Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
9745512c | 25 | |
bf0f6f24 | 26 | /* |
21805085 | 27 | * Targeted preemption latency for CPU-bound tasks: |
172e082a | 28 | * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 29 | * |
21805085 | 30 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
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. | |
bf0f6f24 | 34 | * |
d274a4ce IM |
35 | * (to see the precise effective timeslice length of your workload, |
36 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 37 | */ |
172e082a | 38 | unsigned int sysctl_sched_latency = 5000000ULL; |
0bcdcf28 | 39 | unsigned int normalized_sysctl_sched_latency = 5000000ULL; |
2bd8e6d4 | 40 | |
1983a922 CE |
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 | ||
2bd8e6d4 | 53 | /* |
b2be5e96 | 54 | * Minimal preemption granularity for CPU-bound tasks: |
172e082a | 55 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 56 | */ |
172e082a | 57 | unsigned int sysctl_sched_min_granularity = 1000000ULL; |
0bcdcf28 | 58 | unsigned int normalized_sysctl_sched_min_granularity = 1000000ULL; |
21805085 PZ |
59 | |
60 | /* | |
b2be5e96 PZ |
61 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
62 | */ | |
722aab0c | 63 | static unsigned int sched_nr_latency = 5; |
b2be5e96 PZ |
64 | |
65 | /* | |
2bba22c5 | 66 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 67 | * parent will (try to) run first. |
21805085 | 68 | */ |
2bba22c5 | 69 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 70 | |
1799e35d IM |
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 | ||
bf0f6f24 IM |
79 | /* |
80 | * SCHED_OTHER wake-up granularity. | |
172e082a | 81 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
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 | */ | |
172e082a | 87 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 88 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 89 | |
da84d961 IM |
90 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
91 | ||
a4c2f00f PZ |
92 | static const struct sched_class fair_sched_class; |
93 | ||
bf0f6f24 IM |
94 | /************************************************************** |
95 | * CFS operations on generic schedulable entities: | |
96 | */ | |
97 | ||
62160e3f | 98 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 99 | |
62160e3f | 100 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
101 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
102 | { | |
62160e3f | 103 | return cfs_rq->rq; |
bf0f6f24 IM |
104 | } |
105 | ||
62160e3f IM |
106 | /* An entity is a task if it doesn't "own" a runqueue */ |
107 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 108 | |
8f48894f PZ |
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 | ||
b758149c PZ |
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 | ||
464b7527 PZ |
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 | ||
8f48894f PZ |
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 | } | |
bf0f6f24 | 214 | |
62160e3f IM |
215 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
216 | { | |
217 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
218 | } |
219 | ||
220 | #define entity_is_task(se) 1 | |
221 | ||
b758149c PZ |
222 | #define for_each_sched_entity(se) \ |
223 | for (; se; se = NULL) | |
bf0f6f24 | 224 | |
b758149c | 225 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 226 | { |
b758149c | 227 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
228 | } |
229 | ||
b758149c PZ |
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 | ||
464b7527 PZ |
263 | static inline void |
264 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
265 | { | |
266 | } | |
267 | ||
b758149c PZ |
268 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
269 | ||
bf0f6f24 IM |
270 | |
271 | /************************************************************** | |
272 | * Scheduling class tree data structure manipulation methods: | |
273 | */ | |
274 | ||
0702e3eb | 275 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 276 | { |
368059a9 PZ |
277 | s64 delta = (s64)(vruntime - min_vruntime); |
278 | if (delta > 0) | |
02e0431a PZ |
279 | min_vruntime = vruntime; |
280 | ||
281 | return min_vruntime; | |
282 | } | |
283 | ||
0702e3eb | 284 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
285 | { |
286 | s64 delta = (s64)(vruntime - min_vruntime); | |
287 | if (delta < 0) | |
288 | min_vruntime = vruntime; | |
289 | ||
290 | return min_vruntime; | |
291 | } | |
292 | ||
54fdc581 FC |
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 | ||
0702e3eb | 299 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 300 | { |
30cfdcfc | 301 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
302 | } |
303 | ||
1af5f730 PZ |
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 | ||
e17036da | 316 | if (!cfs_rq->curr) |
1af5f730 PZ |
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 | ||
bf0f6f24 IM |
325 | /* |
326 | * Enqueue an entity into the rb-tree: | |
327 | */ | |
0702e3eb | 328 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
329 | { |
330 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
331 | struct rb_node *parent = NULL; | |
332 | struct sched_entity *entry; | |
9014623c | 333 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
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 | */ | |
9014623c | 346 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
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 | */ | |
1af5f730 | 358 | if (leftmost) |
57cb499d | 359 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
360 | |
361 | rb_link_node(&se->run_node, parent, link); | |
362 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
363 | } |
364 | ||
0702e3eb | 365 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 366 | { |
3fe69747 PZ |
367 | if (cfs_rq->rb_leftmost == &se->run_node) { |
368 | struct rb_node *next_node; | |
3fe69747 PZ |
369 | |
370 | next_node = rb_next(&se->run_node); | |
371 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 372 | } |
e9acbff6 | 373 | |
bf0f6f24 | 374 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
375 | } |
376 | ||
bf0f6f24 IM |
377 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
378 | { | |
f4b6755f PZ |
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); | |
bf0f6f24 IM |
385 | } |
386 | ||
f4b6755f | 387 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 388 | { |
7eee3e67 | 389 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 390 | |
70eee74b BS |
391 | if (!last) |
392 | return NULL; | |
7eee3e67 IM |
393 | |
394 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
395 | } |
396 | ||
bf0f6f24 IM |
397 | /************************************************************** |
398 | * Scheduling class statistics methods: | |
399 | */ | |
400 | ||
b2be5e96 | 401 | #ifdef CONFIG_SCHED_DEBUG |
acb4a848 | 402 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 403 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
404 | loff_t *ppos) |
405 | { | |
8d65af78 | 406 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 407 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
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 | ||
acb4a848 CE |
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 | ||
b2be5e96 PZ |
423 | return 0; |
424 | } | |
425 | #endif | |
647e7cac | 426 | |
a7be37ac | 427 | /* |
f9c0b095 | 428 | * delta /= w |
a7be37ac PZ |
429 | */ |
430 | static inline unsigned long | |
431 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
432 | { | |
f9c0b095 PZ |
433 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
434 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
435 | |
436 | return delta; | |
437 | } | |
438 | ||
647e7cac IM |
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 | */ | |
4d78e7b6 PZ |
447 | static u64 __sched_period(unsigned long nr_running) |
448 | { | |
449 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 450 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
451 | |
452 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 453 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 454 | period *= nr_running; |
4d78e7b6 PZ |
455 | } |
456 | ||
457 | return period; | |
458 | } | |
459 | ||
647e7cac IM |
460 | /* |
461 | * We calculate the wall-time slice from the period by taking a part | |
462 | * proportional to the weight. | |
463 | * | |
f9c0b095 | 464 | * s = p*P[w/rw] |
647e7cac | 465 | */ |
6d0f0ebd | 466 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 467 | { |
0a582440 | 468 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 469 | |
0a582440 | 470 | for_each_sched_entity(se) { |
6272d68c | 471 | struct load_weight *load; |
3104bf03 | 472 | struct load_weight lw; |
6272d68c LM |
473 | |
474 | cfs_rq = cfs_rq_of(se); | |
475 | load = &cfs_rq->load; | |
f9c0b095 | 476 | |
0a582440 | 477 | if (unlikely(!se->on_rq)) { |
3104bf03 | 478 | lw = cfs_rq->load; |
0a582440 MG |
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; | |
bf0f6f24 IM |
486 | } |
487 | ||
647e7cac | 488 | /* |
ac884dec | 489 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 490 | * |
f9c0b095 | 491 | * vs = s/w |
647e7cac | 492 | */ |
f9c0b095 | 493 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 494 | { |
f9c0b095 | 495 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
496 | } |
497 | ||
bf0f6f24 IM |
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 | |
8ebc91d9 IM |
503 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
504 | unsigned long delta_exec) | |
bf0f6f24 | 505 | { |
bbdba7c0 | 506 | unsigned long delta_exec_weighted; |
bf0f6f24 | 507 | |
8179ca23 | 508 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
509 | |
510 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 511 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 512 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 513 | |
e9acbff6 | 514 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 515 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
516 | } |
517 | ||
b7cc0896 | 518 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 519 | { |
429d43bc | 520 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 521 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
522 | unsigned long delta_exec; |
523 | ||
524 | if (unlikely(!curr)) | |
525 | return; | |
526 | ||
527 | /* | |
528 | * Get the amount of time the current task was running | |
529 | * since the last time we changed load (this cannot | |
530 | * overflow on 32 bits): | |
531 | */ | |
8ebc91d9 | 532 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
533 | if (!delta_exec) |
534 | return; | |
bf0f6f24 | 535 | |
8ebc91d9 IM |
536 | __update_curr(cfs_rq, curr, delta_exec); |
537 | curr->exec_start = now; | |
d842de87 SV |
538 | |
539 | if (entity_is_task(curr)) { | |
540 | struct task_struct *curtask = task_of(curr); | |
541 | ||
f977bb49 | 542 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 543 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 544 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 545 | } |
bf0f6f24 IM |
546 | } |
547 | ||
548 | static inline void | |
5870db5b | 549 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 550 | { |
d281918d | 551 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
552 | } |
553 | ||
bf0f6f24 IM |
554 | /* |
555 | * Task is being enqueued - update stats: | |
556 | */ | |
d2417e5a | 557 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 558 | { |
bf0f6f24 IM |
559 | /* |
560 | * Are we enqueueing a waiting task? (for current tasks | |
561 | * a dequeue/enqueue event is a NOP) | |
562 | */ | |
429d43bc | 563 | if (se != cfs_rq->curr) |
5870db5b | 564 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
565 | } |
566 | ||
bf0f6f24 | 567 | static void |
9ef0a961 | 568 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 569 | { |
bbdba7c0 IM |
570 | schedstat_set(se->wait_max, max(se->wait_max, |
571 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
572 | schedstat_set(se->wait_count, se->wait_count + 1); |
573 | schedstat_set(se->wait_sum, se->wait_sum + | |
574 | rq_of(cfs_rq)->clock - se->wait_start); | |
768d0c27 PZ |
575 | #ifdef CONFIG_SCHEDSTATS |
576 | if (entity_is_task(se)) { | |
577 | trace_sched_stat_wait(task_of(se), | |
578 | rq_of(cfs_rq)->clock - se->wait_start); | |
579 | } | |
580 | #endif | |
e1f84508 | 581 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
582 | } |
583 | ||
584 | static inline void | |
19b6a2e3 | 585 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 586 | { |
bf0f6f24 IM |
587 | /* |
588 | * Mark the end of the wait period if dequeueing a | |
589 | * waiting task: | |
590 | */ | |
429d43bc | 591 | if (se != cfs_rq->curr) |
9ef0a961 | 592 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
593 | } |
594 | ||
595 | /* | |
596 | * We are picking a new current task - update its stats: | |
597 | */ | |
598 | static inline void | |
79303e9e | 599 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
600 | { |
601 | /* | |
602 | * We are starting a new run period: | |
603 | */ | |
d281918d | 604 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
605 | } |
606 | ||
bf0f6f24 IM |
607 | /************************************************** |
608 | * Scheduling class queueing methods: | |
609 | */ | |
610 | ||
c09595f6 PZ |
611 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
612 | static void | |
613 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
614 | { | |
615 | cfs_rq->task_weight += weight; | |
616 | } | |
617 | #else | |
618 | static inline void | |
619 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
620 | { | |
621 | } | |
622 | #endif | |
623 | ||
30cfdcfc DA |
624 | static void |
625 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
626 | { | |
627 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
628 | if (!parent_entity(se)) |
629 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 630 | if (entity_is_task(se)) { |
c09595f6 | 631 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
632 | list_add(&se->group_node, &cfs_rq->tasks); |
633 | } | |
30cfdcfc DA |
634 | cfs_rq->nr_running++; |
635 | se->on_rq = 1; | |
636 | } | |
637 | ||
638 | static void | |
639 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
640 | { | |
641 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
642 | if (!parent_entity(se)) |
643 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 644 | if (entity_is_task(se)) { |
c09595f6 | 645 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
646 | list_del_init(&se->group_node); |
647 | } | |
30cfdcfc DA |
648 | cfs_rq->nr_running--; |
649 | se->on_rq = 0; | |
650 | } | |
651 | ||
2396af69 | 652 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 653 | { |
bf0f6f24 | 654 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
655 | struct task_struct *tsk = NULL; |
656 | ||
657 | if (entity_is_task(se)) | |
658 | tsk = task_of(se); | |
659 | ||
bf0f6f24 | 660 | if (se->sleep_start) { |
d281918d | 661 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
662 | |
663 | if ((s64)delta < 0) | |
664 | delta = 0; | |
665 | ||
666 | if (unlikely(delta > se->sleep_max)) | |
667 | se->sleep_max = delta; | |
668 | ||
669 | se->sleep_start = 0; | |
670 | se->sum_sleep_runtime += delta; | |
9745512c | 671 | |
768d0c27 | 672 | if (tsk) { |
e414314c | 673 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
674 | trace_sched_stat_sleep(tsk, delta); |
675 | } | |
bf0f6f24 IM |
676 | } |
677 | if (se->block_start) { | |
d281918d | 678 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
679 | |
680 | if ((s64)delta < 0) | |
681 | delta = 0; | |
682 | ||
683 | if (unlikely(delta > se->block_max)) | |
684 | se->block_max = delta; | |
685 | ||
686 | se->block_start = 0; | |
687 | se->sum_sleep_runtime += delta; | |
30084fbd | 688 | |
e414314c | 689 | if (tsk) { |
8f0dfc34 AV |
690 | if (tsk->in_iowait) { |
691 | se->iowait_sum += delta; | |
692 | se->iowait_count++; | |
768d0c27 | 693 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
694 | } |
695 | ||
e414314c PZ |
696 | /* |
697 | * Blocking time is in units of nanosecs, so shift by | |
698 | * 20 to get a milliseconds-range estimation of the | |
699 | * amount of time that the task spent sleeping: | |
700 | */ | |
701 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
702 | profile_hits(SLEEP_PROFILING, | |
703 | (void *)get_wchan(tsk), | |
704 | delta >> 20); | |
705 | } | |
706 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 707 | } |
bf0f6f24 IM |
708 | } |
709 | #endif | |
710 | } | |
711 | ||
ddc97297 PZ |
712 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
713 | { | |
714 | #ifdef CONFIG_SCHED_DEBUG | |
715 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
716 | ||
717 | if (d < 0) | |
718 | d = -d; | |
719 | ||
720 | if (d > 3*sysctl_sched_latency) | |
721 | schedstat_inc(cfs_rq, nr_spread_over); | |
722 | #endif | |
723 | } | |
724 | ||
aeb73b04 PZ |
725 | static void |
726 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
727 | { | |
1af5f730 | 728 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 729 | |
2cb8600e PZ |
730 | /* |
731 | * The 'current' period is already promised to the current tasks, | |
732 | * however the extra weight of the new task will slow them down a | |
733 | * little, place the new task so that it fits in the slot that | |
734 | * stays open at the end. | |
735 | */ | |
94dfb5e7 | 736 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 737 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 738 | |
a2e7a7eb MG |
739 | /* sleeps up to a single latency don't count. */ |
740 | if (!initial && sched_feat(FAIR_SLEEPERS)) { | |
741 | unsigned long thresh = sysctl_sched_latency; | |
a7be37ac | 742 | |
a2e7a7eb MG |
743 | /* |
744 | * Convert the sleeper threshold into virtual time. | |
745 | * SCHED_IDLE is a special sub-class. We care about | |
746 | * fairness only relative to other SCHED_IDLE tasks, | |
747 | * all of which have the same weight. | |
748 | */ | |
749 | if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) || | |
750 | task_of(se)->policy != SCHED_IDLE)) | |
751 | thresh = calc_delta_fair(thresh, se); | |
a7be37ac | 752 | |
a2e7a7eb MG |
753 | /* |
754 | * Halve their sleep time's effect, to allow | |
755 | * for a gentler effect of sleepers: | |
756 | */ | |
757 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
758 | thresh >>= 1; | |
51e0304c | 759 | |
a2e7a7eb | 760 | vruntime -= thresh; |
aeb73b04 PZ |
761 | } |
762 | ||
b5d9d734 MG |
763 | /* ensure we never gain time by being placed backwards. */ |
764 | vruntime = max_vruntime(se->vruntime, vruntime); | |
765 | ||
67e9fb2a | 766 | se->vruntime = vruntime; |
aeb73b04 PZ |
767 | } |
768 | ||
88ec22d3 PZ |
769 | #define ENQUEUE_WAKEUP 1 |
770 | #define ENQUEUE_MIGRATE 2 | |
771 | ||
bf0f6f24 | 772 | static void |
88ec22d3 | 773 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 774 | { |
88ec22d3 PZ |
775 | /* |
776 | * Update the normalized vruntime before updating min_vruntime | |
777 | * through callig update_curr(). | |
778 | */ | |
779 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATE)) | |
780 | se->vruntime += cfs_rq->min_vruntime; | |
781 | ||
bf0f6f24 | 782 | /* |
a2a2d680 | 783 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 784 | */ |
b7cc0896 | 785 | update_curr(cfs_rq); |
a992241d | 786 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 787 | |
88ec22d3 | 788 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 789 | place_entity(cfs_rq, se, 0); |
2396af69 | 790 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 791 | } |
bf0f6f24 | 792 | |
d2417e5a | 793 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 794 | check_spread(cfs_rq, se); |
83b699ed SV |
795 | if (se != cfs_rq->curr) |
796 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
797 | } |
798 | ||
a571bbea | 799 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 | 800 | { |
de69a80b | 801 | if (!se || cfs_rq->last == se) |
2002c695 PZ |
802 | cfs_rq->last = NULL; |
803 | ||
de69a80b | 804 | if (!se || cfs_rq->next == se) |
2002c695 PZ |
805 | cfs_rq->next = NULL; |
806 | } | |
807 | ||
a571bbea PZ |
808 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
809 | { | |
810 | for_each_sched_entity(se) | |
811 | __clear_buddies(cfs_rq_of(se), se); | |
812 | } | |
813 | ||
bf0f6f24 | 814 | static void |
525c2716 | 815 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 816 | { |
a2a2d680 DA |
817 | /* |
818 | * Update run-time statistics of the 'current'. | |
819 | */ | |
820 | update_curr(cfs_rq); | |
821 | ||
19b6a2e3 | 822 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 823 | if (sleep) { |
67e9fb2a | 824 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
825 | if (entity_is_task(se)) { |
826 | struct task_struct *tsk = task_of(se); | |
827 | ||
828 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 829 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 830 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 831 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 832 | } |
db36cc7d | 833 | #endif |
67e9fb2a PZ |
834 | } |
835 | ||
2002c695 | 836 | clear_buddies(cfs_rq, se); |
4793241b | 837 | |
83b699ed | 838 | if (se != cfs_rq->curr) |
30cfdcfc DA |
839 | __dequeue_entity(cfs_rq, se); |
840 | account_entity_dequeue(cfs_rq, se); | |
1af5f730 | 841 | update_min_vruntime(cfs_rq); |
88ec22d3 PZ |
842 | |
843 | /* | |
844 | * Normalize the entity after updating the min_vruntime because the | |
845 | * update can refer to the ->curr item and we need to reflect this | |
846 | * movement in our normalized position. | |
847 | */ | |
848 | if (!sleep) | |
849 | se->vruntime -= cfs_rq->min_vruntime; | |
bf0f6f24 IM |
850 | } |
851 | ||
852 | /* | |
853 | * Preempt the current task with a newly woken task if needed: | |
854 | */ | |
7c92e54f | 855 | static void |
2e09bf55 | 856 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 857 | { |
11697830 PZ |
858 | unsigned long ideal_runtime, delta_exec; |
859 | ||
6d0f0ebd | 860 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 861 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 862 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 863 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
864 | /* |
865 | * The current task ran long enough, ensure it doesn't get | |
866 | * re-elected due to buddy favours. | |
867 | */ | |
868 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
869 | return; |
870 | } | |
871 | ||
872 | /* | |
873 | * Ensure that a task that missed wakeup preemption by a | |
874 | * narrow margin doesn't have to wait for a full slice. | |
875 | * This also mitigates buddy induced latencies under load. | |
876 | */ | |
877 | if (!sched_feat(WAKEUP_PREEMPT)) | |
878 | return; | |
879 | ||
880 | if (delta_exec < sysctl_sched_min_granularity) | |
881 | return; | |
882 | ||
883 | if (cfs_rq->nr_running > 1) { | |
884 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
885 | s64 delta = curr->vruntime - se->vruntime; | |
886 | ||
887 | if (delta > ideal_runtime) | |
888 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 889 | } |
bf0f6f24 IM |
890 | } |
891 | ||
83b699ed | 892 | static void |
8494f412 | 893 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 894 | { |
83b699ed SV |
895 | /* 'current' is not kept within the tree. */ |
896 | if (se->on_rq) { | |
897 | /* | |
898 | * Any task has to be enqueued before it get to execute on | |
899 | * a CPU. So account for the time it spent waiting on the | |
900 | * runqueue. | |
901 | */ | |
902 | update_stats_wait_end(cfs_rq, se); | |
903 | __dequeue_entity(cfs_rq, se); | |
904 | } | |
905 | ||
79303e9e | 906 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 907 | cfs_rq->curr = se; |
eba1ed4b IM |
908 | #ifdef CONFIG_SCHEDSTATS |
909 | /* | |
910 | * Track our maximum slice length, if the CPU's load is at | |
911 | * least twice that of our own weight (i.e. dont track it | |
912 | * when there are only lesser-weight tasks around): | |
913 | */ | |
495eca49 | 914 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
915 | se->slice_max = max(se->slice_max, |
916 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
917 | } | |
918 | #endif | |
4a55b450 | 919 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
920 | } |
921 | ||
3f3a4904 PZ |
922 | static int |
923 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
924 | ||
f4b6755f | 925 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 926 | { |
f4b6755f | 927 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
f685ceac | 928 | struct sched_entity *left = se; |
f4b6755f | 929 | |
f685ceac MG |
930 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) |
931 | se = cfs_rq->next; | |
aa2ac252 | 932 | |
f685ceac MG |
933 | /* |
934 | * Prefer last buddy, try to return the CPU to a preempted task. | |
935 | */ | |
936 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
937 | se = cfs_rq->last; | |
938 | ||
939 | clear_buddies(cfs_rq, se); | |
4793241b PZ |
940 | |
941 | return se; | |
aa2ac252 PZ |
942 | } |
943 | ||
ab6cde26 | 944 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
945 | { |
946 | /* | |
947 | * If still on the runqueue then deactivate_task() | |
948 | * was not called and update_curr() has to be done: | |
949 | */ | |
950 | if (prev->on_rq) | |
b7cc0896 | 951 | update_curr(cfs_rq); |
bf0f6f24 | 952 | |
ddc97297 | 953 | check_spread(cfs_rq, prev); |
30cfdcfc | 954 | if (prev->on_rq) { |
5870db5b | 955 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
956 | /* Put 'current' back into the tree. */ |
957 | __enqueue_entity(cfs_rq, prev); | |
958 | } | |
429d43bc | 959 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
960 | } |
961 | ||
8f4d37ec PZ |
962 | static void |
963 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 964 | { |
bf0f6f24 | 965 | /* |
30cfdcfc | 966 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 967 | */ |
30cfdcfc | 968 | update_curr(cfs_rq); |
bf0f6f24 | 969 | |
8f4d37ec PZ |
970 | #ifdef CONFIG_SCHED_HRTICK |
971 | /* | |
972 | * queued ticks are scheduled to match the slice, so don't bother | |
973 | * validating it and just reschedule. | |
974 | */ | |
983ed7a6 HH |
975 | if (queued) { |
976 | resched_task(rq_of(cfs_rq)->curr); | |
977 | return; | |
978 | } | |
8f4d37ec PZ |
979 | /* |
980 | * don't let the period tick interfere with the hrtick preemption | |
981 | */ | |
982 | if (!sched_feat(DOUBLE_TICK) && | |
983 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
984 | return; | |
985 | #endif | |
986 | ||
ce6c1311 | 987 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 988 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
989 | } |
990 | ||
991 | /************************************************** | |
992 | * CFS operations on tasks: | |
993 | */ | |
994 | ||
8f4d37ec PZ |
995 | #ifdef CONFIG_SCHED_HRTICK |
996 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
997 | { | |
8f4d37ec PZ |
998 | struct sched_entity *se = &p->se; |
999 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1000 | ||
1001 | WARN_ON(task_rq(p) != rq); | |
1002 | ||
1003 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1004 | u64 slice = sched_slice(cfs_rq, se); | |
1005 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1006 | s64 delta = slice - ran; | |
1007 | ||
1008 | if (delta < 0) { | |
1009 | if (rq->curr == p) | |
1010 | resched_task(p); | |
1011 | return; | |
1012 | } | |
1013 | ||
1014 | /* | |
1015 | * Don't schedule slices shorter than 10000ns, that just | |
1016 | * doesn't make sense. Rely on vruntime for fairness. | |
1017 | */ | |
31656519 | 1018 | if (rq->curr != p) |
157124c1 | 1019 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1020 | |
31656519 | 1021 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1022 | } |
1023 | } | |
a4c2f00f PZ |
1024 | |
1025 | /* | |
1026 | * called from enqueue/dequeue and updates the hrtick when the | |
1027 | * current task is from our class and nr_running is low enough | |
1028 | * to matter. | |
1029 | */ | |
1030 | static void hrtick_update(struct rq *rq) | |
1031 | { | |
1032 | struct task_struct *curr = rq->curr; | |
1033 | ||
1034 | if (curr->sched_class != &fair_sched_class) | |
1035 | return; | |
1036 | ||
1037 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1038 | hrtick_start_fair(rq, curr); | |
1039 | } | |
55e12e5e | 1040 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1041 | static inline void |
1042 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1043 | { | |
1044 | } | |
a4c2f00f PZ |
1045 | |
1046 | static inline void hrtick_update(struct rq *rq) | |
1047 | { | |
1048 | } | |
8f4d37ec PZ |
1049 | #endif |
1050 | ||
bf0f6f24 IM |
1051 | /* |
1052 | * The enqueue_task method is called before nr_running is | |
1053 | * increased. Here we update the fair scheduling stats and | |
1054 | * then put the task into the rbtree: | |
1055 | */ | |
fd390f6a | 1056 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
1057 | { |
1058 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1059 | struct sched_entity *se = &p->se; |
88ec22d3 PZ |
1060 | int flags = 0; |
1061 | ||
1062 | if (wakeup) | |
1063 | flags |= ENQUEUE_WAKEUP; | |
1064 | if (p->state == TASK_WAKING) | |
1065 | flags |= ENQUEUE_MIGRATE; | |
bf0f6f24 IM |
1066 | |
1067 | for_each_sched_entity(se) { | |
62fb1851 | 1068 | if (se->on_rq) |
bf0f6f24 IM |
1069 | break; |
1070 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 PZ |
1071 | enqueue_entity(cfs_rq, se, flags); |
1072 | flags = ENQUEUE_WAKEUP; | |
bf0f6f24 | 1073 | } |
8f4d37ec | 1074 | |
a4c2f00f | 1075 | hrtick_update(rq); |
bf0f6f24 IM |
1076 | } |
1077 | ||
1078 | /* | |
1079 | * The dequeue_task method is called before nr_running is | |
1080 | * decreased. We remove the task from the rbtree and | |
1081 | * update the fair scheduling stats: | |
1082 | */ | |
f02231e5 | 1083 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
1084 | { |
1085 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1086 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1087 | |
1088 | for_each_sched_entity(se) { | |
1089 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 1090 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 1091 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 1092 | if (cfs_rq->load.weight) |
bf0f6f24 | 1093 | break; |
b9fa3df3 | 1094 | sleep = 1; |
bf0f6f24 | 1095 | } |
8f4d37ec | 1096 | |
a4c2f00f | 1097 | hrtick_update(rq); |
bf0f6f24 IM |
1098 | } |
1099 | ||
1100 | /* | |
1799e35d IM |
1101 | * sched_yield() support is very simple - we dequeue and enqueue. |
1102 | * | |
1103 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 1104 | */ |
4530d7ab | 1105 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 1106 | { |
db292ca3 IM |
1107 | struct task_struct *curr = rq->curr; |
1108 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1109 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
1110 | |
1111 | /* | |
1799e35d IM |
1112 | * Are we the only task in the tree? |
1113 | */ | |
1114 | if (unlikely(cfs_rq->nr_running == 1)) | |
1115 | return; | |
1116 | ||
2002c695 PZ |
1117 | clear_buddies(cfs_rq, se); |
1118 | ||
db292ca3 | 1119 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1120 | update_rq_clock(rq); |
1799e35d | 1121 | /* |
a2a2d680 | 1122 | * Update run-time statistics of the 'current'. |
1799e35d | 1123 | */ |
2b1e315d | 1124 | update_curr(cfs_rq); |
1799e35d IM |
1125 | |
1126 | return; | |
1127 | } | |
1128 | /* | |
1129 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1130 | */ |
2b1e315d | 1131 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1132 | /* |
1133 | * Already in the rightmost position? | |
1134 | */ | |
54fdc581 | 1135 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
1799e35d IM |
1136 | return; |
1137 | ||
1138 | /* | |
1139 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1140 | * Upon rescheduling, sched_class::put_prev_task() will place |
1141 | * 'current' within the tree based on its new key value. | |
1799e35d | 1142 | */ |
30cfdcfc | 1143 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1144 | } |
1145 | ||
e7693a36 | 1146 | #ifdef CONFIG_SMP |
098fb9db | 1147 | |
88ec22d3 PZ |
1148 | static void task_waking_fair(struct rq *rq, struct task_struct *p) |
1149 | { | |
1150 | struct sched_entity *se = &p->se; | |
1151 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1152 | ||
1153 | se->vruntime -= cfs_rq->min_vruntime; | |
1154 | } | |
1155 | ||
bb3469ac | 1156 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1157 | /* |
1158 | * effective_load() calculates the load change as seen from the root_task_group | |
1159 | * | |
1160 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1161 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1162 | * can calculate the shift in shares. | |
1163 | * | |
1164 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1165 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1166 | * this change. | |
1167 | * | |
1168 | * We compensate this by not only taking the current delta into account, but | |
1169 | * also considering the delta between when the shares were last adjusted and | |
1170 | * now. | |
1171 | * | |
1172 | * We still saw a performance dip, some tracing learned us that between | |
1173 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1174 | * significantly. Therefore try to bias the error in direction of failing | |
1175 | * the affine wakeup. | |
1176 | * | |
1177 | */ | |
f1d239f7 PZ |
1178 | static long effective_load(struct task_group *tg, int cpu, |
1179 | long wl, long wg) | |
bb3469ac | 1180 | { |
4be9daaa | 1181 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1182 | |
1183 | if (!tg->parent) | |
1184 | return wl; | |
1185 | ||
f5bfb7d9 PZ |
1186 | /* |
1187 | * By not taking the decrease of shares on the other cpu into | |
1188 | * account our error leans towards reducing the affine wakeups. | |
1189 | */ | |
1190 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1191 | return wl; | |
1192 | ||
4be9daaa | 1193 | for_each_sched_entity(se) { |
cb5ef42a | 1194 | long S, rw, s, a, b; |
940959e9 PZ |
1195 | long more_w; |
1196 | ||
1197 | /* | |
1198 | * Instead of using this increment, also add the difference | |
1199 | * between when the shares were last updated and now. | |
1200 | */ | |
1201 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1202 | wl += more_w; | |
1203 | wg += more_w; | |
4be9daaa PZ |
1204 | |
1205 | S = se->my_q->tg->shares; | |
1206 | s = se->my_q->shares; | |
f1d239f7 | 1207 | rw = se->my_q->rq_weight; |
bb3469ac | 1208 | |
cb5ef42a PZ |
1209 | a = S*(rw + wl); |
1210 | b = S*rw + s*wg; | |
4be9daaa | 1211 | |
940959e9 PZ |
1212 | wl = s*(a-b); |
1213 | ||
1214 | if (likely(b)) | |
1215 | wl /= b; | |
1216 | ||
83378269 PZ |
1217 | /* |
1218 | * Assume the group is already running and will | |
1219 | * thus already be accounted for in the weight. | |
1220 | * | |
1221 | * That is, moving shares between CPUs, does not | |
1222 | * alter the group weight. | |
1223 | */ | |
4be9daaa | 1224 | wg = 0; |
4be9daaa | 1225 | } |
bb3469ac | 1226 | |
4be9daaa | 1227 | return wl; |
bb3469ac | 1228 | } |
4be9daaa | 1229 | |
bb3469ac | 1230 | #else |
4be9daaa | 1231 | |
83378269 PZ |
1232 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1233 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1234 | { |
83378269 | 1235 | return wl; |
bb3469ac | 1236 | } |
4be9daaa | 1237 | |
bb3469ac PZ |
1238 | #endif |
1239 | ||
c88d5910 | 1240 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1241 | { |
c88d5910 PZ |
1242 | struct task_struct *curr = current; |
1243 | unsigned long this_load, load; | |
1244 | int idx, this_cpu, prev_cpu; | |
098fb9db | 1245 | unsigned long tl_per_task; |
c88d5910 PZ |
1246 | unsigned int imbalance; |
1247 | struct task_group *tg; | |
83378269 | 1248 | unsigned long weight; |
b3137bc8 | 1249 | int balanced; |
098fb9db | 1250 | |
c88d5910 PZ |
1251 | idx = sd->wake_idx; |
1252 | this_cpu = smp_processor_id(); | |
1253 | prev_cpu = task_cpu(p); | |
1254 | load = source_load(prev_cpu, idx); | |
1255 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1256 | |
e69b0f1b PZ |
1257 | if (sync) { |
1258 | if (sched_feat(SYNC_LESS) && | |
1259 | (curr->se.avg_overlap > sysctl_sched_migration_cost || | |
1260 | p->se.avg_overlap > sysctl_sched_migration_cost)) | |
1261 | sync = 0; | |
1262 | } else { | |
1263 | if (sched_feat(SYNC_MORE) && | |
1264 | (curr->se.avg_overlap < sysctl_sched_migration_cost && | |
1265 | p->se.avg_overlap < sysctl_sched_migration_cost)) | |
1266 | sync = 1; | |
1267 | } | |
fc631c82 | 1268 | |
b3137bc8 MG |
1269 | /* |
1270 | * If sync wakeup then subtract the (maximum possible) | |
1271 | * effect of the currently running task from the load | |
1272 | * of the current CPU: | |
1273 | */ | |
83378269 PZ |
1274 | if (sync) { |
1275 | tg = task_group(current); | |
1276 | weight = current->se.load.weight; | |
1277 | ||
c88d5910 | 1278 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1279 | load += effective_load(tg, prev_cpu, 0, -weight); |
1280 | } | |
b3137bc8 | 1281 | |
83378269 PZ |
1282 | tg = task_group(p); |
1283 | weight = p->se.load.weight; | |
b3137bc8 | 1284 | |
c88d5910 PZ |
1285 | imbalance = 100 + (sd->imbalance_pct - 100) / 2; |
1286 | ||
71a29aa7 PZ |
1287 | /* |
1288 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1289 | * due to the sync cause above having dropped this_load to 0, we'll |
1290 | * always have an imbalance, but there's really nothing you can do | |
1291 | * about that, so that's good too. | |
71a29aa7 PZ |
1292 | * |
1293 | * Otherwise check if either cpus are near enough in load to allow this | |
1294 | * task to be woken on this_cpu. | |
1295 | */ | |
c88d5910 PZ |
1296 | balanced = !this_load || |
1297 | 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <= | |
83378269 | 1298 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); |
b3137bc8 | 1299 | |
098fb9db | 1300 | /* |
4ae7d5ce IM |
1301 | * If the currently running task will sleep within |
1302 | * a reasonable amount of time then attract this newly | |
1303 | * woken task: | |
098fb9db | 1304 | */ |
2fb7635c PZ |
1305 | if (sync && balanced) |
1306 | return 1; | |
098fb9db IM |
1307 | |
1308 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1309 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1310 | ||
c88d5910 PZ |
1311 | if (balanced || |
1312 | (this_load <= load && | |
1313 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1314 | /* |
1315 | * This domain has SD_WAKE_AFFINE and | |
1316 | * p is cache cold in this domain, and | |
1317 | * there is no bad imbalance. | |
1318 | */ | |
c88d5910 | 1319 | schedstat_inc(sd, ttwu_move_affine); |
098fb9db IM |
1320 | schedstat_inc(p, se.nr_wakeups_affine); |
1321 | ||
1322 | return 1; | |
1323 | } | |
1324 | return 0; | |
1325 | } | |
1326 | ||
aaee1203 PZ |
1327 | /* |
1328 | * find_idlest_group finds and returns the least busy CPU group within the | |
1329 | * domain. | |
1330 | */ | |
1331 | static struct sched_group * | |
78e7ed53 | 1332 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1333 | int this_cpu, int load_idx) |
e7693a36 | 1334 | { |
aaee1203 PZ |
1335 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
1336 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
aaee1203 | 1337 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1338 | |
aaee1203 PZ |
1339 | do { |
1340 | unsigned long load, avg_load; | |
1341 | int local_group; | |
1342 | int i; | |
e7693a36 | 1343 | |
aaee1203 PZ |
1344 | /* Skip over this group if it has no CPUs allowed */ |
1345 | if (!cpumask_intersects(sched_group_cpus(group), | |
1346 | &p->cpus_allowed)) | |
1347 | continue; | |
1348 | ||
1349 | local_group = cpumask_test_cpu(this_cpu, | |
1350 | sched_group_cpus(group)); | |
1351 | ||
1352 | /* Tally up the load of all CPUs in the group */ | |
1353 | avg_load = 0; | |
1354 | ||
1355 | for_each_cpu(i, sched_group_cpus(group)) { | |
1356 | /* Bias balancing toward cpus of our domain */ | |
1357 | if (local_group) | |
1358 | load = source_load(i, load_idx); | |
1359 | else | |
1360 | load = target_load(i, load_idx); | |
1361 | ||
1362 | avg_load += load; | |
1363 | } | |
1364 | ||
1365 | /* Adjust by relative CPU power of the group */ | |
1366 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1367 | ||
1368 | if (local_group) { | |
1369 | this_load = avg_load; | |
1370 | this = group; | |
1371 | } else if (avg_load < min_load) { | |
1372 | min_load = avg_load; | |
1373 | idlest = group; | |
1374 | } | |
1375 | } while (group = group->next, group != sd->groups); | |
1376 | ||
1377 | if (!idlest || 100*this_load < imbalance*min_load) | |
1378 | return NULL; | |
1379 | return idlest; | |
1380 | } | |
1381 | ||
1382 | /* | |
1383 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1384 | */ | |
1385 | static int | |
1386 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1387 | { | |
1388 | unsigned long load, min_load = ULONG_MAX; | |
1389 | int idlest = -1; | |
1390 | int i; | |
1391 | ||
1392 | /* Traverse only the allowed CPUs */ | |
1393 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1394 | load = weighted_cpuload(i); | |
1395 | ||
1396 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1397 | min_load = load; | |
1398 | idlest = i; | |
e7693a36 GH |
1399 | } |
1400 | } | |
1401 | ||
aaee1203 PZ |
1402 | return idlest; |
1403 | } | |
e7693a36 | 1404 | |
a50bde51 PZ |
1405 | /* |
1406 | * Try and locate an idle CPU in the sched_domain. | |
1407 | */ | |
1408 | static int | |
1409 | select_idle_sibling(struct task_struct *p, struct sched_domain *sd, int target) | |
1410 | { | |
1411 | int cpu = smp_processor_id(); | |
1412 | int prev_cpu = task_cpu(p); | |
1413 | int i; | |
1414 | ||
1415 | /* | |
1416 | * If this domain spans both cpu and prev_cpu (see the SD_WAKE_AFFINE | |
1417 | * test in select_task_rq_fair) and the prev_cpu is idle then that's | |
1418 | * always a better target than the current cpu. | |
1419 | */ | |
fe3bcfe1 PZ |
1420 | if (target == cpu && !cpu_rq(prev_cpu)->cfs.nr_running) |
1421 | return prev_cpu; | |
a50bde51 PZ |
1422 | |
1423 | /* | |
1424 | * Otherwise, iterate the domain and find an elegible idle cpu. | |
1425 | */ | |
fe3bcfe1 PZ |
1426 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { |
1427 | if (!cpu_rq(i)->cfs.nr_running) { | |
1428 | target = i; | |
1429 | break; | |
a50bde51 PZ |
1430 | } |
1431 | } | |
1432 | ||
1433 | return target; | |
1434 | } | |
1435 | ||
aaee1203 PZ |
1436 | /* |
1437 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1438 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1439 | * SD_BALANCE_EXEC. | |
1440 | * | |
1441 | * Balance, ie. select the least loaded group. | |
1442 | * | |
1443 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1444 | * | |
1445 | * preempt must be disabled. | |
1446 | */ | |
5158f4e4 | 1447 | static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 1448 | { |
29cd8bae | 1449 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1450 | int cpu = smp_processor_id(); |
1451 | int prev_cpu = task_cpu(p); | |
1452 | int new_cpu = cpu; | |
1453 | int want_affine = 0; | |
29cd8bae | 1454 | int want_sd = 1; |
5158f4e4 | 1455 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1456 | |
0763a660 | 1457 | if (sd_flag & SD_BALANCE_WAKE) { |
3f04e8cd MG |
1458 | if (sched_feat(AFFINE_WAKEUPS) && |
1459 | cpumask_test_cpu(cpu, &p->cpus_allowed)) | |
c88d5910 PZ |
1460 | want_affine = 1; |
1461 | new_cpu = prev_cpu; | |
1462 | } | |
aaee1203 PZ |
1463 | |
1464 | for_each_domain(cpu, tmp) { | |
e4f42888 PZ |
1465 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1466 | continue; | |
1467 | ||
aaee1203 | 1468 | /* |
ae154be1 PZ |
1469 | * If power savings logic is enabled for a domain, see if we |
1470 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1471 | */ |
59abf026 | 1472 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1473 | unsigned long power = 0; |
1474 | unsigned long nr_running = 0; | |
1475 | unsigned long capacity; | |
1476 | int i; | |
1477 | ||
1478 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1479 | power += power_of(i); | |
1480 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1481 | } | |
1482 | ||
1483 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1484 | ||
59abf026 PZ |
1485 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1486 | nr_running /= 2; | |
1487 | ||
1488 | if (nr_running < capacity) | |
29cd8bae | 1489 | want_sd = 0; |
ae154be1 | 1490 | } |
aaee1203 | 1491 | |
fe3bcfe1 PZ |
1492 | /* |
1493 | * While iterating the domains looking for a spanning | |
1494 | * WAKE_AFFINE domain, adjust the affine target to any idle cpu | |
1495 | * in cache sharing domains along the way. | |
1496 | */ | |
1497 | if (want_affine) { | |
a50bde51 | 1498 | int target = -1; |
c88d5910 | 1499 | |
a50bde51 PZ |
1500 | /* |
1501 | * If both cpu and prev_cpu are part of this domain, | |
1502 | * cpu is a valid SD_WAKE_AFFINE target. | |
1503 | */ | |
a1f84a3a | 1504 | if (cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) |
a50bde51 | 1505 | target = cpu; |
a1f84a3a MG |
1506 | |
1507 | /* | |
a50bde51 PZ |
1508 | * If there's an idle sibling in this domain, make that |
1509 | * the wake_affine target instead of the current cpu. | |
a1f84a3a | 1510 | */ |
a50bde51 PZ |
1511 | if (tmp->flags & SD_PREFER_SIBLING) |
1512 | target = select_idle_sibling(p, tmp, target); | |
a1f84a3a | 1513 | |
a50bde51 | 1514 | if (target >= 0) { |
fe3bcfe1 PZ |
1515 | if (tmp->flags & SD_WAKE_AFFINE) { |
1516 | affine_sd = tmp; | |
1517 | want_affine = 0; | |
1518 | } | |
a50bde51 | 1519 | cpu = target; |
a1f84a3a | 1520 | } |
c88d5910 PZ |
1521 | } |
1522 | ||
29cd8bae PZ |
1523 | if (!want_sd && !want_affine) |
1524 | break; | |
1525 | ||
0763a660 | 1526 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1527 | continue; |
1528 | ||
29cd8bae PZ |
1529 | if (want_sd) |
1530 | sd = tmp; | |
1531 | } | |
1532 | ||
1533 | if (sched_feat(LB_SHARES_UPDATE)) { | |
1534 | /* | |
1535 | * Pick the largest domain to update shares over | |
1536 | */ | |
1537 | tmp = sd; | |
1538 | if (affine_sd && (!tmp || | |
1539 | cpumask_weight(sched_domain_span(affine_sd)) > | |
1540 | cpumask_weight(sched_domain_span(sd)))) | |
1541 | tmp = affine_sd; | |
1542 | ||
1543 | if (tmp) | |
1544 | update_shares(tmp); | |
c88d5910 | 1545 | } |
aaee1203 | 1546 | |
fb58bac5 PZ |
1547 | if (affine_sd && wake_affine(affine_sd, p, sync)) |
1548 | return cpu; | |
e7693a36 | 1549 | |
aaee1203 | 1550 | while (sd) { |
5158f4e4 | 1551 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1552 | struct sched_group *group; |
c88d5910 | 1553 | int weight; |
098fb9db | 1554 | |
0763a660 | 1555 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1556 | sd = sd->child; |
1557 | continue; | |
1558 | } | |
098fb9db | 1559 | |
5158f4e4 PZ |
1560 | if (sd_flag & SD_BALANCE_WAKE) |
1561 | load_idx = sd->wake_idx; | |
098fb9db | 1562 | |
5158f4e4 | 1563 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1564 | if (!group) { |
1565 | sd = sd->child; | |
1566 | continue; | |
1567 | } | |
4ae7d5ce | 1568 | |
d7c33c49 | 1569 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1570 | if (new_cpu == -1 || new_cpu == cpu) { |
1571 | /* Now try balancing at a lower domain level of cpu */ | |
1572 | sd = sd->child; | |
1573 | continue; | |
e7693a36 | 1574 | } |
aaee1203 PZ |
1575 | |
1576 | /* Now try balancing at a lower domain level of new_cpu */ | |
1577 | cpu = new_cpu; | |
1578 | weight = cpumask_weight(sched_domain_span(sd)); | |
1579 | sd = NULL; | |
1580 | for_each_domain(cpu, tmp) { | |
1581 | if (weight <= cpumask_weight(sched_domain_span(tmp))) | |
1582 | break; | |
0763a660 | 1583 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1584 | sd = tmp; |
1585 | } | |
1586 | /* while loop will break here if sd == NULL */ | |
e7693a36 GH |
1587 | } |
1588 | ||
c88d5910 | 1589 | return new_cpu; |
e7693a36 GH |
1590 | } |
1591 | #endif /* CONFIG_SMP */ | |
1592 | ||
e52fb7c0 PZ |
1593 | /* |
1594 | * Adaptive granularity | |
1595 | * | |
1596 | * se->avg_wakeup gives the average time a task runs until it does a wakeup, | |
1597 | * with the limit of wakeup_gran -- when it never does a wakeup. | |
1598 | * | |
1599 | * So the smaller avg_wakeup is the faster we want this task to preempt, | |
1600 | * but we don't want to treat the preemptee unfairly and therefore allow it | |
1601 | * to run for at least the amount of time we'd like to run. | |
1602 | * | |
1603 | * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one | |
1604 | * | |
1605 | * NOTE: we use *nr_running to scale with load, this nicely matches the | |
1606 | * degrading latency on load. | |
1607 | */ | |
1608 | static unsigned long | |
1609 | adaptive_gran(struct sched_entity *curr, struct sched_entity *se) | |
1610 | { | |
1611 | u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
1612 | u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running; | |
1613 | u64 gran = 0; | |
1614 | ||
1615 | if (this_run < expected_wakeup) | |
1616 | gran = expected_wakeup - this_run; | |
1617 | ||
1618 | return min_t(s64, gran, sysctl_sched_wakeup_granularity); | |
1619 | } | |
1620 | ||
1621 | static unsigned long | |
1622 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1623 | { |
1624 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1625 | ||
e52fb7c0 PZ |
1626 | if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN)) |
1627 | gran = adaptive_gran(curr, se); | |
1628 | ||
0bbd3336 | 1629 | /* |
e52fb7c0 PZ |
1630 | * Since its curr running now, convert the gran from real-time |
1631 | * to virtual-time in his units. | |
0bbd3336 | 1632 | */ |
e52fb7c0 PZ |
1633 | if (sched_feat(ASYM_GRAN)) { |
1634 | /* | |
1635 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1636 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1637 | * the resulting gran will be larger, therefore penalizing the | |
1638 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1639 | * be smaller, again penalizing the lighter task. | |
1640 | * | |
1641 | * This is especially important for buddies when the leftmost | |
1642 | * task is higher priority than the buddy. | |
1643 | */ | |
1644 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1645 | gran = calc_delta_fair(gran, se); | |
1646 | } else { | |
1647 | if (unlikely(curr->load.weight != NICE_0_LOAD)) | |
1648 | gran = calc_delta_fair(gran, curr); | |
1649 | } | |
0bbd3336 PZ |
1650 | |
1651 | return gran; | |
1652 | } | |
1653 | ||
464b7527 PZ |
1654 | /* |
1655 | * Should 'se' preempt 'curr'. | |
1656 | * | |
1657 | * |s1 | |
1658 | * |s2 | |
1659 | * |s3 | |
1660 | * g | |
1661 | * |<--->|c | |
1662 | * | |
1663 | * w(c, s1) = -1 | |
1664 | * w(c, s2) = 0 | |
1665 | * w(c, s3) = 1 | |
1666 | * | |
1667 | */ | |
1668 | static int | |
1669 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1670 | { | |
1671 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1672 | ||
1673 | if (vdiff <= 0) | |
1674 | return -1; | |
1675 | ||
e52fb7c0 | 1676 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1677 | if (vdiff > gran) |
1678 | return 1; | |
1679 | ||
1680 | return 0; | |
1681 | } | |
1682 | ||
02479099 PZ |
1683 | static void set_last_buddy(struct sched_entity *se) |
1684 | { | |
6bc912b7 PZ |
1685 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1686 | for_each_sched_entity(se) | |
1687 | cfs_rq_of(se)->last = se; | |
1688 | } | |
02479099 PZ |
1689 | } |
1690 | ||
1691 | static void set_next_buddy(struct sched_entity *se) | |
1692 | { | |
6bc912b7 PZ |
1693 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1694 | for_each_sched_entity(se) | |
1695 | cfs_rq_of(se)->next = se; | |
1696 | } | |
02479099 PZ |
1697 | } |
1698 | ||
bf0f6f24 IM |
1699 | /* |
1700 | * Preempt the current task with a newly woken task if needed: | |
1701 | */ | |
5a9b86f6 | 1702 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1703 | { |
1704 | struct task_struct *curr = rq->curr; | |
8651a86c | 1705 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1706 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
5a9b86f6 | 1707 | int sync = wake_flags & WF_SYNC; |
f685ceac | 1708 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
bf0f6f24 | 1709 | |
3a7e73a2 PZ |
1710 | if (unlikely(rt_prio(p->prio))) |
1711 | goto preempt; | |
aa2ac252 | 1712 | |
d95f98d0 PZ |
1713 | if (unlikely(p->sched_class != &fair_sched_class)) |
1714 | return; | |
1715 | ||
4ae7d5ce IM |
1716 | if (unlikely(se == pse)) |
1717 | return; | |
1718 | ||
f685ceac | 1719 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) |
3cb63d52 | 1720 | set_next_buddy(pse); |
57fdc26d | 1721 | |
aec0a514 BR |
1722 | /* |
1723 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1724 | * wake up path. | |
1725 | */ | |
1726 | if (test_tsk_need_resched(curr)) | |
1727 | return; | |
1728 | ||
91c234b4 | 1729 | /* |
6bc912b7 | 1730 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1731 | * the tick): |
1732 | */ | |
6bc912b7 | 1733 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1734 | return; |
bf0f6f24 | 1735 | |
6bc912b7 | 1736 | /* Idle tasks are by definition preempted by everybody. */ |
3a7e73a2 PZ |
1737 | if (unlikely(curr->policy == SCHED_IDLE)) |
1738 | goto preempt; | |
bf0f6f24 | 1739 | |
3a7e73a2 PZ |
1740 | if (sched_feat(WAKEUP_SYNC) && sync) |
1741 | goto preempt; | |
15afe09b | 1742 | |
3a7e73a2 PZ |
1743 | if (sched_feat(WAKEUP_OVERLAP) && |
1744 | se->avg_overlap < sysctl_sched_migration_cost && | |
1745 | pse->avg_overlap < sysctl_sched_migration_cost) | |
1746 | goto preempt; | |
1747 | ||
ad4b78bb PZ |
1748 | if (!sched_feat(WAKEUP_PREEMPT)) |
1749 | return; | |
1750 | ||
3a7e73a2 | 1751 | update_curr(cfs_rq); |
464b7527 | 1752 | find_matching_se(&se, &pse); |
002f128b | 1753 | BUG_ON(!pse); |
3a7e73a2 PZ |
1754 | if (wakeup_preempt_entity(se, pse) == 1) |
1755 | goto preempt; | |
464b7527 | 1756 | |
3a7e73a2 | 1757 | return; |
a65ac745 | 1758 | |
3a7e73a2 PZ |
1759 | preempt: |
1760 | resched_task(curr); | |
1761 | /* | |
1762 | * Only set the backward buddy when the current task is still | |
1763 | * on the rq. This can happen when a wakeup gets interleaved | |
1764 | * with schedule on the ->pre_schedule() or idle_balance() | |
1765 | * point, either of which can * drop the rq lock. | |
1766 | * | |
1767 | * Also, during early boot the idle thread is in the fair class, | |
1768 | * for obvious reasons its a bad idea to schedule back to it. | |
1769 | */ | |
1770 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1771 | return; | |
1772 | ||
1773 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1774 | set_last_buddy(se); | |
bf0f6f24 IM |
1775 | } |
1776 | ||
fb8d4724 | 1777 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1778 | { |
8f4d37ec | 1779 | struct task_struct *p; |
bf0f6f24 IM |
1780 | struct cfs_rq *cfs_rq = &rq->cfs; |
1781 | struct sched_entity *se; | |
1782 | ||
36ace27e | 1783 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
1784 | return NULL; |
1785 | ||
1786 | do { | |
9948f4b2 | 1787 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1788 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1789 | cfs_rq = group_cfs_rq(se); |
1790 | } while (cfs_rq); | |
1791 | ||
8f4d37ec PZ |
1792 | p = task_of(se); |
1793 | hrtick_start_fair(rq, p); | |
1794 | ||
1795 | return p; | |
bf0f6f24 IM |
1796 | } |
1797 | ||
1798 | /* | |
1799 | * Account for a descheduled task: | |
1800 | */ | |
31ee529c | 1801 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1802 | { |
1803 | struct sched_entity *se = &prev->se; | |
1804 | struct cfs_rq *cfs_rq; | |
1805 | ||
1806 | for_each_sched_entity(se) { | |
1807 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1808 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1809 | } |
1810 | } | |
1811 | ||
681f3e68 | 1812 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1813 | /************************************************** |
1814 | * Fair scheduling class load-balancing methods: | |
1815 | */ | |
1816 | ||
1817 | /* | |
1818 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1819 | * during the whole iteration, the current task might be | |
1820 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1821 | * achieve that by always pre-iterating before returning | |
1822 | * the current task: | |
1823 | */ | |
a9957449 | 1824 | static struct task_struct * |
4a55bd5e | 1825 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) |
bf0f6f24 | 1826 | { |
354d60c2 DG |
1827 | struct task_struct *p = NULL; |
1828 | struct sched_entity *se; | |
bf0f6f24 | 1829 | |
77ae6513 MG |
1830 | if (next == &cfs_rq->tasks) |
1831 | return NULL; | |
1832 | ||
b87f1724 BR |
1833 | se = list_entry(next, struct sched_entity, group_node); |
1834 | p = task_of(se); | |
1835 | cfs_rq->balance_iterator = next->next; | |
77ae6513 | 1836 | |
bf0f6f24 IM |
1837 | return p; |
1838 | } | |
1839 | ||
1840 | static struct task_struct *load_balance_start_fair(void *arg) | |
1841 | { | |
1842 | struct cfs_rq *cfs_rq = arg; | |
1843 | ||
4a55bd5e | 1844 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); |
bf0f6f24 IM |
1845 | } |
1846 | ||
1847 | static struct task_struct *load_balance_next_fair(void *arg) | |
1848 | { | |
1849 | struct cfs_rq *cfs_rq = arg; | |
1850 | ||
4a55bd5e | 1851 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); |
bf0f6f24 IM |
1852 | } |
1853 | ||
3d45fd80 PZ |
1854 | /* |
1855 | * runqueue iterator, to support SMP load-balancing between different | |
1856 | * scheduling classes, without having to expose their internal data | |
1857 | * structures to the load-balancing proper: | |
1858 | */ | |
1859 | struct rq_iterator { | |
1860 | void *arg; | |
1861 | struct task_struct *(*start)(void *); | |
1862 | struct task_struct *(*next)(void *); | |
1863 | }; | |
1864 | ||
1865 | static unsigned long | |
1866 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1867 | unsigned long max_load_move, struct sched_domain *sd, | |
1868 | enum cpu_idle_type idle, int *all_pinned, | |
1869 | int *this_best_prio, struct rq_iterator *iterator); | |
1870 | ||
1871 | ||
c09595f6 PZ |
1872 | static unsigned long |
1873 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1874 | unsigned long max_load_move, struct sched_domain *sd, | |
1875 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, | |
1876 | struct cfs_rq *cfs_rq) | |
62fb1851 | 1877 | { |
c09595f6 | 1878 | struct rq_iterator cfs_rq_iterator; |
62fb1851 | 1879 | |
c09595f6 PZ |
1880 | cfs_rq_iterator.start = load_balance_start_fair; |
1881 | cfs_rq_iterator.next = load_balance_next_fair; | |
1882 | cfs_rq_iterator.arg = cfs_rq; | |
62fb1851 | 1883 | |
c09595f6 PZ |
1884 | return balance_tasks(this_rq, this_cpu, busiest, |
1885 | max_load_move, sd, idle, all_pinned, | |
1886 | this_best_prio, &cfs_rq_iterator); | |
62fb1851 | 1887 | } |
62fb1851 | 1888 | |
c09595f6 | 1889 | #ifdef CONFIG_FAIR_GROUP_SCHED |
43010659 | 1890 | static unsigned long |
bf0f6f24 | 1891 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1892 | unsigned long max_load_move, |
a4ac01c3 PW |
1893 | struct sched_domain *sd, enum cpu_idle_type idle, |
1894 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 | 1895 | { |
bf0f6f24 | 1896 | long rem_load_move = max_load_move; |
c09595f6 PZ |
1897 | int busiest_cpu = cpu_of(busiest); |
1898 | struct task_group *tg; | |
18d95a28 | 1899 | |
c09595f6 | 1900 | rcu_read_lock(); |
c8cba857 | 1901 | update_h_load(busiest_cpu); |
18d95a28 | 1902 | |
caea8a03 | 1903 | list_for_each_entry_rcu(tg, &task_groups, list) { |
c8cba857 | 1904 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; |
42a3ac7d PZ |
1905 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
1906 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
243e0e7b | 1907 | u64 rem_load, moved_load; |
18d95a28 | 1908 | |
c09595f6 PZ |
1909 | /* |
1910 | * empty group | |
1911 | */ | |
c8cba857 | 1912 | if (!busiest_cfs_rq->task_weight) |
bf0f6f24 IM |
1913 | continue; |
1914 | ||
243e0e7b SV |
1915 | rem_load = (u64)rem_load_move * busiest_weight; |
1916 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
bf0f6f24 | 1917 | |
c09595f6 | 1918 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, |
53fecd8a | 1919 | rem_load, sd, idle, all_pinned, this_best_prio, |
c09595f6 | 1920 | tg->cfs_rq[busiest_cpu]); |
bf0f6f24 | 1921 | |
c09595f6 | 1922 | if (!moved_load) |
bf0f6f24 IM |
1923 | continue; |
1924 | ||
42a3ac7d | 1925 | moved_load *= busiest_h_load; |
243e0e7b | 1926 | moved_load = div_u64(moved_load, busiest_weight + 1); |
bf0f6f24 | 1927 | |
c09595f6 PZ |
1928 | rem_load_move -= moved_load; |
1929 | if (rem_load_move < 0) | |
bf0f6f24 IM |
1930 | break; |
1931 | } | |
c09595f6 | 1932 | rcu_read_unlock(); |
bf0f6f24 | 1933 | |
43010659 | 1934 | return max_load_move - rem_load_move; |
bf0f6f24 | 1935 | } |
c09595f6 PZ |
1936 | #else |
1937 | static unsigned long | |
1938 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1939 | unsigned long max_load_move, | |
1940 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1941 | int *all_pinned, int *this_best_prio) | |
1942 | { | |
1943 | return __load_balance_fair(this_rq, this_cpu, busiest, | |
1944 | max_load_move, sd, idle, all_pinned, | |
1945 | this_best_prio, &busiest->cfs); | |
1946 | } | |
1947 | #endif | |
bf0f6f24 | 1948 | |
e1d1484f | 1949 | static int |
3d45fd80 PZ |
1950 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, |
1951 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1952 | struct rq_iterator *iterator); | |
1953 | ||
1954 | /* | |
1955 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
1956 | * part of active balancing operations within "domain". | |
1957 | * Returns 1 if successful and 0 otherwise. | |
1958 | * | |
1959 | * Called with both runqueues locked. | |
1960 | */ | |
1961 | static int | |
1962 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1963 | struct sched_domain *sd, enum cpu_idle_type idle) | |
e1d1484f PW |
1964 | { |
1965 | struct cfs_rq *busy_cfs_rq; | |
1966 | struct rq_iterator cfs_rq_iterator; | |
1967 | ||
1968 | cfs_rq_iterator.start = load_balance_start_fair; | |
1969 | cfs_rq_iterator.next = load_balance_next_fair; | |
1970 | ||
1971 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1972 | /* | |
1973 | * pass busy_cfs_rq argument into | |
1974 | * load_balance_[start|next]_fair iterators | |
1975 | */ | |
1976 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1977 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1978 | &cfs_rq_iterator)) | |
1979 | return 1; | |
1980 | } | |
1981 | ||
1982 | return 0; | |
1983 | } | |
0bcdcf28 | 1984 | |
1e3c88bd PZ |
1985 | /* |
1986 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1987 | * Both runqueues must be locked. | |
1988 | */ | |
1989 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
1990 | struct rq *this_rq, int this_cpu) | |
1991 | { | |
1992 | deactivate_task(src_rq, p, 0); | |
1993 | set_task_cpu(p, this_cpu); | |
1994 | activate_task(this_rq, p, 0); | |
1995 | check_preempt_curr(this_rq, p, 0); | |
1996 | } | |
1997 | ||
1998 | /* | |
1999 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2000 | */ | |
2001 | static | |
2002 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
2003 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2004 | int *all_pinned) | |
2005 | { | |
2006 | int tsk_cache_hot = 0; | |
2007 | /* | |
2008 | * We do not migrate tasks that are: | |
2009 | * 1) running (obviously), or | |
2010 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2011 | * 3) are cache-hot on their current CPU. | |
2012 | */ | |
2013 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
2014 | schedstat_inc(p, se.nr_failed_migrations_affine); | |
2015 | return 0; | |
2016 | } | |
2017 | *all_pinned = 0; | |
2018 | ||
2019 | if (task_running(rq, p)) { | |
2020 | schedstat_inc(p, se.nr_failed_migrations_running); | |
2021 | return 0; | |
2022 | } | |
2023 | ||
2024 | /* | |
2025 | * Aggressive migration if: | |
2026 | * 1) task is cache cold, or | |
2027 | * 2) too many balance attempts have failed. | |
2028 | */ | |
2029 | ||
2030 | tsk_cache_hot = task_hot(p, rq->clock, sd); | |
2031 | if (!tsk_cache_hot || | |
2032 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
2033 | #ifdef CONFIG_SCHEDSTATS | |
2034 | if (tsk_cache_hot) { | |
2035 | schedstat_inc(sd, lb_hot_gained[idle]); | |
2036 | schedstat_inc(p, se.nr_forced_migrations); | |
2037 | } | |
2038 | #endif | |
2039 | return 1; | |
2040 | } | |
2041 | ||
2042 | if (tsk_cache_hot) { | |
2043 | schedstat_inc(p, se.nr_failed_migrations_hot); | |
2044 | return 0; | |
2045 | } | |
2046 | return 1; | |
2047 | } | |
2048 | ||
2049 | static unsigned long | |
2050 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2051 | unsigned long max_load_move, struct sched_domain *sd, | |
2052 | enum cpu_idle_type idle, int *all_pinned, | |
2053 | int *this_best_prio, struct rq_iterator *iterator) | |
2054 | { | |
2055 | int loops = 0, pulled = 0, pinned = 0; | |
2056 | struct task_struct *p; | |
2057 | long rem_load_move = max_load_move; | |
2058 | ||
2059 | if (max_load_move == 0) | |
2060 | goto out; | |
2061 | ||
2062 | pinned = 1; | |
2063 | ||
2064 | /* | |
2065 | * Start the load-balancing iterator: | |
2066 | */ | |
2067 | p = iterator->start(iterator->arg); | |
2068 | next: | |
2069 | if (!p || loops++ > sysctl_sched_nr_migrate) | |
2070 | goto out; | |
2071 | ||
2072 | if ((p->se.load.weight >> 1) > rem_load_move || | |
2073 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
2074 | p = iterator->next(iterator->arg); | |
2075 | goto next; | |
2076 | } | |
2077 | ||
2078 | pull_task(busiest, p, this_rq, this_cpu); | |
2079 | pulled++; | |
2080 | rem_load_move -= p->se.load.weight; | |
2081 | ||
2082 | #ifdef CONFIG_PREEMPT | |
2083 | /* | |
2084 | * NEWIDLE balancing is a source of latency, so preemptible kernels | |
2085 | * will stop after the first task is pulled to minimize the critical | |
2086 | * section. | |
2087 | */ | |
2088 | if (idle == CPU_NEWLY_IDLE) | |
2089 | goto out; | |
2090 | #endif | |
2091 | ||
2092 | /* | |
2093 | * We only want to steal up to the prescribed amount of weighted load. | |
2094 | */ | |
2095 | if (rem_load_move > 0) { | |
2096 | if (p->prio < *this_best_prio) | |
2097 | *this_best_prio = p->prio; | |
2098 | p = iterator->next(iterator->arg); | |
2099 | goto next; | |
2100 | } | |
2101 | out: | |
2102 | /* | |
2103 | * Right now, this is one of only two places pull_task() is called, | |
2104 | * so we can safely collect pull_task() stats here rather than | |
2105 | * inside pull_task(). | |
2106 | */ | |
2107 | schedstat_add(sd, lb_gained[idle], pulled); | |
2108 | ||
2109 | if (all_pinned) | |
2110 | *all_pinned = pinned; | |
2111 | ||
2112 | return max_load_move - rem_load_move; | |
2113 | } | |
2114 | ||
2115 | /* | |
2116 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2117 | * this_rq, as part of a balancing operation within domain "sd". | |
2118 | * Returns 1 if successful and 0 otherwise. | |
2119 | * | |
2120 | * Called with both runqueues locked. | |
2121 | */ | |
2122 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2123 | unsigned long max_load_move, | |
2124 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2125 | int *all_pinned) | |
2126 | { | |
3d45fd80 | 2127 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2128 | int this_best_prio = this_rq->curr->prio; |
2129 | ||
2130 | do { | |
3d45fd80 | 2131 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd PZ |
2132 | max_load_move - total_load_moved, |
2133 | sd, idle, all_pinned, &this_best_prio); | |
3d45fd80 PZ |
2134 | |
2135 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2136 | |
2137 | #ifdef CONFIG_PREEMPT | |
2138 | /* | |
2139 | * NEWIDLE balancing is a source of latency, so preemptible | |
2140 | * kernels will stop after the first task is pulled to minimize | |
2141 | * the critical section. | |
2142 | */ | |
2143 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2144 | break; | |
2145 | #endif | |
3d45fd80 | 2146 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2147 | |
2148 | return total_load_moved > 0; | |
2149 | } | |
2150 | ||
2151 | static int | |
2152 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2153 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2154 | struct rq_iterator *iterator) | |
2155 | { | |
2156 | struct task_struct *p = iterator->start(iterator->arg); | |
2157 | int pinned = 0; | |
2158 | ||
2159 | while (p) { | |
2160 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
2161 | pull_task(busiest, p, this_rq, this_cpu); | |
2162 | /* | |
2163 | * Right now, this is only the second place pull_task() | |
2164 | * is called, so we can safely collect pull_task() | |
2165 | * stats here rather than inside pull_task(). | |
2166 | */ | |
2167 | schedstat_inc(sd, lb_gained[idle]); | |
2168 | ||
2169 | return 1; | |
2170 | } | |
2171 | p = iterator->next(iterator->arg); | |
2172 | } | |
2173 | ||
2174 | return 0; | |
2175 | } | |
2176 | ||
1e3c88bd PZ |
2177 | /********** Helpers for find_busiest_group ************************/ |
2178 | /* | |
2179 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2180 | * during load balancing. | |
2181 | */ | |
2182 | struct sd_lb_stats { | |
2183 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2184 | struct sched_group *this; /* Local group in this sd */ | |
2185 | unsigned long total_load; /* Total load of all groups in sd */ | |
2186 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2187 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2188 | ||
2189 | /** Statistics of this group */ | |
2190 | unsigned long this_load; | |
2191 | unsigned long this_load_per_task; | |
2192 | unsigned long this_nr_running; | |
2193 | ||
2194 | /* Statistics of the busiest group */ | |
2195 | unsigned long max_load; | |
2196 | unsigned long busiest_load_per_task; | |
2197 | unsigned long busiest_nr_running; | |
2198 | ||
2199 | int group_imb; /* Is there imbalance in this sd */ | |
2200 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2201 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2202 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2203 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2204 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2205 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2206 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2207 | #endif | |
2208 | }; | |
2209 | ||
2210 | /* | |
2211 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2212 | */ | |
2213 | struct sg_lb_stats { | |
2214 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2215 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2216 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2217 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2218 | unsigned long group_capacity; | |
2219 | int group_imb; /* Is there an imbalance in the group ? */ | |
2220 | }; | |
2221 | ||
2222 | /** | |
2223 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2224 | * @group: The group whose first cpu is to be returned. | |
2225 | */ | |
2226 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2227 | { | |
2228 | return cpumask_first(sched_group_cpus(group)); | |
2229 | } | |
2230 | ||
2231 | /** | |
2232 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2233 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2234 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2235 | */ | |
2236 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2237 | enum cpu_idle_type idle) | |
2238 | { | |
2239 | int load_idx; | |
2240 | ||
2241 | switch (idle) { | |
2242 | case CPU_NOT_IDLE: | |
2243 | load_idx = sd->busy_idx; | |
2244 | break; | |
2245 | ||
2246 | case CPU_NEWLY_IDLE: | |
2247 | load_idx = sd->newidle_idx; | |
2248 | break; | |
2249 | default: | |
2250 | load_idx = sd->idle_idx; | |
2251 | break; | |
2252 | } | |
2253 | ||
2254 | return load_idx; | |
2255 | } | |
2256 | ||
2257 | ||
2258 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2259 | /** | |
2260 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2261 | * the given sched_domain, during load balancing. | |
2262 | * | |
2263 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2264 | * @sds: Variable containing the statistics for sd. | |
2265 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2266 | */ | |
2267 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2268 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2269 | { | |
2270 | /* | |
2271 | * Busy processors will not participate in power savings | |
2272 | * balance. | |
2273 | */ | |
2274 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2275 | sds->power_savings_balance = 0; | |
2276 | else { | |
2277 | sds->power_savings_balance = 1; | |
2278 | sds->min_nr_running = ULONG_MAX; | |
2279 | sds->leader_nr_running = 0; | |
2280 | } | |
2281 | } | |
2282 | ||
2283 | /** | |
2284 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2285 | * sched_domain while performing load balancing. | |
2286 | * | |
2287 | * @group: sched_group belonging to the sched_domain under consideration. | |
2288 | * @sds: Variable containing the statistics of the sched_domain | |
2289 | * @local_group: Does group contain the CPU for which we're performing | |
2290 | * load balancing ? | |
2291 | * @sgs: Variable containing the statistics of the group. | |
2292 | */ | |
2293 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2294 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2295 | { | |
2296 | ||
2297 | if (!sds->power_savings_balance) | |
2298 | return; | |
2299 | ||
2300 | /* | |
2301 | * If the local group is idle or completely loaded | |
2302 | * no need to do power savings balance at this domain | |
2303 | */ | |
2304 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2305 | !sds->this_nr_running)) | |
2306 | sds->power_savings_balance = 0; | |
2307 | ||
2308 | /* | |
2309 | * If a group is already running at full capacity or idle, | |
2310 | * don't include that group in power savings calculations | |
2311 | */ | |
2312 | if (!sds->power_savings_balance || | |
2313 | sgs->sum_nr_running >= sgs->group_capacity || | |
2314 | !sgs->sum_nr_running) | |
2315 | return; | |
2316 | ||
2317 | /* | |
2318 | * Calculate the group which has the least non-idle load. | |
2319 | * This is the group from where we need to pick up the load | |
2320 | * for saving power | |
2321 | */ | |
2322 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2323 | (sgs->sum_nr_running == sds->min_nr_running && | |
2324 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2325 | sds->group_min = group; | |
2326 | sds->min_nr_running = sgs->sum_nr_running; | |
2327 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2328 | sgs->sum_nr_running; | |
2329 | } | |
2330 | ||
2331 | /* | |
2332 | * Calculate the group which is almost near its | |
2333 | * capacity but still has some space to pick up some load | |
2334 | * from other group and save more power | |
2335 | */ | |
2336 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2337 | return; | |
2338 | ||
2339 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2340 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2341 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2342 | sds->group_leader = group; | |
2343 | sds->leader_nr_running = sgs->sum_nr_running; | |
2344 | } | |
2345 | } | |
2346 | ||
2347 | /** | |
2348 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2349 | * @sds: Variable containing the statistics of the sched_domain | |
2350 | * under consideration. | |
2351 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2352 | * @imbalance: Variable to store the imbalance. | |
2353 | * | |
2354 | * Description: | |
2355 | * Check if we have potential to perform some power-savings balance. | |
2356 | * If yes, set the busiest group to be the least loaded group in the | |
2357 | * sched_domain, so that it's CPUs can be put to idle. | |
2358 | * | |
2359 | * Returns 1 if there is potential to perform power-savings balance. | |
2360 | * Else returns 0. | |
2361 | */ | |
2362 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2363 | int this_cpu, unsigned long *imbalance) | |
2364 | { | |
2365 | if (!sds->power_savings_balance) | |
2366 | return 0; | |
2367 | ||
2368 | if (sds->this != sds->group_leader || | |
2369 | sds->group_leader == sds->group_min) | |
2370 | return 0; | |
2371 | ||
2372 | *imbalance = sds->min_load_per_task; | |
2373 | sds->busiest = sds->group_min; | |
2374 | ||
2375 | return 1; | |
2376 | ||
2377 | } | |
2378 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2379 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2380 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2381 | { | |
2382 | return; | |
2383 | } | |
2384 | ||
2385 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2386 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2387 | { | |
2388 | return; | |
2389 | } | |
2390 | ||
2391 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2392 | int this_cpu, unsigned long *imbalance) | |
2393 | { | |
2394 | return 0; | |
2395 | } | |
2396 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2397 | ||
2398 | ||
2399 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2400 | { | |
2401 | return SCHED_LOAD_SCALE; | |
2402 | } | |
2403 | ||
2404 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2405 | { | |
2406 | return default_scale_freq_power(sd, cpu); | |
2407 | } | |
2408 | ||
2409 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2410 | { | |
2411 | unsigned long weight = cpumask_weight(sched_domain_span(sd)); | |
2412 | unsigned long smt_gain = sd->smt_gain; | |
2413 | ||
2414 | smt_gain /= weight; | |
2415 | ||
2416 | return smt_gain; | |
2417 | } | |
2418 | ||
2419 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2420 | { | |
2421 | return default_scale_smt_power(sd, cpu); | |
2422 | } | |
2423 | ||
2424 | unsigned long scale_rt_power(int cpu) | |
2425 | { | |
2426 | struct rq *rq = cpu_rq(cpu); | |
2427 | u64 total, available; | |
2428 | ||
2429 | sched_avg_update(rq); | |
2430 | ||
2431 | total = sched_avg_period() + (rq->clock - rq->age_stamp); | |
2432 | available = total - rq->rt_avg; | |
2433 | ||
2434 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
2435 | total = SCHED_LOAD_SCALE; | |
2436 | ||
2437 | total >>= SCHED_LOAD_SHIFT; | |
2438 | ||
2439 | return div_u64(available, total); | |
2440 | } | |
2441 | ||
2442 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2443 | { | |
2444 | unsigned long weight = cpumask_weight(sched_domain_span(sd)); | |
2445 | unsigned long power = SCHED_LOAD_SCALE; | |
2446 | struct sched_group *sdg = sd->groups; | |
2447 | ||
2448 | if (sched_feat(ARCH_POWER)) | |
2449 | power *= arch_scale_freq_power(sd, cpu); | |
2450 | else | |
2451 | power *= default_scale_freq_power(sd, cpu); | |
2452 | ||
2453 | power >>= SCHED_LOAD_SHIFT; | |
2454 | ||
2455 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { | |
2456 | if (sched_feat(ARCH_POWER)) | |
2457 | power *= arch_scale_smt_power(sd, cpu); | |
2458 | else | |
2459 | power *= default_scale_smt_power(sd, cpu); | |
2460 | ||
2461 | power >>= SCHED_LOAD_SHIFT; | |
2462 | } | |
2463 | ||
2464 | power *= scale_rt_power(cpu); | |
2465 | power >>= SCHED_LOAD_SHIFT; | |
2466 | ||
2467 | if (!power) | |
2468 | power = 1; | |
2469 | ||
2470 | sdg->cpu_power = power; | |
2471 | } | |
2472 | ||
2473 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2474 | { | |
2475 | struct sched_domain *child = sd->child; | |
2476 | struct sched_group *group, *sdg = sd->groups; | |
2477 | unsigned long power; | |
2478 | ||
2479 | if (!child) { | |
2480 | update_cpu_power(sd, cpu); | |
2481 | return; | |
2482 | } | |
2483 | ||
2484 | power = 0; | |
2485 | ||
2486 | group = child->groups; | |
2487 | do { | |
2488 | power += group->cpu_power; | |
2489 | group = group->next; | |
2490 | } while (group != child->groups); | |
2491 | ||
2492 | sdg->cpu_power = power; | |
2493 | } | |
2494 | ||
2495 | /** | |
2496 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2497 | * @sd: The sched_domain whose statistics are to be updated. | |
2498 | * @group: sched_group whose statistics are to be updated. | |
2499 | * @this_cpu: Cpu for which load balance is currently performed. | |
2500 | * @idle: Idle status of this_cpu | |
2501 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
2502 | * @sd_idle: Idle status of the sched_domain containing group. | |
2503 | * @local_group: Does group contain this_cpu. | |
2504 | * @cpus: Set of cpus considered for load balancing. | |
2505 | * @balance: Should we balance. | |
2506 | * @sgs: variable to hold the statistics for this group. | |
2507 | */ | |
2508 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2509 | struct sched_group *group, int this_cpu, | |
2510 | enum cpu_idle_type idle, int load_idx, int *sd_idle, | |
2511 | int local_group, const struct cpumask *cpus, | |
2512 | int *balance, struct sg_lb_stats *sgs) | |
2513 | { | |
2514 | unsigned long load, max_cpu_load, min_cpu_load; | |
2515 | int i; | |
2516 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
2517 | unsigned long sum_avg_load_per_task; | |
2518 | unsigned long avg_load_per_task; | |
2519 | ||
2520 | if (local_group) { | |
2521 | balance_cpu = group_first_cpu(group); | |
2522 | if (balance_cpu == this_cpu) | |
2523 | update_group_power(sd, this_cpu); | |
2524 | } | |
2525 | ||
2526 | /* Tally up the load of all CPUs in the group */ | |
2527 | sum_avg_load_per_task = avg_load_per_task = 0; | |
2528 | max_cpu_load = 0; | |
2529 | min_cpu_load = ~0UL; | |
2530 | ||
2531 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2532 | struct rq *rq = cpu_rq(i); | |
2533 | ||
2534 | if (*sd_idle && rq->nr_running) | |
2535 | *sd_idle = 0; | |
2536 | ||
2537 | /* Bias balancing toward cpus of our domain */ | |
2538 | if (local_group) { | |
2539 | if (idle_cpu(i) && !first_idle_cpu) { | |
2540 | first_idle_cpu = 1; | |
2541 | balance_cpu = i; | |
2542 | } | |
2543 | ||
2544 | load = target_load(i, load_idx); | |
2545 | } else { | |
2546 | load = source_load(i, load_idx); | |
2547 | if (load > max_cpu_load) | |
2548 | max_cpu_load = load; | |
2549 | if (min_cpu_load > load) | |
2550 | min_cpu_load = load; | |
2551 | } | |
2552 | ||
2553 | sgs->group_load += load; | |
2554 | sgs->sum_nr_running += rq->nr_running; | |
2555 | sgs->sum_weighted_load += weighted_cpuload(i); | |
2556 | ||
