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 | ||
114 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
115 | ||
116 | /* cpu runqueue to which this cfs_rq is attached */ | |
117 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
118 | { | |
119 | return cfs_rq->rq; | |
120 | } | |
121 | ||
122 | /* currently running entity (if any) on this cfs_rq */ | |
123 | static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq) | |
124 | { | |
125 | return cfs_rq->curr; | |
126 | } | |
127 | ||
128 | /* An entity is a task if it doesn't "own" a runqueue */ | |
129 | #define entity_is_task(se) (!se->my_q) | |
130 | ||
131 | static inline void | |
132 | set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
133 | { | |
134 | cfs_rq->curr = se; | |
135 | } | |
136 | ||
137 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
138 | ||
139 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | |
140 | { | |
141 | return container_of(cfs_rq, struct rq, cfs); | |
142 | } | |
143 | ||
144 | static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq) | |
145 | { | |
146 | struct rq *rq = rq_of(cfs_rq); | |
147 | ||
148 | if (unlikely(rq->curr->sched_class != &fair_sched_class)) | |
149 | return NULL; | |
150 | ||
151 | return &rq->curr->se; | |
152 | } | |
153 | ||
154 | #define entity_is_task(se) 1 | |
155 | ||
156 | static inline void | |
157 | set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
158 | ||
159 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
160 | ||
161 | static inline struct task_struct *task_of(struct sched_entity *se) | |
162 | { | |
163 | return container_of(se, struct task_struct, se); | |
164 | } | |
165 | ||
166 | ||
167 | /************************************************************** | |
168 | * Scheduling class tree data structure manipulation methods: | |
169 | */ | |
170 | ||
171 | /* | |
172 | * Enqueue an entity into the rb-tree: | |
173 | */ | |
174 | static inline void | |
175 | __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
176 | { | |
177 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
178 | struct rb_node *parent = NULL; | |
179 | struct sched_entity *entry; | |
180 | s64 key = se->fair_key; | |
181 | int leftmost = 1; | |
182 | ||
183 | /* | |
184 | * Find the right place in the rbtree: | |
185 | */ | |
186 | while (*link) { | |
187 | parent = *link; | |
188 | entry = rb_entry(parent, struct sched_entity, run_node); | |
189 | /* | |
190 | * We dont care about collisions. Nodes with | |
191 | * the same key stay together. | |
192 | */ | |
193 | if (key - entry->fair_key < 0) { | |
194 | link = &parent->rb_left; | |
195 | } else { | |
196 | link = &parent->rb_right; | |
197 | leftmost = 0; | |
198 | } | |
199 | } | |
200 | ||
201 | /* | |
202 | * Maintain a cache of leftmost tree entries (it is frequently | |
203 | * used): | |
204 | */ | |
205 | if (leftmost) | |
206 | cfs_rq->rb_leftmost = &se->run_node; | |
207 | ||
208 | rb_link_node(&se->run_node, parent, link); | |
209 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
210 | update_load_add(&cfs_rq->load, se->load.weight); | |
211 | cfs_rq->nr_running++; | |
212 | se->on_rq = 1; | |
a206c072 IM |
213 | |
214 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
bf0f6f24 IM |
215 | } |
216 | ||
217 | static inline void | |
218 | __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
219 | { | |
220 | if (cfs_rq->rb_leftmost == &se->run_node) | |
221 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
222 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
223 | update_load_sub(&cfs_rq->load, se->load.weight); | |
224 | cfs_rq->nr_running--; | |
225 | se->on_rq = 0; | |
a206c072 IM |
226 | |
227 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
bf0f6f24 IM |
228 | } |
229 | ||
230 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
231 | { | |
232 | return cfs_rq->rb_leftmost; | |
233 | } | |
234 | ||
235 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
236 | { | |
237 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
238 | } | |
239 | ||
240 | /************************************************************** | |
241 | * Scheduling class statistics methods: | |
242 | */ | |
243 | ||
21805085 PZ |
244 | /* |
245 | * Calculate the preemption granularity needed to schedule every | |
246 | * runnable task once per sysctl_sched_latency amount of time. | |
247 | * (down to a sensible low limit on granularity) | |
248 | * | |
249 | * For example, if there are 2 tasks running and latency is 10 msecs, | |
250 | * we switch tasks every 5 msecs. If we have 3 tasks running, we have | |
251 | * to switch tasks every 3.33 msecs to get a 10 msecs observed latency | |
252 | * for each task. We do finer and finer scheduling up to until we | |
253 | * reach the minimum granularity value. | |
254 | * | |
255 | * To achieve this we use the following dynamic-granularity rule: | |
256 | * | |
257 | * gran = lat/nr - lat/nr/nr | |
