]>
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 | 432 | struct load_weight *load; |
3104bf03 | 433 | struct load_weight lw; |
6272d68c LM |
434 | |
435 | cfs_rq = cfs_rq_of(se); | |
436 | load = &cfs_rq->load; | |
f9c0b095 | 437 | |
0a582440 | 438 | if (unlikely(!se->on_rq)) { |
3104bf03 | 439 | lw = cfs_rq->load; |
0a582440 MG |
440 | |
441 | update_load_add(&lw, se->load.weight); | |
442 | load = &lw; | |
443 | } | |
444 | slice = calc_delta_mine(slice, se->load.weight, load); | |
445 | } | |
446 | return slice; | |
bf0f6f24 IM |
447 | } |
448 | ||
647e7cac | 449 | /* |
ac884dec | 450 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 451 | * |
f9c0b095 | 452 | * vs = s/w |
647e7cac | 453 | */ |
f9c0b095 | 454 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 455 | { |
f9c0b095 | 456 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
457 | } |
458 | ||
bf0f6f24 IM |
459 | /* |
460 | * Update the current task's runtime statistics. Skip current tasks that | |
461 | * are not in our scheduling class. | |
462 | */ | |
463 | static inline void | |
8ebc91d9 IM |
464 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
465 | unsigned long delta_exec) | |
bf0f6f24 | 466 | { |
bbdba7c0 | 467 | unsigned long delta_exec_weighted; |
bf0f6f24 | 468 | |
8179ca23 | 469 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
470 | |
471 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 472 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 473 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
e9acbff6 | 474 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 475 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
476 | } |
477 | ||
b7cc0896 | 478 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 479 | { |
429d43bc | 480 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 481 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
482 | unsigned long delta_exec; |
483 | ||
484 | if (unlikely(!curr)) | |
485 | return; | |
486 | ||
487 | /* | |
488 | * Get the amount of time the current task was running | |
489 | * since the last time we changed load (this cannot | |
490 | * overflow on 32 bits): | |
491 | */ | |
8ebc91d9 | 492 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
493 | if (!delta_exec) |
494 | return; | |
bf0f6f24 | 495 | |
8ebc91d9 IM |
496 | __update_curr(cfs_rq, curr, delta_exec); |
497 | curr->exec_start = now; | |
d842de87 SV |
498 | |
499 | if (entity_is_task(curr)) { | |
500 | struct task_struct *curtask = task_of(curr); | |
501 | ||
502 | cpuacct_charge(curtask, delta_exec); | |
f06febc9 | 503 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 504 | } |
bf0f6f24 IM |
505 | } |
506 | ||
507 | static inline void | |
5870db5b | 508 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 509 | { |
d281918d | 510 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
511 | } |
512 | ||
bf0f6f24 IM |
513 | /* |
514 | * Task is being enqueued - update stats: | |
515 | */ | |
d2417e5a | 516 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 517 | { |
bf0f6f24 IM |
518 | /* |
519 | * Are we enqueueing a waiting task? (for current tasks | |
520 | * a dequeue/enqueue event is a NOP) | |
521 | */ | |
429d43bc | 522 | if (se != cfs_rq->curr) |
5870db5b | 523 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
524 | } |
525 | ||
bf0f6f24 | 526 | static void |
9ef0a961 | 527 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 528 | { |
bbdba7c0 IM |
529 | schedstat_set(se->wait_max, max(se->wait_max, |
530 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
531 | schedstat_set(se->wait_count, se->wait_count + 1); |
532 | schedstat_set(se->wait_sum, se->wait_sum + | |
533 | rq_of(cfs_rq)->clock - se->wait_start); | |
6cfb0d5d | 534 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
535 | } |
536 | ||
537 | static inline void | |
19b6a2e3 | 538 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 539 | { |
bf0f6f24 IM |
540 | /* |
541 | * Mark the end of the wait period if dequeueing a | |
542 | * waiting task: | |
543 | */ | |
429d43bc | 544 | if (se != cfs_rq->curr) |
9ef0a961 | 545 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
546 | } |
547 | ||
548 | /* | |
549 | * We are picking a new current task - update its stats: | |
550 | */ | |
551 | static inline void | |
79303e9e | 552 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
553 | { |
554 | /* | |
555 | * We are starting a new run period: | |
556 | */ | |
d281918d | 557 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
558 | } |
559 | ||
bf0f6f24 IM |
560 | /************************************************** |
561 | * Scheduling class queueing methods: | |
562 | */ | |
563 | ||
c09595f6 PZ |
564 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
565 | static void | |
566 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
567 | { | |
568 | cfs_rq->task_weight += weight; | |
569 | } | |
570 | #else | |
571 | static inline void | |
572 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
573 | { | |
574 | } | |
575 | #endif | |
576 | ||
30cfdcfc DA |
577 | static void |
578 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
579 | { | |
580 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
581 | if (!parent_entity(se)) |
582 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 583 | if (entity_is_task(se)) { |
c09595f6 | 584 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
585 | list_add(&se->group_node, &cfs_rq->tasks); |
586 | } | |
30cfdcfc DA |
587 | cfs_rq->nr_running++; |
588 | se->on_rq = 1; | |
589 | } | |
590 | ||
591 | static void | |
592 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
593 | { | |
594 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
595 | if (!parent_entity(se)) |
596 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 597 | if (entity_is_task(se)) { |
c09595f6 | 598 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
599 | list_del_init(&se->group_node); |
600 | } | |
30cfdcfc DA |
601 | cfs_rq->nr_running--; |
602 | se->on_rq = 0; | |
603 | } | |
604 | ||
2396af69 | 605 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 606 | { |
bf0f6f24 IM |
607 | #ifdef CONFIG_SCHEDSTATS |
608 | if (se->sleep_start) { | |
d281918d | 609 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 610 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
611 | |
612 | if ((s64)delta < 0) | |
