]>
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 | { | |
7eee3e67 | 205 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 206 | |
70eee74b BS |
207 | if (!last) |
208 | return NULL; | |
7eee3e67 IM |
209 | |
210 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
211 | } |
212 | ||
bf0f6f24 IM |
213 | /************************************************************** |
214 | * Scheduling class statistics methods: | |
215 | */ | |
216 | ||
b2be5e96 PZ |
217 | #ifdef CONFIG_SCHED_DEBUG |
218 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
219 | struct file *filp, void __user *buffer, size_t *lenp, | |
220 | loff_t *ppos) | |
221 | { | |
222 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
223 | ||
224 | if (ret || !write) | |
225 | return ret; | |
226 | ||
227 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
228 | sysctl_sched_min_granularity); | |
229 | ||
230 | return 0; | |
231 | } | |
232 | #endif | |
647e7cac IM |
233 | |
234 | /* | |
235 | * The idea is to set a period in which each task runs once. | |
236 | * | |
237 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
238 | * this period because otherwise the slices get too small. | |
239 | * | |
240 | * p = (nr <= nl) ? l : l*nr/nl | |
241 | */ | |
4d78e7b6 PZ |
242 | static u64 __sched_period(unsigned long nr_running) |
243 | { | |
244 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 245 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
246 | |
247 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 248 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 249 | period *= nr_running; |
4d78e7b6 PZ |
250 | } |
251 | ||
252 | return period; | |
253 | } | |
254 | ||
647e7cac IM |
255 | /* |
256 | * We calculate the wall-time slice from the period by taking a part | |
257 | * proportional to the weight. | |
258 | * | |
259 | * s = p*w/rw | |
260 | */ | |
6d0f0ebd | 261 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 262 | { |
647e7cac | 263 | u64 slice = __sched_period(cfs_rq->nr_running); |
21805085 | 264 | |
647e7cac IM |
265 | slice *= se->load.weight; |
266 | do_div(slice, cfs_rq->load.weight); | |
21805085 | 267 | |
647e7cac | 268 | return slice; |
bf0f6f24 IM |
269 | } |
270 | ||
647e7cac IM |
271 | /* |
272 | * We calculate the vruntime slice. | |
273 | * | |
274 | * vs = s/w = p/rw | |
275 | */ | |
276 | static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running) | |
67e9fb2a | 277 | { |
647e7cac | 278 | u64 vslice = __sched_period(nr_running); |
67e9fb2a | 279 | |
10b77724 | 280 | vslice *= NICE_0_LOAD; |
647e7cac | 281 | do_div(vslice, rq_weight); |
67e9fb2a | 282 | |
647e7cac IM |
283 | return vslice; |
284 | } | |
5f6d858e | 285 | |
647e7cac IM |
286 | static u64 sched_vslice(struct cfs_rq *cfs_rq) |
287 | { | |
288 | return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running); | |
289 | } | |
290 | ||
291 | static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
292 | { | |
293 | return __sched_vslice(cfs_rq->load.weight + se->load.weight, | |
294 | cfs_rq->nr_running + 1); | |
67e9fb2a PZ |
295 | } |
296 | ||
bf0f6f24 IM |
297 | /* |
298 | * Update the current task's runtime statistics. Skip current tasks that | |
299 | * are not in our scheduling class. | |
300 | */ | |
301 | static inline void | |
8ebc91d9 IM |
302 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
303 | unsigned long delta_exec) | |
bf0f6f24 | 304 | { |
bbdba7c0 | 305 | unsigned long delta_exec_weighted; |
b0ffd246 | 306 | u64 vruntime; |
bf0f6f24 | 307 | |
8179ca23 | 308 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
309 | |
310 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 311 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
e9acbff6 IM |
312 | delta_exec_weighted = delta_exec; |
313 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { | |
314 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, | |
315 | &curr->load); | |
316 | } | |
317 | curr->vruntime += delta_exec_weighted; | |
02e0431a PZ |
318 | |
319 | /* | |
320 | * maintain cfs_rq->min_vruntime to be a monotonic increasing | |
321 | * value tracking the leftmost vruntime in the tree. | |
322 | */ | |
323 | if (first_fair(cfs_rq)) { | |
b0ffd246 PZ |
324 | vruntime = min_vruntime(curr->vruntime, |
325 | __pick_next_entity(cfs_rq)->vruntime); | |
02e0431a | 326 | } else |
b0ffd246 | 327 | vruntime = curr->vruntime; |
02e0431a PZ |
328 | |
329 | cfs_rq->min_vruntime = | |
