2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
40 * Targeted preemption latency for CPU-bound tasks:
41 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 unsigned int sysctl_sched_latency = 6000000ULL;
52 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
55 * The initial- and re-scaling of tunables is configurable
56 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
59 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
60 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
61 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
63 enum sched_tunable_scaling sysctl_sched_tunable_scaling
64 = SCHED_TUNABLESCALING_LOG;
67 * Minimal preemption granularity for CPU-bound tasks:
68 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
70 unsigned int sysctl_sched_min_granularity = 750000ULL;
71 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
74 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
76 static unsigned int sched_nr_latency = 8;
79 * After fork, child runs first. If set to 0 (default) then
80 * parent will (try to) run first.
82 unsigned int sysctl_sched_child_runs_first __read_mostly;
85 * SCHED_OTHER wake-up granularity.
86 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
88 * This option delays the preemption effects of decoupled workloads
89 * and reduces their over-scheduling. Synchronous workloads will still
90 * have immediate wakeup/sleep latencies.
92 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
93 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
95 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
98 * The exponential sliding window over which load is averaged for shares
102 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
104 #ifdef CONFIG_CFS_BANDWIDTH
106 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
107 * each time a cfs_rq requests quota.
109 * Note: in the case that the slice exceeds the runtime remaining (either due
110 * to consumption or the quota being specified to be smaller than the slice)
111 * we will always only issue the remaining available time.
113 * default: 5 msec, units: microseconds
115 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
118 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
124 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
130 static inline void update_load_set(struct load_weight *lw, unsigned long w)
137 * Increase the granularity value when there are more CPUs,
138 * because with more CPUs the 'effective latency' as visible
139 * to users decreases. But the relationship is not linear,
140 * so pick a second-best guess by going with the log2 of the
143 * This idea comes from the SD scheduler of Con Kolivas:
145 static unsigned int get_update_sysctl_factor(void)
147 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
150 switch (sysctl_sched_tunable_scaling) {
151 case SCHED_TUNABLESCALING_NONE:
154 case SCHED_TUNABLESCALING_LINEAR:
157 case SCHED_TUNABLESCALING_LOG:
159 factor = 1 + ilog2(cpus);
166 static void update_sysctl(void)
168 unsigned int factor = get_update_sysctl_factor();
170 #define SET_SYSCTL(name) \
171 (sysctl_##name = (factor) * normalized_sysctl_##name)
172 SET_SYSCTL(sched_min_granularity);
173 SET_SYSCTL(sched_latency);
174 SET_SYSCTL(sched_wakeup_granularity);
178 void sched_init_granularity(void)
183 #define WMULT_CONST (~0U)
184 #define WMULT_SHIFT 32
186 static void __update_inv_weight(struct load_weight *lw)
190 if (likely(lw->inv_weight))
193 w = scale_load_down(lw->weight);
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
200 lw->inv_weight = WMULT_CONST / w;
204 * delta_exec * weight / lw.weight
206 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
208 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
209 * we're guaranteed shift stays positive because inv_weight is guaranteed to
210 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
212 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
213 * weight/lw.weight <= 1, and therefore our shift will also be positive.
215 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
217 u64 fact = scale_load_down(weight);
218 int shift = WMULT_SHIFT;
220 __update_inv_weight(lw);
222 if (unlikely(fact >> 32)) {
229 /* hint to use a 32x32->64 mul */
230 fact = (u64)(u32)fact * lw->inv_weight;
237 return mul_u64_u32_shr(delta_exec, fact, shift);
241 const struct sched_class fair_sched_class;
243 /**************************************************************
244 * CFS operations on generic schedulable entities:
247 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* cpu runqueue to which this cfs_rq is attached */
250 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
255 /* An entity is a task if it doesn't "own" a runqueue */
256 #define entity_is_task(se) (!se->my_q)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 #ifdef CONFIG_SCHED_DEBUG
261 WARN_ON_ONCE(!entity_is_task(se));
263 return container_of(se, struct task_struct, se);
266 /* Walk up scheduling entities hierarchy */
267 #define for_each_sched_entity(se) \
268 for (; se; se = se->parent)
270 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
275 /* runqueue on which this entity is (to be) queued */
276 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
281 /* runqueue "owned" by this group */
282 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
287 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 if (!cfs_rq->on_list) {
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases.
296 if (cfs_rq->tg->parent &&
297 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
298 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
299 &rq_of(cfs_rq)->leaf_cfs_rq_list);
301 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
302 &rq_of(cfs_rq)->leaf_cfs_rq_list);
309 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
311 if (cfs_rq->on_list) {
312 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
317 /* Iterate thr' all leaf cfs_rq's on a runqueue */
318 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
319 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
321 /* Do the two (enqueued) entities belong to the same group ? */
322 static inline struct cfs_rq *
323 is_same_group(struct sched_entity *se, struct sched_entity *pse)
325 if (se->cfs_rq == pse->cfs_rq)
331 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
339 int se_depth, pse_depth;
342 * preemption test can be made between sibling entities who are in the
343 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
344 * both tasks until we find their ancestors who are siblings of common
348 /* First walk up until both entities are at same depth */
349 se_depth = (*se)->depth;
350 pse_depth = (*pse)->depth;
352 while (se_depth > pse_depth) {
354 *se = parent_entity(*se);
357 while (pse_depth > se_depth) {
359 *pse = parent_entity(*pse);
362 while (!is_same_group(*se, *pse)) {
363 *se = parent_entity(*se);
364 *pse = parent_entity(*pse);
368 #else /* !CONFIG_FAIR_GROUP_SCHED */
370 static inline struct task_struct *task_of(struct sched_entity *se)
372 return container_of(se, struct task_struct, se);
375 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
377 return container_of(cfs_rq, struct rq, cfs);
380 #define entity_is_task(se) 1
382 #define for_each_sched_entity(se) \
383 for (; se; se = NULL)
385 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
387 return &task_rq(p)->cfs;
390 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
392 struct task_struct *p = task_of(se);
393 struct rq *rq = task_rq(p);
398 /* runqueue "owned" by this group */
399 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
404 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
413 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
415 static inline struct sched_entity *parent_entity(struct sched_entity *se)
421 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 #endif /* CONFIG_FAIR_GROUP_SCHED */
427 static __always_inline
428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
430 /**************************************************************
431 * Scheduling class tree data structure manipulation methods:
434 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - max_vruntime);
438 max_vruntime = vruntime;
443 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - min_vruntime);
447 min_vruntime = vruntime;
452 static inline int entity_before(struct sched_entity *a,
453 struct sched_entity *b)
455 return (s64)(a->vruntime - b->vruntime) < 0;
458 static void update_min_vruntime(struct cfs_rq *cfs_rq)
460 u64 vruntime = cfs_rq->min_vruntime;
463 vruntime = cfs_rq->curr->vruntime;
465 if (cfs_rq->rb_leftmost) {
466 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 vruntime = se->vruntime;
473 vruntime = min_vruntime(vruntime, se->vruntime);
476 /* ensure we never gain time by being placed backwards. */
477 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 * Enqueue an entity into the rb-tree:
487 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
490 struct rb_node *parent = NULL;
491 struct sched_entity *entry;
495 * Find the right place in the rbtree:
499 entry = rb_entry(parent, struct sched_entity, run_node);
501 * We dont care about collisions. Nodes with
502 * the same key stay together.
504 if (entity_before(se, entry)) {
505 link = &parent->rb_left;
507 link = &parent->rb_right;
513 * Maintain a cache of leftmost tree entries (it is frequently
517 cfs_rq->rb_leftmost = &se->run_node;
519 rb_link_node(&se->run_node, parent, link);
520 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
523 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 if (cfs_rq->rb_leftmost == &se->run_node) {
526 struct rb_node *next_node;
528 next_node = rb_next(&se->run_node);
529 cfs_rq->rb_leftmost = next_node;
532 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
535 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
537 struct rb_node *left = cfs_rq->rb_leftmost;
542 return rb_entry(left, struct sched_entity, run_node);
545 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
547 struct rb_node *next = rb_next(&se->run_node);
552 return rb_entry(next, struct sched_entity, run_node);
555 #ifdef CONFIG_SCHED_DEBUG
556 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 return rb_entry(last, struct sched_entity, run_node);
566 /**************************************************************
567 * Scheduling class statistics methods:
570 int sched_proc_update_handler(struct ctl_table *table, int write,
571 void __user *buffer, size_t *lenp,
574 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
575 unsigned int factor = get_update_sysctl_factor();
580 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
581 sysctl_sched_min_granularity);
583 #define WRT_SYSCTL(name) \
584 (normalized_sysctl_##name = sysctl_##name / (factor))
585 WRT_SYSCTL(sched_min_granularity);
586 WRT_SYSCTL(sched_latency);
587 WRT_SYSCTL(sched_wakeup_granularity);
597 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
599 if (unlikely(se->load.weight != NICE_0_LOAD))
600 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
606 * The idea is to set a period in which each task runs once.
608 * When there are too many tasks (sched_nr_latency) we have to stretch
609 * this period because otherwise the slices get too small.
611 * p = (nr <= nl) ? l : l*nr/nl
613 static u64 __sched_period(unsigned long nr_running)
615 if (unlikely(nr_running > sched_nr_latency))
616 return nr_running * sysctl_sched_min_granularity;
618 return sysctl_sched_latency;
622 * We calculate the wall-time slice from the period by taking a part
623 * proportional to the weight.
627 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
631 for_each_sched_entity(se) {
632 struct load_weight *load;
633 struct load_weight lw;
635 cfs_rq = cfs_rq_of(se);
636 load = &cfs_rq->load;
638 if (unlikely(!se->on_rq)) {
641 update_load_add(&lw, se->load.weight);
644 slice = __calc_delta(slice, se->load.weight, load);
650 * We calculate the vruntime slice of a to-be-inserted task.
654 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 return calc_delta_fair(sched_slice(cfs_rq, se), se);
660 static int select_idle_sibling(struct task_struct *p, int cpu);
661 static unsigned long task_h_load(struct task_struct *p);
664 * We choose a half-life close to 1 scheduling period.
665 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
666 * dependent on this value.
668 #define LOAD_AVG_PERIOD 32
669 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
670 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
672 /* Give new sched_entity start runnable values to heavy its load in infant time */
673 void init_entity_runnable_average(struct sched_entity *se)
675 struct sched_avg *sa = &se->avg;
677 sa->last_update_time = 0;
679 * sched_avg's period_contrib should be strictly less then 1024, so
680 * we give it 1023 to make sure it is almost a period (1024us), and
681 * will definitely be update (after enqueue).
683 sa->period_contrib = 1023;
684 sa->load_avg = scale_load_down(se->load.weight);
685 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
687 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
688 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
692 void init_entity_runnable_average(struct sched_entity *se)
698 * Update the current task's runtime statistics.
700 static void update_curr(struct cfs_rq *cfs_rq)
702 struct sched_entity *curr = cfs_rq->curr;
703 u64 now = rq_clock_task(rq_of(cfs_rq));
709 delta_exec = now - curr->exec_start;
710 if (unlikely((s64)delta_exec <= 0))
713 curr->exec_start = now;
715 schedstat_set(curr->statistics.exec_max,
716 max(delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
721 curr->vruntime += calc_delta_fair(delta_exec, curr);
722 update_min_vruntime(cfs_rq);
724 if (entity_is_task(curr)) {
725 struct task_struct *curtask = task_of(curr);
727 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
728 cpuacct_charge(curtask, delta_exec);
729 account_group_exec_runtime(curtask, delta_exec);
732 account_cfs_rq_runtime(cfs_rq, delta_exec);
735 static void update_curr_fair(struct rq *rq)
737 update_curr(cfs_rq_of(&rq->curr->se));
741 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
743 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
747 * Task is being enqueued - update stats:
749 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
752 * Are we enqueueing a waiting task? (for current tasks
753 * a dequeue/enqueue event is a NOP)
755 if (se != cfs_rq->curr)
756 update_stats_wait_start(cfs_rq, se);
760 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
762 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
763 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
764 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
765 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
766 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
767 #ifdef CONFIG_SCHEDSTATS
768 if (entity_is_task(se)) {
769 trace_sched_stat_wait(task_of(se),
770 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
773 schedstat_set(se->statistics.wait_start, 0);
777 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 * Mark the end of the wait period if dequeueing a
783 if (se != cfs_rq->curr)
784 update_stats_wait_end(cfs_rq, se);
788 * We are picking a new current task - update its stats:
791 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
794 * We are starting a new run period:
796 se->exec_start = rq_clock_task(rq_of(cfs_rq));
799 /**************************************************
800 * Scheduling class queueing methods:
803 #ifdef CONFIG_NUMA_BALANCING
805 * Approximate time to scan a full NUMA task in ms. The task scan period is
806 * calculated based on the tasks virtual memory size and
807 * numa_balancing_scan_size.
809 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
810 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
812 /* Portion of address space to scan in MB */
813 unsigned int sysctl_numa_balancing_scan_size = 256;
815 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
816 unsigned int sysctl_numa_balancing_scan_delay = 1000;
818 static unsigned int task_nr_scan_windows(struct task_struct *p)
820 unsigned long rss = 0;
821 unsigned long nr_scan_pages;
824 * Calculations based on RSS as non-present and empty pages are skipped
825 * by the PTE scanner and NUMA hinting faults should be trapped based
828 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
829 rss = get_mm_rss(p->mm);
833 rss = round_up(rss, nr_scan_pages);
834 return rss / nr_scan_pages;
837 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
838 #define MAX_SCAN_WINDOW 2560
840 static unsigned int task_scan_min(struct task_struct *p)
842 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
843 unsigned int scan, floor;
844 unsigned int windows = 1;
846 if (scan_size < MAX_SCAN_WINDOW)
847 windows = MAX_SCAN_WINDOW / scan_size;
848 floor = 1000 / windows;
850 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
851 return max_t(unsigned int, floor, scan);
854 static unsigned int task_scan_max(struct task_struct *p)
856 unsigned int smin = task_scan_min(p);
859 /* Watch for min being lower than max due to floor calculations */
860 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
861 return max(smin, smax);
864 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
866 rq->nr_numa_running += (p->numa_preferred_nid != -1);
867 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
870 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
872 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
873 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
879 spinlock_t lock; /* nr_tasks, tasks */
884 nodemask_t active_nodes;
885 unsigned long total_faults;
887 * Faults_cpu is used to decide whether memory should move
888 * towards the CPU. As a consequence, these stats are weighted
889 * more by CPU use than by memory faults.
891 unsigned long *faults_cpu;
892 unsigned long faults[0];
895 /* Shared or private faults. */
896 #define NR_NUMA_HINT_FAULT_TYPES 2
898 /* Memory and CPU locality */
899 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
901 /* Averaged statistics, and temporary buffers. */
902 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
904 pid_t task_numa_group_id(struct task_struct *p)
906 return p->numa_group ? p->numa_group->gid : 0;
910 * The averaged statistics, shared & private, memory & cpu,
911 * occupy the first half of the array. The second half of the
912 * array is for current counters, which are averaged into the
913 * first set by task_numa_placement.
915 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
917 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
920 static inline unsigned long task_faults(struct task_struct *p, int nid)
925 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
926 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
929 static inline unsigned long group_faults(struct task_struct *p, int nid)
934 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
935 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
938 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
940 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
941 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
944 /* Handle placement on systems where not all nodes are directly connected. */
945 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
946 int maxdist, bool task)
948 unsigned long score = 0;
952 * All nodes are directly connected, and the same distance
953 * from each other. No need for fancy placement algorithms.
955 if (sched_numa_topology_type == NUMA_DIRECT)
959 * This code is called for each node, introducing N^2 complexity,
960 * which should be ok given the number of nodes rarely exceeds 8.
962 for_each_online_node(node) {
963 unsigned long faults;
964 int dist = node_distance(nid, node);
967 * The furthest away nodes in the system are not interesting
968 * for placement; nid was already counted.
970 if (dist == sched_max_numa_distance || node == nid)
974 * On systems with a backplane NUMA topology, compare groups
975 * of nodes, and move tasks towards the group with the most
976 * memory accesses. When comparing two nodes at distance
977 * "hoplimit", only nodes closer by than "hoplimit" are part
978 * of each group. Skip other nodes.
980 if (sched_numa_topology_type == NUMA_BACKPLANE &&
984 /* Add up the faults from nearby nodes. */
986 faults = task_faults(p, node);
988 faults = group_faults(p, node);
991 * On systems with a glueless mesh NUMA topology, there are
992 * no fixed "groups of nodes". Instead, nodes that are not
993 * directly connected bounce traffic through intermediate
994 * nodes; a numa_group can occupy any set of nodes.
995 * The further away a node is, the less the faults count.
996 * This seems to result in good task placement.
998 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
999 faults *= (sched_max_numa_distance - dist);
1000 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1010 * These return the fraction of accesses done by a particular task, or
1011 * task group, on a particular numa node. The group weight is given a
1012 * larger multiplier, in order to group tasks together that are almost
1013 * evenly spread out between numa nodes.
1015 static inline unsigned long task_weight(struct task_struct *p, int nid,
1018 unsigned long faults, total_faults;
1020 if (!p->numa_faults)
1023 total_faults = p->total_numa_faults;
1028 faults = task_faults(p, nid);
1029 faults += score_nearby_nodes(p, nid, dist, true);
1031 return 1000 * faults / total_faults;
1034 static inline unsigned long group_weight(struct task_struct *p, int nid,
1037 unsigned long faults, total_faults;
1042 total_faults = p->numa_group->total_faults;
1047 faults = group_faults(p, nid);
1048 faults += score_nearby_nodes(p, nid, dist, false);
1050 return 1000 * faults / total_faults;
1053 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1054 int src_nid, int dst_cpu)
1056 struct numa_group *ng = p->numa_group;
1057 int dst_nid = cpu_to_node(dst_cpu);
1058 int last_cpupid, this_cpupid;
1060 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1063 * Multi-stage node selection is used in conjunction with a periodic
1064 * migration fault to build a temporal task<->page relation. By using
1065 * a two-stage filter we remove short/unlikely relations.
1067 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1068 * a task's usage of a particular page (n_p) per total usage of this
1069 * page (n_t) (in a given time-span) to a probability.
1071 * Our periodic faults will sample this probability and getting the
1072 * same result twice in a row, given these samples are fully
1073 * independent, is then given by P(n)^2, provided our sample period
1074 * is sufficiently short compared to the usage pattern.
1076 * This quadric squishes small probabilities, making it less likely we
1077 * act on an unlikely task<->page relation.
1079 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1080 if (!cpupid_pid_unset(last_cpupid) &&
1081 cpupid_to_nid(last_cpupid) != dst_nid)
1084 /* Always allow migrate on private faults */
1085 if (cpupid_match_pid(p, last_cpupid))
1088 /* A shared fault, but p->numa_group has not been set up yet. */
1093 * Do not migrate if the destination is not a node that
1094 * is actively used by this numa group.
1096 if (!node_isset(dst_nid, ng->active_nodes))
1100 * Source is a node that is not actively used by this
1101 * numa group, while the destination is. Migrate.
1103 if (!node_isset(src_nid, ng->active_nodes))
1107 * Both source and destination are nodes in active
1108 * use by this numa group. Maximize memory bandwidth
1109 * by migrating from more heavily used groups, to less
1110 * heavily used ones, spreading the load around.
1111 * Use a 1/4 hysteresis to avoid spurious page movement.
1113 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1116 static unsigned long weighted_cpuload(const int cpu);
1117 static unsigned long source_load(int cpu, int type);
1118 static unsigned long target_load(int cpu, int type);
1119 static unsigned long capacity_of(int cpu);
1120 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1122 /* Cached statistics for all CPUs within a node */
1124 unsigned long nr_running;
1127 /* Total compute capacity of CPUs on a node */
1128 unsigned long compute_capacity;
1130 /* Approximate capacity in terms of runnable tasks on a node */
1131 unsigned long task_capacity;
1132 int has_free_capacity;
1136 * XXX borrowed from update_sg_lb_stats
1138 static void update_numa_stats(struct numa_stats *ns, int nid)
1140 int smt, cpu, cpus = 0;
1141 unsigned long capacity;
1143 memset(ns, 0, sizeof(*ns));
1144 for_each_cpu(cpu, cpumask_of_node(nid)) {
1145 struct rq *rq = cpu_rq(cpu);
1147 ns->nr_running += rq->nr_running;
1148 ns->load += weighted_cpuload(cpu);
1149 ns->compute_capacity += capacity_of(cpu);
1155 * If we raced with hotplug and there are no CPUs left in our mask
1156 * the @ns structure is NULL'ed and task_numa_compare() will
1157 * not find this node attractive.
1159 * We'll either bail at !has_free_capacity, or we'll detect a huge
1160 * imbalance and bail there.
1165 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1166 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1167 capacity = cpus / smt; /* cores */
1169 ns->task_capacity = min_t(unsigned, capacity,
1170 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1171 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1174 struct task_numa_env {
1175 struct task_struct *p;
1177 int src_cpu, src_nid;
1178 int dst_cpu, dst_nid;
1180 struct numa_stats src_stats, dst_stats;
1185 struct task_struct *best_task;
1190 static void task_numa_assign(struct task_numa_env *env,
1191 struct task_struct *p, long imp)
1194 put_task_struct(env->best_task);
1199 env->best_imp = imp;
1200 env->best_cpu = env->dst_cpu;
1203 static bool load_too_imbalanced(long src_load, long dst_load,
1204 struct task_numa_env *env)
1207 long orig_src_load, orig_dst_load;
1208 long src_capacity, dst_capacity;
1211 * The load is corrected for the CPU capacity available on each node.
1214 * ------------ vs ---------
1215 * src_capacity dst_capacity
1217 src_capacity = env->src_stats.compute_capacity;
1218 dst_capacity = env->dst_stats.compute_capacity;
1220 /* We care about the slope of the imbalance, not the direction. */
1221 if (dst_load < src_load)
1222 swap(dst_load, src_load);
1224 /* Is the difference below the threshold? */
1225 imb = dst_load * src_capacity * 100 -
1226 src_load * dst_capacity * env->imbalance_pct;
1231 * The imbalance is above the allowed threshold.
1232 * Compare it with the old imbalance.
1234 orig_src_load = env->src_stats.load;
1235 orig_dst_load = env->dst_stats.load;
1237 if (orig_dst_load < orig_src_load)
1238 swap(orig_dst_load, orig_src_load);
1240 old_imb = orig_dst_load * src_capacity * 100 -
1241 orig_src_load * dst_capacity * env->imbalance_pct;
1243 /* Would this change make things worse? */
1244 return (imb > old_imb);
1248 * This checks if the overall compute and NUMA accesses of the system would
1249 * be improved if the source tasks was migrated to the target dst_cpu taking
1250 * into account that it might be best if task running on the dst_cpu should
1251 * be exchanged with the source task
1253 static void task_numa_compare(struct task_numa_env *env,
1254 long taskimp, long groupimp)
1256 struct rq *src_rq = cpu_rq(env->src_cpu);
1257 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1258 struct task_struct *cur;
1259 long src_load, dst_load;
1261 long imp = env->p->numa_group ? groupimp : taskimp;
1263 int dist = env->dist;
1267 raw_spin_lock_irq(&dst_rq->lock);
1270 * No need to move the exiting task, and this ensures that ->curr
1271 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1272 * is safe under RCU read lock.
1273 * Note that rcu_read_lock() itself can't protect from the final
1274 * put_task_struct() after the last schedule().
1276 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1278 raw_spin_unlock_irq(&dst_rq->lock);
1281 * Because we have preemption enabled we can get migrated around and
1282 * end try selecting ourselves (current == env->p) as a swap candidate.
1288 * "imp" is the fault differential for the source task between the
1289 * source and destination node. Calculate the total differential for
1290 * the source task and potential destination task. The more negative
1291 * the value is, the more rmeote accesses that would be expected to
1292 * be incurred if the tasks were swapped.
1295 /* Skip this swap candidate if cannot move to the source cpu */
1296 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1300 * If dst and source tasks are in the same NUMA group, or not
1301 * in any group then look only at task weights.
1303 if (cur->numa_group == env->p->numa_group) {
1304 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1305 task_weight(cur, env->dst_nid, dist);
1307 * Add some hysteresis to prevent swapping the
1308 * tasks within a group over tiny differences.
1310 if (cur->numa_group)
1314 * Compare the group weights. If a task is all by
1315 * itself (not part of a group), use the task weight
1318 if (cur->numa_group)
1319 imp += group_weight(cur, env->src_nid, dist) -
1320 group_weight(cur, env->dst_nid, dist);
1322 imp += task_weight(cur, env->src_nid, dist) -
1323 task_weight(cur, env->dst_nid, dist);
1327 if (imp <= env->best_imp && moveimp <= env->best_imp)
1331 /* Is there capacity at our destination? */
1332 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1333 !env->dst_stats.has_free_capacity)
1339 /* Balance doesn't matter much if we're running a task per cpu */
1340 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1341 dst_rq->nr_running == 1)
1345 * In the overloaded case, try and keep the load balanced.
1348 load = task_h_load(env->p);
1349 dst_load = env->dst_stats.load + load;
1350 src_load = env->src_stats.load - load;
1352 if (moveimp > imp && moveimp > env->best_imp) {
1354 * If the improvement from just moving env->p direction is
1355 * better than swapping tasks around, check if a move is
1356 * possible. Store a slightly smaller score than moveimp,
1357 * so an actually idle CPU will win.
