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);
2751 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2753 if (!sched_feat(ATTACH_AGE_LOAD))
2757 * If we got migrated (either between CPUs or between cgroups) we'll
2758 * have aged the average right before clearing @last_update_time.
2760 if (se->avg.last_update_time) {
2761 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2762 &se->avg, 0, 0, NULL);
2765 * XXX: we could have just aged the entire load away if we've been
2766 * absent from the fair class for too long.
2771 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2772 cfs_rq->avg.load_avg += se->avg.load_avg;
2773 cfs_rq->avg.load_sum += se->avg.load_sum;
2774 cfs_rq->avg.util_avg += se->avg.util_avg;
2775 cfs_rq->avg.util_sum += se->avg.util_sum;
2778 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2780 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2781 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2782 cfs_rq->curr == se, NULL);
2784 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2785 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2786 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2787 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2790 /* Add the load generated by se into cfs_rq's load average */
2792 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2794 struct sched_avg *sa = &se->avg;
2795 u64 now = cfs_rq_clock_task(cfs_rq);
2796 int migrated, decayed;
2798 migrated = !sa->last_update_time;
2800 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2801 se->on_rq * scale_load_down(se->load.weight),
2802 cfs_rq->curr == se, NULL);
2805 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2807 cfs_rq->runnable_load_avg += sa->load_avg;
2808 cfs_rq->runnable_load_sum += sa->load_sum;
2811 attach_entity_load_avg(cfs_rq, se);
2813 if (decayed || migrated)
2814 update_tg_load_avg(cfs_rq, 0);
2817 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2819 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2821 update_load_avg(se, 1);
2823 cfs_rq->runnable_load_avg =
2824 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2825 cfs_rq->runnable_load_sum =
2826 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2829 #ifndef CONFIG_64BIT
2830 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2832 u64 last_update_time_copy;
2833 u64 last_update_time;
2836 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2838 last_update_time = cfs_rq->avg.last_update_time;
2839 } while (last_update_time != last_update_time_copy);
2841 return last_update_time;
2844 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2846 return cfs_rq->avg.last_update_time;
2851 * Task first catches up with cfs_rq, and then subtract
2852 * itself from the cfs_rq (task must be off the queue now).
2854 void remove_entity_load_avg(struct sched_entity *se)
2856 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2857 u64 last_update_time;
2860 * Newly created task or never used group entity should not be removed
2861 * from its (source) cfs_rq
2863 if (se->avg.last_update_time == 0)
2866 last_update_time = cfs_rq_last_update_time(cfs_rq);
2868 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2869 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2870 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2874 * Update the rq's load with the elapsed running time before entering
2875 * idle. if the last scheduled task is not a CFS task, idle_enter will
2876 * be the only way to update the runnable statistic.
2878 void idle_enter_fair(struct rq *this_rq)
2883 * Update the rq's load with the elapsed idle time before a task is
2884 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2885 * be the only way to update the runnable statistic.
2887 void idle_exit_fair(struct rq *this_rq)
2891 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2893 return cfs_rq->runnable_load_avg;
2896 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2898 return cfs_rq->avg.load_avg;
2901 static int idle_balance(struct rq *this_rq);
2903 #else /* CONFIG_SMP */
2905 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2907 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2909 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2910 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2913 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2915 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2917 static inline int idle_balance(struct rq *rq)
2922 #endif /* CONFIG_SMP */
2924 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2926 #ifdef CONFIG_SCHEDSTATS
2927 struct task_struct *tsk = NULL;
2929 if (entity_is_task(se))
2932 if (se->statistics.sleep_start) {
2933 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2938 if (unlikely(delta > se->statistics.sleep_max))
2939 se->statistics.sleep_max = delta;
2941 se->statistics.sleep_start = 0;
2942 se->statistics.sum_sleep_runtime += delta;
2945 account_scheduler_latency(tsk, delta >> 10, 1);
2946 trace_sched_stat_sleep(tsk, delta);
2949 if (se->statistics.block_start) {
2950 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2955 if (unlikely(delta > se->statistics.block_max))
2956 se->statistics.block_max = delta;
2958 se->statistics.block_start = 0;
2959 se->statistics.sum_sleep_runtime += delta;
2962 if (tsk->in_iowait) {
2963 se->statistics.iowait_sum += delta;
2964 se->statistics.iowait_count++;
2965 trace_sched_stat_iowait(tsk, delta);
2968 trace_sched_stat_blocked(tsk, delta);
2969 trace_sched_blocked_reason(tsk);
2972 * Blocking time is in units of nanosecs, so shift by
2973 * 20 to get a milliseconds-range estimation of the
2974 * amount of time that the task spent sleeping:
2976 if (unlikely(prof_on == SLEEP_PROFILING)) {
2977 profile_hits(SLEEP_PROFILING,
2978 (void *)get_wchan(tsk),
2981 account_scheduler_latency(tsk, delta >> 10, 0);
2987 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2989 #ifdef CONFIG_SCHED_DEBUG
2990 s64 d = se->vruntime - cfs_rq->min_vruntime;
2995 if (d > 3*sysctl_sched_latency)
2996 schedstat_inc(cfs_rq, nr_spread_over);
3001 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3003 u64 vruntime = cfs_rq->min_vruntime;
3006 * The 'current' period is already promised to the current tasks,
3007 * however the extra weight of the new task will slow them down a
3008 * little, place the new task so that it fits in the slot that
3009 * stays open at the end.
3011 if (initial && sched_feat(START_DEBIT))
3012 vruntime += sched_vslice(cfs_rq, se);
3014 /* sleeps up to a single latency don't count. */
3016 unsigned long thresh = sysctl_sched_latency;
3019 * Halve their sleep time's effect, to allow
3020 * for a gentler effect of sleepers:
3022 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3028 /* ensure we never gain time by being placed backwards. */
3029 se->vruntime = max_vruntime(se->vruntime, vruntime);
3032 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3035 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3038 * Update the normalized vruntime before updating min_vruntime
3039 * through calling update_curr().
3041 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3042 se->vruntime += cfs_rq->min_vruntime;
3045 * Update run-time statistics of the 'current'.
3047 update_curr(cfs_rq);
3048 enqueue_entity_load_avg(cfs_rq, se);
3049 account_entity_enqueue(cfs_rq, se);
3050 update_cfs_shares(cfs_rq);
3052 if (flags & ENQUEUE_WAKEUP) {
3053 place_entity(cfs_rq, se, 0);
3054 enqueue_sleeper(cfs_rq, se);
3057 update_stats_enqueue(cfs_rq, se);
3058 check_spread(cfs_rq, se);
3059 if (se != cfs_rq->curr)
3060 __enqueue_entity(cfs_rq, se);
3063 if (cfs_rq->nr_running == 1) {
3064 list_add_leaf_cfs_rq(cfs_rq);
3065 check_enqueue_throttle(cfs_rq);
3069 static void __clear_buddies_last(struct sched_entity *se)
3071 for_each_sched_entity(se) {
3072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3073 if (cfs_rq->last != se)
3076 cfs_rq->last = NULL;
3080 static void __clear_buddies_next(struct sched_entity *se)
3082 for_each_sched_entity(se) {
3083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3084 if (cfs_rq->next != se)
3087 cfs_rq->next = NULL;
3091 static void __clear_buddies_skip(struct sched_entity *se)
3093 for_each_sched_entity(se) {
3094 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3095 if (cfs_rq->skip != se)
3098 cfs_rq->skip = NULL;
3102 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3104 if (cfs_rq->last == se)
3105 __clear_buddies_last(se);
3107 if (cfs_rq->next == se)
3108 __clear_buddies_next(se);
3110 if (cfs_rq->skip == se)
3111 __clear_buddies_skip(se);
3114 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3117 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3120 * Update run-time statistics of the 'current'.
3122 update_curr(cfs_rq);
3123 dequeue_entity_load_avg(cfs_rq, se);
3125 update_stats_dequeue(cfs_rq, se);
3126 if (flags & DEQUEUE_SLEEP) {
3127 #ifdef CONFIG_SCHEDSTATS
3128 if (entity_is_task(se)) {
3129 struct task_struct *tsk = task_of(se);
3131 if (tsk->state & TASK_INTERRUPTIBLE)
3132 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3133 if (tsk->state & TASK_UNINTERRUPTIBLE)
3134 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3139 clear_buddies(cfs_rq, se);
3141 if (se != cfs_rq->curr)
3142 __dequeue_entity(cfs_rq, se);
3144 account_entity_dequeue(cfs_rq, se);
3147 * Normalize the entity after updating the min_vruntime because the
3148 * update can refer to the ->curr item and we need to reflect this
3149 * movement in our normalized position.
3151 if (!(flags & DEQUEUE_SLEEP))
3152 se->vruntime -= cfs_rq->min_vruntime;
3154 /* return excess runtime on last dequeue */
3155 return_cfs_rq_runtime(cfs_rq);
3157 update_min_vruntime(cfs_rq);
3158 update_cfs_shares(cfs_rq);
3162 * Preempt the current task with a newly woken task if needed:
3165 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3167 unsigned long ideal_runtime, delta_exec;
3168 struct sched_entity *se;
3171 ideal_runtime = sched_slice(cfs_rq, curr);
3172 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3173 if (delta_exec > ideal_runtime) {
3174 resched_curr(rq_of(cfs_rq));
3176 * The current task ran long enough, ensure it doesn't get
3177 * re-elected due to buddy favours.
3179 clear_buddies(cfs_rq, curr);
3184 * Ensure that a task that missed wakeup preemption by a
3185 * narrow margin doesn't have to wait for a full slice.
3186 * This also mitigates buddy induced latencies under load.
3188 if (delta_exec < sysctl_sched_min_granularity)
3191 se = __pick_first_entity(cfs_rq);
3192 delta = curr->vruntime - se->vruntime;
3197 if (delta > ideal_runtime)
3198 resched_curr(rq_of(cfs_rq));
3202 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3204 /* 'current' is not kept within the tree. */
3207 * Any task has to be enqueued before it get to execute on
3208 * a CPU. So account for the time it spent waiting on the
3211 update_stats_wait_end(cfs_rq, se);
3212 __dequeue_entity(cfs_rq, se);
3213 update_load_avg(se, 1);
3216 update_stats_curr_start(cfs_rq, se);
3218 #ifdef CONFIG_SCHEDSTATS
3220 * Track our maximum slice length, if the CPU's load is at
3221 * least twice that of our own weight (i.e. dont track it
3222 * when there are only lesser-weight tasks around):
3224 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3225 se->statistics.slice_max = max(se->statistics.slice_max,
3226 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3229 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3233 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3236 * Pick the next process, keeping these things in mind, in this order:
3237 * 1) keep things fair between processes/task groups
3238 * 2) pick the "next" process, since someone really wants that to run
3239 * 3) pick the "last" process, for cache locality
3240 * 4) do not run the "skip" process, if something else is available
3242 static struct sched_entity *
3243 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3245 struct sched_entity *left = __pick_first_entity(cfs_rq);
3246 struct sched_entity *se;
3249 * If curr is set we have to see if its left of the leftmost entity
3250 * still in the tree, provided there was anything in the tree at all.
3252 if (!left || (curr && entity_before(curr, left)))
3255 se = left; /* ideally we run the leftmost entity */
3258 * Avoid running the skip buddy, if running something else can
3259 * be done without getting too unfair.
3261 if (cfs_rq->skip == se) {
3262 struct sched_entity *second;
3265 second = __pick_first_entity(cfs_rq);
3267 second = __pick_next_entity(se);
3268 if (!second || (curr && entity_before(curr, second)))
3272 if (second && wakeup_preempt_entity(second, left) < 1)
3277 * Prefer last buddy, try to return the CPU to a preempted task.
3279 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3283 * Someone really wants this to run. If it's not unfair, run it.
3285 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3288 clear_buddies(cfs_rq, se);
3293 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3295 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3298 * If still on the runqueue then deactivate_task()
3299 * was not called and update_curr() has to be done:
3302 update_curr(cfs_rq);
3304 /* throttle cfs_rqs exceeding runtime */
3305 check_cfs_rq_runtime(cfs_rq);
3307 check_spread(cfs_rq, prev);
3309 update_stats_wait_start(cfs_rq, prev);
3310 /* Put 'current' back into the tree. */
3311 __enqueue_entity(cfs_rq, prev);
3312 /* in !on_rq case, update occurred at dequeue */
3313 update_load_avg(prev, 0);
3315 cfs_rq->curr = NULL;
3319 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3322 * Update run-time statistics of the 'current'.
3324 update_curr(cfs_rq);
3327 * Ensure that runnable average is periodically updated.
3329 update_load_avg(curr, 1);
3330 update_cfs_shares(cfs_rq);
3332 #ifdef CONFIG_SCHED_HRTICK
3334 * queued ticks are scheduled to match the slice, so don't bother
3335 * validating it and just reschedule.
3338 resched_curr(rq_of(cfs_rq));
3342 * don't let the period tick interfere with the hrtick preemption
3344 if (!sched_feat(DOUBLE_TICK) &&
3345 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3349 if (cfs_rq->nr_running > 1)
3350 check_preempt_tick(cfs_rq, curr);
3354 /**************************************************
3355 * CFS bandwidth control machinery
3358 #ifdef CONFIG_CFS_BANDWIDTH
3360 #ifdef HAVE_JUMP_LABEL
3361 static struct static_key __cfs_bandwidth_used;
3363 static inline bool cfs_bandwidth_used(void)
3365 return static_key_false(&__cfs_bandwidth_used);
3368 void cfs_bandwidth_usage_inc(void)
3370 static_key_slow_inc(&__cfs_bandwidth_used);
3373 void cfs_bandwidth_usage_dec(void)
3375 static_key_slow_dec(&__cfs_bandwidth_used);
3377 #else /* HAVE_JUMP_LABEL */
3378 static bool cfs_bandwidth_used(void)
3383 void cfs_bandwidth_usage_inc(void) {}
3384 void cfs_bandwidth_usage_dec(void) {}
3385 #endif /* HAVE_JUMP_LABEL */
3388 * default period for cfs group bandwidth.
3389 * default: 0.1s, units: nanoseconds
3391 static inline u64 default_cfs_period(void)
3393 return 100000000ULL;
3396 static inline u64 sched_cfs_bandwidth_slice(void)
3398 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3402 * Replenish runtime according to assigned quota and update expiration time.
3403 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3404 * additional synchronization around rq->lock.
3406 * requires cfs_b->lock
3408 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3412 if (cfs_b->quota == RUNTIME_INF)
3415 now = sched_clock_cpu(smp_processor_id());
3416 cfs_b->runtime = cfs_b->quota;
3417 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3420 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3422 return &tg->cfs_bandwidth;
3425 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3426 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3428 if (unlikely(cfs_rq->throttle_count))
3429 return cfs_rq->throttled_clock_task;
3431 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3434 /* returns 0 on failure to allocate runtime */
3435 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3437 struct task_group *tg = cfs_rq->tg;
3438 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3439 u64 amount = 0, min_amount, expires;
3441 /* note: this is a positive sum as runtime_remaining <= 0 */
3442 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3444 raw_spin_lock(&cfs_b->lock);
3445 if (cfs_b->quota == RUNTIME_INF)
3446 amount = min_amount;
3448 start_cfs_bandwidth(cfs_b);
3450 if (cfs_b->runtime > 0) {
3451 amount = min(cfs_b->runtime, min_amount);
3452 cfs_b->runtime -= amount;
3456 expires = cfs_b->runtime_expires;
3457 raw_spin_unlock(&cfs_b->lock);
3459 cfs_rq->runtime_remaining += amount;
3461 * we may have advanced our local expiration to account for allowed
3462 * spread between our sched_clock and the one on which runtime was
3465 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3466 cfs_rq->runtime_expires = expires;
3468 return cfs_rq->runtime_remaining > 0;
3472 * Note: This depends on the synchronization provided by sched_clock and the
3473 * fact that rq->clock snapshots this value.
3475 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3477 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3479 /* if the deadline is ahead of our clock, nothing to do */
3480 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3483 if (cfs_rq->runtime_remaining < 0)
3487 * If the local deadline has passed we have to consider the
3488 * possibility that our sched_clock is 'fast' and the global deadline
3489 * has not truly expired.
3491 * Fortunately we can check determine whether this the case by checking
3492 * whether the global deadline has advanced. It is valid to compare
3493 * cfs_b->runtime_expires without any locks since we only care about
3494 * exact equality, so a partial write will still work.
3497 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3498 /* extend local deadline, drift is bounded above by 2 ticks */
3499 cfs_rq->runtime_expires += TICK_NSEC;
3501 /* global deadline is ahead, expiration has passed */
3502 cfs_rq->runtime_remaining = 0;
3506 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3508 /* dock delta_exec before expiring quota (as it could span periods) */
3509 cfs_rq->runtime_remaining -= delta_exec;
3510 expire_cfs_rq_runtime(cfs_rq);
3512 if (likely(cfs_rq->runtime_remaining > 0))
3516 * if we're unable to extend our runtime we resched so that the active
3517 * hierarchy can be throttled
3519 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3520 resched_curr(rq_of(cfs_rq));
3523 static __always_inline
3524 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3526 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3529 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3532 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3534 return cfs_bandwidth_used() && cfs_rq->throttled;
3537 /* check whether cfs_rq, or any parent, is throttled */
3538 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3540 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3544 * Ensure that neither of the group entities corresponding to src_cpu or
3545 * dest_cpu are members of a throttled hierarchy when performing group
3546 * load-balance operations.
