sched: Use clamp() and clamp_val() to make sys_nice() more readable
[linux-block.git] / kernel / sched / fair.c
CommitLineData
bf0f6f24
IM
1/*
2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3 *
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5 *
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
8 *
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11 *
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15 *
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
21805085
PZ
18 *
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
bf0f6f24
IM
21 */
22
9745512c 23#include <linux/latencytop.h>
1983a922 24#include <linux/sched.h>
3436ae12 25#include <linux/cpumask.h>
029632fb
PZ
26#include <linux/slab.h>
27#include <linux/profile.h>
28#include <linux/interrupt.h>
cbee9f88 29#include <linux/mempolicy.h>
e14808b4 30#include <linux/migrate.h>
cbee9f88 31#include <linux/task_work.h>
029632fb
PZ
32
33#include <trace/events/sched.h>
34
35#include "sched.h"
9745512c 36
bf0f6f24 37/*
21805085 38 * Targeted preemption latency for CPU-bound tasks:
864616ee 39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
bf0f6f24 40 *
21805085 41 * NOTE: this latency value is not the same as the concept of
d274a4ce
IM
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
bf0f6f24 45 *
d274a4ce
IM
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
bf0f6f24 48 */
21406928
MG
49unsigned int sysctl_sched_latency = 6000000ULL;
50unsigned int normalized_sysctl_sched_latency = 6000000ULL;
2bd8e6d4 51
1983a922
CE
52/*
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
55 *
56 * Options are:
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
60 */
61enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
63
2bd8e6d4 64/*
b2be5e96 65 * Minimal preemption granularity for CPU-bound tasks:
864616ee 66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
2bd8e6d4 67 */
0bf377bb
IM
68unsigned int sysctl_sched_min_granularity = 750000ULL;
69unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
21805085
PZ
70
71/*
b2be5e96
PZ
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
73 */
0bf377bb 74static unsigned int sched_nr_latency = 8;
b2be5e96
PZ
75
76/*
2bba22c5 77 * After fork, child runs first. If set to 0 (default) then
b2be5e96 78 * parent will (try to) run first.
21805085 79 */
2bba22c5 80unsigned int sysctl_sched_child_runs_first __read_mostly;
bf0f6f24 81
bf0f6f24
IM
82/*
83 * SCHED_OTHER wake-up granularity.
172e082a 84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
bf0f6f24
IM
85 *
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
89 */
172e082a 90unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
0bcdcf28 91unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
bf0f6f24 92
da84d961
IM
93const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
94
a7a4f8a7
PT
95/*
96 * The exponential sliding window over which load is averaged for shares
97 * distribution.
98 * (default: 10msec)
99 */
100unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
101
ec12cb7f
PT
102#ifdef CONFIG_CFS_BANDWIDTH
103/*
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
106 *
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
110 *
111 * default: 5 msec, units: microseconds
112 */
113unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114#endif
115
8527632d
PG
116static inline void update_load_add(struct load_weight *lw, unsigned long inc)
117{
118 lw->weight += inc;
119 lw->inv_weight = 0;
120}
121
122static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
123{
124 lw->weight -= dec;
125 lw->inv_weight = 0;
126}
127
128static inline void update_load_set(struct load_weight *lw, unsigned long w)
129{
130 lw->weight = w;
131 lw->inv_weight = 0;
132}
133
029632fb
PZ
134/*
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
139 * number of CPUs.
140 *
141 * This idea comes from the SD scheduler of Con Kolivas:
142 */
143static int get_update_sysctl_factor(void)
144{
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
146 unsigned int factor;
147
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
150 factor = 1;
151 break;
152 case SCHED_TUNABLESCALING_LINEAR:
153 factor = cpus;
154 break;
155 case SCHED_TUNABLESCALING_LOG:
156 default:
157 factor = 1 + ilog2(cpus);
158 break;
159 }
160
161 return factor;
162}
163
164static void update_sysctl(void)
165{
166 unsigned int factor = get_update_sysctl_factor();
167
168#define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
173#undef SET_SYSCTL
174}
175
176void sched_init_granularity(void)
177{
178 update_sysctl();
179}
180
9dbdb155 181#define WMULT_CONST (~0U)
029632fb
PZ
182#define WMULT_SHIFT 32
183
9dbdb155
PZ
184static void __update_inv_weight(struct load_weight *lw)
185{
186 unsigned long w;
187
188 if (likely(lw->inv_weight))
189 return;
190
191 w = scale_load_down(lw->weight);
192
193 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
194 lw->inv_weight = 1;
195 else if (unlikely(!w))
196 lw->inv_weight = WMULT_CONST;
197 else
198 lw->inv_weight = WMULT_CONST / w;
199}
029632fb
PZ
200
201/*
9dbdb155
PZ
202 * delta_exec * weight / lw.weight
203 * OR
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
205 *
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
209 *
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
029632fb 212 */
9dbdb155 213static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
029632fb 214{
9dbdb155
PZ
215 u64 fact = scale_load_down(weight);
216 int shift = WMULT_SHIFT;
029632fb 217
9dbdb155 218 __update_inv_weight(lw);
029632fb 219
9dbdb155
PZ
220 if (unlikely(fact >> 32)) {
221 while (fact >> 32) {
222 fact >>= 1;
223 shift--;
224 }
029632fb
PZ
225 }
226
9dbdb155
PZ
227 /* hint to use a 32x32->64 mul */
228 fact = (u64)(u32)fact * lw->inv_weight;
029632fb 229
9dbdb155
PZ
230 while (fact >> 32) {
231 fact >>= 1;
232 shift--;
233 }
029632fb 234
9dbdb155 235 return mul_u64_u32_shr(delta_exec, fact, shift);
029632fb
PZ
236}
237
238
239const struct sched_class fair_sched_class;
a4c2f00f 240
bf0f6f24
IM
241/**************************************************************
242 * CFS operations on generic schedulable entities:
243 */
244
62160e3f 245#ifdef CONFIG_FAIR_GROUP_SCHED
bf0f6f24 246
62160e3f 247/* cpu runqueue to which this cfs_rq is attached */
bf0f6f24
IM
248static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
249{
62160e3f 250 return cfs_rq->rq;
bf0f6f24
IM
251}
252
62160e3f
IM
253/* An entity is a task if it doesn't "own" a runqueue */
254#define entity_is_task(se) (!se->my_q)
bf0f6f24 255
8f48894f
PZ
256static inline struct task_struct *task_of(struct sched_entity *se)
257{
258#ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se));
260#endif
261 return container_of(se, struct task_struct, se);
262}
263
b758149c
PZ
264/* Walk up scheduling entities hierarchy */
265#define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
267
268static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
269{
270 return p->se.cfs_rq;
271}
272
273/* runqueue on which this entity is (to be) queued */
274static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
275{
276 return se->cfs_rq;
277}
278
279/* runqueue "owned" by this group */
280static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281{
282 return grp->my_q;
283}
284
aff3e498
PT
285static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 int force_update);
9ee474f5 287
3d4b47b4
PZ
288static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289{
290 if (!cfs_rq->on_list) {
67e86250
PT
291 /*
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
296 */
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 &rq_of(cfs_rq)->leaf_cfs_rq_list);
301 } else {
302 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
3d4b47b4 303 &rq_of(cfs_rq)->leaf_cfs_rq_list);
67e86250 304 }
3d4b47b4
PZ
305
306 cfs_rq->on_list = 1;
9ee474f5 307 /* We should have no load, but we need to update last_decay. */
aff3e498 308 update_cfs_rq_blocked_load(cfs_rq, 0);
3d4b47b4
PZ
309 }
310}
311
312static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
313{
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 cfs_rq->on_list = 0;
317 }
318}
319
b758149c
PZ
320/* Iterate thr' all leaf cfs_rq's on a runqueue */
321#define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
323
324/* Do the two (enqueued) entities belong to the same group ? */
fed14d45 325static inline struct cfs_rq *
b758149c
PZ
326is_same_group(struct sched_entity *se, struct sched_entity *pse)
327{
328 if (se->cfs_rq == pse->cfs_rq)
fed14d45 329 return se->cfs_rq;
b758149c 330
fed14d45 331 return NULL;
b758149c
PZ
332}
333
334static inline struct sched_entity *parent_entity(struct sched_entity *se)
335{
336 return se->parent;
337}
338
464b7527
PZ
339static void
340find_matching_se(struct sched_entity **se, struct sched_entity **pse)
341{
342 int se_depth, pse_depth;
343
344 /*
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
348 * parent.
349 */
350
351 /* First walk up until both entities are at same depth */
fed14d45
PZ
352 se_depth = (*se)->depth;
353 pse_depth = (*pse)->depth;
464b7527
PZ
354
355 while (se_depth > pse_depth) {
356 se_depth--;
357 *se = parent_entity(*se);
358 }
359
360 while (pse_depth > se_depth) {
361 pse_depth--;
362 *pse = parent_entity(*pse);
363 }
364
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
368 }
369}
370
8f48894f
PZ
371#else /* !CONFIG_FAIR_GROUP_SCHED */
372
373static inline struct task_struct *task_of(struct sched_entity *se)
374{
375 return container_of(se, struct task_struct, se);
376}
bf0f6f24 377
62160e3f
IM
378static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
379{
380 return container_of(cfs_rq, struct rq, cfs);
bf0f6f24
IM
381}
382
383#define entity_is_task(se) 1
384
b758149c
PZ
385#define for_each_sched_entity(se) \
386 for (; se; se = NULL)
bf0f6f24 387
b758149c 388static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
bf0f6f24 389{
b758149c 390 return &task_rq(p)->cfs;
bf0f6f24
IM
391}
392
b758149c
PZ
393static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
394{
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
397
398 return &rq->cfs;
399}
400
401/* runqueue "owned" by this group */
402static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403{
404 return NULL;
405}
406
3d4b47b4
PZ
407static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408{
409}
410
411static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412{
413}
414
b758149c
PZ
415#define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
417
b758149c
PZ
418static inline struct sched_entity *parent_entity(struct sched_entity *se)
419{
420 return NULL;
421}
422
464b7527
PZ
423static inline void
424find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425{
426}
427
b758149c
PZ
428#endif /* CONFIG_FAIR_GROUP_SCHED */
429
6c16a6dc 430static __always_inline
9dbdb155 431void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
bf0f6f24
IM
432
433/**************************************************************
434 * Scheduling class tree data structure manipulation methods:
435 */
436
1bf08230 437static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
02e0431a 438{
1bf08230 439 s64 delta = (s64)(vruntime - max_vruntime);
368059a9 440 if (delta > 0)
1bf08230 441 max_vruntime = vruntime;
02e0431a 442
1bf08230 443 return max_vruntime;
02e0431a
PZ
444}
445
0702e3eb 446static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
b0ffd246
PZ
447{
448 s64 delta = (s64)(vruntime - min_vruntime);
449 if (delta < 0)
450 min_vruntime = vruntime;
451
452 return min_vruntime;
453}
454
54fdc581
FC
455static inline int entity_before(struct sched_entity *a,
456 struct sched_entity *b)
457{
458 return (s64)(a->vruntime - b->vruntime) < 0;
459}
460
1af5f730
PZ
461static void update_min_vruntime(struct cfs_rq *cfs_rq)
462{
463 u64 vruntime = cfs_rq->min_vruntime;
464
465 if (cfs_rq->curr)
466 vruntime = cfs_rq->curr->vruntime;
467
468 if (cfs_rq->rb_leftmost) {
469 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 struct sched_entity,
471 run_node);
472
e17036da 473 if (!cfs_rq->curr)
1af5f730
PZ
474 vruntime = se->vruntime;
475 else
476 vruntime = min_vruntime(vruntime, se->vruntime);
477 }
478
1bf08230 479 /* ensure we never gain time by being placed backwards. */
1af5f730 480 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
3fe1698b
PZ
481#ifndef CONFIG_64BIT
482 smp_wmb();
483 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484#endif
1af5f730
PZ
485}
486
bf0f6f24
IM
487/*
488 * Enqueue an entity into the rb-tree:
489 */
0702e3eb 490static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24
IM
491{
492 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493 struct rb_node *parent = NULL;
494 struct sched_entity *entry;
bf0f6f24
IM
495 int leftmost = 1;
496
497 /*
498 * Find the right place in the rbtree:
499 */
500 while (*link) {
501 parent = *link;
502 entry = rb_entry(parent, struct sched_entity, run_node);
503 /*
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
506 */
2bd2d6f2 507 if (entity_before(se, entry)) {
bf0f6f24
IM
508 link = &parent->rb_left;
509 } else {
510 link = &parent->rb_right;
511 leftmost = 0;
512 }
513 }
514
515 /*
516 * Maintain a cache of leftmost tree entries (it is frequently
517 * used):
518 */
1af5f730 519 if (leftmost)
57cb499d 520 cfs_rq->rb_leftmost = &se->run_node;
bf0f6f24
IM
521
522 rb_link_node(&se->run_node, parent, link);
523 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
bf0f6f24
IM
524}
525
0702e3eb 526static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 527{
3fe69747
PZ
528 if (cfs_rq->rb_leftmost == &se->run_node) {
529 struct rb_node *next_node;
3fe69747
PZ
530
531 next_node = rb_next(&se->run_node);
532 cfs_rq->rb_leftmost = next_node;
3fe69747 533 }
e9acbff6 534
bf0f6f24 535 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
bf0f6f24
IM
536}
537
029632fb 538struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
bf0f6f24 539{
f4b6755f
PZ
540 struct rb_node *left = cfs_rq->rb_leftmost;
541
542 if (!left)
543 return NULL;
544
545 return rb_entry(left, struct sched_entity, run_node);
bf0f6f24
IM
546}
547
ac53db59
RR
548static struct sched_entity *__pick_next_entity(struct sched_entity *se)
549{
550 struct rb_node *next = rb_next(&se->run_node);
551
552 if (!next)
553 return NULL;
554
555 return rb_entry(next, struct sched_entity, run_node);
556}
557
558#ifdef CONFIG_SCHED_DEBUG
029632fb 559struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
aeb73b04 560{
7eee3e67 561 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
aeb73b04 562
70eee74b
BS
563 if (!last)
564 return NULL;
7eee3e67
IM
565
566 return rb_entry(last, struct sched_entity, run_node);
aeb73b04
PZ
567}
568
bf0f6f24
IM
569/**************************************************************
570 * Scheduling class statistics methods:
571 */
572
acb4a848 573int sched_proc_update_handler(struct ctl_table *table, int write,
8d65af78 574 void __user *buffer, size_t *lenp,
b2be5e96
PZ
575 loff_t *ppos)
576{
8d65af78 577 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
acb4a848 578 int factor = get_update_sysctl_factor();
b2be5e96
PZ
579
580 if (ret || !write)
581 return ret;
582
583 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584 sysctl_sched_min_granularity);
585
acb4a848
CE
586#define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity);
589 WRT_SYSCTL(sched_latency);
590 WRT_SYSCTL(sched_wakeup_granularity);
acb4a848
CE
591#undef WRT_SYSCTL
592
b2be5e96
PZ
593 return 0;
594}
595#endif
647e7cac 596
a7be37ac 597/*
f9c0b095 598 * delta /= w
a7be37ac 599 */
9dbdb155 600static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
a7be37ac 601{
f9c0b095 602 if (unlikely(se->load.weight != NICE_0_LOAD))
9dbdb155 603 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
a7be37ac
PZ
604
605 return delta;
606}
607
647e7cac
IM
608/*
609 * The idea is to set a period in which each task runs once.
610 *
532b1858 611 * When there are too many tasks (sched_nr_latency) we have to stretch
647e7cac
IM
612 * this period because otherwise the slices get too small.
613 *
614 * p = (nr <= nl) ? l : l*nr/nl
615 */
4d78e7b6
PZ
616static u64 __sched_period(unsigned long nr_running)
617{
618 u64 period = sysctl_sched_latency;
b2be5e96 619 unsigned long nr_latency = sched_nr_latency;
4d78e7b6
PZ
620
621 if (unlikely(nr_running > nr_latency)) {
4bf0b771 622 period = sysctl_sched_min_granularity;
4d78e7b6 623 period *= nr_running;
4d78e7b6
PZ
624 }
625
626 return period;
627}
628
647e7cac
IM
629/*
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
632 *
f9c0b095 633 * s = p*P[w/rw]
647e7cac 634 */
6d0f0ebd 635static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
21805085 636{
0a582440 637 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
f9c0b095 638
0a582440 639 for_each_sched_entity(se) {
6272d68c 640 struct load_weight *load;
3104bf03 641 struct load_weight lw;
6272d68c
LM
642
643 cfs_rq = cfs_rq_of(se);
644 load = &cfs_rq->load;
f9c0b095 645
0a582440 646 if (unlikely(!se->on_rq)) {
3104bf03 647 lw = cfs_rq->load;
0a582440
MG
648
649 update_load_add(&lw, se->load.weight);
650 load = &lw;
651 }
9dbdb155 652 slice = __calc_delta(slice, se->load.weight, load);
0a582440
MG
653 }
654 return slice;
bf0f6f24
IM
655}
656
647e7cac 657/*
660cc00f 658 * We calculate the vruntime slice of a to-be-inserted task.
647e7cac 659 *
f9c0b095 660 * vs = s/w
647e7cac 661 */
f9c0b095 662static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
67e9fb2a 663{
f9c0b095 664 return calc_delta_fair(sched_slice(cfs_rq, se), se);
a7be37ac
PZ
665}
666
a75cdaa9 667#ifdef CONFIG_SMP
fb13c7ee
MG
668static unsigned long task_h_load(struct task_struct *p);
669
a75cdaa9
AS
670static inline void __update_task_entity_contrib(struct sched_entity *se);
671
672/* Give new task start runnable values to heavy its load in infant time */
673void init_task_runnable_average(struct task_struct *p)
674{
675 u32 slice;
676
677 p->se.avg.decay_count = 0;
678 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679 p->se.avg.runnable_avg_sum = slice;
680 p->se.avg.runnable_avg_period = slice;
681 __update_task_entity_contrib(&p->se);
682}
683#else
684void init_task_runnable_average(struct task_struct *p)
685{
686}
687#endif
688
bf0f6f24 689/*
9dbdb155 690 * Update the current task's runtime statistics.
bf0f6f24 691 */
b7cc0896 692static void update_curr(struct cfs_rq *cfs_rq)
bf0f6f24 693{
429d43bc 694 struct sched_entity *curr = cfs_rq->curr;
78becc27 695 u64 now = rq_clock_task(rq_of(cfs_rq));
9dbdb155 696 u64 delta_exec;
bf0f6f24
IM
697
698 if (unlikely(!curr))
699 return;
700
9dbdb155
PZ
701 delta_exec = now - curr->exec_start;
702 if (unlikely((s64)delta_exec <= 0))
34f28ecd 703 return;
bf0f6f24 704
8ebc91d9 705 curr->exec_start = now;
d842de87 706
9dbdb155
PZ
707 schedstat_set(curr->statistics.exec_max,
708 max(delta_exec, curr->statistics.exec_max));
709
710 curr->sum_exec_runtime += delta_exec;
711 schedstat_add(cfs_rq, exec_clock, delta_exec);
712
713 curr->vruntime += calc_delta_fair(delta_exec, curr);
714 update_min_vruntime(cfs_rq);
715
d842de87
SV
716 if (entity_is_task(curr)) {
717 struct task_struct *curtask = task_of(curr);
718
f977bb49 719 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
d842de87 720 cpuacct_charge(curtask, delta_exec);
f06febc9 721 account_group_exec_runtime(curtask, delta_exec);
d842de87 722 }
ec12cb7f
PT
723
724 account_cfs_rq_runtime(cfs_rq, delta_exec);
bf0f6f24
IM
725}
726
727static inline void
5870db5b 728update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 729{
78becc27 730 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
bf0f6f24
IM
731}
732
bf0f6f24
IM
733/*
734 * Task is being enqueued - update stats:
735 */
d2417e5a 736static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 737{
bf0f6f24
IM
738 /*
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
741 */
429d43bc 742 if (se != cfs_rq->curr)
5870db5b 743 update_stats_wait_start(cfs_rq, se);
bf0f6f24
IM
744}
745
bf0f6f24 746static void
9ef0a961 747update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 748{
41acab88 749 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
78becc27 750 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
41acab88
LDM
751 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
78becc27 753 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
768d0c27
PZ
754#ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se)) {
756 trace_sched_stat_wait(task_of(se),
78becc27 757 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
768d0c27
PZ
758 }
759#endif
41acab88 760 schedstat_set(se->statistics.wait_start, 0);
bf0f6f24
IM
761}
762
763static inline void
19b6a2e3 764update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 765{
bf0f6f24
IM
766 /*
767 * Mark the end of the wait period if dequeueing a
768 * waiting task:
769 */
429d43bc 770 if (se != cfs_rq->curr)
9ef0a961 771 update_stats_wait_end(cfs_rq, se);
bf0f6f24
IM
772}
773
774/*
775 * We are picking a new current task - update its stats:
776 */
777static inline void
79303e9e 778update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24
IM
779{
780 /*
781 * We are starting a new run period:
782 */
78becc27 783 se->exec_start = rq_clock_task(rq_of(cfs_rq));
bf0f6f24
IM
784}
785
bf0f6f24
IM
786/**************************************************
787 * Scheduling class queueing methods:
788 */
789
cbee9f88
PZ
790#ifdef CONFIG_NUMA_BALANCING
791/*
598f0ec0
MG
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
cbee9f88 795 */
598f0ec0
MG
796unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797unsigned int sysctl_numa_balancing_scan_period_max = 60000;
6e5fb223
PZ
798
799/* Portion of address space to scan in MB */
800unsigned int sysctl_numa_balancing_scan_size = 256;
cbee9f88 801
4b96a29b
PZ
802/* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803unsigned int sysctl_numa_balancing_scan_delay = 1000;
804
598f0ec0
MG
805static unsigned int task_nr_scan_windows(struct task_struct *p)
806{
807 unsigned long rss = 0;
808 unsigned long nr_scan_pages;
809
810 /*
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
813 * on resident pages
814 */
815 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816 rss = get_mm_rss(p->mm);
817 if (!rss)
818 rss = nr_scan_pages;
819
820 rss = round_up(rss, nr_scan_pages);
821 return rss / nr_scan_pages;
822}
823
824/* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825#define MAX_SCAN_WINDOW 2560
826
827static unsigned int task_scan_min(struct task_struct *p)
828{
829 unsigned int scan, floor;
830 unsigned int windows = 1;
831
832 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834 floor = 1000 / windows;
835
836 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837 return max_t(unsigned int, floor, scan);
838}
839
840static unsigned int task_scan_max(struct task_struct *p)
841{
842 unsigned int smin = task_scan_min(p);
843 unsigned int smax;
844
845 /* Watch for min being lower than max due to floor calculations */
846 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847 return max(smin, smax);
848}
849
0ec8aa00
PZ
850static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
851{
852 rq->nr_numa_running += (p->numa_preferred_nid != -1);
853 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
854}
855
856static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
857{
858 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
860}
861
8c8a743c
PZ
862struct numa_group {
863 atomic_t refcount;
864
865 spinlock_t lock; /* nr_tasks, tasks */
866 int nr_tasks;
e29cf08b 867 pid_t gid;
8c8a743c
PZ
868 struct list_head task_list;
869
870 struct rcu_head rcu;
20e07dea 871 nodemask_t active_nodes;
989348b5 872 unsigned long total_faults;
7e2703e6
RR
873 /*
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
877 */
50ec8a40 878 unsigned long *faults_cpu;
989348b5 879 unsigned long faults[0];
8c8a743c
PZ
880};
881
be1e4e76
RR
882/* Shared or private faults. */
883#define NR_NUMA_HINT_FAULT_TYPES 2
884
885/* Memory and CPU locality */
886#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
887
888/* Averaged statistics, and temporary buffers. */
889#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
890
e29cf08b
MG
891pid_t task_numa_group_id(struct task_struct *p)
892{
893 return p->numa_group ? p->numa_group->gid : 0;
894}
895
ac8e895b
MG
896static inline int task_faults_idx(int nid, int priv)
897{
be1e4e76 898 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
ac8e895b
MG
899}
900
901static inline unsigned long task_faults(struct task_struct *p, int nid)
902{
ff1df896 903 if (!p->numa_faults_memory)
ac8e895b
MG
904 return 0;
905
ff1df896
RR
906 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907 p->numa_faults_memory[task_faults_idx(nid, 1)];
ac8e895b
MG
908}
909
83e1d2cd
MG
910static inline unsigned long group_faults(struct task_struct *p, int nid)
911{
912 if (!p->numa_group)
913 return 0;
914
82897b4f
WL
915 return p->numa_group->faults[task_faults_idx(nid, 0)] +
916 p->numa_group->faults[task_faults_idx(nid, 1)];
83e1d2cd
MG
917}
918
20e07dea
RR
919static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
920{
921 return group->faults_cpu[task_faults_idx(nid, 0)] +
922 group->faults_cpu[task_faults_idx(nid, 1)];
923}
924
83e1d2cd
MG
925/*
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
930 */
931static inline unsigned long task_weight(struct task_struct *p, int nid)
932{
933 unsigned long total_faults;
934
ff1df896 935 if (!p->numa_faults_memory)
83e1d2cd
MG
936 return 0;
937
938 total_faults = p->total_numa_faults;
939
940 if (!total_faults)
941 return 0;
942
943 return 1000 * task_faults(p, nid) / total_faults;
944}
945
946static inline unsigned long group_weight(struct task_struct *p, int nid)
947{
989348b5 948 if (!p->numa_group || !p->numa_group->total_faults)
83e1d2cd
MG
949 return 0;
950
989348b5 951 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
83e1d2cd
MG
952}
953
10f39042
RR
954bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955 int src_nid, int dst_cpu)
956{
957 struct numa_group *ng = p->numa_group;
958 int dst_nid = cpu_to_node(dst_cpu);
959 int last_cpupid, this_cpupid;
960
961 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
962
963 /*
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
967 *
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
971 *
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
976 *
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
979 */
980 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981 if (!cpupid_pid_unset(last_cpupid) &&
982 cpupid_to_nid(last_cpupid) != dst_nid)
983 return false;
984
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p, last_cpupid))
987 return true;
988
989 /* A shared fault, but p->numa_group has not been set up yet. */
990 if (!ng)
991 return true;
992
993 /*
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
996 */
997 if (!node_isset(dst_nid, ng->active_nodes))
998 return false;
999
1000 /*
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1003 */
1004 if (!node_isset(src_nid, ng->active_nodes))
1005 return true;
1006
1007 /*
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1013 */
1014 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1015}
1016
e6628d5b 1017static unsigned long weighted_cpuload(const int cpu);
58d081b5
MG
1018static unsigned long source_load(int cpu, int type);
1019static unsigned long target_load(int cpu, int type);
1020static unsigned long power_of(int cpu);
1021static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1022
fb13c7ee 1023/* Cached statistics for all CPUs within a node */
58d081b5 1024struct numa_stats {
fb13c7ee 1025 unsigned long nr_running;
58d081b5 1026 unsigned long load;
fb13c7ee
MG
1027
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long power;
1030
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long capacity;
1033 int has_capacity;
58d081b5 1034};
e6628d5b 1035
fb13c7ee
MG
1036/*
1037 * XXX borrowed from update_sg_lb_stats
1038 */
1039static void update_numa_stats(struct numa_stats *ns, int nid)
1040{
5eca82a9 1041 int cpu, cpus = 0;
fb13c7ee
MG
1042
1043 memset(ns, 0, sizeof(*ns));
1044 for_each_cpu(cpu, cpumask_of_node(nid)) {
1045 struct rq *rq = cpu_rq(cpu);
1046
1047 ns->nr_running += rq->nr_running;
1048 ns->load += weighted_cpuload(cpu);
1049 ns->power += power_of(cpu);
5eca82a9
PZ
1050
1051 cpus++;
fb13c7ee
MG
1052 }
1053
5eca82a9
PZ
1054 /*
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1058 *
1059 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1060 * and bail there.
1061 */
1062 if (!cpus)
1063 return;
1064
fb13c7ee
MG
1065 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1066 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1067 ns->has_capacity = (ns->nr_running < ns->capacity);
1068}
1069
58d081b5
MG
1070struct task_numa_env {
1071 struct task_struct *p;
e6628d5b 1072
58d081b5
MG
1073 int src_cpu, src_nid;
1074 int dst_cpu, dst_nid;
e6628d5b 1075
58d081b5 1076 struct numa_stats src_stats, dst_stats;
e6628d5b 1077
40ea2b42 1078 int imbalance_pct;
fb13c7ee
MG
1079
1080 struct task_struct *best_task;
1081 long best_imp;
58d081b5
MG
1082 int best_cpu;
1083};
1084
fb13c7ee
MG
1085static void task_numa_assign(struct task_numa_env *env,
1086 struct task_struct *p, long imp)
1087{
1088 if (env->best_task)
1089 put_task_struct(env->best_task);
1090 if (p)
1091 get_task_struct(p);
1092
1093 env->best_task = p;
1094 env->best_imp = imp;
1095 env->best_cpu = env->dst_cpu;
1096}
1097
1098/*
1099 * This checks if the overall compute and NUMA accesses of the system would
1100 * be improved if the source tasks was migrated to the target dst_cpu taking
1101 * into account that it might be best if task running on the dst_cpu should
1102 * be exchanged with the source task
1103 */
887c290e
RR
1104static void task_numa_compare(struct task_numa_env *env,
1105 long taskimp, long groupimp)
fb13c7ee
MG
1106{
1107 struct rq *src_rq = cpu_rq(env->src_cpu);
1108 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1109 struct task_struct *cur;
1110 long dst_load, src_load;
1111 long load;
887c290e 1112 long imp = (groupimp > 0) ? groupimp : taskimp;
fb13c7ee
MG
1113
1114 rcu_read_lock();
1115 cur = ACCESS_ONCE(dst_rq->curr);
1116 if (cur->pid == 0) /* idle */
1117 cur = NULL;
1118
1119 /*
1120 * "imp" is the fault differential for the source task between the
1121 * source and destination node. Calculate the total differential for
1122 * the source task and potential destination task. The more negative
1123 * the value is, the more rmeote accesses that would be expected to
1124 * be incurred if the tasks were swapped.
1125 */
1126 if (cur) {
1127 /* Skip this swap candidate if cannot move to the source cpu */
1128 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1129 goto unlock;
1130
887c290e
RR
1131 /*
1132 * If dst and source tasks are in the same NUMA group, or not
ca28aa53 1133 * in any group then look only at task weights.
887c290e 1134 */
ca28aa53 1135 if (cur->numa_group == env->p->numa_group) {
887c290e
RR
1136 imp = taskimp + task_weight(cur, env->src_nid) -
1137 task_weight(cur, env->dst_nid);
ca28aa53
RR
1138 /*
1139 * Add some hysteresis to prevent swapping the
1140 * tasks within a group over tiny differences.
1141 */
1142 if (cur->numa_group)
1143 imp -= imp/16;
887c290e 1144 } else {
ca28aa53
RR
1145 /*
1146 * Compare the group weights. If a task is all by
1147 * itself (not part of a group), use the task weight
1148 * instead.
1149 */
1150 if (env->p->numa_group)
1151 imp = groupimp;
1152 else
1153 imp = taskimp;
1154
1155 if (cur->numa_group)
1156 imp += group_weight(cur, env->src_nid) -
1157 group_weight(cur, env->dst_nid);
1158 else
1159 imp += task_weight(cur, env->src_nid) -
1160 task_weight(cur, env->dst_nid);
887c290e 1161 }
fb13c7ee
MG
1162 }
1163
1164 if (imp < env->best_imp)
1165 goto unlock;
1166
1167 if (!cur) {
1168 /* Is there capacity at our destination? */
1169 if (env->src_stats.has_capacity &&
1170 !env->dst_stats.has_capacity)
1171 goto unlock;
1172
1173 goto balance;
1174 }
1175
1176 /* Balance doesn't matter much if we're running a task per cpu */
1177 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1178 goto assign;
1179
1180 /*
1181 * In the overloaded case, try and keep the load balanced.
1182 */
1183balance:
1184 dst_load = env->dst_stats.load;
1185 src_load = env->src_stats.load;
1186
1187 /* XXX missing power terms */
1188 load = task_h_load(env->p);
1189 dst_load += load;
1190 src_load -= load;
1191
1192 if (cur) {
1193 load = task_h_load(cur);
1194 dst_load -= load;
1195 src_load += load;
1196 }
1197
1198 /* make src_load the smaller */
1199 if (dst_load < src_load)
1200 swap(dst_load, src_load);
1201
1202 if (src_load * env->imbalance_pct < dst_load * 100)
1203 goto unlock;
1204
1205assign:
1206 task_numa_assign(env, cur, imp);
1207unlock:
1208 rcu_read_unlock();
1209}
1210
887c290e
RR
1211static void task_numa_find_cpu(struct task_numa_env *env,
1212 long taskimp, long groupimp)
2c8a50aa
MG
1213{
1214 int cpu;
1215
1216 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1217 /* Skip this CPU if the source task cannot migrate */
1218 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1219 continue;
1220
1221 env->dst_cpu = cpu;
887c290e 1222 task_numa_compare(env, taskimp, groupimp);
2c8a50aa
MG
1223 }
1224}
1225
58d081b5
MG
1226static int task_numa_migrate(struct task_struct *p)
1227{
58d081b5
MG
1228 struct task_numa_env env = {
1229 .p = p,
fb13c7ee 1230
58d081b5 1231 .src_cpu = task_cpu(p),
b32e86b4 1232 .src_nid = task_node(p),
fb13c7ee
MG
1233
1234 .imbalance_pct = 112,
1235
1236 .best_task = NULL,
1237 .best_imp = 0,
1238 .best_cpu = -1
58d081b5
MG
1239 };
1240 struct sched_domain *sd;
887c290e 1241 unsigned long taskweight, groupweight;
2c8a50aa 1242 int nid, ret;
887c290e 1243 long taskimp, groupimp;
e6628d5b 1244
58d081b5 1245 /*
fb13c7ee
MG
1246 * Pick the lowest SD_NUMA domain, as that would have the smallest
1247 * imbalance and would be the first to start moving tasks about.
