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)); | |
6d082592 AV |
388 | schedstat_set(se->wait_count, se->wait_count + 1); |
389 | schedstat_set(se->wait_sum, se->wait_sum + | |
390 | rq_of(cfs_rq)->clock - se->wait_start); | |
6cfb0d5d | 391 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
392 | } |
393 | ||
394 | static inline void | |
19b6a2e3 | 395 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 396 | { |
bf0f6f24 IM |
397 | /* |
398 | * Mark the end of the wait period if dequeueing a | |
399 | * waiting task: | |
400 | */ | |
429d43bc | 401 | if (se != cfs_rq->curr) |
9ef0a961 | 402 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
403 | } |
404 | ||
405 | /* | |
406 | * We are picking a new current task - update its stats: | |
407 | */ | |
408 | static inline void | |
79303e9e | 409 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
410 | { |
411 | /* | |
412 | * We are starting a new run period: | |
413 | */ | |
d281918d | 414 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
415 | } |
416 | ||
bf0f6f24 IM |
417 | /************************************************** |
418 | * Scheduling class queueing methods: | |
419 | */ | |
420 | ||
30cfdcfc DA |
421 | static void |
422 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
423 | { | |
424 | update_load_add(&cfs_rq->load, se->load.weight); | |
425 | cfs_rq->nr_running++; | |
426 | se->on_rq = 1; | |
427 | } | |
428 | ||
429 | static void | |
430 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
431 | { | |
432 | update_load_sub(&cfs_rq->load, se->load.weight); | |
433 | cfs_rq->nr_running--; | |
434 | se->on_rq = 0; | |
435 | } | |
436 | ||
2396af69 | 437 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 438 | { |
bf0f6f24 IM |
439 | #ifdef CONFIG_SCHEDSTATS |
440 | if (se->sleep_start) { | |
d281918d | 441 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 442 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
443 | |
444 | if ((s64)delta < 0) | |
445 | delta = 0; | |
446 | ||
447 | if (unlikely(delta > se->sleep_max)) | |
448 | se->sleep_max = delta; | |
449 | ||
450 | se->sleep_start = 0; | |
451 | se->sum_sleep_runtime += delta; | |
9745512c AV |
452 | |
453 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
454 | } |
455 | if (se->block_start) { | |
d281918d | 456 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 457 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
458 | |
459 | if ((s64)delta < 0) | |
460 | delta = 0; | |
461 | ||
462 | if (unlikely(delta > se->block_max)) | |
463 | se->block_max = delta; | |
464 | ||
465 | se->block_start = 0; | |
466 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
467 | |
468 | /* | |
469 | * Blocking time is in units of nanosecs, so shift by 20 to | |
470 | * get a milliseconds-range estimation of the amount of | |
471 | * time that the task spent sleeping: | |
472 | */ | |
473 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 474 | |
30084fbd IM |
475 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
476 | delta >> 20); | |
477 | } | |
9745512c | 478 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
479 | } |
480 | #endif | |
481 | } | |
482 | ||
ddc97297 PZ |
483 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
484 | { | |
485 | #ifdef CONFIG_SCHED_DEBUG | |
486 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
487 | ||
488 | if (d < 0) | |
489 | d = -d; | |
490 | ||
491 | if (d > 3*sysctl_sched_latency) | |
492 | schedstat_inc(cfs_rq, nr_spread_over); | |
493 | #endif | |
494 | } | |
495 | ||
aeb73b04 PZ |
496 | static void |
497 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
498 | { | |
67e9fb2a | 499 | u64 vruntime; |
aeb73b04 | 500 | |
67e9fb2a | 501 | vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 502 | |
06877c33 | 503 | if (sched_feat(TREE_AVG)) { |
94dfb5e7 PZ |
504 | struct sched_entity *last = __pick_last_entity(cfs_rq); |
505 | if (last) { | |
67e9fb2a PZ |
506 | vruntime += last->vruntime; |
507 | vruntime >>= 1; | |
94dfb5e7 | 508 | } |
67e9fb2a | 509 | } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running) |
647e7cac | 510 | vruntime += sched_vslice(cfs_rq)/2; |
94dfb5e7 | 511 | |
2cb8600e PZ |
512 | /* |
513 | * The 'current' period is already promised to the current tasks, | |
514 | * however the extra weight of the new task will slow them down a | |
515 | * little, place the new task so that it fits in the slot that | |
516 | * stays open at the end. | |
517 | */ | |
94dfb5e7 | 518 | if (initial && sched_feat(START_DEBIT)) |
647e7cac | 519 | vruntime += sched_vslice_add(cfs_rq, se); |
aeb73b04 | 520 | |
