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> |
3436ae12 | 25 | #include <linux/cpumask.h> |
9745512c | 26 | |
bf0f6f24 | 27 | /* |
21805085 | 28 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 29 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 30 | * |
21805085 | 31 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
32 | * 'timeslice length' - timeslices in CFS are of variable length |
33 | * and have no persistent notion like in traditional, time-slice | |
34 | * based scheduling concepts. | |
bf0f6f24 | 35 | * |
d274a4ce IM |
36 | * (to see the precise effective timeslice length of your workload, |
37 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 38 | */ |
21406928 MG |
39 | unsigned int sysctl_sched_latency = 6000000ULL; |
40 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 41 | |
1983a922 CE |
42 | /* |
43 | * The initial- and re-scaling of tunables is configurable | |
44 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
45 | * | |
46 | * Options are: | |
47 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
48 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
49 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
50 | */ | |
51 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
52 | = SCHED_TUNABLESCALING_LOG; | |
53 | ||
2bd8e6d4 | 54 | /* |
b2be5e96 | 55 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 56 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 57 | */ |
0bf377bb IM |
58 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
59 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
60 | |
61 | /* | |
b2be5e96 PZ |
62 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
63 | */ | |
0bf377bb | 64 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
65 | |
66 | /* | |
2bba22c5 | 67 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 68 | * parent will (try to) run first. |
21805085 | 69 | */ |
2bba22c5 | 70 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 71 | |
bf0f6f24 IM |
72 | /* |
73 | * SCHED_OTHER wake-up granularity. | |
172e082a | 74 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
75 | * |
76 | * This option delays the preemption effects of decoupled workloads | |
77 | * and reduces their over-scheduling. Synchronous workloads will still | |
78 | * have immediate wakeup/sleep latencies. | |
79 | */ | |
172e082a | 80 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 81 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 82 | |
da84d961 IM |
83 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
84 | ||
a7a4f8a7 PT |
85 | /* |
86 | * The exponential sliding window over which load is averaged for shares | |
87 | * distribution. | |
88 | * (default: 10msec) | |
89 | */ | |
90 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
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 | ||
3d4b47b4 PZ |
138 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
139 | { | |
140 | if (!cfs_rq->on_list) { | |
67e86250 PT |
141 | /* |
142 | * Ensure we either appear before our parent (if already | |
143 | * enqueued) or force our parent to appear after us when it is | |
144 | * enqueued. The fact that we always enqueue bottom-up | |
145 | * reduces this to two cases. | |
146 | */ | |
147 | if (cfs_rq->tg->parent && | |
148 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
149 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
150 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
151 | } else { | |
152 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 153 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 154 | } |
3d4b47b4 PZ |
155 | |
156 | cfs_rq->on_list = 1; | |
157 | } | |
158 | } | |
159 | ||
160 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
161 | { | |
162 | if (cfs_rq->on_list) { | |
163 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
164 | cfs_rq->on_list = 0; | |
165 | } | |
166 | } | |
167 | ||
b758149c PZ |
168 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
169 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
170 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
171 | ||
172 | /* Do the two (enqueued) entities belong to the same group ? */ | |
173 | static inline int | |
174 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
175 | { | |
176 | if (se->cfs_rq == pse->cfs_rq) | |
177 | return 1; | |
178 | ||
179 | return 0; | |
180 | } | |
181 | ||
182 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
183 | { | |
184 | return se->parent; | |
185 | } | |
186 | ||
464b7527 PZ |
187 | /* return depth at which a sched entity is present in the hierarchy */ |
188 | static inline int depth_se(struct sched_entity *se) | |
189 | { | |
190 | int depth = 0; | |
191 | ||
192 | for_each_sched_entity(se) | |
193 | depth++; | |
194 | ||
195 | return depth; | |
196 | } | |
197 | ||
198 | static void | |
199 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
200 | { | |
201 | int se_depth, pse_depth; | |
202 | ||
203 | /* | |
204 | * preemption test can be made between sibling entities who are in the | |
205 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
206 | * both tasks until we find their ancestors who are siblings of common | |
207 | * parent. | |
208 | */ | |
209 | ||
210 | /* First walk up until both entities are at same depth */ | |
211 | se_depth = depth_se(*se); | |
212 | pse_depth = depth_se(*pse); | |
213 | ||
214 | while (se_depth > pse_depth) { | |
215 | se_depth--; | |
216 | *se = parent_entity(*se); | |
217 | } | |
218 | ||
219 | while (pse_depth > se_depth) { | |
220 | pse_depth--; | |
221 | *pse = parent_entity(*pse); | |
222 | } | |
223 | ||
224 | while (!is_same_group(*se, *pse)) { | |
225 | *se = parent_entity(*se); | |
226 | *pse = parent_entity(*pse); | |
227 | } | |
228 | } | |
229 | ||
8f48894f PZ |
230 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
231 | ||
232 | static inline struct task_struct *task_of(struct sched_entity *se) | |
233 | { | |
234 | return container_of(se, struct task_struct, se); | |
235 | } | |
bf0f6f24 | 236 | |
62160e3f IM |
237 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
238 | { | |
239 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
240 | } |
241 | ||
242 | #define entity_is_task(se) 1 | |
243 | ||
b758149c PZ |
244 | #define for_each_sched_entity(se) \ |
245 | for (; se; se = NULL) | |
bf0f6f24 | 246 | |
b758149c | 247 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 248 | { |
b758149c | 249 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
250 | } |
251 | ||
b758149c PZ |
252 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
253 | { | |
254 | struct task_struct *p = task_of(se); | |
255 | struct rq *rq = task_rq(p); | |
256 | ||
257 | return &rq->cfs; | |
258 | } | |
259 | ||
260 | /* runqueue "owned" by this group */ | |
261 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
262 | { | |
263 | return NULL; | |
264 | } | |
265 | ||
3d4b47b4 PZ |
266 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
267 | { | |
268 | } | |
269 | ||
270 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
271 | { | |
272 | } | |
273 | ||
b758149c PZ |
274 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
275 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
276 | ||
277 | static inline int | |
278 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
279 | { | |
280 | return 1; | |
281 | } | |
282 | ||
283 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
284 | { | |
285 | return NULL; | |
286 | } | |
287 | ||
464b7527 PZ |
288 | static inline void |
289 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
290 | { | |
291 | } | |
292 | ||
b758149c PZ |
293 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
294 | ||
bf0f6f24 IM |
295 | |
296 | /************************************************************** | |
297 | * Scheduling class tree data structure manipulation methods: | |
298 | */ | |
299 | ||
0702e3eb | 300 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 301 | { |
368059a9 PZ |
302 | s64 delta = (s64)(vruntime - min_vruntime); |
303 | if (delta > 0) | |
02e0431a PZ |
304 | min_vruntime = vruntime; |
305 | ||
306 | return min_vruntime; | |
307 | } | |
308 | ||
0702e3eb | 309 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
310 | { |
311 | s64 delta = (s64)(vruntime - min_vruntime); | |
312 | if (delta < 0) | |
313 | min_vruntime = vruntime; | |
314 | ||
315 | return min_vruntime; | |
316 | } | |
317 | ||
54fdc581 FC |
318 | static inline int entity_before(struct sched_entity *a, |
319 | struct sched_entity *b) | |
320 | { | |
321 | return (s64)(a->vruntime - b->vruntime) < 0; | |
322 | } | |
323 | ||
1af5f730 PZ |
324 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
325 | { | |
326 | u64 vruntime = cfs_rq->min_vruntime; | |
327 | ||
328 | if (cfs_rq->curr) | |
329 | vruntime = cfs_rq->curr->vruntime; | |
330 | ||
331 | if (cfs_rq->rb_leftmost) { | |
332 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
333 | struct sched_entity, | |
334 | run_node); | |
335 | ||
e17036da | 336 | if (!cfs_rq->curr) |
1af5f730 PZ |
337 | vruntime = se->vruntime; |
338 | else | |
339 | vruntime = min_vruntime(vruntime, se->vruntime); | |
340 | } | |
341 | ||
342 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
3fe1698b PZ |
343 | #ifndef CONFIG_64BIT |
344 | smp_wmb(); | |
345 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
346 | #endif | |
1af5f730 PZ |
347 | } |
348 | ||
bf0f6f24 IM |
349 | /* |
350 | * Enqueue an entity into the rb-tree: | |
351 | */ | |
0702e3eb | 352 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
353 | { |
354 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
355 | struct rb_node *parent = NULL; | |
356 | struct sched_entity *entry; | |
bf0f6f24 IM |
357 | int leftmost = 1; |
358 | ||
359 | /* | |
360 | * Find the right place in the rbtree: | |
361 | */ | |
362 | while (*link) { | |
363 | parent = *link; | |
364 | entry = rb_entry(parent, struct sched_entity, run_node); | |
365 | /* | |
366 | * We dont care about collisions. Nodes with | |
367 | * the same key stay together. | |
368 | */ | |
2bd2d6f2 | 369 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
370 | link = &parent->rb_left; |
371 | } else { | |
372 | link = &parent->rb_right; | |
373 | leftmost = 0; | |
374 | } | |
375 | } | |
376 | ||
377 | /* | |
378 | * Maintain a cache of leftmost tree entries (it is frequently | |
379 | * used): | |
380 | */ | |
1af5f730 | 381 | if (leftmost) |
57cb499d | 382 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
383 | |
384 | rb_link_node(&se->run_node, parent, link); | |
385 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
386 | } |
387 | ||
0702e3eb | 388 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 389 | { |
3fe69747 PZ |
390 | if (cfs_rq->rb_leftmost == &se->run_node) { |
391 | struct rb_node *next_node; | |
3fe69747 PZ |
392 | |
393 | next_node = rb_next(&se->run_node); | |
394 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 395 | } |
e9acbff6 | 396 | |
bf0f6f24 | 397 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
398 | } |
399 | ||
ac53db59 | 400 | static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 401 | { |
f4b6755f PZ |
402 | struct rb_node *left = cfs_rq->rb_leftmost; |
403 | ||
404 | if (!left) | |
405 | return NULL; | |
406 | ||
407 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
408 | } |
409 | ||
ac53db59 RR |
410 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
411 | { | |
412 | struct rb_node *next = rb_next(&se->run_node); | |
413 | ||
414 | if (!next) | |
415 | return NULL; | |
416 | ||
417 | return rb_entry(next, struct sched_entity, run_node); | |
418 | } | |
419 | ||
420 | #ifdef CONFIG_SCHED_DEBUG | |
f4b6755f | 421 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 422 | { |
7eee3e67 | 423 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 424 | |
70eee74b BS |
425 | if (!last) |
426 | return NULL; | |
7eee3e67 IM |
427 | |
428 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
429 | } |
430 | ||
bf0f6f24 IM |
431 | /************************************************************** |
432 | * Scheduling class statistics methods: | |
433 | */ | |
434 | ||
acb4a848 | 435 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 436 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
437 | loff_t *ppos) |
438 | { | |
8d65af78 | 439 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 440 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
441 | |
442 | if (ret || !write) | |
443 | return ret; | |
444 | ||
445 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
446 | sysctl_sched_min_granularity); | |
447 | ||
acb4a848 CE |
448 | #define WRT_SYSCTL(name) \ |
449 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
450 | WRT_SYSCTL(sched_min_granularity); | |
451 | WRT_SYSCTL(sched_latency); | |
452 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
453 | #undef WRT_SYSCTL |
454 | ||
b2be5e96 PZ |
455 | return 0; |
456 | } | |
457 | #endif | |
647e7cac | 458 | |
a7be37ac | 459 | /* |
f9c0b095 | 460 | * delta /= w |
a7be37ac PZ |
461 | */ |
462 | static inline unsigned long | |
463 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
464 | { | |
f9c0b095 PZ |
465 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
466 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
467 | |
468 | return delta; | |
469 | } | |
470 | ||
647e7cac IM |
471 | /* |
472 | * The idea is to set a period in which each task runs once. | |
473 | * | |
474 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
475 | * this period because otherwise the slices get too small. | |
476 | * | |
477 | * p = (nr <= nl) ? l : l*nr/nl | |
478 | */ | |
4d78e7b6 PZ |
479 | static u64 __sched_period(unsigned long nr_running) |
480 | { | |
481 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 482 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
483 | |
484 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 485 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 486 | period *= nr_running; |
4d78e7b6 PZ |
487 | } |
488 | ||
489 | return period; | |
490 | } | |
491 | ||
647e7cac IM |
492 | /* |
493 | * We calculate the wall-time slice from the period by taking a part | |
494 | * proportional to the weight. | |
495 | * | |
f9c0b095 | 496 | * s = p*P[w/rw] |
647e7cac | 497 | */ |
6d0f0ebd | 498 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 499 | { |
0a582440 | 500 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 501 | |
0a582440 | 502 | for_each_sched_entity(se) { |
6272d68c | 503 | struct load_weight *load; |
3104bf03 | 504 | struct load_weight lw; |
6272d68c LM |
505 | |
506 | cfs_rq = cfs_rq_of(se); | |
507 | load = &cfs_rq->load; | |
f9c0b095 | 508 | |
0a582440 | 509 | if (unlikely(!se->on_rq)) { |
3104bf03 | 510 | lw = cfs_rq->load; |
0a582440 MG |
511 | |
512 | update_load_add(&lw, se->load.weight); | |
513 | load = &lw; | |
514 | } | |
515 | slice = calc_delta_mine(slice, se->load.weight, load); | |
516 | } | |
517 | return slice; | |
bf0f6f24 IM |
518 | } |
519 | ||
647e7cac | 520 | /* |
ac884dec | 521 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 522 | * |
f9c0b095 | 523 | * vs = s/w |
647e7cac | 524 | */ |
f9c0b095 | 525 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 526 | { |
f9c0b095 | 527 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
528 | } |
529 | ||
d6b55918 | 530 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); |
6d5ab293 | 531 | static void update_cfs_shares(struct cfs_rq *cfs_rq); |
3b3d190e | 532 | |
bf0f6f24 IM |
533 | /* |
534 | * Update the current task's runtime statistics. Skip current tasks that | |
535 | * are not in our scheduling class. | |
536 | */ | |
537 | static inline void | |
8ebc91d9 IM |
538 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
539 | unsigned long delta_exec) | |
bf0f6f24 | 540 | { |
bbdba7c0 | 541 | unsigned long delta_exec_weighted; |
bf0f6f24 | 542 | |
41acab88 LDM |
543 | schedstat_set(curr->statistics.exec_max, |
544 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
545 | |
546 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 547 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 548 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 549 | |
e9acbff6 | 550 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 551 | update_min_vruntime(cfs_rq); |
3b3d190e | 552 | |
70caf8a6 | 553 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
3b3d190e | 554 | cfs_rq->load_unacc_exec_time += delta_exec; |
3b3d190e | 555 | #endif |
bf0f6f24 IM |
556 | } |
557 | ||
b7cc0896 | 558 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 559 | { |
429d43bc | 560 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 561 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
562 | unsigned long delta_exec; |
563 | ||
564 | if (unlikely(!curr)) | |
565 | return; | |
566 | ||
567 | /* | |
568 | * Get the amount of time the current task was running | |
569 | * since the last time we changed load (this cannot | |
570 | * overflow on 32 bits): | |
571 | */ | |
8ebc91d9 | 572 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
573 | if (!delta_exec) |
574 | return; | |
bf0f6f24 | 575 | |
8ebc91d9 IM |
576 | __update_curr(cfs_rq, curr, delta_exec); |
577 | curr->exec_start = now; | |
d842de87 SV |
578 | |
579 | if (entity_is_task(curr)) { | |
580 | struct task_struct *curtask = task_of(curr); | |
581 | ||
f977bb49 | 582 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 583 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 584 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 585 | } |
bf0f6f24 IM |
586 | } |
587 | ||
588 | static inline void | |
5870db5b | 589 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 590 | { |
41acab88 | 591 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
592 | } |
593 | ||
bf0f6f24 IM |
594 | /* |
595 | * Task is being enqueued - update stats: | |
596 | */ | |
d2417e5a | 597 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 598 | { |
bf0f6f24 IM |
599 | /* |
600 | * Are we enqueueing a waiting task? (for current tasks | |
601 | * a dequeue/enqueue event is a NOP) | |
602 | */ | |
429d43bc | 603 | if (se != cfs_rq->curr) |
5870db5b | 604 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
605 | } |
606 | ||
bf0f6f24 | 607 | static void |
9ef0a961 | 608 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 609 | { |
41acab88 LDM |
610 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
611 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
612 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
613 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
614 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
615 | #ifdef CONFIG_SCHEDSTATS |
616 | if (entity_is_task(se)) { | |
617 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 618 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
619 | } |
620 | #endif | |
41acab88 | 621 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
622 | } |
623 | ||
624 | static inline void | |
19b6a2e3 | 625 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 626 | { |
bf0f6f24 IM |
627 | /* |
628 | * Mark the end of the wait period if dequeueing a | |
629 | * waiting task: | |
630 | */ | |
429d43bc | 631 | if (se != cfs_rq->curr) |
9ef0a961 | 632 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
633 | } |
634 | ||
635 | /* | |
636 | * We are picking a new current task - update its stats: | |
637 | */ | |
638 | static inline void | |
79303e9e | 639 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
640 | { |
641 | /* | |
642 | * We are starting a new run period: | |
643 | */ | |
305e6835 | 644 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
645 | } |
646 | ||
bf0f6f24 IM |
647 | /************************************************** |
648 | * Scheduling class queueing methods: | |
649 | */ | |
650 | ||
c09595f6 PZ |
651 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
652 | static void | |
653 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
654 | { | |
655 | cfs_rq->task_weight += weight; | |
656 | } | |
657 | #else | |
658 | static inline void | |
659 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
660 | { | |
661 | } | |
662 | #endif | |
663 | ||
30cfdcfc DA |
664 | static void |
665 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
666 | { | |
667 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
668 | if (!parent_entity(se)) |
669 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 670 | if (entity_is_task(se)) { |
c09595f6 | 671 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
672 | list_add(&se->group_node, &cfs_rq->tasks); |
673 | } | |
30cfdcfc | 674 | cfs_rq->nr_running++; |
30cfdcfc DA |
675 | } |
676 | ||
677 | static void | |
678 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
679 | { | |
680 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
681 | if (!parent_entity(se)) |
682 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 683 | if (entity_is_task(se)) { |
c09595f6 | 684 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
685 | list_del_init(&se->group_node); |
686 | } | |
30cfdcfc | 687 | cfs_rq->nr_running--; |
30cfdcfc DA |
688 | } |
689 | ||
3ff6dcac YZ |
690 | #ifdef CONFIG_FAIR_GROUP_SCHED |
691 | # ifdef CONFIG_SMP | |
d6b55918 PT |
692 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
693 | int global_update) | |
694 | { | |
695 | struct task_group *tg = cfs_rq->tg; | |
696 | long load_avg; | |
697 | ||
698 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
699 | load_avg -= cfs_rq->load_contribution; | |
700 | ||
701 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
702 | atomic_add(load_avg, &tg->load_weight); | |
703 | cfs_rq->load_contribution += load_avg; | |
704 | } | |
705 | } | |
706 | ||
707 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 708 | { |
a7a4f8a7 | 709 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 710 | u64 now, delta; |
e33078ba | 711 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 712 | |
b815f196 | 713 | if (cfs_rq->tg == &root_task_group) |
2069dd75 PZ |
714 | return; |
715 | ||
05ca62c6 | 716 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
717 | delta = now - cfs_rq->load_stamp; |
718 | ||
e33078ba PT |
719 | /* truncate load history at 4 idle periods */ |
720 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
721 | now - cfs_rq->load_last > 4 * period) { | |
722 | cfs_rq->load_period = 0; | |
723 | cfs_rq->load_avg = 0; | |
f07333bf | 724 | delta = period - 1; |
e33078ba PT |
725 | } |
726 | ||
2069dd75 | 727 | cfs_rq->load_stamp = now; |
3b3d190e | 728 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 729 | cfs_rq->load_period += delta; |
e33078ba PT |
730 | if (load) { |
731 | cfs_rq->load_last = now; | |
732 | cfs_rq->load_avg += delta * load; | |
733 | } | |
2069dd75 | 734 | |
d6b55918 PT |
735 | /* consider updating load contribution on each fold or truncate */ |
736 | if (global_update || cfs_rq->load_period > period | |
737 | || !cfs_rq->load_period) | |
738 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
739 | ||
2069dd75 PZ |
740 | while (cfs_rq->load_period > period) { |
741 | /* | |
742 | * Inline assembly required to prevent the compiler | |
743 | * optimising this loop into a divmod call. | |
744 | * See __iter_div_u64_rem() for another example of this. | |
745 | */ | |
746 | asm("" : "+rm" (cfs_rq->load_period)); | |
747 | cfs_rq->load_period /= 2; | |
748 | cfs_rq->load_avg /= 2; | |
749 | } | |
3d4b47b4 | 750 | |
e33078ba PT |
751 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
752 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
753 | } |
754 | ||
6d5ab293 | 755 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
756 | { |
757 | long load_weight, load, shares; | |
758 | ||
6d5ab293 | 759 | load = cfs_rq->load.weight; |
3ff6dcac YZ |
760 | |
761 | load_weight = atomic_read(&tg->load_weight); | |
3ff6dcac | 762 | load_weight += load; |
6d5ab293 | 763 | load_weight -= cfs_rq->load_contribution; |
3ff6dcac YZ |
764 | |
765 | shares = (tg->shares * load); | |
766 | if (load_weight) | |
767 | shares /= load_weight; | |
768 | ||
769 | if (shares < MIN_SHARES) | |
770 | shares = MIN_SHARES; | |
771 | if (shares > tg->shares) | |
772 | shares = tg->shares; | |
773 | ||
774 | return shares; | |
775 | } | |
776 | ||
777 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
778 | { | |
779 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
780 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 781 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
782 | } |
783 | } | |
784 | # else /* CONFIG_SMP */ | |
785 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
786 | { | |
787 | } | |
788 | ||
6d5ab293 | 789 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
790 | { |
791 | return tg->shares; | |
792 | } | |
793 | ||
794 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
795 | { | |
796 | } | |
797 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
798 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
799 | unsigned long weight) | |
800 | { | |
19e5eebb PT |
801 | if (se->on_rq) { |
802 | /* commit outstanding execution time */ | |
803 | if (cfs_rq->curr == se) | |
804 | update_curr(cfs_rq); | |
2069dd75 | 805 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 806 | } |
2069dd75 PZ |
807 | |
808 | update_load_set(&se->load, weight); | |
809 | ||
810 | if (se->on_rq) | |
811 | account_entity_enqueue(cfs_rq, se); | |
812 | } | |
813 | ||
6d5ab293 | 814 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
815 | { |
816 | struct task_group *tg; | |
817 | struct sched_entity *se; | |
3ff6dcac | 818 | long shares; |
2069dd75 | 819 | |
2069dd75 PZ |
820 | tg = cfs_rq->tg; |
821 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
822 | if (!se) | |
823 | return; | |
3ff6dcac YZ |
824 | #ifndef CONFIG_SMP |
825 | if (likely(se->load.weight == tg->shares)) | |
826 | return; | |
827 | #endif | |
6d5ab293 | 828 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
829 | |
830 | reweight_entity(cfs_rq_of(se), se, shares); | |
831 | } | |
832 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 833 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
834 | { |
835 | } | |
836 | ||
6d5ab293 | 837 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
838 | { |
839 | } | |
43365bd7 PT |
840 | |
841 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
842 | { | |
843 | } | |
2069dd75 PZ |
844 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
845 | ||
2396af69 | 846 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 847 | { |
bf0f6f24 | 848 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
849 | struct task_struct *tsk = NULL; |
850 | ||
851 | if (entity_is_task(se)) | |
852 | tsk = task_of(se); | |
853 | ||
41acab88 LDM |
854 | if (se->statistics.sleep_start) { |
855 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
856 | |
857 | if ((s64)delta < 0) | |
858 | delta = 0; | |
859 | ||
41acab88 LDM |
860 | if (unlikely(delta > se->statistics.sleep_max)) |
861 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 862 | |
41acab88 LDM |
863 | se->statistics.sleep_start = 0; |
864 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 865 | |
768d0c27 | 866 | if (tsk) { |
e414314c | 867 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
868 | trace_sched_stat_sleep(tsk, delta); |
869 | } | |
bf0f6f24 | 870 | } |
41acab88 LDM |
871 | if (se->statistics.block_start) { |
872 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
873 | |
874 | if ((s64)delta < 0) | |
875 | delta = 0; | |
876 | ||
41acab88 LDM |
877 | if (unlikely(delta > se->statistics.block_max)) |
878 | se->statistics.block_max = delta; | |
bf0f6f24 | 879 | |
41acab88 LDM |
880 | se->statistics.block_start = 0; |
881 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 882 | |
e414314c | 883 | if (tsk) { |
8f0dfc34 | 884 | if (tsk->in_iowait) { |
41acab88 LDM |
885 | se->statistics.iowait_sum += delta; |
886 | se->statistics.iowait_count++; | |
768d0c27 | 887 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
888 | } |
889 | ||
e414314c PZ |
890 | /* |
891 | * Blocking time is in units of nanosecs, so shift by | |
892 | * 20 to get a milliseconds-range estimation of the | |
893 | * amount of time that the task spent sleeping: | |
894 | */ | |
895 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
896 | profile_hits(SLEEP_PROFILING, | |
897 | (void *)get_wchan(tsk), | |
898 | delta >> 20); | |
899 | } | |
900 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 901 | } |
bf0f6f24 IM |
902 | } |
903 | #endif | |
904 | } | |
905 | ||
ddc97297 PZ |
906 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
907 | { | |
908 | #ifdef CONFIG_SCHED_DEBUG | |
909 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
910 | ||
911 | if (d < 0) | |
912 | d = -d; | |
913 | ||
914 | if (d > 3*sysctl_sched_latency) | |
915 | schedstat_inc(cfs_rq, nr_spread_over); | |
916 | #endif | |
917 | } | |
918 | ||
aeb73b04 PZ |
919 | static void |
920 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
921 | { | |
1af5f730 | 922 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 923 | |
2cb8600e PZ |
924 | /* |
925 | * The 'current' period is already promised to the current tasks, | |
926 | * however the extra weight of the new task will slow them down a | |
927 | * little, place the new task so that it fits in the slot that | |
928 | * stays open at the end. | |
929 | */ | |
94dfb5e7 | 930 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 931 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 932 | |
a2e7a7eb | 933 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 934 | if (!initial) { |
a2e7a7eb | 935 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 936 | |
a2e7a7eb MG |
937 | /* |
938 | * Halve their sleep time's effect, to allow | |
939 | * for a gentler effect of sleepers: | |
940 | */ | |
941 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
942 | thresh >>= 1; | |
51e0304c | 943 | |
a2e7a7eb | 944 | vruntime -= thresh; |
aeb73b04 PZ |
945 | } |
946 | ||
b5d9d734 MG |
947 | /* ensure we never gain time by being placed backwards. */ |
948 | vruntime = max_vruntime(se->vruntime, vruntime); | |
949 | ||
67e9fb2a | 950 | se->vruntime = vruntime; |
aeb73b04 PZ |
951 | } |
952 | ||
bf0f6f24 | 953 | static void |
88ec22d3 | 954 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 955 | { |
88ec22d3 PZ |
956 | /* |
957 | * Update the normalized vruntime before updating min_vruntime | |
958 | * through callig update_curr(). | |
959 | */ | |
371fd7e7 | 960 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
961 | se->vruntime += cfs_rq->min_vruntime; |
962 | ||
bf0f6f24 | 963 | /* |
a2a2d680 | 964 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 965 | */ |
b7cc0896 | 966 | update_curr(cfs_rq); |
d6b55918 | 967 | update_cfs_load(cfs_rq, 0); |
a992241d | 968 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 969 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 970 | |
88ec22d3 | 971 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 972 | place_entity(cfs_rq, se, 0); |
2396af69 | 973 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 974 | } |
bf0f6f24 | 975 | |
d2417e5a | 976 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 977 | check_spread(cfs_rq, se); |
83b699ed SV |
978 | if (se != cfs_rq->curr) |
979 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 980 | se->on_rq = 1; |
3d4b47b4 PZ |
981 | |
982 | if (cfs_rq->nr_running == 1) | |
983 | list_add_leaf_cfs_rq(cfs_rq); | |
bf0f6f24 IM |
984 | } |
985 | ||
2c13c919 | 986 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 987 | { |
2c13c919 RR |
988 | for_each_sched_entity(se) { |
989 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
990 | if (cfs_rq->last == se) | |
991 | cfs_rq->last = NULL; | |
992 | else | |
993 | break; | |
994 | } | |
995 | } | |
2002c695 | 996 | |
2c13c919 RR |
997 | static void __clear_buddies_next(struct sched_entity *se) |
998 | { | |
999 | for_each_sched_entity(se) { | |
1000 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1001 | if (cfs_rq->next == se) | |
1002 | cfs_rq->next = NULL; | |
1003 | else | |
1004 | break; | |
1005 | } | |
2002c695 PZ |
1006 | } |
1007 | ||
ac53db59 RR |
1008 | static void __clear_buddies_skip(struct sched_entity *se) |
1009 | { | |
1010 | for_each_sched_entity(se) { | |
1011 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1012 | if (cfs_rq->skip == se) | |
1013 | cfs_rq->skip = NULL; | |
1014 | else | |
1015 | break; | |
1016 | } | |
1017 | } | |
1018 | ||
a571bbea PZ |
1019 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1020 | { | |
2c13c919 RR |
1021 | if (cfs_rq->last == se) |
1022 | __clear_buddies_last(se); | |
1023 | ||
1024 | if (cfs_rq->next == se) | |
1025 | __clear_buddies_next(se); | |
ac53db59 RR |
1026 | |
1027 | if (cfs_rq->skip == se) | |
1028 | __clear_buddies_skip(se); | |
a571bbea PZ |
1029 | } |
1030 | ||
bf0f6f24 | 1031 | static void |
371fd7e7 | 1032 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1033 | { |
a2a2d680 DA |
1034 | /* |
1035 | * Update run-time statistics of the 'current'. | |
1036 | */ | |
1037 | update_curr(cfs_rq); | |
1038 | ||
19b6a2e3 | 1039 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1040 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1041 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1042 | if (entity_is_task(se)) { |
1043 | struct task_struct *tsk = task_of(se); | |
1044 | ||
1045 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1046 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1047 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1048 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1049 | } |
db36cc7d | 1050 | #endif |
67e9fb2a PZ |
1051 | } |
1052 | ||
2002c695 | 1053 | clear_buddies(cfs_rq, se); |
4793241b | 1054 | |
83b699ed | 1055 | if (se != cfs_rq->curr) |
30cfdcfc | 1056 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1057 | se->on_rq = 0; |
d6b55918 | 1058 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1059 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1060 | |
1061 | /* | |
1062 | * Normalize the entity after updating the min_vruntime because the | |
1063 | * update can refer to the ->curr item and we need to reflect this | |
1064 | * movement in our normalized position. | |
1065 | */ | |
371fd7e7 | 1066 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1067 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 PZ |
1068 | |
1069 | update_min_vruntime(cfs_rq); | |
1070 | update_cfs_shares(cfs_rq); | |
bf0f6f24 IM |
1071 | } |
1072 | ||
1073 | /* | |
1074 | * Preempt the current task with a newly woken task if needed: | |
1075 | */ | |
7c92e54f | 1076 | static void |
2e09bf55 | 1077 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1078 | { |
11697830 PZ |
1079 | unsigned long ideal_runtime, delta_exec; |
1080 | ||
6d0f0ebd | 1081 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1082 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1083 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1084 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1085 | /* |
1086 | * The current task ran long enough, ensure it doesn't get | |
1087 | * re-elected due to buddy favours. | |
1088 | */ | |
1089 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1090 | return; |
1091 | } | |
1092 | ||
1093 | /* | |
1094 | * Ensure that a task that missed wakeup preemption by a | |
1095 | * narrow margin doesn't have to wait for a full slice. | |
1096 | * This also mitigates buddy induced latencies under load. | |
1097 | */ | |
f685ceac MG |
1098 | if (delta_exec < sysctl_sched_min_granularity) |
1099 | return; | |
1100 | ||
1101 | if (cfs_rq->nr_running > 1) { | |
ac53db59 | 1102 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac MG |
1103 | s64 delta = curr->vruntime - se->vruntime; |
1104 | ||
d7d82944 MG |
1105 | if (delta < 0) |
1106 | return; | |
1107 | ||
f685ceac MG |
1108 | if (delta > ideal_runtime) |
1109 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 1110 | } |
bf0f6f24 IM |
1111 | } |
1112 | ||
83b699ed | 1113 | static void |
8494f412 | 1114 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1115 | { |
83b699ed SV |
1116 | /* 'current' is not kept within the tree. */ |
1117 | if (se->on_rq) { | |
1118 | /* | |
1119 | * Any task has to be enqueued before it get to execute on | |
1120 | * a CPU. So account for the time it spent waiting on the | |
1121 | * runqueue. | |
1122 | */ | |
1123 | update_stats_wait_end(cfs_rq, se); | |
1124 | __dequeue_entity(cfs_rq, se); | |
1125 | } | |
1126 | ||
79303e9e | 1127 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1128 | cfs_rq->curr = se; |
eba1ed4b IM |
1129 | #ifdef CONFIG_SCHEDSTATS |
1130 | /* | |
1131 | * Track our maximum slice length, if the CPU's load is at | |
1132 | * least twice that of our own weight (i.e. dont track it | |
1133 | * when there are only lesser-weight tasks around): | |
1134 | */ | |
495eca49 | 1135 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1136 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1137 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1138 | } | |
1139 | #endif | |
4a55b450 | 1140 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1141 | } |
1142 | ||
3f3a4904 PZ |
1143 | static int |
1144 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1145 | ||
ac53db59 RR |
1146 | /* |
1147 | * Pick the next process, keeping these things in mind, in this order: | |
1148 | * 1) keep things fair between processes/task groups | |
1149 | * 2) pick the "next" process, since someone really wants that to run | |
1150 | * 3) pick the "last" process, for cache locality | |
1151 | * 4) do not run the "skip" process, if something else is available | |
1152 | */ | |
f4b6755f | 1153 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1154 | { |
ac53db59 | 1155 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1156 | struct sched_entity *left = se; |
f4b6755f | 1157 | |
ac53db59 RR |
1158 | /* |
1159 | * Avoid running the skip buddy, if running something else can | |
1160 | * be done without getting too unfair. | |
1161 | */ | |
1162 | if (cfs_rq->skip == se) { | |
1163 | struct sched_entity *second = __pick_next_entity(se); | |
1164 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1165 | se = second; | |
1166 | } | |
aa2ac252 | 1167 | |
f685ceac MG |
1168 | /* |
1169 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1170 | */ | |
1171 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1172 | se = cfs_rq->last; | |
1173 | ||
ac53db59 RR |
1174 | /* |
1175 | * Someone really wants this to run. If it's not unfair, run it. | |
1176 | */ | |
1177 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1178 | se = cfs_rq->next; | |
1179 | ||
f685ceac | 1180 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1181 | |
1182 | return se; | |
aa2ac252 PZ |
1183 | } |
1184 | ||
ab6cde26 | 1185 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1186 | { |
1187 | /* | |
1188 | * If still on the runqueue then deactivate_task() | |
1189 | * was not called and update_curr() has to be done: | |
1190 | */ | |
1191 | if (prev->on_rq) | |
b7cc0896 | 1192 | update_curr(cfs_rq); |
bf0f6f24 | 1193 | |
ddc97297 | 1194 | check_spread(cfs_rq, prev); |
30cfdcfc | 1195 | if (prev->on_rq) { |
5870db5b | 1196 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1197 | /* Put 'current' back into the tree. */ |
1198 | __enqueue_entity(cfs_rq, prev); | |
1199 | } | |
429d43bc | 1200 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1201 | } |
1202 | ||
8f4d37ec PZ |
1203 | static void |
1204 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1205 | { |
bf0f6f24 | 1206 | /* |
30cfdcfc | 1207 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1208 | */ |
30cfdcfc | 1209 | update_curr(cfs_rq); |
bf0f6f24 | 1210 | |
43365bd7 PT |
1211 | /* |
1212 | * Update share accounting for long-running entities. | |
1213 | */ | |
1214 | update_entity_shares_tick(cfs_rq); | |
1215 | ||
8f4d37ec PZ |
1216 | #ifdef CONFIG_SCHED_HRTICK |
1217 | /* | |
1218 | * queued ticks are scheduled to match the slice, so don't bother | |
1219 | * validating it and just reschedule. | |
1220 | */ | |
983ed7a6 HH |
1221 | if (queued) { |
1222 | resched_task(rq_of(cfs_rq)->curr); | |
1223 | return; | |
1224 | } | |
8f4d37ec PZ |
1225 | /* |
1226 | * don't let the period tick interfere with the hrtick preemption | |
1227 | */ | |
1228 | if (!sched_feat(DOUBLE_TICK) && | |
1229 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1230 | return; | |
1231 | #endif | |
1232 | ||
2c2efaed | 1233 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1234 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1235 | } |
1236 | ||
1237 | /************************************************** | |
1238 | * CFS operations on tasks: | |
1239 | */ | |
1240 | ||
8f4d37ec PZ |
1241 | #ifdef CONFIG_SCHED_HRTICK |
1242 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1243 | { | |
8f4d37ec PZ |
1244 | struct sched_entity *se = &p->se; |
1245 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1246 | ||
1247 | WARN_ON(task_rq(p) != rq); | |
1248 | ||
1249 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1250 | u64 slice = sched_slice(cfs_rq, se); | |
1251 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1252 | s64 delta = slice - ran; | |
1253 | ||
1254 | if (delta < 0) { | |
1255 | if (rq->curr == p) | |
1256 | resched_task(p); | |
1257 | return; | |
1258 | } | |
1259 | ||
1260 | /* | |
1261 | * Don't schedule slices shorter than 10000ns, that just | |
1262 | * doesn't make sense. Rely on vruntime for fairness. | |
1263 | */ | |
31656519 | 1264 | if (rq->curr != p) |
157124c1 | 1265 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1266 | |
31656519 | 1267 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1268 | } |
1269 | } | |
a4c2f00f PZ |
1270 | |
1271 | /* | |
1272 | * called from enqueue/dequeue and updates the hrtick when the | |
1273 | * current task is from our class and nr_running is low enough | |
1274 | * to matter. | |
1275 | */ | |
1276 | static void hrtick_update(struct rq *rq) | |
1277 | { | |
1278 | struct task_struct *curr = rq->curr; | |
1279 | ||
1280 | if (curr->sched_class != &fair_sched_class) | |
1281 | return; | |
1282 | ||
1283 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1284 | hrtick_start_fair(rq, curr); | |
1285 | } | |
55e12e5e | 1286 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1287 | static inline void |
1288 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1289 | { | |
1290 | } | |
a4c2f00f PZ |
1291 | |
1292 | static inline void hrtick_update(struct rq *rq) | |
1293 | { | |
1294 | } | |
8f4d37ec PZ |
1295 | #endif |
1296 | ||
bf0f6f24 IM |
1297 | /* |
1298 | * The enqueue_task method is called before nr_running is | |
1299 | * increased. Here we update the fair scheduling stats and | |
1300 | * then put the task into the rbtree: | |
1301 | */ | |
ea87bb78 | 1302 | static void |
371fd7e7 | 1303 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1304 | { |
1305 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1306 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1307 | |
1308 | for_each_sched_entity(se) { | |
62fb1851 | 1309 | if (se->on_rq) |
bf0f6f24 IM |
1310 | break; |
1311 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 PZ |
1312 | enqueue_entity(cfs_rq, se, flags); |
1313 | flags = ENQUEUE_WAKEUP; | |
bf0f6f24 | 1314 | } |
8f4d37ec | 1315 | |
2069dd75 | 1316 | for_each_sched_entity(se) { |
0f317143 | 1317 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 1318 | |
d6b55918 | 1319 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1320 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1321 | } |
1322 | ||
a4c2f00f | 1323 | hrtick_update(rq); |
bf0f6f24 IM |
1324 | } |
1325 | ||
2f36825b VP |
1326 | static void set_next_buddy(struct sched_entity *se); |
1327 | ||
bf0f6f24 IM |
1328 | /* |
1329 | * The dequeue_task method is called before nr_running is | |
1330 | * decreased. We remove the task from the rbtree and | |
1331 | * update the fair scheduling stats: | |
1332 | */ | |
371fd7e7 | 1333 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1334 | { |
1335 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1336 | struct sched_entity *se = &p->se; |
2f36825b | 1337 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
1338 | |
1339 | for_each_sched_entity(se) { | |
1340 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1341 | dequeue_entity(cfs_rq, se, flags); |
2069dd75 | 1342 | |
bf0f6f24 | 1343 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
1344 | if (cfs_rq->load.weight) { |
1345 | /* | |
1346 | * Bias pick_next to pick a task from this cfs_rq, as | |
1347 | * p is sleeping when it is within its sched_slice. | |
1348 | */ | |
1349 | if (task_sleep && parent_entity(se)) | |
1350 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
1351 | |
1352 | /* avoid re-evaluating load for this entity */ | |
1353 | se = parent_entity(se); | |
bf0f6f24 | 1354 | break; |
2f36825b | 1355 | } |
371fd7e7 | 1356 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1357 | } |
8f4d37ec | 1358 | |
2069dd75 | 1359 | for_each_sched_entity(se) { |
0f317143 | 1360 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 1361 | |
d6b55918 | 1362 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1363 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1364 | } |
1365 | ||
a4c2f00f | 1366 | hrtick_update(rq); |
bf0f6f24 IM |
1367 | } |
1368 | ||
e7693a36 | 1369 | #ifdef CONFIG_SMP |
098fb9db | 1370 | |
74f8e4b2 | 1371 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
1372 | { |
1373 | struct sched_entity *se = &p->se; | |
1374 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
1375 | u64 min_vruntime; |
1376 | ||
1377 | #ifndef CONFIG_64BIT | |
1378 | u64 min_vruntime_copy; | |
88ec22d3 | 1379 | |
3fe1698b PZ |
1380 | do { |
1381 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
1382 | smp_rmb(); | |
1383 | min_vruntime = cfs_rq->min_vruntime; | |
1384 | } while (min_vruntime != min_vruntime_copy); | |
1385 | #else | |
1386 | min_vruntime = cfs_rq->min_vruntime; | |
1387 | #endif | |
88ec22d3 | 1388 | |
3fe1698b | 1389 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
1390 | } |
1391 | ||
bb3469ac | 1392 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1393 | /* |
1394 | * effective_load() calculates the load change as seen from the root_task_group | |
1395 | * | |
1396 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1397 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1398 | * can calculate the shift in shares. | |
f5bfb7d9 | 1399 | */ |
2069dd75 | 1400 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 1401 | { |
4be9daaa | 1402 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1403 | |
1404 | if (!tg->parent) | |
1405 | return wl; | |
1406 | ||
4be9daaa | 1407 | for_each_sched_entity(se) { |
977dda7c | 1408 | long lw, w; |
4be9daaa | 1409 | |
977dda7c PT |
1410 | tg = se->my_q->tg; |
1411 | w = se->my_q->load.weight; | |
bb3469ac | 1412 | |
977dda7c PT |
1413 | /* use this cpu's instantaneous contribution */ |
1414 | lw = atomic_read(&tg->load_weight); | |
1415 | lw -= se->my_q->load_contribution; | |
1416 | lw += w + wg; | |
4be9daaa | 1417 | |
977dda7c | 1418 | wl += w; |
940959e9 | 1419 | |
977dda7c PT |
1420 | if (lw > 0 && wl < lw) |
1421 | wl = (wl * tg->shares) / lw; | |
1422 | else | |
1423 | wl = tg->shares; | |
940959e9 | 1424 | |
977dda7c PT |
1425 | /* zero point is MIN_SHARES */ |
1426 | if (wl < MIN_SHARES) | |
1427 | wl = MIN_SHARES; | |
1428 | wl -= se->load.weight; | |
4be9daaa | 1429 | wg = 0; |
4be9daaa | 1430 | } |
bb3469ac | 1431 | |
4be9daaa | 1432 | return wl; |
bb3469ac | 1433 | } |
4be9daaa | 1434 | |
bb3469ac | 1435 | #else |
4be9daaa | 1436 | |
83378269 PZ |
1437 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1438 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1439 | { |
83378269 | 1440 | return wl; |
bb3469ac | 1441 | } |
4be9daaa | 1442 | |
bb3469ac PZ |
1443 | #endif |
1444 | ||
c88d5910 | 1445 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1446 | { |
e37b6a7b | 1447 | s64 this_load, load; |
c88d5910 | 1448 | int idx, this_cpu, prev_cpu; |
098fb9db | 1449 | unsigned long tl_per_task; |
c88d5910 | 1450 | struct task_group *tg; |
83378269 | 1451 | unsigned long weight; |
b3137bc8 | 1452 | int balanced; |
098fb9db | 1453 | |
c88d5910 PZ |
1454 | idx = sd->wake_idx; |
1455 | this_cpu = smp_processor_id(); | |
1456 | prev_cpu = task_cpu(p); | |
1457 | load = source_load(prev_cpu, idx); | |
1458 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1459 | |
b3137bc8 MG |
1460 | /* |
1461 | * If sync wakeup then subtract the (maximum possible) | |
1462 | * effect of the currently running task from the load | |
1463 | * of the current CPU: | |
1464 | */ | |
83378269 PZ |
1465 | if (sync) { |
1466 | tg = task_group(current); | |
1467 | weight = current->se.load.weight; | |
1468 | ||
c88d5910 | 1469 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1470 | load += effective_load(tg, prev_cpu, 0, -weight); |
1471 | } | |
b3137bc8 | 1472 | |
83378269 PZ |
1473 | tg = task_group(p); |
1474 | weight = p->se.load.weight; | |
b3137bc8 | 1475 | |
71a29aa7 PZ |
1476 | /* |
1477 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1478 | * due to the sync cause above having dropped this_load to 0, we'll |
1479 | * always have an imbalance, but there's really nothing you can do | |
1480 | * about that, so that's good too. | |
71a29aa7 PZ |
1481 | * |
1482 | * Otherwise check if either cpus are near enough in load to allow this | |
1483 | * task to be woken on this_cpu. | |
1484 | */ | |
e37b6a7b PT |
1485 | if (this_load > 0) { |
1486 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
1487 | |
1488 | this_eff_load = 100; | |
1489 | this_eff_load *= power_of(prev_cpu); | |
1490 | this_eff_load *= this_load + | |
1491 | effective_load(tg, this_cpu, weight, weight); | |
1492 | ||
1493 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
1494 | prev_eff_load *= power_of(this_cpu); | |
1495 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
1496 | ||
1497 | balanced = this_eff_load <= prev_eff_load; | |
1498 | } else | |
1499 | balanced = true; | |
b3137bc8 | 1500 | |
098fb9db | 1501 | /* |
4ae7d5ce IM |
1502 | * If the currently running task will sleep within |
1503 | * a reasonable amount of time then attract this newly | |
1504 | * woken task: | |
098fb9db | 1505 | */ |
2fb7635c PZ |
1506 | if (sync && balanced) |
1507 | return 1; | |
098fb9db | 1508 | |
41acab88 | 1509 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
1510 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
1511 | ||
c88d5910 PZ |
1512 | if (balanced || |
1513 | (this_load <= load && | |
1514 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1515 | /* |
1516 | * This domain has SD_WAKE_AFFINE and | |
1517 | * p is cache cold in this domain, and | |
1518 | * there is no bad imbalance. | |
1519 | */ | |
c88d5910 | 1520 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 1521 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
1522 | |
1523 | return 1; | |
1524 | } | |
1525 | return 0; | |
1526 | } | |
1527 | ||
aaee1203 PZ |
1528 | /* |
1529 | * find_idlest_group finds and returns the least busy CPU group within the | |
1530 | * domain. | |
1531 | */ | |
1532 | static struct sched_group * | |
78e7ed53 | 1533 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1534 | int this_cpu, int load_idx) |
e7693a36 | 1535 | { |
b3bd3de6 | 1536 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 1537 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 1538 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1539 | |
aaee1203 PZ |
1540 | do { |
1541 | unsigned long load, avg_load; | |
1542 | int local_group; | |
1543 | int i; | |
e7693a36 | 1544 | |
aaee1203 PZ |
1545 | /* Skip over this group if it has no CPUs allowed */ |
1546 | if (!cpumask_intersects(sched_group_cpus(group), | |
1547 | &p->cpus_allowed)) | |
1548 | continue; | |
1549 | ||
1550 | local_group = cpumask_test_cpu(this_cpu, | |
1551 | sched_group_cpus(group)); | |
1552 | ||
1553 | /* Tally up the load of all CPUs in the group */ | |
1554 | avg_load = 0; | |
1555 | ||
1556 | for_each_cpu(i, sched_group_cpus(group)) { | |
1557 | /* Bias balancing toward cpus of our domain */ | |
1558 | if (local_group) | |
1559 | load = source_load(i, load_idx); | |
1560 | else | |
1561 | load = target_load(i, load_idx); | |
1562 | ||
1563 | avg_load += load; | |
1564 | } | |
1565 | ||
1566 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 1567 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
1568 | |
1569 | if (local_group) { | |
1570 | this_load = avg_load; | |
aaee1203 PZ |
1571 | } else if (avg_load < min_load) { |
1572 | min_load = avg_load; | |
1573 | idlest = group; | |
1574 | } | |
1575 | } while (group = group->next, group != sd->groups); | |
1576 | ||
1577 | if (!idlest || 100*this_load < imbalance*min_load) | |
1578 | return NULL; | |
1579 | return idlest; | |
1580 | } | |
1581 | ||
1582 | /* | |
1583 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1584 | */ | |
1585 | static int | |
1586 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1587 | { | |
1588 | unsigned long load, min_load = ULONG_MAX; | |
1589 | int idlest = -1; | |
1590 | int i; | |
1591 | ||
1592 | /* Traverse only the allowed CPUs */ | |
1593 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1594 | load = weighted_cpuload(i); | |
1595 | ||
1596 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1597 | min_load = load; | |
1598 | idlest = i; | |
e7693a36 GH |
1599 | } |
1600 | } | |
1601 | ||
aaee1203 PZ |
1602 | return idlest; |
1603 | } | |
e7693a36 | 1604 | |
a50bde51 PZ |
1605 | /* |
1606 | * Try and locate an idle CPU in the sched_domain. | |
1607 | */ | |
99bd5e2f | 1608 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
1609 | { |
1610 | int cpu = smp_processor_id(); | |
1611 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 1612 | struct sched_domain *sd; |
a50bde51 PZ |
1613 | int i; |
1614 | ||
1615 | /* | |
99bd5e2f SS |
1616 | * If the task is going to be woken-up on this cpu and if it is |
1617 | * already idle, then it is the right target. | |
a50bde51 | 1618 | */ |
99bd5e2f SS |
1619 | if (target == cpu && idle_cpu(cpu)) |
1620 | return cpu; | |
1621 | ||
1622 | /* | |
1623 | * If the task is going to be woken-up on the cpu where it previously | |
1624 | * ran and if it is currently idle, then it the right target. | |
1625 | */ | |
1626 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 1627 | return prev_cpu; |
a50bde51 PZ |
1628 | |
1629 | /* | |
99bd5e2f | 1630 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 1631 | */ |
dce840a0 | 1632 | rcu_read_lock(); |
99bd5e2f SS |
1633 | for_each_domain(target, sd) { |
1634 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 1635 | break; |
99bd5e2f SS |
1636 | |
1637 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1638 | if (idle_cpu(i)) { | |
1639 | target = i; | |
1640 | break; | |
1641 | } | |
a50bde51 | 1642 | } |
99bd5e2f SS |
1643 | |
1644 | /* | |
1645 | * Lets stop looking for an idle sibling when we reached | |
1646 | * the domain that spans the current cpu and prev_cpu. | |
1647 | */ | |
1648 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1649 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1650 | break; | |
a50bde51 | 1651 | } |
dce840a0 | 1652 | rcu_read_unlock(); |
a50bde51 PZ |
1653 | |
1654 | return target; | |
1655 | } | |
1656 | ||
aaee1203 PZ |
1657 | /* |
1658 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1659 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1660 | * SD_BALANCE_EXEC. | |
1661 | * | |
1662 | * Balance, ie. select the least loaded group. | |
1663 | * | |
1664 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1665 | * | |
1666 | * preempt must be disabled. | |
1667 | */ | |
0017d735 | 1668 | static int |
7608dec2 | 1669 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 1670 | { |
29cd8bae | 1671 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1672 | int cpu = smp_processor_id(); |
1673 | int prev_cpu = task_cpu(p); | |
1674 | int new_cpu = cpu; | |
99bd5e2f | 1675 | int want_affine = 0; |
29cd8bae | 1676 | int want_sd = 1; |
5158f4e4 | 1677 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1678 | |
0763a660 | 1679 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 1680 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
1681 | want_affine = 1; |
1682 | new_cpu = prev_cpu; | |
1683 | } | |
aaee1203 | 1684 | |
dce840a0 | 1685 | rcu_read_lock(); |
aaee1203 | 1686 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
1687 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1688 | continue; | |
1689 | ||
aaee1203 | 1690 | /* |
ae154be1 PZ |
1691 | * If power savings logic is enabled for a domain, see if we |
1692 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1693 | */ |
59abf026 | 1694 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1695 | unsigned long power = 0; |
1696 | unsigned long nr_running = 0; | |
1697 | unsigned long capacity; | |
1698 | int i; | |
1699 | ||
1700 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1701 | power += power_of(i); | |
1702 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1703 | } | |
1704 | ||
1399fa78 | 1705 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); |
ae154be1 | 1706 | |
59abf026 PZ |
1707 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1708 | nr_running /= 2; | |
1709 | ||
1710 | if (nr_running < capacity) | |
29cd8bae | 1711 | want_sd = 0; |
ae154be1 | 1712 | } |
aaee1203 | 1713 | |
fe3bcfe1 | 1714 | /* |
99bd5e2f SS |
1715 | * If both cpu and prev_cpu are part of this domain, |
1716 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 1717 | */ |
99bd5e2f SS |
1718 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1719 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1720 | affine_sd = tmp; | |
1721 | want_affine = 0; | |
c88d5910 PZ |
1722 | } |
1723 | ||
29cd8bae PZ |
1724 | if (!want_sd && !want_affine) |
1725 | break; | |
1726 | ||
0763a660 | 1727 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1728 | continue; |
1729 | ||
29cd8bae PZ |
1730 | if (want_sd) |
1731 | sd = tmp; | |
1732 | } | |
1733 | ||
8b911acd | 1734 | if (affine_sd) { |
99bd5e2f | 1735 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
1736 | prev_cpu = cpu; |
1737 | ||
1738 | new_cpu = select_idle_sibling(p, prev_cpu); | |
1739 | goto unlock; | |
8b911acd | 1740 | } |
e7693a36 | 1741 | |
aaee1203 | 1742 | while (sd) { |
5158f4e4 | 1743 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1744 | struct sched_group *group; |
c88d5910 | 1745 | int weight; |
098fb9db | 1746 | |
0763a660 | 1747 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1748 | sd = sd->child; |
1749 | continue; | |
1750 | } | |
098fb9db | 1751 | |
5158f4e4 PZ |
1752 | if (sd_flag & SD_BALANCE_WAKE) |
1753 | load_idx = sd->wake_idx; | |
098fb9db | 1754 | |
5158f4e4 | 1755 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1756 | if (!group) { |
1757 | sd = sd->child; | |
1758 | continue; | |
1759 | } | |
4ae7d5ce | 1760 | |
d7c33c49 | 1761 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1762 | if (new_cpu == -1 || new_cpu == cpu) { |
1763 | /* Now try balancing at a lower domain level of cpu */ | |
1764 | sd = sd->child; | |
1765 | continue; | |
e7693a36 | 1766 | } |
aaee1203 PZ |
1767 | |
1768 | /* Now try balancing at a lower domain level of new_cpu */ | |
1769 | cpu = new_cpu; | |
669c55e9 | 1770 | weight = sd->span_weight; |
aaee1203 PZ |
1771 | sd = NULL; |
1772 | for_each_domain(cpu, tmp) { | |
669c55e9 | 1773 | if (weight <= tmp->span_weight) |
aaee1203 | 1774 | break; |
0763a660 | 1775 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1776 | sd = tmp; |
1777 | } | |
1778 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 1779 | } |
dce840a0 PZ |
1780 | unlock: |
1781 | rcu_read_unlock(); | |
e7693a36 | 1782 | |
c88d5910 | 1783 | return new_cpu; |
e7693a36 GH |
1784 | } |
1785 | #endif /* CONFIG_SMP */ | |
1786 | ||
e52fb7c0 PZ |
1787 | static unsigned long |
1788 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1789 | { |
1790 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1791 | ||
1792 | /* | |
e52fb7c0 PZ |
1793 | * Since its curr running now, convert the gran from real-time |
1794 | * to virtual-time in his units. | |
13814d42 MG |
1795 | * |
1796 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1797 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1798 | * the resulting gran will be larger, therefore penalizing the | |
1799 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1800 | * be smaller, again penalizing the lighter task. | |
1801 | * | |
1802 | * This is especially important for buddies when the leftmost | |
1803 | * task is higher priority than the buddy. | |
0bbd3336 | 1804 | */ |
f4ad9bd2 | 1805 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
1806 | } |
1807 | ||
464b7527 PZ |
1808 | /* |
1809 | * Should 'se' preempt 'curr'. | |
1810 | * | |
1811 | * |s1 | |
1812 | * |s2 | |
1813 | * |s3 | |
1814 | * g | |
1815 | * |<--->|c | |
1816 | * | |
1817 | * w(c, s1) = -1 | |
1818 | * w(c, s2) = 0 | |
1819 | * w(c, s3) = 1 | |
1820 | * | |
1821 | */ | |
1822 | static int | |
1823 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1824 | { | |
1825 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1826 | ||
1827 | if (vdiff <= 0) | |
1828 | return -1; | |
1829 | ||
e52fb7c0 | 1830 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1831 | if (vdiff > gran) |
1832 | return 1; | |
1833 | ||
1834 | return 0; | |
1835 | } | |
1836 | ||
02479099 PZ |
1837 | static void set_last_buddy(struct sched_entity *se) |
1838 | { | |
69c80f3e VP |
1839 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
1840 | return; | |
1841 | ||
1842 | for_each_sched_entity(se) | |
1843 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
1844 | } |
1845 | ||
1846 | static void set_next_buddy(struct sched_entity *se) | |
1847 | { | |
69c80f3e VP |
1848 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
1849 | return; | |
1850 | ||
1851 | for_each_sched_entity(se) | |
1852 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
1853 | } |
1854 | ||
ac53db59 RR |
1855 | static void set_skip_buddy(struct sched_entity *se) |
1856 | { | |
69c80f3e VP |
1857 | for_each_sched_entity(se) |
1858 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
1859 | } |
1860 | ||
bf0f6f24 IM |
1861 | /* |
1862 | * Preempt the current task with a newly woken task if needed: | |
1863 | */ | |
5a9b86f6 | 1864 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1865 | { |
1866 | struct task_struct *curr = rq->curr; | |
8651a86c | 1867 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1868 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 1869 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 1870 | int next_buddy_marked = 0; |
bf0f6f24 | 1871 | |
4ae7d5ce IM |
1872 | if (unlikely(se == pse)) |
1873 | return; | |
1874 | ||
2f36825b | 1875 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 1876 | set_next_buddy(pse); |
2f36825b VP |
1877 | next_buddy_marked = 1; |
1878 | } | |
57fdc26d | 1879 | |
aec0a514 BR |
1880 | /* |
1881 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1882 | * wake up path. | |
1883 | */ | |
1884 | if (test_tsk_need_resched(curr)) | |
1885 | return; | |
1886 | ||
a2f5c9ab DH |
1887 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1888 | if (unlikely(curr->policy == SCHED_IDLE) && | |
1889 | likely(p->policy != SCHED_IDLE)) | |
1890 | goto preempt; | |
1891 | ||
91c234b4 | 1892 | /* |
a2f5c9ab DH |
1893 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
1894 | * is driven by the tick): | |
91c234b4 | 1895 | */ |
6bc912b7 | 1896 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1897 | return; |
bf0f6f24 | 1898 | |
464b7527 | 1899 | find_matching_se(&se, &pse); |
9bbd7374 | 1900 | update_curr(cfs_rq_of(se)); |
002f128b | 1901 | BUG_ON(!pse); |
2f36825b VP |
1902 | if (wakeup_preempt_entity(se, pse) == 1) { |
1903 | /* | |
1904 | * Bias pick_next to pick the sched entity that is | |
1905 | * triggering this preemption. | |
1906 | */ | |
1907 | if (!next_buddy_marked) | |
1908 | set_next_buddy(pse); | |
3a7e73a2 | 1909 | goto preempt; |
2f36825b | 1910 | } |
464b7527 | 1911 | |
3a7e73a2 | 1912 | return; |
a65ac745 | 1913 | |
3a7e73a2 PZ |
1914 | preempt: |
1915 | resched_task(curr); | |
1916 | /* | |
1917 | * Only set the backward buddy when the current task is still | |
1918 | * on the rq. This can happen when a wakeup gets interleaved | |
1919 | * with schedule on the ->pre_schedule() or idle_balance() | |
1920 | * point, either of which can * drop the rq lock. | |
1921 | * | |
1922 | * Also, during early boot the idle thread is in the fair class, | |
1923 | * for obvious reasons its a bad idea to schedule back to it. | |
1924 | */ | |
1925 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1926 | return; | |
1927 | ||
1928 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1929 | set_last_buddy(se); | |
bf0f6f24 IM |
1930 | } |
1931 | ||
fb8d4724 | 1932 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1933 | { |
8f4d37ec | 1934 | struct task_struct *p; |
bf0f6f24 IM |
1935 | struct cfs_rq *cfs_rq = &rq->cfs; |
1936 | struct sched_entity *se; | |
1937 | ||
36ace27e | 1938 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
1939 | return NULL; |
1940 | ||
1941 | do { | |
9948f4b2 | 1942 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1943 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1944 | cfs_rq = group_cfs_rq(se); |
1945 | } while (cfs_rq); | |
1946 | ||
8f4d37ec PZ |
1947 | p = task_of(se); |
1948 | hrtick_start_fair(rq, p); | |
1949 | ||
1950 | return p; | |
bf0f6f24 IM |
1951 | } |
1952 | ||
1953 | /* | |
1954 | * Account for a descheduled task: | |
1955 | */ | |
31ee529c | 1956 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1957 | { |
1958 | struct sched_entity *se = &prev->se; | |
1959 | struct cfs_rq *cfs_rq; | |
1960 | ||
1961 | for_each_sched_entity(se) { | |
1962 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1963 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1964 | } |
1965 | } | |
1966 | ||
ac53db59 RR |
1967 | /* |
1968 | * sched_yield() is very simple | |
1969 | * | |
1970 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
1971 | */ | |
1972 | static void yield_task_fair(struct rq *rq) | |
1973 | { | |
1974 | struct task_struct *curr = rq->curr; | |
1975 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1976 | struct sched_entity *se = &curr->se; | |
1977 | ||
1978 | /* | |
1979 | * Are we the only task in the tree? | |
1980 | */ | |
1981 | if (unlikely(rq->nr_running == 1)) | |
1982 | return; | |
1983 | ||
1984 | clear_buddies(cfs_rq, se); | |
1985 | ||
1986 | if (curr->policy != SCHED_BATCH) { | |
1987 | update_rq_clock(rq); | |
1988 | /* | |
1989 | * Update run-time statistics of the 'current'. | |
1990 | */ | |
1991 | update_curr(cfs_rq); | |
1992 | } | |
1993 | ||
1994 | set_skip_buddy(se); | |
1995 | } | |
1996 | ||
d95f4122 MG |
1997 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
1998 | { | |
1999 | struct sched_entity *se = &p->se; | |
2000 | ||
2001 | if (!se->on_rq) | |
2002 | return false; | |
2003 | ||
2004 | /* Tell the scheduler that we'd really like pse to run next. */ | |
2005 | set_next_buddy(se); | |
2006 | ||
d95f4122 MG |
2007 | yield_task_fair(rq); |
2008 | ||
2009 | return true; | |
2010 | } | |
2011 | ||
681f3e68 | 2012 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
2013 | /************************************************** |
2014 | * Fair scheduling class load-balancing methods: | |
2015 | */ | |
2016 | ||
1e3c88bd PZ |
2017 | /* |
2018 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
2019 | * Both runqueues must be locked. | |
2020 | */ | |
2021 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
2022 | struct rq *this_rq, int this_cpu) | |
2023 | { | |
2024 | deactivate_task(src_rq, p, 0); | |
2025 | set_task_cpu(p, this_cpu); | |
2026 | activate_task(this_rq, p, 0); | |
2027 | check_preempt_curr(this_rq, p, 0); | |
2028 | } | |
2029 | ||
2030 | /* | |
2031 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2032 | */ | |
2033 | static | |
2034 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
2035 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2036 | int *all_pinned) | |
2037 | { | |
2038 | int tsk_cache_hot = 0; | |
2039 | /* | |
2040 | * We do not migrate tasks that are: | |
2041 | * 1) running (obviously), or | |
2042 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2043 | * 3) are cache-hot on their current CPU. | |
2044 | */ | |
2045 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 2046 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
2047 | return 0; |
2048 | } | |
2049 | *all_pinned = 0; | |
2050 | ||
2051 | if (task_running(rq, p)) { | |
41acab88 | 2052 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
2053 | return 0; |
2054 | } | |
2055 | ||
2056 | /* | |
2057 | * Aggressive migration if: | |
2058 | * 1) task is cache cold, or | |
2059 | * 2) too many balance attempts have failed. | |
2060 | */ | |
2061 | ||
305e6835 | 2062 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
2063 | if (!tsk_cache_hot || |
2064 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
2065 | #ifdef CONFIG_SCHEDSTATS | |
2066 | if (tsk_cache_hot) { | |
2067 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 2068 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
2069 | } |
2070 | #endif | |
2071 | return 1; | |
2072 | } | |
2073 | ||
2074 | if (tsk_cache_hot) { | |
41acab88 | 2075 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
2076 | return 0; |
2077 | } | |
2078 | return 1; | |
2079 | } | |
2080 | ||
897c395f PZ |
2081 | /* |
2082 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
2083 | * part of active balancing operations within "domain". | |
2084 | * Returns 1 if successful and 0 otherwise. | |
2085 | * | |
2086 | * Called with both runqueues locked. | |
2087 | */ | |
2088 | static int | |
2089 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2090 | struct sched_domain *sd, enum cpu_idle_type idle) | |
2091 | { | |
2092 | struct task_struct *p, *n; | |
2093 | struct cfs_rq *cfs_rq; | |
2094 | int pinned = 0; | |
2095 | ||
2096 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
2097 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
2098 | ||
2099 | if (!can_migrate_task(p, busiest, this_cpu, | |
2100 | sd, idle, &pinned)) | |
2101 | continue; | |
2102 | ||
2103 | pull_task(busiest, p, this_rq, this_cpu); | |
2104 | /* | |
2105 | * Right now, this is only the second place pull_task() | |
2106 | * is called, so we can safely collect pull_task() | |
2107 | * stats here rather than inside pull_task(). | |
2108 | */ | |
2109 | schedstat_inc(sd, lb_gained[idle]); | |
2110 | return 1; | |
2111 | } | |
2112 | } | |
2113 | ||
2114 | return 0; | |
2115 | } | |
2116 | ||
1e3c88bd PZ |
2117 | static unsigned long |
2118 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2119 | unsigned long max_load_move, struct sched_domain *sd, | |
2120 | enum cpu_idle_type idle, int *all_pinned, | |
931aeeda | 2121 | struct cfs_rq *busiest_cfs_rq) |
1e3c88bd | 2122 | { |
b30aef17 | 2123 | int loops = 0, pulled = 0; |
1e3c88bd | 2124 | long rem_load_move = max_load_move; |
ee00e66f | 2125 | struct task_struct *p, *n; |
1e3c88bd PZ |
2126 | |
2127 | if (max_load_move == 0) | |
2128 | goto out; | |
2129 | ||
ee00e66f PZ |
2130 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
2131 | if (loops++ > sysctl_sched_nr_migrate) | |
2132 | break; | |
1e3c88bd | 2133 | |
ee00e66f | 2134 | if ((p->se.load.weight >> 1) > rem_load_move || |
b30aef17 KC |
2135 | !can_migrate_task(p, busiest, this_cpu, sd, idle, |
2136 | all_pinned)) | |
ee00e66f | 2137 | continue; |
1e3c88bd | 2138 | |
ee00e66f PZ |
2139 | pull_task(busiest, p, this_rq, this_cpu); |
2140 | pulled++; | |
2141 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
2142 | |
2143 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
2144 | /* |
2145 | * NEWIDLE balancing is a source of latency, so preemptible | |
2146 | * kernels will stop after the first task is pulled to minimize | |
2147 | * the critical section. | |
2148 | */ | |
2149 | if (idle == CPU_NEWLY_IDLE) | |
2150 | break; | |
1e3c88bd PZ |
2151 | #endif |
2152 | ||
ee00e66f PZ |
2153 | /* |
2154 | * We only want to steal up to the prescribed amount of | |
2155 | * weighted load. | |
2156 | */ | |
2157 | if (rem_load_move <= 0) | |
2158 | break; | |
1e3c88bd PZ |
2159 | } |
2160 | out: | |
2161 | /* | |
2162 | * Right now, this is one of only two places pull_task() is called, | |
2163 | * so we can safely collect pull_task() stats here rather than | |
2164 | * inside pull_task(). | |
2165 | */ | |
2166 | schedstat_add(sd, lb_gained[idle], pulled); | |
2167 | ||
1e3c88bd PZ |
2168 | return max_load_move - rem_load_move; |
2169 | } | |
2170 | ||
230059de | 2171 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
2172 | /* |
2173 | * update tg->load_weight by folding this cpu's load_avg | |
2174 | */ | |
67e86250 | 2175 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
2176 | { |
2177 | struct cfs_rq *cfs_rq; | |
2178 | unsigned long flags; | |
2179 | struct rq *rq; | |
9e3081ca PZ |
2180 | |
2181 | if (!tg->se[cpu]) | |
2182 | return 0; | |
2183 | ||
2184 | rq = cpu_rq(cpu); | |
2185 | cfs_rq = tg->cfs_rq[cpu]; | |
2186 | ||
2187 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2188 | ||
2189 | update_rq_clock(rq); | |
d6b55918 | 2190 | update_cfs_load(cfs_rq, 1); |
9e3081ca PZ |
2191 | |
2192 | /* | |
2193 | * We need to update shares after updating tg->load_weight in | |
2194 | * order to adjust the weight of groups with long running tasks. | |
2195 | */ | |
6d5ab293 | 2196 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
2197 | |
2198 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2199 | ||
2200 | return 0; | |
2201 | } | |
2202 | ||
2203 | static void update_shares(int cpu) | |
2204 | { | |
2205 | struct cfs_rq *cfs_rq; | |
2206 | struct rq *rq = cpu_rq(cpu); | |
2207 | ||
2208 | rcu_read_lock(); | |
9763b67f PZ |
2209 | /* |
2210 | * Iterates the task_group tree in a bottom up fashion, see | |
2211 | * list_add_leaf_cfs_rq() for details. | |
2212 | */ | |
67e86250 PT |
2213 | for_each_leaf_cfs_rq(rq, cfs_rq) |
2214 | update_shares_cpu(cfs_rq->tg, cpu); | |
9e3081ca PZ |
2215 | rcu_read_unlock(); |
2216 | } | |
2217 | ||
9763b67f PZ |
2218 | /* |
2219 | * Compute the cpu's hierarchical load factor for each task group. | |
2220 | * This needs to be done in a top-down fashion because the load of a child | |
2221 | * group is a fraction of its parents load. | |
2222 | */ | |
2223 | static int tg_load_down(struct task_group *tg, void *data) | |
2224 | { | |
2225 | unsigned long load; | |
2226 | long cpu = (long)data; | |
2227 | ||
2228 | if (!tg->parent) { | |
2229 | load = cpu_rq(cpu)->load.weight; | |
2230 | } else { | |
2231 | load = tg->parent->cfs_rq[cpu]->h_load; | |
2232 | load *= tg->se[cpu]->load.weight; | |
2233 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
2234 | } | |
2235 | ||
2236 | tg->cfs_rq[cpu]->h_load = load; | |
2237 | ||
2238 | return 0; | |
2239 | } | |
2240 | ||
2241 | static void update_h_load(long cpu) | |
2242 | { | |
2243 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); | |
2244 | } | |
2245 | ||
230059de PZ |
2246 | static unsigned long |
2247 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2248 | unsigned long max_load_move, | |
2249 | struct sched_domain *sd, enum cpu_idle_type idle, | |
931aeeda | 2250 | int *all_pinned) |
230059de PZ |
2251 | { |
2252 | long rem_load_move = max_load_move; | |
9763b67f | 2253 | struct cfs_rq *busiest_cfs_rq; |
230059de PZ |
2254 | |
2255 | rcu_read_lock(); | |
9763b67f | 2256 | update_h_load(cpu_of(busiest)); |
230059de | 2257 | |
9763b67f | 2258 | for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) { |
230059de PZ |
2259 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
2260 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
2261 | u64 rem_load, moved_load; | |
2262 | ||
2263 | /* | |
2264 | * empty group | |
2265 | */ | |
2266 | if (!busiest_cfs_rq->task_weight) | |
2267 | continue; | |
2268 | ||
2269 | rem_load = (u64)rem_load_move * busiest_weight; | |
2270 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
2271 | ||
2272 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
931aeeda | 2273 | rem_load, sd, idle, all_pinned, |
230059de PZ |
2274 | busiest_cfs_rq); |
2275 | ||
2276 | if (!moved_load) | |
2277 | continue; | |
2278 | ||
2279 | moved_load *= busiest_h_load; | |
2280 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
2281 | ||
2282 | rem_load_move -= moved_load; | |
2283 | if (rem_load_move < 0) | |
2284 | break; | |
2285 | } | |
2286 | rcu_read_unlock(); | |
2287 | ||
2288 | return max_load_move - rem_load_move; | |
2289 | } | |
2290 | #else | |
9e3081ca PZ |
2291 | static inline void update_shares(int cpu) |
2292 | { | |
2293 | } | |
2294 | ||
230059de PZ |
2295 | static unsigned long |
2296 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2297 | unsigned long max_load_move, | |
2298 | struct sched_domain *sd, enum cpu_idle_type idle, | |
931aeeda | 2299 | int *all_pinned) |
230059de PZ |
2300 | { |
2301 | return balance_tasks(this_rq, this_cpu, busiest, | |
2302 | max_load_move, sd, idle, all_pinned, | |
931aeeda | 2303 | &busiest->cfs); |
230059de PZ |
2304 | } |
2305 | #endif | |
2306 | ||
1e3c88bd PZ |
2307 | /* |
2308 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2309 | * this_rq, as part of a balancing operation within domain "sd". | |
2310 | * Returns 1 if successful and 0 otherwise. | |
2311 | * | |
2312 | * Called with both runqueues locked. | |
2313 | */ | |
2314 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2315 | unsigned long max_load_move, | |
2316 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2317 | int *all_pinned) | |
2318 | { | |
3d45fd80 | 2319 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2320 | |
2321 | do { | |
3d45fd80 | 2322 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd | 2323 | max_load_move - total_load_moved, |
931aeeda | 2324 | sd, idle, all_pinned); |
3d45fd80 PZ |
2325 | |
2326 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2327 | |
2328 | #ifdef CONFIG_PREEMPT | |
2329 | /* | |
2330 | * NEWIDLE balancing is a source of latency, so preemptible | |
2331 | * kernels will stop after the first task is pulled to minimize | |
2332 | * the critical section. | |
2333 | */ | |
2334 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2335 | break; | |
baa8c110 PZ |
2336 | |
2337 | if (raw_spin_is_contended(&this_rq->lock) || | |
2338 | raw_spin_is_contended(&busiest->lock)) | |
2339 | break; | |
1e3c88bd | 2340 | #endif |
3d45fd80 | 2341 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2342 | |
2343 | return total_load_moved > 0; | |
2344 | } | |
2345 | ||
1e3c88bd PZ |
2346 | /********** Helpers for find_busiest_group ************************/ |
2347 | /* | |
2348 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2349 | * during load balancing. | |
2350 | */ | |
2351 | struct sd_lb_stats { | |
2352 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2353 | struct sched_group *this; /* Local group in this sd */ | |
2354 | unsigned long total_load; /* Total load of all groups in sd */ | |
2355 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2356 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2357 | ||
2358 | /** Statistics of this group */ | |
2359 | unsigned long this_load; | |
2360 | unsigned long this_load_per_task; | |
2361 | unsigned long this_nr_running; | |
fab47622 | 2362 | unsigned long this_has_capacity; |
aae6d3dd | 2363 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
2364 | |
2365 | /* Statistics of the busiest group */ | |
aae6d3dd | 2366 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
2367 | unsigned long max_load; |
2368 | unsigned long busiest_load_per_task; | |
2369 | unsigned long busiest_nr_running; | |
dd5feea1 | 2370 | unsigned long busiest_group_capacity; |
fab47622 | 2371 | unsigned long busiest_has_capacity; |
aae6d3dd | 2372 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
2373 | |
2374 | int group_imb; /* Is there imbalance in this sd */ | |
2375 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2376 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2377 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2378 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2379 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2380 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2381 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2382 | #endif | |
2383 | }; | |
2384 | ||
2385 | /* | |
2386 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2387 | */ | |
2388 | struct sg_lb_stats { | |
2389 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2390 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2391 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2392 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2393 | unsigned long group_capacity; | |
aae6d3dd SS |
2394 | unsigned long idle_cpus; |
2395 | unsigned long group_weight; | |
1e3c88bd | 2396 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 2397 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2398 | }; |
2399 | ||
2400 | /** | |
2401 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2402 | * @group: The group whose first cpu is to be returned. | |
2403 | */ | |
2404 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2405 | { | |
2406 | return cpumask_first(sched_group_cpus(group)); | |
2407 | } | |
2408 | ||
2409 | /** | |
2410 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2411 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2412 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2413 | */ | |
2414 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2415 | enum cpu_idle_type idle) | |
2416 | { | |
2417 | int load_idx; | |
2418 | ||
2419 | switch (idle) { | |
2420 | case CPU_NOT_IDLE: | |
2421 | load_idx = sd->busy_idx; | |
2422 | break; | |
2423 | ||
2424 | case CPU_NEWLY_IDLE: | |
2425 | load_idx = sd->newidle_idx; | |
2426 | break; | |
2427 | default: | |
2428 | load_idx = sd->idle_idx; | |
2429 | break; | |
2430 | } | |
2431 | ||
2432 | return load_idx; | |
2433 | } | |
2434 | ||
2435 | ||
2436 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2437 | /** | |
2438 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2439 | * the given sched_domain, during load balancing. | |
2440 | * | |
2441 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2442 | * @sds: Variable containing the statistics for sd. | |
2443 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2444 | */ | |
2445 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2446 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2447 | { | |
2448 | /* | |
2449 | * Busy processors will not participate in power savings | |
2450 | * balance. | |
2451 | */ | |
2452 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2453 | sds->power_savings_balance = 0; | |
2454 | else { | |
2455 | sds->power_savings_balance = 1; | |
2456 | sds->min_nr_running = ULONG_MAX; | |
2457 | sds->leader_nr_running = 0; | |
2458 | } | |
2459 | } | |
2460 | ||
2461 | /** | |
2462 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2463 | * sched_domain while performing load balancing. | |
2464 | * | |
2465 | * @group: sched_group belonging to the sched_domain under consideration. | |
2466 | * @sds: Variable containing the statistics of the sched_domain | |
2467 | * @local_group: Does group contain the CPU for which we're performing | |
2468 | * load balancing ? | |
2469 | * @sgs: Variable containing the statistics of the group. | |
2470 | */ | |
2471 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2472 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2473 | { | |
2474 | ||
2475 | if (!sds->power_savings_balance) | |
2476 | return; | |
2477 | ||
2478 | /* | |
2479 | * If the local group is idle or completely loaded | |
2480 | * no need to do power savings balance at this domain | |
2481 | */ | |
2482 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2483 | !sds->this_nr_running)) | |
2484 | sds->power_savings_balance = 0; | |
2485 | ||
2486 | /* | |
2487 | * If a group is already running at full capacity or idle, | |
2488 | * don't include that group in power savings calculations | |
2489 | */ | |
2490 | if (!sds->power_savings_balance || | |
2491 | sgs->sum_nr_running >= sgs->group_capacity || | |
2492 | !sgs->sum_nr_running) | |
2493 | return; | |
2494 | ||
2495 | /* | |
2496 | * Calculate the group which has the least non-idle load. | |
2497 | * This is the group from where we need to pick up the load | |
2498 | * for saving power | |
2499 | */ | |
2500 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2501 | (sgs->sum_nr_running == sds->min_nr_running && | |
2502 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2503 | sds->group_min = group; | |
2504 | sds->min_nr_running = sgs->sum_nr_running; | |
2505 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2506 | sgs->sum_nr_running; | |
2507 | } | |
2508 | ||
2509 | /* | |
2510 | * Calculate the group which is almost near its | |
2511 | * capacity but still has some space to pick up some load | |
2512 | * from other group and save more power | |
2513 | */ | |
2514 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2515 | return; | |
2516 | ||
2517 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2518 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2519 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2520 | sds->group_leader = group; | |
2521 | sds->leader_nr_running = sgs->sum_nr_running; | |
2522 | } | |
2523 | } | |
2524 | ||
2525 | /** | |
2526 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2527 | * @sds: Variable containing the statistics of the sched_domain | |
2528 | * under consideration. | |
2529 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2530 | * @imbalance: Variable to store the imbalance. | |
2531 | * | |
2532 | * Description: | |
2533 | * Check if we have potential to perform some power-savings balance. | |
2534 | * If yes, set the busiest group to be the least loaded group in the | |
2535 | * sched_domain, so that it's CPUs can be put to idle. | |
2536 | * | |
2537 | * Returns 1 if there is potential to perform power-savings balance. | |
2538 | * Else returns 0. | |
2539 | */ | |
2540 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2541 | int this_cpu, unsigned long *imbalance) | |
2542 | { | |
2543 | if (!sds->power_savings_balance) | |
2544 | return 0; | |
2545 | ||
2546 | if (sds->this != sds->group_leader || | |
2547 | sds->group_leader == sds->group_min) | |
2548 | return 0; | |
2549 | ||
2550 | *imbalance = sds->min_load_per_task; | |
2551 | sds->busiest = sds->group_min; | |
2552 | ||
2553 | return 1; | |
2554 | ||
2555 | } | |
2556 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2557 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2558 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2559 | { | |
2560 | return; | |
2561 | } | |
2562 | ||
2563 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2564 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2565 | { | |
2566 | return; | |
2567 | } | |
2568 | ||
2569 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2570 | int this_cpu, unsigned long *imbalance) | |
2571 | { | |
2572 | return 0; | |
2573 | } | |
2574 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2575 | ||
2576 | ||
2577 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2578 | { | |
1399fa78 | 2579 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
2580 | } |
2581 | ||
2582 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2583 | { | |
2584 | return default_scale_freq_power(sd, cpu); | |
2585 | } | |
2586 | ||
2587 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2588 | { | |
669c55e9 | 2589 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2590 | unsigned long smt_gain = sd->smt_gain; |
2591 | ||
2592 | smt_gain /= weight; | |
2593 | ||
2594 | return smt_gain; | |
2595 | } | |
2596 | ||
2597 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2598 | { | |
2599 | return default_scale_smt_power(sd, cpu); | |
2600 | } | |
2601 | ||
2602 | unsigned long scale_rt_power(int cpu) | |
2603 | { | |
2604 | struct rq *rq = cpu_rq(cpu); | |
2605 | u64 total, available; | |
2606 | ||
1e3c88bd | 2607 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
2608 | |
2609 | if (unlikely(total < rq->rt_avg)) { | |
2610 | /* Ensures that power won't end up being negative */ | |
2611 | available = 0; | |
2612 | } else { | |
2613 | available = total - rq->rt_avg; | |
2614 | } | |
1e3c88bd | 2615 | |
1399fa78 NR |
2616 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
2617 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 2618 | |
1399fa78 | 2619 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
2620 | |
2621 | return div_u64(available, total); | |
2622 | } | |
2623 | ||
2624 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2625 | { | |
669c55e9 | 2626 | unsigned long weight = sd->span_weight; |
1399fa78 | 2627 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
2628 | struct sched_group *sdg = sd->groups; |
2629 | ||
1e3c88bd PZ |
2630 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
2631 | if (sched_feat(ARCH_POWER)) | |
2632 | power *= arch_scale_smt_power(sd, cpu); | |
2633 | else | |
2634 | power *= default_scale_smt_power(sd, cpu); | |
2635 | ||
1399fa78 | 2636 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
2637 | } |
2638 | ||
9c3f75cb | 2639 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
2640 | |
2641 | if (sched_feat(ARCH_POWER)) | |
2642 | power *= arch_scale_freq_power(sd, cpu); | |
2643 | else | |
2644 | power *= default_scale_freq_power(sd, cpu); | |
2645 | ||
1399fa78 | 2646 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 2647 | |
1e3c88bd | 2648 | power *= scale_rt_power(cpu); |
1399fa78 | 2649 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
2650 | |
2651 | if (!power) | |
2652 | power = 1; | |
2653 | ||
e51fd5e2 | 2654 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 2655 | sdg->sgp->power = power; |
1e3c88bd PZ |
2656 | } |
2657 | ||
2658 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2659 | { | |
2660 | struct sched_domain *child = sd->child; | |
2661 | struct sched_group *group, *sdg = sd->groups; | |
2662 | unsigned long power; | |
2663 | ||
2664 | if (!child) { | |
2665 | update_cpu_power(sd, cpu); | |
2666 | return; | |
2667 | } | |
2668 | ||
2669 | power = 0; | |
2670 | ||
2671 | group = child->groups; | |
2672 | do { | |
9c3f75cb | 2673 | power += group->sgp->power; |
1e3c88bd PZ |
2674 | group = group->next; |
2675 | } while (group != child->groups); | |
2676 | ||
9c3f75cb | 2677 | sdg->sgp->power = power; |
1e3c88bd PZ |
2678 | } |
2679 | ||
9d5efe05 SV |
2680 | /* |
2681 | * Try and fix up capacity for tiny siblings, this is needed when | |
2682 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
2683 | * which on its own isn't powerful enough. | |
2684 | * | |
2685 | * See update_sd_pick_busiest() and check_asym_packing(). | |
2686 | */ | |
2687 | static inline int | |
2688 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
2689 | { | |
2690 | /* | |
1399fa78 | 2691 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 2692 | */ |
a6c75f2f | 2693 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
2694 | return 0; |
2695 | ||
2696 | /* | |
2697 | * If ~90% of the cpu_power is still there, we're good. | |
2698 | */ | |
9c3f75cb | 2699 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
2700 | return 1; |
2701 | ||
2702 | return 0; | |
2703 | } | |
2704 | ||
1e3c88bd PZ |
2705 | /** |
2706 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2707 | * @sd: The sched_domain whose statistics are to be updated. | |
2708 | * @group: sched_group whose statistics are to be updated. | |
2709 | * @this_cpu: Cpu for which load balance is currently performed. | |
2710 | * @idle: Idle status of this_cpu | |
2711 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
1e3c88bd PZ |
2712 | * @local_group: Does group contain this_cpu. |
2713 | * @cpus: Set of cpus considered for load balancing. | |
2714 | * @balance: Should we balance. | |
2715 | * @sgs: variable to hold the statistics for this group. | |
2716 | */ | |
2717 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2718 | struct sched_group *group, int this_cpu, | |
46e49b38 | 2719 | enum cpu_idle_type idle, int load_idx, |
1e3c88bd PZ |
2720 | int local_group, const struct cpumask *cpus, |
2721 | int *balance, struct sg_lb_stats *sgs) | |
2722 | { | |
2582f0eb | 2723 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
2724 | int i; |
2725 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 2726 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 2727 | |
871e35bc | 2728 | if (local_group) |
1e3c88bd | 2729 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
2730 | |
2731 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
2732 | max_cpu_load = 0; |
2733 | min_cpu_load = ~0UL; | |
2582f0eb | 2734 | max_nr_running = 0; |
1e3c88bd PZ |
2735 | |
2736 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2737 | struct rq *rq = cpu_rq(i); | |
2738 | ||
1e3c88bd PZ |
2739 | /* Bias balancing toward cpus of our domain */ |
2740 | if (local_group) { | |
2741 | if (idle_cpu(i) && !first_idle_cpu) { | |
2742 | first_idle_cpu = 1; | |
2743 | balance_cpu = i; | |
2744 | } | |
2745 | ||
2746 | load = target_load(i, load_idx); | |
2747 | } else { | |
2748 | load = source_load(i, load_idx); | |
2582f0eb | 2749 | if (load > max_cpu_load) { |
1e3c88bd | 2750 | max_cpu_load = load; |
2582f0eb NR |
2751 | max_nr_running = rq->nr_running; |
2752 | } | |
1e3c88bd PZ |
2753 | if (min_cpu_load > load) |
2754 | min_cpu_load = load; | |
2755 | } | |
2756 | ||
2757 | sgs->group_load += load; | |
2758 | sgs->sum_nr_running += rq->nr_running; | |
2759 | sgs->sum_weighted_load += weighted_cpuload(i); | |
aae6d3dd SS |
2760 | if (idle_cpu(i)) |
2761 | sgs->idle_cpus++; | |
1e3c88bd PZ |
2762 | } |
2763 | ||
2764 | /* | |
2765 | * First idle cpu or the first cpu(busiest) in this sched group | |
2766 | * is eligible for doing load balancing at this and above | |
2767 | * domains. In the newly idle case, we will allow all the cpu's | |
2768 | * to do the newly idle load balance. | |
2769 | */ | |
bbc8cb5b PZ |
2770 | if (idle != CPU_NEWLY_IDLE && local_group) { |
2771 | if (balance_cpu != this_cpu) { | |
2772 | *balance = 0; | |
2773 | return; | |
2774 | } | |
2775 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
2776 | } |
2777 | ||
2778 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2779 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 2780 | |
1e3c88bd PZ |
2781 | /* |
2782 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 2783 | * than the average weight of a task. |
1e3c88bd PZ |
2784 | * |
2785 | * APZ: with cgroup the avg task weight can vary wildly and | |
2786 | * might not be a suitable number - should we keep a | |
2787 | * normalized nr_running number somewhere that negates | |
2788 | * the hierarchy? | |
2789 | */ | |
dd5feea1 SS |
2790 | if (sgs->sum_nr_running) |
2791 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 2792 | |
866ab43e | 2793 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
2794 | sgs->group_imb = 1; |
2795 | ||
9c3f75cb | 2796 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 2797 | SCHED_POWER_SCALE); |
9d5efe05 SV |
2798 | if (!sgs->group_capacity) |
2799 | sgs->group_capacity = fix_small_capacity(sd, group); | |
aae6d3dd | 2800 | sgs->group_weight = group->group_weight; |
fab47622 NR |
2801 | |
2802 | if (sgs->group_capacity > sgs->sum_nr_running) | |
2803 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
2804 | } |
2805 | ||
532cb4c4 MN |
2806 | /** |
2807 | * update_sd_pick_busiest - return 1 on busiest group | |
2808 | * @sd: sched_domain whose statistics are to be checked | |
2809 | * @sds: sched_domain statistics | |
2810 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
2811 | * @sgs: sched_group statistics |
2812 | * @this_cpu: the current cpu | |
532cb4c4 MN |
2813 | * |
2814 | * Determine if @sg is a busier group than the previously selected | |
2815 | * busiest group. | |
2816 | */ | |
2817 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
