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 start = max(start, vma->vm_start);
953 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
954 end = min(end, vma->vm_end);
955 pages -= change_prot_numa(vma, start, end);
960 } while (end != vma->vm_end);
965 * It is possible to reach the end of the VMA list but the last few VMAs are
966 * not guaranteed to the vma_migratable. If they are not, we would find the
967 * !migratable VMA on the next scan but not reset the scanner to the start
971 mm->numa_scan_offset = start;
973 reset_ptenuma_scan(p);
974 up_read(&mm->mmap_sem);
978 * Drive the periodic memory faults..
980 void task_tick_numa(struct rq *rq, struct task_struct *curr)
982 struct callback_head *work = &curr->numa_work;
986 * We don't care about NUMA placement if we don't have memory.
988 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
992 * Using runtime rather than walltime has the dual advantage that
993 * we (mostly) drive the selection from busy threads and that the
994 * task needs to have done some actual work before we bother with
997 now = curr->se.sum_exec_runtime;
998 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1000 if (now - curr->node_stamp > period) {
1001 if (!curr->node_stamp)
1002 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1003 curr->node_stamp = now;
1005 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1006 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1007 task_work_add(curr, work, true);
1012 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1015 #endif /* CONFIG_NUMA_BALANCING */
1018 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1020 update_load_add(&cfs_rq->load, se->load.weight);
1021 if (!parent_entity(se))
1022 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1024 if (entity_is_task(se))
1025 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1027 cfs_rq->nr_running++;
1031 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1033 update_load_sub(&cfs_rq->load, se->load.weight);
1034 if (!parent_entity(se))
1035 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1036 if (entity_is_task(se))
1037 list_del_init(&se->group_node);
1038 cfs_rq->nr_running--;
1041 #ifdef CONFIG_FAIR_GROUP_SCHED
1043 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1048 * Use this CPU's actual weight instead of the last load_contribution
1049 * to gain a more accurate current total weight. See
1050 * update_cfs_rq_load_contribution().
1052 tg_weight = atomic64_read(&tg->load_avg);
1053 tg_weight -= cfs_rq->tg_load_contrib;
1054 tg_weight += cfs_rq->load.weight;
1059 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1061 long tg_weight, load, shares;
1063 tg_weight = calc_tg_weight(tg, cfs_rq);
1064 load = cfs_rq->load.weight;
1066 shares = (tg->shares * load);
1068 shares /= tg_weight;
1070 if (shares < MIN_SHARES)
1071 shares = MIN_SHARES;
1072 if (shares > tg->shares)
1073 shares = tg->shares;
1077 # else /* CONFIG_SMP */
1078 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1082 # endif /* CONFIG_SMP */
1083 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1084 unsigned long weight)
1087 /* commit outstanding execution time */
1088 if (cfs_rq->curr == se)
1089 update_curr(cfs_rq);
1090 account_entity_dequeue(cfs_rq, se);
1093 update_load_set(&se->load, weight);
1096 account_entity_enqueue(cfs_rq, se);
1099 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1101 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1103 struct task_group *tg;
1104 struct sched_entity *se;
1108 se = tg->se[cpu_of(rq_of(cfs_rq))];
1109 if (!se || throttled_hierarchy(cfs_rq))
1112 if (likely(se->load.weight == tg->shares))
1115 shares = calc_cfs_shares(cfs_rq, tg);
1117 reweight_entity(cfs_rq_of(se), se, shares);
1119 #else /* CONFIG_FAIR_GROUP_SCHED */
1120 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1123 #endif /* CONFIG_FAIR_GROUP_SCHED */
1125 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1126 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1128 * We choose a half-life close to 1 scheduling period.
1129 * Note: The tables below are dependent on this value.
1131 #define LOAD_AVG_PERIOD 32
1132 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1133 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1135 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1136 static const u32 runnable_avg_yN_inv[] = {
1137 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1138 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1139 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1140 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1141 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1142 0x85aac367, 0x82cd8698,
1146 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1147 * over-estimates when re-combining.
1149 static const u32 runnable_avg_yN_sum[] = {
1150 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1151 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1152 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1157 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1159 static __always_inline u64 decay_load(u64 val, u64 n)
1161 unsigned int local_n;
1165 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1168 /* after bounds checking we can collapse to 32-bit */
1172 * As y^PERIOD = 1/2, we can combine
1173 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1174 * With a look-up table which covers k^n (n<PERIOD)
1176 * To achieve constant time decay_load.
1178 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1179 val >>= local_n / LOAD_AVG_PERIOD;
1180 local_n %= LOAD_AVG_PERIOD;
1183 val *= runnable_avg_yN_inv[local_n];
1184 /* We don't use SRR here since we always want to round down. */
1189 * For updates fully spanning n periods, the contribution to runnable
1190 * average will be: \Sum 1024*y^n
1192 * We can compute this reasonably efficiently by combining:
1193 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1195 static u32 __compute_runnable_contrib(u64 n)
1199 if (likely(n <= LOAD_AVG_PERIOD))
1200 return runnable_avg_yN_sum[n];
1201 else if (unlikely(n >= LOAD_AVG_MAX_N))
1202 return LOAD_AVG_MAX;
1204 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1206 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1207 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1209 n -= LOAD_AVG_PERIOD;
1210 } while (n > LOAD_AVG_PERIOD);
1212 contrib = decay_load(contrib, n);
1213 return contrib + runnable_avg_yN_sum[n];
1216 #ifdef CONFIG_SCHED_HMP
1217 #define HMP_VARIABLE_SCALE_SHIFT 16ULL
1218 struct hmp_global_attr {
1219 struct attribute attr;
1220 ssize_t (*show)(struct kobject *kobj,
1221 struct attribute *attr, char *buf);
1222 ssize_t (*store)(struct kobject *a, struct attribute *b,
1223 const char *c, size_t count);
1225 int (*to_sysfs)(int);
1226 int (*from_sysfs)(int);
1227 ssize_t (*to_sysfs_text)(char *buf, int buf_size);
1230 #define HMP_DATA_SYSFS_MAX 8
1232 struct hmp_data_struct {
1233 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1234 int freqinvar_load_scale_enabled;
1236 int multiplier; /* used to scale the time delta */
1237 struct attribute_group attr_group;
1238 struct attribute *attributes[HMP_DATA_SYSFS_MAX + 1];
1239 struct hmp_global_attr attr[HMP_DATA_SYSFS_MAX];
1242 static u64 hmp_variable_scale_convert(u64 delta);
1243 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1244 /* Frequency-Invariant Load Modification:
1245 * Loads are calculated as in PJT's patch however we also scale the current
1246 * contribution in line with the frequency of the CPU that the task was
1248 * In this version, we use a simple linear scale derived from the maximum
1249 * frequency reported by CPUFreq. As an example:
1251 * Consider that we ran a task for 100% of the previous interval.
1253 * Our CPU was under asynchronous frequency control through one of the
1254 * CPUFreq governors.
1256 * The CPUFreq governor reports that it is able to scale the CPU between
1259 * During the period, the CPU was running at 1GHz.
1261 * In this case, our load contribution for that period is calculated as
1262 * 1 * (number_of_active_microseconds)
1264 * This results in our task being able to accumulate maximum load as normal.
1267 * Consider now that our CPU was executing at 500MHz.
1269 * We now scale the load contribution such that it is calculated as
1270 * 0.5 * (number_of_active_microseconds)
1272 * Our task can only record 50% maximum load during this period.
1274 * This represents the task consuming 50% of the CPU's *possible* compute
1275 * capacity. However the task did consume 100% of the CPU's *available*
1276 * compute capacity which is the value seen by the CPUFreq governor and
1277 * user-side CPU Utilization tools.
1279 * Restricting tracked load to be scaled by the CPU's frequency accurately
1280 * represents the consumption of possible compute capacity and allows the
1281 * HMP migration's simple threshold migration strategy to interact more
1282 * predictably with CPUFreq's asynchronous compute capacity changes.
1284 #define SCHED_FREQSCALE_SHIFT 10
1285 struct cpufreq_extents {
1291 /* Flag set when the governor in use only allows one frequency.
1294 #define SCHED_LOAD_FREQINVAR_SINGLEFREQ 0x01
1296 static struct cpufreq_extents freq_scale[CONFIG_NR_CPUS];
1297 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1298 #endif /* CONFIG_SCHED_HMP */
1300 /* We can represent the historical contribution to runnable average as the
1301 * coefficients of a geometric series. To do this we sub-divide our runnable
1302 * history into segments of approximately 1ms (1024us); label the segment that
1303 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1305 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1307 * (now) (~1ms ago) (~2ms ago)
1309 * Let u_i denote the fraction of p_i that the entity was runnable.
1311 * We then designate the fractions u_i as our co-efficients, yielding the
1312 * following representation of historical load:
1313 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1315 * We choose y based on the with of a reasonably scheduling period, fixing:
1318 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1319 * approximately half as much as the contribution to load within the last ms
1322 * When a period "rolls over" and we have new u_0`, multiplying the previous
1323 * sum again by y is sufficient to update:
1324 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1325 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1327 static __always_inline int __update_entity_runnable_avg(u64 now,
1328 struct sched_avg *sa,
1334 u32 runnable_contrib;
1335 int delta_w, decayed = 0;
1336 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1338 u32 scaled_runnable_contrib;
1340 u32 curr_scale = 1024;
1341 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1343 delta = now - sa->last_runnable_update;
1344 #ifdef CONFIG_SCHED_HMP
1345 delta = hmp_variable_scale_convert(delta);
1348 * This should only happen when time goes backwards, which it
1349 * unfortunately does during sched clock init when we swap over to TSC.
1351 if ((s64)delta < 0) {
1352 sa->last_runnable_update = now;
1357 * Use 1024ns as the unit of measurement since it's a reasonable
1358 * approximation of 1us and fast to compute.
1363 sa->last_runnable_update = now;
1365 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1366 /* retrieve scale factor for load */
1367 if (hmp_data.freqinvar_load_scale_enabled)
1368 curr_scale = freq_scale[cpu].curr_scale;
1369 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1371 /* delta_w is the amount already accumulated against our next period */
1372 delta_w = sa->runnable_avg_period % 1024;
1373 if (delta + delta_w >= 1024) {
1374 /* period roll-over */
1378 * Now that we know we're crossing a period boundary, figure
1379 * out how much from delta we need to complete the current
1380 * period and accrue it.
1382 delta_w = 1024 - delta_w;
1383 /* scale runnable time if necessary */
1384 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1385 scaled_delta_w = (delta_w * curr_scale)
1386 >> SCHED_FREQSCALE_SHIFT;
1388 sa->runnable_avg_sum += scaled_delta_w;
1390 sa->usage_avg_sum += scaled_delta_w;
1393 sa->runnable_avg_sum += delta_w;
1395 sa->usage_avg_sum += delta_w;
1396 #endif /* #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1397 sa->runnable_avg_period += delta_w;
1401 /* Figure out how many additional periods this update spans */
1402 periods = delta / 1024;
1404 /* decay the load we have accumulated so far */
1405 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1407 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1409 sa->usage_avg_sum = decay_load(sa->usage_avg_sum, periods + 1);
1410 /* add the contribution from this period */
1411 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1412 runnable_contrib = __compute_runnable_contrib(periods);
1413 /* Apply load scaling if necessary.
1414 * Note that multiplying the whole series is same as
1415 * multiplying all terms
1417 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1418 scaled_runnable_contrib = (runnable_contrib * curr_scale)
1419 >> SCHED_FREQSCALE_SHIFT;
1421 sa->runnable_avg_sum += scaled_runnable_contrib;
1423 sa->usage_avg_sum += scaled_runnable_contrib;
1426 sa->runnable_avg_sum += runnable_contrib;
1428 sa->usage_avg_sum += runnable_contrib;
1429 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1430 sa->runnable_avg_period += runnable_contrib;
1433 /* Remainder of delta accrued against u_0` */
1434 /* scale if necessary */
1435 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1436 scaled_delta = ((delta * curr_scale) >> SCHED_FREQSCALE_SHIFT);
1438 sa->runnable_avg_sum += scaled_delta;
1440 sa->usage_avg_sum += scaled_delta;
1443 sa->runnable_avg_sum += delta;
1445 sa->usage_avg_sum += delta;
1446 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1447 sa->runnable_avg_period += delta;
1452 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1453 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1455 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1456 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1458 decays -= se->avg.decay_count;
1460 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1461 se->avg.decay_count = 0;
1465 #ifdef CONFIG_FAIR_GROUP_SCHED
1466 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1469 struct task_group *tg = cfs_rq->tg;
1472 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1473 tg_contrib -= cfs_rq->tg_load_contrib;
1475 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1476 atomic64_add(tg_contrib, &tg->load_avg);
1477 cfs_rq->tg_load_contrib += tg_contrib;
1482 * Aggregate cfs_rq runnable averages into an equivalent task_group
1483 * representation for computing load contributions.
1485 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1486 struct cfs_rq *cfs_rq)
1488 struct task_group *tg = cfs_rq->tg;
1489 long contrib, usage_contrib;
1491 /* The fraction of a cpu used by this cfs_rq */
1492 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1493 sa->runnable_avg_period + 1);
1494 contrib -= cfs_rq->tg_runnable_contrib;
1496 usage_contrib = div_u64(sa->usage_avg_sum << NICE_0_SHIFT,
1497 sa->runnable_avg_period + 1);
1498 usage_contrib -= cfs_rq->tg_usage_contrib;
1501 * contrib/usage at this point represent deltas, only update if they
1504 if ((abs(contrib) > cfs_rq->tg_runnable_contrib / 64) ||
1505 (abs(usage_contrib) > cfs_rq->tg_usage_contrib / 64)) {
1506 atomic_add(contrib, &tg->runnable_avg);
1507 cfs_rq->tg_runnable_contrib += contrib;
1509 atomic_add(usage_contrib, &tg->usage_avg);
1510 cfs_rq->tg_usage_contrib += usage_contrib;
1514 static inline void __update_group_entity_contrib(struct sched_entity *se)
1516 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1517 struct task_group *tg = cfs_rq->tg;
1522 contrib = cfs_rq->tg_load_contrib * tg->shares;
1523 se->avg.load_avg_contrib = div64_u64(contrib,
1524 atomic64_read(&tg->load_avg) + 1);
1527 * For group entities we need to compute a correction term in the case
1528 * that they are consuming <1 cpu so that we would contribute the same
1529 * load as a task of equal weight.
1531 * Explicitly co-ordinating this measurement would be expensive, but
1532 * fortunately the sum of each cpus contribution forms a usable
1533 * lower-bound on the true value.
1535 * Consider the aggregate of 2 contributions. Either they are disjoint
1536 * (and the sum represents true value) or they are disjoint and we are
1537 * understating by the aggregate of their overlap.
1539 * Extending this to N cpus, for a given overlap, the maximum amount we
1540 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1541 * cpus that overlap for this interval and w_i is the interval width.
1543 * On a small machine; the first term is well-bounded which bounds the
1544 * total error since w_i is a subset of the period. Whereas on a
1545 * larger machine, while this first term can be larger, if w_i is the
1546 * of consequential size guaranteed to see n_i*w_i quickly converge to
1547 * our upper bound of 1-cpu.
1549 runnable_avg = atomic_read(&tg->runnable_avg);
1550 if (runnable_avg < NICE_0_LOAD) {
1551 se->avg.load_avg_contrib *= runnable_avg;
1552 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1556 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1557 int force_update) {}
1558 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1559 struct cfs_rq *cfs_rq) {}
1560 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1563 static inline void __update_task_entity_contrib(struct sched_entity *se)
1567 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1568 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1569 contrib /= (se->avg.runnable_avg_period + 1);
1570 se->avg.load_avg_contrib = scale_load(contrib);
1571 trace_sched_task_load_contrib(task_of(se), se->avg.load_avg_contrib);
1572 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1573 contrib /= (se->avg.runnable_avg_period + 1);
1574 se->avg.load_avg_ratio = scale_load(contrib);
1575 trace_sched_task_runnable_ratio(task_of(se), se->avg.load_avg_ratio);
1578 /* Compute the current contribution to load_avg by se, return any delta */
1579 static long __update_entity_load_avg_contrib(struct sched_entity *se, long *ratio)
1581 long old_contrib = se->avg.load_avg_contrib;
1582 long old_ratio = se->avg.load_avg_ratio;
1584 if (entity_is_task(se)) {
1585 __update_task_entity_contrib(se);
1587 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1588 __update_group_entity_contrib(se);
1592 *ratio = se->avg.load_avg_ratio - old_ratio;
1593 return se->avg.load_avg_contrib - old_contrib;
1596 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1599 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1600 cfs_rq->blocked_load_avg -= load_contrib;
1602 cfs_rq->blocked_load_avg = 0;
1605 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1607 /* Update a sched_entity's runnable average */
1608 static inline void update_entity_load_avg(struct sched_entity *se,
1611 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1612 long contrib_delta, ratio_delta;
1614 int cpu = -1; /* not used in normal case */
1616 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1617 cpu = cfs_rq->rq->cpu;
1620 * For a group entity we need to use their owned cfs_rq_clock_task() in
1621 * case they are the parent of a throttled hierarchy.
1623 if (entity_is_task(se))
1624 now = cfs_rq_clock_task(cfs_rq);
1626 now = cfs_rq_clock_task(group_cfs_rq(se));
1628 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1629 cfs_rq->curr == se, cpu))
1632 contrib_delta = __update_entity_load_avg_contrib(se, &ratio_delta);
1638 cfs_rq->runnable_load_avg += contrib_delta;
1639 rq_of(cfs_rq)->avg.load_avg_ratio += ratio_delta;
1641 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1646 * Decay the load contributed by all blocked children and account this so that
1647 * their contribution may appropriately discounted when they wake up.
1649 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1651 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1654 decays = now - cfs_rq->last_decay;
1655 if (!decays && !force_update)
1658 if (atomic64_read(&cfs_rq->removed_load)) {
1659 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1660 subtract_blocked_load_contrib(cfs_rq, removed_load);
1664 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1666 atomic64_add(decays, &cfs_rq->decay_counter);
1667 cfs_rq->last_decay = now;
1670 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1673 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1675 int cpu = -1; /* not used in normal case */
1677 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1680 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1682 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1683 trace_sched_rq_runnable_ratio(cpu_of(rq), rq->avg.load_avg_ratio);
1684 trace_sched_rq_runnable_load(cpu_of(rq), rq->cfs.runnable_load_avg);
1685 trace_sched_rq_nr_running(cpu_of(rq), rq->nr_running, rq->nr_iowait.counter);
1688 /* Add the load generated by se into cfs_rq's child load-average */
1689 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1690 struct sched_entity *se,
1694 * We track migrations using entity decay_count <= 0, on a wake-up
1695 * migration we use a negative decay count to track the remote decays
1696 * accumulated while sleeping.
1698 if (unlikely(se->avg.decay_count <= 0)) {
1699 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1700 if (se->avg.decay_count) {
1702 * In a wake-up migration we have to approximate the
1703 * time sleeping. This is because we can't synchronize
1704 * clock_task between the two cpus, and it is not
1705 * guaranteed to be read-safe. Instead, we can
1706 * approximate this using our carried decays, which are
1707 * explicitly atomically readable.
1709 se->avg.last_runnable_update -= (-se->avg.decay_count)
1711 update_entity_load_avg(se, 0);
1712 /* Indicate that we're now synchronized and on-rq */
1713 se->avg.decay_count = 0;
1717 __synchronize_entity_decay(se);
1720 /* migrated tasks did not contribute to our blocked load */
1722 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1723 update_entity_load_avg(se, 0);
1726 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1727 rq_of(cfs_rq)->avg.load_avg_ratio += se->avg.load_avg_ratio;
1729 /* we force update consideration on load-balancer moves */
1730 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1734 * Remove se's load from this cfs_rq child load-average, if the entity is
1735 * transitioning to a blocked state we track its projected decay using
1738 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1739 struct sched_entity *se,
1742 update_entity_load_avg(se, 1);
1743 /* we force update consideration on load-balancer moves */
1744 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1746 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1747 rq_of(cfs_rq)->avg.load_avg_ratio -= se->avg.load_avg_ratio;
1750 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1751 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1752 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1756 * Update the rq's load with the elapsed running time before entering
1757 * idle. if the last scheduled task is not a CFS task, idle_enter will
1758 * be the only way to update the runnable statistic.
1760 void idle_enter_fair(struct rq *this_rq)
1762 update_rq_runnable_avg(this_rq, 1);
1766 * Update the rq's load with the elapsed idle time before a task is
1767 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1768 * be the only way to update the runnable statistic.