8465e792 | 521 | if (!initial) { |
2cb8600e | 522 | /* sleeps upto a single latency don't count. */ |
6cbf1c12 | 523 | if (sched_feat(NEW_FAIR_SLEEPERS) && entity_is_task(se)) |
94359f05 IM |
524 | vruntime -= sysctl_sched_latency; |
525 | ||
2cb8600e PZ |
526 | /* ensure we never gain time by being placed backwards. */ |
527 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
528 | } |
529 | ||
67e9fb2a | 530 | se->vruntime = vruntime; |
aeb73b04 PZ |
531 | } |
532 | ||
bf0f6f24 | 533 | static void |
83b699ed | 534 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
535 | { |
536 | /* | |
a2a2d680 | 537 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 538 | */ |
b7cc0896 | 539 | update_curr(cfs_rq); |
bf0f6f24 | 540 | |
e9acbff6 | 541 | if (wakeup) { |
aeb73b04 | 542 | place_entity(cfs_rq, se, 0); |
2396af69 | 543 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 544 | } |
bf0f6f24 | 545 | |
d2417e5a | 546 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 547 | check_spread(cfs_rq, se); |
83b699ed SV |
548 | if (se != cfs_rq->curr) |
549 | __enqueue_entity(cfs_rq, se); | |
30cfdcfc | 550 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 IM |
551 | } |
552 | ||
553 | static void | |
525c2716 | 554 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 555 | { |
a2a2d680 DA |
556 | /* |
557 | * Update run-time statistics of the 'current'. | |
558 | */ | |
559 | update_curr(cfs_rq); | |
560 | ||
19b6a2e3 | 561 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 562 | if (sleep) { |
67e9fb2a | 563 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
564 | if (entity_is_task(se)) { |
565 | struct task_struct *tsk = task_of(se); | |
566 | ||
567 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 568 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 569 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 570 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 571 | } |
db36cc7d | 572 | #endif |
67e9fb2a PZ |
573 | } |
574 | ||
83b699ed | 575 | if (se != cfs_rq->curr) |
30cfdcfc DA |
576 | __dequeue_entity(cfs_rq, se); |
577 | account_entity_dequeue(cfs_rq, se); | |
bf0f6f24 IM |
578 | } |
579 | ||
580 | /* | |
581 | * Preempt the current task with a newly woken task if needed: | |
582 | */ | |
7c92e54f | 583 | static void |
2e09bf55 | 584 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 585 | { |
11697830 PZ |
586 | unsigned long ideal_runtime, delta_exec; |
587 | ||
6d0f0ebd | 588 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 589 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
3e3e13f3 | 590 | if (delta_exec > ideal_runtime) |
bf0f6f24 IM |
591 | resched_task(rq_of(cfs_rq)->curr); |
592 | } | |
593 | ||
83b699ed | 594 | static void |
8494f412 | 595 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 596 | { |
83b699ed SV |
597 | /* 'current' is not kept within the tree. */ |
598 | if (se->on_rq) { | |
599 | /* | |
600 | * Any task has to be enqueued before it get to execute on | |
601 | * a CPU. So account for the time it spent waiting on the | |
602 | * runqueue. | |
603 | */ | |
604 | update_stats_wait_end(cfs_rq, se); | |
605 | __dequeue_entity(cfs_rq, se); | |
606 | } | |
607 | ||
79303e9e | 608 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 609 | cfs_rq->curr = se; |
eba1ed4b IM |
610 | #ifdef CONFIG_SCHEDSTATS |
611 | /* | |
612 | * Track our maximum slice length, if the CPU's load is at | |
613 | * least twice that of our own weight (i.e. dont track it | |
614 | * when there are only lesser-weight tasks around): | |
615 | */ | |
495eca49 | 616 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
617 | se->slice_max = max(se->slice_max, |
618 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
619 | } | |
620 | #endif | |
4a55b450 | 621 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
622 | } |
623 | ||
9948f4b2 | 624 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 625 | { |
08ec3df5 | 626 | struct sched_entity *se = NULL; |
bf0f6f24 | 627 | |
08ec3df5 DA |
628 | if (first_fair(cfs_rq)) { |
629 | se = __pick_next_entity(cfs_rq); | |
630 | set_next_entity(cfs_rq, se); | |
631 | } | |
bf0f6f24 IM |
632 | |
633 | return se; | |
634 | } | |
635 | ||
ab6cde26 | 636 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
637 | { |
638 | /* | |
639 | * If still on the runqueue then deactivate_task() | |
640 | * was not called and update_curr() has to be done: | |
641 | */ | |
642 | if (prev->on_rq) | |
b7cc0896 | 643 | update_curr(cfs_rq); |
bf0f6f24 | 644 | |
ddc97297 | 645 | check_spread(cfs_rq, prev); |
30cfdcfc | 646 | if (prev->on_rq) { |