2557 | sum_avg_load_per_task += cpu_avg_load_per_task(i); | |
2558 | } | |
2559 | ||
2560 | /* | |
2561 | * First idle cpu or the first cpu(busiest) in this sched group | |
2562 | * is eligible for doing load balancing at this and above | |
2563 | * domains. In the newly idle case, we will allow all the cpu's | |
2564 | * to do the newly idle load balance. | |
2565 | */ | |
2566 | if (idle != CPU_NEWLY_IDLE && local_group && | |
2567 | balance_cpu != this_cpu && balance) { | |
2568 | *balance = 0; | |
2569 | return; | |
2570 | } | |
2571 | ||
2572 | /* Adjust by relative CPU power of the group */ | |
2573 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2574 | ||
2575 | ||
2576 | /* | |
2577 | * Consider the group unbalanced when the imbalance is larger | |
2578 | * than the average weight of two tasks. | |
2579 | * | |
2580 | * APZ: with cgroup the avg task weight can vary wildly and | |
2581 | * might not be a suitable number - should we keep a | |
2582 | * normalized nr_running number somewhere that negates | |
2583 | * the hierarchy? | |
2584 | */ | |
2585 | avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) / | |
2586 | group->cpu_power; | |
2587 | ||
2588 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task) | |
2589 | sgs->group_imb = 1; | |
2590 | ||
2591 | sgs->group_capacity = | |
2592 | DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); | |
2593 | } | |
2594 | ||
2595 | /** | |
2596 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2597 | * @sd: sched_domain whose statistics are to be updated. | |
2598 | * @this_cpu: Cpu for which load balance is currently performed. | |
2599 | * @idle: Idle status of this_cpu | |
2600 | * @sd_idle: Idle status of the sched_domain containing group. | |
2601 | * @cpus: Set of cpus considered for load balancing. | |
2602 | * @balance: Should we balance. | |
2603 | * @sds: variable to hold the statistics for this sched_domain. | |
2604 | */ | |
2605 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
2606 | enum cpu_idle_type idle, int *sd_idle, | |
2607 | const struct cpumask *cpus, int *balance, | |
2608 | struct sd_lb_stats *sds) | |
2609 | { | |
2610 | struct sched_domain *child = sd->child; | |
2611 | struct sched_group *group = sd->groups; | |
2612 | struct sg_lb_stats sgs; | |
2613 | int load_idx, prefer_sibling = 0; | |
2614 | ||
2615 | if (child && child->flags & SD_PREFER_SIBLING) | |
2616 | prefer_sibling = 1; | |
2617 | ||
2618 | init_sd_power_savings_stats(sd, sds, idle); | |
2619 | load_idx = get_sd_load_idx(sd, idle); | |
2620 | ||
2621 | do { | |
2622 | int local_group; | |
2623 | ||
2624 | local_group = cpumask_test_cpu(this_cpu, | |
2625 | sched_group_cpus(group)); | |
2626 | memset(&sgs, 0, sizeof(sgs)); | |
2627 | update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle, | |
2628 | local_group, cpus, balance, &sgs); | |
2629 | ||
2630 | if (local_group && balance && !(*balance)) | |
2631 | return; | |
2632 | ||
2633 | sds->total_load += sgs.group_load; | |
2634 | sds->total_pwr += group->cpu_power; | |
2635 | ||
2636 | /* | |
2637 | * In case the child domain prefers tasks go to siblings | |
2638 | * first, lower the group capacity to one so that we'll try | |
2639 | * and move all the excess tasks away. | |
2640 | */ | |
2641 | if (prefer_sibling) | |
2642 | sgs.group_capacity = min(sgs.group_capacity, 1UL); | |
2643 | ||
2644 | if (local_group) { | |
2645 | sds->this_load = sgs.avg_load; | |
2646 | sds->this = group; | |
2647 | sds->this_nr_running = sgs.sum_nr_running; | |
2648 | sds->this_load_per_task = sgs.sum_weighted_load; | |
2649 | } else if (sgs.avg_load > sds->max_load && | |
2650 | (sgs.sum_nr_running > sgs.group_capacity || | |
2651 | sgs.group_imb)) { | |
2652 | sds->max_load = sgs.avg_load; | |
2653 | sds->busiest = group; | |
2654 | sds->busiest_nr_running = sgs.sum_nr_running; | |
2655 | sds->busiest_load_per_task = sgs.sum_weighted_load; | |
2656 | sds->group_imb = sgs.group_imb; | |
2657 | } | |
2658 | ||
2659 | update_sd_power_savings_stats(group, sds, local_group, &sgs); | |
2660 | group = group->next; | |
2661 | } while (group != sd->groups); | |
2662 | } | |
2663 | ||
2664 | /** | |
2665 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2666 | * amongst the groups of a sched_domain, during | |
2667 | * load balancing. | |
2668 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2669 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2670 | * @imbalance: Variable to store the imbalance. | |
2671 | */ | |
2672 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2673 | int this_cpu, unsigned long *imbalance) | |
2674 | { | |
2675 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2676 | unsigned int imbn = 2; | |
2677 | ||
2678 | if (sds->this_nr_running) { | |
2679 | sds->this_load_per_task /= sds->this_nr_running; | |
2680 | if (sds->busiest_load_per_task > | |
2681 | sds->this_load_per_task) | |
2682 | imbn = 1; | |
2683 | } else | |
2684 | sds->this_load_per_task = | |
2685 | cpu_avg_load_per_task(this_cpu); | |
2686 | ||
2687 | if (sds->max_load - sds->this_load + sds->busiest_load_per_task >= | |
2688 | sds->busiest_load_per_task * imbn) { | |
2689 | *imbalance = sds->busiest_load_per_task; | |
2690 | return; | |
2691 | } | |
2692 | ||
2693 | /* | |
2694 | * OK, we don't have enough imbalance to justify moving tasks, | |
2695 | * however we may be able to increase total CPU power used by | |
2696 | * moving them. | |
2697 | */ | |
2698 | ||
2699 | pwr_now += sds->busiest->cpu_power * | |
2700 | min(sds->busiest_load_per_task, sds->max_load); | |
2701 | pwr_now += sds->this->cpu_power * | |
2702 | min(sds->this_load_per_task, sds->this_load); | |
2703 | pwr_now /= SCHED_LOAD_SCALE; | |
2704 | ||
2705 | /* Amount of load we'd subtract */ | |
2706 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2707 | sds->busiest->cpu_power; | |
2708 | if (sds->max_load > tmp) | |
2709 | pwr_move += sds->busiest->cpu_power * | |
2710 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
2711 | ||
2712 | /* Amount of load we'd add */ | |
2713 | if (sds->max_load * sds->busiest->cpu_power < | |
2714 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
2715 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
2716 | sds->this->cpu_power; | |
2717 | else | |
2718 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2719 | sds->this->cpu_power; | |
2720 | pwr_move += sds->this->cpu_power * | |
2721 | min(sds->this_load_per_task, sds->this_load + tmp); | |
2722 | pwr_move /= SCHED_LOAD_SCALE; | |
2723 | ||
2724 | /* Move if we gain throughput */ | |
2725 | if (pwr_move > pwr_now) | |
2726 | *imbalance = sds->busiest_load_per_task; | |
2727 | } | |
2728 | ||
2729 | /** | |
2730 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
2731 | * groups of a given sched_domain during load balance. | |
2732 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
2733 | * @this_cpu: Cpu for which currently load balance is being performed. | |
2734 | * @imbalance: The variable to store the imbalance. | |
2735 | */ | |
2736 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
2737 | unsigned long *imbalance) | |
2738 | { | |
2739 | unsigned long max_pull; | |
2740 | /* | |
2741 | * In the presence of smp nice balancing, certain scenarios can have | |
2742 | * max load less than avg load(as we skip the groups at or below | |
2743 | * its cpu_power, while calculating max_load..) | |
2744 | */ | |
2745 | if (sds->max_load < sds->avg_load) { | |
2746 | *imbalance = 0; | |
2747 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2748 | } | |
2749 | ||
2750 | /* Don't want to pull so many tasks that a group would go idle */ | |
2751 | max_pull = min(sds->max_load - sds->avg_load, | |
2752 | sds->max_load - sds->busiest_load_per_task); | |
2753 | ||
2754 | /* How much load to actually move to equalise the imbalance */ | |
2755 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
2756 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
2757 | / SCHED_LOAD_SCALE; | |
2758 | ||
2759 | /* | |
2760 | * if *imbalance is less than the average load per runnable task | |
2761 | * there is no gaurantee that any tasks will be moved so we'll have | |
2762 | * a think about bumping its value to force at least one task to be | |
2763 | * moved | |
2764 | */ | |
2765 | if (*imbalance < sds->busiest_load_per_task) | |
2766 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2767 | ||
2768 | } | |
2769 | /******* find_busiest_group() helpers end here *********************/ | |
2770 | ||
2771 | /** | |
2772 | * find_busiest_group - Returns the busiest group within the sched_domain | |
2773 | * if there is an imbalance. If there isn't an imbalance, and | |
2774 | * the user has opted for power-savings, it returns a group whose | |
2775 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
2776 | * such a group exists. | |
2777 | * | |
2778 | * Also calculates the amount of weighted load which should be moved | |
2779 | * to restore balance. | |
2780 | * | |
2781 | * @sd: The sched_domain whose busiest group is to be returned. | |
2782 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
2783 | * @imbalance: Variable which stores amount of weighted load which should | |
2784 | * be moved to restore balance/put a group to idle. | |
2785 | * @idle: The idle status of this_cpu. | |
2786 | * @sd_idle: The idleness of sd | |
2787 | * @cpus: The set of CPUs under consideration for load-balancing. | |
2788 | * @balance: Pointer to a variable indicating if this_cpu | |
2789 | * is the appropriate cpu to perform load balancing at this_level. | |
2790 | * | |
2791 | * Returns: - the busiest group if imbalance exists. | |
2792 | * - If no imbalance and user has opted for power-savings balance, | |
2793 | * return the least loaded group whose CPUs can be | |
2794 | * put to idle by rebalancing its tasks onto our group. | |
2795 | */ | |
2796 | static struct sched_group * | |
2797 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
2798 | unsigned long *imbalance, enum cpu_idle_type idle, | |
2799 | int *sd_idle, const struct cpumask *cpus, int *balance) | |
2800 | { | |
2801 | struct sd_lb_stats sds; | |
2802 | ||
2803 | memset(&sds, 0, sizeof(sds)); | |
2804 | ||
2805 | /* | |
2806 | * Compute the various statistics relavent for load balancing at | |
2807 | * this level. | |
2808 | */ | |
2809 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, | |
2810 | balance, &sds); | |
2811 | ||
2812 | /* Cases where imbalance does not exist from POV of this_cpu */ | |
2813 | /* 1) this_cpu is not the appropriate cpu to perform load balancing | |
2814 | * at this level. | |
2815 | * 2) There is no busy sibling group to pull from. | |
2816 | * 3) This group is the busiest group. | |
2817 | * 4) This group is more busy than the avg busieness at this | |
2818 | * sched_domain. | |
2819 | * 5) The imbalance is within the specified limit. | |
2820 | * 6) Any rebalance would lead to ping-pong | |
2821 | */ | |
2822 | if (balance && !(*balance)) | |
2823 | goto ret; | |
2824 | ||
2825 | if (!sds.busiest || sds.busiest_nr_running == 0) | |
2826 | goto out_balanced; | |
2827 | ||
2828 | if (sds.this_load >= sds.max_load) | |
2829 | goto out_balanced; | |
2830 | ||
2831 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; | |
2832 | ||
2833 | if (sds.this_load >= sds.avg_load) | |
2834 | goto out_balanced; | |
2835 | ||
2836 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
2837 | goto out_balanced; | |
2838 | ||
2839 | sds.busiest_load_per_task /= sds.busiest_nr_running; | |
2840 | if (sds.group_imb) | |
2841 | sds.busiest_load_per_task = | |
2842 | min(sds.busiest_load_per_task, sds.avg_load); | |
2843 | ||
2844 | /* | |
2845 | * We're trying to get all the cpus to the average_load, so we don't | |
2846 | * want to push ourselves above the average load, nor do we wish to | |
2847 | * reduce the max loaded cpu below the average load, as either of these | |
2848 | * actions would just result in more rebalancing later, and ping-pong | |
2849 | * tasks around. Thus we look for the minimum possible imbalance. | |
2850 | * Negative imbalances (*we* are more loaded than anyone else) will | |
2851 | * be counted as no imbalance for these purposes -- we can't fix that | |
2852 | * by pulling tasks to us. Be careful of negative numbers as they'll | |
2853 | * appear as very large values with unsigned longs. | |
2854 | */ | |
2855 | if (sds.max_load <= sds.busiest_load_per_task) | |
2856 | goto out_balanced; | |
2857 | ||
2858 | /* Looks like there is an imbalance. Compute it */ | |
2859 | calculate_imbalance(&sds, this_cpu, imbalance); | |
2860 | return sds.busiest; | |
2861 | ||
2862 | out_balanced: | |
2863 | /* | |
2864 | * There is no obvious imbalance. But check if we can do some balancing | |
2865 | * to save power. | |
2866 | */ | |
2867 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
2868 | return sds.busiest; | |
2869 | ret: | |
2870 | *imbalance = 0; | |
2871 | return NULL; | |
2872 | } | |
2873 | ||
2874 | /* | |
2875 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2876 | */ | |
2877 | static struct rq * | |
2878 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, | |
2879 | unsigned long imbalance, const struct cpumask *cpus) | |
2880 | { | |
2881 | struct rq *busiest = NULL, *rq; | |
2882 | unsigned long max_load = 0; | |
2883 | int i; | |
2884 | ||
2885 | for_each_cpu(i, sched_group_cpus(group)) { | |
2886 | unsigned long power = power_of(i); | |
2887 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
2888 | unsigned long wl; | |
2889 | ||
2890 | if (!cpumask_test_cpu(i, cpus)) | |
2891 | continue; | |
2892 | ||
2893 | rq = cpu_rq(i); | |
2894 | wl = weighted_cpuload(i) * SCHED_LOAD_SCALE; | |
2895 | wl /= power; | |
2896 | ||
2897 | if (capacity && rq->nr_running == 1 && wl > imbalance) | |
2898 | continue; | |
2899 | ||
2900 | if (wl > max_load) { | |
2901 | max_load = wl; | |
2902 | busiest = rq; | |
2903 | } | |
2904 | } | |
2905 | ||
2906 | return busiest; | |
2907 | } | |
2908 | ||
2909 | /* | |
2910 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2911 | * so long as it is large enough. | |
2912 | */ | |
2913 | #define MAX_PINNED_INTERVAL 512 | |
2914 | ||
2915 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
2916 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
2917 | ||
2918 | /* | |
2919 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
2920 | * tasks if there is an imbalance. | |
2921 | */ | |
2922 | static int load_balance(int this_cpu, struct rq *this_rq, | |
2923 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2924 | int *balance) | |
2925 | { | |
2926 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
2927 | struct sched_group *group; | |
2928 | unsigned long imbalance; | |
2929 | struct rq *busiest; | |
2930 | unsigned long flags; | |
2931 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
2932 | ||
2933 | cpumask_copy(cpus, cpu_active_mask); | |
2934 | ||
2935 | /* | |
2936 | * When power savings policy is enabled for the parent domain, idle | |
2937 | * sibling can pick up load irrespective of busy siblings. In this case, | |
2938 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
2939 | * portraying it as CPU_NOT_IDLE. | |
2940 | */ | |
2941 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
2942 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
2943 | sd_idle = 1; | |
2944 | ||
2945 | schedstat_inc(sd, lb_count[idle]); | |
2946 | ||
2947 | redo: | |
2948 | update_shares(sd); | |
2949 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
2950 | cpus, balance); | |
2951 | ||
2952 | if (*balance == 0) | |
2953 | goto out_balanced; | |
2954 | ||
2955 | if (!group) { | |
2956 | schedstat_inc(sd, lb_nobusyg[idle]); | |
2957 | goto out_balanced; | |
2958 | } | |
2959 | ||
2960 | busiest = find_busiest_queue(group, idle, imbalance, cpus); | |
2961 | if (!busiest) { | |
2962 | schedstat_inc(sd, lb_nobusyq[idle]); | |
2963 | goto out_balanced; | |
2964 | } | |
2965 | ||
2966 | BUG_ON(busiest == this_rq); | |
2967 | ||
2968 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
2969 | ||
2970 | ld_moved = 0; | |
2971 | if (busiest->nr_running > 1) { | |
2972 | /* | |
2973 | * Attempt to move tasks. If find_busiest_group has found | |
2974 | * an imbalance but busiest->nr_running <= 1, the group is | |
2975 | * still unbalanced. ld_moved simply stays zero, so it is | |
2976 | * correctly treated as an imbalance. | |
2977 | */ | |
2978 | local_irq_save(flags); | |
2979 | double_rq_lock(this_rq, busiest); | |
