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 | ||
9745512c AV |
23 | #include <linux/latencytop.h> |
24 | ||
bf0f6f24 | 25 | /* |
21805085 | 26 | * Targeted preemption latency for CPU-bound tasks: |
722aab0c | 27 | * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 28 | * |
21805085 | 29 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
30 | * 'timeslice length' - timeslices in CFS are of variable length |
31 | * and have no persistent notion like in traditional, time-slice | |
32 | * based scheduling concepts. | |
bf0f6f24 | 33 | * |
d274a4ce IM |
34 | * (to see the precise effective timeslice length of your workload, |
35 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 36 | */ |
19978ca6 | 37 | unsigned int sysctl_sched_latency = 20000000ULL; |
2bd8e6d4 IM |
38 | |
39 | /* | |
b2be5e96 | 40 | * Minimal preemption granularity for CPU-bound tasks: |
722aab0c | 41 | * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 42 | */ |
722aab0c | 43 | unsigned int sysctl_sched_min_granularity = 4000000ULL; |
21805085 PZ |
44 | |
45 | /* | |
b2be5e96 PZ |
46 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
47 | */ | |
722aab0c | 48 | static unsigned int sched_nr_latency = 5; |
b2be5e96 PZ |
49 | |
50 | /* | |
51 | * After fork, child runs first. (default) If set to 0 then | |
52 | * parent will (try to) run first. | |
21805085 | 53 | */ |
b2be5e96 | 54 | const_debug unsigned int sysctl_sched_child_runs_first = 1; |
bf0f6f24 | 55 | |
1799e35d IM |
56 | /* |
57 | * sys_sched_yield() compat mode | |
58 | * | |
59 | * This option switches the agressive yield implementation of the | |
60 | * old scheduler back on. | |
61 | */ | |
62 | unsigned int __read_mostly sysctl_sched_compat_yield; | |
63 | ||
bf0f6f24 IM |
64 | /* |
65 | * SCHED_BATCH wake-up granularity. | |
722aab0c | 66 | * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
67 | * |
68 | * This option delays the preemption effects of decoupled workloads | |
69 | * and reduces their over-scheduling. Synchronous workloads will still | |
70 | * have immediate wakeup/sleep latencies. | |
71 | */ | |
19978ca6 | 72 | unsigned int sysctl_sched_batch_wakeup_granularity = 10000000UL; |
bf0f6f24 IM |
73 | |
74 | /* | |
75 | * SCHED_OTHER wake-up granularity. | |
0bbd3336 | 76 | * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
77 | * |
78 | * This option delays the preemption effects of decoupled workloads | |
79 | * and reduces their over-scheduling. Synchronous workloads will still | |
80 | * have immediate wakeup/sleep latencies. | |
81 | */ | |
0bbd3336 | 82 | unsigned int sysctl_sched_wakeup_granularity = 10000000UL; |
bf0f6f24 | 83 | |
da84d961 IM |
84 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
85 | ||
bf0f6f24 IM |
86 | /************************************************************** |
87 | * CFS operations on generic schedulable entities: | |
88 | */ | |
89 | ||
62160e3f | 90 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 91 | |
62160e3f | 92 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
93 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
94 | { | |
62160e3f | 95 | return cfs_rq->rq; |
bf0f6f24 IM |
96 | } |
97 | ||
62160e3f IM |
98 | /* An entity is a task if it doesn't "own" a runqueue */ |
99 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 100 | |
62160e3f | 101 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 102 | |
62160e3f IM |
103 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
104 | { | |
105 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
106 | } |
107 | ||
108 | #define entity_is_task(se) 1 | |
109 | ||
bf0f6f24 IM |
110 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
111 | ||
112 | static inline struct task_struct *task_of(struct sched_entity *se) | |
113 | { | |
114 | return container_of(se, struct task_struct, se); | |
115 | } | |
116 | ||
117 | ||
118 | /************************************************************** | |
119 | * Scheduling class tree data structure manipulation methods: | |
120 | */ | |
121 | ||
0702e3eb | 122 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 123 | { |
368059a9 PZ |
124 | s64 delta = (s64)(vruntime - min_vruntime); |
125 | if (delta > 0) | |
02e0431a PZ |
126 | min_vruntime = vruntime; |
127 | ||
128 | return min_vruntime; | |
129 | } | |
130 | ||
0702e3eb | 131 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
132 | { |
133 | s64 delta = (s64)(vruntime - min_vruntime); | |
134 | if (delta < 0) | |
135 | min_vruntime = vruntime; | |
136 | ||
137 | return min_vruntime; | |
138 | } | |
139 | ||
0702e3eb | 140 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 141 | { |
30cfdcfc | 142 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
143 | } |
144 | ||
bf0f6f24 IM |
145 | /* |
146 | * Enqueue an entity into the rb-tree: | |
147 | */ | |
0702e3eb | 148 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
149 | { |
150 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
151 | struct rb_node *parent = NULL; | |
152 | struct sched_entity *entry; | |
9014623c | 153 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
154 | int leftmost = 1; |
155 | ||
156 | /* | |
157 | * Find the right place in the rbtree: | |
158 | */ | |
159 | while (*link) { | |
160 | parent = *link; | |
161 | entry = rb_entry(parent, struct sched_entity, run_node); | |
162 | /* | |
163 | * We dont care about collisions. Nodes with | |
164 | * the same key stay together. | |
165 | */ | |
9014623c | 166 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
167 | link = &parent->rb_left; |
168 | } else { | |
169 | link = &parent->rb_right; | |
170 | leftmost = 0; | |
171 | } | |
172 | } | |
173 | ||
174 | /* | |
175 | * Maintain a cache of leftmost tree entries (it is frequently | |
176 | * used): | |
177 | */ | |
3fe69747 | 178 | if (leftmost) { |
57cb499d | 179 | cfs_rq->rb_leftmost = &se->run_node; |
3fe69747 PZ |
180 | /* |
181 | * maintain cfs_rq->min_vruntime to be a monotonic increasing | |
182 | * value tracking the leftmost vruntime in the tree. | |
183 | */ | |
184 | cfs_rq->min_vruntime = | |
185 | max_vruntime(cfs_rq->min_vruntime, se->vruntime); | |
186 | } | |
bf0f6f24 IM |
187 | |
188 | rb_link_node(&se->run_node, parent, link); | |
189 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
190 | } |
191 | ||
0702e3eb | 192 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 193 | { |
3fe69747 PZ |
194 | if (cfs_rq->rb_leftmost == &se->run_node) { |