5870db5b | 647 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
648 | /* Put 'current' back into the tree. */ |
649 | __enqueue_entity(cfs_rq, prev); | |
650 | } | |
429d43bc | 651 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
652 | } |
653 | ||
8f4d37ec PZ |
654 | static void |
655 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 656 | { |
bf0f6f24 | 657 | /* |
30cfdcfc | 658 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 659 | */ |
30cfdcfc | 660 | update_curr(cfs_rq); |
bf0f6f24 | 661 | |
8f4d37ec PZ |
662 | #ifdef CONFIG_SCHED_HRTICK |
663 | /* | |
664 | * queued ticks are scheduled to match the slice, so don't bother | |
665 | * validating it and just reschedule. | |
666 | */ | |
667 | if (queued) | |
668 | return resched_task(rq_of(cfs_rq)->curr); | |
669 | /* | |
670 | * don't let the period tick interfere with the hrtick preemption | |
671 | */ | |
672 | if (!sched_feat(DOUBLE_TICK) && | |
673 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
674 | return; | |
675 | #endif | |
676 | ||
ce6c1311 | 677 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 678 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
679 | } |
680 | ||
681 | /************************************************** | |
682 | * CFS operations on tasks: | |
683 | */ | |
684 | ||
685 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
686 | ||
687 | /* Walk up scheduling entities hierarchy */ | |
688 | #define for_each_sched_entity(se) \ | |
689 | for (; se; se = se->parent) | |
690 | ||
691 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
692 | { | |
693 | return p->se.cfs_rq; | |
694 | } | |
695 | ||
696 | /* runqueue on which this entity is (to be) queued */ | |
697 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
698 | { | |
699 | return se->cfs_rq; | |
700 | } | |
701 | ||
702 | /* runqueue "owned" by this group */ | |
703 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
704 | { | |
705 | return grp->my_q; | |
706 | } | |
707 | ||
708 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
709 | * another cpu ('this_cpu') | |
710 | */ | |
711 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
712 | { | |
29f59db3 | 713 | return cfs_rq->tg->cfs_rq[this_cpu]; |
bf0f6f24 IM |
714 | } |
715 | ||
716 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
717 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
ec2c507f | 718 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
bf0f6f24 | 719 | |
fad095a7 SV |
720 | /* Do the two (enqueued) entities belong to the same group ? */ |
721 | static inline int | |
722 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 | 723 | { |
fad095a7 | 724 | if (se->cfs_rq == pse->cfs_rq) |
bf0f6f24 IM |
725 | return 1; |
726 | ||
727 | return 0; | |
728 | } | |
729 | ||
fad095a7 SV |
730 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
731 | { | |
732 | return se->parent; | |
733 | } | |
734 | ||
6b2d7700 SV |
735 | #define GROUP_IMBALANCE_PCT 20 |
736 | ||
bf0f6f24 IM |
737 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
738 | ||
739 | #define for_each_sched_entity(se) \ | |
740 | for (; se; se = NULL) | |
741 | ||
742 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
743 | { | |
744 | return &task_rq(p)->cfs; | |
745 | } | |
746 | ||
747 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
748 | { | |
749 | struct task_struct *p = task_of(se); | |
750 | struct rq *rq = task_rq(p); | |
751 | ||
752 | return &rq->cfs; | |
753 | } | |
754 | ||
755 | /* runqueue "owned" by this group */ | |
756 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
757 | { | |
758 | return NULL; | |
759 | } | |
760 | ||
761 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
762 | { | |
763 | return &cpu_rq(this_cpu)->cfs; | |
764 | } | |
765 | ||
766 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
767 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
768 | ||
fad095a7 SV |
769 | static inline int |
770 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 IM |
771 | { |
772 | return 1; | |
773 | } | |
774 | ||
fad095a7 SV |
775 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
776 | { | |
777 | return NULL; | |
778 | } | |
779 | ||
bf0f6f24 IM |
780 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
781 | ||
8f4d37ec PZ |
782 | #ifdef CONFIG_SCHED_HRTICK |
783 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
784 | { | |
785 | int requeue = rq->curr == p; | |
786 | struct sched_entity *se = &p->se; | |
787 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
788 | ||
789 | WARN_ON(task_rq(p) != rq); | |
790 | ||
791 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