613 | delta = 0; | |
614 | ||
615 | if (unlikely(delta > se->sleep_max)) | |
616 | se->sleep_max = delta; | |
617 | ||
618 | se->sleep_start = 0; | |
619 | se->sum_sleep_runtime += delta; | |
9745512c AV |
620 | |
621 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
622 | } |
623 | if (se->block_start) { | |
d281918d | 624 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 625 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
626 | |
627 | if ((s64)delta < 0) | |
628 | delta = 0; | |
629 | ||
630 | if (unlikely(delta > se->block_max)) | |
631 | se->block_max = delta; | |
632 | ||
633 | se->block_start = 0; | |
634 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
635 | |
636 | /* | |
637 | * Blocking time is in units of nanosecs, so shift by 20 to | |
638 | * get a milliseconds-range estimation of the amount of | |
639 | * time that the task spent sleeping: | |
640 | */ | |
641 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 642 | |
30084fbd IM |
643 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
644 | delta >> 20); | |
645 | } | |
9745512c | 646 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
647 | } |
648 | #endif | |
649 | } | |
650 | ||
ddc97297 PZ |
651 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
652 | { | |
653 | #ifdef CONFIG_SCHED_DEBUG | |
654 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
655 | ||
656 | if (d < 0) | |
657 | d = -d; | |
658 | ||
659 | if (d > 3*sysctl_sched_latency) | |
660 | schedstat_inc(cfs_rq, nr_spread_over); | |
661 | #endif | |
662 | } | |
663 | ||
aeb73b04 PZ |
664 | static void |
665 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
666 | { | |
1af5f730 | 667 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 668 | |
2cb8600e PZ |
669 | /* |
670 | * The 'current' period is already promised to the current tasks, | |
671 | * however the extra weight of the new task will slow them down a | |
672 | * little, place the new task so that it fits in the slot that | |
673 | * stays open at the end. | |
674 | */ | |
94dfb5e7 | 675 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 676 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 677 | |
8465e792 | 678 | if (!initial) { |
2cb8600e | 679 | /* sleeps upto a single latency don't count. */ |
a7be37ac PZ |
680 | if (sched_feat(NEW_FAIR_SLEEPERS)) { |
681 | unsigned long thresh = sysctl_sched_latency; | |
682 | ||
683 | /* | |
6bc912b7 PZ |
684 | * Convert the sleeper threshold into virtual time. |
685 | * SCHED_IDLE is a special sub-class. We care about | |
686 | * fairness only relative to other SCHED_IDLE tasks, | |
687 | * all of which have the same weight. | |
a7be37ac | 688 | */ |
6bc912b7 PZ |
689 | if (sched_feat(NORMALIZED_SLEEPER) && |
690 | task_of(se)->policy != SCHED_IDLE) | |
a7be37ac PZ |
691 | thresh = calc_delta_fair(thresh, se); |
692 | ||
693 | vruntime -= thresh; | |
694 | } | |
94359f05 | 695 | |
2cb8600e PZ |
696 | /* ensure we never gain time by being placed backwards. */ |
697 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
698 | } |
699 | ||
67e9fb2a | 700 | se->vruntime = vruntime; |
aeb73b04 PZ |
701 | } |
702 | ||
bf0f6f24 | 703 | static void |
83b699ed | 704 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
705 | { |
706 | /* | |
a2a2d680 | 707 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 708 | */ |
b7cc0896 | 709 | update_curr(cfs_rq); |
a992241d | 710 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 711 | |
e9acbff6 | 712 | if (wakeup) { |
aeb73b04 | 713 | place_entity(cfs_rq, se, 0); |
2396af69 | 714 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 715 | } |
bf0f6f24 | 716 | |
d2417e5a | 717 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 718 | check_spread(cfs_rq, se); |
83b699ed SV |
719 | if (se != cfs_rq->curr) |
720 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
721 | } |
722 | ||
a571bbea | 723 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 PZ |
724 | { |
725 | if (cfs_rq->last == se) | |
726 | cfs_rq->last = NULL; | |
727 | ||
728 | if (cfs_rq->next == se) | |
729 | cfs_rq->next = NULL; | |
730 | } | |
731 | ||
a571bbea PZ |
732 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
733 | { | |
734 | for_each_sched_entity(se) | |
735 | __clear_buddies(cfs_rq_of(se), se); | |
736 | } | |
737 | ||
bf0f6f24 | 738 | static void |
525c2716 | 739 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 740 | { |
a2a2d680 DA |
741 | /* |
742 | * Update run-time statistics of the 'current'. | |
743 | */ | |
744 | update_curr(cfs_rq); | |
745 | ||
19b6a2e3 | 746 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 747 | if (sleep) { |
67e9fb2a | 748 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
749 | if (entity_is_task(se)) { |
750 | struct task_struct *tsk = task_of(se); | |
751 | ||
752 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 753 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 754 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 755 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 756 | } |
db36cc7d | 757 | #endif |
67e9fb2a PZ |
758 | } |
759 | ||
2002c695 | 760 | clear_buddies(cfs_rq, se); |
4793241b | 761 | |
83b699ed | 762 | if (se != cfs_rq->curr) |
30cfdcfc DA |
763 | __dequeue_entity(cfs_rq, se); |
764 | account_entity_dequeue(cfs_rq, se); | |
1af5f730 | 765 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
766 | } |
767 | ||
768 | /* | |
769 | * Preempt the current task with a newly woken task if needed: | |
770 | */ | |
7c92e54f | 771 | static void |
2e09bf55 | 772 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 773 | { |
11697830 PZ |
774 | unsigned long ideal_runtime, delta_exec; |
775 | ||
6d0f0ebd | 776 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 777 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 778 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 779 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
780 | /* |
781 | * The current task ran long enough, ensure it doesn't get | |
782 | * re-elected due to buddy favours. | |
783 | */ | |
784 | clear_buddies(cfs_rq, curr); | |
785 | } | |
bf0f6f24 IM |
786 | } |
787 | ||
83b699ed | 788 | static void |
8494f412 | 789 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 790 | { |
83b699ed SV |
791 | /* 'current' is not kept within the tree. */ |
792 | if (se->on_rq) { | |