b0ffd246 | 330 | max_vruntime(cfs_rq->min_vruntime, vruntime); |
bf0f6f24 IM |
331 | } |
332 | ||
b7cc0896 | 333 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 334 | { |
429d43bc | 335 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 336 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
337 | unsigned long delta_exec; |
338 | ||
339 | if (unlikely(!curr)) | |
340 | return; | |
341 | ||
342 | /* | |
343 | * Get the amount of time the current task was running | |
344 | * since the last time we changed load (this cannot | |
345 | * overflow on 32 bits): | |
346 | */ | |
8ebc91d9 | 347 | delta_exec = (unsigned long)(now - curr->exec_start); |
bf0f6f24 | 348 | |
8ebc91d9 IM |
349 | __update_curr(cfs_rq, curr, delta_exec); |
350 | curr->exec_start = now; | |
d842de87 SV |
351 | |
352 | if (entity_is_task(curr)) { | |
353 | struct task_struct *curtask = task_of(curr); | |
354 | ||
355 | cpuacct_charge(curtask, delta_exec); | |
356 | } | |
bf0f6f24 IM |
357 | } |
358 | ||
359 | static inline void | |
5870db5b | 360 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 361 | { |
d281918d | 362 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
363 | } |
364 | ||
bf0f6f24 IM |
365 | /* |
366 | * Task is being enqueued - update stats: | |
367 | */ | |
d2417e5a | 368 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 369 | { |
bf0f6f24 IM |
370 | /* |
371 | * Are we enqueueing a waiting task? (for current tasks | |
372 | * a dequeue/enqueue event is a NOP) | |
373 | */ | |
429d43bc | 374 | if (se != cfs_rq->curr) |
5870db5b | 375 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
376 | } |
377 | ||
bf0f6f24 | 378 | static void |
9ef0a961 | 379 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 380 | { |
bbdba7c0 IM |
381 | schedstat_set(se->wait_max, max(se->wait_max, |
382 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
383 | schedstat_set(se->wait_count, se->wait_count + 1); |
384 | schedstat_set(se->wait_sum, se->wait_sum + | |
385 | rq_of(cfs_rq)->clock - se->wait_start); | |
6cfb0d5d | 386 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
387 | } |
388 | ||
389 | static inline void | |
19b6a2e3 | 390 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 391 | { |
bf0f6f24 IM |
392 | /* |
393 | * Mark the end of the wait period if dequeueing a | |
394 | * waiting task: | |
395 | */ | |
429d43bc | 396 | if (se != cfs_rq->curr) |
9ef0a961 | 397 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
398 | } |
399 | ||
400 | /* | |
401 | * We are picking a new current task - update its stats: | |
402 | */ | |
403 | static inline void | |
79303e9e | 404 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
405 | { |
406 | /* | |
407 | * We are starting a new run period: | |
408 | */ | |
d281918d | 409 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
410 | } |
411 | ||
bf0f6f24 IM |
412 | /************************************************** |
413 | * Scheduling class queueing methods: | |
414 | */ | |
415 | ||
30cfdcfc DA |
416 | static void |
417 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
418 | { | |
419 | update_load_add(&cfs_rq->load, se->load.weight); | |
420 | cfs_rq->nr_running++; | |
421 | se->on_rq = 1; | |
422 | } | |
423 | ||
424 | static void | |
425 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
426 | { | |
427 | update_load_sub(&cfs_rq->load, se->load.weight); | |
428 | cfs_rq->nr_running--; | |
429 | se->on_rq = 0; | |
430 | } | |
431 | ||
2396af69 | 432 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 433 | { |
bf0f6f24 IM |
434 | #ifdef CONFIG_SCHEDSTATS |
435 | if (se->sleep_start) { | |
d281918d | 436 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 437 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
438 | |
439 | if ((s64)delta < 0) | |
440 | delta = 0; | |
441 | ||
442 | if (unlikely(delta > se->sleep_max)) | |
443 | se->sleep_max = delta; | |
444 | ||
445 | se->sleep_start = 0; | |
446 | se->sum_sleep_runtime += delta; | |
9745512c AV |
447 | |
448 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
449 | } |
450 | if (se->block_start) { | |
d281918d | 451 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 452 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
453 | |
454 | if ((s64)delta < 0) | |
455 | delta = 0; | |
456 | ||
457 | if (unlikely(delta > se->block_max)) | |
458 | se->block_max = delta; | |
459 | ||
460 | se->block_start = 0; | |
461 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
462 | |