792 | u64 slice = sched_slice(cfs_rq, se); | |
793 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
794 | s64 delta = slice - ran; | |
795 | ||
796 | if (delta < 0) { | |
797 | if (rq->curr == p) | |
798 | resched_task(p); | |
799 | return; | |
800 | } | |
801 | ||
802 | /* | |
803 | * Don't schedule slices shorter than 10000ns, that just | |
804 | * doesn't make sense. Rely on vruntime for fairness. | |
805 | */ | |
806 | if (!requeue) | |
807 | delta = max(10000LL, delta); | |
808 | ||
809 | hrtick_start(rq, delta, requeue); | |
810 | } | |
811 | } | |
812 | #else | |
813 | static inline void | |
814 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
815 | { | |
816 | } | |
817 | #endif | |
818 | ||
bf0f6f24 IM |
819 | /* |
820 | * The enqueue_task method is called before nr_running is | |
821 | * increased. Here we update the fair scheduling stats and | |
822 | * then put the task into the rbtree: | |
823 | */ | |
fd390f6a | 824 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
825 | { |
826 | struct cfs_rq *cfs_rq; | |
58e2d4ca SV |
827 | struct sched_entity *se = &p->se, |
828 | *topse = NULL; /* Highest schedulable entity */ | |
829 | int incload = 1; | |
bf0f6f24 IM |
830 | |
831 | for_each_sched_entity(se) { | |
58e2d4ca SV |
832 | topse = se; |
833 | if (se->on_rq) { | |
834 | incload = 0; | |
bf0f6f24 | 835 | break; |
58e2d4ca | 836 | } |
bf0f6f24 | 837 | cfs_rq = cfs_rq_of(se); |
83b699ed | 838 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 839 | wakeup = 1; |
bf0f6f24 | 840 | } |
58e2d4ca SV |
841 | /* Increment cpu load if we just enqueued the first task of a group on |
842 | * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs | |
843 | * at the highest grouping level. | |
844 | */ | |
845 | if (incload) | |
846 | inc_cpu_load(rq, topse->load.weight); | |
8f4d37ec PZ |
847 | |
848 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
849 | } |
850 | ||
851 | /* | |
852 | * The dequeue_task method is called before nr_running is | |
853 | * decreased. We remove the task from the rbtree and | |
854 | * update the fair scheduling stats: | |
855 | */ | |
f02231e5 | 856 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
857 | { |
858 | struct cfs_rq *cfs_rq; | |
58e2d4ca SV |
859 | struct sched_entity *se = &p->se, |
860 | *topse = NULL; /* Highest schedulable entity */ | |
861 | int decload = 1; | |
bf0f6f24 IM |
862 | |
863 | for_each_sched_entity(se) { | |
58e2d4ca | 864 | topse = se; |
bf0f6f24 | 865 | cfs_rq = cfs_rq_of(se); |
525c2716 | 866 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 867 | /* Don't dequeue parent if it has other entities besides us */ |
58e2d4ca SV |
868 | if (cfs_rq->load.weight) { |
869 | if (parent_entity(se)) | |
870 | decload = 0; | |
bf0f6f24 | 871 | break; |
58e2d4ca | 872 | } |
b9fa3df3 | 873 | sleep = 1; |
bf0f6f24 | 874 | } |
58e2d4ca SV |
875 | /* Decrement cpu load if we just dequeued the last task of a group on |
876 | * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs | |
877 | * at the highest grouping level. | |
878 | */ | |
879 | if (decload) | |
880 | dec_cpu_load(rq, topse->load.weight); | |
8f4d37ec PZ |
881 | |
882 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
883 | } |
884 | ||
885 | /* | |
1799e35d IM |
886 | * sched_yield() support is very simple - we dequeue and enqueue. |
887 | * | |
888 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 889 | */ |
4530d7ab | 890 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 891 | { |
db292ca3 IM |
892 | struct task_struct *curr = rq->curr; |
893 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
894 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
895 | |
896 | /* | |
1799e35d IM |
897 | * Are we the only task in the tree? |
898 | */ | |
899 | if (unlikely(cfs_rq->nr_running == 1)) | |
900 | return; | |
901 | ||
db292ca3 | 902 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
1799e35d IM |
903 | __update_rq_clock(rq); |
904 | /* | |
a2a2d680 | 905 | * Update run-time statistics of the 'current'. |
1799e35d | 906 | */ |
2b1e315d | 907 | update_curr(cfs_rq); |
1799e35d IM |
908 | |
909 | return; | |
910 | } | |
911 | /* | |
912 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 913 | */ |
2b1e315d | 914 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
915 | /* |
916 | * Already in the rightmost position? | |
917 | */ | |
2b1e315d | 918 | if (unlikely(rightmost->vruntime < se->vruntime)) |
1799e35d IM |
919 | return; |
920 | ||
921 | /* | |
922 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
923 | * Upon rescheduling, sched_class::put_prev_task() will place |