1359 if (!load_too_imbalanced(src_load, dst_load, env)) {
1366 if (imp <= env->best_imp)
1370 load = task_h_load(cur);
1375 if (load_too_imbalanced(src_load, dst_load, env))
1379 * One idle CPU per node is evaluated for a task numa move.
1380 * Call select_idle_sibling to maybe find a better one.
1383 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1386 task_numa_assign(env, cur, imp);
1391 static void task_numa_find_cpu(struct task_numa_env *env,
1392 long taskimp, long groupimp)
1396 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1397 /* Skip this CPU if the source task cannot migrate */
1398 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1402 task_numa_compare(env, taskimp, groupimp);
1406 /* Only move tasks to a NUMA node less busy than the current node. */
1407 static bool numa_has_capacity(struct task_numa_env *env)
1409 struct numa_stats *src = &env->src_stats;
1410 struct numa_stats *dst = &env->dst_stats;
1412 if (src->has_free_capacity && !dst->has_free_capacity)
1416 * Only consider a task move if the source has a higher load
1417 * than the destination, corrected for CPU capacity on each node.
1419 * src->load dst->load
1420 * --------------------- vs ---------------------
1421 * src->compute_capacity dst->compute_capacity
1423 if (src->load * dst->compute_capacity * env->imbalance_pct >
1425 dst->load * src->compute_capacity * 100)
1431 static int task_numa_migrate(struct task_struct *p)
1433 struct task_numa_env env = {
1436 .src_cpu = task_cpu(p),
1437 .src_nid = task_node(p),
1439 .imbalance_pct = 112,
1445 struct sched_domain *sd;
1446 unsigned long taskweight, groupweight;
1448 long taskimp, groupimp;
1451 * Pick the lowest SD_NUMA domain, as that would have the smallest
1452 * imbalance and would be the first to start moving tasks about.
1454 * And we want to avoid any moving of tasks about, as that would create
1455 * random movement of tasks -- counter the numa conditions we're trying
1459 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1461 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1465 * Cpusets can break the scheduler domain tree into smaller
1466 * balance domains, some of which do not cross NUMA boundaries.
1467 * Tasks that are "trapped" in such domains cannot be migrated
1468 * elsewhere, so there is no point in (re)trying.
1470 if (unlikely(!sd)) {
1471 p->numa_preferred_nid = task_node(p);
1475 env.dst_nid = p->numa_preferred_nid;
1476 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1477 taskweight = task_weight(p, env.src_nid, dist);
1478 groupweight = group_weight(p, env.src_nid, dist);
1479 update_numa_stats(&env.src_stats, env.src_nid);
1480 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1481 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1482 update_numa_stats(&env.dst_stats, env.dst_nid);
1484 /* Try to find a spot on the preferred nid. */
1485 if (numa_has_capacity(&env))
1486 task_numa_find_cpu(&env, taskimp, groupimp);
1489 * Look at other nodes in these cases:
1490 * - there is no space available on the preferred_nid
1491 * - the task is part of a numa_group that is interleaved across
1492 * multiple NUMA nodes; in order to better consolidate the group,
1493 * we need to check other locations.
1495 if (env.best_cpu == -1 || (p->numa_group &&
1496 nodes_weight(p->numa_group->active_nodes) > 1)) {
1497 for_each_online_node(nid) {
1498 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1501 dist = node_distance(env.src_nid, env.dst_nid);
1502 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1504 taskweight = task_weight(p, env.src_nid, dist);
1505 groupweight = group_weight(p, env.src_nid, dist);
1508 /* Only consider nodes where both task and groups benefit */
1509 taskimp = task_weight(p, nid, dist) - taskweight;
1510 groupimp = group_weight(p, nid, dist) - groupweight;
1511 if (taskimp < 0 && groupimp < 0)
1516 update_numa_stats(&env.dst_stats, env.dst_nid);
1517 if (numa_has_capacity(&env))
1518 task_numa_find_cpu(&env, taskimp, groupimp);
1523 * If the task is part of a workload that spans multiple NUMA nodes,
1524 * and is migrating into one of the workload's active nodes, remember
1525 * this node as the task's preferred numa node, so the workload can
1527 * A task that migrated to a second choice node will be better off
1528 * trying for a better one later. Do not set the preferred node here.
1530 if (p->numa_group) {
1531 if (env.best_cpu == -1)
1536 if (node_isset(nid, p->numa_group->active_nodes))
1537 sched_setnuma(p, env.dst_nid);
1540 /* No better CPU than the current one was found. */
1541 if (env.best_cpu == -1)
1545 * Reset the scan period if the task is being rescheduled on an
1546 * alternative node to recheck if the tasks is now properly placed.
1548 p->numa_scan_period = task_scan_min(p);
1550 if (env.best_task == NULL) {
1551 ret = migrate_task_to(p, env.best_cpu);
1553 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1557 ret = migrate_swap(p, env.best_task);
1559 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1560 put_task_struct(env.best_task);
1564 /* Attempt to migrate a task to a CPU on the preferred node. */
1565 static void numa_migrate_preferred(struct task_struct *p)
1567 unsigned long interval = HZ;
1569 /* This task has no NUMA fault statistics yet */
1570 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1573 /* Periodically retry migrating the task to the preferred node */
1574 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1575 p->numa_migrate_retry = jiffies + interval;
1577 /* Success if task is already running on preferred CPU */
1578 if (task_node(p) == p->numa_preferred_nid)
1581 /* Otherwise, try migrate to a CPU on the preferred node */
1582 task_numa_migrate(p);
1586 * Find the nodes on which the workload is actively running. We do this by
1587 * tracking the nodes from which NUMA hinting faults are triggered. This can
1588 * be different from the set of nodes where the workload's memory is currently
1591 * The bitmask is used to make smarter decisions on when to do NUMA page
1592 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1593 * are added when they cause over 6/16 of the maximum number of faults, but
1594 * only removed when they drop below 3/16.
1596 static void update_numa_active_node_mask(struct numa_group *numa_group)
1598 unsigned long faults, max_faults = 0;
1601 for_each_online_node(nid) {
1602 faults = group_faults_cpu(numa_group, nid);
1603 if (faults > max_faults)
1604 max_faults = faults;
1607 for_each_online_node(nid) {
1608 faults = group_faults_cpu(numa_group, nid);
1609 if (!node_isset(nid, numa_group->active_nodes)) {
1610 if (faults > max_faults * 6 / 16)
1611 node_set(nid, numa_group->active_nodes);
1612 } else if (faults < max_faults * 3 / 16)
1613 node_clear(nid, numa_group->active_nodes);
1618 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1619 * increments. The more local the fault statistics are, the higher the scan
1620 * period will be for the next scan window. If local/(local+remote) ratio is
1621 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1622 * the scan period will decrease. Aim for 70% local accesses.
1624 #define NUMA_PERIOD_SLOTS 10
1625 #define NUMA_PERIOD_THRESHOLD 7
1628 * Increase the scan period (slow down scanning) if the majority of
1629 * our memory is already on our local node, or if the majority of
1630 * the page accesses are shared with other processes.
1631 * Otherwise, decrease the scan period.
1633 static void update_task_scan_period(struct task_struct *p,
1634 unsigned long shared, unsigned long private)
1636 unsigned int period_slot;
1640 unsigned long remote = p->numa_faults_locality[0];
1641 unsigned long local = p->numa_faults_locality[1];
1644 * If there were no record hinting faults then either the task is
1645 * completely idle or all activity is areas that are not of interest
1646 * to automatic numa balancing. Related to that, if there were failed
1647 * migration then it implies we are migrating too quickly or the local
1648 * node is overloaded. In either case, scan slower
1650 if (local + shared == 0 || p->numa_faults_locality[2]) {
1651 p->numa_scan_period = min(p->numa_scan_period_max,
1652 p->numa_scan_period << 1);
1654 p->mm->numa_next_scan = jiffies +
1655 msecs_to_jiffies(p->numa_scan_period);
1661 * Prepare to scale scan period relative to the current period.
1662 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1663 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1664 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1666 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1667 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1668 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1669 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1672 diff = slot * period_slot;
1674 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1677 * Scale scan rate increases based on sharing. There is an
1678 * inverse relationship between the degree of sharing and
1679 * the adjustment made to the scanning period. Broadly
1680 * speaking the intent is that there is little point
1681 * scanning faster if shared accesses dominate as it may
1682 * simply bounce migrations uselessly
1684 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1685 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1688 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1689 task_scan_min(p), task_scan_max(p));
1690 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1694 * Get the fraction of time the task has been running since the last
1695 * NUMA placement cycle. The scheduler keeps similar statistics, but
1696 * decays those on a 32ms period, which is orders of magnitude off
1697 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1698 * stats only if the task is so new there are no NUMA statistics yet.
1700 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1702 u64 runtime, delta, now;
1703 /* Use the start of this time slice to avoid calculations. */
1704 now = p->se.exec_start;
1705 runtime = p->se.sum_exec_runtime;
1707 if (p->last_task_numa_placement) {
1708 delta = runtime - p->last_sum_exec_runtime;
1709 *period = now - p->last_task_numa_placement;
1711 delta = p->se.avg.load_sum / p->se.load.weight;
1712 *period = LOAD_AVG_MAX;
1715 p->last_sum_exec_runtime = runtime;
1716 p->last_task_numa_placement = now;
1722 * Determine the preferred nid for a task in a numa_group. This needs to
1723 * be done in a way that produces consistent results with group_weight,
1724 * otherwise workloads might not converge.
1726 static int preferred_group_nid(struct task_struct *p, int nid)
1731 /* Direct connections between all NUMA nodes. */
1732 if (sched_numa_topology_type == NUMA_DIRECT)
1736 * On a system with glueless mesh NUMA topology, group_weight
1737 * scores nodes according to the number of NUMA hinting faults on
1738 * both the node itself, and on nearby nodes.
1740 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1741 unsigned long score, max_score = 0;
1742 int node, max_node = nid;
1744 dist = sched_max_numa_distance;
1746 for_each_online_node(node) {
1747 score = group_weight(p, node, dist);
1748 if (score > max_score) {
1757 * Finding the preferred nid in a system with NUMA backplane
1758 * interconnect topology is more involved. The goal is to locate
1759 * tasks from numa_groups near each other in the system, and
1760 * untangle workloads from different sides of the system. This requires
1761 * searching down the hierarchy of node groups, recursively searching
1762 * inside the highest scoring group of nodes. The nodemask tricks
1763 * keep the complexity of the search down.
1765 nodes = node_online_map;
1766 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1767 unsigned long max_faults = 0;
1768 nodemask_t max_group = NODE_MASK_NONE;
1771 /* Are there nodes at this distance from each other? */
1772 if (!find_numa_distance(dist))
1775 for_each_node_mask(a, nodes) {
1776 unsigned long faults = 0;
1777 nodemask_t this_group;
1778 nodes_clear(this_group);
1780 /* Sum group's NUMA faults; includes a==b case. */
1781 for_each_node_mask(b, nodes) {
1782 if (node_distance(a, b) < dist) {
1783 faults += group_faults(p, b);
1784 node_set(b, this_group);
1785 node_clear(b, nodes);
1789 /* Remember the top group. */
1790 if (faults > max_faults) {
1791 max_faults = faults;
1792 max_group = this_group;
1794 * subtle: at the smallest distance there is
1795 * just one node left in each "group", the
1796 * winner is the preferred nid.
1801 /* Next round, evaluate the nodes within max_group. */
1809 static void task_numa_placement(struct task_struct *p)
1811 int seq, nid, max_nid = -1, max_group_nid = -1;
1812 unsigned long max_faults = 0, max_group_faults = 0;
1813 unsigned long fault_types[2] = { 0, 0 };
1814 unsigned long total_faults;
1815 u64 runtime, period;
1816 spinlock_t *group_lock = NULL;
1819 * The p->mm->numa_scan_seq field gets updated without
1820 * exclusive access. Use READ_ONCE() here to ensure
1821 * that the field is read in a single access:
1823 seq = READ_ONCE(p->mm->numa_scan_seq);
1824 if (p->numa_scan_seq == seq)
1826 p->numa_scan_seq = seq;
1827 p->numa_scan_period_max = task_scan_max(p);
1829 total_faults = p->numa_faults_locality[0] +
1830 p->numa_faults_locality[1];
1831 runtime = numa_get_avg_runtime(p, &period);
1833 /* If the task is part of a group prevent parallel updates to group stats */
1834 if (p->numa_group) {
1835 group_lock = &p->numa_group->lock;
1836 spin_lock_irq(group_lock);
1839 /* Find the node with the highest number of faults */
1840 for_each_online_node(nid) {
1841 /* Keep track of the offsets in numa_faults array */
1842 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1843 unsigned long faults = 0, group_faults = 0;
1846 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1847 long diff, f_diff, f_weight;
1849 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1850 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1851 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1852 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1854 /* Decay existing window, copy faults since last scan */
1855 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1856 fault_types[priv] += p->numa_faults[membuf_idx];
1857 p->numa_faults[membuf_idx] = 0;
1860 * Normalize the faults_from, so all tasks in a group
1861 * count according to CPU use, instead of by the raw
1862 * number of faults. Tasks with little runtime have
1863 * little over-all impact on throughput, and thus their
1864 * faults are less important.
1866 f_weight = div64_u64(runtime << 16, period + 1);
1867 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1869 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1870 p->numa_faults[cpubuf_idx] = 0;
1872 p->numa_faults[mem_idx] += diff;
1873 p->numa_faults[cpu_idx] += f_diff;
1874 faults += p->numa_faults[mem_idx];
1875 p->total_numa_faults += diff;
1876 if (p->numa_group) {
1878 * safe because we can only change our own group
1880 * mem_idx represents the offset for a given
1881 * nid and priv in a specific region because it
1882 * is at the beginning of the numa_faults array.
1884 p->numa_group->faults[mem_idx] += diff;
1885 p->numa_group->faults_cpu[mem_idx] += f_diff;
1886 p->numa_group->total_faults += diff;
1887 group_faults += p->numa_group->faults[mem_idx];
1891 if (faults > max_faults) {
1892 max_faults = faults;
1896 if (group_faults > max_group_faults) {
1897 max_group_faults = group_faults;
1898 max_group_nid = nid;
1902 update_task_scan_period(p, fault_types[0], fault_types[1]);
1904 if (p->numa_group) {
1905 update_numa_active_node_mask(p->numa_group);
1906 spin_unlock_irq(group_lock);
1907 max_nid = preferred_group_nid(p, max_group_nid);
1911 /* Set the new preferred node */
1912 if (max_nid != p->numa_preferred_nid)
1913 sched_setnuma(p, max_nid);
1915 if (task_node(p) != p->numa_preferred_nid)
1916 numa_migrate_preferred(p);
1920 static inline int get_numa_group(struct numa_group *grp)
1922 return atomic_inc_not_zero(&grp->refcount);
1925 static inline void put_numa_group(struct numa_group *grp)
1927 if (atomic_dec_and_test(&grp->refcount))
1928 kfree_rcu(grp, rcu);
1931 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1934 struct numa_group *grp, *my_grp;
1935 struct task_struct *tsk;
1937 int cpu = cpupid_to_cpu(cpupid);
1940 if (unlikely(!p->numa_group)) {
1941 unsigned int size = sizeof(struct numa_group) +
1942 4*nr_node_ids*sizeof(unsigned long);
1944 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1948 atomic_set(&grp->refcount, 1);
1949 spin_lock_init(&grp->lock);
1951 /* Second half of the array tracks nids where faults happen */
1952 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1955 node_set(task_node(current), grp->active_nodes);
1957 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1958 grp->faults[i] = p->numa_faults[i];
1960 grp->total_faults = p->total_numa_faults;
1963 rcu_assign_pointer(p->numa_group, grp);
1967 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1969 if (!cpupid_match_pid(tsk, cpupid))
1972 grp = rcu_dereference(tsk->numa_group);
1976 my_grp = p->numa_group;
1981 * Only join the other group if its bigger; if we're the bigger group,
1982 * the other task will join us.
1984 if (my_grp->nr_tasks > grp->nr_tasks)
1988 * Tie-break on the grp address.
1990 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1993 /* Always join threads in the same process. */
1994 if (tsk->mm == current->mm)
1997 /* Simple filter to avoid false positives due to PID collisions */
1998 if (flags & TNF_SHARED)
2001 /* Update priv based on whether false sharing was detected */
2004 if (join && !get_numa_group(grp))
2012 BUG_ON(irqs_disabled());
2013 double_lock_irq(&my_grp->lock, &grp->lock);
2015 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2016 my_grp->faults[i] -= p->numa_faults[i];
2017 grp->faults[i] += p->numa_faults[i];
2019 my_grp->total_faults -= p->total_numa_faults;
2020 grp->total_faults += p->total_numa_faults;
2025 spin_unlock(&my_grp->lock);
2026 spin_unlock_irq(&grp->lock);
2028 rcu_assign_pointer(p->numa_group, grp);
2030 put_numa_group(my_grp);
2038 void task_numa_free(struct task_struct *p)
2040 struct numa_group *grp = p->numa_group;
2041 void *numa_faults = p->numa_faults;
2042 unsigned long flags;
2046 spin_lock_irqsave(&grp->lock, flags);
2047 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2048 grp->faults[i] -= p->numa_faults[i];
2049 grp->total_faults -= p->total_numa_faults;
2052 spin_unlock_irqrestore(&grp->lock, flags);
2053 RCU_INIT_POINTER(p->numa_group, NULL);
2054 put_numa_group(grp);
2057 p->numa_faults = NULL;
2062 * Got a PROT_NONE fault for a page on @node.
2064 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2066 struct task_struct *p = current;
2067 bool migrated = flags & TNF_MIGRATED;
2068 int cpu_node = task_node(current);
2069 int local = !!(flags & TNF_FAULT_LOCAL);
2072 if (!static_branch_likely(&sched_numa_balancing))
2075 /* for example, ksmd faulting in a user's mm */
2079 /* Allocate buffer to track faults on a per-node basis */
2080 if (unlikely(!p->numa_faults)) {
2081 int size = sizeof(*p->numa_faults) *
2082 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2084 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2085 if (!p->numa_faults)
2088 p->total_numa_faults = 0;
2089 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2093 * First accesses are treated as private, otherwise consider accesses
2094 * to be private if the accessing pid has not changed
2096 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2099 priv = cpupid_match_pid(p, last_cpupid);
2100 if (!priv && !(flags & TNF_NO_GROUP))
2101 task_numa_group(p, last_cpupid, flags, &priv);
2105 * If a workload spans multiple NUMA nodes, a shared fault that
2106 * occurs wholly within the set of nodes that the workload is
2107 * actively using should be counted as local. This allows the
2108 * scan rate to slow down when a workload has settled down.
2110 if (!priv && !local && p->numa_group &&
2111 node_isset(cpu_node, p->numa_group->active_nodes) &&
2112 node_isset(mem_node, p->numa_group->active_nodes))
2115 task_numa_placement(p);
2118 * Retry task to preferred node migration periodically, in case it
2119 * case it previously failed, or the scheduler moved us.
2121 if (time_after(jiffies, p->numa_migrate_retry))
2122 numa_migrate_preferred(p);
2125 p->numa_pages_migrated += pages;
2126 if (flags & TNF_MIGRATE_FAIL)
2127 p->numa_faults_locality[2] += pages;
2129 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2130 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2131 p->numa_faults_locality[local] += pages;
2134 static void reset_ptenuma_scan(struct task_struct *p)
2137 * We only did a read acquisition of the mmap sem, so
2138 * p->mm->numa_scan_seq is written to without exclusive access
2139 * and the update is not guaranteed to be atomic. That's not
2140 * much of an issue though, since this is just used for
2141 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2142 * expensive, to avoid any form of compiler optimizations:
2144 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2145 p->mm->numa_scan_offset = 0;
2149 * The expensive part of numa migration is done from task_work context.
2150 * Triggered from task_tick_numa().
2152 void task_numa_work(struct callback_head *work)
2154 unsigned long migrate, next_scan, now = jiffies;
2155 struct task_struct *p = current;
2156 struct mm_struct *mm = p->mm;
2157 struct vm_area_struct *vma;
2158 unsigned long start, end;
2159 unsigned long nr_pte_updates = 0;
2160 long pages, virtpages;
2162 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2164 work->next = work; /* protect against double add */
2166 * Who cares about NUMA placement when they're dying.
2168 * NOTE: make sure not to dereference p->mm before this check,
2169 * exit_task_work() happens _after_ exit_mm() so we could be called
2170 * without p->mm even though we still had it when we enqueued this
2173 if (p->flags & PF_EXITING)
2176 if (!mm->numa_next_scan) {
2177 mm->numa_next_scan = now +
2178 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2182 * Enforce maximal scan/migration frequency..
2184 migrate = mm->numa_next_scan;
2185 if (time_before(now, migrate))
2188 if (p->numa_scan_period == 0) {
2189 p->numa_scan_period_max = task_scan_max(p);
2190 p->numa_scan_period = task_scan_min(p);
2193 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2194 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2198 * Delay this task enough that another task of this mm will likely win
2199 * the next time around.
2201 p->node_stamp += 2 * TICK_NSEC;
2203 start = mm->numa_scan_offset;
2204 pages = sysctl_numa_balancing_scan_size;
2205 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2206 virtpages = pages * 8; /* Scan up to this much virtual space */
2211 down_read(&mm->mmap_sem);
2212 vma = find_vma(mm, start);
2214 reset_ptenuma_scan(p);
2218 for (; vma; vma = vma->vm_next) {
2219 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2220 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2225 * Shared library pages mapped by multiple processes are not
2226 * migrated as it is expected they are cache replicated. Avoid
2227 * hinting faults in read-only file-backed mappings or the vdso
2228 * as migrating the pages will be of marginal benefit.
2231 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2235 * Skip inaccessible VMAs to avoid any confusion between
2236 * PROT_NONE and NUMA hinting ptes
2238 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2242 start = max(start, vma->vm_start);
2243 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2244 end = min(end, vma->vm_end);
2245 nr_pte_updates = change_prot_numa(vma, start, end);
2248 * Try to scan sysctl_numa_balancing_size worth of
2249 * hpages that have at least one present PTE that
2250 * is not already pte-numa. If the VMA contains
2251 * areas that are unused or already full of prot_numa
2252 * PTEs, scan up to virtpages, to skip through those
2256 pages -= (end - start) >> PAGE_SHIFT;
2257 virtpages -= (end - start) >> PAGE_SHIFT;
2260 if (pages <= 0 || virtpages <= 0)
2264 } while (end != vma->vm_end);
2269 * It is possible to reach the end of the VMA list but the last few
2270 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2271 * would find the !migratable VMA on the next scan but not reset the
2272 * scanner to the start so check it now.
2275 mm->numa_scan_offset = start;
2277 reset_ptenuma_scan(p);
2278 up_read(&mm->mmap_sem);
2282 * Drive the periodic memory faults..
2284 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2286 struct callback_head *work = &curr->numa_work;
2290 * We don't care about NUMA placement if we don't have memory.
2292 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2296 * Using runtime rather than walltime has the dual advantage that
2297 * we (mostly) drive the selection from busy threads and that the
2298 * task needs to have done some actual work before we bother with
2301 now = curr->se.sum_exec_runtime;
2302 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2304 if (now > curr->node_stamp + period) {
2305 if (!curr->node_stamp)
2306 curr->numa_scan_period = task_scan_min(curr);
2307 curr->node_stamp += period;
2309 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2310 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2311 task_work_add(curr, work, true);
2316 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2320 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2324 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2327 #endif /* CONFIG_NUMA_BALANCING */
2330 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2332 update_load_add(&cfs_rq->load, se->load.weight);
2333 if (!parent_entity(se))
2334 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2336 if (entity_is_task(se)) {
2337 struct rq *rq = rq_of(cfs_rq);
2339 account_numa_enqueue(rq, task_of(se));
2340 list_add(&se->group_node, &rq->cfs_tasks);
2343 cfs_rq->nr_running++;
2347 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2349 update_load_sub(&cfs_rq->load, se->load.weight);
2350 if (!parent_entity(se))
2351 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2352 if (entity_is_task(se)) {
2353 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2354 list_del_init(&se->group_node);
2356 cfs_rq->nr_running--;
2359 #ifdef CONFIG_FAIR_GROUP_SCHED
2361 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2366 * Use this CPU's real-time load instead of the last load contribution
2367 * as the updating of the contribution is delayed, and we will use the
2368 * the real-time load to calc the share. See update_tg_load_avg().
2370 tg_weight = atomic_long_read(&tg->load_avg);
2371 tg_weight -= cfs_rq->tg_load_avg_contrib;
2372 tg_weight += cfs_rq->load.weight;
2377 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2379 long tg_weight, load, shares;
2381 tg_weight = calc_tg_weight(tg, cfs_rq);
2382 load = cfs_rq->load.weight;
2384 shares = (tg->shares * load);
2386 shares /= tg_weight;
2388 if (shares < MIN_SHARES)
2389 shares = MIN_SHARES;
2390 if (shares > tg->shares)
2391 shares = tg->shares;
2395 # else /* CONFIG_SMP */
2396 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2400 # endif /* CONFIG_SMP */
2401 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2402 unsigned long weight)
2405 /* commit outstanding execution time */
2406 if (cfs_rq->curr == se)
2407 update_curr(cfs_rq);
2408 account_entity_dequeue(cfs_rq, se);
2411 update_load_set(&se->load, weight);
2414 account_entity_enqueue(cfs_rq, se);
2417 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2419 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2421 struct task_group *tg;
2422 struct sched_entity *se;
2426 se = tg->se[cpu_of(rq_of(cfs_rq))];
2427 if (!se || throttled_hierarchy(cfs_rq))
2430 if (likely(se->load.weight == tg->shares))
2433 shares = calc_cfs_shares(cfs_rq, tg);
2435 reweight_entity(cfs_rq_of(se), se, shares);
2437 #else /* CONFIG_FAIR_GROUP_SCHED */
2438 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2441 #endif /* CONFIG_FAIR_GROUP_SCHED */
2444 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2445 static const u32 runnable_avg_yN_inv[] = {
2446 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2447 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2448 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2449 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2450 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2451 0x85aac367, 0x82cd8698,
2455 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2456 * over-estimates when re-combining.
