]>
Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
9745512c | 25 | |
bf0f6f24 | 26 | /* |
21805085 | 27 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 28 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 29 | * |
21805085 | 30 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
31 | * 'timeslice length' - timeslices in CFS are of variable length |
32 | * and have no persistent notion like in traditional, time-slice | |
33 | * based scheduling concepts. | |
bf0f6f24 | 34 | * |
d274a4ce IM |
35 | * (to see the precise effective timeslice length of your workload, |
36 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 37 | */ |
21406928 MG |
38 | unsigned int sysctl_sched_latency = 6000000ULL; |
39 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 40 | |
1983a922 CE |
41 | /* |
42 | * The initial- and re-scaling of tunables is configurable | |
43 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
44 | * | |
45 | * Options are: | |
46 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
47 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
48 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
49 | */ | |
50 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
51 | = SCHED_TUNABLESCALING_LOG; | |
52 | ||
2bd8e6d4 | 53 | /* |
b2be5e96 | 54 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 55 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 56 | */ |
0bf377bb IM |
57 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
58 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
59 | |
60 | /* | |
b2be5e96 PZ |
61 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
62 | */ | |
0bf377bb | 63 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
64 | |
65 | /* | |
2bba22c5 | 66 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 67 | * parent will (try to) run first. |
21805085 | 68 | */ |
2bba22c5 | 69 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 70 | |
1799e35d IM |
71 | /* |
72 | * sys_sched_yield() compat mode | |
73 | * | |
74 | * This option switches the agressive yield implementation of the | |
75 | * old scheduler back on. | |
76 | */ | |
77 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
78 | ||
bf0f6f24 IM |
79 | /* |
80 | * SCHED_OTHER wake-up granularity. | |
172e082a | 81 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
82 | * |
83 | * This option delays the preemption effects of decoupled workloads | |
84 | * and reduces their over-scheduling. Synchronous workloads will still | |
85 | * have immediate wakeup/sleep latencies. | |
86 | */ | |
172e082a | 87 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 88 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 89 | |
da84d961 IM |
90 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
91 | ||
a4c2f00f PZ |
92 | static const struct sched_class fair_sched_class; |
93 | ||
bf0f6f24 IM |
94 | /************************************************************** |
95 | * CFS operations on generic schedulable entities: | |
96 | */ | |
97 | ||
62160e3f | 98 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 99 | |
62160e3f | 100 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
101 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
102 | { | |
62160e3f | 103 | return cfs_rq->rq; |
bf0f6f24 IM |
104 | } |
105 | ||
62160e3f IM |
106 | /* An entity is a task if it doesn't "own" a runqueue */ |
107 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 108 | |
8f48894f PZ |
109 | static inline struct task_struct *task_of(struct sched_entity *se) |
110 | { | |
111 | #ifdef CONFIG_SCHED_DEBUG | |
112 | WARN_ON_ONCE(!entity_is_task(se)); | |
113 | #endif | |
114 | return container_of(se, struct task_struct, se); | |
115 | } | |
116 | ||
b758149c PZ |
117 | /* Walk up scheduling entities hierarchy */ |
118 | #define for_each_sched_entity(se) \ | |
119 | for (; se; se = se->parent) | |
120 | ||
121 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
122 | { | |
123 | return p->se.cfs_rq; | |
124 | } | |
125 | ||
126 | /* runqueue on which this entity is (to be) queued */ | |
127 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
128 | { | |
129 | return se->cfs_rq; | |
130 | } | |
131 | ||
132 | /* runqueue "owned" by this group */ | |
133 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
134 | { | |
135 | return grp->my_q; | |
136 | } | |
137 | ||
138 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
139 | * another cpu ('this_cpu') | |
140 | */ | |
141 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
142 | { | |
143 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
144 | } | |
145 | ||
3d4b47b4 PZ |
146 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
147 | { | |
148 | if (!cfs_rq->on_list) { | |
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 | ||
bf0f6f24 IM |
415 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
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 | ||
f4b6755f | 425 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 426 | { |
7eee3e67 | 427 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 428 | |
70eee74b BS |
429 | if (!last) |
430 | return NULL; | |
7eee3e67 IM |
431 | |
432 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
433 | } |
434 | ||
bf0f6f24 IM |
435 | /************************************************************** |
436 | * Scheduling class statistics methods: | |
437 | */ | |
438 | ||
b2be5e96 | 439 | #ifdef CONFIG_SCHED_DEBUG |
acb4a848 | 440 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 441 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
442 | loff_t *ppos) |
443 | { | |
8d65af78 | 444 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 445 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
446 | |
447 | if (ret || !write) | |
448 | return ret; | |
449 | ||
450 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
451 | sysctl_sched_min_granularity); | |
452 | ||
acb4a848 CE |
453 | #define WRT_SYSCTL(name) \ |
454 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
455 | WRT_SYSCTL(sched_min_granularity); | |
456 | WRT_SYSCTL(sched_latency); | |
457 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
458 | #undef WRT_SYSCTL |
459 | ||
b2be5e96 PZ |
460 | return 0; |
461 | } | |
462 | #endif | |
647e7cac | 463 | |
a7be37ac | 464 | /* |
f9c0b095 | 465 | * delta /= w |
a7be37ac PZ |
466 | */ |
467 | static inline unsigned long | |
468 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
469 | { | |
f9c0b095 PZ |
470 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
471 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
472 | |
473 | return delta; | |
474 | } | |
475 | ||
647e7cac IM |
476 | /* |
477 | * The idea is to set a period in which each task runs once. | |
478 | * | |
479 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
480 | * this period because otherwise the slices get too small. | |
481 | * | |
482 | * p = (nr <= nl) ? l : l*nr/nl | |
483 | */ | |
4d78e7b6 PZ |
484 | static u64 __sched_period(unsigned long nr_running) |
485 | { | |
486 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 487 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
488 | |
489 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 490 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 491 | period *= nr_running; |
4d78e7b6 PZ |
492 | } |
493 | ||
494 | return period; | |
495 | } | |
496 | ||
647e7cac IM |
497 | /* |
498 | * We calculate the wall-time slice from the period by taking a part | |
499 | * proportional to the weight. | |
500 | * | |
f9c0b095 | 501 | * s = p*P[w/rw] |
647e7cac | 502 | */ |
6d0f0ebd | 503 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 504 | { |
0a582440 | 505 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 506 | |
0a582440 | 507 | for_each_sched_entity(se) { |
6272d68c | 508 | struct load_weight *load; |
3104bf03 | 509 | struct load_weight lw; |
6272d68c LM |
510 | |
511 | cfs_rq = cfs_rq_of(se); | |
512 | load = &cfs_rq->load; | |
f9c0b095 | 513 | |
0a582440 | 514 | if (unlikely(!se->on_rq)) { |
3104bf03 | 515 | lw = cfs_rq->load; |
0a582440 MG |
516 | |
517 | update_load_add(&lw, se->load.weight); | |
518 | load = &lw; | |
519 | } | |
520 | slice = calc_delta_mine(slice, se->load.weight, load); | |
521 | } | |
522 | return slice; | |
bf0f6f24 IM |
523 | } |
524 | ||
647e7cac | 525 | /* |
ac884dec | 526 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 527 | * |
f9c0b095 | 528 | * vs = s/w |
647e7cac | 529 | */ |
f9c0b095 | 530 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 531 | { |
f9c0b095 | 532 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
533 | } |
534 | ||
bf0f6f24 IM |
535 | /* |
536 | * Update the current task's runtime statistics. Skip current tasks that | |
537 | * are not in our scheduling class. | |
538 | */ | |
539 | static inline void | |
8ebc91d9 IM |
540 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
541 | unsigned long delta_exec) | |
bf0f6f24 | 542 | { |
bbdba7c0 | 543 | unsigned long delta_exec_weighted; |
bf0f6f24 | 544 | |
41acab88 LDM |
545 | schedstat_set(curr->statistics.exec_max, |
546 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
547 | |
548 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 549 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 550 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 551 | |
e9acbff6 | 552 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 553 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
554 | } |
555 | ||
b7cc0896 | 556 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 557 | { |
429d43bc | 558 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 559 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
560 | unsigned long delta_exec; |
561 | ||
562 | if (unlikely(!curr)) | |
563 | return; | |
564 | ||
565 | /* | |
566 | * Get the amount of time the current task was running | |
567 | * since the last time we changed load (this cannot | |
568 | * overflow on 32 bits): | |
569 | */ | |
8ebc91d9 | 570 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
571 | if (!delta_exec) |
572 | return; | |
bf0f6f24 | 573 | |
8ebc91d9 IM |
574 | __update_curr(cfs_rq, curr, delta_exec); |
575 | curr->exec_start = now; | |
d842de87 SV |
576 | |
577 | if (entity_is_task(curr)) { | |
578 | struct task_struct *curtask = task_of(curr); | |
579 | ||
f977bb49 | 580 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 581 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 582 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 583 | } |
bf0f6f24 IM |
584 | } |
585 | ||
586 | static inline void | |
5870db5b | 587 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 588 | { |
41acab88 | 589 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
590 | } |
591 | ||
bf0f6f24 IM |
592 | /* |
593 | * Task is being enqueued - update stats: | |
594 | */ | |
d2417e5a | 595 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 596 | { |
bf0f6f24 IM |
597 | /* |
598 | * Are we enqueueing a waiting task? (for current tasks | |
599 | * a dequeue/enqueue event is a NOP) | |
600 | */ | |
429d43bc | 601 | if (se != cfs_rq->curr) |
5870db5b | 602 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
603 | } |
604 | ||
bf0f6f24 | 605 | static void |
9ef0a961 | 606 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 607 | { |
41acab88 LDM |
608 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
609 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
610 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
611 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
612 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
613 | #ifdef CONFIG_SCHEDSTATS |
614 | if (entity_is_task(se)) { | |
615 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 616 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
617 | } |
618 | #endif | |
41acab88 | 619 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
620 | } |
621 | ||
622 | static inline void | |
19b6a2e3 | 623 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 624 | { |
bf0f6f24 IM |
625 | /* |
626 | * Mark the end of the wait period if dequeueing a | |
627 | * waiting task: | |
628 | */ | |
429d43bc | 629 | if (se != cfs_rq->curr) |
9ef0a961 | 630 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
631 | } |
632 | ||
633 | /* | |
634 | * We are picking a new current task - update its stats: | |
635 | */ | |
636 | static inline void | |
79303e9e | 637 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
638 | { |
639 | /* | |
640 | * We are starting a new run period: | |
641 | */ | |
305e6835 | 642 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
643 | } |
644 | ||
bf0f6f24 IM |
645 | /************************************************** |
646 | * Scheduling class queueing methods: | |
647 | */ | |
648 | ||
c09595f6 PZ |
649 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
650 | static void | |
651 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
652 | { | |
653 | cfs_rq->task_weight += weight; | |
654 | } | |
655 | #else | |
656 | static inline void | |
657 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
658 | { | |
659 | } | |
660 | #endif | |
661 | ||
30cfdcfc DA |
662 | static void |
663 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
664 | { | |
665 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
666 | if (!parent_entity(se)) |
667 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 668 | if (entity_is_task(se)) { |
c09595f6 | 669 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
670 | list_add(&se->group_node, &cfs_rq->tasks); |
671 | } | |
30cfdcfc | 672 | cfs_rq->nr_running++; |
30cfdcfc DA |
673 | } |
674 | ||
675 | static void | |
676 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
677 | { | |
678 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
679 | if (!parent_entity(se)) |
680 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 681 | if (entity_is_task(se)) { |
c09595f6 | 682 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
683 | list_del_init(&se->group_node); |
684 | } | |
30cfdcfc | 685 | cfs_rq->nr_running--; |
30cfdcfc DA |
686 | } |
687 | ||
2069dd75 | 688 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
e33078ba | 689 | static void update_cfs_load(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
690 | { |
691 | u64 period = sched_avg_period(); | |
692 | u64 now, delta; | |
e33078ba | 693 | unsigned long load = cfs_rq->load.weight; |
2069dd75 PZ |
694 | |
695 | if (!cfs_rq) | |
696 | return; | |
697 | ||
698 | now = rq_of(cfs_rq)->clock; | |
699 | delta = now - cfs_rq->load_stamp; | |
700 | ||
e33078ba PT |
701 | /* truncate load history at 4 idle periods */ |
702 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
703 | now - cfs_rq->load_last > 4 * period) { | |
704 | cfs_rq->load_period = 0; | |
705 | cfs_rq->load_avg = 0; | |
706 | } | |
707 | ||
2069dd75 PZ |
708 | cfs_rq->load_stamp = now; |
709 | cfs_rq->load_period += delta; | |
e33078ba PT |
710 | if (load) { |
711 | cfs_rq->load_last = now; | |
712 | cfs_rq->load_avg += delta * load; | |
713 | } | |
2069dd75 PZ |
714 | |
715 | while (cfs_rq->load_period > period) { | |
716 | /* | |
717 | * Inline assembly required to prevent the compiler | |
718 | * optimising this loop into a divmod call. | |
719 | * See __iter_div_u64_rem() for another example of this. | |
720 | */ | |
721 | asm("" : "+rm" (cfs_rq->load_period)); | |
722 | cfs_rq->load_period /= 2; | |
723 | cfs_rq->load_avg /= 2; | |
724 | } | |
3d4b47b4 | 725 | |
e33078ba PT |
726 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
727 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
728 | } |
729 | ||
730 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, | |
731 | unsigned long weight) | |
732 | { | |
733 | if (se->on_rq) | |
734 | account_entity_dequeue(cfs_rq, se); | |
735 | ||
736 | update_load_set(&se->load, weight); | |
737 | ||
738 | if (se->on_rq) | |
739 | account_entity_enqueue(cfs_rq, se); | |
740 | } | |
741 | ||
f0d7442a | 742 | static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta) |
2069dd75 PZ |
743 | { |
744 | struct task_group *tg; | |
745 | struct sched_entity *se; | |
746 | long load_weight, load, shares; | |
747 | ||
748 | if (!cfs_rq) | |
749 | return; | |
750 | ||
751 | tg = cfs_rq->tg; | |
752 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
753 | if (!se) | |
754 | return; | |
755 | ||
f0d7442a | 756 | load = cfs_rq->load.weight + weight_delta; |
2069dd75 PZ |
757 | |
758 | load_weight = atomic_read(&tg->load_weight); | |
759 | load_weight -= cfs_rq->load_contribution; | |
760 | load_weight += load; | |
761 | ||
762 | shares = (tg->shares * load); | |
763 | if (load_weight) | |
764 | shares /= load_weight; | |
765 | ||
766 | if (shares < MIN_SHARES) | |
767 | shares = MIN_SHARES; | |
768 | if (shares > tg->shares) | |
769 | shares = tg->shares; | |
770 | ||
771 | reweight_entity(cfs_rq_of(se), se, shares); | |
772 | } | |
773 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
e33078ba | 774 | static inline void update_cfs_load(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
775 | { |
776 | } | |
777 | ||
f0d7442a | 778 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta) |
2069dd75 PZ |
779 | { |
780 | } | |
781 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
782 | ||
2396af69 | 783 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 784 | { |
bf0f6f24 | 785 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