924 | * 'current' within the tree based on its new key value. | |
1799e35d | 925 | */ |
30cfdcfc | 926 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
927 | } |
928 | ||
e7693a36 GH |
929 | /* |
930 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
931 | * not idle and an idle cpu is available. The span of cpus to | |
932 | * search starts with cpus closest then further out as needed, | |
933 | * so we always favor a closer, idle cpu. | |
934 | * | |
935 | * Returns the CPU we should wake onto. | |
936 | */ | |
937 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
938 | static int wake_idle(int cpu, struct task_struct *p) | |
939 | { | |
940 | cpumask_t tmp; | |
941 | struct sched_domain *sd; | |
942 | int i; | |
943 | ||
944 | /* | |
945 | * If it is idle, then it is the best cpu to run this task. | |
946 | * | |
947 | * This cpu is also the best, if it has more than one task already. | |
948 | * Siblings must be also busy(in most cases) as they didn't already | |
949 | * pickup the extra load from this cpu and hence we need not check | |
950 | * sibling runqueue info. This will avoid the checks and cache miss | |
951 | * penalities associated with that. | |
952 | */ | |
953 | if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1) | |
954 | return cpu; | |
955 | ||
956 | for_each_domain(cpu, sd) { | |
957 | if (sd->flags & SD_WAKE_IDLE) { | |
958 | cpus_and(tmp, sd->span, p->cpus_allowed); | |
959 | for_each_cpu_mask(i, tmp) { | |
960 | if (idle_cpu(i)) { | |
961 | if (i != task_cpu(p)) { | |
962 | schedstat_inc(p, | |
963 | se.nr_wakeups_idle); | |
964 | } | |
965 | return i; | |
966 | } | |
967 | } | |
968 | } else { | |
969 | break; | |
970 | } | |
971 | } | |
972 | return cpu; | |
973 | } | |
974 | #else | |
975 | static inline int wake_idle(int cpu, struct task_struct *p) | |
976 | { | |
977 | return cpu; | |
978 | } | |
979 | #endif | |
980 | ||
981 | #ifdef CONFIG_SMP | |
982 | static int select_task_rq_fair(struct task_struct *p, int sync) | |
983 | { | |
984 | int cpu, this_cpu; | |
985 | struct rq *rq; | |
986 | struct sched_domain *sd, *this_sd = NULL; | |
987 | int new_cpu; | |
988 | ||
989 | cpu = task_cpu(p); | |
990 | rq = task_rq(p); | |
991 | this_cpu = smp_processor_id(); | |
992 | new_cpu = cpu; | |
993 | ||
9ec3b77e DA |
994 | if (cpu == this_cpu) |
995 | goto out_set_cpu; | |
996 | ||
e7693a36 GH |
997 | for_each_domain(this_cpu, sd) { |
998 | if (cpu_isset(cpu, sd->span)) { | |
999 | this_sd = sd; | |
1000 | break; | |
1001 | } | |
1002 | } | |
1003 | ||
1004 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) | |
1005 | goto out_set_cpu; | |
1006 | ||
1007 | /* | |
1008 | * Check for affine wakeup and passive balancing possibilities. | |
1009 | */ | |
1010 | if (this_sd) { | |
1011 | int idx = this_sd->wake_idx; | |
1012 | unsigned int imbalance; | |
1013 | unsigned long load, this_load; | |
1014 | ||
1015 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1016 | ||
1017 | load = source_load(cpu, idx); | |
1018 | this_load = target_load(this_cpu, idx); | |
1019 | ||
1020 | new_cpu = this_cpu; /* Wake to this CPU if we can */ | |
1021 | ||
1022 | if (this_sd->flags & SD_WAKE_AFFINE) { | |
1023 | unsigned long tl = this_load; | |
1024 | unsigned long tl_per_task; | |
1025 | ||
1026 | /* | |
1027 | * Attract cache-cold tasks on sync wakeups: | |
1028 | */ | |
1029 | if (sync && !task_hot(p, rq->clock, this_sd)) | |
1030 | goto out_set_cpu; | |
1031 | ||
1032 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1033 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1034 | ||
1035 | /* | |
1036 | * If sync wakeup then subtract the (maximum possible) | |
1037 | * effect of the currently running task from the load | |
1038 | * of the current CPU: | |
1039 | */ | |
1040 | if (sync) | |
1041 | tl -= current->se.load.weight; | |
1042 | ||
1043 | if ((tl <= load && | |
1044 | tl + target_load(cpu, idx) <= tl_per_task) || | |
1045 | 100*(tl + p->se.load.weight) <= imbalance*load) { | |
1046 | /* | |
1047 | * This domain has SD_WAKE_AFFINE and | |
1048 | * p is cache cold in this domain, and | |
1049 | * there is no bad imbalance. | |
1050 | */ | |
1051 | schedstat_inc(this_sd, ttwu_move_affine); | |
1052 | schedstat_inc(p, se.nr_wakeups_affine); | |
1053 | goto out_set_cpu; | |
1054 | } | |
1055 | } | |
1056 | ||
1057 | /* | |
1058 | * Start passive balancing when half the imbalance_pct | |
1059 | * limit is reached. | |
1060 | */ | |
1061 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1062 | if (imbalance*this_load <= 100*load) { | |
1063 | schedstat_inc(this_sd, ttwu_move_balance); | |
1064 | schedstat_inc(p, se.nr_wakeups_passive); | |