195 | struct rb_node *next_node; | |
196 | struct sched_entity *next; | |
197 | ||
198 | next_node = rb_next(&se->run_node); | |
199 | cfs_rq->rb_leftmost = next_node; | |
200 | ||
201 | if (next_node) { | |
202 | next = rb_entry(next_node, | |
203 | struct sched_entity, run_node); | |
204 | cfs_rq->min_vruntime = | |
205 | max_vruntime(cfs_rq->min_vruntime, | |
206 | next->vruntime); | |
207 | } | |
208 | } | |
e9acbff6 | 209 | |
aa2ac252 PZ |
210 | if (cfs_rq->next == se) |
211 | cfs_rq->next = NULL; | |
212 | ||
bf0f6f24 | 213 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
214 | } |
215 | ||
216 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
217 | { | |
218 | return cfs_rq->rb_leftmost; | |
219 | } | |
220 | ||
221 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
222 | { | |
223 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
224 | } | |
225 | ||
aeb73b04 PZ |
226 | static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
227 | { | |
7eee3e67 | 228 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 229 | |
70eee74b BS |
230 | if (!last) |
231 | return NULL; | |
7eee3e67 IM |
232 | |
233 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
234 | } |
235 | ||
bf0f6f24 IM |
236 | /************************************************************** |
237 | * Scheduling class statistics methods: | |
238 | */ | |
239 | ||
b2be5e96 PZ |
240 | #ifdef CONFIG_SCHED_DEBUG |
241 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
242 | struct file *filp, void __user *buffer, size_t *lenp, | |
243 | loff_t *ppos) | |
244 | { | |
245 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
246 | ||
247 | if (ret || !write) | |
248 | return ret; | |
249 | ||
250 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
251 | sysctl_sched_min_granularity); | |
252 | ||
253 | return 0; | |
254 | } | |
255 | #endif | |
647e7cac IM |
256 | |
257 | /* | |
258 | * The idea is to set a period in which each task runs once. | |
259 | * | |
260 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
261 | * this period because otherwise the slices get too small. | |
262 | * | |
263 | * p = (nr <= nl) ? l : l*nr/nl | |
264 | */ | |
4d78e7b6 PZ |
265 | static u64 __sched_period(unsigned long nr_running) |
266 | { | |
267 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 268 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
269 | |
270 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 271 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 272 | period *= nr_running; |
4d78e7b6 PZ |
273 | } |
274 | ||
275 | return period; | |
276 | } | |
277 | ||
647e7cac IM |
278 | /* |
279 | * We calculate the wall-time slice from the period by taking a part | |
280 | * proportional to the weight. | |
281 | * | |
282 | * s = p*w/rw | |
283 | */ | |
6d0f0ebd | 284 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 285 | { |
6a6029b8 IM |
286 | return calc_delta_mine(__sched_period(cfs_rq->nr_running), |
287 | se->load.weight, &cfs_rq->load); | |
bf0f6f24 IM |
288 | } |
289 | ||
647e7cac IM |
290 | /* |
291 | * We calculate the vruntime slice. | |
292 | * | |
293 | * vs = s/w = p/rw | |
294 | */ | |
295 | static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running) | |
67e9fb2a | 296 | { |
647e7cac | 297 | u64 vslice = __sched_period(nr_running); |
67e9fb2a | 298 | |
10b77724 | 299 | vslice *= NICE_0_LOAD; |
647e7cac | 300 | do_div(vslice, rq_weight); |
67e9fb2a | 301 | |
647e7cac IM |
302 | return vslice; |
303 | } | |
5f6d858e | 304 | |
647e7cac IM |
305 | static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se) |
306 | { | |
307 | return __sched_vslice(cfs_rq->load.weight + se->load.weight, | |
308 | cfs_rq->nr_running + 1); | |
67e9fb2a PZ |
309 | } |
310 | ||
bf0f6f24 IM |
311 | /* |
312 | * Update the current task's runtime statistics. Skip current tasks that | |
313 | * are not in our scheduling class. | |
314 | */ | |
315 | static inline void | |
8ebc91d9 IM |
316 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
317 | unsigned long delta_exec) | |
bf0f6f24 | 318 | { |
bbdba7c0 | 319 | unsigned long delta_exec_weighted; |
bf0f6f24 | 320 | |
8179ca23 | 321 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
322 | |
323 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 324 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
e9acbff6 IM |
325 | delta_exec_weighted = delta_exec; |
326 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { | |
327 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, | |
328 | &curr->load); | |
329 | } | |
330 | curr->vruntime += delta_exec_weighted; | |
bf0f6f24 IM |
331 | } |
332 | ||
b7cc0896 | 333 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 334 | { |
429d43bc | 335 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 336 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
337 | unsigned long delta_exec; |
338 | ||
339 | if (unlikely(!curr)) | |
340 | return; | |
341 | ||
342 | /* | |
343 | * Get the amount of time the current task was running | |
344 | * since the last time we changed load (this cannot | |
345 | * overflow on 32 bits): | |
346 | */ | |
8ebc91d9 | 347 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 348 | |
8ebc91d9 IM |
349 | __update_curr(cfs_rq, curr, delta_exec); |
350 | curr->exec_start = now; | |
d842de87 SV |
351 | |
352 | if (entity_is_task(curr)) { | |
353 | struct task_struct *curtask = task_of(curr); | |
354 | ||
355 | cpuacct_charge(curtask, delta_exec); | |
356 | } | |
bf0f6f24 IM |
357 | } |
358 | ||
359 | static inline void | |
5870db5b | 360 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 361 | { |
d281918d | 362 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
363 | } |
364 | ||
bf0f6f24 IM |
365 | /* |
366 | * Task is being enqueued - update stats: | |
367 | */ | |
d2417e5a | 368 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 369 | { |
bf0f6f24 IM |
370 | /* |
371 | * Are we enqueueing a waiting task? (for current tasks | |
372 | * a dequeue/enqueue event is a NOP) | |
373 | */ | |
429d43bc | 374 | if (se != cfs_rq->curr) |
5870db5b | 375 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
376 | } |
377 | ||
bf0f6f24 | 378 | static void |
9ef0a961 | 379 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 380 | { |
bbdba7c0 IM |
381 | schedstat_set(se->wait_max, max(se->wait_max, |
382 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
