]>
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 | */ | |
151 | static inline void | |
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 | ||
194 | static inline void | |
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 | ||
400 | /* | |
401 | * We calculate fair deltas here, so protect against the random effects | |
402 | * of a multiplication overflow by capping it to the runtime limit: | |
403 | */ | |
404 | #if BITS_PER_LONG == 32 | |
405 | static inline unsigned long | |
406 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
407 | { | |
408 | u64 tmp = (u64)delta * weight >> shift; | |
409 | ||
410 | if (unlikely(tmp > sysctl_sched_runtime_limit*2)) | |
411 | return sysctl_sched_runtime_limit*2; | |
412 | return tmp; | |
413 | } | |
414 | #else | |
415 | static inline unsigned long | |
416 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
417 | { | |
418 | return delta * weight >> shift; | |
419 | } | |
420 | #endif | |
421 | ||
422 | /* | |
423 | * Task is being enqueued - update stats: | |
424 | */ | |
d2417e5a | 425 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
426 | { |
427 | s64 key; | |
428 | ||
429 | /* | |
430 | * Are we enqueueing a waiting task? (for current tasks | |
431 | * a dequeue/enqueue event is a NOP) | |
432 | */ | |
429d43bc | 433 | if (se != cfs_rq->curr) |
5870db5b | 434 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
435 | /* |
436 | * Update the key: | |
437 | */ | |
438 | key = cfs_rq->fair_clock; | |
439 | ||
440 | /* | |
441 | * Optimize the common nice 0 case: | |
442 | */ | |
443 | if (likely(se->load.weight == NICE_0_LOAD)) { | |
444 | key -= se->wait_runtime; | |
445 | } else { | |
446 | u64 tmp; | |
447 | ||
448 | if (se->wait_runtime < 0) { | |
449 | tmp = -se->wait_runtime; | |
450 | key += (tmp * se->load.inv_weight) >> | |
451 | (WMULT_SHIFT - NICE_0_SHIFT); | |
452 | } else { | |
453 | tmp = se->wait_runtime; | |
a69edb55 IM |
454 | key -= (tmp * se->load.inv_weight) >> |
455 | (WMULT_SHIFT - NICE_0_SHIFT); | |
bf0f6f24 IM |
456 | } |
457 | } | |
458 | ||
459 | se->fair_key = key; | |
460 | } | |
461 | ||
462 | /* | |
463 | * Note: must be called with a freshly updated rq->fair_clock. | |
464 | */ | |
465 | static inline void | |
8ebc91d9 IM |
466 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, |
467 | unsigned long delta_fair) | |
bf0f6f24 | 468 | { |
d281918d IM |
469 | schedstat_set(se->wait_max, max(se->wait_max, |
470 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 IM |
471 | |
472 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
473 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
474 | NICE_0_SHIFT); | |
475 | ||
476 | add_wait_runtime(cfs_rq, se, delta_fair); | |
477 | } | |
478 | ||
479 | static void | |
9ef0a961 | 480 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
481 | { |
482 | unsigned long delta_fair; | |
483 | ||
b77d69db IM |
484 | if (unlikely(!se->wait_start_fair)) |
485 | return; | |
486 | ||
bf0f6f24 IM |
487 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
488 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
489 | ||
8ebc91d9 | 490 | __update_stats_wait_end(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
491 | |
492 | se->wait_start_fair = 0; | |
6cfb0d5d | 493 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
494 | } |
495 | ||
496 | static inline void | |
19b6a2e3 | 497 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 498 | { |
b7cc0896 | 499 | update_curr(cfs_rq); |
bf0f6f24 IM |
500 | /* |
501 | * Mark the end of the wait period if dequeueing a | |
502 | * waiting task: | |
503 | */ | |
429d43bc | 504 | if (se != cfs_rq->curr) |
9ef0a961 | 505 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
506 | } |
507 | ||
508 | /* | |
509 | * We are picking a new current task - update its stats: | |
510 | */ | |
511 | static inline void | |
79303e9e | 512 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
513 | { |
514 | /* | |
515 | * We are starting a new run period: | |
516 | */ | |
d281918d | 517 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
518 | } |
519 | ||
520 | /* | |
521 | * We are descheduling a task - update its stats: | |
522 | */ | |
523 | static inline void | |
c7e9b5b2 | 524 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
525 | { |
526 | se->exec_start = 0; | |
527 | } | |
528 | ||
529 | /************************************************** | |
530 | * Scheduling class queueing methods: | |
531 | */ | |
532 | ||
8ebc91d9 IM |
533 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, |
534 | unsigned long delta_fair) | |
bf0f6f24 | 535 | { |
8ebc91d9 | 536 | unsigned long load = cfs_rq->load.weight; |
bf0f6f24 IM |
537 | long prev_runtime; |
