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 start = max(start, vma->vm_start);
950 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
951 end = min(end, vma->vm_end);
952 pages -= change_prot_numa(vma, start, end);
957 } while (end != vma->vm_end);
962 * It is possible to reach the end of the VMA list but the last few VMAs are
963 * not guaranteed to the vma_migratable. If they are not, we would find the
964 * !migratable VMA on the next scan but not reset the scanner to the start
968 mm->numa_scan_offset = start;
970 reset_ptenuma_scan(p);
971 up_read(&mm->mmap_sem);
975 * Drive the periodic memory faults..
977 void task_tick_numa(struct rq *rq, struct task_struct *curr)
979 struct callback_head *work = &curr->numa_work;
983 * We don't care about NUMA placement if we don't have memory.
985 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
989 * Using runtime rather than walltime has the dual advantage that
990 * we (mostly) drive the selection from busy threads and that the
991 * task needs to have done some actual work before we bother with
994 now = curr->se.sum_exec_runtime;
995 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
997 if (now - curr->node_stamp > period) {
998 if (!curr->node_stamp)
999 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1000 curr->node_stamp = now;
1002 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1003 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1004 task_work_add(curr, work, true);
1009 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1012 #endif /* CONFIG_NUMA_BALANCING */
1015 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1017 update_load_add(&cfs_rq->load, se->load.weight);
1018 if (!parent_entity(se))
1019 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1021 if (entity_is_task(se))
1022 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1024 cfs_rq->nr_running++;
1028 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1030 update_load_sub(&cfs_rq->load, se->load.weight);
1031 if (!parent_entity(se))
1032 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1033 if (entity_is_task(se))
1034 list_del_init(&se->group_node);
1035 cfs_rq->nr_running--;
1038 #ifdef CONFIG_FAIR_GROUP_SCHED
1040 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1045 * Use this CPU's actual weight instead of the last load_contribution
1046 * to gain a more accurate current total weight. See
1047 * update_cfs_rq_load_contribution().
1049 tg_weight = atomic64_read(&tg->load_avg);
1050 tg_weight -= cfs_rq->tg_load_contrib;
1051 tg_weight += cfs_rq->load.weight;
1056 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1058 long tg_weight, load, shares;
1060 tg_weight = calc_tg_weight(tg, cfs_rq);
1061 load = cfs_rq->load.weight;
1063 shares = (tg->shares * load);
1065 shares /= tg_weight;
1067 if (shares < MIN_SHARES)
1068 shares = MIN_SHARES;
1069 if (shares > tg->shares)
1070 shares = tg->shares;
1074 # else /* CONFIG_SMP */
1075 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1079 # endif /* CONFIG_SMP */
1080 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1081 unsigned long weight)
1084 /* commit outstanding execution time */
1085 if (cfs_rq->curr == se)
1086 update_curr(cfs_rq);
1087 account_entity_dequeue(cfs_rq, se);
1090 update_load_set(&se->load, weight);
1093 account_entity_enqueue(cfs_rq, se);
1096 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1098 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1100 struct task_group *tg;
1101 struct sched_entity *se;
1105 se = tg->se[cpu_of(rq_of(cfs_rq))];
1106 if (!se || throttled_hierarchy(cfs_rq))
1109 if (likely(se->load.weight == tg->shares))
1112 shares = calc_cfs_shares(cfs_rq, tg);
1114 reweight_entity(cfs_rq_of(se), se, shares);
1116 #else /* CONFIG_FAIR_GROUP_SCHED */
1117 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1120 #endif /* CONFIG_FAIR_GROUP_SCHED */
1122 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1123 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1125 * We choose a half-life close to 1 scheduling period.
1126 * Note: The tables below are dependent on this value.
1128 #define LOAD_AVG_PERIOD 32
1129 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1130 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1132 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1133 static const u32 runnable_avg_yN_inv[] = {
1134 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1135 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1136 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1137 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1138 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1139 0x85aac367, 0x82cd8698,
1143 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1144 * over-estimates when re-combining.
1146 static const u32 runnable_avg_yN_sum[] = {
1147 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1148 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1149 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1154 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1156 static __always_inline u64 decay_load(u64 val, u64 n)
1158 unsigned int local_n;
1162 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1165 /* after bounds checking we can collapse to 32-bit */
1169 * As y^PERIOD = 1/2, we can combine
1170 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1171 * With a look-up table which covers k^n (n<PERIOD)
1173 * To achieve constant time decay_load.
1175 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1176 val >>= local_n / LOAD_AVG_PERIOD;
1177 local_n %= LOAD_AVG_PERIOD;
1180 val *= runnable_avg_yN_inv[local_n];
1181 /* We don't use SRR here since we always want to round down. */
1186 * For updates fully spanning n periods, the contribution to runnable
1187 * average will be: \Sum 1024*y^n
1189 * We can compute this reasonably efficiently by combining:
1190 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1192 static u32 __compute_runnable_contrib(u64 n)
1196 if (likely(n <= LOAD_AVG_PERIOD))
1197 return runnable_avg_yN_sum[n];
1198 else if (unlikely(n >= LOAD_AVG_MAX_N))
1199 return LOAD_AVG_MAX;
1201 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1203 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1204 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1206 n -= LOAD_AVG_PERIOD;
1207 } while (n > LOAD_AVG_PERIOD);
1209 contrib = decay_load(contrib, n);
1210 return contrib + runnable_avg_yN_sum[n];
1213 #define HMP_VARIABLE_SCALE_SHIFT 16ULL
1214 struct hmp_global_attr {
1215 struct attribute attr;
1216 ssize_t (*show)(struct kobject *kobj,
1217 struct attribute *attr, char *buf);
1218 ssize_t (*store)(struct kobject *a, struct attribute *b,
1219 const char *c, size_t count);
1221 int (*to_sysfs)(int);
1222 int (*from_sysfs)(int);
1223 ssize_t (*to_sysfs_text)(char *buf, int buf_size);
1226 #define HMP_DATA_SYSFS_MAX 8
1228 struct hmp_data_struct {
1229 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1230 int freqinvar_load_scale_enabled;
1232 int multiplier; /* used to scale the time delta */
1233 struct attribute_group attr_group;
1234 struct attribute *attributes[HMP_DATA_SYSFS_MAX + 1];
1235 struct hmp_global_attr attr[HMP_DATA_SYSFS_MAX];
1238 static u64 hmp_variable_scale_convert(u64 delta);
1239 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1240 /* Frequency-Invariant Load Modification:
1241 * Loads are calculated as in PJT's patch however we also scale the current
1242 * contribution in line with the frequency of the CPU that the task was
1244 * In this version, we use a simple linear scale derived from the maximum
1245 * frequency reported by CPUFreq. As an example:
1247 * Consider that we ran a task for 100% of the previous interval.
1249 * Our CPU was under asynchronous frequency control through one of the
1250 * CPUFreq governors.
1252 * The CPUFreq governor reports that it is able to scale the CPU between
1255 * During the period, the CPU was running at 1GHz.
1257 * In this case, our load contribution for that period is calculated as
1258 * 1 * (number_of_active_microseconds)
1260 * This results in our task being able to accumulate maximum load as normal.
1263 * Consider now that our CPU was executing at 500MHz.
1265 * We now scale the load contribution such that it is calculated as
1266 * 0.5 * (number_of_active_microseconds)
1268 * Our task can only record 50% maximum load during this period.
1270 * This represents the task consuming 50% of the CPU's *possible* compute
1271 * capacity. However the task did consume 100% of the CPU's *available*
1272 * compute capacity which is the value seen by the CPUFreq governor and
1273 * user-side CPU Utilization tools.
1275 * Restricting tracked load to be scaled by the CPU's frequency accurately
1276 * represents the consumption of possible compute capacity and allows the
1277 * HMP migration's simple threshold migration strategy to interact more
1278 * predictably with CPUFreq's asynchronous compute capacity changes.
1280 #define SCHED_FREQSCALE_SHIFT 10
1281 struct cpufreq_extents {
1287 /* Flag set when the governor in use only allows one frequency.
1290 #define SCHED_LOAD_FREQINVAR_SINGLEFREQ 0x01
1292 static struct cpufreq_extents freq_scale[CONFIG_NR_CPUS];
1293 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1295 /* We can represent the historical contribution to runnable average as the
1296 * coefficients of a geometric series. To do this we sub-divide our runnable
1297 * history into segments of approximately 1ms (1024us); label the segment that
1298 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1300 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1302 * (now) (~1ms ago) (~2ms ago)
1304 * Let u_i denote the fraction of p_i that the entity was runnable.
1306 * We then designate the fractions u_i as our co-efficients, yielding the
1307 * following representation of historical load:
1308 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1310 * We choose y based on the with of a reasonably scheduling period, fixing:
1313 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1314 * approximately half as much as the contribution to load within the last ms
1317 * When a period "rolls over" and we have new u_0`, multiplying the previous
1318 * sum again by y is sufficient to update:
1319 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1320 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1322 static __always_inline int __update_entity_runnable_avg(u64 now,
1323 struct sched_avg *sa,
1329 u32 runnable_contrib;
1330 int delta_w, decayed = 0;
1331 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1333 u32 scaled_runnable_contrib;
1335 u32 curr_scale = 1024;
1336 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1338 delta = now - sa->last_runnable_update;
1340 delta = hmp_variable_scale_convert(delta);
1342 * This should only happen when time goes backwards, which it
1343 * unfortunately does during sched clock init when we swap over to TSC.
1345 if ((s64)delta < 0) {
1346 sa->last_runnable_update = now;
1351 * Use 1024ns as the unit of measurement since it's a reasonable
1352 * approximation of 1us and fast to compute.
1357 sa->last_runnable_update = now;
1359 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1360 /* retrieve scale factor for load */
1361 if (hmp_data.freqinvar_load_scale_enabled)
1362 curr_scale = freq_scale[cpu].curr_scale;
1363 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1365 /* delta_w is the amount already accumulated against our next period */
1366 delta_w = sa->runnable_avg_period % 1024;
1367 if (delta + delta_w >= 1024) {
1368 /* period roll-over */
1372 * Now that we know we're crossing a period boundary, figure
1373 * out how much from delta we need to complete the current
1374 * period and accrue it.
1376 delta_w = 1024 - delta_w;
1377 /* scale runnable time if necessary */
1378 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1379 scaled_delta_w = (delta_w * curr_scale)
1380 >> SCHED_FREQSCALE_SHIFT;
1382 sa->runnable_avg_sum += scaled_delta_w;
1384 sa->usage_avg_sum += scaled_delta_w;
1387 sa->runnable_avg_sum += delta_w;
1389 sa->usage_avg_sum += delta_w;
1390 #endif /* #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1391 sa->runnable_avg_period += delta_w;
1395 /* Figure out how many additional periods this update spans */
1396 periods = delta / 1024;
1398 /* decay the load we have accumulated so far */
1399 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1401 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1403 sa->usage_avg_sum = decay_load(sa->usage_avg_sum, periods + 1);
1404 /* add the contribution from this period */
1405 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1406 runnable_contrib = __compute_runnable_contrib(periods);
1407 /* Apply load scaling if necessary.
1408 * Note that multiplying the whole series is same as
1409 * multiplying all terms
1411 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1412 scaled_runnable_contrib = (runnable_contrib * curr_scale)
1413 >> SCHED_FREQSCALE_SHIFT;
1415 sa->runnable_avg_sum += scaled_runnable_contrib;
1417 sa->usage_avg_sum += scaled_runnable_contrib;
1420 sa->runnable_avg_sum += runnable_contrib;
1422 sa->usage_avg_sum += runnable_contrib;
1423 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1424 sa->runnable_avg_period += runnable_contrib;
1427 /* Remainder of delta accrued against u_0` */
1428 /* scale if necessary */
1429 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1430 scaled_delta = ((delta * curr_scale) >> SCHED_FREQSCALE_SHIFT);
1432 sa->runnable_avg_sum += scaled_delta;
1434 sa->usage_avg_sum += scaled_delta;
1437 sa->runnable_avg_sum += delta;
1439 sa->usage_avg_sum += delta;
1440 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */
1441 sa->runnable_avg_period += delta;
1446 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1447 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1449 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1450 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1452 decays -= se->avg.decay_count;
1456 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1457 se->avg.decay_count = 0;
1462 #ifdef CONFIG_FAIR_GROUP_SCHED
1463 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1466 struct task_group *tg = cfs_rq->tg;
1469 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1470 tg_contrib -= cfs_rq->tg_load_contrib;
1472 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1473 atomic64_add(tg_contrib, &tg->load_avg);
1474 cfs_rq->tg_load_contrib += tg_contrib;
1479 * Aggregate cfs_rq runnable averages into an equivalent task_group
1480 * representation for computing load contributions.
1482 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1483 struct cfs_rq *cfs_rq)
1485 struct task_group *tg = cfs_rq->tg;
1486 long contrib, usage_contrib;
1488 /* The fraction of a cpu used by this cfs_rq */
1489 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1490 sa->runnable_avg_period + 1);
1491 contrib -= cfs_rq->tg_runnable_contrib;
1493 usage_contrib = div_u64(sa->usage_avg_sum << NICE_0_SHIFT,
1494 sa->runnable_avg_period + 1);
1495 usage_contrib -= cfs_rq->tg_usage_contrib;
1498 * contrib/usage at this point represent deltas, only update if they
1501 if ((abs(contrib) > cfs_rq->tg_runnable_contrib / 64) ||
1502 (abs(usage_contrib) > cfs_rq->tg_usage_contrib / 64)) {
1503 atomic_add(contrib, &tg->runnable_avg);
1504 cfs_rq->tg_runnable_contrib += contrib;
1506 atomic_add(usage_contrib, &tg->usage_avg);
1507 cfs_rq->tg_usage_contrib += usage_contrib;
1511 static inline void __update_group_entity_contrib(struct sched_entity *se)
1513 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1514 struct task_group *tg = cfs_rq->tg;
1519 contrib = cfs_rq->tg_load_contrib * tg->shares;
1520 se->avg.load_avg_contrib = div64_u64(contrib,
1521 atomic64_read(&tg->load_avg) + 1);
1524 * For group entities we need to compute a correction term in the case
1525 * that they are consuming <1 cpu so that we would contribute the same
1526 * load as a task of equal weight.
1528 * Explicitly co-ordinating this measurement would be expensive, but
1529 * fortunately the sum of each cpus contribution forms a usable
1530 * lower-bound on the true value.
1532 * Consider the aggregate of 2 contributions. Either they are disjoint
1533 * (and the sum represents true value) or they are disjoint and we are
1534 * understating by the aggregate of their overlap.
1536 * Extending this to N cpus, for a given overlap, the maximum amount we
1537 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1538 * cpus that overlap for this interval and w_i is the interval width.
1540 * On a small machine; the first term is well-bounded which bounds the
1541 * total error since w_i is a subset of the period. Whereas on a
1542 * larger machine, while this first term can be larger, if w_i is the
1543 * of consequential size guaranteed to see n_i*w_i quickly converge to
1544 * our upper bound of 1-cpu.
1546 runnable_avg = atomic_read(&tg->runnable_avg);
1547 if (runnable_avg < NICE_0_LOAD) {
1548 se->avg.load_avg_contrib *= runnable_avg;
1549 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1553 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1554 int force_update) {}
1555 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1556 struct cfs_rq *cfs_rq) {}
1557 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1560 static inline void __update_task_entity_contrib(struct sched_entity *se)
1564 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1565 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1566 contrib /= (se->avg.runnable_avg_period + 1);
1567 se->avg.load_avg_contrib = scale_load(contrib);
1568 trace_sched_task_load_contrib(task_of(se), se->avg.load_avg_contrib);
1569 contrib = se->avg.runnable_avg_sum * scale_load_down(NICE_0_LOAD);
1570 contrib /= (se->avg.runnable_avg_period + 1);
1571 se->avg.load_avg_ratio = scale_load(contrib);
1572 trace_sched_task_runnable_ratio(task_of(se), se->avg.load_avg_ratio);
1575 /* Compute the current contribution to load_avg by se, return any delta */
1576 static long __update_entity_load_avg_contrib(struct sched_entity *se, long *ratio)
1578 long old_contrib = se->avg.load_avg_contrib;
1579 long old_ratio = se->avg.load_avg_ratio;
1581 if (entity_is_task(se)) {
1582 __update_task_entity_contrib(se);
1584 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1585 __update_group_entity_contrib(se);
1589 *ratio = se->avg.load_avg_ratio - old_ratio;
1590 return se->avg.load_avg_contrib - old_contrib;
1593 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1596 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1597 cfs_rq->blocked_load_avg -= load_contrib;
1599 cfs_rq->blocked_load_avg = 0;
1602 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1604 /* Update a sched_entity's runnable average */
1605 static inline void update_entity_load_avg(struct sched_entity *se,
1608 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1609 long contrib_delta, ratio_delta;
1611 int cpu = -1; /* not used in normal case */
1613 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1614 cpu = cfs_rq->rq->cpu;
1617 * For a group entity we need to use their owned cfs_rq_clock_task() in
1618 * case they are the parent of a throttled hierarchy.
1620 if (entity_is_task(se))
1621 now = cfs_rq_clock_task(cfs_rq);
1623 now = cfs_rq_clock_task(group_cfs_rq(se));
1625 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq,
1626 cfs_rq->curr == se, cpu))
1629 contrib_delta = __update_entity_load_avg_contrib(se, &ratio_delta);
1635 cfs_rq->runnable_load_avg += contrib_delta;
1636 rq_of(cfs_rq)->avg.load_avg_ratio += ratio_delta;
1638 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1643 * Decay the load contributed by all blocked children and account this so that
1644 * their contribution may appropriately discounted when they wake up.
1646 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1648 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1651 decays = now - cfs_rq->last_decay;
1652 if (!decays && !force_update)
1655 if (atomic64_read(&cfs_rq->removed_load)) {
1656 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1657 subtract_blocked_load_contrib(cfs_rq, removed_load);
1661 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1663 atomic64_add(decays, &cfs_rq->decay_counter);
1664 cfs_rq->last_decay = now;
1667 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1670 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1672 int cpu = -1; /* not used in normal case */
1674 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
1677 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable,
1679 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1680 trace_sched_rq_runnable_ratio(cpu_of(rq), rq->avg.load_avg_ratio);
1681 trace_sched_rq_runnable_load(cpu_of(rq), rq->cfs.runnable_load_avg);
1682 trace_sched_rq_nr_running(cpu_of(rq), rq->nr_running, rq->nr_iowait.counter);
1685 /* Add the load generated by se into cfs_rq's child load-average */
1686 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1687 struct sched_entity *se,
1691 * We track migrations using entity decay_count <= 0, on a wake-up
1692 * migration we use a negative decay count to track the remote decays
1693 * accumulated while sleeping.
1695 if (unlikely(se->avg.decay_count <= 0)) {
1696 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1697 if (se->avg.decay_count) {
1699 * In a wake-up migration we have to approximate the
1700 * time sleeping. This is because we can't synchronize
1701 * clock_task between the two cpus, and it is not
1702 * guaranteed to be read-safe. Instead, we can
1703 * approximate this using our carried decays, which are
1704 * explicitly atomically readable.
1706 se->avg.last_runnable_update -= (-se->avg.decay_count)
1708 update_entity_load_avg(se, 0);
1709 /* Indicate that we're now synchronized and on-rq */
1710 se->avg.decay_count = 0;
1714 __synchronize_entity_decay(se);
1717 /* migrated tasks did not contribute to our blocked load */
1719 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1720 update_entity_load_avg(se, 0);
1723 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1724 rq_of(cfs_rq)->avg.load_avg_ratio += se->avg.load_avg_ratio;
1726 /* we force update consideration on load-balancer moves */
1727 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1731 * Remove se's load from this cfs_rq child load-average, if the entity is
1732 * transitioning to a blocked state we track its projected decay using
1735 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1736 struct sched_entity *se,
1739 update_entity_load_avg(se, 1);
1740 /* we force update consideration on load-balancer moves */
1741 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1743 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1744 rq_of(cfs_rq)->avg.load_avg_ratio -= se->avg.load_avg_ratio;
1747 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1748 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1749 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1753 * Update the rq's load with the elapsed running time before entering
1754 * idle. if the last scheduled task is not a CFS task, idle_enter will
1755 * be the only way to update the runnable statistic.
1757 void idle_enter_fair(struct rq *this_rq)
1759 update_rq_runnable_avg(this_rq, 1);
1763 * Update the rq's load with the elapsed idle time before a task is
1764 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1765 * be the only way to update the runnable statistic.
