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/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
34 #include <linux/sysfs.h>
35 #include <linux/vmalloc.h>
36 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
37 /* Include cpufreq header to add a notifier so that cpu frequency
38 * scaling can track the current CPU frequency
40 #include <linux/cpufreq.h>
41 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
42 #ifdef CONFIG_SCHED_HMP
43 #include <linux/cpuidle.h>
50 * Targeted preemption latency for CPU-bound tasks:
51 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
53 * NOTE: this latency value is not the same as the concept of
54 * 'timeslice length' - timeslices in CFS are of variable length
55 * and have no persistent notion like in traditional, time-slice
56 * based scheduling concepts.
58 * (to see the precise effective timeslice length of your workload,
59 * run vmstat and monitor the context-switches (cs) field)
61 unsigned int sysctl_sched_latency = 6000000ULL;
62 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
65 * The initial- and re-scaling of tunables is configurable
66 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
69 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
70 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
71 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
73 enum sched_tunable_scaling sysctl_sched_tunable_scaling
74 = SCHED_TUNABLESCALING_LOG;
77 * Minimal preemption granularity for CPU-bound tasks:
78 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
80 unsigned int sysctl_sched_min_granularity = 750000ULL;
81 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
84 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
86 static unsigned int sched_nr_latency = 8;
89 * After fork, child runs first. If set to 0 (default) then
90 * parent will (try to) run first.
92 unsigned int sysctl_sched_child_runs_first __read_mostly;
95 * SCHED_OTHER wake-up granularity.
96 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
98 * This option delays the preemption effects of decoupled workloads
99 * and reduces their over-scheduling. Synchronous workloads will still
100 * have immediate wakeup/sleep latencies.
102 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
103 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
105 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
108 * The exponential sliding window over which load is averaged for shares
112 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
114 #ifdef CONFIG_CFS_BANDWIDTH
116 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
117 * each time a cfs_rq requests quota.
119 * Note: in the case that the slice exceeds the runtime remaining (either due
120 * to consumption or the quota being specified to be smaller than the slice)
121 * we will always only issue the remaining available time.
123 * default: 5 msec, units: microseconds
125 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
129 * Increase the granularity value when there are more CPUs,
130 * because with more CPUs the 'effective latency' as visible
131 * to users decreases. But the relationship is not linear,
132 * so pick a second-best guess by going with the log2 of the
135 * This idea comes from the SD scheduler of Con Kolivas:
137 static int get_update_sysctl_factor(void)
139 unsigned int cpus = min_t(int, num_online_cpus(), 8);
142 switch (sysctl_sched_tunable_scaling) {
143 case SCHED_TUNABLESCALING_NONE:
146 case SCHED_TUNABLESCALING_LINEAR:
149 case SCHED_TUNABLESCALING_LOG:
151 factor = 1 + ilog2(cpus);
158 static void update_sysctl(void)
160 unsigned int factor = get_update_sysctl_factor();
162 #define SET_SYSCTL(name) \
163 (sysctl_##name = (factor) * normalized_sysctl_##name)
164 SET_SYSCTL(sched_min_granularity);
165 SET_SYSCTL(sched_latency);
166 SET_SYSCTL(sched_wakeup_granularity);
170 void sched_init_granularity(void)
175 #if BITS_PER_LONG == 32
176 # define WMULT_CONST (~0UL)
178 # define WMULT_CONST (1UL << 32)
181 #define WMULT_SHIFT 32
184 * Shift right and round:
186 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
189 * delta *= weight / lw
192 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
193 struct load_weight *lw)
198 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
199 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
200 * 2^SCHED_LOAD_RESOLUTION.
202 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
203 tmp = (u64)delta_exec * scale_load_down(weight);
205 tmp = (u64)delta_exec;
207 if (!lw->inv_weight) {
208 unsigned long w = scale_load_down(lw->weight);
210 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
212 else if (unlikely(!w))
213 lw->inv_weight = WMULT_CONST;
215 lw->inv_weight = WMULT_CONST / w;
219 * Check whether we'd overflow the 64-bit multiplication:
221 if (unlikely(tmp > WMULT_CONST))
222 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
225 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
227 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
231 const struct sched_class fair_sched_class;
233 /**************************************************************
234 * CFS operations on generic schedulable entities:
237 #ifdef CONFIG_FAIR_GROUP_SCHED
239 /* cpu runqueue to which this cfs_rq is attached */
240 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
245 /* An entity is a task if it doesn't "own" a runqueue */
246 #define entity_is_task(se) (!se->my_q)
248 static inline struct task_struct *task_of(struct sched_entity *se)
250 #ifdef CONFIG_SCHED_DEBUG
251 WARN_ON_ONCE(!entity_is_task(se));
253 return container_of(se, struct task_struct, se);
256 /* Walk up scheduling entities hierarchy */
257 #define for_each_sched_entity(se) \
258 for (; se; se = se->parent)
260 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
265 /* runqueue on which this entity is (to be) queued */
266 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
271 /* runqueue "owned" by this group */
272 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
277 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
280 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
282 if (!cfs_rq->on_list) {
284 * Ensure we either appear before our parent (if already
285 * enqueued) or force our parent to appear after us when it is
286 * enqueued. The fact that we always enqueue bottom-up
287 * reduces this to two cases.
289 if (cfs_rq->tg->parent &&
290 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
291 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
292 &rq_of(cfs_rq)->leaf_cfs_rq_list);
294 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
295 &rq_of(cfs_rq)->leaf_cfs_rq_list);
299 /* We should have no load, but we need to update last_decay. */
300 update_cfs_rq_blocked_load(cfs_rq, 0);
304 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
306 if (cfs_rq->on_list) {
307 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
312 /* Iterate thr' all leaf cfs_rq's on a runqueue */
313 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
314 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
316 /* Do the two (enqueued) entities belong to the same group ? */
318 is_same_group(struct sched_entity *se, struct sched_entity *pse)
320 if (se->cfs_rq == pse->cfs_rq)
326 static inline struct sched_entity *parent_entity(struct sched_entity *se)
331 /* return depth at which a sched entity is present in the hierarchy */
332 static inline int depth_se(struct sched_entity *se)
336 for_each_sched_entity(se)
343 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
345 int se_depth, pse_depth;
348 * preemption test can be made between sibling entities who are in the
349 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
350 * both tasks until we find their ancestors who are siblings of common
354 /* First walk up until both entities are at same depth */
355 se_depth = depth_se(*se);
356 pse_depth = depth_se(*pse);
358 while (se_depth > pse_depth) {
360 *se = parent_entity(*se);
363 while (pse_depth > se_depth) {
365 *pse = parent_entity(*pse);
368 while (!is_same_group(*se, *pse)) {
369 *se = parent_entity(*se);
370 *pse = parent_entity(*pse);
374 #else /* !CONFIG_FAIR_GROUP_SCHED */
376 static inline struct task_struct *task_of(struct sched_entity *se)
378 return container_of(se, struct task_struct, se);
381 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
383 return container_of(cfs_rq, struct rq, cfs);
386 #define entity_is_task(se) 1
388 #define for_each_sched_entity(se) \
389 for (; se; se = NULL)
391 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
393 return &task_rq(p)->cfs;
396 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
398 struct task_struct *p = task_of(se);
399 struct rq *rq = task_rq(p);
404 /* runqueue "owned" by this group */
405 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
410 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
414 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
418 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
419 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
422 is_same_group(struct sched_entity *se, struct sched_entity *pse)
427 static inline struct sched_entity *parent_entity(struct sched_entity *se)
433 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
437 #endif /* CONFIG_FAIR_GROUP_SCHED */
439 static __always_inline
440 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
442 /**************************************************************
443 * Scheduling class tree data structure manipulation methods:
446 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
448 s64 delta = (s64)(vruntime - max_vruntime);
450 max_vruntime = vruntime;
455 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
457 s64 delta = (s64)(vruntime - min_vruntime);
459 min_vruntime = vruntime;
464 static inline int entity_before(struct sched_entity *a,
465 struct sched_entity *b)
467 return (s64)(a->vruntime - b->vruntime) < 0;
470 static void update_min_vruntime(struct cfs_rq *cfs_rq)
472 u64 vruntime = cfs_rq->min_vruntime;
475 vruntime = cfs_rq->curr->vruntime;
477 if (cfs_rq->rb_leftmost) {
478 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
483 vruntime = se->vruntime;
485 vruntime = min_vruntime(vruntime, se->vruntime);
488 /* ensure we never gain time by being placed backwards. */
489 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
492 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
497 * Enqueue an entity into the rb-tree:
499 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
501 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
502 struct rb_node *parent = NULL;
503 struct sched_entity *entry;
507 * Find the right place in the rbtree:
511 entry = rb_entry(parent, struct sched_entity, run_node);
513 * We dont care about collisions. Nodes with
514 * the same key stay together.
516 if (entity_before(se, entry)) {
517 link = &parent->rb_left;
519 link = &parent->rb_right;
525 * Maintain a cache of leftmost tree entries (it is frequently
529 cfs_rq->rb_leftmost = &se->run_node;
531 rb_link_node(&se->run_node, parent, link);
532 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
535 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
537 if (cfs_rq->rb_leftmost == &se->run_node) {
538 struct rb_node *next_node;
540 next_node = rb_next(&se->run_node);
541 cfs_rq->rb_leftmost = next_node;
544 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
547 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
549 struct rb_node *left = cfs_rq->rb_leftmost;
554 return rb_entry(left, struct sched_entity, run_node);
557 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
559 struct rb_node *next = rb_next(&se->run_node);
564 return rb_entry(next, struct sched_entity, run_node);
567 #ifdef CONFIG_SCHED_DEBUG
568 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
570 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
575 return rb_entry(last, struct sched_entity, run_node);
578 /**************************************************************
579 * Scheduling class statistics methods:
582 int sched_proc_update_handler(struct ctl_table *table, int write,
583 void __user *buffer, size_t *lenp,
586 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
587 int factor = get_update_sysctl_factor();
592 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
593 sysctl_sched_min_granularity);
595 #define WRT_SYSCTL(name) \
596 (normalized_sysctl_##name = sysctl_##name / (factor))
597 WRT_SYSCTL(sched_min_granularity);
598 WRT_SYSCTL(sched_latency);
599 WRT_SYSCTL(sched_wakeup_granularity);
609 static inline unsigned long
610 calc_delta_fair(unsigned long delta, struct sched_entity *se)
612 if (unlikely(se->load.weight != NICE_0_LOAD))
613 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
619 * The idea is to set a period in which each task runs once.
621 * When there are too many tasks (sched_nr_latency) we have to stretch
622 * this period because otherwise the slices get too small.
624 * p = (nr <= nl) ? l : l*nr/nl
626 static u64 __sched_period(unsigned long nr_running)
628 u64 period = sysctl_sched_latency;
629 unsigned long nr_latency = sched_nr_latency;
631 if (unlikely(nr_running > nr_latency)) {
632 period = sysctl_sched_min_granularity;
633 period *= nr_running;
640 * We calculate the wall-time slice from the period by taking a part
641 * proportional to the weight.
645 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
647 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
649 for_each_sched_entity(se) {
650 struct load_weight *load;
651 struct load_weight lw;
653 cfs_rq = cfs_rq_of(se);
654 load = &cfs_rq->load;
656 if (unlikely(!se->on_rq)) {
659 update_load_add(&lw, se->load.weight);
662 slice = calc_delta_mine(slice, se->load.weight, load);
668 * We calculate the vruntime slice of a to-be-inserted task.
672 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674 return calc_delta_fair(sched_slice(cfs_rq, se), se);
678 * Update the current task's runtime statistics. Skip current tasks that
679 * are not in our scheduling class.
682 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
683 unsigned long delta_exec)
685 unsigned long delta_exec_weighted;
687 schedstat_set(curr->statistics.exec_max,
688 max((u64)delta_exec, curr->statistics.exec_max));
690 curr->sum_exec_runtime += delta_exec;
691 schedstat_add(cfs_rq, exec_clock, delta_exec);
692 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
694 curr->vruntime += delta_exec_weighted;
695 update_min_vruntime(cfs_rq);
698 static void update_curr(struct cfs_rq *cfs_rq)
700 struct sched_entity *curr = cfs_rq->curr;
701 u64 now = rq_of(cfs_rq)->clock_task;
702 unsigned long delta_exec;
708 * Get the amount of time the current task was running
709 * since the last time we changed load (this cannot
710 * overflow on 32 bits):
712 delta_exec = (unsigned long)(now - curr->exec_start);
716 __update_curr(cfs_rq, curr, delta_exec);
717 curr->exec_start = now;
719 if (entity_is_task(curr)) {
720 struct task_struct *curtask = task_of(curr);
722 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
723 cpuacct_charge(curtask, delta_exec);
724 account_group_exec_runtime(curtask, delta_exec);
727 account_cfs_rq_runtime(cfs_rq, delta_exec);
731 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
733 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
737 * Task is being enqueued - update stats:
739 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 * Are we enqueueing a waiting task? (for current tasks
743 * a dequeue/enqueue event is a NOP)
745 if (se != cfs_rq->curr)
746 update_stats_wait_start(cfs_rq, se);
750 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
752 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
753 rq_of(cfs_rq)->clock - se->statistics.wait_start));
754 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
755 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
756 rq_of(cfs_rq)->clock - se->statistics.wait_start);
757 #ifdef CONFIG_SCHEDSTATS
758 if (entity_is_task(se)) {
759 trace_sched_stat_wait(task_of(se),
760 rq_of(cfs_rq)->clock - se->statistics.wait_start);
763 schedstat_set(se->statistics.wait_start, 0);
767 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Mark the end of the wait period if dequeueing a
773 if (se != cfs_rq->curr)
774 update_stats_wait_end(cfs_rq, se);
778 * We are picking a new current task - update its stats:
781 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
784 * We are starting a new run period:
786 se->exec_start = rq_of(cfs_rq)->clock_task;
789 /**************************************************
790 * Scheduling class queueing methods:
793 #ifdef CONFIG_NUMA_BALANCING
795 * numa task sample period in ms
797 unsigned int sysctl_numa_balancing_scan_period_min = 100;
798 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
799 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
801 /* Portion of address space to scan in MB */
802 unsigned int sysctl_numa_balancing_scan_size = 256;
804 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
805 unsigned int sysctl_numa_balancing_scan_delay = 1000;
807 static void task_numa_placement(struct task_struct *p)
811 if (!p->mm) /* for example, ksmd faulting in a user's mm */
813 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
814 if (p->numa_scan_seq == seq)
816 p->numa_scan_seq = seq;
818 /* FIXME: Scheduling placement policy hints go here */
822 * Got a PROT_NONE fault for a page on @node.
824 void task_numa_fault(int node, int pages, bool migrated)
826 struct task_struct *p = current;
828 if (!sched_feat_numa(NUMA))
831 /* FIXME: Allocate task-specific structure for placement policy here */
834 * If pages are properly placed (did not migrate) then scan slower.
835 * This is reset periodically in case of phase changes
838 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
839 p->numa_scan_period + jiffies_to_msecs(10));
841 task_numa_placement(p);
844 static void reset_ptenuma_scan(struct task_struct *p)
846 ACCESS_ONCE(p->mm->numa_scan_seq)++;
847 p->mm->numa_scan_offset = 0;
851 * The expensive part of numa migration is done from task_work context.
852 * Triggered from task_tick_numa().
854 void task_numa_work(struct callback_head *work)
856 unsigned long migrate, next_scan, now = jiffies;
857 struct task_struct *p = current;
858 struct mm_struct *mm = p->mm;
859 struct vm_area_struct *vma;
860 unsigned long start, end;
863 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
865 work->next = work; /* protect against double add */
867 * Who cares about NUMA placement when they're dying.
869 * NOTE: make sure not to dereference p->mm before this check,
870 * exit_task_work() happens _after_ exit_mm() so we could be called
871 * without p->mm even though we still had it when we enqueued this
874 if (p->flags & PF_EXITING)
878 * We do not care about task placement until a task runs on a node
879 * other than the first one used by the address space. This is
880 * largely because migrations are driven by what CPU the task
881 * is running on. If it's never scheduled on another node, it'll
882 * not migrate so why bother trapping the fault.
884 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
885 mm->first_nid = numa_node_id();
886 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
887 /* Are we running on a new node yet? */
888 if (numa_node_id() == mm->first_nid &&
889 !sched_feat_numa(NUMA_FORCE))
892 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
896 * Reset the scan period if enough time has gone by. Objective is that
897 * scanning will be reduced if pages are properly placed. As tasks
898 * can enter different phases this needs to be re-examined. Lacking
899 * proper tracking of reference behaviour, this blunt hammer is used.
901 migrate = mm->numa_next_reset;
902 if (time_after(now, migrate)) {
903 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
904 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
905 xchg(&mm->numa_next_reset, next_scan);
909 * Enforce maximal scan/migration frequency..
911 migrate = mm->numa_next_scan;
912 if (time_before(now, migrate))
915 if (p->numa_scan_period == 0)
916 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
918 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
919 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
923 * Do not set pte_numa if the current running node is rate-limited.
924 * This loses statistics on the fault but if we are unwilling to
925 * migrate to this node, it is less likely we can do useful work
927 if (migrate_ratelimited(numa_node_id()))
930 start = mm->numa_scan_offset;
931 pages = sysctl_numa_balancing_scan_size;
932 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
936 down_read(&mm->mmap_sem);
937 vma = find_vma(mm, start);
939 reset_ptenuma_scan(p);
943 for (; vma; vma = vma->vm_next) {
944 if (!vma_migratable(vma))
947 /* Skip small VMAs. They are not likely to be of relevance */
948 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
952 * Skip inaccessible VMAs to avoid any confusion between
953 * PROT_NONE and NUMA hinting ptes
955 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
959 start = max(start, vma->vm_start);
960 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
961 end = min(end, vma->vm_end);
962 pages -= change_prot_numa(vma, start, end);
967 } while (end != vma->vm_end);
972 * It is possible to reach the end of the VMA list but the last few VMAs are
973 * not guaranteed to the vma_migratable. If they are not, we would find the
974 * !migratable VMA on the next scan but not reset the scanner to the start
978 mm->numa_scan_offset = start;
980 reset_ptenuma_scan(p);
981 up_read(&mm->mmap_sem);
985 * Drive the periodic memory faults..
987 void task_tick_numa(struct rq *rq, struct task_struct *curr)
989 struct callback_head *work = &curr->numa_work;
993 * We don't care about NUMA placement if we don't have memory.
995 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
999 * Using runtime rather than walltime has the dual advantage that
1000 * we (mostly) drive the selection from busy threads and that the
1001 * task needs to have done some actual work before we bother with
1004 now = curr->se.sum_exec_runtime;
1005 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1007 if (now - curr->node_stamp > period) {
1008 if (!curr->node_stamp)
1009 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1010 curr->node_stamp = now;
1012 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1013 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1014 task_work_add(curr, work, true);
1019 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1022 #endif /* CONFIG_NUMA_BALANCING */
1025 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1027 update_load_add(&cfs_rq->load, se->load.weight);
1028 if (!parent_entity(se))
1029 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1031 if (entity_is_task(se))
1032 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1034 cfs_rq->nr_running++;
1038 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1040 update_load_sub(&cfs_rq->load, se->load.weight);
1041 if (!parent_entity(se))
1042 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1043 if (entity_is_task(se))
1044 list_del_init(&se->group_node);
1045 cfs_rq->nr_running--;
1048 #ifdef CONFIG_FAIR_GROUP_SCHED
1050 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1055 * Use this CPU's actual weight instead of the last load_contribution
1056 * to gain a more accurate current total weight. See
1057 * update_cfs_rq_load_contribution().
1059 tg_weight = atomic64_read(&tg->load_avg);
1060 tg_weight -= cfs_rq->tg_load_contrib;
1061 tg_weight += cfs_rq->load.weight;
1066 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1068 long tg_weight, load, shares;
1070 tg_weight = calc_tg_weight(tg, cfs_rq);
1071 load = cfs_rq->load.weight;
1073 shares = (tg->shares * load);
1075 shares /= tg_weight;
1077 if (shares < MIN_SHARES)
1078 shares = MIN_SHARES;
1079 if (shares > tg->shares)
1080 shares = tg->shares;
1084 # else /* CONFIG_SMP */
1085 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1089 # endif /* CONFIG_SMP */
1090 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1091 unsigned long weight)
1094 /* commit outstanding execution time */
1095 if (cfs_rq->curr == se)
1096 update_curr(cfs_rq);
1097 account_entity_dequeue(cfs_rq, se);
1100 update_load_set(&se->load, weight);
1103 account_entity_enqueue(cfs_rq, se);
1106 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1108 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1110 struct task_group *tg;
1111 struct sched_entity *se;
1115 se = tg->se[cpu_of(rq_of(cfs_rq))];
1116 if (!se || throttled_hierarchy(cfs_rq))
1119 if (likely(se->load.weight == tg->shares))
1122 shares = calc_cfs_shares(cfs_rq, tg);
1124 reweight_entity(cfs_rq_of(se), se, shares);
1126 #else /* CONFIG_FAIR_GROUP_SCHED */
1127 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1130 #endif /* CONFIG_FAIR_GROUP_SCHED */
1132 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1133 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1135 * We choose a half-life close to 1 scheduling period.
1136 * Note: The tables below are dependent on this value.
1138 #define LOAD_AVG_PERIOD 32
1139 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1140 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1142 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1143 static const u32 runnable_avg_yN_inv[] = {
1144 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1145 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1146 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1147 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1148 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1149 0x85aac367, 0x82cd8698,
1153 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1154 * over-estimates when re-combining.
1156 static const u32 runnable_avg_yN_sum[] = {
1157 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1158 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1159 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1164 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1166 static __always_inline u64 decay_load(u64 val, u64 n)
1168 unsigned int local_n;
1172 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1175 /* after bounds checking we can collapse to 32-bit */
1179 * As y^PERIOD = 1/2, we can combine
1180 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1181 * With a look-up table which covers k^n (n<PERIOD)
1183 * To achieve constant time decay_load.
1185 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1186 val >>= local_n / LOAD_AVG_PERIOD;
1187 local_n %= LOAD_AVG_PERIOD;
1190 val *= runnable_avg_yN_inv[local_n];
1191 /* We don't use SRR here since we always want to round down. */
1196 * For updates fully spanning n periods, the contribution to runnable
1197 * average will be: \Sum 1024*y^n
1199 * We can compute this reasonably efficiently by combining:
1200 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1202 static u32 __compute_runnable_contrib(u64 n)
1206 if (likely(n <= LOAD_AVG_PERIOD))
1207 return runnable_avg_yN_sum[n];
1208 else if (unlikely(n >= LOAD_AVG_MAX_N))
1209 return LOAD_AVG_MAX;
1211 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1213 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1214 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1216 n -= LOAD_AVG_PERIOD;
1217 } while (n > LOAD_AVG_PERIOD);
1219 contrib = decay_load(contrib, n);
1220 return contrib + runnable_avg_yN_sum[n];
1223 #ifdef CONFIG_SCHED_HMP
1224 #define HMP_VARIABLE_SCALE_SHIFT 16ULL
1225 struct hmp_global_attr {
1226 struct attribute attr;
1227 ssize_t (*show)(struct kobject *kobj,
1228 struct attribute *attr, char *buf);
1229 ssize_t (*store)(struct kobject *a, struct attribute *b,
1230 const char *c, size_t count);
1232 int (*to_sysfs)(int);
1233 int (*from_sysfs)(int);
1234 ssize_t (*to_sysfs_text)(char *buf, int buf_size);
1237 #define HMP_DATA_SYSFS_MAX 8
1239 struct hmp_data_struct {
1240 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1241 int freqinvar_load_scale_enabled;
1243 int multiplier; /* used to scale the time delta */
1244 struct attribute_group attr_group;
1245 struct attribute *attributes[HMP_DATA_SYSFS_MAX + 1];
1246 struct hmp_global_attr attr[HMP_DATA_SYSFS_MAX];
1249 static u64 hmp_variable_scale_convert(u64 delta);
1250 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1251 /* Frequency-Invariant Load Modification:
1252 * Loads are calculated as in PJT's patch however we also scale the current
1253 * contribution in line with the frequency of the CPU that the task was
1255 * In this version, we use a simple linear scale derived from the maximum
1256 * frequency reported by CPUFreq. As an example:
1258 * Consider that we ran a task for 100% of the previous interval.
