 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2715 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2716 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2717 sub_positive(&sa->util_avg, r);
2718 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2721 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2722 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2724 #ifndef CONFIG_64BIT
2726 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2729 return decayed || removed;
2732 /* Update task and its cfs_rq load average */
2733 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2735 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2736 u64 now = cfs_rq_clock_task(cfs_rq);
2737 int cpu = cpu_of(rq_of(cfs_rq));
2740 * Track task load average for carrying it to new CPU after migrated, and
2741 * track group sched_entity load average for task_h_load calc in migration
2743 __update_load_avg(now, cpu, &se->avg,
2744 se->on_rq * scale_load_down(se->load.weight),
2745 cfs_rq->curr == se, NULL);
2747 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2748 update_tg_load_avg(cfs_rq, 0);
2750 if (entity_is_task(se))
2751 trace_sched_load_avg_task(task_of(se), &se->avg);
2752 trace_sched_load_avg_cpu(cpu, cfs_rq);
2755 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2757 if (!sched_feat(ATTACH_AGE_LOAD))
2761 * If we got migrated (either between CPUs or between cgroups) we'll
2762 * have aged the average right before clearing @last_update_time.
2764 if (se->avg.last_update_time) {
2765 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2766 &se->avg, 0, 0, NULL);
2769 * XXX: we could have just aged the entire load away if we've been
2770 * absent from the fair class for too long.
2775 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2776 cfs_rq->avg.load_avg += se->avg.load_avg;
2777 cfs_rq->avg.load_sum += se->avg.load_sum;
2778 cfs_rq->avg.util_avg += se->avg.util_avg;
2779 cfs_rq->avg.util_sum += se->avg.util_sum;
2782 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2784 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2785 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2786 cfs_rq->curr == se, NULL);
2788 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2789 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2790 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2791 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2794 /* Add the load generated by se into cfs_rq's load average */
2796 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2798 struct sched_avg *sa = &se->avg;
2799 u64 now = cfs_rq_clock_task(cfs_rq);
2800 int migrated, decayed;
2802 migrated = !sa->last_update_time;
2804 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2805 se->on_rq * scale_load_down(se->load.weight),
2806 cfs_rq->curr == se, NULL);
2809 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2811 cfs_rq->runnable_load_avg += sa->load_avg;
2812 cfs_rq->runnable_load_sum += sa->load_sum;
2815 attach_entity_load_avg(cfs_rq, se);
2817 if (decayed || migrated)
2818 update_tg_load_avg(cfs_rq, 0);
2821 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2823 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2825 update_load_avg(se, 1);
2827 cfs_rq->runnable_load_avg =
2828 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2829 cfs_rq->runnable_load_sum =
2830 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2833 #ifndef CONFIG_64BIT
2834 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2836 u64 last_update_time_copy;
2837 u64 last_update_time;
2840 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2842 last_update_time = cfs_rq->avg.last_update_time;
2843 } while (last_update_time != last_update_time_copy);
2845 return last_update_time;
2848 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2850 return cfs_rq->avg.last_update_time;
2855 * Task first catches up with cfs_rq, and then subtract
2856 * itself from the cfs_rq (task must be off the queue now).
2858 void remove_entity_load_avg(struct sched_entity *se)
2860 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2861 u64 last_update_time;
2864 * Newly created task or never used group entity should not be removed
2865 * from its (source) cfs_rq
2867 if (se->avg.last_update_time == 0)
2870 last_update_time = cfs_rq_last_update_time(cfs_rq);
2872 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2873 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2874 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2878 * Update the rq's load with the elapsed running time before entering
2879 * idle. if the last scheduled task is not a CFS task, idle_enter will
2880 * be the only way to update the runnable statistic.
2882 void idle_enter_fair(struct rq *this_rq)
2887 * Update the rq's load with the elapsed idle time before a task is
2888 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2889 * be the only way to update the runnable statistic.
2891 void idle_exit_fair(struct rq *this_rq)
2895 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2897 return cfs_rq->runnable_load_avg;
2900 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2902 return cfs_rq->avg.load_avg;
2905 static int idle_balance(struct rq *this_rq);
2907 #else /* CONFIG_SMP */
2909 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2911 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2913 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2914 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2917 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2919 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2921 static inline int idle_balance(struct rq *rq)
2926 #endif /* CONFIG_SMP */
2928 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2930 #ifdef CONFIG_SCHEDSTATS
2931 struct task_struct *tsk = NULL;
2933 if (entity_is_task(se))
2936 if (se->statistics.sleep_start) {
2937 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2942 if (unlikely(delta > se->statistics.sleep_max))
2943 se->statistics.sleep_max = delta;
2945 se->statistics.sleep_start = 0;
2946 se->statistics.sum_sleep_runtime += delta;
2949 account_scheduler_latency(tsk, delta >> 10, 1);
2950 trace_sched_stat_sleep(tsk, delta);
2953 if (se->statistics.block_start) {
2954 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2959 if (unlikely(delta > se->statistics.block_max))
2960 se->statistics.block_max = delta;
2962 se->statistics.block_start = 0;
2963 se->statistics.sum_sleep_runtime += delta;
2966 if (tsk->in_iowait) {
2967 se->statistics.iowait_sum += delta;
2968 se->statistics.iowait_count++;
2969 trace_sched_stat_iowait(tsk, delta);
2972 trace_sched_stat_blocked(tsk, delta);
2973 trace_sched_blocked_reason(tsk);
2976 * Blocking time is in units of nanosecs, so shift by
2977 * 20 to get a milliseconds-range estimation of the
2978 * amount of time that the task spent sleeping:
2980 if (unlikely(prof_on == SLEEP_PROFILING)) {
2981 profile_hits(SLEEP_PROFILING,
2982 (void *)get_wchan(tsk),
2985 account_scheduler_latency(tsk, delta >> 10, 0);
2991 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2993 #ifdef CONFIG_SCHED_DEBUG
2994 s64 d = se->vruntime - cfs_rq->min_vruntime;
2999 if (d > 3*sysctl_sched_latency)
3000 schedstat_inc(cfs_rq, nr_spread_over);
3005 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3007 u64 vruntime = cfs_rq->min_vruntime;
3010 * The 'current' period is already promised to the current tasks,
3011 * however the extra weight of the new task will slow them down a
3012 * little, place the new task so that it fits in the slot that
3013 * stays open at the end.
3015 if (initial && sched_feat(START_DEBIT))
3016 vruntime += sched_vslice(cfs_rq, se);
3018 /* sleeps up to a single latency don't count. */
3020 unsigned long thresh = sysctl_sched_latency;
3023 * Halve their sleep time's effect, to allow
3024 * for a gentler effect of sleepers:
3026 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3032 /* ensure we never gain time by being placed backwards. */
3033 se->vruntime = max_vruntime(se->vruntime, vruntime);
3036 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3039 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3042 * Update the normalized vruntime before updating min_vruntime
3043 * through calling update_curr().
3045 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3046 se->vruntime += cfs_rq->min_vruntime;
3049 * Update run-time statistics of the 'current'.
3051 update_curr(cfs_rq);
3052 enqueue_entity_load_avg(cfs_rq, se);
3053 account_entity_enqueue(cfs_rq, se);
3054 update_cfs_shares(cfs_rq);
3056 if (flags & ENQUEUE_WAKEUP) {
3057 place_entity(cfs_rq, se, 0);
3058 enqueue_sleeper(cfs_rq, se);
3061 update_stats_enqueue(cfs_rq, se);
3062 check_spread(cfs_rq, se);
3063 if (se != cfs_rq->curr)
3064 __enqueue_entity(cfs_rq, se);
3067 if (cfs_rq->nr_running == 1) {
3068 list_add_leaf_cfs_rq(cfs_rq);
3069 check_enqueue_throttle(cfs_rq);
3073 static void __clear_buddies_last(struct sched_entity *se)
3075 for_each_sched_entity(se) {
3076 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3077 if (cfs_rq->last != se)
3080 cfs_rq->last = NULL;
3084 static void __clear_buddies_next(struct sched_entity *se)
3086 for_each_sched_entity(se) {
3087 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3088 if (cfs_rq->next != se)
3091 cfs_rq->next = NULL;
3095 static void __clear_buddies_skip(struct sched_entity *se)
3097 for_each_sched_entity(se) {
3098 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3099 if (cfs_rq->skip != se)
3102 cfs_rq->skip = NULL;
3106 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3108 if (cfs_rq->last == se)
3109 __clear_buddies_last(se);
3111 if (cfs_rq->next == se)
3112 __clear_buddies_next(se);
3114 if (cfs_rq->skip == se)
3115 __clear_buddies_skip(se);
3118 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3121 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3124 * Update run-time statistics of the 'current'.
3126 update_curr(cfs_rq);
3127 dequeue_entity_load_avg(cfs_rq, se);
3129 update_stats_dequeue(cfs_rq, se);
3130 if (flags & DEQUEUE_SLEEP) {
3131 #ifdef CONFIG_SCHEDSTATS
3132 if (entity_is_task(se)) {
3133 struct task_struct *tsk = task_of(se);
3135 if (tsk->state & TASK_INTERRUPTIBLE)
3136 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3137 if (tsk->state & TASK_UNINTERRUPTIBLE)
3138 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3143 clear_buddies(cfs_rq, se);
3145 if (se != cfs_rq->curr)
3146 __dequeue_entity(cfs_rq, se);
3148 account_entity_dequeue(cfs_rq, se);
3151 * Normalize the entity after updating the min_vruntime because the
3152 * update can refer to the ->curr item and we need to reflect this
3153 * movement in our normalized position.
3155 if (!(flags & DEQUEUE_SLEEP))
3156 se->vruntime -= cfs_rq->min_vruntime;
3158 /* return excess runtime on last dequeue */
3159 return_cfs_rq_runtime(cfs_rq);
3161 update_min_vruntime(cfs_rq);
3162 update_cfs_shares(cfs_rq);
3166 * Preempt the current task with a newly woken task if needed:
3169 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3171 unsigned long ideal_runtime, delta_exec;
3172 struct sched_entity *se;
3175 ideal_runtime = sched_slice(cfs_rq, curr);
3176 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3177 if (delta_exec > ideal_runtime) {
3178 resched_curr(rq_of(cfs_rq));
3180 * The current task ran long enough, ensure it doesn't get
3181 * re-elected due to buddy favours.
3183 clear_buddies(cfs_rq, curr);
3188 * Ensure that a task that missed wakeup preemption by a
3189 * narrow margin doesn't have to wait for a full slice.
3190 * This also mitigates buddy induced latencies under load.
3192 if (delta_exec < sysctl_sched_min_granularity)
3195 se = __pick_first_entity(cfs_rq);
3196 delta = curr->vruntime - se->vruntime;
3201 if (delta > ideal_runtime)
3202 resched_curr(rq_of(cfs_rq));
3206 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3208 /* 'current' is not kept within the tree. */
3211 * Any task has to be enqueued before it get to execute on
3212 * a CPU. So account for the time it spent waiting on the
3215 update_stats_wait_end(cfs_rq, se);
3216 __dequeue_entity(cfs_rq, se);
3217 update_load_avg(se, 1);
3220 update_stats_curr_start(cfs_rq, se);
3222 #ifdef CONFIG_SCHEDSTATS
3224 * Track our maximum slice length, if the CPU's load is at
3225 * least twice that of our own weight (i.e. dont track it
3226 * when there are only lesser-weight tasks around):
3228 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3229 se->statistics.slice_max = max(se->statistics.slice_max,
3230 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3233 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3237 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3240 * Pick the next process, keeping these things in mind, in this order:
3241 * 1) keep things fair between processes/task groups
3242 * 2) pick the "next" process, since someone really wants that to run
3243 * 3) pick the "last" process, for cache locality
3244 * 4) do not run the "skip" process, if something else is available
3246 static struct sched_entity *
3247 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3249 struct sched_entity *left = __pick_first_entity(cfs_rq);
3250 struct sched_entity *se;
3253 * If curr is set we have to see if its left of the leftmost entity
3254 * still in the tree, provided there was anything in the tree at all.
3256 if (!left || (curr && entity_before(curr, left)))
3259 se = left; /* ideally we run the leftmost entity */
3262 * Avoid running the skip buddy, if running something else can
3263 * be done without getting too unfair.
3265 if (cfs_rq->skip == se) {
3266 struct sched_entity *second;
3269 second = __pick_first_entity(cfs_rq);
3271 second = __pick_next_entity(se);
3272 if (!second || (curr && entity_before(curr, second)))
3276 if (second && wakeup_preempt_entity(second, left) < 1)
3281 * Prefer last buddy, try to return the CPU to a preempted task.
3283 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3287 * Someone really wants this to run. If it's not unfair, run it.
3289 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3292 clear_buddies(cfs_rq, se);
3297 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3299 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3302 * If still on the runqueue then deactivate_task()
3303 * was not called and update_curr() has to be done:
3306 update_curr(cfs_rq);
3308 /* throttle cfs_rqs exceeding runtime */
3309 check_cfs_rq_runtime(cfs_rq);
3311 check_spread(cfs_rq, prev);
3313 update_stats_wait_start(cfs_rq, prev);
3314 /* Put 'current' back into the tree. */
3315 __enqueue_entity(cfs_rq, prev);
3316 /* in !on_rq case, update occurred at dequeue */
3317 update_load_avg(prev, 0);
3319 cfs_rq->curr = NULL;
3323 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3326 * Update run-time statistics of the 'current'.
3328 update_curr(cfs_rq);
3331 * Ensure that runnable average is periodically updated.
3333 update_load_avg(curr, 1);
3334 update_cfs_shares(cfs_rq);
3336 #ifdef CONFIG_SCHED_HRTICK
3338 * queued ticks are scheduled to match the slice, so don't bother
3339 * validating it and just reschedule.
3342 resched_curr(rq_of(cfs_rq));
3346 * don't let the period tick interfere with the hrtick preemption
3348 if (!sched_feat(DOUBLE_TICK) &&
3349 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3353 if (cfs_rq->nr_running > 1)
3354 check_preempt_tick(cfs_rq, curr);
3358 /**************************************************
3359 * CFS bandwidth control machinery
3362 #ifdef CONFIG_CFS_BANDWIDTH
3364 #ifdef HAVE_JUMP_LABEL
3365 static struct static_key __cfs_bandwidth_used;
3367 static inline bool cfs_bandwidth_used(void)
3369 return static_key_false(&__cfs_bandwidth_used);
3372 void cfs_bandwidth_usage_inc(void)
3374 static_key_slow_inc(&__cfs_bandwidth_used);
3377 void cfs_bandwidth_usage_dec(void)
3379 static_key_slow_dec(&__cfs_bandwidth_used);
3381 #else /* HAVE_JUMP_LABEL */
3382 static bool cfs_bandwidth_used(void)
3387 void cfs_bandwidth_usage_inc(void) {}
3388 void cfs_bandwidth_usage_dec(void) {}
3389 #endif /* HAVE_JUMP_LABEL */
3392 * default period for cfs group bandwidth.
3393 * default: 0.1s, units: nanoseconds
3395 static inline u64 default_cfs_period(void)
3397 return 100000000ULL;
3400 static inline u64 sched_cfs_bandwidth_slice(void)
3402 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3406 * Replenish runtime according to assigned quota and update expiration time.
3407 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3408 * additional synchronization around rq->lock.
3410 * requires cfs_b->lock
3412 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3416 if (cfs_b->quota == RUNTIME_INF)
3419 now = sched_clock_cpu(smp_processor_id());
3420 cfs_b->runtime = cfs_b->quota;
3421 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3424 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3426 return &tg->cfs_bandwidth;
3429 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3430 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3432 if (unlikely(cfs_rq->throttle_count))
3433 return cfs_rq->throttled_clock_task;
3435 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3438 /* returns 0 on failure to allocate runtime */
3439 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3441 struct task_group *tg = cfs_rq->tg;
3442 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3443 u64 amount = 0, min_amount, expires;
3445 /* note: this is a positive sum as runtime_remaining <= 0 */
3446 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3448 raw_spin_lock(&cfs_b->lock);
3449 if (cfs_b->quota == RUNTIME_INF)
3450 amount = min_amount;
3452 start_cfs_bandwidth(cfs_b);
3454 if (cfs_b->runtime > 0) {
3455 amount = min(cfs_b->runtime, min_amount);
3456 cfs_b->runtime -= amount;
3460 expires = cfs_b->runtime_expires;
3461 raw_spin_unlock(&cfs_b->lock);
3463 cfs_rq->runtime_remaining += amount;
3465 * we may have advanced our local expiration to account for allowed
3466 * spread between our sched_clock and the one on which runtime was
3469 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3470 cfs_rq->runtime_expires = expires;
3472 return cfs_rq->runtime_remaining > 0;
3476 * Note: This depends on the synchronization provided by sched_clock and the
3477 * fact that rq->clock snapshots this value.
3479 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3481 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3483 /* if the deadline is ahead of our clock, nothing to do */
3484 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3487 if (cfs_rq->runtime_remaining < 0)
3491 * If the local deadline has passed we have to consider the
3492 * possibility that our sched_clock is 'fast' and the global deadline
3493 * has not truly expired.
3495 * Fortunately we can check determine whether this the case by checking
3496 * whether the global deadline has advanced. It is valid to compare
3497 * cfs_b->runtime_expires without any locks since we only care about
3498 * exact equality, so a partial write will still work.
3501 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3502 /* extend local deadline, drift is bounded above by 2 ticks */
3503 cfs_rq->runtime_expires += TICK_NSEC;
3505 /* global deadline is ahead, expiration has passed */
3506 cfs_rq->runtime_remaining = 0;
3510 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3512 /* dock delta_exec before expiring quota (as it could span periods) */
3513 cfs_rq->runtime_remaining -= delta_exec;
3514 expire_cfs_rq_runtime(cfs_rq);
3516 if (likely(cfs_rq->runtime_remaining > 0))
3520 * if we're unable to extend our runtime we resched so that the active
3521 * hierarchy can be throttled
3523 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3524 resched_curr(rq_of(cfs_rq));
3527 static __always_inline
3528 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3530 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3533 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3536 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3538 return cfs_bandwidth_used() && cfs_rq->throttled;
3541 /* check whether cfs_rq, or any parent, is throttled */
3542 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3544 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3548 * Ensure that neither of the group entities corresponding to src_cpu or
3549 * dest_cpu are members of a throttled hierarchy when performing group
3550 * load-balance operations.