3548 static inline int throttled_lb_pair(struct task_group *tg,
3549 int src_cpu, int dest_cpu)
3551 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3553 src_cfs_rq = tg->cfs_rq[src_cpu];
3554 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3556 return throttled_hierarchy(src_cfs_rq) ||
3557 throttled_hierarchy(dest_cfs_rq);
3560 /* updated child weight may affect parent so we have to do this bottom up */
3561 static int tg_unthrottle_up(struct task_group *tg, void *data)
3563 struct rq *rq = data;
3564 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3566 cfs_rq->throttle_count--;
3568 if (!cfs_rq->throttle_count) {
3569 /* adjust cfs_rq_clock_task() */
3570 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3571 cfs_rq->throttled_clock_task;
3578 static int tg_throttle_down(struct task_group *tg, void *data)
3580 struct rq *rq = data;
3581 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3583 /* group is entering throttled state, stop time */
3584 if (!cfs_rq->throttle_count)
3585 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3586 cfs_rq->throttle_count++;
3591 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3593 struct rq *rq = rq_of(cfs_rq);
3594 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3595 struct sched_entity *se;
3596 long task_delta, dequeue = 1;
3599 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3601 /* freeze hierarchy runnable averages while throttled */
3603 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3606 task_delta = cfs_rq->h_nr_running;
3607 for_each_sched_entity(se) {
3608 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3609 /* throttled entity or throttle-on-deactivate */
3614 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3615 qcfs_rq->h_nr_running -= task_delta;
3617 if (qcfs_rq->load.weight)
3622 sub_nr_running(rq, task_delta);
3624 cfs_rq->throttled = 1;
3625 cfs_rq->throttled_clock = rq_clock(rq);
3626 raw_spin_lock(&cfs_b->lock);
3627 empty = list_empty(&cfs_b->throttled_cfs_rq);
3630 * Add to the _head_ of the list, so that an already-started
3631 * distribute_cfs_runtime will not see us
3633 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3636 * If we're the first throttled task, make sure the bandwidth
3640 start_cfs_bandwidth(cfs_b);
3642 raw_spin_unlock(&cfs_b->lock);
3645 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3647 struct rq *rq = rq_of(cfs_rq);
3648 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3649 struct sched_entity *se;
3653 se = cfs_rq->tg->se[cpu_of(rq)];
3655 cfs_rq->throttled = 0;
3657 update_rq_clock(rq);
3659 raw_spin_lock(&cfs_b->lock);
3660 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3661 list_del_rcu(&cfs_rq->throttled_list);
3662 raw_spin_unlock(&cfs_b->lock);
3664 /* update hierarchical throttle state */
3665 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3667 if (!cfs_rq->load.weight)
3670 task_delta = cfs_rq->h_nr_running;
3671 for_each_sched_entity(se) {
3675 cfs_rq = cfs_rq_of(se);
3677 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3678 cfs_rq->h_nr_running += task_delta;
3680 if (cfs_rq_throttled(cfs_rq))
3685 add_nr_running(rq, task_delta);
3687 /* determine whether we need to wake up potentially idle cpu */
3688 if (rq->curr == rq->idle && rq->cfs.nr_running)
3692 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3693 u64 remaining, u64 expires)
3695 struct cfs_rq *cfs_rq;
3697 u64 starting_runtime = remaining;
3700 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3702 struct rq *rq = rq_of(cfs_rq);
3704 raw_spin_lock(&rq->lock);
3705 if (!cfs_rq_throttled(cfs_rq))
3708 runtime = -cfs_rq->runtime_remaining + 1;
3709 if (runtime > remaining)
3710 runtime = remaining;
3711 remaining -= runtime;
3713 cfs_rq->runtime_remaining += runtime;
3714 cfs_rq->runtime_expires = expires;
3716 /* we check whether we're throttled above */
3717 if (cfs_rq->runtime_remaining > 0)
3718 unthrottle_cfs_rq(cfs_rq);
3721 raw_spin_unlock(&rq->lock);
3728 return starting_runtime - remaining;
3732 * Responsible for refilling a task_group's bandwidth and unthrottling its
3733 * cfs_rqs as appropriate. If there has been no activity within the last
3734 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3735 * used to track this state.
3737 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3739 u64 runtime, runtime_expires;
3742 /* no need to continue the timer with no bandwidth constraint */
3743 if (cfs_b->quota == RUNTIME_INF)
3744 goto out_deactivate;
3746 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3747 cfs_b->nr_periods += overrun;
3750 * idle depends on !throttled (for the case of a large deficit), and if
3751 * we're going inactive then everything else can be deferred
3753 if (cfs_b->idle && !throttled)
3754 goto out_deactivate;
3756 __refill_cfs_bandwidth_runtime(cfs_b);
3759 /* mark as potentially idle for the upcoming period */
3764 /* account preceding periods in which throttling occurred */
3765 cfs_b->nr_throttled += overrun;
3767 runtime_expires = cfs_b->runtime_expires;
3770 * This check is repeated as we are holding onto the new bandwidth while
3771 * we unthrottle. This can potentially race with an unthrottled group
3772 * trying to acquire new bandwidth from the global pool. This can result
3773 * in us over-using our runtime if it is all used during this loop, but
3774 * only by limited amounts in that extreme case.
3776 while (throttled && cfs_b->runtime > 0) {
3777 runtime = cfs_b->runtime;
3778 raw_spin_unlock(&cfs_b->lock);
3779 /* we can't nest cfs_b->lock while distributing bandwidth */
3780 runtime = distribute_cfs_runtime(cfs_b, runtime,
3782 raw_spin_lock(&cfs_b->lock);
3784 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3786 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3790 * While we are ensured activity in the period following an
3791 * unthrottle, this also covers the case in which the new bandwidth is
3792 * insufficient to cover the existing bandwidth deficit. (Forcing the
3793 * timer to remain active while there are any throttled entities.)
3803 /* a cfs_rq won't donate quota below this amount */
3804 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3805 /* minimum remaining period time to redistribute slack quota */
3806 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3807 /* how long we wait to gather additional slack before distributing */
3808 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3811 * Are we near the end of the current quota period?
3813 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3814 * hrtimer base being cleared by hrtimer_start. In the case of
3815 * migrate_hrtimers, base is never cleared, so we are fine.
3817 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3819 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3822 /* if the call-back is running a quota refresh is already occurring */
3823 if (hrtimer_callback_running(refresh_timer))
3826 /* is a quota refresh about to occur? */
3827 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3828 if (remaining < min_expire)
3834 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3836 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3838 /* if there's a quota refresh soon don't bother with slack */
3839 if (runtime_refresh_within(cfs_b, min_left))
3842 hrtimer_start(&cfs_b->slack_timer,
3843 ns_to_ktime(cfs_bandwidth_slack_period),
3847 /* we know any runtime found here is valid as update_curr() precedes return */
3848 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3850 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3851 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3853 if (slack_runtime <= 0)
3856 raw_spin_lock(&cfs_b->lock);
3857 if (cfs_b->quota != RUNTIME_INF &&
3858 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3859 cfs_b->runtime += slack_runtime;
3861 /* we are under rq->lock, defer unthrottling using a timer */
3862 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3863 !list_empty(&cfs_b->throttled_cfs_rq))
3864 start_cfs_slack_bandwidth(cfs_b);
3866 raw_spin_unlock(&cfs_b->lock);
3868 /* even if it's not valid for return we don't want to try again */
3869 cfs_rq->runtime_remaining -= slack_runtime;
3872 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3874 if (!cfs_bandwidth_used())
3877 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3880 __return_cfs_rq_runtime(cfs_rq);
3884 * This is done with a timer (instead of inline with bandwidth return) since
3885 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3887 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3889 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3892 /* confirm we're still not at a refresh boundary */
3893 raw_spin_lock(&cfs_b->lock);
3894 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3895 raw_spin_unlock(&cfs_b->lock);
3899 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3900 runtime = cfs_b->runtime;
3902 expires = cfs_b->runtime_expires;
3903 raw_spin_unlock(&cfs_b->lock);
3908 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3910 raw_spin_lock(&cfs_b->lock);
3911 if (expires == cfs_b->runtime_expires)
3912 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3913 raw_spin_unlock(&cfs_b->lock);
3917 * When a group wakes up we want to make sure that its quota is not already
3918 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3919 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3921 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3923 if (!cfs_bandwidth_used())
3926 /* an active group must be handled by the update_curr()->put() path */
3927 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3930 /* ensure the group is not already throttled */
3931 if (cfs_rq_throttled(cfs_rq))
3934 /* update runtime allocation */
3935 account_cfs_rq_runtime(cfs_rq, 0);
3936 if (cfs_rq->runtime_remaining <= 0)
3937 throttle_cfs_rq(cfs_rq);
3940 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3941 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3943 if (!cfs_bandwidth_used())
3946 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3950 * it's possible for a throttled entity to be forced into a running
3951 * state (e.g. set_curr_task), in this case we're finished.
3953 if (cfs_rq_throttled(cfs_rq))
3956 throttle_cfs_rq(cfs_rq);
3960 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3962 struct cfs_bandwidth *cfs_b =
3963 container_of(timer, struct cfs_bandwidth, slack_timer);
3965 do_sched_cfs_slack_timer(cfs_b);
3967 return HRTIMER_NORESTART;
3970 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3972 struct cfs_bandwidth *cfs_b =
3973 container_of(timer, struct cfs_bandwidth, period_timer);
3977 raw_spin_lock(&cfs_b->lock);
3979 overrun = hrtimer_forward_now(timer, cfs_b->period);
3983 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3986 cfs_b->period_active = 0;
3987 raw_spin_unlock(&cfs_b->lock);
3989 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3992 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3994 raw_spin_lock_init(&cfs_b->lock);
3996 cfs_b->quota = RUNTIME_INF;
3997 cfs_b->period = ns_to_ktime(default_cfs_period());
3999 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4000 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4001 cfs_b->period_timer.function = sched_cfs_period_timer;
4002 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4003 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4006 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4008 cfs_rq->runtime_enabled = 0;
4009 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4012 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4014 lockdep_assert_held(&cfs_b->lock);
4016 if (!cfs_b->period_active) {
4017 cfs_b->period_active = 1;
4018 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4019 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4023 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4025 /* init_cfs_bandwidth() was not called */
4026 if (!cfs_b->throttled_cfs_rq.next)
4029 hrtimer_cancel(&cfs_b->period_timer);
4030 hrtimer_cancel(&cfs_b->slack_timer);
4033 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4035 struct cfs_rq *cfs_rq;
4037 for_each_leaf_cfs_rq(rq, cfs_rq) {
4038 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4040 raw_spin_lock(&cfs_b->lock);
4041 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4042 raw_spin_unlock(&cfs_b->lock);
4046 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4048 struct cfs_rq *cfs_rq;
4050 for_each_leaf_cfs_rq(rq, cfs_rq) {
4051 if (!cfs_rq->runtime_enabled)
4055 * clock_task is not advancing so we just need to make sure
4056 * there's some valid quota amount
4058 cfs_rq->runtime_remaining = 1;
4060 * Offline rq is schedulable till cpu is completely disabled
4061 * in take_cpu_down(), so we prevent new cfs throttling here.
4063 cfs_rq->runtime_enabled = 0;
4065 if (cfs_rq_throttled(cfs_rq))
4066 unthrottle_cfs_rq(cfs_rq);
4070 #else /* CONFIG_CFS_BANDWIDTH */
4071 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4073 return rq_clock_task(rq_of(cfs_rq));
4076 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4077 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4078 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4079 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4081 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4086 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4091 static inline int throttled_lb_pair(struct task_group *tg,
4092 int src_cpu, int dest_cpu)
4097 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4099 #ifdef CONFIG_FAIR_GROUP_SCHED
4100 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4103 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4107 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4108 static inline void update_runtime_enabled(struct rq *rq) {}
4109 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4111 #endif /* CONFIG_CFS_BANDWIDTH */
4113 /**************************************************
4114 * CFS operations on tasks:
4117 #ifdef CONFIG_SCHED_HRTICK
4118 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4120 struct sched_entity *se = &p->se;
4121 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4123 WARN_ON(task_rq(p) != rq);
4125 if (cfs_rq->nr_running > 1) {
4126 u64 slice = sched_slice(cfs_rq, se);
4127 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4128 s64 delta = slice - ran;
4135 hrtick_start(rq, delta);
4140 * called from enqueue/dequeue and updates the hrtick when the
4141 * current task is from our class and nr_running is low enough
4144 static void hrtick_update(struct rq *rq)
4146 struct task_struct *curr = rq->curr;
4148 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4151 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4152 hrtick_start_fair(rq, curr);
4154 #else /* !CONFIG_SCHED_HRTICK */
4156 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4160 static inline void hrtick_update(struct rq *rq)
4165 static inline unsigned long boosted_cpu_util(int cpu);
4167 static void update_capacity_of(int cpu)
4169 unsigned long req_cap;
4174 /* Convert scale-invariant capacity to cpu. */
4175 req_cap = boosted_cpu_util(cpu);
4176 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4177 set_cfs_cpu_capacity(cpu, true, req_cap);
4180 static bool cpu_overutilized(int cpu);
4183 * The enqueue_task method is called before nr_running is
4184 * increased. Here we update the fair scheduling stats and
4185 * then put the task into the rbtree:
4188 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4190 struct cfs_rq *cfs_rq;
4191 struct sched_entity *se = &p->se;
4192 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4193 int task_wakeup = flags & ENQUEUE_WAKEUP;
4195 for_each_sched_entity(se) {
4198 cfs_rq = cfs_rq_of(se);
4199 enqueue_entity(cfs_rq, se, flags);
4202 * end evaluation on encountering a throttled cfs_rq
4204 * note: in the case of encountering a throttled cfs_rq we will
4205 * post the final h_nr_running increment below.
4207 if (cfs_rq_throttled(cfs_rq))
4209 cfs_rq->h_nr_running++;
4211 flags = ENQUEUE_WAKEUP;
4214 for_each_sched_entity(se) {
4215 cfs_rq = cfs_rq_of(se);
4216 cfs_rq->h_nr_running++;
4218 if (cfs_rq_throttled(cfs_rq))
4221 update_load_avg(se, 1);
4222 update_cfs_shares(cfs_rq);
4226 add_nr_running(rq, 1);
4227 if (!task_new && !rq->rd->overutilized &&
4228 cpu_overutilized(rq->cpu))
4229 rq->rd->overutilized = true;
4231 schedtune_enqueue_task(p, cpu_of(rq));
4234 * We want to potentially trigger a freq switch
4235 * request only for tasks that are waking up; this is
4236 * because we get here also during load balancing, but
4237 * in these cases it seems wise to trigger as single
4238 * request after load balancing is done.
4240 if (task_new || task_wakeup)
4241 update_capacity_of(cpu_of(rq));
4246 static void set_next_buddy(struct sched_entity *se);
4249 * The dequeue_task method is called before nr_running is
4250 * decreased. We remove the task from the rbtree and
4251 * update the fair scheduling stats:
4253 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4255 struct cfs_rq *cfs_rq;
4256 struct sched_entity *se = &p->se;
4257 int task_sleep = flags & DEQUEUE_SLEEP;
4259 for_each_sched_entity(se) {
4260 cfs_rq = cfs_rq_of(se);
4261 dequeue_entity(cfs_rq, se, flags);
4264 * end evaluation on encountering a throttled cfs_rq
4266 * note: in the case of encountering a throttled cfs_rq we will
4267 * post the final h_nr_running decrement below.
4269 if (cfs_rq_throttled(cfs_rq))
4271 cfs_rq->h_nr_running--;
4273 /* Don't dequeue parent if it has other entities besides us */
4274 if (cfs_rq->load.weight) {
4276 * Bias pick_next to pick a task from this cfs_rq, as
4277 * p is sleeping when it is within its sched_slice.
4279 if (task_sleep && parent_entity(se))
4280 set_next_buddy(parent_entity(se));
4282 /* avoid re-evaluating load for this entity */
4283 se = parent_entity(se);
4286 flags |= DEQUEUE_SLEEP;
4289 for_each_sched_entity(se) {
4290 cfs_rq = cfs_rq_of(se);
4291 cfs_rq->h_nr_running--;
4293 if (cfs_rq_throttled(cfs_rq))
4296 update_load_avg(se, 1);
4297 update_cfs_shares(cfs_rq);
4301 sub_nr_running(rq, 1);
4302 schedtune_dequeue_task(p, cpu_of(rq));
4305 * We want to potentially trigger a freq switch
4306 * request only for tasks that are going to sleep;
4307 * this is because we get here also during load
4308 * balancing, but in these cases it seems wise to
4309 * trigger as single request after load balancing is
4313 if (rq->cfs.nr_running)
4314 update_capacity_of(cpu_of(rq));
4315 else if (sched_freq())
4316 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4325 * per rq 'load' arrray crap; XXX kill this.
4329 * The exact cpuload at various idx values, calculated at every tick would be
4330 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4332 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4333 * on nth tick when cpu may be busy, then we have:
4334 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4335 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4337 * decay_load_missed() below does efficient calculation of
4338 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4339 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4341 * The calculation is approximated on a 128 point scale.
4342 * degrade_zero_ticks is the number of ticks after which load at any
4343 * particular idx is approximated to be zero.
4344 * degrade_factor is a precomputed table, a row for each load idx.
4345 * Each column corresponds to degradation factor for a power of two ticks,
4346 * based on 128 point scale.
4348 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4349 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4351 * With this power of 2 load factors, we can degrade the load n times
4352 * by looking at 1 bits in n and doing as many mult/shift instead of
4353 * n mult/shifts needed by the exact degradation.
4355 #define DEGRADE_SHIFT 7
4356 static const unsigned char
4357 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4358 static const unsigned char
4359 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4360 {0, 0, 0, 0, 0, 0, 0, 0},
4361 {64, 32, 8, 0, 0, 0, 0, 0},
4362 {96, 72, 40, 12, 1, 0, 0},
4363 {112, 98, 75, 43, 15, 1, 0},
4364 {120, 112, 98, 76, 45, 16, 2} };
4367 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4368 * would be when CPU is idle and so we just decay the old load without
4369 * adding any new load.
4371 static unsigned long
4372 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4376 if (!missed_updates)
4379 if (missed_updates >= degrade_zero_ticks[idx])
4383 return load >> missed_updates;
4385 while (missed_updates) {
4386 if (missed_updates % 2)
4387 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4389 missed_updates >>= 1;
4396 * Update rq->cpu_load[] statistics. This function is usually called every
4397 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4398 * every tick. We fix it up based on jiffies.