1260 * Our CPU was under asynchronous frequency control through one of the
1261 * CPUFreq governors.
1263 * The CPUFreq governor reports that it is able to scale the CPU between
1266 * During the period, the CPU was running at 1GHz.
1268 * In this case, our load contribution for that period is calculated as
1269 * 1 * (number_of_active_microseconds)
1271 * This results in our task being able to accumulate maximum load as normal.
1274 * Consider now that our CPU was executing at 500MHz.
1276 * We now scale the load contribution such that it is calculated as
1277 * 0.5 * (number_of_active_microseconds)
1279 * Our task can only record 50% maximum load during this period.
1281 * This represents the task consuming 50% of the CPU's *possible* compute
1282 * capacity. However the task did consume 100% of the CPU's *available*
1283 * compute capacity which is the value seen by the CPUFreq governor and
1284 * user-side CPU Utilization tools.
1286 * Restricting tracked load to be scaled by the CPU's frequency accurately
1287 * represents the consumption of possible compute capacity and allows the
1288 * HMP migration's simple threshold migration strategy to interact more
1289 * predictably with CPUFreq's asynchronous compute capacity changes.
1291 #define SCHED_FREQSCALE_SHIFT 10
1292 struct cpufreq_extents {
1298 /* Flag set when the governor in use only allows one frequency.
1301 #define SCHED_LOAD_FREQINVAR_SINGLEFREQ 0x01
1303 static struct cpufreq_extents freq_scale[CONFIG_NR_CPUS];
1304 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1305 #endif /* CONFIG_SCHED_HMP */
1307 /* We can represent the historical contribution to runnable average as the
1308 * coefficients of a geometric series. To do this we sub-divide our runnable
1309 * history into segments of approximately 1ms (1024us); label the segment that
1310 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1312 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1314 * (now) (~1ms ago) (~2ms ago)
1316 * Let u_i denote the fraction of p_i that the entity was runnable.
1318 * We then designate the fractions u_i as our co-efficients, yielding the
1319 * following representation of historical load:
1320 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1322 * We choose y based on the with of a reasonably scheduling period, fixing:
1325 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1326 * approximately half as much as the contribution to load within the last ms
1329 * When a period "rolls over" and we have new u_0`, multiplying the previous
1330 * sum again by y is sufficient to update:
1331 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1332 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1334 static __always_inline int __update_entity_runnable_avg(u64 now,
1335 struct sched_avg *sa,
1341 u32 runnable_contrib;
1342 int delta_w, decayed = 0;
1343 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1345 u32 scaled_runnable_contrib;
1347 u32 curr_scale = 1024;
1348 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1350 delta = now - sa->last_runnable_update;
1351 #ifdef CONFIG_SCHED_HMP
1352 delta = hmp_variable_scale_convert(delta);
1355 * This should only happen when time goes backwards, which it
1356 * unfortunately does during sched clock init when we swap over to TSC.
1358 if ((s64)delta < 0) {
1359 sa->last_runnable_update = now;
1364 * Use 1024ns as the unit of measurement since it's a reasonable
1365 * approximation of 1us and fast to compute.
1370 sa->last_runnable_update = now;
1372 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1373 /* retrieve scale factor for load */
1374 if (hmp_data.freqinvar_load_scale_enabled)
1375 curr_scale = freq_scale[cpu].curr_scale;
1376 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1378 /* delta_w is the amount already accumulated against our next period */
1379 delta_w = sa->runnable_avg_period % 1024;
1380 if (delta + delta_w >= 1024) {
1381 /* period roll-over */
1385 * Now that we know we're crossing a period boundary, figure
1386 * out how much from delta we need to complete the current
1387 * period and accrue it.
1389 delta_w = 1024 - delta_w;
1390 /* scale runnable time if necessary */
1391 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1392 scaled_delta_w = (delta_w * curr_scale)
1393 >> SCHED_FREQSCALE_SHIFT;
1395 sa->runnable_avg_sum += scaled_delta_w;
1397 sa->usage_avg_sum += scaled_delta_w;
1400 sa->runnable_avg_sum += delta_w;
1402 sa->usage_avg_sum += delta_w;
1403 #endif /* #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1404 sa->runnable_avg_period += delta_w;
1408 /* Figure out how many additional periods this update spans */
1409 periods = delta / 1024;
1411 /* decay the load we have accumulated so far */
1412 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1414 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1416 sa->usage_avg_sum = decay_load(sa->usage_avg_sum, periods + 1);
1417 /* add the contribution from this period */
1418 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1419 runnable_contrib = __compute_runnable_contrib(periods);
1420 /* Apply load scaling if necessary.
1421 * Note that multiplying the whole series is same as
1422 * multiplying all terms
1424 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1425 scaled_runnable_contrib = (runnable_contrib * curr_scale)
1426 >> SCHED_FREQSCALE_SHIFT;
1428 sa->runnable_avg_sum += scaled_runnable_contrib;
1430 sa->usage_avg_sum += scaled_runnable_contrib;
1433 sa->runnable_avg_sum += runnable_contrib;
1435 sa->usage_avg_sum += runnable_contrib;
1436 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1437 sa->runnable_avg_period += runnable_contrib;
1440 /* Remainder of delta accrued against u_0` */
1441 /* scale if necessary */
1442 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1443 scaled_delta = ((delta * curr_scale) >> SCHED_FREQSCALE_SHIFT);
1445 sa->runnable_avg_sum += scaled_delta;
1447 sa->usage_avg_sum += scaled_delta;
1450 sa->runnable_avg_sum += delta;
1452 sa->usage_avg_sum += delta;
1453 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1454 sa->runnable_avg_period += delta;
1459 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1460 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1462 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1463 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1465 decays -= se->avg.decay_count;
1467 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1468 se->avg.decay_count = 0;
1472 #ifdef CONFIG_FAIR_GROUP_SCHED
1473 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1476 struct task_group *tg = cfs_rq->tg;
1479 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1480 tg_contrib -= cfs_rq->tg_load_contrib;
1482 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1483 atomic64_add(tg_contrib, &tg->load_avg);
1484 cfs_rq->tg_load_contrib += tg_contrib;
1489 * Aggregate cfs_rq runnable averages into an equivalent task_group
1490 * representation for computing load contributions.
1492 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1493 struct cfs_rq *cfs_rq)
1495 struct task_group *tg = cfs_rq->tg;
1496 long contrib, usage_contrib;
1498 /* The fraction of a cpu used by this cfs_rq */
1499 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1500 sa->runnable_avg_period + 1);
1501 contrib -= cfs_rq->tg_runnable_contrib;
1503 usage_contrib = div_u64(sa->usage_avg_sum << NICE_0_SHIFT,
1504 sa->runnable_avg_period + 1);
1505 usage_contrib -= cfs_rq->tg_usage_contrib;
1508 * contrib/usage at this point represent deltas, only update if they
1511 if ((abs(contrib) > cfs_rq->tg_runnable_contrib / 64) ||
1512 (abs(usage_contrib) > cfs_rq->tg_usage_contrib / 64)) {
1513 atomic_add(contrib, &tg->runnable_avg);
1514 cfs_rq->tg_runnable_contrib += contrib;
1516 atomic_add(usage_contrib, &tg->usage_avg);
1517 cfs_rq->tg_usage_contrib += usage_contrib;
1521 static inline void __update_group_entity_contrib(struct sched_entity *se)
1523 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1524 struct task_group *tg = cfs_rq->tg;
1529 contrib = cfs_rq->tg_load_contrib * tg->shares;
1530 se->avg.load_avg_contrib = div64_u64(contrib,
1531 atomic64_read(&tg->load_avg) + 1);
1534 * For group entities we need to compute a correction term in the case
1535 * that they are consuming <1 cpu so that we would contribute the same
1536 * load as a task of equal weight.
1538 * Explicitly co-ordinating this measurement would be expensive, but
1539 * fortunately the sum of each cpus contribution forms a usable
1540 * lower-bound on the true value.
1542 * Consider the aggregate of 2 contributions. Either they are disjoint
1543 * (and the sum represents true value) or they are disjoint and we are
1544 * understating by the aggregate of their overlap.
1546 * Extending this to N cpus, for a given overlap, the maximum amount we
1547 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1548 * cpus that overlap for this interval and w_i is the interval width.
1550 * On a small machine; the first term is well-bounded which bounds the
1551 * total error since w_i is a subset of the period. Whereas on a
1552 * larger machine, while this first term can be larger, if w_i is the
1553 * of consequential size guaranteed to see n_i*w_i quickly converge to
1554 * our upper bound of 1-cpu.
1556 runnable_avg = atomic_read(&tg->runnable_avg);
1557 if (runnable_avg < NICE_0_LOAD) {
1558 se->avg.load_avg_contrib *= runnable_avg;
1559 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1563 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1564 int force_update) {}
1565 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1566 struct cfs_rq *cfs_rq) {}
1567 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1570 static inline void __update_task_entity_contrib(struct sched_entity *se)
1574 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1575 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1576 contrib /= (se->avg.runnable_avg_period + 1);
1577 se->avg.load_avg_contrib = scale_load(contrib);
1578 trace_sched_task_load_contrib(task_of(se), se->avg.load_avg_contrib);
1579 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1580 contrib /= (se->avg.runnable_avg_period + 1);
1581 se->avg.load_avg_ratio = scale_load(contrib);
1582 trace_sched_task_runnable_ratio(task_of(se), se->avg.load_avg_ratio);
1585 /* Compute the current contribution to load_avg by se, return any delta */
1586 static long __update_entity_load_avg_contrib(struct sched_entity *se, long *ratio)
1588 long old_contrib = se->avg.load_avg_contrib;
1589 long old_ratio = se->avg.load_avg_ratio;
1591 if (entity_is_task(se)) {
1592 __update_task_entity_contrib(se);
1594 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1595 __update_group_entity_contrib(se);
1599 *ratio = se->avg.load_avg_ratio - old_ratio;
1600 return se->avg.load_avg_contrib - old_contrib;
1603 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1606 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1607 cfs_rq->blocked_load_avg -= load_contrib;
1609 cfs_rq->blocked_load_avg = 0;
1612 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1614 /* Update a sched_entity's runnable average */
1615 static inline void update_entity_load_avg(struct sched_entity *se,
1618 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1619 long contrib_delta, ratio_delta;
1621 int cpu = -1; /* not used in normal case */
1623 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1624 cpu = cfs_rq->rq->cpu;
1627 * For a group entity we need to use their owned cfs_rq_clock_task() in
1628 * case they are the parent of a throttled hierarchy.
1630 if (entity_is_task(se))
1631 now = cfs_rq_clock_task(cfs_rq);
1633 now = cfs_rq_clock_task(group_cfs_rq(se));
1635 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1636 cfs_rq->curr == se, cpu))
1639 contrib_delta = __update_entity_load_avg_contrib(se, &ratio_delta);
1645 cfs_rq->runnable_load_avg += contrib_delta;
1646 rq_of(cfs_rq)->avg.load_avg_ratio += ratio_delta;
1648 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1653 * Decay the load contributed by all blocked children and account this so that
1654 * their contribution may appropriately discounted when they wake up.
1656 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1658 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1661 decays = now - cfs_rq->last_decay;
1662 if (!decays && !force_update)
1665 if (atomic64_read(&cfs_rq->removed_load)) {
1666 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1667 subtract_blocked_load_contrib(cfs_rq, removed_load);
1671 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1673 atomic64_add(decays, &cfs_rq->decay_counter);
1674 cfs_rq->last_decay = now;
1677 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1680 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1682 int cpu = -1; /* not used in normal case */
1684 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1687 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1689 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1690 trace_sched_rq_runnable_ratio(cpu_of(rq), rq->avg.load_avg_ratio);
1691 trace_sched_rq_runnable_load(cpu_of(rq), rq->cfs.runnable_load_avg);
1692 trace_sched_rq_nr_running(cpu_of(rq), rq->nr_running, rq->nr_iowait.counter);
1695 /* Add the load generated by se into cfs_rq's child load-average */
1696 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1697 struct sched_entity *se,
1701 * We track migrations using entity decay_count <= 0, on a wake-up
1702 * migration we use a negative decay count to track the remote decays
1703 * accumulated while sleeping.
1705 if (unlikely(se->avg.decay_count <= 0)) {
1706 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1707 if (se->avg.decay_count) {
1709 * In a wake-up migration we have to approximate the
1710 * time sleeping. This is because we can't synchronize
1711 * clock_task between the two cpus, and it is not
1712 * guaranteed to be read-safe. Instead, we can
1713 * approximate this using our carried decays, which are
1714 * explicitly atomically readable.
1716 se->avg.last_runnable_update -= (-se->avg.decay_count)
1718 update_entity_load_avg(se, 0);
1719 /* Indicate that we're now synchronized and on-rq */
1720 se->avg.decay_count = 0;
1724 __synchronize_entity_decay(se);
1727 /* migrated tasks did not contribute to our blocked load */
1729 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1730 update_entity_load_avg(se, 0);
1733 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1734 rq_of(cfs_rq)->avg.load_avg_ratio += se->avg.load_avg_ratio;
1736 /* we force update consideration on load-balancer moves */
1737 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1741 * Remove se's load from this cfs_rq child load-average, if the entity is
1742 * transitioning to a blocked state we track its projected decay using
1745 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1746 struct sched_entity *se,
1749 update_entity_load_avg(se, 1);
1750 /* we force update consideration on load-balancer moves */
1751 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1753 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1754 rq_of(cfs_rq)->avg.load_avg_ratio -= se->avg.load_avg_ratio;
1757 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1758 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1759 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1763 * Update the rq's load with the elapsed running time before entering
1764 * idle. if the last scheduled task is not a CFS task, idle_enter will
1765 * be the only way to update the runnable statistic.
1767 void idle_enter_fair(struct rq *this_rq)
1769 update_rq_runnable_avg(this_rq, 1);
1773 * Update the rq's load with the elapsed idle time before a task is
1774 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1775 * be the only way to update the runnable statistic.
1777 void idle_exit_fair(struct rq *this_rq)
1779 update_rq_runnable_avg(this_rq, 0);
1783 static inline void update_entity_load_avg(struct sched_entity *se,
1784 int update_cfs_rq) {}
1785 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1786 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1787 struct sched_entity *se,
1789 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1790 struct sched_entity *se,
1792 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1793 int force_update) {}
1796 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1798 #ifdef CONFIG_SCHEDSTATS
1799 struct task_struct *tsk = NULL;
1801 if (entity_is_task(se))
1804 if (se->statistics.sleep_start) {
1805 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1810 if (unlikely(delta > se->statistics.sleep_max))
1811 se->statistics.sleep_max = delta;
1813 se->statistics.sleep_start = 0;
1814 se->statistics.sum_sleep_runtime += delta;
1817 account_scheduler_latency(tsk, delta >> 10, 1);
1818 trace_sched_stat_sleep(tsk, delta);
1821 if (se->statistics.block_start) {
1822 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1827 if (unlikely(delta > se->statistics.block_max))
1828 se->statistics.block_max = delta;
1830 se->statistics.block_start = 0;
1831 se->statistics.sum_sleep_runtime += delta;
1834 if (tsk->in_iowait) {
1835 se->statistics.iowait_sum += delta;
1836 se->statistics.iowait_count++;
1837 trace_sched_stat_iowait(tsk, delta);
1840 trace_sched_stat_blocked(tsk, delta);
1843 * Blocking time is in units of nanosecs, so shift by
1844 * 20 to get a milliseconds-range estimation of the
1845 * amount of time that the task spent sleeping:
1847 if (unlikely(prof_on == SLEEP_PROFILING)) {
1848 profile_hits(SLEEP_PROFILING,
1849 (void *)get_wchan(tsk),
1852 account_scheduler_latency(tsk, delta >> 10, 0);
1858 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1860 #ifdef CONFIG_SCHED_DEBUG
1861 s64 d = se->vruntime - cfs_rq->min_vruntime;
1866 if (d > 3*sysctl_sched_latency)
1867 schedstat_inc(cfs_rq, nr_spread_over);
1872 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1874 u64 vruntime = cfs_rq->min_vruntime;
1877 * The 'current' period is already promised to the current tasks,
1878 * however the extra weight of the new task will slow them down a
1879 * little, place the new task so that it fits in the slot that
1880 * stays open at the end.
1882 if (initial && sched_feat(START_DEBIT))
1883 vruntime += sched_vslice(cfs_rq, se);
1885 /* sleeps up to a single latency don't count. */
1887 unsigned long thresh = sysctl_sched_latency;
1890 * Halve their sleep time's effect, to allow
1891 * for a gentler effect of sleepers:
1893 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1899 /* ensure we never gain time by being placed backwards. */
1900 se->vruntime = max_vruntime(se->vruntime, vruntime);
1903 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1906 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1909 * Update the normalized vruntime before updating min_vruntime
1910 * through callig update_curr().
1912 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1913 se->vruntime += cfs_rq->min_vruntime;
1916 * Update run-time statistics of the 'current'.
1918 update_curr(cfs_rq);
1919 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1920 account_entity_enqueue(cfs_rq, se);
1921 update_cfs_shares(cfs_rq);
1923 if (flags & ENQUEUE_WAKEUP) {
1924 place_entity(cfs_rq, se, 0);
1925 enqueue_sleeper(cfs_rq, se);
1928 update_stats_enqueue(cfs_rq, se);
1929 check_spread(cfs_rq, se);
1930 if (se != cfs_rq->curr)
1931 __enqueue_entity(cfs_rq, se);
1934 if (cfs_rq->nr_running == 1) {
1935 list_add_leaf_cfs_rq(cfs_rq);
1936 check_enqueue_throttle(cfs_rq);
1940 static void __clear_buddies_last(struct sched_entity *se)
1942 for_each_sched_entity(se) {
1943 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1944 if (cfs_rq->last == se)
1945 cfs_rq->last = NULL;
1951 static void __clear_buddies_next(struct sched_entity *se)
1953 for_each_sched_entity(se) {
1954 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1955 if (cfs_rq->next == se)
1956 cfs_rq->next = NULL;
1962 static void __clear_buddies_skip(struct sched_entity *se)
1964 for_each_sched_entity(se) {
1965 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1966 if (cfs_rq->skip == se)
1967 cfs_rq->skip = NULL;
1973 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1975 if (cfs_rq->last == se)
1976 __clear_buddies_last(se);
1978 if (cfs_rq->next == se)
1979 __clear_buddies_next(se);
1981 if (cfs_rq->skip == se)
1982 __clear_buddies_skip(se);
1985 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1988 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1991 * Update run-time statistics of the 'current'.
1993 update_curr(cfs_rq);
1994 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1996 update_stats_dequeue(cfs_rq, se);
1997 if (flags & DEQUEUE_SLEEP) {
1998 #ifdef CONFIG_SCHEDSTATS
1999 if (entity_is_task(se)) {
2000 struct task_struct *tsk = task_of(se);
2002 if (tsk->state & TASK_INTERRUPTIBLE)
2003 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
2004 if (tsk->state & TASK_UNINTERRUPTIBLE)
2005 se->statistics.block_start = rq_of(cfs_rq)->clock;
2010 clear_buddies(cfs_rq, se);
2012 if (se != cfs_rq->curr)
2013 __dequeue_entity(cfs_rq, se);
2015 account_entity_dequeue(cfs_rq, se);
2018 * Normalize the entity after updating the min_vruntime because the
2019 * update can refer to the ->curr item and we need to reflect this
2020 * movement in our normalized position.
2022 if (!(flags & DEQUEUE_SLEEP))
2023 se->vruntime -= cfs_rq->min_vruntime;
2025 /* return excess runtime on last dequeue */
2026 return_cfs_rq_runtime(cfs_rq);
2028 update_min_vruntime(cfs_rq);
2029 update_cfs_shares(cfs_rq);
2033 * Preempt the current task with a newly woken task if needed:
2036 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2038 unsigned long ideal_runtime, delta_exec;
2039 struct sched_entity *se;
2042 ideal_runtime = sched_slice(cfs_rq, curr);
2043 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2044 if (delta_exec > ideal_runtime) {
2045 resched_task(rq_of(cfs_rq)->curr);
2047 * The current task ran long enough, ensure it doesn't get
2048 * re-elected due to buddy favours.
2050 clear_buddies(cfs_rq, curr);
2055 * Ensure that a task that missed wakeup preemption by a
2056 * narrow margin doesn't have to wait for a full slice.
2057 * This also mitigates buddy induced latencies under load.
2059 if (delta_exec < sysctl_sched_min_granularity)
2062 se = __pick_first_entity(cfs_rq);
2063 delta = curr->vruntime - se->vruntime;
2068 if (delta > ideal_runtime)
2069 resched_task(rq_of(cfs_rq)->curr);
2073 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2075 /* 'current' is not kept within the tree. */
2078 * Any task has to be enqueued before it get to execute on
2079 * a CPU. So account for the time it spent waiting on the
2082 update_stats_wait_end(cfs_rq, se);
2083 __dequeue_entity(cfs_rq, se);
2084 update_entity_load_avg(se, 1);
2087 update_stats_curr_start(cfs_rq, se);
2089 #ifdef CONFIG_SCHEDSTATS
2091 * Track our maximum slice length, if the CPU's load is at
2092 * least twice that of our own weight (i.e. dont track it
2093 * when there are only lesser-weight tasks around):
2095 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2096 se->statistics.slice_max = max(se->statistics.slice_max,
2097 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2100 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2104 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2107 * Pick the next process, keeping these things in mind, in this order:
2108 * 1) keep things fair between processes/task groups
2109 * 2) pick the "next" process, since someone really wants that to run
2110 * 3) pick the "last" process, for cache locality
2111 * 4) do not run the "skip" process, if something else is available
2113 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2115 struct sched_entity *se = __pick_first_entity(cfs_rq);
2116 struct sched_entity *left = se;
2119 * Avoid running the skip buddy, if running something else can
2120 * be done without getting too unfair.
2122 if (cfs_rq->skip == se) {
2123 struct sched_entity *second = __pick_next_entity(se);
2124 if (second && wakeup_preempt_entity(second, left) < 1)
2129 * Prefer last buddy, try to return the CPU to a preempted task.
2131 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2135 * Someone really wants this to run. If it's not unfair, run it.
2137 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2140 clear_buddies(cfs_rq, se);
2145 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2147 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2150 * If still on the runqueue then deactivate_task()
2151 * was not called and update_curr() has to be done:
2154 update_curr(cfs_rq);
2156 /* throttle cfs_rqs exceeding runtime */
2157 check_cfs_rq_runtime(cfs_rq);
2159 check_spread(cfs_rq, prev);
2161 update_stats_wait_start(cfs_rq, prev);
2162 /* Put 'current' back into the tree. */
2163 __enqueue_entity(cfs_rq, prev);
2164 /* in !on_rq case, update occurred at dequeue */
2165 update_entity_load_avg(prev, 1);
2167 cfs_rq->curr = NULL;
2171 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2174 * Update run-time statistics of the 'current'.
2176 update_curr(cfs_rq);
2179 * Ensure that runnable average is periodically updated.
