]>
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 | ||
108 | extern struct sched_class fair_sched_class; | |
109 | ||
110 | /************************************************************** | |
111 | * CFS operations on generic schedulable entities: | |
112 | */ | |
113 | ||
62160e3f | 114 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 115 | |
62160e3f | 116 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
117 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
118 | { | |
62160e3f | 119 | return cfs_rq->rq; |
bf0f6f24 IM |
120 | } |
121 | ||
62160e3f IM |
122 | /* An entity is a task if it doesn't "own" a runqueue */ |
123 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 124 | |
62160e3f | 125 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 126 | |
62160e3f IM |
127 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
128 | { | |
129 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
130 | } |
131 | ||
132 | #define entity_is_task(se) 1 | |
133 | ||
bf0f6f24 IM |
134 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
135 | ||
136 | static inline struct task_struct *task_of(struct sched_entity *se) | |
137 | { | |
138 | return container_of(se, struct task_struct, se); | |
139 | } | |
140 | ||
141 | ||
142 | /************************************************************** | |
143 | * Scheduling class tree data structure manipulation methods: | |
144 | */ | |
145 | ||
146 | /* | |
147 | * Enqueue an entity into the rb-tree: | |
148 | */ | |
149 | static inline void | |
150 | __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
151 | { | |
152 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
153 | struct rb_node *parent = NULL; | |
154 | struct sched_entity *entry; | |
155 | s64 key = se->fair_key; | |
156 | int leftmost = 1; | |
157 | ||
158 | /* | |
159 | * Find the right place in the rbtree: | |
160 | */ | |
161 | while (*link) { | |
162 | parent = *link; | |
163 | entry = rb_entry(parent, struct sched_entity, run_node); | |
164 | /* | |
165 | * We dont care about collisions. Nodes with | |
166 | * the same key stay together. | |
167 | */ | |
168 | if (key - entry->fair_key < 0) { | |
169 | link = &parent->rb_left; | |
170 | } else { | |
171 | link = &parent->rb_right; | |
172 | leftmost = 0; | |
173 | } | |
174 | } | |
175 | ||
176 | /* | |
177 | * Maintain a cache of leftmost tree entries (it is frequently | |
178 | * used): | |
179 | */ | |
180 | if (leftmost) | |
181 | cfs_rq->rb_leftmost = &se->run_node; | |
182 | ||
183 | rb_link_node(&se->run_node, parent, link); | |
184 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
185 | update_load_add(&cfs_rq->load, se->load.weight); | |
186 | cfs_rq->nr_running++; | |
187 | se->on_rq = 1; | |
a206c072 IM |
188 | |
189 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
bf0f6f24 IM |
190 | } |
191 | ||
192 | static inline void | |
193 | __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
194 | { | |
195 | if (cfs_rq->rb_leftmost == &se->run_node) | |
196 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
197 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
198 | update_load_sub(&cfs_rq->load, se->load.weight); | |
199 | cfs_rq->nr_running--; | |
200 | se->on_rq = 0; | |
a206c072 IM |
201 | |
202 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
bf0f6f24 IM |
203 | } |
204 | ||
205 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
206 | { | |
207 | return cfs_rq->rb_leftmost; | |
208 | } | |
209 | ||
210 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
211 | { | |
212 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
213 | } | |
214 | ||
215 | /************************************************************** | |
216 | * Scheduling class statistics methods: | |
217 | */ | |
218 | ||
21805085 PZ |
219 | /* |
220 | * Calculate the preemption granularity needed to schedule every | |
221 | * runnable task once per sysctl_sched_latency amount of time. | |
222 | * (down to a sensible low limit on granularity) | |
223 | * | |
224 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
225 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
226 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
227 | * for each task. We do finer and finer scheduling up to until we | |
228 | * reach the minimum granularity value. | |
229 | * | |
230 | * To achieve this we use the following dynamic-granularity rule: | |
231 | * | |
232 | * gran = lat/nr - lat/nr/nr | |
233 | * | |
234 | * This comes out of the following equations: | |
235 | * | |
236 | * kA1 + gran = kB1 | |
237 | * kB2 + gran = kA2 | |
238 | * kA2 = kA1 | |
239 | * kB2 = kB1 - d + d/nr | |
240 | * lat = d * nr | |
241 | * | |
242 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
243 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