258 | * | |
259 | * This comes out of the following equations: | |
260 | * | |
261 | * kA1 + gran = kB1 | |
262 | * kB2 + gran = kA2 | |
263 | * kA2 = kA1 | |
264 | * kB2 = kB1 - d + d/nr | |
265 | * lat = d * nr | |
266 | * | |
267 | * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running), | |
268 | * '1' is start of time, '2' is end of time, 'd' is delay between | |
269 | * 1 and 2 (during which task B was running), 'nr' is number of tasks | |
270 | * running, 'lat' is the the period of each task. ('lat' is the | |
271 | * sched_latency that we aim for.) | |
272 | */ | |
273 | static long | |
274 | sched_granularity(struct cfs_rq *cfs_rq) | |
275 | { | |
276 | unsigned int gran = sysctl_sched_latency; | |
277 | unsigned int nr = cfs_rq->nr_running; | |
278 | ||
279 | if (nr > 1) { | |
280 | gran = gran/nr - gran/nr/nr; | |
172ac3db | 281 | gran = max(gran, sysctl_sched_min_granularity); |
21805085 PZ |
282 | } |
283 | ||
284 | return gran; | |
285 | } | |
286 | ||
bf0f6f24 IM |
287 | /* |
288 | * We rescale the rescheduling granularity of tasks according to their | |
289 | * nice level, but only linearly, not exponentially: | |
290 | */ | |
291 | static long | |
292 | niced_granularity(struct sched_entity *curr, unsigned long granularity) | |
293 | { | |
294 | u64 tmp; | |
295 | ||
7cff8cf6 IM |
296 | if (likely(curr->load.weight == NICE_0_LOAD)) |
297 | return granularity; | |
bf0f6f24 | 298 | /* |
7cff8cf6 | 299 | * Positive nice levels get the same granularity as nice-0: |
bf0f6f24 | 300 | */ |
7cff8cf6 IM |
301 | if (likely(curr->load.weight < NICE_0_LOAD)) { |
302 | tmp = curr->load.weight * (u64)granularity; | |
303 | return (long) (tmp >> NICE_0_SHIFT); | |
304 | } | |
bf0f6f24 | 305 | /* |
7cff8cf6 | 306 | * Negative nice level tasks get linearly finer |
bf0f6f24 IM |
307 | * granularity: |
308 | */ | |
7cff8cf6 | 309 | tmp = curr->load.inv_weight * (u64)granularity; |
bf0f6f24 IM |
310 | |
311 | /* | |
312 | * It will always fit into 'long': | |
313 | */ | |
a0dc7260 | 314 | return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT)); |
bf0f6f24 IM |
315 | } |
316 | ||
317 | static inline void | |
318 | limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
319 | { | |
320 | long limit = sysctl_sched_runtime_limit; | |
321 | ||
322 | /* | |
323 | * Niced tasks have the same history dynamic range as | |
324 | * non-niced tasks: | |
325 | */ | |
326 | if (unlikely(se->wait_runtime > limit)) { | |
327 | se->wait_runtime = limit; | |
328 | schedstat_inc(se, wait_runtime_overruns); | |
329 | schedstat_inc(cfs_rq, wait_runtime_overruns); | |
330 | } | |
331 | if (unlikely(se->wait_runtime < -limit)) { | |
332 | se->wait_runtime = -limit; | |
333 | schedstat_inc(se, wait_runtime_underruns); | |
334 | schedstat_inc(cfs_rq, wait_runtime_underruns); | |
335 | } | |
336 | } | |
337 | ||
338 | static inline void | |
339 | __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
340 | { | |
341 | se->wait_runtime += delta; | |
342 | schedstat_add(se, sum_wait_runtime, delta); | |
343 | limit_wait_runtime(cfs_rq, se); | |
344 | } | |
345 | ||
346 | static void | |
347 | add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta) | |
348 | { | |
349 | schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime); | |
350 | __add_wait_runtime(cfs_rq, se, delta); | |
351 | schedstat_add(cfs_rq, wait_runtime, se->wait_runtime); | |
352 | } | |
353 | ||
354 | /* | |
355 | * Update the current task's runtime statistics. Skip current tasks that | |
356 | * are not in our scheduling class. | |
357 | */ | |
358 | static inline void | |
8ebc91d9 IM |
359 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
360 | unsigned long delta_exec) | |
bf0f6f24 | 361 | { |
8ebc91d9 | 362 | unsigned long delta, delta_fair, delta_mine; |
bf0f6f24 IM |
363 | struct load_weight *lw = &cfs_rq->load; |
364 | unsigned long load = lw->weight; | |
365 | ||
8179ca23 | 366 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
367 | |
368 | curr->sum_exec_runtime += delta_exec; | |
369 | cfs_rq->exec_clock += delta_exec; | |
370 | ||
fd8bb43e IM |
371 | if (unlikely(!load)) |
372 | return; | |
373 | ||
bf0f6f24 IM |
374 | delta_fair = calc_delta_fair(delta_exec, lw); |
375 | delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); | |
376 | ||
5f01d519 | 377 | if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) { |
ea0aa3b2 | 378 | delta = min((u64)delta_mine, cfs_rq->sleeper_bonus); |
b2133c8b IM |
379 | delta = min(delta, (unsigned long)( |
380 | (long)sysctl_sched_runtime_limit - curr->wait_runtime)); | |
bf0f6f24 IM |
381 | cfs_rq->sleeper_bonus -= delta; |
382 | delta_mine -= delta; | |
383 | } | |
384 | ||
385 | cfs_rq->fair_clock += delta_fair; | |
386 | /* | |
387 | * We executed delta_exec amount of time on the CPU, | |
388 | * but we were only entitled to delta_mine amount of | |
389 | * time during that period (if nr_running == 1 then | |