786 | struct task_struct *tsk = NULL; |
787 | ||
788 | if (entity_is_task(se)) | |
789 | tsk = task_of(se); | |
790 | ||
41acab88 LDM |
791 | if (se->statistics.sleep_start) { |
792 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
793 | |
794 | if ((s64)delta < 0) | |
795 | delta = 0; | |
796 | ||
41acab88 LDM |
797 | if (unlikely(delta > se->statistics.sleep_max)) |
798 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 799 | |
41acab88 LDM |
800 | se->statistics.sleep_start = 0; |
801 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 802 | |
768d0c27 | 803 | if (tsk) { |
e414314c | 804 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
805 | trace_sched_stat_sleep(tsk, delta); |
806 | } | |
bf0f6f24 | 807 | } |
41acab88 LDM |
808 | if (se->statistics.block_start) { |
809 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
810 | |
811 | if ((s64)delta < 0) | |
812 | delta = 0; | |
813 | ||
41acab88 LDM |
814 | if (unlikely(delta > se->statistics.block_max)) |
815 | se->statistics.block_max = delta; | |
bf0f6f24 | 816 | |
41acab88 LDM |
817 | se->statistics.block_start = 0; |
818 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 819 | |
e414314c | 820 | if (tsk) { |
8f0dfc34 | 821 | if (tsk->in_iowait) { |
41acab88 LDM |
822 | se->statistics.iowait_sum += delta; |
823 | se->statistics.iowait_count++; | |
768d0c27 | 824 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
825 | } |
826 | ||
e414314c PZ |
827 | /* |
828 | * Blocking time is in units of nanosecs, so shift by | |
829 | * 20 to get a milliseconds-range estimation of the | |
830 | * amount of time that the task spent sleeping: | |
831 | */ | |
832 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
833 | profile_hits(SLEEP_PROFILING, | |
834 | (void *)get_wchan(tsk), | |
835 | delta >> 20); | |
836 | } | |
837 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 838 | } |
bf0f6f24 IM |
839 | } |
840 | #endif | |
841 | } | |
842 | ||
ddc97297 PZ |
843 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
844 | { | |
845 | #ifdef CONFIG_SCHED_DEBUG | |
846 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
847 | ||
848 | if (d < 0) | |
849 | d = -d; | |
850 | ||
851 | if (d > 3*sysctl_sched_latency) | |
852 | schedstat_inc(cfs_rq, nr_spread_over); | |
853 | #endif | |
854 | } | |
855 | ||
aeb73b04 PZ |
856 | static void |
857 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
858 | { | |
1af5f730 | 859 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 860 | |
2cb8600e PZ |
861 | /* |
862 | * The 'current' period is already promised to the current tasks, | |
863 | * however the extra weight of the new task will slow them down a | |
864 | * little, place the new task so that it fits in the slot that | |
865 | * stays open at the end. | |
866 | */ | |
94dfb5e7 | 867 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 868 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 869 | |
a2e7a7eb | 870 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 871 | if (!initial) { |
a2e7a7eb | 872 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 873 | |
a2e7a7eb MG |
874 | /* |
875 | * Halve their sleep time's effect, to allow | |
876 | * for a gentler effect of sleepers: | |
877 | */ | |
878 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
879 | thresh >>= 1; | |
51e0304c | 880 | |
a2e7a7eb | 881 | vruntime -= thresh; |
aeb73b04 PZ |
882 | } |
883 | ||
b5d9d734 MG |
884 | /* ensure we never gain time by being placed backwards. */ |
885 | vruntime = max_vruntime(se->vruntime, vruntime); | |
886 | ||
67e9fb2a | 887 | se->vruntime = vruntime; |
aeb73b04 PZ |
888 | } |
889 | ||
bf0f6f24 | 890 | static void |
88ec22d3 | 891 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 892 | { |
88ec22d3 PZ |
893 | /* |
894 | * Update the normalized vruntime before updating min_vruntime | |
895 | * through callig update_curr(). | |
896 | */ | |
371fd7e7 | 897 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
898 | se->vruntime += cfs_rq->min_vruntime; |
899 | ||
bf0f6f24 | 900 | /* |
a2a2d680 | 901 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 902 | */ |
b7cc0896 | 903 | update_curr(cfs_rq); |
e33078ba | 904 | update_cfs_load(cfs_rq); |
f0d7442a | 905 | update_cfs_shares(cfs_rq, se->load.weight); |
a992241d | 906 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 907 | |
88ec22d3 | 908 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 909 | place_entity(cfs_rq, se, 0); |
2396af69 | 910 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 911 | } |
bf0f6f24 | 912 | |
d2417e5a | 913 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 914 | check_spread(cfs_rq, se); |
83b699ed SV |
915 | if (se != cfs_rq->curr) |
916 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 917 | se->on_rq = 1; |
3d4b47b4 PZ |
918 | |
919 | if (cfs_rq->nr_running == 1) | |
920 | list_add_leaf_cfs_rq(cfs_rq); | |
bf0f6f24 IM |
921 | } |
922 | ||
a571bbea | 923 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 | 924 | { |
de69a80b | 925 | if (!se || cfs_rq->last == se) |
2002c695 PZ |
926 | cfs_rq->last = NULL; |
927 | ||
de69a80b | 928 | if (!se || cfs_rq->next == se) |
2002c695 PZ |
929 | cfs_rq->next = NULL; |
930 | } | |
931 | ||
a571bbea PZ |
932 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
933 | { | |
934 | for_each_sched_entity(se) | |
935 | __clear_buddies(cfs_rq_of(se), se); | |
936 | } | |
937 | ||
bf0f6f24 | 938 | static void |
371fd7e7 | 939 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 940 | { |
a2a2d680 DA |
941 | /* |
942 | * Update run-time statistics of the 'current'. | |
943 | */ | |
944 | update_curr(cfs_rq); | |
945 | ||
19b6a2e3 | 946 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 947 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 948 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
949 | if (entity_is_task(se)) { |
950 | struct task_struct *tsk = task_of(se); | |
951 | ||
952 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 953 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 954 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 955 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 956 | } |
db36cc7d | 957 | #endif |
67e9fb2a PZ |
958 | } |
959 | ||
2002c695 | 960 | clear_buddies(cfs_rq, se); |
4793241b | 961 | |
83b699ed | 962 | if (se != cfs_rq->curr) |
30cfdcfc | 963 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 964 | se->on_rq = 0; |
e33078ba | 965 | update_cfs_load(cfs_rq); |
30cfdcfc | 966 | account_entity_dequeue(cfs_rq, se); |
1af5f730 | 967 | update_min_vruntime(cfs_rq); |
f0d7442a | 968 | update_cfs_shares(cfs_rq, 0); |
88ec22d3 PZ |
969 | |
970 | /* | |
971 | * Normalize the entity after updating the min_vruntime because the | |
972 | * update can refer to the ->curr item and we need to reflect this | |
973 | * movement in our normalized position. | |
974 | */ | |
371fd7e7 | 975 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 976 | se->vruntime -= cfs_rq->min_vruntime; |
bf0f6f24 IM |
977 | } |
978 | ||
979 | /* | |
980 | * Preempt the current task with a newly woken task if needed: | |
981 | */ | |
7c92e54f | 982 | static void |
2e09bf55 | 983 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 984 | { |
11697830 PZ |
985 | unsigned long ideal_runtime, delta_exec; |
986 | ||
6d0f0ebd | 987 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 988 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 989 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 990 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
991 | /* |
992 | * The current task ran long enough, ensure it doesn't get | |
993 | * re-elected due to buddy favours. | |
994 | */ | |
995 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
996 | return; |
997 | } | |
998 | ||
999 | /* | |
1000 | * Ensure that a task that missed wakeup preemption by a | |
1001 | * narrow margin doesn't have to wait for a full slice. | |
1002 | * This also mitigates buddy induced latencies under load. | |
1003 | */ | |
1004 | if (!sched_feat(WAKEUP_PREEMPT)) | |
1005 | return; | |
1006 | ||
1007 | if (delta_exec < sysctl_sched_min_granularity) | |
1008 | return; | |
1009 | ||
1010 | if (cfs_rq->nr_running > 1) { | |
1011 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
1012 | s64 delta = curr->vruntime - se->vruntime; | |
1013 | ||
1014 | if (delta > ideal_runtime) | |
1015 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 1016 | } |
bf0f6f24 IM |
1017 | } |
1018 | ||
83b699ed | 1019 | static void |
8494f412 | 1020 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1021 | { |
83b699ed SV |
1022 | /* 'current' is not kept within the tree. */ |
1023 | if (se->on_rq) { | |
1024 | /* | |
1025 | * Any task has to be enqueued before it get to execute on | |
1026 | * a CPU. So account for the time it spent waiting on the | |
1027 | * runqueue. | |
1028 | */ | |
1029 | update_stats_wait_end(cfs_rq, se); | |
1030 | __dequeue_entity(cfs_rq, se); | |
1031 | } | |
1032 | ||
79303e9e | 1033 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1034 | cfs_rq->curr = se; |
eba1ed4b IM |
1035 | #ifdef CONFIG_SCHEDSTATS |
1036 | /* | |
1037 | * Track our maximum slice length, if the CPU's load is at | |
1038 | * least twice that of our own weight (i.e. dont track it | |
1039 | * when there are only lesser-weight tasks around): | |
1040 | */ | |
495eca49 | 1041 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1042 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1043 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1044 | } | |
1045 | #endif | |
4a55b450 | 1046 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1047 | } |
1048 | ||
3f3a4904 PZ |
1049 | static int |
1050 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1051 | ||
f4b6755f | 1052 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1053 | { |
f4b6755f | 1054 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
f685ceac | 1055 | struct sched_entity *left = se; |
f4b6755f | 1056 | |
f685ceac MG |
1057 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) |
1058 | se = cfs_rq->next; | |
aa2ac252 | 1059 | |
f685ceac MG |
1060 | /* |
1061 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1062 | */ | |
1063 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1064 | se = cfs_rq->last; | |
1065 | ||
1066 | clear_buddies(cfs_rq, se); | |
4793241b PZ |
1067 | |
1068 | return se; | |
aa2ac252 PZ |
1069 | } |
1070 | ||
ab6cde26 | 1071 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1072 | { |
1073 | /* | |
1074 | * If still on the runqueue then deactivate_task() | |
1075 | * was not called and update_curr() has to be done: | |
1076 | */ | |
1077 | if (prev->on_rq) | |
b7cc0896 | 1078 | update_curr(cfs_rq); |
bf0f6f24 | 1079 | |
ddc97297 | 1080 | check_spread(cfs_rq, prev); |
30cfdcfc | 1081 | if (prev->on_rq) { |
5870db5b | 1082 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1083 | /* Put 'current' back into the tree. */ |
1084 | __enqueue_entity(cfs_rq, prev); | |
1085 | } | |
429d43bc | 1086 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1087 | } |
1088 | ||
8f4d37ec PZ |
1089 | static void |
1090 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1091 | { |
bf0f6f24 | 1092 | /* |
30cfdcfc | 1093 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1094 | */ |
30cfdcfc | 1095 | update_curr(cfs_rq); |
bf0f6f24 | 1096 | |
8f4d37ec PZ |
1097 | #ifdef CONFIG_SCHED_HRTICK |
1098 | /* | |
1099 | * queued ticks are scheduled to match the slice, so don't bother | |
1100 | * validating it and just reschedule. | |
1101 | */ | |
983ed7a6 HH |
1102 | if (queued) { |
1103 | resched_task(rq_of(cfs_rq)->curr); | |
1104 | return; | |
1105 | } | |
8f4d37ec PZ |
1106 | /* |
1107 | * don't let the period tick interfere with the hrtick preemption | |
1108 | */ | |
1109 | if (!sched_feat(DOUBLE_TICK) && | |
1110 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1111 | return; | |
1112 | #endif | |
1113 | ||
ce6c1311 | 1114 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 1115 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1116 | } |
1117 | ||
1118 | /************************************************** | |
1119 | * CFS operations on tasks: | |
1120 | */ | |
1121 | ||
8f4d37ec PZ |
1122 | #ifdef CONFIG_SCHED_HRTICK |
1123 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1124 | { | |
8f4d37ec PZ |
1125 | struct sched_entity *se = &p->se; |
1126 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1127 | ||
1128 | WARN_ON(task_rq(p) != rq); | |
1129 | ||
1130 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1131 | u64 slice = sched_slice(cfs_rq, se); | |
1132 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1133 | s64 delta = slice - ran; | |
1134 | ||
1135 | if (delta < 0) { | |
1136 | if (rq->curr == p) | |
1137 | resched_task(p); | |
1138 | return; | |
1139 | } | |
1140 | ||
1141 | /* | |
1142 | * Don't schedule slices shorter than 10000ns, that just | |
1143 | * doesn't make sense. Rely on vruntime for fairness. | |
1144 | */ | |
31656519 | 1145 | if (rq->curr != p) |
157124c1 | 1146 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1147 | |
31656519 | 1148 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1149 | } |
1150 | } | |
a4c2f00f PZ |
1151 | |
1152 | /* | |
1153 | * called from enqueue/dequeue and updates the hrtick when the | |
1154 | * current task is from our class and nr_running is low enough | |
1155 | * to matter. | |
1156 | */ | |
1157 | static void hrtick_update(struct rq *rq) | |
1158 | { | |
1159 | struct task_struct *curr = rq->curr; | |
1160 | ||
1161 | if (curr->sched_class != &fair_sched_class) | |
1162 | return; | |
1163 | ||
1164 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1165 | hrtick_start_fair(rq, curr); | |
1166 | } | |
55e12e5e | 1167 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1168 | static inline void |
1169 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1170 | { | |
1171 | } | |
a4c2f00f PZ |
1172 | |
1173 | static inline void hrtick_update(struct rq *rq) | |
1174 | { | |
1175 | } | |
8f4d37ec PZ |
1176 | #endif |
1177 | ||
bf0f6f24 IM |
1178 | /* |
1179 | * The enqueue_task method is called before nr_running is | |
1180 | * increased. Here we update the fair scheduling stats and | |
1181 | * then put the task into the rbtree: | |
1182 | */ | |
ea87bb78 | 1183 | static void |
371fd7e7 | 1184 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1185 | { |
1186 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1187 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1188 | |
1189 | for_each_sched_entity(se) { | |
62fb1851 | 1190 | if (se->on_rq) |
bf0f6f24 IM |
1191 | break; |
1192 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 PZ |
1193 | enqueue_entity(cfs_rq, se, flags); |
1194 | flags = ENQUEUE_WAKEUP; | |
bf0f6f24 | 1195 | } |
8f4d37ec | 1196 | |
2069dd75 PZ |
1197 | for_each_sched_entity(se) { |
1198 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1199 | ||
e33078ba | 1200 | update_cfs_load(cfs_rq); |
f0d7442a | 1201 | update_cfs_shares(cfs_rq, 0); |
2069dd75 PZ |
1202 | } |
1203 | ||
a4c2f00f | 1204 | hrtick_update(rq); |
bf0f6f24 IM |
1205 | } |
1206 | ||
1207 | /* | |
1208 | * The dequeue_task method is called before nr_running is | |
1209 | * decreased. We remove the task from the rbtree and | |
1210 | * update the fair scheduling stats: | |
1211 | */ | |
371fd7e7 | 1212 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1213 | { |
1214 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1215 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1216 | |
1217 | for_each_sched_entity(se) { | |
1218 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1219 | dequeue_entity(cfs_rq, se, flags); |
2069dd75 | 1220 | |
bf0f6f24 | 1221 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 1222 | if (cfs_rq->load.weight) |
bf0f6f24 | 1223 | break; |
371fd7e7 | 1224 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1225 | } |
8f4d37ec | 1226 | |
2069dd75 PZ |
1227 | for_each_sched_entity(se) { |
1228 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1229 | ||
e33078ba | 1230 | update_cfs_load(cfs_rq); |
f0d7442a | 1231 | update_cfs_shares(cfs_rq, 0); |
2069dd75 PZ |
1232 | } |
1233 | ||
a4c2f00f | 1234 | hrtick_update(rq); |
bf0f6f24 IM |
1235 | } |
1236 | ||
1237 | /* | |
1799e35d IM |
1238 | * sched_yield() support is very simple - we dequeue and enqueue. |
1239 | * | |
1240 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 1241 | */ |
4530d7ab | 1242 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 1243 | { |
db292ca3 IM |
1244 | struct task_struct *curr = rq->curr; |
1245 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1246 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
1247 | |
1248 | /* | |
1799e35d IM |
1249 | * Are we the only task in the tree? |
1250 | */ | |
1251 | if (unlikely(cfs_rq->nr_running == 1)) | |
1252 | return; | |
1253 | ||
2002c695 PZ |
1254 | clear_buddies(cfs_rq, se); |
1255 | ||
db292ca3 | 1256 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1257 | update_rq_clock(rq); |
1799e35d | 1258 | /* |
a2a2d680 | 1259 | * Update run-time statistics of the 'current'. |
1799e35d | 1260 | */ |
2b1e315d | 1261 | update_curr(cfs_rq); |
1799e35d IM |
1262 | |
1263 | return; | |
1264 | } | |
1265 | /* | |
1266 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1267 | */ |