244 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
245 | * running, 'lat' is the the period of each task. ('lat' is the | |
246 | * sched_latency that we aim for.) | |
247 | */ | |
248 | static long | |
249 | sched_granularity(struct cfs_rq *cfs_rq) | |
250 | { | |
251 | unsigned int gran = sysctl_sched_latency; | |
252 | unsigned int nr = cfs_rq->nr_running; | |
253 | ||
254 | if (nr > 1) { | |
255 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 256 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
257 | } |
258 | ||
259 | return gran; | |
260 | } | |
261 | ||
bf0f6f24 IM |
262 | /* |
263 | * We rescale the rescheduling granularity of tasks according to their | |
264 | * nice level, but only linearly, not exponentially: | |
265 | */ | |
266 | static long | |
267 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
268 | { | |
269 | u64 tmp; | |
270 | ||
7cff8cf6 IM |
271 | if (likely(curr->load.weight == NICE_0_LOAD)) |
272 | return granularity; | |
bf0f6f24 | 273 | /* |
7cff8cf6 | 274 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 275 | */ |
7cff8cf6 IM |
276 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
277 | tmp = curr->load.weight * (u64)granularity; | |
278 | return (long) (tmp >> NICE_0_SHIFT); | |
279 | } | |
bf0f6f24 | 280 | /* |
7cff8cf6 | 281 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
282 | * granularity: |
283 | */ | |
7cff8cf6 | 284 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
285 | |
286 | /* | |
287 | * It will always fit into 'long': | |
288 | */ | |
a0dc7260 | 289 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
290 | } |
291 | ||
292 | static inline void | |
293 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
294 | { | |
295 | long limit = sysctl_sched_runtime_limit; | |
296 | ||
297 | /* | |
298 | * Niced tasks have the same history dynamic range as | |
299 | * non-niced tasks: | |
300 | */ | |
301 | if (unlikely(se->wait_runtime > limit)) { | |
302 | se->wait_runtime = limit; | |
303 | schedstat_inc(se, wait_runtime_overruns); | |
304 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
305 | } | |
306 | if (unlikely(se->wait_runtime < -limit)) { | |
307 | se->wait_runtime = -limit; | |
308 | schedstat_inc(se, wait_runtime_underruns); | |
309 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
310 | } | |
311 | } | |
312 | ||
313 | static inline void | |
314 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
315 | { | |
316 | se->wait_runtime += delta; | |
317 | schedstat_add(se, sum_wait_runtime, delta); | |
318 | limit_wait_runtime(cfs_rq, se); | |
319 | } | |
320 | ||
321 | static void | |
322 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
323 | { | |
324 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
325 | __add_wait_runtime(cfs_rq, se, delta); | |
326 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
327 | } | |
328 | ||
329 | /* | |
330 | * Update the current task's runtime statistics. Skip current tasks that | |
331 | * are not in our scheduling class. | |
332 | */ | |
333 | static inline void | |
8ebc91d9 IM |
334 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
335 | unsigned long delta_exec) | |
bf0f6f24 | 336 | { |
8ebc91d9 | 337 | unsigned long delta, delta_fair, delta_mine; |
bf0f6f24 IM |
338 | struct load_weight *lw = &cfs_rq->load; |
339 | unsigned long load = lw->weight; | |
340 | ||
8179ca23 | 341 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
342 | |
343 | curr->sum_exec_runtime += delta_exec; | |
344 | cfs_rq->exec_clock += delta_exec; | |
345 | ||
fd8bb43e IM |
346 | if (unlikely(!load)) |
347 | return; | |
348 | ||
bf0f6f24 IM |
349 | delta_fair = calc_delta_fair(delta_exec, lw); |
350 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
351 | ||
5f01d519 | 352 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 353 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
354 | delta = min(delta, (unsigned long)( |
355 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
356 | cfs_rq->sleeper_bonus -= delta; |
357 | delta_mine -= delta; | |
358 | } | |
359 | ||
360 | cfs_rq->fair_clock += delta_fair; | |
361 | /* | |
362 | * We executed delta_exec amount of time on the CPU, | |
363 | * but we were only entitled to delta_mine amount of | |
364 | * time during that period (if nr_running == 1 then | |
365 | * the two values are equal) | |
366 | * [Note: delta_mine - delta_exec is negative]: | |
367 | */ | |
368 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
369 | } | |
370 | ||
b7cc0896 | 371 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 372 | { |
429d43bc | 373 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 374 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
375 | unsigned long delta_exec; |
376 | ||
