]>
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 | ||
4d78e7b6 PZ |
206 | static u64 __sched_period(unsigned long nr_running) |
207 | { | |
208 | u64 period = sysctl_sched_latency; | |
209 | unsigned long nr_latency = | |
210 | sysctl_sched_latency / sysctl_sched_min_granularity; | |
211 | ||
212 | if (unlikely(nr_running > nr_latency)) { | |
213 | period *= nr_running; | |
214 | do_div(period, nr_latency); | |
215 | } | |
216 | ||
217 | return period; | |
218 | } | |
219 | ||
21805085 PZ |
220 | /* |
221 | * Calculate the preemption granularity needed to schedule every | |
222 | * runnable task once per sysctl_sched_latency amount of time. | |
223 | * (down to a sensible low limit on granularity) | |
224 | * | |
225 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
226 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
227 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
228 | * for each task. We do finer and finer scheduling up to until we | |
229 | * reach the minimum granularity value. | |
230 | * | |
231 | * To achieve this we use the following dynamic-granularity rule: | |
232 | * | |
233 | * gran = lat/nr - lat/nr/nr | |
234 | * | |
235 | * This comes out of the following equations: | |
236 | * | |
237 | * kA1 + gran = kB1 | |
238 | * kB2 + gran = kA2 | |
239 | * kA2 = kA1 | |
240 | * kB2 = kB1 - d + d/nr | |
241 | * lat = d * nr | |
242 | * | |
243 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
244 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
245 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
246 | * running, 'lat' is the the period of each task. ('lat' is the | |
247 | * sched_latency that we aim for.) | |
248 | */ | |
249 | static long | |
250 | sched_granularity(struct cfs_rq *cfs_rq) | |
251 | { | |
252 | unsigned int gran = sysctl_sched_latency; | |
253 | unsigned int nr = cfs_rq->nr_running; | |
254 | ||
255 | if (nr > 1) { | |
256 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 257 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
258 | } |
259 | ||
260 | return gran; | |
261 | } | |
262 | ||
bf0f6f24 IM |
263 | /* |
264 | * We rescale the rescheduling granularity of tasks according to their | |
265 | * nice level, but only linearly, not exponentially: | |
266 | */ | |
267 | static long | |
268 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
269 | { | |
270 | u64 tmp; | |
271 | ||
7cff8cf6 IM |
272 | if (likely(curr->load.weight == NICE_0_LOAD)) |
273 | return granularity; | |
bf0f6f24 | 274 | /* |
7cff8cf6 | 275 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 276 | */ |
7cff8cf6 IM |
277 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
278 | tmp = curr->load.weight * (u64)granularity; | |
279 | return (long) (tmp >> NICE_0_SHIFT); | |
280 | } | |
bf0f6f24 | 281 | /* |
7cff8cf6 | 282 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
283 | * granularity: |
284 | */ | |
7cff8cf6 | 285 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
286 | |
287 | /* | |
288 | * It will always fit into 'long': | |
289 | */ | |
a0dc7260 | 290 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
291 | } |
292 | ||
293 | static inline void | |
294 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
295 | { | |
296 | long limit = sysctl_sched_runtime_limit; | |
297 | ||
298 | /* | |
299 | * Niced tasks have the same history dynamic range as | |
300 | * non-niced tasks: | |
301 | */ | |
302 | if (unlikely(se->wait_runtime > limit)) { | |
303 | se->wait_runtime = limit; | |
304 | schedstat_inc(se, wait_runtime_overruns); | |
305 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
306 | } | |
307 | if (unlikely(se->wait_runtime < -limit)) { | |
308 | se->wait_runtime = -limit; | |
309 | schedstat_inc(se, wait_runtime_underruns); | |
310 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
311 | } | |
312 | } | |
313 | ||
314 | static inline void | |
315 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
316 | { | |
317 | se->wait_runtime += delta; | |
318 | schedstat_add(se, sum_wait_runtime, delta); | |
319 | limit_wait_runtime(cfs_rq, se); | |
320 | } | |
321 | ||
322 | static void | |
323 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
324 | { | |
325 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
326 | __add_wait_runtime(cfs_rq, se, delta); | |
327 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
