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 | ||
146 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
147 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
148 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
149 | ||
150 | /* Do the two (enqueued) entities belong to the same group ? */ | |
151 | static inline int | |
152 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
153 | { | |
154 | if (se->cfs_rq == pse->cfs_rq) | |
155 | return 1; | |
156 | ||
157 | return 0; | |
158 | } | |
159 | ||
160 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
161 | { | |
162 | return se->parent; | |
163 | } | |
164 | ||
464b7527 PZ |
165 | /* return depth at which a sched entity is present in the hierarchy */ |
166 | static inline int depth_se(struct sched_entity *se) | |
167 | { | |
168 | int depth = 0; | |
169 | ||
170 | for_each_sched_entity(se) | |
171 | depth++; | |
172 | ||
173 | return depth; | |
174 | } | |
175 | ||
176 | static void | |
177 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
178 | { | |
179 | int se_depth, pse_depth; | |
180 | ||
181 | /* | |
182 | * preemption test can be made between sibling entities who are in the | |
183 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
184 | * both tasks until we find their ancestors who are siblings of common | |
185 | * parent. | |
186 | */ | |
187 | ||
188 | /* First walk up until both entities are at same depth */ | |
189 | se_depth = depth_se(*se); | |
190 | pse_depth = depth_se(*pse); | |
191 | ||
192 | while (se_depth > pse_depth) { | |
193 | se_depth--; | |
194 | *se = parent_entity(*se); | |
195 | } | |
196 | ||
197 | while (pse_depth > se_depth) { | |
198 | pse_depth--; | |
199 | *pse = parent_entity(*pse); | |
200 | } | |
201 | ||
202 | while (!is_same_group(*se, *pse)) { | |
203 | *se = parent_entity(*se); | |
204 | *pse = parent_entity(*pse); | |
205 | } | |
206 | } | |
207 | ||
8f48894f PZ |
208 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
209 | ||
210 | static inline struct task_struct *task_of(struct sched_entity *se) | |
211 | { | |
212 | return container_of(se, struct task_struct, se); | |
213 | } | |
bf0f6f24 | 214 | |
62160e3f IM |
215 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
216 | { | |
217 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
218 | } |
219 | ||
220 | #define entity_is_task(se) 1 | |
221 | ||
b758149c PZ |
222 | #define for_each_sched_entity(se) \ |
223 | for (; se; se = NULL) | |
bf0f6f24 | 224 | |
b758149c | 225 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 226 | { |
b758149c | 227 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
228 | } |
229 | ||
b758149c PZ |
230 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
231 | { | |
232 | struct task_struct *p = task_of(se); | |
233 | struct rq *rq = task_rq(p); | |
234 | ||
235 | return &rq->cfs; | |
236 | } | |
237 | ||
238 | /* runqueue "owned" by this group */ | |
239 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
240 | { | |
241 | return NULL; | |
242 | } | |
243 | ||
244 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
245 | { | |
246 | return &cpu_rq(this_cpu)->cfs; | |
247 | } | |
248 | ||
249 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
250 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
251 | ||
252 | static inline int | |
253 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
254 | { | |
255 | return 1; | |
256 | } | |
257 | ||
258 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
259 | { | |
260 | return NULL; | |
261 | } | |
262 | ||
464b7527 PZ |
263 | static inline void |
264 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
265 | { | |
266 | } | |
267 | ||
b758149c PZ |
268 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
269 | ||
bf0f6f24 IM |
270 | |
271 | /************************************************************** | |
272 | * Scheduling class tree data structure manipulation methods: | |
273 | */ | |
274 | ||
0702e3eb | 275 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 276 | { |
368059a9 PZ |
277 | s64 delta = (s64)(vruntime - min_vruntime); |
278 | if (delta > 0) | |
02e0431a PZ |
279 | min_vruntime = vruntime; |
280 | ||
281 | return min_vruntime; | |
282 | } | |
283 | ||
0702e3eb | 284 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
285 | { |
286 | s64 delta = (s64)(vruntime - min_vruntime); | |
287 | if (delta < 0) | |
288 | min_vruntime = vruntime; | |
289 | ||
290 | return min_vruntime; | |
291 | } | |
292 | ||
54fdc581 FC |
293 | static inline int entity_before(struct sched_entity *a, |
294 | struct sched_entity *b) | |
295 | { | |
296 | return (s64)(a->vruntime - b->vruntime) < 0; | |
297 | } | |
298 | ||
0702e3eb | 299 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 300 | { |
30cfdcfc | 301 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
302 | } |
303 | ||
1af5f730 PZ |
304 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
305 | { | |
306 | u64 vruntime = cfs_rq->min_vruntime; | |
307 | ||
308 | if (cfs_rq->curr) | |
309 | vruntime = cfs_rq->curr->vruntime; | |
310 | ||
311 | if (cfs_rq->rb_leftmost) { | |
312 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
313 | struct sched_entity, | |
314 | run_node); | |
315 | ||
e17036da | 316 | if (!cfs_rq->curr) |
1af5f730 PZ |
317 | vruntime = se->vruntime; |
318 | else | |
319 | vruntime = min_vruntime(vruntime, se->vruntime); | |
320 | } | |
321 | ||
322 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
323 | } | |
324 | ||
bf0f6f24 IM |
325 | /* |
326 | * Enqueue an entity into the rb-tree: | |
327 | */ | |
0702e3eb | 328 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
329 | { |
330 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
331 | struct rb_node *parent = NULL; | |
332 | struct sched_entity *entry; | |
9014623c | 333 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
334 | int leftmost = 1; |
335 | ||
336 | /* | |
337 | * Find the right place in the rbtree: | |
338 | */ | |
339 | while (*link) { | |
340 | parent = *link; | |
341 | entry = rb_entry(parent, struct sched_entity, run_node); | |
342 | /* | |
343 | * We dont care about collisions. Nodes with | |
344 | * the same key stay together. | |
345 | */ | |
9014623c | 346 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
347 | link = &parent->rb_left; |
348 | } else { | |
349 | link = &parent->rb_right; | |
350 | leftmost = 0; | |
351 | } | |
352 | } | |
353 | ||
354 | /* | |
355 | * Maintain a cache of leftmost tree entries (it is frequently | |
356 | * used): | |
357 | */ | |
1af5f730 | 358 | if (leftmost) |
57cb499d | 359 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
360 | |
361 | rb_link_node(&se->run_node, parent, link); | |
362 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
363 | } |
364 | ||
0702e3eb | 365 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 366 | { |
3fe69747 PZ |
367 | if (cfs_rq->rb_leftmost == &se->run_node) { |
368 | struct rb_node *next_node; | |
3fe69747 PZ |
369 | |
370 | next_node = rb_next(&se->run_node); | |
371 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 372 | } |
e9acbff6 | 373 | |
bf0f6f24 | 374 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
375 | } |
376 | ||
bf0f6f24 IM |
377 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
378 | { | |
f4b6755f PZ |
379 | struct rb_node *left = cfs_rq->rb_leftmost; |
380 | ||
381 | if (!left) | |
382 | return NULL; | |
383 | ||
384 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
385 | } |
386 | ||
f4b6755f | 387 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 388 | { |
7eee3e67 | 389 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 390 | |
70eee74b BS |
391 | if (!last) |
392 | return NULL; | |
7eee3e67 IM |
393 | |
394 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
395 | } |
396 | ||
bf0f6f24 IM |
397 | /************************************************************** |
398 | * Scheduling class statistics methods: | |
399 | */ | |
400 | ||
b2be5e96 | 401 | #ifdef CONFIG_SCHED_DEBUG |
acb4a848 | 402 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 403 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
404 | loff_t *ppos) |
405 | { | |
8d65af78 | 406 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 407 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
408 | |
409 | if (ret || !write) | |
410 | return ret; | |
411 | ||
412 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
413 | sysctl_sched_min_granularity); | |
414 | ||
acb4a848 CE |
415 | #define WRT_SYSCTL(name) \ |
416 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
417 | WRT_SYSCTL(sched_min_granularity); | |
418 | WRT_SYSCTL(sched_latency); | |
419 | WRT_SYSCTL(sched_wakeup_granularity); | |
420 | WRT_SYSCTL(sched_shares_ratelimit); | |
421 | #undef WRT_SYSCTL | |
422 | ||
b2be5e96 PZ |
423 | return 0; |
424 | } | |
425 | #endif | |
647e7cac | 426 | |
a7be37ac | 427 | /* |
f9c0b095 | 428 | * delta /= w |
a7be37ac PZ |
429 | */ |
430 | static inline unsigned long | |
431 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
432 | { | |
f9c0b095 PZ |
433 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
434 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
435 | |
436 | return delta; | |
437 | } | |
438 | ||
647e7cac IM |
439 | /* |
440 | * The idea is to set a period in which each task runs once. | |
441 | * | |
442 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
443 | * this period because otherwise the slices get too small. | |
444 | * | |
445 | * p = (nr <= nl) ? l : l*nr/nl | |
446 | */ | |
4d78e7b6 PZ |
447 | static u64 __sched_period(unsigned long nr_running) |
448 | { | |
449 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 450 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
451 | |
452 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 453 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 454 | period *= nr_running; |
4d78e7b6 PZ |
455 | } |
456 | ||
457 | return period; | |
458 | } | |
459 | ||
647e7cac IM |
460 | /* |
461 | * We calculate the wall-time slice from the period by taking a part | |
462 | * proportional to the weight. | |
463 | * | |
f9c0b095 | 464 | * s = p*P[w/rw] |
647e7cac | 465 | */ |
6d0f0ebd | 466 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 467 | { |
0a582440 | 468 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 469 | |
0a582440 | 470 | for_each_sched_entity(se) { |
6272d68c | 471 | struct load_weight *load; |
3104bf03 | 472 | struct load_weight lw; |
6272d68c LM |
473 | |
474 | cfs_rq = cfs_rq_of(se); | |
475 | load = &cfs_rq->load; | |
f9c0b095 | 476 | |
0a582440 | 477 | if (unlikely(!se->on_rq)) { |
3104bf03 | 478 | lw = cfs_rq->load; |
0a582440 MG |
479 | |
480 | update_load_add(&lw, se->load.weight); | |
481 | load = &lw; | |
482 | } | |
483 | slice = calc_delta_mine(slice, se->load.weight, load); | |
484 | } | |
485 | return slice; | |
bf0f6f24 IM |
486 | } |
487 | ||
647e7cac | 488 | /* |
ac884dec | 489 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 490 | * |
f9c0b095 | 491 | * vs = s/w |
647e7cac | 492 | */ |
f9c0b095 | 493 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 494 | { |
f9c0b095 | 495 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
496 | } |
497 | ||
bf0f6f24 IM |
498 | /* |
499 | * Update the current task's runtime statistics. Skip current tasks that | |
500 | * are not in our scheduling class. | |
501 | */ | |
502 | static inline void | |
8ebc91d9 IM |
503 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
504 | unsigned long delta_exec) | |
bf0f6f24 | 505 | { |
bbdba7c0 | 506 | unsigned long delta_exec_weighted; |
bf0f6f24 | 507 | |
41acab88 LDM |
508 | schedstat_set(curr->statistics.exec_max, |
509 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
510 | |
511 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 512 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 513 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 514 | |
e9acbff6 | 515 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 516 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
517 | } |
518 | ||
b7cc0896 | 519 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 520 | { |
429d43bc | 521 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 522 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
523 | unsigned long delta_exec; |
524 | ||
525 | if (unlikely(!curr)) | |
526 | return; | |
527 | ||
528 | /* | |
529 | * Get the amount of time the current task was running | |
530 | * since the last time we changed load (this cannot | |
531 | * overflow on 32 bits): | |
532 | */ | |
8ebc91d9 | 533 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
534 | if (!delta_exec) |
535 | return; | |
bf0f6f24 | 536 | |
8ebc91d9 IM |
537 | __update_curr(cfs_rq, curr, delta_exec); |
538 | curr->exec_start = now; | |
d842de87 SV |
539 | |
540 | if (entity_is_task(curr)) { | |
541 | struct task_struct *curtask = task_of(curr); | |
542 | ||
f977bb49 | 543 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 544 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 545 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 546 | } |
bf0f6f24 IM |
547 | } |
548 | ||
549 | static inline void | |
5870db5b | 550 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 551 | { |
41acab88 | 552 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
553 | } |
554 | ||
bf0f6f24 IM |
555 | /* |
556 | * Task is being enqueued - update stats: | |
557 | */ | |
d2417e5a | 558 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 559 | { |
bf0f6f24 IM |
560 | /* |
561 | * Are we enqueueing a waiting task? (for current tasks | |
562 | * a dequeue/enqueue event is a NOP) | |
563 | */ | |
429d43bc | 564 | if (se != cfs_rq->curr) |
5870db5b | 565 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
566 | } |
567 | ||
bf0f6f24 | 568 | static void |
9ef0a961 | 569 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 570 | { |
41acab88 LDM |
571 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
572 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
573 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
574 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
575 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
576 | #ifdef CONFIG_SCHEDSTATS |
577 | if (entity_is_task(se)) { | |
578 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 579 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
580 | } |
581 | #endif | |
41acab88 | 582 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
583 | } |
584 | ||
585 | static inline void | |
19b6a2e3 | 586 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 587 | { |
bf0f6f24 IM |
588 | /* |
589 | * Mark the end of the wait period if dequeueing a | |
590 | * waiting task: | |
591 | */ | |
429d43bc | 592 | if (se != cfs_rq->curr) |
9ef0a961 | 593 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
594 | } |
595 | ||
596 | /* | |
597 | * We are picking a new current task - update its stats: | |
598 | */ | |
599 | static inline void | |
79303e9e | 600 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
601 | { |
602 | /* | |
603 | * We are starting a new run period: | |
604 | */ | |
305e6835 | 605 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
606 | } |
607 | ||
bf0f6f24 IM |
608 | /************************************************** |
609 | * Scheduling class queueing methods: | |
610 | */ | |
611 | ||
c09595f6 PZ |
612 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
613 | static void | |
614 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
615 | { | |
616 | cfs_rq->task_weight += weight; | |
617 | } | |
618 | #else | |
619 | static inline void | |
620 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
621 | { | |
622 | } | |
623 | #endif | |
624 | ||
30cfdcfc DA |
625 | static void |
626 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
627 | { | |
628 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
629 | if (!parent_entity(se)) |
630 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 631 | if (entity_is_task(se)) { |
c09595f6 | 632 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
633 | list_add(&se->group_node, &cfs_rq->tasks); |
634 | } | |
30cfdcfc DA |
635 | cfs_rq->nr_running++; |
636 | se->on_rq = 1; | |
637 | } | |
638 | ||
639 | static void | |
640 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
641 | { | |
642 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
643 | if (!parent_entity(se)) |
644 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 645 | if (entity_is_task(se)) { |
c09595f6 | 646 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
647 | list_del_init(&se->group_node); |
648 | } | |
30cfdcfc DA |
649 | cfs_rq->nr_running--; |
650 | se->on_rq = 0; | |
651 | } | |
652 | ||
2396af69 | 653 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 654 | { |
bf0f6f24 | 655 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
656 | struct task_struct *tsk = NULL; |
657 | ||
658 | if (entity_is_task(se)) | |
659 | tsk = task_of(se); | |
660 | ||
41acab88 LDM |
661 | if (se->statistics.sleep_start) { |
662 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
663 | |
664 | if ((s64)delta < 0) | |
665 | delta = 0; | |
666 | ||
41acab88 LDM |
667 | if (unlikely(delta > se->statistics.sleep_max)) |
668 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 669 | |
41acab88 LDM |