1248 *
1249 * And we want to avoid any moving of tasks about, as that would create
1250 * random movement of tasks -- counter the numa conditions we're trying
1251 * to satisfy here.
58d081b5
MG
1252 */
1253 rcu_read_lock();
fb13c7ee 1254 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
46a73e8a
RR
1255 if (sd)
1256 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
e6628d5b
MG
1257 rcu_read_unlock();
1258
46a73e8a
RR
1259 /*
1260 * Cpusets can break the scheduler domain tree into smaller
1261 * balance domains, some of which do not cross NUMA boundaries.
1262 * Tasks that are "trapped" in such domains cannot be migrated
1263 * elsewhere, so there is no point in (re)trying.
1264 */
1265 if (unlikely(!sd)) {
de1b301a 1266 p->numa_preferred_nid = task_node(p);
46a73e8a
RR
1267 return -EINVAL;
1268 }
1269
887c290e
RR
1270 taskweight = task_weight(p, env.src_nid);
1271 groupweight = group_weight(p, env.src_nid);
fb13c7ee 1272 update_numa_stats(&env.src_stats, env.src_nid);
2c8a50aa 1273 env.dst_nid = p->numa_preferred_nid;
887c290e
RR
1274 taskimp = task_weight(p, env.dst_nid) - taskweight;
1275 groupimp = group_weight(p, env.dst_nid) - groupweight;
2c8a50aa 1276 update_numa_stats(&env.dst_stats, env.dst_nid);
58d081b5 1277
e1dda8a7
RR
1278 /* If the preferred nid has capacity, try to use it. */
1279 if (env.dst_stats.has_capacity)
887c290e 1280 task_numa_find_cpu(&env, taskimp, groupimp);
e1dda8a7
RR
1281
1282 /* No space available on the preferred nid. Look elsewhere. */
1283 if (env.best_cpu == -1) {
2c8a50aa
MG
1284 for_each_online_node(nid) {
1285 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1286 continue;
58d081b5 1287
83e1d2cd 1288 /* Only consider nodes where both task and groups benefit */
887c290e
RR
1289 taskimp = task_weight(p, nid) - taskweight;
1290 groupimp = group_weight(p, nid) - groupweight;
1291 if (taskimp < 0 && groupimp < 0)
fb13c7ee
MG
1292 continue;
1293
2c8a50aa
MG
1294 env.dst_nid = nid;
1295 update_numa_stats(&env.dst_stats, env.dst_nid);
887c290e 1296 task_numa_find_cpu(&env, taskimp, groupimp);
58d081b5
MG
1297 }
1298 }
1299
fb13c7ee
MG
1300 /* No better CPU than the current one was found. */
1301 if (env.best_cpu == -1)
1302 return -EAGAIN;
1303
68d1b02a
RR
1304 /*
1305 * If the task is part of a workload that spans multiple NUMA nodes,
1306 * and is migrating into one of the workload's active nodes, remember
1307 * this node as the task's preferred numa node, so the workload can
1308 * settle down.
1309 * A task that migrated to a second choice node will be better off
1310 * trying for a better one later. Do not set the preferred node here.
1311 */
1312 if (p->numa_group && node_isset(env.dst_nid, p->numa_group->active_nodes))
1313 sched_setnuma(p, env.dst_nid);
0ec8aa00 1314
04bb2f94
RR
1315 /*
1316 * Reset the scan period if the task is being rescheduled on an
1317 * alternative node to recheck if the tasks is now properly placed.
1318 */
1319 p->numa_scan_period = task_scan_min(p);
1320
fb13c7ee 1321 if (env.best_task == NULL) {
286549dc
MG
1322 ret = migrate_task_to(p, env.best_cpu);
1323 if (ret != 0)
1324 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
fb13c7ee
MG
1325 return ret;
1326 }
1327
1328 ret = migrate_swap(p, env.best_task);
286549dc
MG
1329 if (ret != 0)
1330 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
fb13c7ee
MG
1331 put_task_struct(env.best_task);
1332 return ret;
e6628d5b
MG
1333}
1334
6b9a7460
MG
1335/* Attempt to migrate a task to a CPU on the preferred node. */
1336static void numa_migrate_preferred(struct task_struct *p)
1337{
5085e2a3
RR
1338 unsigned long interval = HZ;
1339
2739d3ee 1340 /* This task has no NUMA fault statistics yet */
ff1df896 1341 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
6b9a7460
MG
1342 return;
1343
2739d3ee 1344 /* Periodically retry migrating the task to the preferred node */
5085e2a3
RR
1345 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1346 p->numa_migrate_retry = jiffies + interval;
2739d3ee
RR
1347
1348 /* Success if task is already running on preferred CPU */
de1b301a 1349 if (task_node(p) == p->numa_preferred_nid)
6b9a7460
MG
1350 return;
1351
1352 /* Otherwise, try migrate to a CPU on the preferred node */
2739d3ee 1353 task_numa_migrate(p);
6b9a7460
MG
1354}
1355
20e07dea
RR
1356/*
1357 * Find the nodes on which the workload is actively running. We do this by
1358 * tracking the nodes from which NUMA hinting faults are triggered. This can
1359 * be different from the set of nodes where the workload's memory is currently
1360 * located.
1361 *
1362 * The bitmask is used to make smarter decisions on when to do NUMA page
1363 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1364 * are added when they cause over 6/16 of the maximum number of faults, but
1365 * only removed when they drop below 3/16.
1366 */
1367static void update_numa_active_node_mask(struct numa_group *numa_group)
1368{
1369 unsigned long faults, max_faults = 0;
1370 int nid;
1371
1372 for_each_online_node(nid) {
1373 faults = group_faults_cpu(numa_group, nid);
1374 if (faults > max_faults)
1375 max_faults = faults;
1376 }
1377
1378 for_each_online_node(nid) {
1379 faults = group_faults_cpu(numa_group, nid);
1380 if (!node_isset(nid, numa_group->active_nodes)) {
1381 if (faults > max_faults * 6 / 16)
1382 node_set(nid, numa_group->active_nodes);
1383 } else if (faults < max_faults * 3 / 16)
1384 node_clear(nid, numa_group->active_nodes);
1385 }
1386}
1387
04bb2f94
RR
1388/*
1389 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1390 * increments. The more local the fault statistics are, the higher the scan
1391 * period will be for the next scan window. If local/remote ratio is below
1392 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1393 * scan period will decrease
1394 */
1395#define NUMA_PERIOD_SLOTS 10
1396#define NUMA_PERIOD_THRESHOLD 3
1397
1398/*
1399 * Increase the scan period (slow down scanning) if the majority of
1400 * our memory is already on our local node, or if the majority of
1401 * the page accesses are shared with other processes.
1402 * Otherwise, decrease the scan period.
1403 */
1404static void update_task_scan_period(struct task_struct *p,
1405 unsigned long shared, unsigned long private)
1406{
1407 unsigned int period_slot;
1408 int ratio;
1409 int diff;
1410
1411 unsigned long remote = p->numa_faults_locality[0];
1412 unsigned long local = p->numa_faults_locality[1];
1413
1414 /*
1415 * If there were no record hinting faults then either the task is
1416 * completely idle or all activity is areas that are not of interest
1417 * to automatic numa balancing. Scan slower
1418 */
1419 if (local + shared == 0) {
1420 p->numa_scan_period = min(p->numa_scan_period_max,
1421 p->numa_scan_period << 1);
1422
1423 p->mm->numa_next_scan = jiffies +
1424 msecs_to_jiffies(p->numa_scan_period);
1425
1426 return;
1427 }
1428
1429 /*
1430 * Prepare to scale scan period relative to the current period.
1431 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1432 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1433 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1434 */
1435 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1436 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1437 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1438 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1439 if (!slot)
1440 slot = 1;
1441 diff = slot * period_slot;
1442 } else {
1443 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1444
1445 /*
1446 * Scale scan rate increases based on sharing. There is an
1447 * inverse relationship between the degree of sharing and
1448 * the adjustment made to the scanning period. Broadly
1449 * speaking the intent is that there is little point
1450 * scanning faster if shared accesses dominate as it may
1451 * simply bounce migrations uselessly
1452 */
04bb2f94
RR
1453 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1454 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1455 }
1456
1457 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1458 task_scan_min(p), task_scan_max(p));
1459 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1460}
1461
7e2703e6
RR
1462/*
1463 * Get the fraction of time the task has been running since the last
1464 * NUMA placement cycle. The scheduler keeps similar statistics, but
1465 * decays those on a 32ms period, which is orders of magnitude off
1466 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1467 * stats only if the task is so new there are no NUMA statistics yet.
1468 */
1469static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1470{
1471 u64 runtime, delta, now;
1472 /* Use the start of this time slice to avoid calculations. */
1473 now = p->se.exec_start;
1474 runtime = p->se.sum_exec_runtime;
1475
1476 if (p->last_task_numa_placement) {
1477 delta = runtime - p->last_sum_exec_runtime;
1478 *period = now - p->last_task_numa_placement;
1479 } else {
1480 delta = p->se.avg.runnable_avg_sum;
1481 *period = p->se.avg.runnable_avg_period;
1482 }
1483
1484 p->last_sum_exec_runtime = runtime;
1485 p->last_task_numa_placement = now;
1486
1487 return delta;
1488}
1489
cbee9f88
PZ
1490static void task_numa_placement(struct task_struct *p)
1491{
83e1d2cd
MG
1492 int seq, nid, max_nid = -1, max_group_nid = -1;
1493 unsigned long max_faults = 0, max_group_faults = 0;
04bb2f94 1494 unsigned long fault_types[2] = { 0, 0 };
7e2703e6
RR
1495 unsigned long total_faults;
1496 u64 runtime, period;
7dbd13ed 1497 spinlock_t *group_lock = NULL;
cbee9f88 1498
2832bc19 1499 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
cbee9f88
PZ
1500 if (p->numa_scan_seq == seq)
1501 return;
1502 p->numa_scan_seq = seq;
598f0ec0 1503 p->numa_scan_period_max = task_scan_max(p);
cbee9f88 1504
7e2703e6
RR
1505 total_faults = p->numa_faults_locality[0] +
1506 p->numa_faults_locality[1];
1507 runtime = numa_get_avg_runtime(p, &period);
1508
7dbd13ed
MG
1509 /* If the task is part of a group prevent parallel updates to group stats */
1510 if (p->numa_group) {
1511 group_lock = &p->numa_group->lock;
60e69eed 1512 spin_lock_irq(group_lock);
7dbd13ed
MG
1513 }
1514
688b7585
MG
1515 /* Find the node with the highest number of faults */
1516 for_each_online_node(nid) {
83e1d2cd 1517 unsigned long faults = 0, group_faults = 0;
ac8e895b 1518 int priv, i;
745d6147 1519
be1e4e76 1520 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
7e2703e6 1521 long diff, f_diff, f_weight;
8c8a743c 1522
ac8e895b 1523 i = task_faults_idx(nid, priv);
745d6147 1524
ac8e895b 1525 /* Decay existing window, copy faults since last scan */
35664fd4 1526 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
ff1df896
RR
1527 fault_types[priv] += p->numa_faults_buffer_memory[i];
1528 p->numa_faults_buffer_memory[i] = 0;
fb13c7ee 1529
7e2703e6
RR
1530 /*
1531 * Normalize the faults_from, so all tasks in a group
1532 * count according to CPU use, instead of by the raw
1533 * number of faults. Tasks with little runtime have
1534 * little over-all impact on throughput, and thus their
1535 * faults are less important.
1536 */
1537 f_weight = div64_u64(runtime << 16, period + 1);
1538 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1539 (total_faults + 1);
35664fd4 1540 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
50ec8a40
RR
1541 p->numa_faults_buffer_cpu[i] = 0;
1542
35664fd4
RR
1543 p->numa_faults_memory[i] += diff;
1544 p->numa_faults_cpu[i] += f_diff;
ff1df896 1545 faults += p->numa_faults_memory[i];
83e1d2cd 1546 p->total_numa_faults += diff;
8c8a743c
PZ
1547 if (p->numa_group) {
1548 /* safe because we can only change our own group */
989348b5 1549 p->numa_group->faults[i] += diff;
50ec8a40 1550 p->numa_group->faults_cpu[i] += f_diff;
989348b5
MG
1551 p->numa_group->total_faults += diff;
1552 group_faults += p->numa_group->faults[i];
8c8a743c 1553 }
ac8e895b
MG
1554 }
1555
688b7585
MG
1556 if (faults > max_faults) {
1557 max_faults = faults;
1558 max_nid = nid;
1559 }
83e1d2cd
MG
1560
1561 if (group_faults > max_group_faults) {
1562 max_group_faults = group_faults;
1563 max_group_nid = nid;
1564 }
1565 }
1566
04bb2f94
RR
1567 update_task_scan_period(p, fault_types[0], fault_types[1]);
1568
7dbd13ed 1569 if (p->numa_group) {
20e07dea 1570 update_numa_active_node_mask(p->numa_group);
7dbd13ed
MG
1571 /*
1572 * If the preferred task and group nids are different,
1573 * iterate over the nodes again to find the best place.
1574 */
1575 if (max_nid != max_group_nid) {
1576 unsigned long weight, max_weight = 0;
1577
1578 for_each_online_node(nid) {
1579 weight = task_weight(p, nid) + group_weight(p, nid);
1580 if (weight > max_weight) {
1581 max_weight = weight;
1582 max_nid = nid;
1583 }
83e1d2cd
MG
1584 }
1585 }
7dbd13ed 1586
60e69eed 1587 spin_unlock_irq(group_lock);
688b7585
MG
1588 }
1589
6b9a7460 1590 /* Preferred node as the node with the most faults */
3a7053b3 1591 if (max_faults && max_nid != p->numa_preferred_nid) {
e6628d5b 1592 /* Update the preferred nid and migrate task if possible */
0ec8aa00 1593 sched_setnuma(p, max_nid);
6b9a7460 1594 numa_migrate_preferred(p);
3a7053b3 1595 }
cbee9f88
PZ
1596}
1597
8c8a743c
PZ
1598static inline int get_numa_group(struct numa_group *grp)
1599{
1600 return atomic_inc_not_zero(&grp->refcount);
1601}
1602
1603static inline void put_numa_group(struct numa_group *grp)
1604{
1605 if (atomic_dec_and_test(&grp->refcount))
1606 kfree_rcu(grp, rcu);
1607}
1608
3e6a9418
MG
1609static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1610 int *priv)
8c8a743c
PZ
1611{
1612 struct numa_group *grp, *my_grp;
1613 struct task_struct *tsk;
1614 bool join = false;
1615 int cpu = cpupid_to_cpu(cpupid);
1616 int i;
1617
1618 if (unlikely(!p->numa_group)) {
1619 unsigned int size = sizeof(struct numa_group) +
50ec8a40 1620 4*nr_node_ids*sizeof(unsigned long);
8c8a743c
PZ
1621
1622 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1623 if (!grp)
1624 return;
1625
1626 atomic_set(&grp->refcount, 1);
1627 spin_lock_init(&grp->lock);
1628 INIT_LIST_HEAD(&grp->task_list);
e29cf08b 1629 grp->gid = p->pid;
50ec8a40 1630 /* Second half of the array tracks nids where faults happen */
be1e4e76
RR
1631 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1632 nr_node_ids;
8c8a743c 1633
20e07dea
RR
1634 node_set(task_node(current), grp->active_nodes);
1635
be1e4e76 1636 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
ff1df896 1637 grp->faults[i] = p->numa_faults_memory[i];
8c8a743c 1638
989348b5 1639 grp->total_faults = p->total_numa_faults;
83e1d2cd 1640
8c8a743c
PZ
1641 list_add(&p->numa_entry, &grp->task_list);
1642 grp->nr_tasks++;
1643 rcu_assign_pointer(p->numa_group, grp);
1644 }
1645
1646 rcu_read_lock();
1647 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1648
1649 if (!cpupid_match_pid(tsk, cpupid))
3354781a 1650 goto no_join;
8c8a743c
PZ
1651
1652 grp = rcu_dereference(tsk->numa_group);
1653 if (!grp)
3354781a 1654 goto no_join;
8c8a743c
PZ
1655
1656 my_grp = p->numa_group;
1657 if (grp == my_grp)
3354781a 1658 goto no_join;
8c8a743c
PZ
1659
1660 /*
1661 * Only join the other group if its bigger; if we're the bigger group,
1662 * the other task will join us.
1663 */
1664 if (my_grp->nr_tasks > grp->nr_tasks)
3354781a 1665 goto no_join;
8c8a743c
PZ
1666
1667 /*
1668 * Tie-break on the grp address.
1669 */
1670 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
3354781a 1671 goto no_join;
8c8a743c 1672
dabe1d99
RR
1673 /* Always join threads in the same process. */
1674 if (tsk->mm == current->mm)
1675 join = true;
1676
1677 /* Simple filter to avoid false positives due to PID collisions */
1678 if (flags & TNF_SHARED)
1679 join = true;
8c8a743c 1680
3e6a9418
MG
1681 /* Update priv based on whether false sharing was detected */
1682 *priv = !join;
1683
dabe1d99 1684 if (join && !get_numa_group(grp))
3354781a 1685 goto no_join;
8c8a743c 1686
8c8a743c
PZ
1687 rcu_read_unlock();
1688
1689 if (!join)
1690 return;
1691
60e69eed
MG
1692 BUG_ON(irqs_disabled());
1693 double_lock_irq(&my_grp->lock, &grp->lock);
989348b5 1694
be1e4e76 1695 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
ff1df896
RR
1696 my_grp->faults[i] -= p->numa_faults_memory[i];
1697 grp->faults[i] += p->numa_faults_memory[i];
8c8a743c 1698 }
989348b5
MG
1699 my_grp->total_faults -= p->total_numa_faults;
1700 grp->total_faults += p->total_numa_faults;
8c8a743c
PZ
1701
1702 list_move(&p->numa_entry, &grp->task_list);
1703 my_grp->nr_tasks--;
1704 grp->nr_tasks++;
1705
1706 spin_unlock(&my_grp->lock);
60e69eed 1707 spin_unlock_irq(&grp->lock);
8c8a743c
PZ
1708
1709 rcu_assign_pointer(p->numa_group, grp);
1710
1711 put_numa_group(my_grp);
3354781a
PZ
1712 return;
1713
1714no_join:
1715 rcu_read_unlock();
1716 return;
8c8a743c
PZ
1717}
1718
1719void task_numa_free(struct task_struct *p)
1720{
1721 struct numa_group *grp = p->numa_group;
1722 int i;
ff1df896 1723 void *numa_faults = p->numa_faults_memory;
8c8a743c
PZ
1724
1725 if (grp) {
60e69eed 1726 spin_lock_irq(&grp->lock);
be1e4e76 1727 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
ff1df896 1728 grp->faults[i] -= p->numa_faults_memory[i];
989348b5 1729 grp->total_faults -= p->total_numa_faults;
83e1d2cd 1730
8c8a743c
PZ
1731 list_del(&p->numa_entry);
1732 grp->nr_tasks--;
60e69eed 1733 spin_unlock_irq(&grp->lock);
8c8a743c
PZ
1734 rcu_assign_pointer(p->numa_group, NULL);
1735 put_numa_group(grp);
1736 }
1737
ff1df896
RR
1738 p->numa_faults_memory = NULL;
1739 p->numa_faults_buffer_memory = NULL;
50ec8a40
RR
1740 p->numa_faults_cpu= NULL;
1741 p->numa_faults_buffer_cpu = NULL;
82727018 1742 kfree(numa_faults);
8c8a743c
PZ
1743}
1744
cbee9f88
PZ
1745/*
1746 * Got a PROT_NONE fault for a page on @node.
1747 */
58b46da3 1748void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
cbee9f88
PZ
1749{
1750 struct task_struct *p = current;
6688cc05 1751 bool migrated = flags & TNF_MIGRATED;
58b46da3 1752 int cpu_node = task_node(current);
792568ec 1753 int local = !!(flags & TNF_FAULT_LOCAL);
ac8e895b 1754 int priv;
cbee9f88 1755
10e84b97 1756 if (!numabalancing_enabled)
1a687c2e
MG
1757 return;
1758
9ff1d9ff
MG
1759 /* for example, ksmd faulting in a user's mm */
1760 if (!p->mm)
1761 return;
1762
82727018
RR
1763 /* Do not worry about placement if exiting */
1764 if (p->state == TASK_DEAD)
1765 return;
1766
f809ca9a 1767 /* Allocate buffer to track faults on a per-node basis */
ff1df896 1768 if (unlikely(!p->numa_faults_memory)) {
be1e4e76
RR
1769 int size = sizeof(*p->numa_faults_memory) *
1770 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
f809ca9a 1771
be1e4e76 1772 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
ff1df896 1773 if (!p->numa_faults_memory)
f809ca9a 1774 return;
745d6147 1775
ff1df896 1776 BUG_ON(p->numa_faults_buffer_memory);
be1e4e76
RR
1777 /*
1778 * The averaged statistics, shared & private, memory & cpu,
1779 * occupy the first half of the array. The second half of the
1780 * array is for current counters, which are averaged into the
1781 * first set by task_numa_placement.
1782 */
50ec8a40
RR
1783 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1784 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1785 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
83e1d2cd 1786 p->total_numa_faults = 0;
04bb2f94 1787 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
f809ca9a 1788 }
cbee9f88 1789
8c8a743c
PZ
1790 /*
1791 * First accesses are treated as private, otherwise consider accesses
1792 * to be private if the accessing pid has not changed
1793 */
1794 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1795 priv = 1;
1796 } else {
1797 priv = cpupid_match_pid(p, last_cpupid);
6688cc05 1798 if (!priv && !(flags & TNF_NO_GROUP))
3e6a9418 1799 task_numa_group(p, last_cpupid, flags, &priv);
8c8a743c
PZ
1800 }
1801
792568ec
RR
1802 /*
1803 * If a workload spans multiple NUMA nodes, a shared fault that
1804 * occurs wholly within the set of nodes that the workload is
1805 * actively using should be counted as local. This allows the
1806 * scan rate to slow down when a workload has settled down.
1807 */
1808 if (!priv && !local && p->numa_group &&
1809 node_isset(cpu_node, p->numa_group->active_nodes) &&
1810 node_isset(mem_node, p->numa_group->active_nodes))
1811 local = 1;
1812
cbee9f88 1813 task_numa_placement(p);
f809ca9a 1814
2739d3ee
RR
1815 /*
1816 * Retry task to preferred node migration periodically, in case it
1817 * case it previously failed, or the scheduler moved us.
1818 */
1819 if (time_after(jiffies, p->numa_migrate_retry))
6b9a7460
MG
1820 numa_migrate_preferred(p);
1821
b32e86b4
IM
1822 if (migrated)
1823 p->numa_pages_migrated += pages;
1824
58b46da3
RR
1825 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1826 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
792568ec 1827 p->numa_faults_locality[local] += pages;
cbee9f88
PZ
1828}
1829
6e5fb223
PZ
1830static void reset_ptenuma_scan(struct task_struct *p)
1831{
1832 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1833 p->mm->numa_scan_offset = 0;
1834}
1835
cbee9f88
PZ
1836/*
1837 * The expensive part of numa migration is done from task_work context.
1838 * Triggered from task_tick_numa().
1839 */
1840void task_numa_work(struct callback_head *work)
1841{
1842 unsigned long migrate, next_scan, now = jiffies;
1843 struct task_struct *p = current;
1844 struct mm_struct *mm = p->mm;
6e5fb223 1845 struct vm_area_struct *vma;
9f40604c 1846 unsigned long start, end;
598f0ec0 1847 unsigned long nr_pte_updates = 0;
9f40604c 1848 long pages;
cbee9f88
PZ
1849
1850 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1851
1852 work->next = work; /* protect against double add */
1853 /*
1854 * Who cares about NUMA placement when they're dying.
1855 *
1856 * NOTE: make sure not to dereference p->mm before this check,
1857 * exit_task_work() happens _after_ exit_mm() so we could be called
1858 * without p->mm even though we still had it when we enqueued this
1859 * work.
1860 */
1861 if (p->flags & PF_EXITING)
1862 return;
1863
930aa174 1864 if (!mm->numa_next_scan) {
7e8d16b6
MG
1865 mm->numa_next_scan = now +
1866 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
b8593bfd
MG
1867 }
1868
cbee9f88
PZ
1869 /*
1870 * Enforce maximal scan/migration frequency..
1871 */
1872 migrate = mm->numa_next_scan;
1873 if (time_before(now, migrate))
1874 return;
1875
598f0ec0
MG
1876 if (p->numa_scan_period == 0) {
1877 p->numa_scan_period_max = task_scan_max(p);
1878 p->numa_scan_period = task_scan_min(p);
1879 }
cbee9f88 1880
fb003b80 1881 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
cbee9f88
PZ
1882 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1883 return;
1884
19a78d11
PZ
1885 /*
1886 * Delay this task enough that another task of this mm will likely win
1887 * the next time around.
1888 */
1889 p->node_stamp += 2 * TICK_NSEC;
1890
9f40604c
MG
1891 start = mm->numa_scan_offset;
1892 pages = sysctl_numa_balancing_scan_size;
1893 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1894 if (!pages)
1895 return;
cbee9f88 1896
6e5fb223 1897 down_read(&mm->mmap_sem);
9f40604c 1898 vma = find_vma(mm, start);
6e5fb223
PZ
1899 if (!vma) {
1900 reset_ptenuma_scan(p);
9f40604c 1901 start = 0;
6e5fb223
PZ
1902 vma = mm->mmap;
1903 }
9f40604c 1904 for (; vma; vma = vma->vm_next) {
fc314724 1905 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
6e5fb223
PZ
1906 continue;
1907
4591ce4f
MG
1908 /*
1909 * Shared library pages mapped by multiple processes are not
1910 * migrated as it is expected they are cache replicated. Avoid
1911 * hinting faults in read-only file-backed mappings or the vdso
1912 * as migrating the pages will be of marginal benefit.
1913 */
1914 if (!vma->vm_mm ||
1915 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1916 continue;
1917
3c67f474
MG
1918 /*
1919 * Skip inaccessible VMAs to avoid any confusion between
1920 * PROT_NONE and NUMA hinting ptes
1921 */
1922 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1923 continue;
4591ce4f 1924
9f40604c
MG
1925 do {
1926 start = max(start, vma->vm_start);
1927 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1928 end = min(end, vma->vm_end);
598f0ec0
MG
1929 nr_pte_updates += change_prot_numa(vma, start, end);
1930
1931 /*
1932 * Scan sysctl_numa_balancing_scan_size but ensure that
1933 * at least one PTE is updated so that unused virtual
1934 * address space is quickly skipped.
1935 */
1936 if (nr_pte_updates)
1937 pages -= (end - start) >> PAGE_SHIFT;
6e5fb223 1938
9f40604c
MG
1939 start = end;
1940 if (pages <= 0)
1941 goto out;
3cf1962c
RR
1942
1943 cond_resched();
9f40604c 1944 } while (end != vma->vm_end);
cbee9f88 1945 }
6e5fb223 1946
9f40604c 1947out:
6e5fb223 1948 /*
c69307d5
PZ
1949 * It is possible to reach the end of the VMA list but the last few
1950 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1951 * would find the !migratable VMA on the next scan but not reset the
1952 * scanner to the start so check it now.
6e5fb223
PZ
1953 */
1954 if (vma)
9f40604c 1955 mm->numa_scan_offset = start;
6e5fb223
PZ
1956 else
1957 reset_ptenuma_scan(p);
1958 up_read(&mm->mmap_sem);
cbee9f88
PZ
1959}
1960
1961/*
1962 * Drive the periodic memory faults..
1963 */
1964void task_tick_numa(struct rq *rq, struct task_struct *curr)
1965{
1966 struct callback_head *work = &curr->numa_work;
1967 u64 period, now;
1968
1969 /*
1970 * We don't care about NUMA placement if we don't have memory.
1971 */
1972 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1973 return;
1974
1975 /*
1976 * Using runtime rather than walltime has the dual advantage that
1977 * we (mostly) drive the selection from busy threads and that the
1978 * task needs to have done some actual work before we bother with
1979 * NUMA placement.
1980 */
1981 now = curr->se.sum_exec_runtime;
1982 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1983
1984 if (now - curr->node_stamp > period) {
4b96a29b 1985 if (!curr->node_stamp)
598f0ec0 1986 curr->numa_scan_period = task_scan_min(curr);
19a78d11 1987 curr->node_stamp += period;
cbee9f88
PZ
1988
1989 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1990 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1991 task_work_add(curr, work, true);
1992 }
1993 }
1994}
1995#else
1996static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1997{
1998}
0ec8aa00
PZ
1999
2000static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2001{
2002}
2003
2004static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2005{
2006}
cbee9f88
PZ
2007#endif /* CONFIG_NUMA_BALANCING */
2008
30cfdcfc
DA
2009static void
2010account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2011{
2012 update_load_add(&cfs_rq->load, se->load.weight);
c09595f6 2013 if (!parent_entity(se))
029632fb 2014 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
367456c7 2015#ifdef CONFIG_SMP
0ec8aa00
PZ
2016 if (entity_is_task(se)) {
2017 struct rq *rq = rq_of(cfs_rq);
2018
2019 account_numa_enqueue(rq, task_of(se));
2020 list_add(&se->group_node, &rq->cfs_tasks);
2021 }
367456c7 2022#endif
30cfdcfc 2023 cfs_rq->nr_running++;
30cfdcfc
DA
2024}
2025
2026static void
2027account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2028{
2029 update_load_sub(&cfs_rq->load, se->load.weight);
c09595f6 2030 if (!parent_entity(se))
029632fb 2031 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
0ec8aa00
PZ
2032 if (entity_is_task(se)) {
2033 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
b87f1724 2034 list_del_init(&se->group_node);
0ec8aa00 2035 }
30cfdcfc 2036 cfs_rq->nr_running--;
30cfdcfc
DA
2037}
2038
3ff6dcac
YZ
2039#ifdef CONFIG_FAIR_GROUP_SCHED
2040# ifdef CONFIG_SMP
cf5f0acf
PZ
2041static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2042{
2043 long tg_weight;
2044
2045 /*
2046 * Use this CPU's actual weight instead of the last load_contribution
2047 * to gain a more accurate current total weight. See
2048 * update_cfs_rq_load_contribution().
2049 */
bf5b986e 2050 tg_weight = atomic_long_read(&tg->load_avg);
82958366 2051 tg_weight -= cfs_rq->tg_load_contrib;
cf5f0acf
PZ
2052 tg_weight += cfs_rq->load.weight;
2053
2054 return tg_weight;
2055}
2056
6d5ab293 2057static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
3ff6dcac 2058{
cf5f0acf 2059 long tg_weight, load, shares;
3ff6dcac 2060
cf5f0acf 2061 tg_weight = calc_tg_weight(tg, cfs_rq);
6d5ab293 2062 load = cfs_rq->load.weight;
3ff6dcac 2063
3ff6dcac 2064 shares = (tg->shares * load);
cf5f0acf
PZ
2065 if (tg_weight)
2066 shares /= tg_weight;
3ff6dcac
YZ
2067
2068 if (shares < MIN_SHARES)
2069 shares = MIN_SHARES;
2070 if (shares > tg->shares)
2071 shares = tg->shares;
2072
2073 return shares;
2074}
3ff6dcac 2075# else /* CONFIG_SMP */
6d5ab293 2076static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
3ff6dcac
YZ
2077{
2078 return tg->shares;
2079}
3ff6dcac 2080# endif /* CONFIG_SMP */
2069dd75
PZ
2081static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2082 unsigned long weight)
2083{
19e5eebb
PT
2084 if (se->on_rq) {
2085 /* commit outstanding execution time */
2086 if (cfs_rq->curr == se)
2087 update_curr(cfs_rq);
2069dd75 2088 account_entity_dequeue(cfs_rq, se);
19e5eebb 2089 }
2069dd75
PZ
2090
2091 update_load_set(&se->load, weight);
2092
2093 if (se->on_rq)
2094 account_entity_enqueue(cfs_rq, se);
2095}
2096
82958366
PT
2097static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2098
6d5ab293 2099static void update_cfs_shares(struct cfs_rq *cfs_rq)
2069dd75
PZ
2100{
2101 struct task_group *tg;
2102 struct sched_entity *se;
3ff6dcac 2103 long shares;
2069dd75 2104
2069dd75
PZ
2105 tg = cfs_rq->tg;
2106 se = tg->se[cpu_of(rq_of(cfs_rq))];
64660c86 2107 if (!se || throttled_hierarchy(cfs_rq))
2069dd75 2108 return;
3ff6dcac
YZ
2109#ifndef CONFIG_SMP
2110 if (likely(se->load.weight == tg->shares))
2111 return;
2112#endif
6d5ab293 2113 shares = calc_cfs_shares(cfs_rq, tg);
2069dd75
PZ
2114
2115 reweight_entity(cfs_rq_of(se), se, shares);
2116}
2117#else /* CONFIG_FAIR_GROUP_SCHED */
6d5ab293 2118static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2069dd75
PZ
2119{
2120}
2121#endif /* CONFIG_FAIR_GROUP_SCHED */
2122
141965c7 2123#ifdef CONFIG_SMP
5b51f2f8
PT
2124/*
2125 * We choose a half-life close to 1 scheduling period.
2126 * Note: The tables below are dependent on this value.
2127 */
2128#define LOAD_AVG_PERIOD 32
2129#define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2130#define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2131
2132/* Precomputed fixed inverse multiplies for multiplication by y^n */
2133static const u32 runnable_avg_yN_inv[] = {
2134 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2135 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2136 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2137 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2138 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2139 0x85aac367, 0x82cd8698,
2140};
2141
2142/*
2143 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2144 * over-estimates when re-combining.