793 | /* | |
794 | * Any task has to be enqueued before it get to execute on | |
795 | * a CPU. So account for the time it spent waiting on the | |
796 | * runqueue. | |
797 | */ | |
798 | update_stats_wait_end(cfs_rq, se); | |
799 | __dequeue_entity(cfs_rq, se); | |
800 | } | |
801 | ||
79303e9e | 802 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 803 | cfs_rq->curr = se; |
eba1ed4b IM |
804 | #ifdef CONFIG_SCHEDSTATS |
805 | /* | |
806 | * Track our maximum slice length, if the CPU's load is at | |
807 | * least twice that of our own weight (i.e. dont track it | |
808 | * when there are only lesser-weight tasks around): | |
809 | */ | |
495eca49 | 810 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
811 | se->slice_max = max(se->slice_max, |
812 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
813 | } | |
814 | #endif | |
4a55b450 | 815 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
816 | } |
817 | ||
3f3a4904 PZ |
818 | static int |
819 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
820 | ||
f4b6755f | 821 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 822 | { |
f4b6755f PZ |
823 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
824 | ||
4793241b PZ |
825 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1) |
826 | return cfs_rq->next; | |
aa2ac252 | 827 | |
4793241b PZ |
828 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1) |
829 | return cfs_rq->last; | |
830 | ||
831 | return se; | |
aa2ac252 PZ |
832 | } |
833 | ||
ab6cde26 | 834 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
835 | { |
836 | /* | |
837 | * If still on the runqueue then deactivate_task() | |
838 | * was not called and update_curr() has to be done: | |
839 | */ | |
840 | if (prev->on_rq) | |
b7cc0896 | 841 | update_curr(cfs_rq); |
bf0f6f24 | 842 | |
ddc97297 | 843 | check_spread(cfs_rq, prev); |
30cfdcfc | 844 | if (prev->on_rq) { |
5870db5b | 845 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
846 | /* Put 'current' back into the tree. */ |
847 | __enqueue_entity(cfs_rq, prev); | |
848 | } | |
429d43bc | 849 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
850 | } |
851 | ||
8f4d37ec PZ |
852 | static void |
853 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 854 | { |
bf0f6f24 | 855 | /* |
30cfdcfc | 856 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 857 | */ |
30cfdcfc | 858 | update_curr(cfs_rq); |
bf0f6f24 | 859 | |
8f4d37ec PZ |
860 | #ifdef CONFIG_SCHED_HRTICK |
861 | /* | |
862 | * queued ticks are scheduled to match the slice, so don't bother | |
863 | * validating it and just reschedule. | |
864 | */ | |
983ed7a6 HH |
865 | if (queued) { |
866 | resched_task(rq_of(cfs_rq)->curr); | |
867 | return; | |
868 | } | |
8f4d37ec PZ |
869 | /* |
870 | * don't let the period tick interfere with the hrtick preemption | |
871 | */ | |
872 | if (!sched_feat(DOUBLE_TICK) && | |
873 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
874 | return; | |
875 | #endif | |
876 | ||
ce6c1311 | 877 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 878 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
879 | } |
880 | ||
881 | /************************************************** | |
882 | * CFS operations on tasks: | |
883 | */ | |
884 | ||
8f4d37ec PZ |
885 | #ifdef CONFIG_SCHED_HRTICK |
886 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
887 | { | |
8f4d37ec PZ |
888 | struct sched_entity *se = &p->se; |
889 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
890 | ||
891 | WARN_ON(task_rq(p) != rq); | |
892 | ||
893 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
894 | u64 slice = sched_slice(cfs_rq, se); | |
895 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
896 | s64 delta = slice - ran; | |
897 | ||
898 | if (delta < 0) { | |
899 | if (rq->curr == p) | |
900 | resched_task(p); | |
901 | return; | |
902 | } | |
903 | ||
904 | /* | |
905 | * Don't schedule slices shorter than 10000ns, that just | |
906 | * doesn't make sense. Rely on vruntime for fairness. | |
907 | */ | |
31656519 | 908 | if (rq->curr != p) |
157124c1 | 909 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 910 | |
31656519 | 911 | hrtick_start(rq, delta); |
8f4d37ec PZ |
912 | } |
913 | } | |
a4c2f00f PZ |
914 | |
915 | /* | |
916 | * called from enqueue/dequeue and updates the hrtick when the | |
917 | * current task is from our class and nr_running is low enough | |
918 | * to matter. | |
919 | */ | |
920 | static void hrtick_update(struct rq *rq) | |
921 | { | |
922 | struct task_struct *curr = rq->curr; | |
923 | ||
924 | if (curr->sched_class != &fair_sched_class) | |
925 | return; | |
926 | ||
927 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
928 | hrtick_start_fair(rq, curr); | |
929 | } | |
55e12e5e | 930 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
931 | static inline void |
932 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
933 | { | |
934 | } | |
a4c2f00f PZ |
935 | |
936 | static inline void hrtick_update(struct rq *rq) | |
937 | { | |
938 | } | |
8f4d37ec PZ |
939 | #endif |
940 | ||
bf0f6f24 IM |
941 | /* |
942 | * The enqueue_task method is called before nr_running is | |
943 | * increased. Here we update the fair scheduling stats and | |
944 | * then put the task into the rbtree: | |
945 | */ | |
fd390f6a | 946 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
947 | { |
948 | struct cfs_rq *cfs_rq; | |
62fb1851 | 949 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
950 | |
951 | for_each_sched_entity(se) { | |
62fb1851 | 952 | if (se->on_rq) |
bf0f6f24 IM |
953 | break; |
954 | cfs_rq = cfs_rq_of(se); | |
83b699ed | 955 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 956 | wakeup = 1; |
bf0f6f24 | 957 | } |
8f4d37ec | 958 | |
a4c2f00f | 959 | hrtick_update(rq); |
bf0f6f24 IM |
960 | } |
961 | ||
962 | /* | |
963 | * The dequeue_task method is called before nr_running is | |
964 | * decreased. We remove the task from the rbtree and | |
965 | * update the fair scheduling stats: | |
966 | */ | |
f02231e5 | 967 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
968 | { |
969 | struct cfs_rq *cfs_rq; | |
62fb1851 | 970 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
971 | |
972 | for_each_sched_entity(se) { | |
973 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 974 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 975 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 976 | if (cfs_rq->load.weight) |