2b1e315d | 1268 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1269 | /* |
1270 | * Already in the rightmost position? | |
1271 | */ | |
54fdc581 | 1272 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
1799e35d IM |
1273 | return; |
1274 | ||
1275 | /* | |
1276 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1277 | * Upon rescheduling, sched_class::put_prev_task() will place |
1278 | * 'current' within the tree based on its new key value. | |
1799e35d | 1279 | */ |
30cfdcfc | 1280 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1281 | } |
1282 | ||
e7693a36 | 1283 | #ifdef CONFIG_SMP |
098fb9db | 1284 | |
88ec22d3 PZ |
1285 | static void task_waking_fair(struct rq *rq, struct task_struct *p) |
1286 | { | |
1287 | struct sched_entity *se = &p->se; | |
1288 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1289 | ||
1290 | se->vruntime -= cfs_rq->min_vruntime; | |
1291 | } | |
1292 | ||
bb3469ac | 1293 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1294 | /* |
1295 | * effective_load() calculates the load change as seen from the root_task_group | |
1296 | * | |
1297 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1298 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1299 | * can calculate the shift in shares. | |
f5bfb7d9 | 1300 | */ |
2069dd75 | 1301 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 1302 | { |
4be9daaa | 1303 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1304 | |
1305 | if (!tg->parent) | |
1306 | return wl; | |
1307 | ||
4be9daaa | 1308 | for_each_sched_entity(se) { |
cb5ef42a | 1309 | long S, rw, s, a, b; |
4be9daaa PZ |
1310 | |
1311 | S = se->my_q->tg->shares; | |
2069dd75 PZ |
1312 | s = se->load.weight; |
1313 | rw = se->my_q->load.weight; | |
bb3469ac | 1314 | |
cb5ef42a PZ |
1315 | a = S*(rw + wl); |
1316 | b = S*rw + s*wg; | |
4be9daaa | 1317 | |
940959e9 PZ |
1318 | wl = s*(a-b); |
1319 | ||
1320 | if (likely(b)) | |
1321 | wl /= b; | |
1322 | ||
83378269 PZ |
1323 | /* |
1324 | * Assume the group is already running and will | |
1325 | * thus already be accounted for in the weight. | |
1326 | * | |
1327 | * That is, moving shares between CPUs, does not | |
1328 | * alter the group weight. | |
1329 | */ | |
4be9daaa | 1330 | wg = 0; |
4be9daaa | 1331 | } |
bb3469ac | 1332 | |
4be9daaa | 1333 | return wl; |
bb3469ac | 1334 | } |
4be9daaa | 1335 | |
bb3469ac | 1336 | #else |
4be9daaa | 1337 | |
83378269 PZ |
1338 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1339 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1340 | { |
83378269 | 1341 | return wl; |
bb3469ac | 1342 | } |
4be9daaa | 1343 | |
bb3469ac PZ |
1344 | #endif |
1345 | ||
c88d5910 | 1346 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1347 | { |
c88d5910 PZ |
1348 | unsigned long this_load, load; |
1349 | int idx, this_cpu, prev_cpu; | |
098fb9db | 1350 | unsigned long tl_per_task; |
c88d5910 | 1351 | struct task_group *tg; |
83378269 | 1352 | unsigned long weight; |
b3137bc8 | 1353 | int balanced; |
098fb9db | 1354 | |
c88d5910 PZ |
1355 | idx = sd->wake_idx; |
1356 | this_cpu = smp_processor_id(); | |
1357 | prev_cpu = task_cpu(p); | |
1358 | load = source_load(prev_cpu, idx); | |
1359 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1360 | |
b3137bc8 MG |
1361 | /* |
1362 | * If sync wakeup then subtract the (maximum possible) | |
1363 | * effect of the currently running task from the load | |
1364 | * of the current CPU: | |
1365 | */ | |
f3b577de | 1366 | rcu_read_lock(); |
83378269 PZ |
1367 | if (sync) { |
1368 | tg = task_group(current); | |
1369 | weight = current->se.load.weight; | |
1370 | ||
c88d5910 | 1371 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1372 | load += effective_load(tg, prev_cpu, 0, -weight); |
1373 | } | |
b3137bc8 | 1374 | |
83378269 PZ |
1375 | tg = task_group(p); |
1376 | weight = p->se.load.weight; | |
b3137bc8 | 1377 | |
71a29aa7 PZ |
1378 | /* |
1379 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1380 | * due to the sync cause above having dropped this_load to 0, we'll |
1381 | * always have an imbalance, but there's really nothing you can do | |
1382 | * about that, so that's good too. | |
71a29aa7 PZ |
1383 | * |
1384 | * Otherwise check if either cpus are near enough in load to allow this | |
1385 | * task to be woken on this_cpu. | |
1386 | */ | |
e51fd5e2 PZ |
1387 | if (this_load) { |
1388 | unsigned long this_eff_load, prev_eff_load; | |
1389 | ||
1390 | this_eff_load = 100; | |
1391 | this_eff_load *= power_of(prev_cpu); | |
1392 | this_eff_load *= this_load + | |
1393 | effective_load(tg, this_cpu, weight, weight); | |
1394 | ||
1395 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
1396 | prev_eff_load *= power_of(this_cpu); | |
1397 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
1398 | ||
1399 | balanced = this_eff_load <= prev_eff_load; | |
1400 | } else | |
1401 | balanced = true; | |
f3b577de | 1402 | rcu_read_unlock(); |
b3137bc8 | 1403 | |
098fb9db | 1404 | /* |
4ae7d5ce IM |
1405 | * If the currently running task will sleep within |
1406 | * a reasonable amount of time then attract this newly | |
1407 | * woken task: | |
098fb9db | 1408 | */ |
2fb7635c PZ |
1409 | if (sync && balanced) |
1410 | return 1; | |
098fb9db | 1411 | |
41acab88 | 1412 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
1413 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
1414 | ||
c88d5910 PZ |
1415 | if (balanced || |
1416 | (this_load <= load && | |
1417 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1418 | /* |
1419 | * This domain has SD_WAKE_AFFINE and | |
1420 | * p is cache cold in this domain, and | |
1421 | * there is no bad imbalance. | |
1422 | */ | |
c88d5910 | 1423 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 1424 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
1425 | |
1426 | return 1; | |
1427 | } | |
1428 | return 0; | |
1429 | } | |
1430 | ||
aaee1203 PZ |
1431 | /* |
1432 | * find_idlest_group finds and returns the least busy CPU group within the | |
1433 | * domain. | |
1434 | */ | |
1435 | static struct sched_group * | |
78e7ed53 | 1436 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1437 | int this_cpu, int load_idx) |
e7693a36 | 1438 | { |
b3bd3de6 | 1439 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 1440 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 1441 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1442 | |
aaee1203 PZ |
1443 | do { |
1444 | unsigned long load, avg_load; | |
1445 | int local_group; | |
1446 | int i; | |
e7693a36 | 1447 | |
aaee1203 PZ |
1448 | /* Skip over this group if it has no CPUs allowed */ |
1449 | if (!cpumask_intersects(sched_group_cpus(group), | |
1450 | &p->cpus_allowed)) | |
1451 | continue; | |
1452 | ||
1453 | local_group = cpumask_test_cpu(this_cpu, | |
1454 | sched_group_cpus(group)); | |
1455 | ||
1456 | /* Tally up the load of all CPUs in the group */ | |
1457 | avg_load = 0; | |
1458 | ||
1459 | for_each_cpu(i, sched_group_cpus(group)) { | |
1460 | /* Bias balancing toward cpus of our domain */ | |
1461 | if (local_group) | |
1462 | load = source_load(i, load_idx); | |
1463 | else | |
1464 | load = target_load(i, load_idx); | |
1465 | ||
1466 | avg_load += load; | |
1467 | } | |
1468 | ||
1469 | /* Adjust by relative CPU power of the group */ | |
1470 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1471 | ||
1472 | if (local_group) { | |
1473 | this_load = avg_load; | |
aaee1203 PZ |
1474 | } else if (avg_load < min_load) { |
1475 | min_load = avg_load; | |
1476 | idlest = group; | |
1477 | } | |
1478 | } while (group = group->next, group != sd->groups); | |
1479 | ||
1480 | if (!idlest || 100*this_load < imbalance*min_load) | |
1481 | return NULL; | |
1482 | return idlest; | |
1483 | } | |
1484 | ||
1485 | /* | |
1486 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1487 | */ | |
1488 | static int | |
1489 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1490 | { | |
1491 | unsigned long load, min_load = ULONG_MAX; | |
1492 | int idlest = -1; | |
1493 | int i; | |
1494 | ||
1495 | /* Traverse only the allowed CPUs */ | |
1496 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1497 | load = weighted_cpuload(i); | |
1498 | ||
1499 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1500 | min_load = load; | |
1501 | idlest = i; | |
e7693a36 GH |
1502 | } |
1503 | } | |
1504 | ||
aaee1203 PZ |
1505 | return idlest; |
1506 | } | |
e7693a36 | 1507 | |
a50bde51 PZ |
1508 | /* |
1509 | * Try and locate an idle CPU in the sched_domain. | |
1510 | */ | |
99bd5e2f | 1511 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
1512 | { |
1513 | int cpu = smp_processor_id(); | |
1514 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 1515 | struct sched_domain *sd; |
a50bde51 PZ |
1516 | int i; |
1517 | ||
1518 | /* | |
99bd5e2f SS |
1519 | * If the task is going to be woken-up on this cpu and if it is |
1520 | * already idle, then it is the right target. | |
a50bde51 | 1521 | */ |
99bd5e2f SS |
1522 | if (target == cpu && idle_cpu(cpu)) |
1523 | return cpu; | |
1524 | ||
1525 | /* | |
1526 | * If the task is going to be woken-up on the cpu where it previously | |
1527 | * ran and if it is currently idle, then it the right target. | |
1528 | */ | |
1529 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 1530 | return prev_cpu; |
a50bde51 PZ |
1531 | |
1532 | /* | |
99bd5e2f | 1533 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 1534 | */ |
99bd5e2f SS |
1535 | for_each_domain(target, sd) { |
1536 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 1537 | break; |
99bd5e2f SS |
1538 | |
1539 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1540 | if (idle_cpu(i)) { | |
1541 | target = i; | |
1542 | break; | |
1543 | } | |
a50bde51 | 1544 | } |
99bd5e2f SS |
1545 | |
1546 | /* | |
1547 | * Lets stop looking for an idle sibling when we reached | |
1548 | * the domain that spans the current cpu and prev_cpu. | |
1549 | */ | |
1550 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1551 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1552 | break; | |
a50bde51 PZ |
1553 | } |
1554 | ||
1555 | return target; | |
1556 | } | |
1557 | ||
aaee1203 PZ |
1558 | /* |
1559 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1560 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1561 | * SD_BALANCE_EXEC. | |
1562 | * | |
1563 | * Balance, ie. select the least loaded group. | |
1564 | * | |
1565 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1566 | * | |
1567 | * preempt must be disabled. | |
1568 | */ | |
0017d735 PZ |
1569 | static int |
1570 | select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags) | |
aaee1203 | 1571 | { |
29cd8bae | 1572 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1573 | int cpu = smp_processor_id(); |
1574 | int prev_cpu = task_cpu(p); | |
1575 | int new_cpu = cpu; | |
99bd5e2f | 1576 | int want_affine = 0; |
29cd8bae | 1577 | int want_sd = 1; |
5158f4e4 | 1578 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1579 | |
0763a660 | 1580 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 1581 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
1582 | want_affine = 1; |
1583 | new_cpu = prev_cpu; | |
1584 | } | |
aaee1203 PZ |
1585 | |
1586 | for_each_domain(cpu, tmp) { | |
e4f42888 PZ |
1587 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1588 | continue; | |
1589 | ||
aaee1203 | 1590 | /* |
ae154be1 PZ |
1591 | * If power savings logic is enabled for a domain, see if we |
1592 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1593 | */ |
59abf026 | 1594 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1595 | unsigned long power = 0; |
1596 | unsigned long nr_running = 0; | |
1597 | unsigned long capacity; | |
1598 | int i; | |
1599 | ||
1600 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1601 | power += power_of(i); | |
1602 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1603 | } | |
1604 | ||
1605 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1606 | ||
59abf026 PZ |
1607 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1608 | nr_running /= 2; | |
1609 | ||
1610 | if (nr_running < capacity) | |
29cd8bae | 1611 | want_sd = 0; |
ae154be1 | 1612 | } |
aaee1203 | 1613 | |
fe3bcfe1 | 1614 | /* |
99bd5e2f SS |
1615 | * If both cpu and prev_cpu are part of this domain, |
1616 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 1617 | */ |
99bd5e2f SS |
1618 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1619 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1620 | affine_sd = tmp; | |
1621 | want_affine = 0; | |
c88d5910 PZ |
1622 | } |
1623 | ||
29cd8bae PZ |
1624 | if (!want_sd && !want_affine) |
1625 | break; | |
1626 | ||
0763a660 | 1627 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1628 | continue; |
1629 | ||
29cd8bae PZ |
1630 | if (want_sd) |
1631 | sd = tmp; | |
1632 | } | |
1633 | ||
8b911acd | 1634 | if (affine_sd) { |
99bd5e2f SS |
1635 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
1636 | return select_idle_sibling(p, cpu); | |
1637 | else | |
1638 | return select_idle_sibling(p, prev_cpu); | |
8b911acd | 1639 | } |
e7693a36 | 1640 | |
aaee1203 | 1641 | while (sd) { |
5158f4e4 | 1642 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1643 | struct sched_group *group; |
c88d5910 | 1644 | int weight; |
098fb9db | 1645 | |
0763a660 | 1646 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1647 | sd = sd->child; |
1648 | continue; | |
1649 | } | |
098fb9db | 1650 | |
5158f4e4 PZ |
1651 | if (sd_flag & SD_BALANCE_WAKE) |
1652 | load_idx = sd->wake_idx; | |
098fb9db | 1653 | |
5158f4e4 | 1654 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1655 | if (!group) { |
1656 | sd = sd->child; | |
1657 | continue; | |
1658 | } | |
4ae7d5ce | 1659 | |
d7c33c49 | 1660 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1661 | if (new_cpu == -1 || new_cpu == cpu) { |
1662 | /* Now try balancing at a lower domain level of cpu */ | |
1663 | sd = sd->child; | |
1664 | continue; | |
e7693a36 | 1665 | } |
aaee1203 PZ |
1666 | |
1667 | /* Now try balancing at a lower domain level of new_cpu */ | |
1668 | cpu = new_cpu; | |
669c55e9 | 1669 | weight = sd->span_weight; |
aaee1203 PZ |
1670 | sd = NULL; |
1671 | for_each_domain(cpu, tmp) { | |
669c55e9 | 1672 | if (weight <= tmp->span_weight) |
aaee1203 | 1673 | break; |
0763a660 | 1674 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1675 | sd = tmp; |
1676 | } | |
1677 | /* while loop will break here if sd == NULL */ | |
e7693a36 GH |
1678 | } |
1679 | ||
c88d5910 | 1680 | return new_cpu; |
e7693a36 GH |
1681 | } |
1682 | #endif /* CONFIG_SMP */ | |
1683 | ||
e52fb7c0 PZ |
1684 | static unsigned long |
1685 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1686 | { |
1687 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1688 | ||
1689 | /* | |
e52fb7c0 PZ |
1690 | * Since its curr running now, convert the gran from real-time |
1691 | * to virtual-time in his units. | |
13814d42 MG |
1692 | * |
1693 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1694 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1695 | * the resulting gran will be larger, therefore penalizing the | |
1696 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1697 | * be smaller, again penalizing the lighter task. | |
1698 | * | |
1699 | * This is especially important for buddies when the leftmost | |
1700 | * task is higher priority than the buddy. | |
0bbd3336 | 1701 | */ |
13814d42 MG |
1702 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
1703 | gran = calc_delta_fair(gran, se); | |
0bbd3336 PZ |
1704 | |
1705 | return gran; | |
1706 | } | |
1707 | ||
464b7527 PZ |
1708 | /* |
1709 | * Should 'se' preempt 'curr'. | |
1710 | * | |
1711 | * |s1 | |
1712 | * |s2 | |
1713 | * |s3 | |
1714 | * g | |
1715 | * |<--->|c | |
1716 | * | |
1717 | * w(c, s1) = -1 | |
1718 | * w(c, s2) = 0 | |
1719 | * w(c, s3) = 1 | |
1720 | * | |
1721 | */ | |
1722 | static int | |
1723 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1724 | { | |
1725 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1726 | ||
1727 | if (vdiff <= 0) | |
1728 | return -1; | |
1729 | ||
e52fb7c0 | 1730 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1731 | if (vdiff > gran) |
1732 | return 1; | |
1733 | ||
1734 | return 0; | |
1735 | } | |
1736 | ||
02479099 PZ |
1737 | static void set_last_buddy(struct sched_entity *se) |
1738 | { | |
6bc912b7 PZ |
1739 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1740 | for_each_sched_entity(se) | |
1741 | cfs_rq_of(se)->last = se; | |
1742 | } | |
02479099 PZ |
1743 | } |
1744 | ||
1745 | static void set_next_buddy(struct sched_entity *se) | |
1746 | { | |
6bc912b7 PZ |
1747 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1748 | for_each_sched_entity(se) | |
1749 | cfs_rq_of(se)->next = se; | |
1750 | } | |
02479099 PZ |
1751 | } |
1752 | ||
bf0f6f24 IM |
1753 | /* |
1754 | * Preempt the current task with a newly woken task if needed: | |
1755 | */ | |
5a9b86f6 | 1756 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1757 | { |
1758 | struct task_struct *curr = rq->curr; | |
8651a86c | 1759 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1760 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 1761 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
bf0f6f24 | 1762 | |
3a7e73a2 PZ |
1763 | if (unlikely(rt_prio(p->prio))) |
1764 | goto preempt; | |
aa2ac252 | 1765 | |