4400 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4401 unsigned long pending_updates)
4405 this_rq->nr_load_updates++;
4407 /* Update our load: */
4408 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4409 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4410 unsigned long old_load, new_load;
4412 /* scale is effectively 1 << i now, and >> i divides by scale */
4414 old_load = this_rq->cpu_load[i];
4415 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4416 new_load = this_load;
4418 * Round up the averaging division if load is increasing. This
4419 * prevents us from getting stuck on 9 if the load is 10, for
4422 if (new_load > old_load)
4423 new_load += scale - 1;
4425 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4428 sched_avg_update(this_rq);
4431 /* Used instead of source_load when we know the type == 0 */
4432 static unsigned long weighted_cpuload(const int cpu)
4434 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4437 #ifdef CONFIG_NO_HZ_COMMON
4439 * There is no sane way to deal with nohz on smp when using jiffies because the
4440 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4441 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4443 * Therefore we cannot use the delta approach from the regular tick since that
4444 * would seriously skew the load calculation. However we'll make do for those
4445 * updates happening while idle (nohz_idle_balance) or coming out of idle
4446 * (tick_nohz_idle_exit).
4448 * This means we might still be one tick off for nohz periods.
4452 * Called from nohz_idle_balance() to update the load ratings before doing the
4455 static void update_idle_cpu_load(struct rq *this_rq)
4457 unsigned long curr_jiffies = READ_ONCE(jiffies);
4458 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4459 unsigned long pending_updates;
4462 * bail if there's load or we're actually up-to-date.
4464 if (load || curr_jiffies == this_rq->last_load_update_tick)
4467 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4468 this_rq->last_load_update_tick = curr_jiffies;
4470 __update_cpu_load(this_rq, load, pending_updates);
4474 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4476 void update_cpu_load_nohz(void)
4478 struct rq *this_rq = this_rq();
4479 unsigned long curr_jiffies = READ_ONCE(jiffies);
4480 unsigned long pending_updates;
4482 if (curr_jiffies == this_rq->last_load_update_tick)
4485 raw_spin_lock(&this_rq->lock);
4486 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4487 if (pending_updates) {
4488 this_rq->last_load_update_tick = curr_jiffies;
4490 * We were idle, this means load 0, the current load might be
4491 * !0 due to remote wakeups and the sort.
4493 __update_cpu_load(this_rq, 0, pending_updates);
4495 raw_spin_unlock(&this_rq->lock);
4497 #endif /* CONFIG_NO_HZ */
4500 * Called from scheduler_tick()
4502 void update_cpu_load_active(struct rq *this_rq)
4504 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4506 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4508 this_rq->last_load_update_tick = jiffies;
4509 __update_cpu_load(this_rq, load, 1);
4513 * Return a low guess at the load of a migration-source cpu weighted
4514 * according to the scheduling class and "nice" value.
4516 * We want to under-estimate the load of migration sources, to
4517 * balance conservatively.
4519 static unsigned long source_load(int cpu, int type)
4521 struct rq *rq = cpu_rq(cpu);
4522 unsigned long total = weighted_cpuload(cpu);
4524 if (type == 0 || !sched_feat(LB_BIAS))
4527 return min(rq->cpu_load[type-1], total);
4531 * Return a high guess at the load of a migration-target cpu weighted
4532 * according to the scheduling class and "nice" value.
4534 static unsigned long target_load(int cpu, int type)
4536 struct rq *rq = cpu_rq(cpu);
4537 unsigned long total = weighted_cpuload(cpu);
4539 if (type == 0 || !sched_feat(LB_BIAS))
4542 return max(rq->cpu_load[type-1], total);
4546 static unsigned long cpu_avg_load_per_task(int cpu)
4548 struct rq *rq = cpu_rq(cpu);
4549 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4550 unsigned long load_avg = weighted_cpuload(cpu);
4553 return load_avg / nr_running;
4558 static void record_wakee(struct task_struct *p)
4561 * Rough decay (wiping) for cost saving, don't worry
4562 * about the boundary, really active task won't care
4565 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4566 current->wakee_flips >>= 1;
4567 current->wakee_flip_decay_ts = jiffies;
4570 if (current->last_wakee != p) {
4571 current->last_wakee = p;
4572 current->wakee_flips++;
4576 static void task_waking_fair(struct task_struct *p)
4578 struct sched_entity *se = &p->se;
4579 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4582 #ifndef CONFIG_64BIT
4583 u64 min_vruntime_copy;
4586 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4588 min_vruntime = cfs_rq->min_vruntime;
4589 } while (min_vruntime != min_vruntime_copy);
4591 min_vruntime = cfs_rq->min_vruntime;
4594 se->vruntime -= min_vruntime;
4598 #ifdef CONFIG_FAIR_GROUP_SCHED
4600 * effective_load() calculates the load change as seen from the root_task_group
4602 * Adding load to a group doesn't make a group heavier, but can cause movement
4603 * of group shares between cpus. Assuming the shares were perfectly aligned one
4604 * can calculate the shift in shares.
4606 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4607 * on this @cpu and results in a total addition (subtraction) of @wg to the
4608 * total group weight.
4610 * Given a runqueue weight distribution (rw_i) we can compute a shares
4611 * distribution (s_i) using:
4613 * s_i = rw_i / \Sum rw_j (1)
4615 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4616 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4617 * shares distribution (s_i):
4619 * rw_i = { 2, 4, 1, 0 }
4620 * s_i = { 2/7, 4/7, 1/7, 0 }
4622 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4623 * task used to run on and the CPU the waker is running on), we need to
4624 * compute the effect of waking a task on either CPU and, in case of a sync
4625 * wakeup, compute the effect of the current task going to sleep.
4627 * So for a change of @wl to the local @cpu with an overall group weight change
4628 * of @wl we can compute the new shares distribution (s'_i) using:
4630 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4632 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4633 * differences in waking a task to CPU 0. The additional task changes the
4634 * weight and shares distributions like:
4636 * rw'_i = { 3, 4, 1, 0 }
4637 * s'_i = { 3/8, 4/8, 1/8, 0 }
4639 * We can then compute the difference in effective weight by using:
4641 * dw_i = S * (s'_i - s_i) (3)
4643 * Where 'S' is the group weight as seen by its parent.
4645 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4646 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4647 * 4/7) times the weight of the group.
4649 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4651 struct sched_entity *se = tg->se[cpu];
4653 if (!tg->parent) /* the trivial, non-cgroup case */
4656 for_each_sched_entity(se) {
4657 struct cfs_rq *cfs_rq = se->my_q;
4658 long W, w = cfs_rq_load_avg(cfs_rq);
4663 * W = @wg + \Sum rw_j
4665 W = wg + atomic_long_read(&tg->load_avg);
4667 /* Ensure \Sum rw_j >= rw_i */
4668 W -= cfs_rq->tg_load_avg_contrib;
4677 * wl = S * s'_i; see (2)
4680 wl = (w * (long)tg->shares) / W;
4685 * Per the above, wl is the new se->load.weight value; since
4686 * those are clipped to [MIN_SHARES, ...) do so now. See
4687 * calc_cfs_shares().
4689 if (wl < MIN_SHARES)
4693 * wl = dw_i = S * (s'_i - s_i); see (3)
4695 wl -= se->avg.load_avg;
4698 * Recursively apply this logic to all parent groups to compute
4699 * the final effective load change on the root group. Since
4700 * only the @tg group gets extra weight, all parent groups can
4701 * only redistribute existing shares. @wl is the shift in shares
4702 * resulting from this level per the above.
4711 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4718 static inline bool energy_aware(void)
4720 return sched_feat(ENERGY_AWARE);
4724 struct sched_group *sg_top;
4725 struct sched_group *sg_cap;
4732 struct task_struct *task;
4747 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4748 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4749 * energy calculations. Using the scale-invariant util returned by
4750 * cpu_util() and approximating scale-invariant util by:
4752 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4754 * the normalized util can be found using the specific capacity.
4756 * capacity = capacity_orig * curr_freq/max_freq
4758 * norm_util = running_time/time ~ util/capacity
4760 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4762 int util = __cpu_util(cpu, delta);
4764 if (util >= capacity)
4765 return SCHED_CAPACITY_SCALE;
4767 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4770 static int calc_util_delta(struct energy_env *eenv, int cpu)
4772 if (cpu == eenv->src_cpu)
4773 return -eenv->util_delta;
4774 if (cpu == eenv->dst_cpu)
4775 return eenv->util_delta;
4780 unsigned long group_max_util(struct energy_env *eenv)
4783 unsigned long max_util = 0;
4785 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4786 delta = calc_util_delta(eenv, i);
4787 max_util = max(max_util, __cpu_util(i, delta));
4794 * group_norm_util() returns the approximated group util relative to it's
4795 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4796 * energy calculations. Since task executions may or may not overlap in time in
4797 * the group the true normalized util is between max(cpu_norm_util(i)) and
4798 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4799 * latter is used as the estimate as it leads to a more pessimistic energy
4800 * estimate (more busy).
4803 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4806 unsigned long util_sum = 0;
4807 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4809 for_each_cpu(i, sched_group_cpus(sg)) {
4810 delta = calc_util_delta(eenv, i);
4811 util_sum += __cpu_norm_util(i, capacity, delta);
4814 if (util_sum > SCHED_CAPACITY_SCALE)
4815 return SCHED_CAPACITY_SCALE;
4819 static int find_new_capacity(struct energy_env *eenv,
4820 const struct sched_group_energy const *sge)
4823 unsigned long util = group_max_util(eenv);
4825 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4826 if (sge->cap_states[idx].cap >= util)
4830 eenv->cap_idx = idx;
4835 static int group_idle_state(struct sched_group *sg)
4837 int i, state = INT_MAX;
4839 /* Find the shallowest idle state in the sched group. */
4840 for_each_cpu(i, sched_group_cpus(sg))
4841 state = min(state, idle_get_state_idx(cpu_rq(i)));
4843 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4850 * sched_group_energy(): Computes the absolute energy consumption of cpus
4851 * belonging to the sched_group including shared resources shared only by
4852 * members of the group. Iterates over all cpus in the hierarchy below the
4853 * sched_group starting from the bottom working it's way up before going to
4854 * the next cpu until all cpus are covered at all levels. The current
4855 * implementation is likely to gather the same util statistics multiple times.
4856 * This can probably be done in a faster but more complex way.
4857 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4859 static int sched_group_energy(struct energy_env *eenv)
4861 struct sched_domain *sd;
4862 int cpu, total_energy = 0;
4863 struct cpumask visit_cpus;
4864 struct sched_group *sg;
4866 WARN_ON(!eenv->sg_top->sge);
4868 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4870 while (!cpumask_empty(&visit_cpus)) {
4871 struct sched_group *sg_shared_cap = NULL;
4873 cpu = cpumask_first(&visit_cpus);
4876 * Is the group utilization affected by cpus outside this
4879 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4883 * We most probably raced with hotplug; returning a
4884 * wrong energy estimation is better than entering an
4890 sg_shared_cap = sd->parent->groups;
4892 for_each_domain(cpu, sd) {
4895 /* Has this sched_domain already been visited? */
4896 if (sd->child && group_first_cpu(sg) != cpu)
4900 unsigned long group_util;
4901 int sg_busy_energy, sg_idle_energy;
4902 int cap_idx, idle_idx;
4904 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4905 eenv->sg_cap = sg_shared_cap;
4909 cap_idx = find_new_capacity(eenv, sg->sge);
4911 if (sg->group_weight == 1) {
4912 /* Remove capacity of src CPU (before task move) */
4913 if (eenv->util_delta == 0 &&
4914 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4915 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4916 eenv->cap.delta -= eenv->cap.before;
4918 /* Add capacity of dst CPU (after task move) */
4919 if (eenv->util_delta != 0 &&
4920 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4921 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4922 eenv->cap.delta += eenv->cap.after;
4926 idle_idx = group_idle_state(sg);
4927 group_util = group_norm_util(eenv, sg);
4928 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4929 >> SCHED_CAPACITY_SHIFT;
4930 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4931 * sg->sge->idle_states[idle_idx].power)
4932 >> SCHED_CAPACITY_SHIFT;
4934 total_energy += sg_busy_energy + sg_idle_energy;
4937 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4939 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4942 } while (sg = sg->next, sg != sd->groups);
4948 eenv->energy = total_energy;
4952 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4954 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4957 #ifdef CONFIG_SCHED_TUNE
4958 static int energy_diff_evaluate(struct energy_env *eenv)
4963 /* Return energy diff when boost margin is 0 */
4964 #ifdef CONFIG_CGROUP_SCHEDTUNE
4965 boost = schedtune_task_boost(eenv->task);
4967 boost = get_sysctl_sched_cfs_boost();
4970 return eenv->nrg.diff;
4972 /* Compute normalized energy diff */
4973 nrg_delta = schedtune_normalize_energy(eenv->nrg.diff);
4974 eenv->nrg.delta = nrg_delta;
4976 eenv->payoff = schedtune_accept_deltas(
4982 * When SchedTune is enabled, the energy_diff() function will return
4983 * the computed energy payoff value. Since the energy_diff() return
4984 * value is expected to be negative by its callers, this evaluation
4985 * function return a negative value each time the evaluation return a
4986 * positive payoff, which is the condition for the acceptance of
4987 * a scheduling decision
4989 return -eenv->payoff;
4991 #else /* CONFIG_SCHED_TUNE */
4992 #define energy_diff_evaluate(eenv) eenv->nrg.diff
4996 * energy_diff(): Estimate the energy impact of changing the utilization
4997 * distribution. eenv specifies the change: utilisation amount, source, and
4998 * destination cpu. Source or destination cpu may be -1 in which case the
4999 * utilization is removed from or added to the system (e.g. task wake-up). If
5000 * both are specified, the utilization is migrated.
5002 static int energy_diff(struct energy_env *eenv)
5004 struct sched_domain *sd;
5005 struct sched_group *sg;
5006 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5008 struct energy_env eenv_before = {
5010 .src_cpu = eenv->src_cpu,
5011 .dst_cpu = eenv->dst_cpu,
5012 .nrg = { 0, 0, 0, 0},
5016 if (eenv->src_cpu == eenv->dst_cpu)
5019 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5020 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5023 return 0; /* Error */
5028 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5029 eenv_before.sg_top = eenv->sg_top = sg;
5031 if (sched_group_energy(&eenv_before))
5032 return 0; /* Invalid result abort */
5033 energy_before += eenv_before.energy;
5035 /* Keep track of SRC cpu (before) capacity */
5036 eenv->cap.before = eenv_before.cap.before;
5037 eenv->cap.delta = eenv_before.cap.delta;
5039 if (sched_group_energy(eenv))
5040 return 0; /* Invalid result abort */
5041 energy_after += eenv->energy;
5043 } while (sg = sg->next, sg != sd->groups);
5045 eenv->nrg.before = energy_before;
5046 eenv->nrg.after = energy_after;
5047 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5050 return energy_diff_evaluate(eenv);
5054 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5055 * A waker of many should wake a different task than the one last awakened
5056 * at a frequency roughly N times higher than one of its wakees. In order
5057 * to determine whether we should let the load spread vs consolodating to
5058 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5059 * partner, and a factor of lls_size higher frequency in the other. With
5060 * both conditions met, we can be relatively sure that the relationship is
5061 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5062 * being client/server, worker/dispatcher, interrupt source or whatever is
5063 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5065 static int wake_wide(struct task_struct *p)
5067 unsigned int master = current->wakee_flips;
5068 unsigned int slave = p->wakee_flips;
5069 int factor = this_cpu_read(sd_llc_size);
5072 swap(master, slave);
5073 if (slave < factor || master < slave * factor)
5078 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5080 s64 this_load, load;
5081 s64 this_eff_load, prev_eff_load;
5082 int idx, this_cpu, prev_cpu;
5083 struct task_group *tg;
5084 unsigned long weight;
5088 this_cpu = smp_processor_id();
5089 prev_cpu = task_cpu(p);
5090 load = source_load(prev_cpu, idx);
5091 this_load = target_load(this_cpu, idx);
5094 * If sync wakeup then subtract the (maximum possible)
5095 * effect of the currently running task from the load
5096 * of the current CPU:
5099 tg = task_group(current);
5100 weight = current->se.avg.load_avg;
5102 this_load += effective_load(tg, this_cpu, -weight, -weight);
5103 load += effective_load(tg, prev_cpu, 0, -weight);
5107 weight = p->se.avg.load_avg;
5110 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5111 * due to the sync cause above having dropped this_load to 0, we'll
5112 * always have an imbalance, but there's really nothing you can do
5113 * about that, so that's good too.
5115 * Otherwise check if either cpus are near enough in load to allow this
5116 * task to be woken on this_cpu.
5118 this_eff_load = 100;
5119 this_eff_load *= capacity_of(prev_cpu);
5121 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5122 prev_eff_load *= capacity_of(this_cpu);
5124 if (this_load > 0) {
5125 this_eff_load *= this_load +
5126 effective_load(tg, this_cpu, weight, weight);
5128 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5131 balanced = this_eff_load <= prev_eff_load;
5133 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5138 schedstat_inc(sd, ttwu_move_affine);
5139 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5144 static inline unsigned long task_util(struct task_struct *p)
5146 return p->se.avg.util_avg;
5149 unsigned int capacity_margin = 1280; /* ~20% margin */
5151 static inline unsigned long boosted_task_util(struct task_struct *task);
5153 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5155 unsigned long capacity = capacity_of(cpu);
5157 util += boosted_task_util(p);
5159 return (capacity * 1024) > (util * capacity_margin);
5162 static inline bool task_fits_max(struct task_struct *p, int cpu)
5164 unsigned long capacity = capacity_of(cpu);
5165 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5167 if (capacity == max_capacity)
5170 if (capacity * capacity_margin > max_capacity * 1024)
5173 return __task_fits(p, cpu, 0);
5176 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5178 return __task_fits(p, cpu, cpu_util(cpu));
5181 static bool cpu_overutilized(int cpu)
5183 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5186 #ifdef CONFIG_SCHED_TUNE
5188 static unsigned long
5189 schedtune_margin(unsigned long signal, unsigned long boost)
5191 unsigned long long margin = 0;
5194 * Signal proportional compensation (SPC)
5196 * The Boost (B) value is used to compute a Margin (M) which is
5197 * proportional to the complement of the original Signal (S):
5198 * M = B * (SCHED_LOAD_SCALE - S)
5199 * The obtained M could be used by the caller to "boost" S.