463 | /* | |
464 | * Blocking time is in units of nanosecs, so shift by 20 to | |
465 | * get a milliseconds-range estimation of the amount of | |
466 | * time that the task spent sleeping: | |
467 | */ | |
468 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 469 | |
30084fbd IM |
470 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
471 | delta >> 20); | |
472 | } | |
9745512c | 473 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
474 | } |
475 | #endif | |
476 | } | |
477 | ||
ddc97297 PZ |
478 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
479 | { | |
480 | #ifdef CONFIG_SCHED_DEBUG | |
481 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
482 | ||
483 | if (d < 0) | |
484 | d = -d; | |
485 | ||
486 | if (d > 3*sysctl_sched_latency) | |
487 | schedstat_inc(cfs_rq, nr_spread_over); | |
488 | #endif | |
489 | } | |
490 | ||
aeb73b04 PZ |
491 | static void |
492 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
493 | { | |
67e9fb2a | 494 | u64 vruntime; |
aeb73b04 | 495 | |
67e9fb2a | 496 | vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 497 | |
06877c33 | 498 | if (sched_feat(TREE_AVG)) { |
94dfb5e7 PZ |
499 | struct sched_entity *last = __pick_last_entity(cfs_rq); |
500 | if (last) { | |
67e9fb2a PZ |
501 | vruntime += last->vruntime; |
502 | vruntime >>= 1; | |
94dfb5e7 | 503 | } |
67e9fb2a | 504 | } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running) |
647e7cac | 505 | vruntime += sched_vslice(cfs_rq)/2; |
94dfb5e7 | 506 | |
2cb8600e PZ |
507 | /* |
508 | * The 'current' period is already promised to the current tasks, | |
509 | * however the extra weight of the new task will slow them down a | |
510 | * little, place the new task so that it fits in the slot that | |
511 | * stays open at the end. | |
512 | */ | |
94dfb5e7 | 513 | if (initial && sched_feat(START_DEBIT)) |
647e7cac | 514 | vruntime += sched_vslice_add(cfs_rq, se); |
aeb73b04 | 515 | |
8465e792 | 516 | if (!initial) { |
2cb8600e | 517 | /* sleeps upto a single latency don't count. */ |
296825cb | 518 | if (sched_feat(NEW_FAIR_SLEEPERS)) |
94359f05 IM |
519 | vruntime -= sysctl_sched_latency; |
520 | ||
2cb8600e PZ |
521 | /* ensure we never gain time by being placed backwards. */ |
522 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
523 | } |
524 | ||
67e9fb2a | 525 | se->vruntime = vruntime; |
aeb73b04 PZ |
526 | } |
527 | ||
bf0f6f24 | 528 | static void |
83b699ed | 529 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
530 | { |
531 | /* | |
a2a2d680 | 532 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 533 | */ |
b7cc0896 | 534 | update_curr(cfs_rq); |
bf0f6f24 | 535 | |
e9acbff6 | 536 | if (wakeup) { |
aeb73b04 | 537 | place_entity(cfs_rq, se, 0); |
2396af69 | 538 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 539 | } |
bf0f6f24 | 540 | |
d2417e5a | 541 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 542 | check_spread(cfs_rq, se); |
83b699ed SV |
543 | if (se != cfs_rq->curr) |
544 | __enqueue_entity(cfs_rq, se); | |
30cfdcfc | 545 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 IM |
546 | } |
547 | ||
548 | static void | |
525c2716 | 549 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 550 | { |
a2a2d680 DA |
551 | /* |
552 | * Update run-time statistics of the 'current'. | |
553 | */ | |
554 | update_curr(cfs_rq); | |
555 | ||
19b6a2e3 | 556 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 557 | if (sleep) { |
67e9fb2a | 558 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
559 | if (entity_is_task(se)) { |
560 | struct task_struct *tsk = task_of(se); | |
561 | ||
562 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 563 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 564 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 565 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 566 | } |
db36cc7d | 567 | #endif |
67e9fb2a PZ |
568 | } |
569 | ||
83b699ed | 570 | if (se != cfs_rq->curr) |
30cfdcfc DA |
571 | __dequeue_entity(cfs_rq, se); |
572 | account_entity_dequeue(cfs_rq, se); | |
bf0f6f24 IM |
573 | } |
574 | ||
575 | /* | |
576 | * Preempt the current task with a newly woken task if needed: | |
577 | */ | |
7c92e54f | 578 | static void |
2e09bf55 | 579 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 580 | { |
11697830 PZ |
581 | unsigned long ideal_runtime, delta_exec; |
582 | ||
6d0f0ebd | 583 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 584 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
