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 | ||
2bd8e6d4 IM |
23 | /* |
24 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
25 | */ | |
26 | #ifdef CONFIG_SCHED_DEBUG | |
27 | # define const_debug __read_mostly | |
28 | #else | |
29 | # define const_debug static const | |
30 | #endif | |
31 | ||
bf0f6f24 | 32 | /* |
21805085 PZ |
33 | * Targeted preemption latency for CPU-bound tasks: |
34 | * (default: 20ms, units: nanoseconds) | |
bf0f6f24 | 35 | * |
21805085 PZ |
36 | * NOTE: this latency value is not the same as the concept of |
37 | * 'timeslice length' - timeslices in CFS are of variable length. | |
38 | * (to see the precise effective timeslice length of your workload, | |
39 | * run vmstat and monitor the context-switches field) | |
bf0f6f24 IM |
40 | * |
41 | * On SMP systems the value of this is multiplied by the log2 of the | |
42 | * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way | |
43 | * systems, 4x on 8-way systems, 5x on 16-way systems, etc.) | |
21805085 | 44 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 45 | */ |
2bd8e6d4 IM |
46 | const_debug unsigned int sysctl_sched_latency = 20000000ULL; |
47 | ||
48 | /* | |
49 | * After fork, child runs first. (default) If set to 0 then | |
50 | * parent will (try to) run first. | |
51 | */ | |
52 | const_debug unsigned int sysctl_sched_child_runs_first = 1; | |
21805085 PZ |
53 | |
54 | /* | |
55 | * Minimal preemption granularity for CPU-bound tasks: | |
56 | * (default: 2 msec, units: nanoseconds) | |
57 | */ | |
172ac3db | 58 | unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL; |
bf0f6f24 | 59 | |
1799e35d IM |
60 | /* |
61 | * sys_sched_yield() compat mode | |
62 | * | |
63 | * This option switches the agressive yield implementation of the | |
64 | * old scheduler back on. | |
65 | */ | |
66 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
67 | ||
bf0f6f24 IM |
68 | /* |
69 | * SCHED_BATCH wake-up granularity. | |
71fd3714 | 70 | * (default: 25 msec, units: nanoseconds) |
bf0f6f24 IM |
71 | * |
72 | * This option delays the preemption effects of decoupled workloads | |
73 | * and reduces their over-scheduling. Synchronous workloads will still | |
74 | * have immediate wakeup/sleep latencies. | |
75 | */ | |
2bd8e6d4 | 76 | const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 25000000UL; |
bf0f6f24 IM |
77 | |
78 | /* | |
79 | * SCHED_OTHER wake-up granularity. | |
80 | * (default: 1 msec, units: nanoseconds) | |
81 | * | |
82 | * This option delays the preemption effects of decoupled workloads | |
83 | * and reduces their over-scheduling. Synchronous workloads will still | |
84 | * have immediate wakeup/sleep latencies. | |
85 | */ | |
2bd8e6d4 | 86 | const_debug unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 87 | |
bf0f6f24 IM |
88 | unsigned int sysctl_sched_runtime_limit __read_mostly; |
89 | ||
90 | /* | |
91 | * Debugging: various feature bits | |
92 | */ | |
93 | enum { | |
94 | SCHED_FEAT_FAIR_SLEEPERS = 1, | |
95 | SCHED_FEAT_SLEEPER_AVG = 2, | |
96 | SCHED_FEAT_SLEEPER_LOAD_AVG = 4, | |
a25707f3 IM |
97 | SCHED_FEAT_START_DEBIT = 8, |
98 | SCHED_FEAT_SKIP_INITIAL = 16, | |
bf0f6f24 IM |
99 | }; |
100 | ||
2bd8e6d4 | 101 | const_debug unsigned int sysctl_sched_features = |
bf0f6f24 | 102 | SCHED_FEAT_FAIR_SLEEPERS *1 | |
5d2b3d36 | 103 | SCHED_FEAT_SLEEPER_AVG *0 | |
bf0f6f24 | 104 | SCHED_FEAT_SLEEPER_LOAD_AVG *1 | |
bf0f6f24 IM |
105 | SCHED_FEAT_START_DEBIT *1 | |
106 | SCHED_FEAT_SKIP_INITIAL *0; | |
107 | ||
e59c80c5 PZ |
108 | #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x) |
109 | ||
bf0f6f24 IM |
110 | extern struct sched_class fair_sched_class; |
111 | ||
112 | /************************************************************** | |
113 | * CFS operations on generic schedulable entities: | |
114 | */ | |
115 | ||
62160e3f | 116 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 117 | |
62160e3f | 118 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
119 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
120 | { | |
62160e3f | 121 | return cfs_rq->rq; |
bf0f6f24 IM |
122 | } |
123 | ||
62160e3f IM |
124 | /* An entity is a task if it doesn't "own" a runqueue */ |
125 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 126 | |
62160e3f | 127 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 128 | |
62160e3f IM |
129 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
130 | { | |
131 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
132 | } |
133 | ||
134 | #define entity_is_task(se) 1 | |
135 | ||
bf0f6f24 IM |
136 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
137 | ||
138 | static inline struct task_struct *task_of(struct sched_entity *se) | |
139 | { | |
140 | return container_of(se, struct task_struct, se); | |
141 | } | |
142 | ||
143 | ||
144 | /************************************************************** | |
145 | * Scheduling class tree data structure manipulation methods: | |
146 | */ | |
147 | ||
148 | /* | |
149 | * Enqueue an entity into the rb-tree: | |
150 | */ | |
19ccd97a | 151 | static void |
bf0f6f24 IM |
152 | __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
153 | { | |
