]>
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. | |
722aab0c | 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 | */ | |
19978ca6 | 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 | */ | |
178 | if (leftmost) | |
57cb499d | 179 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
180 | |
181 | rb_link_node(&se->run_node, parent, link); | |
182 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
183 | } |
184 | ||
0702e3eb | 185 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
186 | { |
187 | if (cfs_rq->rb_leftmost == &se->run_node) | |
57cb499d | 188 | cfs_rq->rb_leftmost = rb_next(&se->run_node); |
e9acbff6 | 189 | |
bf0f6f24 | 190 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
191 | } |
192 | ||
193 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
194 | { | |
195 | return cfs_rq->rb_leftmost; | |
196 | } | |
197 | ||
198 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
199 | { | |
200 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
201 | } | |
202 | ||
aeb73b04 PZ |
203 | static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
204 | { | |
205 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
206 | struct sched_entity *se = NULL; | |
207 | struct rb_node *parent; | |
208 | ||
209 | while (*link) { | |
210 | parent = *link; | |
211 | se = rb_entry(parent, struct sched_entity, run_node); | |
212 | link = &parent->rb_right; | |
213 | } | |
214 | ||
215 | return se; | |
216 | } | |
217 | ||
bf0f6f24 IM |
218 | /************************************************************** |
219 | * Scheduling class statistics methods: | |
220 | */ | |
221 | ||
b2be5e96 PZ |
222 | #ifdef CONFIG_SCHED_DEBUG |
223 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
224 | struct file *filp, void __user *buffer, size_t *lenp, | |
225 | loff_t *ppos) | |
226 | { | |
227 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
228 | ||
229 | if (ret || !write) | |
230 | return ret; | |
231 | ||
232 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
233 | sysctl_sched_min_granularity); | |
234 | ||
235 | return 0; | |
236 | } | |
237 | #endif | |
647e7cac IM |
238 | |
239 | /* | |
240 | * The idea is to set a period in which each task runs once. | |
241 | * | |
242 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
243 | * this period because otherwise the slices get too small. | |
244 | * | |
245 | * p = (nr <= nl) ? l : l*nr/nl | |
246 | */ | |
4d78e7b6 PZ |
247 | static u64 __sched_period(unsigned long nr_running) |
248 | { | |
249 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 250 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
251 | |
252 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 253 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 254 | period *= nr_running; |
4d78e7b6 PZ |
255 | } |
256 | ||
257 | return period; | |
258 | } | |
259 | ||
647e7cac IM |
260 | /* |
261 | * We calculate the wall-time slice from the period by taking a part | |
262 | * proportional to the weight. | |
263 | * | |
264 | * s = p*w/rw | |
265 | */ | |
6d0f0ebd | 266 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 267 | { |
647e7cac | 268 | u64 slice = __sched_period(cfs_rq->nr_running); |
21805085 | 269 | |
647e7cac IM |
270 | slice *= se->load.weight; |
271 | do_div(slice, cfs_rq->load.weight); | |
21805085 | 272 | |
647e7cac | 273 | return slice; |
bf0f6f24 IM |
274 | } |
275 | ||
647e7cac IM |
276 | /* |
277 | * We calculate the vruntime slice. | |
278 | * | |
279 | * vs = s/w = p/rw | |
280 | */ | |
281 | static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running) | |
67e9fb2a | 282 | { |
647e7cac | 283 | u64 vslice = __sched_period(nr_running); |
67e9fb2a | 284 | |
10b77724 | 285 | vslice *= NICE_0_LOAD; |
647e7cac | 286 | do_div(vslice, rq_weight); |
67e9fb2a | 287 | |
647e7cac IM |
288 | return vslice; |
289 | } | |
5f6d858e | 290 | |
647e7cac IM |
291 | static u64 sched_vslice(struct cfs_rq *cfs_rq) |
292 | { | |
293 | return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running); | |
294 | } | |
295 | ||
296 | static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
297 | { | |
298 | return __sched_vslice(cfs_rq->load.weight + se->load.weight, | |
299 | cfs_rq->nr_running + 1); | |
67e9fb2a PZ |
300 | } |
301 | ||
bf0f6f24 IM |
302 | /* |
303 | * Update the current task's runtime statistics. Skip current tasks that | |
304 | * are not in our scheduling class. | |
305 | */ | |
306 | static inline void | |
8ebc91d9 IM |
307 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
308 | unsigned long delta_exec) | |
bf0f6f24 | 309 | { |
bbdba7c0 | 310 | unsigned long delta_exec_weighted; |
b0ffd246 | 311 | u64 vruntime; |
bf0f6f24 | 312 | |
8179ca23 | 313 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
314 | |
315 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 316 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
e9acbff6 IM |
317 | delta_exec_weighted = delta_exec; |
318 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { | |
319 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, | |
320 | &curr->load); | |
321 | } | |
322 | curr->vruntime += delta_exec_weighted; | |
02e0431a PZ |
323 | |
324 | /* | |
325 | * maintain cfs_rq->min_vruntime to be a monotonic increasing | |
326 | * value tracking the leftmost vruntime in the tree. | |
327 | */ | |
328 | if (first_fair(cfs_rq)) { | |
b0ffd246 PZ |
329 | vruntime = min_vruntime(curr->vruntime, |
330 | __pick_next_entity(cfs_rq)->vruntime); | |
02e0431a | 331 | } else |
b0ffd246 | 332 | vruntime = curr->vruntime; |