bf0f6f24 | 977 | break; |
b9fa3df3 | 978 | sleep = 1; |
bf0f6f24 | 979 | } |
8f4d37ec | 980 | |
a4c2f00f | 981 | hrtick_update(rq); |
bf0f6f24 IM |
982 | } |
983 | ||
984 | /* | |
1799e35d IM |
985 | * sched_yield() support is very simple - we dequeue and enqueue. |
986 | * | |
987 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 988 | */ |
4530d7ab | 989 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 990 | { |
db292ca3 IM |
991 | struct task_struct *curr = rq->curr; |
992 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
993 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
994 | |
995 | /* | |
1799e35d IM |
996 | * Are we the only task in the tree? |
997 | */ | |
998 | if (unlikely(cfs_rq->nr_running == 1)) | |
999 | return; | |
1000 | ||
2002c695 PZ |
1001 | clear_buddies(cfs_rq, se); |
1002 | ||
db292ca3 | 1003 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1004 | update_rq_clock(rq); |
1799e35d | 1005 | /* |
a2a2d680 | 1006 | * Update run-time statistics of the 'current'. |
1799e35d | 1007 | */ |
2b1e315d | 1008 | update_curr(cfs_rq); |
1799e35d IM |
1009 | |
1010 | return; | |
1011 | } | |
1012 | /* | |
1013 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1014 | */ |
2b1e315d | 1015 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1016 | /* |
1017 | * Already in the rightmost position? | |
1018 | */ | |
79b3feff | 1019 | if (unlikely(!rightmost || rightmost->vruntime < se->vruntime)) |
1799e35d IM |
1020 | return; |
1021 | ||
1022 | /* | |
1023 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1024 | * Upon rescheduling, sched_class::put_prev_task() will place |
1025 | * 'current' within the tree based on its new key value. | |
1799e35d | 1026 | */ |
30cfdcfc | 1027 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1028 | } |
1029 | ||
e7693a36 GH |
1030 | /* |
1031 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1032 | * not idle and an idle cpu is available. The span of cpus to | |
1033 | * search starts with cpus closest then further out as needed, | |
1034 | * so we always favor a closer, idle cpu. | |
e761b772 | 1035 | * Domains may include CPUs that are not usable for migration, |
96f874e2 | 1036 | * hence we need to mask them out (cpu_active_mask) |
e7693a36 GH |
1037 | * |
1038 | * Returns the CPU we should wake onto. | |
1039 | */ | |
1040 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1041 | static int wake_idle(int cpu, struct task_struct *p) | |
1042 | { | |
e7693a36 GH |
1043 | struct sched_domain *sd; |
1044 | int i; | |
7eb52dfa VS |
1045 | unsigned int chosen_wakeup_cpu; |
1046 | int this_cpu; | |
1047 | ||
1048 | /* | |
1049 | * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu | |
1050 | * are idle and this is not a kernel thread and this task's affinity | |
1051 | * allows it to be moved to preferred cpu, then just move! | |
1052 | */ | |
1053 | ||
1054 | this_cpu = smp_processor_id(); | |
1055 | chosen_wakeup_cpu = | |
1056 | cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu; | |
1057 | ||
1058 | if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP && | |
1059 | idle_cpu(cpu) && idle_cpu(this_cpu) && | |
1060 | p->mm && !(p->flags & PF_KTHREAD) && | |
1061 | cpu_isset(chosen_wakeup_cpu, p->cpus_allowed)) | |
1062 | return chosen_wakeup_cpu; | |
e7693a36 GH |
1063 | |
1064 | /* | |
1065 | * If it is idle, then it is the best cpu to run this task. | |
1066 | * | |
1067 | * This cpu is also the best, if it has more than one task already. | |
1068 | * Siblings must be also busy(in most cases) as they didn't already | |
1069 | * pickup the extra load from this cpu and hence we need not check | |
1070 | * sibling runqueue info. This will avoid the checks and cache miss | |
1071 | * penalities associated with that. | |
1072 | */ | |
104f6454 | 1073 | if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1) |
e7693a36 GH |
1074 | return cpu; |
1075 | ||
1076 | for_each_domain(cpu, sd) { | |
1d3504fc HS |
1077 | if ((sd->flags & SD_WAKE_IDLE) |
1078 | || ((sd->flags & SD_WAKE_IDLE_FAR) | |
1079 | && !task_hot(p, task_rq(p)->clock, sd))) { | |
758b2cdc RR |
1080 | for_each_cpu_and(i, sched_domain_span(sd), |
1081 | &p->cpus_allowed) { | |
1082 | if (cpu_active(i) && idle_cpu(i)) { | |
e7693a36 GH |
1083 | if (i != task_cpu(p)) { |
1084 | schedstat_inc(p, | |
1085 | se.nr_wakeups_idle); | |
1086 | } | |
1087 | return i; | |
1088 | } | |
1089 | } | |
1090 | } else { | |
1091 | break; | |
1092 | } | |
1093 | } | |
1094 | return cpu; | |
1095 | } | |
55e12e5e | 1096 | #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/ |
e7693a36 GH |
1097 | static inline int wake_idle(int cpu, struct task_struct *p) |
1098 | { | |
1099 | return cpu; | |
1100 | } | |
1101 | #endif | |
1102 | ||
1103 | #ifdef CONFIG_SMP | |
098fb9db | 1104 | |
bb3469ac | 1105 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1106 | /* |
1107 | * effective_load() calculates the load change as seen from the root_task_group | |
1108 | * | |
1109 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1110 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1111 | * can calculate the shift in shares. | |
1112 | * | |
1113 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1114 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1115 | * this change. | |
1116 | * | |
1117 | * We compensate this by not only taking the current delta into account, but | |
1118 | * also considering the delta between when the shares were last adjusted and | |
1119 | * now. | |
1120 | * | |
1121 | * We still saw a performance dip, some tracing learned us that between | |
1122 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1123 | * significantly. Therefore try to bias the error in direction of failing | |
1124 | * the affine wakeup. | |
1125 | * | |
1126 | */ | |
f1d239f7 PZ |
1127 | static long effective_load(struct task_group *tg, int cpu, |
1128 | long wl, long wg) | |
bb3469ac | 1129 | { |
4be9daaa | 1130 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1131 | |
1132 | if (!tg->parent) | |
1133 | return wl; | |
1134 | ||
f5bfb7d9 PZ |
1135 | /* |
1136 | * By not taking the decrease of shares on the other cpu into | |
1137 | * account our error leans towards reducing the affine wakeups. | |
1138 | */ | |