3552 static inline int throttled_lb_pair(struct task_group *tg,
3553 int src_cpu, int dest_cpu)
3555 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3557 src_cfs_rq = tg->cfs_rq[src_cpu];
3558 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3560 return throttled_hierarchy(src_cfs_rq) ||
3561 throttled_hierarchy(dest_cfs_rq);
3564 /* updated child weight may affect parent so we have to do this bottom up */
3565 static int tg_unthrottle_up(struct task_group *tg, void *data)
3567 struct rq *rq = data;
3568 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3570 cfs_rq->throttle_count--;
3572 if (!cfs_rq->throttle_count) {
3573 /* adjust cfs_rq_clock_task() */
3574 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3575 cfs_rq->throttled_clock_task;
3582 static int tg_throttle_down(struct task_group *tg, void *data)
3584 struct rq *rq = data;
3585 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3587 /* group is entering throttled state, stop time */
3588 if (!cfs_rq->throttle_count)
3589 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3590 cfs_rq->throttle_count++;
3595 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3597 struct rq *rq = rq_of(cfs_rq);
3598 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3599 struct sched_entity *se;
3600 long task_delta, dequeue = 1;
3603 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3605 /* freeze hierarchy runnable averages while throttled */
3607 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3610 task_delta = cfs_rq->h_nr_running;
3611 for_each_sched_entity(se) {
3612 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3613 /* throttled entity or throttle-on-deactivate */
3618 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3619 qcfs_rq->h_nr_running -= task_delta;
3621 if (qcfs_rq->load.weight)
3626 sub_nr_running(rq, task_delta);
3628 cfs_rq->throttled = 1;
3629 cfs_rq->throttled_clock = rq_clock(rq);
3630 raw_spin_lock(&cfs_b->lock);
3631 empty = list_empty(&cfs_b->throttled_cfs_rq);
3634 * Add to the _head_ of the list, so that an already-started
3635 * distribute_cfs_runtime will not see us
3637 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3640 * If we're the first throttled task, make sure the bandwidth
3644 start_cfs_bandwidth(cfs_b);
3646 raw_spin_unlock(&cfs_b->lock);
3649 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3651 struct rq *rq = rq_of(cfs_rq);
3652 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3653 struct sched_entity *se;
3657 se = cfs_rq->tg->se[cpu_of(rq)];
3659 cfs_rq->throttled = 0;
3661 update_rq_clock(rq);
3663 raw_spin_lock(&cfs_b->lock);
3664 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3665 list_del_rcu(&cfs_rq->throttled_list);
3666 raw_spin_unlock(&cfs_b->lock);
3668 /* update hierarchical throttle state */
3669 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3671 if (!cfs_rq->load.weight)
3674 task_delta = cfs_rq->h_nr_running;
3675 for_each_sched_entity(se) {
3679 cfs_rq = cfs_rq_of(se);
3681 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3682 cfs_rq->h_nr_running += task_delta;
3684 if (cfs_rq_throttled(cfs_rq))
3689 add_nr_running(rq, task_delta);
3691 /* determine whether we need to wake up potentially idle cpu */
3692 if (rq->curr == rq->idle && rq->cfs.nr_running)
3696 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3697 u64 remaining, u64 expires)
3699 struct cfs_rq *cfs_rq;
3701 u64 starting_runtime = remaining;
3704 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3706 struct rq *rq = rq_of(cfs_rq);
3708 raw_spin_lock(&rq->lock);
3709 if (!cfs_rq_throttled(cfs_rq))
3712 runtime = -cfs_rq->runtime_remaining + 1;
3713 if (runtime > remaining)
3714 runtime = remaining;
3715 remaining -= runtime;
3717 cfs_rq->runtime_remaining += runtime;
3718 cfs_rq->runtime_expires = expires;
3720 /* we check whether we're throttled above */
3721 if (cfs_rq->runtime_remaining > 0)
3722 unthrottle_cfs_rq(cfs_rq);
3725 raw_spin_unlock(&rq->lock);
3732 return starting_runtime - remaining;
3736 * Responsible for refilling a task_group's bandwidth and unthrottling its
3737 * cfs_rqs as appropriate. If there has been no activity within the last
3738 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3739 * used to track this state.
3741 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3743 u64 runtime, runtime_expires;
3746 /* no need to continue the timer with no bandwidth constraint */
3747 if (cfs_b->quota == RUNTIME_INF)
3748 goto out_deactivate;
3750 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3751 cfs_b->nr_periods += overrun;
3754 * idle depends on !throttled (for the case of a large deficit), and if
3755 * we're going inactive then everything else can be deferred
3757 if (cfs_b->idle && !throttled)
3758 goto out_deactivate;
3760 __refill_cfs_bandwidth_runtime(cfs_b);
3763 /* mark as potentially idle for the upcoming period */
3768 /* account preceding periods in which throttling occurred */
3769 cfs_b->nr_throttled += overrun;
3771 runtime_expires = cfs_b->runtime_expires;
3774 * This check is repeated as we are holding onto the new bandwidth while
3775 * we unthrottle. This can potentially race with an unthrottled group
3776 * trying to acquire new bandwidth from the global pool. This can result
3777 * in us over-using our runtime if it is all used during this loop, but
3778 * only by limited amounts in that extreme case.
3780 while (throttled && cfs_b->runtime > 0) {
3781 runtime = cfs_b->runtime;
3782 raw_spin_unlock(&cfs_b->lock);
3783 /* we can't nest cfs_b->lock while distributing bandwidth */
3784 runtime = distribute_cfs_runtime(cfs_b, runtime,
3786 raw_spin_lock(&cfs_b->lock);
3788 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3790 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3794 * While we are ensured activity in the period following an
3795 * unthrottle, this also covers the case in which the new bandwidth is
3796 * insufficient to cover the existing bandwidth deficit. (Forcing the
3797 * timer to remain active while there are any throttled entities.)
3807 /* a cfs_rq won't donate quota below this amount */
3808 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3809 /* minimum remaining period time to redistribute slack quota */
3810 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3811 /* how long we wait to gather additional slack before distributing */
3812 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3815 * Are we near the end of the current quota period?
3817 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3818 * hrtimer base being cleared by hrtimer_start. In the case of
3819 * migrate_hrtimers, base is never cleared, so we are fine.
3821 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3823 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3826 /* if the call-back is running a quota refresh is already occurring */
3827 if (hrtimer_callback_running(refresh_timer))
3830 /* is a quota refresh about to occur? */
3831 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3832 if (remaining < min_expire)
3838 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3840 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3842 /* if there's a quota refresh soon don't bother with slack */
3843 if (runtime_refresh_within(cfs_b, min_left))
3846 hrtimer_start(&cfs_b->slack_timer,
3847 ns_to_ktime(cfs_bandwidth_slack_period),
3851 /* we know any runtime found here is valid as update_curr() precedes return */
3852 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3854 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3855 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3857 if (slack_runtime <= 0)
3860 raw_spin_lock(&cfs_b->lock);
3861 if (cfs_b->quota != RUNTIME_INF &&
3862 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3863 cfs_b->runtime += slack_runtime;
3865 /* we are under rq->lock, defer unthrottling using a timer */
3866 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3867 !list_empty(&cfs_b->throttled_cfs_rq))
3868 start_cfs_slack_bandwidth(cfs_b);
3870 raw_spin_unlock(&cfs_b->lock);
3872 /* even if it's not valid for return we don't want to try again */
3873 cfs_rq->runtime_remaining -= slack_runtime;
3876 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3878 if (!cfs_bandwidth_used())
3881 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3884 __return_cfs_rq_runtime(cfs_rq);
3888 * This is done with a timer (instead of inline with bandwidth return) since
3889 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3891 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3893 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3896 /* confirm we're still not at a refresh boundary */
3897 raw_spin_lock(&cfs_b->lock);
3898 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3899 raw_spin_unlock(&cfs_b->lock);
3903 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3904 runtime = cfs_b->runtime;
3906 expires = cfs_b->runtime_expires;
3907 raw_spin_unlock(&cfs_b->lock);
3912 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3914 raw_spin_lock(&cfs_b->lock);
3915 if (expires == cfs_b->runtime_expires)
3916 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3917 raw_spin_unlock(&cfs_b->lock);
3921 * When a group wakes up we want to make sure that its quota is not already
3922 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3923 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3925 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3927 if (!cfs_bandwidth_used())
3930 /* an active group must be handled by the update_curr()->put() path */
3931 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3934 /* ensure the group is not already throttled */
3935 if (cfs_rq_throttled(cfs_rq))
3938 /* update runtime allocation */
3939 account_cfs_rq_runtime(cfs_rq, 0);
3940 if (cfs_rq->runtime_remaining <= 0)
3941 throttle_cfs_rq(cfs_rq);
3944 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3945 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3947 if (!cfs_bandwidth_used())
3950 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3954 * it's possible for a throttled entity to be forced into a running
3955 * state (e.g. set_curr_task), in this case we're finished.
3957 if (cfs_rq_throttled(cfs_rq))
3960 throttle_cfs_rq(cfs_rq);
3964 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3966 struct cfs_bandwidth *cfs_b =
3967 container_of(timer, struct cfs_bandwidth, slack_timer);
3969 do_sched_cfs_slack_timer(cfs_b);
3971 return HRTIMER_NORESTART;
3974 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3976 struct cfs_bandwidth *cfs_b =
3977 container_of(timer, struct cfs_bandwidth, period_timer);
3981 raw_spin_lock(&cfs_b->lock);
3983 overrun = hrtimer_forward_now(timer, cfs_b->period);
3987 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3990 cfs_b->period_active = 0;
3991 raw_spin_unlock(&cfs_b->lock);
3993 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3996 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3998 raw_spin_lock_init(&cfs_b->lock);
4000 cfs_b->quota = RUNTIME_INF;
4001 cfs_b->period = ns_to_ktime(default_cfs_period());
4003 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4004 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4005 cfs_b->period_timer.function = sched_cfs_period_timer;
4006 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4007 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4010 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4012 cfs_rq->runtime_enabled = 0;
4013 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4016 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4018 lockdep_assert_held(&cfs_b->lock);
4020 if (!cfs_b->period_active) {
4021 cfs_b->period_active = 1;
4022 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4023 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4027 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4029 /* init_cfs_bandwidth() was not called */
4030 if (!cfs_b->throttled_cfs_rq.next)
4033 hrtimer_cancel(&cfs_b->period_timer);
4034 hrtimer_cancel(&cfs_b->slack_timer);
4037 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4039 struct cfs_rq *cfs_rq;
4041 for_each_leaf_cfs_rq(rq, cfs_rq) {
4042 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4044 raw_spin_lock(&cfs_b->lock);
4045 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4046 raw_spin_unlock(&cfs_b->lock);
4050 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4052 struct cfs_rq *cfs_rq;
4054 for_each_leaf_cfs_rq(rq, cfs_rq) {
4055 if (!cfs_rq->runtime_enabled)
4059 * clock_task is not advancing so we just need to make sure
4060 * there's some valid quota amount
4062 cfs_rq->runtime_remaining = 1;
4064 * Offline rq is schedulable till cpu is completely disabled
4065 * in take_cpu_down(), so we prevent new cfs throttling here.
4067 cfs_rq->runtime_enabled = 0;
4069 if (cfs_rq_throttled(cfs_rq))
4070 unthrottle_cfs_rq(cfs_rq);
4074 #else /* CONFIG_CFS_BANDWIDTH */
4075 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4077 return rq_clock_task(rq_of(cfs_rq));
4080 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4081 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4082 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4083 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4085 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4090 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4095 static inline int throttled_lb_pair(struct task_group *tg,
4096 int src_cpu, int dest_cpu)
4101 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4103 #ifdef CONFIG_FAIR_GROUP_SCHED
4104 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4107 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4111 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4112 static inline void update_runtime_enabled(struct rq *rq) {}
4113 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4115 #endif /* CONFIG_CFS_BANDWIDTH */
4117 /**************************************************
4118 * CFS operations on tasks:
4121 #ifdef CONFIG_SCHED_HRTICK
4122 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4124 struct sched_entity *se = &p->se;
4125 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4127 WARN_ON(task_rq(p) != rq);
4129 if (cfs_rq->nr_running > 1) {
4130 u64 slice = sched_slice(cfs_rq, se);
4131 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4132 s64 delta = slice - ran;
4139 hrtick_start(rq, delta);
4144 * called from enqueue/dequeue and updates the hrtick when the
4145 * current task is from our class and nr_running is low enough
4148 static void hrtick_update(struct rq *rq)
4150 struct task_struct *curr = rq->curr;
4152 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4155 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4156 hrtick_start_fair(rq, curr);
4158 #else /* !CONFIG_SCHED_HRTICK */
4160 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4164 static inline void hrtick_update(struct rq *rq)
4169 static inline unsigned long boosted_cpu_util(int cpu);
4171 static void update_capacity_of(int cpu)
4173 unsigned long req_cap;
4178 /* Convert scale-invariant capacity to cpu. */
4179 req_cap = boosted_cpu_util(cpu);
4180 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4181 set_cfs_cpu_capacity(cpu, true, req_cap);
4184 static bool cpu_overutilized(int cpu);
4187 * The enqueue_task method is called before nr_running is
4188 * increased. Here we update the fair scheduling stats and
4189 * then put the task into the rbtree:
4192 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4194 struct cfs_rq *cfs_rq;
4195 struct sched_entity *se = &p->se;
4196 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4197 int task_wakeup = flags & ENQUEUE_WAKEUP;
4199 for_each_sched_entity(se) {
4202 cfs_rq = cfs_rq_of(se);
4203 enqueue_entity(cfs_rq, se, flags);
4206 * end evaluation on encountering a throttled cfs_rq
4208 * note: in the case of encountering a throttled cfs_rq we will
4209 * post the final h_nr_running increment below.
4211 if (cfs_rq_throttled(cfs_rq))
4213 cfs_rq->h_nr_running++;
4215 flags = ENQUEUE_WAKEUP;
4218 for_each_sched_entity(se) {
4219 cfs_rq = cfs_rq_of(se);
4220 cfs_rq->h_nr_running++;
4222 if (cfs_rq_throttled(cfs_rq))
4225 update_load_avg(se, 1);
4226 update_cfs_shares(cfs_rq);
4230 add_nr_running(rq, 1);
4231 if (!task_new && !rq->rd->overutilized &&
4232 cpu_overutilized(rq->cpu))
4233 rq->rd->overutilized = true;
4235 schedtune_enqueue_task(p, cpu_of(rq));
4238 * We want to potentially trigger a freq switch
4239 * request only for tasks that are waking up; this is
4240 * because we get here also during load balancing, but
4241 * in these cases it seems wise to trigger as single
4242 * request after load balancing is done.
4244 if (task_new || task_wakeup)
4245 update_capacity_of(cpu_of(rq));
4250 static void set_next_buddy(struct sched_entity *se);
4253 * The dequeue_task method is called before nr_running is
4254 * decreased. We remove the task from the rbtree and
4255 * update the fair scheduling stats:
4257 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4259 struct cfs_rq *cfs_rq;
4260 struct sched_entity *se = &p->se;
4261 int task_sleep = flags & DEQUEUE_SLEEP;
4263 for_each_sched_entity(se) {
4264 cfs_rq = cfs_rq_of(se);
4265 dequeue_entity(cfs_rq, se, flags);
4268 * end evaluation on encountering a throttled cfs_rq
4270 * note: in the case of encountering a throttled cfs_rq we will
4271 * post the final h_nr_running decrement below.
4273 if (cfs_rq_throttled(cfs_rq))
4275 cfs_rq->h_nr_running--;
4277 /* Don't dequeue parent if it has other entities besides us */
4278 if (cfs_rq->load.weight) {
4280 * Bias pick_next to pick a task from this cfs_rq, as
4281 * p is sleeping when it is within its sched_slice.
4283 if (task_sleep && parent_entity(se))
4284 set_next_buddy(parent_entity(se));
4286 /* avoid re-evaluating load for this entity */
4287 se = parent_entity(se);
4290 flags |= DEQUEUE_SLEEP;
4293 for_each_sched_entity(se) {
4294 cfs_rq = cfs_rq_of(se);
4295 cfs_rq->h_nr_running--;
4297 if (cfs_rq_throttled(cfs_rq))
4300 update_load_avg(se, 1);
4301 update_cfs_shares(cfs_rq);
4305 sub_nr_running(rq, 1);
4306 schedtune_dequeue_task(p, cpu_of(rq));
4309 * We want to potentially trigger a freq switch
4310 * request only for tasks that are going to sleep;
4311 * this is because we get here also during load
4312 * balancing, but in these cases it seems wise to
4313 * trigger as single request after load balancing is
4317 if (rq->cfs.nr_running)
4318 update_capacity_of(cpu_of(rq));
4319 else if (sched_freq())
4320 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4329 * per rq 'load' arrray crap; XXX kill this.
4333 * The exact cpuload at various idx values, calculated at every tick would be
4334 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4336 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4337 * on nth tick when cpu may be busy, then we have:
4338 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4339 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4341 * decay_load_missed() below does efficient calculation of
4342 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4343 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4345 * The calculation is approximated on a 128 point scale.
4346 * degrade_zero_ticks is the number of ticks after which load at any
4347 * particular idx is approximated to be zero.
4348 * degrade_factor is a precomputed table, a row for each load idx.
4349 * Each column corresponds to degradation factor for a power of two ticks,
4350 * based on 128 point scale.
4352 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4353 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4355 * With this power of 2 load factors, we can degrade the load n times
4356 * by looking at 1 bits in n and doing as many mult/shift instead of
4357 * n mult/shifts needed by the exact degradation.
4359 #define DEGRADE_SHIFT 7
4360 static const unsigned char
4361 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4362 static const unsigned char
4363 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4364 {0, 0, 0, 0, 0, 0, 0, 0},
4365 {64, 32, 8, 0, 0, 0, 0, 0},
4366 {96, 72, 40, 12, 1, 0, 0},
4367 {112, 98, 75, 43, 15, 1, 0},
4368 {120, 112, 98, 76, 45, 16, 2} };
4371 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4372 * would be when CPU is idle and so we just decay the old load without
4373 * adding any new load.
