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 */
47 * Targeted preemption latency for CPU-bound tasks:
48 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
50 * NOTE: this latency value is not the same as the concept of
51 * 'timeslice length' - timeslices in CFS are of variable length
52 * and have no persistent notion like in traditional, time-slice
53 * based scheduling concepts.
55 * (to see the precise effective timeslice length of your workload,
56 * run vmstat and monitor the context-switches (cs) field)
58 unsigned int sysctl_sched_latency = 6000000ULL;
59 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
62 * The initial- and re-scaling of tunables is configurable
63 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
66 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
67 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
68 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
70 enum sched_tunable_scaling sysctl_sched_tunable_scaling
71 = SCHED_TUNABLESCALING_LOG;
74 * Minimal preemption granularity for CPU-bound tasks:
75 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
77 unsigned int sysctl_sched_min_granularity = 750000ULL;
78 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
81 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
83 static unsigned int sched_nr_latency = 8;
86 * After fork, child runs first. If set to 0 (default) then
87 * parent will (try to) run first.
89 unsigned int sysctl_sched_child_runs_first __read_mostly;
92 * SCHED_OTHER wake-up granularity.
93 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
95 * This option delays the preemption effects of decoupled workloads
96 * and reduces their over-scheduling. Synchronous workloads will still
97 * have immediate wakeup/sleep latencies.
99 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
100 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
102 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
105 * The exponential sliding window over which load is averaged for shares
109 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
111 #ifdef CONFIG_CFS_BANDWIDTH
113 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
114 * each time a cfs_rq requests quota.
116 * Note: in the case that the slice exceeds the runtime remaining (either due
117 * to consumption or the quota being specified to be smaller than the slice)
118 * we will always only issue the remaining available time.
120 * default: 5 msec, units: microseconds
122 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
126 * Increase the granularity value when there are more CPUs,
127 * because with more CPUs the 'effective latency' as visible
128 * to users decreases. But the relationship is not linear,
129 * so pick a second-best guess by going with the log2 of the
132 * This idea comes from the SD scheduler of Con Kolivas:
134 static int get_update_sysctl_factor(void)
136 unsigned int cpus = min_t(int, num_online_cpus(), 8);
139 switch (sysctl_sched_tunable_scaling) {
140 case SCHED_TUNABLESCALING_NONE:
143 case SCHED_TUNABLESCALING_LINEAR:
146 case SCHED_TUNABLESCALING_LOG:
148 factor = 1 + ilog2(cpus);
155 static void update_sysctl(void)
157 unsigned int factor = get_update_sysctl_factor();
159 #define SET_SYSCTL(name) \
160 (sysctl_##name = (factor) * normalized_sysctl_##name)
161 SET_SYSCTL(sched_min_granularity);
162 SET_SYSCTL(sched_latency);
163 SET_SYSCTL(sched_wakeup_granularity);
167 void sched_init_granularity(void)
172 #if BITS_PER_LONG == 32
173 # define WMULT_CONST (~0UL)
175 # define WMULT_CONST (1UL << 32)
178 #define WMULT_SHIFT 32
181 * Shift right and round:
183 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
186 * delta *= weight / lw
189 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
190 struct load_weight *lw)
195 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
196 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
197 * 2^SCHED_LOAD_RESOLUTION.
199 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
200 tmp = (u64)delta_exec * scale_load_down(weight);
202 tmp = (u64)delta_exec;
204 if (!lw->inv_weight) {
205 unsigned long w = scale_load_down(lw->weight);
207 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
209 else if (unlikely(!w))
210 lw->inv_weight = WMULT_CONST;
212 lw->inv_weight = WMULT_CONST / w;
216 * Check whether we'd overflow the 64-bit multiplication:
218 if (unlikely(tmp > WMULT_CONST))
219 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
222 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
224 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
228 const struct sched_class fair_sched_class;
230 /**************************************************************
231 * CFS operations on generic schedulable entities:
234 #ifdef CONFIG_FAIR_GROUP_SCHED
236 /* cpu runqueue to which this cfs_rq is attached */
237 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
242 /* An entity is a task if it doesn't "own" a runqueue */
243 #define entity_is_task(se) (!se->my_q)
245 static inline struct task_struct *task_of(struct sched_entity *se)
247 #ifdef CONFIG_SCHED_DEBUG
248 WARN_ON_ONCE(!entity_is_task(se));
250 return container_of(se, struct task_struct, se);
253 /* Walk up scheduling entities hierarchy */
254 #define for_each_sched_entity(se) \
255 for (; se; se = se->parent)
257 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
262 /* runqueue on which this entity is (to be) queued */
263 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268 /* runqueue "owned" by this group */
269 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
274 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
277 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
279 if (!cfs_rq->on_list) {
281 * Ensure we either appear before our parent (if already
282 * enqueued) or force our parent to appear after us when it is
283 * enqueued. The fact that we always enqueue bottom-up
284 * reduces this to two cases.
286 if (cfs_rq->tg->parent &&
287 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
288 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
289 &rq_of(cfs_rq)->leaf_cfs_rq_list);
291 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
292 &rq_of(cfs_rq)->leaf_cfs_rq_list);
296 /* We should have no load, but we need to update last_decay. */
297 update_cfs_rq_blocked_load(cfs_rq, 0);
301 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
303 if (cfs_rq->on_list) {
304 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
309 /* Iterate thr' all leaf cfs_rq's on a runqueue */
310 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
311 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
313 /* Do the two (enqueued) entities belong to the same group ? */
315 is_same_group(struct sched_entity *se, struct sched_entity *pse)
317 if (se->cfs_rq == pse->cfs_rq)
323 static inline struct sched_entity *parent_entity(struct sched_entity *se)
328 /* return depth at which a sched entity is present in the hierarchy */
329 static inline int depth_se(struct sched_entity *se)
333 for_each_sched_entity(se)
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 int se_depth, pse_depth;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth = depth_se(*se);
353 pse_depth = depth_se(*pse);
355 while (se_depth > pse_depth) {
357 *se = parent_entity(*se);
360 while (pse_depth > se_depth) {
362 *pse = parent_entity(*pse);
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct *task_of(struct sched_entity *se)
375 return container_of(se, struct task_struct, se);
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 return container_of(cfs_rq, struct rq, cfs);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 return &task_rq(p)->cfs;
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
419 is_same_group(struct sched_entity *se, struct sched_entity *pse)
424 static inline struct sched_entity *parent_entity(struct sched_entity *se)
430 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
434 #endif /* CONFIG_FAIR_GROUP_SCHED */
436 static __always_inline
437 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
439 /**************************************************************
440 * Scheduling class tree data structure manipulation methods:
443 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - max_vruntime);
447 max_vruntime = vruntime;
452 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - min_vruntime);
456 min_vruntime = vruntime;
461 static inline int entity_before(struct sched_entity *a,
462 struct sched_entity *b)
464 return (s64)(a->vruntime - b->vruntime) < 0;
467 static void update_min_vruntime(struct cfs_rq *cfs_rq)
469 u64 vruntime = cfs_rq->min_vruntime;
472 vruntime = cfs_rq->curr->vruntime;
474 if (cfs_rq->rb_leftmost) {
475 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
480 vruntime = se->vruntime;
482 vruntime = min_vruntime(vruntime, se->vruntime);
485 /* ensure we never gain time by being placed backwards. */
486 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
489 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
494 * Enqueue an entity into the rb-tree:
496 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
498 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
499 struct rb_node *parent = NULL;
500 struct sched_entity *entry;
504 * Find the right place in the rbtree:
508 entry = rb_entry(parent, struct sched_entity, run_node);
510 * We dont care about collisions. Nodes with
511 * the same key stay together.
513 if (entity_before(se, entry)) {
514 link = &parent->rb_left;
516 link = &parent->rb_right;
522 * Maintain a cache of leftmost tree entries (it is frequently
526 cfs_rq->rb_leftmost = &se->run_node;
528 rb_link_node(&se->run_node, parent, link);
529 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
532 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
534 if (cfs_rq->rb_leftmost == &se->run_node) {
535 struct rb_node *next_node;
537 next_node = rb_next(&se->run_node);
538 cfs_rq->rb_leftmost = next_node;
541 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
544 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
546 struct rb_node *left = cfs_rq->rb_leftmost;
551 return rb_entry(left, struct sched_entity, run_node);
554 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
556 struct rb_node *next = rb_next(&se->run_node);
561 return rb_entry(next, struct sched_entity, run_node);
564 #ifdef CONFIG_SCHED_DEBUG
565 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
567 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
572 return rb_entry(last, struct sched_entity, run_node);
575 /**************************************************************
576 * Scheduling class statistics methods:
579 int sched_proc_update_handler(struct ctl_table *table, int write,
580 void __user *buffer, size_t *lenp,
583 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
584 int factor = get_update_sysctl_factor();
589 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
590 sysctl_sched_min_granularity);
592 #define WRT_SYSCTL(name) \
593 (normalized_sysctl_##name = sysctl_##name / (factor))
594 WRT_SYSCTL(sched_min_granularity);
595 WRT_SYSCTL(sched_latency);
596 WRT_SYSCTL(sched_wakeup_granularity);
606 static inline unsigned long
607 calc_delta_fair(unsigned long delta, struct sched_entity *se)
609 if (unlikely(se->load.weight != NICE_0_LOAD))
610 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
616 * The idea is to set a period in which each task runs once.
618 * When there are too many tasks (sched_nr_latency) we have to stretch
619 * this period because otherwise the slices get too small.
621 * p = (nr <= nl) ? l : l*nr/nl
623 static u64 __sched_period(unsigned long nr_running)
625 u64 period = sysctl_sched_latency;
626 unsigned long nr_latency = sched_nr_latency;
628 if (unlikely(nr_running > nr_latency)) {
629 period = sysctl_sched_min_granularity;
630 period *= nr_running;
637 * We calculate the wall-time slice from the period by taking a part
638 * proportional to the weight.
642 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
644 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
646 for_each_sched_entity(se) {
647 struct load_weight *load;
648 struct load_weight lw;
650 cfs_rq = cfs_rq_of(se);
651 load = &cfs_rq->load;
653 if (unlikely(!se->on_rq)) {
656 update_load_add(&lw, se->load.weight);
659 slice = calc_delta_mine(slice, se->load.weight, load);
665 * We calculate the vruntime slice of a to-be-inserted task.
669 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
671 return calc_delta_fair(sched_slice(cfs_rq, se), se);
675 * Update the current task's runtime statistics. Skip current tasks that
676 * are not in our scheduling class.
679 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
680 unsigned long delta_exec)
682 unsigned long delta_exec_weighted;
684 schedstat_set(curr->statistics.exec_max,
685 max((u64)delta_exec, curr->statistics.exec_max));
687 curr->sum_exec_runtime += delta_exec;
688 schedstat_add(cfs_rq, exec_clock, delta_exec);
689 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
691 curr->vruntime += delta_exec_weighted;
692 update_min_vruntime(cfs_rq);
695 static void update_curr(struct cfs_rq *cfs_rq)
697 struct sched_entity *curr = cfs_rq->curr;
698 u64 now = rq_of(cfs_rq)->clock_task;
699 unsigned long delta_exec;
705 * Get the amount of time the current task was running
706 * since the last time we changed load (this cannot
707 * overflow on 32 bits):
709 delta_exec = (unsigned long)(now - curr->exec_start);
713 __update_curr(cfs_rq, curr, delta_exec);
714 curr->exec_start = now;
716 if (entity_is_task(curr)) {
717 struct task_struct *curtask = task_of(curr);
719 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720 cpuacct_charge(curtask, delta_exec);
721 account_group_exec_runtime(curtask, delta_exec);
724 account_cfs_rq_runtime(cfs_rq, delta_exec);
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se != cfs_rq->curr)
743 update_stats_wait_start(cfs_rq, se);
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
749 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750 rq_of(cfs_rq)->clock - se->statistics.wait_start));
751 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753 rq_of(cfs_rq)->clock - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se)) {
756 trace_sched_stat_wait(task_of(se),
757 rq_of(cfs_rq)->clock - se->statistics.wait_start);
760 schedstat_set(se->statistics.wait_start, 0);
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Mark the end of the wait period if dequeueing a
770 if (se != cfs_rq->curr)
771 update_stats_wait_end(cfs_rq, se);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
781 * We are starting a new run period:
783 se->exec_start = rq_of(cfs_rq)->clock_task;
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * numa task sample period in ms
794 unsigned int sysctl_numa_balancing_scan_period_min = 100;
795 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
796 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
798 /* Portion of address space to scan in MB */
799 unsigned int sysctl_numa_balancing_scan_size = 256;
801 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
802 unsigned int sysctl_numa_balancing_scan_delay = 1000;
804 static void task_numa_placement(struct task_struct *p)
808 if (!p->mm) /* for example, ksmd faulting in a user's mm */
810 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
811 if (p->numa_scan_seq == seq)
813 p->numa_scan_seq = seq;
815 /* FIXME: Scheduling placement policy hints go here */
819 * Got a PROT_NONE fault for a page on @node.
821 void task_numa_fault(int node, int pages, bool migrated)
823 struct task_struct *p = current;
825 if (!sched_feat_numa(NUMA))
828 /* FIXME: Allocate task-specific structure for placement policy here */
831 * If pages are properly placed (did not migrate) then scan slower.
832 * This is reset periodically in case of phase changes
835 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
836 p->numa_scan_period + jiffies_to_msecs(10));
838 task_numa_placement(p);
841 static void reset_ptenuma_scan(struct task_struct *p)
843 ACCESS_ONCE(p->mm->numa_scan_seq)++;
844 p->mm->numa_scan_offset = 0;
848 * The expensive part of numa migration is done from task_work context.
849 * Triggered from task_tick_numa().
851 void task_numa_work(struct callback_head *work)
853 unsigned long migrate, next_scan, now = jiffies;
854 struct task_struct *p = current;
855 struct mm_struct *mm = p->mm;
856 struct vm_area_struct *vma;
857 unsigned long start, end;
860 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
862 work->next = work; /* protect against double add */
864 * Who cares about NUMA placement when they're dying.
866 * NOTE: make sure not to dereference p->mm before this check,
867 * exit_task_work() happens _after_ exit_mm() so we could be called
868 * without p->mm even though we still had it when we enqueued this
871 if (p->flags & PF_EXITING)
875 * We do not care about task placement until a task runs on a node
876 * other than the first one used by the address space. This is
877 * largely because migrations are driven by what CPU the task
878 * is running on. If it's never scheduled on another node, it'll
879 * not migrate so why bother trapping the fault.
881 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
882 mm->first_nid = numa_node_id();
883 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
884 /* Are we running on a new node yet? */
885 if (numa_node_id() == mm->first_nid &&
886 !sched_feat_numa(NUMA_FORCE))
889 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
893 * Reset the scan period if enough time has gone by. Objective is that
894 * scanning will be reduced if pages are properly placed. As tasks
895 * can enter different phases this needs to be re-examined. Lacking
896 * proper tracking of reference behaviour, this blunt hammer is used.
898 migrate = mm->numa_next_reset;
899 if (time_after(now, migrate)) {
900 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
901 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
902 xchg(&mm->numa_next_reset, next_scan);
906 * Enforce maximal scan/migration frequency..
908 migrate = mm->numa_next_scan;
909 if (time_before(now, migrate))
912 if (p->numa_scan_period == 0)
913 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
915 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
916 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
920 * Do not set pte_numa if the current running node is rate-limited.
921 * This loses statistics on the fault but if we are unwilling to
922 * migrate to this node, it is less likely we can do useful work
924 if (migrate_ratelimited(numa_node_id()))
927 start = mm->numa_scan_offset;
928 pages = sysctl_numa_balancing_scan_size;
929 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
933 down_read(&mm->mmap_sem);
934 vma = find_vma(mm, start);
936 reset_ptenuma_scan(p);
940 for (; vma; vma = vma->vm_next) {
941 if (!vma_migratable(vma))
944 /* Skip small VMAs. They are not likely to be of relevance */
945 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
949 * Skip inaccessible VMAs to avoid any confusion between
950 * PROT_NONE and NUMA hinting ptes
952 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
956 start = max(start, vma->vm_start);
957 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
958 end = min(end, vma->vm_end);
959 pages -= change_prot_numa(vma, start, end);
964 } while (end != vma->vm_end);
969 * It is possible to reach the end of the VMA list but the last few VMAs are
970 * not guaranteed to the vma_migratable. If they are not, we would find the
971 * !migratable VMA on the next scan but not reset the scanner to the start
975 mm->numa_scan_offset = start;
977 reset_ptenuma_scan(p);
978 up_read(&mm->mmap_sem);
982 * Drive the periodic memory faults..
984 void task_tick_numa(struct rq *rq, struct task_struct *curr)
986 struct callback_head *work = &curr->numa_work;
990 * We don't care about NUMA placement if we don't have memory.
992 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
996 * Using runtime rather than walltime has the dual advantage that
997 * we (mostly) drive the selection from busy threads and that the
998 * task needs to have done some actual work before we bother with
1001 now = curr->se.sum_exec_runtime;
1002 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1004 if (now - curr->node_stamp > period) {
1005 if (!curr->node_stamp)
1006 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1007 curr->node_stamp = now;
1009 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1010 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1011 task_work_add(curr, work, true);
1016 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1019 #endif /* CONFIG_NUMA_BALANCING */
1022 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1024 update_load_add(&cfs_rq->load, se->load.weight);
1025 if (!parent_entity(se))
1026 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1028 if (entity_is_task(se))
1029 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1031 cfs_rq->nr_running++;
1035 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1037 update_load_sub(&cfs_rq->load, se->load.weight);
1038 if (!parent_entity(se))
1039 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1040 if (entity_is_task(se))
1041 list_del_init(&se->group_node);
1042 cfs_rq->nr_running--;
1045 #ifdef CONFIG_FAIR_GROUP_SCHED
1047 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1052 * Use this CPU's actual weight instead of the last load_contribution
1053 * to gain a more accurate current total weight. See
1054 * update_cfs_rq_load_contribution().
1056 tg_weight = atomic64_read(&tg->load_avg);
1057 tg_weight -= cfs_rq->tg_load_contrib;
1058 tg_weight += cfs_rq->load.weight;
1063 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1065 long tg_weight, load, shares;
1067 tg_weight = calc_tg_weight(tg, cfs_rq);
1068 load = cfs_rq->load.weight;
1070 shares = (tg->shares * load);
1072 shares /= tg_weight;
1074 if (shares < MIN_SHARES)
1075 shares = MIN_SHARES;
1076 if (shares > tg->shares)
1077 shares = tg->shares;
1081 # else /* CONFIG_SMP */
1082 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1086 # endif /* CONFIG_SMP */
1087 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1088 unsigned long weight)
1091 /* commit outstanding execution time */
1092 if (cfs_rq->curr == se)
1093 update_curr(cfs_rq);
1094 account_entity_dequeue(cfs_rq, se);
1097 update_load_set(&se->load, weight);
1100 account_entity_enqueue(cfs_rq, se);
1103 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1105 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1107 struct task_group *tg;
1108 struct sched_entity *se;
1112 se = tg->se[cpu_of(rq_of(cfs_rq))];
1113 if (!se || throttled_hierarchy(cfs_rq))
1116 if (likely(se->load.weight == tg->shares))
1119 shares = calc_cfs_shares(cfs_rq, tg);
1121 reweight_entity(cfs_rq_of(se), se, shares);
1123 #else /* CONFIG_FAIR_GROUP_SCHED */
1124 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1127 #endif /* CONFIG_FAIR_GROUP_SCHED */
1129 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1130 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1132 * We choose a half-life close to 1 scheduling period.
1133 * Note: The tables below are dependent on this value.
1135 #define LOAD_AVG_PERIOD 32
1136 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1137 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1139 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1140 static const u32 runnable_avg_yN_inv[] = {
1141 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1142 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1143 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1144 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1145 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1146 0x85aac367, 0x82cd8698,
1150 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1151 * over-estimates when re-combining.
1153 static const u32 runnable_avg_yN_sum[] = {
1154 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1155 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1156 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1161 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1163 static __always_inline u64 decay_load(u64 val, u64 n)
1165 unsigned int local_n;
1169 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1172 /* after bounds checking we can collapse to 32-bit */
1176 * As y^PERIOD = 1/2, we can combine
1177 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1178 * With a look-up table which covers k^n (n<PERIOD)
1180 * To achieve constant time decay_load.
1182 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1183 val >>= local_n / LOAD_AVG_PERIOD;
1184 local_n %= LOAD_AVG_PERIOD;
1187 val *= runnable_avg_yN_inv[local_n];
1188 /* We don't use SRR here since we always want to round down. */
1193 * For updates fully spanning n periods, the contribution to runnable
1194 * average will be: \Sum 1024*y^n
1196 * We can compute this reasonably efficiently by combining:
1197 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1199 static u32 __compute_runnable_contrib(u64 n)
1203 if (likely(n <= LOAD_AVG_PERIOD))
1204 return runnable_avg_yN_sum[n];
1205 else if (unlikely(n >= LOAD_AVG_MAX_N))
1206 return LOAD_AVG_MAX;
1208 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1210 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1211 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1213 n -= LOAD_AVG_PERIOD;
1214 } while (n > LOAD_AVG_PERIOD);
1216 contrib = decay_load(contrib, n);
1217 return contrib + runnable_avg_yN_sum[n];
1220 #ifdef CONFIG_SCHED_HMP
1221 #define HMP_VARIABLE_SCALE_SHIFT 16ULL
1222 struct hmp_global_attr {
1223 struct attribute attr;
1224 ssize_t (*show)(struct kobject *kobj,
1225 struct attribute *attr, char *buf);
1226 ssize_t (*store)(struct kobject *a, struct attribute *b,
1227 const char *c, size_t count);
1229 int (*to_sysfs)(int);
1230 int (*from_sysfs)(int);
1231 ssize_t (*to_sysfs_text)(char *buf, int buf_size);
1234 #define HMP_DATA_SYSFS_MAX 8
1236 struct hmp_data_struct {
1237 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1238 int freqinvar_load_scale_enabled;
1240 int multiplier; /* used to scale the time delta */
1241 struct attribute_group attr_group;
1242 struct attribute *attributes[HMP_DATA_SYSFS_MAX + 1];
1243 struct hmp_global_attr attr[HMP_DATA_SYSFS_MAX];
1246 static u64 hmp_variable_scale_convert(u64 delta);
1247 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1248 /* Frequency-Invariant Load Modification:
1249 * Loads are calculated as in PJT's patch however we also scale the current
1250 * contribution in line with the frequency of the CPU that the task was
1252 * In this version, we use a simple linear scale derived from the maximum
1253 * frequency reported by CPUFreq. As an example:
1255 * Consider that we ran a task for 100% of the previous interval.
1257 * Our CPU was under asynchronous frequency control through one of the
1258 * CPUFreq governors.
1260 * The CPUFreq governor reports that it is able to scale the CPU between
1263 * During the period, the CPU was running at 1GHz.