377 | if (unlikely(!curr)) | |
378 | return; | |
379 | ||
380 | /* | |
381 | * Get the amount of time the current task was running | |
382 | * since the last time we changed load (this cannot | |
383 | * overflow on 32 bits): | |
384 | */ | |
8ebc91d9 | 385 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 386 | |
8ebc91d9 IM |
387 | __update_curr(cfs_rq, curr, delta_exec); |
388 | curr->exec_start = now; | |
bf0f6f24 IM |
389 | } |
390 | ||
391 | static inline void | |
5870db5b | 392 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
393 | { |
394 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 395 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
396 | } |
397 | ||
398 | /* | |
399 | * We calculate fair deltas here, so protect against the random effects | |
400 | * of a multiplication overflow by capping it to the runtime limit: | |
401 | */ | |
402 | #if BITS_PER_LONG == 32 | |
403 | static inline unsigned long | |
404 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
405 | { | |
406 | u64 tmp = (u64)delta * weight >> shift; | |
407 | ||
408 | if (unlikely(tmp > sysctl_sched_runtime_limit*2)) | |
409 | return sysctl_sched_runtime_limit*2; | |
410 | return tmp; | |
411 | } | |
412 | #else | |
413 | static inline unsigned long | |
414 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
415 | { | |
416 | return delta * weight >> shift; | |
417 | } | |
418 | #endif | |
419 | ||
420 | /* | |
421 | * Task is being enqueued - update stats: | |
422 | */ | |
d2417e5a | 423 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
424 | { |
425 | s64 key; | |
426 | ||
427 | /* | |
428 | * Are we enqueueing a waiting task? (for current tasks | |
429 | * a dequeue/enqueue event is a NOP) | |
430 | */ | |
429d43bc | 431 | if (se != cfs_rq->curr) |
5870db5b | 432 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
433 | /* |
434 | * Update the key: | |
435 | */ | |
436 | key = cfs_rq->fair_clock; | |
437 | ||
438 | /* | |
439 | * Optimize the common nice 0 case: | |
440 | */ | |
441 | if (likely(se->load.weight == NICE_0_LOAD)) { | |
442 | key -= se->wait_runtime; | |
443 | } else { | |
444 | u64 tmp; | |
445 | ||
446 | if (se->wait_runtime < 0) { | |
447 | tmp = -se->wait_runtime; | |
448 | key += (tmp * se->load.inv_weight) >> | |
449 | (WMULT_SHIFT - NICE_0_SHIFT); | |
450 | } else { | |
451 | tmp = se->wait_runtime; | |
a69edb55 IM |
452 | key -= (tmp * se->load.inv_weight) >> |
453 | (WMULT_SHIFT - NICE_0_SHIFT); | |
bf0f6f24 IM |
454 | } |
455 | } | |
456 | ||
457 | se->fair_key = key; | |
458 | } | |
459 | ||
460 | /* | |
461 | * Note: must be called with a freshly updated rq->fair_clock. | |
462 | */ | |
463 | static inline void | |
8ebc91d9 IM |
464 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, |
465 | unsigned long delta_fair) | |
bf0f6f24 | 466 | { |
d281918d IM |
467 | schedstat_set(se->wait_max, max(se->wait_max, |
468 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 IM |
469 | |
470 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
471 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
472 | NICE_0_SHIFT); | |
473 | ||
474 | add_wait_runtime(cfs_rq, se, delta_fair); | |
475 | } | |
476 | ||
477 | static void | |
9ef0a961 | 478 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
479 | { |
480 | unsigned long delta_fair; | |
481 | ||
b77d69db IM |
482 | if (unlikely(!se->wait_start_fair)) |
483 | return; | |
484 | ||
bf0f6f24 IM |
485 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
486 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
487 | ||
8ebc91d9 | 488 | __update_stats_wait_end(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
489 | |
490 | se->wait_start_fair = 0; | |
6cfb0d5d | 491 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
492 | } |
493 | ||
494 | static inline void | |
19b6a2e3 | 495 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 496 | { |
b7cc0896 | 497 | update_curr(cfs_rq); |
bf0f6f24 IM |
498 | /* |
499 | * Mark the end of the wait period if dequeueing a | |
500 | * waiting task: | |
501 | */ | |
429d43bc | 502 | if (se != cfs_rq->curr) |
9ef0a961 | 503 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
504 | } |
505 | ||
506 | /* | |
507 | * We are picking a new current task - update its stats: | |
508 | */ | |
509 | static inline void | |
79303e9e | 510 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
511 | { |
512 | /* | |
513 | * We are starting a new run period: | |
514 | */ | |
d281918d | 515 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
516 | } |
517 | ||
518 | /* | |
519 | * We are descheduling a task - update its stats: | |
520 | */ | |
521 | static inline void | |
c7e9b5b2 | 522 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
523 | { |
524 | se->exec_start = 0; | |
525 | } | |
526 | ||