2145 */
2146static const u32 runnable_avg_yN_sum[] = {
2147 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2148 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2149 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2150};
2151
9d85f21c
PT
2152/*
2153 * Approximate:
2154 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2155 */
2156static __always_inline u64 decay_load(u64 val, u64 n)
2157{
5b51f2f8
PT
2158 unsigned int local_n;
2159
2160 if (!n)
2161 return val;
2162 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2163 return 0;
2164
2165 /* after bounds checking we can collapse to 32-bit */
2166 local_n = n;
2167
2168 /*
2169 * As y^PERIOD = 1/2, we can combine
2170 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2171 * With a look-up table which covers k^n (n<PERIOD)
2172 *
2173 * To achieve constant time decay_load.
2174 */
2175 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2176 val >>= local_n / LOAD_AVG_PERIOD;
2177 local_n %= LOAD_AVG_PERIOD;
9d85f21c
PT
2178 }
2179
5b51f2f8
PT
2180 val *= runnable_avg_yN_inv[local_n];
2181 /* We don't use SRR here since we always want to round down. */
2182 return val >> 32;
2183}
2184
2185/*
2186 * For updates fully spanning n periods, the contribution to runnable
2187 * average will be: \Sum 1024*y^n
2188 *
2189 * We can compute this reasonably efficiently by combining:
2190 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2191 */
2192static u32 __compute_runnable_contrib(u64 n)
2193{
2194 u32 contrib = 0;
2195
2196 if (likely(n <= LOAD_AVG_PERIOD))
2197 return runnable_avg_yN_sum[n];
2198 else if (unlikely(n >= LOAD_AVG_MAX_N))
2199 return LOAD_AVG_MAX;
2200
2201 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2202 do {
2203 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2204 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2205
2206 n -= LOAD_AVG_PERIOD;
2207 } while (n > LOAD_AVG_PERIOD);
2208
2209 contrib = decay_load(contrib, n);
2210 return contrib + runnable_avg_yN_sum[n];
9d85f21c
PT
2211}
2212
2213/*
2214 * We can represent the historical contribution to runnable average as the
2215 * coefficients of a geometric series. To do this we sub-divide our runnable
2216 * history into segments of approximately 1ms (1024us); label the segment that
2217 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2218 *
2219 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2220 * p0 p1 p2
2221 * (now) (~1ms ago) (~2ms ago)
2222 *
2223 * Let u_i denote the fraction of p_i that the entity was runnable.
2224 *
2225 * We then designate the fractions u_i as our co-efficients, yielding the
2226 * following representation of historical load:
2227 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2228 *
2229 * We choose y based on the with of a reasonably scheduling period, fixing:
2230 * y^32 = 0.5
2231 *
2232 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2233 * approximately half as much as the contribution to load within the last ms
2234 * (u_0).
2235 *
2236 * When a period "rolls over" and we have new u_0`, multiplying the previous
2237 * sum again by y is sufficient to update:
2238 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2239 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2240 */
2241static __always_inline int __update_entity_runnable_avg(u64 now,
2242 struct sched_avg *sa,
2243 int runnable)
2244{
5b51f2f8
PT
2245 u64 delta, periods;
2246 u32 runnable_contrib;
9d85f21c
PT
2247 int delta_w, decayed = 0;
2248
2249 delta = now - sa->last_runnable_update;
2250 /*
2251 * This should only happen when time goes backwards, which it
2252 * unfortunately does during sched clock init when we swap over to TSC.
2253 */
2254 if ((s64)delta < 0) {
2255 sa->last_runnable_update = now;
2256 return 0;
2257 }
2258
2259 /*
2260 * Use 1024ns as the unit of measurement since it's a reasonable
2261 * approximation of 1us and fast to compute.
2262 */
2263 delta >>= 10;
2264 if (!delta)
2265 return 0;
2266 sa->last_runnable_update = now;
2267
2268 /* delta_w is the amount already accumulated against our next period */
2269 delta_w = sa->runnable_avg_period % 1024;
2270 if (delta + delta_w >= 1024) {
2271 /* period roll-over */
2272 decayed = 1;
2273
2274 /*
2275 * Now that we know we're crossing a period boundary, figure
2276 * out how much from delta we need to complete the current
2277 * period and accrue it.
2278 */
2279 delta_w = 1024 - delta_w;
5b51f2f8
PT
2280 if (runnable)
2281 sa->runnable_avg_sum += delta_w;
2282 sa->runnable_avg_period += delta_w;
2283
2284 delta -= delta_w;
2285
2286 /* Figure out how many additional periods this update spans */
2287 periods = delta / 1024;
2288 delta %= 1024;
2289
2290 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2291 periods + 1);
2292 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2293 periods + 1);
2294
2295 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2296 runnable_contrib = __compute_runnable_contrib(periods);
2297 if (runnable)
2298 sa->runnable_avg_sum += runnable_contrib;
2299 sa->runnable_avg_period += runnable_contrib;
9d85f21c
PT
2300 }
2301
2302 /* Remainder of delta accrued against u_0` */
2303 if (runnable)
2304 sa->runnable_avg_sum += delta;
2305 sa->runnable_avg_period += delta;
2306
2307 return decayed;
2308}
2309
9ee474f5 2310/* Synchronize an entity's decay with its parenting cfs_rq.*/
aff3e498 2311static inline u64 __synchronize_entity_decay(struct sched_entity *se)
9ee474f5
PT
2312{
2313 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2314 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2315
2316 decays -= se->avg.decay_count;
2317 if (!decays)
aff3e498 2318 return 0;
9ee474f5
PT
2319
2320 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2321 se->avg.decay_count = 0;
aff3e498
PT
2322
2323 return decays;
9ee474f5
PT
2324}
2325
c566e8e9
PT
2326#ifdef CONFIG_FAIR_GROUP_SCHED
2327static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2328 int force_update)
2329{
2330 struct task_group *tg = cfs_rq->tg;
bf5b986e 2331 long tg_contrib;
c566e8e9
PT
2332
2333 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2334 tg_contrib -= cfs_rq->tg_load_contrib;
2335
bf5b986e
AS
2336 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2337 atomic_long_add(tg_contrib, &tg->load_avg);
c566e8e9
PT
2338 cfs_rq->tg_load_contrib += tg_contrib;
2339 }
2340}
8165e145 2341
bb17f655
PT
2342/*
2343 * Aggregate cfs_rq runnable averages into an equivalent task_group
2344 * representation for computing load contributions.
2345 */
2346static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2347 struct cfs_rq *cfs_rq)
2348{
2349 struct task_group *tg = cfs_rq->tg;
2350 long contrib;
2351
2352 /* The fraction of a cpu used by this cfs_rq */
85b088e9 2353 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
bb17f655
PT
2354 sa->runnable_avg_period + 1);
2355 contrib -= cfs_rq->tg_runnable_contrib;
2356
2357 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2358 atomic_add(contrib, &tg->runnable_avg);
2359 cfs_rq->tg_runnable_contrib += contrib;
2360 }
2361}
2362
8165e145
PT
2363static inline void __update_group_entity_contrib(struct sched_entity *se)
2364{
2365 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2366 struct task_group *tg = cfs_rq->tg;
bb17f655
PT
2367 int runnable_avg;
2368
8165e145
PT
2369 u64 contrib;
2370
2371 contrib = cfs_rq->tg_load_contrib * tg->shares;
bf5b986e
AS
2372 se->avg.load_avg_contrib = div_u64(contrib,
2373 atomic_long_read(&tg->load_avg) + 1);
bb17f655
PT
2374
2375 /*
2376 * For group entities we need to compute a correction term in the case
2377 * that they are consuming <1 cpu so that we would contribute the same
2378 * load as a task of equal weight.
2379 *
2380 * Explicitly co-ordinating this measurement would be expensive, but
2381 * fortunately the sum of each cpus contribution forms a usable
2382 * lower-bound on the true value.
2383 *
2384 * Consider the aggregate of 2 contributions. Either they are disjoint
2385 * (and the sum represents true value) or they are disjoint and we are
2386 * understating by the aggregate of their overlap.
2387 *
2388 * Extending this to N cpus, for a given overlap, the maximum amount we
2389 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2390 * cpus that overlap for this interval and w_i is the interval width.
2391 *
2392 * On a small machine; the first term is well-bounded which bounds the
2393 * total error since w_i is a subset of the period. Whereas on a
2394 * larger machine, while this first term can be larger, if w_i is the
2395 * of consequential size guaranteed to see n_i*w_i quickly converge to
2396 * our upper bound of 1-cpu.
2397 */
2398 runnable_avg = atomic_read(&tg->runnable_avg);
2399 if (runnable_avg < NICE_0_LOAD) {
2400 se->avg.load_avg_contrib *= runnable_avg;
2401 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2402 }
8165e145 2403}
f5f9739d
DE
2404
2405static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2406{
2407 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2408 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2409}
6e83125c 2410#else /* CONFIG_FAIR_GROUP_SCHED */
c566e8e9
PT
2411static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2412 int force_update) {}
bb17f655
PT
2413static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2414 struct cfs_rq *cfs_rq) {}
8165e145 2415static inline void __update_group_entity_contrib(struct sched_entity *se) {}
f5f9739d 2416static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
6e83125c 2417#endif /* CONFIG_FAIR_GROUP_SCHED */
c566e8e9 2418
8165e145
PT
2419static inline void __update_task_entity_contrib(struct sched_entity *se)
2420{
2421 u32 contrib;
2422
2423 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2424 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2425 contrib /= (se->avg.runnable_avg_period + 1);
2426 se->avg.load_avg_contrib = scale_load(contrib);
2427}
2428
2dac754e
PT
2429/* Compute the current contribution to load_avg by se, return any delta */
2430static long __update_entity_load_avg_contrib(struct sched_entity *se)
2431{
2432 long old_contrib = se->avg.load_avg_contrib;
2433
8165e145
PT
2434 if (entity_is_task(se)) {
2435 __update_task_entity_contrib(se);
2436 } else {
bb17f655 2437 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
8165e145
PT
2438 __update_group_entity_contrib(se);
2439 }
2dac754e
PT
2440
2441 return se->avg.load_avg_contrib - old_contrib;
2442}
2443
9ee474f5
PT
2444static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2445 long load_contrib)
2446{
2447 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2448 cfs_rq->blocked_load_avg -= load_contrib;
2449 else
2450 cfs_rq->blocked_load_avg = 0;
2451}
2452
f1b17280
PT
2453static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2454
9d85f21c 2455/* Update a sched_entity's runnable average */
9ee474f5
PT
2456static inline void update_entity_load_avg(struct sched_entity *se,
2457 int update_cfs_rq)
9d85f21c 2458{
2dac754e
PT
2459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2460 long contrib_delta;
f1b17280 2461 u64 now;
2dac754e 2462
f1b17280
PT
2463 /*
2464 * For a group entity we need to use their owned cfs_rq_clock_task() in
2465 * case they are the parent of a throttled hierarchy.
2466 */
2467 if (entity_is_task(se))
2468 now = cfs_rq_clock_task(cfs_rq);
2469 else
2470 now = cfs_rq_clock_task(group_cfs_rq(se));
2471
2472 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2dac754e
PT
2473 return;
2474
2475 contrib_delta = __update_entity_load_avg_contrib(se);
9ee474f5
PT
2476
2477 if (!update_cfs_rq)
2478 return;
2479
2dac754e
PT
2480 if (se->on_rq)
2481 cfs_rq->runnable_load_avg += contrib_delta;
9ee474f5
PT
2482 else
2483 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2484}
2485
2486/*
2487 * Decay the load contributed by all blocked children and account this so that
2488 * their contribution may appropriately discounted when they wake up.
2489 */
aff3e498 2490static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
9ee474f5 2491{
f1b17280 2492 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
9ee474f5
PT
2493 u64 decays;
2494
2495 decays = now - cfs_rq->last_decay;
aff3e498 2496 if (!decays && !force_update)
9ee474f5
PT
2497 return;
2498
2509940f
AS
2499 if (atomic_long_read(&cfs_rq->removed_load)) {
2500 unsigned long removed_load;
2501 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
aff3e498
PT
2502 subtract_blocked_load_contrib(cfs_rq, removed_load);
2503 }
9ee474f5 2504
aff3e498
PT
2505 if (decays) {
2506 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2507 decays);
2508 atomic64_add(decays, &cfs_rq->decay_counter);
2509 cfs_rq->last_decay = now;
2510 }
c566e8e9
PT
2511
2512 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
9d85f21c 2513}
18bf2805 2514
2dac754e
PT
2515/* Add the load generated by se into cfs_rq's child load-average */
2516static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
9ee474f5
PT
2517 struct sched_entity *se,
2518 int wakeup)
2dac754e 2519{
aff3e498
PT
2520 /*
2521 * We track migrations using entity decay_count <= 0, on a wake-up
2522 * migration we use a negative decay count to track the remote decays
2523 * accumulated while sleeping.
a75cdaa9
AS
2524 *
2525 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2526 * are seen by enqueue_entity_load_avg() as a migration with an already
2527 * constructed load_avg_contrib.
aff3e498
PT
2528 */
2529 if (unlikely(se->avg.decay_count <= 0)) {
78becc27 2530 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
aff3e498
PT
2531 if (se->avg.decay_count) {
2532 /*
2533 * In a wake-up migration we have to approximate the
2534 * time sleeping. This is because we can't synchronize
2535 * clock_task between the two cpus, and it is not
2536 * guaranteed to be read-safe. Instead, we can
2537 * approximate this using our carried decays, which are
2538 * explicitly atomically readable.
2539 */
2540 se->avg.last_runnable_update -= (-se->avg.decay_count)
2541 << 20;
2542 update_entity_load_avg(se, 0);
2543 /* Indicate that we're now synchronized and on-rq */
2544 se->avg.decay_count = 0;
2545 }
9ee474f5
PT
2546 wakeup = 0;
2547 } else {
9390675a 2548 __synchronize_entity_decay(se);
9ee474f5
PT
2549 }
2550
aff3e498
PT
2551 /* migrated tasks did not contribute to our blocked load */
2552 if (wakeup) {
9ee474f5 2553 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
aff3e498
PT
2554 update_entity_load_avg(se, 0);
2555 }
9ee474f5 2556
2dac754e 2557 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
aff3e498
PT
2558 /* we force update consideration on load-balancer moves */
2559 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2dac754e
PT
2560}
2561
9ee474f5
PT
2562/*
2563 * Remove se's load from this cfs_rq child load-average, if the entity is
2564 * transitioning to a blocked state we track its projected decay using
2565 * blocked_load_avg.
2566 */
2dac754e 2567static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
9ee474f5
PT
2568 struct sched_entity *se,
2569 int sleep)
2dac754e 2570{
9ee474f5 2571 update_entity_load_avg(se, 1);
aff3e498
PT
2572 /* we force update consideration on load-balancer moves */
2573 update_cfs_rq_blocked_load(cfs_rq, !sleep);
9ee474f5 2574
2dac754e 2575 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
9ee474f5
PT
2576 if (sleep) {
2577 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2578 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2579 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2dac754e 2580}
642dbc39
VG
2581
2582/*
2583 * Update the rq's load with the elapsed running time before entering
2584 * idle. if the last scheduled task is not a CFS task, idle_enter will
2585 * be the only way to update the runnable statistic.
2586 */
2587void idle_enter_fair(struct rq *this_rq)
2588{
2589 update_rq_runnable_avg(this_rq, 1);
2590}
2591
2592/*
2593 * Update the rq's load with the elapsed idle time before a task is
2594 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2595 * be the only way to update the runnable statistic.
2596 */
2597void idle_exit_fair(struct rq *this_rq)
2598{
2599 update_rq_runnable_avg(this_rq, 0);
2600}
2601
6e83125c
PZ
2602static int idle_balance(struct rq *this_rq);
2603
38033c37
PZ
2604#else /* CONFIG_SMP */
2605
9ee474f5
PT
2606static inline void update_entity_load_avg(struct sched_entity *se,
2607 int update_cfs_rq) {}
18bf2805 2608static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2dac754e 2609static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
9ee474f5
PT
2610 struct sched_entity *se,
2611 int wakeup) {}
2dac754e 2612static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
9ee474f5
PT
2613 struct sched_entity *se,
2614 int sleep) {}
aff3e498
PT
2615static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2616 int force_update) {}
6e83125c
PZ
2617
2618static inline int idle_balance(struct rq *rq)
2619{
2620 return 0;
2621}
2622
38033c37 2623#endif /* CONFIG_SMP */
9d85f21c 2624
2396af69 2625static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 2626{
bf0f6f24 2627#ifdef CONFIG_SCHEDSTATS
e414314c
PZ
2628 struct task_struct *tsk = NULL;
2629
2630 if (entity_is_task(se))
2631 tsk = task_of(se);
2632
41acab88 2633 if (se->statistics.sleep_start) {
78becc27 2634 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
bf0f6f24
IM
2635
2636 if ((s64)delta < 0)
2637 delta = 0;
2638
41acab88
LDM
2639 if (unlikely(delta > se->statistics.sleep_max))
2640 se->statistics.sleep_max = delta;
bf0f6f24 2641
8c79a045 2642 se->statistics.sleep_start = 0;
41acab88 2643 se->statistics.sum_sleep_runtime += delta;
9745512c 2644
768d0c27 2645 if (tsk) {
e414314c 2646 account_scheduler_latency(tsk, delta >> 10, 1);
768d0c27
PZ
2647 trace_sched_stat_sleep(tsk, delta);
2648 }
bf0f6f24 2649 }
41acab88 2650 if (se->statistics.block_start) {
78becc27 2651 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
bf0f6f24
IM
2652
2653 if ((s64)delta < 0)
2654 delta = 0;
2655
41acab88
LDM
2656 if (unlikely(delta > se->statistics.block_max))
2657 se->statistics.block_max = delta;
bf0f6f24 2658
8c79a045 2659 se->statistics.block_start = 0;
41acab88 2660 se->statistics.sum_sleep_runtime += delta;
30084fbd 2661
e414314c 2662 if (tsk) {
8f0dfc34 2663 if (tsk->in_iowait) {
41acab88
LDM
2664 se->statistics.iowait_sum += delta;
2665 se->statistics.iowait_count++;
768d0c27 2666 trace_sched_stat_iowait(tsk, delta);
8f0dfc34
AV
2667 }
2668
b781a602
AV
2669 trace_sched_stat_blocked(tsk, delta);
2670
e414314c
PZ
2671 /*
2672 * Blocking time is in units of nanosecs, so shift by
2673 * 20 to get a milliseconds-range estimation of the
2674 * amount of time that the task spent sleeping:
2675 */
2676 if (unlikely(prof_on == SLEEP_PROFILING)) {
2677 profile_hits(SLEEP_PROFILING,
2678 (void *)get_wchan(tsk),
2679 delta >> 20);
2680 }
2681 account_scheduler_latency(tsk, delta >> 10, 0);
30084fbd 2682 }
bf0f6f24
IM
2683 }
2684#endif
2685}
2686
ddc97297
PZ
2687static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2688{
2689#ifdef CONFIG_SCHED_DEBUG
2690 s64 d = se->vruntime - cfs_rq->min_vruntime;
2691
2692 if (d < 0)
2693 d = -d;
2694
2695 if (d > 3*sysctl_sched_latency)
2696 schedstat_inc(cfs_rq, nr_spread_over);
2697#endif
2698}
2699
aeb73b04
PZ
2700static void
2701place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2702{
1af5f730 2703 u64 vruntime = cfs_rq->min_vruntime;
94dfb5e7 2704
2cb8600e
PZ
2705 /*
2706 * The 'current' period is already promised to the current tasks,
2707 * however the extra weight of the new task will slow them down a
2708 * little, place the new task so that it fits in the slot that
2709 * stays open at the end.
2710 */
94dfb5e7 2711 if (initial && sched_feat(START_DEBIT))
f9c0b095 2712 vruntime += sched_vslice(cfs_rq, se);
aeb73b04 2713
a2e7a7eb 2714 /* sleeps up to a single latency don't count. */
5ca9880c 2715 if (!initial) {
a2e7a7eb 2716 unsigned long thresh = sysctl_sched_latency;
a7be37ac 2717
a2e7a7eb
MG
2718 /*
2719 * Halve their sleep time's effect, to allow
2720 * for a gentler effect of sleepers:
2721 */
2722 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2723 thresh >>= 1;
51e0304c 2724
a2e7a7eb 2725 vruntime -= thresh;
aeb73b04
PZ
2726 }
2727
b5d9d734 2728 /* ensure we never gain time by being placed backwards. */
16c8f1c7 2729 se->vruntime = max_vruntime(se->vruntime, vruntime);
aeb73b04
PZ
2730}
2731
d3d9dc33
PT
2732static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2733
bf0f6f24 2734static void
88ec22d3 2735enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
bf0f6f24 2736{
88ec22d3
PZ
2737 /*
2738 * Update the normalized vruntime before updating min_vruntime
0fc576d5 2739 * through calling update_curr().
88ec22d3 2740 */
371fd7e7 2741 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
88ec22d3
PZ
2742 se->vruntime += cfs_rq->min_vruntime;
2743
bf0f6f24 2744 /*
a2a2d680 2745 * Update run-time statistics of the 'current'.
bf0f6f24 2746 */
b7cc0896 2747 update_curr(cfs_rq);
f269ae04 2748 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
17bc14b7
LT
2749 account_entity_enqueue(cfs_rq, se);
2750 update_cfs_shares(cfs_rq);
bf0f6f24 2751
88ec22d3 2752 if (flags & ENQUEUE_WAKEUP) {
aeb73b04 2753 place_entity(cfs_rq, se, 0);
2396af69 2754 enqueue_sleeper(cfs_rq, se);
e9acbff6 2755 }
bf0f6f24 2756
d2417e5a 2757 update_stats_enqueue(cfs_rq, se);
ddc97297 2758 check_spread(cfs_rq, se);
83b699ed
SV
2759 if (se != cfs_rq->curr)
2760 __enqueue_entity(cfs_rq, se);
2069dd75 2761 se->on_rq = 1;
3d4b47b4 2762
d3d9dc33 2763 if (cfs_rq->nr_running == 1) {
3d4b47b4 2764 list_add_leaf_cfs_rq(cfs_rq);
d3d9dc33
PT
2765 check_enqueue_throttle(cfs_rq);
2766 }
bf0f6f24
IM
2767}
2768
2c13c919 2769static void __clear_buddies_last(struct sched_entity *se)
2002c695 2770{
2c13c919
RR
2771 for_each_sched_entity(se) {
2772 struct cfs_rq *cfs_rq = cfs_rq_of(se);
f1044799 2773 if (cfs_rq->last != se)
2c13c919 2774 break;
f1044799
PZ
2775
2776 cfs_rq->last = NULL;
2c13c919
RR
2777 }
2778}
2002c695 2779
2c13c919
RR
2780static void __clear_buddies_next(struct sched_entity *se)
2781{
2782 for_each_sched_entity(se) {
2783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
f1044799 2784 if (cfs_rq->next != se)
2c13c919 2785 break;
f1044799
PZ
2786
2787 cfs_rq->next = NULL;
2c13c919 2788 }
2002c695
PZ
2789}
2790
ac53db59
RR
2791static void __clear_buddies_skip(struct sched_entity *se)
2792{
2793 for_each_sched_entity(se) {
2794 struct cfs_rq *cfs_rq = cfs_rq_of(se);
f1044799 2795 if (cfs_rq->skip != se)
ac53db59 2796 break;
f1044799
PZ
2797
2798 cfs_rq->skip = NULL;
ac53db59
RR
2799 }
2800}
2801
a571bbea
PZ
2802static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2803{
2c13c919
RR
2804 if (cfs_rq->last == se)
2805 __clear_buddies_last(se);
2806
2807 if (cfs_rq->next == se)
2808 __clear_buddies_next(se);
ac53db59
RR
2809
2810 if (cfs_rq->skip == se)
2811 __clear_buddies_skip(se);
a571bbea
PZ
2812}
2813
6c16a6dc 2814static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
d8b4986d 2815
bf0f6f24 2816static void
371fd7e7 2817dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
bf0f6f24 2818{
a2a2d680
DA
2819 /*
2820 * Update run-time statistics of the 'current'.
2821 */
2822 update_curr(cfs_rq);
17bc14b7 2823 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
a2a2d680 2824
19b6a2e3 2825 update_stats_dequeue(cfs_rq, se);
371fd7e7 2826 if (flags & DEQUEUE_SLEEP) {
67e9fb2a 2827#ifdef CONFIG_SCHEDSTATS
bf0f6f24
IM
2828 if (entity_is_task(se)) {
2829 struct task_struct *tsk = task_of(se);
2830
2831 if (tsk->state & TASK_INTERRUPTIBLE)
78becc27 2832 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
bf0f6f24 2833 if (tsk->state & TASK_UNINTERRUPTIBLE)
78becc27 2834 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
bf0f6f24 2835 }
db36cc7d 2836#endif
67e9fb2a
PZ
2837 }
2838
2002c695 2839 clear_buddies(cfs_rq, se);
4793241b 2840
83b699ed 2841 if (se != cfs_rq->curr)
30cfdcfc 2842 __dequeue_entity(cfs_rq, se);
17bc14b7 2843 se->on_rq = 0;
30cfdcfc 2844 account_entity_dequeue(cfs_rq, se);
88ec22d3
PZ
2845
2846 /*
2847 * Normalize the entity after updating the min_vruntime because the
2848 * update can refer to the ->curr item and we need to reflect this
2849 * movement in our normalized position.
2850 */
371fd7e7 2851 if (!(flags & DEQUEUE_SLEEP))
88ec22d3 2852 se->vruntime -= cfs_rq->min_vruntime;
1e876231 2853
d8b4986d
PT
2854 /* return excess runtime on last dequeue */
2855 return_cfs_rq_runtime(cfs_rq);
2856
1e876231 2857 update_min_vruntime(cfs_rq);
17bc14b7 2858 update_cfs_shares(cfs_rq);
bf0f6f24
IM
2859}
2860
2861/*
2862 * Preempt the current task with a newly woken task if needed:
2863 */
7c92e54f 2864static void
2e09bf55 2865check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
bf0f6f24 2866{
11697830 2867 unsigned long ideal_runtime, delta_exec;
f4cfb33e
WX
2868 struct sched_entity *se;
2869 s64 delta;
11697830 2870
6d0f0ebd 2871 ideal_runtime = sched_slice(cfs_rq, curr);
11697830 2872 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
a9f3e2b5 2873 if (delta_exec > ideal_runtime) {
bf0f6f24 2874 resched_task(rq_of(cfs_rq)->curr);
a9f3e2b5
MG
2875 /*
2876 * The current task ran long enough, ensure it doesn't get
2877 * re-elected due to buddy favours.
2878 */
2879 clear_buddies(cfs_rq, curr);
f685ceac
MG
2880 return;
2881 }
2882
2883 /*
2884 * Ensure that a task that missed wakeup preemption by a
2885 * narrow margin doesn't have to wait for a full slice.
2886 * This also mitigates buddy induced latencies under load.
2887 */
f685ceac
MG
2888 if (delta_exec < sysctl_sched_min_granularity)
2889 return;
2890
f4cfb33e
WX
2891 se = __pick_first_entity(cfs_rq);
2892 delta = curr->vruntime - se->vruntime;
f685ceac 2893
f4cfb33e
WX
2894 if (delta < 0)
2895 return;
d7d82944 2896
f4cfb33e
WX
2897 if (delta > ideal_runtime)
2898 resched_task(rq_of(cfs_rq)->curr);
bf0f6f24
IM
2899}
2900
83b699ed 2901static void
8494f412 2902set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 2903{
83b699ed
SV
2904 /* 'current' is not kept within the tree. */
2905 if (se->on_rq) {
2906 /*
2907 * Any task has to be enqueued before it get to execute on
2908 * a CPU. So account for the time it spent waiting on the
2909 * runqueue.
2910 */
2911 update_stats_wait_end(cfs_rq, se);
2912 __dequeue_entity(cfs_rq, se);
2913 }
2914
79303e9e 2915 update_stats_curr_start(cfs_rq, se);
429d43bc 2916 cfs_rq->curr = se;
eba1ed4b
IM
2917#ifdef CONFIG_SCHEDSTATS
2918 /*
2919 * Track our maximum slice length, if the CPU's load is at
2920 * least twice that of our own weight (i.e. dont track it
2921 * when there are only lesser-weight tasks around):
2922 */
495eca49 2923 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
41acab88 2924 se->statistics.slice_max = max(se->statistics.slice_max,
eba1ed4b
IM
2925 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2926 }
2927#endif
4a55b450 2928 se->prev_sum_exec_runtime = se->sum_exec_runtime;
bf0f6f24
IM
2929}
2930
3f3a4904
PZ
2931static int
2932wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2933
ac53db59
RR
2934/*
2935 * Pick the next process, keeping these things in mind, in this order:
2936 * 1) keep things fair between processes/task groups
2937 * 2) pick the "next" process, since someone really wants that to run
2938 * 3) pick the "last" process, for cache locality
2939 * 4) do not run the "skip" process, if something else is available
2940 */
678d5718
PZ
2941static struct sched_entity *
2942pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
aa2ac252 2943{
678d5718
PZ
2944 struct sched_entity *left = __pick_first_entity(cfs_rq);
2945 struct sched_entity *se;
2946
2947 /*
2948 * If curr is set we have to see if its left of the leftmost entity
2949 * still in the tree, provided there was anything in the tree at all.
2950 */
2951 if (!left || (curr && entity_before(curr, left)))
2952 left = curr;
2953
2954 se = left; /* ideally we run the leftmost entity */
f4b6755f 2955
ac53db59
RR
2956 /*
2957 * Avoid running the skip buddy, if running something else can
2958 * be done without getting too unfair.
2959 */
2960 if (cfs_rq->skip == se) {
678d5718
PZ
2961 struct sched_entity *second;
2962
2963 if (se == curr) {
2964 second = __pick_first_entity(cfs_rq);
2965 } else {
2966 second = __pick_next_entity(se);
2967 if (!second || (curr && entity_before(curr, second)))
2968 second = curr;
2969 }
2970
ac53db59
RR
2971 if (second && wakeup_preempt_entity(second, left) < 1)
2972 se = second;
2973 }
aa2ac252 2974
f685ceac
MG
2975 /*
2976 * Prefer last buddy, try to return the CPU to a preempted task.
2977 */
2978 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2979 se = cfs_rq->last;
2980
ac53db59
RR
2981 /*
2982 * Someone really wants this to run. If it's not unfair, run it.
2983 */
2984 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2985 se = cfs_rq->next;
2986
f685ceac 2987 clear_buddies(cfs_rq, se);
4793241b
PZ
2988
2989 return se;
aa2ac252
PZ
2990}
2991
678d5718 2992static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
d3d9dc33 2993
ab6cde26 2994static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
bf0f6f24
IM
2995{
2996 /*
2997 * If still on the runqueue then deactivate_task()
2998 * was not called and update_curr() has to be done:
2999 */
3000 if (prev->on_rq)
b7cc0896 3001 update_curr(cfs_rq);
bf0f6f24 3002
d3d9dc33
PT
3003 /* throttle cfs_rqs exceeding runtime */
3004 check_cfs_rq_runtime(cfs_rq);
3005
ddc97297 3006 check_spread(cfs_rq, prev);
30cfdcfc 3007 if (prev->on_rq) {
5870db5b 3008 update_stats_wait_start(cfs_rq, prev);
30cfdcfc
DA
3009 /* Put 'current' back into the tree. */
3010 __enqueue_entity(cfs_rq, prev);
9d85f21c 3011 /* in !on_rq case, update occurred at dequeue */
9ee474f5 3012 update_entity_load_avg(prev, 1);
30cfdcfc 3013 }
429d43bc 3014 cfs_rq->curr = NULL;
bf0f6f24
IM
3015}
3016
8f4d37ec
PZ
3017static void
3018entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
bf0f6f24 3019{
bf0f6f24 3020 /*
30cfdcfc 3021 * Update run-time statistics of the 'current'.
bf0f6f24 3022 */
30cfdcfc 3023 update_curr(cfs_rq);
bf0f6f24 3024
9d85f21c
PT
3025 /*
3026 * Ensure that runnable average is periodically updated.
3027 */
9ee474f5 3028 update_entity_load_avg(curr, 1);
aff3e498 3029 update_cfs_rq_blocked_load(cfs_rq, 1);
bf0bd948 3030 update_cfs_shares(cfs_rq);
9d85f21c 3031
8f4d37ec
PZ
3032#ifdef CONFIG_SCHED_HRTICK
3033 /*
3034 * queued ticks are scheduled to match the slice, so don't bother
3035 * validating it and just reschedule.
3036 */
983ed7a6
HH
3037 if (queued) {
3038 resched_task(rq_of(cfs_rq)->curr);
3039 return;
3040 }
8f4d37ec
PZ
3041 /*
3042 * don't let the period tick interfere with the hrtick preemption
3043 */
3044 if (!sched_feat(DOUBLE_TICK) &&
3045 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3046 return;
3047#endif
3048
2c2efaed 3049 if (cfs_rq->nr_running > 1)
2e09bf55 3050 check_preempt_tick(cfs_rq, curr);
bf0f6f24
IM
3051}
3052
ab84d31e
PT
3053
3054/**************************************************
3055 * CFS bandwidth control machinery
3056 */
3057
3058#ifdef CONFIG_CFS_BANDWIDTH
029632fb
PZ
3059
3060#ifdef HAVE_JUMP_LABEL
c5905afb 3061static struct static_key __cfs_bandwidth_used;
029632fb
PZ
3062
3063static inline bool cfs_bandwidth_used(void)
3064{
c5905afb 3065 return static_key_false(&__cfs_bandwidth_used);
029632fb
PZ
3066}
3067
1ee14e6c 3068void cfs_bandwidth_usage_inc(void)
029632fb 3069{
1ee14e6c
BS
3070 static_key_slow_inc(&__cfs_bandwidth_used);
3071}
3072
3073void cfs_bandwidth_usage_dec(void)
3074{
3075 static_key_slow_dec(&__cfs_bandwidth_used);
029632fb
PZ
3076}
3077#else /* HAVE_JUMP_LABEL */
3078static bool cfs_bandwidth_used(void)
3079{
3080 return true;
3081}
3082
1ee14e6c
BS
3083void cfs_bandwidth_usage_inc(void) {}
3084void cfs_bandwidth_usage_dec(void) {}
029632fb
PZ
3085#endif /* HAVE_JUMP_LABEL */
3086
ab84d31e
PT
3087/*
3088 * default period for cfs group bandwidth.
