]>
Commit | Line | Data |
---|---|---|
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> | |
18 | * | |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
21 | */ | |
22 | ||
23 | #include <linux/latencytop.h> | |
24 | ||
25 | /* | |
26 | * Targeted preemption latency for CPU-bound tasks: | |
27 | * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds) | |
28 | * | |
29 | * NOTE: this latency value is not the same as the concept of | |
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. | |
33 | * | |
34 | * (to see the precise effective timeslice length of your workload, | |
35 | * run vmstat and monitor the context-switches (cs) field) | |
36 | */ | |
37 | unsigned int sysctl_sched_latency = 20000000ULL; | |
38 | ||
39 | /* | |
40 | * Minimal preemption granularity for CPU-bound tasks: | |
41 | * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
42 | */ | |
43 | unsigned int sysctl_sched_min_granularity = 4000000ULL; | |
44 | ||
45 | /* | |
46 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity | |
47 | */ | |
48 | static unsigned int sched_nr_latency = 5; | |
49 | ||
50 | /* | |
51 | * After fork, child runs first. (default) If set to 0 then | |
52 | * parent will (try to) run first. | |
53 | */ | |
54 | const_debug unsigned int sysctl_sched_child_runs_first = 1; | |
55 | ||
56 | /* | |
57 | * sys_sched_yield() compat mode | |
58 | * | |
59 | * This option switches the agressive yield implementation of the | |
60 | * old scheduler back on. | |
61 | */ | |
62 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
63 | ||
64 | /* | |
65 | * SCHED_OTHER wake-up granularity. | |
66 | * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
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 | */ | |
72 | unsigned int sysctl_sched_wakeup_granularity = 5000000UL; | |
73 | ||
74 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; | |
75 | ||
76 | static const struct sched_class fair_sched_class; | |
77 | ||
78 | /************************************************************** | |
79 | * CFS operations on generic schedulable entities: | |
80 | */ | |
81 | ||
82 | static inline struct task_struct *task_of(struct sched_entity *se) | |
83 | { | |
84 | return container_of(se, struct task_struct, se); | |
85 | } | |
86 | ||
87 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
88 | ||
89 | /* cpu runqueue to which this cfs_rq is attached */ | |
90 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
91 | { | |
92 | return cfs_rq->rq; | |
93 | } | |
94 | ||
95 | /* An entity is a task if it doesn't "own" a runqueue */ | |
96 | #define entity_is_task(se) (!se->my_q) | |
97 | ||
98 | /* Walk up scheduling entities hierarchy */ | |
99 | #define for_each_sched_entity(se) \ | |
100 | for (; se; se = se->parent) | |
101 | ||
102 | static 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 */ | |
108 | static 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 */ | |
114 | static 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 | */ | |
122 | static 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 ? */ | |
132 | static inline int | |
133 | is_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 | ||
141 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
142 | { | |
143 | return se->parent; | |
144 | } | |
145 | ||
146 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
147 | ||
148 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
149 | { | |
150 | return container_of(cfs_rq, struct rq, cfs); | |
151 | } | |
152 | ||
153 | #define entity_is_task(se) 1 | |
154 | ||
155 | #define for_each_sched_entity(se) \ | |
156 | for (; se; se = NULL) | |
157 | ||
158 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
159 | { | |
160 | return &task_rq(p)->cfs; | |
161 | } | |
162 | ||
163 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
164 | { | |
165 | struct task_struct *p = task_of(se); | |
166 | struct rq *rq = task_rq(p); | |
167 | ||
168 | return &rq->cfs; | |
169 | } | |
170 | ||
171 | /* runqueue "owned" by this group */ | |
172 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
173 | { | |
174 | return NULL; | |
175 | } | |
176 | ||
177 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
178 | { | |
179 | return &cpu_rq(this_cpu)->cfs; | |
180 | } | |
181 | ||
182 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
183 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
184 | ||
185 | static inline int | |
186 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
187 | { | |
188 | return 1; | |
189 | } | |
190 | ||
191 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
192 | { | |
193 | return NULL; | |
194 | } | |
195 | ||
196 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
197 | ||
198 | ||
199 | /************************************************************** | |
200 | * Scheduling class tree data structure manipulation methods: | |
201 | */ | |
202 | ||
203 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) | |
204 | { | |
205 | s64 delta = (s64)(vruntime - min_vruntime); | |
206 | if (delta > 0) | |
207 | min_vruntime = vruntime; | |
208 | ||
209 | return min_vruntime; | |
210 | } | |
211 | ||
212 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) | |
213 | { | |
214 | s64 delta = (s64)(vruntime - min_vruntime); | |
215 | if (delta < 0) | |
216 | min_vruntime = vruntime; | |
217 | ||
218 | return min_vruntime; | |
219 | } | |
220 | ||
221 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
222 | { | |
223 | return se->vruntime - cfs_rq->min_vruntime; | |
224 | } | |
225 | ||
226 | static void update_min_vruntime(struct cfs_rq *cfs_rq) | |
227 | { | |
228 | u64 vruntime = cfs_rq->min_vruntime; | |
229 | ||
230 | if (cfs_rq->curr) | |
231 | vruntime = cfs_rq->curr->vruntime; | |
232 | ||
233 | if (cfs_rq->rb_leftmost) { | |
234 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
235 | struct sched_entity, | |
236 | run_node); | |
237 | ||
238 | if (vruntime == cfs_rq->min_vruntime) | |
239 | vruntime = se->vruntime; | |
240 | else | |
241 | vruntime = min_vruntime(vruntime, se->vruntime); | |
242 | } | |
243 | ||
244 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
245 | } | |
246 | ||
247 | /* | |
248 | * Enqueue an entity into the rb-tree: | |
249 | */ | |
250 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
251 | { | |