2181 update_entity_load_avg(curr, 1);
2182 update_cfs_rq_blocked_load(cfs_rq, 1);
2183 update_cfs_shares(cfs_rq);
2185 #ifdef CONFIG_SCHED_HRTICK
2187 * queued ticks are scheduled to match the slice, so don't bother
2188 * validating it and just reschedule.
2191 resched_task(rq_of(cfs_rq)->curr);
2195 * don't let the period tick interfere with the hrtick preemption
2197 if (!sched_feat(DOUBLE_TICK) &&
2198 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2202 if (cfs_rq->nr_running > 1)
2203 check_preempt_tick(cfs_rq, curr);
2207 /**************************************************
2208 * CFS bandwidth control machinery
2211 #ifdef CONFIG_CFS_BANDWIDTH
2213 #ifdef HAVE_JUMP_LABEL
2214 static struct static_key __cfs_bandwidth_used;
2216 static inline bool cfs_bandwidth_used(void)
2218 return static_key_false(&__cfs_bandwidth_used);
2221 void cfs_bandwidth_usage_inc(void)
2223 static_key_slow_inc(&__cfs_bandwidth_used);
2226 void cfs_bandwidth_usage_dec(void)
2228 static_key_slow_dec(&__cfs_bandwidth_used);
2230 #else /* HAVE_JUMP_LABEL */
2231 static bool cfs_bandwidth_used(void)
2236 void cfs_bandwidth_usage_inc(void) {}
2237 void cfs_bandwidth_usage_dec(void) {}
2238 #endif /* HAVE_JUMP_LABEL */
2241 * default period for cfs group bandwidth.
2242 * default: 0.1s, units: nanoseconds
2244 static inline u64 default_cfs_period(void)
2246 return 100000000ULL;
2249 static inline u64 sched_cfs_bandwidth_slice(void)
2251 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2255 * Replenish runtime according to assigned quota and update expiration time.
2256 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2257 * additional synchronization around rq->lock.
2259 * requires cfs_b->lock
2261 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2265 if (cfs_b->quota == RUNTIME_INF)
2268 now = sched_clock_cpu(smp_processor_id());
2269 cfs_b->runtime = cfs_b->quota;
2270 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2273 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2275 return &tg->cfs_bandwidth;
2278 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2279 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2281 if (unlikely(cfs_rq->throttle_count))
2282 return cfs_rq->throttled_clock_task;
2284 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2287 /* returns 0 on failure to allocate runtime */
2288 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2290 struct task_group *tg = cfs_rq->tg;
2291 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2292 u64 amount = 0, min_amount, expires;
2294 /* note: this is a positive sum as runtime_remaining <= 0 */
2295 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2297 raw_spin_lock(&cfs_b->lock);
2298 if (cfs_b->quota == RUNTIME_INF)
2299 amount = min_amount;
2302 * If the bandwidth pool has become inactive, then at least one
2303 * period must have elapsed since the last consumption.
2304 * Refresh the global state and ensure bandwidth timer becomes
2307 if (!cfs_b->timer_active) {
2308 __refill_cfs_bandwidth_runtime(cfs_b);
2309 __start_cfs_bandwidth(cfs_b);
2312 if (cfs_b->runtime > 0) {
2313 amount = min(cfs_b->runtime, min_amount);
2314 cfs_b->runtime -= amount;
2318 expires = cfs_b->runtime_expires;
2319 raw_spin_unlock(&cfs_b->lock);
2321 cfs_rq->runtime_remaining += amount;
2323 * we may have advanced our local expiration to account for allowed
2324 * spread between our sched_clock and the one on which runtime was
2327 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2328 cfs_rq->runtime_expires = expires;
2330 return cfs_rq->runtime_remaining > 0;
2334 * Note: This depends on the synchronization provided by sched_clock and the
2335 * fact that rq->clock snapshots this value.
2337 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2339 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2340 struct rq *rq = rq_of(cfs_rq);
2342 /* if the deadline is ahead of our clock, nothing to do */
2343 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2346 if (cfs_rq->runtime_remaining < 0)
2350 * If the local deadline has passed we have to consider the
2351 * possibility that our sched_clock is 'fast' and the global deadline
2352 * has not truly expired.
2354 * Fortunately we can check determine whether this the case by checking
2355 * whether the global deadline has advanced.
2358 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2359 /* extend local deadline, drift is bounded above by 2 ticks */
2360 cfs_rq->runtime_expires += TICK_NSEC;
2362 /* global deadline is ahead, expiration has passed */
2363 cfs_rq->runtime_remaining = 0;
2367 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2368 unsigned long delta_exec)
2370 /* dock delta_exec before expiring quota (as it could span periods) */
2371 cfs_rq->runtime_remaining -= delta_exec;
2372 expire_cfs_rq_runtime(cfs_rq);
2374 if (likely(cfs_rq->runtime_remaining > 0))
2378 * if we're unable to extend our runtime we resched so that the active
2379 * hierarchy can be throttled
2381 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2382 resched_task(rq_of(cfs_rq)->curr);
2385 static __always_inline
2386 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2388 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2391 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2394 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2396 return cfs_bandwidth_used() && cfs_rq->throttled;
2399 /* check whether cfs_rq, or any parent, is throttled */
2400 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2402 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2406 * Ensure that neither of the group entities corresponding to src_cpu or
2407 * dest_cpu are members of a throttled hierarchy when performing group
2408 * load-balance operations.
2410 static inline int throttled_lb_pair(struct task_group *tg,
2411 int src_cpu, int dest_cpu)
2413 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2415 src_cfs_rq = tg->cfs_rq[src_cpu];
2416 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2418 return throttled_hierarchy(src_cfs_rq) ||
2419 throttled_hierarchy(dest_cfs_rq);
2422 /* updated child weight may affect parent so we have to do this bottom up */
2423 static int tg_unthrottle_up(struct task_group *tg, void *data)
2425 struct rq *rq = data;
2426 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2428 cfs_rq->throttle_count--;
2430 if (!cfs_rq->throttle_count) {
2431 /* adjust cfs_rq_clock_task() */
2432 cfs_rq->throttled_clock_task_time += rq->clock_task -
2433 cfs_rq->throttled_clock_task;
2440 static int tg_throttle_down(struct task_group *tg, void *data)
2442 struct rq *rq = data;
2443 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2445 /* group is entering throttled state, stop time */
2446 if (!cfs_rq->throttle_count)
2447 cfs_rq->throttled_clock_task = rq->clock_task;
2448 cfs_rq->throttle_count++;
2453 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2455 struct rq *rq = rq_of(cfs_rq);
2456 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2457 struct sched_entity *se;
2458 long task_delta, dequeue = 1;
2460 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2462 /* freeze hierarchy runnable averages while throttled */
2464 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2467 task_delta = cfs_rq->h_nr_running;
2468 for_each_sched_entity(se) {
2469 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2470 /* throttled entity or throttle-on-deactivate */
2475 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2476 qcfs_rq->h_nr_running -= task_delta;
2478 if (qcfs_rq->load.weight)
2483 rq->nr_running -= task_delta;
2485 cfs_rq->throttled = 1;
2486 cfs_rq->throttled_clock = rq->clock;
2487 raw_spin_lock(&cfs_b->lock);
2488 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2489 if (!cfs_b->timer_active)
2490 __start_cfs_bandwidth(cfs_b);
2491 raw_spin_unlock(&cfs_b->lock);
2494 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2496 struct rq *rq = rq_of(cfs_rq);
2497 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2498 struct sched_entity *se;
2502 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2504 cfs_rq->throttled = 0;
2505 raw_spin_lock(&cfs_b->lock);
2506 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2507 list_del_rcu(&cfs_rq->throttled_list);
2508 raw_spin_unlock(&cfs_b->lock);
2510 update_rq_clock(rq);
2511 /* update hierarchical throttle state */
2512 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2514 if (!cfs_rq->load.weight)
2517 task_delta = cfs_rq->h_nr_running;
2518 for_each_sched_entity(se) {
2522 cfs_rq = cfs_rq_of(se);
2524 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2525 cfs_rq->h_nr_running += task_delta;
2527 if (cfs_rq_throttled(cfs_rq))
2532 rq->nr_running += task_delta;
2534 /* determine whether we need to wake up potentially idle cpu */
2535 if (rq->curr == rq->idle && rq->cfs.nr_running)
2536 resched_task(rq->curr);
2539 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2540 u64 remaining, u64 expires)
2542 struct cfs_rq *cfs_rq;
2543 u64 runtime = remaining;
2546 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2548 struct rq *rq = rq_of(cfs_rq);
2550 raw_spin_lock(&rq->lock);
2551 if (!cfs_rq_throttled(cfs_rq))
2554 runtime = -cfs_rq->runtime_remaining + 1;
2555 if (runtime > remaining)
2556 runtime = remaining;
2557 remaining -= runtime;
2559 cfs_rq->runtime_remaining += runtime;
2560 cfs_rq->runtime_expires = expires;
2562 /* we check whether we're throttled above */
2563 if (cfs_rq->runtime_remaining > 0)
2564 unthrottle_cfs_rq(cfs_rq);
2567 raw_spin_unlock(&rq->lock);
2578 * Responsible for refilling a task_group's bandwidth and unthrottling its
2579 * cfs_rqs as appropriate. If there has been no activity within the last
2580 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2581 * used to track this state.
2583 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2585 u64 runtime, runtime_expires;
2586 int idle = 1, throttled;
2588 raw_spin_lock(&cfs_b->lock);
2589 /* no need to continue the timer with no bandwidth constraint */
2590 if (cfs_b->quota == RUNTIME_INF)
2593 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2594 /* idle depends on !throttled (for the case of a large deficit) */
2595 idle = cfs_b->idle && !throttled;
2596 cfs_b->nr_periods += overrun;
2598 /* if we're going inactive then everything else can be deferred */
2603 * if we have relooped after returning idle once, we need to update our
2604 * status as actually running, so that other cpus doing
2605 * __start_cfs_bandwidth will stop trying to cancel us.
2607 cfs_b->timer_active = 1;
2609 __refill_cfs_bandwidth_runtime(cfs_b);
2612 /* mark as potentially idle for the upcoming period */
2617 /* account preceding periods in which throttling occurred */
2618 cfs_b->nr_throttled += overrun;
2621 * There are throttled entities so we must first use the new bandwidth
2622 * to unthrottle them before making it generally available. This
2623 * ensures that all existing debts will be paid before a new cfs_rq is
2626 runtime = cfs_b->runtime;
2627 runtime_expires = cfs_b->runtime_expires;
2631 * This check is repeated as we are holding onto the new bandwidth
2632 * while we unthrottle. This can potentially race with an unthrottled
2633 * group trying to acquire new bandwidth from the global pool.
2635 while (throttled && runtime > 0) {
2636 raw_spin_unlock(&cfs_b->lock);
2637 /* we can't nest cfs_b->lock while distributing bandwidth */
2638 runtime = distribute_cfs_runtime(cfs_b, runtime,
2640 raw_spin_lock(&cfs_b->lock);
2642 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2645 /* return (any) remaining runtime */
2646 cfs_b->runtime = runtime;
2648 * While we are ensured activity in the period following an
2649 * unthrottle, this also covers the case in which the new bandwidth is
2650 * insufficient to cover the existing bandwidth deficit. (Forcing the
2651 * timer to remain active while there are any throttled entities.)
2656 cfs_b->timer_active = 0;
2657 raw_spin_unlock(&cfs_b->lock);
2662 /* a cfs_rq won't donate quota below this amount */
2663 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2664 /* minimum remaining period time to redistribute slack quota */
2665 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2666 /* how long we wait to gather additional slack before distributing */
2667 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2670 * Are we near the end of the current quota period?
2672 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
2673 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
2674 * migrate_hrtimers, base is never cleared, so we are fine.
2676 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2678 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2681 /* if the call-back is running a quota refresh is already occurring */
2682 if (hrtimer_callback_running(refresh_timer))
2685 /* is a quota refresh about to occur? */
2686 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2687 if (remaining < min_expire)
2693 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2695 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2697 /* if there's a quota refresh soon don't bother with slack */
2698 if (runtime_refresh_within(cfs_b, min_left))
2701 start_bandwidth_timer(&cfs_b->slack_timer,
2702 ns_to_ktime(cfs_bandwidth_slack_period));
2705 /* we know any runtime found here is valid as update_curr() precedes return */
2706 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2708 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2709 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2711 if (slack_runtime <= 0)
2714 raw_spin_lock(&cfs_b->lock);
2715 if (cfs_b->quota != RUNTIME_INF &&
2716 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2717 cfs_b->runtime += slack_runtime;
2719 /* we are under rq->lock, defer unthrottling using a timer */
2720 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2721 !list_empty(&cfs_b->throttled_cfs_rq))
2722 start_cfs_slack_bandwidth(cfs_b);
2724 raw_spin_unlock(&cfs_b->lock);
2726 /* even if it's not valid for return we don't want to try again */
2727 cfs_rq->runtime_remaining -= slack_runtime;
2730 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2732 if (!cfs_bandwidth_used())
2735 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2738 __return_cfs_rq_runtime(cfs_rq);
2742 * This is done with a timer (instead of inline with bandwidth return) since
2743 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2745 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2747 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2750 /* confirm we're still not at a refresh boundary */
2751 raw_spin_lock(&cfs_b->lock);
2752 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
2753 raw_spin_unlock(&cfs_b->lock);
2757 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2758 runtime = cfs_b->runtime;
2761 expires = cfs_b->runtime_expires;
2762 raw_spin_unlock(&cfs_b->lock);
2767 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2769 raw_spin_lock(&cfs_b->lock);
2770 if (expires == cfs_b->runtime_expires)
2771 cfs_b->runtime = runtime;
2772 raw_spin_unlock(&cfs_b->lock);
2776 * When a group wakes up we want to make sure that its quota is not already
2777 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2778 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2780 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2782 if (!cfs_bandwidth_used())
2785 /* an active group must be handled by the update_curr()->put() path */
2786 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2789 /* ensure the group is not already throttled */
2790 if (cfs_rq_throttled(cfs_rq))
2793 /* update runtime allocation */
2794 account_cfs_rq_runtime(cfs_rq, 0);
2795 if (cfs_rq->runtime_remaining <= 0)
2796 throttle_cfs_rq(cfs_rq);
2799 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2800 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2802 if (!cfs_bandwidth_used())
2805 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2809 * it's possible for a throttled entity to be forced into a running
2810 * state (e.g. set_curr_task), in this case we're finished.
2812 if (cfs_rq_throttled(cfs_rq))
2815 throttle_cfs_rq(cfs_rq);
2818 static inline u64 default_cfs_period(void);
2819 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2820 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2822 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2824 struct cfs_bandwidth *cfs_b =
2825 container_of(timer, struct cfs_bandwidth, slack_timer);
2826 do_sched_cfs_slack_timer(cfs_b);
2828 return HRTIMER_NORESTART;
2831 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2833 struct cfs_bandwidth *cfs_b =
2834 container_of(timer, struct cfs_bandwidth, period_timer);
2840 now = hrtimer_cb_get_time(timer);
2841 overrun = hrtimer_forward(timer, now, cfs_b->period);
2846 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2849 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2852 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2854 raw_spin_lock_init(&cfs_b->lock);
2856 cfs_b->quota = RUNTIME_INF;
2857 cfs_b->period = ns_to_ktime(default_cfs_period());
2859 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2860 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2861 cfs_b->period_timer.function = sched_cfs_period_timer;
2862 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2863 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2866 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2868 cfs_rq->runtime_enabled = 0;
2869 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2872 /* requires cfs_b->lock, may release to reprogram timer */
2873 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2876 * The timer may be active because we're trying to set a new bandwidth
2877 * period or because we're racing with the tear-down path
2878 * (timer_active==0 becomes visible before the hrtimer call-back
2879 * terminates). In either case we ensure that it's re-programmed
2881 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
2882 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
2883 /* bounce the lock to allow do_sched_cfs_period_timer to run */
2884 raw_spin_unlock(&cfs_b->lock);
2886 raw_spin_lock(&cfs_b->lock);
2887 /* if someone else restarted the timer then we're done */
2888 if (cfs_b->timer_active)
2892 cfs_b->timer_active = 1;
2893 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2896 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2898 hrtimer_cancel(&cfs_b->period_timer);
2899 hrtimer_cancel(&cfs_b->slack_timer);
2902 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2904 struct cfs_rq *cfs_rq;
2906 for_each_leaf_cfs_rq(rq, cfs_rq) {
2907 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2909 if (!cfs_rq->runtime_enabled)
2913 * clock_task is not advancing so we just need to make sure
2914 * there's some valid quota amount
2916 cfs_rq->runtime_remaining = cfs_b->quota;
2917 if (cfs_rq_throttled(cfs_rq))
2918 unthrottle_cfs_rq(cfs_rq);
2922 #else /* CONFIG_CFS_BANDWIDTH */
2923 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2925 return rq_of(cfs_rq)->clock_task;
2928 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2929 unsigned long delta_exec) {}
2930 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2931 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2932 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2934 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2939 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2944 static inline int throttled_lb_pair(struct task_group *tg,
2945 int src_cpu, int dest_cpu)
2950 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2952 #ifdef CONFIG_FAIR_GROUP_SCHED
2953 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2956 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2960 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2961 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2963 #endif /* CONFIG_CFS_BANDWIDTH */
2965 /**************************************************
2966 * CFS operations on tasks:
2969 #ifdef CONFIG_SCHED_HRTICK
2970 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2972 struct sched_entity *se = &p->se;
2973 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2975 WARN_ON(task_rq(p) != rq);
2977 if (cfs_rq->nr_running > 1) {
2978 u64 slice = sched_slice(cfs_rq, se);
2979 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2980 s64 delta = slice - ran;
2989 * Don't schedule slices shorter than 10000ns, that just
2990 * doesn't make sense. Rely on vruntime for fairness.
2993 delta = max_t(s64, 10000LL, delta);
2995 hrtick_start(rq, delta);
3000 * called from enqueue/dequeue and updates the hrtick when the
3001 * current task is from our class and nr_running is low enough
3004 static void hrtick_update(struct rq *rq)
3006 struct task_struct *curr = rq->curr;
3008 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3011 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3012 hrtick_start_fair(rq, curr);
3014 #else /* !CONFIG_SCHED_HRTICK */
3016 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3020 static inline void hrtick_update(struct rq *rq)
3026 * The enqueue_task method is called before nr_running is
3027 * increased. Here we update the fair scheduling stats and
3028 * then put the task into the rbtree:
3031 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3033 struct cfs_rq *cfs_rq;
3034 struct sched_entity *se = &p->se;
3036 for_each_sched_entity(se) {
3039 cfs_rq = cfs_rq_of(se);
3040 enqueue_entity(cfs_rq, se, flags);
3043 * end evaluation on encountering a throttled cfs_rq
3045 * note: in the case of encountering a throttled cfs_rq we will
3046 * post the final h_nr_running increment below.
3048 if (cfs_rq_throttled(cfs_rq))
3050 cfs_rq->h_nr_running++;
3052 flags = ENQUEUE_WAKEUP;
3055 for_each_sched_entity(se) {
3056 cfs_rq = cfs_rq_of(se);
3057 cfs_rq->h_nr_running++;
3059 if (cfs_rq_throttled(cfs_rq))
3062 update_cfs_shares(cfs_rq);
3063 update_entity_load_avg(se, 1);
3067 update_rq_runnable_avg(rq, rq->nr_running);
3073 static void set_next_buddy(struct sched_entity *se);
3076 * The dequeue_task method is called before nr_running is
3077 * decreased. We remove the task from the rbtree and
3078 * update the fair scheduling stats:
3080 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3082 struct cfs_rq *cfs_rq;
3083 struct sched_entity *se = &p->se;
3084 int task_sleep = flags & DEQUEUE_SLEEP;
3086 for_each_sched_entity(se) {
3087 cfs_rq = cfs_rq_of(se);
3088 dequeue_entity(cfs_rq, se, flags);
3091 * end evaluation on encountering a throttled cfs_rq
3093 * note: in the case of encountering a throttled cfs_rq we will
3094 * post the final h_nr_running decrement below.
3096 if (cfs_rq_throttled(cfs_rq))
3098 cfs_rq->h_nr_running--;
3100 /* Don't dequeue parent if it has other entities besides us */
3101 if (cfs_rq->load.weight) {
3103 * Bias pick_next to pick a task from this cfs_rq, as
3104 * p is sleeping when it is within its sched_slice.
3106 if (task_sleep && parent_entity(se))
3107 set_next_buddy(parent_entity(se));
3109 /* avoid re-evaluating load for this entity */
3110 se = parent_entity(se);
3113 flags |= DEQUEUE_SLEEP;
3116 for_each_sched_entity(se) {
3117 cfs_rq = cfs_rq_of(se);
3118 cfs_rq->h_nr_running--;
3120 if (cfs_rq_throttled(cfs_rq))
3123 update_cfs_shares(cfs_rq);
3124 update_entity_load_avg(se, 1);
3129 update_rq_runnable_avg(rq, 1);
3135 /* Used instead of source_load when we know the type == 0 */
3136 static unsigned long weighted_cpuload(const int cpu)
3138 return cpu_rq(cpu)->load.weight;
3142 * Return a low guess at the load of a migration-source cpu weighted
3143 * according to the scheduling class and "nice" value.
3145 * We want to under-estimate the load of migration sources, to
3146 * balance conservatively.
3148 static unsigned long source_load(int cpu, int type)
3150 struct rq *rq = cpu_rq(cpu);
3151 unsigned long total = weighted_cpuload(cpu);
3153 if (type == 0 || !sched_feat(LB_BIAS))
3156 return min(rq->cpu_load[type-1], total);
3160 * Return a high guess at the load of a migration-target cpu weighted
3161 * according to the scheduling class and "nice" value.
3163 static unsigned long target_load(int cpu, int type)
3165 struct rq *rq = cpu_rq(cpu);
3166 unsigned long total = weighted_cpuload(cpu);
3168 if (type == 0 || !sched_feat(LB_BIAS))
3171 return max(rq->cpu_load[type-1], total);
3174 static unsigned long power_of(int cpu)
3176 return cpu_rq(cpu)->cpu_power;
3179 static unsigned long cpu_avg_load_per_task(int cpu)
3181 struct rq *rq = cpu_rq(cpu);
3182 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3185 return rq->load.weight / nr_running;
3191 static void task_waking_fair(struct task_struct *p)
3193 struct sched_entity *se = &p->se;
3194 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3197 #ifndef CONFIG_64BIT
3198 u64 min_vruntime_copy;
3201 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3203 min_vruntime = cfs_rq->min_vruntime;
3204 } while (min_vruntime != min_vruntime_copy);
3206 min_vruntime = cfs_rq->min_vruntime;
3209 se->vruntime -= min_vruntime;
3212 #ifdef CONFIG_FAIR_GROUP_SCHED
3214 * effective_load() calculates the load change as seen from the root_task_group
3216 * Adding load to a group doesn't make a group heavier, but can cause movement
3217 * of group shares between cpus. Assuming the shares were perfectly aligned one
3218 * can calculate the shift in shares.
3220 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3221 * on this @cpu and results in a total addition (subtraction) of @wg to the
3222 * total group weight.
3224 * Given a runqueue weight distribution (rw_i) we can compute a shares
3225 * distribution (s_i) using:
3227 * s_i = rw_i / \Sum rw_j (1)
3229 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3230 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3231 * shares distribution (s_i):
3233 * rw_i = { 2, 4, 1, 0 }
3234 * s_i = { 2/7, 4/7, 1/7, 0 }
3236 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3237 * task used to run on and the CPU the waker is running on), we need to
3238 * compute the effect of waking a task on either CPU and, in case of a sync
3239 * wakeup, compute the effect of the current task going to sleep.