3089 * default: 0.1s, units: nanoseconds
3090 */
3091static inline u64 default_cfs_period(void)
3092{
3093 return 100000000ULL;
3094}
ec12cb7f
PT
3095
3096static inline u64 sched_cfs_bandwidth_slice(void)
3097{
3098 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3099}
3100
a9cf55b2
PT
3101/*
3102 * Replenish runtime according to assigned quota and update expiration time.
3103 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3104 * additional synchronization around rq->lock.
3105 *
3106 * requires cfs_b->lock
3107 */
029632fb 3108void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
a9cf55b2
PT
3109{
3110 u64 now;
3111
3112 if (cfs_b->quota == RUNTIME_INF)
3113 return;
3114
3115 now = sched_clock_cpu(smp_processor_id());
3116 cfs_b->runtime = cfs_b->quota;
3117 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3118}
3119
029632fb
PZ
3120static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3121{
3122 return &tg->cfs_bandwidth;
3123}
3124
f1b17280
PT
3125/* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3126static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3127{
3128 if (unlikely(cfs_rq->throttle_count))
3129 return cfs_rq->throttled_clock_task;
3130
78becc27 3131 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
f1b17280
PT
3132}
3133
85dac906
PT
3134/* returns 0 on failure to allocate runtime */
3135static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
ec12cb7f
PT
3136{
3137 struct task_group *tg = cfs_rq->tg;
3138 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
a9cf55b2 3139 u64 amount = 0, min_amount, expires;
ec12cb7f
PT
3140
3141 /* note: this is a positive sum as runtime_remaining <= 0 */
3142 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3143
3144 raw_spin_lock(&cfs_b->lock);
3145 if (cfs_b->quota == RUNTIME_INF)
3146 amount = min_amount;
58088ad0 3147 else {
a9cf55b2
PT
3148 /*
3149 * If the bandwidth pool has become inactive, then at least one
3150 * period must have elapsed since the last consumption.
3151 * Refresh the global state and ensure bandwidth timer becomes
3152 * active.
3153 */
3154 if (!cfs_b->timer_active) {
3155 __refill_cfs_bandwidth_runtime(cfs_b);
58088ad0 3156 __start_cfs_bandwidth(cfs_b);
a9cf55b2 3157 }
58088ad0
PT
3158
3159 if (cfs_b->runtime > 0) {
3160 amount = min(cfs_b->runtime, min_amount);
3161 cfs_b->runtime -= amount;
3162 cfs_b->idle = 0;
3163 }
ec12cb7f 3164 }
a9cf55b2 3165 expires = cfs_b->runtime_expires;
ec12cb7f
PT
3166 raw_spin_unlock(&cfs_b->lock);
3167
3168 cfs_rq->runtime_remaining += amount;
a9cf55b2
PT
3169 /*
3170 * we may have advanced our local expiration to account for allowed
3171 * spread between our sched_clock and the one on which runtime was
3172 * issued.
3173 */
3174 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3175 cfs_rq->runtime_expires = expires;
85dac906
PT
3176
3177 return cfs_rq->runtime_remaining > 0;
ec12cb7f
PT
3178}
3179
a9cf55b2
PT
3180/*
3181 * Note: This depends on the synchronization provided by sched_clock and the
3182 * fact that rq->clock snapshots this value.
3183 */
3184static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
ec12cb7f 3185{
a9cf55b2 3186 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
a9cf55b2
PT
3187
3188 /* if the deadline is ahead of our clock, nothing to do */
78becc27 3189 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
ec12cb7f
PT
3190 return;
3191
a9cf55b2
PT
3192 if (cfs_rq->runtime_remaining < 0)
3193 return;
3194
3195 /*
3196 * If the local deadline has passed we have to consider the
3197 * possibility that our sched_clock is 'fast' and the global deadline
3198 * has not truly expired.
3199 *
3200 * Fortunately we can check determine whether this the case by checking
3201 * whether the global deadline has advanced.
3202 */
3203
3204 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3205 /* extend local deadline, drift is bounded above by 2 ticks */
3206 cfs_rq->runtime_expires += TICK_NSEC;
3207 } else {
3208 /* global deadline is ahead, expiration has passed */
3209 cfs_rq->runtime_remaining = 0;
3210 }
3211}
3212
9dbdb155 3213static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
a9cf55b2
PT
3214{
3215 /* dock delta_exec before expiring quota (as it could span periods) */
ec12cb7f 3216 cfs_rq->runtime_remaining -= delta_exec;
a9cf55b2
PT
3217 expire_cfs_rq_runtime(cfs_rq);
3218
3219 if (likely(cfs_rq->runtime_remaining > 0))
ec12cb7f
PT
3220 return;
3221
85dac906
PT
3222 /*
3223 * if we're unable to extend our runtime we resched so that the active
3224 * hierarchy can be throttled
3225 */
3226 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3227 resched_task(rq_of(cfs_rq)->curr);
ec12cb7f
PT
3228}
3229
6c16a6dc 3230static __always_inline
9dbdb155 3231void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
ec12cb7f 3232{
56f570e5 3233 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
ec12cb7f
PT
3234 return;
3235
3236 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3237}
3238
85dac906
PT
3239static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3240{
56f570e5 3241 return cfs_bandwidth_used() && cfs_rq->throttled;
85dac906
PT
3242}
3243
64660c86
PT
3244/* check whether cfs_rq, or any parent, is throttled */
3245static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3246{
56f570e5 3247 return cfs_bandwidth_used() && cfs_rq->throttle_count;
64660c86
PT
3248}
3249
3250/*
3251 * Ensure that neither of the group entities corresponding to src_cpu or
3252 * dest_cpu are members of a throttled hierarchy when performing group
3253 * load-balance operations.
3254 */
3255static inline int throttled_lb_pair(struct task_group *tg,
3256 int src_cpu, int dest_cpu)
3257{
3258 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3259
3260 src_cfs_rq = tg->cfs_rq[src_cpu];
3261 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3262
3263 return throttled_hierarchy(src_cfs_rq) ||
3264 throttled_hierarchy(dest_cfs_rq);
3265}
3266
3267/* updated child weight may affect parent so we have to do this bottom up */
3268static int tg_unthrottle_up(struct task_group *tg, void *data)
3269{
3270 struct rq *rq = data;
3271 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3272
3273 cfs_rq->throttle_count--;
3274#ifdef CONFIG_SMP
3275 if (!cfs_rq->throttle_count) {
f1b17280 3276 /* adjust cfs_rq_clock_task() */
78becc27 3277 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
f1b17280 3278 cfs_rq->throttled_clock_task;
64660c86
PT
3279 }
3280#endif
3281
3282 return 0;
3283}
3284
3285static int tg_throttle_down(struct task_group *tg, void *data)
3286{
3287 struct rq *rq = data;
3288 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3289
82958366
PT
3290 /* group is entering throttled state, stop time */
3291 if (!cfs_rq->throttle_count)
78becc27 3292 cfs_rq->throttled_clock_task = rq_clock_task(rq);
64660c86
PT
3293 cfs_rq->throttle_count++;
3294
3295 return 0;
3296}
3297
d3d9dc33 3298static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
85dac906
PT
3299{
3300 struct rq *rq = rq_of(cfs_rq);
3301 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3302 struct sched_entity *se;
3303 long task_delta, dequeue = 1;
3304
3305 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3306
f1b17280 3307 /* freeze hierarchy runnable averages while throttled */
64660c86
PT
3308 rcu_read_lock();
3309 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3310 rcu_read_unlock();
85dac906
PT
3311
3312 task_delta = cfs_rq->h_nr_running;
3313 for_each_sched_entity(se) {
3314 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3315 /* throttled entity or throttle-on-deactivate */
3316 if (!se->on_rq)
3317 break;
3318
3319 if (dequeue)
3320 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3321 qcfs_rq->h_nr_running -= task_delta;
3322
3323 if (qcfs_rq->load.weight)
3324 dequeue = 0;
3325 }
3326
3327 if (!se)
3328 rq->nr_running -= task_delta;
3329
3330 cfs_rq->throttled = 1;
78becc27 3331 cfs_rq->throttled_clock = rq_clock(rq);
85dac906
PT
3332 raw_spin_lock(&cfs_b->lock);
3333 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
f9f9ffc2
BS
3334 if (!cfs_b->timer_active)
3335 __start_cfs_bandwidth(cfs_b);
85dac906
PT
3336 raw_spin_unlock(&cfs_b->lock);
3337}
3338
029632fb 3339void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
671fd9da
PT
3340{
3341 struct rq *rq = rq_of(cfs_rq);
3342 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3343 struct sched_entity *se;
3344 int enqueue = 1;
3345 long task_delta;
3346
22b958d8 3347 se = cfs_rq->tg->se[cpu_of(rq)];
671fd9da
PT
3348
3349 cfs_rq->throttled = 0;
1a55af2e
FW
3350
3351 update_rq_clock(rq);
3352
671fd9da 3353 raw_spin_lock(&cfs_b->lock);
78becc27 3354 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
671fd9da
PT
3355 list_del_rcu(&cfs_rq->throttled_list);
3356 raw_spin_unlock(&cfs_b->lock);
3357
64660c86
PT
3358 /* update hierarchical throttle state */
3359 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3360
671fd9da
PT
3361 if (!cfs_rq->load.weight)
3362 return;
3363
3364 task_delta = cfs_rq->h_nr_running;
3365 for_each_sched_entity(se) {
3366 if (se->on_rq)
3367 enqueue = 0;
3368
3369 cfs_rq = cfs_rq_of(se);
3370 if (enqueue)
3371 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3372 cfs_rq->h_nr_running += task_delta;
3373
3374 if (cfs_rq_throttled(cfs_rq))
3375 break;
3376 }
3377
3378 if (!se)
3379 rq->nr_running += task_delta;
3380
3381 /* determine whether we need to wake up potentially idle cpu */
3382 if (rq->curr == rq->idle && rq->cfs.nr_running)
3383 resched_task(rq->curr);
3384}
3385
3386static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3387 u64 remaining, u64 expires)
3388{
3389 struct cfs_rq *cfs_rq;
3390 u64 runtime = remaining;
3391
3392 rcu_read_lock();
3393 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3394 throttled_list) {
3395 struct rq *rq = rq_of(cfs_rq);
3396
3397 raw_spin_lock(&rq->lock);
3398 if (!cfs_rq_throttled(cfs_rq))
3399 goto next;
3400
3401 runtime = -cfs_rq->runtime_remaining + 1;
3402 if (runtime > remaining)
3403 runtime = remaining;
3404 remaining -= runtime;
3405
3406 cfs_rq->runtime_remaining += runtime;
3407 cfs_rq->runtime_expires = expires;
3408
3409 /* we check whether we're throttled above */
3410 if (cfs_rq->runtime_remaining > 0)
3411 unthrottle_cfs_rq(cfs_rq);
3412
3413next:
3414 raw_spin_unlock(&rq->lock);
3415
3416 if (!remaining)
3417 break;
3418 }
3419 rcu_read_unlock();
3420
3421 return remaining;
3422}
3423
58088ad0
PT
3424/*
3425 * Responsible for refilling a task_group's bandwidth and unthrottling its
3426 * cfs_rqs as appropriate. If there has been no activity within the last
3427 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3428 * used to track this state.
3429 */
3430static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3431{
671fd9da
PT
3432 u64 runtime, runtime_expires;
3433 int idle = 1, throttled;
58088ad0
PT
3434
3435 raw_spin_lock(&cfs_b->lock);
3436 /* no need to continue the timer with no bandwidth constraint */
3437 if (cfs_b->quota == RUNTIME_INF)
3438 goto out_unlock;
3439
671fd9da
PT
3440 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3441 /* idle depends on !throttled (for the case of a large deficit) */
3442 idle = cfs_b->idle && !throttled;
e8da1b18 3443 cfs_b->nr_periods += overrun;
671fd9da 3444
a9cf55b2
PT
3445 /* if we're going inactive then everything else can be deferred */
3446 if (idle)
3447 goto out_unlock;
3448
927b54fc
BS
3449 /*
3450 * if we have relooped after returning idle once, we need to update our
3451 * status as actually running, so that other cpus doing
3452 * __start_cfs_bandwidth will stop trying to cancel us.
3453 */
3454 cfs_b->timer_active = 1;
3455
a9cf55b2
PT
3456 __refill_cfs_bandwidth_runtime(cfs_b);
3457
671fd9da
PT
3458 if (!throttled) {
3459 /* mark as potentially idle for the upcoming period */
3460 cfs_b->idle = 1;
3461 goto out_unlock;
3462 }
3463
e8da1b18
NR
3464 /* account preceding periods in which throttling occurred */
3465 cfs_b->nr_throttled += overrun;
3466
671fd9da
PT
3467 /*
3468 * There are throttled entities so we must first use the new bandwidth
3469 * to unthrottle them before making it generally available. This
3470 * ensures that all existing debts will be paid before a new cfs_rq is
3471 * allowed to run.
3472 */
3473 runtime = cfs_b->runtime;
3474 runtime_expires = cfs_b->runtime_expires;
3475 cfs_b->runtime = 0;
3476
3477 /*
3478 * This check is repeated as we are holding onto the new bandwidth
3479 * while we unthrottle. This can potentially race with an unthrottled
3480 * group trying to acquire new bandwidth from the global pool.
3481 */
3482 while (throttled && runtime > 0) {
3483 raw_spin_unlock(&cfs_b->lock);
3484 /* we can't nest cfs_b->lock while distributing bandwidth */
3485 runtime = distribute_cfs_runtime(cfs_b, runtime,
3486 runtime_expires);
3487 raw_spin_lock(&cfs_b->lock);
3488
3489 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3490 }
58088ad0 3491
671fd9da
PT
3492 /* return (any) remaining runtime */
3493 cfs_b->runtime = runtime;
3494 /*
3495 * While we are ensured activity in the period following an
3496 * unthrottle, this also covers the case in which the new bandwidth is
3497 * insufficient to cover the existing bandwidth deficit. (Forcing the
3498 * timer to remain active while there are any throttled entities.)
3499 */
3500 cfs_b->idle = 0;
58088ad0
PT
3501out_unlock:
3502 if (idle)
3503 cfs_b->timer_active = 0;
3504 raw_spin_unlock(&cfs_b->lock);
3505
3506 return idle;
3507}
d3d9dc33 3508
d8b4986d
PT
3509/* a cfs_rq won't donate quota below this amount */
3510static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3511/* minimum remaining period time to redistribute slack quota */
3512static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3513/* how long we wait to gather additional slack before distributing */
3514static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3515
db06e78c
BS
3516/*
3517 * Are we near the end of the current quota period?
3518 *
3519 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3520 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3521 * migrate_hrtimers, base is never cleared, so we are fine.
3522 */
d8b4986d
PT
3523static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3524{
3525 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3526 u64 remaining;
3527
3528 /* if the call-back is running a quota refresh is already occurring */
3529 if (hrtimer_callback_running(refresh_timer))
3530 return 1;
3531
3532 /* is a quota refresh about to occur? */
3533 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3534 if (remaining < min_expire)
3535 return 1;
3536
3537 return 0;
3538}
3539
3540static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3541{
3542 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3543
3544 /* if there's a quota refresh soon don't bother with slack */
3545 if (runtime_refresh_within(cfs_b, min_left))
3546 return;
3547
3548 start_bandwidth_timer(&cfs_b->slack_timer,
3549 ns_to_ktime(cfs_bandwidth_slack_period));
3550}
3551
3552/* we know any runtime found here is valid as update_curr() precedes return */
3553static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3554{
3555 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3556 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3557
3558 if (slack_runtime <= 0)
3559 return;
3560
3561 raw_spin_lock(&cfs_b->lock);
3562 if (cfs_b->quota != RUNTIME_INF &&
3563 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3564 cfs_b->runtime += slack_runtime;
3565
3566 /* we are under rq->lock, defer unthrottling using a timer */
3567 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3568 !list_empty(&cfs_b->throttled_cfs_rq))
3569 start_cfs_slack_bandwidth(cfs_b);
3570 }
3571 raw_spin_unlock(&cfs_b->lock);
3572
3573 /* even if it's not valid for return we don't want to try again */
3574 cfs_rq->runtime_remaining -= slack_runtime;
3575}
3576
3577static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3578{
56f570e5
PT
3579 if (!cfs_bandwidth_used())
3580 return;
3581
fccfdc6f 3582 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
d8b4986d
PT
3583 return;
3584
3585 __return_cfs_rq_runtime(cfs_rq);
3586}
3587
3588/*
3589 * This is done with a timer (instead of inline with bandwidth return) since
3590 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3591 */
3592static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3593{
3594 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3595 u64 expires;
3596
3597 /* confirm we're still not at a refresh boundary */
db06e78c
BS
3598 raw_spin_lock(&cfs_b->lock);
3599 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3600 raw_spin_unlock(&cfs_b->lock);
d8b4986d 3601 return;
db06e78c 3602 }
d8b4986d 3603
d8b4986d
PT
3604 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3605 runtime = cfs_b->runtime;
3606 cfs_b->runtime = 0;
3607 }
3608 expires = cfs_b->runtime_expires;
3609 raw_spin_unlock(&cfs_b->lock);
3610
3611 if (!runtime)
3612 return;
3613
3614 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3615
3616 raw_spin_lock(&cfs_b->lock);
3617 if (expires == cfs_b->runtime_expires)
3618 cfs_b->runtime = runtime;
3619 raw_spin_unlock(&cfs_b->lock);
3620}
3621
d3d9dc33
PT
3622/*
3623 * When a group wakes up we want to make sure that its quota is not already
3624 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3625 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3626 */
3627static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3628{
56f570e5
PT
3629 if (!cfs_bandwidth_used())
3630 return;
3631
d3d9dc33
PT
3632 /* an active group must be handled by the update_curr()->put() path */
3633 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3634 return;
3635
3636 /* ensure the group is not already throttled */
3637 if (cfs_rq_throttled(cfs_rq))
3638 return;
3639
3640 /* update runtime allocation */
3641 account_cfs_rq_runtime(cfs_rq, 0);
3642 if (cfs_rq->runtime_remaining <= 0)
3643 throttle_cfs_rq(cfs_rq);
3644}
3645
3646/* conditionally throttle active cfs_rq's from put_prev_entity() */
678d5718 3647static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
d3d9dc33 3648{
56f570e5 3649 if (!cfs_bandwidth_used())
678d5718 3650 return false;
56f570e5 3651
d3d9dc33 3652 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
678d5718 3653 return false;
d3d9dc33
PT
3654
3655 /*
3656 * it's possible for a throttled entity to be forced into a running
3657 * state (e.g. set_curr_task), in this case we're finished.
3658 */
3659 if (cfs_rq_throttled(cfs_rq))
678d5718 3660 return true;
d3d9dc33
PT
3661
3662 throttle_cfs_rq(cfs_rq);
678d5718 3663 return true;
d3d9dc33 3664}
029632fb 3665
029632fb
PZ
3666static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3667{
3668 struct cfs_bandwidth *cfs_b =
3669 container_of(timer, struct cfs_bandwidth, slack_timer);
3670 do_sched_cfs_slack_timer(cfs_b);
3671
3672 return HRTIMER_NORESTART;
3673}
3674
3675static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3676{
3677 struct cfs_bandwidth *cfs_b =
3678 container_of(timer, struct cfs_bandwidth, period_timer);
3679 ktime_t now;
3680 int overrun;
3681 int idle = 0;
3682
3683 for (;;) {
3684 now = hrtimer_cb_get_time(timer);
3685 overrun = hrtimer_forward(timer, now, cfs_b->period);
3686
3687 if (!overrun)
3688 break;
3689
3690 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3691 }
3692
3693 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3694}
3695
3696void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3697{
3698 raw_spin_lock_init(&cfs_b->lock);
3699 cfs_b->runtime = 0;
3700 cfs_b->quota = RUNTIME_INF;
3701 cfs_b->period = ns_to_ktime(default_cfs_period());
3702
3703 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3704 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3705 cfs_b->period_timer.function = sched_cfs_period_timer;
3706 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3707 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3708}
3709
3710static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3711{
3712 cfs_rq->runtime_enabled = 0;
3713 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3714}
3715
3716/* requires cfs_b->lock, may release to reprogram timer */
3717void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3718{
3719 /*
3720 * The timer may be active because we're trying to set a new bandwidth
3721 * period or because we're racing with the tear-down path
3722 * (timer_active==0 becomes visible before the hrtimer call-back
3723 * terminates). In either case we ensure that it's re-programmed
3724 */
927b54fc
BS
3725 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3726 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3727 /* bounce the lock to allow do_sched_cfs_period_timer to run */
029632fb 3728 raw_spin_unlock(&cfs_b->lock);
927b54fc 3729 cpu_relax();
029632fb
PZ
3730 raw_spin_lock(&cfs_b->lock);
3731 /* if someone else restarted the timer then we're done */
3732 if (cfs_b->timer_active)
3733 return;
3734 }
3735
3736 cfs_b->timer_active = 1;
3737 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3738}
3739
3740static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3741{
3742 hrtimer_cancel(&cfs_b->period_timer);
3743 hrtimer_cancel(&cfs_b->slack_timer);
3744}
3745
38dc3348 3746static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
029632fb
PZ
3747{
3748 struct cfs_rq *cfs_rq;
3749
3750 for_each_leaf_cfs_rq(rq, cfs_rq) {
3751 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3752
3753 if (!cfs_rq->runtime_enabled)
3754 continue;
3755
3756 /*
3757 * clock_task is not advancing so we just need to make sure
3758 * there's some valid quota amount
3759 */
3760 cfs_rq->runtime_remaining = cfs_b->quota;
3761 if (cfs_rq_throttled(cfs_rq))
3762 unthrottle_cfs_rq(cfs_rq);
3763 }
3764}
3765
3766#else /* CONFIG_CFS_BANDWIDTH */
f1b17280
PT
3767static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3768{
78becc27 3769 return rq_clock_task(rq_of(cfs_rq));
f1b17280
PT
3770}
3771
9dbdb155 3772static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
678d5718 3773static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
d3d9dc33 3774static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6c16a6dc 3775static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
85dac906
PT
3776
3777static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3778{
3779 return 0;
3780}
64660c86
PT
3781
3782static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3783{
3784 return 0;
3785}
3786
3787static inline int throttled_lb_pair(struct task_group *tg,
3788 int src_cpu, int dest_cpu)
3789{
3790 return 0;
3791}
029632fb
PZ
3792
3793void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3794
3795#ifdef CONFIG_FAIR_GROUP_SCHED
3796static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
ab84d31e
PT
3797#endif
3798
029632fb
PZ
3799static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3800{
3801 return NULL;
3802}
3803static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
a4c96ae3 3804static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
029632fb
PZ
3805
3806#endif /* CONFIG_CFS_BANDWIDTH */
3807
bf0f6f24
IM
3808/**************************************************
3809 * CFS operations on tasks:
3810 */
3811
8f4d37ec
PZ
3812#ifdef CONFIG_SCHED_HRTICK
3813static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3814{
8f4d37ec
PZ
3815 struct sched_entity *se = &p->se;
3816 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3817
3818 WARN_ON(task_rq(p) != rq);
3819
b39e66ea 3820 if (cfs_rq->nr_running > 1) {
8f4d37ec
PZ
3821 u64 slice = sched_slice(cfs_rq, se);
3822 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3823 s64 delta = slice - ran;
3824
3825 if (delta < 0) {
3826 if (rq->curr == p)
3827 resched_task(p);
3828 return;
3829 }
3830
3831 /*
3832 * Don't schedule slices shorter than 10000ns, that just
3833 * doesn't make sense. Rely on vruntime for fairness.
3834 */
31656519 3835 if (rq->curr != p)
157124c1 3836 delta = max_t(s64, 10000LL, delta);
8f4d37ec 3837
31656519 3838 hrtick_start(rq, delta);
8f4d37ec
PZ
3839 }
3840}
a4c2f00f
PZ
3841
3842/*
3843 * called from enqueue/dequeue and updates the hrtick when the
3844 * current task is from our class and nr_running is low enough
3845 * to matter.
3846 */
3847static void hrtick_update(struct rq *rq)
3848{
3849 struct task_struct *curr = rq->curr;
3850
b39e66ea 3851 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
a4c2f00f
PZ
3852 return;
3853
3854 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3855 hrtick_start_fair(rq, curr);
3856}
55e12e5e 3857#else /* !CONFIG_SCHED_HRTICK */
8f4d37ec
PZ
3858static inline void
3859hrtick_start_fair(struct rq *rq, struct task_struct *p)
3860{
3861}
a4c2f00f
PZ
3862
3863static inline void hrtick_update(struct rq *rq)
3864{
3865}
8f4d37ec
PZ
3866#endif
3867
bf0f6f24
IM
3868/*
3869 * The enqueue_task method is called before nr_running is
3870 * increased. Here we update the fair scheduling stats and
3871 * then put the task into the rbtree:
3872 */
ea87bb78 3873static void
371fd7e7 3874enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
bf0f6f24
IM
3875{
3876 struct cfs_rq *cfs_rq;
62fb1851 3877 struct sched_entity *se = &p->se;
bf0f6f24
IM
3878
3879 for_each_sched_entity(se) {
62fb1851 3880 if (se->on_rq)
bf0f6f24
IM
3881 break;
3882 cfs_rq = cfs_rq_of(se);
88ec22d3 3883 enqueue_entity(cfs_rq, se, flags);
85dac906
PT
3884
3885 /*
3886 * end evaluation on encountering a throttled cfs_rq
3887 *
3888 * note: in the case of encountering a throttled cfs_rq we will
3889 * post the final h_nr_running increment below.
3890 */
3891 if (cfs_rq_throttled(cfs_rq))
3892 break;
953bfcd1 3893 cfs_rq->h_nr_running++;
85dac906 3894
88ec22d3 3895 flags = ENQUEUE_WAKEUP;
bf0f6f24 3896 }
8f4d37ec 3897
2069dd75 3898 for_each_sched_entity(se) {
0f317143 3899 cfs_rq = cfs_rq_of(se);
953bfcd1 3900 cfs_rq->h_nr_running++;
2069dd75 3901
85dac906
PT
3902 if (cfs_rq_throttled(cfs_rq))
3903 break;
3904
17bc14b7 3905 update_cfs_shares(cfs_rq);
9ee474f5 3906 update_entity_load_avg(se, 1);
2069dd75
PZ
3907 }
3908
18bf2805
BS
3909 if (!se) {
3910 update_rq_runnable_avg(rq, rq->nr_running);
85dac906 3911 inc_nr_running(rq);
18bf2805 3912 }
a4c2f00f 3913 hrtick_update(rq);
bf0f6f24
IM
3914}
3915
2f36825b
VP
3916static void set_next_buddy(struct sched_entity *se);
3917
bf0f6f24
IM
3918/*
3919 * The dequeue_task method is called before nr_running is
3920 * decreased. We remove the task from the rbtree and
3921 * update the fair scheduling stats:
3922 */
371fd7e7 3923static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
bf0f6f24
IM
3924{
3925 struct cfs_rq *cfs_rq;
62fb1851 3926 struct sched_entity *se = &p->se;
2f36825b 3927 int task_sleep = flags & DEQUEUE_SLEEP;
bf0f6f24
IM
3928
3929 for_each_sched_entity(se) {
3930 cfs_rq = cfs_rq_of(se);
371fd7e7 3931 dequeue_entity(cfs_rq, se, flags);
85dac906
PT
3932
3933 /*
3934 * end evaluation on encountering a throttled cfs_rq
3935 *
3936 * note: in the case of encountering a throttled cfs_rq we will
3937 * post the final h_nr_running decrement below.
3938 */
3939 if (cfs_rq_throttled(cfs_rq))
3940 break;
953bfcd1 3941 cfs_rq->h_nr_running--;
2069dd75 3942
bf0f6f24 3943 /* Don't dequeue parent if it has other entities besides us */
2f36825b
VP
3944 if (cfs_rq->load.weight) {
3945 /*
3946 * Bias pick_next to pick a task from this cfs_rq, as
3947 * p is sleeping when it is within its sched_slice.
3948 */
3949 if (task_sleep && parent_entity(se))
3950 set_next_buddy(parent_entity(se));
9598c82d
PT
3951
3952 /* avoid re-evaluating load for this entity */
3953 se = parent_entity(se);
bf0f6f24 3954 break;
2f36825b 3955 }
371fd7e7 3956 flags |= DEQUEUE_SLEEP;
bf0f6f24 3957 }
8f4d37ec 3958
2069dd75 3959 for_each_sched_entity(se) {
0f317143 3960 cfs_rq = cfs_rq_of(se);
953bfcd1 3961 cfs_rq->h_nr_running--;
2069dd75 3962
85dac906
PT
3963 if (cfs_rq_throttled(cfs_rq))
3964 break;
3965
17bc14b7 3966 update_cfs_shares(cfs_rq);
9ee474f5 3967 update_entity_load_avg(se, 1);
2069dd75
PZ
3968 }
3969
18bf2805 3970 if (!se) {
85dac906 3971 dec_nr_running(rq);
18bf2805
BS
3972 update_rq_runnable_avg(rq, 1);
3973 }
a4c2f00f 3974 hrtick_update(rq);
bf0f6f24
IM
3975}
3976
e7693a36 3977#ifdef CONFIG_SMP
029632fb
PZ
3978/* Used instead of source_load when we know the type == 0 */
3979static unsigned long weighted_cpuload(const int cpu)
3980{
b92486cb 3981 return cpu_rq(cpu)->cfs.runnable_load_avg;
029632fb
PZ
3982}
3983
3984/*
3985 * Return a low guess at the load of a migration-source cpu weighted
3986 * according to the scheduling class and "nice" value.
3987 *
3988 * We want to under-estimate the load of migration sources, to
3989 * balance conservatively.
3990 */
3991static unsigned long source_load(int cpu, int type)
3992{
3993 struct rq *rq = cpu_rq(cpu);
3994 unsigned long total = weighted_cpuload(cpu);
3995
3996 if (type == 0 || !sched_feat(LB_BIAS))
3997 return total;
3998
3999 return min(rq->cpu_load[type-1], total);
4000}
4001
4002/*
4003 * Return a high guess at the load of a migration-target cpu weighted
4004 * according to the scheduling class and "nice" value.
4005 */
4006static unsigned long target_load(int cpu, int type)
4007{
4008 struct rq *rq = cpu_rq(cpu);
4009 unsigned long total = weighted_cpuload(cpu);
4010
4011 if (type == 0 || !sched_feat(LB_BIAS))
4012 return total;
4013
4014 return max(rq->cpu_load[type-1], total);
4015}
4016
4017static unsigned long power_of(int cpu)
4018{
4019 return cpu_rq(cpu)->cpu_power;
4020}
4021
4022static unsigned long cpu_avg_load_per_task(int cpu)
4023{
4024 struct rq *rq = cpu_rq(cpu);
4025 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
b92486cb 4026 unsigned long load_avg = rq->cfs.runnable_load_avg;
029632fb
PZ
4027
4028 if (nr_running)
b92486cb 4029 return load_avg / nr_running;
029632fb
PZ
4030
4031 return 0;
4032}
4033
62470419
MW
4034static void record_wakee(struct task_struct *p)
4035{
4036 /*
4037 * Rough decay (wiping) for cost saving, don't worry
4038 * about the boundary, really active task won't care
4039 * about the loss.
4040 */
4041 if (jiffies > current->wakee_flip_decay_ts + HZ) {
4042 current->wakee_flips = 0;
4043 current->wakee_flip_decay_ts = jiffies;
4044 }
4045
4046 if (current->last_wakee != p) {
4047 current->last_wakee = p;
4048 current->wakee_flips++;
4049 }
4050}
098fb9db 4051
74f8e4b2 4052static void task_waking_fair(struct task_struct *p)
88ec22d3
PZ
4053{
4054 struct sched_entity *se = &p->se;
4055 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3fe1698b
PZ
4056 u64 min_vruntime;
4057
4058#ifndef CONFIG_64BIT
4059 u64 min_vruntime_copy;
88ec22d3 4060
3fe1698b
PZ
4061 do {
4062 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4063 smp_rmb();
4064 min_vruntime = cfs_rq->min_vruntime;
4065 } while (min_vruntime != min_vruntime_copy);
4066#else
4067 min_vruntime = cfs_rq->min_vruntime;
4068#endif
88ec22d3 4069
3fe1698b 4070 se->vruntime -= min_vruntime;
62470419 4071 record_wakee(p);
88ec22d3
PZ
4072}
4073
bb3469ac 4074#ifdef CONFIG_FAIR_GROUP_SCHED
f5bfb7d9
PZ
4075/*
4076 * effective_load() calculates the load change as seen from the root_task_group
4077 *
4078 * Adding load to a group doesn't make a group heavier, but can cause movement
4079 * of group shares between cpus. Assuming the shares were perfectly aligned one
4080 * can calculate the shift in shares.
cf5f0acf
PZ
4081 *
4082 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4083 * on this @cpu and results in a total addition (subtraction) of @wg to the
4084 * total group weight.
4085 *
4086 * Given a runqueue weight distribution (rw_i) we can compute a shares
4087 * distribution (s_i) using:
4088 *
4089 * s_i = rw_i / \Sum rw_j (1)
4090 *
4091 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4092 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4093 * shares distribution (s_i):
4094 *
4095 * rw_i = { 2, 4, 1, 0 }
4096 * s_i = { 2/7, 4/7, 1/7, 0 }
4097 *
4098 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4099 * task used to run on and the CPU the waker is running on), we need to
4100 * compute the effect of waking a task on either CPU and, in case of a sync
4101 * wakeup, compute the effect of the current task going to sleep.
4102 *
4103 * So for a change of @wl to the local @cpu with an overall group weight change
4104 * of @wl we can compute the new shares distribution (s'_i) using:
4105 *
4106 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4107 *
4108 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4109 * differences in waking a task to CPU 0. The additional task changes the
4110 * weight and shares distributions like:
4111 *
4112 * rw'_i = { 3, 4, 1, 0 }
4113 * s'_i = { 3/8, 4/8, 1/8, 0 }
4114 *
4115 * We can then compute the difference in effective weight by using:
4116 *
4117 * dw_i = S * (s'_i - s_i) (3)
4118 *
4119 * Where 'S' is the group weight as seen by its parent.