3e3e13f3 | 585 | if (delta_exec > ideal_runtime) |
bf0f6f24 IM |
586 | resched_task(rq_of(cfs_rq)->curr); |
587 | } | |
588 | ||
83b699ed | 589 | static void |
8494f412 | 590 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 591 | { |
83b699ed SV |
592 | /* 'current' is not kept within the tree. */ |
593 | if (se->on_rq) { | |
594 | /* | |
595 | * Any task has to be enqueued before it get to execute on | |
596 | * a CPU. So account for the time it spent waiting on the | |
597 | * runqueue. | |
598 | */ | |
599 | update_stats_wait_end(cfs_rq, se); | |
600 | __dequeue_entity(cfs_rq, se); | |
601 | } | |
602 | ||
79303e9e | 603 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 604 | cfs_rq->curr = se; |
eba1ed4b IM |
605 | #ifdef CONFIG_SCHEDSTATS |
606 | /* | |
607 | * Track our maximum slice length, if the CPU's load is at | |
608 | * least twice that of our own weight (i.e. dont track it | |
609 | * when there are only lesser-weight tasks around): | |
610 | */ | |
495eca49 | 611 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
612 | se->slice_max = max(se->slice_max, |
613 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
614 | } | |
615 | #endif | |
4a55b450 | 616 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
617 | } |
618 | ||
9948f4b2 | 619 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 620 | { |
08ec3df5 | 621 | struct sched_entity *se = NULL; |
bf0f6f24 | 622 | |
08ec3df5 DA |
623 | if (first_fair(cfs_rq)) { |
624 | se = __pick_next_entity(cfs_rq); | |
625 | set_next_entity(cfs_rq, se); | |
626 | } | |
bf0f6f24 IM |
627 | |
628 | return se; | |
629 | } | |
630 | ||
ab6cde26 | 631 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
632 | { |
633 | /* | |
634 | * If still on the runqueue then deactivate_task() | |
635 | * was not called and update_curr() has to be done: | |
636 | */ | |
637 | if (prev->on_rq) | |
b7cc0896 | 638 | update_curr(cfs_rq); |
bf0f6f24 | 639 | |
ddc97297 | 640 | check_spread(cfs_rq, prev); |
30cfdcfc | 641 | if (prev->on_rq) { |
5870db5b | 642 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
643 | /* Put 'current' back into the tree. */ |
644 | __enqueue_entity(cfs_rq, prev); | |
645 | } | |
429d43bc | 646 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
647 | } |
648 | ||
8f4d37ec PZ |
649 | static void |
650 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 651 | { |
bf0f6f24 | 652 | /* |
30cfdcfc | 653 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 654 | */ |
30cfdcfc | 655 | update_curr(cfs_rq); |
bf0f6f24 | 656 | |
8f4d37ec PZ |
657 | #ifdef CONFIG_SCHED_HRTICK |
658 | /* | |
659 | * queued ticks are scheduled to match the slice, so don't bother | |
660 | * validating it and just reschedule. | |
661 | */ | |
662 | if (queued) | |
663 | return resched_task(rq_of(cfs_rq)->curr); | |
664 | /* | |
665 | * don't let the period tick interfere with the hrtick preemption | |
666 | */ | |
667 | if (!sched_feat(DOUBLE_TICK) && | |
668 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
669 | return; | |
670 | #endif | |
671 | ||
ce6c1311 | 672 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 673 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
674 | } |
675 | ||
676 | /************************************************** | |
677 | * CFS operations on tasks: | |
678 | */ | |
679 | ||
680 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
681 | ||
682 | /* Walk up scheduling entities hierarchy */ | |
683 | #define for_each_sched_entity(se) \ | |
684 | for (; se; se = se->parent) | |
685 | ||
686 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
687 | { | |
688 | return p->se.cfs_rq; | |
689 | } | |
690 | ||
691 | /* runqueue on which this entity is (to be) queued */ | |
692 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
693 | { | |
694 | return se->cfs_rq; | |
695 | } | |
696 | ||
697 | /* runqueue "owned" by this group */ | |
698 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
699 | { | |
700 | return grp->my_q; | |
701 | } | |
702 | ||
703 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
704 | * another cpu ('this_cpu') | |
705 | */ | |
706 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
707 | { | |
29f59db3 | 708 | return cfs_rq->tg->cfs_rq[this_cpu]; |
bf0f6f24 IM |
709 | } |
710 | ||
711 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
712 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
ec2c507f | 713 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
