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 | |
2bd8e6d4 | 88 | const_debug unsigned int sysctl_sched_stat_granularity; |
bf0f6f24 | 89 | |
bf0f6f24 IM |
90 | unsigned int sysctl_sched_runtime_limit __read_mostly; |
91 | ||
92 | /* | |
93 | * Debugging: various feature bits | |
94 | */ | |
95 | enum { | |
96 | SCHED_FEAT_FAIR_SLEEPERS = 1, | |
97 | SCHED_FEAT_SLEEPER_AVG = 2, | |
98 | SCHED_FEAT_SLEEPER_LOAD_AVG = 4, | |
99 | SCHED_FEAT_PRECISE_CPU_LOAD = 8, | |
100 | SCHED_FEAT_START_DEBIT = 16, | |
101 | SCHED_FEAT_SKIP_INITIAL = 32, | |
102 | }; | |
103 | ||
2bd8e6d4 | 104 | const_debug unsigned int sysctl_sched_features = |
bf0f6f24 | 105 | SCHED_FEAT_FAIR_SLEEPERS *1 | |
5d2b3d36 | 106 | SCHED_FEAT_SLEEPER_AVG *0 | |
bf0f6f24 IM |
107 | SCHED_FEAT_SLEEPER_LOAD_AVG *1 | |
108 | SCHED_FEAT_PRECISE_CPU_LOAD *1 | | |
109 | SCHED_FEAT_START_DEBIT *1 | | |
110 | SCHED_FEAT_SKIP_INITIAL *0; | |
111 | ||
112 | extern struct sched_class fair_sched_class; | |
113 | ||
114 | /************************************************************** | |
115 | * CFS operations on generic schedulable entities: | |
116 | */ | |
117 | ||
118 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
119 | ||
120 | /* cpu runqueue to which this cfs_rq is attached */ | |
121 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
122 | { | |
123 | return cfs_rq->rq; | |
124 | } | |
125 | ||
126 | /* currently running entity (if any) on this cfs_rq */ | |
127 | static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq) | |
128 | { | |
129 | return cfs_rq->curr; | |
130 | } | |
131 | ||
132 | /* An entity is a task if it doesn't "own" a runqueue */ | |
133 | #define entity_is_task(se) (!se->my_q) | |
134 | ||
135 | static inline void | |
136 | set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
137 | { | |
138 | cfs_rq->curr = se; | |
139 | } | |
140 | ||
141 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
142 | ||
143 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
144 | { | |
145 | return container_of(cfs_rq, struct rq, cfs); | |
146 | } | |
147 | ||
148 | static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq) | |
149 | { | |
150 | struct rq *rq = rq_of(cfs_rq); | |
151 | ||
152 | if (unlikely(rq->curr->sched_class != &fair_sched_class)) | |
153 | return NULL; | |
154 | ||
155 | return &rq->curr->se; | |
156 | } | |
157 | ||
158 | #define entity_is_task(se) 1 | |
159 | ||
160 | static inline void | |
161 | set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
162 | ||
163 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
164 | ||
165 | static inline struct task_struct *task_of(struct sched_entity *se) | |
166 | { | |
167 | return container_of(se, struct task_struct, se); | |
168 | } | |
169 | ||
170 | ||
171 | /************************************************************** | |
172 | * Scheduling class tree data structure manipulation methods: | |
173 | */ | |
174 | ||
175 | /* | |
176 | * Enqueue an entity into the rb-tree: | |
177 | */ | |
178 | static inline void | |
179 | __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
180 | { | |
181 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
182 | struct rb_node *parent = NULL; | |
183 | struct sched_entity *entry; | |
184 | s64 key = se->fair_key; | |
185 | int leftmost = 1; | |
186 | ||
187 | /* | |
188 | * Find the right place in the rbtree: | |
189 | */ | |
190 | while (*link) { | |
191 | parent = *link; | |
192 | entry = rb_entry(parent, struct sched_entity, run_node); | |
193 | /* | |
194 | * We dont care about collisions. Nodes with | |
195 | * the same key stay together. | |
196 | */ | |
197 | if (key - entry->fair_key < 0) { | |
198 | link = &parent->rb_left; | |
199 | } else { | |
200 | link = &parent->rb_right; | |
201 | leftmost = 0; | |
202 | } | |
203 | } | |
204 | ||
205 | /* | |
206 | * Maintain a cache of leftmost tree entries (it is frequently | |
207 | * used): | |
208 | */ | |
209 | if (leftmost) | |
210 | cfs_rq->rb_leftmost = &se->run_node; | |
211 | ||
212 | rb_link_node(&se->run_node, parent, link); | |
213 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
214 | update_load_add(&cfs_rq->load, se->load.weight); | |
215 | cfs_rq->nr_running++; | |
216 | se->on_rq = 1; | |
a206c072 IM |
217 | |
218 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
bf0f6f24 IM |
219 | } |
220 | ||
221 | static inline void | |
222 | __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
223 | { | |
224 | if (cfs_rq->rb_leftmost == &se->run_node) | |
225 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
226 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
227 | update_load_sub(&cfs_rq->load, se->load.weight); | |
228 | cfs_rq->nr_running--; | |
229 | se->on_rq = 0; | |
a206c072 IM |
230 | |
231 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
bf0f6f24 IM |
232 | } |
233 | ||
234 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
235 | { | |
236 | return cfs_rq->rb_leftmost; | |
237 | } | |
238 | ||
239 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