02e0431a PZ |
333 | |
334 | cfs_rq->min_vruntime = | |
b0ffd246 | 335 | max_vruntime(cfs_rq->min_vruntime, vruntime); |
bf0f6f24 IM |
336 | } |
337 | ||
b7cc0896 | 338 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 339 | { |
429d43bc | 340 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 341 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
342 | unsigned long delta_exec; |
343 | ||
344 | if (unlikely(!curr)) | |
345 | return; | |
346 | ||
347 | /* | |
348 | * Get the amount of time the current task was running | |
349 | * since the last time we changed load (this cannot | |
350 | * overflow on 32 bits): | |
351 | */ | |
8ebc91d9 | 352 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 353 | |
8ebc91d9 IM |
354 | __update_curr(cfs_rq, curr, delta_exec); |
355 | curr->exec_start = now; | |
d842de87 SV |
356 | |
357 | if (entity_is_task(curr)) { | |
358 | struct task_struct *curtask = task_of(curr); | |
359 | ||
360 | cpuacct_charge(curtask, delta_exec); | |
361 | } | |
bf0f6f24 IM |
362 | } |
363 | ||
364 | static inline void | |
5870db5b | 365 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 366 | { |
d281918d | 367 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
368 | } |
369 | ||
bf0f6f24 IM |
370 | /* |
371 | * Task is being enqueued - update stats: | |
372 | */ | |
d2417e5a | 373 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 374 | { |
bf0f6f24 IM |
375 | /* |
376 | * Are we enqueueing a waiting task? (for current tasks | |
377 | * a dequeue/enqueue event is a NOP) | |
378 | */ | |
429d43bc | 379 | if (se != cfs_rq->curr) |
5870db5b | 380 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
381 | } |
382 | ||
bf0f6f24 | 383 | static void |
9ef0a961 | 384 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 385 | { |
bbdba7c0 IM |
386 | schedstat_set(se->wait_max, max(se->wait_max, |
387 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6cfb0d5d | 388 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
389 | } |
390 | ||
391 | static inline void | |
19b6a2e3 | 392 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 393 | { |
bf0f6f24 IM |
394 | /* |
395 | * Mark the end of the wait period if dequeueing a | |
396 | * waiting task: | |
397 | */ | |
429d43bc | 398 | if (se != cfs_rq->curr) |
9ef0a961 | 399 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
400 | } |
401 | ||
402 | /* | |
403 | * We are picking a new current task - update its stats: | |
404 | */ | |
405 | static inline void | |
79303e9e | 406 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
407 | { |
408 | /* | |
409 | * We are starting a new run period: | |
410 | */ | |
d281918d | 411 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
412 | } |
413 | ||
bf0f6f24 IM |
414 | /************************************************** |
415 | * Scheduling class queueing methods: | |
416 | */ | |
417 | ||
30cfdcfc DA |
418 | static void |
419 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
420 | { | |
421 | update_load_add(&cfs_rq->load, se->load.weight); | |
422 | cfs_rq->nr_running++; | |
423 | se->on_rq = 1; | |
424 | } | |
425 | ||
426 | static void | |
427 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
428 | { | |
429 | update_load_sub(&cfs_rq->load, se->load.weight); | |
430 | cfs_rq->nr_running--; | |
431 | se->on_rq = 0; | |
432 | } | |
433 | ||
2396af69 | 434 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 435 | { |
bf0f6f24 IM |
436 | #ifdef CONFIG_SCHEDSTATS |
437 | if (se->sleep_start) { | |
d281918d | 438 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 439 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
440 | |
441 | if ((s64)delta < 0) | |
442 | delta = 0; | |
443 | ||
444 | if (unlikely(delta > se->sleep_max)) | |
445 | se->sleep_max = delta; | |
446 | ||
447 | se->sleep_start = 0; | |
448 | se->sum_sleep_runtime += delta; | |
9745512c AV |
449 | |
450 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
451 | } |
452 | if (se->block_start) { | |
d281918d | 453 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 454 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
455 | |
456 | if ((s64)delta < 0) | |
457 | delta = 0; | |
458 | ||
459 | if (unlikely(delta > se->block_max)) | |
460 | se->block_max = delta; | |
461 | ||
462 | se->block_start = 0; | |
463 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
464 | |
465 | /* | |
466 | * Blocking time is in units of nanosecs, so shift by 20 to | |
467 | * get a milliseconds-range estimation of the amount of | |
468 | * time that the task spent sleeping: | |
469 | */ | |
470 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 471 | |
30084fbd IM |
472 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
473 | delta >> 20); | |
474 | } | |
9745512c | 475 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
476 | } |
477 | #endif | |
478 | } | |
479 | ||
ddc97297 PZ |
480 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
481 | { | |
482 | #ifdef CONFIG_SCHED_DEBUG | |
483 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
484 | ||
485 | if (d < 0) | |
486 | d = -d; | |
487 | ||
488 | if (d > 3*sysctl_sched_latency) | |
489 | schedstat_inc(cfs_rq, nr_spread_over); | |
490 | #endif | |
491 | } | |
492 | ||
aeb73b04 PZ |
493 | static void |
494 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
495 | { | |
67e9fb2a | 496 | u64 vruntime; |
aeb73b04 | 497 | |
67e9fb2a | 498 | vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 499 | |
06877c33 | 500 | if (sched_feat(TREE_AVG)) { |
94dfb5e7 PZ |