154 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
155 | struct rb_node *parent = NULL; | |
156 | struct sched_entity *entry; | |
157 | s64 key = se->fair_key; | |
158 | int leftmost = 1; | |
159 | ||
160 | /* | |
161 | * Find the right place in the rbtree: | |
162 | */ | |
163 | while (*link) { | |
164 | parent = *link; | |
165 | entry = rb_entry(parent, struct sched_entity, run_node); | |
166 | /* | |
167 | * We dont care about collisions. Nodes with | |
168 | * the same key stay together. | |
169 | */ | |
170 | if (key - entry->fair_key < 0) { | |
171 | link = &parent->rb_left; | |
172 | } else { | |
173 | link = &parent->rb_right; | |
174 | leftmost = 0; | |
175 | } | |
176 | } | |
177 | ||
178 | /* | |
179 | * Maintain a cache of leftmost tree entries (it is frequently | |
180 | * used): | |
181 | */ | |
182 | if (leftmost) | |
183 | cfs_rq->rb_leftmost = &se->run_node; | |
184 | ||
185 | rb_link_node(&se->run_node, parent, link); | |
186 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
187 | update_load_add(&cfs_rq->load, se->load.weight); | |
188 | cfs_rq->nr_running++; | |
189 | se->on_rq = 1; | |
a206c072 IM |
190 | |
191 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
bf0f6f24 IM |
192 | } |
193 | ||
19ccd97a | 194 | static void |
bf0f6f24 IM |
195 | __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
196 | { | |
197 | if (cfs_rq->rb_leftmost == &se->run_node) | |
198 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
199 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
200 | update_load_sub(&cfs_rq->load, se->load.weight); | |
201 | cfs_rq->nr_running--; | |
202 | se->on_rq = 0; | |
a206c072 IM |
203 | |
204 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
bf0f6f24 IM |
205 | } |
206 | ||
207 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
208 | { | |
209 | return cfs_rq->rb_leftmost; | |
210 | } | |
211 | ||
212 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
213 | { | |
214 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
215 | } | |
216 | ||
217 | /************************************************************** | |
218 | * Scheduling class statistics methods: | |
219 | */ | |
220 | ||
21805085 PZ |
221 | /* |
222 | * Calculate the preemption granularity needed to schedule every | |
223 | * runnable task once per sysctl_sched_latency amount of time. | |
224 | * (down to a sensible low limit on granularity) | |
225 | * | |
226 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
227 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
228 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
229 | * for each task. We do finer and finer scheduling up to until we | |
230 | * reach the minimum granularity value. | |
231 | * | |
232 | * To achieve this we use the following dynamic-granularity rule: | |
233 | * | |
234 | * gran = lat/nr - lat/nr/nr | |
235 | * | |
236 | * This comes out of the following equations: | |
237 | * | |
238 | * kA1 + gran = kB1 | |
239 | * kB2 + gran = kA2 | |
240 | * kA2 = kA1 | |
241 | * kB2 = kB1 - d + d/nr | |
242 | * lat = d * nr | |
243 | * | |
244 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
245 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
246 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
247 | * running, 'lat' is the the period of each task. ('lat' is the | |
248 | * sched_latency that we aim for.) | |
249 | */ | |
250 | static long | |
251 | sched_granularity(struct cfs_rq *cfs_rq) | |
252 | { | |
253 | unsigned int gran = sysctl_sched_latency; | |
254 | unsigned int nr = cfs_rq->nr_running; | |
255 | ||
256 | if (nr > 1) { | |
257 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 258 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
259 | } |
260 | ||
261 | return gran; | |
262 | } | |
263 | ||
bf0f6f24 IM |
264 | /* |
265 | * We rescale the rescheduling granularity of tasks according to their | |
266 | * nice level, but only linearly, not exponentially: | |
267 | */ | |
268 | static long | |
269 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
270 | { | |
271 | u64 tmp; | |
272 | ||
7cff8cf6 IM |
273 | if (likely(curr->load.weight == NICE_0_LOAD)) |
274 | return granularity; | |
bf0f6f24 | 275 | /* |
7cff8cf6 | 276 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 277 | */ |
7cff8cf6 IM |
278 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
279 | tmp = curr->load.weight * (u64)granularity; | |
280 | return (long) (tmp >> NICE_0_SHIFT); | |
281 | } | |
bf0f6f24 | 282 | /* |
7cff8cf6 | 283 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
284 | * granularity: |
285 | */ | |
7cff8cf6 | 286 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
287 | |
288 | /* | |
289 | * It will always fit into 'long': | |
290 | */ | |
a0dc7260 | 291 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
292 | } |
293 | ||
294 | static inline void | |
295 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
296 | { | |
297 | long limit = sysctl_sched_runtime_limit; | |
298 | ||
299 | /* | |
300 | * Niced tasks have the same history dynamic range as | |
301 | * non-niced tasks: | |
302 | */ | |
303 | if (unlikely(se->wait_runtime > limit)) { | |