1265 * In this case, our load contribution for that period is calculated as
1266 * 1 * (number_of_active_microseconds)
1268 * This results in our task being able to accumulate maximum load as normal.
1271 * Consider now that our CPU was executing at 500MHz.
1273 * We now scale the load contribution such that it is calculated as
1274 * 0.5 * (number_of_active_microseconds)
1276 * Our task can only record 50% maximum load during this period.
1278 * This represents the task consuming 50% of the CPU's *possible* compute
1279 * capacity. However the task did consume 100% of the CPU's *available*
1280 * compute capacity which is the value seen by the CPUFreq governor and
1281 * user-side CPU Utilization tools.
1283 * Restricting tracked load to be scaled by the CPU's frequency accurately
1284 * represents the consumption of possible compute capacity and allows the
1285 * HMP migration's simple threshold migration strategy to interact more
1286 * predictably with CPUFreq's asynchronous compute capacity changes.
1288 #define SCHED_FREQSCALE_SHIFT 10
1289 struct cpufreq_extents {
1295 /* Flag set when the governor in use only allows one frequency.
1298 #define SCHED_LOAD_FREQINVAR_SINGLEFREQ 0x01
1300 static struct cpufreq_extents freq_scale[CONFIG_NR_CPUS];
1301 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1302 #endif /* CONFIG_SCHED_HMP */
1304 /* We can represent the historical contribution to runnable average as the
1305 * coefficients of a geometric series. To do this we sub-divide our runnable
1306 * history into segments of approximately 1ms (1024us); label the segment that
1307 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1309 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1311 * (now) (~1ms ago) (~2ms ago)
1313 * Let u_i denote the fraction of p_i that the entity was runnable.
1315 * We then designate the fractions u_i as our co-efficients, yielding the
1316 * following representation of historical load:
1317 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1319 * We choose y based on the with of a reasonably scheduling period, fixing:
1322 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1323 * approximately half as much as the contribution to load within the last ms
1326 * When a period "rolls over" and we have new u_0`, multiplying the previous
1327 * sum again by y is sufficient to update:
1328 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1329 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1331 static __always_inline int __update_entity_runnable_avg(u64 now,
1332 struct sched_avg *sa,
1338 u32 runnable_contrib;
1339 int delta_w, decayed = 0;
1340 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1342 u32 scaled_runnable_contrib;
1344 u32 curr_scale = 1024;
1345 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1347 delta = now - sa->last_runnable_update;
1348 #ifdef CONFIG_SCHED_HMP
1349 delta = hmp_variable_scale_convert(delta);
1352 * This should only happen when time goes backwards, which it
1353 * unfortunately does during sched clock init when we swap over to TSC.
1355 if ((s64)delta < 0) {
1356 sa->last_runnable_update = now;
1361 * Use 1024ns as the unit of measurement since it's a reasonable
1362 * approximation of 1us and fast to compute.
1367 sa->last_runnable_update = now;
1369 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1370 /* retrieve scale factor for load */
1371 if (hmp_data.freqinvar_load_scale_enabled)
1372 curr_scale = freq_scale[cpu].curr_scale;
1373 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1375 /* delta_w is the amount already accumulated against our next period */
1376 delta_w = sa->runnable_avg_period % 1024;
1377 if (delta + delta_w >= 1024) {
1378 /* period roll-over */
1382 * Now that we know we're crossing a period boundary, figure
1383 * out how much from delta we need to complete the current
1384 * period and accrue it.
1386 delta_w = 1024 - delta_w;
1387 /* scale runnable time if necessary */
1388 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1389 scaled_delta_w = (delta_w * curr_scale)
1390 >> SCHED_FREQSCALE_SHIFT;
1392 sa->runnable_avg_sum += scaled_delta_w;
1394 sa->usage_avg_sum += scaled_delta_w;
1397 sa->runnable_avg_sum += delta_w;
1399 sa->usage_avg_sum += delta_w;
1400 #endif /* #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1401 sa->runnable_avg_period += delta_w;
1405 /* Figure out how many additional periods this update spans */
1406 periods = delta / 1024;
1408 /* decay the load we have accumulated so far */
1409 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1411 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1413 sa->usage_avg_sum = decay_load(sa->usage_avg_sum, periods + 1);
1414 /* add the contribution from this period */
1415 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1416 runnable_contrib = __compute_runnable_contrib(periods);
1417 /* Apply load scaling if necessary.
1418 * Note that multiplying the whole series is same as
1419 * multiplying all terms
1421 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1422 scaled_runnable_contrib = (runnable_contrib * curr_scale)
1423 >> SCHED_FREQSCALE_SHIFT;
1425 sa->runnable_avg_sum += scaled_runnable_contrib;
1427 sa->usage_avg_sum += scaled_runnable_contrib;
1430 sa->runnable_avg_sum += runnable_contrib;
1432 sa->usage_avg_sum += runnable_contrib;
1433 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1434 sa->runnable_avg_period += runnable_contrib;
1437 /* Remainder of delta accrued against u_0` */
1438 /* scale if necessary */
1439 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1440 scaled_delta = ((delta * curr_scale) >> SCHED_FREQSCALE_SHIFT);
1442 sa->runnable_avg_sum += scaled_delta;
1444 sa->usage_avg_sum += scaled_delta;
1447 sa->runnable_avg_sum += delta;
1449 sa->usage_avg_sum += delta;
1450 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1451 sa->runnable_avg_period += delta;
1456 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1457 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1460 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1462 decays -= se->avg.decay_count;
1464 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1465 se->avg.decay_count = 0;
1469 #ifdef CONFIG_FAIR_GROUP_SCHED
1470 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1473 struct task_group *tg = cfs_rq->tg;
1476 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1477 tg_contrib -= cfs_rq->tg_load_contrib;
1479 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1480 atomic64_add(tg_contrib, &tg->load_avg);
1481 cfs_rq->tg_load_contrib += tg_contrib;
1486 * Aggregate cfs_rq runnable averages into an equivalent task_group
1487 * representation for computing load contributions.
1489 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1490 struct cfs_rq *cfs_rq)
1492 struct task_group *tg = cfs_rq->tg;
1493 long contrib, usage_contrib;
1495 /* The fraction of a cpu used by this cfs_rq */
1496 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1497 sa->runnable_avg_period + 1);
1498 contrib -= cfs_rq->tg_runnable_contrib;
1500 usage_contrib = div_u64(sa->usage_avg_sum << NICE_0_SHIFT,
1501 sa->runnable_avg_period + 1);
1502 usage_contrib -= cfs_rq->tg_usage_contrib;
1505 * contrib/usage at this point represent deltas, only update if they
1508 if ((abs(contrib) > cfs_rq->tg_runnable_contrib / 64) ||
1509 (abs(usage_contrib) > cfs_rq->tg_usage_contrib / 64)) {
1510 atomic_add(contrib, &tg->runnable_avg);
1511 cfs_rq->tg_runnable_contrib += contrib;
1513 atomic_add(usage_contrib, &tg->usage_avg);
1514 cfs_rq->tg_usage_contrib += usage_contrib;
1518 static inline void __update_group_entity_contrib(struct sched_entity *se)
1520 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1521 struct task_group *tg = cfs_rq->tg;
1526 contrib = cfs_rq->tg_load_contrib * tg->shares;
1527 se->avg.load_avg_contrib = div64_u64(contrib,
1528 atomic64_read(&tg->load_avg) + 1);
1531 * For group entities we need to compute a correction term in the case
1532 * that they are consuming <1 cpu so that we would contribute the same
1533 * load as a task of equal weight.
1535 * Explicitly co-ordinating this measurement would be expensive, but
1536 * fortunately the sum of each cpus contribution forms a usable
1537 * lower-bound on the true value.
1539 * Consider the aggregate of 2 contributions. Either they are disjoint
1540 * (and the sum represents true value) or they are disjoint and we are
1541 * understating by the aggregate of their overlap.
1543 * Extending this to N cpus, for a given overlap, the maximum amount we
1544 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1545 * cpus that overlap for this interval and w_i is the interval width.
1547 * On a small machine; the first term is well-bounded which bounds the
1548 * total error since w_i is a subset of the period. Whereas on a
1549 * larger machine, while this first term can be larger, if w_i is the
1550 * of consequential size guaranteed to see n_i*w_i quickly converge to
1551 * our upper bound of 1-cpu.
1553 runnable_avg = atomic_read(&tg->runnable_avg);
1554 if (runnable_avg < NICE_0_LOAD) {
1555 se->avg.load_avg_contrib *= runnable_avg;
1556 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1560 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1561 int force_update) {}
1562 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1563 struct cfs_rq *cfs_rq) {}
1564 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1567 static inline void __update_task_entity_contrib(struct sched_entity *se)
1571 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1572 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1573 contrib /= (se->avg.runnable_avg_period + 1);
1574 se->avg.load_avg_contrib = scale_load(contrib);
1575 trace_sched_task_load_contrib(task_of(se), se->avg.load_avg_contrib);
1576 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1577 contrib /= (se->avg.runnable_avg_period + 1);
1578 se->avg.load_avg_ratio = scale_load(contrib);
1579 trace_sched_task_runnable_ratio(task_of(se), se->avg.load_avg_ratio);
1582 /* Compute the current contribution to load_avg by se, return any delta */
1583 static long __update_entity_load_avg_contrib(struct sched_entity *se, long *ratio)
1585 long old_contrib = se->avg.load_avg_contrib;
1586 long old_ratio = se->avg.load_avg_ratio;
1588 if (entity_is_task(se)) {
1589 __update_task_entity_contrib(se);
1591 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1592 __update_group_entity_contrib(se);
1596 *ratio = se->avg.load_avg_ratio - old_ratio;
1597 return se->avg.load_avg_contrib - old_contrib;
1600 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1603 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1604 cfs_rq->blocked_load_avg -= load_contrib;
1606 cfs_rq->blocked_load_avg = 0;
1609 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1611 /* Update a sched_entity's runnable average */
1612 static inline void update_entity_load_avg(struct sched_entity *se,
1615 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1616 long contrib_delta, ratio_delta;
1618 int cpu = -1; /* not used in normal case */
1620 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1621 cpu = cfs_rq->rq->cpu;
1624 * For a group entity we need to use their owned cfs_rq_clock_task() in
1625 * case they are the parent of a throttled hierarchy.
1627 if (entity_is_task(se))
1628 now = cfs_rq_clock_task(cfs_rq);
1630 now = cfs_rq_clock_task(group_cfs_rq(se));
1632 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1633 cfs_rq->curr == se, cpu))
1636 contrib_delta = __update_entity_load_avg_contrib(se, &ratio_delta);
1642 cfs_rq->runnable_load_avg += contrib_delta;
1643 rq_of(cfs_rq)->avg.load_avg_ratio += ratio_delta;
1645 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1650 * Decay the load contributed by all blocked children and account this so that
1651 * their contribution may appropriately discounted when they wake up.
1653 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1655 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1658 decays = now - cfs_rq->last_decay;
1659 if (!decays && !force_update)
1662 if (atomic64_read(&cfs_rq->removed_load)) {
1663 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1664 subtract_blocked_load_contrib(cfs_rq, removed_load);
1668 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1670 atomic64_add(decays, &cfs_rq->decay_counter);
1671 cfs_rq->last_decay = now;
1674 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1677 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1679 int cpu = -1; /* not used in normal case */
1681 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1684 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1686 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1687 trace_sched_rq_runnable_ratio(cpu_of(rq), rq->avg.load_avg_ratio);
1688 trace_sched_rq_runnable_load(cpu_of(rq), rq->cfs.runnable_load_avg);
1689 trace_sched_rq_nr_running(cpu_of(rq), rq->nr_running, rq->nr_iowait.counter);
1692 /* Add the load generated by se into cfs_rq's child load-average */
1693 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1694 struct sched_entity *se,
1698 * We track migrations using entity decay_count <= 0, on a wake-up
1699 * migration we use a negative decay count to track the remote decays
1700 * accumulated while sleeping.
1702 if (unlikely(se->avg.decay_count <= 0)) {
1703 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1704 if (se->avg.decay_count) {
1706 * In a wake-up migration we have to approximate the
1707 * time sleeping. This is because we can't synchronize
1708 * clock_task between the two cpus, and it is not
1709 * guaranteed to be read-safe. Instead, we can
1710 * approximate this using our carried decays, which are
1711 * explicitly atomically readable.
1713 se->avg.last_runnable_update -= (-se->avg.decay_count)
1715 update_entity_load_avg(se, 0);
1716 /* Indicate that we're now synchronized and on-rq */
1717 se->avg.decay_count = 0;
1721 __synchronize_entity_decay(se);
1724 /* migrated tasks did not contribute to our blocked load */
1726 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1727 update_entity_load_avg(se, 0);
1730 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1731 rq_of(cfs_rq)->avg.load_avg_ratio += se->avg.load_avg_ratio;
1733 /* we force update consideration on load-balancer moves */
1734 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1738 * Remove se's load from this cfs_rq child load-average, if the entity is
1739 * transitioning to a blocked state we track its projected decay using
1742 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1743 struct sched_entity *se,
1746 update_entity_load_avg(se, 1);
1747 /* we force update consideration on load-balancer moves */
1748 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1750 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1751 rq_of(cfs_rq)->avg.load_avg_ratio -= se->avg.load_avg_ratio;
1754 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1755 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1756 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1760 * Update the rq's load with the elapsed running time before entering
1761 * idle. if the last scheduled task is not a CFS task, idle_enter will
1762 * be the only way to update the runnable statistic.
1764 void idle_enter_fair(struct rq *this_rq)
1766 update_rq_runnable_avg(this_rq, 1);
1770 * Update the rq's load with the elapsed idle time before a task is
1771 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1772 * be the only way to update the runnable statistic.
1774 void idle_exit_fair(struct rq *this_rq)
1776 update_rq_runnable_avg(this_rq, 0);
1780 static inline void update_entity_load_avg(struct sched_entity *se,
1781 int update_cfs_rq) {}
1782 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1783 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1784 struct sched_entity *se,
1786 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1787 struct sched_entity *se,
1789 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1790 int force_update) {}
1793 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1795 #ifdef CONFIG_SCHEDSTATS
1796 struct task_struct *tsk = NULL;
1798 if (entity_is_task(se))
1801 if (se->statistics.sleep_start) {
1802 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1807 if (unlikely(delta > se->statistics.sleep_max))
1808 se->statistics.sleep_max = delta;
1810 se->statistics.sleep_start = 0;
1811 se->statistics.sum_sleep_runtime += delta;
1814 account_scheduler_latency(tsk, delta >> 10, 1);
1815 trace_sched_stat_sleep(tsk, delta);
1818 if (se->statistics.block_start) {
1819 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1824 if (unlikely(delta > se->statistics.block_max))
1825 se->statistics.block_max = delta;
1827 se->statistics.block_start = 0;
1828 se->statistics.sum_sleep_runtime += delta;
1831 if (tsk->in_iowait) {
1832 se->statistics.iowait_sum += delta;
1833 se->statistics.iowait_count++;
1834 trace_sched_stat_iowait(tsk, delta);
1837 trace_sched_stat_blocked(tsk, delta);
1840 * Blocking time is in units of nanosecs, so shift by
1841 * 20 to get a milliseconds-range estimation of the
1842 * amount of time that the task spent sleeping:
1844 if (unlikely(prof_on == SLEEP_PROFILING)) {
1845 profile_hits(SLEEP_PROFILING,
1846 (void *)get_wchan(tsk),
1849 account_scheduler_latency(tsk, delta >> 10, 0);
1855 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1857 #ifdef CONFIG_SCHED_DEBUG
1858 s64 d = se->vruntime - cfs_rq->min_vruntime;
1863 if (d > 3*sysctl_sched_latency)
1864 schedstat_inc(cfs_rq, nr_spread_over);
1869 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1871 u64 vruntime = cfs_rq->min_vruntime;
1874 * The 'current' period is already promised to the current tasks,
1875 * however the extra weight of the new task will slow them down a
1876 * little, place the new task so that it fits in the slot that
1877 * stays open at the end.
1879 if (initial && sched_feat(START_DEBIT))
1880 vruntime += sched_vslice(cfs_rq, se);
1882 /* sleeps up to a single latency don't count. */
1884 unsigned long thresh = sysctl_sched_latency;
1887 * Halve their sleep time's effect, to allow
1888 * for a gentler effect of sleepers:
1890 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1896 /* ensure we never gain time by being placed backwards. */
1897 se->vruntime = max_vruntime(se->vruntime, vruntime);
1900 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1903 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1906 * Update the normalized vruntime before updating min_vruntime
1907 * through callig update_curr().
1909 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1910 se->vruntime += cfs_rq->min_vruntime;
1913 * Update run-time statistics of the 'current'.
1915 update_curr(cfs_rq);
1916 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1917 account_entity_enqueue(cfs_rq, se);
1918 update_cfs_shares(cfs_rq);
1920 if (flags & ENQUEUE_WAKEUP) {
1921 place_entity(cfs_rq, se, 0);
1922 enqueue_sleeper(cfs_rq, se);
1925 update_stats_enqueue(cfs_rq, se);
1926 check_spread(cfs_rq, se);
1927 if (se != cfs_rq->curr)
1928 __enqueue_entity(cfs_rq, se);
1931 if (cfs_rq->nr_running == 1) {
1932 list_add_leaf_cfs_rq(cfs_rq);
1933 check_enqueue_throttle(cfs_rq);
1937 static void __clear_buddies_last(struct sched_entity *se)
1939 for_each_sched_entity(se) {
1940 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1941 if (cfs_rq->last == se)
1942 cfs_rq->last = NULL;
1948 static void __clear_buddies_next(struct sched_entity *se)
1950 for_each_sched_entity(se) {
1951 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1952 if (cfs_rq->next == se)
1953 cfs_rq->next = NULL;
1959 static void __clear_buddies_skip(struct sched_entity *se)
1961 for_each_sched_entity(se) {
1962 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1963 if (cfs_rq->skip == se)
1964 cfs_rq->skip = NULL;
1970 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1972 if (cfs_rq->last == se)
1973 __clear_buddies_last(se);
1975 if (cfs_rq->next == se)
1976 __clear_buddies_next(se);
1978 if (cfs_rq->skip == se)
1979 __clear_buddies_skip(se);
1982 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1985 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1988 * Update run-time statistics of the 'current'.
1990 update_curr(cfs_rq);
1991 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1993 update_stats_dequeue(cfs_rq, se);
1994 if (flags & DEQUEUE_SLEEP) {
1995 #ifdef CONFIG_SCHEDSTATS
1996 if (entity_is_task(se)) {
1997 struct task_struct *tsk = task_of(se);
1999 if (tsk->state & TASK_INTERRUPTIBLE)
2000 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
2001 if (tsk->state & TASK_UNINTERRUPTIBLE)
2002 se->statistics.block_start = rq_of(cfs_rq)->clock;
2007 clear_buddies(cfs_rq, se);
2009 if (se != cfs_rq->curr)
2010 __dequeue_entity(cfs_rq, se);
2012 account_entity_dequeue(cfs_rq, se);
2015 * Normalize the entity after updating the min_vruntime because the
2016 * update can refer to the ->curr item and we need to reflect this
2017 * movement in our normalized position.
2019 if (!(flags & DEQUEUE_SLEEP))
2020 se->vruntime -= cfs_rq->min_vruntime;
2022 /* return excess runtime on last dequeue */
2023 return_cfs_rq_runtime(cfs_rq);
2025 update_min_vruntime(cfs_rq);
2026 update_cfs_shares(cfs_rq);
2030 * Preempt the current task with a newly woken task if needed:
2033 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2035 unsigned long ideal_runtime, delta_exec;
2036 struct sched_entity *se;
2039 ideal_runtime = sched_slice(cfs_rq, curr);
2040 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2041 if (delta_exec > ideal_runtime) {
2042 resched_task(rq_of(cfs_rq)->curr);
2044 * The current task ran long enough, ensure it doesn't get
2045 * re-elected due to buddy favours.
2047 clear_buddies(cfs_rq, curr);
2052 * Ensure that a task that missed wakeup preemption by a
2053 * narrow margin doesn't have to wait for a full slice.
2054 * This also mitigates buddy induced latencies under load.
2056 if (delta_exec < sysctl_sched_min_granularity)
2059 se = __pick_first_entity(cfs_rq);
2060 delta = curr->vruntime - se->vruntime;
2065 if (delta > ideal_runtime)
2066 resched_task(rq_of(cfs_rq)->curr);
2070 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2072 /* 'current' is not kept within the tree. */
2075 * Any task has to be enqueued before it get to execute on
2076 * a CPU. So account for the time it spent waiting on the
2079 update_stats_wait_end(cfs_rq, se);
2080 __dequeue_entity(cfs_rq, se);
2081 update_entity_load_avg(se, 1);
2084 update_stats_curr_start(cfs_rq, se);
2086 #ifdef CONFIG_SCHEDSTATS
2088 * Track our maximum slice length, if the CPU's load is at
2089 * least twice that of our own weight (i.e. dont track it
2090 * when there are only lesser-weight tasks around):
2092 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2093 se->statistics.slice_max = max(se->statistics.slice_max,
2094 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2097 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2101 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2104 * Pick the next process, keeping these things in mind, in this order:
2105 * 1) keep things fair between processes/task groups
2106 * 2) pick the "next" process, since someone really wants that to run
2107 * 3) pick the "last" process, for cache locality
2108 * 4) do not run the "skip" process, if something else is available
2110 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2112 struct sched_entity *se = __pick_first_entity(cfs_rq);
2113 struct sched_entity *left = se;
2116 * Avoid running the skip buddy, if running something else can
2117 * be done without getting too unfair.
2119 if (cfs_rq->skip == se) {
2120 struct sched_entity *second = __pick_next_entity(se);
2121 if (second && wakeup_preempt_entity(second, left) < 1)
2126 * Prefer last buddy, try to return the CPU to a preempted task.
2128 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2132 * Someone really wants this to run. If it's not unfair, run it.
2134 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2137 clear_buddies(cfs_rq, se);
2142 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2144 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2147 * If still on the runqueue then deactivate_task()
2148 * was not called and update_curr() has to be done:
2151 update_curr(cfs_rq);
2153 /* throttle cfs_rqs exceeding runtime */
2154 check_cfs_rq_runtime(cfs_rq);
2156 check_spread(cfs_rq, prev);
2158 update_stats_wait_start(cfs_rq, prev);
2159 /* Put 'current' back into the tree. */
2160 __enqueue_entity(cfs_rq, prev);
2161 /* in !on_rq case, update occurred at dequeue */
2162 update_entity_load_avg(prev, 1);
2164 cfs_rq->curr = NULL;
2168 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2171 * Update run-time statistics of the 'current'.
2173 update_curr(cfs_rq);
2176 * Ensure that runnable average is periodically updated.