4375 static unsigned long
4376 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4380 if (!missed_updates)
4383 if (missed_updates >= degrade_zero_ticks[idx])
4387 return load >> missed_updates;
4389 while (missed_updates) {
4390 if (missed_updates % 2)
4391 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4393 missed_updates >>= 1;
4400 * Update rq->cpu_load[] statistics. This function is usually called every
4401 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4402 * every tick. We fix it up based on jiffies.
4404 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4405 unsigned long pending_updates)
4409 this_rq->nr_load_updates++;
4411 /* Update our load: */
4412 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4413 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4414 unsigned long old_load, new_load;
4416 /* scale is effectively 1 << i now, and >> i divides by scale */
4418 old_load = this_rq->cpu_load[i];
4419 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4420 new_load = this_load;
4422 * Round up the averaging division if load is increasing. This
4423 * prevents us from getting stuck on 9 if the load is 10, for
4426 if (new_load > old_load)
4427 new_load += scale - 1;
4429 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4432 sched_avg_update(this_rq);
4435 /* Used instead of source_load when we know the type == 0 */
4436 static unsigned long weighted_cpuload(const int cpu)
4438 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4441 #ifdef CONFIG_NO_HZ_COMMON
4443 * There is no sane way to deal with nohz on smp when using jiffies because the
4444 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4445 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4447 * Therefore we cannot use the delta approach from the regular tick since that
4448 * would seriously skew the load calculation. However we'll make do for those
4449 * updates happening while idle (nohz_idle_balance) or coming out of idle
4450 * (tick_nohz_idle_exit).
4452 * This means we might still be one tick off for nohz periods.
4456 * Called from nohz_idle_balance() to update the load ratings before doing the
4459 static void update_idle_cpu_load(struct rq *this_rq)
4461 unsigned long curr_jiffies = READ_ONCE(jiffies);
4462 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4463 unsigned long pending_updates;
4466 * bail if there's load or we're actually up-to-date.
4468 if (load || curr_jiffies == this_rq->last_load_update_tick)
4471 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4472 this_rq->last_load_update_tick = curr_jiffies;
4474 __update_cpu_load(this_rq, load, pending_updates);
4478 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4480 void update_cpu_load_nohz(void)
4482 struct rq *this_rq = this_rq();
4483 unsigned long curr_jiffies = READ_ONCE(jiffies);
4484 unsigned long pending_updates;
4486 if (curr_jiffies == this_rq->last_load_update_tick)
4489 raw_spin_lock(&this_rq->lock);
4490 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4491 if (pending_updates) {
4492 this_rq->last_load_update_tick = curr_jiffies;
4494 * We were idle, this means load 0, the current load might be
4495 * !0 due to remote wakeups and the sort.
4497 __update_cpu_load(this_rq, 0, pending_updates);
4499 raw_spin_unlock(&this_rq->lock);
4501 #endif /* CONFIG_NO_HZ */
4504 * Called from scheduler_tick()
4506 void update_cpu_load_active(struct rq *this_rq)
4508 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4510 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4512 this_rq->last_load_update_tick = jiffies;
4513 __update_cpu_load(this_rq, load, 1);
4517 * Return a low guess at the load of a migration-source cpu weighted
4518 * according to the scheduling class and "nice" value.
4520 * We want to under-estimate the load of migration sources, to
4521 * balance conservatively.
4523 static unsigned long source_load(int cpu, int type)
4525 struct rq *rq = cpu_rq(cpu);
4526 unsigned long total = weighted_cpuload(cpu);
4528 if (type == 0 || !sched_feat(LB_BIAS))
4531 return min(rq->cpu_load[type-1], total);
4535 * Return a high guess at the load of a migration-target cpu weighted
4536 * according to the scheduling class and "nice" value.
4538 static unsigned long target_load(int cpu, int type)
4540 struct rq *rq = cpu_rq(cpu);
4541 unsigned long total = weighted_cpuload(cpu);
4543 if (type == 0 || !sched_feat(LB_BIAS))
4546 return max(rq->cpu_load[type-1], total);
4550 static unsigned long cpu_avg_load_per_task(int cpu)
4552 struct rq *rq = cpu_rq(cpu);
4553 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4554 unsigned long load_avg = weighted_cpuload(cpu);
4557 return load_avg / nr_running;
4562 static void record_wakee(struct task_struct *p)
4565 * Rough decay (wiping) for cost saving, don't worry
4566 * about the boundary, really active task won't care
4569 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4570 current->wakee_flips >>= 1;
4571 current->wakee_flip_decay_ts = jiffies;
4574 if (current->last_wakee != p) {
4575 current->last_wakee = p;
4576 current->wakee_flips++;
4580 static void task_waking_fair(struct task_struct *p)
4582 struct sched_entity *se = &p->se;
4583 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4586 #ifndef CONFIG_64BIT
4587 u64 min_vruntime_copy;
4590 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4592 min_vruntime = cfs_rq->min_vruntime;
4593 } while (min_vruntime != min_vruntime_copy);
4595 min_vruntime = cfs_rq->min_vruntime;
4598 se->vruntime -= min_vruntime;
4602 #ifdef CONFIG_FAIR_GROUP_SCHED
4604 * effective_load() calculates the load change as seen from the root_task_group
4606 * Adding load to a group doesn't make a group heavier, but can cause movement
4607 * of group shares between cpus. Assuming the shares were perfectly aligned one
4608 * can calculate the shift in shares.
4610 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4611 * on this @cpu and results in a total addition (subtraction) of @wg to the
4612 * total group weight.
4614 * Given a runqueue weight distribution (rw_i) we can compute a shares
4615 * distribution (s_i) using:
4617 * s_i = rw_i / \Sum rw_j (1)
4619 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4620 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4621 * shares distribution (s_i):
4623 * rw_i = { 2, 4, 1, 0 }
4624 * s_i = { 2/7, 4/7, 1/7, 0 }
4626 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4627 * task used to run on and the CPU the waker is running on), we need to
4628 * compute the effect of waking a task on either CPU and, in case of a sync
4629 * wakeup, compute the effect of the current task going to sleep.
4631 * So for a change of @wl to the local @cpu with an overall group weight change
4632 * of @wl we can compute the new shares distribution (s'_i) using:
4634 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4636 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4637 * differences in waking a task to CPU 0. The additional task changes the
4638 * weight and shares distributions like:
4640 * rw'_i = { 3, 4, 1, 0 }
4641 * s'_i = { 3/8, 4/8, 1/8, 0 }
4643 * We can then compute the difference in effective weight by using:
4645 * dw_i = S * (s'_i - s_i) (3)
4647 * Where 'S' is the group weight as seen by its parent.
4649 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4650 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4651 * 4/7) times the weight of the group.
4653 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4655 struct sched_entity *se = tg->se[cpu];
4657 if (!tg->parent) /* the trivial, non-cgroup case */
4660 for_each_sched_entity(se) {
4661 struct cfs_rq *cfs_rq = se->my_q;
4662 long W, w = cfs_rq_load_avg(cfs_rq);
4667 * W = @wg + \Sum rw_j
4669 W = wg + atomic_long_read(&tg->load_avg);
4671 /* Ensure \Sum rw_j >= rw_i */
4672 W -= cfs_rq->tg_load_avg_contrib;
4681 * wl = S * s'_i; see (2)
4684 wl = (w * (long)tg->shares) / W;
4689 * Per the above, wl is the new se->load.weight value; since
4690 * those are clipped to [MIN_SHARES, ...) do so now. See
4691 * calc_cfs_shares().
4693 if (wl < MIN_SHARES)
4697 * wl = dw_i = S * (s'_i - s_i); see (3)
4699 wl -= se->avg.load_avg;
4702 * Recursively apply this logic to all parent groups to compute
4703 * the final effective load change on the root group. Since
4704 * only the @tg group gets extra weight, all parent groups can
4705 * only redistribute existing shares. @wl is the shift in shares
4706 * resulting from this level per the above.
4715 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4722 static inline bool energy_aware(void)
4724 return sched_feat(ENERGY_AWARE);
4728 struct sched_group *sg_top;
4729 struct sched_group *sg_cap;
4736 struct task_struct *task;
4751 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4752 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4753 * energy calculations. Using the scale-invariant util returned by
4754 * cpu_util() and approximating scale-invariant util by:
4756 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4758 * the normalized util can be found using the specific capacity.
4760 * capacity = capacity_orig * curr_freq/max_freq
4762 * norm_util = running_time/time ~ util/capacity
4764 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4766 int util = __cpu_util(cpu, delta);
4768 if (util >= capacity)
4769 return SCHED_CAPACITY_SCALE;
4771 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4774 static int calc_util_delta(struct energy_env *eenv, int cpu)
4776 if (cpu == eenv->src_cpu)
4777 return -eenv->util_delta;
4778 if (cpu == eenv->dst_cpu)
4779 return eenv->util_delta;
4784 unsigned long group_max_util(struct energy_env *eenv)
4787 unsigned long max_util = 0;
4789 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4790 delta = calc_util_delta(eenv, i);
4791 max_util = max(max_util, __cpu_util(i, delta));
4798 * group_norm_util() returns the approximated group util relative to it's
4799 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4800 * energy calculations. Since task executions may or may not overlap in time in
4801 * the group the true normalized util is between max(cpu_norm_util(i)) and
4802 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4803 * latter is used as the estimate as it leads to a more pessimistic energy
4804 * estimate (more busy).
4807 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4810 unsigned long util_sum = 0;
4811 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4813 for_each_cpu(i, sched_group_cpus(sg)) {
4814 delta = calc_util_delta(eenv, i);
4815 util_sum += __cpu_norm_util(i, capacity, delta);
4818 if (util_sum > SCHED_CAPACITY_SCALE)
4819 return SCHED_CAPACITY_SCALE;
4823 static int find_new_capacity(struct energy_env *eenv,
4824 const struct sched_group_energy const *sge)
4827 unsigned long util = group_max_util(eenv);
4829 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4830 if (sge->cap_states[idx].cap >= util)
4834 eenv->cap_idx = idx;
4839 static int group_idle_state(struct sched_group *sg)
4841 int i, state = INT_MAX;
4843 /* Find the shallowest idle state in the sched group. */
4844 for_each_cpu(i, sched_group_cpus(sg))
4845 state = min(state, idle_get_state_idx(cpu_rq(i)));
4847 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4854 * sched_group_energy(): Computes the absolute energy consumption of cpus
4855 * belonging to the sched_group including shared resources shared only by
4856 * members of the group. Iterates over all cpus in the hierarchy below the
4857 * sched_group starting from the bottom working it's way up before going to
4858 * the next cpu until all cpus are covered at all levels. The current
4859 * implementation is likely to gather the same util statistics multiple times.
4860 * This can probably be done in a faster but more complex way.
4861 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4863 static int sched_group_energy(struct energy_env *eenv)
4865 struct sched_domain *sd;
4866 int cpu, total_energy = 0;
4867 struct cpumask visit_cpus;
4868 struct sched_group *sg;
4870 WARN_ON(!eenv->sg_top->sge);
4872 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4874 while (!cpumask_empty(&visit_cpus)) {
4875 struct sched_group *sg_shared_cap = NULL;
4877 cpu = cpumask_first(&visit_cpus);
4880 * Is the group utilization affected by cpus outside this
4883 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4887 * We most probably raced with hotplug; returning a
4888 * wrong energy estimation is better than entering an
4894 sg_shared_cap = sd->parent->groups;
4896 for_each_domain(cpu, sd) {
4899 /* Has this sched_domain already been visited? */
4900 if (sd->child && group_first_cpu(sg) != cpu)
4904 unsigned long group_util;
4905 int sg_busy_energy, sg_idle_energy;
4906 int cap_idx, idle_idx;
4908 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4909 eenv->sg_cap = sg_shared_cap;
4913 cap_idx = find_new_capacity(eenv, sg->sge);
4915 if (sg->group_weight == 1) {
4916 /* Remove capacity of src CPU (before task move) */
4917 if (eenv->util_delta == 0 &&
4918 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4919 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4920 eenv->cap.delta -= eenv->cap.before;
4922 /* Add capacity of dst CPU (after task move) */
4923 if (eenv->util_delta != 0 &&
4924 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4925 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4926 eenv->cap.delta += eenv->cap.after;
4930 idle_idx = group_idle_state(sg);
4931 group_util = group_norm_util(eenv, sg);
4932 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4933 >> SCHED_CAPACITY_SHIFT;
4934 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4935 * sg->sge->idle_states[idle_idx].power)
4936 >> SCHED_CAPACITY_SHIFT;
4938 total_energy += sg_busy_energy + sg_idle_energy;
4941 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4943 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4946 } while (sg = sg->next, sg != sd->groups);
4952 eenv->energy = total_energy;
4956 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4958 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4961 #ifdef CONFIG_SCHED_TUNE
4962 static int energy_diff_evaluate(struct energy_env *eenv)
4967 /* Return energy diff when boost margin is 0 */
4968 #ifdef CONFIG_CGROUP_SCHEDTUNE
4969 boost = schedtune_task_boost(eenv->task);
4971 boost = get_sysctl_sched_cfs_boost();
4974 return eenv->nrg.diff;
4976 /* Compute normalized energy diff */
4977 nrg_delta = schedtune_normalize_energy(eenv->nrg.diff);
4978 eenv->nrg.delta = nrg_delta;
4980 eenv->payoff = schedtune_accept_deltas(
4986 * When SchedTune is enabled, the energy_diff() function will return
4987 * the computed energy payoff value. Since the energy_diff() return
4988 * value is expected to be negative by its callers, this evaluation
4989 * function return a negative value each time the evaluation return a
4990 * positive payoff, which is the condition for the acceptance of
4991 * a scheduling decision
4993 return -eenv->payoff;
4995 #else /* CONFIG_SCHED_TUNE */
4996 #define energy_diff_evaluate(eenv) eenv->nrg.diff
5000 * energy_diff(): Estimate the energy impact of changing the utilization
5001 * distribution. eenv specifies the change: utilisation amount, source, and
5002 * destination cpu. Source or destination cpu may be -1 in which case the
5003 * utilization is removed from or added to the system (e.g. task wake-up). If
5004 * both are specified, the utilization is migrated.
5006 static int energy_diff(struct energy_env *eenv)
5008 struct sched_domain *sd;
5009 struct sched_group *sg;
5010 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5012 struct energy_env eenv_before = {
5014 .src_cpu = eenv->src_cpu,
5015 .dst_cpu = eenv->dst_cpu,
5016 .nrg = { 0, 0, 0, 0},
5020 if (eenv->src_cpu == eenv->dst_cpu)
5023 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5024 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5027 return 0; /* Error */
5032 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5033 eenv_before.sg_top = eenv->sg_top = sg;
5035 if (sched_group_energy(&eenv_before))
5036 return 0; /* Invalid result abort */
5037 energy_before += eenv_before.energy;
5039 /* Keep track of SRC cpu (before) capacity */
5040 eenv->cap.before = eenv_before.cap.before;
5041 eenv->cap.delta = eenv_before.cap.delta;
5043 if (sched_group_energy(eenv))
5044 return 0; /* Invalid result abort */
5045 energy_after += eenv->energy;
5047 } while (sg = sg->next, sg != sd->groups);
5049 eenv->nrg.before = energy_before;
5050 eenv->nrg.after = energy_after;
5051 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5054 return energy_diff_evaluate(eenv);
5058 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5059 * A waker of many should wake a different task than the one last awakened
5060 * at a frequency roughly N times higher than one of its wakees. In order
5061 * to determine whether we should let the load spread vs consolodating to
5062 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5063 * partner, and a factor of lls_size higher frequency in the other. With
5064 * both conditions met, we can be relatively sure that the relationship is
5065 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5066 * being client/server, worker/dispatcher, interrupt source or whatever is
5067 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5069 static int wake_wide(struct task_struct *p)
5071 unsigned int master = current->wakee_flips;
5072 unsigned int slave = p->wakee_flips;
5073 int factor = this_cpu_read(sd_llc_size);
5076 swap(master, slave);
5077 if (slave < factor || master < slave * factor)
5082 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5084 s64 this_load, load;
5085 s64 this_eff_load, prev_eff_load;
5086 int idx, this_cpu, prev_cpu;
5087 struct task_group *tg;
5088 unsigned long weight;
5092 this_cpu = smp_processor_id();
5093 prev_cpu = task_cpu(p);
5094 load = source_load(prev_cpu, idx);
5095 this_load = target_load(this_cpu, idx);
5098 * If sync wakeup then subtract the (maximum possible)
5099 * effect of the currently running task from the load
5100 * of the current CPU:
5103 tg = task_group(current);
5104 weight = current->se.avg.load_avg;
5106 this_load += effective_load(tg, this_cpu, -weight, -weight);
5107 load += effective_load(tg, prev_cpu, 0, -weight);
5111 weight = p->se.avg.load_avg;
5114 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5115 * due to the sync cause above having dropped this_load to 0, we'll
5116 * always have an imbalance, but there's really nothing you can do
5117 * about that, so that's good too.
5119 * Otherwise check if either cpus are near enough in load to allow this
5120 * task to be woken on this_cpu.
5122 this_eff_load = 100;
5123 this_eff_load *= capacity_of(prev_cpu);
5125 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5126 prev_eff_load *= capacity_of(this_cpu);
5128 if (this_load > 0) {
5129 this_eff_load *= this_load +
5130 effective_load(tg, this_cpu, weight, weight);
5132 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5135 balanced = this_eff_load <= prev_eff_load;
5137 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5142 schedstat_inc(sd, ttwu_move_affine);
5143 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5148 static inline unsigned long task_util(struct task_struct *p)
5150 return p->se.avg.util_avg;
5153 unsigned int capacity_margin = 1280; /* ~20% margin */
5155 static inline unsigned long boosted_task_util(struct task_struct *task);
5157 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5159 unsigned long capacity = capacity_of(cpu);
5161 util += boosted_task_util(p);
5163 return (capacity * 1024) > (util * capacity_margin);
5166 static inline bool task_fits_max(struct task_struct *p, int cpu)
5168 unsigned long capacity = capacity_of(cpu);
5169 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5171 if (capacity == max_capacity)
5174 if (capacity * capacity_margin > max_capacity * 1024)
5177 return __task_fits(p, cpu, 0);
5180 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5182 return __task_fits(p, cpu, cpu_util(cpu));
5185 static bool cpu_overutilized(int cpu)
5187 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5190 #ifdef CONFIG_SCHED_TUNE
5192 static unsigned long
5193 schedtune_margin(unsigned long signal, unsigned long boost)
5195 unsigned long long margin = 0;
5198 * Signal proportional compensation (SPC)
5200 * The Boost (B) value is used to compute a Margin (M) which is
5201 * proportional to the complement of the original Signal (S):
5202 * M = B * (SCHED_LOAD_SCALE - S)
5203 * The obtained M could be used by the caller to "boost" S.