4120 *
4121 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4122 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4123 * 4/7) times the weight of the group.
f5bfb7d9 4124 */
2069dd75 4125static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
bb3469ac 4126{
4be9daaa 4127 struct sched_entity *se = tg->se[cpu];
f1d239f7 4128
9722c2da 4129 if (!tg->parent) /* the trivial, non-cgroup case */
f1d239f7
PZ
4130 return wl;
4131
4be9daaa 4132 for_each_sched_entity(se) {
cf5f0acf 4133 long w, W;
4be9daaa 4134
977dda7c 4135 tg = se->my_q->tg;
bb3469ac 4136
cf5f0acf
PZ
4137 /*
4138 * W = @wg + \Sum rw_j
4139 */
4140 W = wg + calc_tg_weight(tg, se->my_q);
4be9daaa 4141
cf5f0acf
PZ
4142 /*
4143 * w = rw_i + @wl
4144 */
4145 w = se->my_q->load.weight + wl;
940959e9 4146
cf5f0acf
PZ
4147 /*
4148 * wl = S * s'_i; see (2)
4149 */
4150 if (W > 0 && w < W)
4151 wl = (w * tg->shares) / W;
977dda7c
PT
4152 else
4153 wl = tg->shares;
940959e9 4154
cf5f0acf
PZ
4155 /*
4156 * Per the above, wl is the new se->load.weight value; since
4157 * those are clipped to [MIN_SHARES, ...) do so now. See
4158 * calc_cfs_shares().
4159 */
977dda7c
PT
4160 if (wl < MIN_SHARES)
4161 wl = MIN_SHARES;
cf5f0acf
PZ
4162
4163 /*
4164 * wl = dw_i = S * (s'_i - s_i); see (3)
4165 */
977dda7c 4166 wl -= se->load.weight;
cf5f0acf
PZ
4167
4168 /*
4169 * Recursively apply this logic to all parent groups to compute
4170 * the final effective load change on the root group. Since
4171 * only the @tg group gets extra weight, all parent groups can
4172 * only redistribute existing shares. @wl is the shift in shares
4173 * resulting from this level per the above.
4174 */
4be9daaa 4175 wg = 0;
4be9daaa 4176 }
bb3469ac 4177
4be9daaa 4178 return wl;
bb3469ac
PZ
4179}
4180#else
4be9daaa 4181
58d081b5 4182static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4be9daaa 4183{
83378269 4184 return wl;
bb3469ac 4185}
4be9daaa 4186
bb3469ac
PZ
4187#endif
4188
62470419
MW
4189static int wake_wide(struct task_struct *p)
4190{
7d9ffa89 4191 int factor = this_cpu_read(sd_llc_size);
62470419
MW
4192
4193 /*
4194 * Yeah, it's the switching-frequency, could means many wakee or
4195 * rapidly switch, use factor here will just help to automatically
4196 * adjust the loose-degree, so bigger node will lead to more pull.
4197 */
4198 if (p->wakee_flips > factor) {
4199 /*
4200 * wakee is somewhat hot, it needs certain amount of cpu
4201 * resource, so if waker is far more hot, prefer to leave
4202 * it alone.
4203 */
4204 if (current->wakee_flips > (factor * p->wakee_flips))
4205 return 1;
4206 }
4207
4208 return 0;
4209}
4210
c88d5910 4211static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
098fb9db 4212{
e37b6a7b 4213 s64 this_load, load;
c88d5910 4214 int idx, this_cpu, prev_cpu;
098fb9db 4215 unsigned long tl_per_task;
c88d5910 4216 struct task_group *tg;
83378269 4217 unsigned long weight;
b3137bc8 4218 int balanced;
098fb9db 4219
62470419
MW
4220 /*
4221 * If we wake multiple tasks be careful to not bounce
4222 * ourselves around too much.
4223 */
4224 if (wake_wide(p))
4225 return 0;
4226
c88d5910
PZ
4227 idx = sd->wake_idx;
4228 this_cpu = smp_processor_id();
4229 prev_cpu = task_cpu(p);
4230 load = source_load(prev_cpu, idx);
4231 this_load = target_load(this_cpu, idx);
098fb9db 4232
b3137bc8
MG
4233 /*
4234 * If sync wakeup then subtract the (maximum possible)
4235 * effect of the currently running task from the load
4236 * of the current CPU:
4237 */
83378269
PZ
4238 if (sync) {
4239 tg = task_group(current);
4240 weight = current->se.load.weight;
4241
c88d5910 4242 this_load += effective_load(tg, this_cpu, -weight, -weight);
83378269
PZ
4243 load += effective_load(tg, prev_cpu, 0, -weight);
4244 }
b3137bc8 4245
83378269
PZ
4246 tg = task_group(p);
4247 weight = p->se.load.weight;
b3137bc8 4248
71a29aa7
PZ
4249 /*
4250 * In low-load situations, where prev_cpu is idle and this_cpu is idle
c88d5910
PZ
4251 * due to the sync cause above having dropped this_load to 0, we'll
4252 * always have an imbalance, but there's really nothing you can do
4253 * about that, so that's good too.
71a29aa7
PZ
4254 *
4255 * Otherwise check if either cpus are near enough in load to allow this
4256 * task to be woken on this_cpu.
4257 */
e37b6a7b
PT
4258 if (this_load > 0) {
4259 s64 this_eff_load, prev_eff_load;
e51fd5e2
PZ
4260
4261 this_eff_load = 100;
4262 this_eff_load *= power_of(prev_cpu);
4263 this_eff_load *= this_load +
4264 effective_load(tg, this_cpu, weight, weight);
4265
4266 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4267 prev_eff_load *= power_of(this_cpu);
4268 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4269
4270 balanced = this_eff_load <= prev_eff_load;
4271 } else
4272 balanced = true;
b3137bc8 4273
098fb9db 4274 /*
4ae7d5ce
IM
4275 * If the currently running task will sleep within
4276 * a reasonable amount of time then attract this newly
4277 * woken task:
098fb9db 4278 */
2fb7635c
PZ
4279 if (sync && balanced)
4280 return 1;
098fb9db 4281
41acab88 4282 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
098fb9db
IM
4283 tl_per_task = cpu_avg_load_per_task(this_cpu);
4284
c88d5910
PZ
4285 if (balanced ||
4286 (this_load <= load &&
4287 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
098fb9db
IM
4288 /*
4289 * This domain has SD_WAKE_AFFINE and
4290 * p is cache cold in this domain, and
4291 * there is no bad imbalance.
4292 */
c88d5910 4293 schedstat_inc(sd, ttwu_move_affine);
41acab88 4294 schedstat_inc(p, se.statistics.nr_wakeups_affine);
098fb9db
IM
4295
4296 return 1;
4297 }
4298 return 0;
4299}
4300
aaee1203
PZ
4301/*
4302 * find_idlest_group finds and returns the least busy CPU group within the
4303 * domain.
4304 */
4305static struct sched_group *
78e7ed53 4306find_idlest_group(struct sched_domain *sd, struct task_struct *p,
c44f2a02 4307 int this_cpu, int sd_flag)
e7693a36 4308{
b3bd3de6 4309 struct sched_group *idlest = NULL, *group = sd->groups;
aaee1203 4310 unsigned long min_load = ULONG_MAX, this_load = 0;
c44f2a02 4311 int load_idx = sd->forkexec_idx;
aaee1203 4312 int imbalance = 100 + (sd->imbalance_pct-100)/2;
e7693a36 4313
c44f2a02
VG
4314 if (sd_flag & SD_BALANCE_WAKE)
4315 load_idx = sd->wake_idx;
4316
aaee1203
PZ
4317 do {
4318 unsigned long load, avg_load;
4319 int local_group;
4320 int i;
e7693a36 4321
aaee1203
PZ
4322 /* Skip over this group if it has no CPUs allowed */
4323 if (!cpumask_intersects(sched_group_cpus(group),
fa17b507 4324 tsk_cpus_allowed(p)))
aaee1203
PZ
4325 continue;
4326
4327 local_group = cpumask_test_cpu(this_cpu,
4328 sched_group_cpus(group));
4329
4330 /* Tally up the load of all CPUs in the group */
4331 avg_load = 0;
4332
4333 for_each_cpu(i, sched_group_cpus(group)) {
4334 /* Bias balancing toward cpus of our domain */
4335 if (local_group)
4336 load = source_load(i, load_idx);
4337 else
4338 load = target_load(i, load_idx);
4339
4340 avg_load += load;
4341 }
4342
4343 /* Adjust by relative CPU power of the group */
9c3f75cb 4344 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
aaee1203
PZ
4345
4346 if (local_group) {
4347 this_load = avg_load;
aaee1203
PZ
4348 } else if (avg_load < min_load) {
4349 min_load = avg_load;
4350 idlest = group;
4351 }
4352 } while (group = group->next, group != sd->groups);
4353
4354 if (!idlest || 100*this_load < imbalance*min_load)
4355 return NULL;
4356 return idlest;
4357}
4358
4359/*
4360 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4361 */
4362static int
4363find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4364{
4365 unsigned long load, min_load = ULONG_MAX;
4366 int idlest = -1;
4367 int i;
4368
4369 /* Traverse only the allowed CPUs */
fa17b507 4370 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
aaee1203
PZ
4371 load = weighted_cpuload(i);
4372
4373 if (load < min_load || (load == min_load && i == this_cpu)) {
4374 min_load = load;
4375 idlest = i;
e7693a36
GH
4376 }
4377 }
4378
aaee1203
PZ
4379 return idlest;
4380}
e7693a36 4381
a50bde51
PZ
4382/*
4383 * Try and locate an idle CPU in the sched_domain.
4384 */
99bd5e2f 4385static int select_idle_sibling(struct task_struct *p, int target)
a50bde51 4386{
99bd5e2f 4387 struct sched_domain *sd;
37407ea7 4388 struct sched_group *sg;
e0a79f52 4389 int i = task_cpu(p);
a50bde51 4390
e0a79f52
MG
4391 if (idle_cpu(target))
4392 return target;
99bd5e2f
SS
4393
4394 /*
e0a79f52 4395 * If the prevous cpu is cache affine and idle, don't be stupid.
99bd5e2f 4396 */
e0a79f52
MG
4397 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4398 return i;
a50bde51
PZ
4399
4400 /*
37407ea7 4401 * Otherwise, iterate the domains and find an elegible idle cpu.
a50bde51 4402 */
518cd623 4403 sd = rcu_dereference(per_cpu(sd_llc, target));
970e1789 4404 for_each_lower_domain(sd) {
37407ea7
LT
4405 sg = sd->groups;
4406 do {
4407 if (!cpumask_intersects(sched_group_cpus(sg),
4408 tsk_cpus_allowed(p)))
4409 goto next;
4410
4411 for_each_cpu(i, sched_group_cpus(sg)) {
e0a79f52 4412 if (i == target || !idle_cpu(i))
37407ea7
LT
4413 goto next;
4414 }
970e1789 4415
37407ea7
LT
4416 target = cpumask_first_and(sched_group_cpus(sg),
4417 tsk_cpus_allowed(p));
4418 goto done;
4419next:
4420 sg = sg->next;
4421 } while (sg != sd->groups);
4422 }
4423done:
a50bde51
PZ
4424 return target;
4425}
4426
aaee1203 4427/*
de91b9cb
MR
4428 * select_task_rq_fair: Select target runqueue for the waking task in domains
4429 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4430 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
aaee1203 4431 *
de91b9cb
MR
4432 * Balances load by selecting the idlest cpu in the idlest group, or under
4433 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
aaee1203 4434 *
de91b9cb 4435 * Returns the target cpu number.
aaee1203
PZ
4436 *
4437 * preempt must be disabled.
4438 */
0017d735 4439static int
ac66f547 4440select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
aaee1203 4441{
29cd8bae 4442 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
c88d5910 4443 int cpu = smp_processor_id();
c88d5910 4444 int new_cpu = cpu;
99bd5e2f 4445 int want_affine = 0;
5158f4e4 4446 int sync = wake_flags & WF_SYNC;
c88d5910 4447
29baa747 4448 if (p->nr_cpus_allowed == 1)
76854c7e
MG
4449 return prev_cpu;
4450
0763a660 4451 if (sd_flag & SD_BALANCE_WAKE) {
fa17b507 4452 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
c88d5910
PZ
4453 want_affine = 1;
4454 new_cpu = prev_cpu;
4455 }
aaee1203 4456
dce840a0 4457 rcu_read_lock();
aaee1203 4458 for_each_domain(cpu, tmp) {
e4f42888
PZ
4459 if (!(tmp->flags & SD_LOAD_BALANCE))
4460 continue;
4461
fe3bcfe1 4462 /*
99bd5e2f
SS
4463 * If both cpu and prev_cpu are part of this domain,
4464 * cpu is a valid SD_WAKE_AFFINE target.
fe3bcfe1 4465 */
99bd5e2f
SS
4466 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4467 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4468 affine_sd = tmp;
29cd8bae 4469 break;
f03542a7 4470 }
29cd8bae 4471
f03542a7 4472 if (tmp->flags & sd_flag)
29cd8bae
PZ
4473 sd = tmp;
4474 }
4475
8bf21433
RR
4476 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4477 prev_cpu = cpu;
dce840a0 4478
8bf21433 4479 if (sd_flag & SD_BALANCE_WAKE) {
dce840a0
PZ
4480 new_cpu = select_idle_sibling(p, prev_cpu);
4481 goto unlock;
8b911acd 4482 }
e7693a36 4483
aaee1203
PZ
4484 while (sd) {
4485 struct sched_group *group;
c88d5910 4486 int weight;
098fb9db 4487
0763a660 4488 if (!(sd->flags & sd_flag)) {
aaee1203
PZ
4489 sd = sd->child;
4490 continue;
4491 }
098fb9db 4492
c44f2a02 4493 group = find_idlest_group(sd, p, cpu, sd_flag);
aaee1203
PZ
4494 if (!group) {
4495 sd = sd->child;
4496 continue;
4497 }
4ae7d5ce 4498
d7c33c49 4499 new_cpu = find_idlest_cpu(group, p, cpu);
aaee1203
PZ
4500 if (new_cpu == -1 || new_cpu == cpu) {
4501 /* Now try balancing at a lower domain level of cpu */
4502 sd = sd->child;
4503 continue;
e7693a36 4504 }
aaee1203
PZ
4505
4506 /* Now try balancing at a lower domain level of new_cpu */
4507 cpu = new_cpu;
669c55e9 4508 weight = sd->span_weight;
aaee1203
PZ
4509 sd = NULL;
4510 for_each_domain(cpu, tmp) {
669c55e9 4511 if (weight <= tmp->span_weight)
aaee1203 4512 break;
0763a660 4513 if (tmp->flags & sd_flag)
aaee1203
PZ
4514 sd = tmp;
4515 }
4516 /* while loop will break here if sd == NULL */
e7693a36 4517 }
dce840a0
PZ
4518unlock:
4519 rcu_read_unlock();
e7693a36 4520
c88d5910 4521 return new_cpu;
e7693a36 4522}
0a74bef8
PT
4523
4524/*
4525 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4526 * cfs_rq_of(p) references at time of call are still valid and identify the
4527 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4528 * other assumptions, including the state of rq->lock, should be made.
4529 */
4530static void
4531migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4532{
aff3e498
PT
4533 struct sched_entity *se = &p->se;
4534 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4535
4536 /*
4537 * Load tracking: accumulate removed load so that it can be processed
4538 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4539 * to blocked load iff they have a positive decay-count. It can never
4540 * be negative here since on-rq tasks have decay-count == 0.
4541 */
4542 if (se->avg.decay_count) {
4543 se->avg.decay_count = -__synchronize_entity_decay(se);
2509940f
AS
4544 atomic_long_add(se->avg.load_avg_contrib,
4545 &cfs_rq->removed_load);
aff3e498 4546 }
3944a927
BS
4547
4548 /* We have migrated, no longer consider this task hot */
4549 se->exec_start = 0;
0a74bef8 4550}
e7693a36
GH
4551#endif /* CONFIG_SMP */
4552
e52fb7c0
PZ
4553static unsigned long
4554wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
0bbd3336
PZ
4555{
4556 unsigned long gran = sysctl_sched_wakeup_granularity;
4557
4558 /*
e52fb7c0
PZ
4559 * Since its curr running now, convert the gran from real-time
4560 * to virtual-time in his units.
13814d42
MG
4561 *
4562 * By using 'se' instead of 'curr' we penalize light tasks, so
4563 * they get preempted easier. That is, if 'se' < 'curr' then
4564 * the resulting gran will be larger, therefore penalizing the
4565 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4566 * be smaller, again penalizing the lighter task.
4567 *
4568 * This is especially important for buddies when the leftmost
4569 * task is higher priority than the buddy.
0bbd3336 4570 */
f4ad9bd2 4571 return calc_delta_fair(gran, se);
0bbd3336
PZ
4572}
4573
464b7527
PZ
4574/*
4575 * Should 'se' preempt 'curr'.
4576 *
4577 * |s1
4578 * |s2
4579 * |s3
4580 * g
4581 * |<--->|c
4582 *
4583 * w(c, s1) = -1
4584 * w(c, s2) = 0
4585 * w(c, s3) = 1
4586 *
4587 */
4588static int
4589wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4590{
4591 s64 gran, vdiff = curr->vruntime - se->vruntime;
4592
4593 if (vdiff <= 0)
4594 return -1;
4595
e52fb7c0 4596 gran = wakeup_gran(curr, se);
464b7527
PZ
4597 if (vdiff > gran)
4598 return 1;
4599
4600 return 0;
4601}
4602
02479099
PZ
4603static void set_last_buddy(struct sched_entity *se)
4604{
69c80f3e
VP
4605 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4606 return;
4607
4608 for_each_sched_entity(se)
4609 cfs_rq_of(se)->last = se;
02479099
PZ
4610}
4611
4612static void set_next_buddy(struct sched_entity *se)
4613{
69c80f3e
VP
4614 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4615 return;
4616
4617 for_each_sched_entity(se)
4618 cfs_rq_of(se)->next = se;
02479099
PZ
4619}
4620
ac53db59
RR
4621static void set_skip_buddy(struct sched_entity *se)
4622{
69c80f3e
VP
4623 for_each_sched_entity(se)
4624 cfs_rq_of(se)->skip = se;
ac53db59
RR
4625}
4626
bf0f6f24
IM
4627/*
4628 * Preempt the current task with a newly woken task if needed:
4629 */
5a9b86f6 4630static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
bf0f6f24
IM
4631{
4632 struct task_struct *curr = rq->curr;
8651a86c 4633 struct sched_entity *se = &curr->se, *pse = &p->se;
03e89e45 4634 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
f685ceac 4635 int scale = cfs_rq->nr_running >= sched_nr_latency;
2f36825b 4636 int next_buddy_marked = 0;
bf0f6f24 4637
4ae7d5ce
IM
4638 if (unlikely(se == pse))
4639 return;
4640
5238cdd3 4641 /*
ddcdf6e7 4642 * This is possible from callers such as move_task(), in which we
5238cdd3
PT
4643 * unconditionally check_prempt_curr() after an enqueue (which may have
4644 * lead to a throttle). This both saves work and prevents false
4645 * next-buddy nomination below.
4646 */
4647 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4648 return;
4649
2f36825b 4650 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3cb63d52 4651 set_next_buddy(pse);
2f36825b
VP
4652 next_buddy_marked = 1;
4653 }
57fdc26d 4654
aec0a514
BR
4655 /*
4656 * We can come here with TIF_NEED_RESCHED already set from new task
4657 * wake up path.
5238cdd3
PT
4658 *
4659 * Note: this also catches the edge-case of curr being in a throttled
4660 * group (e.g. via set_curr_task), since update_curr() (in the
4661 * enqueue of curr) will have resulted in resched being set. This
4662 * prevents us from potentially nominating it as a false LAST_BUDDY
4663 * below.
aec0a514
BR
4664 */
4665 if (test_tsk_need_resched(curr))
4666 return;
4667
a2f5c9ab
DH
4668 /* Idle tasks are by definition preempted by non-idle tasks. */
4669 if (unlikely(curr->policy == SCHED_IDLE) &&
4670 likely(p->policy != SCHED_IDLE))
4671 goto preempt;
4672
91c234b4 4673 /*
a2f5c9ab
DH
4674 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4675 * is driven by the tick):
91c234b4 4676 */
8ed92e51 4677 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
91c234b4 4678 return;
bf0f6f24 4679
464b7527 4680 find_matching_se(&se, &pse);
9bbd7374 4681 update_curr(cfs_rq_of(se));
002f128b 4682 BUG_ON(!pse);
2f36825b
VP
4683 if (wakeup_preempt_entity(se, pse) == 1) {
4684 /*
4685 * Bias pick_next to pick the sched entity that is
4686 * triggering this preemption.
4687 */
4688 if (!next_buddy_marked)
4689 set_next_buddy(pse);
3a7e73a2 4690 goto preempt;
2f36825b 4691 }
464b7527 4692
3a7e73a2 4693 return;
a65ac745 4694
3a7e73a2
PZ
4695preempt:
4696 resched_task(curr);
4697 /*
4698 * Only set the backward buddy when the current task is still
4699 * on the rq. This can happen when a wakeup gets interleaved
4700 * with schedule on the ->pre_schedule() or idle_balance()
4701 * point, either of which can * drop the rq lock.
4702 *
4703 * Also, during early boot the idle thread is in the fair class,
4704 * for obvious reasons its a bad idea to schedule back to it.
4705 */
4706 if (unlikely(!se->on_rq || curr == rq->idle))
4707 return;
4708
4709 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4710 set_last_buddy(se);
bf0f6f24
IM
4711}
4712
606dba2e
PZ
4713static struct task_struct *
4714pick_next_task_fair(struct rq *rq, struct task_struct *prev)
bf0f6f24
IM
4715{
4716 struct cfs_rq *cfs_rq = &rq->cfs;
4717 struct sched_entity *se;
678d5718 4718 struct task_struct *p;
37e117c0 4719 int new_tasks;
678d5718 4720
6e83125c 4721again:
678d5718
PZ
4722#ifdef CONFIG_FAIR_GROUP_SCHED
4723 if (!cfs_rq->nr_running)
38033c37 4724 goto idle;
678d5718 4725
3f1d2a31 4726 if (prev->sched_class != &fair_sched_class)
678d5718
PZ
4727 goto simple;
4728
4729 /*
4730 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4731 * likely that a next task is from the same cgroup as the current.
4732 *
4733 * Therefore attempt to avoid putting and setting the entire cgroup
4734 * hierarchy, only change the part that actually changes.
4735 */
4736
4737 do {
4738 struct sched_entity *curr = cfs_rq->curr;
4739
4740 /*
4741 * Since we got here without doing put_prev_entity() we also
4742 * have to consider cfs_rq->curr. If it is still a runnable
4743 * entity, update_curr() will update its vruntime, otherwise
4744 * forget we've ever seen it.
4745 */
4746 if (curr && curr->on_rq)
4747 update_curr(cfs_rq);
4748 else
4749 curr = NULL;
4750
4751 /*
4752 * This call to check_cfs_rq_runtime() will do the throttle and
4753 * dequeue its entity in the parent(s). Therefore the 'simple'
4754 * nr_running test will indeed be correct.
4755 */
4756 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4757 goto simple;
4758
4759 se = pick_next_entity(cfs_rq, curr);
4760 cfs_rq = group_cfs_rq(se);
4761 } while (cfs_rq);
4762
4763 p = task_of(se);
4764
4765 /*
4766 * Since we haven't yet done put_prev_entity and if the selected task
4767 * is a different task than we started out with, try and touch the
4768 * least amount of cfs_rqs.
4769 */
4770 if (prev != p) {
4771 struct sched_entity *pse = &prev->se;
4772
4773 while (!(cfs_rq = is_same_group(se, pse))) {
4774 int se_depth = se->depth;
4775 int pse_depth = pse->depth;
4776
4777 if (se_depth <= pse_depth) {
4778 put_prev_entity(cfs_rq_of(pse), pse);
4779 pse = parent_entity(pse);
4780 }
4781 if (se_depth >= pse_depth) {
4782 set_next_entity(cfs_rq_of(se), se);
4783 se = parent_entity(se);
4784 }
4785 }
4786
4787 put_prev_entity(cfs_rq, pse);
4788 set_next_entity(cfs_rq, se);
4789 }
4790
4791 if (hrtick_enabled(rq))
4792 hrtick_start_fair(rq, p);
4793
4794 return p;
4795simple:
4796 cfs_rq = &rq->cfs;
4797#endif
bf0f6f24 4798
36ace27e 4799 if (!cfs_rq->nr_running)
38033c37 4800 goto idle;
bf0f6f24 4801
3f1d2a31 4802 put_prev_task(rq, prev);
606dba2e 4803
bf0f6f24 4804 do {
678d5718 4805 se = pick_next_entity(cfs_rq, NULL);
f4b6755f 4806 set_next_entity(cfs_rq, se);
bf0f6f24
IM
4807 cfs_rq = group_cfs_rq(se);
4808 } while (cfs_rq);
4809
8f4d37ec 4810 p = task_of(se);
678d5718 4811
b39e66ea
MG
4812 if (hrtick_enabled(rq))
4813 hrtick_start_fair(rq, p);
8f4d37ec
PZ
4814
4815 return p;
38033c37
PZ
4816
4817idle:
e4aa358b 4818 new_tasks = idle_balance(rq);
37e117c0
PZ
4819 /*
4820 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4821 * possible for any higher priority task to appear. In that case we
4822 * must re-start the pick_next_entity() loop.
4823 */
e4aa358b 4824 if (new_tasks < 0)
37e117c0
PZ
4825 return RETRY_TASK;
4826
e4aa358b 4827 if (new_tasks > 0)
38033c37 4828 goto again;
38033c37
PZ
4829
4830 return NULL;
bf0f6f24
IM
4831}
4832
4833/*
4834 * Account for a descheduled task:
4835 */
31ee529c 4836static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
bf0f6f24
IM
4837{
4838 struct sched_entity *se = &prev->se;
4839 struct cfs_rq *cfs_rq;
4840
4841 for_each_sched_entity(se) {
4842 cfs_rq = cfs_rq_of(se);
ab6cde26 4843 put_prev_entity(cfs_rq, se);
bf0f6f24
IM
4844 }
4845}
4846
ac53db59
RR
4847/*
4848 * sched_yield() is very simple
4849 *
4850 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4851 */
4852static void yield_task_fair(struct rq *rq)
4853{
4854 struct task_struct *curr = rq->curr;
4855 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4856 struct sched_entity *se = &curr->se;
4857
4858 /*
4859 * Are we the only task in the tree?
4860 */
4861 if (unlikely(rq->nr_running == 1))
4862 return;
4863
4864 clear_buddies(cfs_rq, se);
4865
4866 if (curr->policy != SCHED_BATCH) {
4867 update_rq_clock(rq);
4868 /*
4869 * Update run-time statistics of the 'current'.
4870 */
4871 update_curr(cfs_rq);
916671c0
MG
4872 /*
4873 * Tell update_rq_clock() that we've just updated,
4874 * so we don't do microscopic update in schedule()
4875 * and double the fastpath cost.
4876 */
4877 rq->skip_clock_update = 1;
ac53db59
RR
4878 }
4879
4880 set_skip_buddy(se);
4881}
4882
d95f4122
MG
4883static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4884{
4885 struct sched_entity *se = &p->se;
4886
5238cdd3
PT
4887 /* throttled hierarchies are not runnable */
4888 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
d95f4122
MG
4889 return false;
4890
4891 /* Tell the scheduler that we'd really like pse to run next. */
4892 set_next_buddy(se);
4893
d95f4122
MG
4894 yield_task_fair(rq);
4895
4896 return true;
4897}
4898
681f3e68 4899#ifdef CONFIG_SMP
bf0f6f24 4900/**************************************************
e9c84cb8
PZ
4901 * Fair scheduling class load-balancing methods.
4902 *
4903 * BASICS
4904 *
4905 * The purpose of load-balancing is to achieve the same basic fairness the
4906 * per-cpu scheduler provides, namely provide a proportional amount of compute
4907 * time to each task. This is expressed in the following equation:
4908 *
4909 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4910 *
4911 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4912 * W_i,0 is defined as:
4913 *
4914 * W_i,0 = \Sum_j w_i,j (2)
4915 *
4916 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4917 * is derived from the nice value as per prio_to_weight[].
4918 *
4919 * The weight average is an exponential decay average of the instantaneous
4920 * weight:
4921 *
4922 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4923 *
4924 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4925 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4926 * can also include other factors [XXX].
4927 *
4928 * To achieve this balance we define a measure of imbalance which follows
4929 * directly from (1):
4930 *
4931 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4932 *
4933 * We them move tasks around to minimize the imbalance. In the continuous
4934 * function space it is obvious this converges, in the discrete case we get
4935 * a few fun cases generally called infeasible weight scenarios.
4936 *
4937 * [XXX expand on:
4938 * - infeasible weights;
4939 * - local vs global optima in the discrete case. ]
4940 *
4941 *
4942 * SCHED DOMAINS
4943 *
4944 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4945 * for all i,j solution, we create a tree of cpus that follows the hardware
4946 * topology where each level pairs two lower groups (or better). This results
4947 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4948 * tree to only the first of the previous level and we decrease the frequency
4949 * of load-balance at each level inv. proportional to the number of cpus in
4950 * the groups.
4951 *
4952 * This yields:
4953 *
4954 * log_2 n 1 n
4955 * \Sum { --- * --- * 2^i } = O(n) (5)
4956 * i = 0 2^i 2^i
4957 * `- size of each group
4958 * | | `- number of cpus doing load-balance
4959 * | `- freq
4960 * `- sum over all levels
4961 *
4962 * Coupled with a limit on how many tasks we can migrate every balance pass,
4963 * this makes (5) the runtime complexity of the balancer.
4964 *
4965 * An important property here is that each CPU is still (indirectly) connected
4966 * to every other cpu in at most O(log n) steps:
4967 *
4968 * The adjacency matrix of the resulting graph is given by:
4969 *
4970 * log_2 n
4971 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4972 * k = 0
4973 *
4974 * And you'll find that:
4975 *
4976 * A^(log_2 n)_i,j != 0 for all i,j (7)
4977 *
4978 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4979 * The task movement gives a factor of O(m), giving a convergence complexity
4980 * of:
4981 *
4982 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4983 *
4984 *
4985 * WORK CONSERVING
4986 *
4987 * In order to avoid CPUs going idle while there's still work to do, new idle
4988 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4989 * tree itself instead of relying on other CPUs to bring it work.
4990 *
4991 * This adds some complexity to both (5) and (8) but it reduces the total idle
4992 * time.
4993 *
4994 * [XXX more?]
4995 *
4996 *
4997 * CGROUPS
4998 *
4999 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5000 *
5001 * s_k,i
5002 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5003 * S_k
5004 *
5005 * Where
5006 *
5007 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5008 *
5009 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5010 *
5011 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5012 * property.
5013 *
5014 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5015 * rewrite all of this once again.]
5016 */
bf0f6f24 5017
ed387b78
HS
5018static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5019
0ec8aa00
PZ
5020enum fbq_type { regular, remote, all };
5021
ddcdf6e7 5022#define LBF_ALL_PINNED 0x01
367456c7 5023#define LBF_NEED_BREAK 0x02
6263322c
PZ
5024#define LBF_DST_PINNED 0x04
5025#define LBF_SOME_PINNED 0x08
ddcdf6e7
PZ
5026
5027struct lb_env {
5028 struct sched_domain *sd;
5029
ddcdf6e7 5030 struct rq *src_rq;
85c1e7da 5031 int src_cpu;
ddcdf6e7
PZ
5032
5033 int dst_cpu;
5034 struct rq *dst_rq;
5035
88b8dac0
SV
5036 struct cpumask *dst_grpmask;
5037 int new_dst_cpu;
ddcdf6e7 5038 enum cpu_idle_type idle;
bd939f45 5039 long imbalance;
b9403130
MW
5040 /* The set of CPUs under consideration for load-balancing */
5041 struct cpumask *cpus;
5042
ddcdf6e7 5043 unsigned int flags;
367456c7
PZ
5044
5045 unsigned int loop;
5046 unsigned int loop_break;
5047 unsigned int loop_max;
0ec8aa00
PZ
5048
5049 enum fbq_type fbq_type;
ddcdf6e7
PZ
5050};
5051
1e3c88bd 5052/*
ddcdf6e7 5053 * move_task - move a task from one runqueue to another runqueue.
1e3c88bd
PZ
5054 * Both runqueues must be locked.
