]>
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 | ||
23 | /* | |
21805085 PZ |
24 | * Targeted preemption latency for CPU-bound tasks: |
25 | * (default: 20ms, units: nanoseconds) | |
bf0f6f24 | 26 | * |
21805085 PZ |
27 | * NOTE: this latency value is not the same as the concept of |
28 | * 'timeslice length' - timeslices in CFS are of variable length. | |
29 | * (to see the precise effective timeslice length of your workload, | |
30 | * run vmstat and monitor the context-switches field) | |
bf0f6f24 IM |
31 | * |
32 | * On SMP systems the value of this is multiplied by the log2 of the | |
33 | * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way | |
34 | * systems, 4x on 8-way systems, 5x on 16-way systems, etc.) | |
21805085 | 35 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 36 | */ |
21805085 PZ |
37 | unsigned int sysctl_sched_latency __read_mostly = 20000000ULL; |
38 | ||
39 | /* | |
40 | * Minimal preemption granularity for CPU-bound tasks: | |
41 | * (default: 2 msec, units: nanoseconds) | |
42 | */ | |
172ac3db | 43 | unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL; |
bf0f6f24 | 44 | |
1799e35d IM |
45 | /* |
46 | * sys_sched_yield() compat mode | |
47 | * | |
48 | * This option switches the agressive yield implementation of the | |
49 | * old scheduler back on. | |
50 | */ | |
51 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
52 | ||
bf0f6f24 IM |
53 | /* |
54 | * SCHED_BATCH wake-up granularity. | |
71fd3714 | 55 | * (default: 25 msec, units: nanoseconds) |
bf0f6f24 IM |
56 | * |
57 | * This option delays the preemption effects of decoupled workloads | |
58 | * and reduces their over-scheduling. Synchronous workloads will still | |
59 | * have immediate wakeup/sleep latencies. | |
60 | */ | |
71fd3714 | 61 | unsigned int sysctl_sched_batch_wakeup_granularity __read_mostly = 25000000UL; |
bf0f6f24 IM |
62 | |
63 | /* | |
64 | * SCHED_OTHER wake-up granularity. | |
65 | * (default: 1 msec, units: nanoseconds) | |
66 | * | |
67 | * This option delays the preemption effects of decoupled workloads | |
68 | * and reduces their over-scheduling. Synchronous workloads will still | |
69 | * have immediate wakeup/sleep latencies. | |
70 | */ | |
71fd3714 | 71 | unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000UL; |
bf0f6f24 IM |
72 | |
73 | unsigned int sysctl_sched_stat_granularity __read_mostly; | |
74 | ||
75 | /* | |
71fd3714 | 76 | * Initialized in sched_init_granularity() [to 5 times the base granularity]: |
bf0f6f24 IM |
77 | */ |
78 | unsigned int sysctl_sched_runtime_limit __read_mostly; | |
79 | ||
80 | /* | |
81 | * Debugging: various feature bits | |
82 | */ | |
83 | enum { | |
84 | SCHED_FEAT_FAIR_SLEEPERS = 1, | |
85 | SCHED_FEAT_SLEEPER_AVG = 2, | |
86 | SCHED_FEAT_SLEEPER_LOAD_AVG = 4, | |
87 | SCHED_FEAT_PRECISE_CPU_LOAD = 8, | |
88 | SCHED_FEAT_START_DEBIT = 16, | |
89 | SCHED_FEAT_SKIP_INITIAL = 32, | |
90 | }; | |
91 | ||
92 | unsigned int sysctl_sched_features __read_mostly = | |
93 | SCHED_FEAT_FAIR_SLEEPERS *1 | | |
5d2b3d36 | 94 | SCHED_FEAT_SLEEPER_AVG *0 | |
bf0f6f24 IM |
95 | SCHED_FEAT_SLEEPER_LOAD_AVG *1 | |
96 | SCHED_FEAT_PRECISE_CPU_LOAD *1 | | |
97 | SCHED_FEAT_START_DEBIT *1 | | |
98 | SCHED_FEAT_SKIP_INITIAL *0; | |
99 | ||
100 | extern struct sched_class fair_sched_class; | |
101 | ||
102 | /************************************************************** | |
103 | * CFS operations on generic schedulable entities: | |
104 | */ | |
105 | ||
106 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
107 | ||
108 | /* cpu runqueue to which this cfs_rq is attached */ | |
109 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
110 | { | |
111 | return cfs_rq->rq; | |
112 | } | |
113 | ||
114 | /* currently running entity (if any) on this cfs_rq */ | |
115 | static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq) | |
116 | { | |
117 | return cfs_rq->curr; | |
118 | } | |
119 | ||
120 | /* An entity is a task if it doesn't "own" a runqueue */ | |
121 | #define entity_is_task(se) (!se->my_q) | |
122 | ||
123 | static inline void | |
124 | set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
125 | { | |
126 | cfs_rq->curr = se; | |
127 | } | |
128 | ||
129 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
130 | ||
131 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
132 | { | |
133 | return container_of(cfs_rq, struct rq, cfs); | |
134 | } | |
135 | ||
136 | static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq) | |
137 | { | |
138 | struct rq *rq = rq_of(cfs_rq); | |
139 | ||
140 | if (unlikely(rq->curr->sched_class != &fair_sched_class)) | |
141 | return NULL; | |
142 | ||
143 | return &rq->curr->se; | |
144 | } | |
145 | ||
146 | #define entity_is_task(se) 1 | |
147 | ||
148 | static inline void | |
149 | set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
150 | ||
151 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
152 | ||
153 | static inline struct task_struct *task_of(struct sched_entity *se) | |
154 | { | |
155 | return container_of(se, struct task_struct, se); | |
156 | } | |
157 | ||
158 | ||
159 | /************************************************************** | |
160 | * Scheduling class tree data structure manipulation methods: | |
161 | */ | |