527 | /************************************************** | |
528 | * Scheduling class queueing methods: | |
529 | */ | |
530 | ||
8ebc91d9 IM |
531 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, |
532 | unsigned long delta_fair) | |
bf0f6f24 | 533 | { |
8ebc91d9 | 534 | unsigned long load = cfs_rq->load.weight; |
bf0f6f24 IM |
535 | long prev_runtime; |
536 | ||
b2133c8b IM |
537 | /* |
538 | * Do not boost sleepers if there's too much bonus 'in flight' | |
539 | * already: | |
540 | */ | |
541 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
542 | return; | |
543 | ||
bf0f6f24 IM |
544 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG) |
545 | load = rq_of(cfs_rq)->cpu_load[2]; | |
546 | ||
bf0f6f24 IM |
547 | /* |
548 | * Fix up delta_fair with the effect of us running | |
549 | * during the whole sleep period: | |
550 | */ | |
551 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG) | |
552 | delta_fair = div64_likely32((u64)delta_fair * load, | |
553 | load + se->load.weight); | |
554 | ||
555 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
556 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
557 | NICE_0_SHIFT); | |
558 | ||
559 | prev_runtime = se->wait_runtime; | |
560 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
561 | delta_fair = se->wait_runtime - prev_runtime; | |
562 | ||
563 | /* | |
564 | * Track the amount of bonus we've given to sleepers: | |
565 | */ | |
566 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
567 | } |
568 | ||
2396af69 | 569 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
570 | { |
571 | struct task_struct *tsk = task_of(se); | |
572 | unsigned long delta_fair; | |
573 | ||
574 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
575 | !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS)) | |
576 | return; | |
577 | ||
578 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
579 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
580 | ||
8ebc91d9 | 581 | __enqueue_sleeper(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
582 | |
583 | se->sleep_start_fair = 0; | |
584 | ||
585 | #ifdef CONFIG_SCHEDSTATS | |
586 | if (se->sleep_start) { | |
d281918d | 587 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
588 | |
589 | if ((s64)delta < 0) | |
590 | delta = 0; | |
591 | ||
592 | if (unlikely(delta > se->sleep_max)) | |
593 | se->sleep_max = delta; | |
594 | ||
595 | se->sleep_start = 0; | |
596 | se->sum_sleep_runtime += delta; | |
597 | } | |
598 | if (se->block_start) { | |
d281918d | 599 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
600 | |
601 | if ((s64)delta < 0) | |
602 | delta = 0; | |
603 | ||
604 | if (unlikely(delta > se->block_max)) | |
605 | se->block_max = delta; | |
606 | ||
607 | se->block_start = 0; | |
608 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
609 | |
610 | /* | |
611 | * Blocking time is in units of nanosecs, so shift by 20 to | |
612 | * get a milliseconds-range estimation of the amount of | |
613 | * time that the task spent sleeping: | |
614 | */ | |
615 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
616 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
617 | delta >> 20); | |
618 | } | |
bf0f6f24 IM |
619 | } |
620 | #endif | |
621 | } | |
622 | ||
623 | static void | |
668031ca | 624 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
625 | { |
626 | /* | |
627 | * Update the fair clock. | |
628 | */ | |
b7cc0896 | 629 | update_curr(cfs_rq); |
bf0f6f24 IM |
630 | |
631 | if (wakeup) | |
2396af69 | 632 | enqueue_sleeper(cfs_rq, se); |
bf0f6f24 | 633 | |
d2417e5a | 634 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
635 | __enqueue_entity(cfs_rq, se); |
636 | } | |
637 | ||
638 | static void | |
525c2716 | 639 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 640 | { |
19b6a2e3 | 641 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
642 | if (sleep) { |
643 | se->sleep_start_fair = cfs_rq->fair_clock; | |
644 | #ifdef CONFIG_SCHEDSTATS | |
645 | if (entity_is_task(se)) { | |
646 | struct task_struct *tsk = task_of(se); | |
647 | ||
648 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 649 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 650 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 651 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 652 | } |
bf0f6f24 IM |
653 | #endif |
654 | } | |
655 | __dequeue_entity(cfs_rq, se); | |
656 | } | |
657 | ||
658 | /* | |
659 | * Preempt the current task with a newly woken task if needed: | |
660 | */ | |
7c92e54f | 661 | static void |
bf0f6f24 IM |
662 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
663 | struct sched_entity *curr, unsigned long granularity) | |
664 | { | |
665 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
666 | unsigned long ideal_runtime, delta_exec; |