2980 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
2981 | imbalance, sd, idle, &all_pinned); | |
2982 | double_rq_unlock(this_rq, busiest); | |
2983 | local_irq_restore(flags); | |
2984 | ||
2985 | /* | |
2986 | * some other cpu did the load balance for us. | |
2987 | */ | |
2988 | if (ld_moved && this_cpu != smp_processor_id()) | |
2989 | resched_cpu(this_cpu); | |
2990 | ||
2991 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
2992 | if (unlikely(all_pinned)) { | |
2993 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
2994 | if (!cpumask_empty(cpus)) | |
2995 | goto redo; | |
2996 | goto out_balanced; | |
2997 | } | |
2998 | } | |
2999 | ||
3000 | if (!ld_moved) { | |
3001 | schedstat_inc(sd, lb_failed[idle]); | |
3002 | sd->nr_balance_failed++; | |
3003 | ||
3004 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
3005 | ||
3006 | raw_spin_lock_irqsave(&busiest->lock, flags); | |
3007 | ||
3008 | /* don't kick the migration_thread, if the curr | |
3009 | * task on busiest cpu can't be moved to this_cpu | |
3010 | */ | |
3011 | if (!cpumask_test_cpu(this_cpu, | |
3012 | &busiest->curr->cpus_allowed)) { | |
3013 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3014 | flags); | |
3015 | all_pinned = 1; | |
3016 | goto out_one_pinned; | |
3017 | } | |
3018 | ||
3019 | if (!busiest->active_balance) { | |
3020 | busiest->active_balance = 1; | |
3021 | busiest->push_cpu = this_cpu; | |
3022 | active_balance = 1; | |
3023 | } | |
3024 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
3025 | if (active_balance) | |
3026 | wake_up_process(busiest->migration_thread); | |
3027 | ||
3028 | /* | |
3029 | * We've kicked active balancing, reset the failure | |
3030 | * counter. | |
3031 | */ | |
3032 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3033 | } | |
3034 | } else | |
3035 | sd->nr_balance_failed = 0; | |
3036 | ||
3037 | if (likely(!active_balance)) { | |
3038 | /* We were unbalanced, so reset the balancing interval */ | |
3039 | sd->balance_interval = sd->min_interval; | |
3040 | } else { | |
3041 | /* | |
3042 | * If we've begun active balancing, start to back off. This | |
3043 | * case may not be covered by the all_pinned logic if there | |
3044 | * is only 1 task on the busy runqueue (because we don't call | |
3045 | * move_tasks). | |
3046 | */ | |
3047 | if (sd->balance_interval < sd->max_interval) | |
3048 | sd->balance_interval *= 2; | |
3049 | } | |
3050 | ||
3051 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3052 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3053 | ld_moved = -1; | |
3054 | ||
3055 | goto out; | |
3056 | ||
3057 | out_balanced: | |
3058 | schedstat_inc(sd, lb_balanced[idle]); | |
3059 | ||
3060 | sd->nr_balance_failed = 0; | |
3061 | ||
3062 | out_one_pinned: | |
3063 | /* tune up the balancing interval */ | |
3064 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3065 | (sd->balance_interval < sd->max_interval)) | |
3066 | sd->balance_interval *= 2; | |
3067 | ||
3068 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3069 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3070 | ld_moved = -1; | |
3071 | else | |
3072 | ld_moved = 0; | |
3073 | out: | |
3074 | if (ld_moved) | |
3075 | update_shares(sd); | |
3076 | return ld_moved; | |
3077 | } | |
3078 | ||
3079 | /* | |
3080 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3081 | * tasks if there is an imbalance. | |
3082 | * | |
3083 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). | |
3084 | * this_rq is locked. | |
3085 | */ | |
3086 | static int | |
3087 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd) | |
3088 | { | |
3089 | struct sched_group *group; | |
3090 | struct rq *busiest = NULL; | |
3091 | unsigned long imbalance; | |
3092 | int ld_moved = 0; | |
3093 | int sd_idle = 0; | |
3094 | int all_pinned = 0; | |
3095 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3096 | ||
3097 | cpumask_copy(cpus, cpu_active_mask); | |
3098 | ||
3099 | /* | |
3100 | * When power savings policy is enabled for the parent domain, idle | |
3101 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3102 | * let the state of idle sibling percolate up as IDLE, instead of | |
3103 | * portraying it as CPU_NOT_IDLE. | |
3104 | */ | |
3105 | if (sd->flags & SD_SHARE_CPUPOWER && | |
3106 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3107 | sd_idle = 1; | |
3108 | ||
3109 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); | |
3110 | redo: | |
3111 | update_shares_locked(this_rq, sd); | |
3112 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, | |
3113 | &sd_idle, cpus, NULL); | |
3114 | if (!group) { | |
3115 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); | |
3116 | goto out_balanced; | |
3117 | } | |
3118 | ||
3119 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); | |
3120 | if (!busiest) { | |
3121 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); | |
3122 | goto out_balanced; | |
3123 | } | |
3124 | ||
3125 | BUG_ON(busiest == this_rq); | |
3126 | ||
3127 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); | |
3128 | ||
3129 | ld_moved = 0; | |
3130 | if (busiest->nr_running > 1) { | |
3131 | /* Attempt to move tasks */ | |
3132 | double_lock_balance(this_rq, busiest); | |
3133 | /* this_rq->clock is already updated */ | |
3134 | update_rq_clock(busiest); | |
3135 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3136 | imbalance, sd, CPU_NEWLY_IDLE, | |
3137 | &all_pinned); | |
3138 | double_unlock_balance(this_rq, busiest); | |
3139 | ||
3140 | if (unlikely(all_pinned)) { | |
3141 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3142 | if (!cpumask_empty(cpus)) | |
3143 | goto redo; | |
3144 | } | |
3145 | } | |
3146 | ||
3147 | if (!ld_moved) { | |
3148 | int active_balance = 0; | |
3149 | ||
3150 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); | |
3151 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3152 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3153 | return -1; | |
3154 | ||
3155 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) | |
3156 | return -1; | |
3157 | ||
3158 | if (sd->nr_balance_failed++ < 2) | |
3159 | return -1; | |
3160 | ||
3161 | /* | |
3162 | * The only task running in a non-idle cpu can be moved to this | |
3163 | * cpu in an attempt to completely freeup the other CPU | |
3164 | * package. The same method used to move task in load_balance() | |
3165 | * have been extended for load_balance_newidle() to speedup | |
3166 | * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2) | |
3167 | * | |
3168 | * The package power saving logic comes from | |
3169 | * find_busiest_group(). If there are no imbalance, then | |
3170 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3171 | * f_b_g() will select a group from which a running task may be | |
3172 | * pulled to this cpu in order to make the other package idle. | |
3173 | * If there is no opportunity to make a package idle and if | |
3174 | * there are no imbalance, then f_b_g() will return NULL and no | |
3175 | * action will be taken in load_balance_newidle(). | |
3176 | * | |
3177 | * Under normal task pull operation due to imbalance, there | |
3178 | * will be more than one task in the source run queue and | |
3179 | * move_tasks() will succeed. ld_moved will be true and this | |
3180 | * active balance code will not be triggered. | |
3181 | */ | |
3182 | ||
3183 | /* Lock busiest in correct order while this_rq is held */ | |
3184 | double_lock_balance(this_rq, busiest); | |
3185 | ||
3186 | /* | |
3187 | * don't kick the migration_thread, if the curr | |
3188 | * task on busiest cpu can't be moved to this_cpu | |
3189 | */ | |
3190 | if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) { | |
3191 | double_unlock_balance(this_rq, busiest); | |
3192 | all_pinned = 1; | |
3193 | return ld_moved; | |
3194 | } | |
3195 | ||
3196 | if (!busiest->active_balance) { | |
3197 | busiest->active_balance = 1; | |
3198 | busiest->push_cpu = this_cpu; | |
3199 | active_balance = 1; | |
3200 | } | |
3201 | ||
3202 | double_unlock_balance(this_rq, busiest); | |
3203 | /* | |
3204 | * Should not call ttwu while holding a rq->lock | |
3205 | */ | |
3206 | raw_spin_unlock(&this_rq->lock); | |
3207 | if (active_balance) | |
3208 | wake_up_process(busiest->migration_thread); | |
3209 | raw_spin_lock(&this_rq->lock); | |
3210 | ||
3211 | } else | |
3212 | sd->nr_balance_failed = 0; | |
3213 | ||
3214 | update_shares_locked(this_rq, sd); | |
3215 | return ld_moved; | |
3216 | ||
3217 | out_balanced: | |
3218 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); | |
3219 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3220 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3221 | return -1; | |
3222 | sd->nr_balance_failed = 0; | |
3223 | ||
3224 | return 0; | |
3225 | } | |
3226 | ||
3227 | /* | |
3228 | * idle_balance is called by schedule() if this_cpu is about to become | |
3229 | * idle. Attempts to pull tasks from other CPUs. | |
3230 | */ | |
3231 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3232 | { | |
3233 | struct sched_domain *sd; | |
3234 | int pulled_task = 0; | |
3235 | unsigned long next_balance = jiffies + HZ; | |
3236 | ||
3237 | this_rq->idle_stamp = this_rq->clock; | |
3238 | ||
3239 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3240 | return; | |
3241 | ||
3242 | for_each_domain(this_cpu, sd) { | |
3243 | unsigned long interval; | |
3244 | ||
3245 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3246 | continue; | |
3247 | ||
3248 | if (sd->flags & SD_BALANCE_NEWIDLE) | |
3249 | /* If we've pulled tasks over stop searching: */ | |
3250 | pulled_task = load_balance_newidle(this_cpu, this_rq, | |
3251 | sd); | |
3252 | ||
3253 | interval = msecs_to_jiffies(sd->balance_interval); | |
3254 | if (time_after(next_balance, sd->last_balance + interval)) | |
3255 | next_balance = sd->last_balance + interval; | |
3256 | if (pulled_task) { | |
3257 | this_rq->idle_stamp = 0; | |
3258 | break; | |
3259 | } | |
3260 | } | |
3261 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { | |
3262 | /* | |
3263 | * We are going idle. next_balance may be set based on | |
3264 | * a busy processor. So reset next_balance. | |
3265 | */ | |
3266 | this_rq->next_balance = next_balance; | |
3267 | } | |
3268 | } | |
3269 | ||
3270 | /* | |
3271 | * active_load_balance is run by migration threads. It pushes running tasks | |
3272 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
3273 | * running on each physical CPU where possible, and avoids physical / | |
3274 | * logical imbalances. | |
3275 | * | |
3276 | * Called with busiest_rq locked. | |
3277 | */ | |
3278 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) | |
3279 | { | |
3280 | int target_cpu = busiest_rq->push_cpu; | |
3281 | struct sched_domain *sd; | |
3282 | struct rq *target_rq; | |
3283 | ||
3284 | /* Is there any task to move? */ | |
3285 | if (busiest_rq->nr_running <= 1) | |
3286 | return; | |
3287 | ||
3288 | target_rq = cpu_rq(target_cpu); | |
3289 | ||
3290 | /* | |
3291 | * This condition is "impossible", if it occurs | |
3292 | * we need to fix it. Originally reported by | |
3293 | * Bjorn Helgaas on a 128-cpu setup. | |
3294 | */ | |
3295 | BUG_ON(busiest_rq == target_rq); | |
3296 | ||
3297 | /* move a task from busiest_rq to target_rq */ | |
3298 | double_lock_balance(busiest_rq, target_rq); | |
3299 | update_rq_clock(busiest_rq); | |
3300 | update_rq_clock(target_rq); | |
3301 | ||
3302 | /* Search for an sd spanning us and the target CPU. */ | |
3303 | for_each_domain(target_cpu, sd) { | |
3304 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3305 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3306 | break; | |
3307 | } | |
3308 | ||
3309 | if (likely(sd)) { | |
3310 | schedstat_inc(sd, alb_count); | |
3311 | ||
3312 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3313 | sd, CPU_IDLE)) | |
3314 | schedstat_inc(sd, alb_pushed); | |
3315 | else | |
3316 | schedstat_inc(sd, alb_failed); | |
3317 | } | |
3318 | double_unlock_balance(busiest_rq, target_rq); | |
3319 | } | |
3320 | ||
3321 | #ifdef CONFIG_NO_HZ | |
3322 | static struct { | |
3323 | atomic_t load_balancer; | |
3324 | cpumask_var_t cpu_mask; | |
3325 | cpumask_var_t ilb_grp_nohz_mask; | |
3326 | } nohz ____cacheline_aligned = { | |
3327 | .load_balancer = ATOMIC_INIT(-1), | |
3328 | }; | |
3329 | ||
3330 | int get_nohz_load_balancer(void) | |
3331 | { | |
3332 | return atomic_read(&nohz.load_balancer); | |
3333 | } | |
3334 | ||
3335 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3336 | /** | |
3337 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3338 | * @cpu: The cpu whose lowest level of sched domain is to | |
3339 | * be returned. | |
3340 | * @flag: The flag to check for the lowest sched_domain | |
3341 | * for the given cpu. | |
3342 | * | |
3343 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3344 | */ | |
3345 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3346 | { | |
3347 | struct sched_domain *sd; | |
3348 | ||
3349 | for_each_domain(cpu, sd) | |
3350 | if (sd && (sd->flags & flag)) | |
3351 | break; | |
3352 | ||
3353 | return sd; | |
3354 | } | |
3355 | ||
3356 | /** | |
3357 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3358 | * @cpu: The cpu whose domains we're iterating over. | |
3359 | * @sd: variable holding the value of the power_savings_sd | |
3360 | * for cpu. | |
3361 | * @flag: The flag to filter the sched_domains to be iterated. | |
3362 | * | |
3363 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3364 | * set, starting from the lowest sched_domain to the highest. | |
3365 | */ | |
3366 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3367 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3368 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3369 | ||
3370 | /** | |
3371 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3372 | * @ilb_group: group to be checked for semi-idleness | |
3373 | * | |
3374 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3375 | * | |
3376 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3377 | * and atleast one non-idle CPU. This helper function checks if the given | |
3378 | * sched_group is semi-idle or not. | |
3379 | */ | |
3380 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3381 | { | |
3382 | cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask, | |
3383 | sched_group_cpus(ilb_group)); | |
3384 | ||
3385 | /* | |
3386 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3387 | * and atleast one idle cpu. | |
3388 | */ | |
3389 | if (cpumask_empty(nohz.ilb_grp_nohz_mask)) | |
3390 | return 0; | |
3391 | ||
3392 | if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group))) | |
3393 | return 0; | |
3394 | ||
3395 | return 1; | |
3396 | } | |
3397 | /** | |
3398 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3399 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3400 | * | |
3401 | * Returns: Returns the id of the idle load balancer if it exists, | |
3402 | * Else, returns >= nr_cpu_ids. | |
3403 | * | |
3404 | * This algorithm picks the idle load balancer such that it belongs to a | |
3405 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3406 | * completely idle packages/cores just for the purpose of idle load balancing | |
3407 | * when there are other idle cpu's which are better suited for that job. | |
3408 | */ | |
3409 | static int find_new_ilb(int cpu) | |
3410 | { | |
3411 | struct sched_domain *sd; | |
3412 | struct sched_group *ilb_group; | |
3413 | ||
3414 | /* | |
3415 | * Have idle load balancer selection from semi-idle packages only | |
3416 | * when power-aware load balancing is enabled | |
3417 | */ | |
3418 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3419 | goto out_done; | |
3420 | ||
3421 | /* | |
3422 | * Optimize for the case when we have no idle CPUs or only one | |
3423 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3424 | */ | |
3425 | if (cpumask_weight(nohz.cpu_mask) < 2) | |
3426 | goto out_done; | |
3427 | ||
3428 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
3429 | ilb_group = sd->groups; | |
3430 | ||
3431 | do { | |
3432 | if (is_semi_idle_group(ilb_group)) | |
3433 | return cpumask_first(nohz.ilb_grp_nohz_mask); | |
3434 | ||
3435 | ilb_group = ilb_group->next; | |
3436 | ||
3437 | } while (ilb_group != sd->groups); | |
3438 | } | |
3439 | ||
3440 | out_done: | |
3441 | return cpumask_first(nohz.cpu_mask); | |
3442 | } | |