5055 */
ddcdf6e7 5056static void move_task(struct task_struct *p, struct lb_env *env)
1e3c88bd 5057{
ddcdf6e7
PZ
5058 deactivate_task(env->src_rq, p, 0);
5059 set_task_cpu(p, env->dst_cpu);
5060 activate_task(env->dst_rq, p, 0);
5061 check_preempt_curr(env->dst_rq, p, 0);
1e3c88bd
PZ
5062}
5063
029632fb
PZ
5064/*
5065 * Is this task likely cache-hot:
5066 */
5067static int
6037dd1a 5068task_hot(struct task_struct *p, u64 now)
029632fb
PZ
5069{
5070 s64 delta;
5071
5072 if (p->sched_class != &fair_sched_class)
5073 return 0;
5074
5075 if (unlikely(p->policy == SCHED_IDLE))
5076 return 0;
5077
5078 /*
5079 * Buddy candidates are cache hot:
5080 */
5081 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5082 (&p->se == cfs_rq_of(&p->se)->next ||
5083 &p->se == cfs_rq_of(&p->se)->last))
5084 return 1;
5085
5086 if (sysctl_sched_migration_cost == -1)
5087 return 1;
5088 if (sysctl_sched_migration_cost == 0)
5089 return 0;
5090
5091 delta = now - p->se.exec_start;
5092
5093 return delta < (s64)sysctl_sched_migration_cost;
5094}
5095
3a7053b3
MG
5096#ifdef CONFIG_NUMA_BALANCING
5097/* Returns true if the destination node has incurred more faults */
5098static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5099{
5100 int src_nid, dst_nid;
5101
ff1df896 5102 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
3a7053b3
MG
5103 !(env->sd->flags & SD_NUMA)) {
5104 return false;
5105 }
5106
5107 src_nid = cpu_to_node(env->src_cpu);
5108 dst_nid = cpu_to_node(env->dst_cpu);
5109
83e1d2cd 5110 if (src_nid == dst_nid)
3a7053b3
MG
5111 return false;
5112
83e1d2cd
MG
5113 /* Always encourage migration to the preferred node. */
5114 if (dst_nid == p->numa_preferred_nid)
5115 return true;
5116
887c290e
RR
5117 /* If both task and group weight improve, this move is a winner. */
5118 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
5119 group_weight(p, dst_nid) > group_weight(p, src_nid))
3a7053b3
MG
5120 return true;
5121
5122 return false;
5123}
7a0f3083
MG
5124
5125
5126static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5127{
5128 int src_nid, dst_nid;
5129
5130 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5131 return false;
5132
ff1df896 5133 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
7a0f3083
MG
5134 return false;
5135
5136 src_nid = cpu_to_node(env->src_cpu);
5137 dst_nid = cpu_to_node(env->dst_cpu);
5138
83e1d2cd 5139 if (src_nid == dst_nid)
7a0f3083
MG
5140 return false;
5141
83e1d2cd
MG
5142 /* Migrating away from the preferred node is always bad. */
5143 if (src_nid == p->numa_preferred_nid)
5144 return true;
5145
887c290e
RR
5146 /* If either task or group weight get worse, don't do it. */
5147 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
5148 group_weight(p, dst_nid) < group_weight(p, src_nid))
7a0f3083
MG
5149 return true;
5150
5151 return false;
5152}
5153
3a7053b3
MG
5154#else
5155static inline bool migrate_improves_locality(struct task_struct *p,
5156 struct lb_env *env)
5157{
5158 return false;
5159}
7a0f3083
MG
5160
5161static inline bool migrate_degrades_locality(struct task_struct *p,
5162 struct lb_env *env)
5163{
5164 return false;
5165}
3a7053b3
MG
5166#endif
5167
1e3c88bd
PZ
5168/*
5169 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5170 */
5171static
8e45cb54 5172int can_migrate_task(struct task_struct *p, struct lb_env *env)
1e3c88bd
PZ
5173{
5174 int tsk_cache_hot = 0;
5175 /*
5176 * We do not migrate tasks that are:
d3198084 5177 * 1) throttled_lb_pair, or
1e3c88bd 5178 * 2) cannot be migrated to this CPU due to cpus_allowed, or
d3198084
JK
5179 * 3) running (obviously), or
5180 * 4) are cache-hot on their current CPU.
1e3c88bd 5181 */
d3198084
JK
5182 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5183 return 0;
5184
ddcdf6e7 5185 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
e02e60c1 5186 int cpu;
88b8dac0 5187
41acab88 5188 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
88b8dac0 5189
6263322c
PZ
5190 env->flags |= LBF_SOME_PINNED;
5191
88b8dac0
SV
5192 /*
5193 * Remember if this task can be migrated to any other cpu in
5194 * our sched_group. We may want to revisit it if we couldn't
5195 * meet load balance goals by pulling other tasks on src_cpu.
5196 *
5197 * Also avoid computing new_dst_cpu if we have already computed
5198 * one in current iteration.
5199 */
6263322c 5200 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
88b8dac0
SV
5201 return 0;
5202
e02e60c1
JK
5203 /* Prevent to re-select dst_cpu via env's cpus */
5204 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5205 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6263322c 5206 env->flags |= LBF_DST_PINNED;
e02e60c1
JK
5207 env->new_dst_cpu = cpu;
5208 break;
5209 }
88b8dac0 5210 }
e02e60c1 5211
1e3c88bd
PZ
5212 return 0;
5213 }
88b8dac0
SV
5214
5215 /* Record that we found atleast one task that could run on dst_cpu */
8e45cb54 5216 env->flags &= ~LBF_ALL_PINNED;
1e3c88bd 5217
ddcdf6e7 5218 if (task_running(env->src_rq, p)) {
41acab88 5219 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1e3c88bd
PZ
5220 return 0;
5221 }
5222
5223 /*
5224 * Aggressive migration if:
3a7053b3
MG
5225 * 1) destination numa is preferred
5226 * 2) task is cache cold, or
5227 * 3) too many balance attempts have failed.
1e3c88bd 5228 */
6037dd1a 5229 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq));
7a0f3083
MG
5230 if (!tsk_cache_hot)
5231 tsk_cache_hot = migrate_degrades_locality(p, env);
3a7053b3
MG
5232
5233 if (migrate_improves_locality(p, env)) {
5234#ifdef CONFIG_SCHEDSTATS
5235 if (tsk_cache_hot) {
5236 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5237 schedstat_inc(p, se.statistics.nr_forced_migrations);
5238 }
5239#endif
5240 return 1;
5241 }
5242
1e3c88bd 5243 if (!tsk_cache_hot ||
8e45cb54 5244 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4e2dcb73 5245
1e3c88bd 5246 if (tsk_cache_hot) {
8e45cb54 5247 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
41acab88 5248 schedstat_inc(p, se.statistics.nr_forced_migrations);
1e3c88bd 5249 }
4e2dcb73 5250
1e3c88bd
PZ
5251 return 1;
5252 }
5253
4e2dcb73
ZH
5254 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5255 return 0;
1e3c88bd
PZ
5256}
5257
897c395f
PZ
5258/*
5259 * move_one_task tries to move exactly one task from busiest to this_rq, as
5260 * part of active balancing operations within "domain".
5261 * Returns 1 if successful and 0 otherwise.
5262 *
5263 * Called with both runqueues locked.
5264 */
8e45cb54 5265static int move_one_task(struct lb_env *env)
897c395f
PZ
5266{
5267 struct task_struct *p, *n;
897c395f 5268
367456c7 5269 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
367456c7
PZ
5270 if (!can_migrate_task(p, env))
5271 continue;
897c395f 5272
367456c7
PZ
5273 move_task(p, env);
5274 /*
5275 * Right now, this is only the second place move_task()
5276 * is called, so we can safely collect move_task()
5277 * stats here rather than inside move_task().
5278 */
5279 schedstat_inc(env->sd, lb_gained[env->idle]);
5280 return 1;
897c395f 5281 }
897c395f
PZ
5282 return 0;
5283}
5284
eb95308e
PZ
5285static const unsigned int sched_nr_migrate_break = 32;
5286
5d6523eb 5287/*
bd939f45 5288 * move_tasks tries to move up to imbalance weighted load from busiest to
5d6523eb
PZ
5289 * this_rq, as part of a balancing operation within domain "sd".
5290 * Returns 1 if successful and 0 otherwise.
5291 *
5292 * Called with both runqueues locked.
5293 */
5294static int move_tasks(struct lb_env *env)
1e3c88bd 5295{
5d6523eb
PZ
5296 struct list_head *tasks = &env->src_rq->cfs_tasks;
5297 struct task_struct *p;
367456c7
PZ
5298 unsigned long load;
5299 int pulled = 0;
1e3c88bd 5300
bd939f45 5301 if (env->imbalance <= 0)
5d6523eb 5302 return 0;
1e3c88bd 5303
5d6523eb
PZ
5304 while (!list_empty(tasks)) {
5305 p = list_first_entry(tasks, struct task_struct, se.group_node);
1e3c88bd 5306
367456c7
PZ
5307 env->loop++;
5308 /* We've more or less seen every task there is, call it quits */
5d6523eb 5309 if (env->loop > env->loop_max)
367456c7 5310 break;
5d6523eb
PZ
5311
5312 /* take a breather every nr_migrate tasks */
367456c7 5313 if (env->loop > env->loop_break) {
eb95308e 5314 env->loop_break += sched_nr_migrate_break;
8e45cb54 5315 env->flags |= LBF_NEED_BREAK;
ee00e66f 5316 break;
a195f004 5317 }
1e3c88bd 5318
d3198084 5319 if (!can_migrate_task(p, env))
367456c7
PZ
5320 goto next;
5321
5322 load = task_h_load(p);
5d6523eb 5323
eb95308e 5324 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
367456c7
PZ
5325 goto next;
5326
bd939f45 5327 if ((load / 2) > env->imbalance)
367456c7 5328 goto next;
1e3c88bd 5329
ddcdf6e7 5330 move_task(p, env);
ee00e66f 5331 pulled++;
bd939f45 5332 env->imbalance -= load;
1e3c88bd
PZ
5333
5334#ifdef CONFIG_PREEMPT
ee00e66f
PZ
5335 /*
5336 * NEWIDLE balancing is a source of latency, so preemptible
5337 * kernels will stop after the first task is pulled to minimize
5338 * the critical section.
5339 */
5d6523eb 5340 if (env->idle == CPU_NEWLY_IDLE)
ee00e66f 5341 break;
1e3c88bd
PZ
5342#endif
5343
ee00e66f
PZ
5344 /*
5345 * We only want to steal up to the prescribed amount of
5346 * weighted load.
5347 */
bd939f45 5348 if (env->imbalance <= 0)
ee00e66f 5349 break;
367456c7
PZ
5350
5351 continue;
5352next:
5d6523eb 5353 list_move_tail(&p->se.group_node, tasks);
1e3c88bd 5354 }
5d6523eb 5355
1e3c88bd 5356 /*
ddcdf6e7
PZ
5357 * Right now, this is one of only two places move_task() is called,
5358 * so we can safely collect move_task() stats here rather than
5359 * inside move_task().
1e3c88bd 5360 */
8e45cb54 5361 schedstat_add(env->sd, lb_gained[env->idle], pulled);
1e3c88bd 5362
5d6523eb 5363 return pulled;
1e3c88bd
PZ
5364}
5365
230059de 5366#ifdef CONFIG_FAIR_GROUP_SCHED
9e3081ca
PZ
5367/*
5368 * update tg->load_weight by folding this cpu's load_avg
5369 */
48a16753 5370static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
9e3081ca 5371{
48a16753
PT
5372 struct sched_entity *se = tg->se[cpu];
5373 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
9e3081ca 5374
48a16753
PT
5375 /* throttled entities do not contribute to load */
5376 if (throttled_hierarchy(cfs_rq))
5377 return;
9e3081ca 5378
aff3e498 5379 update_cfs_rq_blocked_load(cfs_rq, 1);
9e3081ca 5380
82958366
PT
5381 if (se) {
5382 update_entity_load_avg(se, 1);
5383 /*
5384 * We pivot on our runnable average having decayed to zero for
5385 * list removal. This generally implies that all our children
5386 * have also been removed (modulo rounding error or bandwidth
5387 * control); however, such cases are rare and we can fix these
5388 * at enqueue.
5389 *
5390 * TODO: fix up out-of-order children on enqueue.
5391 */
5392 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5393 list_del_leaf_cfs_rq(cfs_rq);
5394 } else {
48a16753 5395 struct rq *rq = rq_of(cfs_rq);
82958366
PT
5396 update_rq_runnable_avg(rq, rq->nr_running);
5397 }
9e3081ca
PZ
5398}
5399
48a16753 5400static void update_blocked_averages(int cpu)
9e3081ca 5401{
9e3081ca 5402 struct rq *rq = cpu_rq(cpu);
48a16753
PT
5403 struct cfs_rq *cfs_rq;
5404 unsigned long flags;
9e3081ca 5405
48a16753
PT
5406 raw_spin_lock_irqsave(&rq->lock, flags);
5407 update_rq_clock(rq);
9763b67f
PZ
5408 /*
5409 * Iterates the task_group tree in a bottom up fashion, see
5410 * list_add_leaf_cfs_rq() for details.
5411 */
64660c86 5412 for_each_leaf_cfs_rq(rq, cfs_rq) {
48a16753
PT
5413 /*
5414 * Note: We may want to consider periodically releasing
5415 * rq->lock about these updates so that creating many task
5416 * groups does not result in continually extending hold time.
5417 */
5418 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
64660c86 5419 }
48a16753
PT
5420
5421 raw_spin_unlock_irqrestore(&rq->lock, flags);
9e3081ca
PZ
5422}
5423
9763b67f 5424/*
68520796 5425 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9763b67f
PZ
5426 * This needs to be done in a top-down fashion because the load of a child
5427 * group is a fraction of its parents load.
5428 */
68520796 5429static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9763b67f 5430{
68520796
VD
5431 struct rq *rq = rq_of(cfs_rq);
5432 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
a35b6466 5433 unsigned long now = jiffies;
68520796 5434 unsigned long load;
a35b6466 5435
68520796 5436 if (cfs_rq->last_h_load_update == now)
a35b6466
PZ
5437 return;
5438
68520796
VD
5439 cfs_rq->h_load_next = NULL;
5440 for_each_sched_entity(se) {
5441 cfs_rq = cfs_rq_of(se);
5442 cfs_rq->h_load_next = se;
5443 if (cfs_rq->last_h_load_update == now)
5444 break;
5445 }
a35b6466 5446
68520796 5447 if (!se) {
7e3115ef 5448 cfs_rq->h_load = cfs_rq->runnable_load_avg;
68520796
VD
5449 cfs_rq->last_h_load_update = now;
5450 }
5451
5452 while ((se = cfs_rq->h_load_next) != NULL) {
5453 load = cfs_rq->h_load;
5454 load = div64_ul(load * se->avg.load_avg_contrib,
5455 cfs_rq->runnable_load_avg + 1);
5456 cfs_rq = group_cfs_rq(se);
5457 cfs_rq->h_load = load;
5458 cfs_rq->last_h_load_update = now;
5459 }
9763b67f
PZ
5460}
5461
367456c7 5462static unsigned long task_h_load(struct task_struct *p)
230059de 5463{
367456c7 5464 struct cfs_rq *cfs_rq = task_cfs_rq(p);
230059de 5465
68520796 5466 update_cfs_rq_h_load(cfs_rq);
a003a25b
AS
5467 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5468 cfs_rq->runnable_load_avg + 1);
230059de
PZ
5469}
5470#else
48a16753 5471static inline void update_blocked_averages(int cpu)
9e3081ca
PZ
5472{
5473}
5474
367456c7 5475static unsigned long task_h_load(struct task_struct *p)
1e3c88bd 5476{
a003a25b 5477 return p->se.avg.load_avg_contrib;
1e3c88bd 5478}
230059de 5479#endif
1e3c88bd 5480
1e3c88bd 5481/********** Helpers for find_busiest_group ************************/
1e3c88bd
PZ
5482/*
5483 * sg_lb_stats - stats of a sched_group required for load_balancing
5484 */
5485struct sg_lb_stats {
5486 unsigned long avg_load; /*Avg load across the CPUs of the group */
5487 unsigned long group_load; /* Total load over the CPUs of the group */
1e3c88bd 5488 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
56cf515b 5489 unsigned long load_per_task;
3ae11c90 5490 unsigned long group_power;
147c5fc2
PZ
5491 unsigned int sum_nr_running; /* Nr tasks running in the group */
5492 unsigned int group_capacity;
5493 unsigned int idle_cpus;
5494 unsigned int group_weight;
1e3c88bd 5495 int group_imb; /* Is there an imbalance in the group ? */
fab47622 5496 int group_has_capacity; /* Is there extra capacity in the group? */
0ec8aa00
PZ
5497#ifdef CONFIG_NUMA_BALANCING
5498 unsigned int nr_numa_running;
5499 unsigned int nr_preferred_running;
5500#endif
1e3c88bd
PZ
5501};
5502
56cf515b
JK
5503/*
5504 * sd_lb_stats - Structure to store the statistics of a sched_domain
5505 * during load balancing.
5506 */
5507struct sd_lb_stats {
5508 struct sched_group *busiest; /* Busiest group in this sd */
5509 struct sched_group *local; /* Local group in this sd */
5510 unsigned long total_load; /* Total load of all groups in sd */
5511 unsigned long total_pwr; /* Total power of all groups in sd */
5512 unsigned long avg_load; /* Average load across all groups in sd */
5513
56cf515b 5514 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
147c5fc2 5515 struct sg_lb_stats local_stat; /* Statistics of the local group */
56cf515b
JK
5516};
5517
147c5fc2
PZ
5518static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5519{
5520 /*
5521 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5522 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5523 * We must however clear busiest_stat::avg_load because
5524 * update_sd_pick_busiest() reads this before assignment.
5525 */
5526 *sds = (struct sd_lb_stats){
5527 .busiest = NULL,
5528 .local = NULL,
5529 .total_load = 0UL,
5530 .total_pwr = 0UL,
5531 .busiest_stat = {
5532 .avg_load = 0UL,
5533 },
5534 };
5535}
5536
1e3c88bd
PZ
5537/**
5538 * get_sd_load_idx - Obtain the load index for a given sched domain.
5539 * @sd: The sched_domain whose load_idx is to be obtained.
ed1b7732 5540 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
e69f6186
YB
5541 *
5542 * Return: The load index.
1e3c88bd
PZ
5543 */
5544static inline int get_sd_load_idx(struct sched_domain *sd,
5545 enum cpu_idle_type idle)
5546{
5547 int load_idx;
5548
5549 switch (idle) {
5550 case CPU_NOT_IDLE:
5551 load_idx = sd->busy_idx;
5552 break;
5553
5554 case CPU_NEWLY_IDLE:
5555 load_idx = sd->newidle_idx;
5556 break;
5557 default:
5558 load_idx = sd->idle_idx;
5559 break;
5560 }
5561
5562 return load_idx;
5563}
5564
15f803c9 5565static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
1e3c88bd 5566{
1399fa78 5567 return SCHED_POWER_SCALE;
1e3c88bd
PZ
5568}
5569
5570unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5571{
5572 return default_scale_freq_power(sd, cpu);
5573}
5574
15f803c9 5575static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
1e3c88bd 5576{
669c55e9 5577 unsigned long weight = sd->span_weight;
1e3c88bd
PZ
5578 unsigned long smt_gain = sd->smt_gain;
5579
5580 smt_gain /= weight;
5581
5582 return smt_gain;
5583}
5584
5585unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5586{
5587 return default_scale_smt_power(sd, cpu);
5588}
5589
15f803c9 5590static unsigned long scale_rt_power(int cpu)
1e3c88bd
PZ
5591{
5592 struct rq *rq = cpu_rq(cpu);
b654f7de 5593 u64 total, available, age_stamp, avg;
cadefd3d 5594 s64 delta;
1e3c88bd 5595
b654f7de
PZ
5596 /*
5597 * Since we're reading these variables without serialization make sure
5598 * we read them once before doing sanity checks on them.
5599 */
5600 age_stamp = ACCESS_ONCE(rq->age_stamp);
5601 avg = ACCESS_ONCE(rq->rt_avg);
5602
cadefd3d
PZ
5603 delta = rq_clock(rq) - age_stamp;
5604 if (unlikely(delta < 0))
5605 delta = 0;
5606
5607 total = sched_avg_period() + delta;
aa483808 5608
b654f7de 5609 if (unlikely(total < avg)) {
aa483808
VP
5610 /* Ensures that power won't end up being negative */
5611 available = 0;
5612 } else {
b654f7de 5613 available = total - avg;
aa483808 5614 }
1e3c88bd 5615
1399fa78
NR
5616 if (unlikely((s64)total < SCHED_POWER_SCALE))
5617 total = SCHED_POWER_SCALE;
1e3c88bd 5618
1399fa78 5619 total >>= SCHED_POWER_SHIFT;
1e3c88bd
PZ
5620
5621 return div_u64(available, total);
5622}
5623
5624static void update_cpu_power(struct sched_domain *sd, int cpu)
5625{
669c55e9 5626 unsigned long weight = sd->span_weight;
1399fa78 5627 unsigned long power = SCHED_POWER_SCALE;
1e3c88bd
PZ
5628 struct sched_group *sdg = sd->groups;
5629
1e3c88bd
PZ
5630 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5631 if (sched_feat(ARCH_POWER))
5632 power *= arch_scale_smt_power(sd, cpu);
5633 else
5634 power *= default_scale_smt_power(sd, cpu);
5635
1399fa78 5636 power >>= SCHED_POWER_SHIFT;
1e3c88bd
PZ
5637 }
5638
9c3f75cb 5639 sdg->sgp->power_orig = power;
9d5efe05
SV
5640
5641 if (sched_feat(ARCH_POWER))
5642 power *= arch_scale_freq_power(sd, cpu);
5643 else
5644 power *= default_scale_freq_power(sd, cpu);
5645
1399fa78 5646 power >>= SCHED_POWER_SHIFT;
9d5efe05 5647
1e3c88bd 5648 power *= scale_rt_power(cpu);
1399fa78 5649 power >>= SCHED_POWER_SHIFT;
1e3c88bd
PZ
5650
5651 if (!power)
5652 power = 1;
5653
e51fd5e2 5654 cpu_rq(cpu)->cpu_power = power;
9c3f75cb 5655 sdg->sgp->power = power;
1e3c88bd
PZ
5656}
5657
029632fb 5658void update_group_power(struct sched_domain *sd, int cpu)
1e3c88bd
PZ
5659{
5660 struct sched_domain *child = sd->child;
5661 struct sched_group *group, *sdg = sd->groups;
863bffc8 5662 unsigned long power, power_orig;
4ec4412e
VG
5663 unsigned long interval;
5664
5665 interval = msecs_to_jiffies(sd->balance_interval);
5666 interval = clamp(interval, 1UL, max_load_balance_interval);
5667 sdg->sgp->next_update = jiffies + interval;
1e3c88bd
PZ
5668
5669 if (!child) {
5670 update_cpu_power(sd, cpu);
5671 return;
5672 }
5673
863bffc8 5674 power_orig = power = 0;
1e3c88bd 5675
74a5ce20
PZ
5676 if (child->flags & SD_OVERLAP) {
5677 /*
5678 * SD_OVERLAP domains cannot assume that child groups
5679 * span the current group.
5680 */
5681
863bffc8 5682 for_each_cpu(cpu, sched_group_cpus(sdg)) {
9abf24d4
SD
5683 struct sched_group_power *sgp;
5684 struct rq *rq = cpu_rq(cpu);
863bffc8 5685
9abf24d4
SD
5686 /*
5687 * build_sched_domains() -> init_sched_groups_power()
5688 * gets here before we've attached the domains to the
5689 * runqueues.
5690 *
5691 * Use power_of(), which is set irrespective of domains
5692 * in update_cpu_power().
5693 *
5694 * This avoids power/power_orig from being 0 and
5695 * causing divide-by-zero issues on boot.
5696 *
5697 * Runtime updates will correct power_orig.
5698 */
5699 if (unlikely(!rq->sd)) {
5700 power_orig += power_of(cpu);
5701 power += power_of(cpu);
5702 continue;
5703 }
863bffc8 5704
9abf24d4
SD
5705 sgp = rq->sd->groups->sgp;
5706 power_orig += sgp->power_orig;
5707 power += sgp->power;
863bffc8 5708 }
74a5ce20
PZ
5709 } else {
5710 /*
5711 * !SD_OVERLAP domains can assume that child groups
5712 * span the current group.
5713 */
5714
5715 group = child->groups;
5716 do {
863bffc8 5717 power_orig += group->sgp->power_orig;
74a5ce20
PZ
5718 power += group->sgp->power;
5719 group = group->next;
5720 } while (group != child->groups);
5721 }
1e3c88bd 5722
863bffc8
PZ
5723 sdg->sgp->power_orig = power_orig;
5724 sdg->sgp->power = power;
1e3c88bd
PZ
5725}
5726
9d5efe05
SV
5727/*
5728 * Try and fix up capacity for tiny siblings, this is needed when
5729 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5730 * which on its own isn't powerful enough.
5731 *
5732 * See update_sd_pick_busiest() and check_asym_packing().
5733 */
5734static inline int
5735fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5736{
5737 /*
1399fa78 5738 * Only siblings can have significantly less than SCHED_POWER_SCALE
9d5efe05 5739 */
a6c75f2f 5740 if (!(sd->flags & SD_SHARE_CPUPOWER))
9d5efe05
SV
5741 return 0;
5742
5743 /*
5744 * If ~90% of the cpu_power is still there, we're good.
5745 */
9c3f75cb 5746 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
9d5efe05
SV
5747 return 1;
5748
5749 return 0;
5750}
5751
30ce5dab
PZ
5752/*
5753 * Group imbalance indicates (and tries to solve) the problem where balancing
5754 * groups is inadequate due to tsk_cpus_allowed() constraints.
5755 *
5756 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5757 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5758 * Something like:
5759 *
5760 * { 0 1 2 3 } { 4 5 6 7 }
5761 * * * * *
5762 *
5763 * If we were to balance group-wise we'd place two tasks in the first group and
5764 * two tasks in the second group. Clearly this is undesired as it will overload
5765 * cpu 3 and leave one of the cpus in the second group unused.
5766 *
5767 * The current solution to this issue is detecting the skew in the first group
6263322c
PZ
5768 * by noticing the lower domain failed to reach balance and had difficulty
5769 * moving tasks due to affinity constraints.
30ce5dab
PZ
5770 *
5771 * When this is so detected; this group becomes a candidate for busiest; see
ed1b7732 5772 * update_sd_pick_busiest(). And calculate_imbalance() and
6263322c 5773 * find_busiest_group() avoid some of the usual balance conditions to allow it
30ce5dab
PZ
5774 * to create an effective group imbalance.
5775 *
5776 * This is a somewhat tricky proposition since the next run might not find the
5777 * group imbalance and decide the groups need to be balanced again. A most
5778 * subtle and fragile situation.
5779 */
5780
6263322c 5781static inline int sg_imbalanced(struct sched_group *group)
30ce5dab 5782{
6263322c 5783 return group->sgp->imbalance;
30ce5dab
PZ
5784}
5785
b37d9316
PZ
5786/*
5787 * Compute the group capacity.
5788 *
c61037e9
PZ
5789 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5790 * first dividing out the smt factor and computing the actual number of cores
5791 * and limit power unit capacity with that.
b37d9316
PZ
5792 */
5793static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5794{
c61037e9
PZ
5795 unsigned int capacity, smt, cpus;
5796 unsigned int power, power_orig;
5797
5798 power = group->sgp->power;
5799 power_orig = group->sgp->power_orig;
5800 cpus = group->group_weight;
b37d9316 5801
c61037e9
PZ
5802 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5803 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5804 capacity = cpus / smt; /* cores */
b37d9316 5805
c61037e9 5806 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
b37d9316
PZ
5807 if (!capacity)
5808 capacity = fix_small_capacity(env->sd, group);
5809
5810 return capacity;
5811}
5812
1e3c88bd
PZ
5813/**
5814 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
cd96891d 5815 * @env: The load balancing environment.
1e3c88bd 5816 * @group: sched_group whose statistics are to be updated.
1e3c88bd 5817 * @load_idx: Load index of sched_domain of this_cpu for load calc.
1e3c88bd 5818 * @local_group: Does group contain this_cpu.
1e3c88bd
PZ
5819 * @sgs: variable to hold the statistics for this group.
5820 */
bd939f45
PZ
5821static inline void update_sg_lb_stats(struct lb_env *env,
5822 struct sched_group *group, int load_idx,
23f0d209 5823 int local_group, struct sg_lb_stats *sgs)
1e3c88bd 5824{
30ce5dab 5825 unsigned long load;
bd939f45 5826 int i;
1e3c88bd 5827
b72ff13c
PZ
5828 memset(sgs, 0, sizeof(*sgs));
5829
b9403130 5830 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
1e3c88bd
PZ
5831 struct rq *rq = cpu_rq(i);
5832
1e3c88bd 5833 /* Bias balancing toward cpus of our domain */
6263322c 5834 if (local_group)
04f733b4 5835 load = target_load(i, load_idx);
6263322c 5836 else
1e3c88bd 5837 load = source_load(i, load_idx);
1e3c88bd
PZ
5838
5839 sgs->group_load += load;
380c9077 5840 sgs->sum_nr_running += rq->nr_running;
0ec8aa00
PZ
5841#ifdef CONFIG_NUMA_BALANCING
5842 sgs->nr_numa_running += rq->nr_numa_running;
5843 sgs->nr_preferred_running += rq->nr_preferred_running;
5844#endif
1e3c88bd 5845 sgs->sum_weighted_load += weighted_cpuload(i);
aae6d3dd
SS
5846 if (idle_cpu(i))
5847 sgs->idle_cpus++;
1e3c88bd
PZ
5848 }
5849
1e3c88bd 5850 /* Adjust by relative CPU power of the group */
3ae11c90
PZ
5851 sgs->group_power = group->sgp->power;
5852 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
1e3c88bd 5853
dd5feea1 5854 if (sgs->sum_nr_running)
38d0f770 5855 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
1e3c88bd 5856
aae6d3dd 5857 sgs->group_weight = group->group_weight;
fab47622 5858
b37d9316
PZ
5859 sgs->group_imb = sg_imbalanced(group);
5860 sgs->group_capacity = sg_capacity(env, group);
5861
fab47622
NR
5862 if (sgs->group_capacity > sgs->sum_nr_running)
5863 sgs->group_has_capacity = 1;
1e3c88bd
PZ
5864}
5865
532cb4c4
MN
5866/**
5867 * update_sd_pick_busiest - return 1 on busiest group
cd96891d 5868 * @env: The load balancing environment.
532cb4c4
MN
5869 * @sds: sched_domain statistics
5870 * @sg: sched_group candidate to be checked for being the busiest
b6b12294 5871 * @sgs: sched_group statistics
532cb4c4
MN
5872 *
5873 * Determine if @sg is a busier group than the previously selected
5874 * busiest group.
e69f6186
YB
5875 *
5876 * Return: %true if @sg is a busier group than the previously selected
5877 * busiest group. %false otherwise.
532cb4c4 5878 */
bd939f45 5879static bool update_sd_pick_busiest(struct lb_env *env,
532cb4c4
MN
5880 struct sd_lb_stats *sds,
5881 struct sched_group *sg,
bd939f45 5882 struct sg_lb_stats *sgs)
532cb4c4 5883{
56cf515b 5884 if (sgs->avg_load <= sds->busiest_stat.avg_load)
532cb4c4
MN
5885 return false;
5886
5887 if (sgs->sum_nr_running > sgs->group_capacity)
5888 return true;
5889
5890 if (sgs->group_imb)
5891 return true;
5892
5893 /*
5894 * ASYM_PACKING needs to move all the work to the lowest
5895 * numbered CPUs in the group, therefore mark all groups
5896 * higher than ourself as busy.
5897 */
bd939f45
PZ
5898 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5899 env->dst_cpu < group_first_cpu(sg)) {
532cb4c4
MN
5900 if (!sds->busiest)
5901 return true;
5902
5903 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5904 return true;
5905 }
5906
5907 return false;
5908}
5909
0ec8aa00
PZ
5910#ifdef CONFIG_NUMA_BALANCING
5911static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5912{
5913 if (sgs->sum_nr_running > sgs->nr_numa_running)
5914 return regular;
5915 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5916 return remote;
5917 return all;
5918}
5919
5920static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5921{
5922 if (rq->nr_running > rq->nr_numa_running)
5923 return regular;
5924 if (rq->nr_running > rq->nr_preferred_running)
5925 return remote;
5926 return all;
5927}
5928#else
5929static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5930{
5931 return all;
5932}
5933
5934static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5935{
5936 return regular;
5937}
5938#endif /* CONFIG_NUMA_BALANCING */
5939
1e3c88bd 5940/**
461819ac 5941 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
cd96891d 5942 * @env: The load balancing environment.
1e3c88bd
PZ
5943 * @sds: variable to hold the statistics for this sched_domain.
5944 */
0ec8aa00 5945static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
1e3c88bd 5946{
bd939f45
PZ
5947 struct sched_domain *child = env->sd->child;
5948 struct sched_group *sg = env->sd->groups;
56cf515b 5949 struct sg_lb_stats tmp_sgs;
1e3c88bd
PZ
5950 int load_idx, prefer_sibling = 0;
5951
5952 if (child && child->flags & SD_PREFER_SIBLING)
5953 prefer_sibling = 1;
5954
bd939f45 5955 load_idx = get_sd_load_idx(env->sd, env->idle);
1e3c88bd
PZ
5956
5957 do {
56cf515b 5958 struct sg_lb_stats *sgs = &tmp_sgs;
1e3c88bd
PZ
5959 int local_group;
5960
bd939f45 5961 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
56cf515b
JK
5962 if (local_group) {
5963 sds->local = sg;
5964 sgs = &sds->local_stat;
b72ff13c
PZ
5965
5966 if (env->idle != CPU_NEWLY_IDLE ||
5967 time_after_eq(jiffies, sg->sgp->next_update))
5968 update_group_power(env->sd, env->dst_cpu);
56cf515b 5969 }
1e3c88bd 5970
56cf515b 5971 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
1e3c88bd 5972
b72ff13c
PZ
5973 if (local_group)
5974 goto next_group;
5975
1e3c88bd
PZ
5976 /*
5977 * In case the child domain prefers tasks go to siblings
532cb4c4 5978 * first, lower the sg capacity to one so that we'll try
75dd321d
NR
5979 * and move all the excess tasks away. We lower the capacity
5980 * of a group only if the local group has the capacity to fit
5981 * these excess tasks, i.e. nr_running < group_capacity. The
5982 * extra check prevents the case where you always pull from the
5983 * heaviest group when it is already under-utilized (possible
5984 * with a large weight task outweighs the tasks on the system).
1e3c88bd 5985 */
b72ff13c
PZ
5986 if (prefer_sibling && sds->local &&
5987 sds->local_stat.group_has_capacity)
147c5fc2 5988 sgs->group_capacity = min(sgs->group_capacity, 1U);
1e3c88bd 5989
b72ff13c 5990 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
532cb4c4 5991 sds->busiest = sg;
56cf515b 5992 sds->busiest_stat = *sgs;
1e3c88bd
PZ
5993 }
5994
b72ff13c
PZ
5995next_group:
5996 /* Now, start updating sd_lb_stats */
5997 sds->total_load += sgs->group_load;
5998 sds->total_pwr += sgs->group_power;
5999
532cb4c4 6000 sg = sg->next;
bd939f45 6001 } while (sg != env->sd->groups);
0ec8aa00
PZ
6002
6003 if (env->sd->flags & SD_NUMA)
6004 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
532cb4c4
MN
6005}
6006
532cb4c4
MN
6007/**
6008 * check_asym_packing - Check to see if the group is packed into the
6009 * sched doman.