328 | } | |
329 | ||
330 | /* | |
331 | * Update the current task's runtime statistics. Skip current tasks that | |
332 | * are not in our scheduling class. | |
333 | */ | |
334 | static inline void | |
8ebc91d9 IM |
335 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
336 | unsigned long delta_exec) | |
bf0f6f24 | 337 | { |
e9acbff6 | 338 | unsigned long delta, delta_fair, delta_mine, delta_exec_weighted; |
bf0f6f24 IM |
339 | struct load_weight *lw = &cfs_rq->load; |
340 | unsigned long load = lw->weight; | |
341 | ||
8179ca23 | 342 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
343 | |
344 | curr->sum_exec_runtime += delta_exec; | |
345 | cfs_rq->exec_clock += delta_exec; | |
e9acbff6 IM |
346 | delta_exec_weighted = delta_exec; |
347 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { | |
348 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, | |
349 | &curr->load); | |
350 | } | |
351 | curr->vruntime += delta_exec_weighted; | |
bf0f6f24 | 352 | |
6cb58195 IM |
353 | if (!sched_feat(FAIR_SLEEPERS)) |
354 | return; | |
355 | ||
fd8bb43e IM |
356 | if (unlikely(!load)) |
357 | return; | |
358 | ||
bf0f6f24 IM |
359 | delta_fair = calc_delta_fair(delta_exec, lw); |
360 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
361 | ||
5f01d519 | 362 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 363 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
364 | delta = min(delta, (unsigned long)( |
365 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
366 | cfs_rq->sleeper_bonus -= delta; |
367 | delta_mine -= delta; | |
368 | } | |
369 | ||
370 | cfs_rq->fair_clock += delta_fair; | |
371 | /* | |
372 | * We executed delta_exec amount of time on the CPU, | |
373 | * but we were only entitled to delta_mine amount of | |
374 | * time during that period (if nr_running == 1 then | |
375 | * the two values are equal) | |
376 | * [Note: delta_mine - delta_exec is negative]: | |
377 | */ | |
378 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
379 | } | |
380 | ||
b7cc0896 | 381 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 382 | { |
429d43bc | 383 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 384 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
385 | unsigned long delta_exec; |
386 | ||
387 | if (unlikely(!curr)) | |
388 | return; | |
389 | ||
390 | /* | |
391 | * Get the amount of time the current task was running | |
392 | * since the last time we changed load (this cannot | |
393 | * overflow on 32 bits): | |
394 | */ | |
8ebc91d9 | 395 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 396 | |
8ebc91d9 IM |
397 | __update_curr(cfs_rq, curr, delta_exec); |
398 | curr->exec_start = now; | |
bf0f6f24 IM |
399 | } |
400 | ||
401 | static inline void | |
5870db5b | 402 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
403 | { |
404 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 405 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
406 | } |
407 | ||
bf0f6f24 | 408 | static inline unsigned long |
08e2388a | 409 | calc_weighted(unsigned long delta, struct sched_entity *se) |
bf0f6f24 | 410 | { |
08e2388a | 411 | unsigned long weight = se->load.weight; |
bf0f6f24 | 412 | |
08e2388a IM |
413 | if (unlikely(weight != NICE_0_LOAD)) |
414 | return (u64)delta * se->load.weight >> NICE_0_SHIFT; | |
415 | else | |
416 | return delta; | |
bf0f6f24 | 417 | } |
bf0f6f24 IM |
418 | |
419 | /* | |
420 | * Task is being enqueued - update stats: | |
421 | */ | |
d2417e5a | 422 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 423 | { |
bf0f6f24 IM |
424 | /* |
425 | * Are we enqueueing a waiting task? (for current tasks | |
426 | * a dequeue/enqueue event is a NOP) | |
427 | */ | |
429d43bc | 428 | if (se != cfs_rq->curr) |
5870db5b | 429 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
430 | /* |
431 | * Update the key: | |
432 | */ | |
e9acbff6 | 433 | se->fair_key = se->vruntime; |
bf0f6f24 IM |
434 | } |
435 | ||
436 | /* | |
437 | * Note: must be called with a freshly updated rq->fair_clock. | |
438 | */ | |
439 | static inline void | |
8ebc91d9 IM |
440 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, |
441 | unsigned long delta_fair) | |
bf0f6f24 | 442 | { |
d281918d IM |
443 | schedstat_set(se->wait_max, max(se->wait_max, |
444 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 | 445 | |