5205 margin = SCHED_LOAD_SCALE - signal;
5209 * Fast integer division by constant:
5210 * Constant : (C) = 100
5211 * Precision : 0.1% (P) = 0.1
5212 * Reference : C * 100 / P (R) = 100000
5215 * Shift bits : ceil(log(R,2)) (S) = 17
5216 * Mult const : round(2^S/C) (M) = 1311
5226 static inline unsigned int
5227 schedtune_cpu_margin(unsigned long util, int cpu)
5231 #ifdef CONFIG_CGROUP_SCHEDTUNE
5232 boost = schedtune_cpu_boost(cpu);
5234 boost = get_sysctl_sched_cfs_boost();
5239 return schedtune_margin(util, boost);
5242 static inline unsigned long
5243 schedtune_task_margin(struct task_struct *task)
5247 unsigned long margin;
5249 #ifdef CONFIG_CGROUP_SCHEDTUNE
5250 boost = schedtune_task_boost(task);
5252 boost = get_sysctl_sched_cfs_boost();
5257 util = task_util(task);
5258 margin = schedtune_margin(util, boost);
5263 #else /* CONFIG_SCHED_TUNE */
5265 static inline unsigned int
5266 schedtune_cpu_margin(unsigned long util, int cpu)
5271 static inline unsigned int
5272 schedtune_task_margin(struct task_struct *task)
5277 #endif /* CONFIG_SCHED_TUNE */
5279 static inline unsigned long
5280 boosted_cpu_util(int cpu)
5282 unsigned long util = cpu_util(cpu);
5283 unsigned long margin = schedtune_cpu_margin(util, cpu);
5285 return util + margin;
5288 static inline unsigned long
5289 boosted_task_util(struct task_struct *task)
5291 unsigned long util = task_util(task);
5292 unsigned long margin = schedtune_task_margin(task);
5294 return util + margin;
5298 * find_idlest_group finds and returns the least busy CPU group within the
5301 static struct sched_group *
5302 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5303 int this_cpu, int sd_flag)
5305 struct sched_group *idlest = NULL, *group = sd->groups;
5306 struct sched_group *fit_group = NULL, *spare_group = NULL;
5307 unsigned long min_load = ULONG_MAX, this_load = 0;
5308 unsigned long fit_capacity = ULONG_MAX;
5309 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5310 int load_idx = sd->forkexec_idx;
5311 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5313 if (sd_flag & SD_BALANCE_WAKE)
5314 load_idx = sd->wake_idx;
5317 unsigned long load, avg_load, spare_capacity;
5321 /* Skip over this group if it has no CPUs allowed */
5322 if (!cpumask_intersects(sched_group_cpus(group),
5323 tsk_cpus_allowed(p)))
5326 local_group = cpumask_test_cpu(this_cpu,
5327 sched_group_cpus(group));
5329 /* Tally up the load of all CPUs in the group */
5332 for_each_cpu(i, sched_group_cpus(group)) {
5333 /* Bias balancing toward cpus of our domain */
5335 load = source_load(i, load_idx);
5337 load = target_load(i, load_idx);
5342 * Look for most energy-efficient group that can fit
5343 * that can fit the task.
5345 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5346 fit_capacity = capacity_of(i);
5351 * Look for group which has most spare capacity on a
5354 spare_capacity = capacity_of(i) - cpu_util(i);
5355 if (spare_capacity > max_spare_capacity) {
5356 max_spare_capacity = spare_capacity;
5357 spare_group = group;
5361 /* Adjust by relative CPU capacity of the group */
5362 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5365 this_load = avg_load;
5366 } else if (avg_load < min_load) {
5367 min_load = avg_load;
5370 } while (group = group->next, group != sd->groups);
5378 if (!idlest || 100*this_load < imbalance*min_load)
5384 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5387 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5389 unsigned long load, min_load = ULONG_MAX;
5390 unsigned int min_exit_latency = UINT_MAX;
5391 u64 latest_idle_timestamp = 0;
5392 int least_loaded_cpu = this_cpu;
5393 int shallowest_idle_cpu = -1;
5396 /* Traverse only the allowed CPUs */
5397 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5398 if (task_fits_spare(p, i)) {
5399 struct rq *rq = cpu_rq(i);
5400 struct cpuidle_state *idle = idle_get_state(rq);
5401 if (idle && idle->exit_latency < min_exit_latency) {
5403 * We give priority to a CPU whose idle state
5404 * has the smallest exit latency irrespective
5405 * of any idle timestamp.
5407 min_exit_latency = idle->exit_latency;
5408 latest_idle_timestamp = rq->idle_stamp;
5409 shallowest_idle_cpu = i;
5410 } else if (idle_cpu(i) &&
5411 (!idle || idle->exit_latency == min_exit_latency) &&
5412 rq->idle_stamp > latest_idle_timestamp) {
5414 * If equal or no active idle state, then
5415 * the most recently idled CPU might have
5418 latest_idle_timestamp = rq->idle_stamp;
5419 shallowest_idle_cpu = i;
5420 } else if (shallowest_idle_cpu == -1) {
5422 * If we haven't found an idle CPU yet
5423 * pick a non-idle one that can fit the task as
5426 shallowest_idle_cpu = i;
5428 } else if (shallowest_idle_cpu == -1) {
5429 load = weighted_cpuload(i);
5430 if (load < min_load || (load == min_load && i == this_cpu)) {
5432 least_loaded_cpu = i;
5437 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5441 * Try and locate an idle CPU in the sched_domain.
5443 static int select_idle_sibling(struct task_struct *p, int target)
5445 struct sched_domain *sd;
5446 struct sched_group *sg;
5447 int i = task_cpu(p);
5449 if (idle_cpu(target))
5453 * If the prevous cpu is cache affine and idle, don't be stupid.
5455 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5459 * Otherwise, iterate the domains and find an elegible idle cpu.
5461 sd = rcu_dereference(per_cpu(sd_llc, target));
5462 for_each_lower_domain(sd) {
5465 if (!cpumask_intersects(sched_group_cpus(sg),
5466 tsk_cpus_allowed(p)))
5469 for_each_cpu(i, sched_group_cpus(sg)) {
5470 if (i == target || !idle_cpu(i))
5474 target = cpumask_first_and(sched_group_cpus(sg),
5475 tsk_cpus_allowed(p));
5479 } while (sg != sd->groups);
5485 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5487 struct sched_domain *sd;
5488 struct sched_group *sg, *sg_target;
5489 int target_max_cap = INT_MAX;
5490 int target_cpu = task_cpu(p);
5493 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5502 * Find group with sufficient capacity. We only get here if no cpu is
5503 * overutilized. We may end up overutilizing a cpu by adding the task,
5504 * but that should not be any worse than select_idle_sibling().
5505 * load_balance() should sort it out later as we get above the tipping
5509 /* Assuming all cpus are the same in group */
5510 int max_cap_cpu = group_first_cpu(sg);
5513 * Assume smaller max capacity means more energy-efficient.
5514 * Ideally we should query the energy model for the right
5515 * answer but it easily ends up in an exhaustive search.
5517 if (capacity_of(max_cap_cpu) < target_max_cap &&
5518 task_fits_max(p, max_cap_cpu)) {
5520 target_max_cap = capacity_of(max_cap_cpu);
5522 } while (sg = sg->next, sg != sd->groups);
5524 /* Find cpu with sufficient capacity */
5525 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5527 * p's blocked utilization is still accounted for on prev_cpu
5528 * so prev_cpu will receive a negative bias due to the double
5529 * accounting. However, the blocked utilization may be zero.
5531 int new_util = cpu_util(i) + boosted_task_util(p);
5533 if (new_util > capacity_orig_of(i))
5536 if (new_util < capacity_curr_of(i)) {
5538 if (cpu_rq(i)->nr_running)
5542 /* cpu has capacity at higher OPP, keep it as fallback */
5543 if (target_cpu == task_cpu(p))
5547 if (target_cpu != task_cpu(p)) {
5548 struct energy_env eenv = {
5549 .util_delta = task_util(p),
5550 .src_cpu = task_cpu(p),
5551 .dst_cpu = target_cpu,
5555 /* Not enough spare capacity on previous cpu */
5556 if (cpu_overutilized(task_cpu(p)))
5559 if (energy_diff(&eenv) >= 0)
5567 * select_task_rq_fair: Select target runqueue for the waking task in domains
5568 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5569 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5571 * Balances load by selecting the idlest cpu in the idlest group, or under
5572 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5574 * Returns the target cpu number.
5576 * preempt must be disabled.
5579 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5581 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5582 int cpu = smp_processor_id();
5583 int new_cpu = prev_cpu;
5584 int want_affine = 0;
5585 int sync = wake_flags & WF_SYNC;
5587 if (sd_flag & SD_BALANCE_WAKE)
5588 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5589 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5593 for_each_domain(cpu, tmp) {
5594 if (!(tmp->flags & SD_LOAD_BALANCE))
5598 * If both cpu and prev_cpu are part of this domain,
5599 * cpu is a valid SD_WAKE_AFFINE target.
5601 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5602 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5607 if (tmp->flags & sd_flag)
5609 else if (!want_affine)
5614 sd = NULL; /* Prefer wake_affine over balance flags */
5615 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5620 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5621 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5622 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5623 new_cpu = select_idle_sibling(p, new_cpu);
5626 struct sched_group *group;
5629 if (!(sd->flags & sd_flag)) {
5634 group = find_idlest_group(sd, p, cpu, sd_flag);
5640 new_cpu = find_idlest_cpu(group, p, cpu);
5641 if (new_cpu == -1 || new_cpu == cpu) {
5642 /* Now try balancing at a lower domain level of cpu */
5647 /* Now try balancing at a lower domain level of new_cpu */
5649 weight = sd->span_weight;
5651 for_each_domain(cpu, tmp) {
5652 if (weight <= tmp->span_weight)
5654 if (tmp->flags & sd_flag)
5657 /* while loop will break here if sd == NULL */
5665 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5666 * cfs_rq_of(p) references at time of call are still valid and identify the
5667 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5668 * other assumptions, including the state of rq->lock, should be made.
5670 static void migrate_task_rq_fair(struct task_struct *p)
5673 * We are supposed to update the task to "current" time, then its up to date
5674 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5675 * what current time is, so simply throw away the out-of-date time. This
5676 * will result in the wakee task is less decayed, but giving the wakee more
5677 * load sounds not bad.
5679 remove_entity_load_avg(&p->se);
5681 /* Tell new CPU we are migrated */
5682 p->se.avg.last_update_time = 0;
5684 /* We have migrated, no longer consider this task hot */
5685 p->se.exec_start = 0;
5688 static void task_dead_fair(struct task_struct *p)
5690 remove_entity_load_avg(&p->se);
5692 #endif /* CONFIG_SMP */
5694 static unsigned long
5695 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5697 unsigned long gran = sysctl_sched_wakeup_granularity;
5700 * Since its curr running now, convert the gran from real-time
5701 * to virtual-time in his units.
5703 * By using 'se' instead of 'curr' we penalize light tasks, so
5704 * they get preempted easier. That is, if 'se' < 'curr' then
5705 * the resulting gran will be larger, therefore penalizing the
5706 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5707 * be smaller, again penalizing the lighter task.
5709 * This is especially important for buddies when the leftmost
5710 * task is higher priority than the buddy.
5712 return calc_delta_fair(gran, se);
5716 * Should 'se' preempt 'curr'.
5730 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5732 s64 gran, vdiff = curr->vruntime - se->vruntime;
5737 gran = wakeup_gran(curr, se);
5744 static void set_last_buddy(struct sched_entity *se)
5746 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5749 for_each_sched_entity(se)
5750 cfs_rq_of(se)->last = se;
5753 static void set_next_buddy(struct sched_entity *se)
5755 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5758 for_each_sched_entity(se)
5759 cfs_rq_of(se)->next = se;
5762 static void set_skip_buddy(struct sched_entity *se)
5764 for_each_sched_entity(se)
5765 cfs_rq_of(se)->skip = se;
5769 * Preempt the current task with a newly woken task if needed:
5771 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5773 struct task_struct *curr = rq->curr;
5774 struct sched_entity *se = &curr->se, *pse = &p->se;
5775 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5776 int scale = cfs_rq->nr_running >= sched_nr_latency;
5777 int next_buddy_marked = 0;
5779 if (unlikely(se == pse))
5783 * This is possible from callers such as attach_tasks(), in which we
5784 * unconditionally check_prempt_curr() after an enqueue (which may have
5785 * lead to a throttle). This both saves work and prevents false
5786 * next-buddy nomination below.
5788 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5791 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5792 set_next_buddy(pse);
5793 next_buddy_marked = 1;
5797 * We can come here with TIF_NEED_RESCHED already set from new task
5800 * Note: this also catches the edge-case of curr being in a throttled
5801 * group (e.g. via set_curr_task), since update_curr() (in the
5802 * enqueue of curr) will have resulted in resched being set. This
5803 * prevents us from potentially nominating it as a false LAST_BUDDY
5806 if (test_tsk_need_resched(curr))
5809 /* Idle tasks are by definition preempted by non-idle tasks. */
5810 if (unlikely(curr->policy == SCHED_IDLE) &&
5811 likely(p->policy != SCHED_IDLE))
5815 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5816 * is driven by the tick):
5818 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5821 find_matching_se(&se, &pse);
5822 update_curr(cfs_rq_of(se));
5824 if (wakeup_preempt_entity(se, pse) == 1) {
5826 * Bias pick_next to pick the sched entity that is
5827 * triggering this preemption.
5829 if (!next_buddy_marked)
5830 set_next_buddy(pse);
5839 * Only set the backward buddy when the current task is still
5840 * on the rq. This can happen when a wakeup gets interleaved
5841 * with schedule on the ->pre_schedule() or idle_balance()
5842 * point, either of which can * drop the rq lock.
5844 * Also, during early boot the idle thread is in the fair class,
5845 * for obvious reasons its a bad idea to schedule back to it.
5847 if (unlikely(!se->on_rq || curr == rq->idle))
5850 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5854 static struct task_struct *
5855 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5857 struct cfs_rq *cfs_rq = &rq->cfs;
5858 struct sched_entity *se;
5859 struct task_struct *p;
5863 #ifdef CONFIG_FAIR_GROUP_SCHED
5864 if (!cfs_rq->nr_running)
5867 if (prev->sched_class != &fair_sched_class)
5871 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5872 * likely that a next task is from the same cgroup as the current.
5874 * Therefore attempt to avoid putting and setting the entire cgroup
5875 * hierarchy, only change the part that actually changes.
5879 struct sched_entity *curr = cfs_rq->curr;
5882 * Since we got here without doing put_prev_entity() we also
5883 * have to consider cfs_rq->curr. If it is still a runnable
5884 * entity, update_curr() will update its vruntime, otherwise
5885 * forget we've ever seen it.
5889 update_curr(cfs_rq);
5894 * This call to check_cfs_rq_runtime() will do the
5895 * throttle and dequeue its entity in the parent(s).
5896 * Therefore the 'simple' nr_running test will indeed
5899 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5903 se = pick_next_entity(cfs_rq, curr);
5904 cfs_rq = group_cfs_rq(se);
5910 * Since we haven't yet done put_prev_entity and if the selected task
5911 * is a different task than we started out with, try and touch the
5912 * least amount of cfs_rqs.
5915 struct sched_entity *pse = &prev->se;
5917 while (!(cfs_rq = is_same_group(se, pse))) {
5918 int se_depth = se->depth;
5919 int pse_depth = pse->depth;
5921 if (se_depth <= pse_depth) {
5922 put_prev_entity(cfs_rq_of(pse), pse);
5923 pse = parent_entity(pse);
5925 if (se_depth >= pse_depth) {
5926 set_next_entity(cfs_rq_of(se), se);
5927 se = parent_entity(se);
5931 put_prev_entity(cfs_rq, pse);
5932 set_next_entity(cfs_rq, se);
5935 if (hrtick_enabled(rq))
5936 hrtick_start_fair(rq, p);
5938 rq->misfit_task = !task_fits_max(p, rq->cpu);
5945 if (!cfs_rq->nr_running)
5948 put_prev_task(rq, prev);
5951 se = pick_next_entity(cfs_rq, NULL);
5952 set_next_entity(cfs_rq, se);
5953 cfs_rq = group_cfs_rq(se);
5958 if (hrtick_enabled(rq))
5959 hrtick_start_fair(rq, p);
5961 rq->misfit_task = !task_fits_max(p, rq->cpu);
5966 rq->misfit_task = 0;
5968 * This is OK, because current is on_cpu, which avoids it being picked
5969 * for load-balance and preemption/IRQs are still disabled avoiding
5970 * further scheduler activity on it and we're being very careful to
5971 * re-start the picking loop.
5973 lockdep_unpin_lock(&rq->lock);
5974 new_tasks = idle_balance(rq);
5975 lockdep_pin_lock(&rq->lock);
5977 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5978 * possible for any higher priority task to appear. In that case we
5979 * must re-start the pick_next_entity() loop.
5991 * Account for a descheduled task:
5993 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5995 struct sched_entity *se = &prev->se;
5996 struct cfs_rq *cfs_rq;
5998 for_each_sched_entity(se) {
5999 cfs_rq = cfs_rq_of(se);
6000 put_prev_entity(cfs_rq, se);
6005 * sched_yield() is very simple
6007 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6009 static void yield_task_fair(struct rq *rq)
6011 struct task_struct *curr = rq->curr;
6012 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6013 struct sched_entity *se = &curr->se;
6016 * Are we the only task in the tree?
6018 if (unlikely(rq->nr_running == 1))
6021 clear_buddies(cfs_rq, se);
6023 if (curr->policy != SCHED_BATCH) {
6024 update_rq_clock(rq);
6026 * Update run-time statistics of the 'current'.
6028 update_curr(cfs_rq);
6030 * Tell update_rq_clock() that we've just updated,
6031 * so we don't do microscopic update in schedule()
6032 * and double the fastpath cost.
