sched: fix sched_slice()
[deliverable/linux.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
AV
23#include <linux/latencytop.h>
24
bf0f6f24 25/*
21805085 26 * Targeted preemption latency for CPU-bound tasks:
722aab0c 27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
bf0f6f24 28 *
21805085 29 * NOTE: this latency value is not the same as the concept of
d274a4ce
IM
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
bf0f6f24 33 *
d274a4ce
IM
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
bf0f6f24 36 */
19978ca6 37unsigned int sysctl_sched_latency = 20000000ULL;
2bd8e6d4
IM
38
39/*
b2be5e96 40 * Minimal preemption granularity for CPU-bound tasks:
722aab0c 41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
2bd8e6d4 42 */
722aab0c 43unsigned int sysctl_sched_min_granularity = 4000000ULL;
21805085
PZ
44
45/*
b2be5e96
PZ
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
47 */
722aab0c 48static unsigned int sched_nr_latency = 5;
b2be5e96
PZ
49
50/*
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
21805085 53 */
b2be5e96 54const_debug unsigned int sysctl_sched_child_runs_first = 1;
bf0f6f24 55
1799e35d
IM
56/*
57 * sys_sched_yield() compat mode
58 *
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
61 */
62unsigned int __read_mostly sysctl_sched_compat_yield;
63
bf0f6f24
IM
64/*
65 * SCHED_OTHER wake-up granularity.
103638d9 66 * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
bf0f6f24
IM
67 *
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
71 */
103638d9 72unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
bf0f6f24 73
da84d961
IM
74const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
75
a4c2f00f
PZ
76static const struct sched_class fair_sched_class;
77
bf0f6f24
IM
78/**************************************************************
79 * CFS operations on generic schedulable entities:
80 */
81
b758149c
PZ
82static inline struct task_struct *task_of(struct sched_entity *se)
83{
84 return container_of(se, struct task_struct, se);
85}
86
62160e3f 87#ifdef CONFIG_FAIR_GROUP_SCHED
bf0f6f24 88
62160e3f 89/* cpu runqueue to which this cfs_rq is attached */
bf0f6f24
IM
90static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
91{
62160e3f 92 return cfs_rq->rq;
bf0f6f24
IM
93}
94
62160e3f
IM
95/* An entity is a task if it doesn't "own" a runqueue */
96#define entity_is_task(se) (!se->my_q)
bf0f6f24 97
b758149c
PZ
98/* Walk up scheduling entities hierarchy */
99#define for_each_sched_entity(se) \
100 for (; se; se = se->parent)
101
102static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
103{
104 return p->se.cfs_rq;
105}
106
107/* runqueue on which this entity is (to be) queued */
108static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
109{
110 return se->cfs_rq;
111}
112
113/* runqueue "owned" by this group */
114static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
115{
116 return grp->my_q;
117}
118
119/* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
120 * another cpu ('this_cpu')
121 */
122static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
123{
124 return cfs_rq->tg->cfs_rq[this_cpu];
125}
126
127/* Iterate thr' all leaf cfs_rq's on a runqueue */
128#define for_each_leaf_cfs_rq(rq, cfs_rq) \
129 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
130
131/* Do the two (enqueued) entities belong to the same group ? */
132static inline int
133is_same_group(struct sched_entity *se, struct sched_entity *pse)
134{
135 if (se->cfs_rq == pse->cfs_rq)
136 return 1;
137
138 return 0;
139}
140
141static inline struct sched_entity *parent_entity(struct sched_entity *se)
142{
143 return se->parent;
144}
145
464b7527
PZ
146/* return depth at which a sched entity is present in the hierarchy */
147static inline int depth_se(struct sched_entity *se)
148{
149 int depth = 0;
150
151 for_each_sched_entity(se)
152 depth++;
153
154 return depth;
155}
156
157static void
158find_matching_se(struct sched_entity **se, struct sched_entity **pse)
159{
160 int se_depth, pse_depth;
161
162 /*
163 * preemption test can be made between sibling entities who are in the
164 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
165 * both tasks until we find their ancestors who are siblings of common
166 * parent.
167 */
168
169 /* First walk up until both entities are at same depth */
170 se_depth = depth_se(*se);
171 pse_depth = depth_se(*pse);
172
173 while (se_depth > pse_depth) {
174 se_depth--;
175 *se = parent_entity(*se);
176 }
177
178 while (pse_depth > se_depth) {
179 pse_depth--;
180 *pse = parent_entity(*pse);
181 }
182
183 while (!is_same_group(*se, *pse)) {
184 *se = parent_entity(*se);
185 *pse = parent_entity(*pse);
186 }
187}
188
62160e3f 189#else /* CONFIG_FAIR_GROUP_SCHED */
bf0f6f24 190
62160e3f
IM
191static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
192{
193 return container_of(cfs_rq, struct rq, cfs);
bf0f6f24
IM
194}
195
196#define entity_is_task(se) 1
197
b758149c
PZ
198#define for_each_sched_entity(se) \
199 for (; se; se = NULL)
bf0f6f24 200
b758149c 201static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
bf0f6f24 202{
b758149c 203 return &task_rq(p)->cfs;
bf0f6f24
IM
204}
205
b758149c
PZ
206static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
207{
208 struct task_struct *p = task_of(se);
209 struct rq *rq = task_rq(p);
210
211 return &rq->cfs;
212}
213
214/* runqueue "owned" by this group */
215static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
216{
217 return NULL;
218}
219
220static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
221{
222 return &cpu_rq(this_cpu)->cfs;
223}
224
225#define for_each_leaf_cfs_rq(rq, cfs_rq) \
226 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
227
228static inline int
229is_same_group(struct sched_entity *se, struct sched_entity *pse)
230{
231 return 1;
232}
233
234static inline struct sched_entity *parent_entity(struct sched_entity *se)
235{
236 return NULL;
237}
238
464b7527
PZ
239static inline void
240find_matching_se(struct sched_entity **se, struct sched_entity **pse)
241{
242}
243
b758149c
PZ
244#endif /* CONFIG_FAIR_GROUP_SCHED */
245
bf0f6f24
IM
246
247/**************************************************************
248 * Scheduling class tree data structure manipulation methods:
249 */
250
0702e3eb 251static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
02e0431a 252{
368059a9
PZ
253 s64 delta = (s64)(vruntime - min_vruntime);
254 if (delta > 0)
02e0431a
PZ
255 min_vruntime = vruntime;
256
257 return min_vruntime;
258}
259
0702e3eb 260static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
b0ffd246
PZ
261{
262 s64 delta = (s64)(vruntime - min_vruntime);
263 if (delta < 0)
264 min_vruntime = vruntime;
265
266 return min_vruntime;
267}
268
0702e3eb 269static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
9014623c 270{
30cfdcfc 271 return se->vruntime - cfs_rq->min_vruntime;
9014623c
PZ
272}
273
1af5f730
PZ
274static void update_min_vruntime(struct cfs_rq *cfs_rq)
275{
276 u64 vruntime = cfs_rq->min_vruntime;
277
278 if (cfs_rq->curr)
279 vruntime = cfs_rq->curr->vruntime;
280
281 if (cfs_rq->rb_leftmost) {
282 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
283 struct sched_entity,
284 run_node);
285
286 if (vruntime == cfs_rq->min_vruntime)
287 vruntime = se->vruntime;
288 else
289 vruntime = min_vruntime(vruntime, se->vruntime);
290 }
291
292 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
293}
294
bf0f6f24
IM
295/*
296 * Enqueue an entity into the rb-tree:
297 */
0702e3eb 298static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24
IM
299{
300 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
301 struct rb_node *parent = NULL;
302 struct sched_entity *entry;
9014623c 303 s64 key = entity_key(cfs_rq, se);
bf0f6f24
IM
304 int leftmost = 1;
305
306 /*
307 * Find the right place in the rbtree:
308 */
309 while (*link) {
310 parent = *link;
311 entry = rb_entry(parent, struct sched_entity, run_node);
312 /*
313 * We dont care about collisions. Nodes with
314 * the same key stay together.