3241 * So for a change of @wl to the local @cpu with an overall group weight change
3242 * of @wl we can compute the new shares distribution (s'_i) using:
3244 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3246 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3247 * differences in waking a task to CPU 0. The additional task changes the
3248 * weight and shares distributions like:
3250 * rw'_i = { 3, 4, 1, 0 }
3251 * s'_i = { 3/8, 4/8, 1/8, 0 }
3253 * We can then compute the difference in effective weight by using:
3255 * dw_i = S * (s'_i - s_i) (3)
3257 * Where 'S' is the group weight as seen by its parent.
3259 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3260 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3261 * 4/7) times the weight of the group.
3263 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3265 struct sched_entity *se = tg->se[cpu];
3267 if (!tg->parent) /* the trivial, non-cgroup case */
3270 for_each_sched_entity(se) {
3276 * W = @wg + \Sum rw_j
3278 W = wg + calc_tg_weight(tg, se->my_q);
3283 w = se->my_q->load.weight + wl;
3286 * wl = S * s'_i; see (2)
3289 wl = (w * tg->shares) / W;
3294 * Per the above, wl is the new se->load.weight value; since
3295 * those are clipped to [MIN_SHARES, ...) do so now. See
3296 * calc_cfs_shares().
3298 if (wl < MIN_SHARES)
3302 * wl = dw_i = S * (s'_i - s_i); see (3)
3304 wl -= se->load.weight;
3307 * Recursively apply this logic to all parent groups to compute
3308 * the final effective load change on the root group. Since
3309 * only the @tg group gets extra weight, all parent groups can
3310 * only redistribute existing shares. @wl is the shift in shares
3311 * resulting from this level per the above.
3320 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3321 unsigned long wl, unsigned long wg)
3328 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3330 s64 this_load, load;
3331 int idx, this_cpu, prev_cpu;
3332 unsigned long tl_per_task;
3333 struct task_group *tg;
3334 unsigned long weight;
3338 this_cpu = smp_processor_id();
3339 prev_cpu = task_cpu(p);
3340 load = source_load(prev_cpu, idx);
3341 this_load = target_load(this_cpu, idx);
3344 * If sync wakeup then subtract the (maximum possible)
3345 * effect of the currently running task from the load
3346 * of the current CPU:
3349 tg = task_group(current);
3350 weight = current->se.load.weight;
3352 this_load += effective_load(tg, this_cpu, -weight, -weight);
3353 load += effective_load(tg, prev_cpu, 0, -weight);
3357 weight = p->se.load.weight;
3360 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3361 * due to the sync cause above having dropped this_load to 0, we'll
3362 * always have an imbalance, but there's really nothing you can do
3363 * about that, so that's good too.
3365 * Otherwise check if either cpus are near enough in load to allow this
3366 * task to be woken on this_cpu.
3368 if (this_load > 0) {
3369 s64 this_eff_load, prev_eff_load;
3371 this_eff_load = 100;
3372 this_eff_load *= power_of(prev_cpu);
3373 this_eff_load *= this_load +
3374 effective_load(tg, this_cpu, weight, weight);
3376 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3377 prev_eff_load *= power_of(this_cpu);
3378 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3380 balanced = this_eff_load <= prev_eff_load;
3385 * If the currently running task will sleep within
3386 * a reasonable amount of time then attract this newly
3389 if (sync && balanced)
3392 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3393 tl_per_task = cpu_avg_load_per_task(this_cpu);
3396 (this_load <= load &&
3397 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3399 * This domain has SD_WAKE_AFFINE and
3400 * p is cache cold in this domain, and
3401 * there is no bad imbalance.
3403 schedstat_inc(sd, ttwu_move_affine);
3404 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3412 * find_idlest_group finds and returns the least busy CPU group within the
3415 static struct sched_group *
3416 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3417 int this_cpu, int load_idx)
3419 struct sched_group *idlest = NULL, *group = sd->groups;
3420 unsigned long min_load = ULONG_MAX, this_load = 0;
3421 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3424 unsigned long load, avg_load;
3428 /* Skip over this group if it has no CPUs allowed */
3429 if (!cpumask_intersects(sched_group_cpus(group),
3430 tsk_cpus_allowed(p)))
3433 local_group = cpumask_test_cpu(this_cpu,
3434 sched_group_cpus(group));
3436 /* Tally up the load of all CPUs in the group */
3439 for_each_cpu(i, sched_group_cpus(group)) {
3440 /* Bias balancing toward cpus of our domain */
3442 load = source_load(i, load_idx);
3444 load = target_load(i, load_idx);
3449 /* Adjust by relative CPU power of the group */
3450 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3453 this_load = avg_load;
3454 } else if (avg_load < min_load) {
3455 min_load = avg_load;
3458 } while (group = group->next, group != sd->groups);
3460 if (!idlest || 100*this_load < imbalance*min_load)
3466 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3469 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3471 unsigned long load, min_load = ULONG_MAX;
3475 /* Traverse only the allowed CPUs */
3476 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3477 load = weighted_cpuload(i);
3479 if (load < min_load || (load == min_load && i == this_cpu)) {
3489 * Try and locate an idle CPU in the sched_domain.
3491 static int select_idle_sibling(struct task_struct *p, int target)
3493 struct sched_domain *sd;
3494 struct sched_group *sg;
3495 int i = task_cpu(p);
3497 if (idle_cpu(target))
3501 * If the prevous cpu is cache affine and idle, don't be stupid.
3503 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3507 * Otherwise, iterate the domains and find an elegible idle cpu.
3509 sd = rcu_dereference(per_cpu(sd_llc, target));
3510 for_each_lower_domain(sd) {
3513 if (!cpumask_intersects(sched_group_cpus(sg),
3514 tsk_cpus_allowed(p)))
3517 for_each_cpu(i, sched_group_cpus(sg)) {
3518 if (i == target || !idle_cpu(i))
3522 target = cpumask_first_and(sched_group_cpus(sg),
3523 tsk_cpus_allowed(p));
3527 } while (sg != sd->groups);
3533 #ifdef CONFIG_SCHED_HMP
3535 * Heterogenous multiprocessor (HMP) optimizations
3537 * The cpu types are distinguished using a list of hmp_domains
3538 * which each represent one cpu type using a cpumask.
3539 * The list is assumed ordered by compute capacity with the
3540 * fastest domain first.
3542 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3543 static const int hmp_max_tasks = 5;
3545 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3547 #ifdef CONFIG_CPU_IDLE
3551 * In this version we have stopped using forced up migrations when we
3552 * detect that a task running on a little CPU should be moved to a bigger
3553 * CPU. In most cases, the bigger CPU is in a deep sleep state and a forced
3554 * migration means we stop the task immediately but need to wait for the
3555 * target CPU to wake up before we can restart the task which is being
3556 * moved. Instead, we now wake a big CPU with an IPI and ask it to pull
3557 * a task when ready. This allows the task to continue executing on its
3558 * current CPU, reducing the amount of time that the task is stalled for.
3562 * The keepalive timer is used as a way to keep a CPU engaged in an
3563 * idle pull operation out of idle while waiting for the source
3564 * CPU to stop and move the task. Ideally this would not be necessary
3565 * and we could impose a temporary zero-latency requirement on the
3566 * current CPU, but in the current QoS framework this will result in
3567 * all CPUs in the system being unable to enter idle states which is
3568 * not desirable. The timer does not perform any work when it expires.
3570 struct hmp_keepalive {
3572 ktime_t delay; /* if zero, no need for timer */
3573 struct hrtimer timer;
3575 DEFINE_PER_CPU(struct hmp_keepalive, hmp_cpu_keepalive);
3577 /* setup per-cpu keepalive timers */
3578 static enum hrtimer_restart hmp_cpu_keepalive_notify(struct hrtimer *hrtimer)
3580 return HRTIMER_NORESTART;
3584 * Work out if any of the idle states have an exit latency too high for us.
3585 * ns_delay is passed in containing the max we are willing to tolerate.
3586 * If there are none, set ns_delay to zero.
3587 * If there are any, set ns_delay to
3588 * ('target_residency of state with shortest too-big latency' - 1) * 1000.
3590 static void hmp_keepalive_delay(unsigned int *ns_delay)
3592 struct cpuidle_driver *drv;
3593 drv = cpuidle_driver_ref();
3595 unsigned int us_delay = UINT_MAX;
3596 unsigned int us_max_delay = *ns_delay / 1000;
3598 /* if cpuidle states are guaranteed to be sorted we
3599 * could stop at the first match.
3601 for (idx = 0; idx < drv->state_count; idx++) {
3602 if (drv->states[idx].exit_latency > us_max_delay &&
3603 drv->states[idx].target_residency < us_delay) {
3604 us_delay = drv->states[idx].target_residency;
3607 if (us_delay == UINT_MAX)
3608 *ns_delay = 0; /* no timer required */
3610 *ns_delay = 1000 * (us_delay - 1);
3612 cpuidle_driver_unref();
3615 static void hmp_cpu_keepalive_trigger(void)
3617 int cpu = smp_processor_id();
3618 struct hmp_keepalive *keepalive = &per_cpu(hmp_cpu_keepalive, cpu);
3619 if (!keepalive->init) {
3620 unsigned int ns_delay = 100000; /* tolerate 100usec delay */
3622 hrtimer_init(&keepalive->timer,
3623 CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
3624 keepalive->timer.function = hmp_cpu_keepalive_notify;
3626 hmp_keepalive_delay(&ns_delay);
3627 keepalive->delay = ns_to_ktime(ns_delay);
3628 keepalive->init = true;
3630 if (ktime_to_ns(keepalive->delay))
3631 hrtimer_start(&keepalive->timer,
3632 keepalive->delay, HRTIMER_MODE_REL_PINNED);
3635 static void hmp_cpu_keepalive_cancel(int cpu)
3637 struct hmp_keepalive *keepalive = &per_cpu(hmp_cpu_keepalive, cpu);
3638 if (keepalive->init)
3639 hrtimer_cancel(&keepalive->timer);
3641 #else /* !CONFIG_CPU_IDLE */
3642 static void hmp_cpu_keepalive_trigger(void)
3646 static void hmp_cpu_keepalive_cancel(int cpu)
3651 /* Setup hmp_domains */
3652 static int __init hmp_cpu_mask_setup(void)
3655 struct hmp_domain *domain;
3656 struct list_head *pos;
3659 pr_debug("Initializing HMP scheduler:\n");
3661 /* Initialize hmp_domains using platform code */
3662 arch_get_hmp_domains(&hmp_domains);
3663 if (list_empty(&hmp_domains)) {
3664 pr_debug("HMP domain list is empty!\n");
3668 /* Print hmp_domains */
3670 list_for_each(pos, &hmp_domains) {
3671 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3672 cpulist_scnprintf(buf, 64, &domain->possible_cpus);
3673 pr_debug(" HMP domain %d: %s\n", dc, buf);
3675 for_each_cpu_mask(cpu, domain->possible_cpus) {
3676 per_cpu(hmp_cpu_domain, cpu) = domain;
3684 static struct hmp_domain *hmp_get_hmp_domain_for_cpu(int cpu)
3686 struct hmp_domain *domain;
3687 struct list_head *pos;
3689 list_for_each(pos, &hmp_domains) {
3690 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3691 if(cpumask_test_cpu(cpu, &domain->possible_cpus))
3697 static void hmp_online_cpu(int cpu)
3699 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3702 cpumask_set_cpu(cpu, &domain->cpus);
3705 static void hmp_offline_cpu(int cpu)
3707 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3710 cpumask_clear_cpu(cpu, &domain->cpus);
3712 hmp_cpu_keepalive_cancel(cpu);
3715 * Needed to determine heaviest tasks etc.
3717 static inline unsigned int hmp_cpu_is_fastest(int cpu);
3718 static inline unsigned int hmp_cpu_is_slowest(int cpu);
3719 static inline struct hmp_domain *hmp_slower_domain(int cpu);
3720 static inline struct hmp_domain *hmp_faster_domain(int cpu);
3722 /* must hold runqueue lock for queue se is currently on */
3723 static struct sched_entity *hmp_get_heaviest_task(
3724 struct sched_entity *se, int target_cpu)
3726 int num_tasks = hmp_max_tasks;
3727 struct sched_entity *max_se = se;
3728 unsigned long int max_ratio = se->avg.load_avg_ratio;
3729 const struct cpumask *hmp_target_mask = NULL;
3730 struct hmp_domain *hmp;
3732 if (hmp_cpu_is_fastest(cpu_of(se->cfs_rq->rq)))
3735 hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq));
3736 hmp_target_mask = &hmp->cpus;
3737 if (target_cpu >= 0) {
3738 /* idle_balance gets run on a CPU while
3739 * it is in the middle of being hotplugged
3740 * out. Bail early in that case.
3742 if(!cpumask_test_cpu(target_cpu, hmp_target_mask))
3744 hmp_target_mask = cpumask_of(target_cpu);
3746 /* The currently running task is not on the runqueue */
3747 se = __pick_first_entity(cfs_rq_of(se));
3749 while (num_tasks && se) {
3750 if (entity_is_task(se) &&
3751 se->avg.load_avg_ratio > max_ratio &&
3752 cpumask_intersects(hmp_target_mask,
3753 tsk_cpus_allowed(task_of(se)))) {
3755 max_ratio = se->avg.load_avg_ratio;
3757 se = __pick_next_entity(se);
3763 static struct sched_entity *hmp_get_lightest_task(
3764 struct sched_entity *se, int migrate_down)
3766 int num_tasks = hmp_max_tasks;
3767 struct sched_entity *min_se = se;
3768 unsigned long int min_ratio = se->avg.load_avg_ratio;
3769 const struct cpumask *hmp_target_mask = NULL;
3772 struct hmp_domain *hmp;
3773 if (hmp_cpu_is_slowest(cpu_of(se->cfs_rq->rq)))
3775 hmp = hmp_slower_domain(cpu_of(se->cfs_rq->rq));
3776 hmp_target_mask = &hmp->cpus;
3778 /* The currently running task is not on the runqueue */
3779 se = __pick_first_entity(cfs_rq_of(se));
3781 while (num_tasks && se) {
3782 if (entity_is_task(se) &&
3783 (se->avg.load_avg_ratio < min_ratio &&
3785 cpumask_intersects(hmp_target_mask,
3786 tsk_cpus_allowed(task_of(se))))) {
3788 min_ratio = se->avg.load_avg_ratio;
3790 se = __pick_next_entity(se);
3797 * Migration thresholds should be in the range [0..1023]
3798 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3799 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3801 * hmp_up_prio: Only up migrate task with high priority (<hmp_up_prio)
3802 * hmp_next_up_threshold: Delay before next up migration (1024 ~= 1 ms)
3803 * hmp_next_down_threshold: Delay before next down migration (1024 ~= 1 ms)
3805 * Small Task Packing:
3806 * We can choose to fill the littlest CPUs in an HMP system rather than
3807 * the typical spreading mechanic. This behavior is controllable using
3809 * hmp_packing_enabled: runtime control over pack/spread
3810 * hmp_full_threshold: Consider a CPU with this much unweighted load full
3812 unsigned int hmp_up_threshold = 700;
3813 unsigned int hmp_down_threshold = 512;
3814 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
3815 unsigned int hmp_up_prio = NICE_TO_PRIO(CONFIG_SCHED_HMP_PRIO_FILTER_VAL);
3817 unsigned int hmp_next_up_threshold = 4096;
3818 unsigned int hmp_next_down_threshold = 4096;
3820 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3822 * Set the default packing threshold to try to keep little
3823 * CPUs at no more than 80% of their maximum frequency if only
3824 * packing a small number of small tasks. Bigger tasks will
3825 * raise frequency as normal.
3826 * In order to pack a task onto a CPU, the sum of the
3827 * unweighted runnable_avg load of existing tasks plus the
3828 * load of the new task must be less than hmp_full_threshold.
3830 * This works in conjunction with frequency-invariant load
3831 * and DVFS governors. Since most DVFS governors aim for 80%
3832 * utilisation, we arrive at (0.8*0.8*(max_load=1024))=655
3833 * and use a value slightly lower to give a little headroom
3835 * Note that the most efficient frequency is different for
3836 * each system so /sys/kernel/hmp/packing_limit should be
3837 * configured at runtime for any given platform to achieve
3838 * optimal energy usage. Some systems may not benefit from
3839 * packing, so this feature can also be disabled at runtime
3840 * with /sys/kernel/hmp/packing_enable
3842 unsigned int hmp_packing_enabled = 1;
3843 unsigned int hmp_full_threshold = 650;
3846 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se);
3847 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3848 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
3849 int *min_cpu, struct cpumask *affinity);
3851 static inline struct hmp_domain *hmp_smallest_domain(void)
3853 return list_entry(hmp_domains.prev, struct hmp_domain, hmp_domains);
3856 /* Check if cpu is in fastest hmp_domain */
3857 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3859 struct list_head *pos;
3861 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3862 return pos == hmp_domains.next;
3865 /* Check if cpu is in slowest hmp_domain */
3866 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3868 struct list_head *pos;
3870 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3871 return list_is_last(pos, &hmp_domains);
3874 /* Next (slower) hmp_domain relative to cpu */
3875 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3877 struct list_head *pos;
3879 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3880 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3883 /* Previous (faster) hmp_domain relative to cpu */
3884 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3886 struct list_head *pos;
3888 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3889 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3893 * Selects a cpu in previous (faster) hmp_domain
3895 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3898 int lowest_cpu=NR_CPUS;
3899 __always_unused int lowest_ratio;
3900 struct hmp_domain *hmp;
3902 if (hmp_cpu_is_fastest(cpu))
3903 hmp = hmp_cpu_domain(cpu);
3905 hmp = hmp_faster_domain(cpu);
3907 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3908 tsk_cpus_allowed(tsk));
3914 * Selects a cpu in next (slower) hmp_domain
3915 * Note that cpumask_any_and() returns the first cpu in the cpumask
3917 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3920 int lowest_cpu=NR_CPUS;
3921 struct hmp_domain *hmp;
3922 __always_unused int lowest_ratio;
3924 if (hmp_cpu_is_slowest(cpu))
3925 hmp = hmp_cpu_domain(cpu);
3927 hmp = hmp_slower_domain(cpu);
3929 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3930 tsk_cpus_allowed(tsk));
3934 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3936 * Select the 'best' candidate little CPU to wake up on.
3937 * Implements a packing strategy which examines CPU in
3938 * logical CPU order, and selects the first which will
3939 * be loaded less than hmp_full_threshold according to
3940 * the sum of the tracked load of the runqueue and the task.
3942 static inline unsigned int hmp_best_little_cpu(struct task_struct *tsk,
3945 unsigned long estimated_load;
3946 struct hmp_domain *hmp;
3947 struct sched_avg *avg;
3948 struct cpumask allowed_hmp_cpus;
3950 if(!hmp_packing_enabled ||
3951 tsk->se.avg.load_avg_ratio > ((NICE_0_LOAD * 90)/100))
3952 return hmp_select_slower_cpu(tsk, cpu);
3954 if (hmp_cpu_is_slowest(cpu))
3955 hmp = hmp_cpu_domain(cpu);
3957 hmp = hmp_slower_domain(cpu);
3959 /* respect affinity */
3960 cpumask_and(&allowed_hmp_cpus, &hmp->cpus,
3961 tsk_cpus_allowed(tsk));
3963 for_each_cpu_mask(tmp_cpu, allowed_hmp_cpus) {
3964 avg = &cpu_rq(tmp_cpu)->avg;
3965 /* estimate new rq load if we add this task */
3966 estimated_load = avg->load_avg_ratio +
3967 tsk->se.avg.load_avg_ratio;
3968 if (estimated_load <= hmp_full_threshold) {
3973 /* if no match was found, the task uses the initial value */
3977 static inline void hmp_next_up_delay(struct sched_entity *se, int cpu)
3979 /* hack - always use clock from first online CPU */
3980 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3981 se->avg.hmp_last_up_migration = now;
3982 se->avg.hmp_last_down_migration = 0;
3983 cpu_rq(cpu)->avg.hmp_last_up_migration = now;
3984 cpu_rq(cpu)->avg.hmp_last_down_migration = 0;
3987 static inline void hmp_next_down_delay(struct sched_entity *se, int cpu)
3989 /* hack - always use clock from first online CPU */
3990 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3991 se->avg.hmp_last_down_migration = now;
3992 se->avg.hmp_last_up_migration = 0;
3993 cpu_rq(cpu)->avg.hmp_last_down_migration = now;
3994 cpu_rq(cpu)->avg.hmp_last_up_migration = 0;
3998 * Heterogenous multiprocessor (HMP) optimizations
4000 * These functions allow to change the growing speed of the load_avg_ratio
4001 * by default it goes from 0 to 0.5 in LOAD_AVG_PERIOD = 32ms
4002 * This can now be changed with /sys/kernel/hmp/load_avg_period_ms.
4004 * These functions also allow to change the up and down threshold of HMP
4005 * using /sys/kernel/hmp/{up,down}_threshold.
4006 * Both must be between 0 and 1023. The threshold that is compared
4007 * to the load_avg_ratio is up_threshold/1024 and down_threshold/1024.
4009 * For instance, if load_avg_period = 64 and up_threshold = 512, an idle
4010 * task with a load of 0 will reach the threshold after 64ms of busy loop.
4012 * Changing load_avg_periods_ms has the same effect than changing the
4013 * default scaling factor Y=1002/1024 in the load_avg_ratio computation to
4014 * (1002/1024.0)^(LOAD_AVG_PERIOD/load_avg_period_ms), but the last one
4015 * could trigger overflows.
4016 * For instance, with Y = 1023/1024 in __update_task_entity_contrib()
4017 * "contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);"
4018 * could be overflowed for a weight > 2^12 even is the load_avg_contrib
4019 * should still be a 32bits result. This would not happen by multiplicating
4020 * delta time by 1/22 and setting load_avg_period_ms = 706.