383 | schedstat_set(se->wait_count, se->wait_count + 1); |
384 | schedstat_set(se->wait_sum, se->wait_sum + | |
385 | rq_of(cfs_rq)->clock - se->wait_start); | |
6cfb0d5d | 386 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
387 | } |
388 | ||
389 | static inline void | |
19b6a2e3 | 390 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 391 | { |
bf0f6f24 IM |
392 | /* |
393 | * Mark the end of the wait period if dequeueing a | |
394 | * waiting task: | |
395 | */ | |
429d43bc | 396 | if (se != cfs_rq->curr) |
9ef0a961 | 397 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
398 | } |
399 | ||
400 | /* | |
401 | * We are picking a new current task - update its stats: | |
402 | */ | |
403 | static inline void | |
79303e9e | 404 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
405 | { |
406 | /* | |
407 | * We are starting a new run period: | |
408 | */ | |
d281918d | 409 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
410 | } |
411 | ||
bf0f6f24 IM |
412 | /************************************************** |
413 | * Scheduling class queueing methods: | |
414 | */ | |
415 | ||
30cfdcfc DA |
416 | static void |
417 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
418 | { | |
419 | update_load_add(&cfs_rq->load, se->load.weight); | |
420 | cfs_rq->nr_running++; | |
421 | se->on_rq = 1; | |
422 | } | |
423 | ||
424 | static void | |
425 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
426 | { | |
427 | update_load_sub(&cfs_rq->load, se->load.weight); | |
428 | cfs_rq->nr_running--; | |
429 | se->on_rq = 0; | |
430 | } | |
431 | ||
2396af69 | 432 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 433 | { |
bf0f6f24 IM |
434 | #ifdef CONFIG_SCHEDSTATS |
435 | if (se->sleep_start) { | |
d281918d | 436 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 437 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
438 | |
439 | if ((s64)delta < 0) | |
440 | delta = 0; | |
441 | ||
442 | if (unlikely(delta > se->sleep_max)) | |
443 | se->sleep_max = delta; | |
444 | ||
445 | se->sleep_start = 0; | |
446 | se->sum_sleep_runtime += delta; | |
9745512c AV |
447 | |
448 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
449 | } |
450 | if (se->block_start) { | |
d281918d | 451 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 452 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
453 | |
454 | if ((s64)delta < 0) | |
455 | delta = 0; | |
456 | ||
457 | if (unlikely(delta > se->block_max)) | |
458 | se->block_max = delta; | |
459 | ||
460 | se->block_start = 0; | |
461 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
462 | |
463 | /* | |
464 | * Blocking time is in units of nanosecs, so shift by 20 to | |
465 | * get a milliseconds-range estimation of the amount of | |
466 | * time that the task spent sleeping: | |
467 | */ | |
468 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 469 | |
30084fbd IM |
470 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
471 | delta >> 20); | |
472 | } | |
9745512c | 473 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
474 | } |
475 | #endif | |
476 | } | |
477 | ||
ddc97297 PZ |
478 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
479 | { | |
480 | #ifdef CONFIG_SCHED_DEBUG | |
481 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
482 | ||
483 | if (d < 0) | |
484 | d = -d; | |
485 | ||
486 | if (d > 3*sysctl_sched_latency) | |
487 | schedstat_inc(cfs_rq, nr_spread_over); | |
488 | #endif | |
489 | } | |
490 | ||
aeb73b04 PZ |
491 | static void |
492 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
493 | { | |
67e9fb2a | 494 | u64 vruntime; |
aeb73b04 | 495 | |
3fe69747 PZ |
496 | if (first_fair(cfs_rq)) { |
497 | vruntime = min_vruntime(cfs_rq->min_vruntime, | |
498 | __pick_next_entity(cfs_rq)->vruntime); | |
499 | } else | |
500 | vruntime = cfs_rq->min_vruntime; | |
94dfb5e7 | 501 | |
2cb8600e PZ |
502 | /* |
503 | * The 'current' period is already promised to the current tasks, | |
504 | * however the extra weight of the new task will slow them down a | |
505 | * little, place the new task so that it fits in the slot that | |
506 | * stays open at the end. | |
507 | */ | |
94dfb5e7 | 508 | if (initial && sched_feat(START_DEBIT)) |
647e7cac | 509 | vruntime += sched_vslice_add(cfs_rq, se); |
aeb73b04 | 510 | |
8465e792 | 511 | if (!initial) { |
2cb8600e | 512 | /* sleeps upto a single latency don't count. */ |
018d6db4 IM |
513 | if (sched_feat(NEW_FAIR_SLEEPERS)) { |
514 | vruntime -= calc_delta_fair(sysctl_sched_latency, | |
515 | &cfs_rq->load); | |
516 | } | |
94359f05 | 517 | |
2cb8600e PZ |
518 | /* ensure we never gain time by being placed backwards. */ |
519 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
520 | } |
521 | ||
67e9fb2a | 522 | se->vruntime = vruntime; |
aeb73b04 PZ |
523 | } |
524 | ||
bf0f6f24 | 525 | static void |
83b699ed | 526 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
527 | { |
528 | /* | |
a2a2d680 | 529 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 530 | */ |
b7cc0896 | 531 | update_curr(cfs_rq); |
bf0f6f24 | 532 | |
e9acbff6 | 533 | if (wakeup) { |
aeb73b04 | 534 | place_entity(cfs_rq, se, 0); |
2396af69 | 535 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 536 | } |
bf0f6f24 | 537 | |
d2417e5a | 538 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 539 | check_spread(cfs_rq, se); |
83b699ed SV |
540 | if (se != cfs_rq->curr) |
541 | __enqueue_entity(cfs_rq, se); | |
30cfdcfc | 542 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 IM |
543 | } |
544 | ||
4ae7d5ce IM |
545 | static void update_avg(u64 *avg, u64 sample) |
546 | { | |
547 | s64 diff = sample - *avg; | |
548 | *avg += diff >> 3; | |
549 | } | |
550 | ||
551 | static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
552 | { | |
553 | if (!se->last_wakeup) | |
554 | return; | |
555 | ||
556 | update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup); | |
557 | se->last_wakeup = 0; | |
558 | } | |
559 | ||
bf0f6f24 | 560 | static void |
525c2716 | 561 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 562 | { |
a2a2d680 DA |
563 | /* |
564 | * Update run-time statistics of the 'current'. | |
565 | */ | |
566 | update_curr(cfs_rq); | |
567 | ||
19b6a2e3 | 568 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 569 | if (sleep) { |
4ae7d5ce | 570 | update_avg_stats(cfs_rq, se); |
67e9fb2a | 571 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
572 | if (entity_is_task(se)) { |
