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 <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/random.h>
30 #include <linux/mempolicy.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 * Increase the granularity value when there are more CPUs,
118 * because with more CPUs the 'effective latency' as visible
119 * to users decreases. But the relationship is not linear,
120 * so pick a second-best guess by going with the log2 of the
123 * This idea comes from the SD scheduler of Con Kolivas:
125 static int get_update_sysctl_factor(void)
127 unsigned int cpus = min_t(int, num_online_cpus(), 8);
130 switch (sysctl_sched_tunable_scaling) {
131 case SCHED_TUNABLESCALING_NONE:
134 case SCHED_TUNABLESCALING_LINEAR:
137 case SCHED_TUNABLESCALING_LOG:
139 factor = 1 + ilog2(cpus);
146 static void update_sysctl(void)
148 unsigned int factor = get_update_sysctl_factor();
150 #define SET_SYSCTL(name) \
151 (sysctl_##name = (factor) * normalized_sysctl_##name)
152 SET_SYSCTL(sched_min_granularity);
153 SET_SYSCTL(sched_latency);
154 SET_SYSCTL(sched_wakeup_granularity);
158 void sched_init_granularity(void)
163 #if BITS_PER_LONG == 32
164 # define WMULT_CONST (~0UL)
166 # define WMULT_CONST (1UL << 32)
169 #define WMULT_SHIFT 32
172 * Shift right and round:
174 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
177 * delta *= weight / lw
180 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
181 struct load_weight *lw)
186 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
187 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
188 * 2^SCHED_LOAD_RESOLUTION.
190 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
191 tmp = (u64)delta_exec * scale_load_down(weight);
193 tmp = (u64)delta_exec;
195 if (!lw->inv_weight) {
196 unsigned long w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * Check whether we'd overflow the 64-bit multiplication:
209 if (unlikely(tmp > WMULT_CONST))
210 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
213 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
215 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
219 const struct sched_class fair_sched_class;
221 /**************************************************************
222 * CFS operations on generic schedulable entities:
225 #ifdef CONFIG_FAIR_GROUP_SCHED
227 /* cpu runqueue to which this cfs_rq is attached */
228 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
233 /* An entity is a task if it doesn't "own" a runqueue */
234 #define entity_is_task(se) (!se->my_q)
236 static inline struct task_struct *task_of(struct sched_entity *se)
238 #ifdef CONFIG_SCHED_DEBUG
239 WARN_ON_ONCE(!entity_is_task(se));
241 return container_of(se, struct task_struct, se);
244 /* Walk up scheduling entities hierarchy */
245 #define for_each_sched_entity(se) \
246 for (; se; se = se->parent)
248 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
253 /* runqueue on which this entity is (to be) queued */
254 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
259 /* runqueue "owned" by this group */
260 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
265 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
267 if (!cfs_rq->on_list) {
269 * Ensure we either appear before our parent (if already
270 * enqueued) or force our parent to appear after us when it is
271 * enqueued. The fact that we always enqueue bottom-up
272 * reduces this to two cases.
274 if (cfs_rq->tg->parent &&
275 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
276 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
277 &rq_of(cfs_rq)->leaf_cfs_rq_list);
279 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
280 &rq_of(cfs_rq)->leaf_cfs_rq_list);
287 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 if (cfs_rq->on_list) {
290 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
295 /* Iterate thr' all leaf cfs_rq's on a runqueue */
296 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
297 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
299 /* Do the two (enqueued) entities belong to the same group ? */
301 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 if (se->cfs_rq == pse->cfs_rq)
309 static inline struct sched_entity *parent_entity(struct sched_entity *se)
314 /* return depth at which a sched entity is present in the hierarchy */
315 static inline int depth_se(struct sched_entity *se)
319 for_each_sched_entity(se)
326 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
328 int se_depth, pse_depth;
331 * preemption test can be made between sibling entities who are in the
332 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
333 * both tasks until we find their ancestors who are siblings of common
337 /* First walk up until both entities are at same depth */
338 se_depth = depth_se(*se);
339 pse_depth = depth_se(*pse);
341 while (se_depth > pse_depth) {
343 *se = parent_entity(*se);
346 while (pse_depth > se_depth) {
348 *pse = parent_entity(*pse);
351 while (!is_same_group(*se, *pse)) {
352 *se = parent_entity(*se);
353 *pse = parent_entity(*pse);
357 #else /* !CONFIG_FAIR_GROUP_SCHED */
359 static inline struct task_struct *task_of(struct sched_entity *se)
361 return container_of(se, struct task_struct, se);
364 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
366 return container_of(cfs_rq, struct rq, cfs);
369 #define entity_is_task(se) 1
371 #define for_each_sched_entity(se) \
372 for (; se; se = NULL)
374 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
376 return &task_rq(p)->cfs;
379 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
381 struct task_struct *p = task_of(se);
382 struct rq *rq = task_rq(p);
387 /* runqueue "owned" by this group */
388 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
393 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
397 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
401 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
402 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
405 is_same_group(struct sched_entity *se, struct sched_entity *pse)
410 static inline struct sched_entity *parent_entity(struct sched_entity *se)
416 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
420 #endif /* CONFIG_FAIR_GROUP_SCHED */
422 static __always_inline
423 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
425 /**************************************************************
426 * Scheduling class tree data structure manipulation methods:
429 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
431 s64 delta = (s64)(vruntime - min_vruntime);
433 min_vruntime = vruntime;
438 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
440 s64 delta = (s64)(vruntime - min_vruntime);
442 min_vruntime = vruntime;
447 static inline int entity_before(struct sched_entity *a,
448 struct sched_entity *b)
450 return (s64)(a->vruntime - b->vruntime) < 0;
453 static void update_min_vruntime(struct cfs_rq *cfs_rq)
455 u64 vruntime = cfs_rq->min_vruntime;
458 vruntime = cfs_rq->curr->vruntime;
460 if (cfs_rq->rb_leftmost) {
461 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
466 vruntime = se->vruntime;
468 vruntime = min_vruntime(vruntime, se->vruntime);
471 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
474 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
479 * Enqueue an entity into the rb-tree:
481 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
483 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
484 struct rb_node *parent = NULL;
485 struct sched_entity *entry;
489 * Find the right place in the rbtree:
493 entry = rb_entry(parent, struct sched_entity, run_node);
495 * We dont care about collisions. Nodes with
496 * the same key stay together.
498 if (entity_before(se, entry)) {
499 link = &parent->rb_left;
501 link = &parent->rb_right;
507 * Maintain a cache of leftmost tree entries (it is frequently
511 cfs_rq->rb_leftmost = &se->run_node;
513 rb_link_node(&se->run_node, parent, link);
514 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
517 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
519 if (cfs_rq->rb_leftmost == &se->run_node) {
520 struct rb_node *next_node;
522 next_node = rb_next(&se->run_node);
523 cfs_rq->rb_leftmost = next_node;
526 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
529 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
531 struct rb_node *left = cfs_rq->rb_leftmost;
536 return rb_entry(left, struct sched_entity, run_node);
539 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
541 struct rb_node *next = rb_next(&se->run_node);
546 return rb_entry(next, struct sched_entity, run_node);
549 #ifdef CONFIG_SCHED_DEBUG
550 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
552 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
557 return rb_entry(last, struct sched_entity, run_node);
560 /**************************************************************
561 * Scheduling class statistics methods:
564 int sched_proc_update_handler(struct ctl_table *table, int write,
565 void __user *buffer, size_t *lenp,
568 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
569 int factor = get_update_sysctl_factor();
574 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
575 sysctl_sched_min_granularity);
577 #define WRT_SYSCTL(name) \
578 (normalized_sysctl_##name = sysctl_##name / (factor))
579 WRT_SYSCTL(sched_min_granularity);
580 WRT_SYSCTL(sched_latency);
581 WRT_SYSCTL(sched_wakeup_granularity);
591 static inline unsigned long
592 calc_delta_fair(unsigned long delta, struct sched_entity *se)
594 if (unlikely(se->load.weight != NICE_0_LOAD))
595 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
601 * The idea is to set a period in which each task runs once.
603 * When there are too many tasks (sched_nr_latency) we have to stretch
604 * this period because otherwise the slices get too small.
606 * p = (nr <= nl) ? l : l*nr/nl
608 static u64 __sched_period(unsigned long nr_running)
610 u64 period = sysctl_sched_latency;
611 unsigned long nr_latency = sched_nr_latency;
613 if (unlikely(nr_running > nr_latency)) {
614 period = sysctl_sched_min_granularity;
615 period *= nr_running;
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_mine(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);
659 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
660 static void update_cfs_shares(struct cfs_rq *cfs_rq);
663 * Update the current task's runtime statistics. Skip current tasks that
664 * are not in our scheduling class.
667 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
668 unsigned long delta_exec)
670 unsigned long delta_exec_weighted;
672 schedstat_set(curr->statistics.exec_max,
673 max((u64)delta_exec, curr->statistics.exec_max));
675 curr->sum_exec_runtime += delta_exec;
676 schedstat_add(cfs_rq, exec_clock, delta_exec);
677 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
679 curr->vruntime += delta_exec_weighted;
680 update_min_vruntime(cfs_rq);
682 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
683 cfs_rq->load_unacc_exec_time += delta_exec;
687 static void update_curr(struct cfs_rq *cfs_rq)
689 struct sched_entity *curr = cfs_rq->curr;
690 u64 now = rq_of(cfs_rq)->clock_task;
691 unsigned long delta_exec;
697 * Get the amount of time the current task was running
698 * since the last time we changed load (this cannot
699 * overflow on 32 bits):
701 delta_exec = (unsigned long)(now - curr->exec_start);
705 __update_curr(cfs_rq, curr, delta_exec);
706 curr->exec_start = now;
708 if (entity_is_task(curr)) {
709 struct task_struct *curtask = task_of(curr);
711 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
712 cpuacct_charge(curtask, delta_exec);
713 account_group_exec_runtime(curtask, delta_exec);
716 account_cfs_rq_runtime(cfs_rq, delta_exec);
720 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
722 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
726 * Task is being enqueued - update stats:
728 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
731 * Are we enqueueing a waiting task? (for current tasks
732 * a dequeue/enqueue event is a NOP)
734 if (se != cfs_rq->curr)
735 update_stats_wait_start(cfs_rq, se);
739 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
741 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
742 rq_of(cfs_rq)->clock - se->statistics.wait_start));
743 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
744 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
745 rq_of(cfs_rq)->clock - se->statistics.wait_start);
746 #ifdef CONFIG_SCHEDSTATS
747 if (entity_is_task(se)) {
748 trace_sched_stat_wait(task_of(se),
749 rq_of(cfs_rq)->clock - se->statistics.wait_start);
752 schedstat_set(se->statistics.wait_start, 0);
756 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 * Mark the end of the wait period if dequeueing a
762 if (se != cfs_rq->curr)
763 update_stats_wait_end(cfs_rq, se);
767 * We are picking a new current task - update its stats:
770 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 * We are starting a new run period:
775 se->exec_start = rq_of(cfs_rq)->clock_task;
778 /**************************************************
779 * Scheduling class numa methods.
781 * The purpose of the NUMA bits are to maintain compute (task) and data
782 * (memory) locality. We try and achieve this by making tasks stick to
783 * a particular node (their home node) but if fairness mandates they run
784 * elsewhere for long enough, we let the memory follow them.
786 * Tasks start out with their home-node unset (-1) this effectively means
787 * they act !NUMA until we've established the task is busy enough to bother
790 * Once we start doing NUMA placement there's two modes, 'small' process-wide
791 * and 'big' per-task. For the small mode we have a process-wide home node
792 * and lazily mirgrate all memory only when this home-node changes.
794 * For big mode we keep a home-node per task and use periodic fault scans
795 * to try and estalish a task<->page relation. This assumes the task<->page
796 * relation is a compute<->data relation, this is false for things like virt.
797 * and n:m threading solutions but its the best we can do given the
798 * information we have.
801 static unsigned long task_h_load(struct task_struct *p);
803 #ifdef CONFIG_SCHED_NUMA
804 static void account_offnode_enqueue(struct rq *rq, struct task_struct *p)
806 p->numa_contrib = task_h_load(p);
807 rq->offnode_weight += p->numa_contrib;
808 rq->offnode_running++;
811 static void account_offnode_dequeue(struct rq *rq, struct task_struct *p)
813 rq->offnode_weight -= p->numa_contrib;
814 rq->offnode_running--;
818 * numa task sample period in ms: 2.5s
820 unsigned int sysctl_sched_numa_task_period = 2500;
823 * Determine if a process is 'big'.
825 * Currently only looks at CPU-time used, maybe we should also add an RSS
828 static bool task_numa_big(struct task_struct *p)
830 struct sched_domain *sd;
831 struct task_struct *t;
832 u64 walltime = local_clock();
836 if (sched_feat(NUMA_FORCE_BIG))
842 if (t->sched_class == &fair_sched_class)
843 runtime += t->se.sum_exec_runtime;
844 } while ((t = next_thread(t)) != p);
846 sd = rcu_dereference(__raw_get_cpu_var(sd_node));
848 weight = sd->span_weight;
851 runtime -= p->numa_runtime_stamp;
852 walltime -= p->numa_walltime_stamp;
854 p->numa_runtime_stamp += runtime;
855 p->numa_walltime_stamp += walltime;
858 * We're 'big' when we burn more than half a node's worth
861 return runtime > walltime * max(1, weight / 2);
864 static bool had_many_migrate_failures(struct task_struct *p)
866 /* More than 1/4 of the attempted NUMA page migrations failed. */
867 return p->mm->numa_migrate_failed * 3 > p->mm->numa_migrate_success;
870 static inline bool need_numa_migration(struct task_struct *p)
873 * We need to change our home-node, its been different for 2 samples.
874 * See the whole P(n)^2 story in task_tick_numa().
876 return p->node_curr == p->node_last && p->node != p->node_curr;
879 static void sched_setnode_process(struct task_struct *p, int node)
881 struct task_struct *t = p;
885 sched_setnode(t, node);
886 } while ((t = next_thread(t)) != p);
891 * The expensive part of numa migration is done from task_work context.
892 * Triggered from task_tick_numa().
894 void task_numa_work(struct callback_head *work)
896 unsigned long migrate, next_scan, now = jiffies;
897 struct task_struct *p = current;
901 WARN_ON_ONCE(p != container_of(work, struct task_struct, rcu));
904 * Who cares about NUMA placement when they're dying.
906 * NOTE: make sure not to dereference p->mm before this check,
907 * exit_task_work() happens _after_ exit_mm() so we could be called
908 * without p->mm even though we still had it when we enqueued this
911 if (p->flags & PF_EXITING)
914 big = p->mm->numa_big;
915 need_migration = need_numa_migration(p);
918 * Change per-task state before the process wide freq. throttle,
919 * otherwise it might be a long while ere this task wins the
920 * lottery and gets its home-node set.
922 if (big && need_migration)
923 sched_setnode(p, p->node_curr);
926 * Enforce maximal scan/migration frequency..
928 migrate = p->mm->numa_next_scan;
929 if (time_before(now, migrate))
932 next_scan = now + 2*msecs_to_jiffies(sysctl_sched_numa_task_period);
933 if (cmpxchg(&p->mm->numa_next_scan, migrate, next_scan) != migrate)
937 /* Age the numa migrate statistics. */
938 p->mm->numa_migrate_failed /= 2;
939 p->mm->numa_migrate_success /= 2;
941 big = p->mm->numa_big = task_numa_big(p);
944 if (need_migration) {
946 sched_setnode(p, p->node_curr);
948 sched_setnode_process(p, p->node_curr);
951 if (big || need_migration || had_many_migrate_failures(p))
952 lazy_migrate_process(p->mm);
956 * Sample task location from hardirq context (tick), this has minimal bias with
957 * obvious exceptions of frequency interference and tick avoidance techniques.
958 * If this were to become a problem we could move this sampling into the
959 * sleep/wakeup path -- but we'd prefer to avoid that for obvious reasons.
961 void task_tick_numa(struct rq *rq, struct task_struct *curr)
966 * We don't care about NUMA placement if we don't have memory.
972 * Sample our node location every @sysctl_sched_numa_task_period
973 * runtime ms. We use a two stage selection in order to filter
974 * unlikely locations.
976 * If P(n) is the probability we're on node 'n', then the probability
977 * we sample the same node twice is P(n)^2. This quadric squishes small
978 * values and makes it more likely we end up on nodes where we have
979 * significant presence.
981 * Using runtime rather than walltime has the dual advantage that
982 * we (mostly) drive the selection from busy threads and that the
983 * task needs to have done some actual work before we bother with
986 now = curr->se.sum_exec_runtime;
987 period = (u64)sysctl_sched_numa_task_period * NSEC_PER_MSEC;
989 if (now - curr->node_stamp > period) {
990 curr->node_stamp = now;
992 curr->node_last = curr->node_curr;
993 curr->node_curr = numa_node_id();
996 * We need to do expensive work to either migrate or
997 * drive priodic state update or scanning for 'big' processes.
999 if (need_numa_migration(curr) ||
1000 !time_before(jiffies, curr->mm->numa_next_scan)) {
1002 * We can re-use curr->rcu because we checked curr->mm
1003 * != NULL so release_task()->call_rcu() was not called
1004 * yet and exit_task_work() is called before
1007 init_task_work(&curr->rcu, task_numa_work);
1008 task_work_add(curr, &curr->rcu, true);
1013 static void account_offnode_enqueue(struct rq *rq, struct task_struct *p)
1017 static void account_offnode_dequeue(struct rq *rq, struct task_struct *p)
1021 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1024 #endif /* CONFIG_SCHED_NUMA */
1026 /**************************************************
1027 * Scheduling class queueing methods:
1031 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1033 update_load_add(&cfs_rq->load, se->load.weight);
1034 if (!parent_entity(se))
1035 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1037 if (entity_is_task(se)) {
1038 struct rq *rq = rq_of(cfs_rq);
1039 struct task_struct *p = task_of(se);
1040 struct list_head *tasks = &rq->cfs_tasks;
1042 if (offnode_task(p)) {
1043 account_offnode_enqueue(rq, p);
1044 tasks = offnode_tasks(rq);
1047 list_add(&se->group_node, tasks);
1049 #endif /* CONFIG_SMP */
1050 cfs_rq->nr_running++;
1054 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1056 update_load_sub(&cfs_rq->load, se->load.weight);
1057 if (!parent_entity(se))
1058 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1059 if (entity_is_task(se)) {
1060 struct task_struct *p = task_of(se);
1062 list_del_init(&se->group_node);
1064 if (offnode_task(p))
1065 account_offnode_dequeue(rq_of(cfs_rq), p);
1067 cfs_rq->nr_running--;
1070 #ifdef CONFIG_FAIR_GROUP_SCHED
1071 /* we need this in update_cfs_load and load-balance functions below */
1072 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1074 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
1077 struct task_group *tg = cfs_rq->tg;
1080 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
1081 load_avg -= cfs_rq->load_contribution;
1083 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
1084 atomic_add(load_avg, &tg->load_weight);
1085 cfs_rq->load_contribution += load_avg;
1089 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1091 u64 period = sysctl_sched_shares_window;
1093 unsigned long load = cfs_rq->load.weight;
1095 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
1098 now = rq_of(cfs_rq)->clock_task;
1099 delta = now - cfs_rq->load_stamp;
1101 /* truncate load history at 4 idle periods */
1102 if (cfs_rq->load_stamp > cfs_rq->load_last &&
1103 now - cfs_rq->load_last > 4 * period) {
1104 cfs_rq->load_period = 0;
1105 cfs_rq->load_avg = 0;
1109 cfs_rq->load_stamp = now;
1110 cfs_rq->load_unacc_exec_time = 0;
1111 cfs_rq->load_period += delta;
1113 cfs_rq->load_last = now;
1114 cfs_rq->load_avg += delta * load;
1117 /* consider updating load contribution on each fold or truncate */
1118 if (global_update || cfs_rq->load_period > period
1119 || !cfs_rq->load_period)
1120 update_cfs_rq_load_contribution(cfs_rq, global_update);
1122 while (cfs_rq->load_period > period) {
1124 * Inline assembly required to prevent the compiler
1125 * optimising this loop into a divmod call.
1126 * See __iter_div_u64_rem() for another example of this.
1128 asm("" : "+rm" (cfs_rq->load_period));
1129 cfs_rq->load_period /= 2;
1130 cfs_rq->load_avg /= 2;
1133 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
1134 list_del_leaf_cfs_rq(cfs_rq);
1137 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1142 * Use this CPU's actual weight instead of the last load_contribution
1143 * to gain a more accurate current total weight. See
1144 * update_cfs_rq_load_contribution().
1146 tg_weight = atomic_read(&tg->load_weight);
1147 tg_weight -= cfs_rq->load_contribution;
1148 tg_weight += cfs_rq->load.weight;
1153 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1155 long tg_weight, load, shares;
1157 tg_weight = calc_tg_weight(tg, cfs_rq);
1158 load = cfs_rq->load.weight;
1160 shares = (tg->shares * load);
1162 shares /= tg_weight;
1164 if (shares < MIN_SHARES)
1165 shares = MIN_SHARES;
1166 if (shares > tg->shares)
1167 shares = tg->shares;
1172 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1174 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1175 update_cfs_load(cfs_rq, 0);
1176 update_cfs_shares(cfs_rq);
1179 # else /* CONFIG_SMP */
1180 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1184 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1189 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1192 # endif /* CONFIG_SMP */
1193 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1194 unsigned long weight)
1197 /* commit outstanding execution time */
1198 if (cfs_rq->curr == se)
1199 update_curr(cfs_rq);
1200 account_entity_dequeue(cfs_rq, se);
1203 update_load_set(&se->load, weight);
1206 account_entity_enqueue(cfs_rq, se);
1209 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1211 struct task_group *tg;
1212 struct sched_entity *se;
1216 se = tg->se[cpu_of(rq_of(cfs_rq))];
1217 if (!se || throttled_hierarchy(cfs_rq))
1220 if (likely(se->load.weight == tg->shares))
1223 shares = calc_cfs_shares(cfs_rq, tg);
1225 reweight_entity(cfs_rq_of(se), se, shares);
1227 #else /* CONFIG_FAIR_GROUP_SCHED */
1228 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1232 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1236 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1239 #endif /* CONFIG_FAIR_GROUP_SCHED */
1241 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1243 #ifdef CONFIG_SCHEDSTATS
1244 struct task_struct *tsk = NULL;
1246 if (entity_is_task(se))
1249 if (se->statistics.sleep_start) {
1250 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1255 if (unlikely(delta > se->statistics.sleep_max))
1256 se->statistics.sleep_max = delta;
1258 se->statistics.sleep_start = 0;
1259 se->statistics.sum_sleep_runtime += delta;
1262 account_scheduler_latency(tsk, delta >> 10, 1);
1263 trace_sched_stat_sleep(tsk, delta);
1266 if (se->statistics.block_start) {
1267 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1272 if (unlikely(delta > se->statistics.block_max))
1273 se->statistics.block_max = delta;
1275 se->statistics.block_start = 0;
1276 se->statistics.sum_sleep_runtime += delta;
1279 if (tsk->in_iowait) {
1280 se->statistics.iowait_sum += delta;
1281 se->statistics.iowait_count++;
1282 trace_sched_stat_iowait(tsk, delta);
1285 trace_sched_stat_blocked(tsk, delta);
1288 * Blocking time is in units of nanosecs, so shift by
1289 * 20 to get a milliseconds-range estimation of the
1290 * amount of time that the task spent sleeping:
1292 if (unlikely(prof_on == SLEEP_PROFILING)) {
1293 profile_hits(SLEEP_PROFILING,
1294 (void *)get_wchan(tsk),
1297 account_scheduler_latency(tsk, delta >> 10, 0);
1303 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1305 #ifdef CONFIG_SCHED_DEBUG
1306 s64 d = se->vruntime - cfs_rq->min_vruntime;
1311 if (d > 3*sysctl_sched_latency)
1312 schedstat_inc(cfs_rq, nr_spread_over);
1317 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1319 u64 vruntime = cfs_rq->min_vruntime;
1322 * The 'current' period is already promised to the current tasks,
1323 * however the extra weight of the new task will slow them down a
1324 * little, place the new task so that it fits in the slot that
1325 * stays open at the end.
1327 if (initial && sched_feat(START_DEBIT))
1328 vruntime += sched_vslice(cfs_rq, se);
1330 /* sleeps up to a single latency don't count. */
1332 unsigned long thresh = sysctl_sched_latency;
1335 * Halve their sleep time's effect, to allow
1336 * for a gentler effect of sleepers:
1338 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1344 /* ensure we never gain time by being placed backwards. */
1345 vruntime = max_vruntime(se->vruntime, vruntime);
1347 se->vruntime = vruntime;
1350 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1353 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1356 * Update the normalized vruntime before updating min_vruntime
1357 * through callig update_curr().
1359 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1360 se->vruntime += cfs_rq->min_vruntime;
1363 * Update run-time statistics of the 'current'.
1365 update_curr(cfs_rq);
1366 update_cfs_load(cfs_rq, 0);
1367 account_entity_enqueue(cfs_rq, se);
1368 update_cfs_shares(cfs_rq);
1370 if (flags & ENQUEUE_WAKEUP) {
1371 place_entity(cfs_rq, se, 0);
1372 enqueue_sleeper(cfs_rq, se);
1375 update_stats_enqueue(cfs_rq, se);
1376 check_spread(cfs_rq, se);
1377 if (se != cfs_rq->curr)
1378 __enqueue_entity(cfs_rq, se);
1381 if (cfs_rq->nr_running == 1) {
1382 list_add_leaf_cfs_rq(cfs_rq);
1383 check_enqueue_throttle(cfs_rq);
1387 static void __clear_buddies_last(struct sched_entity *se)
1389 for_each_sched_entity(se) {
1390 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1391 if (cfs_rq->last == se)
1392 cfs_rq->last = NULL;
1398 static void __clear_buddies_next(struct sched_entity *se)
1400 for_each_sched_entity(se) {
1401 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1402 if (cfs_rq->next == se)
1403 cfs_rq->next = NULL;
1409 static void __clear_buddies_skip(struct sched_entity *se)
1411 for_each_sched_entity(se) {
1412 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1413 if (cfs_rq->skip == se)
1414 cfs_rq->skip = NULL;
1420 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1422 if (cfs_rq->last == se)
1423 __clear_buddies_last(se);
1425 if (cfs_rq->next == se)
1426 __clear_buddies_next(se);
1428 if (cfs_rq->skip == se)
1429 __clear_buddies_skip(se);
1432 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1435 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1438 * Update run-time statistics of the 'current'.
1440 update_curr(cfs_rq);
1442 update_stats_dequeue(cfs_rq, se);
1443 if (flags & DEQUEUE_SLEEP) {
1444 #ifdef CONFIG_SCHEDSTATS
1445 if (entity_is_task(se)) {
1446 struct task_struct *tsk = task_of(se);
1448 if (tsk->state & TASK_INTERRUPTIBLE)
1449 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1450 if (tsk->state & TASK_UNINTERRUPTIBLE)
1451 se->statistics.block_start = rq_of(cfs_rq)->clock;
1456 clear_buddies(cfs_rq, se);
1458 if (se != cfs_rq->curr)
1459 __dequeue_entity(cfs_rq, se);
1461 update_cfs_load(cfs_rq, 0);
1462 account_entity_dequeue(cfs_rq, se);
1465 * Normalize the entity after updating the min_vruntime because the
1466 * update can refer to the ->curr item and we need to reflect this
1467 * movement in our normalized position.
1469 if (!(flags & DEQUEUE_SLEEP))
1470 se->vruntime -= cfs_rq->min_vruntime;
1472 /* return excess runtime on last dequeue */
1473 return_cfs_rq_runtime(cfs_rq);
1475 update_min_vruntime(cfs_rq);
1476 update_cfs_shares(cfs_rq);
1480 * Preempt the current task with a newly woken task if needed:
1483 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1485 unsigned long ideal_runtime, delta_exec;
1486 struct sched_entity *se;
1489 ideal_runtime = sched_slice(cfs_rq, curr);
1490 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1491 if (delta_exec > ideal_runtime) {
1492 resched_task(rq_of(cfs_rq)->curr);
1494 * The current task ran long enough, ensure it doesn't get
1495 * re-elected due to buddy favours.
1497 clear_buddies(cfs_rq, curr);
1502 * Ensure that a task that missed wakeup preemption by a
1503 * narrow margin doesn't have to wait for a full slice.
1504 * This also mitigates buddy induced latencies under load.
1506 if (delta_exec < sysctl_sched_min_granularity)
1509 se = __pick_first_entity(cfs_rq);
1510 delta = curr->vruntime - se->vruntime;
1515 if (delta > ideal_runtime)
1516 resched_task(rq_of(cfs_rq)->curr);
1520 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1522 /* 'current' is not kept within the tree. */
1525 * Any task has to be enqueued before it get to execute on
1526 * a CPU. So account for the time it spent waiting on the
1529 update_stats_wait_end(cfs_rq, se);
1530 __dequeue_entity(cfs_rq, se);
1533 update_stats_curr_start(cfs_rq, se);
1535 #ifdef CONFIG_SCHEDSTATS
1537 * Track our maximum slice length, if the CPU's load is at
1538 * least twice that of our own weight (i.e. dont track it
1539 * when there are only lesser-weight tasks around):
1541 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1542 se->statistics.slice_max = max(se->statistics.slice_max,
1543 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1546 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1550 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1553 * Pick the next process, keeping these things in mind, in this order:
1554 * 1) keep things fair between processes/task groups
1555 * 2) pick the "next" process, since someone really wants that to run
1556 * 3) pick the "last" process, for cache locality
1557 * 4) do not run the "skip" process, if something else is available
1559 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1561 struct sched_entity *se = __pick_first_entity(cfs_rq);
1562 struct sched_entity *left = se;
1565 * Avoid running the skip buddy, if running something else can
1566 * be done without getting too unfair.
1568 if (cfs_rq->skip == se) {
1569 struct sched_entity *second = __pick_next_entity(se);
1570 if (second && wakeup_preempt_entity(second, left) < 1)
1575 * Prefer last buddy, try to return the CPU to a preempted task.
1577 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1581 * Someone really wants this to run. If it's not unfair, run it.
1583 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1586 clear_buddies(cfs_rq, se);
1591 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1593 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1596 * If still on the runqueue then deactivate_task()
1597 * was not called and update_curr() has to be done:
1600 update_curr(cfs_rq);
1602 /* throttle cfs_rqs exceeding runtime */
1603 check_cfs_rq_runtime(cfs_rq);
1605 check_spread(cfs_rq, prev);
1607 update_stats_wait_start(cfs_rq, prev);
1608 /* Put 'current' back into the tree. */
1609 __enqueue_entity(cfs_rq, prev);
1611 cfs_rq->curr = NULL;
1615 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1618 * Update run-time statistics of the 'current'.
1620 update_curr(cfs_rq);
1623 * Update share accounting for long-running entities.
1625 update_entity_shares_tick(cfs_rq);
1627 #ifdef CONFIG_SCHED_HRTICK
1629 * queued ticks are scheduled to match the slice, so don't bother
1630 * validating it and just reschedule.
1633 resched_task(rq_of(cfs_rq)->curr);
1637 * don't let the period tick interfere with the hrtick preemption
1639 if (!sched_feat(DOUBLE_TICK) &&
1640 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1644 if (cfs_rq->nr_running > 1)
1645 check_preempt_tick(cfs_rq, curr);
1649 /**************************************************
1650 * CFS bandwidth control machinery
1653 #ifdef CONFIG_CFS_BANDWIDTH
1655 #ifdef HAVE_JUMP_LABEL
1656 static struct static_key __cfs_bandwidth_used;
1658 static inline bool cfs_bandwidth_used(void)
1660 return static_key_false(&__cfs_bandwidth_used);
1663 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1665 /* only need to count groups transitioning between enabled/!enabled */
1666 if (enabled && !was_enabled)
1667 static_key_slow_inc(&__cfs_bandwidth_used);
1668 else if (!enabled && was_enabled)
1669 static_key_slow_dec(&__cfs_bandwidth_used);
1671 #else /* HAVE_JUMP_LABEL */
1672 static bool cfs_bandwidth_used(void)
1677 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1678 #endif /* HAVE_JUMP_LABEL */
1681 * default period for cfs group bandwidth.
1682 * default: 0.1s, units: nanoseconds
1684 static inline u64 default_cfs_period(void)
1686 return 100000000ULL;
1689 static inline u64 sched_cfs_bandwidth_slice(void)
1691 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1695 * Replenish runtime according to assigned quota and update expiration time.
1696 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1697 * additional synchronization around rq->lock.
1699 * requires cfs_b->lock
1701 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1705 if (cfs_b->quota == RUNTIME_INF)
1708 now = sched_clock_cpu(smp_processor_id());
1709 cfs_b->runtime = cfs_b->quota;
1710 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1713 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1715 return &tg->cfs_bandwidth;
1718 /* returns 0 on failure to allocate runtime */
1719 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1721 struct task_group *tg = cfs_rq->tg;
1722 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1723 u64 amount = 0, min_amount, expires;
1725 /* note: this is a positive sum as runtime_remaining <= 0 */
1726 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1728 raw_spin_lock(&cfs_b->lock);
1729 if (cfs_b->quota == RUNTIME_INF)
1730 amount = min_amount;
1733 * If the bandwidth pool has become inactive, then at least one
1734 * period must have elapsed since the last consumption.
1735 * Refresh the global state and ensure bandwidth timer becomes
1738 if (!cfs_b->timer_active) {
1739 __refill_cfs_bandwidth_runtime(cfs_b);
1740 __start_cfs_bandwidth(cfs_b);
1743 if (cfs_b->runtime > 0) {
1744 amount = min(cfs_b->runtime, min_amount);
1745 cfs_b->runtime -= amount;
1749 expires = cfs_b->runtime_expires;
1750 raw_spin_unlock(&cfs_b->lock);
1752 cfs_rq->runtime_remaining += amount;
1754 * we may have advanced our local expiration to account for allowed
1755 * spread between our sched_clock and the one on which runtime was
1758 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1759 cfs_rq->runtime_expires = expires;
1761 return cfs_rq->runtime_remaining > 0;
1765 * Note: This depends on the synchronization provided by sched_clock and the
1766 * fact that rq->clock snapshots this value.
1768 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1770 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1771 struct rq *rq = rq_of(cfs_rq);
1773 /* if the deadline is ahead of our clock, nothing to do */
1774 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1777 if (cfs_rq->runtime_remaining < 0)
1781 * If the local deadline has passed we have to consider the
1782 * possibility that our sched_clock is 'fast' and the global deadline
1783 * has not truly expired.
1785 * Fortunately we can check determine whether this the case by checking
1786 * whether the global deadline has advanced.
1789 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1790 /* extend local deadline, drift is bounded above by 2 ticks */
1791 cfs_rq->runtime_expires += TICK_NSEC;
1793 /* global deadline is ahead, expiration has passed */
1794 cfs_rq->runtime_remaining = 0;
1798 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1799 unsigned long delta_exec)
1801 /* dock delta_exec before expiring quota (as it could span periods) */
1802 cfs_rq->runtime_remaining -= delta_exec;
1803 expire_cfs_rq_runtime(cfs_rq);
1805 if (likely(cfs_rq->runtime_remaining > 0))
1809 * if we're unable to extend our runtime we resched so that the active
1810 * hierarchy can be throttled
1812 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1813 resched_task(rq_of(cfs_rq)->curr);
1816 static __always_inline
1817 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1819 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1822 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1825 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1827 return cfs_bandwidth_used() && cfs_rq->throttled;
1830 /* check whether cfs_rq, or any parent, is throttled */
1831 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1833 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1837 * Ensure that neither of the group entities corresponding to src_cpu or
1838 * dest_cpu are members of a throttled hierarchy when performing group
1839 * load-balance operations.
1841 static inline int throttled_lb_pair(struct task_group *tg,
1842 int src_cpu, int dest_cpu)
1844 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1846 src_cfs_rq = tg->cfs_rq[src_cpu];
1847 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1849 return throttled_hierarchy(src_cfs_rq) ||
1850 throttled_hierarchy(dest_cfs_rq);
1853 /* updated child weight may affect parent so we have to do this bottom up */
1854 static int tg_unthrottle_up(struct task_group *tg, void *data)
1856 struct rq *rq = data;
1857 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1859 cfs_rq->throttle_count--;
1861 if (!cfs_rq->throttle_count) {
1862 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1864 /* leaving throttled state, advance shares averaging windows */
1865 cfs_rq->load_stamp += delta;
1866 cfs_rq->load_last += delta;
1868 /* update entity weight now that we are on_rq again */
1869 update_cfs_shares(cfs_rq);
1876 static int tg_throttle_down(struct task_group *tg, void *data)
1878 struct rq *rq = data;
1879 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1881 /* group is entering throttled state, record last load */
1882 if (!cfs_rq->throttle_count)
1883 update_cfs_load(cfs_rq, 0);
1884 cfs_rq->throttle_count++;
1889 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1891 struct rq *rq = rq_of(cfs_rq);
1892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1893 struct sched_entity *se;
1894 long task_delta, dequeue = 1;
1896 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1898 /* account load preceding throttle */
1900 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1903 task_delta = cfs_rq->h_nr_running;
1904 for_each_sched_entity(se) {
1905 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1906 /* throttled entity or throttle-on-deactivate */
1911 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1912 qcfs_rq->h_nr_running -= task_delta;
1914 if (qcfs_rq->load.weight)
1919 rq->nr_running -= task_delta;
1921 cfs_rq->throttled = 1;
1922 cfs_rq->throttled_timestamp = rq->clock;
1923 raw_spin_lock(&cfs_b->lock);
1924 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1925 raw_spin_unlock(&cfs_b->lock);
1928 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1930 struct rq *rq = rq_of(cfs_rq);
1931 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1932 struct sched_entity *se;
1936 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1938 cfs_rq->throttled = 0;
1939 raw_spin_lock(&cfs_b->lock);
1940 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1941 list_del_rcu(&cfs_rq->throttled_list);
1942 raw_spin_unlock(&cfs_b->lock);
1943 cfs_rq->throttled_timestamp = 0;
1945 update_rq_clock(rq);
1946 /* update hierarchical throttle state */
1947 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1949 if (!cfs_rq->load.weight)
1952 task_delta = cfs_rq->h_nr_running;
1953 for_each_sched_entity(se) {
1957 cfs_rq = cfs_rq_of(se);
1959 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1960 cfs_rq->h_nr_running += task_delta;
1962 if (cfs_rq_throttled(cfs_rq))
1967 rq->nr_running += task_delta;
1969 /* determine whether we need to wake up potentially idle cpu */
1970 if (rq->curr == rq->idle && rq->cfs.nr_running)
1971 resched_task(rq->curr);
1974 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1975 u64 remaining, u64 expires)
1977 struct cfs_rq *cfs_rq;
1978 u64 runtime = remaining;
1981 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1983 struct rq *rq = rq_of(cfs_rq);
1985 raw_spin_lock(&rq->lock);
1986 if (!cfs_rq_throttled(cfs_rq))
1989 runtime = -cfs_rq->runtime_remaining + 1;
1990 if (runtime > remaining)
1991 runtime = remaining;
1992 remaining -= runtime;
1994 cfs_rq->runtime_remaining += runtime;
1995 cfs_rq->runtime_expires = expires;
1997 /* we check whether we're throttled above */
1998 if (cfs_rq->runtime_remaining > 0)
1999 unthrottle_cfs_rq(cfs_rq);
2002 raw_spin_unlock(&rq->lock);
2013 * Responsible for refilling a task_group's bandwidth and unthrottling its
2014 * cfs_rqs as appropriate. If there has been no activity within the last
2015 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2016 * used to track this state.
2018 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2020 u64 runtime, runtime_expires;
2021 int idle = 1, throttled;
2023 raw_spin_lock(&cfs_b->lock);
2024 /* no need to continue the timer with no bandwidth constraint */
2025 if (cfs_b->quota == RUNTIME_INF)
2028 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2029 /* idle depends on !throttled (for the case of a large deficit) */
2030 idle = cfs_b->idle && !throttled;
2031 cfs_b->nr_periods += overrun;
2033 /* if we're going inactive then everything else can be deferred */
2037 __refill_cfs_bandwidth_runtime(cfs_b);
2040 /* mark as potentially idle for the upcoming period */
2045 /* account preceding periods in which throttling occurred */
2046 cfs_b->nr_throttled += overrun;
2049 * There are throttled entities so we must first use the new bandwidth
2050 * to unthrottle them before making it generally available. This
2051 * ensures that all existing debts will be paid before a new cfs_rq is
2054 runtime = cfs_b->runtime;
2055 runtime_expires = cfs_b->runtime_expires;
2059 * This check is repeated as we are holding onto the new bandwidth
2060 * while we unthrottle. This can potentially race with an unthrottled
2061 * group trying to acquire new bandwidth from the global pool.
2063 while (throttled && runtime > 0) {
2064 raw_spin_unlock(&cfs_b->lock);
2065 /* we can't nest cfs_b->lock while distributing bandwidth */
2066 runtime = distribute_cfs_runtime(cfs_b, runtime,
2068 raw_spin_lock(&cfs_b->lock);
2070 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2073 /* return (any) remaining runtime */
2074 cfs_b->runtime = runtime;
2076 * While we are ensured activity in the period following an
2077 * unthrottle, this also covers the case in which the new bandwidth is
2078 * insufficient to cover the existing bandwidth deficit. (Forcing the
2079 * timer to remain active while there are any throttled entities.)
2084 cfs_b->timer_active = 0;
2085 raw_spin_unlock(&cfs_b->lock);
2090 /* a cfs_rq won't donate quota below this amount */
2091 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2092 /* minimum remaining period time to redistribute slack quota */
2093 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2094 /* how long we wait to gather additional slack before distributing */
2095 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2097 /* are we near the end of the current quota period? */
2098 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2100 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2103 /* if the call-back is running a quota refresh is already occurring */
2104 if (hrtimer_callback_running(refresh_timer))
2107 /* is a quota refresh about to occur? */
2108 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2109 if (remaining < min_expire)
2115 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2117 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2119 /* if there's a quota refresh soon don't bother with slack */
2120 if (runtime_refresh_within(cfs_b, min_left))
2123 start_bandwidth_timer(&cfs_b->slack_timer,
2124 ns_to_ktime(cfs_bandwidth_slack_period));
2127 /* we know any runtime found here is valid as update_curr() precedes return */
2128 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2130 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2131 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2133 if (slack_runtime <= 0)
2136 raw_spin_lock(&cfs_b->lock);
2137 if (cfs_b->quota != RUNTIME_INF &&
2138 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2139 cfs_b->runtime += slack_runtime;
2141 /* we are under rq->lock, defer unthrottling using a timer */
2142 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2143 !list_empty(&cfs_b->throttled_cfs_rq))
2144 start_cfs_slack_bandwidth(cfs_b);
2146 raw_spin_unlock(&cfs_b->lock);
2148 /* even if it's not valid for return we don't want to try again */
2149 cfs_rq->runtime_remaining -= slack_runtime;
2152 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2154 if (!cfs_bandwidth_used())
2157 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2160 __return_cfs_rq_runtime(cfs_rq);
2164 * This is done with a timer (instead of inline with bandwidth return) since
2165 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2167 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2169 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2172 /* confirm we're still not at a refresh boundary */
2173 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2176 raw_spin_lock(&cfs_b->lock);
2177 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2178 runtime = cfs_b->runtime;
2181 expires = cfs_b->runtime_expires;
2182 raw_spin_unlock(&cfs_b->lock);
2187 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2189 raw_spin_lock(&cfs_b->lock);
2190 if (expires == cfs_b->runtime_expires)
2191 cfs_b->runtime = runtime;
2192 raw_spin_unlock(&cfs_b->lock);
2196 * When a group wakes up we want to make sure that its quota is not already
2197 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2198 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2200 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2202 if (!cfs_bandwidth_used())
2205 /* an active group must be handled by the update_curr()->put() path */
2206 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2209 /* ensure the group is not already throttled */
2210 if (cfs_rq_throttled(cfs_rq))
2213 /* update runtime allocation */
2214 account_cfs_rq_runtime(cfs_rq, 0);
2215 if (cfs_rq->runtime_remaining <= 0)
2216 throttle_cfs_rq(cfs_rq);
2219 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2220 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2222 if (!cfs_bandwidth_used())
2225 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2229 * it's possible for a throttled entity to be forced into a running
2230 * state (e.g. set_curr_task), in this case we're finished.
2232 if (cfs_rq_throttled(cfs_rq))
2235 throttle_cfs_rq(cfs_rq);
2238 static inline u64 default_cfs_period(void);
2239 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2240 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2242 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2244 struct cfs_bandwidth *cfs_b =
2245 container_of(timer, struct cfs_bandwidth, slack_timer);
2246 do_sched_cfs_slack_timer(cfs_b);
2248 return HRTIMER_NORESTART;
2251 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2253 struct cfs_bandwidth *cfs_b =
2254 container_of(timer, struct cfs_bandwidth, period_timer);
2260 now = hrtimer_cb_get_time(timer);
2261 overrun = hrtimer_forward(timer, now, cfs_b->period);
2266 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2269 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2272 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2274 raw_spin_lock_init(&cfs_b->lock);
2276 cfs_b->quota = RUNTIME_INF;
2277 cfs_b->period = ns_to_ktime(default_cfs_period());
2279 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2280 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2281 cfs_b->period_timer.function = sched_cfs_period_timer;
2282 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2283 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2286 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2288 cfs_rq->runtime_enabled = 0;
2289 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2292 /* requires cfs_b->lock, may release to reprogram timer */
2293 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2296 * The timer may be active because we're trying to set a new bandwidth
2297 * period or because we're racing with the tear-down path
2298 * (timer_active==0 becomes visible before the hrtimer call-back
2299 * terminates). In either case we ensure that it's re-programmed
2301 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2302 raw_spin_unlock(&cfs_b->lock);
2303 /* ensure cfs_b->lock is available while we wait */
2304 hrtimer_cancel(&cfs_b->period_timer);
2306 raw_spin_lock(&cfs_b->lock);
2307 /* if someone else restarted the timer then we're done */
2308 if (cfs_b->timer_active)
2312 cfs_b->timer_active = 1;
2313 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2316 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2318 hrtimer_cancel(&cfs_b->period_timer);
2319 hrtimer_cancel(&cfs_b->slack_timer);
2322 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2324 struct cfs_rq *cfs_rq;
2326 for_each_leaf_cfs_rq(rq, cfs_rq) {
2327 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2329 if (!cfs_rq->runtime_enabled)
2333 * clock_task is not advancing so we just need to make sure
2334 * there's some valid quota amount
2336 cfs_rq->runtime_remaining = cfs_b->quota;
2337 if (cfs_rq_throttled(cfs_rq))
2338 unthrottle_cfs_rq(cfs_rq);
2342 #else /* CONFIG_CFS_BANDWIDTH */
2343 static __always_inline
2344 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2345 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2346 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2347 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2349 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2354 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2359 static inline int throttled_lb_pair(struct task_group *tg,
2360 int src_cpu, int dest_cpu)
2365 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2367 #ifdef CONFIG_FAIR_GROUP_SCHED
2368 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2371 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2375 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2376 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2378 #endif /* CONFIG_CFS_BANDWIDTH */
2380 /**************************************************
2381 * CFS operations on tasks:
2384 #ifdef CONFIG_SCHED_HRTICK
2385 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2387 struct sched_entity *se = &p->se;
2388 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2390 WARN_ON(task_rq(p) != rq);
2392 if (cfs_rq->nr_running > 1) {
2393 u64 slice = sched_slice(cfs_rq, se);
2394 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2395 s64 delta = slice - ran;
2404 * Don't schedule slices shorter than 10000ns, that just
2405 * doesn't make sense. Rely on vruntime for fairness.
2408 delta = max_t(s64, 10000LL, delta);
2410 hrtick_start(rq, delta);
2415 * called from enqueue/dequeue and updates the hrtick when the
2416 * current task is from our class and nr_running is low enough
2419 static void hrtick_update(struct rq *rq)
2421 struct task_struct *curr = rq->curr;
2423 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2426 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2427 hrtick_start_fair(rq, curr);
2429 #else /* !CONFIG_SCHED_HRTICK */
2431 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2435 static inline void hrtick_update(struct rq *rq)
2441 * The enqueue_task method is called before nr_running is
2442 * increased. Here we update the fair scheduling stats and
2443 * then put the task into the rbtree:
2446 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2448 struct cfs_rq *cfs_rq;
2449 struct sched_entity *se = &p->se;
2451 for_each_sched_entity(se) {
2454 cfs_rq = cfs_rq_of(se);
2455 enqueue_entity(cfs_rq, se, flags);
2458 * end evaluation on encountering a throttled cfs_rq
2460 * note: in the case of encountering a throttled cfs_rq we will
2461 * post the final h_nr_running increment below.
2463 if (cfs_rq_throttled(cfs_rq))
2465 cfs_rq->h_nr_running++;
2467 flags = ENQUEUE_WAKEUP;
2470 for_each_sched_entity(se) {
2471 cfs_rq = cfs_rq_of(se);
2472 cfs_rq->h_nr_running++;
2474 if (cfs_rq_throttled(cfs_rq))
2477 update_cfs_load(cfs_rq, 0);
2478 update_cfs_shares(cfs_rq);
2486 static void set_next_buddy(struct sched_entity *se);
2489 * The dequeue_task method is called before nr_running is
2490 * decreased. We remove the task from the rbtree and
2491 * update the fair scheduling stats:
2493 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2495 struct cfs_rq *cfs_rq;
2496 struct sched_entity *se = &p->se;
2497 int task_sleep = flags & DEQUEUE_SLEEP;
2499 for_each_sched_entity(se) {
2500 cfs_rq = cfs_rq_of(se);
2501 dequeue_entity(cfs_rq, se, flags);
2504 * end evaluation on encountering a throttled cfs_rq
2506 * note: in the case of encountering a throttled cfs_rq we will
2507 * post the final h_nr_running decrement below.
2509 if (cfs_rq_throttled(cfs_rq))
2511 cfs_rq->h_nr_running--;
2513 /* Don't dequeue parent if it has other entities besides us */
2514 if (cfs_rq->load.weight) {
2516 * Bias pick_next to pick a task from this cfs_rq, as
2517 * p is sleeping when it is within its sched_slice.
2519 if (task_sleep && parent_entity(se))
2520 set_next_buddy(parent_entity(se));
2522 /* avoid re-evaluating load for this entity */
2523 se = parent_entity(se);
2526 flags |= DEQUEUE_SLEEP;
2529 for_each_sched_entity(se) {
2530 cfs_rq = cfs_rq_of(se);
2531 cfs_rq->h_nr_running--;
2533 if (cfs_rq_throttled(cfs_rq))
2536 update_cfs_load(cfs_rq, 0);
2537 update_cfs_shares(cfs_rq);
2546 /* Used instead of source_load when we know the type == 0 */
2547 static unsigned long weighted_cpuload(const int cpu)
2549 return cpu_rq(cpu)->load.weight;
2553 * Return a low guess at the load of a migration-source cpu weighted
2554 * according to the scheduling class and "nice" value.
2556 * We want to under-estimate the load of migration sources, to
2557 * balance conservatively.
2559 static unsigned long source_load(int cpu, int type)
2561 struct rq *rq = cpu_rq(cpu);
2562 unsigned long total = weighted_cpuload(cpu);
2564 if (type == 0 || !sched_feat(LB_BIAS))
2567 return min(rq->cpu_load[type-1], total);
2571 * Return a high guess at the load of a migration-target cpu weighted
2572 * according to the scheduling class and "nice" value.
2574 static unsigned long target_load(int cpu, int type)
2576 struct rq *rq = cpu_rq(cpu);
2577 unsigned long total = weighted_cpuload(cpu);
2579 if (type == 0 || !sched_feat(LB_BIAS))
2582 return max(rq->cpu_load[type-1], total);
2585 static unsigned long power_of(int cpu)
2587 return cpu_rq(cpu)->cpu_power;
2590 static unsigned long cpu_avg_load_per_task(int cpu)
2592 struct rq *rq = cpu_rq(cpu);
2593 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2596 return rq->load.weight / nr_running;
2602 static void task_waking_fair(struct task_struct *p)
2604 struct sched_entity *se = &p->se;
2605 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2608 #ifndef CONFIG_64BIT
2609 u64 min_vruntime_copy;
2612 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2614 min_vruntime = cfs_rq->min_vruntime;
2615 } while (min_vruntime != min_vruntime_copy);
2617 min_vruntime = cfs_rq->min_vruntime;
2620 se->vruntime -= min_vruntime;
2623 #ifdef CONFIG_FAIR_GROUP_SCHED
2625 * effective_load() calculates the load change as seen from the root_task_group
2627 * Adding load to a group doesn't make a group heavier, but can cause movement
2628 * of group shares between cpus. Assuming the shares were perfectly aligned one
2629 * can calculate the shift in shares.
2631 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2632 * on this @cpu and results in a total addition (subtraction) of @wg to the
2633 * total group weight.
2635 * Given a runqueue weight distribution (rw_i) we can compute a shares
2636 * distribution (s_i) using:
2638 * s_i = rw_i / \Sum rw_j (1)
2640 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2641 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2642 * shares distribution (s_i):
2644 * rw_i = { 2, 4, 1, 0 }
2645 * s_i = { 2/7, 4/7, 1/7, 0 }
2647 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2648 * task used to run on and the CPU the waker is running on), we need to
2649 * compute the effect of waking a task on either CPU and, in case of a sync
2650 * wakeup, compute the effect of the current task going to sleep.
2652 * So for a change of @wl to the local @cpu with an overall group weight change
2653 * of @wl we can compute the new shares distribution (s'_i) using:
2655 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2657 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2658 * differences in waking a task to CPU 0. The additional task changes the
2659 * weight and shares distributions like:
2661 * rw'_i = { 3, 4, 1, 0 }
2662 * s'_i = { 3/8, 4/8, 1/8, 0 }
2664 * We can then compute the difference in effective weight by using:
2666 * dw_i = S * (s'_i - s_i) (3)
2668 * Where 'S' is the group weight as seen by its parent.
2670 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2671 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2672 * 4/7) times the weight of the group.
2674 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2676 struct sched_entity *se = tg->se[cpu];
2678 if (!tg->parent) /* the trivial, non-cgroup case */
2681 for_each_sched_entity(se) {
2687 * W = @wg + \Sum rw_j
2689 W = wg + calc_tg_weight(tg, se->my_q);
2694 w = se->my_q->load.weight + wl;
2697 * wl = S * s'_i; see (2)
2700 wl = (w * tg->shares) / W;
2705 * Per the above, wl is the new se->load.weight value; since
2706 * those are clipped to [MIN_SHARES, ...) do so now. See
2707 * calc_cfs_shares().
2709 if (wl < MIN_SHARES)
2713 * wl = dw_i = S * (s'_i - s_i); see (3)
2715 wl -= se->load.weight;
2718 * Recursively apply this logic to all parent groups to compute
2719 * the final effective load change on the root group. Since
2720 * only the @tg group gets extra weight, all parent groups can
2721 * only redistribute existing shares. @wl is the shift in shares
2722 * resulting from this level per the above.
2731 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2732 unsigned long wl, unsigned long wg)
2739 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2741 s64 this_load, load;
2742 int idx, this_cpu, prev_cpu;
2743 unsigned long tl_per_task;
2744 struct task_group *tg;
2745 unsigned long weight;
2749 this_cpu = smp_processor_id();
2750 prev_cpu = task_cpu(p);
2751 load = source_load(prev_cpu, idx);
2752 this_load = target_load(this_cpu, idx);
2755 * If sync wakeup then subtract the (maximum possible)
2756 * effect of the currently running task from the load
2757 * of the current CPU:
2760 tg = task_group(current);
2761 weight = current->se.load.weight;
2763 this_load += effective_load(tg, this_cpu, -weight, -weight);
2764 load += effective_load(tg, prev_cpu, 0, -weight);
2768 weight = p->se.load.weight;
2771 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2772 * due to the sync cause above having dropped this_load to 0, we'll
2773 * always have an imbalance, but there's really nothing you can do
2774 * about that, so that's good too.
2776 * Otherwise check if either cpus are near enough in load to allow this
2777 * task to be woken on this_cpu.
2779 if (this_load > 0) {
2780 s64 this_eff_load, prev_eff_load;
2782 this_eff_load = 100;
2783 this_eff_load *= power_of(prev_cpu);
2784 this_eff_load *= this_load +
2785 effective_load(tg, this_cpu, weight, weight);
2787 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2788 prev_eff_load *= power_of(this_cpu);
2789 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2791 balanced = this_eff_load <= prev_eff_load;
2796 * If the currently running task will sleep within
2797 * a reasonable amount of time then attract this newly
2800 if (sync && balanced)
2803 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2804 tl_per_task = cpu_avg_load_per_task(this_cpu);
2807 (this_load <= load &&
2808 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2810 * This domain has SD_WAKE_AFFINE and
2811 * p is cache cold in this domain, and
2812 * there is no bad imbalance.
2814 schedstat_inc(sd, ttwu_move_affine);
2815 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2823 * find_idlest_group finds and returns the least busy CPU group within the
2826 static struct sched_group *
2827 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2828 int this_cpu, int load_idx)
2830 struct sched_group *idlest = NULL, *group = sd->groups;
2831 unsigned long min_load = ULONG_MAX, this_load = 0;
2832 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2835 unsigned long load, avg_load;
2839 /* Skip over this group if it has no CPUs allowed */
2840 if (!cpumask_intersects(sched_group_cpus(group),
2841 tsk_cpus_allowed(p)))
2844 local_group = cpumask_test_cpu(this_cpu,
2845 sched_group_cpus(group));
2847 /* Tally up the load of all CPUs in the group */
2850 for_each_cpu(i, sched_group_cpus(group)) {
2851 /* Bias balancing toward cpus of our domain */
2853 load = source_load(i, load_idx);
2855 load = target_load(i, load_idx);
2860 /* Adjust by relative CPU power of the group */
2861 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2864 this_load = avg_load;
2865 } else if (avg_load < min_load) {
2866 min_load = avg_load;
2869 } while (group = group->next, group != sd->groups);
2871 if (!idlest || 100*this_load < imbalance*min_load)
2877 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2880 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2882 unsigned long load, min_load = ULONG_MAX;
2886 /* Traverse only the allowed CPUs */
2887 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2888 load = weighted_cpuload(i);
2890 if (load < min_load || (load == min_load && i == this_cpu)) {
2900 * Try and locate an idle CPU in the sched_domain.
2902 static int select_idle_sibling(struct task_struct *p, int target)
2904 int cpu = smp_processor_id();
2905 int prev_cpu = task_cpu(p);
2906 struct sched_domain *sd;
2907 struct sched_group *sg;
2911 * If the task is going to be woken-up on this cpu and if it is
2912 * already idle, then it is the right target.
2914 if (target == cpu && idle_cpu(cpu))
2918 * If the task is going to be woken-up on the cpu where it previously
2919 * ran and if it is currently idle, then it the right target.
2921 if (target == prev_cpu && idle_cpu(prev_cpu))
2925 * Otherwise, iterate the domains and find an elegible idle cpu.
2927 sd = rcu_dereference(per_cpu(sd_llc, target));
2928 for_each_lower_domain(sd) {
2931 if (!cpumask_intersects(sched_group_cpus(sg),
2932 tsk_cpus_allowed(p)))
2935 for_each_cpu(i, sched_group_cpus(sg)) {
2940 target = cpumask_first_and(sched_group_cpus(sg),
2941 tsk_cpus_allowed(p));
2945 } while (sg != sd->groups);
2951 #ifdef CONFIG_SCHED_NUMA
2952 static inline bool pick_numa_rand(int n)
2954 return !(get_random_int() % n);
2958 * Pick a random elegible CPU in the target node, hopefully faster
2959 * than doing a least-loaded scan.
2961 static int numa_select_node_cpu(struct task_struct *p, int node)
2963 int weight = cpumask_weight(cpumask_of_node(node));
2966 for_each_cpu_and(i, cpumask_of_node(node), tsk_cpus_allowed(p)) {
2967 if (cpu < 0 || pick_numa_rand(weight))
2974 static int numa_select_node_cpu(struct task_struct *p, int node)
2978 #endif /* CONFIG_SCHED_NUMA */
2981 * sched_balance_self: balance the current task (running on cpu) in domains
2982 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2985 * Balance, ie. select the least loaded group.
2987 * Returns the target CPU number, or the same CPU if no balancing is needed.
2989 * preempt must be disabled.
2992 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2994 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2995 int cpu = smp_processor_id();
2996 int prev_cpu = task_cpu(p);
2998 int want_affine = 0;
2999 int sync = wake_flags & WF_SYNC;
3000 int node = tsk_home_node(p);
3002 if (p->nr_cpus_allowed == 1)
3005 if (sd_flag & SD_BALANCE_WAKE) {
3006 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3012 if (sched_feat_numa(NUMA_TTWU_BIAS) && node != -1) {
3014 * For fork,exec find the idlest cpu in the home-node.
3016 if (sd_flag & (SD_BALANCE_FORK|SD_BALANCE_EXEC)) {
3017 int node_cpu = numa_select_node_cpu(p, node);
3021 new_cpu = cpu = node_cpu;
3022 sd = per_cpu(sd_node, cpu);
3027 * For wake, pretend we were running in the home-node.
3029 if (cpu_to_node(prev_cpu) != node) {
3030 int node_cpu = numa_select_node_cpu(p, node);
3034 if (sched_feat_numa(NUMA_TTWU_TO))
3037 prev_cpu = node_cpu;
3042 for_each_domain(cpu, tmp) {
3043 if (!(tmp->flags & SD_LOAD_BALANCE))
3047 * If both cpu and prev_cpu are part of this domain,
3048 * cpu is a valid SD_WAKE_AFFINE target.
3050 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3051 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3056 if (tmp->flags & sd_flag)
3061 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3064 new_cpu = select_idle_sibling(p, prev_cpu);
3070 int load_idx = sd->forkexec_idx;
3071 struct sched_group *group;
3074 if (!(sd->flags & sd_flag)) {
3079 if (sd_flag & SD_BALANCE_WAKE)
3080 load_idx = sd->wake_idx;
3082 group = find_idlest_group(sd, p, cpu, load_idx);
3088 new_cpu = find_idlest_cpu(group, p, cpu);
3089 if (new_cpu == -1 || new_cpu == cpu) {
3090 /* Now try balancing at a lower domain level of cpu */
3095 /* Now try balancing at a lower domain level of new_cpu */
3097 weight = sd->span_weight;
3099 for_each_domain(cpu, tmp) {
3100 if (weight <= tmp->span_weight)
3102 if (tmp->flags & sd_flag)
3105 /* while loop will break here if sd == NULL */
3112 #endif /* CONFIG_SMP */
3114 static unsigned long
3115 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3117 unsigned long gran = sysctl_sched_wakeup_granularity;
3120 * Since its curr running now, convert the gran from real-time
3121 * to virtual-time in his units.
3123 * By using 'se' instead of 'curr' we penalize light tasks, so
3124 * they get preempted easier. That is, if 'se' < 'curr' then
3125 * the resulting gran will be larger, therefore penalizing the
3126 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3127 * be smaller, again penalizing the lighter task.
3129 * This is especially important for buddies when the leftmost
3130 * task is higher priority than the buddy.
3132 return calc_delta_fair(gran, se);
3136 * Should 'se' preempt 'curr'.
3150 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3152 s64 gran, vdiff = curr->vruntime - se->vruntime;
3157 gran = wakeup_gran(curr, se);
3164 static void set_last_buddy(struct sched_entity *se)
3166 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3169 for_each_sched_entity(se)
3170 cfs_rq_of(se)->last = se;
3173 static void set_next_buddy(struct sched_entity *se)
3175 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3178 for_each_sched_entity(se)
3179 cfs_rq_of(se)->next = se;
3182 static void set_skip_buddy(struct sched_entity *se)
3184 for_each_sched_entity(se)
3185 cfs_rq_of(se)->skip = se;
3189 * Preempt the current task with a newly woken task if needed:
3191 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3193 struct task_struct *curr = rq->curr;
3194 struct sched_entity *se = &curr->se, *pse = &p->se;
3195 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3196 int scale = cfs_rq->nr_running >= sched_nr_latency;
3197 int next_buddy_marked = 0;
3199 if (unlikely(se == pse))
3203 * This is possible from callers such as move_task(), in which we
3204 * unconditionally check_prempt_curr() after an enqueue (which may have
3205 * lead to a throttle). This both saves work and prevents false
3206 * next-buddy nomination below.
3208 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3211 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3212 set_next_buddy(pse);
3213 next_buddy_marked = 1;
3217 * We can come here with TIF_NEED_RESCHED already set from new task
3220 * Note: this also catches the edge-case of curr being in a throttled
3221 * group (e.g. via set_curr_task), since update_curr() (in the
3222 * enqueue of curr) will have resulted in resched being set. This
3223 * prevents us from potentially nominating it as a false LAST_BUDDY
3226 if (test_tsk_need_resched(curr))
3229 /* Idle tasks are by definition preempted by non-idle tasks. */
3230 if (unlikely(curr->policy == SCHED_IDLE) &&
3231 likely(p->policy != SCHED_IDLE))
3235 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3236 * is driven by the tick):
3238 if (unlikely(p->policy != SCHED_NORMAL))
3241 find_matching_se(&se, &pse);
3242 update_curr(cfs_rq_of(se));
3244 if (wakeup_preempt_entity(se, pse) == 1) {
3246 * Bias pick_next to pick the sched entity that is
3247 * triggering this preemption.
3249 if (!next_buddy_marked)
3250 set_next_buddy(pse);
3259 * Only set the backward buddy when the current task is still
3260 * on the rq. This can happen when a wakeup gets interleaved
3261 * with schedule on the ->pre_schedule() or idle_balance()
3262 * point, either of which can * drop the rq lock.
3264 * Also, during early boot the idle thread is in the fair class,
3265 * for obvious reasons its a bad idea to schedule back to it.
3267 if (unlikely(!se->on_rq || curr == rq->idle))
3270 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3274 static struct task_struct *pick_next_task_fair(struct rq *rq)
3276 struct task_struct *p;
3277 struct cfs_rq *cfs_rq = &rq->cfs;
3278 struct sched_entity *se;
3280 if (!cfs_rq->nr_running)
3284 se = pick_next_entity(cfs_rq);
3285 set_next_entity(cfs_rq, se);
3286 cfs_rq = group_cfs_rq(se);
3290 if (hrtick_enabled(rq))
3291 hrtick_start_fair(rq, p);
3297 * Account for a descheduled task:
3299 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3301 struct sched_entity *se = &prev->se;
3302 struct cfs_rq *cfs_rq;
3304 for_each_sched_entity(se) {
3305 cfs_rq = cfs_rq_of(se);
3306 put_prev_entity(cfs_rq, se);
3311 * sched_yield() is very simple
3313 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3315 static void yield_task_fair(struct rq *rq)
3317 struct task_struct *curr = rq->curr;
3318 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3319 struct sched_entity *se = &curr->se;
3322 * Are we the only task in the tree?
3324 if (unlikely(rq->nr_running == 1))
3327 clear_buddies(cfs_rq, se);
3329 if (curr->policy != SCHED_BATCH) {
3330 update_rq_clock(rq);
3332 * Update run-time statistics of the 'current'.
3334 update_curr(cfs_rq);
3336 * Tell update_rq_clock() that we've just updated,
3337 * so we don't do microscopic update in schedule()
3338 * and double the fastpath cost.
3340 rq->skip_clock_update = 1;
3346 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3348 struct sched_entity *se = &p->se;
3350 /* throttled hierarchies are not runnable */
3351 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3354 /* Tell the scheduler that we'd really like pse to run next. */
3357 yield_task_fair(rq);
3363 /**************************************************
3364 * Fair scheduling class load-balancing methods:
3367 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3369 #define LBF_ALL_PINNED 0x01
3370 #define LBF_NEED_BREAK 0x02
3371 #define LBF_SOME_PINNED 0x04
3374 struct sched_domain *sd;
3382 struct cpumask *dst_grpmask;
3384 enum cpu_idle_type idle;
3386 /* The set of CPUs under consideration for load-balancing */
3387 struct cpumask *cpus;
3391 struct list_head *tasks;
3394 unsigned int loop_break;
3395 unsigned int loop_max;
3397 struct rq * (*find_busiest_queue)(struct lb_env *,
3398 struct sched_group *);
3402 * move_task - move a task from one runqueue to another runqueue.
3403 * Both runqueues must be locked.
3405 static void move_task(struct task_struct *p, struct lb_env *env)
3407 deactivate_task(env->src_rq, p, 0);
3408 set_task_cpu(p, env->dst_cpu);
3409 activate_task(env->dst_rq, p, 0);
3410 check_preempt_curr(env->dst_rq, p, 0);
3413 static int task_numa_hot(struct task_struct *p, struct lb_env *env)
3415 int from_dist, to_dist;
3416 int node = tsk_home_node(p);
3418 if (!sched_feat_numa(NUMA_HOT) || node == -1)
3419 return 0; /* no node preference */
3421 from_dist = node_distance(cpu_to_node(env->src_cpu), node);
3422 to_dist = node_distance(cpu_to_node(env->dst_cpu), node);
3424 if (to_dist < from_dist)
3425 return 0; /* getting closer is ok */
3427 return 1; /* stick to where we are */
3431 * Is this task likely cache-hot:
3434 task_hot(struct task_struct *p, struct lb_env *env)
3438 if (p->sched_class != &fair_sched_class)
3441 if (unlikely(p->policy == SCHED_IDLE))
3445 * Buddy candidates are cache hot:
3447 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3448 (&p->se == cfs_rq_of(&p->se)->next ||
3449 &p->se == cfs_rq_of(&p->se)->last))
3452 if (sysctl_sched_migration_cost == -1)
3454 if (sysctl_sched_migration_cost == 0)
3457 delta = env->src_rq->clock_task - p->se.exec_start;
3459 return delta < (s64)sysctl_sched_migration_cost;
3463 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3466 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3468 int tsk_cache_hot = 0;
3470 * We do not migrate tasks that are:
3471 * 1) running (obviously), or
3472 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3473 * 3) are cache-hot on their current CPU.
3475 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3478 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3481 * Remember if this task can be migrated to any other cpu in
3482 * our sched_group. We may want to revisit it if we couldn't
3483 * meet load balance goals by pulling other tasks on src_cpu.
3485 * Also avoid computing new_dst_cpu if we have already computed
3486 * one in current iteration.
3488 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3491 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3492 tsk_cpus_allowed(p));
3493 if (new_dst_cpu < nr_cpu_ids) {
3494 env->flags |= LBF_SOME_PINNED;
3495 env->new_dst_cpu = new_dst_cpu;
3500 /* Record that we found atleast one task that could run on dst_cpu */
3501 env->flags &= ~LBF_ALL_PINNED;
3503 if (task_running(env->src_rq, p)) {
3504 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3509 * Aggressive migration if:
3510 * 1) task is cache cold, or
3511 * 2) too many balance attempts have failed.
3514 tsk_cache_hot = task_hot(p, env);
3515 if (env->idle == CPU_NOT_IDLE)
3516 tsk_cache_hot |= task_numa_hot(p, env);
3517 if (!tsk_cache_hot ||
3518 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3519 #ifdef CONFIG_SCHEDSTATS
3520 if (tsk_cache_hot) {
3521 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3522 schedstat_inc(p, se.statistics.nr_forced_migrations);
3528 if (tsk_cache_hot) {
3529 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3536 * move_one_task tries to move exactly one task from busiest to this_rq, as
3537 * part of active balancing operations within "domain".
3538 * Returns 1 if successful and 0 otherwise.
3540 * Called with both runqueues locked.
3542 static int __move_one_task(struct lb_env *env)
3544 struct task_struct *p, *n;
3546 list_for_each_entry_safe(p, n, env->tasks, se.group_node) {
3547 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3550 if (!can_migrate_task(p, env))
3555 * Right now, this is only the second place move_task()
3556 * is called, so we can safely collect move_task()
3557 * stats here rather than inside move_task().
3559 schedstat_inc(env->sd, lb_gained[env->idle]);
3565 static int move_one_task(struct lb_env *env)
3567 if (sched_feat_numa(NUMA_PULL)) {
3568 env->tasks = offnode_tasks(env->src_rq);
3569 if (__move_one_task(env))
3573 env->tasks = &env->src_rq->cfs_tasks;
3574 if (__move_one_task(env))
3580 static const unsigned int sched_nr_migrate_break = 32;
3583 * move_tasks tries to move up to imbalance weighted load from busiest to
3584 * this_rq, as part of a balancing operation within domain "sd".
3585 * Returns 1 if successful and 0 otherwise.
3587 * Called with both runqueues locked.
3589 static int move_tasks(struct lb_env *env)
3591 struct task_struct *p;
3595 if (env->imbalance <= 0)
3599 while (!list_empty(env->tasks)) {
3600 p = list_first_entry(env->tasks, struct task_struct, se.group_node);
3603 /* We've more or less seen every task there is, call it quits */
3604 if (env->loop > env->loop_max)
3607 /* take a breather every nr_migrate tasks */
3608 if (env->loop > env->loop_break) {
3609 env->loop_break += sched_nr_migrate_break;
3610 env->flags |= LBF_NEED_BREAK;
3614 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3617 load = task_h_load(p);
3619 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3622 if ((load / 2) > env->imbalance)
3625 if (!can_migrate_task(p, env))
3630 env->imbalance -= load;
3632 #ifdef CONFIG_PREEMPT
3634 * NEWIDLE balancing is a source of latency, so preemptible
3635 * kernels will stop after the first task is pulled to minimize
3636 * the critical section.
3638 if (env->idle == CPU_NEWLY_IDLE)
3643 * We only want to steal up to the prescribed amount of
3646 if (env->imbalance <= 0)
3651 list_move_tail(&p->se.group_node, env->tasks);
3654 if (env->tasks == offnode_tasks(env->src_rq)) {
3655 env->tasks = &env->src_rq->cfs_tasks;
3662 * Right now, this is one of only two places move_task() is called,
3663 * so we can safely collect move_task() stats here rather than
3664 * inside move_task().
3666 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3671 #ifdef CONFIG_FAIR_GROUP_SCHED
3673 * update tg->load_weight by folding this cpu's load_avg
3675 static int update_shares_cpu(struct task_group *tg, int cpu)
3677 struct cfs_rq *cfs_rq;
3678 unsigned long flags;
3685 cfs_rq = tg->cfs_rq[cpu];
3687 raw_spin_lock_irqsave(&rq->lock, flags);
3689 update_rq_clock(rq);
3690 update_cfs_load(cfs_rq, 1);
3693 * We need to update shares after updating tg->load_weight in
3694 * order to adjust the weight of groups with long running tasks.
3696 update_cfs_shares(cfs_rq);
3698 raw_spin_unlock_irqrestore(&rq->lock, flags);
3703 static void update_shares(int cpu)
3705 struct cfs_rq *cfs_rq;
3706 struct rq *rq = cpu_rq(cpu);
3710 * Iterates the task_group tree in a bottom up fashion, see
3711 * list_add_leaf_cfs_rq() for details.
3713 for_each_leaf_cfs_rq(rq, cfs_rq) {
3714 /* throttled entities do not contribute to load */
3715 if (throttled_hierarchy(cfs_rq))
3718 update_shares_cpu(cfs_rq->tg, cpu);
3724 * Compute the cpu's hierarchical load factor for each task group.
3725 * This needs to be done in a top-down fashion because the load of a child
3726 * group is a fraction of its parents load.
3728 static int tg_load_down(struct task_group *tg, void *data)
3731 long cpu = (long)data;
3734 load = cpu_rq(cpu)->load.weight;
3736 load = tg->parent->cfs_rq[cpu]->h_load;
3737 load *= tg->se[cpu]->load.weight;
3738 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3741 tg->cfs_rq[cpu]->h_load = load;
3746 static void update_h_load(long cpu)
3748 struct rq *rq = cpu_rq(cpu);
3749 unsigned long now = jiffies;
3751 if (rq->h_load_throttle == now)
3754 rq->h_load_throttle = now;
3757 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3761 static unsigned long task_h_load(struct task_struct *p)
3763 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3766 load = p->se.load.weight;
3767 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3772 static inline void update_shares(int cpu)
3776 static inline void update_h_load(long cpu)
3780 static unsigned long task_h_load(struct task_struct *p)
3782 return p->se.load.weight;
3786 /********** Helpers for find_busiest_group ************************/
3788 * sd_lb_stats - Structure to store the statistics of a sched_domain
3789 * during load balancing.
3791 struct sd_lb_stats {
3792 struct sched_group *busiest; /* Busiest group in this sd */
3793 struct sched_group *this; /* Local group in this sd */
3794 unsigned long total_load; /* Total load of all groups in sd */
3795 unsigned long total_pwr; /* Total power of all groups in sd */
3796 unsigned long avg_load; /* Average load across all groups in sd */
3798 /** Statistics of this group */
3799 unsigned long this_load;
3800 unsigned long this_load_per_task;
3801 unsigned long this_nr_running;
3802 unsigned long this_has_capacity;
3803 unsigned int this_idle_cpus;
3805 /* Statistics of the busiest group */
3806 unsigned int busiest_idle_cpus;
3807 unsigned long max_load;
3808 unsigned long busiest_load_per_task;
3809 unsigned long busiest_nr_running;
3810 unsigned long busiest_group_capacity;
3811 unsigned long busiest_has_capacity;
3812 unsigned int busiest_group_weight;
3814 int group_imb; /* Is there imbalance in this sd */
3815 #ifdef CONFIG_SCHED_NUMA
3816 struct sched_group *numa_group; /* group which has offnode_tasks */
3817 unsigned long numa_group_weight;
3818 unsigned long numa_group_running;
3823 * sg_lb_stats - stats of a sched_group required for load_balancing
3825 struct sg_lb_stats {
3826 unsigned long avg_load; /*Avg load across the CPUs of the group */
3827 unsigned long group_load; /* Total load over the CPUs of the group */
3828 unsigned long sum_nr_running; /* Nr tasks running in the group */
3829 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3830 unsigned long group_capacity;
3831 unsigned long idle_cpus;
3832 unsigned long group_weight;
3833 int group_imb; /* Is there an imbalance in the group ? */
3834 int group_has_capacity; /* Is there extra capacity in the group? */
3835 #ifdef CONFIG_SCHED_NUMA
3836 unsigned long numa_weight;
3837 unsigned long numa_running;
3842 * get_sd_load_idx - Obtain the load index for a given sched domain.
3843 * @sd: The sched_domain whose load_idx is to be obtained.
3844 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3846 static inline int get_sd_load_idx(struct sched_domain *sd,
3847 enum cpu_idle_type idle)
3853 load_idx = sd->busy_idx;
3856 case CPU_NEWLY_IDLE:
3857 load_idx = sd->newidle_idx;
3860 load_idx = sd->idle_idx;
3867 #ifdef CONFIG_SCHED_NUMA
3868 static inline void update_sg_numa_stats(struct sg_lb_stats *sgs, struct rq *rq)
3870 sgs->numa_weight += rq->offnode_weight;
3871 sgs->numa_running += rq->offnode_running;
3875 * Since the offnode lists are indiscriminate (they contain tasks for all other
3876 * nodes) it is impossible to say if there's any task on there that wants to
3877 * move towards the pulling cpu. Therefore select a random offnode list to pull
3878 * from such that eventually we'll try them all.
3880 * Select a random group that has offnode tasks as sds->numa_group
3882 static inline void update_sd_numa_stats(struct sched_domain *sd,
3883 struct sched_group *group, struct sd_lb_stats *sds,
3884 int local_group, struct sg_lb_stats *sgs)
3886 if (!(sd->flags & SD_NUMA))
3892 if (!sgs->numa_running)
3895 if (!sds->numa_group || pick_numa_rand(sd->span_weight / group->group_weight)) {
3896 sds->numa_group = group;
3897 sds->numa_group_weight = sgs->numa_weight;
3898 sds->numa_group_running = sgs->numa_running;
3903 * Pick a random queue from the group that has offnode tasks.
3905 static struct rq *find_busiest_numa_queue(struct lb_env *env,
3906 struct sched_group *group)
3908 struct rq *busiest = NULL, *rq;
3911 for_each_cpu_and(cpu, sched_group_cpus(group), env->cpus) {
3913 if (!rq->offnode_running)
3915 if (!busiest || pick_numa_rand(group->group_weight))
3923 * Called in case of no other imbalance, if there is a queue running offnode
3924 * tasksk we'll say we're imbalanced anyway to nudge these tasks towards their
3927 static inline int check_numa_busiest_group(struct lb_env *env, struct sd_lb_stats *sds)
3929 if (!sched_feat(NUMA_PULL_BIAS))
3932 if (!sds->numa_group)
3935 env->imbalance = sds->numa_group_weight / sds->numa_group_running;
3936 sds->busiest = sds->numa_group;
3937 env->find_busiest_queue = find_busiest_numa_queue;
3941 static inline bool need_active_numa_balance(struct lb_env *env)
3943 return env->find_busiest_queue == find_busiest_numa_queue &&
3944 env->src_rq->offnode_running == 1 &&
3945 env->src_rq->nr_running == 1;
3948 #else /* CONFIG_SCHED_NUMA */
3950 static inline void update_sg_numa_stats(struct sg_lb_stats *sgs, struct rq *rq)
3954 static inline void update_sd_numa_stats(struct sched_domain *sd,
3955 struct sched_group *group, struct sd_lb_stats *sds,
3956 int local_group, struct sg_lb_stats *sgs)
3960 static inline int check_numa_busiest_group(struct lb_env *env, struct sd_lb_stats *sds)
3965 static inline bool need_active_numa_balance(struct lb_env *env)
3969 #endif /* CONFIG_SCHED_NUMA */
3971 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3973 return SCHED_POWER_SCALE;
3976 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3978 return default_scale_freq_power(sd, cpu);
3981 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3983 unsigned long weight = sd->span_weight;
3984 unsigned long smt_gain = sd->smt_gain;
3991 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3993 return default_scale_smt_power(sd, cpu);
3996 unsigned long scale_rt_power(int cpu)
3998 struct rq *rq = cpu_rq(cpu);
3999 u64 total, available, age_stamp, avg;
4002 * Since we're reading these variables without serialization make sure
4003 * we read them once before doing sanity checks on them.
4005 age_stamp = ACCESS_ONCE(rq->age_stamp);
4006 avg = ACCESS_ONCE(rq->rt_avg);
4008 total = sched_avg_period() + (rq->clock - age_stamp);
4010 if (unlikely(total < avg)) {
4011 /* Ensures that power won't end up being negative */
4014 available = total - avg;
4017 if (unlikely((s64)total < SCHED_POWER_SCALE))
4018 total = SCHED_POWER_SCALE;
4020 total >>= SCHED_POWER_SHIFT;
4022 return div_u64(available, total);
4025 static void update_cpu_power(struct sched_domain *sd, int cpu)
4027 unsigned long weight = sd->span_weight;
4028 unsigned long power = SCHED_POWER_SCALE;
4029 struct sched_group *sdg = sd->groups;
4031 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4032 if (sched_feat(ARCH_POWER))
4033 power *= arch_scale_smt_power(sd, cpu);
4035 power *= default_scale_smt_power(sd, cpu);
4037 power >>= SCHED_POWER_SHIFT;
4040 sdg->sgp->power_orig = power;
4042 if (sched_feat(ARCH_POWER))
4043 power *= arch_scale_freq_power(sd, cpu);
4045 power *= default_scale_freq_power(sd, cpu);
4047 power >>= SCHED_POWER_SHIFT;
4049 power *= scale_rt_power(cpu);
4050 power >>= SCHED_POWER_SHIFT;
4055 cpu_rq(cpu)->cpu_power = power;
4056 sdg->sgp->power = power;
4059 void update_group_power(struct sched_domain *sd, int cpu)
4061 struct sched_domain *child = sd->child;
4062 struct sched_group *group, *sdg = sd->groups;
4063 unsigned long power;
4064 unsigned long interval;
4066 interval = msecs_to_jiffies(sd->balance_interval);
4067 interval = clamp(interval, 1UL, max_load_balance_interval);
4068 sdg->sgp->next_update = jiffies + interval;
4071 update_cpu_power(sd, cpu);
4077 if (child->flags & SD_OVERLAP) {
4079 * SD_OVERLAP domains cannot assume that child groups
4080 * span the current group.
4083 for_each_cpu(cpu, sched_group_cpus(sdg))
4084 power += power_of(cpu);
4087 * !SD_OVERLAP domains can assume that child groups
4088 * span the current group.
4091 group = child->groups;
4093 power += group->sgp->power;
4094 group = group->next;
4095 } while (group != child->groups);
4098 sdg->sgp->power_orig = sdg->sgp->power = power;
4102 * Try and fix up capacity for tiny siblings, this is needed when
4103 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4104 * which on its own isn't powerful enough.
4106 * See update_sd_pick_busiest() and check_asym_packing().
4109 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4112 * Only siblings can have significantly less than SCHED_POWER_SCALE
4114 if (!(sd->flags & SD_SHARE_CPUPOWER))
4118 * If ~90% of the cpu_power is still there, we're good.
4120 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4127 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4128 * @env: The load balancing environment.
4129 * @group: sched_group whose statistics are to be updated.
4130 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4131 * @local_group: Does group contain this_cpu.
4132 * @balance: Should we balance.
4133 * @sgs: variable to hold the statistics for this group.
4135 static inline void update_sg_lb_stats(struct lb_env *env,
4136 struct sched_group *group, int load_idx,
4137 int local_group, int *balance, struct sg_lb_stats *sgs)
4139 unsigned long nr_running, max_nr_running, min_nr_running;
4140 unsigned long load, max_cpu_load, min_cpu_load;
4141 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4142 unsigned long avg_load_per_task = 0;
4146 balance_cpu = group_balance_cpu(group);
4148 /* Tally up the load of all CPUs in the group */
4150 min_cpu_load = ~0UL;
4152 min_nr_running = ~0UL;
4154 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4155 struct rq *rq = cpu_rq(i);
4157 nr_running = rq->nr_running;
4159 /* Bias balancing toward cpus of our domain */
4161 if (idle_cpu(i) && !first_idle_cpu &&
4162 cpumask_test_cpu(i, sched_group_mask(group))) {
4167 load = target_load(i, load_idx);
4169 load = source_load(i, load_idx);
4170 if (load > max_cpu_load)
4171 max_cpu_load = load;
4172 if (min_cpu_load > load)
4173 min_cpu_load = load;
4175 if (nr_running > max_nr_running)
4176 max_nr_running = nr_running;
4177 if (min_nr_running > nr_running)
4178 min_nr_running = nr_running;
4181 sgs->group_load += load;
4182 sgs->sum_nr_running += nr_running;
4183 sgs->sum_weighted_load += weighted_cpuload(i);
4187 update_sg_numa_stats(sgs, rq);
4191 * First idle cpu or the first cpu(busiest) in this sched group
4192 * is eligible for doing load balancing at this and above
4193 * domains. In the newly idle case, we will allow all the cpu's
4194 * to do the newly idle load balance.
4197 if (env->idle != CPU_NEWLY_IDLE) {
4198 if (balance_cpu != env->dst_cpu) {
4202 update_group_power(env->sd, env->dst_cpu);
4203 } else if (time_after_eq(jiffies, group->sgp->next_update))
4204 update_group_power(env->sd, env->dst_cpu);
4207 /* Adjust by relative CPU power of the group */
4208 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4211 * Consider the group unbalanced when the imbalance is larger
4212 * than the average weight of a task.
4214 * APZ: with cgroup the avg task weight can vary wildly and
4215 * might not be a suitable number - should we keep a
4216 * normalized nr_running number somewhere that negates
4219 if (sgs->sum_nr_running)
4220 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4222 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4223 (max_nr_running - min_nr_running) > 1)
4226 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4228 if (!sgs->group_capacity)
4229 sgs->group_capacity = fix_small_capacity(env->sd, group);
4230 sgs->group_weight = group->group_weight;
4232 if (sgs->group_capacity > sgs->sum_nr_running)
4233 sgs->group_has_capacity = 1;
4237 * update_sd_pick_busiest - return 1 on busiest group
4238 * @env: The load balancing environment.
4239 * @sds: sched_domain statistics
4240 * @sg: sched_group candidate to be checked for being the busiest
4241 * @sgs: sched_group statistics
4243 * Determine if @sg is a busier group than the previously selected
4246 static bool update_sd_pick_busiest(struct lb_env *env,
4247 struct sd_lb_stats *sds,
4248 struct sched_group *sg,
4249 struct sg_lb_stats *sgs)
4251 if (sgs->avg_load <= sds->max_load)
4254 if (sgs->sum_nr_running > sgs->group_capacity)
4261 * ASYM_PACKING needs to move all the work to the lowest
4262 * numbered CPUs in the group, therefore mark all groups
4263 * higher than ourself as busy.
4265 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4266 env->dst_cpu < group_first_cpu(sg)) {
4270 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4278 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4279 * @env: The load balancing environment.
4280 * @balance: Should we balance.
4281 * @sds: variable to hold the statistics for this sched_domain.
4283 static inline void update_sd_lb_stats(struct lb_env *env,
4284 int *balance, struct sd_lb_stats *sds)
4286 struct sched_domain *child = env->sd->child;
4287 struct sched_group *sg = env->sd->groups;
4288 struct sg_lb_stats sgs;
4289 int load_idx, prefer_sibling = 0;
4291 if (child && child->flags & SD_PREFER_SIBLING)
4294 load_idx = get_sd_load_idx(env->sd, env->idle);
4299 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4300 memset(&sgs, 0, sizeof(sgs));
4301 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4303 if (local_group && !(*balance))
4306 sds->total_load += sgs.group_load;
4307 sds->total_pwr += sg->sgp->power;
4310 * In case the child domain prefers tasks go to siblings
4311 * first, lower the sg capacity to one so that we'll try
4312 * and move all the excess tasks away. We lower the capacity
4313 * of a group only if the local group has the capacity to fit
4314 * these excess tasks, i.e. nr_running < group_capacity. The
4315 * extra check prevents the case where you always pull from the
4316 * heaviest group when it is already under-utilized (possible
4317 * with a large weight task outweighs the tasks on the system).
4319 if (prefer_sibling && !local_group && sds->this_has_capacity)
4320 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4323 sds->this_load = sgs.avg_load;
4325 sds->this_nr_running = sgs.sum_nr_running;
4326 sds->this_load_per_task = sgs.sum_weighted_load;
4327 sds->this_has_capacity = sgs.group_has_capacity;
4328 sds->this_idle_cpus = sgs.idle_cpus;
4329 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4330 sds->max_load = sgs.avg_load;
4332 sds->busiest_nr_running = sgs.sum_nr_running;
4333 sds->busiest_idle_cpus = sgs.idle_cpus;
4334 sds->busiest_group_capacity = sgs.group_capacity;
4335 sds->busiest_load_per_task = sgs.sum_weighted_load;
4336 sds->busiest_has_capacity = sgs.group_has_capacity;
4337 sds->busiest_group_weight = sgs.group_weight;
4338 sds->group_imb = sgs.group_imb;
4341 update_sd_numa_stats(env->sd, sg, sds, local_group, &sgs);
4344 } while (sg != env->sd->groups);
4348 * check_asym_packing - Check to see if the group is packed into the
4351 * This is primarily intended to used at the sibling level. Some
4352 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4353 * case of POWER7, it can move to lower SMT modes only when higher
4354 * threads are idle. When in lower SMT modes, the threads will
4355 * perform better since they share less core resources. Hence when we
4356 * have idle threads, we want them to be the higher ones.
4358 * This packing function is run on idle threads. It checks to see if
4359 * the busiest CPU in this domain (core in the P7 case) has a higher
4360 * CPU number than the packing function is being run on. Here we are
4361 * assuming lower CPU number will be equivalent to lower a SMT thread
4364 * Returns 1 when packing is required and a task should be moved to
4365 * this CPU. The amount of the imbalance is returned in *imbalance.
4367 * @env: The load balancing environment.
4368 * @sds: Statistics of the sched_domain which is to be packed
4370 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4374 if (!(env->sd->flags & SD_ASYM_PACKING))
4380 busiest_cpu = group_first_cpu(sds->busiest);
4381 if (env->dst_cpu > busiest_cpu)
4384 env->imbalance = DIV_ROUND_CLOSEST(
4385 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4391 * fix_small_imbalance - Calculate the minor imbalance that exists
4392 * amongst the groups of a sched_domain, during
4394 * @env: The load balancing environment.
4395 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4398 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4400 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4401 unsigned int imbn = 2;
4402 unsigned long scaled_busy_load_per_task;
4404 if (sds->this_nr_running) {
4405 sds->this_load_per_task /= sds->this_nr_running;
4406 if (sds->busiest_load_per_task >
4407 sds->this_load_per_task)
4410 sds->this_load_per_task =
4411 cpu_avg_load_per_task(env->dst_cpu);
4414 scaled_busy_load_per_task = sds->busiest_load_per_task
4415 * SCHED_POWER_SCALE;
4416 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4418 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4419 (scaled_busy_load_per_task * imbn)) {
4420 env->imbalance = sds->busiest_load_per_task;
4425 * OK, we don't have enough imbalance to justify moving tasks,
4426 * however we may be able to increase total CPU power used by
4430 pwr_now += sds->busiest->sgp->power *
4431 min(sds->busiest_load_per_task, sds->max_load);
4432 pwr_now += sds->this->sgp->power *
4433 min(sds->this_load_per_task, sds->this_load);
4434 pwr_now /= SCHED_POWER_SCALE;
4436 /* Amount of load we'd subtract */
4437 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4438 sds->busiest->sgp->power;
4439 if (sds->max_load > tmp)
4440 pwr_move += sds->busiest->sgp->power *
4441 min(sds->busiest_load_per_task, sds->max_load - tmp);
4443 /* Amount of load we'd add */
4444 if (sds->max_load * sds->busiest->sgp->power <
4445 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4446 tmp = (sds->max_load * sds->busiest->sgp->power) /
4447 sds->this->sgp->power;
4449 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4450 sds->this->sgp->power;
4451 pwr_move += sds->this->sgp->power *
4452 min(sds->this_load_per_task, sds->this_load + tmp);
4453 pwr_move /= SCHED_POWER_SCALE;
4455 /* Move if we gain throughput */
4456 if (pwr_move > pwr_now)
4457 env->imbalance = sds->busiest_load_per_task;
4461 * calculate_imbalance - Calculate the amount of imbalance present within the
4462 * groups of a given sched_domain during load balance.
4463 * @env: load balance environment
4464 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4466 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4468 unsigned long max_pull, load_above_capacity = ~0UL;
4470 sds->busiest_load_per_task /= sds->busiest_nr_running;
4471 if (sds->group_imb) {
4472 sds->busiest_load_per_task =
4473 min(sds->busiest_load_per_task, sds->avg_load);
4477 * In the presence of smp nice balancing, certain scenarios can have
4478 * max load less than avg load(as we skip the groups at or below
4479 * its cpu_power, while calculating max_load..)
4481 if (sds->max_load < sds->avg_load) {
4483 return fix_small_imbalance(env, sds);
4486 if (!sds->group_imb) {
4488 * Don't want to pull so many tasks that a group would go idle.
4490 load_above_capacity = (sds->busiest_nr_running -
4491 sds->busiest_group_capacity);
4493 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4495 load_above_capacity /= sds->busiest->sgp->power;
4499 * We're trying to get all the cpus to the average_load, so we don't
4500 * want to push ourselves above the average load, nor do we wish to
4501 * reduce the max loaded cpu below the average load. At the same time,
4502 * we also don't want to reduce the group load below the group capacity
4503 * (so that we can implement power-savings policies etc). Thus we look
4504 * for the minimum possible imbalance.
4505 * Be careful of negative numbers as they'll appear as very large values
4506 * with unsigned longs.
4508 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4510 /* How much load to actually move to equalise the imbalance */
4511 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4512 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4513 / SCHED_POWER_SCALE;
4516 * if *imbalance is less than the average load per runnable task
4517 * there is no guarantee that any tasks will be moved so we'll have
4518 * a think about bumping its value to force at least one task to be
4521 if (env->imbalance < sds->busiest_load_per_task)
4522 return fix_small_imbalance(env, sds);
4526 /******* find_busiest_group() helpers end here *********************/
4529 * find_busiest_group - Returns the busiest group within the sched_domain
4530 * if there is an imbalance. If there isn't an imbalance, and
4531 * the user has opted for power-savings, it returns a group whose
4532 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4533 * such a group exists.
4535 * Also calculates the amount of weighted load which should be moved
4536 * to restore balance.
4538 * @env: The load balancing environment.
4539 * @balance: Pointer to a variable indicating if this_cpu
4540 * is the appropriate cpu to perform load balancing at this_level.
4542 * Returns: - the busiest group if imbalance exists.
4543 * - If no imbalance and user has opted for power-savings balance,
4544 * return the least loaded group whose CPUs can be
4545 * put to idle by rebalancing its tasks onto our group.
4547 static struct sched_group *
4548 find_busiest_group(struct lb_env *env, int *balance)
4550 struct sd_lb_stats sds;
4552 memset(&sds, 0, sizeof(sds));
4555 * Compute the various statistics relavent for load balancing at
4558 update_sd_lb_stats(env, balance, &sds);
4561 * this_cpu is not the appropriate cpu to perform load balancing at
4567 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4568 check_asym_packing(env, &sds))
4571 /* There is no busy sibling group to pull tasks from */
4572 if (!sds.busiest || sds.busiest_nr_running == 0)
4575 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4578 * If the busiest group is imbalanced the below checks don't
4579 * work because they assumes all things are equal, which typically
4580 * isn't true due to cpus_allowed constraints and the like.
4585 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4586 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4587 !sds.busiest_has_capacity)
4591 * If the local group is more busy than the selected busiest group
4592 * don't try and pull any tasks.
4594 if (sds.this_load >= sds.max_load)
4598 * Don't pull any tasks if this group is already above the domain
4601 if (sds.this_load >= sds.avg_load)
4604 if (env->idle == CPU_IDLE) {
4606 * This cpu is idle. If the busiest group load doesn't
4607 * have more tasks than the number of available cpu's and
4608 * there is no imbalance between this and busiest group
4609 * wrt to idle cpu's, it is balanced.
4611 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4612 sds.busiest_nr_running <= sds.busiest_group_weight)
4616 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4617 * imbalance_pct to be conservative.
4619 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4624 /* Looks like there is an imbalance. Compute it */
4625 calculate_imbalance(env, &sds);
4629 if (check_numa_busiest_group(env, &sds))
4638 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4640 static struct rq *find_busiest_queue(struct lb_env *env,
4641 struct sched_group *group)
4643 struct rq *busiest = NULL, *rq;
4644 unsigned long max_load = 0;
4647 for_each_cpu(i, sched_group_cpus(group)) {
4648 unsigned long power = power_of(i);
4649 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4654 capacity = fix_small_capacity(env->sd, group);
4656 if (!cpumask_test_cpu(i, env->cpus))
4660 wl = weighted_cpuload(i);
4663 * When comparing with imbalance, use weighted_cpuload()
4664 * which is not scaled with the cpu power.
4666 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4670 * For the load comparisons with the other cpu's, consider
4671 * the weighted_cpuload() scaled with the cpu power, so that
4672 * the load can be moved away from the cpu that is potentially
4673 * running at a lower capacity.
4675 wl = (wl * SCHED_POWER_SCALE) / power;
4677 if (wl > max_load) {
4687 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4688 * so long as it is large enough.
4690 #define MAX_PINNED_INTERVAL 512
4692 /* Working cpumask for load_balance and load_balance_newidle. */
4693 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4695 static int need_active_balance(struct lb_env *env)
4697 struct sched_domain *sd = env->sd;
4699 if (env->idle == CPU_NEWLY_IDLE) {
4702 * ASYM_PACKING needs to force migrate tasks from busy but
4703 * higher numbered CPUs in order to pack all tasks in the
4704 * lowest numbered CPUs.
4706 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4710 if (need_active_numa_balance(env))
4713 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4716 static int active_load_balance_cpu_stop(void *data);
4719 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4720 * tasks if there is an imbalance.
4722 static int load_balance(int this_cpu, struct rq *this_rq,
4723 struct sched_domain *sd, enum cpu_idle_type idle,
4726 int ld_moved, cur_ld_moved, active_balance = 0;
4727 int lb_iterations, max_lb_iterations;
4728 struct sched_group *group;
4730 unsigned long flags;
4731 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4733 struct lb_env env = {
4735 .dst_cpu = this_cpu,
4737 .dst_grpmask = sched_group_cpus(sd->groups),
4739 .loop_break = sched_nr_migrate_break,
4741 .find_busiest_queue = find_busiest_queue,
4744 cpumask_copy(cpus, cpu_active_mask);
4745 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4747 schedstat_inc(sd, lb_count[idle]);
4750 group = find_busiest_group(&env, balance);
4756 schedstat_inc(sd, lb_nobusyg[idle]);
4760 busiest = env.find_busiest_queue(&env, group);
4762 schedstat_inc(sd, lb_nobusyq[idle]);
4765 env.src_rq = busiest;
4766 env.src_cpu = busiest->cpu;
4768 BUG_ON(busiest == env.dst_rq);
4770 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4774 if (busiest->nr_running > 1) {
4776 * Attempt to move tasks. If find_busiest_group has found
4777 * an imbalance but busiest->nr_running <= 1, the group is
4778 * still unbalanced. ld_moved simply stays zero, so it is
4779 * correctly treated as an imbalance.
4781 env.flags |= LBF_ALL_PINNED;
4782 env.src_cpu = busiest->cpu;
4783 env.src_rq = busiest;
4784 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4785 if (sched_feat_numa(NUMA_PULL))
4786 env.tasks = offnode_tasks(busiest);
4788 env.tasks = &busiest->cfs_tasks;
4790 update_h_load(env.src_cpu);
4792 local_irq_save(flags);
4793 double_rq_lock(env.dst_rq, busiest);
4796 * cur_ld_moved - load moved in current iteration
4797 * ld_moved - cumulative load moved across iterations
4799 cur_ld_moved = move_tasks(&env);
4800 ld_moved += cur_ld_moved;
4801 double_rq_unlock(env.dst_rq, busiest);
4802 local_irq_restore(flags);
4804 if (env.flags & LBF_NEED_BREAK) {
4805 env.flags &= ~LBF_NEED_BREAK;
4810 * some other cpu did the load balance for us.
4812 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4813 resched_cpu(env.dst_cpu);
4816 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4817 * us and move them to an alternate dst_cpu in our sched_group
4818 * where they can run. The upper limit on how many times we
4819 * iterate on same src_cpu is dependent on number of cpus in our
4822 * This changes load balance semantics a bit on who can move
4823 * load to a given_cpu. In addition to the given_cpu itself
4824 * (or a ilb_cpu acting on its behalf where given_cpu is
4825 * nohz-idle), we now have balance_cpu in a position to move
4826 * load to given_cpu. In rare situations, this may cause
4827 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4828 * _independently_ and at _same_ time to move some load to
4829 * given_cpu) causing exceess load to be moved to given_cpu.
4830 * This however should not happen so much in practice and
4831 * moreover subsequent load balance cycles should correct the
4832 * excess load moved.
4834 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4835 lb_iterations++ < max_lb_iterations) {
4837 env.dst_rq = cpu_rq(env.new_dst_cpu);
4838 env.dst_cpu = env.new_dst_cpu;
4839 env.flags &= ~LBF_SOME_PINNED;
4841 env.loop_break = sched_nr_migrate_break;
4843 * Go back to "more_balance" rather than "redo" since we
4844 * need to continue with same src_cpu.
4849 /* All tasks on this runqueue were pinned by CPU affinity */
4850 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4851 cpumask_clear_cpu(cpu_of(busiest), cpus);
4852 if (!cpumask_empty(cpus)) {
4854 env.loop_break = sched_nr_migrate_break;
4862 schedstat_inc(sd, lb_failed[idle]);
4864 * Increment the failure counter only on periodic balance.
4865 * We do not want newidle balance, which can be very
4866 * frequent, pollute the failure counter causing
4867 * excessive cache_hot migrations and active balances.
4869 if (idle != CPU_NEWLY_IDLE)
4870 sd->nr_balance_failed++;
4872 if (need_active_balance(&env)) {
4873 raw_spin_lock_irqsave(&busiest->lock, flags);
4875 /* don't kick the active_load_balance_cpu_stop,
4876 * if the curr task on busiest cpu can't be
4879 if (!cpumask_test_cpu(this_cpu,
4880 tsk_cpus_allowed(busiest->curr))) {
4881 raw_spin_unlock_irqrestore(&busiest->lock,
4883 env.flags |= LBF_ALL_PINNED;
4884 goto out_one_pinned;
4888 * ->active_balance synchronizes accesses to
4889 * ->active_balance_work. Once set, it's cleared
4890 * only after active load balance is finished.
4892 if (!busiest->active_balance) {
4893 busiest->active_balance = 1;
4894 busiest->push_cpu = this_cpu;
4897 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4899 if (active_balance) {
4900 stop_one_cpu_nowait(cpu_of(busiest),
4901 active_load_balance_cpu_stop, busiest,
4902 &busiest->active_balance_work);
4906 * We've kicked active balancing, reset the failure
4909 sd->nr_balance_failed = sd->cache_nice_tries+1;
4912 sd->nr_balance_failed = 0;
4914 if (likely(!active_balance)) {
4915 /* We were unbalanced, so reset the balancing interval */
4916 sd->balance_interval = sd->min_interval;
4919 * If we've begun active balancing, start to back off. This
4920 * case may not be covered by the all_pinned logic if there
4921 * is only 1 task on the busy runqueue (because we don't call
4924 if (sd->balance_interval < sd->max_interval)
4925 sd->balance_interval *= 2;
4931 schedstat_inc(sd, lb_balanced[idle]);
4933 sd->nr_balance_failed = 0;
4936 /* tune up the balancing interval */
4937 if (((env.flags & LBF_ALL_PINNED) &&
4938 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4939 (sd->balance_interval < sd->max_interval))
4940 sd->balance_interval *= 2;
4948 * idle_balance is called by schedule() if this_cpu is about to become
4949 * idle. Attempts to pull tasks from other CPUs.
4951 void idle_balance(int this_cpu, struct rq *this_rq)
4953 struct sched_domain *sd;
4954 int pulled_task = 0;
4955 unsigned long next_balance = jiffies + HZ;
4957 this_rq->idle_stamp = this_rq->clock;
4959 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4963 * Drop the rq->lock, but keep IRQ/preempt disabled.
4965 raw_spin_unlock(&this_rq->lock);
4967 update_shares(this_cpu);
4969 for_each_domain(this_cpu, sd) {
4970 unsigned long interval;
4973 if (!(sd->flags & SD_LOAD_BALANCE))
4976 if (sd->flags & SD_BALANCE_NEWIDLE) {
4977 /* If we've pulled tasks over stop searching: */
4978 pulled_task = load_balance(this_cpu, this_rq,
4979 sd, CPU_NEWLY_IDLE, &balance);
4982 interval = msecs_to_jiffies(sd->balance_interval);
4983 if (time_after(next_balance, sd->last_balance + interval))
4984 next_balance = sd->last_balance + interval;
4986 this_rq->idle_stamp = 0;
4992 raw_spin_lock(&this_rq->lock);
4994 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4996 * We are going idle. next_balance may be set based on
4997 * a busy processor. So reset next_balance.
4999 this_rq->next_balance = next_balance;
5004 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5005 * running tasks off the busiest CPU onto idle CPUs. It requires at
5006 * least 1 task to be running on each physical CPU where possible, and
5007 * avoids physical / logical imbalances.
5009 static int active_load_balance_cpu_stop(void *data)
5011 struct rq *busiest_rq = data;
5012 int busiest_cpu = cpu_of(busiest_rq);
5013 int target_cpu = busiest_rq->push_cpu;
5014 struct rq *target_rq = cpu_rq(target_cpu);
5015 struct sched_domain *sd;
5017 raw_spin_lock_irq(&busiest_rq->lock);
5019 /* make sure the requested cpu hasn't gone down in the meantime */
5020 if (unlikely(busiest_cpu != smp_processor_id() ||
5021 !busiest_rq->active_balance))
5024 /* Is there any task to move? */
5025 if (busiest_rq->nr_running <= 1)
5029 * This condition is "impossible", if it occurs
5030 * we need to fix it. Originally reported by
5031 * Bjorn Helgaas on a 128-cpu setup.
5033 BUG_ON(busiest_rq == target_rq);
5035 /* move a task from busiest_rq to target_rq */
5036 double_lock_balance(busiest_rq, target_rq);
5038 /* Search for an sd spanning us and the target CPU. */
5040 for_each_domain(target_cpu, sd) {
5041 if ((sd->flags & SD_LOAD_BALANCE) &&
5042 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5047 struct lb_env env = {
5049 .dst_cpu = target_cpu,
5050 .dst_rq = target_rq,
5051 .src_cpu = busiest_rq->cpu,
5052 .src_rq = busiest_rq,
5056 schedstat_inc(sd, alb_count);
5058 if (move_one_task(&env))
5059 schedstat_inc(sd, alb_pushed);
5061 schedstat_inc(sd, alb_failed);
5064 double_unlock_balance(busiest_rq, target_rq);
5066 busiest_rq->active_balance = 0;
5067 raw_spin_unlock_irq(&busiest_rq->lock);
5073 * idle load balancing details
5074 * - When one of the busy CPUs notice that there may be an idle rebalancing
5075 * needed, they will kick the idle load balancer, which then does idle
5076 * load balancing for all the idle CPUs.
5079 cpumask_var_t idle_cpus_mask;
5081 unsigned long next_balance; /* in jiffy units */
5082 } nohz ____cacheline_aligned;
5084 static inline int find_new_ilb(int call_cpu)
5086 int ilb = cpumask_first(nohz.idle_cpus_mask);
5088 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5095 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5096 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5097 * CPU (if there is one).
5099 static void nohz_balancer_kick(int cpu)
5103 nohz.next_balance++;
5105 ilb_cpu = find_new_ilb(cpu);
5107 if (ilb_cpu >= nr_cpu_ids)
5110 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5113 * Use smp_send_reschedule() instead of resched_cpu().
5114 * This way we generate a sched IPI on the target cpu which
5115 * is idle. And the softirq performing nohz idle load balance
5116 * will be run before returning from the IPI.
5118 smp_send_reschedule(ilb_cpu);
5122 static inline void nohz_balance_exit_idle(int cpu)
5124 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5125 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5126 atomic_dec(&nohz.nr_cpus);
5127 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5131 static inline void set_cpu_sd_state_busy(void)
5133 struct sched_domain *sd;
5134 int cpu = smp_processor_id();
5136 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5138 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5141 for_each_domain(cpu, sd)
5142 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5146 void set_cpu_sd_state_idle(void)
5148 struct sched_domain *sd;
5149 int cpu = smp_processor_id();
5151 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5153 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5156 for_each_domain(cpu, sd)
5157 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5162 * This routine will record that the cpu is going idle with tick stopped.
5163 * This info will be used in performing idle load balancing in the future.
5165 void nohz_balance_enter_idle(int cpu)
5168 * If this cpu is going down, then nothing needs to be done.
5170 if (!cpu_active(cpu))
5173 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5176 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5177 atomic_inc(&nohz.nr_cpus);
5178 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5181 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5182 unsigned long action, void *hcpu)
5184 switch (action & ~CPU_TASKS_FROZEN) {
5186 nohz_balance_exit_idle(smp_processor_id());
5194 static DEFINE_SPINLOCK(balancing);
5197 * Scale the max load_balance interval with the number of CPUs in the system.
5198 * This trades load-balance latency on larger machines for less cross talk.
5200 void update_max_interval(void)
5202 max_load_balance_interval = HZ*num_online_cpus()/10;
5206 * It checks each scheduling domain to see if it is due to be balanced,
5207 * and initiates a balancing operation if so.
5209 * Balancing parameters are set up in arch_init_sched_domains.
5211 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5214 struct rq *rq = cpu_rq(cpu);
5215 unsigned long interval;
5216 struct sched_domain *sd;
5217 /* Earliest time when we have to do rebalance again */
5218 unsigned long next_balance = jiffies + 60*HZ;
5219 int update_next_balance = 0;
5225 for_each_domain(cpu, sd) {
5226 if (!(sd->flags & SD_LOAD_BALANCE))
5229 interval = sd->balance_interval;
5230 if (idle != CPU_IDLE)
5231 interval *= sd->busy_factor;
5233 /* scale ms to jiffies */
5234 interval = msecs_to_jiffies(interval);
5235 interval = clamp(interval, 1UL, max_load_balance_interval);
5237 need_serialize = sd->flags & SD_SERIALIZE;
5239 if (need_serialize) {
5240 if (!spin_trylock(&balancing))
5244 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5245 if (load_balance(cpu, rq, sd, idle, &balance)) {
5247 * We've pulled tasks over so either we're no
5250 idle = CPU_NOT_IDLE;
5252 sd->last_balance = jiffies;
5255 spin_unlock(&balancing);
5257 if (time_after(next_balance, sd->last_balance + interval)) {
5258 next_balance = sd->last_balance + interval;
5259 update_next_balance = 1;
5263 * Stop the load balance at this level. There is another
5264 * CPU in our sched group which is doing load balancing more
5273 * next_balance will be updated only when there is a need.
5274 * When the cpu is attached to null domain for ex, it will not be
5277 if (likely(update_next_balance))
5278 rq->next_balance = next_balance;
5283 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5284 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5286 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5288 struct rq *this_rq = cpu_rq(this_cpu);
5292 if (idle != CPU_IDLE ||
5293 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5296 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5297 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5301 * If this cpu gets work to do, stop the load balancing
5302 * work being done for other cpus. Next load
5303 * balancing owner will pick it up.
5308 rq = cpu_rq(balance_cpu);
5310 raw_spin_lock_irq(&rq->lock);
5311 update_rq_clock(rq);
5312 update_idle_cpu_load(rq);
5313 raw_spin_unlock_irq(&rq->lock);
5315 rebalance_domains(balance_cpu, CPU_IDLE);
5317 if (time_after(this_rq->next_balance, rq->next_balance))
5318 this_rq->next_balance = rq->next_balance;
5320 nohz.next_balance = this_rq->next_balance;
5322 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5326 * Current heuristic for kicking the idle load balancer in the presence
5327 * of an idle cpu is the system.
5328 * - This rq has more than one task.
5329 * - At any scheduler domain level, this cpu's scheduler group has multiple
5330 * busy cpu's exceeding the group's power.
5331 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5332 * domain span are idle.
5334 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5336 unsigned long now = jiffies;
5337 struct sched_domain *sd;
5339 if (unlikely(idle_cpu(cpu)))
5343 * We may be recently in ticked or tickless idle mode. At the first
5344 * busy tick after returning from idle, we will update the busy stats.
5346 set_cpu_sd_state_busy();
5347 nohz_balance_exit_idle(cpu);
5350 * None are in tickless mode and hence no need for NOHZ idle load
5353 if (likely(!atomic_read(&nohz.nr_cpus)))
5356 if (time_before(now, nohz.next_balance))
5359 if (rq->nr_running >= 2)
5363 for_each_domain(cpu, sd) {
5364 struct sched_group *sg = sd->groups;
5365 struct sched_group_power *sgp = sg->sgp;
5366 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5368 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5369 goto need_kick_unlock;
5371 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5372 && (cpumask_first_and(nohz.idle_cpus_mask,
5373 sched_domain_span(sd)) < cpu))
5374 goto need_kick_unlock;
5376 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5388 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5392 * run_rebalance_domains is triggered when needed from the scheduler tick.
5393 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5395 static void run_rebalance_domains(struct softirq_action *h)
5397 int this_cpu = smp_processor_id();
5398 struct rq *this_rq = cpu_rq(this_cpu);
5399 enum cpu_idle_type idle = this_rq->idle_balance ?
5400 CPU_IDLE : CPU_NOT_IDLE;
5402 rebalance_domains(this_cpu, idle);
5405 * If this cpu has a pending nohz_balance_kick, then do the
5406 * balancing on behalf of the other idle cpus whose ticks are
5409 nohz_idle_balance(this_cpu, idle);
5412 static inline int on_null_domain(int cpu)
5414 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5418 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5420 void trigger_load_balance(struct rq *rq, int cpu)
5422 /* Don't need to rebalance while attached to NULL domain */
5423 if (time_after_eq(jiffies, rq->next_balance) &&
5424 likely(!on_null_domain(cpu)))
5425 raise_softirq(SCHED_SOFTIRQ);
5427 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5428 nohz_balancer_kick(cpu);
5432 static void rq_online_fair(struct rq *rq)
5437 static void rq_offline_fair(struct rq *rq)
5441 /* Ensure any throttled groups are reachable by pick_next_task */
5442 unthrottle_offline_cfs_rqs(rq);
5445 #endif /* CONFIG_SMP */
5448 * scheduler tick hitting a task of our scheduling class:
5450 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5452 struct cfs_rq *cfs_rq;
5453 struct sched_entity *se = &curr->se;
5455 for_each_sched_entity(se) {
5456 cfs_rq = cfs_rq_of(se);
5457 entity_tick(cfs_rq, se, queued);
5460 if (sched_feat_numa(NUMA))
5461 task_tick_numa(rq, curr);
5465 * called on fork with the child task as argument from the parent's context
5466 * - child not yet on the tasklist
5467 * - preemption disabled
5469 static void task_fork_fair(struct task_struct *p)
5471 struct cfs_rq *cfs_rq;
5472 struct sched_entity *se = &p->se, *curr;
5473 int this_cpu = smp_processor_id();
5474 struct rq *rq = this_rq();
5475 unsigned long flags;
5477 raw_spin_lock_irqsave(&rq->lock, flags);
5479 update_rq_clock(rq);
5481 cfs_rq = task_cfs_rq(current);
5482 curr = cfs_rq->curr;
5484 if (unlikely(task_cpu(p) != this_cpu)) {
5486 __set_task_cpu(p, this_cpu);
5490 update_curr(cfs_rq);
5493 se->vruntime = curr->vruntime;
5494 place_entity(cfs_rq, se, 1);
5496 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5498 * Upon rescheduling, sched_class::put_prev_task() will place
5499 * 'current' within the tree based on its new key value.
5501 swap(curr->vruntime, se->vruntime);
5502 resched_task(rq->curr);
5505 se->vruntime -= cfs_rq->min_vruntime;
5507 raw_spin_unlock_irqrestore(&rq->lock, flags);
5511 * Priority of the task has changed. Check to see if we preempt
5515 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5521 * Reschedule if we are currently running on this runqueue and
5522 * our priority decreased, or if we are not currently running on
5523 * this runqueue and our priority is higher than the current's
5525 if (rq->curr == p) {
5526 if (p->prio > oldprio)
5527 resched_task(rq->curr);
5529 check_preempt_curr(rq, p, 0);
5532 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5534 struct sched_entity *se = &p->se;
5535 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5538 * Ensure the task's vruntime is normalized, so that when its
5539 * switched back to the fair class the enqueue_entity(.flags=0) will
5540 * do the right thing.
5542 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5543 * have normalized the vruntime, if it was !on_rq, then only when
5544 * the task is sleeping will it still have non-normalized vruntime.
5546 if (!se->on_rq && p->state != TASK_RUNNING) {
5548 * Fix up our vruntime so that the current sleep doesn't
5549 * cause 'unlimited' sleep bonus.
5551 place_entity(cfs_rq, se, 0);
5552 se->vruntime -= cfs_rq->min_vruntime;
5557 * We switched to the sched_fair class.
5559 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5565 * We were most likely switched from sched_rt, so
5566 * kick off the schedule if running, otherwise just see
5567 * if we can still preempt the current task.
5570 resched_task(rq->curr);
5572 check_preempt_curr(rq, p, 0);
5575 /* Account for a task changing its policy or group.
5577 * This routine is mostly called to set cfs_rq->curr field when a task
5578 * migrates between groups/classes.
5580 static void set_curr_task_fair(struct rq *rq)
5582 struct sched_entity *se = &rq->curr->se;
5584 for_each_sched_entity(se) {
5585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5587 set_next_entity(cfs_rq, se);
5588 /* ensure bandwidth has been allocated on our new cfs_rq */
5589 account_cfs_rq_runtime(cfs_rq, 0);
5593 void init_cfs_rq(struct cfs_rq *cfs_rq)
5595 cfs_rq->tasks_timeline = RB_ROOT;
5596 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5597 #ifndef CONFIG_64BIT
5598 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5602 #ifdef CONFIG_FAIR_GROUP_SCHED
5603 static void task_move_group_fair(struct task_struct *p, int on_rq)
5606 * If the task was not on the rq at the time of this cgroup movement
5607 * it must have been asleep, sleeping tasks keep their ->vruntime
5608 * absolute on their old rq until wakeup (needed for the fair sleeper
5609 * bonus in place_entity()).
5611 * If it was on the rq, we've just 'preempted' it, which does convert
5612 * ->vruntime to a relative base.
5614 * Make sure both cases convert their relative position when migrating
5615 * to another cgroup's rq. This does somewhat interfere with the
5616 * fair sleeper stuff for the first placement, but who cares.
5619 * When !on_rq, vruntime of the task has usually NOT been normalized.
5620 * But there are some cases where it has already been normalized:
5622 * - Moving a forked child which is waiting for being woken up by
5623 * wake_up_new_task().
5624 * - Moving a task which has been woken up by try_to_wake_up() and
5625 * waiting for actually being woken up by sched_ttwu_pending().
5627 * To prevent boost or penalty in the new cfs_rq caused by delta
5628 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5630 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5634 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5635 set_task_rq(p, task_cpu(p));
5637 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5640 void free_fair_sched_group(struct task_group *tg)
5644 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5646 for_each_possible_cpu(i) {
5648 kfree(tg->cfs_rq[i]);
5657 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5659 struct cfs_rq *cfs_rq;
5660 struct sched_entity *se;
5663 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5666 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5670 tg->shares = NICE_0_LOAD;
5672 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5674 for_each_possible_cpu(i) {
5675 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5676 GFP_KERNEL, cpu_to_node(i));
5680 se = kzalloc_node(sizeof(struct sched_entity),
5681 GFP_KERNEL, cpu_to_node(i));
5685 init_cfs_rq(cfs_rq);
5686 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5697 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5699 struct rq *rq = cpu_rq(cpu);
5700 unsigned long flags;
5703 * Only empty task groups can be destroyed; so we can speculatively
5704 * check on_list without danger of it being re-added.
5706 if (!tg->cfs_rq[cpu]->on_list)
5709 raw_spin_lock_irqsave(&rq->lock, flags);
5710 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5711 raw_spin_unlock_irqrestore(&rq->lock, flags);
5714 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5715 struct sched_entity *se, int cpu,
5716 struct sched_entity *parent)
5718 struct rq *rq = cpu_rq(cpu);
5723 /* allow initial update_cfs_load() to truncate */
5724 cfs_rq->load_stamp = 1;
5726 init_cfs_rq_runtime(cfs_rq);
5728 tg->cfs_rq[cpu] = cfs_rq;
5731 /* se could be NULL for root_task_group */
5736 se->cfs_rq = &rq->cfs;
5738 se->cfs_rq = parent->my_q;
5741 update_load_set(&se->load, 0);
5742 se->parent = parent;
5745 static DEFINE_MUTEX(shares_mutex);
5747 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5750 unsigned long flags;
5753 * We can't change the weight of the root cgroup.
5758 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5760 mutex_lock(&shares_mutex);
5761 if (tg->shares == shares)
5764 tg->shares = shares;
5765 for_each_possible_cpu(i) {
5766 struct rq *rq = cpu_rq(i);
5767 struct sched_entity *se;
5770 /* Propagate contribution to hierarchy */
5771 raw_spin_lock_irqsave(&rq->lock, flags);
5772 for_each_sched_entity(se)
5773 update_cfs_shares(group_cfs_rq(se));
5774 raw_spin_unlock_irqrestore(&rq->lock, flags);
5778 mutex_unlock(&shares_mutex);
5781 #else /* CONFIG_FAIR_GROUP_SCHED */
5783 void free_fair_sched_group(struct task_group *tg) { }
5785 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5790 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5792 #endif /* CONFIG_FAIR_GROUP_SCHED */
5795 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5797 struct sched_entity *se = &task->se;
5798 unsigned int rr_interval = 0;
5801 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5804 if (rq->cfs.load.weight)
5805 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5811 * All the scheduling class methods:
5813 const struct sched_class fair_sched_class = {
5814 .next = &idle_sched_class,
5815 .enqueue_task = enqueue_task_fair,
5816 .dequeue_task = dequeue_task_fair,
5817 .yield_task = yield_task_fair,
5818 .yield_to_task = yield_to_task_fair,
5820 .check_preempt_curr = check_preempt_wakeup,
5822 .pick_next_task = pick_next_task_fair,
5823 .put_prev_task = put_prev_task_fair,
5826 .select_task_rq = select_task_rq_fair,
5828 .rq_online = rq_online_fair,
5829 .rq_offline = rq_offline_fair,
5831 .task_waking = task_waking_fair,
5834 .set_curr_task = set_curr_task_fair,
5835 .task_tick = task_tick_fair,
5836 .task_fork = task_fork_fair,
5838 .prio_changed = prio_changed_fair,
5839 .switched_from = switched_from_fair,
5840 .switched_to = switched_to_fair,
5842 .get_rr_interval = get_rr_interval_fair,
5844 #ifdef CONFIG_FAIR_GROUP_SCHED
5845 .task_move_group = task_move_group_fair,
5849 #ifdef CONFIG_SCHED_DEBUG
5850 void print_cfs_stats(struct seq_file *m, int cpu)
5852 struct cfs_rq *cfs_rq;
5855 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5856 print_cfs_rq(m, cpu, cfs_rq);
5861 __init void init_sched_fair_class(void)
5864 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5867 nohz.next_balance = jiffies;
5868 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5869 cpu_notifier(sched_ilb_notifier, 0);