1767 void idle_exit_fair(struct rq *this_rq)
1769 update_rq_runnable_avg(this_rq, 0);
1773 static inline void update_entity_load_avg(struct sched_entity *se,
1774 int update_cfs_rq) {}
1775 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1776 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1777 struct sched_entity *se,
1779 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1780 struct sched_entity *se,
1782 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1783 int force_update) {}
1786 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1788 #ifdef CONFIG_SCHEDSTATS
1789 struct task_struct *tsk = NULL;
1791 if (entity_is_task(se))
1794 if (se->statistics.sleep_start) {
1795 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1800 if (unlikely(delta > se->statistics.sleep_max))
1801 se->statistics.sleep_max = delta;
1803 se->statistics.sleep_start = 0;
1804 se->statistics.sum_sleep_runtime += delta;
1807 account_scheduler_latency(tsk, delta >> 10, 1);
1808 trace_sched_stat_sleep(tsk, delta);
1811 if (se->statistics.block_start) {
1812 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1817 if (unlikely(delta > se->statistics.block_max))
1818 se->statistics.block_max = delta;
1820 se->statistics.block_start = 0;
1821 se->statistics.sum_sleep_runtime += delta;
1824 if (tsk->in_iowait) {
1825 se->statistics.iowait_sum += delta;
1826 se->statistics.iowait_count++;
1827 trace_sched_stat_iowait(tsk, delta);
1830 trace_sched_stat_blocked(tsk, delta);
1833 * Blocking time is in units of nanosecs, so shift by
1834 * 20 to get a milliseconds-range estimation of the
1835 * amount of time that the task spent sleeping:
1837 if (unlikely(prof_on == SLEEP_PROFILING)) {
1838 profile_hits(SLEEP_PROFILING,
1839 (void *)get_wchan(tsk),
1842 account_scheduler_latency(tsk, delta >> 10, 0);
1848 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1850 #ifdef CONFIG_SCHED_DEBUG
1851 s64 d = se->vruntime - cfs_rq->min_vruntime;
1856 if (d > 3*sysctl_sched_latency)
1857 schedstat_inc(cfs_rq, nr_spread_over);
1862 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1864 u64 vruntime = cfs_rq->min_vruntime;
1867 * The 'current' period is already promised to the current tasks,
1868 * however the extra weight of the new task will slow them down a
1869 * little, place the new task so that it fits in the slot that
1870 * stays open at the end.
1872 if (initial && sched_feat(START_DEBIT))
1873 vruntime += sched_vslice(cfs_rq, se);
1875 /* sleeps up to a single latency don't count. */
1877 unsigned long thresh = sysctl_sched_latency;
1880 * Halve their sleep time's effect, to allow
1881 * for a gentler effect of sleepers:
1883 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1889 /* ensure we never gain time by being placed backwards. */
1890 se->vruntime = max_vruntime(se->vruntime, vruntime);
1893 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1896 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1899 * Update the normalized vruntime before updating min_vruntime
1900 * through callig update_curr().
1902 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1903 se->vruntime += cfs_rq->min_vruntime;
1906 * Update run-time statistics of the 'current'.
1908 update_curr(cfs_rq);
1909 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1910 account_entity_enqueue(cfs_rq, se);
1911 update_cfs_shares(cfs_rq);
1913 if (flags & ENQUEUE_WAKEUP) {
1914 place_entity(cfs_rq, se, 0);
1915 enqueue_sleeper(cfs_rq, se);
1918 update_stats_enqueue(cfs_rq, se);
1919 check_spread(cfs_rq, se);
1920 if (se != cfs_rq->curr)
1921 __enqueue_entity(cfs_rq, se);
1924 if (cfs_rq->nr_running == 1) {
1925 list_add_leaf_cfs_rq(cfs_rq);
1926 check_enqueue_throttle(cfs_rq);
1930 static void __clear_buddies_last(struct sched_entity *se)
1932 for_each_sched_entity(se) {
1933 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1934 if (cfs_rq->last == se)
1935 cfs_rq->last = NULL;
1941 static void __clear_buddies_next(struct sched_entity *se)
1943 for_each_sched_entity(se) {
1944 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1945 if (cfs_rq->next == se)
1946 cfs_rq->next = NULL;
1952 static void __clear_buddies_skip(struct sched_entity *se)
1954 for_each_sched_entity(se) {
1955 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1956 if (cfs_rq->skip == se)
1957 cfs_rq->skip = NULL;
1963 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1965 if (cfs_rq->last == se)
1966 __clear_buddies_last(se);
1968 if (cfs_rq->next == se)
1969 __clear_buddies_next(se);
1971 if (cfs_rq->skip == se)
1972 __clear_buddies_skip(se);
1975 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1978 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1981 * Update run-time statistics of the 'current'.
1983 update_curr(cfs_rq);
1984 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1986 update_stats_dequeue(cfs_rq, se);
1987 if (flags & DEQUEUE_SLEEP) {
1988 #ifdef CONFIG_SCHEDSTATS
1989 if (entity_is_task(se)) {
1990 struct task_struct *tsk = task_of(se);
1992 if (tsk->state & TASK_INTERRUPTIBLE)
1993 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1994 if (tsk->state & TASK_UNINTERRUPTIBLE)
1995 se->statistics.block_start = rq_of(cfs_rq)->clock;
2000 clear_buddies(cfs_rq, se);
2002 if (se != cfs_rq->curr)
2003 __dequeue_entity(cfs_rq, se);
2005 account_entity_dequeue(cfs_rq, se);
2008 * Normalize the entity after updating the min_vruntime because the
2009 * update can refer to the ->curr item and we need to reflect this
2010 * movement in our normalized position.
2012 if (!(flags & DEQUEUE_SLEEP))
2013 se->vruntime -= cfs_rq->min_vruntime;
2015 /* return excess runtime on last dequeue */
2016 return_cfs_rq_runtime(cfs_rq);
2018 update_min_vruntime(cfs_rq);
2019 update_cfs_shares(cfs_rq);
2023 * Preempt the current task with a newly woken task if needed:
2026 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2028 unsigned long ideal_runtime, delta_exec;
2029 struct sched_entity *se;
2032 ideal_runtime = sched_slice(cfs_rq, curr);
2033 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2034 if (delta_exec > ideal_runtime) {
2035 resched_task(rq_of(cfs_rq)->curr);
2037 * The current task ran long enough, ensure it doesn't get
2038 * re-elected due to buddy favours.
2040 clear_buddies(cfs_rq, curr);
2045 * Ensure that a task that missed wakeup preemption by a
2046 * narrow margin doesn't have to wait for a full slice.
2047 * This also mitigates buddy induced latencies under load.
2049 if (delta_exec < sysctl_sched_min_granularity)
2052 se = __pick_first_entity(cfs_rq);
2053 delta = curr->vruntime - se->vruntime;
2058 if (delta > ideal_runtime)
2059 resched_task(rq_of(cfs_rq)->curr);
2063 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2065 /* 'current' is not kept within the tree. */
2068 * Any task has to be enqueued before it get to execute on
2069 * a CPU. So account for the time it spent waiting on the
2072 update_stats_wait_end(cfs_rq, se);
2073 __dequeue_entity(cfs_rq, se);
2074 update_entity_load_avg(se, 1);
2077 update_stats_curr_start(cfs_rq, se);
2079 #ifdef CONFIG_SCHEDSTATS
2081 * Track our maximum slice length, if the CPU's load is at
2082 * least twice that of our own weight (i.e. dont track it
2083 * when there are only lesser-weight tasks around):
2085 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2086 se->statistics.slice_max = max(se->statistics.slice_max,
2087 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2090 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2094 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2097 * Pick the next process, keeping these things in mind, in this order:
2098 * 1) keep things fair between processes/task groups
2099 * 2) pick the "next" process, since someone really wants that to run
2100 * 3) pick the "last" process, for cache locality
2101 * 4) do not run the "skip" process, if something else is available
2103 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2105 struct sched_entity *se = __pick_first_entity(cfs_rq);
2106 struct sched_entity *left = se;
2109 * Avoid running the skip buddy, if running something else can
2110 * be done without getting too unfair.
2112 if (cfs_rq->skip == se) {
2113 struct sched_entity *second = __pick_next_entity(se);
2114 if (second && wakeup_preempt_entity(second, left) < 1)
2119 * Prefer last buddy, try to return the CPU to a preempted task.
2121 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2125 * Someone really wants this to run. If it's not unfair, run it.
2127 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2130 clear_buddies(cfs_rq, se);
2135 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2137 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2140 * If still on the runqueue then deactivate_task()
2141 * was not called and update_curr() has to be done:
2144 update_curr(cfs_rq);
2146 /* throttle cfs_rqs exceeding runtime */
2147 check_cfs_rq_runtime(cfs_rq);
2149 check_spread(cfs_rq, prev);
2151 update_stats_wait_start(cfs_rq, prev);
2152 /* Put 'current' back into the tree. */
2153 __enqueue_entity(cfs_rq, prev);
2154 /* in !on_rq case, update occurred at dequeue */
2155 update_entity_load_avg(prev, 1);
2157 cfs_rq->curr = NULL;
2161 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2164 * Update run-time statistics of the 'current'.
2166 update_curr(cfs_rq);
2169 * Ensure that runnable average is periodically updated.
2171 update_entity_load_avg(curr, 1);
2172 update_cfs_rq_blocked_load(cfs_rq, 1);
2173 update_cfs_shares(cfs_rq);
2175 #ifdef CONFIG_SCHED_HRTICK
2177 * queued ticks are scheduled to match the slice, so don't bother
2178 * validating it and just reschedule.
2181 resched_task(rq_of(cfs_rq)->curr);
2185 * don't let the period tick interfere with the hrtick preemption
2187 if (!sched_feat(DOUBLE_TICK) &&
2188 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2192 if (cfs_rq->nr_running > 1)
2193 check_preempt_tick(cfs_rq, curr);
2197 /**************************************************
2198 * CFS bandwidth control machinery
2201 #ifdef CONFIG_CFS_BANDWIDTH
2203 #ifdef HAVE_JUMP_LABEL
2204 static struct static_key __cfs_bandwidth_used;
2206 static inline bool cfs_bandwidth_used(void)
2208 return static_key_false(&__cfs_bandwidth_used);
2211 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2213 /* only need to count groups transitioning between enabled/!enabled */
2214 if (enabled && !was_enabled)
2215 static_key_slow_inc(&__cfs_bandwidth_used);
2216 else if (!enabled && was_enabled)
2217 static_key_slow_dec(&__cfs_bandwidth_used);
2219 #else /* HAVE_JUMP_LABEL */
2220 static bool cfs_bandwidth_used(void)
2225 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2226 #endif /* HAVE_JUMP_LABEL */
2229 * default period for cfs group bandwidth.
2230 * default: 0.1s, units: nanoseconds
2232 static inline u64 default_cfs_period(void)
2234 return 100000000ULL;
2237 static inline u64 sched_cfs_bandwidth_slice(void)
2239 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2243 * Replenish runtime according to assigned quota and update expiration time.
2244 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2245 * additional synchronization around rq->lock.
2247 * requires cfs_b->lock
2249 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2253 if (cfs_b->quota == RUNTIME_INF)
2256 now = sched_clock_cpu(smp_processor_id());
2257 cfs_b->runtime = cfs_b->quota;
2258 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2261 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2263 return &tg->cfs_bandwidth;
2266 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2267 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2269 if (unlikely(cfs_rq->throttle_count))
2270 return cfs_rq->throttled_clock_task;
2272 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2275 /* returns 0 on failure to allocate runtime */
2276 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2278 struct task_group *tg = cfs_rq->tg;
2279 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2280 u64 amount = 0, min_amount, expires;
2282 /* note: this is a positive sum as runtime_remaining <= 0 */
2283 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2285 raw_spin_lock(&cfs_b->lock);
2286 if (cfs_b->quota == RUNTIME_INF)
2287 amount = min_amount;
2290 * If the bandwidth pool has become inactive, then at least one
2291 * period must have elapsed since the last consumption.
2292 * Refresh the global state and ensure bandwidth timer becomes
2295 if (!cfs_b->timer_active) {
2296 __refill_cfs_bandwidth_runtime(cfs_b);
2297 __start_cfs_bandwidth(cfs_b);
2300 if (cfs_b->runtime > 0) {
2301 amount = min(cfs_b->runtime, min_amount);
2302 cfs_b->runtime -= amount;
2306 expires = cfs_b->runtime_expires;
2307 raw_spin_unlock(&cfs_b->lock);
2309 cfs_rq->runtime_remaining += amount;
2311 * we may have advanced our local expiration to account for allowed
2312 * spread between our sched_clock and the one on which runtime was
2315 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2316 cfs_rq->runtime_expires = expires;
2318 return cfs_rq->runtime_remaining > 0;
2322 * Note: This depends on the synchronization provided by sched_clock and the
2323 * fact that rq->clock snapshots this value.
2325 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2327 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2328 struct rq *rq = rq_of(cfs_rq);
2330 /* if the deadline is ahead of our clock, nothing to do */
2331 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2334 if (cfs_rq->runtime_remaining < 0)
2338 * If the local deadline has passed we have to consider the
2339 * possibility that our sched_clock is 'fast' and the global deadline
2340 * has not truly expired.
2342 * Fortunately we can check determine whether this the case by checking
2343 * whether the global deadline has advanced.
2346 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2347 /* extend local deadline, drift is bounded above by 2 ticks */
2348 cfs_rq->runtime_expires += TICK_NSEC;
2350 /* global deadline is ahead, expiration has passed */
2351 cfs_rq->runtime_remaining = 0;
2355 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2356 unsigned long delta_exec)
2358 /* dock delta_exec before expiring quota (as it could span periods) */
2359 cfs_rq->runtime_remaining -= delta_exec;
2360 expire_cfs_rq_runtime(cfs_rq);
2362 if (likely(cfs_rq->runtime_remaining > 0))
2366 * if we're unable to extend our runtime we resched so that the active
2367 * hierarchy can be throttled
2369 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2370 resched_task(rq_of(cfs_rq)->curr);
2373 static __always_inline
2374 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2376 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2379 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2382 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2384 return cfs_bandwidth_used() && cfs_rq->throttled;
2387 /* check whether cfs_rq, or any parent, is throttled */
2388 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2390 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2394 * Ensure that neither of the group entities corresponding to src_cpu or
2395 * dest_cpu are members of a throttled hierarchy when performing group
2396 * load-balance operations.
2398 static inline int throttled_lb_pair(struct task_group *tg,
2399 int src_cpu, int dest_cpu)
2401 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2403 src_cfs_rq = tg->cfs_rq[src_cpu];
2404 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2406 return throttled_hierarchy(src_cfs_rq) ||
2407 throttled_hierarchy(dest_cfs_rq);
2410 /* updated child weight may affect parent so we have to do this bottom up */
2411 static int tg_unthrottle_up(struct task_group *tg, void *data)
2413 struct rq *rq = data;
2414 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2416 cfs_rq->throttle_count--;
2418 if (!cfs_rq->throttle_count) {
2419 /* adjust cfs_rq_clock_task() */
2420 cfs_rq->throttled_clock_task_time += rq->clock_task -
2421 cfs_rq->throttled_clock_task;
2428 static int tg_throttle_down(struct task_group *tg, void *data)
2430 struct rq *rq = data;
2431 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2433 /* group is entering throttled state, stop time */
2434 if (!cfs_rq->throttle_count)
2435 cfs_rq->throttled_clock_task = rq->clock_task;
2436 cfs_rq->throttle_count++;
2441 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2443 struct rq *rq = rq_of(cfs_rq);
2444 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2445 struct sched_entity *se;
2446 long task_delta, dequeue = 1;
2448 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2450 /* freeze hierarchy runnable averages while throttled */
2452 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2455 task_delta = cfs_rq->h_nr_running;
2456 for_each_sched_entity(se) {
2457 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2458 /* throttled entity or throttle-on-deactivate */
2463 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2464 qcfs_rq->h_nr_running -= task_delta;
2466 if (qcfs_rq->load.weight)
2471 rq->nr_running -= task_delta;
2473 cfs_rq->throttled = 1;
2474 cfs_rq->throttled_clock = rq->clock;
2475 raw_spin_lock(&cfs_b->lock);
2476 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2477 raw_spin_unlock(&cfs_b->lock);
2480 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2482 struct rq *rq = rq_of(cfs_rq);
2483 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2484 struct sched_entity *se;
2488 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2490 cfs_rq->throttled = 0;
2491 raw_spin_lock(&cfs_b->lock);
2492 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2493 list_del_rcu(&cfs_rq->throttled_list);
2494 raw_spin_unlock(&cfs_b->lock);
2496 update_rq_clock(rq);
2497 /* update hierarchical throttle state */
2498 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2500 if (!cfs_rq->load.weight)
2503 task_delta = cfs_rq->h_nr_running;
2504 for_each_sched_entity(se) {
2508 cfs_rq = cfs_rq_of(se);
2510 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2511 cfs_rq->h_nr_running += task_delta;
2513 if (cfs_rq_throttled(cfs_rq))
2518 rq->nr_running += task_delta;
2520 /* determine whether we need to wake up potentially idle cpu */
2521 if (rq->curr == rq->idle && rq->cfs.nr_running)
2522 resched_task(rq->curr);
2525 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2526 u64 remaining, u64 expires)
2528 struct cfs_rq *cfs_rq;
2529 u64 runtime = remaining;
2532 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2534 struct rq *rq = rq_of(cfs_rq);
2536 raw_spin_lock(&rq->lock);
2537 if (!cfs_rq_throttled(cfs_rq))
2540 runtime = -cfs_rq->runtime_remaining + 1;
2541 if (runtime > remaining)
2542 runtime = remaining;
2543 remaining -= runtime;
2545 cfs_rq->runtime_remaining += runtime;
2546 cfs_rq->runtime_expires = expires;
2548 /* we check whether we're throttled above */
2549 if (cfs_rq->runtime_remaining > 0)
2550 unthrottle_cfs_rq(cfs_rq);
2553 raw_spin_unlock(&rq->lock);
2564 * Responsible for refilling a task_group's bandwidth and unthrottling its
2565 * cfs_rqs as appropriate. If there has been no activity within the last
2566 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2567 * used to track this state.
2569 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2571 u64 runtime, runtime_expires;
2572 int idle = 1, throttled;
2574 raw_spin_lock(&cfs_b->lock);
2575 /* no need to continue the timer with no bandwidth constraint */
2576 if (cfs_b->quota == RUNTIME_INF)
2579 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2580 /* idle depends on !throttled (for the case of a large deficit) */
2581 idle = cfs_b->idle && !throttled;
2582 cfs_b->nr_periods += overrun;
2584 /* if we're going inactive then everything else can be deferred */
2588 __refill_cfs_bandwidth_runtime(cfs_b);
2591 /* mark as potentially idle for the upcoming period */
2596 /* account preceding periods in which throttling occurred */
2597 cfs_b->nr_throttled += overrun;
2600 * There are throttled entities so we must first use the new bandwidth
2601 * to unthrottle them before making it generally available. This
2602 * ensures that all existing debts will be paid before a new cfs_rq is
2605 runtime = cfs_b->runtime;
2606 runtime_expires = cfs_b->runtime_expires;
2610 * This check is repeated as we are holding onto the new bandwidth
2611 * while we unthrottle. This can potentially race with an unthrottled
2612 * group trying to acquire new bandwidth from the global pool.
2614 while (throttled && runtime > 0) {
2615 raw_spin_unlock(&cfs_b->lock);
2616 /* we can't nest cfs_b->lock while distributing bandwidth */
2617 runtime = distribute_cfs_runtime(cfs_b, runtime,
2619 raw_spin_lock(&cfs_b->lock);
2621 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2624 /* return (any) remaining runtime */
2625 cfs_b->runtime = runtime;
2627 * While we are ensured activity in the period following an
2628 * unthrottle, this also covers the case in which the new bandwidth is
2629 * insufficient to cover the existing bandwidth deficit. (Forcing the
2630 * timer to remain active while there are any throttled entities.)
2635 cfs_b->timer_active = 0;
2636 raw_spin_unlock(&cfs_b->lock);
2641 /* a cfs_rq won't donate quota below this amount */
2642 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2643 /* minimum remaining period time to redistribute slack quota */
2644 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2645 /* how long we wait to gather additional slack before distributing */
2646 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2648 /* are we near the end of the current quota period? */
2649 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2651 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2654 /* if the call-back is running a quota refresh is already occurring */
2655 if (hrtimer_callback_running(refresh_timer))
2658 /* is a quota refresh about to occur? */
2659 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2660 if (remaining < min_expire)
2666 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2668 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2670 /* if there's a quota refresh soon don't bother with slack */
2671 if (runtime_refresh_within(cfs_b, min_left))
2674 start_bandwidth_timer(&cfs_b->slack_timer,
2675 ns_to_ktime(cfs_bandwidth_slack_period));
2678 /* we know any runtime found here is valid as update_curr() precedes return */
2679 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2681 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2682 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2684 if (slack_runtime <= 0)
2687 raw_spin_lock(&cfs_b->lock);
2688 if (cfs_b->quota != RUNTIME_INF &&
2689 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2690 cfs_b->runtime += slack_runtime;
2692 /* we are under rq->lock, defer unthrottling using a timer */
2693 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2694 !list_empty(&cfs_b->throttled_cfs_rq))
2695 start_cfs_slack_bandwidth(cfs_b);
2697 raw_spin_unlock(&cfs_b->lock);
2699 /* even if it's not valid for return we don't want to try again */
2700 cfs_rq->runtime_remaining -= slack_runtime;
2703 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2705 if (!cfs_bandwidth_used())
2708 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2711 __return_cfs_rq_runtime(cfs_rq);
2715 * This is done with a timer (instead of inline with bandwidth return) since
2716 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2718 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2720 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2723 /* confirm we're still not at a refresh boundary */
2724 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2727 raw_spin_lock(&cfs_b->lock);
2728 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2729 runtime = cfs_b->runtime;
2732 expires = cfs_b->runtime_expires;
2733 raw_spin_unlock(&cfs_b->lock);
2738 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2740 raw_spin_lock(&cfs_b->lock);
2741 if (expires == cfs_b->runtime_expires)
2742 cfs_b->runtime = runtime;
2743 raw_spin_unlock(&cfs_b->lock);
2747 * When a group wakes up we want to make sure that its quota is not already
2748 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2749 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2751 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2753 if (!cfs_bandwidth_used())
2756 /* an active group must be handled by the update_curr()->put() path */
2757 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2760 /* ensure the group is not already throttled */
2761 if (cfs_rq_throttled(cfs_rq))
2764 /* update runtime allocation */
2765 account_cfs_rq_runtime(cfs_rq, 0);
2766 if (cfs_rq->runtime_remaining <= 0)
2767 throttle_cfs_rq(cfs_rq);
2770 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2771 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2773 if (!cfs_bandwidth_used())
2776 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2780 * it's possible for a throttled entity to be forced into a running
2781 * state (e.g. set_curr_task), in this case we're finished.