6034 rq_clock_skip_update(rq, true);
6040 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6042 struct sched_entity *se = &p->se;
6044 /* throttled hierarchies are not runnable */
6045 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6048 /* Tell the scheduler that we'd really like pse to run next. */
6051 yield_task_fair(rq);
6057 /**************************************************
6058 * Fair scheduling class load-balancing methods.
6062 * The purpose of load-balancing is to achieve the same basic fairness the
6063 * per-cpu scheduler provides, namely provide a proportional amount of compute
6064 * time to each task. This is expressed in the following equation:
6066 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6068 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6069 * W_i,0 is defined as:
6071 * W_i,0 = \Sum_j w_i,j (2)
6073 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6074 * is derived from the nice value as per prio_to_weight[].
6076 * The weight average is an exponential decay average of the instantaneous
6079 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6081 * C_i is the compute capacity of cpu i, typically it is the
6082 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6083 * can also include other factors [XXX].
6085 * To achieve this balance we define a measure of imbalance which follows
6086 * directly from (1):
6088 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6090 * We them move tasks around to minimize the imbalance. In the continuous
6091 * function space it is obvious this converges, in the discrete case we get
6092 * a few fun cases generally called infeasible weight scenarios.
6095 * - infeasible weights;
6096 * - local vs global optima in the discrete case. ]
6101 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6102 * for all i,j solution, we create a tree of cpus that follows the hardware
6103 * topology where each level pairs two lower groups (or better). This results
6104 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6105 * tree to only the first of the previous level and we decrease the frequency
6106 * of load-balance at each level inv. proportional to the number of cpus in
6112 * \Sum { --- * --- * 2^i } = O(n) (5)
6114 * `- size of each group
6115 * | | `- number of cpus doing load-balance
6117 * `- sum over all levels
6119 * Coupled with a limit on how many tasks we can migrate every balance pass,
6120 * this makes (5) the runtime complexity of the balancer.
6122 * An important property here is that each CPU is still (indirectly) connected
6123 * to every other cpu in at most O(log n) steps:
6125 * The adjacency matrix of the resulting graph is given by:
6128 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6131 * And you'll find that:
6133 * A^(log_2 n)_i,j != 0 for all i,j (7)
6135 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6136 * The task movement gives a factor of O(m), giving a convergence complexity
6139 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6144 * In order to avoid CPUs going idle while there's still work to do, new idle
6145 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6146 * tree itself instead of relying on other CPUs to bring it work.
6148 * This adds some complexity to both (5) and (8) but it reduces the total idle
6156 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6159 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6164 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6166 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6168 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6171 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6172 * rewrite all of this once again.]
6175 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6177 enum fbq_type { regular, remote, all };
6186 #define LBF_ALL_PINNED 0x01
6187 #define LBF_NEED_BREAK 0x02
6188 #define LBF_DST_PINNED 0x04
6189 #define LBF_SOME_PINNED 0x08
6192 struct sched_domain *sd;
6200 struct cpumask *dst_grpmask;
6202 enum cpu_idle_type idle;
6204 unsigned int src_grp_nr_running;
6205 /* The set of CPUs under consideration for load-balancing */
6206 struct cpumask *cpus;
6211 unsigned int loop_break;
6212 unsigned int loop_max;
6214 enum fbq_type fbq_type;
6215 enum group_type busiest_group_type;
6216 struct list_head tasks;
6220 * Is this task likely cache-hot:
6222 static int task_hot(struct task_struct *p, struct lb_env *env)
6226 lockdep_assert_held(&env->src_rq->lock);
6228 if (p->sched_class != &fair_sched_class)
6231 if (unlikely(p->policy == SCHED_IDLE))
6235 * Buddy candidates are cache hot:
6237 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6238 (&p->se == cfs_rq_of(&p->se)->next ||
6239 &p->se == cfs_rq_of(&p->se)->last))
6242 if (sysctl_sched_migration_cost == -1)
6244 if (sysctl_sched_migration_cost == 0)
6247 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6249 return delta < (s64)sysctl_sched_migration_cost;
6252 #ifdef CONFIG_NUMA_BALANCING
6254 * Returns 1, if task migration degrades locality
6255 * Returns 0, if task migration improves locality i.e migration preferred.
6256 * Returns -1, if task migration is not affected by locality.
6258 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6260 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6261 unsigned long src_faults, dst_faults;
6262 int src_nid, dst_nid;
6264 if (!static_branch_likely(&sched_numa_balancing))
6267 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6270 src_nid = cpu_to_node(env->src_cpu);
6271 dst_nid = cpu_to_node(env->dst_cpu);
6273 if (src_nid == dst_nid)
6276 /* Migrating away from the preferred node is always bad. */
6277 if (src_nid == p->numa_preferred_nid) {
6278 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6284 /* Encourage migration to the preferred node. */
6285 if (dst_nid == p->numa_preferred_nid)
6289 src_faults = group_faults(p, src_nid);
6290 dst_faults = group_faults(p, dst_nid);
6292 src_faults = task_faults(p, src_nid);
6293 dst_faults = task_faults(p, dst_nid);
6296 return dst_faults < src_faults;
6300 static inline int migrate_degrades_locality(struct task_struct *p,
6308 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6311 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6315 lockdep_assert_held(&env->src_rq->lock);
6318 * We do not migrate tasks that are:
6319 * 1) throttled_lb_pair, or
6320 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6321 * 3) running (obviously), or
6322 * 4) are cache-hot on their current CPU.
6324 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6327 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6330 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6332 env->flags |= LBF_SOME_PINNED;
6335 * Remember if this task can be migrated to any other cpu in
6336 * our sched_group. We may want to revisit it if we couldn't
6337 * meet load balance goals by pulling other tasks on src_cpu.
6339 * Also avoid computing new_dst_cpu if we have already computed
6340 * one in current iteration.
6342 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6345 /* Prevent to re-select dst_cpu via env's cpus */
6346 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6347 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6348 env->flags |= LBF_DST_PINNED;
6349 env->new_dst_cpu = cpu;
6357 /* Record that we found atleast one task that could run on dst_cpu */
6358 env->flags &= ~LBF_ALL_PINNED;
6360 if (task_running(env->src_rq, p)) {
6361 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6366 * Aggressive migration if:
6367 * 1) destination numa is preferred
6368 * 2) task is cache cold, or
6369 * 3) too many balance attempts have failed.
6371 tsk_cache_hot = migrate_degrades_locality(p, env);
6372 if (tsk_cache_hot == -1)
6373 tsk_cache_hot = task_hot(p, env);
6375 if (tsk_cache_hot <= 0 ||
6376 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6377 if (tsk_cache_hot == 1) {
6378 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6379 schedstat_inc(p, se.statistics.nr_forced_migrations);
6384 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6389 * detach_task() -- detach the task for the migration specified in env
6391 static void detach_task(struct task_struct *p, struct lb_env *env)
6393 lockdep_assert_held(&env->src_rq->lock);
6395 deactivate_task(env->src_rq, p, 0);
6396 p->on_rq = TASK_ON_RQ_MIGRATING;
6397 set_task_cpu(p, env->dst_cpu);
6401 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6402 * part of active balancing operations within "domain".
6404 * Returns a task if successful and NULL otherwise.
6406 static struct task_struct *detach_one_task(struct lb_env *env)
6408 struct task_struct *p, *n;
6410 lockdep_assert_held(&env->src_rq->lock);
6412 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6413 if (!can_migrate_task(p, env))
6416 detach_task(p, env);
6419 * Right now, this is only the second place where
6420 * lb_gained[env->idle] is updated (other is detach_tasks)
6421 * so we can safely collect stats here rather than
6422 * inside detach_tasks().
6424 schedstat_inc(env->sd, lb_gained[env->idle]);
6430 static const unsigned int sched_nr_migrate_break = 32;
6433 * detach_tasks() -- tries to detach up to imbalance weighted load from
6434 * busiest_rq, as part of a balancing operation within domain "sd".
6436 * Returns number of detached tasks if successful and 0 otherwise.
6438 static int detach_tasks(struct lb_env *env)
6440 struct list_head *tasks = &env->src_rq->cfs_tasks;
6441 struct task_struct *p;
6445 lockdep_assert_held(&env->src_rq->lock);
6447 if (env->imbalance <= 0)
6450 while (!list_empty(tasks)) {
6452 * We don't want to steal all, otherwise we may be treated likewise,
6453 * which could at worst lead to a livelock crash.
6455 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6458 p = list_first_entry(tasks, struct task_struct, se.group_node);
6461 /* We've more or less seen every task there is, call it quits */
6462 if (env->loop > env->loop_max)
6465 /* take a breather every nr_migrate tasks */
6466 if (env->loop > env->loop_break) {
6467 env->loop_break += sched_nr_migrate_break;
6468 env->flags |= LBF_NEED_BREAK;
6472 if (!can_migrate_task(p, env))
6475 load = task_h_load(p);
6477 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6480 if ((load / 2) > env->imbalance)
6483 detach_task(p, env);
6484 list_add(&p->se.group_node, &env->tasks);
6487 env->imbalance -= load;
6489 #ifdef CONFIG_PREEMPT
6491 * NEWIDLE balancing is a source of latency, so preemptible
6492 * kernels will stop after the first task is detached to minimize
6493 * the critical section.
6495 if (env->idle == CPU_NEWLY_IDLE)
6500 * We only want to steal up to the prescribed amount of
6503 if (env->imbalance <= 0)
6508 list_move_tail(&p->se.group_node, tasks);
6512 * Right now, this is one of only two places we collect this stat
6513 * so we can safely collect detach_one_task() stats here rather
6514 * than inside detach_one_task().
6516 schedstat_add(env->sd, lb_gained[env->idle], detached);
6522 * attach_task() -- attach the task detached by detach_task() to its new rq.
6524 static void attach_task(struct rq *rq, struct task_struct *p)
6526 lockdep_assert_held(&rq->lock);
6528 BUG_ON(task_rq(p) != rq);
6529 p->on_rq = TASK_ON_RQ_QUEUED;
6530 activate_task(rq, p, 0);
6531 check_preempt_curr(rq, p, 0);
6535 * attach_one_task() -- attaches the task returned from detach_one_task() to
6538 static void attach_one_task(struct rq *rq, struct task_struct *p)
6540 raw_spin_lock(&rq->lock);
6543 * We want to potentially raise target_cpu's OPP.
6545 update_capacity_of(cpu_of(rq));
6546 raw_spin_unlock(&rq->lock);
6550 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6553 static void attach_tasks(struct lb_env *env)
6555 struct list_head *tasks = &env->tasks;
6556 struct task_struct *p;
6558 raw_spin_lock(&env->dst_rq->lock);
6560 while (!list_empty(tasks)) {
6561 p = list_first_entry(tasks, struct task_struct, se.group_node);
6562 list_del_init(&p->se.group_node);
6564 attach_task(env->dst_rq, p);
6568 * We want to potentially raise env.dst_cpu's OPP.
6570 update_capacity_of(env->dst_cpu);
6572 raw_spin_unlock(&env->dst_rq->lock);
6575 #ifdef CONFIG_FAIR_GROUP_SCHED
6576 static void update_blocked_averages(int cpu)
6578 struct rq *rq = cpu_rq(cpu);
6579 struct cfs_rq *cfs_rq;
6580 unsigned long flags;
6582 raw_spin_lock_irqsave(&rq->lock, flags);
6583 update_rq_clock(rq);
6586 * Iterates the task_group tree in a bottom up fashion, see
6587 * list_add_leaf_cfs_rq() for details.
6589 for_each_leaf_cfs_rq(rq, cfs_rq) {
6590 /* throttled entities do not contribute to load */
6591 if (throttled_hierarchy(cfs_rq))
6594 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6595 update_tg_load_avg(cfs_rq, 0);
6597 raw_spin_unlock_irqrestore(&rq->lock, flags);
6601 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6602 * This needs to be done in a top-down fashion because the load of a child
6603 * group is a fraction of its parents load.
6605 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6607 struct rq *rq = rq_of(cfs_rq);
6608 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6609 unsigned long now = jiffies;
6612 if (cfs_rq->last_h_load_update == now)
6615 cfs_rq->h_load_next = NULL;
6616 for_each_sched_entity(se) {
6617 cfs_rq = cfs_rq_of(se);
6618 cfs_rq->h_load_next = se;
6619 if (cfs_rq->last_h_load_update == now)
6624 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6625 cfs_rq->last_h_load_update = now;
6628 while ((se = cfs_rq->h_load_next) != NULL) {
6629 load = cfs_rq->h_load;
6630 load = div64_ul(load * se->avg.load_avg,
6631 cfs_rq_load_avg(cfs_rq) + 1);
6632 cfs_rq = group_cfs_rq(se);
6633 cfs_rq->h_load = load;
6634 cfs_rq->last_h_load_update = now;
6638 static unsigned long task_h_load(struct task_struct *p)
6640 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6642 update_cfs_rq_h_load(cfs_rq);
6643 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6644 cfs_rq_load_avg(cfs_rq) + 1);
6647 static inline void update_blocked_averages(int cpu)
6649 struct rq *rq = cpu_rq(cpu);
6650 struct cfs_rq *cfs_rq = &rq->cfs;
6651 unsigned long flags;
6653 raw_spin_lock_irqsave(&rq->lock, flags);
6654 update_rq_clock(rq);
6655 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6656 raw_spin_unlock_irqrestore(&rq->lock, flags);
6659 static unsigned long task_h_load(struct task_struct *p)
6661 return p->se.avg.load_avg;
6665 /********** Helpers for find_busiest_group ************************/
6668 * sg_lb_stats - stats of a sched_group required for load_balancing
6670 struct sg_lb_stats {
6671 unsigned long avg_load; /*Avg load across the CPUs of the group */
6672 unsigned long group_load; /* Total load over the CPUs of the group */
6673 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6674 unsigned long load_per_task;
6675 unsigned long group_capacity;
6676 unsigned long group_util; /* Total utilization of the group */
6677 unsigned int sum_nr_running; /* Nr tasks running in the group */
6678 unsigned int idle_cpus;
6679 unsigned int group_weight;
6680 enum group_type group_type;
6681 int group_no_capacity;
6682 int group_misfit_task; /* A cpu has a task too big for its capacity */
6683 #ifdef CONFIG_NUMA_BALANCING
6684 unsigned int nr_numa_running;
6685 unsigned int nr_preferred_running;
6690 * sd_lb_stats - Structure to store the statistics of a sched_domain
6691 * during load balancing.
6693 struct sd_lb_stats {
6694 struct sched_group *busiest; /* Busiest group in this sd */
6695 struct sched_group *local; /* Local group in this sd */
6696 unsigned long total_load; /* Total load of all groups in sd */
6697 unsigned long total_capacity; /* Total capacity of all groups in sd */
6698 unsigned long avg_load; /* Average load across all groups in sd */
6700 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6701 struct sg_lb_stats local_stat; /* Statistics of the local group */
6704 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6707 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6708 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6709 * We must however clear busiest_stat::avg_load because
6710 * update_sd_pick_busiest() reads this before assignment.
6712 *sds = (struct sd_lb_stats){
6716 .total_capacity = 0UL,
6719 .sum_nr_running = 0,
6720 .group_type = group_other,
6726 * get_sd_load_idx - Obtain the load index for a given sched domain.
6727 * @sd: The sched_domain whose load_idx is to be obtained.
6728 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6730 * Return: The load index.
6732 static inline int get_sd_load_idx(struct sched_domain *sd,
6733 enum cpu_idle_type idle)
6739 load_idx = sd->busy_idx;
6742 case CPU_NEWLY_IDLE:
6743 load_idx = sd->newidle_idx;
6746 load_idx = sd->idle_idx;
6753 static unsigned long scale_rt_capacity(int cpu)
6755 struct rq *rq = cpu_rq(cpu);
6756 u64 total, used, age_stamp, avg;
6760 * Since we're reading these variables without serialization make sure
6761 * we read them once before doing sanity checks on them.
6763 age_stamp = READ_ONCE(rq->age_stamp);
6764 avg = READ_ONCE(rq->rt_avg);
6765 delta = __rq_clock_broken(rq) - age_stamp;
6767 if (unlikely(delta < 0))
6770 total = sched_avg_period() + delta;
6772 used = div_u64(avg, total);
6775 * deadline bandwidth is defined at system level so we must
6776 * weight this bandwidth with the max capacity of the system.
6777 * As a reminder, avg_bw is 20bits width and
6778 * scale_cpu_capacity is 10 bits width
6780 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6782 if (likely(used < SCHED_CAPACITY_SCALE))
6783 return SCHED_CAPACITY_SCALE - used;
6788 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6790 raw_spin_lock_init(&mcc->lock);
6795 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6797 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6798 struct sched_group *sdg = sd->groups;
6799 struct max_cpu_capacity *mcc;
6800 unsigned long max_capacity;
6802 unsigned long flags;
6804 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6806 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6808 raw_spin_lock_irqsave(&mcc->lock, flags);
6809 max_capacity = mcc->val;
6810 max_cap_cpu = mcc->cpu;
6812 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6813 (max_capacity < capacity)) {
6814 mcc->val = capacity;
6816 #ifdef CONFIG_SCHED_DEBUG
6817 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6818 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6822 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6824 skip_unlock: __attribute__ ((unused));
6825 capacity *= scale_rt_capacity(cpu);
6826 capacity >>= SCHED_CAPACITY_SHIFT;
6831 cpu_rq(cpu)->cpu_capacity = capacity;
6832 sdg->sgc->capacity = capacity;
6833 sdg->sgc->max_capacity = capacity;
6836 void update_group_capacity(struct sched_domain *sd, int cpu)
6838 struct sched_domain *child = sd->child;
6839 struct sched_group *group, *sdg = sd->groups;
6840 unsigned long capacity, max_capacity;
6841 unsigned long interval;
6843 interval = msecs_to_jiffies(sd->balance_interval);
6844 interval = clamp(interval, 1UL, max_load_balance_interval);
6845 sdg->sgc->next_update = jiffies + interval;
6848 update_cpu_capacity(sd, cpu);
6855 if (child->flags & SD_OVERLAP) {
6857 * SD_OVERLAP domains cannot assume that child groups
6858 * span the current group.
6861 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6862 struct sched_group_capacity *sgc;
6863 struct rq *rq = cpu_rq(cpu);
6866 * build_sched_domains() -> init_sched_groups_capacity()
6867 * gets here before we've attached the domains to the
6870 * Use capacity_of(), which is set irrespective of domains
6871 * in update_cpu_capacity().
