]>
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 AV |
23 | #include <linux/latencytop.h> |
24 | ||
bf0f6f24 | 25 | /* |
21805085 | 26 | * Targeted preemption latency for CPU-bound tasks: |
722aab0c | 27 | * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 28 | * |
21805085 | 29 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
30 | * 'timeslice length' - timeslices in CFS are of variable length |
31 | * and have no persistent notion like in traditional, time-slice | |
32 | * based scheduling concepts. | |
bf0f6f24 | 33 | * |
d274a4ce IM |
34 | * (to see the precise effective timeslice length of your workload, |
35 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 36 | */ |
19978ca6 | 37 | unsigned int sysctl_sched_latency = 20000000ULL; |
2bd8e6d4 IM |
38 | |
39 | /* | |
b2be5e96 | 40 | * Minimal preemption granularity for CPU-bound tasks: |
722aab0c | 41 | * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 42 | */ |
722aab0c | 43 | unsigned int sysctl_sched_min_granularity = 4000000ULL; |
21805085 PZ |
44 | |
45 | /* | |
b2be5e96 PZ |
46 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
47 | */ | |
722aab0c | 48 | static unsigned int sched_nr_latency = 5; |
b2be5e96 PZ |
49 | |
50 | /* | |
51 | * After fork, child runs first. (default) If set to 0 then | |
52 | * parent will (try to) run first. | |
21805085 | 53 | */ |
b2be5e96 | 54 | const_debug unsigned int sysctl_sched_child_runs_first = 1; |
bf0f6f24 | 55 | |
1799e35d IM |
56 | /* |
57 | * sys_sched_yield() compat mode | |
58 | * | |
59 | * This option switches the agressive yield implementation of the | |
60 | * old scheduler back on. | |
61 | */ | |
62 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
63 | ||
bf0f6f24 IM |
64 | /* |
65 | * SCHED_OTHER wake-up granularity. | |
103638d9 | 66 | * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
67 | * |
68 | * This option delays the preemption effects of decoupled workloads | |
69 | * and reduces their over-scheduling. Synchronous workloads will still | |
70 | * have immediate wakeup/sleep latencies. | |
71 | */ | |
103638d9 | 72 | unsigned int sysctl_sched_wakeup_granularity = 5000000UL; |
bf0f6f24 | 73 | |
da84d961 IM |
74 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
75 | ||
a4c2f00f PZ |
76 | static const struct sched_class fair_sched_class; |
77 | ||
bf0f6f24 IM |
78 | /************************************************************** |
79 | * CFS operations on generic schedulable entities: | |
80 | */ | |
81 | ||
b758149c PZ |
82 | static inline struct task_struct *task_of(struct sched_entity *se) |
83 | { | |
84 | return container_of(se, struct task_struct, se); | |
85 | } | |
86 | ||
62160e3f | 87 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 88 | |
62160e3f | 89 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
90 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
91 | { | |
62160e3f | 92 | return cfs_rq->rq; |
bf0f6f24 IM |
93 | } |
94 | ||
62160e3f IM |
95 | /* An entity is a task if it doesn't "own" a runqueue */ |
96 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 97 | |
b758149c PZ |
98 | /* Walk up scheduling entities hierarchy */ |
99 | #define for_each_sched_entity(se) \ | |
100 | for (; se; se = se->parent) | |
101 | ||
102 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
103 | { | |
104 | return p->se.cfs_rq; | |
105 | } | |
106 | ||
107 | /* runqueue on which this entity is (to be) queued */ | |
108 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
109 | { | |
110 | return se->cfs_rq; | |
111 | } | |
112 | ||
113 | /* runqueue "owned" by this group */ | |
114 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
115 | { | |
116 | return grp->my_q; | |
117 | } | |
118 | ||
119 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
120 | * another cpu ('this_cpu') | |
121 | */ | |
122 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
123 | { | |
124 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
125 | } | |
126 | ||
127 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
128 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
129 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
130 | ||
131 | /* Do the two (enqueued) entities belong to the same group ? */ | |
132 | static inline int | |
133 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
134 | { | |
135 | if (se->cfs_rq == pse->cfs_rq) | |
136 | return 1; | |
137 | ||
138 | return 0; | |
139 | } | |
140 | ||
141 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
142 | { | |
143 | return se->parent; | |
144 | } | |
145 | ||
464b7527 PZ |
146 | /* return depth at which a sched entity is present in the hierarchy */ |
147 | static inline int depth_se(struct sched_entity *se) | |
148 | { | |
149 | int depth = 0; | |
150 | ||
151 | for_each_sched_entity(se) | |
152 | depth++; | |
153 | ||
154 | return depth; | |
155 | } | |
156 | ||
157 | static void | |
158 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
159 | { | |
160 | int se_depth, pse_depth; | |
161 | ||
162 | /* | |
163 | * preemption test can be made between sibling entities who are in the | |
164 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
165 | * both tasks until we find their ancestors who are siblings of common | |
166 | * parent. | |
167 | */ | |
168 | ||
169 | /* First walk up until both entities are at same depth */ | |
170 | se_depth = depth_se(*se); | |
171 | pse_depth = depth_se(*pse); | |
172 | ||
173 | while (se_depth > pse_depth) { | |
174 | se_depth--; | |
175 | *se = parent_entity(*se); | |
176 | } | |
177 | ||
178 | while (pse_depth > se_depth) { | |
179 | pse_depth--; | |
180 | *pse = parent_entity(*pse); | |
181 | } | |
182 | ||
183 | while (!is_same_group(*se, *pse)) { | |
184 | *se = parent_entity(*se); | |
185 | *pse = parent_entity(*pse); | |
186 | } | |
187 | } | |
188 | ||
62160e3f | 189 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 190 | |
62160e3f IM |
191 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
192 | { | |
193 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
194 | } |
195 | ||
196 | #define entity_is_task(se) 1 | |
197 | ||
b758149c PZ |
198 | #define for_each_sched_entity(se) \ |
199 | for (; se; se = NULL) | |
bf0f6f24 | 200 | |
b758149c | 201 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 202 | { |
b758149c | 203 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
204 | } |
205 | ||
b758149c PZ |
206 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
207 | { | |
208 | struct task_struct *p = task_of(se); | |
209 | struct rq *rq = task_rq(p); | |
210 | ||
211 | return &rq->cfs; | |
212 | } | |
213 | ||
214 | /* runqueue "owned" by this group */ | |
215 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
216 | { | |
217 | return NULL; | |
218 | } | |
219 | ||
220 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
221 | { | |
222 | return &cpu_rq(this_cpu)->cfs; | |
223 | } | |
224 | ||
225 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
226 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
227 | ||
228 | static inline int | |
229 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
230 | { | |
231 | return 1; | |
232 | } | |
233 | ||
234 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
235 | { | |
236 | return NULL; | |
237 | } | |
238 | ||
464b7527 PZ |
239 | static inline void |
240 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
241 | { | |
242 | } | |
243 | ||
b758149c PZ |
244 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
245 | ||
bf0f6f24 IM |
246 | |
247 | /************************************************************** | |
248 | * Scheduling class tree data structure manipulation methods: | |
249 | */ | |
250 | ||