2783 if (cfs_rq_throttled(cfs_rq))
2786 throttle_cfs_rq(cfs_rq);
2789 static inline u64 default_cfs_period(void);
2790 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2791 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2793 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2795 struct cfs_bandwidth *cfs_b =
2796 container_of(timer, struct cfs_bandwidth, slack_timer);
2797 do_sched_cfs_slack_timer(cfs_b);
2799 return HRTIMER_NORESTART;
2802 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2804 struct cfs_bandwidth *cfs_b =
2805 container_of(timer, struct cfs_bandwidth, period_timer);
2811 now = hrtimer_cb_get_time(timer);
2812 overrun = hrtimer_forward(timer, now, cfs_b->period);
2817 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2820 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2823 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2825 raw_spin_lock_init(&cfs_b->lock);
2827 cfs_b->quota = RUNTIME_INF;
2828 cfs_b->period = ns_to_ktime(default_cfs_period());
2830 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2831 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2832 cfs_b->period_timer.function = sched_cfs_period_timer;
2833 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2834 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2837 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2839 cfs_rq->runtime_enabled = 0;
2840 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2843 /* requires cfs_b->lock, may release to reprogram timer */
2844 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2847 * The timer may be active because we're trying to set a new bandwidth
2848 * period or because we're racing with the tear-down path
2849 * (timer_active==0 becomes visible before the hrtimer call-back
2850 * terminates). In either case we ensure that it's re-programmed
2852 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2853 raw_spin_unlock(&cfs_b->lock);
2854 /* ensure cfs_b->lock is available while we wait */
2855 hrtimer_cancel(&cfs_b->period_timer);
2857 raw_spin_lock(&cfs_b->lock);
2858 /* if someone else restarted the timer then we're done */
2859 if (cfs_b->timer_active)
2863 cfs_b->timer_active = 1;
2864 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2867 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2869 hrtimer_cancel(&cfs_b->period_timer);
2870 hrtimer_cancel(&cfs_b->slack_timer);
2873 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2875 struct cfs_rq *cfs_rq;
2877 for_each_leaf_cfs_rq(rq, cfs_rq) {
2878 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2880 if (!cfs_rq->runtime_enabled)
2884 * clock_task is not advancing so we just need to make sure
2885 * there's some valid quota amount
2887 cfs_rq->runtime_remaining = cfs_b->quota;
2888 if (cfs_rq_throttled(cfs_rq))
2889 unthrottle_cfs_rq(cfs_rq);
2893 #else /* CONFIG_CFS_BANDWIDTH */
2894 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2896 return rq_of(cfs_rq)->clock_task;
2899 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2900 unsigned long delta_exec) {}
2901 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2902 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2903 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2905 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2910 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2915 static inline int throttled_lb_pair(struct task_group *tg,
2916 int src_cpu, int dest_cpu)
2921 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2923 #ifdef CONFIG_FAIR_GROUP_SCHED
2924 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2927 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2931 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2932 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2934 #endif /* CONFIG_CFS_BANDWIDTH */
2936 /**************************************************
2937 * CFS operations on tasks:
2940 #ifdef CONFIG_SCHED_HRTICK
2941 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2943 struct sched_entity *se = &p->se;
2944 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2946 WARN_ON(task_rq(p) != rq);
2948 if (cfs_rq->nr_running > 1) {
2949 u64 slice = sched_slice(cfs_rq, se);
2950 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2951 s64 delta = slice - ran;
2960 * Don't schedule slices shorter than 10000ns, that just
2961 * doesn't make sense. Rely on vruntime for fairness.
2964 delta = max_t(s64, 10000LL, delta);
2966 hrtick_start(rq, delta);
2971 * called from enqueue/dequeue and updates the hrtick when the
2972 * current task is from our class and nr_running is low enough
2975 static void hrtick_update(struct rq *rq)
2977 struct task_struct *curr = rq->curr;
2979 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2982 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2983 hrtick_start_fair(rq, curr);
2985 #else /* !CONFIG_SCHED_HRTICK */
2987 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2991 static inline void hrtick_update(struct rq *rq)
2997 * The enqueue_task method is called before nr_running is
2998 * increased. Here we update the fair scheduling stats and
2999 * then put the task into the rbtree:
3002 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3004 struct cfs_rq *cfs_rq;
3005 struct sched_entity *se = &p->se;
3007 for_each_sched_entity(se) {
3010 cfs_rq = cfs_rq_of(se);
3011 enqueue_entity(cfs_rq, se, flags);
3014 * end evaluation on encountering a throttled cfs_rq
3016 * note: in the case of encountering a throttled cfs_rq we will
3017 * post the final h_nr_running increment below.
3019 if (cfs_rq_throttled(cfs_rq))
3021 cfs_rq->h_nr_running++;
3023 flags = ENQUEUE_WAKEUP;
3026 for_each_sched_entity(se) {
3027 cfs_rq = cfs_rq_of(se);
3028 cfs_rq->h_nr_running++;
3030 if (cfs_rq_throttled(cfs_rq))
3033 update_cfs_shares(cfs_rq);
3034 update_entity_load_avg(se, 1);
3038 update_rq_runnable_avg(rq, rq->nr_running);
3044 static void set_next_buddy(struct sched_entity *se);
3047 * The dequeue_task method is called before nr_running is
3048 * decreased. We remove the task from the rbtree and
3049 * update the fair scheduling stats:
3051 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3053 struct cfs_rq *cfs_rq;
3054 struct sched_entity *se = &p->se;
3055 int task_sleep = flags & DEQUEUE_SLEEP;
3057 for_each_sched_entity(se) {
3058 cfs_rq = cfs_rq_of(se);
3059 dequeue_entity(cfs_rq, se, flags);
3062 * end evaluation on encountering a throttled cfs_rq
3064 * note: in the case of encountering a throttled cfs_rq we will
3065 * post the final h_nr_running decrement below.
3067 if (cfs_rq_throttled(cfs_rq))
3069 cfs_rq->h_nr_running--;
3071 /* Don't dequeue parent if it has other entities besides us */
3072 if (cfs_rq->load.weight) {
3074 * Bias pick_next to pick a task from this cfs_rq, as
3075 * p is sleeping when it is within its sched_slice.
3077 if (task_sleep && parent_entity(se))
3078 set_next_buddy(parent_entity(se));
3080 /* avoid re-evaluating load for this entity */
3081 se = parent_entity(se);
3084 flags |= DEQUEUE_SLEEP;
3087 for_each_sched_entity(se) {
3088 cfs_rq = cfs_rq_of(se);
3089 cfs_rq->h_nr_running--;
3091 if (cfs_rq_throttled(cfs_rq))
3094 update_cfs_shares(cfs_rq);
3095 update_entity_load_avg(se, 1);
3100 update_rq_runnable_avg(rq, 1);
3106 /* Used instead of source_load when we know the type == 0 */
3107 static unsigned long weighted_cpuload(const int cpu)
3109 return cpu_rq(cpu)->load.weight;
3113 * Return a low guess at the load of a migration-source cpu weighted
3114 * according to the scheduling class and "nice" value.
3116 * We want to under-estimate the load of migration sources, to
3117 * balance conservatively.
3119 static unsigned long source_load(int cpu, int type)
3121 struct rq *rq = cpu_rq(cpu);
3122 unsigned long total = weighted_cpuload(cpu);
3124 if (type == 0 || !sched_feat(LB_BIAS))
3127 return min(rq->cpu_load[type-1], total);
3131 * Return a high guess at the load of a migration-target cpu weighted
3132 * according to the scheduling class and "nice" value.
3134 static unsigned long target_load(int cpu, int type)
3136 struct rq *rq = cpu_rq(cpu);
3137 unsigned long total = weighted_cpuload(cpu);
3139 if (type == 0 || !sched_feat(LB_BIAS))
3142 return max(rq->cpu_load[type-1], total);
3145 static unsigned long power_of(int cpu)
3147 return cpu_rq(cpu)->cpu_power;
3150 static unsigned long cpu_avg_load_per_task(int cpu)
3152 struct rq *rq = cpu_rq(cpu);
3153 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3156 return rq->load.weight / nr_running;
3162 static void task_waking_fair(struct task_struct *p)
3164 struct sched_entity *se = &p->se;
3165 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3168 #ifndef CONFIG_64BIT
3169 u64 min_vruntime_copy;
3172 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3174 min_vruntime = cfs_rq->min_vruntime;
3175 } while (min_vruntime != min_vruntime_copy);
3177 min_vruntime = cfs_rq->min_vruntime;
3180 se->vruntime -= min_vruntime;
3183 #ifdef CONFIG_FAIR_GROUP_SCHED
3185 * effective_load() calculates the load change as seen from the root_task_group
3187 * Adding load to a group doesn't make a group heavier, but can cause movement
3188 * of group shares between cpus. Assuming the shares were perfectly aligned one
3189 * can calculate the shift in shares.
3191 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3192 * on this @cpu and results in a total addition (subtraction) of @wg to the
3193 * total group weight.
3195 * Given a runqueue weight distribution (rw_i) we can compute a shares
3196 * distribution (s_i) using:
3198 * s_i = rw_i / \Sum rw_j (1)
3200 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3201 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3202 * shares distribution (s_i):
3204 * rw_i = { 2, 4, 1, 0 }
3205 * s_i = { 2/7, 4/7, 1/7, 0 }
3207 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3208 * task used to run on and the CPU the waker is running on), we need to
3209 * compute the effect of waking a task on either CPU and, in case of a sync
3210 * wakeup, compute the effect of the current task going to sleep.
3212 * So for a change of @wl to the local @cpu with an overall group weight change
3213 * of @wl we can compute the new shares distribution (s'_i) using:
3215 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3217 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3218 * differences in waking a task to CPU 0. The additional task changes the
3219 * weight and shares distributions like:
3221 * rw'_i = { 3, 4, 1, 0 }
3222 * s'_i = { 3/8, 4/8, 1/8, 0 }
3224 * We can then compute the difference in effective weight by using:
3226 * dw_i = S * (s'_i - s_i) (3)
3228 * Where 'S' is the group weight as seen by its parent.
3230 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3231 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3232 * 4/7) times the weight of the group.
3234 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3236 struct sched_entity *se = tg->se[cpu];
3238 if (!tg->parent) /* the trivial, non-cgroup case */
3241 for_each_sched_entity(se) {
3247 * W = @wg + \Sum rw_j
3249 W = wg + calc_tg_weight(tg, se->my_q);
3254 w = se->my_q->load.weight + wl;
3257 * wl = S * s'_i; see (2)
3260 wl = (w * tg->shares) / W;
3265 * Per the above, wl is the new se->load.weight value; since
3266 * those are clipped to [MIN_SHARES, ...) do so now. See
3267 * calc_cfs_shares().
3269 if (wl < MIN_SHARES)
3273 * wl = dw_i = S * (s'_i - s_i); see (3)
3275 wl -= se->load.weight;
3278 * Recursively apply this logic to all parent groups to compute
3279 * the final effective load change on the root group. Since
3280 * only the @tg group gets extra weight, all parent groups can
3281 * only redistribute existing shares. @wl is the shift in shares
3282 * resulting from this level per the above.
3291 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3292 unsigned long wl, unsigned long wg)
3299 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3301 s64 this_load, load;
3302 int idx, this_cpu, prev_cpu;
3303 unsigned long tl_per_task;
3304 struct task_group *tg;
3305 unsigned long weight;
3309 this_cpu = smp_processor_id();
3310 prev_cpu = task_cpu(p);
3311 load = source_load(prev_cpu, idx);
3312 this_load = target_load(this_cpu, idx);
3315 * If sync wakeup then subtract the (maximum possible)
3316 * effect of the currently running task from the load
3317 * of the current CPU:
3320 tg = task_group(current);
3321 weight = current->se.load.weight;
3323 this_load += effective_load(tg, this_cpu, -weight, -weight);
3324 load += effective_load(tg, prev_cpu, 0, -weight);
3328 weight = p->se.load.weight;
3331 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3332 * due to the sync cause above having dropped this_load to 0, we'll
3333 * always have an imbalance, but there's really nothing you can do
3334 * about that, so that's good too.
3336 * Otherwise check if either cpus are near enough in load to allow this
3337 * task to be woken on this_cpu.
3339 if (this_load > 0) {
3340 s64 this_eff_load, prev_eff_load;
3342 this_eff_load = 100;
3343 this_eff_load *= power_of(prev_cpu);
3344 this_eff_load *= this_load +
3345 effective_load(tg, this_cpu, weight, weight);
3347 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3348 prev_eff_load *= power_of(this_cpu);
3349 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3351 balanced = this_eff_load <= prev_eff_load;
3356 * If the currently running task will sleep within
3357 * a reasonable amount of time then attract this newly
3360 if (sync && balanced)
3363 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3364 tl_per_task = cpu_avg_load_per_task(this_cpu);
3367 (this_load <= load &&
3368 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3370 * This domain has SD_WAKE_AFFINE and
3371 * p is cache cold in this domain, and
3372 * there is no bad imbalance.
3374 schedstat_inc(sd, ttwu_move_affine);
3375 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3383 * find_idlest_group finds and returns the least busy CPU group within the
3386 static struct sched_group *
3387 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3388 int this_cpu, int load_idx)
3390 struct sched_group *idlest = NULL, *group = sd->groups;
3391 unsigned long min_load = ULONG_MAX, this_load = 0;
3392 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3395 unsigned long load, avg_load;
3399 /* Skip over this group if it has no CPUs allowed */
3400 if (!cpumask_intersects(sched_group_cpus(group),
3401 tsk_cpus_allowed(p)))
3404 local_group = cpumask_test_cpu(this_cpu,
3405 sched_group_cpus(group));
3407 /* Tally up the load of all CPUs in the group */
3410 for_each_cpu(i, sched_group_cpus(group)) {
3411 /* Bias balancing toward cpus of our domain */
3413 load = source_load(i, load_idx);
3415 load = target_load(i, load_idx);
3420 /* Adjust by relative CPU power of the group */
3421 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3424 this_load = avg_load;
3425 } else if (avg_load < min_load) {
3426 min_load = avg_load;
3429 } while (group = group->next, group != sd->groups);
3431 if (!idlest || 100*this_load < imbalance*min_load)
3437 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3440 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3442 unsigned long load, min_load = ULONG_MAX;
3446 /* Traverse only the allowed CPUs */
3447 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3448 load = weighted_cpuload(i);
3450 if (load < min_load || (load == min_load && i == this_cpu)) {
3460 * Try and locate an idle CPU in the sched_domain.
3462 static int select_idle_sibling(struct task_struct *p, int target)
3464 struct sched_domain *sd;
3465 struct sched_group *sg;
3466 int i = task_cpu(p);
3468 if (idle_cpu(target))
3472 * If the prevous cpu is cache affine and idle, don't be stupid.
3474 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3478 * Otherwise, iterate the domains and find an elegible idle cpu.
3480 sd = rcu_dereference(per_cpu(sd_llc, target));
3481 for_each_lower_domain(sd) {
3484 if (!cpumask_intersects(sched_group_cpus(sg),
3485 tsk_cpus_allowed(p)))
3488 for_each_cpu(i, sched_group_cpus(sg)) {
3489 if (i == target || !idle_cpu(i))
3493 target = cpumask_first_and(sched_group_cpus(sg),
3494 tsk_cpus_allowed(p));
3498 } while (sg != sd->groups);
3504 #ifdef CONFIG_SCHED_HMP
3506 * Heterogenous multiprocessor (HMP) optimizations
3508 * The cpu types are distinguished using a list of hmp_domains
3509 * which each represent one cpu type using a cpumask.
3510 * The list is assumed ordered by compute capacity with the
3511 * fastest domain first.
3513 DEFINE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
3514 static const int hmp_max_tasks = 5;
3516 extern void __init arch_get_hmp_domains(struct list_head *hmp_domains_list);
3518 /* Setup hmp_domains */
3519 static int __init hmp_cpu_mask_setup(void)
3522 struct hmp_domain *domain;
3523 struct list_head *pos;
3526 pr_debug("Initializing HMP scheduler:\n");
3528 /* Initialize hmp_domains using platform code */
3529 arch_get_hmp_domains(&hmp_domains);
3530 if (list_empty(&hmp_domains)) {
3531 pr_debug("HMP domain list is empty!\n");
3535 /* Print hmp_domains */
3537 list_for_each(pos, &hmp_domains) {
3538 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3539 cpulist_scnprintf(buf, 64, &domain->possible_cpus);
3540 pr_debug(" HMP domain %d: %s\n", dc, buf);
3542 for_each_cpu_mask(cpu, domain->possible_cpus) {
3543 per_cpu(hmp_cpu_domain, cpu) = domain;
3551 static struct hmp_domain *hmp_get_hmp_domain_for_cpu(int cpu)
3553 struct hmp_domain *domain;
3554 struct list_head *pos;
3556 list_for_each(pos, &hmp_domains) {
3557 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3558 if(cpumask_test_cpu(cpu, &domain->possible_cpus))
3564 static void hmp_online_cpu(int cpu)
3566 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3569 cpumask_set_cpu(cpu, &domain->cpus);
3572 static void hmp_offline_cpu(int cpu)
3574 struct hmp_domain *domain = hmp_get_hmp_domain_for_cpu(cpu);
3577 cpumask_clear_cpu(cpu, &domain->cpus);
3580 * Needed to determine heaviest tasks etc.