5201 margin = SCHED_LOAD_SCALE - signal;
5205 * Fast integer division by constant:
5206 * Constant : (C) = 100
5207 * Precision : 0.1% (P) = 0.1
5208 * Reference : C * 100 / P (R) = 100000
5211 * Shift bits : ceil(log(R,2)) (S) = 17
5212 * Mult const : round(2^S/C) (M) = 1311
5222 static inline unsigned int
5223 schedtune_cpu_margin(unsigned long util, int cpu)
5227 #ifdef CONFIG_CGROUP_SCHEDTUNE
5228 boost = schedtune_cpu_boost(cpu);
5230 boost = get_sysctl_sched_cfs_boost();
5235 return schedtune_margin(util, boost);
5238 static inline unsigned long
5239 schedtune_task_margin(struct task_struct *task)
5243 unsigned long margin;
5245 #ifdef CONFIG_CGROUP_SCHEDTUNE
5246 boost = schedtune_task_boost(task);
5248 boost = get_sysctl_sched_cfs_boost();
5253 util = task_util(task);
5254 margin = schedtune_margin(util, boost);
5259 #else /* CONFIG_SCHED_TUNE */
5261 static inline unsigned int
5262 schedtune_cpu_margin(unsigned long util, int cpu)
5267 static inline unsigned int
5268 schedtune_task_margin(struct task_struct *task)
5273 #endif /* CONFIG_SCHED_TUNE */
5275 static inline unsigned long
5276 boosted_cpu_util(int cpu)
5278 unsigned long util = cpu_util(cpu);
5279 unsigned long margin = schedtune_cpu_margin(util, cpu);
5281 return util + margin;
5284 static inline unsigned long
5285 boosted_task_util(struct task_struct *task)
5287 unsigned long util = task_util(task);
5288 unsigned long margin = schedtune_task_margin(task);
5290 return util + margin;
5294 * find_idlest_group finds and returns the least busy CPU group within the
5297 static struct sched_group *
5298 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5299 int this_cpu, int sd_flag)
5301 struct sched_group *idlest = NULL, *group = sd->groups;
5302 struct sched_group *fit_group = NULL, *spare_group = NULL;
5303 unsigned long min_load = ULONG_MAX, this_load = 0;
5304 unsigned long fit_capacity = ULONG_MAX;
5305 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5306 int load_idx = sd->forkexec_idx;
5307 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5309 if (sd_flag & SD_BALANCE_WAKE)
5310 load_idx = sd->wake_idx;
5313 unsigned long load, avg_load, spare_capacity;
5317 /* Skip over this group if it has no CPUs allowed */
5318 if (!cpumask_intersects(sched_group_cpus(group),
5319 tsk_cpus_allowed(p)))
5322 local_group = cpumask_test_cpu(this_cpu,
5323 sched_group_cpus(group));
5325 /* Tally up the load of all CPUs in the group */
5328 for_each_cpu(i, sched_group_cpus(group)) {
5329 /* Bias balancing toward cpus of our domain */
5331 load = source_load(i, load_idx);
5333 load = target_load(i, load_idx);
5338 * Look for most energy-efficient group that can fit
5339 * that can fit the task.
5341 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5342 fit_capacity = capacity_of(i);
5347 * Look for group which has most spare capacity on a
5350 spare_capacity = capacity_of(i) - cpu_util(i);
5351 if (spare_capacity > max_spare_capacity) {
5352 max_spare_capacity = spare_capacity;
5353 spare_group = group;
5357 /* Adjust by relative CPU capacity of the group */
5358 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5361 this_load = avg_load;
5362 } else if (avg_load < min_load) {
5363 min_load = avg_load;
5366 } while (group = group->next, group != sd->groups);
5374 if (!idlest || 100*this_load < imbalance*min_load)
5380 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5383 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5385 unsigned long load, min_load = ULONG_MAX;
5386 unsigned int min_exit_latency = UINT_MAX;
5387 u64 latest_idle_timestamp = 0;
5388 int least_loaded_cpu = this_cpu;
5389 int shallowest_idle_cpu = -1;
5392 /* Traverse only the allowed CPUs */
5393 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5394 if (task_fits_spare(p, i)) {
5395 struct rq *rq = cpu_rq(i);
5396 struct cpuidle_state *idle = idle_get_state(rq);
5397 if (idle && idle->exit_latency < min_exit_latency) {
5399 * We give priority to a CPU whose idle state
5400 * has the smallest exit latency irrespective
5401 * of any idle timestamp.
5403 min_exit_latency = idle->exit_latency;
5404 latest_idle_timestamp = rq->idle_stamp;
5405 shallowest_idle_cpu = i;
5406 } else if (idle_cpu(i) &&
5407 (!idle || idle->exit_latency == min_exit_latency) &&
5408 rq->idle_stamp > latest_idle_timestamp) {
5410 * If equal or no active idle state, then
5411 * the most recently idled CPU might have
5414 latest_idle_timestamp = rq->idle_stamp;
5415 shallowest_idle_cpu = i;
5416 } else if (shallowest_idle_cpu == -1) {
5418 * If we haven't found an idle CPU yet
5419 * pick a non-idle one that can fit the task as
5422 shallowest_idle_cpu = i;
5424 } else if (shallowest_idle_cpu == -1) {
5425 load = weighted_cpuload(i);
5426 if (load < min_load || (load == min_load && i == this_cpu)) {
5428 least_loaded_cpu = i;
5433 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5437 * Try and locate an idle CPU in the sched_domain.
5439 static int select_idle_sibling(struct task_struct *p, int target)
5441 struct sched_domain *sd;
5442 struct sched_group *sg;
5443 int i = task_cpu(p);
5445 if (idle_cpu(target))
5449 * If the prevous cpu is cache affine and idle, don't be stupid.
5451 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5455 * Otherwise, iterate the domains and find an elegible idle cpu.
5457 sd = rcu_dereference(per_cpu(sd_llc, target));
5458 for_each_lower_domain(sd) {
5461 if (!cpumask_intersects(sched_group_cpus(sg),
5462 tsk_cpus_allowed(p)))
5465 for_each_cpu(i, sched_group_cpus(sg)) {
5466 if (i == target || !idle_cpu(i))
5470 target = cpumask_first_and(sched_group_cpus(sg),
5471 tsk_cpus_allowed(p));
5475 } while (sg != sd->groups);
5481 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5483 struct sched_domain *sd;
5484 struct sched_group *sg, *sg_target;
5485 int target_max_cap = INT_MAX;
5486 int target_cpu = task_cpu(p);
5489 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5498 * Find group with sufficient capacity. We only get here if no cpu is
5499 * overutilized. We may end up overutilizing a cpu by adding the task,
5500 * but that should not be any worse than select_idle_sibling().
5501 * load_balance() should sort it out later as we get above the tipping
5505 /* Assuming all cpus are the same in group */
5506 int max_cap_cpu = group_first_cpu(sg);
5509 * Assume smaller max capacity means more energy-efficient.
5510 * Ideally we should query the energy model for the right
5511 * answer but it easily ends up in an exhaustive search.
5513 if (capacity_of(max_cap_cpu) < target_max_cap &&
5514 task_fits_max(p, max_cap_cpu)) {
5516 target_max_cap = capacity_of(max_cap_cpu);
5518 } while (sg = sg->next, sg != sd->groups);
5520 /* Find cpu with sufficient capacity */
5521 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5523 * p's blocked utilization is still accounted for on prev_cpu
5524 * so prev_cpu will receive a negative bias due to the double
5525 * accounting. However, the blocked utilization may be zero.
5527 int new_util = cpu_util(i) + boosted_task_util(p);
5529 if (new_util > capacity_orig_of(i))
5532 if (new_util < capacity_curr_of(i)) {
5534 if (cpu_rq(i)->nr_running)
5538 /* cpu has capacity at higher OPP, keep it as fallback */
5539 if (target_cpu == task_cpu(p))
5543 if (target_cpu != task_cpu(p)) {
5544 struct energy_env eenv = {
5545 .util_delta = task_util(p),
5546 .src_cpu = task_cpu(p),
5547 .dst_cpu = target_cpu,
5551 /* Not enough spare capacity on previous cpu */
5552 if (cpu_overutilized(task_cpu(p)))
5555 if (energy_diff(&eenv) >= 0)
5563 * select_task_rq_fair: Select target runqueue for the waking task in domains
5564 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5565 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5567 * Balances load by selecting the idlest cpu in the idlest group, or under
5568 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5570 * Returns the target cpu number.
5572 * preempt must be disabled.
5575 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5577 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5578 int cpu = smp_processor_id();
5579 int new_cpu = prev_cpu;
5580 int want_affine = 0;
5581 int sync = wake_flags & WF_SYNC;
5583 if (sd_flag & SD_BALANCE_WAKE)
5584 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5585 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5589 for_each_domain(cpu, tmp) {
5590 if (!(tmp->flags & SD_LOAD_BALANCE))
5594 * If both cpu and prev_cpu are part of this domain,
5595 * cpu is a valid SD_WAKE_AFFINE target.
5597 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5598 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5603 if (tmp->flags & sd_flag)
5605 else if (!want_affine)
5610 sd = NULL; /* Prefer wake_affine over balance flags */
5611 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5616 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5617 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5618 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5619 new_cpu = select_idle_sibling(p, new_cpu);
5622 struct sched_group *group;
5625 if (!(sd->flags & sd_flag)) {
5630 group = find_idlest_group(sd, p, cpu, sd_flag);
5636 new_cpu = find_idlest_cpu(group, p, cpu);
5637 if (new_cpu == -1 || new_cpu == cpu) {
5638 /* Now try balancing at a lower domain level of cpu */
5643 /* Now try balancing at a lower domain level of new_cpu */
5645 weight = sd->span_weight;
5647 for_each_domain(cpu, tmp) {
5648 if (weight <= tmp->span_weight)
5650 if (tmp->flags & sd_flag)
5653 /* while loop will break here if sd == NULL */
5661 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5662 * cfs_rq_of(p) references at time of call are still valid and identify the
5663 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5664 * other assumptions, including the state of rq->lock, should be made.
5666 static void migrate_task_rq_fair(struct task_struct *p)
5669 * We are supposed to update the task to "current" time, then its up to date
5670 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5671 * what current time is, so simply throw away the out-of-date time. This
5672 * will result in the wakee task is less decayed, but giving the wakee more
5673 * load sounds not bad.
5675 remove_entity_load_avg(&p->se);
5677 /* Tell new CPU we are migrated */
5678 p->se.avg.last_update_time = 0;
5680 /* We have migrated, no longer consider this task hot */
5681 p->se.exec_start = 0;
5684 static void task_dead_fair(struct task_struct *p)
5686 remove_entity_load_avg(&p->se);
5688 #endif /* CONFIG_SMP */
5690 static unsigned long
5691 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5693 unsigned long gran = sysctl_sched_wakeup_granularity;
5696 * Since its curr running now, convert the gran from real-time
5697 * to virtual-time in his units.
5699 * By using 'se' instead of 'curr' we penalize light tasks, so
5700 * they get preempted easier. That is, if 'se' < 'curr' then
5701 * the resulting gran will be larger, therefore penalizing the
5702 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5703 * be smaller, again penalizing the lighter task.
5705 * This is especially important for buddies when the leftmost
5706 * task is higher priority than the buddy.
5708 return calc_delta_fair(gran, se);
5712 * Should 'se' preempt 'curr'.
5726 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5728 s64 gran, vdiff = curr->vruntime - se->vruntime;
5733 gran = wakeup_gran(curr, se);
5740 static void set_last_buddy(struct sched_entity *se)
5742 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5745 for_each_sched_entity(se)
5746 cfs_rq_of(se)->last = se;
5749 static void set_next_buddy(struct sched_entity *se)
5751 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5754 for_each_sched_entity(se)
5755 cfs_rq_of(se)->next = se;
5758 static void set_skip_buddy(struct sched_entity *se)
5760 for_each_sched_entity(se)
5761 cfs_rq_of(se)->skip = se;
5765 * Preempt the current task with a newly woken task if needed:
5767 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5769 struct task_struct *curr = rq->curr;
5770 struct sched_entity *se = &curr->se, *pse = &p->se;
5771 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5772 int scale = cfs_rq->nr_running >= sched_nr_latency;
5773 int next_buddy_marked = 0;
5775 if (unlikely(se == pse))
5779 * This is possible from callers such as attach_tasks(), in which we
5780 * unconditionally check_prempt_curr() after an enqueue (which may have
5781 * lead to a throttle). This both saves work and prevents false
5782 * next-buddy nomination below.
5784 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5787 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5788 set_next_buddy(pse);
5789 next_buddy_marked = 1;
5793 * We can come here with TIF_NEED_RESCHED already set from new task
5796 * Note: this also catches the edge-case of curr being in a throttled
5797 * group (e.g. via set_curr_task), since update_curr() (in the
5798 * enqueue of curr) will have resulted in resched being set. This
5799 * prevents us from potentially nominating it as a false LAST_BUDDY
5802 if (test_tsk_need_resched(curr))
5805 /* Idle tasks are by definition preempted by non-idle tasks. */
5806 if (unlikely(curr->policy == SCHED_IDLE) &&
5807 likely(p->policy != SCHED_IDLE))
5811 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5812 * is driven by the tick):
5814 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5817 find_matching_se(&se, &pse);
5818 update_curr(cfs_rq_of(se));
5820 if (wakeup_preempt_entity(se, pse) == 1) {
5822 * Bias pick_next to pick the sched entity that is
5823 * triggering this preemption.
5825 if (!next_buddy_marked)
5826 set_next_buddy(pse);
5835 * Only set the backward buddy when the current task is still
5836 * on the rq. This can happen when a wakeup gets interleaved
5837 * with schedule on the ->pre_schedule() or idle_balance()
5838 * point, either of which can * drop the rq lock.
5840 * Also, during early boot the idle thread is in the fair class,
5841 * for obvious reasons its a bad idea to schedule back to it.
5843 if (unlikely(!se->on_rq || curr == rq->idle))
5846 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5850 static struct task_struct *
5851 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5853 struct cfs_rq *cfs_rq = &rq->cfs;
5854 struct sched_entity *se;
5855 struct task_struct *p;
5859 #ifdef CONFIG_FAIR_GROUP_SCHED
5860 if (!cfs_rq->nr_running)
5863 if (prev->sched_class != &fair_sched_class)
5867 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5868 * likely that a next task is from the same cgroup as the current.
5870 * Therefore attempt to avoid putting and setting the entire cgroup
5871 * hierarchy, only change the part that actually changes.
5875 struct sched_entity *curr = cfs_rq->curr;
5878 * Since we got here without doing put_prev_entity() we also
5879 * have to consider cfs_rq->curr. If it is still a runnable
5880 * entity, update_curr() will update its vruntime, otherwise
5881 * forget we've ever seen it.
5885 update_curr(cfs_rq);
5890 * This call to check_cfs_rq_runtime() will do the
5891 * throttle and dequeue its entity in the parent(s).
5892 * Therefore the 'simple' nr_running test will indeed
5895 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5899 se = pick_next_entity(cfs_rq, curr);
5900 cfs_rq = group_cfs_rq(se);
5906 * Since we haven't yet done put_prev_entity and if the selected task
5907 * is a different task than we started out with, try and touch the
5908 * least amount of cfs_rqs.
5911 struct sched_entity *pse = &prev->se;
5913 while (!(cfs_rq = is_same_group(se, pse))) {
5914 int se_depth = se->depth;
5915 int pse_depth = pse->depth;
5917 if (se_depth <= pse_depth) {
5918 put_prev_entity(cfs_rq_of(pse), pse);
5919 pse = parent_entity(pse);
5921 if (se_depth >= pse_depth) {
5922 set_next_entity(cfs_rq_of(se), se);
5923 se = parent_entity(se);
5927 put_prev_entity(cfs_rq, pse);
5928 set_next_entity(cfs_rq, se);
5931 if (hrtick_enabled(rq))
5932 hrtick_start_fair(rq, p);
5934 rq->misfit_task = !task_fits_max(p, rq->cpu);
5941 if (!cfs_rq->nr_running)
5944 put_prev_task(rq, prev);
5947 se = pick_next_entity(cfs_rq, NULL);
5948 set_next_entity(cfs_rq, se);
5949 cfs_rq = group_cfs_rq(se);
5954 if (hrtick_enabled(rq))
5955 hrtick_start_fair(rq, p);
5957 rq->misfit_task = !task_fits_max(p, rq->cpu);
5962 rq->misfit_task = 0;
5964 * This is OK, because current is on_cpu, which avoids it being picked
5965 * for load-balance and preemption/IRQs are still disabled avoiding
5966 * further scheduler activity on it and we're being very careful to
5967 * re-start the picking loop.
5969 lockdep_unpin_lock(&rq->lock);
5970 new_tasks = idle_balance(rq);
5971 lockdep_pin_lock(&rq->lock);
5973 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5974 * possible for any higher priority task to appear. In that case we
5975 * must re-start the pick_next_entity() loop.
5987 * Account for a descheduled task:
5989 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5991 struct sched_entity *se = &prev->se;
5992 struct cfs_rq *cfs_rq;
5994 for_each_sched_entity(se) {
5995 cfs_rq = cfs_rq_of(se);
5996 put_prev_entity(cfs_rq, se);
6001 * sched_yield() is very simple
6003 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6005 static void yield_task_fair(struct rq *rq)
6007 struct task_struct *curr = rq->curr;
6008 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6009 struct sched_entity *se = &curr->se;
6012 * Are we the only task in the tree?
6014 if (unlikely(rq->nr_running == 1))
6017 clear_buddies(cfs_rq, se);
6019 if (curr->policy != SCHED_BATCH) {
6020 update_rq_clock(rq);
6022 * Update run-time statistics of the 'current'.
6024 update_curr(cfs_rq);
6026 * Tell update_rq_clock() that we've just updated,
6027 * so we don't do microscopic update in schedule()
6028 * and double the fastpath cost.
6030 rq_clock_skip_update(rq, true);
6036 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6038 struct sched_entity *se = &p->se;
6040 /* throttled hierarchies are not runnable */
6041 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6044 /* Tell the scheduler that we'd really like pse to run next. */
6047 yield_task_fair(rq);
6053 /**************************************************
6054 * Fair scheduling class load-balancing methods.
6058 * The purpose of load-balancing is to achieve the same basic fairness the
6059 * per-cpu scheduler provides, namely provide a proportional amount of compute
6060 * time to each task. This is expressed in the following equation:
6062 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6064 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6065 * W_i,0 is defined as:
6067 * W_i,0 = \Sum_j w_i,j (2)
6069 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6070 * is derived from the nice value as per prio_to_weight[].