1065 | goto out_set_cpu; | |
1066 | } | |
1067 | } | |
1068 | } | |
1069 | ||
1070 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1071 | out_set_cpu: | |
1072 | return wake_idle(new_cpu, p); | |
1073 | } | |
1074 | #endif /* CONFIG_SMP */ | |
1075 | ||
1076 | ||
bf0f6f24 IM |
1077 | /* |
1078 | * Preempt the current task with a newly woken task if needed: | |
1079 | */ | |
2e09bf55 | 1080 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1081 | { |
1082 | struct task_struct *curr = rq->curr; | |
fad095a7 | 1083 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
8651a86c | 1084 | struct sched_entity *se = &curr->se, *pse = &p->se; |
502d26b5 | 1085 | unsigned long gran; |
bf0f6f24 IM |
1086 | |
1087 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 1088 | update_rq_clock(rq); |
b7cc0896 | 1089 | update_curr(cfs_rq); |
bf0f6f24 IM |
1090 | resched_task(curr); |
1091 | return; | |
1092 | } | |
91c234b4 IM |
1093 | /* |
1094 | * Batch tasks do not preempt (their preemption is driven by | |
1095 | * the tick): | |
1096 | */ | |
1097 | if (unlikely(p->policy == SCHED_BATCH)) | |
1098 | return; | |
bf0f6f24 | 1099 | |
77d9cc44 IM |
1100 | if (!sched_feat(WAKEUP_PREEMPT)) |
1101 | return; | |
8651a86c | 1102 | |
77d9cc44 IM |
1103 | while (!is_same_group(se, pse)) { |
1104 | se = parent_entity(se); | |
1105 | pse = parent_entity(pse); | |
ce6c1311 | 1106 | } |
77d9cc44 | 1107 | |
77d9cc44 IM |
1108 | gran = sysctl_sched_wakeup_granularity; |
1109 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1110 | gran = calc_delta_fair(gran, &se->load); | |
1111 | ||
502d26b5 | 1112 | if (pse->vruntime + gran < se->vruntime) |
77d9cc44 | 1113 | resched_task(curr); |
bf0f6f24 IM |
1114 | } |
1115 | ||
fb8d4724 | 1116 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1117 | { |
8f4d37ec | 1118 | struct task_struct *p; |
bf0f6f24 IM |
1119 | struct cfs_rq *cfs_rq = &rq->cfs; |
1120 | struct sched_entity *se; | |
1121 | ||
1122 | if (unlikely(!cfs_rq->nr_running)) | |
1123 | return NULL; | |
1124 | ||
1125 | do { | |
9948f4b2 | 1126 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1127 | cfs_rq = group_cfs_rq(se); |
1128 | } while (cfs_rq); | |
1129 | ||
8f4d37ec PZ |
1130 | p = task_of(se); |
1131 | hrtick_start_fair(rq, p); | |
1132 | ||
1133 | return p; | |
bf0f6f24 IM |
1134 | } |
1135 | ||
1136 | /* | |
1137 | * Account for a descheduled task: | |
1138 | */ | |
31ee529c | 1139 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1140 | { |
1141 | struct sched_entity *se = &prev->se; | |
1142 | struct cfs_rq *cfs_rq; | |
1143 | ||
1144 | for_each_sched_entity(se) { | |
1145 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1146 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1147 | } |
1148 | } | |
1149 | ||
681f3e68 | 1150 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1151 | /************************************************** |
1152 | * Fair scheduling class load-balancing methods: | |
1153 | */ | |
1154 | ||
1155 | /* | |
1156 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1157 | * during the whole iteration, the current task might be | |
1158 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1159 | * achieve that by always pre-iterating before returning | |
1160 | * the current task: | |
1161 | */ | |
a9957449 | 1162 | static struct task_struct * |
bf0f6f24 IM |
1163 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) |
1164 | { | |
1165 | struct task_struct *p; | |
1166 | ||
1167 | if (!curr) | |
1168 | return NULL; | |
1169 | ||
1170 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1171 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1172 | ||
1173 | return p; | |
1174 | } | |
1175 | ||
1176 | static struct task_struct *load_balance_start_fair(void *arg) | |
1177 | { | |
1178 | struct cfs_rq *cfs_rq = arg; | |
1179 | ||
1180 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1181 | } | |
1182 | ||
1183 | static struct task_struct *load_balance_next_fair(void *arg) | |
1184 | { | |
1185 | struct cfs_rq *cfs_rq = arg; | |
1186 | ||
1187 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1188 | } | |
1189 | ||
43010659 | 1190 | static unsigned long |
bf0f6f24 | 1191 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1192 | unsigned long max_load_move, |
a4ac01c3 PW |
1193 | struct sched_domain *sd, enum cpu_idle_type idle, |
1194 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1195 | { |
1196 | struct cfs_rq *busy_cfs_rq; | |
bf0f6f24 IM |
1197 | long rem_load_move = max_load_move; |
1198 | struct rq_iterator cfs_rq_iterator; | |
6b2d7700 | 1199 | unsigned long load_moved; |