1770 void idle_exit_fair(struct rq *this_rq)
1772 update_rq_runnable_avg(this_rq, 0);
1776 static inline void update_entity_load_avg(struct sched_entity *se,
1777 int update_cfs_rq) {}
1778 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1779 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1780 struct sched_entity *se,
1782 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1783 struct sched_entity *se,
1785 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1786 int force_update) {}
1789 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1791 #ifdef CONFIG_SCHEDSTATS
1792 struct task_struct *tsk = NULL;
1794 if (entity_is_task(se))
1797 if (se->statistics.sleep_start) {
1798 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1803 if (unlikely(delta > se->statistics.sleep_max))
1804 se->statistics.sleep_max = delta;
1806 se->statistics.sleep_start = 0;
1807 se->statistics.sum_sleep_runtime += delta;
1810 account_scheduler_latency(tsk, delta >> 10, 1);
1811 trace_sched_stat_sleep(tsk, delta);
1814 if (se->statistics.block_start) {
1815 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1820 if (unlikely(delta > se->statistics.block_max))
1821 se->statistics.block_max = delta;
1823 se->statistics.block_start = 0;
1824 se->statistics.sum_sleep_runtime += delta;
1827 if (tsk->in_iowait) {
1828 se->statistics.iowait_sum += delta;
1829 se->statistics.iowait_count++;
1830 trace_sched_stat_iowait(tsk, delta);
1833 trace_sched_stat_blocked(tsk, delta);
1836 * Blocking time is in units of nanosecs, so shift by
1837 * 20 to get a milliseconds-range estimation of the
1838 * amount of time that the task spent sleeping:
1840 if (unlikely(prof_on == SLEEP_PROFILING)) {
1841 profile_hits(SLEEP_PROFILING,
1842 (void *)get_wchan(tsk),
1845 account_scheduler_latency(tsk, delta >> 10, 0);
1851 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1853 #ifdef CONFIG_SCHED_DEBUG
1854 s64 d = se->vruntime - cfs_rq->min_vruntime;
1859 if (d > 3*sysctl_sched_latency)
1860 schedstat_inc(cfs_rq, nr_spread_over);
1865 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1867 u64 vruntime = cfs_rq->min_vruntime;
1870 * The 'current' period is already promised to the current tasks,
1871 * however the extra weight of the new task will slow them down a
1872 * little, place the new task so that it fits in the slot that
1873 * stays open at the end.
1875 if (initial && sched_feat(START_DEBIT))
1876 vruntime += sched_vslice(cfs_rq, se);
1878 /* sleeps up to a single latency don't count. */
1880 unsigned long thresh = sysctl_sched_latency;
1883 * Halve their sleep time's effect, to allow
1884 * for a gentler effect of sleepers:
1886 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1892 /* ensure we never gain time by being placed backwards. */
1893 se->vruntime = max_vruntime(se->vruntime, vruntime);
1896 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1899 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1902 * Update the normalized vruntime before updating min_vruntime
1903 * through callig update_curr().
1905 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1906 se->vruntime += cfs_rq->min_vruntime;
1909 * Update run-time statistics of the 'current'.
1911 update_curr(cfs_rq);
1912 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1913 account_entity_enqueue(cfs_rq, se);
1914 update_cfs_shares(cfs_rq);
1916 if (flags & ENQUEUE_WAKEUP) {
1917 place_entity(cfs_rq, se, 0);
1918 enqueue_sleeper(cfs_rq, se);
1921 update_stats_enqueue(cfs_rq, se);
1922 check_spread(cfs_rq, se);
1923 if (se != cfs_rq->curr)
1924 __enqueue_entity(cfs_rq, se);
1927 if (cfs_rq->nr_running == 1) {
1928 list_add_leaf_cfs_rq(cfs_rq);
1929 check_enqueue_throttle(cfs_rq);
1933 static void __clear_buddies_last(struct sched_entity *se)
1935 for_each_sched_entity(se) {
1936 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1937 if (cfs_rq->last == se)
1938 cfs_rq->last = NULL;
1944 static void __clear_buddies_next(struct sched_entity *se)
1946 for_each_sched_entity(se) {
1947 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1948 if (cfs_rq->next == se)
1949 cfs_rq->next = NULL;
1955 static void __clear_buddies_skip(struct sched_entity *se)
1957 for_each_sched_entity(se) {
1958 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1959 if (cfs_rq->skip == se)
1960 cfs_rq->skip = NULL;
1966 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1968 if (cfs_rq->last == se)
1969 __clear_buddies_last(se);
1971 if (cfs_rq->next == se)
1972 __clear_buddies_next(se);
1974 if (cfs_rq->skip == se)
1975 __clear_buddies_skip(se);
1978 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1981 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1984 * Update run-time statistics of the 'current'.
1986 update_curr(cfs_rq);
1987 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1989 update_stats_dequeue(cfs_rq, se);
1990 if (flags & DEQUEUE_SLEEP) {
1991 #ifdef CONFIG_SCHEDSTATS
1992 if (entity_is_task(se)) {
1993 struct task_struct *tsk = task_of(se);
1995 if (tsk->state & TASK_INTERRUPTIBLE)
1996 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1997 if (tsk->state & TASK_UNINTERRUPTIBLE)
1998 se->statistics.block_start = rq_of(cfs_rq)->clock;
2003 clear_buddies(cfs_rq, se);
2005 if (se != cfs_rq->curr)
2006 __dequeue_entity(cfs_rq, se);
2008 account_entity_dequeue(cfs_rq, se);
2011 * Normalize the entity after updating the min_vruntime because the
2012 * update can refer to the ->curr item and we need to reflect this
2013 * movement in our normalized position.
2015 if (!(flags & DEQUEUE_SLEEP))
2016 se->vruntime -= cfs_rq->min_vruntime;
2018 /* return excess runtime on last dequeue */
2019 return_cfs_rq_runtime(cfs_rq);
2021 update_min_vruntime(cfs_rq);
2022 update_cfs_shares(cfs_rq);
2026 * Preempt the current task with a newly woken task if needed:
2029 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2031 unsigned long ideal_runtime, delta_exec;
2032 struct sched_entity *se;
2035 ideal_runtime = sched_slice(cfs_rq, curr);
2036 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2037 if (delta_exec > ideal_runtime) {
2038 resched_task(rq_of(cfs_rq)->curr);
2040 * The current task ran long enough, ensure it doesn't get
2041 * re-elected due to buddy favours.
2043 clear_buddies(cfs_rq, curr);
2048 * Ensure that a task that missed wakeup preemption by a
2049 * narrow margin doesn't have to wait for a full slice.
2050 * This also mitigates buddy induced latencies under load.
2052 if (delta_exec < sysctl_sched_min_granularity)
2055 se = __pick_first_entity(cfs_rq);
2056 delta = curr->vruntime - se->vruntime;
2061 if (delta > ideal_runtime)
2062 resched_task(rq_of(cfs_rq)->curr);
2066 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2068 /* 'current' is not kept within the tree. */
2071 * Any task has to be enqueued before it get to execute on
2072 * a CPU. So account for the time it spent waiting on the
2075 update_stats_wait_end(cfs_rq, se);
2076 __dequeue_entity(cfs_rq, se);
2077 update_entity_load_avg(se, 1);
2080 update_stats_curr_start(cfs_rq, se);
2082 #ifdef CONFIG_SCHEDSTATS
2084 * Track our maximum slice length, if the CPU's load is at
2085 * least twice that of our own weight (i.e. dont track it
2086 * when there are only lesser-weight tasks around):
2088 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2089 se->statistics.slice_max = max(se->statistics.slice_max,
2090 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2093 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2097 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2100 * Pick the next process, keeping these things in mind, in this order:
2101 * 1) keep things fair between processes/task groups
2102 * 2) pick the "next" process, since someone really wants that to run
2103 * 3) pick the "last" process, for cache locality
2104 * 4) do not run the "skip" process, if something else is available
2106 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2108 struct sched_entity *se = __pick_first_entity(cfs_rq);
2109 struct sched_entity *left = se;
2112 * Avoid running the skip buddy, if running something else can
2113 * be done without getting too unfair.
2115 if (cfs_rq->skip == se) {
2116 struct sched_entity *second = __pick_next_entity(se);
2117 if (second && wakeup_preempt_entity(second, left) < 1)
2122 * Prefer last buddy, try to return the CPU to a preempted task.
2124 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2128 * Someone really wants this to run. If it's not unfair, run it.
2130 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2133 clear_buddies(cfs_rq, se);
2138 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2140 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2143 * If still on the runqueue then deactivate_task()
2144 * was not called and update_curr() has to be done:
2147 update_curr(cfs_rq);
2149 /* throttle cfs_rqs exceeding runtime */
2150 check_cfs_rq_runtime(cfs_rq);
2152 check_spread(cfs_rq, prev);
2154 update_stats_wait_start(cfs_rq, prev);
2155 /* Put 'current' back into the tree. */
2156 __enqueue_entity(cfs_rq, prev);
2157 /* in !on_rq case, update occurred at dequeue */
2158 update_entity_load_avg(prev, 1);
2160 cfs_rq->curr = NULL;
2164 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2167 * Update run-time statistics of the 'current'.
2169 update_curr(cfs_rq);
2172 * Ensure that runnable average is periodically updated.
2174 update_entity_load_avg(curr, 1);
2175 update_cfs_rq_blocked_load(cfs_rq, 1);
2177 #ifdef CONFIG_SCHED_HRTICK
2179 * queued ticks are scheduled to match the slice, so don't bother
2180 * validating it and just reschedule.
2183 resched_task(rq_of(cfs_rq)->curr);
2187 * don't let the period tick interfere with the hrtick preemption
2189 if (!sched_feat(DOUBLE_TICK) &&
2190 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2194 if (cfs_rq->nr_running > 1)
2195 check_preempt_tick(cfs_rq, curr);
2199 /**************************************************
2200 * CFS bandwidth control machinery
2203 #ifdef CONFIG_CFS_BANDWIDTH
2205 #ifdef HAVE_JUMP_LABEL
2206 static struct static_key __cfs_bandwidth_used;
2208 static inline bool cfs_bandwidth_used(void)
2210 return static_key_false(&__cfs_bandwidth_used);
2213 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2215 /* only need to count groups transitioning between enabled/!enabled */
2216 if (enabled && !was_enabled)
2217 static_key_slow_inc(&__cfs_bandwidth_used);
2218 else if (!enabled && was_enabled)
2219 static_key_slow_dec(&__cfs_bandwidth_used);
2221 #else /* HAVE_JUMP_LABEL */
2222 static bool cfs_bandwidth_used(void)
2227 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2228 #endif /* HAVE_JUMP_LABEL */
2231 * default period for cfs group bandwidth.
2232 * default: 0.1s, units: nanoseconds
2234 static inline u64 default_cfs_period(void)
2236 return 100000000ULL;
2239 static inline u64 sched_cfs_bandwidth_slice(void)
2241 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2245 * Replenish runtime according to assigned quota and update expiration time.
2246 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2247 * additional synchronization around rq->lock.
2249 * requires cfs_b->lock
2251 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2255 if (cfs_b->quota == RUNTIME_INF)
2258 now = sched_clock_cpu(smp_processor_id());
2259 cfs_b->runtime = cfs_b->quota;
2260 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2263 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2265 return &tg->cfs_bandwidth;
2268 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2269 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2271 if (unlikely(cfs_rq->throttle_count))
2272 return cfs_rq->throttled_clock_task;
2274 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2277 /* returns 0 on failure to allocate runtime */
2278 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2280 struct task_group *tg = cfs_rq->tg;
2281 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2282 u64 amount = 0, min_amount, expires;
2284 /* note: this is a positive sum as runtime_remaining <= 0 */
2285 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2287 raw_spin_lock(&cfs_b->lock);
2288 if (cfs_b->quota == RUNTIME_INF)
2289 amount = min_amount;
2292 * If the bandwidth pool has become inactive, then at least one
2293 * period must have elapsed since the last consumption.
2294 * Refresh the global state and ensure bandwidth timer becomes
2297 if (!cfs_b->timer_active) {
2298 __refill_cfs_bandwidth_runtime(cfs_b);
2299 __start_cfs_bandwidth(cfs_b);
2302 if (cfs_b->runtime > 0) {
2303 amount = min(cfs_b->runtime, min_amount);
2304 cfs_b->runtime -= amount;
2308 expires = cfs_b->runtime_expires;
2309 raw_spin_unlock(&cfs_b->lock);
2311 cfs_rq->runtime_remaining += amount;
2313 * we may have advanced our local expiration to account for allowed
2314 * spread between our sched_clock and the one on which runtime was
2317 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2318 cfs_rq->runtime_expires = expires;
2320 return cfs_rq->runtime_remaining > 0;
2324 * Note: This depends on the synchronization provided by sched_clock and the
2325 * fact that rq->clock snapshots this value.
2327 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2329 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2330 struct rq *rq = rq_of(cfs_rq);
2332 /* if the deadline is ahead of our clock, nothing to do */
2333 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2336 if (cfs_rq->runtime_remaining < 0)
2340 * If the local deadline has passed we have to consider the
2341 * possibility that our sched_clock is 'fast' and the global deadline
2342 * has not truly expired.
2344 * Fortunately we can check determine whether this the case by checking
2345 * whether the global deadline has advanced.
2348 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2349 /* extend local deadline, drift is bounded above by 2 ticks */
2350 cfs_rq->runtime_expires += TICK_NSEC;
2352 /* global deadline is ahead, expiration has passed */
2353 cfs_rq->runtime_remaining = 0;
2357 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2358 unsigned long delta_exec)
2360 /* dock delta_exec before expiring quota (as it could span periods) */
2361 cfs_rq->runtime_remaining -= delta_exec;
2362 expire_cfs_rq_runtime(cfs_rq);
2364 if (likely(cfs_rq->runtime_remaining > 0))
2368 * if we're unable to extend our runtime we resched so that the active
2369 * hierarchy can be throttled
2371 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2372 resched_task(rq_of(cfs_rq)->curr);
2375 static __always_inline
2376 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2378 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2381 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2384 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2386 return cfs_bandwidth_used() && cfs_rq->throttled;
2389 /* check whether cfs_rq, or any parent, is throttled */
2390 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2392 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2396 * Ensure that neither of the group entities corresponding to src_cpu or
2397 * dest_cpu are members of a throttled hierarchy when performing group
2398 * load-balance operations.
2400 static inline int throttled_lb_pair(struct task_group *tg,
2401 int src_cpu, int dest_cpu)
2403 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2405 src_cfs_rq = tg->cfs_rq[src_cpu];
2406 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2408 return throttled_hierarchy(src_cfs_rq) ||
2409 throttled_hierarchy(dest_cfs_rq);
2412 /* updated child weight may affect parent so we have to do this bottom up */
2413 static int tg_unthrottle_up(struct task_group *tg, void *data)
2415 struct rq *rq = data;
2416 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2418 cfs_rq->throttle_count--;
2420 if (!cfs_rq->throttle_count) {
2421 /* adjust cfs_rq_clock_task() */
2422 cfs_rq->throttled_clock_task_time += rq->clock_task -
2423 cfs_rq->throttled_clock_task;
2430 static int tg_throttle_down(struct task_group *tg, void *data)
2432 struct rq *rq = data;
2433 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2435 /* group is entering throttled state, stop time */
2436 if (!cfs_rq->throttle_count)
2437 cfs_rq->throttled_clock_task = rq->clock_task;
2438 cfs_rq->throttle_count++;
2443 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2445 struct rq *rq = rq_of(cfs_rq);
2446 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2447 struct sched_entity *se;
2448 long task_delta, dequeue = 1;
2450 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2452 /* freeze hierarchy runnable averages while throttled */
2454 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2457 task_delta = cfs_rq->h_nr_running;
2458 for_each_sched_entity(se) {
2459 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2460 /* throttled entity or throttle-on-deactivate */
2465 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2466 qcfs_rq->h_nr_running -= task_delta;
2468 if (qcfs_rq->load.weight)
2473 rq->nr_running -= task_delta;
2475 cfs_rq->throttled = 1;
2476 cfs_rq->throttled_clock = rq->clock;
2477 raw_spin_lock(&cfs_b->lock);
2478 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2479 raw_spin_unlock(&cfs_b->lock);
2482 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2484 struct rq *rq = rq_of(cfs_rq);
2485 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2486 struct sched_entity *se;
2490 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2492 cfs_rq->throttled = 0;
2493 raw_spin_lock(&cfs_b->lock);
2494 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2495 list_del_rcu(&cfs_rq->throttled_list);
2496 raw_spin_unlock(&cfs_b->lock);
2498 update_rq_clock(rq);
2499 /* update hierarchical throttle state */
2500 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2502 if (!cfs_rq->load.weight)
2505 task_delta = cfs_rq->h_nr_running;
2506 for_each_sched_entity(se) {
2510 cfs_rq = cfs_rq_of(se);
2512 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2513 cfs_rq->h_nr_running += task_delta;
2515 if (cfs_rq_throttled(cfs_rq))
2520 rq->nr_running += task_delta;
2522 /* determine whether we need to wake up potentially idle cpu */
2523 if (rq->curr == rq->idle && rq->cfs.nr_running)
2524 resched_task(rq->curr);
2527 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2528 u64 remaining, u64 expires)
2530 struct cfs_rq *cfs_rq;
2531 u64 runtime = remaining;
2534 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2536 struct rq *rq = rq_of(cfs_rq);
2538 raw_spin_lock(&rq->lock);
2539 if (!cfs_rq_throttled(cfs_rq))
2542 runtime = -cfs_rq->runtime_remaining + 1;
2543 if (runtime > remaining)
2544 runtime = remaining;
2545 remaining -= runtime;
2547 cfs_rq->runtime_remaining += runtime;
2548 cfs_rq->runtime_expires = expires;
2550 /* we check whether we're throttled above */
2551 if (cfs_rq->runtime_remaining > 0)
2552 unthrottle_cfs_rq(cfs_rq);
2555 raw_spin_unlock(&rq->lock);
2566 * Responsible for refilling a task_group's bandwidth and unthrottling its
2567 * cfs_rqs as appropriate. If there has been no activity within the last
2568 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2569 * used to track this state.
2571 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2573 u64 runtime, runtime_expires;
2574 int idle = 1, throttled;
2576 raw_spin_lock(&cfs_b->lock);
2577 /* no need to continue the timer with no bandwidth constraint */
2578 if (cfs_b->quota == RUNTIME_INF)
2581 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2582 /* idle depends on !throttled (for the case of a large deficit) */
2583 idle = cfs_b->idle && !throttled;
2584 cfs_b->nr_periods += overrun;
2586 /* if we're going inactive then everything else can be deferred */
2590 __refill_cfs_bandwidth_runtime(cfs_b);
2593 /* mark as potentially idle for the upcoming period */
2598 /* account preceding periods in which throttling occurred */
2599 cfs_b->nr_throttled += overrun;
2602 * There are throttled entities so we must first use the new bandwidth
2603 * to unthrottle them before making it generally available. This
2604 * ensures that all existing debts will be paid before a new cfs_rq is
2607 runtime = cfs_b->runtime;
2608 runtime_expires = cfs_b->runtime_expires;
2612 * This check is repeated as we are holding onto the new bandwidth
2613 * while we unthrottle. This can potentially race with an unthrottled
2614 * group trying to acquire new bandwidth from the global pool.
2616 while (throttled && runtime > 0) {
2617 raw_spin_unlock(&cfs_b->lock);
2618 /* we can't nest cfs_b->lock while distributing bandwidth */
2619 runtime = distribute_cfs_runtime(cfs_b, runtime,
2621 raw_spin_lock(&cfs_b->lock);
2623 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2626 /* return (any) remaining runtime */
2627 cfs_b->runtime = runtime;
2629 * While we are ensured activity in the period following an
2630 * unthrottle, this also covers the case in which the new bandwidth is
2631 * insufficient to cover the existing bandwidth deficit. (Forcing the
2632 * timer to remain active while there are any throttled entities.)
2637 cfs_b->timer_active = 0;
2638 raw_spin_unlock(&cfs_b->lock);
2643 /* a cfs_rq won't donate quota below this amount */
2644 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2645 /* minimum remaining period time to redistribute slack quota */
2646 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2647 /* how long we wait to gather additional slack before distributing */
2648 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2650 /* are we near the end of the current quota period? */
2651 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2653 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2656 /* if the call-back is running a quota refresh is already occurring */
2657 if (hrtimer_callback_running(refresh_timer))
2660 /* is a quota refresh about to occur? */
2661 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2662 if (remaining < min_expire)
2668 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2670 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2672 /* if there's a quota refresh soon don't bother with slack */
2673 if (runtime_refresh_within(cfs_b, min_left))
2676 start_bandwidth_timer(&cfs_b->slack_timer,
2677 ns_to_ktime(cfs_bandwidth_slack_period));
2680 /* we know any runtime found here is valid as update_curr() precedes return */
2681 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2683 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2684 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2686 if (slack_runtime <= 0)
2689 raw_spin_lock(&cfs_b->lock);
2690 if (cfs_b->quota != RUNTIME_INF &&
2691 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2692 cfs_b->runtime += slack_runtime;
2694 /* we are under rq->lock, defer unthrottling using a timer */
2695 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2696 !list_empty(&cfs_b->throttled_cfs_rq))
2697 start_cfs_slack_bandwidth(cfs_b);
2699 raw_spin_unlock(&cfs_b->lock);
2701 /* even if it's not valid for return we don't want to try again */
2702 cfs_rq->runtime_remaining -= slack_runtime;
2705 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2707 if (!cfs_bandwidth_used())
2710 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2713 __return_cfs_rq_runtime(cfs_rq);
2717 * This is done with a timer (instead of inline with bandwidth return) since
2718 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2720 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2722 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2725 /* confirm we're still not at a refresh boundary */
2726 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2729 raw_spin_lock(&cfs_b->lock);
2730 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2731 runtime = cfs_b->runtime;
2734 expires = cfs_b->runtime_expires;
2735 raw_spin_unlock(&cfs_b->lock);
2740 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2742 raw_spin_lock(&cfs_b->lock);
2743 if (expires == cfs_b->runtime_expires)
2744 cfs_b->runtime = runtime;
2745 raw_spin_unlock(&cfs_b->lock);
2749 * When a group wakes up we want to make sure that its quota is not already
2750 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2751 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2753 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2755 if (!cfs_bandwidth_used())
2758 /* an active group must be handled by the update_curr()->put() path */
2759 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2762 /* ensure the group is not already throttled */
2763 if (cfs_rq_throttled(cfs_rq))
2766 /* update runtime allocation */
2767 account_cfs_rq_runtime(cfs_rq, 0);
2768 if (cfs_rq->runtime_remaining <= 0)
2769 throttle_cfs_rq(cfs_rq);
2772 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2773 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2775 if (!cfs_bandwidth_used())
2778 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2782 * it's possible for a throttled entity to be forced into a running
2783 * state (e.g. set_curr_task), in this case we're finished.