2178 update_entity_load_avg(curr, 1);
2179 update_cfs_rq_blocked_load(cfs_rq, 1);
2180 update_cfs_shares(cfs_rq);
2182 #ifdef CONFIG_SCHED_HRTICK
2184 * queued ticks are scheduled to match the slice, so don't bother
2185 * validating it and just reschedule.
2188 resched_task(rq_of(cfs_rq)->curr);
2192 * don't let the period tick interfere with the hrtick preemption
2194 if (!sched_feat(DOUBLE_TICK) &&
2195 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2199 if (cfs_rq->nr_running > 1)
2200 check_preempt_tick(cfs_rq, curr);
2204 /**************************************************
2205 * CFS bandwidth control machinery
2208 #ifdef CONFIG_CFS_BANDWIDTH
2210 #ifdef HAVE_JUMP_LABEL
2211 static struct static_key __cfs_bandwidth_used;
2213 static inline bool cfs_bandwidth_used(void)
2215 return static_key_false(&__cfs_bandwidth_used);
2218 void cfs_bandwidth_usage_inc(void)
2220 static_key_slow_inc(&__cfs_bandwidth_used);
2223 void cfs_bandwidth_usage_dec(void)
2225 static_key_slow_dec(&__cfs_bandwidth_used);
2227 #else /* HAVE_JUMP_LABEL */
2228 static bool cfs_bandwidth_used(void)
2233 void cfs_bandwidth_usage_inc(void) {}
2234 void cfs_bandwidth_usage_dec(void) {}
2235 #endif /* HAVE_JUMP_LABEL */
2238 * default period for cfs group bandwidth.
2239 * default: 0.1s, units: nanoseconds
2241 static inline u64 default_cfs_period(void)
2243 return 100000000ULL;
2246 static inline u64 sched_cfs_bandwidth_slice(void)
2248 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2252 * Replenish runtime according to assigned quota and update expiration time.
2253 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2254 * additional synchronization around rq->lock.
2256 * requires cfs_b->lock
2258 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2262 if (cfs_b->quota == RUNTIME_INF)
2265 now = sched_clock_cpu(smp_processor_id());
2266 cfs_b->runtime = cfs_b->quota;
2267 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2270 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2272 return &tg->cfs_bandwidth;
2275 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2276 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2278 if (unlikely(cfs_rq->throttle_count))
2279 return cfs_rq->throttled_clock_task;
2281 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2284 /* returns 0 on failure to allocate runtime */
2285 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2287 struct task_group *tg = cfs_rq->tg;
2288 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2289 u64 amount = 0, min_amount, expires;
2291 /* note: this is a positive sum as runtime_remaining <= 0 */
2292 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2294 raw_spin_lock(&cfs_b->lock);
2295 if (cfs_b->quota == RUNTIME_INF)
2296 amount = min_amount;
2299 * If the bandwidth pool has become inactive, then at least one
2300 * period must have elapsed since the last consumption.
2301 * Refresh the global state and ensure bandwidth timer becomes
2304 if (!cfs_b->timer_active) {
2305 __refill_cfs_bandwidth_runtime(cfs_b);
2306 __start_cfs_bandwidth(cfs_b);
2309 if (cfs_b->runtime > 0) {
2310 amount = min(cfs_b->runtime, min_amount);
2311 cfs_b->runtime -= amount;
2315 expires = cfs_b->runtime_expires;
2316 raw_spin_unlock(&cfs_b->lock);
2318 cfs_rq->runtime_remaining += amount;
2320 * we may have advanced our local expiration to account for allowed
2321 * spread between our sched_clock and the one on which runtime was
2324 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2325 cfs_rq->runtime_expires = expires;
2327 return cfs_rq->runtime_remaining > 0;
2331 * Note: This depends on the synchronization provided by sched_clock and the
2332 * fact that rq->clock snapshots this value.
2334 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2336 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2337 struct rq *rq = rq_of(cfs_rq);
2339 /* if the deadline is ahead of our clock, nothing to do */
2340 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2343 if (cfs_rq->runtime_remaining < 0)
2347 * If the local deadline has passed we have to consider the
2348 * possibility that our sched_clock is 'fast' and the global deadline
2349 * has not truly expired.
2351 * Fortunately we can check determine whether this the case by checking
2352 * whether the global deadline has advanced.
2355 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2356 /* extend local deadline, drift is bounded above by 2 ticks */
2357 cfs_rq->runtime_expires += TICK_NSEC;
2359 /* global deadline is ahead, expiration has passed */
2360 cfs_rq->runtime_remaining = 0;
2364 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2365 unsigned long delta_exec)
2367 /* dock delta_exec before expiring quota (as it could span periods) */
2368 cfs_rq->runtime_remaining -= delta_exec;
2369 expire_cfs_rq_runtime(cfs_rq);
2371 if (likely(cfs_rq->runtime_remaining > 0))
2375 * if we're unable to extend our runtime we resched so that the active
2376 * hierarchy can be throttled
2378 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2379 resched_task(rq_of(cfs_rq)->curr);
2382 static __always_inline
2383 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2385 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2388 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2391 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2393 return cfs_bandwidth_used() && cfs_rq->throttled;
2396 /* check whether cfs_rq, or any parent, is throttled */
2397 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2399 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2403 * Ensure that neither of the group entities corresponding to src_cpu or
2404 * dest_cpu are members of a throttled hierarchy when performing group
2405 * load-balance operations.
2407 static inline int throttled_lb_pair(struct task_group *tg,
2408 int src_cpu, int dest_cpu)
2410 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2412 src_cfs_rq = tg->cfs_rq[src_cpu];
2413 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2415 return throttled_hierarchy(src_cfs_rq) ||
2416 throttled_hierarchy(dest_cfs_rq);
2419 /* updated child weight may affect parent so we have to do this bottom up */
2420 static int tg_unthrottle_up(struct task_group *tg, void *data)
2422 struct rq *rq = data;
2423 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2425 cfs_rq->throttle_count--;
2427 if (!cfs_rq->throttle_count) {
2428 /* adjust cfs_rq_clock_task() */
2429 cfs_rq->throttled_clock_task_time += rq->clock_task -
2430 cfs_rq->throttled_clock_task;
2437 static int tg_throttle_down(struct task_group *tg, void *data)
2439 struct rq *rq = data;
2440 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2442 /* group is entering throttled state, stop time */
2443 if (!cfs_rq->throttle_count)
2444 cfs_rq->throttled_clock_task = rq->clock_task;
2445 cfs_rq->throttle_count++;
2450 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2452 struct rq *rq = rq_of(cfs_rq);
2453 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2454 struct sched_entity *se;
2455 long task_delta, dequeue = 1;
2457 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2459 /* freeze hierarchy runnable averages while throttled */
2461 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2464 task_delta = cfs_rq->h_nr_running;
2465 for_each_sched_entity(se) {
2466 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2467 /* throttled entity or throttle-on-deactivate */
2472 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2473 qcfs_rq->h_nr_running -= task_delta;
2475 if (qcfs_rq->load.weight)
2480 rq->nr_running -= task_delta;
2482 cfs_rq->throttled = 1;
2483 cfs_rq->throttled_clock = rq->clock;
2484 raw_spin_lock(&cfs_b->lock);
2485 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2486 if (!cfs_b->timer_active)
2487 __start_cfs_bandwidth(cfs_b);
2488 raw_spin_unlock(&cfs_b->lock);
2491 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2493 struct rq *rq = rq_of(cfs_rq);
2494 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2495 struct sched_entity *se;
2499 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2501 cfs_rq->throttled = 0;
2502 raw_spin_lock(&cfs_b->lock);
2503 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2504 list_del_rcu(&cfs_rq->throttled_list);
2505 raw_spin_unlock(&cfs_b->lock);
2507 update_rq_clock(rq);
2508 /* update hierarchical throttle state */
2509 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2511 if (!cfs_rq->load.weight)
2514 task_delta = cfs_rq->h_nr_running;
2515 for_each_sched_entity(se) {
2519 cfs_rq = cfs_rq_of(se);
2521 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2522 cfs_rq->h_nr_running += task_delta;
2524 if (cfs_rq_throttled(cfs_rq))
2529 rq->nr_running += task_delta;
2531 /* determine whether we need to wake up potentially idle cpu */
2532 if (rq->curr == rq->idle && rq->cfs.nr_running)
2533 resched_task(rq->curr);
2536 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2537 u64 remaining, u64 expires)
2539 struct cfs_rq *cfs_rq;
2540 u64 runtime = remaining;
2543 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2545 struct rq *rq = rq_of(cfs_rq);
2547 raw_spin_lock(&rq->lock);
2548 if (!cfs_rq_throttled(cfs_rq))
2551 runtime = -cfs_rq->runtime_remaining + 1;
2552 if (runtime > remaining)
2553 runtime = remaining;
2554 remaining -= runtime;
2556 cfs_rq->runtime_remaining += runtime;
2557 cfs_rq->runtime_expires = expires;
2559 /* we check whether we're throttled above */
2560 if (cfs_rq->runtime_remaining > 0)
2561 unthrottle_cfs_rq(cfs_rq);
2564 raw_spin_unlock(&rq->lock);
2575 * Responsible for refilling a task_group's bandwidth and unthrottling its
2576 * cfs_rqs as appropriate. If there has been no activity within the last
2577 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2578 * used to track this state.
2580 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2582 u64 runtime, runtime_expires;
2583 int idle = 1, throttled;
2585 raw_spin_lock(&cfs_b->lock);
2586 /* no need to continue the timer with no bandwidth constraint */
2587 if (cfs_b->quota == RUNTIME_INF)
2590 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2591 /* idle depends on !throttled (for the case of a large deficit) */
2592 idle = cfs_b->idle && !throttled;
2593 cfs_b->nr_periods += overrun;
2595 /* if we're going inactive then everything else can be deferred */
2600 * if we have relooped after returning idle once, we need to update our
2601 * status as actually running, so that other cpus doing
2602 * __start_cfs_bandwidth will stop trying to cancel us.
2604 cfs_b->timer_active = 1;
2606 __refill_cfs_bandwidth_runtime(cfs_b);
2609 /* mark as potentially idle for the upcoming period */
2614 /* account preceding periods in which throttling occurred */
2615 cfs_b->nr_throttled += overrun;
2618 * There are throttled entities so we must first use the new bandwidth
2619 * to unthrottle them before making it generally available. This
2620 * ensures that all existing debts will be paid before a new cfs_rq is
2623 runtime = cfs_b->runtime;
2624 runtime_expires = cfs_b->runtime_expires;
2628 * This check is repeated as we are holding onto the new bandwidth
2629 * while we unthrottle. This can potentially race with an unthrottled
2630 * group trying to acquire new bandwidth from the global pool.
2632 while (throttled && runtime > 0) {
2633 raw_spin_unlock(&cfs_b->lock);
2634 /* we can't nest cfs_b->lock while distributing bandwidth */
2635 runtime = distribute_cfs_runtime(cfs_b, runtime,
2637 raw_spin_lock(&cfs_b->lock);
2639 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2642 /* return (any) remaining runtime */
2643 cfs_b->runtime = runtime;
2645 * While we are ensured activity in the period following an
2646 * unthrottle, this also covers the case in which the new bandwidth is
2647 * insufficient to cover the existing bandwidth deficit. (Forcing the
2648 * timer to remain active while there are any throttled entities.)
2653 cfs_b->timer_active = 0;
2654 raw_spin_unlock(&cfs_b->lock);
2659 /* a cfs_rq won't donate quota below this amount */
2660 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2661 /* minimum remaining period time to redistribute slack quota */
2662 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2663 /* how long we wait to gather additional slack before distributing */
2664 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2667 * Are we near the end of the current quota period?
2669 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
2670 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
2671 * migrate_hrtimers, base is never cleared, so we are fine.
2673 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2675 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2678 /* if the call-back is running a quota refresh is already occurring */
2679 if (hrtimer_callback_running(refresh_timer))
2682 /* is a quota refresh about to occur? */
2683 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2684 if (remaining < min_expire)
2690 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2692 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2694 /* if there's a quota refresh soon don't bother with slack */
2695 if (runtime_refresh_within(cfs_b, min_left))
2698 start_bandwidth_timer(&cfs_b->slack_timer,
2699 ns_to_ktime(cfs_bandwidth_slack_period));
2702 /* we know any runtime found here is valid as update_curr() precedes return */
2703 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2705 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2706 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2708 if (slack_runtime <= 0)
2711 raw_spin_lock(&cfs_b->lock);
2712 if (cfs_b->quota != RUNTIME_INF &&
2713 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2714 cfs_b->runtime += slack_runtime;
2716 /* we are under rq->lock, defer unthrottling using a timer */
2717 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2718 !list_empty(&cfs_b->throttled_cfs_rq))
2719 start_cfs_slack_bandwidth(cfs_b);
2721 raw_spin_unlock(&cfs_b->lock);
2723 /* even if it's not valid for return we don't want to try again */
2724 cfs_rq->runtime_remaining -= slack_runtime;
2727 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2729 if (!cfs_bandwidth_used())
2732 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2735 __return_cfs_rq_runtime(cfs_rq);
2739 * This is done with a timer (instead of inline with bandwidth return) since
2740 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2742 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2744 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2747 /* confirm we're still not at a refresh boundary */
2748 raw_spin_lock(&cfs_b->lock);
2749 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
2750 raw_spin_unlock(&cfs_b->lock);
2754 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2755 runtime = cfs_b->runtime;
2758 expires = cfs_b->runtime_expires;
2759 raw_spin_unlock(&cfs_b->lock);
2764 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2766 raw_spin_lock(&cfs_b->lock);
2767 if (expires == cfs_b->runtime_expires)
2768 cfs_b->runtime = runtime;
2769 raw_spin_unlock(&cfs_b->lock);
2773 * When a group wakes up we want to make sure that its quota is not already
2774 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2775 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2777 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2779 if (!cfs_bandwidth_used())
2782 /* an active group must be handled by the update_curr()->put() path */
2783 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2786 /* ensure the group is not already throttled */
2787 if (cfs_rq_throttled(cfs_rq))
2790 /* update runtime allocation */
2791 account_cfs_rq_runtime(cfs_rq, 0);
2792 if (cfs_rq->runtime_remaining <= 0)
2793 throttle_cfs_rq(cfs_rq);
2796 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2797 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2799 if (!cfs_bandwidth_used())
2802 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2806 * it's possible for a throttled entity to be forced into a running
2807 * state (e.g. set_curr_task), in this case we're finished.
2809 if (cfs_rq_throttled(cfs_rq))
2812 throttle_cfs_rq(cfs_rq);
2815 static inline u64 default_cfs_period(void);
2816 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2817 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2819 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2821 struct cfs_bandwidth *cfs_b =
2822 container_of(timer, struct cfs_bandwidth, slack_timer);
2823 do_sched_cfs_slack_timer(cfs_b);
2825 return HRTIMER_NORESTART;
2828 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2830 struct cfs_bandwidth *cfs_b =
2831 container_of(timer, struct cfs_bandwidth, period_timer);
2837 now = hrtimer_cb_get_time(timer);
2838 overrun = hrtimer_forward(timer, now, cfs_b->period);
2843 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2846 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2849 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2851 raw_spin_lock_init(&cfs_b->lock);
2853 cfs_b->quota = RUNTIME_INF;
2854 cfs_b->period = ns_to_ktime(default_cfs_period());
2856 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2857 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2858 cfs_b->period_timer.function = sched_cfs_period_timer;
2859 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2860 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2863 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2865 cfs_rq->runtime_enabled = 0;
2866 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2869 /* requires cfs_b->lock, may release to reprogram timer */
2870 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2873 * The timer may be active because we're trying to set a new bandwidth
2874 * period or because we're racing with the tear-down path
2875 * (timer_active==0 becomes visible before the hrtimer call-back
2876 * terminates). In either case we ensure that it's re-programmed
2878 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
2879 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
2880 /* bounce the lock to allow do_sched_cfs_period_timer to run */
2881 raw_spin_unlock(&cfs_b->lock);
2883 raw_spin_lock(&cfs_b->lock);
2884 /* if someone else restarted the timer then we're done */
2885 if (cfs_b->timer_active)
2889 cfs_b->timer_active = 1;
2890 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2893 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2895 hrtimer_cancel(&cfs_b->period_timer);
2896 hrtimer_cancel(&cfs_b->slack_timer);
2899 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2901 struct cfs_rq *cfs_rq;
2903 for_each_leaf_cfs_rq(rq, cfs_rq) {
2904 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2906 if (!cfs_rq->runtime_enabled)
2910 * clock_task is not advancing so we just need to make sure
2911 * there's some valid quota amount
2913 cfs_rq->runtime_remaining = cfs_b->quota;
2914 if (cfs_rq_throttled(cfs_rq))
2915 unthrottle_cfs_rq(cfs_rq);
2919 #else /* CONFIG_CFS_BANDWIDTH */
2920 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2922 return rq_of(cfs_rq)->clock_task;
2925 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2926 unsigned long delta_exec) {}
2927 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2928 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2929 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2931 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2936 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2941 static inline int throttled_lb_pair(struct task_group *tg,
2942 int src_cpu, int dest_cpu)
2947 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2949 #ifdef CONFIG_FAIR_GROUP_SCHED
2950 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2953 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2957 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2958 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2960 #endif /* CONFIG_CFS_BANDWIDTH */
2962 /**************************************************
2963 * CFS operations on tasks:
2966 #ifdef CONFIG_SCHED_HRTICK
2967 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2969 struct sched_entity *se = &p->se;
2970 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2972 WARN_ON(task_rq(p) != rq);
2974 if (cfs_rq->nr_running > 1) {
2975 u64 slice = sched_slice(cfs_rq, se);
2976 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2977 s64 delta = slice - ran;
2986 * Don't schedule slices shorter than 10000ns, that just
2987 * doesn't make sense. Rely on vruntime for fairness.
2990 delta = max_t(s64, 10000LL, delta);
2992 hrtick_start(rq, delta);
2997 * called from enqueue/dequeue and updates the hrtick when the
2998 * current task is from our class and nr_running is low enough
3001 static void hrtick_update(struct rq *rq)
3003 struct task_struct *curr = rq->curr;
3005 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3008 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3009 hrtick_start_fair(rq, curr);
3011 #else /* !CONFIG_SCHED_HRTICK */
3013 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3017 static inline void hrtick_update(struct rq *rq)
3023 * The enqueue_task method is called before nr_running is
3024 * increased. Here we update the fair scheduling stats and
3025 * then put the task into the rbtree:
3028 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3030 struct cfs_rq *cfs_rq;
3031 struct sched_entity *se = &p->se;
3033 for_each_sched_entity(se) {
3036 cfs_rq = cfs_rq_of(se);
3037 enqueue_entity(cfs_rq, se, flags);
3040 * end evaluation on encountering a throttled cfs_rq
3042 * note: in the case of encountering a throttled cfs_rq we will
3043 * post the final h_nr_running increment below.
3045 if (cfs_rq_throttled(cfs_rq))
3047 cfs_rq->h_nr_running++;
3049 flags = ENQUEUE_WAKEUP;
3052 for_each_sched_entity(se) {
3053 cfs_rq = cfs_rq_of(se);
3054 cfs_rq->h_nr_running++;
3056 if (cfs_rq_throttled(cfs_rq))
3059 update_cfs_shares(cfs_rq);
3060 update_entity_load_avg(se, 1);
3064 update_rq_runnable_avg(rq, rq->nr_running);
3070 static void set_next_buddy(struct sched_entity *se);
3073 * The dequeue_task method is called before nr_running is
3074 * decreased. We remove the task from the rbtree and
3075 * update the fair scheduling stats:
3077 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3079 struct cfs_rq *cfs_rq;
3080 struct sched_entity *se = &p->se;
3081 int task_sleep = flags & DEQUEUE_SLEEP;
3083 for_each_sched_entity(se) {
3084 cfs_rq = cfs_rq_of(se);
3085 dequeue_entity(cfs_rq, se, flags);
3088 * end evaluation on encountering a throttled cfs_rq
3090 * note: in the case of encountering a throttled cfs_rq we will
3091 * post the final h_nr_running decrement below.
3093 if (cfs_rq_throttled(cfs_rq))
3095 cfs_rq->h_nr_running--;
3097 /* Don't dequeue parent if it has other entities besides us */
3098 if (cfs_rq->load.weight) {
3100 * Bias pick_next to pick a task from this cfs_rq, as
3101 * p is sleeping when it is within its sched_slice.
3103 if (task_sleep && parent_entity(se))
3104 set_next_buddy(parent_entity(se));
3106 /* avoid re-evaluating load for this entity */
3107 se = parent_entity(se);
3110 flags |= DEQUEUE_SLEEP;
3113 for_each_sched_entity(se) {
3114 cfs_rq = cfs_rq_of(se);
3115 cfs_rq->h_nr_running--;
3117 if (cfs_rq_throttled(cfs_rq))
3120 update_cfs_shares(cfs_rq);
3121 update_entity_load_avg(se, 1);
3126 update_rq_runnable_avg(rq, 1);
3132 /* Used instead of source_load when we know the type == 0 */
3133 static unsigned long weighted_cpuload(const int cpu)
3135 return cpu_rq(cpu)->load.weight;
3139 * Return a low guess at the load of a migration-source cpu weighted
3140 * according to the scheduling class and "nice" value.
3142 * We want to under-estimate the load of migration sources, to
3143 * balance conservatively.
3145 static unsigned long source_load(int cpu, int type)
3147 struct rq *rq = cpu_rq(cpu);
3148 unsigned long total = weighted_cpuload(cpu);
3150 if (type == 0 || !sched_feat(LB_BIAS))
3153 return min(rq->cpu_load[type-1], total);
3157 * Return a high guess at the load of a migration-target cpu weighted
3158 * according to the scheduling class and "nice" value.
3160 static unsigned long target_load(int cpu, int type)
3162 struct rq *rq = cpu_rq(cpu);
3163 unsigned long total = weighted_cpuload(cpu);
3165 if (type == 0 || !sched_feat(LB_BIAS))
3168 return max(rq->cpu_load[type-1], total);
3171 static unsigned long power_of(int cpu)
3173 return cpu_rq(cpu)->cpu_power;
3176 static unsigned long cpu_avg_load_per_task(int cpu)
3178 struct rq *rq = cpu_rq(cpu);
3179 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3182 return rq->load.weight / nr_running;
3188 static void task_waking_fair(struct task_struct *p)
3190 struct sched_entity *se = &p->se;
3191 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3194 #ifndef CONFIG_64BIT
3195 u64 min_vruntime_copy;
3198 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3200 min_vruntime = cfs_rq->min_vruntime;
3201 } while (min_vruntime != min_vruntime_copy);
3203 min_vruntime = cfs_rq->min_vruntime;
3206 se->vruntime -= min_vruntime;
3209 #ifdef CONFIG_FAIR_GROUP_SCHED
3211 * effective_load() calculates the load change as seen from the root_task_group
3213 * Adding load to a group doesn't make a group heavier, but can cause movement
3214 * of group shares between cpus. Assuming the shares were perfectly aligned one
3215 * can calculate the shift in shares.