315 */
9014623c 316 if (key < entity_key(cfs_rq, entry)) {
bf0f6f24
IM
317 link = &parent->rb_left;
318 } else {
319 link = &parent->rb_right;
320 leftmost = 0;
321 }
322 }
323
324 /*
325 * Maintain a cache of leftmost tree entries (it is frequently
326 * used):
327 */
1af5f730 328 if (leftmost)
57cb499d 329 cfs_rq->rb_leftmost = &se->run_node;
bf0f6f24
IM
330
331 rb_link_node(&se->run_node, parent, link);
332 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
bf0f6f24
IM
333}
334
0702e3eb 335static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 336{
3fe69747
PZ
337 if (cfs_rq->rb_leftmost == &se->run_node) {
338 struct rb_node *next_node;
3fe69747
PZ
339
340 next_node = rb_next(&se->run_node);
341 cfs_rq->rb_leftmost = next_node;
3fe69747 342 }
e9acbff6 343
bf0f6f24 344 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
bf0f6f24
IM
345}
346
bf0f6f24
IM
347static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
348{
f4b6755f
PZ
349 struct rb_node *left = cfs_rq->rb_leftmost;
350
351 if (!left)
352 return NULL;
353
354 return rb_entry(left, struct sched_entity, run_node);
bf0f6f24
IM
355}
356
f4b6755f 357static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
aeb73b04 358{
7eee3e67 359 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
aeb73b04 360
70eee74b
BS
361 if (!last)
362 return NULL;
7eee3e67
IM
363
364 return rb_entry(last, struct sched_entity, run_node);
aeb73b04
PZ
365}
366
bf0f6f24
IM
367/**************************************************************
368 * Scheduling class statistics methods:
369 */
370
b2be5e96
PZ
371#ifdef CONFIG_SCHED_DEBUG
372int sched_nr_latency_handler(struct ctl_table *table, int write,
373 struct file *filp, void __user *buffer, size_t *lenp,
374 loff_t *ppos)
375{
376 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
377
378 if (ret || !write)
379 return ret;
380
381 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
382 sysctl_sched_min_granularity);
383
384 return 0;
385}
386#endif
647e7cac 387
a7be37ac 388/*
f9c0b095 389 * delta /= w
a7be37ac
PZ
390 */
391static inline unsigned long
392calc_delta_fair(unsigned long delta, struct sched_entity *se)
393{
f9c0b095
PZ
394 if (unlikely(se->load.weight != NICE_0_LOAD))
395 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
a7be37ac
PZ
396
397 return delta;
398}
399
647e7cac
IM
400/*
401 * The idea is to set a period in which each task runs once.
402 *
403 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
404 * this period because otherwise the slices get too small.
405 *
406 * p = (nr <= nl) ? l : l*nr/nl
407 */
4d78e7b6
PZ
408static u64 __sched_period(unsigned long nr_running)
409{
410 u64 period = sysctl_sched_latency;
b2be5e96 411 unsigned long nr_latency = sched_nr_latency;
4d78e7b6
PZ
412
413 if (unlikely(nr_running > nr_latency)) {
4bf0b771 414 period = sysctl_sched_min_granularity;
4d78e7b6 415 period *= nr_running;
4d78e7b6
PZ
416 }
417
418 return period;
419}
420
647e7cac
IM
421/*
422 * We calculate the wall-time slice from the period by taking a part
423 * proportional to the weight.
424 *
f9c0b095 425 * s = p*P[w/rw]
647e7cac 426 */
6d0f0ebd 427static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
21805085 428{
0a582440 429 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
f9c0b095 430
0a582440
MG
431 for_each_sched_entity(se) {
432 struct load_weight *load = &cfs_rq->load;
f9c0b095 433
0a582440
MG
434 if (unlikely(!se->on_rq)) {
435 struct load_weight lw = cfs_rq->load;
436
437 update_load_add(&lw, se->load.weight);
438 load = &lw;
439 }
440 slice = calc_delta_mine(slice, se->load.weight, load);
441 }
442 return slice;
bf0f6f24
IM
443}
444
647e7cac 445/*
ac884dec 446 * We calculate the vruntime slice of a to be inserted task
647e7cac 447 *
f9c0b095 448 * vs = s/w
647e7cac 449 */
f9c0b095 450static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
67e9fb2a 451{
f9c0b095 452 return calc_delta_fair(sched_slice(cfs_rq, se), se);
a7be37ac
PZ
453}
454
bf0f6f24
IM
455/*
456 * Update the current task's runtime statistics. Skip current tasks that
457 * are not in our scheduling class.
458 */
459static inline void
8ebc91d9
IM
460__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
461 unsigned long delta_exec)
bf0f6f24 462{
bbdba7c0 463 unsigned long delta_exec_weighted;
bf0f6f24 464
8179ca23 465 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
bf0f6f24
IM
466
467 curr->sum_exec_runtime += delta_exec;
7a62eabc 468 schedstat_add(cfs_rq, exec_clock, delta_exec);
a7be37ac 469 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
e9acbff6 470 curr->vruntime += delta_exec_weighted;
1af5f730 471 update_min_vruntime(cfs_rq);
bf0f6f24
IM
472}
473
b7cc0896 474static void update_curr(struct cfs_rq *cfs_rq)
bf0f6f24 475{
429d43bc 476 struct sched_entity *curr = cfs_rq->curr;
8ebc91d9 477 u64 now = rq_of(cfs_rq)->clock;
bf0f6f24
IM
478 unsigned long delta_exec;
479
480 if (unlikely(!curr))
481 return;
482
483 /*
484 * Get the amount of time the current task was running
485 * since the last time we changed load (this cannot
486 * overflow on 32 bits):
487 */
8ebc91d9 488 delta_exec = (unsigned long)(now - curr->exec_start);
34f28ecd
PZ
489 if (!delta_exec)
490 return;
bf0f6f24 491
8ebc91d9
IM
492 __update_curr(cfs_rq, curr, delta_exec);
493 curr->exec_start = now;
d842de87
SV
494
495 if (entity_is_task(curr)) {
496 struct task_struct *curtask = task_of(curr);
497
498 cpuacct_charge(curtask, delta_exec);
f06febc9 499 account_group_exec_runtime(curtask, delta_exec);
d842de87 500 }
bf0f6f24
IM
501}
502
503static inline void
5870db5b 504update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 505{
d281918d 506 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
bf0f6f24
IM
507}
508
bf0f6f24
IM
509/*
510 * Task is being enqueued - update stats:
511 */
d2417e5a 512static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 513{
bf0f6f24
IM
514 /*
515 * Are we enqueueing a waiting task? (for current tasks
516 * a dequeue/enqueue event is a NOP)
517 */
429d43bc 518 if (se != cfs_rq->curr)
5870db5b 519 update_stats_wait_start(cfs_rq, se);
bf0f6f24
IM
520}
521
bf0f6f24 522static void