bf0f6f24 IM |
1200 | |
1201 | cfs_rq_iterator.start = load_balance_start_fair; | |
1202 | cfs_rq_iterator.next = load_balance_next_fair; | |
1203 | ||
1204 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1205 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6b2d7700 SV |
1206 | struct cfs_rq *this_cfs_rq = busy_cfs_rq->tg->cfs_rq[this_cpu]; |
1207 | unsigned long maxload, task_load, group_weight; | |
1208 | unsigned long thisload, per_task_load; | |
1209 | struct sched_entity *se = busy_cfs_rq->tg->se[busiest->cpu]; | |
bf0f6f24 | 1210 | |
6b2d7700 SV |
1211 | task_load = busy_cfs_rq->load.weight; |
1212 | group_weight = se->load.weight; | |
bf0f6f24 | 1213 | |
6b2d7700 SV |
1214 | /* |
1215 | * 'group_weight' is contributed by tasks of total weight | |
1216 | * 'task_load'. To move 'rem_load_move' worth of weight only, | |
1217 | * we need to move a maximum task load of: | |
1218 | * | |
1219 | * maxload = (remload / group_weight) * task_load; | |
1220 | */ | |
1221 | maxload = (rem_load_move * task_load) / group_weight; | |
1222 | ||
1223 | if (!maxload || !task_load) | |
bf0f6f24 IM |
1224 | continue; |
1225 | ||
6b2d7700 SV |
1226 | per_task_load = task_load / busy_cfs_rq->nr_running; |
1227 | /* | |
1228 | * balance_tasks will try to forcibly move atleast one task if | |
1229 | * possible (because of SCHED_LOAD_SCALE_FUZZ). Avoid that if | |
1230 | * maxload is less than GROUP_IMBALANCE_FUZZ% the per_task_load. | |
1231 | */ | |
1232 | if (100 * maxload < GROUP_IMBALANCE_PCT * per_task_load) | |
1233 | continue; | |
bf0f6f24 | 1234 | |
6b2d7700 SV |
1235 | /* Disable priority-based load balance */ |
1236 | *this_best_prio = 0; | |
1237 | thisload = this_cfs_rq->load.weight; | |
a4ac01c3 | 1238 | #else |
e56f31aa | 1239 | # define maxload rem_load_move |
a4ac01c3 | 1240 | #endif |
e1d1484f PW |
1241 | /* |
1242 | * pass busy_cfs_rq argument into | |
bf0f6f24 IM |
1243 | * load_balance_[start|next]_fair iterators |
1244 | */ | |
1245 | cfs_rq_iterator.arg = busy_cfs_rq; | |
6b2d7700 | 1246 | load_moved = balance_tasks(this_rq, this_cpu, busiest, |
e1d1484f PW |
1247 | maxload, sd, idle, all_pinned, |
1248 | this_best_prio, | |
1249 | &cfs_rq_iterator); | |
bf0f6f24 | 1250 | |
6b2d7700 SV |
1251 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1252 | /* | |
1253 | * load_moved holds the task load that was moved. The | |
1254 | * effective (group) weight moved would be: | |
1255 | * load_moved_eff = load_moved/task_load * group_weight; | |
1256 | */ | |
1257 | load_moved = (group_weight * load_moved) / task_load; | |
1258 | ||
1259 | /* Adjust shares on both cpus to reflect load_moved */ | |
1260 | group_weight -= load_moved; | |
1261 | set_se_shares(se, group_weight); | |
1262 | ||
1263 | se = busy_cfs_rq->tg->se[this_cpu]; | |
1264 | if (!thisload) | |
1265 | group_weight = load_moved; | |
1266 | else | |
1267 | group_weight = se->load.weight + load_moved; | |
1268 | set_se_shares(se, group_weight); | |
1269 | #endif | |
1270 | ||
1271 | rem_load_move -= load_moved; | |
1272 | ||
e1d1484f | 1273 | if (rem_load_move <= 0) |
bf0f6f24 IM |
1274 | break; |
1275 | } | |
1276 | ||
43010659 | 1277 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1278 | } |
1279 | ||
e1d1484f PW |
1280 | static int |
1281 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1282 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1283 | { | |
1284 | struct cfs_rq *busy_cfs_rq; | |
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) { | |
1291 | /* | |
1292 | * pass busy_cfs_rq argument into | |
1293 | * load_balance_[start|next]_fair iterators | |
1294 | */ | |
1295 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1296 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1297 | &cfs_rq_iterator)) | |
1298 | return 1; | |
1299 | } | |
1300 | ||
1301 | return 0; | |
1302 | } | |
681f3e68 | 1303 | #endif |
e1d1484f | 1304 | |
bf0f6f24 IM |
1305 | /* |
1306 | * scheduler tick hitting a task of our scheduling class: | |
1307 | */ | |
8f4d37ec | 1308 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1309 | { |
1310 | struct cfs_rq *cfs_rq; | |
1311 | struct sched_entity *se = &curr->se; | |
1312 | ||
1313 | for_each_sched_entity(se) { | |
1314 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1315 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1316 | } |
1317 | } | |
1318 | ||
8eb172d9 | 1319 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
4d78e7b6 | 1320 | |
bf0f6f24 IM |
1321 | /* |
1322 | * Share the fairness runtime between parent and child, thus the | |
1323 | * total amount of pressure for CPU stays equal - new tasks | |