4024 * By scaling the delta time it end-up increasing or decrease the
4025 * growing speed of the per entity load_avg_ratio
4026 * The scale factor hmp_data.multiplier is a fixed point
4027 * number: (32-HMP_VARIABLE_SCALE_SHIFT).HMP_VARIABLE_SCALE_SHIFT
4029 static inline u64 hmp_variable_scale_convert(u64 delta)
4031 #ifdef CONFIG_HMP_VARIABLE_SCALE
4032 u64 high = delta >> 32ULL;
4033 u64 low = delta & 0xffffffffULL;
4034 low *= hmp_data.multiplier;
4035 high *= hmp_data.multiplier;
4036 return (low >> HMP_VARIABLE_SCALE_SHIFT)
4037 + (high << (32ULL - HMP_VARIABLE_SCALE_SHIFT));
4043 static ssize_t hmp_show(struct kobject *kobj,
4044 struct attribute *attr, char *buf)
4046 struct hmp_global_attr *hmp_attr =
4047 container_of(attr, struct hmp_global_attr, attr);
4050 if (hmp_attr->to_sysfs_text != NULL)
4051 return hmp_attr->to_sysfs_text(buf, PAGE_SIZE);
4053 temp = *(hmp_attr->value);
4054 if (hmp_attr->to_sysfs != NULL)
4055 temp = hmp_attr->to_sysfs(temp);
4057 return (ssize_t)sprintf(buf, "%d\n", temp);
4060 static ssize_t hmp_store(struct kobject *a, struct attribute *attr,
4061 const char *buf, size_t count)
4064 ssize_t ret = count;
4065 struct hmp_global_attr *hmp_attr =
4066 container_of(attr, struct hmp_global_attr, attr);
4067 char *str = vmalloc(count + 1);
4070 memcpy(str, buf, count);
4072 if (sscanf(str, "%d", &temp) < 1)
4075 if (hmp_attr->from_sysfs != NULL)
4076 temp = hmp_attr->from_sysfs(temp);
4080 *(hmp_attr->value) = temp;
4086 static ssize_t hmp_print_domains(char *outbuf, int outbufsize)
4089 const char nospace[] = "%s", space[] = " %s";
4090 const char *fmt = nospace;
4091 struct hmp_domain *domain;
4092 struct list_head *pos;
4094 list_for_each(pos, &hmp_domains) {
4095 domain = list_entry(pos, struct hmp_domain, hmp_domains);
4096 if (cpumask_scnprintf(buf, 64, &domain->possible_cpus)) {
4097 outpos += sprintf(outbuf+outpos, fmt, buf);
4101 strcat(outbuf, "\n");
4105 #ifdef CONFIG_HMP_VARIABLE_SCALE
4106 static int hmp_period_tofrom_sysfs(int value)
4108 return (LOAD_AVG_PERIOD << HMP_VARIABLE_SCALE_SHIFT) / value;
4111 /* max value for threshold is 1024 */
4112 static int hmp_theshold_from_sysfs(int value)
4118 #if defined(CONFIG_SCHED_HMP_LITTLE_PACKING) || \
4119 defined(CONFIG_HMP_FREQUENCY_INVARIANT_SCALE)
4120 /* toggle control is only 0,1 off/on */
4121 static int hmp_toggle_from_sysfs(int value)
4123 if (value < 0 || value > 1)
4128 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4129 /* packing value must be non-negative */
4130 static int hmp_packing_from_sysfs(int value)
4137 static void hmp_attr_add(
4140 int (*to_sysfs)(int),
4141 int (*from_sysfs)(int),
4142 ssize_t (*to_sysfs_text)(char *, int),
4146 while (hmp_data.attributes[i] != NULL) {
4148 if (i >= HMP_DATA_SYSFS_MAX)
4152 hmp_data.attr[i].attr.mode = mode;
4154 hmp_data.attr[i].attr.mode = 0644;
4155 hmp_data.attr[i].show = hmp_show;
4156 hmp_data.attr[i].store = hmp_store;
4157 hmp_data.attr[i].attr.name = name;
4158 hmp_data.attr[i].value = value;
4159 hmp_data.attr[i].to_sysfs = to_sysfs;
4160 hmp_data.attr[i].from_sysfs = from_sysfs;
4161 hmp_data.attr[i].to_sysfs_text = to_sysfs_text;
4162 hmp_data.attributes[i] = &hmp_data.attr[i].attr;
4163 hmp_data.attributes[i + 1] = NULL;
4166 static int hmp_attr_init(void)
4169 memset(&hmp_data, sizeof(hmp_data), 0);
4170 hmp_attr_add("hmp_domains",
4176 hmp_attr_add("up_threshold",
4179 hmp_theshold_from_sysfs,
4182 hmp_attr_add("down_threshold",
4183 &hmp_down_threshold,
4185 hmp_theshold_from_sysfs,
4188 #ifdef CONFIG_HMP_VARIABLE_SCALE
4189 /* by default load_avg_period_ms == LOAD_AVG_PERIOD
4192 hmp_data.multiplier = hmp_period_tofrom_sysfs(LOAD_AVG_PERIOD);
4193 hmp_attr_add("load_avg_period_ms",
4194 &hmp_data.multiplier,
4195 hmp_period_tofrom_sysfs,
4196 hmp_period_tofrom_sysfs,
4200 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
4201 /* default frequency-invariant scaling ON */
4202 hmp_data.freqinvar_load_scale_enabled = 1;
4203 hmp_attr_add("frequency_invariant_load_scale",
4204 &hmp_data.freqinvar_load_scale_enabled,
4206 hmp_toggle_from_sysfs,
4210 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4211 hmp_attr_add("packing_enable",
4212 &hmp_packing_enabled,
4214 hmp_toggle_from_sysfs,
4217 hmp_attr_add("packing_limit",
4218 &hmp_full_threshold,
4220 hmp_packing_from_sysfs,
4224 hmp_data.attr_group.name = "hmp";
4225 hmp_data.attr_group.attrs = hmp_data.attributes;
4226 ret = sysfs_create_group(kernel_kobj,
4227 &hmp_data.attr_group);
4230 late_initcall(hmp_attr_init);
4232 * return the load of the lowest-loaded CPU in a given HMP domain
4233 * min_cpu optionally points to an int to receive the CPU.
4234 * affinity optionally points to a cpumask containing the
4235 * CPUs to be considered. note:
4236 * + min_cpu = NR_CPUS only if no CPUs are in the set of
4237 * affinity && hmp_domain cpus
4238 * + min_cpu will always otherwise equal one of the CPUs in
4240 * + when more than one CPU has the same load, the one which
4241 * is least-recently-disturbed by an HMP migration will be
4243 * + if all CPUs are equally loaded or idle and the times are
4244 * all the same, the first in the set will be used
4245 * + if affinity is not set, cpu_online_mask is used
4247 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
4248 int *min_cpu, struct cpumask *affinity)
4251 int min_cpu_runnable_temp = NR_CPUS;
4252 u64 min_target_last_migration = ULLONG_MAX;
4253 u64 curr_last_migration;
4254 unsigned long min_runnable_load = INT_MAX;
4255 unsigned long contrib;
4256 struct sched_avg *avg;
4257 struct cpumask temp_cpumask;
4259 * only look at CPUs allowed if specified,
4260 * otherwise look at all online CPUs in the
4263 cpumask_and(&temp_cpumask, &hmpd->cpus, affinity ? affinity : cpu_online_mask);
4265 for_each_cpu_mask(cpu, temp_cpumask) {
4266 avg = &cpu_rq(cpu)->avg;
4267 /* used for both up and down migration */
4268 curr_last_migration = avg->hmp_last_up_migration ?
4269 avg->hmp_last_up_migration : avg->hmp_last_down_migration;
4271 contrib = avg->load_avg_ratio;
4273 * Consider a runqueue completely busy if there is any load
4274 * on it. Definitely not the best for overall fairness, but
4275 * does well in typical Android use cases.
4280 if ((contrib < min_runnable_load) ||
4281 (contrib == min_runnable_load &&
4282 curr_last_migration < min_target_last_migration)) {
4284 * if the load is the same target the CPU with
4285 * the longest time since a migration.
4286 * This is to spread migration load between
4287 * members of a domain more evenly when the
4288 * domain is fully loaded
4290 min_runnable_load = contrib;
4291 min_cpu_runnable_temp = cpu;
4292 min_target_last_migration = curr_last_migration;
4297 *min_cpu = min_cpu_runnable_temp;
4299 return min_runnable_load;
4303 * Calculate the task starvation
4304 * This is the ratio of actually running time vs. runnable time.
4305 * If the two are equal the task is getting the cpu time it needs or
4306 * it is alone on the cpu and the cpu is fully utilized.
4308 static inline unsigned int hmp_task_starvation(struct sched_entity *se)
4312 starvation = se->avg.usage_avg_sum * scale_load_down(NICE_0_LOAD);
4313 starvation /= (se->avg.runnable_avg_sum + 1);
4315 return scale_load(starvation);
4318 static inline unsigned int hmp_offload_down(int cpu, struct sched_entity *se)
4321 int dest_cpu = NR_CPUS;
4323 if (hmp_cpu_is_slowest(cpu))
4326 /* Is there an idle CPU in the current domain */
4327 min_usage = hmp_domain_min_load(hmp_cpu_domain(cpu), NULL, NULL);
4328 if (min_usage == 0) {
4329 trace_sched_hmp_offload_abort(cpu, min_usage, "load");
4333 /* Is the task alone on the cpu? */
4334 if (cpu_rq(cpu)->cfs.h_nr_running < 2) {
4335 trace_sched_hmp_offload_abort(cpu,
4336 cpu_rq(cpu)->cfs.h_nr_running, "nr_running");
4340 /* Is the task actually starving? */
4341 /* >=25% ratio running/runnable = starving */
4342 if (hmp_task_starvation(se) > 768) {
4343 trace_sched_hmp_offload_abort(cpu, hmp_task_starvation(se),
4348 /* Does the slower domain have any idle CPUs? */
4349 min_usage = hmp_domain_min_load(hmp_slower_domain(cpu), &dest_cpu,
4350 tsk_cpus_allowed(task_of(se)));
4352 if (min_usage == 0) {
4353 trace_sched_hmp_offload_succeed(cpu, dest_cpu);
4356 trace_sched_hmp_offload_abort(cpu,min_usage,"slowdomain");
4359 #endif /* CONFIG_SCHED_HMP */
4362 * sched_balance_self: balance the current task (running on cpu) in domains
4363 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4366 * Balance, ie. select the least loaded group.
4368 * Returns the target CPU number, or the same CPU if no balancing is needed.
4370 * preempt must be disabled.
4373 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
4375 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4376 int cpu = smp_processor_id();
4377 int prev_cpu = task_cpu(p);
4379 int want_affine = 0;
4380 int sync = wake_flags & WF_SYNC;
4382 if (p->nr_cpus_allowed == 1)
4385 #ifdef CONFIG_SCHED_HMP
4386 /* always put non-kernel forking tasks on a big domain */
4387 if (p->mm && (sd_flag & SD_BALANCE_FORK)) {
4388 new_cpu = hmp_select_faster_cpu(p, prev_cpu);
4389 if (new_cpu != NR_CPUS) {
4390 hmp_next_up_delay(&p->se, new_cpu);
4393 /* failed to perform HMP fork balance, use normal balance */
4398 if (sd_flag & SD_BALANCE_WAKE) {
4399 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4405 for_each_domain(cpu, tmp) {
4406 if (!(tmp->flags & SD_LOAD_BALANCE))
4410 * If both cpu and prev_cpu are part of this domain,
4411 * cpu is a valid SD_WAKE_AFFINE target.
4413 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4414 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4419 if (tmp->flags & sd_flag)
4424 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4427 new_cpu = select_idle_sibling(p, prev_cpu);
4432 int load_idx = sd->forkexec_idx;
4433 struct sched_group *group;
4436 if (!(sd->flags & sd_flag)) {
4441 if (sd_flag & SD_BALANCE_WAKE)
4442 load_idx = sd->wake_idx;
4444 group = find_idlest_group(sd, p, cpu, load_idx);
4450 new_cpu = find_idlest_cpu(group, p, cpu);
4451 if (new_cpu == -1 || new_cpu == cpu) {
4452 /* Now try balancing at a lower domain level of cpu */
4457 /* Now try balancing at a lower domain level of new_cpu */
4459 weight = sd->span_weight;
4461 for_each_domain(cpu, tmp) {
4462 if (weight <= tmp->span_weight)
4464 if (tmp->flags & sd_flag)
4467 /* while loop will break here if sd == NULL */
4472 #ifdef CONFIG_SCHED_HMP
4473 prev_cpu = task_cpu(p);
4475 if (hmp_up_migration(prev_cpu, &new_cpu, &p->se)) {
4476 hmp_next_up_delay(&p->se, new_cpu);
4477 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4480 if (hmp_down_migration(prev_cpu, &p->se)) {
4481 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4482 new_cpu = hmp_best_little_cpu(p, prev_cpu);
4484 new_cpu = hmp_select_slower_cpu(p, prev_cpu);
4486 if (new_cpu != prev_cpu) {
4487 hmp_next_down_delay(&p->se, new_cpu);
4488 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4492 /* Make sure that the task stays in its previous hmp domain */
4493 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
4501 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
4502 * removed when useful for applications beyond shares distribution (e.g.
4505 #ifdef CONFIG_FAIR_GROUP_SCHED
4507 #ifdef CONFIG_NO_HZ_COMMON
4508 static int nohz_test_cpu(int cpu);
4510 static inline int nohz_test_cpu(int cpu)
4517 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4518 * cfs_rq_of(p) references at time of call are still valid and identify the
4519 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4520 * other assumptions, including the state of rq->lock, should be made.
4523 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4525 struct sched_entity *se = &p->se;
4526 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4529 * Load tracking: accumulate removed load so that it can be processed
4530 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4531 * to blocked load iff they have a positive decay-count. It can never
4532 * be negative here since on-rq tasks have decay-count == 0.
4534 if (se->avg.decay_count) {
4536 * If we migrate a sleeping task away from a CPU
4537 * which has the tick stopped, then both the clock_task
4538 * and decay_counter will be out of date for that CPU
4539 * and we will not decay load correctly.
4541 if (!se->on_rq && nohz_test_cpu(task_cpu(p))) {
4542 struct rq *rq = cpu_rq(task_cpu(p));
4543 unsigned long flags;
4545 * Current CPU cannot be holding rq->lock in this
4546 * circumstance, but another might be. We must hold
4547 * rq->lock before we go poking around in its clocks
4549 raw_spin_lock_irqsave(&rq->lock, flags);
4550 update_rq_clock(rq);
4551 update_cfs_rq_blocked_load(cfs_rq, 0);
4552 raw_spin_unlock_irqrestore(&rq->lock, flags);
4554 se->avg.decay_count = -__synchronize_entity_decay(se);
4555 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
4559 #endif /* CONFIG_SMP */
4561 static unsigned long
4562 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4564 unsigned long gran = sysctl_sched_wakeup_granularity;
4567 * Since its curr running now, convert the gran from real-time
4568 * to virtual-time in his units.
4570 * By using 'se' instead of 'curr' we penalize light tasks, so
4571 * they get preempted easier. That is, if 'se' < 'curr' then
4572 * the resulting gran will be larger, therefore penalizing the
4573 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4574 * be smaller, again penalizing the lighter task.
4576 * This is especially important for buddies when the leftmost
4577 * task is higher priority than the buddy.
4579 return calc_delta_fair(gran, se);
4583 * Should 'se' preempt 'curr'.
4597 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4599 s64 gran, vdiff = curr->vruntime - se->vruntime;
4604 gran = wakeup_gran(curr, se);
4611 static void set_last_buddy(struct sched_entity *se)
4613 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4616 for_each_sched_entity(se)
4617 cfs_rq_of(se)->last = se;
4620 static void set_next_buddy(struct sched_entity *se)
4622 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4625 for_each_sched_entity(se)
4626 cfs_rq_of(se)->next = se;
4629 static void set_skip_buddy(struct sched_entity *se)
4631 for_each_sched_entity(se)
4632 cfs_rq_of(se)->skip = se;
4636 * Preempt the current task with a newly woken task if needed:
4638 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4640 struct task_struct *curr = rq->curr;
4641 struct sched_entity *se = &curr->se, *pse = &p->se;
4642 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4643 int scale = cfs_rq->nr_running >= sched_nr_latency;
4644 int next_buddy_marked = 0;
4646 if (unlikely(se == pse))
4650 * This is possible from callers such as move_task(), in which we
4651 * unconditionally check_prempt_curr() after an enqueue (which may have
4652 * lead to a throttle). This both saves work and prevents false
4653 * next-buddy nomination below.
4655 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4658 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4659 set_next_buddy(pse);
4660 next_buddy_marked = 1;
4664 * We can come here with TIF_NEED_RESCHED already set from new task
4667 * Note: this also catches the edge-case of curr being in a throttled
4668 * group (e.g. via set_curr_task), since update_curr() (in the
4669 * enqueue of curr) will have resulted in resched being set. This
4670 * prevents us from potentially nominating it as a false LAST_BUDDY
4673 if (test_tsk_need_resched(curr))
4676 /* Idle tasks are by definition preempted by non-idle tasks. */
4677 if (unlikely(curr->policy == SCHED_IDLE) &&
4678 likely(p->policy != SCHED_IDLE))
4682 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4683 * is driven by the tick):
4685 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4688 find_matching_se(&se, &pse);
4689 update_curr(cfs_rq_of(se));
4691 if (wakeup_preempt_entity(se, pse) == 1) {
4693 * Bias pick_next to pick the sched entity that is
4694 * triggering this preemption.
4696 if (!next_buddy_marked)
4697 set_next_buddy(pse);
4706 * Only set the backward buddy when the current task is still
4707 * on the rq. This can happen when a wakeup gets interleaved
4708 * with schedule on the ->pre_schedule() or idle_balance()
4709 * point, either of which can * drop the rq lock.
4711 * Also, during early boot the idle thread is in the fair class,
4712 * for obvious reasons its a bad idea to schedule back to it.
4714 if (unlikely(!se->on_rq || curr == rq->idle))
4717 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4721 static struct task_struct *pick_next_task_fair(struct rq *rq)
4723 struct task_struct *p;
4724 struct cfs_rq *cfs_rq = &rq->cfs;
4725 struct sched_entity *se;
4727 if (!cfs_rq->nr_running)
4731 se = pick_next_entity(cfs_rq);
4732 set_next_entity(cfs_rq, se);
4733 cfs_rq = group_cfs_rq(se);
4737 if (hrtick_enabled(rq))
4738 hrtick_start_fair(rq, p);
4744 * Account for a descheduled task:
4746 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4748 struct sched_entity *se = &prev->se;
4749 struct cfs_rq *cfs_rq;
4751 for_each_sched_entity(se) {
4752 cfs_rq = cfs_rq_of(se);
4753 put_prev_entity(cfs_rq, se);
4758 * sched_yield() is very simple
4760 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4762 static void yield_task_fair(struct rq *rq)
4764 struct task_struct *curr = rq->curr;
4765 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4766 struct sched_entity *se = &curr->se;
4769 * Are we the only task in the tree?
4771 if (unlikely(rq->nr_running == 1))
4774 clear_buddies(cfs_rq, se);
4776 if (curr->policy != SCHED_BATCH) {
4777 update_rq_clock(rq);
4779 * Update run-time statistics of the 'current'.
4781 update_curr(cfs_rq);
4783 * Tell update_rq_clock() that we've just updated,
4784 * so we don't do microscopic update in schedule()
4785 * and double the fastpath cost.
4787 rq->skip_clock_update = 1;
4793 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4795 struct sched_entity *se = &p->se;
4797 /* throttled hierarchies are not runnable */
4798 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4801 /* Tell the scheduler that we'd really like pse to run next. */
4804 yield_task_fair(rq);
4810 /**************************************************
4811 * Fair scheduling class load-balancing methods.
4815 * The purpose of load-balancing is to achieve the same basic fairness the
4816 * per-cpu scheduler provides, namely provide a proportional amount of compute
4817 * time to each task. This is expressed in the following equation:
4819 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4821 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4822 * W_i,0 is defined as:
4824 * W_i,0 = \Sum_j w_i,j (2)
4826 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4827 * is derived from the nice value as per prio_to_weight[].
4829 * The weight average is an exponential decay average of the instantaneous
4832 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4834 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4835 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4836 * can also include other factors [XXX].
4838 * To achieve this balance we define a measure of imbalance which follows
4839 * directly from (1):
4841 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4843 * We them move tasks around to minimize the imbalance. In the continuous
4844 * function space it is obvious this converges, in the discrete case we get
4845 * a few fun cases generally called infeasible weight scenarios.
4848 * - infeasible weights;
4849 * - local vs global optima in the discrete case. ]
4854 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4855 * for all i,j solution, we create a tree of cpus that follows the hardware
4856 * topology where each level pairs two lower groups (or better). This results
4857 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4858 * tree to only the first of the previous level and we decrease the frequency
4859 * of load-balance at each level inv. proportional to the number of cpus in
4865 * \Sum { --- * --- * 2^i } = O(n) (5)
4867 * `- size of each group
4868 * | | `- number of cpus doing load-balance
4870 * `- sum over all levels
4872 * Coupled with a limit on how many tasks we can migrate every balance pass,
4873 * this makes (5) the runtime complexity of the balancer.
4875 * An important property here is that each CPU is still (indirectly) connected
4876 * to every other cpu in at most O(log n) steps:
4878 * The adjacency matrix of the resulting graph is given by:
4881 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4884 * And you'll find that:
4886 * A^(log_2 n)_i,j != 0 for all i,j (7)
4888 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4889 * The task movement gives a factor of O(m), giving a convergence complexity
4892 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4897 * In order to avoid CPUs going idle while there's still work to do, new idle
4898 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4899 * tree itself instead of relying on other CPUs to bring it work.
4901 * This adds some complexity to both (5) and (8) but it reduces the total idle
4909 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4912 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4917 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4919 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4921 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4924 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4925 * rewrite all of this once again.]
4928 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4930 #define LBF_ALL_PINNED 0x01
4931 #define LBF_NEED_BREAK 0x02
4932 #define LBF_SOME_PINNED 0x04
4935 struct sched_domain *sd;
4943 struct cpumask *dst_grpmask;
4945 enum cpu_idle_type idle;
4947 /* The set of CPUs under consideration for load-balancing */
4948 struct cpumask *cpus;
4953 unsigned int loop_break;
4954 unsigned int loop_max;
4958 * move_task - move a task from one runqueue to another runqueue.
4959 * Both runqueues must be locked.
4961 static void move_task(struct task_struct *p, struct lb_env *env)
4963 deactivate_task(env->src_rq, p, 0);
4964 set_task_cpu(p, env->dst_cpu);
4965 activate_task(env->dst_rq, p, 0);
4966 check_preempt_curr(env->dst_rq, p, 0);
4970 * Is this task likely cache-hot:
4973 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4977 if (p->sched_class != &fair_sched_class)
4980 if (unlikely(p->policy == SCHED_IDLE))
4984 * Buddy candidates are cache hot:
4986 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4987 (&p->se == cfs_rq_of(&p->se)->next ||
4988 &p->se == cfs_rq_of(&p->se)->last))
4991 if (sysctl_sched_migration_cost == -1)
4993 if (sysctl_sched_migration_cost == 0)
4996 delta = now - p->se.exec_start;
4998 return delta < (s64)sysctl_sched_migration_cost;
5002 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5005 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5007 int tsk_cache_hot = 0;
5009 * We do not migrate tasks that are:
5010 * 1) throttled_lb_pair, or
5011 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5012 * 3) running (obviously), or
5013 * 4) are cache-hot on their current CPU.
5015 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5018 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5021 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5024 * Remember if this task can be migrated to any other cpu in
5025 * our sched_group. We may want to revisit it if we couldn't
5026 * meet load balance goals by pulling other tasks on src_cpu.
5028 * Also avoid computing new_dst_cpu if we have already computed
5029 * one in current iteration.
5031 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
5034 /* Prevent to re-select dst_cpu via env's cpus */
5035 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5036 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5037 env->flags |= LBF_SOME_PINNED;
5038 env->new_dst_cpu = cpu;
5046 /* Record that we found atleast one task that could run on dst_cpu */
5047 env->flags &= ~LBF_ALL_PINNED;
5049 if (task_running(env->src_rq, p)) {
5050 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5055 * Aggressive migration if:
5056 * 1) task is cache cold, or
5057 * 2) too many balance attempts have failed.
5059 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
5060 if (!tsk_cache_hot ||
5061 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5063 if (tsk_cache_hot) {
5064 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5065 schedstat_inc(p, se.statistics.nr_forced_migrations);
5071 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5076 * move_one_task tries to move exactly one task from busiest to this_rq, as
5077 * part of active balancing operations within "domain".
5078 * Returns 1 if successful and 0 otherwise.
5080 * Called with both runqueues locked.
5082 static int move_one_task(struct lb_env *env)
5084 struct task_struct *p, *n;
5086 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5087 if (!can_migrate_task(p, env))
5092 * Right now, this is only the second place move_task()
5093 * is called, so we can safely collect move_task()
5094 * stats here rather than inside move_task().