573 | struct task_struct *tsk = task_of(se); | |
574 | ||
575 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 576 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 577 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 578 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 579 | } |
db36cc7d | 580 | #endif |
67e9fb2a PZ |
581 | } |
582 | ||
83b699ed | 583 | if (se != cfs_rq->curr) |
30cfdcfc DA |
584 | __dequeue_entity(cfs_rq, se); |
585 | account_entity_dequeue(cfs_rq, se); | |
bf0f6f24 IM |
586 | } |
587 | ||
588 | /* | |
589 | * Preempt the current task with a newly woken task if needed: | |
590 | */ | |
7c92e54f | 591 | static void |
2e09bf55 | 592 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 593 | { |
11697830 PZ |
594 | unsigned long ideal_runtime, delta_exec; |
595 | ||
6d0f0ebd | 596 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 597 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
3e3e13f3 | 598 | if (delta_exec > ideal_runtime) |
bf0f6f24 IM |
599 | resched_task(rq_of(cfs_rq)->curr); |
600 | } | |
601 | ||
83b699ed | 602 | static void |
8494f412 | 603 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 604 | { |
83b699ed SV |
605 | /* 'current' is not kept within the tree. */ |
606 | if (se->on_rq) { | |
607 | /* | |
608 | * Any task has to be enqueued before it get to execute on | |
609 | * a CPU. So account for the time it spent waiting on the | |
610 | * runqueue. | |
611 | */ | |
612 | update_stats_wait_end(cfs_rq, se); | |
613 | __dequeue_entity(cfs_rq, se); | |
614 | } | |
615 | ||
79303e9e | 616 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 617 | cfs_rq->curr = se; |
eba1ed4b IM |
618 | #ifdef CONFIG_SCHEDSTATS |
619 | /* | |
620 | * Track our maximum slice length, if the CPU's load is at | |
621 | * least twice that of our own weight (i.e. dont track it | |
622 | * when there are only lesser-weight tasks around): | |
623 | */ | |
495eca49 | 624 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
625 | se->slice_max = max(se->slice_max, |
626 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
627 | } | |
628 | #endif | |
4a55b450 | 629 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
630 | } |
631 | ||
0bbd3336 PZ |
632 | static int |
633 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
634 | ||
aa2ac252 PZ |
635 | static struct sched_entity * |
636 | pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
637 | { | |
aa2ac252 PZ |
638 | if (!cfs_rq->next) |
639 | return se; | |
640 | ||
0bbd3336 | 641 | if (wakeup_preempt_entity(cfs_rq->next, se) != 0) |
aa2ac252 PZ |
642 | return se; |
643 | ||
644 | return cfs_rq->next; | |
645 | } | |
646 | ||
9948f4b2 | 647 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 648 | { |
08ec3df5 | 649 | struct sched_entity *se = NULL; |
bf0f6f24 | 650 | |
08ec3df5 DA |
651 | if (first_fair(cfs_rq)) { |
652 | se = __pick_next_entity(cfs_rq); | |
aa2ac252 | 653 | se = pick_next(cfs_rq, se); |
08ec3df5 DA |
654 | set_next_entity(cfs_rq, se); |
655 | } | |
bf0f6f24 IM |
656 | |
657 | return se; | |
658 | } | |
659 | ||
ab6cde26 | 660 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
661 | { |
662 | /* | |
663 | * If still on the runqueue then deactivate_task() | |
664 | * was not called and update_curr() has to be done: | |
665 | */ | |
666 | if (prev->on_rq) | |
b7cc0896 | 667 | update_curr(cfs_rq); |
bf0f6f24 | 668 | |
ddc97297 | 669 | check_spread(cfs_rq, prev); |
30cfdcfc | 670 | if (prev->on_rq) { |
5870db5b | 671 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
672 | /* Put 'current' back into the tree. */ |
673 | __enqueue_entity(cfs_rq, prev); | |
674 | } | |
429d43bc | 675 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
676 | } |
677 | ||
8f4d37ec PZ |
678 | static void |
679 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 680 | { |
bf0f6f24 | 681 | /* |
30cfdcfc | 682 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 683 | */ |
30cfdcfc | 684 | update_curr(cfs_rq); |
bf0f6f24 | 685 | |
8f4d37ec PZ |
686 | #ifdef CONFIG_SCHED_HRTICK |
687 | /* | |
688 | * queued ticks are scheduled to match the slice, so don't bother | |
689 | * validating it and just reschedule. | |
690 | */ | |
691 | if (queued) | |
692 | return resched_task(rq_of(cfs_rq)->curr); | |
693 | /* | |
694 | * don't let the period tick interfere with the hrtick preemption | |
695 | */ | |
696 | if (!sched_feat(DOUBLE_TICK) && | |
697 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
698 | return; | |
699 | #endif | |
700 | ||
ce6c1311 | 701 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 702 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
703 | } |
704 | ||
705 | /************************************************** | |
706 | * CFS operations on tasks: | |
707 | */ | |
708 | ||
709 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
710 | ||
711 | /* Walk up scheduling entities hierarchy */ | |
712 | #define for_each_sched_entity(se) \ | |
713 | for (; se; se = se->parent) | |
714 | ||
715 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
716 | { | |
717 | return p->se.cfs_rq; | |
718 | } | |
719 | ||
720 | /* runqueue on which this entity is (to be) queued */ | |
721 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
722 | { | |
723 | return se->cfs_rq; | |
724 | } | |
725 | ||
726 | /* runqueue "owned" by this group */ | |
727 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
728 | { | |
729 | return grp->my_q; | |
730 | } | |
731 | ||
732 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
733 | * another cpu ('this_cpu') | |
734 | */ | |
735 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
736 | { | |
29f59db3 | 737 | return cfs_rq->tg->cfs_rq[this_cpu]; |
bf0f6f24 IM |
738 | } |
739 | ||
740 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
741 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
ec2c507f | 742 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
bf0f6f24 | 743 | |
fad095a7 SV |
744 | /* Do the two (enqueued) entities belong to the same group ? */ |
745 | static inline int | |
746 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 | 747 | { |
fad095a7 | 748 | if (se->cfs_rq == pse->cfs_rq) |
bf0f6f24 IM |
749 | return 1; |
750 | ||
751 | return 0; | |
752 | } | |
753 | ||
fad095a7 SV |
754 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
755 | { | |
756 | return se->parent; | |
757 | } | |
758 | ||
bf0f6f24 IM |