1324 | * get a chance to run but frequent forkers are not allowed to | |
1325 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1326 | * the child is not running yet. | |
1327 | */ | |
ee0827d8 | 1328 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1329 | { |
1330 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1331 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1332 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1333 | |
1334 | sched_info_queued(p); | |
1335 | ||
7109c442 | 1336 | update_curr(cfs_rq); |
aeb73b04 | 1337 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1338 | |
3c90e6e9 | 1339 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1340 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1341 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1342 | /* |
edcb60a3 IM |
1343 | * Upon rescheduling, sched_class::put_prev_task() will place |
1344 | * 'current' within the tree based on its new key value. | |
1345 | */ | |
4d78e7b6 | 1346 | swap(curr->vruntime, se->vruntime); |
4d78e7b6 | 1347 | } |
bf0f6f24 | 1348 | |
b9dca1e0 | 1349 | enqueue_task_fair(rq, p, 0); |
bb61c210 | 1350 | resched_task(rq->curr); |
bf0f6f24 IM |
1351 | } |
1352 | ||
cb469845 SR |
1353 | /* |
1354 | * Priority of the task has changed. Check to see if we preempt | |
1355 | * the current task. | |
1356 | */ | |
1357 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1358 | int oldprio, int running) | |
1359 | { | |
1360 | /* | |
1361 | * Reschedule if we are currently running on this runqueue and | |
1362 | * our priority decreased, or if we are not currently running on | |
1363 | * this runqueue and our priority is higher than the current's | |
1364 | */ | |
1365 | if (running) { | |
1366 | if (p->prio > oldprio) | |
1367 | resched_task(rq->curr); | |
1368 | } else | |
1369 | check_preempt_curr(rq, p); | |
1370 | } | |
1371 | ||
1372 | /* | |
1373 | * We switched to the sched_fair class. | |
1374 | */ | |
1375 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1376 | int running) | |
1377 | { | |
1378 | /* | |
1379 | * We were most likely switched from sched_rt, so | |
1380 | * kick off the schedule if running, otherwise just see | |
1381 | * if we can still preempt the current task. | |
1382 | */ | |
1383 | if (running) | |
1384 | resched_task(rq->curr); | |
1385 | else | |
1386 | check_preempt_curr(rq, p); | |
1387 | } | |
1388 | ||
83b699ed SV |
1389 | /* Account for a task changing its policy or group. |
1390 | * | |
1391 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1392 | * migrates between groups/classes. | |
1393 | */ | |
1394 | static void set_curr_task_fair(struct rq *rq) | |
1395 | { | |
1396 | struct sched_entity *se = &rq->curr->se; | |
1397 | ||
1398 | for_each_sched_entity(se) | |
1399 | set_next_entity(cfs_rq_of(se), se); | |
1400 | } | |
1401 | ||
bf0f6f24 IM |
1402 | /* |
1403 | * All the scheduling class methods: | |
1404 | */ | |
5522d5d5 IM |
1405 | static const struct sched_class fair_sched_class = { |
1406 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1407 | .enqueue_task = enqueue_task_fair, |
1408 | .dequeue_task = dequeue_task_fair, | |
1409 | .yield_task = yield_task_fair, | |
e7693a36 GH |
1410 | #ifdef CONFIG_SMP |
1411 | .select_task_rq = select_task_rq_fair, | |
1412 | #endif /* CONFIG_SMP */ | |
bf0f6f24 | 1413 | |
2e09bf55 | 1414 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1415 | |
1416 | .pick_next_task = pick_next_task_fair, | |
1417 | .put_prev_task = put_prev_task_fair, | |
1418 | ||
681f3e68 | 1419 | #ifdef CONFIG_SMP |
bf0f6f24 | 1420 | .load_balance = load_balance_fair, |
e1d1484f | 1421 | .move_one_task = move_one_task_fair, |
681f3e68 | 1422 | #endif |
bf0f6f24 | 1423 | |
83b699ed | 1424 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1425 | .task_tick = task_tick_fair, |
1426 | .task_new = task_new_fair, | |
cb469845 SR |
1427 | |
1428 | .prio_changed = prio_changed_fair, | |
1429 | .switched_to = switched_to_fair, | |
bf0f6f24 IM |
1430 | }; |
1431 | ||
1432 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1433 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1434 | { |
bf0f6f24 IM |
1435 | struct cfs_rq *cfs_rq; |
1436 | ||
75c28ace SV |
1437 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1438 | print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs); | |
1439 | #endif | |
5973e5b9 | 1440 | rcu_read_lock(); |
c3b64f1e | 1441 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1442 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1443 | rcu_read_unlock(); |
bf0f6f24 IM |
1444 | } |
1445 | #endif |