2458 static const u32 runnable_avg_yN_sum[] = {
2459 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2460 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2461 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2466 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2468 static __always_inline u64 decay_load(u64 val, u64 n)
2470 unsigned int local_n;
2474 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2477 /* after bounds checking we can collapse to 32-bit */
2481 * As y^PERIOD = 1/2, we can combine
2482 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2483 * With a look-up table which covers y^n (n<PERIOD)
2485 * To achieve constant time decay_load.
2487 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2488 val >>= local_n / LOAD_AVG_PERIOD;
2489 local_n %= LOAD_AVG_PERIOD;
2492 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2497 * For updates fully spanning n periods, the contribution to runnable
2498 * average will be: \Sum 1024*y^n
2500 * We can compute this reasonably efficiently by combining:
2501 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2503 static u32 __compute_runnable_contrib(u64 n)
2507 if (likely(n <= LOAD_AVG_PERIOD))
2508 return runnable_avg_yN_sum[n];
2509 else if (unlikely(n >= LOAD_AVG_MAX_N))
2510 return LOAD_AVG_MAX;
2512 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2514 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2515 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2517 n -= LOAD_AVG_PERIOD;
2518 } while (n > LOAD_AVG_PERIOD);
2520 contrib = decay_load(contrib, n);
2521 return contrib + runnable_avg_yN_sum[n];
2524 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2525 #error "load tracking assumes 2^10 as unit"
2528 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2531 * We can represent the historical contribution to runnable average as the
2532 * coefficients of a geometric series. To do this we sub-divide our runnable
2533 * history into segments of approximately 1ms (1024us); label the segment that
2534 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2536 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2538 * (now) (~1ms ago) (~2ms ago)
2540 * Let u_i denote the fraction of p_i that the entity was runnable.
2542 * We then designate the fractions u_i as our co-efficients, yielding the
2543 * following representation of historical load:
2544 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2546 * We choose y based on the with of a reasonably scheduling period, fixing:
2549 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2550 * approximately half as much as the contribution to load within the last ms
2553 * When a period "rolls over" and we have new u_0`, multiplying the previous
2554 * sum again by y is sufficient to update:
2555 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2556 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2558 static __always_inline int
2559 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2560 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2562 u64 delta, scaled_delta, periods;
2564 unsigned int delta_w, scaled_delta_w, decayed = 0;
2565 unsigned long scale_freq, scale_cpu;
2567 delta = now - sa->last_update_time;
2569 * This should only happen when time goes backwards, which it
2570 * unfortunately does during sched clock init when we swap over to TSC.
2572 if ((s64)delta < 0) {
2573 sa->last_update_time = now;
2578 * Use 1024ns as the unit of measurement since it's a reasonable
2579 * approximation of 1us and fast to compute.
2584 sa->last_update_time = now;
2586 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2587 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2588 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2590 /* delta_w is the amount already accumulated against our next period */
2591 delta_w = sa->period_contrib;
2592 if (delta + delta_w >= 1024) {
2595 /* how much left for next period will start over, we don't know yet */
2596 sa->period_contrib = 0;
2599 * Now that we know we're crossing a period boundary, figure
2600 * out how much from delta we need to complete the current
2601 * period and accrue it.
2603 delta_w = 1024 - delta_w;
2604 scaled_delta_w = cap_scale(delta_w, scale_freq);
2606 sa->load_sum += weight * scaled_delta_w;
2608 cfs_rq->runnable_load_sum +=
2609 weight * scaled_delta_w;
2613 sa->util_sum += scaled_delta_w * scale_cpu;
2617 /* Figure out how many additional periods this update spans */
2618 periods = delta / 1024;
2621 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2623 cfs_rq->runnable_load_sum =
2624 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2626 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2628 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2629 contrib = __compute_runnable_contrib(periods);
2630 contrib = cap_scale(contrib, scale_freq);
2632 sa->load_sum += weight * contrib;
2634 cfs_rq->runnable_load_sum += weight * contrib;
2637 sa->util_sum += contrib * scale_cpu;
2640 /* Remainder of delta accrued against u_0` */
2641 scaled_delta = cap_scale(delta, scale_freq);
2643 sa->load_sum += weight * scaled_delta;
2645 cfs_rq->runnable_load_sum += weight * scaled_delta;
2648 sa->util_sum += scaled_delta * scale_cpu;
2650 sa->period_contrib += delta;
2653 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2655 cfs_rq->runnable_load_avg =
2656 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2658 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2664 #ifdef CONFIG_FAIR_GROUP_SCHED
2666 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2667 * and effective_load (which is not done because it is too costly).
2669 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2671 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2673 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2674 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2675 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2679 #else /* CONFIG_FAIR_GROUP_SCHED */
2680 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2681 #endif /* CONFIG_FAIR_GROUP_SCHED */
2683 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2686 * Unsigned subtract and clamp on underflow.
2688 * Explicitly do a load-store to ensure the intermediate value never hits
2689 * memory. This allows lockless observations without ever seeing the negative
2692 #define sub_positive(_ptr, _val) do { \
2693 typeof(_ptr) ptr = (_ptr); \
2694 typeof(*ptr) val = (_val); \
2695 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2699 WRITE_ONCE(*ptr, res); \
2702 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2703 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2705 struct sched_avg *sa = &cfs_rq->avg;
2706 int decayed, removed = 0;
2708 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2709 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2710 sub_positive(&sa->load_avg, r);
2711 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2715 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2716 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2717 sub_positive(&sa->util_avg, r);
2718 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2721 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2722 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2724 #ifndef CONFIG_64BIT
2726 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2729 return decayed || removed;
2732 /* Update task and its cfs_rq load average */
2733 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2735 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2736 u64 now = cfs_rq_clock_task(cfs_rq);
2737 int cpu = cpu_of(rq_of(cfs_rq));
2740 * Track task load average for carrying it to new CPU after migrated, and
2741 * track group sched_entity load average for task_h_load calc in migration
2743 __update_load_avg(now, cpu, &se->avg,
2744 se->on_rq * scale_load_down(se->load.weight),
2745 cfs_rq->curr == se, NULL);
2747 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2748 update_tg_load_avg(cfs_rq, 0);
2750 if (entity_is_task(se))
2751 trace_sched_load_avg_task(task_of(se), &se->avg);
2752 trace_sched_load_avg_cpu(cpu, cfs_rq);
2755 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2757 if (!sched_feat(ATTACH_AGE_LOAD))
2761 * If we got migrated (either between CPUs or between cgroups) we'll
2762 * have aged the average right before clearing @last_update_time.
2764 if (se->avg.last_update_time) {
2765 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2766 &se->avg, 0, 0, NULL);
2769 * XXX: we could have just aged the entire load away if we've been
2770 * absent from the fair class for too long.
2775 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2776 cfs_rq->avg.load_avg += se->avg.load_avg;
2777 cfs_rq->avg.load_sum += se->avg.load_sum;
2778 cfs_rq->avg.util_avg += se->avg.util_avg;
2779 cfs_rq->avg.util_sum += se->avg.util_sum;
2782 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2784 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2785 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2786 cfs_rq->curr == se, NULL);
2788 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2789 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2790 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2791 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2794 /* Add the load generated by se into cfs_rq's load average */
2796 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2798 struct sched_avg *sa = &se->avg;
2799 u64 now = cfs_rq_clock_task(cfs_rq);
2800 int migrated, decayed;
2802 migrated = !sa->last_update_time;
2804 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2805 se->on_rq * scale_load_down(se->load.weight),
2806 cfs_rq->curr == se, NULL);
2809 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2811 cfs_rq->runnable_load_avg += sa->load_avg;
2812 cfs_rq->runnable_load_sum += sa->load_sum;
2815 attach_entity_load_avg(cfs_rq, se);
2817 if (decayed || migrated)
2818 update_tg_load_avg(cfs_rq, 0);
2821 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2823 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2825 update_load_avg(se, 1);
2827 cfs_rq->runnable_load_avg =
2828 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2829 cfs_rq->runnable_load_sum =
2830 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2833 #ifndef CONFIG_64BIT
2834 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2836 u64 last_update_time_copy;
2837 u64 last_update_time;
2840 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2842 last_update_time = cfs_rq->avg.last_update_time;
2843 } while (last_update_time != last_update_time_copy);
2845 return last_update_time;
2848 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2850 return cfs_rq->avg.last_update_time;
2855 * Task first catches up with cfs_rq, and then subtract
2856 * itself from the cfs_rq (task must be off the queue now).
2858 void remove_entity_load_avg(struct sched_entity *se)
2860 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2861 u64 last_update_time;
2864 * Newly created task or never used group entity should not be removed
2865 * from its (source) cfs_rq
2867 if (se->avg.last_update_time == 0)
2870 last_update_time = cfs_rq_last_update_time(cfs_rq);
2872 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2873 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2874 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2878 * Update the rq's load with the elapsed running time before entering
2879 * idle. if the last scheduled task is not a CFS task, idle_enter will
2880 * be the only way to update the runnable statistic.
2882 void idle_enter_fair(struct rq *this_rq)
2887 * Update the rq's load with the elapsed idle time before a task is
2888 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2889 * be the only way to update the runnable statistic.
2891 void idle_exit_fair(struct rq *this_rq)
2895 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2897 return cfs_rq->runnable_load_avg;
2900 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2902 return cfs_rq->avg.load_avg;
2905 static int idle_balance(struct rq *this_rq);
2907 #else /* CONFIG_SMP */
2909 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2911 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2913 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2914 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2917 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2919 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2921 static inline int idle_balance(struct rq *rq)
2926 #endif /* CONFIG_SMP */
2928 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2930 #ifdef CONFIG_SCHEDSTATS
2931 struct task_struct *tsk = NULL;
2933 if (entity_is_task(se))
2936 if (se->statistics.sleep_start) {
2937 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2942 if (unlikely(delta > se->statistics.sleep_max))
2943 se->statistics.sleep_max = delta;
2945 se->statistics.sleep_start = 0;
2946 se->statistics.sum_sleep_runtime += delta;
2949 account_scheduler_latency(tsk, delta >> 10, 1);
2950 trace_sched_stat_sleep(tsk, delta);
2953 if (se->statistics.block_start) {
2954 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2959 if (unlikely(delta > se->statistics.block_max))
2960 se->statistics.block_max = delta;
2962 se->statistics.block_start = 0;
2963 se->statistics.sum_sleep_runtime += delta;
2966 if (tsk->in_iowait) {
2967 se->statistics.iowait_sum += delta;
2968 se->statistics.iowait_count++;
2969 trace_sched_stat_iowait(tsk, delta);
2972 trace_sched_stat_blocked(tsk, delta);
2973 trace_sched_blocked_reason(tsk);
2976 * Blocking time is in units of nanosecs, so shift by
2977 * 20 to get a milliseconds-range estimation of the
2978 * amount of time that the task spent sleeping:
2980 if (unlikely(prof_on == SLEEP_PROFILING)) {
2981 profile_hits(SLEEP_PROFILING,
2982 (void *)get_wchan(tsk),
2985 account_scheduler_latency(tsk, delta >> 10, 0);
2991 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2993 #ifdef CONFIG_SCHED_DEBUG
2994 s64 d = se->vruntime - cfs_rq->min_vruntime;
2999 if (d > 3*sysctl_sched_latency)
3000 schedstat_inc(cfs_rq, nr_spread_over);
3005 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3007 u64 vruntime = cfs_rq->min_vruntime;
3010 * The 'current' period is already promised to the current tasks,
3011 * however the extra weight of the new task will slow them down a
3012 * little, place the new task so that it fits in the slot that
3013 * stays open at the end.
3015 if (initial && sched_feat(START_DEBIT))
3016 vruntime += sched_vslice(cfs_rq, se);
3018 /* sleeps up to a single latency don't count. */
3020 unsigned long thresh = sysctl_sched_latency;
3023 * Halve their sleep time's effect, to allow
3024 * for a gentler effect of sleepers:
3026 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3032 /* ensure we never gain time by being placed backwards. */
3033 se->vruntime = max_vruntime(se->vruntime, vruntime);
3036 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3039 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3042 * Update the normalized vruntime before updating min_vruntime
3043 * through calling update_curr().
3045 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3046 se->vruntime += cfs_rq->min_vruntime;
3049 * Update run-time statistics of the 'current'.
3051 update_curr(cfs_rq);
3052 enqueue_entity_load_avg(cfs_rq, se);
3053 account_entity_enqueue(cfs_rq, se);
3054 update_cfs_shares(cfs_rq);
3056 if (flags & ENQUEUE_WAKEUP) {
3057 place_entity(cfs_rq, se, 0);
3058 enqueue_sleeper(cfs_rq, se);
3061 update_stats_enqueue(cfs_rq, se);
3062 check_spread(cfs_rq, se);
3063 if (se != cfs_rq->curr)
3064 __enqueue_entity(cfs_rq, se);
3067 if (cfs_rq->nr_running == 1) {
3068 list_add_leaf_cfs_rq(cfs_rq);
3069 check_enqueue_throttle(cfs_rq);
3073 static void __clear_buddies_last(struct sched_entity *se)
3075 for_each_sched_entity(se) {
3076 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3077 if (cfs_rq->last != se)
3080 cfs_rq->last = NULL;
3084 static void __clear_buddies_next(struct sched_entity *se)
3086 for_each_sched_entity(se) {
3087 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3088 if (cfs_rq->next != se)
3091 cfs_rq->next = NULL;
3095 static void __clear_buddies_skip(struct sched_entity *se)
3097 for_each_sched_entity(se) {
3098 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3099 if (cfs_rq->skip != se)
3102 cfs_rq->skip = NULL;
3106 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3108 if (cfs_rq->last == se)
3109 __clear_buddies_last(se);
3111 if (cfs_rq->next == se)
3112 __clear_buddies_next(se);
3114 if (cfs_rq->skip == se)
3115 __clear_buddies_skip(se);
3118 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3121 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3124 * Update run-time statistics of the 'current'.
3126 update_curr(cfs_rq);
3127 dequeue_entity_load_avg(cfs_rq, se);
3129 update_stats_dequeue(cfs_rq, se);
3130 if (flags & DEQUEUE_SLEEP) {
3131 #ifdef CONFIG_SCHEDSTATS
3132 if (entity_is_task(se)) {
3133 struct task_struct *tsk = task_of(se);
3135 if (tsk->state & TASK_INTERRUPTIBLE)
3136 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3137 if (tsk->state & TASK_UNINTERRUPTIBLE)
3138 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3143 clear_buddies(cfs_rq, se);
3145 if (se != cfs_rq->curr)
3146 __dequeue_entity(cfs_rq, se);
3148 account_entity_dequeue(cfs_rq, se);
3151 * Normalize the entity after updating the min_vruntime because the
3152 * update can refer to the ->curr item and we need to reflect this
3153 * movement in our normalized position.
3155 if (!(flags & DEQUEUE_SLEEP))
3156 se->vruntime -= cfs_rq->min_vruntime;
3158 /* return excess runtime on last dequeue */
3159 return_cfs_rq_runtime(cfs_rq);
3161 update_min_vruntime(cfs_rq);
3162 update_cfs_shares(cfs_rq);
3166 * Preempt the current task with a newly woken task if needed:
3169 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3171 unsigned long ideal_runtime, delta_exec;
3172 struct sched_entity *se;
3175 ideal_runtime = sched_slice(cfs_rq, curr);
3176 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3177 if (delta_exec > ideal_runtime) {
3178 resched_curr(rq_of(cfs_rq));
3180 * The current task ran long enough, ensure it doesn't get
3181 * re-elected due to buddy favours.
3183 clear_buddies(cfs_rq, curr);
3188 * Ensure that a task that missed wakeup preemption by a
3189 * narrow margin doesn't have to wait for a full slice.
3190 * This also mitigates buddy induced latencies under load.
3192 if (delta_exec < sysctl_sched_min_granularity)
3195 se = __pick_first_entity(cfs_rq);
3196 delta = curr->vruntime - se->vruntime;
3201 if (delta > ideal_runtime)
3202 resched_curr(rq_of(cfs_rq));
3206 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3208 /* 'current' is not kept within the tree. */
3211 * Any task has to be enqueued before it get to execute on
3212 * a CPU. So account for the time it spent waiting on the
3215 update_stats_wait_end(cfs_rq, se);
3216 __dequeue_entity(cfs_rq, se);
3217 update_load_avg(se, 1);
3220 update_stats_curr_start(cfs_rq, se);
3222 #ifdef CONFIG_SCHEDSTATS
3224 * Track our maximum slice length, if the CPU's load is at
3225 * least twice that of our own weight (i.e. dont track it
3226 * when there are only lesser-weight tasks around):
3228 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3229 se->statistics.slice_max = max(se->statistics.slice_max,
3230 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3233 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3237 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3240 * Pick the next process, keeping these things in mind, in this order:
3241 * 1) keep things fair between processes/task groups
3242 * 2) pick the "next" process, since someone really wants that to run
3243 * 3) pick the "last" process, for cache locality
3244 * 4) do not run the "skip" process, if something else is available
3246 static struct sched_entity *
3247 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3249 struct sched_entity *left = __pick_first_entity(cfs_rq);
3250 struct sched_entity *se;
3253 * If curr is set we have to see if its left of the leftmost entity
3254 * still in the tree, provided there was anything in the tree at all.
3256 if (!left || (curr && entity_before(curr, left)))
3259 se = left; /* ideally we run the leftmost entity */
3262 * Avoid running the skip buddy, if running something else can
3263 * be done without getting too unfair.
3265 if (cfs_rq->skip == se) {
3266 struct sched_entity *second;
3269 second = __pick_first_entity(cfs_rq);
3271 second = __pick_next_entity(se);
3272 if (!second || (curr && entity_before(curr, second)))
3276 if (second && wakeup_preempt_entity(second, left) < 1)
3281 * Prefer last buddy, try to return the CPU to a preempted task.
3283 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3287 * Someone really wants this to run. If it's not unfair, run it.
3289 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3292 clear_buddies(cfs_rq, se);
3297 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3299 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3302 * If still on the runqueue then deactivate_task()
3303 * was not called and update_curr() has to be done:
3306 update_curr(cfs_rq);
3308 /* throttle cfs_rqs exceeding runtime */
3309 check_cfs_rq_runtime(cfs_rq);
3311 check_spread(cfs_rq, prev);
3313 update_stats_wait_start(cfs_rq, prev);
3314 /* Put 'current' back into the tree. */
3315 __enqueue_entity(cfs_rq, prev);
3316 /* in !on_rq case, update occurred at dequeue */
3317 update_load_avg(prev, 0);
3319 cfs_rq->curr = NULL;
3323 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3326 * Update run-time statistics of the 'current'.
3328 update_curr(cfs_rq);
3331 * Ensure that runnable average is periodically updated.
3333 update_load_avg(curr, 1);
3334 update_cfs_shares(cfs_rq);
3336 #ifdef CONFIG_SCHED_HRTICK
3338 * queued ticks are scheduled to match the slice, so don't bother
3339 * validating it and just reschedule.
3342 resched_curr(rq_of(cfs_rq));
3346 * don't let the period tick interfere with the hrtick preemption
3348 if (!sched_feat(DOUBLE_TICK) &&
3349 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3353 if (cfs_rq->nr_running > 1)
3354 check_preempt_tick(cfs_rq, curr);
3358 /**************************************************
3359 * CFS bandwidth control machinery
3362 #ifdef CONFIG_CFS_BANDWIDTH
3364 #ifdef HAVE_JUMP_LABEL
3365 static struct static_key __cfs_bandwidth_used;
3367 static inline bool cfs_bandwidth_used(void)
3369 return static_key_false(&__cfs_bandwidth_used);
3372 void cfs_bandwidth_usage_inc(void)
3374 static_key_slow_inc(&__cfs_bandwidth_used);
3377 void cfs_bandwidth_usage_dec(void)
3379 static_key_slow_dec(&__cfs_bandwidth_used);
3381 #else /* HAVE_JUMP_LABEL */
3382 static bool cfs_bandwidth_used(void)
3387 void cfs_bandwidth_usage_inc(void) {}
3388 void cfs_bandwidth_usage_dec(void) {}
3389 #endif /* HAVE_JUMP_LABEL */
3392 * default period for cfs group bandwidth.
3393 * default: 0.1s, units: nanoseconds
3395 static inline u64 default_cfs_period(void)
3397 return 100000000ULL;
3400 static inline u64 sched_cfs_bandwidth_slice(void)
3402 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3406 * Replenish runtime according to assigned quota and update expiration time.
3407 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3408 * additional synchronization around rq->lock.
3410 * requires cfs_b->lock
3412 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3416 if (cfs_b->quota == RUNTIME_INF)
3419 now = sched_clock_cpu(smp_processor_id());
3420 cfs_b->runtime = cfs_b->quota;
3421 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3424 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3426 return &tg->cfs_bandwidth;
3429 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3430 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3432 if (unlikely(cfs_rq->throttle_count))
3433 return cfs_rq->throttled_clock_task;
3435 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3438 /* returns 0 on failure to allocate runtime */
3439 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3441 struct task_group *tg = cfs_rq->tg;
3442 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3443 u64 amount = 0, min_amount, expires;
3445 /* note: this is a positive sum as runtime_remaining <= 0 */
3446 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3448 raw_spin_lock(&cfs_b->lock);
3449 if (cfs_b->quota == RUNTIME_INF)
3450 amount = min_amount;
3452 start_cfs_bandwidth(cfs_b);
3454 if (cfs_b->runtime > 0) {
3455 amount = min(cfs_b->runtime, min_amount);
3456 cfs_b->runtime -= amount;
3460 expires = cfs_b->runtime_expires;
3461 raw_spin_unlock(&cfs_b->lock);
3463 cfs_rq->runtime_remaining += amount;
3465 * we may have advanced our local expiration to account for allowed
3466 * spread between our sched_clock and the one on which runtime was
3469 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3470 cfs_rq->runtime_expires = expires;
3472 return cfs_rq->runtime_remaining > 0;
3476 * Note: This depends on the synchronization provided by sched_clock and the
3477 * fact that rq->clock snapshots this value.
3479 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3481 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3483 /* if the deadline is ahead of our clock, nothing to do */
3484 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3487 if (cfs_rq->runtime_remaining < 0)
3491 * If the local deadline has passed we have to consider the
3492 * possibility that our sched_clock is 'fast' and the global deadline
3493 * has not truly expired.
3495 * Fortunately we can check determine whether this the case by checking
3496 * whether the global deadline has advanced. It is valid to compare
3497 * cfs_b->runtime_expires without any locks since we only care about
3498 * exact equality, so a partial write will still work.
3501 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3502 /* extend local deadline, drift is bounded above by 2 ticks */
3503 cfs_rq->runtime_expires += TICK_NSEC;
3505 /* global deadline is ahead, expiration has passed */
3506 cfs_rq->runtime_remaining = 0;
3510 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3512 /* dock delta_exec before expiring quota (as it could span periods) */
3513 cfs_rq->runtime_remaining -= delta_exec;
3514 expire_cfs_rq_runtime(cfs_rq);
3516 if (likely(cfs_rq->runtime_remaining > 0))
3520 * if we're unable to extend our runtime we resched so that the active
3521 * hierarchy can be throttled
3523 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3524 resched_curr(rq_of(cfs_rq));
3527 static __always_inline
3528 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3530 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3533 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3536 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3538 return cfs_bandwidth_used() && cfs_rq->throttled;
3541 /* check whether cfs_rq, or any parent, is throttled */
3542 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3544 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3548 * Ensure that neither of the group entities corresponding to src_cpu or
3549 * dest_cpu are members of a throttled hierarchy when performing group
3550 * load-balance operations.