390 | * the two values are equal) | |
391 | * [Note: delta_mine - delta_exec is negative]: | |
392 | */ | |
393 | add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec); | |
394 | } | |
395 | ||
b7cc0896 | 396 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
397 | { |
398 | struct sched_entity *curr = cfs_rq_curr(cfs_rq); | |
8ebc91d9 | 399 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
400 | unsigned long delta_exec; |
401 | ||
402 | if (unlikely(!curr)) | |
403 | return; | |
404 | ||
405 | /* | |
406 | * Get the amount of time the current task was running | |
407 | * since the last time we changed load (this cannot | |
408 | * overflow on 32 bits): | |
409 | */ | |
8ebc91d9 | 410 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 411 | |
8ebc91d9 IM |
412 | __update_curr(cfs_rq, curr, delta_exec); |
413 | curr->exec_start = now; | |
bf0f6f24 IM |
414 | } |
415 | ||
416 | static inline void | |
5870db5b | 417 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
418 | { |
419 | se->wait_start_fair = cfs_rq->fair_clock; | |
d281918d | 420 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
421 | } |
422 | ||
423 | /* | |
424 | * We calculate fair deltas here, so protect against the random effects | |
425 | * of a multiplication overflow by capping it to the runtime limit: | |
426 | */ | |
427 | #if BITS_PER_LONG == 32 | |
428 | static inline unsigned long | |
429 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
430 | { | |
431 | u64 tmp = (u64)delta * weight >> shift; | |
432 | ||
433 | if (unlikely(tmp > sysctl_sched_runtime_limit*2)) | |
434 | return sysctl_sched_runtime_limit*2; | |
435 | return tmp; | |
436 | } | |
437 | #else | |
438 | static inline unsigned long | |
439 | calc_weighted(unsigned long delta, unsigned long weight, int shift) | |
440 | { | |
441 | return delta * weight >> shift; | |
442 | } | |
443 | #endif | |
444 | ||
445 | /* | |
446 | * Task is being enqueued - update stats: | |
447 | */ | |
d2417e5a | 448 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
449 | { |
450 | s64 key; | |
451 | ||
452 | /* | |
453 | * Are we enqueueing a waiting task? (for current tasks | |
454 | * a dequeue/enqueue event is a NOP) | |
455 | */ | |
456 | if (se != cfs_rq_curr(cfs_rq)) | |
5870db5b | 457 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
458 | /* |
459 | * Update the key: | |
460 | */ | |
461 | key = cfs_rq->fair_clock; | |
462 | ||
463 | /* | |
464 | * Optimize the common nice 0 case: | |
465 | */ | |
466 | if (likely(se->load.weight == NICE_0_LOAD)) { | |
467 | key -= se->wait_runtime; | |
468 | } else { | |
469 | u64 tmp; | |
470 | ||
471 | if (se->wait_runtime < 0) { | |
472 | tmp = -se->wait_runtime; | |
473 | key += (tmp * se->load.inv_weight) >> | |
474 | (WMULT_SHIFT - NICE_0_SHIFT); | |
475 | } else { | |
476 | tmp = se->wait_runtime; | |
a69edb55 IM |
477 | key -= (tmp * se->load.inv_weight) >> |
478 | (WMULT_SHIFT - NICE_0_SHIFT); | |
bf0f6f24 IM |
479 | } |
480 | } | |
481 | ||
482 | se->fair_key = key; | |
483 | } | |
484 | ||
485 | /* | |
486 | * Note: must be called with a freshly updated rq->fair_clock. | |
487 | */ | |
488 | static inline void | |
8ebc91d9 IM |
489 | __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, |
490 | unsigned long delta_fair) | |
bf0f6f24 | 491 | { |
d281918d IM |
492 | schedstat_set(se->wait_max, max(se->wait_max, |
493 | rq_of(cfs_rq)->clock - se->wait_start)); | |
bf0f6f24 IM |
494 | |
495 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
496 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
497 | NICE_0_SHIFT); | |
498 | ||
499 | add_wait_runtime(cfs_rq, se, delta_fair); | |
500 | } | |
501 | ||
502 | static void | |
9ef0a961 | 503 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
504 | { |
505 | unsigned long delta_fair; | |
506 | ||
b77d69db IM |
507 | if (unlikely(!se->wait_start_fair)) |
508 | return; | |
509 | ||
bf0f6f24 IM |
510 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), |
511 | (u64)(cfs_rq->fair_clock - se->wait_start_fair)); | |
512 | ||
8ebc91d9 | 513 | __update_stats_wait_end(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
514 | |
515 | se->wait_start_fair = 0; | |
6cfb0d5d | 516 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
517 | } |
518 | ||
519 | static inline void | |
19b6a2e3 | 520 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 521 | { |
b7cc0896 | 522 | update_curr(cfs_rq); |
bf0f6f24 IM |
523 | /* |
524 | * Mark the end of the wait period if dequeueing a | |
525 | * waiting task: | |
526 | */ | |
527 | if (se != cfs_rq_curr(cfs_rq)) | |
9ef0a961 | 528 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
529 | } |
530 | ||
531 | /* | |
532 | * We are picking a new current task - update its stats: | |
533 | */ | |
534 | static inline void | |
79303e9e | 535 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
536 | { |
537 | /* | |
538 | * We are starting a new run period: | |
539 | */ | |