240 | { | |
241 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
242 | } | |
243 | ||
244 | /************************************************************** | |
245 | * Scheduling class statistics methods: | |
246 | */ | |
247 | ||
21805085 PZ |
248 | /* |
249 | * Calculate the preemption granularity needed to schedule every | |
250 | * runnable task once per sysctl_sched_latency amount of time. | |
251 | * (down to a sensible low limit on granularity) | |
252 | * | |
253 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
254 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
255 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
256 | * for each task. We do finer and finer scheduling up to until we | |
257 | * reach the minimum granularity value. | |
258 | * | |
259 | * To achieve this we use the following dynamic-granularity rule: | |
260 | * | |
261 | * gran = lat/nr - lat/nr/nr | |
262 | * | |
263 | * This comes out of the following equations: | |
264 | * | |
265 | * kA1 + gran = kB1 | |
266 | * kB2 + gran = kA2 | |
267 | * kA2 = kA1 | |
268 | * kB2 = kB1 - d + d/nr | |
269 | * lat = d * nr | |
270 | * | |
271 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
272 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
273 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
274 | * running, 'lat' is the the period of each task. ('lat' is the | |
275 | * sched_latency that we aim for.) | |
276 | */ | |
277 | static long | |
278 | sched_granularity(struct cfs_rq *cfs_rq) | |
279 | { | |
280 | unsigned int gran = sysctl_sched_latency; | |
281 | unsigned int nr = cfs_rq->nr_running; | |
282 | ||
283 | if (nr > 1) { | |
284 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 285 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
286 | } |
287 | ||
288 | return gran; | |
289 | } | |
290 | ||
bf0f6f24 IM |
291 | /* |
292 | * We rescale the rescheduling granularity of tasks according to their | |
293 | * nice level, but only linearly, not exponentially: | |
294 | */ | |
295 | static long | |
296 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
297 | { | |
298 | u64 tmp; | |
299 | ||
7cff8cf6 IM |
300 | if (likely(curr->load.weight == NICE_0_LOAD)) |
301 | return granularity; | |
bf0f6f24 | 302 | /* |
7cff8cf6 | 303 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 304 | */ |
7cff8cf6 IM |
305 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
306 | tmp = curr->load.weight * (u64)granularity; | |
307 | return (long) (tmp >> NICE_0_SHIFT); | |
308 | } | |
bf0f6f24 | 309 | /* |
7cff8cf6 | 310 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
311 | * granularity: |
312 | */ | |
7cff8cf6 | 313 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
314 | |
315 | /* | |
316 | * It will always fit into 'long': | |
317 | */ | |
a0dc7260 | 318 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
319 | } |
320 | ||
321 | static inline void | |
322 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
323 | { | |
324 | long limit = sysctl_sched_runtime_limit; | |
325 | ||
326 | /* | |
327 | * Niced tasks have the same history dynamic range as | |
328 | * non-niced tasks: | |
329 | */ | |
330 | if (unlikely(se->wait_runtime > limit)) { | |
331 | se->wait_runtime = limit; | |
332 | schedstat_inc(se, wait_runtime_overruns); | |
333 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
334 | } | |
335 | if (unlikely(se->wait_runtime < -limit)) { | |
336 | se->wait_runtime = -limit; | |
337 | schedstat_inc(se, wait_runtime_underruns); | |
338 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
339 | } | |
340 | } | |
341 | ||
342 | static inline void | |
343 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
344 | { | |
345 | se->wait_runtime += delta; | |
346 | schedstat_add(se, sum_wait_runtime, delta); | |
347 | limit_wait_runtime(cfs_rq, se); | |
348 | } | |
349 | ||
350 | static void | |
351 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
352 | { | |
353 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
354 | __add_wait_runtime(cfs_rq, se, delta); | |
355 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
356 | } | |
357 | ||
358 | /* | |
359 | * Update the current task's runtime statistics. Skip current tasks that | |
360 | * are not in our scheduling class. | |
361 | */ | |
362 | static inline void | |
b7cc0896 | 363 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 364 | { |
c5dcfe72 | 365 | unsigned long delta, delta_exec, delta_fair, delta_mine; |
bf0f6f24 IM |
366 | struct load_weight *lw = &cfs_rq->load; |
367 | unsigned long load = lw->weight; | |
368 | ||
bf0f6f24 | 369 | delta_exec = curr->delta_exec; |
8179ca23 | 370 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
371 | |
372 | curr->sum_exec_runtime += delta_exec; | |
373 | cfs_rq->exec_clock += delta_exec; | |
374 | ||
fd8bb43e IM |
375 | if (unlikely(!load)) |
376 | return; | |
377 | ||
bf0f6f24 IM |
378 | delta_fair = calc_delta_fair(delta_exec, lw); |
379 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
380 | ||