304 | se->wait_runtime = limit; | |
305 | schedstat_inc(se, wait_runtime_overruns); | |
306 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
307 | } | |
308 | if (unlikely(se->wait_runtime < -limit)) { | |
309 | se->wait_runtime = -limit; | |
310 | schedstat_inc(se, wait_runtime_underruns); | |
311 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
312 | } | |
313 | } | |
314 | ||
315 | static inline void | |
316 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
317 | { | |
318 | se->wait_runtime += delta; | |
319 | schedstat_add(se, sum_wait_runtime, delta); | |
320 | limit_wait_runtime(cfs_rq, se); | |
321 | } | |
322 | ||
323 | static void | |
324 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
325 | { | |
326 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
327 | __add_wait_runtime(cfs_rq, se, delta); | |
328 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
329 | } | |
330 | ||
331 | /* | |
332 | * Update the current task's runtime statistics. Skip current tasks that | |
333 | * are not in our scheduling class. | |
334 | */ | |
335 | static inline void | |
8ebc91d9 IM |
336 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
337 | unsigned long delta_exec) | |
bf0f6f24 | 338 | { |
8ebc91d9 | 339 | unsigned long delta, delta_fair, delta_mine; |
bf0f6f24 IM |
340 | struct load_weight *lw = &cfs_rq->load; |
341 | unsigned long load = lw->weight; | |
342 | ||
8179ca23 | 343 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
344 | |
345 | curr->sum_exec_runtime += delta_exec; | |
346 | cfs_rq->exec_clock += delta_exec; | |
347 | ||
fd8bb43e IM |
348 | if (unlikely(!load)) |
349 | return; | |
350 | ||
bf0f6f24 IM |
351 | delta_fair = calc_delta_fair(delta_exec, lw); |
352 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
353 | ||
5f01d519 | 354 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 355 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
356 | delta = min(delta, (unsigned long)( |
357 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
358 | cfs_rq->sleeper_bonus -= delta; |
359 | delta_mine -= delta; | |
360 | } | |
361 | ||
362 | cfs_rq->fair_clock += delta_fair; | |
363 | /* | |
364 | * We executed delta_exec amount of time on the CPU, | |
365 | * but we were only entitled to delta_mine amount of | |
366 | * time during that period (if nr_running == 1 then | |
367 | * the two values are equal) | |
368 | * [Note: delta_mine - delta_exec is negative]: | |
369 | */ | |
370 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
371 | } | |
372 | ||
b7cc0896 | 373 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 374 | { |
429d43bc | 375 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 376 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
377 | unsigned long delta_exec; |
378 | ||
379 | if (unlikely(!curr)) | |
380 | return; | |
381 | ||
382 | /* | |
383 | * Get the amount of time the current task was running | |
384 | * since the last time we changed load (this cannot | |
385 | * overflow on 32 bits): | |
386 | */ | |
8ebc91d9 | 387 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 388 | |
8ebc91d9 IM |
389 | __update_curr(cfs_rq, curr, delta_exec); |
390 | curr->exec_start = now; | |
bf0f6f24 IM |
391 | } |
392 | ||
393 | static inline void | |
5870db5b | 394 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
395 | { |
396 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 397 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
398 | } |
399 | ||
bf0f6f24 | 400 | static inline unsigned long |
08e2388a | 401 | calc_weighted(unsigned long delta, struct sched_entity *se) |
bf0f6f24 | 402 | { |
08e2388a | 403 | unsigned long weight = se->load.weight; |
bf0f6f24 | 404 | |
08e2388a IM |
405 | if (unlikely(weight != NICE_0_LOAD)) |
406 | return (u64)delta * se->load.weight >> NICE_0_SHIFT; | |
407 | else | |
408 | return delta; | |
bf0f6f24 | 409 | } |
bf0f6f24 IM |
410 | |
411 | /* | |
412 | * Task is being enqueued - update stats: | |
413 | */ | |
d2417e5a | 414 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
415 | { |
416 | s64 key; | |
417 | ||
418 | /* | |
419 | * Are we enqueueing a waiting task? (for current tasks | |
420 | * a dequeue/enqueue event is a NOP) | |
421 | */ | |
429d43bc | 422 | if (se != cfs_rq->curr) |
5870db5b | 423 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
424 | /* |
425 | * Update the key: | |
426 | */ | |
427 | key = cfs_rq->fair_clock; | |
428 | ||
429 | /* | |
430 | * Optimize the common nice 0 case: | |
431 | */ | |
432 | if (likely(se->load.weight == NICE_0_LOAD)) { | |
433 | key -= se->wait_runtime; | |
434 | } else { | |
435 | u64 tmp; | |
436 | ||
437 | if (se->wait_runtime < 0) { | |
438 | tmp = -se->wait_runtime; | |
439 | key += (tmp * se->load.inv_weight) >> | |
440 | (WMULT_SHIFT - NICE_0_SHIFT); | |
441 | } else { | |
442 | tmp = se->wait_runtime; | |
a69edb55 IM |
443 | key -= (tmp * se->load.inv_weight) >> |
444 | (WMULT_SHIFT - NICE_0_SHIFT); | |
bf0f6f24 IM |
445 | } |
446 | } | |
447 | ||
448 | se->fair_key = key; | |
449 | } | |
450 | ||
451 | /* | |
452 | * Note: must be called with a freshly updated rq->fair_clock. | |
453 | */ | |
