]>
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 | */ |
2bd8e6d4 IM |
37 | const_debug unsigned int sysctl_sched_latency = 20000000ULL; |
38 | ||
39 | /* | |
40 | * After fork, child runs first. (default) If set to 0 then | |
41 | * parent will (try to) run first. | |
42 | */ | |
43 | const_debug unsigned int sysctl_sched_child_runs_first = 1; | |
21805085 PZ |
44 | |
45 | /* | |
46 | * Minimal preemption granularity for CPU-bound tasks: | |
47 | * (default: 2 msec, units: nanoseconds) | |
48 | */ | |
172ac3db | 49 | unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL; |
bf0f6f24 | 50 | |
1799e35d IM |
51 | /* |
52 | * sys_sched_yield() compat mode | |
53 | * | |
54 | * This option switches the agressive yield implementation of the | |
55 | * old scheduler back on. | |
56 | */ | |
57 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
58 | ||
bf0f6f24 IM |
59 | /* |
60 | * SCHED_BATCH wake-up granularity. | |
71fd3714 | 61 | * (default: 25 msec, units: nanoseconds) |
bf0f6f24 IM |
62 | * |
63 | * This option delays the preemption effects of decoupled workloads | |
64 | * and reduces their over-scheduling. Synchronous workloads will still | |
65 | * have immediate wakeup/sleep latencies. | |
66 | */ | |
2bd8e6d4 | 67 | const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 25000000UL; |
bf0f6f24 IM |
68 | |
69 | /* | |
70 | * SCHED_OTHER wake-up granularity. | |
71 | * (default: 1 msec, units: nanoseconds) | |
72 | * | |
73 | * This option delays the preemption effects of decoupled workloads | |
74 | * and reduces their over-scheduling. Synchronous workloads will still | |
75 | * have immediate wakeup/sleep latencies. | |
76 | */ | |
2bd8e6d4 | 77 | const_debug unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 78 | |
bf0f6f24 IM |
79 | unsigned int sysctl_sched_runtime_limit __read_mostly; |
80 | ||
bf0f6f24 IM |
81 | extern struct sched_class fair_sched_class; |
82 | ||
83 | /************************************************************** | |
84 | * CFS operations on generic schedulable entities: | |
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 | |
62160e3f | 98 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 99 | |
62160e3f IM |
100 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
101 | { | |
102 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
103 | } |
104 | ||
105 | #define entity_is_task(se) 1 | |
106 | ||
bf0f6f24 IM |
107 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
108 | ||
109 | static inline struct task_struct *task_of(struct sched_entity *se) | |
110 | { | |
111 | return container_of(se, struct task_struct, se); | |
112 | } | |
113 | ||
114 | ||
115 | /************************************************************** | |
116 | * Scheduling class tree data structure manipulation methods: | |
117 | */ | |
118 | ||
e9acbff6 IM |
119 | static inline void |
120 | set_leftmost(struct cfs_rq *cfs_rq, struct rb_node *leftmost) | |
121 | { | |
122 | struct sched_entity *se; | |
123 | ||
124 | cfs_rq->rb_leftmost = leftmost; | |
125 | if (leftmost) { | |
126 | se = rb_entry(leftmost, struct sched_entity, run_node); | |
127 | cfs_rq->min_vruntime = max(se->vruntime, | |
128 | cfs_rq->min_vruntime); | |
129 | } | |
130 | } | |
131 | ||
bf0f6f24 IM |
132 | /* |
133 | * Enqueue an entity into the rb-tree: | |
134 | */ | |
19ccd97a | 135 | static void |
bf0f6f24 IM |
136 | __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
137 | { | |
138 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
139 | struct rb_node *parent = NULL; | |
140 | struct sched_entity *entry; | |
141 | s64 key = se->fair_key; | |
142 | int leftmost = 1; | |
143 | ||
144 | /* | |
145 | * Find the right place in the rbtree: | |
146 | */ | |
147 | while (*link) { | |
148 | parent = *link; | |
149 | entry = rb_entry(parent, struct sched_entity, run_node); | |
150 | /* | |
151 | * We dont care about collisions. Nodes with | |
152 | * the same key stay together. | |
153 | */ | |
154 | if (key - entry->fair_key < 0) { | |
155 | link = &parent->rb_left; | |
156 | } else { | |
157 | link = &parent->rb_right; | |
158 | leftmost = 0; | |
159 | } | |
160 | } | |
161 | ||
162 | /* | |
163 | * Maintain a cache of leftmost tree entries (it is frequently | |
164 | * used): | |
165 | */ | |
166 | if (leftmost) | |
e9acbff6 | 167 | set_leftmost(cfs_rq, &se->run_node); |
bf0f6f24 IM |
168 | |
169 | rb_link_node(&se->run_node, parent, link); | |
170 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
171 | update_load_add(&cfs_rq->load, se->load.weight); | |
172 | cfs_rq->nr_running++; | |
173 | se->on_rq = 1; | |
a206c072 IM |
174 | |
175 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
bf0f6f24 IM |
176 | } |
177 | ||
19ccd97a | 178 | static void |
bf0f6f24 IM |
179 | __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
180 | { | |
181 | if (cfs_rq->rb_leftmost == &se->run_node) | |
e9acbff6 IM |
182 | set_leftmost(cfs_rq, rb_next(&se->run_node)); |
183 | ||
bf0f6f24 IM |