d281918d | 540 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
541 | } |
542 | ||
543 | /* | |
544 | * We are descheduling a task - update its stats: | |
545 | */ | |
546 | static inline void | |
c7e9b5b2 | 547 | update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
548 | { |
549 | se->exec_start = 0; | |
550 | } | |
551 | ||
552 | /************************************************** | |
553 | * Scheduling class queueing methods: | |
554 | */ | |
555 | ||
8ebc91d9 IM |
556 | static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, |
557 | unsigned long delta_fair) | |
bf0f6f24 | 558 | { |
8ebc91d9 | 559 | unsigned long load = cfs_rq->load.weight; |
bf0f6f24 IM |
560 | long prev_runtime; |
561 | ||
b2133c8b IM |
562 | /* |
563 | * Do not boost sleepers if there's too much bonus 'in flight' | |
564 | * already: | |
565 | */ | |
566 | if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit)) | |
567 | return; | |
568 | ||
bf0f6f24 IM |
569 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG) |
570 | load = rq_of(cfs_rq)->cpu_load[2]; | |
571 | ||
bf0f6f24 IM |
572 | /* |
573 | * Fix up delta_fair with the effect of us running | |
574 | * during the whole sleep period: | |
575 | */ | |
576 | if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG) | |
577 | delta_fair = div64_likely32((u64)delta_fair * load, | |
578 | load + se->load.weight); | |
579 | ||
580 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
581 | delta_fair = calc_weighted(delta_fair, se->load.weight, | |
582 | NICE_0_SHIFT); | |
583 | ||
584 | prev_runtime = se->wait_runtime; | |
585 | __add_wait_runtime(cfs_rq, se, delta_fair); | |
586 | delta_fair = se->wait_runtime - prev_runtime; | |
587 | ||
588 | /* | |
589 | * Track the amount of bonus we've given to sleepers: | |
590 | */ | |
591 | cfs_rq->sleeper_bonus += delta_fair; | |
bf0f6f24 IM |
592 | } |
593 | ||
2396af69 | 594 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
595 | { |
596 | struct task_struct *tsk = task_of(se); | |
597 | unsigned long delta_fair; | |
598 | ||
599 | if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) || | |
600 | !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS)) | |
601 | return; | |
602 | ||
603 | delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit), | |
604 | (u64)(cfs_rq->fair_clock - se->sleep_start_fair)); | |
605 | ||
8ebc91d9 | 606 | __enqueue_sleeper(cfs_rq, se, delta_fair); |
bf0f6f24 IM |
607 | |
608 | se->sleep_start_fair = 0; | |
609 | ||
610 | #ifdef CONFIG_SCHEDSTATS | |
611 | if (se->sleep_start) { | |
d281918d | 612 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
613 | |
614 | if ((s64)delta < 0) | |
615 | delta = 0; | |
616 | ||
617 | if (unlikely(delta > se->sleep_max)) | |
618 | se->sleep_max = delta; | |
619 | ||
620 | se->sleep_start = 0; | |
621 | se->sum_sleep_runtime += delta; | |
622 | } | |
623 | if (se->block_start) { | |
d281918d | 624 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
625 | |
626 | if ((s64)delta < 0) | |
627 | delta = 0; | |
628 | ||
629 | if (unlikely(delta > se->block_max)) | |
630 | se->block_max = delta; | |
631 | ||
632 | se->block_start = 0; | |
633 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
634 | |
635 | /* | |
636 | * Blocking time is in units of nanosecs, so shift by 20 to | |
637 | * get a milliseconds-range estimation of the amount of | |
638 | * time that the task spent sleeping: | |
639 | */ | |
640 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
641 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), | |
642 | delta >> 20); | |
643 | } | |
bf0f6f24 IM |
644 | } |
645 | #endif | |
646 | } | |
647 | ||
648 | static void | |
668031ca | 649 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
650 | { |
651 | /* | |
652 | * Update the fair clock. | |
653 | */ | |
b7cc0896 | 654 | update_curr(cfs_rq); |
bf0f6f24 IM |
655 | |
656 | if (wakeup) | |
2396af69 | 657 | enqueue_sleeper(cfs_rq, se); |
bf0f6f24 | 658 | |
d2417e5a | 659 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
660 | __enqueue_entity(cfs_rq, se); |
661 | } | |
662 | ||
663 | static void | |
525c2716 | 664 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 665 | { |
19b6a2e3 | 666 | update_stats_dequeue(cfs_rq, se); |
bf0f6f24 IM |
667 | if (sleep) { |
668 | se->sleep_start_fair = cfs_rq->fair_clock; | |
669 | #ifdef CONFIG_SCHEDSTATS | |
670 | if (entity_is_task(se)) { | |
671 | struct task_struct *tsk = task_of(se); | |
672 | ||
673 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 674 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 675 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 676 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 677 | } |
bf0f6f24 IM |
678 | #endif |
679 | } | |
680 | __dequeue_entity(cfs_rq, se); | |
681 | } | |
682 | ||
683 | /* | |
684 | * Preempt the current task with a newly woken task if needed: | |
685 | */ | |
7c92e54f | 686 | static void |