08e2388a | 446 | delta_fair = calc_weighted(delta_fair, se); |
bf0f6f24 IM |
447 | |
448 | add_wait_runtime(cfs_rq, se, delta_fair); | |
449 | } | |
450 | ||
451 | static void | |
9ef0a961 | 452 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
453 | { |
454 | unsigned long delta_fair; | |
455 | ||
b77d69db IM |
456 | if (unlikely(!se->wait_start_fair)) |
457 | return; | |
458 | ||
bf0f6f24 IM |
459 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
460 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
461 | ||
8ebc91d9 | 462 | __update_stats_wait_end(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
463 | |
464 | se->wait_start_fair = 0; | |
6cfb0d5d | 465 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
466 | } |
467 | ||
468 | static inline void | |
19b6a2e3 | 469 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 470 | { |
b7cc0896 | 471 | update_curr(cfs_rq); |
bf0f6f24 IM |
472 | /* |
473 | * Mark the end of the wait period if dequeueing a | |
474 | * waiting task: | |
475 | */ | |
429d43bc | 476 | if (se != cfs_rq->curr) |
9ef0a961 | 477 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
478 | } |
479 | ||
480 | /* | |
481 | * We are picking a new current task - update its stats: | |
482 | */ | |
483 | static inline void | |
79303e9e | 484 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
485 | { |
486 | /* | |
487 | * We are starting a new run period: | |
488 | */ | |
d281918d | 489 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
490 | } |
491 | ||
492 | /* | |
493 | * We are descheduling a task - update its stats: | |
494 | */ | |
495 | static inline void | |
c7e9b5b2 | 496 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
497 | { |
498 | se->exec_start = 0; | |
499 | } | |
500 | ||
501 | /************************************************** | |
502 | * Scheduling class queueing methods: | |
503 | */ | |
504 | ||
8ebc91d9 IM |
505 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, |
506 | unsigned long delta_fair) | |
bf0f6f24 | 507 | { |
8ebc91d9 | 508 | unsigned long load = cfs_rq->load.weight; |
bf0f6f24 IM |
509 | long prev_runtime; |
510 | ||
b2133c8b IM |
511 | /* |
512 | * Do not boost sleepers if there's too much bonus 'in flight' | |
513 | * already: | |
514 | */ | |
515 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
516 | return; | |
517 | ||
e59c80c5 | 518 | if (sched_feat(SLEEPER_LOAD_AVG)) |
bf0f6f24 IM |
519 | load = rq_of(cfs_rq)->cpu_load[2]; |
520 | ||
bf0f6f24 IM |
521 | /* |
522 | * Fix up delta_fair with the effect of us running | |
523 | * during the whole sleep period: | |
524 | */ | |
e59c80c5 | 525 | if (sched_feat(SLEEPER_AVG)) |
bf0f6f24 IM |
526 | delta_fair = div64_likely32((u64)delta_fair * load, |
527 | load + se->load.weight); | |
528 | ||
08e2388a | 529 | delta_fair = calc_weighted(delta_fair, se); |
bf0f6f24 IM |
530 | |
531 | prev_runtime = se->wait_runtime; | |
532 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
533 | delta_fair = se->wait_runtime - prev_runtime; | |
534 | ||
535 | /* | |
536 | * Track the amount of bonus we've given to sleepers: | |
537 | */ | |
538 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
539 | } |
540 | ||
2396af69 | 541 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
542 | { |
543 | struct task_struct *tsk = task_of(se); | |
544 | unsigned long delta_fair; | |
545 | ||
546 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
e59c80c5 | 547 | !sched_feat(FAIR_SLEEPERS)) |
bf0f6f24 IM |
548 | return; |
549 | ||
550 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
551 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
552 | ||
8ebc91d9 | 553 | __enqueue_sleeper(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
554 | |
555 | se->sleep_start_fair = 0; | |
556 | ||
557 | #ifdef CONFIG_SCHEDSTATS | |
558 | if (se->sleep_start) { | |
d281918d | 559 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
560 | |
561 | if ((s64)delta < 0) | |
562 | delta = 0; | |
563 | ||
564 | if (unlikely(delta > se->sleep_max)) | |
565 | se->sleep_max = delta; | |
566 | ||
567 | se->sleep_start = 0; | |
568 | se->sum_sleep_runtime += delta; | |
569 | } | |
570 | if (se->block_start) { | |
d281918d | 571 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
572 | |
573 | if ((s64)delta < 0) | |
574 | delta = 0; | |
575 | ||
576 | if (unlikely(delta > se->block_max)) | |