2785 if (cfs_rq_throttled(cfs_rq))
2788 throttle_cfs_rq(cfs_rq);
2791 static inline u64 default_cfs_period(void);
2792 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2793 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2795 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2797 struct cfs_bandwidth *cfs_b =
2798 container_of(timer, struct cfs_bandwidth, slack_timer);
2799 do_sched_cfs_slack_timer(cfs_b);
2801 return HRTIMER_NORESTART;
2804 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2806 struct cfs_bandwidth *cfs_b =
2807 container_of(timer, struct cfs_bandwidth, period_timer);
2813 now = hrtimer_cb_get_time(timer);
2814 overrun = hrtimer_forward(timer, now, cfs_b->period);
2819 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2822 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2825 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2827 raw_spin_lock_init(&cfs_b->lock);
2829 cfs_b->quota = RUNTIME_INF;
2830 cfs_b->period = ns_to_ktime(default_cfs_period());
2832 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2833 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2834 cfs_b->period_timer.function = sched_cfs_period_timer;
2835 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2836 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2839 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2841 cfs_rq->runtime_enabled = 0;
2842 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2845 /* requires cfs_b->lock, may release to reprogram timer */
2846 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2849 * The timer may be active because we're trying to set a new bandwidth
2850 * period or because we're racing with the tear-down path
2851 * (timer_active==0 becomes visible before the hrtimer call-back
2852 * terminates). In either case we ensure that it's re-programmed
2854 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2855 raw_spin_unlock(&cfs_b->lock);
2856 /* ensure cfs_b->lock is available while we wait */
2857 hrtimer_cancel(&cfs_b->period_timer);
2859 raw_spin_lock(&cfs_b->lock);
2860 /* if someone else restarted the timer then we're done */
2861 if (cfs_b->timer_active)
2865 cfs_b->timer_active = 1;
2866 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2869 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2871 hrtimer_cancel(&cfs_b->period_timer);
2872 hrtimer_cancel(&cfs_b->slack_timer);
2875 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2877 struct cfs_rq *cfs_rq;
2879 for_each_leaf_cfs_rq(rq, cfs_rq) {
2880 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2882 if (!cfs_rq->runtime_enabled)
2886 * clock_task is not advancing so we just need to make sure
2887 * there's some valid quota amount
2889 cfs_rq->runtime_remaining = cfs_b->quota;
2890 if (cfs_rq_throttled(cfs_rq))
2891 unthrottle_cfs_rq(cfs_rq);
2895 #else /* CONFIG_CFS_BANDWIDTH */
2896 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2898 return rq_of(cfs_rq)->clock_task;
2901 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2902 unsigned long delta_exec) {}
2903 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2904 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2905 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2907 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2912 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2917 static inline int throttled_lb_pair(struct task_group *tg,
2918 int src_cpu, int dest_cpu)
2923 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2925 #ifdef CONFIG_FAIR_GROUP_SCHED
2926 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2929 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2933 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2934 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2936 #endif /* CONFIG_CFS_BANDWIDTH */
2938 /**************************************************
2939 * CFS operations on tasks:
2942 #ifdef CONFIG_SCHED_HRTICK
2943 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2945 struct sched_entity *se = &p->se;
2946 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2948 WARN_ON(task_rq(p) != rq);
2950 if (cfs_rq->nr_running > 1) {
2951 u64 slice = sched_slice(cfs_rq, se);
2952 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2953 s64 delta = slice - ran;
2962 * Don't schedule slices shorter than 10000ns, that just
2963 * doesn't make sense. Rely on vruntime for fairness.
2966 delta = max_t(s64, 10000LL, delta);
2968 hrtick_start(rq, delta);
2973 * called from enqueue/dequeue and updates the hrtick when the
2974 * current task is from our class and nr_running is low enough
2977 static void hrtick_update(struct rq *rq)
2979 struct task_struct *curr = rq->curr;
2981 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2984 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2985 hrtick_start_fair(rq, curr);
2987 #else /* !CONFIG_SCHED_HRTICK */
2989 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2993 static inline void hrtick_update(struct rq *rq)
2999 * The enqueue_task method is called before nr_running is
3000 * increased. Here we update the fair scheduling stats and
3001 * then put the task into the rbtree:
3004 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3006 struct cfs_rq *cfs_rq;
3007 struct sched_entity *se = &p->se;
3009 for_each_sched_entity(se) {
3012 cfs_rq = cfs_rq_of(se);
3013 enqueue_entity(cfs_rq, se, flags);
3016 * end evaluation on encountering a throttled cfs_rq
3018 * note: in the case of encountering a throttled cfs_rq we will
3019 * post the final h_nr_running increment below.
3021 if (cfs_rq_throttled(cfs_rq))
3023 cfs_rq->h_nr_running++;
3025 flags = ENQUEUE_WAKEUP;
3028 for_each_sched_entity(se) {
3029 cfs_rq = cfs_rq_of(se);
3030 cfs_rq->h_nr_running++;
3032 if (cfs_rq_throttled(cfs_rq))
3035 update_cfs_shares(cfs_rq);
3036 update_entity_load_avg(se, 1);
3040 update_rq_runnable_avg(rq, rq->nr_running);
3046 static void set_next_buddy(struct sched_entity *se);
3049 * The dequeue_task method is called before nr_running is
3050 * decreased. We remove the task from the rbtree and
3051 * update the fair scheduling stats:
3053 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3055 struct cfs_rq *cfs_rq;
3056 struct sched_entity *se = &p->se;
3057 int task_sleep = flags & DEQUEUE_SLEEP;
3059 for_each_sched_entity(se) {
3060 cfs_rq = cfs_rq_of(se);
3061 dequeue_entity(cfs_rq, se, flags);
3064 * end evaluation on encountering a throttled cfs_rq
3066 * note: in the case of encountering a throttled cfs_rq we will
3067 * post the final h_nr_running decrement below.
3069 if (cfs_rq_throttled(cfs_rq))
3071 cfs_rq->h_nr_running--;
3073 /* Don't dequeue parent if it has other entities besides us */
3074 if (cfs_rq->load.weight) {
3076 * Bias pick_next to pick a task from this cfs_rq, as
3077 * p is sleeping when it is within its sched_slice.
3079 if (task_sleep && parent_entity(se))
3080 set_next_buddy(parent_entity(se));
3082 /* avoid re-evaluating load for this entity */
3083 se = parent_entity(se);
3086 flags |= DEQUEUE_SLEEP;
3089 for_each_sched_entity(se) {
3090 cfs_rq = cfs_rq_of(se);
3091 cfs_rq->h_nr_running--;
3093 if (cfs_rq_throttled(cfs_rq))
3096 update_cfs_shares(cfs_rq);
3097 update_entity_load_avg(se, 1);
3102 update_rq_runnable_avg(rq, 1);
3108 /* Used instead of source_load when we know the type == 0 */
3109 static unsigned long weighted_cpuload(const int cpu)
3111 return cpu_rq(cpu)->load.weight;
3115 * Return a low guess at the load of a migration-source cpu weighted
3116 * according to the scheduling class and "nice" value.
3118 * We want to under-estimate the load of migration sources, to
3119 * balance conservatively.
3121 static unsigned long source_load(int cpu, int type)
3123 struct rq *rq = cpu_rq(cpu);
3124 unsigned long total = weighted_cpuload(cpu);
3126 if (type == 0 || !sched_feat(LB_BIAS))
3129 return min(rq->cpu_load[type-1], total);
3133 * Return a high guess at the load of a migration-target cpu weighted
3134 * according to the scheduling class and "nice" value.
3136 static unsigned long target_load(int cpu, int type)
3138 struct rq *rq = cpu_rq(cpu);
3139 unsigned long total = weighted_cpuload(cpu);
3141 if (type == 0 || !sched_feat(LB_BIAS))
3144 return max(rq->cpu_load[type-1], total);
3147 static unsigned long power_of(int cpu)
3149 return cpu_rq(cpu)->cpu_power;
3152 static unsigned long cpu_avg_load_per_task(int cpu)
3154 struct rq *rq = cpu_rq(cpu);
3155 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3158 return rq->load.weight / nr_running;
3164 static void task_waking_fair(struct task_struct *p)
3166 struct sched_entity *se = &p->se;
3167 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3170 #ifndef CONFIG_64BIT
3171 u64 min_vruntime_copy;
3174 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3176 min_vruntime = cfs_rq->min_vruntime;
3177 } while (min_vruntime != min_vruntime_copy);
3179 min_vruntime = cfs_rq->min_vruntime;
3182 se->vruntime -= min_vruntime;
3185 #ifdef CONFIG_FAIR_GROUP_SCHED
3187 * effective_load() calculates the load change as seen from the root_task_group
3189 * Adding load to a group doesn't make a group heavier, but can cause movement
3190 * of group shares between cpus. Assuming the shares were perfectly aligned one
3191 * can calculate the shift in shares.
3193 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3194 * on this @cpu and results in a total addition (subtraction) of @wg to the
3195 * total group weight.
3197 * Given a runqueue weight distribution (rw_i) we can compute a shares
3198 * distribution (s_i) using:
3200 * s_i = rw_i / \Sum rw_j (1)
3202 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3203 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3204 * shares distribution (s_i):
3206 * rw_i = { 2, 4, 1, 0 }
3207 * s_i = { 2/7, 4/7, 1/7, 0 }
3209 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3210 * task used to run on and the CPU the waker is running on), we need to
3211 * compute the effect of waking a task on either CPU and, in case of a sync
3212 * wakeup, compute the effect of the current task going to sleep.
3214 * So for a change of @wl to the local @cpu with an overall group weight change
3215 * of @wl we can compute the new shares distribution (s'_i) using:
3217 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3219 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3220 * differences in waking a task to CPU 0. The additional task changes the
3221 * weight and shares distributions like:
3223 * rw'_i = { 3, 4, 1, 0 }
3224 * s'_i = { 3/8, 4/8, 1/8, 0 }
3226 * We can then compute the difference in effective weight by using:
3228 * dw_i = S * (s'_i - s_i) (3)
3230 * Where 'S' is the group weight as seen by its parent.
3232 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3233 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3234 * 4/7) times the weight of the group.
3236 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3238 struct sched_entity *se = tg->se[cpu];
3240 if (!tg->parent) /* the trivial, non-cgroup case */
3243 for_each_sched_entity(se) {
3249 * W = @wg + \Sum rw_j
3251 W = wg + calc_tg_weight(tg, se->my_q);
3256 w = se->my_q->load.weight + wl;
3259 * wl = S * s'_i; see (2)
3262 wl = (w * tg->shares) / W;
3267 * Per the above, wl is the new se->load.weight value; since
3268 * those are clipped to [MIN_SHARES, ...) do so now. See
3269 * calc_cfs_shares().
3271 if (wl < MIN_SHARES)
3275 * wl = dw_i = S * (s'_i - s_i); see (3)
3277 wl -= se->load.weight;
3280 * Recursively apply this logic to all parent groups to compute
3281 * the final effective load change on the root group. Since
3282 * only the @tg group gets extra weight, all parent groups can
3283 * only redistribute existing shares. @wl is the shift in shares
3284 * resulting from this level per the above.
3293 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3294 unsigned long wl, unsigned long wg)
3301 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3303 s64 this_load, load;
3304 int idx, this_cpu, prev_cpu;
3305 unsigned long tl_per_task;
3306 struct task_group *tg;
3307 unsigned long weight;
3311 this_cpu = smp_processor_id();
3312 prev_cpu = task_cpu(p);
3313 load = source_load(prev_cpu, idx);
3314 this_load = target_load(this_cpu, idx);
3317 * If sync wakeup then subtract the (maximum possible)
3318 * effect of the currently running task from the load
3319 * of the current CPU:
3322 tg = task_group(current);
3323 weight = current->se.load.weight;
3325 this_load += effective_load(tg, this_cpu, -weight, -weight);
3326 load += effective_load(tg, prev_cpu, 0, -weight);
3330 weight = p->se.load.weight;
3333 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3334 * due to the sync cause above having dropped this_load to 0, we'll
3335 * always have an imbalance, but there's really nothing you can do
3336 * about that, so that's good too.
3338 * Otherwise check if either cpus are near enough in load to allow this
3339 * task to be woken on this_cpu.
3341 if (this_load > 0) {
3342 s64 this_eff_load, prev_eff_load;
3344 this_eff_load = 100;
3345 this_eff_load *= power_of(prev_cpu);
3346 this_eff_load *= this_load +
3347 effective_load(tg, this_cpu, weight, weight);
3349 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3350 prev_eff_load *= power_of(this_cpu);
3351 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3353 balanced = this_eff_load <= prev_eff_load;
3358 * If the currently running task will sleep within
3359 * a reasonable amount of time then attract this newly
3362 if (sync && balanced)
3365 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3366 tl_per_task = cpu_avg_load_per_task(this_cpu);
3369 (this_load <= load &&
3370 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3372 * This domain has SD_WAKE_AFFINE and
3373 * p is cache cold in this domain, and
3374 * there is no bad imbalance.
3376 schedstat_inc(sd, ttwu_move_affine);
3377 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3385 * find_idlest_group finds and returns the least busy CPU group within the
3388 static struct sched_group *
3389 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3390 int this_cpu, int load_idx)
3392 struct sched_group *idlest = NULL, *group = sd->groups;
3393 unsigned long min_load = ULONG_MAX, this_load = 0;
3394 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3397 unsigned long load, avg_load;
3401 /* Skip over this group if it has no CPUs allowed */
3402 if (!cpumask_intersects(sched_group_cpus(group),
3403 tsk_cpus_allowed(p)))
3406 local_group = cpumask_test_cpu(this_cpu,
3407 sched_group_cpus(group));
3409 /* Tally up the load of all CPUs in the group */
3412 for_each_cpu(i, sched_group_cpus(group)) {
3413 /* Bias balancing toward cpus of our domain */
3415 load = source_load(i, load_idx);
3417 load = target_load(i, load_idx);
3422 /* Adjust by relative CPU power of the group */
3423 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3426 this_load = avg_load;
3427 } else if (avg_load < min_load) {
3428 min_load = avg_load;
3431 } while (group = group->next, group != sd->groups);
3433 if (!idlest || 100*this_load < imbalance*min_load)
3439 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3442 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3444 unsigned long load, min_load = ULONG_MAX;
3448 /* Traverse only the allowed CPUs */
3449 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3450 load = weighted_cpuload(i);
3452 if (load < min_load || (load == min_load && i == this_cpu)) {
3462 * Try and locate an idle CPU in the sched_domain.
3464 static int select_idle_sibling(struct task_struct *p, int target)
3466 struct sched_domain *sd;
3467 struct sched_group *sg;
3468 int i = task_cpu(p);
3470 if (idle_cpu(target))
3474 * If the prevous cpu is cache affine and idle, don't be stupid.
3476 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3480 * Otherwise, iterate the domains and find an elegible idle cpu.
3482 sd = rcu_dereference(per_cpu(sd_llc, target));
3483 for_each_lower_domain(sd) {
3486 if (!cpumask_intersects(sched_group_cpus(sg),
3487 tsk_cpus_allowed(p)))
3490 for_each_cpu(i, sched_group_cpus(sg)) {
3491 if (i == target || !idle_cpu(i))
3495 target = cpumask_first_and(sched_group_cpus(sg),
3496 tsk_cpus_allowed(p));
3500 } while (sg != sd->groups);
3506 #ifdef CONFIG_SCHED_HMP
3508 * Heterogenous multiprocessor (HMP) optimizations
3510 * The cpu types are distinguished using a list of hmp_domains
3511 * which each represent one cpu type using a cpumask.
3512 * The list is assumed ordered by compute capacity with the
3513 * fastest domain first.
3515 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3516 static const int hmp_max_tasks = 5;
3518 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3520 #ifdef CONFIG_CPU_IDLE
3524 * In this version we have stopped using forced up migrations when we
3525 * detect that a task running on a little CPU should be moved to a bigger
3526 * CPU. In most cases, the bigger CPU is in a deep sleep state and a forced
3527 * migration means we stop the task immediately but need to wait for the
3528 * target CPU to wake up before we can restart the task which is being
3529 * moved. Instead, we now wake a big CPU with an IPI and ask it to pull
3530 * a task when ready. This allows the task to continue executing on its
3531 * current CPU, reducing the amount of time that the task is stalled for.
3535 * The keepalive timer is used as a way to keep a CPU engaged in an
3536 * idle pull operation out of idle while waiting for the source
3537 * CPU to stop and move the task. Ideally this would not be necessary
3538 * and we could impose a temporary zero-latency requirement on the
3539 * current CPU, but in the current QoS framework this will result in
3540 * all CPUs in the system being unable to enter idle states which is
3541 * not desirable. The timer does not perform any work when it expires.
3543 struct hmp_keepalive {
3545 ktime_t delay; /* if zero, no need for timer */
3546 struct hrtimer timer;
3548 DEFINE_PER_CPU(struct hmp_keepalive, hmp_cpu_keepalive);
3550 /* setup per-cpu keepalive timers */
3551 static enum hrtimer_restart hmp_cpu_keepalive_notify(struct hrtimer *hrtimer)
3553 return HRTIMER_NORESTART;
3557 * Work out if any of the idle states have an exit latency too high for us.
3558 * ns_delay is passed in containing the max we are willing to tolerate.
3559 * If there are none, set ns_delay to zero.
3560 * If there are any, set ns_delay to
3561 * ('target_residency of state with shortest too-big latency' - 1) * 1000.
3563 static void hmp_keepalive_delay(unsigned int *ns_delay)
3565 struct cpuidle_driver *drv;
3566 drv = cpuidle_driver_ref();
3568 unsigned int us_delay = UINT_MAX;
3569 unsigned int us_max_delay = *ns_delay / 1000;
3571 /* if cpuidle states are guaranteed to be sorted we
3572 * could stop at the first match.
3574 for (idx = 0; idx < drv->state_count; idx++) {
3575 if (drv->states[idx].exit_latency > us_max_delay &&
3576 drv->states[idx].target_residency < us_delay) {
3577 us_delay = drv->states[idx].target_residency;
3580 if (us_delay == UINT_MAX)
3581 *ns_delay = 0; /* no timer required */
3583 *ns_delay = 1000 * (us_delay - 1);
3585 cpuidle_driver_unref();
3588 static void hmp_cpu_keepalive_trigger(void)
3590 int cpu = smp_processor_id();
3591 struct hmp_keepalive *keepalive = &per_cpu(hmp_cpu_keepalive, cpu);
3592 if (!keepalive->init) {
3593 unsigned int ns_delay = 100000; /* tolerate 100usec delay */
3595 hrtimer_init(&keepalive->timer,
3596 CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
3597 keepalive->timer.function = hmp_cpu_keepalive_notify;
3599 hmp_keepalive_delay(&ns_delay);
3600 keepalive->delay = ns_to_ktime(ns_delay);
3601 keepalive->init = true;
3603 if (ktime_to_ns(keepalive->delay))
3604 hrtimer_start(&keepalive->timer,
3605 keepalive->delay, HRTIMER_MODE_REL_PINNED);
3608 static void hmp_cpu_keepalive_cancel(int cpu)
3610 struct hmp_keepalive *keepalive = &per_cpu(hmp_cpu_keepalive, cpu);
3611 if (keepalive->init)
3612 hrtimer_cancel(&keepalive->timer);
3614 #else /* !CONFIG_CPU_IDLE */
3615 static void hmp_cpu_keepalive_trigger(void)
3619 static void hmp_cpu_keepalive_cancel(int cpu)
3624 /* Setup hmp_domains */
3625 static int __init hmp_cpu_mask_setup(void)
3628 struct hmp_domain *domain;
3629 struct list_head *pos;
3632 pr_debug("Initializing HMP scheduler:\n");
3634 /* Initialize hmp_domains using platform code */
3635 arch_get_hmp_domains(&hmp_domains);
3636 if (list_empty(&hmp_domains)) {
3637 pr_debug("HMP domain list is empty!\n");
3641 /* Print hmp_domains */
3643 list_for_each(pos, &hmp_domains) {
3644 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3645 cpulist_scnprintf(buf, 64, &domain->possible_cpus);
3646 pr_debug(" HMP domain %d: %s\n", dc, buf);
3648 for_each_cpu_mask(cpu, domain->possible_cpus) {
3649 per_cpu(hmp_cpu_domain, cpu) = domain;
3657 static struct hmp_domain *hmp_get_hmp_domain_for_cpu(int cpu)
3659 struct hmp_domain *domain;
3660 struct list_head *pos;
3662 list_for_each(pos, &hmp_domains) {
3663 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3664 if(cpumask_test_cpu(cpu, &domain->possible_cpus))
3670 static void hmp_online_cpu(int cpu)
3672 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3675 cpumask_set_cpu(cpu, &domain->cpus);
3678 static void hmp_offline_cpu(int cpu)
3680 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3683 cpumask_clear_cpu(cpu, &domain->cpus);
3685 hmp_cpu_keepalive_cancel(cpu);
3688 * Needed to determine heaviest tasks etc.