0702e3eb | 251 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 252 | { |
368059a9 PZ |
253 | s64 delta = (s64)(vruntime - min_vruntime); |
254 | if (delta > 0) | |
02e0431a PZ |
255 | min_vruntime = vruntime; |
256 | ||
257 | return min_vruntime; | |
258 | } | |
259 | ||
0702e3eb | 260 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
261 | { |
262 | s64 delta = (s64)(vruntime - min_vruntime); | |
263 | if (delta < 0) | |
264 | min_vruntime = vruntime; | |
265 | ||
266 | return min_vruntime; | |
267 | } | |
268 | ||
0702e3eb | 269 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 270 | { |
30cfdcfc | 271 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
272 | } |
273 | ||
1af5f730 PZ |
274 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
275 | { | |
276 | u64 vruntime = cfs_rq->min_vruntime; | |
277 | ||
278 | if (cfs_rq->curr) | |
279 | vruntime = cfs_rq->curr->vruntime; | |
280 | ||
281 | if (cfs_rq->rb_leftmost) { | |
282 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
283 | struct sched_entity, | |
284 | run_node); | |
285 | ||
e17036da | 286 | if (!cfs_rq->curr) |
1af5f730 PZ |
287 | vruntime = se->vruntime; |
288 | else | |
289 | vruntime = min_vruntime(vruntime, se->vruntime); | |
290 | } | |
291 | ||
292 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
293 | } | |
294 | ||
bf0f6f24 IM |
295 | /* |
296 | * Enqueue an entity into the rb-tree: | |
297 | */ | |
0702e3eb | 298 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
299 | { |
300 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
301 | struct rb_node *parent = NULL; | |
302 | struct sched_entity *entry; | |
9014623c | 303 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
304 | int leftmost = 1; |
305 | ||
306 | /* | |
307 | * Find the right place in the rbtree: | |
308 | */ | |
309 | while (*link) { | |
310 | parent = *link; | |
311 | entry = rb_entry(parent, struct sched_entity, run_node); | |
312 | /* | |
313 | * We dont care about collisions. Nodes with | |
314 | * the same key stay together. | |
315 | */ | |
9014623c | 316 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
317 | link = &parent->rb_left; |
318 | } else { | |
319 | link = &parent->rb_right; | |
320 | leftmost = 0; | |
321 | } | |
322 | } | |
323 | ||
324 | /* | |
325 | * Maintain a cache of leftmost tree entries (it is frequently | |
326 | * used): | |
327 | */ | |
1af5f730 | 328 | if (leftmost) |
57cb499d | 329 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
330 | |
331 | rb_link_node(&se->run_node, parent, link); | |
332 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
333 | } |
334 | ||
0702e3eb | 335 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 336 | { |
3fe69747 PZ |
337 | if (cfs_rq->rb_leftmost == &se->run_node) { |
338 | struct rb_node *next_node; | |
3fe69747 PZ |
339 | |
340 | next_node = rb_next(&se->run_node); | |
341 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 342 | } |
e9acbff6 | 343 | |
bf0f6f24 | 344 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
345 | } |
346 | ||
bf0f6f24 IM |
347 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
348 | { | |
f4b6755f PZ |
349 | struct rb_node *left = cfs_rq->rb_leftmost; |
350 | ||
351 | if (!left) | |
352 | return NULL; | |
353 | ||
354 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
355 | } |
356 | ||
f4b6755f | 357 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 358 | { |
7eee3e67 | 359 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 360 | |
70eee74b BS |
361 | if (!last) |
362 | return NULL; | |
7eee3e67 IM |
363 | |
364 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
365 | } |
366 | ||
bf0f6f24 IM |
367 | /************************************************************** |
368 | * Scheduling class statistics methods: | |
369 | */ | |
370 | ||
b2be5e96 PZ |
371 | #ifdef CONFIG_SCHED_DEBUG |
372 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
373 | struct file *filp, void __user *buffer, size_t *lenp, | |
374 | loff_t *ppos) | |
375 | { | |
376 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
377 | ||
378 | if (ret || !write) | |
379 | return ret; | |
380 | ||
381 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
382 | sysctl_sched_min_granularity); | |
383 | ||
384 | return 0; | |
385 | } | |
386 | #endif | |
647e7cac | 387 | |
a7be37ac | 388 | /* |
f9c0b095 | 389 | * delta /= w |
a7be37ac PZ |
390 | */ |
391 | static inline unsigned long | |
392 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
393 | { | |
f9c0b095 PZ |
394 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
395 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
396 | |
397 | return delta; | |
398 | } | |
399 | ||
647e7cac IM |
400 | /* |
401 | * The idea is to set a period in which each task runs once. | |
402 | * | |
403 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
404 | * this period because otherwise the slices get too small. | |
405 | * | |
406 | * p = (nr <= nl) ? l : l*nr/nl | |
407 | */ | |
4d78e7b6 PZ |
408 | static u64 __sched_period(unsigned long nr_running) |
409 | { | |
410 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 411 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
412 | |
413 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 414 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 415 | period *= nr_running; |
4d78e7b6 PZ |
416 | } |
417 | ||
418 | return period; | |
419 | } | |
420 | ||
647e7cac IM |
421 | /* |
422 | * We calculate the wall-time slice from the period by taking a part | |
423 | * proportional to the weight. | |
424 | * | |
f9c0b095 | 425 | * s = p*P[w/rw] |
647e7cac | 426 | */ |
6d0f0ebd | 427 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 428 | { |
0a582440 | 429 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 430 | |
0a582440 | 431 | for_each_sched_entity(se) { |
6272d68c LM |
432 | struct load_weight *load; |
433 | ||
434 | cfs_rq = cfs_rq_of(se); | |
435 | load = &cfs_rq->load; | |
f9c0b095 | 436 | |
0a582440 MG |
437 | if (unlikely(!se->on_rq)) { |
438 | struct load_weight lw = cfs_rq->load; | |
439 | ||
440 | update_load_add(&lw, se->load.weight); | |
441 | load = &lw; | |
442 | } | |
443 | slice = calc_delta_mine(slice, se->load.weight, load); | |
444 | } | |
445 | return slice; | |
bf0f6f24 IM |
446 | } |
447 | ||
647e7cac | 448 | /* |
ac884dec | 449 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 450 | * |
f9c0b095 | 451 | * vs = s/w |
647e7cac | 452 | */ |
f9c0b095 | 453 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 454 | { |
f9c0b095 | 455 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
456 | } |
457 | ||
bf0f6f24 IM |
458 | /* |
459 | * Update the current task's runtime statistics. Skip current tasks that | |
460 | * are not in our scheduling class. | |
461 | */ | |
462 | static inline void | |
8ebc91d9 IM |
463 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
464 | unsigned long delta_exec) | |
bf0f6f24 | 465 | { |
bbdba7c0 | 466 | unsigned long delta_exec_weighted; |
bf0f6f24 | 467 | |
8179ca23 | 468 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
469 | |
470 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 471 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 472 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
e9acbff6 | 473 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 474 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
475 | } |
476 | ||
b7cc0896 | 477 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 478 | { |
429d43bc | 479 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 480 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
481 | unsigned long delta_exec; |
482 | ||
483 | if (unlikely(!curr)) | |
484 | return; | |
485 | ||
486 | /* | |