162 | ||
163 | /* | |
164 | * Enqueue an entity into the rb-tree: | |
165 | */ | |
166 | static inline void | |
167 | __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
168 | { | |
169 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
170 | struct rb_node *parent = NULL; | |
171 | struct sched_entity *entry; | |
172 | s64 key = se->fair_key; | |
173 | int leftmost = 1; | |
174 | ||
175 | /* | |
176 | * Find the right place in the rbtree: | |
177 | */ | |
178 | while (*link) { | |
179 | parent = *link; | |
180 | entry = rb_entry(parent, struct sched_entity, run_node); | |
181 | /* | |
182 | * We dont care about collisions. Nodes with | |
183 | * the same key stay together. | |
184 | */ | |
185 | if (key - entry->fair_key < 0) { | |
186 | link = &parent->rb_left; | |
187 | } else { | |
188 | link = &parent->rb_right; | |
189 | leftmost = 0; | |
190 | } | |
191 | } | |
192 | ||
193 | /* | |
194 | * Maintain a cache of leftmost tree entries (it is frequently | |
195 | * used): | |
196 | */ | |
197 | if (leftmost) | |
198 | cfs_rq->rb_leftmost = &se->run_node; | |
199 | ||
200 | rb_link_node(&se->run_node, parent, link); | |
201 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
202 | update_load_add(&cfs_rq->load, se->load.weight); | |
203 | cfs_rq->nr_running++; | |
204 | se->on_rq = 1; | |
a206c072 IM |
205 | |
206 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
bf0f6f24 IM |
207 | } |
208 | ||
209 | static inline void | |
210 | __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
211 | { | |
212 | if (cfs_rq->rb_leftmost == &se->run_node) | |
213 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
214 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
215 | update_load_sub(&cfs_rq->load, se->load.weight); | |
216 | cfs_rq->nr_running--; | |
217 | se->on_rq = 0; | |
a206c072 IM |
218 | |
219 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
bf0f6f24 IM |
220 | } |
221 | ||
222 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
223 | { | |
224 | return cfs_rq->rb_leftmost; | |
225 | } | |
226 | ||
227 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
228 | { | |
229 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
230 | } | |
231 | ||
232 | /************************************************************** | |
233 | * Scheduling class statistics methods: | |
234 | */ | |
235 | ||
21805085 PZ |
236 | /* |
237 | * Calculate the preemption granularity needed to schedule every | |
238 | * runnable task once per sysctl_sched_latency amount of time. | |
239 | * (down to a sensible low limit on granularity) | |
240 | * | |
241 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
242 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
243 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
244 | * for each task. We do finer and finer scheduling up to until we | |
245 | * reach the minimum granularity value. | |
246 | * | |
247 | * To achieve this we use the following dynamic-granularity rule: | |
248 | * | |
249 | * gran = lat/nr - lat/nr/nr | |
250 | * | |
251 | * This comes out of the following equations: | |
252 | * | |
253 | * kA1 + gran = kB1 | |
254 | * kB2 + gran = kA2 | |
255 | * kA2 = kA1 | |
256 | * kB2 = kB1 - d + d/nr | |
257 | * lat = d * nr | |
258 | * | |
259 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
260 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
261 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
262 | * running, 'lat' is the the period of each task. ('lat' is the | |
263 | * sched_latency that we aim for.) | |
264 | */ | |
265 | static long | |
266 | sched_granularity(struct cfs_rq *cfs_rq) | |
267 | { | |
268 | unsigned int gran = sysctl_sched_latency; | |
269 | unsigned int nr = cfs_rq->nr_running; | |
270 | ||
271 | if (nr > 1) { | |
272 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 273 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
274 | } |
275 | ||
276 | return gran; | |
277 | } | |
278 | ||
bf0f6f24 IM |
279 | /* |
280 | * We rescale the rescheduling granularity of tasks according to their | |
281 | * nice level, but only linearly, not exponentially: | |
282 | */ | |
283 | static long | |
284 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
285 | { | |
286 | u64 tmp; | |
287 | ||
7cff8cf6 IM |
288 | if (likely(curr->load.weight == NICE_0_LOAD)) |
289 | return granularity; | |
bf0f6f24 | 290 | /* |
7cff8cf6 | 291 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 292 | */ |
7cff8cf6 IM |
293 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
294 | tmp = curr->load.weight * (u64)granularity; | |
295 | return (long) (tmp >> NICE_0_SHIFT); | |
296 | } | |
bf0f6f24 | 297 | /* |
7cff8cf6 | 298 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
299 | * granularity: |
300 | */ | |
7cff8cf6 | 301 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
302 | |
303 | /* | |
304 | * It will always fit into 'long': | |
305 | */ | |
a0dc7260 | 306 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
307 | } |
308 | ||
309 | static inline void | |
310 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
311 | { | |
312 | long limit = sysctl_sched_runtime_limit; | |
313 | ||
314 | /* | |
315 | * Niced tasks have the same history dynamic range as | |
316 | * non-niced tasks: | |