3690 static inline unsigned int hmp_cpu_is_fastest(int cpu);
3691 static inline unsigned int hmp_cpu_is_slowest(int cpu);
3692 static inline struct hmp_domain *hmp_slower_domain(int cpu);
3693 static inline struct hmp_domain *hmp_faster_domain(int cpu);
3695 /* must hold runqueue lock for queue se is currently on */
3696 static struct sched_entity *hmp_get_heaviest_task(
3697 struct sched_entity *se, int target_cpu)
3699 int num_tasks = hmp_max_tasks;
3700 struct sched_entity *max_se = se;
3701 unsigned long int max_ratio = se->avg.load_avg_ratio;
3702 const struct cpumask *hmp_target_mask = NULL;
3703 struct hmp_domain *hmp;
3705 if (hmp_cpu_is_fastest(cpu_of(se->cfs_rq->rq)))
3708 hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq));
3709 hmp_target_mask = &hmp->cpus;
3710 if (target_cpu >= 0) {
3711 /* idle_balance gets run on a CPU while
3712 * it is in the middle of being hotplugged
3713 * out. Bail early in that case.
3715 if(!cpumask_test_cpu(target_cpu, hmp_target_mask))
3717 hmp_target_mask = cpumask_of(target_cpu);
3719 /* The currently running task is not on the runqueue */
3720 se = __pick_first_entity(cfs_rq_of(se));
3722 while (num_tasks && se) {
3723 if (entity_is_task(se) &&
3724 se->avg.load_avg_ratio > max_ratio &&
3725 cpumask_intersects(hmp_target_mask,
3726 tsk_cpus_allowed(task_of(se)))) {
3728 max_ratio = se->avg.load_avg_ratio;
3730 se = __pick_next_entity(se);
3736 static struct sched_entity *hmp_get_lightest_task(
3737 struct sched_entity *se, int migrate_down)
3739 int num_tasks = hmp_max_tasks;
3740 struct sched_entity *min_se = se;
3741 unsigned long int min_ratio = se->avg.load_avg_ratio;
3742 const struct cpumask *hmp_target_mask = NULL;
3745 struct hmp_domain *hmp;
3746 if (hmp_cpu_is_slowest(cpu_of(se->cfs_rq->rq)))
3748 hmp = hmp_slower_domain(cpu_of(se->cfs_rq->rq));
3749 hmp_target_mask = &hmp->cpus;
3751 /* The currently running task is not on the runqueue */
3752 se = __pick_first_entity(cfs_rq_of(se));
3754 while (num_tasks && se) {
3755 if (entity_is_task(se) &&
3756 (se->avg.load_avg_ratio < min_ratio &&
3758 cpumask_intersects(hmp_target_mask,
3759 tsk_cpus_allowed(task_of(se))))) {
3761 min_ratio = se->avg.load_avg_ratio;
3763 se = __pick_next_entity(se);
3770 * Migration thresholds should be in the range [0..1023]
3771 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3772 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3774 * hmp_up_prio: Only up migrate task with high priority (<hmp_up_prio)
3775 * hmp_next_up_threshold: Delay before next up migration (1024 ~= 1 ms)
3776 * hmp_next_down_threshold: Delay before next down migration (1024 ~= 1 ms)
3778 * Small Task Packing:
3779 * We can choose to fill the littlest CPUs in an HMP system rather than
3780 * the typical spreading mechanic. This behavior is controllable using
3782 * hmp_packing_enabled: runtime control over pack/spread
3783 * hmp_full_threshold: Consider a CPU with this much unweighted load full
3785 unsigned int hmp_up_threshold = 700;
3786 unsigned int hmp_down_threshold = 512;
3787 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
3788 unsigned int hmp_up_prio = NICE_TO_PRIO(CONFIG_SCHED_HMP_PRIO_FILTER_VAL);
3790 unsigned int hmp_next_up_threshold = 4096;
3791 unsigned int hmp_next_down_threshold = 4096;
3793 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3795 * Set the default packing threshold to try to keep little
3796 * CPUs at no more than 80% of their maximum frequency if only
3797 * packing a small number of small tasks. Bigger tasks will
3798 * raise frequency as normal.
3799 * In order to pack a task onto a CPU, the sum of the
3800 * unweighted runnable_avg load of existing tasks plus the
3801 * load of the new task must be less than hmp_full_threshold.
3803 * This works in conjunction with frequency-invariant load
3804 * and DVFS governors. Since most DVFS governors aim for 80%
3805 * utilisation, we arrive at (0.8*0.8*(max_load=1024))=655
3806 * and use a value slightly lower to give a little headroom
3808 * Note that the most efficient frequency is different for
3809 * each system so /sys/kernel/hmp/packing_limit should be
3810 * configured at runtime for any given platform to achieve
3811 * optimal energy usage. Some systems may not benefit from
3812 * packing, so this feature can also be disabled at runtime
3813 * with /sys/kernel/hmp/packing_enable
3815 unsigned int hmp_packing_enabled = 1;
3816 unsigned int hmp_full_threshold = 650;
3819 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se);
3820 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3821 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
3822 int *min_cpu, struct cpumask *affinity);
3824 static inline struct hmp_domain *hmp_smallest_domain(void)
3826 return list_entry(hmp_domains.prev, struct hmp_domain, hmp_domains);
3829 /* Check if cpu is in fastest hmp_domain */
3830 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3832 struct list_head *pos;
3834 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3835 return pos == hmp_domains.next;
3838 /* Check if cpu is in slowest hmp_domain */
3839 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3841 struct list_head *pos;
3843 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3844 return list_is_last(pos, &hmp_domains);
3847 /* Next (slower) hmp_domain relative to cpu */
3848 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3850 struct list_head *pos;
3852 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3853 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3856 /* Previous (faster) hmp_domain relative to cpu */
3857 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3859 struct list_head *pos;
3861 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3862 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3866 * Selects a cpu in previous (faster) hmp_domain
3868 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3871 int lowest_cpu=NR_CPUS;
3872 __always_unused int lowest_ratio;
3873 struct hmp_domain *hmp;
3875 if (hmp_cpu_is_fastest(cpu))
3876 hmp = hmp_cpu_domain(cpu);
3878 hmp = hmp_faster_domain(cpu);
3880 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3881 tsk_cpus_allowed(tsk));
3887 * Selects a cpu in next (slower) hmp_domain
3888 * Note that cpumask_any_and() returns the first cpu in the cpumask
3890 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3893 int lowest_cpu=NR_CPUS;
3894 struct hmp_domain *hmp;
3895 __always_unused int lowest_ratio;
3897 if (hmp_cpu_is_slowest(cpu))
3898 hmp = hmp_cpu_domain(cpu);
3900 hmp = hmp_slower_domain(cpu);
3902 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3903 tsk_cpus_allowed(tsk));
3907 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3909 * Select the 'best' candidate little CPU to wake up on.
3910 * Implements a packing strategy which examines CPU in
3911 * logical CPU order, and selects the first which will
3912 * be loaded less than hmp_full_threshold according to
3913 * the sum of the tracked load of the runqueue and the task.
3915 static inline unsigned int hmp_best_little_cpu(struct task_struct *tsk,
3918 unsigned long estimated_load;
3919 struct hmp_domain *hmp;
3920 struct sched_avg *avg;
3921 struct cpumask allowed_hmp_cpus;
3923 if(!hmp_packing_enabled ||
3924 tsk->se.avg.load_avg_ratio > ((NICE_0_LOAD * 90)/100))
3925 return hmp_select_slower_cpu(tsk, cpu);
3927 if (hmp_cpu_is_slowest(cpu))
3928 hmp = hmp_cpu_domain(cpu);
3930 hmp = hmp_slower_domain(cpu);
3932 /* respect affinity */
3933 cpumask_and(&allowed_hmp_cpus, &hmp->cpus,
3934 tsk_cpus_allowed(tsk));
3936 for_each_cpu_mask(tmp_cpu, allowed_hmp_cpus) {
3937 avg = &cpu_rq(tmp_cpu)->avg;
3938 /* estimate new rq load if we add this task */
3939 estimated_load = avg->load_avg_ratio +
3940 tsk->se.avg.load_avg_ratio;
3941 if (estimated_load <= hmp_full_threshold) {
3946 /* if no match was found, the task uses the initial value */
3950 static inline void hmp_next_up_delay(struct sched_entity *se, int cpu)
3952 /* hack - always use clock from first online CPU */
3953 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3954 se->avg.hmp_last_up_migration = now;
3955 se->avg.hmp_last_down_migration = 0;
3956 cpu_rq(cpu)->avg.hmp_last_up_migration = now;
3957 cpu_rq(cpu)->avg.hmp_last_down_migration = 0;
3960 static inline void hmp_next_down_delay(struct sched_entity *se, int cpu)
3962 /* hack - always use clock from first online CPU */
3963 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3964 se->avg.hmp_last_down_migration = now;
3965 se->avg.hmp_last_up_migration = 0;
3966 cpu_rq(cpu)->avg.hmp_last_down_migration = now;
3967 cpu_rq(cpu)->avg.hmp_last_up_migration = 0;
3971 * Heterogenous multiprocessor (HMP) optimizations
3973 * These functions allow to change the growing speed of the load_avg_ratio
3974 * by default it goes from 0 to 0.5 in LOAD_AVG_PERIOD = 32ms
3975 * This can now be changed with /sys/kernel/hmp/load_avg_period_ms.
3977 * These functions also allow to change the up and down threshold of HMP
3978 * using /sys/kernel/hmp/{up,down}_threshold.
3979 * Both must be between 0 and 1023. The threshold that is compared
3980 * to the load_avg_ratio is up_threshold/1024 and down_threshold/1024.
3982 * For instance, if load_avg_period = 64 and up_threshold = 512, an idle
3983 * task with a load of 0 will reach the threshold after 64ms of busy loop.
3985 * Changing load_avg_periods_ms has the same effect than changing the
3986 * default scaling factor Y=1002/1024 in the load_avg_ratio computation to
3987 * (1002/1024.0)^(LOAD_AVG_PERIOD/load_avg_period_ms), but the last one
3988 * could trigger overflows.
3989 * For instance, with Y = 1023/1024 in __update_task_entity_contrib()
3990 * "contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);"
3991 * could be overflowed for a weight > 2^12 even is the load_avg_contrib
3992 * should still be a 32bits result. This would not happen by multiplicating
3993 * delta time by 1/22 and setting load_avg_period_ms = 706.
3997 * By scaling the delta time it end-up increasing or decrease the
3998 * growing speed of the per entity load_avg_ratio
3999 * The scale factor hmp_data.multiplier is a fixed point
4000 * number: (32-HMP_VARIABLE_SCALE_SHIFT).HMP_VARIABLE_SCALE_SHIFT
4002 static inline u64 hmp_variable_scale_convert(u64 delta)
4004 #ifdef CONFIG_HMP_VARIABLE_SCALE
4005 u64 high = delta >> 32ULL;
4006 u64 low = delta & 0xffffffffULL;
4007 low *= hmp_data.multiplier;
4008 high *= hmp_data.multiplier;
4009 return (low >> HMP_VARIABLE_SCALE_SHIFT)
4010 + (high << (32ULL - HMP_VARIABLE_SCALE_SHIFT));
4016 static ssize_t hmp_show(struct kobject *kobj,
4017 struct attribute *attr, char *buf)
4019 struct hmp_global_attr *hmp_attr =
4020 container_of(attr, struct hmp_global_attr, attr);
4023 if (hmp_attr->to_sysfs_text != NULL)
4024 return hmp_attr->to_sysfs_text(buf, PAGE_SIZE);
4026 temp = *(hmp_attr->value);
4027 if (hmp_attr->to_sysfs != NULL)
4028 temp = hmp_attr->to_sysfs(temp);
4030 return (ssize_t)sprintf(buf, "%d\n", temp);
4033 static ssize_t hmp_store(struct kobject *a, struct attribute *attr,
4034 const char *buf, size_t count)
4037 ssize_t ret = count;
4038 struct hmp_global_attr *hmp_attr =
4039 container_of(attr, struct hmp_global_attr, attr);
4040 char *str = vmalloc(count + 1);
4043 memcpy(str, buf, count);
4045 if (sscanf(str, "%d", &temp) < 1)
4048 if (hmp_attr->from_sysfs != NULL)
4049 temp = hmp_attr->from_sysfs(temp);
4053 *(hmp_attr->value) = temp;
4059 static ssize_t hmp_print_domains(char *outbuf, int outbufsize)
4062 const char nospace[] = "%s", space[] = " %s";
4063 const char *fmt = nospace;
4064 struct hmp_domain *domain;
4065 struct list_head *pos;
4067 list_for_each(pos, &hmp_domains) {
4068 domain = list_entry(pos, struct hmp_domain, hmp_domains);
4069 if (cpumask_scnprintf(buf, 64, &domain->possible_cpus)) {
4070 outpos += sprintf(outbuf+outpos, fmt, buf);
4074 strcat(outbuf, "\n");
4078 #ifdef CONFIG_HMP_VARIABLE_SCALE
4079 static int hmp_period_tofrom_sysfs(int value)
4081 return (LOAD_AVG_PERIOD << HMP_VARIABLE_SCALE_SHIFT) / value;
4084 /* max value for threshold is 1024 */
4085 static int hmp_theshold_from_sysfs(int value)
4091 #if defined(CONFIG_SCHED_HMP_LITTLE_PACKING) || \
4092 defined(CONFIG_HMP_FREQUENCY_INVARIANT_SCALE)
4093 /* toggle control is only 0,1 off/on */
4094 static int hmp_toggle_from_sysfs(int value)
4096 if (value < 0 || value > 1)
4101 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4102 /* packing value must be non-negative */
4103 static int hmp_packing_from_sysfs(int value)
4110 static void hmp_attr_add(
4113 int (*to_sysfs)(int),
4114 int (*from_sysfs)(int),
4115 ssize_t (*to_sysfs_text)(char *, int),
4119 while (hmp_data.attributes[i] != NULL) {
4121 if (i >= HMP_DATA_SYSFS_MAX)
4125 hmp_data.attr[i].attr.mode = mode;
4127 hmp_data.attr[i].attr.mode = 0644;
4128 hmp_data.attr[i].show = hmp_show;
4129 hmp_data.attr[i].store = hmp_store;
4130 hmp_data.attr[i].attr.name = name;
4131 hmp_data.attr[i].value = value;
4132 hmp_data.attr[i].to_sysfs = to_sysfs;
4133 hmp_data.attr[i].from_sysfs = from_sysfs;
4134 hmp_data.attr[i].to_sysfs_text = to_sysfs_text;
4135 hmp_data.attributes[i] = &hmp_data.attr[i].attr;
4136 hmp_data.attributes[i + 1] = NULL;
4139 static int hmp_attr_init(void)
4142 memset(&hmp_data, sizeof(hmp_data), 0);
4143 hmp_attr_add("hmp_domains",
4149 hmp_attr_add("up_threshold",
4152 hmp_theshold_from_sysfs,
4155 hmp_attr_add("down_threshold",
4156 &hmp_down_threshold,
4158 hmp_theshold_from_sysfs,
4161 #ifdef CONFIG_HMP_VARIABLE_SCALE
4162 /* by default load_avg_period_ms == LOAD_AVG_PERIOD
4165 hmp_data.multiplier = hmp_period_tofrom_sysfs(LOAD_AVG_PERIOD);
4166 hmp_attr_add("load_avg_period_ms",
4167 &hmp_data.multiplier,
4168 hmp_period_tofrom_sysfs,
4169 hmp_period_tofrom_sysfs,
4173 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
4174 /* default frequency-invariant scaling ON */
4175 hmp_data.freqinvar_load_scale_enabled = 1;
4176 hmp_attr_add("frequency_invariant_load_scale",
4177 &hmp_data.freqinvar_load_scale_enabled,
4179 hmp_toggle_from_sysfs,
4183 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4184 hmp_attr_add("packing_enable",
4185 &hmp_packing_enabled,
4187 hmp_toggle_from_sysfs,
4190 hmp_attr_add("packing_limit",
4191 &hmp_full_threshold,
4193 hmp_packing_from_sysfs,
4197 hmp_data.attr_group.name = "hmp";
4198 hmp_data.attr_group.attrs = hmp_data.attributes;
4199 ret = sysfs_create_group(kernel_kobj,
4200 &hmp_data.attr_group);
4203 late_initcall(hmp_attr_init);
4205 * return the load of the lowest-loaded CPU in a given HMP domain
4206 * min_cpu optionally points to an int to receive the CPU.
4207 * affinity optionally points to a cpumask containing the
4208 * CPUs to be considered. note:
4209 * + min_cpu = NR_CPUS only if no CPUs are in the set of
4210 * affinity && hmp_domain cpus
4211 * + min_cpu will always otherwise equal one of the CPUs in
4213 * + when more than one CPU has the same load, the one which
4214 * is least-recently-disturbed by an HMP migration will be
4216 * + if all CPUs are equally loaded or idle and the times are
4217 * all the same, the first in the set will be used
4218 * + if affinity is not set, cpu_online_mask is used
4220 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
4221 int *min_cpu, struct cpumask *affinity)
4224 int min_cpu_runnable_temp = NR_CPUS;
4225 u64 min_target_last_migration = ULLONG_MAX;
4226 u64 curr_last_migration;
4227 unsigned long min_runnable_load = INT_MAX;
4228 unsigned long contrib;
4229 struct sched_avg *avg;
4230 struct cpumask temp_cpumask;
4232 * only look at CPUs allowed if specified,
4233 * otherwise look at all online CPUs in the
4236 cpumask_and(&temp_cpumask, &hmpd->cpus, affinity ? affinity : cpu_online_mask);
4238 for_each_cpu_mask(cpu, temp_cpumask) {
4239 avg = &cpu_rq(cpu)->avg;
4240 /* used for both up and down migration */
4241 curr_last_migration = avg->hmp_last_up_migration ?
4242 avg->hmp_last_up_migration : avg->hmp_last_down_migration;
4244 contrib = avg->load_avg_ratio;
4246 * Consider a runqueue completely busy if there is any load
4247 * on it. Definitely not the best for overall fairness, but
4248 * does well in typical Android use cases.
4253 if ((contrib < min_runnable_load) ||
4254 (contrib == min_runnable_load &&
4255 curr_last_migration < min_target_last_migration)) {
4257 * if the load is the same target the CPU with
4258 * the longest time since a migration.
4259 * This is to spread migration load between
4260 * members of a domain more evenly when the
4261 * domain is fully loaded
4263 min_runnable_load = contrib;
4264 min_cpu_runnable_temp = cpu;
4265 min_target_last_migration = curr_last_migration;
4270 *min_cpu = min_cpu_runnable_temp;
4272 return min_runnable_load;
4276 * Calculate the task starvation
4277 * This is the ratio of actually running time vs. runnable time.
4278 * If the two are equal the task is getting the cpu time it needs or
4279 * it is alone on the cpu and the cpu is fully utilized.
4281 static inline unsigned int hmp_task_starvation(struct sched_entity *se)
4285 starvation = se->avg.usage_avg_sum * scale_load_down(NICE_0_LOAD);
4286 starvation /= (se->avg.runnable_avg_sum + 1);
4288 return scale_load(starvation);
4291 static inline unsigned int hmp_offload_down(int cpu, struct sched_entity *se)
4294 int dest_cpu = NR_CPUS;
4296 if (hmp_cpu_is_slowest(cpu))
4299 /* Is there an idle CPU in the current domain */
4300 min_usage = hmp_domain_min_load(hmp_cpu_domain(cpu), NULL, NULL);
4301 if (min_usage == 0) {
4302 trace_sched_hmp_offload_abort(cpu, min_usage, "load");
4306 /* Is the task alone on the cpu? */
4307 if (cpu_rq(cpu)->cfs.h_nr_running < 2) {
4308 trace_sched_hmp_offload_abort(cpu,
4309 cpu_rq(cpu)->cfs.h_nr_running, "nr_running");
4313 /* Is the task actually starving? */
4314 /* >=25% ratio running/runnable = starving */
4315 if (hmp_task_starvation(se) > 768) {
4316 trace_sched_hmp_offload_abort(cpu, hmp_task_starvation(se),
4321 /* Does the slower domain have any idle CPUs? */
4322 min_usage = hmp_domain_min_load(hmp_slower_domain(cpu), &dest_cpu,
4323 tsk_cpus_allowed(task_of(se)));
4325 if (min_usage == 0) {
4326 trace_sched_hmp_offload_succeed(cpu, dest_cpu);
4329 trace_sched_hmp_offload_abort(cpu,min_usage,"slowdomain");
4332 #endif /* CONFIG_SCHED_HMP */
4335 * sched_balance_self: balance the current task (running on cpu) in domains
4336 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4339 * Balance, ie. select the least loaded group.
4341 * Returns the target CPU number, or the same CPU if no balancing is needed.
4343 * preempt must be disabled.