3217 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3218 * on this @cpu and results in a total addition (subtraction) of @wg to the
3219 * total group weight.
3221 * Given a runqueue weight distribution (rw_i) we can compute a shares
3222 * distribution (s_i) using:
3224 * s_i = rw_i / \Sum rw_j (1)
3226 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3227 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3228 * shares distribution (s_i):
3230 * rw_i = { 2, 4, 1, 0 }
3231 * s_i = { 2/7, 4/7, 1/7, 0 }
3233 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3234 * task used to run on and the CPU the waker is running on), we need to
3235 * compute the effect of waking a task on either CPU and, in case of a sync
3236 * wakeup, compute the effect of the current task going to sleep.
3238 * So for a change of @wl to the local @cpu with an overall group weight change
3239 * of @wl we can compute the new shares distribution (s'_i) using:
3241 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3243 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3244 * differences in waking a task to CPU 0. The additional task changes the
3245 * weight and shares distributions like:
3247 * rw'_i = { 3, 4, 1, 0 }
3248 * s'_i = { 3/8, 4/8, 1/8, 0 }
3250 * We can then compute the difference in effective weight by using:
3252 * dw_i = S * (s'_i - s_i) (3)
3254 * Where 'S' is the group weight as seen by its parent.
3256 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3257 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3258 * 4/7) times the weight of the group.
3260 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3262 struct sched_entity *se = tg->se[cpu];
3264 if (!tg->parent) /* the trivial, non-cgroup case */
3267 for_each_sched_entity(se) {
3273 * W = @wg + \Sum rw_j
3275 W = wg + calc_tg_weight(tg, se->my_q);
3280 w = se->my_q->load.weight + wl;
3283 * wl = S * s'_i; see (2)
3286 wl = (w * tg->shares) / W;
3291 * Per the above, wl is the new se->load.weight value; since
3292 * those are clipped to [MIN_SHARES, ...) do so now. See
3293 * calc_cfs_shares().
3295 if (wl < MIN_SHARES)
3299 * wl = dw_i = S * (s'_i - s_i); see (3)
3301 wl -= se->load.weight;
3304 * Recursively apply this logic to all parent groups to compute
3305 * the final effective load change on the root group. Since
3306 * only the @tg group gets extra weight, all parent groups can
3307 * only redistribute existing shares. @wl is the shift in shares
3308 * resulting from this level per the above.
3317 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3318 unsigned long wl, unsigned long wg)
3325 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3327 s64 this_load, load;
3328 int idx, this_cpu, prev_cpu;
3329 unsigned long tl_per_task;
3330 struct task_group *tg;
3331 unsigned long weight;
3335 this_cpu = smp_processor_id();
3336 prev_cpu = task_cpu(p);
3337 load = source_load(prev_cpu, idx);
3338 this_load = target_load(this_cpu, idx);
3341 * If sync wakeup then subtract the (maximum possible)
3342 * effect of the currently running task from the load
3343 * of the current CPU:
3346 tg = task_group(current);
3347 weight = current->se.load.weight;
3349 this_load += effective_load(tg, this_cpu, -weight, -weight);
3350 load += effective_load(tg, prev_cpu, 0, -weight);
3354 weight = p->se.load.weight;
3357 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3358 * due to the sync cause above having dropped this_load to 0, we'll
3359 * always have an imbalance, but there's really nothing you can do
3360 * about that, so that's good too.
3362 * Otherwise check if either cpus are near enough in load to allow this
3363 * task to be woken on this_cpu.
3365 if (this_load > 0) {
3366 s64 this_eff_load, prev_eff_load;
3368 this_eff_load = 100;
3369 this_eff_load *= power_of(prev_cpu);
3370 this_eff_load *= this_load +
3371 effective_load(tg, this_cpu, weight, weight);
3373 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3374 prev_eff_load *= power_of(this_cpu);
3375 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3377 balanced = this_eff_load <= prev_eff_load;
3382 * If the currently running task will sleep within
3383 * a reasonable amount of time then attract this newly
3386 if (sync && balanced)
3389 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3390 tl_per_task = cpu_avg_load_per_task(this_cpu);
3393 (this_load <= load &&
3394 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3396 * This domain has SD_WAKE_AFFINE and
3397 * p is cache cold in this domain, and
3398 * there is no bad imbalance.
3400 schedstat_inc(sd, ttwu_move_affine);
3401 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3409 * find_idlest_group finds and returns the least busy CPU group within the
3412 static struct sched_group *
3413 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3414 int this_cpu, int load_idx)
3416 struct sched_group *idlest = NULL, *group = sd->groups;
3417 unsigned long min_load = ULONG_MAX, this_load = 0;
3418 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3421 unsigned long load, avg_load;
3425 /* Skip over this group if it has no CPUs allowed */
3426 if (!cpumask_intersects(sched_group_cpus(group),
3427 tsk_cpus_allowed(p)))
3430 local_group = cpumask_test_cpu(this_cpu,
3431 sched_group_cpus(group));
3433 /* Tally up the load of all CPUs in the group */
3436 for_each_cpu(i, sched_group_cpus(group)) {
3437 /* Bias balancing toward cpus of our domain */
3439 load = source_load(i, load_idx);
3441 load = target_load(i, load_idx);
3446 /* Adjust by relative CPU power of the group */
3447 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3450 this_load = avg_load;
3451 } else if (avg_load < min_load) {
3452 min_load = avg_load;
3455 } while (group = group->next, group != sd->groups);
3457 if (!idlest || 100*this_load < imbalance*min_load)
3463 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3466 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3468 unsigned long load, min_load = ULONG_MAX;
3472 /* Traverse only the allowed CPUs */
3473 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3474 load = weighted_cpuload(i);
3476 if (load < min_load || (load == min_load && i == this_cpu)) {
3486 * Try and locate an idle CPU in the sched_domain.
3488 static int select_idle_sibling(struct task_struct *p, int target)
3490 struct sched_domain *sd;
3491 struct sched_group *sg;
3492 int i = task_cpu(p);
3494 if (idle_cpu(target))
3498 * If the prevous cpu is cache affine and idle, don't be stupid.
3500 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3504 * Otherwise, iterate the domains and find an elegible idle cpu.
3506 sd = rcu_dereference(per_cpu(sd_llc, target));
3507 for_each_lower_domain(sd) {
3510 if (!cpumask_intersects(sched_group_cpus(sg),
3511 tsk_cpus_allowed(p)))
3514 for_each_cpu(i, sched_group_cpus(sg)) {
3515 if (i == target || !idle_cpu(i))
3519 target = cpumask_first_and(sched_group_cpus(sg),
3520 tsk_cpus_allowed(p));
3524 } while (sg != sd->groups);
3530 #ifdef CONFIG_SCHED_HMP
3532 * Heterogenous multiprocessor (HMP) optimizations
3534 * The cpu types are distinguished using a list of hmp_domains
3535 * which each represent one cpu type using a cpumask.
3536 * The list is assumed ordered by compute capacity with the
3537 * fastest domain first.
3539 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3540 static const int hmp_max_tasks = 5;
3542 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3544 /* Setup hmp_domains */
3545 static int __init hmp_cpu_mask_setup(void)
3548 struct hmp_domain *domain;
3549 struct list_head *pos;
3552 pr_debug("Initializing HMP scheduler:\n");
3554 /* Initialize hmp_domains using platform code */
3555 arch_get_hmp_domains(&hmp_domains);
3556 if (list_empty(&hmp_domains)) {
3557 pr_debug("HMP domain list is empty!\n");
3561 /* Print hmp_domains */
3563 list_for_each(pos, &hmp_domains) {
3564 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3565 cpulist_scnprintf(buf, 64, &domain->possible_cpus);
3566 pr_debug(" HMP domain %d: %s\n", dc, buf);
3568 for_each_cpu_mask(cpu, domain->possible_cpus) {
3569 per_cpu(hmp_cpu_domain, cpu) = domain;
3577 static struct hmp_domain *hmp_get_hmp_domain_for_cpu(int cpu)
3579 struct hmp_domain *domain;
3580 struct list_head *pos;
3582 list_for_each(pos, &hmp_domains) {
3583 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3584 if(cpumask_test_cpu(cpu, &domain->possible_cpus))
3590 static void hmp_online_cpu(int cpu)
3592 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3595 cpumask_set_cpu(cpu, &domain->cpus);
3598 static void hmp_offline_cpu(int cpu)
3600 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3603 cpumask_clear_cpu(cpu, &domain->cpus);
3606 * Needed to determine heaviest tasks etc.
3608 static inline unsigned int hmp_cpu_is_fastest(int cpu);
3609 static inline unsigned int hmp_cpu_is_slowest(int cpu);
3610 static inline struct hmp_domain *hmp_slower_domain(int cpu);
3611 static inline struct hmp_domain *hmp_faster_domain(int cpu);
3613 /* must hold runqueue lock for queue se is currently on */
3614 static struct sched_entity *hmp_get_heaviest_task(
3615 struct sched_entity *se, int migrate_up)
3617 int num_tasks = hmp_max_tasks;
3618 struct sched_entity *max_se = se;
3619 unsigned long int max_ratio = se->avg.load_avg_ratio;
3620 const struct cpumask *hmp_target_mask = NULL;
3623 struct hmp_domain *hmp;
3624 if (hmp_cpu_is_fastest(cpu_of(se->cfs_rq->rq)))
3627 hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq));
3628 hmp_target_mask = &hmp->cpus;
3630 /* The currently running task is not on the runqueue */
3631 se = __pick_first_entity(cfs_rq_of(se));
3633 while (num_tasks && se) {
3634 if (entity_is_task(se) &&
3635 (se->avg.load_avg_ratio > max_ratio &&
3637 cpumask_intersects(hmp_target_mask,
3638 tsk_cpus_allowed(task_of(se))))) {
3640 max_ratio = se->avg.load_avg_ratio;
3642 se = __pick_next_entity(se);
3648 static struct sched_entity *hmp_get_lightest_task(
3649 struct sched_entity *se, int migrate_down)
3651 int num_tasks = hmp_max_tasks;
3652 struct sched_entity *min_se = se;
3653 unsigned long int min_ratio = se->avg.load_avg_ratio;
3654 const struct cpumask *hmp_target_mask = NULL;
3657 struct hmp_domain *hmp;
3658 if (hmp_cpu_is_slowest(cpu_of(se->cfs_rq->rq)))
3660 hmp = hmp_slower_domain(cpu_of(se->cfs_rq->rq));
3661 hmp_target_mask = &hmp->cpus;
3663 /* The currently running task is not on the runqueue */
3664 se = __pick_first_entity(cfs_rq_of(se));
3666 while (num_tasks && se) {
3667 if (entity_is_task(se) &&
3668 (se->avg.load_avg_ratio < min_ratio &&
3670 cpumask_intersects(hmp_target_mask,
3671 tsk_cpus_allowed(task_of(se))))) {
3673 min_ratio = se->avg.load_avg_ratio;
3675 se = __pick_next_entity(se);
3682 * Migration thresholds should be in the range [0..1023]
3683 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3684 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3686 * hmp_up_prio: Only up migrate task with high priority (<hmp_up_prio)
3687 * hmp_next_up_threshold: Delay before next up migration (1024 ~= 1 ms)
3688 * hmp_next_down_threshold: Delay before next down migration (1024 ~= 1 ms)
3690 * Small Task Packing:
3691 * We can choose to fill the littlest CPUs in an HMP system rather than
3692 * the typical spreading mechanic. This behavior is controllable using
3694 * hmp_packing_enabled: runtime control over pack/spread
3695 * hmp_full_threshold: Consider a CPU with this much unweighted load full
3697 unsigned int hmp_up_threshold = 700;
3698 unsigned int hmp_down_threshold = 512;
3699 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
3700 unsigned int hmp_up_prio = NICE_TO_PRIO(CONFIG_SCHED_HMP_PRIO_FILTER_VAL);
3702 unsigned int hmp_next_up_threshold = 4096;
3703 unsigned int hmp_next_down_threshold = 4096;
3705 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3707 * Set the default packing threshold to try to keep little
3708 * CPUs at no more than 80% of their maximum frequency if only
3709 * packing a small number of small tasks. Bigger tasks will
3710 * raise frequency as normal.
3711 * In order to pack a task onto a CPU, the sum of the
3712 * unweighted runnable_avg load of existing tasks plus the
3713 * load of the new task must be less than hmp_full_threshold.
3715 * This works in conjunction with frequency-invariant load
3716 * and DVFS governors. Since most DVFS governors aim for 80%
3717 * utilisation, we arrive at (0.8*0.8*(max_load=1024))=655
3718 * and use a value slightly lower to give a little headroom
3720 * Note that the most efficient frequency is different for
3721 * each system so /sys/kernel/hmp/packing_limit should be
3722 * configured at runtime for any given platform to achieve
3723 * optimal energy usage. Some systems may not benefit from
3724 * packing, so this feature can also be disabled at runtime
3725 * with /sys/kernel/hmp/packing_enable
3727 unsigned int hmp_packing_enabled = 1;
3728 unsigned int hmp_full_threshold = 650;
3731 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se);
3732 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3733 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
3734 int *min_cpu, struct cpumask *affinity);
3736 static inline struct hmp_domain *hmp_smallest_domain(void)
3738 return list_entry(hmp_domains.prev, struct hmp_domain, hmp_domains);
3741 /* Check if cpu is in fastest hmp_domain */
3742 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3744 struct list_head *pos;
3746 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3747 return pos == hmp_domains.next;
3750 /* Check if cpu is in slowest hmp_domain */
3751 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3753 struct list_head *pos;
3755 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3756 return list_is_last(pos, &hmp_domains);
3759 /* Next (slower) hmp_domain relative to cpu */
3760 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3762 struct list_head *pos;
3764 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3765 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3768 /* Previous (faster) hmp_domain relative to cpu */
3769 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3771 struct list_head *pos;
3773 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3774 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3778 * Selects a cpu in previous (faster) hmp_domain
3780 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3783 int lowest_cpu=NR_CPUS;
3784 __always_unused int lowest_ratio;
3785 struct hmp_domain *hmp;
3787 if (hmp_cpu_is_fastest(cpu))
3788 hmp = hmp_cpu_domain(cpu);
3790 hmp = hmp_faster_domain(cpu);
3792 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3793 tsk_cpus_allowed(tsk));
3799 * Selects a cpu in next (slower) hmp_domain
3800 * Note that cpumask_any_and() returns the first cpu in the cpumask
3802 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3805 int lowest_cpu=NR_CPUS;
3806 struct hmp_domain *hmp;
3807 __always_unused int lowest_ratio;
3809 if (hmp_cpu_is_slowest(cpu))
3810 hmp = hmp_cpu_domain(cpu);
3812 hmp = hmp_slower_domain(cpu);
3814 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3815 tsk_cpus_allowed(tsk));
3819 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3821 * Select the 'best' candidate little CPU to wake up on.
3822 * Implements a packing strategy which examines CPU in
3823 * logical CPU order, and selects the first which will
3824 * be loaded less than hmp_full_threshold according to
3825 * the sum of the tracked load of the runqueue and the task.
3827 static inline unsigned int hmp_best_little_cpu(struct task_struct *tsk,
3830 unsigned long estimated_load;
3831 struct hmp_domain *hmp;
3832 struct sched_avg *avg;
3833 struct cpumask allowed_hmp_cpus;
3835 if(!hmp_packing_enabled ||
3836 tsk->se.avg.load_avg_ratio > ((NICE_0_LOAD * 90)/100))
3837 return hmp_select_slower_cpu(tsk, cpu);
3839 if (hmp_cpu_is_slowest(cpu))
3840 hmp = hmp_cpu_domain(cpu);
3842 hmp = hmp_slower_domain(cpu);
3844 /* respect affinity */
3845 cpumask_and(&allowed_hmp_cpus, &hmp->cpus,
3846 tsk_cpus_allowed(tsk));
3848 for_each_cpu_mask(tmp_cpu, allowed_hmp_cpus) {
3849 avg = &cpu_rq(tmp_cpu)->avg;
3850 /* estimate new rq load if we add this task */
3851 estimated_load = avg->load_avg_ratio +
3852 tsk->se.avg.load_avg_ratio;
3853 if (estimated_load <= hmp_full_threshold) {
3858 /* if no match was found, the task uses the initial value */
3862 static inline void hmp_next_up_delay(struct sched_entity *se, int cpu)
3864 /* hack - always use clock from first online CPU */
3865 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3866 se->avg.hmp_last_up_migration = now;
3867 se->avg.hmp_last_down_migration = 0;
3868 cpu_rq(cpu)->avg.hmp_last_up_migration = now;
3869 cpu_rq(cpu)->avg.hmp_last_down_migration = 0;
3872 static inline void hmp_next_down_delay(struct sched_entity *se, int cpu)
3874 /* hack - always use clock from first online CPU */
3875 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3876 se->avg.hmp_last_down_migration = now;
3877 se->avg.hmp_last_up_migration = 0;
3878 cpu_rq(cpu)->avg.hmp_last_down_migration = now;
3879 cpu_rq(cpu)->avg.hmp_last_up_migration = 0;
3883 * Heterogenous multiprocessor (HMP) optimizations
3885 * These functions allow to change the growing speed of the load_avg_ratio
3886 * by default it goes from 0 to 0.5 in LOAD_AVG_PERIOD = 32ms
3887 * This can now be changed with /sys/kernel/hmp/load_avg_period_ms.
3889 * These functions also allow to change the up and down threshold of HMP
3890 * using /sys/kernel/hmp/{up,down}_threshold.
3891 * Both must be between 0 and 1023. The threshold that is compared
3892 * to the load_avg_ratio is up_threshold/1024 and down_threshold/1024.
3894 * For instance, if load_avg_period = 64 and up_threshold = 512, an idle
3895 * task with a load of 0 will reach the threshold after 64ms of busy loop.
3897 * Changing load_avg_periods_ms has the same effect than changing the
3898 * default scaling factor Y=1002/1024 in the load_avg_ratio computation to
3899 * (1002/1024.0)^(LOAD_AVG_PERIOD/load_avg_period_ms), but the last one
3900 * could trigger overflows.
3901 * For instance, with Y = 1023/1024 in __update_task_entity_contrib()
3902 * "contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);"
3903 * could be overflowed for a weight > 2^12 even is the load_avg_contrib
3904 * should still be a 32bits result. This would not happen by multiplicating
3905 * delta time by 1/22 and setting load_avg_period_ms = 706.
3909 * By scaling the delta time it end-up increasing or decrease the
3910 * growing speed of the per entity load_avg_ratio
3911 * The scale factor hmp_data.multiplier is a fixed point
3912 * number: (32-HMP_VARIABLE_SCALE_SHIFT).HMP_VARIABLE_SCALE_SHIFT
3914 static inline u64 hmp_variable_scale_convert(u64 delta)
3916 #ifdef CONFIG_HMP_VARIABLE_SCALE
3917 u64 high = delta >> 32ULL;
3918 u64 low = delta & 0xffffffffULL;
3919 low *= hmp_data.multiplier;
3920 high *= hmp_data.multiplier;
3921 return (low >> HMP_VARIABLE_SCALE_SHIFT)
3922 + (high << (32ULL - HMP_VARIABLE_SCALE_SHIFT));
3928 static ssize_t hmp_show(struct kobject *kobj,
3929 struct attribute *attr, char *buf)
3931 struct hmp_global_attr *hmp_attr =
3932 container_of(attr, struct hmp_global_attr, attr);
3935 if (hmp_attr->to_sysfs_text != NULL)
3936 return hmp_attr->to_sysfs_text(buf, PAGE_SIZE);
3938 temp = *(hmp_attr->value);
3939 if (hmp_attr->to_sysfs != NULL)
3940 temp = hmp_attr->to_sysfs(temp);
3942 return (ssize_t)sprintf(buf, "%d\n", temp);
3945 static ssize_t hmp_store(struct kobject *a, struct attribute *attr,
3946 const char *buf, size_t count)
3949 ssize_t ret = count;
3950 struct hmp_global_attr *hmp_attr =
3951 container_of(attr, struct hmp_global_attr, attr);
3952 char *str = vmalloc(count + 1);
3955 memcpy(str, buf, count);
3957 if (sscanf(str, "%d", &temp) < 1)
3960 if (hmp_attr->from_sysfs != NULL)
3961 temp = hmp_attr->from_sysfs(temp);
3965 *(hmp_attr->value) = temp;
3971 static ssize_t hmp_print_domains(char *outbuf, int outbufsize)
3974 const char nospace[] = "%s", space[] = " %s";
3975 const char *fmt = nospace;
3976 struct hmp_domain *domain;
3977 struct list_head *pos;
3979 list_for_each(pos, &hmp_domains) {
3980 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3981 if (cpumask_scnprintf(buf, 64, &domain->possible_cpus)) {
3982 outpos += sprintf(outbuf+outpos, fmt, buf);
3986 strcat(outbuf, "\n");
3990 #ifdef CONFIG_HMP_VARIABLE_SCALE
3991 static int hmp_period_tofrom_sysfs(int value)
3993 return (LOAD_AVG_PERIOD << HMP_VARIABLE_SCALE_SHIFT) / value;
3996 /* max value for threshold is 1024 */
3997 static int hmp_theshold_from_sysfs(int value)
4003 #if defined(CONFIG_SCHED_HMP_LITTLE_PACKING) || \
4004 defined(CONFIG_HMP_FREQUENCY_INVARIANT_SCALE)
4005 /* toggle control is only 0,1 off/on */
4006 static int hmp_toggle_from_sysfs(int value)
4008 if (value < 0 || value > 1)
4013 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4014 /* packing value must be non-negative */
4015 static int hmp_packing_from_sysfs(int value)
4022 static void hmp_attr_add(
4025 int (*to_sysfs)(int),
4026 int (*from_sysfs)(int),
4027 ssize_t (*to_sysfs_text)(char *, int),
4031 while (hmp_data.attributes[i] != NULL) {
4033 if (i >= HMP_DATA_SYSFS_MAX)
4037 hmp_data.attr[i].attr.mode = mode;
4039 hmp_data.attr[i].attr.mode = 0644;
4040 hmp_data.attr[i].show = hmp_show;
4041 hmp_data.attr[i].store = hmp_store;
4042 hmp_data.attr[i].attr.name = name;
4043 hmp_data.attr[i].value = value;
4044 hmp_data.attr[i].to_sysfs = to_sysfs;
4045 hmp_data.attr[i].from_sysfs = from_sysfs;
4046 hmp_data.attr[i].to_sysfs_text = to_sysfs_text;
4047 hmp_data.attributes[i] = &hmp_data.attr[i].attr;
4048 hmp_data.attributes[i + 1] = NULL;
4051 static int hmp_attr_init(void)
4054 memset(&hmp_data, sizeof(hmp_data), 0);
4055 hmp_attr_add("hmp_domains",
4061 hmp_attr_add("up_threshold",
4064 hmp_theshold_from_sysfs,
4067 hmp_attr_add("down_threshold",
4068 &hmp_down_threshold,
4070 hmp_theshold_from_sysfs,
4073 #ifdef CONFIG_HMP_VARIABLE_SCALE
4074 /* by default load_avg_period_ms == LOAD_AVG_PERIOD
4077 hmp_data.multiplier = hmp_period_tofrom_sysfs(LOAD_AVG_PERIOD);
4078 hmp_attr_add("load_avg_period_ms",
4079 &hmp_data.multiplier,
4080 hmp_period_tofrom_sysfs,
4081 hmp_period_tofrom_sysfs,
4085 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
4086 /* default frequency-invariant scaling ON */
4087 hmp_data.freqinvar_load_scale_enabled = 1;
4088 hmp_attr_add("frequency_invariant_load_scale",
4089 &hmp_data.freqinvar_load_scale_enabled,
4091 hmp_toggle_from_sysfs,
4095 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4096 hmp_attr_add("packing_enable",
4097 &hmp_packing_enabled,
4099 hmp_toggle_from_sysfs,
4102 hmp_attr_add("packing_limit",
4103 &hmp_full_threshold,
4105 hmp_packing_from_sysfs,
4109 hmp_data.attr_group.name = "hmp";
4110 hmp_data.attr_group.attrs = hmp_data.attributes;
4111 ret = sysfs_create_group(kernel_kobj,
4112 &hmp_data.attr_group);
4115 late_initcall(hmp_attr_init);
4117 * return the load of the lowest-loaded CPU in a given HMP domain
4118 * min_cpu optionally points to an int to receive the CPU.