9ef0a961 523update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 524{
bbdba7c0
IM
525 schedstat_set(se->wait_max, max(se->wait_max,
526 rq_of(cfs_rq)->clock - se->wait_start));
6d082592
AV
527 schedstat_set(se->wait_count, se->wait_count + 1);
528 schedstat_set(se->wait_sum, se->wait_sum +
529 rq_of(cfs_rq)->clock - se->wait_start);
6cfb0d5d 530 schedstat_set(se->wait_start, 0);
bf0f6f24
IM
531}
532
533static inline void
19b6a2e3 534update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 535{
bf0f6f24
IM
536 /*
537 * Mark the end of the wait period if dequeueing a
538 * waiting task:
539 */
429d43bc 540 if (se != cfs_rq->curr)
9ef0a961 541 update_stats_wait_end(cfs_rq, se);
bf0f6f24
IM
542}
543
544/*
545 * We are picking a new current task - update its stats:
546 */
547static inline void
79303e9e 548update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24
IM
549{
550 /*
551 * We are starting a new run period:
552 */
d281918d 553 se->exec_start = rq_of(cfs_rq)->clock;
bf0f6f24
IM
554}
555
bf0f6f24
IM
556/**************************************************
557 * Scheduling class queueing methods:
558 */
559
c09595f6
PZ
560#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
561static void
562add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
563{
564 cfs_rq->task_weight += weight;
565}
566#else
567static inline void
568add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
569{
570}
571#endif
572
30cfdcfc
DA
573static void
574account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
575{
576 update_load_add(&cfs_rq->load, se->load.weight);
c09595f6
PZ
577 if (!parent_entity(se))
578 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
b87f1724 579 if (entity_is_task(se)) {
c09595f6 580 add_cfs_task_weight(cfs_rq, se->load.weight);
b87f1724
BR
581 list_add(&se->group_node, &cfs_rq->tasks);
582 }
30cfdcfc
DA
583 cfs_rq->nr_running++;
584 se->on_rq = 1;
585}
586
587static void
588account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
589{
590 update_load_sub(&cfs_rq->load, se->load.weight);
c09595f6
PZ
591 if (!parent_entity(se))
592 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
b87f1724 593 if (entity_is_task(se)) {
c09595f6 594 add_cfs_task_weight(cfs_rq, -se->load.weight);
b87f1724
BR
595 list_del_init(&se->group_node);
596 }
30cfdcfc
DA
597 cfs_rq->nr_running--;
598 se->on_rq = 0;
599}
600
2396af69 601static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 602{
bf0f6f24
IM
603#ifdef CONFIG_SCHEDSTATS
604 if (se->sleep_start) {
d281918d 605 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
9745512c 606 struct task_struct *tsk = task_of(se);
bf0f6f24
IM
607
608 if ((s64)delta < 0)
609 delta = 0;
610
611 if (unlikely(delta > se->sleep_max))
612 se->sleep_max = delta;
613
614 se->sleep_start = 0;
615 se->sum_sleep_runtime += delta;
9745512c
AV
616
617 account_scheduler_latency(tsk, delta >> 10, 1);
bf0f6f24
IM
618 }
619 if (se->block_start) {
d281918d 620 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
9745512c 621 struct task_struct *tsk = task_of(se);
bf0f6f24
IM
622
623 if ((s64)delta < 0)
624 delta = 0;
625
626 if (unlikely(delta > se->block_max))
627 se->block_max = delta;
628
629 se->block_start = 0;
630 se->sum_sleep_runtime += delta;
30084fbd
IM
631
632 /*
633 * Blocking time is in units of nanosecs, so shift by 20 to
634 * get a milliseconds-range estimation of the amount of
635 * time that the task spent sleeping:
636 */
637 if (unlikely(prof_on == SLEEP_PROFILING)) {
e22f5bbf 638
30084fbd
IM
639 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
640 delta >> 20);
641 }
9745512c 642 account_scheduler_latency(tsk, delta >> 10, 0);
bf0f6f24
IM
643 }
644#endif
645}
646
ddc97297
PZ
647static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
648{
649#ifdef CONFIG_SCHED_DEBUG
650 s64 d = se->vruntime - cfs_rq->min_vruntime;
651
652 if (d < 0)
653 d = -d;
654
655 if (d > 3*sysctl_sched_latency)
656 schedstat_inc(cfs_rq, nr_spread_over);
657#endif
658}
659
aeb73b04
PZ
660static void
661place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
662{
1af5f730 663 u64 vruntime = cfs_rq->min_vruntime;
94dfb5e7 664
2cb8600e
PZ
665 /*
666 * The 'current' period is already promised to the current tasks,
667 * however the extra weight of the new task will slow them down a
668 * little, place the new task so that it fits in the slot that
669 * stays open at the end.
670 */
94dfb5e7 671 if (initial && sched_feat(START_DEBIT))
f9c0b095 672 vruntime += sched_vslice(cfs_rq, se);
aeb73b04 673
8465e792 674 if (!initial) {
2cb8600e 675 /* sleeps upto a single latency don't count. */
a7be37ac
PZ
676 if (sched_feat(NEW_FAIR_SLEEPERS)) {
677 unsigned long thresh = sysctl_sched_latency;
678
679 /*
680 * convert the sleeper threshold into virtual time
681 */
682 if (sched_feat(NORMALIZED_SLEEPER))
683 thresh = calc_delta_fair(thresh, se);
684
685 vruntime -= thresh;
686 }
94359f05 687
2cb8600e
PZ
688 /* ensure we never gain time by being placed backwards. */
689 vruntime = max_vruntime(se->vruntime, vruntime);
aeb73b04
PZ
690 }
691
67e9fb2a 692 se->vruntime = vruntime;
aeb73b04
PZ
693}
694
bf0f6f24 695static void
83b699ed 696enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
bf0f6f24
IM
697{
698 /*
a2a2d680 699 * Update run-time statistics of the 'current'.
bf0f6f24 700 */
b7cc0896 701 update_curr(cfs_rq);
a992241d 702 account_entity_enqueue(cfs_rq, se);
bf0f6f24 703
e9acbff6 704 if (wakeup) {
aeb73b04 705 place_entity(cfs_rq, se, 0);
2396af69 706 enqueue_sleeper(cfs_rq, se);
e9acbff6 707 }
bf0f6f24 708
d2417e5a 709 update_stats_enqueue(cfs_rq, se);
ddc97297 710 check_spread(cfs_rq, se);
83b699ed
SV
711 if (se != cfs_rq->curr)
712 __enqueue_entity(cfs_rq, se);
bf0f6f24
IM
713}
714
2002c695
PZ
715static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
716{
717 if (cfs_rq->last == se)
718 cfs_rq->last = NULL;
719
720 if (cfs_rq->next == se)
721 cfs_rq->next = NULL;
722}
723
bf0f6f24 724static void
525c2716 725dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
bf0f6f24 726{
a2a2d680
DA
727 /*
728 * Update run-time statistics of the 'current'.