501 | struct sched_entity *last = __pick_last_entity(cfs_rq); |
502 | if (last) { | |
67e9fb2a PZ |
503 | vruntime += last->vruntime; |
504 | vruntime >>= 1; | |
94dfb5e7 | 505 | } |
67e9fb2a | 506 | } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running) |
647e7cac | 507 | vruntime += sched_vslice(cfs_rq)/2; |
94dfb5e7 | 508 | |
2cb8600e PZ |
509 | /* |
510 | * The 'current' period is already promised to the current tasks, | |
511 | * however the extra weight of the new task will slow them down a | |
512 | * little, place the new task so that it fits in the slot that | |
513 | * stays open at the end. | |
514 | */ | |
94dfb5e7 | 515 | if (initial && sched_feat(START_DEBIT)) |
647e7cac | 516 | vruntime += sched_vslice_add(cfs_rq, se); |
aeb73b04 | 517 | |
8465e792 | 518 | if (!initial) { |
2cb8600e | 519 | /* sleeps upto a single latency don't count. */ |
6cbf1c12 | 520 | if (sched_feat(NEW_FAIR_SLEEPERS) && entity_is_task(se)) |
94359f05 IM |
521 | vruntime -= sysctl_sched_latency; |
522 | ||
2cb8600e PZ |
523 | /* ensure we never gain time by being placed backwards. */ |
524 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
525 | } |
526 | ||
67e9fb2a | 527 | se->vruntime = vruntime; |
aeb73b04 PZ |
528 | } |
529 | ||
bf0f6f24 | 530 | static void |
83b699ed | 531 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
532 | { |
533 | /* | |
a2a2d680 | 534 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 535 | */ |
b7cc0896 | 536 | update_curr(cfs_rq); |
bf0f6f24 | 537 | |
e9acbff6 | 538 | if (wakeup) { |
aeb73b04 | 539 | place_entity(cfs_rq, se, 0); |
2396af69 | 540 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 541 | } |
bf0f6f24 | 542 | |
d2417e5a | 543 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 544 | check_spread(cfs_rq, se); |
83b699ed SV |
545 | if (se != cfs_rq->curr) |
546 | __enqueue_entity(cfs_rq, se); | |
30cfdcfc | 547 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 IM |
548 | } |
549 | ||
550 | static void | |
525c2716 | 551 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 552 | { |
a2a2d680 DA |
553 | /* |
554 | * Update run-time statistics of the 'current'. | |
555 | */ | |
556 | update_curr(cfs_rq); | |
557 | ||
19b6a2e3 | 558 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 559 | if (sleep) { |
67e9fb2a | 560 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
561 | if (entity_is_task(se)) { |
562 | struct task_struct *tsk = task_of(se); | |
563 | ||
564 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 565 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 566 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 567 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 568 | } |
db36cc7d | 569 | #endif |
67e9fb2a PZ |
570 | } |
571 | ||
83b699ed | 572 | if (se != cfs_rq->curr) |
30cfdcfc DA |
573 | __dequeue_entity(cfs_rq, se); |
574 | account_entity_dequeue(cfs_rq, se); | |
bf0f6f24 IM |
575 | } |
576 | ||
577 | /* | |
578 | * Preempt the current task with a newly woken task if needed: | |
579 | */ | |
7c92e54f | 580 | static void |
2e09bf55 | 581 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 582 | { |
11697830 PZ |
583 | unsigned long ideal_runtime, delta_exec; |
584 | ||
6d0f0ebd | 585 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 586 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
3e3e13f3 | 587 | if (delta_exec > ideal_runtime) |
bf0f6f24 IM |
588 | resched_task(rq_of(cfs_rq)->curr); |
589 | } | |
590 | ||
83b699ed | 591 | static void |
8494f412 | 592 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 593 | { |
83b699ed SV |
594 | /* 'current' is not kept within the tree. */ |
595 | if (se->on_rq) { | |
596 | /* | |
597 | * Any task has to be enqueued before it get to execute on | |
598 | * a CPU. So account for the time it spent waiting on the | |
599 | * runqueue. | |
600 | */ | |
601 | update_stats_wait_end(cfs_rq, se); | |
602 | __dequeue_entity(cfs_rq, se); | |
603 | } | |
604 | ||
79303e9e | 605 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 606 | cfs_rq->curr = se; |
eba1ed4b IM |
607 | #ifdef CONFIG_SCHEDSTATS |
608 | /* | |
609 | * Track our maximum slice length, if the CPU's load is at | |
610 | * least twice that of our own weight (i.e. dont track it | |
611 | * when there are only lesser-weight tasks around): | |
612 | */ | |
495eca49 | 613 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
614 | se->slice_max = max(se->slice_max, |
615 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
616 | } | |
617 | #endif | |
4a55b450 | 618 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
619 | } |
620 | ||
9948f4b2 | 621 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 622 | { |
08ec3df5 | 623 | struct sched_entity *se = NULL; |
bf0f6f24 | 624 | |
08ec3df5 DA |
625 | if (first_fair(cfs_rq)) { |
626 | se = __pick_next_entity(cfs_rq); | |
627 | set_next_entity(cfs_rq, se); | |
628 | } | |
bf0f6f24 IM |
629 | |
630 | return se; | |
631 | } | |
632 | ||
ab6cde26 | 633 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
634 | { |
635 | /* | |
636 | * If still on the runqueue then deactivate_task() | |
637 | * was not called and update_curr() has to be done: | |
638 | */ | |
639 | if (prev->on_rq) | |
b7cc0896 | 640 | update_curr(cfs_rq); |
bf0f6f24 | 641 | |
ddc97297 | 642 | check_spread(cfs_rq, prev); |
30cfdcfc | 643 | if (prev->on_rq) { |
5870db5b | 644 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
645 | /* Put 'current' back into the tree. */ |
646 | __enqueue_entity(cfs_rq, prev); | |
647 | } | |