4119 * affinity optionally points to a cpumask containing the
4120 * CPUs to be considered. note:
4121 * + min_cpu = NR_CPUS only if no CPUs are in the set of
4122 * affinity && hmp_domain cpus
4123 * + min_cpu will always otherwise equal one of the CPUs in
4125 * + when more than one CPU has the same load, the one which
4126 * is least-recently-disturbed by an HMP migration will be
4128 * + if all CPUs are equally loaded or idle and the times are
4129 * all the same, the first in the set will be used
4130 * + if affinity is not set, cpu_online_mask is used
4132 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
4133 int *min_cpu, struct cpumask *affinity)
4136 int min_cpu_runnable_temp = NR_CPUS;
4137 u64 min_target_last_migration = ULLONG_MAX;
4138 u64 curr_last_migration;
4139 unsigned long min_runnable_load = INT_MAX;
4140 unsigned long contrib;
4141 struct sched_avg *avg;
4142 struct cpumask temp_cpumask;
4144 * only look at CPUs allowed if specified,
4145 * otherwise look at all online CPUs in the
4148 cpumask_and(&temp_cpumask, &hmpd->cpus, affinity ? affinity : cpu_online_mask);
4150 for_each_cpu_mask(cpu, temp_cpumask) {
4151 avg = &cpu_rq(cpu)->avg;
4152 /* used for both up and down migration */
4153 curr_last_migration = avg->hmp_last_up_migration ?
4154 avg->hmp_last_up_migration : avg->hmp_last_down_migration;
4156 contrib = avg->load_avg_ratio;
4158 * Consider a runqueue completely busy if there is any load
4159 * on it. Definitely not the best for overall fairness, but
4160 * does well in typical Android use cases.
4165 if ((contrib < min_runnable_load) ||
4166 (contrib == min_runnable_load &&
4167 curr_last_migration < min_target_last_migration)) {
4169 * if the load is the same target the CPU with
4170 * the longest time since a migration.
4171 * This is to spread migration load between
4172 * members of a domain more evenly when the
4173 * domain is fully loaded
4175 min_runnable_load = contrib;
4176 min_cpu_runnable_temp = cpu;
4177 min_target_last_migration = curr_last_migration;
4182 *min_cpu = min_cpu_runnable_temp;
4184 return min_runnable_load;
4188 * Calculate the task starvation
4189 * This is the ratio of actually running time vs. runnable time.
4190 * If the two are equal the task is getting the cpu time it needs or
4191 * it is alone on the cpu and the cpu is fully utilized.
4193 static inline unsigned int hmp_task_starvation(struct sched_entity *se)
4197 starvation = se->avg.usage_avg_sum * scale_load_down(NICE_0_LOAD);
4198 starvation /= (se->avg.runnable_avg_sum + 1);
4200 return scale_load(starvation);
4203 static inline unsigned int hmp_offload_down(int cpu, struct sched_entity *se)
4206 int dest_cpu = NR_CPUS;
4208 if (hmp_cpu_is_slowest(cpu))
4211 /* Is there an idle CPU in the current domain */
4212 min_usage = hmp_domain_min_load(hmp_cpu_domain(cpu), NULL, NULL);
4213 if (min_usage == 0) {
4214 trace_sched_hmp_offload_abort(cpu, min_usage, "load");
4218 /* Is the task alone on the cpu? */
4219 if (cpu_rq(cpu)->cfs.h_nr_running < 2) {
4220 trace_sched_hmp_offload_abort(cpu,
4221 cpu_rq(cpu)->cfs.h_nr_running, "nr_running");
4225 /* Is the task actually starving? */
4226 /* >=25% ratio running/runnable = starving */
4227 if (hmp_task_starvation(se) > 768) {
4228 trace_sched_hmp_offload_abort(cpu, hmp_task_starvation(se),
4233 /* Does the slower domain have any idle CPUs? */
4234 min_usage = hmp_domain_min_load(hmp_slower_domain(cpu), &dest_cpu,
4235 tsk_cpus_allowed(task_of(se)));
4237 if (min_usage == 0) {
4238 trace_sched_hmp_offload_succeed(cpu, dest_cpu);
4241 trace_sched_hmp_offload_abort(cpu,min_usage,"slowdomain");
4244 #endif /* CONFIG_SCHED_HMP */
4247 * sched_balance_self: balance the current task (running on cpu) in domains
4248 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4251 * Balance, ie. select the least loaded group.
4253 * Returns the target CPU number, or the same CPU if no balancing is needed.
4255 * preempt must be disabled.
4258 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
4260 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4261 int cpu = smp_processor_id();
4262 int prev_cpu = task_cpu(p);
4264 int want_affine = 0;
4265 int sync = wake_flags & WF_SYNC;
4267 if (p->nr_cpus_allowed == 1)
4270 #ifdef CONFIG_SCHED_HMP
4271 /* always put non-kernel forking tasks on a big domain */
4272 if (p->mm && (sd_flag & SD_BALANCE_FORK)) {
4273 new_cpu = hmp_select_faster_cpu(p, prev_cpu);
4274 if (new_cpu != NR_CPUS) {
4275 hmp_next_up_delay(&p->se, new_cpu);
4278 /* failed to perform HMP fork balance, use normal balance */
4283 if (sd_flag & SD_BALANCE_WAKE) {
4284 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4290 for_each_domain(cpu, tmp) {
4291 if (!(tmp->flags & SD_LOAD_BALANCE))
4295 * If both cpu and prev_cpu are part of this domain,
4296 * cpu is a valid SD_WAKE_AFFINE target.
4298 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4299 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4304 if (tmp->flags & sd_flag)
4309 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4312 new_cpu = select_idle_sibling(p, prev_cpu);
4317 int load_idx = sd->forkexec_idx;
4318 struct sched_group *group;
4321 if (!(sd->flags & sd_flag)) {
4326 if (sd_flag & SD_BALANCE_WAKE)
4327 load_idx = sd->wake_idx;
4329 group = find_idlest_group(sd, p, cpu, load_idx);
4335 new_cpu = find_idlest_cpu(group, p, cpu);
4336 if (new_cpu == -1 || new_cpu == cpu) {
4337 /* Now try balancing at a lower domain level of cpu */
4342 /* Now try balancing at a lower domain level of new_cpu */
4344 weight = sd->span_weight;
4346 for_each_domain(cpu, tmp) {
4347 if (weight <= tmp->span_weight)
4349 if (tmp->flags & sd_flag)
4352 /* while loop will break here if sd == NULL */
4357 #ifdef CONFIG_SCHED_HMP
4358 prev_cpu = task_cpu(p);
4360 if (hmp_up_migration(prev_cpu, &new_cpu, &p->se)) {
4361 hmp_next_up_delay(&p->se, new_cpu);
4362 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4365 if (hmp_down_migration(prev_cpu, &p->se)) {
4366 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4367 new_cpu = hmp_best_little_cpu(p, prev_cpu);
4369 new_cpu = hmp_select_slower_cpu(p, prev_cpu);
4371 if (new_cpu != prev_cpu) {
4372 hmp_next_down_delay(&p->se, new_cpu);
4373 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4377 /* Make sure that the task stays in its previous hmp domain */
4378 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
4386 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
4387 * removed when useful for applications beyond shares distribution (e.g.
4390 #ifdef CONFIG_FAIR_GROUP_SCHED
4392 #ifdef CONFIG_NO_HZ_COMMON
4393 static int nohz_test_cpu(int cpu);
4395 static inline int nohz_test_cpu(int cpu)
4402 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4403 * cfs_rq_of(p) references at time of call are still valid and identify the
4404 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4405 * other assumptions, including the state of rq->lock, should be made.
4408 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4410 struct sched_entity *se = &p->se;
4411 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4414 * Load tracking: accumulate removed load so that it can be processed
4415 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4416 * to blocked load iff they have a positive decay-count. It can never
4417 * be negative here since on-rq tasks have decay-count == 0.
4419 if (se->avg.decay_count) {
4421 * If we migrate a sleeping task away from a CPU
4422 * which has the tick stopped, then both the clock_task
4423 * and decay_counter will be out of date for that CPU
4424 * and we will not decay load correctly.
4426 if (!se->on_rq && nohz_test_cpu(task_cpu(p))) {
4427 struct rq *rq = cpu_rq(task_cpu(p));
4428 unsigned long flags;
4430 * Current CPU cannot be holding rq->lock in this
4431 * circumstance, but another might be. We must hold
4432 * rq->lock before we go poking around in its clocks
4434 raw_spin_lock_irqsave(&rq->lock, flags);
4435 update_rq_clock(rq);
4436 update_cfs_rq_blocked_load(cfs_rq, 0);
4437 raw_spin_unlock_irqrestore(&rq->lock, flags);
4439 se->avg.decay_count = -__synchronize_entity_decay(se);
4440 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
4444 #endif /* CONFIG_SMP */
4446 static unsigned long
4447 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4449 unsigned long gran = sysctl_sched_wakeup_granularity;
4452 * Since its curr running now, convert the gran from real-time
4453 * to virtual-time in his units.
4455 * By using 'se' instead of 'curr' we penalize light tasks, so
4456 * they get preempted easier. That is, if 'se' < 'curr' then
4457 * the resulting gran will be larger, therefore penalizing the
4458 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4459 * be smaller, again penalizing the lighter task.
4461 * This is especially important for buddies when the leftmost
4462 * task is higher priority than the buddy.
4464 return calc_delta_fair(gran, se);
4468 * Should 'se' preempt 'curr'.
4482 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4484 s64 gran, vdiff = curr->vruntime - se->vruntime;
4489 gran = wakeup_gran(curr, se);
4496 static void set_last_buddy(struct sched_entity *se)
4498 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4501 for_each_sched_entity(se)
4502 cfs_rq_of(se)->last = se;
4505 static void set_next_buddy(struct sched_entity *se)
4507 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4510 for_each_sched_entity(se)
4511 cfs_rq_of(se)->next = se;
4514 static void set_skip_buddy(struct sched_entity *se)
4516 for_each_sched_entity(se)
4517 cfs_rq_of(se)->skip = se;
4521 * Preempt the current task with a newly woken task if needed:
4523 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4525 struct task_struct *curr = rq->curr;
4526 struct sched_entity *se = &curr->se, *pse = &p->se;
4527 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4528 int scale = cfs_rq->nr_running >= sched_nr_latency;
4529 int next_buddy_marked = 0;
4531 if (unlikely(se == pse))
4535 * This is possible from callers such as move_task(), in which we
4536 * unconditionally check_prempt_curr() after an enqueue (which may have
4537 * lead to a throttle). This both saves work and prevents false
4538 * next-buddy nomination below.
4540 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4543 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4544 set_next_buddy(pse);
4545 next_buddy_marked = 1;
4549 * We can come here with TIF_NEED_RESCHED already set from new task
4552 * Note: this also catches the edge-case of curr being in a throttled
4553 * group (e.g. via set_curr_task), since update_curr() (in the
4554 * enqueue of curr) will have resulted in resched being set. This
4555 * prevents us from potentially nominating it as a false LAST_BUDDY
4558 if (test_tsk_need_resched(curr))
4561 /* Idle tasks are by definition preempted by non-idle tasks. */
4562 if (unlikely(curr->policy == SCHED_IDLE) &&
4563 likely(p->policy != SCHED_IDLE))
4567 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4568 * is driven by the tick):
4570 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4573 find_matching_se(&se, &pse);
4574 update_curr(cfs_rq_of(se));
4576 if (wakeup_preempt_entity(se, pse) == 1) {
4578 * Bias pick_next to pick the sched entity that is
4579 * triggering this preemption.
4581 if (!next_buddy_marked)
4582 set_next_buddy(pse);
4591 * Only set the backward buddy when the current task is still
4592 * on the rq. This can happen when a wakeup gets interleaved
4593 * with schedule on the ->pre_schedule() or idle_balance()
4594 * point, either of which can * drop the rq lock.
4596 * Also, during early boot the idle thread is in the fair class,
4597 * for obvious reasons its a bad idea to schedule back to it.
4599 if (unlikely(!se->on_rq || curr == rq->idle))
4602 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4606 static struct task_struct *pick_next_task_fair(struct rq *rq)
4608 struct task_struct *p;
4609 struct cfs_rq *cfs_rq = &rq->cfs;
4610 struct sched_entity *se;
4612 if (!cfs_rq->nr_running)
4616 se = pick_next_entity(cfs_rq);
4617 set_next_entity(cfs_rq, se);
4618 cfs_rq = group_cfs_rq(se);
4622 if (hrtick_enabled(rq))
4623 hrtick_start_fair(rq, p);
4629 * Account for a descheduled task:
4631 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4633 struct sched_entity *se = &prev->se;
4634 struct cfs_rq *cfs_rq;
4636 for_each_sched_entity(se) {
4637 cfs_rq = cfs_rq_of(se);
4638 put_prev_entity(cfs_rq, se);
4643 * sched_yield() is very simple
4645 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4647 static void yield_task_fair(struct rq *rq)
4649 struct task_struct *curr = rq->curr;
4650 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4651 struct sched_entity *se = &curr->se;
4654 * Are we the only task in the tree?
4656 if (unlikely(rq->nr_running == 1))
4659 clear_buddies(cfs_rq, se);
4661 if (curr->policy != SCHED_BATCH) {
4662 update_rq_clock(rq);
4664 * Update run-time statistics of the 'current'.
4666 update_curr(cfs_rq);
4668 * Tell update_rq_clock() that we've just updated,
4669 * so we don't do microscopic update in schedule()
4670 * and double the fastpath cost.
4672 rq->skip_clock_update = 1;
4678 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4680 struct sched_entity *se = &p->se;
4682 /* throttled hierarchies are not runnable */
4683 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4686 /* Tell the scheduler that we'd really like pse to run next. */
4689 yield_task_fair(rq);
4695 /**************************************************
4696 * Fair scheduling class load-balancing methods.
4700 * The purpose of load-balancing is to achieve the same basic fairness the
4701 * per-cpu scheduler provides, namely provide a proportional amount of compute
4702 * time to each task. This is expressed in the following equation:
4704 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4706 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4707 * W_i,0 is defined as:
4709 * W_i,0 = \Sum_j w_i,j (2)
4711 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4712 * is derived from the nice value as per prio_to_weight[].
4714 * The weight average is an exponential decay average of the instantaneous
4717 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4719 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4720 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4721 * can also include other factors [XXX].
4723 * To achieve this balance we define a measure of imbalance which follows
4724 * directly from (1):
4726 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4728 * We them move tasks around to minimize the imbalance. In the continuous
4729 * function space it is obvious this converges, in the discrete case we get
4730 * a few fun cases generally called infeasible weight scenarios.
4733 * - infeasible weights;
4734 * - local vs global optima in the discrete case. ]
4739 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4740 * for all i,j solution, we create a tree of cpus that follows the hardware
4741 * topology where each level pairs two lower groups (or better). This results
4742 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4743 * tree to only the first of the previous level and we decrease the frequency
4744 * of load-balance at each level inv. proportional to the number of cpus in
4750 * \Sum { --- * --- * 2^i } = O(n) (5)
4752 * `- size of each group
4753 * | | `- number of cpus doing load-balance
4755 * `- sum over all levels
4757 * Coupled with a limit on how many tasks we can migrate every balance pass,
4758 * this makes (5) the runtime complexity of the balancer.
4760 * An important property here is that each CPU is still (indirectly) connected
4761 * to every other cpu in at most O(log n) steps:
4763 * The adjacency matrix of the resulting graph is given by:
4766 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4769 * And you'll find that:
4771 * A^(log_2 n)_i,j != 0 for all i,j (7)
4773 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4774 * The task movement gives a factor of O(m), giving a convergence complexity
4777 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4782 * In order to avoid CPUs going idle while there's still work to do, new idle
4783 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4784 * tree itself instead of relying on other CPUs to bring it work.
4786 * This adds some complexity to both (5) and (8) but it reduces the total idle
4794 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4797 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4802 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4804 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4806 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4809 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4810 * rewrite all of this once again.]
4813 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4815 #define LBF_ALL_PINNED 0x01
4816 #define LBF_NEED_BREAK 0x02
4817 #define LBF_SOME_PINNED 0x04
4820 struct sched_domain *sd;
4828 struct cpumask *dst_grpmask;
4830 enum cpu_idle_type idle;
4832 /* The set of CPUs under consideration for load-balancing */
4833 struct cpumask *cpus;
4838 unsigned int loop_break;
4839 unsigned int loop_max;
4843 * move_task - move a task from one runqueue to another runqueue.
4844 * Both runqueues must be locked.
4846 static void move_task(struct task_struct *p, struct lb_env *env)
4848 deactivate_task(env->src_rq, p, 0);
4849 set_task_cpu(p, env->dst_cpu);
4850 activate_task(env->dst_rq, p, 0);
4851 check_preempt_curr(env->dst_rq, p, 0);
4855 * Is this task likely cache-hot:
4858 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4862 if (p->sched_class != &fair_sched_class)
4865 if (unlikely(p->policy == SCHED_IDLE))
4869 * Buddy candidates are cache hot:
4871 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4872 (&p->se == cfs_rq_of(&p->se)->next ||
4873 &p->se == cfs_rq_of(&p->se)->last))
4876 if (sysctl_sched_migration_cost == -1)
4878 if (sysctl_sched_migration_cost == 0)
4881 delta = now - p->se.exec_start;
4883 return delta < (s64)sysctl_sched_migration_cost;
4887 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4890 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4892 int tsk_cache_hot = 0;
4894 * We do not migrate tasks that are:
4895 * 1) throttled_lb_pair, or
4896 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4897 * 3) running (obviously), or
4898 * 4) are cache-hot on their current CPU.
4900 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4903 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4906 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4909 * Remember if this task can be migrated to any other cpu in
4910 * our sched_group. We may want to revisit it if we couldn't
4911 * meet load balance goals by pulling other tasks on src_cpu.
4913 * Also avoid computing new_dst_cpu if we have already computed
4914 * one in current iteration.
4916 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
4919 /* Prevent to re-select dst_cpu via env's cpus */
4920 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4921 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4922 env->flags |= LBF_SOME_PINNED;
4923 env->new_dst_cpu = cpu;
4931 /* Record that we found atleast one task that could run on dst_cpu */
4932 env->flags &= ~LBF_ALL_PINNED;
4934 if (task_running(env->src_rq, p)) {
4935 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4940 * Aggressive migration if:
4941 * 1) task is cache cold, or
4942 * 2) too many balance attempts have failed.
4944 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
4945 if (!tsk_cache_hot ||
4946 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4948 if (tsk_cache_hot) {
4949 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4950 schedstat_inc(p, se.statistics.nr_forced_migrations);
4956 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4961 * move_one_task tries to move exactly one task from busiest to this_rq, as
4962 * part of active balancing operations within "domain".
4963 * Returns 1 if successful and 0 otherwise.
4965 * Called with both runqueues locked.
4967 static int move_one_task(struct lb_env *env)
4969 struct task_struct *p, *n;
4971 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4972 if (!can_migrate_task(p, env))
4977 * Right now, this is only the second place move_task()
4978 * is called, so we can safely collect move_task()
4979 * stats here rather than inside move_task().
4981 schedstat_inc(env->sd, lb_gained[env->idle]);
4987 static unsigned long task_h_load(struct task_struct *p);
4989 static const unsigned int sched_nr_migrate_break = 32;
4992 * move_tasks tries to move up to imbalance weighted load from busiest to
4993 * this_rq, as part of a balancing operation within domain "sd".
4994 * Returns 1 if successful and 0 otherwise.
4996 * Called with both runqueues locked.
4998 static int move_tasks(struct lb_env *env)
5000 struct list_head *tasks = &env->src_rq->cfs_tasks;
5001 struct task_struct *p;
5005 if (env->imbalance <= 0)
5008 while (!list_empty(tasks)) {
5009 p = list_first_entry(tasks, struct task_struct, se.group_node);
5012 /* We've more or less seen every task there is, call it quits */
5013 if (env->loop > env->loop_max)
5016 /* take a breather every nr_migrate tasks */
5017 if (env->loop > env->loop_break) {
5018 env->loop_break += sched_nr_migrate_break;
5019 env->flags |= LBF_NEED_BREAK;
5023 if (!can_migrate_task(p, env))
5026 load = task_h_load(p);
5028 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5031 if ((load / 2) > env->imbalance)
5036 env->imbalance -= load;
5038 #ifdef CONFIG_PREEMPT
5040 * NEWIDLE balancing is a source of latency, so preemptible
5041 * kernels will stop after the first task is pulled to minimize
5042 * the critical section.
5044 if (env->idle == CPU_NEWLY_IDLE)
5049 * We only want to steal up to the prescribed amount of
5052 if (env->imbalance <= 0)
5057 list_move_tail(&p->se.group_node, tasks);
5061 * Right now, this is one of only two places move_task() is called,
5062 * so we can safely collect move_task() stats here rather than
5063 * inside move_task().
5065 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5070 #ifdef CONFIG_FAIR_GROUP_SCHED
5072 * update tg->load_weight by folding this cpu's load_avg
5074 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5076 struct sched_entity *se = tg->se[cpu];
5077 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5079 /* throttled entities do not contribute to load */
5080 if (throttled_hierarchy(cfs_rq))
5083 update_cfs_rq_blocked_load(cfs_rq, 1);
5086 update_entity_load_avg(se, 1);
5088 * We pivot on our runnable average having decayed to zero for
5089 * list removal. This generally implies that all our children
5090 * have also been removed (modulo rounding error or bandwidth
5091 * control); however, such cases are rare and we can fix these
5094 * TODO: fix up out-of-order children on enqueue.