1139 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1140 | return wl; | |
1141 | ||
4be9daaa | 1142 | for_each_sched_entity(se) { |
cb5ef42a | 1143 | long S, rw, s, a, b; |
940959e9 PZ |
1144 | long more_w; |
1145 | ||
1146 | /* | |
1147 | * Instead of using this increment, also add the difference | |
1148 | * between when the shares were last updated and now. | |
1149 | */ | |
1150 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1151 | wl += more_w; | |
1152 | wg += more_w; | |
4be9daaa PZ |
1153 | |
1154 | S = se->my_q->tg->shares; | |
1155 | s = se->my_q->shares; | |
f1d239f7 | 1156 | rw = se->my_q->rq_weight; |
bb3469ac | 1157 | |
cb5ef42a PZ |
1158 | a = S*(rw + wl); |
1159 | b = S*rw + s*wg; | |
4be9daaa | 1160 | |
940959e9 PZ |
1161 | wl = s*(a-b); |
1162 | ||
1163 | if (likely(b)) | |
1164 | wl /= b; | |
1165 | ||
83378269 PZ |
1166 | /* |
1167 | * Assume the group is already running and will | |
1168 | * thus already be accounted for in the weight. | |
1169 | * | |
1170 | * That is, moving shares between CPUs, does not | |
1171 | * alter the group weight. | |
1172 | */ | |
4be9daaa | 1173 | wg = 0; |
4be9daaa | 1174 | } |
bb3469ac | 1175 | |
4be9daaa | 1176 | return wl; |
bb3469ac | 1177 | } |
4be9daaa | 1178 | |
bb3469ac | 1179 | #else |
4be9daaa | 1180 | |
83378269 PZ |
1181 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1182 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1183 | { |
83378269 | 1184 | return wl; |
bb3469ac | 1185 | } |
4be9daaa | 1186 | |
bb3469ac PZ |
1187 | #endif |
1188 | ||
098fb9db | 1189 | static int |
64b9e029 | 1190 | wake_affine(struct sched_domain *this_sd, struct rq *this_rq, |
4ae7d5ce IM |
1191 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, |
1192 | int idx, unsigned long load, unsigned long this_load, | |
098fb9db IM |
1193 | unsigned int imbalance) |
1194 | { | |
fc631c82 PZ |
1195 | struct task_struct *curr = this_rq->curr; |
1196 | struct task_group *tg; | |
098fb9db IM |
1197 | unsigned long tl = this_load; |
1198 | unsigned long tl_per_task; | |
83378269 | 1199 | unsigned long weight; |
b3137bc8 | 1200 | int balanced; |
098fb9db | 1201 | |
b3137bc8 | 1202 | if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS)) |
098fb9db IM |
1203 | return 0; |
1204 | ||
fc631c82 PZ |
1205 | if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost || |
1206 | p->se.avg_overlap > sysctl_sched_migration_cost)) | |
1207 | sync = 0; | |
1208 | ||
b3137bc8 MG |
1209 | /* |
1210 | * If sync wakeup then subtract the (maximum possible) | |
1211 | * effect of the currently running task from the load | |
1212 | * of the current CPU: | |
1213 | */ | |
83378269 PZ |
1214 | if (sync) { |
1215 | tg = task_group(current); | |
1216 | weight = current->se.load.weight; | |
1217 | ||
1218 | tl += effective_load(tg, this_cpu, -weight, -weight); | |
1219 | load += effective_load(tg, prev_cpu, 0, -weight); | |
1220 | } | |
b3137bc8 | 1221 | |
83378269 PZ |
1222 | tg = task_group(p); |
1223 | weight = p->se.load.weight; | |
b3137bc8 | 1224 | |
83378269 PZ |
1225 | balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <= |
1226 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); | |
b3137bc8 | 1227 | |
098fb9db | 1228 | /* |
4ae7d5ce IM |
1229 | * If the currently running task will sleep within |
1230 | * a reasonable amount of time then attract this newly | |
1231 | * woken task: | |
098fb9db | 1232 | */ |
2fb7635c PZ |
1233 | if (sync && balanced) |
1234 | return 1; | |
098fb9db IM |
1235 | |
1236 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1237 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1238 | ||
64b9e029 AA |
1239 | if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <= |
1240 | tl_per_task)) { | |
098fb9db IM |
1241 | /* |
1242 | * This domain has SD_WAKE_AFFINE and | |
1243 | * p is cache cold in this domain, and | |
1244 | * there is no bad imbalance. | |
1245 | */ | |
1246 | schedstat_inc(this_sd, ttwu_move_affine); | |
1247 | schedstat_inc(p, se.nr_wakeups_affine); | |
1248 | ||
1249 | return 1; | |
1250 | } | |
1251 | return 0; | |
1252 | } | |
1253 | ||
e7693a36 GH |
1254 | static int select_task_rq_fair(struct task_struct *p, int sync) |
1255 | { | |
e7693a36 | 1256 | struct sched_domain *sd, *this_sd = NULL; |
ac192d39 | 1257 | int prev_cpu, this_cpu, new_cpu; |
098fb9db | 1258 | unsigned long load, this_load; |
64b9e029 | 1259 | struct rq *this_rq; |
098fb9db | 1260 | unsigned int imbalance; |
098fb9db | 1261 | int idx; |
e7693a36 | 1262 | |
ac192d39 | 1263 | prev_cpu = task_cpu(p); |
ac192d39 | 1264 | this_cpu = smp_processor_id(); |
4ae7d5ce | 1265 | this_rq = cpu_rq(this_cpu); |
ac192d39 | 1266 | new_cpu = prev_cpu; |
e7693a36 | 1267 | |
64b9e029 AA |
1268 | if (prev_cpu == this_cpu) |
1269 | goto out; | |
ac192d39 IM |
1270 | /* |
1271 | * 'this_sd' is the first domain that both | |
1272 | * this_cpu and prev_cpu are present in: | |
1273 | */ | |
e7693a36 | 1274 | for_each_domain(this_cpu, sd) { |
758b2cdc | 1275 | if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) { |
e7693a36 GH |
1276 | this_sd = sd; |
1277 | break; | |
1278 | } | |
1279 | } | |
1280 | ||
96f874e2 | 1281 | if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed))) |
f4827386 | 1282 | goto out; |
e7693a36 GH |
1283 | |
1284 | /* | |
1285 | * Check for affine wakeup and passive balancing possibilities. | |
1286 | */ | |
098fb9db | 1287 | if (!this_sd) |
f4827386 | 1288 | goto out; |
e7693a36 | 1289 | |
098fb9db IM |
1290 | idx = this_sd->wake_idx; |
1291 | ||
1292 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1293 | ||
ac192d39 | 1294 | load = source_load(prev_cpu, idx); |
098fb9db IM |
1295 | this_load = target_load(this_cpu, idx); |
1296 | ||
64b9e029 | 1297 | if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, |
4ae7d5ce IM |
1298 | load, this_load, imbalance)) |
1299 | return this_cpu; | |
1300 | ||
098fb9db IM |
1301 | /* |
1302 | * Start passive balancing when half the imbalance_pct | |
1303 | * limit is reached. | |
1304 | */ | |
1305 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1306 | if (imbalance*this_load <= 100*load) { | |
1307 | schedstat_inc(this_sd, ttwu_move_balance); | |
1308 | schedstat_inc(p, se.nr_wakeups_passive); | |
4ae7d5ce | 1309 | return this_cpu; |
e7693a36 GH |
1310 | } |
1311 | } | |
1312 | ||
f4827386 | 1313 | out: |
e7693a36 GH |