bf0f6f24 | 714 | |
fad095a7 SV |
715 | /* Do the two (enqueued) entities belong to the same group ? */ |
716 | static inline int | |
717 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 | 718 | { |
fad095a7 | 719 | if (se->cfs_rq == pse->cfs_rq) |
bf0f6f24 IM |
720 | return 1; |
721 | ||
722 | return 0; | |
723 | } | |
724 | ||
fad095a7 SV |
725 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
726 | { | |
727 | return se->parent; | |
728 | } | |
729 | ||
6b2d7700 SV |
730 | #define GROUP_IMBALANCE_PCT 20 |
731 | ||
bf0f6f24 IM |
732 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
733 | ||
734 | #define for_each_sched_entity(se) \ | |
735 | for (; se; se = NULL) | |
736 | ||
737 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
738 | { | |
739 | return &task_rq(p)->cfs; | |
740 | } | |
741 | ||
742 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
743 | { | |
744 | struct task_struct *p = task_of(se); | |
745 | struct rq *rq = task_rq(p); | |
746 | ||
747 | return &rq->cfs; | |
748 | } | |
749 | ||
750 | /* runqueue "owned" by this group */ | |
751 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
752 | { | |
753 | return NULL; | |
754 | } | |
755 | ||
756 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
757 | { | |
758 | return &cpu_rq(this_cpu)->cfs; | |
759 | } | |
760 | ||
761 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
762 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
763 | ||
fad095a7 SV |
764 | static inline int |
765 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 IM |
766 | { |
767 | return 1; | |
768 | } | |
769 | ||
fad095a7 SV |
770 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
771 | { | |
772 | return NULL; | |
773 | } | |
774 | ||
bf0f6f24 IM |
775 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
776 | ||
8f4d37ec PZ |
777 | #ifdef CONFIG_SCHED_HRTICK |
778 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
779 | { | |
780 | int requeue = rq->curr == p; | |
781 | struct sched_entity *se = &p->se; | |
782 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
783 | ||
784 | WARN_ON(task_rq(p) != rq); | |
785 | ||
786 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
787 | u64 slice = sched_slice(cfs_rq, se); | |
788 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
789 | s64 delta = slice - ran; | |
790 | ||
791 | if (delta < 0) { | |
792 | if (rq->curr == p) | |
793 | resched_task(p); | |
794 | return; | |
795 | } | |
796 | ||
797 | /* | |
798 | * Don't schedule slices shorter than 10000ns, that just | |
799 | * doesn't make sense. Rely on vruntime for fairness. | |
800 | */ | |
801 | if (!requeue) | |
802 | delta = max(10000LL, delta); | |
803 | ||
804 | hrtick_start(rq, delta, requeue); | |
805 | } | |
806 | } | |
807 | #else | |
808 | static inline void | |
809 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
810 | { | |
811 | } | |
812 | #endif | |
813 | ||
bf0f6f24 IM |
814 | /* |
815 | * The enqueue_task method is called before nr_running is | |
816 | * increased. Here we update the fair scheduling stats and | |
817 | * then put the task into the rbtree: | |
818 | */ | |
fd390f6a | 819 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
820 | { |
821 | struct cfs_rq *cfs_rq; | |
58e2d4ca SV |
822 | struct sched_entity *se = &p->se, |
823 | *topse = NULL; /* Highest schedulable entity */ | |
824 | int incload = 1; | |
bf0f6f24 IM |
825 | |
826 | for_each_sched_entity(se) { | |
58e2d4ca SV |
827 | topse = se; |
828 | if (se->on_rq) { | |
829 | incload = 0; | |
bf0f6f24 | 830 | break; |
58e2d4ca | 831 | } |
bf0f6f24 | 832 | cfs_rq = cfs_rq_of(se); |
83b699ed | 833 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 834 | wakeup = 1; |
bf0f6f24 | 835 | } |
58e2d4ca SV |
836 | /* Increment cpu load if we just enqueued the first task of a group on |
837 | * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs | |
838 | * at the highest grouping level. | |
839 | */ | |
840 | if (incload) | |
841 | inc_cpu_load(rq, topse->load.weight); | |
8f4d37ec PZ |
842 | |
843 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
844 | } |
845 | ||
846 | /* | |
847 | * The dequeue_task method is called before nr_running is | |
848 | * decreased. We remove the task from the rbtree and | |
849 | * update the fair scheduling stats: | |
850 | */ | |
f02231e5 | 851 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
852 | { |
853 | struct cfs_rq *cfs_rq; | |
58e2d4ca SV |