d95f98d0 PZ |
1766 | if (unlikely(p->sched_class != &fair_sched_class)) |
1767 | return; | |
1768 | ||
4ae7d5ce IM |
1769 | if (unlikely(se == pse)) |
1770 | return; | |
1771 | ||
f685ceac | 1772 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) |
3cb63d52 | 1773 | set_next_buddy(pse); |
57fdc26d | 1774 | |
aec0a514 BR |
1775 | /* |
1776 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1777 | * wake up path. | |
1778 | */ | |
1779 | if (test_tsk_need_resched(curr)) | |
1780 | return; | |
1781 | ||
91c234b4 | 1782 | /* |
6bc912b7 | 1783 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1784 | * the tick): |
1785 | */ | |
6bc912b7 | 1786 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1787 | return; |
bf0f6f24 | 1788 | |
6bc912b7 | 1789 | /* Idle tasks are by definition preempted by everybody. */ |
3a7e73a2 PZ |
1790 | if (unlikely(curr->policy == SCHED_IDLE)) |
1791 | goto preempt; | |
bf0f6f24 | 1792 | |
ad4b78bb PZ |
1793 | if (!sched_feat(WAKEUP_PREEMPT)) |
1794 | return; | |
1795 | ||
3a7e73a2 | 1796 | update_curr(cfs_rq); |
464b7527 | 1797 | find_matching_se(&se, &pse); |
002f128b | 1798 | BUG_ON(!pse); |
3a7e73a2 PZ |
1799 | if (wakeup_preempt_entity(se, pse) == 1) |
1800 | goto preempt; | |
464b7527 | 1801 | |
3a7e73a2 | 1802 | return; |
a65ac745 | 1803 | |
3a7e73a2 PZ |
1804 | preempt: |
1805 | resched_task(curr); | |
1806 | /* | |
1807 | * Only set the backward buddy when the current task is still | |
1808 | * on the rq. This can happen when a wakeup gets interleaved | |
1809 | * with schedule on the ->pre_schedule() or idle_balance() | |
1810 | * point, either of which can * drop the rq lock. | |
1811 | * | |
1812 | * Also, during early boot the idle thread is in the fair class, | |
1813 | * for obvious reasons its a bad idea to schedule back to it. | |
1814 | */ | |
1815 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
1816 | return; | |
1817 | ||
1818 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
1819 | set_last_buddy(se); | |
bf0f6f24 IM |
1820 | } |
1821 | ||
fb8d4724 | 1822 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1823 | { |
8f4d37ec | 1824 | struct task_struct *p; |
bf0f6f24 IM |
1825 | struct cfs_rq *cfs_rq = &rq->cfs; |
1826 | struct sched_entity *se; | |
1827 | ||
36ace27e | 1828 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
1829 | return NULL; |
1830 | ||
1831 | do { | |
9948f4b2 | 1832 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1833 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1834 | cfs_rq = group_cfs_rq(se); |
1835 | } while (cfs_rq); | |
1836 | ||
8f4d37ec PZ |
1837 | p = task_of(se); |
1838 | hrtick_start_fair(rq, p); | |
1839 | ||
1840 | return p; | |
bf0f6f24 IM |
1841 | } |
1842 | ||
1843 | /* | |
1844 | * Account for a descheduled task: | |
1845 | */ | |
31ee529c | 1846 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1847 | { |
1848 | struct sched_entity *se = &prev->se; | |
1849 | struct cfs_rq *cfs_rq; | |
1850 | ||
1851 | for_each_sched_entity(se) { | |
1852 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1853 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1854 | } |
1855 | } | |
1856 | ||
681f3e68 | 1857 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1858 | /************************************************** |
1859 | * Fair scheduling class load-balancing methods: | |
1860 | */ | |
1861 | ||
1e3c88bd PZ |
1862 | /* |
1863 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
1864 | * Both runqueues must be locked. | |
1865 | */ | |
1866 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
1867 | struct rq *this_rq, int this_cpu) | |
1868 | { | |
1869 | deactivate_task(src_rq, p, 0); | |
1870 | set_task_cpu(p, this_cpu); | |
1871 | activate_task(this_rq, p, 0); | |
1872 | check_preempt_curr(this_rq, p, 0); | |
fab47622 NR |
1873 | |
1874 | /* re-arm NEWIDLE balancing when moving tasks */ | |
1875 | src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost; | |
1876 | this_rq->idle_stamp = 0; | |
1e3c88bd PZ |
1877 | } |
1878 | ||
1879 | /* | |
1880 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
1881 | */ | |
1882 | static | |
1883 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
1884 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1885 | int *all_pinned) | |
1886 | { | |
1887 | int tsk_cache_hot = 0; | |
1888 | /* | |
1889 | * We do not migrate tasks that are: | |
1890 | * 1) running (obviously), or | |
1891 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
1892 | * 3) are cache-hot on their current CPU. | |
1893 | */ | |
1894 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 1895 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
1896 | return 0; |
1897 | } | |
1898 | *all_pinned = 0; | |
1899 | ||
1900 | if (task_running(rq, p)) { | |
41acab88 | 1901 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
1902 | return 0; |
1903 | } | |
1904 | ||
1905 | /* | |
1906 | * Aggressive migration if: | |
1907 | * 1) task is cache cold, or | |
1908 | * 2) too many balance attempts have failed. | |
1909 | */ | |
1910 | ||
305e6835 | 1911 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
1912 | if (!tsk_cache_hot || |
1913 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
1914 | #ifdef CONFIG_SCHEDSTATS | |
1915 | if (tsk_cache_hot) { | |
1916 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 1917 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
1918 | } |
1919 | #endif | |
1920 | return 1; | |
1921 | } | |
1922 | ||
1923 | if (tsk_cache_hot) { | |
41acab88 | 1924 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
1925 | return 0; |
1926 | } | |
1927 | return 1; | |
1928 | } | |
1929 | ||
897c395f PZ |
1930 | /* |
1931 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
1932 | * part of active balancing operations within "domain". | |
1933 | * Returns 1 if successful and 0 otherwise. | |
1934 | * | |
1935 | * Called with both runqueues locked. | |
1936 | */ | |
1937 | static int | |
1938 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1939 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1940 | { | |
1941 | struct task_struct *p, *n; | |
1942 | struct cfs_rq *cfs_rq; | |
1943 | int pinned = 0; | |
1944 | ||
1945 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
1946 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
1947 | ||
1948 | if (!can_migrate_task(p, busiest, this_cpu, | |
1949 | sd, idle, &pinned)) | |
1950 | continue; | |
1951 | ||
1952 | pull_task(busiest, p, this_rq, this_cpu); | |
1953 | /* | |
1954 | * Right now, this is only the second place pull_task() | |
1955 | * is called, so we can safely collect pull_task() | |
1956 | * stats here rather than inside pull_task(). | |
1957 | */ | |
1958 | schedstat_inc(sd, lb_gained[idle]); | |
1959 | return 1; | |
1960 | } | |
1961 | } | |
1962 | ||
1963 | return 0; | |
1964 | } | |
1965 | ||
1e3c88bd PZ |
1966 | static unsigned long |
1967 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1968 | unsigned long max_load_move, struct sched_domain *sd, | |
1969 | enum cpu_idle_type idle, int *all_pinned, | |
ee00e66f | 1970 | int *this_best_prio, struct cfs_rq *busiest_cfs_rq) |
1e3c88bd PZ |
1971 | { |
1972 | int loops = 0, pulled = 0, pinned = 0; | |
1e3c88bd | 1973 | long rem_load_move = max_load_move; |
ee00e66f | 1974 | struct task_struct *p, *n; |
1e3c88bd PZ |
1975 | |
1976 | if (max_load_move == 0) | |
1977 | goto out; | |
1978 | ||
1979 | pinned = 1; | |
1980 | ||
ee00e66f PZ |
1981 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
1982 | if (loops++ > sysctl_sched_nr_migrate) | |
1983 | break; | |
1e3c88bd | 1984 | |
ee00e66f PZ |
1985 | if ((p->se.load.weight >> 1) > rem_load_move || |
1986 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) | |
1987 | continue; | |
1e3c88bd | 1988 | |
ee00e66f PZ |
1989 | pull_task(busiest, p, this_rq, this_cpu); |
1990 | pulled++; | |
1991 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
1992 | |
1993 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
1994 | /* |
1995 | * NEWIDLE balancing is a source of latency, so preemptible | |
1996 | * kernels will stop after the first task is pulled to minimize | |
1997 | * the critical section. | |
1998 | */ | |
1999 | if (idle == CPU_NEWLY_IDLE) | |
2000 | break; | |
1e3c88bd PZ |
2001 | #endif |
2002 | ||
ee00e66f PZ |
2003 | /* |
2004 | * We only want to steal up to the prescribed amount of | |
2005 | * weighted load. | |
2006 | */ | |
2007 | if (rem_load_move <= 0) | |
2008 | break; | |
2009 | ||
1e3c88bd PZ |
2010 | if (p->prio < *this_best_prio) |
2011 | *this_best_prio = p->prio; | |
1e3c88bd PZ |
2012 | } |
2013 | out: | |
2014 | /* | |
2015 | * Right now, this is one of only two places pull_task() is called, | |
2016 | * so we can safely collect pull_task() stats here rather than | |
2017 | * inside pull_task(). | |
2018 | */ | |
2019 | schedstat_add(sd, lb_gained[idle], pulled); | |
2020 | ||
2021 | if (all_pinned) | |
2022 | *all_pinned = pinned; | |
2023 | ||
2024 | return max_load_move - rem_load_move; | |
2025 | } | |
2026 | ||
230059de | 2027 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
2028 | /* |
2029 | * update tg->load_weight by folding this cpu's load_avg | |
2030 | */ | |
67e86250 | 2031 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
2032 | { |
2033 | struct cfs_rq *cfs_rq; | |
2034 | unsigned long flags; | |
2035 | struct rq *rq; | |
2036 | long load_avg; | |
2037 | ||
2038 | if (!tg->se[cpu]) | |
2039 | return 0; | |
2040 | ||
2041 | rq = cpu_rq(cpu); | |
2042 | cfs_rq = tg->cfs_rq[cpu]; | |
2043 | ||
2044 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2045 | ||
2046 | update_rq_clock(rq); | |
e33078ba | 2047 | update_cfs_load(cfs_rq); |
9e3081ca PZ |
2048 | |
2049 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
2050 | load_avg -= cfs_rq->load_contribution; | |
2051 | atomic_add(load_avg, &tg->load_weight); | |
2052 | cfs_rq->load_contribution += load_avg; | |
2053 | ||
2054 | /* | |
2055 | * We need to update shares after updating tg->load_weight in | |
2056 | * order to adjust the weight of groups with long running tasks. | |
2057 | */ | |
f0d7442a | 2058 | update_cfs_shares(cfs_rq, 0); |
9e3081ca PZ |
2059 | |
2060 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2061 | ||
2062 | return 0; | |
2063 | } | |
2064 | ||
2065 | static void update_shares(int cpu) | |
2066 | { | |
2067 | struct cfs_rq *cfs_rq; | |
2068 | struct rq *rq = cpu_rq(cpu); | |
2069 | ||
2070 | rcu_read_lock(); | |
67e86250 PT |
2071 | for_each_leaf_cfs_rq(rq, cfs_rq) |
2072 | update_shares_cpu(cfs_rq->tg, cpu); | |
9e3081ca PZ |
2073 | rcu_read_unlock(); |
2074 | } | |
2075 | ||
230059de PZ |
2076 | static unsigned long |
2077 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2078 | unsigned long max_load_move, | |
2079 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2080 | int *all_pinned, int *this_best_prio) | |
2081 | { | |
2082 | long rem_load_move = max_load_move; | |
2083 | int busiest_cpu = cpu_of(busiest); | |
2084 | struct task_group *tg; | |
2085 | ||
2086 | rcu_read_lock(); | |
2087 | update_h_load(busiest_cpu); | |
2088 | ||
2089 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
2090 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; | |
2091 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; | |
2092 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
2093 | u64 rem_load, moved_load; | |
2094 | ||
2095 | /* | |
2096 | * empty group | |
2097 | */ | |
2098 | if (!busiest_cfs_rq->task_weight) | |
2099 | continue; | |
2100 | ||
2101 | rem_load = (u64)rem_load_move * busiest_weight; | |
2102 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
2103 | ||
2104 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
2105 | rem_load, sd, idle, all_pinned, this_best_prio, | |
2106 | busiest_cfs_rq); | |
2107 | ||
2108 | if (!moved_load) | |
2109 | continue; | |
2110 | ||
2111 | moved_load *= busiest_h_load; | |
2112 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
2113 | ||
2114 | rem_load_move -= moved_load; | |
2115 | if (rem_load_move < 0) | |
2116 | break; | |
2117 | } | |
2118 | rcu_read_unlock(); | |
2119 | ||
2120 | return max_load_move - rem_load_move; | |
2121 | } | |
2122 | #else | |
9e3081ca PZ |
2123 | static inline void update_shares(int cpu) |
2124 | { | |
2125 | } | |
2126 | ||
230059de PZ |
2127 | static unsigned long |
2128 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2129 | unsigned long max_load_move, | |
2130 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2131 | int *all_pinned, int *this_best_prio) | |
2132 | { | |
2133 | return balance_tasks(this_rq, this_cpu, busiest, | |
2134 | max_load_move, sd, idle, all_pinned, | |
2135 | this_best_prio, &busiest->cfs); | |
2136 | } | |
2137 | #endif | |
2138 | ||
1e3c88bd PZ |
2139 | /* |
2140 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2141 | * this_rq, as part of a balancing operation within domain "sd". | |
2142 | * Returns 1 if successful and 0 otherwise. | |
2143 | * | |
2144 | * Called with both runqueues locked. | |
2145 | */ | |
2146 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2147 | unsigned long max_load_move, | |
2148 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2149 | int *all_pinned) | |
2150 | { | |
3d45fd80 | 2151 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2152 | int this_best_prio = this_rq->curr->prio; |
2153 | ||
2154 | do { | |
3d45fd80 | 2155 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd PZ |
2156 | max_load_move - total_load_moved, |
2157 | sd, idle, all_pinned, &this_best_prio); | |
3d45fd80 PZ |
2158 | |
2159 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2160 | |
2161 | #ifdef CONFIG_PREEMPT | |
2162 | /* | |
2163 | * NEWIDLE balancing is a source of latency, so preemptible | |
2164 | * kernels will stop after the first task is pulled to minimize | |
2165 | * the critical section. | |
2166 | */ | |
2167 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2168 | break; | |
baa8c110 PZ |
2169 | |
2170 | if (raw_spin_is_contended(&this_rq->lock) || | |
2171 | raw_spin_is_contended(&busiest->lock)) | |
2172 | break; | |
1e3c88bd | 2173 | #endif |
3d45fd80 | 2174 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2175 | |
2176 | return total_load_moved > 0; | |
2177 | } | |
2178 | ||
1e3c88bd PZ |
2179 | /********** Helpers for find_busiest_group ************************/ |
2180 | /* | |
2181 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2182 | * during load balancing. | |
2183 | */ | |
2184 | struct sd_lb_stats { | |
2185 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2186 | struct sched_group *this; /* Local group in this sd */ | |
2187 | unsigned long total_load; /* Total load of all groups in sd */ | |
2188 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2189 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2190 | ||
2191 | /** Statistics of this group */ | |
2192 | unsigned long this_load; | |
2193 | unsigned long this_load_per_task; | |
2194 | unsigned long this_nr_running; | |
fab47622 | 2195 | unsigned long this_has_capacity; |
1e3c88bd PZ |
2196 | |
2197 | /* Statistics of the busiest group */ | |
2198 | unsigned long max_load; | |
2199 | unsigned long busiest_load_per_task; | |
2200 | unsigned long busiest_nr_running; | |
dd5feea1 | 2201 | unsigned long busiest_group_capacity; |
fab47622 | 2202 | unsigned long busiest_has_capacity; |
1e3c88bd PZ |
2203 | |
2204 | int group_imb; /* Is there imbalance in this sd */ | |
2205 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2206 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2207 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2208 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2209 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2210 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2211 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2212 | #endif | |
2213 | }; | |
2214 | ||
2215 | /* | |
2216 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2217 | */ | |
2218 | struct sg_lb_stats { | |
2219 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2220 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2221 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2222 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2223 | unsigned long group_capacity; | |
2224 | int group_imb; /* Is there an imbalance in the group ? */ | |
fab47622 | 2225 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2226 | }; |
2227 | ||
2228 | /** | |
2229 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2230 | * @group: The group whose first cpu is to be returned. | |
2231 | */ | |
2232 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2233 | { | |
2234 | return cpumask_first(sched_group_cpus(group)); | |
2235 | } | |
2236 | ||
2237 | /** | |
2238 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2239 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2240 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2241 | */ | |
2242 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2243 | enum cpu_idle_type idle) | |
2244 | { | |
2245 | int load_idx; | |
2246 | ||
2247 | switch (idle) { | |
2248 | case CPU_NOT_IDLE: | |
2249 | load_idx = sd->busy_idx; | |
2250 | break; | |
2251 | ||
2252 | case CPU_NEWLY_IDLE: | |
2253 | load_idx = sd->newidle_idx; | |
2254 | break; | |
2255 | default: | |
2256 | load_idx = sd->idle_idx; | |