1314 | return wake_idle(new_cpu, p); |
1315 | } | |
1316 | #endif /* CONFIG_SMP */ | |
1317 | ||
e52fb7c0 PZ |
1318 | /* |
1319 | * Adaptive granularity | |
1320 | * | |
1321 | * se->avg_wakeup gives the average time a task runs until it does a wakeup, | |
1322 | * with the limit of wakeup_gran -- when it never does a wakeup. | |
1323 | * | |
1324 | * So the smaller avg_wakeup is the faster we want this task to preempt, | |
1325 | * but we don't want to treat the preemptee unfairly and therefore allow it | |
1326 | * to run for at least the amount of time we'd like to run. | |
1327 | * | |
1328 | * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one | |
1329 | * | |
1330 | * NOTE: we use *nr_running to scale with load, this nicely matches the | |
1331 | * degrading latency on load. | |
1332 | */ | |
1333 | static unsigned long | |
1334 | adaptive_gran(struct sched_entity *curr, struct sched_entity *se) | |
1335 | { | |
1336 | u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
1337 | u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running; | |
1338 | u64 gran = 0; | |
1339 | ||
1340 | if (this_run < expected_wakeup) | |
1341 | gran = expected_wakeup - this_run; | |
1342 | ||
1343 | return min_t(s64, gran, sysctl_sched_wakeup_granularity); | |
1344 | } | |
1345 | ||
1346 | static unsigned long | |
1347 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1348 | { |
1349 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1350 | ||
e52fb7c0 PZ |
1351 | if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN)) |
1352 | gran = adaptive_gran(curr, se); | |
1353 | ||
0bbd3336 | 1354 | /* |
e52fb7c0 PZ |
1355 | * Since its curr running now, convert the gran from real-time |
1356 | * to virtual-time in his units. | |
0bbd3336 | 1357 | */ |
e52fb7c0 PZ |
1358 | if (sched_feat(ASYM_GRAN)) { |
1359 | /* | |
1360 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1361 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1362 | * the resulting gran will be larger, therefore penalizing the | |
1363 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1364 | * be smaller, again penalizing the lighter task. | |
1365 | * | |
1366 | * This is especially important for buddies when the leftmost | |
1367 | * task is higher priority than the buddy. | |
1368 | */ | |
1369 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1370 | gran = calc_delta_fair(gran, se); | |
1371 | } else { | |
1372 | if (unlikely(curr->load.weight != NICE_0_LOAD)) | |
1373 | gran = calc_delta_fair(gran, curr); | |
1374 | } | |
0bbd3336 PZ |
1375 | |
1376 | return gran; | |
1377 | } | |
1378 | ||
464b7527 PZ |
1379 | /* |
1380 | * Should 'se' preempt 'curr'. | |
1381 | * | |
1382 | * |s1 | |
1383 | * |s2 | |
1384 | * |s3 | |
1385 | * g | |
1386 | * |<--->|c | |
1387 | * | |
1388 | * w(c, s1) = -1 | |
1389 | * w(c, s2) = 0 | |
1390 | * w(c, s3) = 1 | |
1391 | * | |
1392 | */ | |
1393 | static int | |
1394 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1395 | { | |
1396 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1397 | ||
1398 | if (vdiff <= 0) | |
1399 | return -1; | |
1400 | ||
e52fb7c0 | 1401 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1402 | if (vdiff > gran) |
1403 | return 1; | |
1404 | ||
1405 | return 0; | |
1406 | } | |
1407 | ||
02479099 PZ |
1408 | static void set_last_buddy(struct sched_entity *se) |
1409 | { | |
6bc912b7 PZ |
1410 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1411 | for_each_sched_entity(se) | |
1412 | cfs_rq_of(se)->last = se; | |
1413 | } | |
02479099 PZ |
1414 | } |
1415 | ||
1416 | static void set_next_buddy(struct sched_entity *se) | |
1417 | { | |
6bc912b7 PZ |
1418 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1419 | for_each_sched_entity(se) | |
1420 | cfs_rq_of(se)->next = se; | |
1421 | } | |
02479099 PZ |
1422 | } |
1423 | ||
bf0f6f24 IM |
1424 | /* |
1425 | * Preempt the current task with a newly woken task if needed: | |
1426 | */ | |
15afe09b | 1427 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync) |
bf0f6f24 IM |
1428 | { |
1429 | struct task_struct *curr = rq->curr; | |
8651a86c | 1430 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1431 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
bf0f6f24 | 1432 | |
03e89e45 | 1433 | update_curr(cfs_rq); |
4793241b | 1434 | |
03e89e45 | 1435 | if (unlikely(rt_prio(p->prio))) { |
bf0f6f24 IM |
1436 | resched_task(curr); |
1437 | return; | |
1438 | } | |
aa2ac252 | 1439 | |
d95f98d0 PZ |
1440 | if (unlikely(p->sched_class != &fair_sched_class)) |
1441 | return; | |
1442 | ||
4ae7d5ce IM |
1443 | if (unlikely(se == pse)) |
1444 | return; | |
1445 | ||
4793241b PZ |
1446 | /* |
1447 | * Only set the backward buddy when the current task is still on the | |
1448 | * rq. This can happen when a wakeup gets interleaved with schedule on | |
1449 | * the ->pre_schedule() or idle_balance() point, either of which can | |
1450 | * drop the rq lock. | |
1451 | * | |
1452 | * Also, during early boot the idle thread is in the fair class, for | |
1453 | * obvious reasons its a bad idea to schedule back to the idle thread. | |
1454 | */ | |
1455 | if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle)) | |
02479099 PZ |
1456 | set_last_buddy(se); |
1457 | set_next_buddy(pse); | |
57fdc26d | 1458 | |
aec0a514 BR |
1459 | /* |
1460 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1461 | * wake up path. | |
1462 | */ | |
1463 | if (test_tsk_need_resched(curr)) | |
1464 | return; | |
1465 | ||
91c234b4 | 1466 | /* |
6bc912b7 | 1467 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1468 | * the tick): |
1469 | */ | |
6bc912b7 | 1470 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1471 | return; |
bf0f6f24 | 1472 | |
6bc912b7 PZ |
1473 | /* Idle tasks are by definition preempted by everybody. */ |
1474 | if (unlikely(curr->policy == SCHED_IDLE)) { | |
1475 | resched_task(curr); | |
91c234b4 | 1476 | return; |
6bc912b7 | 1477 | } |
bf0f6f24 | 1478 | |
77d9cc44 IM |
1479 | if (!sched_feat(WAKEUP_PREEMPT)) |
1480 | return; | |
8651a86c | 1481 | |
fc631c82 PZ |
1482 | if (sched_feat(WAKEUP_OVERLAP) && (sync || |
1483 | (se->avg_overlap < sysctl_sched_migration_cost && | |
1484 | pse->avg_overlap < sysctl_sched_migration_cost))) { | |
15afe09b PZ |
1485 | resched_task(curr); |
1486 | return; | |
1487 | } | |
1488 | ||
464b7527 PZ |