667 | ||
668 | /* | |
669 | * ideal_runtime is compared against sum_exec_runtime, which is | |
670 | * walltime, hence do not scale. | |
671 | */ | |
672 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
673 | (unsigned long)sysctl_sched_min_granularity); | |
674 | ||
675 | /* | |
676 | * If we executed more than what the latency constraint suggests, | |
677 | * reduce the rescheduling granularity. This way the total latency | |
678 | * of how much a task is not scheduled converges to | |
679 | * sysctl_sched_latency: | |
680 | */ | |
681 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
682 | if (delta_exec > ideal_runtime) | |
683 | granularity = 0; | |
bf0f6f24 IM |
684 | |
685 | /* | |
686 | * Take scheduling granularity into account - do not | |
687 | * preempt the current task unless the best task has | |
688 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
689 | * |
690 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 691 | */ |
4a55b450 | 692 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
693 | resched_task(rq_of(cfs_rq)->curr); |
694 | } | |
695 | ||
696 | static inline void | |
8494f412 | 697 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
698 | { |
699 | /* | |
700 | * Any task has to be enqueued before it get to execute on | |
701 | * a CPU. So account for the time it spent waiting on the | |
702 | * runqueue. (note, here we rely on pick_next_task() having | |
703 | * done a put_prev_task_fair() shortly before this, which | |
704 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
705 | */ | |
9ef0a961 | 706 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 707 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 708 | cfs_rq->curr = se; |
eba1ed4b IM |
709 | #ifdef CONFIG_SCHEDSTATS |
710 | /* | |
711 | * Track our maximum slice length, if the CPU's load is at | |
712 | * least twice that of our own weight (i.e. dont track it | |
713 | * when there are only lesser-weight tasks around): | |
714 | */ | |
715 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
716 | se->slice_max = max(se->slice_max, | |
717 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
718 | } | |
719 | #endif | |
4a55b450 | 720 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
721 | } |
722 | ||
9948f4b2 | 723 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
724 | { |
725 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
726 | ||
8494f412 | 727 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
728 | |
729 | return se; | |
730 | } | |
731 | ||
ab6cde26 | 732 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
733 | { |
734 | /* | |
735 | * If still on the runqueue then deactivate_task() | |
736 | * was not called and update_curr() has to be done: | |
737 | */ | |
738 | if (prev->on_rq) | |
b7cc0896 | 739 | update_curr(cfs_rq); |
bf0f6f24 | 740 | |
c7e9b5b2 | 741 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
742 | |
743 | if (prev->on_rq) | |
5870db5b | 744 | update_stats_wait_start(cfs_rq, prev); |
429d43bc | 745 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
746 | } |
747 | ||
748 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
749 | { | |
bf0f6f24 | 750 | struct sched_entity *next; |
c1b3da3e | 751 | |
bf0f6f24 IM |
752 | /* |
753 | * Dequeue and enqueue the task to update its | |
754 | * position within the tree: | |
755 | */ | |
525c2716 | 756 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 757 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
758 | |
759 | /* | |
760 | * Reschedule if another task tops the current one. | |
761 | */ | |
762 | next = __pick_next_entity(cfs_rq); | |
763 | if (next == curr) | |
764 | return; | |
765 | ||
11697830 PZ |
766 | __check_preempt_curr_fair(cfs_rq, next, curr, |
767 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
768 | } |
769 | ||
770 | /************************************************** | |
771 | * CFS operations on tasks: | |
772 | */ | |
773 | ||
774 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
775 | ||
776 | /* Walk up scheduling entities hierarchy */ | |
777 | #define for_each_sched_entity(se) \ | |
778 | for (; se; se = se->parent) | |
779 | ||
780 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
781 | { | |
782 | return p->se.cfs_rq; | |
783 | } | |
784 | ||
785 | /* runqueue on which this entity is (to be) queued */ | |
786 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
787 | { | |
788 | return se->cfs_rq; | |
789 | } | |
790 | ||
791 | /* runqueue "owned" by this group */ | |
792 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
793 | { | |
794 | return grp->my_q; | |
795 | } | |
796 | ||
797 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
798 | * another cpu ('this_cpu') | |
799 | */ | |