6873 * This avoids capacity from being 0 and
6874 * causing divide-by-zero issues on boot.
6876 if (unlikely(!rq->sd)) {
6877 capacity += capacity_of(cpu);
6879 sgc = rq->sd->groups->sgc;
6880 capacity += sgc->capacity;
6883 max_capacity = max(capacity, max_capacity);
6887 * !SD_OVERLAP domains can assume that child groups
6888 * span the current group.
6891 group = child->groups;
6893 struct sched_group_capacity *sgc = group->sgc;
6895 capacity += sgc->capacity;
6896 max_capacity = max(sgc->max_capacity, max_capacity);
6897 group = group->next;
6898 } while (group != child->groups);
6901 sdg->sgc->capacity = capacity;
6902 sdg->sgc->max_capacity = max_capacity;
6906 * Check whether the capacity of the rq has been noticeably reduced by side
6907 * activity. The imbalance_pct is used for the threshold.
6908 * Return true is the capacity is reduced
6911 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6913 return ((rq->cpu_capacity * sd->imbalance_pct) <
6914 (rq->cpu_capacity_orig * 100));
6918 * Group imbalance indicates (and tries to solve) the problem where balancing
6919 * groups is inadequate due to tsk_cpus_allowed() constraints.
6921 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6922 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6925 * { 0 1 2 3 } { 4 5 6 7 }
6928 * If we were to balance group-wise we'd place two tasks in the first group and
6929 * two tasks in the second group. Clearly this is undesired as it will overload
6930 * cpu 3 and leave one of the cpus in the second group unused.
6932 * The current solution to this issue is detecting the skew in the first group
6933 * by noticing the lower domain failed to reach balance and had difficulty
6934 * moving tasks due to affinity constraints.
6936 * When this is so detected; this group becomes a candidate for busiest; see
6937 * update_sd_pick_busiest(). And calculate_imbalance() and
6938 * find_busiest_group() avoid some of the usual balance conditions to allow it
6939 * to create an effective group imbalance.
6941 * This is a somewhat tricky proposition since the next run might not find the
6942 * group imbalance and decide the groups need to be balanced again. A most
6943 * subtle and fragile situation.
6946 static inline int sg_imbalanced(struct sched_group *group)
6948 return group->sgc->imbalance;
6952 * group_has_capacity returns true if the group has spare capacity that could
6953 * be used by some tasks.
6954 * We consider that a group has spare capacity if the * number of task is
6955 * smaller than the number of CPUs or if the utilization is lower than the
6956 * available capacity for CFS tasks.
6957 * For the latter, we use a threshold to stabilize the state, to take into
6958 * account the variance of the tasks' load and to return true if the available
6959 * capacity in meaningful for the load balancer.
6960 * As an example, an available capacity of 1% can appear but it doesn't make
6961 * any benefit for the load balance.
6964 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6966 if (sgs->sum_nr_running < sgs->group_weight)
6969 if ((sgs->group_capacity * 100) >
6970 (sgs->group_util * env->sd->imbalance_pct))
6977 * group_is_overloaded returns true if the group has more tasks than it can
6979 * group_is_overloaded is not equals to !group_has_capacity because a group
6980 * with the exact right number of tasks, has no more spare capacity but is not
6981 * overloaded so both group_has_capacity and group_is_overloaded return
6985 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6987 if (sgs->sum_nr_running <= sgs->group_weight)
6990 if ((sgs->group_capacity * 100) <
6991 (sgs->group_util * env->sd->imbalance_pct))
6999 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7000 * per-cpu capacity than sched_group ref.
7003 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7005 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7006 ref->sgc->max_capacity;
7010 group_type group_classify(struct sched_group *group,
7011 struct sg_lb_stats *sgs)
7013 if (sgs->group_no_capacity)
7014 return group_overloaded;
7016 if (sg_imbalanced(group))
7017 return group_imbalanced;
7019 if (sgs->group_misfit_task)
7020 return group_misfit_task;
7026 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7027 * @env: The load balancing environment.
7028 * @group: sched_group whose statistics are to be updated.
7029 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7030 * @local_group: Does group contain this_cpu.
7031 * @sgs: variable to hold the statistics for this group.
7032 * @overload: Indicate more than one runnable task for any CPU.
7033 * @overutilized: Indicate overutilization for any CPU.
7035 static inline void update_sg_lb_stats(struct lb_env *env,
7036 struct sched_group *group, int load_idx,
7037 int local_group, struct sg_lb_stats *sgs,
7038 bool *overload, bool *overutilized)
7043 memset(sgs, 0, sizeof(*sgs));
7045 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7046 struct rq *rq = cpu_rq(i);
7048 /* Bias balancing toward cpus of our domain */
7050 load = target_load(i, load_idx);
7052 load = source_load(i, load_idx);
7054 sgs->group_load += load;
7055 sgs->group_util += cpu_util(i);
7056 sgs->sum_nr_running += rq->cfs.h_nr_running;
7058 if (rq->nr_running > 1)
7061 #ifdef CONFIG_NUMA_BALANCING
7062 sgs->nr_numa_running += rq->nr_numa_running;
7063 sgs->nr_preferred_running += rq->nr_preferred_running;
7065 sgs->sum_weighted_load += weighted_cpuload(i);
7069 if (cpu_overutilized(i)) {
7070 *overutilized = true;
7071 if (!sgs->group_misfit_task && rq->misfit_task)
7072 sgs->group_misfit_task = capacity_of(i);
7076 /* Adjust by relative CPU capacity of the group */
7077 sgs->group_capacity = group->sgc->capacity;
7078 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7080 if (sgs->sum_nr_running)
7081 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7083 sgs->group_weight = group->group_weight;
7085 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7086 sgs->group_type = group_classify(group, sgs);
7090 * update_sd_pick_busiest - return 1 on busiest group
7091 * @env: The load balancing environment.
7092 * @sds: sched_domain statistics
7093 * @sg: sched_group candidate to be checked for being the busiest
7094 * @sgs: sched_group statistics
7096 * Determine if @sg is a busier group than the previously selected
7099 * Return: %true if @sg is a busier group than the previously selected
7100 * busiest group. %false otherwise.
7102 static bool update_sd_pick_busiest(struct lb_env *env,
7103 struct sd_lb_stats *sds,
7104 struct sched_group *sg,
7105 struct sg_lb_stats *sgs)
7107 struct sg_lb_stats *busiest = &sds->busiest_stat;
7109 if (sgs->group_type > busiest->group_type)
7112 if (sgs->group_type < busiest->group_type)
7116 * Candidate sg doesn't face any serious load-balance problems
7117 * so don't pick it if the local sg is already filled up.
7119 if (sgs->group_type == group_other &&
7120 !group_has_capacity(env, &sds->local_stat))
7123 if (sgs->avg_load <= busiest->avg_load)
7127 * Candiate sg has no more than one task per cpu and has higher
7128 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7130 if (sgs->sum_nr_running <= sgs->group_weight &&
7131 group_smaller_cpu_capacity(sds->local, sg))
7134 /* This is the busiest node in its class. */
7135 if (!(env->sd->flags & SD_ASYM_PACKING))
7139 * ASYM_PACKING needs to move all the work to the lowest
7140 * numbered CPUs in the group, therefore mark all groups
7141 * higher than ourself as busy.
7143 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7147 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7154 #ifdef CONFIG_NUMA_BALANCING
7155 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7157 if (sgs->sum_nr_running > sgs->nr_numa_running)
7159 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7164 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7166 if (rq->nr_running > rq->nr_numa_running)
7168 if (rq->nr_running > rq->nr_preferred_running)
7173 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7178 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7182 #endif /* CONFIG_NUMA_BALANCING */
7185 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7186 * @env: The load balancing environment.
7187 * @sds: variable to hold the statistics for this sched_domain.
7189 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7191 struct sched_domain *child = env->sd->child;
7192 struct sched_group *sg = env->sd->groups;
7193 struct sg_lb_stats tmp_sgs;
7194 int load_idx, prefer_sibling = 0;
7195 bool overload = false, overutilized = false;
7197 if (child && child->flags & SD_PREFER_SIBLING)
7200 load_idx = get_sd_load_idx(env->sd, env->idle);
7203 struct sg_lb_stats *sgs = &tmp_sgs;
7206 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7209 sgs = &sds->local_stat;
7211 if (env->idle != CPU_NEWLY_IDLE ||
7212 time_after_eq(jiffies, sg->sgc->next_update))
7213 update_group_capacity(env->sd, env->dst_cpu);
7216 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7217 &overload, &overutilized);
7223 * In case the child domain prefers tasks go to siblings
7224 * first, lower the sg capacity so that we'll try
7225 * and move all the excess tasks away. We lower the capacity
7226 * of a group only if the local group has the capacity to fit
7227 * these excess tasks. The extra check prevents the case where
7228 * you always pull from the heaviest group when it is already
7229 * under-utilized (possible with a large weight task outweighs
7230 * the tasks on the system).
7232 if (prefer_sibling && sds->local &&
7233 group_has_capacity(env, &sds->local_stat) &&
7234 (sgs->sum_nr_running > 1)) {
7235 sgs->group_no_capacity = 1;
7236 sgs->group_type = group_classify(sg, sgs);
7240 * Ignore task groups with misfit tasks if local group has no
7241 * capacity or if per-cpu capacity isn't higher.
7243 if (sgs->group_type == group_misfit_task &&
7244 (!group_has_capacity(env, &sds->local_stat) ||
7245 !group_smaller_cpu_capacity(sg, sds->local)))
7246 sgs->group_type = group_other;
7248 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7250 sds->busiest_stat = *sgs;
7254 /* Now, start updating sd_lb_stats */
7255 sds->total_load += sgs->group_load;
7256 sds->total_capacity += sgs->group_capacity;
7259 } while (sg != env->sd->groups);
7261 if (env->sd->flags & SD_NUMA)
7262 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7264 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7266 if (!env->sd->parent) {
7267 /* update overload indicator if we are at root domain */
7268 if (env->dst_rq->rd->overload != overload)
7269 env->dst_rq->rd->overload = overload;
7271 /* Update over-utilization (tipping point, U >= 0) indicator */
7272 if (env->dst_rq->rd->overutilized != overutilized)
7273 env->dst_rq->rd->overutilized = overutilized;
7275 if (!env->dst_rq->rd->overutilized && overutilized)
7276 env->dst_rq->rd->overutilized = true;
7281 * check_asym_packing - Check to see if the group is packed into the
7284 * This is primarily intended to used at the sibling level. Some
7285 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7286 * case of POWER7, it can move to lower SMT modes only when higher
7287 * threads are idle. When in lower SMT modes, the threads will
7288 * perform better since they share less core resources. Hence when we
7289 * have idle threads, we want them to be the higher ones.
7291 * This packing function is run on idle threads. It checks to see if
7292 * the busiest CPU in this domain (core in the P7 case) has a higher
7293 * CPU number than the packing function is being run on. Here we are
7294 * assuming lower CPU number will be equivalent to lower a SMT thread
7297 * Return: 1 when packing is required and a task should be moved to
7298 * this CPU. The amount of the imbalance is returned in *imbalance.
7300 * @env: The load balancing environment.
7301 * @sds: Statistics of the sched_domain which is to be packed
7303 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7307 if (!(env->sd->flags & SD_ASYM_PACKING))
7313 busiest_cpu = group_first_cpu(sds->busiest);
7314 if (env->dst_cpu > busiest_cpu)
7317 env->imbalance = DIV_ROUND_CLOSEST(
7318 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7319 SCHED_CAPACITY_SCALE);
7325 * fix_small_imbalance - Calculate the minor imbalance that exists
7326 * amongst the groups of a sched_domain, during
7328 * @env: The load balancing environment.
7329 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7332 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7334 unsigned long tmp, capa_now = 0, capa_move = 0;
7335 unsigned int imbn = 2;
7336 unsigned long scaled_busy_load_per_task;
7337 struct sg_lb_stats *local, *busiest;
7339 local = &sds->local_stat;
7340 busiest = &sds->busiest_stat;
7342 if (!local->sum_nr_running)
7343 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7344 else if (busiest->load_per_task > local->load_per_task)
7347 scaled_busy_load_per_task =
7348 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7349 busiest->group_capacity;
7351 if (busiest->avg_load + scaled_busy_load_per_task >=
7352 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7353 env->imbalance = busiest->load_per_task;
7358 * OK, we don't have enough imbalance to justify moving tasks,
7359 * however we may be able to increase total CPU capacity used by
7363 capa_now += busiest->group_capacity *
7364 min(busiest->load_per_task, busiest->avg_load);
7365 capa_now += local->group_capacity *
7366 min(local->load_per_task, local->avg_load);
7367 capa_now /= SCHED_CAPACITY_SCALE;
7369 /* Amount of load we'd subtract */
7370 if (busiest->avg_load > scaled_busy_load_per_task) {
7371 capa_move += busiest->group_capacity *
7372 min(busiest->load_per_task,
7373 busiest->avg_load - scaled_busy_load_per_task);
7376 /* Amount of load we'd add */
7377 if (busiest->avg_load * busiest->group_capacity <
7378 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7379 tmp = (busiest->avg_load * busiest->group_capacity) /
7380 local->group_capacity;
7382 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7383 local->group_capacity;
7385 capa_move += local->group_capacity *
7386 min(local->load_per_task, local->avg_load + tmp);
7387 capa_move /= SCHED_CAPACITY_SCALE;
7389 /* Move if we gain throughput */
7390 if (capa_move > capa_now)
7391 env->imbalance = busiest->load_per_task;
7395 * calculate_imbalance - Calculate the amount of imbalance present within the
7396 * groups of a given sched_domain during load balance.
7397 * @env: load balance environment
7398 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7400 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7402 unsigned long max_pull, load_above_capacity = ~0UL;
7403 struct sg_lb_stats *local, *busiest;
7405 local = &sds->local_stat;
7406 busiest = &sds->busiest_stat;
7408 if (busiest->group_type == group_imbalanced) {
7410 * In the group_imb case we cannot rely on group-wide averages
7411 * to ensure cpu-load equilibrium, look at wider averages. XXX
7413 busiest->load_per_task =
7414 min(busiest->load_per_task, sds->avg_load);
7418 * In the presence of smp nice balancing, certain scenarios can have
7419 * max load less than avg load(as we skip the groups at or below
7420 * its cpu_capacity, while calculating max_load..)
7422 if (busiest->avg_load <= sds->avg_load ||
7423 local->avg_load >= sds->avg_load) {
7424 /* Misfitting tasks should be migrated in any case */
7425 if (busiest->group_type == group_misfit_task) {
7426 env->imbalance = busiest->group_misfit_task;
7431 * Busiest group is overloaded, local is not, use the spare
7432 * cycles to maximize throughput
7434 if (busiest->group_type == group_overloaded &&
7435 local->group_type <= group_misfit_task) {
7436 env->imbalance = busiest->load_per_task;
7441 return fix_small_imbalance(env, sds);
7445 * If there aren't any idle cpus, avoid creating some.
7447 if (busiest->group_type == group_overloaded &&
7448 local->group_type == group_overloaded) {
7449 load_above_capacity = busiest->sum_nr_running *
7451 if (load_above_capacity > busiest->group_capacity)
7452 load_above_capacity -= busiest->group_capacity;
7454 load_above_capacity = ~0UL;
7458 * We're trying to get all the cpus to the average_load, so we don't
7459 * want to push ourselves above the average load, nor do we wish to
7460 * reduce the max loaded cpu below the average load. At the same time,
7461 * we also don't want to reduce the group load below the group capacity
7462 * (so that we can implement power-savings policies etc). Thus we look
7463 * for the minimum possible imbalance.
7465 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7467 /* How much load to actually move to equalise the imbalance */
7468 env->imbalance = min(
7469 max_pull * busiest->group_capacity,
7470 (sds->avg_load - local->avg_load) * local->group_capacity
7471 ) / SCHED_CAPACITY_SCALE;
7473 /* Boost imbalance to allow misfit task to be balanced. */
7474 if (busiest->group_type == group_misfit_task)
7475 env->imbalance = max_t(long, env->imbalance,
7476 busiest->group_misfit_task);
7479 * if *imbalance is less than the average load per runnable task
7480 * there is no guarantee that any tasks will be moved so we'll have
7481 * a think about bumping its value to force at least one task to be
7484 if (env->imbalance < busiest->load_per_task)
7485 return fix_small_imbalance(env, sds);
7488 /******* find_busiest_group() helpers end here *********************/
7491 * find_busiest_group - Returns the busiest group within the sched_domain
7492 * if there is an imbalance. If there isn't an imbalance, and
7493 * the user has opted for power-savings, it returns a group whose
7494 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7495 * such a group exists.
7497 * Also calculates the amount of weighted load which should be moved
7498 * to restore balance.
7500 * @env: The load balancing environment.
7502 * Return: - The busiest group if imbalance exists.
7503 * - If no imbalance and user has opted for power-savings balance,
7504 * return the least loaded group whose CPUs can be
7505 * put to idle by rebalancing its tasks onto our group.
7507 static struct sched_group *find_busiest_group(struct lb_env *env)
7509 struct sg_lb_stats *local, *busiest;
7510 struct sd_lb_stats sds;
7512 init_sd_lb_stats(&sds);
7515 * Compute the various statistics relavent for load balancing at
7518 update_sd_lb_stats(env, &sds);
7520 if (energy_aware() && !env->dst_rq->rd->overutilized)
7523 local = &sds.local_stat;
7524 busiest = &sds.busiest_stat;
7526 /* ASYM feature bypasses nice load balance check */
7527 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7528 check_asym_packing(env, &sds))
7531 /* There is no busy sibling group to pull tasks from */
7532 if (!sds.busiest || busiest->sum_nr_running == 0)
7535 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7536 / sds.total_capacity;
7539 * If the busiest group is imbalanced the below checks don't
7540 * work because they assume all things are equal, which typically
7541 * isn't true due to cpus_allowed constraints and the like.
7543 if (busiest->group_type == group_imbalanced)
7546 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7547 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7548 busiest->group_no_capacity)
7551 /* Misfitting tasks should be dealt with regardless of the avg load */
7552 if (busiest->group_type == group_misfit_task) {
7557 * If the local group is busier than the selected busiest group
7558 * don't try and pull any tasks.