1489 | find_matching_se(&se, &pse); |
1490 | ||
002f128b | 1491 | BUG_ON(!pse); |
464b7527 | 1492 | |
002f128b PT |
1493 | if (wakeup_preempt_entity(se, pse) == 1) |
1494 | resched_task(curr); | |
bf0f6f24 IM |
1495 | } |
1496 | ||
fb8d4724 | 1497 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1498 | { |
8f4d37ec | 1499 | struct task_struct *p; |
bf0f6f24 IM |
1500 | struct cfs_rq *cfs_rq = &rq->cfs; |
1501 | struct sched_entity *se; | |
1502 | ||
1503 | if (unlikely(!cfs_rq->nr_running)) | |
1504 | return NULL; | |
1505 | ||
1506 | do { | |
9948f4b2 | 1507 | se = pick_next_entity(cfs_rq); |
a9f3e2b5 MG |
1508 | /* |
1509 | * If se was a buddy, clear it so that it will have to earn | |
1510 | * the favour again. | |
1511 | */ | |
a571bbea | 1512 | __clear_buddies(cfs_rq, se); |
f4b6755f | 1513 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1514 | cfs_rq = group_cfs_rq(se); |
1515 | } while (cfs_rq); | |
1516 | ||
8f4d37ec PZ |
1517 | p = task_of(se); |
1518 | hrtick_start_fair(rq, p); | |
1519 | ||
1520 | return p; | |
bf0f6f24 IM |
1521 | } |
1522 | ||
1523 | /* | |
1524 | * Account for a descheduled task: | |
1525 | */ | |
31ee529c | 1526 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1527 | { |
1528 | struct sched_entity *se = &prev->se; | |
1529 | struct cfs_rq *cfs_rq; | |
1530 | ||
1531 | for_each_sched_entity(se) { | |
1532 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1533 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1534 | } |
1535 | } | |
1536 | ||
681f3e68 | 1537 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1538 | /************************************************** |
1539 | * Fair scheduling class load-balancing methods: | |
1540 | */ | |
1541 | ||
1542 | /* | |
1543 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1544 | * during the whole iteration, the current task might be | |
1545 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1546 | * achieve that by always pre-iterating before returning | |
1547 | * the current task: | |
1548 | */ | |
a9957449 | 1549 | static struct task_struct * |
4a55bd5e | 1550 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) |
bf0f6f24 | 1551 | { |
354d60c2 DG |
1552 | struct task_struct *p = NULL; |
1553 | struct sched_entity *se; | |
bf0f6f24 | 1554 | |
77ae6513 MG |
1555 | if (next == &cfs_rq->tasks) |
1556 | return NULL; | |
1557 | ||
b87f1724 BR |
1558 | se = list_entry(next, struct sched_entity, group_node); |
1559 | p = task_of(se); | |
1560 | cfs_rq->balance_iterator = next->next; | |
77ae6513 | 1561 | |
bf0f6f24 IM |
1562 | return p; |
1563 | } | |
1564 | ||
1565 | static struct task_struct *load_balance_start_fair(void *arg) | |
1566 | { | |
1567 | struct cfs_rq *cfs_rq = arg; | |
1568 | ||
4a55bd5e | 1569 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); |
bf0f6f24 IM |
1570 | } |
1571 | ||
1572 | static struct task_struct *load_balance_next_fair(void *arg) | |
1573 | { | |
1574 | struct cfs_rq *cfs_rq = arg; | |
1575 | ||
4a55bd5e | 1576 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); |
bf0f6f24 IM |
1577 | } |
1578 | ||
c09595f6 PZ |
1579 | static unsigned long |
1580 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1581 | unsigned long max_load_move, struct sched_domain *sd, | |
1582 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, | |
1583 | struct cfs_rq *cfs_rq) | |
62fb1851 | 1584 | { |
c09595f6 | 1585 | struct rq_iterator cfs_rq_iterator; |
62fb1851 | 1586 | |
c09595f6 PZ |
1587 | cfs_rq_iterator.start = load_balance_start_fair; |
1588 | cfs_rq_iterator.next = load_balance_next_fair; | |
1589 | cfs_rq_iterator.arg = cfs_rq; | |
62fb1851 | 1590 | |
c09595f6 PZ |
1591 | return balance_tasks(this_rq, this_cpu, busiest, |
1592 | max_load_move, sd, idle, all_pinned, | |
1593 | this_best_prio, &cfs_rq_iterator); | |
62fb1851 | 1594 | } |
62fb1851 | 1595 | |
c09595f6 | 1596 | #ifdef CONFIG_FAIR_GROUP_SCHED |
43010659 | 1597 | static unsigned long |
bf0f6f24 | 1598 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1599 | unsigned long max_load_move, |
a4ac01c3 PW |
1600 | struct sched_domain *sd, enum cpu_idle_type idle, |
1601 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 | 1602 | { |
bf0f6f24 | 1603 | long rem_load_move = max_load_move; |
c09595f6 PZ |
1604 | int busiest_cpu = cpu_of(busiest); |
1605 | struct task_group *tg; | |
18d95a28 | 1606 | |
c09595f6 | 1607 | rcu_read_lock(); |
c8cba857 | 1608 | update_h_load(busiest_cpu); |
18d95a28 | 1609 | |
caea8a03 | 1610 | list_for_each_entry_rcu(tg, &task_groups, list) { |
c8cba857 | 1611 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; |
42a3ac7d PZ |
1612 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
1613 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
243e0e7b | 1614 | u64 rem_load, moved_load; |
18d95a28 | 1615 | |
c09595f6 PZ |
1616 | /* |
1617 | * empty group | |
1618 | */ | |
c8cba857 | 1619 | if (!busiest_cfs_rq->task_weight) |
bf0f6f24 IM |
1620 | continue; |
1621 | ||
243e0e7b SV |
1622 | rem_load = (u64)rem_load_move * busiest_weight; |
1623 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
bf0f6f24 | 1624 | |
c09595f6 | 1625 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, |
53fecd8a | 1626 | rem_load, sd, idle, all_pinned, this_best_prio, |
c09595f6 | 1627 | tg->cfs_rq[busiest_cpu]); |
bf0f6f24 | 1628 | |
c09595f6 | 1629 | if (!moved_load) |
bf0f6f24 IM |
1630 | continue; |
1631 | ||
42a3ac7d | 1632 | moved_load *= busiest_h_load; |
243e0e7b | 1633 | moved_load = div_u64(moved_load, busiest_weight + 1); |
bf0f6f24 | 1634 | |
c09595f6 PZ |
1635 | rem_load_move -= moved_load; |
1636 | if (rem_load_move < 0) | |
bf0f6f24 IM |
1637 | break; |
1638 | } | |
c09595f6 | 1639 | rcu_read_unlock(); |
bf0f6f24 | 1640 | |
43010659 | 1641 | return max_load_move - rem_load_move; |
bf0f6f24 | 1642 | } |
c09595f6 PZ |
1643 | #else |
1644 | static unsigned long | |
1645 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1646 | unsigned long max_load_move, | |
1647 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1648 | int *all_pinned, int *this_best_prio) | |
1649 | { | |
1650 | return __load_balance_fair(this_rq, this_cpu, busiest, | |
1651 | max_load_move, sd, idle, all_pinned, | |
1652 | this_best_prio, &busiest->cfs); | |
1653 | } | |
1654 | #endif | |
bf0f6f24 | 1655 | |
e1d1484f PW |
1656 | static int |