487 | * Get the amount of time the current task was running | |
488 | * since the last time we changed load (this cannot | |
489 | * overflow on 32 bits): | |
490 | */ | |
8ebc91d9 | 491 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
492 | if (!delta_exec) |
493 | return; | |
bf0f6f24 | 494 | |
8ebc91d9 IM |
495 | __update_curr(cfs_rq, curr, delta_exec); |
496 | curr->exec_start = now; | |
d842de87 SV |
497 | |
498 | if (entity_is_task(curr)) { | |
499 | struct task_struct *curtask = task_of(curr); | |
500 | ||
501 | cpuacct_charge(curtask, delta_exec); | |
f06febc9 | 502 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 503 | } |
bf0f6f24 IM |
504 | } |
505 | ||
506 | static inline void | |
5870db5b | 507 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 508 | { |
d281918d | 509 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
510 | } |
511 | ||
bf0f6f24 IM |
512 | /* |
513 | * Task is being enqueued - update stats: | |
514 | */ | |
d2417e5a | 515 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 516 | { |
bf0f6f24 IM |
517 | /* |
518 | * Are we enqueueing a waiting task? (for current tasks | |
519 | * a dequeue/enqueue event is a NOP) | |
520 | */ | |
429d43bc | 521 | if (se != cfs_rq->curr) |
5870db5b | 522 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
523 | } |
524 | ||
bf0f6f24 | 525 | static void |
9ef0a961 | 526 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 527 | { |
bbdba7c0 IM |
528 | schedstat_set(se->wait_max, max(se->wait_max, |
529 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
530 | schedstat_set(se->wait_count, se->wait_count + 1); |
531 | schedstat_set(se->wait_sum, se->wait_sum + | |
532 | rq_of(cfs_rq)->clock - se->wait_start); | |
6cfb0d5d | 533 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
534 | } |
535 | ||
536 | static inline void | |
19b6a2e3 | 537 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 538 | { |
bf0f6f24 IM |
539 | /* |
540 | * Mark the end of the wait period if dequeueing a | |
541 | * waiting task: | |
542 | */ | |
429d43bc | 543 | if (se != cfs_rq->curr) |
9ef0a961 | 544 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
545 | } |
546 | ||
547 | /* | |
548 | * We are picking a new current task - update its stats: | |
549 | */ | |
550 | static inline void | |
79303e9e | 551 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
552 | { |
553 | /* | |
554 | * We are starting a new run period: | |
555 | */ | |
d281918d | 556 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
557 | } |
558 | ||
bf0f6f24 IM |
559 | /************************************************** |
560 | * Scheduling class queueing methods: | |
561 | */ | |
562 | ||
c09595f6 PZ |
563 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
564 | static void | |
565 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
566 | { | |
567 | cfs_rq->task_weight += weight; | |
568 | } | |
569 | #else | |
570 | static inline void | |
571 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
572 | { | |
573 | } | |
574 | #endif | |
575 | ||
30cfdcfc DA |
576 | static void |
577 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
578 | { | |
579 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
580 | if (!parent_entity(se)) |
581 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 582 | if (entity_is_task(se)) { |
c09595f6 | 583 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
584 | list_add(&se->group_node, &cfs_rq->tasks); |
585 | } | |
30cfdcfc DA |
586 | cfs_rq->nr_running++; |
587 | se->on_rq = 1; | |
588 | } | |
589 | ||
590 | static void | |
591 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
592 | { | |
593 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
594 | if (!parent_entity(se)) |
595 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 596 | if (entity_is_task(se)) { |
c09595f6 | 597 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
598 | list_del_init(&se->group_node); |
599 | } | |
30cfdcfc DA |
600 | cfs_rq->nr_running--; |
601 | se->on_rq = 0; | |
602 | } | |
603 | ||
2396af69 | 604 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 605 | { |
bf0f6f24 IM |
606 | #ifdef CONFIG_SCHEDSTATS |
607 | if (se->sleep_start) { | |
d281918d | 608 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 609 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
610 | |
611 | if ((s64)delta < 0) | |
612 | delta = 0; | |
613 | ||
614 | if (unlikely(delta > se->sleep_max)) | |
615 | se->sleep_max = delta; | |
616 | ||
617 | se->sleep_start = 0; | |
618 | se->sum_sleep_runtime += delta; | |
9745512c AV |
619 | |
620 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
621 | } |
622 | if (se->block_start) { | |
d281918d | 623 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 624 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
625 | |
626 | if ((s64)delta < 0) | |
627 | delta = 0; | |
628 | ||
629 | if (unlikely(delta > se->block_max)) | |
630 | se->block_max = delta; | |
631 | ||
632 | se->block_start = 0; | |
633 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
634 | |
635 | /* | |
636 | * Blocking time is in units of nanosecs, so shift by 20 to | |
637 | * get a milliseconds-range estimation of the amount of | |
638 | * time that the task spent sleeping: | |
639 | */ | |
640 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 641 | |
30084fbd IM |
642 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
643 | delta >> 20); | |
644 | } | |
9745512c | 645 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
646 | } |
647 | #endif | |
648 | } | |
649 | ||
ddc97297 PZ |
650 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
651 | { | |
652 | #ifdef CONFIG_SCHED_DEBUG | |
653 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
654 | ||
655 | if (d < 0) | |
656 | d = -d; | |
657 | ||
658 | if (d > 3*sysctl_sched_latency) | |
659 | schedstat_inc(cfs_rq, nr_spread_over); | |
660 | #endif | |
661 | } | |
662 | ||
aeb73b04 PZ |
663 | static void |
664 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
665 | { | |
1af5f730 | 666 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 667 | |
2cb8600e PZ |
668 | /* |
669 | * The 'current' period is already promised to the current tasks, | |
670 | * however the extra weight of the new task will slow them down a | |
671 | * little, place the new task so that it fits in the slot that | |
672 | * stays open at the end. | |
673 | */ | |
94dfb5e7 | 674 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 675 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 676 | |
8465e792 | 677 | if (!initial) { |
2cb8600e | 678 | /* sleeps upto a single latency don't count. */ |
a7be37ac PZ |
679 | if (sched_feat(NEW_FAIR_SLEEPERS)) { |
680 | unsigned long thresh = sysctl_sched_latency; | |
681 | ||
682 | /* | |
6bc912b7 PZ |
683 | * Convert the sleeper threshold into virtual time. |
684 | * SCHED_IDLE is a special sub-class. We care about | |
685 | * fairness only relative to other SCHED_IDLE tasks, | |
686 | * all of which have the same weight. | |
a7be37ac | 687 | */ |
6bc912b7 PZ |
688 | if (sched_feat(NORMALIZED_SLEEPER) && |
689 | task_of(se)->policy != SCHED_IDLE) | |
a7be37ac PZ |
690 | thresh = calc_delta_fair(thresh, se); |
691 | ||
692 | vruntime -= thresh; | |
693 | } | |
94359f05 | 694 | |
2cb8600e PZ |
695 | /* ensure we never gain time by being placed backwards. */ |
696 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
697 | } |
698 | ||
67e9fb2a | 699 | se->vruntime = vruntime; |
aeb73b04 PZ |
700 | } |
701 | ||
bf0f6f24 | 702 | static void |
83b699ed | 703 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
704 | { |
705 | /* | |
a2a2d680 | 706 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 707 | */ |