252 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
253 | struct rb_node *parent = NULL; | |
254 | struct sched_entity *entry; | |
255 | s64 key = entity_key(cfs_rq, se); | |
256 | int leftmost = 1; | |
257 | ||
258 | /* | |
259 | * Find the right place in the rbtree: | |
260 | */ | |
261 | while (*link) { | |
262 | parent = *link; | |
263 | entry = rb_entry(parent, struct sched_entity, run_node); | |
264 | /* | |
265 | * We dont care about collisions. Nodes with | |
266 | * the same key stay together. | |
267 | */ | |
268 | if (key < entity_key(cfs_rq, entry)) { | |
269 | link = &parent->rb_left; | |
270 | } else { | |
271 | link = &parent->rb_right; | |
272 | leftmost = 0; | |
273 | } | |
274 | } | |
275 | ||
276 | /* | |
277 | * Maintain a cache of leftmost tree entries (it is frequently | |
278 | * used): | |
279 | */ | |
280 | if (leftmost) | |
281 | cfs_rq->rb_leftmost = &se->run_node; | |
282 | ||
283 | rb_link_node(&se->run_node, parent, link); | |
284 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
285 | } | |
286 | ||
287 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
288 | { | |
289 | if (cfs_rq->rb_leftmost == &se->run_node) { | |
290 | struct rb_node *next_node; | |
291 | ||
292 | next_node = rb_next(&se->run_node); | |
293 | cfs_rq->rb_leftmost = next_node; | |
294 | } | |
295 | ||
296 | if (cfs_rq->next == se) | |
297 | cfs_rq->next = NULL; | |
298 | ||
299 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
300 | } | |
301 | ||
302 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
303 | { | |
304 | return cfs_rq->rb_leftmost; | |
305 | } | |
306 | ||
307 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
308 | { | |
309 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
310 | } | |
311 | ||
312 | static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) | |
313 | { | |
314 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); | |
315 | ||
316 | if (!last) | |
317 | return NULL; | |
318 | ||
319 | return rb_entry(last, struct sched_entity, run_node); | |
320 | } | |
321 | ||
322 | /************************************************************** | |
323 | * Scheduling class statistics methods: | |
324 | */ | |
325 | ||
326 | #ifdef CONFIG_SCHED_DEBUG | |
327 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
328 | struct file *filp, void __user *buffer, size_t *lenp, | |
329 | loff_t *ppos) | |
330 | { | |
331 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
332 | ||
333 | if (ret || !write) | |
334 | return ret; | |
335 | ||
336 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
337 | sysctl_sched_min_granularity); | |
338 | ||
339 | return 0; | |
340 | } | |
341 | #endif | |
342 | ||
343 | /* | |
344 | * delta *= P[w / rw] | |
345 | */ | |
346 | static inline unsigned long | |
347 | calc_delta_weight(unsigned long delta, struct sched_entity *se) | |
348 | { | |
349 | for_each_sched_entity(se) { | |
350 | delta = calc_delta_mine(delta, | |
351 | se->load.weight, &cfs_rq_of(se)->load); | |
352 | } | |
353 | ||
354 | return delta; | |
355 | } | |
356 | ||
357 | /* | |
358 | * delta /= w | |
359 | */ | |
360 | static inline unsigned long | |
361 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
362 | { | |
363 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
364 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
365 | ||
366 | return delta; | |
367 | } | |
368 | ||
369 | /* | |
370 | * The idea is to set a period in which each task runs once. | |
371 | * | |
372 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
373 | * this period because otherwise the slices get too small. | |
374 | * | |
375 | * p = (nr <= nl) ? l : l*nr/nl | |
376 | */ | |
377 | static u64 __sched_period(unsigned long nr_running) | |
378 | { | |
379 | u64 period = sysctl_sched_latency; | |
380 | unsigned long nr_latency = sched_nr_latency; | |
381 | ||
382 | if (unlikely(nr_running > nr_latency)) { | |
383 | period = sysctl_sched_min_granularity; | |
384 | period *= nr_running; | |
385 | } | |
386 | ||
387 | return period; | |
388 | } | |
389 | ||
390 | /* | |
391 | * We calculate the wall-time slice from the period by taking a part | |
392 | * proportional to the weight. | |
393 | * | |
394 | * s = p*P[w/rw] | |
395 | */ | |
396 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
397 | { | |
398 | unsigned long nr_running = cfs_rq->nr_running; | |
399 | ||
400 | if (unlikely(!se->on_rq)) | |
401 | nr_running++; | |
402 | ||
403 | return calc_delta_weight(__sched_period(nr_running), se); | |
404 | } | |
405 | ||
406 | /* | |
407 | * We calculate the vruntime slice of a to be inserted task | |
408 | * | |
409 | * vs = s/w | |
410 | */ | |
411 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
412 | { | |
413 | return calc_delta_fair(sched_slice(cfs_rq, se), se); | |
414 | } | |
415 | ||
416 | /* | |
417 | * Update the current task's runtime statistics. Skip current tasks that | |
418 | * are not in our scheduling class. | |
419 | */ | |
420 | static inline void | |
421 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, | |
422 | unsigned long delta_exec) | |
423 | { | |
424 | unsigned long delta_exec_weighted; | |
425 | ||
426 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); | |
427 | ||
428 | curr->sum_exec_runtime += delta_exec; | |
429 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
430 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); | |
431 | curr->vruntime += delta_exec_weighted; | |
432 | update_min_vruntime(cfs_rq); | |
433 | } | |
434 | ||
435 | static void update_curr(struct cfs_rq *cfs_rq) | |
436 | { | |
437 | struct sched_entity *curr = cfs_rq->curr; | |
438 | u64 now = rq_of(cfs_rq)->clock; | |
439 | unsigned long delta_exec; | |
440 | ||
441 | if (unlikely(!curr)) | |
442 | return; | |
443 | ||
444 | /* | |
445 | * Get the amount of time the current task was running | |
446 | * since the last time we changed load (this cannot | |
447 | * overflow on 32 bits): | |
448 | */ | |
449 | delta_exec = (unsigned long)(now - curr->exec_start); | |
450 | ||
451 | __update_curr(cfs_rq, curr, delta_exec); | |
452 | curr->exec_start = now; | |
453 | ||
454 | if (entity_is_task(curr)) { | |
455 | struct task_struct *curtask = task_of(curr); | |
456 | ||
457 | cpuacct_charge(curtask, delta_exec); | |
458 | account_group_exec_runtime(curtask, delta_exec); | |
459 | } | |
460 | } | |
461 | ||
462 | static inline void | |
463 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
464 | { | |
465 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); | |
466 | } | |
467 | ||
468 | /* | |
469 | * Task is being enqueued - update stats: | |
470 | */ | |
471 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
472 | { | |
473 | /* | |
474 | * Are we enqueueing a waiting task? (for current tasks | |
475 | * a dequeue/enqueue event is a NOP) | |
476 | */ | |
477 | if (se != cfs_rq->curr) | |
478 | update_stats_wait_start(cfs_rq, se); | |
479 | } | |
480 | ||
481 | static void | |
482 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
483 | { | |
484 | schedstat_set(se->wait_max, max(se->wait_max, | |
485 | rq_of(cfs_rq)->clock - se->wait_start)); | |
486 | schedstat_set(se->wait_count, se->wait_count + 1); | |
487 | schedstat_set(se->wait_sum, se->wait_sum + | |
488 | rq_of(cfs_rq)->clock - se->wait_start); | |
489 | schedstat_set(se->wait_start, 0); | |
490 | } | |
491 | ||
492 | static inline void | |
493 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
494 | { | |
495 | /* | |
496 | * Mark the end of the wait period if dequeueing a | |
497 | * waiting task: | |
498 | */ | |
499 | if (se != cfs_rq->curr) | |
500 | update_stats_wait_end(cfs_rq, se); | |
501 | } | |
502 | ||
503 | /* | |
504 | * We are picking a new current task - update its stats: | |
505 | */ | |
506 | static inline void | |
507 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
508 | { | |
509 | /* | |
510 | * We are starting a new run period: | |
511 | */ | |
512 | se->exec_start = rq_of(cfs_rq)->clock; | |
513 | } | |
514 | ||
515 | /************************************************** | |
516 | * Scheduling class queueing methods: | |
517 | */ | |
518 | ||
519 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED | |
520 | static void | |
521 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
522 | { | |
523 | cfs_rq->task_weight += weight; | |
524 | } | |
525 | #else | |
526 | static inline void | |
527 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
528 | { | |
529 | } | |
530 | #endif | |
531 | ||
532 | static void | |
533 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
534 | { | |
535 | update_load_add(&cfs_rq->load, se->load.weight); | |
536 | if (!parent_entity(se)) | |
537 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
538 | if (entity_is_task(se)) { | |
539 | add_cfs_task_weight(cfs_rq, se->load.weight); | |
540 | list_add(&se->group_node, &cfs_rq->tasks); | |
541 | } | |
542 | cfs_rq->nr_running++; | |
543 | se->on_rq = 1; | |
544 | } | |
545 | ||
546 | static void | |
547 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
548 | { | |
549 | update_load_sub(&cfs_rq->load, se->load.weight); | |
550 | if (!parent_entity(se)) | |
551 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
552 | if (entity_is_task(se)) { | |
553 | add_cfs_task_weight(cfs_rq, -se->load.weight); | |
554 | list_del_init(&se->group_node); | |
555 | } | |
556 | cfs_rq->nr_running--; | |
557 | se->on_rq = 0; | |
558 | } | |
559 | ||
560 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
561 | { | |
562 | #ifdef CONFIG_SCHEDSTATS | |
563 | if (se->sleep_start) { | |
564 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; | |
565 | struct task_struct *tsk = task_of(se); | |
566 | ||
567 | if ((s64)delta < 0) | |
568 | delta = 0; | |
569 | ||
570 | if (unlikely(delta > se->sleep_max)) | |
571 | se->sleep_max = delta; | |
572 | ||
573 | se->sleep_start = 0; | |
574 | se->sum_sleep_runtime += delta; | |
575 | ||
576 | account_scheduler_latency(tsk, delta >> 10, 1); | |
577 | } | |
578 | if (se->block_start) { | |
579 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; | |
580 | struct task_struct *tsk = task_of(se); | |
581 | ||
582 | if ((s64)delta < 0) | |
583 | delta = 0; | |
584 | ||
585 | if (unlikely(delta > se->block_max)) | |
586 | se->block_max = delta; | |
587 | ||
588 | se->block_start = 0; | |
589 | se->sum_sleep_runtime += delta; | |
590 | ||
591 | /* | |
592 | * Blocking time is in units of nanosecs, so shift by 20 to | |
593 | * get a milliseconds-range estimation of the amount of | |
594 | * time that the task spent sleeping: | |
595 | */ | |
596 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
597 | ||
598 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
599 | delta >> 20); | |
600 | } | |
601 | account_scheduler_latency(tsk, delta >> 10, 0); | |
602 | } | |
603 | #endif | |
604 | } | |
605 | ||
606 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
607 | { | |
608 | #ifdef CONFIG_SCHED_DEBUG | |
609 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
610 | ||
611 | if (d < 0) | |
612 | d = -d; | |
613 | ||
614 | if (d > 3*sysctl_sched_latency) | |
615 | schedstat_inc(cfs_rq, nr_spread_over); | |
616 | #endif | |
617 | } | |
618 | ||
619 | static void | |
620 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
621 | { | |
622 | u64 vruntime = cfs_rq->min_vruntime; | |
623 | ||
624 | /* | |
625 | * The 'current' period is already promised to the current tasks, | |
626 | * however the extra weight of the new task will slow them down a | |
627 | * little, place the new task so that it fits in the slot that | |
628 | * stays open at the end. | |
629 | */ | |
630 | if (initial && sched_feat(START_DEBIT)) | |
631 | vruntime += sched_vslice(cfs_rq, se); | |
632 | ||
633 | if (!initial) { | |
634 | /* sleeps upto a single latency don't count. */ | |
635 | if (sched_feat(NEW_FAIR_SLEEPERS)) { | |
636 | unsigned long thresh = sysctl_sched_latency; | |
637 | ||
638 | /* | |
639 | * convert the sleeper threshold into virtual time | |
640 | */ | |
641 | if (sched_feat(NORMALIZED_SLEEPER)) | |
642 | thresh = calc_delta_fair(thresh, se); | |
643 | ||
644 | vruntime -= thresh; | |
645 | } | |
646 | ||
647 | /* ensure we never gain time by being placed backwards. */ | |
648 | vruntime = max_vruntime(se->vruntime, vruntime); | |
649 | } | |
650 | ||
651 | se->vruntime = vruntime; | |
652 | } | |
653 | ||
654 | static void | |
655 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) | |
656 | { | |
657 | /* | |
658 | * Update run-time statistics of the 'current'. | |
659 | */ | |
660 | update_curr(cfs_rq); | |
661 | account_entity_enqueue(cfs_rq, se); | |
662 | ||
663 | if (wakeup) { | |
664 | place_entity(cfs_rq, se, 0); | |
665 | enqueue_sleeper(cfs_rq, se); | |
666 | } | |
667 | ||
668 | update_stats_enqueue(cfs_rq, se); | |
669 | check_spread(cfs_rq, se); | |
670 | if (se != cfs_rq->curr) | |
671 | __enqueue_entity(cfs_rq, se); | |
672 | } | |
673 | ||
674 | static void | |
675 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) | |
676 | { | |
677 | /* | |
678 | * Update run-time statistics of the 'current'. | |
679 | */ | |
680 | update_curr(cfs_rq); | |
681 | ||
682 | update_stats_dequeue(cfs_rq, se); | |
683 | if (sleep) { | |
684 | #ifdef CONFIG_SCHEDSTATS | |
685 | if (entity_is_task(se)) { | |
686 | struct task_struct *tsk = task_of(se); | |
687 | ||
688 | if (tsk->state & TASK_INTERRUPTIBLE) | |
689 | se->sleep_start = rq_of(cfs_rq)->clock; | |
690 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
691 | se->block_start = rq_of(cfs_rq)->clock; | |
692 | } | |
693 | #endif | |
694 | } | |
695 | ||
696 | if (se != cfs_rq->curr) | |
697 | __dequeue_entity(cfs_rq, se); | |
698 | account_entity_dequeue(cfs_rq, se); | |
699 | update_min_vruntime(cfs_rq); | |
700 | } | |
701 | ||
702 | /* | |
703 | * Preempt the current task with a newly woken task if needed: | |
704 | */ | |
705 | static void | |
706 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
707 | { | |
708 | unsigned long ideal_runtime, delta_exec; | |
709 | ||
710 | ideal_runtime = sched_slice(cfs_rq, curr); | |
711 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
712 | if (delta_exec > ideal_runtime) | |
713 | resched_task(rq_of(cfs_rq)->curr); | |
714 | } | |
715 | ||
716 | static void | |
717 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
718 | { | |
719 | /* 'current' is not kept within the tree. */ | |
720 | if (se->on_rq) { | |
721 | /* | |
722 | * Any task has to be enqueued before it get to execute on | |
723 | * a CPU. So account for the time it spent waiting on the | |
724 | * runqueue. | |
725 | */ | |
726 | update_stats_wait_end(cfs_rq, se); | |
727 | __dequeue_entity(cfs_rq, se); | |
728 | } | |
729 | ||
730 | update_stats_curr_start(cfs_rq, se); | |
731 | cfs_rq->curr = se; | |
732 | #ifdef CONFIG_SCHEDSTATS | |
733 | /* | |
734 | * Track our maximum slice length, if the CPU's load is at | |
735 | * least twice that of our own weight (i.e. dont track it | |
736 | * when there are only lesser-weight tasks around): | |
737 | */ | |
738 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { | |
739 | se->slice_max = max(se->slice_max, | |
740 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
741 | } | |
742 | #endif | |
743 | se->prev_sum_exec_runtime = se->sum_exec_runtime; | |
744 | } | |
745 | ||
746 | static struct sched_entity * | |
747 | pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
748 | { | |
749 | struct rq *rq = rq_of(cfs_rq); | |
750 | u64 pair_slice = rq->clock - cfs_rq->pair_start; | |
751 | ||
752 | if (!cfs_rq->next || pair_slice > sysctl_sched_min_granularity) { | |
753 | cfs_rq->pair_start = rq->clock; | |
754 | return se; | |
755 | } | |
756 | ||
757 | return cfs_rq->next; | |
758 | } | |
759 | ||
760 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) | |
761 | { | |
762 | struct sched_entity *se = NULL; | |
763 | ||
764 | if (first_fair(cfs_rq)) { | |
765 | se = __pick_next_entity(cfs_rq); | |
766 | se = pick_next(cfs_rq, se); | |
767 | set_next_entity(cfs_rq, se); | |
768 | } | |
769 | ||
770 | return se; | |
771 | } | |
772 | ||
773 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) | |
774 | { | |
775 | /* | |
776 | * If still on the runqueue then deactivate_task() | |
777 | * was not called and update_curr() has to be done: | |
778 | */ | |
779 | if (prev->on_rq) | |
780 | update_curr(cfs_rq); | |
781 | ||
782 | check_spread(cfs_rq, prev); | |
783 | if (prev->on_rq) { | |
784 | update_stats_wait_start(cfs_rq, prev); | |
785 | /* Put 'current' back into the tree. */ | |
786 | __enqueue_entity(cfs_rq, prev); | |
787 | } | |
788 | cfs_rq->curr = NULL; | |
789 | } | |
790 | ||
791 | static void | |
792 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
793 | { | |
794 | /* | |
795 | * Update run-time statistics of the 'current'. | |
796 | */ | |
797 | update_curr(cfs_rq); | |
798 | ||
799 | #ifdef CONFIG_SCHED_HRTICK | |
800 | /* | |
801 | * queued ticks are scheduled to match the slice, so don't bother | |
802 | * validating it and just reschedule. | |
803 | */ | |
804 | if (queued) { | |
805 | resched_task(rq_of(cfs_rq)->curr); | |
806 | return; | |
807 | } | |
808 | /* | |
809 | * don't let the period tick interfere with the hrtick preemption | |
810 | */ | |
811 | if (!sched_feat(DOUBLE_TICK) && | |
812 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
813 | return; | |
814 | #endif | |
815 | ||
816 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) | |
817 | check_preempt_tick(cfs_rq, curr); | |
818 | } | |
819 | ||
820 | /************************************************** | |
821 | * CFS operations on tasks: | |
822 | */ | |
823 | ||
824 | #ifdef CONFIG_SCHED_HRTICK | |
825 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
826 | { | |
827 | struct sched_entity *se = &p->se; | |
828 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
829 | ||
830 | WARN_ON(task_rq(p) != rq); | |
831 | ||
832 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
833 | u64 slice = sched_slice(cfs_rq, se); | |
834 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
835 | s64 delta = slice - ran; | |
836 | ||
837 | if (delta < 0) { | |
838 | if (rq->curr == p) | |
839 | resched_task(p); | |
840 | return; | |
841 | } | |
842 | ||
843 | /* | |
844 | * Don't schedule slices shorter than 10000ns, that just | |
845 | * doesn't make sense. Rely on vruntime for fairness. | |
846 | */ | |
847 | if (rq->curr != p) | |
848 | delta = max_t(s64, 10000LL, delta); | |
849 | ||
850 | hrtick_start(rq, delta); | |
851 | } | |
852 | } | |
853 | ||
854 | /* | |
855 | * called from enqueue/dequeue and updates the hrtick when the | |
856 | * current task is from our class and nr_running is low enough | |
857 | * to matter. | |
858 | */ | |
859 | static void hrtick_update(struct rq *rq) | |
860 | { | |
861 | struct task_struct *curr = rq->curr; | |
862 | ||
863 | if (curr->sched_class != &fair_sched_class) | |
864 | return; | |
865 | ||