3582 static inline unsigned int hmp_cpu_is_fastest(int cpu);
3583 static inline unsigned int hmp_cpu_is_slowest(int cpu);
3584 static inline struct hmp_domain *hmp_slower_domain(int cpu);
3585 static inline struct hmp_domain *hmp_faster_domain(int cpu);
3587 /* must hold runqueue lock for queue se is currently on */
3588 static struct sched_entity *hmp_get_heaviest_task(
3589 struct sched_entity *se, int migrate_up)
3591 int num_tasks = hmp_max_tasks;
3592 struct sched_entity *max_se = se;
3593 unsigned long int max_ratio = se->avg.load_avg_ratio;
3594 const struct cpumask *hmp_target_mask = NULL;
3597 struct hmp_domain *hmp;
3598 if (hmp_cpu_is_fastest(cpu_of(se->cfs_rq->rq)))
3601 hmp = hmp_faster_domain(cpu_of(se->cfs_rq->rq));
3602 hmp_target_mask = &hmp->cpus;
3604 /* The currently running task is not on the runqueue */
3605 se = __pick_first_entity(cfs_rq_of(se));
3607 while (num_tasks && se) {
3608 if (entity_is_task(se) &&
3609 (se->avg.load_avg_ratio > max_ratio &&
3611 cpumask_intersects(hmp_target_mask,
3612 tsk_cpus_allowed(task_of(se))))) {
3614 max_ratio = se->avg.load_avg_ratio;
3616 se = __pick_next_entity(se);
3622 static struct sched_entity *hmp_get_lightest_task(
3623 struct sched_entity *se, int migrate_down)
3625 int num_tasks = hmp_max_tasks;
3626 struct sched_entity *min_se = se;
3627 unsigned long int min_ratio = se->avg.load_avg_ratio;
3628 const struct cpumask *hmp_target_mask = NULL;
3631 struct hmp_domain *hmp;
3632 if (hmp_cpu_is_slowest(cpu_of(se->cfs_rq->rq)))
3634 hmp = hmp_slower_domain(cpu_of(se->cfs_rq->rq));
3635 hmp_target_mask = &hmp->cpus;
3637 /* The currently running task is not on the runqueue */
3638 se = __pick_first_entity(cfs_rq_of(se));
3640 while (num_tasks && se) {
3641 if (entity_is_task(se) &&
3642 (se->avg.load_avg_ratio < min_ratio &&
3644 cpumask_intersects(hmp_target_mask,
3645 tsk_cpus_allowed(task_of(se))))) {
3647 min_ratio = se->avg.load_avg_ratio;
3649 se = __pick_next_entity(se);
3656 * Migration thresholds should be in the range [0..1023]
3657 * hmp_up_threshold: min. load required for migrating tasks to a faster cpu
3658 * hmp_down_threshold: max. load allowed for tasks migrating to a slower cpu
3660 * hmp_up_prio: Only up migrate task with high priority (<hmp_up_prio)
3661 * hmp_next_up_threshold: Delay before next up migration (1024 ~= 1 ms)
3662 * hmp_next_down_threshold: Delay before next down migration (1024 ~= 1 ms)
3664 * Small Task Packing:
3665 * We can choose to fill the littlest CPUs in an HMP system rather than
3666 * the typical spreading mechanic. This behavior is controllable using
3668 * hmp_packing_enabled: runtime control over pack/spread
3669 * hmp_full_threshold: Consider a CPU with this much unweighted load full
3671 unsigned int hmp_up_threshold = 700;
3672 unsigned int hmp_down_threshold = 512;
3673 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
3674 unsigned int hmp_up_prio = NICE_TO_PRIO(CONFIG_SCHED_HMP_PRIO_FILTER_VAL);
3676 unsigned int hmp_next_up_threshold = 4096;
3677 unsigned int hmp_next_down_threshold = 4096;
3679 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3680 unsigned int hmp_packing_enabled = 1;
3681 #ifndef CONFIG_ARCH_VEXPRESS_TC2
3682 unsigned int hmp_full_threshold = (NICE_0_LOAD * 9) / 8;
3684 /* TC2 has a sharp consumption curve @ around 800Mhz, so
3685 we aim to spread the load around that frequency. */
3686 unsigned int hmp_full_threshold = 650; /* 80% of the 800Mhz freq * NICE_0_LOAD */
3690 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se);
3691 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se);
3692 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
3693 int *min_cpu, struct cpumask *affinity);
3695 static inline struct hmp_domain *hmp_smallest_domain(void)
3697 return list_entry(hmp_domains.prev, struct hmp_domain, hmp_domains);
3700 /* Check if cpu is in fastest hmp_domain */
3701 static inline unsigned int hmp_cpu_is_fastest(int cpu)
3703 struct list_head *pos;
3705 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3706 return pos == hmp_domains.next;
3709 /* Check if cpu is in slowest hmp_domain */
3710 static inline unsigned int hmp_cpu_is_slowest(int cpu)
3712 struct list_head *pos;
3714 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3715 return list_is_last(pos, &hmp_domains);
3718 /* Next (slower) hmp_domain relative to cpu */
3719 static inline struct hmp_domain *hmp_slower_domain(int cpu)
3721 struct list_head *pos;
3723 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3724 return list_entry(pos->next, struct hmp_domain, hmp_domains);
3727 /* Previous (faster) hmp_domain relative to cpu */
3728 static inline struct hmp_domain *hmp_faster_domain(int cpu)
3730 struct list_head *pos;
3732 pos = &hmp_cpu_domain(cpu)->hmp_domains;
3733 return list_entry(pos->prev, struct hmp_domain, hmp_domains);
3737 * Selects a cpu in previous (faster) hmp_domain
3739 static inline unsigned int hmp_select_faster_cpu(struct task_struct *tsk,
3742 int lowest_cpu=NR_CPUS;
3743 __always_unused int lowest_ratio;
3744 struct hmp_domain *hmp;
3746 if (hmp_cpu_is_fastest(cpu))
3747 hmp = hmp_cpu_domain(cpu);
3749 hmp = hmp_faster_domain(cpu);
3751 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3752 tsk_cpus_allowed(tsk));
3758 * Selects a cpu in next (slower) hmp_domain
3759 * Note that cpumask_any_and() returns the first cpu in the cpumask
3761 static inline unsigned int hmp_select_slower_cpu(struct task_struct *tsk,
3764 int lowest_cpu=NR_CPUS;
3765 struct hmp_domain *hmp;
3766 __always_unused int lowest_ratio;
3768 if (hmp_cpu_is_slowest(cpu))
3769 hmp = hmp_cpu_domain(cpu);
3771 hmp = hmp_slower_domain(cpu);
3773 lowest_ratio = hmp_domain_min_load(hmp, &lowest_cpu,
3774 tsk_cpus_allowed(tsk));
3778 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3780 * Select the 'best' candidate little CPU to wake up on.
3781 * Implements a packing strategy which examines CPU in
3782 * logical CPU order, and selects the first which will
3783 * have at least 10% capacity available, according to
3784 * both tracked load of the runqueue and the task.
3786 static inline unsigned int hmp_best_little_cpu(struct task_struct *tsk,
3789 unsigned long estimated_load;
3790 struct hmp_domain *hmp;
3791 struct sched_avg *avg;
3792 struct cpumask allowed_hmp_cpus;
3794 if(!hmp_packing_enabled ||
3795 tsk->se.avg.load_avg_ratio > ((NICE_0_LOAD * 90)/100))
3796 return hmp_select_slower_cpu(tsk, cpu);
3798 if (hmp_cpu_is_slowest(cpu))
3799 hmp = hmp_cpu_domain(cpu);
3801 hmp = hmp_slower_domain(cpu);
3803 /* respect affinity */
3804 cpumask_and(&allowed_hmp_cpus, &hmp->cpus,
3805 tsk_cpus_allowed(tsk));
3807 for_each_cpu_mask(tmp_cpu, allowed_hmp_cpus) {
3808 avg = &cpu_rq(tmp_cpu)->avg;
3809 /* estimate new rq load if we add this task */
3810 estimated_load = avg->load_avg_ratio +
3811 tsk->se.avg.load_avg_ratio;
3812 if (estimated_load <= hmp_full_threshold) {
3817 /* if no match was found, the task uses the initial value */
3821 static inline void hmp_next_up_delay(struct sched_entity *se, int cpu)
3823 /* hack - always use clock from first online CPU */
3824 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3825 se->avg.hmp_last_up_migration = now;
3826 se->avg.hmp_last_down_migration = 0;
3827 cpu_rq(cpu)->avg.hmp_last_up_migration = now;
3828 cpu_rq(cpu)->avg.hmp_last_down_migration = 0;
3831 static inline void hmp_next_down_delay(struct sched_entity *se, int cpu)
3833 /* hack - always use clock from first online CPU */
3834 u64 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
3835 se->avg.hmp_last_down_migration = now;
3836 se->avg.hmp_last_up_migration = 0;
3837 cpu_rq(cpu)->avg.hmp_last_down_migration = now;
3838 cpu_rq(cpu)->avg.hmp_last_up_migration = 0;
3842 * Heterogenous multiprocessor (HMP) optimizations
3844 * These functions allow to change the growing speed of the load_avg_ratio
3845 * by default it goes from 0 to 0.5 in LOAD_AVG_PERIOD = 32ms
3846 * This can now be changed with /sys/kernel/hmp/load_avg_period_ms.
3848 * These functions also allow to change the up and down threshold of HMP
3849 * using /sys/kernel/hmp/{up,down}_threshold.
3850 * Both must be between 0 and 1023. The threshold that is compared
3851 * to the load_avg_ratio is up_threshold/1024 and down_threshold/1024.
3853 * For instance, if load_avg_period = 64 and up_threshold = 512, an idle
3854 * task with a load of 0 will reach the threshold after 64ms of busy loop.
3856 * Changing load_avg_periods_ms has the same effect than changing the
3857 * default scaling factor Y=1002/1024 in the load_avg_ratio computation to
3858 * (1002/1024.0)^(LOAD_AVG_PERIOD/load_avg_period_ms), but the last one
3859 * could trigger overflows.
3860 * For instance, with Y = 1023/1024 in __update_task_entity_contrib()
3861 * "contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);"
3862 * could be overflowed for a weight > 2^12 even is the load_avg_contrib
3863 * should still be a 32bits result. This would not happen by multiplicating
3864 * delta time by 1/22 and setting load_avg_period_ms = 706.
3868 * By scaling the delta time it end-up increasing or decrease the
3869 * growing speed of the per entity load_avg_ratio
3870 * The scale factor hmp_data.multiplier is a fixed point
3871 * number: (32-HMP_VARIABLE_SCALE_SHIFT).HMP_VARIABLE_SCALE_SHIFT
3873 static inline u64 hmp_variable_scale_convert(u64 delta)
3875 #ifdef CONFIG_HMP_VARIABLE_SCALE
3876 u64 high = delta >> 32ULL;
3877 u64 low = delta & 0xffffffffULL;
3878 low *= hmp_data.multiplier;
3879 high *= hmp_data.multiplier;
3880 return (low >> HMP_VARIABLE_SCALE_SHIFT)
3881 + (high << (32ULL - HMP_VARIABLE_SCALE_SHIFT));
3887 static ssize_t hmp_show(struct kobject *kobj,
3888 struct attribute *attr, char *buf)
3890 struct hmp_global_attr *hmp_attr =
3891 container_of(attr, struct hmp_global_attr, attr);
3894 if (hmp_attr->to_sysfs_text != NULL)
3895 return hmp_attr->to_sysfs_text(buf, PAGE_SIZE);
3897 temp = *(hmp_attr->value);
3898 if (hmp_attr->to_sysfs != NULL)
3899 temp = hmp_attr->to_sysfs(temp);
3901 return (ssize_t)sprintf(buf, "%d\n", temp);
3904 static ssize_t hmp_store(struct kobject *a, struct attribute *attr,
3905 const char *buf, size_t count)
3908 ssize_t ret = count;
3909 struct hmp_global_attr *hmp_attr =
3910 container_of(attr, struct hmp_global_attr, attr);
3911 char *str = vmalloc(count + 1);
3914 memcpy(str, buf, count);
3916 if (sscanf(str, "%d", &temp) < 1)
3919 if (hmp_attr->from_sysfs != NULL)
3920 temp = hmp_attr->from_sysfs(temp);
3924 *(hmp_attr->value) = temp;
3930 static ssize_t hmp_print_domains(char *outbuf, int outbufsize)
3933 const char nospace[] = "%s", space[] = " %s";
3934 const char *fmt = nospace;
3935 struct hmp_domain *domain;
3936 struct list_head *pos;
3938 list_for_each(pos, &hmp_domains) {
3939 domain = list_entry(pos, struct hmp_domain, hmp_domains);
3940 if (cpumask_scnprintf(buf, 64, &domain->possible_cpus)) {
3941 outpos += sprintf(outbuf+outpos, fmt, buf);
3945 strcat(outbuf, "\n");
3949 #ifdef CONFIG_HMP_VARIABLE_SCALE
3950 static int hmp_period_tofrom_sysfs(int value)
3952 return (LOAD_AVG_PERIOD << HMP_VARIABLE_SCALE_SHIFT) / value;
3955 /* max value for threshold is 1024 */
3956 static int hmp_theshold_from_sysfs(int value)
3962 #if defined(CONFIG_SCHED_HMP_LITTLE_PACKING) || \
3963 defined(CONFIG_HMP_FREQUENCY_INVARIANT_SCALE)
3964 /* toggle control is only 0,1 off/on */
3965 static int hmp_toggle_from_sysfs(int value)
3967 if (value < 0 || value > 1)
3972 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
3973 /* packing value must be non-negative */
3974 static int hmp_packing_from_sysfs(int value)
3981 static void hmp_attr_add(
3984 int (*to_sysfs)(int),
3985 int (*from_sysfs)(int),
3986 ssize_t (*to_sysfs_text)(char *, int),
3990 while (hmp_data.attributes[i] != NULL) {
3992 if (i >= HMP_DATA_SYSFS_MAX)
3996 hmp_data.attr[i].attr.mode = mode;
3998 hmp_data.attr[i].attr.mode = 0644;
3999 hmp_data.attr[i].show = hmp_show;
4000 hmp_data.attr[i].store = hmp_store;
4001 hmp_data.attr[i].attr.name = name;
4002 hmp_data.attr[i].value = value;
4003 hmp_data.attr[i].to_sysfs = to_sysfs;
4004 hmp_data.attr[i].from_sysfs = from_sysfs;
4005 hmp_data.attr[i].to_sysfs_text = to_sysfs_text;
4006 hmp_data.attributes[i] = &hmp_data.attr[i].attr;
4007 hmp_data.attributes[i + 1] = NULL;
4010 static int hmp_attr_init(void)
4013 memset(&hmp_data, sizeof(hmp_data), 0);
4014 hmp_attr_add("hmp_domains",
4020 hmp_attr_add("up_threshold",
4023 hmp_theshold_from_sysfs,
4026 hmp_attr_add("down_threshold",
4027 &hmp_down_threshold,
4029 hmp_theshold_from_sysfs,
4032 #ifdef CONFIG_HMP_VARIABLE_SCALE
4033 /* by default load_avg_period_ms == LOAD_AVG_PERIOD
4036 hmp_data.multiplier = hmp_period_tofrom_sysfs(LOAD_AVG_PERIOD);
4037 hmp_attr_add("load_avg_period_ms",
4038 &hmp_data.multiplier,
4039 hmp_period_tofrom_sysfs,
4040 hmp_period_tofrom_sysfs,
4044 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
4045 /* default frequency-invariant scaling ON */
4046 hmp_data.freqinvar_load_scale_enabled = 1;
4047 hmp_attr_add("frequency_invariant_load_scale",
4048 &hmp_data.freqinvar_load_scale_enabled,
4050 hmp_toggle_from_sysfs,
4054 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4055 hmp_attr_add("packing_enable",
4056 &hmp_packing_enabled,
4058 hmp_toggle_from_sysfs,
4061 hmp_attr_add("packing_limit",
4062 &hmp_full_threshold,
4064 hmp_packing_from_sysfs,
4068 hmp_data.attr_group.name = "hmp";
4069 hmp_data.attr_group.attrs = hmp_data.attributes;
4070 ret = sysfs_create_group(kernel_kobj,
4071 &hmp_data.attr_group);
4074 late_initcall(hmp_attr_init);
4076 * return the load of the lowest-loaded CPU in a given HMP domain
4077 * min_cpu optionally points to an int to receive the CPU.
4078 * affinity optionally points to a cpumask containing the
4079 * CPUs to be considered. note:
4080 * + min_cpu = NR_CPUS only if no CPUs are in the set of
4081 * affinity && hmp_domain cpus
4082 * + min_cpu will always otherwise equal one of the CPUs in
4084 * + when more than one CPU has the same load, the one which
4085 * is least-recently-disturbed by an HMP migration will be
4087 * + if all CPUs are equally loaded or idle and the times are
4088 * all the same, the first in the set will be used
4089 * + if affinity is not set, cpu_online_mask is used
4091 static inline unsigned int hmp_domain_min_load(struct hmp_domain *hmpd,
4092 int *min_cpu, struct cpumask *affinity)
4095 int min_cpu_runnable_temp = NR_CPUS;
4096 u64 min_target_last_migration = ULLONG_MAX;
4097 u64 curr_last_migration;
4098 unsigned long min_runnable_load = INT_MAX;
4099 unsigned long contrib;
4100 struct sched_avg *avg;
4101 struct cpumask temp_cpumask;
4103 * only look at CPUs allowed if specified,
4104 * otherwise look at all online CPUs in the
4107 cpumask_and(&temp_cpumask, &hmpd->cpus, affinity ? affinity : cpu_online_mask);
4109 for_each_cpu_mask(cpu, temp_cpumask) {
4110 avg = &cpu_rq(cpu)->avg;
4111 /* used for both up and down migration */
4112 curr_last_migration = avg->hmp_last_up_migration ?
4113 avg->hmp_last_up_migration : avg->hmp_last_down_migration;
4115 contrib = avg->load_avg_ratio;
4117 * Consider a runqueue completely busy if there is any load
4118 * on it. Definitely not the best for overall fairness, but
4119 * does well in typical Android use cases.
4124 if ((contrib < min_runnable_load) ||
4125 (contrib == min_runnable_load &&
4126 curr_last_migration < min_target_last_migration)) {
4128 * if the load is the same target the CPU with
4129 * the longest time since a migration.
4130 * This is to spread migration load between
4131 * members of a domain more evenly when the
4132 * domain is fully loaded
4134 min_runnable_load = contrib;
4135 min_cpu_runnable_temp = cpu;
4136 min_target_last_migration = curr_last_migration;
4141 *min_cpu = min_cpu_runnable_temp;
4143 return min_runnable_load;
4147 * Calculate the task starvation
4148 * This is the ratio of actually running time vs. runnable time.
4149 * If the two are equal the task is getting the cpu time it needs or
4150 * it is alone on the cpu and the cpu is fully utilized.
4152 static inline unsigned int hmp_task_starvation(struct sched_entity *se)
4156 starvation = se->avg.usage_avg_sum * scale_load_down(NICE_0_LOAD);
4157 starvation /= (se->avg.runnable_avg_sum + 1);
4159 return scale_load(starvation);
4162 static inline unsigned int hmp_offload_down(int cpu, struct sched_entity *se)
4165 int dest_cpu = NR_CPUS;
4167 if (hmp_cpu_is_slowest(cpu))
4170 /* Is there an idle CPU in the current domain */
4171 min_usage = hmp_domain_min_load(hmp_cpu_domain(cpu), NULL, NULL);
4172 if (min_usage == 0) {
4173 trace_sched_hmp_offload_abort(cpu, min_usage, "load");
4177 /* Is the task alone on the cpu? */
4178 if (cpu_rq(cpu)->cfs.h_nr_running < 2) {
4179 trace_sched_hmp_offload_abort(cpu,
4180 cpu_rq(cpu)->cfs.h_nr_running, "nr_running");
4184 /* Is the task actually starving? */
4185 /* >=25% ratio running/runnable = starving */
4186 if (hmp_task_starvation(se) > 768) {
4187 trace_sched_hmp_offload_abort(cpu, hmp_task_starvation(se),
4192 /* Does the slower domain have any idle CPUs? */
4193 min_usage = hmp_domain_min_load(hmp_slower_domain(cpu), &dest_cpu,
4194 tsk_cpus_allowed(task_of(se)));
4196 if (min_usage == 0) {
4197 trace_sched_hmp_offload_succeed(cpu, dest_cpu);
4200 trace_sched_hmp_offload_abort(cpu,min_usage,"slowdomain");
4203 #endif /* CONFIG_SCHED_HMP */
4206 * sched_balance_self: balance the current task (running on cpu) in domains
4207 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4210 * Balance, ie. select the least loaded group.
4212 * Returns the target CPU number, or the same CPU if no balancing is needed.
4214 * preempt must be disabled.
4217 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
4219 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4220 int cpu = smp_processor_id();
4221 int prev_cpu = task_cpu(p);
4223 int want_affine = 0;
4224 int sync = wake_flags & WF_SYNC;
4226 if (p->nr_cpus_allowed == 1)
4229 #ifdef CONFIG_SCHED_HMP
4230 /* always put non-kernel forking tasks on a big domain */
4231 if (p->mm && (sd_flag & SD_BALANCE_FORK)) {
4232 new_cpu = hmp_select_faster_cpu(p, prev_cpu);
4233 if (new_cpu != NR_CPUS) {
4234 hmp_next_up_delay(&p->se, new_cpu);
4237 /* failed to perform HMP fork balance, use normal balance */
4242 if (sd_flag & SD_BALANCE_WAKE) {
4243 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4249 for_each_domain(cpu, tmp) {
4250 if (!(tmp->flags & SD_LOAD_BALANCE))
4254 * If both cpu and prev_cpu are part of this domain,
4255 * cpu is a valid SD_WAKE_AFFINE target.
4257 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4258 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4263 if (tmp->flags & sd_flag)
4268 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4271 new_cpu = select_idle_sibling(p, prev_cpu);
4276 int load_idx = sd->forkexec_idx;
4277 struct sched_group *group;
4280 if (!(sd->flags & sd_flag)) {
4285 if (sd_flag & SD_BALANCE_WAKE)
4286 load_idx = sd->wake_idx;
4288 group = find_idlest_group(sd, p, cpu, load_idx);
4294 new_cpu = find_idlest_cpu(group, p, cpu);
4295 if (new_cpu == -1 || new_cpu == cpu) {
4296 /* Now try balancing at a lower domain level of cpu */
4301 /* Now try balancing at a lower domain level of new_cpu */
4303 weight = sd->span_weight;
4305 for_each_domain(cpu, tmp) {
4306 if (weight <= tmp->span_weight)
4308 if (tmp->flags & sd_flag)
4311 /* while loop will break here if sd == NULL */
4316 #ifdef CONFIG_SCHED_HMP
4317 prev_cpu = task_cpu(p);
4319 if (hmp_up_migration(prev_cpu, &new_cpu, &p->se)) {
4320 hmp_next_up_delay(&p->se, new_cpu);
4321 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4324 if (hmp_down_migration(prev_cpu, &p->se)) {
4325 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
4326 new_cpu = hmp_best_little_cpu(p, prev_cpu);
4328 new_cpu = hmp_select_slower_cpu(p, prev_cpu);
4330 if (new_cpu != prev_cpu) {
4331 hmp_next_down_delay(&p->se, new_cpu);
4332 trace_sched_hmp_migrate(p, new_cpu, HMP_MIGRATE_WAKEUP);
4336 /* Make sure that the task stays in its previous hmp domain */
4337 if (!cpumask_test_cpu(new_cpu, &hmp_cpu_domain(prev_cpu)->cpus))
4345 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
4346 * removed when useful for applications beyond shares distribution (e.g.