317 | */ | |
318 | if (unlikely(se->wait_runtime > limit)) { | |
319 | se->wait_runtime = limit; | |
320 | schedstat_inc(se, wait_runtime_overruns); | |
321 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
322 | } | |
323 | if (unlikely(se->wait_runtime < -limit)) { | |
324 | se->wait_runtime = -limit; | |
325 | schedstat_inc(se, wait_runtime_underruns); | |
326 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
327 | } | |
328 | } | |
329 | ||
330 | static inline void | |
331 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
332 | { | |
333 | se->wait_runtime += delta; | |
334 | schedstat_add(se, sum_wait_runtime, delta); | |
335 | limit_wait_runtime(cfs_rq, se); | |
336 | } | |
337 | ||
338 | static void | |
339 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
340 | { | |
341 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
342 | __add_wait_runtime(cfs_rq, se, delta); | |
343 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
344 | } | |
345 | ||
346 | /* | |
347 | * Update the current task's runtime statistics. Skip current tasks that | |
348 | * are not in our scheduling class. | |
349 | */ | |
350 | static inline void | |
b7cc0896 | 351 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 352 | { |
c5dcfe72 | 353 | unsigned long delta, delta_exec, delta_fair, delta_mine; |
bf0f6f24 IM |
354 | struct load_weight *lw = &cfs_rq->load; |
355 | unsigned long load = lw->weight; | |
356 | ||
bf0f6f24 | 357 | delta_exec = curr->delta_exec; |
8179ca23 | 358 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
359 | |
360 | curr->sum_exec_runtime += delta_exec; | |
361 | cfs_rq->exec_clock += delta_exec; | |
362 | ||
fd8bb43e IM |
363 | if (unlikely(!load)) |
364 | return; | |
365 | ||
bf0f6f24 IM |
366 | delta_fair = calc_delta_fair(delta_exec, lw); |
367 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
368 | ||
5f01d519 | 369 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 370 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
371 | delta = min(delta, (unsigned long)( |
372 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
373 | cfs_rq->sleeper_bonus -= delta; |
374 | delta_mine -= delta; | |
375 | } | |
376 | ||
377 | cfs_rq->fair_clock += delta_fair; | |
378 | /* | |
379 | * We executed delta_exec amount of time on the CPU, | |
380 | * but we were only entitled to delta_mine amount of | |
381 | * time during that period (if nr_running == 1 then | |
382 | * the two values are equal) | |
383 | * [Note: delta_mine - delta_exec is negative]: | |
384 | */ | |
385 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
386 | } | |
387 | ||
b7cc0896 | 388 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
389 | { |
390 | struct sched_entity *curr = cfs_rq_curr(cfs_rq); | |
391 | unsigned long delta_exec; | |
392 | ||
393 | if (unlikely(!curr)) | |
394 | return; | |
395 | ||
396 | /* | |
397 | * Get the amount of time the current task was running | |
398 | * since the last time we changed load (this cannot | |
399 | * overflow on 32 bits): | |
400 | */ | |
d281918d | 401 | delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start); |
bf0f6f24 IM |
402 | |
403 | curr->delta_exec += delta_exec; | |
404 | ||
405 | if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) { | |
b7cc0896 | 406 | __update_curr(cfs_rq, curr); |
bf0f6f24 IM |
407 | curr->delta_exec = 0; |
408 | } | |
d281918d | 409 | curr->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
410 | } |
411 | ||
412 | static inline void | |
5870db5b | 413 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
414 | { |
415 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 416 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
417 | } |
418 | ||
419 | /* | |
420 | * We calculate fair deltas here, so protect against the random effects | |
421 | * of a multiplication overflow by capping it to the runtime limit: | |
422 | */ | |
423 | #if BITS_PER_LONG == 32 | |
424 | static inline unsigned long | |
425 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
426 | { | |
427 | u64 tmp = (u64)delta * weight >> shift; | |
428 | ||
429 | if (unlikely(tmp > sysctl_sched_runtime_limit*2)) | |
430 | return sysctl_sched_runtime_limit*2; | |
431 | return tmp; | |
432 | } | |
433 | #else | |
434 | static inline unsigned long | |
435 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
436 | { | |
437 | return delta * weight >> shift; | |
438 | } | |
439 | #endif | |
440 | ||
441 | /* | |
442 | * Task is being enqueued - update stats: | |
443 | */ | |
d2417e5a | 444 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
445 | { |
446 | s64 key; | |
447 | ||
448 | /* | |
449 | * Are we enqueueing a waiting task? (for current tasks | |
450 | * a dequeue/enqueue event is a NOP) | |
451 | */ | |
452 | if (se != cfs_rq_curr(cfs_rq)) | |
5870db5b | 453 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
454 | /* |
455 | * Update the key: | |
456 | */ | |
457 | key = cfs_rq->fair_clock; | |
458 | ||
459 | /* | |
460 | * Optimize the common nice 0 case: | |
461 | */ | |
462 | if (likely(se->load.weight == NICE_0_LOAD)) { | |
463 | key -= se->wait_runtime; | |
464 | } else { | |
465 | u64 tmp; | |
466 | ||
467 | if (se->wait_runtime < 0) { | |
468 | tmp = -se->wait_runtime; | |