759 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
760 | ||
761 | #define for_each_sched_entity(se) \ | |
762 | for (; se; se = NULL) | |
763 | ||
764 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
765 | { | |
766 | return &task_rq(p)->cfs; | |
767 | } | |
768 | ||
769 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
770 | { | |
771 | struct task_struct *p = task_of(se); | |
772 | struct rq *rq = task_rq(p); | |
773 | ||
774 | return &rq->cfs; | |
775 | } | |
776 | ||
777 | /* runqueue "owned" by this group */ | |
778 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
779 | { | |
780 | return NULL; | |
781 | } | |
782 | ||
783 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
784 | { | |
785 | return &cpu_rq(this_cpu)->cfs; | |
786 | } | |
787 | ||
788 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
789 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
790 | ||
fad095a7 SV |
791 | static inline int |
792 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 IM |
793 | { |
794 | return 1; | |
795 | } | |
796 | ||
fad095a7 SV |
797 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
798 | { | |
799 | return NULL; | |
800 | } | |
801 | ||
bf0f6f24 IM |
802 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
803 | ||
8f4d37ec PZ |
804 | #ifdef CONFIG_SCHED_HRTICK |
805 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
806 | { | |
807 | int requeue = rq->curr == p; | |
808 | struct sched_entity *se = &p->se; | |
809 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
810 | ||
811 | WARN_ON(task_rq(p) != rq); | |
812 | ||
813 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
814 | u64 slice = sched_slice(cfs_rq, se); | |
815 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
816 | s64 delta = slice - ran; | |
817 | ||
818 | if (delta < 0) { | |
819 | if (rq->curr == p) | |
820 | resched_task(p); | |
821 | return; | |
822 | } | |
823 | ||
824 | /* | |
825 | * Don't schedule slices shorter than 10000ns, that just | |
826 | * doesn't make sense. Rely on vruntime for fairness. | |
827 | */ | |
828 | if (!requeue) | |
829 | delta = max(10000LL, delta); | |
830 | ||
831 | hrtick_start(rq, delta, requeue); | |
832 | } | |
833 | } | |
834 | #else | |
835 | static inline void | |
836 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
837 | { | |
838 | } | |
839 | #endif | |
840 | ||
bf0f6f24 IM |
841 | /* |
842 | * The enqueue_task method is called before nr_running is | |
843 | * increased. Here we update the fair scheduling stats and | |
844 | * then put the task into the rbtree: | |
845 | */ | |
fd390f6a | 846 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
847 | { |
848 | struct cfs_rq *cfs_rq; | |
62fb1851 | 849 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
850 | |
851 | for_each_sched_entity(se) { | |
62fb1851 | 852 | if (se->on_rq) |
bf0f6f24 IM |
853 | break; |
854 | cfs_rq = cfs_rq_of(se); | |
83b699ed | 855 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 856 | wakeup = 1; |
bf0f6f24 | 857 | } |
8f4d37ec PZ |
858 | |
859 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
860 | } |
861 | ||
862 | /* | |
863 | * The dequeue_task method is called before nr_running is | |
864 | * decreased. We remove the task from the rbtree and | |
865 | * update the fair scheduling stats: | |
866 | */ | |
f02231e5 | 867 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
868 | { |
869 | struct cfs_rq *cfs_rq; | |
62fb1851 | 870 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
871 | |
872 | for_each_sched_entity(se) { | |
873 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 874 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 875 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 876 | if (cfs_rq->load.weight) |
bf0f6f24 | 877 | break; |
b9fa3df3 | 878 | sleep = 1; |
bf0f6f24 | 879 | } |
8f4d37ec PZ |
880 | |
881 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
882 | } |
883 | ||
884 | /* | |
1799e35d IM |
885 | * sched_yield() support is very simple - we dequeue and enqueue. |
886 | * | |
887 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 888 | */ |
4530d7ab | 889 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 890 | { |
db292ca3 IM |
891 | struct task_struct *curr = rq->curr; |
892 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
893 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
894 | |
895 | /* | |
1799e35d IM |
896 | * Are we the only task in the tree? |
897 | */ | |
898 | if (unlikely(cfs_rq->nr_running == 1)) | |
899 | return; | |
900 | ||
db292ca3 | 901 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
1799e35d IM |
902 | __update_rq_clock(rq); |
903 | /* | |
a2a2d680 | 904 | * Update run-time statistics of the 'current'. |
1799e35d | 905 | */ |
2b1e315d | 906 | update_curr(cfs_rq); |
1799e35d IM |
907 | |
908 | return; | |
909 | } | |
910 | /* | |
911 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 912 | */ |
2b1e315d | 913 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
914 | /* |
915 | * Already in the rightmost position? | |
916 | */ | |
2b1e315d | 917 | if (unlikely(rightmost->vruntime < se->vruntime)) |
1799e35d IM |
918 | return; |
919 | ||
920 | /* | |
921 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
922 | * Upon rescheduling, sched_class::put_prev_task() will place |
923 | * 'current' within the tree based on its new key value. | |
1799e35d | 924 | */ |
30cfdcfc | 925 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
926 | } |
927 | ||
e7693a36 GH |
928 | /* |
929 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
930 | * not idle and an idle cpu is available. The span of cpus to | |
931 | * search starts with cpus closest then further out as needed, | |
932 | * so we always favor a closer, idle cpu. | |
933 | * | |
934 | * Returns the CPU we should wake onto. | |
935 | */ | |
936 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
937 | static int wake_idle(int cpu, struct task_struct *p) | |
938 | { | |
939 | cpumask_t tmp; | |
940 | struct sched_domain *sd; | |
941 | int i; | |
942 | ||
943 | /* | |
944 | * If it is idle, then it is the best cpu to run this task. | |
945 | * | |
946 | * This cpu is also the best, if it has more than one task already. | |
947 | * Siblings must be also busy(in most cases) as they didn't already | |
948 | * pickup the extra load from this cpu and hence we need not check | |
949 | * sibling runqueue info. This will avoid the checks and cache miss | |