729 */
730 update_curr(cfs_rq);
731
19b6a2e3 732 update_stats_dequeue(cfs_rq, se);
db36cc7d 733 if (sleep) {
67e9fb2a 734#ifdef CONFIG_SCHEDSTATS
bf0f6f24
IM
735 if (entity_is_task(se)) {
736 struct task_struct *tsk = task_of(se);
737
738 if (tsk->state & TASK_INTERRUPTIBLE)
d281918d 739 se->sleep_start = rq_of(cfs_rq)->clock;
bf0f6f24 740 if (tsk->state & TASK_UNINTERRUPTIBLE)
d281918d 741 se->block_start = rq_of(cfs_rq)->clock;
bf0f6f24 742 }
db36cc7d 743#endif
67e9fb2a
PZ
744 }
745
2002c695 746 clear_buddies(cfs_rq, se);
4793241b 747
83b699ed 748 if (se != cfs_rq->curr)
30cfdcfc
DA
749 __dequeue_entity(cfs_rq, se);
750 account_entity_dequeue(cfs_rq, se);
1af5f730 751 update_min_vruntime(cfs_rq);
bf0f6f24
IM
752}
753
754/*
755 * Preempt the current task with a newly woken task if needed:
756 */
7c92e54f 757static void
2e09bf55 758check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
bf0f6f24 759{
11697830
PZ
760 unsigned long ideal_runtime, delta_exec;
761
6d0f0ebd 762 ideal_runtime = sched_slice(cfs_rq, curr);
11697830 763 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3e3e13f3 764 if (delta_exec > ideal_runtime)
bf0f6f24
IM
765 resched_task(rq_of(cfs_rq)->curr);
766}
767
83b699ed 768static void
8494f412 769set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
bf0f6f24 770{
83b699ed
SV
771 /* 'current' is not kept within the tree. */
772 if (se->on_rq) {
773 /*
774 * Any task has to be enqueued before it get to execute on
775 * a CPU. So account for the time it spent waiting on the
776 * runqueue.
777 */
778 update_stats_wait_end(cfs_rq, se);
779 __dequeue_entity(cfs_rq, se);
780 }
781
79303e9e 782 update_stats_curr_start(cfs_rq, se);
429d43bc 783 cfs_rq->curr = se;
eba1ed4b
IM
784#ifdef CONFIG_SCHEDSTATS
785 /*
786 * Track our maximum slice length, if the CPU's load is at
787 * least twice that of our own weight (i.e. dont track it
788 * when there are only lesser-weight tasks around):
789 */
495eca49 790 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
eba1ed4b
IM
791 se->slice_max = max(se->slice_max,
792 se->sum_exec_runtime - se->prev_sum_exec_runtime);
793 }
794#endif
4a55b450 795 se->prev_sum_exec_runtime = se->sum_exec_runtime;
bf0f6f24
IM
796}
797
3f3a4904
PZ
798static int
799wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
800
f4b6755f 801static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
aa2ac252 802{
f4b6755f
PZ
803 struct sched_entity *se = __pick_next_entity(cfs_rq);
804
4793241b
PZ
805 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
806 return cfs_rq->next;
aa2ac252 807
4793241b
PZ
808 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
809 return cfs_rq->last;
810
811 return se;
aa2ac252
PZ
812}
813
ab6cde26 814static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
bf0f6f24
IM
815{
816 /*
817 * If still on the runqueue then deactivate_task()
818 * was not called and update_curr() has to be done:
819 */
820 if (prev->on_rq)
b7cc0896 821 update_curr(cfs_rq);
bf0f6f24 822
ddc97297 823 check_spread(cfs_rq, prev);
30cfdcfc 824 if (prev->on_rq) {
5870db5b 825 update_stats_wait_start(cfs_rq, prev);
30cfdcfc
DA
826 /* Put 'current' back into the tree. */
827 __enqueue_entity(cfs_rq, prev);
828 }
429d43bc 829 cfs_rq->curr = NULL;
bf0f6f24
IM
830}
831
8f4d37ec
PZ
832static void
833entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
bf0f6f24 834{
bf0f6f24 835 /*
30cfdcfc 836 * Update run-time statistics of the 'current'.
bf0f6f24 837 */
30cfdcfc 838 update_curr(cfs_rq);
bf0f6f24 839
8f4d37ec
PZ
840#ifdef CONFIG_SCHED_HRTICK
841 /*
842 * queued ticks are scheduled to match the slice, so don't bother
843 * validating it and just reschedule.
844 */
983ed7a6
HH
845 if (queued) {
846 resched_task(rq_of(cfs_rq)->curr);
847 return;
848 }
8f4d37ec
PZ
849 /*
850 * don't let the period tick interfere with the hrtick preemption
851 */
852 if (!sched_feat(DOUBLE_TICK) &&
853 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
854 return;
855#endif
856
ce6c1311 857 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
2e09bf55 858 check_preempt_tick(cfs_rq, curr);
bf0f6f24
IM
859}
860
861/**************************************************
862 * CFS operations on tasks:
863 */
864
8f4d37ec
PZ
865#ifdef CONFIG_SCHED_HRTICK
866static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
867{
8f4d37ec
PZ
868 struct sched_entity *se = &p->se;
869 struct cfs_rq *cfs_rq = cfs_rq_of(se);
870
871 WARN_ON(task_rq(p) != rq);
872
873 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
874 u64 slice = sched_slice(cfs_rq, se);
875 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
876 s64 delta = slice - ran;
877
878 if (delta < 0) {
879 if (rq->curr == p)
880 resched_task(p);
881 return;
882 }
883
884 /*
885 * Don't schedule slices shorter than 10000ns, that just
886 * doesn't make sense. Rely on vruntime for fairness.
887 */
31656519 888 if (rq->curr != p)
157124c1 889 delta = max_t(s64, 10000LL, delta);
8f4d37ec 890
31656519 891 hrtick_start(rq, delta);
8f4d37ec
PZ
892 }
893}
a4c2f00f
PZ
894
895/*
896 * called from enqueue/dequeue and updates the hrtick when the
897 * current task is from our class and nr_running is low enough
898 * to matter.
899 */
900static void hrtick_update(struct rq *rq)
901{
902 struct task_struct *curr = rq->curr;
903
904 if (curr->sched_class != &fair_sched_class)
905 return;
906
907 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
908 hrtick_start_fair(rq, curr);
909}
55e12e5e 910#else /* !CONFIG_SCHED_HRTICK */
8f4d37ec
PZ
911static inline void
912hrtick_start_fair(struct rq *rq, struct task_struct *p)
913{
914}
a4c2f00f
PZ
915
916static inline void hrtick_update(struct rq *rq)
917{
918}
8f4d37ec
PZ
919#endif
920
bf0f6f24
IM
921/*
922 * The enqueue_task method is called before nr_running is
923 * increased. Here we update the fair scheduling stats and
924 * then put the task into the rbtree:
925 */
fd390f6a 926static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
bf0f6f24
IM
927{
928 struct cfs_rq *cfs_rq;
62fb1851 929 struct sched_entity *se = &p->se;
bf0f6f24
IM
930
931 for_each_sched_entity(se) {
62fb1851 932 if (se->on_rq)
bf0f6f24
IM
933 break;
934 cfs_rq = cfs_rq_of(se);
83b699ed 935 enqueue_entity(cfs_rq, se, wakeup);
b9fa3df3 936 wakeup = 1;
bf0f6f24 937 }
8f4d37ec 938
a4c2f00f 939 hrtick_update(rq);
bf0f6f24
IM
940}
941
942/*
943 * The dequeue_task method is called before nr_running is
944 * decreased. We remove the task from the rbtree and
945 * update the fair scheduling stats:
946 */
f02231e5 947static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
bf0f6f24
IM
948{
949 struct cfs_rq *cfs_rq;
62fb1851 950 struct sched_entity *se = &p->se;
bf0f6f24
IM
951
952 for_each_sched_entity(se) {
953 cfs_rq = cfs_rq_of(se);
525c2716 954 dequeue_entity(cfs_rq, se, sleep);
bf0f6f24 955 /* Don't dequeue parent if it has other entities besides us */
62fb1851 956 if (cfs_rq->load.weight)
bf0f6f24 957 break;
b9fa3df3 958 sleep = 1;
bf0f6f24 959 }
8f4d37ec 960
a4c2f00f 961 hrtick_update(rq);
bf0f6f24
IM
962}
963
964/*
1799e35d
IM
965 * sched_yield() support is very simple - we dequeue and enqueue.