5096 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5097 list_del_leaf_cfs_rq(cfs_rq);
5099 struct rq *rq = rq_of(cfs_rq);
5100 update_rq_runnable_avg(rq, rq->nr_running);
5104 static void update_blocked_averages(int cpu)
5106 struct rq *rq = cpu_rq(cpu);
5107 struct cfs_rq *cfs_rq;
5108 unsigned long flags;
5110 raw_spin_lock_irqsave(&rq->lock, flags);
5111 update_rq_clock(rq);
5113 * Iterates the task_group tree in a bottom up fashion, see
5114 * list_add_leaf_cfs_rq() for details.
5116 for_each_leaf_cfs_rq(rq, cfs_rq) {
5118 * Note: We may want to consider periodically releasing
5119 * rq->lock about these updates so that creating many task
5120 * groups does not result in continually extending hold time.
5122 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5125 raw_spin_unlock_irqrestore(&rq->lock, flags);
5129 * Compute the cpu's hierarchical load factor for each task group.
5130 * This needs to be done in a top-down fashion because the load of a child
5131 * group is a fraction of its parents load.
5133 static int tg_load_down(struct task_group *tg, void *data)
5136 long cpu = (long)data;
5139 load = cpu_rq(cpu)->load.weight;
5141 load = tg->parent->cfs_rq[cpu]->h_load;
5142 load *= tg->se[cpu]->load.weight;
5143 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
5146 tg->cfs_rq[cpu]->h_load = load;
5151 static void update_h_load(long cpu)
5153 struct rq *rq = cpu_rq(cpu);
5154 unsigned long now = jiffies;
5156 if (rq->h_load_throttle == now)
5159 rq->h_load_throttle = now;
5162 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
5166 static unsigned long task_h_load(struct task_struct *p)
5168 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5171 load = p->se.load.weight;
5172 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
5177 static inline void update_blocked_averages(int cpu)
5181 static inline void update_h_load(long cpu)
5185 static unsigned long task_h_load(struct task_struct *p)
5187 return p->se.load.weight;
5191 /********** Helpers for find_busiest_group ************************/
5193 * sd_lb_stats - Structure to store the statistics of a sched_domain
5194 * during load balancing.
5196 struct sd_lb_stats {
5197 struct sched_group *busiest; /* Busiest group in this sd */
5198 struct sched_group *this; /* Local group in this sd */
5199 unsigned long total_load; /* Total load of all groups in sd */
5200 unsigned long total_pwr; /* Total power of all groups in sd */
5201 unsigned long avg_load; /* Average load across all groups in sd */
5203 /** Statistics of this group */
5204 unsigned long this_load;
5205 unsigned long this_load_per_task;
5206 unsigned long this_nr_running;
5207 unsigned long this_has_capacity;
5208 unsigned int this_idle_cpus;
5210 /* Statistics of the busiest group */
5211 unsigned int busiest_idle_cpus;
5212 unsigned long max_load;
5213 unsigned long busiest_load_per_task;
5214 unsigned long busiest_nr_running;
5215 unsigned long busiest_group_capacity;
5216 unsigned long busiest_has_capacity;
5217 unsigned int busiest_group_weight;
5219 int group_imb; /* Is there imbalance in this sd */
5223 * sg_lb_stats - stats of a sched_group required for load_balancing
5225 struct sg_lb_stats {
5226 unsigned long avg_load; /*Avg load across the CPUs of the group */
5227 unsigned long group_load; /* Total load over the CPUs of the group */
5228 unsigned long sum_nr_running; /* Nr tasks running in the group */
5229 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5230 unsigned long group_capacity;
5231 unsigned long idle_cpus;
5232 unsigned long group_weight;
5233 int group_imb; /* Is there an imbalance in the group ? */
5234 int group_has_capacity; /* Is there extra capacity in the group? */
5238 * get_sd_load_idx - Obtain the load index for a given sched domain.
5239 * @sd: The sched_domain whose load_idx is to be obtained.
5240 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5242 static inline int get_sd_load_idx(struct sched_domain *sd,
5243 enum cpu_idle_type idle)
5249 load_idx = sd->busy_idx;
5252 case CPU_NEWLY_IDLE:
5253 load_idx = sd->newidle_idx;
5256 load_idx = sd->idle_idx;
5263 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5265 return SCHED_POWER_SCALE;
5268 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5270 return default_scale_freq_power(sd, cpu);
5273 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5275 unsigned long weight = sd->span_weight;
5276 unsigned long smt_gain = sd->smt_gain;
5283 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5285 return default_scale_smt_power(sd, cpu);
5288 static unsigned long scale_rt_power(int cpu)
5290 struct rq *rq = cpu_rq(cpu);
5291 u64 total, available, age_stamp, avg;
5294 * Since we're reading these variables without serialization make sure
5295 * we read them once before doing sanity checks on them.
5297 age_stamp = ACCESS_ONCE(rq->age_stamp);
5298 avg = ACCESS_ONCE(rq->rt_avg);
5300 total = sched_avg_period() + (rq->clock - age_stamp);
5302 if (unlikely(total < avg)) {
5303 /* Ensures that power won't end up being negative */
5306 available = total - avg;
5309 if (unlikely((s64)total < SCHED_POWER_SCALE))
5310 total = SCHED_POWER_SCALE;
5312 total >>= SCHED_POWER_SHIFT;
5314 return div_u64(available, total);
5317 static void update_cpu_power(struct sched_domain *sd, int cpu)
5319 unsigned long weight = sd->span_weight;
5320 unsigned long power = SCHED_POWER_SCALE;
5321 struct sched_group *sdg = sd->groups;
5323 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5324 if (sched_feat(ARCH_POWER))
5325 power *= arch_scale_smt_power(sd, cpu);
5327 power *= default_scale_smt_power(sd, cpu);
5329 power >>= SCHED_POWER_SHIFT;
5332 sdg->sgp->power_orig = power;
5334 if (sched_feat(ARCH_POWER))
5335 power *= arch_scale_freq_power(sd, cpu);
5337 power *= default_scale_freq_power(sd, cpu);
5339 power >>= SCHED_POWER_SHIFT;
5341 power *= scale_rt_power(cpu);
5342 power >>= SCHED_POWER_SHIFT;
5347 cpu_rq(cpu)->cpu_power = power;
5348 sdg->sgp->power = power;
5351 void update_group_power(struct sched_domain *sd, int cpu)
5353 struct sched_domain *child = sd->child;
5354 struct sched_group *group, *sdg = sd->groups;
5355 unsigned long power;
5356 unsigned long interval;
5358 interval = msecs_to_jiffies(sd->balance_interval);
5359 interval = clamp(interval, 1UL, max_load_balance_interval);
5360 sdg->sgp->next_update = jiffies + interval;
5363 update_cpu_power(sd, cpu);
5369 if (child->flags & SD_OVERLAP) {
5371 * SD_OVERLAP domains cannot assume that child groups
5372 * span the current group.
5375 for_each_cpu(cpu, sched_group_cpus(sdg))
5376 power += power_of(cpu);
5379 * !SD_OVERLAP domains can assume that child groups
5380 * span the current group.
5383 group = child->groups;
5385 power += group->sgp->power;
5386 group = group->next;
5387 } while (group != child->groups);
5390 sdg->sgp->power_orig = sdg->sgp->power = power;
5394 * Try and fix up capacity for tiny siblings, this is needed when
5395 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5396 * which on its own isn't powerful enough.
5398 * See update_sd_pick_busiest() and check_asym_packing().
5401 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5404 * Only siblings can have significantly less than SCHED_POWER_SCALE
5406 if (!(sd->flags & SD_SHARE_CPUPOWER))
5410 * If ~90% of the cpu_power is still there, we're good.
5412 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5419 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5420 * @env: The load balancing environment.
5421 * @group: sched_group whose statistics are to be updated.
5422 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5423 * @local_group: Does group contain this_cpu.
5424 * @balance: Should we balance.
5425 * @sgs: variable to hold the statistics for this group.
5427 static inline void update_sg_lb_stats(struct lb_env *env,
5428 struct sched_group *group, int load_idx,
5429 int local_group, int *balance, struct sg_lb_stats *sgs)
5431 unsigned long nr_running, max_nr_running, min_nr_running;
5432 unsigned long load, max_cpu_load, min_cpu_load;
5433 unsigned int balance_cpu = -1, first_idle_cpu = 0;
5434 unsigned long avg_load_per_task = 0;
5438 balance_cpu = group_balance_cpu(group);
5440 /* Tally up the load of all CPUs in the group */
5442 min_cpu_load = ~0UL;
5444 min_nr_running = ~0UL;
5446 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5447 struct rq *rq = cpu_rq(i);
5449 nr_running = rq->nr_running;
5451 /* Bias balancing toward cpus of our domain */
5453 if (idle_cpu(i) && !first_idle_cpu &&
5454 cpumask_test_cpu(i, sched_group_mask(group))) {
5459 load = target_load(i, load_idx);
5461 load = source_load(i, load_idx);
5462 if (load > max_cpu_load)
5463 max_cpu_load = load;
5464 if (min_cpu_load > load)
5465 min_cpu_load = load;
5467 if (nr_running > max_nr_running)
5468 max_nr_running = nr_running;
5469 if (min_nr_running > nr_running)
5470 min_nr_running = nr_running;
5473 sgs->group_load += load;
5474 sgs->sum_nr_running += nr_running;
5475 sgs->sum_weighted_load += weighted_cpuload(i);
5481 * First idle cpu or the first cpu(busiest) in this sched group
5482 * is eligible for doing load balancing at this and above
5483 * domains. In the newly idle case, we will allow all the cpu's
5484 * to do the newly idle load balance.
5487 if (env->idle != CPU_NEWLY_IDLE) {
5488 if (balance_cpu != env->dst_cpu) {
5492 update_group_power(env->sd, env->dst_cpu);
5493 } else if (time_after_eq(jiffies, group->sgp->next_update))
5494 update_group_power(env->sd, env->dst_cpu);
5497 /* Adjust by relative CPU power of the group */
5498 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
5501 * Consider the group unbalanced when the imbalance is larger
5502 * than the average weight of a task.
5504 * APZ: with cgroup the avg task weight can vary wildly and
5505 * might not be a suitable number - should we keep a
5506 * normalized nr_running number somewhere that negates
5509 if (sgs->sum_nr_running)
5510 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5512 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
5513 (max_nr_running - min_nr_running) > 1)
5516 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
5518 if (!sgs->group_capacity)
5519 sgs->group_capacity = fix_small_capacity(env->sd, group);
5520 sgs->group_weight = group->group_weight;
5522 if (sgs->group_capacity > sgs->sum_nr_running)
5523 sgs->group_has_capacity = 1;
5527 * update_sd_pick_busiest - return 1 on busiest group
5528 * @env: The load balancing environment.
5529 * @sds: sched_domain statistics
5530 * @sg: sched_group candidate to be checked for being the busiest
5531 * @sgs: sched_group statistics
5533 * Determine if @sg is a busier group than the previously selected
5536 static bool update_sd_pick_busiest(struct lb_env *env,
5537 struct sd_lb_stats *sds,
5538 struct sched_group *sg,
5539 struct sg_lb_stats *sgs)
5541 if (sgs->avg_load <= sds->max_load)
5544 if (sgs->sum_nr_running > sgs->group_capacity)
5551 * ASYM_PACKING needs to move all the work to the lowest
5552 * numbered CPUs in the group, therefore mark all groups
5553 * higher than ourself as busy.
5555 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5556 env->dst_cpu < group_first_cpu(sg)) {
5560 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5568 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5569 * @env: The load balancing environment.
5570 * @balance: Should we balance.
5571 * @sds: variable to hold the statistics for this sched_domain.
5573 static inline void update_sd_lb_stats(struct lb_env *env,
5574 int *balance, struct sd_lb_stats *sds)
5576 struct sched_domain *child = env->sd->child;
5577 struct sched_group *sg = env->sd->groups;
5578 struct sg_lb_stats sgs;
5579 int load_idx, prefer_sibling = 0;
5581 if (child && child->flags & SD_PREFER_SIBLING)
5584 load_idx = get_sd_load_idx(env->sd, env->idle);
5589 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5590 memset(&sgs, 0, sizeof(sgs));
5591 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
5593 if (local_group && !(*balance))
5596 sds->total_load += sgs.group_load;
5597 sds->total_pwr += sg->sgp->power;
5600 * In case the child domain prefers tasks go to siblings
5601 * first, lower the sg capacity to one so that we'll try
5602 * and move all the excess tasks away. We lower the capacity
5603 * of a group only if the local group has the capacity to fit
5604 * these excess tasks, i.e. nr_running < group_capacity. The
5605 * extra check prevents the case where you always pull from the
5606 * heaviest group when it is already under-utilized (possible
5607 * with a large weight task outweighs the tasks on the system).
5609 if (prefer_sibling && !local_group && sds->this_has_capacity)
5610 sgs.group_capacity = min(sgs.group_capacity, 1UL);
5613 sds->this_load = sgs.avg_load;
5615 sds->this_nr_running = sgs.sum_nr_running;
5616 sds->this_load_per_task = sgs.sum_weighted_load;
5617 sds->this_has_capacity = sgs.group_has_capacity;
5618 sds->this_idle_cpus = sgs.idle_cpus;
5619 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
5620 sds->max_load = sgs.avg_load;
5622 sds->busiest_nr_running = sgs.sum_nr_running;
5623 sds->busiest_idle_cpus = sgs.idle_cpus;
5624 sds->busiest_group_capacity = sgs.group_capacity;
5625 sds->busiest_load_per_task = sgs.sum_weighted_load;
5626 sds->busiest_has_capacity = sgs.group_has_capacity;
5627 sds->busiest_group_weight = sgs.group_weight;
5628 sds->group_imb = sgs.group_imb;
5632 } while (sg != env->sd->groups);
5636 * check_asym_packing - Check to see if the group is packed into the
5639 * This is primarily intended to used at the sibling level. Some
5640 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5641 * case of POWER7, it can move to lower SMT modes only when higher
5642 * threads are idle. When in lower SMT modes, the threads will
5643 * perform better since they share less core resources. Hence when we
5644 * have idle threads, we want them to be the higher ones.
5646 * This packing function is run on idle threads. It checks to see if
5647 * the busiest CPU in this domain (core in the P7 case) has a higher
5648 * CPU number than the packing function is being run on. Here we are
5649 * assuming lower CPU number will be equivalent to lower a SMT thread
5652 * Returns 1 when packing is required and a task should be moved to
5653 * this CPU. The amount of the imbalance is returned in *imbalance.
5655 * @env: The load balancing environment.
5656 * @sds: Statistics of the sched_domain which is to be packed
5658 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5662 if (!(env->sd->flags & SD_ASYM_PACKING))
5668 busiest_cpu = group_first_cpu(sds->busiest);
5669 if (env->dst_cpu > busiest_cpu)
5672 env->imbalance = DIV_ROUND_CLOSEST(
5673 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
5679 * fix_small_imbalance - Calculate the minor imbalance that exists
5680 * amongst the groups of a sched_domain, during
5682 * @env: The load balancing environment.
5683 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5686 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5688 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5689 unsigned int imbn = 2;
5690 unsigned long scaled_busy_load_per_task;
5692 if (sds->this_nr_running) {
5693 sds->this_load_per_task /= sds->this_nr_running;
5694 if (sds->busiest_load_per_task >
5695 sds->this_load_per_task)
5698 sds->this_load_per_task =
5699 cpu_avg_load_per_task(env->dst_cpu);
5702 scaled_busy_load_per_task = sds->busiest_load_per_task
5703 * SCHED_POWER_SCALE;
5704 scaled_busy_load_per_task /= sds->busiest->sgp->power;
5706 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
5707 (scaled_busy_load_per_task * imbn)) {
5708 env->imbalance = sds->busiest_load_per_task;
5713 * OK, we don't have enough imbalance to justify moving tasks,
5714 * however we may be able to increase total CPU power used by
5718 pwr_now += sds->busiest->sgp->power *
5719 min(sds->busiest_load_per_task, sds->max_load);
5720 pwr_now += sds->this->sgp->power *
5721 min(sds->this_load_per_task, sds->this_load);
5722 pwr_now /= SCHED_POWER_SCALE;
5724 /* Amount of load we'd subtract */
5725 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5726 sds->busiest->sgp->power;
5727 if (sds->max_load > tmp)
5728 pwr_move += sds->busiest->sgp->power *
5729 min(sds->busiest_load_per_task, sds->max_load - tmp);
5731 /* Amount of load we'd add */
5732 if (sds->max_load * sds->busiest->sgp->power <
5733 sds->busiest_load_per_task * SCHED_POWER_SCALE)
5734 tmp = (sds->max_load * sds->busiest->sgp->power) /
5735 sds->this->sgp->power;
5737 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5738 sds->this->sgp->power;
5739 pwr_move += sds->this->sgp->power *
5740 min(sds->this_load_per_task, sds->this_load + tmp);
5741 pwr_move /= SCHED_POWER_SCALE;
5743 /* Move if we gain throughput */
5744 if (pwr_move > pwr_now)
5745 env->imbalance = sds->busiest_load_per_task;
5749 * calculate_imbalance - Calculate the amount of imbalance present within the
5750 * groups of a given sched_domain during load balance.
5751 * @env: load balance environment
5752 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5754 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5756 unsigned long max_pull, load_above_capacity = ~0UL;
5758 sds->busiest_load_per_task /= sds->busiest_nr_running;
5759 if (sds->group_imb) {
5760 sds->busiest_load_per_task =
5761 min(sds->busiest_load_per_task, sds->avg_load);
5765 * In the presence of smp nice balancing, certain scenarios can have
5766 * max load less than avg load(as we skip the groups at or below
5767 * its cpu_power, while calculating max_load..)
5769 if (sds->max_load < sds->avg_load) {
5771 return fix_small_imbalance(env, sds);
5774 if (!sds->group_imb) {
5776 * Don't want to pull so many tasks that a group would go idle.
5778 load_above_capacity = (sds->busiest_nr_running -
5779 sds->busiest_group_capacity);
5781 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5783 load_above_capacity /= sds->busiest->sgp->power;
5787 * We're trying to get all the cpus to the average_load, so we don't
5788 * want to push ourselves above the average load, nor do we wish to
5789 * reduce the max loaded cpu below the average load. At the same time,
5790 * we also don't want to reduce the group load below the group capacity
5791 * (so that we can implement power-savings policies etc). Thus we look
5792 * for the minimum possible imbalance.
5793 * Be careful of negative numbers as they'll appear as very large values
5794 * with unsigned longs.
5796 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
5798 /* How much load to actually move to equalise the imbalance */
5799 env->imbalance = min(max_pull * sds->busiest->sgp->power,
5800 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
5801 / SCHED_POWER_SCALE;
5804 * if *imbalance is less than the average load per runnable task
5805 * there is no guarantee that any tasks will be moved so we'll have
5806 * a think about bumping its value to force at least one task to be
5809 if (env->imbalance < sds->busiest_load_per_task)
5810 return fix_small_imbalance(env, sds);
5814 /******* find_busiest_group() helpers end here *********************/
5817 * find_busiest_group - Returns the busiest group within the sched_domain
5818 * if there is an imbalance. If there isn't an imbalance, and
5819 * the user has opted for power-savings, it returns a group whose
5820 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5821 * such a group exists.
5823 * Also calculates the amount of weighted load which should be moved
5824 * to restore balance.
5826 * @env: The load balancing environment.
5827 * @balance: Pointer to a variable indicating if this_cpu
5828 * is the appropriate cpu to perform load balancing at this_level.
5830 * Returns: - the busiest group if imbalance exists.
5831 * - If no imbalance and user has opted for power-savings balance,
5832 * return the least loaded group whose CPUs can be
5833 * put to idle by rebalancing its tasks onto our group.
5835 static struct sched_group *
5836 find_busiest_group(struct lb_env *env, int *balance)
5838 struct sd_lb_stats sds;
5840 memset(&sds, 0, sizeof(sds));
5843 * Compute the various statistics relavent for load balancing at
5846 update_sd_lb_stats(env, balance, &sds);
5849 * this_cpu is not the appropriate cpu to perform load balancing at
5855 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5856 check_asym_packing(env, &sds))
5859 /* There is no busy sibling group to pull tasks from */
5860 if (!sds.busiest || sds.busiest_nr_running == 0)
5863 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5866 * If the busiest group is imbalanced the below checks don't
5867 * work because they assumes all things are equal, which typically
5868 * isn't true due to cpus_allowed constraints and the like.
5873 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5874 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5875 !sds.busiest_has_capacity)
5879 * If the local group is more busy than the selected busiest group
5880 * don't try and pull any tasks.
5882 if (sds.this_load >= sds.max_load)
5886 * Don't pull any tasks if this group is already above the domain
5889 if (sds.this_load >= sds.avg_load)
5892 if (env->idle == CPU_IDLE) {
5894 * This cpu is idle. If the busiest group load doesn't
5895 * have more tasks than the number of available cpu's and
5896 * there is no imbalance between this and busiest group
5897 * wrt to idle cpu's, it is balanced.
5899 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
5900 sds.busiest_nr_running <= sds.busiest_group_weight)
5904 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5905 * imbalance_pct to be conservative.
5907 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
5912 /* Looks like there is an imbalance. Compute it */
5913 calculate_imbalance(env, &sds);
5923 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5925 static struct rq *find_busiest_queue(struct lb_env *env,
5926 struct sched_group *group)
5928 struct rq *busiest = NULL, *rq;
5929 unsigned long max_load = 0;
5932 for_each_cpu(i, sched_group_cpus(group)) {
5933 unsigned long power = power_of(i);
5934 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5939 capacity = fix_small_capacity(env->sd, group);
5941 if (!cpumask_test_cpu(i, env->cpus))
5945 wl = weighted_cpuload(i);
5948 * When comparing with imbalance, use weighted_cpuload()
5949 * which is not scaled with the cpu power.
5951 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5955 * For the load comparisons with the other cpu's, consider
5956 * the weighted_cpuload() scaled with the cpu power, so that
5957 * the load can be moved away from the cpu that is potentially
5958 * running at a lower capacity.
5960 wl = (wl * SCHED_POWER_SCALE) / power;
5962 if (wl > max_load) {
5972 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5973 * so long as it is large enough.
5975 #define MAX_PINNED_INTERVAL 512
5977 /* Working cpumask for load_balance and load_balance_newidle. */
5978 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5980 static int need_active_balance(struct lb_env *env)
5982 struct sched_domain *sd = env->sd;
5984 if (env->idle == CPU_NEWLY_IDLE) {
5987 * ASYM_PACKING needs to force migrate tasks from busy but
5988 * higher numbered CPUs in order to pack all tasks in the
5989 * lowest numbered CPUs.