6072 * The weight average is an exponential decay average of the instantaneous
6075 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6077 * C_i is the compute capacity of cpu i, typically it is the
6078 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6079 * can also include other factors [XXX].
6081 * To achieve this balance we define a measure of imbalance which follows
6082 * directly from (1):
6084 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6086 * We them move tasks around to minimize the imbalance. In the continuous
6087 * function space it is obvious this converges, in the discrete case we get
6088 * a few fun cases generally called infeasible weight scenarios.
6091 * - infeasible weights;
6092 * - local vs global optima in the discrete case. ]
6097 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6098 * for all i,j solution, we create a tree of cpus that follows the hardware
6099 * topology where each level pairs two lower groups (or better). This results
6100 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6101 * tree to only the first of the previous level and we decrease the frequency
6102 * of load-balance at each level inv. proportional to the number of cpus in
6108 * \Sum { --- * --- * 2^i } = O(n) (5)
6110 * `- size of each group
6111 * | | `- number of cpus doing load-balance
6113 * `- sum over all levels
6115 * Coupled with a limit on how many tasks we can migrate every balance pass,
6116 * this makes (5) the runtime complexity of the balancer.
6118 * An important property here is that each CPU is still (indirectly) connected
6119 * to every other cpu in at most O(log n) steps:
6121 * The adjacency matrix of the resulting graph is given by:
6124 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6127 * And you'll find that:
6129 * A^(log_2 n)_i,j != 0 for all i,j (7)
6131 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6132 * The task movement gives a factor of O(m), giving a convergence complexity
6135 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6140 * In order to avoid CPUs going idle while there's still work to do, new idle
6141 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6142 * tree itself instead of relying on other CPUs to bring it work.
6144 * This adds some complexity to both (5) and (8) but it reduces the total idle
6152 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6155 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6160 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6162 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6164 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6167 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6168 * rewrite all of this once again.]
6171 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6173 enum fbq_type { regular, remote, all };
6182 #define LBF_ALL_PINNED 0x01
6183 #define LBF_NEED_BREAK 0x02
6184 #define LBF_DST_PINNED 0x04
6185 #define LBF_SOME_PINNED 0x08
6188 struct sched_domain *sd;
6196 struct cpumask *dst_grpmask;
6198 enum cpu_idle_type idle;
6200 unsigned int src_grp_nr_running;
6201 /* The set of CPUs under consideration for load-balancing */
6202 struct cpumask *cpus;
6207 unsigned int loop_break;
6208 unsigned int loop_max;
6210 enum fbq_type fbq_type;
6211 enum group_type busiest_group_type;
6212 struct list_head tasks;
6216 * Is this task likely cache-hot:
6218 static int task_hot(struct task_struct *p, struct lb_env *env)
6222 lockdep_assert_held(&env->src_rq->lock);
6224 if (p->sched_class != &fair_sched_class)
6227 if (unlikely(p->policy == SCHED_IDLE))
6231 * Buddy candidates are cache hot:
6233 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6234 (&p->se == cfs_rq_of(&p->se)->next ||
6235 &p->se == cfs_rq_of(&p->se)->last))
6238 if (sysctl_sched_migration_cost == -1)
6240 if (sysctl_sched_migration_cost == 0)
6243 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6245 return delta < (s64)sysctl_sched_migration_cost;
6248 #ifdef CONFIG_NUMA_BALANCING
6250 * Returns 1, if task migration degrades locality
6251 * Returns 0, if task migration improves locality i.e migration preferred.
6252 * Returns -1, if task migration is not affected by locality.
6254 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6256 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6257 unsigned long src_faults, dst_faults;
6258 int src_nid, dst_nid;
6260 if (!static_branch_likely(&sched_numa_balancing))
6263 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6266 src_nid = cpu_to_node(env->src_cpu);
6267 dst_nid = cpu_to_node(env->dst_cpu);
6269 if (src_nid == dst_nid)
6272 /* Migrating away from the preferred node is always bad. */
6273 if (src_nid == p->numa_preferred_nid) {
6274 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6280 /* Encourage migration to the preferred node. */
6281 if (dst_nid == p->numa_preferred_nid)
6285 src_faults = group_faults(p, src_nid);
6286 dst_faults = group_faults(p, dst_nid);
6288 src_faults = task_faults(p, src_nid);
6289 dst_faults = task_faults(p, dst_nid);
6292 return dst_faults < src_faults;
6296 static inline int migrate_degrades_locality(struct task_struct *p,
6304 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6307 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6311 lockdep_assert_held(&env->src_rq->lock);
6314 * We do not migrate tasks that are:
6315 * 1) throttled_lb_pair, or
6316 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6317 * 3) running (obviously), or
6318 * 4) are cache-hot on their current CPU.
6320 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6323 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6326 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6328 env->flags |= LBF_SOME_PINNED;
6331 * Remember if this task can be migrated to any other cpu in
6332 * our sched_group. We may want to revisit it if we couldn't
6333 * meet load balance goals by pulling other tasks on src_cpu.
6335 * Also avoid computing new_dst_cpu if we have already computed
6336 * one in current iteration.
6338 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6341 /* Prevent to re-select dst_cpu via env's cpus */
6342 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6343 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6344 env->flags |= LBF_DST_PINNED;
6345 env->new_dst_cpu = cpu;
6353 /* Record that we found atleast one task that could run on dst_cpu */
6354 env->flags &= ~LBF_ALL_PINNED;
6356 if (task_running(env->src_rq, p)) {
6357 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6362 * Aggressive migration if:
6363 * 1) destination numa is preferred
6364 * 2) task is cache cold, or
6365 * 3) too many balance attempts have failed.
6367 tsk_cache_hot = migrate_degrades_locality(p, env);
6368 if (tsk_cache_hot == -1)
6369 tsk_cache_hot = task_hot(p, env);
6371 if (tsk_cache_hot <= 0 ||
6372 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6373 if (tsk_cache_hot == 1) {
6374 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6375 schedstat_inc(p, se.statistics.nr_forced_migrations);
6380 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6385 * detach_task() -- detach the task for the migration specified in env
6387 static void detach_task(struct task_struct *p, struct lb_env *env)
6389 lockdep_assert_held(&env->src_rq->lock);
6391 deactivate_task(env->src_rq, p, 0);
6392 p->on_rq = TASK_ON_RQ_MIGRATING;
6393 set_task_cpu(p, env->dst_cpu);
6397 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6398 * part of active balancing operations within "domain".
6400 * Returns a task if successful and NULL otherwise.
6402 static struct task_struct *detach_one_task(struct lb_env *env)
6404 struct task_struct *p, *n;
6406 lockdep_assert_held(&env->src_rq->lock);
6408 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6409 if (!can_migrate_task(p, env))
6412 detach_task(p, env);
6415 * Right now, this is only the second place where
6416 * lb_gained[env->idle] is updated (other is detach_tasks)
6417 * so we can safely collect stats here rather than
6418 * inside detach_tasks().
6420 schedstat_inc(env->sd, lb_gained[env->idle]);
6426 static const unsigned int sched_nr_migrate_break = 32;
6429 * detach_tasks() -- tries to detach up to imbalance weighted load from
6430 * busiest_rq, as part of a balancing operation within domain "sd".
6432 * Returns number of detached tasks if successful and 0 otherwise.
6434 static int detach_tasks(struct lb_env *env)
6436 struct list_head *tasks = &env->src_rq->cfs_tasks;
6437 struct task_struct *p;
6441 lockdep_assert_held(&env->src_rq->lock);
6443 if (env->imbalance <= 0)
6446 while (!list_empty(tasks)) {
6448 * We don't want to steal all, otherwise we may be treated likewise,
6449 * which could at worst lead to a livelock crash.
6451 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6454 p = list_first_entry(tasks, struct task_struct, se.group_node);
6457 /* We've more or less seen every task there is, call it quits */
6458 if (env->loop > env->loop_max)
6461 /* take a breather every nr_migrate tasks */
6462 if (env->loop > env->loop_break) {
6463 env->loop_break += sched_nr_migrate_break;
6464 env->flags |= LBF_NEED_BREAK;
6468 if (!can_migrate_task(p, env))
6471 load = task_h_load(p);
6473 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6476 if ((load / 2) > env->imbalance)
6479 detach_task(p, env);
6480 list_add(&p->se.group_node, &env->tasks);
6483 env->imbalance -= load;
6485 #ifdef CONFIG_PREEMPT
6487 * NEWIDLE balancing is a source of latency, so preemptible
6488 * kernels will stop after the first task is detached to minimize
6489 * the critical section.
6491 if (env->idle == CPU_NEWLY_IDLE)
6496 * We only want to steal up to the prescribed amount of
6499 if (env->imbalance <= 0)
6504 list_move_tail(&p->se.group_node, tasks);
6508 * Right now, this is one of only two places we collect this stat
6509 * so we can safely collect detach_one_task() stats here rather
6510 * than inside detach_one_task().
6512 schedstat_add(env->sd, lb_gained[env->idle], detached);
6518 * attach_task() -- attach the task detached by detach_task() to its new rq.
6520 static void attach_task(struct rq *rq, struct task_struct *p)
6522 lockdep_assert_held(&rq->lock);
6524 BUG_ON(task_rq(p) != rq);
6525 p->on_rq = TASK_ON_RQ_QUEUED;
6526 activate_task(rq, p, 0);
6527 check_preempt_curr(rq, p, 0);
6531 * attach_one_task() -- attaches the task returned from detach_one_task() to
6534 static void attach_one_task(struct rq *rq, struct task_struct *p)
6536 raw_spin_lock(&rq->lock);
6539 * We want to potentially raise target_cpu's OPP.
6541 update_capacity_of(cpu_of(rq));
6542 raw_spin_unlock(&rq->lock);
6546 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6549 static void attach_tasks(struct lb_env *env)
6551 struct list_head *tasks = &env->tasks;
6552 struct task_struct *p;
6554 raw_spin_lock(&env->dst_rq->lock);
6556 while (!list_empty(tasks)) {
6557 p = list_first_entry(tasks, struct task_struct, se.group_node);
6558 list_del_init(&p->se.group_node);
6560 attach_task(env->dst_rq, p);
6564 * We want to potentially raise env.dst_cpu's OPP.
6566 update_capacity_of(env->dst_cpu);
6568 raw_spin_unlock(&env->dst_rq->lock);
6571 #ifdef CONFIG_FAIR_GROUP_SCHED
6572 static void update_blocked_averages(int cpu)
6574 struct rq *rq = cpu_rq(cpu);
6575 struct cfs_rq *cfs_rq;
6576 unsigned long flags;
6578 raw_spin_lock_irqsave(&rq->lock, flags);
6579 update_rq_clock(rq);
6582 * Iterates the task_group tree in a bottom up fashion, see
6583 * list_add_leaf_cfs_rq() for details.
6585 for_each_leaf_cfs_rq(rq, cfs_rq) {
6586 /* throttled entities do not contribute to load */
6587 if (throttled_hierarchy(cfs_rq))
6590 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6591 update_tg_load_avg(cfs_rq, 0);
6593 raw_spin_unlock_irqrestore(&rq->lock, flags);
6597 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6598 * This needs to be done in a top-down fashion because the load of a child
6599 * group is a fraction of its parents load.
6601 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6603 struct rq *rq = rq_of(cfs_rq);
6604 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6605 unsigned long now = jiffies;
6608 if (cfs_rq->last_h_load_update == now)
6611 cfs_rq->h_load_next = NULL;
6612 for_each_sched_entity(se) {
6613 cfs_rq = cfs_rq_of(se);
6614 cfs_rq->h_load_next = se;
6615 if (cfs_rq->last_h_load_update == now)
6620 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6621 cfs_rq->last_h_load_update = now;
6624 while ((se = cfs_rq->h_load_next) != NULL) {
6625 load = cfs_rq->h_load;
6626 load = div64_ul(load * se->avg.load_avg,
6627 cfs_rq_load_avg(cfs_rq) + 1);
6628 cfs_rq = group_cfs_rq(se);
6629 cfs_rq->h_load = load;
6630 cfs_rq->last_h_load_update = now;
6634 static unsigned long task_h_load(struct task_struct *p)
6636 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6638 update_cfs_rq_h_load(cfs_rq);
6639 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6640 cfs_rq_load_avg(cfs_rq) + 1);
6643 static inline void update_blocked_averages(int cpu)
6645 struct rq *rq = cpu_rq(cpu);
6646 struct cfs_rq *cfs_rq = &rq->cfs;
6647 unsigned long flags;
6649 raw_spin_lock_irqsave(&rq->lock, flags);
6650 update_rq_clock(rq);
6651 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6652 raw_spin_unlock_irqrestore(&rq->lock, flags);
6655 static unsigned long task_h_load(struct task_struct *p)
6657 return p->se.avg.load_avg;
6661 /********** Helpers for find_busiest_group ************************/
6664 * sg_lb_stats - stats of a sched_group required for load_balancing
6666 struct sg_lb_stats {
6667 unsigned long avg_load; /*Avg load across the CPUs of the group */
6668 unsigned long group_load; /* Total load over the CPUs of the group */
6669 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6670 unsigned long load_per_task;
6671 unsigned long group_capacity;
6672 unsigned long group_util; /* Total utilization of the group */
6673 unsigned int sum_nr_running; /* Nr tasks running in the group */
6674 unsigned int idle_cpus;
6675 unsigned int group_weight;
6676 enum group_type group_type;
6677 int group_no_capacity;
6678 int group_misfit_task; /* A cpu has a task too big for its capacity */
6679 #ifdef CONFIG_NUMA_BALANCING
6680 unsigned int nr_numa_running;
6681 unsigned int nr_preferred_running;
6686 * sd_lb_stats - Structure to store the statistics of a sched_domain
6687 * during load balancing.
6689 struct sd_lb_stats {
6690 struct sched_group *busiest; /* Busiest group in this sd */
6691 struct sched_group *local; /* Local group in this sd */
6692 unsigned long total_load; /* Total load of all groups in sd */
6693 unsigned long total_capacity; /* Total capacity of all groups in sd */
6694 unsigned long avg_load; /* Average load across all groups in sd */
6696 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6697 struct sg_lb_stats local_stat; /* Statistics of the local group */
6700 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6703 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6704 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6705 * We must however clear busiest_stat::avg_load because
6706 * update_sd_pick_busiest() reads this before assignment.
6708 *sds = (struct sd_lb_stats){
6712 .total_capacity = 0UL,
6715 .sum_nr_running = 0,
6716 .group_type = group_other,
6722 * get_sd_load_idx - Obtain the load index for a given sched domain.
6723 * @sd: The sched_domain whose load_idx is to be obtained.
6724 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6726 * Return: The load index.
6728 static inline int get_sd_load_idx(struct sched_domain *sd,
6729 enum cpu_idle_type idle)
6735 load_idx = sd->busy_idx;
6738 case CPU_NEWLY_IDLE:
6739 load_idx = sd->newidle_idx;
6742 load_idx = sd->idle_idx;
6749 static unsigned long scale_rt_capacity(int cpu)
6751 struct rq *rq = cpu_rq(cpu);
6752 u64 total, used, age_stamp, avg;
6756 * Since we're reading these variables without serialization make sure
6757 * we read them once before doing sanity checks on them.
6759 age_stamp = READ_ONCE(rq->age_stamp);
6760 avg = READ_ONCE(rq->rt_avg);
6761 delta = __rq_clock_broken(rq) - age_stamp;
6763 if (unlikely(delta < 0))
6766 total = sched_avg_period() + delta;
6768 used = div_u64(avg, total);
6771 * deadline bandwidth is defined at system level so we must
6772 * weight this bandwidth with the max capacity of the system.
6773 * As a reminder, avg_bw is 20bits width and
6774 * scale_cpu_capacity is 10 bits width
6776 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6778 if (likely(used < SCHED_CAPACITY_SCALE))
6779 return SCHED_CAPACITY_SCALE - used;
6784 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6786 raw_spin_lock_init(&mcc->lock);
6791 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6793 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6794 struct sched_group *sdg = sd->groups;
6795 struct max_cpu_capacity *mcc;
6796 unsigned long max_capacity;
6798 unsigned long flags;
6800 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6802 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6804 raw_spin_lock_irqsave(&mcc->lock, flags);
6805 max_capacity = mcc->val;
6806 max_cap_cpu = mcc->cpu;
6808 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6809 (max_capacity < capacity)) {
6810 mcc->val = capacity;
6812 #ifdef CONFIG_SCHED_DEBUG
6813 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6814 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6818 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6820 skip_unlock: __attribute__ ((unused));
6821 capacity *= scale_rt_capacity(cpu);
6822 capacity >>= SCHED_CAPACITY_SHIFT;
6827 cpu_rq(cpu)->cpu_capacity = capacity;
6828 sdg->sgc->capacity = capacity;
6829 sdg->sgc->max_capacity = capacity;
6832 void update_group_capacity(struct sched_domain *sd, int cpu)
6834 struct sched_domain *child = sd->child;
6835 struct sched_group *group, *sdg = sd->groups;
6836 unsigned long capacity, max_capacity;
6837 unsigned long interval;
6839 interval = msecs_to_jiffies(sd->balance_interval);
6840 interval = clamp(interval, 1UL, max_load_balance_interval);
6841 sdg->sgc->next_update = jiffies + interval;
6844 update_cpu_capacity(sd, cpu);
6851 if (child->flags & SD_OVERLAP) {
6853 * SD_OVERLAP domains cannot assume that child groups
6854 * span the current group.
6857 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6858 struct sched_group_capacity *sgc;
6859 struct rq *rq = cpu_rq(cpu);
6862 * build_sched_domains() -> init_sched_groups_capacity()
6863 * gets here before we've attached the domains to the
6866 * Use capacity_of(), which is set irrespective of domains
6867 * in update_cpu_capacity().
6869 * This avoids capacity from being 0 and
6870 * causing divide-by-zero issues on boot.
6872 if (unlikely(!rq->sd)) {
6873 capacity += capacity_of(cpu);
6875 sgc = rq->sd->groups->sgc;
6876 capacity += sgc->capacity;
6879 max_capacity = max(capacity, max_capacity);
6883 * !SD_OVERLAP domains can assume that child groups
6884 * span the current group.