429d43bc | 648 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
649 | } |
650 | ||
8f4d37ec PZ |
651 | static void |
652 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 653 | { |
bf0f6f24 | 654 | /* |
30cfdcfc | 655 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 656 | */ |
30cfdcfc | 657 | update_curr(cfs_rq); |
bf0f6f24 | 658 | |
8f4d37ec PZ |
659 | #ifdef CONFIG_SCHED_HRTICK |
660 | /* | |
661 | * queued ticks are scheduled to match the slice, so don't bother | |
662 | * validating it and just reschedule. | |
663 | */ | |
664 | if (queued) | |
665 | return resched_task(rq_of(cfs_rq)->curr); | |
666 | /* | |
667 | * don't let the period tick interfere with the hrtick preemption | |
668 | */ | |
669 | if (!sched_feat(DOUBLE_TICK) && | |
670 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
671 | return; | |
672 | #endif | |
673 | ||
ce6c1311 | 674 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 675 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
676 | } |
677 | ||
678 | /************************************************** | |
679 | * CFS operations on tasks: | |
680 | */ | |
681 | ||
682 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
683 | ||
684 | /* Walk up scheduling entities hierarchy */ | |
685 | #define for_each_sched_entity(se) \ | |
686 | for (; se; se = se->parent) | |
687 | ||
688 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
689 | { | |
690 | return p->se.cfs_rq; | |
691 | } | |
692 | ||
693 | /* runqueue on which this entity is (to be) queued */ | |
694 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
695 | { | |
696 | return se->cfs_rq; | |
697 | } | |
698 | ||
699 | /* runqueue "owned" by this group */ | |
700 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
701 | { | |
702 | return grp->my_q; | |
703 | } | |
704 | ||
705 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
706 | * another cpu ('this_cpu') | |
707 | */ | |
708 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
709 | { | |
29f59db3 | 710 | return cfs_rq->tg->cfs_rq[this_cpu]; |
bf0f6f24 IM |
711 | } |
712 | ||
713 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
714 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
ec2c507f | 715 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
bf0f6f24 | 716 | |
fad095a7 SV |
717 | /* Do the two (enqueued) entities belong to the same group ? */ |
718 | static inline int | |
719 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 | 720 | { |
fad095a7 | 721 | if (se->cfs_rq == pse->cfs_rq) |
bf0f6f24 IM |
722 | return 1; |
723 | ||
724 | return 0; | |
725 | } | |
726 | ||
fad095a7 SV |
727 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
728 | { | |
729 | return se->parent; | |
730 | } | |
731 | ||
6b2d7700 SV |
732 | #define GROUP_IMBALANCE_PCT 20 |
733 | ||
bf0f6f24 IM |
734 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
735 | ||
736 | #define for_each_sched_entity(se) \ | |
737 | for (; se; se = NULL) | |
738 | ||
739 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
740 | { | |
741 | return &task_rq(p)->cfs; | |
742 | } | |
743 | ||
744 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
745 | { | |
746 | struct task_struct *p = task_of(se); | |
747 | struct rq *rq = task_rq(p); | |
748 | ||
749 | return &rq->cfs; | |
750 | } | |
751 | ||
752 | /* runqueue "owned" by this group */ | |
753 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
754 | { | |
755 | return NULL; | |
756 | } | |
757 | ||
758 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
759 | { | |
760 | return &cpu_rq(this_cpu)->cfs; | |
761 | } | |
762 | ||
763 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
764 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
765 | ||
fad095a7 SV |
766 | static inline int |
767 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 IM |
768 | { |
769 | return 1; | |
770 | } | |
771 | ||
fad095a7 SV |
772 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
773 | { | |
774 | return NULL; | |
775 | } | |
776 | ||
bf0f6f24 IM |
777 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
778 | ||
8f4d37ec PZ |
779 | #ifdef CONFIG_SCHED_HRTICK |
780 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
781 | { | |
782 | int requeue = rq->curr == p; | |
783 | struct sched_entity *se = &p->se; | |
784 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
785 | ||
786 | WARN_ON(task_rq(p) != rq); | |
787 | ||
788 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
789 | u64 slice = sched_slice(cfs_rq, se); | |
790 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
791 | s64 delta = slice - ran; | |
792 | ||
793 | if (delta < 0) { | |
794 | if (rq->curr == p) | |
795 | resched_task(p); | |
796 | return; | |
797 | } | |
798 | ||
799 | /* | |
800 | * Don't schedule slices shorter than 10000ns, that just | |
801 | * doesn't make sense. Rely on vruntime for fairness. | |
802 | */ | |
803 | if (!requeue) | |
804 | delta = max(10000LL, delta); | |
805 | ||
806 | hrtick_start(rq, delta, requeue); | |
807 | } | |
808 | } | |
809 | #else | |
810 | static inline void | |
811 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
812 | { | |
813 | } | |
814 | #endif | |
815 | ||
bf0f6f24 IM |
816 | /* |
817 | * The enqueue_task method is called before nr_running is | |
818 | * increased. Here we update the fair scheduling stats and | |
819 | * then put the task into the rbtree: | |
820 | */ | |
fd390f6a | 821 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
822 | { |
823 | struct cfs_rq *cfs_rq; | |
58e2d4ca SV |
824 | struct sched_entity *se = &p->se, |