184 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
185 | update_load_sub(&cfs_rq->load, se->load.weight); | |
186 | cfs_rq->nr_running--; | |
187 | se->on_rq = 0; | |
a206c072 IM |
188 | |
189 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
bf0f6f24 IM |
190 | } |
191 | ||
192 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
193 | { | |
194 | return cfs_rq->rb_leftmost; | |
195 | } | |
196 | ||
197 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
198 | { | |
199 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
200 | } | |
201 | ||
202 | /************************************************************** | |
203 | * Scheduling class statistics methods: | |
204 | */ | |
205 | ||
21805085 PZ |
206 | /* |
207 | * Calculate the preemption granularity needed to schedule every | |
208 | * runnable task once per sysctl_sched_latency amount of time. | |
209 | * (down to a sensible low limit on granularity) | |
210 | * | |
211 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
212 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
213 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
214 | * for each task. We do finer and finer scheduling up to until we | |
215 | * reach the minimum granularity value. | |
216 | * | |
217 | * To achieve this we use the following dynamic-granularity rule: | |
218 | * | |
219 | * gran = lat/nr - lat/nr/nr | |
220 | * | |
221 | * This comes out of the following equations: | |
222 | * | |
223 | * kA1 + gran = kB1 | |
224 | * kB2 + gran = kA2 | |
225 | * kA2 = kA1 | |
226 | * kB2 = kB1 - d + d/nr | |
227 | * lat = d * nr | |
228 | * | |
229 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
230 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
231 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
232 | * running, 'lat' is the the period of each task. ('lat' is the | |
233 | * sched_latency that we aim for.) | |
234 | */ | |
235 | static long | |
236 | sched_granularity(struct cfs_rq *cfs_rq) | |
237 | { | |
238 | unsigned int gran = sysctl_sched_latency; | |
239 | unsigned int nr = cfs_rq->nr_running; | |
240 | ||
241 | if (nr > 1) { | |
242 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 243 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
244 | } |
245 | ||
246 | return gran; | |
247 | } | |
248 | ||
bf0f6f24 IM |
249 | /* |
250 | * We rescale the rescheduling granularity of tasks according to their | |
251 | * nice level, but only linearly, not exponentially: | |
252 | */ | |
253 | static long | |
254 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
255 | { | |
256 | u64 tmp; | |
257 | ||
7cff8cf6 IM |
258 | if (likely(curr->load.weight == NICE_0_LOAD)) |
259 | return granularity; | |
bf0f6f24 | 260 | /* |
7cff8cf6 | 261 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 262 | */ |
7cff8cf6 IM |
263 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
264 | tmp = curr->load.weight * (u64)granularity; | |
265 | return (long) (tmp >> NICE_0_SHIFT); | |
266 | } | |
bf0f6f24 | 267 | /* |
7cff8cf6 | 268 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
269 | * granularity: |
270 | */ | |
7cff8cf6 | 271 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
272 | |
273 | /* | |
274 | * It will always fit into 'long': | |
275 | */ | |
a0dc7260 | 276 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
277 | } |
278 | ||
279 | static inline void | |
280 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
281 | { | |
282 | long limit = sysctl_sched_runtime_limit; | |
283 | ||
284 | /* | |
285 | * Niced tasks have the same history dynamic range as | |
286 | * non-niced tasks: | |
287 | */ | |
288 | if (unlikely(se->wait_runtime > limit)) { | |
289 | se->wait_runtime = limit; | |
290 | schedstat_inc(se, wait_runtime_overruns); | |
291 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
292 | } | |
293 | if (unlikely(se->wait_runtime < -limit)) { | |
294 | se->wait_runtime = -limit; | |
295 | schedstat_inc(se, wait_runtime_underruns); | |
296 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
297 | } | |
298 | } | |
299 | ||
300 | static inline void | |
301 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
302 | { | |
303 | se->wait_runtime += delta; | |
304 | schedstat_add(se, sum_wait_runtime, delta); | |
305 | limit_wait_runtime(cfs_rq, se); | |
306 | } | |
307 | ||
308 | static void | |
309 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
310 | { | |
311 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
312 | __add_wait_runtime(cfs_rq, se, delta); | |
313 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
314 | } | |
315 | ||
316 | /* | |
317 | * Update the current task's runtime statistics. Skip current tasks that | |
318 | * are not in our scheduling class. | |
319 | */ | |
320 | static inline void | |
8ebc91d9 IM |
321 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
322 | unsigned long delta_exec) | |
bf0f6f24 | 323 | { |
e9acbff6 | 324 | unsigned long delta, delta_fair, delta_mine, delta_exec_weighted; |
bf0f6f24 IM |