454 | static inline void | |
8ebc91d9 IM |
455 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, |
456 | unsigned long delta_fair) | |
bf0f6f24 | 457 | { |
d281918d IM |
458 | schedstat_set(se->wait_max, max(se->wait_max, |
459 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 | 460 | |
08e2388a | 461 | delta_fair = calc_weighted(delta_fair, se); |
bf0f6f24 IM |
462 | |
463 | add_wait_runtime(cfs_rq, se, delta_fair); | |
464 | } | |
465 | ||
466 | static void | |
9ef0a961 | 467 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
468 | { |
469 | unsigned long delta_fair; | |
470 | ||
b77d69db IM |
471 | if (unlikely(!se->wait_start_fair)) |
472 | return; | |
473 | ||
bf0f6f24 IM |
474 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
475 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
476 | ||
8ebc91d9 | 477 | __update_stats_wait_end(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
478 | |
479 | se->wait_start_fair = 0; | |
6cfb0d5d | 480 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
481 | } |
482 | ||
483 | static inline void | |
19b6a2e3 | 484 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 485 | { |
b7cc0896 | 486 | update_curr(cfs_rq); |
bf0f6f24 IM |
487 | /* |
488 | * Mark the end of the wait period if dequeueing a | |
489 | * waiting task: | |
490 | */ | |
429d43bc | 491 | if (se != cfs_rq->curr) |
9ef0a961 | 492 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
493 | } |
494 | ||
495 | /* | |
496 | * We are picking a new current task - update its stats: | |
497 | */ | |
498 | static inline void | |
79303e9e | 499 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
500 | { |
501 | /* | |
502 | * We are starting a new run period: | |
503 | */ | |
d281918d | 504 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
505 | } |
506 | ||
507 | /* | |
508 | * We are descheduling a task - update its stats: | |
509 | */ | |
510 | static inline void | |
c7e9b5b2 | 511 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
512 | { |
513 | se->exec_start = 0; | |
514 | } | |
515 | ||
516 | /************************************************** | |
517 | * Scheduling class queueing methods: | |
518 | */ | |
519 | ||
8ebc91d9 IM |
520 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, |
521 | unsigned long delta_fair) | |
bf0f6f24 | 522 | { |
8ebc91d9 | 523 | unsigned long load = cfs_rq->load.weight; |
bf0f6f24 IM |
524 | long prev_runtime; |
525 | ||
b2133c8b IM |
526 | /* |
527 | * Do not boost sleepers if there's too much bonus 'in flight' | |
528 | * already: | |
529 | */ | |
530 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
531 | return; | |
532 | ||
e59c80c5 | 533 | if (sched_feat(SLEEPER_LOAD_AVG)) |
bf0f6f24 IM |
534 | load = rq_of(cfs_rq)->cpu_load[2]; |
535 | ||
bf0f6f24 IM |
536 | /* |
537 | * Fix up delta_fair with the effect of us running | |
538 | * during the whole sleep period: | |
539 | */ | |
e59c80c5 | 540 | if (sched_feat(SLEEPER_AVG)) |
bf0f6f24 IM |
541 | delta_fair = div64_likely32((u64)delta_fair * load, |
542 | load + se->load.weight); | |
543 | ||
08e2388a | 544 | delta_fair = calc_weighted(delta_fair, se); |
bf0f6f24 IM |
545 | |
546 | prev_runtime = se->wait_runtime; | |
547 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
548 | delta_fair = se->wait_runtime - prev_runtime; | |
549 | ||
550 | /* | |
551 | * Track the amount of bonus we've given to sleepers: | |
552 | */ | |
553 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
554 | } |
555 | ||
2396af69 | 556 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
557 | { |
558 | struct task_struct *tsk = task_of(se); | |
559 | unsigned long delta_fair; | |
560 | ||
561 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
e59c80c5 | 562 | !sched_feat(FAIR_SLEEPERS)) |
bf0f6f24 IM |
563 | return; |
564 | ||
565 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
566 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
567 | ||
8ebc91d9 | 568 | __enqueue_sleeper(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
569 | |
570 | se->sleep_start_fair = 0; | |
571 | ||
572 | #ifdef CONFIG_SCHEDSTATS | |
573 | if (se->sleep_start) { | |
d281918d | 574 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
575 | |
576 | if ((s64)delta < 0) | |
577 | delta = 0; | |
578 | ||
579 | if (unlikely(delta > se->sleep_max)) | |
580 | se->sleep_max = delta; | |
581 | ||
582 | se->sleep_start = 0; | |
583 | se->sum_sleep_runtime += delta; | |
584 | } | |
585 | if (se->block_start) { | |
d281918d | 586 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
587 | |
588 | if ((s64)delta < 0) | |
589 | delta = 0; | |
590 | ||
591 | if (unlikely(delta > se->block_max)) | |
592 | se->block_max = delta; | |
593 | ||
594 | se->block_start = 0; | |
595 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
596 | |
597 | /* | |
598 | * Blocking time is in units of nanosecs, so shift by 20 to | |
599 | * get a milliseconds-range estimation of the amount of | |
600 | * time that the task spent sleeping: | |
601 | */ | |
602 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
603 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
604 | delta >> 20); | |
605 | } | |