800 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
801 | { | |
802 | /* A later patch will take group into account */ | |
803 | return &cpu_rq(this_cpu)->cfs; | |
804 | } | |
805 | ||
806 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
807 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
808 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
809 | ||
810 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
811 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
812 | { | |
813 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
814 | return 1; | |
815 | ||
816 | return 0; | |
817 | } | |
818 | ||
819 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
820 | ||
821 | #define for_each_sched_entity(se) \ | |
822 | for (; se; se = NULL) | |
823 | ||
824 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
825 | { | |
826 | return &task_rq(p)->cfs; | |
827 | } | |
828 | ||
829 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
830 | { | |
831 | struct task_struct *p = task_of(se); | |
832 | struct rq *rq = task_rq(p); | |
833 | ||
834 | return &rq->cfs; | |
835 | } | |
836 | ||
837 | /* runqueue "owned" by this group */ | |
838 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
839 | { | |
840 | return NULL; | |
841 | } | |
842 | ||
843 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
844 | { | |
845 | return &cpu_rq(this_cpu)->cfs; | |
846 | } | |
847 | ||
848 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
849 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
850 | ||
851 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
852 | { | |
853 | return 1; | |
854 | } | |
855 | ||
856 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
857 | ||
858 | /* | |
859 | * The enqueue_task method is called before nr_running is | |
860 | * increased. Here we update the fair scheduling stats and | |
861 | * then put the task into the rbtree: | |
862 | */ | |
fd390f6a | 863 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
864 | { |
865 | struct cfs_rq *cfs_rq; | |
866 | struct sched_entity *se = &p->se; | |
867 | ||
868 | for_each_sched_entity(se) { | |
869 | if (se->on_rq) | |
870 | break; | |
871 | cfs_rq = cfs_rq_of(se); | |
668031ca | 872 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
873 | } |
874 | } | |
875 | ||
876 | /* | |
877 | * The dequeue_task method is called before nr_running is | |
878 | * decreased. We remove the task from the rbtree and | |
879 | * update the fair scheduling stats: | |
880 | */ | |
f02231e5 | 881 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
882 | { |
883 | struct cfs_rq *cfs_rq; | |
884 | struct sched_entity *se = &p->se; | |
885 | ||
886 | for_each_sched_entity(se) { | |
887 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 888 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
889 | /* Don't dequeue parent if it has other entities besides us */ |
890 | if (cfs_rq->load.weight) | |
891 | break; | |
892 | } | |
893 | } | |
894 | ||
895 | /* | |
1799e35d IM |
896 | * sched_yield() support is very simple - we dequeue and enqueue. |
897 | * | |
898 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
899 | */ |
900 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
901 | { | |
902 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
903 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
904 | struct sched_entity *rightmost, *se = &p->se; | |
905 | struct rb_node *parent; | |
bf0f6f24 IM |
906 | |
907 | /* | |
1799e35d IM |
908 | * Are we the only task in the tree? |
909 | */ | |
910 | if (unlikely(cfs_rq->nr_running == 1)) | |
911 | return; | |
912 | ||
913 | if (likely(!sysctl_sched_compat_yield)) { | |
914 | __update_rq_clock(rq); | |
915 | /* | |
916 | * Dequeue and enqueue the task to update its | |
917 | * position within the tree: | |
918 | */ | |
919 | dequeue_entity(cfs_rq, &p->se, 0); | |
920 | enqueue_entity(cfs_rq, &p->se, 0); | |
921 | ||
922 | return; | |
923 | } | |
924 | /* | |
925 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 926 | */ |
1799e35d IM |
927 | do { |
928 | parent = *link; | |
929 | link = &parent->rb_right; | |
930 | } while (*link); | |
931 | ||
932 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
933 | /* | |
934 | * Already in the rightmost position? | |
935 | */ | |
936 | if (unlikely(rightmost == se)) | |
937 | return; | |
938 | ||
939 | /* | |
940 | * Minimally necessary key value to be last in the tree: | |
941 | */ | |
942 | se->fair_key = rightmost->fair_key + 1; | |
943 | ||
944 | if (cfs_rq->rb_leftmost == &se->run_node) | |
945 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
946 | /* | |
947 | * Relink the task to the rightmost position: | |
948 | */ | |
949 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
950 | rb_link_node(&se->run_node, parent, link); | |