670 | se->statistics.sleep_start = 0; |
671 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 672 | |
768d0c27 | 673 | if (tsk) { |
e414314c | 674 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
675 | trace_sched_stat_sleep(tsk, delta); |
676 | } | |
bf0f6f24 | 677 | } |
41acab88 LDM |
678 | if (se->statistics.block_start) { |
679 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
680 | |
681 | if ((s64)delta < 0) | |
682 | delta = 0; | |
683 | ||
41acab88 LDM |
684 | if (unlikely(delta > se->statistics.block_max)) |
685 | se->statistics.block_max = delta; | |
bf0f6f24 | 686 | |
41acab88 LDM |
687 | se->statistics.block_start = 0; |
688 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 689 | |
e414314c | 690 | if (tsk) { |
8f0dfc34 | 691 | if (tsk->in_iowait) { |
41acab88 LDM |
692 | se->statistics.iowait_sum += delta; |
693 | se->statistics.iowait_count++; | |
768d0c27 | 694 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
695 | } |
696 | ||
e414314c PZ |
697 | /* |
698 | * Blocking time is in units of nanosecs, so shift by | |
699 | * 20 to get a milliseconds-range estimation of the | |
700 | * amount of time that the task spent sleeping: | |
701 | */ | |
702 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
703 | profile_hits(SLEEP_PROFILING, | |
704 | (void *)get_wchan(tsk), | |
705 | delta >> 20); | |
706 | } | |
707 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 708 | } |
bf0f6f24 IM |
709 | } |
710 | #endif | |
711 | } | |
712 | ||
ddc97297 PZ |
713 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
714 | { | |
715 | #ifdef CONFIG_SCHED_DEBUG | |
716 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
717 | ||
718 | if (d < 0) | |
719 | d = -d; | |
720 | ||
721 | if (d > 3*sysctl_sched_latency) | |
722 | schedstat_inc(cfs_rq, nr_spread_over); | |
723 | #endif | |
724 | } | |
725 | ||
aeb73b04 PZ |
726 | static void |
727 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
728 | { | |
1af5f730 | 729 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 730 | |
2cb8600e PZ |
731 | /* |
732 | * The 'current' period is already promised to the current tasks, | |
733 | * however the extra weight of the new task will slow them down a | |
734 | * little, place the new task so that it fits in the slot that | |
735 | * stays open at the end. | |
736 | */ | |
94dfb5e7 | 737 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 738 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 739 | |
a2e7a7eb | 740 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 741 | if (!initial) { |
a2e7a7eb | 742 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 743 | |
a2e7a7eb MG |
744 | /* |
745 | * Halve their sleep time's effect, to allow | |
746 | * for a gentler effect of sleepers: | |
747 | */ | |
748 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
749 | thresh >>= 1; | |
51e0304c | 750 | |
a2e7a7eb | 751 | vruntime -= thresh; |
aeb73b04 PZ |
752 | } |
753 | ||
b5d9d734 MG |
754 | /* ensure we never gain time by being placed backwards. */ |
755 | vruntime = max_vruntime(se->vruntime, vruntime); | |
756 | ||
67e9fb2a | 757 | se->vruntime = vruntime; |
aeb73b04 PZ |
758 | } |
759 | ||
bf0f6f24 | 760 | static void |
88ec22d3 | 761 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 762 | { |
88ec22d3 PZ |
763 | /* |
764 | * Update the normalized vruntime before updating min_vruntime | |
765 | * through callig update_curr(). | |
766 | */ | |
371fd7e7 | 767 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
768 | se->vruntime += cfs_rq->min_vruntime; |
769 | ||
bf0f6f24 | 770 | /* |
a2a2d680 | 771 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 772 | */ |
b7cc0896 | 773 | update_curr(cfs_rq); |
a992241d | 774 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 775 | |
88ec22d3 | 776 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 777 | place_entity(cfs_rq, se, 0); |
2396af69 | 778 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 779 | } |
bf0f6f24 | 780 | |
d2417e5a | 781 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 782 | check_spread(cfs_rq, se); |
83b699ed SV |
783 | if (se != cfs_rq->curr) |
784 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
785 | } |
786 | ||
a571bbea | 787 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 | 788 | { |
de69a80b | 789 | if (!se || cfs_rq->last == se) |
2002c695 PZ |
790 | cfs_rq->last = NULL; |
791 | ||
de69a80b | 792 | if (!se || cfs_rq->next == se) |
2002c695 PZ |
793 | cfs_rq->next = NULL; |
794 | } | |
795 | ||
a571bbea PZ |
796 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
797 | { | |
798 | for_each_sched_entity(se) | |
799 | __clear_buddies(cfs_rq_of(se), se); | |
800 | } | |
801 | ||
bf0f6f24 | 802 | static void |
371fd7e7 | 803 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 804 | { |
a2a2d680 DA |
805 | /* |
806 | * Update run-time statistics of the 'current'. | |
807 | */ | |
808 | update_curr(cfs_rq); | |
809 | ||
19b6a2e3 | 810 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 811 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 812 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
813 | if (entity_is_task(se)) { |
814 | struct task_struct *tsk = task_of(se); | |
815 | ||
816 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 817 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 818 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 819 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 820 | } |
db36cc7d | 821 | #endif |
67e9fb2a PZ |
822 | } |
823 | ||
2002c695 | 824 | clear_buddies(cfs_rq, se); |
4793241b | 825 | |
83b699ed | 826 | if (se != cfs_rq->curr) |
30cfdcfc DA |
827 | __dequeue_entity(cfs_rq, se); |
828 | account_entity_dequeue(cfs_rq, se); | |
1af5f730 | 829 | update_min_vruntime(cfs_rq); |
88ec22d3 PZ |
830 | |
831 | /* | |
832 | * Normalize the entity after updating the min_vruntime because the | |
833 | * update can refer to the ->curr item and we need to reflect this | |
834 | * movement in our normalized position. | |
835 | */ | |
371fd7e7 | 836 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 837 | se->vruntime -= cfs_rq->min_vruntime; |
bf0f6f24 IM |
838 | } |
839 | ||
840 | /* | |
841 | * Preempt the current task with a newly woken task if needed: | |
842 | */ | |
7c92e54f | 843 | static void |
2e09bf55 | 844 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 845 | { |
11697830 PZ |
846 | unsigned long ideal_runtime, delta_exec; |
847 | ||
6d0f0ebd | 848 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 849 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 850 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 851 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
852 | /* |
853 | * The current task ran long enough, ensure it doesn't get | |
854 | * re-elected due to buddy favours. | |
855 | */ | |
856 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
857 | return; |
858 | } | |
859 | ||
860 | /* | |
861 | * Ensure that a task that missed wakeup preemption by a | |
862 | * narrow margin doesn't have to wait for a full slice. | |
863 | * This also mitigates buddy induced latencies under load. | |
864 | */ | |
865 | if (!sched_feat(WAKEUP_PREEMPT)) | |
866 | return; | |
867 | ||
868 | if (delta_exec < sysctl_sched_min_granularity) | |
869 | return; | |
870 | ||
871 | if (cfs_rq->nr_running > 1) { | |
872 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
873 | s64 delta = curr->vruntime - se->vruntime; | |
874 | ||
875 | if (delta > ideal_runtime) | |
876 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 877 | } |
bf0f6f24 IM |
878 | } |
879 | ||
83b699ed | 880 | static void |
8494f412 | 881 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 882 | { |
83b699ed SV |
883 | /* 'current' is not kept within the tree. */ |
884 | if (se->on_rq) { | |
885 | /* | |
886 | * Any task has to be enqueued before it get to execute on | |
887 | * a CPU. So account for the time it spent waiting on the | |
888 | * runqueue. | |
889 | */ | |
890 | update_stats_wait_end(cfs_rq, se); | |
891 | __dequeue_entity(cfs_rq, se); | |
892 | } | |
893 | ||
79303e9e | 894 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 895 | cfs_rq->curr = se; |
eba1ed4b IM |
896 | #ifdef CONFIG_SCHEDSTATS |
897 | /* | |
898 | * Track our maximum slice length, if the CPU's load is at | |
899 | * least twice that of our own weight (i.e. dont track it | |
900 | * when there are only lesser-weight tasks around): | |
901 | */ | |
495eca49 | 902 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 903 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
904 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
905 | } | |
906 | #endif | |
4a55b450 | 907 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
908 | } |
909 | ||
3f3a4904 PZ |
910 | static int |
911 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
912 | ||
f4b6755f | 913 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 914 | { |
f4b6755f | 915 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
f685ceac | 916 | struct sched_entity *left = se; |
f4b6755f | 917 | |
f685ceac MG |
918 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) |
919 | se = cfs_rq->next; | |
aa2ac252 | 920 | |
f685ceac MG |
921 | /* |
922 | * Prefer last buddy, try to return the CPU to a preempted task. | |
923 | */ | |
924 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
925 | se = cfs_rq->last; | |
926 | ||
927 | clear_buddies(cfs_rq, se); | |
4793241b PZ |
928 | |
929 | return se; | |
aa2ac252 PZ |
930 | } |
931 | ||
ab6cde26 | 932 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
933 | { |
934 | /* | |
935 | * If still on the runqueue then deactivate_task() | |
936 | * was not called and update_curr() has to be done: | |
937 | */ | |
938 | if (prev->on_rq) | |
b7cc0896 | 939 | update_curr(cfs_rq); |
bf0f6f24 | 940 | |
ddc97297 | 941 | check_spread(cfs_rq, prev); |
30cfdcfc | 942 | if (prev->on_rq) { |
5870db5b | 943 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
944 | /* Put 'current' back into the tree. */ |
945 | __enqueue_entity(cfs_rq, prev); | |
946 | } | |
429d43bc | 947 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
948 | } |
949 | ||
8f4d37ec PZ |
950 | static void |
951 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 952 | { |
bf0f6f24 | 953 | /* |
30cfdcfc | 954 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 955 | */ |
30cfdcfc | 956 | update_curr(cfs_rq); |
bf0f6f24 | 957 | |
8f4d37ec PZ |
958 | #ifdef CONFIG_SCHED_HRTICK |
959 | /* | |
960 | * queued ticks are scheduled to match the slice, so don't bother | |
961 | * validating it and just reschedule. | |
962 | */ | |
983ed7a6 HH |
963 | if (queued) { |
964 | resched_task(rq_of(cfs_rq)->curr); | |
965 | return; | |
966 | } | |
8f4d37ec PZ |
967 | /* |
968 | * don't let the period tick interfere with the hrtick preemption | |
969 | */ | |
970 | if (!sched_feat(DOUBLE_TICK) && | |
971 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
972 | return; | |
973 | #endif | |
974 | ||
ce6c1311 | 975 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 976 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
977 | } |
978 | ||
979 | /************************************************** | |
980 | * CFS operations on tasks: | |
981 | */ | |
982 | ||
8f4d37ec PZ |
983 | #ifdef CONFIG_SCHED_HRTICK |
984 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
985 | { | |
8f4d37ec PZ |
986 | struct sched_entity *se = &p->se; |
987 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
988 | ||
989 | WARN_ON(task_rq(p) != rq); | |
990 | ||
991 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
992 | u64 slice = sched_slice(cfs_rq, se); | |
993 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
994 | s64 delta = slice - ran; | |
995 | ||
996 | if (delta < 0) { | |
997 | if (rq->curr == p) | |
998 | resched_task(p); | |
999 | return; | |
1000 | } | |
1001 | ||
1002 | /* | |
1003 | * Don't schedule slices shorter than 10000ns, that just | |
1004 | * doesn't make sense. Rely on vruntime for fairness. | |
1005 | */ | |
31656519 | 1006 | if (rq->curr != p) |
157124c1 | 1007 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1008 | |
31656519 | 1009 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1010 | } |
1011 | } | |
a4c2f00f PZ |
1012 | |
1013 | /* | |
1014 | * called from enqueue/dequeue and updates the hrtick when the | |
1015 | * current task is from our class and nr_running is low enough | |
1016 | * to matter. | |
1017 | */ | |
1018 | static void hrtick_update(struct rq *rq) | |
1019 | { | |
1020 | struct task_struct *curr = rq->curr; | |
1021 | ||
1022 | if (curr->sched_class != &fair_sched_class) | |
1023 | return; | |
1024 | ||
1025 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1026 | hrtick_start_fair(rq, curr); | |
1027 | } | |
55e12e5e | 1028 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1029 | static inline void |
1030 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1031 | { | |
1032 | } | |
a4c2f00f PZ |
1033 | |
1034 | static inline void hrtick_update(struct rq *rq) | |
1035 | { | |
1036 | } | |
8f4d37ec PZ |
1037 | #endif |
1038 | ||
bf0f6f24 IM |
1039 | /* |
1040 | * The enqueue_task method is called before nr_running is | |
1041 | * increased. Here we update the fair scheduling stats and | |
1042 | * then put the task into the rbtree: | |
1043 | */ | |
ea87bb78 | 1044 | static void |
371fd7e7 | 1045 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1046 | { |
1047 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1048 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1049 | |
1050 | for_each_sched_entity(se) { | |
62fb1851 | 1051 | if (se->on_rq) |
bf0f6f24 IM |
1052 | break; |
1053 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 PZ |
1054 | enqueue_entity(cfs_rq, se, flags); |
1055 | flags = ENQUEUE_WAKEUP; | |
bf0f6f24 | 1056 | } |
8f4d37ec | 1057 | |
a4c2f00f | 1058 | hrtick_update(rq); |
bf0f6f24 IM |
1059 | } |
1060 | ||
1061 | /* | |
1062 | * The dequeue_task method is called before nr_running is | |
1063 | * decreased. We remove the task from the rbtree and | |
1064 | * update the fair scheduling stats: | |
1065 | */ | |
371fd7e7 | 1066 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1067 | { |
1068 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1069 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1070 | |
1071 | for_each_sched_entity(se) { | |
1072 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1073 | dequeue_entity(cfs_rq, se, flags); |
bf0f6f24 | 1074 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 1075 | if (cfs_rq->load.weight) |
bf0f6f24 | 1076 | break; |
371fd7e7 | 1077 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1078 | } |
8f4d37ec | 1079 | |
a4c2f00f | 1080 | hrtick_update(rq); |
bf0f6f24 IM |
1081 | } |
1082 | ||
1083 | /* | |
1799e35d IM |
1084 | * sched_yield() support is very simple - we dequeue and enqueue. |
1085 | * | |
1086 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 1087 | */ |
4530d7ab | 1088 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 1089 | { |
db292ca3 IM |
1090 | struct task_struct *curr = rq->curr; |
1091 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1092 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
1093 | |
1094 | /* | |
1799e35d IM |
1095 | * Are we the only task in the tree? |
1096 | */ | |
1097 | if (unlikely(cfs_rq->nr_running == 1)) | |
1098 | return; | |
1099 | ||
2002c695 PZ |
1100 | clear_buddies(cfs_rq, se); |
1101 | ||
db292ca3 | 1102 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1103 | update_rq_clock(rq); |
1799e35d | 1104 | /* |
a2a2d680 | 1105 | * Update run-time statistics of the 'current'. |
1799e35d | 1106 | */ |
2b1e315d | 1107 | update_curr(cfs_rq); |
1799e35d IM |
1108 | |
1109 | return; | |
1110 | } | |
1111 | /* | |
1112 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1113 | */ |
2b1e315d | 1114 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1115 | /* |
1116 | * Already in the rightmost position? | |
1117 | */ | |
54fdc581 | 1118 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
1799e35d IM |
1119 | return; |
1120 | ||
1121 | /* | |
1122 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1123 | * Upon rescheduling, sched_class::put_prev_task() will place |
1124 | * 'current' within the tree based on its new key value. | |
1799e35d | 1125 | */ |
30cfdcfc | 1126 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1127 | } |
1128 | ||
e7693a36 | 1129 | #ifdef CONFIG_SMP |
098fb9db | 1130 | |
88ec22d3 PZ |
1131 | static void task_waking_fair(struct rq *rq, struct task_struct *p) |
1132 | { | |
1133 | struct sched_entity *se = &p->se; | |
1134 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1135 | ||
1136 | se->vruntime -= cfs_rq->min_vruntime; | |
1137 | } | |
1138 | ||
bb3469ac | 1139 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1140 | /* |
1141 | * effective_load() calculates the load change as seen from the root_task_group | |
1142 | * | |
1143 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1144 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1145 | * can calculate the shift in shares. | |
1146 | * | |
1147 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1148 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1149 | * this change. | |
1150 | * | |
1151 | * We compensate this by not only taking the current delta into account, but | |
1152 | * also considering the delta between when the shares were last adjusted and | |
1153 | * now. | |
1154 | * | |
1155 | * We still saw a performance dip, some tracing learned us that between | |
1156 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1157 | * significantly. Therefore try to bias the error in direction of failing | |
1158 | * the affine wakeup. | |
1159 | * | |
1160 | */ | |
f1d239f7 PZ |
1161 | static long effective_load(struct task_group *tg, int cpu, |
1162 | long wl, long wg) | |
bb3469ac | 1163 | { |
4be9daaa | 1164 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1165 | |
1166 | if (!tg->parent) | |
1167 | return wl; | |
1168 | ||
f5bfb7d9 PZ |
1169 | /* |
1170 | * By not taking the decrease of shares on the other cpu into | |
1171 | * account our error leans towards reducing the affine wakeups. | |
1172 | */ | |
1173 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1174 | return wl; | |
1175 | ||
4be9daaa | 1176 | for_each_sched_entity(se) { |
cb5ef42a | 1177 | long S, rw, s, a, b; |
940959e9 PZ |
1178 | long more_w; |
1179 | ||
1180 | /* | |
1181 | * Instead of using this increment, also add the difference | |
1182 | * between when the shares were last updated and now. | |
1183 | */ | |
1184 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1185 | wl += more_w; | |
1186 | wg += more_w; | |
4be9daaa PZ |
1187 | |
1188 | S = se->my_q->tg->shares; | |
1189 | s = se->my_q->shares; | |
f1d239f7 | 1190 | rw = se->my_q->rq_weight; |
bb3469ac | 1191 | |
cb5ef42a PZ |
1192 | a = S*(rw + wl); |
1193 | b = S*rw + s*wg; | |
4be9daaa | 1194 | |
940959e9 PZ |
1195 | wl = s*(a-b); |
1196 | ||
1197 | if (likely(b)) | |
1198 | wl /= b; | |
1199 | ||
83378269 PZ |
1200 | /* |
1201 | * Assume the group is already running and will | |
1202 | * thus already be accounted for in the weight. | |
1203 | * | |
1204 | * That is, moving shares between CPUs, does not | |
1205 | * alter the group weight. | |
1206 | */ | |
4be9daaa | 1207 | wg = 0; |
4be9daaa | 1208 | } |
bb3469ac | 1209 | |
4be9daaa | 1210 | return wl; |
bb3469ac | 1211 | } |
4be9daaa | 1212 | |
bb3469ac | 1213 | #else |
4be9daaa | 1214 | |
83378269 PZ |
1215 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1216 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1217 | { |
83378269 | 1218 | return wl; |
bb3469ac | 1219 | } |
4be9daaa | 1220 | |
bb3469ac PZ |
1221 | #endif |
1222 | ||
c88d5910 | 1223 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1224 | { |
c88d5910 PZ |
1225 | unsigned long this_load, load; |
1226 | int idx, this_cpu, prev_cpu; | |
098fb9db | 1227 | unsigned long tl_per_task; |
c88d5910 | 1228 | struct task_group *tg; |
83378269 | 1229 | unsigned long weight; |
b3137bc8 | 1230 | int balanced; |
098fb9db | 1231 | |
c88d5910 PZ |
1232 | idx = sd->wake_idx; |
1233 | this_cpu = smp_processor_id(); | |
1234 | prev_cpu = task_cpu(p); | |
1235 | load = source_load(prev_cpu, idx); | |
1236 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1237 | |
b3137bc8 MG |
1238 | /* |
1239 | * If sync wakeup then subtract the (maximum possible) | |
1240 | * effect of the currently running task from the load | |
1241 | * of the current CPU: | |
1242 | */ | |
f3b577de | 1243 | rcu_read_lock(); |
83378269 PZ |
1244 | if (sync) { |
1245 | tg = task_group(current); | |
1246 | weight = current->se.load.weight; | |
1247 | ||
c88d5910 | 1248 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1249 | load += effective_load(tg, prev_cpu, 0, -weight); |
1250 | } | |
b3137bc8 | 1251 | |
83378269 PZ |
1252 | tg = task_group(p); |
1253 | weight = p->se.load.weight; | |
b3137bc8 | 1254 | |
71a29aa7 PZ |
1255 | /* |
1256 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1257 | * due to the sync cause above having dropped this_load to 0, we'll |
1258 | * always have an imbalance, but there's really nothing you can do | |
1259 | * about that, so that's good too. | |
71a29aa7 PZ |
1260 | * |
1261 | * Otherwise check if either cpus are near enough in load to allow this | |
1262 | * task to be woken on this_cpu. | |
1263 | */ | |
e51fd5e2 PZ |
1264 | if (this_load) { |
1265 | unsigned long this_eff_load, prev_eff_load; | |
1266 | ||
1267 | this_eff_load = 100; | |
1268 | this_eff_load *= power_of(prev_cpu); | |
1269 | this_eff_load *= this_load + | |
1270 | effective_load(tg, this_cpu, weight, weight); | |
1271 | ||
1272 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
1273 | prev_eff_load *= power_of(this_cpu); | |
1274 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
1275 | ||
1276 | balanced = this_eff_load <= prev_eff_load; | |
1277 | } else | |
1278 | balanced = true; | |
f3b577de | 1279 | rcu_read_unlock(); |
b3137bc8 | 1280 | |
098fb9db | 1281 | /* |
4ae7d5ce IM |
1282 | * If the currently running task will sleep within |
1283 | * a reasonable amount of time then attract this newly | |
1284 | * woken task: | |
098fb9db | 1285 | */ |
2fb7635c PZ |
1286 | if (sync && balanced) |
1287 | return 1; | |
098fb9db | 1288 | |
41acab88 | 1289 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
1290 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
1291 | ||
c88d5910 PZ |
1292 | if (balanced || |
1293 | (this_load <= load && | |
1294 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1295 | /* |
1296 | * This domain has SD_WAKE_AFFINE and | |
1297 | * p is cache cold in this domain, and | |
1298 | * there is no bad imbalance. | |
1299 | */ | |
c88d5910 | 1300 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 1301 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
1302 | |
1303 | return 1; | |
1304 | } | |
1305 | return 0; | |
1306 | } | |
1307 | ||
aaee1203 PZ |
1308 | /* |
1309 | * find_idlest_group finds and returns the least busy CPU group within the | |
1310 | * domain. | |
1311 | */ | |
1312 | static struct sched_group * | |
78e7ed53 | 1313 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1314 | int this_cpu, int load_idx) |
e7693a36 | 1315 | { |
b3bd3de6 | 1316 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 1317 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 1318 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1319 | |
aaee1203 PZ |
1320 | do { |
1321 | unsigned long load, avg_load; | |
1322 | int local_group; | |
1323 | int i; | |
e7693a36 | 1324 | |
aaee1203 PZ |
1325 | /* Skip over this group if it has no CPUs allowed */ |
1326 | if (!cpumask_intersects(sched_group_cpus(group), | |
1327 | &p->cpus_allowed)) | |
1328 | continue; | |
1329 | ||
1330 | local_group = cpumask_test_cpu(this_cpu, | |
1331 | sched_group_cpus(group)); | |
1332 | ||
1333 | /* Tally up the load of all CPUs in the group */ | |
1334 | avg_load = 0; | |
1335 | ||
1336 | for_each_cpu(i, sched_group_cpus(group)) { | |
1337 | /* Bias balancing toward cpus of our domain */ | |
1338 | if (local_group) | |
1339 | load = source_load(i, load_idx); | |
1340 | else | |
1341 | load = target_load(i, load_idx); | |
1342 | ||
1343 | avg_load += load; | |
1344 | } | |
1345 | ||
1346 | /* Adjust by relative CPU power of the group */ | |
1347 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1348 | ||
1349 | if (local_group) { | |
1350 | this_load = avg_load; | |
aaee1203 PZ |
1351 | } else if (avg_load < min_load) { |
1352 | min_load = avg_load; | |
1353 | idlest = group; | |
1354 | } | |
1355 | } while (group = group->next, group != sd->groups); | |
1356 | ||
1357 | if (!idlest || 100*this_load < imbalance*min_load) | |
1358 | return NULL; | |
1359 | return idlest; | |
1360 | } | |
1361 | ||
1362 | /* | |
1363 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1364 | */ | |
1365 | static int | |
1366 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1367 | { | |
1368 | unsigned long load, min_load = ULONG_MAX; | |
1369 | int idlest = -1; | |
1370 | int i; | |
1371 | ||
1372 | /* Traverse only the allowed CPUs */ | |
1373 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1374 | load = weighted_cpuload(i); | |
1375 | ||
1376 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1377 | min_load = load; | |
1378 | idlest = i; | |
e7693a36 GH |
1379 | } |
1380 | } | |
1381 | ||
aaee1203 PZ |
1382 | return idlest; |
1383 | } | |
e7693a36 | 1384 | |
a50bde51 PZ |
1385 | /* |
1386 | * Try and locate an idle CPU in the sched_domain. | |
1387 | */ | |
99bd5e2f | 1388 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
1389 | { |
1390 | int cpu = smp_processor_id(); | |
1391 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 1392 | struct sched_domain *sd; |
a50bde51 PZ |
1393 | int i; |
1394 | ||
1395 | /* | |
99bd5e2f SS |
1396 | * If the task is going to be woken-up on this cpu and if it is |
1397 | * already idle, then it is the right target. | |
a50bde51 | 1398 | */ |
99bd5e2f SS |
1399 | if (target == cpu && idle_cpu(cpu)) |
1400 | return cpu; | |
1401 | ||
1402 | /* | |
1403 | * If the task is going to be woken-up on the cpu where it previously | |
1404 | * ran and if it is currently idle, then it the right target. | |
1405 | */ | |
1406 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 1407 | return prev_cpu; |
a50bde51 PZ |
1408 | |
1409 | /* | |
99bd5e2f | 1410 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 1411 | */ |
99bd5e2f SS |
1412 | for_each_domain(target, sd) { |
1413 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 1414 | break; |
99bd5e2f SS |
1415 | |
1416 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1417 | if (idle_cpu(i)) { | |
1418 | target = i; | |
1419 | break; | |
1420 | } | |
a50bde51 | 1421 | } |
99bd5e2f SS |
1422 | |
1423 | /* | |
1424 | * Lets stop looking for an idle sibling when we reached | |
1425 | * the domain that spans the current cpu and prev_cpu. | |
1426 | */ | |
1427 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1428 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1429 | break; | |
a50bde51 PZ |
1430 | } |
1431 | ||
1432 | return target; | |
1433 | } | |
1434 | ||
aaee1203 PZ |
1435 | /* |
1436 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1437 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1438 | * SD_BALANCE_EXEC. | |
1439 | * | |
1440 | * Balance, ie. select the least loaded group. | |
1441 | * | |
1442 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1443 | * | |
1444 | * preempt must be disabled. | |
1445 | */ | |
0017d735 PZ |
1446 | static int |
1447 | select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags) | |
aaee1203 | 1448 | { |
29cd8bae | 1449 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1450 | int cpu = smp_processor_id(); |
1451 | int prev_cpu = task_cpu(p); | |
1452 | int new_cpu = cpu; | |
99bd5e2f | 1453 | int want_affine = 0; |
29cd8bae | 1454 | int want_sd = 1; |
5158f4e4 | 1455 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1456 | |
0763a660 | 1457 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 1458 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
1459 | want_affine = 1; |
1460 | new_cpu = prev_cpu; | |
1461 | } | |
aaee1203 PZ |
1462 | |
1463 | for_each_domain(cpu, tmp) { | |
e4f42888 PZ |
1464 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1465 | continue; | |
1466 | ||
aaee1203 | 1467 | /* |
ae154be1 PZ |
1468 | * If power savings logic is enabled for a domain, see if we |
1469 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1470 | */ |
59abf026 | 1471 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1472 | unsigned long power = 0; |
1473 | unsigned long nr_running = 0; | |
1474 | unsigned long capacity; | |
1475 | int i; | |
1476 | ||
1477 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1478 | power += power_of(i); | |
1479 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1480 | } | |
1481 | ||
1482 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1483 | ||
59abf026 PZ |
1484 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1485 | nr_running /= 2; | |
1486 | ||
1487 | if (nr_running < capacity) | |
29cd8bae | 1488 | want_sd = 0; |
ae154be1 | 1489 | } |
aaee1203 | 1490 | |
fe3bcfe1 | 1491 | /* |
99bd5e2f SS |
1492 | * If both cpu and prev_cpu are part of this domain, |
1493 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 1494 | */ |
99bd5e2f SS |
1495 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1496 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1497 | affine_sd = tmp; | |
1498 | want_affine = 0; | |
c88d5910 PZ |
1499 | } |
1500 | ||
29cd8bae PZ |
1501 | if (!want_sd && !want_affine) |
1502 | break; | |
1503 | ||
0763a660 | 1504 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1505 | continue; |
1506 | ||
29cd8bae PZ |
1507 | if (want_sd) |
1508 | sd = tmp; | |
1509 | } | |
1510 | ||
8b911acd | 1511 | #ifdef CONFIG_FAIR_GROUP_SCHED |
29cd8bae PZ |
1512 | if (sched_feat(LB_SHARES_UPDATE)) { |
1513 | /* | |
1514 | * Pick the largest domain to update shares over | |
1515 | */ | |
1516 | tmp = sd; | |
669c55e9 | 1517 | if (affine_sd && (!tmp || affine_sd->span_weight > sd->span_weight)) |
29cd8bae PZ |
1518 | tmp = affine_sd; |
1519 | ||
0017d735 PZ |
1520 | if (tmp) { |
1521 | raw_spin_unlock(&rq->lock); | |
29cd8bae | 1522 | update_shares(tmp); |
0017d735 PZ |
1523 | raw_spin_lock(&rq->lock); |
1524 | } | |
c88d5910 | 1525 | } |
8b911acd | 1526 | #endif |
aaee1203 | 1527 | |
8b911acd | 1528 | if (affine_sd) { |
99bd5e2f SS |
1529 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
1530 | return select_idle_sibling(p, cpu); | |
1531 | else | |
1532 | return select_idle_sibling(p, prev_cpu); | |
8b911acd | 1533 | } |
e7693a36 | 1534 | |
aaee1203 | 1535 | while (sd) { |
5158f4e4 | 1536 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1537 | struct sched_group *group; |
c88d5910 | 1538 | int weight; |
098fb9db | 1539 | |
0763a660 | 1540 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1541 | sd = sd->child; |
1542 | continue; | |
1543 | } | |
098fb9db | 1544 | |
5158f4e4 PZ |
1545 | if (sd_flag & SD_BALANCE_WAKE) |
1546 | load_idx = sd->wake_idx; | |
098fb9db | 1547 | |
5158f4e4 | 1548 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1549 | if (!group) { |
1550 | sd = sd->child; | |
1551 | continue; | |
1552 | } | |
4ae7d5ce | 1553 | |
d7c33c49 | 1554 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1555 | if (new_cpu == -1 || new_cpu == cpu) { |
1556 | /* Now try balancing at a lower domain level of cpu */ | |
1557 | sd = sd->child; | |
1558 | continue; | |
e7693a36 | 1559 | } |
aaee1203 PZ |
1560 | |
1561 | /* Now try balancing at a lower domain level of new_cpu */ | |
1562 | cpu = new_cpu; | |
669c55e9 | 1563 | weight = sd->span_weight; |
aaee1203 PZ |
1564 | sd = NULL; |
1565 | for_each_domain(cpu, tmp) { | |
669c55e9 | 1566 | if (weight <= tmp->span_weight) |
aaee1203 | 1567 | break; |
0763a660 | 1568 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1569 | sd = tmp; |
1570 | } | |
1571 | /* while loop will break here if sd == NULL */ | |
e7693a36 GH |
1572 | } |
1573 | ||
c88d5910 | 1574 | return new_cpu; |
e7693a36 GH |
1575 | } |
1576 | #endif /* CONFIG_SMP */ | |
1577 | ||
e52fb7c0 PZ |
1578 | static unsigned long |
1579 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1580 | { |
1581 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1582 | ||
1583 | /* | |
e52fb7c0 PZ |
1584 | * Since its curr running now, convert the gran from real-time |
1585 | * to virtual-time in his units. | |
13814d42 MG |
1586 | * |
1587 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1588 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1589 | * the resulting gran will be larger, therefore penalizing the | |
1590 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1591 | * be smaller, again penalizing the lighter task. | |
1592 | * | |
1593 | * This is especially important for buddies when the leftmost | |
1594 | * task is higher priority than the buddy. | |
0bbd3336 | 1595 | */ |
13814d42 MG |
1596 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
1597 | gran = calc_delta_fair(gran, se); | |
0bbd3336 PZ |
1598 | |
1599 | return gran; | |
1600 | } | |
1601 | ||
464b7527 PZ |
1602 | /* |
1603 | * Should 'se' preempt 'curr'. | |
1604 | * | |
1605 | * |s1 | |
1606 | * |s2 | |
1607 | * |s3 | |
1608 | * g | |
1609 | * |<--->|c | |
1610 | * | |
1611 | * w(c, s1) = -1 | |
1612 | * w(c, s2) = 0 | |
1613 | * w(c, s3) = 1 | |
1614 | * | |
1615 | */ | |
1616 | static int | |
1617 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1618 | { | |
1619 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1620 | ||
1621 | if (vdiff <= 0) | |
1622 | return -1; | |
1623 | ||
e52fb7c0 | 1624 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1625 | if (vdiff > gran) |
1626 | return 1; | |
1627 | ||
1628 | return 0; | |
1629 | } | |
1630 | ||
02479099 PZ |
1631 | static void set_last_buddy(struct sched_entity *se) |
1632 | { | |
6bc912b7 PZ |
1633 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1634 | for_each_sched_entity(se) | |
1635 | cfs_rq_of(se)->last = se; | |
1636 | } | |
02479099 PZ |
1637 | } |
1638 | ||
1639 | static void set_next_buddy(struct sched_entity *se) | |
1640 | { | |
6bc912b7 PZ |
1641 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1642 | for_each_sched_entity(se) | |
1643 | cfs_rq_of(se)->next = se; | |
1644 | } | |
02479099 PZ |
1645 | } |
1646 | ||
bf0f6f24 IM |
1647 | /* |
1648 | * Preempt the current task with a newly woken task if needed: | |
1649 | */ | |