325 | struct load_weight *lw = &cfs_rq->load; |
326 | unsigned long load = lw->weight; | |
327 | ||
8179ca23 | 328 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
329 | |
330 | curr->sum_exec_runtime += delta_exec; | |
331 | cfs_rq->exec_clock += delta_exec; | |
e9acbff6 IM |
332 | delta_exec_weighted = delta_exec; |
333 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { | |
334 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, | |
335 | &curr->load); | |
336 | } | |
337 | curr->vruntime += delta_exec_weighted; | |
bf0f6f24 | 338 | |
6cb58195 IM |
339 | if (!sched_feat(FAIR_SLEEPERS)) |
340 | return; | |
341 | ||
fd8bb43e IM |
342 | if (unlikely(!load)) |
343 | return; | |
344 | ||
bf0f6f24 IM |
345 | delta_fair = calc_delta_fair(delta_exec, lw); |
346 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
347 | ||
5f01d519 | 348 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 349 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
350 | delta = min(delta, (unsigned long)( |
351 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
352 | cfs_rq->sleeper_bonus -= delta; |
353 | delta_mine -= delta; | |
354 | } | |
355 | ||
356 | cfs_rq->fair_clock += delta_fair; | |
357 | /* | |
358 | * We executed delta_exec amount of time on the CPU, | |
359 | * but we were only entitled to delta_mine amount of | |
360 | * time during that period (if nr_running == 1 then | |
361 | * the two values are equal) | |
362 | * [Note: delta_mine - delta_exec is negative]: | |
363 | */ | |
364 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
365 | } | |
366 | ||
b7cc0896 | 367 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 368 | { |
429d43bc | 369 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 370 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
371 | unsigned long delta_exec; |
372 | ||
373 | if (unlikely(!curr)) | |
374 | return; | |
375 | ||
376 | /* | |
377 | * Get the amount of time the current task was running | |
378 | * since the last time we changed load (this cannot | |
379 | * overflow on 32 bits): | |
380 | */ | |
8ebc91d9 | 381 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 382 | |
8ebc91d9 IM |
383 | __update_curr(cfs_rq, curr, delta_exec); |
384 | curr->exec_start = now; | |
bf0f6f24 IM |
385 | } |
386 | ||
387 | static inline void | |
5870db5b | 388 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
389 | { |
390 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 391 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
392 | } |
393 | ||
bf0f6f24 | 394 | static inline unsigned long |
08e2388a | 395 | calc_weighted(unsigned long delta, struct sched_entity *se) |
bf0f6f24 | 396 | { |
08e2388a | 397 | unsigned long weight = se->load.weight; |
bf0f6f24 | 398 | |
08e2388a IM |
399 | if (unlikely(weight != NICE_0_LOAD)) |
400 | return (u64)delta * se->load.weight >> NICE_0_SHIFT; | |
401 | else | |
402 | return delta; | |
bf0f6f24 | 403 | } |
bf0f6f24 IM |
404 | |
405 | /* | |
406 | * Task is being enqueued - update stats: | |
407 | */ | |
d2417e5a | 408 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 409 | { |
bf0f6f24 IM |
410 | /* |
411 | * Are we enqueueing a waiting task? (for current tasks | |
412 | * a dequeue/enqueue event is a NOP) | |
413 | */ | |
429d43bc | 414 | if (se != cfs_rq->curr) |
5870db5b | 415 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
416 | /* |
417 | * Update the key: | |
418 | */ | |
e9acbff6 | 419 | se->fair_key = se->vruntime; |
bf0f6f24 IM |
420 | } |
421 | ||
422 | /* | |
423 | * Note: must be called with a freshly updated rq->fair_clock. | |
424 | */ | |
425 | static inline void | |
8ebc91d9 IM |
426 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, |
427 | unsigned long delta_fair) | |
bf0f6f24 | 428 | { |
d281918d IM |
429 | schedstat_set(se->wait_max, max(se->wait_max, |
430 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 | 431 | |
08e2388a | 432 | delta_fair = calc_weighted(delta_fair, se); |
bf0f6f24 IM |
433 | |
434 | add_wait_runtime(cfs_rq, se, delta_fair); | |
435 | } | |
436 | ||
437 | static void | |
9ef0a961 | 438 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
439 | { |
440 | unsigned long delta_fair; | |
441 | ||
b77d69db IM |
442 | if (unlikely(!se->wait_start_fair)) |
443 | return; | |
444 | ||
bf0f6f24 IM |
445 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
446 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
447 | ||
8ebc91d9 | 448 | __update_stats_wait_end(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
449 | |
450 | se->wait_start_fair = 0; | |
6cfb0d5d | 451 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
452 | } |
453 | ||
454 | static inline void | |
19b6a2e3 | 455 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 456 | { |
b7cc0896 | 457 | update_curr(cfs_rq); |
bf0f6f24 IM |
458 | /* |
459 | * Mark the end of the wait period if dequeueing a | |
460 | * waiting task: | |
461 | */ | |
429d43bc | 462 | if (se != cfs_rq->curr) |