538 | ||
b2133c8b IM |
539 | /* |
540 | * Do not boost sleepers if there's too much bonus 'in flight' | |
541 | * already: | |
542 | */ | |
543 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
544 | return; | |
545 | ||
e59c80c5 | 546 | if (sched_feat(SLEEPER_LOAD_AVG)) |
bf0f6f24 IM |
547 | load = rq_of(cfs_rq)->cpu_load[2]; |
548 | ||
bf0f6f24 IM |
549 | /* |
550 | * Fix up delta_fair with the effect of us running | |
551 | * during the whole sleep period: | |
552 | */ | |
e59c80c5 | 553 | if (sched_feat(SLEEPER_AVG)) |
bf0f6f24 IM |
554 | delta_fair = div64_likely32((u64)delta_fair * load, |
555 | load + se->load.weight); | |
556 | ||
557 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
558 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
559 | NICE_0_SHIFT); | |
560 | ||
561 | prev_runtime = se->wait_runtime; | |
562 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
563 | delta_fair = se->wait_runtime - prev_runtime; | |
564 | ||
565 | /* | |
566 | * Track the amount of bonus we've given to sleepers: | |
567 | */ | |
568 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
569 | } |
570 | ||
2396af69 | 571 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
572 | { |
573 | struct task_struct *tsk = task_of(se); | |
574 | unsigned long delta_fair; | |
575 | ||
576 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
e59c80c5 | 577 | !sched_feat(FAIR_SLEEPERS)) |
bf0f6f24 IM |
578 | return; |
579 | ||
580 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
581 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
582 | ||
8ebc91d9 | 583 | __enqueue_sleeper(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
584 | |
585 | se->sleep_start_fair = 0; | |
586 | ||
587 | #ifdef CONFIG_SCHEDSTATS | |
588 | if (se->sleep_start) { | |
d281918d | 589 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
590 | |
591 | if ((s64)delta < 0) | |
592 | delta = 0; | |
593 | ||
594 | if (unlikely(delta > se->sleep_max)) | |
595 | se->sleep_max = delta; | |
596 | ||
597 | se->sleep_start = 0; | |
598 | se->sum_sleep_runtime += delta; | |
599 | } | |
600 | if (se->block_start) { | |
d281918d | 601 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
602 | |
603 | if ((s64)delta < 0) | |
604 | delta = 0; | |
605 | ||
606 | if (unlikely(delta > se->block_max)) | |
607 | se->block_max = delta; | |
608 | ||
609 | se->block_start = 0; | |
610 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
611 | |
612 | /* | |
613 | * Blocking time is in units of nanosecs, so shift by 20 to | |
614 | * get a milliseconds-range estimation of the amount of | |
615 | * time that the task spent sleeping: | |
616 | */ | |
617 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
618 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
619 | delta >> 20); | |
620 | } | |
bf0f6f24 IM |
621 | } |
622 | #endif | |
623 | } | |
624 | ||
625 | static void | |
668031ca | 626 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
627 | { |
628 | /* | |
629 | * Update the fair clock. | |
630 | */ | |
b7cc0896 | 631 | update_curr(cfs_rq); |
bf0f6f24 IM |
632 | |
633 | if (wakeup) | |
2396af69 | 634 | enqueue_sleeper(cfs_rq, se); |
bf0f6f24 | 635 | |
d2417e5a | 636 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
637 | __enqueue_entity(cfs_rq, se); |
638 | } | |
639 | ||
640 | static void | |
525c2716 | 641 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 642 | { |
19b6a2e3 | 643 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
644 | if (sleep) { |
645 | se->sleep_start_fair = cfs_rq->fair_clock; | |
646 | #ifdef CONFIG_SCHEDSTATS | |
647 | if (entity_is_task(se)) { | |
648 | struct task_struct *tsk = task_of(se); | |
649 | ||
650 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 651 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 652 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 653 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 654 | } |
bf0f6f24 IM |
655 | #endif |
656 | } | |
657 | __dequeue_entity(cfs_rq, se); | |
658 | } | |
659 | ||
660 | /* | |
661 | * Preempt the current task with a newly woken task if needed: | |
662 | */ | |
7c92e54f | 663 | static void |
bf0f6f24 IM |
664 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
665 | struct sched_entity *curr, unsigned long granularity) | |
666 | { | |
667 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
668 | unsigned long ideal_runtime, delta_exec; |
669 | ||
670 | /* | |
671 | * ideal_runtime is compared against sum_exec_runtime, which is | |
672 | * walltime, hence do not scale. | |
673 | */ | |
674 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
675 | (unsigned long)sysctl_sched_min_granularity); | |
676 | ||