866 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
867 | hrtick_start_fair(rq, curr); | |
868 | } | |
869 | #else /* !CONFIG_SCHED_HRTICK */ | |
870 | static inline void | |
871 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
872 | { | |
873 | } | |
874 | ||
875 | static inline void hrtick_update(struct rq *rq) | |
876 | { | |
877 | } | |
878 | #endif | |
879 | ||
880 | /* | |
881 | * The enqueue_task method is called before nr_running is | |
882 | * increased. Here we update the fair scheduling stats and | |
883 | * then put the task into the rbtree: | |
884 | */ | |
885 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) | |
886 | { | |
887 | struct cfs_rq *cfs_rq; | |
888 | struct sched_entity *se = &p->se; | |
889 | ||
890 | for_each_sched_entity(se) { | |
891 | if (se->on_rq) | |
892 | break; | |
893 | cfs_rq = cfs_rq_of(se); | |
894 | enqueue_entity(cfs_rq, se, wakeup); | |
895 | wakeup = 1; | |
896 | } | |
897 | ||
898 | hrtick_update(rq); | |
899 | } | |
900 | ||
901 | /* | |
902 | * The dequeue_task method is called before nr_running is | |
903 | * decreased. We remove the task from the rbtree and | |
904 | * update the fair scheduling stats: | |
905 | */ | |
906 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) | |
907 | { | |
908 | struct cfs_rq *cfs_rq; | |
909 | struct sched_entity *se = &p->se; | |
910 | ||
911 | for_each_sched_entity(se) { | |
912 | cfs_rq = cfs_rq_of(se); | |
913 | dequeue_entity(cfs_rq, se, sleep); | |
914 | /* Don't dequeue parent if it has other entities besides us */ | |
915 | if (cfs_rq->load.weight) | |
916 | break; | |
917 | sleep = 1; | |
918 | } | |
919 | ||
920 | hrtick_update(rq); | |
921 | } | |
922 | ||
923 | /* | |
924 | * sched_yield() support is very simple - we dequeue and enqueue. | |
925 | * | |
926 | * If compat_yield is turned on then we requeue to the end of the tree. | |
927 | */ | |
928 | static void yield_task_fair(struct rq *rq) | |
929 | { | |
930 | struct task_struct *curr = rq->curr; | |
931 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
932 | struct sched_entity *rightmost, *se = &curr->se; | |
933 | ||
934 | /* | |
935 | * Are we the only task in the tree? | |
936 | */ | |
937 | if (unlikely(cfs_rq->nr_running == 1)) | |
938 | return; | |
939 | ||
940 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { | |
941 | update_rq_clock(rq); | |
942 | /* | |
943 | * Update run-time statistics of the 'current'. | |
944 | */ | |
945 | update_curr(cfs_rq); | |
946 | ||
947 | return; | |
948 | } | |
949 | /* | |
950 | * Find the rightmost entry in the rbtree: | |
951 | */ | |
952 | rightmost = __pick_last_entity(cfs_rq); | |
953 | /* | |
954 | * Already in the rightmost position? | |
955 | */ | |
956 | if (unlikely(!rightmost || rightmost->vruntime < se->vruntime)) | |
957 | return; | |
958 | ||
959 | /* | |
960 | * Minimally necessary key value to be last in the tree: | |
961 | * Upon rescheduling, sched_class::put_prev_task() will place | |
962 | * 'current' within the tree based on its new key value. | |
963 | */ | |
964 | se->vruntime = rightmost->vruntime + 1; | |
965 | } | |
966 | ||
967 | /* | |
968 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
969 | * not idle and an idle cpu is available. The span of cpus to | |
970 | * search starts with cpus closest then further out as needed, | |
971 | * so we always favor a closer, idle cpu. | |
972 | * Domains may include CPUs that are not usable for migration, | |
973 | * hence we need to mask them out (cpu_active_map) | |
974 | * | |
975 | * Returns the CPU we should wake onto. | |
976 | */ | |
977 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
978 | static int wake_idle(int cpu, struct task_struct *p) | |
979 | { | |
980 | cpumask_t tmp; | |
981 | struct sched_domain *sd; | |
982 | int i; | |
983 | ||
984 | /* | |
985 | * If it is idle, then it is the best cpu to run this task. | |
986 | * | |
987 | * This cpu is also the best, if it has more than one task already. | |
988 | * Siblings must be also busy(in most cases) as they didn't already | |
989 | * pickup the extra load from this cpu and hence we need not check | |
990 | * sibling runqueue info. This will avoid the checks and cache miss | |
991 | * penalities associated with that. | |
992 | */ | |
993 | if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1) | |
994 | return cpu; | |
995 | ||
996 | for_each_domain(cpu, sd) { | |
997 | if ((sd->flags & SD_WAKE_IDLE) | |
998 | || ((sd->flags & SD_WAKE_IDLE_FAR) | |
999 | && !task_hot(p, task_rq(p)->clock, sd))) { | |
1000 | cpus_and(tmp, sd->span, p->cpus_allowed); | |
1001 | cpus_and(tmp, tmp, cpu_active_map); | |
1002 | for_each_cpu_mask_nr(i, tmp) { | |
1003 | if (idle_cpu(i)) { | |
1004 | if (i != task_cpu(p)) { | |
1005 | schedstat_inc(p, | |
1006 | se.nr_wakeups_idle); | |
1007 | } | |
1008 | return i; | |
1009 | } | |
1010 | } | |
1011 | } else { | |
1012 | break; | |
1013 | } | |
1014 | } | |
1015 | return cpu; | |
1016 | } | |
1017 | #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/ | |
1018 | static inline int wake_idle(int cpu, struct task_struct *p) | |
1019 | { | |
1020 | return cpu; | |
1021 | } | |
1022 | #endif | |
1023 | ||
1024 | #ifdef CONFIG_SMP | |
1025 | ||
1026 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1027 | /* | |
1028 | * effective_load() calculates the load change as seen from the root_task_group | |
1029 | * | |
1030 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1031 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1032 | * can calculate the shift in shares. | |
1033 | * | |
1034 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1035 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1036 | * this change. | |
1037 | * | |
1038 | * We compensate this by not only taking the current delta into account, but | |
1039 | * also considering the delta between when the shares were last adjusted and | |
1040 | * now. | |
1041 | * | |
1042 | * We still saw a performance dip, some tracing learned us that between | |
1043 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1044 | * significantly. Therefore try to bias the error in direction of failing | |
1045 | * the affine wakeup. | |
1046 | * | |
1047 | */ | |
1048 | static long effective_load(struct task_group *tg, int cpu, | |
1049 | long wl, long wg) | |
1050 | { | |
1051 | struct sched_entity *se = tg->se[cpu]; | |
1052 | ||
1053 | if (!tg->parent) | |
1054 | return wl; | |
1055 | ||
1056 | /* | |
1057 | * By not taking the decrease of shares on the other cpu into | |
1058 | * account our error leans towards reducing the affine wakeups. | |
1059 | */ | |
1060 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1061 | return wl; | |
1062 | ||
1063 | for_each_sched_entity(se) { | |
1064 | long S, rw, s, a, b; | |
1065 | long more_w; | |
1066 | ||
1067 | /* | |
1068 | * Instead of using this increment, also add the difference | |
1069 | * between when the shares were last updated and now. | |
1070 | */ | |
1071 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1072 | wl += more_w; | |
1073 | wg += more_w; | |
1074 | ||
1075 | S = se->my_q->tg->shares; | |
1076 | s = se->my_q->shares; | |
1077 | rw = se->my_q->rq_weight; | |
1078 | ||
1079 | a = S*(rw + wl); | |
1080 | b = S*rw + s*wg; | |
1081 | ||
1082 | wl = s*(a-b); | |
1083 | ||
1084 | if (likely(b)) | |
1085 | wl /= b; | |
1086 | ||
1087 | /* | |
1088 | * Assume the group is already running and will | |
1089 | * thus already be accounted for in the weight. | |
1090 | * | |
1091 | * That is, moving shares between CPUs, does not | |
1092 | * alter the group weight. | |
1093 | */ | |
1094 | wg = 0; | |
1095 | } | |
1096 | ||
1097 | return wl; | |
1098 | } | |
1099 | ||
1100 | #else | |
1101 | ||
1102 | static inline unsigned long effective_load(struct task_group *tg, int cpu, | |
1103 | unsigned long wl, unsigned long wg) | |
1104 | { | |
1105 | return wl; | |
1106 | } | |
1107 | ||
1108 | #endif | |
1109 | ||
1110 | static int | |
1111 | wake_affine(struct sched_domain *this_sd, struct rq *this_rq, | |
1112 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, | |
1113 | int idx, unsigned long load, unsigned long this_load, | |
1114 | unsigned int imbalance) | |
1115 | { | |
1116 | struct task_struct *curr = this_rq->curr; | |
1117 | struct task_group *tg; | |
1118 | unsigned long tl = this_load; | |
1119 | unsigned long tl_per_task; | |
1120 | unsigned long weight; | |
1121 | int balanced; | |
1122 | ||
1123 | if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS)) | |
1124 | return 0; | |
1125 | ||
1126 | if (!sync && sched_feat(SYNC_WAKEUPS) && | |
1127 | curr->se.avg_overlap < sysctl_sched_migration_cost && | |
1128 | p->se.avg_overlap < sysctl_sched_migration_cost) | |
1129 | sync = 1; | |
1130 | ||
1131 | /* | |
1132 | * If sync wakeup then subtract the (maximum possible) | |
1133 | * effect of the currently running task from the load | |
1134 | * of the current CPU: | |
1135 | */ | |
1136 | if (sync) { | |
1137 | tg = task_group(current); | |
1138 | weight = current->se.load.weight; | |
1139 | ||
1140 | tl += effective_load(tg, this_cpu, -weight, -weight); | |
1141 | load += effective_load(tg, prev_cpu, 0, -weight); | |
1142 | } | |
1143 | ||
1144 | tg = task_group(p); | |
1145 | weight = p->se.load.weight; | |
1146 | ||
1147 | balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <= | |
1148 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); | |
1149 | ||
1150 | /* | |
1151 | * If the currently running task will sleep within | |
1152 | * a reasonable amount of time then attract this newly | |
1153 | * woken task: | |
1154 | */ | |
1155 | if (sync && balanced) | |
1156 | return 1; | |
1157 | ||
1158 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1159 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1160 | ||
1161 | if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <= | |
1162 | tl_per_task)) { | |
1163 | /* | |
1164 | * This domain has SD_WAKE_AFFINE and | |
1165 | * p is cache cold in this domain, and | |
1166 | * there is no bad imbalance. | |
1167 | */ | |
1168 | schedstat_inc(this_sd, ttwu_move_affine); | |
1169 | schedstat_inc(p, se.nr_wakeups_affine); | |
1170 | ||
1171 | return 1; | |
1172 | } | |
1173 | return 0; | |
1174 | } | |
1175 | ||
1176 | static int select_task_rq_fair(struct task_struct *p, int sync) | |
1177 | { | |
1178 | struct sched_domain *sd, *this_sd = NULL; | |
1179 | int prev_cpu, this_cpu, new_cpu; | |
1180 | unsigned long load, this_load; | |
1181 | struct rq *this_rq; | |
1182 | unsigned int imbalance; | |
1183 | int idx; | |
1184 | ||
1185 | prev_cpu = task_cpu(p); | |
1186 | this_cpu = smp_processor_id(); | |
1187 | this_rq = cpu_rq(this_cpu); | |
1188 | new_cpu = prev_cpu; | |
1189 | ||
1190 | if (prev_cpu == this_cpu) | |
1191 | goto out; | |
1192 | /* | |
1193 | * 'this_sd' is the first domain that both | |
1194 | * this_cpu and prev_cpu are present in: | |
1195 | */ | |
1196 | for_each_domain(this_cpu, sd) { | |
1197 | if (cpu_isset(prev_cpu, sd->span)) { | |
1198 | this_sd = sd; | |
1199 | break; | |
1200 | } | |
1201 | } | |
1202 | ||
1203 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) | |
1204 | goto out; | |
1205 | ||
1206 | /* | |
1207 | * Check for affine wakeup and passive balancing possibilities. | |
1208 | */ | |
1209 | if (!this_sd) | |
1210 | goto out; | |
1211 | ||
1212 | idx = this_sd->wake_idx; | |
1213 | ||
1214 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1215 | ||
1216 | load = source_load(prev_cpu, idx); | |
1217 | this_load = target_load(this_cpu, idx); | |
1218 | ||
1219 | if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, | |
1220 | load, this_load, imbalance)) | |
1221 | return this_cpu; | |
1222 | ||
1223 | /* | |
1224 | * Start passive balancing when half the imbalance_pct | |
1225 | * limit is reached. | |
1226 | */ | |
1227 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1228 | if (imbalance*this_load <= 100*load) { | |
1229 | schedstat_inc(this_sd, ttwu_move_balance); | |
1230 | schedstat_inc(p, se.nr_wakeups_passive); | |
1231 | return this_cpu; | |
1232 | } | |
1233 | } | |
1234 | ||
1235 | out: | |
1236 | return wake_idle(new_cpu, p); | |
1237 | } | |
1238 | #endif /* CONFIG_SMP */ | |
1239 | ||
1240 | static unsigned long wakeup_gran(struct sched_entity *se) | |
1241 | { | |
1242 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1243 | ||
1244 | /* | |
1245 | * More easily preempt - nice tasks, while not making it harder for | |
1246 | * + nice tasks. | |
1247 | */ | |
1248 | if (sched_feat(ASYM_GRAN)) | |
1249 | gran = calc_delta_mine(gran, NICE_0_LOAD, &se->load); | |
1250 | ||
1251 | return gran; | |
1252 | } | |
1253 | ||
1254 | /* | |
1255 | * Preempt the current task with a newly woken task if needed: | |
1256 | */ | |
1257 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync) | |
1258 | { | |
1259 | struct task_struct *curr = rq->curr; | |
1260 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1261 | struct sched_entity *se = &curr->se, *pse = &p->se; | |
1262 | s64 delta_exec; | |
1263 | ||
1264 | if (unlikely(rt_prio(p->prio))) { | |
1265 | update_rq_clock(rq); | |
1266 | update_curr(cfs_rq); | |
1267 | resched_task(curr); | |
1268 | return; | |
1269 | } | |
1270 | ||
1271 | if (unlikely(se == pse)) | |
1272 | return; | |
1273 | ||
1274 | cfs_rq_of(pse)->next = pse; | |
1275 | ||
1276 | /* | |
1277 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1278 | * wake up path. | |
1279 | */ | |
1280 | if (test_tsk_need_resched(curr)) | |
1281 | return; | |
1282 | ||
1283 | /* | |
1284 | * Batch tasks do not preempt (their preemption is driven by | |
1285 | * the tick): | |
1286 | */ | |
1287 | if (unlikely(p->policy == SCHED_BATCH)) | |
1288 | return; | |
1289 | ||
1290 | if (!sched_feat(WAKEUP_PREEMPT)) | |
1291 | return; | |
1292 | ||
1293 | if (sched_feat(WAKEUP_OVERLAP) && (sync || | |
1294 | (se->avg_overlap < sysctl_sched_migration_cost && | |
1295 | pse->avg_overlap < sysctl_sched_migration_cost))) { | |
1296 | resched_task(curr); | |
1297 | return; | |
1298 | } | |
1299 | ||
1300 | delta_exec = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1301 | if (delta_exec > wakeup_gran(pse)) | |
1302 | resched_task(curr); | |
1303 | } | |
1304 | ||
1305 | static struct task_struct *pick_next_task_fair(struct rq *rq) | |
1306 | { | |
1307 | struct task_struct *p; | |
1308 | struct cfs_rq *cfs_rq = &rq->cfs; | |
1309 | struct sched_entity *se; | |
1310 | ||
1311 | if (unlikely(!cfs_rq->nr_running)) | |
1312 | return NULL; | |
1313 | ||
1314 | do { | |
1315 | se = pick_next_entity(cfs_rq); | |
1316 | cfs_rq = group_cfs_rq(se); | |
1317 | } while (cfs_rq); | |
1318 | ||
1319 | p = task_of(se); | |
1320 | hrtick_start_fair(rq, p); | |
1321 | ||
1322 | return p; | |
1323 | } | |
1324 | ||
1325 | /* | |
1326 | * Account for a descheduled task: | |
1327 | */ | |
1328 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) | |
1329 | { | |
1330 | struct sched_entity *se = &prev->se; | |
1331 | struct cfs_rq *cfs_rq; | |
1332 | ||
1333 | for_each_sched_entity(se) { | |
1334 | cfs_rq = cfs_rq_of(se); | |
1335 | put_prev_entity(cfs_rq, se); | |
1336 | } | |
1337 | } | |
1338 | ||
1339 | #ifdef CONFIG_SMP | |
1340 | /************************************************** | |
1341 | * Fair scheduling class load-balancing methods: | |
1342 | */ | |
1343 | ||
1344 | /* | |
1345 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1346 | * during the whole iteration, the current task might be | |
1347 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1348 | * achieve that by always pre-iterating before returning | |
1349 | * the current task: | |
1350 | */ | |
1351 | static struct task_struct * | |
1352 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) | |
1353 | { | |
1354 | struct task_struct *p = NULL; | |
1355 | struct sched_entity *se; | |
1356 | ||
1357 | if (next == &cfs_rq->tasks) | |
1358 | return NULL; | |
1359 | ||
1360 | se = list_entry(next, struct sched_entity, group_node); | |
1361 | p = task_of(se); | |
1362 | cfs_rq->balance_iterator = next->next; | |
1363 | ||
1364 | return p; | |
1365 | } | |
1366 | ||
1367 | static struct task_struct *load_balance_start_fair(void *arg) | |
1368 | { | |
1369 | struct cfs_rq *cfs_rq = arg; | |
1370 | ||
1371 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); | |
1372 | } | |
1373 | ||
1374 | static struct task_struct *load_balance_next_fair(void *arg) | |
1375 | { | |
1376 | struct cfs_rq *cfs_rq = arg; | |
1377 | ||
1378 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); | |
1379 | } | |
1380 | ||
1381 | static unsigned long | |
1382 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1383 | unsigned long max_load_move, struct sched_domain *sd, | |
1384 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, | |
1385 | struct cfs_rq *cfs_rq) | |
1386 | { | |
1387 | struct rq_iterator cfs_rq_iterator; | |
1388 | ||
1389 | cfs_rq_iterator.start = load_balance_start_fair; | |
1390 | cfs_rq_iterator.next = load_balance_next_fair; | |
1391 | cfs_rq_iterator.arg = cfs_rq; | |
1392 | ||
1393 | return balance_tasks(this_rq, this_cpu, busiest, | |
1394 | max_load_move, sd, idle, all_pinned, | |
1395 | this_best_prio, &cfs_rq_iterator); | |
1396 | } | |
1397 | ||
1398 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1399 | static unsigned long | |
1400 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1401 | unsigned long max_load_move, | |
1402 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1403 | int *all_pinned, int *this_best_prio) | |
1404 | { | |
1405 | long rem_load_move = max_load_move; | |
1406 | int busiest_cpu = cpu_of(busiest); | |
1407 | struct task_group *tg; | |
1408 | ||
1409 | rcu_read_lock(); | |
1410 | update_h_load(busiest_cpu); | |
1411 | ||
1412 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
1413 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; | |
1414 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; | |
1415 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
1416 | u64 rem_load, moved_load; | |
1417 | ||
1418 | /* | |
1419 | * empty group | |
1420 | */ | |
1421 | if (!busiest_cfs_rq->task_weight) | |
1422 | continue; | |
1423 | ||
1424 | rem_load = (u64)rem_load_move * busiest_weight; | |
1425 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
1426 | ||
1427 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, | |
1428 | rem_load, sd, idle, all_pinned, this_best_prio, | |
1429 | tg->cfs_rq[busiest_cpu]); | |
1430 | ||
1431 | if (!moved_load) | |
1432 | continue; | |
1433 | ||
1434 | moved_load *= busiest_h_load; | |
1435 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
1436 | ||
1437 | rem_load_move -= moved_load; | |
1438 | if (rem_load_move < 0) | |
1439 | break; | |
1440 | } | |
1441 | rcu_read_unlock(); | |
1442 | ||
1443 | return max_load_move - rem_load_move; | |
1444 | } | |
1445 | #else | |
1446 | static unsigned long | |
1447 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1448 | unsigned long max_load_move, | |
1449 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1450 | int *all_pinned, int *this_best_prio) | |
1451 | { | |
1452 | return __load_balance_fair(this_rq, this_cpu, busiest, | |
1453 | max_load_move, sd, idle, all_pinned, | |
1454 | this_best_prio, &busiest->cfs); | |
1455 | } | |
1456 | #endif | |
1457 | ||
1458 | static int | |
1459 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1460 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1461 | { | |
1462 | struct cfs_rq *busy_cfs_rq; | |
1463 | struct rq_iterator cfs_rq_iterator; | |
1464 | ||
1465 | cfs_rq_iterator.start = load_balance_start_fair; | |
1466 | cfs_rq_iterator.next = load_balance_next_fair; | |
1467 | ||
1468 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1469 | /* | |
1470 | * pass busy_cfs_rq argument into | |
1471 | * load_balance_[start|next]_fair iterators | |
1472 | */ | |
1473 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1474 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1475 | &cfs_rq_iterator)) | |
1476 | return 1; | |
1477 | } | |
1478 | ||
1479 | return 0; | |
1480 | } | |
1481 | #endif /* CONFIG_SMP */ | |
1482 | ||
1483 | /* | |
1484 | * scheduler tick hitting a task of our scheduling class: | |
1485 | */ | |
1486 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) | |
1487 | { | |
1488 | struct cfs_rq *cfs_rq; | |
1489 | struct sched_entity *se = &curr->se; | |
1490 | ||
1491 | for_each_sched_entity(se) { | |
1492 | cfs_rq = cfs_rq_of(se); | |
1493 | entity_tick(cfs_rq, se, queued); | |
1494 | } | |
1495 | } | |
1496 | ||
1497 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) | |
1498 | ||
1499 | /* | |
1500 | * Share the fairness runtime between parent and child, thus the | |
1501 | * total amount of pressure for CPU stays equal - new tasks | |
1502 | * get a chance to run but frequent forkers are not allowed to | |
1503 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1504 | * the child is not running yet. | |
1505 | */ | |
1506 | static void task_new_fair(struct rq *rq, struct task_struct *p) | |
1507 | { | |
1508 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1509 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; | |
1510 | int this_cpu = smp_processor_id(); | |
1511 | ||
1512 | sched_info_queued(p); | |
1513 | ||
1514 | update_curr(cfs_rq); | |
1515 | place_entity(cfs_rq, se, 1); | |
1516 | ||
1517 | /* 'curr' will be NULL if the child belongs to a different group */ | |
1518 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && | |
1519 | curr && curr->vruntime < se->vruntime) { | |
1520 | /* | |
1521 | * Upon rescheduling, sched_class::put_prev_task() will place | |
1522 | * 'current' within the tree based on its new key value. | |
1523 | */ | |
1524 | swap(curr->vruntime, se->vruntime); | |
1525 | resched_task(rq->curr); | |
1526 | } | |
1527 | ||
1528 | enqueue_task_fair(rq, p, 0); | |
1529 | } | |
1530 | ||
1531 | /* | |
1532 | * Priority of the task has changed. Check to see if we preempt | |
1533 | * the current task. | |
1534 | */ | |
1535 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1536 | int oldprio, int running) | |
1537 | { | |
1538 | /* | |
1539 | * Reschedule if we are currently running on this runqueue and | |
1540 | * our priority decreased, or if we are not currently running on | |
1541 | * this runqueue and our priority is higher than the current's | |
1542 | */ | |
1543 | if (running) { | |
1544 | if (p->prio > oldprio) | |
1545 | resched_task(rq->curr); | |
1546 | } else | |
1547 | check_preempt_curr(rq, p, 0); | |
1548 | } | |
1549 | ||
1550 | /* | |
1551 | * We switched to the sched_fair class. | |
1552 | */ | |
1553 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1554 | int running) | |
1555 | { | |
1556 | /* | |
1557 | * We were most likely switched from sched_rt, so | |
1558 | * kick off the schedule if running, otherwise just see | |
1559 | * if we can still preempt the current task. | |
1560 | */ | |
1561 | if (running) | |
1562 | resched_task(rq->curr); | |
1563 | else | |
1564 | check_preempt_curr(rq, p, 0); | |
1565 | } | |
1566 | ||
1567 | /* Account for a task changing its policy or group. | |
1568 | * | |
1569 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1570 | * migrates between groups/classes. | |
1571 | */ | |
1572 | static void set_curr_task_fair(struct rq *rq) | |
1573 | { | |
1574 | struct sched_entity *se = &rq->curr->se; | |
1575 | ||
1576 | for_each_sched_entity(se) | |
1577 | set_next_entity(cfs_rq_of(se), se); | |
1578 | } | |
1579 | ||
1580 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1581 | static void moved_group_fair(struct task_struct *p) | |
1582 | { | |
1583 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1584 | ||
1585 | update_curr(cfs_rq); | |
1586 | place_entity(cfs_rq, &p->se, 1); | |
1587 | } | |
1588 | #endif | |
1589 | ||
1590 | /* | |
1591 | * All the scheduling class methods: | |
1592 | */ | |
1593 | static const struct sched_class fair_sched_class = { | |
1594 | .next = &idle_sched_class, | |
1595 | .enqueue_task = enqueue_task_fair, | |
1596 | .dequeue_task = dequeue_task_fair, | |
1597 | .yield_task = yield_task_fair, | |
1598 | ||
1599 | .check_preempt_curr = check_preempt_wakeup, | |
1600 | ||
1601 | .pick_next_task = pick_next_task_fair, | |
1602 | .put_prev_task = put_prev_task_fair, | |
1603 | ||
1604 | #ifdef CONFIG_SMP | |
1605 | .select_task_rq = select_task_rq_fair, | |
1606 | ||
1607 | .load_balance = load_balance_fair, | |
1608 | .move_one_task = move_one_task_fair, | |
1609 | #endif | |
1610 | ||
1611 | .set_curr_task = set_curr_task_fair, | |
1612 | .task_tick = task_tick_fair, | |
1613 | .task_new = task_new_fair, | |
1614 | ||
1615 | .prio_changed = prio_changed_fair, | |
1616 | .switched_to = switched_to_fair, | |
1617 | ||
1618 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1619 | .moved_group = moved_group_fair, | |
1620 | #endif | |
1621 | }; | |
1622 | ||
1623 | #ifdef CONFIG_SCHED_DEBUG | |
1624 | static void print_cfs_stats(struct seq_file *m, int cpu) | |
1625 | { | |
1626 | struct cfs_rq *cfs_rq; | |
1627 | ||
1628 | rcu_read_lock(); | |
1629 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) | |
1630 | print_cfs_rq(m, cpu, cfs_rq); | |
1631 | rcu_read_unlock(); | |
1632 | } | |
1633 | #endif |