6887 group = child->groups;
6889 struct sched_group_capacity *sgc = group->sgc;
6891 capacity += sgc->capacity;
6892 max_capacity = max(sgc->max_capacity, max_capacity);
6893 group = group->next;
6894 } while (group != child->groups);
6897 sdg->sgc->capacity = capacity;
6898 sdg->sgc->max_capacity = max_capacity;
6902 * Check whether the capacity of the rq has been noticeably reduced by side
6903 * activity. The imbalance_pct is used for the threshold.
6904 * Return true is the capacity is reduced
6907 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6909 return ((rq->cpu_capacity * sd->imbalance_pct) <
6910 (rq->cpu_capacity_orig * 100));
6914 * Group imbalance indicates (and tries to solve) the problem where balancing
6915 * groups is inadequate due to tsk_cpus_allowed() constraints.
6917 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6918 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6921 * { 0 1 2 3 } { 4 5 6 7 }
6924 * If we were to balance group-wise we'd place two tasks in the first group and
6925 * two tasks in the second group. Clearly this is undesired as it will overload
6926 * cpu 3 and leave one of the cpus in the second group unused.
6928 * The current solution to this issue is detecting the skew in the first group
6929 * by noticing the lower domain failed to reach balance and had difficulty
6930 * moving tasks due to affinity constraints.
6932 * When this is so detected; this group becomes a candidate for busiest; see
6933 * update_sd_pick_busiest(). And calculate_imbalance() and
6934 * find_busiest_group() avoid some of the usual balance conditions to allow it
6935 * to create an effective group imbalance.
6937 * This is a somewhat tricky proposition since the next run might not find the
6938 * group imbalance and decide the groups need to be balanced again. A most
6939 * subtle and fragile situation.
6942 static inline int sg_imbalanced(struct sched_group *group)
6944 return group->sgc->imbalance;
6948 * group_has_capacity returns true if the group has spare capacity that could
6949 * be used by some tasks.
6950 * We consider that a group has spare capacity if the * number of task is
6951 * smaller than the number of CPUs or if the utilization is lower than the
6952 * available capacity for CFS tasks.
6953 * For the latter, we use a threshold to stabilize the state, to take into
6954 * account the variance of the tasks' load and to return true if the available
6955 * capacity in meaningful for the load balancer.
6956 * As an example, an available capacity of 1% can appear but it doesn't make
6957 * any benefit for the load balance.
6960 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6962 if (sgs->sum_nr_running < sgs->group_weight)
6965 if ((sgs->group_capacity * 100) >
6966 (sgs->group_util * env->sd->imbalance_pct))
6973 * group_is_overloaded returns true if the group has more tasks than it can
6975 * group_is_overloaded is not equals to !group_has_capacity because a group
6976 * with the exact right number of tasks, has no more spare capacity but is not
6977 * overloaded so both group_has_capacity and group_is_overloaded return
6981 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6983 if (sgs->sum_nr_running <= sgs->group_weight)
6986 if ((sgs->group_capacity * 100) <
6987 (sgs->group_util * env->sd->imbalance_pct))
6995 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
6996 * per-cpu capacity than sched_group ref.
6999 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7001 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7002 ref->sgc->max_capacity;
7006 group_type group_classify(struct sched_group *group,
7007 struct sg_lb_stats *sgs)
7009 if (sgs->group_no_capacity)
7010 return group_overloaded;
7012 if (sg_imbalanced(group))
7013 return group_imbalanced;
7015 if (sgs->group_misfit_task)
7016 return group_misfit_task;
7022 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7023 * @env: The load balancing environment.
7024 * @group: sched_group whose statistics are to be updated.
7025 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7026 * @local_group: Does group contain this_cpu.
7027 * @sgs: variable to hold the statistics for this group.
7028 * @overload: Indicate more than one runnable task for any CPU.
7029 * @overutilized: Indicate overutilization for any CPU.
7031 static inline void update_sg_lb_stats(struct lb_env *env,
7032 struct sched_group *group, int load_idx,
7033 int local_group, struct sg_lb_stats *sgs,
7034 bool *overload, bool *overutilized)
7039 memset(sgs, 0, sizeof(*sgs));
7041 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7042 struct rq *rq = cpu_rq(i);
7044 /* Bias balancing toward cpus of our domain */
7046 load = target_load(i, load_idx);
7048 load = source_load(i, load_idx);
7050 sgs->group_load += load;
7051 sgs->group_util += cpu_util(i);
7052 sgs->sum_nr_running += rq->cfs.h_nr_running;
7054 if (rq->nr_running > 1)
7057 #ifdef CONFIG_NUMA_BALANCING
7058 sgs->nr_numa_running += rq->nr_numa_running;
7059 sgs->nr_preferred_running += rq->nr_preferred_running;
7061 sgs->sum_weighted_load += weighted_cpuload(i);
7065 if (cpu_overutilized(i)) {
7066 *overutilized = true;
7067 if (!sgs->group_misfit_task && rq->misfit_task)
7068 sgs->group_misfit_task = capacity_of(i);
7072 /* Adjust by relative CPU capacity of the group */
7073 sgs->group_capacity = group->sgc->capacity;
7074 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7076 if (sgs->sum_nr_running)
7077 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7079 sgs->group_weight = group->group_weight;
7081 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7082 sgs->group_type = group_classify(group, sgs);
7086 * update_sd_pick_busiest - return 1 on busiest group
7087 * @env: The load balancing environment.
7088 * @sds: sched_domain statistics
7089 * @sg: sched_group candidate to be checked for being the busiest
7090 * @sgs: sched_group statistics
7092 * Determine if @sg is a busier group than the previously selected
7095 * Return: %true if @sg is a busier group than the previously selected
7096 * busiest group. %false otherwise.
7098 static bool update_sd_pick_busiest(struct lb_env *env,
7099 struct sd_lb_stats *sds,
7100 struct sched_group *sg,
7101 struct sg_lb_stats *sgs)
7103 struct sg_lb_stats *busiest = &sds->busiest_stat;
7105 if (sgs->group_type > busiest->group_type)
7108 if (sgs->group_type < busiest->group_type)
7112 * Candidate sg doesn't face any serious load-balance problems
7113 * so don't pick it if the local sg is already filled up.
7115 if (sgs->group_type == group_other &&
7116 !group_has_capacity(env, &sds->local_stat))
7119 if (sgs->avg_load <= busiest->avg_load)
7123 * Candiate sg has no more than one task per cpu and has higher
7124 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7126 if (sgs->sum_nr_running <= sgs->group_weight &&
7127 group_smaller_cpu_capacity(sds->local, sg))
7130 /* This is the busiest node in its class. */
7131 if (!(env->sd->flags & SD_ASYM_PACKING))
7135 * ASYM_PACKING needs to move all the work to the lowest
7136 * numbered CPUs in the group, therefore mark all groups
7137 * higher than ourself as busy.
7139 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7143 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7150 #ifdef CONFIG_NUMA_BALANCING
7151 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7153 if (sgs->sum_nr_running > sgs->nr_numa_running)
7155 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7160 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7162 if (rq->nr_running > rq->nr_numa_running)
7164 if (rq->nr_running > rq->nr_preferred_running)
7169 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7174 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7178 #endif /* CONFIG_NUMA_BALANCING */
7181 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7182 * @env: The load balancing environment.
7183 * @sds: variable to hold the statistics for this sched_domain.
7185 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7187 struct sched_domain *child = env->sd->child;
7188 struct sched_group *sg = env->sd->groups;
7189 struct sg_lb_stats tmp_sgs;
7190 int load_idx, prefer_sibling = 0;
7191 bool overload = false, overutilized = false;
7193 if (child && child->flags & SD_PREFER_SIBLING)
7196 load_idx = get_sd_load_idx(env->sd, env->idle);
7199 struct sg_lb_stats *sgs = &tmp_sgs;
7202 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7205 sgs = &sds->local_stat;
7207 if (env->idle != CPU_NEWLY_IDLE ||
7208 time_after_eq(jiffies, sg->sgc->next_update))
7209 update_group_capacity(env->sd, env->dst_cpu);
7212 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7213 &overload, &overutilized);
7219 * In case the child domain prefers tasks go to siblings
7220 * first, lower the sg capacity so that we'll try
7221 * and move all the excess tasks away. We lower the capacity
7222 * of a group only if the local group has the capacity to fit
7223 * these excess tasks. The extra check prevents the case where
7224 * you always pull from the heaviest group when it is already
7225 * under-utilized (possible with a large weight task outweighs
7226 * the tasks on the system).
7228 if (prefer_sibling && sds->local &&
7229 group_has_capacity(env, &sds->local_stat) &&
7230 (sgs->sum_nr_running > 1)) {
7231 sgs->group_no_capacity = 1;
7232 sgs->group_type = group_classify(sg, sgs);
7236 * Ignore task groups with misfit tasks if local group has no
7237 * capacity or if per-cpu capacity isn't higher.
7239 if (sgs->group_type == group_misfit_task &&
7240 (!group_has_capacity(env, &sds->local_stat) ||
7241 !group_smaller_cpu_capacity(sg, sds->local)))
7242 sgs->group_type = group_other;
7244 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7246 sds->busiest_stat = *sgs;
7250 /* Now, start updating sd_lb_stats */
7251 sds->total_load += sgs->group_load;
7252 sds->total_capacity += sgs->group_capacity;
7255 } while (sg != env->sd->groups);
7257 if (env->sd->flags & SD_NUMA)
7258 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7260 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7262 if (!env->sd->parent) {
7263 /* update overload indicator if we are at root domain */
7264 if (env->dst_rq->rd->overload != overload)
7265 env->dst_rq->rd->overload = overload;
7267 /* Update over-utilization (tipping point, U >= 0) indicator */
7268 if (env->dst_rq->rd->overutilized != overutilized)
7269 env->dst_rq->rd->overutilized = overutilized;
7271 if (!env->dst_rq->rd->overutilized && overutilized)
7272 env->dst_rq->rd->overutilized = true;
7277 * check_asym_packing - Check to see if the group is packed into the
7280 * This is primarily intended to used at the sibling level. Some
7281 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7282 * case of POWER7, it can move to lower SMT modes only when higher
7283 * threads are idle. When in lower SMT modes, the threads will
7284 * perform better since they share less core resources. Hence when we
7285 * have idle threads, we want them to be the higher ones.
7287 * This packing function is run on idle threads. It checks to see if
7288 * the busiest CPU in this domain (core in the P7 case) has a higher
7289 * CPU number than the packing function is being run on. Here we are
7290 * assuming lower CPU number will be equivalent to lower a SMT thread
7293 * Return: 1 when packing is required and a task should be moved to
7294 * this CPU. The amount of the imbalance is returned in *imbalance.
7296 * @env: The load balancing environment.
7297 * @sds: Statistics of the sched_domain which is to be packed
7299 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7303 if (!(env->sd->flags & SD_ASYM_PACKING))
7309 busiest_cpu = group_first_cpu(sds->busiest);
7310 if (env->dst_cpu > busiest_cpu)
7313 env->imbalance = DIV_ROUND_CLOSEST(
7314 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7315 SCHED_CAPACITY_SCALE);
7321 * fix_small_imbalance - Calculate the minor imbalance that exists
7322 * amongst the groups of a sched_domain, during
7324 * @env: The load balancing environment.
7325 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7328 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7330 unsigned long tmp, capa_now = 0, capa_move = 0;
7331 unsigned int imbn = 2;
7332 unsigned long scaled_busy_load_per_task;
7333 struct sg_lb_stats *local, *busiest;
7335 local = &sds->local_stat;
7336 busiest = &sds->busiest_stat;
7338 if (!local->sum_nr_running)
7339 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7340 else if (busiest->load_per_task > local->load_per_task)
7343 scaled_busy_load_per_task =
7344 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7345 busiest->group_capacity;
7347 if (busiest->avg_load + scaled_busy_load_per_task >=
7348 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7349 env->imbalance = busiest->load_per_task;
7354 * OK, we don't have enough imbalance to justify moving tasks,
7355 * however we may be able to increase total CPU capacity used by
7359 capa_now += busiest->group_capacity *
7360 min(busiest->load_per_task, busiest->avg_load);
7361 capa_now += local->group_capacity *
7362 min(local->load_per_task, local->avg_load);
7363 capa_now /= SCHED_CAPACITY_SCALE;
7365 /* Amount of load we'd subtract */
7366 if (busiest->avg_load > scaled_busy_load_per_task) {
7367 capa_move += busiest->group_capacity *
7368 min(busiest->load_per_task,
7369 busiest->avg_load - scaled_busy_load_per_task);
7372 /* Amount of load we'd add */
7373 if (busiest->avg_load * busiest->group_capacity <
7374 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7375 tmp = (busiest->avg_load * busiest->group_capacity) /
7376 local->group_capacity;
7378 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7379 local->group_capacity;
7381 capa_move += local->group_capacity *
7382 min(local->load_per_task, local->avg_load + tmp);
7383 capa_move /= SCHED_CAPACITY_SCALE;
7385 /* Move if we gain throughput */
7386 if (capa_move > capa_now)
7387 env->imbalance = busiest->load_per_task;
7391 * calculate_imbalance - Calculate the amount of imbalance present within the
7392 * groups of a given sched_domain during load balance.
7393 * @env: load balance environment
7394 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7396 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7398 unsigned long max_pull, load_above_capacity = ~0UL;
7399 struct sg_lb_stats *local, *busiest;
7401 local = &sds->local_stat;
7402 busiest = &sds->busiest_stat;
7404 if (busiest->group_type == group_imbalanced) {
7406 * In the group_imb case we cannot rely on group-wide averages
7407 * to ensure cpu-load equilibrium, look at wider averages. XXX
7409 busiest->load_per_task =
7410 min(busiest->load_per_task, sds->avg_load);
7414 * In the presence of smp nice balancing, certain scenarios can have
7415 * max load less than avg load(as we skip the groups at or below
7416 * its cpu_capacity, while calculating max_load..)
7418 if (busiest->avg_load <= sds->avg_load ||
7419 local->avg_load >= sds->avg_load) {
7420 /* Misfitting tasks should be migrated in any case */
7421 if (busiest->group_type == group_misfit_task) {
7422 env->imbalance = busiest->group_misfit_task;
7427 * Busiest group is overloaded, local is not, use the spare
7428 * cycles to maximize throughput
7430 if (busiest->group_type == group_overloaded &&
7431 local->group_type <= group_misfit_task) {
7432 env->imbalance = busiest->load_per_task;
7437 return fix_small_imbalance(env, sds);
7441 * If there aren't any idle cpus, avoid creating some.
7443 if (busiest->group_type == group_overloaded &&
7444 local->group_type == group_overloaded) {
7445 load_above_capacity = busiest->sum_nr_running *
7447 if (load_above_capacity > busiest->group_capacity)
7448 load_above_capacity -= busiest->group_capacity;
7450 load_above_capacity = ~0UL;
7454 * We're trying to get all the cpus to the average_load, so we don't
7455 * want to push ourselves above the average load, nor do we wish to
7456 * reduce the max loaded cpu below the average load. At the same time,
7457 * we also don't want to reduce the group load below the group capacity
7458 * (so that we can implement power-savings policies etc). Thus we look
7459 * for the minimum possible imbalance.
7461 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7463 /* How much load to actually move to equalise the imbalance */
7464 env->imbalance = min(
7465 max_pull * busiest->group_capacity,
7466 (sds->avg_load - local->avg_load) * local->group_capacity
7467 ) / SCHED_CAPACITY_SCALE;
7469 /* Boost imbalance to allow misfit task to be balanced. */
7470 if (busiest->group_type == group_misfit_task)
7471 env->imbalance = max_t(long, env->imbalance,
7472 busiest->group_misfit_task);
7475 * if *imbalance is less than the average load per runnable task
7476 * there is no guarantee that any tasks will be moved so we'll have
7477 * a think about bumping its value to force at least one task to be
7480 if (env->imbalance < busiest->load_per_task)
7481 return fix_small_imbalance(env, sds);
7484 /******* find_busiest_group() helpers end here *********************/
7487 * find_busiest_group - Returns the busiest group within the sched_domain
7488 * if there is an imbalance. If there isn't an imbalance, and
7489 * the user has opted for power-savings, it returns a group whose
7490 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7491 * such a group exists.
7493 * Also calculates the amount of weighted load which should be moved
7494 * to restore balance.
7496 * @env: The load balancing environment.
7498 * Return: - The busiest group if imbalance exists.
7499 * - If no imbalance and user has opted for power-savings balance,
7500 * return the least loaded group whose CPUs can be
7501 * put to idle by rebalancing its tasks onto our group.
7503 static struct sched_group *find_busiest_group(struct lb_env *env)
7505 struct sg_lb_stats *local, *busiest;
7506 struct sd_lb_stats sds;
7508 init_sd_lb_stats(&sds);
7511 * Compute the various statistics relavent for load balancing at
7514 update_sd_lb_stats(env, &sds);
7516 if (energy_aware() && !env->dst_rq->rd->overutilized)
7519 local = &sds.local_stat;
7520 busiest = &sds.busiest_stat;
7522 /* ASYM feature bypasses nice load balance check */
7523 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7524 check_asym_packing(env, &sds))
7527 /* There is no busy sibling group to pull tasks from */
7528 if (!sds.busiest || busiest->sum_nr_running == 0)
7531 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7532 / sds.total_capacity;
7535 * If the busiest group is imbalanced the below checks don't
7536 * work because they assume all things are equal, which typically
7537 * isn't true due to cpus_allowed constraints and the like.
7539 if (busiest->group_type == group_imbalanced)
7542 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7543 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7544 busiest->group_no_capacity)
7547 /* Misfitting tasks should be dealt with regardless of the avg load */
7548 if (busiest->group_type == group_misfit_task) {
7553 * If the local group is busier than the selected busiest group
7554 * don't try and pull any tasks.