854 | struct sched_entity *se = &p->se, |
855 | *topse = NULL; /* Highest schedulable entity */ | |
856 | int decload = 1; | |
bf0f6f24 IM |
857 | |
858 | for_each_sched_entity(se) { | |
58e2d4ca | 859 | topse = se; |
bf0f6f24 | 860 | cfs_rq = cfs_rq_of(se); |
525c2716 | 861 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 862 | /* Don't dequeue parent if it has other entities besides us */ |
58e2d4ca SV |
863 | if (cfs_rq->load.weight) { |
864 | if (parent_entity(se)) | |
865 | decload = 0; | |
bf0f6f24 | 866 | break; |
58e2d4ca | 867 | } |
b9fa3df3 | 868 | sleep = 1; |
bf0f6f24 | 869 | } |
58e2d4ca SV |
870 | /* Decrement cpu load if we just dequeued the last task of a group on |
871 | * 'rq->cpu'. 'topse' represents the group to which task 'p' belongs | |
872 | * at the highest grouping level. | |
873 | */ | |
874 | if (decload) | |
875 | dec_cpu_load(rq, topse->load.weight); | |
8f4d37ec PZ |
876 | |
877 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
878 | } |
879 | ||
880 | /* | |
1799e35d IM |
881 | * sched_yield() support is very simple - we dequeue and enqueue. |
882 | * | |
883 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 884 | */ |
4530d7ab | 885 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 886 | { |
db292ca3 IM |
887 | struct task_struct *curr = rq->curr; |
888 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
889 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
890 | |
891 | /* | |
1799e35d IM |
892 | * Are we the only task in the tree? |
893 | */ | |
894 | if (unlikely(cfs_rq->nr_running == 1)) | |
895 | return; | |
896 | ||
db292ca3 | 897 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
1799e35d IM |
898 | __update_rq_clock(rq); |
899 | /* | |
a2a2d680 | 900 | * Update run-time statistics of the 'current'. |
1799e35d | 901 | */ |
2b1e315d | 902 | update_curr(cfs_rq); |
1799e35d IM |
903 | |
904 | return; | |
905 | } | |
906 | /* | |
907 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 908 | */ |
2b1e315d | 909 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
910 | /* |
911 | * Already in the rightmost position? | |
912 | */ | |
2b1e315d | 913 | if (unlikely(rightmost->vruntime < se->vruntime)) |
1799e35d IM |
914 | return; |
915 | ||
916 | /* | |
917 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
918 | * Upon rescheduling, sched_class::put_prev_task() will place |
919 | * 'current' within the tree based on its new key value. | |
1799e35d | 920 | */ |
30cfdcfc | 921 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
922 | } |
923 | ||
e7693a36 GH |
924 | /* |
925 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
926 | * not idle and an idle cpu is available. The span of cpus to | |
927 | * search starts with cpus closest then further out as needed, | |
928 | * so we always favor a closer, idle cpu. | |
929 | * | |
930 | * Returns the CPU we should wake onto. | |
931 | */ | |
932 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
933 | static int wake_idle(int cpu, struct task_struct *p) | |
934 | { | |
935 | cpumask_t tmp; | |
936 | struct sched_domain *sd; | |
937 | int i; | |
938 | ||
939 | /* | |
940 | * If it is idle, then it is the best cpu to run this task. | |
941 | * | |
942 | * This cpu is also the best, if it has more than one task already. | |
943 | * Siblings must be also busy(in most cases) as they didn't already | |
944 | * pickup the extra load from this cpu and hence we need not check | |
945 | * sibling runqueue info. This will avoid the checks and cache miss | |
946 | * penalities associated with that. | |
947 | */ | |
948 | if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1) | |
949 | return cpu; | |
950 | ||
951 | for_each_domain(cpu, sd) { | |
952 | if (sd->flags & SD_WAKE_IDLE) { | |
953 | cpus_and(tmp, sd->span, p->cpus_allowed); | |
954 | for_each_cpu_mask(i, tmp) { | |
955 | if (idle_cpu(i)) { | |
956 | if (i != task_cpu(p)) { | |
957 | schedstat_inc(p, | |
958 | se.nr_wakeups_idle); | |
959 | } | |
960 | return i; | |
961 | } | |
962 | } | |
963 | } else { | |
964 | break; | |
965 | } | |
966 | } | |
967 | return cpu; | |
968 | } | |
969 | #else | |
970 | static inline int wake_idle(int cpu, struct task_struct *p) | |
971 | { | |
972 | return cpu; | |
973 | } | |
974 | #endif | |
975 | ||
976 | #ifdef CONFIG_SMP | |
977 | static int select_task_rq_fair(struct task_struct *p, int sync) | |
978 | { | |
979 | int cpu, this_cpu; | |
980 | struct rq *rq; | |