3552 static inline int throttled_lb_pair(struct task_group *tg,
3553 int src_cpu, int dest_cpu)
3555 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3557 src_cfs_rq = tg->cfs_rq[src_cpu];
3558 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3560 return throttled_hierarchy(src_cfs_rq) ||
3561 throttled_hierarchy(dest_cfs_rq);
3564 /* updated child weight may affect parent so we have to do this bottom up */
3565 static int tg_unthrottle_up(struct task_group *tg, void *data)
3567 struct rq *rq = data;
3568 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3570 cfs_rq->throttle_count--;
3572 if (!cfs_rq->throttle_count) {
3573 /* adjust cfs_rq_clock_task() */
3574 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3575 cfs_rq->throttled_clock_task;
3582 static int tg_throttle_down(struct task_group *tg, void *data)
3584 struct rq *rq = data;
3585 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3587 /* group is entering throttled state, stop time */
3588 if (!cfs_rq->throttle_count)
3589 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3590 cfs_rq->throttle_count++;
3595 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3597 struct rq *rq = rq_of(cfs_rq);
3598 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3599 struct sched_entity *se;
3600 long task_delta, dequeue = 1;
3603 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3605 /* freeze hierarchy runnable averages while throttled */
3607 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3610 task_delta = cfs_rq->h_nr_running;
3611 for_each_sched_entity(se) {
3612 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3613 /* throttled entity or throttle-on-deactivate */
3618 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3619 qcfs_rq->h_nr_running -= task_delta;
3621 if (qcfs_rq->load.weight)
3626 sub_nr_running(rq, task_delta);
3628 cfs_rq->throttled = 1;
3629 cfs_rq->throttled_clock = rq_clock(rq);
3630 raw_spin_lock(&cfs_b->lock);
3631 empty = list_empty(&cfs_b->throttled_cfs_rq);
3634 * Add to the _head_ of the list, so that an already-started
3635 * distribute_cfs_runtime will not see us
3637 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3640 * If we're the first throttled task, make sure the bandwidth
3644 start_cfs_bandwidth(cfs_b);
3646 raw_spin_unlock(&cfs_b->lock);
3649 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3651 struct rq *rq = rq_of(cfs_rq);
3652 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3653 struct sched_entity *se;
3657 se = cfs_rq->tg->se[cpu_of(rq)];
3659 cfs_rq->throttled = 0;
3661 update_rq_clock(rq);
3663 raw_spin_lock(&cfs_b->lock);
3664 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3665 list_del_rcu(&cfs_rq->throttled_list);
3666 raw_spin_unlock(&cfs_b->lock);
3668 /* update hierarchical throttle state */
3669 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3671 if (!cfs_rq->load.weight)
3674 task_delta = cfs_rq->h_nr_running;
3675 for_each_sched_entity(se) {
3679 cfs_rq = cfs_rq_of(se);
3681 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3682 cfs_rq->h_nr_running += task_delta;
3684 if (cfs_rq_throttled(cfs_rq))
3689 add_nr_running(rq, task_delta);
3691 /* determine whether we need to wake up potentially idle cpu */
3692 if (rq->curr == rq->idle && rq->cfs.nr_running)
3696 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3697 u64 remaining, u64 expires)
3699 struct cfs_rq *cfs_rq;
3701 u64 starting_runtime = remaining;
3704 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3706 struct rq *rq = rq_of(cfs_rq);
3708 raw_spin_lock(&rq->lock);
3709 if (!cfs_rq_throttled(cfs_rq))
3712 runtime = -cfs_rq->runtime_remaining + 1;
3713 if (runtime > remaining)
3714 runtime = remaining;
3715 remaining -= runtime;
3717 cfs_rq->runtime_remaining += runtime;
3718 cfs_rq->runtime_expires = expires;
3720 /* we check whether we're throttled above */
3721 if (cfs_rq->runtime_remaining > 0)
3722 unthrottle_cfs_rq(cfs_rq);
3725 raw_spin_unlock(&rq->lock);
3732 return starting_runtime - remaining;
3736 * Responsible for refilling a task_group's bandwidth and unthrottling its
3737 * cfs_rqs as appropriate. If there has been no activity within the last
3738 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3739 * used to track this state.
3741 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3743 u64 runtime, runtime_expires;
3746 /* no need to continue the timer with no bandwidth constraint */
3747 if (cfs_b->quota == RUNTIME_INF)
3748 goto out_deactivate;
3750 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3751 cfs_b->nr_periods += overrun;
3754 * idle depends on !throttled (for the case of a large deficit), and if
3755 * we're going inactive then everything else can be deferred
3757 if (cfs_b->idle && !throttled)
3758 goto out_deactivate;
3760 __refill_cfs_bandwidth_runtime(cfs_b);
3763 /* mark as potentially idle for the upcoming period */
3768 /* account preceding periods in which throttling occurred */
3769 cfs_b->nr_throttled += overrun;
3771 runtime_expires = cfs_b->runtime_expires;
3774 * This check is repeated as we are holding onto the new bandwidth while
3775 * we unthrottle. This can potentially race with an unthrottled group
3776 * trying to acquire new bandwidth from the global pool. This can result
3777 * in us over-using our runtime if it is all used during this loop, but
3778 * only by limited amounts in that extreme case.
3780 while (throttled && cfs_b->runtime > 0) {
3781 runtime = cfs_b->runtime;
3782 raw_spin_unlock(&cfs_b->lock);
3783 /* we can't nest cfs_b->lock while distributing bandwidth */
3784 runtime = distribute_cfs_runtime(cfs_b, runtime,
3786 raw_spin_lock(&cfs_b->lock);
3788 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3790 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3794 * While we are ensured activity in the period following an
3795 * unthrottle, this also covers the case in which the new bandwidth is
3796 * insufficient to cover the existing bandwidth deficit. (Forcing the
3797 * timer to remain active while there are any throttled entities.)
3807 /* a cfs_rq won't donate quota below this amount */
3808 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3809 /* minimum remaining period time to redistribute slack quota */
3810 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3811 /* how long we wait to gather additional slack before distributing */
3812 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3815 * Are we near the end of the current quota period?
3817 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3818 * hrtimer base being cleared by hrtimer_start. In the case of
3819 * migrate_hrtimers, base is never cleared, so we are fine.
3821 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3823 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3826 /* if the call-back is running a quota refresh is already occurring */
3827 if (hrtimer_callback_running(refresh_timer))
3830 /* is a quota refresh about to occur? */
3831 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3832 if (remaining < min_expire)
3838 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3840 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3842 /* if there's a quota refresh soon don't bother with slack */
3843 if (runtime_refresh_within(cfs_b, min_left))
3846 hrtimer_start(&cfs_b->slack_timer,
3847 ns_to_ktime(cfs_bandwidth_slack_period),
3851 /* we know any runtime found here is valid as update_curr() precedes return */
3852 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3854 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3855 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3857 if (slack_runtime <= 0)
3860 raw_spin_lock(&cfs_b->lock);
3861 if (cfs_b->quota != RUNTIME_INF &&
3862 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3863 cfs_b->runtime += slack_runtime;
3865 /* we are under rq->lock, defer unthrottling using a timer */
3866 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3867 !list_empty(&cfs_b->throttled_cfs_rq))
3868 start_cfs_slack_bandwidth(cfs_b);
3870 raw_spin_unlock(&cfs_b->lock);
3872 /* even if it's not valid for return we don't want to try again */
3873 cfs_rq->runtime_remaining -= slack_runtime;
3876 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3878 if (!cfs_bandwidth_used())
3881 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3884 __return_cfs_rq_runtime(cfs_rq);
3888 * This is done with a timer (instead of inline with bandwidth return) since
3889 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3891 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3893 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3896 /* confirm we're still not at a refresh boundary */
3897 raw_spin_lock(&cfs_b->lock);
3898 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3899 raw_spin_unlock(&cfs_b->lock);
3903 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3904 runtime = cfs_b->runtime;
3906 expires = cfs_b->runtime_expires;
3907 raw_spin_unlock(&cfs_b->lock);
3912 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3914 raw_spin_lock(&cfs_b->lock);
3915 if (expires == cfs_b->runtime_expires)
3916 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3917 raw_spin_unlock(&cfs_b->lock);
3921 * When a group wakes up we want to make sure that its quota is not already
3922 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3923 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3925 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3927 if (!cfs_bandwidth_used())
3930 /* an active group must be handled by the update_curr()->put() path */
3931 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3934 /* ensure the group is not already throttled */
3935 if (cfs_rq_throttled(cfs_rq))
3938 /* update runtime allocation */
3939 account_cfs_rq_runtime(cfs_rq, 0);
3940 if (cfs_rq->runtime_remaining <= 0)
3941 throttle_cfs_rq(cfs_rq);
3944 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3945 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3947 if (!cfs_bandwidth_used())
3950 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3954 * it's possible for a throttled entity to be forced into a running
3955 * state (e.g. set_curr_task), in this case we're finished.
3957 if (cfs_rq_throttled(cfs_rq))
3960 throttle_cfs_rq(cfs_rq);
3964 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3966 struct cfs_bandwidth *cfs_b =
3967 container_of(timer, struct cfs_bandwidth, slack_timer);
3969 do_sched_cfs_slack_timer(cfs_b);
3971 return HRTIMER_NORESTART;
3974 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3976 struct cfs_bandwidth *cfs_b =
3977 container_of(timer, struct cfs_bandwidth, period_timer);
3981 raw_spin_lock(&cfs_b->lock);
3983 overrun = hrtimer_forward_now(timer, cfs_b->period);
3987 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3990 cfs_b->period_active = 0;
3991 raw_spin_unlock(&cfs_b->lock);
3993 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3996 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3998 raw_spin_lock_init(&cfs_b->lock);
4000 cfs_b->quota = RUNTIME_INF;
4001 cfs_b->period = ns_to_ktime(default_cfs_period());
4003 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4004 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4005 cfs_b->period_timer.function = sched_cfs_period_timer;
4006 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4007 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4010 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4012 cfs_rq->runtime_enabled = 0;
4013 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4016 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4018 lockdep_assert_held(&cfs_b->lock);
4020 if (!cfs_b->period_active) {
4021 cfs_b->period_active = 1;
4022 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4023 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4027 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4029 /* init_cfs_bandwidth() was not called */
4030 if (!cfs_b->throttled_cfs_rq.next)
4033 hrtimer_cancel(&cfs_b->period_timer);
4034 hrtimer_cancel(&cfs_b->slack_timer);
4037 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4039 struct cfs_rq *cfs_rq;
4041 for_each_leaf_cfs_rq(rq, cfs_rq) {
4042 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4044 raw_spin_lock(&cfs_b->lock);
4045 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4046 raw_spin_unlock(&cfs_b->lock);
4050 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4052 struct cfs_rq *cfs_rq;
4054 for_each_leaf_cfs_rq(rq, cfs_rq) {
4055 if (!cfs_rq->runtime_enabled)
4059 * clock_task is not advancing so we just need to make sure
4060 * there's some valid quota amount
4062 cfs_rq->runtime_remaining = 1;
4064 * Offline rq is schedulable till cpu is completely disabled
4065 * in take_cpu_down(), so we prevent new cfs throttling here.
4067 cfs_rq->runtime_enabled = 0;
4069 if (cfs_rq_throttled(cfs_rq))
4070 unthrottle_cfs_rq(cfs_rq);
4074 #else /* CONFIG_CFS_BANDWIDTH */
4075 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4077 return rq_clock_task(rq_of(cfs_rq));
4080 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4081 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4082 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4083 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4085 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4090 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4095 static inline int throttled_lb_pair(struct task_group *tg,
4096 int src_cpu, int dest_cpu)
4101 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4103 #ifdef CONFIG_FAIR_GROUP_SCHED
4104 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4107 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4111 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4112 static inline void update_runtime_enabled(struct rq *rq) {}
4113 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4115 #endif /* CONFIG_CFS_BANDWIDTH */
4117 /**************************************************
4118 * CFS operations on tasks:
4121 #ifdef CONFIG_SCHED_HRTICK
4122 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4124 struct sched_entity *se = &p->se;
4125 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4127 WARN_ON(task_rq(p) != rq);
4129 if (cfs_rq->nr_running > 1) {
4130 u64 slice = sched_slice(cfs_rq, se);
4131 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4132 s64 delta = slice - ran;
4139 hrtick_start(rq, delta);
4144 * called from enqueue/dequeue and updates the hrtick when the
4145 * current task is from our class and nr_running is low enough
4148 static void hrtick_update(struct rq *rq)
4150 struct task_struct *curr = rq->curr;
4152 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4155 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4156 hrtick_start_fair(rq, curr);
4158 #else /* !CONFIG_SCHED_HRTICK */
4160 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4164 static inline void hrtick_update(struct rq *rq)
4170 static bool cpu_overutilized(int cpu);
4171 static inline unsigned long boosted_cpu_util(int cpu);
4173 #define boosted_cpu_util(cpu) cpu_util(cpu)
4177 static void update_capacity_of(int cpu)
4179 unsigned long req_cap;
4184 /* Convert scale-invariant capacity to cpu. */
4185 req_cap = boosted_cpu_util(cpu);
4186 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4187 set_cfs_cpu_capacity(cpu, true, req_cap);
4192 * The enqueue_task method is called before nr_running is
4193 * increased. Here we update the fair scheduling stats and
4194 * then put the task into the rbtree:
4197 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4199 struct cfs_rq *cfs_rq;
4200 struct sched_entity *se = &p->se;
4202 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4203 int task_wakeup = flags & ENQUEUE_WAKEUP;
4206 for_each_sched_entity(se) {
4209 cfs_rq = cfs_rq_of(se);
4210 enqueue_entity(cfs_rq, se, flags);
4213 * end evaluation on encountering a throttled cfs_rq
4215 * note: in the case of encountering a throttled cfs_rq we will
4216 * post the final h_nr_running increment below.
4218 if (cfs_rq_throttled(cfs_rq))
4220 cfs_rq->h_nr_running++;
4222 flags = ENQUEUE_WAKEUP;
4225 for_each_sched_entity(se) {
4226 cfs_rq = cfs_rq_of(se);
4227 cfs_rq->h_nr_running++;
4229 if (cfs_rq_throttled(cfs_rq))
4232 update_load_avg(se, 1);
4233 update_cfs_shares(cfs_rq);
4237 add_nr_running(rq, 1);
4242 if (!task_new && !rq->rd->overutilized &&
4243 cpu_overutilized(rq->cpu))
4244 rq->rd->overutilized = true;
4246 schedtune_enqueue_task(p, cpu_of(rq));
4249 * We want to potentially trigger a freq switch
4250 * request only for tasks that are waking up; this is
4251 * because we get here also during load balancing, but
4252 * in these cases it seems wise to trigger as single
4253 * request after load balancing is done.
4255 if (task_new || task_wakeup)
4256 update_capacity_of(cpu_of(rq));
4258 #endif /* CONFIG_SMP */
4263 static void set_next_buddy(struct sched_entity *se);
4266 * The dequeue_task method is called before nr_running is
4267 * decreased. We remove the task from the rbtree and
4268 * update the fair scheduling stats:
4270 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4272 struct cfs_rq *cfs_rq;
4273 struct sched_entity *se = &p->se;
4274 int task_sleep = flags & DEQUEUE_SLEEP;
4276 for_each_sched_entity(se) {
4277 cfs_rq = cfs_rq_of(se);
4278 dequeue_entity(cfs_rq, se, flags);
4281 * end evaluation on encountering a throttled cfs_rq
4283 * note: in the case of encountering a throttled cfs_rq we will
4284 * post the final h_nr_running decrement below.
4286 if (cfs_rq_throttled(cfs_rq))
4288 cfs_rq->h_nr_running--;
4290 /* Don't dequeue parent if it has other entities besides us */
4291 if (cfs_rq->load.weight) {
4293 * Bias pick_next to pick a task from this cfs_rq, as
4294 * p is sleeping when it is within its sched_slice.
4296 if (task_sleep && parent_entity(se))
4297 set_next_buddy(parent_entity(se));
4299 /* avoid re-evaluating load for this entity */
4300 se = parent_entity(se);
4303 flags |= DEQUEUE_SLEEP;
4306 for_each_sched_entity(se) {
4307 cfs_rq = cfs_rq_of(se);
4308 cfs_rq->h_nr_running--;
4310 if (cfs_rq_throttled(cfs_rq))
4313 update_load_avg(se, 1);
4314 update_cfs_shares(cfs_rq);
4318 sub_nr_running(rq, 1);
4323 schedtune_dequeue_task(p, cpu_of(rq));
4326 * We want to potentially trigger a freq switch
4327 * request only for tasks that are going to sleep;
4328 * this is because we get here also during load
4329 * balancing, but in these cases it seems wise to
4330 * trigger as single request after load balancing is
4334 if (rq->cfs.nr_running)
4335 update_capacity_of(cpu_of(rq));
4336 else if (sched_freq())
4337 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4341 #endif /* CONFIG_SMP */
4349 * per rq 'load' arrray crap; XXX kill this.
4353 * The exact cpuload at various idx values, calculated at every tick would be
4354 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4356 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4357 * on nth tick when cpu may be busy, then we have:
4358 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4359 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4361 * decay_load_missed() below does efficient calculation of
4362 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4363 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4365 * The calculation is approximated on a 128 point scale.
4366 * degrade_zero_ticks is the number of ticks after which load at any
4367 * particular idx is approximated to be zero.
4368 * degrade_factor is a precomputed table, a row for each load idx.
4369 * Each column corresponds to degradation factor for a power of two ticks,
4370 * based on 128 point scale.
4372 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4373 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4375 * With this power of 2 load factors, we can degrade the load n times
4376 * by looking at 1 bits in n and doing as many mult/shift instead of
4377 * n mult/shifts needed by the exact degradation.
4379 #define DEGRADE_SHIFT 7
4380 static const unsigned char
4381 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4382 static const unsigned char
4383 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4384 {0, 0, 0, 0, 0, 0, 0, 0},
4385 {64, 32, 8, 0, 0, 0, 0, 0},
4386 {96, 72, 40, 12, 1, 0, 0},
4387 {112, 98, 75, 43, 15, 1, 0},
4388 {120, 112, 98, 76, 45, 16, 2} };
4391 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4392 * would be when CPU is idle and so we just decay the old load without
4393 * adding any new load.
4395 static unsigned long
4396 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4400 if (!missed_updates)
4403 if (missed_updates >= degrade_zero_ticks[idx])
4407 return load >> missed_updates;
4409 while (missed_updates) {
4410 if (missed_updates % 2)
4411 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4413 missed_updates >>= 1;
4420 * Update rq->cpu_load[] statistics. This function is usually called every
4421 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4422 * every tick. We fix it up based on jiffies.
4424 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4425 unsigned long pending_updates)
4429 this_rq->nr_load_updates++;
4431 /* Update our load: */
4432 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4433 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4434 unsigned long old_load, new_load;
4436 /* scale is effectively 1 << i now, and >> i divides by scale */
4438 old_load = this_rq->cpu_load[i];
4439 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4440 new_load = this_load;
4442 * Round up the averaging division if load is increasing. This
4443 * prevents us from getting stuck on 9 if the load is 10, for
4446 if (new_load > old_load)
4447 new_load += scale - 1;
4449 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4452 sched_avg_update(this_rq);
4455 /* Used instead of source_load when we know the type == 0 */
4456 static unsigned long weighted_cpuload(const int cpu)
4458 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4461 #ifdef CONFIG_NO_HZ_COMMON
4463 * There is no sane way to deal with nohz on smp when using jiffies because the
4464 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4465 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4467 * Therefore we cannot use the delta approach from the regular tick since that
4468 * would seriously skew the load calculation. However we'll make do for those
4469 * updates happening while idle (nohz_idle_balance) or coming out of idle
4470 * (tick_nohz_idle_exit).
4472 * This means we might still be one tick off for nohz periods.
4476 * Called from nohz_idle_balance() to update the load ratings before doing the
4479 static void update_idle_cpu_load(struct rq *this_rq)
4481 unsigned long curr_jiffies = READ_ONCE(jiffies);
4482 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4483 unsigned long pending_updates;
4486 * bail if there's load or we're actually up-to-date.
4488 if (load || curr_jiffies == this_rq->last_load_update_tick)
4491 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4492 this_rq->last_load_update_tick = curr_jiffies;
4494 __update_cpu_load(this_rq, load, pending_updates);
4498 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4500 void update_cpu_load_nohz(void)
4502 struct rq *this_rq = this_rq();
4503 unsigned long curr_jiffies = READ_ONCE(jiffies);
4504 unsigned long pending_updates;
4506 if (curr_jiffies == this_rq->last_load_update_tick)
4509 raw_spin_lock(&this_rq->lock);
4510 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4511 if (pending_updates) {
4512 this_rq->last_load_update_tick = curr_jiffies;
4514 * We were idle, this means load 0, the current load might be
4515 * !0 due to remote wakeups and the sort.
4517 __update_cpu_load(this_rq, 0, pending_updates);
4519 raw_spin_unlock(&this_rq->lock);
4521 #endif /* CONFIG_NO_HZ */
4524 * Called from scheduler_tick()
4526 void update_cpu_load_active(struct rq *this_rq)
4528 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4530 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4532 this_rq->last_load_update_tick = jiffies;
4533 __update_cpu_load(this_rq, load, 1);
4537 * Return a low guess at the load of a migration-source cpu weighted
4538 * according to the scheduling class and "nice" value.
4540 * We want to under-estimate the load of migration sources, to
4541 * balance conservatively.
4543 static unsigned long source_load(int cpu, int type)
4545 struct rq *rq = cpu_rq(cpu);
4546 unsigned long total = weighted_cpuload(cpu);
4548 if (type == 0 || !sched_feat(LB_BIAS))
4551 return min(rq->cpu_load[type-1], total);
4555 * Return a high guess at the load of a migration-target cpu weighted
4556 * according to the scheduling class and "nice" value.
4558 static unsigned long target_load(int cpu, int type)
4560 struct rq *rq = cpu_rq(cpu);
4561 unsigned long total = weighted_cpuload(cpu);
4563 if (type == 0 || !sched_feat(LB_BIAS))
4566 return max(rq->cpu_load[type-1], total);
4570 static unsigned long cpu_avg_load_per_task(int cpu)
4572 struct rq *rq = cpu_rq(cpu);
4573 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4574 unsigned long load_avg = weighted_cpuload(cpu);
4577 return load_avg / nr_running;
4582 static void record_wakee(struct task_struct *p)
4585 * Rough decay (wiping) for cost saving, don't worry
4586 * about the boundary, really active task won't care
4589 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4590 current->wakee_flips >>= 1;
4591 current->wakee_flip_decay_ts = jiffies;
4594 if (current->last_wakee != p) {
4595 current->last_wakee = p;
4596 current->wakee_flips++;
4600 static void task_waking_fair(struct task_struct *p)
4602 struct sched_entity *se = &p->se;
4603 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4606 #ifndef CONFIG_64BIT
4607 u64 min_vruntime_copy;
4610 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4612 min_vruntime = cfs_rq->min_vruntime;
4613 } while (min_vruntime != min_vruntime_copy);
4615 min_vruntime = cfs_rq->min_vruntime;
4618 se->vruntime -= min_vruntime;
4622 #ifdef CONFIG_FAIR_GROUP_SCHED
4624 * effective_load() calculates the load change as seen from the root_task_group
4626 * Adding load to a group doesn't make a group heavier, but can cause movement
4627 * of group shares between cpus. Assuming the shares were perfectly aligned one
4628 * can calculate the shift in shares.
4630 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4631 * on this @cpu and results in a total addition (subtraction) of @wg to the
4632 * total group weight.
4634 * Given a runqueue weight distribution (rw_i) we can compute a shares
4635 * distribution (s_i) using:
4637 * s_i = rw_i / \Sum rw_j (1)
4639 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4640 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4641 * shares distribution (s_i):
4643 * rw_i = { 2, 4, 1, 0 }
4644 * s_i = { 2/7, 4/7, 1/7, 0 }
4646 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4647 * task used to run on and the CPU the waker is running on), we need to
4648 * compute the effect of waking a task on either CPU and, in case of a sync
4649 * wakeup, compute the effect of the current task going to sleep.
4651 * So for a change of @wl to the local @cpu with an overall group weight change
4652 * of @wl we can compute the new shares distribution (s'_i) using:
4654 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4656 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4657 * differences in waking a task to CPU 0. The additional task changes the
4658 * weight and shares distributions like:
4660 * rw'_i = { 3, 4, 1, 0 }
4661 * s'_i = { 3/8, 4/8, 1/8, 0 }
4663 * We can then compute the difference in effective weight by using:
4665 * dw_i = S * (s'_i - s_i) (3)
4667 * Where 'S' is the group weight as seen by its parent.
4669 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4670 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4671 * 4/7) times the weight of the group.
4673 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4675 struct sched_entity *se = tg->se[cpu];
4677 if (!tg->parent) /* the trivial, non-cgroup case */
4680 for_each_sched_entity(se) {
4681 struct cfs_rq *cfs_rq = se->my_q;
4682 long W, w = cfs_rq_load_avg(cfs_rq);
4687 * W = @wg + \Sum rw_j
4689 W = wg + atomic_long_read(&tg->load_avg);
4691 /* Ensure \Sum rw_j >= rw_i */
4692 W -= cfs_rq->tg_load_avg_contrib;
4701 * wl = S * s'_i; see (2)
4704 wl = (w * (long)tg->shares) / W;
4709 * Per the above, wl is the new se->load.weight value; since
4710 * those are clipped to [MIN_SHARES, ...) do so now. See
4711 * calc_cfs_shares().
4713 if (wl < MIN_SHARES)
4717 * wl = dw_i = S * (s'_i - s_i); see (3)
4719 wl -= se->avg.load_avg;
4722 * Recursively apply this logic to all parent groups to compute
4723 * the final effective load change on the root group. Since
4724 * only the @tg group gets extra weight, all parent groups can
4725 * only redistribute existing shares. @wl is the shift in shares
4726 * resulting from this level per the above.
4735 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4743 * Returns the current capacity of cpu after applying both
4744 * cpu and freq scaling.
4746 unsigned long capacity_curr_of(int cpu)
4748 return cpu_rq(cpu)->cpu_capacity_orig *
4749 arch_scale_freq_capacity(NULL, cpu)
4750 >> SCHED_CAPACITY_SHIFT;
4753 static inline bool energy_aware(void)
4755 return sched_feat(ENERGY_AWARE);
4759 struct sched_group *sg_top;
4760 struct sched_group *sg_cap;
4767 struct task_struct *task;
4782 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4783 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4784 * energy calculations. Using the scale-invariant util returned by
4785 * cpu_util() and approximating scale-invariant util by:
4787 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4789 * the normalized util can be found using the specific capacity.