951 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
952 | } |
953 | ||
954 | /* | |
955 | * Preempt the current task with a newly woken task if needed: | |
956 | */ | |
957 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
958 | { | |
959 | struct task_struct *curr = rq->curr; | |
960 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
961 | unsigned long gran; | |
962 | ||
963 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 964 | update_rq_clock(rq); |
b7cc0896 | 965 | update_curr(cfs_rq); |
bf0f6f24 IM |
966 | resched_task(curr); |
967 | return; | |
968 | } | |
969 | ||
970 | gran = sysctl_sched_wakeup_granularity; | |
971 | /* | |
972 | * Batch tasks prefer throughput over latency: | |
973 | */ | |
974 | if (unlikely(p->policy == SCHED_BATCH)) | |
975 | gran = sysctl_sched_batch_wakeup_granularity; | |
976 | ||
977 | if (is_same_group(curr, p)) | |
978 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
979 | } | |
980 | ||
fb8d4724 | 981 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
982 | { |
983 | struct cfs_rq *cfs_rq = &rq->cfs; | |
984 | struct sched_entity *se; | |
985 | ||
986 | if (unlikely(!cfs_rq->nr_running)) | |
987 | return NULL; | |
988 | ||
989 | do { | |
9948f4b2 | 990 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
991 | cfs_rq = group_cfs_rq(se); |
992 | } while (cfs_rq); | |
993 | ||
994 | return task_of(se); | |
995 | } | |
996 | ||
997 | /* | |
998 | * Account for a descheduled task: | |
999 | */ | |
31ee529c | 1000 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1001 | { |
1002 | struct sched_entity *se = &prev->se; | |
1003 | struct cfs_rq *cfs_rq; | |
1004 | ||
1005 | for_each_sched_entity(se) { | |
1006 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1007 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1008 | } |
1009 | } | |
1010 | ||
1011 | /************************************************** | |
1012 | * Fair scheduling class load-balancing methods: | |
1013 | */ | |
1014 | ||
1015 | /* | |
1016 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1017 | * during the whole iteration, the current task might be | |
1018 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1019 | * achieve that by always pre-iterating before returning | |
1020 | * the current task: | |
1021 | */ | |
1022 | static inline struct task_struct * | |
1023 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1024 | { | |
1025 | struct task_struct *p; | |
1026 | ||
1027 | if (!curr) | |
1028 | return NULL; | |
1029 | ||
1030 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1031 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1032 | ||
1033 | return p; | |
1034 | } | |
1035 | ||
1036 | static struct task_struct *load_balance_start_fair(void *arg) | |
1037 | { | |
1038 | struct cfs_rq *cfs_rq = arg; | |
1039 | ||
1040 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1041 | } | |
1042 | ||
1043 | static struct task_struct *load_balance_next_fair(void *arg) | |
1044 | { | |
1045 | struct cfs_rq *cfs_rq = arg; | |
1046 | ||
1047 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1048 | } | |
1049 | ||
a4ac01c3 | 1050 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1051 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1052 | { | |
1053 | struct sched_entity *curr; | |
1054 | struct task_struct *p; | |
1055 | ||
1056 | if (!cfs_rq->nr_running) | |
1057 | return MAX_PRIO; | |
1058 | ||
1059 | curr = __pick_next_entity(cfs_rq); | |
1060 | p = task_of(curr); | |
1061 | ||
1062 | return p->prio; | |
1063 | } | |
a4ac01c3 | 1064 | #endif |
bf0f6f24 | 1065 | |
43010659 | 1066 | static unsigned long |
bf0f6f24 | 1067 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1068 | unsigned long max_nr_move, unsigned long max_load_move, |
1069 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1070 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1071 | { |
1072 | struct cfs_rq *busy_cfs_rq; | |
1073 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1074 | long rem_load_move = max_load_move; | |
1075 | struct rq_iterator cfs_rq_iterator; | |
1076 | ||
1077 | cfs_rq_iterator.start = load_balance_start_fair; | |
1078 | cfs_rq_iterator.next = load_balance_next_fair; | |
1079 | ||
1080 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1081 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1082 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1083 | long imbalance; |
bf0f6f24 | 1084 | unsigned long maxload; |
bf0f6f24 IM |
1085 | |
1086 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1087 | ||
e56f31aa | 1088 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1089 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1090 | if (imbalance <= 0) | |
1091 | continue; | |
1092 | ||
1093 | /* Don't pull more than imbalance/2 */ | |