950 | * penalities associated with that. | |
951 | */ | |
952 | if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1) | |
953 | return cpu; | |
954 | ||
955 | for_each_domain(cpu, sd) { | |
956 | if (sd->flags & SD_WAKE_IDLE) { | |
957 | cpus_and(tmp, sd->span, p->cpus_allowed); | |
958 | for_each_cpu_mask(i, tmp) { | |
959 | if (idle_cpu(i)) { | |
960 | if (i != task_cpu(p)) { | |
961 | schedstat_inc(p, | |
962 | se.nr_wakeups_idle); | |
963 | } | |
964 | return i; | |
965 | } | |
966 | } | |
967 | } else { | |
968 | break; | |
969 | } | |
970 | } | |
971 | return cpu; | |
972 | } | |
973 | #else | |
974 | static inline int wake_idle(int cpu, struct task_struct *p) | |
975 | { | |
976 | return cpu; | |
977 | } | |
978 | #endif | |
979 | ||
980 | #ifdef CONFIG_SMP | |
098fb9db | 981 | |
4ae7d5ce IM |
982 | static const struct sched_class fair_sched_class; |
983 | ||
098fb9db | 984 | static int |
4ae7d5ce IM |
985 | wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq, |
986 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, | |
987 | int idx, unsigned long load, unsigned long this_load, | |
098fb9db IM |
988 | unsigned int imbalance) |
989 | { | |
4ae7d5ce | 990 | struct task_struct *curr = this_rq->curr; |
098fb9db IM |
991 | unsigned long tl = this_load; |
992 | unsigned long tl_per_task; | |
993 | ||
994 | if (!(this_sd->flags & SD_WAKE_AFFINE)) | |
995 | return 0; | |
996 | ||
997 | /* | |
4ae7d5ce IM |
998 | * If the currently running task will sleep within |
999 | * a reasonable amount of time then attract this newly | |
1000 | * woken task: | |
098fb9db | 1001 | */ |
4ae7d5ce IM |
1002 | if (sync && curr->sched_class == &fair_sched_class) { |
1003 | if (curr->se.avg_overlap < sysctl_sched_migration_cost && | |
1004 | p->se.avg_overlap < sysctl_sched_migration_cost) | |
1005 | return 1; | |
1006 | } | |
098fb9db IM |
1007 | |
1008 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1009 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1010 | ||
1011 | /* | |
1012 | * If sync wakeup then subtract the (maximum possible) | |
1013 | * effect of the currently running task from the load | |
1014 | * of the current CPU: | |
1015 | */ | |
1016 | if (sync) | |
1017 | tl -= current->se.load.weight; | |
1018 | ||
ac192d39 | 1019 | if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) || |
098fb9db IM |
1020 | 100*(tl + p->se.load.weight) <= imbalance*load) { |
1021 | /* | |
1022 | * This domain has SD_WAKE_AFFINE and | |
1023 | * p is cache cold in this domain, and | |
1024 | * there is no bad imbalance. | |
1025 | */ | |
1026 | schedstat_inc(this_sd, ttwu_move_affine); | |
1027 | schedstat_inc(p, se.nr_wakeups_affine); | |
1028 | ||
1029 | return 1; | |
1030 | } | |
1031 | return 0; | |
1032 | } | |
1033 | ||
e7693a36 GH |
1034 | static int select_task_rq_fair(struct task_struct *p, int sync) |
1035 | { | |
e7693a36 | 1036 | struct sched_domain *sd, *this_sd = NULL; |
ac192d39 | 1037 | int prev_cpu, this_cpu, new_cpu; |
098fb9db | 1038 | unsigned long load, this_load; |
4ae7d5ce | 1039 | struct rq *rq, *this_rq; |
098fb9db | 1040 | unsigned int imbalance; |
098fb9db | 1041 | int idx; |
e7693a36 | 1042 | |
ac192d39 IM |
1043 | prev_cpu = task_cpu(p); |
1044 | rq = task_rq(p); | |
1045 | this_cpu = smp_processor_id(); | |
4ae7d5ce | 1046 | this_rq = cpu_rq(this_cpu); |
ac192d39 | 1047 | new_cpu = prev_cpu; |
e7693a36 | 1048 | |
ac192d39 IM |
1049 | /* |
1050 | * 'this_sd' is the first domain that both | |
1051 | * this_cpu and prev_cpu are present in: | |
1052 | */ | |
e7693a36 | 1053 | for_each_domain(this_cpu, sd) { |
ac192d39 | 1054 | if (cpu_isset(prev_cpu, sd->span)) { |
e7693a36 GH |
1055 | this_sd = sd; |
1056 | break; | |
1057 | } | |
1058 | } | |
1059 | ||
1060 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) | |
f4827386 | 1061 | goto out; |
e7693a36 GH |
1062 | |
1063 | /* | |
1064 | * Check for affine wakeup and passive balancing possibilities. | |
1065 | */ | |
098fb9db | 1066 | if (!this_sd) |
f4827386 | 1067 | goto out; |
e7693a36 | 1068 | |
098fb9db IM |
1069 | idx = this_sd->wake_idx; |
1070 | ||
1071 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1072 | ||
ac192d39 | 1073 | load = source_load(prev_cpu, idx); |
098fb9db IM |
1074 | this_load = target_load(this_cpu, idx); |
1075 | ||
4ae7d5ce IM |
1076 | if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, |
1077 | load, this_load, imbalance)) | |
1078 | return this_cpu; | |
1079 | ||
1080 | if (prev_cpu == this_cpu) | |
f4827386 | 1081 | goto out; |
098fb9db IM |
1082 | |
1083 | /* | |
1084 | * Start passive balancing when half the imbalance_pct | |
1085 | * limit is reached. | |
1086 | */ | |
1087 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1088 | if (imbalance*this_load <= 100*load) { | |
1089 | schedstat_inc(this_sd, ttwu_move_balance); | |
1090 | schedstat_inc(p, se.nr_wakeups_passive); | |
4ae7d5ce | 1091 | return this_cpu; |
e7693a36 GH |
1092 | } |
1093 | } | |
1094 | ||
f4827386 | 1095 | out: |
e7693a36 GH |
1096 | return wake_idle(new_cpu, p); |
1097 | } | |
1098 | #endif /* CONFIG_SMP */ | |
1099 | ||
0bbd3336 PZ |
1100 | static unsigned long wakeup_gran(struct sched_entity *se) |
1101 | { | |
1102 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1103 | ||
1104 | /* | |
1105 | * More easily preempt - nice tasks, while not making | |
1106 | * it harder for + nice tasks. | |
1107 | */ | |
1108 | if (unlikely(se->load.weight > NICE_0_LOAD)) | |
1109 | gran = calc_delta_fair(gran, &se->load); | |
1110 | ||
1111 | return gran; | |
1112 | } | |
1113 | ||
1114 | /* | |
1115 | * Should 'se' preempt 'curr'. | |
1116 | * | |
1117 | * |s1 | |
1118 | * |s2 | |
1119 | * |s3 | |
1120 | * g | |
1121 | * |<--->|c | |
1122 | * | |
1123 | * w(c, s1) = -1 | |
1124 | * w(c, s2) = 0 | |
1125 | * w(c, s3) = 1 | |
1126 | * | |
1127 | */ | |
1128 | static int | |
1129 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1130 | { | |
1131 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1132 | ||
1133 | if (vdiff < 0) | |
1134 | return -1; | |
1135 | ||
1136 | gran = wakeup_gran(curr); | |
1137 | if (vdiff > gran) | |
1138 | return 1; | |
1139 | ||
1140 | return 0; | |
1141 | } | |
e7693a36 | 1142 | |
bf0f6f24 IM |
1143 | /* |
1144 | * Preempt the current task with a newly woken task if needed: | |
1145 | */ | |
2e09bf55 | 1146 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1147 | { |