4791 * capacity = capacity_orig * curr_freq/max_freq
4793 * norm_util = running_time/time ~ util/capacity
4795 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4797 int util = __cpu_util(cpu, delta);
4799 if (util >= capacity)
4800 return SCHED_CAPACITY_SCALE;
4802 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4805 static int calc_util_delta(struct energy_env *eenv, int cpu)
4807 if (cpu == eenv->src_cpu)
4808 return -eenv->util_delta;
4809 if (cpu == eenv->dst_cpu)
4810 return eenv->util_delta;
4815 unsigned long group_max_util(struct energy_env *eenv)
4818 unsigned long max_util = 0;
4820 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4821 delta = calc_util_delta(eenv, i);
4822 max_util = max(max_util, __cpu_util(i, delta));
4829 * group_norm_util() returns the approximated group util relative to it's
4830 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4831 * energy calculations. Since task executions may or may not overlap in time in
4832 * the group the true normalized util is between max(cpu_norm_util(i)) and
4833 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4834 * latter is used as the estimate as it leads to a more pessimistic energy
4835 * estimate (more busy).
4838 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4841 unsigned long util_sum = 0;
4842 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4844 for_each_cpu(i, sched_group_cpus(sg)) {
4845 delta = calc_util_delta(eenv, i);
4846 util_sum += __cpu_norm_util(i, capacity, delta);
4849 if (util_sum > SCHED_CAPACITY_SCALE)
4850 return SCHED_CAPACITY_SCALE;
4854 static int find_new_capacity(struct energy_env *eenv,
4855 const struct sched_group_energy const *sge)
4858 unsigned long util = group_max_util(eenv);
4860 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4861 if (sge->cap_states[idx].cap >= util)
4865 eenv->cap_idx = idx;
4870 static int group_idle_state(struct sched_group *sg)
4872 int i, state = INT_MAX;
4874 /* Find the shallowest idle state in the sched group. */
4875 for_each_cpu(i, sched_group_cpus(sg))
4876 state = min(state, idle_get_state_idx(cpu_rq(i)));
4878 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4885 * sched_group_energy(): Computes the absolute energy consumption of cpus
4886 * belonging to the sched_group including shared resources shared only by
4887 * members of the group. Iterates over all cpus in the hierarchy below the
4888 * sched_group starting from the bottom working it's way up before going to
4889 * the next cpu until all cpus are covered at all levels. The current
4890 * implementation is likely to gather the same util statistics multiple times.
4891 * This can probably be done in a faster but more complex way.
4892 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4894 static int sched_group_energy(struct energy_env *eenv)
4896 struct sched_domain *sd;
4897 int cpu, total_energy = 0;
4898 struct cpumask visit_cpus;
4899 struct sched_group *sg;
4901 WARN_ON(!eenv->sg_top->sge);
4903 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4905 while (!cpumask_empty(&visit_cpus)) {
4906 struct sched_group *sg_shared_cap = NULL;
4908 cpu = cpumask_first(&visit_cpus);
4911 * Is the group utilization affected by cpus outside this
4914 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4918 * We most probably raced with hotplug; returning a
4919 * wrong energy estimation is better than entering an
4925 sg_shared_cap = sd->parent->groups;
4927 for_each_domain(cpu, sd) {
4930 /* Has this sched_domain already been visited? */
4931 if (sd->child && group_first_cpu(sg) != cpu)
4935 unsigned long group_util;
4936 int sg_busy_energy, sg_idle_energy;
4937 int cap_idx, idle_idx;
4939 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4940 eenv->sg_cap = sg_shared_cap;
4944 cap_idx = find_new_capacity(eenv, sg->sge);
4946 if (sg->group_weight == 1) {
4947 /* Remove capacity of src CPU (before task move) */
4948 if (eenv->util_delta == 0 &&
4949 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4950 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4951 eenv->cap.delta -= eenv->cap.before;
4953 /* Add capacity of dst CPU (after task move) */
4954 if (eenv->util_delta != 0 &&
4955 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4956 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4957 eenv->cap.delta += eenv->cap.after;
4961 idle_idx = group_idle_state(sg);
4962 group_util = group_norm_util(eenv, sg);
4963 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4964 >> SCHED_CAPACITY_SHIFT;
4965 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4966 * sg->sge->idle_states[idle_idx].power)
4967 >> SCHED_CAPACITY_SHIFT;
4969 total_energy += sg_busy_energy + sg_idle_energy;
4972 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4974 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4977 } while (sg = sg->next, sg != sd->groups);
4983 eenv->energy = total_energy;
4987 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4989 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4992 #ifdef CONFIG_SCHED_TUNE
4993 static int energy_diff_evaluate(struct energy_env *eenv)
4998 /* Return energy diff when boost margin is 0 */
4999 #ifdef CONFIG_CGROUP_SCHEDTUNE
5000 boost = schedtune_task_boost(eenv->task);
5002 boost = get_sysctl_sched_cfs_boost();
5005 return eenv->nrg.diff;
5007 /* Compute normalized energy diff */
5008 nrg_delta = schedtune_normalize_energy(eenv->nrg.diff);
5009 eenv->nrg.delta = nrg_delta;
5011 eenv->payoff = schedtune_accept_deltas(
5017 * When SchedTune is enabled, the energy_diff() function will return
5018 * the computed energy payoff value. Since the energy_diff() return
5019 * value is expected to be negative by its callers, this evaluation
5020 * function return a negative value each time the evaluation return a
5021 * positive payoff, which is the condition for the acceptance of
5022 * a scheduling decision
5024 return -eenv->payoff;
5026 #else /* CONFIG_SCHED_TUNE */
5027 #define energy_diff_evaluate(eenv) eenv->nrg.diff
5031 * energy_diff(): Estimate the energy impact of changing the utilization
5032 * distribution. eenv specifies the change: utilisation amount, source, and
5033 * destination cpu. Source or destination cpu may be -1 in which case the
5034 * utilization is removed from or added to the system (e.g. task wake-up). If
5035 * both are specified, the utilization is migrated.
5037 static int energy_diff(struct energy_env *eenv)
5039 struct sched_domain *sd;
5040 struct sched_group *sg;
5041 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5044 struct energy_env eenv_before = {
5046 .src_cpu = eenv->src_cpu,
5047 .dst_cpu = eenv->dst_cpu,
5048 .nrg = { 0, 0, 0, 0},
5052 if (eenv->src_cpu == eenv->dst_cpu)
5055 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5056 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5059 return 0; /* Error */
5064 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5065 eenv_before.sg_top = eenv->sg_top = sg;
5067 if (sched_group_energy(&eenv_before))
5068 return 0; /* Invalid result abort */
5069 energy_before += eenv_before.energy;
5071 /* Keep track of SRC cpu (before) capacity */
5072 eenv->cap.before = eenv_before.cap.before;
5073 eenv->cap.delta = eenv_before.cap.delta;
5075 if (sched_group_energy(eenv))
5076 return 0; /* Invalid result abort */
5077 energy_after += eenv->energy;
5079 } while (sg = sg->next, sg != sd->groups);
5081 eenv->nrg.before = energy_before;
5082 eenv->nrg.after = energy_after;
5083 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5086 result = energy_diff_evaluate(eenv);
5088 trace_sched_energy_diff(eenv->task,
5089 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5090 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5091 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5092 eenv->nrg.delta, eenv->payoff);
5098 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5099 * A waker of many should wake a different task than the one last awakened
5100 * at a frequency roughly N times higher than one of its wakees. In order
5101 * to determine whether we should let the load spread vs consolodating to
5102 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5103 * partner, and a factor of lls_size higher frequency in the other. With
5104 * both conditions met, we can be relatively sure that the relationship is
5105 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5106 * being client/server, worker/dispatcher, interrupt source or whatever is
5107 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5109 static int wake_wide(struct task_struct *p)
5111 unsigned int master = current->wakee_flips;
5112 unsigned int slave = p->wakee_flips;
5113 int factor = this_cpu_read(sd_llc_size);
5116 swap(master, slave);
5117 if (slave < factor || master < slave * factor)
5122 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5124 s64 this_load, load;
5125 s64 this_eff_load, prev_eff_load;
5126 int idx, this_cpu, prev_cpu;
5127 struct task_group *tg;
5128 unsigned long weight;
5132 this_cpu = smp_processor_id();
5133 prev_cpu = task_cpu(p);
5134 load = source_load(prev_cpu, idx);
5135 this_load = target_load(this_cpu, idx);
5138 * If sync wakeup then subtract the (maximum possible)
5139 * effect of the currently running task from the load
5140 * of the current CPU:
5143 tg = task_group(current);
5144 weight = current->se.avg.load_avg;
5146 this_load += effective_load(tg, this_cpu, -weight, -weight);
5147 load += effective_load(tg, prev_cpu, 0, -weight);
5151 weight = p->se.avg.load_avg;
5154 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5155 * due to the sync cause above having dropped this_load to 0, we'll
5156 * always have an imbalance, but there's really nothing you can do
5157 * about that, so that's good too.
5159 * Otherwise check if either cpus are near enough in load to allow this
5160 * task to be woken on this_cpu.
5162 this_eff_load = 100;
5163 this_eff_load *= capacity_of(prev_cpu);
5165 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5166 prev_eff_load *= capacity_of(this_cpu);
5168 if (this_load > 0) {
5169 this_eff_load *= this_load +
5170 effective_load(tg, this_cpu, weight, weight);
5172 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5175 balanced = this_eff_load <= prev_eff_load;
5177 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5182 schedstat_inc(sd, ttwu_move_affine);
5183 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5188 static inline unsigned long task_util(struct task_struct *p)
5190 return p->se.avg.util_avg;
5193 unsigned int capacity_margin = 1280; /* ~20% margin */
5195 static inline unsigned long boosted_task_util(struct task_struct *task);
5197 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5199 unsigned long capacity = capacity_of(cpu);
5201 util += boosted_task_util(p);
5203 return (capacity * 1024) > (util * capacity_margin);
5206 static inline bool task_fits_max(struct task_struct *p, int cpu)
5208 unsigned long capacity = capacity_of(cpu);
5209 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5211 if (capacity == max_capacity)
5214 if (capacity * capacity_margin > max_capacity * 1024)
5217 return __task_fits(p, cpu, 0);
5220 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5222 return __task_fits(p, cpu, cpu_util(cpu));
5225 static bool cpu_overutilized(int cpu)
5227 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5230 #ifdef CONFIG_SCHED_TUNE
5232 static unsigned long
5233 schedtune_margin(unsigned long signal, unsigned long boost)
5235 unsigned long long margin = 0;
5238 * Signal proportional compensation (SPC)
5240 * The Boost (B) value is used to compute a Margin (M) which is
5241 * proportional to the complement of the original Signal (S):
5242 * M = B * (SCHED_LOAD_SCALE - S)
5243 * The obtained M could be used by the caller to "boost" S.
5245 margin = SCHED_LOAD_SCALE - signal;
5249 * Fast integer division by constant:
5250 * Constant : (C) = 100
5251 * Precision : 0.1% (P) = 0.1
5252 * Reference : C * 100 / P (R) = 100000
5255 * Shift bits : ceil(log(R,2)) (S) = 17
5256 * Mult const : round(2^S/C) (M) = 1311
5266 static inline unsigned int
5267 schedtune_cpu_margin(unsigned long util, int cpu)
5271 #ifdef CONFIG_CGROUP_SCHEDTUNE
5272 boost = schedtune_cpu_boost(cpu);
5274 boost = get_sysctl_sched_cfs_boost();
5279 return schedtune_margin(util, boost);
5282 static inline unsigned long
5283 schedtune_task_margin(struct task_struct *task)
5287 unsigned long margin;
5289 #ifdef CONFIG_CGROUP_SCHEDTUNE
5290 boost = schedtune_task_boost(task);
5292 boost = get_sysctl_sched_cfs_boost();
5297 util = task_util(task);
5298 margin = schedtune_margin(util, boost);
5303 #else /* CONFIG_SCHED_TUNE */
5305 static inline unsigned int
5306 schedtune_cpu_margin(unsigned long util, int cpu)
5311 static inline unsigned int
5312 schedtune_task_margin(struct task_struct *task)
5317 #endif /* CONFIG_SCHED_TUNE */
5319 static inline unsigned long
5320 boosted_cpu_util(int cpu)
5322 unsigned long util = cpu_util(cpu);
5323 unsigned long margin = schedtune_cpu_margin(util, cpu);
5325 trace_sched_boost_cpu(cpu, util, margin);
5327 return util + margin;
5330 static inline unsigned long
5331 boosted_task_util(struct task_struct *task)
5333 unsigned long util = task_util(task);
5334 unsigned long margin = schedtune_task_margin(task);
5336 trace_sched_boost_task(task, util, margin);
5338 return util + margin;
5342 * find_idlest_group finds and returns the least busy CPU group within the
5345 static struct sched_group *
5346 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5347 int this_cpu, int sd_flag)
5349 struct sched_group *idlest = NULL, *group = sd->groups;
5350 struct sched_group *fit_group = NULL, *spare_group = NULL;
5351 unsigned long min_load = ULONG_MAX, this_load = 0;
5352 unsigned long fit_capacity = ULONG_MAX;
5353 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5354 int load_idx = sd->forkexec_idx;
5355 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5357 if (sd_flag & SD_BALANCE_WAKE)
5358 load_idx = sd->wake_idx;
5361 unsigned long load, avg_load, spare_capacity;
5365 /* Skip over this group if it has no CPUs allowed */
5366 if (!cpumask_intersects(sched_group_cpus(group),
5367 tsk_cpus_allowed(p)))
5370 local_group = cpumask_test_cpu(this_cpu,
5371 sched_group_cpus(group));
5373 /* Tally up the load of all CPUs in the group */
5376 for_each_cpu(i, sched_group_cpus(group)) {
5377 /* Bias balancing toward cpus of our domain */
5379 load = source_load(i, load_idx);
5381 load = target_load(i, load_idx);
5386 * Look for most energy-efficient group that can fit
5387 * that can fit the task.
5389 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5390 fit_capacity = capacity_of(i);
5395 * Look for group which has most spare capacity on a
5398 spare_capacity = capacity_of(i) - cpu_util(i);
5399 if (spare_capacity > max_spare_capacity) {
5400 max_spare_capacity = spare_capacity;
5401 spare_group = group;
5405 /* Adjust by relative CPU capacity of the group */
5406 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5409 this_load = avg_load;
5410 } else if (avg_load < min_load) {
5411 min_load = avg_load;
5414 } while (group = group->next, group != sd->groups);
5422 if (!idlest || 100*this_load < imbalance*min_load)
5428 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5431 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5433 unsigned long load, min_load = ULONG_MAX;
5434 unsigned int min_exit_latency = UINT_MAX;
5435 u64 latest_idle_timestamp = 0;
5436 int least_loaded_cpu = this_cpu;
5437 int shallowest_idle_cpu = -1;
5440 /* Traverse only the allowed CPUs */
5441 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5442 if (task_fits_spare(p, i)) {
5443 struct rq *rq = cpu_rq(i);
5444 struct cpuidle_state *idle = idle_get_state(rq);
5445 if (idle && idle->exit_latency < min_exit_latency) {
5447 * We give priority to a CPU whose idle state
5448 * has the smallest exit latency irrespective
5449 * of any idle timestamp.
5451 min_exit_latency = idle->exit_latency;
5452 latest_idle_timestamp = rq->idle_stamp;
5453 shallowest_idle_cpu = i;
5454 } else if (idle_cpu(i) &&
5455 (!idle || idle->exit_latency == min_exit_latency) &&
5456 rq->idle_stamp > latest_idle_timestamp) {
5458 * If equal or no active idle state, then
5459 * the most recently idled CPU might have
5462 latest_idle_timestamp = rq->idle_stamp;
5463 shallowest_idle_cpu = i;
5464 } else if (shallowest_idle_cpu == -1) {
5466 * If we haven't found an idle CPU yet
5467 * pick a non-idle one that can fit the task as
5470 shallowest_idle_cpu = i;
5472 } else if (shallowest_idle_cpu == -1) {
5473 load = weighted_cpuload(i);
5474 if (load < min_load || (load == min_load && i == this_cpu)) {
5476 least_loaded_cpu = i;
5481 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5485 * Try and locate an idle CPU in the sched_domain.
5487 static int select_idle_sibling(struct task_struct *p, int target)
5489 struct sched_domain *sd;
5490 struct sched_group *sg;
5491 int i = task_cpu(p);
5493 if (idle_cpu(target))
5497 * If the prevous cpu is cache affine and idle, don't be stupid.
5499 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5503 * Otherwise, iterate the domains and find an elegible idle cpu.
5505 sd = rcu_dereference(per_cpu(sd_llc, target));
5506 for_each_lower_domain(sd) {
5509 if (!cpumask_intersects(sched_group_cpus(sg),
5510 tsk_cpus_allowed(p)))
5513 for_each_cpu(i, sched_group_cpus(sg)) {
5514 if (i == target || !idle_cpu(i))
5518 target = cpumask_first_and(sched_group_cpus(sg),
5519 tsk_cpus_allowed(p));
5523 } while (sg != sd->groups);
5529 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5531 struct sched_domain *sd;
5532 struct sched_group *sg, *sg_target;
5533 int target_max_cap = INT_MAX;
5534 int target_cpu = task_cpu(p);
5537 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5546 * Find group with sufficient capacity. We only get here if no cpu is
5547 * overutilized. We may end up overutilizing a cpu by adding the task,
5548 * but that should not be any worse than select_idle_sibling().
5549 * load_balance() should sort it out later as we get above the tipping
5553 /* Assuming all cpus are the same in group */
5554 int max_cap_cpu = group_first_cpu(sg);
5557 * Assume smaller max capacity means more energy-efficient.
5558 * Ideally we should query the energy model for the right
5559 * answer but it easily ends up in an exhaustive search.
5561 if (capacity_of(max_cap_cpu) < target_max_cap &&
5562 task_fits_max(p, max_cap_cpu)) {
5564 target_max_cap = capacity_of(max_cap_cpu);
5566 } while (sg = sg->next, sg != sd->groups);
5568 /* Find cpu with sufficient capacity */
5569 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5571 * p's blocked utilization is still accounted for on prev_cpu
5572 * so prev_cpu will receive a negative bias due to the double
5573 * accounting. However, the blocked utilization may be zero.
5575 int new_util = cpu_util(i) + boosted_task_util(p);
5577 if (new_util > capacity_orig_of(i))
5580 if (new_util < capacity_curr_of(i)) {
5582 if (cpu_rq(i)->nr_running)
5586 /* cpu has capacity at higher OPP, keep it as fallback */
5587 if (target_cpu == task_cpu(p))
5591 if (target_cpu != task_cpu(p)) {
5592 struct energy_env eenv = {
5593 .util_delta = task_util(p),
5594 .src_cpu = task_cpu(p),
5595 .dst_cpu = target_cpu,
5599 /* Not enough spare capacity on previous cpu */
5600 if (cpu_overutilized(task_cpu(p)))
5603 if (energy_diff(&eenv) >= 0)
5611 * select_task_rq_fair: Select target runqueue for the waking task in domains
5612 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5613 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5615 * Balances load by selecting the idlest cpu in the idlest group, or under
5616 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5618 * Returns the target cpu number.
5620 * preempt must be disabled.
5623 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5625 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5626 int cpu = smp_processor_id();
5627 int new_cpu = prev_cpu;
5628 int want_affine = 0;
5629 int sync = wake_flags & WF_SYNC;
5631 if (sd_flag & SD_BALANCE_WAKE)
5632 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5633 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5637 for_each_domain(cpu, tmp) {
5638 if (!(tmp->flags & SD_LOAD_BALANCE))
5642 * If both cpu and prev_cpu are part of this domain,
5643 * cpu is a valid SD_WAKE_AFFINE target.
5645 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5646 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5651 if (tmp->flags & sd_flag)
5653 else if (!want_affine)
5658 sd = NULL; /* Prefer wake_affine over balance flags */
5659 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5664 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5665 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5666 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5667 new_cpu = select_idle_sibling(p, new_cpu);
5670 struct sched_group *group;
5673 if (!(sd->flags & sd_flag)) {
5678 group = find_idlest_group(sd, p, cpu, sd_flag);
5684 new_cpu = find_idlest_cpu(group, p, cpu);
5685 if (new_cpu == -1 || new_cpu == cpu) {
5686 /* Now try balancing at a lower domain level of cpu */
5691 /* Now try balancing at a lower domain level of new_cpu */
5693 weight = sd->span_weight;
5695 for_each_domain(cpu, tmp) {
5696 if (weight <= tmp->span_weight)
5698 if (tmp->flags & sd_flag)
5701 /* while loop will break here if sd == NULL */
5709 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5710 * cfs_rq_of(p) references at time of call are still valid and identify the
5711 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5712 * other assumptions, including the state of rq->lock, should be made.
5714 static void migrate_task_rq_fair(struct task_struct *p)
5717 * We are supposed to update the task to "current" time, then its up to date
5718 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5719 * what current time is, so simply throw away the out-of-date time. This
5720 * will result in the wakee task is less decayed, but giving the wakee more
5721 * load sounds not bad.
5723 remove_entity_load_avg(&p->se);
5725 /* Tell new CPU we are migrated */
5726 p->se.avg.last_update_time = 0;
5728 /* We have migrated, no longer consider this task hot */
5729 p->se.exec_start = 0;
5732 static void task_dead_fair(struct task_struct *p)
5734 remove_entity_load_avg(&p->se);
5737 #define task_fits_max(p, cpu) true
5738 #endif /* CONFIG_SMP */
5740 static unsigned long
5741 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5743 unsigned long gran = sysctl_sched_wakeup_granularity;
5746 * Since its curr running now, convert the gran from real-time
5747 * to virtual-time in his units.
5749 * By using 'se' instead of 'curr' we penalize light tasks, so
5750 * they get preempted easier. That is, if 'se' < 'curr' then
5751 * the resulting gran will be larger, therefore penalizing the
5752 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5753 * be smaller, again penalizing the lighter task.
5755 * This is especially important for buddies when the leftmost
5756 * task is higher priority than the buddy.
5758 return calc_delta_fair(gran, se);
5762 * Should 'se' preempt 'curr'.
5776 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5778 s64 gran, vdiff = curr->vruntime - se->vruntime;
5783 gran = wakeup_gran(curr, se);
5790 static void set_last_buddy(struct sched_entity *se)
5792 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5795 for_each_sched_entity(se)
5796 cfs_rq_of(se)->last = se;
5799 static void set_next_buddy(struct sched_entity *se)
5801 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5804 for_each_sched_entity(se)
5805 cfs_rq_of(se)->next = se;
5808 static void set_skip_buddy(struct sched_entity *se)
5810 for_each_sched_entity(se)
5811 cfs_rq_of(se)->skip = se;
5815 * Preempt the current task with a newly woken task if needed:
5817 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5819 struct task_struct *curr = rq->curr;
5820 struct sched_entity *se = &curr->se, *pse = &p->se;
5821 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5822 int scale = cfs_rq->nr_running >= sched_nr_latency;
5823 int next_buddy_marked = 0;
5825 if (unlikely(se == pse))
5829 * This is possible from callers such as attach_tasks(), in which we
5830 * unconditionally check_prempt_curr() after an enqueue (which may have
5831 * lead to a throttle). This both saves work and prevents false
5832 * next-buddy nomination below.
5834 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5837 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5838 set_next_buddy(pse);
5839 next_buddy_marked = 1;
5843 * We can come here with TIF_NEED_RESCHED already set from new task
5846 * Note: this also catches the edge-case of curr being in a throttled
5847 * group (e.g. via set_curr_task), since update_curr() (in the
5848 * enqueue of curr) will have resulted in resched being set. This
5849 * prevents us from potentially nominating it as a false LAST_BUDDY
5852 if (test_tsk_need_resched(curr))
5855 /* Idle tasks are by definition preempted by non-idle tasks. */
5856 if (unlikely(curr->policy == SCHED_IDLE) &&
5857 likely(p->policy != SCHED_IDLE))
5861 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5862 * is driven by the tick):
5864 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5867 find_matching_se(&se, &pse);
5868 update_curr(cfs_rq_of(se));
5870 if (wakeup_preempt_entity(se, pse) == 1) {
5872 * Bias pick_next to pick the sched entity that is
5873 * triggering this preemption.
5875 if (!next_buddy_marked)
5876 set_next_buddy(pse);
5885 * Only set the backward buddy when the current task is still
5886 * on the rq. This can happen when a wakeup gets interleaved
5887 * with schedule on the ->pre_schedule() or idle_balance()
5888 * point, either of which can * drop the rq lock.
5890 * Also, during early boot the idle thread is in the fair class,
5891 * for obvious reasons its a bad idea to schedule back to it.
5893 if (unlikely(!se->on_rq || curr == rq->idle))
5896 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5900 static struct task_struct *
5901 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5903 struct cfs_rq *cfs_rq = &rq->cfs;
5904 struct sched_entity *se;
5905 struct task_struct *p;
5909 #ifdef CONFIG_FAIR_GROUP_SCHED
5910 if (!cfs_rq->nr_running)
5913 if (prev->sched_class != &fair_sched_class)
5917 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5918 * likely that a next task is from the same cgroup as the current.
5920 * Therefore attempt to avoid putting and setting the entire cgroup
5921 * hierarchy, only change the part that actually changes.