2818 | struct sd_lb_stats *sds, | |
2819 | struct sched_group *sg, | |
2820 | struct sg_lb_stats *sgs, | |
2821 | int this_cpu) | |
2822 | { | |
2823 | if (sgs->avg_load <= sds->max_load) | |
2824 | return false; | |
2825 | ||
2826 | if (sgs->sum_nr_running > sgs->group_capacity) | |
2827 | return true; | |
2828 | ||
2829 | if (sgs->group_imb) | |
2830 | return true; | |
2831 | ||
2832 | /* | |
2833 | * ASYM_PACKING needs to move all the work to the lowest | |
2834 | * numbered CPUs in the group, therefore mark all groups | |
2835 | * higher than ourself as busy. | |
2836 | */ | |
2837 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
2838 | this_cpu < group_first_cpu(sg)) { | |
2839 | if (!sds->busiest) | |
2840 | return true; | |
2841 | ||
2842 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
2843 | return true; | |
2844 | } | |
2845 | ||
2846 | return false; | |
2847 | } | |
2848 | ||
1e3c88bd PZ |
2849 | /** |
2850 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2851 | * @sd: sched_domain whose statistics are to be updated. | |
2852 | * @this_cpu: Cpu for which load balance is currently performed. | |
2853 | * @idle: Idle status of this_cpu | |
1e3c88bd PZ |
2854 | * @cpus: Set of cpus considered for load balancing. |
2855 | * @balance: Should we balance. | |
2856 | * @sds: variable to hold the statistics for this sched_domain. | |
2857 | */ | |
2858 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
46e49b38 VP |
2859 | enum cpu_idle_type idle, const struct cpumask *cpus, |
2860 | int *balance, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
2861 | { |
2862 | struct sched_domain *child = sd->child; | |
532cb4c4 | 2863 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
2864 | struct sg_lb_stats sgs; |
2865 | int load_idx, prefer_sibling = 0; | |
2866 | ||
2867 | if (child && child->flags & SD_PREFER_SIBLING) | |
2868 | prefer_sibling = 1; | |
2869 | ||
2870 | init_sd_power_savings_stats(sd, sds, idle); | |
2871 | load_idx = get_sd_load_idx(sd, idle); | |
2872 | ||
2873 | do { | |
2874 | int local_group; | |
2875 | ||
532cb4c4 | 2876 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 2877 | memset(&sgs, 0, sizeof(sgs)); |
46e49b38 | 2878 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, |
1e3c88bd PZ |
2879 | local_group, cpus, balance, &sgs); |
2880 | ||
8f190fb3 | 2881 | if (local_group && !(*balance)) |
1e3c88bd PZ |
2882 | return; |
2883 | ||
2884 | sds->total_load += sgs.group_load; | |
9c3f75cb | 2885 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
2886 | |
2887 | /* | |
2888 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 2889 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
2890 | * and move all the excess tasks away. We lower the capacity |
2891 | * of a group only if the local group has the capacity to fit | |
2892 | * these excess tasks, i.e. nr_running < group_capacity. The | |
2893 | * extra check prevents the case where you always pull from the | |
2894 | * heaviest group when it is already under-utilized (possible | |
2895 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 2896 | */ |
75dd321d | 2897 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
2898 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
2899 | ||
2900 | if (local_group) { | |
2901 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 2902 | sds->this = sg; |
1e3c88bd PZ |
2903 | sds->this_nr_running = sgs.sum_nr_running; |
2904 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 2905 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 2906 | sds->this_idle_cpus = sgs.idle_cpus; |
532cb4c4 | 2907 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 2908 | sds->max_load = sgs.avg_load; |
532cb4c4 | 2909 | sds->busiest = sg; |
1e3c88bd | 2910 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 2911 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 2912 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 2913 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 2914 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 2915 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
2916 | sds->group_imb = sgs.group_imb; |
2917 | } | |
2918 | ||
532cb4c4 MN |
2919 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
2920 | sg = sg->next; | |
2921 | } while (sg != sd->groups); | |
2922 | } | |
2923 | ||
2ec57d44 | 2924 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
2925 | { |
2926 | return 0*SD_ASYM_PACKING; | |
2927 | } | |
2928 | ||
2929 | /** | |
2930 | * check_asym_packing - Check to see if the group is packed into the | |
2931 | * sched doman. | |
2932 | * | |
2933 | * This is primarily intended to used at the sibling level. Some | |
2934 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
2935 | * case of POWER7, it can move to lower SMT modes only when higher | |
2936 | * threads are idle. When in lower SMT modes, the threads will | |
2937 | * perform better since they share less core resources. Hence when we | |
2938 | * have idle threads, we want them to be the higher ones. | |
2939 | * | |
2940 | * This packing function is run on idle threads. It checks to see if | |
2941 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
2942 | * CPU number than the packing function is being run on. Here we are | |
2943 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
2944 | * number. | |
2945 | * | |
b6b12294 MN |
2946 | * Returns 1 when packing is required and a task should be moved to |
2947 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
2948 | * | |
532cb4c4 MN |
2949 | * @sd: The sched_domain whose packing is to be checked. |
2950 | * @sds: Statistics of the sched_domain which is to be packed | |
2951 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2952 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
2953 | */ |
2954 | static int check_asym_packing(struct sched_domain *sd, | |
2955 | struct sd_lb_stats *sds, | |
2956 | int this_cpu, unsigned long *imbalance) | |
2957 | { | |
2958 | int busiest_cpu; | |
2959 | ||
2960 | if (!(sd->flags & SD_ASYM_PACKING)) | |
2961 | return 0; | |
2962 | ||
2963 | if (!sds->busiest) | |
2964 | return 0; | |
2965 | ||
2966 | busiest_cpu = group_first_cpu(sds->busiest); | |
2967 | if (this_cpu > busiest_cpu) | |
2968 | return 0; | |
2969 | ||
9c3f75cb | 2970 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power, |
1399fa78 | 2971 | SCHED_POWER_SCALE); |
532cb4c4 | 2972 | return 1; |
1e3c88bd PZ |
2973 | } |
2974 | ||
2975 | /** | |
2976 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2977 | * amongst the groups of a sched_domain, during | |
2978 | * load balancing. | |
2979 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2980 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2981 | * @imbalance: Variable to store the imbalance. | |
2982 | */ | |
2983 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2984 | int this_cpu, unsigned long *imbalance) | |
2985 | { | |
2986 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2987 | unsigned int imbn = 2; | |
dd5feea1 | 2988 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
2989 | |
2990 | if (sds->this_nr_running) { | |
2991 | sds->this_load_per_task /= sds->this_nr_running; | |
2992 | if (sds->busiest_load_per_task > | |
2993 | sds->this_load_per_task) | |
2994 | imbn = 1; | |
2995 | } else | |
2996 | sds->this_load_per_task = | |
2997 | cpu_avg_load_per_task(this_cpu); | |
2998 | ||
dd5feea1 | 2999 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 3000 | * SCHED_POWER_SCALE; |
9c3f75cb | 3001 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
3002 | |
3003 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
3004 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
3005 | *imbalance = sds->busiest_load_per_task; |
3006 | return; | |
3007 | } | |
3008 | ||
3009 | /* | |
3010 | * OK, we don't have enough imbalance to justify moving tasks, | |
3011 | * however we may be able to increase total CPU power used by | |
3012 | * moving them. | |
3013 | */ | |
3014 | ||
9c3f75cb | 3015 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 3016 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 3017 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 3018 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 3019 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
3020 | |
3021 | /* Amount of load we'd subtract */ | |
1399fa78 | 3022 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 3023 | sds->busiest->sgp->power; |
1e3c88bd | 3024 | if (sds->max_load > tmp) |
9c3f75cb | 3025 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
3026 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
3027 | ||
3028 | /* Amount of load we'd add */ | |
9c3f75cb | 3029 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 3030 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
3031 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
3032 | sds->this->sgp->power; | |
1e3c88bd | 3033 | else |
1399fa78 | 3034 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
3035 | sds->this->sgp->power; |
3036 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 3037 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 3038 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
3039 | |
3040 | /* Move if we gain throughput */ | |
3041 | if (pwr_move > pwr_now) | |
3042 | *imbalance = sds->busiest_load_per_task; | |
3043 | } | |
3044 | ||
3045 | /** | |
3046 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
3047 | * groups of a given sched_domain during load balance. | |
3048 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
3049 | * @this_cpu: Cpu for which currently load balance is being performed. | |
3050 | * @imbalance: The variable to store the imbalance. | |
3051 | */ | |
3052 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
3053 | unsigned long *imbalance) | |
3054 | { | |
dd5feea1 SS |
3055 | unsigned long max_pull, load_above_capacity = ~0UL; |
3056 | ||
3057 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
3058 | if (sds->group_imb) { | |
3059 | sds->busiest_load_per_task = | |
3060 | min(sds->busiest_load_per_task, sds->avg_load); | |
3061 | } | |
3062 | ||
1e3c88bd PZ |
3063 | /* |
3064 | * In the presence of smp nice balancing, certain scenarios can have | |
3065 | * max load less than avg load(as we skip the groups at or below | |
3066 | * its cpu_power, while calculating max_load..) | |
3067 | */ | |
3068 | if (sds->max_load < sds->avg_load) { | |
3069 | *imbalance = 0; | |
3070 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3071 | } | |
3072 | ||
dd5feea1 SS |
3073 | if (!sds->group_imb) { |
3074 | /* | |
3075 | * Don't want to pull so many tasks that a group would go idle. | |
3076 | */ | |
3077 | load_above_capacity = (sds->busiest_nr_running - | |
3078 | sds->busiest_group_capacity); | |
3079 | ||
1399fa78 | 3080 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 3081 | |
9c3f75cb | 3082 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
3083 | } |
3084 | ||
3085 | /* | |
3086 | * We're trying to get all the cpus to the average_load, so we don't | |
3087 | * want to push ourselves above the average load, nor do we wish to | |
3088 | * reduce the max loaded cpu below the average load. At the same time, | |
3089 | * we also don't want to reduce the group load below the group capacity | |
3090 | * (so that we can implement power-savings policies etc). Thus we look | |
3091 | * for the minimum possible imbalance. | |
3092 | * Be careful of negative numbers as they'll appear as very large values | |
3093 | * with unsigned longs. | |
3094 | */ | |
3095 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
3096 | |
3097 | /* How much load to actually move to equalise the imbalance */ | |
9c3f75cb PZ |
3098 | *imbalance = min(max_pull * sds->busiest->sgp->power, |
3099 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) | |
1399fa78 | 3100 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
3101 | |
3102 | /* | |
3103 | * if *imbalance is less than the average load per runnable task | |
25985edc | 3104 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
3105 | * a think about bumping its value to force at least one task to be |
3106 | * moved | |
3107 | */ | |
3108 | if (*imbalance < sds->busiest_load_per_task) | |
3109 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3110 | ||
3111 | } | |
fab47622 | 3112 | |
1e3c88bd PZ |
3113 | /******* find_busiest_group() helpers end here *********************/ |
3114 | ||
3115 | /** | |
3116 | * find_busiest_group - Returns the busiest group within the sched_domain | |
3117 | * if there is an imbalance. If there isn't an imbalance, and | |
3118 | * the user has opted for power-savings, it returns a group whose | |
3119 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
3120 | * such a group exists. | |
3121 | * | |
3122 | * Also calculates the amount of weighted load which should be moved | |
3123 | * to restore balance. | |
3124 | * | |
3125 | * @sd: The sched_domain whose busiest group is to be returned. | |
3126 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
3127 | * @imbalance: Variable which stores amount of weighted load which should | |
3128 | * be moved to restore balance/put a group to idle. | |
3129 | * @idle: The idle status of this_cpu. | |
1e3c88bd PZ |
3130 | * @cpus: The set of CPUs under consideration for load-balancing. |
3131 | * @balance: Pointer to a variable indicating if this_cpu | |
3132 | * is the appropriate cpu to perform load balancing at this_level. | |
3133 | * | |
3134 | * Returns: - the busiest group if imbalance exists. | |
3135 | * - If no imbalance and user has opted for power-savings balance, | |
3136 | * return the least loaded group whose CPUs can be | |
3137 | * put to idle by rebalancing its tasks onto our group. | |
3138 | */ | |
3139 | static struct sched_group * | |
3140 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
3141 | unsigned long *imbalance, enum cpu_idle_type idle, | |
46e49b38 | 3142 | const struct cpumask *cpus, int *balance) |
1e3c88bd PZ |
3143 | { |
3144 | struct sd_lb_stats sds; | |
3145 | ||
3146 | memset(&sds, 0, sizeof(sds)); | |
3147 | ||
3148 | /* | |
3149 | * Compute the various statistics relavent for load balancing at | |
3150 | * this level. | |
3151 | */ | |
46e49b38 | 3152 | update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds); |
1e3c88bd | 3153 | |
cc57aa8f PZ |
3154 | /* |
3155 | * this_cpu is not the appropriate cpu to perform load balancing at | |
3156 | * this level. | |
1e3c88bd | 3157 | */ |
8f190fb3 | 3158 | if (!(*balance)) |
1e3c88bd PZ |
3159 | goto ret; |
3160 | ||
532cb4c4 MN |
3161 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
3162 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
3163 | return sds.busiest; | |
3164 | ||
cc57aa8f | 3165 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
3166 | if (!sds.busiest || sds.busiest_nr_running == 0) |
3167 | goto out_balanced; | |
3168 | ||
1399fa78 | 3169 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 3170 | |
866ab43e PZ |
3171 | /* |
3172 | * If the busiest group is imbalanced the below checks don't | |
3173 | * work because they assumes all things are equal, which typically | |
3174 | * isn't true due to cpus_allowed constraints and the like. | |
3175 | */ | |
3176 | if (sds.group_imb) | |
3177 | goto force_balance; | |
3178 | ||
cc57aa8f | 3179 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
fab47622 NR |
3180 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
3181 | !sds.busiest_has_capacity) | |
3182 | goto force_balance; | |
3183 | ||
cc57aa8f PZ |
3184 | /* |
3185 | * If the local group is more busy than the selected busiest group | |
3186 | * don't try and pull any tasks. | |
3187 | */ | |
1e3c88bd PZ |
3188 | if (sds.this_load >= sds.max_load) |
3189 | goto out_balanced; | |
3190 | ||
cc57aa8f PZ |
3191 | /* |
3192 | * Don't pull any tasks if this group is already above the domain | |
3193 | * average load. | |
3194 | */ | |
1e3c88bd PZ |
3195 | if (sds.this_load >= sds.avg_load) |
3196 | goto out_balanced; | |
3197 | ||
c186fafe | 3198 | if (idle == CPU_IDLE) { |
aae6d3dd SS |
3199 | /* |
3200 | * This cpu is idle. If the busiest group load doesn't | |
3201 | * have more tasks than the number of available cpu's and | |
3202 | * there is no imbalance between this and busiest group | |
3203 | * wrt to idle cpu's, it is balanced. | |
3204 | */ | |
c186fafe | 3205 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
3206 | sds.busiest_nr_running <= sds.busiest_group_weight) |
3207 | goto out_balanced; | |
c186fafe PZ |
3208 | } else { |
3209 | /* | |
3210 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
3211 | * imbalance_pct to be conservative. | |
3212 | */ | |
3213 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
3214 | goto out_balanced; | |
aae6d3dd | 3215 | } |
1e3c88bd | 3216 | |
fab47622 | 3217 | force_balance: |
1e3c88bd PZ |
3218 | /* Looks like there is an imbalance. Compute it */ |
3219 | calculate_imbalance(&sds, this_cpu, imbalance); | |
3220 | return sds.busiest; | |
3221 | ||
3222 | out_balanced: | |
3223 | /* | |
3224 | * There is no obvious imbalance. But check if we can do some balancing | |
3225 | * to save power. | |
3226 | */ | |
3227 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
3228 | return sds.busiest; | |
3229 | ret: | |
3230 | *imbalance = 0; | |
3231 | return NULL; | |
3232 | } | |
3233 | ||
3234 | /* | |
3235 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3236 | */ | |
3237 | static struct rq * | |
9d5efe05 SV |
3238 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
3239 | enum cpu_idle_type idle, unsigned long imbalance, | |
3240 | const struct cpumask *cpus) | |
1e3c88bd PZ |
3241 | { |
3242 | struct rq *busiest = NULL, *rq; | |
3243 | unsigned long max_load = 0; | |
3244 | int i; | |
3245 | ||
3246 | for_each_cpu(i, sched_group_cpus(group)) { | |
3247 | unsigned long power = power_of(i); | |
1399fa78 NR |
3248 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
3249 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
3250 | unsigned long wl; |
3251 | ||
9d5efe05 SV |
3252 | if (!capacity) |
3253 | capacity = fix_small_capacity(sd, group); | |
3254 | ||
1e3c88bd PZ |
3255 | if (!cpumask_test_cpu(i, cpus)) |
3256 | continue; | |
3257 | ||
3258 | rq = cpu_rq(i); | |
6e40f5bb | 3259 | wl = weighted_cpuload(i); |
1e3c88bd | 3260 | |
6e40f5bb TG |
3261 | /* |
3262 | * When comparing with imbalance, use weighted_cpuload() | |
3263 | * which is not scaled with the cpu power. | |
3264 | */ | |
1e3c88bd PZ |
3265 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
3266 | continue; | |
3267 | ||
6e40f5bb TG |
3268 | /* |
3269 | * For the load comparisons with the other cpu's, consider | |
3270 | * the weighted_cpuload() scaled with the cpu power, so that | |
3271 | * the load can be moved away from the cpu that is potentially | |
3272 | * running at a lower capacity. | |
3273 | */ | |
1399fa78 | 3274 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 3275 | |
1e3c88bd PZ |
3276 | if (wl > max_load) { |
3277 | max_load = wl; | |
3278 | busiest = rq; | |
3279 | } | |
3280 | } | |
3281 | ||
3282 | return busiest; | |
3283 | } | |
3284 | ||
3285 | /* | |
3286 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3287 | * so long as it is large enough. | |
3288 | */ | |