7556 if (local->avg_load >= busiest->avg_load)
7560 * Don't pull any tasks if this group is already above the domain
7563 if (local->avg_load >= sds.avg_load)
7566 if (env->idle == CPU_IDLE) {
7568 * This cpu is idle. If the busiest group is not overloaded
7569 * and there is no imbalance between this and busiest group
7570 * wrt idle cpus, it is balanced. The imbalance becomes
7571 * significant if the diff is greater than 1 otherwise we
7572 * might end up to just move the imbalance on another group
7574 if ((busiest->group_type != group_overloaded) &&
7575 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7576 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7580 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7581 * imbalance_pct to be conservative.
7583 if (100 * busiest->avg_load <=
7584 env->sd->imbalance_pct * local->avg_load)
7589 env->busiest_group_type = busiest->group_type;
7590 /* Looks like there is an imbalance. Compute it */
7591 calculate_imbalance(env, &sds);
7600 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7602 static struct rq *find_busiest_queue(struct lb_env *env,
7603 struct sched_group *group)
7605 struct rq *busiest = NULL, *rq;
7606 unsigned long busiest_load = 0, busiest_capacity = 1;
7609 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7610 unsigned long capacity, wl;
7614 rt = fbq_classify_rq(rq);
7617 * We classify groups/runqueues into three groups:
7618 * - regular: there are !numa tasks
7619 * - remote: there are numa tasks that run on the 'wrong' node
7620 * - all: there is no distinction
7622 * In order to avoid migrating ideally placed numa tasks,
7623 * ignore those when there's better options.
7625 * If we ignore the actual busiest queue to migrate another
7626 * task, the next balance pass can still reduce the busiest
7627 * queue by moving tasks around inside the node.
7629 * If we cannot move enough load due to this classification
7630 * the next pass will adjust the group classification and
7631 * allow migration of more tasks.
7633 * Both cases only affect the total convergence complexity.
7635 if (rt > env->fbq_type)
7638 capacity = capacity_of(i);
7640 wl = weighted_cpuload(i);
7643 * When comparing with imbalance, use weighted_cpuload()
7644 * which is not scaled with the cpu capacity.
7647 if (rq->nr_running == 1 && wl > env->imbalance &&
7648 !check_cpu_capacity(rq, env->sd) &&
7649 env->busiest_group_type != group_misfit_task)
7653 * For the load comparisons with the other cpu's, consider
7654 * the weighted_cpuload() scaled with the cpu capacity, so
7655 * that the load can be moved away from the cpu that is
7656 * potentially running at a lower capacity.
7658 * Thus we're looking for max(wl_i / capacity_i), crosswise
7659 * multiplication to rid ourselves of the division works out
7660 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7661 * our previous maximum.
7663 if (wl * busiest_capacity > busiest_load * capacity) {
7665 busiest_capacity = capacity;
7674 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7675 * so long as it is large enough.
7677 #define MAX_PINNED_INTERVAL 512
7679 /* Working cpumask for load_balance and load_balance_newidle. */
7680 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7682 static int need_active_balance(struct lb_env *env)
7684 struct sched_domain *sd = env->sd;
7686 if (env->idle == CPU_NEWLY_IDLE) {
7689 * ASYM_PACKING needs to force migrate tasks from busy but
7690 * higher numbered CPUs in order to pack all tasks in the
7691 * lowest numbered CPUs.
7693 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7698 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7699 * It's worth migrating the task if the src_cpu's capacity is reduced
7700 * because of other sched_class or IRQs if more capacity stays
7701 * available on dst_cpu.
7703 if ((env->idle != CPU_NOT_IDLE) &&
7704 (env->src_rq->cfs.h_nr_running == 1)) {
7705 if ((check_cpu_capacity(env->src_rq, sd)) &&
7706 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7710 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7711 env->src_rq->cfs.h_nr_running == 1 &&
7712 cpu_overutilized(env->src_cpu) &&
7713 !cpu_overutilized(env->dst_cpu)) {
7717 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7720 static int active_load_balance_cpu_stop(void *data);
7722 static int should_we_balance(struct lb_env *env)
7724 struct sched_group *sg = env->sd->groups;
7725 struct cpumask *sg_cpus, *sg_mask;
7726 int cpu, balance_cpu = -1;
7729 * In the newly idle case, we will allow all the cpu's
7730 * to do the newly idle load balance.
7732 if (env->idle == CPU_NEWLY_IDLE)
7735 sg_cpus = sched_group_cpus(sg);
7736 sg_mask = sched_group_mask(sg);
7737 /* Try to find first idle cpu */
7738 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7739 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7746 if (balance_cpu == -1)
7747 balance_cpu = group_balance_cpu(sg);
7750 * First idle cpu or the first cpu(busiest) in this sched group
7751 * is eligible for doing load balancing at this and above domains.
7753 return balance_cpu == env->dst_cpu;
7757 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7758 * tasks if there is an imbalance.
7760 static int load_balance(int this_cpu, struct rq *this_rq,
7761 struct sched_domain *sd, enum cpu_idle_type idle,
7762 int *continue_balancing)
7764 int ld_moved, cur_ld_moved, active_balance = 0;
7765 struct sched_domain *sd_parent = sd->parent;
7766 struct sched_group *group;
7768 unsigned long flags;
7769 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7771 struct lb_env env = {
7773 .dst_cpu = this_cpu,
7775 .dst_grpmask = sched_group_cpus(sd->groups),
7777 .loop_break = sched_nr_migrate_break,
7780 .tasks = LIST_HEAD_INIT(env.tasks),
7784 * For NEWLY_IDLE load_balancing, we don't need to consider
7785 * other cpus in our group
7787 if (idle == CPU_NEWLY_IDLE)
7788 env.dst_grpmask = NULL;
7790 cpumask_copy(cpus, cpu_active_mask);
7792 schedstat_inc(sd, lb_count[idle]);
7795 if (!should_we_balance(&env)) {
7796 *continue_balancing = 0;
7800 group = find_busiest_group(&env);
7802 schedstat_inc(sd, lb_nobusyg[idle]);
7806 busiest = find_busiest_queue(&env, group);
7808 schedstat_inc(sd, lb_nobusyq[idle]);
7812 BUG_ON(busiest == env.dst_rq);
7814 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7816 env.src_cpu = busiest->cpu;
7817 env.src_rq = busiest;
7820 if (busiest->nr_running > 1) {
7822 * Attempt to move tasks. If find_busiest_group has found
7823 * an imbalance but busiest->nr_running <= 1, the group is
7824 * still unbalanced. ld_moved simply stays zero, so it is
7825 * correctly treated as an imbalance.
7827 env.flags |= LBF_ALL_PINNED;
7828 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7831 raw_spin_lock_irqsave(&busiest->lock, flags);
7834 * cur_ld_moved - load moved in current iteration
7835 * ld_moved - cumulative load moved across iterations
7837 cur_ld_moved = detach_tasks(&env);
7839 * We want to potentially lower env.src_cpu's OPP.
7842 update_capacity_of(env.src_cpu);
7845 * We've detached some tasks from busiest_rq. Every
7846 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7847 * unlock busiest->lock, and we are able to be sure
7848 * that nobody can manipulate the tasks in parallel.
7849 * See task_rq_lock() family for the details.
7852 raw_spin_unlock(&busiest->lock);
7856 ld_moved += cur_ld_moved;
7859 local_irq_restore(flags);
7861 if (env.flags & LBF_NEED_BREAK) {
7862 env.flags &= ~LBF_NEED_BREAK;
7867 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7868 * us and move them to an alternate dst_cpu in our sched_group
7869 * where they can run. The upper limit on how many times we
7870 * iterate on same src_cpu is dependent on number of cpus in our
7873 * This changes load balance semantics a bit on who can move
7874 * load to a given_cpu. In addition to the given_cpu itself
7875 * (or a ilb_cpu acting on its behalf where given_cpu is
7876 * nohz-idle), we now have balance_cpu in a position to move
7877 * load to given_cpu. In rare situations, this may cause
7878 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7879 * _independently_ and at _same_ time to move some load to
7880 * given_cpu) causing exceess load to be moved to given_cpu.
7881 * This however should not happen so much in practice and
7882 * moreover subsequent load balance cycles should correct the
7883 * excess load moved.
7885 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7887 /* Prevent to re-select dst_cpu via env's cpus */
7888 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7890 env.dst_rq = cpu_rq(env.new_dst_cpu);
7891 env.dst_cpu = env.new_dst_cpu;
7892 env.flags &= ~LBF_DST_PINNED;
7894 env.loop_break = sched_nr_migrate_break;
7897 * Go back to "more_balance" rather than "redo" since we
7898 * need to continue with same src_cpu.
7904 * We failed to reach balance because of affinity.
7907 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7909 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7910 *group_imbalance = 1;
7913 /* All tasks on this runqueue were pinned by CPU affinity */
7914 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7915 cpumask_clear_cpu(cpu_of(busiest), cpus);
7916 if (!cpumask_empty(cpus)) {
7918 env.loop_break = sched_nr_migrate_break;
7921 goto out_all_pinned;
7926 schedstat_inc(sd, lb_failed[idle]);
7928 * Increment the failure counter only on periodic balance.
7929 * We do not want newidle balance, which can be very
7930 * frequent, pollute the failure counter causing
7931 * excessive cache_hot migrations and active balances.
7933 if (idle != CPU_NEWLY_IDLE)
7934 if (env.src_grp_nr_running > 1)
7935 sd->nr_balance_failed++;
7937 if (need_active_balance(&env)) {
7938 raw_spin_lock_irqsave(&busiest->lock, flags);
7940 /* don't kick the active_load_balance_cpu_stop,
7941 * if the curr task on busiest cpu can't be
7944 if (!cpumask_test_cpu(this_cpu,
7945 tsk_cpus_allowed(busiest->curr))) {
7946 raw_spin_unlock_irqrestore(&busiest->lock,
7948 env.flags |= LBF_ALL_PINNED;
7949 goto out_one_pinned;
7953 * ->active_balance synchronizes accesses to
7954 * ->active_balance_work. Once set, it's cleared
7955 * only after active load balance is finished.
7957 if (!busiest->active_balance) {
7958 busiest->active_balance = 1;
7959 busiest->push_cpu = this_cpu;
7962 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7964 if (active_balance) {
7965 stop_one_cpu_nowait(cpu_of(busiest),
7966 active_load_balance_cpu_stop, busiest,
7967 &busiest->active_balance_work);
7971 * We've kicked active balancing, reset the failure
7974 sd->nr_balance_failed = sd->cache_nice_tries+1;
7977 sd->nr_balance_failed = 0;
7979 if (likely(!active_balance)) {
7980 /* We were unbalanced, so reset the balancing interval */
7981 sd->balance_interval = sd->min_interval;
7984 * If we've begun active balancing, start to back off. This
7985 * case may not be covered by the all_pinned logic if there
7986 * is only 1 task on the busy runqueue (because we don't call
7989 if (sd->balance_interval < sd->max_interval)
7990 sd->balance_interval *= 2;
7997 * We reach balance although we may have faced some affinity
7998 * constraints. Clear the imbalance flag if it was set.
8001 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8003 if (*group_imbalance)
8004 *group_imbalance = 0;
8009 * We reach balance because all tasks are pinned at this level so
8010 * we can't migrate them. Let the imbalance flag set so parent level
8011 * can try to migrate them.
8013 schedstat_inc(sd, lb_balanced[idle]);
8015 sd->nr_balance_failed = 0;
8018 /* tune up the balancing interval */
8019 if (((env.flags & LBF_ALL_PINNED) &&
8020 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8021 (sd->balance_interval < sd->max_interval))
8022 sd->balance_interval *= 2;
8029 static inline unsigned long
8030 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8032 unsigned long interval = sd->balance_interval;
8035 interval *= sd->busy_factor;
8037 /* scale ms to jiffies */
8038 interval = msecs_to_jiffies(interval);
8039 interval = clamp(interval, 1UL, max_load_balance_interval);
8045 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8047 unsigned long interval, next;
8049 interval = get_sd_balance_interval(sd, cpu_busy);
8050 next = sd->last_balance + interval;
8052 if (time_after(*next_balance, next))
8053 *next_balance = next;
8057 * idle_balance is called by schedule() if this_cpu is about to become
8058 * idle. Attempts to pull tasks from other CPUs.
8060 static int idle_balance(struct rq *this_rq)
8062 unsigned long next_balance = jiffies + HZ;
8063 int this_cpu = this_rq->cpu;
8064 struct sched_domain *sd;
8065 int pulled_task = 0;
8068 idle_enter_fair(this_rq);
8071 * We must set idle_stamp _before_ calling idle_balance(), such that we
8072 * measure the duration of idle_balance() as idle time.
8074 this_rq->idle_stamp = rq_clock(this_rq);
8076 if (!energy_aware() &&
8077 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8078 !this_rq->rd->overload)) {
8080 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8082 update_next_balance(sd, 0, &next_balance);
8088 raw_spin_unlock(&this_rq->lock);
8090 update_blocked_averages(this_cpu);
8092 for_each_domain(this_cpu, sd) {
8093 int continue_balancing = 1;
8094 u64 t0, domain_cost;
8096 if (!(sd->flags & SD_LOAD_BALANCE))
8099 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8100 update_next_balance(sd, 0, &next_balance);
8104 if (sd->flags & SD_BALANCE_NEWIDLE) {
8105 t0 = sched_clock_cpu(this_cpu);
8107 pulled_task = load_balance(this_cpu, this_rq,
8109 &continue_balancing);
8111 domain_cost = sched_clock_cpu(this_cpu) - t0;
8112 if (domain_cost > sd->max_newidle_lb_cost)
8113 sd->max_newidle_lb_cost = domain_cost;
8115 curr_cost += domain_cost;
8118 update_next_balance(sd, 0, &next_balance);
8121 * Stop searching for tasks to pull if there are
8122 * now runnable tasks on this rq.
8124 if (pulled_task || this_rq->nr_running > 0)
8129 raw_spin_lock(&this_rq->lock);
8131 if (curr_cost > this_rq->max_idle_balance_cost)
8132 this_rq->max_idle_balance_cost = curr_cost;
8135 * While browsing the domains, we released the rq lock, a task could
8136 * have been enqueued in the meantime. Since we're not going idle,
8137 * pretend we pulled a task.
8139 if (this_rq->cfs.h_nr_running && !pulled_task)
8143 /* Move the next balance forward */
8144 if (time_after(this_rq->next_balance, next_balance))
8145 this_rq->next_balance = next_balance;
8147 /* Is there a task of a high priority class? */
8148 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8152 idle_exit_fair(this_rq);
8153 this_rq->idle_stamp = 0;
8160 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8161 * running tasks off the busiest CPU onto idle CPUs. It requires at
8162 * least 1 task to be running on each physical CPU where possible, and
8163 * avoids physical / logical imbalances.
8165 static int active_load_balance_cpu_stop(void *data)
8167 struct rq *busiest_rq = data;
8168 int busiest_cpu = cpu_of(busiest_rq);
8169 int target_cpu = busiest_rq->push_cpu;
8170 struct rq *target_rq = cpu_rq(target_cpu);
8171 struct sched_domain *sd;
8172 struct task_struct *p = NULL;
8174 raw_spin_lock_irq(&busiest_rq->lock);
8176 /* make sure the requested cpu hasn't gone down in the meantime */
8177 if (unlikely(busiest_cpu != smp_processor_id() ||
8178 !busiest_rq->active_balance))
8181 /* Is there any task to move? */
8182 if (busiest_rq->nr_running <= 1)
8186 * This condition is "impossible", if it occurs
8187 * we need to fix it. Originally reported by
8188 * Bjorn Helgaas on a 128-cpu setup.
8190 BUG_ON(busiest_rq == target_rq);
8192 /* Search for an sd spanning us and the target CPU. */
8194 for_each_domain(target_cpu, sd) {
8195 if ((sd->flags & SD_LOAD_BALANCE) &&
8196 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8201 struct lb_env env = {
8203 .dst_cpu = target_cpu,
8204 .dst_rq = target_rq,
8205 .src_cpu = busiest_rq->cpu,
8206 .src_rq = busiest_rq,
8210 schedstat_inc(sd, alb_count);
8212 p = detach_one_task(&env);
8214 schedstat_inc(sd, alb_pushed);
8216 * We want to potentially lower env.src_cpu's OPP.
8218 update_capacity_of(env.src_cpu);
8221 schedstat_inc(sd, alb_failed);
8225 busiest_rq->active_balance = 0;
8226 raw_spin_unlock(&busiest_rq->lock);
8229 attach_one_task(target_rq, p);
8236 static inline int on_null_domain(struct rq *rq)
8238 return unlikely(!rcu_dereference_sched(rq->sd));
8241 #ifdef CONFIG_NO_HZ_COMMON
8243 * idle load balancing details
8244 * - When one of the busy CPUs notice that there may be an idle rebalancing
8245 * needed, they will kick the idle load balancer, which then does idle
8246 * load balancing for all the idle CPUs.
8249 cpumask_var_t idle_cpus_mask;
8251 unsigned long next_balance; /* in jiffy units */
8252 } nohz ____cacheline_aligned;
8254 static inline int find_new_ilb(void)
8256 int ilb = cpumask_first(nohz.idle_cpus_mask);
8258 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8265 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8266 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8267 * CPU (if there is one).
8269 static void nohz_balancer_kick(void)
8273 nohz.next_balance++;
8275 ilb_cpu = find_new_ilb();
8277 if (ilb_cpu >= nr_cpu_ids)
8280 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8283 * Use smp_send_reschedule() instead of resched_cpu().
8284 * This way we generate a sched IPI on the target cpu which
8285 * is idle. And the softirq performing nohz idle load balance
8286 * will be run before returning from the IPI.
8288 smp_send_reschedule(ilb_cpu);
8292 static inline void nohz_balance_exit_idle(int cpu)
8294 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8296 * Completely isolated CPUs don't ever set, so we must test.
8298 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8299 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8300 atomic_dec(&nohz.nr_cpus);
8302 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8306 static inline void set_cpu_sd_state_busy(void)
8308 struct sched_domain *sd;
8309 int cpu = smp_processor_id();
8312 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8314 if (!sd || !sd->nohz_idle)
8318 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8323 void set_cpu_sd_state_idle(void)
8325 struct sched_domain *sd;
8326 int cpu = smp_processor_id();
8329 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8331 if (!sd || sd->nohz_idle)
8335 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8341 * This routine will record that the cpu is going idle with tick stopped.
8342 * This info will be used in performing idle load balancing in the future.