4346 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
4348 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4349 int cpu = smp_processor_id();
4350 int prev_cpu = task_cpu(p);
4352 int want_affine = 0;
4353 int sync = wake_flags & WF_SYNC;
4355 if (p->nr_cpus_allowed == 1)
4358 #ifdef CONFIG_SCHED_HMP
4359 /* always put non-kernel forking tasks on a big domain */
4360 if (p->mm && (sd_flag & SD_BALANCE_FORK)) {
4361 new_cpu = hmp_select_faster_cpu(p, prev_cpu);
4362 if (new_cpu != NR_CPUS) {
4363 hmp_next_up_delay(&p->se, new_cpu);
4366 /* failed to perform HMP fork balance, use normal balance */
4371 if (sd_flag & SD_BALANCE_WAKE) {
4372 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4378 for_each_domain(cpu, tmp) {
4379 if (!(tmp->flags & SD_LOAD_BALANCE))
4383 * If both cpu and prev_cpu are part of this domain,
4384 * cpu is a valid SD_WAKE_AFFINE target.
4386 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4387 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4392 if (tmp->flags & sd_flag)
4397 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4400 new_cpu = select_idle_sibling(p, prev_cpu);
4405 int load_idx = sd->forkexec_idx;
4406 struct sched_group *group;
4409 if (!(sd->flags & sd_flag)) {
4414 if (sd_flag & SD_BALANCE_WAKE)
4415 load_idx = sd->wake_idx;
4417 group = find_idlest_group(sd, p, cpu, load_idx);
4423 new_cpu = find_idlest_cpu(group, p, cpu);
4424 if (new_cpu == -1 || new_cpu == cpu) {
4425 /* Now try balancing at a lower domain level of cpu */
4430 /* Now try balancing at a lower domain level of new_cpu */
4432 weight = sd->span_weight;
4434 for_each_domain(cpu, tmp) {
4435 if (weight <= tmp->span_weight)
4437 if (tmp->flags & sd_flag)
4440 /* while loop will break here if sd == NULL */
4445 #ifdef CONFIG_SCHED_HMP
4446 prev_cpu = task_cpu(p);
4448 if (hmp_up_migration(prev_cpu, &new_cpu, &p->se)) {
4449 hmp_next_up_delay(&p->se, new_cpu);
4450 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4453 if (hmp_down_migration(prev_cpu, &p->se)) {
4454 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4455 new_cpu = hmp_best_little_cpu(p, prev_cpu);
4457 new_cpu = hmp_select_slower_cpu(p, prev_cpu);
4460 * we might have no suitable CPU
4461 * in which case new_cpu == NR_CPUS
4463 if (new_cpu < NR_CPUS && new_cpu != prev_cpu) {
4464 hmp_next_down_delay(&p->se, new_cpu);
4465 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4469 /* Make sure that the task stays in its previous hmp domain */
4470 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
4478 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
4479 * removed when useful for applications beyond shares distribution (e.g.
4482 #ifdef CONFIG_FAIR_GROUP_SCHED
4484 #ifdef CONFIG_NO_HZ_COMMON
4485 static int nohz_test_cpu(int cpu);
4487 static inline int nohz_test_cpu(int cpu)
4494 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4495 * cfs_rq_of(p) references at time of call are still valid and identify the
4496 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4497 * other assumptions, including the state of rq->lock, should be made.
4500 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4502 struct sched_entity *se = &p->se;
4503 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4506 * Load tracking: accumulate removed load so that it can be processed
4507 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4508 * to blocked load iff they have a positive decay-count. It can never
4509 * be negative here since on-rq tasks have decay-count == 0.
4511 if (se->avg.decay_count) {
4513 * If we migrate a sleeping task away from a CPU
4514 * which has the tick stopped, then both the clock_task
4515 * and decay_counter will be out of date for that CPU
4516 * and we will not decay load correctly.
4518 if (!se->on_rq && nohz_test_cpu(task_cpu(p))) {
4519 struct rq *rq = cpu_rq(task_cpu(p));
4520 unsigned long flags;
4522 * Current CPU cannot be holding rq->lock in this
4523 * circumstance, but another might be. We must hold
4524 * rq->lock before we go poking around in its clocks
4526 raw_spin_lock_irqsave(&rq->lock, flags);
4527 update_rq_clock(rq);
4528 update_cfs_rq_blocked_load(cfs_rq, 0);
4529 raw_spin_unlock_irqrestore(&rq->lock, flags);
4531 se->avg.decay_count = -__synchronize_entity_decay(se);
4532 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
4536 #endif /* CONFIG_SMP */
4538 static unsigned long
4539 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4541 unsigned long gran = sysctl_sched_wakeup_granularity;
4544 * Since its curr running now, convert the gran from real-time
4545 * to virtual-time in his units.
4547 * By using 'se' instead of 'curr' we penalize light tasks, so
4548 * they get preempted easier. That is, if 'se' < 'curr' then
4549 * the resulting gran will be larger, therefore penalizing the
4550 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4551 * be smaller, again penalizing the lighter task.
4553 * This is especially important for buddies when the leftmost
4554 * task is higher priority than the buddy.
4556 return calc_delta_fair(gran, se);
4560 * Should 'se' preempt 'curr'.
4574 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4576 s64 gran, vdiff = curr->vruntime - se->vruntime;
4581 gran = wakeup_gran(curr, se);
4588 static void set_last_buddy(struct sched_entity *se)
4590 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4593 for_each_sched_entity(se)
4594 cfs_rq_of(se)->last = se;
4597 static void set_next_buddy(struct sched_entity *se)
4599 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4602 for_each_sched_entity(se)
4603 cfs_rq_of(se)->next = se;
4606 static void set_skip_buddy(struct sched_entity *se)
4608 for_each_sched_entity(se)
4609 cfs_rq_of(se)->skip = se;
4613 * Preempt the current task with a newly woken task if needed:
4615 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4617 struct task_struct *curr = rq->curr;
4618 struct sched_entity *se = &curr->se, *pse = &p->se;
4619 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4620 int scale = cfs_rq->nr_running >= sched_nr_latency;
4621 int next_buddy_marked = 0;
4623 if (unlikely(se == pse))
4627 * This is possible from callers such as move_task(), in which we
4628 * unconditionally check_prempt_curr() after an enqueue (which may have
4629 * lead to a throttle). This both saves work and prevents false
4630 * next-buddy nomination below.
4632 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4635 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4636 set_next_buddy(pse);
4637 next_buddy_marked = 1;
4641 * We can come here with TIF_NEED_RESCHED already set from new task
4644 * Note: this also catches the edge-case of curr being in a throttled
4645 * group (e.g. via set_curr_task), since update_curr() (in the
4646 * enqueue of curr) will have resulted in resched being set. This
4647 * prevents us from potentially nominating it as a false LAST_BUDDY
4650 if (test_tsk_need_resched(curr))
4653 /* Idle tasks are by definition preempted by non-idle tasks. */
4654 if (unlikely(curr->policy == SCHED_IDLE) &&
4655 likely(p->policy != SCHED_IDLE))
4659 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4660 * is driven by the tick):
4662 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4665 find_matching_se(&se, &pse);
4666 update_curr(cfs_rq_of(se));
4668 if (wakeup_preempt_entity(se, pse) == 1) {
4670 * Bias pick_next to pick the sched entity that is
4671 * triggering this preemption.
4673 if (!next_buddy_marked)
4674 set_next_buddy(pse);
4683 * Only set the backward buddy when the current task is still
4684 * on the rq. This can happen when a wakeup gets interleaved
4685 * with schedule on the ->pre_schedule() or idle_balance()
4686 * point, either of which can * drop the rq lock.
4688 * Also, during early boot the idle thread is in the fair class,
4689 * for obvious reasons its a bad idea to schedule back to it.
4691 if (unlikely(!se->on_rq || curr == rq->idle))
4694 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4698 static struct task_struct *pick_next_task_fair(struct rq *rq)
4700 struct task_struct *p;
4701 struct cfs_rq *cfs_rq = &rq->cfs;
4702 struct sched_entity *se;
4704 if (!cfs_rq->nr_running)
4708 se = pick_next_entity(cfs_rq);
4709 set_next_entity(cfs_rq, se);
4710 cfs_rq = group_cfs_rq(se);
4714 if (hrtick_enabled(rq))
4715 hrtick_start_fair(rq, p);
4721 * Account for a descheduled task:
4723 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4725 struct sched_entity *se = &prev->se;
4726 struct cfs_rq *cfs_rq;
4728 for_each_sched_entity(se) {
4729 cfs_rq = cfs_rq_of(se);
4730 put_prev_entity(cfs_rq, se);
4735 * sched_yield() is very simple
4737 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4739 static void yield_task_fair(struct rq *rq)
4741 struct task_struct *curr = rq->curr;
4742 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4743 struct sched_entity *se = &curr->se;
4746 * Are we the only task in the tree?
4748 if (unlikely(rq->nr_running == 1))
4751 clear_buddies(cfs_rq, se);
4753 if (curr->policy != SCHED_BATCH) {
4754 update_rq_clock(rq);
4756 * Update run-time statistics of the 'current'.
4758 update_curr(cfs_rq);
4760 * Tell update_rq_clock() that we've just updated,
4761 * so we don't do microscopic update in schedule()
4762 * and double the fastpath cost.
4764 rq->skip_clock_update = 1;
4770 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4772 struct sched_entity *se = &p->se;
4774 /* throttled hierarchies are not runnable */
4775 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4778 /* Tell the scheduler that we'd really like pse to run next. */
4781 yield_task_fair(rq);
4787 /**************************************************
4788 * Fair scheduling class load-balancing methods.
4792 * The purpose of load-balancing is to achieve the same basic fairness the
4793 * per-cpu scheduler provides, namely provide a proportional amount of compute
4794 * time to each task. This is expressed in the following equation:
4796 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4798 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4799 * W_i,0 is defined as:
4801 * W_i,0 = \Sum_j w_i,j (2)
4803 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4804 * is derived from the nice value as per prio_to_weight[].
4806 * The weight average is an exponential decay average of the instantaneous
4809 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4811 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4812 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4813 * can also include other factors [XXX].
4815 * To achieve this balance we define a measure of imbalance which follows
4816 * directly from (1):
4818 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4820 * We them move tasks around to minimize the imbalance. In the continuous
4821 * function space it is obvious this converges, in the discrete case we get
4822 * a few fun cases generally called infeasible weight scenarios.
4825 * - infeasible weights;
4826 * - local vs global optima in the discrete case. ]
4831 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4832 * for all i,j solution, we create a tree of cpus that follows the hardware
4833 * topology where each level pairs two lower groups (or better). This results
4834 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4835 * tree to only the first of the previous level and we decrease the frequency
4836 * of load-balance at each level inv. proportional to the number of cpus in
4842 * \Sum { --- * --- * 2^i } = O(n) (5)
4844 * `- size of each group
4845 * | | `- number of cpus doing load-balance
4847 * `- sum over all levels
4849 * Coupled with a limit on how many tasks we can migrate every balance pass,
4850 * this makes (5) the runtime complexity of the balancer.
4852 * An important property here is that each CPU is still (indirectly) connected
4853 * to every other cpu in at most O(log n) steps:
4855 * The adjacency matrix of the resulting graph is given by:
4858 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4861 * And you'll find that:
4863 * A^(log_2 n)_i,j != 0 for all i,j (7)
4865 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4866 * The task movement gives a factor of O(m), giving a convergence complexity
4869 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4874 * In order to avoid CPUs going idle while there's still work to do, new idle
4875 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4876 * tree itself instead of relying on other CPUs to bring it work.
4878 * This adds some complexity to both (5) and (8) but it reduces the total idle
4886 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4889 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4894 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4896 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4898 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4901 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4902 * rewrite all of this once again.]
4905 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4907 #define LBF_ALL_PINNED 0x01
4908 #define LBF_NEED_BREAK 0x02
4909 #define LBF_SOME_PINNED 0x04
4912 struct sched_domain *sd;
4920 struct cpumask *dst_grpmask;
4922 enum cpu_idle_type idle;
4924 /* The set of CPUs under consideration for load-balancing */
4925 struct cpumask *cpus;
4930 unsigned int loop_break;
4931 unsigned int loop_max;
4935 * move_task - move a task from one runqueue to another runqueue.
4936 * Both runqueues must be locked.
4938 static void move_task(struct task_struct *p, struct lb_env *env)
4940 deactivate_task(env->src_rq, p, 0);
4941 set_task_cpu(p, env->dst_cpu);
4942 activate_task(env->dst_rq, p, 0);
4943 check_preempt_curr(env->dst_rq, p, 0);
4947 * Is this task likely cache-hot:
4950 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4954 if (p->sched_class != &fair_sched_class)
4957 if (unlikely(p->policy == SCHED_IDLE))
4961 * Buddy candidates are cache hot:
4963 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4964 (&p->se == cfs_rq_of(&p->se)->next ||
4965 &p->se == cfs_rq_of(&p->se)->last))
4968 if (sysctl_sched_migration_cost == -1)
4970 if (sysctl_sched_migration_cost == 0)
4973 delta = now - p->se.exec_start;
4975 return delta < (s64)sysctl_sched_migration_cost;
4979 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4982 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4984 int tsk_cache_hot = 0;
4986 * We do not migrate tasks that are:
4987 * 1) throttled_lb_pair, or
4988 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4989 * 3) running (obviously), or
4990 * 4) are cache-hot on their current CPU.
4992 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4995 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4998 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5001 * Remember if this task can be migrated to any other cpu in
5002 * our sched_group. We may want to revisit it if we couldn't
5003 * meet load balance goals by pulling other tasks on src_cpu.
5005 * Also avoid computing new_dst_cpu if we have already computed
5006 * one in current iteration.
5008 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
5011 /* Prevent to re-select dst_cpu via env's cpus */
5012 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5013 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5014 env->flags |= LBF_SOME_PINNED;
5015 env->new_dst_cpu = cpu;
5023 /* Record that we found atleast one task that could run on dst_cpu */
5024 env->flags &= ~LBF_ALL_PINNED;
5026 if (task_running(env->src_rq, p)) {
5027 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5032 * Aggressive migration if:
5033 * 1) task is cache cold, or
5034 * 2) too many balance attempts have failed.
5036 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
5037 if (!tsk_cache_hot ||
5038 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5040 if (tsk_cache_hot) {
5041 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5042 schedstat_inc(p, se.statistics.nr_forced_migrations);
5048 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5053 * move_one_task tries to move exactly one task from busiest to this_rq, as
5054 * part of active balancing operations within "domain".
5055 * Returns 1 if successful and 0 otherwise.
5057 * Called with both runqueues locked.
5059 static int move_one_task(struct lb_env *env)
5061 struct task_struct *p, *n;
5063 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5064 if (!can_migrate_task(p, env))
5069 * Right now, this is only the second place move_task()
5070 * is called, so we can safely collect move_task()
5071 * stats here rather than inside move_task().
5073 schedstat_inc(env->sd, lb_gained[env->idle]);
5079 static unsigned long task_h_load(struct task_struct *p);
5081 static const unsigned int sched_nr_migrate_break = 32;
5084 * move_tasks tries to move up to imbalance weighted load from busiest to
5085 * this_rq, as part of a balancing operation within domain "sd".
5086 * Returns 1 if successful and 0 otherwise.
5088 * Called with both runqueues locked.
5090 static int move_tasks(struct lb_env *env)
5092 struct list_head *tasks = &env->src_rq->cfs_tasks;
5093 struct task_struct *p;
5097 if (env->imbalance <= 0)
5100 while (!list_empty(tasks)) {
5101 p = list_first_entry(tasks, struct task_struct, se.group_node);
5104 /* We've more or less seen every task there is, call it quits */
5105 if (env->loop > env->loop_max)
5108 /* take a breather every nr_migrate tasks */
5109 if (env->loop > env->loop_break) {
5110 env->loop_break += sched_nr_migrate_break;
5111 env->flags |= LBF_NEED_BREAK;
5115 if (!can_migrate_task(p, env))
5118 load = task_h_load(p);
5120 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5123 if ((load / 2) > env->imbalance)
5128 env->imbalance -= load;
5130 #ifdef CONFIG_PREEMPT
5132 * NEWIDLE balancing is a source of latency, so preemptible
5133 * kernels will stop after the first task is pulled to minimize
5134 * the critical section.
5136 if (env->idle == CPU_NEWLY_IDLE)
5141 * We only want to steal up to the prescribed amount of
5144 if (env->imbalance <= 0)
5149 list_move_tail(&p->se.group_node, tasks);
5153 * Right now, this is one of only two places move_task() is called,
5154 * so we can safely collect move_task() stats here rather than
5155 * inside move_task().
5157 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5162 #ifdef CONFIG_FAIR_GROUP_SCHED
5164 * update tg->load_weight by folding this cpu's load_avg
5166 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5168 struct sched_entity *se = tg->se[cpu];
5169 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5171 /* throttled entities do not contribute to load */
5172 if (throttled_hierarchy(cfs_rq))
5175 update_cfs_rq_blocked_load(cfs_rq, 1);
5178 update_entity_load_avg(se, 1);
5180 * We pivot on our runnable average having decayed to zero for
5181 * list removal. This generally implies that all our children
5182 * have also been removed (modulo rounding error or bandwidth
5183 * control); however, such cases are rare and we can fix these
5186 * TODO: fix up out-of-order children on enqueue.
5188 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5189 list_del_leaf_cfs_rq(cfs_rq);
5191 struct rq *rq = rq_of(cfs_rq);
5192 update_rq_runnable_avg(rq, rq->nr_running);
5196 static void update_blocked_averages(int cpu)
5198 struct rq *rq = cpu_rq(cpu);
5199 struct cfs_rq *cfs_rq;
5200 unsigned long flags;
5202 raw_spin_lock_irqsave(&rq->lock, flags);
5203 update_rq_clock(rq);
5205 * Iterates the task_group tree in a bottom up fashion, see
5206 * list_add_leaf_cfs_rq() for details.
5208 for_each_leaf_cfs_rq(rq, cfs_rq) {
5210 * Note: We may want to consider periodically releasing
5211 * rq->lock about these updates so that creating many task
5212 * groups does not result in continually extending hold time.
5214 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5217 raw_spin_unlock_irqrestore(&rq->lock, flags);
5221 * Compute the cpu's hierarchical load factor for each task group.
5222 * This needs to be done in a top-down fashion because the load of a child
5223 * group is a fraction of its parents load.
5225 static int tg_load_down(struct task_group *tg, void *data)
5228 long cpu = (long)data;
5231 load = cpu_rq(cpu)->load.weight;
5233 load = tg->parent->cfs_rq[cpu]->h_load;
5234 load *= tg->se[cpu]->load.weight;
5235 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
5238 tg->cfs_rq[cpu]->h_load = load;
5243 static void update_h_load(long cpu)
5245 struct rq *rq = cpu_rq(cpu);
5246 unsigned long now = jiffies;
5248 if (rq->h_load_throttle == now)
5251 rq->h_load_throttle = now;
5254 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
5258 static unsigned long task_h_load(struct task_struct *p)
5260 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5263 load = p->se.load.weight;
5264 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
5269 static inline void update_blocked_averages(int cpu)
5273 static inline void update_h_load(long cpu)
5277 static unsigned long task_h_load(struct task_struct *p)
5279 return p->se.load.weight;
5283 /********** Helpers for find_busiest_group ************************/
5285 * sd_lb_stats - Structure to store the statistics of a sched_domain
5286 * during load balancing.
5288 struct sd_lb_stats {
5289 struct sched_group *busiest; /* Busiest group in this sd */
5290 struct sched_group *this; /* Local group in this sd */
5291 unsigned long total_load; /* Total load of all groups in sd */
5292 unsigned long total_pwr; /* Total power of all groups in sd */
5293 unsigned long avg_load; /* Average load across all groups in sd */
5295 /** Statistics of this group */
5296 unsigned long this_load;
5297 unsigned long this_load_per_task;
5298 unsigned long this_nr_running;
5299 unsigned long this_has_capacity;
5300 unsigned int this_idle_cpus;
5302 /* Statistics of the busiest group */
5303 unsigned int busiest_idle_cpus;
5304 unsigned long max_load;
5305 unsigned long busiest_load_per_task;
5306 unsigned long busiest_nr_running;
5307 unsigned long busiest_group_capacity;
5308 unsigned long busiest_has_capacity;
5309 unsigned int busiest_group_weight;
5311 int group_imb; /* Is there imbalance in this sd */
5315 * sg_lb_stats - stats of a sched_group required for load_balancing
5317 struct sg_lb_stats {
5318 unsigned long avg_load; /*Avg load across the CPUs of the group */
5319 unsigned long group_load; /* Total load over the CPUs of the group */
5320 unsigned long sum_nr_running; /* Nr tasks running in the group */
5321 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5322 unsigned long group_capacity;
5323 unsigned long idle_cpus;
5324 unsigned long group_weight;
5325 int group_imb; /* Is there an imbalance in the group ? */
5326 int group_has_capacity; /* Is there extra capacity in the group? */
5330 * get_sd_load_idx - Obtain the load index for a given sched domain.
5331 * @sd: The sched_domain whose load_idx is to be obtained.
5332 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5334 static inline int get_sd_load_idx(struct sched_domain *sd,
5335 enum cpu_idle_type idle)
5341 load_idx = sd->busy_idx;
5344 case CPU_NEWLY_IDLE:
5345 load_idx = sd->newidle_idx;
5348 load_idx = sd->idle_idx;
5355 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5357 return SCHED_POWER_SCALE;
5360 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5362 return default_scale_freq_power(sd, cpu);
5365 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5367 unsigned long weight = sd->span_weight;
5368 unsigned long smt_gain = sd->smt_gain;
5375 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5377 return default_scale_smt_power(sd, cpu);
5380 static unsigned long scale_rt_power(int cpu)
5382 struct rq *rq = cpu_rq(cpu);
5383 u64 total, available, age_stamp, avg;
5386 * Since we're reading these variables without serialization make sure
5387 * we read them once before doing sanity checks on them.