825 | *topse = NULL; /* Highest schedulable entity */ | |
826 | int incload = 1; | |
bf0f6f24 IM |
827 | |
828 | for_each_sched_entity(se) { | |
58e2d4ca SV |
829 | topse = se; |
830 | if (se->on_rq) { | |
831 | incload = 0; | |
bf0f6f24 | 832 | break; |
58e2d4ca | 833 | } |
bf0f6f24 | 834 | cfs_rq = cfs_rq_of(se); |
83b699ed | 835 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 836 | wakeup = 1; |
bf0f6f24 | 837 | } |
58e2d4ca SV |
838 | /* Increment cpu load if we just enqueued the first task of a group on |
839 | * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs | |
840 | * at the highest grouping level. | |
841 | */ | |
842 | if (incload) | |
843 | inc_cpu_load(rq, topse->load.weight); | |
8f4d37ec PZ |
844 | |
845 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
846 | } |
847 | ||
848 | /* | |
849 | * The dequeue_task method is called before nr_running is | |
850 | * decreased. We remove the task from the rbtree and | |
851 | * update the fair scheduling stats: | |
852 | */ | |
f02231e5 | 853 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
854 | { |
855 | struct cfs_rq *cfs_rq; | |
58e2d4ca SV |
856 | struct sched_entity *se = &p->se, |
857 | *topse = NULL; /* Highest schedulable entity */ | |
858 | int decload = 1; | |
bf0f6f24 IM |
859 | |
860 | for_each_sched_entity(se) { | |
58e2d4ca | 861 | topse = se; |
bf0f6f24 | 862 | cfs_rq = cfs_rq_of(se); |
525c2716 | 863 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 864 | /* Don't dequeue parent if it has other entities besides us */ |
58e2d4ca SV |
865 | if (cfs_rq->load.weight) { |
866 | if (parent_entity(se)) | |
867 | decload = 0; | |
bf0f6f24 | 868 | break; |
58e2d4ca | 869 | } |
b9fa3df3 | 870 | sleep = 1; |
bf0f6f24 | 871 | } |
58e2d4ca SV |
872 | /* Decrement cpu load if we just dequeued the last task of a group on |
873 | * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs | |
874 | * at the highest grouping level. | |
875 | */ | |
876 | if (decload) | |
877 | dec_cpu_load(rq, topse->load.weight); | |
8f4d37ec PZ |
878 | |
879 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
880 | } |
881 | ||
882 | /* | |
1799e35d IM |
883 | * sched_yield() support is very simple - we dequeue and enqueue. |
884 | * | |
885 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 886 | */ |
4530d7ab | 887 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 888 | { |
db292ca3 IM |
889 | struct task_struct *curr = rq->curr; |
890 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
891 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
892 | |
893 | /* | |
1799e35d IM |
894 | * Are we the only task in the tree? |
895 | */ | |
896 | if (unlikely(cfs_rq->nr_running == 1)) | |
897 | return; | |
898 | ||
db292ca3 | 899 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
1799e35d IM |
900 | __update_rq_clock(rq); |
901 | /* | |
a2a2d680 | 902 | * Update run-time statistics of the 'current'. |
1799e35d | 903 | */ |
2b1e315d | 904 | update_curr(cfs_rq); |
1799e35d IM |
905 | |
906 | return; | |
907 | } | |
908 | /* | |
909 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 910 | */ |
2b1e315d | 911 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
912 | /* |
913 | * Already in the rightmost position? | |
914 | */ | |
2b1e315d | 915 | if (unlikely(rightmost->vruntime < se->vruntime)) |
1799e35d IM |
916 | return; |
917 | ||
918 | /* | |
919 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
920 | * Upon rescheduling, sched_class::put_prev_task() will place |
921 | * 'current' within the tree based on its new key value. | |
1799e35d | 922 | */ |
30cfdcfc | 923 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
924 | } |
925 | ||
e7693a36 GH |
926 | /* |
927 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
928 | * not idle and an idle cpu is available. The span of cpus to | |
929 | * search starts with cpus closest then further out as needed, | |
930 | * so we always favor a closer, idle cpu. | |
931 | * | |
932 | * Returns the CPU we should wake onto. | |
933 | */ | |
934 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
935 | static int wake_idle(int cpu, struct task_struct *p) | |
936 | { | |
937 | cpumask_t tmp; | |
938 | struct sched_domain *sd; | |
939 | int i; | |
940 | ||
941 | /* | |
942 | * If it is idle, then it is the best cpu to run this task. | |
943 | * | |
944 | * This cpu is also the best, if it has more than one task already. | |
945 | * Siblings must be also busy(in most cases) as they didn't already | |
946 | * pickup the extra load from this cpu and hence we need not check | |
947 | * sibling runqueue info. This will avoid the checks and cache miss | |
948 | * penalities associated with that. | |
949 | */ | |
950 | if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1) | |
951 | return cpu; | |
952 | ||
953 | for_each_domain(cpu, sd) { | |
954 | if (sd->flags & SD_WAKE_IDLE) { | |
955 | cpus_and(tmp, sd->span, p->cpus_allowed); | |
956 | for_each_cpu_mask(i, tmp) { | |
957 | if (idle_cpu(i)) { | |
958 | if (i != task_cpu(p)) { | |
959 | schedstat_inc(p, | |
960 | se.nr_wakeups_idle); | |
961 | } | |
962 | return i; | |
963 | } | |
964 | } | |
965 | } else { | |
966 | break; | |
967 | } | |
968 | } | |
969 | return cpu; | |
970 | } | |
971 | #else | |
972 | static inline int wake_idle(int cpu, struct task_struct *p) | |
973 | { | |
974 | return cpu; | |
975 | } | |
976 | #endif | |
977 | ||
978 | #ifdef CONFIG_SMP | |
979 | static int select_task_rq_fair(struct task_struct *p, int sync) | |
980 | { | |
981 | int cpu, this_cpu; | |