469 | key += (tmp * se->load.inv_weight) >> | |
470 | (WMULT_SHIFT - NICE_0_SHIFT); | |
471 | } else { | |
472 | tmp = se->wait_runtime; | |
a69edb55 IM |
473 | key -= (tmp * se->load.inv_weight) >> |
474 | (WMULT_SHIFT - NICE_0_SHIFT); | |
bf0f6f24 IM |
475 | } |
476 | } | |
477 | ||
478 | se->fair_key = key; | |
479 | } | |
480 | ||
481 | /* | |
482 | * Note: must be called with a freshly updated rq->fair_clock. | |
483 | */ | |
484 | static inline void | |
eac55ea3 | 485 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
486 | { |
487 | unsigned long delta_fair = se->delta_fair_run; | |
488 | ||
d281918d IM |
489 | schedstat_set(se->wait_max, max(se->wait_max, |
490 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 IM |
491 | |
492 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
493 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
494 | NICE_0_SHIFT); | |
495 | ||
496 | add_wait_runtime(cfs_rq, se, delta_fair); | |
497 | } | |
498 | ||
499 | static void | |
9ef0a961 | 500 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
501 | { |
502 | unsigned long delta_fair; | |
503 | ||
b77d69db IM |
504 | if (unlikely(!se->wait_start_fair)) |
505 | return; | |
506 | ||
bf0f6f24 IM |
507 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
508 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
509 | ||
510 | se->delta_fair_run += delta_fair; | |
511 | if (unlikely(abs(se->delta_fair_run) >= | |
512 | sysctl_sched_stat_granularity)) { | |
eac55ea3 | 513 | __update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
514 | se->delta_fair_run = 0; |
515 | } | |
516 | ||
517 | se->wait_start_fair = 0; | |
6cfb0d5d | 518 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
519 | } |
520 | ||
521 | static inline void | |
19b6a2e3 | 522 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 523 | { |
b7cc0896 | 524 | update_curr(cfs_rq); |
bf0f6f24 IM |
525 | /* |
526 | * Mark the end of the wait period if dequeueing a | |
527 | * waiting task: | |
528 | */ | |
529 | if (se != cfs_rq_curr(cfs_rq)) | |
9ef0a961 | 530 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
531 | } |
532 | ||
533 | /* | |
534 | * We are picking a new current task - update its stats: | |
535 | */ | |
536 | static inline void | |
79303e9e | 537 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
538 | { |
539 | /* | |
540 | * We are starting a new run period: | |
541 | */ | |
d281918d | 542 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
543 | } |
544 | ||
545 | /* | |
546 | * We are descheduling a task - update its stats: | |
547 | */ | |
548 | static inline void | |
c7e9b5b2 | 549 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
550 | { |
551 | se->exec_start = 0; | |
552 | } | |
553 | ||
554 | /************************************************** | |
555 | * Scheduling class queueing methods: | |
556 | */ | |
557 | ||
dfdc119e | 558 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
559 | { |
560 | unsigned long load = cfs_rq->load.weight, delta_fair; | |
561 | long prev_runtime; | |
562 | ||
b2133c8b IM |
563 | /* |
564 | * Do not boost sleepers if there's too much bonus 'in flight' | |
565 | * already: | |
566 | */ | |
567 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
568 | return; | |
569 | ||
bf0f6f24 IM |
570 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG) |
571 | load = rq_of(cfs_rq)->cpu_load[2]; | |
572 | ||
573 | delta_fair = se->delta_fair_sleep; | |
574 | ||
575 | /* | |
576 | * Fix up delta_fair with the effect of us running | |
577 | * during the whole sleep period: | |
578 | */ | |
579 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG) | |
580 | delta_fair = div64_likely32((u64)delta_fair * load, | |
581 | load + se->load.weight); | |
582 | ||
583 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
584 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
585 | NICE_0_SHIFT); | |
586 | ||
587 | prev_runtime = se->wait_runtime; | |
588 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
589 | delta_fair = se->wait_runtime - prev_runtime; | |
590 | ||
591 | /* | |
592 | * Track the amount of bonus we've given to sleepers: | |
593 | */ | |
594 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
595 | } |
596 | ||
2396af69 | 597 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
598 | { |
599 | struct task_struct *tsk = task_of(se); | |
600 | unsigned long delta_fair; | |
601 | ||
602 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
603 | !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS)) | |
604 | return; | |
605 | ||
606 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
607 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
608 | ||
609 | se->delta_fair_sleep += delta_fair; | |
610 | if (unlikely(abs(se->delta_fair_sleep) >= | |
611 | sysctl_sched_stat_granularity)) { | |
dfdc119e | 612 | __enqueue_sleeper(cfs_rq, se); |
bf0f6f24 IM |
613 | se->delta_fair_sleep = 0; |
614 | } | |
615 | ||
616 | se->sleep_start_fair = 0; | |
617 | ||
618 | #ifdef CONFIG_SCHEDSTATS | |
619 | if (se->sleep_start) { | |
d281918d | 620 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
621 | |
622 | if ((s64)delta < 0) | |
623 | delta = 0; | |
624 | ||
625 | if (unlikely(delta > se->sleep_max)) | |
626 | se->sleep_max = delta; | |