bf0f6f24 IM |
687 | __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, |
688 | struct sched_entity *curr, unsigned long granularity) | |
689 | { | |
690 | s64 __delta = curr->fair_key - se->fair_key; | |
11697830 PZ |
691 | unsigned long ideal_runtime, delta_exec; |
692 | ||
693 | /* | |
694 | * ideal_runtime is compared against sum_exec_runtime, which is | |
695 | * walltime, hence do not scale. | |
696 | */ | |
697 | ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running, | |
698 | (unsigned long)sysctl_sched_min_granularity); | |
699 | ||
700 | /* | |
701 | * If we executed more than what the latency constraint suggests, | |
702 | * reduce the rescheduling granularity. This way the total latency | |
703 | * of how much a task is not scheduled converges to | |
704 | * sysctl_sched_latency: | |
705 | */ | |
706 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
707 | if (delta_exec > ideal_runtime) | |
708 | granularity = 0; | |
bf0f6f24 IM |
709 | |
710 | /* | |
711 | * Take scheduling granularity into account - do not | |
712 | * preempt the current task unless the best task has | |
713 | * a larger than sched_granularity fairness advantage: | |
11697830 PZ |
714 | * |
715 | * scale granularity as key space is in fair_clock. | |
bf0f6f24 | 716 | */ |
4a55b450 | 717 | if (__delta > niced_granularity(curr, granularity)) |
bf0f6f24 IM |
718 | resched_task(rq_of(cfs_rq)->curr); |
719 | } | |
720 | ||
721 | static inline void | |
8494f412 | 722 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
723 | { |
724 | /* | |
725 | * Any task has to be enqueued before it get to execute on | |
726 | * a CPU. So account for the time it spent waiting on the | |
727 | * runqueue. (note, here we rely on pick_next_task() having | |
728 | * done a put_prev_task_fair() shortly before this, which | |
729 | * updated rq->fair_clock - used by update_stats_wait_end()) | |
730 | */ | |
9ef0a961 | 731 | update_stats_wait_end(cfs_rq, se); |
79303e9e | 732 | update_stats_curr_start(cfs_rq, se); |
bf0f6f24 | 733 | set_cfs_rq_curr(cfs_rq, se); |
eba1ed4b IM |
734 | #ifdef CONFIG_SCHEDSTATS |
735 | /* | |
736 | * Track our maximum slice length, if the CPU's load is at | |
737 | * least twice that of our own weight (i.e. dont track it | |
738 | * when there are only lesser-weight tasks around): | |
739 | */ | |
740 | if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) { | |
741 | se->slice_max = max(se->slice_max, | |
742 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
743 | } | |
744 | #endif | |
4a55b450 | 745 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
746 | } |
747 | ||
9948f4b2 | 748 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 IM |
749 | { |
750 | struct sched_entity *se = __pick_next_entity(cfs_rq); | |
751 | ||
8494f412 | 752 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
753 | |
754 | return se; | |
755 | } | |
756 | ||
ab6cde26 | 757 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
758 | { |
759 | /* | |
760 | * If still on the runqueue then deactivate_task() | |
761 | * was not called and update_curr() has to be done: | |
762 | */ | |
763 | if (prev->on_rq) | |
b7cc0896 | 764 | update_curr(cfs_rq); |
bf0f6f24 | 765 | |
c7e9b5b2 | 766 | update_stats_curr_end(cfs_rq, prev); |
bf0f6f24 IM |
767 | |
768 | if (prev->on_rq) | |
5870db5b | 769 | update_stats_wait_start(cfs_rq, prev); |
bf0f6f24 IM |
770 | set_cfs_rq_curr(cfs_rq, NULL); |
771 | } | |
772 | ||
773 | static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
774 | { | |
bf0f6f24 | 775 | struct sched_entity *next; |
c1b3da3e | 776 | |
bf0f6f24 IM |
777 | /* |
778 | * Dequeue and enqueue the task to update its | |
779 | * position within the tree: | |
780 | */ | |
525c2716 | 781 | dequeue_entity(cfs_rq, curr, 0); |
668031ca | 782 | enqueue_entity(cfs_rq, curr, 0); |
bf0f6f24 IM |
783 | |
784 | /* | |
785 | * Reschedule if another task tops the current one. | |
786 | */ | |
787 | next = __pick_next_entity(cfs_rq); | |
788 | if (next == curr) | |
789 | return; | |
790 | ||
11697830 PZ |
791 | __check_preempt_curr_fair(cfs_rq, next, curr, |
792 | sched_granularity(cfs_rq)); | |
bf0f6f24 IM |
793 | } |
794 | ||
795 | /************************************************** | |
796 | * CFS operations on tasks: | |
797 | */ | |
798 | ||
799 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
800 | ||
801 | /* Walk up scheduling entities hierarchy */ | |
802 | #define for_each_sched_entity(se) \ | |
803 | for (; se; se = se->parent) | |
804 | ||
805 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
806 | { | |
807 | return p->se.cfs_rq; | |
808 | } | |
809 | ||
810 | /* runqueue on which this entity is (to be) queued */ | |
811 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
812 | { | |
813 | return se->cfs_rq; | |
814 | } | |
815 | ||
816 | /* runqueue "owned" by this group */ | |
817 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
818 | { | |