4349 #ifdef CONFIG_FAIR_GROUP_SCHED
4351 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4352 * cfs_rq_of(p) references at time of call are still valid and identify the
4353 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4354 * other assumptions, including the state of rq->lock, should be made.
4357 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4359 struct sched_entity *se = &p->se;
4360 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4363 * Load tracking: accumulate removed load so that it can be processed
4364 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4365 * to blocked load iff they have a positive decay-count. It can never
4366 * be negative here since on-rq tasks have decay-count == 0.
4368 if (se->avg.decay_count) {
4369 se->avg.decay_count = -__synchronize_entity_decay(se);
4370 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
4374 #endif /* CONFIG_SMP */
4376 static unsigned long
4377 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4379 unsigned long gran = sysctl_sched_wakeup_granularity;
4382 * Since its curr running now, convert the gran from real-time
4383 * to virtual-time in his units.
4385 * By using 'se' instead of 'curr' we penalize light tasks, so
4386 * they get preempted easier. That is, if 'se' < 'curr' then
4387 * the resulting gran will be larger, therefore penalizing the
4388 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4389 * be smaller, again penalizing the lighter task.
4391 * This is especially important for buddies when the leftmost
4392 * task is higher priority than the buddy.
4394 return calc_delta_fair(gran, se);
4398 * Should 'se' preempt 'curr'.
4412 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4414 s64 gran, vdiff = curr->vruntime - se->vruntime;
4419 gran = wakeup_gran(curr, se);
4426 static void set_last_buddy(struct sched_entity *se)
4428 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4431 for_each_sched_entity(se)
4432 cfs_rq_of(se)->last = se;
4435 static void set_next_buddy(struct sched_entity *se)
4437 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4440 for_each_sched_entity(se)
4441 cfs_rq_of(se)->next = se;
4444 static void set_skip_buddy(struct sched_entity *se)
4446 for_each_sched_entity(se)
4447 cfs_rq_of(se)->skip = se;
4451 * Preempt the current task with a newly woken task if needed:
4453 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4455 struct task_struct *curr = rq->curr;
4456 struct sched_entity *se = &curr->se, *pse = &p->se;
4457 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4458 int scale = cfs_rq->nr_running >= sched_nr_latency;
4459 int next_buddy_marked = 0;
4461 if (unlikely(se == pse))
4465 * This is possible from callers such as move_task(), in which we
4466 * unconditionally check_prempt_curr() after an enqueue (which may have
4467 * lead to a throttle). This both saves work and prevents false
4468 * next-buddy nomination below.
4470 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4473 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4474 set_next_buddy(pse);
4475 next_buddy_marked = 1;
4479 * We can come here with TIF_NEED_RESCHED already set from new task
4482 * Note: this also catches the edge-case of curr being in a throttled
4483 * group (e.g. via set_curr_task), since update_curr() (in the
4484 * enqueue of curr) will have resulted in resched being set. This
4485 * prevents us from potentially nominating it as a false LAST_BUDDY
4488 if (test_tsk_need_resched(curr))
4491 /* Idle tasks are by definition preempted by non-idle tasks. */
4492 if (unlikely(curr->policy == SCHED_IDLE) &&
4493 likely(p->policy != SCHED_IDLE))
4497 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4498 * is driven by the tick):
4500 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4503 find_matching_se(&se, &pse);
4504 update_curr(cfs_rq_of(se));
4506 if (wakeup_preempt_entity(se, pse) == 1) {
4508 * Bias pick_next to pick the sched entity that is
4509 * triggering this preemption.
4511 if (!next_buddy_marked)
4512 set_next_buddy(pse);
4521 * Only set the backward buddy when the current task is still
4522 * on the rq. This can happen when a wakeup gets interleaved
4523 * with schedule on the ->pre_schedule() or idle_balance()
4524 * point, either of which can * drop the rq lock.
4526 * Also, during early boot the idle thread is in the fair class,
4527 * for obvious reasons its a bad idea to schedule back to it.
4529 if (unlikely(!se->on_rq || curr == rq->idle))
4532 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4536 static struct task_struct *pick_next_task_fair(struct rq *rq)
4538 struct task_struct *p;
4539 struct cfs_rq *cfs_rq = &rq->cfs;
4540 struct sched_entity *se;
4542 if (!cfs_rq->nr_running)
4546 se = pick_next_entity(cfs_rq);
4547 set_next_entity(cfs_rq, se);
4548 cfs_rq = group_cfs_rq(se);
4552 if (hrtick_enabled(rq))
4553 hrtick_start_fair(rq, p);
4559 * Account for a descheduled task:
4561 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4563 struct sched_entity *se = &prev->se;
4564 struct cfs_rq *cfs_rq;
4566 for_each_sched_entity(se) {
4567 cfs_rq = cfs_rq_of(se);
4568 put_prev_entity(cfs_rq, se);
4573 * sched_yield() is very simple
4575 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4577 static void yield_task_fair(struct rq *rq)
4579 struct task_struct *curr = rq->curr;
4580 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4581 struct sched_entity *se = &curr->se;
4584 * Are we the only task in the tree?
4586 if (unlikely(rq->nr_running == 1))
4589 clear_buddies(cfs_rq, se);
4591 if (curr->policy != SCHED_BATCH) {
4592 update_rq_clock(rq);
4594 * Update run-time statistics of the 'current'.
4596 update_curr(cfs_rq);
4598 * Tell update_rq_clock() that we've just updated,
4599 * so we don't do microscopic update in schedule()
4600 * and double the fastpath cost.
4602 rq->skip_clock_update = 1;
4608 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4610 struct sched_entity *se = &p->se;
4612 /* throttled hierarchies are not runnable */
4613 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4616 /* Tell the scheduler that we'd really like pse to run next. */
4619 yield_task_fair(rq);
4625 /**************************************************
4626 * Fair scheduling class load-balancing methods.
4630 * The purpose of load-balancing is to achieve the same basic fairness the
4631 * per-cpu scheduler provides, namely provide a proportional amount of compute
4632 * time to each task. This is expressed in the following equation:
4634 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4636 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4637 * W_i,0 is defined as:
4639 * W_i,0 = \Sum_j w_i,j (2)
4641 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4642 * is derived from the nice value as per prio_to_weight[].
4644 * The weight average is an exponential decay average of the instantaneous
4647 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4649 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4650 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4651 * can also include other factors [XXX].
4653 * To achieve this balance we define a measure of imbalance which follows
4654 * directly from (1):
4656 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4658 * We them move tasks around to minimize the imbalance. In the continuous
4659 * function space it is obvious this converges, in the discrete case we get
4660 * a few fun cases generally called infeasible weight scenarios.
4663 * - infeasible weights;
4664 * - local vs global optima in the discrete case. ]
4669 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4670 * for all i,j solution, we create a tree of cpus that follows the hardware
4671 * topology where each level pairs two lower groups (or better). This results
4672 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4673 * tree to only the first of the previous level and we decrease the frequency
4674 * of load-balance at each level inv. proportional to the number of cpus in
4680 * \Sum { --- * --- * 2^i } = O(n) (5)
4682 * `- size of each group
4683 * | | `- number of cpus doing load-balance
4685 * `- sum over all levels
4687 * Coupled with a limit on how many tasks we can migrate every balance pass,
4688 * this makes (5) the runtime complexity of the balancer.
4690 * An important property here is that each CPU is still (indirectly) connected
4691 * to every other cpu in at most O(log n) steps:
4693 * The adjacency matrix of the resulting graph is given by:
4696 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4699 * And you'll find that:
4701 * A^(log_2 n)_i,j != 0 for all i,j (7)
4703 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4704 * The task movement gives a factor of O(m), giving a convergence complexity
4707 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4712 * In order to avoid CPUs going idle while there's still work to do, new idle
4713 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4714 * tree itself instead of relying on other CPUs to bring it work.
4716 * This adds some complexity to both (5) and (8) but it reduces the total idle
4724 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4727 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4732 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4734 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4736 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4739 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4740 * rewrite all of this once again.]
4743 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4745 #define LBF_ALL_PINNED 0x01
4746 #define LBF_NEED_BREAK 0x02
4747 #define LBF_SOME_PINNED 0x04
4750 struct sched_domain *sd;
4758 struct cpumask *dst_grpmask;
4760 enum cpu_idle_type idle;
4762 /* The set of CPUs under consideration for load-balancing */
4763 struct cpumask *cpus;
4768 unsigned int loop_break;
4769 unsigned int loop_max;
4773 * move_task - move a task from one runqueue to another runqueue.
4774 * Both runqueues must be locked.
4776 static void move_task(struct task_struct *p, struct lb_env *env)
4778 deactivate_task(env->src_rq, p, 0);
4779 set_task_cpu(p, env->dst_cpu);
4780 activate_task(env->dst_rq, p, 0);
4781 check_preempt_curr(env->dst_rq, p, 0);
4785 * Is this task likely cache-hot:
4788 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4792 if (p->sched_class != &fair_sched_class)
4795 if (unlikely(p->policy == SCHED_IDLE))
4799 * Buddy candidates are cache hot:
4801 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4802 (&p->se == cfs_rq_of(&p->se)->next ||
4803 &p->se == cfs_rq_of(&p->se)->last))
4806 if (sysctl_sched_migration_cost == -1)
4808 if (sysctl_sched_migration_cost == 0)
4811 delta = now - p->se.exec_start;
4813 return delta < (s64)sysctl_sched_migration_cost;
4817 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4820 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4822 int tsk_cache_hot = 0;
4824 * We do not migrate tasks that are:
4825 * 1) throttled_lb_pair, or
4826 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4827 * 3) running (obviously), or
4828 * 4) are cache-hot on their current CPU.
4830 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4833 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4836 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4839 * Remember if this task can be migrated to any other cpu in
4840 * our sched_group. We may want to revisit it if we couldn't
4841 * meet load balance goals by pulling other tasks on src_cpu.
4843 * Also avoid computing new_dst_cpu if we have already computed
4844 * one in current iteration.
4846 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
4849 /* Prevent to re-select dst_cpu via env's cpus */
4850 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4851 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4852 env->flags |= LBF_SOME_PINNED;
4853 env->new_dst_cpu = cpu;
4861 /* Record that we found atleast one task that could run on dst_cpu */
4862 env->flags &= ~LBF_ALL_PINNED;
4864 if (task_running(env->src_rq, p)) {
4865 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4870 * Aggressive migration if:
4871 * 1) task is cache cold, or
4872 * 2) too many balance attempts have failed.
4874 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
4875 if (!tsk_cache_hot ||
4876 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4878 if (tsk_cache_hot) {
4879 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4880 schedstat_inc(p, se.statistics.nr_forced_migrations);
4886 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4891 * move_one_task tries to move exactly one task from busiest to this_rq, as
4892 * part of active balancing operations within "domain".
4893 * Returns 1 if successful and 0 otherwise.
4895 * Called with both runqueues locked.
4897 static int move_one_task(struct lb_env *env)
4899 struct task_struct *p, *n;
4901 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4902 if (!can_migrate_task(p, env))
4907 * Right now, this is only the second place move_task()
4908 * is called, so we can safely collect move_task()
4909 * stats here rather than inside move_task().
4911 schedstat_inc(env->sd, lb_gained[env->idle]);
4917 static unsigned long task_h_load(struct task_struct *p);
4919 static const unsigned int sched_nr_migrate_break = 32;
4922 * move_tasks tries to move up to imbalance weighted load from busiest to
4923 * this_rq, as part of a balancing operation within domain "sd".
4924 * Returns 1 if successful and 0 otherwise.
4926 * Called with both runqueues locked.
4928 static int move_tasks(struct lb_env *env)
4930 struct list_head *tasks = &env->src_rq->cfs_tasks;
4931 struct task_struct *p;
4935 if (env->imbalance <= 0)
4938 while (!list_empty(tasks)) {
4939 p = list_first_entry(tasks, struct task_struct, se.group_node);
4942 /* We've more or less seen every task there is, call it quits */
4943 if (env->loop > env->loop_max)
4946 /* take a breather every nr_migrate tasks */
4947 if (env->loop > env->loop_break) {
4948 env->loop_break += sched_nr_migrate_break;
4949 env->flags |= LBF_NEED_BREAK;
4953 if (!can_migrate_task(p, env))
4956 load = task_h_load(p);
4958 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4961 if ((load / 2) > env->imbalance)
4966 env->imbalance -= load;
4968 #ifdef CONFIG_PREEMPT
4970 * NEWIDLE balancing is a source of latency, so preemptible
4971 * kernels will stop after the first task is pulled to minimize
4972 * the critical section.
4974 if (env->idle == CPU_NEWLY_IDLE)
4979 * We only want to steal up to the prescribed amount of
4982 if (env->imbalance <= 0)
4987 list_move_tail(&p->se.group_node, tasks);
4991 * Right now, this is one of only two places move_task() is called,
4992 * so we can safely collect move_task() stats here rather than
4993 * inside move_task().
4995 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5000 #ifdef CONFIG_FAIR_GROUP_SCHED
5002 * update tg->load_weight by folding this cpu's load_avg
5004 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5006 struct sched_entity *se = tg->se[cpu];
5007 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5009 /* throttled entities do not contribute to load */
5010 if (throttled_hierarchy(cfs_rq))
5013 update_cfs_rq_blocked_load(cfs_rq, 1);
5016 update_entity_load_avg(se, 1);
5018 * We pivot on our runnable average having decayed to zero for
5019 * list removal. This generally implies that all our children
5020 * have also been removed (modulo rounding error or bandwidth
5021 * control); however, such cases are rare and we can fix these
5024 * TODO: fix up out-of-order children on enqueue.
5026 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5027 list_del_leaf_cfs_rq(cfs_rq);
5029 struct rq *rq = rq_of(cfs_rq);
5030 update_rq_runnable_avg(rq, rq->nr_running);
5034 static void update_blocked_averages(int cpu)
5036 struct rq *rq = cpu_rq(cpu);
5037 struct cfs_rq *cfs_rq;
5038 unsigned long flags;
5040 raw_spin_lock_irqsave(&rq->lock, flags);
5041 update_rq_clock(rq);
5043 * Iterates the task_group tree in a bottom up fashion, see
5044 * list_add_leaf_cfs_rq() for details.
5046 for_each_leaf_cfs_rq(rq, cfs_rq) {
5048 * Note: We may want to consider periodically releasing
5049 * rq->lock about these updates so that creating many task
5050 * groups does not result in continually extending hold time.
5052 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5055 raw_spin_unlock_irqrestore(&rq->lock, flags);
5059 * Compute the cpu's hierarchical load factor for each task group.
5060 * This needs to be done in a top-down fashion because the load of a child
5061 * group is a fraction of its parents load.
5063 static int tg_load_down(struct task_group *tg, void *data)
5066 long cpu = (long)data;
5069 load = cpu_rq(cpu)->load.weight;
5071 load = tg->parent->cfs_rq[cpu]->h_load;
5072 load *= tg->se[cpu]->load.weight;
5073 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
5076 tg->cfs_rq[cpu]->h_load = load;
5081 static void update_h_load(long cpu)
5083 struct rq *rq = cpu_rq(cpu);
5084 unsigned long now = jiffies;
5086 if (rq->h_load_throttle == now)
5089 rq->h_load_throttle = now;
5092 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
5096 static unsigned long task_h_load(struct task_struct *p)
5098 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5101 load = p->se.load.weight;
5102 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
5107 static inline void update_blocked_averages(int cpu)
5111 static inline void update_h_load(long cpu)
5115 static unsigned long task_h_load(struct task_struct *p)
5117 return p->se.load.weight;
5121 /********** Helpers for find_busiest_group ************************/
5123 * sd_lb_stats - Structure to store the statistics of a sched_domain
5124 * during load balancing.
5126 struct sd_lb_stats {
5127 struct sched_group *busiest; /* Busiest group in this sd */
5128 struct sched_group *this; /* Local group in this sd */
5129 unsigned long total_load; /* Total load of all groups in sd */
5130 unsigned long total_pwr; /* Total power of all groups in sd */
5131 unsigned long avg_load; /* Average load across all groups in sd */
5133 /** Statistics of this group */
5134 unsigned long this_load;
5135 unsigned long this_load_per_task;
5136 unsigned long this_nr_running;
5137 unsigned long this_has_capacity;
5138 unsigned int this_idle_cpus;
5140 /* Statistics of the busiest group */
5141 unsigned int busiest_idle_cpus;
5142 unsigned long max_load;
5143 unsigned long busiest_load_per_task;
5144 unsigned long busiest_nr_running;
5145 unsigned long busiest_group_capacity;
5146 unsigned long busiest_has_capacity;
5147 unsigned int busiest_group_weight;
5149 int group_imb; /* Is there imbalance in this sd */
5153 * sg_lb_stats - stats of a sched_group required for load_balancing
5155 struct sg_lb_stats {
5156 unsigned long avg_load; /*Avg load across the CPUs of the group */
5157 unsigned long group_load; /* Total load over the CPUs of the group */
5158 unsigned long sum_nr_running; /* Nr tasks running in the group */
5159 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5160 unsigned long group_capacity;
5161 unsigned long idle_cpus;
5162 unsigned long group_weight;
5163 int group_imb; /* Is there an imbalance in the group ? */
5164 int group_has_capacity; /* Is there extra capacity in the group? */
5168 * get_sd_load_idx - Obtain the load index for a given sched domain.
5169 * @sd: The sched_domain whose load_idx is to be obtained.
5170 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5172 static inline int get_sd_load_idx(struct sched_domain *sd,
5173 enum cpu_idle_type idle)
5179 load_idx = sd->busy_idx;
5182 case CPU_NEWLY_IDLE:
5183 load_idx = sd->newidle_idx;
5186 load_idx = sd->idle_idx;
5193 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5195 return SCHED_POWER_SCALE;
5198 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5200 return default_scale_freq_power(sd, cpu);
5203 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5205 unsigned long weight = sd->span_weight;
5206 unsigned long smt_gain = sd->smt_gain;
5213 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5215 return default_scale_smt_power(sd, cpu);
5218 static unsigned long scale_rt_power(int cpu)
5220 struct rq *rq = cpu_rq(cpu);
5221 u64 total, available, age_stamp, avg;
5224 * Since we're reading these variables without serialization make sure
5225 * we read them once before doing sanity checks on them.
5227 age_stamp = ACCESS_ONCE(rq->age_stamp);
5228 avg = ACCESS_ONCE(rq->rt_avg);
5230 total = sched_avg_period() + (rq->clock - age_stamp);
5232 if (unlikely(total < avg)) {
5233 /* Ensures that power won't end up being negative */
5236 available = total - avg;
5239 if (unlikely((s64)total < SCHED_POWER_SCALE))
5240 total = SCHED_POWER_SCALE;
5242 total >>= SCHED_POWER_SHIFT;
5244 return div_u64(available, total);
5247 static void update_cpu_power(struct sched_domain *sd, int cpu)
5249 unsigned long weight = sd->span_weight;
5250 unsigned long power = SCHED_POWER_SCALE;
5251 struct sched_group *sdg = sd->groups;
5253 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5254 if (sched_feat(ARCH_POWER))
5255 power *= arch_scale_smt_power(sd, cpu);
5257 power *= default_scale_smt_power(sd, cpu);
5259 power >>= SCHED_POWER_SHIFT;
5262 sdg->sgp->power_orig = power;
5264 if (sched_feat(ARCH_POWER))
5265 power *= arch_scale_freq_power(sd, cpu);
5267 power *= default_scale_freq_power(sd, cpu);
5269 power >>= SCHED_POWER_SHIFT;
5271 power *= scale_rt_power(cpu);
5272 power >>= SCHED_POWER_SHIFT;
5277 cpu_rq(cpu)->cpu_power = power;
5278 sdg->sgp->power = power;
5281 void update_group_power(struct sched_domain *sd, int cpu)
5283 struct sched_domain *child = sd->child;
5284 struct sched_group *group, *sdg = sd->groups;
5285 unsigned long power;
5286 unsigned long interval;
5288 interval = msecs_to_jiffies(sd->balance_interval);
5289 interval = clamp(interval, 1UL, max_load_balance_interval);
5290 sdg->sgp->next_update = jiffies + interval;
5293 update_cpu_power(sd, cpu);
5299 if (child->flags & SD_OVERLAP) {
5301 * SD_OVERLAP domains cannot assume that child groups
5302 * span the current group.
5305 for_each_cpu(cpu, sched_group_cpus(sdg))
5306 power += power_of(cpu);
5309 * !SD_OVERLAP domains can assume that child groups
5310 * span the current group.
5313 group = child->groups;
5315 power += group->sgp->power;
5316 group = group->next;
5317 } while (group != child->groups);
5320 sdg->sgp->power_orig = sdg->sgp->power = power;
5324 * Try and fix up capacity for tiny siblings, this is needed when
5325 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5326 * which on its own isn't powerful enough.
5328 * See update_sd_pick_busiest() and check_asym_packing().
5331 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5334 * Only siblings can have significantly less than SCHED_POWER_SCALE
5336 if (!(sd->flags & SD_SHARE_CPUPOWER))
5340 * If ~90% of the cpu_power is still there, we're good.
5342 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5349 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5350 * @env: The load balancing environment.
5351 * @group: sched_group whose statistics are to be updated.