5f01d519 | 381 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 382 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
383 | delta = min(delta, (unsigned long)( |
384 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
385 | cfs_rq->sleeper_bonus -= delta; |
386 | delta_mine -= delta; | |
387 | } | |
388 | ||
389 | cfs_rq->fair_clock += delta_fair; | |
390 | /* | |
391 | * We executed delta_exec amount of time on the CPU, | |
392 | * but we were only entitled to delta_mine amount of | |
393 | * time during that period (if nr_running == 1 then | |
394 | * the two values are equal) | |
395 | * [Note: delta_mine - delta_exec is negative]: | |
396 | */ | |
397 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
398 | } | |
399 | ||
b7cc0896 | 400 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
401 | { |
402 | struct sched_entity *curr = cfs_rq_curr(cfs_rq); | |
403 | unsigned long delta_exec; | |
404 | ||
405 | if (unlikely(!curr)) | |
406 | return; | |
407 | ||
408 | /* | |
409 | * Get the amount of time the current task was running | |
410 | * since the last time we changed load (this cannot | |
411 | * overflow on 32 bits): | |
412 | */ | |
d281918d | 413 | delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start); |
bf0f6f24 IM |
414 | |
415 | curr->delta_exec += delta_exec; | |
416 | ||
417 | if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) { | |
b7cc0896 | 418 | __update_curr(cfs_rq, curr); |
bf0f6f24 IM |
419 | curr->delta_exec = 0; |
420 | } | |
d281918d | 421 | curr->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
422 | } |
423 | ||
424 | static inline void | |
5870db5b | 425 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
426 | { |
427 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 428 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
429 | } |
430 | ||
431 | /* | |
432 | * We calculate fair deltas here, so protect against the random effects | |
433 | * of a multiplication overflow by capping it to the runtime limit: | |
434 | */ | |
435 | #if BITS_PER_LONG == 32 | |
436 | static inline unsigned long | |
437 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
438 | { | |
439 | u64 tmp = (u64)delta * weight >> shift; | |
440 | ||
441 | if (unlikely(tmp > sysctl_sched_runtime_limit*2)) | |
442 | return sysctl_sched_runtime_limit*2; | |
443 | return tmp; | |
444 | } | |
445 | #else | |
446 | static inline unsigned long | |
447 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
448 | { | |
449 | return delta * weight >> shift; | |
450 | } | |
451 | #endif | |
452 | ||
453 | /* | |
454 | * Task is being enqueued - update stats: | |
455 | */ | |
d2417e5a | 456 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
457 | { |
458 | s64 key; | |
459 | ||
460 | /* | |
461 | * Are we enqueueing a waiting task? (for current tasks | |
462 | * a dequeue/enqueue event is a NOP) | |
463 | */ | |
464 | if (se != cfs_rq_curr(cfs_rq)) | |
5870db5b | 465 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
466 | /* |
467 | * Update the key: | |
468 | */ | |
469 | key = cfs_rq->fair_clock; | |
470 | ||
471 | /* | |
472 | * Optimize the common nice 0 case: | |
473 | */ | |
474 | if (likely(se->load.weight == NICE_0_LOAD)) { | |
475 | key -= se->wait_runtime; | |
476 | } else { | |
477 | u64 tmp; | |
478 | ||
479 | if (se->wait_runtime < 0) { | |
480 | tmp = -se->wait_runtime; | |
481 | key += (tmp * se->load.inv_weight) >> | |
482 | (WMULT_SHIFT - NICE_0_SHIFT); | |
483 | } else { | |
484 | tmp = se->wait_runtime; | |
a69edb55 IM |
485 | key -= (tmp * se->load.inv_weight) >> |
486 | (WMULT_SHIFT - NICE_0_SHIFT); | |
bf0f6f24 IM |
487 | } |
488 | } | |
489 | ||
490 | se->fair_key = key; | |
491 | } | |
492 | ||
493 | /* | |
494 | * Note: must be called with a freshly updated rq->fair_clock. | |
495 | */ | |
496 | static inline void | |
eac55ea3 | 497 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
498 | { |
499 | unsigned long delta_fair = se->delta_fair_run; | |
500 | ||
d281918d IM |
501 | schedstat_set(se->wait_max, max(se->wait_max, |
502 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 IM |
503 | |
504 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
505 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
506 | NICE_0_SHIFT); | |
507 | ||
508 | add_wait_runtime(cfs_rq, se, delta_fair); | |
509 | } | |
510 | ||
511 | static void | |
9ef0a961 | 512 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
513 | { |
514 | unsigned long delta_fair; | |
515 | ||
b77d69db IM |
516 | if (unlikely(!se->wait_start_fair)) |
517 | return; | |
518 | ||
bf0f6f24 IM |
519 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
520 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
521 | ||
522 | se->delta_fair_run += delta_fair; | |
523 | if (unlikely(abs(se->delta_fair_run) >= | |
524 | sysctl_sched_stat_granularity)) { | |
eac55ea3 | 525 | __update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
526 | se->delta_fair_run = 0; |
527 | } | |
528 | ||
529 | se->wait_start_fair = 0; | |