2257 | break; | |
2258 | } | |
2259 | ||
2260 | return load_idx; | |
2261 | } | |
2262 | ||
2263 | ||
2264 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2265 | /** | |
2266 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2267 | * the given sched_domain, during load balancing. | |
2268 | * | |
2269 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2270 | * @sds: Variable containing the statistics for sd. | |
2271 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2272 | */ | |
2273 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2274 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2275 | { | |
2276 | /* | |
2277 | * Busy processors will not participate in power savings | |
2278 | * balance. | |
2279 | */ | |
2280 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2281 | sds->power_savings_balance = 0; | |
2282 | else { | |
2283 | sds->power_savings_balance = 1; | |
2284 | sds->min_nr_running = ULONG_MAX; | |
2285 | sds->leader_nr_running = 0; | |
2286 | } | |
2287 | } | |
2288 | ||
2289 | /** | |
2290 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2291 | * sched_domain while performing load balancing. | |
2292 | * | |
2293 | * @group: sched_group belonging to the sched_domain under consideration. | |
2294 | * @sds: Variable containing the statistics of the sched_domain | |
2295 | * @local_group: Does group contain the CPU for which we're performing | |
2296 | * load balancing ? | |
2297 | * @sgs: Variable containing the statistics of the group. | |
2298 | */ | |
2299 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2300 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2301 | { | |
2302 | ||
2303 | if (!sds->power_savings_balance) | |
2304 | return; | |
2305 | ||
2306 | /* | |
2307 | * If the local group is idle or completely loaded | |
2308 | * no need to do power savings balance at this domain | |
2309 | */ | |
2310 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2311 | !sds->this_nr_running)) | |
2312 | sds->power_savings_balance = 0; | |
2313 | ||
2314 | /* | |
2315 | * If a group is already running at full capacity or idle, | |
2316 | * don't include that group in power savings calculations | |
2317 | */ | |
2318 | if (!sds->power_savings_balance || | |
2319 | sgs->sum_nr_running >= sgs->group_capacity || | |
2320 | !sgs->sum_nr_running) | |
2321 | return; | |
2322 | ||
2323 | /* | |
2324 | * Calculate the group which has the least non-idle load. | |
2325 | * This is the group from where we need to pick up the load | |
2326 | * for saving power | |
2327 | */ | |
2328 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2329 | (sgs->sum_nr_running == sds->min_nr_running && | |
2330 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2331 | sds->group_min = group; | |
2332 | sds->min_nr_running = sgs->sum_nr_running; | |
2333 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2334 | sgs->sum_nr_running; | |
2335 | } | |
2336 | ||
2337 | /* | |
2338 | * Calculate the group which is almost near its | |
2339 | * capacity but still has some space to pick up some load | |
2340 | * from other group and save more power | |
2341 | */ | |
2342 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2343 | return; | |
2344 | ||
2345 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2346 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2347 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2348 | sds->group_leader = group; | |
2349 | sds->leader_nr_running = sgs->sum_nr_running; | |
2350 | } | |
2351 | } | |
2352 | ||
2353 | /** | |
2354 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2355 | * @sds: Variable containing the statistics of the sched_domain | |
2356 | * under consideration. | |
2357 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2358 | * @imbalance: Variable to store the imbalance. | |
2359 | * | |
2360 | * Description: | |
2361 | * Check if we have potential to perform some power-savings balance. | |
2362 | * If yes, set the busiest group to be the least loaded group in the | |
2363 | * sched_domain, so that it's CPUs can be put to idle. | |
2364 | * | |
2365 | * Returns 1 if there is potential to perform power-savings balance. | |
2366 | * Else returns 0. | |
2367 | */ | |
2368 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2369 | int this_cpu, unsigned long *imbalance) | |
2370 | { | |
2371 | if (!sds->power_savings_balance) | |
2372 | return 0; | |
2373 | ||
2374 | if (sds->this != sds->group_leader || | |
2375 | sds->group_leader == sds->group_min) | |
2376 | return 0; | |
2377 | ||
2378 | *imbalance = sds->min_load_per_task; | |
2379 | sds->busiest = sds->group_min; | |
2380 | ||
2381 | return 1; | |
2382 | ||
2383 | } | |
2384 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2385 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2386 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2387 | { | |
2388 | return; | |
2389 | } | |
2390 | ||
2391 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2392 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2393 | { | |
2394 | return; | |
2395 | } | |
2396 | ||
2397 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2398 | int this_cpu, unsigned long *imbalance) | |
2399 | { | |
2400 | return 0; | |
2401 | } | |
2402 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2403 | ||
2404 | ||
2405 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2406 | { | |
2407 | return SCHED_LOAD_SCALE; | |
2408 | } | |
2409 | ||
2410 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2411 | { | |
2412 | return default_scale_freq_power(sd, cpu); | |
2413 | } | |
2414 | ||
2415 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2416 | { | |
669c55e9 | 2417 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2418 | unsigned long smt_gain = sd->smt_gain; |
2419 | ||
2420 | smt_gain /= weight; | |
2421 | ||
2422 | return smt_gain; | |
2423 | } | |
2424 | ||
2425 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2426 | { | |
2427 | return default_scale_smt_power(sd, cpu); | |
2428 | } | |
2429 | ||
2430 | unsigned long scale_rt_power(int cpu) | |
2431 | { | |
2432 | struct rq *rq = cpu_rq(cpu); | |
2433 | u64 total, available; | |
2434 | ||
1e3c88bd | 2435 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
2436 | |
2437 | if (unlikely(total < rq->rt_avg)) { | |
2438 | /* Ensures that power won't end up being negative */ | |
2439 | available = 0; | |
2440 | } else { | |
2441 | available = total - rq->rt_avg; | |
2442 | } | |
1e3c88bd PZ |
2443 | |
2444 | if (unlikely((s64)total < SCHED_LOAD_SCALE)) | |
2445 | total = SCHED_LOAD_SCALE; | |
2446 | ||
2447 | total >>= SCHED_LOAD_SHIFT; | |
2448 | ||
2449 | return div_u64(available, total); | |
2450 | } | |
2451 | ||
2452 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2453 | { | |
669c55e9 | 2454 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2455 | unsigned long power = SCHED_LOAD_SCALE; |
2456 | struct sched_group *sdg = sd->groups; | |
2457 | ||
1e3c88bd PZ |
2458 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
2459 | if (sched_feat(ARCH_POWER)) | |
2460 | power *= arch_scale_smt_power(sd, cpu); | |
2461 | else | |
2462 | power *= default_scale_smt_power(sd, cpu); | |
2463 | ||
2464 | power >>= SCHED_LOAD_SHIFT; | |
2465 | } | |
2466 | ||
9d5efe05 SV |
2467 | sdg->cpu_power_orig = power; |
2468 | ||
2469 | if (sched_feat(ARCH_POWER)) | |
2470 | power *= arch_scale_freq_power(sd, cpu); | |
2471 | else | |
2472 | power *= default_scale_freq_power(sd, cpu); | |
2473 | ||
2474 | power >>= SCHED_LOAD_SHIFT; | |
2475 | ||
1e3c88bd PZ |
2476 | power *= scale_rt_power(cpu); |
2477 | power >>= SCHED_LOAD_SHIFT; | |
2478 | ||
2479 | if (!power) | |
2480 | power = 1; | |
2481 | ||
e51fd5e2 | 2482 | cpu_rq(cpu)->cpu_power = power; |
1e3c88bd PZ |
2483 | sdg->cpu_power = power; |
2484 | } | |
2485 | ||
2486 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2487 | { | |
2488 | struct sched_domain *child = sd->child; | |
2489 | struct sched_group *group, *sdg = sd->groups; | |
2490 | unsigned long power; | |
2491 | ||
2492 | if (!child) { | |
2493 | update_cpu_power(sd, cpu); | |
2494 | return; | |
2495 | } | |
2496 | ||
2497 | power = 0; | |
2498 | ||
2499 | group = child->groups; | |
2500 | do { | |
2501 | power += group->cpu_power; | |
2502 | group = group->next; | |
2503 | } while (group != child->groups); | |
2504 | ||
2505 | sdg->cpu_power = power; | |
2506 | } | |
2507 | ||
9d5efe05 SV |
2508 | /* |
2509 | * Try and fix up capacity for tiny siblings, this is needed when | |
2510 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
2511 | * which on its own isn't powerful enough. | |
2512 | * | |
2513 | * See update_sd_pick_busiest() and check_asym_packing(). | |
2514 | */ | |
2515 | static inline int | |
2516 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
2517 | { | |
2518 | /* | |
2519 | * Only siblings can have significantly less than SCHED_LOAD_SCALE | |
2520 | */ | |
2521 | if (sd->level != SD_LV_SIBLING) | |
2522 | return 0; | |
2523 | ||
2524 | /* | |
2525 | * If ~90% of the cpu_power is still there, we're good. | |
2526 | */ | |
694f5a11 | 2527 | if (group->cpu_power * 32 > group->cpu_power_orig * 29) |
9d5efe05 SV |
2528 | return 1; |
2529 | ||
2530 | return 0; | |
2531 | } | |
2532 | ||
1e3c88bd PZ |
2533 | /** |
2534 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2535 | * @sd: The sched_domain whose statistics are to be updated. | |
2536 | * @group: sched_group whose statistics are to be updated. | |
2537 | * @this_cpu: Cpu for which load balance is currently performed. | |
2538 | * @idle: Idle status of this_cpu | |
2539 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
2540 | * @sd_idle: Idle status of the sched_domain containing group. | |
2541 | * @local_group: Does group contain this_cpu. | |
2542 | * @cpus: Set of cpus considered for load balancing. | |
2543 | * @balance: Should we balance. | |
2544 | * @sgs: variable to hold the statistics for this group. | |
2545 | */ | |
2546 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2547 | struct sched_group *group, int this_cpu, | |
2548 | enum cpu_idle_type idle, int load_idx, int *sd_idle, | |
2549 | int local_group, const struct cpumask *cpus, | |
2550 | int *balance, struct sg_lb_stats *sgs) | |
2551 | { | |
2582f0eb | 2552 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
2553 | int i; |
2554 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 2555 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 2556 | |
871e35bc | 2557 | if (local_group) |
1e3c88bd | 2558 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
2559 | |
2560 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
2561 | max_cpu_load = 0; |
2562 | min_cpu_load = ~0UL; | |
2582f0eb | 2563 | max_nr_running = 0; |
1e3c88bd PZ |
2564 | |
2565 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2566 | struct rq *rq = cpu_rq(i); | |
2567 | ||
2568 | if (*sd_idle && rq->nr_running) | |
2569 | *sd_idle = 0; | |
2570 | ||
2571 | /* Bias balancing toward cpus of our domain */ | |
2572 | if (local_group) { | |
2573 | if (idle_cpu(i) && !first_idle_cpu) { | |
2574 | first_idle_cpu = 1; | |
2575 | balance_cpu = i; | |
2576 | } | |
2577 | ||
2578 | load = target_load(i, load_idx); | |
2579 | } else { | |
2580 | load = source_load(i, load_idx); | |
2582f0eb | 2581 | if (load > max_cpu_load) { |
1e3c88bd | 2582 | max_cpu_load = load; |
2582f0eb NR |
2583 | max_nr_running = rq->nr_running; |
2584 | } | |
1e3c88bd PZ |
2585 | if (min_cpu_load > load) |
2586 | min_cpu_load = load; | |
2587 | } | |
2588 | ||
2589 | sgs->group_load += load; | |
2590 | sgs->sum_nr_running += rq->nr_running; | |
2591 | sgs->sum_weighted_load += weighted_cpuload(i); | |
2592 | ||
1e3c88bd PZ |
2593 | } |
2594 | ||
2595 | /* | |
2596 | * First idle cpu or the first cpu(busiest) in this sched group | |
2597 | * is eligible for doing load balancing at this and above | |
2598 | * domains. In the newly idle case, we will allow all the cpu's | |
2599 | * to do the newly idle load balance. | |
2600 | */ | |
bbc8cb5b PZ |
2601 | if (idle != CPU_NEWLY_IDLE && local_group) { |
2602 | if (balance_cpu != this_cpu) { | |
2603 | *balance = 0; | |
2604 | return; | |
2605 | } | |
2606 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
2607 | } |
2608 | ||
2609 | /* Adjust by relative CPU power of the group */ | |
2610 | sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
2611 | ||
1e3c88bd PZ |
2612 | /* |
2613 | * Consider the group unbalanced when the imbalance is larger | |
2614 | * than the average weight of two tasks. | |
2615 | * | |
2616 | * APZ: with cgroup the avg task weight can vary wildly and | |
2617 | * might not be a suitable number - should we keep a | |
2618 | * normalized nr_running number somewhere that negates | |
2619 | * the hierarchy? | |
2620 | */ | |
dd5feea1 SS |
2621 | if (sgs->sum_nr_running) |
2622 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 2623 | |
2582f0eb | 2624 | if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
2625 | sgs->group_imb = 1; |
2626 | ||
2582f0eb | 2627 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE); |
9d5efe05 SV |
2628 | if (!sgs->group_capacity) |
2629 | sgs->group_capacity = fix_small_capacity(sd, group); | |
fab47622 NR |
2630 | |
2631 | if (sgs->group_capacity > sgs->sum_nr_running) | |
2632 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
2633 | } |
2634 | ||
532cb4c4 MN |
2635 | /** |
2636 | * update_sd_pick_busiest - return 1 on busiest group | |
2637 | * @sd: sched_domain whose statistics are to be checked | |
2638 | * @sds: sched_domain statistics | |
2639 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
2640 | * @sgs: sched_group statistics |
2641 | * @this_cpu: the current cpu | |
532cb4c4 MN |
2642 | * |
2643 | * Determine if @sg is a busier group than the previously selected | |
2644 | * busiest group. | |
2645 | */ | |
2646 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
2647 | struct sd_lb_stats *sds, | |
2648 | struct sched_group *sg, | |
2649 | struct sg_lb_stats *sgs, | |
2650 | int this_cpu) | |
2651 | { | |
2652 | if (sgs->avg_load <= sds->max_load) | |
2653 | return false; | |
2654 | ||
2655 | if (sgs->sum_nr_running > sgs->group_capacity) | |
2656 | return true; | |
2657 | ||
2658 | if (sgs->group_imb) | |
2659 | return true; | |
2660 | ||
2661 | /* | |
2662 | * ASYM_PACKING needs to move all the work to the lowest | |
2663 | * numbered CPUs in the group, therefore mark all groups | |
2664 | * higher than ourself as busy. | |
2665 | */ | |
2666 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
2667 | this_cpu < group_first_cpu(sg)) { | |
2668 | if (!sds->busiest) | |
2669 | return true; | |
2670 | ||
2671 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
2672 | return true; | |
2673 | } | |
2674 | ||
2675 | return false; | |
2676 | } | |
2677 | ||
1e3c88bd PZ |
2678 | /** |
2679 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2680 | * @sd: sched_domain whose statistics are to be updated. | |
2681 | * @this_cpu: Cpu for which load balance is currently performed. | |
2682 | * @idle: Idle status of this_cpu | |
532cb4c4 | 2683 | * @sd_idle: Idle status of the sched_domain containing sg. |
1e3c88bd PZ |
2684 | * @cpus: Set of cpus considered for load balancing. |
2685 | * @balance: Should we balance. | |
2686 | * @sds: variable to hold the statistics for this sched_domain. | |
2687 | */ | |
2688 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
2689 | enum cpu_idle_type idle, int *sd_idle, | |
2690 | const struct cpumask *cpus, int *balance, | |
2691 | struct sd_lb_stats *sds) | |
2692 | { | |
2693 | struct sched_domain *child = sd->child; | |
532cb4c4 | 2694 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
2695 | struct sg_lb_stats sgs; |
2696 | int load_idx, prefer_sibling = 0; | |
2697 | ||
2698 | if (child && child->flags & SD_PREFER_SIBLING) | |
2699 | prefer_sibling = 1; | |
2700 | ||
2701 | init_sd_power_savings_stats(sd, sds, idle); | |
2702 | load_idx = get_sd_load_idx(sd, idle); | |
2703 | ||
2704 | do { | |
2705 | int local_group; | |
2706 | ||
532cb4c4 | 2707 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 2708 | memset(&sgs, 0, sizeof(sgs)); |
532cb4c4 | 2709 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle, |
1e3c88bd PZ |
2710 | local_group, cpus, balance, &sgs); |
2711 | ||
8f190fb3 | 2712 | if (local_group && !(*balance)) |
1e3c88bd PZ |
2713 | return; |
2714 | ||
2715 | sds->total_load += sgs.group_load; | |
532cb4c4 | 2716 | sds->total_pwr += sg->cpu_power; |
1e3c88bd PZ |
2717 | |
2718 | /* | |
2719 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 2720 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
2721 | * and move all the excess tasks away. We lower the capacity |
2722 | * of a group only if the local group has the capacity to fit | |
2723 | * these excess tasks, i.e. nr_running < group_capacity. The | |
2724 | * extra check prevents the case where you always pull from the | |
2725 | * heaviest group when it is already under-utilized (possible | |
2726 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 2727 | */ |
75dd321d | 2728 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
2729 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
2730 | ||
2731 | if (local_group) { | |
2732 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 2733 | sds->this = sg; |
1e3c88bd PZ |
2734 | sds->this_nr_running = sgs.sum_nr_running; |
2735 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 2736 | sds->this_has_capacity = sgs.group_has_capacity; |
532cb4c4 | 2737 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 2738 | sds->max_load = sgs.avg_load; |
532cb4c4 | 2739 | sds->busiest = sg; |
1e3c88bd | 2740 | sds->busiest_nr_running = sgs.sum_nr_running; |
dd5feea1 | 2741 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 2742 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 2743 | sds->busiest_has_capacity = sgs.group_has_capacity; |
1e3c88bd PZ |
2744 | sds->group_imb = sgs.group_imb; |
2745 | } | |
2746 | ||
532cb4c4 MN |
2747 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
2748 | sg = sg->next; | |
2749 | } while (sg != sd->groups); | |
2750 | } | |
2751 | ||
2ec57d44 | 2752 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
2753 | { |
2754 | return 0*SD_ASYM_PACKING; | |
2755 | } | |
2756 | ||
2757 | /** | |
2758 | * check_asym_packing - Check to see if the group is packed into the | |
2759 | * sched doman. | |
2760 | * | |
2761 | * This is primarily intended to used at the sibling level. Some | |
2762 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
2763 | * case of POWER7, it can move to lower SMT modes only when higher | |
2764 | * threads are idle. When in lower SMT modes, the threads will | |
2765 | * perform better since they share less core resources. Hence when we | |
2766 | * have idle threads, we want them to be the higher ones. | |
2767 | * | |
2768 | * This packing function is run on idle threads. It checks to see if | |
2769 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
2770 | * CPU number than the packing function is being run on. Here we are | |
2771 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
2772 | * number. | |
2773 | * | |
b6b12294 MN |
2774 | * Returns 1 when packing is required and a task should be moved to |
2775 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
2776 | * | |
532cb4c4 MN |
2777 | * @sd: The sched_domain whose packing is to be checked. |
2778 | * @sds: Statistics of the sched_domain which is to be packed | |
2779 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2780 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
2781 | */ |
2782 | static int check_asym_packing(struct sched_domain *sd, | |
2783 | struct sd_lb_stats *sds, | |
2784 | int this_cpu, unsigned long *imbalance) | |
2785 | { | |
2786 | int busiest_cpu; | |
2787 | ||
2788 | if (!(sd->flags & SD_ASYM_PACKING)) | |
2789 | return 0; | |
2790 | ||
2791 | if (!sds->busiest) | |
2792 | return 0; | |
2793 | ||
2794 | busiest_cpu = group_first_cpu(sds->busiest); | |
2795 | if (this_cpu > busiest_cpu) | |
2796 | return 0; | |
2797 | ||
2798 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power, | |
2799 | SCHED_LOAD_SCALE); | |
2800 | return 1; | |
1e3c88bd PZ |
2801 | } |
2802 | ||
2803 | /** | |
2804 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
2805 | * amongst the groups of a sched_domain, during | |
2806 | * load balancing. | |
2807 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
2808 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
2809 | * @imbalance: Variable to store the imbalance. | |
2810 | */ | |
2811 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
2812 | int this_cpu, unsigned long *imbalance) | |
2813 | { | |
2814 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
2815 | unsigned int imbn = 2; | |
dd5feea1 | 2816 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
2817 | |
2818 | if (sds->this_nr_running) { | |
2819 | sds->this_load_per_task /= sds->this_nr_running; | |
2820 | if (sds->busiest_load_per_task > | |
2821 | sds->this_load_per_task) | |
2822 | imbn = 1; | |
2823 | } else | |
2824 | sds->this_load_per_task = | |
2825 | cpu_avg_load_per_task(this_cpu); | |
2826 | ||
dd5feea1 SS |
2827 | scaled_busy_load_per_task = sds->busiest_load_per_task |
2828 | * SCHED_LOAD_SCALE; | |
2829 | scaled_busy_load_per_task /= sds->busiest->cpu_power; | |
2830 | ||
2831 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
2832 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
2833 | *imbalance = sds->busiest_load_per_task; |
2834 | return; | |
2835 | } | |
2836 | ||
2837 | /* | |
2838 | * OK, we don't have enough imbalance to justify moving tasks, | |
2839 | * however we may be able to increase total CPU power used by | |
2840 | * moving them. | |
2841 | */ | |
2842 | ||
2843 | pwr_now += sds->busiest->cpu_power * | |
2844 | min(sds->busiest_load_per_task, sds->max_load); | |
2845 | pwr_now += sds->this->cpu_power * | |
2846 | min(sds->this_load_per_task, sds->this_load); | |
2847 | pwr_now /= SCHED_LOAD_SCALE; | |
2848 | ||
2849 | /* Amount of load we'd subtract */ | |
2850 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2851 | sds->busiest->cpu_power; | |
2852 | if (sds->max_load > tmp) | |
2853 | pwr_move += sds->busiest->cpu_power * | |
2854 | min(sds->busiest_load_per_task, sds->max_load - tmp); | |
2855 | ||
2856 | /* Amount of load we'd add */ | |
2857 | if (sds->max_load * sds->busiest->cpu_power < | |
2858 | sds->busiest_load_per_task * SCHED_LOAD_SCALE) | |
2859 | tmp = (sds->max_load * sds->busiest->cpu_power) / | |
2860 | sds->this->cpu_power; | |
2861 | else | |
2862 | tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) / | |
2863 | sds->this->cpu_power; | |
2864 | pwr_move += sds->this->cpu_power * | |
2865 | min(sds->this_load_per_task, sds->this_load + tmp); | |
2866 | pwr_move /= SCHED_LOAD_SCALE; | |
2867 | ||
2868 | /* Move if we gain throughput */ | |
2869 | if (pwr_move > pwr_now) | |
2870 | *imbalance = sds->busiest_load_per_task; | |
2871 | } | |
2872 | ||
2873 | /** | |
2874 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
2875 | * groups of a given sched_domain during load balance. | |
2876 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
2877 | * @this_cpu: Cpu for which currently load balance is being performed. | |
2878 | * @imbalance: The variable to store the imbalance. | |
2879 | */ | |
2880 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
2881 | unsigned long *imbalance) | |
2882 | { | |
dd5feea1 SS |
2883 | unsigned long max_pull, load_above_capacity = ~0UL; |
2884 | ||
2885 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
2886 | if (sds->group_imb) { | |
2887 | sds->busiest_load_per_task = | |
2888 | min(sds->busiest_load_per_task, sds->avg_load); | |
2889 | } | |
2890 | ||
1e3c88bd PZ |
2891 | /* |
2892 | * In the presence of smp nice balancing, certain scenarios can have | |
2893 | * max load less than avg load(as we skip the groups at or below | |
2894 | * its cpu_power, while calculating max_load..) | |
2895 | */ | |
2896 | if (sds->max_load < sds->avg_load) { | |
2897 | *imbalance = 0; | |
2898 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2899 | } | |
2900 | ||
dd5feea1 SS |
2901 | if (!sds->group_imb) { |
2902 | /* | |
2903 | * Don't want to pull so many tasks that a group would go idle. | |
2904 | */ | |
2905 | load_above_capacity = (sds->busiest_nr_running - | |
2906 | sds->busiest_group_capacity); | |
2907 | ||
2908 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE); | |
2909 | ||
2910 | load_above_capacity /= sds->busiest->cpu_power; | |
2911 | } | |
2912 | ||
2913 | /* | |
2914 | * We're trying to get all the cpus to the average_load, so we don't | |
2915 | * want to push ourselves above the average load, nor do we wish to | |
2916 | * reduce the max loaded cpu below the average load. At the same time, | |
2917 | * we also don't want to reduce the group load below the group capacity | |
2918 | * (so that we can implement power-savings policies etc). Thus we look | |
2919 | * for the minimum possible imbalance. | |
2920 | * Be careful of negative numbers as they'll appear as very large values | |
2921 | * with unsigned longs. | |
2922 | */ | |
2923 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
2924 | |
2925 | /* How much load to actually move to equalise the imbalance */ | |
2926 | *imbalance = min(max_pull * sds->busiest->cpu_power, | |
2927 | (sds->avg_load - sds->this_load) * sds->this->cpu_power) | |
2928 | / SCHED_LOAD_SCALE; | |
2929 | ||
2930 | /* | |
2931 | * if *imbalance is less than the average load per runnable task | |
2932 | * there is no gaurantee that any tasks will be moved so we'll have | |
2933 | * a think about bumping its value to force at least one task to be | |
2934 | * moved | |
2935 | */ | |
2936 | if (*imbalance < sds->busiest_load_per_task) | |
2937 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
2938 | ||
2939 | } | |
fab47622 | 2940 | |
1e3c88bd PZ |
2941 | /******* find_busiest_group() helpers end here *********************/ |
2942 | ||
2943 | /** | |
2944 | * find_busiest_group - Returns the busiest group within the sched_domain | |
2945 | * if there is an imbalance. If there isn't an imbalance, and | |
2946 | * the user has opted for power-savings, it returns a group whose | |
2947 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
2948 | * such a group exists. | |
2949 | * | |
2950 | * Also calculates the amount of weighted load which should be moved | |
2951 | * to restore balance. | |
2952 | * | |
2953 | * @sd: The sched_domain whose busiest group is to be returned. | |
2954 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
2955 | * @imbalance: Variable which stores amount of weighted load which should | |
2956 | * be moved to restore balance/put a group to idle. | |
2957 | * @idle: The idle status of this_cpu. | |
2958 | * @sd_idle: The idleness of sd | |
2959 | * @cpus: The set of CPUs under consideration for load-balancing. | |
2960 | * @balance: Pointer to a variable indicating if this_cpu | |
2961 | * is the appropriate cpu to perform load balancing at this_level. | |
2962 | * | |
2963 | * Returns: - the busiest group if imbalance exists. | |
2964 | * - If no imbalance and user has opted for power-savings balance, | |
2965 | * return the least loaded group whose CPUs can be | |
2966 | * put to idle by rebalancing its tasks onto our group. | |
2967 | */ | |
2968 | static struct sched_group * | |
2969 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
2970 | unsigned long *imbalance, enum cpu_idle_type idle, | |
2971 | int *sd_idle, const struct cpumask *cpus, int *balance) | |
2972 | { | |
2973 | struct sd_lb_stats sds; | |
2974 | ||
2975 | memset(&sds, 0, sizeof(sds)); | |
2976 | ||
2977 | /* | |
2978 | * Compute the various statistics relavent for load balancing at | |
2979 | * this level. | |
2980 | */ | |
2981 | update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus, | |
2982 | balance, &sds); | |
2983 | ||
2984 | /* Cases where imbalance does not exist from POV of this_cpu */ | |
2985 | /* 1) this_cpu is not the appropriate cpu to perform load balancing | |
2986 | * at this level. | |
2987 | * 2) There is no busy sibling group to pull from. | |
2988 | * 3) This group is the busiest group. | |
2989 | * 4) This group is more busy than the avg busieness at this | |
2990 | * sched_domain. | |
2991 | * 5) The imbalance is within the specified limit. | |
fab47622 NR |
2992 | * |
2993 | * Note: when doing newidle balance, if the local group has excess | |
2994 | * capacity (i.e. nr_running < group_capacity) and the busiest group | |
2995 | * does not have any capacity, we force a load balance to pull tasks | |
2996 | * to the local group. In this case, we skip past checks 3, 4 and 5. | |
1e3c88bd | 2997 | */ |
8f190fb3 | 2998 | if (!(*balance)) |
1e3c88bd PZ |
2999 | goto ret; |
3000 | ||
532cb4c4 MN |
3001 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
3002 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
3003 | return sds.busiest; | |
3004 | ||
1e3c88bd PZ |
3005 | if (!sds.busiest || sds.busiest_nr_running == 0) |
3006 | goto out_balanced; | |
3007 | ||
fab47622 NR |
3008 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
3009 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && | |
3010 | !sds.busiest_has_capacity) | |
3011 | goto force_balance; | |
3012 | ||
1e3c88bd PZ |
3013 | if (sds.this_load >= sds.max_load) |
3014 | goto out_balanced; | |
3015 | ||
3016 | sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr; | |
3017 | ||
3018 | if (sds.this_load >= sds.avg_load) | |
3019 | goto out_balanced; | |
3020 | ||
3021 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
3022 | goto out_balanced; | |
3023 | ||
fab47622 | 3024 | force_balance: |
1e3c88bd PZ |
3025 | /* Looks like there is an imbalance. Compute it */ |
3026 | calculate_imbalance(&sds, this_cpu, imbalance); | |
3027 | return sds.busiest; | |
3028 | ||
3029 | out_balanced: | |
3030 | /* | |
3031 | * There is no obvious imbalance. But check if we can do some balancing | |
3032 | * to save power. | |
3033 | */ | |
3034 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
3035 | return sds.busiest; | |
3036 | ret: | |
3037 | *imbalance = 0; | |
3038 | return NULL; | |
3039 | } | |
3040 | ||
3041 | /* | |
3042 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3043 | */ | |
3044 | static struct rq * | |
9d5efe05 SV |
3045 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
3046 | enum cpu_idle_type idle, unsigned long imbalance, | |
3047 | const struct cpumask *cpus) | |
1e3c88bd PZ |
3048 | { |
3049 | struct rq *busiest = NULL, *rq; | |
3050 | unsigned long max_load = 0; | |
3051 | int i; | |
3052 | ||
3053 | for_each_cpu(i, sched_group_cpus(group)) { | |
3054 | unsigned long power = power_of(i); | |
3055 | unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
3056 | unsigned long wl; | |
3057 | ||
9d5efe05 SV |
3058 | if (!capacity) |
3059 | capacity = fix_small_capacity(sd, group); | |
3060 | ||
1e3c88bd PZ |
3061 | if (!cpumask_test_cpu(i, cpus)) |
3062 | continue; | |
3063 | ||
3064 | rq = cpu_rq(i); | |
6e40f5bb | 3065 | wl = weighted_cpuload(i); |
1e3c88bd | 3066 | |
6e40f5bb TG |
3067 | /* |
3068 | * When comparing with imbalance, use weighted_cpuload() | |
3069 | * which is not scaled with the cpu power. | |
3070 | */ | |
1e3c88bd PZ |
3071 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
3072 | continue; | |
3073 | ||
6e40f5bb TG |
3074 | /* |
3075 | * For the load comparisons with the other cpu's, consider | |
3076 | * the weighted_cpuload() scaled with the cpu power, so that | |
3077 | * the load can be moved away from the cpu that is potentially | |
3078 | * running at a lower capacity. | |
3079 | */ | |
3080 | wl = (wl * SCHED_LOAD_SCALE) / power; | |
3081 | ||
1e3c88bd PZ |
3082 | if (wl > max_load) { |
3083 | max_load = wl; | |
3084 | busiest = rq; | |
3085 | } | |
3086 | } | |
3087 | ||
3088 | return busiest; | |
3089 | } | |
3090 | ||
3091 | /* | |
3092 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3093 | * so long as it is large enough. | |
3094 | */ | |
3095 | #define MAX_PINNED_INTERVAL 512 | |
3096 | ||
3097 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
3098 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
3099 | ||
532cb4c4 MN |
3100 | static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle, |
3101 | int busiest_cpu, int this_cpu) | |
1af3ed3d PZ |
3102 | { |
3103 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
3104 | |
3105 | /* | |
3106 | * ASYM_PACKING needs to force migrate tasks from busy but | |
3107 | * higher numbered CPUs in order to pack all tasks in the | |
3108 | * lowest numbered CPUs. | |
3109 | */ | |
3110 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
3111 | return 1; | |
3112 | ||
1af3ed3d PZ |
3113 | /* |
3114 | * The only task running in a non-idle cpu can be moved to this | |
3115 | * cpu in an attempt to completely freeup the other CPU | |
3116 | * package. | |
3117 | * | |
3118 | * The package power saving logic comes from | |
3119 | * find_busiest_group(). If there are no imbalance, then | |
3120 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3121 | * f_b_g() will select a group from which a running task may be | |
3122 | * pulled to this cpu in order to make the other package idle. | |
3123 | * If there is no opportunity to make a package idle and if | |
3124 | * there are no imbalance, then f_b_g() will return NULL and no | |
3125 | * action will be taken in load_balance_newidle(). | |
3126 | * | |
3127 | * Under normal task pull operation due to imbalance, there | |
3128 | * will be more than one task in the source run queue and | |
3129 | * move_tasks() will succeed. ld_moved will be true and this | |
3130 | * active balance code will not be triggered. | |
3131 | */ | |
3132 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3133 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3134 | return 0; | |
3135 | ||
3136 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) | |
3137 | return 0; | |
3138 | } | |
3139 | ||
3140 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3141 | } | |
3142 | ||
969c7921 TH |
3143 | static int active_load_balance_cpu_stop(void *data); |
3144 | ||
1e3c88bd PZ |
3145 | /* |
3146 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3147 | * tasks if there is an imbalance. | |
3148 | */ | |
3149 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3150 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3151 | int *balance) | |
3152 | { | |
3153 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; | |
3154 | struct sched_group *group; | |
3155 | unsigned long imbalance; | |