1657 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1658 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1659 | { | |
1660 | struct cfs_rq *busy_cfs_rq; | |
1661 | struct rq_iterator cfs_rq_iterator; | |
1662 | ||
1663 | cfs_rq_iterator.start = load_balance_start_fair; | |
1664 | cfs_rq_iterator.next = load_balance_next_fair; | |
1665 | ||
1666 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1667 | /* | |
1668 | * pass busy_cfs_rq argument into | |
1669 | * load_balance_[start|next]_fair iterators | |
1670 | */ | |
1671 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1672 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1673 | &cfs_rq_iterator)) | |
1674 | return 1; | |
1675 | } | |
1676 | ||
1677 | return 0; | |
1678 | } | |
55e12e5e | 1679 | #endif /* CONFIG_SMP */ |
e1d1484f | 1680 | |
bf0f6f24 IM |
1681 | /* |
1682 | * scheduler tick hitting a task of our scheduling class: | |
1683 | */ | |
8f4d37ec | 1684 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1685 | { |
1686 | struct cfs_rq *cfs_rq; | |
1687 | struct sched_entity *se = &curr->se; | |
1688 | ||
1689 | for_each_sched_entity(se) { | |
1690 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1691 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1692 | } |
1693 | } | |
1694 | ||
1695 | /* | |
1696 | * Share the fairness runtime between parent and child, thus the | |
1697 | * total amount of pressure for CPU stays equal - new tasks | |
1698 | * get a chance to run but frequent forkers are not allowed to | |
1699 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1700 | * the child is not running yet. | |
1701 | */ | |
ee0827d8 | 1702 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1703 | { |
1704 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1705 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1706 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1707 | |
1708 | sched_info_queued(p); | |
1709 | ||
7109c442 | 1710 | update_curr(cfs_rq); |
aeb73b04 | 1711 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1712 | |
3c90e6e9 | 1713 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1714 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1715 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1716 | /* |
edcb60a3 IM |
1717 | * Upon rescheduling, sched_class::put_prev_task() will place |
1718 | * 'current' within the tree based on its new key value. | |
1719 | */ | |
4d78e7b6 | 1720 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 1721 | resched_task(rq->curr); |
4d78e7b6 | 1722 | } |
bf0f6f24 | 1723 | |
b9dca1e0 | 1724 | enqueue_task_fair(rq, p, 0); |
bf0f6f24 IM |
1725 | } |
1726 | ||
cb469845 SR |
1727 | /* |
1728 | * Priority of the task has changed. Check to see if we preempt | |
1729 | * the current task. | |
1730 | */ | |
1731 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1732 | int oldprio, int running) | |
1733 | { | |
1734 | /* | |
1735 | * Reschedule if we are currently running on this runqueue and | |
1736 | * our priority decreased, or if we are not currently running on | |
1737 | * this runqueue and our priority is higher than the current's | |
1738 | */ | |
1739 | if (running) { | |
1740 | if (p->prio > oldprio) | |
1741 | resched_task(rq->curr); | |
1742 | } else | |
15afe09b | 1743 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1744 | } |
1745 | ||
1746 | /* | |
1747 | * We switched to the sched_fair class. | |
1748 | */ | |
1749 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1750 | int running) | |
1751 | { | |
1752 | /* | |
1753 | * We were most likely switched from sched_rt, so | |
1754 | * kick off the schedule if running, otherwise just see | |
1755 | * if we can still preempt the current task. | |
1756 | */ | |
1757 | if (running) | |
1758 | resched_task(rq->curr); | |
1759 | else | |
15afe09b | 1760 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1761 | } |
1762 | ||
83b699ed SV |
1763 | /* Account for a task changing its policy or group. |
1764 | * | |
1765 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1766 | * migrates between groups/classes. | |
1767 | */ | |
1768 | static void set_curr_task_fair(struct rq *rq) | |
1769 | { | |
1770 | struct sched_entity *se = &rq->curr->se; | |
1771 | ||
1772 | for_each_sched_entity(se) | |
1773 | set_next_entity(cfs_rq_of(se), se); | |
1774 | } | |
1775 | ||
810b3817 PZ |
1776 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1777 | static void moved_group_fair(struct task_struct *p) | |
1778 | { | |
1779 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1780 | ||
1781 | update_curr(cfs_rq); | |
1782 | place_entity(cfs_rq, &p->se, 1); | |
1783 | } | |
1784 | #endif | |
1785 | ||
bf0f6f24 IM |
1786 | /* |
1787 | * All the scheduling class methods: | |
1788 | */ | |
5522d5d5 IM |
1789 | static const struct sched_class fair_sched_class = { |
1790 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1791 | .enqueue_task = enqueue_task_fair, |
1792 | .dequeue_task = dequeue_task_fair, | |
1793 | .yield_task = yield_task_fair, | |
1794 | ||
2e09bf55 | 1795 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1796 | |
1797 | .pick_next_task = pick_next_task_fair, | |
1798 | .put_prev_task = put_prev_task_fair, | |
1799 | ||
681f3e68 | 1800 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
1801 | .select_task_rq = select_task_rq_fair, |
1802 | ||
bf0f6f24 | 1803 | .load_balance = load_balance_fair, |
e1d1484f | 1804 | .move_one_task = move_one_task_fair, |
681f3e68 | 1805 | #endif |
bf0f6f24 | 1806 | |
83b699ed | 1807 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1808 | .task_tick = task_tick_fair, |
1809 | .task_new = task_new_fair, | |
cb469845 SR |
1810 | |
1811 | .prio_changed = prio_changed_fair, | |
1812 | .switched_to = switched_to_fair, | |
810b3817 PZ |
1813 | |
1814 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1815 | .moved_group = moved_group_fair, | |
1816 | #endif | |
bf0f6f24 IM |
1817 | }; |
1818 | ||
1819 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1820 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1821 | { |
bf0f6f24 IM |
1822 | struct cfs_rq *cfs_rq; |
1823 | ||
5973e5b9 | 1824 | rcu_read_lock(); |
c3b64f1e | 1825 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1826 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1827 | rcu_read_unlock(); |
bf0f6f24 IM |
1828 | } |
1829 | #endif |