1148 | struct task_struct *curr = rq->curr; | |
fad095a7 | 1149 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
8651a86c | 1150 | struct sched_entity *se = &curr->se, *pse = &p->se; |
bf0f6f24 IM |
1151 | |
1152 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 1153 | update_rq_clock(rq); |
b7cc0896 | 1154 | update_curr(cfs_rq); |
bf0f6f24 IM |
1155 | resched_task(curr); |
1156 | return; | |
1157 | } | |
aa2ac252 | 1158 | |
4ae7d5ce IM |
1159 | se->last_wakeup = se->sum_exec_runtime; |
1160 | if (unlikely(se == pse)) | |
1161 | return; | |
1162 | ||
aa2ac252 PZ |
1163 | cfs_rq_of(pse)->next = pse; |
1164 | ||
91c234b4 IM |
1165 | /* |
1166 | * Batch tasks do not preempt (their preemption is driven by | |
1167 | * the tick): | |
1168 | */ | |
1169 | if (unlikely(p->policy == SCHED_BATCH)) | |
1170 | return; | |
bf0f6f24 | 1171 | |
77d9cc44 IM |
1172 | if (!sched_feat(WAKEUP_PREEMPT)) |
1173 | return; | |
8651a86c | 1174 | |
77d9cc44 IM |
1175 | while (!is_same_group(se, pse)) { |
1176 | se = parent_entity(se); | |
1177 | pse = parent_entity(pse); | |
ce6c1311 | 1178 | } |
77d9cc44 | 1179 | |
0bbd3336 | 1180 | if (wakeup_preempt_entity(se, pse) == 1) |
77d9cc44 | 1181 | resched_task(curr); |
bf0f6f24 IM |
1182 | } |
1183 | ||
fb8d4724 | 1184 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1185 | { |
8f4d37ec | 1186 | struct task_struct *p; |
bf0f6f24 IM |
1187 | struct cfs_rq *cfs_rq = &rq->cfs; |
1188 | struct sched_entity *se; | |
1189 | ||
1190 | if (unlikely(!cfs_rq->nr_running)) | |
1191 | return NULL; | |
1192 | ||
1193 | do { | |
9948f4b2 | 1194 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1195 | cfs_rq = group_cfs_rq(se); |
1196 | } while (cfs_rq); | |
1197 | ||
8f4d37ec PZ |
1198 | p = task_of(se); |
1199 | hrtick_start_fair(rq, p); | |
1200 | ||
1201 | return p; | |
bf0f6f24 IM |
1202 | } |
1203 | ||
1204 | /* | |
1205 | * Account for a descheduled task: | |
1206 | */ | |
31ee529c | 1207 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1208 | { |
1209 | struct sched_entity *se = &prev->se; | |
1210 | struct cfs_rq *cfs_rq; | |
1211 | ||
1212 | for_each_sched_entity(se) { | |
1213 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1214 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1215 | } |
1216 | } | |
1217 | ||
681f3e68 | 1218 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1219 | /************************************************** |
1220 | * Fair scheduling class load-balancing methods: | |
1221 | */ | |
1222 | ||
1223 | /* | |
1224 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1225 | * during the whole iteration, the current task might be | |
1226 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1227 | * achieve that by always pre-iterating before returning | |
1228 | * the current task: | |
1229 | */ | |
a9957449 | 1230 | static struct task_struct * |
bf0f6f24 IM |
1231 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) |
1232 | { | |
1233 | struct task_struct *p; | |
1234 | ||
1235 | if (!curr) | |
1236 | return NULL; | |
1237 | ||
1238 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1239 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1240 | ||
1241 | return p; | |
1242 | } | |
1243 | ||
1244 | static struct task_struct *load_balance_start_fair(void *arg) | |
1245 | { | |
1246 | struct cfs_rq *cfs_rq = arg; | |
1247 | ||
1248 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1249 | } | |
1250 | ||
1251 | static struct task_struct *load_balance_next_fair(void *arg) | |
1252 | { | |
1253 | struct cfs_rq *cfs_rq = arg; | |
1254 | ||
1255 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1256 | } | |
1257 | ||
62fb1851 PZ |
1258 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1259 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) | |
1260 | { | |
1261 | struct sched_entity *curr; | |
1262 | struct task_struct *p; | |
1263 | ||
1264 | if (!cfs_rq->nr_running || !first_fair(cfs_rq)) | |
1265 | return MAX_PRIO; | |
1266 | ||
1267 | curr = cfs_rq->curr; | |
1268 | if (!curr) | |
1269 | curr = __pick_next_entity(cfs_rq); | |
1270 | ||
1271 | p = task_of(curr); | |
1272 | ||
1273 | return p->prio; | |
1274 | } | |
1275 | #endif | |
1276 | ||
43010659 | 1277 | static unsigned long |
bf0f6f24 | 1278 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1279 | unsigned long max_load_move, |
a4ac01c3 PW |
1280 | struct sched_domain *sd, enum cpu_idle_type idle, |
1281 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1282 | { |
1283 | struct cfs_rq *busy_cfs_rq; | |
bf0f6f24 IM |
1284 | long rem_load_move = max_load_move; |
1285 | struct rq_iterator cfs_rq_iterator; | |
1286 | ||
1287 | cfs_rq_iterator.start = load_balance_start_fair; | |
1288 | cfs_rq_iterator.next = load_balance_next_fair; | |
1289 | ||
1290 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1291 | #ifdef CONFIG_FAIR_GROUP_SCHED |
62fb1851 PZ |
1292 | struct cfs_rq *this_cfs_rq; |
1293 | long imbalance; | |
1294 | unsigned long maxload; | |
bf0f6f24 | 1295 | |
62fb1851 | 1296 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); |
6b2d7700 | 1297 | |
62fb1851 PZ |
1298 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
1299 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ | |
1300 | if (imbalance <= 0) | |
bf0f6f24 IM |
1301 | continue; |
1302 | ||
62fb1851 PZ |
1303 | /* Don't pull more than imbalance/2 */ |
1304 | imbalance /= 2; | |
1305 | maxload = min(rem_load_move, imbalance); | |
bf0f6f24 | 1306 | |
62fb1851 | 1307 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
a4ac01c3 | 1308 | #else |
e56f31aa | 1309 | # define maxload rem_load_move |
a4ac01c3 | 1310 | #endif |
e1d1484f PW |
1311 | /* |
1312 | * pass busy_cfs_rq argument into | |
bf0f6f24 IM |
1313 | * load_balance_[start|next]_fair iterators |
1314 | */ | |
1315 | cfs_rq_iterator.arg = busy_cfs_rq; | |
62fb1851 | 1316 | rem_load_move -= balance_tasks(this_rq, this_cpu, busiest, |
e1d1484f PW |
1317 | maxload, sd, idle, all_pinned, |
1318 | this_best_prio, | |
1319 | &cfs_rq_iterator); | |
bf0f6f24 | 1320 | |
e1d1484f | 1321 | if (rem_load_move <= 0) |
bf0f6f24 IM |
1322 | break; |
1323 | } | |
1324 | ||
43010659 | 1325 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1326 | } |
1327 | ||
e1d1484f PW |
1328 | static int |
1329 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1330 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1331 | { | |
1332 | struct cfs_rq *busy_cfs_rq; | |