9ef0a961 | 463 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
464 | } |
465 | ||
466 | /* | |
467 | * We are picking a new current task - update its stats: | |
468 | */ | |
469 | static inline void | |
79303e9e | 470 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
471 | { |
472 | /* | |
473 | * We are starting a new run period: | |
474 | */ | |
d281918d | 475 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
476 | } |
477 | ||
478 | /* | |
479 | * We are descheduling a task - update its stats: | |
480 | */ | |
481 | static inline void | |
c7e9b5b2 | 482 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
483 | { |
484 | se->exec_start = 0; | |
485 | } | |
486 | ||
487 | /************************************************** | |
488 | * Scheduling class queueing methods: | |
489 | */ | |
490 | ||
8ebc91d9 IM |
491 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, |
492 | unsigned long delta_fair) | |
bf0f6f24 | 493 | { |
8ebc91d9 | 494 | unsigned long load = cfs_rq->load.weight; |
bf0f6f24 IM |
495 | long prev_runtime; |
496 | ||
b2133c8b IM |
497 | /* |
498 | * Do not boost sleepers if there's too much bonus 'in flight' | |
499 | * already: | |
500 | */ | |
501 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
502 | return; | |
503 | ||
e59c80c5 | 504 | if (sched_feat(SLEEPER_LOAD_AVG)) |
bf0f6f24 IM |
505 | load = rq_of(cfs_rq)->cpu_load[2]; |
506 | ||
bf0f6f24 IM |
507 | /* |
508 | * Fix up delta_fair with the effect of us running | |
509 | * during the whole sleep period: | |
510 | */ | |
e59c80c5 | 511 | if (sched_feat(SLEEPER_AVG)) |
bf0f6f24 IM |
512 | delta_fair = div64_likely32((u64)delta_fair * load, |
513 | load + se->load.weight); | |
514 | ||
08e2388a | 515 | delta_fair = calc_weighted(delta_fair, se); |
bf0f6f24 IM |
516 | |
517 | prev_runtime = se->wait_runtime; | |
518 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
519 | delta_fair = se->wait_runtime - prev_runtime; | |
520 | ||
521 | /* | |
522 | * Track the amount of bonus we've given to sleepers: | |
523 | */ | |
524 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
525 | } |
526 | ||
2396af69 | 527 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
528 | { |
529 | struct task_struct *tsk = task_of(se); | |
530 | unsigned long delta_fair; | |
531 | ||
532 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
e59c80c5 | 533 | !sched_feat(FAIR_SLEEPERS)) |
bf0f6f24 IM |
534 | return; |
535 | ||
536 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
537 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
538 | ||
8ebc91d9 | 539 | __enqueue_sleeper(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
540 | |
541 | se->sleep_start_fair = 0; | |
542 | ||
543 | #ifdef CONFIG_SCHEDSTATS | |
544 | if (se->sleep_start) { | |
d281918d | 545 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
546 | |
547 | if ((s64)delta < 0) | |
548 | delta = 0; | |
549 | ||
550 | if (unlikely(delta > se->sleep_max)) | |
551 | se->sleep_max = delta; | |
552 | ||
553 | se->sleep_start = 0; | |
554 | se->sum_sleep_runtime += delta; | |
555 | } | |
556 | if (se->block_start) { | |
d281918d | 557 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
558 | |
559 | if ((s64)delta < 0) | |
560 | delta = 0; | |
561 | ||
562 | if (unlikely(delta > se->block_max)) | |
563 | se->block_max = delta; | |
564 | ||
565 | se->block_start = 0; | |
566 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
567 | |
568 | /* | |
569 | * Blocking time is in units of nanosecs, so shift by 20 to | |
570 | * get a milliseconds-range estimation of the amount of | |
571 | * time that the task spent sleeping: | |
572 | */ | |
573 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
574 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
575 | delta >> 20); | |
576 | } | |
bf0f6f24 IM |
577 | } |
578 | #endif | |
579 | } | |
580 | ||
581 | static void | |
668031ca | 582 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
583 | { |
584 | /* | |
585 | * Update the fair clock. | |
586 | */ | |
b7cc0896 | 587 | update_curr(cfs_rq); |
bf0f6f24 | 588 | |
e9acbff6 IM |
589 | if (wakeup) { |
590 | u64 min_runtime, latency; | |
591 | ||
592 | min_runtime = cfs_rq->min_vruntime; | |
593 | min_runtime += sysctl_sched_latency/2; | |
594 | ||
595 | if (sched_feat(NEW_FAIR_SLEEPERS)) { | |
596 | latency = calc_weighted(sysctl_sched_latency, se); | |
597 | if (min_runtime > latency) | |
598 | min_runtime -= latency; | |
599 | } | |
600 | ||
601 | se->vruntime = max(se->vruntime, min_runtime); | |
602 | ||
2396af69 | 603 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 604 | } |
bf0f6f24 | 605 | |
d2417e5a | 606 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
607 | __enqueue_entity(cfs_rq, se); |
608 | } | |
609 | ||
610 | static void | |
525c2716 | 611 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 612 | { |
19b6a2e3 | 613 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
614 | if (sleep) { |
615 | se->sleep_start_fair = cfs_rq->fair_clock; | |
616 | #ifdef CONFIG_SCHEDSTATS | |