3443 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3444 | static inline int find_new_ilb(int call_cpu) | |
3445 | { | |
3446 | return cpumask_first(nohz.cpu_mask); | |
3447 | } | |
3448 | #endif | |
3449 | ||
3450 | /* | |
3451 | * This routine will try to nominate the ilb (idle load balancing) | |
3452 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
3453 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
3454 | * go into this tickless mode, then there will be no ilb owner (as there is | |
3455 | * no need for one) and all the cpus will sleep till the next wakeup event | |
3456 | * arrives... | |
3457 | * | |
3458 | * For the ilb owner, tick is not stopped. And this tick will be used | |
3459 | * for idle load balancing. ilb owner will still be part of | |
3460 | * nohz.cpu_mask.. | |
3461 | * | |
3462 | * While stopping the tick, this cpu will become the ilb owner if there | |
3463 | * is no other owner. And will be the owner till that cpu becomes busy | |
3464 | * or if all cpus in the system stop their ticks at which point | |
3465 | * there is no need for ilb owner. | |
3466 | * | |
3467 | * When the ilb owner becomes busy, it nominates another owner, during the | |
3468 | * next busy scheduler_tick() | |
3469 | */ | |
3470 | int select_nohz_load_balancer(int stop_tick) | |
3471 | { | |
3472 | int cpu = smp_processor_id(); | |
3473 | ||
3474 | if (stop_tick) { | |
3475 | cpu_rq(cpu)->in_nohz_recently = 1; | |
3476 | ||
3477 | if (!cpu_active(cpu)) { | |
3478 | if (atomic_read(&nohz.load_balancer) != cpu) | |
3479 | return 0; | |
3480 | ||
3481 | /* | |
3482 | * If we are going offline and still the leader, | |
3483 | * give up! | |
3484 | */ | |
3485 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3486 | BUG(); | |
3487 | ||
3488 | return 0; | |
3489 | } | |
3490 | ||
3491 | cpumask_set_cpu(cpu, nohz.cpu_mask); | |
3492 | ||
3493 | /* time for ilb owner also to sleep */ | |
3494 | if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) { | |
3495 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3496 | atomic_set(&nohz.load_balancer, -1); | |
3497 | return 0; | |
3498 | } | |
3499 | ||
3500 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3501 | /* make me the ilb owner */ | |
3502 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
3503 | return 1; | |
3504 | } else if (atomic_read(&nohz.load_balancer) == cpu) { | |
3505 | int new_ilb; | |
3506 | ||
3507 | if (!(sched_smt_power_savings || | |
3508 | sched_mc_power_savings)) | |
3509 | return 1; | |
3510 | /* | |
3511 | * Check to see if there is a more power-efficient | |
3512 | * ilb. | |
3513 | */ | |
3514 | new_ilb = find_new_ilb(cpu); | |
3515 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
3516 | atomic_set(&nohz.load_balancer, -1); | |
3517 | resched_cpu(new_ilb); | |
3518 | return 0; | |
3519 | } | |
3520 | return 1; | |
3521 | } | |
3522 | } else { | |
3523 | if (!cpumask_test_cpu(cpu, nohz.cpu_mask)) | |
3524 | return 0; | |
3525 | ||
3526 | cpumask_clear_cpu(cpu, nohz.cpu_mask); | |
3527 | ||
3528 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3529 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3530 | BUG(); | |
3531 | } | |
3532 | return 0; | |
3533 | } | |
3534 | #endif | |
3535 | ||
3536 | static DEFINE_SPINLOCK(balancing); | |
3537 | ||
3538 | /* | |
3539 | * It checks each scheduling domain to see if it is due to be balanced, | |
3540 | * and initiates a balancing operation if so. | |
3541 | * | |
3542 | * Balancing parameters are set up in arch_init_sched_domains. | |
3543 | */ | |
3544 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3545 | { | |
3546 | int balance = 1; | |
3547 | struct rq *rq = cpu_rq(cpu); | |
3548 | unsigned long interval; | |
3549 | struct sched_domain *sd; | |
3550 | /* Earliest time when we have to do rebalance again */ | |
3551 | unsigned long next_balance = jiffies + 60*HZ; | |
3552 | int update_next_balance = 0; | |
3553 | int need_serialize; | |
3554 | ||
3555 | for_each_domain(cpu, sd) { | |
3556 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3557 | continue; | |
3558 | ||
3559 | interval = sd->balance_interval; | |
3560 | if (idle != CPU_IDLE) | |
3561 | interval *= sd->busy_factor; | |
3562 | ||
3563 | /* scale ms to jiffies */ | |
3564 | interval = msecs_to_jiffies(interval); | |
3565 | if (unlikely(!interval)) | |
3566 | interval = 1; | |
3567 | if (interval > HZ*NR_CPUS/10) | |
3568 | interval = HZ*NR_CPUS/10; | |
3569 | ||
3570 | need_serialize = sd->flags & SD_SERIALIZE; | |
3571 | ||
3572 | if (need_serialize) { | |
3573 | if (!spin_trylock(&balancing)) | |
3574 | goto out; | |
3575 | } | |
3576 | ||
3577 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3578 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3579 | /* | |
3580 | * We've pulled tasks over so either we're no | |
3581 | * longer idle, or one of our SMT siblings is | |
3582 | * not idle. | |
3583 | */ | |
3584 | idle = CPU_NOT_IDLE; | |
3585 | } | |
3586 | sd->last_balance = jiffies; | |
3587 | } | |
3588 | if (need_serialize) | |
3589 | spin_unlock(&balancing); | |
3590 | out: | |
3591 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3592 | next_balance = sd->last_balance + interval; | |
3593 | update_next_balance = 1; | |
3594 | } | |
3595 | ||
3596 | /* | |
3597 | * Stop the load balance at this level. There is another | |
3598 | * CPU in our sched group which is doing load balancing more | |
3599 | * actively. | |
3600 | */ | |
3601 | if (!balance) | |
3602 | break; | |
3603 | } | |
3604 | ||
3605 | /* | |
3606 | * next_balance will be updated only when there is a need. | |
3607 | * When the cpu is attached to null domain for ex, it will not be | |
3608 | * updated. | |
3609 | */ | |
3610 | if (likely(update_next_balance)) | |
3611 | rq->next_balance = next_balance; | |
3612 | } | |
3613 | ||
3614 | /* | |
3615 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
3616 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
3617 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
3618 | */ | |
3619 | static void run_rebalance_domains(struct softirq_action *h) | |
3620 | { | |
3621 | int this_cpu = smp_processor_id(); | |
3622 | struct rq *this_rq = cpu_rq(this_cpu); | |
3623 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
3624 | CPU_IDLE : CPU_NOT_IDLE; | |
3625 | ||
3626 | rebalance_domains(this_cpu, idle); | |
3627 | ||
3628 | #ifdef CONFIG_NO_HZ | |
3629 | /* | |
3630 | * If this cpu is the owner for idle load balancing, then do the | |
3631 | * balancing on behalf of the other idle cpus whose ticks are | |
3632 | * stopped. | |
3633 | */ | |
3634 | if (this_rq->idle_at_tick && | |
3635 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
3636 | struct rq *rq; | |
3637 | int balance_cpu; | |
3638 | ||
3639 | for_each_cpu(balance_cpu, nohz.cpu_mask) { | |
3640 | if (balance_cpu == this_cpu) | |
3641 | continue; | |
3642 | ||
3643 | /* | |
3644 | * If this cpu gets work to do, stop the load balancing | |
3645 | * work being done for other cpus. Next load | |
3646 | * balancing owner will pick it up. | |
3647 | */ | |
3648 | if (need_resched()) | |
3649 | break; | |
3650 | ||
3651 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3652 | ||
3653 | rq = cpu_rq(balance_cpu); | |
3654 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3655 | this_rq->next_balance = rq->next_balance; | |
3656 | } | |
3657 | } | |
3658 | #endif | |
3659 | } | |
3660 | ||
3661 | static inline int on_null_domain(int cpu) | |
3662 | { | |
3663 | return !rcu_dereference(cpu_rq(cpu)->sd); | |
3664 | } | |
3665 | ||
3666 | /* | |
3667 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
3668 | * | |
3669 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
3670 | * idle load balancing owner or decide to stop the periodic load balancing, | |
3671 | * if the whole system is idle. | |
3672 | */ | |
3673 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
3674 | { | |
3675 | #ifdef CONFIG_NO_HZ | |
3676 | /* | |
3677 | * If we were in the nohz mode recently and busy at the current | |
3678 | * scheduler tick, then check if we need to nominate new idle | |
3679 | * load balancer. | |
3680 | */ | |
3681 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
3682 | rq->in_nohz_recently = 0; | |
3683 | ||
3684 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
3685 | cpumask_clear_cpu(cpu, nohz.cpu_mask); | |
3686 | atomic_set(&nohz.load_balancer, -1); | |
3687 | } | |
3688 | ||
3689 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3690 | int ilb = find_new_ilb(cpu); | |
3691 | ||
3692 | if (ilb < nr_cpu_ids) | |
3693 | resched_cpu(ilb); | |
3694 | } | |
3695 | } | |
3696 | ||
3697 | /* | |
3698 | * If this cpu is idle and doing idle load balancing for all the | |
3699 | * cpus with ticks stopped, is it time for that to stop? | |
3700 | */ | |
3701 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
3702 | cpumask_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3703 | resched_cpu(cpu); | |
3704 | return; | |
3705 | } | |
3706 | ||
3707 | /* | |
3708 | * If this cpu is idle and the idle load balancing is done by | |
3709 | * someone else, then no need raise the SCHED_SOFTIRQ | |
3710 | */ | |
3711 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
3712 | cpumask_test_cpu(cpu, nohz.cpu_mask)) | |
3713 | return; | |
3714 | #endif | |
3715 | /* Don't need to rebalance while attached to NULL domain */ | |
3716 | if (time_after_eq(jiffies, rq->next_balance) && | |
3717 | likely(!on_null_domain(cpu))) | |
3718 | raise_softirq(SCHED_SOFTIRQ); | |
3719 | } | |
3720 | ||
0bcdcf28 CE |
3721 | static void rq_online_fair(struct rq *rq) |
3722 | { | |
3723 | update_sysctl(); | |
3724 | } | |
3725 | ||
3726 | static void rq_offline_fair(struct rq *rq) | |
3727 | { | |
3728 | update_sysctl(); | |
3729 | } | |
3730 | ||
1e3c88bd PZ |
3731 | #else /* CONFIG_SMP */ |
3732 | ||
3733 | /* | |
3734 | * on UP we do not need to balance between CPUs: | |
3735 | */ | |
3736 | static inline void idle_balance(int cpu, struct rq *rq) | |
3737 | { | |
3738 | } | |
3739 | ||
55e12e5e | 3740 | #endif /* CONFIG_SMP */ |
e1d1484f | 3741 | |
bf0f6f24 IM |
3742 | /* |
3743 | * scheduler tick hitting a task of our scheduling class: | |
3744 | */ | |
8f4d37ec | 3745 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
3746 | { |
3747 | struct cfs_rq *cfs_rq; | |
3748 | struct sched_entity *se = &curr->se; | |
3749 | ||
3750 | for_each_sched_entity(se) { | |
3751 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 3752 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
3753 | } |
3754 | } | |
3755 | ||
3756 | /* | |
cd29fe6f PZ |
3757 | * called on fork with the child task as argument from the parent's context |
3758 | * - child not yet on the tasklist | |
3759 | * - preemption disabled | |
bf0f6f24 | 3760 | */ |
cd29fe6f | 3761 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 3762 | { |
cd29fe6f | 3763 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 3764 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 3765 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
3766 | struct rq *rq = this_rq(); |
3767 | unsigned long flags; | |
3768 | ||
05fa785c | 3769 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 3770 | |
cd29fe6f PZ |
3771 | if (unlikely(task_cpu(p) != this_cpu)) |
3772 | __set_task_cpu(p, this_cpu); | |
bf0f6f24 | 3773 | |
7109c442 | 3774 | update_curr(cfs_rq); |
cd29fe6f | 3775 | |
b5d9d734 MG |
3776 | if (curr) |
3777 | se->vruntime = curr->vruntime; | |
aeb73b04 | 3778 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 3779 | |
cd29fe6f | 3780 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 3781 | /* |
edcb60a3 IM |
3782 | * Upon rescheduling, sched_class::put_prev_task() will place |
3783 | * 'current' within the tree based on its new key value. | |
3784 | */ | |
4d78e7b6 | 3785 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 3786 | resched_task(rq->curr); |
4d78e7b6 | 3787 | } |
bf0f6f24 | 3788 | |
88ec22d3 PZ |
3789 | se->vruntime -= cfs_rq->min_vruntime; |
3790 | ||
05fa785c | 3791 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
3792 | } |
3793 | ||
cb469845 SR |
3794 | /* |
3795 | * Priority of the task has changed. Check to see if we preempt | |
3796 | * the current task. | |
3797 | */ | |
3798 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
3799 | int oldprio, int running) | |
3800 | { | |
3801 | /* | |
3802 | * Reschedule if we are currently running on this runqueue and | |
3803 | * our priority decreased, or if we are not currently running on | |
3804 | * this runqueue and our priority is higher than the current's | |
3805 | */ | |
3806 | if (running) { | |
3807 | if (p->prio > oldprio) | |
3808 | resched_task(rq->curr); | |
3809 | } else | |
15afe09b | 3810 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
3811 | } |
3812 | ||
3813 | /* | |
3814 | * We switched to the sched_fair class. | |
3815 | */ | |
3816 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
3817 | int running) | |
3818 | { | |
3819 | /* | |
3820 | * We were most likely switched from sched_rt, so | |
3821 | * kick off the schedule if running, otherwise just see | |
3822 | * if we can still preempt the current task. | |
3823 | */ | |
3824 | if (running) | |
3825 | resched_task(rq->curr); | |
3826 | else | |
15afe09b | 3827 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
3828 | } |
3829 | ||
83b699ed SV |
3830 | /* Account for a task changing its policy or group. |
3831 | * | |
3832 | * This routine is mostly called to set cfs_rq->curr field when a task | |
3833 | * migrates between groups/classes. | |
3834 | */ | |
3835 | static void set_curr_task_fair(struct rq *rq) | |
3836 | { | |
3837 | struct sched_entity *se = &rq->curr->se; | |
3838 | ||
3839 | for_each_sched_entity(se) | |
3840 | set_next_entity(cfs_rq_of(se), se); | |
3841 | } | |
3842 | ||
810b3817 | 3843 | #ifdef CONFIG_FAIR_GROUP_SCHED |
88ec22d3 | 3844 | static void moved_group_fair(struct task_struct *p, int on_rq) |
810b3817 PZ |
3845 | { |
3846 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
3847 | ||
3848 | update_curr(cfs_rq); | |
88ec22d3 PZ |
3849 | if (!on_rq) |
3850 | place_entity(cfs_rq, &p->se, 1); | |
810b3817 PZ |
3851 | } |
3852 | #endif | |
3853 | ||
6d686f45 | 3854 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
3855 | { |
3856 | struct sched_entity *se = &task->se; | |
0d721cea PW |
3857 | unsigned int rr_interval = 0; |
3858 | ||
3859 | /* | |
3860 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
3861 | * idle runqueue: | |
3862 | */ | |
0d721cea PW |
3863 | if (rq->cfs.load.weight) |
3864 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
3865 | |
3866 | return rr_interval; | |
3867 | } | |
3868 | ||
bf0f6f24 IM |
3869 | /* |
3870 | * All the scheduling class methods: | |
3871 | */ | |
5522d5d5 IM |
3872 | static const struct sched_class fair_sched_class = { |
3873 | .next = &idle_sched_class, | |
bf0f6f24 IM |
3874 | .enqueue_task = enqueue_task_fair, |
3875 | .dequeue_task = dequeue_task_fair, | |
3876 | .yield_task = yield_task_fair, | |
3877 | ||
2e09bf55 | 3878 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
3879 | |
3880 | .pick_next_task = pick_next_task_fair, | |
3881 | .put_prev_task = put_prev_task_fair, | |
3882 | ||
681f3e68 | 3883 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
3884 | .select_task_rq = select_task_rq_fair, |
3885 | ||
0bcdcf28 CE |
3886 | .rq_online = rq_online_fair, |
3887 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
3888 | |
3889 | .task_waking = task_waking_fair, | |
681f3e68 | 3890 | #endif |
bf0f6f24 | 3891 | |
83b699ed | 3892 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 3893 | .task_tick = task_tick_fair, |
cd29fe6f | 3894 | .task_fork = task_fork_fair, |
cb469845 SR |
3895 | |
3896 | .prio_changed = prio_changed_fair, | |
3897 | .switched_to = switched_to_fair, | |
810b3817 | 3898 | |
0d721cea PW |
3899 | .get_rr_interval = get_rr_interval_fair, |
3900 | ||
810b3817 PZ |
3901 | #ifdef CONFIG_FAIR_GROUP_SCHED |
3902 | .moved_group = moved_group_fair, | |
3903 | #endif | |
bf0f6f24 IM |
3904 | }; |
3905 | ||
3906 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 3907 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 3908 | { |
bf0f6f24 IM |
3909 | struct cfs_rq *cfs_rq; |
3910 | ||
5973e5b9 | 3911 | rcu_read_lock(); |
c3b64f1e | 3912 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 3913 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 3914 | rcu_read_unlock(); |
bf0f6f24 IM |
3915 | } |
3916 | #endif |