5991 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5995 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5998 static int active_load_balance_cpu_stop(void *data);
6001 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6002 * tasks if there is an imbalance.
6004 static int load_balance(int this_cpu, struct rq *this_rq,
6005 struct sched_domain *sd, enum cpu_idle_type idle,
6008 int ld_moved, cur_ld_moved, active_balance = 0;
6009 struct sched_group *group;
6011 unsigned long flags;
6012 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6014 struct lb_env env = {
6016 .dst_cpu = this_cpu,
6018 .dst_grpmask = sched_group_cpus(sd->groups),
6020 .loop_break = sched_nr_migrate_break,
6025 * For NEWLY_IDLE load_balancing, we don't need to consider
6026 * other cpus in our group
6028 if (idle == CPU_NEWLY_IDLE)
6029 env.dst_grpmask = NULL;
6031 cpumask_copy(cpus, cpu_active_mask);
6033 schedstat_inc(sd, lb_count[idle]);
6036 group = find_busiest_group(&env, balance);
6042 schedstat_inc(sd, lb_nobusyg[idle]);
6046 busiest = find_busiest_queue(&env, group);
6048 schedstat_inc(sd, lb_nobusyq[idle]);
6052 BUG_ON(busiest == env.dst_rq);
6054 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6057 if (busiest->nr_running > 1) {
6059 * Attempt to move tasks. If find_busiest_group has found
6060 * an imbalance but busiest->nr_running <= 1, the group is
6061 * still unbalanced. ld_moved simply stays zero, so it is
6062 * correctly treated as an imbalance.
6064 env.flags |= LBF_ALL_PINNED;
6065 env.src_cpu = busiest->cpu;
6066 env.src_rq = busiest;
6067 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6069 update_h_load(env.src_cpu);
6071 local_irq_save(flags);
6072 double_rq_lock(env.dst_rq, busiest);
6075 * cur_ld_moved - load moved in current iteration
6076 * ld_moved - cumulative load moved across iterations
6078 cur_ld_moved = move_tasks(&env);
6079 ld_moved += cur_ld_moved;
6080 double_rq_unlock(env.dst_rq, busiest);
6081 local_irq_restore(flags);
6084 * some other cpu did the load balance for us.
6086 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6087 resched_cpu(env.dst_cpu);
6089 if (env.flags & LBF_NEED_BREAK) {
6090 env.flags &= ~LBF_NEED_BREAK;
6095 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6096 * us and move them to an alternate dst_cpu in our sched_group
6097 * where they can run. The upper limit on how many times we
6098 * iterate on same src_cpu is dependent on number of cpus in our
6101 * This changes load balance semantics a bit on who can move
6102 * load to a given_cpu. In addition to the given_cpu itself
6103 * (or a ilb_cpu acting on its behalf where given_cpu is
6104 * nohz-idle), we now have balance_cpu in a position to move
6105 * load to given_cpu. In rare situations, this may cause
6106 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6107 * _independently_ and at _same_ time to move some load to
6108 * given_cpu) causing exceess load to be moved to given_cpu.
6109 * This however should not happen so much in practice and
6110 * moreover subsequent load balance cycles should correct the
6111 * excess load moved.
6113 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6115 env.dst_rq = cpu_rq(env.new_dst_cpu);
6116 env.dst_cpu = env.new_dst_cpu;
6117 env.flags &= ~LBF_SOME_PINNED;
6119 env.loop_break = sched_nr_migrate_break;
6121 /* Prevent to re-select dst_cpu via env's cpus */
6122 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6125 * Go back to "more_balance" rather than "redo" since we
6126 * need to continue with same src_cpu.
6131 /* All tasks on this runqueue were pinned by CPU affinity */
6132 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6133 cpumask_clear_cpu(cpu_of(busiest), cpus);
6134 if (!cpumask_empty(cpus)) {
6136 env.loop_break = sched_nr_migrate_break;
6144 schedstat_inc(sd, lb_failed[idle]);
6146 * Increment the failure counter only on periodic balance.
6147 * We do not want newidle balance, which can be very
6148 * frequent, pollute the failure counter causing
6149 * excessive cache_hot migrations and active balances.
6151 if (idle != CPU_NEWLY_IDLE)
6152 sd->nr_balance_failed++;
6154 if (need_active_balance(&env)) {
6155 raw_spin_lock_irqsave(&busiest->lock, flags);
6157 /* don't kick the active_load_balance_cpu_stop,
6158 * if the curr task on busiest cpu can't be
6161 if (!cpumask_test_cpu(this_cpu,
6162 tsk_cpus_allowed(busiest->curr))) {
6163 raw_spin_unlock_irqrestore(&busiest->lock,
6165 env.flags |= LBF_ALL_PINNED;
6166 goto out_one_pinned;
6170 * ->active_balance synchronizes accesses to
6171 * ->active_balance_work. Once set, it's cleared
6172 * only after active load balance is finished.
6174 if (!busiest->active_balance) {
6175 busiest->active_balance = 1;
6176 busiest->push_cpu = this_cpu;
6179 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6181 if (active_balance) {
6182 stop_one_cpu_nowait(cpu_of(busiest),
6183 active_load_balance_cpu_stop, busiest,
6184 &busiest->active_balance_work);
6188 * We've kicked active balancing, reset the failure
6191 sd->nr_balance_failed = sd->cache_nice_tries+1;
6194 sd->nr_balance_failed = 0;
6196 if (likely(!active_balance)) {
6197 /* We were unbalanced, so reset the balancing interval */
6198 sd->balance_interval = sd->min_interval;
6201 * If we've begun active balancing, start to back off. This
6202 * case may not be covered by the all_pinned logic if there
6203 * is only 1 task on the busy runqueue (because we don't call
6206 if (sd->balance_interval < sd->max_interval)
6207 sd->balance_interval *= 2;
6213 schedstat_inc(sd, lb_balanced[idle]);
6215 sd->nr_balance_failed = 0;
6218 /* tune up the balancing interval */
6219 if (((env.flags & LBF_ALL_PINNED) &&
6220 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6221 (sd->balance_interval < sd->max_interval))
6222 sd->balance_interval *= 2;
6228 #ifdef CONFIG_SCHED_HMP
6229 static unsigned int hmp_idle_pull(int this_cpu);
6232 * idle_balance is called by schedule() if this_cpu is about to become
6233 * idle. Attempts to pull tasks from other CPUs.
6235 void idle_balance(int this_cpu, struct rq *this_rq)
6237 struct sched_domain *sd;
6238 int pulled_task = 0;
6239 unsigned long next_balance = jiffies + HZ;
6241 this_rq->idle_stamp = this_rq->clock;
6243 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6247 * Drop the rq->lock, but keep IRQ/preempt disabled.
6249 raw_spin_unlock(&this_rq->lock);
6251 update_blocked_averages(this_cpu);
6253 for_each_domain(this_cpu, sd) {
6254 unsigned long interval;
6257 if (!(sd->flags & SD_LOAD_BALANCE))
6260 if (sd->flags & SD_BALANCE_NEWIDLE) {
6261 /* If we've pulled tasks over stop searching: */
6262 pulled_task = load_balance(this_cpu, this_rq,
6263 sd, CPU_NEWLY_IDLE, &balance);
6266 interval = msecs_to_jiffies(sd->balance_interval);
6267 if (time_after(next_balance, sd->last_balance + interval))
6268 next_balance = sd->last_balance + interval;
6270 this_rq->idle_stamp = 0;
6275 #ifdef CONFIG_SCHED_HMP
6277 pulled_task = hmp_idle_pull(this_cpu);
6279 raw_spin_lock(&this_rq->lock);
6281 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6283 * We are going idle. next_balance may be set based on
6284 * a busy processor. So reset next_balance.
6286 this_rq->next_balance = next_balance;
6291 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6292 * running tasks off the busiest CPU onto idle CPUs. It requires at
6293 * least 1 task to be running on each physical CPU where possible, and
6294 * avoids physical / logical imbalances.
6296 static int active_load_balance_cpu_stop(void *data)
6298 struct rq *busiest_rq = data;
6299 int busiest_cpu = cpu_of(busiest_rq);
6300 int target_cpu = busiest_rq->push_cpu;
6301 struct rq *target_rq = cpu_rq(target_cpu);
6302 struct sched_domain *sd;
6304 raw_spin_lock_irq(&busiest_rq->lock);
6306 /* make sure the requested cpu hasn't gone down in the meantime */
6307 if (unlikely(busiest_cpu != smp_processor_id() ||
6308 !busiest_rq->active_balance))
6311 /* Is there any task to move? */
6312 if (busiest_rq->nr_running <= 1)
6316 * This condition is "impossible", if it occurs
6317 * we need to fix it. Originally reported by
6318 * Bjorn Helgaas on a 128-cpu setup.
6320 BUG_ON(busiest_rq == target_rq);
6322 /* move a task from busiest_rq to target_rq */
6323 double_lock_balance(busiest_rq, target_rq);
6325 /* Search for an sd spanning us and the target CPU. */
6327 for_each_domain(target_cpu, sd) {
6328 if ((sd->flags & SD_LOAD_BALANCE) &&
6329 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6334 struct lb_env env = {
6336 .dst_cpu = target_cpu,
6337 .dst_rq = target_rq,
6338 .src_cpu = busiest_rq->cpu,
6339 .src_rq = busiest_rq,
6343 schedstat_inc(sd, alb_count);
6345 if (move_one_task(&env))
6346 schedstat_inc(sd, alb_pushed);
6348 schedstat_inc(sd, alb_failed);
6351 double_unlock_balance(busiest_rq, target_rq);
6353 busiest_rq->active_balance = 0;
6354 raw_spin_unlock_irq(&busiest_rq->lock);
6358 #ifdef CONFIG_NO_HZ_COMMON
6360 * idle load balancing details
6361 * - When one of the busy CPUs notice that there may be an idle rebalancing
6362 * needed, they will kick the idle load balancer, which then does idle
6363 * load balancing for all the idle CPUs.
6366 cpumask_var_t idle_cpus_mask;
6368 unsigned long next_balance; /* in jiffy units */
6369 } nohz ____cacheline_aligned;
6372 * nohz_test_cpu used when load tracking is enabled. FAIR_GROUP_SCHED
6373 * dependency below may be removed when load tracking guards are
6376 #ifdef CONFIG_FAIR_GROUP_SCHED
6377 static int nohz_test_cpu(int cpu)
6379 return cpumask_test_cpu(cpu, nohz.idle_cpus_mask);
6383 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6385 * Decide if the tasks on the busy CPUs in the
6386 * littlest domain would benefit from an idle balance
6388 static int hmp_packing_ilb_needed(int cpu)
6390 struct hmp_domain *hmp;
6391 /* always allow ilb on non-slowest domain */
6392 if (!hmp_cpu_is_slowest(cpu))
6395 /* if disabled, use normal ILB behaviour */
6396 if (!hmp_packing_enabled)
6399 hmp = hmp_cpu_domain(cpu);
6400 for_each_cpu_and(cpu, &hmp->cpus, nohz.idle_cpus_mask) {
6401 /* only idle balance if a CPU is loaded over threshold */
6402 if (cpu_rq(cpu)->avg.load_avg_ratio > hmp_full_threshold)
6409 static inline int find_new_ilb(int call_cpu)
6411 int ilb = cpumask_first(nohz.idle_cpus_mask);
6412 #ifdef CONFIG_SCHED_HMP
6415 /* restrict nohz balancing to occur in the same hmp domain */
6416 ilb = cpumask_first_and(nohz.idle_cpus_mask,
6417 &((struct hmp_domain *)hmp_cpu_domain(call_cpu))->cpus);
6419 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6420 if (ilb < nr_cpu_ids)
6421 ilb_needed = hmp_packing_ilb_needed(ilb);
6424 if (ilb_needed && ilb < nr_cpu_ids && idle_cpu(ilb))
6427 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6435 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6436 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6437 * CPU (if there is one).
6439 static void nohz_balancer_kick(int cpu)
6443 nohz.next_balance++;
6445 ilb_cpu = find_new_ilb(cpu);
6447 if (ilb_cpu >= nr_cpu_ids)
6450 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6453 * Use smp_send_reschedule() instead of resched_cpu().
6454 * This way we generate a sched IPI on the target cpu which
6455 * is idle. And the softirq performing nohz idle load balance
6456 * will be run before returning from the IPI.
6458 smp_send_reschedule(ilb_cpu);
6462 static inline void nohz_balance_exit_idle(int cpu)
6464 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6465 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6466 atomic_dec(&nohz.nr_cpus);
6467 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6471 static inline void set_cpu_sd_state_busy(void)
6473 struct sched_domain *sd;
6474 int cpu = smp_processor_id();
6477 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6479 if (!sd || !sd->nohz_idle)
6483 for (; sd; sd = sd->parent)
6484 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6489 void set_cpu_sd_state_idle(void)
6491 struct sched_domain *sd;
6492 int cpu = smp_processor_id();
6495 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6497 if (!sd || sd->nohz_idle)
6501 for (; sd; sd = sd->parent)
6502 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6508 * This routine will record that the cpu is going idle with tick stopped.
6509 * This info will be used in performing idle load balancing in the future.
6511 void nohz_balance_enter_idle(int cpu)
6514 * If this cpu is going down, then nothing needs to be done.
6516 if (!cpu_active(cpu))
6519 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6522 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6523 atomic_inc(&nohz.nr_cpus);
6524 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6527 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
6528 unsigned long action, void *hcpu)
6530 switch (action & ~CPU_TASKS_FROZEN) {
6532 nohz_balance_exit_idle(smp_processor_id());
6540 static DEFINE_SPINLOCK(balancing);
6543 * Scale the max load_balance interval with the number of CPUs in the system.
6544 * This trades load-balance latency on larger machines for less cross talk.
6546 void update_max_interval(void)
6548 max_load_balance_interval = HZ*num_online_cpus()/10;
6552 * It checks each scheduling domain to see if it is due to be balanced,
6553 * and initiates a balancing operation if so.
6555 * Balancing parameters are set up in init_sched_domains.
6557 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6560 struct rq *rq = cpu_rq(cpu);
6561 unsigned long interval;
6562 struct sched_domain *sd;
6563 /* Earliest time when we have to do rebalance again */
6564 unsigned long next_balance = jiffies + 60*HZ;
6565 int update_next_balance = 0;
6568 update_blocked_averages(cpu);
6571 for_each_domain(cpu, sd) {
6572 if (!(sd->flags & SD_LOAD_BALANCE))
6575 interval = sd->balance_interval;
6576 if (idle != CPU_IDLE)
6577 interval *= sd->busy_factor;
6579 /* scale ms to jiffies */
6580 interval = msecs_to_jiffies(interval);
6581 interval = clamp(interval, 1UL, max_load_balance_interval);
6583 need_serialize = sd->flags & SD_SERIALIZE;
6585 if (need_serialize) {
6586 if (!spin_trylock(&balancing))
6590 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6591 if (load_balance(cpu, rq, sd, idle, &balance)) {
6593 * The LBF_SOME_PINNED logic could have changed
6594 * env->dst_cpu, so we can't know our idle
6595 * state even if we migrated tasks. Update it.
6597 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6599 sd->last_balance = jiffies;
6602 spin_unlock(&balancing);
6604 if (time_after(next_balance, sd->last_balance + interval)) {
6605 next_balance = sd->last_balance + interval;
6606 update_next_balance = 1;
6610 * Stop the load balance at this level. There is another
6611 * CPU in our sched group which is doing load balancing more
6620 * next_balance will be updated only when there is a need.
6621 * When the cpu is attached to null domain for ex, it will not be
6624 if (likely(update_next_balance))
6625 rq->next_balance = next_balance;
6628 #ifdef CONFIG_NO_HZ_COMMON
6630 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6631 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6633 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6635 struct rq *this_rq = cpu_rq(this_cpu);
6639 if (idle != CPU_IDLE ||
6640 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6643 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6644 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6648 * If this cpu gets work to do, stop the load balancing
6649 * work being done for other cpus. Next load
6650 * balancing owner will pick it up.
6655 rq = cpu_rq(balance_cpu);
6657 raw_spin_lock_irq(&rq->lock);
6658 update_rq_clock(rq);
6659 update_idle_cpu_load(rq);
6660 raw_spin_unlock_irq(&rq->lock);
6662 rebalance_domains(balance_cpu, CPU_IDLE);
6664 if (time_after(this_rq->next_balance, rq->next_balance))
6665 this_rq->next_balance = rq->next_balance;
6667 nohz.next_balance = this_rq->next_balance;
6669 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6673 * Current heuristic for kicking the idle load balancer in the presence
6674 * of an idle cpu is the system.
6675 * - This rq has more than one task.
6676 * - At any scheduler domain level, this cpu's scheduler group has multiple
6677 * busy cpu's exceeding the group's power.
6678 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6679 * domain span are idle.
6681 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6683 unsigned long now = jiffies;
6684 struct sched_domain *sd;
6686 if (unlikely(idle_cpu(cpu)))
6690 * We may be recently in ticked or tickless idle mode. At the first
6691 * busy tick after returning from idle, we will update the busy stats.
6693 set_cpu_sd_state_busy();
6694 nohz_balance_exit_idle(cpu);
6697 * None are in tickless mode and hence no need for NOHZ idle load
6700 if (likely(!atomic_read(&nohz.nr_cpus)))
6703 if (time_before(now, nohz.next_balance))
6706 #ifdef CONFIG_SCHED_HMP
6708 * Bail out if there are no nohz CPUs in our
6709 * HMP domain, since we will move tasks between
6710 * domains through wakeup and force balancing
6711 * as necessary based upon task load.
6713 if (cpumask_first_and(nohz.idle_cpus_mask,
6714 &((struct hmp_domain *)hmp_cpu_domain(cpu))->cpus) >= nr_cpu_ids)
6718 if (rq->nr_running >= 2)
6722 for_each_domain(cpu, sd) {
6723 struct sched_group *sg = sd->groups;
6724 struct sched_group_power *sgp = sg->sgp;
6725 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6727 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6728 goto need_kick_unlock;
6730 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6731 && (cpumask_first_and(nohz.idle_cpus_mask,
6732 sched_domain_span(sd)) < cpu))
6733 goto need_kick_unlock;
6735 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6747 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6750 #ifdef CONFIG_SCHED_HMP
6751 /* Check if task should migrate to a faster cpu */
6752 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se)
6754 struct task_struct *p = task_of(se);
6755 int temp_target_cpu;
6758 if (hmp_cpu_is_fastest(cpu))
6761 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6762 /* Filter by task priority */
6763 if (p->prio >= hmp_up_prio)
6766 if (se->avg.load_avg_ratio < hmp_up_threshold)
6769 /* Let the task load settle before doing another up migration */
6770 /* hack - always use clock from first online CPU */
6771 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6772 if (((now - se->avg.hmp_last_up_migration) >> 10)
6773 < hmp_next_up_threshold)
6776 /* hmp_domain_min_load only returns 0 for an
6777 * idle CPU or 1023 for any partly-busy one.
6778 * Be explicit about requirement for an idle CPU.
6780 if (hmp_domain_min_load(hmp_faster_domain(cpu), &temp_target_cpu,
6781 tsk_cpus_allowed(p)) == 0 && temp_target_cpu != NR_CPUS) {
6783 *target_cpu = temp_target_cpu;
6789 /* Check if task should migrate to a slower cpu */
6790 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
6792 struct task_struct *p = task_of(se);
6795 if (hmp_cpu_is_slowest(cpu)) {
6796 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6797 if(hmp_packing_enabled)
6804 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6805 /* Filter by task priority */
6806 if ((p->prio >= hmp_up_prio) &&
6807 cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6808 tsk_cpus_allowed(p))) {
6813 /* Let the task load settle before doing another down migration */
6814 /* hack - always use clock from first online CPU */
6815 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6816 if (((now - se->avg.hmp_last_down_migration) >> 10)
6817 < hmp_next_down_threshold)
6820 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6821 tsk_cpus_allowed(p))
6822 && se->avg.load_avg_ratio < hmp_down_threshold) {
6829 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6830 * Ideally this function should be merged with can_migrate_task() to avoid
6833 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
6835 int tsk_cache_hot = 0;
6838 * We do not migrate tasks that are:
6839 * 1) running (obviously), or
6840 * 2) cannot be migrated to this CPU due to cpus_allowed
6842 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6843 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6846 env->flags &= ~LBF_ALL_PINNED;
6848 if (task_running(env->src_rq, p)) {
6849 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6854 * Aggressive migration if:
6855 * 1) task is cache cold, or
6856 * 2) too many balance attempts have failed.
6859 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
6860 if (!tsk_cache_hot ||
6861 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6862 #ifdef CONFIG_SCHEDSTATS
6863 if (tsk_cache_hot) {
6864 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6865 schedstat_inc(p, se.statistics.nr_forced_migrations);
6875 * move_specific_task tries to move a specific task.
6876 * Returns 1 if successful and 0 otherwise.
6877 * Called with both runqueues locked.
6879 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
6881 struct task_struct *p, *n;
6883 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6884 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
6888 if (!hmp_can_migrate_task(p, env))
6890 /* Check if we found the right task */
6896 * Right now, this is only the third place move_task()
6897 * is called, so we can safely collect move_task()
6898 * stats here rather than inside move_task().
6900 schedstat_inc(env->sd, lb_gained[env->idle]);
6907 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
6908 * migrate a specific task from one runqueue to another.
6909 * hmp_force_up_migration uses this to push a currently running task
6911 * Based on active_load_balance_stop_cpu and can potentially be merged.
6913 static int hmp_active_task_migration_cpu_stop(void *data)
6915 struct rq *busiest_rq = data;
6916 struct task_struct *p = busiest_rq->migrate_task;
6917 int busiest_cpu = cpu_of(busiest_rq);
6918 int target_cpu = busiest_rq->push_cpu;
6919 struct rq *target_rq = cpu_rq(target_cpu);
6920 struct sched_domain *sd;
6922 raw_spin_lock_irq(&busiest_rq->lock);
6923 /* make sure the requested cpu hasn't gone down in the meantime */
6924 if (unlikely(busiest_cpu != smp_processor_id() ||
6925 !busiest_rq->active_balance)) {
6928 /* Is there any task to move? */
6929 if (busiest_rq->nr_running <= 1)
6931 /* Task has migrated meanwhile, abort forced migration */
6932 if (task_rq(p) != busiest_rq)
6935 * This condition is "impossible", if it occurs
6936 * we need to fix it. Originally reported by
6937 * Bjorn Helgaas on a 128-cpu setup.