6cfb0d5d | 530 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
531 | } |
532 | ||
533 | static inline void | |
19b6a2e3 | 534 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 535 | { |
b7cc0896 | 536 | update_curr(cfs_rq); |
bf0f6f24 IM |
537 | /* |
538 | * Mark the end of the wait period if dequeueing a | |
539 | * waiting task: | |
540 | */ | |
541 | if (se != cfs_rq_curr(cfs_rq)) | |
9ef0a961 | 542 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
543 | } |
544 | ||
545 | /* | |
546 | * We are picking a new current task - update its stats: | |
547 | */ | |
548 | static inline void | |
79303e9e | 549 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
550 | { |
551 | /* | |
552 | * We are starting a new run period: | |
553 | */ | |
d281918d | 554 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
555 | } |
556 | ||
557 | /* | |
558 | * We are descheduling a task - update its stats: | |
559 | */ | |
560 | static inline void | |
c7e9b5b2 | 561 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
562 | { |
563 | se->exec_start = 0; | |
564 | } | |
565 | ||
566 | /************************************************** | |
567 | * Scheduling class queueing methods: | |
568 | */ | |
569 | ||
dfdc119e | 570 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
571 | { |
572 | unsigned long load = cfs_rq->load.weight, delta_fair; | |
573 | long prev_runtime; | |
574 | ||
b2133c8b IM |
575 | /* |
576 | * Do not boost sleepers if there's too much bonus 'in flight' | |
577 | * already: | |
578 | */ | |
579 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
580 | return; | |
581 | ||
bf0f6f24 IM |
582 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG) |
583 | load = rq_of(cfs_rq)->cpu_load[2]; | |
584 | ||
585 | delta_fair = se->delta_fair_sleep; | |
586 | ||
587 | /* | |
588 | * Fix up delta_fair with the effect of us running | |
589 | * during the whole sleep period: | |
590 | */ | |
591 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG) | |
592 | delta_fair = div64_likely32((u64)delta_fair * load, | |
593 | load + se->load.weight); | |
594 | ||
595 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
596 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
597 | NICE_0_SHIFT); | |
598 | ||
599 | prev_runtime = se->wait_runtime; | |
600 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
601 | delta_fair = se->wait_runtime - prev_runtime; | |
602 | ||
603 | /* | |
604 | * Track the amount of bonus we've given to sleepers: | |
605 | */ | |
606 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
607 | } |
608 | ||
2396af69 | 609 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
610 | { |
611 | struct task_struct *tsk = task_of(se); | |
612 | unsigned long delta_fair; | |
613 | ||
614 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
615 | !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS)) | |
616 | return; | |
617 | ||
618 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
619 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
620 | ||
621 | se->delta_fair_sleep += delta_fair; | |
622 | if (unlikely(abs(se->delta_fair_sleep) >= | |
623 | sysctl_sched_stat_granularity)) { | |
dfdc119e | 624 | __enqueue_sleeper(cfs_rq, se); |
bf0f6f24 IM |
625 | se->delta_fair_sleep = 0; |
626 | } | |
627 | ||
628 | se->sleep_start_fair = 0; | |
629 | ||
630 | #ifdef CONFIG_SCHEDSTATS | |
631 | if (se->sleep_start) { | |
d281918d | 632 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
633 | |
634 | if ((s64)delta < 0) | |
635 | delta = 0; | |
636 | ||
637 | if (unlikely(delta > se->sleep_max)) | |
638 | se->sleep_max = delta; | |
639 | ||
640 | se->sleep_start = 0; | |
641 | se->sum_sleep_runtime += delta; | |
642 | } | |
643 | if (se->block_start) { | |
d281918d | 644 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
645 | |
646 | if ((s64)delta < 0) | |
647 | delta = 0; | |
648 | ||
649 | if (unlikely(delta > se->block_max)) | |
650 | se->block_max = delta; | |
651 | ||
652 | se->block_start = 0; | |
653 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
654 | |
655 | /* | |
656 | * Blocking time is in units of nanosecs, so shift by 20 to | |
657 | * get a milliseconds-range estimation of the amount of | |
658 | * time that the task spent sleeping: | |
659 | */ | |
660 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
661 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
662 | delta >> 20); | |
663 | } | |
bf0f6f24 IM |
664 | } |
665 | #endif | |
666 | } | |
667 | ||
668 | static void | |
668031ca | 669 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
670 | { |
671 | /* | |
672 | * Update the fair clock. | |
673 | */ | |
b7cc0896 | 674 | update_curr(cfs_rq); |
bf0f6f24 IM |
675 | |
676 | if (wakeup) | |
2396af69 | 677 | enqueue_sleeper(cfs_rq, se); |
bf0f6f24 | 678 | |
d2417e5a | 679 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
680 | __enqueue_entity(cfs_rq, se); |
681 | } | |
682 | ||
683 | static void | |
525c2716 | 684 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 685 | { |