5389 age_stamp = ACCESS_ONCE(rq->age_stamp);
5390 avg = ACCESS_ONCE(rq->rt_avg);
5392 total = sched_avg_period() + (rq->clock - age_stamp);
5394 if (unlikely(total < avg)) {
5395 /* Ensures that power won't end up being negative */
5398 available = total - avg;
5401 if (unlikely((s64)total < SCHED_POWER_SCALE))
5402 total = SCHED_POWER_SCALE;
5404 total >>= SCHED_POWER_SHIFT;
5406 return div_u64(available, total);
5409 static void update_cpu_power(struct sched_domain *sd, int cpu)
5411 unsigned long weight = sd->span_weight;
5412 unsigned long power = SCHED_POWER_SCALE;
5413 struct sched_group *sdg = sd->groups;
5415 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5416 if (sched_feat(ARCH_POWER))
5417 power *= arch_scale_smt_power(sd, cpu);
5419 power *= default_scale_smt_power(sd, cpu);
5421 power >>= SCHED_POWER_SHIFT;
5424 sdg->sgp->power_orig = power;
5426 if (sched_feat(ARCH_POWER))
5427 power *= arch_scale_freq_power(sd, cpu);
5429 power *= default_scale_freq_power(sd, cpu);
5431 power >>= SCHED_POWER_SHIFT;
5433 power *= scale_rt_power(cpu);
5434 power >>= SCHED_POWER_SHIFT;
5439 cpu_rq(cpu)->cpu_power = power;
5440 sdg->sgp->power = power;
5443 void update_group_power(struct sched_domain *sd, int cpu)
5445 struct sched_domain *child = sd->child;
5446 struct sched_group *group, *sdg = sd->groups;
5447 unsigned long power;
5448 unsigned long interval;
5450 interval = msecs_to_jiffies(sd->balance_interval);
5451 interval = clamp(interval, 1UL, max_load_balance_interval);
5452 sdg->sgp->next_update = jiffies + interval;
5455 update_cpu_power(sd, cpu);
5461 if (child->flags & SD_OVERLAP) {
5463 * SD_OVERLAP domains cannot assume that child groups
5464 * span the current group.
5467 for_each_cpu(cpu, sched_group_cpus(sdg))
5468 power += power_of(cpu);
5471 * !SD_OVERLAP domains can assume that child groups
5472 * span the current group.
5475 group = child->groups;
5477 power += group->sgp->power;
5478 group = group->next;
5479 } while (group != child->groups);
5482 sdg->sgp->power_orig = sdg->sgp->power = power;
5486 * Try and fix up capacity for tiny siblings, this is needed when
5487 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5488 * which on its own isn't powerful enough.
5490 * See update_sd_pick_busiest() and check_asym_packing().
5493 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5496 * Only siblings can have significantly less than SCHED_POWER_SCALE
5498 if (!(sd->flags & SD_SHARE_CPUPOWER))
5502 * If ~90% of the cpu_power is still there, we're good.
5504 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5511 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5512 * @env: The load balancing environment.
5513 * @group: sched_group whose statistics are to be updated.
5514 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5515 * @local_group: Does group contain this_cpu.
5516 * @balance: Should we balance.
5517 * @sgs: variable to hold the statistics for this group.
5519 static inline void update_sg_lb_stats(struct lb_env *env,
5520 struct sched_group *group, int load_idx,
5521 int local_group, int *balance, struct sg_lb_stats *sgs)
5523 unsigned long nr_running, max_nr_running, min_nr_running;
5524 unsigned long load, max_cpu_load, min_cpu_load;
5525 unsigned int balance_cpu = -1, first_idle_cpu = 0;
5526 unsigned long avg_load_per_task = 0;
5530 balance_cpu = group_balance_cpu(group);
5532 /* Tally up the load of all CPUs in the group */
5534 min_cpu_load = ~0UL;
5536 min_nr_running = ~0UL;
5538 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5539 struct rq *rq = cpu_rq(i);
5541 nr_running = rq->nr_running;
5543 /* Bias balancing toward cpus of our domain */
5545 if (idle_cpu(i) && !first_idle_cpu &&
5546 cpumask_test_cpu(i, sched_group_mask(group))) {
5551 load = target_load(i, load_idx);
5553 load = source_load(i, load_idx);
5554 if (load > max_cpu_load)
5555 max_cpu_load = load;
5556 if (min_cpu_load > load)
5557 min_cpu_load = load;
5559 if (nr_running > max_nr_running)
5560 max_nr_running = nr_running;
5561 if (min_nr_running > nr_running)
5562 min_nr_running = nr_running;
5565 sgs->group_load += load;
5566 sgs->sum_nr_running += nr_running;
5567 sgs->sum_weighted_load += weighted_cpuload(i);
5573 * First idle cpu or the first cpu(busiest) in this sched group
5574 * is eligible for doing load balancing at this and above
5575 * domains. In the newly idle case, we will allow all the cpu's
5576 * to do the newly idle load balance.
5579 if (env->idle != CPU_NEWLY_IDLE) {
5580 if (balance_cpu != env->dst_cpu) {
5584 update_group_power(env->sd, env->dst_cpu);
5585 } else if (time_after_eq(jiffies, group->sgp->next_update))
5586 update_group_power(env->sd, env->dst_cpu);
5589 /* Adjust by relative CPU power of the group */
5590 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
5593 * Consider the group unbalanced when the imbalance is larger
5594 * than the average weight of a task.
5596 * APZ: with cgroup the avg task weight can vary wildly and
5597 * might not be a suitable number - should we keep a
5598 * normalized nr_running number somewhere that negates
5601 if (sgs->sum_nr_running)
5602 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5604 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
5605 (max_nr_running - min_nr_running) > 1)
5608 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
5610 if (!sgs->group_capacity)
5611 sgs->group_capacity = fix_small_capacity(env->sd, group);
5612 sgs->group_weight = group->group_weight;
5614 if (sgs->group_capacity > sgs->sum_nr_running)
5615 sgs->group_has_capacity = 1;
5619 * update_sd_pick_busiest - return 1 on busiest group
5620 * @env: The load balancing environment.
5621 * @sds: sched_domain statistics
5622 * @sg: sched_group candidate to be checked for being the busiest
5623 * @sgs: sched_group statistics
5625 * Determine if @sg is a busier group than the previously selected
5628 static bool update_sd_pick_busiest(struct lb_env *env,
5629 struct sd_lb_stats *sds,
5630 struct sched_group *sg,
5631 struct sg_lb_stats *sgs)
5633 if (sgs->avg_load <= sds->max_load)
5636 if (sgs->sum_nr_running > sgs->group_capacity)
5643 * ASYM_PACKING needs to move all the work to the lowest
5644 * numbered CPUs in the group, therefore mark all groups
5645 * higher than ourself as busy.
5647 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5648 env->dst_cpu < group_first_cpu(sg)) {
5652 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5660 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5661 * @env: The load balancing environment.
5662 * @balance: Should we balance.
5663 * @sds: variable to hold the statistics for this sched_domain.
5665 static inline void update_sd_lb_stats(struct lb_env *env,
5666 int *balance, struct sd_lb_stats *sds)
5668 struct sched_domain *child = env->sd->child;
5669 struct sched_group *sg = env->sd->groups;
5670 struct sg_lb_stats sgs;
5671 int load_idx, prefer_sibling = 0;
5673 if (child && child->flags & SD_PREFER_SIBLING)
5676 load_idx = get_sd_load_idx(env->sd, env->idle);
5681 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5682 memset(&sgs, 0, sizeof(sgs));
5683 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
5685 if (local_group && !(*balance))
5688 sds->total_load += sgs.group_load;
5689 sds->total_pwr += sg->sgp->power;
5692 * In case the child domain prefers tasks go to siblings
5693 * first, lower the sg capacity to one so that we'll try
5694 * and move all the excess tasks away. We lower the capacity
5695 * of a group only if the local group has the capacity to fit
5696 * these excess tasks, i.e. nr_running < group_capacity. The
5697 * extra check prevents the case where you always pull from the
5698 * heaviest group when it is already under-utilized (possible
5699 * with a large weight task outweighs the tasks on the system).
5701 if (prefer_sibling && !local_group && sds->this_has_capacity)
5702 sgs.group_capacity = min(sgs.group_capacity, 1UL);
5705 sds->this_load = sgs.avg_load;
5707 sds->this_nr_running = sgs.sum_nr_running;
5708 sds->this_load_per_task = sgs.sum_weighted_load;
5709 sds->this_has_capacity = sgs.group_has_capacity;
5710 sds->this_idle_cpus = sgs.idle_cpus;
5711 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
5712 sds->max_load = sgs.avg_load;
5714 sds->busiest_nr_running = sgs.sum_nr_running;
5715 sds->busiest_idle_cpus = sgs.idle_cpus;
5716 sds->busiest_group_capacity = sgs.group_capacity;
5717 sds->busiest_load_per_task = sgs.sum_weighted_load;
5718 sds->busiest_has_capacity = sgs.group_has_capacity;
5719 sds->busiest_group_weight = sgs.group_weight;
5720 sds->group_imb = sgs.group_imb;
5724 } while (sg != env->sd->groups);
5728 * check_asym_packing - Check to see if the group is packed into the
5731 * This is primarily intended to used at the sibling level. Some
5732 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5733 * case of POWER7, it can move to lower SMT modes only when higher
5734 * threads are idle. When in lower SMT modes, the threads will
5735 * perform better since they share less core resources. Hence when we
5736 * have idle threads, we want them to be the higher ones.
5738 * This packing function is run on idle threads. It checks to see if
5739 * the busiest CPU in this domain (core in the P7 case) has a higher
5740 * CPU number than the packing function is being run on. Here we are
5741 * assuming lower CPU number will be equivalent to lower a SMT thread
5744 * Returns 1 when packing is required and a task should be moved to
5745 * this CPU. The amount of the imbalance is returned in *imbalance.
5747 * @env: The load balancing environment.
5748 * @sds: Statistics of the sched_domain which is to be packed
5750 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5754 if (!(env->sd->flags & SD_ASYM_PACKING))
5760 busiest_cpu = group_first_cpu(sds->busiest);
5761 if (env->dst_cpu > busiest_cpu)
5764 env->imbalance = DIV_ROUND_CLOSEST(
5765 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
5771 * fix_small_imbalance - Calculate the minor imbalance that exists
5772 * amongst the groups of a sched_domain, during
5774 * @env: The load balancing environment.
5775 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5778 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5780 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5781 unsigned int imbn = 2;
5782 unsigned long scaled_busy_load_per_task;
5784 if (sds->this_nr_running) {
5785 sds->this_load_per_task /= sds->this_nr_running;
5786 if (sds->busiest_load_per_task >
5787 sds->this_load_per_task)
5790 sds->this_load_per_task =
5791 cpu_avg_load_per_task(env->dst_cpu);
5794 scaled_busy_load_per_task = sds->busiest_load_per_task
5795 * SCHED_POWER_SCALE;
5796 scaled_busy_load_per_task /= sds->busiest->sgp->power;
5798 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
5799 (scaled_busy_load_per_task * imbn)) {
5800 env->imbalance = sds->busiest_load_per_task;
5805 * OK, we don't have enough imbalance to justify moving tasks,
5806 * however we may be able to increase total CPU power used by
5810 pwr_now += sds->busiest->sgp->power *
5811 min(sds->busiest_load_per_task, sds->max_load);
5812 pwr_now += sds->this->sgp->power *
5813 min(sds->this_load_per_task, sds->this_load);
5814 pwr_now /= SCHED_POWER_SCALE;
5816 /* Amount of load we'd subtract */
5817 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5818 sds->busiest->sgp->power;
5819 if (sds->max_load > tmp)
5820 pwr_move += sds->busiest->sgp->power *
5821 min(sds->busiest_load_per_task, sds->max_load - tmp);
5823 /* Amount of load we'd add */
5824 if (sds->max_load * sds->busiest->sgp->power <
5825 sds->busiest_load_per_task * SCHED_POWER_SCALE)
5826 tmp = (sds->max_load * sds->busiest->sgp->power) /
5827 sds->this->sgp->power;
5829 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5830 sds->this->sgp->power;
5831 pwr_move += sds->this->sgp->power *
5832 min(sds->this_load_per_task, sds->this_load + tmp);
5833 pwr_move /= SCHED_POWER_SCALE;
5835 /* Move if we gain throughput */
5836 if (pwr_move > pwr_now)
5837 env->imbalance = sds->busiest_load_per_task;
5841 * calculate_imbalance - Calculate the amount of imbalance present within the
5842 * groups of a given sched_domain during load balance.
5843 * @env: load balance environment
5844 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5846 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5848 unsigned long max_pull, load_above_capacity = ~0UL;
5850 sds->busiest_load_per_task /= sds->busiest_nr_running;
5851 if (sds->group_imb) {
5852 sds->busiest_load_per_task =
5853 min(sds->busiest_load_per_task, sds->avg_load);
5857 * In the presence of smp nice balancing, certain scenarios can have
5858 * max load less than avg load(as we skip the groups at or below
5859 * its cpu_power, while calculating max_load..)
5861 if (sds->max_load < sds->avg_load) {
5863 return fix_small_imbalance(env, sds);
5866 if (!sds->group_imb) {
5868 * Don't want to pull so many tasks that a group would go idle.
5870 load_above_capacity = (sds->busiest_nr_running -
5871 sds->busiest_group_capacity);
5873 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5875 load_above_capacity /= sds->busiest->sgp->power;
5879 * We're trying to get all the cpus to the average_load, so we don't
5880 * want to push ourselves above the average load, nor do we wish to
5881 * reduce the max loaded cpu below the average load. At the same time,
5882 * we also don't want to reduce the group load below the group capacity
5883 * (so that we can implement power-savings policies etc). Thus we look
5884 * for the minimum possible imbalance.
5885 * Be careful of negative numbers as they'll appear as very large values
5886 * with unsigned longs.
5888 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
5890 /* How much load to actually move to equalise the imbalance */
5891 env->imbalance = min(max_pull * sds->busiest->sgp->power,
5892 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
5893 / SCHED_POWER_SCALE;
5896 * if *imbalance is less than the average load per runnable task
5897 * there is no guarantee that any tasks will be moved so we'll have
5898 * a think about bumping its value to force at least one task to be
5901 if (env->imbalance < sds->busiest_load_per_task)
5902 return fix_small_imbalance(env, sds);
5906 /******* find_busiest_group() helpers end here *********************/
5909 * find_busiest_group - Returns the busiest group within the sched_domain
5910 * if there is an imbalance. If there isn't an imbalance, and
5911 * the user has opted for power-savings, it returns a group whose
5912 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5913 * such a group exists.
5915 * Also calculates the amount of weighted load which should be moved
5916 * to restore balance.
5918 * @env: The load balancing environment.
5919 * @balance: Pointer to a variable indicating if this_cpu
5920 * is the appropriate cpu to perform load balancing at this_level.
5922 * Returns: - the busiest group if imbalance exists.
5923 * - If no imbalance and user has opted for power-savings balance,
5924 * return the least loaded group whose CPUs can be
5925 * put to idle by rebalancing its tasks onto our group.
5927 static struct sched_group *
5928 find_busiest_group(struct lb_env *env, int *balance)
5930 struct sd_lb_stats sds;
5932 memset(&sds, 0, sizeof(sds));
5935 * Compute the various statistics relavent for load balancing at
5938 update_sd_lb_stats(env, balance, &sds);
5941 * this_cpu is not the appropriate cpu to perform load balancing at
5947 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5948 check_asym_packing(env, &sds))
5951 /* There is no busy sibling group to pull tasks from */
5952 if (!sds.busiest || sds.busiest_nr_running == 0)
5955 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5958 * If the busiest group is imbalanced the below checks don't
5959 * work because they assumes all things are equal, which typically
5960 * isn't true due to cpus_allowed constraints and the like.
5965 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5966 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5967 !sds.busiest_has_capacity)
5971 * If the local group is more busy than the selected busiest group
5972 * don't try and pull any tasks.
5974 if (sds.this_load >= sds.max_load)
5978 * Don't pull any tasks if this group is already above the domain
5981 if (sds.this_load >= sds.avg_load)
5984 if (env->idle == CPU_IDLE) {
5986 * This cpu is idle. If the busiest group load doesn't
5987 * have more tasks than the number of available cpu's and
5988 * there is no imbalance between this and busiest group
5989 * wrt to idle cpu's, it is balanced.
5991 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
5992 sds.busiest_nr_running <= sds.busiest_group_weight)
5996 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5997 * imbalance_pct to be conservative.
5999 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
6004 /* Looks like there is an imbalance. Compute it */
6005 calculate_imbalance(env, &sds);
6015 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6017 static struct rq *find_busiest_queue(struct lb_env *env,
6018 struct sched_group *group)
6020 struct rq *busiest = NULL, *rq;
6021 unsigned long max_load = 0;
6024 for_each_cpu(i, sched_group_cpus(group)) {
6025 unsigned long power = power_of(i);
6026 unsigned long capacity = DIV_ROUND_CLOSEST(power,
6031 capacity = fix_small_capacity(env->sd, group);
6033 if (!cpumask_test_cpu(i, env->cpus))
6037 wl = weighted_cpuload(i);
6040 * When comparing with imbalance, use weighted_cpuload()
6041 * which is not scaled with the cpu power.
6043 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6047 * For the load comparisons with the other cpu's, consider
6048 * the weighted_cpuload() scaled with the cpu power, so that
6049 * the load can be moved away from the cpu that is potentially
6050 * running at a lower capacity.
6052 wl = (wl * SCHED_POWER_SCALE) / power;
6054 if (wl > max_load) {
6064 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6065 * so long as it is large enough.
6067 #define MAX_PINNED_INTERVAL 512
6069 /* Working cpumask for load_balance and load_balance_newidle. */
6070 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6072 static int need_active_balance(struct lb_env *env)
6074 struct sched_domain *sd = env->sd;
6076 if (env->idle == CPU_NEWLY_IDLE) {
6079 * ASYM_PACKING needs to force migrate tasks from busy but
6080 * higher numbered CPUs in order to pack all tasks in the
6081 * lowest numbered CPUs.
6083 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6087 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6090 static int active_load_balance_cpu_stop(void *data);
6093 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6094 * tasks if there is an imbalance.
6096 static int load_balance(int this_cpu, struct rq *this_rq,
6097 struct sched_domain *sd, enum cpu_idle_type idle,
6100 int ld_moved, cur_ld_moved, active_balance = 0;
6101 struct sched_group *group;
6103 unsigned long flags;
6104 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6106 struct lb_env env = {
6108 .dst_cpu = this_cpu,
6110 .dst_grpmask = sched_group_cpus(sd->groups),
6112 .loop_break = sched_nr_migrate_break,
6117 * For NEWLY_IDLE load_balancing, we don't need to consider
6118 * other cpus in our group
6120 if (idle == CPU_NEWLY_IDLE)
6121 env.dst_grpmask = NULL;
6123 cpumask_copy(cpus, cpu_active_mask);
6125 schedstat_inc(sd, lb_count[idle]);
6128 group = find_busiest_group(&env, balance);
6134 schedstat_inc(sd, lb_nobusyg[idle]);
6138 busiest = find_busiest_queue(&env, group);
6140 schedstat_inc(sd, lb_nobusyq[idle]);
6144 BUG_ON(busiest == env.dst_rq);
6146 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6149 if (busiest->nr_running > 1) {
6151 * Attempt to move tasks. If find_busiest_group has found
6152 * an imbalance but busiest->nr_running <= 1, the group is
6153 * still unbalanced. ld_moved simply stays zero, so it is
6154 * correctly treated as an imbalance.
6156 env.flags |= LBF_ALL_PINNED;
6157 env.src_cpu = busiest->cpu;
6158 env.src_rq = busiest;
6159 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6161 update_h_load(env.src_cpu);
6163 local_irq_save(flags);
6164 double_rq_lock(env.dst_rq, busiest);
6167 * cur_ld_moved - load moved in current iteration
6168 * ld_moved - cumulative load moved across iterations
6170 cur_ld_moved = move_tasks(&env);
6171 ld_moved += cur_ld_moved;
6172 double_rq_unlock(env.dst_rq, busiest);
6173 local_irq_restore(flags);
6176 * some other cpu did the load balance for us.
6178 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6179 resched_cpu(env.dst_cpu);
6181 if (env.flags & LBF_NEED_BREAK) {
6182 env.flags &= ~LBF_NEED_BREAK;
6187 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6188 * us and move them to an alternate dst_cpu in our sched_group
6189 * where they can run. The upper limit on how many times we
6190 * iterate on same src_cpu is dependent on number of cpus in our
6193 * This changes load balance semantics a bit on who can move
6194 * load to a given_cpu. In addition to the given_cpu itself
6195 * (or a ilb_cpu acting on its behalf where given_cpu is
6196 * nohz-idle), we now have balance_cpu in a position to move
6197 * load to given_cpu. In rare situations, this may cause
6198 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6199 * _independently_ and at _same_ time to move some load to
6200 * given_cpu) causing exceess load to be moved to given_cpu.