6939 BUG_ON(busiest_rq == target_rq);
6941 /* move a task from busiest_rq to target_rq */
6942 double_lock_balance(busiest_rq, target_rq);
6944 /* Search for an sd spanning us and the target CPU. */
6946 for_each_domain(target_cpu, sd) {
6947 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6952 struct lb_env env = {
6954 .dst_cpu = target_cpu,
6955 .dst_rq = target_rq,
6956 .src_cpu = busiest_rq->cpu,
6957 .src_rq = busiest_rq,
6961 schedstat_inc(sd, alb_count);
6963 if (move_specific_task(&env, p))
6964 schedstat_inc(sd, alb_pushed);
6966 schedstat_inc(sd, alb_failed);
6969 double_unlock_balance(busiest_rq, target_rq);
6972 busiest_rq->active_balance = 0;
6973 raw_spin_unlock_irq(&busiest_rq->lock);
6978 * hmp_idle_pull_cpu_stop is run by cpu stopper and used to
6979 * migrate a specific task from one runqueue to another.
6980 * hmp_idle_pull uses this to push a currently running task
6981 * off a runqueue to a faster CPU.
6982 * Locking is slightly different than usual.
6983 * Based on active_load_balance_stop_cpu and can potentially be merged.
6985 static int hmp_idle_pull_cpu_stop(void *data)
6987 struct rq *busiest_rq = data;
6988 struct task_struct *p = busiest_rq->migrate_task;
6989 int busiest_cpu = cpu_of(busiest_rq);
6990 int target_cpu = busiest_rq->push_cpu;
6991 struct rq *target_rq = cpu_rq(target_cpu);
6992 struct sched_domain *sd;
6994 raw_spin_lock_irq(&busiest_rq->lock);
6996 /* make sure the requested cpu hasn't gone down in the meantime */
6997 if (unlikely(busiest_cpu != smp_processor_id() ||
6998 !busiest_rq->active_balance))
7001 /* Is there any task to move? */
7002 if (busiest_rq->nr_running <= 1)
7005 /* Task has migrated meanwhile, abort forced migration */
7006 if (task_rq(p) != busiest_rq)
7010 * This condition is "impossible", if it occurs
7011 * we need to fix it. Originally reported by
7012 * Bjorn Helgaas on a 128-cpu setup.
7014 BUG_ON(busiest_rq == target_rq);
7016 /* move a task from busiest_rq to target_rq */
7017 double_lock_balance(busiest_rq, target_rq);
7019 /* Search for an sd spanning us and the target CPU. */
7021 for_each_domain(target_cpu, sd) {
7022 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7026 struct lb_env env = {
7028 .dst_cpu = target_cpu,
7029 .dst_rq = target_rq,
7030 .src_cpu = busiest_rq->cpu,
7031 .src_rq = busiest_rq,
7035 schedstat_inc(sd, alb_count);
7037 if (move_specific_task(&env, p))
7038 schedstat_inc(sd, alb_pushed);
7040 schedstat_inc(sd, alb_failed);
7043 double_unlock_balance(busiest_rq, target_rq);
7046 busiest_rq->active_balance = 0;
7047 raw_spin_unlock_irq(&busiest_rq->lock);
7052 * Move task in a runnable state to another CPU.
7054 * Tailored on 'active_load_balance_stop_cpu' 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;
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 if (hmp_up_migration(cpu, &target_cpu, curr)) {
7156 if (!target->active_balance) {
7158 target->push_cpu = target_cpu;
7159 target->migrate_task = p;
7161 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_FORCE);
7162 hmp_next_up_delay(&p->se, target->push_cpu);
7165 if (!got_target && !target->active_balance) {
7167 * For now we just check the currently running task.
7168 * Selecting the lightest task for offloading will
7169 * require extensive book keeping.
7171 curr = hmp_get_lightest_task(orig, 1);
7173 target->push_cpu = hmp_offload_down(cpu, curr);
7174 if (target->push_cpu < NR_CPUS) {
7176 target->migrate_task = p;
7178 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_OFFLOAD);
7179 hmp_next_down_delay(&p->se, target->push_cpu);
7183 * We have a target with no active_balance. If the task
7184 * is not currently running move it, otherwise let the
7185 * CPU stopper take care of it.
7187 if (got_target && !target->active_balance) {
7188 if (!task_running(target, p)) {
7189 trace_sched_hmp_migrate_force_running(p, 0);
7190 hmp_migrate_runnable_task(target);
7192 target->active_balance = 1;
7197 raw_spin_unlock_irqrestore(&target->lock, flags);
7200 stop_one_cpu_nowait(cpu_of(target),
7201 hmp_active_task_migration_cpu_stop,
7202 target, &target->active_balance_work);
7204 spin_unlock(&hmp_force_migration);
7207 * hmp_idle_pull looks at little domain runqueues to see
7208 * if a task should be pulled.
7210 * Reuses hmp_force_migration spinlock.
7213 static unsigned int hmp_idle_pull(int this_cpu)
7216 struct sched_entity *curr, *orig;
7217 struct hmp_domain *hmp_domain = NULL;
7218 struct rq *target = NULL, *rq;
7219 unsigned long flags, ratio = 0;
7220 unsigned int force = 0;
7221 struct task_struct *p = NULL;
7223 if (!hmp_cpu_is_slowest(this_cpu))
7224 hmp_domain = hmp_slower_domain(this_cpu);
7228 if (!spin_trylock(&hmp_force_migration))
7231 /* first select a task */
7232 for_each_cpu(cpu, &hmp_domain->cpus) {
7234 raw_spin_lock_irqsave(&rq->lock, flags);
7235 curr = rq->cfs.curr;
7237 raw_spin_unlock_irqrestore(&rq->lock, flags);
7240 if (!entity_is_task(curr)) {
7241 struct cfs_rq *cfs_rq;
7243 cfs_rq = group_cfs_rq(curr);
7245 curr = cfs_rq->curr;
7246 if (!entity_is_task(curr))
7247 cfs_rq = group_cfs_rq(curr);
7253 curr = hmp_get_heaviest_task(curr, 1);
7254 if (curr->avg.load_avg_ratio > hmp_up_threshold &&
7255 curr->avg.load_avg_ratio > ratio) {
7258 ratio = curr->avg.load_avg_ratio;
7260 raw_spin_unlock_irqrestore(&rq->lock, flags);
7266 /* now we have a candidate */
7267 raw_spin_lock_irqsave(&target->lock, flags);
7268 if (!target->active_balance && task_rq(p) == target) {
7270 target->push_cpu = this_cpu;
7271 target->migrate_task = p;
7272 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_IDLE_PULL);
7273 hmp_next_up_delay(&p->se, target->push_cpu);
7275 * if the task isn't running move it right away.
7276 * Otherwise setup the active_balance mechanic and let
7277 * the CPU stopper do its job.
7279 if (!task_running(target, p)) {
7280 trace_sched_hmp_migrate_idle_running(p, 0);
7281 hmp_migrate_runnable_task(target);
7283 target->active_balance = 1;
7287 raw_spin_unlock_irqrestore(&target->lock, flags);
7290 stop_one_cpu_nowait(cpu_of(target),
7291 hmp_idle_pull_cpu_stop,
7292 target, &target->active_balance_work);
7295 spin_unlock(&hmp_force_migration);
7299 static void hmp_force_up_migration(int this_cpu) { }
7300 #endif /* CONFIG_SCHED_HMP */
7303 * run_rebalance_domains is triggered when needed from the scheduler tick.
7304 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7306 static void run_rebalance_domains(struct softirq_action *h)
7308 int this_cpu = smp_processor_id();
7309 struct rq *this_rq = cpu_rq(this_cpu);
7310 enum cpu_idle_type idle = this_rq->idle_balance ?
7311 CPU_IDLE : CPU_NOT_IDLE;
7313 hmp_force_up_migration(this_cpu);
7315 rebalance_domains(this_cpu, idle);
7318 * If this cpu has a pending nohz_balance_kick, then do the
7319 * balancing on behalf of the other idle cpus whose ticks are
7322 nohz_idle_balance(this_cpu, idle);
7325 static inline int on_null_domain(int cpu)
7327 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
7331 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7333 void trigger_load_balance(struct rq *rq, int cpu)
7335 /* Don't need to rebalance while attached to NULL domain */
7336 if (time_after_eq(jiffies, rq->next_balance) &&
7337 likely(!on_null_domain(cpu)))
7338 raise_softirq(SCHED_SOFTIRQ);
7339 #ifdef CONFIG_NO_HZ_COMMON
7340 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
7341 nohz_balancer_kick(cpu);
7345 static void rq_online_fair(struct rq *rq)
7347 #ifdef CONFIG_SCHED_HMP
7348 hmp_online_cpu(rq->cpu);
7353 static void rq_offline_fair(struct rq *rq)
7355 #ifdef CONFIG_SCHED_HMP
7356 hmp_offline_cpu(rq->cpu);
7360 /* Ensure any throttled groups are reachable by pick_next_task */
7361 unthrottle_offline_cfs_rqs(rq);
7364 #endif /* CONFIG_SMP */
7367 * scheduler tick hitting a task of our scheduling class:
7369 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7371 struct cfs_rq *cfs_rq;
7372 struct sched_entity *se = &curr->se;
7374 for_each_sched_entity(se) {
7375 cfs_rq = cfs_rq_of(se);
7376 entity_tick(cfs_rq, se, queued);
7379 if (sched_feat_numa(NUMA))
7380 task_tick_numa(rq, curr);
7382 update_rq_runnable_avg(rq, 1);
7386 * called on fork with the child task as argument from the parent's context
7387 * - child not yet on the tasklist
7388 * - preemption disabled
7390 static void task_fork_fair(struct task_struct *p)
7392 struct cfs_rq *cfs_rq;
7393 struct sched_entity *se = &p->se, *curr;
7394 int this_cpu = smp_processor_id();
7395 struct rq *rq = this_rq();
7396 unsigned long flags;
7398 raw_spin_lock_irqsave(&rq->lock, flags);
7400 update_rq_clock(rq);
7402 cfs_rq = task_cfs_rq(current);
7403 curr = cfs_rq->curr;
7406 * Not only the cpu but also the task_group of the parent might have
7407 * been changed after parent->se.parent,cfs_rq were copied to
7408 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7409 * of child point to valid ones.
7412 __set_task_cpu(p, this_cpu);
7415 update_curr(cfs_rq);
7418 se->vruntime = curr->vruntime;
7419 place_entity(cfs_rq, se, 1);
7421 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7423 * Upon rescheduling, sched_class::put_prev_task() will place
7424 * 'current' within the tree based on its new key value.
7426 swap(curr->vruntime, se->vruntime);
7427 resched_task(rq->curr);
7430 se->vruntime -= cfs_rq->min_vruntime;
7432 raw_spin_unlock_irqrestore(&rq->lock, flags);
7436 * Priority of the task has changed. Check to see if we preempt
7440 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7446 * Reschedule if we are currently running on this runqueue and
7447 * our priority decreased, or if we are not currently running on
7448 * this runqueue and our priority is higher than the current's
7450 if (rq->curr == p) {
7451 if (p->prio > oldprio)
7452 resched_task(rq->curr);
7454 check_preempt_curr(rq, p, 0);
7457 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7459 struct sched_entity *se = &p->se;
7460 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7463 * Ensure the task's vruntime is normalized, so that when it's
7464 * switched back to the fair class the enqueue_entity(.flags=0) will
7465 * do the right thing.
7467 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7468 * have normalized the vruntime, if it's !on_rq, then only when
7469 * the task is sleeping will it still have non-normalized vruntime.
7471 if (!p->on_rq && p->state != TASK_RUNNING) {
7473 * Fix up our vruntime so that the current sleep doesn't
7474 * cause 'unlimited' sleep bonus.
7476 place_entity(cfs_rq, se, 0);
7477 se->vruntime -= cfs_rq->min_vruntime;
7480 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7482 * Remove our load from contribution when we leave sched_fair
7483 * and ensure we don't carry in an old decay_count if we
7486 if (p->se.avg.decay_count) {
7487 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
7488 __synchronize_entity_decay(&p->se);
7489 subtract_blocked_load_contrib(cfs_rq,
7490 p->se.avg.load_avg_contrib);
7496 * We switched to the sched_fair class.
7498 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7504 * We were most likely switched from sched_rt, so
7505 * kick off the schedule if running, otherwise just see
7506 * if we can still preempt the current task.
7509 resched_task(rq->curr);
7511 check_preempt_curr(rq, p, 0);
7514 /* Account for a task changing its policy or group.
7516 * This routine is mostly called to set cfs_rq->curr field when a task
7517 * migrates between groups/classes.
7519 static void set_curr_task_fair(struct rq *rq)
7521 struct sched_entity *se = &rq->curr->se;
7523 for_each_sched_entity(se) {
7524 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7526 set_next_entity(cfs_rq, se);
7527 /* ensure bandwidth has been allocated on our new cfs_rq */
7528 account_cfs_rq_runtime(cfs_rq, 0);
7532 void init_cfs_rq(struct cfs_rq *cfs_rq)
7534 cfs_rq->tasks_timeline = RB_ROOT;
7535 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7536 #ifndef CONFIG_64BIT
7537 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7539 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7540 atomic64_set(&cfs_rq->decay_counter, 1);
7541 atomic64_set(&cfs_rq->removed_load, 0);
7545 #ifdef CONFIG_FAIR_GROUP_SCHED
7546 static void task_move_group_fair(struct task_struct *p, int on_rq)
7548 struct cfs_rq *cfs_rq;
7550 * If the task was not on the rq at the time of this cgroup movement
7551 * it must have been asleep, sleeping tasks keep their ->vruntime
7552 * absolute on their old rq until wakeup (needed for the fair sleeper
7553 * bonus in place_entity()).
7555 * If it was on the rq, we've just 'preempted' it, which does convert
7556 * ->vruntime to a relative base.
7558 * Make sure both cases convert their relative position when migrating
7559 * to another cgroup's rq. This does somewhat interfere with the
7560 * fair sleeper stuff for the first placement, but who cares.
7563 * When !on_rq, vruntime of the task has usually NOT been normalized.
7564 * But there are some cases where it has already been normalized:
7566 * - Moving a forked child which is waiting for being woken up by
7567 * wake_up_new_task().
7568 * - Moving a task which has been woken up by try_to_wake_up() and
7569 * waiting for actually being woken up by sched_ttwu_pending().
7571 * To prevent boost or penalty in the new cfs_rq caused by delta
7572 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7574 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7578 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7579 set_task_rq(p, task_cpu(p));
7581 cfs_rq = cfs_rq_of(&p->se);
7582 p->se.vruntime += cfs_rq->min_vruntime;
7585 * migrate_task_rq_fair() will have removed our previous
7586 * contribution, but we must synchronize for ongoing future
7589 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7590 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7595 void free_fair_sched_group(struct task_group *tg)
7599 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7601 for_each_possible_cpu(i) {
7603 kfree(tg->cfs_rq[i]);
7612 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7614 struct cfs_rq *cfs_rq;
7615 struct sched_entity *se;
7618 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7621 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7625 tg->shares = NICE_0_LOAD;
7627 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7629 for_each_possible_cpu(i) {
7630 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7631 GFP_KERNEL, cpu_to_node(i));
7635 se = kzalloc_node(sizeof(struct sched_entity),
7636 GFP_KERNEL, cpu_to_node(i));
7640 init_cfs_rq(cfs_rq);
7641 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7652 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7654 struct rq *rq = cpu_rq(cpu);
7655 unsigned long flags;
7658 * Only empty task groups can be destroyed; so we can speculatively
7659 * check on_list without danger of it being re-added.
7661 if (!tg->cfs_rq[cpu]->on_list)
7664 raw_spin_lock_irqsave(&rq->lock, flags);
7665 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7666 raw_spin_unlock_irqrestore(&rq->lock, flags);
7669 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7670 struct sched_entity *se, int cpu,
7671 struct sched_entity *parent)
7673 struct rq *rq = cpu_rq(cpu);
7677 init_cfs_rq_runtime(cfs_rq);
7679 tg->cfs_rq[cpu] = cfs_rq;
7682 /* se could be NULL for root_task_group */
7687 se->cfs_rq = &rq->cfs;
7689 se->cfs_rq = parent->my_q;
7692 /* guarantee group entities always have weight */
7693 update_load_set(&se->load, NICE_0_LOAD);
7694 se->parent = parent;
7697 static DEFINE_MUTEX(shares_mutex);
7699 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7702 unsigned long flags;
7705 * We can't change the weight of the root cgroup.
7710 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7712 mutex_lock(&shares_mutex);
7713 if (tg->shares == shares)
7716 tg->shares = shares;
7717 for_each_possible_cpu(i) {
7718 struct rq *rq = cpu_rq(i);
7719 struct sched_entity *se;
7722 /* Propagate contribution to hierarchy */
7723 raw_spin_lock_irqsave(&rq->lock, flags);
7724 for_each_sched_entity(se)
7725 update_cfs_shares(group_cfs_rq(se));
7726 raw_spin_unlock_irqrestore(&rq->lock, flags);
7730 mutex_unlock(&shares_mutex);
7733 #else /* CONFIG_FAIR_GROUP_SCHED */
7735 void free_fair_sched_group(struct task_group *tg) { }
7737 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7742 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7744 #endif /* CONFIG_FAIR_GROUP_SCHED */
7747 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7749 struct sched_entity *se = &task->se;
7750 unsigned int rr_interval = 0;
7753 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7756 if (rq->cfs.load.weight)
7757 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7763 * All the scheduling class methods:
7765 const struct sched_class fair_sched_class = {
7766 .next = &idle_sched_class,
7767 .enqueue_task = enqueue_task_fair,
7768 .dequeue_task = dequeue_task_fair,
7769 .yield_task = yield_task_fair,
7770 .yield_to_task = yield_to_task_fair,
7772 .check_preempt_curr = check_preempt_wakeup,
7774 .pick_next_task = pick_next_task_fair,
7775 .put_prev_task = put_prev_task_fair,
7778 .select_task_rq = select_task_rq_fair,
7779 #ifdef CONFIG_FAIR_GROUP_SCHED
7780 .migrate_task_rq = migrate_task_rq_fair,
7782 .rq_online = rq_online_fair,
7783 .rq_offline = rq_offline_fair,
7785 .task_waking = task_waking_fair,
7788 .set_curr_task = set_curr_task_fair,
7789 .task_tick = task_tick_fair,
7790 .task_fork = task_fork_fair,
7792 .prio_changed = prio_changed_fair,
7793 .switched_from = switched_from_fair,
7794 .switched_to = switched_to_fair,
7796 .get_rr_interval = get_rr_interval_fair,
7798 #ifdef CONFIG_FAIR_GROUP_SCHED
7799 .task_move_group = task_move_group_fair,
7803 #ifdef CONFIG_SCHED_DEBUG
7804 void print_cfs_stats(struct seq_file *m, int cpu)
7806 struct cfs_rq *cfs_rq;
7809 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7810 print_cfs_rq(m, cpu, cfs_rq);
7815 __init void init_sched_fair_class(void)
7818 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7820 #ifdef CONFIG_NO_HZ_COMMON
7821 nohz.next_balance = jiffies;
7822 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7823 cpu_notifier(sched_ilb_notifier, 0);
7826 #ifdef CONFIG_SCHED_HMP
7827 hmp_cpu_mask_setup();
7833 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
7834 static u32 cpufreq_calc_scale(u32 min, u32 max, u32 curr)
7836 u32 result = curr / max;
7840 /* Called when the CPU Frequency is changed.
7841 * Once for each CPU.
7843 static int cpufreq_callback(struct notifier_block *nb,
7844 unsigned long val, void *data)
7846 struct cpufreq_freqs *freq = data;
7847 int cpu = freq->cpu;
7848 struct cpufreq_extents *extents;
7850 if (freq->flags & CPUFREQ_CONST_LOOPS)
7853 if (val != CPUFREQ_POSTCHANGE)
7856 /* if dynamic load scale is disabled, set the load scale to 1.0 */
7857 if (!hmp_data.freqinvar_load_scale_enabled) {
7858 freq_scale[cpu].curr_scale = 1024;
7862 extents = &freq_scale[cpu];
7863 if (extents->flags & SCHED_LOAD_FREQINVAR_SINGLEFREQ) {
7864 /* If our governor was recognised as a single-freq governor,
7867 extents->curr_scale = 1024;
7869 extents->curr_scale = cpufreq_calc_scale(extents->min,
7870 extents->max, freq->new);
7876 /* Called when the CPUFreq governor is changed.
7877 * Only called for the CPUs which are actually changed by the
7880 static int cpufreq_policy_callback(struct notifier_block *nb,
7881 unsigned long event, void *data)
7883 struct cpufreq_policy *policy = data;
7884 struct cpufreq_extents *extents;
7885 int cpu, singleFreq = 0;
7886 static const char performance_governor[] = "performance";
7887 static const char powersave_governor[] = "powersave";
7889 if (event == CPUFREQ_START)
7892 if (event != CPUFREQ_INCOMPATIBLE)
7895 /* CPUFreq governors do not accurately report the range of
7896 * CPU Frequencies they will choose from.
7897 * We recognise performance and powersave governors as
7898 * single-frequency only.
7900 if (!strncmp(policy->governor->name, performance_governor,
7901 strlen(performance_governor)) ||
7902 !strncmp(policy->governor->name, powersave_governor,
7903 strlen(powersave_governor)))
7906 /* Make sure that all CPUs impacted by this policy are
7907 * updated since we will only get a notification when the
7908 * user explicitly changes the policy on a CPU.
7910 for_each_cpu(cpu, policy->cpus) {
7911 extents = &freq_scale[cpu];
7912 extents->max = policy->max >> SCHED_FREQSCALE_SHIFT;
7913 extents->min = policy->min >> SCHED_FREQSCALE_SHIFT;
7914 if (!hmp_data.freqinvar_load_scale_enabled) {
7915 extents->curr_scale = 1024;
7916 } else if (singleFreq) {
7917 extents->flags |= SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7918 extents->curr_scale = 1024;
7920 extents->flags &= ~SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7921 extents->curr_scale = cpufreq_calc_scale(extents->min,
7922 extents->max, policy->cur);
7929 static struct notifier_block cpufreq_notifier = {
7930 .notifier_call = cpufreq_callback,
7932 static struct notifier_block cpufreq_policy_notifier = {
7933 .notifier_call = cpufreq_policy_callback,
7936 static int __init register_sched_cpufreq_notifier(void)
7940 /* init safe defaults since there are no policies at registration */
7941 for (ret = 0; ret < CONFIG_NR_CPUS; ret++) {
7943 freq_scale[ret].max = 1024;
7944 freq_scale[ret].min = 1024;
7945 freq_scale[ret].curr_scale = 1024;
7948 pr_info("sched: registering cpufreq notifiers for scale-invariant loads\n");
7949 ret = cpufreq_register_notifier(&cpufreq_policy_notifier,
7950 CPUFREQ_POLICY_NOTIFIER);
7953 ret = cpufreq_register_notifier(&cpufreq_notifier,
7954 CPUFREQ_TRANSITION_NOTIFIER);
7959 core_initcall(register_sched_cpufreq_notifier);
7960 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */