]>
Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
9745512c | 25 | |
bf0f6f24 | 26 | /* |
21805085 | 27 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 28 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 29 | * |
21805085 | 30 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
31 | * 'timeslice length' - timeslices in CFS are of variable length |
32 | * and have no persistent notion like in traditional, time-slice | |
33 | * based scheduling concepts. | |
bf0f6f24 | 34 | * |
d274a4ce IM |
35 | * (to see the precise effective timeslice length of your workload, |
36 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 37 | */ |
21406928 MG |
38 | unsigned int sysctl_sched_latency = 6000000ULL; |
39 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 40 | |
1983a922 CE |
41 | /* |
42 | * The initial- and re-scaling of tunables is configurable | |
43 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
44 | * | |
45 | * Options are: | |
46 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
47 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
48 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
49 | */ | |
50 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
51 | = SCHED_TUNABLESCALING_LOG; | |
52 | ||
2bd8e6d4 | 53 | /* |
b2be5e96 | 54 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 55 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 56 | */ |
0bf377bb IM |
57 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
58 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
59 | |
60 | /* | |
b2be5e96 PZ |
61 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
62 | */ | |
0bf377bb | 63 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
64 | |
65 | /* | |
2bba22c5 | 66 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 67 | * parent will (try to) run first. |
21805085 | 68 | */ |
2bba22c5 | 69 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 70 | |
1799e35d IM |
71 | /* |
72 | * sys_sched_yield() compat mode | |
73 | * | |
74 | * This option switches the agressive yield implementation of the | |
75 | * old scheduler back on. | |
76 | */ | |
77 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
78 | ||
bf0f6f24 IM |
79 | /* |
80 | * SCHED_OTHER wake-up granularity. | |
172e082a | 81 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
82 | * |
83 | * This option delays the preemption effects of decoupled workloads | |
84 | * and reduces their over-scheduling. Synchronous workloads will still | |
85 | * have immediate wakeup/sleep latencies. | |
86 | */ | |
172e082a | 87 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 88 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 89 | |
da84d961 IM |
90 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
91 | ||
a4c2f00f PZ |
92 | static const struct sched_class fair_sched_class; |
93 | ||
bf0f6f24 IM |
94 | /************************************************************** |
95 | * CFS operations on generic schedulable entities: | |
96 | */ | |
97 | ||
62160e3f | 98 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 99 | |
62160e3f | 100 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
101 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
102 | { | |
62160e3f | 103 | return cfs_rq->rq; |
bf0f6f24 IM |
104 | } |
105 | ||
62160e3f IM |
106 | /* An entity is a task if it doesn't "own" a runqueue */ |
107 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 108 | |
8f48894f PZ |
109 | static inline struct task_struct *task_of(struct sched_entity *se) |
110 | { | |
111 | #ifdef CONFIG_SCHED_DEBUG | |
112 | WARN_ON_ONCE(!entity_is_task(se)); | |
113 | #endif | |
114 | return container_of(se, struct task_struct, se); | |
115 | } | |
116 | ||
b758149c PZ |
117 | /* Walk up scheduling entities hierarchy */ |
118 | #define for_each_sched_entity(se) \ | |
119 | for (; se; se = se->parent) | |
120 | ||
121 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
122 | { | |
123 | return p->se.cfs_rq; | |
124 | } | |
125 | ||
126 | /* runqueue on which this entity is (to be) queued */ | |
127 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
128 | { | |
129 | return se->cfs_rq; | |
130 | } | |
131 | ||
132 | /* runqueue "owned" by this group */ | |
133 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
134 | { | |
135 | return grp->my_q; | |
136 | } | |
137 | ||
138 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
139 | * another cpu ('this_cpu') | |
140 | */ | |
141 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
142 | { | |
143 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
144 | } | |
145 | ||
3d4b47b4 PZ |
146 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
147 | { | |
148 | if (!cfs_rq->on_list) { | |
149 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
150 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
151 | ||
152 | cfs_rq->on_list = 1; | |
153 | } | |
154 | } | |
155 | ||
156 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
157 | { | |
158 | if (cfs_rq->on_list) { | |
159 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
160 | cfs_rq->on_list = 0; | |
161 | } | |
162 | } | |
163 | ||
b758149c PZ |
164 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
165 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
166 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
167 | ||
168 | /* Do the two (enqueued) entities belong to the same group ? */ | |
169 | static inline int | |
170 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
171 | { | |
172 | if (se->cfs_rq == pse->cfs_rq) | |
173 | return 1; | |
174 | ||
175 | return 0; | |
176 | } | |
177 | ||
178 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
179 | { | |
180 | return se->parent; | |
181 | } | |
182 | ||
464b7527 PZ |
183 | /* return depth at which a sched entity is present in the hierarchy */ |
184 | static inline int depth_se(struct sched_entity *se) | |
185 | { | |
186 | int depth = 0; | |
187 | ||
188 | for_each_sched_entity(se) | |
189 | depth++; | |
190 | ||
191 | return depth; | |
192 | } | |
193 | ||
194 | static void | |
195 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
196 | { | |
197 | int se_depth, pse_depth; | |
198 | ||
199 | /* | |
200 | * preemption test can be made between sibling entities who are in the | |
201 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
202 | * both tasks until we find their ancestors who are siblings of common | |
203 | * parent. | |
204 | */ | |
205 | ||
206 | /* First walk up until both entities are at same depth */ | |
207 | se_depth = depth_se(*se); | |
208 | pse_depth = depth_se(*pse); | |
209 | ||
210 | while (se_depth > pse_depth) { | |
211 | se_depth--; | |
212 | *se = parent_entity(*se); | |
213 | } | |
214 | ||
215 | while (pse_depth > se_depth) { | |
216 | pse_depth--; | |
217 | *pse = parent_entity(*pse); | |
218 | } | |
219 | ||
220 | while (!is_same_group(*se, *pse)) { | |
221 | *se = parent_entity(*se); | |
222 | *pse = parent_entity(*pse); | |
223 | } | |
224 | } | |
225 | ||
8f48894f PZ |
226 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
227 | ||
228 | static inline struct task_struct *task_of(struct sched_entity *se) | |
229 | { | |
230 | return container_of(se, struct task_struct, se); | |
231 | } | |
bf0f6f24 | 232 | |
62160e3f IM |
233 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
234 | { | |
235 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
236 | } |
237 | ||
238 | #define entity_is_task(se) 1 | |
239 | ||
b758149c PZ |
240 | #define for_each_sched_entity(se) \ |
241 | for (; se; se = NULL) | |
bf0f6f24 | 242 | |
b758149c | 243 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 244 | { |
b758149c | 245 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
246 | } |
247 | ||
b758149c PZ |
248 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
249 | { | |
250 | struct task_struct *p = task_of(se); | |
251 | struct rq *rq = task_rq(p); | |
252 | ||
253 | return &rq->cfs; | |
254 | } | |
255 | ||
256 | /* runqueue "owned" by this group */ | |
257 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
258 | { | |
259 | return NULL; | |
260 | } | |
261 | ||
262 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
263 | { | |
264 | return &cpu_rq(this_cpu)->cfs; | |
265 | } | |
266 | ||
3d4b47b4 PZ |
267 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
268 | { | |
269 | } | |
270 | ||
271 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
272 | { | |
273 | } | |
274 | ||
b758149c PZ |
275 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
276 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
277 | ||
278 | static inline int | |
279 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
280 | { | |
281 | return 1; | |
282 | } | |
283 | ||
284 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
285 | { | |
286 | return NULL; | |
287 | } | |
288 | ||
464b7527 PZ |
289 | static inline void |
290 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
291 | { | |
292 | } | |
293 | ||
b758149c PZ |
294 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
295 | ||
bf0f6f24 IM |
296 | |
297 | /************************************************************** | |
298 | * Scheduling class tree data structure manipulation methods: | |
299 | */ | |
300 | ||
0702e3eb | 301 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 302 | { |
368059a9 PZ |
303 | s64 delta = (s64)(vruntime - min_vruntime); |
304 | if (delta > 0) | |
02e0431a PZ |
305 | min_vruntime = vruntime; |
306 | ||
307 | return min_vruntime; | |
308 | } | |
309 | ||
0702e3eb | 310 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
311 | { |
312 | s64 delta = (s64)(vruntime - min_vruntime); | |
313 | if (delta < 0) | |
314 | min_vruntime = vruntime; | |
315 | ||
316 | return min_vruntime; | |
317 | } | |
318 | ||
54fdc581 FC |
319 | static inline int entity_before(struct sched_entity *a, |
320 | struct sched_entity *b) | |
321 | { | |
322 | return (s64)(a->vruntime - b->vruntime) < 0; | |
323 | } | |
324 | ||
0702e3eb | 325 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 326 | { |
30cfdcfc | 327 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
328 | } |
329 | ||
1af5f730 PZ |
330 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
331 | { | |
332 | u64 vruntime = cfs_rq->min_vruntime; | |
333 | ||
334 | if (cfs_rq->curr) | |
335 | vruntime = cfs_rq->curr->vruntime; | |
336 | ||
337 | if (cfs_rq->rb_leftmost) { | |
338 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
339 | struct sched_entity, | |
340 | run_node); | |
341 | ||
e17036da | 342 | if (!cfs_rq->curr) |
1af5f730 PZ |
343 | vruntime = se->vruntime; |
344 | else | |
345 | vruntime = min_vruntime(vruntime, se->vruntime); | |
346 | } | |
347 | ||
348 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
349 | } | |
350 | ||
bf0f6f24 IM |
351 | /* |
352 | * Enqueue an entity into the rb-tree: | |
353 | */ | |
0702e3eb | 354 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
355 | { |
356 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
357 | struct rb_node *parent = NULL; | |
358 | struct sched_entity *entry; | |
9014623c | 359 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
360 | int leftmost = 1; |
361 | ||
362 | /* | |
363 | * Find the right place in the rbtree: | |
364 | */ | |
365 | while (*link) { | |
366 | parent = *link; | |
367 | entry = rb_entry(parent, struct sched_entity, run_node); | |
368 | /* | |
369 | * We dont care about collisions. Nodes with | |
370 | * the same key stay together. | |
371 | */ | |
9014623c | 372 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
373 | link = &parent->rb_left; |
374 | } else { | |
375 | link = &parent->rb_right; | |
376 | leftmost = 0; | |
377 | } | |
378 | } | |
379 | ||
380 | /* | |
381 | * Maintain a cache of leftmost tree entries (it is frequently | |
382 | * used): | |
383 | */ | |
1af5f730 | 384 | if (leftmost) |
57cb499d | 385 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
386 | |
387 | rb_link_node(&se->run_node, parent, link); | |
388 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
389 | } |
390 | ||
0702e3eb | 391 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 392 | { |
3fe69747 PZ |
393 | if (cfs_rq->rb_leftmost == &se->run_node) { |
394 | struct rb_node *next_node; | |
3fe69747 PZ |
395 | |
396 | next_node = rb_next(&se->run_node); | |
397 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 398 | } |
e9acbff6 | 399 | |
bf0f6f24 | 400 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
401 | } |
402 | ||
bf0f6f24 IM |
403 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
404 | { | |
f4b6755f PZ |
405 | struct rb_node *left = cfs_rq->rb_leftmost; |
406 | ||
407 | if (!left) | |
408 | return NULL; | |
409 | ||
410 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
411 | } |
412 | ||
f4b6755f | 413 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 414 | { |
7eee3e67 | 415 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 416 | |
70eee74b BS |
417 | if (!last) |
418 | return NULL; | |
7eee3e67 IM |
419 | |
420 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
421 | } |
422 | ||
bf0f6f24 IM |
423 | /************************************************************** |
424 | * Scheduling class statistics methods: | |
425 | */ | |
426 | ||
b2be5e96 | 427 | #ifdef CONFIG_SCHED_DEBUG |
acb4a848 | 428 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 429 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
430 | loff_t *ppos) |
431 | { | |
8d65af78 | 432 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 433 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
434 | |
435 | if (ret || !write) | |
436 | return ret; | |
437 | ||
438 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
439 | sysctl_sched_min_granularity); | |
440 | ||
acb4a848 CE |
441 | #define WRT_SYSCTL(name) \ |
442 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
443 | WRT_SYSCTL(sched_min_granularity); | |
444 | WRT_SYSCTL(sched_latency); | |
445 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
446 | #undef WRT_SYSCTL |
447 | ||
b2be5e96 PZ |
448 | return 0; |
449 | } | |
450 | #endif | |
647e7cac | 451 | |
a7be37ac | 452 | /* |
f9c0b095 | 453 | * delta /= w |
a7be37ac PZ |
454 | */ |
455 | static inline unsigned long | |
456 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
457 | { | |
f9c0b095 PZ |
458 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
459 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
460 | |
461 | return delta; | |
462 | } | |
463 | ||
647e7cac IM |
464 | /* |
465 | * The idea is to set a period in which each task runs once. | |
466 | * | |
467 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
468 | * this period because otherwise the slices get too small. | |
469 | * | |
470 | * p = (nr <= nl) ? l : l*nr/nl | |
471 | */ | |
4d78e7b6 PZ |
472 | static u64 __sched_period(unsigned long nr_running) |
473 | { | |
474 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 475 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
476 | |
477 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 478 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 479 | period *= nr_running; |
4d78e7b6 PZ |
480 | } |
481 | ||
482 | return period; | |
483 | } | |
484 | ||
647e7cac IM |
485 | /* |
486 | * We calculate the wall-time slice from the period by taking a part | |
487 | * proportional to the weight. | |
488 | * | |
f9c0b095 | 489 | * s = p*P[w/rw] |
647e7cac | 490 | */ |
6d0f0ebd | 491 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 492 | { |
0a582440 | 493 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 494 | |
0a582440 | 495 | for_each_sched_entity(se) { |
6272d68c | 496 | struct load_weight *load; |
3104bf03 | 497 | struct load_weight lw; |
6272d68c LM |
498 | |
499 | cfs_rq = cfs_rq_of(se); | |
500 | load = &cfs_rq->load; | |
f9c0b095 | 501 | |
0a582440 | 502 | if (unlikely(!se->on_rq)) { |
3104bf03 | 503 | lw = cfs_rq->load; |
0a582440 MG |
504 | |
505 | update_load_add(&lw, se->load.weight); | |
506 | load = &lw; | |
507 | } | |
508 | slice = calc_delta_mine(slice, se->load.weight, load); | |
509 | } | |
510 | return slice; | |
bf0f6f24 IM |
511 | } |
512 | ||
647e7cac | 513 | /* |
ac884dec | 514 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 515 | * |
f9c0b095 | 516 | * vs = s/w |
647e7cac | 517 | */ |
f9c0b095 | 518 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 519 | { |
f9c0b095 | 520 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
521 | } |
522 | ||
bf0f6f24 IM |
523 | /* |
524 | * Update the current task's runtime statistics. Skip current tasks that | |
525 | * are not in our scheduling class. | |
526 | */ | |
527 | static inline void | |
8ebc91d9 IM |
528 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
529 | unsigned long delta_exec) | |
bf0f6f24 | 530 | { |
bbdba7c0 | 531 | unsigned long delta_exec_weighted; |
bf0f6f24 | 532 | |
41acab88 LDM |
533 | schedstat_set(curr->statistics.exec_max, |
534 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
535 | |
536 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 537 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 538 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 539 | |
e9acbff6 | 540 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 541 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
542 | } |
543 | ||
b7cc0896 | 544 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 545 | { |
429d43bc | 546 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 547 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
548 | unsigned long delta_exec; |
549 | ||
550 | if (unlikely(!curr)) | |
551 | return; | |
552 | ||
553 | /* | |
554 | * Get the amount of time the current task was running | |
555 | * since the last time we changed load (this cannot | |
556 | * overflow on 32 bits): | |
557 | */ | |
8ebc91d9 | 558 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
559 | if (!delta_exec) |
560 | return; | |
bf0f6f24 | 561 | |
8ebc91d9 IM |
562 | __update_curr(cfs_rq, curr, delta_exec); |
563 | curr->exec_start = now; | |
d842de87 SV |
564 | |
565 | if (entity_is_task(curr)) { | |
566 | struct task_struct *curtask = task_of(curr); | |
567 | ||
f977bb49 | 568 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 569 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 570 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 571 | } |
bf0f6f24 IM |
572 | } |
573 | ||
574 | static inline void | |
5870db5b | 575 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 576 | { |
41acab88 | 577 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
578 | } |
579 | ||
bf0f6f24 IM |
580 | /* |
581 | * Task is being enqueued - update stats: | |
582 | */ | |
d2417e5a | 583 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 584 | { |
bf0f6f24 IM |
585 | /* |
586 | * Are we enqueueing a waiting task? (for current tasks | |
587 | * a dequeue/enqueue event is a NOP) | |
588 | */ | |
429d43bc | 589 | if (se != cfs_rq->curr) |
5870db5b | 590 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
591 | } |
592 | ||
bf0f6f24 | 593 | static void |
9ef0a961 | 594 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 595 | { |
41acab88 LDM |
596 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
597 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
598 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
599 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
600 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
601 | #ifdef CONFIG_SCHEDSTATS |
602 | if (entity_is_task(se)) { | |
603 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 604 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
605 | } |
606 | #endif | |
41acab88 | 607 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
608 | } |
609 | ||
610 | static inline void | |
19b6a2e3 | 611 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 612 | { |
bf0f6f24 IM |
613 | /* |
614 | * Mark the end of the wait period if dequeueing a | |
615 | * waiting task: | |
616 | */ | |
429d43bc | 617 | if (se != cfs_rq->curr) |
9ef0a961 | 618 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
619 | } |
620 | ||
621 | /* | |
622 | * We are picking a new current task - update its stats: | |
623 | */ | |
624 | static inline void | |
79303e9e | 625 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
626 | { |
627 | /* | |
628 | * We are starting a new run period: | |
629 | */ | |
305e6835 | 630 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
631 | } |
632 | ||
bf0f6f24 IM |
633 | /************************************************** |
634 | * Scheduling class queueing methods: | |
635 | */ | |
636 | ||
c09595f6 PZ |
637 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
638 | static void | |
639 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
640 | { | |
641 | cfs_rq->task_weight += weight; | |
642 | } | |
643 | #else | |
644 | static inline void | |
645 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
646 | { | |
647 | } | |
648 | #endif | |
649 | ||
30cfdcfc DA |
650 | static void |
651 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
652 | { | |
653 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
654 | if (!parent_entity(se)) |
655 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 656 | if (entity_is_task(se)) { |
c09595f6 | 657 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
658 | list_add(&se->group_node, &cfs_rq->tasks); |
659 | } | |
30cfdcfc | 660 | cfs_rq->nr_running++; |
30cfdcfc DA |
661 | } |
662 | ||
663 | static void | |
664 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
665 | { | |
666 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
667 | if (!parent_entity(se)) |
668 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 669 | if (entity_is_task(se)) { |
c09595f6 | 670 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
671 | list_del_init(&se->group_node); |
672 | } | |
30cfdcfc | 673 | cfs_rq->nr_running--; |
30cfdcfc DA |
674 | } |
675 | ||
2069dd75 | 676 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
e33078ba | 677 | static void update_cfs_load(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
678 | { |
679 | u64 period = sched_avg_period(); | |
680 | u64 now, delta; | |
e33078ba | 681 | unsigned long load = cfs_rq->load.weight; |
2069dd75 PZ |
682 | |
683 | if (!cfs_rq) | |
684 | return; | |
685 | ||
686 | now = rq_of(cfs_rq)->clock; | |
687 | delta = now - cfs_rq->load_stamp; | |
688 | ||
e33078ba PT |
689 | /* truncate load history at 4 idle periods */ |
690 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
691 | now - cfs_rq->load_last > 4 * period) { | |
692 | cfs_rq->load_period = 0; | |
693 | cfs_rq->load_avg = 0; | |
694 | } | |
695 | ||
2069dd75 PZ |
696 | cfs_rq->load_stamp = now; |
697 | cfs_rq->load_period += delta; | |
e33078ba PT |
698 | if (load) { |
699 | cfs_rq->load_last = now; | |
700 | cfs_rq->load_avg += delta * load; | |
701 | } | |
2069dd75 PZ |
702 | |
703 | while (cfs_rq->load_period > period) { | |
704 | /* | |
705 | * Inline assembly required to prevent the compiler | |
706 | * optimising this loop into a divmod call. | |
707 | * See __iter_div_u64_rem() for another example of this. | |
708 | */ | |
709 | asm("" : "+rm" (cfs_rq->load_period)); | |
710 | cfs_rq->load_period /= 2; | |
711 | cfs_rq->load_avg /= 2; | |
712 | } | |
3d4b47b4 | 713 | |
e33078ba PT |
714 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
715 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
716 | } |
717 | ||
718 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, | |
719 | unsigned long weight) | |
720 | { | |
721 | if (se->on_rq) | |
722 | account_entity_dequeue(cfs_rq, se); | |
723 | ||
724 | update_load_set(&se->load, weight); | |
725 | ||
726 | if (se->on_rq) | |
727 | account_entity_enqueue(cfs_rq, se); | |
728 | } | |
729 | ||
f0d7442a | 730 | static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta) |
2069dd75 PZ |
731 | { |
732 | struct task_group *tg; | |
733 | struct sched_entity *se; | |
734 | long load_weight, load, shares; | |
735 | ||
736 | if (!cfs_rq) | |
737 | return; | |
738 | ||
739 | tg = cfs_rq->tg; | |
740 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
741 | if (!se) | |
742 | return; | |
743 | ||
f0d7442a | 744 | load = cfs_rq->load.weight + weight_delta; |
2069dd75 PZ |
745 | |
746 | load_weight = atomic_read(&tg->load_weight); | |
747 | load_weight -= cfs_rq->load_contribution; | |
748 | load_weight += load; | |
749 | ||
750 | shares = (tg->shares * load); | |
751 | if (load_weight) | |
752 | shares /= load_weight; | |
753 | ||
754 | if (shares < MIN_SHARES) | |
755 | shares = MIN_SHARES; | |
756 | if (shares > tg->shares) | |
757 | shares = tg->shares; | |
758 | ||
759 | reweight_entity(cfs_rq_of(se), se, shares); | |
760 | } | |
761 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
e33078ba | 762 | static inline void update_cfs_load(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
763 | { |
764 | } | |
765 | ||
f0d7442a | 766 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta) |
2069dd75 PZ |
767 | { |
768 | } | |
769 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
770 | ||
2396af69 | 771 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 772 | { |
bf0f6f24 | 773 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
774 | struct task_struct *tsk = NULL; |
775 | ||
776 | if (entity_is_task(se)) | |
777 | tsk = task_of(se); | |
778 | ||
41acab88 LDM |
779 | if (se->statistics.sleep_start) { |
780 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
781 | |
782 | if ((s64)delta < 0) | |
783 | delta = 0; | |
784 | ||
41acab88 LDM |
785 | if (unlikely(delta > se->statistics.sleep_max)) |
786 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 787 | |
41acab88 LDM |
788 | se->statistics.sleep_start = 0; |
789 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 790 | |
768d0c27 | 791 | if (tsk) { |
e414314c | 792 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
793 | trace_sched_stat_sleep(tsk, delta); |
794 | } | |
bf0f6f24 | 795 | } |
41acab88 LDM |
796 | if (se->statistics.block_start) { |
797 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
798 | |
799 | if ((s64)delta < 0) | |
800 | delta = 0; | |
801 | ||
41acab88 LDM |
802 | if (unlikely(delta > se->statistics.block_max)) |
803 | se->statistics.block_max = delta; | |
bf0f6f24 | 804 | |
41acab88 LDM |
805 | se->statistics.block_start = 0; |
806 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 807 | |
e414314c | 808 | if (tsk) { |
8f0dfc34 | 809 | if (tsk->in_iowait) { |
41acab88 LDM |
810 | se->statistics.iowait_sum += delta; |
811 | se->statistics.iowait_count++; | |
768d0c27 | 812 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
813 | } |
814 | ||
e414314c PZ |
815 | /* |
816 | * Blocking time is in units of nanosecs, so shift by | |
817 | * 20 to get a milliseconds-range estimation of the | |
818 | * amount of time that the task spent sleeping: | |
819 | */ | |
820 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
821 | profile_hits(SLEEP_PROFILING, | |
822 | (void *)get_wchan(tsk), | |
823 | delta >> 20); | |
824 | } | |
825 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 826 | } |
bf0f6f24 IM |
827 | } |
828 | #endif | |
829 | } | |
830 | ||
ddc97297 PZ |
831 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
832 | { | |
833 | #ifdef CONFIG_SCHED_DEBUG | |
834 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
835 | ||
836 | if (d < 0) | |
837 | d = -d; | |
838 | ||
839 | if (d > 3*sysctl_sched_latency) | |
840 | schedstat_inc(cfs_rq, nr_spread_over); | |
841 | #endif | |
842 | } | |
843 | ||
aeb73b04 PZ |
844 | static void |
845 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
846 | { | |
1af5f730 | 847 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 848 | |
2cb8600e PZ |
849 | /* |
850 | * The 'current' period is already promised to the current tasks, | |
851 | * however the extra weight of the new task will slow them down a | |
852 | * little, place the new task so that it fits in the slot that | |
853 | * stays open at the end. | |
854 | */ | |
94dfb5e7 | 855 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 856 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 857 | |
a2e7a7eb | 858 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 859 | if (!initial) { |
a2e7a7eb | 860 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 861 | |
a2e7a7eb MG |
862 | /* |
863 | * Halve their sleep time's effect, to allow | |
864 | * for a gentler effect of sleepers: | |
865 | */ | |
866 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
867 | thresh >>= 1; | |
51e0304c | 868 | |
a2e7a7eb | 869 | vruntime -= thresh; |
aeb73b04 PZ |
870 | } |
871 | ||
b5d9d734 MG |
872 | /* ensure we never gain time by being placed backwards. */ |
873 | vruntime = max_vruntime(se->vruntime, vruntime); | |
874 | ||
67e9fb2a | 875 | se->vruntime = vruntime; |
aeb73b04 PZ |
876 | } |
877 | ||
bf0f6f24 | 878 | static void |
88ec22d3 | 879 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 880 | { |
88ec22d3 PZ |
881 | /* |
882 | * Update the normalized vruntime before updating min_vruntime | |
883 | * through callig update_curr(). | |
884 | */ | |
371fd7e7 | 885 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
886 | se->vruntime += cfs_rq->min_vruntime; |
887 | ||
bf0f6f24 | 888 | /* |
a2a2d680 | 889 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 890 | */ |
b7cc0896 | 891 | update_curr(cfs_rq); |
e33078ba | 892 | update_cfs_load(cfs_rq); |
f0d7442a | 893 | update_cfs_shares(cfs_rq, se->load.weight); |
a992241d | 894 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 895 | |
88ec22d3 | 896 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 897 | place_entity(cfs_rq, se, 0); |
2396af69 | 898 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 899 | } |
bf0f6f24 | 900 | |
d2417e5a | 901 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 902 | check_spread(cfs_rq, se); |
83b699ed SV |
903 | if (se != cfs_rq->curr) |
904 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 905 | se->on_rq = 1; |
3d4b47b4 PZ |
906 | |
907 | if (cfs_rq->nr_running == 1) | |
908 | list_add_leaf_cfs_rq(cfs_rq); | |
bf0f6f24 IM |
909 | } |
910 | ||
a571bbea | 911 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 | 912 | { |
de69a80b | 913 | if (!se || cfs_rq->last == se) |
2002c695 PZ |
914 | cfs_rq->last = NULL; |
915 | ||
de69a80b | 916 | if (!se || cfs_rq->next == se) |
2002c695 PZ |
917 | cfs_rq->next = NULL; |
918 | } | |
919 | ||
a571bbea PZ |
920 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
921 | { | |
922 | for_each_sched_entity(se) | |
923 | __clear_buddies(cfs_rq_of(se), se); | |
924 | } | |
925 | ||
bf0f6f24 | 926 | static void |
371fd7e7 | 927 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 928 | { |
a2a2d680 DA |
929 | /* |
930 | * Update run-time statistics of the 'current'. | |
931 | */ | |
932 | update_curr(cfs_rq); | |
933 | ||
19b6a2e3 | 934 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 935 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 936 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
937 | if (entity_is_task(se)) { |
938 | struct task_struct *tsk = task_of(se); | |
939 | ||
940 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 941 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 942 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 943 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 944 | } |
db36cc7d | 945 | #endif |
67e9fb2a PZ |
946 | } |
947 | ||
2002c695 | 948 | clear_buddies(cfs_rq, se); |
4793241b | 949 | |
83b699ed | 950 | if (se != cfs_rq->curr) |
30cfdcfc | 951 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 952 | se->on_rq = 0; |
e33078ba | 953 | update_cfs_load(cfs_rq); |
30cfdcfc | 954 | account_entity_dequeue(cfs_rq, se); |
1af5f730 | 955 | update_min_vruntime(cfs_rq); |
f0d7442a | 956 | update_cfs_shares(cfs_rq, 0); |
88ec22d3 PZ |
957 | |
958 | /* | |
959 | * Normalize the entity after updating the min_vruntime because the | |
960 | * update can refer to the ->curr item and we need to reflect this | |
961 | * movement in our normalized position. | |
962 | */ | |
371fd7e7 | 963 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 964 | se->vruntime -= cfs_rq->min_vruntime; |
bf0f6f24 IM |
965 | } |
966 | ||
967 | /* | |
968 | * Preempt the current task with a newly woken task if needed: | |
969 | */ | |
7c92e54f | 970 | static void |
2e09bf55 | 971 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 972 | { |
11697830 PZ |
973 | unsigned long ideal_runtime, delta_exec; |
974 | ||
6d0f0ebd | 975 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 976 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 977 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 978 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
979 | /* |
980 | * The current task ran long enough, ensure it doesn't get | |
981 | * re-elected due to buddy favours. | |
982 | */ | |
983 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
984 | return; |
985 | } | |
986 | ||
987 | /* | |
988 | * Ensure that a task that missed wakeup preemption by a | |
989 | * narrow margin doesn't have to wait for a full slice. | |
990 | * This also mitigates buddy induced latencies under load. | |
991 | */ | |
992 | if (!sched_feat(WAKEUP_PREEMPT)) | |
993 | return; | |
994 | ||
995 | if (delta_exec < sysctl_sched_min_granularity) | |
996 | return; | |
997 | ||
998 | if (cfs_rq->nr_running > 1) { | |
999 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
1000 | s64 delta = curr->vruntime - se->vruntime; | |
1001 | ||
1002 | if (delta > ideal_runtime) | |
1003 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 1004 | } |
bf0f6f24 IM |
1005 | } |
1006 | ||
83b699ed | 1007 | static void |
8494f412 | 1008 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1009 | { |
83b699ed SV |
1010 | /* 'current' is not kept within the tree. */ |
1011 | if (se->on_rq) { | |
1012 | /* | |
1013 | * Any task has to be enqueued before it get to execute on | |
1014 | * a CPU. So account for the time it spent waiting on the | |
1015 | * runqueue. | |
1016 | */ | |
1017 | update_stats_wait_end(cfs_rq, se); | |
1018 | __dequeue_entity(cfs_rq, se); | |
1019 | } | |
1020 | ||
79303e9e | 1021 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1022 | cfs_rq->curr = se; |
eba1ed4b IM |
1023 | #ifdef CONFIG_SCHEDSTATS |
1024 | /* | |
1025 | * Track our maximum slice length, if the CPU's load is at | |
1026 | * least twice that of our own weight (i.e. dont track it | |
1027 | * when there are only lesser-weight tasks around): | |
1028 | */ | |
495eca49 | 1029 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1030 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1031 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1032 | } | |
1033 | #endif | |
4a55b450 | 1034 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1035 | } |
1036 | ||
3f3a4904 PZ |
1037 | static int |
1038 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1039 | ||
f4b6755f | 1040 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1041 | { |
f4b6755f | 1042 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
f685ceac | 1043 | struct sched_entity *left = se; |
f4b6755f | 1044 | |
f685ceac MG |
1045 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) |
1046 | se = cfs_rq->next; | |
aa2ac252 | 1047 | |
f685ceac MG |
1048 | /* |
1049 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1050 | */ | |
1051 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1052 | se = cfs_rq->last; | |
1053 | ||
1054 | clear_buddies(cfs_rq, se); | |
4793241b PZ |
1055 | |
1056 | return se; | |
aa2ac252 PZ |
1057 | } |
1058 | ||
ab6cde26 | 1059 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1060 | { |
1061 | /* | |
1062 | * If still on the runqueue then deactivate_task() | |
1063 | * was not called and update_curr() has to be done: | |
1064 | */ | |
1065 | if (prev->on_rq) | |
b7cc0896 | 1066 | update_curr(cfs_rq); |
bf0f6f24 | 1067 | |
ddc97297 | 1068 | check_spread(cfs_rq, prev); |
30cfdcfc | 1069 | if (prev->on_rq) { |
5870db5b | 1070 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1071 | /* Put 'current' back into the tree. */ |
1072 | __enqueue_entity(cfs_rq, prev); | |
1073 | } | |
429d43bc | 1074 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1075 | } |
1076 | ||
8f4d37ec PZ |
1077 | static void |
1078 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1079 | { |
bf0f6f24 | 1080 | /* |
30cfdcfc | 1081 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1082 | */ |
30cfdcfc | 1083 | update_curr(cfs_rq); |
bf0f6f24 | 1084 | |
8f4d37ec PZ |
1085 | #ifdef CONFIG_SCHED_HRTICK |
1086 | /* | |
1087 | * queued ticks are scheduled to match the slice, so don't bother | |
1088 | * validating it and just reschedule. | |
1089 | */ | |
983ed7a6 HH |
1090 | if (queued) { |
1091 | resched_task(rq_of(cfs_rq)->curr); | |
1092 | return; | |
1093 | } | |
8f4d37ec PZ |
1094 | /* |
1095 | * don't let the period tick interfere with the hrtick preemption | |
1096 | */ | |
1097 | if (!sched_feat(DOUBLE_TICK) && | |
1098 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1099 | return; | |
1100 | #endif | |
1101 | ||
ce6c1311 | 1102 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 1103 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1104 | } |
1105 | ||
1106 | /************************************************** | |
1107 | * CFS operations on tasks: | |
1108 | */ | |
1109 | ||
8f4d37ec PZ |
1110 | #ifdef CONFIG_SCHED_HRTICK |
1111 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1112 | { | |
8f4d37ec PZ |
1113 | struct sched_entity *se = &p->se; |
1114 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1115 | ||
1116 | WARN_ON(task_rq(p) != rq); | |
1117 | ||
1118 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1119 | u64 slice = sched_slice(cfs_rq, se); | |
1120 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1121 | s64 delta = slice - ran; | |
1122 | ||
1123 | if (delta < 0) { | |
1124 | if (rq->curr == p) | |
1125 | resched_task(p); | |
1126 | return; | |
1127 | } | |
1128 | ||
1129 | /* | |
1130 | * Don't schedule slices shorter than 10000ns, that just | |
1131 | * doesn't make sense. Rely on vruntime for fairness. | |
1132 | */ | |
31656519 | 1133 | if (rq->curr != p) |
157124c1 | 1134 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1135 | |
31656519 | 1136 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1137 | } |
1138 | } | |
a4c2f00f PZ |
1139 | |
1140 | /* | |
1141 | * called from enqueue/dequeue and updates the hrtick when the | |
1142 | * current task is from our class and nr_running is low enough | |
1143 | * to matter. | |
1144 | */ | |
1145 | static void hrtick_update(struct rq *rq) | |
1146 | { | |
1147 | struct task_struct *curr = rq->curr; | |
1148 | ||
1149 | if (curr->sched_class != &fair_sched_class) | |
1150 | return; | |
1151 | ||
1152 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1153 | hrtick_start_fair(rq, curr); | |
1154 | } | |
55e12e5e | 1155 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1156 | static inline void |
1157 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1158 | { | |
1159 | } | |
a4c2f00f PZ |
1160 | |
1161 | static inline void hrtick_update(struct rq *rq) | |
1162 | { | |
1163 | } | |
8f4d37ec PZ |
1164 | #endif |
1165 | ||
bf0f6f24 IM |
1166 | /* |
1167 | * The enqueue_task method is called before nr_running is | |
1168 | * increased. Here we update the fair scheduling stats and | |
1169 | * then put the task into the rbtree: | |
1170 | */ | |
ea87bb78 | 1171 | static void |
371fd7e7 | 1172 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1173 | { |
1174 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1175 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1176 | |
1177 | for_each_sched_entity(se) { | |
62fb1851 | 1178 | if (se->on_rq) |
bf0f6f24 IM |
1179 | break; |
1180 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 PZ |
1181 | enqueue_entity(cfs_rq, se, flags); |
1182 | flags = ENQUEUE_WAKEUP; | |
bf0f6f24 | 1183 | } |
8f4d37ec | 1184 | |
2069dd75 PZ |
1185 | for_each_sched_entity(se) { |
1186 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1187 | ||
e33078ba | 1188 | update_cfs_load(cfs_rq); |
f0d7442a | 1189 | update_cfs_shares(cfs_rq, 0); |
2069dd75 PZ |
1190 | } |
1191 | ||
a4c2f00f | 1192 | hrtick_update(rq); |
bf0f6f24 IM |
1193 | } |
1194 | ||
1195 | /* | |
1196 | * The dequeue_task method is called before nr_running is | |
1197 | * decreased. We remove the task from the rbtree and | |
1198 | * update the fair scheduling stats: | |
1199 | */ | |
371fd7e7 | 1200 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1201 | { |
1202 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1203 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1204 | |
1205 | for_each_sched_entity(se) { | |
1206 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1207 | dequeue_entity(cfs_rq, se, flags); |
2069dd75 | 1208 | |
bf0f6f24 | 1209 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 1210 | if (cfs_rq->load.weight) |
bf0f6f24 | 1211 | break; |
371fd7e7 | 1212 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1213 | } |
8f4d37ec | 1214 | |
2069dd75 PZ |
1215 | for_each_sched_entity(se) { |
1216 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1217 | ||
e33078ba | 1218 | update_cfs_load(cfs_rq); |
f0d7442a | 1219 | update_cfs_shares(cfs_rq, 0); |
2069dd75 PZ |
1220 | } |
1221 | ||
a4c2f00f | 1222 | hrtick_update(rq); |
bf0f6f24 IM |
1223 | } |
1224 | ||
1225 | /* | |
1799e35d IM |
1226 | * sched_yield() support is very simple - we dequeue and enqueue. |
1227 | * | |
1228 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 1229 | */ |
4530d7ab | 1230 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 1231 | { |
db292ca3 IM |
1232 | struct task_struct *curr = rq->curr; |
1233 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1234 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
1235 | |
1236 | /* | |
1799e35d IM |
1237 | * Are we the only task in the tree? |
1238 | */ | |
1239 | if (unlikely(cfs_rq->nr_running == 1)) | |
1240 | return; | |
1241 | ||
2002c695 PZ |
1242 | clear_buddies(cfs_rq, se); |
1243 | ||
db292ca3 | 1244 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1245 | update_rq_clock(rq); |
1799e35d | 1246 | /* |
a2a2d680 | 1247 | * Update run-time statistics of the 'current'. |
1799e35d | 1248 | */ |
2b1e315d | 1249 | update_curr(cfs_rq); |
1799e35d IM |
1250 | |
1251 | return; | |
1252 | } | |
1253 | /* | |
1254 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1255 | */ |
2b1e315d | 1256 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1257 | /* |
1258 | * Already in the rightmost position? | |
1259 | */ | |
54fdc581 | 1260 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
1799e35d IM |
1261 | return; |
1262 | ||
1263 | /* | |
1264 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1265 | * Upon rescheduling, sched_class::put_prev_task() will place |
1266 | * 'current' within the tree based on its new key value. | |
1799e35d | 1267 | */ |
30cfdcfc | 1268 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1269 | } |
1270 | ||
e7693a36 | 1271 | #ifdef CONFIG_SMP |
098fb9db | 1272 | |
88ec22d3 PZ |
1273 | static void task_waking_fair(struct rq *rq, struct task_struct *p) |
1274 | { | |
1275 | struct sched_entity *se = &p->se; | |
1276 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1277 | ||
1278 | se->vruntime -= cfs_rq->min_vruntime; | |
1279 | } | |
1280 | ||
bb3469ac | 1281 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1282 | /* |
1283 | * effective_load() calculates the load change as seen from the root_task_group | |
1284 | * | |
1285 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1286 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1287 | * can calculate the shift in shares. | |
f5bfb7d9 | 1288 | */ |
2069dd75 | 1289 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 1290 | { |
4be9daaa | 1291 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1292 | |
1293 | if (!tg->parent) | |
1294 | return wl; | |
1295 | ||
4be9daaa | 1296 | for_each_sched_entity(se) { |
cb5ef42a | 1297 | long S, rw, s, a, b; |
4be9daaa PZ |
1298 | |
1299 | S = se->my_q->tg->shares; | |
2069dd75 PZ |
1300 | s = se->load.weight; |
1301 | rw = se->my_q->load.weight; | |
bb3469ac | 1302 | |
cb5ef42a PZ |
1303 | a = S*(rw + wl); |
1304 | b = S*rw + s*wg; | |
4be9daaa | 1305 | |
940959e9 PZ |
1306 | wl = s*(a-b); |
1307 | ||
1308 | if (likely(b)) | |
1309 | wl /= b; | |
1310 | ||
83378269 PZ |
1311 | /* |
1312 | * Assume the group is already running and will | |
1313 | * thus already be accounted for in the weight. | |
1314 | * | |
1315 | * That is, moving shares between CPUs, does not | |
1316 | * alter the group weight. | |
1317 | */ | |
4be9daaa | 1318 | wg = 0; |
4be9daaa | 1319 | } |
bb3469ac | 1320 | |
4be9daaa | 1321 | return wl; |
bb3469ac | 1322 | } |
4be9daaa | 1323 | |
bb3469ac | 1324 | #else |
4be9daaa | 1325 | |
83378269 PZ |
1326 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1327 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1328 | { |
83378269 | 1329 | return wl; |
bb3469ac | 1330 | } |
4be9daaa | 1331 | |
bb3469ac PZ |
1332 | #endif |
1333 | ||
c88d5910 | 1334 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1335 | { |
c88d5910 PZ |
1336 | unsigned long this_load, load; |
1337 | int idx, this_cpu, prev_cpu; | |
098fb9db | 1338 | unsigned long tl_per_task; |
c88d5910 | 1339 | struct task_group *tg; |
83378269 | 1340 | unsigned long weight; |
b3137bc8 | 1341 | int balanced; |
098fb9db | 1342 | |
c88d5910 PZ |
1343 | idx = sd->wake_idx; |
1344 | this_cpu = smp_processor_id(); | |
1345 | prev_cpu = task_cpu(p); | |
1346 | load = source_load(prev_cpu, idx); | |
1347 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1348 | |
b3137bc8 MG |
1349 | /* |
1350 | * If sync wakeup then subtract the (maximum possible) | |
1351 | * effect of the currently running task from the load | |
1352 | * of the current CPU: | |
1353 | */ | |
f3b577de | 1354 | rcu_read_lock(); |
83378269 PZ |
1355 | if (sync) { |
1356 | tg = task_group(current); | |
1357 | weight = current->se.load.weight; | |
1358 | ||
c88d5910 | 1359 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1360 | load += effective_load(tg, prev_cpu, 0, -weight); |
1361 | } | |
b3137bc8 | 1362 | |
83378269 PZ |
1363 | tg = task_group(p); |
1364 | weight = p->se.load.weight; | |
b3137bc8 | 1365 | |
71a29aa7 PZ |
1366 | /* |
1367 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1368 | * due to the sync cause above having dropped this_load to 0, we'll |
1369 | * always have an imbalance, but there's really nothing you can do | |
1370 | * about that, so that's good too. | |
71a29aa7 PZ |
1371 | * |
1372 | * Otherwise check if either cpus are near enough in load to allow this | |
1373 | * task to be woken on this_cpu. | |
1374 | */ | |
e51fd5e2 PZ |
1375 | if (this_load) { |
1376 | unsigned long this_eff_load, prev_eff_load; | |
1377 | ||
1378 | this_eff_load = 100; | |
1379 | this_eff_load *= power_of(prev_cpu); | |
1380 | this_eff_load *= this_load + | |
1381 | effective_load(tg, this_cpu, weight, weight); | |
1382 | ||
1383 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
1384 | prev_eff_load *= power_of(this_cpu); | |
1385 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
1386 | ||
1387 | balanced = this_eff_load <= prev_eff_load; | |
1388 | } else | |
1389 | balanced = true; | |
f3b577de | 1390 | rcu_read_unlock(); |
b3137bc8 | 1391 | |
098fb9db | 1392 | /* |
4ae7d5ce IM |
1393 | * If the currently running task will sleep within |
1394 | * a reasonable amount of time then attract this newly | |
1395 | * woken task: | |
098fb9db | 1396 | */ |
2fb7635c PZ |
1397 | if (sync && balanced) |
1398 | return 1; | |
098fb9db | 1399 | |
41acab88 | 1400 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
1401 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
1402 | ||
c88d5910 PZ |
1403 | if (balanced || |
1404 | (this_load <= load && | |
1405 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1406 | /* |
1407 | * This domain has SD_WAKE_AFFINE and | |
1408 | * p is cache cold in this domain, and | |
1409 | * there is no bad imbalance. | |
1410 | */ | |
c88d5910 | 1411 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 1412 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
1413 | |
1414 | return 1; | |
1415 | } | |
1416 | return 0; | |
1417 | } | |
1418 | ||
aaee1203 PZ |
1419 | /* |
1420 | * find_idlest_group finds and returns the least busy CPU group within the | |
1421 | * domain. | |
1422 | */ | |
1423 | static struct sched_group * | |
78e7ed53 | 1424 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1425 | int this_cpu, int load_idx) |
e7693a36 | 1426 | { |
b3bd3de6 | 1427 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 1428 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 1429 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1430 | |
aaee1203 PZ |
1431 | do { |
1432 | unsigned long load, avg_load; | |
1433 | int local_group; | |
1434 | int i; | |
e7693a36 | 1435 | |
aaee1203 PZ |
1436 | /* Skip over this group if it has no CPUs allowed */ |
1437 | if (!cpumask_intersects(sched_group_cpus(group), | |
1438 | &p->cpus_allowed)) | |
1439 | continue; | |
1440 | ||
1441 | local_group = cpumask_test_cpu(this_cpu, | |
1442 | sched_group_cpus(group)); | |
1443 | ||
1444 | /* Tally up the load of all CPUs in the group */ | |
1445 | avg_load = 0; | |
1446 | ||
1447 | for_each_cpu(i, sched_group_cpus(group)) { | |
1448 | /* Bias balancing toward cpus of our domain */ | |
1449 | if (local_group) | |
1450 | load = source_load(i, load_idx); | |
1451 | else | |
1452 | load = target_load(i, load_idx); | |
1453 | ||
1454 | avg_load += load; | |
1455 | } | |
1456 | ||
1457 | /* Adjust by relative CPU power of the group */ | |
1458 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1459 | ||
1460 | if (local_group) { | |
1461 | this_load = avg_load; | |
aaee1203 PZ |
1462 | } else if (avg_load < min_load) { |
1463 | min_load = avg_load; | |
1464 | idlest = group; | |
1465 | } | |
1466 | } while (group = group->next, group != sd->groups); | |
1467 | ||
1468 | if (!idlest || 100*this_load < imbalance*min_load) | |
1469 | return NULL; | |
1470 | return idlest; | |
1471 | } | |
1472 | ||
1473 | /* | |
1474 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1475 | */ | |
1476 | static int | |
1477 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1478 | { | |
1479 | unsigned long load, min_load = ULONG_MAX; | |
1480 | int idlest = -1; | |
1481 | int i; | |
1482 | ||
1483 | /* Traverse only the allowed CPUs */ | |
1484 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1485 | load = weighted_cpuload(i); | |
1486 | ||
1487 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1488 | min_load = load; | |
1489 | idlest = i; | |
e7693a36 GH |
1490 | } |
1491 | } | |
1492 | ||
aaee1203 PZ |
1493 | return idlest; |
1494 | } | |
e7693a36 | 1495 | |
a50bde51 PZ |
1496 | /* |
1497 | * Try and locate an idle CPU in the sched_domain. | |
1498 | */ | |
99bd5e2f | 1499 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
1500 | { |
1501 | int cpu = smp_processor_id(); | |
1502 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 1503 | struct sched_domain *sd; |
a50bde51 PZ |
1504 | int i; |
1505 | ||
1506 | /* | |
99bd5e2f SS |
1507 | * If the task is going to be woken-up on this cpu and if it is |
1508 | * already idle, then it is the right target. | |
a50bde51 | 1509 | */ |
99bd5e2f SS |
1510 | if (target == cpu && idle_cpu(cpu)) |
1511 | return cpu; | |
1512 | ||
1513 | /* | |
1514 | * If the task is going to be woken-up on the cpu where it previously | |
1515 | * ran and if it is currently idle, then it the right target. | |
1516 | */ | |
1517 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 1518 | return prev_cpu; |
a50bde51 PZ |
1519 | |
1520 | /* | |
99bd5e2f | 1521 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 1522 | */ |
99bd5e2f SS |
1523 | for_each_domain(target, sd) { |
1524 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 1525 | break; |
99bd5e2f SS |
1526 | |
1527 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1528 | if (idle_cpu(i)) { | |
1529 | target = i; | |
1530 | break; | |
1531 | } | |
a50bde51 | 1532 | } |
99bd5e2f SS |
1533 | |
1534 | /* | |
1535 | * Lets stop looking for an idle sibling when we reached | |
1536 | * the domain that spans the current cpu and prev_cpu. | |
1537 | */ | |
1538 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1539 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1540 | break; | |
a50bde51 PZ |
1541 | } |
1542 | ||
1543 | return target; | |
1544 | } | |
1545 | ||
aaee1203 PZ |
1546 | /* |
1547 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1548 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1549 | * SD_BALANCE_EXEC. | |
1550 | * | |
1551 | * Balance, ie. select the least loaded group. | |
1552 | * | |
1553 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1554 | * | |
1555 | * preempt must be disabled. | |
1556 | */ | |
0017d735 PZ |
1557 | static int |
1558 | select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags) | |
aaee1203 | 1559 | { |
29cd8bae | 1560 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1561 | int cpu = smp_processor_id(); |
1562 | int prev_cpu = task_cpu(p); | |
1563 | int new_cpu = cpu; | |
99bd5e2f | 1564 | int want_affine = 0; |
29cd8bae | 1565 | int want_sd = 1; |
5158f4e4 | 1566 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1567 | |
0763a660 | 1568 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 1569 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
1570 | want_affine = 1; |
1571 | new_cpu = prev_cpu; | |
1572 | } | |
aaee1203 PZ |
1573 | |
1574 | for_each_domain(cpu, tmp) { | |
e4f42888 PZ |
1575 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1576 | continue; | |
1577 | ||
aaee1203 | 1578 | /* |
ae154be1 PZ |
1579 | * If power savings logic is enabled for a domain, see if we |
1580 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1581 | */ |
59abf026 | 1582 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1583 | unsigned long power = 0; |
1584 | unsigned long nr_running = 0; | |
1585 | unsigned long capacity; | |
1586 | int i; | |
1587 | ||
1588 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1589 | power += power_of(i); | |
1590 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1591 | } | |
1592 | ||
1593 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1594 | ||
59abf026 PZ |
1595 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1596 | nr_running /= 2; | |
1597 | ||
1598 | if (nr_running < capacity) | |
29cd8bae | 1599 | want_sd = 0; |
ae154be1 | 1600 | } |
aaee1203 | 1601 | |
fe3bcfe1 | 1602 | /* |
99bd5e2f SS |
1603 | * If both cpu and prev_cpu are part of this domain, |
1604 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 1605 | */ |
99bd5e2f SS |
1606 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1607 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1608 | affine_sd = tmp; | |
1609 | want_affine = 0; | |
c88d5910 PZ |
1610 | } |
1611 | ||
29cd8bae PZ |
1612 | if (!want_sd && !want_affine) |
1613 | break; | |
1614 | ||
0763a660 | 1615 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1616 | continue; |
1617 | ||
29cd8bae PZ |
1618 | if (want_sd) |
1619 | sd = tmp; | |
1620 | } | |
1621 | ||
8b911acd | 1622 | if (affine_sd) { |
99bd5e2f SS |
1623 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
1624 | return select_idle_sibling(p, cpu); | |
1625 | else | |
1626 | return select_idle_sibling(p, prev_cpu); | |
8b911acd | 1627 | } |
e7693a36 | 1628 | |
aaee1203 | 1629 | while (sd) { |
5158f4e4 | 1630 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1631 | struct sched_group *group; |
c88d5910 | 1632 | int weight; |
098fb9db | 1633 | |
0763a660 | 1634 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1635 | sd = sd->child; |
1636 | continue; | |
1637 | } | |
098fb9db | 1638 | |
5158f4e4 PZ |
1639 | if (sd_flag & SD_BALANCE_WAKE) |
1640 | load_idx = sd->wake_idx; | |
098fb9db | 1641 | |
5158f4e4 | 1642 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1643 | if (!group) { |
1644 | sd = sd->child; | |
1645 | continue; | |
1646 | } | |
4ae7d5ce | 1647 | |
d7c33c49 | 1648 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1649 | if (new_cpu == -1 || new_cpu == cpu) { |
1650 | /* Now try balancing at a lower domain level of cpu */ | |
1651 | sd = sd->child; | |
1652 | continue; | |
e7693a36 | 1653 | } |
aaee1203 PZ |
1654 | |
1655 | /* Now try balancing at a lower domain level of new_cpu */ | |
1656 | cpu = new_cpu; | |
669c55e9 | 1657 | weight = sd->span_weight; |
aaee1203 PZ |
1658 | sd = NULL; |
1659 | for_each_domain(cpu, tmp) { | |
669c55e9 | 1660 | if (weight <= tmp->span_weight) |
aaee1203 | 1661 | break; |
0763a660 | 1662 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1663 | sd = tmp; |
1664 | } | |
1665 | /* while loop will break here if sd == NULL */ | |
e7693a36 GH |
1666 | } |
1667 | ||
c88d5910 | 1668 | return new_cpu; |
e7693a36 GH |
1669 | } |
1670 | #endif /* CONFIG_SMP */ | |
1671 | ||
e52fb7c0 PZ |
1672 | static unsigned long |
1673 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1674 | { |
1675 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1676 | ||
1677 | /* | |
e52fb7c0 PZ |
1678 | * Since its curr running now, convert the gran from real-time |
1679 | * to virtual-time in his units. | |
13814d42 MG |
1680 | * |
1681 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1682 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1683 | * the resulting gran will be larger, therefore penalizing the | |
1684 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1685 | * be smaller, again penalizing the lighter task. | |
1686 | * | |
1687 | * This is especially important for buddies when the leftmost | |
1688 | * task is higher priority than the buddy. | |
0bbd3336 | 1689 | */ |
13814d42 MG |
1690 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
1691 | gran = calc_delta_fair(gran, se); | |
0bbd3336 PZ |
1692 | |
1693 | return gran; | |
1694 | } | |
1695 | ||
464b7527 PZ |
1696 | /* |
1697 | * Should 'se' preempt 'curr'. | |
1698 | * | |
1699 | * |s1 | |
1700 | * |s2 | |
1701 | * |s3 | |
1702 | * g | |
1703 | * |<--->|c | |
1704 | * | |
1705 | * w(c, s1) = -1 | |
1706 | * w(c, s2) = 0 | |
1707 | * w(c, s3) = 1 | |
1708 | * | |
1709 | */ | |
1710 | static int | |
1711 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1712 | { | |
1713 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1714 | ||
1715 | if (vdiff <= 0) | |
1716 | return -1; | |
1717 | ||
e52fb7c0 | 1718 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1719 | if (vdiff > gran) |
1720 | return 1; | |
1721 | ||
1722 | return 0; | |
1723 | } | |
1724 | ||
02479099 PZ |
1725 | static void set_last_buddy(struct sched_entity *se) |
1726 | { | |
6bc912b7 PZ |
1727 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1728 | for_each_sched_entity(se) | |
1729 | cfs_rq_of(se)->last = se; | |
1730 | } | |
02479099 PZ |
1731 | } |
1732 | ||
1733 | static void set_next_buddy(struct sched_entity *se) | |
1734 | { | |
6bc912b7 PZ |
1735 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1736 | for_each_sched_entity(se) | |
1737 | cfs_rq_of(se)->next = se; | |
1738 | } | |
02479099 PZ |
1739 | } |
1740 | ||
bf0f6f24 IM |
1741 | /* |
1742 | * Preempt the current task with a newly woken task if needed: | |
1743 | */ | |
5a9b86f6 | 1744 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1745 | { |
1746 | struct task_struct *curr = rq->curr; | |
8651a86c | 1747 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1748 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 1749 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
bf0f6f24 | 1750 | |
3a7e73a2 PZ |
1751 | if (unlikely(rt_prio(p->prio))) |
1752 | goto preempt; | |
aa2ac252 | 1753 | |
d95f98d0 PZ |
1754 | if (unlikely(p->sched_class != &fair_sched_class)) |
1755 | return; | |
1756 | ||
4ae7d5ce IM |
1757 | if (unlikely(se == pse)) |
1758 | return; | |
1759 | ||
f685ceac | 1760 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) |
3cb63d52 | 1761 | set_next_buddy(pse); |
57fdc26d | 1762 | |
aec0a514 BR |
1763 | /* |
1764 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1765 | * wake up path. | |
1766 | */ | |
1767 | if (test_tsk_need_resched(curr)) | |
1768 | return; | |
1769 | ||
91c234b4 | 1770 | /* |
6bc912b7 | 1771 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1772 | * the tick): |
1773 | */ | |
6bc912b7 | 1774 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1775 | return; |
bf0f6f24 | 1776 | |
6bc912b7 | 1777 | /* Idle tasks are by definition preempted by everybody. */ |
3a7e73a2 PZ |
1778 | if (unlikely(curr->policy == SCHED_IDLE)) |
1779 | goto preempt; | |
bf0f6f24 | 1780 | |
ad4b78bb PZ |
1781 | if (!sched_feat(WAKEUP_PREEMPT)) |
1782 | return; | |
1783 | ||
3a7e73a2 | 1784 | update_curr(cfs_rq); |
464b7527 | 1785 | find_matching_se(&se, &pse); |
002f128b | 1786 | BUG_ON(!pse); |
3a7e73a2 PZ |
1787 | if (wakeup_preempt_entity(se, pse) == 1) |
1788 | goto preempt; | |
464b7527 | 1789 | |
3a7e73a2 | 1790 | return; |
a65ac745 | 1791 | |
3a7e73a2 PZ |
1792 | preempt: |
1793 | resched_task(curr); | |
1794 | /* | |
1795 | * Only set the backward buddy when the current task is still | |
1796 | * on the rq. This can happen when a wakeup gets interleaved | |
1797 | * with schedule on the ->pre_schedule() or idle_balance() | |
1798 | * point, either of which can * drop the rq lock. | |
1799 | * | |
1800 | * Also, during early boot the idle thread is in the fair class, | |
1801 | * for obvious reasons its a bad idea to schedule back to it. | |
1802 | */ | |
1803 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1804 | return; | |
1805 | ||
1806 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1807 | set_last_buddy(se); | |
bf0f6f24 IM |
1808 | } |
1809 | ||
fb8d4724 | 1810 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1811 | { |
8f4d37ec | 1812 | struct task_struct *p; |
bf0f6f24 IM |
1813 | struct cfs_rq *cfs_rq = &rq->cfs; |
1814 | struct sched_entity *se; | |
1815 | ||
36ace27e | 1816 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
1817 | return NULL; |
1818 | ||
1819 | do { | |
9948f4b2 | 1820 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1821 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1822 | cfs_rq = group_cfs_rq(se); |
1823 | } while (cfs_rq); | |
1824 | ||
8f4d37ec PZ |
1825 | p = task_of(se); |
1826 | hrtick_start_fair(rq, p); | |
1827 | ||
1828 | return p; | |
bf0f6f24 IM |
1829 | } |
1830 | ||
1831 | /* | |
1832 | * Account for a descheduled task: | |
1833 | */ | |
31ee529c | 1834 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1835 | { |
1836 | struct sched_entity *se = &prev->se; | |
1837 | struct cfs_rq *cfs_rq; | |
1838 | ||
1839 | for_each_sched_entity(se) { | |
1840 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1841 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1842 | } |
1843 | } | |
1844 | ||
681f3e68 | 1845 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1846 | /************************************************** |
1847 | * Fair scheduling class load-balancing methods: | |
1848 | */ | |
1849 | ||
1e3c88bd PZ |
1850 | /* |
1851 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1852 | * Both runqueues must be locked. | |
1853 | */ | |
1854 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
1855 | struct rq *this_rq, int this_cpu) | |
1856 | { | |
1857 | deactivate_task(src_rq, p, 0); | |
1858 | set_task_cpu(p, this_cpu); | |
1859 | activate_task(this_rq, p, 0); | |
1860 | check_preempt_curr(this_rq, p, 0); | |
fab47622 NR |
1861 | |
1862 | /* re-arm NEWIDLE balancing when moving tasks */ | |
1863 | src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost; | |
1864 | this_rq->idle_stamp = 0; | |
1e3c88bd PZ |
1865 | } |
1866 | ||
1867 | /* | |
1868 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1869 | */ | |
1870 | static | |
1871 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
1872 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1873 | int *all_pinned) | |
1874 | { | |
1875 | int tsk_cache_hot = 0; | |
1876 | /* | |
1877 | * We do not migrate tasks that are: | |
1878 | * 1) running (obviously), or | |
1879 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1880 | * 3) are cache-hot on their current CPU. | |
1881 | */ | |
1882 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 1883 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
1884 | return 0; |
1885 | } | |
1886 | *all_pinned = 0; | |
1887 | ||
1888 | if (task_running(rq, p)) { | |
41acab88 | 1889 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
1890 | return 0; |
1891 | } | |
1892 | ||
1893 | /* | |
1894 | * Aggressive migration if: | |
1895 | * 1) task is cache cold, or | |
1896 | * 2) too many balance attempts have failed. | |
1897 | */ | |
1898 | ||
305e6835 | 1899 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
1900 | if (!tsk_cache_hot || |
1901 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
1902 | #ifdef CONFIG_SCHEDSTATS | |
1903 | if (tsk_cache_hot) { | |
1904 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 1905 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
1906 | } |
1907 | #endif | |
1908 | return 1; | |
1909 | } | |
1910 | ||
1911 | if (tsk_cache_hot) { | |
41acab88 | 1912 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
1913 | return 0; |
1914 | } | |
1915 | return 1; | |
1916 | } | |
1917 | ||
897c395f PZ |
1918 | /* |
1919 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
1920 | * part of active balancing operations within "domain". | |
1921 | * Returns 1 if successful and 0 otherwise. | |
1922 | * | |
1923 | * Called with both runqueues locked. | |
1924 | */ | |
1925 | static int | |
1926 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1927 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1928 | { | |
1929 | struct task_struct *p, *n; | |
1930 | struct cfs_rq *cfs_rq; | |
1931 | int pinned = 0; | |
1932 | ||
1933 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
1934 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
1935 | ||
1936 | if (!can_migrate_task(p, busiest, this_cpu, | |
1937 | sd, idle, &pinned)) | |
1938 | continue; | |
1939 | ||
1940 | pull_task(busiest, p, this_rq, this_cpu); | |
1941 | /* | |
1942 | * Right now, this is only the second place pull_task() | |
1943 | * is called, so we can safely collect pull_task() | |
1944 | * stats here rather than inside pull_task(). | |
1945 | */ | |
1946 | schedstat_inc(sd, lb_gained[idle]); | |
1947 | return 1; | |
1948 | } | |
1949 | } | |
1950 | ||
1951 | return 0; | |
1952 | } | |
1953 | ||
1e3c88bd PZ |
1954 | static unsigned long |
1955 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1956 | unsigned long max_load_move, struct sched_domain *sd, | |
1957 | enum cpu_idle_type idle, int *all_pinned, | |
ee00e66f | 1958 | int *this_best_prio, struct cfs_rq *busiest_cfs_rq) |
1e3c88bd PZ |
1959 | { |
1960 | int loops = 0, pulled = 0, pinned = 0; | |
1e3c88bd | 1961 | long rem_load_move = max_load_move; |
ee00e66f | 1962 | struct task_struct *p, *n; |
1e3c88bd PZ |
1963 | |
1964 | if (max_load_move == 0) | |
1965 | goto out; | |
1966 | ||
1967 | pinned = 1; | |
1968 | ||
ee00e66f PZ |
1969 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
1970 | if (loops++ > sysctl_sched_nr_migrate) | |
1971 | break; | |
1e3c88bd | 1972 | |
ee00e66f PZ |
1973 | if ((p->se.load.weight >> 1) > rem_load_move || |
1974 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) | |
1975 | continue; | |
1e3c88bd | 1976 | |
ee00e66f PZ |
1977 | pull_task(busiest, p, this_rq, this_cpu); |
1978 | pulled++; | |
1979 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
1980 | |
1981 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
1982 | /* |
1983 | * NEWIDLE balancing is a source of latency, so preemptible | |
1984 | * kernels will stop after the first task is pulled to minimize | |
1985 | * the critical section. | |
1986 | */ | |
1987 | if (idle == CPU_NEWLY_IDLE) | |
1988 | break; | |
1e3c88bd PZ |
1989 | #endif |
1990 | ||
ee00e66f PZ |
1991 | /* |
1992 | * We only want to steal up to the prescribed amount of | |
1993 | * weighted load. | |
1994 | */ | |
1995 | if (rem_load_move <= 0) | |
1996 | break; | |
1997 | ||
1e3c88bd PZ |
1998 | if (p->prio < *this_best_prio) |
1999 | *this_best_prio = p->prio; | |
1e3c88bd PZ |
2000 | } |
2001 | out: | |
2002 | /* | |
2003 | * Right now, this is one of only two places pull_task() is called, | |
2004 | * so we can safely collect pull_task() stats here rather than | |
2005 | * inside pull_task(). | |
2006 | */ | |
2007 | schedstat_add(sd, lb_gained[idle], pulled); | |
2008 | ||
2009 | if (all_pinned) | |
2010 | *all_pinned = pinned; | |
2011 | ||
2012 | return max_load_move - rem_load_move; | |
2013 | } | |
2014 | ||
230059de | 2015 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
2016 | /* |
2017 | * update tg->load_weight by folding this cpu's load_avg | |
2018 | */ | |
2019 | static int tg_shares_up(struct task_group *tg, int cpu) | |
2020 | { | |
2021 | struct cfs_rq *cfs_rq; | |
2022 | unsigned long flags; | |
2023 | struct rq *rq; | |
2024 | long load_avg; | |
2025 | ||
2026 | if (!tg->se[cpu]) | |
2027 | return 0; | |
2028 | ||
2029 | rq = cpu_rq(cpu); | |
2030 | cfs_rq = tg->cfs_rq[cpu]; | |
2031 | ||
2032 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2033 | ||
2034 | update_rq_clock(rq); | |
e33078ba | 2035 | update_cfs_load(cfs_rq); |
9e3081ca PZ |
2036 | |
2037 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
2038 | load_avg -= cfs_rq->load_contribution; | |
2039 | atomic_add(load_avg, &tg->load_weight); | |
2040 | cfs_rq->load_contribution += load_avg; | |
2041 | ||
2042 | /* | |
2043 | * We need to update shares after updating tg->load_weight in | |
2044 | * order to adjust the weight of groups with long running tasks. | |
2045 | */ | |
f0d7442a | 2046 | update_cfs_shares(cfs_rq, 0); |
9e3081ca PZ |
2047 | |
2048 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2049 | ||
2050 | return 0; | |
2051 | } | |
2052 | ||
2053 | static void update_shares(int cpu) | |
2054 | { | |
2055 | struct cfs_rq *cfs_rq; | |
2056 | struct rq *rq = cpu_rq(cpu); | |
2057 | ||
2058 | rcu_read_lock(); | |
2059 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2060 | struct task_group *tg = cfs_rq->tg; | |
2061 | ||
2062 | do { | |
2063 | tg_shares_up(tg, cpu); | |
2064 | tg = tg->parent; | |
2065 | } while (tg); | |
2066 | } | |
2067 | rcu_read_unlock(); | |
2068 | } | |
2069 | ||
230059de PZ |
2070 | static unsigned long |
2071 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2072 | unsigned long max_load_move, | |
2073 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2074 | int *all_pinned, int *this_best_prio) | |
2075 | { | |
2076 | long rem_load_move = max_load_move; | |
2077 | int busiest_cpu = cpu_of(busiest); | |
2078 | struct task_group *tg; | |
2079 | ||
2080 | rcu_read_lock(); | |
2081 | update_h_load(busiest_cpu); | |
2082 | ||
2083 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
2084 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; | |
2085 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; | |
2086 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
2087 | u64 rem_load, moved_load; | |
2088 | ||
2089 | /* | |
2090 | * empty group | |
2091 | */ | |
2092 | if (!busiest_cfs_rq->task_weight) | |
2093 | continue; | |
2094 | ||
2095 | rem_load = (u64)rem_load_move * busiest_weight; | |
2096 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
2097 | ||
2098 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
2099 | rem_load, sd, idle, all_pinned, this_best_prio, | |
2100 | busiest_cfs_rq); | |
2101 | ||
2102 | if (!moved_load) | |
2103 | continue; | |
2104 | ||
2105 | moved_load *= busiest_h_load; | |
2106 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
2107 | ||
2108 | rem_load_move -= moved_load; | |
2109 | if (rem_load_move < 0) | |
2110 | break; | |
2111 | } | |
2112 | rcu_read_unlock(); | |
2113 | ||
2114 | return max_load_move - rem_load_move; | |
2115 | } | |
2116 | #else | |
9e3081ca PZ |
2117 | static inline void update_shares(int cpu) |
2118 | { | |
2119 | } | |
2120 | ||
230059de PZ |
2121 | static unsigned long |
2122 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2123 | unsigned long max_load_move, | |
2124 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2125 | int *all_pinned, int *this_best_prio) | |
2126 | { | |
2127 | return balance_tasks(this_rq, this_cpu, busiest, | |
2128 | max_load_move, sd, idle, all_pinned, | |
2129 | this_best_prio, &busiest->cfs); | |
2130 | } | |
2131 | #endif | |
2132 | ||
1e3c88bd PZ |
2133 | /* |
2134 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2135 | * this_rq, as part of a balancing operation within domain "sd". | |
2136 | * Returns 1 if successful and 0 otherwise. | |
2137 | * | |
2138 | * Called with both runqueues locked. | |
2139 | */ | |
2140 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2141 | unsigned long max_load_move, | |
2142 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2143 | int *all_pinned) | |
2144 | { | |
3d45fd80 | 2145 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2146 | int this_best_prio = this_rq->curr->prio; |
2147 | ||
2148 | do { | |
3d45fd80 | 2149 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd PZ |
2150 | max_load_move - total_load_moved, |
2151 | sd, idle, all_pinned, &this_best_prio); | |
3d45fd80 PZ |
2152 | |
2153 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2154 | |
2155 | #ifdef CONFIG_PREEMPT | |
2156 | /* | |
2157 | * NEWIDLE balancing is a source of latency, so preemptible | |
2158 | * kernels will stop after the first task is pulled to minimize | |
2159 | * the critical section. | |
2160 | */ | |
2161 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2162 | break; | |
baa8c110 PZ |
2163 | |
2164 | if (raw_spin_is_contended(&this_rq->lock) || | |
2165 | raw_spin_is_contended(&busiest->lock)) | |
2166 | break; | |
1e3c88bd | 2167 | #endif |
3d45fd80 | 2168 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2169 | |
2170 | return total_load_moved > 0; | |
2171 | } | |
2172 | ||
1e3c88bd PZ |
2173 | /********** Helpers for find_busiest_group ************************/ |
2174 | /* | |
2175 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2176 | * during load balancing. | |
2177 | */ | |
2178 | struct sd_lb_stats { | |
2179 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2180 | struct sched_group *this; /* Local group in this sd */ | |
2181 | unsigned long total_load; /* Total load of all groups in sd */ | |
2182 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2183 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2184 | ||
2185 | /** Statistics of this group */ | |
2186 | unsigned long this_load; | |
2187 | unsigned long this_load_per_task; | |
2188 | unsigned long this_nr_running; | |
fab47622 | 2189 | unsigned long this_has_capacity; |
1e3c88bd PZ |
2190 | |
2191 | /* Statistics of the busiest group */ | |
2192 | unsigned long max_load; | |
2193 | unsigned long busiest_load_per_task; | |
2194 | unsigned long busiest_nr_running; | |
dd5feea1 | 2195 | unsigned long busiest_group_capacity; |
fab47622 | 2196 | unsigned long busiest_has_capacity; |
1e3c88bd PZ |
2197 | |
2198 | int group_imb; /* Is there imbalance in this sd */ | |
2199 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2200 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2201 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2202 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2203 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2204 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2205 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2206 | #endif | |
2207 | }; | |
2208 | ||
2209 | /* | |
2210 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2211 | */ | |
2212 | struct sg_lb_stats { | |
2213 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2214 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2215 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2216 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2217 | unsigned long group_capacity; | |
2218 | int group_imb; /* Is there an imbalance in the group ? */ | |
fab47622 | 2219 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2220 | }; |
2221 | ||
2222 | /** | |
2223 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2224 | * @group: The group whose first cpu is to be returned. | |
2225 | */ | |
2226 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2227 | { | |
2228 | return cpumask_first(sched_group_cpus(group)); | |
2229 | } | |
2230 | ||
2231 | /** | |
2232 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2233 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2234 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2235 | */ | |
2236 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2237 | enum cpu_idle_type idle) | |
2238 | { | |
2239 | int load_idx; | |
2240 | ||
2241 | switch (idle) { | |
2242 | case CPU_NOT_IDLE: | |
2243 | load_idx = sd->busy_idx; | |
2244 | break; | |
2245 | ||
2246 | case CPU_NEWLY_IDLE: | |
2247 | load_idx = sd->newidle_idx; | |
2248 | break; | |
2249 | default: | |
2250 | load_idx = sd->idle_idx; | |
2251 | break; | |
2252 | } | |
2253 | ||
2254 | return load_idx; | |
2255 | } | |
2256 | ||
2257 | ||
2258 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2259 | /** | |
2260 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2261 | * the given sched_domain, during load balancing. | |
2262 | * | |
2263 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2264 | * @sds: Variable containing the statistics for sd. | |
2265 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2266 | */ | |
2267 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2268 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2269 | { | |
2270 | /* | |
2271 | * Busy processors will not participate in power savings | |
2272 | * balance. | |
2273 | */ | |
2274 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2275 | sds->power_savings_balance = 0; | |
2276 | else { | |
2277 | sds->power_savings_balance = 1; | |
2278 | sds->min_nr_running = ULONG_MAX; | |
2279 | sds->leader_nr_running = 0; | |
2280 | } | |
2281 | } | |
2282 | ||
2283 | /** | |
2284 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2285 | * sched_domain while performing load balancing. | |
2286 | * | |
2287 | * @group: sched_group belonging to the sched_domain under consideration. | |
2288 | * @sds: Variable containing the statistics of the sched_domain | |
2289 | * @local_group: Does group contain the CPU for which we're performing | |
2290 | * load balancing ? | |
2291 | * @sgs: Variable containing the statistics of the group. | |
2292 | */ | |
2293 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2294 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2295 | { | |
2296 | ||
2297 | if (!sds->power_savings_balance) | |
2298 | return; | |
2299 | ||
2300 | /* | |
2301 | * If the local group is idle or completely loaded | |
2302 | * no need to do power savings balance at this domain | |
2303 | */ | |
2304 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2305 | !sds->this_nr_running)) | |
2306 | sds->power_savings_balance = 0; | |
2307 | ||
2308 | /* | |
2309 | * If a group is already running at full capacity or idle, | |
2310 | * don't include that group in power savings calculations | |
2311 | */ | |
2312 | if (!sds->power_savings_balance || | |
2313 | sgs->sum_nr_running >= sgs->group_capacity || | |
2314 | !sgs->sum_nr_running) | |
2315 | return; | |
2316 | ||
2317 | /* | |
2318 | * Calculate the group which has the least non-idle load. | |
2319 | * This is the group from where we need to pick up the load | |
2320 | * for saving power | |
2321 | */ | |
2322 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2323 | (sgs->sum_nr_running == sds->min_nr_running && | |
2324 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2325 | sds->group_min = group; | |
2326 | sds->min_nr_running = sgs->sum_nr_running; | |
2327 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2328 | sgs->sum_nr_running; | |
2329 | } | |
2330 | ||
2331 | /* | |
2332 | * Calculate the group which is almost near its | |
2333 | * capacity but still has some space to pick up some load | |
2334 | * from other group and save more power | |
2335 | */ | |
2336 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2337 | return; | |
2338 | ||
2339 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2340 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2341 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2342 | sds->group_leader = group; | |
2343 | sds->leader_nr_running = sgs->sum_nr_running; | |
2344 | } | |
2345 | } | |
2346 | ||
2347 | /** | |
2348 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2349 | * @sds: Variable containing the statistics of the sched_domain | |
2350 | * under consideration. | |
2351 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2352 | * @imbalance: Variable to store the imbalance. | |
2353 | * | |
2354 | * Description: | |
2355 | * Check if we have potential to perform some power-savings balance. | |
2356 | * If yes, set the busiest group to be the least loaded group in the | |
2357 | * sched_domain, so that it's CPUs can be put to idle. | |
2358 | * | |
2359 | * Returns 1 if there is potential to perform power-savings balance. | |
2360 | * Else returns 0. | |
2361 | */ | |
2362 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2363 | int this_cpu, unsigned long *imbalance) | |
2364 | { | |
2365 | if (!sds->power_savings_balance) | |
2366 | return 0; | |
2367 | ||
2368 | if (sds->this != sds->group_leader || | |
2369 | sds->group_leader == sds->group_min) | |
2370 | return 0; | |
2371 | ||
2372 | *imbalance = sds->min_load_per_task; | |
2373 | sds->busiest = sds->group_min; | |
2374 | ||
2375 | return 1; | |
2376 | ||
2377 | } | |
2378 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2379 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2380 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2381 | { | |
2382 | return; | |
2383 | } | |
2384 | ||
2385 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2386 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2387 | { | |
2388 | return; | |
2389 | } | |
2390 | ||
2391 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2392 | int this_cpu, unsigned long *imbalance) | |
2393 | { | |
2394 | return 0; | |
2395 | } | |
2396 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2397 | ||
2398 | ||
2399 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2400 | { | |
2401 | return SCHED_LOAD_SCALE; | |
2402 | } | |
2403 | ||
2404 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2405 | { | |
2406 | return default_scale_freq_power(sd, cpu); | |
2407 | } | |
2408 | ||
2409 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2410 | { | |
669c55e9 | 2411 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2412 | unsigned long smt_gain = sd->smt_gain; |
2413 | ||
2414 | smt_gain /= weight; | |
2415 | ||
2416 | return smt_gain; | |
2417 | } | |
2418 | ||
2419 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2420 | { | |
2421 | return default_scale_smt_power(sd, cpu); | |
2422 | } | |
2423 | ||
2424 | unsigned long scale_rt_power(int cpu) | |
2425 | { | |
2426 | struct rq *rq = cpu_rq(cpu); | |
2427 | u64 total, available; | |
2428 | ||
1e3c88bd | 2429 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
2430 | |
2431 | if (unlikely(total < rq->rt_avg)) { | |
2432 | /* Ensures that power won't end up being negative */ | |
2433 | available = 0; | |
2434 | } else { | |
2435 | available = total - rq->rt_avg; | |
2436 | } | |
1e3c88bd PZ |
2437 | |
2438 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
2439 | total = SCHED_LOAD_SCALE; | |
2440 | ||
2441 | total >>= SCHED_LOAD_SHIFT; | |
2442 | ||
2443 | return div_u64(available, total); | |
2444 | } | |
2445 | ||
2446 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2447 | { | |
669c55e9 | 2448 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2449 | unsigned long power = SCHED_LOAD_SCALE; |
2450 | struct sched_group *sdg = sd->groups; | |
2451 | ||
1e3c88bd PZ |
2452 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
2453 | if (sched_feat(ARCH_POWER)) | |
2454 | power *= arch_scale_smt_power(sd, cpu); | |
2455 | else | |
2456 | power *= default_scale_smt_power(sd, cpu); | |
2457 | ||
2458 | power >>= SCHED_LOAD_SHIFT; | |
2459 | } | |
2460 | ||
9d5efe05 SV |
2461 | sdg->cpu_power_orig = power; |
2462 | ||
2463 | if (sched_feat(ARCH_POWER)) | |
2464 | power *= arch_scale_freq_power(sd, cpu); | |
2465 | else | |
2466 | power *= default_scale_freq_power(sd, cpu); | |
2467 | ||
2468 | power >>= SCHED_LOAD_SHIFT; | |
2469 | ||
1e3c88bd PZ |
2470 | power *= scale_rt_power(cpu); |
2471 | power >>= SCHED_LOAD_SHIFT; | |
2472 | ||
2473 | if (!power) | |
2474 | power = 1; | |
2475 | ||
e51fd5e2 | 2476 | cpu_rq(cpu)->cpu_power = power; |
1e3c88bd PZ |
2477 | sdg->cpu_power = power; |
2478 | } | |
2479 | ||
2480 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2481 | { | |
2482 | struct sched_domain *child = sd->child; | |
2483 | struct sched_group *group, *sdg = sd->groups; | |
2484 | unsigned long power; | |
2485 | ||
2486 | if (!child) { | |
2487 | update_cpu_power(sd, cpu); | |
2488 | return; | |
2489 | } | |
2490 | ||
2491 | power = 0; | |
2492 | ||
2493 | group = child->groups; | |
2494 | do { | |
2495 | power += group->cpu_power; | |
2496 | group = group->next; | |
2497 | } while (group != child->groups); | |
2498 | ||
2499 | sdg->cpu_power = power; | |
2500 | } | |
2501 | ||
9d5efe05 SV |
2502 | /* |
2503 | * Try and fix up capacity for tiny siblings, this is needed when | |
2504 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
2505 | * which on its own isn't powerful enough. | |
2506 | * | |
2507 | * See update_sd_pick_busiest() and check_asym_packing(). | |
2508 | */ | |
2509 | static inline int | |
2510 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
2511 | { | |
2512 | /* | |
2513 | * Only siblings can have significantly less than SCHED_LOAD_SCALE | |
2514 | */ | |
2515 | if (sd->level != SD_LV_SIBLING) | |
2516 | return 0; | |
2517 | ||
2518 | /* | |
2519 | * If ~90% of the cpu_power is still there, we're good. | |
2520 | */ | |
694f5a11 | 2521 | if (group->cpu_power * 32 > group->cpu_power_orig * 29) |
9d5efe05 SV |
2522 | return 1; |
2523 | ||
2524 | return 0; | |
2525 | } | |
2526 | ||
1e3c88bd PZ |
2527 | /** |
2528 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2529 | * @sd: The sched_domain whose statistics are to be updated. | |
2530 | * @group: sched_group whose statistics are to be updated. | |
2531 | * @this_cpu: Cpu for which load balance is currently performed. | |
2532 | * @idle: Idle status of this_cpu | |
2533 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
2534 | * @sd_idle: Idle status of the sched_domain containing group. | |
2535 | * @local_group: Does group contain this_cpu. | |
2536 | * @cpus: Set of cpus considered for load balancing. | |
2537 | * @balance: Should we balance. | |
2538 | * @sgs: variable to hold the statistics for this group. | |
2539 | */ | |
2540 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2541 | struct sched_group *group, int this_cpu, | |
2542 | enum cpu_idle_type idle, int load_idx, int *sd_idle, | |
2543 | int local_group, const struct cpumask *cpus, | |
2544 | int *balance, struct sg_lb_stats *sgs) | |
2545 | { | |
2582f0eb | 2546 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
2547 | int i; |
2548 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 2549 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 2550 | |
871e35bc | 2551 | if (local_group) |
1e3c88bd | 2552 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
2553 | |
2554 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
2555 | max_cpu_load = 0; |
2556 | min_cpu_load = ~0UL; | |
2582f0eb | 2557 | max_nr_running = 0; |
1e3c88bd PZ |
2558 | |
2559 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2560 | struct rq *rq = cpu_rq(i); | |
2561 | ||
2562 | if (*sd_idle && rq->nr_running) | |
2563 | *sd_idle = 0; | |
2564 | ||
2565 | /* Bias balancing toward cpus of our domain */ | |
2566 | if (local_group) { | |
2567 | if (idle_cpu(i) && !first_idle_cpu) { | |
2568 | first_idle_cpu = 1; | |
2569 | balance_cpu = i; | |
2570 | } | |
2571 | ||
2572 | load = target_load(i, load_idx); | |
2573 | } else { | |
2574 | load = source_load(i, load_idx); | |
2582f0eb | 2575 | if (load > max_cpu_load) { |
1e3c88bd | 2576 | max_cpu_load = load; |
2582f0eb NR |
2577 | max_nr_running = rq->nr_running; |
2578 | } | |
1e3c88bd PZ |
2579 | if (min_cpu_load > load) |
2580 | min_cpu_load = load; | |
2581 | } | |
2582 | ||
2583 | sgs->group_load += load; | |
2584 | sgs->sum_nr_running += rq->nr_running; | |
2585 | sgs->sum_weighted_load += weighted_cpuload(i); | |
2586 | ||
1e3c88bd PZ |
2587 | } |
2588 | ||
2589 | /* | |
2590 | * First idle cpu or the first cpu(busiest) in this sched group | |
2591 | * is eligible for doing load balancing at this and above | |
2592 | * domains. In the newly idle case, we will allow all the cpu's | |
2593 | * to do the newly idle load balance. | |
2594 | */ | |
bbc8cb5b PZ |
2595 | if (idle != CPU_NEWLY_IDLE && local_group) { |
2596 | if (balance_cpu != this_cpu) { | |
2597 | *balance = 0; | |
2598 | return; | |
2599 | } | |
2600 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
2601 | } |
2602 | ||
2603 | /* Adjust by relative CPU power of the group */ | |
2604 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2605 | ||
1e3c88bd PZ |
2606 | /* |
2607 | * Consider the group unbalanced when the imbalance is larger | |
2608 | * than the average weight of two tasks. | |
2609 | * | |
2610 | * APZ: with cgroup the avg task weight can vary wildly and | |
2611 | * might not be a suitable number - should we keep a | |
2612 | * normalized nr_running number somewhere that negates | |
2613 | * the hierarchy? | |
2614 | */ | |
dd5feea1 SS |
2615 | if (sgs->sum_nr_running) |
2616 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 2617 | |
2582f0eb | 2618 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
2619 | sgs->group_imb = 1; |
2620 | ||
2582f0eb | 2621 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); |
9d5efe05 SV |
2622 | if (!sgs->group_capacity) |
2623 | sgs->group_capacity = fix_small_capacity(sd, group); | |
fab47622 NR |
2624 | |
2625 | if (sgs->group_capacity > sgs->sum_nr_running) | |
2626 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
2627 | } |
2628 | ||
532cb4c4 MN |
2629 | /** |
2630 | * update_sd_pick_busiest - return 1 on busiest group | |
2631 | * @sd: sched_domain whose statistics are to be checked | |
2632 | * @sds: sched_domain statistics | |
2633 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
2634 | * @sgs: sched_group statistics |
2635 | * @this_cpu: the current cpu | |
532cb4c4 MN |
2636 | * |
2637 | * Determine if @sg is a busier group than the previously selected | |
2638 | * busiest group. | |
2639 | */ | |
2640 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
2641 | struct sd_lb_stats *sds, | |
2642 | struct sched_group *sg, | |
2643 | struct sg_lb_stats *sgs, | |
2644 | int this_cpu) | |
2645 | { | |
2646 | if (sgs->avg_load <= sds->max_load) | |
2647 | return false; | |
2648 | ||
2649 | if (sgs->sum_nr_running > sgs->group_capacity) | |
2650 | return true; | |
2651 | ||
2652 | if (sgs->group_imb) | |
2653 | return true; | |
2654 | ||
2655 | /* | |
2656 | * ASYM_PACKING needs to move all the work to the lowest | |
2657 | * numbered CPUs in the group, therefore mark all groups | |
2658 | * higher than ourself as busy. | |
2659 | */ | |
2660 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
2661 | this_cpu < group_first_cpu(sg)) { | |
2662 | if (!sds->busiest) | |
2663 | return true; | |
2664 | ||
2665 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
2666 | return true; | |
2667 | } | |
2668 | ||
2669 | return false; | |
2670 | } | |
2671 | ||
1e3c88bd PZ |
2672 | /** |
2673 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2674 | * @sd: sched_domain whose statistics are to be updated. | |
2675 | * @this_cpu: Cpu for which load balance is currently performed. | |
2676 | * @idle: Idle status of this_cpu | |
532cb4c4 | 2677 | * @sd_idle: Idle status of the sched_domain containing sg. |
1e3c88bd PZ |
2678 | * @cpus: Set of cpus considered for load balancing. |
2679 | * @balance: Should we balance. | |
2680 | * @sds: variable to hold the statistics for this sched_domain. | |
2681 | */ | |
2682 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
2683 | enum cpu_idle_type idle, int *sd_idle, | |
2684 | const struct cpumask *cpus, int *balance, | |
2685 | struct sd_lb_stats *sds) | |
2686 | { | |
2687 | struct sched_domain *child = sd->child; | |
532cb4c4 | 2688 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
2689 | struct sg_lb_stats sgs; |
2690 | int load_idx, prefer_sibling = 0; | |
2691 | ||
2692 | if (child && child->flags & SD_PREFER_SIBLING) | |
2693 | prefer_sibling = 1; | |
2694 | ||
2695 | init_sd_power_savings_stats(sd, sds, idle); | |
2696 | load_idx = get_sd_load_idx(sd, idle); | |
2697 | ||
2698 | do { | |
2699 | int local_group; | |
2700 | ||
532cb4c4 | 2701 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 2702 | memset(&sgs, 0, sizeof(sgs)); |
532cb4c4 | 2703 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle, |
1e3c88bd PZ |
2704 | local_group, cpus, balance, &sgs); |
2705 | ||
8f190fb3 | 2706 | if (local_group && !(*balance)) |
1e3c88bd PZ |
2707 | return; |
2708 | ||
2709 | sds->total_load += sgs.group_load; | |
532cb4c4 | 2710 | sds->total_pwr += sg->cpu_power; |
1e3c88bd PZ |
2711 | |
2712 | /* | |
2713 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 2714 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
2715 | * and move all the excess tasks away. We lower the capacity |
2716 | * of a group only if the local group has the capacity to fit | |
2717 | * these excess tasks, i.e. nr_running < group_capacity. The | |
2718 | * extra check prevents the case where you always pull from the | |
2719 | * heaviest group when it is already under-utilized (possible | |
2720 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 2721 | */ |
75dd321d | 2722 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
2723 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
2724 | ||
2725 | if (local_group) { | |
2726 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 2727 | sds->this = sg; |
1e3c88bd PZ |
2728 | sds->this_nr_running = sgs.sum_nr_running; |
2729 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 2730 | sds->this_has_capacity = sgs.group_has_capacity; |
532cb4c4 | 2731 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 2732 | sds->max_load = sgs.avg_load; |
532cb4c4 | 2733 | sds->busiest = sg; |
1e3c88bd | 2734 | sds->busiest_nr_running = sgs.sum_nr_running; |
dd5feea1 | 2735 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 2736 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 2737 | sds->busiest_has_capacity = sgs.group_has_capacity; |
1e3c88bd PZ |
2738 | sds->group_imb = sgs.group_imb; |
2739 | } | |
2740 | ||
532cb4c4 MN |
2741 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
2742 | sg = sg->next; | |
2743 | } while (sg != sd->groups); | |
2744 | } | |
2745 | ||
2ec57d44 | 2746 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
2747 | { |
2748 | return 0*SD_ASYM_PACKING; | |
2749 | } | |
2750 | ||
2751 | /** | |
2752 | * check_asym_packing - Check to see if the group is packed into the | |
2753 | * sched doman. | |
2754 | * | |
2755 | * This is primarily intended to used at the sibling level. Some | |
2756 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
2757 | * case of POWER7, it can move to lower SMT modes only when higher | |
2758 | * threads are idle. When in lower SMT modes, the threads will | |
2759 | * perform better since they share less core resources. Hence when we | |
2760 | * have idle threads, we want them to be the higher ones. | |
2761 | * | |
2762 | * This packing function is run on idle threads. It checks to see if | |
2763 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
2764 | * CPU number than the packing function is being run on. Here we are | |
2765 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
2766 | * number. | |
2767 | * | |
b6b12294 MN |
2768 | * Returns 1 when packing is required and a task should be moved to |
2769 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
2770 | * | |
532cb4c4 MN |
2771 | * @sd: The sched_domain whose packing is to be checked. |
2772 | * @sds: Statistics of the sched_domain which is to be packed | |
2773 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2774 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
2775 | */ |
2776 | static int check_asym_packing(struct sched_domain *sd, | |
2777 | struct sd_lb_stats *sds, | |
2778 | int this_cpu, unsigned long *imbalance) | |
2779 | { | |
2780 | int busiest_cpu; | |
2781 | ||
2782 | if (!(sd->flags & SD_ASYM_PACKING)) | |
2783 | return 0; | |
2784 | ||
2785 | if (!sds->busiest) | |
2786 | return 0; | |
2787 | ||
2788 | busiest_cpu = group_first_cpu(sds->busiest); | |
2789 | if (this_cpu > busiest_cpu) | |
2790 | return 0; | |
2791 | ||
2792 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power, | |
2793 | SCHED_LOAD_SCALE); | |
2794 | return 1; | |
1e3c88bd PZ |
2795 | } |
2796 | ||
2797 | /** | |
2798 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2799 | * amongst the groups of a sched_domain, during | |
2800 | * load balancing. | |
2801 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2802 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2803 | * @imbalance: Variable to store the imbalance. | |
2804 | */ | |
2805 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2806 | int this_cpu, unsigned long *imbalance) | |
2807 | { | |
2808 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2809 | unsigned int imbn = 2; | |
dd5feea1 | 2810 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
2811 | |
2812 | if (sds->this_nr_running) { | |
2813 | sds->this_load_per_task /= sds->this_nr_running; | |
2814 | if (sds->busiest_load_per_task > | |
2815 | sds->this_load_per_task) | |
2816 | imbn = 1; | |
2817 | } else | |
2818 | sds->this_load_per_task = | |
2819 | cpu_avg_load_per_task(this_cpu); | |
2820 | ||
dd5feea1 SS |
2821 | scaled_busy_load_per_task = sds->busiest_load_per_task |
2822 | * SCHED_LOAD_SCALE; | |
2823 | scaled_busy_load_per_task /= sds->busiest->cpu_power; | |
2824 | ||
2825 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
2826 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
2827 | *imbalance = sds->busiest_load_per_task; |
2828 | return; | |
2829 | } | |
2830 | ||
2831 | /* | |
2832 | * OK, we don't have enough imbalance to justify moving tasks, | |
2833 | * however we may be able to increase total CPU power used by | |
2834 | * moving them. | |
2835 | */ | |
2836 | ||
2837 | pwr_now += sds->busiest->cpu_power * | |
2838 | min(sds->busiest_load_per_task, sds->max_load); | |
2839 | pwr_now += sds->this->cpu_power * | |
2840 | min(sds->this_load_per_task, sds->this_load); | |
2841 | pwr_now /= SCHED_LOAD_SCALE; | |
2842 | ||
2843 | /* Amount of load we'd subtract */ | |
2844 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2845 | sds->busiest->cpu_power; | |
2846 | if (sds->max_load > tmp) | |
2847 | pwr_move += sds->busiest->cpu_power * | |
2848 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
2849 | ||
2850 | /* Amount of load we'd add */ | |
2851 | if (sds->max_load * sds->busiest->cpu_power < | |
2852 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
2853 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
2854 | sds->this->cpu_power; | |
2855 | else | |
2856 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2857 | sds->this->cpu_power; | |
2858 | pwr_move += sds->this->cpu_power * | |
2859 | min(sds->this_load_per_task, sds->this_load + tmp); | |
2860 | pwr_move /= SCHED_LOAD_SCALE; | |
2861 | ||
2862 | /* Move if we gain throughput */ | |
2863 | if (pwr_move > pwr_now) | |
2864 | *imbalance = sds->busiest_load_per_task; | |
2865 | } | |
2866 | ||
2867 | /** | |
2868 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
2869 | * groups of a given sched_domain during load balance. | |
2870 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
2871 | * @this_cpu: Cpu for which currently load balance is being performed. | |
2872 | * @imbalance: The variable to store the imbalance. | |
2873 | */ | |
2874 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
2875 | unsigned long *imbalance) | |
2876 | { | |
dd5feea1 SS |
2877 | unsigned long max_pull, load_above_capacity = ~0UL; |
2878 | ||
2879 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
2880 | if (sds->group_imb) { | |
2881 | sds->busiest_load_per_task = | |
2882 | min(sds->busiest_load_per_task, sds->avg_load); | |
2883 | } | |
2884 | ||
1e3c88bd PZ |
2885 | /* |
2886 | * In the presence of smp nice balancing, certain scenarios can have | |
2887 | * max load less than avg load(as we skip the groups at or below | |
2888 | * its cpu_power, while calculating max_load..) | |
2889 | */ | |
2890 | if (sds->max_load < sds->avg_load) { | |
2891 | *imbalance = 0; | |
2892 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2893 | } | |
2894 | ||
dd5feea1 SS |
2895 | if (!sds->group_imb) { |
2896 | /* | |
2897 | * Don't want to pull so many tasks that a group would go idle. | |
2898 | */ | |
2899 | load_above_capacity = (sds->busiest_nr_running - | |
2900 | sds->busiest_group_capacity); | |
2901 | ||
2902 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE); | |
2903 | ||
2904 | load_above_capacity /= sds->busiest->cpu_power; | |
2905 | } | |
2906 | ||
2907 | /* | |
2908 | * We're trying to get all the cpus to the average_load, so we don't | |
2909 | * want to push ourselves above the average load, nor do we wish to | |
2910 | * reduce the max loaded cpu below the average load. At the same time, | |
2911 | * we also don't want to reduce the group load below the group capacity | |
2912 | * (so that we can implement power-savings policies etc). Thus we look | |
2913 | * for the minimum possible imbalance. | |
2914 | * Be careful of negative numbers as they'll appear as very large values | |
2915 | * with unsigned longs. | |
2916 | */ | |
2917 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
2918 | |
2919 | /* How much load to actually move to equalise the imbalance */ | |
2920 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
2921 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
2922 | / SCHED_LOAD_SCALE; | |
2923 | ||
2924 | /* | |
2925 | * if *imbalance is less than the average load per runnable task | |
2926 | * there is no gaurantee that any tasks will be moved so we'll have | |
2927 | * a think about bumping its value to force at least one task to be | |
2928 | * moved | |
2929 | */ | |
2930 | if (*imbalance < sds->busiest_load_per_task) | |
2931 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2932 | ||
2933 | } | |
fab47622 | 2934 | |
1e3c88bd PZ |
2935 | /******* find_busiest_group() helpers end here *********************/ |
2936 | ||
2937 | /** | |
2938 | * find_busiest_group - Returns the busiest group within the sched_domain | |
2939 | * if there is an imbalance. If there isn't an imbalance, and | |
2940 | * the user has opted for power-savings, it returns a group whose | |
2941 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
2942 | * such a group exists. | |
2943 | * | |
2944 | * Also calculates the amount of weighted load which should be moved | |
2945 | * to restore balance. | |
2946 | * | |
2947 | * @sd: The sched_domain whose busiest group is to be returned. | |
2948 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
2949 | * @imbalance: Variable which stores amount of weighted load which should | |
2950 | * be moved to restore balance/put a group to idle. | |
2951 | * @idle: The idle status of this_cpu. | |
2952 | * @sd_idle: The idleness of sd | |
2953 | * @cpus: The set of CPUs under consideration for load-balancing. | |
2954 | * @balance: Pointer to a variable indicating if this_cpu | |
2955 | * is the appropriate cpu to perform load balancing at this_level. | |
2956 | * | |
2957 | * Returns: - the busiest group if imbalance exists. | |
2958 | * - If no imbalance and user has opted for power-savings balance, | |
2959 | * return the least loaded group whose CPUs can be | |
2960 | * put to idle by rebalancing its tasks onto our group. | |
2961 | */ | |
2962 | static struct sched_group * | |
2963 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
2964 | unsigned long *imbalance, enum cpu_idle_type idle, | |
2965 | int *sd_idle, const struct cpumask *cpus, int *balance) | |
2966 | { | |
2967 | struct sd_lb_stats sds; | |
2968 | ||
2969 | memset(&sds, 0, sizeof(sds)); | |
2970 | ||
2971 | /* | |
2972 | * Compute the various statistics relavent for load balancing at | |
2973 | * this level. | |
2974 | */ | |
2975 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, | |
2976 | balance, &sds); | |
2977 | ||
2978 | /* Cases where imbalance does not exist from POV of this_cpu */ | |
2979 | /* 1) this_cpu is not the appropriate cpu to perform load balancing | |
2980 | * at this level. | |
2981 | * 2) There is no busy sibling group to pull from. | |
2982 | * 3) This group is the busiest group. | |
2983 | * 4) This group is more busy than the avg busieness at this | |
2984 | * sched_domain. | |
2985 | * 5) The imbalance is within the specified limit. | |
fab47622 NR |
2986 | * |
2987 | * Note: when doing newidle balance, if the local group has excess | |
2988 | * capacity (i.e. nr_running < group_capacity) and the busiest group | |
2989 | * does not have any capacity, we force a load balance to pull tasks | |
2990 | * to the local group. In this case, we skip past checks 3, 4 and 5. | |
1e3c88bd | 2991 | */ |
8f190fb3 | 2992 | if (!(*balance)) |
1e3c88bd PZ |
2993 | goto ret; |
2994 | ||
532cb4c4 MN |
2995 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
2996 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
2997 | return sds.busiest; | |
2998 | ||
1e3c88bd PZ |
2999 | if (!sds.busiest || sds.busiest_nr_running == 0) |
3000 | goto out_balanced; | |
3001 | ||
fab47622 NR |
3002 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
3003 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && | |
3004 | !sds.busiest_has_capacity) | |
3005 | goto force_balance; | |
3006 | ||
1e3c88bd PZ |
3007 | if (sds.this_load >= sds.max_load) |
3008 | goto out_balanced; | |
3009 | ||
3010 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; | |
3011 | ||
3012 | if (sds.this_load >= sds.avg_load) | |
3013 | goto out_balanced; | |
3014 | ||
3015 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
3016 | goto out_balanced; | |
3017 | ||
fab47622 | 3018 | force_balance: |
1e3c88bd PZ |
3019 | /* Looks like there is an imbalance. Compute it */ |
3020 | calculate_imbalance(&sds, this_cpu, imbalance); | |
3021 | return sds.busiest; | |
3022 | ||
3023 | out_balanced: | |
3024 | /* | |
3025 | * There is no obvious imbalance. But check if we can do some balancing | |
3026 | * to save power. | |
3027 | */ | |
3028 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
3029 | return sds.busiest; | |
3030 | ret: | |
3031 | *imbalance = 0; | |
3032 | return NULL; | |
3033 | } | |
3034 | ||
3035 | /* | |
3036 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3037 | */ | |
3038 | static struct rq * | |
9d5efe05 SV |
3039 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
3040 | enum cpu_idle_type idle, unsigned long imbalance, | |
3041 | const struct cpumask *cpus) | |
1e3c88bd PZ |
3042 | { |
3043 | struct rq *busiest = NULL, *rq; | |
3044 | unsigned long max_load = 0; | |
3045 | int i; | |
3046 | ||
3047 | for_each_cpu(i, sched_group_cpus(group)) { | |
3048 | unsigned long power = power_of(i); | |
3049 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
3050 | unsigned long wl; | |
3051 | ||
9d5efe05 SV |
3052 | if (!capacity) |
3053 | capacity = fix_small_capacity(sd, group); | |
3054 | ||
1e3c88bd PZ |
3055 | if (!cpumask_test_cpu(i, cpus)) |
3056 | continue; | |
3057 | ||
3058 | rq = cpu_rq(i); | |
6e40f5bb | 3059 | wl = weighted_cpuload(i); |
1e3c88bd | 3060 | |
6e40f5bb TG |
3061 | /* |
3062 | * When comparing with imbalance, use weighted_cpuload() | |
3063 | * which is not scaled with the cpu power. | |
3064 | */ | |
1e3c88bd PZ |
3065 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
3066 | continue; | |
3067 | ||
6e40f5bb TG |
3068 | /* |
3069 | * For the load comparisons with the other cpu's, consider | |
3070 | * the weighted_cpuload() scaled with the cpu power, so that | |
3071 | * the load can be moved away from the cpu that is potentially | |
3072 | * running at a lower capacity. | |
3073 | */ | |
3074 | wl = (wl * SCHED_LOAD_SCALE) / power; | |
3075 | ||
1e3c88bd PZ |
3076 | if (wl > max_load) { |
3077 | max_load = wl; | |
3078 | busiest = rq; | |
3079 | } | |
3080 | } | |
3081 | ||
3082 | return busiest; | |
3083 | } | |
3084 | ||
3085 | /* | |
3086 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3087 | * so long as it is large enough. | |
3088 | */ | |
3089 | #define MAX_PINNED_INTERVAL 512 | |
3090 | ||
3091 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
3092 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
3093 | ||
532cb4c4 MN |
3094 | static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle, |
3095 | int busiest_cpu, int this_cpu) | |
1af3ed3d PZ |
3096 | { |
3097 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
3098 | |
3099 | /* | |
3100 | * ASYM_PACKING needs to force migrate tasks from busy but | |
3101 | * higher numbered CPUs in order to pack all tasks in the | |