5925 struct sched_entity *curr = cfs_rq->curr;
5928 * Since we got here without doing put_prev_entity() we also
5929 * have to consider cfs_rq->curr. If it is still a runnable
5930 * entity, update_curr() will update its vruntime, otherwise
5931 * forget we've ever seen it.
5935 update_curr(cfs_rq);
5940 * This call to check_cfs_rq_runtime() will do the
5941 * throttle and dequeue its entity in the parent(s).
5942 * Therefore the 'simple' nr_running test will indeed
5945 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5949 se = pick_next_entity(cfs_rq, curr);
5950 cfs_rq = group_cfs_rq(se);
5956 * Since we haven't yet done put_prev_entity and if the selected task
5957 * is a different task than we started out with, try and touch the
5958 * least amount of cfs_rqs.
5961 struct sched_entity *pse = &prev->se;
5963 while (!(cfs_rq = is_same_group(se, pse))) {
5964 int se_depth = se->depth;
5965 int pse_depth = pse->depth;
5967 if (se_depth <= pse_depth) {
5968 put_prev_entity(cfs_rq_of(pse), pse);
5969 pse = parent_entity(pse);
5971 if (se_depth >= pse_depth) {
5972 set_next_entity(cfs_rq_of(se), se);
5973 se = parent_entity(se);
5977 put_prev_entity(cfs_rq, pse);
5978 set_next_entity(cfs_rq, se);
5981 if (hrtick_enabled(rq))
5982 hrtick_start_fair(rq, p);
5984 rq->misfit_task = !task_fits_max(p, rq->cpu);
5991 if (!cfs_rq->nr_running)
5994 put_prev_task(rq, prev);
5997 se = pick_next_entity(cfs_rq, NULL);
5998 set_next_entity(cfs_rq, se);
5999 cfs_rq = group_cfs_rq(se);
6004 if (hrtick_enabled(rq))
6005 hrtick_start_fair(rq, p);
6007 rq->misfit_task = !task_fits_max(p, rq->cpu);
6012 rq->misfit_task = 0;
6014 * This is OK, because current is on_cpu, which avoids it being picked
6015 * for load-balance and preemption/IRQs are still disabled avoiding
6016 * further scheduler activity on it and we're being very careful to
6017 * re-start the picking loop.
6019 lockdep_unpin_lock(&rq->lock);
6020 new_tasks = idle_balance(rq);
6021 lockdep_pin_lock(&rq->lock);
6023 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6024 * possible for any higher priority task to appear. In that case we
6025 * must re-start the pick_next_entity() loop.
6037 * Account for a descheduled task:
6039 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6041 struct sched_entity *se = &prev->se;
6042 struct cfs_rq *cfs_rq;
6044 for_each_sched_entity(se) {
6045 cfs_rq = cfs_rq_of(se);
6046 put_prev_entity(cfs_rq, se);
6051 * sched_yield() is very simple
6053 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6055 static void yield_task_fair(struct rq *rq)
6057 struct task_struct *curr = rq->curr;
6058 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6059 struct sched_entity *se = &curr->se;
6062 * Are we the only task in the tree?
6064 if (unlikely(rq->nr_running == 1))
6067 clear_buddies(cfs_rq, se);
6069 if (curr->policy != SCHED_BATCH) {
6070 update_rq_clock(rq);
6072 * Update run-time statistics of the 'current'.
6074 update_curr(cfs_rq);
6076 * Tell update_rq_clock() that we've just updated,
6077 * so we don't do microscopic update in schedule()
6078 * and double the fastpath cost.
6080 rq_clock_skip_update(rq, true);
6086 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6088 struct sched_entity *se = &p->se;
6090 /* throttled hierarchies are not runnable */
6091 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6094 /* Tell the scheduler that we'd really like pse to run next. */
6097 yield_task_fair(rq);
6103 /**************************************************
6104 * Fair scheduling class load-balancing methods.
6108 * The purpose of load-balancing is to achieve the same basic fairness the
6109 * per-cpu scheduler provides, namely provide a proportional amount of compute
6110 * time to each task. This is expressed in the following equation:
6112 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6114 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6115 * W_i,0 is defined as:
6117 * W_i,0 = \Sum_j w_i,j (2)
6119 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6120 * is derived from the nice value as per prio_to_weight[].
6122 * The weight average is an exponential decay average of the instantaneous
6125 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6127 * C_i is the compute capacity of cpu i, typically it is the
6128 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6129 * can also include other factors [XXX].
6131 * To achieve this balance we define a measure of imbalance which follows
6132 * directly from (1):
6134 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6136 * We them move tasks around to minimize the imbalance. In the continuous
6137 * function space it is obvious this converges, in the discrete case we get
6138 * a few fun cases generally called infeasible weight scenarios.
6141 * - infeasible weights;
6142 * - local vs global optima in the discrete case. ]
6147 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6148 * for all i,j solution, we create a tree of cpus that follows the hardware
6149 * topology where each level pairs two lower groups (or better). This results
6150 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6151 * tree to only the first of the previous level and we decrease the frequency
6152 * of load-balance at each level inv. proportional to the number of cpus in
6158 * \Sum { --- * --- * 2^i } = O(n) (5)
6160 * `- size of each group
6161 * | | `- number of cpus doing load-balance
6163 * `- sum over all levels
6165 * Coupled with a limit on how many tasks we can migrate every balance pass,
6166 * this makes (5) the runtime complexity of the balancer.
6168 * An important property here is that each CPU is still (indirectly) connected
6169 * to every other cpu in at most O(log n) steps:
6171 * The adjacency matrix of the resulting graph is given by:
6174 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6177 * And you'll find that:
6179 * A^(log_2 n)_i,j != 0 for all i,j (7)
6181 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6182 * The task movement gives a factor of O(m), giving a convergence complexity
6185 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6190 * In order to avoid CPUs going idle while there's still work to do, new idle
6191 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6192 * tree itself instead of relying on other CPUs to bring it work.
6194 * This adds some complexity to both (5) and (8) but it reduces the total idle
6202 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6205 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6210 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6212 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6214 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6217 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6218 * rewrite all of this once again.]
6221 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6223 enum fbq_type { regular, remote, all };
6232 #define LBF_ALL_PINNED 0x01
6233 #define LBF_NEED_BREAK 0x02
6234 #define LBF_DST_PINNED 0x04
6235 #define LBF_SOME_PINNED 0x08
6238 struct sched_domain *sd;
6246 struct cpumask *dst_grpmask;
6248 enum cpu_idle_type idle;
6250 unsigned int src_grp_nr_running;
6251 /* The set of CPUs under consideration for load-balancing */
6252 struct cpumask *cpus;
6257 unsigned int loop_break;
6258 unsigned int loop_max;
6260 enum fbq_type fbq_type;
6261 enum group_type busiest_group_type;
6262 struct list_head tasks;
6266 * Is this task likely cache-hot:
6268 static int task_hot(struct task_struct *p, struct lb_env *env)
6272 lockdep_assert_held(&env->src_rq->lock);
6274 if (p->sched_class != &fair_sched_class)
6277 if (unlikely(p->policy == SCHED_IDLE))
6281 * Buddy candidates are cache hot:
6283 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6284 (&p->se == cfs_rq_of(&p->se)->next ||
6285 &p->se == cfs_rq_of(&p->se)->last))
6288 if (sysctl_sched_migration_cost == -1)
6290 if (sysctl_sched_migration_cost == 0)
6293 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6295 return delta < (s64)sysctl_sched_migration_cost;
6298 #ifdef CONFIG_NUMA_BALANCING
6300 * Returns 1, if task migration degrades locality
6301 * Returns 0, if task migration improves locality i.e migration preferred.
6302 * Returns -1, if task migration is not affected by locality.
6304 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6306 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6307 unsigned long src_faults, dst_faults;
6308 int src_nid, dst_nid;
6310 if (!static_branch_likely(&sched_numa_balancing))
6313 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6316 src_nid = cpu_to_node(env->src_cpu);
6317 dst_nid = cpu_to_node(env->dst_cpu);
6319 if (src_nid == dst_nid)
6322 /* Migrating away from the preferred node is always bad. */
6323 if (src_nid == p->numa_preferred_nid) {
6324 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6330 /* Encourage migration to the preferred node. */
6331 if (dst_nid == p->numa_preferred_nid)
6335 src_faults = group_faults(p, src_nid);
6336 dst_faults = group_faults(p, dst_nid);
6338 src_faults = task_faults(p, src_nid);
6339 dst_faults = task_faults(p, dst_nid);
6342 return dst_faults < src_faults;
6346 static inline int migrate_degrades_locality(struct task_struct *p,
6354 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6357 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6361 lockdep_assert_held(&env->src_rq->lock);
6364 * We do not migrate tasks that are:
6365 * 1) throttled_lb_pair, or
6366 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6367 * 3) running (obviously), or
6368 * 4) are cache-hot on their current CPU.
6370 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6373 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6376 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6378 env->flags |= LBF_SOME_PINNED;
6381 * Remember if this task can be migrated to any other cpu in
6382 * our sched_group. We may want to revisit it if we couldn't
6383 * meet load balance goals by pulling other tasks on src_cpu.
6385 * Also avoid computing new_dst_cpu if we have already computed
6386 * one in current iteration.
6388 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6391 /* Prevent to re-select dst_cpu via env's cpus */
6392 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6393 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6394 env->flags |= LBF_DST_PINNED;
6395 env->new_dst_cpu = cpu;
6403 /* Record that we found atleast one task that could run on dst_cpu */
6404 env->flags &= ~LBF_ALL_PINNED;
6406 if (task_running(env->src_rq, p)) {
6407 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6412 * Aggressive migration if:
6413 * 1) destination numa is preferred
6414 * 2) task is cache cold, or
6415 * 3) too many balance attempts have failed.
6417 tsk_cache_hot = migrate_degrades_locality(p, env);
6418 if (tsk_cache_hot == -1)
6419 tsk_cache_hot = task_hot(p, env);
6421 if (tsk_cache_hot <= 0 ||
6422 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6423 if (tsk_cache_hot == 1) {
6424 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6425 schedstat_inc(p, se.statistics.nr_forced_migrations);
6430 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6435 * detach_task() -- detach the task for the migration specified in env
6437 static void detach_task(struct task_struct *p, struct lb_env *env)
6439 lockdep_assert_held(&env->src_rq->lock);
6441 deactivate_task(env->src_rq, p, 0);
6442 p->on_rq = TASK_ON_RQ_MIGRATING;
6443 set_task_cpu(p, env->dst_cpu);
6447 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6448 * part of active balancing operations within "domain".
6450 * Returns a task if successful and NULL otherwise.
6452 static struct task_struct *detach_one_task(struct lb_env *env)
6454 struct task_struct *p, *n;
6456 lockdep_assert_held(&env->src_rq->lock);
6458 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6459 if (!can_migrate_task(p, env))
6462 detach_task(p, env);
6465 * Right now, this is only the second place where
6466 * lb_gained[env->idle] is updated (other is detach_tasks)
6467 * so we can safely collect stats here rather than
6468 * inside detach_tasks().
6470 schedstat_inc(env->sd, lb_gained[env->idle]);
6476 static const unsigned int sched_nr_migrate_break = 32;
6479 * detach_tasks() -- tries to detach up to imbalance weighted load from
6480 * busiest_rq, as part of a balancing operation within domain "sd".
6482 * Returns number of detached tasks if successful and 0 otherwise.
6484 static int detach_tasks(struct lb_env *env)
6486 struct list_head *tasks = &env->src_rq->cfs_tasks;
6487 struct task_struct *p;
6491 lockdep_assert_held(&env->src_rq->lock);
6493 if (env->imbalance <= 0)
6496 while (!list_empty(tasks)) {
6498 * We don't want to steal all, otherwise we may be treated likewise,
6499 * which could at worst lead to a livelock crash.
6501 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6504 p = list_first_entry(tasks, struct task_struct, se.group_node);
6507 /* We've more or less seen every task there is, call it quits */
6508 if (env->loop > env->loop_max)
6511 /* take a breather every nr_migrate tasks */
6512 if (env->loop > env->loop_break) {
6513 env->loop_break += sched_nr_migrate_break;
6514 env->flags |= LBF_NEED_BREAK;
6518 if (!can_migrate_task(p, env))
6521 load = task_h_load(p);
6523 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6526 if ((load / 2) > env->imbalance)
6529 detach_task(p, env);
6530 list_add(&p->se.group_node, &env->tasks);
6533 env->imbalance -= load;
6535 #ifdef CONFIG_PREEMPT
6537 * NEWIDLE balancing is a source of latency, so preemptible
6538 * kernels will stop after the first task is detached to minimize
6539 * the critical section.
6541 if (env->idle == CPU_NEWLY_IDLE)
6546 * We only want to steal up to the prescribed amount of
6549 if (env->imbalance <= 0)
6554 list_move_tail(&p->se.group_node, tasks);
6558 * Right now, this is one of only two places we collect this stat
6559 * so we can safely collect detach_one_task() stats here rather
6560 * than inside detach_one_task().
6562 schedstat_add(env->sd, lb_gained[env->idle], detached);
6568 * attach_task() -- attach the task detached by detach_task() to its new rq.
6570 static void attach_task(struct rq *rq, struct task_struct *p)
6572 lockdep_assert_held(&rq->lock);
6574 BUG_ON(task_rq(p) != rq);
6575 p->on_rq = TASK_ON_RQ_QUEUED;
6576 activate_task(rq, p, 0);
6577 check_preempt_curr(rq, p, 0);
6581 * attach_one_task() -- attaches the task returned from detach_one_task() to
6584 static void attach_one_task(struct rq *rq, struct task_struct *p)
6586 raw_spin_lock(&rq->lock);
6589 * We want to potentially raise target_cpu's OPP.
6591 update_capacity_of(cpu_of(rq));
6592 raw_spin_unlock(&rq->lock);
6596 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6599 static void attach_tasks(struct lb_env *env)
6601 struct list_head *tasks = &env->tasks;
6602 struct task_struct *p;
6604 raw_spin_lock(&env->dst_rq->lock);
6606 while (!list_empty(tasks)) {
6607 p = list_first_entry(tasks, struct task_struct, se.group_node);
6608 list_del_init(&p->se.group_node);
6610 attach_task(env->dst_rq, p);
6614 * We want to potentially raise env.dst_cpu's OPP.
6616 update_capacity_of(env->dst_cpu);
6618 raw_spin_unlock(&env->dst_rq->lock);
6621 #ifdef CONFIG_FAIR_GROUP_SCHED
6622 static void update_blocked_averages(int cpu)
6624 struct rq *rq = cpu_rq(cpu);
6625 struct cfs_rq *cfs_rq;
6626 unsigned long flags;
6628 raw_spin_lock_irqsave(&rq->lock, flags);
6629 update_rq_clock(rq);
6632 * Iterates the task_group tree in a bottom up fashion, see
6633 * list_add_leaf_cfs_rq() for details.
6635 for_each_leaf_cfs_rq(rq, cfs_rq) {
6636 /* throttled entities do not contribute to load */
6637 if (throttled_hierarchy(cfs_rq))
6640 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6641 update_tg_load_avg(cfs_rq, 0);
6643 raw_spin_unlock_irqrestore(&rq->lock, flags);
6647 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6648 * This needs to be done in a top-down fashion because the load of a child
6649 * group is a fraction of its parents load.
6651 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6653 struct rq *rq = rq_of(cfs_rq);
6654 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6655 unsigned long now = jiffies;
6658 if (cfs_rq->last_h_load_update == now)
6661 cfs_rq->h_load_next = NULL;
6662 for_each_sched_entity(se) {
6663 cfs_rq = cfs_rq_of(se);
6664 cfs_rq->h_load_next = se;
6665 if (cfs_rq->last_h_load_update == now)
6670 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6671 cfs_rq->last_h_load_update = now;
6674 while ((se = cfs_rq->h_load_next) != NULL) {
6675 load = cfs_rq->h_load;
6676 load = div64_ul(load * se->avg.load_avg,
6677 cfs_rq_load_avg(cfs_rq) + 1);
6678 cfs_rq = group_cfs_rq(se);
6679 cfs_rq->h_load = load;
6680 cfs_rq->last_h_load_update = now;
6684 static unsigned long task_h_load(struct task_struct *p)
6686 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6688 update_cfs_rq_h_load(cfs_rq);
6689 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6690 cfs_rq_load_avg(cfs_rq) + 1);
6693 static inline void update_blocked_averages(int cpu)
6695 struct rq *rq = cpu_rq(cpu);
6696 struct cfs_rq *cfs_rq = &rq->cfs;
6697 unsigned long flags;
6699 raw_spin_lock_irqsave(&rq->lock, flags);
6700 update_rq_clock(rq);
6701 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6702 raw_spin_unlock_irqrestore(&rq->lock, flags);
6705 static unsigned long task_h_load(struct task_struct *p)
6707 return p->se.avg.load_avg;
6711 /********** Helpers for find_busiest_group ************************/
6714 * sg_lb_stats - stats of a sched_group required for load_balancing
6716 struct sg_lb_stats {
6717 unsigned long avg_load; /*Avg load across the CPUs of the group */
6718 unsigned long group_load; /* Total load over the CPUs of the group */
6719 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6720 unsigned long load_per_task;
6721 unsigned long group_capacity;
6722 unsigned long group_util; /* Total utilization of the group */
6723 unsigned int sum_nr_running; /* Nr tasks running in the group */
6724 unsigned int idle_cpus;
6725 unsigned int group_weight;
6726 enum group_type group_type;
6727 int group_no_capacity;
6728 int group_misfit_task; /* A cpu has a task too big for its capacity */
6729 #ifdef CONFIG_NUMA_BALANCING
6730 unsigned int nr_numa_running;
6731 unsigned int nr_preferred_running;
6736 * sd_lb_stats - Structure to store the statistics of a sched_domain
6737 * during load balancing.
6739 struct sd_lb_stats {
6740 struct sched_group *busiest; /* Busiest group in this sd */
6741 struct sched_group *local; /* Local group in this sd */
6742 unsigned long total_load; /* Total load of all groups in sd */
6743 unsigned long total_capacity; /* Total capacity of all groups in sd */
6744 unsigned long avg_load; /* Average load across all groups in sd */
6746 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6747 struct sg_lb_stats local_stat; /* Statistics of the local group */
6750 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6753 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6754 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6755 * We must however clear busiest_stat::avg_load because
6756 * update_sd_pick_busiest() reads this before assignment.
6758 *sds = (struct sd_lb_stats){
6762 .total_capacity = 0UL,
6765 .sum_nr_running = 0,
6766 .group_type = group_other,
6772 * get_sd_load_idx - Obtain the load index for a given sched domain.
6773 * @sd: The sched_domain whose load_idx is to be obtained.
6774 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6776 * Return: The load index.
6778 static inline int get_sd_load_idx(struct sched_domain *sd,
6779 enum cpu_idle_type idle)
6785 load_idx = sd->busy_idx;
6788 case CPU_NEWLY_IDLE:
6789 load_idx = sd->newidle_idx;
6792 load_idx = sd->idle_idx;
6799 static unsigned long scale_rt_capacity(int cpu)
6801 struct rq *rq = cpu_rq(cpu);
6802 u64 total, used, age_stamp, avg;
6806 * Since we're reading these variables without serialization make sure
6807 * we read them once before doing sanity checks on them.
6809 age_stamp = READ_ONCE(rq->age_stamp);
6810 avg = READ_ONCE(rq->rt_avg);
6811 delta = __rq_clock_broken(rq) - age_stamp;
6813 if (unlikely(delta < 0))
6816 total = sched_avg_period() + delta;
6818 used = div_u64(avg, total);
6821 * deadline bandwidth is defined at system level so we must
6822 * weight this bandwidth with the max capacity of the system.
6823 * As a reminder, avg_bw is 20bits width and
6824 * scale_cpu_capacity is 10 bits width
6826 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6828 if (likely(used < SCHED_CAPACITY_SCALE))
6829 return SCHED_CAPACITY_SCALE - used;
6834 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6836 raw_spin_lock_init(&mcc->lock);
6841 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6843 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6844 struct sched_group *sdg = sd->groups;
6845 struct max_cpu_capacity *mcc;
6846 unsigned long max_capacity;
6848 unsigned long flags;
6850 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6852 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6854 raw_spin_lock_irqsave(&mcc->lock, flags);
6855 max_capacity = mcc->val;
6856 max_cap_cpu = mcc->cpu;
6858 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6859 (max_capacity < capacity)) {
6860 mcc->val = capacity;
6862 #ifdef CONFIG_SCHED_DEBUG
6863 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6864 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6868 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6870 skip_unlock: __attribute__ ((unused));
6871 capacity *= scale_rt_capacity(cpu);
6872 capacity >>= SCHED_CAPACITY_SHIFT;
6877 cpu_rq(cpu)->cpu_capacity = capacity;
6878 sdg->sgc->capacity = capacity;
6879 sdg->sgc->max_capacity = capacity;
6882 void update_group_capacity(struct sched_domain *sd, int cpu)
6884 struct sched_domain *child = sd->child;
6885 struct sched_group *group, *sdg = sd->groups;
6886 unsigned long capacity, max_capacity;
6887 unsigned long interval;
6889 interval = msecs_to_jiffies(sd->balance_interval);
6890 interval = clamp(interval, 1UL, max_load_balance_interval);
6891 sdg->sgc->next_update = jiffies + interval;
6894 update_cpu_capacity(sd, cpu);
6901 if (child->flags & SD_OVERLAP) {
6903 * SD_OVERLAP domains cannot assume that child groups
6904 * span the current group.
6907 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6908 struct sched_group_capacity *sgc;
6909 struct rq *rq = cpu_rq(cpu);
6912 * build_sched_domains() -> init_sched_groups_capacity()
6913 * gets here before we've attached the domains to the
6916 * Use capacity_of(), which is set irrespective of domains
6917 * in update_cpu_capacity().
6919 * This avoids capacity from being 0 and
6920 * causing divide-by-zero issues on boot.
6922 if (unlikely(!rq->sd)) {
6923 capacity += capacity_of(cpu);
6925 sgc = rq->sd->groups->sgc;
6926 capacity += sgc->capacity;
6929 max_capacity = max(capacity, max_capacity);
6933 * !SD_OVERLAP domains can assume that child groups
6934 * span the current group.
6937 group = child->groups;
6939 struct sched_group_capacity *sgc = group->sgc;
6941 capacity += sgc->capacity;
6942 max_capacity = max(sgc->max_capacity, max_capacity);
6943 group = group->next;
6944 } while (group != child->groups);
6947 sdg->sgc->capacity = capacity;
6948 sdg->sgc->max_capacity = max_capacity;
6952 * Check whether the capacity of the rq has been noticeably reduced by side
6953 * activity. The imbalance_pct is used for the threshold.
6954 * Return true is the capacity is reduced
6957 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6959 return ((rq->cpu_capacity * sd->imbalance_pct) <
6960 (rq->cpu_capacity_orig * 100));
6964 * Group imbalance indicates (and tries to solve) the problem where balancing
6965 * groups is inadequate due to tsk_cpus_allowed() constraints.
6967 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6968 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6971 * { 0 1 2 3 } { 4 5 6 7 }
6974 * If we were to balance group-wise we'd place two tasks in the first group and
6975 * two tasks in the second group. Clearly this is undesired as it will overload
6976 * cpu 3 and leave one of the cpus in the second group unused.
6978 * The current solution to this issue is detecting the skew in the first group
6979 * by noticing the lower domain failed to reach balance and had difficulty
6980 * moving tasks due to affinity constraints.
6982 * When this is so detected; this group becomes a candidate for busiest; see
6983 * update_sd_pick_busiest(). And calculate_imbalance() and
6984 * find_busiest_group() avoid some of the usual balance conditions to allow it
6985 * to create an effective group imbalance.
6987 * This is a somewhat tricky proposition since the next run might not find the
6988 * group imbalance and decide the groups need to be balanced again. A most
6989 * subtle and fragile situation.
6992 static inline int sg_imbalanced(struct sched_group *group)
6994 return group->sgc->imbalance;
6998 * group_has_capacity returns true if the group has spare capacity that could
6999 * be used by some tasks.
7000 * We consider that a group has spare capacity if the * number of task is
7001 * smaller than the number of CPUs or if the utilization is lower than the
7002 * available capacity for CFS tasks.
7003 * For the latter, we use a threshold to stabilize the state, to take into
7004 * account the variance of the tasks' load and to return true if the available
7005 * capacity in meaningful for the load balancer.
7006 * As an example, an available capacity of 1% can appear but it doesn't make
7007 * any benefit for the load balance.
7010 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7012 if (sgs->sum_nr_running < sgs->group_weight)
7015 if ((sgs->group_capacity * 100) >
7016 (sgs->group_util * env->sd->imbalance_pct))
7023 * group_is_overloaded returns true if the group has more tasks than it can
7025 * group_is_overloaded is not equals to !group_has_capacity because a group
7026 * with the exact right number of tasks, has no more spare capacity but is not
7027 * overloaded so both group_has_capacity and group_is_overloaded return
7031 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7033 if (sgs->sum_nr_running <= sgs->group_weight)
7036 if ((sgs->group_capacity * 100) <
7037 (sgs->group_util * env->sd->imbalance_pct))
7045 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7046 * per-cpu capacity than sched_group ref.
7049 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7051 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7052 ref->sgc->max_capacity;
7056 group_type group_classify(struct sched_group *group,
7057 struct sg_lb_stats *sgs)
7059 if (sgs->group_no_capacity)
7060 return group_overloaded;
7062 if (sg_imbalanced(group))
7063 return group_imbalanced;
7065 if (sgs->group_misfit_task)
7066 return group_misfit_task;
7072 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7073 * @env: The load balancing environment.
7074 * @group: sched_group whose statistics are to be updated.
7075 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7076 * @local_group: Does group contain this_cpu.
7077 * @sgs: variable to hold the statistics for this group.
7078 * @overload: Indicate more than one runnable task for any CPU.
7079 * @overutilized: Indicate overutilization for any CPU.
7081 static inline void update_sg_lb_stats(struct lb_env *env,
7082 struct sched_group *group, int load_idx,
7083 int local_group, struct sg_lb_stats *sgs,
7084 bool *overload, bool *overutilized)
7089 memset(sgs, 0, sizeof(*sgs));
7091 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7092 struct rq *rq = cpu_rq(i);
7094 /* Bias balancing toward cpus of our domain */
7096 load = target_load(i, load_idx);
7098 load = source_load(i, load_idx);
7100 sgs->group_load += load;
7101 sgs->group_util += cpu_util(i);
7102 sgs->sum_nr_running += rq->cfs.h_nr_running;
7104 if (rq->nr_running > 1)
7107 #ifdef CONFIG_NUMA_BALANCING
7108 sgs->nr_numa_running += rq->nr_numa_running;
7109 sgs->nr_preferred_running += rq->nr_preferred_running;
7111 sgs->sum_weighted_load += weighted_cpuload(i);
7115 if (cpu_overutilized(i)) {
7116 *overutilized = true;
7117 if (!sgs->group_misfit_task && rq->misfit_task)
7118 sgs->group_misfit_task = capacity_of(i);
7122 /* Adjust by relative CPU capacity of the group */
7123 sgs->group_capacity = group->sgc->capacity;
7124 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7126 if (sgs->sum_nr_running)
7127 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7129 sgs->group_weight = group->group_weight;
7131 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7132 sgs->group_type = group_classify(group, sgs);
7136 * update_sd_pick_busiest - return 1 on busiest group
7137 * @env: The load balancing environment.
7138 * @sds: sched_domain statistics
7139 * @sg: sched_group candidate to be checked for being the busiest
7140 * @sgs: sched_group statistics
7142 * Determine if @sg is a busier group than the previously selected
7145 * Return: %true if @sg is a busier group than the previously selected
7146 * busiest group. %false otherwise.
7148 static bool update_sd_pick_busiest(struct lb_env *env,
7149 struct sd_lb_stats *sds,
7150 struct sched_group *sg,
7151 struct sg_lb_stats *sgs)
7153 struct sg_lb_stats *busiest = &sds->busiest_stat;
7155 if (sgs->group_type > busiest->group_type)
7158 if (sgs->group_type < busiest->group_type)
7162 * Candidate sg doesn't face any serious load-balance problems
7163 * so don't pick it if the local sg is already filled up.
7165 if (sgs->group_type == group_other &&
7166 !group_has_capacity(env, &sds->local_stat))
7169 if (sgs->avg_load <= busiest->avg_load)
7173 * Candiate sg has no more than one task per cpu and has higher
7174 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7176 if (sgs->sum_nr_running <= sgs->group_weight &&
7177 group_smaller_cpu_capacity(sds->local, sg))
7180 /* This is the busiest node in its class. */
7181 if (!(env->sd->flags & SD_ASYM_PACKING))
7185 * ASYM_PACKING needs to move all the work to the lowest
7186 * numbered CPUs in the group, therefore mark all groups
7187 * higher than ourself as busy.
7189 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7193 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7200 #ifdef CONFIG_NUMA_BALANCING
7201 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7203 if (sgs->sum_nr_running > sgs->nr_numa_running)
7205 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7210 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7212 if (rq->nr_running > rq->nr_numa_running)
7214 if (rq->nr_running > rq->nr_preferred_running)
7219 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7224 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7228 #endif /* CONFIG_NUMA_BALANCING */
7231 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7232 * @env: The load balancing environment.
7233 * @sds: variable to hold the statistics for this sched_domain.
7235 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7237 struct sched_domain *child = env->sd->child;
7238 struct sched_group *sg = env->sd->groups;
7239 struct sg_lb_stats tmp_sgs;
7240 int load_idx, prefer_sibling = 0;
7241 bool overload = false, overutilized = false;
7243 if (child && child->flags & SD_PREFER_SIBLING)
7246 load_idx = get_sd_load_idx(env->sd, env->idle);
7249 struct sg_lb_stats *sgs = &tmp_sgs;
7252 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7255 sgs = &sds->local_stat;
7257 if (env->idle != CPU_NEWLY_IDLE ||
7258 time_after_eq(jiffies, sg->sgc->next_update))
7259 update_group_capacity(env->sd, env->dst_cpu);
7262 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7263 &overload, &overutilized);
7269 * In case the child domain prefers tasks go to siblings
7270 * first, lower the sg capacity so that we'll try
7271 * and move all the excess tasks away. We lower the capacity
7272 * of a group only if the local group has the capacity to fit
7273 * these excess tasks. The extra check prevents the case where
7274 * you always pull from the heaviest group when it is already
7275 * under-utilized (possible with a large weight task outweighs
7276 * the tasks on the system).
7278 if (prefer_sibling && sds->local &&
7279 group_has_capacity(env, &sds->local_stat) &&
7280 (sgs->sum_nr_running > 1)) {
7281 sgs->group_no_capacity = 1;
7282 sgs->group_type = group_classify(sg, sgs);
7286 * Ignore task groups with misfit tasks if local group has no
7287 * capacity or if per-cpu capacity isn't higher.
7289 if (sgs->group_type == group_misfit_task &&
7290 (!group_has_capacity(env, &sds->local_stat) ||
7291 !group_smaller_cpu_capacity(sg, sds->local)))
7292 sgs->group_type = group_other;
7294 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7296 sds->busiest_stat = *sgs;
7300 /* Now, start updating sd_lb_stats */
7301 sds->total_load += sgs->group_load;
7302 sds->total_capacity += sgs->group_capacity;
7305 } while (sg != env->sd->groups);
7307 if (env->sd->flags & SD_NUMA)
7308 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7310 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7312 if (!env->sd->parent) {
7313 /* update overload indicator if we are at root domain */
7314 if (env->dst_rq->rd->overload != overload)
7315 env->dst_rq->rd->overload = overload;
7317 /* Update over-utilization (tipping point, U >= 0) indicator */
7318 if (env->dst_rq->rd->overutilized != overutilized)
7319 env->dst_rq->rd->overutilized = overutilized;
7321 if (!env->dst_rq->rd->overutilized && overutilized)
7322 env->dst_rq->rd->overutilized = true;
7327 * check_asym_packing - Check to see if the group is packed into the
7330 * This is primarily intended to used at the sibling level. Some
7331 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7332 * case of POWER7, it can move to lower SMT modes only when higher
7333 * threads are idle. When in lower SMT modes, the threads will
7334 * perform better since they share less core resources. Hence when we
7335 * have idle threads, we want them to be the higher ones.
7337 * This packing function is run on idle threads. It checks to see if
7338 * the busiest CPU in this domain (core in the P7 case) has a higher
7339 * CPU number than the packing function is being run on. Here we are
7340 * assuming lower CPU number will be equivalent to lower a SMT thread
7343 * Return: 1 when packing is required and a task should be moved to
7344 * this CPU. The amount of the imbalance is returned in *imbalance.
7346 * @env: The load balancing environment.
7347 * @sds: Statistics of the sched_domain which is to be packed
7349 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7353 if (!(env->sd->flags & SD_ASYM_PACKING))
7359 busiest_cpu = group_first_cpu(sds->busiest);
7360 if (env->dst_cpu > busiest_cpu)
7363 env->imbalance = DIV_ROUND_CLOSEST(
7364 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7365 SCHED_CAPACITY_SCALE);
7371 * fix_small_imbalance - Calculate the minor imbalance that exists
7372 * amongst the groups of a sched_domain, during
7374 * @env: The load balancing environment.
7375 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7378 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7380 unsigned long tmp, capa_now = 0, capa_move = 0;
7381 unsigned int imbn = 2;
7382 unsigned long scaled_busy_load_per_task;
7383 struct sg_lb_stats *local, *busiest;
7385 local = &sds->local_stat;
7386 busiest = &sds->busiest_stat;
7388 if (!local->sum_nr_running)
7389 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7390 else if (busiest->load_per_task > local->load_per_task)
7393 scaled_busy_load_per_task =
7394 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7395 busiest->group_capacity;
7397 if (busiest->avg_load + scaled_busy_load_per_task >=
7398 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7399 env->imbalance = busiest->load_per_task;
7404 * OK, we don't have enough imbalance to justify moving tasks,
7405 * however we may be able to increase total CPU capacity used by
7409 capa_now += busiest->group_capacity *
7410 min(busiest->load_per_task, busiest->avg_load);
7411 capa_now += local->group_capacity *
7412 min(local->load_per_task, local->avg_load);
7413 capa_now /= SCHED_CAPACITY_SCALE;
7415 /* Amount of load we'd subtract */
7416 if (busiest->avg_load > scaled_busy_load_per_task) {
7417 capa_move += busiest->group_capacity *
7418 min(busiest->load_per_task,
7419 busiest->avg_load - scaled_busy_load_per_task);
7422 /* Amount of load we'd add */
7423 if (busiest->avg_load * busiest->group_capacity <
7424 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7425 tmp = (busiest->avg_load * busiest->group_capacity) /
7426 local->group_capacity;
7428 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7429 local->group_capacity;
7431 capa_move += local->group_capacity *
7432 min(local->load_per_task, local->avg_load + tmp);
7433 capa_move /= SCHED_CAPACITY_SCALE;
7435 /* Move if we gain throughput */
7436 if (capa_move > capa_now)
7437 env->imbalance = busiest->load_per_task;
7441 * calculate_imbalance - Calculate the amount of imbalance present within the
7442 * groups of a given sched_domain during load balance.
7443 * @env: load balance environment
7444 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7446 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7448 unsigned long max_pull, load_above_capacity = ~0UL;
7449 struct sg_lb_stats *local, *busiest;
7451 local = &sds->local_stat;
7452 busiest = &sds->busiest_stat;
7454 if (busiest->group_type == group_imbalanced) {
7456 * In the group_imb case we cannot rely on group-wide averages
7457 * to ensure cpu-load equilibrium, look at wider averages. XXX
7459 busiest->load_per_task =
7460 min(busiest->load_per_task, sds->avg_load);
7464 * In the presence of smp nice balancing, certain scenarios can have
7465 * max load less than avg load(as we skip the groups at or below
7466 * its cpu_capacity, while calculating max_load..)
7468 if (busiest->avg_load <= sds->avg_load ||
7469 local->avg_load >= sds->avg_load) {
7470 /* Misfitting tasks should be migrated in any case */
7471 if (busiest->group_type == group_misfit_task) {
7472 env->imbalance = busiest->group_misfit_task;
7477 * Busiest group is overloaded, local is not, use the spare
7478 * cycles to maximize throughput
7480 if (busiest->group_type == group_overloaded &&
7481 local->group_type <= group_misfit_task) {
7482 env->imbalance = busiest->load_per_task;
7487 return fix_small_imbalance(env, sds);
7491 * If there aren't any idle cpus, avoid creating some.
7493 if (busiest->group_type == group_overloaded &&
7494 local->group_type == group_overloaded) {
7495 load_above_capacity = busiest->sum_nr_running *
7497 if (load_above_capacity > busiest->group_capacity)
7498 load_above_capacity -= busiest->group_capacity;
7500 load_above_capacity = ~0UL;
7504 * We're trying to get all the cpus to the average_load, so we don't
7505 * want to push ourselves above the average load, nor do we wish to
7506 * reduce the max loaded cpu below the average load. At the same time,
7507 * we also don't want to reduce the group load below the group capacity
7508 * (so that we can implement power-savings policies etc). Thus we look
7509 * for the minimum possible imbalance.
7511 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7513 /* How much load to actually move to equalise the imbalance */
7514 env->imbalance = min(
7515 max_pull * busiest->group_capacity,
7516 (sds->avg_load - local->avg_load) * local->group_capacity
7517 ) / SCHED_CAPACITY_SCALE;
7519 /* Boost imbalance to allow misfit task to be balanced. */
7520 if (busiest->group_type == group_misfit_task)
7521 env->imbalance = max_t(long, env->imbalance,
7522 busiest->group_misfit_task);
7525 * if *imbalance is less than the average load per runnable task
7526 * there is no guarantee that any tasks will be moved so we'll have
7527 * a think about bumping its value to force at least one task to be
7530 if (env->imbalance < busiest->load_per_task)
7531 return fix_small_imbalance(env, sds);
7534 /******* find_busiest_group() helpers end here *********************/
7537 * find_busiest_group - Returns the busiest group within the sched_domain
7538 * if there is an imbalance. If there isn't an imbalance, and
7539 * the user has opted for power-savings, it returns a group whose
7540 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7541 * such a group exists.
7543 * Also calculates the amount of weighted load which should be moved
7544 * to restore balance.
7546 * @env: The load balancing environment.
7548 * Return: - The busiest group if imbalance exists.
7549 * - If no imbalance and user has opted for power-savings balance,
7550 * return the least loaded group whose CPUs can be
7551 * put to idle by rebalancing its tasks onto our group.
7553 static struct sched_group *find_busiest_group(struct lb_env *env)
7555 struct sg_lb_stats *local, *busiest;
7556 struct sd_lb_stats sds;
7558 init_sd_lb_stats(&sds);
7561 * Compute the various statistics relavent for load balancing at
7564 update_sd_lb_stats(env, &sds);
7566 if (energy_aware() && !env->dst_rq->rd->overutilized)
7569 local = &sds.local_stat;
7570 busiest = &sds.busiest_stat;
7572 /* ASYM feature bypasses nice load balance check */
7573 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7574 check_asym_packing(env, &sds))
7577 /* There is no busy sibling group to pull tasks from */
7578 if (!sds.busiest || busiest->sum_nr_running == 0)
7581 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7582 / sds.total_capacity;
7585 * If the busiest group is imbalanced the below checks don't
7586 * work because they assume all things are equal, which typically
7587 * isn't true due to cpus_allowed constraints and the like.
7589 if (busiest->group_type == group_imbalanced)
7592 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7593 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7594 busiest->group_no_capacity)
7597 /* Misfitting tasks should be dealt with regardless of the avg load */
7598 if (busiest->group_type == group_misfit_task) {
7603 * If the local group is busier than the selected busiest group
7604 * don't try and pull any tasks.
7606 if (local->avg_load >= busiest->avg_load)
7610 * Don't pull any tasks if this group is already above the domain
7613 if (local->avg_load >= sds.avg_load)
7616 if (env->idle == CPU_IDLE) {
7618 * This cpu is idle. If the busiest group is not overloaded
7619 * and there is no imbalance between this and busiest group
7620 * wrt idle cpus, it is balanced. The imbalance becomes
7621 * significant if the diff is greater than 1 otherwise we
7622 * might end up to just move the imbalance on another group
7624 if ((busiest->group_type != group_overloaded) &&
7625 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7626 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7630 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7631 * imbalance_pct to be conservative.
7633 if (100 * busiest->avg_load <=
7634 env->sd->imbalance_pct * local->avg_load)
7639 env->busiest_group_type = busiest->group_type;
7640 /* Looks like there is an imbalance. Compute it */
7641 calculate_imbalance(env, &sds);
7650 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7652 static struct rq *find_busiest_queue(struct lb_env *env,
7653 struct sched_group *group)
7655 struct rq *busiest = NULL, *rq;
7656 unsigned long busiest_load = 0, busiest_capacity = 1;
7659 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7660 unsigned long capacity, wl;
7664 rt = fbq_classify_rq(rq);
7667 * We classify groups/runqueues into three groups:
7668 * - regular: there are !numa tasks
7669 * - remote: there are numa tasks that run on the 'wrong' node
7670 * - all: there is no distinction
7672 * In order to avoid migrating ideally placed numa tasks,
7673 * ignore those when there's better options.
7675 * If we ignore the actual busiest queue to migrate another
7676 * task, the next balance pass can still reduce the busiest
7677 * queue by moving tasks around inside the node.
7679 * If we cannot move enough load due to this classification
7680 * the next pass will adjust the group classification and
7681 * allow migration of more tasks.
7683 * Both cases only affect the total convergence complexity.
7685 if (rt > env->fbq_type)
7688 capacity = capacity_of(i);
7690 wl = weighted_cpuload(i);
7693 * When comparing with imbalance, use weighted_cpuload()
7694 * which is not scaled with the cpu capacity.
7697 if (rq->nr_running == 1 && wl > env->imbalance &&
7698 !check_cpu_capacity(rq, env->sd) &&
7699 env->busiest_group_type != group_misfit_task)
7703 * For the load comparisons with the other cpu's, consider
7704 * the weighted_cpuload() scaled with the cpu capacity, so
7705 * that the load can be moved away from the cpu that is
7706 * potentially running at a lower capacity.
7708 * Thus we're looking for max(wl_i / capacity_i), crosswise
7709 * multiplication to rid ourselves of the division works out
7710 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7711 * our previous maximum.
7713 if (wl * busiest_capacity > busiest_load * capacity) {
7715 busiest_capacity = capacity;
7724 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7725 * so long as it is large enough.
7727 #define MAX_PINNED_INTERVAL 512
7729 /* Working cpumask for load_balance and load_balance_newidle. */
7730 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7732 static int need_active_balance(struct lb_env *env)
7734 struct sched_domain *sd = env->sd;
7736 if (env->idle == CPU_NEWLY_IDLE) {
7739 * ASYM_PACKING needs to force migrate tasks from busy but
7740 * higher numbered CPUs in order to pack all tasks in the
7741 * lowest numbered CPUs.
7743 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7748 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7749 * It's worth migrating the task if the src_cpu's capacity is reduced
7750 * because of other sched_class or IRQs if more capacity stays
7751 * available on dst_cpu.
7753 if ((env->idle != CPU_NOT_IDLE) &&
7754 (env->src_rq->cfs.h_nr_running == 1)) {
7755 if ((check_cpu_capacity(env->src_rq, sd)) &&
7756 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7760 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7761 env->src_rq->cfs.h_nr_running == 1 &&
7762 cpu_overutilized(env->src_cpu) &&
7763 !cpu_overutilized(env->dst_cpu)) {
7767 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7770 static int active_load_balance_cpu_stop(void *data);
7772 static int should_we_balance(struct lb_env *env)
7774 struct sched_group *sg = env->sd->groups;
7775 struct cpumask *sg_cpus, *sg_mask;
7776 int cpu, balance_cpu = -1;
7779 * In the newly idle case, we will allow all the cpu's
7780 * to do the newly idle load balance.
7782 if (env->idle == CPU_NEWLY_IDLE)
7785 sg_cpus = sched_group_cpus(sg);
7786 sg_mask = sched_group_mask(sg);
7787 /* Try to find first idle cpu */
7788 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7789 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7796 if (balance_cpu == -1)
7797 balance_cpu = group_balance_cpu(sg);
7800 * First idle cpu or the first cpu(busiest) in this sched group
7801 * is eligible for doing load balancing at this and above domains.
7803 return balance_cpu == env->dst_cpu;
7807 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7808 * tasks if there is an imbalance.
7810 static int load_balance(int this_cpu, struct rq *this_rq,
7811 struct sched_domain *sd, enum cpu_idle_type idle,
7812 int *continue_balancing)
7814 int ld_moved, cur_ld_moved, active_balance = 0;
7815 struct sched_domain *sd_parent = sd->parent;
7816 struct sched_group *group;
7818 unsigned long flags;
7819 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7821 struct lb_env env = {
7823 .dst_cpu = this_cpu,
7825 .dst_grpmask = sched_group_cpus(sd->groups),
7827 .loop_break = sched_nr_migrate_break,
7830 .tasks = LIST_HEAD_INIT(env.tasks),
7834 * For NEWLY_IDLE load_balancing, we don't need to consider
7835 * other cpus in our group
7837 if (idle == CPU_NEWLY_IDLE)
7838 env.dst_grpmask = NULL;
7840 cpumask_copy(cpus, cpu_active_mask);
7842 schedstat_inc(sd, lb_count[idle]);
7845 if (!should_we_balance(&env)) {
7846 *continue_balancing = 0;
7850 group = find_busiest_group(&env);
7852 schedstat_inc(sd, lb_nobusyg[idle]);
7856 busiest = find_busiest_queue(&env, group);
7858 schedstat_inc(sd, lb_nobusyq[idle]);
7862 BUG_ON(busiest == env.dst_rq);
7864 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7866 env.src_cpu = busiest->cpu;
7867 env.src_rq = busiest;
7870 if (busiest->nr_running > 1) {
7872 * Attempt to move tasks. If find_busiest_group has found
7873 * an imbalance but busiest->nr_running <= 1, the group is
7874 * still unbalanced. ld_moved simply stays zero, so it is
7875 * correctly treated as an imbalance.
7877 env.flags |= LBF_ALL_PINNED;
7878 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7881 raw_spin_lock_irqsave(&busiest->lock, flags);
7884 * cur_ld_moved - load moved in current iteration
7885 * ld_moved - cumulative load moved across iterations
7887 cur_ld_moved = detach_tasks(&env);
7889 * We want to potentially lower env.src_cpu's OPP.
7892 update_capacity_of(env.src_cpu);
7895 * We've detached some tasks from busiest_rq. Every
7896 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7897 * unlock busiest->lock, and we are able to be sure
7898 * that nobody can manipulate the tasks in parallel.
7899 * See task_rq_lock() family for the details.
7902 raw_spin_unlock(&busiest->lock);
7906 ld_moved += cur_ld_moved;
7909 local_irq_restore(flags);
7911 if (env.flags & LBF_NEED_BREAK) {
7912 env.flags &= ~LBF_NEED_BREAK;
7917 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7918 * us and move them to an alternate dst_cpu in our sched_group
7919 * where they can run. The upper limit on how many times we
7920 * iterate on same src_cpu is dependent on number of cpus in our
7923 * This changes load balance semantics a bit on who can move
7924 * load to a given_cpu. In addition to the given_cpu itself
7925 * (or a ilb_cpu acting on its behalf where given_cpu is
7926 * nohz-idle), we now have balance_cpu in a position to move
7927 * load to given_cpu. In rare situations, this may cause
7928 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7929 * _independently_ and at _same_ time to move some load to
7930 * given_cpu) causing exceess load to be moved to given_cpu.
7931 * This however should not happen so much in practice and
7932 * moreover subsequent load balance cycles should correct the
7933 * excess load moved.
7935 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7937 /* Prevent to re-select dst_cpu via env's cpus */
7938 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7940 env.dst_rq = cpu_rq(env.new_dst_cpu);
7941 env.dst_cpu = env.new_dst_cpu;
7942 env.flags &= ~LBF_DST_PINNED;
7944 env.loop_break = sched_nr_migrate_break;
7947 * Go back to "more_balance" rather than "redo" since we
7948 * need to continue with same src_cpu.
7954 * We failed to reach balance because of affinity.
7957 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7959 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7960 *group_imbalance = 1;
7963 /* All tasks on this runqueue were pinned by CPU affinity */
7964 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7965 cpumask_clear_cpu(cpu_of(busiest), cpus);
7966 if (!cpumask_empty(cpus)) {
7968 env.loop_break = sched_nr_migrate_break;
7971 goto out_all_pinned;
7976 schedstat_inc(sd, lb_failed[idle]);
7978 * Increment the failure counter only on periodic balance.
7979 * We do not want newidle balance, which can be very
7980 * frequent, pollute the failure counter causing
7981 * excessive cache_hot migrations and active balances.
7983 if (idle != CPU_NEWLY_IDLE)
7984 if (env.src_grp_nr_running > 1)
7985 sd->nr_balance_failed++;
7987 if (need_active_balance(&env)) {
7988 raw_spin_lock_irqsave(&busiest->lock, flags);
7990 /* don't kick the active_load_balance_cpu_stop,
7991 * if the curr task on busiest cpu can't be
7994 if (!cpumask_test_cpu(this_cpu,
7995 tsk_cpus_allowed(busiest->curr))) {
7996 raw_spin_unlock_irqrestore(&busiest->lock,
7998 env.flags |= LBF_ALL_PINNED;
7999 goto out_one_pinned;
8003 * ->active_balance synchronizes accesses to
8004 * ->active_balance_work. Once set, it's cleared
8005 * only after active load balance is finished.
8007 if (!busiest->active_balance) {
8008 busiest->active_balance = 1;
8009 busiest->push_cpu = this_cpu;
8012 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8014 if (active_balance) {
8015 stop_one_cpu_nowait(cpu_of(busiest),
8016 active_load_balance_cpu_stop, busiest,
8017 &busiest->active_balance_work);
8021 * We've kicked active balancing, reset the failure
8024 sd->nr_balance_failed = sd->cache_nice_tries+1;
8027 sd->nr_balance_failed = 0;
8029 if (likely(!active_balance)) {
8030 /* We were unbalanced, so reset the balancing interval */
8031 sd->balance_interval = sd->min_interval;
8034 * If we've begun active balancing, start to back off. This
8035 * case may not be covered by the all_pinned logic if there
8036 * is only 1 task on the busy runqueue (because we don't call
8039 if (sd->balance_interval < sd->max_interval)
8040 sd->balance_interval *= 2;
8047 * We reach balance although we may have faced some affinity
8048 * constraints. Clear the imbalance flag if it was set.