577 | se->block_max = delta; | |
578 | ||
579 | se->block_start = 0; | |
580 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
581 | |
582 | /* | |
583 | * Blocking time is in units of nanosecs, so shift by 20 to | |
584 | * get a milliseconds-range estimation of the amount of | |
585 | * time that the task spent sleeping: | |
586 | */ | |
587 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
588 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
589 | delta >> 20); | |
590 | } | |
bf0f6f24 IM |
591 | } |
592 | #endif | |
593 | } | |
594 | ||
595 | static void | |
668031ca | 596 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
597 | { |
598 | /* | |
599 | * Update the fair clock. | |
600 | */ | |
b7cc0896 | 601 | update_curr(cfs_rq); |
bf0f6f24 | 602 | |
e9acbff6 IM |
603 | if (wakeup) { |
604 | u64 min_runtime, latency; | |
605 | ||
606 | min_runtime = cfs_rq->min_vruntime; | |
607 | min_runtime += sysctl_sched_latency/2; | |
608 | ||
609 | if (sched_feat(NEW_FAIR_SLEEPERS)) { | |
610 | latency = calc_weighted(sysctl_sched_latency, se); | |
611 | if (min_runtime > latency) | |
612 | min_runtime -= latency; | |
613 | } | |
614 | ||
615 | se->vruntime = max(se->vruntime, min_runtime); | |
616 | ||
2396af69 | 617 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 618 | } |
bf0f6f24 | 619 | |
d2417e5a | 620 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
621 | __enqueue_entity(cfs_rq, se); |
622 | } | |
623 | ||
624 | static void | |
525c2716 | 625 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 626 | { |
19b6a2e3 | 627 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
628 | if (sleep) { |
629 | se->sleep_start_fair = cfs_rq->fair_clock; | |
630 | #ifdef CONFIG_SCHEDSTATS | |
631 | if (entity_is_task(se)) { | |
632 | struct task_struct *tsk = task_of(se); | |
633 | ||
634 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 635 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 636 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 637 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 638 | } |
bf0f6f24 IM |
639 | #endif |
640 | } | |
641 | __dequeue_entity(cfs_rq, se); | |
642 | } | |
643 | ||
644 | /* | |
645 | * Preempt the current task with a newly woken task if needed: | |
646 | */ | |
7c92e54f | 647 | static void |
bf0f6f24 IM |
648 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
649 | struct sched_entity *curr, unsigned long granularity) | |
650 | { | |
651 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
652 | unsigned long ideal_runtime, delta_exec; |
653 | ||
654 | /* | |
655 | * ideal_runtime is compared against sum_exec_runtime, which is | |
656 | * walltime, hence do not scale. | |
657 | */ | |
658 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
659 | (unsigned long)sysctl_sched_min_granularity); | |
660 | ||
661 | /* | |
662 | * If we executed more than what the latency constraint suggests, | |
663 | * reduce the rescheduling granularity. This way the total latency | |
664 | * of how much a task is not scheduled converges to | |
665 | * sysctl_sched_latency: | |
666 | */ | |
667 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
668 | if (delta_exec > ideal_runtime) | |
669 | granularity = 0; | |
bf0f6f24 IM |
670 | |
671 | /* | |
672 | * Take scheduling granularity into account - do not | |
673 | * preempt the current task unless the best task has | |
674 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
675 | * |
676 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 677 | */ |
4a55b450 | 678 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
679 | resched_task(rq_of(cfs_rq)->curr); |
680 | } | |
681 | ||
682 | static inline void | |
8494f412 | 683 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
684 | { |
685 | /* | |
686 | * Any task has to be enqueued before it get to execute on | |
687 | * a CPU. So account for the time it spent waiting on the | |
688 | * runqueue. (note, here we rely on pick_next_task() having | |
689 | * done a put_prev_task_fair() shortly before this, which | |
690 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
691 | */ | |
9ef0a961 | 692 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 693 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 694 | cfs_rq->curr = se; |
eba1ed4b IM |
695 | #ifdef CONFIG_SCHEDSTATS |
696 | /* | |