5352 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5353 * @local_group: Does group contain this_cpu.
5354 * @balance: Should we balance.
5355 * @sgs: variable to hold the statistics for this group.
5357 static inline void update_sg_lb_stats(struct lb_env *env,
5358 struct sched_group *group, int load_idx,
5359 int local_group, int *balance, struct sg_lb_stats *sgs)
5361 unsigned long nr_running, max_nr_running, min_nr_running;
5362 unsigned long load, max_cpu_load, min_cpu_load;
5363 unsigned int balance_cpu = -1, first_idle_cpu = 0;
5364 unsigned long avg_load_per_task = 0;
5368 balance_cpu = group_balance_cpu(group);
5370 /* Tally up the load of all CPUs in the group */
5372 min_cpu_load = ~0UL;
5374 min_nr_running = ~0UL;
5376 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5377 struct rq *rq = cpu_rq(i);
5379 nr_running = rq->nr_running;
5381 /* Bias balancing toward cpus of our domain */
5383 if (idle_cpu(i) && !first_idle_cpu &&
5384 cpumask_test_cpu(i, sched_group_mask(group))) {
5389 load = target_load(i, load_idx);
5391 load = source_load(i, load_idx);
5392 if (load > max_cpu_load)
5393 max_cpu_load = load;
5394 if (min_cpu_load > load)
5395 min_cpu_load = load;
5397 if (nr_running > max_nr_running)
5398 max_nr_running = nr_running;
5399 if (min_nr_running > nr_running)
5400 min_nr_running = nr_running;
5403 sgs->group_load += load;
5404 sgs->sum_nr_running += nr_running;
5405 sgs->sum_weighted_load += weighted_cpuload(i);
5411 * First idle cpu or the first cpu(busiest) in this sched group
5412 * is eligible for doing load balancing at this and above
5413 * domains. In the newly idle case, we will allow all the cpu's
5414 * to do the newly idle load balance.
5417 if (env->idle != CPU_NEWLY_IDLE) {
5418 if (balance_cpu != env->dst_cpu) {
5422 update_group_power(env->sd, env->dst_cpu);
5423 } else if (time_after_eq(jiffies, group->sgp->next_update))
5424 update_group_power(env->sd, env->dst_cpu);
5427 /* Adjust by relative CPU power of the group */
5428 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
5431 * Consider the group unbalanced when the imbalance is larger
5432 * than the average weight of a task.
5434 * APZ: with cgroup the avg task weight can vary wildly and
5435 * might not be a suitable number - should we keep a
5436 * normalized nr_running number somewhere that negates
5439 if (sgs->sum_nr_running)
5440 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5442 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
5443 (max_nr_running - min_nr_running) > 1)
5446 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
5448 if (!sgs->group_capacity)
5449 sgs->group_capacity = fix_small_capacity(env->sd, group);
5450 sgs->group_weight = group->group_weight;
5452 if (sgs->group_capacity > sgs->sum_nr_running)
5453 sgs->group_has_capacity = 1;
5457 * update_sd_pick_busiest - return 1 on busiest group
5458 * @env: The load balancing environment.
5459 * @sds: sched_domain statistics
5460 * @sg: sched_group candidate to be checked for being the busiest
5461 * @sgs: sched_group statistics
5463 * Determine if @sg is a busier group than the previously selected
5466 static bool update_sd_pick_busiest(struct lb_env *env,
5467 struct sd_lb_stats *sds,
5468 struct sched_group *sg,
5469 struct sg_lb_stats *sgs)
5471 if (sgs->avg_load <= sds->max_load)
5474 if (sgs->sum_nr_running > sgs->group_capacity)
5481 * ASYM_PACKING needs to move all the work to the lowest
5482 * numbered CPUs in the group, therefore mark all groups
5483 * higher than ourself as busy.
5485 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5486 env->dst_cpu < group_first_cpu(sg)) {
5490 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5498 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5499 * @env: The load balancing environment.
5500 * @balance: Should we balance.
5501 * @sds: variable to hold the statistics for this sched_domain.
5503 static inline void update_sd_lb_stats(struct lb_env *env,
5504 int *balance, struct sd_lb_stats *sds)
5506 struct sched_domain *child = env->sd->child;
5507 struct sched_group *sg = env->sd->groups;
5508 struct sg_lb_stats sgs;
5509 int load_idx, prefer_sibling = 0;
5511 if (child && child->flags & SD_PREFER_SIBLING)
5514 load_idx = get_sd_load_idx(env->sd, env->idle);
5519 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5520 memset(&sgs, 0, sizeof(sgs));
5521 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
5523 if (local_group && !(*balance))
5526 sds->total_load += sgs.group_load;
5527 sds->total_pwr += sg->sgp->power;
5530 * In case the child domain prefers tasks go to siblings
5531 * first, lower the sg capacity to one so that we'll try
5532 * and move all the excess tasks away. We lower the capacity
5533 * of a group only if the local group has the capacity to fit
5534 * these excess tasks, i.e. nr_running < group_capacity. The
5535 * extra check prevents the case where you always pull from the
5536 * heaviest group when it is already under-utilized (possible
5537 * with a large weight task outweighs the tasks on the system).
5539 if (prefer_sibling && !local_group && sds->this_has_capacity)
5540 sgs.group_capacity = min(sgs.group_capacity, 1UL);
5543 sds->this_load = sgs.avg_load;
5545 sds->this_nr_running = sgs.sum_nr_running;
5546 sds->this_load_per_task = sgs.sum_weighted_load;
5547 sds->this_has_capacity = sgs.group_has_capacity;
5548 sds->this_idle_cpus = sgs.idle_cpus;
5549 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
5550 sds->max_load = sgs.avg_load;
5552 sds->busiest_nr_running = sgs.sum_nr_running;
5553 sds->busiest_idle_cpus = sgs.idle_cpus;
5554 sds->busiest_group_capacity = sgs.group_capacity;
5555 sds->busiest_load_per_task = sgs.sum_weighted_load;
5556 sds->busiest_has_capacity = sgs.group_has_capacity;
5557 sds->busiest_group_weight = sgs.group_weight;
5558 sds->group_imb = sgs.group_imb;
5562 } while (sg != env->sd->groups);
5566 * check_asym_packing - Check to see if the group is packed into the
5569 * This is primarily intended to used at the sibling level. Some
5570 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5571 * case of POWER7, it can move to lower SMT modes only when higher
5572 * threads are idle. When in lower SMT modes, the threads will
5573 * perform better since they share less core resources. Hence when we
5574 * have idle threads, we want them to be the higher ones.
5576 * This packing function is run on idle threads. It checks to see if
5577 * the busiest CPU in this domain (core in the P7 case) has a higher
5578 * CPU number than the packing function is being run on. Here we are
5579 * assuming lower CPU number will be equivalent to lower a SMT thread
5582 * Returns 1 when packing is required and a task should be moved to
5583 * this CPU. The amount of the imbalance is returned in *imbalance.
5585 * @env: The load balancing environment.
5586 * @sds: Statistics of the sched_domain which is to be packed
5588 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5592 if (!(env->sd->flags & SD_ASYM_PACKING))
5598 busiest_cpu = group_first_cpu(sds->busiest);
5599 if (env->dst_cpu > busiest_cpu)
5602 env->imbalance = DIV_ROUND_CLOSEST(
5603 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
5609 * fix_small_imbalance - Calculate the minor imbalance that exists
5610 * amongst the groups of a sched_domain, during
5612 * @env: The load balancing environment.
5613 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5616 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5618 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5619 unsigned int imbn = 2;
5620 unsigned long scaled_busy_load_per_task;
5622 if (sds->this_nr_running) {
5623 sds->this_load_per_task /= sds->this_nr_running;
5624 if (sds->busiest_load_per_task >
5625 sds->this_load_per_task)
5628 sds->this_load_per_task =
5629 cpu_avg_load_per_task(env->dst_cpu);
5632 scaled_busy_load_per_task = sds->busiest_load_per_task
5633 * SCHED_POWER_SCALE;
5634 scaled_busy_load_per_task /= sds->busiest->sgp->power;
5636 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
5637 (scaled_busy_load_per_task * imbn)) {
5638 env->imbalance = sds->busiest_load_per_task;
5643 * OK, we don't have enough imbalance to justify moving tasks,
5644 * however we may be able to increase total CPU power used by
5648 pwr_now += sds->busiest->sgp->power *
5649 min(sds->busiest_load_per_task, sds->max_load);
5650 pwr_now += sds->this->sgp->power *
5651 min(sds->this_load_per_task, sds->this_load);
5652 pwr_now /= SCHED_POWER_SCALE;
5654 /* Amount of load we'd subtract */
5655 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5656 sds->busiest->sgp->power;
5657 if (sds->max_load > tmp)
5658 pwr_move += sds->busiest->sgp->power *
5659 min(sds->busiest_load_per_task, sds->max_load - tmp);
5661 /* Amount of load we'd add */
5662 if (sds->max_load * sds->busiest->sgp->power <
5663 sds->busiest_load_per_task * SCHED_POWER_SCALE)
5664 tmp = (sds->max_load * sds->busiest->sgp->power) /
5665 sds->this->sgp->power;
5667 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
5668 sds->this->sgp->power;
5669 pwr_move += sds->this->sgp->power *
5670 min(sds->this_load_per_task, sds->this_load + tmp);
5671 pwr_move /= SCHED_POWER_SCALE;
5673 /* Move if we gain throughput */
5674 if (pwr_move > pwr_now)
5675 env->imbalance = sds->busiest_load_per_task;
5679 * calculate_imbalance - Calculate the amount of imbalance present within the
5680 * groups of a given sched_domain during load balance.
5681 * @env: load balance environment
5682 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5684 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5686 unsigned long max_pull, load_above_capacity = ~0UL;
5688 sds->busiest_load_per_task /= sds->busiest_nr_running;
5689 if (sds->group_imb) {
5690 sds->busiest_load_per_task =
5691 min(sds->busiest_load_per_task, sds->avg_load);
5695 * In the presence of smp nice balancing, certain scenarios can have
5696 * max load less than avg load(as we skip the groups at or below
5697 * its cpu_power, while calculating max_load..)
5699 if (sds->max_load < sds->avg_load) {
5701 return fix_small_imbalance(env, sds);
5704 if (!sds->group_imb) {
5706 * Don't want to pull so many tasks that a group would go idle.
5708 load_above_capacity = (sds->busiest_nr_running -
5709 sds->busiest_group_capacity);
5711 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5713 load_above_capacity /= sds->busiest->sgp->power;
5717 * We're trying to get all the cpus to the average_load, so we don't
5718 * want to push ourselves above the average load, nor do we wish to
5719 * reduce the max loaded cpu below the average load. At the same time,
5720 * we also don't want to reduce the group load below the group capacity
5721 * (so that we can implement power-savings policies etc). Thus we look
5722 * for the minimum possible imbalance.
5723 * Be careful of negative numbers as they'll appear as very large values
5724 * with unsigned longs.
5726 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
5728 /* How much load to actually move to equalise the imbalance */
5729 env->imbalance = min(max_pull * sds->busiest->sgp->power,
5730 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
5731 / SCHED_POWER_SCALE;
5734 * if *imbalance is less than the average load per runnable task
5735 * there is no guarantee that any tasks will be moved so we'll have
5736 * a think about bumping its value to force at least one task to be
5739 if (env->imbalance < sds->busiest_load_per_task)
5740 return fix_small_imbalance(env, sds);
5744 /******* find_busiest_group() helpers end here *********************/
5747 * find_busiest_group - Returns the busiest group within the sched_domain
5748 * if there is an imbalance. If there isn't an imbalance, and
5749 * the user has opted for power-savings, it returns a group whose
5750 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5751 * such a group exists.
5753 * Also calculates the amount of weighted load which should be moved
5754 * to restore balance.
5756 * @env: The load balancing environment.
5757 * @balance: Pointer to a variable indicating if this_cpu
5758 * is the appropriate cpu to perform load balancing at this_level.
5760 * Returns: - the busiest group if imbalance exists.
5761 * - If no imbalance and user has opted for power-savings balance,
5762 * return the least loaded group whose CPUs can be
5763 * put to idle by rebalancing its tasks onto our group.
5765 static struct sched_group *
5766 find_busiest_group(struct lb_env *env, int *balance)
5768 struct sd_lb_stats sds;
5770 memset(&sds, 0, sizeof(sds));
5773 * Compute the various statistics relavent for load balancing at
5776 update_sd_lb_stats(env, balance, &sds);
5779 * this_cpu is not the appropriate cpu to perform load balancing at
5785 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5786 check_asym_packing(env, &sds))
5789 /* There is no busy sibling group to pull tasks from */
5790 if (!sds.busiest || sds.busiest_nr_running == 0)
5793 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5796 * If the busiest group is imbalanced the below checks don't
5797 * work because they assumes all things are equal, which typically
5798 * isn't true due to cpus_allowed constraints and the like.
5803 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5804 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
5805 !sds.busiest_has_capacity)
5809 * If the local group is more busy than the selected busiest group
5810 * don't try and pull any tasks.
5812 if (sds.this_load >= sds.max_load)
5816 * Don't pull any tasks if this group is already above the domain
5819 if (sds.this_load >= sds.avg_load)
5822 if (env->idle == CPU_IDLE) {
5824 * This cpu is idle. If the busiest group load doesn't
5825 * have more tasks than the number of available cpu's and
5826 * there is no imbalance between this and busiest group
5827 * wrt to idle cpu's, it is balanced.
5829 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
5830 sds.busiest_nr_running <= sds.busiest_group_weight)
5834 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5835 * imbalance_pct to be conservative.
5837 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
5842 /* Looks like there is an imbalance. Compute it */
5843 calculate_imbalance(env, &sds);
5853 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5855 static struct rq *find_busiest_queue(struct lb_env *env,
5856 struct sched_group *group)
5858 struct rq *busiest = NULL, *rq;
5859 unsigned long max_load = 0;
5862 for_each_cpu(i, sched_group_cpus(group)) {
5863 unsigned long power = power_of(i);
5864 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5869 capacity = fix_small_capacity(env->sd, group);
5871 if (!cpumask_test_cpu(i, env->cpus))
5875 wl = weighted_cpuload(i);
5878 * When comparing with imbalance, use weighted_cpuload()
5879 * which is not scaled with the cpu power.
5881 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5885 * For the load comparisons with the other cpu's, consider
5886 * the weighted_cpuload() scaled with the cpu power, so that
5887 * the load can be moved away from the cpu that is potentially
5888 * running at a lower capacity.
5890 wl = (wl * SCHED_POWER_SCALE) / power;
5892 if (wl > max_load) {
5902 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5903 * so long as it is large enough.
5905 #define MAX_PINNED_INTERVAL 512
5907 /* Working cpumask for load_balance and load_balance_newidle. */
5908 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5910 static int need_active_balance(struct lb_env *env)
5912 struct sched_domain *sd = env->sd;
5914 if (env->idle == CPU_NEWLY_IDLE) {
5917 * ASYM_PACKING needs to force migrate tasks from busy but
5918 * higher numbered CPUs in order to pack all tasks in the
5919 * lowest numbered CPUs.
5921 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5925 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5928 static int active_load_balance_cpu_stop(void *data);
5931 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5932 * tasks if there is an imbalance.
5934 static int load_balance(int this_cpu, struct rq *this_rq,
5935 struct sched_domain *sd, enum cpu_idle_type idle,
5938 int ld_moved, cur_ld_moved, active_balance = 0;
5939 struct sched_group *group;
5941 unsigned long flags;
5942 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5944 struct lb_env env = {
5946 .dst_cpu = this_cpu,
5948 .dst_grpmask = sched_group_cpus(sd->groups),
5950 .loop_break = sched_nr_migrate_break,
5955 * For NEWLY_IDLE load_balancing, we don't need to consider
5956 * other cpus in our group
5958 if (idle == CPU_NEWLY_IDLE)
5959 env.dst_grpmask = NULL;
5961 cpumask_copy(cpus, cpu_active_mask);
5963 schedstat_inc(sd, lb_count[idle]);
5966 group = find_busiest_group(&env, balance);
5972 schedstat_inc(sd, lb_nobusyg[idle]);
5976 busiest = find_busiest_queue(&env, group);
5978 schedstat_inc(sd, lb_nobusyq[idle]);
5982 BUG_ON(busiest == env.dst_rq);
5984 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5987 if (busiest->nr_running > 1) {
5989 * Attempt to move tasks. If find_busiest_group has found
5990 * an imbalance but busiest->nr_running <= 1, the group is
5991 * still unbalanced. ld_moved simply stays zero, so it is
5992 * correctly treated as an imbalance.
5994 env.flags |= LBF_ALL_PINNED;
5995 env.src_cpu = busiest->cpu;
5996 env.src_rq = busiest;
5997 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5999 update_h_load(env.src_cpu);
6001 local_irq_save(flags);
6002 double_rq_lock(env.dst_rq, busiest);
6005 * cur_ld_moved - load moved in current iteration
6006 * ld_moved - cumulative load moved across iterations
6008 cur_ld_moved = move_tasks(&env);
6009 ld_moved += cur_ld_moved;
6010 double_rq_unlock(env.dst_rq, busiest);
6011 local_irq_restore(flags);
6014 * some other cpu did the load balance for us.
6016 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6017 resched_cpu(env.dst_cpu);
6019 if (env.flags & LBF_NEED_BREAK) {
6020 env.flags &= ~LBF_NEED_BREAK;
6025 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6026 * us and move them to an alternate dst_cpu in our sched_group
6027 * where they can run. The upper limit on how many times we
6028 * iterate on same src_cpu is dependent on number of cpus in our
6031 * This changes load balance semantics a bit on who can move
6032 * load to a given_cpu. In addition to the given_cpu itself
6033 * (or a ilb_cpu acting on its behalf where given_cpu is
6034 * nohz-idle), we now have balance_cpu in a position to move
6035 * load to given_cpu. In rare situations, this may cause
6036 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6037 * _independently_ and at _same_ time to move some load to
6038 * given_cpu) causing exceess load to be moved to given_cpu.
6039 * This however should not happen so much in practice and
6040 * moreover subsequent load balance cycles should correct the
6041 * excess load moved.
6043 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6045 env.dst_rq = cpu_rq(env.new_dst_cpu);
6046 env.dst_cpu = env.new_dst_cpu;
6047 env.flags &= ~LBF_SOME_PINNED;
6049 env.loop_break = sched_nr_migrate_break;
6051 /* Prevent to re-select dst_cpu via env's cpus */
6052 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6055 * Go back to "more_balance" rather than "redo" since we
6056 * need to continue with same src_cpu.
6061 /* All tasks on this runqueue were pinned by CPU affinity */
6062 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6063 cpumask_clear_cpu(cpu_of(busiest), cpus);
6064 if (!cpumask_empty(cpus)) {
6066 env.loop_break = sched_nr_migrate_break;
6074 schedstat_inc(sd, lb_failed[idle]);
6076 * Increment the failure counter only on periodic balance.
6077 * We do not want newidle balance, which can be very
6078 * frequent, pollute the failure counter causing
6079 * excessive cache_hot migrations and active balances.
6081 if (idle != CPU_NEWLY_IDLE)
6082 sd->nr_balance_failed++;
6084 if (need_active_balance(&env)) {
6085 raw_spin_lock_irqsave(&busiest->lock, flags);
6087 /* don't kick the active_load_balance_cpu_stop,
6088 * if the curr task on busiest cpu can't be
6091 if (!cpumask_test_cpu(this_cpu,
6092 tsk_cpus_allowed(busiest->curr))) {
6093 raw_spin_unlock_irqrestore(&busiest->lock,
6095 env.flags |= LBF_ALL_PINNED;
6096 goto out_one_pinned;
6100 * ->active_balance synchronizes accesses to
6101 * ->active_balance_work. Once set, it's cleared
6102 * only after active load balance is finished.
6104 if (!busiest->active_balance) {
6105 busiest->active_balance = 1;
6106 busiest->push_cpu = this_cpu;
6109 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6111 if (active_balance) {
6112 stop_one_cpu_nowait(cpu_of(busiest),
6113 active_load_balance_cpu_stop, busiest,
6114 &busiest->active_balance_work);
6118 * We've kicked active balancing, reset the failure
6121 sd->nr_balance_failed = sd->cache_nice_tries+1;
6124 sd->nr_balance_failed = 0;
6126 if (likely(!active_balance)) {
6127 /* We were unbalanced, so reset the balancing interval */
6128 sd->balance_interval = sd->min_interval;
6131 * If we've begun active balancing, start to back off. This
6132 * case may not be covered by the all_pinned logic if there
6133 * is only 1 task on the busy runqueue (because we don't call
6136 if (sd->balance_interval < sd->max_interval)
6137 sd->balance_interval *= 2;
6143 schedstat_inc(sd, lb_balanced[idle]);
6145 sd->nr_balance_failed = 0;
6148 /* tune up the balancing interval */
6149 if (((env.flags & LBF_ALL_PINNED) &&
6150 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6151 (sd->balance_interval < sd->max_interval))
6152 sd->balance_interval *= 2;
6158 #ifdef CONFIG_SCHED_HMP
6159 static unsigned int hmp_idle_pull(int this_cpu);
6162 * idle_balance is called by schedule() if this_cpu is about to become
6163 * idle. Attempts to pull tasks from other CPUs.