982 | struct rq *rq; | |
983 | struct sched_domain *sd, *this_sd = NULL; | |
984 | int new_cpu; | |
985 | ||
986 | cpu = task_cpu(p); | |
987 | rq = task_rq(p); | |
988 | this_cpu = smp_processor_id(); | |
989 | new_cpu = cpu; | |
990 | ||
9ec3b77e DA |
991 | if (cpu == this_cpu) |
992 | goto out_set_cpu; | |
993 | ||
e7693a36 GH |
994 | for_each_domain(this_cpu, sd) { |
995 | if (cpu_isset(cpu, sd->span)) { | |
996 | this_sd = sd; | |
997 | break; | |
998 | } | |
999 | } | |
1000 | ||
1001 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) | |
1002 | goto out_set_cpu; | |
1003 | ||
1004 | /* | |
1005 | * Check for affine wakeup and passive balancing possibilities. | |
1006 | */ | |
1007 | if (this_sd) { | |
1008 | int idx = this_sd->wake_idx; | |
1009 | unsigned int imbalance; | |
1010 | unsigned long load, this_load; | |
1011 | ||
1012 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1013 | ||
1014 | load = source_load(cpu, idx); | |
1015 | this_load = target_load(this_cpu, idx); | |
1016 | ||
1017 | new_cpu = this_cpu; /* Wake to this CPU if we can */ | |
1018 | ||
1019 | if (this_sd->flags & SD_WAKE_AFFINE) { | |
1020 | unsigned long tl = this_load; | |
1021 | unsigned long tl_per_task; | |
1022 | ||
1023 | /* | |
1024 | * Attract cache-cold tasks on sync wakeups: | |
1025 | */ | |
1026 | if (sync && !task_hot(p, rq->clock, this_sd)) | |
1027 | goto out_set_cpu; | |
1028 | ||
1029 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1030 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1031 | ||
1032 | /* | |
1033 | * If sync wakeup then subtract the (maximum possible) | |
1034 | * effect of the currently running task from the load | |
1035 | * of the current CPU: | |
1036 | */ | |
1037 | if (sync) | |
1038 | tl -= current->se.load.weight; | |
1039 | ||
1040 | if ((tl <= load && | |
1041 | tl + target_load(cpu, idx) <= tl_per_task) || | |
1042 | 100*(tl + p->se.load.weight) <= imbalance*load) { | |
1043 | /* | |
1044 | * This domain has SD_WAKE_AFFINE and | |
1045 | * p is cache cold in this domain, and | |
1046 | * there is no bad imbalance. | |
1047 | */ | |
1048 | schedstat_inc(this_sd, ttwu_move_affine); | |
1049 | schedstat_inc(p, se.nr_wakeups_affine); | |
1050 | goto out_set_cpu; | |
1051 | } | |
1052 | } | |
1053 | ||
1054 | /* | |
1055 | * Start passive balancing when half the imbalance_pct | |
1056 | * limit is reached. | |
1057 | */ | |
1058 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1059 | if (imbalance*this_load <= 100*load) { | |
1060 | schedstat_inc(this_sd, ttwu_move_balance); | |
1061 | schedstat_inc(p, se.nr_wakeups_passive); | |
1062 | goto out_set_cpu; | |
1063 | } | |
1064 | } | |
1065 | } | |
1066 | ||
1067 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1068 | out_set_cpu: | |
1069 | return wake_idle(new_cpu, p); | |
1070 | } | |
1071 | #endif /* CONFIG_SMP */ | |
1072 | ||
1073 | ||
bf0f6f24 IM |
1074 | /* |
1075 | * Preempt the current task with a newly woken task if needed: | |
1076 | */ | |
2e09bf55 | 1077 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1078 | { |
1079 | struct task_struct *curr = rq->curr; | |
fad095a7 | 1080 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
8651a86c | 1081 | struct sched_entity *se = &curr->se, *pse = &p->se; |
502d26b5 | 1082 | unsigned long gran; |
bf0f6f24 IM |
1083 | |
1084 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 1085 | update_rq_clock(rq); |
b7cc0896 | 1086 | update_curr(cfs_rq); |
bf0f6f24 IM |
1087 | resched_task(curr); |
1088 | return; | |
1089 | } | |
91c234b4 IM |
1090 | /* |
1091 | * Batch tasks do not preempt (their preemption is driven by | |
1092 | * the tick): | |
1093 | */ | |
1094 | if (unlikely(p->policy == SCHED_BATCH)) | |
1095 | return; | |
bf0f6f24 | 1096 | |
77d9cc44 IM |
1097 | if (!sched_feat(WAKEUP_PREEMPT)) |
1098 | return; | |
8651a86c | 1099 | |
77d9cc44 IM |
1100 | while (!is_same_group(se, pse)) { |
1101 | se = parent_entity(se); | |
1102 | pse = parent_entity(pse); | |
ce6c1311 | 1103 | } |
77d9cc44 | 1104 | |
77d9cc44 IM |
1105 | gran = sysctl_sched_wakeup_granularity; |
1106 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1107 | gran = calc_delta_fair(gran, &se->load); | |
1108 | ||
502d26b5 | 1109 | if (pse->vruntime + gran < se->vruntime) |
77d9cc44 | 1110 | resched_task(curr); |
bf0f6f24 IM |
1111 | } |
1112 | ||
fb8d4724 | 1113 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1114 | { |
8f4d37ec | 1115 | struct task_struct *p; |
bf0f6f24 IM |
1116 | struct cfs_rq *cfs_rq = &rq->cfs; |
1117 | struct sched_entity *se; | |
1118 | ||
1119 | if (unlikely(!cfs_rq->nr_running)) | |
1120 | return NULL; | |
1121 | ||
1122 | do { | |
9948f4b2 | 1123 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1124 | cfs_rq = group_cfs_rq(se); |
1125 | } while (cfs_rq); | |
1126 | ||
8f4d37ec PZ |
1127 | p = task_of(se); |
1128 | hrtick_start_fair(rq, p); | |
1129 | ||
1130 | return p; | |
bf0f6f24 IM |
1131 | } |
1132 | ||
1133 | /* | |
1134 | * Account for a descheduled task: | |
1135 | */ | |
31ee529c | 1136 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1137 | { |
1138 | struct sched_entity *se = &prev->se; | |
1139 | struct cfs_rq *cfs_rq; | |
1140 | ||
1141 | for_each_sched_entity(se) { | |
1142 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1143 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1144 | } |
1145 | } | |
1146 | ||
681f3e68 | 1147 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1148 | /************************************************** |