19b6a2e3 | 686 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
687 | if (sleep) { |
688 | se->sleep_start_fair = cfs_rq->fair_clock; | |
689 | #ifdef CONFIG_SCHEDSTATS | |
690 | if (entity_is_task(se)) { | |
691 | struct task_struct *tsk = task_of(se); | |
692 | ||
693 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 694 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 695 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 696 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 697 | } |
bf0f6f24 IM |
698 | #endif |
699 | } | |
700 | __dequeue_entity(cfs_rq, se); | |
701 | } | |
702 | ||
703 | /* | |
704 | * Preempt the current task with a newly woken task if needed: | |
705 | */ | |
7c92e54f | 706 | static void |
bf0f6f24 IM |
707 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
708 | struct sched_entity *curr, unsigned long granularity) | |
709 | { | |
710 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
711 | unsigned long ideal_runtime, delta_exec; |
712 | ||
713 | /* | |
714 | * ideal_runtime is compared against sum_exec_runtime, which is | |
715 | * walltime, hence do not scale. | |
716 | */ | |
717 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
718 | (unsigned long)sysctl_sched_min_granularity); | |
719 | ||
720 | /* | |
721 | * If we executed more than what the latency constraint suggests, | |
722 | * reduce the rescheduling granularity. This way the total latency | |
723 | * of how much a task is not scheduled converges to | |
724 | * sysctl_sched_latency: | |
725 | */ | |
726 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
727 | if (delta_exec > ideal_runtime) | |
728 | granularity = 0; | |
bf0f6f24 IM |
729 | |
730 | /* | |
731 | * Take scheduling granularity into account - do not | |
732 | * preempt the current task unless the best task has | |
733 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
734 | * |
735 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 736 | */ |
4a55b450 | 737 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
738 | resched_task(rq_of(cfs_rq)->curr); |
739 | } | |
740 | ||
741 | static inline void | |
8494f412 | 742 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
743 | { |
744 | /* | |
745 | * Any task has to be enqueued before it get to execute on | |
746 | * a CPU. So account for the time it spent waiting on the | |
747 | * runqueue. (note, here we rely on pick_next_task() having | |
748 | * done a put_prev_task_fair() shortly before this, which | |
749 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
750 | */ | |
9ef0a961 | 751 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 752 | update_stats_curr_start(cfs_rq, se); |
bf0f6f24 | 753 | set_cfs_rq_curr(cfs_rq, se); |
eba1ed4b IM |
754 | #ifdef CONFIG_SCHEDSTATS |
755 | /* | |
756 | * Track our maximum slice length, if the CPU's load is at | |
757 | * least twice that of our own weight (i.e. dont track it | |
758 | * when there are only lesser-weight tasks around): | |
759 | */ | |
760 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
761 | se->slice_max = max(se->slice_max, | |
762 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
763 | } | |
764 | #endif | |
4a55b450 | 765 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
766 | } |
767 | ||
9948f4b2 | 768 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
769 | { |
770 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
771 | ||
8494f412 | 772 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
773 | |
774 | return se; | |
775 | } | |
776 | ||
ab6cde26 | 777 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
778 | { |
779 | /* | |
780 | * If still on the runqueue then deactivate_task() | |
781 | * was not called and update_curr() has to be done: | |
782 | */ | |
783 | if (prev->on_rq) | |
b7cc0896 | 784 | update_curr(cfs_rq); |
bf0f6f24 | 785 | |
c7e9b5b2 | 786 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
787 | |
788 | if (prev->on_rq) | |
5870db5b | 789 | update_stats_wait_start(cfs_rq, prev); |
bf0f6f24 IM |
790 | set_cfs_rq_curr(cfs_rq, NULL); |
791 | } | |
792 | ||
793 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
794 | { | |
bf0f6f24 | 795 | struct sched_entity *next; |
c1b3da3e | 796 | |
bf0f6f24 IM |
797 | /* |
798 | * Dequeue and enqueue the task to update its | |
799 | * position within the tree: | |
800 | */ | |
525c2716 | 801 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 802 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
803 | |
804 | /* | |
805 | * Reschedule if another task tops the current one. | |
806 | */ | |
807 | next = __pick_next_entity(cfs_rq); | |
808 | if (next == curr) | |
809 | return; | |
810 | ||
11697830 PZ |
811 | __check_preempt_curr_fair(cfs_rq, next, curr, |
812 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
813 | } |
814 | ||
815 | /************************************************** | |
816 | * CFS operations on tasks: | |
817 | */ | |
818 | ||
819 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
820 | ||
821 | /* Walk up scheduling entities hierarchy */ | |