617 | if (entity_is_task(se)) { | |
618 | struct task_struct *tsk = task_of(se); | |
619 | ||
620 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 621 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 622 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 623 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 624 | } |
bf0f6f24 IM |
625 | #endif |
626 | } | |
627 | __dequeue_entity(cfs_rq, se); | |
628 | } | |
629 | ||
630 | /* | |
631 | * Preempt the current task with a newly woken task if needed: | |
632 | */ | |
7c92e54f | 633 | static void |
bf0f6f24 IM |
634 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
635 | struct sched_entity *curr, unsigned long granularity) | |
636 | { | |
637 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
638 | unsigned long ideal_runtime, delta_exec; |
639 | ||
640 | /* | |
641 | * ideal_runtime is compared against sum_exec_runtime, which is | |
642 | * walltime, hence do not scale. | |
643 | */ | |
644 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
645 | (unsigned long)sysctl_sched_min_granularity); | |
646 | ||
647 | /* | |
648 | * If we executed more than what the latency constraint suggests, | |
649 | * reduce the rescheduling granularity. This way the total latency | |
650 | * of how much a task is not scheduled converges to | |
651 | * sysctl_sched_latency: | |
652 | */ | |
653 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
654 | if (delta_exec > ideal_runtime) | |
655 | granularity = 0; | |
bf0f6f24 IM |
656 | |
657 | /* | |
658 | * Take scheduling granularity into account - do not | |
659 | * preempt the current task unless the best task has | |
660 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
661 | * |
662 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 663 | */ |
4a55b450 | 664 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
665 | resched_task(rq_of(cfs_rq)->curr); |
666 | } | |
667 | ||
668 | static inline void | |
8494f412 | 669 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
670 | { |
671 | /* | |
672 | * Any task has to be enqueued before it get to execute on | |
673 | * a CPU. So account for the time it spent waiting on the | |
674 | * runqueue. (note, here we rely on pick_next_task() having | |
675 | * done a put_prev_task_fair() shortly before this, which | |
676 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
677 | */ | |
9ef0a961 | 678 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 679 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 680 | cfs_rq->curr = se; |
eba1ed4b IM |
681 | #ifdef CONFIG_SCHEDSTATS |
682 | /* | |
683 | * Track our maximum slice length, if the CPU's load is at | |
684 | * least twice that of our own weight (i.e. dont track it | |
685 | * when there are only lesser-weight tasks around): | |
686 | */ | |
687 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
688 | se->slice_max = max(se->slice_max, | |
689 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
690 | } | |
691 | #endif | |
4a55b450 | 692 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
693 | } |
694 | ||
9948f4b2 | 695 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
696 | { |
697 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
698 | ||
8494f412 | 699 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
700 | |
701 | return se; | |
702 | } | |
703 | ||
ab6cde26 | 704 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
705 | { |
706 | /* | |
707 | * If still on the runqueue then deactivate_task() | |
708 | * was not called and update_curr() has to be done: | |
709 | */ | |
710 | if (prev->on_rq) | |
b7cc0896 | 711 | update_curr(cfs_rq); |
bf0f6f24 | 712 | |
c7e9b5b2 | 713 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
714 | |
715 | if (prev->on_rq) | |
5870db5b | 716 | update_stats_wait_start(cfs_rq, prev); |
429d43bc | 717 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
718 | } |
719 | ||
720 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
721 | { | |
bf0f6f24 | 722 | struct sched_entity *next; |
c1b3da3e | 723 | |
bf0f6f24 IM |
724 | /* |
725 | * Dequeue and enqueue the task to update its | |
726 | * position within the tree: | |
727 | */ | |
525c2716 | 728 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 729 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
730 | |
731 | /* | |
732 | * Reschedule if another task tops the current one. | |
733 | */ | |
734 | next = __pick_next_entity(cfs_rq); | |
735 | if (next == curr) | |
736 | return; | |
737 | ||
11697830 PZ |
738 | __check_preempt_curr_fair(cfs_rq, next, curr, |
739 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
740 | } |
741 | ||
742 | /************************************************** | |
743 | * CFS operations on tasks: | |
744 | */ | |
745 | ||
746 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
747 | ||
748 | /* Walk up scheduling entities hierarchy */ | |
749 | #define for_each_sched_entity(se) \ | |
750 | for (; se; se = se->parent) | |
751 | ||