981 | struct sched_domain *sd, *this_sd = NULL; | |
982 | int new_cpu; | |
983 | ||
984 | cpu = task_cpu(p); | |
985 | rq = task_rq(p); | |
986 | this_cpu = smp_processor_id(); | |
987 | new_cpu = cpu; | |
988 | ||
9ec3b77e DA |
989 | if (cpu == this_cpu) |
990 | goto out_set_cpu; | |
991 | ||
e7693a36 GH |
992 | for_each_domain(this_cpu, sd) { |
993 | if (cpu_isset(cpu, sd->span)) { | |
994 | this_sd = sd; | |
995 | break; | |
996 | } | |
997 | } | |
998 | ||
999 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) | |
1000 | goto out_set_cpu; | |
1001 | ||
1002 | /* | |
1003 | * Check for affine wakeup and passive balancing possibilities. | |
1004 | */ | |
1005 | if (this_sd) { | |
1006 | int idx = this_sd->wake_idx; | |
1007 | unsigned int imbalance; | |
1008 | unsigned long load, this_load; | |
1009 | ||
1010 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1011 | ||
1012 | load = source_load(cpu, idx); | |
1013 | this_load = target_load(this_cpu, idx); | |
1014 | ||
1015 | new_cpu = this_cpu; /* Wake to this CPU if we can */ | |
1016 | ||
1017 | if (this_sd->flags & SD_WAKE_AFFINE) { | |
1018 | unsigned long tl = this_load; | |
1019 | unsigned long tl_per_task; | |
1020 | ||
1021 | /* | |
1022 | * Attract cache-cold tasks on sync wakeups: | |
1023 | */ | |
1024 | if (sync && !task_hot(p, rq->clock, this_sd)) | |
1025 | goto out_set_cpu; | |
1026 | ||
1027 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1028 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1029 | ||
1030 | /* | |
1031 | * If sync wakeup then subtract the (maximum possible) | |
1032 | * effect of the currently running task from the load | |
1033 | * of the current CPU: | |
1034 | */ | |
1035 | if (sync) | |
1036 | tl -= current->se.load.weight; | |
1037 | ||
1038 | if ((tl <= load && | |
1039 | tl + target_load(cpu, idx) <= tl_per_task) || | |
1040 | 100*(tl + p->se.load.weight) <= imbalance*load) { | |
1041 | /* | |
1042 | * This domain has SD_WAKE_AFFINE and | |
1043 | * p is cache cold in this domain, and | |
1044 | * there is no bad imbalance. | |
1045 | */ | |
1046 | schedstat_inc(this_sd, ttwu_move_affine); | |
1047 | schedstat_inc(p, se.nr_wakeups_affine); | |
1048 | goto out_set_cpu; | |
1049 | } | |
1050 | } | |
1051 | ||
1052 | /* | |
1053 | * Start passive balancing when half the imbalance_pct | |
1054 | * limit is reached. | |
1055 | */ | |
1056 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1057 | if (imbalance*this_load <= 100*load) { | |
1058 | schedstat_inc(this_sd, ttwu_move_balance); | |
1059 | schedstat_inc(p, se.nr_wakeups_passive); | |
1060 | goto out_set_cpu; | |
1061 | } | |
1062 | } | |
1063 | } | |
1064 | ||
1065 | new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */ | |
1066 | out_set_cpu: | |
1067 | return wake_idle(new_cpu, p); | |
1068 | } | |
1069 | #endif /* CONFIG_SMP */ | |
1070 | ||
1071 | ||
bf0f6f24 IM |
1072 | /* |
1073 | * Preempt the current task with a newly woken task if needed: | |
1074 | */ | |
2e09bf55 | 1075 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1076 | { |
1077 | struct task_struct *curr = rq->curr; | |
fad095a7 | 1078 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
8651a86c | 1079 | struct sched_entity *se = &curr->se, *pse = &p->se; |
502d26b5 | 1080 | unsigned long gran; |
bf0f6f24 IM |
1081 | |
1082 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 1083 | update_rq_clock(rq); |
b7cc0896 | 1084 | update_curr(cfs_rq); |
bf0f6f24 IM |
1085 | resched_task(curr); |
1086 | return; | |
1087 | } | |
91c234b4 IM |
1088 | /* |
1089 | * Batch tasks do not preempt (their preemption is driven by | |
1090 | * the tick): | |
1091 | */ | |
1092 | if (unlikely(p->policy == SCHED_BATCH)) | |
1093 | return; | |
bf0f6f24 | 1094 | |
77d9cc44 IM |
1095 | if (!sched_feat(WAKEUP_PREEMPT)) |
1096 | return; | |
8651a86c | 1097 | |
77d9cc44 IM |
1098 | while (!is_same_group(se, pse)) { |
1099 | se = parent_entity(se); | |
1100 | pse = parent_entity(pse); | |
ce6c1311 | 1101 | } |
77d9cc44 | 1102 | |
77d9cc44 | 1103 | gran = sysctl_sched_wakeup_granularity; |
ef9884e6 PZ |
1104 | /* |
1105 | * More easily preempt - nice tasks, while not making | |
1106 | * it harder for + nice tasks. | |
1107 | */ | |
1108 | if (unlikely(se->load.weight > NICE_0_LOAD)) | |
77d9cc44 IM |
1109 | gran = calc_delta_fair(gran, &se->load); |
1110 | ||
502d26b5 | 1111 | if (pse->vruntime + gran < se->vruntime) |
77d9cc44 | 1112 | resched_task(curr); |
bf0f6f24 IM |
1113 | } |
1114 | ||