3156 | struct rq *busiest; | |
3157 | unsigned long flags; | |
3158 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3159 | ||
3160 | cpumask_copy(cpus, cpu_active_mask); | |
3161 | ||
3162 | /* | |
3163 | * When power savings policy is enabled for the parent domain, idle | |
3164 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3165 | * let the state of idle sibling percolate up as CPU_IDLE, instead of | |
3166 | * portraying it as CPU_NOT_IDLE. | |
3167 | */ | |
3168 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && | |
3169 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3170 | sd_idle = 1; | |
3171 | ||
3172 | schedstat_inc(sd, lb_count[idle]); | |
3173 | ||
3174 | redo: | |
1e3c88bd PZ |
3175 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, |
3176 | cpus, balance); | |
3177 | ||
3178 | if (*balance == 0) | |
3179 | goto out_balanced; | |
3180 | ||
3181 | if (!group) { | |
3182 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3183 | goto out_balanced; | |
3184 | } | |
3185 | ||
9d5efe05 | 3186 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3187 | if (!busiest) { |
3188 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3189 | goto out_balanced; | |
3190 | } | |
3191 | ||
3192 | BUG_ON(busiest == this_rq); | |
3193 | ||
3194 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3195 | ||
3196 | ld_moved = 0; | |
3197 | if (busiest->nr_running > 1) { | |
3198 | /* | |
3199 | * Attempt to move tasks. If find_busiest_group has found | |
3200 | * an imbalance but busiest->nr_running <= 1, the group is | |
3201 | * still unbalanced. ld_moved simply stays zero, so it is | |
3202 | * correctly treated as an imbalance. | |
3203 | */ | |
3204 | local_irq_save(flags); | |
3205 | double_rq_lock(this_rq, busiest); | |
3206 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3207 | imbalance, sd, idle, &all_pinned); | |
3208 | double_rq_unlock(this_rq, busiest); | |
3209 | local_irq_restore(flags); | |
3210 | ||
3211 | /* | |
3212 | * some other cpu did the load balance for us. | |
3213 | */ | |
3214 | if (ld_moved && this_cpu != smp_processor_id()) | |
3215 | resched_cpu(this_cpu); | |
3216 | ||
3217 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3218 | if (unlikely(all_pinned)) { | |
3219 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3220 | if (!cpumask_empty(cpus)) | |
3221 | goto redo; | |
3222 | goto out_balanced; | |
3223 | } | |
3224 | } | |
3225 | ||
3226 | if (!ld_moved) { | |
3227 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3228 | /* |
3229 | * Increment the failure counter only on periodic balance. | |
3230 | * We do not want newidle balance, which can be very | |
3231 | * frequent, pollute the failure counter causing | |
3232 | * excessive cache_hot migrations and active balances. | |
3233 | */ | |
3234 | if (idle != CPU_NEWLY_IDLE) | |
3235 | sd->nr_balance_failed++; | |
1e3c88bd | 3236 | |
532cb4c4 MN |
3237 | if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest), |
3238 | this_cpu)) { | |
1e3c88bd PZ |
3239 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3240 | ||
969c7921 TH |
3241 | /* don't kick the active_load_balance_cpu_stop, |
3242 | * if the curr task on busiest cpu can't be | |
3243 | * moved to this_cpu | |
1e3c88bd PZ |
3244 | */ |
3245 | if (!cpumask_test_cpu(this_cpu, | |
3246 | &busiest->curr->cpus_allowed)) { | |
3247 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3248 | flags); | |
3249 | all_pinned = 1; | |
3250 | goto out_one_pinned; | |
3251 | } | |
3252 | ||
969c7921 TH |
3253 | /* |
3254 | * ->active_balance synchronizes accesses to | |
3255 | * ->active_balance_work. Once set, it's cleared | |
3256 | * only after active load balance is finished. | |
3257 | */ | |
1e3c88bd PZ |
3258 | if (!busiest->active_balance) { |
3259 | busiest->active_balance = 1; | |
3260 | busiest->push_cpu = this_cpu; | |
3261 | active_balance = 1; | |
3262 | } | |
3263 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 3264 | |
1e3c88bd | 3265 | if (active_balance) |
969c7921 TH |
3266 | stop_one_cpu_nowait(cpu_of(busiest), |
3267 | active_load_balance_cpu_stop, busiest, | |
3268 | &busiest->active_balance_work); | |
1e3c88bd PZ |
3269 | |
3270 | /* | |
3271 | * We've kicked active balancing, reset the failure | |
3272 | * counter. | |
3273 | */ | |
3274 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3275 | } | |
3276 | } else | |
3277 | sd->nr_balance_failed = 0; | |
3278 | ||
3279 | if (likely(!active_balance)) { | |
3280 | /* We were unbalanced, so reset the balancing interval */ | |
3281 | sd->balance_interval = sd->min_interval; | |
3282 | } else { | |
3283 | /* | |
3284 | * If we've begun active balancing, start to back off. This | |
3285 | * case may not be covered by the all_pinned logic if there | |
3286 | * is only 1 task on the busy runqueue (because we don't call | |
3287 | * move_tasks). | |
3288 | */ | |
3289 | if (sd->balance_interval < sd->max_interval) | |
3290 | sd->balance_interval *= 2; | |
3291 | } | |
3292 | ||
3293 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3294 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3295 | ld_moved = -1; | |
3296 | ||
3297 | goto out; | |
3298 | ||
3299 | out_balanced: | |
3300 | schedstat_inc(sd, lb_balanced[idle]); | |
3301 | ||
3302 | sd->nr_balance_failed = 0; | |
3303 | ||
3304 | out_one_pinned: | |
3305 | /* tune up the balancing interval */ | |
3306 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3307 | (sd->balance_interval < sd->max_interval)) | |
3308 | sd->balance_interval *= 2; | |
3309 | ||
3310 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && | |
3311 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
3312 | ld_moved = -1; | |
3313 | else | |
3314 | ld_moved = 0; | |
3315 | out: | |
1e3c88bd PZ |
3316 | return ld_moved; |
3317 | } | |
3318 | ||
1e3c88bd PZ |
3319 | /* |
3320 | * idle_balance is called by schedule() if this_cpu is about to become | |
3321 | * idle. Attempts to pull tasks from other CPUs. | |
3322 | */ | |
3323 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3324 | { | |
3325 | struct sched_domain *sd; | |
3326 | int pulled_task = 0; | |
3327 | unsigned long next_balance = jiffies + HZ; | |
3328 | ||
3329 | this_rq->idle_stamp = this_rq->clock; | |
3330 | ||
3331 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3332 | return; | |
3333 | ||
f492e12e PZ |
3334 | /* |
3335 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
3336 | */ | |
3337 | raw_spin_unlock(&this_rq->lock); | |
3338 | ||
1e3c88bd PZ |
3339 | for_each_domain(this_cpu, sd) { |
3340 | unsigned long interval; | |
f492e12e | 3341 | int balance = 1; |
1e3c88bd PZ |
3342 | |
3343 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3344 | continue; | |
3345 | ||
f492e12e | 3346 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 3347 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
3348 | pulled_task = load_balance(this_cpu, this_rq, |
3349 | sd, CPU_NEWLY_IDLE, &balance); | |
3350 | } | |
1e3c88bd PZ |
3351 | |
3352 | interval = msecs_to_jiffies(sd->balance_interval); | |
3353 | if (time_after(next_balance, sd->last_balance + interval)) | |
3354 | next_balance = sd->last_balance + interval; | |
fab47622 | 3355 | if (pulled_task) |
1e3c88bd | 3356 | break; |
1e3c88bd | 3357 | } |
f492e12e PZ |
3358 | |
3359 | raw_spin_lock(&this_rq->lock); | |
3360 | ||
1e3c88bd PZ |
3361 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
3362 | /* | |
3363 | * We are going idle. next_balance may be set based on | |
3364 | * a busy processor. So reset next_balance. | |
3365 | */ | |
3366 | this_rq->next_balance = next_balance; | |
3367 | } | |
3368 | } | |
3369 | ||
3370 | /* | |
969c7921 TH |
3371 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
3372 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
3373 | * least 1 task to be running on each physical CPU where possible, and | |
3374 | * avoids physical / logical imbalances. | |
1e3c88bd | 3375 | */ |
969c7921 | 3376 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 3377 | { |
969c7921 TH |
3378 | struct rq *busiest_rq = data; |
3379 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 3380 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 3381 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 3382 | struct sched_domain *sd; |
969c7921 TH |
3383 | |
3384 | raw_spin_lock_irq(&busiest_rq->lock); | |
3385 | ||
3386 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
3387 | if (unlikely(busiest_cpu != smp_processor_id() || | |
3388 | !busiest_rq->active_balance)) | |
3389 | goto out_unlock; | |
1e3c88bd PZ |
3390 | |
3391 | /* Is there any task to move? */ | |
3392 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 3393 | goto out_unlock; |
1e3c88bd PZ |
3394 | |
3395 | /* | |
3396 | * This condition is "impossible", if it occurs | |
3397 | * we need to fix it. Originally reported by | |
3398 | * Bjorn Helgaas on a 128-cpu setup. | |
3399 | */ | |
3400 | BUG_ON(busiest_rq == target_rq); | |
3401 | ||
3402 | /* move a task from busiest_rq to target_rq */ | |
3403 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
3404 | |
3405 | /* Search for an sd spanning us and the target CPU. */ | |
3406 | for_each_domain(target_cpu, sd) { | |
3407 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3408 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3409 | break; | |
3410 | } | |
3411 | ||
3412 | if (likely(sd)) { | |
3413 | schedstat_inc(sd, alb_count); | |
3414 | ||
3415 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3416 | sd, CPU_IDLE)) | |
3417 | schedstat_inc(sd, alb_pushed); | |
3418 | else | |
3419 | schedstat_inc(sd, alb_failed); | |
3420 | } | |
3421 | double_unlock_balance(busiest_rq, target_rq); | |
969c7921 TH |
3422 | out_unlock: |
3423 | busiest_rq->active_balance = 0; | |
3424 | raw_spin_unlock_irq(&busiest_rq->lock); | |
3425 | return 0; | |
1e3c88bd PZ |
3426 | } |
3427 | ||
3428 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
3429 | |
3430 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
3431 | ||
3432 | static void trigger_sched_softirq(void *data) | |
3433 | { | |
3434 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
3435 | } | |
3436 | ||
3437 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
3438 | { | |
3439 | csd->func = trigger_sched_softirq; | |
3440 | csd->info = NULL; | |
3441 | csd->flags = 0; | |
3442 | csd->priv = 0; | |
3443 | } | |
3444 | ||
3445 | /* | |
3446 | * idle load balancing details | |
3447 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
3448 | * entering idle. | |
3449 | * - This idle load balancer CPU will also go into tickless mode when | |
3450 | * it is idle, just like all other idle CPUs | |
3451 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
3452 | * needed, they will kick the idle load balancer, which then does idle | |
3453 | * load balancing for all the idle CPUs. | |
3454 | */ | |
1e3c88bd PZ |
3455 | static struct { |
3456 | atomic_t load_balancer; | |
83cd4fe2 VP |
3457 | atomic_t first_pick_cpu; |
3458 | atomic_t second_pick_cpu; | |
3459 | cpumask_var_t idle_cpus_mask; | |
3460 | cpumask_var_t grp_idle_mask; | |
3461 | unsigned long next_balance; /* in jiffy units */ | |
3462 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
3463 | |
3464 | int get_nohz_load_balancer(void) | |
3465 | { | |
3466 | return atomic_read(&nohz.load_balancer); | |
3467 | } | |
3468 | ||
3469 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3470 | /** | |
3471 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3472 | * @cpu: The cpu whose lowest level of sched domain is to | |
3473 | * be returned. | |
3474 | * @flag: The flag to check for the lowest sched_domain | |
3475 | * for the given cpu. | |
3476 | * | |
3477 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3478 | */ | |
3479 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3480 | { | |
3481 | struct sched_domain *sd; | |
3482 | ||
3483 | for_each_domain(cpu, sd) | |
3484 | if (sd && (sd->flags & flag)) | |
3485 | break; | |
3486 | ||
3487 | return sd; | |
3488 | } | |
3489 | ||
3490 | /** | |
3491 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3492 | * @cpu: The cpu whose domains we're iterating over. | |
3493 | * @sd: variable holding the value of the power_savings_sd | |
3494 | * for cpu. | |
3495 | * @flag: The flag to filter the sched_domains to be iterated. | |
3496 | * | |
3497 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3498 | * set, starting from the lowest sched_domain to the highest. | |
3499 | */ | |
3500 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3501 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3502 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3503 | ||
3504 | /** | |
3505 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3506 | * @ilb_group: group to be checked for semi-idleness | |
3507 | * | |
3508 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3509 | * | |
3510 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3511 | * and atleast one non-idle CPU. This helper function checks if the given | |
3512 | * sched_group is semi-idle or not. | |
3513 | */ | |
3514 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3515 | { | |
83cd4fe2 | 3516 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
3517 | sched_group_cpus(ilb_group)); |
3518 | ||
3519 | /* | |
3520 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3521 | * and atleast one idle cpu. | |
3522 | */ | |
83cd4fe2 | 3523 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
3524 | return 0; |
3525 | ||
83cd4fe2 | 3526 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
3527 | return 0; |
3528 | ||
3529 | return 1; | |
3530 | } | |
3531 | /** | |
3532 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3533 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3534 | * | |
3535 | * Returns: Returns the id of the idle load balancer if it exists, | |
3536 | * Else, returns >= nr_cpu_ids. | |
3537 | * | |
3538 | * This algorithm picks the idle load balancer such that it belongs to a | |
3539 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3540 | * completely idle packages/cores just for the purpose of idle load balancing | |
3541 | * when there are other idle cpu's which are better suited for that job. | |
3542 | */ | |
3543 | static int find_new_ilb(int cpu) | |
3544 | { | |
3545 | struct sched_domain *sd; | |
3546 | struct sched_group *ilb_group; | |
3547 | ||
3548 | /* | |
3549 | * Have idle load balancer selection from semi-idle packages only | |
3550 | * when power-aware load balancing is enabled | |
3551 | */ | |
3552 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3553 | goto out_done; | |
3554 | ||
3555 | /* | |
3556 | * Optimize for the case when we have no idle CPUs or only one | |
3557 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3558 | */ | |
83cd4fe2 | 3559 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
3560 | goto out_done; |
3561 | ||
3562 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { | |
3563 | ilb_group = sd->groups; | |
3564 | ||
3565 | do { | |
3566 | if (is_semi_idle_group(ilb_group)) | |
83cd4fe2 | 3567 | return cpumask_first(nohz.grp_idle_mask); |
1e3c88bd PZ |
3568 | |
3569 | ilb_group = ilb_group->next; | |
3570 | ||
3571 | } while (ilb_group != sd->groups); | |
3572 | } | |
3573 | ||
3574 | out_done: | |
83cd4fe2 | 3575 | return nr_cpu_ids; |
1e3c88bd PZ |
3576 | } |
3577 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3578 | static inline int find_new_ilb(int call_cpu) | |
3579 | { | |
83cd4fe2 | 3580 | return nr_cpu_ids; |
1e3c88bd PZ |
3581 | } |
3582 | #endif | |
3583 | ||
83cd4fe2 VP |
3584 | /* |
3585 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
3586 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
3587 | * CPU (if there is one). | |
3588 | */ | |
3589 | static void nohz_balancer_kick(int cpu) | |
3590 | { | |
3591 | int ilb_cpu; | |
3592 | ||
3593 | nohz.next_balance++; | |
3594 | ||
3595 | ilb_cpu = get_nohz_load_balancer(); | |
3596 | ||
3597 | if (ilb_cpu >= nr_cpu_ids) { | |
3598 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
3599 | if (ilb_cpu >= nr_cpu_ids) | |
3600 | return; | |
3601 | } | |
3602 | ||
3603 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
3604 | struct call_single_data *cp; | |
3605 | ||
3606 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
3607 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
3608 | __smp_call_function_single(ilb_cpu, cp, 0); | |
3609 | } | |
3610 | return; | |
3611 | } | |
3612 | ||
1e3c88bd PZ |
3613 | /* |
3614 | * This routine will try to nominate the ilb (idle load balancing) | |
3615 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 3616 | * load balancing on behalf of all those cpus. |
1e3c88bd | 3617 | * |
83cd4fe2 VP |
3618 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
3619 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
3620 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 3621 | * |
83cd4fe2 VP |
3622 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
3623 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
3624 | * behalf of all idle CPUs). | |
1e3c88bd | 3625 | */ |
83cd4fe2 | 3626 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
3627 | { |
3628 | int cpu = smp_processor_id(); | |
3629 | ||
3630 | if (stop_tick) { | |
1e3c88bd PZ |
3631 | if (!cpu_active(cpu)) { |