752 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
753 | { | |
754 | return p->se.cfs_rq; | |
755 | } | |
756 | ||
757 | /* runqueue on which this entity is (to be) queued */ | |
758 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
759 | { | |
760 | return se->cfs_rq; | |
761 | } | |
762 | ||
763 | /* runqueue "owned" by this group */ | |
764 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
765 | { | |
766 | return grp->my_q; | |
767 | } | |
768 | ||
769 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
770 | * another cpu ('this_cpu') | |
771 | */ | |
772 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
773 | { | |
774 | /* A later patch will take group into account */ | |
775 | return &cpu_rq(this_cpu)->cfs; | |
776 | } | |
777 | ||
778 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
779 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
780 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
781 | ||
782 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
783 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
784 | { | |
785 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
786 | return 1; | |
787 | ||
788 | return 0; | |
789 | } | |
790 | ||
791 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
792 | ||
793 | #define for_each_sched_entity(se) \ | |
794 | for (; se; se = NULL) | |
795 | ||
796 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
797 | { | |
798 | return &task_rq(p)->cfs; | |
799 | } | |
800 | ||
801 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
802 | { | |
803 | struct task_struct *p = task_of(se); | |
804 | struct rq *rq = task_rq(p); | |
805 | ||
806 | return &rq->cfs; | |
807 | } | |
808 | ||
809 | /* runqueue "owned" by this group */ | |
810 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
811 | { | |
812 | return NULL; | |
813 | } | |
814 | ||
815 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
816 | { | |
817 | return &cpu_rq(this_cpu)->cfs; | |
818 | } | |
819 | ||
820 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
821 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
822 | ||
823 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
824 | { | |
825 | return 1; | |
826 | } | |
827 | ||
828 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
829 | ||
830 | /* | |
831 | * The enqueue_task method is called before nr_running is | |
832 | * increased. Here we update the fair scheduling stats and | |
833 | * then put the task into the rbtree: | |
834 | */ | |
fd390f6a | 835 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
836 | { |
837 | struct cfs_rq *cfs_rq; | |
838 | struct sched_entity *se = &p->se; | |
839 | ||
840 | for_each_sched_entity(se) { | |
841 | if (se->on_rq) | |
842 | break; | |
843 | cfs_rq = cfs_rq_of(se); | |
668031ca | 844 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
845 | } |
846 | } | |
847 | ||
848 | /* | |
849 | * The dequeue_task method is called before nr_running is | |
850 | * decreased. We remove the task from the rbtree and | |
851 | * update the fair scheduling stats: | |
852 | */ | |
f02231e5 | 853 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
854 | { |
855 | struct cfs_rq *cfs_rq; | |
856 | struct sched_entity *se = &p->se; | |
857 | ||
858 | for_each_sched_entity(se) { | |
859 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 860 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
861 | /* Don't dequeue parent if it has other entities besides us */ |
862 | if (cfs_rq->load.weight) | |
863 | break; | |
864 | } | |
865 | } | |
866 | ||
867 | /* | |
1799e35d IM |
868 | * sched_yield() support is very simple - we dequeue and enqueue. |
869 | * | |
870 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
871 | */ |
872 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
873 | { | |
874 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
875 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
876 | struct sched_entity *rightmost, *se = &p->se; | |
877 | struct rb_node *parent; | |
bf0f6f24 IM |
878 | |
879 | /* | |
1799e35d IM |
880 | * Are we the only task in the tree? |
881 | */ | |
882 | if (unlikely(cfs_rq->nr_running == 1)) | |
883 | return; | |
884 | ||
885 | if (likely(!sysctl_sched_compat_yield)) { | |
886 | __update_rq_clock(rq); | |
887 | /* | |
888 | * Dequeue and enqueue the task to update its | |
889 | * position within the tree: | |
890 | */ | |
891 | dequeue_entity(cfs_rq, &p->se, 0); | |
892 | enqueue_entity(cfs_rq, &p->se, 0); | |
893 | ||
894 | return; | |
895 | } | |
896 | /* | |
897 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 898 | */ |
1799e35d IM |
899 | do { |
900 | parent = *link; | |
901 | link = &parent->rb_right; | |
902 | } while (*link); | |
903 | ||
904 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
905 | /* | |
906 | * Already in the rightmost position? | |
907 | */ | |
908 | if (unlikely(rightmost == se)) | |
909 | return; | |
910 | ||
911 | /* | |
912 | * Minimally necessary key value to be last in the tree: | |
913 | */ | |
914 | se->fair_key = rightmost->fair_key + 1; | |
915 | ||
916 | if (cfs_rq->rb_leftmost == &se->run_node) | |
917 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
918 | /* | |
919 | * Relink the task to the rightmost position: | |
920 | */ | |
921 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
922 | rb_link_node(&se->run_node, parent, link); | |
923 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
924 | } |
925 | ||
926 | /* | |
927 | * Preempt the current task with a newly woken task if needed: | |
928 | */ | |
929 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
930 | { | |
931 | struct task_struct *curr = rq->curr; | |
932 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
933 | unsigned long gran; | |
934 | ||
935 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 936 | update_rq_clock(rq); |
b7cc0896 | 937 | update_curr(cfs_rq); |
bf0f6f24 IM |
938 | resched_task(curr); |
939 | return; | |
940 | } | |
941 | ||
942 | gran = sysctl_sched_wakeup_granularity; | |
943 | /* | |
944 | * Batch tasks prefer throughput over latency: | |
945 | */ | |
946 | if (unlikely(p->policy == SCHED_BATCH)) | |
947 | gran = sysctl_sched_batch_wakeup_granularity; | |
948 | ||
949 | if (is_same_group(curr, p)) | |
950 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
951 | } | |
952 | ||
fb8d4724 | 953 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
954 | { |
955 | struct cfs_rq *cfs_rq = &rq->cfs; | |
956 | struct sched_entity *se; | |
957 | ||
958 | if (unlikely(!cfs_rq->nr_running)) | |
959 | return NULL; | |
960 | ||
961 | do { | |
9948f4b2 | 962 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
963 | cfs_rq = group_cfs_rq(se); |
964 | } while (cfs_rq); | |
965 | ||
966 | return task_of(se); | |
967 | } | |
968 | ||
969 | /* | |
970 | * Account for a descheduled task: | |
971 | */ | |
31ee529c | 972 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
973 | { |
974 | struct sched_entity *se = &prev->se; | |
975 | struct cfs_rq *cfs_rq; | |
976 | ||
977 | for_each_sched_entity(se) { | |
978 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 979 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
980 | } |
981 | } | |
982 | ||
983 | /************************************************** | |
984 | * Fair scheduling class load-balancing methods: | |
985 | */ | |
986 | ||
987 | /* | |
988 | * Load-balancing iterator. Note: while the runqueue stays locked | |
989 | * during the whole iteration, the current task might be | |
990 | * dequeued so the iterator has to be dequeue-safe. Here we | |
991 | * achieve that by always pre-iterating before returning | |
992 | * the current task: | |
993 | */ | |
994 | static inline struct task_struct * | |
995 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
996 | { | |
997 | struct task_struct *p; | |
998 | ||
999 | if (!curr) | |
1000 | return NULL; | |
1001 | ||
1002 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1003 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1004 | ||
1005 | return p; | |
1006 | } | |
1007 | ||
1008 | static struct task_struct *load_balance_start_fair(void *arg) | |
1009 | { | |
1010 | struct cfs_rq *cfs_rq = arg; | |
1011 | ||
1012 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1013 | } | |
1014 | ||
1015 | static struct task_struct *load_balance_next_fair(void *arg) | |
1016 | { | |
1017 | struct cfs_rq *cfs_rq = arg; | |
1018 | ||
1019 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1020 | } | |
1021 | ||
a4ac01c3 | 1022 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1023 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1024 | { | |
1025 | struct sched_entity *curr; | |
1026 | struct task_struct *p; | |
1027 | ||
1028 | if (!cfs_rq->nr_running) | |
1029 | return MAX_PRIO; | |
1030 | ||
1031 | curr = __pick_next_entity(cfs_rq); | |
1032 | p = task_of(curr); | |
1033 | ||
1034 | return p->prio; | |
1035 | } | |
a4ac01c3 | 1036 | #endif |
bf0f6f24 | 1037 | |
43010659 | 1038 | static unsigned long |
bf0f6f24 | 1039 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1040 | unsigned long max_nr_move, unsigned long max_load_move, |
1041 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1042 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1043 | { |
1044 | struct cfs_rq *busy_cfs_rq; | |
1045 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1046 | long rem_load_move = max_load_move; | |
1047 | struct rq_iterator cfs_rq_iterator; | |
1048 | ||
1049 | cfs_rq_iterator.start = load_balance_start_fair; | |
1050 | cfs_rq_iterator.next = load_balance_next_fair; | |
1051 | ||
1052 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1053 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1054 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1055 | long imbalance; |
bf0f6f24 | 1056 | unsigned long maxload; |
bf0f6f24 IM |
1057 | |
1058 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1059 | ||