5096 schedstat_inc(env->sd, lb_gained[env->idle]);
5102 static unsigned long task_h_load(struct task_struct *p);
5104 static const unsigned int sched_nr_migrate_break = 32;
5107 * move_tasks tries to move up to imbalance weighted load from busiest to
5108 * this_rq, as part of a balancing operation within domain "sd".
5109 * Returns 1 if successful and 0 otherwise.
5111 * Called with both runqueues locked.
5113 static int move_tasks(struct lb_env *env)
5115 struct list_head *tasks = &env->src_rq->cfs_tasks;
5116 struct task_struct *p;
5120 if (env->imbalance <= 0)
5123 while (!list_empty(tasks)) {
5124 p = list_first_entry(tasks, struct task_struct, se.group_node);
5127 /* We've more or less seen every task there is, call it quits */
5128 if (env->loop > env->loop_max)
5131 /* take a breather every nr_migrate tasks */
5132 if (env->loop > env->loop_break) {
5133 env->loop_break += sched_nr_migrate_break;
5134 env->flags |= LBF_NEED_BREAK;
5138 if (!can_migrate_task(p, env))
5141 load = task_h_load(p);
5143 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5146 if ((load / 2) > env->imbalance)
5151 env->imbalance -= load;
5153 #ifdef CONFIG_PREEMPT
5155 * NEWIDLE balancing is a source of latency, so preemptible
5156 * kernels will stop after the first task is pulled to minimize
5157 * the critical section.
5159 if (env->idle == CPU_NEWLY_IDLE)
5164 * We only want to steal up to the prescribed amount of
5167 if (env->imbalance <= 0)
5172 list_move_tail(&p->se.group_node, tasks);
5176 * Right now, this is one of only two places move_task() is called,
5177 * so we can safely collect move_task() stats here rather than
5178 * inside move_task().
5180 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5185 #ifdef CONFIG_FAIR_GROUP_SCHED
5187 * update tg->load_weight by folding this cpu's load_avg
5189 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5191 struct sched_entity *se = tg->se[cpu];
5192 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5194 /* throttled entities do not contribute to load */
5195 if (throttled_hierarchy(cfs_rq))
5198 update_cfs_rq_blocked_load(cfs_rq, 1);
5201 update_entity_load_avg(se, 1);
5203 * We pivot on our runnable average having decayed to zero for
5204 * list removal. This generally implies that all our children
5205 * have also been removed (modulo rounding error or bandwidth
5206 * control); however, such cases are rare and we can fix these
5209 * TODO: fix up out-of-order children on enqueue.
5211 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5212 list_del_leaf_cfs_rq(cfs_rq);
5214 struct rq *rq = rq_of(cfs_rq);
5215 update_rq_runnable_avg(rq, rq->nr_running);
5219 static void update_blocked_averages(int cpu)
5221 struct rq *rq = cpu_rq(cpu);
5222 struct cfs_rq *cfs_rq;
5223 unsigned long flags;
5225 raw_spin_lock_irqsave(&rq->lock, flags);
5226 update_rq_clock(rq);
5228 * Iterates the task_group tree in a bottom up fashion, see
5229 * list_add_leaf_cfs_rq() for details.
5231 for_each_leaf_cfs_rq(rq, cfs_rq) {
5233 * Note: We may want to consider periodically releasing
5234 * rq->lock about these updates so that creating many task
5235 * groups does not result in continually extending hold time.
5237 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5240 raw_spin_unlock_irqrestore(&rq->lock, flags);
5244 * Compute the cpu's hierarchical load factor for each task group.
5245 * This needs to be done in a top-down fashion because the load of a child
5246 * group is a fraction of its parents load.
5248 static int tg_load_down(struct task_group *tg, void *data)
5251 long cpu = (long)data;
5254 load = cpu_rq(cpu)->load.weight;
5256 load = tg->parent->cfs_rq[cpu]->h_load;
5257 load *= tg->se[cpu]->load.weight;
5258 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
5261 tg->cfs_rq[cpu]->h_load = load;
5266 static void update_h_load(long cpu)
5268 struct rq *rq = cpu_rq(cpu);
5269 unsigned long now = jiffies;
5271 if (rq->h_load_throttle == now)
5274 rq->h_load_throttle = now;
5277 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
5281 static unsigned long task_h_load(struct task_struct *p)
5283 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5286 load = p->se.load.weight;
5287 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
5292 static inline void update_blocked_averages(int cpu)
5296 static inline void update_h_load(long cpu)
5300 static unsigned long task_h_load(struct task_struct *p)
5302 return p->se.load.weight;
5306 /********** Helpers for find_busiest_group ************************/
5308 * sd_lb_stats - Structure to store the statistics of a sched_domain
5309 * during load balancing.
5311 struct sd_lb_stats {
5312 struct sched_group *busiest; /* Busiest group in this sd */
5313 struct sched_group *this; /* Local group in this sd */
5314 unsigned long total_load; /* Total load of all groups in sd */
5315 unsigned long total_pwr; /* Total power of all groups in sd */
5316 unsigned long avg_load; /* Average load across all groups in sd */
5318 /** Statistics of this group */
5319 unsigned long this_load;
5320 unsigned long this_load_per_task;
5321 unsigned long this_nr_running;
5322 unsigned long this_has_capacity;
5323 unsigned int this_idle_cpus;
5325 /* Statistics of the busiest group */
5326 unsigned int busiest_idle_cpus;
5327 unsigned long max_load;
5328 unsigned long busiest_load_per_task;
5329 unsigned long busiest_nr_running;
5330 unsigned long busiest_group_capacity;
5331 unsigned long busiest_has_capacity;
5332 unsigned int busiest_group_weight;
5334 int group_imb; /* Is there imbalance in this sd */
5338 * sg_lb_stats - stats of a sched_group required for load_balancing
5340 struct sg_lb_stats {
5341 unsigned long avg_load; /*Avg load across the CPUs of the group */
5342 unsigned long group_load; /* Total load over the CPUs of the group */
5343 unsigned long sum_nr_running; /* Nr tasks running in the group */
5344 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5345 unsigned long group_capacity;
5346 unsigned long idle_cpus;
5347 unsigned long group_weight;
5348 int group_imb; /* Is there an imbalance in the group ? */
5349 int group_has_capacity; /* Is there extra capacity in the group? */
5353 * get_sd_load_idx - Obtain the load index for a given sched domain.
5354 * @sd: The sched_domain whose load_idx is to be obtained.
5355 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5357 static inline int get_sd_load_idx(struct sched_domain *sd,
5358 enum cpu_idle_type idle)
5364 load_idx = sd->busy_idx;
5367 case CPU_NEWLY_IDLE:
5368 load_idx = sd->newidle_idx;
5371 load_idx = sd->idle_idx;
5378 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5380 return SCHED_POWER_SCALE;
5383 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5385 return default_scale_freq_power(sd, cpu);
5388 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5390 unsigned long weight = sd->span_weight;
5391 unsigned long smt_gain = sd->smt_gain;
5398 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5400 return default_scale_smt_power(sd, cpu);
5403 static unsigned long scale_rt_power(int cpu)
5405 struct rq *rq = cpu_rq(cpu);
5406 u64 total, available, age_stamp, avg;
5409 * Since we're reading these variables without serialization make sure
5410 * we read them once before doing sanity checks on them.
5412 age_stamp = ACCESS_ONCE(rq->age_stamp);
5413 avg = ACCESS_ONCE(rq->rt_avg);
5415 total = sched_avg_period() + (rq->clock - age_stamp);
5417 if (unlikely(total < avg)) {
5418 /* Ensures that power won't end up being negative */
5421 available = total - avg;
5424 if (unlikely((s64)total < SCHED_POWER_SCALE))
5425 total = SCHED_POWER_SCALE;
5427 total >>= SCHED_POWER_SHIFT;
5429 return div_u64(available, total);
5432 static void update_cpu_power(struct sched_domain *sd, int cpu)
5434 unsigned long weight = sd->span_weight;
5435 unsigned long power = SCHED_POWER_SCALE;
5436 struct sched_group *sdg = sd->groups;
5438 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5439 if (sched_feat(ARCH_POWER))
5440 power *= arch_scale_smt_power(sd, cpu);
5442 power *= default_scale_smt_power(sd, cpu);
5444 power >>= SCHED_POWER_SHIFT;
5447 sdg->sgp->power_orig = power;
5449 if (sched_feat(ARCH_POWER))
5450 power *= arch_scale_freq_power(sd, cpu);
5452 power *= default_scale_freq_power(sd, cpu);
5454 power >>= SCHED_POWER_SHIFT;
5456 power *= scale_rt_power(cpu);
5457 power >>= SCHED_POWER_SHIFT;
5462 cpu_rq(cpu)->cpu_power = power;
5463 sdg->sgp->power = power;
5466 void update_group_power(struct sched_domain *sd, int cpu)
5468 struct sched_domain *child = sd->child;
5469 struct sched_group *group, *sdg = sd->groups;
5470 unsigned long power;
5471 unsigned long interval;
5473 interval = msecs_to_jiffies(sd->balance_interval);
5474 interval = clamp(interval, 1UL, max_load_balance_interval);
5475 sdg->sgp->next_update = jiffies + interval;
5478 update_cpu_power(sd, cpu);
5484 if (child->flags & SD_OVERLAP) {
5486 * SD_OVERLAP domains cannot assume that child groups
5487 * span the current group.
5490 for_each_cpu(cpu, sched_group_cpus(sdg))
5491 power += power_of(cpu);
5494 * !SD_OVERLAP domains can assume that child groups
5495 * span the current group.
5498 group = child->groups;
5500 power += group->sgp->power;
5501 group = group->next;
5502 } while (group != child->groups);
5505 sdg->sgp->power_orig = sdg->sgp->power = power;
5509 * Try and fix up capacity for tiny siblings, this is needed when
5510 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5511 * which on its own isn't powerful enough.
5513 * See update_sd_pick_busiest() and check_asym_packing().
5516 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5519 * Only siblings can have significantly less than SCHED_POWER_SCALE
5521 if (!(sd->flags & SD_SHARE_CPUPOWER))
5525 * If ~90% of the cpu_power is still there, we're good.
5527 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5534 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5535 * @env: The load balancing environment.
5536 * @group: sched_group whose statistics are to be updated.
5537 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5538 * @local_group: Does group contain this_cpu.
5539 * @balance: Should we balance.
5540 * @sgs: variable to hold the statistics for this group.
5542 static inline void update_sg_lb_stats(struct lb_env *env,
5543 struct sched_group *group, int load_idx,
5544 int local_group, int *balance, struct sg_lb_stats *sgs)
5546 unsigned long nr_running, max_nr_running, min_nr_running;
5547 unsigned long load, max_cpu_load, min_cpu_load;
5548 unsigned int balance_cpu = -1, first_idle_cpu = 0;
5549 unsigned long avg_load_per_task = 0;
5553 balance_cpu = group_balance_cpu(group);
5555 /* Tally up the load of all CPUs in the group */
5557 min_cpu_load = ~0UL;
5559 min_nr_running = ~0UL;
5561 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5562 struct rq *rq = cpu_rq(i);
5564 nr_running = rq->nr_running;
5566 /* Bias balancing toward cpus of our domain */
5568 if (idle_cpu(i) && !first_idle_cpu &&
5569 cpumask_test_cpu(i, sched_group_mask(group))) {
5574 load = target_load(i, load_idx);
5576 load = source_load(i, load_idx);
5577 if (load > max_cpu_load)
5578 max_cpu_load = load;
5579 if (min_cpu_load > load)
5580 min_cpu_load = load;
5582 if (nr_running > max_nr_running)
5583 max_nr_running = nr_running;
5584 if (min_nr_running > nr_running)
5585 min_nr_running = nr_running;
5588 sgs->group_load += load;
5589 sgs->sum_nr_running += nr_running;
5590 sgs->sum_weighted_load += weighted_cpuload(i);
5596 * First idle cpu or the first cpu(busiest) in this sched group
5597 * is eligible for doing load balancing at this and above
5598 * domains. In the newly idle case, we will allow all the cpu's
5599 * to do the newly idle load balance.
5602 if (env->idle != CPU_NEWLY_IDLE) {
5603 if (balance_cpu != env->dst_cpu) {
5607 update_group_power(env->sd, env->dst_cpu);
5608 } else if (time_after_eq(jiffies, group->sgp->next_update))
5609 update_group_power(env->sd, env->dst_cpu);
5612 /* Adjust by relative CPU power of the group */
5613 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
5616 * Consider the group unbalanced when the imbalance is larger
5617 * than the average weight of a task.
5619 * APZ: with cgroup the avg task weight can vary wildly and
5620 * might not be a suitable number - should we keep a
5621 * normalized nr_running number somewhere that negates
5624 if (sgs->sum_nr_running)
5625 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5627 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
5628 (max_nr_running - min_nr_running) > 1)
5631 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
5633 if (!sgs->group_capacity)
5634 sgs->group_capacity = fix_small_capacity(env->sd, group);
5635 sgs->group_weight = group->group_weight;
5637 if (sgs->group_capacity > sgs->sum_nr_running)
5638 sgs->group_has_capacity = 1;
5642 * update_sd_pick_busiest - return 1 on busiest group
5643 * @env: The load balancing environment.
5644 * @sds: sched_domain statistics
5645 * @sg: sched_group candidate to be checked for being the busiest
5646 * @sgs: sched_group statistics
5648 * Determine if @sg is a busier group than the previously selected
5651 static bool update_sd_pick_busiest(struct lb_env *env,
5652 struct sd_lb_stats *sds,
5653 struct sched_group *sg,
5654 struct sg_lb_stats *sgs)
5656 if (sgs->avg_load <= sds->max_load)
5659 if (sgs->sum_nr_running > sgs->group_capacity)
5666 * ASYM_PACKING needs to move all the work to the lowest
5667 * numbered CPUs in the group, therefore mark all groups
5668 * higher than ourself as busy.
5670 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5671 env->dst_cpu < group_first_cpu(sg)) {
5675 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5683 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5684 * @env: The load balancing environment.
5685 * @balance: Should we balance.
5686 * @sds: variable to hold the statistics for this sched_domain.
5688 static inline void update_sd_lb_stats(struct lb_env *env,
5689 int *balance, struct sd_lb_stats *sds)
5691 struct sched_domain *child = env->sd->child;
5692 struct sched_group *sg = env->sd->groups;
5693 struct sg_lb_stats sgs;
5694 int load_idx, prefer_sibling = 0;
5696 if (child && child->flags & SD_PREFER_SIBLING)
5699 load_idx = get_sd_load_idx(env->sd, env->idle);
5704 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5705 memset(&sgs, 0, sizeof(sgs));
5706 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
5708 if (local_group && !(*balance))
5711 sds->total_load += sgs.group_load;
5712 sds->total_pwr += sg->sgp->power;
5715 * In case the child domain prefers tasks go to siblings
5716 * first, lower the sg capacity to one so that we'll try
5717 * and move all the excess tasks away. We lower the capacity
5718 * of a group only if the local group has the capacity to fit
5719 * these excess tasks, i.e. nr_running < group_capacity. The
5720 * extra check prevents the case where you always pull from the
5721 * heaviest group when it is already under-utilized (possible
5722 * with a large weight task outweighs the tasks on the system).
5724 if (prefer_sibling && !local_group && sds->this_has_capacity)
5725 sgs.group_capacity = min(sgs.group_capacity, 1UL);
5728 sds->this_load = sgs.avg_load;
5730 sds->this_nr_running = sgs.sum_nr_running;
5731 sds->this_load_per_task = sgs.sum_weighted_load;
5732 sds->this_has_capacity = sgs.group_has_capacity;
5733 sds->this_idle_cpus = sgs.idle_cpus;
5734 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
5735 sds->max_load = sgs.avg_load;
5737 sds->busiest_nr_running = sgs.sum_nr_running;
5738 sds->busiest_idle_cpus = sgs.idle_cpus;
5739 sds->busiest_group_capacity = sgs.group_capacity;
5740 sds->busiest_load_per_task = sgs.sum_weighted_load;
5741 sds->busiest_has_capacity = sgs.group_has_capacity;
5742 sds->busiest_group_weight = sgs.group_weight;
5743 sds->group_imb = sgs.group_imb;
5747 } while (sg != env->sd->groups);
5751 * check_asym_packing - Check to see if the group is packed into the
5754 * This is primarily intended to used at the sibling level. Some
5755 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5756 * case of POWER7, it can move to lower SMT modes only when higher
5757 * threads are idle. When in lower SMT modes, the threads will
5758 * perform better since they share less core resources. Hence when we
5759 * have idle threads, we want them to be the higher ones.
5761 * This packing function is run on idle threads. It checks to see if
5762 * the busiest CPU in this domain (core in the P7 case) has a higher
5763 * CPU number than the packing function is being run on. Here we are
5764 * assuming lower CPU number will be equivalent to lower a SMT thread
5767 * Returns 1 when packing is required and a task should be moved to
5768 * this CPU. The amount of the imbalance is returned in *imbalance.
5770 * @env: The load balancing environment.
5771 * @sds: Statistics of the sched_domain which is to be packed
5773 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5777 if (!(env->sd->flags & SD_ASYM_PACKING))
5783 busiest_cpu = group_first_cpu(sds->busiest);
5784 if (env->dst_cpu > busiest_cpu)
5787 env->imbalance = DIV_ROUND_CLOSEST(
5788 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
5794 * fix_small_imbalance - Calculate the minor imbalance that exists
5795 * amongst the groups of a sched_domain, during
5797 * @env: The load balancing environment.
5798 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5801 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5803 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5804 unsigned int imbn = 2;
5805 unsigned long scaled_busy_load_per_task;
5807 if (sds->this_nr_running) {
5808 sds->this_load_per_task /= sds->this_nr_running;
5809 if (sds->busiest_load_per_task >
5810 sds->this_load_per_task)
5813 sds->this_load_per_task =
5814 cpu_avg_load_per_task(env->dst_cpu);
5817 scaled_busy_load_per_task = sds->busiest_load_per_task
5818 * SCHED_POWER_SCALE;
5819 scaled_busy_load_per_task /= sds->busiest->sgp->power;
5821 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
5822 (scaled_busy_load_per_task * imbn)) {
5823 env->imbalance = sds->busiest_load_per_task;
5828 * OK, we don't have enough imbalance to justify moving tasks,
5829 * however we may be able to increase total CPU power used by
5833 pwr_now += sds->busiest->sgp->power *
5834 min(sds->busiest_load_per_task, sds->max_load);
5835 pwr_now += sds->this->sgp->power *
5836 min(sds->this_load_per_task, sds->this_load);
5837 pwr_now /= SCHED_POWER_SCALE;
5839 /* Amount of load we'd subtract */
5840 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5841 sds->busiest->sgp->power;
5842 if (sds->max_load > tmp)
5843 pwr_move += sds->busiest->sgp->power *
5844 min(sds->busiest_load_per_task, sds->max_load - tmp);
5846 /* Amount of load we'd add */
5847 if (sds->max_load * sds->busiest->sgp->power <
5848 sds->busiest_load_per_task * SCHED_POWER_SCALE)
5849 tmp = (sds->max_load * sds->busiest->sgp->power) /
5850 sds->this->sgp->power;
5852 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5853 sds->this->sgp->power;
5854 pwr_move += sds->this->sgp->power *
5855 min(sds->this_load_per_task, sds->this_load + tmp);
5856 pwr_move /= SCHED_POWER_SCALE;
5858 /* Move if we gain throughput */
5859 if (pwr_move > pwr_now)
5860 env->imbalance = sds->busiest_load_per_task;
5864 * calculate_imbalance - Calculate the amount of imbalance present within the
5865 * groups of a given sched_domain during load balance.
5866 * @env: load balance environment
5867 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5869 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5871 unsigned long max_pull, load_above_capacity = ~0UL;
5873 sds->busiest_load_per_task /= sds->busiest_nr_running;
5874 if (sds->group_imb) {
5875 sds->busiest_load_per_task =
5876 min(sds->busiest_load_per_task, sds->avg_load);
5880 * In the presence of smp nice balancing, certain scenarios can have
5881 * max load less than avg load(as we skip the groups at or below
5882 * its cpu_power, while calculating max_load..)
5884 if (sds->max_load < sds->avg_load) {
5886 return fix_small_imbalance(env, sds);
5889 if (!sds->group_imb) {
5891 * Don't want to pull so many tasks that a group would go idle.
5893 load_above_capacity = (sds->busiest_nr_running -
5894 sds->busiest_group_capacity);
5896 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5898 load_above_capacity /= sds->busiest->sgp->power;
5902 * We're trying to get all the cpus to the average_load, so we don't
5903 * want to push ourselves above the average load, nor do we wish to
5904 * reduce the max loaded cpu below the average load. At the same time,
5905 * we also don't want to reduce the group load below the group capacity
5906 * (so that we can implement power-savings policies etc). Thus we look
5907 * for the minimum possible imbalance.
5908 * Be careful of negative numbers as they'll appear as very large values
5909 * with unsigned longs.
5911 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
5913 /* How much load to actually move to equalise the imbalance */
5914 env->imbalance = min(max_pull * sds->busiest->sgp->power,
5915 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
5916 / SCHED_POWER_SCALE;
5919 * if *imbalance is less than the average load per runnable task
5920 * there is no guarantee that any tasks will be moved so we'll have
5921 * a think about bumping its value to force at least one task to be
5924 if (env->imbalance < sds->busiest_load_per_task)
5925 return fix_small_imbalance(env, sds);
5929 /******* find_busiest_group() helpers end here *********************/
5932 * find_busiest_group - Returns the busiest group within the sched_domain
5933 * if there is an imbalance. If there isn't an imbalance, and
5934 * the user has opted for power-savings, it returns a group whose
5935 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5936 * such a group exists.
5938 * Also calculates the amount of weighted load which should be moved
5939 * to restore balance.
5941 * @env: The load balancing environment.
5942 * @balance: Pointer to a variable indicating if this_cpu
5943 * is the appropriate cpu to perform load balancing at this_level.
5945 * Returns: - the busiest group if imbalance exists.
5946 * - If no imbalance and user has opted for power-savings balance,
5947 * return the least loaded group whose CPUs can be
5948 * put to idle by rebalancing its tasks onto our group.
5950 static struct sched_group *
5951 find_busiest_group(struct lb_env *env, int *balance)
5953 struct sd_lb_stats sds;
5955 memset(&sds, 0, sizeof(sds));
5958 * Compute the various statistics relavent for load balancing at
5961 update_sd_lb_stats(env, balance, &sds);
5964 * this_cpu is not the appropriate cpu to perform load balancing at
5970 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5971 check_asym_packing(env, &sds))
5974 /* There is no busy sibling group to pull tasks from */
5975 if (!sds.busiest || sds.busiest_nr_running == 0)
5978 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5981 * If the busiest group is imbalanced the below checks don't
5982 * work because they assumes all things are equal, which typically
5983 * isn't true due to cpus_allowed constraints and the like.
5988 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5989 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5990 !sds.busiest_has_capacity)
5994 * If the local group is more busy than the selected busiest group
5995 * don't try and pull any tasks.
5997 if (sds.this_load >= sds.max_load)
6001 * Don't pull any tasks if this group is already above the domain
6004 if (sds.this_load >= sds.avg_load)
6007 if (env->idle == CPU_IDLE) {
6009 * This cpu is idle. If the busiest group load doesn't
6010 * have more tasks than the number of available cpu's and
6011 * there is no imbalance between this and busiest group
6012 * wrt to idle cpu's, it is balanced.
6014 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
6015 sds.busiest_nr_running <= sds.busiest_group_weight)
6019 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6020 * imbalance_pct to be conservative.