6201 * This however should not happen so much in practice and
6202 * moreover subsequent load balance cycles should correct the
6203 * excess load moved.
6205 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6207 env.dst_rq = cpu_rq(env.new_dst_cpu);
6208 env.dst_cpu = env.new_dst_cpu;
6209 env.flags &= ~LBF_SOME_PINNED;
6211 env.loop_break = sched_nr_migrate_break;
6213 /* Prevent to re-select dst_cpu via env's cpus */
6214 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6217 * Go back to "more_balance" rather than "redo" since we
6218 * need to continue with same src_cpu.
6223 /* All tasks on this runqueue were pinned by CPU affinity */
6224 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6225 cpumask_clear_cpu(cpu_of(busiest), cpus);
6226 if (!cpumask_empty(cpus)) {
6228 env.loop_break = sched_nr_migrate_break;
6236 schedstat_inc(sd, lb_failed[idle]);
6238 * Increment the failure counter only on periodic balance.
6239 * We do not want newidle balance, which can be very
6240 * frequent, pollute the failure counter causing
6241 * excessive cache_hot migrations and active balances.
6243 if (idle != CPU_NEWLY_IDLE)
6244 sd->nr_balance_failed++;
6246 if (need_active_balance(&env)) {
6247 raw_spin_lock_irqsave(&busiest->lock, flags);
6249 /* don't kick the active_load_balance_cpu_stop,
6250 * if the curr task on busiest cpu can't be
6253 if (!cpumask_test_cpu(this_cpu,
6254 tsk_cpus_allowed(busiest->curr))) {
6255 raw_spin_unlock_irqrestore(&busiest->lock,
6257 env.flags |= LBF_ALL_PINNED;
6258 goto out_one_pinned;
6262 * ->active_balance synchronizes accesses to
6263 * ->active_balance_work. Once set, it's cleared
6264 * only after active load balance is finished.
6266 if (!busiest->active_balance) {
6267 busiest->active_balance = 1;
6268 busiest->push_cpu = this_cpu;
6271 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6273 if (active_balance) {
6274 stop_one_cpu_nowait(cpu_of(busiest),
6275 active_load_balance_cpu_stop, busiest,
6276 &busiest->active_balance_work);
6280 * We've kicked active balancing, reset the failure
6283 sd->nr_balance_failed = sd->cache_nice_tries+1;
6286 sd->nr_balance_failed = 0;
6288 if (likely(!active_balance)) {
6289 /* We were unbalanced, so reset the balancing interval */
6290 sd->balance_interval = sd->min_interval;
6293 * If we've begun active balancing, start to back off. This
6294 * case may not be covered by the all_pinned logic if there
6295 * is only 1 task on the busy runqueue (because we don't call
6298 if (sd->balance_interval < sd->max_interval)
6299 sd->balance_interval *= 2;
6305 schedstat_inc(sd, lb_balanced[idle]);
6307 sd->nr_balance_failed = 0;
6310 /* tune up the balancing interval */
6311 if (((env.flags & LBF_ALL_PINNED) &&
6312 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6313 (sd->balance_interval < sd->max_interval))
6314 sd->balance_interval *= 2;
6321 #ifdef CONFIG_SCHED_HMP
6322 static unsigned int hmp_idle_pull(int this_cpu);
6323 static int move_specific_task(struct lb_env *env, struct task_struct *pm);
6325 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
6332 * idle_balance is called by schedule() if this_cpu is about to become
6333 * idle. Attempts to pull tasks from other CPUs.
6335 void idle_balance(int this_cpu, struct rq *this_rq)
6337 struct sched_domain *sd;
6338 int pulled_task = 0;
6339 unsigned long next_balance = jiffies + HZ;
6341 this_rq->idle_stamp = this_rq->clock;
6343 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6347 * Drop the rq->lock, but keep IRQ/preempt disabled.
6349 raw_spin_unlock(&this_rq->lock);
6351 update_blocked_averages(this_cpu);
6353 for_each_domain(this_cpu, sd) {
6354 unsigned long interval;
6357 if (!(sd->flags & SD_LOAD_BALANCE))
6360 if (sd->flags & SD_BALANCE_NEWIDLE) {
6361 /* If we've pulled tasks over stop searching: */
6362 pulled_task = load_balance(this_cpu, this_rq,
6363 sd, CPU_NEWLY_IDLE, &balance);
6366 interval = msecs_to_jiffies(sd->balance_interval);
6367 if (time_after(next_balance, sd->last_balance + interval))
6368 next_balance = sd->last_balance + interval;
6370 this_rq->idle_stamp = 0;
6375 #ifdef CONFIG_SCHED_HMP
6377 pulled_task = hmp_idle_pull(this_cpu);
6379 raw_spin_lock(&this_rq->lock);
6381 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6383 * We are going idle. next_balance may be set based on
6384 * a busy processor. So reset next_balance.
6386 this_rq->next_balance = next_balance;
6390 static int __do_active_load_balance_cpu_stop(void *data, bool check_sd_lb_flag)
6392 struct rq *busiest_rq = data;
6393 int busiest_cpu = cpu_of(busiest_rq);
6394 int target_cpu = busiest_rq->push_cpu;
6395 struct rq *target_rq = cpu_rq(target_cpu);
6396 struct sched_domain *sd;
6397 struct task_struct *p = NULL;
6399 raw_spin_lock_irq(&busiest_rq->lock);
6400 #ifdef CONFIG_SCHED_HMP
6401 p = busiest_rq->migrate_task;
6403 /* make sure the requested cpu hasn't gone down in the meantime */
6404 if (unlikely(busiest_cpu != smp_processor_id() ||
6405 !busiest_rq->active_balance))
6408 /* Is there any task to move? */
6409 if (busiest_rq->nr_running <= 1)
6412 if (!check_sd_lb_flag) {
6413 /* Task has migrated meanwhile, abort forced migration */
6414 if (task_rq(p) != busiest_rq)
6418 * This condition is "impossible", if it occurs
6419 * we need to fix it. Originally reported by
6420 * Bjorn Helgaas on a 128-cpu setup.
6422 BUG_ON(busiest_rq == target_rq);
6424 /* move a task from busiest_rq to target_rq */
6425 double_lock_balance(busiest_rq, target_rq);
6427 /* Search for an sd spanning us and the target CPU. */
6429 for_each_domain(target_cpu, sd) {
6430 if (((check_sd_lb_flag && sd->flags & SD_LOAD_BALANCE) ||
6431 !check_sd_lb_flag) &&
6432 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6437 bool success = false;
6438 struct lb_env env = {
6440 .dst_cpu = target_cpu,
6441 .dst_rq = target_rq,
6442 .src_cpu = busiest_rq->cpu,
6443 .src_rq = busiest_rq,
6447 schedstat_inc(sd, alb_count);
6449 if (check_sd_lb_flag) {
6450 if (move_one_task(&env))
6453 if (move_specific_task(&env, p))
6457 schedstat_inc(sd, alb_pushed);
6459 schedstat_inc(sd, alb_failed);
6462 double_unlock_balance(busiest_rq, target_rq);
6464 if (!check_sd_lb_flag)
6466 busiest_rq->active_balance = 0;
6467 raw_spin_unlock_irq(&busiest_rq->lock);
6472 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6473 * running tasks off the busiest CPU onto idle CPUs. It requires at
6474 * least 1 task to be running on each physical CPU where possible, and
6475 * avoids physical / logical imbalances.
6477 static int active_load_balance_cpu_stop(void *data)
6479 return __do_active_load_balance_cpu_stop(data, true);
6482 #ifdef CONFIG_NO_HZ_COMMON
6484 * idle load balancing details
6485 * - When one of the busy CPUs notice that there may be an idle rebalancing
6486 * needed, they will kick the idle load balancer, which then does idle
6487 * load balancing for all the idle CPUs.
6490 cpumask_var_t idle_cpus_mask;
6492 unsigned long next_balance; /* in jiffy units */
6493 } nohz ____cacheline_aligned;
6496 * nohz_test_cpu used when load tracking is enabled. FAIR_GROUP_SCHED
6497 * dependency below may be removed when load tracking guards are
6500 #ifdef CONFIG_FAIR_GROUP_SCHED
6501 static int nohz_test_cpu(int cpu)
6503 return cpumask_test_cpu(cpu, nohz.idle_cpus_mask);
6507 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6509 * Decide if the tasks on the busy CPUs in the
6510 * littlest domain would benefit from an idle balance
6512 static int hmp_packing_ilb_needed(int cpu)
6514 struct hmp_domain *hmp;
6515 /* always allow ilb on non-slowest domain */
6516 if (!hmp_cpu_is_slowest(cpu))
6519 /* if disabled, use normal ILB behaviour */
6520 if (!hmp_packing_enabled)
6523 hmp = hmp_cpu_domain(cpu);
6524 for_each_cpu_and(cpu, &hmp->cpus, nohz.idle_cpus_mask) {
6525 /* only idle balance if a CPU is loaded over threshold */
6526 if (cpu_rq(cpu)->avg.load_avg_ratio > hmp_full_threshold)
6533 static inline int find_new_ilb(int call_cpu)
6535 int ilb = cpumask_first(nohz.idle_cpus_mask);
6536 #ifdef CONFIG_SCHED_HMP
6539 /* restrict nohz balancing to occur in the same hmp domain */
6540 ilb = cpumask_first_and(nohz.idle_cpus_mask,
6541 &((struct hmp_domain *)hmp_cpu_domain(call_cpu))->cpus);
6543 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6544 if (ilb < nr_cpu_ids)
6545 ilb_needed = hmp_packing_ilb_needed(ilb);
6548 if (ilb_needed && ilb < nr_cpu_ids && idle_cpu(ilb))
6551 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6559 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6560 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6561 * CPU (if there is one).
6563 static void nohz_balancer_kick(int cpu)
6567 nohz.next_balance++;
6569 ilb_cpu = find_new_ilb(cpu);
6571 if (ilb_cpu >= nr_cpu_ids)
6574 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6577 * Use smp_send_reschedule() instead of resched_cpu().
6578 * This way we generate a sched IPI on the target cpu which
6579 * is idle. And the softirq performing nohz idle load balance
6580 * will be run before returning from the IPI.
6582 smp_send_reschedule(ilb_cpu);
6586 static inline void nohz_balance_exit_idle(int cpu)
6588 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6589 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6590 atomic_dec(&nohz.nr_cpus);
6591 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6595 static inline void set_cpu_sd_state_busy(void)
6597 struct sched_domain *sd;
6598 int cpu = smp_processor_id();
6601 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6603 if (!sd || !sd->nohz_idle)
6607 for (; sd; sd = sd->parent)
6608 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6613 void set_cpu_sd_state_idle(void)
6615 struct sched_domain *sd;
6616 int cpu = smp_processor_id();
6619 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6621 if (!sd || sd->nohz_idle)
6625 for (; sd; sd = sd->parent)
6626 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6632 * This routine will record that the cpu is going idle with tick stopped.
6633 * This info will be used in performing idle load balancing in the future.
6635 void nohz_balance_enter_idle(int cpu)
6638 * If this cpu is going down, then nothing needs to be done.
6640 if (!cpu_active(cpu))
6643 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6646 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6647 atomic_inc(&nohz.nr_cpus);
6648 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6651 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
6652 unsigned long action, void *hcpu)
6654 switch (action & ~CPU_TASKS_FROZEN) {
6656 nohz_balance_exit_idle(smp_processor_id());
6664 static DEFINE_SPINLOCK(balancing);
6667 * Scale the max load_balance interval with the number of CPUs in the system.
6668 * This trades load-balance latency on larger machines for less cross talk.
6670 void update_max_interval(void)
6672 max_load_balance_interval = HZ*num_online_cpus()/10;
6676 * It checks each scheduling domain to see if it is due to be balanced,
6677 * and initiates a balancing operation if so.
6679 * Balancing parameters are set up in init_sched_domains.
6681 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6684 struct rq *rq = cpu_rq(cpu);
6685 unsigned long interval;
6686 struct sched_domain *sd;
6687 /* Earliest time when we have to do rebalance again */
6688 unsigned long next_balance = jiffies + 60*HZ;
6689 int update_next_balance = 0;
6692 update_blocked_averages(cpu);
6695 for_each_domain(cpu, sd) {
6696 if (!(sd->flags & SD_LOAD_BALANCE))
6699 interval = sd->balance_interval;
6700 if (idle != CPU_IDLE)
6701 interval *= sd->busy_factor;
6703 /* scale ms to jiffies */
6704 interval = msecs_to_jiffies(interval);
6705 interval = clamp(interval, 1UL, max_load_balance_interval);
6707 need_serialize = sd->flags & SD_SERIALIZE;
6709 if (need_serialize) {
6710 if (!spin_trylock(&balancing))
6714 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6715 if (load_balance(cpu, rq, sd, idle, &balance)) {
6717 * The LBF_SOME_PINNED logic could have changed
6718 * env->dst_cpu, so we can't know our idle
6719 * state even if we migrated tasks. Update it.
6721 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6723 sd->last_balance = jiffies;
6726 spin_unlock(&balancing);
6728 if (time_after(next_balance, sd->last_balance + interval)) {
6729 next_balance = sd->last_balance + interval;
6730 update_next_balance = 1;
6734 * Stop the load balance at this level. There is another
6735 * CPU in our sched group which is doing load balancing more
6744 * next_balance will be updated only when there is a need.
6745 * When the cpu is attached to null domain for ex, it will not be
6748 if (likely(update_next_balance))
6749 rq->next_balance = next_balance;
6752 #ifdef CONFIG_NO_HZ_COMMON
6754 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6755 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6757 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6759 struct rq *this_rq = cpu_rq(this_cpu);
6763 if (idle != CPU_IDLE ||
6764 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6767 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6768 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6772 * If this cpu gets work to do, stop the load balancing
6773 * work being done for other cpus. Next load
6774 * balancing owner will pick it up.
6779 rq = cpu_rq(balance_cpu);
6781 raw_spin_lock_irq(&rq->lock);
6782 update_rq_clock(rq);
6783 update_idle_cpu_load(rq);
6784 raw_spin_unlock_irq(&rq->lock);
6786 rebalance_domains(balance_cpu, CPU_IDLE);
6788 if (time_after(this_rq->next_balance, rq->next_balance))
6789 this_rq->next_balance = rq->next_balance;
6791 nohz.next_balance = this_rq->next_balance;
6793 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6797 * Current heuristic for kicking the idle load balancer in the presence
6798 * of an idle cpu is the system.
6799 * - This rq has more than one task.
6800 * - At any scheduler domain level, this cpu's scheduler group has multiple
6801 * busy cpu's exceeding the group's power.
6802 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6803 * domain span are idle.
6805 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6807 unsigned long now = jiffies;
6808 struct sched_domain *sd;
6810 if (unlikely(idle_cpu(cpu)))
6814 * We may be recently in ticked or tickless idle mode. At the first
6815 * busy tick after returning from idle, we will update the busy stats.
6817 set_cpu_sd_state_busy();
6818 nohz_balance_exit_idle(cpu);
6821 * None are in tickless mode and hence no need for NOHZ idle load
6824 if (likely(!atomic_read(&nohz.nr_cpus)))
6827 if (time_before(now, nohz.next_balance))
6830 #ifdef CONFIG_SCHED_HMP
6832 * Bail out if there are no nohz CPUs in our
6833 * HMP domain, since we will move tasks between
6834 * domains through wakeup and force balancing
6835 * as necessary based upon task load.
6837 if (cpumask_first_and(nohz.idle_cpus_mask,
6838 &((struct hmp_domain *)hmp_cpu_domain(cpu))->cpus) >= nr_cpu_ids)
6842 if (rq->nr_running >= 2)
6846 for_each_domain(cpu, sd) {
6847 struct sched_group *sg = sd->groups;
6848 struct sched_group_power *sgp = sg->sgp;
6849 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6851 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6852 goto need_kick_unlock;
6854 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6855 && (cpumask_first_and(nohz.idle_cpus_mask,
6856 sched_domain_span(sd)) < cpu))
6857 goto need_kick_unlock;
6859 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6871 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6874 #ifdef CONFIG_SCHED_HMP
6875 static unsigned int hmp_task_eligible_for_up_migration(struct sched_entity *se)
6877 /* below hmp_up_threshold, never eligible */
6878 if (se->avg.load_avg_ratio < hmp_up_threshold)
6883 /* Check if task should migrate to a faster cpu */
6884 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se)
6886 struct task_struct *p = task_of(se);
6887 int temp_target_cpu;
6890 if (hmp_cpu_is_fastest(cpu))
6893 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6894 /* Filter by task priority */
6895 if (p->prio >= hmp_up_prio)
6898 if (!hmp_task_eligible_for_up_migration(se))
6901 /* Let the task load settle before doing another up migration */
6902 /* hack - always use clock from first online CPU */
6903 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6904 if (((now - se->avg.hmp_last_up_migration) >> 10)
6905 < hmp_next_up_threshold)
6908 /* hmp_domain_min_load only returns 0 for an
6909 * idle CPU or 1023 for any partly-busy one.
6910 * Be explicit about requirement for an idle CPU.
6912 if (hmp_domain_min_load(hmp_faster_domain(cpu), &temp_target_cpu,
6913 tsk_cpus_allowed(p)) == 0 && temp_target_cpu != NR_CPUS) {
6915 *target_cpu = temp_target_cpu;
6921 /* Check if task should migrate to a slower cpu */
6922 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
6924 struct task_struct *p = task_of(se);
6927 if (hmp_cpu_is_slowest(cpu)) {
6928 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6929 if(hmp_packing_enabled)
6936 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6937 /* Filter by task priority */
6938 if ((p->prio >= hmp_up_prio) &&
6939 cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6940 tsk_cpus_allowed(p))) {
6945 /* Let the task load settle before doing another down migration */
6946 /* hack - always use clock from first online CPU */
6947 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6948 if (((now - se->avg.hmp_last_down_migration) >> 10)
6949 < hmp_next_down_threshold)
6952 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6953 tsk_cpus_allowed(p))
6954 && se->avg.load_avg_ratio < hmp_down_threshold) {
6961 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6962 * Ideally this function should be merged with can_migrate_task() to avoid
6965 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
6967 int tsk_cache_hot = 0;
6970 * We do not migrate tasks that are:
6971 * 1) running (obviously), or
6972 * 2) cannot be migrated to this CPU due to cpus_allowed
6974 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6975 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6978 env->flags &= ~LBF_ALL_PINNED;
6980 if (task_running(env->src_rq, p)) {
6981 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6986 * Aggressive migration if:
6987 * 1) task is cache cold, or
6988 * 2) too many balance attempts have failed.
6991 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
6992 if (!tsk_cache_hot ||
6993 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6994 #ifdef CONFIG_SCHEDSTATS
6995 if (tsk_cache_hot) {
6996 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6997 schedstat_inc(p, se.statistics.nr_forced_migrations);
7007 * move_specific_task tries to move a specific task.
7008 * Returns 1 if successful and 0 otherwise.
7009 * Called with both runqueues locked.
7011 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
7013 struct task_struct *p, *n;
7015 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7016 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
7020 if (!hmp_can_migrate_task(p, env))
7022 /* Check if we found the right task */
7028 * Right now, this is only the third place move_task()
7029 * is called, so we can safely collect move_task()
7030 * stats here rather than inside move_task().
7032 schedstat_inc(env->sd, lb_gained[env->idle]);
7039 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
7040 * migrate a specific task from one runqueue to another.
7041 * hmp_force_up_migration uses this to push a currently running task
7042 * off a runqueue. hmp_idle_pull uses this to pull a currently
7043 * running task to an idle runqueue.
7044 * Reuses __do_active_load_balance_cpu_stop to actually do the work.
7046 static int hmp_active_task_migration_cpu_stop(void *data)
7048 return __do_active_load_balance_cpu_stop(data, false);
7052 * Move task in a runnable state to another CPU.
7054 * Tailored on 'active_load_balance_cpu_stop' with slight
7055 * modification to locking and pre-transfer checks. Note
7056 * rq->lock must be held before calling.
7058 static void hmp_migrate_runnable_task(struct rq *rq)
7060 struct sched_domain *sd;
7061 int src_cpu = cpu_of(rq);
7062 struct rq *src_rq = rq;
7063 int dst_cpu = rq->push_cpu;
7064 struct rq *dst_rq = cpu_rq(dst_cpu);
7065 struct task_struct *p = rq->migrate_task;
7067 * One last check to make sure nobody else is playing
7068 * with the source rq.
7070 if (src_rq->active_balance)
7073 if (src_rq->nr_running <= 1)
7076 if (task_rq(p) != src_rq)
7079 * Not sure if this applies here but one can never
7082 BUG_ON(src_rq == dst_rq);
7084 double_lock_balance(src_rq, dst_rq);
7087 for_each_domain(dst_cpu, sd) {
7088 if (cpumask_test_cpu(src_cpu, sched_domain_span(sd)))
7093 struct lb_env env = {
7102 schedstat_inc(sd, alb_count);
7104 if (move_specific_task(&env, p))
7105 schedstat_inc(sd, alb_pushed);
7107 schedstat_inc(sd, alb_failed);
7111 double_unlock_balance(src_rq, dst_rq);
7116 static DEFINE_SPINLOCK(hmp_force_migration);
7119 * hmp_force_up_migration checks runqueues for tasks that need to
7120 * be actively migrated to a faster cpu.