819 | return grp->my_q; | |
820 | } | |
821 | ||
822 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
823 | * another cpu ('this_cpu') | |
824 | */ | |
825 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
826 | { | |
827 | /* A later patch will take group into account */ | |
828 | return &cpu_rq(this_cpu)->cfs; | |
829 | } | |
830 | ||
831 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
832 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
833 | list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
834 | ||
835 | /* Do the two (enqueued) tasks belong to the same group ? */ | |
836 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
837 | { | |
838 | if (curr->se.cfs_rq == p->se.cfs_rq) | |
839 | return 1; | |
840 | ||
841 | return 0; | |
842 | } | |
843 | ||
844 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
845 | ||
846 | #define for_each_sched_entity(se) \ | |
847 | for (; se; se = NULL) | |
848 | ||
849 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
850 | { | |
851 | return &task_rq(p)->cfs; | |
852 | } | |
853 | ||
854 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
855 | { | |
856 | struct task_struct *p = task_of(se); | |
857 | struct rq *rq = task_rq(p); | |
858 | ||
859 | return &rq->cfs; | |
860 | } | |
861 | ||
862 | /* runqueue "owned" by this group */ | |
863 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
864 | { | |
865 | return NULL; | |
866 | } | |
867 | ||
868 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
869 | { | |
870 | return &cpu_rq(this_cpu)->cfs; | |
871 | } | |
872 | ||
873 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
874 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
875 | ||
876 | static inline int is_same_group(struct task_struct *curr, struct task_struct *p) | |
877 | { | |
878 | return 1; | |
879 | } | |
880 | ||
881 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
882 | ||
883 | /* | |
884 | * The enqueue_task method is called before nr_running is | |
885 | * increased. Here we update the fair scheduling stats and | |
886 | * then put the task into the rbtree: | |
887 | */ | |
fd390f6a | 888 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
889 | { |
890 | struct cfs_rq *cfs_rq; | |
891 | struct sched_entity *se = &p->se; | |
892 | ||
893 | for_each_sched_entity(se) { | |
894 | if (se->on_rq) | |
895 | break; | |
896 | cfs_rq = cfs_rq_of(se); | |
668031ca | 897 | enqueue_entity(cfs_rq, se, wakeup); |
bf0f6f24 IM |
898 | } |
899 | } | |
900 | ||
901 | /* | |
902 | * The dequeue_task method is called before nr_running is | |
903 | * decreased. We remove the task from the rbtree and | |
904 | * update the fair scheduling stats: | |
905 | */ | |
f02231e5 | 906 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
907 | { |
908 | struct cfs_rq *cfs_rq; | |
909 | struct sched_entity *se = &p->se; | |
910 | ||
911 | for_each_sched_entity(se) { | |
912 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 913 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 IM |
914 | /* Don't dequeue parent if it has other entities besides us */ |
915 | if (cfs_rq->load.weight) | |
916 | break; | |
917 | } | |
918 | } | |
919 | ||
920 | /* | |
1799e35d IM |
921 | * sched_yield() support is very simple - we dequeue and enqueue. |
922 | * | |
923 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 IM |
924 | */ |
925 | static void yield_task_fair(struct rq *rq, struct task_struct *p) | |
926 | { | |
927 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1799e35d IM |
928 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
929 | struct sched_entity *rightmost, *se = &p->se; | |
930 | struct rb_node *parent; | |
bf0f6f24 IM |
931 | |
932 | /* | |
1799e35d IM |
933 | * Are we the only task in the tree? |
934 | */ | |
935 | if (unlikely(cfs_rq->nr_running == 1)) | |
936 | return; | |
937 | ||
938 | if (likely(!sysctl_sched_compat_yield)) { | |
939 | __update_rq_clock(rq); | |
940 | /* | |
941 | * Dequeue and enqueue the task to update its | |
942 | * position within the tree: | |
943 | */ | |
944 | dequeue_entity(cfs_rq, &p->se, 0); | |
945 | enqueue_entity(cfs_rq, &p->se, 0); | |
946 | ||
947 | return; | |
948 | } | |
949 | /* | |
950 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 951 | */ |
1799e35d IM |
952 | do { |
953 | parent = *link; | |
954 | link = &parent->rb_right; | |
955 | } while (*link); | |
956 | ||
957 | rightmost = rb_entry(parent, struct sched_entity, run_node); | |
958 | /* | |
959 | * Already in the rightmost position? | |
960 | */ | |
961 | if (unlikely(rightmost == se)) | |
962 | return; | |
963 | ||
964 | /* | |
965 | * Minimally necessary key value to be last in the tree: | |
966 | */ | |
967 | se->fair_key = rightmost->fair_key + 1; | |
968 | ||
969 | if (cfs_rq->rb_leftmost == &se->run_node) | |
970 | cfs_rq->rb_leftmost = rb_next(&se->run_node); | |
971 | /* | |
972 | * Relink the task to the rightmost position: | |