b7cc0896 | 708 | update_curr(cfs_rq); |
a992241d | 709 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 710 | |
e9acbff6 | 711 | if (wakeup) { |
aeb73b04 | 712 | place_entity(cfs_rq, se, 0); |
2396af69 | 713 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 714 | } |
bf0f6f24 | 715 | |
d2417e5a | 716 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 717 | check_spread(cfs_rq, se); |
83b699ed SV |
718 | if (se != cfs_rq->curr) |
719 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
720 | } |
721 | ||
2002c695 PZ |
722 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
723 | { | |
724 | if (cfs_rq->last == se) | |
725 | cfs_rq->last = NULL; | |
726 | ||
727 | if (cfs_rq->next == se) | |
728 | cfs_rq->next = NULL; | |
729 | } | |
730 | ||
bf0f6f24 | 731 | static void |
525c2716 | 732 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 733 | { |
a2a2d680 DA |
734 | /* |
735 | * Update run-time statistics of the 'current'. | |
736 | */ | |
737 | update_curr(cfs_rq); | |
738 | ||
19b6a2e3 | 739 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 740 | if (sleep) { |
67e9fb2a | 741 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
742 | if (entity_is_task(se)) { |
743 | struct task_struct *tsk = task_of(se); | |
744 | ||
745 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 746 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 747 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 748 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 749 | } |
db36cc7d | 750 | #endif |
67e9fb2a PZ |
751 | } |
752 | ||
2002c695 | 753 | clear_buddies(cfs_rq, se); |
4793241b | 754 | |
83b699ed | 755 | if (se != cfs_rq->curr) |
30cfdcfc DA |
756 | __dequeue_entity(cfs_rq, se); |
757 | account_entity_dequeue(cfs_rq, se); | |
1af5f730 | 758 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
759 | } |
760 | ||
761 | /* | |
762 | * Preempt the current task with a newly woken task if needed: | |
763 | */ | |
7c92e54f | 764 | static void |
2e09bf55 | 765 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 766 | { |
11697830 PZ |
767 | unsigned long ideal_runtime, delta_exec; |
768 | ||
6d0f0ebd | 769 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 770 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 771 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 772 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
773 | /* |
774 | * The current task ran long enough, ensure it doesn't get | |
775 | * re-elected due to buddy favours. | |
776 | */ | |
777 | clear_buddies(cfs_rq, curr); | |
778 | } | |
bf0f6f24 IM |
779 | } |
780 | ||
83b699ed | 781 | static void |
8494f412 | 782 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 783 | { |
83b699ed SV |
784 | /* 'current' is not kept within the tree. */ |
785 | if (se->on_rq) { | |
786 | /* | |
787 | * Any task has to be enqueued before it get to execute on | |
788 | * a CPU. So account for the time it spent waiting on the | |
789 | * runqueue. | |
790 | */ | |
791 | update_stats_wait_end(cfs_rq, se); | |
792 | __dequeue_entity(cfs_rq, se); | |
793 | } | |
794 | ||
79303e9e | 795 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 796 | cfs_rq->curr = se; |
eba1ed4b IM |
797 | #ifdef CONFIG_SCHEDSTATS |
798 | /* | |
799 | * Track our maximum slice length, if the CPU's load is at | |
800 | * least twice that of our own weight (i.e. dont track it | |
801 | * when there are only lesser-weight tasks around): | |
802 | */ | |
495eca49 | 803 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
804 | se->slice_max = max(se->slice_max, |
805 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
806 | } | |
807 | #endif | |
4a55b450 | 808 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
809 | } |
810 | ||
3f3a4904 PZ |
811 | static int |
812 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
813 | ||
f4b6755f | 814 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 815 | { |
f4b6755f PZ |
816 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
817 | ||
4793241b PZ |
818 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1) |
819 | return cfs_rq->next; | |
aa2ac252 | 820 | |
4793241b PZ |
821 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1) |
822 | return cfs_rq->last; | |
823 | ||
824 | return se; | |
aa2ac252 PZ |
825 | } |
826 | ||
ab6cde26 | 827 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
828 | { |
829 | /* | |
830 | * If still on the runqueue then deactivate_task() | |
831 | * was not called and update_curr() has to be done: | |
832 | */ | |
833 | if (prev->on_rq) | |
b7cc0896 | 834 | update_curr(cfs_rq); |
bf0f6f24 | 835 | |
ddc97297 | 836 | check_spread(cfs_rq, prev); |
30cfdcfc | 837 | if (prev->on_rq) { |
5870db5b | 838 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
839 | /* Put 'current' back into the tree. */ |
840 | __enqueue_entity(cfs_rq, prev); | |
841 | } | |
429d43bc | 842 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
843 | } |
844 | ||
8f4d37ec PZ |
845 | static void |
846 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 847 | { |
bf0f6f24 | 848 | /* |
30cfdcfc | 849 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 850 | */ |
30cfdcfc | 851 | update_curr(cfs_rq); |
bf0f6f24 | 852 | |
8f4d37ec PZ |
853 | #ifdef CONFIG_SCHED_HRTICK |
854 | /* | |
855 | * queued ticks are scheduled to match the slice, so don't bother | |
856 | * validating it and just reschedule. | |
857 | */ | |
983ed7a6 HH |
858 | if (queued) { |
859 | resched_task(rq_of(cfs_rq)->curr); | |
860 | return; | |
861 | } | |
8f4d37ec PZ |
862 | /* |
863 | * don't let the period tick interfere with the hrtick preemption | |
864 | */ | |
865 | if (!sched_feat(DOUBLE_TICK) && | |
866 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
867 | return; | |
868 | #endif | |
869 | ||
ce6c1311 | 870 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 871 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
872 | } |
873 | ||
874 | /************************************************** | |
875 | * CFS operations on tasks: | |
876 | */ | |
877 | ||
8f4d37ec PZ |
878 | #ifdef CONFIG_SCHED_HRTICK |
879 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
880 | { | |
8f4d37ec PZ |
881 | struct sched_entity *se = &p->se; |
882 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
883 | ||
884 | WARN_ON(task_rq(p) != rq); | |
885 | ||
886 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
887 | u64 slice = sched_slice(cfs_rq, se); | |
888 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
889 | s64 delta = slice - ran; | |
890 | ||
891 | if (delta < 0) { | |
892 | if (rq->curr == p) | |
893 | resched_task(p); | |
894 | return; | |
895 | } | |
896 | ||
897 | /* | |
898 | * Don't schedule slices shorter than 10000ns, that just | |
899 | * doesn't make sense. Rely on vruntime for fairness. | |
900 | */ | |
31656519 | 901 | if (rq->curr != p) |
157124c1 | 902 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 903 | |
31656519 | 904 | hrtick_start(rq, delta); |
8f4d37ec PZ |
905 | } |
906 | } | |
a4c2f00f PZ |
907 | |
908 | /* | |
909 | * called from enqueue/dequeue and updates the hrtick when the | |
910 | * current task is from our class and nr_running is low enough | |
911 | * to matter. | |
912 | */ | |
913 | static void hrtick_update(struct rq *rq) | |
914 | { | |
915 | struct task_struct *curr = rq->curr; | |
916 | ||
917 | if (curr->sched_class != &fair_sched_class) | |
918 | return; | |
919 | ||
920 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
921 | hrtick_start_fair(rq, curr); | |
922 | } | |
55e12e5e | 923 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
924 | static inline void |
925 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
926 | { | |
927 | } | |
a4c2f00f PZ |
928 | |