7122 static void hmp_force_up_migration(int this_cpu)
7124 int cpu, target_cpu;
7125 struct sched_entity *curr, *orig;
7127 unsigned long flags;
7128 unsigned int force, got_target;
7129 struct task_struct *p;
7131 if (!spin_trylock(&hmp_force_migration))
7133 for_each_online_cpu(cpu) {
7136 target = cpu_rq(cpu);
7137 raw_spin_lock_irqsave(&target->lock, flags);
7138 curr = target->cfs.curr;
7139 if (!curr || target->active_balance) {
7140 raw_spin_unlock_irqrestore(&target->lock, flags);
7143 if (!entity_is_task(curr)) {
7144 struct cfs_rq *cfs_rq;
7146 cfs_rq = group_cfs_rq(curr);
7148 curr = cfs_rq->curr;
7149 cfs_rq = group_cfs_rq(curr);
7153 curr = hmp_get_heaviest_task(curr, -1);
7155 raw_spin_unlock_irqrestore(&target->lock, flags);
7159 if (hmp_up_migration(cpu, &target_cpu, curr)) {
7160 cpu_rq(target_cpu)->wake_for_idle_pull = 1;
7161 raw_spin_unlock_irqrestore(&target->lock, flags);
7162 spin_unlock(&hmp_force_migration);
7163 smp_send_reschedule(target_cpu);
7168 * For now we just check the currently running task.
7169 * Selecting the lightest task for offloading will
7170 * require extensive book keeping.
7172 curr = hmp_get_lightest_task(orig, 1);
7174 target->push_cpu = hmp_offload_down(cpu, curr);
7175 if (target->push_cpu < NR_CPUS) {
7177 target->migrate_task = p;
7179 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_OFFLOAD);
7180 hmp_next_down_delay(&p->se, target->push_cpu);
7184 * We have a target with no active_balance. If the task
7185 * is not currently running move it, otherwise let the
7186 * CPU stopper take care of it.
7189 if (!task_running(target, p)) {
7190 trace_sched_hmp_migrate_force_running(p, 0);
7191 hmp_migrate_runnable_task(target);
7193 target->active_balance = 1;
7198 raw_spin_unlock_irqrestore(&target->lock, flags);
7201 stop_one_cpu_nowait(cpu_of(target),
7202 hmp_active_task_migration_cpu_stop,
7203 target, &target->active_balance_work);
7205 spin_unlock(&hmp_force_migration);
7208 * hmp_idle_pull looks at little domain runqueues to see
7209 * if a task should be pulled.
7211 * Reuses hmp_force_migration spinlock.
7214 static unsigned int hmp_idle_pull(int this_cpu)
7217 struct sched_entity *curr, *orig;
7218 struct hmp_domain *hmp_domain = NULL;
7219 struct rq *target = NULL, *rq;
7220 unsigned long flags, ratio = 0;
7221 unsigned int force = 0;
7222 struct task_struct *p = NULL;
7224 if (!hmp_cpu_is_slowest(this_cpu))
7225 hmp_domain = hmp_slower_domain(this_cpu);
7229 if (!spin_trylock(&hmp_force_migration))
7232 /* first select a task */
7233 for_each_cpu(cpu, &hmp_domain->cpus) {
7235 raw_spin_lock_irqsave(&rq->lock, flags);
7236 curr = rq->cfs.curr;
7238 raw_spin_unlock_irqrestore(&rq->lock, flags);
7241 if (!entity_is_task(curr)) {
7242 struct cfs_rq *cfs_rq;
7244 cfs_rq = group_cfs_rq(curr);
7246 curr = cfs_rq->curr;
7247 if (!entity_is_task(curr))
7248 cfs_rq = group_cfs_rq(curr);
7254 curr = hmp_get_heaviest_task(curr, this_cpu);
7255 /* check if heaviest eligible task on this
7256 * CPU is heavier than previous task
7258 if (curr && hmp_task_eligible_for_up_migration(curr) &&
7259 curr->avg.load_avg_ratio > ratio &&
7260 cpumask_test_cpu(this_cpu,
7261 tsk_cpus_allowed(task_of(curr)))) {
7264 ratio = curr->avg.load_avg_ratio;
7266 raw_spin_unlock_irqrestore(&rq->lock, flags);
7272 /* now we have a candidate */
7273 raw_spin_lock_irqsave(&target->lock, flags);
7274 if (!target->active_balance && task_rq(p) == target) {
7276 target->push_cpu = this_cpu;
7277 target->migrate_task = p;
7278 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_IDLE_PULL);
7279 hmp_next_up_delay(&p->se, target->push_cpu);
7281 * if the task isn't running move it right away.
7282 * Otherwise setup the active_balance mechanic and let
7283 * the CPU stopper do its job.
7285 if (!task_running(target, p)) {
7286 trace_sched_hmp_migrate_idle_running(p, 0);
7287 hmp_migrate_runnable_task(target);
7289 target->active_balance = 1;
7293 raw_spin_unlock_irqrestore(&target->lock, flags);
7296 /* start timer to keep us awake */
7297 hmp_cpu_keepalive_trigger();
7298 stop_one_cpu_nowait(cpu_of(target),
7299 hmp_active_task_migration_cpu_stop,
7300 target, &target->active_balance_work);
7303 spin_unlock(&hmp_force_migration);
7307 static void hmp_force_up_migration(int this_cpu) { }
7308 #endif /* CONFIG_SCHED_HMP */
7311 * run_rebalance_domains is triggered when needed from the scheduler tick.
7312 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7314 static void run_rebalance_domains(struct softirq_action *h)
7316 int this_cpu = smp_processor_id();
7317 struct rq *this_rq = cpu_rq(this_cpu);
7318 enum cpu_idle_type idle = this_rq->idle_balance ?
7319 CPU_IDLE : CPU_NOT_IDLE;
7321 #ifdef CONFIG_SCHED_HMP
7322 /* shortcut for hmp idle pull wakeups */
7323 if (unlikely(this_rq->wake_for_idle_pull)) {
7324 this_rq->wake_for_idle_pull = 0;
7325 if (hmp_idle_pull(this_cpu)) {
7326 /* break out unless running nohz idle as well */
7327 if (idle != CPU_IDLE)
7333 hmp_force_up_migration(this_cpu);
7335 rebalance_domains(this_cpu, idle);
7338 * If this cpu has a pending nohz_balance_kick, then do the
7339 * balancing on behalf of the other idle cpus whose ticks are
7342 nohz_idle_balance(this_cpu, idle);
7345 static inline int on_null_domain(int cpu)
7347 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
7351 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7353 void trigger_load_balance(struct rq *rq, int cpu)
7355 /* Don't need to rebalance while attached to NULL domain */
7356 if (time_after_eq(jiffies, rq->next_balance) &&
7357 likely(!on_null_domain(cpu)))
7358 raise_softirq(SCHED_SOFTIRQ);
7359 #ifdef CONFIG_NO_HZ_COMMON
7360 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
7361 nohz_balancer_kick(cpu);
7365 static void rq_online_fair(struct rq *rq)
7367 #ifdef CONFIG_SCHED_HMP
7368 hmp_online_cpu(rq->cpu);
7373 static void rq_offline_fair(struct rq *rq)
7375 #ifdef CONFIG_SCHED_HMP
7376 hmp_offline_cpu(rq->cpu);
7380 /* Ensure any throttled groups are reachable by pick_next_task */
7381 unthrottle_offline_cfs_rqs(rq);
7384 #endif /* CONFIG_SMP */
7387 * scheduler tick hitting a task of our scheduling class:
7389 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7391 struct cfs_rq *cfs_rq;
7392 struct sched_entity *se = &curr->se;
7394 for_each_sched_entity(se) {
7395 cfs_rq = cfs_rq_of(se);
7396 entity_tick(cfs_rq, se, queued);
7399 if (sched_feat_numa(NUMA))
7400 task_tick_numa(rq, curr);
7402 update_rq_runnable_avg(rq, 1);
7406 * called on fork with the child task as argument from the parent's context
7407 * - child not yet on the tasklist
7408 * - preemption disabled
7410 static void task_fork_fair(struct task_struct *p)
7412 struct cfs_rq *cfs_rq;
7413 struct sched_entity *se = &p->se, *curr;
7414 int this_cpu = smp_processor_id();
7415 struct rq *rq = this_rq();
7416 unsigned long flags;
7418 raw_spin_lock_irqsave(&rq->lock, flags);
7420 update_rq_clock(rq);
7422 cfs_rq = task_cfs_rq(current);
7423 curr = cfs_rq->curr;
7425 if (unlikely(task_cpu(p) != this_cpu)) {
7427 __set_task_cpu(p, this_cpu);
7431 update_curr(cfs_rq);
7434 se->vruntime = curr->vruntime;
7435 place_entity(cfs_rq, se, 1);
7437 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7439 * Upon rescheduling, sched_class::put_prev_task() will place
7440 * 'current' within the tree based on its new key value.
7442 swap(curr->vruntime, se->vruntime);
7443 resched_task(rq->curr);
7446 se->vruntime -= cfs_rq->min_vruntime;
7448 raw_spin_unlock_irqrestore(&rq->lock, flags);
7452 * Priority of the task has changed. Check to see if we preempt
7456 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7462 * Reschedule if we are currently running on this runqueue and
7463 * our priority decreased, or if we are not currently running on
7464 * this runqueue and our priority is higher than the current's
7466 if (rq->curr == p) {
7467 if (p->prio > oldprio)
7468 resched_task(rq->curr);
7470 check_preempt_curr(rq, p, 0);
7473 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7475 struct sched_entity *se = &p->se;
7476 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7479 * Ensure the task's vruntime is normalized, so that when its
7480 * switched back to the fair class the enqueue_entity(.flags=0) will
7481 * do the right thing.
7483 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7484 * have normalized the vruntime, if it was !on_rq, then only when
7485 * the task is sleeping will it still have non-normalized vruntime.
7487 if (!se->on_rq && p->state != TASK_RUNNING) {
7489 * Fix up our vruntime so that the current sleep doesn't
7490 * cause 'unlimited' sleep bonus.
7492 place_entity(cfs_rq, se, 0);
7493 se->vruntime -= cfs_rq->min_vruntime;
7496 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7498 * Remove our load from contribution when we leave sched_fair
7499 * and ensure we don't carry in an old decay_count if we
7502 if (p->se.avg.decay_count) {
7503 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
7504 __synchronize_entity_decay(&p->se);
7505 subtract_blocked_load_contrib(cfs_rq,
7506 p->se.avg.load_avg_contrib);
7512 * We switched to the sched_fair class.
7514 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7520 * We were most likely switched from sched_rt, so
7521 * kick off the schedule if running, otherwise just see
7522 * if we can still preempt the current task.
7525 resched_task(rq->curr);
7527 check_preempt_curr(rq, p, 0);
7530 /* Account for a task changing its policy or group.
7532 * This routine is mostly called to set cfs_rq->curr field when a task
7533 * migrates between groups/classes.
7535 static void set_curr_task_fair(struct rq *rq)
7537 struct sched_entity *se = &rq->curr->se;
7539 for_each_sched_entity(se) {
7540 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7542 set_next_entity(cfs_rq, se);
7543 /* ensure bandwidth has been allocated on our new cfs_rq */
7544 account_cfs_rq_runtime(cfs_rq, 0);
7548 void init_cfs_rq(struct cfs_rq *cfs_rq)
7550 cfs_rq->tasks_timeline = RB_ROOT;
7551 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7552 #ifndef CONFIG_64BIT
7553 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7555 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7556 atomic64_set(&cfs_rq->decay_counter, 1);
7557 atomic64_set(&cfs_rq->removed_load, 0);
7561 #ifdef CONFIG_FAIR_GROUP_SCHED
7562 static void task_move_group_fair(struct task_struct *p, int on_rq)
7564 struct cfs_rq *cfs_rq;
7566 * If the task was not on the rq at the time of this cgroup movement
7567 * it must have been asleep, sleeping tasks keep their ->vruntime
7568 * absolute on their old rq until wakeup (needed for the fair sleeper
7569 * bonus in place_entity()).
7571 * If it was on the rq, we've just 'preempted' it, which does convert
7572 * ->vruntime to a relative base.
7574 * Make sure both cases convert their relative position when migrating
7575 * to another cgroup's rq. This does somewhat interfere with the
7576 * fair sleeper stuff for the first placement, but who cares.
7579 * When !on_rq, vruntime of the task has usually NOT been normalized.
7580 * But there are some cases where it has already been normalized:
7582 * - Moving a forked child which is waiting for being woken up by
7583 * wake_up_new_task().
7584 * - Moving a task which has been woken up by try_to_wake_up() and
7585 * waiting for actually being woken up by sched_ttwu_pending().
7587 * To prevent boost or penalty in the new cfs_rq caused by delta
7588 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7590 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7594 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7595 set_task_rq(p, task_cpu(p));
7597 cfs_rq = cfs_rq_of(&p->se);
7598 p->se.vruntime += cfs_rq->min_vruntime;
7601 * migrate_task_rq_fair() will have removed our previous
7602 * contribution, but we must synchronize for ongoing future
7605 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7606 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7611 void free_fair_sched_group(struct task_group *tg)
7615 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7617 for_each_possible_cpu(i) {
7619 kfree(tg->cfs_rq[i]);
7628 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7630 struct cfs_rq *cfs_rq;
7631 struct sched_entity *se;
7634 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7637 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7641 tg->shares = NICE_0_LOAD;
7643 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7645 for_each_possible_cpu(i) {
7646 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7647 GFP_KERNEL, cpu_to_node(i));
7651 se = kzalloc_node(sizeof(struct sched_entity),
7652 GFP_KERNEL, cpu_to_node(i));
7656 init_cfs_rq(cfs_rq);
7657 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7668 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7670 struct rq *rq = cpu_rq(cpu);
7671 unsigned long flags;
7674 * Only empty task groups can be destroyed; so we can speculatively
7675 * check on_list without danger of it being re-added.
7677 if (!tg->cfs_rq[cpu]->on_list)
7680 raw_spin_lock_irqsave(&rq->lock, flags);
7681 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7682 raw_spin_unlock_irqrestore(&rq->lock, flags);
7685 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7686 struct sched_entity *se, int cpu,
7687 struct sched_entity *parent)
7689 struct rq *rq = cpu_rq(cpu);
7693 init_cfs_rq_runtime(cfs_rq);
7695 tg->cfs_rq[cpu] = cfs_rq;
7698 /* se could be NULL for root_task_group */
7703 se->cfs_rq = &rq->cfs;
7705 se->cfs_rq = parent->my_q;
7708 update_load_set(&se->load, 0);
7709 se->parent = parent;
7712 static DEFINE_MUTEX(shares_mutex);
7714 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7717 unsigned long flags;
7720 * We can't change the weight of the root cgroup.
7725 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7727 mutex_lock(&shares_mutex);
7728 if (tg->shares == shares)
7731 tg->shares = shares;
7732 for_each_possible_cpu(i) {
7733 struct rq *rq = cpu_rq(i);
7734 struct sched_entity *se;
7737 /* Propagate contribution to hierarchy */
7738 raw_spin_lock_irqsave(&rq->lock, flags);
7739 for_each_sched_entity(se)
7740 update_cfs_shares(group_cfs_rq(se));
7741 raw_spin_unlock_irqrestore(&rq->lock, flags);
7745 mutex_unlock(&shares_mutex);
7748 #else /* CONFIG_FAIR_GROUP_SCHED */
7750 void free_fair_sched_group(struct task_group *tg) { }
7752 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7757 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7759 #endif /* CONFIG_FAIR_GROUP_SCHED */
7762 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7764 struct sched_entity *se = &task->se;
7765 unsigned int rr_interval = 0;
7768 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7771 if (rq->cfs.load.weight)
7772 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7778 * All the scheduling class methods:
7780 const struct sched_class fair_sched_class = {
7781 .next = &idle_sched_class,
7782 .enqueue_task = enqueue_task_fair,
7783 .dequeue_task = dequeue_task_fair,
7784 .yield_task = yield_task_fair,
7785 .yield_to_task = yield_to_task_fair,
7787 .check_preempt_curr = check_preempt_wakeup,
7789 .pick_next_task = pick_next_task_fair,
7790 .put_prev_task = put_prev_task_fair,
7793 .select_task_rq = select_task_rq_fair,
7794 #ifdef CONFIG_FAIR_GROUP_SCHED
7795 .migrate_task_rq = migrate_task_rq_fair,
7797 .rq_online = rq_online_fair,
7798 .rq_offline = rq_offline_fair,
7800 .task_waking = task_waking_fair,
7803 .set_curr_task = set_curr_task_fair,
7804 .task_tick = task_tick_fair,
7805 .task_fork = task_fork_fair,
7807 .prio_changed = prio_changed_fair,
7808 .switched_from = switched_from_fair,
7809 .switched_to = switched_to_fair,
7811 .get_rr_interval = get_rr_interval_fair,
7813 #ifdef CONFIG_FAIR_GROUP_SCHED
7814 .task_move_group = task_move_group_fair,
7818 #ifdef CONFIG_SCHED_DEBUG
7819 void print_cfs_stats(struct seq_file *m, int cpu)
7821 struct cfs_rq *cfs_rq;
7824 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7825 print_cfs_rq(m, cpu, cfs_rq);
7830 __init void init_sched_fair_class(void)
7833 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7835 #ifdef CONFIG_NO_HZ_COMMON
7836 nohz.next_balance = jiffies;
7837 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7838 cpu_notifier(sched_ilb_notifier, 0);
7841 #ifdef CONFIG_SCHED_HMP
7842 hmp_cpu_mask_setup();
7848 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
7849 static u32 cpufreq_calc_scale(u32 min, u32 max, u32 curr)
7851 u32 result = curr / max;
7855 /* Called when the CPU Frequency is changed.
7856 * Once for each CPU.
7858 static int cpufreq_callback(struct notifier_block *nb,
7859 unsigned long val, void *data)
7861 struct cpufreq_freqs *freq = data;
7862 int cpu = freq->cpu;
7863 struct cpufreq_extents *extents;
7865 if (freq->flags & CPUFREQ_CONST_LOOPS)
7868 if (val != CPUFREQ_POSTCHANGE)
7871 /* if dynamic load scale is disabled, set the load scale to 1.0 */
7872 if (!hmp_data.freqinvar_load_scale_enabled) {
7873 freq_scale[cpu].curr_scale = 1024;
7877 extents = &freq_scale[cpu];
7878 if (extents->flags & SCHED_LOAD_FREQINVAR_SINGLEFREQ) {
7879 /* If our governor was recognised as a single-freq governor,
7882 extents->curr_scale = 1024;
7884 extents->curr_scale = cpufreq_calc_scale(extents->min,
7885 extents->max, freq->new);
7891 /* Called when the CPUFreq governor is changed.
7892 * Only called for the CPUs which are actually changed by the
7895 static int cpufreq_policy_callback(struct notifier_block *nb,
7896 unsigned long event, void *data)
7898 struct cpufreq_policy *policy = data;
7899 struct cpufreq_extents *extents;
7900 int cpu, singleFreq = 0;
7901 static const char performance_governor[] = "performance";
7902 static const char powersave_governor[] = "powersave";
7904 if (event == CPUFREQ_START)
7907 if (event != CPUFREQ_INCOMPATIBLE)
7910 /* CPUFreq governors do not accurately report the range of
7911 * CPU Frequencies they will choose from.
7912 * We recognise performance and powersave governors as
7913 * single-frequency only.
7915 if (!strncmp(policy->governor->name, performance_governor,
7916 strlen(performance_governor)) ||
7917 !strncmp(policy->governor->name, powersave_governor,
7918 strlen(powersave_governor)))
7921 /* Make sure that all CPUs impacted by this policy are
7922 * updated since we will only get a notification when the
7923 * user explicitly changes the policy on a CPU.
7925 for_each_cpu(cpu, policy->cpus) {
7926 extents = &freq_scale[cpu];
7927 extents->max = policy->max >> SCHED_FREQSCALE_SHIFT;
7928 extents->min = policy->min >> SCHED_FREQSCALE_SHIFT;
7929 if (!hmp_data.freqinvar_load_scale_enabled) {
7930 extents->curr_scale = 1024;
7931 } else if (singleFreq) {
7932 extents->flags |= SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7933 extents->curr_scale = 1024;
7935 extents->flags &= ~SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7936 extents->curr_scale = cpufreq_calc_scale(extents->min,
7937 extents->max, policy->cur);
7944 static struct notifier_block cpufreq_notifier = {
7945 .notifier_call = cpufreq_callback,
7947 static struct notifier_block cpufreq_policy_notifier = {
7948 .notifier_call = cpufreq_policy_callback,
7951 static int __init register_sched_cpufreq_notifier(void)
7955 /* init safe defaults since there are no policies at registration */
7956 for (ret = 0; ret < CONFIG_NR_CPUS; ret++) {
7958 freq_scale[ret].max = 1024;
7959 freq_scale[ret].min = 1024;
7960 freq_scale[ret].curr_scale = 1024;
7963 pr_info("sched: registering cpufreq notifiers for scale-invariant loads\n");
7964 ret = cpufreq_register_notifier(&cpufreq_policy_notifier,
7965 CPUFREQ_POLICY_NOTIFIER);
7968 ret = cpufreq_register_notifier(&cpufreq_notifier,
7969 CPUFREQ_TRANSITION_NOTIFIER);
7974 core_initcall(register_sched_cpufreq_notifier);
7975 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */