Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
3436ae12 | 25 | #include <linux/cpumask.h> |
9745512c | 26 | |
bf0f6f24 | 27 | /* |
21805085 | 28 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 29 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 30 | * |
21805085 | 31 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
32 | * 'timeslice length' - timeslices in CFS are of variable length |
33 | * and have no persistent notion like in traditional, time-slice | |
34 | * based scheduling concepts. | |
bf0f6f24 | 35 | * |
d274a4ce IM |
36 | * (to see the precise effective timeslice length of your workload, |
37 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 38 | */ |
21406928 MG |
39 | unsigned int sysctl_sched_latency = 6000000ULL; |
40 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 41 | |
1983a922 CE |
42 | /* |
43 | * The initial- and re-scaling of tunables is configurable | |
44 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
45 | * | |
46 | * Options are: | |
47 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
48 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
49 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
50 | */ | |
51 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
52 | = SCHED_TUNABLESCALING_LOG; | |
53 | ||
2bd8e6d4 | 54 | /* |
b2be5e96 | 55 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 56 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 57 | */ |
0bf377bb IM |
58 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
59 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
60 | |
61 | /* | |
b2be5e96 PZ |
62 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
63 | */ | |
0bf377bb | 64 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
65 | |
66 | /* | |
2bba22c5 | 67 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 68 | * parent will (try to) run first. |
21805085 | 69 | */ |
2bba22c5 | 70 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 71 | |
bf0f6f24 IM |
72 | /* |
73 | * SCHED_OTHER wake-up granularity. | |
172e082a | 74 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
75 | * |
76 | * This option delays the preemption effects of decoupled workloads | |
77 | * and reduces their over-scheduling. Synchronous workloads will still | |
78 | * have immediate wakeup/sleep latencies. | |
79 | */ | |
172e082a | 80 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 81 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 82 | |
da84d961 IM |
83 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
84 | ||
a7a4f8a7 PT |
85 | /* |
86 | * The exponential sliding window over which load is averaged for shares | |
87 | * distribution. | |
88 | * (default: 10msec) | |
89 | */ | |
90 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
91 | ||
a4c2f00f PZ |
92 | static const struct sched_class fair_sched_class; |
93 | ||
bf0f6f24 IM |
94 | /************************************************************** |
95 | * CFS operations on generic schedulable entities: | |
96 | */ | |
97 | ||
62160e3f | 98 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 99 | |
62160e3f | 100 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
101 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
102 | { | |
62160e3f | 103 | return cfs_rq->rq; |
bf0f6f24 IM |
104 | } |
105 | ||
62160e3f IM |
106 | /* An entity is a task if it doesn't "own" a runqueue */ |
107 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 108 | |
8f48894f PZ |
109 | static inline struct task_struct *task_of(struct sched_entity *se) |
110 | { | |
111 | #ifdef CONFIG_SCHED_DEBUG | |
112 | WARN_ON_ONCE(!entity_is_task(se)); | |
113 | #endif | |
114 | return container_of(se, struct task_struct, se); | |
115 | } | |
116 | ||
b758149c PZ |
117 | /* Walk up scheduling entities hierarchy */ |
118 | #define for_each_sched_entity(se) \ | |
119 | for (; se; se = se->parent) | |
120 | ||
121 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
122 | { | |
123 | return p->se.cfs_rq; | |
124 | } | |
125 | ||
126 | /* runqueue on which this entity is (to be) queued */ | |
127 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
128 | { | |
129 | return se->cfs_rq; | |
130 | } | |
131 | ||
132 | /* runqueue "owned" by this group */ | |
133 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
134 | { | |
135 | return grp->my_q; | |
136 | } | |
137 | ||
138 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
139 | * another cpu ('this_cpu') | |
140 | */ | |
141 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
142 | { | |
143 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
144 | } | |
145 | ||
3d4b47b4 PZ |
146 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
147 | { | |
148 | if (!cfs_rq->on_list) { | |
67e86250 PT |
149 | /* |
150 | * Ensure we either appear before our parent (if already | |
151 | * enqueued) or force our parent to appear after us when it is | |
152 | * enqueued. The fact that we always enqueue bottom-up | |
153 | * reduces this to two cases. | |
154 | */ | |
155 | if (cfs_rq->tg->parent && | |
156 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
157 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
158 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
159 | } else { | |
160 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 161 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 162 | } |
3d4b47b4 PZ |
163 | |
164 | cfs_rq->on_list = 1; | |
165 | } | |
166 | } | |
167 | ||
168 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
169 | { | |
170 | if (cfs_rq->on_list) { | |
171 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
172 | cfs_rq->on_list = 0; | |
173 | } | |
174 | } | |
175 | ||
b758149c PZ |
176 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
177 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
178 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
179 | ||
180 | /* Do the two (enqueued) entities belong to the same group ? */ | |
181 | static inline int | |
182 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
183 | { | |
184 | if (se->cfs_rq == pse->cfs_rq) | |
185 | return 1; | |
186 | ||
187 | return 0; | |
188 | } | |
189 | ||
190 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
191 | { | |
192 | return se->parent; | |
193 | } | |
194 | ||
464b7527 PZ |
195 | /* return depth at which a sched entity is present in the hierarchy */ |
196 | static inline int depth_se(struct sched_entity *se) | |
197 | { | |
198 | int depth = 0; | |
199 | ||
200 | for_each_sched_entity(se) | |
201 | depth++; | |
202 | ||
203 | return depth; | |
204 | } | |
205 | ||
206 | static void | |
207 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
208 | { | |
209 | int se_depth, pse_depth; | |
210 | ||
211 | /* | |
212 | * preemption test can be made between sibling entities who are in the | |
213 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
214 | * both tasks until we find their ancestors who are siblings of common | |
215 | * parent. | |
216 | */ | |
217 | ||
218 | /* First walk up until both entities are at same depth */ | |
219 | se_depth = depth_se(*se); | |
220 | pse_depth = depth_se(*pse); | |
221 | ||
222 | while (se_depth > pse_depth) { | |
223 | se_depth--; | |
224 | *se = parent_entity(*se); | |
225 | } | |
226 | ||
227 | while (pse_depth > se_depth) { | |
228 | pse_depth--; | |
229 | *pse = parent_entity(*pse); | |
230 | } | |
231 | ||
232 | while (!is_same_group(*se, *pse)) { | |
233 | *se = parent_entity(*se); | |
234 | *pse = parent_entity(*pse); | |
235 | } | |
236 | } | |
237 | ||
8f48894f PZ |
238 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
239 | ||
240 | static inline struct task_struct *task_of(struct sched_entity *se) | |
241 | { | |
242 | return container_of(se, struct task_struct, se); | |
243 | } | |
bf0f6f24 | 244 | |
62160e3f IM |
245 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
246 | { | |
247 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
248 | } |
249 | ||
250 | #define entity_is_task(se) 1 | |
251 | ||
b758149c PZ |
252 | #define for_each_sched_entity(se) \ |
253 | for (; se; se = NULL) | |
bf0f6f24 | 254 | |
b758149c | 255 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 256 | { |
b758149c | 257 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
258 | } |
259 | ||
b758149c PZ |
260 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
261 | { | |
262 | struct task_struct *p = task_of(se); | |
263 | struct rq *rq = task_rq(p); | |
264 | ||
265 | return &rq->cfs; | |
266 | } | |
267 | ||
268 | /* runqueue "owned" by this group */ | |
269 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
270 | { | |
271 | return NULL; | |
272 | } | |
273 | ||
274 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
275 | { | |
276 | return &cpu_rq(this_cpu)->cfs; | |
277 | } | |
278 | ||
3d4b47b4 PZ |
279 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
280 | { | |
281 | } | |
282 | ||
283 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
284 | { | |
285 | } | |
286 | ||
b758149c PZ |
287 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
288 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
289 | ||
290 | static inline int | |
291 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
292 | { | |
293 | return 1; | |
294 | } | |
295 | ||
296 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
297 | { | |
298 | return NULL; | |
299 | } | |
300 | ||
464b7527 PZ |
301 | static inline void |
302 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
303 | { | |
304 | } | |
305 | ||
b758149c PZ |
306 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
307 | ||
bf0f6f24 IM |
308 | |
309 | /************************************************************** | |
310 | * Scheduling class tree data structure manipulation methods: | |
311 | */ | |
312 | ||
0702e3eb | 313 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 314 | { |
368059a9 PZ |
315 | s64 delta = (s64)(vruntime - min_vruntime); |
316 | if (delta > 0) | |
02e0431a PZ |
317 | min_vruntime = vruntime; |
318 | ||
319 | return min_vruntime; | |
320 | } | |
321 | ||
0702e3eb | 322 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
323 | { |
324 | s64 delta = (s64)(vruntime - min_vruntime); | |
325 | if (delta < 0) | |
326 | min_vruntime = vruntime; | |
327 | ||
328 | return min_vruntime; | |
329 | } | |
330 | ||
54fdc581 FC |
331 | static inline int entity_before(struct sched_entity *a, |
332 | struct sched_entity *b) | |
333 | { | |
334 | return (s64)(a->vruntime - b->vruntime) < 0; | |
335 | } | |
336 | ||
0702e3eb | 337 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 338 | { |
30cfdcfc | 339 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
340 | } |
341 | ||
1af5f730 PZ |
342 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
343 | { | |
344 | u64 vruntime = cfs_rq->min_vruntime; | |
345 | ||
346 | if (cfs_rq->curr) | |
347 | vruntime = cfs_rq->curr->vruntime; | |
348 | ||
349 | if (cfs_rq->rb_leftmost) { | |
350 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
351 | struct sched_entity, | |
352 | run_node); | |
353 | ||
e17036da | 354 | if (!cfs_rq->curr) |
1af5f730 PZ |
355 | vruntime = se->vruntime; |
356 | else | |
357 | vruntime = min_vruntime(vruntime, se->vruntime); | |
358 | } | |
359 | ||
360 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
361 | } | |
362 | ||
bf0f6f24 IM |
363 | /* |
364 | * Enqueue an entity into the rb-tree: | |
365 | */ | |
0702e3eb | 366 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
367 | { |
368 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
369 | struct rb_node *parent = NULL; | |
370 | struct sched_entity *entry; | |
9014623c | 371 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
372 | int leftmost = 1; |
373 | ||
374 | /* | |
375 | * Find the right place in the rbtree: | |
376 | */ | |
377 | while (*link) { | |
378 | parent = *link; | |
379 | entry = rb_entry(parent, struct sched_entity, run_node); | |
380 | /* | |
381 | * We dont care about collisions. Nodes with | |
382 | * the same key stay together. | |
383 | */ | |
9014623c | 384 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
385 | link = &parent->rb_left; |
386 | } else { | |
387 | link = &parent->rb_right; | |
388 | leftmost = 0; | |
389 | } | |
390 | } | |
391 | ||
392 | /* | |
393 | * Maintain a cache of leftmost tree entries (it is frequently | |
394 | * used): | |
395 | */ | |
1af5f730 | 396 | if (leftmost) |
57cb499d | 397 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
398 | |
399 | rb_link_node(&se->run_node, parent, link); | |
400 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
401 | } |
402 | ||
0702e3eb | 403 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 404 | { |
3fe69747 PZ |
405 | if (cfs_rq->rb_leftmost == &se->run_node) { |
406 | struct rb_node *next_node; | |
3fe69747 PZ |
407 | |
408 | next_node = rb_next(&se->run_node); | |
409 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 410 | } |
e9acbff6 | 411 | |
bf0f6f24 | 412 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
413 | } |
414 | ||
ac53db59 | 415 | static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 416 | { |
f4b6755f PZ |
417 | struct rb_node *left = cfs_rq->rb_leftmost; |
418 | ||
419 | if (!left) | |
420 | return NULL; | |
421 | ||
422 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
423 | } |
424 | ||
ac53db59 RR |
425 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
426 | { | |
427 | struct rb_node *next = rb_next(&se->run_node); | |
428 | ||
429 | if (!next) | |
430 | return NULL; | |
431 | ||
432 | return rb_entry(next, struct sched_entity, run_node); | |
433 | } | |
434 | ||
435 | #ifdef CONFIG_SCHED_DEBUG | |
f4b6755f | 436 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 437 | { |
7eee3e67 | 438 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 439 | |
70eee74b BS |
440 | if (!last) |
441 | return NULL; | |
7eee3e67 IM |
442 | |
443 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
444 | } |
445 | ||
bf0f6f24 IM |
446 | /************************************************************** |
447 | * Scheduling class statistics methods: | |
448 | */ | |
449 | ||
acb4a848 | 450 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 451 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
452 | loff_t *ppos) |
453 | { | |
8d65af78 | 454 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 455 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
456 | |
457 | if (ret || !write) | |
458 | return ret; | |
459 | ||
460 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
461 | sysctl_sched_min_granularity); | |
462 | ||
acb4a848 CE |
463 | #define WRT_SYSCTL(name) \ |
464 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
465 | WRT_SYSCTL(sched_min_granularity); | |
466 | WRT_SYSCTL(sched_latency); | |
467 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
468 | #undef WRT_SYSCTL |
469 | ||
b2be5e96 PZ |
470 | return 0; |
471 | } | |
472 | #endif | |
647e7cac | 473 | |
a7be37ac | 474 | /* |
f9c0b095 | 475 | * delta /= w |
a7be37ac PZ |
476 | */ |
477 | static inline unsigned long | |
478 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
479 | { | |
f9c0b095 PZ |
480 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
481 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
482 | |
483 | return delta; | |
484 | } | |
485 | ||
647e7cac IM |
486 | /* |
487 | * The idea is to set a period in which each task runs once. | |
488 | * | |
489 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
490 | * this period because otherwise the slices get too small. | |
491 | * | |
492 | * p = (nr <= nl) ? l : l*nr/nl | |
493 | */ | |
4d78e7b6 PZ |
494 | static u64 __sched_period(unsigned long nr_running) |
495 | { | |
496 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 497 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
498 | |
499 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 500 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 501 | period *= nr_running; |
4d78e7b6 PZ |
502 | } |
503 | ||
504 | return period; | |
505 | } | |
506 | ||
647e7cac IM |
507 | /* |
508 | * We calculate the wall-time slice from the period by taking a part | |
509 | * proportional to the weight. | |
510 | * | |
f9c0b095 | 511 | * s = p*P[w/rw] |
647e7cac | 512 | */ |
6d0f0ebd | 513 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 514 | { |
0a582440 | 515 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 516 | |
0a582440 | 517 | for_each_sched_entity(se) { |
6272d68c | 518 | struct load_weight *load; |
3104bf03 | 519 | struct load_weight lw; |
6272d68c LM |
520 | |
521 | cfs_rq = cfs_rq_of(se); | |
522 | load = &cfs_rq->load; | |
f9c0b095 | 523 | |
0a582440 | 524 | if (unlikely(!se->on_rq)) { |
3104bf03 | 525 | lw = cfs_rq->load; |
0a582440 MG |
526 | |
527 | update_load_add(&lw, se->load.weight); | |
528 | load = &lw; | |
529 | } | |
530 | slice = calc_delta_mine(slice, se->load.weight, load); | |
531 | } | |
532 | return slice; | |
bf0f6f24 IM |
533 | } |
534 | ||
647e7cac | 535 | /* |
ac884dec | 536 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 537 | * |
f9c0b095 | 538 | * vs = s/w |
647e7cac | 539 | */ |
f9c0b095 | 540 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 541 | { |
f9c0b095 | 542 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
543 | } |
544 | ||
d6b55918 | 545 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); |
6d5ab293 | 546 | static void update_cfs_shares(struct cfs_rq *cfs_rq); |
3b3d190e | 547 | |
bf0f6f24 IM |
548 | /* |
549 | * Update the current task's runtime statistics. Skip current tasks that | |
550 | * are not in our scheduling class. | |
551 | */ | |
552 | static inline void | |
8ebc91d9 IM |
553 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
554 | unsigned long delta_exec) | |
bf0f6f24 | 555 | { |
bbdba7c0 | 556 | unsigned long delta_exec_weighted; |
bf0f6f24 | 557 | |
41acab88 LDM |
558 | schedstat_set(curr->statistics.exec_max, |
559 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
560 | |
561 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 562 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 563 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 564 | |
e9acbff6 | 565 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 566 | update_min_vruntime(cfs_rq); |
3b3d190e | 567 | |
70caf8a6 | 568 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
3b3d190e | 569 | cfs_rq->load_unacc_exec_time += delta_exec; |
3b3d190e | 570 | #endif |
bf0f6f24 IM |
571 | } |
572 | ||
b7cc0896 | 573 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 574 | { |
429d43bc | 575 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 576 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
577 | unsigned long delta_exec; |
578 | ||
579 | if (unlikely(!curr)) | |
580 | return; | |
581 | ||
582 | /* | |
583 | * Get the amount of time the current task was running | |
584 | * since the last time we changed load (this cannot | |
585 | * overflow on 32 bits): | |
586 | */ | |
8ebc91d9 | 587 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
588 | if (!delta_exec) |
589 | return; | |
bf0f6f24 | 590 | |
8ebc91d9 IM |
591 | __update_curr(cfs_rq, curr, delta_exec); |
592 | curr->exec_start = now; | |
d842de87 SV |
593 | |
594 | if (entity_is_task(curr)) { | |
595 | struct task_struct *curtask = task_of(curr); | |
596 | ||
f977bb49 | 597 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 598 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 599 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 600 | } |
bf0f6f24 IM |
601 | } |
602 | ||
603 | static inline void | |
5870db5b | 604 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 605 | { |
41acab88 | 606 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
607 | } |
608 | ||
bf0f6f24 IM |
609 | /* |
610 | * Task is being enqueued - update stats: | |
611 | */ | |
d2417e5a | 612 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 613 | { |
bf0f6f24 IM |
614 | /* |
615 | * Are we enqueueing a waiting task? (for current tasks | |
616 | * a dequeue/enqueue event is a NOP) | |
617 | */ | |
429d43bc | 618 | if (se != cfs_rq->curr) |
5870db5b | 619 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
620 | } |
621 | ||
bf0f6f24 | 622 | static void |
9ef0a961 | 623 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 624 | { |
41acab88 LDM |
625 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
626 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
627 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
628 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
629 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
630 | #ifdef CONFIG_SCHEDSTATS |
631 | if (entity_is_task(se)) { | |
632 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 633 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
634 | } |
635 | #endif | |
41acab88 | 636 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
637 | } |
638 | ||
639 | static inline void | |
19b6a2e3 | 640 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 641 | { |
bf0f6f24 IM |
642 | /* |
643 | * Mark the end of the wait period if dequeueing a | |
644 | * waiting task: | |
645 | */ | |
429d43bc | 646 | if (se != cfs_rq->curr) |
9ef0a961 | 647 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
648 | } |
649 | ||
650 | /* | |
651 | * We are picking a new current task - update its stats: | |
652 | */ | |
653 | static inline void | |
79303e9e | 654 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
655 | { |
656 | /* | |
657 | * We are starting a new run period: | |
658 | */ | |
305e6835 | 659 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
660 | } |
661 | ||
bf0f6f24 IM |
662 | /************************************************** |
663 | * Scheduling class queueing methods: | |
664 | */ | |
665 | ||
c09595f6 PZ |
666 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
667 | static void | |
668 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
669 | { | |
670 | cfs_rq->task_weight += weight; | |
671 | } | |
672 | #else | |
673 | static inline void | |
674 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
675 | { | |
676 | } | |
677 | #endif | |
678 | ||
30cfdcfc DA |
679 | static void |
680 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
681 | { | |
682 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
683 | if (!parent_entity(se)) |
684 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 685 | if (entity_is_task(se)) { |
c09595f6 | 686 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
687 | list_add(&se->group_node, &cfs_rq->tasks); |
688 | } | |
30cfdcfc | 689 | cfs_rq->nr_running++; |
30cfdcfc DA |
690 | } |
691 | ||
692 | static void | |
693 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
694 | { | |
695 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
696 | if (!parent_entity(se)) |
697 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 698 | if (entity_is_task(se)) { |
c09595f6 | 699 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
700 | list_del_init(&se->group_node); |
701 | } | |
30cfdcfc | 702 | cfs_rq->nr_running--; |
30cfdcfc DA |
703 | } |
704 | ||
3ff6dcac YZ |
705 | #ifdef CONFIG_FAIR_GROUP_SCHED |
706 | # ifdef CONFIG_SMP | |
d6b55918 PT |
707 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
708 | int global_update) | |
709 | { | |
710 | struct task_group *tg = cfs_rq->tg; | |
711 | long load_avg; | |
712 | ||
713 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
714 | load_avg -= cfs_rq->load_contribution; | |
715 | ||
716 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
717 | atomic_add(load_avg, &tg->load_weight); | |
718 | cfs_rq->load_contribution += load_avg; | |
719 | } | |
720 | } | |
721 | ||
722 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 723 | { |
a7a4f8a7 | 724 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 725 | u64 now, delta; |
e33078ba | 726 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 727 | |
b815f196 | 728 | if (cfs_rq->tg == &root_task_group) |
2069dd75 PZ |
729 | return; |
730 | ||
05ca62c6 | 731 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
732 | delta = now - cfs_rq->load_stamp; |
733 | ||
e33078ba PT |
734 | /* truncate load history at 4 idle periods */ |
735 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
736 | now - cfs_rq->load_last > 4 * period) { | |
737 | cfs_rq->load_period = 0; | |
738 | cfs_rq->load_avg = 0; | |
f07333bf | 739 | delta = period - 1; |
e33078ba PT |
740 | } |
741 | ||
2069dd75 | 742 | cfs_rq->load_stamp = now; |
3b3d190e | 743 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 744 | cfs_rq->load_period += delta; |
e33078ba PT |
745 | if (load) { |
746 | cfs_rq->load_last = now; | |
747 | cfs_rq->load_avg += delta * load; | |
748 | } | |
2069dd75 | 749 | |
d6b55918 PT |
750 | /* consider updating load contribution on each fold or truncate */ |
751 | if (global_update || cfs_rq->load_period > period | |
752 | || !cfs_rq->load_period) | |
753 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
754 | ||
2069dd75 PZ |
755 | while (cfs_rq->load_period > period) { |
756 | /* | |
757 | * Inline assembly required to prevent the compiler | |
758 | * optimising this loop into a divmod call. | |
759 | * See __iter_div_u64_rem() for another example of this. | |
760 | */ | |
761 | asm("" : "+rm" (cfs_rq->load_period)); | |
762 | cfs_rq->load_period /= 2; | |
763 | cfs_rq->load_avg /= 2; | |
764 | } | |
3d4b47b4 | 765 | |
e33078ba PT |
766 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
767 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
768 | } |
769 | ||
6d5ab293 | 770 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
771 | { |
772 | long load_weight, load, shares; | |
773 | ||
6d5ab293 | 774 | load = cfs_rq->load.weight; |
3ff6dcac YZ |
775 | |
776 | load_weight = atomic_read(&tg->load_weight); | |
3ff6dcac | 777 | load_weight += load; |
6d5ab293 | 778 | load_weight -= cfs_rq->load_contribution; |
3ff6dcac YZ |
779 | |
780 | shares = (tg->shares * load); | |
781 | if (load_weight) | |
782 | shares /= load_weight; | |
783 | ||
784 | if (shares < MIN_SHARES) | |
785 | shares = MIN_SHARES; | |
786 | if (shares > tg->shares) | |
787 | shares = tg->shares; | |
788 | ||
789 | return shares; | |
790 | } | |
791 | ||
792 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
793 | { | |
794 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
795 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 796 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
797 | } |
798 | } | |
799 | # else /* CONFIG_SMP */ | |
800 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
801 | { | |
802 | } | |
803 | ||
6d5ab293 | 804 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
805 | { |
806 | return tg->shares; | |
807 | } | |
808 | ||
809 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
810 | { | |
811 | } | |
812 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
813 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
814 | unsigned long weight) | |
815 | { | |
19e5eebb PT |
816 | if (se->on_rq) { |
817 | /* commit outstanding execution time */ | |
818 | if (cfs_rq->curr == se) | |
819 | update_curr(cfs_rq); | |
2069dd75 | 820 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 821 | } |
2069dd75 PZ |
822 | |
823 | update_load_set(&se->load, weight); | |
824 | ||
825 | if (se->on_rq) | |
826 | account_entity_enqueue(cfs_rq, se); | |
827 | } | |
828 | ||
6d5ab293 | 829 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
830 | { |
831 | struct task_group *tg; | |
832 | struct sched_entity *se; | |
3ff6dcac | 833 | long shares; |
2069dd75 | 834 | |
2069dd75 PZ |
835 | tg = cfs_rq->tg; |
836 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
837 | if (!se) | |
838 | return; | |
3ff6dcac YZ |
839 | #ifndef CONFIG_SMP |
840 | if (likely(se->load.weight == tg->shares)) | |
841 | return; | |
842 | #endif | |
6d5ab293 | 843 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
844 | |
845 | reweight_entity(cfs_rq_of(se), se, shares); | |
846 | } | |
847 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 848 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
849 | { |
850 | } | |
851 | ||
6d5ab293 | 852 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
853 | { |
854 | } | |
43365bd7 PT |
855 | |
856 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
857 | { | |
858 | } | |
2069dd75 PZ |
859 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
860 | ||
2396af69 | 861 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 862 | { |
bf0f6f24 | 863 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
864 | struct task_struct *tsk = NULL; |
865 | ||
866 | if (entity_is_task(se)) | |
867 | tsk = task_of(se); | |
868 | ||
41acab88 LDM |
869 | if (se->statistics.sleep_start) { |
870 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
871 | |
872 | if ((s64)delta < 0) | |
873 | delta = 0; | |
874 | ||
41acab88 LDM |
875 | if (unlikely(delta > se->statistics.sleep_max)) |
876 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 877 | |
41acab88 LDM |
878 | se->statistics.sleep_start = 0; |
879 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 880 | |
768d0c27 | 881 | if (tsk) { |
e414314c | 882 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
883 | trace_sched_stat_sleep(tsk, delta); |
884 | } | |
bf0f6f24 | 885 | } |
41acab88 LDM |
886 | if (se->statistics.block_start) { |
887 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
888 | |
889 | if ((s64)delta < 0) | |
890 | delta = 0; | |
891 | ||
41acab88 LDM |
892 | if (unlikely(delta > se->statistics.block_max)) |
893 | se->statistics.block_max = delta; | |
bf0f6f24 | 894 | |
41acab88 LDM |
895 | se->statistics.block_start = 0; |
896 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 897 | |
e414314c | 898 | if (tsk) { |
8f0dfc34 | 899 | if (tsk->in_iowait) { |
41acab88 LDM |
900 | se->statistics.iowait_sum += delta; |
901 | se->statistics.iowait_count++; | |
768d0c27 | 902 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
903 | } |
904 | ||
e414314c PZ |
905 | /* |
906 | * Blocking time is in units of nanosecs, so shift by | |
907 | * 20 to get a milliseconds-range estimation of the | |
908 | * amount of time that the task spent sleeping: | |
909 | */ | |
910 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
911 | profile_hits(SLEEP_PROFILING, | |
912 | (void *)get_wchan(tsk), | |
913 | delta >> 20); | |
914 | } | |
915 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 916 | } |
bf0f6f24 IM |
917 | } |
918 | #endif | |
919 | } | |
920 | ||
ddc97297 PZ |
921 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
922 | { | |
923 | #ifdef CONFIG_SCHED_DEBUG | |
924 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
925 | ||
926 | if (d < 0) | |
927 | d = -d; | |
928 | ||
929 | if (d > 3*sysctl_sched_latency) | |
930 | schedstat_inc(cfs_rq, nr_spread_over); | |
931 | #endif | |
932 | } | |
933 | ||
aeb73b04 PZ |
934 | static void |
935 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
936 | { | |
1af5f730 | 937 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 938 | |
2cb8600e PZ |
939 | /* |
940 | * The 'current' period is already promised to the current tasks, | |
941 | * however the extra weight of the new task will slow them down a | |
942 | * little, place the new task so that it fits in the slot that | |
943 | * stays open at the end. | |
944 | */ | |
94dfb5e7 | 945 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 946 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 947 | |
a2e7a7eb | 948 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 949 | if (!initial) { |
a2e7a7eb | 950 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 951 | |
a2e7a7eb MG |
952 | /* |
953 | * Halve their sleep time's effect, to allow | |
954 | * for a gentler effect of sleepers: | |
955 | */ | |
956 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
957 | thresh >>= 1; | |
51e0304c | 958 | |
a2e7a7eb | 959 | vruntime -= thresh; |
aeb73b04 PZ |
960 | } |
961 | ||
b5d9d734 MG |
962 | /* ensure we never gain time by being placed backwards. */ |
963 | vruntime = max_vruntime(se->vruntime, vruntime); | |
964 | ||
67e9fb2a | 965 | se->vruntime = vruntime; |
aeb73b04 PZ |
966 | } |
967 | ||
bf0f6f24 | 968 | static void |
88ec22d3 | 969 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 970 | { |
88ec22d3 PZ |
971 | /* |
972 | * Update the normalized vruntime before updating min_vruntime | |
973 | * through callig update_curr(). | |
974 | */ | |
371fd7e7 | 975 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
976 | se->vruntime += cfs_rq->min_vruntime; |
977 | ||
bf0f6f24 | 978 | /* |
a2a2d680 | 979 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 980 | */ |
b7cc0896 | 981 | update_curr(cfs_rq); |
d6b55918 | 982 | update_cfs_load(cfs_rq, 0); |
a992241d | 983 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 984 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 985 | |
88ec22d3 | 986 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 987 | place_entity(cfs_rq, se, 0); |
2396af69 | 988 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 989 | } |
bf0f6f24 | 990 | |
d2417e5a | 991 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 992 | check_spread(cfs_rq, se); |
83b699ed SV |
993 | if (se != cfs_rq->curr) |
994 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 995 | se->on_rq = 1; |
3d4b47b4 PZ |
996 | |
997 | if (cfs_rq->nr_running == 1) | |
998 | list_add_leaf_cfs_rq(cfs_rq); | |
bf0f6f24 IM |
999 | } |
1000 | ||
2c13c919 | 1001 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1002 | { |
2c13c919 RR |
1003 | for_each_sched_entity(se) { |
1004 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1005 | if (cfs_rq->last == se) | |
1006 | cfs_rq->last = NULL; | |
1007 | else | |
1008 | break; | |
1009 | } | |
1010 | } | |
2002c695 | 1011 | |
2c13c919 RR |
1012 | static void __clear_buddies_next(struct sched_entity *se) |
1013 | { | |
1014 | for_each_sched_entity(se) { | |
1015 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1016 | if (cfs_rq->next == se) | |
1017 | cfs_rq->next = NULL; | |
1018 | else | |
1019 | break; | |
1020 | } | |
2002c695 PZ |
1021 | } |
1022 | ||
ac53db59 RR |
1023 | static void __clear_buddies_skip(struct sched_entity *se) |
1024 | { | |
1025 | for_each_sched_entity(se) { | |
1026 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1027 | if (cfs_rq->skip == se) | |
1028 | cfs_rq->skip = NULL; | |
1029 | else | |
1030 | break; | |
1031 | } | |
1032 | } | |
1033 | ||
a571bbea PZ |
1034 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1035 | { | |
2c13c919 RR |
1036 | if (cfs_rq->last == se) |
1037 | __clear_buddies_last(se); | |
1038 | ||
1039 | if (cfs_rq->next == se) | |
1040 | __clear_buddies_next(se); | |
ac53db59 RR |
1041 | |
1042 | if (cfs_rq->skip == se) | |
1043 | __clear_buddies_skip(se); | |
a571bbea PZ |
1044 | } |
1045 | ||
bf0f6f24 | 1046 | static void |
371fd7e7 | 1047 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1048 | { |
a2a2d680 DA |
1049 | /* |
1050 | * Update run-time statistics of the 'current'. | |
1051 | */ | |
1052 | update_curr(cfs_rq); | |
1053 | ||
19b6a2e3 | 1054 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1055 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1056 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1057 | if (entity_is_task(se)) { |
1058 | struct task_struct *tsk = task_of(se); | |
1059 | ||
1060 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1061 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1062 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1063 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1064 | } |
db36cc7d | 1065 | #endif |
67e9fb2a PZ |
1066 | } |
1067 | ||
2002c695 | 1068 | clear_buddies(cfs_rq, se); |
4793241b | 1069 | |
83b699ed | 1070 | if (se != cfs_rq->curr) |
30cfdcfc | 1071 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1072 | se->on_rq = 0; |
d6b55918 | 1073 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1074 | account_entity_dequeue(cfs_rq, se); |
1af5f730 | 1075 | update_min_vruntime(cfs_rq); |
6d5ab293 | 1076 | update_cfs_shares(cfs_rq); |
88ec22d3 PZ |
1077 | |
1078 | /* | |
1079 | * Normalize the entity after updating the min_vruntime because the | |
1080 | * update can refer to the ->curr item and we need to reflect this | |
1081 | * movement in our normalized position. | |
1082 | */ | |
371fd7e7 | 1083 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1084 | se->vruntime -= cfs_rq->min_vruntime; |
bf0f6f24 IM |
1085 | } |
1086 | ||
1087 | /* | |
1088 | * Preempt the current task with a newly woken task if needed: | |
1089 | */ | |
7c92e54f | 1090 | static void |
2e09bf55 | 1091 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1092 | { |
11697830 PZ |
1093 | unsigned long ideal_runtime, delta_exec; |
1094 | ||
6d0f0ebd | 1095 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1096 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1097 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1098 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1099 | /* |
1100 | * The current task ran long enough, ensure it doesn't get | |
1101 | * re-elected due to buddy favours. | |
1102 | */ | |
1103 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1104 | return; |
1105 | } | |
1106 | ||
1107 | /* | |
1108 | * Ensure that a task that missed wakeup preemption by a | |
1109 | * narrow margin doesn't have to wait for a full slice. | |
1110 | * This also mitigates buddy induced latencies under load. | |
1111 | */ | |
1112 | if (!sched_feat(WAKEUP_PREEMPT)) | |
1113 | return; | |
1114 | ||
1115 | if (delta_exec < sysctl_sched_min_granularity) | |
1116 | return; | |
1117 | ||
1118 | if (cfs_rq->nr_running > 1) { | |
ac53db59 | 1119 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac MG |
1120 | s64 delta = curr->vruntime - se->vruntime; |
1121 | ||
d7d82944 MG |
1122 | if (delta < 0) |
1123 | return; | |
1124 | ||
f685ceac MG |
1125 | if (delta > ideal_runtime) |
1126 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 1127 | } |
bf0f6f24 IM |
1128 | } |
1129 | ||
83b699ed | 1130 | static void |
8494f412 | 1131 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1132 | { |
83b699ed SV |
1133 | /* 'current' is not kept within the tree. */ |
1134 | if (se->on_rq) { | |
1135 | /* | |
1136 | * Any task has to be enqueued before it get to execute on | |
1137 | * a CPU. So account for the time it spent waiting on the | |
1138 | * runqueue. | |
1139 | */ | |
1140 | update_stats_wait_end(cfs_rq, se); | |
1141 | __dequeue_entity(cfs_rq, se); | |
1142 | } | |
1143 | ||
79303e9e | 1144 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1145 | cfs_rq->curr = se; |
eba1ed4b IM |
1146 | #ifdef CONFIG_SCHEDSTATS |
1147 | /* | |
1148 | * Track our maximum slice length, if the CPU's load is at | |
1149 | * least twice that of our own weight (i.e. dont track it | |
1150 | * when there are only lesser-weight tasks around): | |
1151 | */ | |
495eca49 | 1152 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1153 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1154 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1155 | } | |
1156 | #endif | |
4a55b450 | 1157 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1158 | } |
1159 | ||
3f3a4904 PZ |
1160 | static int |
1161 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1162 | ||
ac53db59 RR |
1163 | /* |
1164 | * Pick the next process, keeping these things in mind, in this order: | |
1165 | * 1) keep things fair between processes/task groups | |
1166 | * 2) pick the "next" process, since someone really wants that to run | |
1167 | * 3) pick the "last" process, for cache locality | |
1168 | * 4) do not run the "skip" process, if something else is available | |
1169 | */ | |
f4b6755f | 1170 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1171 | { |
ac53db59 | 1172 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1173 | struct sched_entity *left = se; |
f4b6755f | 1174 | |
ac53db59 RR |
1175 | /* |
1176 | * Avoid running the skip buddy, if running something else can | |
1177 | * be done without getting too unfair. | |
1178 | */ | |
1179 | if (cfs_rq->skip == se) { | |
1180 | struct sched_entity *second = __pick_next_entity(se); | |
1181 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1182 | se = second; | |
1183 | } | |
aa2ac252 | 1184 | |
f685ceac MG |
1185 | /* |
1186 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1187 | */ | |
1188 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1189 | se = cfs_rq->last; | |
1190 | ||
ac53db59 RR |
1191 | /* |
1192 | * Someone really wants this to run. If it's not unfair, run it. | |
1193 | */ | |
1194 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1195 | se = cfs_rq->next; | |
1196 | ||
f685ceac | 1197 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1198 | |
1199 | return se; | |
aa2ac252 PZ |
1200 | } |
1201 | ||
ab6cde26 | 1202 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1203 | { |
1204 | /* | |
1205 | * If still on the runqueue then deactivate_task() | |
1206 | * was not called and update_curr() has to be done: | |
1207 | */ | |
1208 | if (prev->on_rq) | |
b7cc0896 | 1209 | update_curr(cfs_rq); |
bf0f6f24 | 1210 | |
ddc97297 | 1211 | check_spread(cfs_rq, prev); |
30cfdcfc | 1212 | if (prev->on_rq) { |
5870db5b | 1213 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1214 | /* Put 'current' back into the tree. */ |
1215 | __enqueue_entity(cfs_rq, prev); | |
1216 | } | |
429d43bc | 1217 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1218 | } |
1219 | ||
8f4d37ec PZ |
1220 | static void |
1221 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1222 | { |
bf0f6f24 | 1223 | /* |
30cfdcfc | 1224 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1225 | */ |
30cfdcfc | 1226 | update_curr(cfs_rq); |
bf0f6f24 | 1227 | |
43365bd7 PT |
1228 | /* |
1229 | * Update share accounting for long-running entities. | |
1230 | */ | |
1231 | update_entity_shares_tick(cfs_rq); | |
1232 | ||
8f4d37ec PZ |
1233 | #ifdef CONFIG_SCHED_HRTICK |
1234 | /* | |
1235 | * queued ticks are scheduled to match the slice, so don't bother | |
1236 | * validating it and just reschedule. | |
1237 | */ | |
983ed7a6 HH |
1238 | if (queued) { |
1239 | resched_task(rq_of(cfs_rq)->curr); | |
1240 | return; | |
1241 | } | |
8f4d37ec PZ |
1242 | /* |
1243 | * don't let the period tick interfere with the hrtick preemption | |
1244 | */ | |
1245 | if (!sched_feat(DOUBLE_TICK) && | |
1246 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1247 | return; | |
1248 | #endif | |
1249 | ||
ce6c1311 | 1250 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 1251 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1252 | } |
1253 | ||
1254 | /************************************************** | |
1255 | * CFS operations on tasks: | |
1256 | */ | |
1257 | ||
8f4d37ec PZ |
1258 | #ifdef CONFIG_SCHED_HRTICK |
1259 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1260 | { | |
8f4d37ec PZ |
1261 | struct sched_entity *se = &p->se; |
1262 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1263 | ||
1264 | WARN_ON(task_rq(p) != rq); | |
1265 | ||
1266 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1267 | u64 slice = sched_slice(cfs_rq, se); | |
1268 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1269 | s64 delta = slice - ran; | |
1270 | ||
1271 | if (delta < 0) { | |
1272 | if (rq->curr == p) | |
1273 | resched_task(p); | |
1274 | return; | |
1275 | } | |
1276 | ||
1277 | /* | |
1278 | * Don't schedule slices shorter than 10000ns, that just | |
1279 | * doesn't make sense. Rely on vruntime for fairness. | |
1280 | */ | |
31656519 | 1281 | if (rq->curr != p) |
157124c1 | 1282 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1283 | |
31656519 | 1284 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1285 | } |
1286 | } | |
a4c2f00f PZ |
1287 | |
1288 | /* | |
1289 | * called from enqueue/dequeue and updates the hrtick when the | |
1290 | * current task is from our class and nr_running is low enough | |
1291 | * to matter. | |
1292 | */ | |
1293 | static void hrtick_update(struct rq *rq) | |
1294 | { | |
1295 | struct task_struct *curr = rq->curr; | |
1296 | ||
1297 | if (curr->sched_class != &fair_sched_class) | |
1298 | return; | |
1299 | ||
1300 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1301 | hrtick_start_fair(rq, curr); | |
1302 | } | |
55e12e5e | 1303 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1304 | static inline void |
1305 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1306 | { | |
1307 | } | |
a4c2f00f PZ |
1308 | |
1309 | static inline void hrtick_update(struct rq *rq) | |
1310 | { | |
1311 | } | |
8f4d37ec PZ |
1312 | #endif |
1313 | ||
bf0f6f24 IM |
1314 | /* |
1315 | * The enqueue_task method is called before nr_running is | |
1316 | * increased. Here we update the fair scheduling stats and | |
1317 | * then put the task into the rbtree: | |
1318 | */ | |
ea87bb78 | 1319 | static void |
371fd7e7 | 1320 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1321 | { |
1322 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1323 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1324 | |
1325 | for_each_sched_entity(se) { | |
62fb1851 | 1326 | if (se->on_rq) |
bf0f6f24 IM |
1327 | break; |
1328 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 PZ |
1329 | enqueue_entity(cfs_rq, se, flags); |
1330 | flags = ENQUEUE_WAKEUP; | |
bf0f6f24 | 1331 | } |
8f4d37ec | 1332 | |
2069dd75 PZ |
1333 | for_each_sched_entity(se) { |
1334 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1335 | ||
d6b55918 | 1336 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1337 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1338 | } |
1339 | ||
a4c2f00f | 1340 | hrtick_update(rq); |
bf0f6f24 IM |
1341 | } |
1342 | ||
1343 | /* | |
1344 | * The dequeue_task method is called before nr_running is | |
1345 | * decreased. We remove the task from the rbtree and | |
1346 | * update the fair scheduling stats: | |
1347 | */ | |
371fd7e7 | 1348 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1349 | { |
1350 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1351 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1352 | |
1353 | for_each_sched_entity(se) { | |
1354 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1355 | dequeue_entity(cfs_rq, se, flags); |
2069dd75 | 1356 | |
bf0f6f24 | 1357 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 1358 | if (cfs_rq->load.weight) |
bf0f6f24 | 1359 | break; |
371fd7e7 | 1360 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1361 | } |
8f4d37ec | 1362 | |
2069dd75 PZ |
1363 | for_each_sched_entity(se) { |
1364 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1365 | ||
d6b55918 | 1366 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1367 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1368 | } |
1369 | ||
a4c2f00f | 1370 | hrtick_update(rq); |
bf0f6f24 IM |
1371 | } |
1372 | ||
e7693a36 | 1373 | #ifdef CONFIG_SMP |
098fb9db | 1374 | |
88ec22d3 PZ |
1375 | static void task_waking_fair(struct rq *rq, struct task_struct *p) |
1376 | { | |
1377 | struct sched_entity *se = &p->se; | |
1378 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1379 | ||
1380 | se->vruntime -= cfs_rq->min_vruntime; | |
1381 | } | |
1382 | ||
bb3469ac | 1383 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1384 | /* |
1385 | * effective_load() calculates the load change as seen from the root_task_group | |
1386 | * | |
1387 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1388 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1389 | * can calculate the shift in shares. | |
f5bfb7d9 | 1390 | */ |
2069dd75 | 1391 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 1392 | { |
4be9daaa | 1393 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1394 | |
1395 | if (!tg->parent) | |
1396 | return wl; | |
1397 | ||
4be9daaa | 1398 | for_each_sched_entity(se) { |
977dda7c | 1399 | long lw, w; |
4be9daaa | 1400 | |
977dda7c PT |
1401 | tg = se->my_q->tg; |
1402 | w = se->my_q->load.weight; | |
bb3469ac | 1403 | |
977dda7c PT |
1404 | /* use this cpu's instantaneous contribution */ |
1405 | lw = atomic_read(&tg->load_weight); | |
1406 | lw -= se->my_q->load_contribution; | |
1407 | lw += w + wg; | |
4be9daaa | 1408 | |
977dda7c | 1409 | wl += w; |
940959e9 | 1410 | |
977dda7c PT |
1411 | if (lw > 0 && wl < lw) |
1412 | wl = (wl * tg->shares) / lw; | |
1413 | else | |
1414 | wl = tg->shares; | |
940959e9 | 1415 | |
977dda7c PT |
1416 | /* zero point is MIN_SHARES */ |
1417 | if (wl < MIN_SHARES) | |
1418 | wl = MIN_SHARES; | |
1419 | wl -= se->load.weight; | |
4be9daaa | 1420 | wg = 0; |
4be9daaa | 1421 | } |
bb3469ac | 1422 | |
4be9daaa | 1423 | return wl; |
bb3469ac | 1424 | } |
4be9daaa | 1425 | |
bb3469ac | 1426 | #else |
4be9daaa | 1427 | |
83378269 PZ |
1428 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1429 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1430 | { |
83378269 | 1431 | return wl; |
bb3469ac | 1432 | } |
4be9daaa | 1433 | |
bb3469ac PZ |
1434 | #endif |
1435 | ||
c88d5910 | 1436 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1437 | { |
e37b6a7b | 1438 | s64 this_load, load; |
c88d5910 | 1439 | int idx, this_cpu, prev_cpu; |
098fb9db | 1440 | unsigned long tl_per_task; |
c88d5910 | 1441 | struct task_group *tg; |
83378269 | 1442 | unsigned long weight; |
b3137bc8 | 1443 | int balanced; |
098fb9db | 1444 | |
c88d5910 PZ |
1445 | idx = sd->wake_idx; |
1446 | this_cpu = smp_processor_id(); | |
1447 | prev_cpu = task_cpu(p); | |
1448 | load = source_load(prev_cpu, idx); | |
1449 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1450 | |
b3137bc8 MG |
1451 | /* |
1452 | * If sync wakeup then subtract the (maximum possible) | |
1453 | * effect of the currently running task from the load | |
1454 | * of the current CPU: | |
1455 | */ | |
f3b577de | 1456 | rcu_read_lock(); |
83378269 PZ |
1457 | if (sync) { |
1458 | tg = task_group(current); | |
1459 | weight = current->se.load.weight; | |
1460 | ||
c88d5910 | 1461 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1462 | load += effective_load(tg, prev_cpu, 0, -weight); |
1463 | } | |
b3137bc8 | 1464 | |
83378269 PZ |
1465 | tg = task_group(p); |
1466 | weight = p->se.load.weight; | |
b3137bc8 | 1467 | |
71a29aa7 PZ |
1468 | /* |
1469 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1470 | * due to the sync cause above having dropped this_load to 0, we'll |
1471 | * always have an imbalance, but there's really nothing you can do | |
1472 | * about that, so that's good too. | |
71a29aa7 PZ |
1473 | * |
1474 | * Otherwise check if either cpus are near enough in load to allow this | |
1475 | * task to be woken on this_cpu. | |
1476 | */ | |
e37b6a7b PT |
1477 | if (this_load > 0) { |
1478 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
1479 | |
1480 | this_eff_load = 100; | |
1481 | this_eff_load *= power_of(prev_cpu); | |
1482 | this_eff_load *= this_load + | |
1483 | effective_load(tg, this_cpu, weight, weight); | |
1484 | ||
1485 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
1486 | prev_eff_load *= power_of(this_cpu); | |
1487 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
1488 | ||
1489 | balanced = this_eff_load <= prev_eff_load; | |
1490 | } else | |
1491 | balanced = true; | |
f3b577de | 1492 | rcu_read_unlock(); |
b3137bc8 | 1493 | |
098fb9db | 1494 | /* |
4ae7d5ce IM |
1495 | * If the currently running task will sleep within |
1496 | * a reasonable amount of time then attract this newly | |
1497 | * woken task: | |
098fb9db | 1498 | */ |
2fb7635c PZ |
1499 | if (sync && balanced) |
1500 | return 1; | |
098fb9db | 1501 | |
41acab88 | 1502 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
1503 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
1504 | ||
c88d5910 PZ |
1505 | if (balanced || |
1506 | (this_load <= load && | |
1507 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1508 | /* |
1509 | * This domain has SD_WAKE_AFFINE and | |
1510 | * p is cache cold in this domain, and | |
1511 | * there is no bad imbalance. | |
1512 | */ | |
c88d5910 | 1513 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 1514 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
1515 | |
1516 | return 1; | |
1517 | } | |
1518 | return 0; | |
1519 | } | |
1520 | ||
aaee1203 PZ |
1521 | /* |
1522 | * find_idlest_group finds and returns the least busy CPU group within the | |
1523 | * domain. | |
1524 | */ | |
1525 | static struct sched_group * | |
78e7ed53 | 1526 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1527 | int this_cpu, int load_idx) |
e7693a36 | 1528 | { |
b3bd3de6 | 1529 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 1530 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 1531 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1532 | |
aaee1203 PZ |
1533 | do { |
1534 | unsigned long load, avg_load; | |
1535 | int local_group; | |
1536 | int i; | |
e7693a36 | 1537 | |
aaee1203 PZ |
1538 | /* Skip over this group if it has no CPUs allowed */ |
1539 | if (!cpumask_intersects(sched_group_cpus(group), | |
1540 | &p->cpus_allowed)) | |
1541 | continue; | |
1542 | ||
1543 | local_group = cpumask_test_cpu(this_cpu, | |
1544 | sched_group_cpus(group)); | |
1545 | ||
1546 | /* Tally up the load of all CPUs in the group */ | |
1547 | avg_load = 0; | |
1548 | ||
1549 | for_each_cpu(i, sched_group_cpus(group)) { | |
1550 | /* Bias balancing toward cpus of our domain */ | |
1551 | if (local_group) | |
1552 | load = source_load(i, load_idx); | |
1553 | else | |
1554 | load = target_load(i, load_idx); | |
1555 | ||
1556 | avg_load += load; | |
1557 | } | |
1558 | ||
1559 | /* Adjust by relative CPU power of the group */ | |
1560 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1561 | ||
1562 | if (local_group) { | |
1563 | this_load = avg_load; | |
aaee1203 PZ |
1564 | } else if (avg_load < min_load) { |
1565 | min_load = avg_load; | |
1566 | idlest = group; | |
1567 | } | |
1568 | } while (group = group->next, group != sd->groups); | |
1569 | ||
1570 | if (!idlest || 100*this_load < imbalance*min_load) | |
1571 | return NULL; | |
1572 | return idlest; | |
1573 | } | |
1574 | ||
1575 | /* | |
1576 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1577 | */ | |
1578 | static int | |
1579 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1580 | { | |
1581 | unsigned long load, min_load = ULONG_MAX; | |
1582 | int idlest = -1; | |
1583 | int i; | |
1584 | ||
1585 | /* Traverse only the allowed CPUs */ | |
1586 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1587 | load = weighted_cpuload(i); | |
1588 | ||
1589 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1590 | min_load = load; | |
1591 | idlest = i; | |
e7693a36 GH |
1592 | } |
1593 | } | |
1594 | ||
aaee1203 PZ |
1595 | return idlest; |
1596 | } | |
e7693a36 | 1597 | |
a50bde51 PZ |
1598 | /* |
1599 | * Try and locate an idle CPU in the sched_domain. | |
1600 | */ | |
99bd5e2f | 1601 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
1602 | { |
1603 | int cpu = smp_processor_id(); | |
1604 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 1605 | struct sched_domain *sd; |
a50bde51 PZ |
1606 | int i; |
1607 | ||
1608 | /* | |
99bd5e2f SS |
1609 | * If the task is going to be woken-up on this cpu and if it is |
1610 | * already idle, then it is the right target. | |
a50bde51 | 1611 | */ |
99bd5e2f SS |
1612 | if (target == cpu && idle_cpu(cpu)) |
1613 | return cpu; | |
1614 | ||
1615 | /* | |
1616 | * If the task is going to be woken-up on the cpu where it previously | |
1617 | * ran and if it is currently idle, then it the right target. | |
1618 | */ | |
1619 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 1620 | return prev_cpu; |
a50bde51 PZ |
1621 | |
1622 | /* | |
99bd5e2f | 1623 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 1624 | */ |
99bd5e2f SS |
1625 | for_each_domain(target, sd) { |
1626 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 1627 | break; |
99bd5e2f SS |
1628 | |
1629 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1630 | if (idle_cpu(i)) { | |
1631 | target = i; | |
1632 | break; | |
1633 | } | |
a50bde51 | 1634 | } |
99bd5e2f SS |
1635 | |
1636 | /* | |
1637 | * Lets stop looking for an idle sibling when we reached | |
1638 | * the domain that spans the current cpu and prev_cpu. | |
1639 | */ | |
1640 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1641 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1642 | break; | |
a50bde51 PZ |
1643 | } |
1644 | ||
1645 | return target; | |
1646 | } | |
1647 | ||
aaee1203 PZ |
1648 | /* |
1649 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1650 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1651 | * SD_BALANCE_EXEC. | |
1652 | * | |
1653 | * Balance, ie. select the least loaded group. | |
1654 | * | |
1655 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1656 | * | |
1657 | * preempt must be disabled. | |
1658 | */ | |
0017d735 PZ |
1659 | static int |
1660 | select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags) | |
aaee1203 | 1661 | { |
29cd8bae | 1662 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1663 | int cpu = smp_processor_id(); |
1664 | int prev_cpu = task_cpu(p); | |
1665 | int new_cpu = cpu; | |
99bd5e2f | 1666 | int want_affine = 0; |
29cd8bae | 1667 | int want_sd = 1; |
5158f4e4 | 1668 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1669 | |
0763a660 | 1670 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 1671 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
1672 | want_affine = 1; |
1673 | new_cpu = prev_cpu; | |
1674 | } | |
aaee1203 PZ |
1675 | |
1676 | for_each_domain(cpu, tmp) { | |
e4f42888 PZ |
1677 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1678 | continue; | |
1679 | ||
aaee1203 | 1680 | /* |
ae154be1 PZ |
1681 | * If power savings logic is enabled for a domain, see if we |
1682 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1683 | */ |
59abf026 | 1684 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1685 | unsigned long power = 0; |
1686 | unsigned long nr_running = 0; | |
1687 | unsigned long capacity; | |
1688 | int i; | |
1689 | ||
1690 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1691 | power += power_of(i); | |
1692 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1693 | } | |
1694 | ||
1695 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1696 | ||
59abf026 PZ |
1697 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1698 | nr_running /= 2; | |
1699 | ||
1700 | if (nr_running < capacity) | |
29cd8bae | 1701 | want_sd = 0; |
ae154be1 | 1702 | } |
aaee1203 | 1703 | |
fe3bcfe1 | 1704 | /* |
99bd5e2f SS |
1705 | * If both cpu and prev_cpu are part of this domain, |
1706 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 1707 | */ |
99bd5e2f SS |
1708 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1709 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1710 | affine_sd = tmp; | |
1711 | want_affine = 0; | |
c88d5910 PZ |
1712 | } |
1713 | ||
29cd8bae PZ |
1714 | if (!want_sd && !want_affine) |
1715 | break; | |
1716 | ||
0763a660 | 1717 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1718 | continue; |
1719 | ||
29cd8bae PZ |
1720 | if (want_sd) |
1721 | sd = tmp; | |
1722 | } | |
1723 | ||
8b911acd | 1724 | if (affine_sd) { |
99bd5e2f SS |
1725 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
1726 | return select_idle_sibling(p, cpu); | |
1727 | else | |
1728 | return select_idle_sibling(p, prev_cpu); | |
8b911acd | 1729 | } |
e7693a36 | 1730 | |
aaee1203 | 1731 | while (sd) { |
5158f4e4 | 1732 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1733 | struct sched_group *group; |
c88d5910 | 1734 | int weight; |
098fb9db | 1735 | |
0763a660 | 1736 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1737 | sd = sd->child; |
1738 | continue; | |
1739 | } | |
098fb9db | 1740 | |
5158f4e4 PZ |
1741 | if (sd_flag & SD_BALANCE_WAKE) |
1742 | load_idx = sd->wake_idx; | |
098fb9db | 1743 | |
5158f4e4 | 1744 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1745 | if (!group) { |
1746 | sd = sd->child; | |
1747 | continue; | |
1748 | } | |
4ae7d5ce | 1749 | |
d7c33c49 | 1750 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1751 | if (new_cpu == -1 || new_cpu == cpu) { |
1752 | /* Now try balancing at a lower domain level of cpu */ | |
1753 | sd = sd->child; | |
1754 | continue; | |
e7693a36 | 1755 | } |
aaee1203 PZ |
1756 | |
1757 | /* Now try balancing at a lower domain level of new_cpu */ | |
1758 | cpu = new_cpu; | |
669c55e9 | 1759 | weight = sd->span_weight; |
aaee1203 PZ |
1760 | sd = NULL; |
1761 | for_each_domain(cpu, tmp) { | |
669c55e9 | 1762 | if (weight <= tmp->span_weight) |
aaee1203 | 1763 | break; |
0763a660 | 1764 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1765 | sd = tmp; |
1766 | } | |
1767 | /* while loop will break here if sd == NULL */ | |
e7693a36 GH |
1768 | } |
1769 | ||
c88d5910 | 1770 | return new_cpu; |
e7693a36 GH |
1771 | } |
1772 | #endif /* CONFIG_SMP */ | |
1773 | ||
e52fb7c0 PZ |
1774 | static unsigned long |
1775 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1776 | { |
1777 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1778 | ||
1779 | /* | |
e52fb7c0 PZ |
1780 | * Since its curr running now, convert the gran from real-time |
1781 | * to virtual-time in his units. | |
13814d42 MG |
1782 | * |
1783 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1784 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1785 | * the resulting gran will be larger, therefore penalizing the | |
1786 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1787 | * be smaller, again penalizing the lighter task. | |
1788 | * | |
1789 | * This is especially important for buddies when the leftmost | |
1790 | * task is higher priority than the buddy. | |
0bbd3336 | 1791 | */ |
13814d42 MG |
1792 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
1793 | gran = calc_delta_fair(gran, se); | |
0bbd3336 PZ |
1794 | |
1795 | return gran; | |
1796 | } | |
1797 | ||
464b7527 PZ |
1798 | /* |
1799 | * Should 'se' preempt 'curr'. | |
1800 | * | |
1801 | * |s1 | |
1802 | * |s2 | |
1803 | * |s3 | |
1804 | * g | |
1805 | * |<--->|c | |
1806 | * | |
1807 | * w(c, s1) = -1 | |
1808 | * w(c, s2) = 0 | |
1809 | * w(c, s3) = 1 | |
1810 | * | |
1811 | */ | |
1812 | static int | |
1813 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1814 | { | |
1815 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1816 | ||
1817 | if (vdiff <= 0) | |
1818 | return -1; | |
1819 | ||
e52fb7c0 | 1820 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1821 | if (vdiff > gran) |
1822 | return 1; | |
1823 | ||
1824 | return 0; | |
1825 | } | |
1826 | ||
02479099 PZ |
1827 | static void set_last_buddy(struct sched_entity *se) |
1828 | { | |
6bc912b7 PZ |
1829 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1830 | for_each_sched_entity(se) | |
1831 | cfs_rq_of(se)->last = se; | |
1832 | } | |
02479099 PZ |
1833 | } |
1834 | ||
1835 | static void set_next_buddy(struct sched_entity *se) | |
1836 | { | |
6bc912b7 PZ |
1837 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1838 | for_each_sched_entity(se) | |
1839 | cfs_rq_of(se)->next = se; | |
1840 | } | |
02479099 PZ |
1841 | } |
1842 | ||
ac53db59 RR |
1843 | static void set_skip_buddy(struct sched_entity *se) |
1844 | { | |
1845 | if (likely(task_of(se)->policy != SCHED_IDLE)) { | |
1846 | for_each_sched_entity(se) | |
1847 | cfs_rq_of(se)->skip = se; | |
1848 | } | |
1849 | } | |
1850 | ||
bf0f6f24 IM |
1851 | /* |
1852 | * Preempt the current task with a newly woken task if needed: | |
1853 | */ | |
5a9b86f6 | 1854 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1855 | { |
1856 | struct task_struct *curr = rq->curr; | |
8651a86c | 1857 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1858 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 1859 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
bf0f6f24 | 1860 | |
4ae7d5ce IM |
1861 | if (unlikely(se == pse)) |
1862 | return; | |
1863 | ||
f685ceac | 1864 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) |
3cb63d52 | 1865 | set_next_buddy(pse); |
57fdc26d | 1866 | |
aec0a514 BR |
1867 | /* |
1868 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1869 | * wake up path. | |
1870 | */ | |
1871 | if (test_tsk_need_resched(curr)) | |
1872 | return; | |
1873 | ||
a2f5c9ab DH |
1874 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1875 | if (unlikely(curr->policy == SCHED_IDLE) && | |
1876 | likely(p->policy != SCHED_IDLE)) | |
1877 | goto preempt; | |
1878 | ||
91c234b4 | 1879 | /* |
a2f5c9ab DH |
1880 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
1881 | * is driven by the tick): | |
91c234b4 | 1882 | */ |
6bc912b7 | 1883 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1884 | return; |
bf0f6f24 | 1885 | |
bf0f6f24 | 1886 | |
ad4b78bb PZ |
1887 | if (!sched_feat(WAKEUP_PREEMPT)) |
1888 | return; | |
1889 | ||
3a7e73a2 | 1890 | update_curr(cfs_rq); |
464b7527 | 1891 | find_matching_se(&se, &pse); |
002f128b | 1892 | BUG_ON(!pse); |
3a7e73a2 PZ |
1893 | if (wakeup_preempt_entity(se, pse) == 1) |
1894 | goto preempt; | |
464b7527 | 1895 | |
3a7e73a2 | 1896 | return; |
a65ac745 | 1897 | |
3a7e73a2 PZ |
1898 | preempt: |
1899 | resched_task(curr); | |
1900 | /* | |
1901 | * Only set the backward buddy when the current task is still | |
1902 | * on the rq. This can happen when a wakeup gets interleaved | |
1903 | * with schedule on the ->pre_schedule() or idle_balance() | |
1904 | * point, either of which can * drop the rq lock. | |
1905 | * | |
1906 | * Also, during early boot the idle thread is in the fair class, | |
1907 | * for obvious reasons its a bad idea to schedule back to it. | |
1908 | */ | |
1909 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1910 | return; | |
1911 | ||
1912 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1913 | set_last_buddy(se); | |
bf0f6f24 IM |
1914 | } |
1915 | ||
fb8d4724 | 1916 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1917 | { |
8f4d37ec | 1918 | struct task_struct *p; |
bf0f6f24 IM |
1919 | struct cfs_rq *cfs_rq = &rq->cfs; |
1920 | struct sched_entity *se; | |
1921 | ||
36ace27e | 1922 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
1923 | return NULL; |
1924 | ||
1925 | do { | |
9948f4b2 | 1926 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1927 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1928 | cfs_rq = group_cfs_rq(se); |
1929 | } while (cfs_rq); | |
1930 | ||
8f4d37ec PZ |
1931 | p = task_of(se); |
1932 | hrtick_start_fair(rq, p); | |
1933 | ||
1934 | return p; | |
bf0f6f24 IM |
1935 | } |
1936 | ||
1937 | /* | |
1938 | * Account for a descheduled task: | |
1939 | */ | |
31ee529c | 1940 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1941 | { |
1942 | struct sched_entity *se = &prev->se; | |
1943 | struct cfs_rq *cfs_rq; | |
1944 | ||
1945 | for_each_sched_entity(se) { | |
1946 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1947 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1948 | } |
1949 | } | |
1950 | ||
ac53db59 RR |
1951 | /* |
1952 | * sched_yield() is very simple | |
1953 | * | |
1954 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
1955 | */ | |
1956 | static void yield_task_fair(struct rq *rq) | |
1957 | { | |
1958 | struct task_struct *curr = rq->curr; | |
1959 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1960 | struct sched_entity *se = &curr->se; | |
1961 | ||
1962 | /* | |
1963 | * Are we the only task in the tree? | |
1964 | */ | |
1965 | if (unlikely(rq->nr_running == 1)) | |
1966 | return; | |
1967 | ||
1968 | clear_buddies(cfs_rq, se); | |
1969 | ||
1970 | if (curr->policy != SCHED_BATCH) { | |
1971 | update_rq_clock(rq); | |
1972 | /* | |
1973 | * Update run-time statistics of the 'current'. | |
1974 | */ | |
1975 | update_curr(cfs_rq); | |
1976 | } | |
1977 | ||
1978 | set_skip_buddy(se); | |
1979 | } | |
1980 | ||
d95f4122 MG |
1981 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
1982 | { | |
1983 | struct sched_entity *se = &p->se; | |
1984 | ||
1985 | if (!se->on_rq) | |
1986 | return false; | |
1987 | ||
1988 | /* Tell the scheduler that we'd really like pse to run next. */ | |
1989 | set_next_buddy(se); | |
1990 | ||
d95f4122 MG |
1991 | yield_task_fair(rq); |
1992 | ||
1993 | return true; | |
1994 | } | |
1995 | ||
681f3e68 | 1996 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1997 | /************************************************** |
1998 | * Fair scheduling class load-balancing methods: | |
1999 | */ | |
2000 | ||
1e3c88bd PZ |
2001 | /* |
2002 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
2003 | * Both runqueues must be locked. | |
2004 | */ | |
2005 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
2006 | struct rq *this_rq, int this_cpu) | |
2007 | { | |
2008 | deactivate_task(src_rq, p, 0); | |
2009 | set_task_cpu(p, this_cpu); | |
2010 | activate_task(this_rq, p, 0); | |
2011 | check_preempt_curr(this_rq, p, 0); | |
2012 | } | |
2013 | ||
2014 | /* | |
2015 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2016 | */ | |
2017 | static | |
2018 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
2019 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2020 | int *all_pinned) | |
2021 | { | |
2022 | int tsk_cache_hot = 0; | |
2023 | /* | |
2024 | * We do not migrate tasks that are: | |
2025 | * 1) running (obviously), or | |
2026 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2027 | * 3) are cache-hot on their current CPU. | |
2028 | */ | |
2029 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 2030 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
2031 | return 0; |
2032 | } | |
2033 | *all_pinned = 0; | |
2034 | ||
2035 | if (task_running(rq, p)) { | |
41acab88 | 2036 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
2037 | return 0; |
2038 | } | |
2039 | ||
2040 | /* | |
2041 | * Aggressive migration if: | |
2042 | * 1) task is cache cold, or | |
2043 | * 2) too many balance attempts have failed. | |
2044 | */ | |
2045 | ||
305e6835 | 2046 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
2047 | if (!tsk_cache_hot || |
2048 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
2049 | #ifdef CONFIG_SCHEDSTATS | |
2050 | if (tsk_cache_hot) { | |
2051 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 2052 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
2053 | } |
2054 | #endif | |
2055 | return 1; | |
2056 | } | |
2057 | ||
2058 | if (tsk_cache_hot) { | |
41acab88 | 2059 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
2060 | return 0; |
2061 | } | |
2062 | return 1; | |
2063 | } | |
2064 | ||
897c395f PZ |
2065 | /* |
2066 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
2067 | * part of active balancing operations within "domain". | |
2068 | * Returns 1 if successful and 0 otherwise. | |
2069 | * | |
2070 | * Called with both runqueues locked. | |
2071 | */ | |
2072 | static int | |
2073 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2074 | struct sched_domain *sd, enum cpu_idle_type idle) | |
2075 | { | |
2076 | struct task_struct *p, *n; | |
2077 | struct cfs_rq *cfs_rq; | |
2078 | int pinned = 0; | |
2079 | ||
2080 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
2081 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
2082 | ||
2083 | if (!can_migrate_task(p, busiest, this_cpu, | |
2084 | sd, idle, &pinned)) | |
2085 | continue; | |
2086 | ||
2087 | pull_task(busiest, p, this_rq, this_cpu); | |
2088 | /* | |
2089 | * Right now, this is only the second place pull_task() | |
2090 | * is called, so we can safely collect pull_task() | |
2091 | * stats here rather than inside pull_task(). | |
2092 | */ | |
2093 | schedstat_inc(sd, lb_gained[idle]); | |
2094 | return 1; | |
2095 | } | |
2096 | } | |
2097 | ||
2098 | return 0; | |
2099 | } | |
2100 | ||
1e3c88bd PZ |
2101 | static unsigned long |
2102 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2103 | unsigned long max_load_move, struct sched_domain *sd, | |
2104 | enum cpu_idle_type idle, int *all_pinned, | |
ee00e66f | 2105 | int *this_best_prio, struct cfs_rq *busiest_cfs_rq) |
1e3c88bd | 2106 | { |
b30aef17 | 2107 | int loops = 0, pulled = 0; |
1e3c88bd | 2108 | long rem_load_move = max_load_move; |
ee00e66f | 2109 | struct task_struct *p, *n; |
1e3c88bd PZ |
2110 | |
2111 | if (max_load_move == 0) | |
2112 | goto out; | |
2113 | ||
ee00e66f PZ |
2114 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
2115 | if (loops++ > sysctl_sched_nr_migrate) | |
2116 | break; | |
1e3c88bd | 2117 | |
ee00e66f | 2118 | if ((p->se.load.weight >> 1) > rem_load_move || |
b30aef17 KC |
2119 | !can_migrate_task(p, busiest, this_cpu, sd, idle, |
2120 | all_pinned)) | |
ee00e66f | 2121 | continue; |
1e3c88bd | 2122 | |
ee00e66f PZ |
2123 | pull_task(busiest, p, this_rq, this_cpu); |
2124 | pulled++; | |
2125 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
2126 | |
2127 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
2128 | /* |
2129 | * NEWIDLE balancing is a source of latency, so preemptible | |
2130 | * kernels will stop after the first task is pulled to minimize | |
2131 | * the critical section. | |
2132 | */ | |
2133 | if (idle == CPU_NEWLY_IDLE) | |
2134 | break; | |
1e3c88bd PZ |
2135 | #endif |
2136 | ||
ee00e66f PZ |
2137 | /* |
2138 | * We only want to steal up to the prescribed amount of | |
2139 | * weighted load. | |
2140 | */ | |
2141 | if (rem_load_move <= 0) | |
2142 | break; | |
2143 | ||
1e3c88bd PZ |
2144 | if (p->prio < *this_best_prio) |
2145 | *this_best_prio = p->prio; | |
1e3c88bd PZ |
2146 | } |
2147 | out: | |
2148 | /* | |
2149 | * Right now, this is one of only two places pull_task() is called, | |
2150 | * so we can safely collect pull_task() stats here rather than | |
2151 | * inside pull_task(). | |
2152 | */ | |
2153 | schedstat_add(sd, lb_gained[idle], pulled); | |
2154 | ||
1e3c88bd PZ |
2155 | return max_load_move - rem_load_move; |
2156 | } | |
2157 | ||
230059de | 2158 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
2159 | /* |
2160 | * update tg->load_weight by folding this cpu's load_avg | |
2161 | */ | |
67e86250 | 2162 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
2163 | { |
2164 | struct cfs_rq *cfs_rq; | |
2165 | unsigned long flags; | |
2166 | struct rq *rq; | |
9e3081ca PZ |
2167 | |
2168 | if (!tg->se[cpu]) | |
2169 | return 0; | |
2170 | ||
2171 | rq = cpu_rq(cpu); | |
2172 | cfs_rq = tg->cfs_rq[cpu]; | |
2173 | ||
2174 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2175 | ||
2176 | update_rq_clock(rq); | |
d6b55918 | 2177 | update_cfs_load(cfs_rq, 1); |
9e3081ca PZ |
2178 | |
2179 | /* | |
2180 | * We need to update shares after updating tg->load_weight in | |
2181 | * order to adjust the weight of groups with long running tasks. | |
2182 | */ | |
6d5ab293 | 2183 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
2184 | |
2185 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2186 | ||
2187 | return 0; | |
2188 | } | |
2189 | ||
2190 | static void update_shares(int cpu) | |
2191 | { | |
2192 | struct cfs_rq *cfs_rq; | |
2193 | struct rq *rq = cpu_rq(cpu); | |
2194 | ||
2195 | rcu_read_lock(); | |
67e86250 PT |
2196 | for_each_leaf_cfs_rq(rq, cfs_rq) |
2197 | update_shares_cpu(cfs_rq->tg, cpu); | |
9e3081ca PZ |
2198 | rcu_read_unlock(); |
2199 | } | |
2200 | ||
230059de PZ |
2201 | static unsigned long |
2202 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2203 | unsigned long max_load_move, | |
2204 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2205 | int *all_pinned, int *this_best_prio) | |
2206 | { | |
2207 | long rem_load_move = max_load_move; | |
2208 | int busiest_cpu = cpu_of(busiest); | |
2209 | struct task_group *tg; | |
2210 | ||
2211 | rcu_read_lock(); | |
2212 | update_h_load(busiest_cpu); | |
2213 | ||
2214 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
2215 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; | |
2216 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; | |
2217 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
2218 | u64 rem_load, moved_load; | |
2219 | ||
2220 | /* | |
2221 | * empty group | |
2222 | */ | |
2223 | if (!busiest_cfs_rq->task_weight) | |
2224 | continue; | |
2225 | ||
2226 | rem_load = (u64)rem_load_move * busiest_weight; | |
2227 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
2228 | ||
2229 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
2230 | rem_load, sd, idle, all_pinned, this_best_prio, | |
2231 | busiest_cfs_rq); | |
2232 | ||
2233 | if (!moved_load) | |
2234 | continue; | |
2235 | ||
2236 | moved_load *= busiest_h_load; | |
2237 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
2238 | ||
2239 | rem_load_move -= moved_load; | |
2240 | if (rem_load_move < 0) | |
2241 | break; | |
2242 | } | |
2243 | rcu_read_unlock(); | |
2244 | ||
2245 | return max_load_move - rem_load_move; | |
2246 | } | |
2247 | #else | |
9e3081ca PZ |
2248 | static inline void update_shares(int cpu) |
2249 | { | |
2250 | } | |
2251 | ||
230059de PZ |
2252 | static unsigned long |
2253 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2254 | unsigned long max_load_move, | |
2255 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2256 | int *all_pinned, int *this_best_prio) | |
2257 | { | |
2258 | return balance_tasks(this_rq, this_cpu, busiest, | |
2259 | max_load_move, sd, idle, all_pinned, | |
2260 | this_best_prio, &busiest->cfs); | |
2261 | } | |
2262 | #endif | |
2263 | ||
1e3c88bd PZ |
2264 | /* |
2265 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2266 | * this_rq, as part of a balancing operation within domain "sd". | |
2267 | * Returns 1 if successful and 0 otherwise. | |
2268 | * | |
2269 | * Called with both runqueues locked. | |
2270 | */ | |
2271 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2272 | unsigned long max_load_move, | |
2273 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2274 | int *all_pinned) | |
2275 | { | |
3d45fd80 | 2276 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2277 | int this_best_prio = this_rq->curr->prio; |
2278 | ||
2279 | do { | |
3d45fd80 | 2280 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd PZ |
2281 | max_load_move - total_load_moved, |
2282 | sd, idle, all_pinned, &this_best_prio); | |
3d45fd80 PZ |
2283 | |
2284 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2285 | |
2286 | #ifdef CONFIG_PREEMPT | |
2287 | /* | |
2288 | * NEWIDLE balancing is a source of latency, so preemptible | |
2289 | * kernels will stop after the first task is pulled to minimize | |
2290 | * the critical section. | |
2291 | */ | |
2292 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2293 | break; | |
baa8c110 PZ |
2294 | |
2295 | if (raw_spin_is_contended(&this_rq->lock) || | |
2296 | raw_spin_is_contended(&busiest->lock)) | |
2297 | break; | |
1e3c88bd | 2298 | #endif |
3d45fd80 | 2299 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2300 | |
2301 | return total_load_moved > 0; | |
2302 | } | |
2303 | ||
1e3c88bd PZ |
2304 | /********** Helpers for find_busiest_group ************************/ |
2305 | /* | |
2306 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2307 | * during load balancing. | |
2308 | */ | |
2309 | struct sd_lb_stats { | |
2310 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2311 | struct sched_group *this; /* Local group in this sd */ | |
2312 | unsigned long total_load; /* Total load of all groups in sd */ | |
2313 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2314 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2315 | ||
2316 | /** Statistics of this group */ | |
2317 | unsigned long this_load; | |
2318 | unsigned long this_load_per_task; | |
2319 | unsigned long this_nr_running; | |
fab47622 | 2320 | unsigned long this_has_capacity; |
aae6d3dd | 2321 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
2322 | |
2323 | /* Statistics of the busiest group */ | |
aae6d3dd | 2324 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
2325 | unsigned long max_load; |
2326 | unsigned long busiest_load_per_task; | |
2327 | unsigned long busiest_nr_running; | |
dd5feea1 | 2328 | unsigned long busiest_group_capacity; |
fab47622 | 2329 | unsigned long busiest_has_capacity; |
aae6d3dd | 2330 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
2331 | |
2332 | int group_imb; /* Is there imbalance in this sd */ | |
2333 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2334 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2335 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2336 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2337 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2338 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2339 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2340 | #endif | |
2341 | }; | |
2342 | ||
2343 | /* | |
2344 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2345 | */ | |
2346 | struct sg_lb_stats { | |
2347 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2348 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2349 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2350 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2351 | unsigned long group_capacity; | |
aae6d3dd SS |
2352 | unsigned long idle_cpus; |
2353 | unsigned long group_weight; | |
1e3c88bd | 2354 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 2355 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2356 | }; |
2357 | ||
2358 | /** | |
2359 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2360 | * @group: The group whose first cpu is to be returned. | |
2361 | */ | |
2362 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2363 | { | |
2364 | return cpumask_first(sched_group_cpus(group)); | |
2365 | } | |
2366 | ||
2367 | /** | |
2368 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2369 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2370 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2371 | */ | |
2372 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2373 | enum cpu_idle_type idle) | |
2374 | { | |
2375 | int load_idx; | |
2376 | ||
2377 | switch (idle) { | |
2378 | case CPU_NOT_IDLE: | |
2379 | load_idx = sd->busy_idx; | |
2380 | break; | |
2381 | ||
2382 | case CPU_NEWLY_IDLE: | |
2383 | load_idx = sd->newidle_idx; | |
2384 | break; | |
2385 | default: | |
2386 | load_idx = sd->idle_idx; | |
2387 | break; | |
2388 | } | |
2389 | ||
2390 | return load_idx; | |
2391 | } | |
2392 | ||
2393 | ||
2394 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2395 | /** | |
2396 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2397 | * the given sched_domain, during load balancing. | |
2398 | * | |
2399 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2400 | * @sds: Variable containing the statistics for sd. | |
2401 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2402 | */ | |
2403 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2404 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2405 | { | |
2406 | /* | |
2407 | * Busy processors will not participate in power savings | |
2408 | * balance. | |
2409 | */ | |
2410 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2411 | sds->power_savings_balance = 0; | |
2412 | else { | |
2413 | sds->power_savings_balance = 1; | |
2414 | sds->min_nr_running = ULONG_MAX; | |
2415 | sds->leader_nr_running = 0; | |
2416 | } | |
2417 | } | |
2418 | ||
2419 | /** | |
2420 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2421 | * sched_domain while performing load balancing. | |
2422 | * | |
2423 | * @group: sched_group belonging to the sched_domain under consideration. | |
2424 | * @sds: Variable containing the statistics of the sched_domain | |
2425 | * @local_group: Does group contain the CPU for which we're performing | |
2426 | * load balancing ? | |
2427 | * @sgs: Variable containing the statistics of the group. | |
2428 | */ | |
2429 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2430 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2431 | { | |
2432 | ||
2433 | if (!sds->power_savings_balance) | |
2434 | return; | |
2435 | ||
2436 | /* | |
2437 | * If the local group is idle or completely loaded | |
2438 | * no need to do power savings balance at this domain | |
2439 | */ | |
2440 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2441 | !sds->this_nr_running)) | |
2442 | sds->power_savings_balance = 0; | |
2443 | ||
2444 | /* | |
2445 | * If a group is already running at full capacity or idle, | |
2446 | * don't include that group in power savings calculations | |
2447 | */ | |
2448 | if (!sds->power_savings_balance || | |
2449 | sgs->sum_nr_running >= sgs->group_capacity || | |
2450 | !sgs->sum_nr_running) | |
2451 | return; | |
2452 | ||
2453 | /* | |
2454 | * Calculate the group which has the least non-idle load. | |
2455 | * This is the group from where we need to pick up the load | |
2456 | * for saving power | |
2457 | */ | |
2458 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2459 | (sgs->sum_nr_running == sds->min_nr_running && | |
2460 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2461 | sds->group_min = group; | |
2462 | sds->min_nr_running = sgs->sum_nr_running; | |
2463 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2464 | sgs->sum_nr_running; | |
2465 | } | |
2466 | ||
2467 | /* | |
2468 | * Calculate the group which is almost near its | |
2469 | * capacity but still has some space to pick up some load | |
2470 | * from other group and save more power | |
2471 | */ | |
2472 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2473 | return; | |
2474 | ||
2475 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2476 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2477 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2478 | sds->group_leader = group; | |
2479 | sds->leader_nr_running = sgs->sum_nr_running; | |
2480 | } | |
2481 | } | |
2482 | ||
2483 | /** | |
2484 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2485 | * @sds: Variable containing the statistics of the sched_domain | |
2486 | * under consideration. | |
2487 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2488 | * @imbalance: Variable to store the imbalance. | |
2489 | * | |
2490 | * Description: | |
2491 | * Check if we have potential to perform some power-savings balance. | |
2492 | * If yes, set the busiest group to be the least loaded group in the | |
2493 | * sched_domain, so that it's CPUs can be put to idle. | |
2494 | * | |
2495 | * Returns 1 if there is potential to perform power-savings balance. | |
2496 | * Else returns 0. | |
2497 | */ | |
2498 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2499 | int this_cpu, unsigned long *imbalance) | |
2500 | { | |
2501 | if (!sds->power_savings_balance) | |
2502 | return 0; | |
2503 | ||
2504 | if (sds->this != sds->group_leader || | |
2505 | sds->group_leader == sds->group_min) | |
2506 | return 0; | |
2507 | ||
2508 | *imbalance = sds->min_load_per_task; | |
2509 | sds->busiest = sds->group_min; | |
2510 | ||
2511 | return 1; | |
2512 | ||
2513 | } | |
2514 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2515 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2516 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2517 | { | |
2518 | return; | |
2519 | } | |
2520 | ||
2521 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2522 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2523 | { | |
2524 | return; | |
2525 | } | |
2526 | ||
2527 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2528 | int this_cpu, unsigned long *imbalance) | |
2529 | { | |
2530 | return 0; | |
2531 | } | |
2532 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2533 | ||
2534 | ||
2535 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2536 | { | |
2537 | return SCHED_LOAD_SCALE; | |
2538 | } | |
2539 | ||
2540 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2541 | { | |
2542 | return default_scale_freq_power(sd, cpu); | |
2543 | } | |
2544 | ||
2545 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2546 | { | |
669c55e9 | 2547 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2548 | unsigned long smt_gain = sd->smt_gain; |
2549 | ||
2550 | smt_gain /= weight; | |
2551 | ||
2552 | return smt_gain; | |
2553 | } | |
2554 | ||
2555 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2556 | { | |
2557 | return default_scale_smt_power(sd, cpu); | |
2558 | } | |
2559 | ||
2560 | unsigned long scale_rt_power(int cpu) | |
2561 | { | |
2562 | struct rq *rq = cpu_rq(cpu); | |
2563 | u64 total, available; | |
2564 | ||
1e3c88bd | 2565 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
2566 | |
2567 | if (unlikely(total < rq->rt_avg)) { | |
2568 | /* Ensures that power won't end up being negative */ | |
2569 | available = 0; | |
2570 | } else { | |
2571 | available = total - rq->rt_avg; | |
2572 | } | |
1e3c88bd PZ |
2573 | |
2574 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
2575 | total = SCHED_LOAD_SCALE; | |
2576 | ||
2577 | total >>= SCHED_LOAD_SHIFT; | |
2578 | ||
2579 | return div_u64(available, total); | |
2580 | } | |
2581 | ||
2582 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2583 | { | |
669c55e9 | 2584 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2585 | unsigned long power = SCHED_LOAD_SCALE; |
2586 | struct sched_group *sdg = sd->groups; | |
2587 | ||
1e3c88bd PZ |
2588 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
2589 | if (sched_feat(ARCH_POWER)) | |
2590 | power *= arch_scale_smt_power(sd, cpu); | |
2591 | else | |
2592 | power *= default_scale_smt_power(sd, cpu); | |
2593 | ||
2594 | power >>= SCHED_LOAD_SHIFT; | |
2595 | } | |
2596 | ||
9d5efe05 SV |
2597 | sdg->cpu_power_orig = power; |
2598 | ||
2599 | if (sched_feat(ARCH_POWER)) | |
2600 | power *= arch_scale_freq_power(sd, cpu); | |
2601 | else | |
2602 | power *= default_scale_freq_power(sd, cpu); | |
2603 | ||
2604 | power >>= SCHED_LOAD_SHIFT; | |
2605 | ||
1e3c88bd PZ |
2606 | power *= scale_rt_power(cpu); |
2607 | power >>= SCHED_LOAD_SHIFT; | |
2608 | ||
2609 | if (!power) | |
2610 | power = 1; | |
2611 | ||
e51fd5e2 | 2612 | cpu_rq(cpu)->cpu_power = power; |
1e3c88bd PZ |
2613 | sdg->cpu_power = power; |
2614 | } | |
2615 | ||
2616 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2617 | { | |
2618 | struct sched_domain *child = sd->child; | |
2619 | struct sched_group *group, *sdg = sd->groups; | |
2620 | unsigned long power; | |
2621 | ||
2622 | if (!child) { | |
2623 | update_cpu_power(sd, cpu); | |
2624 | return; | |
2625 | } | |
2626 | ||
2627 | power = 0; | |
2628 | ||
2629 | group = child->groups; | |
2630 | do { | |
2631 | power += group->cpu_power; | |
2632 | group = group->next; | |
2633 | } while (group != child->groups); | |
2634 | ||
2635 | sdg->cpu_power = power; | |
2636 | } | |
2637 | ||
9d5efe05 SV |
2638 | /* |
2639 | * Try and fix up capacity for tiny siblings, this is needed when | |
2640 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
2641 | * which on its own isn't powerful enough. | |
2642 | * | |
2643 | * See update_sd_pick_busiest() and check_asym_packing(). | |
2644 | */ | |
2645 | static inline int | |
2646 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
2647 | { | |
2648 | /* | |
2649 | * Only siblings can have significantly less than SCHED_LOAD_SCALE | |
2650 | */ | |
2651 | if (sd->level != SD_LV_SIBLING) | |
2652 | return 0; | |
2653 | ||
2654 | /* | |
2655 | * If ~90% of the cpu_power is still there, we're good. | |
2656 | */ | |
694f5a11 | 2657 | if (group->cpu_power * 32 > group->cpu_power_orig * 29) |
9d5efe05 SV |
2658 | return 1; |
2659 | ||
2660 | return 0; | |
2661 | } | |
2662 | ||
1e3c88bd PZ |
2663 | /** |
2664 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2665 | * @sd: The sched_domain whose statistics are to be updated. | |
2666 | * @group: sched_group whose statistics are to be updated. | |
2667 | * @this_cpu: Cpu for which load balance is currently performed. | |
2668 | * @idle: Idle status of this_cpu | |
2669 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
1e3c88bd PZ |
2670 | * @local_group: Does group contain this_cpu. |
2671 | * @cpus: Set of cpus considered for load balancing. | |
2672 | * @balance: Should we balance. | |
2673 | * @sgs: variable to hold the statistics for this group. | |
2674 | */ | |
2675 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2676 | struct sched_group *group, int this_cpu, | |
46e49b38 | 2677 | enum cpu_idle_type idle, int load_idx, |
1e3c88bd PZ |
2678 | int local_group, const struct cpumask *cpus, |
2679 | int *balance, struct sg_lb_stats *sgs) | |
2680 | { | |
2582f0eb | 2681 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
2682 | int i; |
2683 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 2684 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 2685 | |
871e35bc | 2686 | if (local_group) |
1e3c88bd | 2687 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
2688 | |
2689 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
2690 | max_cpu_load = 0; |
2691 | min_cpu_load = ~0UL; | |
2582f0eb | 2692 | max_nr_running = 0; |
1e3c88bd PZ |
2693 | |
2694 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2695 | struct rq *rq = cpu_rq(i); | |
2696 | ||
1e3c88bd PZ |
2697 | /* Bias balancing toward cpus of our domain */ |
2698 | if (local_group) { | |
2699 | if (idle_cpu(i) && !first_idle_cpu) { | |
2700 | first_idle_cpu = 1; | |
2701 | balance_cpu = i; | |
2702 | } | |
2703 | ||
2704 | load = target_load(i, load_idx); | |
2705 | } else { | |
2706 | load = source_load(i, load_idx); | |
2582f0eb | 2707 | if (load > max_cpu_load) { |
1e3c88bd | 2708 | max_cpu_load = load; |
2582f0eb NR |
2709 | max_nr_running = rq->nr_running; |
2710 | } | |
1e3c88bd PZ |
2711 | if (min_cpu_load > load) |
2712 | min_cpu_load = load; | |
2713 | } | |
2714 | ||
2715 | sgs->group_load += load; | |
2716 | sgs->sum_nr_running += rq->nr_running; | |
2717 | sgs->sum_weighted_load += weighted_cpuload(i); | |
aae6d3dd SS |
2718 | if (idle_cpu(i)) |
2719 | sgs->idle_cpus++; | |
1e3c88bd PZ |
2720 | } |
2721 | ||
2722 | /* | |
2723 | * First idle cpu or the first cpu(busiest) in this sched group | |
2724 | * is eligible for doing load balancing at this and above | |
2725 | * domains. In the newly idle case, we will allow all the cpu's | |
2726 | * to do the newly idle load balance. | |
2727 | */ | |
bbc8cb5b PZ |
2728 | if (idle != CPU_NEWLY_IDLE && local_group) { |
2729 | if (balance_cpu != this_cpu) { | |
2730 | *balance = 0; | |
2731 | return; | |
2732 | } | |
2733 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
2734 | } |
2735 | ||
2736 | /* Adjust by relative CPU power of the group */ | |
2737 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2738 | ||
1e3c88bd PZ |
2739 | /* |
2740 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 2741 | * than the average weight of a task. |
1e3c88bd PZ |
2742 | * |
2743 | * APZ: with cgroup the avg task weight can vary wildly and | |
2744 | * might not be a suitable number - should we keep a | |
2745 | * normalized nr_running number somewhere that negates | |
2746 | * the hierarchy? | |
2747 | */ | |
dd5feea1 SS |
2748 | if (sgs->sum_nr_running) |
2749 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 2750 | |
866ab43e | 2751 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
2752 | sgs->group_imb = 1; |
2753 | ||
2582f0eb | 2754 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); |
9d5efe05 SV |
2755 | if (!sgs->group_capacity) |
2756 | sgs->group_capacity = fix_small_capacity(sd, group); | |
aae6d3dd | 2757 | sgs->group_weight = group->group_weight; |
fab47622 NR |
2758 | |
2759 | if (sgs->group_capacity > sgs->sum_nr_running) | |
2760 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
2761 | } |
2762 | ||
532cb4c4 MN |
2763 | /** |
2764 | * update_sd_pick_busiest - return 1 on busiest group | |
2765 | * @sd: sched_domain whose statistics are to be checked | |
2766 | * @sds: sched_domain statistics | |
2767 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
2768 | * @sgs: sched_group statistics |
2769 | * @this_cpu: the current cpu | |
532cb4c4 MN |
2770 | * |
2771 | * Determine if @sg is a busier group than the previously selected | |
2772 | * busiest group. | |
2773 | */ | |
2774 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
2775 | struct sd_lb_stats *sds, | |
2776 | struct sched_group *sg, | |
2777 | struct sg_lb_stats *sgs, | |
2778 | int this_cpu) | |
2779 | { | |
2780 | if (sgs->avg_load <= sds->max_load) | |
2781 | return false; | |
2782 | ||
2783 | if (sgs->sum_nr_running > sgs->group_capacity) | |
2784 | return true; | |
2785 | ||
2786 | if (sgs->group_imb) | |
2787 | return true; | |
2788 | ||
2789 | /* | |
2790 | * ASYM_PACKING needs to move all the work to the lowest | |
2791 | * numbered CPUs in the group, therefore mark all groups | |
2792 | * higher than ourself as busy. | |
2793 | */ | |
2794 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
2795 | this_cpu < group_first_cpu(sg)) { | |
2796 | if (!sds->busiest) | |
2797 | return true; | |
2798 | ||
2799 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
2800 | return true; | |
2801 | } | |
2802 | ||
2803 | return false; | |
2804 | } | |
2805 | ||
1e3c88bd PZ |
2806 | /** |
2807 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2808 | * @sd: sched_domain whose statistics are to be updated. | |
2809 | * @this_cpu: Cpu for which load balance is currently performed. | |
2810 | * @idle: Idle status of this_cpu | |
1e3c88bd PZ |
2811 | * @cpus: Set of cpus considered for load balancing. |
2812 | * @balance: Should we balance. | |
2813 | * @sds: variable to hold the statistics for this sched_domain. | |
2814 | */ | |
2815 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
46e49b38 VP |
2816 | enum cpu_idle_type idle, const struct cpumask *cpus, |
2817 | int *balance, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
2818 | { |
2819 | struct sched_domain *child = sd->child; | |
532cb4c4 | 2820 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
2821 | struct sg_lb_stats sgs; |
2822 | int load_idx, prefer_sibling = 0; | |
2823 | ||
2824 | if (child && child->flags & SD_PREFER_SIBLING) | |
2825 | prefer_sibling = 1; | |
2826 | ||
2827 | init_sd_power_savings_stats(sd, sds, idle); | |
2828 | load_idx = get_sd_load_idx(sd, idle); | |
2829 | ||
2830 | do { | |
2831 | int local_group; | |
2832 | ||
532cb4c4 | 2833 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 2834 | memset(&sgs, 0, sizeof(sgs)); |
46e49b38 | 2835 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, |
1e3c88bd PZ |
2836 | local_group, cpus, balance, &sgs); |
2837 | ||
8f190fb3 | 2838 | if (local_group && !(*balance)) |
1e3c88bd PZ |
2839 | return; |
2840 | ||
2841 | sds->total_load += sgs.group_load; | |
532cb4c4 | 2842 | sds->total_pwr += sg->cpu_power; |
1e3c88bd PZ |
2843 | |
2844 | /* | |
2845 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 2846 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
2847 | * and move all the excess tasks away. We lower the capacity |
2848 | * of a group only if the local group has the capacity to fit | |
2849 | * these excess tasks, i.e. nr_running < group_capacity. The | |
2850 | * extra check prevents the case where you always pull from the | |
2851 | * heaviest group when it is already under-utilized (possible | |
2852 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 2853 | */ |
75dd321d | 2854 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
2855 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
2856 | ||
2857 | if (local_group) { | |
2858 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 2859 | sds->this = sg; |
1e3c88bd PZ |
2860 | sds->this_nr_running = sgs.sum_nr_running; |
2861 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 2862 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 2863 | sds->this_idle_cpus = sgs.idle_cpus; |
532cb4c4 | 2864 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 2865 | sds->max_load = sgs.avg_load; |
532cb4c4 | 2866 | sds->busiest = sg; |
1e3c88bd | 2867 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 2868 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 2869 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 2870 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 2871 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 2872 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
2873 | sds->group_imb = sgs.group_imb; |
2874 | } | |
2875 | ||
532cb4c4 MN |
2876 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
2877 | sg = sg->next; | |
2878 | } while (sg != sd->groups); | |
2879 | } | |
2880 | ||
2ec57d44 | 2881 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
2882 | { |
2883 | return 0*SD_ASYM_PACKING; | |
2884 | } | |
2885 | ||
2886 | /** | |
2887 | * check_asym_packing - Check to see if the group is packed into the | |
2888 | * sched doman. | |
2889 | * | |
2890 | * This is primarily intended to used at the sibling level. Some | |
2891 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
2892 | * case of POWER7, it can move to lower SMT modes only when higher | |
2893 | * threads are idle. When in lower SMT modes, the threads will | |
2894 | * perform better since they share less core resources. Hence when we | |
2895 | * have idle threads, we want them to be the higher ones. | |
2896 | * | |
2897 | * This packing function is run on idle threads. It checks to see if | |
2898 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
2899 | * CPU number than the packing function is being run on. Here we are | |
2900 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
2901 | * number. | |
2902 | * | |
b6b12294 MN |
2903 | * Returns 1 when packing is required and a task should be moved to |
2904 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
2905 | * | |
532cb4c4 MN |
2906 | * @sd: The sched_domain whose packing is to be checked. |
2907 | * @sds: Statistics of the sched_domain which is to be packed | |
2908 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2909 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
2910 | */ |
2911 | static int check_asym_packing(struct sched_domain *sd, | |
2912 | struct sd_lb_stats *sds, | |
2913 | int this_cpu, unsigned long *imbalance) | |
2914 | { | |
2915 | int busiest_cpu; | |
2916 | ||
2917 | if (!(sd->flags & SD_ASYM_PACKING)) | |
2918 | return 0; | |
2919 | ||
2920 | if (!sds->busiest) | |
2921 | return 0; | |
2922 | ||
2923 | busiest_cpu = group_first_cpu(sds->busiest); | |
2924 | if (this_cpu > busiest_cpu) | |
2925 | return 0; | |
2926 | ||
2927 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power, | |
2928 | SCHED_LOAD_SCALE); | |
2929 | return 1; | |
1e3c88bd PZ |
2930 | } |
2931 | ||
2932 | /** | |
2933 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2934 | * amongst the groups of a sched_domain, during | |
2935 | * load balancing. | |
2936 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2937 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2938 | * @imbalance: Variable to store the imbalance. | |
2939 | */ | |
2940 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2941 | int this_cpu, unsigned long *imbalance) | |
2942 | { | |
2943 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2944 | unsigned int imbn = 2; | |
dd5feea1 | 2945 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
2946 | |
2947 | if (sds->this_nr_running) { | |
2948 | sds->this_load_per_task /= sds->this_nr_running; | |
2949 | if (sds->busiest_load_per_task > | |
2950 | sds->this_load_per_task) | |
2951 | imbn = 1; | |
2952 | } else | |
2953 | sds->this_load_per_task = | |
2954 | cpu_avg_load_per_task(this_cpu); | |
2955 | ||
dd5feea1 SS |
2956 | scaled_busy_load_per_task = sds->busiest_load_per_task |
2957 | * SCHED_LOAD_SCALE; | |
2958 | scaled_busy_load_per_task /= sds->busiest->cpu_power; | |
2959 | ||
2960 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
2961 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
2962 | *imbalance = sds->busiest_load_per_task; |
2963 | return; | |
2964 | } | |
2965 | ||
2966 | /* | |
2967 | * OK, we don't have enough imbalance to justify moving tasks, | |
2968 | * however we may be able to increase total CPU power used by | |
2969 | * moving them. | |
2970 | */ | |
2971 | ||
2972 | pwr_now += sds->busiest->cpu_power * | |
2973 | min(sds->busiest_load_per_task, sds->max_load); | |
2974 | pwr_now += sds->this->cpu_power * | |
2975 | min(sds->this_load_per_task, sds->this_load); | |
2976 | pwr_now /= SCHED_LOAD_SCALE; | |
2977 | ||
2978 | /* Amount of load we'd subtract */ | |
2979 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2980 | sds->busiest->cpu_power; | |
2981 | if (sds->max_load > tmp) | |
2982 | pwr_move += sds->busiest->cpu_power * | |
2983 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
2984 | ||
2985 | /* Amount of load we'd add */ | |
2986 | if (sds->max_load * sds->busiest->cpu_power < | |
2987 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
2988 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
2989 | sds->this->cpu_power; | |
2990 | else | |
2991 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2992 | sds->this->cpu_power; | |
2993 | pwr_move += sds->this->cpu_power * | |
2994 | min(sds->this_load_per_task, sds->this_load + tmp); | |
2995 | pwr_move /= SCHED_LOAD_SCALE; | |
2996 | ||
2997 | /* Move if we gain throughput */ | |
2998 | if (pwr_move > pwr_now) | |
2999 | *imbalance = sds->busiest_load_per_task; | |
3000 | } | |
3001 | ||
3002 | /** | |
3003 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
3004 | * groups of a given sched_domain during load balance. | |
3005 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
3006 | * @this_cpu: Cpu for which currently load balance is being performed. | |
3007 | * @imbalance: The variable to store the imbalance. | |
3008 | */ | |
3009 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
3010 | unsigned long *imbalance) | |
3011 | { | |
dd5feea1 SS |
3012 | unsigned long max_pull, load_above_capacity = ~0UL; |
3013 | ||
3014 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
3015 | if (sds->group_imb) { | |
3016 | sds->busiest_load_per_task = | |
3017 | min(sds->busiest_load_per_task, sds->avg_load); | |
3018 | } | |
3019 | ||
1e3c88bd PZ |
3020 | /* |
3021 | * In the presence of smp nice balancing, certain scenarios can have | |
3022 | * max load less than avg load(as we skip the groups at or below | |
3023 | * its cpu_power, while calculating max_load..) | |
3024 | */ | |
3025 | if (sds->max_load < sds->avg_load) { | |
3026 | *imbalance = 0; | |
3027 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3028 | } | |
3029 | ||
dd5feea1 SS |
3030 | if (!sds->group_imb) { |
3031 | /* | |
3032 | * Don't want to pull so many tasks that a group would go idle. | |
3033 | */ | |
3034 | load_above_capacity = (sds->busiest_nr_running - | |
3035 | sds->busiest_group_capacity); | |
3036 | ||
3037 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE); | |
3038 | ||
3039 | load_above_capacity /= sds->busiest->cpu_power; | |
3040 | } | |
3041 | ||
3042 | /* | |
3043 | * We're trying to get all the cpus to the average_load, so we don't | |
3044 | * want to push ourselves above the average load, nor do we wish to | |
3045 | * reduce the max loaded cpu below the average load. At the same time, | |
3046 | * we also don't want to reduce the group load below the group capacity | |
3047 | * (so that we can implement power-savings policies etc). Thus we look | |
3048 | * for the minimum possible imbalance. | |
3049 | * Be careful of negative numbers as they'll appear as very large values | |
3050 | * with unsigned longs. | |
3051 | */ | |
3052 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
3053 | |
3054 | /* How much load to actually move to equalise the imbalance */ | |
3055 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
3056 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
3057 | / SCHED_LOAD_SCALE; | |
3058 | ||
3059 | /* | |
3060 | * if *imbalance is less than the average load per runnable task | |
25985edc | 3061 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
3062 | * a think about bumping its value to force at least one task to be |
3063 | * moved | |
3064 | */ | |
3065 | if (*imbalance < sds->busiest_load_per_task) | |
3066 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3067 | ||
3068 | } | |
fab47622 | 3069 | |
1e3c88bd PZ |
3070 | /******* find_busiest_group() helpers end here *********************/ |
3071 | ||
3072 | /** | |
3073 | * find_busiest_group - Returns the busiest group within the sched_domain | |
3074 | * if there is an imbalance. If there isn't an imbalance, and | |
3075 | * the user has opted for power-savings, it returns a group whose | |
3076 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
3077 | * such a group exists. | |
3078 | * | |
3079 | * Also calculates the amount of weighted load which should be moved | |
3080 | * to restore balance. | |
3081 | * | |
3082 | * @sd: The sched_domain whose busiest group is to be returned. | |
3083 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
3084 | * @imbalance: Variable which stores amount of weighted load which should | |
3085 | * be moved to restore balance/put a group to idle. | |
3086 | * @idle: The idle status of this_cpu. | |
1e3c88bd PZ |
3087 | * @cpus: The set of CPUs under consideration for load-balancing. |
3088 | * @balance: Pointer to a variable indicating if this_cpu | |
3089 | * is the appropriate cpu to perform load balancing at this_level. | |
3090 | * | |
3091 | * Returns: - the busiest group if imbalance exists. | |
3092 | * - If no imbalance and user has opted for power-savings balance, | |
3093 | * return the least loaded group whose CPUs can be | |
3094 | * put to idle by rebalancing its tasks onto our group. | |
3095 | */ | |
3096 | static struct sched_group * | |
3097 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
3098 | unsigned long *imbalance, enum cpu_idle_type idle, | |
46e49b38 | 3099 | const struct cpumask *cpus, int *balance) |
1e3c88bd PZ |
3100 | { |
3101 | struct sd_lb_stats sds; | |
3102 | ||
3103 | memset(&sds, 0, sizeof(sds)); | |
3104 | ||
3105 | /* | |
3106 | * Compute the various statistics relavent for load balancing at | |
3107 | * this level. | |
3108 | */ | |
46e49b38 | 3109 | update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds); |
1e3c88bd | 3110 | |
cc57aa8f PZ |
3111 | /* |
3112 | * this_cpu is not the appropriate cpu to perform load balancing at | |
3113 | * this level. | |
1e3c88bd | 3114 | */ |
8f190fb3 | 3115 | if (!(*balance)) |
1e3c88bd PZ |
3116 | goto ret; |
3117 | ||
532cb4c4 MN |
3118 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
3119 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
3120 | return sds.busiest; | |
3121 | ||
cc57aa8f | 3122 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
3123 | if (!sds.busiest || sds.busiest_nr_running == 0) |
3124 | goto out_balanced; | |
3125 | ||
b0432d8f KC |
3126 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; |
3127 | ||
866ab43e PZ |
3128 | /* |
3129 | * If the busiest group is imbalanced the below checks don't | |
3130 | * work because they assumes all things are equal, which typically | |
3131 | * isn't true due to cpus_allowed constraints and the like. | |
3132 | */ | |
3133 | if (sds.group_imb) | |
3134 | goto force_balance; | |
3135 | ||
cc57aa8f | 3136 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
fab47622 NR |
3137 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
3138 | !sds.busiest_has_capacity) | |
3139 | goto force_balance; | |
3140 | ||
cc57aa8f PZ |
3141 | /* |
3142 | * If the local group is more busy than the selected busiest group | |
3143 | * don't try and pull any tasks. | |
3144 | */ | |
1e3c88bd PZ |
3145 | if (sds.this_load >= sds.max_load) |
3146 | goto out_balanced; | |
3147 | ||
cc57aa8f PZ |
3148 | /* |
3149 | * Don't pull any tasks if this group is already above the domain | |
3150 | * average load. | |
3151 | */ | |
1e3c88bd PZ |
3152 | if (sds.this_load >= sds.avg_load) |
3153 | goto out_balanced; | |
3154 | ||
c186fafe | 3155 | if (idle == CPU_IDLE) { |
aae6d3dd SS |
3156 | /* |
3157 | * This cpu is idle. If the busiest group load doesn't | |
3158 | * have more tasks than the number of available cpu's and | |
3159 | * there is no imbalance between this and busiest group | |
3160 | * wrt to idle cpu's, it is balanced. | |
3161 | */ | |
c186fafe | 3162 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
3163 | sds.busiest_nr_running <= sds.busiest_group_weight) |
3164 | goto out_balanced; | |
c186fafe PZ |
3165 | } else { |
3166 | /* | |
3167 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
3168 | * imbalance_pct to be conservative. | |
3169 | */ | |
3170 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
3171 | goto out_balanced; | |
aae6d3dd | 3172 | } |
1e3c88bd | 3173 | |
fab47622 | 3174 | force_balance: |
1e3c88bd PZ |
3175 | /* Looks like there is an imbalance. Compute it */ |
3176 | calculate_imbalance(&sds, this_cpu, imbalance); | |
3177 | return sds.busiest; | |
3178 | ||
3179 | out_balanced: | |
3180 | /* | |
3181 | * There is no obvious imbalance. But check if we can do some balancing | |
3182 | * to save power. | |
3183 | */ | |
3184 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
3185 | return sds.busiest; | |
3186 | ret: | |
3187 | *imbalance = 0; | |
3188 | return NULL; | |
3189 | } | |
3190 | ||
3191 | /* | |
3192 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3193 | */ | |
3194 | static struct rq * | |
9d5efe05 SV |
3195 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
3196 | enum cpu_idle_type idle, unsigned long imbalance, | |
3197 | const struct cpumask *cpus) | |
1e3c88bd PZ |
3198 | { |
3199 | struct rq *busiest = NULL, *rq; | |
3200 | unsigned long max_load = 0; | |
3201 | int i; | |
3202 | ||
3203 | for_each_cpu(i, sched_group_cpus(group)) { | |
3204 | unsigned long power = power_of(i); | |
3205 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
3206 | unsigned long wl; | |
3207 | ||
9d5efe05 SV |
3208 | if (!capacity) |
3209 | capacity = fix_small_capacity(sd, group); | |
3210 | ||
1e3c88bd PZ |
3211 | if (!cpumask_test_cpu(i, cpus)) |
3212 | continue; | |
3213 | ||
3214 | rq = cpu_rq(i); | |
6e40f5bb | 3215 | wl = weighted_cpuload(i); |
1e3c88bd | 3216 | |
6e40f5bb TG |
3217 | /* |
3218 | * When comparing with imbalance, use weighted_cpuload() | |
3219 | * which is not scaled with the cpu power. | |
3220 | */ | |
1e3c88bd PZ |
3221 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
3222 | continue; | |
3223 | ||
6e40f5bb TG |
3224 | /* |
3225 | * For the load comparisons with the other cpu's, consider | |
3226 | * the weighted_cpuload() scaled with the cpu power, so that | |
3227 | * the load can be moved away from the cpu that is potentially | |
3228 | * running at a lower capacity. | |
3229 | */ | |
3230 | wl = (wl * SCHED_LOAD_SCALE) / power; | |
3231 | ||
1e3c88bd PZ |
3232 | if (wl > max_load) { |
3233 | max_load = wl; | |
3234 | busiest = rq; | |
3235 | } | |
3236 | } | |
3237 | ||
3238 | return busiest; | |
3239 | } | |
3240 | ||
3241 | /* | |
3242 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3243 | * so long as it is large enough. | |
3244 | */ | |
3245 | #define MAX_PINNED_INTERVAL 512 | |
3246 | ||
3247 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
3248 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
3249 | ||
46e49b38 | 3250 | static int need_active_balance(struct sched_domain *sd, int idle, |
532cb4c4 | 3251 | int busiest_cpu, int this_cpu) |
1af3ed3d PZ |
3252 | { |
3253 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
3254 | |
3255 | /* | |
3256 | * ASYM_PACKING needs to force migrate tasks from busy but | |
3257 | * higher numbered CPUs in order to pack all tasks in the | |
3258 | * lowest numbered CPUs. | |
3259 | */ | |
3260 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
3261 | return 1; | |
3262 | ||
1af3ed3d PZ |
3263 | /* |
3264 | * The only task running in a non-idle cpu can be moved to this | |
3265 | * cpu in an attempt to completely freeup the other CPU | |
3266 | * package. | |
3267 | * | |
3268 | * The package power saving logic comes from | |
3269 | * find_busiest_group(). If there are no imbalance, then | |
3270 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3271 | * f_b_g() will select a group from which a running task may be | |
3272 | * pulled to this cpu in order to make the other package idle. | |
3273 | * If there is no opportunity to make a package idle and if | |
3274 | * there are no imbalance, then f_b_g() will return NULL and no | |
3275 | * action will be taken in load_balance_newidle(). | |
3276 | * | |
3277 | * Under normal task pull operation due to imbalance, there | |
3278 | * will be more than one task in the source run queue and | |
3279 | * move_tasks() will succeed. ld_moved will be true and this | |
3280 | * active balance code will not be triggered. | |
3281 | */ | |
1af3ed3d PZ |
3282 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) |
3283 | return 0; | |
3284 | } | |
3285 | ||
3286 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3287 | } | |
3288 | ||
969c7921 TH |
3289 | static int active_load_balance_cpu_stop(void *data); |
3290 | ||
1e3c88bd PZ |
3291 | /* |
3292 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3293 | * tasks if there is an imbalance. | |
3294 | */ | |
3295 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3296 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3297 | int *balance) | |
3298 | { | |
46e49b38 | 3299 | int ld_moved, all_pinned = 0, active_balance = 0; |
1e3c88bd PZ |
3300 | struct sched_group *group; |
3301 | unsigned long imbalance; | |
3302 | struct rq *busiest; | |
3303 | unsigned long flags; | |
3304 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3305 | ||
3306 | cpumask_copy(cpus, cpu_active_mask); | |
3307 | ||
1e3c88bd PZ |
3308 | schedstat_inc(sd, lb_count[idle]); |
3309 | ||
3310 | redo: | |
46e49b38 | 3311 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, |
1e3c88bd PZ |
3312 | cpus, balance); |
3313 | ||
3314 | if (*balance == 0) | |
3315 | goto out_balanced; | |
3316 | ||
3317 | if (!group) { | |
3318 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3319 | goto out_balanced; | |
3320 | } | |
3321 | ||
9d5efe05 | 3322 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3323 | if (!busiest) { |
3324 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3325 | goto out_balanced; | |
3326 | } | |
3327 | ||
3328 | BUG_ON(busiest == this_rq); | |
3329 | ||
3330 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3331 | ||
3332 | ld_moved = 0; | |
3333 | if (busiest->nr_running > 1) { | |
3334 | /* | |
3335 | * Attempt to move tasks. If find_busiest_group has found | |
3336 | * an imbalance but busiest->nr_running <= 1, the group is | |
3337 | * still unbalanced. ld_moved simply stays zero, so it is | |
3338 | * correctly treated as an imbalance. | |
3339 | */ | |
b30aef17 | 3340 | all_pinned = 1; |
1e3c88bd PZ |
3341 | local_irq_save(flags); |
3342 | double_rq_lock(this_rq, busiest); | |
3343 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3344 | imbalance, sd, idle, &all_pinned); | |
3345 | double_rq_unlock(this_rq, busiest); | |
3346 | local_irq_restore(flags); | |
3347 | ||
3348 | /* | |
3349 | * some other cpu did the load balance for us. | |
3350 | */ | |
3351 | if (ld_moved && this_cpu != smp_processor_id()) | |
3352 | resched_cpu(this_cpu); | |
3353 | ||
3354 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3355 | if (unlikely(all_pinned)) { | |
3356 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3357 | if (!cpumask_empty(cpus)) | |
3358 | goto redo; | |
3359 | goto out_balanced; | |
3360 | } | |
3361 | } | |
3362 | ||
3363 | if (!ld_moved) { | |
3364 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3365 | /* |
3366 | * Increment the failure counter only on periodic balance. | |
3367 | * We do not want newidle balance, which can be very | |
3368 | * frequent, pollute the failure counter causing | |
3369 | * excessive cache_hot migrations and active balances. | |
3370 | */ | |
3371 | if (idle != CPU_NEWLY_IDLE) | |
3372 | sd->nr_balance_failed++; | |
1e3c88bd | 3373 | |
46e49b38 | 3374 | if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) { |
1e3c88bd PZ |
3375 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3376 | ||
969c7921 TH |
3377 | /* don't kick the active_load_balance_cpu_stop, |
3378 | * if the curr task on busiest cpu can't be | |
3379 | * moved to this_cpu | |
1e3c88bd PZ |
3380 | */ |
3381 | if (!cpumask_test_cpu(this_cpu, | |
3382 | &busiest->curr->cpus_allowed)) { | |
3383 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3384 | flags); | |
3385 | all_pinned = 1; | |
3386 | goto out_one_pinned; | |
3387 | } | |
3388 | ||
969c7921 TH |
3389 | /* |
3390 | * ->active_balance synchronizes accesses to | |
3391 | * ->active_balance_work. Once set, it's cleared | |
3392 | * only after active load balance is finished. | |
3393 | */ | |
1e3c88bd PZ |
3394 | if (!busiest->active_balance) { |
3395 | busiest->active_balance = 1; | |
3396 | busiest->push_cpu = this_cpu; | |
3397 | active_balance = 1; | |
3398 | } | |
3399 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 3400 | |
1e3c88bd | 3401 | if (active_balance) |
969c7921 TH |
3402 | stop_one_cpu_nowait(cpu_of(busiest), |
3403 | active_load_balance_cpu_stop, busiest, | |
3404 | &busiest->active_balance_work); | |
1e3c88bd PZ |
3405 | |
3406 | /* | |
3407 | * We've kicked active balancing, reset the failure | |
3408 | * counter. | |
3409 | */ | |
3410 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3411 | } | |
3412 | } else | |
3413 | sd->nr_balance_failed = 0; | |
3414 | ||
3415 | if (likely(!active_balance)) { | |
3416 | /* We were unbalanced, so reset the balancing interval */ | |
3417 | sd->balance_interval = sd->min_interval; | |
3418 | } else { | |
3419 | /* | |
3420 | * If we've begun active balancing, start to back off. This | |
3421 | * case may not be covered by the all_pinned logic if there | |
3422 | * is only 1 task on the busy runqueue (because we don't call | |
3423 | * move_tasks). | |
3424 | */ | |
3425 | if (sd->balance_interval < sd->max_interval) | |
3426 | sd->balance_interval *= 2; | |
3427 | } | |
3428 | ||
1e3c88bd PZ |
3429 | goto out; |
3430 | ||
3431 | out_balanced: | |
3432 | schedstat_inc(sd, lb_balanced[idle]); | |
3433 | ||
3434 | sd->nr_balance_failed = 0; | |
3435 | ||
3436 | out_one_pinned: | |
3437 | /* tune up the balancing interval */ | |
3438 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3439 | (sd->balance_interval < sd->max_interval)) | |
3440 | sd->balance_interval *= 2; | |
3441 | ||
46e49b38 | 3442 | ld_moved = 0; |
1e3c88bd | 3443 | out: |
1e3c88bd PZ |
3444 | return ld_moved; |
3445 | } | |
3446 | ||
1e3c88bd PZ |
3447 | /* |
3448 | * idle_balance is called by schedule() if this_cpu is about to become | |
3449 | * idle. Attempts to pull tasks from other CPUs. | |
3450 | */ | |
3451 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3452 | { | |
3453 | struct sched_domain *sd; | |
3454 | int pulled_task = 0; | |
3455 | unsigned long next_balance = jiffies + HZ; | |
3456 | ||
3457 | this_rq->idle_stamp = this_rq->clock; | |
3458 | ||
3459 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3460 | return; | |
3461 | ||
f492e12e PZ |
3462 | /* |
3463 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
3464 | */ | |
3465 | raw_spin_unlock(&this_rq->lock); | |
3466 | ||
c66eaf61 | 3467 | update_shares(this_cpu); |
1e3c88bd PZ |
3468 | for_each_domain(this_cpu, sd) { |
3469 | unsigned long interval; | |
f492e12e | 3470 | int balance = 1; |
1e3c88bd PZ |
3471 | |
3472 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3473 | continue; | |
3474 | ||
f492e12e | 3475 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 3476 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
3477 | pulled_task = load_balance(this_cpu, this_rq, |
3478 | sd, CPU_NEWLY_IDLE, &balance); | |
3479 | } | |
1e3c88bd PZ |
3480 | |
3481 | interval = msecs_to_jiffies(sd->balance_interval); | |
3482 | if (time_after(next_balance, sd->last_balance + interval)) | |
3483 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
3484 | if (pulled_task) { |
3485 | this_rq->idle_stamp = 0; | |
1e3c88bd | 3486 | break; |
d5ad140b | 3487 | } |
1e3c88bd | 3488 | } |
f492e12e PZ |
3489 | |
3490 | raw_spin_lock(&this_rq->lock); | |
3491 | ||
1e3c88bd PZ |
3492 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
3493 | /* | |
3494 | * We are going idle. next_balance may be set based on | |
3495 | * a busy processor. So reset next_balance. | |
3496 | */ | |
3497 | this_rq->next_balance = next_balance; | |
3498 | } | |
3499 | } | |
3500 | ||
3501 | /* | |
969c7921 TH |
3502 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
3503 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
3504 | * least 1 task to be running on each physical CPU where possible, and | |
3505 | * avoids physical / logical imbalances. | |
1e3c88bd | 3506 | */ |
969c7921 | 3507 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 3508 | { |
969c7921 TH |
3509 | struct rq *busiest_rq = data; |
3510 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 3511 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 3512 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 3513 | struct sched_domain *sd; |
969c7921 TH |
3514 | |
3515 | raw_spin_lock_irq(&busiest_rq->lock); | |
3516 | ||
3517 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
3518 | if (unlikely(busiest_cpu != smp_processor_id() || | |
3519 | !busiest_rq->active_balance)) | |
3520 | goto out_unlock; | |
1e3c88bd PZ |
3521 | |
3522 | /* Is there any task to move? */ | |
3523 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 3524 | goto out_unlock; |
1e3c88bd PZ |
3525 | |
3526 | /* | |
3527 | * This condition is "impossible", if it occurs | |
3528 | * we need to fix it. Originally reported by | |
3529 | * Bjorn Helgaas on a 128-cpu setup. | |
3530 | */ | |
3531 | BUG_ON(busiest_rq == target_rq); | |
3532 | ||
3533 | /* move a task from busiest_rq to target_rq */ | |
3534 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
3535 | |
3536 | /* Search for an sd spanning us and the target CPU. */ | |
3537 | for_each_domain(target_cpu, sd) { | |
3538 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3539 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3540 | break; | |
3541 | } | |
3542 | ||
3543 | if (likely(sd)) { | |
3544 | schedstat_inc(sd, alb_count); | |
3545 | ||
3546 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3547 | sd, CPU_IDLE)) | |
3548 | schedstat_inc(sd, alb_pushed); | |
3549 | else | |
3550 | schedstat_inc(sd, alb_failed); | |
3551 | } | |
3552 | double_unlock_balance(busiest_rq, target_rq); | |
969c7921 TH |
3553 | out_unlock: |
3554 | busiest_rq->active_balance = 0; | |
3555 | raw_spin_unlock_irq(&busiest_rq->lock); | |
3556 | return 0; | |
1e3c88bd PZ |
3557 | } |
3558 | ||
3559 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
3560 | |
3561 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
3562 | ||
3563 | static void trigger_sched_softirq(void *data) | |
3564 | { | |
3565 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
3566 | } | |
3567 | ||
3568 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
3569 | { | |
3570 | csd->func = trigger_sched_softirq; | |
3571 | csd->info = NULL; | |
3572 | csd->flags = 0; | |
3573 | csd->priv = 0; | |
3574 | } | |
3575 | ||
3576 | /* | |
3577 | * idle load balancing details | |
3578 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
3579 | * entering idle. | |
3580 | * - This idle load balancer CPU will also go into tickless mode when | |
3581 | * it is idle, just like all other idle CPUs | |
3582 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
3583 | * needed, they will kick the idle load balancer, which then does idle | |
3584 | * load balancing for all the idle CPUs. | |
3585 | */ | |
1e3c88bd PZ |
3586 | static struct { |
3587 | atomic_t load_balancer; | |
83cd4fe2 VP |
3588 | atomic_t first_pick_cpu; |
3589 | atomic_t second_pick_cpu; | |
3590 | cpumask_var_t idle_cpus_mask; | |
3591 | cpumask_var_t grp_idle_mask; | |
3592 | unsigned long next_balance; /* in jiffy units */ | |
3593 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
3594 | |
3595 | int get_nohz_load_balancer(void) | |
3596 | { | |
3597 | return atomic_read(&nohz.load_balancer); | |
3598 | } | |
3599 | ||
3600 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3601 | /** | |
3602 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3603 | * @cpu: The cpu whose lowest level of sched domain is to | |
3604 | * be returned. | |
3605 | * @flag: The flag to check for the lowest sched_domain | |
3606 | * for the given cpu. | |
3607 | * | |
3608 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3609 | */ | |
3610 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3611 | { | |
3612 | struct sched_domain *sd; | |
3613 | ||
3614 | for_each_domain(cpu, sd) | |
3615 | if (sd && (sd->flags & flag)) | |
3616 | break; | |
3617 | ||
3618 | return sd; | |
3619 | } | |
3620 | ||
3621 | /** | |
3622 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3623 | * @cpu: The cpu whose domains we're iterating over. | |
3624 | * @sd: variable holding the value of the power_savings_sd | |
3625 | * for cpu. | |
3626 | * @flag: The flag to filter the sched_domains to be iterated. | |
3627 | * | |
3628 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3629 | * set, starting from the lowest sched_domain to the highest. | |
3630 | */ | |
3631 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3632 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3633 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3634 | ||
3635 | /** | |
3636 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3637 | * @ilb_group: group to be checked for semi-idleness | |
3638 | * | |
3639 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3640 | * | |
3641 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3642 | * and atleast one non-idle CPU. This helper function checks if the given | |
3643 | * sched_group is semi-idle or not. | |
3644 | */ | |
3645 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3646 | { | |
83cd4fe2 | 3647 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
3648 | sched_group_cpus(ilb_group)); |
3649 | ||
3650 | /* | |
3651 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3652 | * and atleast one idle cpu. | |
3653 | */ | |
83cd4fe2 | 3654 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
3655 | return 0; |
3656 | ||
83cd4fe2 | 3657 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
3658 | return 0; |
3659 | ||
3660 | return 1; | |
3661 | } | |
3662 | /** | |
3663 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3664 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3665 | * | |
3666 | * Returns: Returns the id of the idle load balancer if it exists, | |
3667 | * Else, returns >= nr_cpu_ids. | |
3668 | * | |
3669 | * This algorithm picks the idle load balancer such that it belongs to a | |
3670 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3671 | * completely idle packages/cores just for the purpose of idle load balancing | |
3672 | * when there are other idle cpu's which are better suited for that job. | |
3673 | */ | |
3674 | static int find_new_ilb(int cpu) | |
3675 | { | |
3676 | struct sched_domain *sd; | |
3677 | struct sched_group *ilb_group; | |
3678 | ||
3679 | /* | |
3680 | * Have idle load balancer selection from semi-idle packages only | |
3681 | * when power-aware load balancing is enabled | |
3682 | */ | |
3683 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3684 | goto out_done; | |
3685 | ||
3686 | /* | |
3687 | * Optimize for the case when we have no idle CPUs or only one | |
3688 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3689 | */ | |
83cd4fe2 | 3690 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
3691 | goto out_done; |
3692 | ||
3693 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
3694 | ilb_group = sd->groups; | |
3695 | ||
3696 | do { | |
3697 | if (is_semi_idle_group(ilb_group)) | |
83cd4fe2 | 3698 | return cpumask_first(nohz.grp_idle_mask); |
1e3c88bd PZ |
3699 | |
3700 | ilb_group = ilb_group->next; | |
3701 | ||
3702 | } while (ilb_group != sd->groups); | |
3703 | } | |
3704 | ||
3705 | out_done: | |
83cd4fe2 | 3706 | return nr_cpu_ids; |
1e3c88bd PZ |
3707 | } |
3708 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3709 | static inline int find_new_ilb(int call_cpu) | |
3710 | { | |
83cd4fe2 | 3711 | return nr_cpu_ids; |
1e3c88bd PZ |
3712 | } |
3713 | #endif | |
3714 | ||
83cd4fe2 VP |
3715 | /* |
3716 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
3717 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
3718 | * CPU (if there is one). | |
3719 | */ | |
3720 | static void nohz_balancer_kick(int cpu) | |
3721 | { | |
3722 | int ilb_cpu; | |
3723 | ||
3724 | nohz.next_balance++; | |
3725 | ||
3726 | ilb_cpu = get_nohz_load_balancer(); | |
3727 | ||
3728 | if (ilb_cpu >= nr_cpu_ids) { | |
3729 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
3730 | if (ilb_cpu >= nr_cpu_ids) | |
3731 | return; | |
3732 | } | |
3733 | ||
3734 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
3735 | struct call_single_data *cp; | |
3736 | ||
3737 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
3738 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
3739 | __smp_call_function_single(ilb_cpu, cp, 0); | |
3740 | } | |
3741 | return; | |
3742 | } | |
3743 | ||
1e3c88bd PZ |
3744 | /* |
3745 | * This routine will try to nominate the ilb (idle load balancing) | |
3746 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 3747 | * load balancing on behalf of all those cpus. |
1e3c88bd | 3748 | * |
83cd4fe2 VP |
3749 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
3750 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
3751 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 3752 | * |
83cd4fe2 VP |
3753 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
3754 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
3755 | * behalf of all idle CPUs). | |
1e3c88bd | 3756 | */ |
83cd4fe2 | 3757 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
3758 | { |
3759 | int cpu = smp_processor_id(); | |
3760 | ||
3761 | if (stop_tick) { | |
1e3c88bd PZ |
3762 | if (!cpu_active(cpu)) { |
3763 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 3764 | return; |
1e3c88bd PZ |
3765 | |
3766 | /* | |
3767 | * If we are going offline and still the leader, | |
3768 | * give up! | |
3769 | */ | |
83cd4fe2 VP |
3770 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3771 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3772 | BUG(); |
3773 | ||
83cd4fe2 | 3774 | return; |
1e3c88bd PZ |
3775 | } |
3776 | ||
83cd4fe2 | 3777 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 3778 | |
83cd4fe2 VP |
3779 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
3780 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
3781 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
3782 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 3783 | |
83cd4fe2 | 3784 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
3785 | int new_ilb; |
3786 | ||
83cd4fe2 VP |
3787 | /* make me the ilb owner */ |
3788 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
3789 | cpu) != nr_cpu_ids) | |
3790 | return; | |
3791 | ||
1e3c88bd PZ |
3792 | /* |
3793 | * Check to see if there is a more power-efficient | |
3794 | * ilb. | |
3795 | */ | |
3796 | new_ilb = find_new_ilb(cpu); | |
3797 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 3798 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 3799 | resched_cpu(new_ilb); |
83cd4fe2 | 3800 | return; |
1e3c88bd | 3801 | } |
83cd4fe2 | 3802 | return; |
1e3c88bd PZ |
3803 | } |
3804 | } else { | |
83cd4fe2 VP |
3805 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
3806 | return; | |
1e3c88bd | 3807 | |
83cd4fe2 | 3808 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
3809 | |
3810 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
3811 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3812 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3813 | BUG(); |
3814 | } | |
83cd4fe2 | 3815 | return; |
1e3c88bd PZ |
3816 | } |
3817 | #endif | |
3818 | ||
3819 | static DEFINE_SPINLOCK(balancing); | |
3820 | ||
49c022e6 PZ |
3821 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3822 | ||
3823 | /* | |
3824 | * Scale the max load_balance interval with the number of CPUs in the system. | |
3825 | * This trades load-balance latency on larger machines for less cross talk. | |
3826 | */ | |
3827 | static void update_max_interval(void) | |
3828 | { | |
3829 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
3830 | } | |
3831 | ||
1e3c88bd PZ |
3832 | /* |
3833 | * It checks each scheduling domain to see if it is due to be balanced, | |
3834 | * and initiates a balancing operation if so. | |
3835 | * | |
3836 | * Balancing parameters are set up in arch_init_sched_domains. | |
3837 | */ | |
3838 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3839 | { | |
3840 | int balance = 1; | |
3841 | struct rq *rq = cpu_rq(cpu); | |
3842 | unsigned long interval; | |
3843 | struct sched_domain *sd; | |
3844 | /* Earliest time when we have to do rebalance again */ | |
3845 | unsigned long next_balance = jiffies + 60*HZ; | |
3846 | int update_next_balance = 0; | |
3847 | int need_serialize; | |
3848 | ||
2069dd75 PZ |
3849 | update_shares(cpu); |
3850 | ||
1e3c88bd PZ |
3851 | for_each_domain(cpu, sd) { |
3852 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3853 | continue; | |
3854 | ||
3855 | interval = sd->balance_interval; | |
3856 | if (idle != CPU_IDLE) | |
3857 | interval *= sd->busy_factor; | |
3858 | ||
3859 | /* scale ms to jiffies */ | |
3860 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 3861 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
3862 | |
3863 | need_serialize = sd->flags & SD_SERIALIZE; | |
3864 | ||
3865 | if (need_serialize) { | |
3866 | if (!spin_trylock(&balancing)) | |
3867 | goto out; | |
3868 | } | |
3869 | ||
3870 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3871 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3872 | /* | |
3873 | * We've pulled tasks over so either we're no | |
c186fafe | 3874 | * longer idle. |
1e3c88bd PZ |
3875 | */ |
3876 | idle = CPU_NOT_IDLE; | |
3877 | } | |
3878 | sd->last_balance = jiffies; | |
3879 | } | |
3880 | if (need_serialize) | |
3881 | spin_unlock(&balancing); | |
3882 | out: | |
3883 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3884 | next_balance = sd->last_balance + interval; | |
3885 | update_next_balance = 1; | |
3886 | } | |
3887 | ||
3888 | /* | |
3889 | * Stop the load balance at this level. There is another | |
3890 | * CPU in our sched group which is doing load balancing more | |
3891 | * actively. | |
3892 | */ | |
3893 | if (!balance) | |
3894 | break; | |
3895 | } | |
3896 | ||
3897 | /* | |
3898 | * next_balance will be updated only when there is a need. | |
3899 | * When the cpu is attached to null domain for ex, it will not be | |
3900 | * updated. | |
3901 | */ | |
3902 | if (likely(update_next_balance)) | |
3903 | rq->next_balance = next_balance; | |
3904 | } | |
3905 | ||
83cd4fe2 | 3906 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 3907 | /* |
83cd4fe2 | 3908 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
3909 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
3910 | */ | |
83cd4fe2 VP |
3911 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
3912 | { | |
3913 | struct rq *this_rq = cpu_rq(this_cpu); | |
3914 | struct rq *rq; | |
3915 | int balance_cpu; | |
3916 | ||
3917 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
3918 | return; | |
3919 | ||
3920 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
3921 | if (balance_cpu == this_cpu) | |
3922 | continue; | |
3923 | ||
3924 | /* | |
3925 | * If this cpu gets work to do, stop the load balancing | |
3926 | * work being done for other cpus. Next load | |
3927 | * balancing owner will pick it up. | |
3928 | */ | |
3929 | if (need_resched()) { | |
3930 | this_rq->nohz_balance_kick = 0; | |
3931 | break; | |
3932 | } | |
3933 | ||
3934 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 3935 | update_rq_clock(this_rq); |
83cd4fe2 VP |
3936 | update_cpu_load(this_rq); |
3937 | raw_spin_unlock_irq(&this_rq->lock); | |
3938 | ||
3939 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3940 | ||
3941 | rq = cpu_rq(balance_cpu); | |
3942 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3943 | this_rq->next_balance = rq->next_balance; | |
3944 | } | |
3945 | nohz.next_balance = this_rq->next_balance; | |
3946 | this_rq->nohz_balance_kick = 0; | |
3947 | } | |
3948 | ||
3949 | /* | |
3950 | * Current heuristic for kicking the idle load balancer | |
3951 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
3952 | * idle load balancer when it has more than one process active. This | |
3953 | * eliminates the need for idle load balancing altogether when we have | |
3954 | * only one running process in the system (common case). | |
3955 | * - If there are more than one busy CPU, idle load balancer may have | |
3956 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
3957 | * SMT or core siblings and can run better if they move to different | |
3958 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
3959 | * which will kick idle load balancer as soon as it has any load. | |
3960 | */ | |
3961 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
3962 | { | |
3963 | unsigned long now = jiffies; | |
3964 | int ret; | |
3965 | int first_pick_cpu, second_pick_cpu; | |
3966 | ||
3967 | if (time_before(now, nohz.next_balance)) | |
3968 | return 0; | |
3969 | ||
f6c3f168 | 3970 | if (rq->idle_at_tick) |
83cd4fe2 VP |
3971 | return 0; |
3972 | ||
3973 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
3974 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
3975 | ||
3976 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
3977 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
3978 | return 0; | |
3979 | ||
3980 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
3981 | if (ret == nr_cpu_ids || ret == cpu) { | |
3982 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
3983 | if (rq->nr_running > 1) | |
3984 | return 1; | |
3985 | } else { | |
3986 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
3987 | if (ret == nr_cpu_ids || ret == cpu) { | |
3988 | if (rq->nr_running) | |
3989 | return 1; | |
3990 | } | |
3991 | } | |
3992 | return 0; | |
3993 | } | |
3994 | #else | |
3995 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
3996 | #endif | |
3997 | ||
3998 | /* | |
3999 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4000 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
4001 | */ | |
1e3c88bd PZ |
4002 | static void run_rebalance_domains(struct softirq_action *h) |
4003 | { | |
4004 | int this_cpu = smp_processor_id(); | |
4005 | struct rq *this_rq = cpu_rq(this_cpu); | |
4006 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4007 | CPU_IDLE : CPU_NOT_IDLE; | |
4008 | ||
4009 | rebalance_domains(this_cpu, idle); | |
4010 | ||
1e3c88bd | 4011 | /* |
83cd4fe2 | 4012 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
4013 | * balancing on behalf of the other idle cpus whose ticks are |
4014 | * stopped. | |
4015 | */ | |
83cd4fe2 | 4016 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
4017 | } |
4018 | ||
4019 | static inline int on_null_domain(int cpu) | |
4020 | { | |
90a6501f | 4021 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
4022 | } |
4023 | ||
4024 | /* | |
4025 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
4026 | */ |
4027 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
4028 | { | |
1e3c88bd PZ |
4029 | /* Don't need to rebalance while attached to NULL domain */ |
4030 | if (time_after_eq(jiffies, rq->next_balance) && | |
4031 | likely(!on_null_domain(cpu))) | |
4032 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
4033 | #ifdef CONFIG_NO_HZ |
4034 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
4035 | nohz_balancer_kick(cpu); | |
4036 | #endif | |
1e3c88bd PZ |
4037 | } |
4038 | ||
0bcdcf28 CE |
4039 | static void rq_online_fair(struct rq *rq) |
4040 | { | |
4041 | update_sysctl(); | |
4042 | } | |
4043 | ||
4044 | static void rq_offline_fair(struct rq *rq) | |
4045 | { | |
4046 | update_sysctl(); | |
4047 | } | |
4048 | ||
1e3c88bd PZ |
4049 | #else /* CONFIG_SMP */ |
4050 | ||
4051 | /* | |
4052 | * on UP we do not need to balance between CPUs: | |
4053 | */ | |
4054 | static inline void idle_balance(int cpu, struct rq *rq) | |
4055 | { | |
4056 | } | |
4057 | ||
55e12e5e | 4058 | #endif /* CONFIG_SMP */ |
e1d1484f | 4059 | |
bf0f6f24 IM |
4060 | /* |
4061 | * scheduler tick hitting a task of our scheduling class: | |
4062 | */ | |
8f4d37ec | 4063 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
4064 | { |
4065 | struct cfs_rq *cfs_rq; | |
4066 | struct sched_entity *se = &curr->se; | |
4067 | ||
4068 | for_each_sched_entity(se) { | |
4069 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 4070 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
4071 | } |
4072 | } | |
4073 | ||
4074 | /* | |
cd29fe6f PZ |
4075 | * called on fork with the child task as argument from the parent's context |
4076 | * - child not yet on the tasklist | |
4077 | * - preemption disabled | |
bf0f6f24 | 4078 | */ |
cd29fe6f | 4079 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 4080 | { |
cd29fe6f | 4081 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 4082 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 4083 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
4084 | struct rq *rq = this_rq(); |
4085 | unsigned long flags; | |
4086 | ||
05fa785c | 4087 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 4088 | |
861d034e PZ |
4089 | update_rq_clock(rq); |
4090 | ||
b0a0f667 PM |
4091 | if (unlikely(task_cpu(p) != this_cpu)) { |
4092 | rcu_read_lock(); | |
cd29fe6f | 4093 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
4094 | rcu_read_unlock(); |
4095 | } | |
bf0f6f24 | 4096 | |
7109c442 | 4097 | update_curr(cfs_rq); |
cd29fe6f | 4098 | |
b5d9d734 MG |
4099 | if (curr) |
4100 | se->vruntime = curr->vruntime; | |
aeb73b04 | 4101 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 4102 | |
cd29fe6f | 4103 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 4104 | /* |
edcb60a3 IM |
4105 | * Upon rescheduling, sched_class::put_prev_task() will place |
4106 | * 'current' within the tree based on its new key value. | |
4107 | */ | |
4d78e7b6 | 4108 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 4109 | resched_task(rq->curr); |
4d78e7b6 | 4110 | } |
bf0f6f24 | 4111 | |
88ec22d3 PZ |
4112 | se->vruntime -= cfs_rq->min_vruntime; |
4113 | ||
05fa785c | 4114 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
4115 | } |
4116 | ||
cb469845 SR |
4117 | /* |
4118 | * Priority of the task has changed. Check to see if we preempt | |
4119 | * the current task. | |
4120 | */ | |
da7a735e PZ |
4121 | static void |
4122 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 4123 | { |
da7a735e PZ |
4124 | if (!p->se.on_rq) |
4125 | return; | |
4126 | ||
cb469845 SR |
4127 | /* |
4128 | * Reschedule if we are currently running on this runqueue and | |
4129 | * our priority decreased, or if we are not currently running on | |
4130 | * this runqueue and our priority is higher than the current's | |
4131 | */ | |
da7a735e | 4132 | if (rq->curr == p) { |
cb469845 SR |
4133 | if (p->prio > oldprio) |
4134 | resched_task(rq->curr); | |
4135 | } else | |
15afe09b | 4136 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4137 | } |
4138 | ||
da7a735e PZ |
4139 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
4140 | { | |
4141 | struct sched_entity *se = &p->se; | |
4142 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4143 | ||
4144 | /* | |
4145 | * Ensure the task's vruntime is normalized, so that when its | |
4146 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
4147 | * do the right thing. | |
4148 | * | |
4149 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
4150 | * have normalized the vruntime, if it was !on_rq, then only when | |
4151 | * the task is sleeping will it still have non-normalized vruntime. | |
4152 | */ | |
4153 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
4154 | /* | |
4155 | * Fix up our vruntime so that the current sleep doesn't | |
4156 | * cause 'unlimited' sleep bonus. | |
4157 | */ | |
4158 | place_entity(cfs_rq, se, 0); | |
4159 | se->vruntime -= cfs_rq->min_vruntime; | |
4160 | } | |
4161 | } | |
4162 | ||
cb469845 SR |
4163 | /* |
4164 | * We switched to the sched_fair class. | |
4165 | */ | |
da7a735e | 4166 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 4167 | { |
da7a735e PZ |
4168 | if (!p->se.on_rq) |
4169 | return; | |
4170 | ||
cb469845 SR |
4171 | /* |
4172 | * We were most likely switched from sched_rt, so | |
4173 | * kick off the schedule if running, otherwise just see | |
4174 | * if we can still preempt the current task. | |
4175 | */ | |
da7a735e | 4176 | if (rq->curr == p) |
cb469845 SR |
4177 | resched_task(rq->curr); |
4178 | else | |
15afe09b | 4179 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4180 | } |
4181 | ||
83b699ed SV |
4182 | /* Account for a task changing its policy or group. |
4183 | * | |
4184 | * This routine is mostly called to set cfs_rq->curr field when a task | |
4185 | * migrates between groups/classes. | |
4186 | */ | |
4187 | static void set_curr_task_fair(struct rq *rq) | |
4188 | { | |
4189 | struct sched_entity *se = &rq->curr->se; | |
4190 | ||
4191 | for_each_sched_entity(se) | |
4192 | set_next_entity(cfs_rq_of(se), se); | |
4193 | } | |
4194 | ||
810b3817 | 4195 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4196 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 4197 | { |
b2b5ce02 PZ |
4198 | /* |
4199 | * If the task was not on the rq at the time of this cgroup movement | |
4200 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
4201 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
4202 | * bonus in place_entity()). | |
4203 | * | |
4204 | * If it was on the rq, we've just 'preempted' it, which does convert | |
4205 | * ->vruntime to a relative base. | |
4206 | * | |
4207 | * Make sure both cases convert their relative position when migrating | |
4208 | * to another cgroup's rq. This does somewhat interfere with the | |
4209 | * fair sleeper stuff for the first placement, but who cares. | |
4210 | */ | |
4211 | if (!on_rq) | |
4212 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
4213 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 4214 | if (!on_rq) |
b2b5ce02 | 4215 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
4216 | } |
4217 | #endif | |
4218 | ||
6d686f45 | 4219 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
4220 | { |
4221 | struct sched_entity *se = &task->se; | |
0d721cea PW |
4222 | unsigned int rr_interval = 0; |
4223 | ||
4224 | /* | |
4225 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
4226 | * idle runqueue: | |
4227 | */ | |
0d721cea PW |
4228 | if (rq->cfs.load.weight) |
4229 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
4230 | |
4231 | return rr_interval; | |
4232 | } | |
4233 | ||
bf0f6f24 IM |
4234 | /* |
4235 | * All the scheduling class methods: | |
4236 | */ | |
5522d5d5 IM |
4237 | static const struct sched_class fair_sched_class = { |
4238 | .next = &idle_sched_class, | |
bf0f6f24 IM |
4239 | .enqueue_task = enqueue_task_fair, |
4240 | .dequeue_task = dequeue_task_fair, | |
4241 | .yield_task = yield_task_fair, | |
d95f4122 | 4242 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 4243 | |
2e09bf55 | 4244 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
4245 | |
4246 | .pick_next_task = pick_next_task_fair, | |
4247 | .put_prev_task = put_prev_task_fair, | |
4248 | ||
681f3e68 | 4249 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
4250 | .select_task_rq = select_task_rq_fair, |
4251 | ||
0bcdcf28 CE |
4252 | .rq_online = rq_online_fair, |
4253 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
4254 | |
4255 | .task_waking = task_waking_fair, | |
681f3e68 | 4256 | #endif |
bf0f6f24 | 4257 | |
83b699ed | 4258 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 4259 | .task_tick = task_tick_fair, |
cd29fe6f | 4260 | .task_fork = task_fork_fair, |
cb469845 SR |
4261 | |
4262 | .prio_changed = prio_changed_fair, | |
da7a735e | 4263 | .switched_from = switched_from_fair, |
cb469845 | 4264 | .switched_to = switched_to_fair, |
810b3817 | 4265 | |
0d721cea PW |
4266 | .get_rr_interval = get_rr_interval_fair, |
4267 | ||
810b3817 | 4268 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4269 | .task_move_group = task_move_group_fair, |
810b3817 | 4270 | #endif |
bf0f6f24 IM |
4271 | }; |
4272 | ||
4273 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 4274 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 4275 | { |
bf0f6f24 IM |
4276 | struct cfs_rq *cfs_rq; |
4277 | ||
5973e5b9 | 4278 | rcu_read_lock(); |
c3b64f1e | 4279 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 4280 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 4281 | rcu_read_unlock(); |
bf0f6f24 IM |
4282 | } |
4283 | #endif |