8051 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8053 if (*group_imbalance)
8054 *group_imbalance = 0;
8059 * We reach balance because all tasks are pinned at this level so
8060 * we can't migrate them. Let the imbalance flag set so parent level
8061 * can try to migrate them.
8063 schedstat_inc(sd, lb_balanced[idle]);
8065 sd->nr_balance_failed = 0;
8068 /* tune up the balancing interval */
8069 if (((env.flags & LBF_ALL_PINNED) &&
8070 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8071 (sd->balance_interval < sd->max_interval))
8072 sd->balance_interval *= 2;
8079 static inline unsigned long
8080 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8082 unsigned long interval = sd->balance_interval;
8085 interval *= sd->busy_factor;
8087 /* scale ms to jiffies */
8088 interval = msecs_to_jiffies(interval);
8089 interval = clamp(interval, 1UL, max_load_balance_interval);
8095 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8097 unsigned long interval, next;
8099 interval = get_sd_balance_interval(sd, cpu_busy);
8100 next = sd->last_balance + interval;
8102 if (time_after(*next_balance, next))
8103 *next_balance = next;
8107 * idle_balance is called by schedule() if this_cpu is about to become
8108 * idle. Attempts to pull tasks from other CPUs.
8110 static int idle_balance(struct rq *this_rq)
8112 unsigned long next_balance = jiffies + HZ;
8113 int this_cpu = this_rq->cpu;
8114 struct sched_domain *sd;
8115 int pulled_task = 0;
8118 idle_enter_fair(this_rq);
8121 * We must set idle_stamp _before_ calling idle_balance(), such that we
8122 * measure the duration of idle_balance() as idle time.
8124 this_rq->idle_stamp = rq_clock(this_rq);
8126 if (!energy_aware() &&
8127 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8128 !this_rq->rd->overload)) {
8130 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8132 update_next_balance(sd, 0, &next_balance);
8138 raw_spin_unlock(&this_rq->lock);
8140 update_blocked_averages(this_cpu);
8142 for_each_domain(this_cpu, sd) {
8143 int continue_balancing = 1;
8144 u64 t0, domain_cost;
8146 if (!(sd->flags & SD_LOAD_BALANCE))
8149 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8150 update_next_balance(sd, 0, &next_balance);
8154 if (sd->flags & SD_BALANCE_NEWIDLE) {
8155 t0 = sched_clock_cpu(this_cpu);
8157 pulled_task = load_balance(this_cpu, this_rq,
8159 &continue_balancing);
8161 domain_cost = sched_clock_cpu(this_cpu) - t0;
8162 if (domain_cost > sd->max_newidle_lb_cost)
8163 sd->max_newidle_lb_cost = domain_cost;
8165 curr_cost += domain_cost;
8168 update_next_balance(sd, 0, &next_balance);
8171 * Stop searching for tasks to pull if there are
8172 * now runnable tasks on this rq.
8174 if (pulled_task || this_rq->nr_running > 0)
8179 raw_spin_lock(&this_rq->lock);
8181 if (curr_cost > this_rq->max_idle_balance_cost)
8182 this_rq->max_idle_balance_cost = curr_cost;
8185 * While browsing the domains, we released the rq lock, a task could
8186 * have been enqueued in the meantime. Since we're not going idle,
8187 * pretend we pulled a task.
8189 if (this_rq->cfs.h_nr_running && !pulled_task)
8193 /* Move the next balance forward */
8194 if (time_after(this_rq->next_balance, next_balance))
8195 this_rq->next_balance = next_balance;
8197 /* Is there a task of a high priority class? */
8198 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8202 idle_exit_fair(this_rq);
8203 this_rq->idle_stamp = 0;
8210 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8211 * running tasks off the busiest CPU onto idle CPUs. It requires at
8212 * least 1 task to be running on each physical CPU where possible, and
8213 * avoids physical / logical imbalances.
8215 static int active_load_balance_cpu_stop(void *data)
8217 struct rq *busiest_rq = data;
8218 int busiest_cpu = cpu_of(busiest_rq);
8219 int target_cpu = busiest_rq->push_cpu;
8220 struct rq *target_rq = cpu_rq(target_cpu);
8221 struct sched_domain *sd;
8222 struct task_struct *p = NULL;
8224 raw_spin_lock_irq(&busiest_rq->lock);
8226 /* make sure the requested cpu hasn't gone down in the meantime */
8227 if (unlikely(busiest_cpu != smp_processor_id() ||
8228 !busiest_rq->active_balance))
8231 /* Is there any task to move? */
8232 if (busiest_rq->nr_running <= 1)
8236 * This condition is "impossible", if it occurs
8237 * we need to fix it. Originally reported by
8238 * Bjorn Helgaas on a 128-cpu setup.
8240 BUG_ON(busiest_rq == target_rq);
8242 /* Search for an sd spanning us and the target CPU. */
8244 for_each_domain(target_cpu, sd) {
8245 if ((sd->flags & SD_LOAD_BALANCE) &&
8246 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8251 struct lb_env env = {
8253 .dst_cpu = target_cpu,
8254 .dst_rq = target_rq,
8255 .src_cpu = busiest_rq->cpu,
8256 .src_rq = busiest_rq,
8260 schedstat_inc(sd, alb_count);
8262 p = detach_one_task(&env);
8264 schedstat_inc(sd, alb_pushed);
8266 * We want to potentially lower env.src_cpu's OPP.
8268 update_capacity_of(env.src_cpu);
8271 schedstat_inc(sd, alb_failed);
8275 busiest_rq->active_balance = 0;
8276 raw_spin_unlock(&busiest_rq->lock);
8279 attach_one_task(target_rq, p);
8286 static inline int on_null_domain(struct rq *rq)
8288 return unlikely(!rcu_dereference_sched(rq->sd));
8291 #ifdef CONFIG_NO_HZ_COMMON
8293 * idle load balancing details
8294 * - When one of the busy CPUs notice that there may be an idle rebalancing
8295 * needed, they will kick the idle load balancer, which then does idle
8296 * load balancing for all the idle CPUs.
8299 cpumask_var_t idle_cpus_mask;
8301 unsigned long next_balance; /* in jiffy units */
8302 } nohz ____cacheline_aligned;
8304 static inline int find_new_ilb(void)
8306 int ilb = cpumask_first(nohz.idle_cpus_mask);
8308 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8315 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8316 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8317 * CPU (if there is one).
8319 static void nohz_balancer_kick(void)
8323 nohz.next_balance++;
8325 ilb_cpu = find_new_ilb();
8327 if (ilb_cpu >= nr_cpu_ids)
8330 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8333 * Use smp_send_reschedule() instead of resched_cpu().
8334 * This way we generate a sched IPI on the target cpu which
8335 * is idle. And the softirq performing nohz idle load balance
8336 * will be run before returning from the IPI.
8338 smp_send_reschedule(ilb_cpu);
8342 static inline void nohz_balance_exit_idle(int cpu)
8344 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8346 * Completely isolated CPUs don't ever set, so we must test.
8348 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8349 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8350 atomic_dec(&nohz.nr_cpus);
8352 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8356 static inline void set_cpu_sd_state_busy(void)
8358 struct sched_domain *sd;
8359 int cpu = smp_processor_id();
8362 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8364 if (!sd || !sd->nohz_idle)
8368 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8373 void set_cpu_sd_state_idle(void)
8375 struct sched_domain *sd;
8376 int cpu = smp_processor_id();
8379 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8381 if (!sd || sd->nohz_idle)
8385 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8391 * This routine will record that the cpu is going idle with tick stopped.
8392 * This info will be used in performing idle load balancing in the future.
8394 void nohz_balance_enter_idle(int cpu)
8397 * If this cpu is going down, then nothing needs to be done.
8399 if (!cpu_active(cpu))
8402 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8406 * If we're a completely isolated CPU, we don't play.
8408 if (on_null_domain(cpu_rq(cpu)))
8411 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8412 atomic_inc(&nohz.nr_cpus);
8413 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8416 static int sched_ilb_notifier(struct notifier_block *nfb,
8417 unsigned long action, void *hcpu)
8419 switch (action & ~CPU_TASKS_FROZEN) {
8421 nohz_balance_exit_idle(smp_processor_id());
8429 static DEFINE_SPINLOCK(balancing);
8432 * Scale the max load_balance interval with the number of CPUs in the system.
8433 * This trades load-balance latency on larger machines for less cross talk.
8435 void update_max_interval(void)
8437 max_load_balance_interval = HZ*num_online_cpus()/10;
8441 * It checks each scheduling domain to see if it is due to be balanced,
8442 * and initiates a balancing operation if so.
8444 * Balancing parameters are set up in init_sched_domains.
8446 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8448 int continue_balancing = 1;
8450 unsigned long interval;
8451 struct sched_domain *sd;
8452 /* Earliest time when we have to do rebalance again */
8453 unsigned long next_balance = jiffies + 60*HZ;
8454 int update_next_balance = 0;
8455 int need_serialize, need_decay = 0;
8458 update_blocked_averages(cpu);
8461 for_each_domain(cpu, sd) {
8463 * Decay the newidle max times here because this is a regular
8464 * visit to all the domains. Decay ~1% per second.
8466 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8467 sd->max_newidle_lb_cost =
8468 (sd->max_newidle_lb_cost * 253) / 256;
8469 sd->next_decay_max_lb_cost = jiffies + HZ;
8472 max_cost += sd->max_newidle_lb_cost;
8474 if (!(sd->flags & SD_LOAD_BALANCE))
8478 * Stop the load balance at this level. There is another
8479 * CPU in our sched group which is doing load balancing more
8482 if (!continue_balancing) {
8488 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8490 need_serialize = sd->flags & SD_SERIALIZE;
8491 if (need_serialize) {
8492 if (!spin_trylock(&balancing))
8496 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8497 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8499 * The LBF_DST_PINNED logic could have changed
8500 * env->dst_cpu, so we can't know our idle
8501 * state even if we migrated tasks. Update it.
8503 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8505 sd->last_balance = jiffies;
8506 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8509 spin_unlock(&balancing);
8511 if (time_after(next_balance, sd->last_balance + interval)) {
8512 next_balance = sd->last_balance + interval;
8513 update_next_balance = 1;
8518 * Ensure the rq-wide value also decays but keep it at a
8519 * reasonable floor to avoid funnies with rq->avg_idle.
8521 rq->max_idle_balance_cost =
8522 max((u64)sysctl_sched_migration_cost, max_cost);
8527 * next_balance will be updated only when there is a need.
8528 * When the cpu is attached to null domain for ex, it will not be
8531 if (likely(update_next_balance)) {
8532 rq->next_balance = next_balance;
8534 #ifdef CONFIG_NO_HZ_COMMON
8536 * If this CPU has been elected to perform the nohz idle
8537 * balance. Other idle CPUs have already rebalanced with
8538 * nohz_idle_balance() and nohz.next_balance has been
8539 * updated accordingly. This CPU is now running the idle load
8540 * balance for itself and we need to update the
8541 * nohz.next_balance accordingly.
8543 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8544 nohz.next_balance = rq->next_balance;
8549 #ifdef CONFIG_NO_HZ_COMMON
8551 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8552 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8554 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8556 int this_cpu = this_rq->cpu;
8559 /* Earliest time when we have to do rebalance again */
8560 unsigned long next_balance = jiffies + 60*HZ;
8561 int update_next_balance = 0;
8563 if (idle != CPU_IDLE ||
8564 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8567 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8568 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8572 * If this cpu gets work to do, stop the load balancing
8573 * work being done for other cpus. Next load
8574 * balancing owner will pick it up.
8579 rq = cpu_rq(balance_cpu);
8582 * If time for next balance is due,
8585 if (time_after_eq(jiffies, rq->next_balance)) {
8586 raw_spin_lock_irq(&rq->lock);
8587 update_rq_clock(rq);
8588 update_idle_cpu_load(rq);
8589 raw_spin_unlock_irq(&rq->lock);
8590 rebalance_domains(rq, CPU_IDLE);
8593 if (time_after(next_balance, rq->next_balance)) {
8594 next_balance = rq->next_balance;
8595 update_next_balance = 1;
8600 * next_balance will be updated only when there is a need.
8601 * When the CPU is attached to null domain for ex, it will not be
8604 if (likely(update_next_balance))
8605 nohz.next_balance = next_balance;
8607 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8611 * Current heuristic for kicking the idle load balancer in the presence
8612 * of an idle cpu in the system.
8613 * - This rq has more than one task.
8614 * - This rq has at least one CFS task and the capacity of the CPU is
8615 * significantly reduced because of RT tasks or IRQs.
8616 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8617 * multiple busy cpu.
8618 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8619 * domain span are idle.
8621 static inline bool nohz_kick_needed(struct rq *rq)
8623 unsigned long now = jiffies;
8624 struct sched_domain *sd;
8625 struct sched_group_capacity *sgc;
8626 int nr_busy, cpu = rq->cpu;
8629 if (unlikely(rq->idle_balance))
8633 * We may be recently in ticked or tickless idle mode. At the first
8634 * busy tick after returning from idle, we will update the busy stats.
8636 set_cpu_sd_state_busy();
8637 nohz_balance_exit_idle(cpu);
8640 * None are in tickless mode and hence no need for NOHZ idle load
8643 if (likely(!atomic_read(&nohz.nr_cpus)))
8646 if (time_before(now, nohz.next_balance))
8649 if (rq->nr_running >= 2 &&
8650 (!energy_aware() || cpu_overutilized(cpu)))
8654 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8655 if (sd && !energy_aware()) {
8656 sgc = sd->groups->sgc;
8657 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8666 sd = rcu_dereference(rq->sd);
8668 if ((rq->cfs.h_nr_running >= 1) &&
8669 check_cpu_capacity(rq, sd)) {
8675 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8676 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8677 sched_domain_span(sd)) < cpu)) {
8687 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8691 * run_rebalance_domains is triggered when needed from the scheduler tick.
8692 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8694 static void run_rebalance_domains(struct softirq_action *h)
8696 struct rq *this_rq = this_rq();
8697 enum cpu_idle_type idle = this_rq->idle_balance ?
8698 CPU_IDLE : CPU_NOT_IDLE;
8701 * If this cpu has a pending nohz_balance_kick, then do the
8702 * balancing on behalf of the other idle cpus whose ticks are
8703 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8704 * give the idle cpus a chance to load balance. Else we may
8705 * load balance only within the local sched_domain hierarchy
8706 * and abort nohz_idle_balance altogether if we pull some load.
8708 nohz_idle_balance(this_rq, idle);
8709 rebalance_domains(this_rq, idle);
8713 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8715 void trigger_load_balance(struct rq *rq)
8717 /* Don't need to rebalance while attached to NULL domain */
8718 if (unlikely(on_null_domain(rq)))
8721 if (time_after_eq(jiffies, rq->next_balance))
8722 raise_softirq(SCHED_SOFTIRQ);
8723 #ifdef CONFIG_NO_HZ_COMMON
8724 if (nohz_kick_needed(rq))
8725 nohz_balancer_kick();
8729 static void rq_online_fair(struct rq *rq)
8733 update_runtime_enabled(rq);
8736 static void rq_offline_fair(struct rq *rq)
8740 /* Ensure any throttled groups are reachable by pick_next_task */
8741 unthrottle_offline_cfs_rqs(rq);
8744 #endif /* CONFIG_SMP */
8747 * scheduler tick hitting a task of our scheduling class:
8749 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8751 struct cfs_rq *cfs_rq;
8752 struct sched_entity *se = &curr->se;
8754 for_each_sched_entity(se) {
8755 cfs_rq = cfs_rq_of(se);
8756 entity_tick(cfs_rq, se, queued);
8759 if (static_branch_unlikely(&sched_numa_balancing))
8760 task_tick_numa(rq, curr);
8763 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8764 rq->rd->overutilized = true;
8766 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8772 * called on fork with the child task as argument from the parent's context
8773 * - child not yet on the tasklist
8774 * - preemption disabled
8776 static void task_fork_fair(struct task_struct *p)
8778 struct cfs_rq *cfs_rq;
8779 struct sched_entity *se = &p->se, *curr;
8780 int this_cpu = smp_processor_id();
8781 struct rq *rq = this_rq();
8782 unsigned long flags;
8784 raw_spin_lock_irqsave(&rq->lock, flags);
8786 update_rq_clock(rq);
8788 cfs_rq = task_cfs_rq(current);
8789 curr = cfs_rq->curr;
8792 * Not only the cpu but also the task_group of the parent might have
8793 * been changed after parent->se.parent,cfs_rq were copied to
8794 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8795 * of child point to valid ones.
8798 __set_task_cpu(p, this_cpu);
8801 update_curr(cfs_rq);
8804 se->vruntime = curr->vruntime;
8805 place_entity(cfs_rq, se, 1);
8807 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8809 * Upon rescheduling, sched_class::put_prev_task() will place
8810 * 'current' within the tree based on its new key value.
8812 swap(curr->vruntime, se->vruntime);
8816 se->vruntime -= cfs_rq->min_vruntime;
8818 raw_spin_unlock_irqrestore(&rq->lock, flags);
8822 * Priority of the task has changed. Check to see if we preempt
8826 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8828 if (!task_on_rq_queued(p))
8832 * Reschedule if we are currently running on this runqueue and
8833 * our priority decreased, or if we are not currently running on
8834 * this runqueue and our priority is higher than the current's
8836 if (rq->curr == p) {
8837 if (p->prio > oldprio)
8840 check_preempt_curr(rq, p, 0);
8843 static inline bool vruntime_normalized(struct task_struct *p)
8845 struct sched_entity *se = &p->se;
8848 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8849 * the dequeue_entity(.flags=0) will already have normalized the
8856 * When !on_rq, vruntime of the task has usually NOT been normalized.
8857 * But there are some cases where it has already been normalized:
8859 * - A forked child which is waiting for being woken up by
8860 * wake_up_new_task().
8861 * - A task which has been woken up by try_to_wake_up() and
8862 * waiting for actually being woken up by sched_ttwu_pending().
8864 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8870 static void detach_task_cfs_rq(struct task_struct *p)
8872 struct sched_entity *se = &p->se;
8873 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8875 if (!vruntime_normalized(p)) {
8877 * Fix up our vruntime so that the current sleep doesn't
8878 * cause 'unlimited' sleep bonus.
8880 place_entity(cfs_rq, se, 0);
8881 se->vruntime -= cfs_rq->min_vruntime;
8884 /* Catch up with the cfs_rq and remove our load when we leave */
8885 detach_entity_load_avg(cfs_rq, se);
8888 static void attach_task_cfs_rq(struct task_struct *p)
8890 struct sched_entity *se = &p->se;
8891 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8893 #ifdef CONFIG_FAIR_GROUP_SCHED
8895 * Since the real-depth could have been changed (only FAIR
8896 * class maintain depth value), reset depth properly.
8898 se->depth = se->parent ? se->parent->depth + 1 : 0;
8901 /* Synchronize task with its cfs_rq */
8902 attach_entity_load_avg(cfs_rq, se);
8904 if (!vruntime_normalized(p))
8905 se->vruntime += cfs_rq->min_vruntime;
8908 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8910 detach_task_cfs_rq(p);
8913 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8915 attach_task_cfs_rq(p);
8917 if (task_on_rq_queued(p)) {
8919 * We were most likely switched from sched_rt, so
8920 * kick off the schedule if running, otherwise just see
8921 * if we can still preempt the current task.
8926 check_preempt_curr(rq, p, 0);
8930 /* Account for a task changing its policy or group.
8932 * This routine is mostly called to set cfs_rq->curr field when a task
8933 * migrates between groups/classes.
8935 static void set_curr_task_fair(struct rq *rq)
8937 struct sched_entity *se = &rq->curr->se;
8939 for_each_sched_entity(se) {
8940 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8942 set_next_entity(cfs_rq, se);
8943 /* ensure bandwidth has been allocated on our new cfs_rq */
8944 account_cfs_rq_runtime(cfs_rq, 0);
8948 void init_cfs_rq(struct cfs_rq *cfs_rq)
8950 cfs_rq->tasks_timeline = RB_ROOT;
8951 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8952 #ifndef CONFIG_64BIT
8953 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8956 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8957 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8961 #ifdef CONFIG_FAIR_GROUP_SCHED
8962 static void task_move_group_fair(struct task_struct *p)
8964 detach_task_cfs_rq(p);
8965 set_task_rq(p, task_cpu(p));
8968 /* Tell se's cfs_rq has been changed -- migrated */
8969 p->se.avg.last_update_time = 0;
8971 attach_task_cfs_rq(p);
8974 void free_fair_sched_group(struct task_group *tg)
8978 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8980 for_each_possible_cpu(i) {
8982 kfree(tg->cfs_rq[i]);
8985 remove_entity_load_avg(tg->se[i]);
8994 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8996 struct cfs_rq *cfs_rq;
8997 struct sched_entity *se;
9000 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9003 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9007 tg->shares = NICE_0_LOAD;
9009 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9011 for_each_possible_cpu(i) {
9012 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9013 GFP_KERNEL, cpu_to_node(i));
9017 se = kzalloc_node(sizeof(struct sched_entity),
9018 GFP_KERNEL, cpu_to_node(i));
9022 init_cfs_rq(cfs_rq);
9023 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9024 init_entity_runnable_average(se);
9035 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9037 struct rq *rq = cpu_rq(cpu);
9038 unsigned long flags;
9041 * Only empty task groups can be destroyed; so we can speculatively
9042 * check on_list without danger of it being re-added.
9044 if (!tg->cfs_rq[cpu]->on_list)
9047 raw_spin_lock_irqsave(&rq->lock, flags);
9048 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9049 raw_spin_unlock_irqrestore(&rq->lock, flags);
9052 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9053 struct sched_entity *se, int cpu,
9054 struct sched_entity *parent)
9056 struct rq *rq = cpu_rq(cpu);
9060 init_cfs_rq_runtime(cfs_rq);
9062 tg->cfs_rq[cpu] = cfs_rq;
9065 /* se could be NULL for root_task_group */
9070 se->cfs_rq = &rq->cfs;
9073 se->cfs_rq = parent->my_q;
9074 se->depth = parent->depth + 1;
9078 /* guarantee group entities always have weight */
9079 update_load_set(&se->load, NICE_0_LOAD);
9080 se->parent = parent;
9083 static DEFINE_MUTEX(shares_mutex);
9085 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9088 unsigned long flags;
9091 * We can't change the weight of the root cgroup.
9096 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9098 mutex_lock(&shares_mutex);
9099 if (tg->shares == shares)
9102 tg->shares = shares;
9103 for_each_possible_cpu(i) {
9104 struct rq *rq = cpu_rq(i);
9105 struct sched_entity *se;
9108 /* Propagate contribution to hierarchy */
9109 raw_spin_lock_irqsave(&rq->lock, flags);
9111 /* Possible calls to update_curr() need rq clock */
9112 update_rq_clock(rq);
9113 for_each_sched_entity(se)
9114 update_cfs_shares(group_cfs_rq(se));
9115 raw_spin_unlock_irqrestore(&rq->lock, flags);
9119 mutex_unlock(&shares_mutex);
9122 #else /* CONFIG_FAIR_GROUP_SCHED */
9124 void free_fair_sched_group(struct task_group *tg) { }
9126 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9131 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9133 #endif /* CONFIG_FAIR_GROUP_SCHED */
9136 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9138 struct sched_entity *se = &task->se;
9139 unsigned int rr_interval = 0;
9142 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9145 if (rq->cfs.load.weight)
9146 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9152 * All the scheduling class methods:
9154 const struct sched_class fair_sched_class = {
9155 .next = &idle_sched_class,
9156 .enqueue_task = enqueue_task_fair,
9157 .dequeue_task = dequeue_task_fair,
9158 .yield_task = yield_task_fair,
9159 .yield_to_task = yield_to_task_fair,
9161 .check_preempt_curr = check_preempt_wakeup,
9163 .pick_next_task = pick_next_task_fair,
9164 .put_prev_task = put_prev_task_fair,
9167 .select_task_rq = select_task_rq_fair,
9168 .migrate_task_rq = migrate_task_rq_fair,
9170 .rq_online = rq_online_fair,
9171 .rq_offline = rq_offline_fair,
9173 .task_waking = task_waking_fair,
9174 .task_dead = task_dead_fair,
9175 .set_cpus_allowed = set_cpus_allowed_common,
9178 .set_curr_task = set_curr_task_fair,
9179 .task_tick = task_tick_fair,
9180 .task_fork = task_fork_fair,
9182 .prio_changed = prio_changed_fair,
9183 .switched_from = switched_from_fair,
9184 .switched_to = switched_to_fair,
9186 .get_rr_interval = get_rr_interval_fair,
9188 .update_curr = update_curr_fair,
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9191 .task_move_group = task_move_group_fair,
9195 #ifdef CONFIG_SCHED_DEBUG
9196 void print_cfs_stats(struct seq_file *m, int cpu)
9198 struct cfs_rq *cfs_rq;
9201 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9202 print_cfs_rq(m, cpu, cfs_rq);
9206 #ifdef CONFIG_NUMA_BALANCING
9207 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9210 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9212 for_each_online_node(node) {
9213 if (p->numa_faults) {
9214 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9215 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9217 if (p->numa_group) {
9218 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9219 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9221 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9224 #endif /* CONFIG_NUMA_BALANCING */
9225 #endif /* CONFIG_SCHED_DEBUG */
9227 __init void init_sched_fair_class(void)
9230 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9232 #ifdef CONFIG_NO_HZ_COMMON
9233 nohz.next_balance = jiffies;
9234 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9235 cpu_notifier(sched_ilb_notifier, 0);