973 | */ | |
974 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | |
975 | rb_link_node(&se->run_node, parent, link); | |
976 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
977 | } |
978 | ||
979 | /* | |
980 | * Preempt the current task with a newly woken task if needed: | |
981 | */ | |
982 | static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p) | |
983 | { | |
984 | struct task_struct *curr = rq->curr; | |
985 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
986 | unsigned long gran; | |
987 | ||
988 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 989 | update_rq_clock(rq); |
b7cc0896 | 990 | update_curr(cfs_rq); |
bf0f6f24 IM |
991 | resched_task(curr); |
992 | return; | |
993 | } | |
994 | ||
995 | gran = sysctl_sched_wakeup_granularity; | |
996 | /* | |
997 | * Batch tasks prefer throughput over latency: | |
998 | */ | |
999 | if (unlikely(p->policy == SCHED_BATCH)) | |
1000 | gran = sysctl_sched_batch_wakeup_granularity; | |
1001 | ||
1002 | if (is_same_group(curr, p)) | |
1003 | __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran); | |
1004 | } | |
1005 | ||
fb8d4724 | 1006 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 IM |
1007 | { |
1008 | struct cfs_rq *cfs_rq = &rq->cfs; | |
1009 | struct sched_entity *se; | |
1010 | ||
1011 | if (unlikely(!cfs_rq->nr_running)) | |
1012 | return NULL; | |
1013 | ||
1014 | do { | |
9948f4b2 | 1015 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1016 | cfs_rq = group_cfs_rq(se); |
1017 | } while (cfs_rq); | |
1018 | ||
1019 | return task_of(se); | |
1020 | } | |
1021 | ||
1022 | /* | |
1023 | * Account for a descheduled task: | |
1024 | */ | |
31ee529c | 1025 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1026 | { |
1027 | struct sched_entity *se = &prev->se; | |
1028 | struct cfs_rq *cfs_rq; | |
1029 | ||
1030 | for_each_sched_entity(se) { | |
1031 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1032 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1033 | } |
1034 | } | |
1035 | ||
1036 | /************************************************** | |
1037 | * Fair scheduling class load-balancing methods: | |
1038 | */ | |
1039 | ||
1040 | /* | |
1041 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1042 | * during the whole iteration, the current task might be | |
1043 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1044 | * achieve that by always pre-iterating before returning | |
1045 | * the current task: | |
1046 | */ | |
1047 | static inline struct task_struct * | |
1048 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) | |
1049 | { | |
1050 | struct task_struct *p; | |
1051 | ||
1052 | if (!curr) | |
1053 | return NULL; | |
1054 | ||
1055 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1056 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1057 | ||
1058 | return p; | |
1059 | } | |
1060 | ||
1061 | static struct task_struct *load_balance_start_fair(void *arg) | |
1062 | { | |
1063 | struct cfs_rq *cfs_rq = arg; | |
1064 | ||
1065 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1066 | } | |
1067 | ||
1068 | static struct task_struct *load_balance_next_fair(void *arg) | |
1069 | { | |
1070 | struct cfs_rq *cfs_rq = arg; | |
1071 | ||
1072 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1073 | } | |
1074 | ||
a4ac01c3 | 1075 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 IM |
1076 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
1077 | { | |
1078 | struct sched_entity *curr; | |
1079 | struct task_struct *p; | |
1080 | ||
1081 | if (!cfs_rq->nr_running) | |
1082 | return MAX_PRIO; | |
1083 | ||
1084 | curr = __pick_next_entity(cfs_rq); | |
1085 | p = task_of(curr); | |
1086 | ||
1087 | return p->prio; | |
1088 | } | |
a4ac01c3 | 1089 | #endif |
bf0f6f24 | 1090 | |
43010659 | 1091 | static unsigned long |
bf0f6f24 | 1092 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
a4ac01c3 PW |
1093 | unsigned long max_nr_move, unsigned long max_load_move, |
1094 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1095 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1096 | { |
1097 | struct cfs_rq *busy_cfs_rq; | |
1098 | unsigned long load_moved, total_nr_moved = 0, nr_moved; | |
1099 | long rem_load_move = max_load_move; | |
1100 | struct rq_iterator cfs_rq_iterator; | |
1101 | ||
1102 | cfs_rq_iterator.start = load_balance_start_fair; | |
1103 | cfs_rq_iterator.next = load_balance_next_fair; | |
1104 | ||
1105 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1106 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 1107 | struct cfs_rq *this_cfs_rq; |
e56f31aa | 1108 | long imbalance; |
bf0f6f24 | 1109 | unsigned long maxload; |
bf0f6f24 IM |
1110 | |
1111 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); | |
1112 | ||
e56f31aa | 1113 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
bf0f6f24 IM |
1114 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
1115 | if (imbalance <= 0) | |
1116 | continue; | |
1117 | ||
1118 | /* Don't pull more than imbalance/2 */ | |
1119 | imbalance /= 2; | |
1120 | maxload = min(rem_load_move, imbalance); | |
1121 | ||
a4ac01c3 PW |
1122 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
1123 | #else | |
e56f31aa | 1124 | # define maxload rem_load_move |
a4ac01c3 | 1125 | #endif |
bf0f6f24 IM |
1126 | /* pass busy_cfs_rq argument into |
1127 | * load_balance_[start|next]_fair iterators | |
1128 | */ | |
1129 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1130 | nr_moved = balance_tasks(this_rq, this_cpu, busiest, | |
1131 | max_nr_move, maxload, sd, idle, all_pinned, | |
a4ac01c3 | 1132 | &load_moved, this_best_prio, &cfs_rq_iterator); |
bf0f6f24 IM |
1133 | |
1134 | total_nr_moved += nr_moved; | |
1135 | max_nr_move -= nr_moved; | |
1136 | rem_load_move -= load_moved; | |
1137 | ||
1138 | if (max_nr_move <= 0 || rem_load_move <= 0) | |
1139 | break; | |
1140 | } | |
1141 | ||
43010659 | 1142 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1143 | } |
1144 | ||
1145 | /* | |
1146 | * scheduler tick hitting a task of our scheduling class: | |
1147 | */ | |
1148 | static void task_tick_fair(struct rq *rq, struct task_struct *curr) | |
1149 | { | |
1150 | struct cfs_rq *cfs_rq; | |
1151 | struct sched_entity *se = &curr->se; | |
1152 | ||
1153 | for_each_sched_entity(se) { | |
1154 | cfs_rq = cfs_rq_of(se); | |
1155 | entity_tick(cfs_rq, se); | |
1156 | } | |
1157 | } | |
1158 | ||
1159 | /* | |
1160 | * Share the fairness runtime between parent and child, thus the | |
1161 | * total amount of pressure for CPU stays equal - new tasks | |
1162 | * get a chance to run but frequent forkers are not allowed to | |
1163 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1164 | * the child is not running yet. | |
1165 | */ | |
ee0827d8 | 1166 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1167 | { |
1168 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
7109c442 | 1169 | struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq); |
bf0f6f24 IM |
1170 | |
1171 | sched_info_queued(p); | |
1172 | ||
7109c442 | 1173 | update_curr(cfs_rq); |
d2417e5a | 1174 | update_stats_enqueue(cfs_rq, se); |
bf0f6f24 IM |
1175 | /* |
1176 | * Child runs first: we let it run before the parent | |
1177 | * until it reschedules once. We set up the key so that | |
1178 | * it will preempt the parent: | |
1179 | */ | |
9f508f82 | 1180 | se->fair_key = curr->fair_key - |
7109c442 | 1181 | niced_granularity(curr, sched_granularity(cfs_rq)) - 1; |
bf0f6f24 IM |
1182 | /* |
1183 | * The first wait is dominated by the child-runs-first logic, | |
1184 | * so do not credit it with that waiting time yet: | |
1185 | */ | |
1186 | if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL) | |
9f508f82 | 1187 | se->wait_start_fair = 0; |
bf0f6f24 IM |
1188 | |
1189 | /* | |
1190 | * The statistical average of wait_runtime is about | |
1191 | * -granularity/2, so initialize the task with that: | |
1192 | */ | |
a206c072 | 1193 | if (sysctl_sched_features & SCHED_FEAT_START_DEBIT) |
9f508f82 | 1194 | se->wait_runtime = -(sched_granularity(cfs_rq) / 2); |
bf0f6f24 IM |
1195 | |
1196 | __enqueue_entity(cfs_rq, se); | |
bb61c210 | 1197 | resched_task(rq->curr); |
bf0f6f24 IM |
1198 | } |
1199 | ||
1200 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1201 | /* Account for a task changing its policy or group. | |
1202 | * | |
1203 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1204 | * migrates between groups/classes. | |
1205 | */ | |
1206 | static void set_curr_task_fair(struct rq *rq) | |
1207 | { | |
7c6c16f3 | 1208 | struct sched_entity *se = &rq->curr->se; |
a8e504d2 | 1209 | |
c3b64f1e IM |
1210 | for_each_sched_entity(se) |
1211 | set_next_entity(cfs_rq_of(se), se); | |
bf0f6f24 IM |
1212 | } |
1213 | #else | |
1214 | static void set_curr_task_fair(struct rq *rq) | |
1215 | { | |
1216 | } | |
1217 | #endif | |
1218 | ||
1219 | /* | |
1220 | * All the scheduling class methods: | |
1221 | */ | |
1222 | struct sched_class fair_sched_class __read_mostly = { | |
1223 | .enqueue_task = enqueue_task_fair, | |
1224 | .dequeue_task = dequeue_task_fair, | |
1225 | .yield_task = yield_task_fair, | |
1226 | ||
1227 | .check_preempt_curr = check_preempt_curr_fair, | |
1228 | ||
1229 | .pick_next_task = pick_next_task_fair, | |
1230 | .put_prev_task = put_prev_task_fair, | |
1231 | ||
1232 | .load_balance = load_balance_fair, | |
1233 | ||
1234 | .set_curr_task = set_curr_task_fair, | |
1235 | .task_tick = task_tick_fair, | |
1236 | .task_new = task_new_fair, | |
1237 | }; | |
1238 | ||
1239 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1240 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1241 | { |
bf0f6f24 IM |
1242 | struct cfs_rq *cfs_rq; |
1243 | ||
c3b64f1e | 1244 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1245 | print_cfs_rq(m, cpu, cfs_rq); |
bf0f6f24 IM |
1246 | } |
1247 | #endif |