966 *
967 * If compat_yield is turned on then we requeue to the end of the tree.
bf0f6f24 968 */
4530d7ab 969static void yield_task_fair(struct rq *rq)
bf0f6f24 970{
db292ca3
IM
971 struct task_struct *curr = rq->curr;
972 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
973 struct sched_entity *rightmost, *se = &curr->se;
bf0f6f24
IM
974
975 /*
1799e35d
IM
976 * Are we the only task in the tree?
977 */
978 if (unlikely(cfs_rq->nr_running == 1))
979 return;
980
2002c695
PZ
981 clear_buddies(cfs_rq, se);
982
db292ca3 983 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
3e51f33f 984 update_rq_clock(rq);
1799e35d 985 /*
a2a2d680 986 * Update run-time statistics of the 'current'.
1799e35d 987 */
2b1e315d 988 update_curr(cfs_rq);
1799e35d
IM
989
990 return;
991 }
992 /*
993 * Find the rightmost entry in the rbtree:
bf0f6f24 994 */
2b1e315d 995 rightmost = __pick_last_entity(cfs_rq);
1799e35d
IM
996 /*
997 * Already in the rightmost position?
998 */
79b3feff 999 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
1799e35d
IM
1000 return;
1001
1002 /*
1003 * Minimally necessary key value to be last in the tree:
2b1e315d
DA
1004 * Upon rescheduling, sched_class::put_prev_task() will place
1005 * 'current' within the tree based on its new key value.
1799e35d 1006 */
30cfdcfc 1007 se->vruntime = rightmost->vruntime + 1;
bf0f6f24
IM
1008}
1009
e7693a36
GH
1010/*
1011 * wake_idle() will wake a task on an idle cpu if task->cpu is
1012 * not idle and an idle cpu is available. The span of cpus to
1013 * search starts with cpus closest then further out as needed,
1014 * so we always favor a closer, idle cpu.
e761b772
MK
1015 * Domains may include CPUs that are not usable for migration,
1016 * hence we need to mask them out (cpu_active_map)
e7693a36
GH
1017 *
1018 * Returns the CPU we should wake onto.
1019 */
1020#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1021static int wake_idle(int cpu, struct task_struct *p)
1022{
1023 cpumask_t tmp;
1024 struct sched_domain *sd;
1025 int i;
1026
1027 /*
1028 * If it is idle, then it is the best cpu to run this task.
1029 *
1030 * This cpu is also the best, if it has more than one task already.
1031 * Siblings must be also busy(in most cases) as they didn't already
1032 * pickup the extra load from this cpu and hence we need not check
1033 * sibling runqueue info. This will avoid the checks and cache miss
1034 * penalities associated with that.
1035 */
104f6454 1036 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
e7693a36
GH
1037 return cpu;
1038
1039 for_each_domain(cpu, sd) {
1d3504fc
HS
1040 if ((sd->flags & SD_WAKE_IDLE)
1041 || ((sd->flags & SD_WAKE_IDLE_FAR)
1042 && !task_hot(p, task_rq(p)->clock, sd))) {
e7693a36 1043 cpus_and(tmp, sd->span, p->cpus_allowed);
e761b772 1044 cpus_and(tmp, tmp, cpu_active_map);
363ab6f1 1045 for_each_cpu_mask_nr(i, tmp) {
e7693a36
GH
1046 if (idle_cpu(i)) {
1047 if (i != task_cpu(p)) {
1048 schedstat_inc(p,
1049 se.nr_wakeups_idle);
1050 }
1051 return i;
1052 }
1053 }
1054 } else {
1055 break;
1056 }
1057 }
1058 return cpu;
1059}
55e12e5e 1060#else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
e7693a36
GH
1061static inline int wake_idle(int cpu, struct task_struct *p)
1062{
1063 return cpu;
1064}
1065#endif
1066
1067#ifdef CONFIG_SMP
098fb9db 1068
bb3469ac 1069#ifdef CONFIG_FAIR_GROUP_SCHED
f5bfb7d9
PZ
1070/*
1071 * effective_load() calculates the load change as seen from the root_task_group
1072 *
1073 * Adding load to a group doesn't make a group heavier, but can cause movement
1074 * of group shares between cpus. Assuming the shares were perfectly aligned one
1075 * can calculate the shift in shares.
1076 *
1077 * The problem is that perfectly aligning the shares is rather expensive, hence
1078 * we try to avoid doing that too often - see update_shares(), which ratelimits
1079 * this change.
1080 *
1081 * We compensate this by not only taking the current delta into account, but
1082 * also considering the delta between when the shares were last adjusted and
1083 * now.
1084 *
1085 * We still saw a performance dip, some tracing learned us that between
1086 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1087 * significantly. Therefore try to bias the error in direction of failing
1088 * the affine wakeup.
1089 *
1090 */
f1d239f7
PZ
1091static long effective_load(struct task_group *tg, int cpu,
1092 long wl, long wg)
bb3469ac 1093{
4be9daaa 1094 struct sched_entity *se = tg->se[cpu];
f1d239f7
PZ
1095
1096 if (!tg->parent)
1097 return wl;
1098
f5bfb7d9
PZ
1099 /*
1100 * By not taking the decrease of shares on the other cpu into
1101 * account our error leans towards reducing the affine wakeups.
1102 */
1103 if (!wl && sched_feat(ASYM_EFF_LOAD))
1104 return wl;
1105
4be9daaa 1106 for_each_sched_entity(se) {
cb5ef42a 1107 long S, rw, s, a, b;
940959e9
PZ
1108 long more_w;
1109
1110 /*
1111 * Instead of using this increment, also add the difference
1112 * between when the shares were last updated and now.
1113 */
1114 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1115 wl += more_w;
1116 wg += more_w;
4be9daaa
PZ
1117
1118 S = se->my_q->tg->shares;
1119 s = se->my_q->shares;
f1d239f7 1120 rw = se->my_q->rq_weight;
bb3469ac 1121
cb5ef42a
PZ
1122 a = S*(rw + wl);
1123 b = S*rw + s*wg;
4be9daaa 1124
940959e9
PZ
1125 wl = s*(a-b);
1126
1127 if (likely(b))
1128 wl /= b;
1129
83378269
PZ
1130 /*
1131 * Assume the group is already running and will
1132 * thus already be accounted for in the weight.
1133 *
1134 * That is, moving shares between CPUs, does not
1135 * alter the group weight.