3289 | #define MAX_PINNED_INTERVAL 512 | |
3290 | ||
3291 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
3292 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
3293 | ||
46e49b38 | 3294 | static int need_active_balance(struct sched_domain *sd, int idle, |
532cb4c4 | 3295 | int busiest_cpu, int this_cpu) |
1af3ed3d PZ |
3296 | { |
3297 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
3298 | |
3299 | /* | |
3300 | * ASYM_PACKING needs to force migrate tasks from busy but | |
3301 | * higher numbered CPUs in order to pack all tasks in the | |
3302 | * lowest numbered CPUs. | |
3303 | */ | |
3304 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
3305 | return 1; | |
3306 | ||
1af3ed3d PZ |
3307 | /* |
3308 | * The only task running in a non-idle cpu can be moved to this | |
3309 | * cpu in an attempt to completely freeup the other CPU | |
3310 | * package. | |
3311 | * | |
3312 | * The package power saving logic comes from | |
3313 | * find_busiest_group(). If there are no imbalance, then | |
3314 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3315 | * f_b_g() will select a group from which a running task may be | |
3316 | * pulled to this cpu in order to make the other package idle. | |
3317 | * If there is no opportunity to make a package idle and if | |
3318 | * there are no imbalance, then f_b_g() will return NULL and no | |
3319 | * action will be taken in load_balance_newidle(). | |
3320 | * | |
3321 | * Under normal task pull operation due to imbalance, there | |
3322 | * will be more than one task in the source run queue and | |
3323 | * move_tasks() will succeed. ld_moved will be true and this | |
3324 | * active balance code will not be triggered. | |
3325 | */ | |
1af3ed3d PZ |
3326 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) |
3327 | return 0; | |
3328 | } | |
3329 | ||
3330 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3331 | } | |
3332 | ||
969c7921 TH |
3333 | static int active_load_balance_cpu_stop(void *data); |
3334 | ||
1e3c88bd PZ |
3335 | /* |
3336 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3337 | * tasks if there is an imbalance. | |
3338 | */ | |
3339 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3340 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3341 | int *balance) | |
3342 | { | |
46e49b38 | 3343 | int ld_moved, all_pinned = 0, active_balance = 0; |
1e3c88bd PZ |
3344 | struct sched_group *group; |
3345 | unsigned long imbalance; | |
3346 | struct rq *busiest; | |
3347 | unsigned long flags; | |
3348 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3349 | ||
3350 | cpumask_copy(cpus, cpu_active_mask); | |
3351 | ||
1e3c88bd PZ |
3352 | schedstat_inc(sd, lb_count[idle]); |
3353 | ||
3354 | redo: | |
46e49b38 | 3355 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, |
1e3c88bd PZ |
3356 | cpus, balance); |
3357 | ||
3358 | if (*balance == 0) | |
3359 | goto out_balanced; | |
3360 | ||
3361 | if (!group) { | |
3362 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3363 | goto out_balanced; | |
3364 | } | |
3365 | ||
9d5efe05 | 3366 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3367 | if (!busiest) { |
3368 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3369 | goto out_balanced; | |
3370 | } | |
3371 | ||
3372 | BUG_ON(busiest == this_rq); | |
3373 | ||
3374 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3375 | ||
3376 | ld_moved = 0; | |
3377 | if (busiest->nr_running > 1) { | |
3378 | /* | |
3379 | * Attempt to move tasks. If find_busiest_group has found | |
3380 | * an imbalance but busiest->nr_running <= 1, the group is | |
3381 | * still unbalanced. ld_moved simply stays zero, so it is | |
3382 | * correctly treated as an imbalance. | |
3383 | */ | |
b30aef17 | 3384 | all_pinned = 1; |
1e3c88bd PZ |
3385 | local_irq_save(flags); |
3386 | double_rq_lock(this_rq, busiest); | |
3387 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3388 | imbalance, sd, idle, &all_pinned); | |
3389 | double_rq_unlock(this_rq, busiest); | |
3390 | local_irq_restore(flags); | |
3391 | ||
3392 | /* | |
3393 | * some other cpu did the load balance for us. | |
3394 | */ | |
3395 | if (ld_moved && this_cpu != smp_processor_id()) | |
3396 | resched_cpu(this_cpu); | |
3397 | ||
3398 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3399 | if (unlikely(all_pinned)) { | |
3400 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3401 | if (!cpumask_empty(cpus)) | |
3402 | goto redo; | |
3403 | goto out_balanced; | |
3404 | } | |
3405 | } | |
3406 | ||
3407 | if (!ld_moved) { | |
3408 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3409 | /* |
3410 | * Increment the failure counter only on periodic balance. | |
3411 | * We do not want newidle balance, which can be very | |
3412 | * frequent, pollute the failure counter causing | |
3413 | * excessive cache_hot migrations and active balances. | |
3414 | */ | |
3415 | if (idle != CPU_NEWLY_IDLE) | |
3416 | sd->nr_balance_failed++; | |
1e3c88bd | 3417 | |
46e49b38 | 3418 | if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) { |
1e3c88bd PZ |
3419 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3420 | ||
969c7921 TH |
3421 | /* don't kick the active_load_balance_cpu_stop, |
3422 | * if the curr task on busiest cpu can't be | |
3423 | * moved to this_cpu | |
1e3c88bd PZ |
3424 | */ |
3425 | if (!cpumask_test_cpu(this_cpu, | |
3426 | &busiest->curr->cpus_allowed)) { | |
3427 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3428 | flags); | |
3429 | all_pinned = 1; | |
3430 | goto out_one_pinned; | |
3431 | } | |
3432 | ||
969c7921 TH |
3433 | /* |
3434 | * ->active_balance synchronizes accesses to | |
3435 | * ->active_balance_work. Once set, it's cleared | |
3436 | * only after active load balance is finished. | |
3437 | */ | |
1e3c88bd PZ |
3438 | if (!busiest->active_balance) { |
3439 | busiest->active_balance = 1; | |
3440 | busiest->push_cpu = this_cpu; | |
3441 | active_balance = 1; | |
3442 | } | |
3443 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 3444 | |
1e3c88bd | 3445 | if (active_balance) |
969c7921 TH |
3446 | stop_one_cpu_nowait(cpu_of(busiest), |
3447 | active_load_balance_cpu_stop, busiest, | |
3448 | &busiest->active_balance_work); | |
1e3c88bd PZ |
3449 | |
3450 | /* | |
3451 | * We've kicked active balancing, reset the failure | |
3452 | * counter. | |
3453 | */ | |
3454 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3455 | } | |
3456 | } else | |
3457 | sd->nr_balance_failed = 0; | |
3458 | ||
3459 | if (likely(!active_balance)) { | |
3460 | /* We were unbalanced, so reset the balancing interval */ | |
3461 | sd->balance_interval = sd->min_interval; | |
3462 | } else { | |
3463 | /* | |
3464 | * If we've begun active balancing, start to back off. This | |
3465 | * case may not be covered by the all_pinned logic if there | |
3466 | * is only 1 task on the busy runqueue (because we don't call | |
3467 | * move_tasks). | |
3468 | */ | |
3469 | if (sd->balance_interval < sd->max_interval) | |
3470 | sd->balance_interval *= 2; | |
3471 | } | |
3472 | ||
1e3c88bd PZ |
3473 | goto out; |
3474 | ||
3475 | out_balanced: | |
3476 | schedstat_inc(sd, lb_balanced[idle]); | |
3477 | ||
3478 | sd->nr_balance_failed = 0; | |
3479 | ||
3480 | out_one_pinned: | |
3481 | /* tune up the balancing interval */ | |
3482 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3483 | (sd->balance_interval < sd->max_interval)) | |
3484 | sd->balance_interval *= 2; | |
3485 | ||
46e49b38 | 3486 | ld_moved = 0; |
1e3c88bd | 3487 | out: |
1e3c88bd PZ |
3488 | return ld_moved; |
3489 | } | |
3490 | ||
1e3c88bd PZ |
3491 | /* |
3492 | * idle_balance is called by schedule() if this_cpu is about to become | |
3493 | * idle. Attempts to pull tasks from other CPUs. | |
3494 | */ | |
3495 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3496 | { | |
3497 | struct sched_domain *sd; | |
3498 | int pulled_task = 0; | |
3499 | unsigned long next_balance = jiffies + HZ; | |
3500 | ||
3501 | this_rq->idle_stamp = this_rq->clock; | |
3502 | ||
3503 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3504 | return; | |
3505 | ||
f492e12e PZ |
3506 | /* |
3507 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
3508 | */ | |
3509 | raw_spin_unlock(&this_rq->lock); | |
3510 | ||
c66eaf61 | 3511 | update_shares(this_cpu); |
dce840a0 | 3512 | rcu_read_lock(); |
1e3c88bd PZ |
3513 | for_each_domain(this_cpu, sd) { |
3514 | unsigned long interval; | |
f492e12e | 3515 | int balance = 1; |
1e3c88bd PZ |
3516 | |
3517 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3518 | continue; | |
3519 | ||
f492e12e | 3520 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 3521 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
3522 | pulled_task = load_balance(this_cpu, this_rq, |
3523 | sd, CPU_NEWLY_IDLE, &balance); | |
3524 | } | |
1e3c88bd PZ |
3525 | |
3526 | interval = msecs_to_jiffies(sd->balance_interval); | |
3527 | if (time_after(next_balance, sd->last_balance + interval)) | |
3528 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
3529 | if (pulled_task) { |
3530 | this_rq->idle_stamp = 0; | |
1e3c88bd | 3531 | break; |
d5ad140b | 3532 | } |
1e3c88bd | 3533 | } |
dce840a0 | 3534 | rcu_read_unlock(); |
f492e12e PZ |
3535 | |
3536 | raw_spin_lock(&this_rq->lock); | |
3537 | ||
1e3c88bd PZ |
3538 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
3539 | /* | |
3540 | * We are going idle. next_balance may be set based on | |
3541 | * a busy processor. So reset next_balance. | |
3542 | */ | |
3543 | this_rq->next_balance = next_balance; | |
3544 | } | |
3545 | } | |
3546 | ||
3547 | /* | |
969c7921 TH |
3548 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
3549 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
3550 | * least 1 task to be running on each physical CPU where possible, and | |
3551 | * avoids physical / logical imbalances. | |
1e3c88bd | 3552 | */ |
969c7921 | 3553 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 3554 | { |
969c7921 TH |
3555 | struct rq *busiest_rq = data; |
3556 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 3557 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 3558 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 3559 | struct sched_domain *sd; |
969c7921 TH |
3560 | |
3561 | raw_spin_lock_irq(&busiest_rq->lock); | |
3562 | ||
3563 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
3564 | if (unlikely(busiest_cpu != smp_processor_id() || | |
3565 | !busiest_rq->active_balance)) | |
3566 | goto out_unlock; | |
1e3c88bd PZ |
3567 | |
3568 | /* Is there any task to move? */ | |
3569 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 3570 | goto out_unlock; |
1e3c88bd PZ |
3571 | |
3572 | /* | |
3573 | * This condition is "impossible", if it occurs | |
3574 | * we need to fix it. Originally reported by | |
3575 | * Bjorn Helgaas on a 128-cpu setup. | |
3576 | */ | |
3577 | BUG_ON(busiest_rq == target_rq); | |
3578 | ||
3579 | /* move a task from busiest_rq to target_rq */ | |
3580 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
3581 | |
3582 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 3583 | rcu_read_lock(); |
1e3c88bd PZ |
3584 | for_each_domain(target_cpu, sd) { |
3585 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3586 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3587 | break; | |
3588 | } | |
3589 | ||
3590 | if (likely(sd)) { | |
3591 | schedstat_inc(sd, alb_count); | |
3592 | ||
3593 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3594 | sd, CPU_IDLE)) | |
3595 | schedstat_inc(sd, alb_pushed); | |
3596 | else | |
3597 | schedstat_inc(sd, alb_failed); | |
3598 | } | |
dce840a0 | 3599 | rcu_read_unlock(); |
1e3c88bd | 3600 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
3601 | out_unlock: |
3602 | busiest_rq->active_balance = 0; | |
3603 | raw_spin_unlock_irq(&busiest_rq->lock); | |
3604 | return 0; | |
1e3c88bd PZ |
3605 | } |
3606 | ||
3607 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
3608 | |
3609 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
3610 | ||
3611 | static void trigger_sched_softirq(void *data) | |
3612 | { | |
3613 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
3614 | } | |
3615 | ||
3616 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
3617 | { | |
3618 | csd->func = trigger_sched_softirq; | |
3619 | csd->info = NULL; | |
3620 | csd->flags = 0; | |
3621 | csd->priv = 0; | |
3622 | } | |
3623 | ||
3624 | /* | |
3625 | * idle load balancing details | |
3626 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
3627 | * entering idle. | |
3628 | * - This idle load balancer CPU will also go into tickless mode when | |
3629 | * it is idle, just like all other idle CPUs | |
3630 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
3631 | * needed, they will kick the idle load balancer, which then does idle | |
3632 | * load balancing for all the idle CPUs. | |
3633 | */ | |
1e3c88bd PZ |
3634 | static struct { |
3635 | atomic_t load_balancer; | |
83cd4fe2 VP |
3636 | atomic_t first_pick_cpu; |
3637 | atomic_t second_pick_cpu; | |
3638 | cpumask_var_t idle_cpus_mask; | |
3639 | cpumask_var_t grp_idle_mask; | |
3640 | unsigned long next_balance; /* in jiffy units */ | |
3641 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
3642 | |
3643 | int get_nohz_load_balancer(void) | |
3644 | { | |
3645 | return atomic_read(&nohz.load_balancer); | |
3646 | } | |
3647 | ||
3648 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3649 | /** | |
3650 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3651 | * @cpu: The cpu whose lowest level of sched domain is to | |
3652 | * be returned. | |
3653 | * @flag: The flag to check for the lowest sched_domain | |
3654 | * for the given cpu. | |
3655 | * | |
3656 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3657 | */ | |
3658 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3659 | { | |
3660 | struct sched_domain *sd; | |
3661 | ||
3662 | for_each_domain(cpu, sd) | |
08354716 | 3663 | if (sd->flags & flag) |
1e3c88bd PZ |
3664 | break; |
3665 | ||
3666 | return sd; | |
3667 | } | |
3668 | ||
3669 | /** | |
3670 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3671 | * @cpu: The cpu whose domains we're iterating over. | |
3672 | * @sd: variable holding the value of the power_savings_sd | |
3673 | * for cpu. | |
3674 | * @flag: The flag to filter the sched_domains to be iterated. | |
3675 | * | |
3676 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3677 | * set, starting from the lowest sched_domain to the highest. | |
3678 | */ | |
3679 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3680 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3681 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3682 | ||
3683 | /** | |
3684 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3685 | * @ilb_group: group to be checked for semi-idleness | |
3686 | * | |
3687 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3688 | * | |
3689 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3690 | * and atleast one non-idle CPU. This helper function checks if the given | |
3691 | * sched_group is semi-idle or not. | |
3692 | */ | |
3693 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3694 | { | |
83cd4fe2 | 3695 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
3696 | sched_group_cpus(ilb_group)); |
3697 | ||
3698 | /* | |
3699 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3700 | * and atleast one idle cpu. | |
3701 | */ | |
83cd4fe2 | 3702 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
3703 | return 0; |
3704 | ||
83cd4fe2 | 3705 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
3706 | return 0; |
3707 | ||
3708 | return 1; | |
3709 | } | |
3710 | /** | |
3711 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3712 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3713 | * | |
3714 | * Returns: Returns the id of the idle load balancer if it exists, | |
3715 | * Else, returns >= nr_cpu_ids. | |
3716 | * | |
3717 | * This algorithm picks the idle load balancer such that it belongs to a | |
3718 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3719 | * completely idle packages/cores just for the purpose of idle load balancing | |
3720 | * when there are other idle cpu's which are better suited for that job. | |
3721 | */ | |
3722 | static int find_new_ilb(int cpu) | |
3723 | { | |
3724 | struct sched_domain *sd; | |
3725 | struct sched_group *ilb_group; | |
dce840a0 | 3726 | int ilb = nr_cpu_ids; |
1e3c88bd PZ |
3727 | |
3728 | /* | |
3729 | * Have idle load balancer selection from semi-idle packages only | |
3730 | * when power-aware load balancing is enabled | |
3731 | */ | |
3732 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3733 | goto out_done; | |
3734 | ||
3735 | /* | |
3736 | * Optimize for the case when we have no idle CPUs or only one | |
3737 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3738 | */ | |
83cd4fe2 | 3739 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
3740 | goto out_done; |
3741 | ||
dce840a0 | 3742 | rcu_read_lock(); |
1e3c88bd PZ |
3743 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { |
3744 | ilb_group = sd->groups; | |
3745 | ||
3746 | do { | |
dce840a0 PZ |
3747 | if (is_semi_idle_group(ilb_group)) { |
3748 | ilb = cpumask_first(nohz.grp_idle_mask); | |
3749 | goto unlock; | |
3750 | } | |
1e3c88bd PZ |
3751 | |
3752 | ilb_group = ilb_group->next; | |
3753 | ||
3754 | } while (ilb_group != sd->groups); | |
3755 | } | |
dce840a0 PZ |
3756 | unlock: |
3757 | rcu_read_unlock(); | |
1e3c88bd PZ |
3758 | |
3759 | out_done: | |
dce840a0 | 3760 | return ilb; |
1e3c88bd PZ |
3761 | } |
3762 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3763 | static inline int find_new_ilb(int call_cpu) | |
3764 | { | |
83cd4fe2 | 3765 | return nr_cpu_ids; |
1e3c88bd PZ |
3766 | } |
3767 | #endif | |
3768 | ||
83cd4fe2 VP |
3769 | /* |
3770 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
3771 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
3772 | * CPU (if there is one). | |
3773 | */ | |
3774 | static void nohz_balancer_kick(int cpu) | |
3775 | { | |
3776 | int ilb_cpu; | |
3777 | ||
3778 | nohz.next_balance++; | |
3779 | ||
3780 | ilb_cpu = get_nohz_load_balancer(); | |
3781 | ||
3782 | if (ilb_cpu >= nr_cpu_ids) { | |
3783 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
3784 | if (ilb_cpu >= nr_cpu_ids) | |
3785 | return; | |
3786 | } | |
3787 | ||
3788 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
3789 | struct call_single_data *cp; | |
3790 | ||
3791 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
3792 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
3793 | __smp_call_function_single(ilb_cpu, cp, 0); | |
3794 | } | |
3795 | return; | |
3796 | } | |
3797 | ||
1e3c88bd PZ |
3798 | /* |
3799 | * This routine will try to nominate the ilb (idle load balancing) | |
3800 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 3801 | * load balancing on behalf of all those cpus. |
1e3c88bd | 3802 | * |
83cd4fe2 VP |
3803 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
3804 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
3805 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 3806 | * |
83cd4fe2 VP |
3807 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
3808 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