6165 void idle_balance(int this_cpu, struct rq *this_rq)
6167 struct sched_domain *sd;
6168 int pulled_task = 0;
6169 unsigned long next_balance = jiffies + HZ;
6171 this_rq->idle_stamp = this_rq->clock;
6173 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6177 * Drop the rq->lock, but keep IRQ/preempt disabled.
6179 raw_spin_unlock(&this_rq->lock);
6181 update_blocked_averages(this_cpu);
6183 for_each_domain(this_cpu, sd) {
6184 unsigned long interval;
6187 if (!(sd->flags & SD_LOAD_BALANCE))
6190 if (sd->flags & SD_BALANCE_NEWIDLE) {
6191 /* If we've pulled tasks over stop searching: */
6192 pulled_task = load_balance(this_cpu, this_rq,
6193 sd, CPU_NEWLY_IDLE, &balance);
6196 interval = msecs_to_jiffies(sd->balance_interval);
6197 if (time_after(next_balance, sd->last_balance + interval))
6198 next_balance = sd->last_balance + interval;
6200 this_rq->idle_stamp = 0;
6205 #ifdef CONFIG_SCHED_HMP
6207 pulled_task = hmp_idle_pull(this_cpu);
6209 raw_spin_lock(&this_rq->lock);
6211 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6213 * We are going idle. next_balance may be set based on
6214 * a busy processor. So reset next_balance.
6216 this_rq->next_balance = next_balance;
6221 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6222 * running tasks off the busiest CPU onto idle CPUs. It requires at
6223 * least 1 task to be running on each physical CPU where possible, and
6224 * avoids physical / logical imbalances.
6226 static int active_load_balance_cpu_stop(void *data)
6228 struct rq *busiest_rq = data;
6229 int busiest_cpu = cpu_of(busiest_rq);
6230 int target_cpu = busiest_rq->push_cpu;
6231 struct rq *target_rq = cpu_rq(target_cpu);
6232 struct sched_domain *sd;
6234 raw_spin_lock_irq(&busiest_rq->lock);
6236 /* make sure the requested cpu hasn't gone down in the meantime */
6237 if (unlikely(busiest_cpu != smp_processor_id() ||
6238 !busiest_rq->active_balance))
6241 /* Is there any task to move? */
6242 if (busiest_rq->nr_running <= 1)
6246 * This condition is "impossible", if it occurs
6247 * we need to fix it. Originally reported by
6248 * Bjorn Helgaas on a 128-cpu setup.
6250 BUG_ON(busiest_rq == target_rq);
6252 /* move a task from busiest_rq to target_rq */
6253 double_lock_balance(busiest_rq, target_rq);
6255 /* Search for an sd spanning us and the target CPU. */
6257 for_each_domain(target_cpu, sd) {
6258 if ((sd->flags & SD_LOAD_BALANCE) &&
6259 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6264 struct lb_env env = {
6266 .dst_cpu = target_cpu,
6267 .dst_rq = target_rq,
6268 .src_cpu = busiest_rq->cpu,
6269 .src_rq = busiest_rq,
6273 schedstat_inc(sd, alb_count);
6275 if (move_one_task(&env))
6276 schedstat_inc(sd, alb_pushed);
6278 schedstat_inc(sd, alb_failed);
6281 double_unlock_balance(busiest_rq, target_rq);
6283 busiest_rq->active_balance = 0;
6284 raw_spin_unlock_irq(&busiest_rq->lock);
6288 #ifdef CONFIG_NO_HZ_COMMON
6290 * idle load balancing details
6291 * - When one of the busy CPUs notice that there may be an idle rebalancing
6292 * needed, they will kick the idle load balancer, which then does idle
6293 * load balancing for all the idle CPUs.
6296 cpumask_var_t idle_cpus_mask;
6298 unsigned long next_balance; /* in jiffy units */
6299 } nohz ____cacheline_aligned;
6301 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6303 * Decide if the tasks on the busy CPUs in the
6304 * littlest domain would benefit from an idle balance
6306 static int hmp_packing_ilb_needed(int cpu)
6308 struct hmp_domain *hmp;
6309 /* always allow ilb on non-slowest domain */
6310 if (!hmp_cpu_is_slowest(cpu))
6313 hmp = hmp_cpu_domain(cpu);
6314 for_each_cpu_and(cpu, &hmp->cpus, nohz.idle_cpus_mask) {
6315 /* only idle balance if a CPU is loaded over threshold */
6316 if (cpu_rq(cpu)->avg.load_avg_ratio > hmp_full_threshold)
6323 static inline int find_new_ilb(int call_cpu)
6325 int ilb = cpumask_first(nohz.idle_cpus_mask);
6326 #ifdef CONFIG_SCHED_HMP
6329 /* restrict nohz balancing to occur in the same hmp domain */
6330 ilb = cpumask_first_and(nohz.idle_cpus_mask,
6331 &((struct hmp_domain *)hmp_cpu_domain(call_cpu))->cpus);
6333 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6334 if (ilb < nr_cpu_ids)
6335 ilb_needed = hmp_packing_ilb_needed(ilb);
6338 if (ilb_needed && ilb < nr_cpu_ids && idle_cpu(ilb))
6341 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6349 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6350 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6351 * CPU (if there is one).
6353 static void nohz_balancer_kick(int cpu)
6357 nohz.next_balance++;
6359 ilb_cpu = find_new_ilb(cpu);
6361 if (ilb_cpu >= nr_cpu_ids)
6364 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6367 * Use smp_send_reschedule() instead of resched_cpu().
6368 * This way we generate a sched IPI on the target cpu which
6369 * is idle. And the softirq performing nohz idle load balance
6370 * will be run before returning from the IPI.
6372 smp_send_reschedule(ilb_cpu);
6376 static inline void nohz_balance_exit_idle(int cpu)
6378 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6379 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6380 atomic_dec(&nohz.nr_cpus);
6381 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6385 static inline void set_cpu_sd_state_busy(void)
6387 struct sched_domain *sd;
6388 int cpu = smp_processor_id();
6391 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6393 if (!sd || !sd->nohz_idle)
6397 for (; sd; sd = sd->parent)
6398 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6403 void set_cpu_sd_state_idle(void)
6405 struct sched_domain *sd;
6406 int cpu = smp_processor_id();
6409 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
6411 if (!sd || sd->nohz_idle)
6415 for (; sd; sd = sd->parent)
6416 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6422 * This routine will record that the cpu is going idle with tick stopped.
6423 * This info will be used in performing idle load balancing in the future.
6425 void nohz_balance_enter_idle(int cpu)
6428 * If this cpu is going down, then nothing needs to be done.
6430 if (!cpu_active(cpu))
6433 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6436 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6437 atomic_inc(&nohz.nr_cpus);
6438 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6441 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
6442 unsigned long action, void *hcpu)
6444 switch (action & ~CPU_TASKS_FROZEN) {
6446 nohz_balance_exit_idle(smp_processor_id());
6454 static DEFINE_SPINLOCK(balancing);
6457 * Scale the max load_balance interval with the number of CPUs in the system.
6458 * This trades load-balance latency on larger machines for less cross talk.
6460 void update_max_interval(void)
6462 max_load_balance_interval = HZ*num_online_cpus()/10;
6466 * It checks each scheduling domain to see if it is due to be balanced,
6467 * and initiates a balancing operation if so.
6469 * Balancing parameters are set up in init_sched_domains.
6471 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6474 struct rq *rq = cpu_rq(cpu);
6475 unsigned long interval;
6476 struct sched_domain *sd;
6477 /* Earliest time when we have to do rebalance again */
6478 unsigned long next_balance = jiffies + 60*HZ;
6479 int update_next_balance = 0;
6482 update_blocked_averages(cpu);
6485 for_each_domain(cpu, sd) {
6486 if (!(sd->flags & SD_LOAD_BALANCE))
6489 interval = sd->balance_interval;
6490 if (idle != CPU_IDLE)
6491 interval *= sd->busy_factor;
6493 /* scale ms to jiffies */
6494 interval = msecs_to_jiffies(interval);
6495 interval = clamp(interval, 1UL, max_load_balance_interval);
6497 need_serialize = sd->flags & SD_SERIALIZE;
6499 if (need_serialize) {
6500 if (!spin_trylock(&balancing))
6504 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6505 if (load_balance(cpu, rq, sd, idle, &balance)) {
6507 * The LBF_SOME_PINNED logic could have changed
6508 * env->dst_cpu, so we can't know our idle
6509 * state even if we migrated tasks. Update it.
6511 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6513 sd->last_balance = jiffies;
6516 spin_unlock(&balancing);
6518 if (time_after(next_balance, sd->last_balance + interval)) {
6519 next_balance = sd->last_balance + interval;
6520 update_next_balance = 1;
6524 * Stop the load balance at this level. There is another
6525 * CPU in our sched group which is doing load balancing more
6534 * next_balance will be updated only when there is a need.
6535 * When the cpu is attached to null domain for ex, it will not be
6538 if (likely(update_next_balance))
6539 rq->next_balance = next_balance;
6542 #ifdef CONFIG_NO_HZ_COMMON
6544 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6545 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6547 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6549 struct rq *this_rq = cpu_rq(this_cpu);
6553 if (idle != CPU_IDLE ||
6554 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6557 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6558 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6562 * If this cpu gets work to do, stop the load balancing
6563 * work being done for other cpus. Next load
6564 * balancing owner will pick it up.
6569 rq = cpu_rq(balance_cpu);
6571 raw_spin_lock_irq(&rq->lock);
6572 update_rq_clock(rq);
6573 update_idle_cpu_load(rq);
6574 raw_spin_unlock_irq(&rq->lock);
6576 rebalance_domains(balance_cpu, CPU_IDLE);
6578 if (time_after(this_rq->next_balance, rq->next_balance))
6579 this_rq->next_balance = rq->next_balance;
6581 nohz.next_balance = this_rq->next_balance;
6583 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6587 * Current heuristic for kicking the idle load balancer in the presence
6588 * of an idle cpu is the system.
6589 * - This rq has more than one task.
6590 * - At any scheduler domain level, this cpu's scheduler group has multiple
6591 * busy cpu's exceeding the group's power.
6592 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6593 * domain span are idle.
6595 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6597 unsigned long now = jiffies;
6598 struct sched_domain *sd;
6600 if (unlikely(idle_cpu(cpu)))
6604 * We may be recently in ticked or tickless idle mode. At the first
6605 * busy tick after returning from idle, we will update the busy stats.
6607 set_cpu_sd_state_busy();
6608 nohz_balance_exit_idle(cpu);
6611 * None are in tickless mode and hence no need for NOHZ idle load
6614 if (likely(!atomic_read(&nohz.nr_cpus)))
6617 if (time_before(now, nohz.next_balance))
6620 #ifdef CONFIG_SCHED_HMP
6622 * Bail out if there are no nohz CPUs in our
6623 * HMP domain, since we will move tasks between
6624 * domains through wakeup and force balancing
6625 * as necessary based upon task load.
6627 if (cpumask_first_and(nohz.idle_cpus_mask,
6628 &((struct hmp_domain *)hmp_cpu_domain(cpu))->cpus) >= nr_cpu_ids)
6632 if (rq->nr_running >= 2)
6636 for_each_domain(cpu, sd) {
6637 struct sched_group *sg = sd->groups;
6638 struct sched_group_power *sgp = sg->sgp;
6639 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6641 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6642 goto need_kick_unlock;
6644 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6645 && (cpumask_first_and(nohz.idle_cpus_mask,
6646 sched_domain_span(sd)) < cpu))
6647 goto need_kick_unlock;
6649 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6661 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6664 #ifdef CONFIG_SCHED_HMP
6665 /* Check if task should migrate to a faster cpu */
6666 static unsigned int hmp_up_migration(int cpu, int *target_cpu, struct sched_entity *se)
6668 struct task_struct *p = task_of(se);
6669 int temp_target_cpu;
6672 if (hmp_cpu_is_fastest(cpu))
6675 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6676 /* Filter by task priority */
6677 if (p->prio >= hmp_up_prio)
6680 if (se->avg.load_avg_ratio < hmp_up_threshold)
6683 /* Let the task load settle before doing another up migration */
6684 /* hack - always use clock from first online CPU */
6685 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6686 if (((now - se->avg.hmp_last_up_migration) >> 10)
6687 < hmp_next_up_threshold)
6690 /* hmp_domain_min_load only returns 0 for an
6691 * idle CPU or 1023 for any partly-busy one.
6692 * Be explicit about requirement for an idle CPU.
6694 if (hmp_domain_min_load(hmp_faster_domain(cpu), &temp_target_cpu,
6695 tsk_cpus_allowed(p)) == 0 && temp_target_cpu != NR_CPUS) {
6697 *target_cpu = temp_target_cpu;
6703 /* Check if task should migrate to a slower cpu */
6704 static unsigned int hmp_down_migration(int cpu, struct sched_entity *se)
6706 struct task_struct *p = task_of(se);
6709 if (hmp_cpu_is_slowest(cpu)) {
6710 #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING
6711 if(hmp_packing_enabled)
6718 #ifdef CONFIG_SCHED_HMP_PRIO_FILTER
6719 /* Filter by task priority */
6720 if ((p->prio >= hmp_up_prio) &&
6721 cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6722 tsk_cpus_allowed(p))) {
6727 /* Let the task load settle before doing another down migration */
6728 /* hack - always use clock from first online CPU */
6729 now = cpu_rq(cpumask_first(cpu_online_mask))->clock_task;
6730 if (((now - se->avg.hmp_last_down_migration) >> 10)
6731 < hmp_next_down_threshold)
6734 if (cpumask_intersects(&hmp_slower_domain(cpu)->cpus,
6735 tsk_cpus_allowed(p))
6736 && se->avg.load_avg_ratio < hmp_down_threshold) {
6743 * hmp_can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6744 * Ideally this function should be merged with can_migrate_task() to avoid
6747 static int hmp_can_migrate_task(struct task_struct *p, struct lb_env *env)
6749 int tsk_cache_hot = 0;
6752 * We do not migrate tasks that are:
6753 * 1) running (obviously), or
6754 * 2) cannot be migrated to this CPU due to cpus_allowed
6756 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6757 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6760 env->flags &= ~LBF_ALL_PINNED;
6762 if (task_running(env->src_rq, p)) {
6763 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6768 * Aggressive migration if:
6769 * 1) task is cache cold, or
6770 * 2) too many balance attempts have failed.
6773 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
6774 if (!tsk_cache_hot ||
6775 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6776 #ifdef CONFIG_SCHEDSTATS
6777 if (tsk_cache_hot) {
6778 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6779 schedstat_inc(p, se.statistics.nr_forced_migrations);
6789 * move_specific_task tries to move a specific task.
6790 * Returns 1 if successful and 0 otherwise.
6791 * Called with both runqueues locked.
6793 static int move_specific_task(struct lb_env *env, struct task_struct *pm)
6795 struct task_struct *p, *n;
6797 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6798 if (throttled_lb_pair(task_group(p), env->src_rq->cpu,
6802 if (!hmp_can_migrate_task(p, env))
6804 /* Check if we found the right task */
6810 * Right now, this is only the third place move_task()
6811 * is called, so we can safely collect move_task()
6812 * stats here rather than inside move_task().
6814 schedstat_inc(env->sd, lb_gained[env->idle]);
6821 * hmp_active_task_migration_cpu_stop is run by cpu stopper and used to
6822 * migrate a specific task from one runqueue to another.
6823 * hmp_force_up_migration uses this to push a currently running task
6825 * Based on active_load_balance_stop_cpu and can potentially be merged.
6827 static int hmp_active_task_migration_cpu_stop(void *data)
6829 struct rq *busiest_rq = data;
6830 struct task_struct *p = busiest_rq->migrate_task;
6831 int busiest_cpu = cpu_of(busiest_rq);
6832 int target_cpu = busiest_rq->push_cpu;
6833 struct rq *target_rq = cpu_rq(target_cpu);
6834 struct sched_domain *sd;
6836 raw_spin_lock_irq(&busiest_rq->lock);
6837 /* make sure the requested cpu hasn't gone down in the meantime */
6838 if (unlikely(busiest_cpu != smp_processor_id() ||
6839 !busiest_rq->active_balance)) {
6842 /* Is there any task to move? */
6843 if (busiest_rq->nr_running <= 1)
6845 /* Task has migrated meanwhile, abort forced migration */
6846 if (task_rq(p) != busiest_rq)
6849 * This condition is "impossible", if it occurs
6850 * we need to fix it. Originally reported by
6851 * Bjorn Helgaas on a 128-cpu setup.
6853 BUG_ON(busiest_rq == target_rq);
6855 /* move a task from busiest_rq to target_rq */
6856 double_lock_balance(busiest_rq, target_rq);
6858 /* Search for an sd spanning us and the target CPU. */
6860 for_each_domain(target_cpu, sd) {
6861 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6866 struct lb_env env = {
6868 .dst_cpu = target_cpu,
6869 .dst_rq = target_rq,
6870 .src_cpu = busiest_rq->cpu,
6871 .src_rq = busiest_rq,
6875 schedstat_inc(sd, alb_count);
6877 if (move_specific_task(&env, p))
6878 schedstat_inc(sd, alb_pushed);
6880 schedstat_inc(sd, alb_failed);
6883 double_unlock_balance(busiest_rq, target_rq);
6886 busiest_rq->active_balance = 0;
6887 raw_spin_unlock_irq(&busiest_rq->lock);
6892 * hmp_idle_pull_cpu_stop is run by cpu stopper and used to
6893 * migrate a specific task from one runqueue to another.
6894 * hmp_idle_pull uses this to push a currently running task
6895 * off a runqueue to a faster CPU.
6896 * Locking is slightly different than usual.
6897 * Based on active_load_balance_stop_cpu and can potentially be merged.
6899 static int hmp_idle_pull_cpu_stop(void *data)
6901 struct rq *busiest_rq = data;
6902 struct task_struct *p = busiest_rq->migrate_task;
6903 int busiest_cpu = cpu_of(busiest_rq);
6904 int target_cpu = busiest_rq->push_cpu;
6905 struct rq *target_rq = cpu_rq(target_cpu);
6906 struct sched_domain *sd;
6908 raw_spin_lock_irq(&busiest_rq->lock);
6910 /* make sure the requested cpu hasn't gone down in the meantime */
6911 if (unlikely(busiest_cpu != smp_processor_id() ||
6912 !busiest_rq->active_balance))
6915 /* Is there any task to move? */
6916 if (busiest_rq->nr_running <= 1)
6919 /* Task has migrated meanwhile, abort forced migration */
6920 if (task_rq(p) != busiest_rq)
6924 * This condition is "impossible", if it occurs
6925 * we need to fix it. Originally reported by
6926 * Bjorn Helgaas on a 128-cpu setup.
6928 BUG_ON(busiest_rq == target_rq);
6930 /* move a task from busiest_rq to target_rq */
6931 double_lock_balance(busiest_rq, target_rq);
6933 /* Search for an sd spanning us and the target CPU. */
6935 for_each_domain(target_cpu, sd) {
6936 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6940 struct lb_env env = {
6942 .dst_cpu = target_cpu,
6943 .dst_rq = target_rq,
6944 .src_cpu = busiest_rq->cpu,
6945 .src_rq = busiest_rq,
6949 schedstat_inc(sd, alb_count);
6951 if (move_specific_task(&env, p))
6952 schedstat_inc(sd, alb_pushed);
6954 schedstat_inc(sd, alb_failed);
6957 double_unlock_balance(busiest_rq, target_rq);
6960 busiest_rq->active_balance = 0;
6961 raw_spin_unlock_irq(&busiest_rq->lock);
6966 * Move task in a runnable state to another CPU.
6968 * Tailored on 'active_load_balance_stop_cpu' with slight
6969 * modification to locking and pre-transfer checks. Note
6970 * rq->lock must be held before calling.
6972 static void hmp_migrate_runnable_task(struct rq *rq)
6974 struct sched_domain *sd;
6975 int src_cpu = cpu_of(rq);
6976 struct rq *src_rq = rq;
6977 int dst_cpu = rq->push_cpu;
6978 struct rq *dst_rq = cpu_rq(dst_cpu);
6979 struct task_struct *p = rq->migrate_task;
6981 * One last check to make sure nobody else is playing
6982 * with the source rq.
6984 if (src_rq->active_balance)
6987 if (src_rq->nr_running <= 1)
6990 if (task_rq(p) != src_rq)
6993 * Not sure if this applies here but one can never
6996 BUG_ON(src_rq == dst_rq);
6998 double_lock_balance(src_rq, dst_rq);
7001 for_each_domain(dst_cpu, sd) {
7002 if (cpumask_test_cpu(src_cpu, sched_domain_span(sd)))
7007 struct lb_env env = {
7016 schedstat_inc(sd, alb_count);
7018 if (move_specific_task(&env, p))
7019 schedstat_inc(sd, alb_pushed);
7021 schedstat_inc(sd, alb_failed);
7025 double_unlock_balance(src_rq, dst_rq);
7028 static DEFINE_SPINLOCK(hmp_force_migration);
7031 * hmp_force_up_migration checks runqueues for tasks that need to
7032 * be actively migrated to a faster cpu.