bf0f6f24 IM |
606 | } |
607 | #endif | |
608 | } | |
609 | ||
610 | static void | |
668031ca | 611 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
612 | { |
613 | /* | |
614 | * Update the fair clock. | |
615 | */ | |
b7cc0896 | 616 | update_curr(cfs_rq); |
bf0f6f24 IM |
617 | |
618 | if (wakeup) | |
2396af69 | 619 | enqueue_sleeper(cfs_rq, se); |
bf0f6f24 | 620 | |
d2417e5a | 621 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
622 | __enqueue_entity(cfs_rq, se); |
623 | } | |
624 | ||
625 | static void | |
525c2716 | 626 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 627 | { |
19b6a2e3 | 628 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
629 | if (sleep) { |
630 | se->sleep_start_fair = cfs_rq->fair_clock; | |
631 | #ifdef CONFIG_SCHEDSTATS | |
632 | if (entity_is_task(se)) { | |
633 | struct task_struct *tsk = task_of(se); | |
634 | ||
635 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 636 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 637 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 638 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 639 | } |
bf0f6f24 IM |
640 | #endif |
641 | } | |
642 | __dequeue_entity(cfs_rq, se); | |
643 | } | |
644 | ||
645 | /* | |
646 | * Preempt the current task with a newly woken task if needed: | |
647 | */ | |
7c92e54f | 648 | static void |
bf0f6f24 IM |
649 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
650 | struct sched_entity *curr, unsigned long granularity) | |
651 | { | |
652 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
653 | unsigned long ideal_runtime, delta_exec; |
654 | ||
655 | /* | |
656 | * ideal_runtime is compared against sum_exec_runtime, which is | |
657 | * walltime, hence do not scale. | |
658 | */ | |
659 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
660 | (unsigned long)sysctl_sched_min_granularity); | |
661 | ||
662 | /* | |
663 | * If we executed more than what the latency constraint suggests, | |
664 | * reduce the rescheduling granularity. This way the total latency | |
665 | * of how much a task is not scheduled converges to | |
666 | * sysctl_sched_latency: | |
667 | */ | |
668 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
669 | if (delta_exec > ideal_runtime) | |
670 | granularity = 0; | |
bf0f6f24 IM |
671 | |
672 | /* | |
673 | * Take scheduling granularity into account - do not | |
674 | * preempt the current task unless the best task has | |
675 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
676 | * |
677 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 678 | */ |
4a55b450 | 679 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
680 | resched_task(rq_of(cfs_rq)->curr); |
681 | } | |
682 | ||
683 | static inline void | |
8494f412 | 684 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
685 | { |
686 | /* | |
687 | * Any task has to be enqueued before it get to execute on | |
688 | * a CPU. So account for the time it spent waiting on the | |
689 | * runqueue. (note, here we rely on pick_next_task() having | |
690 | * done a put_prev_task_fair() shortly before this, which | |
691 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
692 | */ | |
9ef0a961 | 693 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 694 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 695 | cfs_rq->curr = se; |
eba1ed4b IM |
696 | #ifdef CONFIG_SCHEDSTATS |
697 | /* | |
698 | * Track our maximum slice length, if the CPU's load is at | |
699 | * least twice that of our own weight (i.e. dont track it | |
700 | * when there are only lesser-weight tasks around): | |
701 | */ | |
702 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
703 | se->slice_max = max(se->slice_max, | |
704 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
705 | } | |
706 | #endif | |
4a55b450 | 707 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
708 | } |
709 | ||
9948f4b2 | 710 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
711 | { |
712 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
713 | ||
8494f412 | 714 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
715 | |
716 | return se; | |
717 | } | |
718 | ||
ab6cde26 | 719 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
720 | { |
721 | /* | |
722 | * If still on the runqueue then deactivate_task() | |
723 | * was not called and update_curr() has to be done: | |
724 | */ | |
725 | if (prev->on_rq) | |
b7cc0896 | 726 | update_curr(cfs_rq); |
bf0f6f24 | 727 | |
c7e9b5b2 | 728 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
729 | |
730 | if (prev->on_rq) | |
5870db5b | 731 | update_stats_wait_start(cfs_rq, prev); |
429d43bc | 732 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
733 | } |
734 | ||
735 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
736 | { | |
bf0f6f24 | 737 | struct sched_entity *next; |
c1b3da3e | 738 | |
bf0f6f24 IM |
739 | /* |
740 | * Dequeue and enqueue the task to update its | |
741 | * position within the tree: | |
742 | */ | |
525c2716 | 743 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 744 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
745 | |
746 | /* | |
747 | * Reschedule if another task tops the current one. | |