5a9b86f6 | 1650 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1651 | { |
1652 | struct task_struct *curr = rq->curr; | |
8651a86c | 1653 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1654 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 1655 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
bf0f6f24 | 1656 | |
4ae7d5ce IM |
1657 | if (unlikely(se == pse)) |
1658 | return; | |
1659 | ||
f685ceac | 1660 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) |
3cb63d52 | 1661 | set_next_buddy(pse); |
57fdc26d | 1662 | |
aec0a514 BR |
1663 | /* |
1664 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1665 | * wake up path. | |
1666 | */ | |
1667 | if (test_tsk_need_resched(curr)) | |
1668 | return; | |
1669 | ||
91c234b4 | 1670 | /* |
6bc912b7 | 1671 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1672 | * the tick): |
1673 | */ | |
6bc912b7 | 1674 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1675 | return; |
bf0f6f24 | 1676 | |
6bc912b7 | 1677 | /* Idle tasks are by definition preempted by everybody. */ |
3a7e73a2 PZ |
1678 | if (unlikely(curr->policy == SCHED_IDLE)) |
1679 | goto preempt; | |
bf0f6f24 | 1680 | |
ad4b78bb PZ |
1681 | if (!sched_feat(WAKEUP_PREEMPT)) |
1682 | return; | |
1683 | ||
3a7e73a2 | 1684 | update_curr(cfs_rq); |
464b7527 | 1685 | find_matching_se(&se, &pse); |
002f128b | 1686 | BUG_ON(!pse); |
3a7e73a2 PZ |
1687 | if (wakeup_preempt_entity(se, pse) == 1) |
1688 | goto preempt; | |
464b7527 | 1689 | |
3a7e73a2 | 1690 | return; |
a65ac745 | 1691 | |
3a7e73a2 PZ |
1692 | preempt: |
1693 | resched_task(curr); | |
1694 | /* | |
1695 | * Only set the backward buddy when the current task is still | |
1696 | * on the rq. This can happen when a wakeup gets interleaved | |
1697 | * with schedule on the ->pre_schedule() or idle_balance() | |
1698 | * point, either of which can * drop the rq lock. | |
1699 | * | |
1700 | * Also, during early boot the idle thread is in the fair class, | |
1701 | * for obvious reasons its a bad idea to schedule back to it. | |
1702 | */ | |
1703 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1704 | return; | |
1705 | ||
1706 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1707 | set_last_buddy(se); | |
bf0f6f24 IM |
1708 | } |
1709 | ||
fb8d4724 | 1710 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1711 | { |
8f4d37ec | 1712 | struct task_struct *p; |
bf0f6f24 IM |
1713 | struct cfs_rq *cfs_rq = &rq->cfs; |
1714 | struct sched_entity *se; | |
1715 | ||
36ace27e | 1716 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
1717 | return NULL; |
1718 | ||
1719 | do { | |
9948f4b2 | 1720 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1721 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1722 | cfs_rq = group_cfs_rq(se); |
1723 | } while (cfs_rq); | |
1724 | ||
8f4d37ec PZ |
1725 | p = task_of(se); |
1726 | hrtick_start_fair(rq, p); | |
1727 | ||
1728 | return p; | |
bf0f6f24 IM |
1729 | } |
1730 | ||
1731 | /* | |
1732 | * Account for a descheduled task: | |
1733 | */ | |
31ee529c | 1734 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1735 | { |
1736 | struct sched_entity *se = &prev->se; | |
1737 | struct cfs_rq *cfs_rq; | |
1738 | ||
1739 | for_each_sched_entity(se) { | |
1740 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1741 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1742 | } |
1743 | } | |
1744 | ||
681f3e68 | 1745 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1746 | /************************************************** |
1747 | * Fair scheduling class load-balancing methods: | |
1748 | */ | |
1749 | ||
1e3c88bd PZ |
1750 | /* |
1751 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1752 | * Both runqueues must be locked. | |
1753 | */ | |
1754 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
1755 | struct rq *this_rq, int this_cpu) | |
1756 | { | |
1757 | deactivate_task(src_rq, p, 0); | |
1758 | set_task_cpu(p, this_cpu); | |
1759 | activate_task(this_rq, p, 0); | |
1760 | check_preempt_curr(this_rq, p, 0); | |
fab47622 NR |
1761 | |
1762 | /* re-arm NEWIDLE balancing when moving tasks */ | |
1763 | src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost; | |
1764 | this_rq->idle_stamp = 0; | |
1e3c88bd PZ |
1765 | } |
1766 | ||
1767 | /* | |
1768 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1769 | */ | |
1770 | static | |
1771 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
1772 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1773 | int *all_pinned) | |
1774 | { | |
1775 | int tsk_cache_hot = 0; | |
1776 | /* | |
1777 | * We do not migrate tasks that are: | |
1778 | * 1) running (obviously), or | |
1779 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1780 | * 3) are cache-hot on their current CPU. | |
1781 | */ | |
1782 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 1783 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
1784 | return 0; |
1785 | } | |
1786 | *all_pinned = 0; | |
1787 | ||
1788 | if (task_running(rq, p)) { | |
41acab88 | 1789 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
1790 | return 0; |
1791 | } | |
1792 | ||
1793 | /* | |
1794 | * Aggressive migration if: | |
1795 | * 1) task is cache cold, or | |
1796 | * 2) too many balance attempts have failed. | |
1797 | */ | |
1798 | ||
305e6835 | 1799 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
1800 | if (!tsk_cache_hot || |
1801 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
1802 | #ifdef CONFIG_SCHEDSTATS | |
1803 | if (tsk_cache_hot) { | |
1804 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 1805 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
1806 | } |
1807 | #endif | |
1808 | return 1; | |
1809 | } | |
1810 | ||
1811 | if (tsk_cache_hot) { | |
41acab88 | 1812 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
1813 | return 0; |
1814 | } | |
1815 | return 1; | |
1816 | } | |
1817 | ||
897c395f PZ |
1818 | /* |
1819 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
1820 | * part of active balancing operations within "domain". | |
1821 | * Returns 1 if successful and 0 otherwise. | |
1822 | * | |
1823 | * Called with both runqueues locked. | |
1824 | */ | |
1825 | static int | |
1826 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1827 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1828 | { | |
1829 | struct task_struct *p, *n; | |
1830 | struct cfs_rq *cfs_rq; | |
1831 | int pinned = 0; | |
1832 | ||
1833 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
1834 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
1835 | ||
1836 | if (!can_migrate_task(p, busiest, this_cpu, | |
1837 | sd, idle, &pinned)) | |
1838 | continue; | |
1839 | ||
1840 | pull_task(busiest, p, this_rq, this_cpu); | |
1841 | /* | |
1842 | * Right now, this is only the second place pull_task() | |
1843 | * is called, so we can safely collect pull_task() | |
1844 | * stats here rather than inside pull_task(). | |
1845 | */ | |
1846 | schedstat_inc(sd, lb_gained[idle]); | |
1847 | return 1; | |
1848 | } | |
1849 | } | |
1850 | ||
1851 | return 0; | |
1852 | } | |
1853 | ||
1e3c88bd PZ |
1854 | static unsigned long |
1855 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1856 | unsigned long max_load_move, struct sched_domain *sd, | |
1857 | enum cpu_idle_type idle, int *all_pinned, | |
ee00e66f | 1858 | int *this_best_prio, struct cfs_rq *busiest_cfs_rq) |
1e3c88bd PZ |
1859 | { |
1860 | int loops = 0, pulled = 0, pinned = 0; | |
1e3c88bd | 1861 | long rem_load_move = max_load_move; |
ee00e66f | 1862 | struct task_struct *p, *n; |
1e3c88bd PZ |
1863 | |
1864 | if (max_load_move == 0) | |
1865 | goto out; | |
1866 | ||
1867 | pinned = 1; | |
1868 | ||
ee00e66f PZ |
1869 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
1870 | if (loops++ > sysctl_sched_nr_migrate) | |
1871 | break; | |
1e3c88bd | 1872 | |
ee00e66f PZ |
1873 | if ((p->se.load.weight >> 1) > rem_load_move || |
1874 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) | |
1875 | continue; | |
1e3c88bd | 1876 | |
ee00e66f PZ |
1877 | pull_task(busiest, p, this_rq, this_cpu); |
1878 | pulled++; | |
1879 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
1880 | |
1881 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
1882 | /* |
1883 | * NEWIDLE balancing is a source of latency, so preemptible | |
1884 | * kernels will stop after the first task is pulled to minimize | |
1885 | * the critical section. | |
1886 | */ | |
1887 | if (idle == CPU_NEWLY_IDLE) | |
1888 | break; | |
1e3c88bd PZ |
1889 | #endif |
1890 | ||
ee00e66f PZ |
1891 | /* |
1892 | * We only want to steal up to the prescribed amount of | |
1893 | * weighted load. | |
1894 | */ | |
1895 | if (rem_load_move <= 0) | |
1896 | break; | |
1897 | ||
1e3c88bd PZ |
1898 | if (p->prio < *this_best_prio) |
1899 | *this_best_prio = p->prio; | |
1e3c88bd PZ |
1900 | } |
1901 | out: | |
1902 | /* | |
1903 | * Right now, this is one of only two places pull_task() is called, | |
1904 | * so we can safely collect pull_task() stats here rather than | |
1905 | * inside pull_task(). | |
1906 | */ | |
1907 | schedstat_add(sd, lb_gained[idle], pulled); | |
1908 | ||
1909 | if (all_pinned) | |
1910 | *all_pinned = pinned; | |
1911 | ||
1912 | return max_load_move - rem_load_move; | |
1913 | } | |
1914 | ||
230059de PZ |
1915 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1916 | static unsigned long | |
1917 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1918 | unsigned long max_load_move, | |
1919 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1920 | int *all_pinned, int *this_best_prio) | |
1921 | { | |
1922 | long rem_load_move = max_load_move; | |
1923 | int busiest_cpu = cpu_of(busiest); | |
1924 | struct task_group *tg; | |
1925 | ||
1926 | rcu_read_lock(); | |
1927 | update_h_load(busiest_cpu); | |
1928 | ||
1929 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
1930 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; | |
1931 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; | |
1932 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
1933 | u64 rem_load, moved_load; | |
1934 | ||
1935 | /* | |
1936 | * empty group | |
1937 | */ | |
1938 | if (!busiest_cfs_rq->task_weight) | |
1939 | continue; | |
1940 | ||
1941 | rem_load = (u64)rem_load_move * busiest_weight; | |
1942 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
1943 | ||
1944 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
1945 | rem_load, sd, idle, all_pinned, this_best_prio, | |
1946 | busiest_cfs_rq); | |
1947 | ||
1948 | if (!moved_load) | |
1949 | continue; | |
1950 | ||
1951 | moved_load *= busiest_h_load; | |
1952 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
1953 | ||
1954 | rem_load_move -= moved_load; | |
1955 | if (rem_load_move < 0) | |
1956 | break; | |
1957 | } | |
1958 | rcu_read_unlock(); | |
1959 | ||
1960 | return max_load_move - rem_load_move; | |
1961 | } | |
1962 | #else | |
1963 | static unsigned long | |
1964 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1965 | unsigned long max_load_move, | |
1966 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1967 | int *all_pinned, int *this_best_prio) | |
1968 | { | |
1969 | return balance_tasks(this_rq, this_cpu, busiest, | |
1970 | max_load_move, sd, idle, all_pinned, | |
1971 | this_best_prio, &busiest->cfs); | |
1972 | } | |
1973 | #endif | |
1974 | ||
1e3c88bd PZ |
1975 | /* |
1976 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
1977 | * this_rq, as part of a balancing operation within domain "sd". | |
1978 | * Returns 1 if successful and 0 otherwise. | |
1979 | * | |
1980 | * Called with both runqueues locked. | |
1981 | */ | |
1982 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1983 | unsigned long max_load_move, | |
1984 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1985 | int *all_pinned) | |
1986 | { | |
3d45fd80 | 1987 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
1988 | int this_best_prio = this_rq->curr->prio; |
1989 | ||
1990 | do { | |
3d45fd80 | 1991 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd PZ |
1992 | max_load_move - total_load_moved, |
1993 | sd, idle, all_pinned, &this_best_prio); | |
3d45fd80 PZ |
1994 | |
1995 | total_load_moved += load_moved; | |
1e3c88bd PZ |
1996 | |
1997 | #ifdef CONFIG_PREEMPT | |
1998 | /* | |
1999 | * NEWIDLE balancing is a source of latency, so preemptible | |
2000 | * kernels will stop after the first task is pulled to minimize | |
2001 | * the critical section. | |
2002 | */ | |
2003 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2004 | break; | |
baa8c110 PZ |
2005 | |
2006 | if (raw_spin_is_contended(&this_rq->lock) || | |
2007 | raw_spin_is_contended(&busiest->lock)) | |
2008 | break; | |
1e3c88bd | 2009 | #endif |
3d45fd80 | 2010 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2011 | |
2012 | return total_load_moved > 0; | |
2013 | } | |
2014 | ||
1e3c88bd PZ |
2015 | /********** Helpers for find_busiest_group ************************/ |
2016 | /* | |
2017 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2018 | * during load balancing. | |
2019 | */ | |
2020 | struct sd_lb_stats { | |
2021 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2022 | struct sched_group *this; /* Local group in this sd */ | |
2023 | unsigned long total_load; /* Total load of all groups in sd */ | |
2024 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2025 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2026 | ||
2027 | /** Statistics of this group */ | |
2028 | unsigned long this_load; | |
2029 | unsigned long this_load_per_task; | |
2030 | unsigned long this_nr_running; | |
fab47622 | 2031 | unsigned long this_has_capacity; |
aae6d3dd | 2032 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
2033 | |
2034 | /* Statistics of the busiest group */ | |
aae6d3dd | 2035 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
2036 | unsigned long max_load; |
2037 | unsigned long busiest_load_per_task; | |
2038 | unsigned long busiest_nr_running; | |
dd5feea1 | 2039 | unsigned long busiest_group_capacity; |
fab47622 | 2040 | unsigned long busiest_has_capacity; |
aae6d3dd | 2041 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
2042 | |
2043 | int group_imb; /* Is there imbalance in this sd */ | |
2044 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2045 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2046 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2047 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2048 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2049 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2050 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2051 | #endif | |
2052 | }; | |
2053 | ||
2054 | /* | |
2055 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2056 | */ | |
2057 | struct sg_lb_stats { | |
2058 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2059 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2060 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2061 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2062 | unsigned long group_capacity; | |
aae6d3dd SS |
2063 | unsigned long idle_cpus; |
2064 | unsigned long group_weight; | |
1e3c88bd | 2065 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 2066 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2067 | }; |
2068 | ||
2069 | /** | |
2070 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2071 | * @group: The group whose first cpu is to be returned. | |
2072 | */ | |
2073 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2074 | { | |
2075 | return cpumask_first(sched_group_cpus(group)); | |
2076 | } | |
2077 | ||
2078 | /** | |
2079 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2080 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2081 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2082 | */ | |
2083 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2084 | enum cpu_idle_type idle) | |
2085 | { | |
2086 | int load_idx; | |
2087 | ||
2088 | switch (idle) { | |
2089 | case CPU_NOT_IDLE: | |
2090 | load_idx = sd->busy_idx; | |
2091 | break; | |
2092 | ||
2093 | case CPU_NEWLY_IDLE: | |
2094 | load_idx = sd->newidle_idx; | |
2095 | break; | |
2096 | default: | |
2097 | load_idx = sd->idle_idx; | |
2098 | break; | |
2099 | } | |
2100 | ||
2101 | return load_idx; | |
2102 | } | |
2103 | ||
2104 | ||
2105 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2106 | /** | |
2107 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2108 | * the given sched_domain, during load balancing. | |
2109 | * | |
2110 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2111 | * @sds: Variable containing the statistics for sd. | |
2112 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2113 | */ | |
2114 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2115 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2116 | { | |
2117 | /* | |
2118 | * Busy processors will not participate in power savings | |
2119 | * balance. | |
2120 | */ | |
2121 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2122 | sds->power_savings_balance = 0; | |
2123 | else { | |
2124 | sds->power_savings_balance = 1; | |
2125 | sds->min_nr_running = ULONG_MAX; | |
2126 | sds->leader_nr_running = 0; | |
2127 | } | |
2128 | } | |
2129 | ||
2130 | /** | |
2131 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2132 | * sched_domain while performing load balancing. | |
2133 | * | |
2134 | * @group: sched_group belonging to the sched_domain under consideration. | |
2135 | * @sds: Variable containing the statistics of the sched_domain | |
2136 | * @local_group: Does group contain the CPU for which we're performing | |
2137 | * load balancing ? | |
2138 | * @sgs: Variable containing the statistics of the group. | |
2139 | */ | |
2140 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2141 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2142 | { | |
2143 | ||
2144 | if (!sds->power_savings_balance) | |
2145 | return; | |
2146 | ||
2147 | /* | |
2148 | * If the local group is idle or completely loaded | |
2149 | * no need to do power savings balance at this domain | |
2150 | */ | |
2151 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2152 | !sds->this_nr_running)) | |
2153 | sds->power_savings_balance = 0; | |
2154 | ||
2155 | /* | |
2156 | * If a group is already running at full capacity or idle, | |
2157 | * don't include that group in power savings calculations | |
2158 | */ | |
2159 | if (!sds->power_savings_balance || | |
2160 | sgs->sum_nr_running >= sgs->group_capacity || | |
2161 | !sgs->sum_nr_running) | |
2162 | return; | |
2163 | ||
2164 | /* | |
2165 | * Calculate the group which has the least non-idle load. | |
2166 | * This is the group from where we need to pick up the load | |
2167 | * for saving power | |
2168 | */ | |
2169 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2170 | (sgs->sum_nr_running == sds->min_nr_running && | |
2171 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2172 | sds->group_min = group; | |
2173 | sds->min_nr_running = sgs->sum_nr_running; | |
2174 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2175 | sgs->sum_nr_running; | |
2176 | } | |
2177 | ||
2178 | /* | |
2179 | * Calculate the group which is almost near its | |
2180 | * capacity but still has some space to pick up some load | |
2181 | * from other group and save more power | |
2182 | */ | |
2183 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2184 | return; | |
2185 | ||
2186 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2187 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2188 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2189 | sds->group_leader = group; | |
2190 | sds->leader_nr_running = sgs->sum_nr_running; | |
2191 | } | |
2192 | } | |
2193 | ||
2194 | /** | |
2195 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2196 | * @sds: Variable containing the statistics of the sched_domain | |
2197 | * under consideration. | |
2198 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2199 | * @imbalance: Variable to store the imbalance. | |
2200 | * | |
2201 | * Description: | |
2202 | * Check if we have potential to perform some power-savings balance. | |
2203 | * If yes, set the busiest group to be the least loaded group in the | |
2204 | * sched_domain, so that it's CPUs can be put to idle. | |
2205 | * | |
2206 | * Returns 1 if there is potential to perform power-savings balance. | |
2207 | * Else returns 0. | |
2208 | */ | |
2209 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2210 | int this_cpu, unsigned long *imbalance) | |
2211 | { | |
2212 | if (!sds->power_savings_balance) | |
2213 | return 0; | |
2214 | ||
2215 | if (sds->this != sds->group_leader || | |
2216 | sds->group_leader == sds->group_min) | |
2217 | return 0; | |
2218 | ||
2219 | *imbalance = sds->min_load_per_task; | |
2220 | sds->busiest = sds->group_min; | |
2221 | ||
2222 | return 1; | |
2223 | ||
2224 | } | |
2225 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2226 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2227 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2228 | { | |
2229 | return; | |
2230 | } | |
2231 | ||
2232 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2233 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2234 | { | |
2235 | return; | |
2236 | } | |
2237 | ||
2238 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2239 | int this_cpu, unsigned long *imbalance) | |
2240 | { | |
2241 | return 0; | |
2242 | } | |
2243 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2244 | ||
2245 | ||
2246 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2247 | { | |
2248 | return SCHED_LOAD_SCALE; | |
2249 | } | |
2250 | ||
2251 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2252 | { | |
2253 | return default_scale_freq_power(sd, cpu); | |
2254 | } | |
2255 | ||
2256 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2257 | { | |
669c55e9 | 2258 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2259 | unsigned long smt_gain = sd->smt_gain; |
2260 | ||
2261 | smt_gain /= weight; | |
2262 | ||
2263 | return smt_gain; | |
2264 | } | |
2265 | ||
2266 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2267 | { | |
2268 | return default_scale_smt_power(sd, cpu); | |
2269 | } | |
2270 | ||
2271 | unsigned long scale_rt_power(int cpu) | |
2272 | { | |
2273 | struct rq *rq = cpu_rq(cpu); | |
2274 | u64 total, available; | |
2275 | ||
1e3c88bd | 2276 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
2277 | |
2278 | if (unlikely(total < rq->rt_avg)) { | |
2279 | /* Ensures that power won't end up being negative */ | |
2280 | available = 0; | |
2281 | } else { | |
2282 | available = total - rq->rt_avg; | |
2283 | } | |
1e3c88bd PZ |
2284 | |
2285 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
2286 | total = SCHED_LOAD_SCALE; | |
2287 | ||
2288 | total >>= SCHED_LOAD_SHIFT; | |
2289 | ||
2290 | return div_u64(available, total); | |
2291 | } | |
2292 | ||
2293 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2294 | { | |
669c55e9 | 2295 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2296 | unsigned long power = SCHED_LOAD_SCALE; |
2297 | struct sched_group *sdg = sd->groups; | |
2298 | ||
1e3c88bd PZ |
2299 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
2300 | if (sched_feat(ARCH_POWER)) | |
2301 | power *= arch_scale_smt_power(sd, cpu); | |
2302 | else | |
2303 | power *= default_scale_smt_power(sd, cpu); | |
2304 | ||
2305 | power >>= SCHED_LOAD_SHIFT; | |
2306 | } | |
2307 | ||
9d5efe05 SV |
2308 | sdg->cpu_power_orig = power; |
2309 | ||
2310 | if (sched_feat(ARCH_POWER)) | |
2311 | power *= arch_scale_freq_power(sd, cpu); | |
2312 | else | |
2313 | power *= default_scale_freq_power(sd, cpu); | |
2314 | ||
2315 | power >>= SCHED_LOAD_SHIFT; | |
2316 | ||
1e3c88bd PZ |
2317 | power *= scale_rt_power(cpu); |
2318 | power >>= SCHED_LOAD_SHIFT; | |
2319 | ||
2320 | if (!power) | |
2321 | power = 1; | |
2322 | ||
e51fd5e2 | 2323 | cpu_rq(cpu)->cpu_power = power; |
1e3c88bd PZ |
2324 | sdg->cpu_power = power; |
2325 | } | |
2326 | ||
2327 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2328 | { | |
2329 | struct sched_domain *child = sd->child; | |
2330 | struct sched_group *group, *sdg = sd->groups; | |
2331 | unsigned long power; | |
2332 | ||
2333 | if (!child) { | |
2334 | update_cpu_power(sd, cpu); | |
2335 | return; | |
2336 | } | |
2337 | ||
2338 | power = 0; | |
2339 | ||
2340 | group = child->groups; | |
2341 | do { | |
2342 | power += group->cpu_power; | |
2343 | group = group->next; | |
2344 | } while (group != child->groups); | |
2345 | ||
2346 | sdg->cpu_power = power; | |
2347 | } | |
2348 | ||
9d5efe05 SV |
2349 | /* |
2350 | * Try and fix up capacity for tiny siblings, this is needed when | |
2351 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
2352 | * which on its own isn't powerful enough. | |
2353 | * | |
2354 | * See update_sd_pick_busiest() and check_asym_packing(). | |
2355 | */ | |
2356 | static inline int | |
2357 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
2358 | { | |
2359 | /* | |
2360 | * Only siblings can have significantly less than SCHED_LOAD_SCALE | |
2361 | */ | |
2362 | if (sd->level != SD_LV_SIBLING) | |
2363 | return 0; | |
2364 | ||
2365 | /* | |
2366 | * If ~90% of the cpu_power is still there, we're good. | |
2367 | */ | |
694f5a11 | 2368 | if (group->cpu_power * 32 > group->cpu_power_orig * 29) |
9d5efe05 SV |
2369 | return 1; |
2370 | ||
2371 | return 0; | |
2372 | } | |
2373 | ||
1e3c88bd PZ |
2374 | /** |
2375 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2376 | * @sd: The sched_domain whose statistics are to be updated. | |
2377 | * @group: sched_group whose statistics are to be updated. | |
2378 | * @this_cpu: Cpu for which load balance is currently performed. | |
2379 | * @idle: Idle status of this_cpu | |
2380 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
2381 | * @sd_idle: Idle status of the sched_domain containing group. | |
2382 | * @local_group: Does group contain this_cpu. | |
2383 | * @cpus: Set of cpus considered for load balancing. | |
2384 | * @balance: Should we balance. | |
2385 | * @sgs: variable to hold the statistics for this group. | |
2386 | */ | |
2387 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2388 | struct sched_group *group, int this_cpu, | |
2389 | enum cpu_idle_type idle, int load_idx, int *sd_idle, | |
2390 | int local_group, const struct cpumask *cpus, | |
2391 | int *balance, struct sg_lb_stats *sgs) | |
2392 | { | |
2582f0eb | 2393 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
2394 | int i; |
2395 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 2396 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 2397 | |
871e35bc | 2398 | if (local_group) |
1e3c88bd | 2399 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
2400 | |
2401 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
2402 | max_cpu_load = 0; |
2403 | min_cpu_load = ~0UL; | |
2582f0eb | 2404 | max_nr_running = 0; |
1e3c88bd PZ |
2405 | |
2406 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2407 | struct rq *rq = cpu_rq(i); | |
2408 | ||
2409 | if (*sd_idle && rq->nr_running) | |
2410 | *sd_idle = 0; | |
2411 | ||
2412 | /* Bias balancing toward cpus of our domain */ | |
2413 | if (local_group) { | |
2414 | if (idle_cpu(i) && !first_idle_cpu) { | |
2415 | first_idle_cpu = 1; | |
2416 | balance_cpu = i; | |
2417 | } | |
2418 | ||
2419 | load = target_load(i, load_idx); | |
2420 | } else { | |
2421 | load = source_load(i, load_idx); | |
2582f0eb | 2422 | if (load > max_cpu_load) { |
1e3c88bd | 2423 | max_cpu_load = load; |
2582f0eb NR |
2424 | max_nr_running = rq->nr_running; |
2425 | } | |
1e3c88bd PZ |
2426 | if (min_cpu_load > load) |
2427 | min_cpu_load = load; | |
2428 | } | |
2429 | ||
2430 | sgs->group_load += load; | |
2431 | sgs->sum_nr_running += rq->nr_running; | |
2432 | sgs->sum_weighted_load += weighted_cpuload(i); | |
aae6d3dd SS |
2433 | if (idle_cpu(i)) |
2434 | sgs->idle_cpus++; | |
1e3c88bd PZ |
2435 | } |
2436 | ||
2437 | /* | |
2438 | * First idle cpu or the first cpu(busiest) in this sched group | |
2439 | * is eligible for doing load balancing at this and above | |
2440 | * domains. In the newly idle case, we will allow all the cpu's | |
2441 | * to do the newly idle load balance. | |
2442 | */ | |
bbc8cb5b PZ |
2443 | if (idle != CPU_NEWLY_IDLE && local_group) { |
2444 | if (balance_cpu != this_cpu) { | |
2445 | *balance = 0; | |
2446 | return; | |
2447 | } | |
2448 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
2449 | } |
2450 | ||
2451 | /* Adjust by relative CPU power of the group */ | |
2452 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2453 | ||
1e3c88bd PZ |
2454 | /* |
2455 | * Consider the group unbalanced when the imbalance is larger | |
2456 | * than the average weight of two tasks. | |
2457 | * | |
2458 | * APZ: with cgroup the avg task weight can vary wildly and | |
2459 | * might not be a suitable number - should we keep a | |
2460 | * normalized nr_running number somewhere that negates | |
2461 | * the hierarchy? | |
2462 | */ | |
dd5feea1 SS |
2463 | if (sgs->sum_nr_running) |
2464 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 2465 | |
2582f0eb | 2466 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
2467 | sgs->group_imb = 1; |
2468 | ||
2582f0eb | 2469 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); |
9d5efe05 SV |
2470 | if (!sgs->group_capacity) |
2471 | sgs->group_capacity = fix_small_capacity(sd, group); | |
aae6d3dd | 2472 | sgs->group_weight = group->group_weight; |
fab47622 NR |
2473 | |
2474 | if (sgs->group_capacity > sgs->sum_nr_running) | |
2475 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
2476 | } |
2477 | ||
532cb4c4 MN |
2478 | /** |
2479 | * update_sd_pick_busiest - return 1 on busiest group | |
2480 | * @sd: sched_domain whose statistics are to be checked | |
2481 | * @sds: sched_domain statistics | |
2482 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
2483 | * @sgs: sched_group statistics |
2484 | * @this_cpu: the current cpu | |
532cb4c4 MN |
2485 | * |
2486 | * Determine if @sg is a busier group than the previously selected | |
2487 | * busiest group. | |
2488 | */ | |
2489 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
2490 | struct sd_lb_stats *sds, | |
2491 | struct sched_group *sg, | |
2492 | struct sg_lb_stats *sgs, | |
2493 | int this_cpu) | |
2494 | { | |
2495 | if (sgs->avg_load <= sds->max_load) | |
2496 | return false; | |
2497 | ||
2498 | if (sgs->sum_nr_running > sgs->group_capacity) | |
2499 | return true; | |
2500 | ||
2501 | if (sgs->group_imb) | |
2502 | return true; | |
2503 | ||
2504 | /* | |
2505 | * ASYM_PACKING needs to move all the work to the lowest | |
2506 | * numbered CPUs in the group, therefore mark all groups | |
2507 | * higher than ourself as busy. | |
2508 | */ | |
2509 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
2510 | this_cpu < group_first_cpu(sg)) { | |
2511 | if (!sds->busiest) | |
2512 | return true; | |
2513 | ||
2514 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
2515 | return true; | |
2516 | } | |
2517 | ||
2518 | return false; | |
2519 | } | |
2520 | ||
1e3c88bd PZ |
2521 | /** |
2522 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2523 | * @sd: sched_domain whose statistics are to be updated. | |
2524 | * @this_cpu: Cpu for which load balance is currently performed. | |
2525 | * @idle: Idle status of this_cpu | |
532cb4c4 | 2526 | * @sd_idle: Idle status of the sched_domain containing sg. |
1e3c88bd PZ |
2527 | * @cpus: Set of cpus considered for load balancing. |
2528 | * @balance: Should we balance. | |
2529 | * @sds: variable to hold the statistics for this sched_domain. | |
2530 | */ | |
2531 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
2532 | enum cpu_idle_type idle, int *sd_idle, | |
2533 | const struct cpumask *cpus, int *balance, | |
2534 | struct sd_lb_stats *sds) | |
2535 | { | |
2536 | struct sched_domain *child = sd->child; | |
532cb4c4 | 2537 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
2538 | struct sg_lb_stats sgs; |
2539 | int load_idx, prefer_sibling = 0; | |
2540 | ||
2541 | if (child && child->flags & SD_PREFER_SIBLING) | |
2542 | prefer_sibling = 1; | |
2543 | ||
2544 | init_sd_power_savings_stats(sd, sds, idle); | |
2545 | load_idx = get_sd_load_idx(sd, idle); | |
2546 | ||
2547 | do { | |
2548 | int local_group; | |
2549 | ||
532cb4c4 | 2550 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 2551 | memset(&sgs, 0, sizeof(sgs)); |
532cb4c4 | 2552 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle, |
1e3c88bd PZ |
2553 | local_group, cpus, balance, &sgs); |
2554 | ||
8f190fb3 | 2555 | if (local_group && !(*balance)) |
1e3c88bd PZ |
2556 | return; |
2557 | ||
2558 | sds->total_load += sgs.group_load; | |
532cb4c4 | 2559 | sds->total_pwr += sg->cpu_power; |
1e3c88bd PZ |
2560 | |
2561 | /* | |
2562 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 2563 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
2564 | * and move all the excess tasks away. We lower the capacity |
2565 | * of a group only if the local group has the capacity to fit | |
2566 | * these excess tasks, i.e. nr_running < group_capacity. The | |
2567 | * extra check prevents the case where you always pull from the | |
2568 | * heaviest group when it is already under-utilized (possible | |
2569 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 2570 | */ |
75dd321d | 2571 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
2572 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
2573 | ||
2574 | if (local_group) { | |
2575 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 2576 | sds->this = sg; |
1e3c88bd PZ |
2577 | sds->this_nr_running = sgs.sum_nr_running; |
2578 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 2579 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 2580 | sds->this_idle_cpus = sgs.idle_cpus; |
532cb4c4 | 2581 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 2582 | sds->max_load = sgs.avg_load; |
532cb4c4 | 2583 | sds->busiest = sg; |
1e3c88bd | 2584 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 2585 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 2586 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 2587 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 2588 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 2589 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
2590 | sds->group_imb = sgs.group_imb; |
2591 | } | |
2592 | ||
532cb4c4 MN |
2593 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
2594 | sg = sg->next; | |
2595 | } while (sg != sd->groups); | |
2596 | } | |
2597 | ||
2ec57d44 | 2598 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
2599 | { |
2600 | return 0*SD_ASYM_PACKING; | |
2601 | } | |
2602 | ||
2603 | /** | |
2604 | * check_asym_packing - Check to see if the group is packed into the | |
2605 | * sched doman. | |
2606 | * | |
2607 | * This is primarily intended to used at the sibling level. Some | |
2608 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
2609 | * case of POWER7, it can move to lower SMT modes only when higher | |
2610 | * threads are idle. When in lower SMT modes, the threads will | |
2611 | * perform better since they share less core resources. Hence when we | |
2612 | * have idle threads, we want them to be the higher ones. | |
2613 | * | |
2614 | * This packing function is run on idle threads. It checks to see if | |
2615 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
2616 | * CPU number than the packing function is being run on. Here we are | |
2617 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
2618 | * number. | |
2619 | * | |
b6b12294 MN |
2620 | * Returns 1 when packing is required and a task should be moved to |
2621 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
2622 | * | |
532cb4c4 MN |
2623 | * @sd: The sched_domain whose packing is to be checked. |
2624 | * @sds: Statistics of the sched_domain which is to be packed | |
2625 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2626 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
2627 | */ |
2628 | static int check_asym_packing(struct sched_domain *sd, | |
2629 | struct sd_lb_stats *sds, | |
2630 | int this_cpu, unsigned long *imbalance) | |
2631 | { | |
2632 | int busiest_cpu; | |
2633 | ||
2634 | if (!(sd->flags & SD_ASYM_PACKING)) | |
2635 | return 0; | |
2636 | ||
2637 | if (!sds->busiest) | |
2638 | return 0; | |
2639 | ||
2640 | busiest_cpu = group_first_cpu(sds->busiest); | |
2641 | if (this_cpu > busiest_cpu) | |
2642 | return 0; | |
2643 | ||
2644 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power, | |
2645 | SCHED_LOAD_SCALE); | |
2646 | return 1; | |
1e3c88bd PZ |
2647 | } |
2648 | ||
2649 | /** | |
2650 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2651 | * amongst the groups of a sched_domain, during | |
2652 | * load balancing. | |
2653 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2654 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2655 | * @imbalance: Variable to store the imbalance. | |
2656 | */ | |
2657 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2658 | int this_cpu, unsigned long *imbalance) | |
2659 | { | |
2660 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2661 | unsigned int imbn = 2; | |
dd5feea1 | 2662 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
2663 | |
2664 | if (sds->this_nr_running) { | |
2665 | sds->this_load_per_task /= sds->this_nr_running; | |
2666 | if (sds->busiest_load_per_task > | |
2667 | sds->this_load_per_task) | |
2668 | imbn = 1; | |
2669 | } else | |
2670 | sds->this_load_per_task = | |
2671 | cpu_avg_load_per_task(this_cpu); | |
2672 | ||
dd5feea1 SS |
2673 | scaled_busy_load_per_task = sds->busiest_load_per_task |
2674 | * SCHED_LOAD_SCALE; | |
2675 | scaled_busy_load_per_task /= sds->busiest->cpu_power; | |
2676 | ||
2677 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
2678 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
2679 | *imbalance = sds->busiest_load_per_task; |
2680 | return; | |
2681 | } | |
2682 | ||
2683 | /* | |
2684 | * OK, we don't have enough imbalance to justify moving tasks, | |
2685 | * however we may be able to increase total CPU power used by | |
2686 | * moving them. | |
2687 | */ | |
2688 | ||
2689 | pwr_now += sds->busiest->cpu_power * | |
2690 | min(sds->busiest_load_per_task, sds->max_load); | |
2691 | pwr_now += sds->this->cpu_power * | |
2692 | min(sds->this_load_per_task, sds->this_load); | |
2693 | pwr_now /= SCHED_LOAD_SCALE; | |
2694 | ||
2695 | /* Amount of load we'd subtract */ | |
2696 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2697 | sds->busiest->cpu_power; | |
2698 | if (sds->max_load > tmp) | |
2699 | pwr_move += sds->busiest->cpu_power * | |
2700 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
2701 | ||
2702 | /* Amount of load we'd add */ | |
2703 | if (sds->max_load * sds->busiest->cpu_power < | |
2704 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
2705 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
2706 | sds->this->cpu_power; | |
2707 | else | |
2708 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2709 | sds->this->cpu_power; | |
2710 | pwr_move += sds->this->cpu_power * | |
2711 | min(sds->this_load_per_task, sds->this_load + tmp); | |
2712 | pwr_move /= SCHED_LOAD_SCALE; | |
2713 | ||
2714 | /* Move if we gain throughput */ | |
2715 | if (pwr_move > pwr_now) | |
2716 | *imbalance = sds->busiest_load_per_task; | |
2717 | } | |
2718 | ||
2719 | /** | |
2720 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
2721 | * groups of a given sched_domain during load balance. | |
2722 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
2723 | * @this_cpu: Cpu for which currently load balance is being performed. | |
2724 | * @imbalance: The variable to store the imbalance. | |
2725 | */ | |
2726 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
2727 | unsigned long *imbalance) | |
2728 | { | |
dd5feea1 SS |
2729 | unsigned long max_pull, load_above_capacity = ~0UL; |
2730 | ||
2731 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
2732 | if (sds->group_imb) { | |
2733 | sds->busiest_load_per_task = | |
2734 | min(sds->busiest_load_per_task, sds->avg_load); | |
2735 | } | |
2736 | ||
1e3c88bd PZ |
2737 | /* |
2738 | * In the presence of smp nice balancing, certain scenarios can have | |
2739 | * max load less than avg load(as we skip the groups at or below | |
2740 | * its cpu_power, while calculating max_load..) | |
2741 | */ | |
2742 | if (sds->max_load < sds->avg_load) { | |
2743 | *imbalance = 0; | |
2744 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2745 | } | |
2746 | ||
dd5feea1 SS |
2747 | if (!sds->group_imb) { |
2748 | /* | |
2749 | * Don't want to pull so many tasks that a group would go idle. | |
2750 | */ | |
2751 | load_above_capacity = (sds->busiest_nr_running - | |
2752 | sds->busiest_group_capacity); | |
2753 | ||
2754 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE); | |
2755 | ||
2756 | load_above_capacity /= sds->busiest->cpu_power; | |
2757 | } | |
2758 | ||
2759 | /* | |
2760 | * We're trying to get all the cpus to the average_load, so we don't | |
2761 | * want to push ourselves above the average load, nor do we wish to | |
2762 | * reduce the max loaded cpu below the average load. At the same time, | |
2763 | * we also don't want to reduce the group load below the group capacity | |
2764 | * (so that we can implement power-savings policies etc). Thus we look | |
2765 | * for the minimum possible imbalance. | |
2766 | * Be careful of negative numbers as they'll appear as very large values | |
2767 | * with unsigned longs. | |
2768 | */ | |
2769 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
2770 | |
2771 | /* How much load to actually move to equalise the imbalance */ | |
2772 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
2773 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
2774 | / SCHED_LOAD_SCALE; | |
2775 | ||
2776 | /* | |
2777 | * if *imbalance is less than the average load per runnable task | |
2778 | * there is no gaurantee that any tasks will be moved so we'll have | |
2779 | * a think about bumping its value to force at least one task to be | |
2780 | * moved | |
2781 | */ | |
2782 | if (*imbalance < sds->busiest_load_per_task) | |
2783 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2784 | ||
2785 | } | |
fab47622 | 2786 | |
1e3c88bd PZ |
2787 | /******* find_busiest_group() helpers end here *********************/ |
2788 | ||
2789 | /** | |
2790 | * find_busiest_group - Returns the busiest group within the sched_domain | |
2791 | * if there is an imbalance. If there isn't an imbalance, and | |
2792 | * the user has opted for power-savings, it returns a group whose | |
2793 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
2794 | * such a group exists. | |
2795 | * | |
2796 | * Also calculates the amount of weighted load which should be moved | |
2797 | * to restore balance. | |
2798 | * | |
2799 | * @sd: The sched_domain whose busiest group is to be returned. | |
2800 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
2801 | * @imbalance: Variable which stores amount of weighted load which should | |
2802 | * be moved to restore balance/put a group to idle. | |
2803 | * @idle: The idle status of this_cpu. | |
2804 | * @sd_idle: The idleness of sd | |
2805 | * @cpus: The set of CPUs under consideration for load-balancing. | |
2806 | * @balance: Pointer to a variable indicating if this_cpu | |
2807 | * is the appropriate cpu to perform load balancing at this_level. | |
2808 | * | |
2809 | * Returns: - the busiest group if imbalance exists. | |
2810 | * - If no imbalance and user has opted for power-savings balance, | |
2811 | * return the least loaded group whose CPUs can be | |
2812 | * put to idle by rebalancing its tasks onto our group. | |
2813 | */ | |
2814 | static struct sched_group * | |
2815 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
2816 | unsigned long *imbalance, enum cpu_idle_type idle, | |
2817 | int *sd_idle, const struct cpumask *cpus, int *balance) | |
2818 | { | |
2819 | struct sd_lb_stats sds; | |
2820 | ||
2821 | memset(&sds, 0, sizeof(sds)); | |
2822 | ||
2823 | /* | |
2824 | * Compute the various statistics relavent for load balancing at | |
2825 | * this level. | |
2826 | */ | |
2827 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, | |
2828 | balance, &sds); | |
2829 | ||
2830 | /* Cases where imbalance does not exist from POV of this_cpu */ | |
2831 | /* 1) this_cpu is not the appropriate cpu to perform load balancing | |
2832 | * at this level. | |
2833 | * 2) There is no busy sibling group to pull from. | |
2834 | * 3) This group is the busiest group. | |
2835 | * 4) This group is more busy than the avg busieness at this | |
2836 | * sched_domain. | |
2837 | * 5) The imbalance is within the specified limit. | |
fab47622 NR |
2838 | * |
2839 | * Note: when doing newidle balance, if the local group has excess | |
2840 | * capacity (i.e. nr_running < group_capacity) and the busiest group | |
2841 | * does not have any capacity, we force a load balance to pull tasks | |
2842 | * to the local group. In this case, we skip past checks 3, 4 and 5. | |
1e3c88bd | 2843 | */ |
8f190fb3 | 2844 | if (!(*balance)) |
1e3c88bd PZ |
2845 | goto ret; |
2846 | ||
532cb4c4 MN |
2847 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
2848 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
2849 | return sds.busiest; | |
2850 | ||
1e3c88bd PZ |
2851 | if (!sds.busiest || sds.busiest_nr_running == 0) |
2852 | goto out_balanced; | |
2853 | ||
fab47622 NR |
2854 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
2855 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && | |
2856 | !sds.busiest_has_capacity) | |
2857 | goto force_balance; | |
2858 | ||
1e3c88bd PZ |
2859 | if (sds.this_load >= sds.max_load) |
2860 | goto out_balanced; | |
2861 | ||
2862 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; | |
2863 | ||
2864 | if (sds.this_load >= sds.avg_load) | |
2865 | goto out_balanced; | |
2866 | ||
aae6d3dd SS |
2867 | /* |
2868 | * In the CPU_NEWLY_IDLE, use imbalance_pct to be conservative. | |
2869 | * And to check for busy balance use !idle_cpu instead of | |
2870 | * CPU_NOT_IDLE. This is because HT siblings will use CPU_NOT_IDLE | |
2871 | * even when they are idle. | |
2872 | */ | |
2873 | if (idle == CPU_NEWLY_IDLE || !idle_cpu(this_cpu)) { | |
2874 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
2875 | goto out_balanced; | |
2876 | } else { | |
2877 | /* | |
2878 | * This cpu is idle. If the busiest group load doesn't | |
2879 | * have more tasks than the number of available cpu's and | |
2880 | * there is no imbalance between this and busiest group | |
2881 | * wrt to idle cpu's, it is balanced. | |
2882 | */ | |
2883 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && | |
2884 | sds.busiest_nr_running <= sds.busiest_group_weight) | |
2885 | goto out_balanced; | |
2886 | } | |
1e3c88bd | 2887 | |
fab47622 | 2888 | force_balance: |
1e3c88bd PZ |
2889 | /* Looks like there is an imbalance. Compute it */ |
2890 | calculate_imbalance(&sds, this_cpu, imbalance); | |
2891 | return sds.busiest; | |
2892 | ||
2893 | out_balanced: | |
2894 | /* | |
2895 | * There is no obvious imbalance. But check if we can do some balancing | |
2896 | * to save power. | |
2897 | */ | |
2898 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
2899 | return sds.busiest; | |
2900 | ret: | |
2901 | *imbalance = 0; | |
2902 | return NULL; | |
2903 | } | |
2904 | ||
2905 | /* | |
2906 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
2907 | */ | |
2908 | static struct rq * | |
9d5efe05 SV |
2909 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
2910 | enum cpu_idle_type idle, unsigned long imbalance, | |
2911 | const struct cpumask *cpus) | |
1e3c88bd PZ |
2912 | { |
2913 | struct rq *busiest = NULL, *rq; | |
2914 | unsigned long max_load = 0; | |
2915 | int i; | |
2916 | ||
2917 | for_each_cpu(i, sched_group_cpus(group)) { | |
2918 | unsigned long power = power_of(i); | |
2919 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
2920 | unsigned long wl; | |
2921 | ||
9d5efe05 SV |
2922 | if (!capacity) |
2923 | capacity = fix_small_capacity(sd, group); | |
2924 | ||
1e3c88bd PZ |
2925 | if (!cpumask_test_cpu(i, cpus)) |
2926 | continue; | |
2927 | ||
2928 | rq = cpu_rq(i); | |
6e40f5bb | 2929 | wl = weighted_cpuload(i); |
1e3c88bd | 2930 | |
6e40f5bb TG |
2931 | /* |
2932 | * When comparing with imbalance, use weighted_cpuload() | |
2933 | * which is not scaled with the cpu power. | |
2934 | */ | |
1e3c88bd PZ |
2935 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
2936 | continue; | |
2937 | ||
6e40f5bb TG |
2938 | /* |
2939 | * For the load comparisons with the other cpu's, consider | |
2940 | * the weighted_cpuload() scaled with the cpu power, so that | |
2941 | * the load can be moved away from the cpu that is potentially | |
2942 | * running at a lower capacity. | |
2943 | */ | |
2944 | wl = (wl * SCHED_LOAD_SCALE) / power; | |
2945 | ||
1e3c88bd PZ |
2946 | if (wl > max_load) { |
2947 | max_load = wl; | |
2948 | busiest = rq; | |
2949 | } | |
2950 | } | |
2951 | ||
2952 | return busiest; | |
2953 | } | |
2954 | ||
2955 | /* | |
2956 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
2957 | * so long as it is large enough. | |
2958 | */ | |
2959 | #define MAX_PINNED_INTERVAL 512 | |
2960 | ||
2961 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
2962 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
2963 | ||
532cb4c4 MN |
2964 | static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle, |
2965 | int busiest_cpu, int this_cpu) | |
1af3ed3d PZ |
2966 | { |
2967 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
2968 | |
2969 | /* | |
2970 | * ASYM_PACKING needs to force migrate tasks from busy but | |
2971 | * higher numbered CPUs in order to pack all tasks in the | |
2972 | * lowest numbered CPUs. | |
2973 | */ | |
2974 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
2975 | return 1; | |
2976 | ||
1af3ed3d PZ |
2977 | /* |
2978 | * The only task running in a non-idle cpu can be moved to this | |
2979 | * cpu in an attempt to completely freeup the other CPU | |
2980 | * package. | |
2981 | * | |
2982 | * The package power saving logic comes from | |
2983 | * find_busiest_group(). If there are no imbalance, then | |
2984 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
2985 | * f_b_g() will select a group from which a running task may be | |
2986 | * pulled to this cpu in order to make the other package idle. | |
2987 | * If there is no opportunity to make a package idle and if | |
2988 | * there are no imbalance, then f_b_g() will return NULL and no | |
2989 | * action will be taken in load_balance_newidle(). | |
2990 | * | |
2991 | * Under normal task pull operation due to imbalance, there | |
2992 | * will be more than one task in the source run queue and | |
2993 | * move_tasks() will succeed. ld_moved will be true and this | |
2994 | * active balance code will not be triggered. | |
2995 | */ | |
2996 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
2997 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
2998 | return 0; | |
2999 | ||
3000 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) | |
3001 | return 0; | |
3002 | } | |
3003 | ||
3004 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3005 | } | |
3006 | ||
969c7921 TH |
3007 | static int active_load_balance_cpu_stop(void *data); |
3008 | ||
1e3c88bd PZ |
3009 | /* |
3010 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3011 | * tasks if there is an imbalance. | |
3012 | */ | |
3013 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3014 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3015 | int *balance) | |
3016 | { | |
3017 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
3018 | struct sched_group *group; | |
3019 | unsigned long imbalance; | |
3020 | struct rq *busiest; | |
3021 | unsigned long flags; | |
3022 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3023 | ||
3024 | cpumask_copy(cpus, cpu_active_mask); | |
3025 | ||
3026 | /* | |
3027 | * When power savings policy is enabled for the parent domain, idle | |
3028 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3029 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
3030 | * portraying it as CPU_NOT_IDLE. | |
3031 | */ | |
3032 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
3033 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3034 | sd_idle = 1; | |
3035 | ||
3036 | schedstat_inc(sd, lb_count[idle]); | |
3037 | ||
3038 | redo: | |
3039 | update_shares(sd); | |
3040 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
3041 | cpus, balance); | |
3042 | ||
3043 | if (*balance == 0) | |
3044 | goto out_balanced; | |
3045 | ||
3046 | if (!group) { | |
3047 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3048 | goto out_balanced; | |
3049 | } | |
3050 | ||
9d5efe05 | 3051 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3052 | if (!busiest) { |
3053 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3054 | goto out_balanced; | |
3055 | } | |
3056 | ||
3057 | BUG_ON(busiest == this_rq); | |
3058 | ||
3059 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3060 | ||
3061 | ld_moved = 0; | |
3062 | if (busiest->nr_running > 1) { | |
3063 | /* | |
3064 | * Attempt to move tasks. If find_busiest_group has found | |
3065 | * an imbalance but busiest->nr_running <= 1, the group is | |
3066 | * still unbalanced. ld_moved simply stays zero, so it is | |
3067 | * correctly treated as an imbalance. | |
3068 | */ | |
3069 | local_irq_save(flags); | |
3070 | double_rq_lock(this_rq, busiest); | |
3071 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3072 | imbalance, sd, idle, &all_pinned); | |
3073 | double_rq_unlock(this_rq, busiest); | |
3074 | local_irq_restore(flags); | |
3075 | ||
3076 | /* | |
3077 | * some other cpu did the load balance for us. | |
3078 | */ | |
3079 | if (ld_moved && this_cpu != smp_processor_id()) | |
3080 | resched_cpu(this_cpu); | |
3081 | ||
3082 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3083 | if (unlikely(all_pinned)) { | |
3084 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3085 | if (!cpumask_empty(cpus)) | |
3086 | goto redo; | |
3087 | goto out_balanced; | |
3088 | } | |
3089 | } | |
3090 | ||
3091 | if (!ld_moved) { | |
3092 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3093 | /* |
3094 | * Increment the failure counter only on periodic balance. | |
3095 | * We do not want newidle balance, which can be very | |
3096 | * frequent, pollute the failure counter causing | |
3097 | * excessive cache_hot migrations and active balances. | |
3098 | */ | |
3099 | if (idle != CPU_NEWLY_IDLE) | |
3100 | sd->nr_balance_failed++; | |
1e3c88bd | 3101 | |
532cb4c4 MN |
3102 | if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest), |
3103 | this_cpu)) { | |
1e3c88bd PZ |
3104 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3105 | ||
969c7921 TH |
3106 | /* don't kick the active_load_balance_cpu_stop, |
3107 | * if the curr task on busiest cpu can't be | |
3108 | * moved to this_cpu | |
1e3c88bd PZ |
3109 | */ |
3110 | if (!cpumask_test_cpu(this_cpu, | |
3111 | &busiest->curr->cpus_allowed)) { | |
3112 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3113 | flags); | |
3114 | all_pinned = 1; | |
3115 | goto out_one_pinned; | |
3116 | } | |
3117 | ||
969c7921 TH |
3118 | /* |
3119 | * ->active_balance synchronizes accesses to | |
3120 | * ->active_balance_work. Once set, it's cleared | |
3121 | * only after active load balance is finished. | |
3122 | */ | |
1e3c88bd PZ |
3123 | if (!busiest->active_balance) { |
3124 | busiest->active_balance = 1; | |
3125 | busiest->push_cpu = this_cpu; | |
3126 | active_balance = 1; | |
3127 | } | |
3128 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 3129 | |
1e3c88bd | 3130 | if (active_balance) |
969c7921 TH |
3131 | stop_one_cpu_nowait(cpu_of(busiest), |
3132 | active_load_balance_cpu_stop, busiest, | |
3133 | &busiest->active_balance_work); | |
1e3c88bd PZ |
3134 | |
3135 | /* | |
3136 | * We've kicked active balancing, reset the failure | |
3137 | * counter. | |
3138 | */ | |
3139 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3140 | } | |
3141 | } else | |
3142 | sd->nr_balance_failed = 0; | |
3143 | ||
3144 | if (likely(!active_balance)) { | |
3145 | /* We were unbalanced, so reset the balancing interval */ | |
3146 | sd->balance_interval = sd->min_interval; | |
3147 | } else { | |
3148 | /* | |
3149 | * If we've begun active balancing, start to back off. This | |
3150 | * case may not be covered by the all_pinned logic if there | |
3151 | * is only 1 task on the busy runqueue (because we don't call | |
3152 | * move_tasks). | |
3153 | */ | |
3154 | if (sd->balance_interval < sd->max_interval) | |
3155 | sd->balance_interval *= 2; | |
3156 | } | |
3157 | ||
3158 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3159 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3160 | ld_moved = -1; | |
3161 | ||
3162 | goto out; | |
3163 | ||
3164 | out_balanced: | |
3165 | schedstat_inc(sd, lb_balanced[idle]); | |
3166 | ||
3167 | sd->nr_balance_failed = 0; | |
3168 | ||
3169 | out_one_pinned: | |
3170 | /* tune up the balancing interval */ | |
3171 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3172 | (sd->balance_interval < sd->max_interval)) | |
3173 | sd->balance_interval *= 2; | |
3174 | ||
3175 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3176 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3177 | ld_moved = -1; | |
3178 | else | |
3179 | ld_moved = 0; | |
3180 | out: | |
3181 | if (ld_moved) | |
3182 | update_shares(sd); | |
3183 | return ld_moved; | |
3184 | } | |
3185 | ||
1e3c88bd PZ |
3186 | /* |
3187 | * idle_balance is called by schedule() if this_cpu is about to become | |
3188 | * idle. Attempts to pull tasks from other CPUs. | |
3189 | */ | |
3190 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3191 | { | |
3192 | struct sched_domain *sd; | |
3193 | int pulled_task = 0; | |
3194 | unsigned long next_balance = jiffies + HZ; | |
3195 | ||
3196 | this_rq->idle_stamp = this_rq->clock; | |
3197 | ||
3198 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3199 | return; | |
3200 | ||
f492e12e PZ |
3201 | /* |
3202 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
3203 | */ | |
3204 | raw_spin_unlock(&this_rq->lock); | |
3205 | ||
1e3c88bd PZ |
3206 | for_each_domain(this_cpu, sd) { |
3207 | unsigned long interval; | |
f492e12e | 3208 | int balance = 1; |
1e3c88bd PZ |
3209 | |
3210 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3211 | continue; | |
3212 | ||
f492e12e | 3213 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 3214 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
3215 | pulled_task = load_balance(this_cpu, this_rq, |
3216 | sd, CPU_NEWLY_IDLE, &balance); | |
3217 | } | |
1e3c88bd PZ |
3218 | |
3219 | interval = msecs_to_jiffies(sd->balance_interval); | |
3220 | if (time_after(next_balance, sd->last_balance + interval)) | |
3221 | next_balance = sd->last_balance + interval; | |
fab47622 | 3222 | if (pulled_task) |
1e3c88bd | 3223 | break; |
1e3c88bd | 3224 | } |
f492e12e PZ |
3225 | |
3226 | raw_spin_lock(&this_rq->lock); | |
3227 | ||
1e3c88bd PZ |
3228 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
3229 | /* | |
3230 | * We are going idle. next_balance may be set based on | |
3231 | * a busy processor. So reset next_balance. | |
3232 | */ | |
3233 | this_rq->next_balance = next_balance; | |
3234 | } | |
3235 | } | |
3236 | ||
3237 | /* | |
969c7921 TH |
3238 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
3239 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
3240 | * least 1 task to be running on each physical CPU where possible, and | |
3241 | * avoids physical / logical imbalances. | |
1e3c88bd | 3242 | */ |
969c7921 | 3243 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 3244 | { |
969c7921 TH |
3245 | struct rq *busiest_rq = data; |
3246 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 3247 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 3248 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 3249 | struct sched_domain *sd; |
969c7921 TH |
3250 | |
3251 | raw_spin_lock_irq(&busiest_rq->lock); | |
3252 | ||
3253 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
3254 | if (unlikely(busiest_cpu != smp_processor_id() || | |
3255 | !busiest_rq->active_balance)) | |
3256 | goto out_unlock; | |
1e3c88bd PZ |
3257 | |
3258 | /* Is there any task to move? */ | |
3259 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 3260 | goto out_unlock; |
1e3c88bd PZ |
3261 | |
3262 | /* | |
3263 | * This condition is "impossible", if it occurs | |
3264 | * we need to fix it. Originally reported by | |
3265 | * Bjorn Helgaas on a 128-cpu setup. | |
3266 | */ | |
3267 | BUG_ON(busiest_rq == target_rq); | |
3268 | ||
3269 | /* move a task from busiest_rq to target_rq */ | |
3270 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
3271 | |
3272 | /* Search for an sd spanning us and the target CPU. */ | |
3273 | for_each_domain(target_cpu, sd) { | |
3274 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3275 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3276 | break; | |
3277 | } | |
3278 | ||
3279 | if (likely(sd)) { | |
3280 | schedstat_inc(sd, alb_count); | |
3281 | ||
3282 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3283 | sd, CPU_IDLE)) | |
3284 | schedstat_inc(sd, alb_pushed); | |
3285 | else | |
3286 | schedstat_inc(sd, alb_failed); | |
3287 | } | |
3288 | double_unlock_balance(busiest_rq, target_rq); | |
969c7921 TH |
3289 | out_unlock: |
3290 | busiest_rq->active_balance = 0; | |
3291 | raw_spin_unlock_irq(&busiest_rq->lock); | |
3292 | return 0; | |
1e3c88bd PZ |
3293 | } |
3294 | ||
3295 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
3296 | |
3297 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
3298 | ||
3299 | static void trigger_sched_softirq(void *data) | |
3300 | { | |
3301 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
3302 | } | |
3303 | ||
3304 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
3305 | { | |
3306 | csd->func = trigger_sched_softirq; | |
3307 | csd->info = NULL; | |
3308 | csd->flags = 0; | |
3309 | csd->priv = 0; | |
3310 | } | |
3311 | ||
3312 | /* | |
3313 | * idle load balancing details | |
3314 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
3315 | * entering idle. | |
3316 | * - This idle load balancer CPU will also go into tickless mode when | |
3317 | * it is idle, just like all other idle CPUs | |
3318 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
3319 | * needed, they will kick the idle load balancer, which then does idle | |
3320 | * load balancing for all the idle CPUs. | |
3321 | */ | |
1e3c88bd PZ |
3322 | static struct { |
3323 | atomic_t load_balancer; | |
83cd4fe2 VP |
3324 | atomic_t first_pick_cpu; |
3325 | atomic_t second_pick_cpu; | |
3326 | cpumask_var_t idle_cpus_mask; | |
3327 | cpumask_var_t grp_idle_mask; | |
3328 | unsigned long next_balance; /* in jiffy units */ | |
3329 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
3330 | |
3331 | int get_nohz_load_balancer(void) | |
3332 | { | |
3333 | return atomic_read(&nohz.load_balancer); | |
3334 | } | |
3335 | ||
3336 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3337 | /** | |
3338 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3339 | * @cpu: The cpu whose lowest level of sched domain is to | |
3340 | * be returned. | |
3341 | * @flag: The flag to check for the lowest sched_domain | |
3342 | * for the given cpu. | |
3343 | * | |
3344 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3345 | */ | |
3346 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3347 | { | |
3348 | struct sched_domain *sd; | |
3349 | ||
3350 | for_each_domain(cpu, sd) | |
3351 | if (sd && (sd->flags & flag)) | |
3352 | break; | |
3353 | ||
3354 | return sd; | |
3355 | } | |
3356 | ||
3357 | /** | |
3358 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3359 | * @cpu: The cpu whose domains we're iterating over. | |
3360 | * @sd: variable holding the value of the power_savings_sd | |
3361 | * for cpu. | |
3362 | * @flag: The flag to filter the sched_domains to be iterated. | |
3363 | * | |
3364 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3365 | * set, starting from the lowest sched_domain to the highest. | |
3366 | */ | |
3367 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3368 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3369 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3370 | ||
3371 | /** | |
3372 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3373 | * @ilb_group: group to be checked for semi-idleness | |
3374 | * | |
3375 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3376 | * | |
3377 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3378 | * and atleast one non-idle CPU. This helper function checks if the given | |
3379 | * sched_group is semi-idle or not. | |
3380 | */ | |
3381 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3382 | { | |
83cd4fe2 | 3383 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
3384 | sched_group_cpus(ilb_group)); |
3385 | ||
3386 | /* | |
3387 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3388 | * and atleast one idle cpu. | |
3389 | */ | |
83cd4fe2 | 3390 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
3391 | return 0; |
3392 | ||
83cd4fe2 | 3393 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
3394 | return 0; |
3395 | ||
3396 | return 1; | |
3397 | } | |
3398 | /** | |
3399 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3400 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3401 | * | |
3402 | * Returns: Returns the id of the idle load balancer if it exists, | |
3403 | * Else, returns >= nr_cpu_ids. | |
3404 | * | |
3405 | * This algorithm picks the idle load balancer such that it belongs to a | |
3406 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3407 | * completely idle packages/cores just for the purpose of idle load balancing | |
3408 | * when there are other idle cpu's which are better suited for that job. | |
3409 | */ | |
3410 | static int find_new_ilb(int cpu) | |
3411 | { | |
3412 | struct sched_domain *sd; | |
3413 | struct sched_group *ilb_group; | |
3414 | ||
3415 | /* | |
3416 | * Have idle load balancer selection from semi-idle packages only | |
3417 | * when power-aware load balancing is enabled | |
3418 | */ | |
3419 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3420 | goto out_done; | |
3421 | ||
3422 | /* | |
3423 | * Optimize for the case when we have no idle CPUs or only one | |
3424 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3425 | */ | |
83cd4fe2 | 3426 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
3427 | goto out_done; |
3428 | ||
3429 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
3430 | ilb_group = sd->groups; | |
3431 | ||
3432 | do { | |
3433 | if (is_semi_idle_group(ilb_group)) | |
83cd4fe2 | 3434 | return cpumask_first(nohz.grp_idle_mask); |
1e3c88bd PZ |
3435 | |
3436 | ilb_group = ilb_group->next; | |
3437 | ||
3438 | } while (ilb_group != sd->groups); | |
3439 | } | |
3440 | ||
3441 | out_done: | |
83cd4fe2 | 3442 | return nr_cpu_ids; |
1e3c88bd PZ |
3443 | } |
3444 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3445 | static inline int find_new_ilb(int call_cpu) | |
3446 | { | |
83cd4fe2 | 3447 | return nr_cpu_ids; |
1e3c88bd PZ |
3448 | } |
3449 | #endif | |
3450 | ||
83cd4fe2 VP |
3451 | /* |
3452 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
3453 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
3454 | * CPU (if there is one). | |
3455 | */ | |
3456 | static void nohz_balancer_kick(int cpu) | |
3457 | { | |
3458 | int ilb_cpu; | |
3459 | ||
3460 | nohz.next_balance++; | |
3461 | ||
3462 | ilb_cpu = get_nohz_load_balancer(); | |
3463 | ||
3464 | if (ilb_cpu >= nr_cpu_ids) { | |
3465 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
3466 | if (ilb_cpu >= nr_cpu_ids) | |
3467 | return; | |
3468 | } | |
3469 | ||
3470 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
3471 | struct call_single_data *cp; | |
3472 | ||
3473 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
3474 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
3475 | __smp_call_function_single(ilb_cpu, cp, 0); | |
3476 | } | |
3477 | return; | |
3478 | } | |
3479 | ||
1e3c88bd PZ |
3480 | /* |
3481 | * This routine will try to nominate the ilb (idle load balancing) | |
3482 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 3483 | * load balancing on behalf of all those cpus. |
1e3c88bd | 3484 | * |
83cd4fe2 VP |
3485 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
3486 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
3487 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 3488 | * |
83cd4fe2 VP |
3489 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
3490 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
3491 | * behalf of all idle CPUs). | |
1e3c88bd | 3492 | */ |
83cd4fe2 | 3493 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
3494 | { |
3495 | int cpu = smp_processor_id(); | |
3496 | ||
3497 | if (stop_tick) { | |
1e3c88bd PZ |