1149 | * Fair scheduling class load-balancing methods: | |
1150 | */ | |
1151 | ||
1152 | /* | |
1153 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1154 | * during the whole iteration, the current task might be | |
1155 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1156 | * achieve that by always pre-iterating before returning | |
1157 | * the current task: | |
1158 | */ | |
a9957449 | 1159 | static struct task_struct * |
bf0f6f24 IM |
1160 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) |
1161 | { | |
1162 | struct task_struct *p; | |
1163 | ||
1164 | if (!curr) | |
1165 | return NULL; | |
1166 | ||
1167 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1168 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1169 | ||
1170 | return p; | |
1171 | } | |
1172 | ||
1173 | static struct task_struct *load_balance_start_fair(void *arg) | |
1174 | { | |
1175 | struct cfs_rq *cfs_rq = arg; | |
1176 | ||
1177 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1178 | } | |
1179 | ||
1180 | static struct task_struct *load_balance_next_fair(void *arg) | |
1181 | { | |
1182 | struct cfs_rq *cfs_rq = arg; | |
1183 | ||
1184 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1185 | } | |
1186 | ||
43010659 | 1187 | static unsigned long |
bf0f6f24 | 1188 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1189 | unsigned long max_load_move, |
a4ac01c3 PW |
1190 | struct sched_domain *sd, enum cpu_idle_type idle, |
1191 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1192 | { |
1193 | struct cfs_rq *busy_cfs_rq; | |
bf0f6f24 IM |
1194 | long rem_load_move = max_load_move; |
1195 | struct rq_iterator cfs_rq_iterator; | |
6b2d7700 | 1196 | unsigned long load_moved; |
bf0f6f24 IM |
1197 | |
1198 | cfs_rq_iterator.start = load_balance_start_fair; | |
1199 | cfs_rq_iterator.next = load_balance_next_fair; | |
1200 | ||
1201 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1202 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6b2d7700 SV |
1203 | struct cfs_rq *this_cfs_rq = busy_cfs_rq->tg->cfs_rq[this_cpu]; |
1204 | unsigned long maxload, task_load, group_weight; | |
1205 | unsigned long thisload, per_task_load; | |
1206 | struct sched_entity *se = busy_cfs_rq->tg->se[busiest->cpu]; | |
bf0f6f24 | 1207 | |
6b2d7700 SV |
1208 | task_load = busy_cfs_rq->load.weight; |
1209 | group_weight = se->load.weight; | |
bf0f6f24 | 1210 | |
6b2d7700 SV |
1211 | /* |
1212 | * 'group_weight' is contributed by tasks of total weight | |
1213 | * 'task_load'. To move 'rem_load_move' worth of weight only, | |
1214 | * we need to move a maximum task load of: | |
1215 | * | |
1216 | * maxload = (remload / group_weight) * task_load; | |
1217 | */ | |
1218 | maxload = (rem_load_move * task_load) / group_weight; | |
1219 | ||
1220 | if (!maxload || !task_load) | |
bf0f6f24 IM |
1221 | continue; |
1222 | ||
6b2d7700 SV |
1223 | per_task_load = task_load / busy_cfs_rq->nr_running; |
1224 | /* | |
1225 | * balance_tasks will try to forcibly move atleast one task if | |
1226 | * possible (because of SCHED_LOAD_SCALE_FUZZ). Avoid that if | |
1227 | * maxload is less than GROUP_IMBALANCE_FUZZ% the per_task_load. | |
1228 | */ | |
1229 | if (100 * maxload < GROUP_IMBALANCE_PCT * per_task_load) | |
1230 | continue; | |
bf0f6f24 | 1231 | |
6b2d7700 SV |
1232 | /* Disable priority-based load balance */ |
1233 | *this_best_prio = 0; | |
1234 | thisload = this_cfs_rq->load.weight; | |
a4ac01c3 | 1235 | #else |
e56f31aa | 1236 | # define maxload rem_load_move |
a4ac01c3 | 1237 | #endif |
e1d1484f PW |
1238 | /* |
1239 | * pass busy_cfs_rq argument into | |
bf0f6f24 IM |
1240 | * load_balance_[start|next]_fair iterators |
1241 | */ | |
1242 | cfs_rq_iterator.arg = busy_cfs_rq; | |
6b2d7700 | 1243 | load_moved = balance_tasks(this_rq, this_cpu, busiest, |
e1d1484f PW |
1244 | maxload, sd, idle, all_pinned, |
1245 | this_best_prio, | |
1246 | &cfs_rq_iterator); | |
bf0f6f24 | 1247 | |
6b2d7700 SV |
1248 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1249 | /* | |
1250 | * load_moved holds the task load that was moved. The | |
1251 | * effective (group) weight moved would be: | |
1252 | * load_moved_eff = load_moved/task_load * group_weight; | |
1253 | */ | |
1254 | load_moved = (group_weight * load_moved) / task_load; | |
1255 | ||
1256 | /* Adjust shares on both cpus to reflect load_moved */ | |
1257 | group_weight -= load_moved; | |
1258 | set_se_shares(se, group_weight); | |
1259 | ||
1260 | se = busy_cfs_rq->tg->se[this_cpu]; | |
1261 | if (!thisload) | |
1262 | group_weight = load_moved; | |
1263 | else | |
1264 | group_weight = se->load.weight + load_moved; | |
1265 | set_se_shares(se, group_weight); | |
1266 | #endif | |
1267 | ||
1268 | rem_load_move -= load_moved; | |
1269 | ||
e1d1484f | 1270 | if (rem_load_move <= 0) |
bf0f6f24 IM |
1271 | break; |
1272 | } | |
1273 | ||
43010659 | 1274 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1275 | } |
1276 | ||
e1d1484f PW |
1277 | static int |
1278 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1279 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1280 | { | |
1281 | struct cfs_rq *busy_cfs_rq; | |
1282 | struct rq_iterator cfs_rq_iterator; | |
1283 | ||
1284 | cfs_rq_iterator.start = load_balance_start_fair; | |
1285 | cfs_rq_iterator.next = load_balance_next_fair; | |
1286 | ||
1287 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1288 | /* | |
1289 | * pass busy_cfs_rq argument into | |
1290 | * load_balance_[start|next]_fair iterators | |
1291 | */ | |