7034 static void hmp_force_up_migration(int this_cpu)
7036 int cpu, target_cpu;
7037 struct sched_entity *curr, *orig;
7039 unsigned long flags;
7040 unsigned int force, got_target;
7041 struct task_struct *p;
7043 if (!spin_trylock(&hmp_force_migration))
7045 for_each_online_cpu(cpu) {
7048 target = cpu_rq(cpu);
7049 raw_spin_lock_irqsave(&target->lock, flags);
7050 curr = target->cfs.curr;
7052 raw_spin_unlock_irqrestore(&target->lock, flags);
7055 if (!entity_is_task(curr)) {
7056 struct cfs_rq *cfs_rq;
7058 cfs_rq = group_cfs_rq(curr);
7060 curr = cfs_rq->curr;
7061 cfs_rq = group_cfs_rq(curr);
7065 curr = hmp_get_heaviest_task(curr, 1);
7067 if (hmp_up_migration(cpu, &target_cpu, curr)) {
7068 if (!target->active_balance) {
7070 target->push_cpu = target_cpu;
7071 target->migrate_task = p;
7073 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_FORCE);
7074 hmp_next_up_delay(&p->se, target->push_cpu);
7077 if (!got_target && !target->active_balance) {
7079 * For now we just check the currently running task.
7080 * Selecting the lightest task for offloading will
7081 * require extensive book keeping.
7083 curr = hmp_get_lightest_task(orig, 1);
7085 target->push_cpu = hmp_offload_down(cpu, curr);
7086 if (target->push_cpu < NR_CPUS) {
7088 target->migrate_task = p;
7090 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_OFFLOAD);
7091 hmp_next_down_delay(&p->se, target->push_cpu);
7095 * We have a target with no active_balance. If the task
7096 * is not currently running move it, otherwise let the
7097 * CPU stopper take care of it.
7099 if (got_target && !target->active_balance) {
7100 if (!task_running(target, p)) {
7101 trace_sched_hmp_migrate_force_running(p, 0);
7102 hmp_migrate_runnable_task(target);
7104 target->active_balance = 1;
7109 raw_spin_unlock_irqrestore(&target->lock, flags);
7112 stop_one_cpu_nowait(cpu_of(target),
7113 hmp_active_task_migration_cpu_stop,
7114 target, &target->active_balance_work);
7116 spin_unlock(&hmp_force_migration);
7119 * hmp_idle_pull looks at little domain runqueues to see
7120 * if a task should be pulled.
7122 * Reuses hmp_force_migration spinlock.
7125 static unsigned int hmp_idle_pull(int this_cpu)
7128 struct sched_entity *curr, *orig;
7129 struct hmp_domain *hmp_domain = NULL;
7130 struct rq *target = NULL, *rq;
7131 unsigned long flags, ratio = 0;
7132 unsigned int force = 0;
7133 struct task_struct *p = NULL;
7135 if (!hmp_cpu_is_slowest(this_cpu))
7136 hmp_domain = hmp_slower_domain(this_cpu);
7140 if (!spin_trylock(&hmp_force_migration))
7143 /* first select a task */
7144 for_each_cpu(cpu, &hmp_domain->cpus) {
7146 raw_spin_lock_irqsave(&rq->lock, flags);
7147 curr = rq->cfs.curr;
7149 raw_spin_unlock_irqrestore(&rq->lock, flags);
7152 if (!entity_is_task(curr)) {
7153 struct cfs_rq *cfs_rq;
7155 cfs_rq = group_cfs_rq(curr);
7157 curr = cfs_rq->curr;
7158 if (!entity_is_task(curr))
7159 cfs_rq = group_cfs_rq(curr);
7165 curr = hmp_get_heaviest_task(curr, 1);
7166 if (curr->avg.load_avg_ratio > hmp_up_threshold &&
7167 curr->avg.load_avg_ratio > ratio) {
7170 ratio = curr->avg.load_avg_ratio;
7172 raw_spin_unlock_irqrestore(&rq->lock, flags);
7178 /* now we have a candidate */
7179 raw_spin_lock_irqsave(&target->lock, flags);
7180 if (!target->active_balance && task_rq(p) == target) {
7182 target->push_cpu = this_cpu;
7183 target->migrate_task = p;
7184 trace_sched_hmp_migrate(p, target->push_cpu, HMP_MIGRATE_IDLE_PULL);
7185 hmp_next_up_delay(&p->se, target->push_cpu);
7187 * if the task isn't running move it right away.
7188 * Otherwise setup the active_balance mechanic and let
7189 * the CPU stopper do its job.
7191 if (!task_running(target, p)) {
7192 trace_sched_hmp_migrate_idle_running(p, 0);
7193 hmp_migrate_runnable_task(target);
7195 target->active_balance = 1;
7199 raw_spin_unlock_irqrestore(&target->lock, flags);
7202 stop_one_cpu_nowait(cpu_of(target),
7203 hmp_idle_pull_cpu_stop,
7204 target, &target->active_balance_work);
7207 spin_unlock(&hmp_force_migration);
7211 static void hmp_force_up_migration(int this_cpu) { }
7212 #endif /* CONFIG_SCHED_HMP */
7215 * run_rebalance_domains is triggered when needed from the scheduler tick.
7216 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7218 static void run_rebalance_domains(struct softirq_action *h)
7220 int this_cpu = smp_processor_id();
7221 struct rq *this_rq = cpu_rq(this_cpu);
7222 enum cpu_idle_type idle = this_rq->idle_balance ?
7223 CPU_IDLE : CPU_NOT_IDLE;
7225 hmp_force_up_migration(this_cpu);
7227 rebalance_domains(this_cpu, idle);
7230 * If this cpu has a pending nohz_balance_kick, then do the
7231 * balancing on behalf of the other idle cpus whose ticks are
7234 nohz_idle_balance(this_cpu, idle);
7237 static inline int on_null_domain(int cpu)
7239 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
7243 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7245 void trigger_load_balance(struct rq *rq, int cpu)
7247 /* Don't need to rebalance while attached to NULL domain */
7248 if (time_after_eq(jiffies, rq->next_balance) &&
7249 likely(!on_null_domain(cpu)))
7250 raise_softirq(SCHED_SOFTIRQ);
7251 #ifdef CONFIG_NO_HZ_COMMON
7252 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
7253 nohz_balancer_kick(cpu);
7257 static void rq_online_fair(struct rq *rq)
7259 #ifdef CONFIG_SCHED_HMP
7260 hmp_online_cpu(rq->cpu);
7265 static void rq_offline_fair(struct rq *rq)
7267 #ifdef CONFIG_SCHED_HMP
7268 hmp_offline_cpu(rq->cpu);
7272 /* Ensure any throttled groups are reachable by pick_next_task */
7273 unthrottle_offline_cfs_rqs(rq);
7276 #endif /* CONFIG_SMP */
7279 * scheduler tick hitting a task of our scheduling class:
7281 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7283 struct cfs_rq *cfs_rq;
7284 struct sched_entity *se = &curr->se;
7286 for_each_sched_entity(se) {
7287 cfs_rq = cfs_rq_of(se);
7288 entity_tick(cfs_rq, se, queued);
7291 if (sched_feat_numa(NUMA))
7292 task_tick_numa(rq, curr);
7294 update_rq_runnable_avg(rq, 1);
7298 * called on fork with the child task as argument from the parent's context
7299 * - child not yet on the tasklist
7300 * - preemption disabled
7302 static void task_fork_fair(struct task_struct *p)
7304 struct cfs_rq *cfs_rq;
7305 struct sched_entity *se = &p->se, *curr;
7306 int this_cpu = smp_processor_id();
7307 struct rq *rq = this_rq();
7308 unsigned long flags;
7310 raw_spin_lock_irqsave(&rq->lock, flags);
7312 update_rq_clock(rq);
7314 cfs_rq = task_cfs_rq(current);
7315 curr = cfs_rq->curr;
7318 * Not only the cpu but also the task_group of the parent might have
7319 * been changed after parent->se.parent,cfs_rq were copied to
7320 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7321 * of child point to valid ones.
7324 __set_task_cpu(p, this_cpu);
7327 update_curr(cfs_rq);
7330 se->vruntime = curr->vruntime;
7331 place_entity(cfs_rq, se, 1);
7333 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7335 * Upon rescheduling, sched_class::put_prev_task() will place
7336 * 'current' within the tree based on its new key value.
7338 swap(curr->vruntime, se->vruntime);
7339 resched_task(rq->curr);
7342 se->vruntime -= cfs_rq->min_vruntime;
7344 raw_spin_unlock_irqrestore(&rq->lock, flags);
7348 * Priority of the task has changed. Check to see if we preempt
7352 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7358 * Reschedule if we are currently running on this runqueue and
7359 * our priority decreased, or if we are not currently running on
7360 * this runqueue and our priority is higher than the current's
7362 if (rq->curr == p) {
7363 if (p->prio > oldprio)
7364 resched_task(rq->curr);
7366 check_preempt_curr(rq, p, 0);
7369 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7371 struct sched_entity *se = &p->se;
7372 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7375 * Ensure the task's vruntime is normalized, so that when its
7376 * switched back to the fair class the enqueue_entity(.flags=0) will
7377 * do the right thing.
7379 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7380 * have normalized the vruntime, if it was !on_rq, then only when
7381 * the task is sleeping will it still have non-normalized vruntime.
7383 if (!se->on_rq && p->state != TASK_RUNNING) {
7385 * Fix up our vruntime so that the current sleep doesn't
7386 * cause 'unlimited' sleep bonus.
7388 place_entity(cfs_rq, se, 0);
7389 se->vruntime -= cfs_rq->min_vruntime;
7392 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7394 * Remove our load from contribution when we leave sched_fair
7395 * and ensure we don't carry in an old decay_count if we
7398 if (p->se.avg.decay_count) {
7399 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
7400 __synchronize_entity_decay(&p->se);
7401 subtract_blocked_load_contrib(cfs_rq,
7402 p->se.avg.load_avg_contrib);
7408 * We switched to the sched_fair class.
7410 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7416 * We were most likely switched from sched_rt, so
7417 * kick off the schedule if running, otherwise just see
7418 * if we can still preempt the current task.
7421 resched_task(rq->curr);
7423 check_preempt_curr(rq, p, 0);
7426 /* Account for a task changing its policy or group.
7428 * This routine is mostly called to set cfs_rq->curr field when a task
7429 * migrates between groups/classes.
7431 static void set_curr_task_fair(struct rq *rq)
7433 struct sched_entity *se = &rq->curr->se;
7435 for_each_sched_entity(se) {
7436 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7438 set_next_entity(cfs_rq, se);
7439 /* ensure bandwidth has been allocated on our new cfs_rq */
7440 account_cfs_rq_runtime(cfs_rq, 0);
7444 void init_cfs_rq(struct cfs_rq *cfs_rq)
7446 cfs_rq->tasks_timeline = RB_ROOT;
7447 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7448 #ifndef CONFIG_64BIT
7449 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7451 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
7452 atomic64_set(&cfs_rq->decay_counter, 1);
7453 atomic64_set(&cfs_rq->removed_load, 0);
7457 #ifdef CONFIG_FAIR_GROUP_SCHED
7458 static void task_move_group_fair(struct task_struct *p, int on_rq)
7460 struct cfs_rq *cfs_rq;
7462 * If the task was not on the rq at the time of this cgroup movement
7463 * it must have been asleep, sleeping tasks keep their ->vruntime
7464 * absolute on their old rq until wakeup (needed for the fair sleeper
7465 * bonus in place_entity()).
7467 * If it was on the rq, we've just 'preempted' it, which does convert
7468 * ->vruntime to a relative base.
7470 * Make sure both cases convert their relative position when migrating
7471 * to another cgroup's rq. This does somewhat interfere with the
7472 * fair sleeper stuff for the first placement, but who cares.
7475 * When !on_rq, vruntime of the task has usually NOT been normalized.
7476 * But there are some cases where it has already been normalized:
7478 * - Moving a forked child which is waiting for being woken up by
7479 * wake_up_new_task().
7480 * - Moving a task which has been woken up by try_to_wake_up() and
7481 * waiting for actually being woken up by sched_ttwu_pending().
7483 * To prevent boost or penalty in the new cfs_rq caused by delta
7484 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7486 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7490 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7491 set_task_rq(p, task_cpu(p));
7493 cfs_rq = cfs_rq_of(&p->se);
7494 p->se.vruntime += cfs_rq->min_vruntime;
7497 * migrate_task_rq_fair() will have removed our previous
7498 * contribution, but we must synchronize for ongoing future
7501 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7502 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7507 void free_fair_sched_group(struct task_group *tg)
7511 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7513 for_each_possible_cpu(i) {
7515 kfree(tg->cfs_rq[i]);
7524 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7526 struct cfs_rq *cfs_rq;
7527 struct sched_entity *se;
7530 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7533 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7537 tg->shares = NICE_0_LOAD;
7539 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7541 for_each_possible_cpu(i) {
7542 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7543 GFP_KERNEL, cpu_to_node(i));
7547 se = kzalloc_node(sizeof(struct sched_entity),
7548 GFP_KERNEL, cpu_to_node(i));
7552 init_cfs_rq(cfs_rq);
7553 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7564 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7566 struct rq *rq = cpu_rq(cpu);
7567 unsigned long flags;
7570 * Only empty task groups can be destroyed; so we can speculatively
7571 * check on_list without danger of it being re-added.
7573 if (!tg->cfs_rq[cpu]->on_list)
7576 raw_spin_lock_irqsave(&rq->lock, flags);
7577 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7578 raw_spin_unlock_irqrestore(&rq->lock, flags);
7581 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7582 struct sched_entity *se, int cpu,
7583 struct sched_entity *parent)
7585 struct rq *rq = cpu_rq(cpu);
7589 init_cfs_rq_runtime(cfs_rq);
7591 tg->cfs_rq[cpu] = cfs_rq;
7594 /* se could be NULL for root_task_group */
7599 se->cfs_rq = &rq->cfs;
7601 se->cfs_rq = parent->my_q;
7604 update_load_set(&se->load, 0);
7605 se->parent = parent;
7608 static DEFINE_MUTEX(shares_mutex);
7610 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7613 unsigned long flags;
7616 * We can't change the weight of the root cgroup.
7621 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7623 mutex_lock(&shares_mutex);
7624 if (tg->shares == shares)
7627 tg->shares = shares;
7628 for_each_possible_cpu(i) {
7629 struct rq *rq = cpu_rq(i);
7630 struct sched_entity *se;
7633 /* Propagate contribution to hierarchy */
7634 raw_spin_lock_irqsave(&rq->lock, flags);
7635 for_each_sched_entity(se)
7636 update_cfs_shares(group_cfs_rq(se));
7637 raw_spin_unlock_irqrestore(&rq->lock, flags);
7641 mutex_unlock(&shares_mutex);
7644 #else /* CONFIG_FAIR_GROUP_SCHED */
7646 void free_fair_sched_group(struct task_group *tg) { }
7648 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7653 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7655 #endif /* CONFIG_FAIR_GROUP_SCHED */
7658 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7660 struct sched_entity *se = &task->se;
7661 unsigned int rr_interval = 0;
7664 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7667 if (rq->cfs.load.weight)
7668 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7674 * All the scheduling class methods:
7676 const struct sched_class fair_sched_class = {
7677 .next = &idle_sched_class,
7678 .enqueue_task = enqueue_task_fair,
7679 .dequeue_task = dequeue_task_fair,
7680 .yield_task = yield_task_fair,
7681 .yield_to_task = yield_to_task_fair,
7683 .check_preempt_curr = check_preempt_wakeup,
7685 .pick_next_task = pick_next_task_fair,
7686 .put_prev_task = put_prev_task_fair,
7689 .select_task_rq = select_task_rq_fair,
7690 #ifdef CONFIG_FAIR_GROUP_SCHED
7691 .migrate_task_rq = migrate_task_rq_fair,
7693 .rq_online = rq_online_fair,
7694 .rq_offline = rq_offline_fair,
7696 .task_waking = task_waking_fair,
7699 .set_curr_task = set_curr_task_fair,
7700 .task_tick = task_tick_fair,
7701 .task_fork = task_fork_fair,
7703 .prio_changed = prio_changed_fair,
7704 .switched_from = switched_from_fair,
7705 .switched_to = switched_to_fair,
7707 .get_rr_interval = get_rr_interval_fair,
7709 #ifdef CONFIG_FAIR_GROUP_SCHED
7710 .task_move_group = task_move_group_fair,
7714 #ifdef CONFIG_SCHED_DEBUG
7715 void print_cfs_stats(struct seq_file *m, int cpu)
7717 struct cfs_rq *cfs_rq;
7720 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7721 print_cfs_rq(m, cpu, cfs_rq);
7726 __init void init_sched_fair_class(void)
7729 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7731 #ifdef CONFIG_NO_HZ_COMMON
7732 nohz.next_balance = jiffies;
7733 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7734 cpu_notifier(sched_ilb_notifier, 0);
7737 #ifdef CONFIG_SCHED_HMP
7738 hmp_cpu_mask_setup();
7744 #ifdef CONFIG_HMP_FREQUENCY_INVARIANT_SCALE
7745 static u32 cpufreq_calc_scale(u32 min, u32 max, u32 curr)
7747 u32 result = curr / max;
7751 /* Called when the CPU Frequency is changed.
7752 * Once for each CPU.
7754 static int cpufreq_callback(struct notifier_block *nb,
7755 unsigned long val, void *data)
7757 struct cpufreq_freqs *freq = data;
7758 int cpu = freq->cpu;
7759 struct cpufreq_extents *extents;
7761 if (freq->flags & CPUFREQ_CONST_LOOPS)
7764 if (val != CPUFREQ_POSTCHANGE)
7767 /* if dynamic load scale is disabled, set the load scale to 1.0 */
7768 if (!hmp_data.freqinvar_load_scale_enabled) {
7769 freq_scale[cpu].curr_scale = 1024;
7773 extents = &freq_scale[cpu];
7774 if (extents->flags & SCHED_LOAD_FREQINVAR_SINGLEFREQ) {
7775 /* If our governor was recognised as a single-freq governor,
7778 extents->curr_scale = 1024;
7780 extents->curr_scale = cpufreq_calc_scale(extents->min,
7781 extents->max, freq->new);
7787 /* Called when the CPUFreq governor is changed.
7788 * Only called for the CPUs which are actually changed by the
7791 static int cpufreq_policy_callback(struct notifier_block *nb,
7792 unsigned long event, void *data)
7794 struct cpufreq_policy *policy = data;
7795 struct cpufreq_extents *extents;
7796 int cpu, singleFreq = 0;
7797 static const char performance_governor[] = "performance";
7798 static const char powersave_governor[] = "powersave";
7800 if (event == CPUFREQ_START)
7803 if (event != CPUFREQ_INCOMPATIBLE)
7806 /* CPUFreq governors do not accurately report the range of
7807 * CPU Frequencies they will choose from.
7808 * We recognise performance and powersave governors as
7809 * single-frequency only.
7811 if (!strncmp(policy->governor->name, performance_governor,
7812 strlen(performance_governor)) ||
7813 !strncmp(policy->governor->name, powersave_governor,
7814 strlen(powersave_governor)))
7817 /* Make sure that all CPUs impacted by this policy are
7818 * updated since we will only get a notification when the
7819 * user explicitly changes the policy on a CPU.
7821 for_each_cpu(cpu, policy->cpus) {
7822 extents = &freq_scale[cpu];
7823 extents->max = policy->max >> SCHED_FREQSCALE_SHIFT;
7824 extents->min = policy->min >> SCHED_FREQSCALE_SHIFT;
7825 if (!hmp_data.freqinvar_load_scale_enabled) {
7826 extents->curr_scale = 1024;
7827 } else if (singleFreq) {
7828 extents->flags |= SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7829 extents->curr_scale = 1024;
7831 extents->flags &= ~SCHED_LOAD_FREQINVAR_SINGLEFREQ;
7832 extents->curr_scale = cpufreq_calc_scale(extents->min,
7833 extents->max, policy->cur);
7840 static struct notifier_block cpufreq_notifier = {
7841 .notifier_call = cpufreq_callback,
7843 static struct notifier_block cpufreq_policy_notifier = {
7844 .notifier_call = cpufreq_policy_callback,
7847 static int __init register_sched_cpufreq_notifier(void)
7851 /* init safe defaults since there are no policies at registration */
7852 for (ret = 0; ret < CONFIG_NR_CPUS; ret++) {
7854 freq_scale[ret].max = 1024;
7855 freq_scale[ret].min = 1024;
7856 freq_scale[ret].curr_scale = 1024;
7859 pr_info("sched: registering cpufreq notifiers for scale-invariant loads\n");
7860 ret = cpufreq_register_notifier(&cpufreq_policy_notifier,
7861 CPUFREQ_POLICY_NOTIFIER);
7864 ret = cpufreq_register_notifier(&cpufreq_notifier,
7865 CPUFREQ_TRANSITION_NOTIFIER);
7870 core_initcall(register_sched_cpufreq_notifier);
7871 #endif /* CONFIG_HMP_FREQUENCY_INVARIANT_SCALE */