6022 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
6027 /* Looks like there is an imbalance. Compute it */
6028 calculate_imbalance(env, &sds);
6038 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6040 static struct rq *find_busiest_queue(struct lb_env *env,
6041 struct sched_group *group)
6043 struct rq *busiest = NULL, *rq;
6044 unsigned long max_load = 0;
6047 for_each_cpu(i, sched_group_cpus(group)) {
6048 unsigned long power = power_of(i);
6049 unsigned long capacity = DIV_ROUND_CLOSEST(power,
6054 capacity = fix_small_capacity(env->sd, group);
6056 if (!cpumask_test_cpu(i, env->cpus))
6060 wl = weighted_cpuload(i);
6063 * When comparing with imbalance, use weighted_cpuload()
6064 * which is not scaled with the cpu power.
6066 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6070 * For the load comparisons with the other cpu's, consider
6071 * the weighted_cpuload() scaled with the cpu power, so that
6072 * the load can be moved away from the cpu that is potentially
6073 * running at a lower capacity.
6075 wl = (wl * SCHED_POWER_SCALE) / power;
6077 if (wl > max_load) {
6087 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6088 * so long as it is large enough.
6090 #define MAX_PINNED_INTERVAL 512
6092 /* Working cpumask for load_balance and load_balance_newidle. */
6093 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6095 static int need_active_balance(struct lb_env *env)
6097 struct sched_domain *sd = env->sd;
6099 if (env->idle == CPU_NEWLY_IDLE) {
6102 * ASYM_PACKING needs to force migrate tasks from busy but
6103 * higher numbered CPUs in order to pack all tasks in the
6104 * lowest numbered CPUs.
6106 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6110 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6113 static int active_load_balance_cpu_stop(void *data);
6116 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6117 * tasks if there is an imbalance.
6119 static int load_balance(int this_cpu, struct rq *this_rq,
6120 struct sched_domain *sd, enum cpu_idle_type idle,
6123 int ld_moved, cur_ld_moved, active_balance = 0;
6124 struct sched_group *group;
6126 unsigned long flags;
6127 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6129 struct lb_env env = {
6131 .dst_cpu = this_cpu,
6133 .dst_grpmask = sched_group_cpus(sd->groups),
6135 .loop_break = sched_nr_migrate_break,
6140 * For NEWLY_IDLE load_balancing, we don't need to consider
6141 * other cpus in our group
6143 if (idle == CPU_NEWLY_IDLE)
6144 env.dst_grpmask = NULL;
6146 cpumask_copy(cpus, cpu_active_mask);
6148 schedstat_inc(sd, lb_count[idle]);
6151 group = find_busiest_group(&env, balance);
6157 schedstat_inc(sd, lb_nobusyg[idle]);
6161 busiest = find_busiest_queue(&env, group);
6163 schedstat_inc(sd, lb_nobusyq[idle]);
6167 BUG_ON(busiest == env.dst_rq);
6169 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6172 if (busiest->nr_running > 1) {
6174 * Attempt to move tasks. If find_busiest_group has found
6175 * an imbalance but busiest->nr_running <= 1, the group is
6176 * still unbalanced. ld_moved simply stays zero, so it is
6177 * correctly treated as an imbalance.
6179 env.flags |= LBF_ALL_PINNED;
6180 env.src_cpu = busiest->cpu;
6181 env.src_rq = busiest;
6182 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6184 update_h_load(env.src_cpu);
6186 local_irq_save(flags);
6187 double_rq_lock(env.dst_rq, busiest);
6190 * cur_ld_moved - load moved in current iteration
6191 * ld_moved - cumulative load moved across iterations
6193 cur_ld_moved = move_tasks(&env);
6194 ld_moved += cur_ld_moved;
6195 double_rq_unlock(env.dst_rq, busiest);
6196 local_irq_restore(flags);
6199 * some other cpu did the load balance for us.
6201 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6202 resched_cpu(env.dst_cpu);
6204 if (env.flags & LBF_NEED_BREAK) {
6205 env.flags &= ~LBF_NEED_BREAK;
6210 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6211 * us and move them to an alternate dst_cpu in our sched_group
6212 * where they can run. The upper limit on how many times we
6213 * iterate on same src_cpu is dependent on number of cpus in our
6216 * This changes load balance semantics a bit on who can move
6217 * load to a given_cpu. In addition to the given_cpu itself
6218 * (or a ilb_cpu acting on its behalf where given_cpu is
6219 * nohz-idle), we now have balance_cpu in a position to move
6220 * load to given_cpu. In rare situations, this may cause
6221 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6222 * _independently_ and at _same_ time to move some load to
6223 * given_cpu) causing exceess load to be moved to given_cpu.
6224 * This however should not happen so much in practice and
6225 * moreover subsequent load balance cycles should correct the
6226 * excess load moved.
6228 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6230 env.dst_rq = cpu_rq(env.new_dst_cpu);
6231 env.dst_cpu = env.new_dst_cpu;
6232 env.flags &= ~LBF_SOME_PINNED;
6234 env.loop_break = sched_nr_migrate_break;
6236 /* Prevent to re-select dst_cpu via env's cpus */
6237 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6240 * Go back to "more_balance" rather than "redo" since we
6241 * need to continue with same src_cpu.
6246 /* All tasks on this runqueue were pinned by CPU affinity */
6247 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6248 cpumask_clear_cpu(cpu_of(busiest), cpus);
6249 if (!cpumask_empty(cpus)) {
6251 env.loop_break = sched_nr_migrate_break;
6259 schedstat_inc(sd, lb_failed[idle]);
6261 * Increment the failure counter only on periodic balance.
6262 * We do not want newidle balance, which can be very
6263 * frequent, pollute the failure counter causing
6264 * excessive cache_hot migrations and active balances.
6266 if (idle != CPU_NEWLY_IDLE)
6267 sd->nr_balance_failed++;
6269 if (need_active_balance(&env)) {
6270 raw_spin_lock_irqsave(&busiest->lock, flags);
6272 /* don't kick the active_load_balance_cpu_stop,
6273 * if the curr task on busiest cpu can't be
6276 if (!cpumask_test_cpu(this_cpu,
6277 tsk_cpus_allowed(busiest->curr))) {
6278 raw_spin_unlock_irqrestore(&busiest->lock,
6280 env.flags |= LBF_ALL_PINNED;
6281 goto out_one_pinned;
6285 * ->active_balance synchronizes accesses to
6286 * ->active_balance_work. Once set, it's cleared
6287 * only after active load balance is finished.
6289 if (!busiest->active_balance) {
6290 busiest->active_balance = 1;
6291 busiest->push_cpu = this_cpu;
6294 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6296 if (active_balance) {
6297 stop_one_cpu_nowait(cpu_of(busiest),
6298 active_load_balance_cpu_stop, busiest,
6299 &busiest->active_balance_work);
6303 * We've kicked active balancing, reset the failure
6306 sd->nr_balance_failed = sd->cache_nice_tries+1;
6309 sd->nr_balance_failed = 0;
6311 if (likely(!active_balance)) {
6312 /* We were unbalanced, so reset the balancing interval */
6313 sd->balance_interval = sd->min_interval;
6316 * If we've begun active balancing, start to back off. This
6317 * case may not be covered by the all_pinned logic if there
6318 * is only 1 task on the busy runqueue (because we don't call
6321 if (sd->balance_interval < sd->max_interval)
6322 sd->balance_interval *= 2;
6328 schedstat_inc(sd, lb_balanced[idle]);
6330 sd->nr_balance_failed = 0;
6333 /* tune up the balancing interval */
6334 if (((env.flags & LBF_ALL_PINNED) &&
6335 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6336 (sd->balance_interval < sd->max_interval))
6337 sd->balance_interval *= 2;
6344 #ifdef CONFIG_SCHED_HMP
6345 static unsigned int hmp_idle_pull(int this_cpu);
6346 static int move_specific_task(struct lb_env *env, struct task_struct *pm);
6348 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
6355 * idle_balance is called by schedule() if this_cpu is about to become
6356 * idle. Attempts to pull tasks from other CPUs.
6358 void idle_balance(int this_cpu, struct rq *this_rq)
6360 struct sched_domain *sd;
6361 int pulled_task = 0;
6362 unsigned long next_balance = jiffies + HZ;
6364 this_rq->idle_stamp = this_rq->clock;
6366 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6370 * Drop the rq->lock, but keep IRQ/preempt disabled.
6372 raw_spin_unlock(&this_rq->lock);
6374 update_blocked_averages(this_cpu);
6376 for_each_domain(this_cpu, sd) {
6377 unsigned long interval;
6380 if (!(sd->flags & SD_LOAD_BALANCE))
6383 if (sd->flags & SD_BALANCE_NEWIDLE) {
6384 /* If we've pulled tasks over stop searching: */
6385 pulled_task = load_balance(this_cpu, this_rq,
6386 sd, CPU_NEWLY_IDLE, &balance);
6389 interval = msecs_to_jiffies(sd->balance_interval);
6390 if (time_after(next_balance, sd->last_balance + interval))
6391 next_balance = sd->last_balance + interval;
6393 this_rq->idle_stamp = 0;
6398 #ifdef CONFIG_SCHED_HMP
6400 pulled_task = hmp_idle_pull(this_cpu);
6402 raw_spin_lock(&this_rq->lock);
6404 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6406 * We are going idle. next_balance may be set based on
6407 * a busy processor. So reset next_balance.
6409 this_rq->next_balance = next_balance;
6413 static int __do_active_load_balance_cpu_stop(void *data, bool check_sd_lb_flag)
6415 struct rq *busiest_rq = data;
6416 int busiest_cpu = cpu_of(busiest_rq);
6417 int target_cpu = busiest_rq->push_cpu;
6418 struct rq *target_rq = cpu_rq(target_cpu);
6419 struct sched_domain *sd;
6420 struct task_struct *p = NULL;
6422 raw_spin_lock_irq(&busiest_rq->lock);
6423 #ifdef CONFIG_SCHED_HMP
6424 p = busiest_rq->migrate_task;
6426 /* make sure the requested cpu hasn't gone down in the meantime */
6427 if (unlikely(busiest_cpu != smp_processor_id() ||
6428 !busiest_rq->active_balance))
6431 /* Is there any task to move? */
6432 if (busiest_rq->nr_running <= 1)
6435 if (!check_sd_lb_flag) {
6436 /* Task has migrated meanwhile, abort forced migration */
6437 if (task_rq(p) != busiest_rq)
6441 * This condition is "impossible", if it occurs
6442 * we need to fix it. Originally reported by
6443 * Bjorn Helgaas on a 128-cpu setup.
6445 BUG_ON(busiest_rq == target_rq);
6447 /* move a task from busiest_rq to target_rq */
6448 double_lock_balance(busiest_rq, target_rq);
6450 /* Search for an sd spanning us and the target CPU. */
6452 for_each_domain(target_cpu, sd) {
6453 if (((check_sd_lb_flag && sd->flags & SD_LOAD_BALANCE) ||
6454 !check_sd_lb_flag) &&
6455 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6460 bool success = false;
6461 struct lb_env env = {
6463 .dst_cpu = target_cpu,
6464 .dst_rq = target_rq,
6465 .src_cpu = busiest_rq->cpu,
6466 .src_rq = busiest_rq,
6470 schedstat_inc(sd, alb_count);
6472 if (check_sd_lb_flag) {
6473 if (move_one_task(&env))
6476 if (move_specific_task(&env, p))
6480 schedstat_inc(sd, alb_pushed);
6482 schedstat_inc(sd, alb_failed);
6485 double_unlock_balance(busiest_rq, target_rq);
6487 if (!check_sd_lb_flag)
6489 busiest_rq->active_balance = 0;
6490 raw_spin_unlock_irq(&busiest_rq->lock);
6495 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6496 * running tasks off the busiest CPU onto idle CPUs. It requires at
6497 * least 1 task to be running on each physical CPU where possible, and
6498 * avoids physical / logical imbalances.
6500 static int active_load_balance_cpu_stop(void *data)
6502 return __do_active_load_balance_cpu_stop(data, true);
6505 #ifdef CONFIG_NO_HZ_COMMON
6507 * idle load balancing details
6508 * - When one of the busy CPUs notice that there may be an idle rebalancing
6509 * needed, they will kick the idle load balancer, which then does idle
6510 * load balancing for all the idle CPUs.
6513 cpumask_var_t idle_cpus_mask;
6515 unsigned long next_balance; /* in jiffy units */
6516 } nohz ____cacheline_aligned;
6519 * nohz_test_cpu used when load tracking is enabled. FAIR_GROUP_SCHED
6520 * dependency below may be removed when load tracking guards are
6523 #ifdef CONFIG_FAIR_GROUP_SCHED
6524 static int nohz_test_cpu(int cpu)
6526 return cpumask_test_cpu(cpu, nohz.idle_cpus_mask);
6530 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6532 * Decide if the tasks on the busy CPUs in the
6533 * littlest domain would benefit from an idle balance
6535 static int hmp_packing_ilb_needed(int cpu)
6537 struct hmp_domain *hmp;
6538 /* always allow ilb on non-slowest domain */
6539 if (!hmp_cpu_is_slowest(cpu))
6542 /* if disabled, use normal ILB behaviour */
6543 if (!hmp_packing_enabled)
6546 hmp = hmp_cpu_domain(cpu);
6547 for_each_cpu_and(cpu, &hmp->cpus, nohz.idle_cpus_mask) {
6548 /* only idle balance if a CPU is loaded over threshold */
6549 if (cpu_rq(cpu)->avg.load_avg_ratio > hmp_full_threshold)
6556 static inline int find_new_ilb(int call_cpu)
6558 int ilb = cpumask_first(nohz.idle_cpus_mask);
6559 #ifdef CONFIG_SCHED_HMP
6562 /* restrict nohz balancing to occur in the same hmp domain */
6563 ilb = cpumask_first_and(nohz.idle_cpus_mask,
6564 &((struct hmp_domain *)hmp_cpu_domain(call_cpu))->cpus);
6566 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6567 if (ilb < nr_cpu_ids)
6568 ilb_needed = hmp_packing_ilb_needed(ilb);
6571 if (ilb_needed && ilb < nr_cpu_ids && idle_cpu(ilb))
6574 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6582 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6583 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6584 * CPU (if there is one).
6586 static void nohz_balancer_kick(int cpu)
6590 nohz.next_balance++;
6592 ilb_cpu = find_new_ilb(cpu);
6594 if (ilb_cpu >= nr_cpu_ids)
6597 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6600 * Use smp_send_reschedule() instead of resched_cpu().
6601 * This way we generate a sched IPI on the target cpu which
6602 * is idle. And the softirq performing nohz idle load balance
6603 * will be run before returning from the IPI.
6605 smp_send_reschedule(ilb_cpu);
6609 static inline void nohz_balance_exit_idle(int cpu)
6611 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6612 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6613 atomic_dec(&nohz.nr_cpus);
6614 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6618 static inline void set_cpu_sd_state_busy(void)
6620 struct sched_domain *sd;
6621 int cpu = smp_processor_id();
6624 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6626 if (!sd || !sd->nohz_idle)
6630 for (; sd; sd = sd->parent)
6631 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6636 void set_cpu_sd_state_idle(void)
6638 struct sched_domain *sd;
6639 int cpu = smp_processor_id();
6642 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6644 if (!sd || sd->nohz_idle)
6648 for (; sd; sd = sd->parent)
6649 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6655 * This routine will record that the cpu is going idle with tick stopped.
6656 * This info will be used in performing idle load balancing in the future.
6658 void nohz_balance_enter_idle(int cpu)
6661 * If this cpu is going down, then nothing needs to be done.
6663 if (!cpu_active(cpu))
6666 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6669 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6670 atomic_inc(&nohz.nr_cpus);
6671 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6674 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
6675 unsigned long action, void *hcpu)
6677 switch (action & ~CPU_TASKS_FROZEN) {
6679 nohz_balance_exit_idle(smp_processor_id());
6687 static DEFINE_SPINLOCK(balancing);
6690 * Scale the max load_balance interval with the number of CPUs in the system.
6691 * This trades load-balance latency on larger machines for less cross talk.
6693 void update_max_interval(void)
6695 max_load_balance_interval = HZ*num_online_cpus()/10;
6699 * It checks each scheduling domain to see if it is due to be balanced,
6700 * and initiates a balancing operation if so.
6702 * Balancing parameters are set up in init_sched_domains.
6704 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6707 struct rq *rq = cpu_rq(cpu);
6708 unsigned long interval;
6709 struct sched_domain *sd;
6710 /* Earliest time when we have to do rebalance again */
6711 unsigned long next_balance = jiffies + 60*HZ;
6712 int update_next_balance = 0;
6715 update_blocked_averages(cpu);
6718 for_each_domain(cpu, sd) {
6719 if (!(sd->flags & SD_LOAD_BALANCE))
6722 interval = sd->balance_interval;
6723 if (idle != CPU_IDLE)
6724 interval *= sd->busy_factor;
6726 /* scale ms to jiffies */
6727 interval = msecs_to_jiffies(interval);
6728 interval = clamp(interval, 1UL, max_load_balance_interval);
6730 need_serialize = sd->flags & SD_SERIALIZE;
6732 if (need_serialize) {
6733 if (!spin_trylock(&balancing))
6737 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6738 if (load_balance(cpu, rq, sd, idle, &balance)) {
6740 * The LBF_SOME_PINNED logic could have changed
6741 * env->dst_cpu, so we can't know our idle
6742 * state even if we migrated tasks. Update it.
6744 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6746 sd->last_balance = jiffies;
6749 spin_unlock(&balancing);
6751 if (time_after(next_balance, sd->last_balance + interval)) {
6752 next_balance = sd->last_balance + interval;
6753 update_next_balance = 1;
6757 * Stop the load balance at this level. There is another
6758 * CPU in our sched group which is doing load balancing more
6767 * next_balance will be updated only when there is a need.
6768 * When the cpu is attached to null domain for ex, it will not be
6771 if (likely(update_next_balance))
6772 rq->next_balance = next_balance;
6775 #ifdef CONFIG_NO_HZ_COMMON
6777 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6778 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6780 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6782 struct rq *this_rq = cpu_rq(this_cpu);
6786 if (idle != CPU_IDLE ||
6787 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6790 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6791 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6795 * If this cpu gets work to do, stop the load balancing
6796 * work being done for other cpus. Next load
6797 * balancing owner will pick it up.
6802 rq = cpu_rq(balance_cpu);
6804 raw_spin_lock_irq(&rq->lock);
6805 update_rq_clock(rq);
6806 update_idle_cpu_load(rq);
6807 raw_spin_unlock_irq(&rq->lock);
6809 rebalance_domains(balance_cpu, CPU_IDLE);
6811 if (time_after(this_rq->next_balance, rq->next_balance))
6812 this_rq->next_balance = rq->next_balance;
6814 nohz.next_balance = this_rq->next_balance;
6816 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6820 * Current heuristic for kicking the idle load balancer in the presence
6821 * of an idle cpu is the system.
6822 * - This rq has more than one task.
6823 * - At any scheduler domain level, this cpu's scheduler group has multiple
6824 * busy cpu's exceeding the group's power.
6825 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6826 * domain span are idle.
6828 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6830 unsigned long now = jiffies;
6831 struct sched_domain *sd;
6833 if (unlikely(idle_cpu(cpu)))
6837 * We may be recently in ticked or tickless idle mode. At the first
6838 * busy tick after returning from idle, we will update the busy stats.
6840 set_cpu_sd_state_busy();
6841 nohz_balance_exit_idle(cpu);
6844 * None are in tickless mode and hence no need for NOHZ idle load
6847 if (likely(!atomic_read(&nohz.nr_cpus)))
6850 if (time_before(now, nohz.next_balance))
6853 #ifdef CONFIG_SCHED_HMP
6855 * Bail out if there are no nohz CPUs in our
6856 * HMP domain, since we will move tasks between
6857 * domains through wakeup and force balancing
6858 * as necessary based upon task load.
6860 if (cpumask_first_and(nohz.idle_cpus_mask,
6861 &((struct hmp_domain *)hmp_cpu_domain(cpu))->cpus) >= nr_cpu_ids)
6865 if (rq->nr_running >= 2)
6869 for_each_domain(cpu, sd) {
6870 struct sched_group *sg = sd->groups;
6871 struct sched_group_power *sgp = sg->sgp;
6872 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6874 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6875 goto need_kick_unlock;
6877 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6878 && (cpumask_first_and(nohz.idle_cpus_mask,
6879 sched_domain_span(sd)) < cpu))
6880 goto need_kick_unlock;
6882 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6894 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6897 #ifdef CONFIG_SCHED_HMP
6898 static unsigned int hmp_task_eligible_for_up_migration(struct sched_entity *se)
6900 /* below hmp_up_threshold, never eligible */
6901 if (se->avg.load_avg_ratio < hmp_up_threshold)
6906 /* Check if task should migrate to a faster cpu */
6907 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se)
6909 struct task_struct *p = task_of(se);
6910 int temp_target_cpu;
6913 if (hmp_cpu_is_fastest(cpu))
6916 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6917 /* Filter by task priority */
6918 if (p->prio >= hmp_up_prio)
6921 if (!hmp_task_eligible_for_up_migration(se))
6924 /* Let the task load settle before doing another up migration */
6925 /* hack - always use clock from first online CPU */
6926 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6927 if (((now - se->avg.hmp_last_up_migration) >> 10)
6928 < hmp_next_up_threshold)
6931 /* hmp_domain_min_load only returns 0 for an
6932 * idle CPU or 1023 for any partly-busy one.
6933 * Be explicit about requirement for an idle CPU.
6935 if (hmp_domain_min_load(hmp_faster_domain(cpu), &temp_target_cpu,
6936 tsk_cpus_allowed(p)) == 0 && temp_target_cpu != NR_CPUS) {
6938 *target_cpu = temp_target_cpu;
6944 /* Check if task should migrate to a slower cpu */
6945 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
6947 struct task_struct *p = task_of(se);
6950 if (hmp_cpu_is_slowest(cpu)) {
6951 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6952 if(hmp_packing_enabled)
6959 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6960 /* Filter by task priority */
6961 if ((p->prio >= hmp_up_prio) &&
6962 cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6963 tsk_cpus_allowed(p))) {
6968 /* Let the task load settle before doing another down migration */
6969 /* hack - always use clock from first online CPU */
6970 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6971 if (((now - se->avg.hmp_last_down_migration) >> 10)
6972 < hmp_next_down_threshold)
6975 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6976 tsk_cpus_allowed(p))
6977 && se->avg.load_avg_ratio < hmp_down_threshold) {
6984 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6985 * Ideally this function should be merged with can_migrate_task() to avoid
6988 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
6990 int tsk_cache_hot = 0;
6993 * We do not migrate tasks that are:
6994 * 1) running (obviously), or
6995 * 2) cannot be migrated to this CPU due to cpus_allowed
6997 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6998 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
7001 env->flags &= ~LBF_ALL_PINNED;
7003 if (task_running(env->src_rq, p)) {
7004 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
7009 * Aggressive migration if:
7010 * 1) task is cache cold, or
7011 * 2) too many balance attempts have failed.
7014 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
7015 if (!tsk_cache_hot ||
7016 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7017 #ifdef CONFIG_SCHEDSTATS
7018 if (tsk_cache_hot) {
7019 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
7020 schedstat_inc(p, se.statistics.nr_forced_migrations);
7030 * move_specific_task tries to move a specific task.
7031 * Returns 1 if successful and 0 otherwise.
7032 * Called with both runqueues locked.
7034 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
7036 struct task_struct *p, *n;
7038 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7039 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
7043 if (!hmp_can_migrate_task(p, env))
7045 /* Check if we found the right task */
7051 * Right now, this is only the third place move_task()
7052 * is called, so we can safely collect move_task()
7053 * stats here rather than inside move_task().