822 | #define for_each_sched_entity(se) \ | |
823 | for (; se; se = se->parent) | |
824 | ||
825 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
826 | { | |
827 | return p->se.cfs_rq; | |
828 | } | |
829 | ||
830 | /* runqueue on which this entity is (to be) queued */ | |
831 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
832 | { | |
833 | return se->cfs_rq; | |
834 | } | |
835 | ||
836 | /* runqueue "owned" by this group */ | |
837 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
838 | { | |
839 | return grp->my_q; | |
840 | } | |
841 | ||
842 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
843 | * another cpu ('this_cpu') | |
844 | */ | |
845 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
846 | { | |
847 | /* A later patch will take group into account */ | |
848 | return &cpu_rq(this_cpu)->cfs; | |
849 | } | |
850 | ||
851 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
852 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
853 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
854 | ||
855 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
856 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
857 | { | |
858 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
859 | return 1; | |
860 | ||
861 | return 0; | |
862 | } | |
863 | ||
864 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
865 | ||
866 | #define for_each_sched_entity(se) \ | |
867 | for (; se; se = NULL) | |
868 | ||
869 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
870 | { | |
871 | return &task_rq(p)->cfs; | |
872 | } | |
873 | ||
874 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
875 | { | |
876 | struct task_struct *p = task_of(se); | |
877 | struct rq *rq = task_rq(p); | |
878 | ||
879 | return &rq->cfs; | |
880 | } | |
881 | ||
882 | /* runqueue "owned" by this group */ | |
883 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
884 | { | |
885 | return NULL; | |
886 | } | |
887 | ||
888 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
889 | { | |
890 | return &cpu_rq(this_cpu)->cfs; | |
891 | } | |
892 | ||
893 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
894 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
895 | ||
896 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
897 | { | |
898 | return 1; | |
899 | } | |
900 | ||
901 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
902 | ||
903 | /* | |
904 | * The enqueue_task method is called before nr_running is | |
905 | * increased. Here we update the fair scheduling stats and | |
906 | * then put the task into the rbtree: | |
907 | */ | |
fd390f6a | 908 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
909 | { |
910 | struct cfs_rq *cfs_rq; | |
911 | struct sched_entity *se = &p->se; | |
912 | ||
913 | for_each_sched_entity(se) { | |
914 | if (se->on_rq) | |
915 | break; | |
916 | cfs_rq = cfs_rq_of(se); | |
668031ca | 917 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
918 | } |
919 | } | |
920 | ||
921 | /* | |
922 | * The dequeue_task method is called before nr_running is | |
923 | * decreased. We remove the task from the rbtree and | |
924 | * update the fair scheduling stats: | |
925 | */ | |
f02231e5 | 926 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
927 | { |
928 | struct cfs_rq *cfs_rq; | |
929 | struct sched_entity *se = &p->se; | |
930 | ||
931 | for_each_sched_entity(se) { | |
932 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 933 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
934 | /* Don't dequeue parent if it has other entities besides us */ |
935 | if (cfs_rq->load.weight) | |
936 | break; | |
937 | } | |
938 | } | |
939 | ||
940 | /* | |
1799e35d IM |
941 | * sched_yield() support is very simple - we dequeue and enqueue. |
942 | * | |
943 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
944 | */ |
945 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
946 | { | |
947 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
948 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
949 | struct sched_entity *rightmost, *se = &p->se; | |
950 | struct rb_node *parent; | |
bf0f6f24 IM |
951 | |
952 | /* | |
1799e35d IM |
953 | * Are we the only task in the tree? |
954 | */ | |
955 | if (unlikely(cfs_rq->nr_running == 1)) | |
956 | return; | |
957 | ||
958 | if (likely(!sysctl_sched_compat_yield)) { | |
959 | __update_rq_clock(rq); | |
960 | /* | |
961 | * Dequeue and enqueue the task to update its | |
962 | * position within the tree: | |
963 | */ | |
964 | dequeue_entity(cfs_rq, &p->se, 0); | |
965 | enqueue_entity(cfs_rq, &p->se, 0); | |
966 | ||
967 | return; | |
968 | } | |
969 | /* | |
970 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 971 | */ |
1799e35d IM |
972 | do { |
973 | parent = *link; | |
974 | link = &parent->rb_right; | |
975 | } while (*link); | |
976 | ||
977 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
978 | /* | |
979 | * Already in the rightmost position? | |
980 | */ | |
981 | if (unlikely(rightmost == se)) | |
982 | return; | |