677 | /* | |
678 | * If we executed more than what the latency constraint suggests, | |
679 | * reduce the rescheduling granularity. This way the total latency | |
680 | * of how much a task is not scheduled converges to | |
681 | * sysctl_sched_latency: | |
682 | */ | |
683 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
684 | if (delta_exec > ideal_runtime) | |
685 | granularity = 0; | |
bf0f6f24 IM |
686 | |
687 | /* | |
688 | * Take scheduling granularity into account - do not | |
689 | * preempt the current task unless the best task has | |
690 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
691 | * |
692 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 693 | */ |
4a55b450 | 694 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
695 | resched_task(rq_of(cfs_rq)->curr); |
696 | } | |
697 | ||
698 | static inline void | |
8494f412 | 699 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
700 | { |
701 | /* | |
702 | * Any task has to be enqueued before it get to execute on | |
703 | * a CPU. So account for the time it spent waiting on the | |
704 | * runqueue. (note, here we rely on pick_next_task() having | |
705 | * done a put_prev_task_fair() shortly before this, which | |
706 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
707 | */ | |
9ef0a961 | 708 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 709 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 710 | cfs_rq->curr = se; |
eba1ed4b IM |
711 | #ifdef CONFIG_SCHEDSTATS |
712 | /* | |
713 | * Track our maximum slice length, if the CPU's load is at | |
714 | * least twice that of our own weight (i.e. dont track it | |
715 | * when there are only lesser-weight tasks around): | |
716 | */ | |
717 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
718 | se->slice_max = max(se->slice_max, | |
719 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
720 | } | |
721 | #endif | |
4a55b450 | 722 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
723 | } |
724 | ||
9948f4b2 | 725 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
726 | { |
727 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
728 | ||
8494f412 | 729 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
730 | |
731 | return se; | |
732 | } | |
733 | ||
ab6cde26 | 734 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
735 | { |
736 | /* | |
737 | * If still on the runqueue then deactivate_task() | |
738 | * was not called and update_curr() has to be done: | |
739 | */ | |
740 | if (prev->on_rq) | |
b7cc0896 | 741 | update_curr(cfs_rq); |
bf0f6f24 | 742 | |
c7e9b5b2 | 743 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
744 | |
745 | if (prev->on_rq) | |
5870db5b | 746 | update_stats_wait_start(cfs_rq, prev); |
429d43bc | 747 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
748 | } |
749 | ||
750 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
751 | { | |
bf0f6f24 | 752 | struct sched_entity *next; |
c1b3da3e | 753 | |
bf0f6f24 IM |
754 | /* |
755 | * Dequeue and enqueue the task to update its | |
756 | * position within the tree: | |
757 | */ | |
525c2716 | 758 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 759 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
760 | |
761 | /* | |
762 | * Reschedule if another task tops the current one. | |
763 | */ | |
764 | next = __pick_next_entity(cfs_rq); | |
765 | if (next == curr) | |
766 | return; | |
767 | ||
11697830 PZ |
768 | __check_preempt_curr_fair(cfs_rq, next, curr, |
769 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
770 | } |
771 | ||
772 | /************************************************** | |
773 | * CFS operations on tasks: | |
774 | */ | |
775 | ||
776 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
777 | ||
778 | /* Walk up scheduling entities hierarchy */ | |
779 | #define for_each_sched_entity(se) \ | |
780 | for (; se; se = se->parent) | |
781 | ||
782 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
783 | { | |
784 | return p->se.cfs_rq; | |
785 | } | |
786 | ||
787 | /* runqueue on which this entity is (to be) queued */ | |
788 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
789 | { | |
790 | return se->cfs_rq; | |
791 | } | |
792 | ||
793 | /* runqueue "owned" by this group */ | |
794 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
795 | { | |
796 | return grp->my_q; | |
797 | } | |
798 | ||
799 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
800 | * another cpu ('this_cpu') | |
801 | */ | |
802 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
803 | { | |
804 | /* A later patch will take group into account */ | |
805 | return &cpu_rq(this_cpu)->cfs; | |
806 | } | |
807 | ||
808 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