929 | static inline void hrtick_update(struct rq *rq) | |
930 | { | |
931 | } | |
8f4d37ec PZ |
932 | #endif |
933 | ||
bf0f6f24 IM |
934 | /* |
935 | * The enqueue_task method is called before nr_running is | |
936 | * increased. Here we update the fair scheduling stats and | |
937 | * then put the task into the rbtree: | |
938 | */ | |
fd390f6a | 939 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
940 | { |
941 | struct cfs_rq *cfs_rq; | |
62fb1851 | 942 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
943 | |
944 | for_each_sched_entity(se) { | |
62fb1851 | 945 | if (se->on_rq) |
bf0f6f24 IM |
946 | break; |
947 | cfs_rq = cfs_rq_of(se); | |
83b699ed | 948 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 949 | wakeup = 1; |
bf0f6f24 | 950 | } |
8f4d37ec | 951 | |
a4c2f00f | 952 | hrtick_update(rq); |
bf0f6f24 IM |
953 | } |
954 | ||
955 | /* | |
956 | * The dequeue_task method is called before nr_running is | |
957 | * decreased. We remove the task from the rbtree and | |
958 | * update the fair scheduling stats: | |
959 | */ | |
f02231e5 | 960 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
961 | { |
962 | struct cfs_rq *cfs_rq; | |
62fb1851 | 963 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
964 | |
965 | for_each_sched_entity(se) { | |
966 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 967 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 968 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 969 | if (cfs_rq->load.weight) |
bf0f6f24 | 970 | break; |
b9fa3df3 | 971 | sleep = 1; |
bf0f6f24 | 972 | } |
8f4d37ec | 973 | |
a4c2f00f | 974 | hrtick_update(rq); |
bf0f6f24 IM |
975 | } |
976 | ||
977 | /* | |
1799e35d IM |
978 | * sched_yield() support is very simple - we dequeue and enqueue. |
979 | * | |
980 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 981 | */ |
4530d7ab | 982 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 983 | { |
db292ca3 IM |
984 | struct task_struct *curr = rq->curr; |
985 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
986 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
987 | |
988 | /* | |
1799e35d IM |
989 | * Are we the only task in the tree? |
990 | */ | |
991 | if (unlikely(cfs_rq->nr_running == 1)) | |
992 | return; | |
993 | ||
2002c695 PZ |
994 | clear_buddies(cfs_rq, se); |
995 | ||
db292ca3 | 996 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 997 | update_rq_clock(rq); |
1799e35d | 998 | /* |
a2a2d680 | 999 | * Update run-time statistics of the 'current'. |
1799e35d | 1000 | */ |
2b1e315d | 1001 | update_curr(cfs_rq); |
1799e35d IM |
1002 | |
1003 | return; | |
1004 | } | |
1005 | /* | |
1006 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1007 | */ |
2b1e315d | 1008 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1009 | /* |
1010 | * Already in the rightmost position? | |
1011 | */ | |
79b3feff | 1012 | if (unlikely(!rightmost || rightmost->vruntime < se->vruntime)) |
1799e35d IM |
1013 | return; |
1014 | ||
1015 | /* | |
1016 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1017 | * Upon rescheduling, sched_class::put_prev_task() will place |
1018 | * 'current' within the tree based on its new key value. | |
1799e35d | 1019 | */ |
30cfdcfc | 1020 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1021 | } |
1022 | ||
e7693a36 GH |
1023 | /* |
1024 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1025 | * not idle and an idle cpu is available. The span of cpus to | |
1026 | * search starts with cpus closest then further out as needed, | |
1027 | * so we always favor a closer, idle cpu. | |
e761b772 | 1028 | * Domains may include CPUs that are not usable for migration, |
96f874e2 | 1029 | * hence we need to mask them out (cpu_active_mask) |
e7693a36 GH |
1030 | * |
1031 | * Returns the CPU we should wake onto. | |
1032 | */ | |
1033 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1034 | static int wake_idle(int cpu, struct task_struct *p) | |
1035 | { | |
e7693a36 GH |
1036 | struct sched_domain *sd; |
1037 | int i; | |
7eb52dfa VS |
1038 | unsigned int chosen_wakeup_cpu; |
1039 | int this_cpu; | |
1040 | ||
1041 | /* | |
1042 | * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu | |
1043 | * are idle and this is not a kernel thread and this task's affinity | |
1044 | * allows it to be moved to preferred cpu, then just move! | |
1045 | */ | |
1046 | ||
1047 | this_cpu = smp_processor_id(); | |
1048 | chosen_wakeup_cpu = | |
1049 | cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu; | |
1050 | ||
1051 | if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP && | |
1052 | idle_cpu(cpu) && idle_cpu(this_cpu) && | |
1053 | p->mm && !(p->flags & PF_KTHREAD) && | |
1054 | cpu_isset(chosen_wakeup_cpu, p->cpus_allowed)) | |
1055 | return chosen_wakeup_cpu; | |
e7693a36 GH |
1056 | |
1057 | /* | |
1058 | * If it is idle, then it is the best cpu to run this task. | |
1059 | * | |
1060 | * This cpu is also the best, if it has more than one task already. | |
1061 | * Siblings must be also busy(in most cases) as they didn't already | |
1062 | * pickup the extra load from this cpu and hence we need not check | |
1063 | * sibling runqueue info. This will avoid the checks and cache miss | |
1064 | * penalities associated with that. | |
1065 | */ | |
104f6454 | 1066 | if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1) |
e7693a36 GH |
1067 | return cpu; |
1068 | ||
1069 | for_each_domain(cpu, sd) { | |
1d3504fc HS |
1070 | if ((sd->flags & SD_WAKE_IDLE) |
1071 | || ((sd->flags & SD_WAKE_IDLE_FAR) | |
1072 | && !task_hot(p, task_rq(p)->clock, sd))) { | |
758b2cdc RR |
1073 | for_each_cpu_and(i, sched_domain_span(sd), |
1074 | &p->cpus_allowed) { | |
1075 | if (cpu_active(i) && idle_cpu(i)) { | |
e7693a36 GH |
1076 | if (i != task_cpu(p)) { |
1077 | schedstat_inc(p, | |
1078 | se.nr_wakeups_idle); | |
1079 | } | |
1080 | return i; | |
1081 | } | |
1082 | } | |
1083 | } else { | |
1084 | break; | |
1085 | } | |
1086 | } | |
1087 | return cpu; | |
1088 | } | |
55e12e5e | 1089 | #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/ |
e7693a36 GH |
1090 | static inline int wake_idle(int cpu, struct task_struct *p) |
1091 | { | |
1092 | return cpu; | |
1093 | } | |
1094 | #endif | |
1095 | ||
1096 | #ifdef CONFIG_SMP | |
098fb9db | 1097 | |
bb3469ac | 1098 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1099 | /* |
1100 | * effective_load() calculates the load change as seen from the root_task_group | |
1101 | * | |
1102 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1103 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1104 | * can calculate the shift in shares. | |
1105 | * | |
1106 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1107 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1108 | * this change. | |
1109 | * | |
1110 | * We compensate this by not only taking the current delta into account, but | |
1111 | * also considering the delta between when the shares were last adjusted and | |
1112 | * now. | |
1113 | * | |
1114 | * We still saw a performance dip, some tracing learned us that between | |
1115 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1116 | * significantly. Therefore try to bias the error in direction of failing | |
1117 | * the affine wakeup. | |
1118 | * | |
1119 | */ | |
f1d239f7 PZ |
1120 | static long effective_load(struct task_group *tg, int cpu, |
1121 | long wl, long wg) | |
bb3469ac | 1122 | { |
4be9daaa | 1123 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1124 | |
1125 | if (!tg->parent) | |
1126 | return wl; | |
1127 | ||
f5bfb7d9 PZ |
1128 | /* |
1129 | * By not taking the decrease of shares on the other cpu into | |
1130 | * account our error leans towards reducing the affine wakeups. | |