e56f31aa | 1060 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1061 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1062 | if (imbalance <= 0) | |
1063 | continue; | |
1064 | ||
1065 | /* Don't pull more than imbalance/2 */ | |
1066 | imbalance /= 2; | |
1067 | maxload = min(rem_load_move, imbalance); | |
1068 | ||
a4ac01c3 PW |
1069 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1070 | #else | |
e56f31aa | 1071 | # define maxload rem_load_move |
a4ac01c3 | 1072 | #endif |
bf0f6f24 IM |
1073 | /* pass busy_cfs_rq argument into |
1074 | * load_balance_[start|next]_fair iterators | |
1075 | */ | |
1076 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1077 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1078 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1079 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1080 | |
1081 | total_nr_moved += nr_moved; | |
1082 | max_nr_move -= nr_moved; | |
1083 | rem_load_move -= load_moved; | |
1084 | ||
1085 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1086 | break; | |
1087 | } | |
1088 | ||
43010659 | 1089 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1090 | } |
1091 | ||
1092 | /* | |
1093 | * scheduler tick hitting a task of our scheduling class: | |
1094 | */ | |
1095 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1096 | { | |
1097 | struct cfs_rq *cfs_rq; | |
1098 | struct sched_entity *se = &curr->se; | |
1099 | ||
1100 | for_each_sched_entity(se) { | |
1101 | cfs_rq = cfs_rq_of(se); | |
1102 | entity_tick(cfs_rq, se); | |
1103 | } | |
1104 | } | |
1105 | ||
1106 | /* | |
1107 | * Share the fairness runtime between parent and child, thus the | |
1108 | * total amount of pressure for CPU stays equal - new tasks | |
1109 | * get a chance to run but frequent forkers are not allowed to | |
1110 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1111 | * the child is not running yet. | |
1112 | */ | |
ee0827d8 | 1113 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1114 | { |
1115 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1116 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
bf0f6f24 IM |
1117 | |
1118 | sched_info_queued(p); | |
1119 | ||
7109c442 | 1120 | update_curr(cfs_rq); |
d2417e5a | 1121 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1122 | /* |
1123 | * Child runs first: we let it run before the parent | |
1124 | * until it reschedules once. We set up the key so that | |
1125 | * it will preempt the parent: | |
1126 | */ | |
9f508f82 | 1127 | se->fair_key = curr->fair_key - |
7109c442 | 1128 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1129 | /* |
1130 | * The first wait is dominated by the child-runs-first logic, | |
1131 | * so do not credit it with that waiting time yet: | |
1132 | */ | |
e59c80c5 | 1133 | if (sched_feat(SKIP_INITIAL)) |
9f508f82 | 1134 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1135 | |
1136 | /* | |
1137 | * The statistical average of wait_runtime is about | |
1138 | * -granularity/2, so initialize the task with that: | |
1139 | */ | |
e59c80c5 | 1140 | if (sched_feat(START_DEBIT)) |
9f508f82 | 1141 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 | 1142 | |
e9acbff6 IM |
1143 | se->vruntime = cfs_rq->min_vruntime; |
1144 | update_stats_enqueue(cfs_rq, se); | |
bf0f6f24 | 1145 | __enqueue_entity(cfs_rq, se); |
bb61c210 | 1146 | resched_task(rq->curr); |
bf0f6f24 IM |
1147 | } |
1148 | ||
1149 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1150 | /* Account for a task changing its policy or group. | |
1151 | * | |
1152 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1153 | * migrates between groups/classes. | |
1154 | */ | |
1155 | static void set_curr_task_fair(struct rq *rq) | |
1156 | { | |
7c6c16f3 | 1157 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1158 | |
c3b64f1e IM |
1159 | for_each_sched_entity(se) |
1160 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1161 | } |
1162 | #else | |
1163 | static void set_curr_task_fair(struct rq *rq) | |
1164 | { | |
1165 | } | |
1166 | #endif | |
1167 | ||
1168 | /* | |
1169 | * All the scheduling class methods: | |
1170 | */ | |
1171 | struct sched_class fair_sched_class __read_mostly = { | |
1172 | .enqueue_task = enqueue_task_fair, | |
1173 | .dequeue_task = dequeue_task_fair, | |
1174 | .yield_task = yield_task_fair, | |
1175 | ||
1176 | .check_preempt_curr = check_preempt_curr_fair, | |
1177 | ||
1178 | .pick_next_task = pick_next_task_fair, | |
1179 | .put_prev_task = put_prev_task_fair, | |
1180 | ||
1181 | .load_balance = load_balance_fair, | |
1182 | ||
1183 | .set_curr_task = set_curr_task_fair, | |
1184 | .task_tick = task_tick_fair, | |
1185 | .task_new = task_new_fair, | |
1186 | }; | |
1187 | ||
1188 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1189 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1190 | { |
bf0f6f24 IM |
1191 | struct cfs_rq *cfs_rq; |
1192 | ||
c3b64f1e | 1193 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1194 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1195 | } |
1196 | #endif |