748 | */ | |
749 | next = __pick_next_entity(cfs_rq); | |
750 | if (next == curr) | |
751 | return; | |
752 | ||
11697830 PZ |
753 | __check_preempt_curr_fair(cfs_rq, next, curr, |
754 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
755 | } |
756 | ||
757 | /************************************************** | |
758 | * CFS operations on tasks: | |
759 | */ | |
760 | ||
761 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
762 | ||
763 | /* Walk up scheduling entities hierarchy */ | |
764 | #define for_each_sched_entity(se) \ | |
765 | for (; se; se = se->parent) | |
766 | ||
767 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
768 | { | |
769 | return p->se.cfs_rq; | |
770 | } | |
771 | ||
772 | /* runqueue on which this entity is (to be) queued */ | |
773 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
774 | { | |
775 | return se->cfs_rq; | |
776 | } | |
777 | ||
778 | /* runqueue "owned" by this group */ | |
779 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
780 | { | |
781 | return grp->my_q; | |
782 | } | |
783 | ||
784 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
785 | * another cpu ('this_cpu') | |
786 | */ | |
787 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
788 | { | |
789 | /* A later patch will take group into account */ | |
790 | return &cpu_rq(this_cpu)->cfs; | |
791 | } | |
792 | ||
793 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
794 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
795 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
796 | ||
797 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
798 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
799 | { | |
800 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
801 | return 1; | |
802 | ||
803 | return 0; | |
804 | } | |
805 | ||
806 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
807 | ||
808 | #define for_each_sched_entity(se) \ | |
809 | for (; se; se = NULL) | |
810 | ||
811 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
812 | { | |
813 | return &task_rq(p)->cfs; | |
814 | } | |
815 | ||
816 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
817 | { | |
818 | struct task_struct *p = task_of(se); | |
819 | struct rq *rq = task_rq(p); | |
820 | ||
821 | return &rq->cfs; | |
822 | } | |
823 | ||
824 | /* runqueue "owned" by this group */ | |
825 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
826 | { | |
827 | return NULL; | |
828 | } | |
829 | ||
830 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
831 | { | |
832 | return &cpu_rq(this_cpu)->cfs; | |
833 | } | |
834 | ||
835 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
836 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
837 | ||
838 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
839 | { | |
840 | return 1; | |
841 | } | |
842 | ||
843 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
844 | ||
845 | /* | |
846 | * The enqueue_task method is called before nr_running is | |
847 | * increased. Here we update the fair scheduling stats and | |
848 | * then put the task into the rbtree: | |
849 | */ | |
fd390f6a | 850 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
851 | { |
852 | struct cfs_rq *cfs_rq; | |
853 | struct sched_entity *se = &p->se; | |
854 | ||
855 | for_each_sched_entity(se) { | |
856 | if (se->on_rq) | |
857 | break; | |
858 | cfs_rq = cfs_rq_of(se); | |
668031ca | 859 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
860 | } |
861 | } | |
862 | ||
863 | /* | |
864 | * The dequeue_task method is called before nr_running is | |
865 | * decreased. We remove the task from the rbtree and | |
866 | * update the fair scheduling stats: | |
867 | */ | |
f02231e5 | 868 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
869 | { |
870 | struct cfs_rq *cfs_rq; | |
871 | struct sched_entity *se = &p->se; | |
872 | ||
873 | for_each_sched_entity(se) { | |
874 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 875 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
876 | /* Don't dequeue parent if it has other entities besides us */ |
877 | if (cfs_rq->load.weight) | |
878 | break; | |
879 | } | |
880 | } | |
881 | ||
882 | /* | |
1799e35d IM |
883 | * sched_yield() support is very simple - we dequeue and enqueue. |
884 | * | |
885 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
886 | */ |
887 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
888 | { | |
889 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
890 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
891 | struct sched_entity *rightmost, *se = &p->se; | |
892 | struct rb_node *parent; | |
bf0f6f24 IM |
893 | |
894 | /* | |
1799e35d IM |
895 | * Are we the only task in the tree? |
896 | */ | |
897 | if (unlikely(cfs_rq->nr_running == 1)) | |
898 | return; | |
899 | ||
900 | if (likely(!sysctl_sched_compat_yield)) { | |
901 | __update_rq_clock(rq); | |
902 | /* | |
903 | * Dequeue and enqueue the task to update its | |
904 | * position within the tree: | |
905 | */ | |
906 | dequeue_entity(cfs_rq, &p->se, 0); | |
907 | enqueue_entity(cfs_rq, &p->se, 0); | |
908 | ||
909 | return; | |
910 | } | |
911 | /* | |