1131 | */ | |
1132 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1133 | return wl; | |
1134 | ||
4be9daaa | 1135 | for_each_sched_entity(se) { |
cb5ef42a | 1136 | long S, rw, s, a, b; |
940959e9 PZ |
1137 | long more_w; |
1138 | ||
1139 | /* | |
1140 | * Instead of using this increment, also add the difference | |
1141 | * between when the shares were last updated and now. | |
1142 | */ | |
1143 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1144 | wl += more_w; | |
1145 | wg += more_w; | |
4be9daaa PZ |
1146 | |
1147 | S = se->my_q->tg->shares; | |
1148 | s = se->my_q->shares; | |
f1d239f7 | 1149 | rw = se->my_q->rq_weight; |
bb3469ac | 1150 | |
cb5ef42a PZ |
1151 | a = S*(rw + wl); |
1152 | b = S*rw + s*wg; | |
4be9daaa | 1153 | |
940959e9 PZ |
1154 | wl = s*(a-b); |
1155 | ||
1156 | if (likely(b)) | |
1157 | wl /= b; | |
1158 | ||
83378269 PZ |
1159 | /* |
1160 | * Assume the group is already running and will | |
1161 | * thus already be accounted for in the weight. | |
1162 | * | |
1163 | * That is, moving shares between CPUs, does not | |
1164 | * alter the group weight. | |
1165 | */ | |
4be9daaa | 1166 | wg = 0; |
4be9daaa | 1167 | } |
bb3469ac | 1168 | |
4be9daaa | 1169 | return wl; |
bb3469ac | 1170 | } |
4be9daaa | 1171 | |
bb3469ac | 1172 | #else |
4be9daaa | 1173 | |
83378269 PZ |
1174 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1175 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1176 | { |
83378269 | 1177 | return wl; |
bb3469ac | 1178 | } |
4be9daaa | 1179 | |
bb3469ac PZ |
1180 | #endif |
1181 | ||
098fb9db | 1182 | static int |
64b9e029 | 1183 | wake_affine(struct sched_domain *this_sd, struct rq *this_rq, |
4ae7d5ce IM |
1184 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, |
1185 | int idx, unsigned long load, unsigned long this_load, | |
098fb9db IM |
1186 | unsigned int imbalance) |
1187 | { | |
1188 | unsigned long tl = this_load; | |
1189 | unsigned long tl_per_task; | |
d942fb6c | 1190 | struct task_group *tg; |
83378269 | 1191 | unsigned long weight; |
b3137bc8 | 1192 | int balanced; |
098fb9db | 1193 | |
b3137bc8 | 1194 | if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS)) |
098fb9db IM |
1195 | return 0; |
1196 | ||
b3137bc8 MG |
1197 | /* |
1198 | * If sync wakeup then subtract the (maximum possible) | |
1199 | * effect of the currently running task from the load | |
1200 | * of the current CPU: | |
1201 | */ | |
83378269 PZ |
1202 | if (sync) { |
1203 | tg = task_group(current); | |
1204 | weight = current->se.load.weight; | |
1205 | ||
1206 | tl += effective_load(tg, this_cpu, -weight, -weight); | |
1207 | load += effective_load(tg, prev_cpu, 0, -weight); | |
1208 | } | |
b3137bc8 | 1209 | |
83378269 PZ |
1210 | tg = task_group(p); |
1211 | weight = p->se.load.weight; | |
b3137bc8 | 1212 | |
83378269 PZ |
1213 | balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <= |
1214 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); | |
b3137bc8 | 1215 | |
098fb9db | 1216 | /* |
4ae7d5ce IM |
1217 | * If the currently running task will sleep within |
1218 | * a reasonable amount of time then attract this newly | |
1219 | * woken task: | |
098fb9db | 1220 | */ |
2fb7635c PZ |
1221 | if (sync && balanced) |
1222 | return 1; | |
098fb9db IM |
1223 | |
1224 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1225 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1226 | ||
64b9e029 AA |
1227 | if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <= |
1228 | tl_per_task)) { | |
098fb9db IM |
1229 | /* |
1230 | * This domain has SD_WAKE_AFFINE and | |
1231 | * p is cache cold in this domain, and | |
1232 | * there is no bad imbalance. | |
1233 | */ | |
1234 | schedstat_inc(this_sd, ttwu_move_affine); | |
1235 | schedstat_inc(p, se.nr_wakeups_affine); | |
1236 | ||
1237 | return 1; | |
1238 | } | |
1239 | return 0; | |
1240 | } | |
1241 | ||
e7693a36 GH |
1242 | static int select_task_rq_fair(struct task_struct *p, int sync) |
1243 | { | |
e7693a36 | 1244 | struct sched_domain *sd, *this_sd = NULL; |
ac192d39 | 1245 | int prev_cpu, this_cpu, new_cpu; |
098fb9db | 1246 | unsigned long load, this_load; |
64b9e029 | 1247 | struct rq *this_rq; |
098fb9db | 1248 | unsigned int imbalance; |
098fb9db | 1249 | int idx; |
e7693a36 | 1250 | |
ac192d39 | 1251 | prev_cpu = task_cpu(p); |
ac192d39 | 1252 | this_cpu = smp_processor_id(); |
4ae7d5ce | 1253 | this_rq = cpu_rq(this_cpu); |
ac192d39 | 1254 | new_cpu = prev_cpu; |
e7693a36 | 1255 | |
64b9e029 AA |
1256 | if (prev_cpu == this_cpu) |
1257 | goto out; | |
ac192d39 IM |
1258 | /* |
1259 | * 'this_sd' is the first domain that both | |
1260 | * this_cpu and prev_cpu are present in: | |
1261 | */ | |
e7693a36 | 1262 | for_each_domain(this_cpu, sd) { |
758b2cdc | 1263 | if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) { |
e7693a36 GH |
1264 | this_sd = sd; |
1265 | break; | |
1266 | } | |
1267 | } | |
1268 | ||
96f874e2 | 1269 | if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed))) |
f4827386 | 1270 | goto out; |
e7693a36 GH |
1271 | |
1272 | /* | |
1273 | * Check for affine wakeup and passive balancing possibilities. | |
1274 | */ | |
098fb9db | 1275 | if (!this_sd) |
f4827386 | 1276 | goto out; |
e7693a36 | 1277 | |
098fb9db IM |
1278 | idx = this_sd->wake_idx; |
1279 | ||
1280 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1281 | ||
ac192d39 | 1282 | load = source_load(prev_cpu, idx); |
098fb9db IM |
1283 | this_load = target_load(this_cpu, idx); |
1284 | ||
64b9e029 | 1285 | if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, |
4ae7d5ce IM |
1286 | load, this_load, imbalance)) |
1287 | return this_cpu; | |
1288 | ||
098fb9db IM |
1289 | /* |
1290 | * Start passive balancing when half the imbalance_pct | |
1291 | * limit is reached. | |
1292 | */ | |
1293 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1294 | if (imbalance*this_load <= 100*load) { | |
1295 | schedstat_inc(this_sd, ttwu_move_balance); | |
1296 | schedstat_inc(p, se.nr_wakeups_passive); | |
4ae7d5ce | 1297 | return this_cpu; |
e7693a36 GH |
1298 | } |
1299 | } | |
1300 | ||
f4827386 | 1301 | out: |
e7693a36 GH |
1302 | return wake_idle(new_cpu, p); |
1303 | } | |
1304 | #endif /* CONFIG_SMP */ | |
1305 | ||
0bbd3336 PZ |
1306 | static unsigned long wakeup_gran(struct sched_entity *se) |
1307 | { | |
1308 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1309 | ||
1310 | /* | |
a7be37ac PZ |
1311 | * More easily preempt - nice tasks, while not making it harder for |
1312 | * + nice tasks. | |
0bbd3336 | 1313 | */ |
464b7527 PZ |
1314 | if (!sched_feat(ASYM_GRAN) || se->load.weight > NICE_0_LOAD) |
1315 | gran = calc_delta_fair(sysctl_sched_wakeup_granularity, se); | |
0bbd3336 PZ |
1316 | |
1317 | return gran; | |
1318 | } | |
1319 | ||
464b7527 PZ |
1320 | /* |
1321 | * Should 'se' preempt 'curr'. | |
1322 | * | |
1323 | * |s1 | |
1324 | * |s2 | |
1325 | * |s3 | |
1326 | * g | |
1327 | * |<--->|c | |
1328 | * | |
1329 | * w(c, s1) = -1 | |
1330 | * w(c, s2) = 0 | |
1331 | * w(c, s3) = 1 | |
1332 | * | |
1333 | */ | |
1334 | static int | |
1335 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1336 | { | |
1337 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1338 | ||
1339 | if (vdiff <= 0) | |
1340 | return -1; | |
1341 | ||
1342 | gran = wakeup_gran(curr); | |
1343 | if (vdiff > gran) | |
1344 | return 1; | |
1345 | ||
1346 | return 0; | |
1347 | } | |
1348 | ||
02479099 PZ |
1349 | static void set_last_buddy(struct sched_entity *se) |
1350 | { | |
6bc912b7 PZ |
1351 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1352 | for_each_sched_entity(se) | |
1353 | cfs_rq_of(se)->last = se; | |
1354 | } | |
02479099 PZ |
1355 | } |
1356 | ||