6010 *
6011 * This is primarily intended to used at the sibling level. Some
6012 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6013 * case of POWER7, it can move to lower SMT modes only when higher
6014 * threads are idle. When in lower SMT modes, the threads will
6015 * perform better since they share less core resources. Hence when we
6016 * have idle threads, we want them to be the higher ones.
6017 *
6018 * This packing function is run on idle threads. It checks to see if
6019 * the busiest CPU in this domain (core in the P7 case) has a higher
6020 * CPU number than the packing function is being run on. Here we are
6021 * assuming lower CPU number will be equivalent to lower a SMT thread
6022 * number.
6023 *
e69f6186 6024 * Return: 1 when packing is required and a task should be moved to
b6b12294
MN
6025 * this CPU. The amount of the imbalance is returned in *imbalance.
6026 *
cd96891d 6027 * @env: The load balancing environment.
532cb4c4 6028 * @sds: Statistics of the sched_domain which is to be packed
532cb4c4 6029 */
bd939f45 6030static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
532cb4c4
MN
6031{
6032 int busiest_cpu;
6033
bd939f45 6034 if (!(env->sd->flags & SD_ASYM_PACKING))
532cb4c4
MN
6035 return 0;
6036
6037 if (!sds->busiest)
6038 return 0;
6039
6040 busiest_cpu = group_first_cpu(sds->busiest);
bd939f45 6041 if (env->dst_cpu > busiest_cpu)
532cb4c4
MN
6042 return 0;
6043
bd939f45 6044 env->imbalance = DIV_ROUND_CLOSEST(
3ae11c90
PZ
6045 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
6046 SCHED_POWER_SCALE);
bd939f45 6047
532cb4c4 6048 return 1;
1e3c88bd
PZ
6049}
6050
6051/**
6052 * fix_small_imbalance - Calculate the minor imbalance that exists
6053 * amongst the groups of a sched_domain, during
6054 * load balancing.
cd96891d 6055 * @env: The load balancing environment.
1e3c88bd 6056 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
1e3c88bd 6057 */
bd939f45
PZ
6058static inline
6059void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
1e3c88bd
PZ
6060{
6061 unsigned long tmp, pwr_now = 0, pwr_move = 0;
6062 unsigned int imbn = 2;
dd5feea1 6063 unsigned long scaled_busy_load_per_task;
56cf515b 6064 struct sg_lb_stats *local, *busiest;
1e3c88bd 6065
56cf515b
JK
6066 local = &sds->local_stat;
6067 busiest = &sds->busiest_stat;
1e3c88bd 6068
56cf515b
JK
6069 if (!local->sum_nr_running)
6070 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6071 else if (busiest->load_per_task > local->load_per_task)
6072 imbn = 1;
dd5feea1 6073
56cf515b
JK
6074 scaled_busy_load_per_task =
6075 (busiest->load_per_task * SCHED_POWER_SCALE) /
3ae11c90 6076 busiest->group_power;
56cf515b 6077
3029ede3
VD
6078 if (busiest->avg_load + scaled_busy_load_per_task >=
6079 local->avg_load + (scaled_busy_load_per_task * imbn)) {
56cf515b 6080 env->imbalance = busiest->load_per_task;
1e3c88bd
PZ
6081 return;
6082 }
6083
6084 /*
6085 * OK, we don't have enough imbalance to justify moving tasks,
6086 * however we may be able to increase total CPU power used by
6087 * moving them.
6088 */
6089
3ae11c90 6090 pwr_now += busiest->group_power *
56cf515b 6091 min(busiest->load_per_task, busiest->avg_load);
3ae11c90 6092 pwr_now += local->group_power *
56cf515b 6093 min(local->load_per_task, local->avg_load);
1399fa78 6094 pwr_now /= SCHED_POWER_SCALE;
1e3c88bd
PZ
6095
6096 /* Amount of load we'd subtract */
a2cd4260 6097 if (busiest->avg_load > scaled_busy_load_per_task) {
3ae11c90 6098 pwr_move += busiest->group_power *
56cf515b 6099 min(busiest->load_per_task,
a2cd4260 6100 busiest->avg_load - scaled_busy_load_per_task);
56cf515b 6101 }
1e3c88bd
PZ
6102
6103 /* Amount of load we'd add */
3ae11c90 6104 if (busiest->avg_load * busiest->group_power <
56cf515b 6105 busiest->load_per_task * SCHED_POWER_SCALE) {
3ae11c90
PZ
6106 tmp = (busiest->avg_load * busiest->group_power) /
6107 local->group_power;
56cf515b
JK
6108 } else {
6109 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
3ae11c90 6110 local->group_power;
56cf515b 6111 }
3ae11c90
PZ
6112 pwr_move += local->group_power *
6113 min(local->load_per_task, local->avg_load + tmp);
1399fa78 6114 pwr_move /= SCHED_POWER_SCALE;
1e3c88bd
PZ
6115
6116 /* Move if we gain throughput */
6117 if (pwr_move > pwr_now)
56cf515b 6118 env->imbalance = busiest->load_per_task;
1e3c88bd
PZ
6119}
6120
6121/**
6122 * calculate_imbalance - Calculate the amount of imbalance present within the
6123 * groups of a given sched_domain during load balance.
bd939f45 6124 * @env: load balance environment
1e3c88bd 6125 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
1e3c88bd 6126 */
bd939f45 6127static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
1e3c88bd 6128{
dd5feea1 6129 unsigned long max_pull, load_above_capacity = ~0UL;
56cf515b
JK
6130 struct sg_lb_stats *local, *busiest;
6131
6132 local = &sds->local_stat;
56cf515b 6133 busiest = &sds->busiest_stat;
dd5feea1 6134
56cf515b 6135 if (busiest->group_imb) {
30ce5dab
PZ
6136 /*
6137 * In the group_imb case we cannot rely on group-wide averages
6138 * to ensure cpu-load equilibrium, look at wider averages. XXX
6139 */
56cf515b
JK
6140 busiest->load_per_task =
6141 min(busiest->load_per_task, sds->avg_load);
dd5feea1
SS
6142 }
6143
1e3c88bd
PZ
6144 /*
6145 * In the presence of smp nice balancing, certain scenarios can have
6146 * max load less than avg load(as we skip the groups at or below
6147 * its cpu_power, while calculating max_load..)
6148 */
b1885550
VD
6149 if (busiest->avg_load <= sds->avg_load ||
6150 local->avg_load >= sds->avg_load) {
bd939f45
PZ
6151 env->imbalance = 0;
6152 return fix_small_imbalance(env, sds);
1e3c88bd
PZ
6153 }
6154
56cf515b 6155 if (!busiest->group_imb) {
dd5feea1
SS
6156 /*
6157 * Don't want to pull so many tasks that a group would go idle.
30ce5dab
PZ
6158 * Except of course for the group_imb case, since then we might
6159 * have to drop below capacity to reach cpu-load equilibrium.
dd5feea1 6160 */
56cf515b
JK
6161 load_above_capacity =
6162 (busiest->sum_nr_running - busiest->group_capacity);
dd5feea1 6163
1399fa78 6164 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3ae11c90 6165 load_above_capacity /= busiest->group_power;
dd5feea1
SS
6166 }
6167
6168 /*
6169 * We're trying to get all the cpus to the average_load, so we don't
6170 * want to push ourselves above the average load, nor do we wish to
6171 * reduce the max loaded cpu below the average load. At the same time,
6172 * we also don't want to reduce the group load below the group capacity
6173 * (so that we can implement power-savings policies etc). Thus we look
6174 * for the minimum possible imbalance.
dd5feea1 6175 */
30ce5dab 6176 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
1e3c88bd
PZ
6177
6178 /* How much load to actually move to equalise the imbalance */
56cf515b 6179 env->imbalance = min(
3ae11c90
PZ
6180 max_pull * busiest->group_power,
6181 (sds->avg_load - local->avg_load) * local->group_power
56cf515b 6182 ) / SCHED_POWER_SCALE;
1e3c88bd
PZ
6183
6184 /*
6185 * if *imbalance is less than the average load per runnable task
25985edc 6186 * there is no guarantee that any tasks will be moved so we'll have
1e3c88bd
PZ
6187 * a think about bumping its value to force at least one task to be
6188 * moved
6189 */
56cf515b 6190 if (env->imbalance < busiest->load_per_task)
bd939f45 6191 return fix_small_imbalance(env, sds);
1e3c88bd 6192}
fab47622 6193
1e3c88bd
PZ
6194/******* find_busiest_group() helpers end here *********************/
6195
6196/**
6197 * find_busiest_group - Returns the busiest group within the sched_domain
6198 * if there is an imbalance. If there isn't an imbalance, and
6199 * the user has opted for power-savings, it returns a group whose
6200 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6201 * such a group exists.
6202 *
6203 * Also calculates the amount of weighted load which should be moved
6204 * to restore balance.
6205 *
cd96891d 6206 * @env: The load balancing environment.
1e3c88bd 6207 *
e69f6186 6208 * Return: - The busiest group if imbalance exists.
1e3c88bd
PZ
6209 * - If no imbalance and user has opted for power-savings balance,
6210 * return the least loaded group whose CPUs can be
6211 * put to idle by rebalancing its tasks onto our group.
6212 */
56cf515b 6213static struct sched_group *find_busiest_group(struct lb_env *env)
1e3c88bd 6214{
56cf515b 6215 struct sg_lb_stats *local, *busiest;
1e3c88bd
PZ
6216 struct sd_lb_stats sds;
6217
147c5fc2 6218 init_sd_lb_stats(&sds);
1e3c88bd
PZ
6219
6220 /*
6221 * Compute the various statistics relavent for load balancing at
6222 * this level.
6223 */
23f0d209 6224 update_sd_lb_stats(env, &sds);
56cf515b
JK
6225 local = &sds.local_stat;
6226 busiest = &sds.busiest_stat;
1e3c88bd 6227
bd939f45
PZ
6228 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6229 check_asym_packing(env, &sds))
532cb4c4
MN
6230 return sds.busiest;
6231
cc57aa8f 6232 /* There is no busy sibling group to pull tasks from */
56cf515b 6233 if (!sds.busiest || busiest->sum_nr_running == 0)
1e3c88bd
PZ
6234 goto out_balanced;
6235
1399fa78 6236 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
b0432d8f 6237
866ab43e
PZ
6238 /*
6239 * If the busiest group is imbalanced the below checks don't
30ce5dab 6240 * work because they assume all things are equal, which typically
866ab43e
PZ
6241 * isn't true due to cpus_allowed constraints and the like.
6242 */
56cf515b 6243 if (busiest->group_imb)
866ab43e
PZ
6244 goto force_balance;
6245
cc57aa8f 6246 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
56cf515b
JK
6247 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6248 !busiest->group_has_capacity)
fab47622
NR
6249 goto force_balance;
6250
cc57aa8f
PZ
6251 /*
6252 * If the local group is more busy than the selected busiest group
6253 * don't try and pull any tasks.
6254 */
56cf515b 6255 if (local->avg_load >= busiest->avg_load)
1e3c88bd
PZ
6256 goto out_balanced;
6257
cc57aa8f
PZ
6258 /*
6259 * Don't pull any tasks if this group is already above the domain
6260 * average load.
6261 */
56cf515b 6262 if (local->avg_load >= sds.avg_load)
1e3c88bd
PZ
6263 goto out_balanced;
6264
bd939f45 6265 if (env->idle == CPU_IDLE) {
aae6d3dd
SS
6266 /*
6267 * This cpu is idle. If the busiest group load doesn't
6268 * have more tasks than the number of available cpu's and
6269 * there is no imbalance between this and busiest group
6270 * wrt to idle cpu's, it is balanced.
6271 */
56cf515b
JK
6272 if ((local->idle_cpus < busiest->idle_cpus) &&
6273 busiest->sum_nr_running <= busiest->group_weight)
aae6d3dd 6274 goto out_balanced;
c186fafe
PZ
6275 } else {
6276 /*
6277 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6278 * imbalance_pct to be conservative.
6279 */
56cf515b
JK
6280 if (100 * busiest->avg_load <=
6281 env->sd->imbalance_pct * local->avg_load)
c186fafe 6282 goto out_balanced;
aae6d3dd 6283 }
1e3c88bd 6284
fab47622 6285force_balance:
1e3c88bd 6286 /* Looks like there is an imbalance. Compute it */
bd939f45 6287 calculate_imbalance(env, &sds);
1e3c88bd
PZ
6288 return sds.busiest;
6289
6290out_balanced:
bd939f45 6291 env->imbalance = 0;
1e3c88bd
PZ
6292 return NULL;
6293}
6294
6295/*
6296 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6297 */
bd939f45 6298static struct rq *find_busiest_queue(struct lb_env *env,
b9403130 6299 struct sched_group *group)
1e3c88bd
PZ
6300{
6301 struct rq *busiest = NULL, *rq;
95a79b80 6302 unsigned long busiest_load = 0, busiest_power = 1;
1e3c88bd
PZ
6303 int i;
6304
6906a408 6305 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
0ec8aa00
PZ
6306 unsigned long power, capacity, wl;
6307 enum fbq_type rt;
6308
6309 rq = cpu_rq(i);
6310 rt = fbq_classify_rq(rq);
1e3c88bd 6311
0ec8aa00
PZ
6312 /*
6313 * We classify groups/runqueues into three groups:
6314 * - regular: there are !numa tasks
6315 * - remote: there are numa tasks that run on the 'wrong' node
6316 * - all: there is no distinction
6317 *
6318 * In order to avoid migrating ideally placed numa tasks,
6319 * ignore those when there's better options.
6320 *
6321 * If we ignore the actual busiest queue to migrate another
6322 * task, the next balance pass can still reduce the busiest
6323 * queue by moving tasks around inside the node.
6324 *
6325 * If we cannot move enough load due to this classification
6326 * the next pass will adjust the group classification and
6327 * allow migration of more tasks.
6328 *
6329 * Both cases only affect the total convergence complexity.
6330 */
6331 if (rt > env->fbq_type)
6332 continue;
6333
6334 power = power_of(i);
6335 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
9d5efe05 6336 if (!capacity)
bd939f45 6337 capacity = fix_small_capacity(env->sd, group);
9d5efe05 6338
6e40f5bb 6339 wl = weighted_cpuload(i);
1e3c88bd 6340
6e40f5bb
TG
6341 /*
6342 * When comparing with imbalance, use weighted_cpuload()
6343 * which is not scaled with the cpu power.
6344 */
bd939f45 6345 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
1e3c88bd
PZ
6346 continue;
6347
6e40f5bb
TG
6348 /*
6349 * For the load comparisons with the other cpu's, consider
6350 * the weighted_cpuload() scaled with the cpu power, so that
6351 * the load can be moved away from the cpu that is potentially
6352 * running at a lower capacity.
95a79b80
JK
6353 *
6354 * Thus we're looking for max(wl_i / power_i), crosswise
6355 * multiplication to rid ourselves of the division works out
6356 * to: wl_i * power_j > wl_j * power_i; where j is our
6357 * previous maximum.
6e40f5bb 6358 */
95a79b80
JK
6359 if (wl * busiest_power > busiest_load * power) {
6360 busiest_load = wl;
6361 busiest_power = power;
1e3c88bd
PZ
6362 busiest = rq;
6363 }
6364 }
6365
6366 return busiest;
6367}
6368
6369/*
6370 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6371 * so long as it is large enough.
6372 */
6373#define MAX_PINNED_INTERVAL 512
6374
6375/* Working cpumask for load_balance and load_balance_newidle. */
e6252c3e 6376DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
1e3c88bd 6377
bd939f45 6378static int need_active_balance(struct lb_env *env)
1af3ed3d 6379{
bd939f45
PZ
6380 struct sched_domain *sd = env->sd;
6381
6382 if (env->idle == CPU_NEWLY_IDLE) {
532cb4c4
MN
6383
6384 /*
6385 * ASYM_PACKING needs to force migrate tasks from busy but
6386 * higher numbered CPUs in order to pack all tasks in the
6387 * lowest numbered CPUs.
6388 */
bd939f45 6389 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
532cb4c4 6390 return 1;
1af3ed3d
PZ
6391 }
6392
6393 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6394}
6395
969c7921
TH
6396static int active_load_balance_cpu_stop(void *data);
6397
23f0d209
JK
6398static int should_we_balance(struct lb_env *env)
6399{
6400 struct sched_group *sg = env->sd->groups;
6401 struct cpumask *sg_cpus, *sg_mask;
6402 int cpu, balance_cpu = -1;
6403
6404 /*
6405 * In the newly idle case, we will allow all the cpu's
6406 * to do the newly idle load balance.
6407 */
6408 if (env->idle == CPU_NEWLY_IDLE)
6409 return 1;
6410
6411 sg_cpus = sched_group_cpus(sg);
6412 sg_mask = sched_group_mask(sg);
6413 /* Try to find first idle cpu */
6414 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6415 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6416 continue;
6417
6418 balance_cpu = cpu;
6419 break;
6420 }
6421
6422 if (balance_cpu == -1)
6423 balance_cpu = group_balance_cpu(sg);
6424
6425 /*
6426 * First idle cpu or the first cpu(busiest) in this sched group
6427 * is eligible for doing load balancing at this and above domains.
6428 */
b0cff9d8 6429 return balance_cpu == env->dst_cpu;
23f0d209
JK
6430}
6431
1e3c88bd
PZ
6432/*
6433 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6434 * tasks if there is an imbalance.
6435 */
6436static int load_balance(int this_cpu, struct rq *this_rq,
6437 struct sched_domain *sd, enum cpu_idle_type idle,
23f0d209 6438 int *continue_balancing)
1e3c88bd 6439{
88b8dac0 6440 int ld_moved, cur_ld_moved, active_balance = 0;
6263322c 6441 struct sched_domain *sd_parent = sd->parent;
1e3c88bd 6442 struct sched_group *group;
1e3c88bd
PZ
6443 struct rq *busiest;
6444 unsigned long flags;
e6252c3e 6445 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
1e3c88bd 6446
8e45cb54
PZ
6447 struct lb_env env = {
6448 .sd = sd,
ddcdf6e7
PZ
6449 .dst_cpu = this_cpu,
6450 .dst_rq = this_rq,
88b8dac0 6451 .dst_grpmask = sched_group_cpus(sd->groups),
8e45cb54 6452 .idle = idle,
eb95308e 6453 .loop_break = sched_nr_migrate_break,
b9403130 6454 .cpus = cpus,
0ec8aa00 6455 .fbq_type = all,
8e45cb54
PZ
6456 };
6457
cfc03118
JK
6458 /*
6459 * For NEWLY_IDLE load_balancing, we don't need to consider
6460 * other cpus in our group
6461 */
e02e60c1 6462 if (idle == CPU_NEWLY_IDLE)
cfc03118 6463 env.dst_grpmask = NULL;
cfc03118 6464
1e3c88bd
PZ
6465 cpumask_copy(cpus, cpu_active_mask);
6466
1e3c88bd
PZ
6467 schedstat_inc(sd, lb_count[idle]);
6468
6469redo:
23f0d209
JK
6470 if (!should_we_balance(&env)) {
6471 *continue_balancing = 0;
1e3c88bd 6472 goto out_balanced;
23f0d209 6473 }
1e3c88bd 6474
23f0d209 6475 group = find_busiest_group(&env);
1e3c88bd
PZ
6476 if (!group) {
6477 schedstat_inc(sd, lb_nobusyg[idle]);
6478 goto out_balanced;
6479 }
6480
b9403130 6481 busiest = find_busiest_queue(&env, group);
1e3c88bd
PZ
6482 if (!busiest) {
6483 schedstat_inc(sd, lb_nobusyq[idle]);
6484 goto out_balanced;
6485 }
6486
78feefc5 6487 BUG_ON(busiest == env.dst_rq);
1e3c88bd 6488
bd939f45 6489 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
1e3c88bd
PZ
6490
6491 ld_moved = 0;
6492 if (busiest->nr_running > 1) {
6493 /*
6494 * Attempt to move tasks. If find_busiest_group has found
6495 * an imbalance but busiest->nr_running <= 1, the group is
6496 * still unbalanced. ld_moved simply stays zero, so it is
6497 * correctly treated as an imbalance.
6498 */
8e45cb54 6499 env.flags |= LBF_ALL_PINNED;
c82513e5
PZ
6500 env.src_cpu = busiest->cpu;
6501 env.src_rq = busiest;
6502 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8e45cb54 6503
5d6523eb 6504more_balance:
1e3c88bd 6505 local_irq_save(flags);
78feefc5 6506 double_rq_lock(env.dst_rq, busiest);
88b8dac0
SV
6507
6508 /*
6509 * cur_ld_moved - load moved in current iteration
6510 * ld_moved - cumulative load moved across iterations
6511 */
6512 cur_ld_moved = move_tasks(&env);
6513 ld_moved += cur_ld_moved;
78feefc5 6514 double_rq_unlock(env.dst_rq, busiest);
1e3c88bd
PZ
6515 local_irq_restore(flags);
6516
6517 /*
6518 * some other cpu did the load balance for us.
6519 */
88b8dac0
SV
6520 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6521 resched_cpu(env.dst_cpu);
6522
f1cd0858
JK
6523 if (env.flags & LBF_NEED_BREAK) {
6524 env.flags &= ~LBF_NEED_BREAK;
6525 goto more_balance;
6526 }
6527
88b8dac0
SV
6528 /*
6529 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6530 * us and move them to an alternate dst_cpu in our sched_group
6531 * where they can run. The upper limit on how many times we
6532 * iterate on same src_cpu is dependent on number of cpus in our
6533 * sched_group.
6534 *
6535 * This changes load balance semantics a bit on who can move
6536 * load to a given_cpu. In addition to the given_cpu itself
6537 * (or a ilb_cpu acting on its behalf where given_cpu is
6538 * nohz-idle), we now have balance_cpu in a position to move
6539 * load to given_cpu. In rare situations, this may cause
6540 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6541 * _independently_ and at _same_ time to move some load to
6542 * given_cpu) causing exceess load to be moved to given_cpu.
6543 * This however should not happen so much in practice and
6544 * moreover subsequent load balance cycles should correct the
6545 * excess load moved.
6546 */
6263322c 6547 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
88b8dac0 6548
7aff2e3a
VD
6549 /* Prevent to re-select dst_cpu via env's cpus */
6550 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6551
78feefc5 6552 env.dst_rq = cpu_rq(env.new_dst_cpu);
88b8dac0 6553 env.dst_cpu = env.new_dst_cpu;
6263322c 6554 env.flags &= ~LBF_DST_PINNED;
88b8dac0
SV
6555 env.loop = 0;
6556 env.loop_break = sched_nr_migrate_break;
e02e60c1 6557
88b8dac0
SV
6558 /*
6559 * Go back to "more_balance" rather than "redo" since we
6560 * need to continue with same src_cpu.
6561 */
6562 goto more_balance;
6563 }
1e3c88bd 6564
6263322c
PZ
6565 /*
6566 * We failed to reach balance because of affinity.
6567 */
6568 if (sd_parent) {
6569 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6570
6571 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6572 *group_imbalance = 1;
6573 } else if (*group_imbalance)
6574 *group_imbalance = 0;
6575 }
6576
1e3c88bd 6577 /* All tasks on this runqueue were pinned by CPU affinity */
8e45cb54 6578 if (unlikely(env.flags & LBF_ALL_PINNED)) {
1e3c88bd 6579 cpumask_clear_cpu(cpu_of(busiest), cpus);
bbf18b19
PN
6580 if (!cpumask_empty(cpus)) {
6581 env.loop = 0;
6582 env.loop_break = sched_nr_migrate_break;
1e3c88bd 6583 goto redo;
bbf18b19 6584 }
1e3c88bd
PZ
6585 goto out_balanced;
6586 }
6587 }
6588
6589 if (!ld_moved) {
6590 schedstat_inc(sd, lb_failed[idle]);
58b26c4c
VP
6591 /*
6592 * Increment the failure counter only on periodic balance.
6593 * We do not want newidle balance, which can be very
6594 * frequent, pollute the failure counter causing
6595 * excessive cache_hot migrations and active balances.
6596 */
6597 if (idle != CPU_NEWLY_IDLE)
6598 sd->nr_balance_failed++;
1e3c88bd 6599
bd939f45 6600 if (need_active_balance(&env)) {
1e3c88bd
PZ
6601 raw_spin_lock_irqsave(&busiest->lock, flags);
6602
969c7921
TH
6603 /* don't kick the active_load_balance_cpu_stop,
6604 * if the curr task on busiest cpu can't be
6605 * moved to this_cpu
1e3c88bd
PZ
6606 */
6607 if (!cpumask_test_cpu(this_cpu,
fa17b507 6608 tsk_cpus_allowed(busiest->curr))) {
1e3c88bd
PZ
6609 raw_spin_unlock_irqrestore(&busiest->lock,
6610 flags);
8e45cb54 6611 env.flags |= LBF_ALL_PINNED;
1e3c88bd
PZ
6612 goto out_one_pinned;
6613 }
6614
969c7921
TH
6615 /*
6616 * ->active_balance synchronizes accesses to
6617 * ->active_balance_work. Once set, it's cleared
6618 * only after active load balance is finished.
6619 */
1e3c88bd
PZ
6620 if (!busiest->active_balance) {
6621 busiest->active_balance = 1;
6622 busiest->push_cpu = this_cpu;
6623 active_balance = 1;
6624 }
6625 raw_spin_unlock_irqrestore(&busiest->lock, flags);
969c7921 6626
bd939f45 6627 if (active_balance) {
969c7921
TH
6628 stop_one_cpu_nowait(cpu_of(busiest),
6629 active_load_balance_cpu_stop, busiest,
6630 &busiest->active_balance_work);
bd939f45 6631 }
1e3c88bd
PZ
6632
6633 /*
6634 * We've kicked active balancing, reset the failure
6635 * counter.
6636 */
6637 sd->nr_balance_failed = sd->cache_nice_tries+1;
6638 }
6639 } else
6640 sd->nr_balance_failed = 0;
6641
6642 if (likely(!active_balance)) {
6643 /* We were unbalanced, so reset the balancing interval */
6644 sd->balance_interval = sd->min_interval;
6645 } else {
6646 /*
6647 * If we've begun active balancing, start to back off. This
6648 * case may not be covered by the all_pinned logic if there
6649 * is only 1 task on the busy runqueue (because we don't call
6650 * move_tasks).
6651 */
6652 if (sd->balance_interval < sd->max_interval)
6653 sd->balance_interval *= 2;
6654 }
6655
1e3c88bd
PZ
6656 goto out;
6657
6658out_balanced:
6659 schedstat_inc(sd, lb_balanced[idle]);
6660
6661 sd->nr_balance_failed = 0;
6662
6663out_one_pinned:
6664 /* tune up the balancing interval */
8e45cb54 6665 if (((env.flags & LBF_ALL_PINNED) &&
5b54b56b 6666 sd->balance_interval < MAX_PINNED_INTERVAL) ||
1e3c88bd
PZ
6667 (sd->balance_interval < sd->max_interval))
6668 sd->balance_interval *= 2;
6669
46e49b38 6670 ld_moved = 0;
1e3c88bd 6671out:
1e3c88bd
PZ
6672 return ld_moved;
6673}
6674
1e3c88bd
PZ
6675/*
6676 * idle_balance is called by schedule() if this_cpu is about to become
6677 * idle. Attempts to pull tasks from other CPUs.
6678 */
6e83125c 6679static int idle_balance(struct rq *this_rq)
1e3c88bd
PZ
6680{
6681 struct sched_domain *sd;
6682 int pulled_task = 0;
6683 unsigned long next_balance = jiffies + HZ;
9bd721c5 6684 u64 curr_cost = 0;
b4f2ab43 6685 int this_cpu = this_rq->cpu;
1e3c88bd 6686
6e83125c 6687 idle_enter_fair(this_rq);
0e5b5337 6688
6e83125c
PZ
6689 /*
6690 * We must set idle_stamp _before_ calling idle_balance(), such that we
6691 * measure the duration of idle_balance() as idle time.
6692 */
6693 this_rq->idle_stamp = rq_clock(this_rq);
6694
1e3c88bd 6695 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6e83125c 6696 goto out;
1e3c88bd 6697
f492e12e
PZ
6698 /*
6699 * Drop the rq->lock, but keep IRQ/preempt disabled.
6700 */
6701 raw_spin_unlock(&this_rq->lock);
6702
48a16753 6703 update_blocked_averages(this_cpu);
dce840a0 6704 rcu_read_lock();
1e3c88bd
PZ
6705 for_each_domain(this_cpu, sd) {
6706 unsigned long interval;
23f0d209 6707 int continue_balancing = 1;
9bd721c5 6708 u64 t0, domain_cost;
1e3c88bd
PZ
6709
6710 if (!(sd->flags & SD_LOAD_BALANCE))
6711 continue;
6712
9bd721c5
JL
6713 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6714 break;
6715
f492e12e 6716 if (sd->flags & SD_BALANCE_NEWIDLE) {
9bd721c5
JL
6717 t0 = sched_clock_cpu(this_cpu);
6718
f492e12e 6719 pulled_task = load_balance(this_cpu, this_rq,
23f0d209
JK
6720 sd, CPU_NEWLY_IDLE,
6721 &continue_balancing);
9bd721c5
JL
6722
6723 domain_cost = sched_clock_cpu(this_cpu) - t0;
6724 if (domain_cost > sd->max_newidle_lb_cost)
6725 sd->max_newidle_lb_cost = domain_cost;
6726
6727 curr_cost += domain_cost;
f492e12e 6728 }
1e3c88bd
PZ
6729
6730 interval = msecs_to_jiffies(sd->balance_interval);
6731 if (time_after(next_balance, sd->last_balance + interval))
6732 next_balance = sd->last_balance + interval;
39a4d9ca
JL
6733
6734 /*
6735 * Stop searching for tasks to pull if there are
6736 * now runnable tasks on this rq.
6737 */
6738 if (pulled_task || this_rq->nr_running > 0)
1e3c88bd 6739 break;
1e3c88bd 6740 }
dce840a0 6741 rcu_read_unlock();
f492e12e
PZ
6742
6743 raw_spin_lock(&this_rq->lock);
6744
0e5b5337
JL
6745 if (curr_cost > this_rq->max_idle_balance_cost)
6746 this_rq->max_idle_balance_cost = curr_cost;
6747
e5fc6611 6748 /*
0e5b5337
JL
6749 * While browsing the domains, we released the rq lock, a task could
6750 * have been enqueued in the meantime. Since we're not going idle,
6751 * pretend we pulled a task.
e5fc6611 6752 */
0e5b5337 6753 if (this_rq->cfs.h_nr_running && !pulled_task)
6e83125c 6754 pulled_task = 1;
e5fc6611 6755
1e3c88bd
PZ
6756 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6757 /*
6758 * We are going idle. next_balance may be set based on
6759 * a busy processor. So reset next_balance.
6760 */
6761 this_rq->next_balance = next_balance;
6762 }
9bd721c5 6763
6e83125c 6764out:
e4aa358b 6765 /* Is there a task of a high priority class? */
46383648 6766 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
e4aa358b
KT
6767 pulled_task = -1;
6768
6769 if (pulled_task) {
6770 idle_exit_fair(this_rq);
6e83125c 6771 this_rq->idle_stamp = 0;
e4aa358b 6772 }
6e83125c 6773
3c4017c1 6774 return pulled_task;
1e3c88bd
PZ
6775}
6776
6777/*
969c7921
TH
6778 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6779 * running tasks off the busiest CPU onto idle CPUs. It requires at
6780 * least 1 task to be running on each physical CPU where possible, and
6781 * avoids physical / logical imbalances.
1e3c88bd 6782 */
969c7921 6783static int active_load_balance_cpu_stop(void *data)
1e3c88bd 6784{
969c7921
TH
6785 struct rq *busiest_rq = data;
6786 int busiest_cpu = cpu_of(busiest_rq);
1e3c88bd 6787 int target_cpu = busiest_rq->push_cpu;
969c7921 6788 struct rq *target_rq = cpu_rq(target_cpu);
1e3c88bd 6789 struct sched_domain *sd;
969c7921
TH
6790
6791 raw_spin_lock_irq(&busiest_rq->lock);
6792
6793 /* make sure the requested cpu hasn't gone down in the meantime */
6794 if (unlikely(busiest_cpu != smp_processor_id() ||
6795 !busiest_rq->active_balance))
6796 goto out_unlock;
1e3c88bd
PZ
6797
6798 /* Is there any task to move? */
6799 if (busiest_rq->nr_running <= 1)
969c7921 6800 goto out_unlock;
1e3c88bd
PZ
6801
6802 /*
6803 * This condition is "impossible", if it occurs
6804 * we need to fix it. Originally reported by
6805 * Bjorn Helgaas on a 128-cpu setup.