697 | * Track our maximum slice length, if the CPU's load is at | |
698 | * least twice that of our own weight (i.e. dont track it | |
699 | * when there are only lesser-weight tasks around): | |
700 | */ | |
701 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
702 | se->slice_max = max(se->slice_max, | |
703 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
704 | } | |
705 | #endif | |
4a55b450 | 706 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
707 | } |
708 | ||
9948f4b2 | 709 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
710 | { |
711 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
712 | ||
8494f412 | 713 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
714 | |
715 | return se; | |
716 | } | |
717 | ||
ab6cde26 | 718 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
719 | { |
720 | /* | |
721 | * If still on the runqueue then deactivate_task() | |
722 | * was not called and update_curr() has to be done: | |
723 | */ | |
724 | if (prev->on_rq) | |
b7cc0896 | 725 | update_curr(cfs_rq); |
bf0f6f24 | 726 | |
c7e9b5b2 | 727 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
728 | |
729 | if (prev->on_rq) | |
5870db5b | 730 | update_stats_wait_start(cfs_rq, prev); |
429d43bc | 731 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
732 | } |
733 | ||
734 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
735 | { | |
bf0f6f24 | 736 | struct sched_entity *next; |
c1b3da3e | 737 | |
bf0f6f24 IM |
738 | /* |
739 | * Dequeue and enqueue the task to update its | |
740 | * position within the tree: | |
741 | */ | |
525c2716 | 742 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 743 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
744 | |
745 | /* | |
746 | * Reschedule if another task tops the current one. | |
747 | */ | |
748 | next = __pick_next_entity(cfs_rq); | |
749 | if (next == curr) | |
750 | return; | |
751 | ||
11697830 PZ |
752 | __check_preempt_curr_fair(cfs_rq, next, curr, |
753 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
754 | } |
755 | ||
756 | /************************************************** | |
757 | * CFS operations on tasks: | |
758 | */ | |
759 | ||
760 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
761 | ||
762 | /* Walk up scheduling entities hierarchy */ | |
763 | #define for_each_sched_entity(se) \ | |
764 | for (; se; se = se->parent) | |
765 | ||
766 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
767 | { | |
768 | return p->se.cfs_rq; | |
769 | } | |
770 | ||
771 | /* runqueue on which this entity is (to be) queued */ | |
772 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
773 | { | |
774 | return se->cfs_rq; | |
775 | } | |
776 | ||
777 | /* runqueue "owned" by this group */ | |
778 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
779 | { | |
780 | return grp->my_q; | |
781 | } | |
782 | ||
783 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
784 | * another cpu ('this_cpu') | |
785 | */ | |
786 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
787 | { | |
788 | /* A later patch will take group into account */ | |
789 | return &cpu_rq(this_cpu)->cfs; | |
790 | } | |
791 | ||
792 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
793 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
794 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
795 | ||
796 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
797 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
798 | { | |
799 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
800 | return 1; | |
801 | ||
802 | return 0; | |
803 | } | |
804 | ||
805 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
806 | ||
807 | #define for_each_sched_entity(se) \ | |
808 | for (; se; se = NULL) | |
809 | ||
810 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
811 | { | |
812 | return &task_rq(p)->cfs; | |
813 | } | |
814 | ||
815 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
816 | { | |
817 | struct task_struct *p = task_of(se); | |
818 | struct rq *rq = task_rq(p); | |
819 | ||
820 | return &rq->cfs; | |
821 | } | |
822 | ||
823 | /* runqueue "owned" by this group */ | |
824 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
825 | { | |
826 | return NULL; | |
827 | } | |
828 | ||
829 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
830 | { | |
831 | return &cpu_rq(this_cpu)->cfs; | |
832 | } | |
833 | ||