1357 | static void set_next_buddy(struct sched_entity *se) | |
1358 | { | |
6bc912b7 PZ |
1359 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1360 | for_each_sched_entity(se) | |
1361 | cfs_rq_of(se)->next = se; | |
1362 | } | |
02479099 PZ |
1363 | } |
1364 | ||
bf0f6f24 IM |
1365 | /* |
1366 | * Preempt the current task with a newly woken task if needed: | |
1367 | */ | |
15afe09b | 1368 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync) |
bf0f6f24 IM |
1369 | { |
1370 | struct task_struct *curr = rq->curr; | |
8651a86c | 1371 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1372 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
bf0f6f24 | 1373 | |
03e89e45 | 1374 | update_curr(cfs_rq); |
4793241b | 1375 | |
03e89e45 | 1376 | if (unlikely(rt_prio(p->prio))) { |
bf0f6f24 IM |
1377 | resched_task(curr); |
1378 | return; | |
1379 | } | |
aa2ac252 | 1380 | |
d95f98d0 PZ |
1381 | if (unlikely(p->sched_class != &fair_sched_class)) |
1382 | return; | |
1383 | ||
4ae7d5ce IM |
1384 | if (unlikely(se == pse)) |
1385 | return; | |
1386 | ||
4793241b PZ |
1387 | /* |
1388 | * Only set the backward buddy when the current task is still on the | |
1389 | * rq. This can happen when a wakeup gets interleaved with schedule on | |
1390 | * the ->pre_schedule() or idle_balance() point, either of which can | |
1391 | * drop the rq lock. | |
1392 | * | |
1393 | * Also, during early boot the idle thread is in the fair class, for | |
1394 | * obvious reasons its a bad idea to schedule back to the idle thread. | |
1395 | */ | |
1396 | if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle)) | |
02479099 PZ |
1397 | set_last_buddy(se); |
1398 | set_next_buddy(pse); | |
57fdc26d | 1399 | |
aec0a514 BR |
1400 | /* |
1401 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1402 | * wake up path. | |
1403 | */ | |
1404 | if (test_tsk_need_resched(curr)) | |
1405 | return; | |
1406 | ||
91c234b4 | 1407 | /* |
6bc912b7 | 1408 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1409 | * the tick): |
1410 | */ | |
6bc912b7 | 1411 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1412 | return; |
bf0f6f24 | 1413 | |
6bc912b7 PZ |
1414 | /* Idle tasks are by definition preempted by everybody. */ |
1415 | if (unlikely(curr->policy == SCHED_IDLE)) { | |
1416 | resched_task(curr); | |
1417 | return; | |
1418 | } | |
1419 | ||
77d9cc44 IM |
1420 | if (!sched_feat(WAKEUP_PREEMPT)) |
1421 | return; | |
8651a86c | 1422 | |
d942fb6c | 1423 | if (sched_feat(WAKEUP_OVERLAP) && sync) { |
15afe09b PZ |
1424 | resched_task(curr); |
1425 | return; | |
1426 | } | |
1427 | ||
464b7527 PZ |
1428 | find_matching_se(&se, &pse); |
1429 | ||
1430 | while (se) { | |
1431 | BUG_ON(!pse); | |
1432 | ||
1433 | if (wakeup_preempt_entity(se, pse) == 1) { | |
1434 | resched_task(curr); | |
1435 | break; | |
1436 | } | |
1437 | ||
1438 | se = parent_entity(se); | |
1439 | pse = parent_entity(pse); | |
1440 | } | |
bf0f6f24 IM |
1441 | } |
1442 | ||
fb8d4724 | 1443 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1444 | { |
8f4d37ec | 1445 | struct task_struct *p; |
bf0f6f24 IM |
1446 | struct cfs_rq *cfs_rq = &rq->cfs; |
1447 | struct sched_entity *se; | |
1448 | ||
1449 | if (unlikely(!cfs_rq->nr_running)) | |
1450 | return NULL; | |
1451 | ||
1452 | do { | |
9948f4b2 | 1453 | se = pick_next_entity(cfs_rq); |
a9f3e2b5 MG |
1454 | /* |
1455 | * If se was a buddy, clear it so that it will have to earn | |
1456 | * the favour again. | |
1457 | */ | |
1458 | clear_buddies(cfs_rq, se); | |
f4b6755f | 1459 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1460 | cfs_rq = group_cfs_rq(se); |
1461 | } while (cfs_rq); | |
1462 | ||
8f4d37ec PZ |
1463 | p = task_of(se); |
1464 | hrtick_start_fair(rq, p); | |
1465 | ||
1466 | return p; | |
bf0f6f24 IM |
1467 | } |
1468 | ||
1469 | /* | |
1470 | * Account for a descheduled task: | |
1471 | */ | |
31ee529c | 1472 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1473 | { |
1474 | struct sched_entity *se = &prev->se; | |
1475 | struct cfs_rq *cfs_rq; | |
1476 | ||
1477 | for_each_sched_entity(se) { | |
1478 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1479 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1480 | } |
1481 | } | |
1482 | ||
681f3e68 | 1483 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1484 | /************************************************** |
1485 | * Fair scheduling class load-balancing methods: | |
1486 | */ | |
1487 | ||
1488 | /* | |
1489 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1490 | * during the whole iteration, the current task might be | |
1491 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1492 | * achieve that by always pre-iterating before returning | |
1493 | * the current task: | |
1494 | */ | |
a9957449 | 1495 | static struct task_struct * |
4a55bd5e | 1496 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) |
bf0f6f24 | 1497 | { |
354d60c2 DG |
1498 | struct task_struct *p = NULL; |
1499 | struct sched_entity *se; | |
bf0f6f24 | 1500 | |
77ae6513 MG |
1501 | if (next == &cfs_rq->tasks) |
1502 | return NULL; | |
1503 | ||
b87f1724 BR |
1504 | se = list_entry(next, struct sched_entity, group_node); |
1505 | p = task_of(se); | |
1506 | cfs_rq->balance_iterator = next->next; | |
77ae6513 | 1507 | |
bf0f6f24 IM |
1508 | return p; |
1509 | } | |
1510 | ||
1511 | static struct task_struct *load_balance_start_fair(void *arg) | |
1512 | { | |
1513 | struct cfs_rq *cfs_rq = arg; | |
1514 | ||
4a55bd5e | 1515 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); |
bf0f6f24 IM |
1516 | } |
1517 | ||
1518 | static struct task_struct *load_balance_next_fair(void *arg) | |
1519 | { | |
1520 | struct cfs_rq *cfs_rq = arg; | |
1521 | ||
4a55bd5e | 1522 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); |
bf0f6f24 IM |
1523 | } |
1524 | ||
c09595f6 PZ |
1525 | static unsigned long |
1526 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1527 | unsigned long max_load_move, struct sched_domain *sd, | |
1528 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, | |
1529 | struct cfs_rq *cfs_rq) | |
62fb1851 | 1530 | { |
c09595f6 | 1531 | struct rq_iterator cfs_rq_iterator; |
62fb1851 | 1532 | |
c09595f6 PZ |
1533 | cfs_rq_iterator.start = load_balance_start_fair; |
1534 | cfs_rq_iterator.next = load_balance_next_fair; | |
1535 | cfs_rq_iterator.arg = cfs_rq; | |
62fb1851 | 1536 | |
c09595f6 PZ |
1537 | return balance_tasks(this_rq, this_cpu, busiest, |
1538 | max_load_move, sd, idle, all_pinned, | |
1539 | this_best_prio, &cfs_rq_iterator); | |
62fb1851 | 1540 | } |
62fb1851 | 1541 | |
c09595f6 | 1542 | #ifdef CONFIG_FAIR_GROUP_SCHED |
43010659 | 1543 | static unsigned long |
bf0f6f24 | 1544 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1545 | unsigned long max_load_move, |
a4ac01c3 PW |
1546 | struct sched_domain *sd, enum cpu_idle_type idle, |
1547 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 | 1548 | { |
bf0f6f24 | 1549 | long rem_load_move = max_load_move; |
c09595f6 PZ |
1550 | int busiest_cpu = cpu_of(busiest); |
1551 | struct task_group *tg; | |
18d95a28 | 1552 | |
c09595f6 | 1553 | rcu_read_lock(); |
c8cba857 | 1554 | update_h_load(busiest_cpu); |
18d95a28 | 1555 | |
caea8a03 | 1556 | list_for_each_entry_rcu(tg, &task_groups, list) { |
c8cba857 | 1557 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; |
42a3ac7d PZ |
1558 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
1559 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
243e0e7b | 1560 | u64 rem_load, moved_load; |
18d95a28 | 1561 | |
c09595f6 PZ |
1562 | /* |
1563 | * empty group | |
1564 | */ | |
c8cba857 | 1565 | if (!busiest_cfs_rq->task_weight) |
bf0f6f24 IM |
1566 | continue; |
1567 | ||
243e0e7b SV |