1333 | struct rq_iterator cfs_rq_iterator; | |
1334 | ||
1335 | cfs_rq_iterator.start = load_balance_start_fair; | |
1336 | cfs_rq_iterator.next = load_balance_next_fair; | |
1337 | ||
1338 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1339 | /* | |
1340 | * pass busy_cfs_rq argument into | |
1341 | * load_balance_[start|next]_fair iterators | |
1342 | */ | |
1343 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1344 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1345 | &cfs_rq_iterator)) | |
1346 | return 1; | |
1347 | } | |
1348 | ||
1349 | return 0; | |
1350 | } | |
681f3e68 | 1351 | #endif |
e1d1484f | 1352 | |
bf0f6f24 IM |
1353 | /* |
1354 | * scheduler tick hitting a task of our scheduling class: | |
1355 | */ | |
8f4d37ec | 1356 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1357 | { |
1358 | struct cfs_rq *cfs_rq; | |
1359 | struct sched_entity *se = &curr->se; | |
1360 | ||
1361 | for_each_sched_entity(se) { | |
1362 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1363 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1364 | } |
1365 | } | |
1366 | ||
8eb172d9 | 1367 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
4d78e7b6 | 1368 | |
bf0f6f24 IM |
1369 | /* |
1370 | * Share the fairness runtime between parent and child, thus the | |
1371 | * total amount of pressure for CPU stays equal - new tasks | |
1372 | * get a chance to run but frequent forkers are not allowed to | |
1373 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1374 | * the child is not running yet. | |
1375 | */ | |
ee0827d8 | 1376 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1377 | { |
1378 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1379 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1380 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1381 | |
1382 | sched_info_queued(p); | |
1383 | ||
7109c442 | 1384 | update_curr(cfs_rq); |
aeb73b04 | 1385 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1386 | |
3c90e6e9 | 1387 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1388 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1389 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1390 | /* |
edcb60a3 IM |
1391 | * Upon rescheduling, sched_class::put_prev_task() will place |
1392 | * 'current' within the tree based on its new key value. | |
1393 | */ | |
4d78e7b6 | 1394 | swap(curr->vruntime, se->vruntime); |
4d78e7b6 | 1395 | } |
bf0f6f24 | 1396 | |
b9dca1e0 | 1397 | enqueue_task_fair(rq, p, 0); |
bb61c210 | 1398 | resched_task(rq->curr); |
bf0f6f24 IM |
1399 | } |
1400 | ||
cb469845 SR |
1401 | /* |
1402 | * Priority of the task has changed. Check to see if we preempt | |
1403 | * the current task. | |
1404 | */ | |
1405 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1406 | int oldprio, int running) | |
1407 | { | |
1408 | /* | |
1409 | * Reschedule if we are currently running on this runqueue and | |
1410 | * our priority decreased, or if we are not currently running on | |
1411 | * this runqueue and our priority is higher than the current's | |
1412 | */ | |
1413 | if (running) { | |
1414 | if (p->prio > oldprio) | |
1415 | resched_task(rq->curr); | |
1416 | } else | |
1417 | check_preempt_curr(rq, p); | |
1418 | } | |
1419 | ||
1420 | /* | |
1421 | * We switched to the sched_fair class. | |
1422 | */ | |
1423 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1424 | int running) | |
1425 | { | |
1426 | /* | |
1427 | * We were most likely switched from sched_rt, so | |
1428 | * kick off the schedule if running, otherwise just see | |
1429 | * if we can still preempt the current task. | |
1430 | */ | |
1431 | if (running) | |
1432 | resched_task(rq->curr); | |
1433 | else | |
1434 | check_preempt_curr(rq, p); | |
1435 | } | |
1436 | ||
83b699ed SV |
1437 | /* Account for a task changing its policy or group. |
1438 | * | |
1439 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1440 | * migrates between groups/classes. | |
1441 | */ | |
1442 | static void set_curr_task_fair(struct rq *rq) | |
1443 | { | |
1444 | struct sched_entity *se = &rq->curr->se; | |
1445 | ||
1446 | for_each_sched_entity(se) | |
1447 | set_next_entity(cfs_rq_of(se), se); | |
1448 | } | |
1449 | ||
810b3817 PZ |
1450 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1451 | static void moved_group_fair(struct task_struct *p) | |
1452 | { | |
1453 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1454 | ||
1455 | update_curr(cfs_rq); | |
1456 | place_entity(cfs_rq, &p->se, 1); | |
1457 | } | |
1458 | #endif | |
1459 | ||
bf0f6f24 IM |
1460 | /* |
1461 | * All the scheduling class methods: | |
1462 | */ | |
5522d5d5 IM |
1463 | static const struct sched_class fair_sched_class = { |
1464 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1465 | .enqueue_task = enqueue_task_fair, |
1466 | .dequeue_task = dequeue_task_fair, | |
1467 | .yield_task = yield_task_fair, | |
e7693a36 GH |
1468 | #ifdef CONFIG_SMP |
1469 | .select_task_rq = select_task_rq_fair, | |
1470 | #endif /* CONFIG_SMP */ | |
bf0f6f24 | 1471 | |
2e09bf55 | 1472 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1473 | |
1474 | .pick_next_task = pick_next_task_fair, | |
1475 | .put_prev_task = put_prev_task_fair, | |
1476 | ||
681f3e68 | 1477 | #ifdef CONFIG_SMP |
bf0f6f24 | 1478 | .load_balance = load_balance_fair, |
e1d1484f | 1479 | .move_one_task = move_one_task_fair, |
681f3e68 | 1480 | #endif |
bf0f6f24 | 1481 | |
83b699ed | 1482 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1483 | .task_tick = task_tick_fair, |
1484 | .task_new = task_new_fair, | |
cb469845 SR |
1485 | |
1486 | .prio_changed = prio_changed_fair, | |
1487 | .switched_to = switched_to_fair, | |
810b3817 PZ |
1488 | |
1489 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1490 | .moved_group = moved_group_fair, | |
1491 | #endif | |
bf0f6f24 IM |
1492 | }; |
1493 | ||
1494 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1495 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1496 | { |
bf0f6f24 IM |
1497 | struct cfs_rq *cfs_rq; |
1498 | ||
75c28ace SV |
1499 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1500 | print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs); | |
1501 | #endif | |
5973e5b9 | 1502 | rcu_read_lock(); |
c3b64f1e | 1503 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1504 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1505 | rcu_read_unlock(); |
bf0f6f24 IM |
1506 | } |
1507 | #endif |