627 | ||
628 | se->sleep_start = 0; | |
629 | se->sum_sleep_runtime += delta; | |
630 | } | |
631 | if (se->block_start) { | |
d281918d | 632 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
633 | |
634 | if ((s64)delta < 0) | |
635 | delta = 0; | |
636 | ||
637 | if (unlikely(delta > se->block_max)) | |
638 | se->block_max = delta; | |
639 | ||
640 | se->block_start = 0; | |
641 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
642 | |
643 | /* | |
644 | * Blocking time is in units of nanosecs, so shift by 20 to | |
645 | * get a milliseconds-range estimation of the amount of | |
646 | * time that the task spent sleeping: | |
647 | */ | |
648 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
649 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
650 | delta >> 20); | |
651 | } | |
bf0f6f24 IM |
652 | } |
653 | #endif | |
654 | } | |
655 | ||
656 | static void | |
668031ca | 657 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
658 | { |
659 | /* | |
660 | * Update the fair clock. | |
661 | */ | |
b7cc0896 | 662 | update_curr(cfs_rq); |
bf0f6f24 IM |
663 | |
664 | if (wakeup) | |
2396af69 | 665 | enqueue_sleeper(cfs_rq, se); |
bf0f6f24 | 666 | |
d2417e5a | 667 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
668 | __enqueue_entity(cfs_rq, se); |
669 | } | |
670 | ||
671 | static void | |
525c2716 | 672 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 673 | { |
19b6a2e3 | 674 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
675 | if (sleep) { |
676 | se->sleep_start_fair = cfs_rq->fair_clock; | |
677 | #ifdef CONFIG_SCHEDSTATS | |
678 | if (entity_is_task(se)) { | |
679 | struct task_struct *tsk = task_of(se); | |
680 | ||
681 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 682 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 683 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 684 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 685 | } |
bf0f6f24 IM |
686 | #endif |
687 | } | |
688 | __dequeue_entity(cfs_rq, se); | |
689 | } | |
690 | ||
691 | /* | |
692 | * Preempt the current task with a newly woken task if needed: | |
693 | */ | |
7c92e54f | 694 | static void |
bf0f6f24 IM |
695 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
696 | struct sched_entity *curr, unsigned long granularity) | |
697 | { | |
698 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
699 | unsigned long ideal_runtime, delta_exec; |
700 | ||
701 | /* | |
702 | * ideal_runtime is compared against sum_exec_runtime, which is | |
703 | * walltime, hence do not scale. | |
704 | */ | |
705 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
706 | (unsigned long)sysctl_sched_min_granularity); | |
707 | ||
708 | /* | |
709 | * If we executed more than what the latency constraint suggests, | |
710 | * reduce the rescheduling granularity. This way the total latency | |
711 | * of how much a task is not scheduled converges to | |
712 | * sysctl_sched_latency: | |
713 | */ | |
714 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
715 | if (delta_exec > ideal_runtime) | |
716 | granularity = 0; | |
bf0f6f24 IM |
717 | |
718 | /* | |
719 | * Take scheduling granularity into account - do not | |
720 | * preempt the current task unless the best task has | |
721 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
722 | * |
723 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 724 | */ |
4a55b450 | 725 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
726 | resched_task(rq_of(cfs_rq)->curr); |
727 | } | |
728 | ||
729 | static inline void | |
8494f412 | 730 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
731 | { |
732 | /* | |
733 | * Any task has to be enqueued before it get to execute on | |
734 | * a CPU. So account for the time it spent waiting on the | |
735 | * runqueue. (note, here we rely on pick_next_task() having | |
736 | * done a put_prev_task_fair() shortly before this, which | |
737 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
738 | */ | |
9ef0a961 | 739 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 740 | update_stats_curr_start(cfs_rq, se); |
bf0f6f24 | 741 | set_cfs_rq_curr(cfs_rq, se); |
4a55b450 | 742 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
743 | } |
744 | ||
9948f4b2 | 745 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
746 | { |
747 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
748 | ||
8494f412 | 749 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
750 | |
751 | return se; | |
752 | } | |
753 | ||
ab6cde26 | 754 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
755 | { |
756 | /* | |
757 | * If still on the runqueue then deactivate_task() | |
758 | * was not called and update_curr() has to be done: | |
759 | */ | |
760 | if (prev->on_rq) | |
b7cc0896 | 761 | update_curr(cfs_rq); |
bf0f6f24 | 762 | |
c7e9b5b2 | 763 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
764 | |
765 | if (prev->on_rq) | |
5870db5b | 766 | update_stats_wait_start(cfs_rq, prev); |
bf0f6f24 IM |
767 | set_cfs_rq_curr(cfs_rq, NULL); |
768 | } | |
769 | ||
770 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
771 | { | |
bf0f6f24 | 772 | struct sched_entity *next; |
c1b3da3e | 773 | |
bf0f6f24 IM |
774 | /* |
775 | * Dequeue and enqueue the task to update its | |
776 | * position within the tree: | |
777 | */ | |
525c2716 | 778 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 779 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
780 | |
781 | /* | |
782 | * Reschedule if another task tops the current one. | |
783 | */ | |
784 | next = __pick_next_entity(cfs_rq); | |
785 | if (next == curr) | |
786 | return; | |
787 | ||
11697830 PZ |
788 | __check_preempt_curr_fair(cfs_rq, next, curr, |
789 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
790 | } |
791 | ||
792 | /************************************************** | |
793 | * CFS operations on tasks: | |
794 | */ | |
795 | ||
796 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
797 | ||
798 | /* Walk up scheduling entities hierarchy */ | |
799 | #define for_each_sched_entity(se) \ | |
800 | for (; se; se = se->parent) | |
801 | ||
802 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
803 | { | |
804 | return p->se.cfs_rq; | |
805 | } | |
806 | ||
807 | /* runqueue on which this entity is (to be) queued */ | |
808 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
809 | { | |
810 | return se->cfs_rq; | |
811 | } | |
812 | ||
813 | /* runqueue "owned" by this group */ | |
814 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
815 | { | |
816 | return grp->my_q; | |
817 | } | |
818 | ||
819 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
820 | * another cpu ('this_cpu') | |
821 | */ | |
822 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
823 | { | |
824 | /* A later patch will take group into account */ | |
825 | return &cpu_rq(this_cpu)->cfs; | |
826 | } | |
827 | ||
828 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
829 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
830 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
831 | ||
832 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
833 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
834 | { | |
835 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
836 | return 1; | |
837 | ||
838 | return 0; | |
839 | } | |
840 | ||
841 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
842 | ||
843 | #define for_each_sched_entity(se) \ | |
844 | for (; se; se = NULL) | |
845 | ||
846 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
847 | { | |
848 | return &task_rq(p)->cfs; | |
849 | } | |
850 | ||
851 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
852 | { | |
853 | struct task_struct *p = task_of(se); | |
854 | struct rq *rq = task_rq(p); | |
855 | ||
856 | return &rq->cfs; | |
857 | } | |
858 | ||
859 | /* runqueue "owned" by this group */ | |
860 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
861 | { | |
862 | return NULL; | |
863 | } | |
864 | ||
865 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
866 | { | |
867 | return &cpu_rq(this_cpu)->cfs; | |
868 | } | |
869 | ||
870 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
871 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
872 | ||
873 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
874 | { | |
875 | return 1; | |
876 | } | |
877 | ||
878 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
879 | ||
880 | /* | |
881 | * The enqueue_task method is called before nr_running is | |
882 | * increased. Here we update the fair scheduling stats and | |
883 | * then put the task into the rbtree: | |
884 | */ | |
fd390f6a | 885 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
886 | { |
887 | struct cfs_rq *cfs_rq; | |
888 | struct sched_entity *se = &p->se; | |
889 | ||
890 | for_each_sched_entity(se) { | |
891 | if (se->on_rq) | |
892 | break; | |
893 | cfs_rq = cfs_rq_of(se); | |
668031ca | 894 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
895 | } |
896 | } | |
897 | ||
898 | /* | |
899 | * The dequeue_task method is called before nr_running is | |
900 | * decreased. We remove the task from the rbtree and | |
901 | * update the fair scheduling stats: | |
902 | */ | |
f02231e5 | 903 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
904 | { |
905 | struct cfs_rq *cfs_rq; | |
906 | struct sched_entity *se = &p->se; | |
907 | ||
908 | for_each_sched_entity(se) { | |
909 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 910 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
911 | /* Don't dequeue parent if it has other entities besides us */ |
912 | if (cfs_rq->load.weight) | |
913 | break; | |
914 | } | |
915 | } | |
916 | ||
917 | /* | |
1799e35d IM |
918 | * sched_yield() support is very simple - we dequeue and enqueue. |
919 | * | |
920 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
921 | */ |
922 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
923 | { | |
924 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
925 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
926 | struct sched_entity *rightmost, *se = &p->se; | |
927 | struct rb_node *parent; | |
bf0f6f24 IM |
928 | |
929 | /* | |
1799e35d IM |
930 | * Are we the only task in the tree? |
931 | */ | |
932 | if (unlikely(cfs_rq->nr_running == 1)) | |
933 | return; | |
934 | ||
935 | if (likely(!sysctl_sched_compat_yield)) { | |
936 | __update_rq_clock(rq); | |
937 | /* | |
938 | * Dequeue and enqueue the task to update its | |
939 | * position within the tree: | |
940 | */ | |