3809 | * behalf of all idle CPUs). | |
1e3c88bd | 3810 | */ |
83cd4fe2 | 3811 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
3812 | { |
3813 | int cpu = smp_processor_id(); | |
3814 | ||
3815 | if (stop_tick) { | |
1e3c88bd PZ |
3816 | if (!cpu_active(cpu)) { |
3817 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 3818 | return; |
1e3c88bd PZ |
3819 | |
3820 | /* | |
3821 | * If we are going offline and still the leader, | |
3822 | * give up! | |
3823 | */ | |
83cd4fe2 VP |
3824 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3825 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3826 | BUG(); |
3827 | ||
83cd4fe2 | 3828 | return; |
1e3c88bd PZ |
3829 | } |
3830 | ||
83cd4fe2 | 3831 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 3832 | |
83cd4fe2 VP |
3833 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
3834 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
3835 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
3836 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 3837 | |
83cd4fe2 | 3838 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
3839 | int new_ilb; |
3840 | ||
83cd4fe2 VP |
3841 | /* make me the ilb owner */ |
3842 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
3843 | cpu) != nr_cpu_ids) | |
3844 | return; | |
3845 | ||
1e3c88bd PZ |
3846 | /* |
3847 | * Check to see if there is a more power-efficient | |
3848 | * ilb. | |
3849 | */ | |
3850 | new_ilb = find_new_ilb(cpu); | |
3851 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 3852 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 3853 | resched_cpu(new_ilb); |
83cd4fe2 | 3854 | return; |
1e3c88bd | 3855 | } |
83cd4fe2 | 3856 | return; |
1e3c88bd PZ |
3857 | } |
3858 | } else { | |
83cd4fe2 VP |
3859 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
3860 | return; | |
1e3c88bd | 3861 | |
83cd4fe2 | 3862 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
3863 | |
3864 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
3865 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3866 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3867 | BUG(); |
3868 | } | |
83cd4fe2 | 3869 | return; |
1e3c88bd PZ |
3870 | } |
3871 | #endif | |
3872 | ||
3873 | static DEFINE_SPINLOCK(balancing); | |
3874 | ||
49c022e6 PZ |
3875 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3876 | ||
3877 | /* | |
3878 | * Scale the max load_balance interval with the number of CPUs in the system. | |
3879 | * This trades load-balance latency on larger machines for less cross talk. | |
3880 | */ | |
3881 | static void update_max_interval(void) | |
3882 | { | |
3883 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
3884 | } | |
3885 | ||
1e3c88bd PZ |
3886 | /* |
3887 | * It checks each scheduling domain to see if it is due to be balanced, | |
3888 | * and initiates a balancing operation if so. | |
3889 | * | |
3890 | * Balancing parameters are set up in arch_init_sched_domains. | |
3891 | */ | |
3892 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3893 | { | |
3894 | int balance = 1; | |
3895 | struct rq *rq = cpu_rq(cpu); | |
3896 | unsigned long interval; | |
3897 | struct sched_domain *sd; | |
3898 | /* Earliest time when we have to do rebalance again */ | |
3899 | unsigned long next_balance = jiffies + 60*HZ; | |
3900 | int update_next_balance = 0; | |
3901 | int need_serialize; | |
3902 | ||
2069dd75 PZ |
3903 | update_shares(cpu); |
3904 | ||
dce840a0 | 3905 | rcu_read_lock(); |
1e3c88bd PZ |
3906 | for_each_domain(cpu, sd) { |
3907 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3908 | continue; | |
3909 | ||
3910 | interval = sd->balance_interval; | |
3911 | if (idle != CPU_IDLE) | |
3912 | interval *= sd->busy_factor; | |
3913 | ||
3914 | /* scale ms to jiffies */ | |
3915 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 3916 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
3917 | |
3918 | need_serialize = sd->flags & SD_SERIALIZE; | |
3919 | ||
3920 | if (need_serialize) { | |
3921 | if (!spin_trylock(&balancing)) | |
3922 | goto out; | |
3923 | } | |
3924 | ||
3925 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3926 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3927 | /* | |
3928 | * We've pulled tasks over so either we're no | |
c186fafe | 3929 | * longer idle. |
1e3c88bd PZ |
3930 | */ |
3931 | idle = CPU_NOT_IDLE; | |
3932 | } | |
3933 | sd->last_balance = jiffies; | |
3934 | } | |
3935 | if (need_serialize) | |
3936 | spin_unlock(&balancing); | |
3937 | out: | |
3938 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3939 | next_balance = sd->last_balance + interval; | |
3940 | update_next_balance = 1; | |
3941 | } | |
3942 | ||
3943 | /* | |
3944 | * Stop the load balance at this level. There is another | |
3945 | * CPU in our sched group which is doing load balancing more | |
3946 | * actively. | |
3947 | */ | |
3948 | if (!balance) | |
3949 | break; | |
3950 | } | |
dce840a0 | 3951 | rcu_read_unlock(); |
1e3c88bd PZ |
3952 | |
3953 | /* | |
3954 | * next_balance will be updated only when there is a need. | |
3955 | * When the cpu is attached to null domain for ex, it will not be | |
3956 | * updated. | |
3957 | */ | |
3958 | if (likely(update_next_balance)) | |
3959 | rq->next_balance = next_balance; | |
3960 | } | |
3961 | ||
83cd4fe2 | 3962 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 3963 | /* |
83cd4fe2 | 3964 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
3965 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
3966 | */ | |
83cd4fe2 VP |
3967 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
3968 | { | |
3969 | struct rq *this_rq = cpu_rq(this_cpu); | |
3970 | struct rq *rq; | |
3971 | int balance_cpu; | |
3972 | ||
3973 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
3974 | return; | |
3975 | ||
3976 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
3977 | if (balance_cpu == this_cpu) | |
3978 | continue; | |
3979 | ||
3980 | /* | |
3981 | * If this cpu gets work to do, stop the load balancing | |
3982 | * work being done for other cpus. Next load | |
3983 | * balancing owner will pick it up. | |
3984 | */ | |
3985 | if (need_resched()) { | |
3986 | this_rq->nohz_balance_kick = 0; | |
3987 | break; | |
3988 | } | |
3989 | ||
3990 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 3991 | update_rq_clock(this_rq); |
83cd4fe2 VP |
3992 | update_cpu_load(this_rq); |
3993 | raw_spin_unlock_irq(&this_rq->lock); | |
3994 | ||
3995 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3996 | ||
3997 | rq = cpu_rq(balance_cpu); | |
3998 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3999 | this_rq->next_balance = rq->next_balance; | |
4000 | } | |
4001 | nohz.next_balance = this_rq->next_balance; | |
4002 | this_rq->nohz_balance_kick = 0; | |
4003 | } | |
4004 | ||
4005 | /* | |
4006 | * Current heuristic for kicking the idle load balancer | |
4007 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
4008 | * idle load balancer when it has more than one process active. This | |
4009 | * eliminates the need for idle load balancing altogether when we have | |
4010 | * only one running process in the system (common case). | |
4011 | * - If there are more than one busy CPU, idle load balancer may have | |
4012 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
4013 | * SMT or core siblings and can run better if they move to different | |
4014 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
4015 | * which will kick idle load balancer as soon as it has any load. | |
4016 | */ | |
4017 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
4018 | { | |
4019 | unsigned long now = jiffies; | |
4020 | int ret; | |
4021 | int first_pick_cpu, second_pick_cpu; | |
4022 | ||
4023 | if (time_before(now, nohz.next_balance)) | |
4024 | return 0; | |
4025 | ||
f6c3f168 | 4026 | if (rq->idle_at_tick) |
83cd4fe2 VP |
4027 | return 0; |
4028 | ||
4029 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
4030 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
4031 | ||
4032 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
4033 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
4034 | return 0; | |
4035 | ||
4036 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
4037 | if (ret == nr_cpu_ids || ret == cpu) { | |
4038 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
4039 | if (rq->nr_running > 1) | |
4040 | return 1; | |
4041 | } else { | |
4042 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
4043 | if (ret == nr_cpu_ids || ret == cpu) { | |
4044 | if (rq->nr_running) | |
4045 | return 1; | |
4046 | } | |
4047 | } | |
4048 | return 0; | |
4049 | } | |
4050 | #else | |
4051 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
4052 | #endif | |
4053 | ||
4054 | /* | |
4055 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4056 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
4057 | */ | |
1e3c88bd PZ |
4058 | static void run_rebalance_domains(struct softirq_action *h) |
4059 | { | |
4060 | int this_cpu = smp_processor_id(); | |
4061 | struct rq *this_rq = cpu_rq(this_cpu); | |
4062 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4063 | CPU_IDLE : CPU_NOT_IDLE; | |
4064 | ||
4065 | rebalance_domains(this_cpu, idle); | |
4066 | ||
1e3c88bd | 4067 | /* |
83cd4fe2 | 4068 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
4069 | * balancing on behalf of the other idle cpus whose ticks are |
4070 | * stopped. | |
4071 | */ | |
83cd4fe2 | 4072 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
4073 | } |
4074 | ||
4075 | static inline int on_null_domain(int cpu) | |
4076 | { | |
90a6501f | 4077 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
4078 | } |
4079 | ||
4080 | /* | |
4081 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
4082 | */ |
4083 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
4084 | { | |
1e3c88bd PZ |
4085 | /* Don't need to rebalance while attached to NULL domain */ |
4086 | if (time_after_eq(jiffies, rq->next_balance) && | |
4087 | likely(!on_null_domain(cpu))) | |
4088 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
4089 | #ifdef CONFIG_NO_HZ |
4090 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
4091 | nohz_balancer_kick(cpu); | |
4092 | #endif | |
1e3c88bd PZ |
4093 | } |
4094 | ||
0bcdcf28 CE |
4095 | static void rq_online_fair(struct rq *rq) |
4096 | { | |
4097 | update_sysctl(); | |
4098 | } | |
4099 | ||
4100 | static void rq_offline_fair(struct rq *rq) | |
4101 | { | |
4102 | update_sysctl(); | |
4103 | } | |
4104 | ||
1e3c88bd PZ |
4105 | #else /* CONFIG_SMP */ |
4106 | ||
4107 | /* | |
4108 | * on UP we do not need to balance between CPUs: | |
4109 | */ | |
4110 | static inline void idle_balance(int cpu, struct rq *rq) | |
4111 | { | |
4112 | } | |
4113 | ||
55e12e5e | 4114 | #endif /* CONFIG_SMP */ |
e1d1484f | 4115 | |
bf0f6f24 IM |
4116 | /* |
4117 | * scheduler tick hitting a task of our scheduling class: | |
4118 | */ | |
8f4d37ec | 4119 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
4120 | { |
4121 | struct cfs_rq *cfs_rq; | |
4122 | struct sched_entity *se = &curr->se; | |
4123 | ||
4124 | for_each_sched_entity(se) { | |
4125 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 4126 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
4127 | } |
4128 | } | |
4129 | ||
4130 | /* | |
cd29fe6f PZ |
4131 | * called on fork with the child task as argument from the parent's context |
4132 | * - child not yet on the tasklist | |
4133 | * - preemption disabled | |
bf0f6f24 | 4134 | */ |
cd29fe6f | 4135 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 4136 | { |
cd29fe6f | 4137 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 4138 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 4139 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
4140 | struct rq *rq = this_rq(); |
4141 | unsigned long flags; | |
4142 | ||
05fa785c | 4143 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 4144 | |
861d034e PZ |
4145 | update_rq_clock(rq); |
4146 | ||
b0a0f667 PM |
4147 | if (unlikely(task_cpu(p) != this_cpu)) { |
4148 | rcu_read_lock(); | |
cd29fe6f | 4149 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
4150 | rcu_read_unlock(); |
4151 | } | |
bf0f6f24 | 4152 | |
7109c442 | 4153 | update_curr(cfs_rq); |
cd29fe6f | 4154 | |
b5d9d734 MG |
4155 | if (curr) |
4156 | se->vruntime = curr->vruntime; | |
aeb73b04 | 4157 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 4158 | |
cd29fe6f | 4159 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 4160 | /* |
edcb60a3 IM |
4161 | * Upon rescheduling, sched_class::put_prev_task() will place |
4162 | * 'current' within the tree based on its new key value. | |
4163 | */ | |
4d78e7b6 | 4164 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 4165 | resched_task(rq->curr); |
4d78e7b6 | 4166 | } |
bf0f6f24 | 4167 | |
88ec22d3 PZ |
4168 | se->vruntime -= cfs_rq->min_vruntime; |
4169 | ||
05fa785c | 4170 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
4171 | } |
4172 | ||
cb469845 SR |
4173 | /* |
4174 | * Priority of the task has changed. Check to see if we preempt | |
4175 | * the current task. | |
4176 | */ | |
da7a735e PZ |
4177 | static void |
4178 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 4179 | { |
da7a735e PZ |
4180 | if (!p->se.on_rq) |
4181 | return; | |
4182 | ||
cb469845 SR |
4183 | /* |
4184 | * Reschedule if we are currently running on this runqueue and | |
4185 | * our priority decreased, or if we are not currently running on | |
4186 | * this runqueue and our priority is higher than the current's | |
4187 | */ | |
da7a735e | 4188 | if (rq->curr == p) { |
cb469845 SR |
4189 | if (p->prio > oldprio) |
4190 | resched_task(rq->curr); | |
4191 | } else | |
15afe09b | 4192 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4193 | } |
4194 | ||
da7a735e PZ |
4195 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
4196 | { | |
4197 | struct sched_entity *se = &p->se; | |
4198 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4199 | ||
4200 | /* | |
4201 | * Ensure the task's vruntime is normalized, so that when its | |
4202 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
4203 | * do the right thing. | |
4204 | * | |
4205 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
4206 | * have normalized the vruntime, if it was !on_rq, then only when | |
4207 | * the task is sleeping will it still have non-normalized vruntime. | |
4208 | */ | |
4209 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
4210 | /* | |
4211 | * Fix up our vruntime so that the current sleep doesn't | |
4212 | * cause 'unlimited' sleep bonus. | |
4213 | */ | |
4214 | place_entity(cfs_rq, se, 0); | |
4215 | se->vruntime -= cfs_rq->min_vruntime; | |
4216 | } | |
4217 | } | |
4218 | ||
cb469845 SR |
4219 | /* |
4220 | * We switched to the sched_fair class. | |
4221 | */ | |
da7a735e | 4222 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 4223 | { |
da7a735e PZ |
4224 | if (!p->se.on_rq) |
4225 | return; | |
4226 | ||
cb469845 SR |
4227 | /* |
4228 | * We were most likely switched from sched_rt, so | |
4229 | * kick off the schedule if running, otherwise just see | |
4230 | * if we can still preempt the current task. | |
4231 | */ | |
da7a735e | 4232 | if (rq->curr == p) |
cb469845 SR |
4233 | resched_task(rq->curr); |
4234 | else | |
15afe09b | 4235 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4236 | } |
4237 | ||
83b699ed SV |
4238 | /* Account for a task changing its policy or group. |
4239 | * | |
4240 | * This routine is mostly called to set cfs_rq->curr field when a task | |
4241 | * migrates between groups/classes. | |
4242 | */ | |
4243 | static void set_curr_task_fair(struct rq *rq) | |
4244 | { | |
4245 | struct sched_entity *se = &rq->curr->se; | |
4246 | ||
4247 | for_each_sched_entity(se) | |
4248 | set_next_entity(cfs_rq_of(se), se); | |
4249 | } | |
4250 | ||
810b3817 | 4251 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4252 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 4253 | { |
b2b5ce02 PZ |
4254 | /* |
4255 | * If the task was not on the rq at the time of this cgroup movement | |
4256 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
4257 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
4258 | * bonus in place_entity()). | |
4259 | * | |
4260 | * If it was on the rq, we've just 'preempted' it, which does convert | |
4261 | * ->vruntime to a relative base. | |
4262 | * | |
4263 | * Make sure both cases convert their relative position when migrating | |
4264 | * to another cgroup's rq. This does somewhat interfere with the | |
4265 | * fair sleeper stuff for the first placement, but who cares. | |
4266 | */ | |
4267 | if (!on_rq) | |
4268 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
4269 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 4270 | if (!on_rq) |
b2b5ce02 | 4271 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
4272 | } |
4273 | #endif | |
4274 | ||
6d686f45 | 4275 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
4276 | { |
4277 | struct sched_entity *se = &task->se; | |
0d721cea PW |
4278 | unsigned int rr_interval = 0; |
4279 | ||
4280 | /* | |
4281 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
4282 | * idle runqueue: | |
4283 | */ | |
0d721cea PW |
4284 | if (rq->cfs.load.weight) |
4285 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
4286 | |
4287 | return rr_interval; | |
4288 | } | |
4289 | ||
bf0f6f24 IM |
4290 | /* |
4291 | * All the scheduling class methods: | |
4292 | */ | |
5522d5d5 IM |
4293 | static const struct sched_class fair_sched_class = { |
4294 | .next = &idle_sched_class, | |
bf0f6f24 IM |
4295 | .enqueue_task = enqueue_task_fair, |
4296 | .dequeue_task = dequeue_task_fair, | |
4297 | .yield_task = yield_task_fair, | |
d95f4122 | 4298 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 4299 | |
2e09bf55 | 4300 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
4301 | |
4302 | .pick_next_task = pick_next_task_fair, | |
4303 | .put_prev_task = put_prev_task_fair, | |
4304 | ||
681f3e68 | 4305 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
4306 | .select_task_rq = select_task_rq_fair, |
4307 | ||
0bcdcf28 CE |
4308 | .rq_online = rq_online_fair, |
4309 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
4310 | |
4311 | .task_waking = task_waking_fair, | |
681f3e68 | 4312 | #endif |
bf0f6f24 | 4313 | |
83b699ed | 4314 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 4315 | .task_tick = task_tick_fair, |
cd29fe6f | 4316 | .task_fork = task_fork_fair, |
cb469845 SR |
4317 | |
4318 | .prio_changed = prio_changed_fair, | |
da7a735e | 4319 | .switched_from = switched_from_fair, |
cb469845 | 4320 | .switched_to = switched_to_fair, |
810b3817 | 4321 | |
0d721cea PW |
4322 | .get_rr_interval = get_rr_interval_fair, |
4323 | ||
810b3817 | 4324 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4325 | .task_move_group = task_move_group_fair, |
810b3817 | 4326 | #endif |
bf0f6f24 IM |
4327 | }; |
4328 | ||
4329 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 4330 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 4331 | { |
bf0f6f24 IM |
4332 | struct cfs_rq *cfs_rq; |
4333 | ||
5973e5b9 | 4334 | rcu_read_lock(); |
c3b64f1e | 4335 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 4336 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 4337 | rcu_read_unlock(); |
bf0f6f24 IM |
4338 | } |
4339 | #endif |