1136 */
4be9daaa 1137 wg = 0;
4be9daaa 1138 }
bb3469ac 1139
4be9daaa 1140 return wl;
bb3469ac 1141}
4be9daaa 1142
bb3469ac 1143#else
4be9daaa 1144
83378269
PZ
1145static inline unsigned long effective_load(struct task_group *tg, int cpu,
1146 unsigned long wl, unsigned long wg)
4be9daaa 1147{
83378269 1148 return wl;
bb3469ac 1149}
4be9daaa 1150
bb3469ac
PZ
1151#endif
1152
098fb9db 1153static int
64b9e029 1154wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
4ae7d5ce
IM
1155 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1156 int idx, unsigned long load, unsigned long this_load,
098fb9db
IM
1157 unsigned int imbalance)
1158{
4ae7d5ce 1159 struct task_struct *curr = this_rq->curr;
83378269 1160 struct task_group *tg;
098fb9db
IM
1161 unsigned long tl = this_load;
1162 unsigned long tl_per_task;
83378269 1163 unsigned long weight;
b3137bc8 1164 int balanced;
098fb9db 1165
b3137bc8 1166 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
098fb9db
IM
1167 return 0;
1168
0d13033b
MG
1169 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1170 p->se.avg_overlap > sysctl_sched_migration_cost))
1171 sync = 0;
2fb7635c 1172
b3137bc8
MG
1173 /*
1174 * If sync wakeup then subtract the (maximum possible)
1175 * effect of the currently running task from the load
1176 * of the current CPU:
1177 */
83378269
PZ
1178 if (sync) {
1179 tg = task_group(current);
1180 weight = current->se.load.weight;
1181
1182 tl += effective_load(tg, this_cpu, -weight, -weight);
1183 load += effective_load(tg, prev_cpu, 0, -weight);
1184 }
b3137bc8 1185
83378269
PZ
1186 tg = task_group(p);
1187 weight = p->se.load.weight;
b3137bc8 1188
83378269
PZ
1189 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1190 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
b3137bc8 1191
098fb9db 1192 /*
4ae7d5ce
IM
1193 * If the currently running task will sleep within
1194 * a reasonable amount of time then attract this newly
1195 * woken task:
098fb9db 1196 */
2fb7635c
PZ
1197 if (sync && balanced)
1198 return 1;
098fb9db
IM
1199
1200 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1201 tl_per_task = cpu_avg_load_per_task(this_cpu);
1202
64b9e029
AA
1203 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1204 tl_per_task)) {
098fb9db
IM
1205 /*
1206 * This domain has SD_WAKE_AFFINE and
1207 * p is cache cold in this domain, and
1208 * there is no bad imbalance.
1209 */
1210 schedstat_inc(this_sd, ttwu_move_affine);
1211 schedstat_inc(p, se.nr_wakeups_affine);
1212
1213 return 1;
1214 }
1215 return 0;
1216}
1217
e7693a36
GH
1218static int select_task_rq_fair(struct task_struct *p, int sync)
1219{
e7693a36 1220 struct sched_domain *sd, *this_sd = NULL;
ac192d39 1221 int prev_cpu, this_cpu, new_cpu;
098fb9db 1222 unsigned long load, this_load;
64b9e029 1223 struct rq *this_rq;
098fb9db 1224 unsigned int imbalance;
098fb9db 1225 int idx;
e7693a36 1226
ac192d39 1227 prev_cpu = task_cpu(p);
ac192d39 1228 this_cpu = smp_processor_id();
4ae7d5ce 1229 this_rq = cpu_rq(this_cpu);
ac192d39 1230 new_cpu = prev_cpu;
e7693a36 1231
64b9e029
AA
1232 if (prev_cpu == this_cpu)
1233 goto out;
ac192d39
IM
1234 /*
1235 * 'this_sd' is the first domain that both
1236 * this_cpu and prev_cpu are present in:
1237 */
e7693a36 1238 for_each_domain(this_cpu, sd) {
ac192d39 1239 if (cpu_isset(prev_cpu, sd->span)) {
e7693a36
GH
1240 this_sd = sd;
1241 break;
1242 }
1243 }
1244
1245 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
f4827386 1246 goto out;
e7693a36
GH
1247
1248 /*
1249 * Check for affine wakeup and passive balancing possibilities.
1250 */
098fb9db 1251 if (!this_sd)
f4827386 1252 goto out;
e7693a36 1253
098fb9db
IM
1254 idx = this_sd->wake_idx;
1255
1256 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1257
ac192d39 1258 load = source_load(prev_cpu, idx);
098fb9db
IM
1259 this_load = target_load(this_cpu, idx);
1260
64b9e029 1261 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
4ae7d5ce
IM
1262 load, this_load, imbalance))
1263 return this_cpu;
1264
098fb9db
IM
1265 /*
1266 * Start passive balancing when half the imbalance_pct
1267 * limit is reached.
1268 */
1269 if (this_sd->flags & SD_WAKE_BALANCE) {
1270 if (imbalance*this_load <= 100*load) {
1271 schedstat_inc(this_sd, ttwu_move_balance);
1272 schedstat_inc(p, se.nr_wakeups_passive);
4ae7d5ce 1273 return this_cpu;
e7693a36
GH
1274 }
1275 }
1276
f4827386 1277out:
e7693a36
GH
1278 return wake_idle(new_cpu, p);
1279}
1280#endif /* CONFIG_SMP */
1281
0bbd3336
PZ
1282static unsigned long wakeup_gran(struct sched_entity *se)
1283{
1284 unsigned long gran = sysctl_sched_wakeup_granularity;
1285
1286 /*
a7be37ac
PZ
1287 * More easily preempt - nice tasks, while not making it harder for
1288 * + nice tasks.
0bbd3336 1289 */
464b7527
PZ
1290 if (!sched_feat(ASYM_GRAN) || se->load.weight > NICE_0_LOAD)
1291 gran = calc_delta_fair(sysctl_sched_wakeup_granularity, se);
0bbd3336
PZ
1292
1293 return gran;
1294}
1295
464b7527
PZ
1296/*
1297 * Should 'se' preempt 'curr'.
1298 *
1299 * |s1
1300 * |s2
1301 * |s3
1302 * g
1303 * |<--->|c
1304 *
1305 * w(c, s1) = -1
1306 * w(c, s2) = 0
1307 * w(c, s3) = 1
1308 *
1309 */
1310static int
1311wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1312{
1313 s64 gran, vdiff = curr->vruntime - se->vruntime;
1314
1315 if (vdiff <= 0)
1316 return -1;
1317
1318 gran = wakeup_gran(curr);
1319 if (vdiff > gran)
1320 return 1;
1321
1322 return 0;
1323}
1324
02479099
PZ
1325static void set_last_buddy(struct sched_entity *se)
1326{
1327 for_each_sched_entity(se)
1328 cfs_rq_of(se)->last = se;
1329}
1330
1331static void set_next_buddy(struct sched_entity *se)
1332{
1333 for_each_sched_entity(se)
1334 cfs_rq_of(se)->next = se;
1335}
1336
bf0f6f24
IM
1337/*
1338 * Preempt the current task with a newly woken task if needed:
1339 */
15afe09b 1340static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
bf0f6f24
IM
1341{
1342 struct task_struct *curr = rq->curr;
8651a86c 1343 struct sched_entity *se = &curr->se, *pse = &p->se;
03e89e45 1344 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
bf0f6f24 1345
03e89e45 1346 update_curr(cfs_rq);
4793241b 1347
03e89e45 1348 if (unlikely(rt_prio(p->prio))) {
bf0f6f24
IM
1349 resched_task(curr);
1350 return;
1351 }
aa2ac252 1352
d95f98d0
PZ
1353 if (unlikely(p->sched_class != &fair_sched_class))
1354 return;
1355
4ae7d5ce
IM
1356 if (unlikely(se == pse))
1357 return;
1358
4793241b
PZ
1359 /*
1360 * Only set the backward buddy when the current task is still on the
1361 * rq. This can happen when a wakeup gets interleaved with schedule on
1362 * the ->pre_schedule() or idle_balance() point, either of which can
1363 * drop the rq lock.