6806 */
6807 BUG_ON(busiest_rq == target_rq);
6808
6809 /* move a task from busiest_rq to target_rq */
6810 double_lock_balance(busiest_rq, target_rq);
1e3c88bd
PZ
6811
6812 /* Search for an sd spanning us and the target CPU. */
dce840a0 6813 rcu_read_lock();
1e3c88bd
PZ
6814 for_each_domain(target_cpu, sd) {
6815 if ((sd->flags & SD_LOAD_BALANCE) &&
6816 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6817 break;
6818 }
6819
6820 if (likely(sd)) {
8e45cb54
PZ
6821 struct lb_env env = {
6822 .sd = sd,
ddcdf6e7
PZ
6823 .dst_cpu = target_cpu,
6824 .dst_rq = target_rq,
6825 .src_cpu = busiest_rq->cpu,
6826 .src_rq = busiest_rq,
8e45cb54
PZ
6827 .idle = CPU_IDLE,
6828 };
6829
1e3c88bd
PZ
6830 schedstat_inc(sd, alb_count);
6831
8e45cb54 6832 if (move_one_task(&env))
1e3c88bd
PZ
6833 schedstat_inc(sd, alb_pushed);
6834 else
6835 schedstat_inc(sd, alb_failed);
6836 }
dce840a0 6837 rcu_read_unlock();
1e3c88bd 6838 double_unlock_balance(busiest_rq, target_rq);
969c7921
TH
6839out_unlock:
6840 busiest_rq->active_balance = 0;
6841 raw_spin_unlock_irq(&busiest_rq->lock);
6842 return 0;
1e3c88bd
PZ
6843}
6844
d987fc7f
MG
6845static inline int on_null_domain(struct rq *rq)
6846{
6847 return unlikely(!rcu_dereference_sched(rq->sd));
6848}
6849
3451d024 6850#ifdef CONFIG_NO_HZ_COMMON
83cd4fe2
VP
6851/*
6852 * idle load balancing details
83cd4fe2
VP
6853 * - When one of the busy CPUs notice that there may be an idle rebalancing
6854 * needed, they will kick the idle load balancer, which then does idle
6855 * load balancing for all the idle CPUs.
6856 */
1e3c88bd 6857static struct {
83cd4fe2 6858 cpumask_var_t idle_cpus_mask;
0b005cf5 6859 atomic_t nr_cpus;
83cd4fe2
VP
6860 unsigned long next_balance; /* in jiffy units */
6861} nohz ____cacheline_aligned;
1e3c88bd 6862
3dd0337d 6863static inline int find_new_ilb(void)
1e3c88bd 6864{
0b005cf5 6865 int ilb = cpumask_first(nohz.idle_cpus_mask);
1e3c88bd 6866
786d6dc7
SS
6867 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6868 return ilb;
6869
6870 return nr_cpu_ids;
1e3c88bd 6871}
1e3c88bd 6872
83cd4fe2
VP
6873/*
6874 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6875 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6876 * CPU (if there is one).
6877 */
0aeeeeba 6878static void nohz_balancer_kick(void)
83cd4fe2
VP
6879{
6880 int ilb_cpu;
6881
6882 nohz.next_balance++;
6883
3dd0337d 6884 ilb_cpu = find_new_ilb();
83cd4fe2 6885
0b005cf5
SS
6886 if (ilb_cpu >= nr_cpu_ids)
6887 return;
83cd4fe2 6888
cd490c5b 6889 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
1c792db7
SS
6890 return;
6891 /*
6892 * Use smp_send_reschedule() instead of resched_cpu().
6893 * This way we generate a sched IPI on the target cpu which
6894 * is idle. And the softirq performing nohz idle load balance
6895 * will be run before returning from the IPI.
6896 */
6897 smp_send_reschedule(ilb_cpu);
83cd4fe2
VP
6898 return;
6899}
6900
c1cc017c 6901static inline void nohz_balance_exit_idle(int cpu)
71325960
SS
6902{
6903 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
d987fc7f
MG
6904 /*
6905 * Completely isolated CPUs don't ever set, so we must test.
6906 */
6907 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6908 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6909 atomic_dec(&nohz.nr_cpus);
6910 }
71325960
SS
6911 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6912 }
6913}
6914
69e1e811
SS
6915static inline void set_cpu_sd_state_busy(void)
6916{
6917 struct sched_domain *sd;
37dc6b50 6918 int cpu = smp_processor_id();
69e1e811 6919
69e1e811 6920 rcu_read_lock();
37dc6b50 6921 sd = rcu_dereference(per_cpu(sd_busy, cpu));
25f55d9d
VG
6922
6923 if (!sd || !sd->nohz_idle)
6924 goto unlock;
6925 sd->nohz_idle = 0;
6926
37dc6b50 6927 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
25f55d9d 6928unlock:
69e1e811
SS
6929 rcu_read_unlock();
6930}
6931
6932void set_cpu_sd_state_idle(void)
6933{
6934 struct sched_domain *sd;
37dc6b50 6935 int cpu = smp_processor_id();
69e1e811 6936
69e1e811 6937 rcu_read_lock();
37dc6b50 6938 sd = rcu_dereference(per_cpu(sd_busy, cpu));
25f55d9d
VG
6939
6940 if (!sd || sd->nohz_idle)
6941 goto unlock;
6942 sd->nohz_idle = 1;
6943
37dc6b50 6944 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
25f55d9d 6945unlock:
69e1e811
SS
6946 rcu_read_unlock();
6947}
6948
1e3c88bd 6949/*
c1cc017c 6950 * This routine will record that the cpu is going idle with tick stopped.
0b005cf5 6951 * This info will be used in performing idle load balancing in the future.
1e3c88bd 6952 */
c1cc017c 6953void nohz_balance_enter_idle(int cpu)
1e3c88bd 6954{
71325960
SS
6955 /*
6956 * If this cpu is going down, then nothing needs to be done.
6957 */
6958 if (!cpu_active(cpu))
6959 return;
6960
c1cc017c
AS
6961 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6962 return;
1e3c88bd 6963
d987fc7f
MG
6964 /*
6965 * If we're a completely isolated CPU, we don't play.
6966 */
6967 if (on_null_domain(cpu_rq(cpu)))
6968 return;
6969
c1cc017c
AS
6970 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6971 atomic_inc(&nohz.nr_cpus);
6972 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
1e3c88bd 6973}
71325960 6974
0db0628d 6975static int sched_ilb_notifier(struct notifier_block *nfb,
71325960
SS
6976 unsigned long action, void *hcpu)
6977{
6978 switch (action & ~CPU_TASKS_FROZEN) {
6979 case CPU_DYING:
c1cc017c 6980 nohz_balance_exit_idle(smp_processor_id());
71325960
SS
6981 return NOTIFY_OK;
6982 default:
6983 return NOTIFY_DONE;
6984 }
6985}
1e3c88bd
PZ
6986#endif
6987
6988static DEFINE_SPINLOCK(balancing);
6989
49c022e6
PZ
6990/*
6991 * Scale the max load_balance interval with the number of CPUs in the system.
6992 * This trades load-balance latency on larger machines for less cross talk.
6993 */
029632fb 6994void update_max_interval(void)
49c022e6
PZ
6995{
6996 max_load_balance_interval = HZ*num_online_cpus()/10;
6997}
6998
1e3c88bd
PZ
6999/*
7000 * It checks each scheduling domain to see if it is due to be balanced,
7001 * and initiates a balancing operation if so.
7002 *
b9b0853a 7003 * Balancing parameters are set up in init_sched_domains.
1e3c88bd 7004 */
f7ed0a89 7005static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
1e3c88bd 7006{
23f0d209 7007 int continue_balancing = 1;
f7ed0a89 7008 int cpu = rq->cpu;
1e3c88bd 7009 unsigned long interval;
04f733b4 7010 struct sched_domain *sd;
1e3c88bd
PZ
7011 /* Earliest time when we have to do rebalance again */
7012 unsigned long next_balance = jiffies + 60*HZ;
7013 int update_next_balance = 0;
f48627e6
JL
7014 int need_serialize, need_decay = 0;
7015 u64 max_cost = 0;
1e3c88bd 7016
48a16753 7017 update_blocked_averages(cpu);
2069dd75 7018
dce840a0 7019 rcu_read_lock();
1e3c88bd 7020 for_each_domain(cpu, sd) {
f48627e6
JL
7021 /*
7022 * Decay the newidle max times here because this is a regular
7023 * visit to all the domains. Decay ~1% per second.
7024 */
7025 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7026 sd->max_newidle_lb_cost =
7027 (sd->max_newidle_lb_cost * 253) / 256;
7028 sd->next_decay_max_lb_cost = jiffies + HZ;
7029 need_decay = 1;
7030 }
7031 max_cost += sd->max_newidle_lb_cost;
7032
1e3c88bd
PZ
7033 if (!(sd->flags & SD_LOAD_BALANCE))
7034 continue;
7035
f48627e6
JL
7036 /*
7037 * Stop the load balance at this level. There is another
7038 * CPU in our sched group which is doing load balancing more
7039 * actively.
7040 */
7041 if (!continue_balancing) {
7042 if (need_decay)
7043 continue;
7044 break;
7045 }
7046
1e3c88bd
PZ
7047 interval = sd->balance_interval;
7048 if (idle != CPU_IDLE)
7049 interval *= sd->busy_factor;
7050
7051 /* scale ms to jiffies */
7052 interval = msecs_to_jiffies(interval);
49c022e6 7053 interval = clamp(interval, 1UL, max_load_balance_interval);
1e3c88bd
PZ
7054
7055 need_serialize = sd->flags & SD_SERIALIZE;
7056
7057 if (need_serialize) {
7058 if (!spin_trylock(&balancing))
7059 goto out;
7060 }
7061
7062 if (time_after_eq(jiffies, sd->last_balance + interval)) {
23f0d209 7063 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
1e3c88bd 7064 /*
6263322c 7065 * The LBF_DST_PINNED logic could have changed
de5eb2dd
JK
7066 * env->dst_cpu, so we can't know our idle
7067 * state even if we migrated tasks. Update it.
1e3c88bd 7068 */
de5eb2dd 7069 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
1e3c88bd
PZ
7070 }
7071 sd->last_balance = jiffies;
7072 }
7073 if (need_serialize)
7074 spin_unlock(&balancing);
7075out:
7076 if (time_after(next_balance, sd->last_balance + interval)) {
7077 next_balance = sd->last_balance + interval;
7078 update_next_balance = 1;
7079 }
f48627e6
JL
7080 }
7081 if (need_decay) {
1e3c88bd 7082 /*
f48627e6
JL
7083 * Ensure the rq-wide value also decays but keep it at a
7084 * reasonable floor to avoid funnies with rq->avg_idle.
1e3c88bd 7085 */
f48627e6
JL
7086 rq->max_idle_balance_cost =
7087 max((u64)sysctl_sched_migration_cost, max_cost);
1e3c88bd 7088 }
dce840a0 7089 rcu_read_unlock();
1e3c88bd
PZ
7090
7091 /*
7092 * next_balance will be updated only when there is a need.
7093 * When the cpu is attached to null domain for ex, it will not be
7094 * updated.
7095 */
7096 if (likely(update_next_balance))
7097 rq->next_balance = next_balance;
7098}
7099
3451d024 7100#ifdef CONFIG_NO_HZ_COMMON
1e3c88bd 7101/*
3451d024 7102 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
1e3c88bd
PZ
7103 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7104 */
208cb16b 7105static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
83cd4fe2 7106{
208cb16b 7107 int this_cpu = this_rq->cpu;
83cd4fe2
VP
7108 struct rq *rq;
7109 int balance_cpu;
7110
1c792db7
SS
7111 if (idle != CPU_IDLE ||
7112 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7113 goto end;
83cd4fe2
VP
7114
7115 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8a6d42d1 7116 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
83cd4fe2
VP
7117 continue;
7118
7119 /*
7120 * If this cpu gets work to do, stop the load balancing
7121 * work being done for other cpus. Next load
7122 * balancing owner will pick it up.
7123 */
1c792db7 7124 if (need_resched())
83cd4fe2 7125 break;
83cd4fe2 7126
5ed4f1d9
VG
7127 rq = cpu_rq(balance_cpu);
7128
7129 raw_spin_lock_irq(&rq->lock);
7130 update_rq_clock(rq);
7131 update_idle_cpu_load(rq);
7132 raw_spin_unlock_irq(&rq->lock);
83cd4fe2 7133
f7ed0a89 7134 rebalance_domains(rq, CPU_IDLE);
83cd4fe2 7135
83cd4fe2
VP
7136 if (time_after(this_rq->next_balance, rq->next_balance))
7137 this_rq->next_balance = rq->next_balance;
7138 }
7139 nohz.next_balance = this_rq->next_balance;
1c792db7
SS
7140end:
7141 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
83cd4fe2
VP
7142}
7143
7144/*
0b005cf5
SS
7145 * Current heuristic for kicking the idle load balancer in the presence
7146 * of an idle cpu is the system.
7147 * - This rq has more than one task.
7148 * - At any scheduler domain level, this cpu's scheduler group has multiple
7149 * busy cpu's exceeding the group's power.
7150 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7151 * domain span are idle.
83cd4fe2 7152 */
4a725627 7153static inline int nohz_kick_needed(struct rq *rq)
83cd4fe2
VP
7154{
7155 unsigned long now = jiffies;
0b005cf5 7156 struct sched_domain *sd;
37dc6b50 7157 struct sched_group_power *sgp;
4a725627 7158 int nr_busy, cpu = rq->cpu;
83cd4fe2 7159
4a725627 7160 if (unlikely(rq->idle_balance))
83cd4fe2
VP
7161 return 0;
7162
1c792db7
SS
7163 /*
7164 * We may be recently in ticked or tickless idle mode. At the first
7165 * busy tick after returning from idle, we will update the busy stats.
7166 */
69e1e811 7167 set_cpu_sd_state_busy();
c1cc017c 7168 nohz_balance_exit_idle(cpu);
0b005cf5
SS
7169
7170 /*
7171 * None are in tickless mode and hence no need for NOHZ idle load
7172 * balancing.
7173 */
7174 if (likely(!atomic_read(&nohz.nr_cpus)))
7175 return 0;
1c792db7
SS
7176
7177 if (time_before(now, nohz.next_balance))
83cd4fe2
VP
7178 return 0;
7179
0b005cf5
SS
7180 if (rq->nr_running >= 2)
7181 goto need_kick;
83cd4fe2 7182
067491b7 7183 rcu_read_lock();
37dc6b50 7184 sd = rcu_dereference(per_cpu(sd_busy, cpu));
83cd4fe2 7185
37dc6b50
PM
7186 if (sd) {
7187 sgp = sd->groups->sgp;
7188 nr_busy = atomic_read(&sgp->nr_busy_cpus);
0b005cf5 7189
37dc6b50 7190 if (nr_busy > 1)
067491b7 7191 goto need_kick_unlock;
83cd4fe2 7192 }
37dc6b50
PM
7193
7194 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7195
7196 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7197 sched_domain_span(sd)) < cpu))
7198 goto need_kick_unlock;
7199
067491b7 7200 rcu_read_unlock();
83cd4fe2 7201 return 0;
067491b7
PZ
7202
7203need_kick_unlock:
7204 rcu_read_unlock();
0b005cf5
SS
7205need_kick:
7206 return 1;
83cd4fe2
VP
7207}
7208#else
208cb16b 7209static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
83cd4fe2
VP
7210#endif
7211
7212/*
7213 * run_rebalance_domains is triggered when needed from the scheduler tick.
7214 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7215 */
1e3c88bd
PZ
7216static void run_rebalance_domains(struct softirq_action *h)
7217{
208cb16b 7218 struct rq *this_rq = this_rq();
6eb57e0d 7219 enum cpu_idle_type idle = this_rq->idle_balance ?
1e3c88bd
PZ
7220 CPU_IDLE : CPU_NOT_IDLE;
7221
f7ed0a89 7222 rebalance_domains(this_rq, idle);
1e3c88bd 7223
1e3c88bd 7224 /*
83cd4fe2 7225 * If this cpu has a pending nohz_balance_kick, then do the
1e3c88bd
PZ
7226 * balancing on behalf of the other idle cpus whose ticks are
7227 * stopped.
7228 */
208cb16b 7229 nohz_idle_balance(this_rq, idle);
1e3c88bd
PZ
7230}
7231
1e3c88bd
PZ
7232/*
7233 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
1e3c88bd 7234 */
7caff66f 7235void trigger_load_balance(struct rq *rq)
1e3c88bd 7236{
1e3c88bd 7237 /* Don't need to rebalance while attached to NULL domain */
c726099e
DL
7238 if (unlikely(on_null_domain(rq)))
7239 return;
7240
7241 if (time_after_eq(jiffies, rq->next_balance))
1e3c88bd 7242 raise_softirq(SCHED_SOFTIRQ);
3451d024 7243#ifdef CONFIG_NO_HZ_COMMON
c726099e 7244 if (nohz_kick_needed(rq))
0aeeeeba 7245 nohz_balancer_kick();
83cd4fe2 7246#endif
1e3c88bd
PZ
7247}
7248
0bcdcf28
CE
7249static void rq_online_fair(struct rq *rq)
7250{
7251 update_sysctl();
7252}
7253
7254static void rq_offline_fair(struct rq *rq)
7255{
7256 update_sysctl();
a4c96ae3
PB
7257
7258 /* Ensure any throttled groups are reachable by pick_next_task */
7259 unthrottle_offline_cfs_rqs(rq);
0bcdcf28
CE
7260}
7261
55e12e5e 7262#endif /* CONFIG_SMP */
e1d1484f 7263
bf0f6f24
IM
7264/*
7265 * scheduler tick hitting a task of our scheduling class:
7266 */
8f4d37ec 7267static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
bf0f6f24
IM
7268{
7269 struct cfs_rq *cfs_rq;
7270 struct sched_entity *se = &curr->se;
7271
7272 for_each_sched_entity(se) {
7273 cfs_rq = cfs_rq_of(se);
8f4d37ec 7274 entity_tick(cfs_rq, se, queued);
bf0f6f24 7275 }
18bf2805 7276
10e84b97 7277 if (numabalancing_enabled)
cbee9f88 7278 task_tick_numa(rq, curr);
3d59eebc 7279
18bf2805 7280 update_rq_runnable_avg(rq, 1);
bf0f6f24
IM
7281}
7282
7283/*
cd29fe6f
PZ
7284 * called on fork with the child task as argument from the parent's context
7285 * - child not yet on the tasklist
7286 * - preemption disabled
bf0f6f24 7287 */
cd29fe6f 7288static void task_fork_fair(struct task_struct *p)
bf0f6f24 7289{
4fc420c9
DN
7290 struct cfs_rq *cfs_rq;
7291 struct sched_entity *se = &p->se, *curr;
00bf7bfc 7292 int this_cpu = smp_processor_id();
cd29fe6f
PZ
7293 struct rq *rq = this_rq();
7294 unsigned long flags;
7295
05fa785c 7296 raw_spin_lock_irqsave(&rq->lock, flags);
bf0f6f24 7297
861d034e
PZ
7298 update_rq_clock(rq);
7299
4fc420c9
DN
7300 cfs_rq = task_cfs_rq(current);
7301 curr = cfs_rq->curr;
7302
6c9a27f5
DN
7303 /*
7304 * Not only the cpu but also the task_group of the parent might have
7305 * been changed after parent->se.parent,cfs_rq were copied to
7306 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7307 * of child point to valid ones.
7308 */
7309 rcu_read_lock();
7310 __set_task_cpu(p, this_cpu);
7311 rcu_read_unlock();
bf0f6f24 7312
7109c442 7313 update_curr(cfs_rq);
cd29fe6f 7314
b5d9d734
MG
7315 if (curr)
7316 se->vruntime = curr->vruntime;
aeb73b04 7317 place_entity(cfs_rq, se, 1);
4d78e7b6 7318
cd29fe6f 7319 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
87fefa38 7320 /*
edcb60a3
IM
7321 * Upon rescheduling, sched_class::put_prev_task() will place
7322 * 'current' within the tree based on its new key value.
7323 */
4d78e7b6 7324 swap(curr->vruntime, se->vruntime);
aec0a514 7325 resched_task(rq->curr);
4d78e7b6 7326 }
bf0f6f24 7327
88ec22d3
PZ
7328 se->vruntime -= cfs_rq->min_vruntime;
7329
05fa785c 7330 raw_spin_unlock_irqrestore(&rq->lock, flags);
bf0f6f24
IM
7331}
7332
cb469845
SR
7333/*
7334 * Priority of the task has changed. Check to see if we preempt
7335 * the current task.
7336 */
da7a735e
PZ
7337static void
7338prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
cb469845 7339{
da7a735e
PZ
7340 if (!p->se.on_rq)
7341 return;
7342
cb469845
SR
7343 /*
7344 * Reschedule if we are currently running on this runqueue and
7345 * our priority decreased, or if we are not currently running on
7346 * this runqueue and our priority is higher than the current's
7347 */
da7a735e 7348 if (rq->curr == p) {
cb469845
SR
7349 if (p->prio > oldprio)
7350 resched_task(rq->curr);
7351 } else
15afe09b 7352 check_preempt_curr(rq, p, 0);
cb469845
SR
7353}
7354
da7a735e
PZ
7355static void switched_from_fair(struct rq *rq, struct task_struct *p)
7356{
7357 struct sched_entity *se = &p->se;
7358 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7359
7360 /*
791c9e02 7361 * Ensure the task's vruntime is normalized, so that when it's
da7a735e
PZ
7362 * switched back to the fair class the enqueue_entity(.flags=0) will
7363 * do the right thing.
7364 *
791c9e02
GM
7365 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7366 * have normalized the vruntime, if it's !on_rq, then only when
da7a735e
PZ
7367 * the task is sleeping will it still have non-normalized vruntime.
7368 */
791c9e02 7369 if (!p->on_rq && p->state != TASK_RUNNING) {
da7a735e
PZ
7370 /*
7371 * Fix up our vruntime so that the current sleep doesn't
7372 * cause 'unlimited' sleep bonus.
7373 */
7374 place_entity(cfs_rq, se, 0);
7375 se->vruntime -= cfs_rq->min_vruntime;
7376 }
9ee474f5 7377
141965c7 7378#ifdef CONFIG_SMP
9ee474f5
PT
7379 /*
7380 * Remove our load from contribution when we leave sched_fair
7381 * and ensure we don't carry in an old decay_count if we
7382 * switch back.
7383 */
87e3c8ae
KT
7384 if (se->avg.decay_count) {
7385 __synchronize_entity_decay(se);
7386 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
9ee474f5
PT
7387 }
7388#endif
da7a735e
PZ
7389}
7390
cb469845
SR
7391/*
7392 * We switched to the sched_fair class.
7393 */
da7a735e 7394static void switched_to_fair(struct rq *rq, struct task_struct *p)
cb469845 7395{
eb7a59b2
M
7396 struct sched_entity *se = &p->se;
7397#ifdef CONFIG_FAIR_GROUP_SCHED
7398 /*
7399 * Since the real-depth could have been changed (only FAIR
7400 * class maintain depth value), reset depth properly.
7401 */
7402 se->depth = se->parent ? se->parent->depth + 1 : 0;
7403#endif
7404 if (!se->on_rq)
da7a735e
PZ
7405 return;
7406
cb469845
SR
7407 /*
7408 * We were most likely switched from sched_rt, so
7409 * kick off the schedule if running, otherwise just see
7410 * if we can still preempt the current task.
7411 */
da7a735e 7412 if (rq->curr == p)
cb469845
SR
7413 resched_task(rq->curr);
7414 else
15afe09b 7415 check_preempt_curr(rq, p, 0);
cb469845
SR
7416}
7417
83b699ed
SV
7418/* Account for a task changing its policy or group.
7419 *
7420 * This routine is mostly called to set cfs_rq->curr field when a task
7421 * migrates between groups/classes.
7422 */
7423static void set_curr_task_fair(struct rq *rq)
7424{
7425 struct sched_entity *se = &rq->curr->se;
7426
ec12cb7f
PT
7427 for_each_sched_entity(se) {
7428 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7429
7430 set_next_entity(cfs_rq, se);
7431 /* ensure bandwidth has been allocated on our new cfs_rq */
7432 account_cfs_rq_runtime(cfs_rq, 0);
7433 }
83b699ed
SV
7434}
7435
029632fb
PZ
7436void init_cfs_rq(struct cfs_rq *cfs_rq)
7437{
7438 cfs_rq->tasks_timeline = RB_ROOT;
029632fb
PZ
7439 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7440#ifndef CONFIG_64BIT
7441 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7442#endif
141965c7 7443#ifdef CONFIG_SMP
9ee474f5 7444 atomic64_set(&cfs_rq->decay_counter, 1);
2509940f 7445 atomic_long_set(&cfs_rq->removed_load, 0);
9ee474f5 7446#endif
029632fb
PZ
7447}
7448
810b3817 7449#ifdef CONFIG_FAIR_GROUP_SCHED
b2b5ce02 7450static void task_move_group_fair(struct task_struct *p, int on_rq)
810b3817 7451{
fed14d45 7452 struct sched_entity *se = &p->se;
aff3e498 7453 struct cfs_rq *cfs_rq;
fed14d45 7454
b2b5ce02
PZ
7455 /*
7456 * If the task was not on the rq at the time of this cgroup movement
7457 * it must have been asleep, sleeping tasks keep their ->vruntime
7458 * absolute on their old rq until wakeup (needed for the fair sleeper
7459 * bonus in place_entity()).
7460 *
7461 * If it was on the rq, we've just 'preempted' it, which does convert
7462 * ->vruntime to a relative base.
7463 *
7464 * Make sure both cases convert their relative position when migrating
7465 * to another cgroup's rq. This does somewhat interfere with the
7466 * fair sleeper stuff for the first placement, but who cares.
7467 */
7ceff013
DN
7468 /*
7469 * When !on_rq, vruntime of the task has usually NOT been normalized.
7470 * But there are some cases where it has already been normalized:
7471 *
7472 * - Moving a forked child which is waiting for being woken up by
7473 * wake_up_new_task().
62af3783
DN
7474 * - Moving a task which has been woken up by try_to_wake_up() and
7475 * waiting for actually being woken up by sched_ttwu_pending().
7ceff013
DN
7476 *
7477 * To prevent boost or penalty in the new cfs_rq caused by delta
7478 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7479 */
fed14d45 7480 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7ceff013
DN
7481 on_rq = 1;
7482
b2b5ce02 7483 if (!on_rq)
fed14d45 7484 se->vruntime -= cfs_rq_of(se)->min_vruntime;
b2b5ce02 7485 set_task_rq(p, task_cpu(p));
fed14d45 7486 se->depth = se->parent ? se->parent->depth + 1 : 0;
aff3e498 7487 if (!on_rq) {
fed14d45
PZ
7488 cfs_rq = cfs_rq_of(se);
7489 se->vruntime += cfs_rq->min_vruntime;
aff3e498
PT
7490#ifdef CONFIG_SMP
7491 /*
7492 * migrate_task_rq_fair() will have removed our previous
7493 * contribution, but we must synchronize for ongoing future
7494 * decay.
7495 */
fed14d45
PZ
7496 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7497 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
aff3e498
PT
7498#endif
7499 }
810b3817 7500}
029632fb
PZ
7501
7502void free_fair_sched_group(struct task_group *tg)
7503{
7504 int i;
7505
7506 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7507
7508 for_each_possible_cpu(i) {
7509 if (tg->cfs_rq)
7510 kfree(tg->cfs_rq[i]);
7511 if (tg->se)
7512 kfree(tg->se[i]);
7513 }
7514
7515 kfree(tg->cfs_rq);
7516 kfree(tg->se);
7517}
7518
7519int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7520{
7521 struct cfs_rq *cfs_rq;
7522 struct sched_entity *se;
7523 int i;
7524
7525 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7526 if (!tg->cfs_rq)
7527 goto err;
7528 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7529 if (!tg->se)
7530 goto err;
7531
7532 tg->shares = NICE_0_LOAD;
7533
7534 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7535
7536 for_each_possible_cpu(i) {
7537 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7538 GFP_KERNEL, cpu_to_node(i));
7539 if (!cfs_rq)
7540 goto err;
7541
7542 se = kzalloc_node(sizeof(struct sched_entity),
7543 GFP_KERNEL, cpu_to_node(i));
7544 if (!se)
7545 goto err_free_rq;
7546
7547 init_cfs_rq(cfs_rq);
7548 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7549 }
7550
7551 return 1;
7552
7553err_free_rq:
7554 kfree(cfs_rq);
7555err:
7556 return 0;
7557}
7558
7559void unregister_fair_sched_group(struct task_group *tg, int cpu)
7560{
7561 struct rq *rq = cpu_rq(cpu);
7562 unsigned long flags;
7563
7564 /*
7565 * Only empty task groups can be destroyed; so we can speculatively
7566 * check on_list without danger of it being re-added.
7567 */
7568 if (!tg->cfs_rq[cpu]->on_list)
7569 return;
7570
7571 raw_spin_lock_irqsave(&rq->lock, flags);
7572 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7573 raw_spin_unlock_irqrestore(&rq->lock, flags);
7574}
7575
7576void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7577 struct sched_entity *se, int cpu,
7578 struct sched_entity *parent)
7579{
7580 struct rq *rq = cpu_rq(cpu);
7581
7582 cfs_rq->tg = tg;
7583 cfs_rq->rq = rq;
029632fb
PZ
7584 init_cfs_rq_runtime(cfs_rq);
7585
7586 tg->cfs_rq[cpu] = cfs_rq;
7587 tg->se[cpu] = se;
7588
7589 /* se could be NULL for root_task_group */
7590 if (!se)
7591 return;
7592
fed14d45 7593 if (!parent) {
029632fb 7594 se->cfs_rq = &rq->cfs;
fed14d45
PZ
7595 se->depth = 0;
7596 } else {
029632fb 7597 se->cfs_rq = parent->my_q;
fed14d45
PZ
7598 se->depth = parent->depth + 1;
7599 }
029632fb
PZ
7600
7601 se->my_q = cfs_rq;
0ac9b1c2
PT
7602 /* guarantee group entities always have weight */
7603 update_load_set(&se->load, NICE_0_LOAD);
029632fb
PZ
7604 se->parent = parent;
7605}
7606
7607static DEFINE_MUTEX(shares_mutex);
7608
7609int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7610{
7611 int i;
7612 unsigned long flags;
7613
7614 /*
7615 * We can't change the weight of the root cgroup.
7616 */
7617 if (!tg->se[0])
7618 return -EINVAL;
7619
7620 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7621
7622 mutex_lock(&shares_mutex);
7623 if (tg->shares == shares)
7624 goto done;
7625
7626 tg->shares = shares;
7627 for_each_possible_cpu(i) {
7628 struct rq *rq = cpu_rq(i);
7629 struct sched_entity *se;
7630
7631 se = tg->se[i];
7632 /* Propagate contribution to hierarchy */
7633 raw_spin_lock_irqsave(&rq->lock, flags);
71b1da46
FW
7634
7635 /* Possible calls to update_curr() need rq clock */
7636 update_rq_clock(rq);
17bc14b7 7637 for_each_sched_entity(se)
029632fb
PZ
7638 update_cfs_shares(group_cfs_rq(se));
7639 raw_spin_unlock_irqrestore(&rq->lock, flags);
7640 }
7641
7642done:
7643 mutex_unlock(&shares_mutex);
7644 return 0;
7645}
7646#else /* CONFIG_FAIR_GROUP_SCHED */
7647
7648void free_fair_sched_group(struct task_group *tg) { }
7649
7650int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7651{
7652 return 1;
7653}
7654
7655void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7656
7657#endif /* CONFIG_FAIR_GROUP_SCHED */
7658
810b3817 7659
6d686f45 7660static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
0d721cea
PW
7661{
7662 struct sched_entity *se = &task->se;
0d721cea
PW
7663 unsigned int rr_interval = 0;
7664
7665 /*
7666 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7667 * idle runqueue:
7668 */
0d721cea 7669 if (rq->cfs.load.weight)
a59f4e07 7670 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
0d721cea
PW
7671
7672 return rr_interval;
7673}
7674
bf0f6f24
IM
7675/*
7676 * All the scheduling class methods:
7677 */
029632fb 7678const struct sched_class fair_sched_class = {
5522d5d5 7679 .next = &idle_sched_class,
bf0f6f24
IM
7680 .enqueue_task = enqueue_task_fair,
7681 .dequeue_task = dequeue_task_fair,
7682 .yield_task = yield_task_fair,
d95f4122 7683 .yield_to_task = yield_to_task_fair,
bf0f6f24 7684
2e09bf55 7685 .check_preempt_curr = check_preempt_wakeup,
bf0f6f24
IM
7686
7687 .pick_next_task = pick_next_task_fair,
7688 .put_prev_task = put_prev_task_fair,
7689
681f3e68 7690#ifdef CONFIG_SMP
4ce72a2c 7691 .select_task_rq = select_task_rq_fair,
0a74bef8 7692 .migrate_task_rq = migrate_task_rq_fair,
141965c7 7693
0bcdcf28
CE
7694 .rq_online = rq_online_fair,
7695 .rq_offline = rq_offline_fair,
88ec22d3
PZ
7696
7697 .task_waking = task_waking_fair,
681f3e68 7698#endif
bf0f6f24 7699
83b699ed 7700 .set_curr_task = set_curr_task_fair,
bf0f6f24 7701 .task_tick = task_tick_fair,
cd29fe6f 7702 .task_fork = task_fork_fair,
cb469845
SR
7703
7704 .prio_changed = prio_changed_fair,
da7a735e 7705 .switched_from = switched_from_fair,
cb469845 7706 .switched_to = switched_to_fair,
810b3817 7707
0d721cea
PW
7708 .get_rr_interval = get_rr_interval_fair,
7709
810b3817 7710#ifdef CONFIG_FAIR_GROUP_SCHED
b2b5ce02 7711 .task_move_group = task_move_group_fair,
810b3817 7712#endif
bf0f6f24
IM
7713};
7714
7715#ifdef CONFIG_SCHED_DEBUG
029632fb 7716void print_cfs_stats(struct seq_file *m, int cpu)
bf0f6f24 7717{
bf0f6f24
IM
7718 struct cfs_rq *cfs_rq;
7719
5973e5b9 7720 rcu_read_lock();
c3b64f1e 7721 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5cef9eca 7722 print_cfs_rq(m, cpu, cfs_rq);
5973e5b9 7723 rcu_read_unlock();
bf0f6f24
IM
7724}
7725#endif
029632fb
PZ
7726
7727__init void init_sched_fair_class(void)
7728{
7729#ifdef CONFIG_SMP
7730 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7731
3451d024 7732#ifdef CONFIG_NO_HZ_COMMON
554cecaf 7733 nohz.next_balance = jiffies;
029632fb 7734 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
71325960 7735 cpu_notifier(sched_ilb_notifier, 0);
029632fb
PZ
7736#endif
7737#endif /* SMP */
7738
7739}