1292 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1293 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1294 | &cfs_rq_iterator)) | |
1295 | return 1; | |
1296 | } | |
1297 | ||
1298 | return 0; | |
1299 | } | |
681f3e68 | 1300 | #endif |
e1d1484f | 1301 | |
bf0f6f24 IM |
1302 | /* |
1303 | * scheduler tick hitting a task of our scheduling class: | |
1304 | */ | |
8f4d37ec | 1305 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1306 | { |
1307 | struct cfs_rq *cfs_rq; | |
1308 | struct sched_entity *se = &curr->se; | |
1309 | ||
1310 | for_each_sched_entity(se) { | |
1311 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1312 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1313 | } |
1314 | } | |
1315 | ||
8eb172d9 | 1316 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
4d78e7b6 | 1317 | |
bf0f6f24 IM |
1318 | /* |
1319 | * Share the fairness runtime between parent and child, thus the | |
1320 | * total amount of pressure for CPU stays equal - new tasks | |
1321 | * get a chance to run but frequent forkers are not allowed to | |
1322 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1323 | * the child is not running yet. | |
1324 | */ | |
ee0827d8 | 1325 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1326 | { |
1327 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1328 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1329 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1330 | |
1331 | sched_info_queued(p); | |
1332 | ||
7109c442 | 1333 | update_curr(cfs_rq); |
aeb73b04 | 1334 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1335 | |
3c90e6e9 | 1336 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1337 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1338 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1339 | /* |
edcb60a3 IM |
1340 | * Upon rescheduling, sched_class::put_prev_task() will place |
1341 | * 'current' within the tree based on its new key value. | |
1342 | */ | |
4d78e7b6 | 1343 | swap(curr->vruntime, se->vruntime); |
4d78e7b6 | 1344 | } |
bf0f6f24 | 1345 | |
b9dca1e0 | 1346 | enqueue_task_fair(rq, p, 0); |
bb61c210 | 1347 | resched_task(rq->curr); |
bf0f6f24 IM |
1348 | } |
1349 | ||
cb469845 SR |
1350 | /* |
1351 | * Priority of the task has changed. Check to see if we preempt | |
1352 | * the current task. | |
1353 | */ | |
1354 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1355 | int oldprio, int running) | |
1356 | { | |
1357 | /* | |
1358 | * Reschedule if we are currently running on this runqueue and | |
1359 | * our priority decreased, or if we are not currently running on | |
1360 | * this runqueue and our priority is higher than the current's | |
1361 | */ | |
1362 | if (running) { | |
1363 | if (p->prio > oldprio) | |
1364 | resched_task(rq->curr); | |
1365 | } else | |
1366 | check_preempt_curr(rq, p); | |
1367 | } | |
1368 | ||
1369 | /* | |
1370 | * We switched to the sched_fair class. | |
1371 | */ | |
1372 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1373 | int running) | |
1374 | { | |
1375 | /* | |
1376 | * We were most likely switched from sched_rt, so | |
1377 | * kick off the schedule if running, otherwise just see | |
1378 | * if we can still preempt the current task. | |
1379 | */ | |
1380 | if (running) | |
1381 | resched_task(rq->curr); | |
1382 | else | |
1383 | check_preempt_curr(rq, p); | |
1384 | } | |
1385 | ||
83b699ed SV |
1386 | /* Account for a task changing its policy or group. |
1387 | * | |
1388 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1389 | * migrates between groups/classes. | |
1390 | */ | |
1391 | static void set_curr_task_fair(struct rq *rq) | |
1392 | { | |
1393 | struct sched_entity *se = &rq->curr->se; | |
1394 | ||
1395 | for_each_sched_entity(se) | |
1396 | set_next_entity(cfs_rq_of(se), se); | |
1397 | } | |
1398 | ||
bf0f6f24 IM |
1399 | /* |
1400 | * All the scheduling class methods: | |
1401 | */ | |
5522d5d5 IM |
1402 | static const struct sched_class fair_sched_class = { |
1403 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1404 | .enqueue_task = enqueue_task_fair, |
1405 | .dequeue_task = dequeue_task_fair, | |
1406 | .yield_task = yield_task_fair, | |
e7693a36 GH |
1407 | #ifdef CONFIG_SMP |
1408 | .select_task_rq = select_task_rq_fair, | |
1409 | #endif /* CONFIG_SMP */ | |
bf0f6f24 | 1410 | |
2e09bf55 | 1411 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1412 | |
1413 | .pick_next_task = pick_next_task_fair, | |
1414 | .put_prev_task = put_prev_task_fair, | |
1415 | ||
681f3e68 | 1416 | #ifdef CONFIG_SMP |
bf0f6f24 | 1417 | .load_balance = load_balance_fair, |
e1d1484f | 1418 | .move_one_task = move_one_task_fair, |
681f3e68 | 1419 | #endif |
bf0f6f24 | 1420 | |
83b699ed | 1421 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1422 | .task_tick = task_tick_fair, |
1423 | .task_new = task_new_fair, | |
cb469845 SR |
1424 | |
1425 | .prio_changed = prio_changed_fair, | |
1426 | .switched_to = switched_to_fair, | |
bf0f6f24 IM |
1427 | }; |
1428 | ||
1429 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1430 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1431 | { |
bf0f6f24 IM |
1432 | struct cfs_rq *cfs_rq; |
1433 | ||
75c28ace SV |
1434 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1435 | print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs); | |
1436 | #endif | |
ec2c507f | 1437 | lock_task_group_list(); |
c3b64f1e | 1438 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1439 | print_cfs_rq(m, cpu, cfs_rq); |
ec2c507f | 1440 | unlock_task_group_list(); |
bf0f6f24 IM |
1441 | } |
1442 | #endif |