941 | dequeue_entity(cfs_rq, &p->se, 0); | |
942 | enqueue_entity(cfs_rq, &p->se, 0); | |
943 | ||
944 | return; | |
945 | } | |
946 | /* | |
947 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 948 | */ |
1799e35d IM |
949 | do { |
950 | parent = *link; | |
951 | link = &parent->rb_right; | |
952 | } while (*link); | |
953 | ||
954 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
955 | /* | |
956 | * Already in the rightmost position? | |
957 | */ | |
958 | if (unlikely(rightmost == se)) | |
959 | return; | |
960 | ||
961 | /* | |
962 | * Minimally necessary key value to be last in the tree: | |
963 | */ | |
964 | se->fair_key = rightmost->fair_key + 1; | |
965 | ||
966 | if (cfs_rq->rb_leftmost == &se->run_node) | |
967 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
968 | /* | |
969 | * Relink the task to the rightmost position: | |
970 | */ | |
971 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
972 | rb_link_node(&se->run_node, parent, link); | |
973 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
974 | } |
975 | ||
976 | /* | |
977 | * Preempt the current task with a newly woken task if needed: | |
978 | */ | |
979 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
980 | { | |
981 | struct task_struct *curr = rq->curr; | |
982 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
983 | unsigned long gran; | |
984 | ||
985 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 986 | update_rq_clock(rq); |
b7cc0896 | 987 | update_curr(cfs_rq); |
bf0f6f24 IM |
988 | resched_task(curr); |
989 | return; | |
990 | } | |
991 | ||
992 | gran = sysctl_sched_wakeup_granularity; | |
993 | /* | |
994 | * Batch tasks prefer throughput over latency: | |
995 | */ | |
996 | if (unlikely(p->policy == SCHED_BATCH)) | |
997 | gran = sysctl_sched_batch_wakeup_granularity; | |
998 | ||
999 | if (is_same_group(curr, p)) | |
1000 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
1001 | } | |
1002 | ||
fb8d4724 | 1003 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
1004 | { |
1005 | struct cfs_rq *cfs_rq = &rq->cfs; | |
1006 | struct sched_entity *se; | |
1007 | ||
1008 | if (unlikely(!cfs_rq->nr_running)) | |
1009 | return NULL; | |
1010 | ||
1011 | do { | |
9948f4b2 | 1012 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1013 | cfs_rq = group_cfs_rq(se); |
1014 | } while (cfs_rq); | |
1015 | ||
1016 | return task_of(se); | |
1017 | } | |
1018 | ||
1019 | /* | |
1020 | * Account for a descheduled task: | |
1021 | */ | |
31ee529c | 1022 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1023 | { |
1024 | struct sched_entity *se = &prev->se; | |
1025 | struct cfs_rq *cfs_rq; | |
1026 | ||
1027 | for_each_sched_entity(se) { | |
1028 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1029 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1030 | } |
1031 | } | |
1032 | ||
1033 | /************************************************** | |
1034 | * Fair scheduling class load-balancing methods: | |
1035 | */ | |
1036 | ||
1037 | /* | |
1038 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1039 | * during the whole iteration, the current task might be | |
1040 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1041 | * achieve that by always pre-iterating before returning | |
1042 | * the current task: | |
1043 | */ | |
1044 | static inline struct task_struct * | |
1045 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1046 | { | |
1047 | struct task_struct *p; | |
1048 | ||
1049 | if (!curr) | |
1050 | return NULL; | |
1051 | ||
1052 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1053 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1054 | ||
1055 | return p; | |
1056 | } | |
1057 | ||
1058 | static struct task_struct *load_balance_start_fair(void *arg) | |
1059 | { | |
1060 | struct cfs_rq *cfs_rq = arg; | |
1061 | ||
1062 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1063 | } | |
1064 | ||
1065 | static struct task_struct *load_balance_next_fair(void *arg) | |
1066 | { | |
1067 | struct cfs_rq *cfs_rq = arg; | |
1068 | ||
1069 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1070 | } | |
1071 | ||
a4ac01c3 | 1072 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1073 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1074 | { | |
1075 | struct sched_entity *curr; | |
1076 | struct task_struct *p; | |
1077 | ||
1078 | if (!cfs_rq->nr_running) | |
1079 | return MAX_PRIO; | |
1080 | ||
1081 | curr = __pick_next_entity(cfs_rq); | |
1082 | p = task_of(curr); | |
1083 | ||
1084 | return p->prio; | |
1085 | } | |
a4ac01c3 | 1086 | #endif |
bf0f6f24 | 1087 | |
43010659 | 1088 | static unsigned long |
bf0f6f24 | 1089 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1090 | unsigned long max_nr_move, unsigned long max_load_move, |
1091 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1092 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1093 | { |
1094 | struct cfs_rq *busy_cfs_rq; | |
1095 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1096 | long rem_load_move = max_load_move; | |
1097 | struct rq_iterator cfs_rq_iterator; | |
1098 | ||
1099 | cfs_rq_iterator.start = load_balance_start_fair; | |
1100 | cfs_rq_iterator.next = load_balance_next_fair; | |
1101 | ||