912 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 913 | */ |
1799e35d IM |
914 | do { |
915 | parent = *link; | |
916 | link = &parent->rb_right; | |
917 | } while (*link); | |
918 | ||
919 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
920 | /* | |
921 | * Already in the rightmost position? | |
922 | */ | |
923 | if (unlikely(rightmost == se)) | |
924 | return; | |
925 | ||
926 | /* | |
927 | * Minimally necessary key value to be last in the tree: | |
928 | */ | |
929 | se->fair_key = rightmost->fair_key + 1; | |
930 | ||
931 | if (cfs_rq->rb_leftmost == &se->run_node) | |
932 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
933 | /* | |
934 | * Relink the task to the rightmost position: | |
935 | */ | |
936 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
937 | rb_link_node(&se->run_node, parent, link); | |
938 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
939 | } |
940 | ||
941 | /* | |
942 | * Preempt the current task with a newly woken task if needed: | |
943 | */ | |
944 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
945 | { | |
946 | struct task_struct *curr = rq->curr; | |
947 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
948 | unsigned long gran; | |
949 | ||
950 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 951 | update_rq_clock(rq); |
b7cc0896 | 952 | update_curr(cfs_rq); |
bf0f6f24 IM |
953 | resched_task(curr); |
954 | return; | |
955 | } | |
956 | ||
957 | gran = sysctl_sched_wakeup_granularity; | |
958 | /* | |
959 | * Batch tasks prefer throughput over latency: | |
960 | */ | |
961 | if (unlikely(p->policy == SCHED_BATCH)) | |
962 | gran = sysctl_sched_batch_wakeup_granularity; | |
963 | ||
964 | if (is_same_group(curr, p)) | |
965 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
966 | } | |
967 | ||
fb8d4724 | 968 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
969 | { |
970 | struct cfs_rq *cfs_rq = &rq->cfs; | |
971 | struct sched_entity *se; | |
972 | ||
973 | if (unlikely(!cfs_rq->nr_running)) | |
974 | return NULL; | |
975 | ||
976 | do { | |
9948f4b2 | 977 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
978 | cfs_rq = group_cfs_rq(se); |
979 | } while (cfs_rq); | |
980 | ||
981 | return task_of(se); | |
982 | } | |
983 | ||
984 | /* | |
985 | * Account for a descheduled task: | |
986 | */ | |
31ee529c | 987 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
988 | { |
989 | struct sched_entity *se = &prev->se; | |
990 | struct cfs_rq *cfs_rq; | |
991 | ||
992 | for_each_sched_entity(se) { | |
993 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 994 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
995 | } |
996 | } | |
997 | ||
998 | /************************************************** | |
999 | * Fair scheduling class load-balancing methods: | |
1000 | */ | |
1001 | ||
1002 | /* | |
1003 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1004 | * during the whole iteration, the current task might be | |
1005 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1006 | * achieve that by always pre-iterating before returning | |
1007 | * the current task: | |
1008 | */ | |
1009 | static inline struct task_struct * | |
1010 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1011 | { | |
1012 | struct task_struct *p; | |
1013 | ||
1014 | if (!curr) | |
1015 | return NULL; | |
1016 | ||
1017 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1018 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1019 | ||
1020 | return p; | |
1021 | } | |
1022 | ||
1023 | static struct task_struct *load_balance_start_fair(void *arg) | |
1024 | { | |
1025 | struct cfs_rq *cfs_rq = arg; | |
1026 | ||
1027 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1028 | } | |
1029 | ||
1030 | static struct task_struct *load_balance_next_fair(void *arg) | |
1031 | { | |
1032 | struct cfs_rq *cfs_rq = arg; | |
1033 | ||
1034 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1035 | } | |
1036 | ||
a4ac01c3 | 1037 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1038 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1039 | { | |
1040 | struct sched_entity *curr; | |
1041 | struct task_struct *p; | |
1042 | ||
1043 | if (!cfs_rq->nr_running) | |
1044 | return MAX_PRIO; | |
1045 | ||
1046 | curr = __pick_next_entity(cfs_rq); | |
1047 | p = task_of(curr); | |
1048 | ||
1049 | return p->prio; | |
1050 | } | |
a4ac01c3 | 1051 | #endif |
bf0f6f24 | 1052 | |
43010659 | 1053 | static unsigned long |
bf0f6f24 | 1054 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1055 | unsigned long max_nr_move, unsigned long max_load_move, |
1056 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1057 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1058 | { |
1059 | struct cfs_rq *busy_cfs_rq; | |
1060 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1061 | long rem_load_move = max_load_move; | |
1062 | struct rq_iterator cfs_rq_iterator; | |
1063 | ||
1064 | cfs_rq_iterator.start = load_balance_start_fair; | |
1065 | cfs_rq_iterator.next = load_balance_next_fair; | |
1066 | ||
1067 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1068 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1069 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1070 | long imbalance; |