8344 void nohz_balance_enter_idle(int cpu)
8347 * If this cpu is going down, then nothing needs to be done.
8349 if (!cpu_active(cpu))
8352 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8356 * If we're a completely isolated CPU, we don't play.
8358 if (on_null_domain(cpu_rq(cpu)))
8361 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8362 atomic_inc(&nohz.nr_cpus);
8363 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8366 static int sched_ilb_notifier(struct notifier_block *nfb,
8367 unsigned long action, void *hcpu)
8369 switch (action & ~CPU_TASKS_FROZEN) {
8371 nohz_balance_exit_idle(smp_processor_id());
8379 static DEFINE_SPINLOCK(balancing);
8382 * Scale the max load_balance interval with the number of CPUs in the system.
8383 * This trades load-balance latency on larger machines for less cross talk.
8385 void update_max_interval(void)
8387 max_load_balance_interval = HZ*num_online_cpus()/10;
8391 * It checks each scheduling domain to see if it is due to be balanced,
8392 * and initiates a balancing operation if so.
8394 * Balancing parameters are set up in init_sched_domains.
8396 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8398 int continue_balancing = 1;
8400 unsigned long interval;
8401 struct sched_domain *sd;
8402 /* Earliest time when we have to do rebalance again */
8403 unsigned long next_balance = jiffies + 60*HZ;
8404 int update_next_balance = 0;
8405 int need_serialize, need_decay = 0;
8408 update_blocked_averages(cpu);
8411 for_each_domain(cpu, sd) {
8413 * Decay the newidle max times here because this is a regular
8414 * visit to all the domains. Decay ~1% per second.
8416 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8417 sd->max_newidle_lb_cost =
8418 (sd->max_newidle_lb_cost * 253) / 256;
8419 sd->next_decay_max_lb_cost = jiffies + HZ;
8422 max_cost += sd->max_newidle_lb_cost;
8424 if (!(sd->flags & SD_LOAD_BALANCE))
8428 * Stop the load balance at this level. There is another
8429 * CPU in our sched group which is doing load balancing more
8432 if (!continue_balancing) {
8438 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8440 need_serialize = sd->flags & SD_SERIALIZE;
8441 if (need_serialize) {
8442 if (!spin_trylock(&balancing))
8446 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8447 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8449 * The LBF_DST_PINNED logic could have changed
8450 * env->dst_cpu, so we can't know our idle
8451 * state even if we migrated tasks. Update it.
8453 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8455 sd->last_balance = jiffies;
8456 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8459 spin_unlock(&balancing);
8461 if (time_after(next_balance, sd->last_balance + interval)) {
8462 next_balance = sd->last_balance + interval;
8463 update_next_balance = 1;
8468 * Ensure the rq-wide value also decays but keep it at a
8469 * reasonable floor to avoid funnies with rq->avg_idle.
8471 rq->max_idle_balance_cost =
8472 max((u64)sysctl_sched_migration_cost, max_cost);
8477 * next_balance will be updated only when there is a need.
8478 * When the cpu is attached to null domain for ex, it will not be
8481 if (likely(update_next_balance)) {
8482 rq->next_balance = next_balance;
8484 #ifdef CONFIG_NO_HZ_COMMON
8486 * If this CPU has been elected to perform the nohz idle
8487 * balance. Other idle CPUs have already rebalanced with
8488 * nohz_idle_balance() and nohz.next_balance has been
8489 * updated accordingly. This CPU is now running the idle load
8490 * balance for itself and we need to update the
8491 * nohz.next_balance accordingly.
8493 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8494 nohz.next_balance = rq->next_balance;
8499 #ifdef CONFIG_NO_HZ_COMMON
8501 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8502 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8504 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8506 int this_cpu = this_rq->cpu;
8509 /* Earliest time when we have to do rebalance again */
8510 unsigned long next_balance = jiffies + 60*HZ;
8511 int update_next_balance = 0;
8513 if (idle != CPU_IDLE ||
8514 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8517 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8518 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8522 * If this cpu gets work to do, stop the load balancing
8523 * work being done for other cpus. Next load
8524 * balancing owner will pick it up.
8529 rq = cpu_rq(balance_cpu);
8532 * If time for next balance is due,
8535 if (time_after_eq(jiffies, rq->next_balance)) {
8536 raw_spin_lock_irq(&rq->lock);
8537 update_rq_clock(rq);
8538 update_idle_cpu_load(rq);
8539 raw_spin_unlock_irq(&rq->lock);
8540 rebalance_domains(rq, CPU_IDLE);
8543 if (time_after(next_balance, rq->next_balance)) {
8544 next_balance = rq->next_balance;
8545 update_next_balance = 1;
8550 * next_balance will be updated only when there is a need.
8551 * When the CPU is attached to null domain for ex, it will not be
8554 if (likely(update_next_balance))
8555 nohz.next_balance = next_balance;
8557 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8561 * Current heuristic for kicking the idle load balancer in the presence
8562 * of an idle cpu in the system.
8563 * - This rq has more than one task.
8564 * - This rq has at least one CFS task and the capacity of the CPU is
8565 * significantly reduced because of RT tasks or IRQs.
8566 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8567 * multiple busy cpu.
8568 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8569 * domain span are idle.
8571 static inline bool nohz_kick_needed(struct rq *rq)
8573 unsigned long now = jiffies;
8574 struct sched_domain *sd;
8575 struct sched_group_capacity *sgc;
8576 int nr_busy, cpu = rq->cpu;
8579 if (unlikely(rq->idle_balance))
8583 * We may be recently in ticked or tickless idle mode. At the first
8584 * busy tick after returning from idle, we will update the busy stats.
8586 set_cpu_sd_state_busy();
8587 nohz_balance_exit_idle(cpu);
8590 * None are in tickless mode and hence no need for NOHZ idle load
8593 if (likely(!atomic_read(&nohz.nr_cpus)))
8596 if (time_before(now, nohz.next_balance))
8599 if (rq->nr_running >= 2 &&
8600 (!energy_aware() || cpu_overutilized(cpu)))
8604 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8605 if (sd && !energy_aware()) {
8606 sgc = sd->groups->sgc;
8607 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8616 sd = rcu_dereference(rq->sd);
8618 if ((rq->cfs.h_nr_running >= 1) &&
8619 check_cpu_capacity(rq, sd)) {
8625 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8626 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8627 sched_domain_span(sd)) < cpu)) {
8637 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8641 * run_rebalance_domains is triggered when needed from the scheduler tick.
8642 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8644 static void run_rebalance_domains(struct softirq_action *h)
8646 struct rq *this_rq = this_rq();
8647 enum cpu_idle_type idle = this_rq->idle_balance ?
8648 CPU_IDLE : CPU_NOT_IDLE;
8651 * If this cpu has a pending nohz_balance_kick, then do the
8652 * balancing on behalf of the other idle cpus whose ticks are
8653 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8654 * give the idle cpus a chance to load balance. Else we may
8655 * load balance only within the local sched_domain hierarchy
8656 * and abort nohz_idle_balance altogether if we pull some load.
8658 nohz_idle_balance(this_rq, idle);
8659 rebalance_domains(this_rq, idle);
8663 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8665 void trigger_load_balance(struct rq *rq)
8667 /* Don't need to rebalance while attached to NULL domain */
8668 if (unlikely(on_null_domain(rq)))
8671 if (time_after_eq(jiffies, rq->next_balance))
8672 raise_softirq(SCHED_SOFTIRQ);
8673 #ifdef CONFIG_NO_HZ_COMMON
8674 if (nohz_kick_needed(rq))
8675 nohz_balancer_kick();
8679 static void rq_online_fair(struct rq *rq)
8683 update_runtime_enabled(rq);
8686 static void rq_offline_fair(struct rq *rq)
8690 /* Ensure any throttled groups are reachable by pick_next_task */
8691 unthrottle_offline_cfs_rqs(rq);
8694 #endif /* CONFIG_SMP */
8697 * scheduler tick hitting a task of our scheduling class:
8699 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8701 struct cfs_rq *cfs_rq;
8702 struct sched_entity *se = &curr->se;
8704 for_each_sched_entity(se) {
8705 cfs_rq = cfs_rq_of(se);
8706 entity_tick(cfs_rq, se, queued);
8709 if (static_branch_unlikely(&sched_numa_balancing))
8710 task_tick_numa(rq, curr);
8712 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8713 rq->rd->overutilized = true;
8715 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8719 * called on fork with the child task as argument from the parent's context
8720 * - child not yet on the tasklist
8721 * - preemption disabled
8723 static void task_fork_fair(struct task_struct *p)
8725 struct cfs_rq *cfs_rq;
8726 struct sched_entity *se = &p->se, *curr;
8727 int this_cpu = smp_processor_id();
8728 struct rq *rq = this_rq();
8729 unsigned long flags;
8731 raw_spin_lock_irqsave(&rq->lock, flags);
8733 update_rq_clock(rq);
8735 cfs_rq = task_cfs_rq(current);
8736 curr = cfs_rq->curr;
8739 * Not only the cpu but also the task_group of the parent might have
8740 * been changed after parent->se.parent,cfs_rq were copied to
8741 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8742 * of child point to valid ones.
8745 __set_task_cpu(p, this_cpu);
8748 update_curr(cfs_rq);
8751 se->vruntime = curr->vruntime;
8752 place_entity(cfs_rq, se, 1);
8754 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8756 * Upon rescheduling, sched_class::put_prev_task() will place
8757 * 'current' within the tree based on its new key value.
8759 swap(curr->vruntime, se->vruntime);
8763 se->vruntime -= cfs_rq->min_vruntime;
8765 raw_spin_unlock_irqrestore(&rq->lock, flags);
8769 * Priority of the task has changed. Check to see if we preempt
8773 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8775 if (!task_on_rq_queued(p))
8779 * Reschedule if we are currently running on this runqueue and
8780 * our priority decreased, or if we are not currently running on
8781 * this runqueue and our priority is higher than the current's
8783 if (rq->curr == p) {
8784 if (p->prio > oldprio)
8787 check_preempt_curr(rq, p, 0);
8790 static inline bool vruntime_normalized(struct task_struct *p)
8792 struct sched_entity *se = &p->se;
8795 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8796 * the dequeue_entity(.flags=0) will already have normalized the
8803 * When !on_rq, vruntime of the task has usually NOT been normalized.
8804 * But there are some cases where it has already been normalized:
8806 * - A forked child which is waiting for being woken up by
8807 * wake_up_new_task().
8808 * - A task which has been woken up by try_to_wake_up() and
8809 * waiting for actually being woken up by sched_ttwu_pending().
8811 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8817 static void detach_task_cfs_rq(struct task_struct *p)
8819 struct sched_entity *se = &p->se;
8820 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8822 if (!vruntime_normalized(p)) {
8824 * Fix up our vruntime so that the current sleep doesn't
8825 * cause 'unlimited' sleep bonus.
8827 place_entity(cfs_rq, se, 0);
8828 se->vruntime -= cfs_rq->min_vruntime;
8831 /* Catch up with the cfs_rq and remove our load when we leave */
8832 detach_entity_load_avg(cfs_rq, se);
8835 static void attach_task_cfs_rq(struct task_struct *p)
8837 struct sched_entity *se = &p->se;
8838 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8840 #ifdef CONFIG_FAIR_GROUP_SCHED
8842 * Since the real-depth could have been changed (only FAIR
8843 * class maintain depth value), reset depth properly.
8845 se->depth = se->parent ? se->parent->depth + 1 : 0;
8848 /* Synchronize task with its cfs_rq */
8849 attach_entity_load_avg(cfs_rq, se);
8851 if (!vruntime_normalized(p))
8852 se->vruntime += cfs_rq->min_vruntime;
8855 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8857 detach_task_cfs_rq(p);
8860 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8862 attach_task_cfs_rq(p);
8864 if (task_on_rq_queued(p)) {
8866 * We were most likely switched from sched_rt, so
8867 * kick off the schedule if running, otherwise just see
8868 * if we can still preempt the current task.
8873 check_preempt_curr(rq, p, 0);
8877 /* Account for a task changing its policy or group.
8879 * This routine is mostly called to set cfs_rq->curr field when a task
8880 * migrates between groups/classes.
8882 static void set_curr_task_fair(struct rq *rq)
8884 struct sched_entity *se = &rq->curr->se;
8886 for_each_sched_entity(se) {
8887 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8889 set_next_entity(cfs_rq, se);
8890 /* ensure bandwidth has been allocated on our new cfs_rq */
8891 account_cfs_rq_runtime(cfs_rq, 0);
8895 void init_cfs_rq(struct cfs_rq *cfs_rq)
8897 cfs_rq->tasks_timeline = RB_ROOT;
8898 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8899 #ifndef CONFIG_64BIT
8900 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8903 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8904 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8908 #ifdef CONFIG_FAIR_GROUP_SCHED
8909 static void task_move_group_fair(struct task_struct *p)
8911 detach_task_cfs_rq(p);
8912 set_task_rq(p, task_cpu(p));
8915 /* Tell se's cfs_rq has been changed -- migrated */
8916 p->se.avg.last_update_time = 0;
8918 attach_task_cfs_rq(p);
8921 void free_fair_sched_group(struct task_group *tg)
8925 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8927 for_each_possible_cpu(i) {
8929 kfree(tg->cfs_rq[i]);
8932 remove_entity_load_avg(tg->se[i]);
8941 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8943 struct cfs_rq *cfs_rq;
8944 struct sched_entity *se;
8947 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8950 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8954 tg->shares = NICE_0_LOAD;
8956 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8958 for_each_possible_cpu(i) {
8959 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8960 GFP_KERNEL, cpu_to_node(i));
8964 se = kzalloc_node(sizeof(struct sched_entity),
8965 GFP_KERNEL, cpu_to_node(i));
8969 init_cfs_rq(cfs_rq);
8970 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8971 init_entity_runnable_average(se);
8982 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8984 struct rq *rq = cpu_rq(cpu);
8985 unsigned long flags;
8988 * Only empty task groups can be destroyed; so we can speculatively
8989 * check on_list without danger of it being re-added.
8991 if (!tg->cfs_rq[cpu]->on_list)
8994 raw_spin_lock_irqsave(&rq->lock, flags);
8995 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8996 raw_spin_unlock_irqrestore(&rq->lock, flags);
8999 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9000 struct sched_entity *se, int cpu,
9001 struct sched_entity *parent)
9003 struct rq *rq = cpu_rq(cpu);
9007 init_cfs_rq_runtime(cfs_rq);
9009 tg->cfs_rq[cpu] = cfs_rq;
9012 /* se could be NULL for root_task_group */
9017 se->cfs_rq = &rq->cfs;
9020 se->cfs_rq = parent->my_q;
9021 se->depth = parent->depth + 1;
9025 /* guarantee group entities always have weight */
9026 update_load_set(&se->load, NICE_0_LOAD);
9027 se->parent = parent;
9030 static DEFINE_MUTEX(shares_mutex);
9032 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9035 unsigned long flags;
9038 * We can't change the weight of the root cgroup.
9043 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9045 mutex_lock(&shares_mutex);
9046 if (tg->shares == shares)
9049 tg->shares = shares;
9050 for_each_possible_cpu(i) {
9051 struct rq *rq = cpu_rq(i);
9052 struct sched_entity *se;
9055 /* Propagate contribution to hierarchy */
9056 raw_spin_lock_irqsave(&rq->lock, flags);
9058 /* Possible calls to update_curr() need rq clock */
9059 update_rq_clock(rq);
9060 for_each_sched_entity(se)
9061 update_cfs_shares(group_cfs_rq(se));
9062 raw_spin_unlock_irqrestore(&rq->lock, flags);
9066 mutex_unlock(&shares_mutex);
9069 #else /* CONFIG_FAIR_GROUP_SCHED */
9071 void free_fair_sched_group(struct task_group *tg) { }
9073 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9078 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9080 #endif /* CONFIG_FAIR_GROUP_SCHED */
9083 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9085 struct sched_entity *se = &task->se;
9086 unsigned int rr_interval = 0;
9089 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9092 if (rq->cfs.load.weight)
9093 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9099 * All the scheduling class methods:
9101 const struct sched_class fair_sched_class = {
9102 .next = &idle_sched_class,
9103 .enqueue_task = enqueue_task_fair,
9104 .dequeue_task = dequeue_task_fair,
9105 .yield_task = yield_task_fair,
9106 .yield_to_task = yield_to_task_fair,
9108 .check_preempt_curr = check_preempt_wakeup,
9110 .pick_next_task = pick_next_task_fair,
9111 .put_prev_task = put_prev_task_fair,
9114 .select_task_rq = select_task_rq_fair,
9115 .migrate_task_rq = migrate_task_rq_fair,
9117 .rq_online = rq_online_fair,
9118 .rq_offline = rq_offline_fair,
9120 .task_waking = task_waking_fair,
9121 .task_dead = task_dead_fair,
9122 .set_cpus_allowed = set_cpus_allowed_common,
9125 .set_curr_task = set_curr_task_fair,
9126 .task_tick = task_tick_fair,
9127 .task_fork = task_fork_fair,
9129 .prio_changed = prio_changed_fair,
9130 .switched_from = switched_from_fair,
9131 .switched_to = switched_to_fair,
9133 .get_rr_interval = get_rr_interval_fair,
9135 .update_curr = update_curr_fair,
9137 #ifdef CONFIG_FAIR_GROUP_SCHED
9138 .task_move_group = task_move_group_fair,
9142 #ifdef CONFIG_SCHED_DEBUG
9143 void print_cfs_stats(struct seq_file *m, int cpu)
9145 struct cfs_rq *cfs_rq;
9148 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9149 print_cfs_rq(m, cpu, cfs_rq);
9153 #ifdef CONFIG_NUMA_BALANCING
9154 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9157 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9159 for_each_online_node(node) {
9160 if (p->numa_faults) {
9161 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9162 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9164 if (p->numa_group) {
9165 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9166 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9168 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9171 #endif /* CONFIG_NUMA_BALANCING */
9172 #endif /* CONFIG_SCHED_DEBUG */
9174 __init void init_sched_fair_class(void)
9177 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9179 #ifdef CONFIG_NO_HZ_COMMON
9180 nohz.next_balance = jiffies;
9181 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9182 cpu_notifier(sched_ilb_notifier, 0);