1102 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1103 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1104 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1105 | long imbalance; |
bf0f6f24 | 1106 | unsigned long maxload; |
bf0f6f24 IM |
1107 | |
1108 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1109 | ||
e56f31aa | 1110 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1111 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1112 | if (imbalance <= 0) | |
1113 | continue; | |
1114 | ||
1115 | /* Don't pull more than imbalance/2 */ | |
1116 | imbalance /= 2; | |
1117 | maxload = min(rem_load_move, imbalance); | |
1118 | ||
a4ac01c3 PW |
1119 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1120 | #else | |
e56f31aa | 1121 | # define maxload rem_load_move |
a4ac01c3 | 1122 | #endif |
bf0f6f24 IM |
1123 | /* pass busy_cfs_rq argument into |
1124 | * load_balance_[start|next]_fair iterators | |
1125 | */ | |
1126 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1127 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1128 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1129 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1130 | |
1131 | total_nr_moved += nr_moved; | |
1132 | max_nr_move -= nr_moved; | |
1133 | rem_load_move -= load_moved; | |
1134 | ||
1135 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1136 | break; | |
1137 | } | |
1138 | ||
43010659 | 1139 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1140 | } |
1141 | ||
1142 | /* | |
1143 | * scheduler tick hitting a task of our scheduling class: | |
1144 | */ | |
1145 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1146 | { | |
1147 | struct cfs_rq *cfs_rq; | |
1148 | struct sched_entity *se = &curr->se; | |
1149 | ||
1150 | for_each_sched_entity(se) { | |
1151 | cfs_rq = cfs_rq_of(se); | |
1152 | entity_tick(cfs_rq, se); | |
1153 | } | |
1154 | } | |
1155 | ||
1156 | /* | |
1157 | * Share the fairness runtime between parent and child, thus the | |
1158 | * total amount of pressure for CPU stays equal - new tasks | |
1159 | * get a chance to run but frequent forkers are not allowed to | |
1160 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1161 | * the child is not running yet. | |
1162 | */ | |
ee0827d8 | 1163 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1164 | { |
1165 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
7109c442 | 1166 | struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq); |
bf0f6f24 IM |
1167 | |
1168 | sched_info_queued(p); | |
1169 | ||
7109c442 | 1170 | update_curr(cfs_rq); |
d2417e5a | 1171 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1172 | /* |
1173 | * Child runs first: we let it run before the parent | |
1174 | * until it reschedules once. We set up the key so that | |
1175 | * it will preempt the parent: | |
1176 | */ | |
9f508f82 | 1177 | se->fair_key = curr->fair_key - |
7109c442 | 1178 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1179 | /* |
1180 | * The first wait is dominated by the child-runs-first logic, | |
1181 | * so do not credit it with that waiting time yet: | |
1182 | */ | |
1183 | if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL) | |
9f508f82 | 1184 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1185 | |
1186 | /* | |
1187 | * The statistical average of wait_runtime is about | |
1188 | * -granularity/2, so initialize the task with that: | |
1189 | */ | |
a206c072 | 1190 | if (sysctl_sched_features & SCHED_FEAT_START_DEBIT) |
9f508f82 | 1191 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 IM |
1192 | |
1193 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
1194 | } |
1195 | ||
1196 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1197 | /* Account for a task changing its policy or group. | |
1198 | * | |
1199 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1200 | * migrates between groups/classes. | |
1201 | */ | |
1202 | static void set_curr_task_fair(struct rq *rq) | |
1203 | { | |
7c6c16f3 | 1204 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1205 | |
c3b64f1e IM |
1206 | for_each_sched_entity(se) |
1207 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1208 | } |
1209 | #else | |
1210 | static void set_curr_task_fair(struct rq *rq) | |
1211 | { | |
1212 | } | |
1213 | #endif | |
1214 | ||
1215 | /* | |
1216 | * All the scheduling class methods: | |
1217 | */ | |
1218 | struct sched_class fair_sched_class __read_mostly = { | |
1219 | .enqueue_task = enqueue_task_fair, | |
1220 | .dequeue_task = dequeue_task_fair, | |
1221 | .yield_task = yield_task_fair, | |
1222 | ||
1223 | .check_preempt_curr = check_preempt_curr_fair, | |
1224 | ||
1225 | .pick_next_task = pick_next_task_fair, | |
1226 | .put_prev_task = put_prev_task_fair, | |
1227 | ||
1228 | .load_balance = load_balance_fair, | |
1229 | ||
1230 | .set_curr_task = set_curr_task_fair, | |
1231 | .task_tick = task_tick_fair, | |
1232 | .task_new = task_new_fair, | |
1233 | }; | |
1234 | ||
1235 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1236 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1237 | { |
bf0f6f24 IM |
1238 | struct cfs_rq *cfs_rq; |
1239 | ||
c3b64f1e | 1240 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1241 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1242 | } |
1243 | #endif |