7055 schedstat_inc(env->sd, lb_gained[env->idle]);
7062 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
7063 * migrate a specific task from one runqueue to another.
7064 * hmp_force_up_migration uses this to push a currently running task
7065 * off a runqueue. hmp_idle_pull uses this to pull a currently
7066 * running task to an idle runqueue.
7067 * Reuses __do_active_load_balance_cpu_stop to actually do the work.
7069 static int hmp_active_task_migration_cpu_stop(void *data)
7071 return __do_active_load_balance_cpu_stop(data, false);
7075 * Move task in a runnable state to another CPU.
7077 * Tailored on 'active_load_balance_cpu_stop' with slight
7078 * modification to locking and pre-transfer checks. Note
7079 * rq->lock must be held before calling.
7081 static void hmp_migrate_runnable_task(struct rq *rq)
7083 struct sched_domain *sd;
7084 int src_cpu = cpu_of(rq);
7085 struct rq *src_rq = rq;
7086 int dst_cpu = rq->push_cpu;
7087 struct rq *dst_rq = cpu_rq(dst_cpu);
7088 struct task_struct *p = rq->migrate_task;
7090 * One last check to make sure nobody else is playing
7091 * with the source rq.
7093 if (src_rq->active_balance)
7096 if (src_rq->nr_running <= 1)
7099 if (task_rq(p) != src_rq)
7102 * Not sure if this applies here but one can never
7105 BUG_ON(src_rq == dst_rq);
7107 double_lock_balance(src_rq, dst_rq);
7110 for_each_domain(dst_cpu, sd) {
7111 if (cpumask_test_cpu(src_cpu, sched_domain_span(sd)))
7116 struct lb_env env = {
7125 schedstat_inc(sd, alb_count);
7127 if (move_specific_task(&env, p))
7128 schedstat_inc(sd, alb_pushed);
7130 schedstat_inc(sd, alb_failed);
7134 double_unlock_balance(src_rq, dst_rq);
7139 static DEFINE_SPINLOCK(hmp_force_migration);
7142 * hmp_force_up_migration checks runqueues for tasks that need to
7143 * be actively migrated to a faster cpu.
7145 static void hmp_force_up_migration(int this_cpu)
7147 int cpu, target_cpu;
7148 struct sched_entity *curr, *orig;
7150 unsigned long flags;
7151 unsigned int force, got_target;
7152 struct task_struct *p;
7154 if (!spin_trylock(&hmp_force_migration))
7156 for_each_online_cpu(cpu) {
7159 target = cpu_rq(cpu);
7160 raw_spin_lock_irqsave(&target->lock, flags);
7161 curr = target->cfs.curr;
7162 if (!curr || target->active_balance) {
7163 raw_spin_unlock_irqrestore(&target->lock, flags);
7166 if (!entity_is_task(curr)) {
7167 struct cfs_rq *cfs_rq;
7169 cfs_rq = group_cfs_rq(curr);
7171 curr = cfs_rq->curr;
7172 cfs_rq = group_cfs_rq(curr);
7176 curr = hmp_get_heaviest_task(curr, -1);
7178 raw_spin_unlock_irqrestore(&target->lock, flags);
7182 if (hmp_up_migration(cpu, &target_cpu, curr)) {
7183 cpu_rq(target_cpu)->wake_for_idle_pull = 1;
7184 raw_spin_unlock_irqrestore(&target->lock, flags);
7185 spin_unlock(&hmp_force_migration);
7186 smp_send_reschedule(target_cpu);
7191 * For now we just check the currently running task.
7192 * Selecting the lightest task for offloading will
7193 * require extensive book keeping.
7195 curr = hmp_get_lightest_task(orig, 1);
7197 target->push_cpu = hmp_offload_down(cpu, curr);
7198 if (target->push_cpu < NR_CPUS) {
7200 target->migrate_task = p;
7202 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_OFFLOAD);
7203 hmp_next_down_delay(&p->se, target->push_cpu);
7207 * We have a target with no active_balance. If the task
7208 * is not currently running move it, otherwise let the
7209 * CPU stopper take care of it.
7212 if (!task_running(target, p)) {
7213 trace_sched_hmp_migrate_force_running(p, 0);
7214 hmp_migrate_runnable_task(target);
7216 target->active_balance = 1;
7221 raw_spin_unlock_irqrestore(&target->lock, flags);
7224 stop_one_cpu_nowait(cpu_of(target),
7225 hmp_active_task_migration_cpu_stop,
7226 target, &target->active_balance_work);
7228 spin_unlock(&hmp_force_migration);
7231 * hmp_idle_pull looks at little domain runqueues to see
7232 * if a task should be pulled.
7234 * Reuses hmp_force_migration spinlock.
7237 static unsigned int hmp_idle_pull(int this_cpu)
7240 struct sched_entity *curr, *orig;
7241 struct hmp_domain *hmp_domain = NULL;
7242 struct rq *target = NULL, *rq;
7243 unsigned long flags, ratio = 0;
7244 unsigned int force = 0;
7245 struct task_struct *p = NULL;
7247 if (!hmp_cpu_is_slowest(this_cpu))
7248 hmp_domain = hmp_slower_domain(this_cpu);
7252 if (!spin_trylock(&hmp_force_migration))
7255 /* first select a task */
7256 for_each_cpu(cpu, &hmp_domain->cpus) {
7258 raw_spin_lock_irqsave(&rq->lock, flags);
7259 curr = rq->cfs.curr;
7261 raw_spin_unlock_irqrestore(&rq->lock, flags);
7264 if (!entity_is_task(curr)) {
7265 struct cfs_rq *cfs_rq;
7267 cfs_rq = group_cfs_rq(curr);
7269 curr = cfs_rq->curr;
7270 if (!entity_is_task(curr))
7271 cfs_rq = group_cfs_rq(curr);
7277 curr = hmp_get_heaviest_task(curr, this_cpu);
7278 /* check if heaviest eligible task on this
7279 * CPU is heavier than previous task
7281 if (curr && hmp_task_eligible_for_up_migration(curr) &&
7282 curr->avg.load_avg_ratio > ratio &&
7283 cpumask_test_cpu(this_cpu,
7284 tsk_cpus_allowed(task_of(curr)))) {
7287 ratio = curr->avg.load_avg_ratio;
7289 raw_spin_unlock_irqrestore(&rq->lock, flags);
7295 /* now we have a candidate */
7296 raw_spin_lock_irqsave(&target->lock, flags);
7297 if (!target->active_balance && task_rq(p) == target) {
7299 target->push_cpu = this_cpu;
7300 target->migrate_task = p;
7301 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_IDLE_PULL);
7302 hmp_next_up_delay(&p->se, target->push_cpu);
7304 * if the task isn't running move it right away.
7305 * Otherwise setup the active_balance mechanic and let
7306 * the CPU stopper do its job.
7308 if (!task_running(target, p)) {
7309 trace_sched_hmp_migrate_idle_running(p, 0);
7310 hmp_migrate_runnable_task(target);
7312 target->active_balance = 1;
7316 raw_spin_unlock_irqrestore(&target->lock, flags);
7319 /* start timer to keep us awake */
7320 hmp_cpu_keepalive_trigger();
7321 stop_one_cpu_nowait(cpu_of(target),
7322 hmp_active_task_migration_cpu_stop,
7323 target, &target->active_balance_work);
7326 spin_unlock(&hmp_force_migration);
7330 static void hmp_force_up_migration(int this_cpu) { }
7331 #endif /* CONFIG_SCHED_HMP */
7334 * run_rebalance_domains is triggered when needed from the scheduler tick.
7335 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7337 static void run_rebalance_domains(struct softirq_action *h)
7339 int this_cpu = smp_processor_id();
7340 struct rq *this_rq = cpu_rq(this_cpu);
7341 enum cpu_idle_type idle = this_rq->idle_balance ?
7342 CPU_IDLE : CPU_NOT_IDLE;
7344 #ifdef CONFIG_SCHED_HMP
7345 /* shortcut for hmp idle pull wakeups */
7346 if (unlikely(this_rq->wake_for_idle_pull)) {
7347 this_rq->wake_for_idle_pull = 0;
7348 if (hmp_idle_pull(this_cpu)) {
7349 /* break out unless running nohz idle as well */
7350 if (idle != CPU_IDLE)
7356 hmp_force_up_migration(this_cpu);
7358 rebalance_domains(this_cpu, idle);
7361 * If this cpu has a pending nohz_balance_kick, then do the
7362 * balancing on behalf of the other idle cpus whose ticks are
7365 nohz_idle_balance(this_cpu, idle);
7368 static inline int on_null_domain(int cpu)
7370 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
7374 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7376 void trigger_load_balance(struct rq *rq, int cpu)
7378 /* Don't need to rebalance while attached to NULL domain */
7379 if (time_after_eq(jiffies, rq->next_balance) &&
7380 likely(!on_null_domain(cpu)))
7381 raise_softirq(SCHED_SOFTIRQ);
7382 #ifdef CONFIG_NO_HZ_COMMON
7383 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
7384 nohz_balancer_kick(cpu);
7388 static void rq_online_fair(struct rq *rq)
7390 #ifdef CONFIG_SCHED_HMP
7391 hmp_online_cpu(rq->cpu);
7396 static void rq_offline_fair(struct rq *rq)
7398 #ifdef CONFIG_SCHED_HMP
7399 hmp_offline_cpu(rq->cpu);
7403 /* Ensure any throttled groups are reachable by pick_next_task */
7404 unthrottle_offline_cfs_rqs(rq);
7407 #endif /* CONFIG_SMP */
7410 * scheduler tick hitting a task of our scheduling class:
7412 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7414 struct cfs_rq *cfs_rq;
7415 struct sched_entity *se = &curr->se;
7417 for_each_sched_entity(se) {
7418 cfs_rq = cfs_rq_of(se);
7419 entity_tick(cfs_rq, se, queued);
7422 if (sched_feat_numa(NUMA))
7423 task_tick_numa(rq, curr);
7425 update_rq_runnable_avg(rq, 1);
7429 * called on fork with the child task as argument from the parent's context
7430 * - child not yet on the tasklist
7431 * - preemption disabled
7433 static void task_fork_fair(struct task_struct *p)
7435 struct cfs_rq *cfs_rq;
7436 struct sched_entity *se = &p->se, *curr;
7437 int this_cpu = smp_processor_id();
7438 struct rq *rq = this_rq();
7439 unsigned long flags;
7441 raw_spin_lock_irqsave(&rq->lock, flags);
7443 update_rq_clock(rq);
7445 cfs_rq = task_cfs_rq(current);
7446 curr = cfs_rq->curr;
7449 * Not only the cpu but also the task_group of the parent might have
7450 * been changed after parent->se.parent,cfs_rq were copied to
7451 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7452 * of child point to valid ones.
7455 __set_task_cpu(p, this_cpu);
7458 update_curr(cfs_rq);
7461 se->vruntime = curr->vruntime;
7462 place_entity(cfs_rq, se, 1);
7464 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7466 * Upon rescheduling, sched_class::put_prev_task() will place
7467 * 'current' within the tree based on its new key value.
7469 swap(curr->vruntime, se->vruntime);
7470 resched_task(rq->curr);
7473 se->vruntime -= cfs_rq->min_vruntime;
7475 raw_spin_unlock_irqrestore(&rq->lock, flags);
7479 * Priority of the task has changed. Check to see if we preempt
7483 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7489 * Reschedule if we are currently running on this runqueue and
7490 * our priority decreased, or if we are not currently running on
7491 * this runqueue and our priority is higher than the current's
7493 if (rq->curr == p) {
7494 if (p->prio > oldprio)
7495 resched_task(rq->curr);
7497 check_preempt_curr(rq, p, 0);
7500 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7502 struct sched_entity *se = &p->se;
7503 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7506 * Ensure the task's vruntime is normalized, so that when it's
7507 * switched back to the fair class the enqueue_entity(.flags=0) will
7508 * do the right thing.
7510 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7511 * have normalized the vruntime, if it's !on_rq, then only when
7512 * the task is sleeping will it still have non-normalized vruntime.
7514 if (!p->on_rq && p->state != TASK_RUNNING) {
7516 * Fix up our vruntime so that the current sleep doesn't
7517 * cause 'unlimited' sleep bonus.
7519 place_entity(cfs_rq, se, 0);
7520 se->vruntime -= cfs_rq->min_vruntime;
7523 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7525 * Remove our load from contribution when we leave sched_fair
7526 * and ensure we don't carry in an old decay_count if we
7529 if (p->se.avg.decay_count) {
7530 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
7531 __synchronize_entity_decay(&p->se);
7532 subtract_blocked_load_contrib(cfs_rq,
7533 p->se.avg.load_avg_contrib);
7539 * We switched to the sched_fair class.
7541 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7547 * We were most likely switched from sched_rt, so
7548 * kick off the schedule if running, otherwise just see
7549 * if we can still preempt the current task.
7552 resched_task(rq->curr);
7554 check_preempt_curr(rq, p, 0);
7557 /* Account for a task changing its policy or group.
7559 * This routine is mostly called to set cfs_rq->curr field when a task
7560 * migrates between groups/classes.
7562 static void set_curr_task_fair(struct rq *rq)
7564 struct sched_entity *se = &rq->curr->se;
7566 for_each_sched_entity(se) {
7567 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7569 set_next_entity(cfs_rq, se);
7570 /* ensure bandwidth has been allocated on our new cfs_rq */
7571 account_cfs_rq_runtime(cfs_rq, 0);
7575 void init_cfs_rq(struct cfs_rq *cfs_rq)
7577 cfs_rq->tasks_timeline = RB_ROOT;
7578 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7579 #ifndef CONFIG_64BIT
7580 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7582 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7583 atomic64_set(&cfs_rq->decay_counter, 1);
7584 atomic64_set(&cfs_rq->removed_load, 0);
7588 #ifdef CONFIG_FAIR_GROUP_SCHED
7589 static void task_move_group_fair(struct task_struct *p, int on_rq)
7591 struct cfs_rq *cfs_rq;
7593 * If the task was not on the rq at the time of this cgroup movement
7594 * it must have been asleep, sleeping tasks keep their ->vruntime
7595 * absolute on their old rq until wakeup (needed for the fair sleeper
7596 * bonus in place_entity()).
7598 * If it was on the rq, we've just 'preempted' it, which does convert
7599 * ->vruntime to a relative base.
7601 * Make sure both cases convert their relative position when migrating
7602 * to another cgroup's rq. This does somewhat interfere with the
7603 * fair sleeper stuff for the first placement, but who cares.
7606 * When !on_rq, vruntime of the task has usually NOT been normalized.
7607 * But there are some cases where it has already been normalized:
7609 * - Moving a forked child which is waiting for being woken up by
7610 * wake_up_new_task().
7611 * - Moving a task which has been woken up by try_to_wake_up() and
7612 * waiting for actually being woken up by sched_ttwu_pending().
7614 * To prevent boost or penalty in the new cfs_rq caused by delta
7615 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7617 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7621 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7622 set_task_rq(p, task_cpu(p));
7624 cfs_rq = cfs_rq_of(&p->se);
7625 p->se.vruntime += cfs_rq->min_vruntime;
7628 * migrate_task_rq_fair() will have removed our previous
7629 * contribution, but we must synchronize for ongoing future
7632 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7633 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7638 void free_fair_sched_group(struct task_group *tg)
7642 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7644 for_each_possible_cpu(i) {
7646 kfree(tg->cfs_rq[i]);
7655 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7657 struct cfs_rq *cfs_rq;
7658 struct sched_entity *se;
7661 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7664 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7668 tg->shares = NICE_0_LOAD;
7670 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7672 for_each_possible_cpu(i) {
7673 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7674 GFP_KERNEL, cpu_to_node(i));
7678 se = kzalloc_node(sizeof(struct sched_entity),
7679 GFP_KERNEL, cpu_to_node(i));
7683 init_cfs_rq(cfs_rq);
7684 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7695 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7697 struct rq *rq = cpu_rq(cpu);
7698 unsigned long flags;
7701 * Only empty task groups can be destroyed; so we can speculatively
7702 * check on_list without danger of it being re-added.
7704 if (!tg->cfs_rq[cpu]->on_list)
7707 raw_spin_lock_irqsave(&rq->lock, flags);
7708 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7709 raw_spin_unlock_irqrestore(&rq->lock, flags);
7712 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7713 struct sched_entity *se, int cpu,
7714 struct sched_entity *parent)
7716 struct rq *rq = cpu_rq(cpu);
7720 init_cfs_rq_runtime(cfs_rq);
7722 tg->cfs_rq[cpu] = cfs_rq;
7725 /* se could be NULL for root_task_group */
7730 se->cfs_rq = &rq->cfs;
7732 se->cfs_rq = parent->my_q;
7735 /* guarantee group entities always have weight */
7736 update_load_set(&se->load, NICE_0_LOAD);
7737 se->parent = parent;
7740 static DEFINE_MUTEX(shares_mutex);
7742 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7745 unsigned long flags;
7748 * We can't change the weight of the root cgroup.
7753 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7755 mutex_lock(&shares_mutex);
7756 if (tg->shares == shares)
7759 tg->shares = shares;
7760 for_each_possible_cpu(i) {
7761 struct rq *rq = cpu_rq(i);
7762 struct sched_entity *se;
7765 /* Propagate contribution to hierarchy */
7766 raw_spin_lock_irqsave(&rq->lock, flags);
7767 for_each_sched_entity(se)
7768 update_cfs_shares(group_cfs_rq(se));
7769 raw_spin_unlock_irqrestore(&rq->lock, flags);
7773 mutex_unlock(&shares_mutex);
7776 #else /* CONFIG_FAIR_GROUP_SCHED */
7778 void free_fair_sched_group(struct task_group *tg) { }
7780 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7785 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7787 #endif /* CONFIG_FAIR_GROUP_SCHED */
7790 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7792 struct sched_entity *se = &task->se;
7793 unsigned int rr_interval = 0;
7796 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7799 if (rq->cfs.load.weight)
7800 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7806 * All the scheduling class methods:
7808 const struct sched_class fair_sched_class = {
7809 .next = &idle_sched_class,
7810 .enqueue_task = enqueue_task_fair,
7811 .dequeue_task = dequeue_task_fair,
7812 .yield_task = yield_task_fair,
7813 .yield_to_task = yield_to_task_fair,
7815 .check_preempt_curr = check_preempt_wakeup,
7817 .pick_next_task = pick_next_task_fair,
7818 .put_prev_task = put_prev_task_fair,
7821 .select_task_rq = select_task_rq_fair,
7822 #ifdef CONFIG_FAIR_GROUP_SCHED
7823 .migrate_task_rq = migrate_task_rq_fair,
7825 .rq_online = rq_online_fair,
7826 .rq_offline = rq_offline_fair,
7828 .task_waking = task_waking_fair,
7831 .set_curr_task = set_curr_task_fair,
7832 .task_tick = task_tick_fair,
7833 .task_fork = task_fork_fair,
7835 .prio_changed = prio_changed_fair,
7836 .switched_from = switched_from_fair,
7837 .switched_to = switched_to_fair,
7839 .get_rr_interval = get_rr_interval_fair,
7841 #ifdef CONFIG_FAIR_GROUP_SCHED
7842 .task_move_group = task_move_group_fair,
7846 #ifdef CONFIG_SCHED_DEBUG
7847 void print_cfs_stats(struct seq_file *m, int cpu)
7849 struct cfs_rq *cfs_rq;
7852 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7853 print_cfs_rq(m, cpu, cfs_rq);
7858 __init void init_sched_fair_class(void)
7861 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7863 #ifdef CONFIG_NO_HZ_COMMON
7864 nohz.next_balance = jiffies;
7865 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7866 cpu_notifier(sched_ilb_notifier, 0);
7869 #ifdef CONFIG_SCHED_HMP
7870 hmp_cpu_mask_setup();
7876 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
7877 static u32 cpufreq_calc_scale(u32 min, u32 max, u32 curr)
7879 u32 result = curr / max;
7883 /* Called when the CPU Frequency is changed.
7884 * Once for each CPU.
7886 static int cpufreq_callback(struct notifier_block *nb,
7887 unsigned long val, void *data)
7889 struct cpufreq_freqs *freq = data;
7890 int cpu = freq->cpu;
7891 struct cpufreq_extents *extents;
7893 if (freq->flags & CPUFREQ_CONST_LOOPS)
7896 if (val != CPUFREQ_POSTCHANGE)
7899 /* if dynamic load scale is disabled, set the load scale to 1.0 */
7900 if (!hmp_data.freqinvar_load_scale_enabled) {
7901 freq_scale[cpu].curr_scale = 1024;
7905 extents = &freq_scale[cpu];
7906 if (extents->flags & SCHED_LOAD_FREQINVAR_SINGLEFREQ) {
7907 /* If our governor was recognised as a single-freq governor,
7910 extents->curr_scale = 1024;
7912 extents->curr_scale = cpufreq_calc_scale(extents->min,
7913 extents->max, freq->new);
7919 /* Called when the CPUFreq governor is changed.
7920 * Only called for the CPUs which are actually changed by the
7923 static int cpufreq_policy_callback(struct notifier_block *nb,
7924 unsigned long event, void *data)
7926 struct cpufreq_policy *policy = data;
7927 struct cpufreq_extents *extents;
7928 int cpu, singleFreq = 0;
7929 static const char performance_governor[] = "performance";
7930 static const char powersave_governor[] = "powersave";
7932 if (event == CPUFREQ_START)
7935 if (event != CPUFREQ_INCOMPATIBLE)
7938 /* CPUFreq governors do not accurately report the range of
7939 * CPU Frequencies they will choose from.
7940 * We recognise performance and powersave governors as
7941 * single-frequency only.
7943 if (!strncmp(policy->governor->name, performance_governor,
7944 strlen(performance_governor)) ||
7945 !strncmp(policy->governor->name, powersave_governor,
7946 strlen(powersave_governor)))
7949 /* Make sure that all CPUs impacted by this policy are
7950 * updated since we will only get a notification when the
7951 * user explicitly changes the policy on a CPU.
7953 for_each_cpu(cpu, policy->cpus) {
7954 extents = &freq_scale[cpu];
7955 extents->max = policy->max >> SCHED_FREQSCALE_SHIFT;
7956 extents->min = policy->min >> SCHED_FREQSCALE_SHIFT;
7957 if (!hmp_data.freqinvar_load_scale_enabled) {
7958 extents->curr_scale = 1024;
7959 } else if (singleFreq) {
7960 extents->flags |= SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7961 extents->curr_scale = 1024;
7963 extents->flags &= ~SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7964 extents->curr_scale = cpufreq_calc_scale(extents->min,
7965 extents->max, policy->cur);
7972 static struct notifier_block cpufreq_notifier = {
7973 .notifier_call = cpufreq_callback,
7975 static struct notifier_block cpufreq_policy_notifier = {
7976 .notifier_call = cpufreq_policy_callback,
7979 static int __init register_sched_cpufreq_notifier(void)
7983 /* init safe defaults since there are no policies at registration */
7984 for (ret = 0; ret < CONFIG_NR_CPUS; ret++) {
7986 freq_scale[ret].max = 1024;
7987 freq_scale[ret].min = 1024;
7988 freq_scale[ret].curr_scale = 1024;
7991 pr_info("sched: registering cpufreq notifiers for scale-invariant loads\n");
7992 ret = cpufreq_register_notifier(&cpufreq_policy_notifier,
7993 CPUFREQ_POLICY_NOTIFIER);
7996 ret = cpufreq_register_notifier(&cpufreq_notifier,
7997 CPUFREQ_TRANSITION_NOTIFIER);
8002 core_initcall(register_sched_cpufreq_notifier);
8003 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */