]>
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 | #include <linux/sched.h> | |
25 | ||
26 | /* | |
27 | * Targeted preemption latency for CPU-bound tasks: | |
28 | * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds) | |
29 | * | |
30 | * NOTE: this latency value is not the same as the concept of | |
31 | * 'timeslice length' - timeslices in CFS are of variable length | |
32 | * and have no persistent notion like in traditional, time-slice | |
33 | * based scheduling concepts. | |
34 | * | |
35 | * (to see the precise effective timeslice length of your workload, | |
36 | * run vmstat and monitor the context-switches (cs) field) | |
37 | */ | |
38 | unsigned int sysctl_sched_latency = 6000000ULL; | |
39 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
40 | ||
41 | /* | |
42 | * The initial- and re-scaling of tunables is configurable | |
43 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
44 | * | |
45 | * Options are: | |
46 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
47 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
48 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
49 | */ | |
50 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
51 | = SCHED_TUNABLESCALING_LOG; | |
52 | ||
53 | /* | |
54 | * Minimal preemption granularity for CPU-bound tasks: | |
55 | * (default: 2 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
56 | */ | |
57 | unsigned int sysctl_sched_min_granularity = 2000000ULL; | |
58 | unsigned int normalized_sysctl_sched_min_granularity = 2000000ULL; | |
59 | ||
60 | /* | |
61 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity | |
62 | */ | |
63 | static unsigned int sched_nr_latency = 3; | |
64 | ||
65 | /* | |
66 | * After fork, child runs first. If set to 0 (default) then | |
67 | * parent will (try to) run first. | |
68 | */ | |
69 | unsigned int sysctl_sched_child_runs_first __read_mostly; | |
70 | ||
71 | /* | |
72 | * sys_sched_yield() compat mode | |
73 | * | |
74 | * This option switches the agressive yield implementation of the | |
75 | * old scheduler back on. | |
76 | */ | |
77 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
78 | ||
79 | /* | |
80 | * SCHED_OTHER wake-up granularity. | |
81 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
82 | * | |
83 | * This option delays the preemption effects of decoupled workloads | |
84 | * and reduces their over-scheduling. Synchronous workloads will still | |
85 | * have immediate wakeup/sleep latencies. | |
86 | */ | |
87 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; | |
88 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
89 | ||
90 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; | |
91 | ||
92 | static const struct sched_class fair_sched_class; | |
93 | ||
94 | /************************************************************** | |
95 | * CFS operations on generic schedulable entities: | |
96 | */ | |
97 | ||
98 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
99 | ||
100 | /* cpu runqueue to which this cfs_rq is attached */ | |
101 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
102 | { | |
103 | return cfs_rq->rq; | |
104 | } | |
105 | ||
106 | /* An entity is a task if it doesn't "own" a runqueue */ | |
107 | #define entity_is_task(se) (!se->my_q) | |
108 | ||
109 | static inline struct task_struct *task_of(struct sched_entity *se) | |
110 | { | |
111 | #ifdef CONFIG_SCHED_DEBUG | |
112 | WARN_ON_ONCE(!entity_is_task(se)); | |
113 | #endif | |
114 | return container_of(se, struct task_struct, se); | |
115 | } | |
116 | ||
117 | /* Walk up scheduling entities hierarchy */ | |
118 | #define for_each_sched_entity(se) \ | |
119 | for (; se; se = se->parent) | |
120 | ||
121 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
122 | { | |
123 | return p->se.cfs_rq; | |
124 | } | |
125 | ||
126 | /* runqueue on which this entity is (to be) queued */ | |
127 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
128 | { | |
129 | return se->cfs_rq; | |
130 | } | |
131 | ||
132 | /* runqueue "owned" by this group */ | |
133 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
134 | { | |
135 | return grp->my_q; | |
136 | } | |
137 | ||
138 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
139 | * another cpu ('this_cpu') | |
140 | */ | |
141 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
142 | { | |
143 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
144 | } | |
145 | ||
146 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
147 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
148 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
149 | ||
150 | /* Do the two (enqueued) entities belong to the same group ? */ | |
151 | static inline int | |
152 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
153 | { | |
154 | if (se->cfs_rq == pse->cfs_rq) | |
155 | return 1; | |
156 | ||
157 | return 0; | |
158 | } | |
159 | ||
160 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
161 | { | |
162 | return se->parent; | |
163 | } | |
164 | ||
165 | /* return depth at which a sched entity is present in the hierarchy */ | |
166 | static inline int depth_se(struct sched_entity *se) | |
167 | { | |
168 | int depth = 0; | |
169 | ||
170 | for_each_sched_entity(se) | |
171 | depth++; | |
172 | ||
173 | return depth; | |
174 | } | |
175 | ||
176 | static void | |
177 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
178 | { | |
179 | int se_depth, pse_depth; | |
180 | ||
181 | /* | |
182 | * preemption test can be made between sibling entities who are in the | |
183 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
184 | * both tasks until we find their ancestors who are siblings of common | |
185 | * parent. | |
186 | */ | |
187 | ||
188 | /* First walk up until both entities are at same depth */ | |
189 | se_depth = depth_se(*se); | |
190 | pse_depth = depth_se(*pse); | |
191 | ||
192 | while (se_depth > pse_depth) { | |
193 | se_depth--; | |
194 | *se = parent_entity(*se); | |
195 | } | |
196 | ||
197 | while (pse_depth > se_depth) { | |
198 | pse_depth--; | |
199 | *pse = parent_entity(*pse); | |
200 | } | |
201 | ||
202 | while (!is_same_group(*se, *pse)) { | |
203 | *se = parent_entity(*se); | |
204 | *pse = parent_entity(*pse); | |
205 | } | |
206 | } | |
207 | ||
208 | #else /* !CONFIG_FAIR_GROUP_SCHED */ | |
209 | ||
210 | static inline struct task_struct *task_of(struct sched_entity *se) | |
211 | { | |
212 | return container_of(se, struct task_struct, se); | |
213 | } | |
214 | ||
215 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
216 | { | |
217 | return container_of(cfs_rq, struct rq, cfs); | |
218 | } | |
219 | ||
220 | #define entity_is_task(se) 1 | |
221 | ||
222 | #define for_each_sched_entity(se) \ | |
223 | for (; se; se = NULL) | |
224 | ||
225 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
226 | { | |
227 | return &task_rq(p)->cfs; | |
228 | } | |
229 | ||
230 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
231 | { | |
232 | struct task_struct *p = task_of(se); | |
233 | struct rq *rq = task_rq(p); | |
234 | ||
235 | return &rq->cfs; | |
236 | } | |
237 | ||
238 | /* runqueue "owned" by this group */ | |
239 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
240 | { | |
241 | return NULL; | |
242 | } | |
243 | ||
244 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
245 | { | |
246 | return &cpu_rq(this_cpu)->cfs; | |
247 | } | |
248 | ||
249 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
250 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
251 | ||
252 | static inline int | |
253 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
254 | { | |
255 | return 1; | |
256 | } | |
257 | ||
258 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
259 | { | |
260 | return NULL; | |
261 | } | |
262 | ||
263 | static inline void | |
264 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
265 | { | |
266 | } | |
267 | ||
268 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
269 | ||
270 | ||
271 | /************************************************************** | |
272 | * Scheduling class tree data structure manipulation methods: | |
273 | */ | |
274 | ||
275 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) | |
276 | { | |
277 | s64 delta = (s64)(vruntime - min_vruntime); | |
278 | if (delta > 0) | |
279 | min_vruntime = vruntime; | |
280 | ||
281 | return min_vruntime; | |
282 | } | |
283 | ||
284 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) | |
285 | { | |
286 | s64 delta = (s64)(vruntime - min_vruntime); | |
287 | if (delta < 0) | |
288 | min_vruntime = vruntime; | |
289 | ||
290 | return min_vruntime; | |
291 | } | |
292 | ||
293 | static inline int entity_before(struct sched_entity *a, | |
294 | struct sched_entity *b) | |
295 | { | |
296 | return (s64)(a->vruntime - b->vruntime) < 0; | |
297 | } | |
298 | ||
299 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
300 | { | |
301 | return se->vruntime - cfs_rq->min_vruntime; | |
302 | } | |
303 | ||
304 | static void update_min_vruntime(struct cfs_rq *cfs_rq) | |
305 | { | |
306 | u64 vruntime = cfs_rq->min_vruntime; | |
307 | ||
308 | if (cfs_rq->curr) | |
309 | vruntime = cfs_rq->curr->vruntime; | |
310 | ||
311 | if (cfs_rq->rb_leftmost) { | |
312 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
313 | struct sched_entity, | |
314 | run_node); | |
315 | ||
316 | if (!cfs_rq->curr) | |
317 | vruntime = se->vruntime; | |
318 | else | |
319 | vruntime = min_vruntime(vruntime, se->vruntime); | |
320 | } | |
321 | ||
322 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
323 | } | |
324 | ||
325 | /* | |
326 | * Enqueue an entity into the rb-tree: | |
327 | */ | |
328 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
329 | { | |
330 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
331 | struct rb_node *parent = NULL; | |
332 | struct sched_entity *entry; | |
333 | s64 key = entity_key(cfs_rq, se); | |
334 | int leftmost = 1; | |
335 | ||
336 | /* | |
337 | * Find the right place in the rbtree: | |
338 | */ | |
339 | while (*link) { | |
340 | parent = *link; | |
341 | entry = rb_entry(parent, struct sched_entity, run_node); | |
342 | /* | |
343 | * We dont care about collisions. Nodes with | |
344 | * the same key stay together. | |
345 | */ | |
346 | if (key < entity_key(cfs_rq, entry)) { | |
347 | link = &parent->rb_left; | |
348 | } else { | |
349 | link = &parent->rb_right; | |
350 | leftmost = 0; | |
351 | } | |
352 | } | |
353 | ||
354 | /* | |
355 | * Maintain a cache of leftmost tree entries (it is frequently | |
356 | * used): | |
357 | */ | |
358 | if (leftmost) | |
359 | cfs_rq->rb_leftmost = &se->run_node; | |
360 | ||
361 | rb_link_node(&se->run_node, parent, link); | |
362 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
363 | } | |
364 | ||
365 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
366 | { | |
367 | if (cfs_rq->rb_leftmost == &se->run_node) { | |
368 | struct rb_node *next_node; | |
369 | ||
370 | next_node = rb_next(&se->run_node); | |
371 | cfs_rq->rb_leftmost = next_node; | |
372 | } | |
373 | ||
374 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
375 | } | |
376 | ||
377 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
378 | { | |
379 | struct rb_node *left = cfs_rq->rb_leftmost; | |
380 | ||
381 | if (!left) | |
382 | return NULL; | |
383 | ||
384 | return rb_entry(left, struct sched_entity, run_node); | |
385 | } | |
386 | ||
387 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) | |
388 | { | |
389 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); | |
390 | ||
391 | if (!last) | |
392 | return NULL; | |
393 | ||
394 | return rb_entry(last, struct sched_entity, run_node); | |
395 | } | |
396 | ||
397 | /************************************************************** | |
398 | * Scheduling class statistics methods: | |
399 | */ | |
400 | ||
401 | #ifdef CONFIG_SCHED_DEBUG | |
402 | int sched_proc_update_handler(struct ctl_table *table, int write, | |
403 | void __user *buffer, size_t *lenp, | |
404 | loff_t *ppos) | |
405 | { | |
406 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
407 | int factor = get_update_sysctl_factor(); | |
408 | ||
409 | if (ret || !write) | |
410 | return ret; | |
411 | ||
412 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
413 | sysctl_sched_min_granularity); | |
414 | ||
415 | #define WRT_SYSCTL(name) \ | |
416 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
417 | WRT_SYSCTL(sched_min_granularity); | |
418 | WRT_SYSCTL(sched_latency); | |
419 | WRT_SYSCTL(sched_wakeup_granularity); | |
420 | WRT_SYSCTL(sched_shares_ratelimit); | |
421 | #undef WRT_SYSCTL | |
422 | ||
423 | return 0; | |
424 | } | |
425 | #endif | |
426 | ||
427 | /* | |
428 | * delta /= w | |
429 | */ | |
430 | static inline unsigned long | |
431 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
432 | { | |
433 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
434 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
435 | ||
436 | return delta; | |
437 | } | |
438 | ||
439 | /* | |
440 | * The idea is to set a period in which each task runs once. | |
441 | * | |
442 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
443 | * this period because otherwise the slices get too small. | |
444 | * | |
445 | * p = (nr <= nl) ? l : l*nr/nl | |
446 | */ | |
447 | static u64 __sched_period(unsigned long nr_running) | |
448 | { | |
449 | u64 period = sysctl_sched_latency; | |
450 | unsigned long nr_latency = sched_nr_latency; | |
451 | ||
452 | if (unlikely(nr_running > nr_latency)) { | |
453 | period = sysctl_sched_min_granularity; | |
454 | period *= nr_running; | |
455 | } | |
456 | ||
457 | return period; | |
458 | } | |
459 | ||
460 | /* | |
461 | * We calculate the wall-time slice from the period by taking a part | |
462 | * proportional to the weight. | |
463 | * | |
464 | * s = p*P[w/rw] | |
465 | */ | |
466 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
467 | { | |
468 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); | |
469 | ||
470 | for_each_sched_entity(se) { | |
471 | struct load_weight *load; | |
472 | struct load_weight lw; | |
473 | ||
474 | cfs_rq = cfs_rq_of(se); | |
475 | load = &cfs_rq->load; | |
476 | ||
477 | if (unlikely(!se->on_rq)) { | |
478 | lw = cfs_rq->load; | |
479 | ||
480 | update_load_add(&lw, se->load.weight); | |
481 | load = &lw; | |
482 | } | |
483 | slice = calc_delta_mine(slice, se->load.weight, load); | |
484 | } | |
485 | return slice; | |
486 | } | |
487 | ||
488 | /* | |
489 | * We calculate the vruntime slice of a to be inserted task | |
490 | * | |
491 | * vs = s/w | |
492 | */ | |
493 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
494 | { | |
495 | return calc_delta_fair(sched_slice(cfs_rq, se), se); | |
496 | } | |
497 | ||
498 | /* | |
499 | * Update the current task's runtime statistics. Skip current tasks that | |
500 | * are not in our scheduling class. | |
501 | */ | |
502 | static inline void | |
503 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, | |
504 | unsigned long delta_exec) | |
505 | { | |
506 | unsigned long delta_exec_weighted; | |
507 | ||
508 | schedstat_set(curr->statistics.exec_max, | |
509 | max((u64)delta_exec, curr->statistics.exec_max)); | |
510 | ||
511 | curr->sum_exec_runtime += delta_exec; | |
512 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
513 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); | |
514 | ||
515 | curr->vruntime += delta_exec_weighted; | |
516 | update_min_vruntime(cfs_rq); | |
517 | } | |
518 | ||
519 | static void update_curr(struct cfs_rq *cfs_rq) | |
520 | { | |
521 | struct sched_entity *curr = cfs_rq->curr; | |
522 | u64 now = rq_of(cfs_rq)->clock; | |
523 | unsigned long delta_exec; | |
524 | ||
525 | if (unlikely(!curr)) | |
526 | return; | |
527 | ||
528 | /* | |
529 | * Get the amount of time the current task was running | |
530 | * since the last time we changed load (this cannot | |
531 | * overflow on 32 bits): | |
532 | */ | |
533 | delta_exec = (unsigned long)(now - curr->exec_start); | |
534 | if (!delta_exec) | |
535 | return; | |
536 | ||
537 | __update_curr(cfs_rq, curr, delta_exec); | |
538 | curr->exec_start = now; | |
539 | ||
540 | if (entity_is_task(curr)) { | |
541 | struct task_struct *curtask = task_of(curr); | |
542 | ||
543 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); | |
544 | cpuacct_charge(curtask, delta_exec); | |
545 | account_group_exec_runtime(curtask, delta_exec); | |
546 | } | |
547 | } | |
548 | ||
549 | static inline void | |
550 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
551 | { | |
552 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); | |
553 | } | |
554 | ||
555 | /* | |
556 | * Task is being enqueued - update stats: | |
557 | */ | |
558 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
559 | { | |
560 | /* | |
561 | * Are we enqueueing a waiting task? (for current tasks | |
562 | * a dequeue/enqueue event is a NOP) | |
563 | */ | |
564 | if (se != cfs_rq->curr) | |
565 | update_stats_wait_start(cfs_rq, se); | |
566 | } | |
567 | ||
568 | static void | |
569 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
570 | { | |
571 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, | |
572 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
573 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
574 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
575 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
576 | #ifdef CONFIG_SCHEDSTATS | |
577 | if (entity_is_task(se)) { | |
578 | trace_sched_stat_wait(task_of(se), | |
579 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
580 | } | |
581 | #endif | |
582 | schedstat_set(se->statistics.wait_start, 0); | |
583 | } | |
584 | ||
585 | static inline void | |
586 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
587 | { | |
588 | /* | |
589 | * Mark the end of the wait period if dequeueing a | |
590 | * waiting task: | |
591 | */ | |
592 | if (se != cfs_rq->curr) | |
593 | update_stats_wait_end(cfs_rq, se); | |
594 | } | |
595 | ||
596 | /* | |
597 | * We are picking a new current task - update its stats: | |
598 | */ | |
599 | static inline void | |
600 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
601 | { | |
602 | /* | |
603 | * We are starting a new run period: | |
604 | */ | |
605 | se->exec_start = rq_of(cfs_rq)->clock; | |
606 | } | |
607 | ||
608 | /************************************************** | |
609 | * Scheduling class queueing methods: | |
610 | */ | |
611 | ||
612 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED | |
613 | static void | |
614 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
615 | { | |
616 | cfs_rq->task_weight += weight; | |
617 | } | |
618 | #else | |
619 | static inline void | |
620 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
621 | { | |
622 | } | |
623 | #endif | |
624 | ||
625 | static void | |
626 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
627 | { | |
628 | update_load_add(&cfs_rq->load, se->load.weight); | |
629 | if (!parent_entity(se)) | |
630 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
631 | if (entity_is_task(se)) { | |
632 | add_cfs_task_weight(cfs_rq, se->load.weight); | |
633 | list_add(&se->group_node, &cfs_rq->tasks); | |
634 | } | |
635 | cfs_rq->nr_running++; | |
636 | se->on_rq = 1; | |
637 | } | |
638 | ||
639 | static void | |
640 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
641 | { | |
642 | update_load_sub(&cfs_rq->load, se->load.weight); | |
643 | if (!parent_entity(se)) | |
644 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
645 | if (entity_is_task(se)) { | |
646 | add_cfs_task_weight(cfs_rq, -se->load.weight); | |
647 | list_del_init(&se->group_node); | |
648 | } | |
649 | cfs_rq->nr_running--; | |
650 | se->on_rq = 0; | |
651 | } | |
652 | ||
653 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
654 | { | |
655 | #ifdef CONFIG_SCHEDSTATS | |
656 | struct task_struct *tsk = NULL; | |
657 | ||
658 | if (entity_is_task(se)) | |
659 | tsk = task_of(se); | |
660 | ||
661 | if (se->statistics.sleep_start) { | |
662 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
663 | ||
664 | if ((s64)delta < 0) | |
665 | delta = 0; | |
666 | ||
667 | if (unlikely(delta > se->statistics.sleep_max)) | |
668 | se->statistics.sleep_max = delta; | |
669 | ||
670 | se->statistics.sleep_start = 0; | |
671 | se->statistics.sum_sleep_runtime += delta; | |
672 | ||
673 | if (tsk) { | |
674 | account_scheduler_latency(tsk, delta >> 10, 1); | |
675 | trace_sched_stat_sleep(tsk, delta); | |
676 | } | |
677 | } | |
678 | if (se->statistics.block_start) { | |
679 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
680 | ||
681 | if ((s64)delta < 0) | |
682 | delta = 0; | |
683 | ||
684 | if (unlikely(delta > se->statistics.block_max)) | |
685 | se->statistics.block_max = delta; | |
686 | ||
687 | se->statistics.block_start = 0; | |
688 | se->statistics.sum_sleep_runtime += delta; | |
689 | ||
690 | if (tsk) { | |
691 | if (tsk->in_iowait) { | |
692 | se->statistics.iowait_sum += delta; | |
693 | se->statistics.iowait_count++; | |
694 | trace_sched_stat_iowait(tsk, delta); | |
695 | } | |
696 | ||
697 | /* | |
698 | * Blocking time is in units of nanosecs, so shift by | |
699 | * 20 to get a milliseconds-range estimation of the | |
700 | * amount of time that the task spent sleeping: | |
701 | */ | |
702 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
703 | profile_hits(SLEEP_PROFILING, | |
704 | (void *)get_wchan(tsk), | |
705 | delta >> 20); | |
706 | } | |
707 | account_scheduler_latency(tsk, delta >> 10, 0); | |
708 | } | |
709 | } | |
710 | #endif | |
711 | } | |
712 | ||
713 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
714 | { | |
715 | #ifdef CONFIG_SCHED_DEBUG | |
716 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
717 | ||
718 | if (d < 0) | |
719 | d = -d; | |
720 | ||
721 | if (d > 3*sysctl_sched_latency) | |
722 | schedstat_inc(cfs_rq, nr_spread_over); | |
723 | #endif | |
724 | } | |
725 | ||
726 | static void | |
727 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
728 | { | |
729 | u64 vruntime = cfs_rq->min_vruntime; | |
730 | ||
731 | /* | |
732 | * The 'current' period is already promised to the current tasks, | |
733 | * however the extra weight of the new task will slow them down a | |
734 | * little, place the new task so that it fits in the slot that | |
735 | * stays open at the end. | |
736 | */ | |
737 | if (initial && sched_feat(START_DEBIT)) | |
738 | vruntime += sched_vslice(cfs_rq, se); | |
739 | ||
740 | /* sleeps up to a single latency don't count. */ | |
741 | if (!initial) { | |
742 | unsigned long thresh = sysctl_sched_latency; | |
743 | ||
744 | /* | |
745 | * Halve their sleep time's effect, to allow | |
746 | * for a gentler effect of sleepers: | |
747 | */ | |
748 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
749 | thresh >>= 1; | |
750 | ||
751 | vruntime -= thresh; | |
752 | } | |
753 | ||
754 | /* ensure we never gain time by being placed backwards. */ | |
755 | vruntime = max_vruntime(se->vruntime, vruntime); | |
756 | ||
757 | se->vruntime = vruntime; | |
758 | } | |
759 | ||
760 | static void | |
761 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
762 | { | |
763 | /* | |
764 | * Update the normalized vruntime before updating min_vruntime | |
765 | * through callig update_curr(). | |
766 | */ | |
767 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) | |
768 | se->vruntime += cfs_rq->min_vruntime; | |
769 | ||
770 | /* | |
771 | * Update run-time statistics of the 'current'. | |
772 | */ | |
773 | update_curr(cfs_rq); | |
774 | account_entity_enqueue(cfs_rq, se); | |
775 | ||
776 | if (flags & ENQUEUE_WAKEUP) { | |
777 | place_entity(cfs_rq, se, 0); | |
778 | enqueue_sleeper(cfs_rq, se); | |
779 | } | |
780 | ||
781 | update_stats_enqueue(cfs_rq, se); | |
782 | check_spread(cfs_rq, se); | |
783 | if (se != cfs_rq->curr) | |
784 | __enqueue_entity(cfs_rq, se); | |
785 | } | |
786 | ||
787 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
788 | { | |
789 | if (!se || cfs_rq->last == se) | |
790 | cfs_rq->last = NULL; | |
791 | ||
792 | if (!se || cfs_rq->next == se) | |
793 | cfs_rq->next = NULL; | |
794 | } | |
795 | ||
796 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
797 | { | |
798 | for_each_sched_entity(se) | |
799 | __clear_buddies(cfs_rq_of(se), se); | |
800 | } | |
801 | ||
802 | static void | |
803 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
804 | { | |
805 | /* | |
806 | * Update run-time statistics of the 'current'. | |
807 | */ | |
808 | update_curr(cfs_rq); | |
809 | ||
810 | update_stats_dequeue(cfs_rq, se); | |
811 | if (flags & DEQUEUE_SLEEP) { | |
812 | #ifdef CONFIG_SCHEDSTATS | |
813 | if (entity_is_task(se)) { | |
814 | struct task_struct *tsk = task_of(se); | |
815 | ||
816 | if (tsk->state & TASK_INTERRUPTIBLE) | |
817 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; | |
818 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
819 | se->statistics.block_start = rq_of(cfs_rq)->clock; | |
820 | } | |
821 | #endif | |
822 | } | |
823 | ||
824 | clear_buddies(cfs_rq, se); | |
825 | ||
826 | if (se != cfs_rq->curr) | |
827 | __dequeue_entity(cfs_rq, se); | |
828 | account_entity_dequeue(cfs_rq, se); | |
829 | update_min_vruntime(cfs_rq); | |
830 | ||
831 | /* | |
832 | * Normalize the entity after updating the min_vruntime because the | |
833 | * update can refer to the ->curr item and we need to reflect this | |
834 | * movement in our normalized position. | |
835 | */ | |
836 | if (!(flags & DEQUEUE_SLEEP)) | |
837 | se->vruntime -= cfs_rq->min_vruntime; | |
838 | } | |
839 | ||
840 | /* | |
841 | * Preempt the current task with a newly woken task if needed: | |
842 | */ | |
843 | static void | |
844 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
845 | { | |
846 | unsigned long ideal_runtime, delta_exec; | |
847 | ||
848 | ideal_runtime = sched_slice(cfs_rq, curr); | |
849 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
850 | if (delta_exec > ideal_runtime) { | |
851 | resched_task(rq_of(cfs_rq)->curr); | |
852 | /* | |
853 | * The current task ran long enough, ensure it doesn't get | |
854 | * re-elected due to buddy favours. | |
855 | */ | |
856 | clear_buddies(cfs_rq, curr); | |
857 | return; | |
858 | } | |
859 | ||
860 | /* | |
861 | * Ensure that a task that missed wakeup preemption by a | |
862 | * narrow margin doesn't have to wait for a full slice. | |
863 | * This also mitigates buddy induced latencies under load. | |
864 | */ | |
865 | if (!sched_feat(WAKEUP_PREEMPT)) | |
866 | return; | |
867 | ||
868 | if (delta_exec < sysctl_sched_min_granularity) | |
869 | return; | |
870 | ||
871 | if (cfs_rq->nr_running > 1) { | |
872 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
873 | s64 delta = curr->vruntime - se->vruntime; | |
874 | ||
875 | if (delta > ideal_runtime) | |
876 | resched_task(rq_of(cfs_rq)->curr); | |
877 | } | |
878 | } | |
879 | ||
880 | static void | |
881 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
882 | { | |
883 | /* 'current' is not kept within the tree. */ | |
884 | if (se->on_rq) { | |
885 | /* | |
886 | * Any task has to be enqueued before it get to execute on | |
887 | * a CPU. So account for the time it spent waiting on the | |
888 | * runqueue. | |
889 | */ | |
890 | update_stats_wait_end(cfs_rq, se); | |
891 | __dequeue_entity(cfs_rq, se); | |
892 | } | |
893 | ||
894 | update_stats_curr_start(cfs_rq, se); | |
895 | cfs_rq->curr = se; | |
896 | #ifdef CONFIG_SCHEDSTATS | |
897 | /* | |
898 | * Track our maximum slice length, if the CPU's load is at | |
899 | * least twice that of our own weight (i.e. dont track it | |
900 | * when there are only lesser-weight tasks around): | |
901 | */ | |
902 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { | |
903 | se->statistics.slice_max = max(se->statistics.slice_max, | |
904 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
905 | } | |
906 | #endif | |
907 | se->prev_sum_exec_runtime = se->sum_exec_runtime; | |
908 | } | |
909 | ||
910 | static int | |
911 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
912 | ||
913 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) | |
914 | { | |
915 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
916 | struct sched_entity *left = se; | |
917 | ||
918 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
919 | se = cfs_rq->next; | |
920 | ||
921 | /* | |
922 | * Prefer last buddy, try to return the CPU to a preempted task. | |
923 | */ | |
924 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
925 | se = cfs_rq->last; | |
926 | ||
927 | clear_buddies(cfs_rq, se); | |
928 | ||
929 | return se; | |
930 | } | |
931 | ||
932 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) | |
933 | { | |
934 | /* | |
935 | * If still on the runqueue then deactivate_task() | |
936 | * was not called and update_curr() has to be done: | |
937 | */ | |
938 | if (prev->on_rq) | |
939 | update_curr(cfs_rq); | |
940 | ||
941 | check_spread(cfs_rq, prev); | |
942 | if (prev->on_rq) { | |
943 | update_stats_wait_start(cfs_rq, prev); | |
944 | /* Put 'current' back into the tree. */ | |
945 | __enqueue_entity(cfs_rq, prev); | |
946 | } | |
947 | cfs_rq->curr = NULL; | |
948 | } | |
949 | ||
950 | static void | |
951 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
952 | { | |
953 | /* | |
954 | * Update run-time statistics of the 'current'. | |
955 | */ | |
956 | update_curr(cfs_rq); | |
957 | ||
958 | #ifdef CONFIG_SCHED_HRTICK | |
959 | /* | |
960 | * queued ticks are scheduled to match the slice, so don't bother | |
961 | * validating it and just reschedule. | |
962 | */ | |
963 | if (queued) { | |
964 | resched_task(rq_of(cfs_rq)->curr); | |
965 | return; | |
966 | } | |
967 | /* | |
968 | * don't let the period tick interfere with the hrtick preemption | |
969 | */ | |
970 | if (!sched_feat(DOUBLE_TICK) && | |
971 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
972 | return; | |
973 | #endif | |
974 | ||
975 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) | |
976 | check_preempt_tick(cfs_rq, curr); | |
977 | } | |
978 | ||
979 | /************************************************** | |
980 | * CFS operations on tasks: | |
981 | */ | |
982 | ||
983 | #ifdef CONFIG_SCHED_HRTICK | |
984 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
985 | { | |
986 | struct sched_entity *se = &p->se; | |
987 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
988 | ||
989 | WARN_ON(task_rq(p) != rq); | |
990 | ||
991 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
992 | u64 slice = sched_slice(cfs_rq, se); | |
993 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
994 | s64 delta = slice - ran; | |
995 | ||
996 | if (delta < 0) { | |
997 | if (rq->curr == p) | |
998 | resched_task(p); | |
999 | return; | |
1000 | } | |
1001 | ||
1002 | /* | |
1003 | * Don't schedule slices shorter than 10000ns, that just | |
1004 | * doesn't make sense. Rely on vruntime for fairness. | |
1005 | */ | |
1006 | if (rq->curr != p) | |
1007 | delta = max_t(s64, 10000LL, delta); | |
1008 | ||
1009 | hrtick_start(rq, delta); | |
1010 | } | |
1011 | } | |
1012 | ||
1013 | /* | |
1014 | * called from enqueue/dequeue and updates the hrtick when the | |
1015 | * current task is from our class and nr_running is low enough | |
1016 | * to matter. | |
1017 | */ | |
1018 | static void hrtick_update(struct rq *rq) | |
1019 | { | |
1020 | struct task_struct *curr = rq->curr; | |
1021 | ||
1022 | if (curr->sched_class != &fair_sched_class) | |
1023 | return; | |
1024 | ||
1025 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1026 | hrtick_start_fair(rq, curr); | |
1027 | } | |
1028 | #else /* !CONFIG_SCHED_HRTICK */ | |
1029 | static inline void | |
1030 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1031 | { | |
1032 | } | |
1033 | ||
1034 | static inline void hrtick_update(struct rq *rq) | |
1035 | { | |
1036 | } | |
1037 | #endif | |
1038 | ||
1039 | /* | |
1040 | * The enqueue_task method is called before nr_running is | |
1041 | * increased. Here we update the fair scheduling stats and | |
1042 | * then put the task into the rbtree: | |
1043 | */ | |
1044 | static void | |
1045 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) | |
1046 | { | |
1047 | struct cfs_rq *cfs_rq; | |
1048 | struct sched_entity *se = &p->se; | |
1049 | ||
1050 | for_each_sched_entity(se) { | |
1051 | if (se->on_rq) | |
1052 | break; | |
1053 | cfs_rq = cfs_rq_of(se); | |
1054 | enqueue_entity(cfs_rq, se, flags); | |
1055 | flags = ENQUEUE_WAKEUP; | |
1056 | } | |
1057 | ||
1058 | hrtick_update(rq); | |
1059 | } | |
1060 | ||
1061 | /* | |
1062 | * The dequeue_task method is called before nr_running is | |
1063 | * decreased. We remove the task from the rbtree and | |
1064 | * update the fair scheduling stats: | |
1065 | */ | |
1066 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) | |
1067 | { | |
1068 | struct cfs_rq *cfs_rq; | |
1069 | struct sched_entity *se = &p->se; | |
1070 | ||
1071 | for_each_sched_entity(se) { | |
1072 | cfs_rq = cfs_rq_of(se); | |
1073 | dequeue_entity(cfs_rq, se, flags); | |
1074 | /* Don't dequeue parent if it has other entities besides us */ | |
1075 | if (cfs_rq->load.weight) | |
1076 | break; | |
1077 | flags |= DEQUEUE_SLEEP; | |
1078 | } | |
1079 | ||
1080 | hrtick_update(rq); | |
1081 | } | |
1082 | ||
1083 | /* | |
1084 | * sched_yield() support is very simple - we dequeue and enqueue. | |
1085 | * | |
1086 | * If compat_yield is turned on then we requeue to the end of the tree. | |
1087 | */ | |
1088 | static void yield_task_fair(struct rq *rq) | |
1089 | { | |
1090 | struct task_struct *curr = rq->curr; | |
1091 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1092 | struct sched_entity *rightmost, *se = &curr->se; | |
1093 | ||
1094 | /* | |
1095 | * Are we the only task in the tree? | |
1096 | */ | |
1097 | if (unlikely(cfs_rq->nr_running == 1)) | |
1098 | return; | |
1099 | ||
1100 | clear_buddies(cfs_rq, se); | |
1101 | ||
1102 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { | |
1103 | update_rq_clock(rq); | |
1104 | /* | |
1105 | * Update run-time statistics of the 'current'. | |
1106 | */ | |
1107 | update_curr(cfs_rq); | |
1108 | ||
1109 | return; | |
1110 | } | |
1111 | /* | |
1112 | * Find the rightmost entry in the rbtree: | |
1113 | */ | |
1114 | rightmost = __pick_last_entity(cfs_rq); | |
1115 | /* | |
1116 | * Already in the rightmost position? | |
1117 | */ | |
1118 | if (unlikely(!rightmost || entity_before(rightmost, se))) | |
1119 | return; | |
1120 | ||
1121 | /* | |
1122 | * Minimally necessary key value to be last in the tree: | |
1123 | * Upon rescheduling, sched_class::put_prev_task() will place | |
1124 | * 'current' within the tree based on its new key value. | |
1125 | */ | |
1126 | se->vruntime = rightmost->vruntime + 1; | |
1127 | } | |
1128 | ||
1129 | #ifdef CONFIG_SMP | |
1130 | ||
1131 | static void task_waking_fair(struct rq *rq, struct task_struct *p) | |
1132 | { | |
1133 | struct sched_entity *se = &p->se; | |
1134 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1135 | ||
1136 | se->vruntime -= cfs_rq->min_vruntime; | |
1137 | } | |
1138 | ||
1139 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1140 | /* | |
1141 | * effective_load() calculates the load change as seen from the root_task_group | |
1142 | * | |
1143 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1144 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1145 | * can calculate the shift in shares. | |
1146 | * | |
1147 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1148 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1149 | * this change. | |
1150 | * | |
1151 | * We compensate this by not only taking the current delta into account, but | |
1152 | * also considering the delta between when the shares were last adjusted and | |
1153 | * now. | |
1154 | * | |
1155 | * We still saw a performance dip, some tracing learned us that between | |
1156 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1157 | * significantly. Therefore try to bias the error in direction of failing | |
1158 | * the affine wakeup. | |
1159 | * | |
1160 | */ | |
1161 | static long effective_load(struct task_group *tg, int cpu, | |
1162 | long wl, long wg) | |
1163 | { | |
1164 | struct sched_entity *se = tg->se[cpu]; | |
1165 | ||
1166 | if (!tg->parent) | |
1167 | return wl; | |
1168 | ||
1169 | /* | |
1170 | * By not taking the decrease of shares on the other cpu into | |
1171 | * account our error leans towards reducing the affine wakeups. | |
1172 | */ | |
1173 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1174 | return wl; | |
1175 | ||
1176 | for_each_sched_entity(se) { | |
1177 | long S, rw, s, a, b; | |
1178 | long more_w; | |
1179 | ||
1180 | /* | |
1181 | * Instead of using this increment, also add the difference | |
1182 | * between when the shares were last updated and now. | |
1183 | */ | |
1184 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1185 | wl += more_w; | |
1186 | wg += more_w; | |
1187 | ||
1188 | S = se->my_q->tg->shares; | |
1189 | s = se->my_q->shares; | |
1190 | rw = se->my_q->rq_weight; | |
1191 | ||
1192 | a = S*(rw + wl); | |
1193 | b = S*rw + s*wg; | |
1194 | ||
1195 | wl = s*(a-b); | |
1196 | ||
1197 | if (likely(b)) | |
1198 | wl /= b; | |
1199 | ||
1200 | /* | |
1201 | * Assume the group is already running and will | |
1202 | * thus already be accounted for in the weight. | |
1203 | * | |
1204 | * That is, moving shares between CPUs, does not | |
1205 | * alter the group weight. | |
1206 | */ | |
1207 | wg = 0; | |
1208 | } | |
1209 | ||
1210 | return wl; | |
1211 | } | |
1212 | ||
1213 | #else | |
1214 | ||
1215 | static inline unsigned long effective_load(struct task_group *tg, int cpu, | |
1216 | unsigned long wl, unsigned long wg) | |
1217 | { | |
1218 | return wl; | |
1219 | } | |
1220 | ||
1221 | #endif | |
1222 | ||
1223 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) | |
1224 | { | |
1225 | unsigned long this_load, load; | |
1226 | int idx, this_cpu, prev_cpu; | |
1227 | unsigned long tl_per_task; | |
1228 | unsigned int imbalance; | |
1229 | struct task_group *tg; | |
1230 | unsigned long weight; | |
1231 | int balanced; | |
1232 | ||
1233 | idx = sd->wake_idx; | |
1234 | this_cpu = smp_processor_id(); | |
1235 | prev_cpu = task_cpu(p); | |
1236 | load = source_load(prev_cpu, idx); | |
1237 | this_load = target_load(this_cpu, idx); | |
1238 | ||
1239 | /* | |
1240 | * If sync wakeup then subtract the (maximum possible) | |
1241 | * effect of the currently running task from the load | |
1242 | * of the current CPU: | |
1243 | */ | |
1244 | if (sync) { | |
1245 | tg = task_group(current); | |
1246 | weight = current->se.load.weight; | |
1247 | ||
1248 | this_load += effective_load(tg, this_cpu, -weight, -weight); | |
1249 | load += effective_load(tg, prev_cpu, 0, -weight); | |
1250 | } | |
1251 | ||
1252 | tg = task_group(p); | |
1253 | weight = p->se.load.weight; | |
1254 | ||
1255 | imbalance = 100 + (sd->imbalance_pct - 100) / 2; | |
1256 | ||
1257 | /* | |
1258 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
1259 | * due to the sync cause above having dropped this_load to 0, we'll | |
1260 | * always have an imbalance, but there's really nothing you can do | |
1261 | * about that, so that's good too. | |
1262 | * | |
1263 | * Otherwise check if either cpus are near enough in load to allow this | |
1264 | * task to be woken on this_cpu. | |
1265 | */ | |
1266 | balanced = !this_load || | |
1267 | 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <= | |
1268 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); | |
1269 | ||
1270 | /* | |
1271 | * If the currently running task will sleep within | |
1272 | * a reasonable amount of time then attract this newly | |
1273 | * woken task: | |
1274 | */ | |
1275 | if (sync && balanced) | |
1276 | return 1; | |
1277 | ||
1278 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); | |
1279 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1280 | ||
1281 | if (balanced || | |
1282 | (this_load <= load && | |
1283 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
1284 | /* | |
1285 | * This domain has SD_WAKE_AFFINE and | |
1286 | * p is cache cold in this domain, and | |
1287 | * there is no bad imbalance. | |
1288 | */ | |
1289 | schedstat_inc(sd, ttwu_move_affine); | |
1290 | schedstat_inc(p, se.statistics.nr_wakeups_affine); | |
1291 | ||
1292 | return 1; | |
1293 | } | |
1294 | return 0; | |
1295 | } | |
1296 | ||
1297 | /* | |
1298 | * find_idlest_group finds and returns the least busy CPU group within the | |
1299 | * domain. | |
1300 | */ | |
1301 | static struct sched_group * | |
1302 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, | |
1303 | int this_cpu, int load_idx) | |
1304 | { | |
1305 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
1306 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
1307 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
1308 | ||
1309 | do { | |
1310 | unsigned long load, avg_load; | |
1311 | int local_group; | |
1312 | int i; | |
1313 | ||
1314 | /* Skip over this group if it has no CPUs allowed */ | |
1315 | if (!cpumask_intersects(sched_group_cpus(group), | |
1316 | &p->cpus_allowed)) | |
1317 | continue; | |
1318 | ||
1319 | local_group = cpumask_test_cpu(this_cpu, | |
1320 | sched_group_cpus(group)); | |
1321 | ||
1322 | /* Tally up the load of all CPUs in the group */ | |
1323 | avg_load = 0; | |
1324 | ||
1325 | for_each_cpu(i, sched_group_cpus(group)) { | |
1326 | /* Bias balancing toward cpus of our domain */ | |
1327 | if (local_group) | |
1328 | load = source_load(i, load_idx); | |
1329 | else | |
1330 | load = target_load(i, load_idx); | |
1331 | ||
1332 | avg_load += load; | |
1333 | } | |
1334 | ||
1335 | /* Adjust by relative CPU power of the group */ | |
1336 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1337 | ||
1338 | if (local_group) { | |
1339 | this_load = avg_load; | |
1340 | this = group; | |
1341 | } else if (avg_load < min_load) { | |
1342 | min_load = avg_load; | |
1343 | idlest = group; | |
1344 | } | |
1345 | } while (group = group->next, group != sd->groups); | |
1346 | ||
1347 | if (!idlest || 100*this_load < imbalance*min_load) | |
1348 | return NULL; | |
1349 | return idlest; | |
1350 | } | |
1351 | ||
1352 | /* | |
1353 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1354 | */ | |
1355 | static int | |
1356 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1357 | { | |
1358 | unsigned long load, min_load = ULONG_MAX; | |
1359 | int idlest = -1; | |
1360 | int i; | |
1361 | ||
1362 | /* Traverse only the allowed CPUs */ | |
1363 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1364 | load = weighted_cpuload(i); | |
1365 | ||
1366 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1367 | min_load = load; | |
1368 | idlest = i; | |
1369 | } | |
1370 | } | |
1371 | ||
1372 | return idlest; | |
1373 | } | |
1374 | ||
1375 | /* | |
1376 | * Try and locate an idle CPU in the sched_domain. | |
1377 | */ | |
1378 | static int select_idle_sibling(struct task_struct *p, int target) | |
1379 | { | |
1380 | int cpu = smp_processor_id(); | |
1381 | int prev_cpu = task_cpu(p); | |
1382 | struct sched_domain *sd; | |
1383 | int i; | |
1384 | ||
1385 | /* | |
1386 | * If the task is going to be woken-up on this cpu and if it is | |
1387 | * already idle, then it is the right target. | |
1388 | */ | |
1389 | if (target == cpu && idle_cpu(cpu)) | |
1390 | return cpu; | |
1391 | ||
1392 | /* | |
1393 | * If the task is going to be woken-up on the cpu where it previously | |
1394 | * ran and if it is currently idle, then it the right target. | |
1395 | */ | |
1396 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
1397 | return prev_cpu; | |
1398 | ||
1399 | /* | |
1400 | * Otherwise, iterate the domains and find an elegible idle cpu. | |
1401 | */ | |
1402 | for_each_domain(target, sd) { | |
1403 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
1404 | break; | |
1405 | ||
1406 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1407 | if (idle_cpu(i)) { | |
1408 | target = i; | |
1409 | break; | |
1410 | } | |
1411 | } | |
1412 | ||
1413 | /* | |
1414 | * Lets stop looking for an idle sibling when we reached | |
1415 | * the domain that spans the current cpu and prev_cpu. | |
1416 | */ | |
1417 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1418 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1419 | break; | |
1420 | } | |
1421 | ||
1422 | return target; | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1427 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1428 | * SD_BALANCE_EXEC. | |
1429 | * | |
1430 | * Balance, ie. select the least loaded group. | |
1431 | * | |
1432 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1433 | * | |
1434 | * preempt must be disabled. | |
1435 | */ | |
1436 | static int | |
1437 | select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags) | |
1438 | { | |
1439 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; | |
1440 | int cpu = smp_processor_id(); | |
1441 | int prev_cpu = task_cpu(p); | |
1442 | int new_cpu = cpu; | |
1443 | int want_affine = 0; | |
1444 | int want_sd = 1; | |
1445 | int sync = wake_flags & WF_SYNC; | |
1446 | ||
1447 | if (sd_flag & SD_BALANCE_WAKE) { | |
1448 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) | |
1449 | want_affine = 1; | |
1450 | new_cpu = prev_cpu; | |
1451 | } | |
1452 | ||
1453 | for_each_domain(cpu, tmp) { | |
1454 | if (!(tmp->flags & SD_LOAD_BALANCE)) | |
1455 | continue; | |
1456 | ||
1457 | /* | |
1458 | * If power savings logic is enabled for a domain, see if we | |
1459 | * are not overloaded, if so, don't balance wider. | |
1460 | */ | |
1461 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { | |
1462 | unsigned long power = 0; | |
1463 | unsigned long nr_running = 0; | |
1464 | unsigned long capacity; | |
1465 | int i; | |
1466 | ||
1467 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1468 | power += power_of(i); | |
1469 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1470 | } | |
1471 | ||
1472 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1473 | ||
1474 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) | |
1475 | nr_running /= 2; | |
1476 | ||
1477 | if (nr_running < capacity) | |
1478 | want_sd = 0; | |
1479 | } | |
1480 | ||
1481 | /* | |
1482 | * If both cpu and prev_cpu are part of this domain, | |
1483 | * cpu is a valid SD_WAKE_AFFINE target. | |
1484 | */ | |
1485 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && | |
1486 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1487 | affine_sd = tmp; | |
1488 | want_affine = 0; | |
1489 | } | |
1490 | ||
1491 | if (!want_sd && !want_affine) | |
1492 | break; | |
1493 | ||
1494 | if (!(tmp->flags & sd_flag)) | |
1495 | continue; | |
1496 | ||
1497 | if (want_sd) | |
1498 | sd = tmp; | |
1499 | } | |
1500 | ||
1501 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1502 | if (sched_feat(LB_SHARES_UPDATE)) { | |
1503 | /* | |
1504 | * Pick the largest domain to update shares over | |
1505 | */ | |
1506 | tmp = sd; | |
1507 | if (affine_sd && (!tmp || affine_sd->span_weight > sd->span_weight)) | |
1508 | tmp = affine_sd; | |
1509 | ||
1510 | if (tmp) { | |
1511 | raw_spin_unlock(&rq->lock); | |
1512 | update_shares(tmp); | |
1513 | raw_spin_lock(&rq->lock); | |
1514 | } | |
1515 | } | |
1516 | #endif | |
1517 | ||
1518 | if (affine_sd) { | |
1519 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) | |
1520 | return select_idle_sibling(p, cpu); | |
1521 | else | |
1522 | return select_idle_sibling(p, prev_cpu); | |
1523 | } | |
1524 | ||
1525 | while (sd) { | |
1526 | int load_idx = sd->forkexec_idx; | |
1527 | struct sched_group *group; | |
1528 | int weight; | |
1529 | ||
1530 | if (!(sd->flags & sd_flag)) { | |
1531 | sd = sd->child; | |
1532 | continue; | |
1533 | } | |
1534 | ||
1535 | if (sd_flag & SD_BALANCE_WAKE) | |
1536 | load_idx = sd->wake_idx; | |
1537 | ||
1538 | group = find_idlest_group(sd, p, cpu, load_idx); | |
1539 | if (!group) { | |
1540 | sd = sd->child; | |
1541 | continue; | |
1542 | } | |
1543 | ||
1544 | new_cpu = find_idlest_cpu(group, p, cpu); | |
1545 | if (new_cpu == -1 || new_cpu == cpu) { | |
1546 | /* Now try balancing at a lower domain level of cpu */ | |
1547 | sd = sd->child; | |
1548 | continue; | |
1549 | } | |
1550 | ||
1551 | /* Now try balancing at a lower domain level of new_cpu */ | |
1552 | cpu = new_cpu; | |
1553 | weight = sd->span_weight; | |
1554 | sd = NULL; | |
1555 | for_each_domain(cpu, tmp) { | |
1556 | if (weight <= tmp->span_weight) | |
1557 | break; | |
1558 | if (tmp->flags & sd_flag) | |
1559 | sd = tmp; | |
1560 | } | |
1561 | /* while loop will break here if sd == NULL */ | |
1562 | } | |
1563 | ||
1564 | return new_cpu; | |
1565 | } | |
1566 | #endif /* CONFIG_SMP */ | |
1567 | ||
1568 | static unsigned long | |
1569 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
1570 | { | |
1571 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1572 | ||
1573 | /* | |
1574 | * Since its curr running now, convert the gran from real-time | |
1575 | * to virtual-time in his units. | |
1576 | * | |
1577 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1578 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1579 | * the resulting gran will be larger, therefore penalizing the | |
1580 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1581 | * be smaller, again penalizing the lighter task. | |
1582 | * | |
1583 | * This is especially important for buddies when the leftmost | |
1584 | * task is higher priority than the buddy. | |
1585 | */ | |
1586 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1587 | gran = calc_delta_fair(gran, se); | |
1588 | ||
1589 | return gran; | |
1590 | } | |
1591 | ||
1592 | /* | |
1593 | * Should 'se' preempt 'curr'. | |
1594 | * | |
1595 | * |s1 | |
1596 | * |s2 | |
1597 | * |s3 | |
1598 | * g | |
1599 | * |<--->|c | |
1600 | * | |
1601 | * w(c, s1) = -1 | |
1602 | * w(c, s2) = 0 | |
1603 | * w(c, s3) = 1 | |
1604 | * | |
1605 | */ | |
1606 | static int | |
1607 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1608 | { | |
1609 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1610 | ||
1611 | if (vdiff <= 0) | |
1612 | return -1; | |
1613 | ||
1614 | gran = wakeup_gran(curr, se); | |
1615 | if (vdiff > gran) | |
1616 | return 1; | |
1617 | ||
1618 | return 0; | |
1619 | } | |
1620 | ||
1621 | static void set_last_buddy(struct sched_entity *se) | |
1622 | { | |
1623 | if (likely(task_of(se)->policy != SCHED_IDLE)) { | |
1624 | for_each_sched_entity(se) | |
1625 | cfs_rq_of(se)->last = se; | |
1626 | } | |
1627 | } | |
1628 | ||
1629 | static void set_next_buddy(struct sched_entity *se) | |
1630 | { | |
1631 | if (likely(task_of(se)->policy != SCHED_IDLE)) { | |
1632 | for_each_sched_entity(se) | |
1633 | cfs_rq_of(se)->next = se; | |
1634 | } | |
1635 | } | |
1636 | ||
1637 | /* | |
1638 | * Preempt the current task with a newly woken task if needed: | |
1639 | */ | |
1640 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) | |
1641 | { | |
1642 | struct task_struct *curr = rq->curr; | |
1643 | struct sched_entity *se = &curr->se, *pse = &p->se; | |
1644 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1645 | int scale = cfs_rq->nr_running >= sched_nr_latency; | |
1646 | ||
1647 | if (unlikely(rt_prio(p->prio))) | |
1648 | goto preempt; | |
1649 | ||
1650 | if (unlikely(p->sched_class != &fair_sched_class)) | |
1651 | return; | |
1652 | ||
1653 | if (unlikely(se == pse)) | |
1654 | return; | |
1655 | ||
1656 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) | |
1657 | set_next_buddy(pse); | |
1658 | ||
1659 | /* | |
1660 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1661 | * wake up path. | |
1662 | */ | |
1663 | if (test_tsk_need_resched(curr)) | |
1664 | return; | |
1665 | ||
1666 | /* | |
1667 | * Batch and idle tasks do not preempt (their preemption is driven by | |
1668 | * the tick): | |
1669 | */ | |
1670 | if (unlikely(p->policy != SCHED_NORMAL)) | |
1671 | return; | |
1672 | ||
1673 | /* Idle tasks are by definition preempted by everybody. */ | |
1674 | if (unlikely(curr->policy == SCHED_IDLE)) | |
1675 | goto preempt; | |
1676 | ||
1677 | if (!sched_feat(WAKEUP_PREEMPT)) | |
1678 | return; | |
1679 | ||
1680 | update_curr(cfs_rq); | |
1681 | find_matching_se(&se, &pse); | |
1682 | BUG_ON(!pse); | |
1683 | if (wakeup_preempt_entity(se, pse) == 1) | |
1684 | goto preempt; | |
1685 | ||
1686 | return; | |
1687 | ||
1688 | preempt: | |
1689 | resched_task(curr); | |
1690 | /* | |
1691 | * Only set the backward buddy when the current task is still | |
1692 | * on the rq. This can happen when a wakeup gets interleaved | |
1693 | * with schedule on the ->pre_schedule() or idle_balance() | |
1694 | * point, either of which can * drop the rq lock. | |
1695 | * | |
1696 | * Also, during early boot the idle thread is in the fair class, | |
1697 | * for obvious reasons its a bad idea to schedule back to it. | |
1698 | */ | |
1699 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1700 | return; | |
1701 | ||
1702 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1703 | set_last_buddy(se); | |
1704 | } | |
1705 | ||
1706 | static struct task_struct *pick_next_task_fair(struct rq *rq) | |
1707 | { | |
1708 | struct task_struct *p; | |
1709 | struct cfs_rq *cfs_rq = &rq->cfs; | |
1710 | struct sched_entity *se; | |
1711 | ||
1712 | if (!cfs_rq->nr_running) | |
1713 | return NULL; | |
1714 | ||
1715 | do { | |
1716 | se = pick_next_entity(cfs_rq); | |
1717 | set_next_entity(cfs_rq, se); | |
1718 | cfs_rq = group_cfs_rq(se); | |
1719 | } while (cfs_rq); | |
1720 | ||
1721 | p = task_of(se); | |
1722 | hrtick_start_fair(rq, p); | |
1723 | ||
1724 | return p; | |
1725 | } | |
1726 | ||
1727 | /* | |
1728 | * Account for a descheduled task: | |
1729 | */ | |
1730 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) | |
1731 | { | |
1732 | struct sched_entity *se = &prev->se; | |
1733 | struct cfs_rq *cfs_rq; | |
1734 | ||
1735 | for_each_sched_entity(se) { | |
1736 | cfs_rq = cfs_rq_of(se); | |
1737 | put_prev_entity(cfs_rq, se); | |
1738 | } | |
1739 | } | |
1740 | ||
1741 | #ifdef CONFIG_SMP | |
1742 | /************************************************** | |
1743 | * Fair scheduling class load-balancing methods: | |
1744 | */ | |
1745 | ||
1746 | /* | |
1747 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1748 | * Both runqueues must be locked. | |
1749 | */ | |
1750 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
1751 | struct rq *this_rq, int this_cpu) | |
1752 | { | |
1753 | deactivate_task(src_rq, p, 0); | |
1754 | set_task_cpu(p, this_cpu); | |
1755 | activate_task(this_rq, p, 0); | |
1756 | check_preempt_curr(this_rq, p, 0); | |
1757 | } | |
1758 | ||
1759 | /* | |
1760 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1761 | */ | |
1762 | static | |
1763 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
1764 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1765 | int *all_pinned) | |
1766 | { | |
1767 | int tsk_cache_hot = 0; | |
1768 | /* | |
1769 | * We do not migrate tasks that are: | |
1770 | * 1) running (obviously), or | |
1771 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1772 | * 3) are cache-hot on their current CPU. | |
1773 | */ | |
1774 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
1775 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); | |
1776 | return 0; | |
1777 | } | |
1778 | *all_pinned = 0; | |
1779 | ||
1780 | if (task_running(rq, p)) { | |
1781 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); | |
1782 | return 0; | |
1783 | } | |
1784 | ||
1785 | /* | |
1786 | * Aggressive migration if: | |
1787 | * 1) task is cache cold, or | |
1788 | * 2) too many balance attempts have failed. | |
1789 | */ | |
1790 | ||
1791 | tsk_cache_hot = task_hot(p, rq->clock, sd); | |
1792 | if (!tsk_cache_hot || | |
1793 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
1794 | #ifdef CONFIG_SCHEDSTATS | |
1795 | if (tsk_cache_hot) { | |
1796 | schedstat_inc(sd, lb_hot_gained[idle]); | |
1797 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
1798 | } | |
1799 | #endif | |
1800 | return 1; | |
1801 | } | |
1802 | ||
1803 | if (tsk_cache_hot) { | |
1804 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); | |
1805 | return 0; | |
1806 | } | |
1807 | return 1; | |
1808 | } | |
1809 | ||
1810 | /* | |
1811 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
1812 | * part of active balancing operations within "domain". | |
1813 | * Returns 1 if successful and 0 otherwise. | |
1814 | * | |
1815 | * Called with both runqueues locked. | |
1816 | */ | |
1817 | static int | |
1818 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1819 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1820 | { | |
1821 | struct task_struct *p, *n; | |
1822 | struct cfs_rq *cfs_rq; | |
1823 | int pinned = 0; | |
1824 | ||
1825 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
1826 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
1827 | ||
1828 | if (!can_migrate_task(p, busiest, this_cpu, | |
1829 | sd, idle, &pinned)) | |
1830 | continue; | |
1831 | ||
1832 | pull_task(busiest, p, this_rq, this_cpu); | |
1833 | /* | |
1834 | * Right now, this is only the second place pull_task() | |
1835 | * is called, so we can safely collect pull_task() | |
1836 | * stats here rather than inside pull_task(). | |
1837 | */ | |
1838 | schedstat_inc(sd, lb_gained[idle]); | |
1839 | return 1; | |
1840 | } | |
1841 | } | |
1842 | ||
1843 | return 0; | |
1844 | } | |
1845 | ||
1846 | static unsigned long | |
1847 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1848 | unsigned long max_load_move, struct sched_domain *sd, | |
1849 | enum cpu_idle_type idle, int *all_pinned, | |
1850 | int *this_best_prio, struct cfs_rq *busiest_cfs_rq) | |
1851 | { | |
1852 | int loops = 0, pulled = 0, pinned = 0; | |
1853 | long rem_load_move = max_load_move; | |
1854 | struct task_struct *p, *n; | |
1855 | ||
1856 | if (max_load_move == 0) | |
1857 | goto out; | |
1858 | ||
1859 | pinned = 1; | |
1860 | ||
1861 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { | |
1862 | if (loops++ > sysctl_sched_nr_migrate) | |
1863 | break; | |
1864 | ||
1865 | if ((p->se.load.weight >> 1) > rem_load_move || | |
1866 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) | |
1867 | continue; | |
1868 | ||
1869 | pull_task(busiest, p, this_rq, this_cpu); | |
1870 | pulled++; | |
1871 | rem_load_move -= p->se.load.weight; | |
1872 | ||
1873 | #ifdef CONFIG_PREEMPT | |
1874 | /* | |
1875 | * NEWIDLE balancing is a source of latency, so preemptible | |
1876 | * kernels will stop after the first task is pulled to minimize | |
1877 | * the critical section. | |
1878 | */ | |
1879 | if (idle == CPU_NEWLY_IDLE) | |
1880 | break; | |
1881 | #endif | |
1882 | ||
1883 | /* | |
1884 | * We only want to steal up to the prescribed amount of | |
1885 | * weighted load. | |
1886 | */ | |
1887 | if (rem_load_move <= 0) | |
1888 | break; | |
1889 | ||
1890 | if (p->prio < *this_best_prio) | |
1891 | *this_best_prio = p->prio; | |
1892 | } | |
1893 | out: | |
1894 | /* | |
1895 | * Right now, this is one of only two places pull_task() is called, | |
1896 | * so we can safely collect pull_task() stats here rather than | |
1897 | * inside pull_task(). | |
1898 | */ | |
1899 | schedstat_add(sd, lb_gained[idle], pulled); | |
1900 | ||
1901 | if (all_pinned) | |
1902 | *all_pinned = pinned; | |
1903 | ||
1904 | return max_load_move - rem_load_move; | |
1905 | } | |
1906 | ||
1907 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1908 | static unsigned long | |
1909 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1910 | unsigned long max_load_move, | |
1911 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1912 | int *all_pinned, int *this_best_prio) | |
1913 | { | |
1914 | long rem_load_move = max_load_move; | |
1915 | int busiest_cpu = cpu_of(busiest); | |
1916 | struct task_group *tg; | |
1917 | ||
1918 | rcu_read_lock(); | |
1919 | update_h_load(busiest_cpu); | |
1920 | ||
1921 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
1922 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; | |
1923 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; | |
1924 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
1925 | u64 rem_load, moved_load; | |
1926 | ||
1927 | /* | |
1928 | * empty group | |
1929 | */ | |
1930 | if (!busiest_cfs_rq->task_weight) | |
1931 | continue; | |
1932 | ||
1933 | rem_load = (u64)rem_load_move * busiest_weight; | |
1934 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
1935 | ||
1936 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
1937 | rem_load, sd, idle, all_pinned, this_best_prio, | |
1938 | busiest_cfs_rq); | |
1939 | ||
1940 | if (!moved_load) | |
1941 | continue; | |
1942 | ||
1943 | moved_load *= busiest_h_load; | |
1944 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
1945 | ||
1946 | rem_load_move -= moved_load; | |
1947 | if (rem_load_move < 0) | |
1948 | break; | |
1949 | } | |
1950 | rcu_read_unlock(); | |
1951 | ||
1952 | return max_load_move - rem_load_move; | |
1953 | } | |
1954 | #else | |
1955 | static unsigned long | |
1956 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1957 | unsigned long max_load_move, | |
1958 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1959 | int *all_pinned, int *this_best_prio) | |
1960 | { | |
1961 | return balance_tasks(this_rq, this_cpu, busiest, | |
1962 | max_load_move, sd, idle, all_pinned, | |
1963 | this_best_prio, &busiest->cfs); | |
1964 | } | |
1965 | #endif | |
1966 | ||
1967 | /* | |
1968 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
1969 | * this_rq, as part of a balancing operation within domain "sd". | |
1970 | * Returns 1 if successful and 0 otherwise. | |
1971 | * | |
1972 | * Called with both runqueues locked. | |
1973 | */ | |
1974 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1975 | unsigned long max_load_move, | |
1976 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1977 | int *all_pinned) | |
1978 | { | |
1979 | unsigned long total_load_moved = 0, load_moved; | |
1980 | int this_best_prio = this_rq->curr->prio; | |
1981 | ||
1982 | do { | |
1983 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, | |
1984 | max_load_move - total_load_moved, | |
1985 | sd, idle, all_pinned, &this_best_prio); | |
1986 | ||
1987 | total_load_moved += load_moved; | |
1988 | ||
1989 | #ifdef CONFIG_PREEMPT | |
1990 | /* | |
1991 | * NEWIDLE balancing is a source of latency, so preemptible | |
1992 | * kernels will stop after the first task is pulled to minimize | |
1993 | * the critical section. | |
1994 | */ | |
1995 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
1996 | break; | |
1997 | ||
1998 | if (raw_spin_is_contended(&this_rq->lock) || | |
1999 | raw_spin_is_contended(&busiest->lock)) | |
2000 | break; | |
2001 | #endif | |
2002 | } while (load_moved && max_load_move > total_load_moved); | |
2003 | ||
2004 | return total_load_moved > 0; | |
2005 | } | |
2006 | ||
2007 | /********** Helpers for find_busiest_group ************************/ | |
2008 | /* | |
2009 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2010 | * during load balancing. | |
2011 | */ | |
2012 | struct sd_lb_stats { | |
2013 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2014 | struct sched_group *this; /* Local group in this sd */ | |
2015 | unsigned long total_load; /* Total load of all groups in sd */ | |
2016 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2017 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2018 | ||
2019 | /** Statistics of this group */ | |
2020 | unsigned long this_load; | |
2021 | unsigned long this_load_per_task; | |
2022 | unsigned long this_nr_running; | |
2023 | ||
2024 | /* Statistics of the busiest group */ | |
2025 | unsigned long max_load; | |
2026 | unsigned long busiest_load_per_task; | |
2027 | unsigned long busiest_nr_running; | |
2028 | unsigned long busiest_group_capacity; | |
2029 | ||
2030 | int group_imb; /* Is there imbalance in this sd */ | |
2031 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2032 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2033 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2034 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2035 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2036 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2037 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2038 | #endif | |
2039 | }; | |
2040 | ||
2041 | /* | |
2042 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2043 | */ | |
2044 | struct sg_lb_stats { | |
2045 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2046 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2047 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2048 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2049 | unsigned long group_capacity; | |
2050 | int group_imb; /* Is there an imbalance in the group ? */ | |
2051 | }; | |
2052 | ||
2053 | /** | |
2054 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2055 | * @group: The group whose first cpu is to be returned. | |
2056 | */ | |
2057 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2058 | { | |
2059 | return cpumask_first(sched_group_cpus(group)); | |
2060 | } | |
2061 | ||
2062 | /** | |
2063 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2064 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2065 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2066 | */ | |
2067 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2068 | enum cpu_idle_type idle) | |
2069 | { | |
2070 | int load_idx; | |
2071 | ||
2072 | switch (idle) { | |
2073 | case CPU_NOT_IDLE: | |
2074 | load_idx = sd->busy_idx; | |
2075 | break; | |
2076 | ||
2077 | case CPU_NEWLY_IDLE: | |
2078 | load_idx = sd->newidle_idx; | |
2079 | break; | |
2080 | default: | |
2081 | load_idx = sd->idle_idx; | |
2082 | break; | |
2083 | } | |
2084 | ||
2085 | return load_idx; | |
2086 | } | |
2087 | ||
2088 | ||
2089 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2090 | /** | |
2091 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2092 | * the given sched_domain, during load balancing. | |
2093 | * | |
2094 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2095 | * @sds: Variable containing the statistics for sd. | |
2096 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2097 | */ | |
2098 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2099 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2100 | { | |
2101 | /* | |
2102 | * Busy processors will not participate in power savings | |
2103 | * balance. | |
2104 | */ | |
2105 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2106 | sds->power_savings_balance = 0; | |
2107 | else { | |
2108 | sds->power_savings_balance = 1; | |
2109 | sds->min_nr_running = ULONG_MAX; | |
2110 | sds->leader_nr_running = 0; | |
2111 | } | |
2112 | } | |
2113 | ||
2114 | /** | |
2115 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2116 | * sched_domain while performing load balancing. | |
2117 | * | |
2118 | * @group: sched_group belonging to the sched_domain under consideration. | |
2119 | * @sds: Variable containing the statistics of the sched_domain | |
2120 | * @local_group: Does group contain the CPU for which we're performing | |
2121 | * load balancing ? | |
2122 | * @sgs: Variable containing the statistics of the group. | |
2123 | */ | |
2124 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2125 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2126 | { | |
2127 | ||
2128 | if (!sds->power_savings_balance) | |
2129 | return; | |
2130 | ||
2131 | /* | |
2132 | * If the local group is idle or completely loaded | |
2133 | * no need to do power savings balance at this domain | |
2134 | */ | |
2135 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2136 | !sds->this_nr_running)) | |
2137 | sds->power_savings_balance = 0; | |
2138 | ||
2139 | /* | |
2140 | * If a group is already running at full capacity or idle, | |
2141 | * don't include that group in power savings calculations | |
2142 | */ | |
2143 | if (!sds->power_savings_balance || | |
2144 | sgs->sum_nr_running >= sgs->group_capacity || | |
2145 | !sgs->sum_nr_running) | |
2146 | return; | |
2147 | ||
2148 | /* | |
2149 | * Calculate the group which has the least non-idle load. | |
2150 | * This is the group from where we need to pick up the load | |
2151 | * for saving power | |
2152 | */ | |
2153 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2154 | (sgs->sum_nr_running == sds->min_nr_running && | |
2155 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2156 | sds->group_min = group; | |
2157 | sds->min_nr_running = sgs->sum_nr_running; | |
2158 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2159 | sgs->sum_nr_running; | |
2160 | } | |
2161 | ||
2162 | /* | |
2163 | * Calculate the group which is almost near its | |
2164 | * capacity but still has some space to pick up some load | |
2165 | * from other group and save more power | |
2166 | */ | |
2167 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2168 | return; | |
2169 | ||
2170 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2171 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2172 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2173 | sds->group_leader = group; | |
2174 | sds->leader_nr_running = sgs->sum_nr_running; | |
2175 | } | |
2176 | } | |
2177 | ||
2178 | /** | |
2179 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2180 | * @sds: Variable containing the statistics of the sched_domain | |
2181 | * under consideration. | |
2182 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2183 | * @imbalance: Variable to store the imbalance. | |
2184 | * | |
2185 | * Description: | |
2186 | * Check if we have potential to perform some power-savings balance. | |
2187 | * If yes, set the busiest group to be the least loaded group in the | |
2188 | * sched_domain, so that it's CPUs can be put to idle. | |
2189 | * | |
2190 | * Returns 1 if there is potential to perform power-savings balance. | |
2191 | * Else returns 0. | |
2192 | */ | |
2193 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2194 | int this_cpu, unsigned long *imbalance) | |
2195 | { | |
2196 | if (!sds->power_savings_balance) | |
2197 | return 0; | |
2198 | ||
2199 | if (sds->this != sds->group_leader || | |
2200 | sds->group_leader == sds->group_min) | |
2201 | return 0; | |
2202 | ||
2203 | *imbalance = sds->min_load_per_task; | |
2204 | sds->busiest = sds->group_min; | |
2205 | ||
2206 | return 1; | |
2207 | ||
2208 | } | |
2209 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2210 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2211 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2212 | { | |
2213 | return; | |
2214 | } | |
2215 | ||
2216 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2217 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2218 | { | |
2219 | return; | |
2220 | } | |
2221 | ||
2222 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2223 | int this_cpu, unsigned long *imbalance) | |
2224 | { | |
2225 | return 0; | |
2226 | } | |
2227 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2228 | ||
2229 | ||
2230 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2231 | { | |
2232 | return SCHED_LOAD_SCALE; | |
2233 | } | |
2234 | ||
2235 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2236 | { | |
2237 | return default_scale_freq_power(sd, cpu); | |
2238 | } | |
2239 | ||
2240 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2241 | { | |
2242 | unsigned long weight = sd->span_weight; | |
2243 | unsigned long smt_gain = sd->smt_gain; | |
2244 | ||
2245 | smt_gain /= weight; | |
2246 | ||
2247 | return smt_gain; | |
2248 | } | |
2249 | ||
2250 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2251 | { | |
2252 | return default_scale_smt_power(sd, cpu); | |
2253 | } | |
2254 | ||
2255 | unsigned long scale_rt_power(int cpu) | |
2256 | { | |
2257 | struct rq *rq = cpu_rq(cpu); | |
2258 | u64 total, available; | |
2259 | ||
2260 | sched_avg_update(rq); | |
2261 | ||
2262 | total = sched_avg_period() + (rq->clock - rq->age_stamp); | |
2263 | available = total - rq->rt_avg; | |
2264 | ||
2265 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
2266 | total = SCHED_LOAD_SCALE; | |
2267 | ||
2268 | total >>= SCHED_LOAD_SHIFT; | |
2269 | ||
2270 | return div_u64(available, total); | |
2271 | } | |
2272 | ||
2273 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2274 | { | |
2275 | unsigned long weight = sd->span_weight; | |
2276 | unsigned long power = SCHED_LOAD_SCALE; | |
2277 | struct sched_group *sdg = sd->groups; | |
2278 | ||
2279 | if (sched_feat(ARCH_POWER)) | |
2280 | power *= arch_scale_freq_power(sd, cpu); | |
2281 | else | |
2282 | power *= default_scale_freq_power(sd, cpu); | |
2283 | ||
2284 | power >>= SCHED_LOAD_SHIFT; | |
2285 | ||
2286 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { | |
2287 | if (sched_feat(ARCH_POWER)) | |
2288 | power *= arch_scale_smt_power(sd, cpu); | |
2289 | else | |
2290 | power *= default_scale_smt_power(sd, cpu); | |
2291 | ||
2292 | power >>= SCHED_LOAD_SHIFT; | |
2293 | } | |
2294 | ||
2295 | power *= scale_rt_power(cpu); | |
2296 | power >>= SCHED_LOAD_SHIFT; | |
2297 | ||
2298 | if (!power) | |
2299 | power = 1; | |
2300 | ||
2301 | sdg->cpu_power = power; | |
2302 | } | |
2303 | ||
2304 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2305 | { | |
2306 | struct sched_domain *child = sd->child; | |
2307 | struct sched_group *group, *sdg = sd->groups; | |
2308 | unsigned long power; | |
2309 | ||
2310 | if (!child) { | |
2311 | update_cpu_power(sd, cpu); | |
2312 | return; | |
2313 | } | |
2314 | ||
2315 | power = 0; | |
2316 | ||
2317 | group = child->groups; | |
2318 | do { | |
2319 | power += group->cpu_power; | |
2320 | group = group->next; | |
2321 | } while (group != child->groups); | |
2322 | ||
2323 | sdg->cpu_power = power; | |
2324 | } | |
2325 | ||
2326 | /** | |
2327 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2328 | * @sd: The sched_domain whose statistics are to be updated. | |
2329 | * @group: sched_group whose statistics are to be updated. | |
2330 | * @this_cpu: Cpu for which load balance is currently performed. | |
2331 | * @idle: Idle status of this_cpu | |
2332 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
2333 | * @sd_idle: Idle status of the sched_domain containing group. | |
2334 | * @local_group: Does group contain this_cpu. | |
2335 | * @cpus: Set of cpus considered for load balancing. | |
2336 | * @balance: Should we balance. | |
2337 | * @sgs: variable to hold the statistics for this group. | |
2338 | */ | |
2339 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2340 | struct sched_group *group, int this_cpu, | |
2341 | enum cpu_idle_type idle, int load_idx, int *sd_idle, | |
2342 | int local_group, const struct cpumask *cpus, | |
2343 | int *balance, struct sg_lb_stats *sgs) | |
2344 | { | |
2345 | unsigned long load, max_cpu_load, min_cpu_load; | |
2346 | int i; | |
2347 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
2348 | unsigned long avg_load_per_task = 0; | |
2349 | ||
2350 | if (local_group) | |
2351 | balance_cpu = group_first_cpu(group); | |
2352 | ||
2353 | /* Tally up the load of all CPUs in the group */ | |
2354 | max_cpu_load = 0; | |
2355 | min_cpu_load = ~0UL; | |
2356 | ||
2357 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2358 | struct rq *rq = cpu_rq(i); | |
2359 | ||
2360 | if (*sd_idle && rq->nr_running) | |
2361 | *sd_idle = 0; | |
2362 | ||
2363 | /* Bias balancing toward cpus of our domain */ | |
2364 | if (local_group) { | |
2365 | if (idle_cpu(i) && !first_idle_cpu) { | |
2366 | first_idle_cpu = 1; | |
2367 | balance_cpu = i; | |
2368 | } | |
2369 | ||
2370 | load = target_load(i, load_idx); | |
2371 | } else { | |
2372 | load = source_load(i, load_idx); | |
2373 | if (load > max_cpu_load) | |
2374 | max_cpu_load = load; | |
2375 | if (min_cpu_load > load) | |
2376 | min_cpu_load = load; | |
2377 | } | |
2378 | ||
2379 | sgs->group_load += load; | |
2380 | sgs->sum_nr_running += rq->nr_running; | |
2381 | sgs->sum_weighted_load += weighted_cpuload(i); | |
2382 | ||
2383 | } | |
2384 | ||
2385 | /* | |
2386 | * First idle cpu or the first cpu(busiest) in this sched group | |
2387 | * is eligible for doing load balancing at this and above | |
2388 | * domains. In the newly idle case, we will allow all the cpu's | |
2389 | * to do the newly idle load balance. | |
2390 | */ | |
2391 | if (idle != CPU_NEWLY_IDLE && local_group && | |
2392 | balance_cpu != this_cpu) { | |
2393 | *balance = 0; | |
2394 | return; | |
2395 | } | |
2396 | ||
2397 | update_group_power(sd, this_cpu); | |
2398 | ||
2399 | /* Adjust by relative CPU power of the group */ | |
2400 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2401 | ||
2402 | /* | |
2403 | * Consider the group unbalanced when the imbalance is larger | |
2404 | * than the average weight of two tasks. | |
2405 | * | |
2406 | * APZ: with cgroup the avg task weight can vary wildly and | |
2407 | * might not be a suitable number - should we keep a | |
2408 | * normalized nr_running number somewhere that negates | |
2409 | * the hierarchy? | |
2410 | */ | |
2411 | if (sgs->sum_nr_running) | |
2412 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
2413 | ||
2414 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task) | |
2415 | sgs->group_imb = 1; | |
2416 | ||
2417 | sgs->group_capacity = | |
2418 | DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); | |
2419 | } | |
2420 | ||
2421 | /** | |
2422 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2423 | * @sd: sched_domain whose statistics are to be updated. | |
2424 | * @this_cpu: Cpu for which load balance is currently performed. | |
2425 | * @idle: Idle status of this_cpu | |
2426 | * @sd_idle: Idle status of the sched_domain containing group. | |
2427 | * @cpus: Set of cpus considered for load balancing. | |
2428 | * @balance: Should we balance. | |
2429 | * @sds: variable to hold the statistics for this sched_domain. | |
2430 | */ | |
2431 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
2432 | enum cpu_idle_type idle, int *sd_idle, | |
2433 | const struct cpumask *cpus, int *balance, | |
2434 | struct sd_lb_stats *sds) | |
2435 | { | |
2436 | struct sched_domain *child = sd->child; | |
2437 | struct sched_group *group = sd->groups; | |
2438 | struct sg_lb_stats sgs; | |
2439 | int load_idx, prefer_sibling = 0; | |
2440 | ||
2441 | if (child && child->flags & SD_PREFER_SIBLING) | |
2442 | prefer_sibling = 1; | |
2443 | ||
2444 | init_sd_power_savings_stats(sd, sds, idle); | |
2445 | load_idx = get_sd_load_idx(sd, idle); | |
2446 | ||
2447 | do { | |
2448 | int local_group; | |
2449 | ||
2450 | local_group = cpumask_test_cpu(this_cpu, | |
2451 | sched_group_cpus(group)); | |
2452 | memset(&sgs, 0, sizeof(sgs)); | |
2453 | update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle, | |
2454 | local_group, cpus, balance, &sgs); | |
2455 | ||
2456 | if (local_group && !(*balance)) | |
2457 | return; | |
2458 | ||
2459 | sds->total_load += sgs.group_load; | |
2460 | sds->total_pwr += group->cpu_power; | |
2461 | ||
2462 | /* | |
2463 | * In case the child domain prefers tasks go to siblings | |
2464 | * first, lower the group capacity to one so that we'll try | |
2465 | * and move all the excess tasks away. | |
2466 | */ | |
2467 | if (prefer_sibling) | |
2468 | sgs.group_capacity = min(sgs.group_capacity, 1UL); | |
2469 | ||
2470 | if (local_group) { | |
2471 | sds->this_load = sgs.avg_load; | |
2472 | sds->this = group; | |
2473 | sds->this_nr_running = sgs.sum_nr_running; | |
2474 | sds->this_load_per_task = sgs.sum_weighted_load; | |
2475 | } else if (sgs.avg_load > sds->max_load && | |
2476 | (sgs.sum_nr_running > sgs.group_capacity || | |
2477 | sgs.group_imb)) { | |
2478 | sds->max_load = sgs.avg_load; | |
2479 | sds->busiest = group; | |
2480 | sds->busiest_nr_running = sgs.sum_nr_running; | |
2481 | sds->busiest_group_capacity = sgs.group_capacity; | |
2482 | sds->busiest_load_per_task = sgs.sum_weighted_load; | |
2483 | sds->group_imb = sgs.group_imb; | |
2484 | } | |
2485 | ||
2486 | update_sd_power_savings_stats(group, sds, local_group, &sgs); | |
2487 | group = group->next; | |
2488 | } while (group != sd->groups); | |
2489 | } | |
2490 | ||
2491 | /** | |
2492 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2493 | * amongst the groups of a sched_domain, during | |
2494 | * load balancing. | |
2495 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2496 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2497 | * @imbalance: Variable to store the imbalance. | |
2498 | */ | |
2499 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2500 | int this_cpu, unsigned long *imbalance) | |
2501 | { | |
2502 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2503 | unsigned int imbn = 2; | |
2504 | unsigned long scaled_busy_load_per_task; | |
2505 | ||
2506 | if (sds->this_nr_running) { | |
2507 | sds->this_load_per_task /= sds->this_nr_running; | |
2508 | if (sds->busiest_load_per_task > | |
2509 | sds->this_load_per_task) | |
2510 | imbn = 1; | |
2511 | } else | |
2512 | sds->this_load_per_task = | |
2513 | cpu_avg_load_per_task(this_cpu); | |
2514 | ||
2515 | scaled_busy_load_per_task = sds->busiest_load_per_task | |
2516 | * SCHED_LOAD_SCALE; | |
2517 | scaled_busy_load_per_task /= sds->busiest->cpu_power; | |
2518 | ||
2519 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
2520 | (scaled_busy_load_per_task * imbn)) { | |
2521 | *imbalance = sds->busiest_load_per_task; | |
2522 | return; | |
2523 | } | |
2524 | ||
2525 | /* | |
2526 | * OK, we don't have enough imbalance to justify moving tasks, | |
2527 | * however we may be able to increase total CPU power used by | |
2528 | * moving them. | |
2529 | */ | |
2530 | ||
2531 | pwr_now += sds->busiest->cpu_power * | |
2532 | min(sds->busiest_load_per_task, sds->max_load); | |
2533 | pwr_now += sds->this->cpu_power * | |
2534 | min(sds->this_load_per_task, sds->this_load); | |
2535 | pwr_now /= SCHED_LOAD_SCALE; | |
2536 | ||
2537 | /* Amount of load we'd subtract */ | |
2538 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2539 | sds->busiest->cpu_power; | |
2540 | if (sds->max_load > tmp) | |
2541 | pwr_move += sds->busiest->cpu_power * | |
2542 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
2543 | ||
2544 | /* Amount of load we'd add */ | |
2545 | if (sds->max_load * sds->busiest->cpu_power < | |
2546 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
2547 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
2548 | sds->this->cpu_power; | |
2549 | else | |
2550 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2551 | sds->this->cpu_power; | |
2552 | pwr_move += sds->this->cpu_power * | |
2553 | min(sds->this_load_per_task, sds->this_load + tmp); | |
2554 | pwr_move /= SCHED_LOAD_SCALE; | |
2555 | ||
2556 | /* Move if we gain throughput */ | |
2557 | if (pwr_move > pwr_now) | |
2558 | *imbalance = sds->busiest_load_per_task; | |
2559 | } | |
2560 | ||
2561 | /** | |
2562 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
2563 | * groups of a given sched_domain during load balance. | |
2564 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
2565 | * @this_cpu: Cpu for which currently load balance is being performed. | |
2566 | * @imbalance: The variable to store the imbalance. | |
2567 | */ | |
2568 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
2569 | unsigned long *imbalance) | |
2570 | { | |
2571 | unsigned long max_pull, load_above_capacity = ~0UL; | |
2572 | ||
2573 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
2574 | if (sds->group_imb) { | |
2575 | sds->busiest_load_per_task = | |
2576 | min(sds->busiest_load_per_task, sds->avg_load); | |
2577 | } | |
2578 | ||
2579 | /* | |
2580 | * In the presence of smp nice balancing, certain scenarios can have | |
2581 | * max load less than avg load(as we skip the groups at or below | |
2582 | * its cpu_power, while calculating max_load..) | |
2583 | */ | |
2584 | if (sds->max_load < sds->avg_load) { | |
2585 | *imbalance = 0; | |
2586 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2587 | } | |
2588 | ||
2589 | if (!sds->group_imb) { | |
2590 | /* | |
2591 | * Don't want to pull so many tasks that a group would go idle. | |
2592 | */ | |
2593 | load_above_capacity = (sds->busiest_nr_running - | |
2594 | sds->busiest_group_capacity); | |
2595 | ||
2596 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE); | |
2597 | ||
2598 | load_above_capacity /= sds->busiest->cpu_power; | |
2599 | } | |
2600 | ||
2601 | /* | |
2602 | * We're trying to get all the cpus to the average_load, so we don't | |
2603 | * want to push ourselves above the average load, nor do we wish to | |
2604 | * reduce the max loaded cpu below the average load. At the same time, | |
2605 | * we also don't want to reduce the group load below the group capacity | |
2606 | * (so that we can implement power-savings policies etc). Thus we look | |
2607 | * for the minimum possible imbalance. | |
2608 | * Be careful of negative numbers as they'll appear as very large values | |
2609 | * with unsigned longs. | |
2610 | */ | |
2611 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
2612 | ||
2613 | /* How much load to actually move to equalise the imbalance */ | |
2614 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
2615 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
2616 | / SCHED_LOAD_SCALE; | |
2617 | ||
2618 | /* | |
2619 | * if *imbalance is less than the average load per runnable task | |
2620 | * there is no gaurantee that any tasks will be moved so we'll have | |
2621 | * a think about bumping its value to force at least one task to be | |
2622 | * moved | |
2623 | */ | |
2624 | if (*imbalance < sds->busiest_load_per_task) | |
2625 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2626 | ||
2627 | } | |
2628 | /******* find_busiest_group() helpers end here *********************/ | |
2629 | ||
2630 | /** | |
2631 | * find_busiest_group - Returns the busiest group within the sched_domain | |
2632 | * if there is an imbalance. If there isn't an imbalance, and | |
2633 | * the user has opted for power-savings, it returns a group whose | |
2634 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
2635 | * such a group exists. | |
2636 | * | |
2637 | * Also calculates the amount of weighted load which should be moved | |
2638 | * to restore balance. | |
2639 | * | |
2640 | * @sd: The sched_domain whose busiest group is to be returned. | |
2641 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
2642 | * @imbalance: Variable which stores amount of weighted load which should | |
2643 | * be moved to restore balance/put a group to idle. | |
2644 | * @idle: The idle status of this_cpu. | |
2645 | * @sd_idle: The idleness of sd | |
2646 | * @cpus: The set of CPUs under consideration for load-balancing. | |
2647 | * @balance: Pointer to a variable indicating if this_cpu | |
2648 | * is the appropriate cpu to perform load balancing at this_level. | |
2649 | * | |
2650 | * Returns: - the busiest group if imbalance exists. | |
2651 | * - If no imbalance and user has opted for power-savings balance, | |
2652 | * return the least loaded group whose CPUs can be | |
2653 | * put to idle by rebalancing its tasks onto our group. | |
2654 | */ | |
2655 | static struct sched_group * | |
2656 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
2657 | unsigned long *imbalance, enum cpu_idle_type idle, | |
2658 | int *sd_idle, const struct cpumask *cpus, int *balance) | |
2659 | { | |
2660 | struct sd_lb_stats sds; | |
2661 | ||
2662 | memset(&sds, 0, sizeof(sds)); | |
2663 | ||
2664 | /* | |
2665 | * Compute the various statistics relavent for load balancing at | |
2666 | * this level. | |
2667 | */ | |
2668 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, | |
2669 | balance, &sds); | |
2670 | ||
2671 | /* Cases where imbalance does not exist from POV of this_cpu */ | |
2672 | /* 1) this_cpu is not the appropriate cpu to perform load balancing | |
2673 | * at this level. | |
2674 | * 2) There is no busy sibling group to pull from. | |
2675 | * 3) This group is the busiest group. | |
2676 | * 4) This group is more busy than the avg busieness at this | |
2677 | * sched_domain. | |
2678 | * 5) The imbalance is within the specified limit. | |
2679 | */ | |
2680 | if (!(*balance)) | |
2681 | goto ret; | |
2682 | ||
2683 | if (!sds.busiest || sds.busiest_nr_running == 0) | |
2684 | goto out_balanced; | |
2685 | ||
2686 | if (sds.this_load >= sds.max_load) | |
2687 | goto out_balanced; | |
2688 | ||
2689 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; | |
2690 | ||
2691 | if (sds.this_load >= sds.avg_load) | |
2692 | goto out_balanced; | |
2693 | ||
2694 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
2695 | goto out_balanced; | |
2696 | ||
2697 | /* Looks like there is an imbalance. Compute it */ | |
2698 | calculate_imbalance(&sds, this_cpu, imbalance); | |
2699 | return sds.busiest; | |
2700 | ||
2701 | out_balanced: | |
2702 | /* | |
2703 | * There is no obvious imbalance. But check if we can do some balancing | |
2704 | * to save power. | |
2705 | */ | |
2706 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
2707 | return sds.busiest; | |
2708 | ret: | |
2709 | *imbalance = 0; | |
2710 | return NULL; | |
2711 | } | |
2712 | ||
2713 | /* | |
2714 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2715 | */ | |
2716 | static struct rq * | |
2717 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, | |
2718 | unsigned long imbalance, const struct cpumask *cpus) | |
2719 | { | |
2720 | struct rq *busiest = NULL, *rq; | |
2721 | unsigned long max_load = 0; | |
2722 | int i; | |
2723 | ||
2724 | for_each_cpu(i, sched_group_cpus(group)) { | |
2725 | unsigned long power = power_of(i); | |
2726 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
2727 | unsigned long wl; | |
2728 | ||
2729 | if (!cpumask_test_cpu(i, cpus)) | |
2730 | continue; | |
2731 | ||
2732 | rq = cpu_rq(i); | |
2733 | wl = weighted_cpuload(i); | |
2734 | ||
2735 | /* | |
2736 | * When comparing with imbalance, use weighted_cpuload() | |
2737 | * which is not scaled with the cpu power. | |
2738 | */ | |
2739 | if (capacity && rq->nr_running == 1 && wl > imbalance) | |
2740 | continue; | |
2741 | ||
2742 | /* | |
2743 | * For the load comparisons with the other cpu's, consider | |
2744 | * the weighted_cpuload() scaled with the cpu power, so that | |
2745 | * the load can be moved away from the cpu that is potentially | |
2746 | * running at a lower capacity. | |
2747 | */ | |
2748 | wl = (wl * SCHED_LOAD_SCALE) / power; | |
2749 | ||
2750 | if (wl > max_load) { | |
2751 | max_load = wl; | |
2752 | busiest = rq; | |
2753 | } | |
2754 | } | |
2755 | ||
2756 | return busiest; | |
2757 | } | |
2758 | ||
2759 | /* | |
2760 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2761 | * so long as it is large enough. | |
2762 | */ | |
2763 | #define MAX_PINNED_INTERVAL 512 | |
2764 | ||
2765 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
2766 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
2767 | ||
2768 | static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle) | |
2769 | { | |
2770 | if (idle == CPU_NEWLY_IDLE) { | |
2771 | /* | |
2772 | * The only task running in a non-idle cpu can be moved to this | |
2773 | * cpu in an attempt to completely freeup the other CPU | |
2774 | * package. | |
2775 | * | |
2776 | * The package power saving logic comes from | |
2777 | * find_busiest_group(). If there are no imbalance, then | |
2778 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
2779 | * f_b_g() will select a group from which a running task may be | |
2780 | * pulled to this cpu in order to make the other package idle. | |
2781 | * If there is no opportunity to make a package idle and if | |
2782 | * there are no imbalance, then f_b_g() will return NULL and no | |
2783 | * action will be taken in load_balance_newidle(). | |
2784 | * | |
2785 | * Under normal task pull operation due to imbalance, there | |
2786 | * will be more than one task in the source run queue and | |
2787 | * move_tasks() will succeed. ld_moved will be true and this | |
2788 | * active balance code will not be triggered. | |
2789 | */ | |
2790 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
2791 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
2792 | return 0; | |
2793 | ||
2794 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) | |
2795 | return 0; | |
2796 | } | |
2797 | ||
2798 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
2799 | } | |
2800 | ||
2801 | /* | |
2802 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2803 | * tasks if there is an imbalance. | |
2804 | */ | |
2805 | static int load_balance(int this_cpu, struct rq *this_rq, | |
2806 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2807 | int *balance) | |
2808 | { | |
2809 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
2810 | struct sched_group *group; | |
2811 | unsigned long imbalance; | |
2812 | struct rq *busiest; | |
2813 | unsigned long flags; | |
2814 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
2815 | ||
2816 | cpumask_copy(cpus, cpu_active_mask); | |
2817 | ||
2818 | /* | |
2819 | * When power savings policy is enabled for the parent domain, idle | |
2820 | * sibling can pick up load irrespective of busy siblings. In this case, | |
2821 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
2822 | * portraying it as CPU_NOT_IDLE. | |
2823 | */ | |
2824 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
2825 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
2826 | sd_idle = 1; | |
2827 | ||
2828 | schedstat_inc(sd, lb_count[idle]); | |
2829 | ||
2830 | redo: | |
2831 | update_shares(sd); | |
2832 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
2833 | cpus, balance); | |
2834 | ||
2835 | if (*balance == 0) | |
2836 | goto out_balanced; | |
2837 | ||
2838 | if (!group) { | |
2839 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2840 | goto out_balanced; | |
2841 | } | |
2842 | ||
2843 | busiest = find_busiest_queue(group, idle, imbalance, cpus); | |
2844 | if (!busiest) { | |
2845 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2846 | goto out_balanced; | |
2847 | } | |
2848 | ||
2849 | BUG_ON(busiest == this_rq); | |
2850 | ||
2851 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2852 | ||
2853 | ld_moved = 0; | |
2854 | if (busiest->nr_running > 1) { | |
2855 | /* | |
2856 | * Attempt to move tasks. If find_busiest_group has found | |
2857 | * an imbalance but busiest->nr_running <= 1, the group is | |
2858 | * still unbalanced. ld_moved simply stays zero, so it is | |
2859 | * correctly treated as an imbalance. | |
2860 | */ | |
2861 | local_irq_save(flags); | |
2862 | double_rq_lock(this_rq, busiest); | |
2863 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
2864 | imbalance, sd, idle, &all_pinned); | |
2865 | double_rq_unlock(this_rq, busiest); | |
2866 | local_irq_restore(flags); | |
2867 | ||
2868 | /* | |
2869 | * some other cpu did the load balance for us. | |
2870 | */ | |
2871 | if (ld_moved && this_cpu != smp_processor_id()) | |
2872 | resched_cpu(this_cpu); | |
2873 | ||
2874 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2875 | if (unlikely(all_pinned)) { | |
2876 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
2877 | if (!cpumask_empty(cpus)) | |
2878 | goto redo; | |
2879 | goto out_balanced; | |
2880 | } | |
2881 | } | |
2882 | ||
2883 | if (!ld_moved) { | |
2884 | schedstat_inc(sd, lb_failed[idle]); | |
2885 | sd->nr_balance_failed++; | |
2886 | ||
2887 | if (need_active_balance(sd, sd_idle, idle)) { | |
2888 | raw_spin_lock_irqsave(&busiest->lock, flags); | |
2889 | ||
2890 | /* don't kick the migration_thread, if the curr | |
2891 | * task on busiest cpu can't be moved to this_cpu | |
2892 | */ | |
2893 | if (!cpumask_test_cpu(this_cpu, | |
2894 | &busiest->curr->cpus_allowed)) { | |
2895 | raw_spin_unlock_irqrestore(&busiest->lock, | |
2896 | flags); | |
2897 | all_pinned = 1; | |
2898 | goto out_one_pinned; | |
2899 | } | |
2900 | ||
2901 | if (!busiest->active_balance) { | |
2902 | busiest->active_balance = 1; | |
2903 | busiest->push_cpu = this_cpu; | |
2904 | active_balance = 1; | |
2905 | } | |
2906 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
2907 | if (active_balance) | |
2908 | wake_up_process(busiest->migration_thread); | |
2909 | ||
2910 | /* | |
2911 | * We've kicked active balancing, reset the failure | |
2912 | * counter. | |
2913 | */ | |
2914 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
2915 | } | |
2916 | } else | |
2917 | sd->nr_balance_failed = 0; | |
2918 | ||
2919 | if (likely(!active_balance)) { | |
2920 | /* We were unbalanced, so reset the balancing interval */ | |
2921 | sd->balance_interval = sd->min_interval; | |
2922 | } else { | |
2923 | /* | |
2924 | * If we've begun active balancing, start to back off. This | |
2925 | * case may not be covered by the all_pinned logic if there | |
2926 | * is only 1 task on the busy runqueue (because we don't call | |
2927 | * move_tasks). | |
2928 | */ | |
2929 | if (sd->balance_interval < sd->max_interval) | |
2930 | sd->balance_interval *= 2; | |
2931 | } | |
2932 | ||
2933 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
2934 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
2935 | ld_moved = -1; | |
2936 | ||
2937 | goto out; | |
2938 | ||
2939 | out_balanced: | |
2940 | schedstat_inc(sd, lb_balanced[idle]); | |
2941 | ||
2942 | sd->nr_balance_failed = 0; | |
2943 | ||
2944 | out_one_pinned: | |
2945 | /* tune up the balancing interval */ | |
2946 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
2947 | (sd->balance_interval < sd->max_interval)) | |
2948 | sd->balance_interval *= 2; | |
2949 | ||
2950 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
2951 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
2952 | ld_moved = -1; | |
2953 | else | |
2954 | ld_moved = 0; | |
2955 | out: | |
2956 | if (ld_moved) | |
2957 | update_shares(sd); | |
2958 | return ld_moved; | |
2959 | } | |
2960 | ||
2961 | /* | |
2962 | * idle_balance is called by schedule() if this_cpu is about to become | |
2963 | * idle. Attempts to pull tasks from other CPUs. | |
2964 | */ | |
2965 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
2966 | { | |
2967 | struct sched_domain *sd; | |
2968 | int pulled_task = 0; | |
2969 | unsigned long next_balance = jiffies + HZ; | |
2970 | ||
2971 | this_rq->idle_stamp = this_rq->clock; | |
2972 | ||
2973 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
2974 | return; | |
2975 | ||
2976 | /* | |
2977 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
2978 | */ | |
2979 | raw_spin_unlock(&this_rq->lock); | |
2980 | ||
2981 | for_each_domain(this_cpu, sd) { | |
2982 | unsigned long interval; | |
2983 | int balance = 1; | |
2984 | ||
2985 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
2986 | continue; | |
2987 | ||
2988 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
2989 | /* If we've pulled tasks over stop searching: */ | |
2990 | pulled_task = load_balance(this_cpu, this_rq, | |
2991 | sd, CPU_NEWLY_IDLE, &balance); | |
2992 | } | |
2993 | ||
2994 | interval = msecs_to_jiffies(sd->balance_interval); | |
2995 | if (time_after(next_balance, sd->last_balance + interval)) | |
2996 | next_balance = sd->last_balance + interval; | |
2997 | if (pulled_task) { | |
2998 | this_rq->idle_stamp = 0; | |
2999 | break; | |
3000 | } | |
3001 | } | |
3002 | ||
3003 | raw_spin_lock(&this_rq->lock); | |
3004 | ||
3005 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { | |
3006 | /* | |
3007 | * We are going idle. next_balance may be set based on | |
3008 | * a busy processor. So reset next_balance. | |
3009 | */ | |
3010 | this_rq->next_balance = next_balance; | |
3011 | } | |
3012 | } | |
3013 | ||
3014 | /* | |
3015 | * active_load_balance is run by migration threads. It pushes running tasks | |
3016 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
3017 | * running on each physical CPU where possible, and avoids physical / | |
3018 | * logical imbalances. | |
3019 | * | |
3020 | * Called with busiest_rq locked. | |
3021 | */ | |
3022 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) | |
3023 | { | |
3024 | int target_cpu = busiest_rq->push_cpu; | |
3025 | struct sched_domain *sd; | |
3026 | struct rq *target_rq; | |
3027 | ||
3028 | /* Is there any task to move? */ | |
3029 | if (busiest_rq->nr_running <= 1) | |
3030 | return; | |
3031 | ||
3032 | target_rq = cpu_rq(target_cpu); | |
3033 | ||
3034 | /* | |
3035 | * This condition is "impossible", if it occurs | |
3036 | * we need to fix it. Originally reported by | |
3037 | * Bjorn Helgaas on a 128-cpu setup. | |
3038 | */ | |
3039 | BUG_ON(busiest_rq == target_rq); | |
3040 | ||
3041 | /* move a task from busiest_rq to target_rq */ | |
3042 | double_lock_balance(busiest_rq, target_rq); | |
3043 | ||
3044 | /* Search for an sd spanning us and the target CPU. */ | |
3045 | for_each_domain(target_cpu, sd) { | |
3046 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3047 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3048 | break; | |
3049 | } | |
3050 | ||
3051 | if (likely(sd)) { | |
3052 | schedstat_inc(sd, alb_count); | |
3053 | ||
3054 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3055 | sd, CPU_IDLE)) | |
3056 | schedstat_inc(sd, alb_pushed); | |
3057 | else | |
3058 | schedstat_inc(sd, alb_failed); | |
3059 | } | |
3060 | double_unlock_balance(busiest_rq, target_rq); | |
3061 | } | |
3062 | ||
3063 | #ifdef CONFIG_NO_HZ | |
3064 | static struct { | |
3065 | atomic_t load_balancer; | |
3066 | cpumask_var_t cpu_mask; | |
3067 | cpumask_var_t ilb_grp_nohz_mask; | |
3068 | } nohz ____cacheline_aligned = { | |
3069 | .load_balancer = ATOMIC_INIT(-1), | |
3070 | }; | |
3071 | ||
3072 | int get_nohz_load_balancer(void) | |
3073 | { | |
3074 | return atomic_read(&nohz.load_balancer); | |
3075 | } | |
3076 | ||
3077 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3078 | /** | |
3079 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3080 | * @cpu: The cpu whose lowest level of sched domain is to | |
3081 | * be returned. | |
3082 | * @flag: The flag to check for the lowest sched_domain | |
3083 | * for the given cpu. | |
3084 | * | |
3085 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3086 | */ | |
3087 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3088 | { | |
3089 | struct sched_domain *sd; | |
3090 | ||
3091 | for_each_domain(cpu, sd) | |
3092 | if (sd && (sd->flags & flag)) | |
3093 | break; | |
3094 | ||
3095 | return sd; | |
3096 | } | |
3097 | ||
3098 | /** | |
3099 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3100 | * @cpu: The cpu whose domains we're iterating over. | |
3101 | * @sd: variable holding the value of the power_savings_sd | |
3102 | * for cpu. | |
3103 | * @flag: The flag to filter the sched_domains to be iterated. | |
3104 | * | |
3105 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3106 | * set, starting from the lowest sched_domain to the highest. | |
3107 | */ | |
3108 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3109 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3110 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3111 | ||
3112 | /** | |
3113 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3114 | * @ilb_group: group to be checked for semi-idleness | |
3115 | * | |
3116 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3117 | * | |
3118 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3119 | * and atleast one non-idle CPU. This helper function checks if the given | |
3120 | * sched_group is semi-idle or not. | |
3121 | */ | |
3122 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3123 | { | |
3124 | cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask, | |
3125 | sched_group_cpus(ilb_group)); | |
3126 | ||
3127 | /* | |
3128 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3129 | * and atleast one idle cpu. | |
3130 | */ | |
3131 | if (cpumask_empty(nohz.ilb_grp_nohz_mask)) | |
3132 | return 0; | |
3133 | ||
3134 | if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group))) | |
3135 | return 0; | |
3136 | ||
3137 | return 1; | |
3138 | } | |
3139 | /** | |
3140 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3141 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3142 | * | |
3143 | * Returns: Returns the id of the idle load balancer if it exists, | |
3144 | * Else, returns >= nr_cpu_ids. | |
3145 | * | |
3146 | * This algorithm picks the idle load balancer such that it belongs to a | |
3147 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3148 | * completely idle packages/cores just for the purpose of idle load balancing | |
3149 | * when there are other idle cpu's which are better suited for that job. | |
3150 | */ | |
3151 | static int find_new_ilb(int cpu) | |
3152 | { | |
3153 | struct sched_domain *sd; | |
3154 | struct sched_group *ilb_group; | |
3155 | ||
3156 | /* | |
3157 | * Have idle load balancer selection from semi-idle packages only | |
3158 | * when power-aware load balancing is enabled | |
3159 | */ | |
3160 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3161 | goto out_done; | |
3162 | ||
3163 | /* | |
3164 | * Optimize for the case when we have no idle CPUs or only one | |
3165 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3166 | */ | |
3167 | if (cpumask_weight(nohz.cpu_mask) < 2) | |
3168 | goto out_done; | |
3169 | ||
3170 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
3171 | ilb_group = sd->groups; | |
3172 | ||
3173 | do { | |
3174 | if (is_semi_idle_group(ilb_group)) | |
3175 | return cpumask_first(nohz.ilb_grp_nohz_mask); | |
3176 | ||
3177 | ilb_group = ilb_group->next; | |
3178 | ||
3179 | } while (ilb_group != sd->groups); | |
3180 | } | |
3181 | ||
3182 | out_done: | |
3183 | return cpumask_first(nohz.cpu_mask); | |
3184 | } | |
3185 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3186 | static inline int find_new_ilb(int call_cpu) | |
3187 | { | |
3188 | return cpumask_first(nohz.cpu_mask); | |
3189 | } | |
3190 | #endif | |
3191 | ||
3192 | /* | |
3193 | * This routine will try to nominate the ilb (idle load balancing) | |
3194 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
3195 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
3196 | * go into this tickless mode, then there will be no ilb owner (as there is | |
3197 | * no need for one) and all the cpus will sleep till the next wakeup event | |
3198 | * arrives... | |
3199 | * | |
3200 | * For the ilb owner, tick is not stopped. And this tick will be used | |
3201 | * for idle load balancing. ilb owner will still be part of | |
3202 | * nohz.cpu_mask.. | |
3203 | * | |
3204 | * While stopping the tick, this cpu will become the ilb owner if there | |
3205 | * is no other owner. And will be the owner till that cpu becomes busy | |
3206 | * or if all cpus in the system stop their ticks at which point | |
3207 | * there is no need for ilb owner. | |
3208 | * | |
3209 | * When the ilb owner becomes busy, it nominates another owner, during the | |
3210 | * next busy scheduler_tick() | |
3211 | */ | |
3212 | int select_nohz_load_balancer(int stop_tick) | |
3213 | { | |
3214 | int cpu = smp_processor_id(); | |
3215 | ||
3216 | if (stop_tick) { | |
3217 | cpu_rq(cpu)->in_nohz_recently = 1; | |
3218 | ||
3219 | if (!cpu_active(cpu)) { | |
3220 | if (atomic_read(&nohz.load_balancer) != cpu) | |
3221 | return 0; | |
3222 | ||
3223 | /* | |
3224 | * If we are going offline and still the leader, | |
3225 | * give up! | |
3226 | */ | |
3227 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3228 | BUG(); | |
3229 | ||
3230 | return 0; | |
3231 | } | |
3232 | ||
3233 | cpumask_set_cpu(cpu, nohz.cpu_mask); | |
3234 | ||
3235 | /* time for ilb owner also to sleep */ | |
3236 | if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) { | |
3237 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3238 | atomic_set(&nohz.load_balancer, -1); | |
3239 | return 0; | |
3240 | } | |
3241 | ||
3242 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3243 | /* make me the ilb owner */ | |
3244 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
3245 | return 1; | |
3246 | } else if (atomic_read(&nohz.load_balancer) == cpu) { | |
3247 | int new_ilb; | |
3248 | ||
3249 | if (!(sched_smt_power_savings || | |
3250 | sched_mc_power_savings)) | |
3251 | return 1; | |
3252 | /* | |
3253 | * Check to see if there is a more power-efficient | |
3254 | * ilb. | |
3255 | */ | |
3256 | new_ilb = find_new_ilb(cpu); | |
3257 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
3258 | atomic_set(&nohz.load_balancer, -1); | |
3259 | resched_cpu(new_ilb); | |
3260 | return 0; | |
3261 | } | |
3262 | return 1; | |
3263 | } | |
3264 | } else { | |
3265 | if (!cpumask_test_cpu(cpu, nohz.cpu_mask)) | |
3266 | return 0; | |
3267 | ||
3268 | cpumask_clear_cpu(cpu, nohz.cpu_mask); | |
3269 | ||
3270 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3271 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3272 | BUG(); | |
3273 | } | |
3274 | return 0; | |
3275 | } | |
3276 | #endif | |
3277 | ||
3278 | static DEFINE_SPINLOCK(balancing); | |
3279 | ||
3280 | /* | |
3281 | * It checks each scheduling domain to see if it is due to be balanced, | |
3282 | * and initiates a balancing operation if so. | |
3283 | * | |
3284 | * Balancing parameters are set up in arch_init_sched_domains. | |
3285 | */ | |
3286 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3287 | { | |
3288 | int balance = 1; | |
3289 | struct rq *rq = cpu_rq(cpu); | |
3290 | unsigned long interval; | |
3291 | struct sched_domain *sd; | |
3292 | /* Earliest time when we have to do rebalance again */ | |
3293 | unsigned long next_balance = jiffies + 60*HZ; | |
3294 | int update_next_balance = 0; | |
3295 | int need_serialize; | |
3296 | ||
3297 | for_each_domain(cpu, sd) { | |
3298 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3299 | continue; | |
3300 | ||
3301 | interval = sd->balance_interval; | |
3302 | if (idle != CPU_IDLE) | |
3303 | interval *= sd->busy_factor; | |
3304 | ||
3305 | /* scale ms to jiffies */ | |
3306 | interval = msecs_to_jiffies(interval); | |
3307 | if (unlikely(!interval)) | |
3308 | interval = 1; | |
3309 | if (interval > HZ*NR_CPUS/10) | |
3310 | interval = HZ*NR_CPUS/10; | |
3311 | ||
3312 | need_serialize = sd->flags & SD_SERIALIZE; | |
3313 | ||
3314 | if (need_serialize) { | |
3315 | if (!spin_trylock(&balancing)) | |
3316 | goto out; | |
3317 | } | |
3318 | ||
3319 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3320 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3321 | /* | |
3322 | * We've pulled tasks over so either we're no | |
3323 | * longer idle, or one of our SMT siblings is | |
3324 | * not idle. | |
3325 | */ | |
3326 | idle = CPU_NOT_IDLE; | |
3327 | } | |
3328 | sd->last_balance = jiffies; | |
3329 | } | |
3330 | if (need_serialize) | |
3331 | spin_unlock(&balancing); | |
3332 | out: | |
3333 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3334 | next_balance = sd->last_balance + interval; | |
3335 | update_next_balance = 1; | |
3336 | } | |
3337 | ||
3338 | /* | |
3339 | * Stop the load balance at this level. There is another | |
3340 | * CPU in our sched group which is doing load balancing more | |
3341 | * actively. | |
3342 | */ | |
3343 | if (!balance) | |
3344 | break; | |
3345 | } | |
3346 | ||
3347 | /* | |
3348 | * next_balance will be updated only when there is a need. | |
3349 | * When the cpu is attached to null domain for ex, it will not be | |
3350 | * updated. | |
3351 | */ | |
3352 | if (likely(update_next_balance)) | |
3353 | rq->next_balance = next_balance; | |
3354 | } | |
3355 | ||
3356 | /* | |
3357 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
3358 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
3359 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
3360 | */ | |
3361 | static void run_rebalance_domains(struct softirq_action *h) | |
3362 | { | |
3363 | int this_cpu = smp_processor_id(); | |
3364 | struct rq *this_rq = cpu_rq(this_cpu); | |
3365 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
3366 | CPU_IDLE : CPU_NOT_IDLE; | |
3367 | ||
3368 | rebalance_domains(this_cpu, idle); | |
3369 | ||
3370 | #ifdef CONFIG_NO_HZ | |
3371 | /* | |
3372 | * If this cpu is the owner for idle load balancing, then do the | |
3373 | * balancing on behalf of the other idle cpus whose ticks are | |
3374 | * stopped. | |
3375 | */ | |
3376 | if (this_rq->idle_at_tick && | |
3377 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
3378 | struct rq *rq; | |
3379 | int balance_cpu; | |
3380 | ||
3381 | for_each_cpu(balance_cpu, nohz.cpu_mask) { | |
3382 | if (balance_cpu == this_cpu) | |
3383 | continue; | |
3384 | ||
3385 | /* | |
3386 | * If this cpu gets work to do, stop the load balancing | |
3387 | * work being done for other cpus. Next load | |
3388 | * balancing owner will pick it up. | |
3389 | */ | |
3390 | if (need_resched()) | |
3391 | break; | |
3392 | ||
3393 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3394 | ||
3395 | rq = cpu_rq(balance_cpu); | |
3396 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3397 | this_rq->next_balance = rq->next_balance; | |
3398 | } | |
3399 | } | |
3400 | #endif | |
3401 | } | |
3402 | ||
3403 | static inline int on_null_domain(int cpu) | |
3404 | { | |
3405 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); | |
3406 | } | |
3407 | ||
3408 | /* | |
3409 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
3410 | * | |
3411 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
3412 | * idle load balancing owner or decide to stop the periodic load balancing, | |
3413 | * if the whole system is idle. | |
3414 | */ | |
3415 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
3416 | { | |
3417 | #ifdef CONFIG_NO_HZ | |
3418 | /* | |
3419 | * If we were in the nohz mode recently and busy at the current | |
3420 | * scheduler tick, then check if we need to nominate new idle | |
3421 | * load balancer. | |
3422 | */ | |
3423 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
3424 | rq->in_nohz_recently = 0; | |
3425 | ||
3426 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
3427 | cpumask_clear_cpu(cpu, nohz.cpu_mask); | |
3428 | atomic_set(&nohz.load_balancer, -1); | |
3429 | } | |
3430 | ||
3431 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3432 | int ilb = find_new_ilb(cpu); | |
3433 | ||
3434 | if (ilb < nr_cpu_ids) | |
3435 | resched_cpu(ilb); | |
3436 | } | |
3437 | } | |
3438 | ||
3439 | /* | |
3440 | * If this cpu is idle and doing idle load balancing for all the | |
3441 | * cpus with ticks stopped, is it time for that to stop? | |
3442 | */ | |
3443 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
3444 | cpumask_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3445 | resched_cpu(cpu); | |
3446 | return; | |
3447 | } | |
3448 | ||
3449 | /* | |
3450 | * If this cpu is idle and the idle load balancing is done by | |
3451 | * someone else, then no need raise the SCHED_SOFTIRQ | |
3452 | */ | |
3453 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
3454 | cpumask_test_cpu(cpu, nohz.cpu_mask)) | |
3455 | return; | |
3456 | #endif | |
3457 | /* Don't need to rebalance while attached to NULL domain */ | |
3458 | if (time_after_eq(jiffies, rq->next_balance) && | |
3459 | likely(!on_null_domain(cpu))) | |
3460 | raise_softirq(SCHED_SOFTIRQ); | |
3461 | } | |
3462 | ||
3463 | static void rq_online_fair(struct rq *rq) | |
3464 | { | |
3465 | update_sysctl(); | |
3466 | } | |
3467 | ||
3468 | static void rq_offline_fair(struct rq *rq) | |
3469 | { | |
3470 | update_sysctl(); | |
3471 | } | |
3472 | ||
3473 | #else /* CONFIG_SMP */ | |
3474 | ||
3475 | /* | |
3476 | * on UP we do not need to balance between CPUs: | |
3477 | */ | |
3478 | static inline void idle_balance(int cpu, struct rq *rq) | |
3479 | { | |
3480 | } | |
3481 | ||
3482 | #endif /* CONFIG_SMP */ | |
3483 | ||
3484 | /* | |
3485 | * scheduler tick hitting a task of our scheduling class: | |
3486 | */ | |
3487 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) | |
3488 | { | |
3489 | struct cfs_rq *cfs_rq; | |
3490 | struct sched_entity *se = &curr->se; | |
3491 | ||
3492 | for_each_sched_entity(se) { | |
3493 | cfs_rq = cfs_rq_of(se); | |
3494 | entity_tick(cfs_rq, se, queued); | |
3495 | } | |
3496 | } | |
3497 | ||
3498 | /* | |
3499 | * called on fork with the child task as argument from the parent's context | |
3500 | * - child not yet on the tasklist | |
3501 | * - preemption disabled | |
3502 | */ | |
3503 | static void task_fork_fair(struct task_struct *p) | |
3504 | { | |
3505 | struct cfs_rq *cfs_rq = task_cfs_rq(current); | |
3506 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; | |
3507 | int this_cpu = smp_processor_id(); | |
3508 | struct rq *rq = this_rq(); | |
3509 | unsigned long flags; | |
3510 | ||
3511 | raw_spin_lock_irqsave(&rq->lock, flags); | |
3512 | ||
3513 | if (unlikely(task_cpu(p) != this_cpu)) | |
3514 | __set_task_cpu(p, this_cpu); | |
3515 | ||
3516 | update_curr(cfs_rq); | |
3517 | ||
3518 | if (curr) | |
3519 | se->vruntime = curr->vruntime; | |
3520 | place_entity(cfs_rq, se, 1); | |
3521 | ||
3522 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { | |
3523 | /* | |
3524 | * Upon rescheduling, sched_class::put_prev_task() will place | |
3525 | * 'current' within the tree based on its new key value. | |
3526 | */ | |
3527 | swap(curr->vruntime, se->vruntime); | |
3528 | resched_task(rq->curr); | |
3529 | } | |
3530 | ||
3531 | se->vruntime -= cfs_rq->min_vruntime; | |
3532 | ||
3533 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
3534 | } | |
3535 | ||
3536 | /* | |
3537 | * Priority of the task has changed. Check to see if we preempt | |
3538 | * the current task. | |
3539 | */ | |
3540 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
3541 | int oldprio, int running) | |
3542 | { | |
3543 | /* | |
3544 | * Reschedule if we are currently running on this runqueue and | |
3545 | * our priority decreased, or if we are not currently running on | |
3546 | * this runqueue and our priority is higher than the current's | |
3547 | */ | |
3548 | if (running) { | |
3549 | if (p->prio > oldprio) | |
3550 | resched_task(rq->curr); | |
3551 | } else | |
3552 | check_preempt_curr(rq, p, 0); | |
3553 | } | |
3554 | ||
3555 | /* | |
3556 | * We switched to the sched_fair class. | |
3557 | */ | |
3558 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
3559 | int running) | |
3560 | { | |
3561 | /* | |
3562 | * We were most likely switched from sched_rt, so | |
3563 | * kick off the schedule if running, otherwise just see | |
3564 | * if we can still preempt the current task. | |
3565 | */ | |
3566 | if (running) | |
3567 | resched_task(rq->curr); | |
3568 | else | |
3569 | check_preempt_curr(rq, p, 0); | |
3570 | } | |
3571 | ||
3572 | /* Account for a task changing its policy or group. | |
3573 | * | |
3574 | * This routine is mostly called to set cfs_rq->curr field when a task | |
3575 | * migrates between groups/classes. | |
3576 | */ | |
3577 | static void set_curr_task_fair(struct rq *rq) | |
3578 | { | |
3579 | struct sched_entity *se = &rq->curr->se; | |
3580 | ||
3581 | for_each_sched_entity(se) | |
3582 | set_next_entity(cfs_rq_of(se), se); | |
3583 | } | |
3584 | ||
3585 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3586 | static void moved_group_fair(struct task_struct *p, int on_rq) | |
3587 | { | |
3588 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
3589 | ||
3590 | update_curr(cfs_rq); | |
3591 | if (!on_rq) | |
3592 | place_entity(cfs_rq, &p->se, 1); | |
3593 | } | |
3594 | #endif | |
3595 | ||
3596 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) | |
3597 | { | |
3598 | struct sched_entity *se = &task->se; | |
3599 | unsigned int rr_interval = 0; | |
3600 | ||
3601 | /* | |
3602 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
3603 | * idle runqueue: | |
3604 | */ | |
3605 | if (rq->cfs.load.weight) | |
3606 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
3607 | ||
3608 | return rr_interval; | |
3609 | } | |
3610 | ||
3611 | /* | |
3612 | * All the scheduling class methods: | |
3613 | */ | |
3614 | static const struct sched_class fair_sched_class = { | |
3615 | .next = &idle_sched_class, | |
3616 | .enqueue_task = enqueue_task_fair, | |
3617 | .dequeue_task = dequeue_task_fair, | |
3618 | .yield_task = yield_task_fair, | |
3619 | ||
3620 | .check_preempt_curr = check_preempt_wakeup, | |
3621 | ||
3622 | .pick_next_task = pick_next_task_fair, | |
3623 | .put_prev_task = put_prev_task_fair, | |
3624 | ||
3625 | #ifdef CONFIG_SMP | |
3626 | .select_task_rq = select_task_rq_fair, | |
3627 | ||
3628 | .rq_online = rq_online_fair, | |
3629 | .rq_offline = rq_offline_fair, | |
3630 | ||
3631 | .task_waking = task_waking_fair, | |
3632 | #endif | |
3633 | ||
3634 | .set_curr_task = set_curr_task_fair, | |
3635 | .task_tick = task_tick_fair, | |
3636 | .task_fork = task_fork_fair, | |
3637 | ||
3638 | .prio_changed = prio_changed_fair, | |
3639 | .switched_to = switched_to_fair, | |
3640 | ||
3641 | .get_rr_interval = get_rr_interval_fair, | |
3642 | ||
3643 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3644 | .moved_group = moved_group_fair, | |
3645 | #endif | |
3646 | }; | |
3647 | ||
3648 | #ifdef CONFIG_SCHED_DEBUG | |
3649 | static void print_cfs_stats(struct seq_file *m, int cpu) | |
3650 | { | |
3651 | struct cfs_rq *cfs_rq; | |
3652 | ||
3653 | rcu_read_lock(); | |
3654 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) | |
3655 | print_cfs_rq(m, cpu, cfs_rq); | |
3656 | rcu_read_unlock(); | |
3657 | } | |
3658 | #endif |