1364 *
1365 * Also, during early boot the idle thread is in the fair class, for
1366 * obvious reasons its a bad idea to schedule back to the idle thread.
1367 */
1368 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
02479099
PZ
1369 set_last_buddy(se);
1370 set_next_buddy(pse);
57fdc26d 1371
aec0a514
BR
1372 /*
1373 * We can come here with TIF_NEED_RESCHED already set from new task
1374 * wake up path.
1375 */
1376 if (test_tsk_need_resched(curr))
1377 return;
1378
91c234b4
IM
1379 /*
1380 * Batch tasks do not preempt (their preemption is driven by
1381 * the tick):
1382 */
1383 if (unlikely(p->policy == SCHED_BATCH))
1384 return;
bf0f6f24 1385
77d9cc44
IM
1386 if (!sched_feat(WAKEUP_PREEMPT))
1387 return;
8651a86c 1388
2fb7635c
PZ
1389 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1390 (se->avg_overlap < sysctl_sched_migration_cost &&
1391 pse->avg_overlap < sysctl_sched_migration_cost))) {
15afe09b
PZ
1392 resched_task(curr);
1393 return;
1394 }
1395
464b7527
PZ
1396 find_matching_se(&se, &pse);
1397
1398 while (se) {
1399 BUG_ON(!pse);
1400
1401 if (wakeup_preempt_entity(se, pse) == 1) {
1402 resched_task(curr);
1403 break;
1404 }
1405
1406 se = parent_entity(se);
1407 pse = parent_entity(pse);
1408 }
bf0f6f24
IM
1409}
1410
fb8d4724 1411static struct task_struct *pick_next_task_fair(struct rq *rq)
bf0f6f24 1412{
8f4d37ec 1413 struct task_struct *p;
bf0f6f24
IM
1414 struct cfs_rq *cfs_rq = &rq->cfs;
1415 struct sched_entity *se;
1416
1417 if (unlikely(!cfs_rq->nr_running))
1418 return NULL;
1419
1420 do {
9948f4b2 1421 se = pick_next_entity(cfs_rq);
f4b6755f 1422 set_next_entity(cfs_rq, se);
bf0f6f24
IM
1423 cfs_rq = group_cfs_rq(se);
1424 } while (cfs_rq);
1425
8f4d37ec
PZ
1426 p = task_of(se);
1427 hrtick_start_fair(rq, p);
1428
1429 return p;
bf0f6f24
IM
1430}
1431
1432/*
1433 * Account for a descheduled task:
1434 */
31ee529c 1435static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
bf0f6f24
IM
1436{
1437 struct sched_entity *se = &prev->se;
1438 struct cfs_rq *cfs_rq;
1439
1440 for_each_sched_entity(se) {
1441 cfs_rq = cfs_rq_of(se);
ab6cde26 1442 put_prev_entity(cfs_rq, se);
bf0f6f24
IM
1443 }
1444}
1445
681f3e68 1446#ifdef CONFIG_SMP
bf0f6f24
IM
1447/**************************************************
1448 * Fair scheduling class load-balancing methods:
1449 */
1450
1451/*
1452 * Load-balancing iterator. Note: while the runqueue stays locked
1453 * during the whole iteration, the current task might be
1454 * dequeued so the iterator has to be dequeue-safe. Here we
1455 * achieve that by always pre-iterating before returning
1456 * the current task:
1457 */
a9957449 1458static struct task_struct *
4a55bd5e 1459__load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
bf0f6f24 1460{
354d60c2
DG
1461 struct task_struct *p = NULL;
1462 struct sched_entity *se;
bf0f6f24 1463
77ae6513
MG
1464 if (next == &cfs_rq->tasks)
1465 return NULL;
1466
b87f1724
BR
1467 se = list_entry(next, struct sched_entity, group_node);
1468 p = task_of(se);
1469 cfs_rq->balance_iterator = next->next;
77ae6513 1470
bf0f6f24
IM
1471 return p;
1472}
1473
1474static struct task_struct *load_balance_start_fair(void *arg)
1475{
1476 struct cfs_rq *cfs_rq = arg;
1477
4a55bd5e 1478 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
bf0f6f24
IM
1479}
1480
1481static struct task_struct *load_balance_next_fair(void *arg)
1482{
1483 struct cfs_rq *cfs_rq = arg;
1484
4a55bd5e 1485 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
bf0f6f24
IM
1486}
1487
c09595f6
PZ
1488static unsigned long
1489__load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1490 unsigned long max_load_move, struct sched_domain *sd,
1491 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1492 struct cfs_rq *cfs_rq)
62fb1851 1493{
c09595f6 1494 struct rq_iterator cfs_rq_iterator;
62fb1851 1495
c09595f6
PZ
1496 cfs_rq_iterator.start = load_balance_start_fair;
1497 cfs_rq_iterator.next = load_balance_next_fair;
1498 cfs_rq_iterator.arg = cfs_rq;
62fb1851 1499
c09595f6
PZ
1500 return balance_tasks(this_rq, this_cpu, busiest,
1501 max_load_move, sd, idle, all_pinned,
1502 this_best_prio, &cfs_rq_iterator);
62fb1851 1503}
62fb1851 1504
c09595f6 1505#ifdef CONFIG_FAIR_GROUP_SCHED
43010659 1506static unsigned long
bf0f6f24 1507load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
e1d1484f 1508 unsigned long max_load_move,
a4ac01c3
PW
1509 struct sched_domain *sd, enum cpu_idle_type idle,
1510 int *all_pinned, int *this_best_prio)
bf0f6f24 1511{
bf0f6f24 1512 long rem_load_move = max_load_move;
c09595f6
PZ
1513 int busiest_cpu = cpu_of(busiest);
1514 struct task_group *tg;
18d95a28 1515
c09595f6 1516 rcu_read_lock();
c8cba857 1517 update_h_load(busiest_cpu);
18d95a28 1518
caea8a03 1519 list_for_each_entry_rcu(tg, &task_groups, list) {
c8cba857 1520 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
42a3ac7d
PZ
1521 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1522 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
243e0e7b 1523 u64 rem_load, moved_load;
18d95a28 1524
c09595f6
PZ
1525 /*
1526 * empty group
1527 */
c8cba857 1528 if (!busiest_cfs_rq->task_weight)
bf0f6f24
IM
1529 continue;
1530
243e0e7b
SV
1531 rem_load = (u64)rem_load_move * busiest_weight;
1532 rem_load = div_u64(rem_load, busiest_h_load + 1);
bf0f6f24 1533
c09595f6 1534 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
53fecd8a 1535 rem_load, sd, idle, all_pinned, this_best_prio,
c09595f6 1536 tg->cfs_rq[busiest_cpu]);
bf0f6f24 1537
c09595f6 1538 if (!moved_load)
bf0f6f24
IM
1539 continue;
1540
42a3ac7d 1541 moved_load *= busiest_h_load;
243e0e7b 1542 moved_load = div_u64(moved_load, busiest_weight + 1);
bf0f6f24 1543
c09595f6
PZ
1544 rem_load_move -= moved_load;
1545 if (rem_load_move < 0)
bf0f6f24
IM
1546 break;
1547 }
c09595f6 1548 rcu_read_unlock();
bf0f6f24 1549
43010659 1550 return max_load_move - rem_load_move;
bf0f6f24 1551}
c09595f6
PZ
1552#else
1553static unsigned long
1554load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1555 unsigned long max_load_move,
1556 struct sched_domain *sd, enum cpu_idle_type idle,
1557 int *all_pinned, int *this_best_prio)
1558{
1559 return __load_balance_fair(this_rq, this_cpu, busiest,
1560 max_load_move, sd, idle, all_pinned,
1561 this_best_prio, &busiest->cfs);
1562}
1563#endif
bf0f6f24 1564
e1d1484f
PW
1565static int
1566move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1567 struct sched_domain *sd, enum cpu_idle_type idle)
1568{
1569 struct cfs_rq *busy_cfs_rq;
1570 struct rq_iterator cfs_rq_iterator;
1571
1572 cfs_rq_iterator.start = load_balance_start_fair;
1573 cfs_rq_iterator.next = load_balance_next_fair;
1574
1575 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1576 /*
1577 * pass busy_cfs_rq argument into
1578 * load_balance_[start|next]_fair iterators
1579 */
1580 cfs_rq_iterator.arg = busy_cfs_rq;
1581 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1582 &cfs_rq_iterator))
1583 return 1;
1584 }
1585
1586 return 0;
1587}
55e12e5e 1588#endif /* CONFIG_SMP */
e1d1484f 1589
bf0f6f24
IM
1590/*
1591 * scheduler tick hitting a task of our scheduling class:
1592 */
8f4d37ec 1593static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
bf0f6f24
IM
1594{
1595 struct cfs_rq *cfs_rq;
1596 struct sched_entity *se = &curr->se;
1597
1598 for_each_sched_entity(se) {
1599 cfs_rq = cfs_rq_of(se);
8f4d37ec 1600 entity_tick(cfs_rq, se, queued);
bf0f6f24
IM
1601 }
1602}
1603
8eb172d9 1604#define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
4d78e7b6 1605
bf0f6f24
IM
1606/*
1607 * Share the fairness runtime between parent and child, thus the
1608 * total amount of pressure for CPU stays equal - new tasks
1609 * get a chance to run but frequent forkers are not allowed to
1610 * monopolize the CPU. Note: the parent runqueue is locked,
1611 * the child is not running yet.
1612 */
ee0827d8 1613static void task_new_fair(struct rq *rq, struct task_struct *p)
bf0f6f24
IM
1614{
1615 struct cfs_rq *cfs_rq = task_cfs_rq(p);
429d43bc 1616 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
00bf7bfc 1617 int this_cpu = smp_processor_id();
bf0f6f24
IM
1618
1619 sched_info_queued(p);
1620
7109c442 1621 update_curr(cfs_rq);
aeb73b04 1622 place_entity(cfs_rq, se, 1);
4d78e7b6 1623
3c90e6e9 1624 /* 'curr' will be NULL if the child belongs to a different group */
00bf7bfc 1625 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
3c90e6e9 1626 curr && curr->vruntime < se->vruntime) {
87fefa38 1627 /*
edcb60a3
IM
1628 * Upon rescheduling, sched_class::put_prev_task() will place
1629 * 'current' within the tree based on its new key value.
1630 */
4d78e7b6 1631 swap(curr->vruntime, se->vruntime);
aec0a514 1632 resched_task(rq->curr);
4d78e7b6 1633 }
bf0f6f24 1634
b9dca1e0 1635 enqueue_task_fair(rq, p, 0);
bf0f6f24
IM
1636}
1637
cb469845
SR
1638/*
1639 * Priority of the task has changed. Check to see if we preempt
1640 * the current task.
1641 */
1642static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1643 int oldprio, int running)
1644{
1645 /*
1646 * Reschedule if we are currently running on this runqueue and
1647 * our priority decreased, or if we are not currently running on
1648 * this runqueue and our priority is higher than the current's
1649 */
1650 if (running) {
1651 if (p->prio > oldprio)
1652 resched_task(rq->curr);
1653 } else
15afe09b 1654 check_preempt_curr(rq, p, 0);
cb469845
SR
1655}
1656
1657/*
1658 * We switched to the sched_fair class.
1659 */
1660static void switched_to_fair(struct rq *rq, struct task_struct *p,
1661 int running)
1662{
1663 /*
1664 * We were most likely switched from sched_rt, so
1665 * kick off the schedule if running, otherwise just see
1666 * if we can still preempt the current task.
1667 */
1668 if (running)
1669 resched_task(rq->curr);
1670 else
15afe09b 1671 check_preempt_curr(rq, p, 0);
cb469845
SR
1672}
1673
83b699ed
SV
1674/* Account for a task changing its policy or group.
1675 *
1676 * This routine is mostly called to set cfs_rq->curr field when a task
1677 * migrates between groups/classes.
1678 */
1679static void set_curr_task_fair(struct rq *rq)
1680{
1681 struct sched_entity *se = &rq->curr->se;
1682
1683 for_each_sched_entity(se)
1684 set_next_entity(cfs_rq_of(se), se);
1685}
1686
810b3817
PZ
1687#ifdef CONFIG_FAIR_GROUP_SCHED
1688static void moved_group_fair(struct task_struct *p)
1689{
1690 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1691
1692 update_curr(cfs_rq);
1693 place_entity(cfs_rq, &p->se, 1);
1694}
1695#endif
1696
bf0f6f24
IM
1697/*
1698 * All the scheduling class methods:
1699 */
5522d5d5
IM
1700static const struct sched_class fair_sched_class = {
1701 .next = &idle_sched_class,
bf0f6f24
IM
1702 .enqueue_task = enqueue_task_fair,
1703 .dequeue_task = dequeue_task_fair,
1704 .yield_task = yield_task_fair,
1705
2e09bf55 1706 .check_preempt_curr = check_preempt_wakeup,
bf0f6f24
IM
1707
1708 .pick_next_task = pick_next_task_fair,
1709 .put_prev_task = put_prev_task_fair,
1710
681f3e68 1711#ifdef CONFIG_SMP
4ce72a2c
LZ
1712 .select_task_rq = select_task_rq_fair,
1713
bf0f6f24 1714 .load_balance = load_balance_fair,
e1d1484f 1715 .move_one_task = move_one_task_fair,
681f3e68 1716#endif
bf0f6f24 1717
83b699ed 1718 .set_curr_task = set_curr_task_fair,
bf0f6f24
IM
1719 .task_tick = task_tick_fair,
1720 .task_new = task_new_fair,
cb469845
SR
1721
1722 .prio_changed = prio_changed_fair,
1723 .switched_to = switched_to_fair,
810b3817
PZ
1724
1725#ifdef CONFIG_FAIR_GROUP_SCHED
1726 .moved_group = moved_group_fair,
1727#endif
bf0f6f24
IM
1728};
1729
1730#ifdef CONFIG_SCHED_DEBUG
5cef9eca 1731static void print_cfs_stats(struct seq_file *m, int cpu)
bf0f6f24 1732{
bf0f6f24
IM
1733 struct cfs_rq *cfs_rq;
1734
5973e5b9 1735 rcu_read_lock();
c3b64f1e 1736 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5cef9eca 1737 print_cfs_rq(m, cpu, cfs_rq);
5973e5b9 1738 rcu_read_unlock();
bf0f6f24
IM
1739}
1740#endif
This page took 0.313195 seconds and 5 git commands to generate.