3498 | if (!cpu_active(cpu)) { |
3499 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 3500 | return; |
1e3c88bd PZ |
3501 | |
3502 | /* | |
3503 | * If we are going offline and still the leader, | |
3504 | * give up! | |
3505 | */ | |
83cd4fe2 VP |
3506 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3507 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3508 | BUG(); |
3509 | ||
83cd4fe2 | 3510 | return; |
1e3c88bd PZ |
3511 | } |
3512 | ||
83cd4fe2 | 3513 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 3514 | |
83cd4fe2 VP |
3515 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
3516 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
3517 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
3518 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 3519 | |
83cd4fe2 | 3520 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
3521 | int new_ilb; |
3522 | ||
83cd4fe2 VP |
3523 | /* make me the ilb owner */ |
3524 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
3525 | cpu) != nr_cpu_ids) | |
3526 | return; | |
3527 | ||
1e3c88bd PZ |
3528 | /* |
3529 | * Check to see if there is a more power-efficient | |
3530 | * ilb. | |
3531 | */ | |
3532 | new_ilb = find_new_ilb(cpu); | |
3533 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 3534 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 3535 | resched_cpu(new_ilb); |
83cd4fe2 | 3536 | return; |
1e3c88bd | 3537 | } |
83cd4fe2 | 3538 | return; |
1e3c88bd PZ |
3539 | } |
3540 | } else { | |
83cd4fe2 VP |
3541 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
3542 | return; | |
1e3c88bd | 3543 | |
83cd4fe2 | 3544 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
3545 | |
3546 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
3547 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3548 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3549 | BUG(); |
3550 | } | |
83cd4fe2 | 3551 | return; |
1e3c88bd PZ |
3552 | } |
3553 | #endif | |
3554 | ||
3555 | static DEFINE_SPINLOCK(balancing); | |
3556 | ||
3557 | /* | |
3558 | * It checks each scheduling domain to see if it is due to be balanced, | |
3559 | * and initiates a balancing operation if so. | |
3560 | * | |
3561 | * Balancing parameters are set up in arch_init_sched_domains. | |
3562 | */ | |
3563 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3564 | { | |
3565 | int balance = 1; | |
3566 | struct rq *rq = cpu_rq(cpu); | |
3567 | unsigned long interval; | |
3568 | struct sched_domain *sd; | |
3569 | /* Earliest time when we have to do rebalance again */ | |
3570 | unsigned long next_balance = jiffies + 60*HZ; | |
3571 | int update_next_balance = 0; | |
3572 | int need_serialize; | |
3573 | ||
3574 | for_each_domain(cpu, sd) { | |
3575 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3576 | continue; | |
3577 | ||
3578 | interval = sd->balance_interval; | |
3579 | if (idle != CPU_IDLE) | |
3580 | interval *= sd->busy_factor; | |
3581 | ||
3582 | /* scale ms to jiffies */ | |
3583 | interval = msecs_to_jiffies(interval); | |
3584 | if (unlikely(!interval)) | |
3585 | interval = 1; | |
3586 | if (interval > HZ*NR_CPUS/10) | |
3587 | interval = HZ*NR_CPUS/10; | |
3588 | ||
3589 | need_serialize = sd->flags & SD_SERIALIZE; | |
3590 | ||
3591 | if (need_serialize) { | |
3592 | if (!spin_trylock(&balancing)) | |
3593 | goto out; | |
3594 | } | |
3595 | ||
3596 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3597 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3598 | /* | |
3599 | * We've pulled tasks over so either we're no | |
3600 | * longer idle, or one of our SMT siblings is | |
3601 | * not idle. | |
3602 | */ | |
3603 | idle = CPU_NOT_IDLE; | |
3604 | } | |
3605 | sd->last_balance = jiffies; | |
3606 | } | |
3607 | if (need_serialize) | |
3608 | spin_unlock(&balancing); | |
3609 | out: | |
3610 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3611 | next_balance = sd->last_balance + interval; | |
3612 | update_next_balance = 1; | |
3613 | } | |
3614 | ||
3615 | /* | |
3616 | * Stop the load balance at this level. There is another | |
3617 | * CPU in our sched group which is doing load balancing more | |
3618 | * actively. | |
3619 | */ | |
3620 | if (!balance) | |
3621 | break; | |
3622 | } | |
3623 | ||
3624 | /* | |
3625 | * next_balance will be updated only when there is a need. | |
3626 | * When the cpu is attached to null domain for ex, it will not be | |
3627 | * updated. | |
3628 | */ | |
3629 | if (likely(update_next_balance)) | |
3630 | rq->next_balance = next_balance; | |
3631 | } | |
3632 | ||
83cd4fe2 | 3633 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 3634 | /* |
83cd4fe2 | 3635 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
3636 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
3637 | */ | |
83cd4fe2 VP |
3638 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
3639 | { | |
3640 | struct rq *this_rq = cpu_rq(this_cpu); | |
3641 | struct rq *rq; | |
3642 | int balance_cpu; | |
3643 | ||
3644 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
3645 | return; | |
3646 | ||
3647 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
3648 | if (balance_cpu == this_cpu) | |
3649 | continue; | |
3650 | ||
3651 | /* | |
3652 | * If this cpu gets work to do, stop the load balancing | |
3653 | * work being done for other cpus. Next load | |
3654 | * balancing owner will pick it up. | |
3655 | */ | |
3656 | if (need_resched()) { | |
3657 | this_rq->nohz_balance_kick = 0; | |
3658 | break; | |
3659 | } | |
3660 | ||
3661 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 3662 | update_rq_clock(this_rq); |
83cd4fe2 VP |
3663 | update_cpu_load(this_rq); |
3664 | raw_spin_unlock_irq(&this_rq->lock); | |
3665 | ||
3666 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3667 | ||
3668 | rq = cpu_rq(balance_cpu); | |
3669 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3670 | this_rq->next_balance = rq->next_balance; | |
3671 | } | |
3672 | nohz.next_balance = this_rq->next_balance; | |
3673 | this_rq->nohz_balance_kick = 0; | |
3674 | } | |
3675 | ||
3676 | /* | |
3677 | * Current heuristic for kicking the idle load balancer | |
3678 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
3679 | * idle load balancer when it has more than one process active. This | |
3680 | * eliminates the need for idle load balancing altogether when we have | |
3681 | * only one running process in the system (common case). | |
3682 | * - If there are more than one busy CPU, idle load balancer may have | |
3683 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
3684 | * SMT or core siblings and can run better if they move to different | |
3685 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
3686 | * which will kick idle load balancer as soon as it has any load. | |
3687 | */ | |
3688 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
3689 | { | |
3690 | unsigned long now = jiffies; | |
3691 | int ret; | |
3692 | int first_pick_cpu, second_pick_cpu; | |
3693 | ||
3694 | if (time_before(now, nohz.next_balance)) | |
3695 | return 0; | |
3696 | ||
f6c3f168 | 3697 | if (rq->idle_at_tick) |
83cd4fe2 VP |
3698 | return 0; |
3699 | ||
3700 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
3701 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
3702 | ||
3703 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
3704 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
3705 | return 0; | |
3706 | ||
3707 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
3708 | if (ret == nr_cpu_ids || ret == cpu) { | |
3709 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
3710 | if (rq->nr_running > 1) | |
3711 | return 1; | |
3712 | } else { | |
3713 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
3714 | if (ret == nr_cpu_ids || ret == cpu) { | |
3715 | if (rq->nr_running) | |
3716 | return 1; | |
3717 | } | |
3718 | } | |
3719 | return 0; | |
3720 | } | |
3721 | #else | |
3722 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
3723 | #endif | |
3724 | ||
3725 | /* | |
3726 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
3727 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
3728 | */ | |
1e3c88bd PZ |
3729 | static void run_rebalance_domains(struct softirq_action *h) |
3730 | { | |
3731 | int this_cpu = smp_processor_id(); | |
3732 | struct rq *this_rq = cpu_rq(this_cpu); | |
3733 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
3734 | CPU_IDLE : CPU_NOT_IDLE; | |
3735 | ||
3736 | rebalance_domains(this_cpu, idle); | |
3737 | ||
1e3c88bd | 3738 | /* |
83cd4fe2 | 3739 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
3740 | * balancing on behalf of the other idle cpus whose ticks are |
3741 | * stopped. | |
3742 | */ | |
83cd4fe2 | 3743 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
3744 | } |
3745 | ||
3746 | static inline int on_null_domain(int cpu) | |
3747 | { | |
90a6501f | 3748 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
3749 | } |
3750 | ||
3751 | /* | |
3752 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
3753 | */ |
3754 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
3755 | { | |
1e3c88bd PZ |
3756 | /* Don't need to rebalance while attached to NULL domain */ |
3757 | if (time_after_eq(jiffies, rq->next_balance) && | |
3758 | likely(!on_null_domain(cpu))) | |
3759 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
3760 | #ifdef CONFIG_NO_HZ |
3761 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
3762 | nohz_balancer_kick(cpu); | |
3763 | #endif | |
1e3c88bd PZ |
3764 | } |
3765 | ||
0bcdcf28 CE |
3766 | static void rq_online_fair(struct rq *rq) |
3767 | { | |
3768 | update_sysctl(); | |
3769 | } | |
3770 | ||
3771 | static void rq_offline_fair(struct rq *rq) | |
3772 | { | |
3773 | update_sysctl(); | |
3774 | } | |
3775 | ||
1e3c88bd PZ |
3776 | #else /* CONFIG_SMP */ |
3777 | ||
3778 | /* | |
3779 | * on UP we do not need to balance between CPUs: | |
3780 | */ | |
3781 | static inline void idle_balance(int cpu, struct rq *rq) | |
3782 | { | |
3783 | } | |
3784 | ||
55e12e5e | 3785 | #endif /* CONFIG_SMP */ |
e1d1484f | 3786 | |
bf0f6f24 IM |
3787 | /* |
3788 | * scheduler tick hitting a task of our scheduling class: | |
3789 | */ | |
8f4d37ec | 3790 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
3791 | { |
3792 | struct cfs_rq *cfs_rq; | |
3793 | struct sched_entity *se = &curr->se; | |
3794 | ||
3795 | for_each_sched_entity(se) { | |
3796 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 3797 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
3798 | } |
3799 | } | |
3800 | ||
3801 | /* | |
cd29fe6f PZ |
3802 | * called on fork with the child task as argument from the parent's context |
3803 | * - child not yet on the tasklist | |
3804 | * - preemption disabled | |
bf0f6f24 | 3805 | */ |
cd29fe6f | 3806 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 3807 | { |
cd29fe6f | 3808 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 3809 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 3810 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
3811 | struct rq *rq = this_rq(); |
3812 | unsigned long flags; | |
3813 | ||
05fa785c | 3814 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 3815 | |
861d034e PZ |
3816 | update_rq_clock(rq); |
3817 | ||
b0a0f667 PM |
3818 | if (unlikely(task_cpu(p) != this_cpu)) { |
3819 | rcu_read_lock(); | |
cd29fe6f | 3820 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
3821 | rcu_read_unlock(); |
3822 | } | |
bf0f6f24 | 3823 | |
7109c442 | 3824 | update_curr(cfs_rq); |
cd29fe6f | 3825 | |
b5d9d734 MG |
3826 | if (curr) |
3827 | se->vruntime = curr->vruntime; | |
aeb73b04 | 3828 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 3829 | |
cd29fe6f | 3830 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 3831 | /* |
edcb60a3 IM |
3832 | * Upon rescheduling, sched_class::put_prev_task() will place |
3833 | * 'current' within the tree based on its new key value. | |
3834 | */ | |
4d78e7b6 | 3835 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 3836 | resched_task(rq->curr); |
4d78e7b6 | 3837 | } |
bf0f6f24 | 3838 | |
88ec22d3 PZ |
3839 | se->vruntime -= cfs_rq->min_vruntime; |
3840 | ||
05fa785c | 3841 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
3842 | } |
3843 | ||
cb469845 SR |
3844 | /* |
3845 | * Priority of the task has changed. Check to see if we preempt | |
3846 | * the current task. | |
3847 | */ | |
3848 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
3849 | int oldprio, int running) | |
3850 | { | |
3851 | /* | |
3852 | * Reschedule if we are currently running on this runqueue and | |
3853 | * our priority decreased, or if we are not currently running on | |
3854 | * this runqueue and our priority is higher than the current's | |
3855 | */ | |
3856 | if (running) { | |
3857 | if (p->prio > oldprio) | |
3858 | resched_task(rq->curr); | |
3859 | } else | |
15afe09b | 3860 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
3861 | } |
3862 | ||
3863 | /* | |
3864 | * We switched to the sched_fair class. | |
3865 | */ | |
3866 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
3867 | int running) | |
3868 | { | |
3869 | /* | |
3870 | * We were most likely switched from sched_rt, so | |
3871 | * kick off the schedule if running, otherwise just see | |
3872 | * if we can still preempt the current task. | |
3873 | */ | |
3874 | if (running) | |
3875 | resched_task(rq->curr); | |
3876 | else | |
15afe09b | 3877 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
3878 | } |
3879 | ||
83b699ed SV |
3880 | /* Account for a task changing its policy or group. |
3881 | * | |
3882 | * This routine is mostly called to set cfs_rq->curr field when a task | |
3883 | * migrates between groups/classes. | |
3884 | */ | |
3885 | static void set_curr_task_fair(struct rq *rq) | |
3886 | { | |
3887 | struct sched_entity *se = &rq->curr->se; | |
3888 | ||
3889 | for_each_sched_entity(se) | |
3890 | set_next_entity(cfs_rq_of(se), se); | |
3891 | } | |
3892 | ||
810b3817 | 3893 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 3894 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 3895 | { |
b2b5ce02 PZ |
3896 | /* |
3897 | * If the task was not on the rq at the time of this cgroup movement | |
3898 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
3899 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
3900 | * bonus in place_entity()). | |
3901 | * | |
3902 | * If it was on the rq, we've just 'preempted' it, which does convert | |
3903 | * ->vruntime to a relative base. | |
3904 | * | |
3905 | * Make sure both cases convert their relative position when migrating | |
3906 | * to another cgroup's rq. This does somewhat interfere with the | |
3907 | * fair sleeper stuff for the first placement, but who cares. | |
3908 | */ | |
3909 | if (!on_rq) | |
3910 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
3911 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 3912 | if (!on_rq) |
b2b5ce02 | 3913 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
3914 | } |
3915 | #endif | |
3916 | ||
6d686f45 | 3917 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
3918 | { |
3919 | struct sched_entity *se = &task->se; | |
0d721cea PW |
3920 | unsigned int rr_interval = 0; |
3921 | ||
3922 | /* | |
3923 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
3924 | * idle runqueue: | |
3925 | */ | |
0d721cea PW |
3926 | if (rq->cfs.load.weight) |
3927 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
3928 | |
3929 | return rr_interval; | |
3930 | } | |
3931 | ||
bf0f6f24 IM |
3932 | /* |
3933 | * All the scheduling class methods: | |
3934 | */ | |
5522d5d5 IM |
3935 | static const struct sched_class fair_sched_class = { |
3936 | .next = &idle_sched_class, | |
bf0f6f24 IM |
3937 | .enqueue_task = enqueue_task_fair, |
3938 | .dequeue_task = dequeue_task_fair, | |
3939 | .yield_task = yield_task_fair, | |
3940 | ||
2e09bf55 | 3941 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
3942 | |
3943 | .pick_next_task = pick_next_task_fair, | |
3944 | .put_prev_task = put_prev_task_fair, | |
3945 | ||
681f3e68 | 3946 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
3947 | .select_task_rq = select_task_rq_fair, |
3948 | ||
0bcdcf28 CE |
3949 | .rq_online = rq_online_fair, |
3950 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
3951 | |
3952 | .task_waking = task_waking_fair, | |
681f3e68 | 3953 | #endif |
bf0f6f24 | 3954 | |
83b699ed | 3955 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 3956 | .task_tick = task_tick_fair, |
cd29fe6f | 3957 | .task_fork = task_fork_fair, |
cb469845 SR |
3958 | |
3959 | .prio_changed = prio_changed_fair, | |
3960 | .switched_to = switched_to_fair, | |
810b3817 | 3961 | |
0d721cea PW |
3962 | .get_rr_interval = get_rr_interval_fair, |
3963 | ||
810b3817 | 3964 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 3965 | .task_move_group = task_move_group_fair, |
810b3817 | 3966 | #endif |
bf0f6f24 IM |
3967 | }; |
3968 | ||
3969 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 3970 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 3971 | { |
bf0f6f24 IM |
3972 | struct cfs_rq *cfs_rq; |
3973 | ||
5973e5b9 | 3974 | rcu_read_lock(); |
c3b64f1e | 3975 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 3976 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 3977 | rcu_read_unlock(); |
bf0f6f24 IM |
3978 | } |
3979 | #endif |