1568 | rem_load = (u64)rem_load_move * busiest_weight; |
1569 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
bf0f6f24 | 1570 | |
c09595f6 | 1571 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, |
53fecd8a | 1572 | rem_load, sd, idle, all_pinned, this_best_prio, |
c09595f6 | 1573 | tg->cfs_rq[busiest_cpu]); |
bf0f6f24 | 1574 | |
c09595f6 | 1575 | if (!moved_load) |
bf0f6f24 IM |
1576 | continue; |
1577 | ||
42a3ac7d | 1578 | moved_load *= busiest_h_load; |
243e0e7b | 1579 | moved_load = div_u64(moved_load, busiest_weight + 1); |
bf0f6f24 | 1580 | |
c09595f6 PZ |
1581 | rem_load_move -= moved_load; |
1582 | if (rem_load_move < 0) | |
bf0f6f24 IM |
1583 | break; |
1584 | } | |
c09595f6 | 1585 | rcu_read_unlock(); |
bf0f6f24 | 1586 | |
43010659 | 1587 | return max_load_move - rem_load_move; |
bf0f6f24 | 1588 | } |
c09595f6 PZ |
1589 | #else |
1590 | static unsigned long | |
1591 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1592 | unsigned long max_load_move, | |
1593 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1594 | int *all_pinned, int *this_best_prio) | |
1595 | { | |
1596 | return __load_balance_fair(this_rq, this_cpu, busiest, | |
1597 | max_load_move, sd, idle, all_pinned, | |
1598 | this_best_prio, &busiest->cfs); | |
1599 | } | |
1600 | #endif | |
bf0f6f24 | 1601 | |
e1d1484f PW |
1602 | static int |
1603 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1604 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1605 | { | |
1606 | struct cfs_rq *busy_cfs_rq; | |
1607 | struct rq_iterator cfs_rq_iterator; | |
1608 | ||
1609 | cfs_rq_iterator.start = load_balance_start_fair; | |
1610 | cfs_rq_iterator.next = load_balance_next_fair; | |
1611 | ||
1612 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1613 | /* | |
1614 | * pass busy_cfs_rq argument into | |
1615 | * load_balance_[start|next]_fair iterators | |
1616 | */ | |
1617 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1618 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1619 | &cfs_rq_iterator)) | |
1620 | return 1; | |
1621 | } | |
1622 | ||
1623 | return 0; | |
1624 | } | |
55e12e5e | 1625 | #endif /* CONFIG_SMP */ |
e1d1484f | 1626 | |
bf0f6f24 IM |
1627 | /* |
1628 | * scheduler tick hitting a task of our scheduling class: | |
1629 | */ | |
8f4d37ec | 1630 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1631 | { |
1632 | struct cfs_rq *cfs_rq; | |
1633 | struct sched_entity *se = &curr->se; | |
1634 | ||
1635 | for_each_sched_entity(se) { | |
1636 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1637 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1638 | } |
1639 | } | |
1640 | ||
1641 | /* | |
1642 | * Share the fairness runtime between parent and child, thus the | |
1643 | * total amount of pressure for CPU stays equal - new tasks | |
1644 | * get a chance to run but frequent forkers are not allowed to | |
1645 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1646 | * the child is not running yet. | |
1647 | */ | |
ee0827d8 | 1648 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1649 | { |
1650 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1651 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1652 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1653 | |
1654 | sched_info_queued(p); | |
1655 | ||
7109c442 | 1656 | update_curr(cfs_rq); |
aeb73b04 | 1657 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1658 | |
3c90e6e9 | 1659 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1660 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1661 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1662 | /* |
edcb60a3 IM |
1663 | * Upon rescheduling, sched_class::put_prev_task() will place |
1664 | * 'current' within the tree based on its new key value. | |
1665 | */ | |
4d78e7b6 | 1666 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 1667 | resched_task(rq->curr); |
4d78e7b6 | 1668 | } |
bf0f6f24 | 1669 | |
b9dca1e0 | 1670 | enqueue_task_fair(rq, p, 0); |
bf0f6f24 IM |
1671 | } |
1672 | ||
cb469845 SR |
1673 | /* |
1674 | * Priority of the task has changed. Check to see if we preempt | |
1675 | * the current task. | |
1676 | */ | |
1677 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1678 | int oldprio, int running) | |
1679 | { | |
1680 | /* | |
1681 | * Reschedule if we are currently running on this runqueue and | |
1682 | * our priority decreased, or if we are not currently running on | |
1683 | * this runqueue and our priority is higher than the current's | |
1684 | */ | |
1685 | if (running) { | |
1686 | if (p->prio > oldprio) | |
1687 | resched_task(rq->curr); | |
1688 | } else | |
15afe09b | 1689 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1690 | } |
1691 | ||
1692 | /* | |
1693 | * We switched to the sched_fair class. | |
1694 | */ | |
1695 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1696 | int running) | |
1697 | { | |
1698 | /* | |
1699 | * We were most likely switched from sched_rt, so | |
1700 | * kick off the schedule if running, otherwise just see | |
1701 | * if we can still preempt the current task. | |
1702 | */ | |
1703 | if (running) | |
1704 | resched_task(rq->curr); | |
1705 | else | |
15afe09b | 1706 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1707 | } |
1708 | ||
83b699ed SV |
1709 | /* Account for a task changing its policy or group. |
1710 | * | |
1711 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1712 | * migrates between groups/classes. | |
1713 | */ | |
1714 | static void set_curr_task_fair(struct rq *rq) | |
1715 | { | |
1716 | struct sched_entity *se = &rq->curr->se; | |
1717 | ||
1718 | for_each_sched_entity(se) | |
1719 | set_next_entity(cfs_rq_of(se), se); | |
1720 | } | |
1721 | ||
810b3817 PZ |
1722 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1723 | static void moved_group_fair(struct task_struct *p) | |
1724 | { | |
1725 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1726 | ||
1727 | update_curr(cfs_rq); | |
1728 | place_entity(cfs_rq, &p->se, 1); | |
1729 | } | |
1730 | #endif | |
1731 | ||
bf0f6f24 IM |
1732 | /* |
1733 | * All the scheduling class methods: | |
1734 | */ | |
5522d5d5 IM |
1735 | static const struct sched_class fair_sched_class = { |
1736 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1737 | .enqueue_task = enqueue_task_fair, |
1738 | .dequeue_task = dequeue_task_fair, | |
1739 | .yield_task = yield_task_fair, | |
1740 | ||
2e09bf55 | 1741 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1742 | |
1743 | .pick_next_task = pick_next_task_fair, | |
1744 | .put_prev_task = put_prev_task_fair, | |
1745 | ||
681f3e68 | 1746 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
1747 | .select_task_rq = select_task_rq_fair, |
1748 | ||
bf0f6f24 | 1749 | .load_balance = load_balance_fair, |
e1d1484f | 1750 | .move_one_task = move_one_task_fair, |
681f3e68 | 1751 | #endif |
bf0f6f24 | 1752 | |
83b699ed | 1753 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1754 | .task_tick = task_tick_fair, |
1755 | .task_new = task_new_fair, | |
cb469845 SR |
1756 | |
1757 | .prio_changed = prio_changed_fair, | |
1758 | .switched_to = switched_to_fair, | |
810b3817 PZ |
1759 | |
1760 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1761 | .moved_group = moved_group_fair, | |
1762 | #endif | |
bf0f6f24 IM |
1763 | }; |
1764 | ||
1765 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1766 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1767 | { |
bf0f6f24 IM |
1768 | struct cfs_rq *cfs_rq; |
1769 | ||
5973e5b9 | 1770 | rcu_read_lock(); |
c3b64f1e | 1771 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1772 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1773 | rcu_read_unlock(); |
bf0f6f24 IM |
1774 | } |
1775 | #endif |