3632 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 3633 | return; |
1e3c88bd PZ |
3634 | |
3635 | /* | |
3636 | * If we are going offline and still the leader, | |
3637 | * give up! | |
3638 | */ | |
83cd4fe2 VP |
3639 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3640 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3641 | BUG(); |
3642 | ||
83cd4fe2 | 3643 | return; |
1e3c88bd PZ |
3644 | } |
3645 | ||
83cd4fe2 | 3646 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 3647 | |
83cd4fe2 VP |
3648 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
3649 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
3650 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
3651 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 3652 | |
83cd4fe2 | 3653 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
3654 | int new_ilb; |
3655 | ||
83cd4fe2 VP |
3656 | /* make me the ilb owner */ |
3657 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
3658 | cpu) != nr_cpu_ids) | |
3659 | return; | |
3660 | ||
1e3c88bd PZ |
3661 | /* |
3662 | * Check to see if there is a more power-efficient | |
3663 | * ilb. | |
3664 | */ | |
3665 | new_ilb = find_new_ilb(cpu); | |
3666 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 3667 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 3668 | resched_cpu(new_ilb); |
83cd4fe2 | 3669 | return; |
1e3c88bd | 3670 | } |
83cd4fe2 | 3671 | return; |
1e3c88bd PZ |
3672 | } |
3673 | } else { | |
83cd4fe2 VP |
3674 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
3675 | return; | |
1e3c88bd | 3676 | |
83cd4fe2 | 3677 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
3678 | |
3679 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
3680 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3681 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3682 | BUG(); |
3683 | } | |
83cd4fe2 | 3684 | return; |
1e3c88bd PZ |
3685 | } |
3686 | #endif | |
3687 | ||
3688 | static DEFINE_SPINLOCK(balancing); | |
3689 | ||
3690 | /* | |
3691 | * It checks each scheduling domain to see if it is due to be balanced, | |
3692 | * and initiates a balancing operation if so. | |
3693 | * | |
3694 | * Balancing parameters are set up in arch_init_sched_domains. | |
3695 | */ | |
3696 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
3697 | { | |
3698 | int balance = 1; | |
3699 | struct rq *rq = cpu_rq(cpu); | |
3700 | unsigned long interval; | |
3701 | struct sched_domain *sd; | |
3702 | /* Earliest time when we have to do rebalance again */ | |
3703 | unsigned long next_balance = jiffies + 60*HZ; | |
3704 | int update_next_balance = 0; | |
3705 | int need_serialize; | |
3706 | ||
2069dd75 PZ |
3707 | update_shares(cpu); |
3708 | ||
1e3c88bd PZ |
3709 | for_each_domain(cpu, sd) { |
3710 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3711 | continue; | |
3712 | ||
3713 | interval = sd->balance_interval; | |
3714 | if (idle != CPU_IDLE) | |
3715 | interval *= sd->busy_factor; | |
3716 | ||
3717 | /* scale ms to jiffies */ | |
3718 | interval = msecs_to_jiffies(interval); | |
3719 | if (unlikely(!interval)) | |
3720 | interval = 1; | |
3721 | if (interval > HZ*NR_CPUS/10) | |
3722 | interval = HZ*NR_CPUS/10; | |
3723 | ||
3724 | need_serialize = sd->flags & SD_SERIALIZE; | |
3725 | ||
3726 | if (need_serialize) { | |
3727 | if (!spin_trylock(&balancing)) | |
3728 | goto out; | |
3729 | } | |
3730 | ||
3731 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
3732 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
3733 | /* | |
3734 | * We've pulled tasks over so either we're no | |
3735 | * longer idle, or one of our SMT siblings is | |
3736 | * not idle. | |
3737 | */ | |
3738 | idle = CPU_NOT_IDLE; | |
3739 | } | |
3740 | sd->last_balance = jiffies; | |
3741 | } | |
3742 | if (need_serialize) | |
3743 | spin_unlock(&balancing); | |
3744 | out: | |
3745 | if (time_after(next_balance, sd->last_balance + interval)) { | |
3746 | next_balance = sd->last_balance + interval; | |
3747 | update_next_balance = 1; | |
3748 | } | |
3749 | ||
3750 | /* | |
3751 | * Stop the load balance at this level. There is another | |
3752 | * CPU in our sched group which is doing load balancing more | |
3753 | * actively. | |
3754 | */ | |
3755 | if (!balance) | |
3756 | break; | |
3757 | } | |
3758 | ||
3759 | /* | |
3760 | * next_balance will be updated only when there is a need. | |
3761 | * When the cpu is attached to null domain for ex, it will not be | |
3762 | * updated. | |
3763 | */ | |
3764 | if (likely(update_next_balance)) | |
3765 | rq->next_balance = next_balance; | |
3766 | } | |
3767 | ||
83cd4fe2 | 3768 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 3769 | /* |
83cd4fe2 | 3770 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
3771 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
3772 | */ | |
83cd4fe2 VP |
3773 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
3774 | { | |
3775 | struct rq *this_rq = cpu_rq(this_cpu); | |
3776 | struct rq *rq; | |
3777 | int balance_cpu; | |
3778 | ||
3779 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
3780 | return; | |
3781 | ||
3782 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
3783 | if (balance_cpu == this_cpu) | |
3784 | continue; | |
3785 | ||
3786 | /* | |
3787 | * If this cpu gets work to do, stop the load balancing | |
3788 | * work being done for other cpus. Next load | |
3789 | * balancing owner will pick it up. | |
3790 | */ | |
3791 | if (need_resched()) { | |
3792 | this_rq->nohz_balance_kick = 0; | |
3793 | break; | |
3794 | } | |
3795 | ||
3796 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 3797 | update_rq_clock(this_rq); |
83cd4fe2 VP |
3798 | update_cpu_load(this_rq); |
3799 | raw_spin_unlock_irq(&this_rq->lock); | |
3800 | ||
3801 | rebalance_domains(balance_cpu, CPU_IDLE); | |
3802 | ||
3803 | rq = cpu_rq(balance_cpu); | |
3804 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
3805 | this_rq->next_balance = rq->next_balance; | |
3806 | } | |
3807 | nohz.next_balance = this_rq->next_balance; | |
3808 | this_rq->nohz_balance_kick = 0; | |
3809 | } | |
3810 | ||
3811 | /* | |
3812 | * Current heuristic for kicking the idle load balancer | |
3813 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
3814 | * idle load balancer when it has more than one process active. This | |
3815 | * eliminates the need for idle load balancing altogether when we have | |
3816 | * only one running process in the system (common case). | |
3817 | * - If there are more than one busy CPU, idle load balancer may have | |
3818 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
3819 | * SMT or core siblings and can run better if they move to different | |
3820 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
3821 | * which will kick idle load balancer as soon as it has any load. | |
3822 | */ | |
3823 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
3824 | { | |
3825 | unsigned long now = jiffies; | |
3826 | int ret; | |
3827 | int first_pick_cpu, second_pick_cpu; | |
3828 | ||
3829 | if (time_before(now, nohz.next_balance)) | |
3830 | return 0; | |
3831 | ||
f6c3f168 | 3832 | if (rq->idle_at_tick) |
83cd4fe2 VP |
3833 | return 0; |
3834 | ||
3835 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
3836 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
3837 | ||
3838 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
3839 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
3840 | return 0; | |
3841 | ||
3842 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
3843 | if (ret == nr_cpu_ids || ret == cpu) { | |
3844 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
3845 | if (rq->nr_running > 1) | |
3846 | return 1; | |
3847 | } else { | |
3848 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
3849 | if (ret == nr_cpu_ids || ret == cpu) { | |
3850 | if (rq->nr_running) | |
3851 | return 1; | |
3852 | } | |
3853 | } | |
3854 | return 0; | |
3855 | } | |
3856 | #else | |
3857 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
3858 | #endif | |
3859 | ||
3860 | /* | |
3861 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
3862 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
3863 | */ | |
1e3c88bd PZ |
3864 | static void run_rebalance_domains(struct softirq_action *h) |
3865 | { | |
3866 | int this_cpu = smp_processor_id(); | |
3867 | struct rq *this_rq = cpu_rq(this_cpu); | |
3868 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
3869 | CPU_IDLE : CPU_NOT_IDLE; | |
3870 | ||
3871 | rebalance_domains(this_cpu, idle); | |
3872 | ||
1e3c88bd | 3873 | /* |
83cd4fe2 | 3874 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
3875 | * balancing on behalf of the other idle cpus whose ticks are |
3876 | * stopped. | |
3877 | */ | |
83cd4fe2 | 3878 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
3879 | } |
3880 | ||
3881 | static inline int on_null_domain(int cpu) | |
3882 | { | |
90a6501f | 3883 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
3884 | } |
3885 | ||
3886 | /* | |
3887 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
3888 | */ |
3889 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
3890 | { | |
1e3c88bd PZ |
3891 | /* Don't need to rebalance while attached to NULL domain */ |
3892 | if (time_after_eq(jiffies, rq->next_balance) && | |
3893 | likely(!on_null_domain(cpu))) | |
3894 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
3895 | #ifdef CONFIG_NO_HZ |
3896 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
3897 | nohz_balancer_kick(cpu); | |
3898 | #endif | |
1e3c88bd PZ |
3899 | } |
3900 | ||
0bcdcf28 CE |
3901 | static void rq_online_fair(struct rq *rq) |
3902 | { | |
3903 | update_sysctl(); | |
3904 | } | |
3905 | ||
3906 | static void rq_offline_fair(struct rq *rq) | |
3907 | { | |
3908 | update_sysctl(); | |
3909 | } | |
3910 | ||
1e3c88bd PZ |
3911 | #else /* CONFIG_SMP */ |
3912 | ||
3913 | /* | |
3914 | * on UP we do not need to balance between CPUs: | |
3915 | */ | |
3916 | static inline void idle_balance(int cpu, struct rq *rq) | |
3917 | { | |
3918 | } | |
3919 | ||
55e12e5e | 3920 | #endif /* CONFIG_SMP */ |
e1d1484f | 3921 | |
bf0f6f24 IM |
3922 | /* |
3923 | * scheduler tick hitting a task of our scheduling class: | |
3924 | */ | |
8f4d37ec | 3925 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
3926 | { |
3927 | struct cfs_rq *cfs_rq; | |
3928 | struct sched_entity *se = &curr->se; | |
3929 | ||
3930 | for_each_sched_entity(se) { | |
3931 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 3932 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
3933 | } |
3934 | } | |
3935 | ||
3936 | /* | |
cd29fe6f PZ |
3937 | * called on fork with the child task as argument from the parent's context |
3938 | * - child not yet on the tasklist | |
3939 | * - preemption disabled | |
bf0f6f24 | 3940 | */ |
cd29fe6f | 3941 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 3942 | { |
cd29fe6f | 3943 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 3944 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 3945 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
3946 | struct rq *rq = this_rq(); |
3947 | unsigned long flags; | |
3948 | ||
05fa785c | 3949 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 3950 | |
861d034e PZ |
3951 | update_rq_clock(rq); |
3952 | ||
b0a0f667 PM |
3953 | if (unlikely(task_cpu(p) != this_cpu)) { |
3954 | rcu_read_lock(); | |
cd29fe6f | 3955 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
3956 | rcu_read_unlock(); |
3957 | } | |
bf0f6f24 | 3958 | |
7109c442 | 3959 | update_curr(cfs_rq); |
cd29fe6f | 3960 | |
b5d9d734 MG |
3961 | if (curr) |
3962 | se->vruntime = curr->vruntime; | |
aeb73b04 | 3963 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 3964 | |
cd29fe6f | 3965 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 3966 | /* |
edcb60a3 IM |
3967 | * Upon rescheduling, sched_class::put_prev_task() will place |
3968 | * 'current' within the tree based on its new key value. | |
3969 | */ | |
4d78e7b6 | 3970 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 3971 | resched_task(rq->curr); |
4d78e7b6 | 3972 | } |
bf0f6f24 | 3973 | |
88ec22d3 PZ |
3974 | se->vruntime -= cfs_rq->min_vruntime; |
3975 | ||
05fa785c | 3976 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
3977 | } |
3978 | ||
cb469845 SR |
3979 | /* |
3980 | * Priority of the task has changed. Check to see if we preempt | |
3981 | * the current task. | |
3982 | */ | |
3983 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
3984 | int oldprio, int running) | |
3985 | { | |
3986 | /* | |
3987 | * Reschedule if we are currently running on this runqueue and | |
3988 | * our priority decreased, or if we are not currently running on | |
3989 | * this runqueue and our priority is higher than the current's | |
3990 | */ | |
3991 | if (running) { | |
3992 | if (p->prio > oldprio) | |
3993 | resched_task(rq->curr); | |
3994 | } else | |
15afe09b | 3995 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
3996 | } |
3997 | ||
3998 | /* | |
3999 | * We switched to the sched_fair class. | |
4000 | */ | |
4001 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
4002 | int running) | |
4003 | { | |
4004 | /* | |
4005 | * We were most likely switched from sched_rt, so | |
4006 | * kick off the schedule if running, otherwise just see | |
4007 | * if we can still preempt the current task. | |
4008 | */ | |
4009 | if (running) | |
4010 | resched_task(rq->curr); | |
4011 | else | |
15afe09b | 4012 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4013 | } |
4014 | ||
83b699ed SV |
4015 | /* Account for a task changing its policy or group. |
4016 | * | |
4017 | * This routine is mostly called to set cfs_rq->curr field when a task | |
4018 | * migrates between groups/classes. | |
4019 | */ | |
4020 | static void set_curr_task_fair(struct rq *rq) | |
4021 | { | |
4022 | struct sched_entity *se = &rq->curr->se; | |
4023 | ||
4024 | for_each_sched_entity(se) | |
4025 | set_next_entity(cfs_rq_of(se), se); | |
4026 | } | |
4027 | ||
810b3817 | 4028 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4029 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 4030 | { |
b2b5ce02 PZ |
4031 | /* |
4032 | * If the task was not on the rq at the time of this cgroup movement | |
4033 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
4034 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
4035 | * bonus in place_entity()). | |
4036 | * | |
4037 | * If it was on the rq, we've just 'preempted' it, which does convert | |
4038 | * ->vruntime to a relative base. | |
4039 | * | |
4040 | * Make sure both cases convert their relative position when migrating | |
4041 | * to another cgroup's rq. This does somewhat interfere with the | |
4042 | * fair sleeper stuff for the first placement, but who cares. | |
4043 | */ | |
4044 | if (!on_rq) | |
4045 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
4046 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 4047 | if (!on_rq) |
b2b5ce02 | 4048 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
4049 | } |
4050 | #endif | |
4051 | ||
6d686f45 | 4052 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
4053 | { |
4054 | struct sched_entity *se = &task->se; | |
0d721cea PW |
4055 | unsigned int rr_interval = 0; |
4056 | ||
4057 | /* | |
4058 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
4059 | * idle runqueue: | |
4060 | */ | |
0d721cea PW |
4061 | if (rq->cfs.load.weight) |
4062 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
4063 | |
4064 | return rr_interval; | |
4065 | } | |
4066 | ||
bf0f6f24 IM |
4067 | /* |
4068 | * All the scheduling class methods: | |
4069 | */ | |
5522d5d5 IM |
4070 | static const struct sched_class fair_sched_class = { |
4071 | .next = &idle_sched_class, | |
bf0f6f24 IM |
4072 | .enqueue_task = enqueue_task_fair, |
4073 | .dequeue_task = dequeue_task_fair, | |
4074 | .yield_task = yield_task_fair, | |
4075 | ||
2e09bf55 | 4076 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
4077 | |
4078 | .pick_next_task = pick_next_task_fair, | |
4079 | .put_prev_task = put_prev_task_fair, | |
4080 | ||
681f3e68 | 4081 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
4082 | .select_task_rq = select_task_rq_fair, |
4083 | ||
0bcdcf28 CE |
4084 | .rq_online = rq_online_fair, |
4085 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
4086 | |
4087 | .task_waking = task_waking_fair, | |
681f3e68 | 4088 | #endif |
bf0f6f24 | 4089 | |
83b699ed | 4090 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 4091 | .task_tick = task_tick_fair, |
cd29fe6f | 4092 | .task_fork = task_fork_fair, |
cb469845 SR |
4093 | |
4094 | .prio_changed = prio_changed_fair, | |
4095 | .switched_to = switched_to_fair, | |
810b3817 | 4096 | |
0d721cea PW |
4097 | .get_rr_interval = get_rr_interval_fair, |
4098 | ||
810b3817 | 4099 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4100 | .task_move_group = task_move_group_fair, |
810b3817 | 4101 | #endif |
bf0f6f24 IM |
4102 | }; |
4103 | ||
4104 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 4105 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 4106 | { |
bf0f6f24 IM |
4107 | struct cfs_rq *cfs_rq; |
4108 | ||
5973e5b9 | 4109 | rcu_read_lock(); |
c3b64f1e | 4110 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 4111 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 4112 | rcu_read_unlock(); |
bf0f6f24 IM |
4113 | } |
4114 | #endif |