bf0f6f24 | 1071 | unsigned long maxload; |
bf0f6f24 IM |
1072 | |
1073 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1074 | ||
e56f31aa | 1075 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1076 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1077 | if (imbalance <= 0) | |
1078 | continue; | |
1079 | ||
1080 | /* Don't pull more than imbalance/2 */ | |
1081 | imbalance /= 2; | |
1082 | maxload = min(rem_load_move, imbalance); | |
1083 | ||
a4ac01c3 PW |
1084 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1085 | #else | |
e56f31aa | 1086 | # define maxload rem_load_move |
a4ac01c3 | 1087 | #endif |
bf0f6f24 IM |
1088 | /* pass busy_cfs_rq argument into |
1089 | * load_balance_[start|next]_fair iterators | |
1090 | */ | |
1091 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1092 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1093 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1094 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1095 | |
1096 | total_nr_moved += nr_moved; | |
1097 | max_nr_move -= nr_moved; | |
1098 | rem_load_move -= load_moved; | |
1099 | ||
1100 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1101 | break; | |
1102 | } | |
1103 | ||
43010659 | 1104 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1105 | } |
1106 | ||
1107 | /* | |
1108 | * scheduler tick hitting a task of our scheduling class: | |
1109 | */ | |
1110 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1111 | { | |
1112 | struct cfs_rq *cfs_rq; | |
1113 | struct sched_entity *se = &curr->se; | |
1114 | ||
1115 | for_each_sched_entity(se) { | |
1116 | cfs_rq = cfs_rq_of(se); | |
1117 | entity_tick(cfs_rq, se); | |
1118 | } | |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * Share the fairness runtime between parent and child, thus the | |
1123 | * total amount of pressure for CPU stays equal - new tasks | |
1124 | * get a chance to run but frequent forkers are not allowed to | |
1125 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1126 | * the child is not running yet. | |
1127 | */ | |
ee0827d8 | 1128 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1129 | { |
1130 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1131 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
bf0f6f24 IM |
1132 | |
1133 | sched_info_queued(p); | |
1134 | ||
7109c442 | 1135 | update_curr(cfs_rq); |
d2417e5a | 1136 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1137 | /* |
1138 | * Child runs first: we let it run before the parent | |
1139 | * until it reschedules once. We set up the key so that | |
1140 | * it will preempt the parent: | |
1141 | */ | |
9f508f82 | 1142 | se->fair_key = curr->fair_key - |
7109c442 | 1143 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1144 | /* |
1145 | * The first wait is dominated by the child-runs-first logic, | |
1146 | * so do not credit it with that waiting time yet: | |
1147 | */ | |
e59c80c5 | 1148 | if (sched_feat(SKIP_INITIAL)) |
9f508f82 | 1149 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1150 | |
1151 | /* | |
1152 | * The statistical average of wait_runtime is about | |
1153 | * -granularity/2, so initialize the task with that: | |
1154 | */ | |
e59c80c5 | 1155 | if (sched_feat(START_DEBIT)) |
9f508f82 | 1156 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 IM |
1157 | |
1158 | __enqueue_entity(cfs_rq, se); | |
bb61c210 | 1159 | resched_task(rq->curr); |
bf0f6f24 IM |
1160 | } |
1161 | ||
1162 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1163 | /* Account for a task changing its policy or group. | |
1164 | * | |
1165 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1166 | * migrates between groups/classes. | |
1167 | */ | |
1168 | static void set_curr_task_fair(struct rq *rq) | |
1169 | { | |
7c6c16f3 | 1170 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1171 | |
c3b64f1e IM |
1172 | for_each_sched_entity(se) |
1173 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1174 | } |
1175 | #else | |
1176 | static void set_curr_task_fair(struct rq *rq) | |
1177 | { | |
1178 | } | |
1179 | #endif | |
1180 | ||
1181 | /* | |
1182 | * All the scheduling class methods: | |
1183 | */ | |
1184 | struct sched_class fair_sched_class __read_mostly = { | |
1185 | .enqueue_task = enqueue_task_fair, | |
1186 | .dequeue_task = dequeue_task_fair, | |
1187 | .yield_task = yield_task_fair, | |
1188 | ||
1189 | .check_preempt_curr = check_preempt_curr_fair, | |
1190 | ||
1191 | .pick_next_task = pick_next_task_fair, | |
1192 | .put_prev_task = put_prev_task_fair, | |
1193 | ||
1194 | .load_balance = load_balance_fair, | |
1195 | ||
1196 | .set_curr_task = set_curr_task_fair, | |
1197 | .task_tick = task_tick_fair, | |
1198 | .task_new = task_new_fair, | |
1199 | }; | |
1200 | ||
1201 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1202 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1203 | { |
bf0f6f24 IM |
1204 | struct cfs_rq *cfs_rq; |
1205 | ||
c3b64f1e | 1206 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1207 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1208 | } |
1209 | #endif |