3102 | * lowest numbered CPUs. | |
3103 | */ | |
3104 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
3105 | return 1; | |
3106 | ||
1af3ed3d PZ |
3107 | /* |
3108 | * The only task running in a non-idle cpu can be moved to this | |
3109 | * cpu in an attempt to completely freeup the other CPU | |
3110 | * package. | |
3111 | * | |
3112 | * The package power saving logic comes from | |
3113 | * find_busiest_group(). If there are no imbalance, then | |
3114 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3115 | * f_b_g() will select a group from which a running task may be | |
3116 | * pulled to this cpu in order to make the other package idle. | |
3117 | * If there is no opportunity to make a package idle and if | |
3118 | * there are no imbalance, then f_b_g() will return NULL and no | |
3119 | * action will be taken in load_balance_newidle(). | |
3120 | * | |
3121 | * Under normal task pull operation due to imbalance, there | |
3122 | * will be more than one task in the source run queue and | |
3123 | * move_tasks() will succeed. ld_moved will be true and this | |
3124 | * active balance code will not be triggered. | |
3125 | */ | |
3126 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3127 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3128 | return 0; | |
3129 | ||
3130 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) | |
3131 | return 0; | |
3132 | } | |
3133 | ||
3134 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3135 | } | |
3136 | ||
969c7921 TH |
3137 | static int active_load_balance_cpu_stop(void *data); |
3138 | ||
1e3c88bd PZ |
3139 | /* |
3140 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3141 | * tasks if there is an imbalance. | |
3142 | */ | |
3143 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3144 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3145 | int *balance) | |
3146 | { | |
3147 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
3148 | struct sched_group *group; | |
3149 | unsigned long imbalance; | |
3150 | struct rq *busiest; | |
3151 | unsigned long flags; | |
3152 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3153 | ||
3154 | cpumask_copy(cpus, cpu_active_mask); | |
3155 | ||
3156 | /* | |
3157 | * When power savings policy is enabled for the parent domain, idle | |
3158 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3159 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
3160 | * portraying it as CPU_NOT_IDLE. | |
3161 | */ | |
3162 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
3163 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3164 | sd_idle = 1; | |
3165 | ||
3166 | schedstat_inc(sd, lb_count[idle]); | |
3167 | ||
3168 | redo: | |
1e3c88bd PZ |
3169 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, |
3170 | cpus, balance); | |
3171 | ||
3172 | if (*balance == 0) | |
3173 | goto out_balanced; | |
3174 | ||
3175 | if (!group) { | |
3176 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3177 | goto out_balanced; | |
3178 | } | |
3179 | ||
9d5efe05 | 3180 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3181 | if (!busiest) { |
3182 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3183 | goto out_balanced; | |
3184 | } | |
3185 | ||
3186 | BUG_ON(busiest == this_rq); | |
3187 | ||
3188 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3189 | ||
3190 | ld_moved = 0; | |
3191 | if (busiest->nr_running > 1) { | |
3192 | /* | |
3193 | * Attempt to move tasks. If find_busiest_group has found | |
3194 | * an imbalance but busiest->nr_running <= 1, the group is | |
3195 | * still unbalanced. ld_moved simply stays zero, so it is | |
3196 | * correctly treated as an imbalance. | |
3197 | */ | |
3198 | local_irq_save(flags); | |
3199 | double_rq_lock(this_rq, busiest); | |
3200 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3201 | imbalance, sd, idle, &all_pinned); | |
3202 | double_rq_unlock(this_rq, busiest); | |
3203 | local_irq_restore(flags); | |
3204 | ||
3205 | /* | |
3206 | * some other cpu did the load balance for us. | |
3207 | */ | |
3208 | if (ld_moved && this_cpu != smp_processor_id()) | |
3209 | resched_cpu(this_cpu); | |
3210 | ||
3211 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3212 | if (unlikely(all_pinned)) { | |
3213 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3214 | if (!cpumask_empty(cpus)) | |
3215 | goto redo; | |
3216 | goto out_balanced; | |
3217 | } | |
3218 | } | |
3219 | ||
3220 | if (!ld_moved) { | |
3221 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3222 | /* |
3223 | * Increment the failure counter only on periodic balance. | |
3224 | * We do not want newidle balance, which can be very | |
3225 | * frequent, pollute the failure counter causing | |
3226 | * excessive cache_hot migrations and active balances. | |
3227 | */ | |
3228 | if (idle != CPU_NEWLY_IDLE) | |
3229 | sd->nr_balance_failed++; | |
1e3c88bd | 3230 | |
532cb4c4 MN |
3231 | if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest), |
3232 | this_cpu)) { | |
1e3c88bd PZ |
3233 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3234 | ||
969c7921 TH |
3235 | /* don't kick the active_load_balance_cpu_stop, |
3236 | * if the curr task on busiest cpu can't be | |
3237 | * moved to this_cpu | |
1e3c88bd PZ |
3238 | */ |
3239 | if (!cpumask_test_cpu(this_cpu, | |
3240 | &busiest->curr->cpus_allowed)) { | |
3241 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3242 | flags); | |
3243 | all_pinned = 1; | |
3244 | goto out_one_pinned; | |
3245 | } | |
3246 | ||
969c7921 TH |
3247 | /* |
3248 | * ->active_balance synchronizes accesses to | |
3249 | * ->active_balance_work. Once set, it's cleared | |
3250 | * only after active load balance is finished. | |
3251 | */ | |
1e3c88bd PZ |
3252 | if (!busiest->active_balance) { |
3253 | busiest->active_balance = 1; | |
3254 | busiest->push_cpu = this_cpu; | |
3255 | active_balance = 1; | |
3256 | } | |
3257 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 3258 | |
1e3c88bd | 3259 | if (active_balance) |
969c7921 TH |
3260 | stop_one_cpu_nowait(cpu_of(busiest), |
3261 | active_load_balance_cpu_stop, busiest, | |
3262 | &busiest->active_balance_work); | |
1e3c88bd PZ |
3263 | |
3264 | /* | |
3265 | * We've kicked active balancing, reset the failure | |
3266 | * counter. | |
3267 | */ | |
3268 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3269 | } | |
3270 | } else | |
3271 | sd->nr_balance_failed = 0; | |
3272 | ||
3273 | if (likely(!active_balance)) { | |
3274 | /* We were unbalanced, so reset the balancing interval */ | |
3275 | sd->balance_interval = sd->min_interval; | |
3276 | } else { | |
3277 | /* | |
3278 | * If we've begun active balancing, start to back off. This | |
3279 | * case may not be covered by the all_pinned logic if there | |
3280 | * is only 1 task on the busy runqueue (because we don't call | |
3281 | * move_tasks). | |
3282 | */ | |
3283 | if (sd->balance_interval < sd->max_interval) | |
3284 | sd->balance_interval *= 2; | |
3285 | } | |
3286 | ||
3287 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3288 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3289 | ld_moved = -1; | |
3290 | ||
3291 | goto out; | |
3292 | ||
3293 | out_balanced: | |
3294 | schedstat_inc(sd, lb_balanced[idle]); | |
3295 | ||
3296 | sd->nr_balance_failed = 0; | |
3297 | ||
3298 | out_one_pinned: | |
3299 | /* tune up the balancing interval */ | |
3300 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3301 | (sd->balance_interval < sd->max_interval)) | |
3302 | sd->balance_interval *= 2; | |
3303 | ||
3304 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3305 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3306 | ld_moved = -1; | |
3307 | else | |
3308 | ld_moved = 0; | |
3309 | out: | |
1e3c88bd PZ |
3310 | return ld_moved; |
3311 | } | |
3312 | ||
1e3c88bd PZ |
3313 | /* |
3314 | * idle_balance is called by schedule() if this_cpu is about to become | |
3315 | * idle. Attempts to pull tasks from other CPUs. | |
3316 | */ | |
3317 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3318 | { | |
3319 | struct sched_domain *sd; | |
3320 | int pulled_task = 0; | |
3321 | unsigned long next_balance = jiffies + HZ; | |
3322 | ||
3323 | this_rq->idle_stamp = this_rq->clock; | |
3324 | ||
3325 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3326 | return; | |
3327 | ||
f492e12e PZ |
3328 | /* |
3329 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
3330 | */ | |
3331 | raw_spin_unlock(&this_rq->lock); | |
3332 | ||
1e3c88bd PZ |
3333 | for_each_domain(this_cpu, sd) { |
3334 | unsigned long interval; | |
f492e12e | 3335 | int balance = 1; |
1e3c88bd PZ |
3336 | |
3337 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3338 | continue; | |
3339 | ||
f492e12e | 3340 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 3341 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
3342 | pulled_task = load_balance(this_cpu, this_rq, |
3343 | sd, CPU_NEWLY_IDLE, &balance); | |
3344 | } | |
1e3c88bd PZ |
3345 | |
3346 | interval = msecs_to_jiffies(sd->balance_interval); | |
3347 | if (time_after(next_balance, sd->last_balance + interval)) | |
3348 | next_balance = sd->last_balance + interval; | |
fab47622 | 3349 | if (pulled_task) |
1e3c88bd | 3350 | break; |
1e3c88bd | 3351 | } |
f492e12e PZ |
3352 | |
3353 | raw_spin_lock(&this_rq->lock); | |
3354 | ||
1e3c88bd PZ |
3355 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
3356 | /* | |
3357 | * We are going idle. next_balance may be set based on | |
3358 | * a busy processor. So reset next_balance. | |
3359 | */ | |
3360 | this_rq->next_balance = next_balance; | |
3361 | } | |
3362 | } | |
3363 | ||
3364 | /* | |
969c7921 TH |
3365 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
3366 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
3367 | * least 1 task to be running on each physical CPU where possible, and | |
3368 | * avoids physical / logical imbalances. | |
1e3c88bd | 3369 | */ |
969c7921 | 3370 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 3371 | { |
969c7921 TH |
3372 | struct rq *busiest_rq = data; |
3373 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 3374 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 3375 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 3376 | struct sched_domain *sd; |
969c7921 TH |
3377 | |
3378 | raw_spin_lock_irq(&busiest_rq->lock); | |
3379 | ||
3380 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
3381 | if (unlikely(busiest_cpu != smp_processor_id() || | |
3382 | !busiest_rq->active_balance)) | |
3383 | goto out_unlock; | |
1e3c88bd PZ |
3384 | |
3385 | /* Is there any task to move? */ | |
3386 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 3387 | goto out_unlock; |
1e3c88bd PZ |
3388 | |
3389 | /* | |
3390 | * This condition is "impossible", if it occurs | |
3391 | * we need to fix it. Originally reported by | |
3392 | * Bjorn Helgaas on a 128-cpu setup. | |
3393 | */ | |
3394 | BUG_ON(busiest_rq == target_rq); | |
3395 | ||
3396 | /* move a task from busiest_rq to target_rq */ | |
3397 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
3398 | |
3399 | /* Search for an sd spanning us and the target CPU. */ | |
3400 | for_each_domain(target_cpu, sd) { | |
3401 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3402 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3403 | break; | |
3404 | } | |
3405 | ||
3406 | if (likely(sd)) { | |
3407 | schedstat_inc(sd, alb_count); | |
3408 | ||
3409 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3410 | sd, CPU_IDLE)) | |
3411 | schedstat_inc(sd, alb_pushed); | |
3412 | else | |
3413 | schedstat_inc(sd, alb_failed); | |
3414 | } | |
3415 | double_unlock_balance(busiest_rq, target_rq); | |
969c7921 TH |
3416 | out_unlock: |
3417 | busiest_rq->active_balance = 0; | |
3418 | raw_spin_unlock_irq(&busiest_rq->lock); | |
3419 | return 0; | |
1e3c88bd PZ |
3420 | } |
3421 | ||
3422 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
3423 | |
3424 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
3425 | ||
3426 | static void trigger_sched_softirq(void *data) | |
3427 | { | |
3428 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
3429 | } | |
3430 | ||
3431 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
3432 | { | |
3433 | csd->func = trigger_sched_softirq; | |
3434 | csd->info = NULL; | |
3435 | csd->flags = 0; | |
3436 | csd->priv = 0; | |
3437 | } | |
3438 | ||
3439 | /* | |
3440 | * idle load balancing details | |
3441 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
3442 | * entering idle. | |
3443 | * - This idle load balancer CPU will also go into tickless mode when | |
3444 | * it is idle, just like all other idle CPUs | |
3445 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
3446 | * needed, they will kick the idle load balancer, which then does idle | |
3447 | * load balancing for all the idle CPUs. | |
3448 | */ | |
1e3c88bd PZ |
3449 | static struct { |
3450 | atomic_t load_balancer; | |
83cd4fe2 VP |
3451 | atomic_t first_pick_cpu; |
3452 | atomic_t second_pick_cpu; | |
3453 | cpumask_var_t idle_cpus_mask; | |
3454 | cpumask_var_t grp_idle_mask; | |
3455 | unsigned long next_balance; /* in jiffy units */ | |
3456 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
3457 | |
3458 | int get_nohz_load_balancer(void) | |
3459 | { | |
3460 | return atomic_read(&nohz.load_balancer); | |
3461 | } | |
3462 | ||
3463 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3464 | /** | |
3465 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3466 | * @cpu: The cpu whose lowest level of sched domain is to | |
3467 | * be returned. | |
3468 | * @flag: The flag to check for the lowest sched_domain | |
3469 | * for the given cpu. | |
3470 | * | |
3471 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3472 | */ | |
3473 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3474 | { | |
3475 | struct sched_domain *sd; | |
3476 | ||
3477 | for_each_domain(cpu, sd) | |
3478 | if (sd && (sd->flags & flag)) | |
3479 | break; | |
3480 | ||
3481 | return sd; | |
3482 | } | |
3483 | ||
3484 | /** | |
3485 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3486 | * @cpu: The cpu whose domains we're iterating over. | |
3487 | * @sd: variable holding the value of the power_savings_sd | |
3488 | * for cpu. | |
3489 | * @flag: The flag to filter the sched_domains to be iterated. | |
3490 | * | |
3491 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3492 | * set, starting from the lowest sched_domain to the highest. | |
3493 | */ | |
3494 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3495 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3496 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3497 | ||
3498 | /** | |
3499 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3500 | * @ilb_group: group to be checked for semi-idleness | |
3501 | * | |
3502 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3503 | * | |
3504 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3505 | * and atleast one non-idle CPU. This helper function checks if the given | |
3506 | * sched_group is semi-idle or not. | |
3507 | */ | |
3508 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3509 | { | |
83cd4fe2 | 3510 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
3511 | sched_group_cpus(ilb_group)); |
3512 | ||
3513 | /* | |
3514 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3515 | * and atleast one idle cpu. | |
3516 | */ | |
83cd4fe2 | 3517 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
3518 | return 0; |
3519 | ||
83cd4fe2 | 3520 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
3521 | return 0; |
3522 | ||
3523 | return 1; | |
3524 | } | |
3525 | /** | |
3526 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3527 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3528 | * | |
3529 | * Returns: Returns the id of the idle load balancer if it exists, | |
3530 | * Else, returns >= nr_cpu_ids. | |
3531 | * | |
3532 | * This algorithm picks the idle load balancer such that it belongs to a | |
3533 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3534 | * completely idle packages/cores just for the purpose of idle load balancing | |
3535 | * when there are other idle cpu's which are better suited for that job. | |
3536 | */ | |
3537 | static int find_new_ilb(int cpu) | |
3538 | { | |
3539 | struct sched_domain *sd; | |
3540 | struct sched_group *ilb_group; | |
3541 | ||
3542 | /* | |
3543 | * Have idle load balancer selection from semi-idle packages only | |
3544 | * when power-aware load balancing is enabled | |
3545 | */ | |
3546 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3547 | goto out_done; | |
3548 | ||
3549 | /* | |
3550 | * Optimize for the case when we have no idle CPUs or only one | |
3551 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3552 | */ | |
83cd4fe2 | 3553 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
3554 | goto out_done; |
3555 | ||
3556 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
3557 | ilb_group = sd->groups; | |
3558 | ||
3559 | do { | |
3560 | if (is_semi_idle_group(ilb_group)) | |
83cd4fe2 | 3561 | return cpumask_first(nohz.grp_idle_mask); |
1e3c88bd PZ |
3562 | |
3563 | ilb_group = ilb_group->next; | |
3564 | ||
3565 | } while (ilb_group != sd->groups); | |
3566 | } | |
3567 | ||
3568 | out_done: | |
83cd4fe2 | 3569 | return nr_cpu_ids; |
1e3c88bd PZ |
3570 | } |
3571 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3572 | static inline int find_new_ilb(int call_cpu) | |
3573 | { | |
83cd4fe2 | 3574 | return nr_cpu_ids; |
1e3c88bd PZ |
3575 | } |
3576 | #endif | |
3577 | ||
83cd4fe2 VP |
3578 | /* |
3579 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
3580 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
3581 | * CPU (if there is one). | |
3582 | */ | |
3583 | static void nohz_balancer_kick(int cpu) | |
3584 | { | |
3585 | int ilb_cpu; | |
3586 | ||
3587 | nohz.next_balance++; | |
3588 | ||
3589 | ilb_cpu = get_nohz_load_balancer(); | |
3590 | ||
3591 | if (ilb_cpu >= nr_cpu_ids) { | |
3592 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
3593 | if (ilb_cpu >= nr_cpu_ids) | |
3594 | return; | |
3595 | } | |
3596 | ||
3597 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
3598 | struct call_single_data *cp; | |
3599 | ||
3600 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