809 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
810 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
811 | ||
812 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
813 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
814 | { | |
815 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
816 | return 1; | |
817 | ||
818 | return 0; | |
819 | } | |
820 | ||
821 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
822 | ||
823 | #define for_each_sched_entity(se) \ | |
824 | for (; se; se = NULL) | |
825 | ||
826 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
827 | { | |
828 | return &task_rq(p)->cfs; | |
829 | } | |
830 | ||
831 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
832 | { | |
833 | struct task_struct *p = task_of(se); | |
834 | struct rq *rq = task_rq(p); | |
835 | ||
836 | return &rq->cfs; | |
837 | } | |
838 | ||
839 | /* runqueue "owned" by this group */ | |
840 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
841 | { | |
842 | return NULL; | |
843 | } | |
844 | ||
845 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
846 | { | |
847 | return &cpu_rq(this_cpu)->cfs; | |
848 | } | |
849 | ||
850 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
851 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
852 | ||
853 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
854 | { | |
855 | return 1; | |
856 | } | |
857 | ||
858 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
859 | ||
860 | /* | |
861 | * The enqueue_task method is called before nr_running is | |
862 | * increased. Here we update the fair scheduling stats and | |
863 | * then put the task into the rbtree: | |
864 | */ | |
fd390f6a | 865 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
866 | { |
867 | struct cfs_rq *cfs_rq; | |
868 | struct sched_entity *se = &p->se; | |
869 | ||
870 | for_each_sched_entity(se) { | |
871 | if (se->on_rq) | |
872 | break; | |
873 | cfs_rq = cfs_rq_of(se); | |
668031ca | 874 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
875 | } |
876 | } | |
877 | ||
878 | /* | |
879 | * The dequeue_task method is called before nr_running is | |
880 | * decreased. We remove the task from the rbtree and | |
881 | * update the fair scheduling stats: | |
882 | */ | |
f02231e5 | 883 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
884 | { |
885 | struct cfs_rq *cfs_rq; | |
886 | struct sched_entity *se = &p->se; | |
887 | ||
888 | for_each_sched_entity(se) { | |
889 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 890 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
891 | /* Don't dequeue parent if it has other entities besides us */ |
892 | if (cfs_rq->load.weight) | |
893 | break; | |
894 | } | |
895 | } | |
896 | ||
897 | /* | |
1799e35d IM |
898 | * sched_yield() support is very simple - we dequeue and enqueue. |
899 | * | |
900 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
901 | */ |
902 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
903 | { | |
904 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
905 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
906 | struct sched_entity *rightmost, *se = &p->se; | |
907 | struct rb_node *parent; | |
bf0f6f24 IM |
908 | |
909 | /* | |
1799e35d IM |
910 | * Are we the only task in the tree? |
911 | */ | |
912 | if (unlikely(cfs_rq->nr_running == 1)) | |
913 | return; | |
914 | ||
915 | if (likely(!sysctl_sched_compat_yield)) { | |
916 | __update_rq_clock(rq); | |
917 | /* | |
918 | * Dequeue and enqueue the task to update its | |
919 | * position within the tree: | |
920 | */ | |
921 | dequeue_entity(cfs_rq, &p->se, 0); | |
922 | enqueue_entity(cfs_rq, &p->se, 0); | |
923 | ||
924 | return; | |
925 | } | |
926 | /* | |
927 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 928 | */ |
1799e35d IM |
929 | do { |
930 | parent = *link; | |
931 | link = &parent->rb_right; | |
932 | } while (*link); | |
933 | ||
934 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
935 | /* | |
936 | * Already in the rightmost position? | |
937 | */ | |
938 | if (unlikely(rightmost == se)) | |
939 | return; | |
940 | ||
941 | /* | |
942 | * Minimally necessary key value to be last in the tree: | |
943 | */ | |
944 | se->fair_key = rightmost->fair_key + 1; | |
945 | ||
946 | if (cfs_rq->rb_leftmost == &se->run_node) | |
947 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
948 | /* | |
949 | * Relink the task to the rightmost position: | |
950 | */ | |
951 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
952 | rb_link_node(&se->run_node, parent, link); | |
953 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
954 | } |
955 | ||
956 | /* | |
957 | * Preempt the current task with a newly woken task if needed: | |
958 | */ | |
959 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
960 | { | |
961 | struct task_struct *curr = rq->curr; | |
962 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
963 | unsigned long gran; | |
964 | ||
965 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 966 | update_rq_clock(rq); |
b7cc0896 | 967 | update_curr(cfs_rq); |
bf0f6f24 IM |
968 | resched_task(curr); |
969 | return; | |
970 | } | |
971 | ||
972 | gran = sysctl_sched_wakeup_granularity; | |
973 | /* | |
974 | * Batch tasks prefer throughput over latency: | |
975 | */ | |
976 | if (unlikely(p->policy == SCHED_BATCH)) | |
977 | gran = sysctl_sched_batch_wakeup_granularity; | |
978 | ||
979 | if (is_same_group(curr, p)) | |
980 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
981 | } | |
982 | ||
fb8d4724 | 983 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
984 | { |
985 | struct cfs_rq *cfs_rq = &rq->cfs; | |
986 | struct sched_entity *se; | |
987 | ||
988 | if (unlikely(!cfs_rq->nr_running)) | |
989 | return NULL; | |
990 | ||
991 | do { | |
9948f4b2 | 992 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
993 | cfs_rq = group_cfs_rq(se); |
994 | } while (cfs_rq); | |
995 | ||
996 | return task_of(se); | |
997 | } | |
998 | ||
999 | /* | |
1000 | * Account for a descheduled task: | |
1001 | */ | |
31ee529c | 1002 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1003 | { |
1004 | struct sched_entity *se = &prev->se; | |
1005 | struct cfs_rq *cfs_rq; | |
1006 | ||
1007 | for_each_sched_entity(se) { | |
1008 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1009 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1010 | } |
1011 | } | |
1012 | ||
1013 | /************************************************** | |
1014 | * Fair scheduling class load-balancing methods: | |
1015 | */ | |
1016 | ||
1017 | /* | |
1018 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1019 | * during the whole iteration, the current task might be | |
1020 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1021 | * achieve that by always pre-iterating before returning | |
1022 | * the current task: | |
1023 | */ | |
1024 | static inline struct task_struct * | |
1025 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1026 | { | |
1027 | struct task_struct *p; | |
1028 | ||
1029 | if (!curr) | |
1030 | return NULL; | |
1031 | ||
1032 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1033 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1034 | ||
1035 | return p; | |
1036 | } | |
1037 | ||
1038 | static struct task_struct *load_balance_start_fair(void *arg) | |
1039 | { | |
1040 | struct cfs_rq *cfs_rq = arg; | |
1041 | ||
1042 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1043 | } | |
1044 | ||
1045 | static struct task_struct *load_balance_next_fair(void *arg) | |
1046 | { | |
1047 | struct cfs_rq *cfs_rq = arg; | |
1048 | ||
1049 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1050 | } | |
1051 | ||
a4ac01c3 | 1052 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1053 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1054 | { | |
1055 | struct sched_entity *curr; | |
1056 | struct task_struct *p; | |
1057 | ||
1058 | if (!cfs_rq->nr_running) | |
1059 | return MAX_PRIO; | |
1060 | ||
1061 | curr = __pick_next_entity(cfs_rq); | |
1062 | p = task_of(curr); | |
1063 | ||
1064 | return p->prio; | |
1065 | } | |
a4ac01c3 | 1066 | #endif |
bf0f6f24 | 1067 | |
43010659 | 1068 | static unsigned long |
bf0f6f24 | 1069 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1070 | unsigned long max_nr_move, unsigned long max_load_move, |
1071 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1072 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1073 | { |
1074 | struct cfs_rq *busy_cfs_rq; | |
1075 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1076 | long rem_load_move = max_load_move; | |
1077 | struct rq_iterator cfs_rq_iterator; | |
1078 | ||
1079 | cfs_rq_iterator.start = load_balance_start_fair; | |
1080 | cfs_rq_iterator.next = load_balance_next_fair; | |
1081 | ||
1082 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1083 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1084 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1085 | long imbalance; |
bf0f6f24 | 1086 | unsigned long maxload; |
bf0f6f24 IM |
1087 | |
1088 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1089 | ||
e56f31aa | 1090 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1091 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1092 | if (imbalance <= 0) | |
1093 | continue; | |
1094 | ||
1095 | /* Don't pull more than imbalance/2 */ | |
1096 | imbalance /= 2; | |
1097 | maxload = min(rem_load_move, imbalance); | |
1098 | ||
a4ac01c3 PW |
1099 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1100 | #else | |
e56f31aa | 1101 | # define maxload rem_load_move |
a4ac01c3 | 1102 | #endif |
bf0f6f24 IM |
1103 | /* pass busy_cfs_rq argument into |
1104 | * load_balance_[start|next]_fair iterators | |
1105 | */ | |
1106 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1107 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1108 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1109 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1110 | |
1111 | total_nr_moved += nr_moved; | |
1112 | max_nr_move -= nr_moved; | |
1113 | rem_load_move -= load_moved; | |
1114 | ||
1115 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1116 | break; | |
1117 | } | |
1118 | ||
43010659 | 1119 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1120 | } |
1121 | ||
1122 | /* | |
1123 | * scheduler tick hitting a task of our scheduling class: | |
1124 | */ | |
1125 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1126 | { | |
1127 | struct cfs_rq *cfs_rq; | |
1128 | struct sched_entity *se = &curr->se; | |
1129 | ||
1130 | for_each_sched_entity(se) { | |
1131 | cfs_rq = cfs_rq_of(se); | |
1132 | entity_tick(cfs_rq, se); | |
1133 | } | |
1134 | } | |
1135 | ||
1136 | /* | |
1137 | * Share the fairness runtime between parent and child, thus the | |
1138 | * total amount of pressure for CPU stays equal - new tasks | |
1139 | * get a chance to run but frequent forkers are not allowed to | |
1140 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1141 | * the child is not running yet. | |
1142 | */ | |
ee0827d8 | 1143 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1144 | { |
1145 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1146 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
bf0f6f24 IM |
1147 | |
1148 | sched_info_queued(p); | |
1149 | ||
7109c442 | 1150 | update_curr(cfs_rq); |
d2417e5a | 1151 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1152 | /* |
1153 | * Child runs first: we let it run before the parent | |
1154 | * until it reschedules once. We set up the key so that | |
1155 | * it will preempt the parent: | |
1156 | */ | |
9f508f82 | 1157 | se->fair_key = curr->fair_key - |
7109c442 | 1158 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1159 | /* |
1160 | * The first wait is dominated by the child-runs-first logic, | |
1161 | * so do not credit it with that waiting time yet: | |
1162 | */ | |
e59c80c5 | 1163 | if (sched_feat(SKIP_INITIAL)) |
9f508f82 | 1164 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1165 | |
1166 | /* | |
1167 | * The statistical average of wait_runtime is about | |
1168 | * -granularity/2, so initialize the task with that: | |
1169 | */ | |
e59c80c5 | 1170 | if (sched_feat(START_DEBIT)) |
9f508f82 | 1171 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 IM |
1172 | |
1173 | __enqueue_entity(cfs_rq, se); | |
bb61c210 | 1174 | resched_task(rq->curr); |
bf0f6f24 IM |
1175 | } |
1176 | ||
1177 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1178 | /* Account for a task changing its policy or group. | |
1179 | * | |
1180 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1181 | * migrates between groups/classes. | |
1182 | */ | |
1183 | static void set_curr_task_fair(struct rq *rq) | |
1184 | { | |
7c6c16f3 | 1185 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1186 | |
c3b64f1e IM |
1187 | for_each_sched_entity(se) |
1188 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1189 | } |
1190 | #else | |
1191 | static void set_curr_task_fair(struct rq *rq) | |
1192 | { | |
1193 | } | |
1194 | #endif | |
1195 | ||
1196 | /* | |
1197 | * All the scheduling class methods: | |
1198 | */ | |
1199 | struct sched_class fair_sched_class __read_mostly = { | |
1200 | .enqueue_task = enqueue_task_fair, | |
1201 | .dequeue_task = dequeue_task_fair, | |
1202 | .yield_task = yield_task_fair, | |
1203 | ||
1204 | .check_preempt_curr = check_preempt_curr_fair, | |
1205 | ||
1206 | .pick_next_task = pick_next_task_fair, | |
1207 | .put_prev_task = put_prev_task_fair, | |
1208 | ||
1209 | .load_balance = load_balance_fair, | |
1210 | ||
1211 | .set_curr_task = set_curr_task_fair, | |
1212 | .task_tick = task_tick_fair, | |
1213 | .task_new = task_new_fair, | |
1214 | }; | |
1215 | ||
1216 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1217 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1218 | { |
bf0f6f24 IM |
1219 | struct cfs_rq *cfs_rq; |
1220 | ||
c3b64f1e | 1221 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1222 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1223 | } |
1224 | #endif |