7560 if (local->avg_load >= busiest->avg_load)
7564 * Don't pull any tasks if this group is already above the domain
7567 if (local->avg_load >= sds.avg_load)
7570 if (env->idle == CPU_IDLE) {
7572 * This cpu is idle. If the busiest group is not overloaded
7573 * and there is no imbalance between this and busiest group
7574 * wrt idle cpus, it is balanced. The imbalance becomes
7575 * significant if the diff is greater than 1 otherwise we
7576 * might end up to just move the imbalance on another group
7578 if ((busiest->group_type != group_overloaded) &&
7579 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7580 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7584 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7585 * imbalance_pct to be conservative.
7587 if (100 * busiest->avg_load <=
7588 env->sd->imbalance_pct * local->avg_load)
7593 env->busiest_group_type = busiest->group_type;
7594 /* Looks like there is an imbalance. Compute it */
7595 calculate_imbalance(env, &sds);
7604 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7606 static struct rq *find_busiest_queue(struct lb_env *env,
7607 struct sched_group *group)
7609 struct rq *busiest = NULL, *rq;
7610 unsigned long busiest_load = 0, busiest_capacity = 1;
7613 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7614 unsigned long capacity, wl;
7618 rt = fbq_classify_rq(rq);
7621 * We classify groups/runqueues into three groups:
7622 * - regular: there are !numa tasks
7623 * - remote: there are numa tasks that run on the 'wrong' node
7624 * - all: there is no distinction
7626 * In order to avoid migrating ideally placed numa tasks,
7627 * ignore those when there's better options.
7629 * If we ignore the actual busiest queue to migrate another
7630 * task, the next balance pass can still reduce the busiest
7631 * queue by moving tasks around inside the node.
7633 * If we cannot move enough load due to this classification
7634 * the next pass will adjust the group classification and
7635 * allow migration of more tasks.
7637 * Both cases only affect the total convergence complexity.
7639 if (rt > env->fbq_type)
7642 capacity = capacity_of(i);
7644 wl = weighted_cpuload(i);
7647 * When comparing with imbalance, use weighted_cpuload()
7648 * which is not scaled with the cpu capacity.
7651 if (rq->nr_running == 1 && wl > env->imbalance &&
7652 !check_cpu_capacity(rq, env->sd) &&
7653 env->busiest_group_type != group_misfit_task)
7657 * For the load comparisons with the other cpu's, consider
7658 * the weighted_cpuload() scaled with the cpu capacity, so
7659 * that the load can be moved away from the cpu that is
7660 * potentially running at a lower capacity.
7662 * Thus we're looking for max(wl_i / capacity_i), crosswise
7663 * multiplication to rid ourselves of the division works out
7664 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7665 * our previous maximum.
7667 if (wl * busiest_capacity > busiest_load * capacity) {
7669 busiest_capacity = capacity;
7678 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7679 * so long as it is large enough.
7681 #define MAX_PINNED_INTERVAL 512
7683 /* Working cpumask for load_balance and load_balance_newidle. */
7684 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7686 static int need_active_balance(struct lb_env *env)
7688 struct sched_domain *sd = env->sd;
7690 if (env->idle == CPU_NEWLY_IDLE) {
7693 * ASYM_PACKING needs to force migrate tasks from busy but
7694 * higher numbered CPUs in order to pack all tasks in the
7695 * lowest numbered CPUs.
7697 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7702 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7703 * It's worth migrating the task if the src_cpu's capacity is reduced
7704 * because of other sched_class or IRQs if more capacity stays
7705 * available on dst_cpu.
7707 if ((env->idle != CPU_NOT_IDLE) &&
7708 (env->src_rq->cfs.h_nr_running == 1)) {
7709 if ((check_cpu_capacity(env->src_rq, sd)) &&
7710 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7714 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7715 env->src_rq->cfs.h_nr_running == 1 &&
7716 cpu_overutilized(env->src_cpu) &&
7717 !cpu_overutilized(env->dst_cpu)) {
7721 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7724 static int active_load_balance_cpu_stop(void *data);
7726 static int should_we_balance(struct lb_env *env)
7728 struct sched_group *sg = env->sd->groups;
7729 struct cpumask *sg_cpus, *sg_mask;
7730 int cpu, balance_cpu = -1;
7733 * In the newly idle case, we will allow all the cpu's
7734 * to do the newly idle load balance.
7736 if (env->idle == CPU_NEWLY_IDLE)
7739 sg_cpus = sched_group_cpus(sg);
7740 sg_mask = sched_group_mask(sg);
7741 /* Try to find first idle cpu */
7742 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7743 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7750 if (balance_cpu == -1)
7751 balance_cpu = group_balance_cpu(sg);
7754 * First idle cpu or the first cpu(busiest) in this sched group
7755 * is eligible for doing load balancing at this and above domains.
7757 return balance_cpu == env->dst_cpu;
7761 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7762 * tasks if there is an imbalance.
7764 static int load_balance(int this_cpu, struct rq *this_rq,
7765 struct sched_domain *sd, enum cpu_idle_type idle,
7766 int *continue_balancing)
7768 int ld_moved, cur_ld_moved, active_balance = 0;
7769 struct sched_domain *sd_parent = sd->parent;
7770 struct sched_group *group;
7772 unsigned long flags;
7773 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7775 struct lb_env env = {
7777 .dst_cpu = this_cpu,
7779 .dst_grpmask = sched_group_cpus(sd->groups),
7781 .loop_break = sched_nr_migrate_break,
7784 .tasks = LIST_HEAD_INIT(env.tasks),
7788 * For NEWLY_IDLE load_balancing, we don't need to consider
7789 * other cpus in our group
7791 if (idle == CPU_NEWLY_IDLE)
7792 env.dst_grpmask = NULL;
7794 cpumask_copy(cpus, cpu_active_mask);
7796 schedstat_inc(sd, lb_count[idle]);
7799 if (!should_we_balance(&env)) {
7800 *continue_balancing = 0;
7804 group = find_busiest_group(&env);
7806 schedstat_inc(sd, lb_nobusyg[idle]);
7810 busiest = find_busiest_queue(&env, group);
7812 schedstat_inc(sd, lb_nobusyq[idle]);
7816 BUG_ON(busiest == env.dst_rq);
7818 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7820 env.src_cpu = busiest->cpu;
7821 env.src_rq = busiest;
7824 if (busiest->nr_running > 1) {
7826 * Attempt to move tasks. If find_busiest_group has found
7827 * an imbalance but busiest->nr_running <= 1, the group is
7828 * still unbalanced. ld_moved simply stays zero, so it is
7829 * correctly treated as an imbalance.
7831 env.flags |= LBF_ALL_PINNED;
7832 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7835 raw_spin_lock_irqsave(&busiest->lock, flags);
7838 * cur_ld_moved - load moved in current iteration
7839 * ld_moved - cumulative load moved across iterations
7841 cur_ld_moved = detach_tasks(&env);
7843 * We want to potentially lower env.src_cpu's OPP.
7846 update_capacity_of(env.src_cpu);
7849 * We've detached some tasks from busiest_rq. Every
7850 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7851 * unlock busiest->lock, and we are able to be sure
7852 * that nobody can manipulate the tasks in parallel.
7853 * See task_rq_lock() family for the details.
7856 raw_spin_unlock(&busiest->lock);
7860 ld_moved += cur_ld_moved;
7863 local_irq_restore(flags);
7865 if (env.flags & LBF_NEED_BREAK) {
7866 env.flags &= ~LBF_NEED_BREAK;
7871 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7872 * us and move them to an alternate dst_cpu in our sched_group
7873 * where they can run. The upper limit on how many times we
7874 * iterate on same src_cpu is dependent on number of cpus in our
7877 * This changes load balance semantics a bit on who can move
7878 * load to a given_cpu. In addition to the given_cpu itself
7879 * (or a ilb_cpu acting on its behalf where given_cpu is
7880 * nohz-idle), we now have balance_cpu in a position to move
7881 * load to given_cpu. In rare situations, this may cause
7882 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7883 * _independently_ and at _same_ time to move some load to
7884 * given_cpu) causing exceess load to be moved to given_cpu.
7885 * This however should not happen so much in practice and
7886 * moreover subsequent load balance cycles should correct the
7887 * excess load moved.
7889 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7891 /* Prevent to re-select dst_cpu via env's cpus */
7892 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7894 env.dst_rq = cpu_rq(env.new_dst_cpu);
7895 env.dst_cpu = env.new_dst_cpu;
7896 env.flags &= ~LBF_DST_PINNED;
7898 env.loop_break = sched_nr_migrate_break;
7901 * Go back to "more_balance" rather than "redo" since we
7902 * need to continue with same src_cpu.
7908 * We failed to reach balance because of affinity.
7911 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7913 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7914 *group_imbalance = 1;
7917 /* All tasks on this runqueue were pinned by CPU affinity */
7918 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7919 cpumask_clear_cpu(cpu_of(busiest), cpus);
7920 if (!cpumask_empty(cpus)) {
7922 env.loop_break = sched_nr_migrate_break;
7925 goto out_all_pinned;
7930 schedstat_inc(sd, lb_failed[idle]);
7932 * Increment the failure counter only on periodic balance.
7933 * We do not want newidle balance, which can be very
7934 * frequent, pollute the failure counter causing
7935 * excessive cache_hot migrations and active balances.
7937 if (idle != CPU_NEWLY_IDLE)
7938 if (env.src_grp_nr_running > 1)
7939 sd->nr_balance_failed++;
7941 if (need_active_balance(&env)) {
7942 raw_spin_lock_irqsave(&busiest->lock, flags);
7944 /* don't kick the active_load_balance_cpu_stop,
7945 * if the curr task on busiest cpu can't be
7948 if (!cpumask_test_cpu(this_cpu,
7949 tsk_cpus_allowed(busiest->curr))) {
7950 raw_spin_unlock_irqrestore(&busiest->lock,
7952 env.flags |= LBF_ALL_PINNED;
7953 goto out_one_pinned;
7957 * ->active_balance synchronizes accesses to
7958 * ->active_balance_work. Once set, it's cleared
7959 * only after active load balance is finished.
7961 if (!busiest->active_balance) {
7962 busiest->active_balance = 1;
7963 busiest->push_cpu = this_cpu;
7966 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7968 if (active_balance) {
7969 stop_one_cpu_nowait(cpu_of(busiest),
7970 active_load_balance_cpu_stop, busiest,
7971 &busiest->active_balance_work);
7975 * We've kicked active balancing, reset the failure
7978 sd->nr_balance_failed = sd->cache_nice_tries+1;
7981 sd->nr_balance_failed = 0;
7983 if (likely(!active_balance)) {
7984 /* We were unbalanced, so reset the balancing interval */
7985 sd->balance_interval = sd->min_interval;
7988 * If we've begun active balancing, start to back off. This
7989 * case may not be covered by the all_pinned logic if there
7990 * is only 1 task on the busy runqueue (because we don't call
7993 if (sd->balance_interval < sd->max_interval)
7994 sd->balance_interval *= 2;
8001 * We reach balance although we may have faced some affinity
8002 * constraints. Clear the imbalance flag if it was set.
8005 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8007 if (*group_imbalance)
8008 *group_imbalance = 0;
8013 * We reach balance because all tasks are pinned at this level so
8014 * we can't migrate them. Let the imbalance flag set so parent level
8015 * can try to migrate them.
8017 schedstat_inc(sd, lb_balanced[idle]);
8019 sd->nr_balance_failed = 0;
8022 /* tune up the balancing interval */
8023 if (((env.flags & LBF_ALL_PINNED) &&
8024 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8025 (sd->balance_interval < sd->max_interval))
8026 sd->balance_interval *= 2;
8033 static inline unsigned long
8034 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8036 unsigned long interval = sd->balance_interval;
8039 interval *= sd->busy_factor;
8041 /* scale ms to jiffies */
8042 interval = msecs_to_jiffies(interval);
8043 interval = clamp(interval, 1UL, max_load_balance_interval);
8049 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8051 unsigned long interval, next;
8053 interval = get_sd_balance_interval(sd, cpu_busy);
8054 next = sd->last_balance + interval;
8056 if (time_after(*next_balance, next))
8057 *next_balance = next;
8061 * idle_balance is called by schedule() if this_cpu is about to become
8062 * idle. Attempts to pull tasks from other CPUs.
8064 static int idle_balance(struct rq *this_rq)
8066 unsigned long next_balance = jiffies + HZ;
8067 int this_cpu = this_rq->cpu;
8068 struct sched_domain *sd;
8069 int pulled_task = 0;
8072 idle_enter_fair(this_rq);
8075 * We must set idle_stamp _before_ calling idle_balance(), such that we
8076 * measure the duration of idle_balance() as idle time.
8078 this_rq->idle_stamp = rq_clock(this_rq);
8080 if (!energy_aware() &&
8081 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8082 !this_rq->rd->overload)) {
8084 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8086 update_next_balance(sd, 0, &next_balance);
8092 raw_spin_unlock(&this_rq->lock);
8094 update_blocked_averages(this_cpu);
8096 for_each_domain(this_cpu, sd) {
8097 int continue_balancing = 1;
8098 u64 t0, domain_cost;
8100 if (!(sd->flags & SD_LOAD_BALANCE))
8103 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8104 update_next_balance(sd, 0, &next_balance);
8108 if (sd->flags & SD_BALANCE_NEWIDLE) {
8109 t0 = sched_clock_cpu(this_cpu);
8111 pulled_task = load_balance(this_cpu, this_rq,
8113 &continue_balancing);
8115 domain_cost = sched_clock_cpu(this_cpu) - t0;
8116 if (domain_cost > sd->max_newidle_lb_cost)
8117 sd->max_newidle_lb_cost = domain_cost;
8119 curr_cost += domain_cost;
8122 update_next_balance(sd, 0, &next_balance);
8125 * Stop searching for tasks to pull if there are
8126 * now runnable tasks on this rq.
8128 if (pulled_task || this_rq->nr_running > 0)
8133 raw_spin_lock(&this_rq->lock);
8135 if (curr_cost > this_rq->max_idle_balance_cost)
8136 this_rq->max_idle_balance_cost = curr_cost;
8139 * While browsing the domains, we released the rq lock, a task could
8140 * have been enqueued in the meantime. Since we're not going idle,
8141 * pretend we pulled a task.
8143 if (this_rq->cfs.h_nr_running && !pulled_task)
8147 /* Move the next balance forward */
8148 if (time_after(this_rq->next_balance, next_balance))
8149 this_rq->next_balance = next_balance;
8151 /* Is there a task of a high priority class? */
8152 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8156 idle_exit_fair(this_rq);
8157 this_rq->idle_stamp = 0;
8164 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8165 * running tasks off the busiest CPU onto idle CPUs. It requires at
8166 * least 1 task to be running on each physical CPU where possible, and
8167 * avoids physical / logical imbalances.
8169 static int active_load_balance_cpu_stop(void *data)
8171 struct rq *busiest_rq = data;
8172 int busiest_cpu = cpu_of(busiest_rq);
8173 int target_cpu = busiest_rq->push_cpu;
8174 struct rq *target_rq = cpu_rq(target_cpu);
8175 struct sched_domain *sd;
8176 struct task_struct *p = NULL;
8178 raw_spin_lock_irq(&busiest_rq->lock);
8180 /* make sure the requested cpu hasn't gone down in the meantime */
8181 if (unlikely(busiest_cpu != smp_processor_id() ||
8182 !busiest_rq->active_balance))
8185 /* Is there any task to move? */
8186 if (busiest_rq->nr_running <= 1)
8190 * This condition is "impossible", if it occurs
8191 * we need to fix it. Originally reported by
8192 * Bjorn Helgaas on a 128-cpu setup.
8194 BUG_ON(busiest_rq == target_rq);
8196 /* Search for an sd spanning us and the target CPU. */
8198 for_each_domain(target_cpu, sd) {
8199 if ((sd->flags & SD_LOAD_BALANCE) &&
8200 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8205 struct lb_env env = {
8207 .dst_cpu = target_cpu,
8208 .dst_rq = target_rq,
8209 .src_cpu = busiest_rq->cpu,
8210 .src_rq = busiest_rq,
8214 schedstat_inc(sd, alb_count);
8216 p = detach_one_task(&env);
8218 schedstat_inc(sd, alb_pushed);
8220 * We want to potentially lower env.src_cpu's OPP.
8222 update_capacity_of(env.src_cpu);
8225 schedstat_inc(sd, alb_failed);
8229 busiest_rq->active_balance = 0;
8230 raw_spin_unlock(&busiest_rq->lock);
8233 attach_one_task(target_rq, p);
8240 static inline int on_null_domain(struct rq *rq)
8242 return unlikely(!rcu_dereference_sched(rq->sd));
8245 #ifdef CONFIG_NO_HZ_COMMON
8247 * idle load balancing details
8248 * - When one of the busy CPUs notice that there may be an idle rebalancing
8249 * needed, they will kick the idle load balancer, which then does idle
8250 * load balancing for all the idle CPUs.
8253 cpumask_var_t idle_cpus_mask;
8255 unsigned long next_balance; /* in jiffy units */
8256 } nohz ____cacheline_aligned;
8258 static inline int find_new_ilb(void)
8260 int ilb = cpumask_first(nohz.idle_cpus_mask);
8262 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8269 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8270 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8271 * CPU (if there is one).
8273 static void nohz_balancer_kick(void)
8277 nohz.next_balance++;
8279 ilb_cpu = find_new_ilb();
8281 if (ilb_cpu >= nr_cpu_ids)
8284 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8287 * Use smp_send_reschedule() instead of resched_cpu().
8288 * This way we generate a sched IPI on the target cpu which
8289 * is idle. And the softirq performing nohz idle load balance
8290 * will be run before returning from the IPI.
8292 smp_send_reschedule(ilb_cpu);
8296 static inline void nohz_balance_exit_idle(int cpu)
8298 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8300 * Completely isolated CPUs don't ever set, so we must test.
8302 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8303 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8304 atomic_dec(&nohz.nr_cpus);
8306 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8310 static inline void set_cpu_sd_state_busy(void)
8312 struct sched_domain *sd;
8313 int cpu = smp_processor_id();
8316 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8318 if (!sd || !sd->nohz_idle)
8322 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8327 void set_cpu_sd_state_idle(void)
8329 struct sched_domain *sd;
8330 int cpu = smp_processor_id();
8333 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8335 if (!sd || sd->nohz_idle)
8339 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8345 * This routine will record that the cpu is going idle with tick stopped.