983 | ||
984 | /* | |
985 | * Minimally necessary key value to be last in the tree: | |
986 | */ | |
987 | se->fair_key = rightmost->fair_key + 1; | |
988 | ||
989 | if (cfs_rq->rb_leftmost == &se->run_node) | |
990 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
991 | /* | |
992 | * Relink the task to the rightmost position: | |
993 | */ | |
994 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
995 | rb_link_node(&se->run_node, parent, link); | |
996 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
997 | } |
998 | ||
999 | /* | |
1000 | * Preempt the current task with a newly woken task if needed: | |
1001 | */ | |
1002 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
1003 | { | |
1004 | struct task_struct *curr = rq->curr; | |
1005 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1006 | unsigned long gran; | |
1007 | ||
1008 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 1009 | update_rq_clock(rq); |
b7cc0896 | 1010 | update_curr(cfs_rq); |
bf0f6f24 IM |
1011 | resched_task(curr); |
1012 | return; | |
1013 | } | |
1014 | ||
1015 | gran = sysctl_sched_wakeup_granularity; | |
1016 | /* | |
1017 | * Batch tasks prefer throughput over latency: | |
1018 | */ | |
1019 | if (unlikely(p->policy == SCHED_BATCH)) | |
1020 | gran = sysctl_sched_batch_wakeup_granularity; | |
1021 | ||
1022 | if (is_same_group(curr, p)) | |
1023 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
1024 | } | |
1025 | ||
fb8d4724 | 1026 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
1027 | { |
1028 | struct cfs_rq *cfs_rq = &rq->cfs; | |
1029 | struct sched_entity *se; | |
1030 | ||
1031 | if (unlikely(!cfs_rq->nr_running)) | |
1032 | return NULL; | |
1033 | ||
1034 | do { | |
9948f4b2 | 1035 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1036 | cfs_rq = group_cfs_rq(se); |
1037 | } while (cfs_rq); | |
1038 | ||
1039 | return task_of(se); | |
1040 | } | |
1041 | ||
1042 | /* | |
1043 | * Account for a descheduled task: | |
1044 | */ | |
31ee529c | 1045 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1046 | { |
1047 | struct sched_entity *se = &prev->se; | |
1048 | struct cfs_rq *cfs_rq; | |
1049 | ||
1050 | for_each_sched_entity(se) { | |
1051 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1052 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1053 | } |
1054 | } | |
1055 | ||
1056 | /************************************************** | |
1057 | * Fair scheduling class load-balancing methods: | |
1058 | */ | |
1059 | ||
1060 | /* | |
1061 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1062 | * during the whole iteration, the current task might be | |
1063 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1064 | * achieve that by always pre-iterating before returning | |
1065 | * the current task: | |
1066 | */ | |
1067 | static inline struct task_struct * | |
1068 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1069 | { | |
1070 | struct task_struct *p; | |
1071 | ||
1072 | if (!curr) | |
1073 | return NULL; | |
1074 | ||
1075 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1076 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1077 | ||
1078 | return p; | |
1079 | } | |
1080 | ||
1081 | static struct task_struct *load_balance_start_fair(void *arg) | |
1082 | { | |
1083 | struct cfs_rq *cfs_rq = arg; | |
1084 | ||
1085 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1086 | } | |
1087 | ||
1088 | static struct task_struct *load_balance_next_fair(void *arg) | |
1089 | { | |
1090 | struct cfs_rq *cfs_rq = arg; | |
1091 | ||
1092 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1093 | } | |
1094 | ||
a4ac01c3 | 1095 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1096 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1097 | { | |
1098 | struct sched_entity *curr; | |
1099 | struct task_struct *p; | |
1100 | ||
1101 | if (!cfs_rq->nr_running) | |
1102 | return MAX_PRIO; | |
1103 | ||
1104 | curr = __pick_next_entity(cfs_rq); | |
1105 | p = task_of(curr); | |
1106 | ||
1107 | return p->prio; | |
1108 | } | |
a4ac01c3 | 1109 | #endif |
bf0f6f24 | 1110 | |
43010659 | 1111 | static unsigned long |
bf0f6f24 | 1112 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1113 | unsigned long max_nr_move, unsigned long max_load_move, |
1114 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1115 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1116 | { |
1117 | struct cfs_rq *busy_cfs_rq; | |
1118 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1119 | long rem_load_move = max_load_move; | |
1120 | struct rq_iterator cfs_rq_iterator; | |
1121 | ||
1122 | cfs_rq_iterator.start = load_balance_start_fair; | |
1123 | cfs_rq_iterator.next = load_balance_next_fair; | |
1124 | ||
1125 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1126 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1127 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1128 | long imbalance; |
bf0f6f24 | 1129 | unsigned long maxload; |
bf0f6f24 IM |
1130 | |
1131 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1132 | ||