3601 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
3602 | __smp_call_function_single(ilb_cpu, cp, 0); | |
3603 | } | |
3604 | return; | |
3605 | } | |
3606 | ||
1e3c88bd PZ |
3607 | /* |
3608 | * This routine will try to nominate the ilb (idle load balancing) | |
3609 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 3610 | * load balancing on behalf of all those cpus. |
1e3c88bd | 3611 | * |
83cd4fe2 VP |
3612 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
3613 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
3614 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 3615 | * |
83cd4fe2 VP |
3616 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
3617 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
3618 | * behalf of all idle CPUs). | |
1e3c88bd | 3619 | */ |
83cd4fe2 | 3620 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
3621 | { |
3622 | int cpu = smp_processor_id(); | |
3623 | ||
3624 | if (stop_tick) { | |
1e3c88bd PZ |
3625 | if (!cpu_active(cpu)) { |
3626 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 3627 | return; |
1e3c88bd PZ |
3628 | |
3629 | /* | |
3630 | * If we are going offline and still the leader, | |
3631 | * give up! | |
3632 | */ | |
83cd4fe2 VP |
3633 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3634 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3635 | BUG(); |
3636 | ||
83cd4fe2 | 3637 | return; |
1e3c88bd PZ |
3638 | } |
3639 | ||
83cd4fe2 | 3640 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 3641 | |
83cd4fe2 VP |
3642 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
3643 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
3644 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
3645 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 3646 | |
83cd4fe2 | 3647 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
3648 | int new_ilb; |
3649 | ||
83cd4fe2 VP |
3650 | /* make me the ilb owner */ |
3651 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
3652 | cpu) != nr_cpu_ids) | |
3653 | return; | |
3654 | ||
1e3c88bd PZ |
3655 | /* |
3656 | * Check to see if there is a more power-efficient | |
3657 | * ilb. | |
3658 | */ | |
3659 | new_ilb = find_new_ilb(cpu); | |
3660 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 3661 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 3662 | resched_cpu(new_ilb); |
83cd4fe2 | 3663 | return; |
1e3c88bd | 3664 | } |
83cd4fe2 | 3665 | return; |
1e3c88bd PZ |
3666 | } |
3667 | } else { | |
83cd4fe2 VP |
3668 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
3669 | return; | |
1e3c88bd | 3670 | |
83cd4fe2 | 3671 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
3672 | |
3673 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
3674 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3675 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3676 | BUG(); |
3677 | } | |
83cd4fe2 | 3678 | return; |
1e3c88bd PZ |
3679 | } |
3680 | #endif | |
3681 | ||
3682 | static DEFINE_SPINLOCK(balancing); | |
3683 | ||
3684 | /* | |
3685 | * It checks each scheduling domain to see if it is due to be balanced, | |
3686 | * and initiates a balancing operation if so. | |
3687 | * | |
3688 | * Balancing parameters are set up in arch_init_sched_domains. | |
3689 | */ | |
3690 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3691 | { | |
3692 | int balance = 1; | |
3693 | struct rq *rq = cpu_rq(cpu); | |
3694 | unsigned long interval; | |
3695 | struct sched_domain *sd; | |
3696 | /* Earliest time when we have to do rebalance again */ | |
3697 | unsigned long next_balance = jiffies + 60*HZ; | |
3698 | int update_next_balance = 0; | |
3699 | int need_serialize; | |
3700 | ||
2069dd75 PZ |
3701 | update_shares(cpu); |
3702 | ||
1e3c88bd PZ |
3703 | for_each_domain(cpu, sd) { |
3704 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3705 | continue; | |
3706 | ||
3707 | interval = sd->balance_interval; | |
3708 | if (idle != CPU_IDLE) | |
3709 | interval *= sd->busy_factor; | |
3710 | ||
3711 | /* scale ms to jiffies */ | |
3712 | interval = msecs_to_jiffies(interval); | |
3713 | if (unlikely(!interval)) | |
3714 | interval = 1; | |
3715 | if (interval > HZ*NR_CPUS/10) | |
3716 | interval = HZ*NR_CPUS/10; | |
3717 | ||
3718 | need_serialize = sd->flags & SD_SERIALIZE; | |
3719 | ||
3720 | if (need_serialize) { | |
3721 | if (!spin_trylock(&balancing)) | |
3722 | goto out; | |
3723 | } | |
3724 | ||
3725 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3726 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3727 | /* | |
3728 | * We've pulled tasks over so either we're no | |
3729 | * longer idle, or one of our SMT siblings is | |
3730 | * not idle. | |
3731 | */ | |
3732 | idle = CPU_NOT_IDLE; | |
3733 | } | |
3734 | sd->last_balance = jiffies; | |
3735 | } | |
3736 | if (need_serialize) | |
3737 | spin_unlock(&balancing); | |
3738 | out: | |
3739 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3740 | next_balance = sd->last_balance + interval; | |
3741 | update_next_balance = 1; | |
3742 | } | |
3743 | ||
3744 | /* | |
3745 | * Stop the load balance at this level. There is another | |
3746 | * CPU in our sched group which is doing load balancing more | |
3747 | * actively. | |
3748 | */ | |
3749 | if (!balance) | |
3750 | break; | |
3751 | } | |
3752 | ||
3753 | /* | |
3754 | * next_balance will be updated only when there is a need. | |
3755 | * When the cpu is attached to null domain for ex, it will not be | |
3756 | * updated. | |
3757 | */ | |
3758 | if (likely(update_next_balance)) | |
3759 | rq->next_balance = next_balance; | |
3760 | } | |
3761 | ||
83cd4fe2 | 3762 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 3763 | /* |
83cd4fe2 | 3764 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
3765 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
3766 | */ | |
83cd4fe2 VP |
3767 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
3768 | { | |
3769 | struct rq *this_rq = cpu_rq(this_cpu); | |
3770 | struct rq *rq; | |
3771 | int balance_cpu; | |
3772 | ||
3773 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
3774 | return; | |
3775 | ||
3776 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
3777 | if (balance_cpu == this_cpu) | |
3778 | continue; | |
3779 | ||
3780 | /* | |
3781 | * If this cpu gets work to do, stop the load balancing | |
3782 | * work being done for other cpus. Next load | |
3783 | * balancing owner will pick it up. | |
3784 | */ | |
3785 | if (need_resched()) { | |
3786 | this_rq->nohz_balance_kick = 0; | |
3787 | break; | |
3788 | } | |
3789 | ||
3790 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 3791 | update_rq_clock(this_rq); |
83cd4fe2 VP |
3792 | update_cpu_load(this_rq); |
3793 | raw_spin_unlock_irq(&this_rq->lock); | |
3794 | ||
3795 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3796 | ||
3797 | rq = cpu_rq(balance_cpu); | |
3798 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3799 | this_rq->next_balance = rq->next_balance; | |
3800 | } | |
3801 | nohz.next_balance = this_rq->next_balance; | |
3802 | this_rq->nohz_balance_kick = 0; | |
3803 | } | |
3804 | ||
3805 | /* | |
3806 | * Current heuristic for kicking the idle load balancer | |
3807 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
3808 | * idle load balancer when it has more than one process active. This | |
3809 | * eliminates the need for idle load balancing altogether when we have | |
3810 | * only one running process in the system (common case). | |
3811 | * - If there are more than one busy CPU, idle load balancer may have | |
3812 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
3813 | * SMT or core siblings and can run better if they move to different | |
3814 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
3815 | * which will kick idle load balancer as soon as it has any load. | |
3816 | */ | |
3817 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
3818 | { | |
3819 | unsigned long now = jiffies; | |
3820 | int ret; | |
3821 | int first_pick_cpu, second_pick_cpu; | |
3822 | ||
3823 | if (time_before(now, nohz.next_balance)) | |
3824 | return 0; | |
3825 | ||
f6c3f168 | 3826 | if (rq->idle_at_tick) |
83cd4fe2 VP |
3827 | return 0; |
3828 | ||
3829 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
3830 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
3831 | ||
3832 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
3833 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
3834 | return 0; | |
3835 | ||
3836 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
3837 | if (ret == nr_cpu_ids || ret == cpu) { | |
3838 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
3839 | if (rq->nr_running > 1) | |
3840 | return 1; | |
3841 | } else { | |
3842 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
3843 | if (ret == nr_cpu_ids || ret == cpu) { | |
3844 | if (rq->nr_running) | |
3845 | return 1; | |
3846 | } | |
3847 | } | |
3848 | return 0; | |
3849 | } | |
3850 | #else | |
3851 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
3852 | #endif | |
3853 | ||
3854 | /* | |
3855 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
3856 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
3857 | */ | |
1e3c88bd PZ |
3858 | static void run_rebalance_domains(struct softirq_action *h) |
3859 | { | |
3860 | int this_cpu = smp_processor_id(); | |
3861 | struct rq *this_rq = cpu_rq(this_cpu); | |
3862 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
3863 | CPU_IDLE : CPU_NOT_IDLE; | |
3864 | ||
3865 | rebalance_domains(this_cpu, idle); | |
3866 | ||
1e3c88bd | 3867 | /* |
83cd4fe2 | 3868 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
3869 | * balancing on behalf of the other idle cpus whose ticks are |
3870 | * stopped. | |
3871 | */ | |
83cd4fe2 | 3872 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
3873 | } |
3874 | ||
3875 | static inline int on_null_domain(int cpu) | |
3876 | { | |
90a6501f | 3877 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
3878 | } |
3879 | ||
3880 | /* | |
3881 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
3882 | */ |
3883 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
3884 | { | |
1e3c88bd PZ |
3885 | /* Don't need to rebalance while attached to NULL domain */ |
3886 | if (time_after_eq(jiffies, rq->next_balance) && | |
3887 | likely(!on_null_domain(cpu))) | |
3888 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
3889 | #ifdef CONFIG_NO_HZ |
3890 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
3891 | nohz_balancer_kick(cpu); | |
3892 | #endif | |
1e3c88bd PZ |
3893 | } |
3894 | ||
0bcdcf28 CE |
3895 | static void rq_online_fair(struct rq *rq) |
3896 | { | |
3897 | update_sysctl(); | |
3898 | } | |
3899 | ||
3900 | static void rq_offline_fair(struct rq *rq) | |
3901 | { | |
3902 | update_sysctl(); | |
3903 | } | |
3904 | ||
1e3c88bd PZ |
3905 | #else /* CONFIG_SMP */ |
3906 | ||
3907 | /* | |
3908 | * on UP we do not need to balance between CPUs: | |
3909 | */ | |
3910 | static inline void idle_balance(int cpu, struct rq *rq) | |
3911 | { | |
3912 | } | |
3913 | ||
55e12e5e | 3914 | #endif /* CONFIG_SMP */ |
e1d1484f | 3915 | |
bf0f6f24 IM |
3916 | /* |
3917 | * scheduler tick hitting a task of our scheduling class: | |
3918 | */ | |
8f4d37ec | 3919 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
3920 | { |
3921 | struct cfs_rq *cfs_rq; | |
3922 | struct sched_entity *se = &curr->se; | |
3923 | ||
3924 | for_each_sched_entity(se) { | |
3925 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 3926 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
3927 | } |
3928 | } | |
3929 | ||
3930 | /* | |
cd29fe6f PZ |
3931 | * called on fork with the child task as argument from the parent's context |
3932 | * - child not yet on the tasklist | |
3933 | * - preemption disabled | |
bf0f6f24 | 3934 | */ |
cd29fe6f | 3935 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 3936 | { |
cd29fe6f | 3937 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 3938 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 3939 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
3940 | struct rq *rq = this_rq(); |
3941 | unsigned long flags; | |
3942 | ||
05fa785c | 3943 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 3944 | |
861d034e PZ |
3945 | update_rq_clock(rq); |
3946 | ||
b0a0f667 PM |
3947 | if (unlikely(task_cpu(p) != this_cpu)) { |
3948 | rcu_read_lock(); | |
cd29fe6f | 3949 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
3950 | rcu_read_unlock(); |
3951 | } | |
bf0f6f24 | 3952 | |
7109c442 | 3953 | update_curr(cfs_rq); |
cd29fe6f | 3954 | |
b5d9d734 MG |
3955 | if (curr) |
3956 | se->vruntime = curr->vruntime; | |
aeb73b04 | 3957 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 3958 | |
cd29fe6f | 3959 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 3960 | /* |
edcb60a3 IM |
3961 | * Upon rescheduling, sched_class::put_prev_task() will place |
3962 | * 'current' within the tree based on its new key value. | |
3963 | */ | |
4d78e7b6 | 3964 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 3965 | resched_task(rq->curr); |
4d78e7b6 | 3966 | } |
bf0f6f24 | 3967 | |
88ec22d3 PZ |
3968 | se->vruntime -= cfs_rq->min_vruntime; |
3969 | ||
05fa785c | 3970 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
3971 | } |
3972 | ||
cb469845 SR |
3973 | /* |
3974 | * Priority of the task has changed. Check to see if we preempt | |
3975 | * the current task. | |
3976 | */ | |
3977 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
3978 | int oldprio, int running) | |
3979 | { | |
3980 | /* | |
3981 | * Reschedule if we are currently running on this runqueue and | |
3982 | * our priority decreased, or if we are not currently running on | |
3983 | * this runqueue and our priority is higher than the current's | |
3984 | */ | |
3985 | if (running) { | |
3986 | if (p->prio > oldprio) | |
3987 | resched_task(rq->curr); | |
3988 | } else | |
15afe09b | 3989 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
3990 | } |
3991 | ||
3992 | /* | |
3993 | * We switched to the sched_fair class. | |
3994 | */ | |
3995 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
3996 | int running) | |
3997 | { | |
3998 | /* | |
3999 | * We were most likely switched from sched_rt, so | |
4000 | * kick off the schedule if running, otherwise just see | |
4001 | * if we can still preempt the current task. | |
4002 | */ | |
4003 | if (running) | |
4004 | resched_task(rq->curr); | |
4005 | else | |
15afe09b | 4006 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4007 | } |
4008 | ||
83b699ed SV |
4009 | /* Account for a task changing its policy or group. |
4010 | * | |
4011 | * This routine is mostly called to set cfs_rq->curr field when a task | |
4012 | * migrates between groups/classes. | |
4013 | */ | |
4014 | static void set_curr_task_fair(struct rq *rq) | |
4015 | { | |
4016 | struct sched_entity *se = &rq->curr->se; | |
4017 | ||
4018 | for_each_sched_entity(se) | |
4019 | set_next_entity(cfs_rq_of(se), se); | |
4020 | } | |
4021 | ||
810b3817 | 4022 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4023 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 4024 | { |
b2b5ce02 PZ |
4025 | /* |
4026 | * If the task was not on the rq at the time of this cgroup movement | |
4027 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
4028 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
4029 | * bonus in place_entity()). | |
4030 | * | |
4031 | * If it was on the rq, we've just 'preempted' it, which does convert | |
4032 | * ->vruntime to a relative base. | |
4033 | * | |
4034 | * Make sure both cases convert their relative position when migrating | |
4035 | * to another cgroup's rq. This does somewhat interfere with the | |
4036 | * fair sleeper stuff for the first placement, but who cares. | |
4037 | */ | |
4038 | if (!on_rq) | |
4039 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
4040 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 4041 | if (!on_rq) |
b2b5ce02 | 4042 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
4043 | } |
4044 | #endif | |
4045 | ||
6d686f45 | 4046 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
4047 | { |
4048 | struct sched_entity *se = &task->se; | |
0d721cea PW |
4049 | unsigned int rr_interval = 0; |
4050 | ||
4051 | /* | |
4052 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
4053 | * idle runqueue: | |
4054 | */ | |
0d721cea PW |
4055 | if (rq->cfs.load.weight) |
4056 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
4057 | |
4058 | return rr_interval; | |
4059 | } | |
4060 | ||
bf0f6f24 IM |
4061 | /* |
4062 | * All the scheduling class methods: | |
4063 | */ | |
5522d5d5 IM |
4064 | static const struct sched_class fair_sched_class = { |
4065 | .next = &idle_sched_class, | |
bf0f6f24 IM |
4066 | .enqueue_task = enqueue_task_fair, |
4067 | .dequeue_task = dequeue_task_fair, | |
4068 | .yield_task = yield_task_fair, | |
4069 | ||
2e09bf55 | 4070 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
4071 | |
4072 | .pick_next_task = pick_next_task_fair, | |
4073 | .put_prev_task = put_prev_task_fair, | |
4074 | ||
681f3e68 | 4075 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
4076 | .select_task_rq = select_task_rq_fair, |
4077 | ||
0bcdcf28 CE |
4078 | .rq_online = rq_online_fair, |
4079 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
4080 | |
4081 | .task_waking = task_waking_fair, | |
681f3e68 | 4082 | #endif |
bf0f6f24 | 4083 | |
83b699ed | 4084 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 4085 | .task_tick = task_tick_fair, |
cd29fe6f | 4086 | .task_fork = task_fork_fair, |
cb469845 SR |
4087 | |
4088 | .prio_changed = prio_changed_fair, | |
4089 | .switched_to = switched_to_fair, | |
810b3817 | 4090 | |
0d721cea PW |
4091 | .get_rr_interval = get_rr_interval_fair, |
4092 | ||
810b3817 | 4093 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4094 | .task_move_group = task_move_group_fair, |
810b3817 | 4095 | #endif |
bf0f6f24 IM |
4096 | }; |
4097 | ||
4098 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 4099 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 4100 | { |
bf0f6f24 IM |
4101 | struct cfs_rq *cfs_rq; |
4102 | ||
5973e5b9 | 4103 | rcu_read_lock(); |
c3b64f1e | 4104 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 4105 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 4106 | rcu_read_unlock(); |
bf0f6f24 IM |
4107 | } |
4108 | #endif |