8346 * This info will be used in performing idle load balancing in the future.
8348 void nohz_balance_enter_idle(int cpu)
8351 * If this cpu is going down, then nothing needs to be done.
8353 if (!cpu_active(cpu))
8356 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8360 * If we're a completely isolated CPU, we don't play.
8362 if (on_null_domain(cpu_rq(cpu)))
8365 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8366 atomic_inc(&nohz.nr_cpus);
8367 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8370 static int sched_ilb_notifier(struct notifier_block *nfb,
8371 unsigned long action, void *hcpu)
8373 switch (action & ~CPU_TASKS_FROZEN) {
8375 nohz_balance_exit_idle(smp_processor_id());
8383 static DEFINE_SPINLOCK(balancing);
8386 * Scale the max load_balance interval with the number of CPUs in the system.
8387 * This trades load-balance latency on larger machines for less cross talk.
8389 void update_max_interval(void)
8391 max_load_balance_interval = HZ*num_online_cpus()/10;
8395 * It checks each scheduling domain to see if it is due to be balanced,
8396 * and initiates a balancing operation if so.
8398 * Balancing parameters are set up in init_sched_domains.
8400 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8402 int continue_balancing = 1;
8404 unsigned long interval;
8405 struct sched_domain *sd;
8406 /* Earliest time when we have to do rebalance again */
8407 unsigned long next_balance = jiffies + 60*HZ;
8408 int update_next_balance = 0;
8409 int need_serialize, need_decay = 0;
8412 update_blocked_averages(cpu);
8415 for_each_domain(cpu, sd) {
8417 * Decay the newidle max times here because this is a regular
8418 * visit to all the domains. Decay ~1% per second.
8420 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8421 sd->max_newidle_lb_cost =
8422 (sd->max_newidle_lb_cost * 253) / 256;
8423 sd->next_decay_max_lb_cost = jiffies + HZ;
8426 max_cost += sd->max_newidle_lb_cost;
8428 if (!(sd->flags & SD_LOAD_BALANCE))
8432 * Stop the load balance at this level. There is another
8433 * CPU in our sched group which is doing load balancing more
8436 if (!continue_balancing) {
8442 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8444 need_serialize = sd->flags & SD_SERIALIZE;
8445 if (need_serialize) {
8446 if (!spin_trylock(&balancing))
8450 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8451 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8453 * The LBF_DST_PINNED logic could have changed
8454 * env->dst_cpu, so we can't know our idle
8455 * state even if we migrated tasks. Update it.
8457 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8459 sd->last_balance = jiffies;
8460 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8463 spin_unlock(&balancing);
8465 if (time_after(next_balance, sd->last_balance + interval)) {
8466 next_balance = sd->last_balance + interval;
8467 update_next_balance = 1;
8472 * Ensure the rq-wide value also decays but keep it at a
8473 * reasonable floor to avoid funnies with rq->avg_idle.
8475 rq->max_idle_balance_cost =
8476 max((u64)sysctl_sched_migration_cost, max_cost);
8481 * next_balance will be updated only when there is a need.
8482 * When the cpu is attached to null domain for ex, it will not be
8485 if (likely(update_next_balance)) {
8486 rq->next_balance = next_balance;
8488 #ifdef CONFIG_NO_HZ_COMMON
8490 * If this CPU has been elected to perform the nohz idle
8491 * balance. Other idle CPUs have already rebalanced with
8492 * nohz_idle_balance() and nohz.next_balance has been
8493 * updated accordingly. This CPU is now running the idle load
8494 * balance for itself and we need to update the
8495 * nohz.next_balance accordingly.
8497 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8498 nohz.next_balance = rq->next_balance;
8503 #ifdef CONFIG_NO_HZ_COMMON
8505 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8506 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8508 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8510 int this_cpu = this_rq->cpu;
8513 /* Earliest time when we have to do rebalance again */
8514 unsigned long next_balance = jiffies + 60*HZ;
8515 int update_next_balance = 0;
8517 if (idle != CPU_IDLE ||
8518 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8521 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8522 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8526 * If this cpu gets work to do, stop the load balancing
8527 * work being done for other cpus. Next load
8528 * balancing owner will pick it up.
8533 rq = cpu_rq(balance_cpu);
8536 * If time for next balance is due,
8539 if (time_after_eq(jiffies, rq->next_balance)) {
8540 raw_spin_lock_irq(&rq->lock);
8541 update_rq_clock(rq);
8542 update_idle_cpu_load(rq);
8543 raw_spin_unlock_irq(&rq->lock);
8544 rebalance_domains(rq, CPU_IDLE);
8547 if (time_after(next_balance, rq->next_balance)) {
8548 next_balance = rq->next_balance;
8549 update_next_balance = 1;
8554 * next_balance will be updated only when there is a need.
8555 * When the CPU is attached to null domain for ex, it will not be
8558 if (likely(update_next_balance))
8559 nohz.next_balance = next_balance;
8561 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8565 * Current heuristic for kicking the idle load balancer in the presence
8566 * of an idle cpu in the system.
8567 * - This rq has more than one task.
8568 * - This rq has at least one CFS task and the capacity of the CPU is
8569 * significantly reduced because of RT tasks or IRQs.
8570 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8571 * multiple busy cpu.
8572 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8573 * domain span are idle.
8575 static inline bool nohz_kick_needed(struct rq *rq)
8577 unsigned long now = jiffies;
8578 struct sched_domain *sd;
8579 struct sched_group_capacity *sgc;
8580 int nr_busy, cpu = rq->cpu;
8583 if (unlikely(rq->idle_balance))
8587 * We may be recently in ticked or tickless idle mode. At the first
8588 * busy tick after returning from idle, we will update the busy stats.
8590 set_cpu_sd_state_busy();
8591 nohz_balance_exit_idle(cpu);
8594 * None are in tickless mode and hence no need for NOHZ idle load
8597 if (likely(!atomic_read(&nohz.nr_cpus)))
8600 if (time_before(now, nohz.next_balance))
8603 if (rq->nr_running >= 2 &&
8604 (!energy_aware() || cpu_overutilized(cpu)))
8608 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8609 if (sd && !energy_aware()) {
8610 sgc = sd->groups->sgc;
8611 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8620 sd = rcu_dereference(rq->sd);
8622 if ((rq->cfs.h_nr_running >= 1) &&
8623 check_cpu_capacity(rq, sd)) {
8629 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8630 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8631 sched_domain_span(sd)) < cpu)) {
8641 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8645 * run_rebalance_domains is triggered when needed from the scheduler tick.
8646 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8648 static void run_rebalance_domains(struct softirq_action *h)
8650 struct rq *this_rq = this_rq();
8651 enum cpu_idle_type idle = this_rq->idle_balance ?
8652 CPU_IDLE : CPU_NOT_IDLE;
8655 * If this cpu has a pending nohz_balance_kick, then do the
8656 * balancing on behalf of the other idle cpus whose ticks are
8657 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8658 * give the idle cpus a chance to load balance. Else we may
8659 * load balance only within the local sched_domain hierarchy
8660 * and abort nohz_idle_balance altogether if we pull some load.
8662 nohz_idle_balance(this_rq, idle);
8663 rebalance_domains(this_rq, idle);
8667 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8669 void trigger_load_balance(struct rq *rq)
8671 /* Don't need to rebalance while attached to NULL domain */
8672 if (unlikely(on_null_domain(rq)))
8675 if (time_after_eq(jiffies, rq->next_balance))
8676 raise_softirq(SCHED_SOFTIRQ);
8677 #ifdef CONFIG_NO_HZ_COMMON
8678 if (nohz_kick_needed(rq))
8679 nohz_balancer_kick();
8683 static void rq_online_fair(struct rq *rq)
8687 update_runtime_enabled(rq);
8690 static void rq_offline_fair(struct rq *rq)
8694 /* Ensure any throttled groups are reachable by pick_next_task */
8695 unthrottle_offline_cfs_rqs(rq);
8698 #endif /* CONFIG_SMP */
8701 * scheduler tick hitting a task of our scheduling class:
8703 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8705 struct cfs_rq *cfs_rq;
8706 struct sched_entity *se = &curr->se;
8708 for_each_sched_entity(se) {
8709 cfs_rq = cfs_rq_of(se);
8710 entity_tick(cfs_rq, se, queued);
8713 if (static_branch_unlikely(&sched_numa_balancing))
8714 task_tick_numa(rq, curr);
8716 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8717 rq->rd->overutilized = true;
8719 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8723 * called on fork with the child task as argument from the parent's context
8724 * - child not yet on the tasklist
8725 * - preemption disabled
8727 static void task_fork_fair(struct task_struct *p)
8729 struct cfs_rq *cfs_rq;
8730 struct sched_entity *se = &p->se, *curr;
8731 int this_cpu = smp_processor_id();
8732 struct rq *rq = this_rq();
8733 unsigned long flags;
8735 raw_spin_lock_irqsave(&rq->lock, flags);
8737 update_rq_clock(rq);
8739 cfs_rq = task_cfs_rq(current);
8740 curr = cfs_rq->curr;
8743 * Not only the cpu but also the task_group of the parent might have
8744 * been changed after parent->se.parent,cfs_rq were copied to
8745 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8746 * of child point to valid ones.
8749 __set_task_cpu(p, this_cpu);
8752 update_curr(cfs_rq);
8755 se->vruntime = curr->vruntime;
8756 place_entity(cfs_rq, se, 1);
8758 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8760 * Upon rescheduling, sched_class::put_prev_task() will place
8761 * 'current' within the tree based on its new key value.
8763 swap(curr->vruntime, se->vruntime);
8767 se->vruntime -= cfs_rq->min_vruntime;
8769 raw_spin_unlock_irqrestore(&rq->lock, flags);
8773 * Priority of the task has changed. Check to see if we preempt
8777 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8779 if (!task_on_rq_queued(p))
8783 * Reschedule if we are currently running on this runqueue and
8784 * our priority decreased, or if we are not currently running on
8785 * this runqueue and our priority is higher than the current's
8787 if (rq->curr == p) {
8788 if (p->prio > oldprio)
8791 check_preempt_curr(rq, p, 0);
8794 static inline bool vruntime_normalized(struct task_struct *p)
8796 struct sched_entity *se = &p->se;
8799 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8800 * the dequeue_entity(.flags=0) will already have normalized the
8807 * When !on_rq, vruntime of the task has usually NOT been normalized.
8808 * But there are some cases where it has already been normalized:
8810 * - A forked child which is waiting for being woken up by
8811 * wake_up_new_task().
8812 * - A task which has been woken up by try_to_wake_up() and
8813 * waiting for actually being woken up by sched_ttwu_pending().
8815 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8821 static void detach_task_cfs_rq(struct task_struct *p)
8823 struct sched_entity *se = &p->se;
8824 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8826 if (!vruntime_normalized(p)) {
8828 * Fix up our vruntime so that the current sleep doesn't
8829 * cause 'unlimited' sleep bonus.
8831 place_entity(cfs_rq, se, 0);
8832 se->vruntime -= cfs_rq->min_vruntime;
8835 /* Catch up with the cfs_rq and remove our load when we leave */
8836 detach_entity_load_avg(cfs_rq, se);
8839 static void attach_task_cfs_rq(struct task_struct *p)
8841 struct sched_entity *se = &p->se;
8842 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8844 #ifdef CONFIG_FAIR_GROUP_SCHED
8846 * Since the real-depth could have been changed (only FAIR
8847 * class maintain depth value), reset depth properly.
8849 se->depth = se->parent ? se->parent->depth + 1 : 0;
8852 /* Synchronize task with its cfs_rq */
8853 attach_entity_load_avg(cfs_rq, se);
8855 if (!vruntime_normalized(p))
8856 se->vruntime += cfs_rq->min_vruntime;
8859 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8861 detach_task_cfs_rq(p);
8864 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8866 attach_task_cfs_rq(p);
8868 if (task_on_rq_queued(p)) {
8870 * We were most likely switched from sched_rt, so
8871 * kick off the schedule if running, otherwise just see
8872 * if we can still preempt the current task.
8877 check_preempt_curr(rq, p, 0);
8881 /* Account for a task changing its policy or group.
8883 * This routine is mostly called to set cfs_rq->curr field when a task
8884 * migrates between groups/classes.
8886 static void set_curr_task_fair(struct rq *rq)
8888 struct sched_entity *se = &rq->curr->se;
8890 for_each_sched_entity(se) {
8891 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8893 set_next_entity(cfs_rq, se);
8894 /* ensure bandwidth has been allocated on our new cfs_rq */
8895 account_cfs_rq_runtime(cfs_rq, 0);
8899 void init_cfs_rq(struct cfs_rq *cfs_rq)
8901 cfs_rq->tasks_timeline = RB_ROOT;
8902 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8903 #ifndef CONFIG_64BIT
8904 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8907 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8908 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8912 #ifdef CONFIG_FAIR_GROUP_SCHED
8913 static void task_move_group_fair(struct task_struct *p)
8915 detach_task_cfs_rq(p);
8916 set_task_rq(p, task_cpu(p));
8919 /* Tell se's cfs_rq has been changed -- migrated */
8920 p->se.avg.last_update_time = 0;
8922 attach_task_cfs_rq(p);
8925 void free_fair_sched_group(struct task_group *tg)
8929 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8931 for_each_possible_cpu(i) {
8933 kfree(tg->cfs_rq[i]);
8936 remove_entity_load_avg(tg->se[i]);
8945 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8947 struct cfs_rq *cfs_rq;
8948 struct sched_entity *se;
8951 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8954 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8958 tg->shares = NICE_0_LOAD;
8960 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8962 for_each_possible_cpu(i) {
8963 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8964 GFP_KERNEL, cpu_to_node(i));
8968 se = kzalloc_node(sizeof(struct sched_entity),
8969 GFP_KERNEL, cpu_to_node(i));
8973 init_cfs_rq(cfs_rq);
8974 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8975 init_entity_runnable_average(se);
8986 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8988 struct rq *rq = cpu_rq(cpu);
8989 unsigned long flags;
8992 * Only empty task groups can be destroyed; so we can speculatively
8993 * check on_list without danger of it being re-added.
8995 if (!tg->cfs_rq[cpu]->on_list)
8998 raw_spin_lock_irqsave(&rq->lock, flags);
8999 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9000 raw_spin_unlock_irqrestore(&rq->lock, flags);
9003 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9004 struct sched_entity *se, int cpu,
9005 struct sched_entity *parent)
9007 struct rq *rq = cpu_rq(cpu);
9011 init_cfs_rq_runtime(cfs_rq);
9013 tg->cfs_rq[cpu] = cfs_rq;
9016 /* se could be NULL for root_task_group */
9021 se->cfs_rq = &rq->cfs;
9024 se->cfs_rq = parent->my_q;
9025 se->depth = parent->depth + 1;
9029 /* guarantee group entities always have weight */
9030 update_load_set(&se->load, NICE_0_LOAD);
9031 se->parent = parent;
9034 static DEFINE_MUTEX(shares_mutex);
9036 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9039 unsigned long flags;
9042 * We can't change the weight of the root cgroup.
9047 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9049 mutex_lock(&shares_mutex);
9050 if (tg->shares == shares)
9053 tg->shares = shares;
9054 for_each_possible_cpu(i) {
9055 struct rq *rq = cpu_rq(i);
9056 struct sched_entity *se;
9059 /* Propagate contribution to hierarchy */
9060 raw_spin_lock_irqsave(&rq->lock, flags);
9062 /* Possible calls to update_curr() need rq clock */
9063 update_rq_clock(rq);
9064 for_each_sched_entity(se)
9065 update_cfs_shares(group_cfs_rq(se));
9066 raw_spin_unlock_irqrestore(&rq->lock, flags);
9070 mutex_unlock(&shares_mutex);
9073 #else /* CONFIG_FAIR_GROUP_SCHED */
9075 void free_fair_sched_group(struct task_group *tg) { }
9077 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9082 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9084 #endif /* CONFIG_FAIR_GROUP_SCHED */
9087 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9089 struct sched_entity *se = &task->se;
9090 unsigned int rr_interval = 0;
9093 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9096 if (rq->cfs.load.weight)
9097 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9103 * All the scheduling class methods:
9105 const struct sched_class fair_sched_class = {
9106 .next = &idle_sched_class,
9107 .enqueue_task = enqueue_task_fair,
9108 .dequeue_task = dequeue_task_fair,
9109 .yield_task = yield_task_fair,
9110 .yield_to_task = yield_to_task_fair,
9112 .check_preempt_curr = check_preempt_wakeup,
9114 .pick_next_task = pick_next_task_fair,
9115 .put_prev_task = put_prev_task_fair,
9118 .select_task_rq = select_task_rq_fair,
9119 .migrate_task_rq = migrate_task_rq_fair,
9121 .rq_online = rq_online_fair,
9122 .rq_offline = rq_offline_fair,
9124 .task_waking = task_waking_fair,
9125 .task_dead = task_dead_fair,
9126 .set_cpus_allowed = set_cpus_allowed_common,
9129 .set_curr_task = set_curr_task_fair,
9130 .task_tick = task_tick_fair,
9131 .task_fork = task_fork_fair,
9133 .prio_changed = prio_changed_fair,
9134 .switched_from = switched_from_fair,
9135 .switched_to = switched_to_fair,
9137 .get_rr_interval = get_rr_interval_fair,
9139 .update_curr = update_curr_fair,
9141 #ifdef CONFIG_FAIR_GROUP_SCHED
9142 .task_move_group = task_move_group_fair,
9146 #ifdef CONFIG_SCHED_DEBUG
9147 void print_cfs_stats(struct seq_file *m, int cpu)
9149 struct cfs_rq *cfs_rq;
9152 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9153 print_cfs_rq(m, cpu, cfs_rq);
9157 #ifdef CONFIG_NUMA_BALANCING
9158 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9161 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9163 for_each_online_node(node) {
9164 if (p->numa_faults) {
9165 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9166 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9168 if (p->numa_group) {
9169 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9170 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9172 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9175 #endif /* CONFIG_NUMA_BALANCING */
9176 #endif /* CONFIG_SCHED_DEBUG */
9178 __init void init_sched_fair_class(void)
9181 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9183 #ifdef CONFIG_NO_HZ_COMMON
9184 nohz.next_balance = jiffies;
9185 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9186 cpu_notifier(sched_ilb_notifier, 0);