e56f31aa | 1133 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1134 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1135 | if (imbalance <= 0) | |
1136 | continue; | |
1137 | ||
1138 | /* Don't pull more than imbalance/2 */ | |
1139 | imbalance /= 2; | |
1140 | maxload = min(rem_load_move, imbalance); | |
1141 | ||
a4ac01c3 PW |
1142 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1143 | #else | |
e56f31aa | 1144 | # define maxload rem_load_move |
a4ac01c3 | 1145 | #endif |
bf0f6f24 IM |
1146 | /* pass busy_cfs_rq argument into |
1147 | * load_balance_[start|next]_fair iterators | |
1148 | */ | |
1149 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1150 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1151 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1152 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1153 | |
1154 | total_nr_moved += nr_moved; | |
1155 | max_nr_move -= nr_moved; | |
1156 | rem_load_move -= load_moved; | |
1157 | ||
1158 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1159 | break; | |
1160 | } | |
1161 | ||
43010659 | 1162 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1163 | } |
1164 | ||
1165 | /* | |
1166 | * scheduler tick hitting a task of our scheduling class: | |
1167 | */ | |
1168 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1169 | { | |
1170 | struct cfs_rq *cfs_rq; | |
1171 | struct sched_entity *se = &curr->se; | |
1172 | ||
1173 | for_each_sched_entity(se) { | |
1174 | cfs_rq = cfs_rq_of(se); | |
1175 | entity_tick(cfs_rq, se); | |
1176 | } | |
1177 | } | |
1178 | ||
1179 | /* | |
1180 | * Share the fairness runtime between parent and child, thus the | |
1181 | * total amount of pressure for CPU stays equal - new tasks | |
1182 | * get a chance to run but frequent forkers are not allowed to | |
1183 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1184 | * the child is not running yet. | |
1185 | */ | |
ee0827d8 | 1186 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1187 | { |
1188 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
7109c442 | 1189 | struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq); |
bf0f6f24 IM |
1190 | |
1191 | sched_info_queued(p); | |
1192 | ||
7109c442 | 1193 | update_curr(cfs_rq); |
d2417e5a | 1194 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1195 | /* |
1196 | * Child runs first: we let it run before the parent | |
1197 | * until it reschedules once. We set up the key so that | |
1198 | * it will preempt the parent: | |
1199 | */ | |
9f508f82 | 1200 | se->fair_key = curr->fair_key - |
7109c442 | 1201 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1202 | /* |
1203 | * The first wait is dominated by the child-runs-first logic, | |
1204 | * so do not credit it with that waiting time yet: | |
1205 | */ | |
1206 | if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL) | |
9f508f82 | 1207 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1208 | |
1209 | /* | |
1210 | * The statistical average of wait_runtime is about | |
1211 | * -granularity/2, so initialize the task with that: | |
1212 | */ | |
a206c072 | 1213 | if (sysctl_sched_features & SCHED_FEAT_START_DEBIT) |
9f508f82 | 1214 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 IM |
1215 | |
1216 | __enqueue_entity(cfs_rq, se); | |
bb61c210 | 1217 | resched_task(rq->curr); |
bf0f6f24 IM |
1218 | } |
1219 | ||
1220 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1221 | /* Account for a task changing its policy or group. | |
1222 | * | |
1223 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1224 | * migrates between groups/classes. | |
1225 | */ | |
1226 | static void set_curr_task_fair(struct rq *rq) | |
1227 | { | |
7c6c16f3 | 1228 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1229 | |
c3b64f1e IM |
1230 | for_each_sched_entity(se) |
1231 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1232 | } |
1233 | #else | |
1234 | static void set_curr_task_fair(struct rq *rq) | |
1235 | { | |
1236 | } | |
1237 | #endif | |
1238 | ||
1239 | /* | |
1240 | * All the scheduling class methods: | |
1241 | */ | |
1242 | struct sched_class fair_sched_class __read_mostly = { | |
1243 | .enqueue_task = enqueue_task_fair, | |
1244 | .dequeue_task = dequeue_task_fair, | |
1245 | .yield_task = yield_task_fair, | |
1246 | ||
1247 | .check_preempt_curr = check_preempt_curr_fair, | |
1248 | ||
1249 | .pick_next_task = pick_next_task_fair, | |
1250 | .put_prev_task = put_prev_task_fair, | |
1251 | ||
1252 | .load_balance = load_balance_fair, | |
1253 | ||
1254 | .set_curr_task = set_curr_task_fair, | |
1255 | .task_tick = task_tick_fair, | |
1256 | .task_new = task_new_fair, | |
1257 | }; | |
1258 | ||
1259 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1260 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1261 | { |
bf0f6f24 IM |
1262 | struct cfs_rq *cfs_rq; |
1263 | ||
c3b64f1e | 1264 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1265 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1266 | } |
1267 | #endif |