1094 | imbalance /= 2; | |
1095 | maxload = min(rem_load_move, imbalance); | |
1096 | ||
a4ac01c3 PW |
1097 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1098 | #else | |
e56f31aa | 1099 | # define maxload rem_load_move |
a4ac01c3 | 1100 | #endif |
bf0f6f24 IM |
1101 | /* pass busy_cfs_rq argument into |
1102 | * load_balance_[start|next]_fair iterators | |
1103 | */ | |
1104 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1105 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1106 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1107 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1108 | |
1109 | total_nr_moved += nr_moved; | |
1110 | max_nr_move -= nr_moved; | |
1111 | rem_load_move -= load_moved; | |
1112 | ||
1113 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1114 | break; | |
1115 | } | |
1116 | ||
43010659 | 1117 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1118 | } |
1119 | ||
1120 | /* | |
1121 | * scheduler tick hitting a task of our scheduling class: | |
1122 | */ | |
1123 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1124 | { | |
1125 | struct cfs_rq *cfs_rq; | |
1126 | struct sched_entity *se = &curr->se; | |
1127 | ||
1128 | for_each_sched_entity(se) { | |
1129 | cfs_rq = cfs_rq_of(se); | |
1130 | entity_tick(cfs_rq, se); | |
1131 | } | |
1132 | } | |
1133 | ||
1134 | /* | |
1135 | * Share the fairness runtime between parent and child, thus the | |
1136 | * total amount of pressure for CPU stays equal - new tasks | |
1137 | * get a chance to run but frequent forkers are not allowed to | |
1138 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1139 | * the child is not running yet. | |
1140 | */ | |
ee0827d8 | 1141 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1142 | { |
1143 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1144 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
bf0f6f24 IM |
1145 | |
1146 | sched_info_queued(p); | |
1147 | ||
7109c442 | 1148 | update_curr(cfs_rq); |
d2417e5a | 1149 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1150 | /* |
1151 | * Child runs first: we let it run before the parent | |
1152 | * until it reschedules once. We set up the key so that | |
1153 | * it will preempt the parent: | |
1154 | */ | |
9f508f82 | 1155 | se->fair_key = curr->fair_key - |
7109c442 | 1156 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1157 | /* |
1158 | * The first wait is dominated by the child-runs-first logic, | |
1159 | * so do not credit it with that waiting time yet: | |
1160 | */ | |
1161 | if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL) | |
9f508f82 | 1162 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1163 | |
1164 | /* | |
1165 | * The statistical average of wait_runtime is about | |
1166 | * -granularity/2, so initialize the task with that: | |
1167 | */ | |
a206c072 | 1168 | if (sysctl_sched_features & SCHED_FEAT_START_DEBIT) |
9f508f82 | 1169 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 IM |
1170 | |
1171 | __enqueue_entity(cfs_rq, se); | |
bb61c210 | 1172 | resched_task(rq->curr); |
bf0f6f24 IM |
1173 | } |
1174 | ||
1175 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1176 | /* Account for a task changing its policy or group. | |
1177 | * | |
1178 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1179 | * migrates between groups/classes. | |
1180 | */ | |
1181 | static void set_curr_task_fair(struct rq *rq) | |
1182 | { | |
7c6c16f3 | 1183 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1184 | |
c3b64f1e IM |
1185 | for_each_sched_entity(se) |
1186 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1187 | } |
1188 | #else | |
1189 | static void set_curr_task_fair(struct rq *rq) | |
1190 | { | |
1191 | } | |
1192 | #endif | |
1193 | ||
1194 | /* | |
1195 | * All the scheduling class methods: | |
1196 | */ | |
1197 | struct sched_class fair_sched_class __read_mostly = { | |
1198 | .enqueue_task = enqueue_task_fair, | |
1199 | .dequeue_task = dequeue_task_fair, | |
1200 | .yield_task = yield_task_fair, | |
1201 | ||
1202 | .check_preempt_curr = check_preempt_curr_fair, | |
1203 | ||
1204 | .pick_next_task = pick_next_task_fair, | |
1205 | .put_prev_task = put_prev_task_fair, | |
1206 | ||
1207 | .load_balance = load_balance_fair, | |
1208 | ||
1209 | .set_curr_task = set_curr_task_fair, | |
1210 | .task_tick = task_tick_fair, | |
1211 | .task_new = task_new_fair, | |
1212 | }; | |
1213 | ||
1214 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1215 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1216 | { |
bf0f6f24 IM |
1217 | struct cfs_rq *cfs_rq; |
1218 | ||
c3b64f1e | 1219 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1220 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1221 | } |
1222 | #endif |