834 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
835 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
836 | ||
837 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
838 | { | |
839 | return 1; | |
840 | } | |
841 | ||
842 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
843 | ||
844 | /* | |
845 | * The enqueue_task method is called before nr_running is | |
846 | * increased. Here we update the fair scheduling stats and | |
847 | * then put the task into the rbtree: | |
848 | */ | |
fd390f6a | 849 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
850 | { |
851 | struct cfs_rq *cfs_rq; | |
852 | struct sched_entity *se = &p->se; | |
853 | ||
854 | for_each_sched_entity(se) { | |
855 | if (se->on_rq) | |
856 | break; | |
857 | cfs_rq = cfs_rq_of(se); | |
668031ca | 858 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
859 | } |
860 | } | |
861 | ||
862 | /* | |
863 | * The dequeue_task method is called before nr_running is | |
864 | * decreased. We remove the task from the rbtree and | |
865 | * update the fair scheduling stats: | |
866 | */ | |
f02231e5 | 867 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
868 | { |
869 | struct cfs_rq *cfs_rq; | |
870 | struct sched_entity *se = &p->se; | |
871 | ||
872 | for_each_sched_entity(se) { | |
873 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 874 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
875 | /* Don't dequeue parent if it has other entities besides us */ |
876 | if (cfs_rq->load.weight) | |
877 | break; | |
878 | } | |
879 | } | |
880 | ||
881 | /* | |
1799e35d IM |
882 | * sched_yield() support is very simple - we dequeue and enqueue. |
883 | * | |
884 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
885 | */ |
886 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
887 | { | |
888 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
889 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
890 | struct sched_entity *rightmost, *se = &p->se; | |
891 | struct rb_node *parent; | |
bf0f6f24 IM |
892 | |
893 | /* | |
1799e35d IM |
894 | * Are we the only task in the tree? |
895 | */ | |
896 | if (unlikely(cfs_rq->nr_running == 1)) | |
897 | return; | |
898 | ||
899 | if (likely(!sysctl_sched_compat_yield)) { | |
900 | __update_rq_clock(rq); | |
901 | /* | |
902 | * Dequeue and enqueue the task to update its | |
903 | * position within the tree: | |
904 | */ | |
905 | dequeue_entity(cfs_rq, &p->se, 0); | |
906 | enqueue_entity(cfs_rq, &p->se, 0); | |
907 | ||
908 | return; | |
909 | } | |
910 | /* | |
911 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 912 | */ |
1799e35d IM |
913 | do { |
914 | parent = *link; | |
915 | link = &parent->rb_right; | |
916 | } while (*link); | |
917 | ||
918 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
919 | /* | |
920 | * Already in the rightmost position? | |
921 | */ | |
922 | if (unlikely(rightmost == se)) | |
923 | return; | |
924 | ||
925 | /* | |
926 | * Minimally necessary key value to be last in the tree: | |
927 | */ | |
928 | se->fair_key = rightmost->fair_key + 1; | |
929 | ||
930 | if (cfs_rq->rb_leftmost == &se->run_node) | |
931 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
932 | /* | |
933 | * Relink the task to the rightmost position: | |
934 | */ | |
935 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
936 | rb_link_node(&se->run_node, parent, link); | |
937 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
938 | } |
939 | ||
940 | /* | |
941 | * Preempt the current task with a newly woken task if needed: | |
942 | */ | |
943 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
944 | { | |
945 | struct task_struct *curr = rq->curr; | |
946 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
947 | unsigned long gran; | |
948 | ||
949 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 950 | update_rq_clock(rq); |
b7cc0896 | 951 | update_curr(cfs_rq); |
bf0f6f24 IM |
952 | resched_task(curr); |
953 | return; | |
954 | } | |
955 | ||
956 | gran = sysctl_sched_wakeup_granularity; | |
957 | /* | |
958 | * Batch tasks prefer throughput over latency: | |
959 | */ | |
960 | if (unlikely(p->policy == SCHED_BATCH)) | |
961 | gran = sysctl_sched_batch_wakeup_granularity; | |
962 | ||
963 | if (is_same_group(curr, p)) | |
964 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
965 | } | |
966 | ||
fb8d4724 | 967 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
968 | { |
969 | struct cfs_rq *cfs_rq = &rq->cfs; | |
970 | struct sched_entity *se; | |
971 | ||
972 | if (unlikely(!cfs_rq->nr_running)) | |
973 | return NULL; | |
974 | ||
975 | do { | |
9948f4b2 | 976 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
977 | cfs_rq = group_cfs_rq(se); |
978 | } while (cfs_rq); | |
979 | ||
980 | return task_of(se); | |
981 | } | |
982 | ||
983 | /* | |
984 | * Account for a descheduled task: | |
985 | */ | |
31ee529c | 986 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
987 | { |
988 | struct sched_entity *se = &prev->se; | |
989 | struct cfs_rq *cfs_rq; | |
990 | ||
991 | for_each_sched_entity(se) { | |
992 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 993 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
994 | } |
995 | } | |
996 | ||
997 | /************************************************** | |
998 | * Fair scheduling class load-balancing methods: | |
999 | */ | |
1000 | ||
1001 | /* | |
1002 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1003 | * during the whole iteration, the current task might be | |
1004 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1005 | * achieve that by always pre-iterating before returning | |
1006 | * the current task: | |
1007 | */ | |
1008 | static inline struct task_struct * | |
1009 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1010 | { | |
1011 | struct task_struct *p; | |
1012 | ||
1013 | if (!curr) | |
1014 | return NULL; | |
1015 | ||
1016 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1017 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1018 | ||
1019 | return p; | |
1020 | } | |
1021 | ||
1022 | static struct task_struct *load_balance_start_fair(void *arg) | |
1023 | { | |
1024 | struct cfs_rq *cfs_rq = arg; | |
1025 | ||
1026 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1027 | } | |
1028 | ||
1029 | static struct task_struct *load_balance_next_fair(void *arg) | |
1030 | { | |
1031 | struct cfs_rq *cfs_rq = arg; | |
1032 | ||
1033 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1034 | } | |
1035 | ||
a4ac01c3 | 1036 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1037 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1038 | { | |
1039 | struct sched_entity *curr; | |
1040 | struct task_struct *p; | |
1041 | ||
1042 | if (!cfs_rq->nr_running) | |
1043 | return MAX_PRIO; | |
1044 | ||
1045 | curr = __pick_next_entity(cfs_rq); | |
1046 | p = task_of(curr); | |
1047 | ||
1048 | return p->prio; | |
1049 | } | |
a4ac01c3 | 1050 | #endif |
bf0f6f24 | 1051 | |
43010659 | 1052 | static unsigned long |
bf0f6f24 | 1053 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1054 | unsigned long max_nr_move, unsigned long max_load_move, |
1055 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1056 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1057 | { |
1058 | struct cfs_rq *busy_cfs_rq; | |
1059 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1060 | long rem_load_move = max_load_move; | |
1061 | struct rq_iterator cfs_rq_iterator; | |
1062 | ||
1063 | cfs_rq_iterator.start = load_balance_start_fair; | |
1064 | cfs_rq_iterator.next = load_balance_next_fair; | |
1065 | ||
1066 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1067 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1068 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1069 | long imbalance; |
bf0f6f24 | 1070 | unsigned long maxload; |
bf0f6f24 IM |
1071 | |
1072 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1073 | ||
e56f31aa | 1074 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1075 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1076 | if (imbalance <= 0) | |
1077 | continue; | |
1078 | ||
1079 | /* Don't pull more than imbalance/2 */ | |
1080 | imbalance /= 2; | |
1081 | maxload = min(rem_load_move, imbalance); | |
1082 | ||
a4ac01c3 PW |
1083 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1084 | #else | |
e56f31aa | 1085 | # define maxload rem_load_move |
a4ac01c3 | 1086 | #endif |
bf0f6f24 IM |
1087 | /* pass busy_cfs_rq argument into |
1088 | * load_balance_[start|next]_fair iterators | |
1089 | */ | |
1090 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1091 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1092 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1093 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1094 | |
1095 | total_nr_moved += nr_moved; | |
1096 | max_nr_move -= nr_moved; | |
1097 | rem_load_move -= load_moved; | |
1098 | ||
1099 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1100 | break; | |
1101 | } | |
1102 | ||
43010659 | 1103 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1104 | } |
1105 | ||
1106 | /* | |
1107 | * scheduler tick hitting a task of our scheduling class: | |
1108 | */ | |
1109 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1110 | { | |
1111 | struct cfs_rq *cfs_rq; | |
1112 | struct sched_entity *se = &curr->se; | |
1113 | ||
1114 | for_each_sched_entity(se) { | |
1115 | cfs_rq = cfs_rq_of(se); | |
1116 | entity_tick(cfs_rq, se); | |
1117 | } | |
1118 | } | |
1119 | ||
4d78e7b6 PZ |
1120 | #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
1121 | ||
bf0f6f24 IM |
1122 | /* |
1123 | * Share the fairness runtime between parent and child, thus the | |
1124 | * total amount of pressure for CPU stays equal - new tasks | |
1125 | * get a chance to run but frequent forkers are not allowed to | |
1126 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1127 | * the child is not running yet. | |
1128 | */ | |
ee0827d8 | 1129 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1130 | { |
1131 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1132 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
bf0f6f24 IM |
1133 | |
1134 | sched_info_queued(p); | |
1135 | ||
7109c442 | 1136 | update_curr(cfs_rq); |
4d78e7b6 | 1137 | se->vruntime = cfs_rq->min_vruntime; |
d2417e5a | 1138 | update_stats_enqueue(cfs_rq, se); |
4d78e7b6 | 1139 | |
bf0f6f24 IM |
1140 | /* |
1141 | * The first wait is dominated by the child-runs-first logic, | |
1142 | * so do not credit it with that waiting time yet: | |
1143 | */ | |
e59c80c5 | 1144 | if (sched_feat(SKIP_INITIAL)) |
9f508f82 | 1145 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1146 | |
1147 | /* | |
1148 | * The statistical average of wait_runtime is about | |
1149 | * -granularity/2, so initialize the task with that: | |
1150 | */ | |
e59c80c5 | 1151 | if (sched_feat(START_DEBIT)) |
4d78e7b6 PZ |
1152 | se->wait_runtime = -(__sched_period(cfs_rq->nr_running+1) / 2); |
1153 | ||
1154 | if (sysctl_sched_child_runs_first && | |
1155 | curr->vruntime < se->vruntime) { | |
1156 | ||
1157 | dequeue_entity(cfs_rq, curr, 0); | |
1158 | swap(curr->vruntime, se->vruntime); | |
1159 | enqueue_entity(cfs_rq, curr, 0); | |
1160 | } | |
bf0f6f24 | 1161 | |
e9acbff6 | 1162 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 | 1163 | __enqueue_entity(cfs_rq, se); |
bb61c210 | 1164 | resched_task(rq->curr); |
bf0f6f24 IM |
1165 | } |
1166 | ||
1167 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1168 | /* Account for a task changing its policy or group. | |
1169 | * | |
1170 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1171 | * migrates between groups/classes. | |
1172 | */ | |
1173 | static void set_curr_task_fair(struct rq *rq) | |
1174 | { | |
7c6c16f3 | 1175 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1176 | |
c3b64f1e IM |
1177 | for_each_sched_entity(se) |
1178 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1179 | } |
1180 | #else | |
1181 | static void set_curr_task_fair(struct rq *rq) | |
1182 | { | |
1183 | } | |
1184 | #endif | |
1185 | ||
1186 | /* | |
1187 | * All the scheduling class methods: | |
1188 | */ | |
1189 | struct sched_class fair_sched_class __read_mostly = { | |
1190 | .enqueue_task = enqueue_task_fair, | |
1191 | .dequeue_task = dequeue_task_fair, | |
1192 | .yield_task = yield_task_fair, | |
1193 | ||
1194 | .check_preempt_curr = check_preempt_curr_fair, | |
1195 | ||
1196 | .pick_next_task = pick_next_task_fair, | |
1197 | .put_prev_task = put_prev_task_fair, | |
1198 | ||
1199 | .load_balance = load_balance_fair, | |
1200 | ||
1201 | .set_curr_task = set_curr_task_fair, | |
1202 | .task_tick = task_tick_fair, | |
1203 | .task_new = task_new_fair, | |
1204 | }; | |
1205 | ||
1206 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1207 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1208 | { |
bf0f6f24 IM |
1209 | struct cfs_rq *cfs_rq; |
1210 | ||
c3b64f1e | 1211 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1212 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1213 | } |
1214 | #endif |