fb8d4724 | 1115 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1116 | { |
8f4d37ec | 1117 | struct task_struct *p; |
bf0f6f24 IM |
1118 | struct cfs_rq *cfs_rq = &rq->cfs; |
1119 | struct sched_entity *se; | |
1120 | ||
1121 | if (unlikely(!cfs_rq->nr_running)) | |
1122 | return NULL; | |
1123 | ||
1124 | do { | |
9948f4b2 | 1125 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1126 | cfs_rq = group_cfs_rq(se); |
1127 | } while (cfs_rq); | |
1128 | ||
8f4d37ec PZ |
1129 | p = task_of(se); |
1130 | hrtick_start_fair(rq, p); | |
1131 | ||
1132 | return p; | |
bf0f6f24 IM |
1133 | } |
1134 | ||
1135 | /* | |
1136 | * Account for a descheduled task: | |
1137 | */ | |
31ee529c | 1138 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1139 | { |
1140 | struct sched_entity *se = &prev->se; | |
1141 | struct cfs_rq *cfs_rq; | |
1142 | ||
1143 | for_each_sched_entity(se) { | |
1144 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1145 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1146 | } |
1147 | } | |
1148 | ||
681f3e68 | 1149 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1150 | /************************************************** |
1151 | * Fair scheduling class load-balancing methods: | |
1152 | */ | |
1153 | ||
1154 | /* | |
1155 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1156 | * during the whole iteration, the current task might be | |
1157 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1158 | * achieve that by always pre-iterating before returning | |
1159 | * the current task: | |
1160 | */ | |
a9957449 | 1161 | static struct task_struct * |
bf0f6f24 IM |
1162 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) |
1163 | { | |
1164 | struct task_struct *p; | |
1165 | ||
1166 | if (!curr) | |
1167 | return NULL; | |
1168 | ||
1169 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1170 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1171 | ||
1172 | return p; | |
1173 | } | |
1174 | ||
1175 | static struct task_struct *load_balance_start_fair(void *arg) | |
1176 | { | |
1177 | struct cfs_rq *cfs_rq = arg; | |
1178 | ||
1179 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1180 | } | |
1181 | ||
1182 | static struct task_struct *load_balance_next_fair(void *arg) | |
1183 | { | |
1184 | struct cfs_rq *cfs_rq = arg; | |
1185 | ||
1186 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1187 | } | |
1188 | ||
43010659 | 1189 | static unsigned long |
bf0f6f24 | 1190 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1191 | unsigned long max_load_move, |
a4ac01c3 PW |
1192 | struct sched_domain *sd, enum cpu_idle_type idle, |
1193 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1194 | { |
1195 | struct cfs_rq *busy_cfs_rq; | |
bf0f6f24 IM |
1196 | long rem_load_move = max_load_move; |
1197 | struct rq_iterator cfs_rq_iterator; | |
6b2d7700 | 1198 | unsigned long load_moved; |
bf0f6f24 IM |
1199 | |
1200 | cfs_rq_iterator.start = load_balance_start_fair; | |
1201 | cfs_rq_iterator.next = load_balance_next_fair; | |
1202 | ||
1203 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1204 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6b2d7700 SV |
1205 | struct cfs_rq *this_cfs_rq = busy_cfs_rq->tg->cfs_rq[this_cpu]; |
1206 | unsigned long maxload, task_load, group_weight; | |
1207 | unsigned long thisload, per_task_load; | |
1208 | struct sched_entity *se = busy_cfs_rq->tg->se[busiest->cpu]; | |
bf0f6f24 | 1209 | |
6b2d7700 SV |
1210 | task_load = busy_cfs_rq->load.weight; |
1211 | group_weight = se->load.weight; | |
bf0f6f24 | 1212 | |
6b2d7700 SV |
1213 | /* |
1214 | * 'group_weight' is contributed by tasks of total weight | |
1215 | * 'task_load'. To move 'rem_load_move' worth of weight only, | |
1216 | * we need to move a maximum task load of: | |
1217 | * | |
1218 | * maxload = (remload / group_weight) * task_load; | |
1219 | */ | |
1220 | maxload = (rem_load_move * task_load) / group_weight; | |
1221 | ||
1222 | if (!maxload || !task_load) | |
bf0f6f24 IM |
1223 | continue; |
1224 | ||
6b2d7700 SV |
1225 | per_task_load = task_load / busy_cfs_rq->nr_running; |
1226 | /* | |
1227 | * balance_tasks will try to forcibly move atleast one task if | |
1228 | * possible (because of SCHED_LOAD_SCALE_FUZZ). Avoid that if | |
1229 | * maxload is less than GROUP_IMBALANCE_FUZZ% the per_task_load. | |
1230 | */ | |
1231 | if (100 * maxload < GROUP_IMBALANCE_PCT * per_task_load) | |
1232 | continue; | |
bf0f6f24 | 1233 | |
6b2d7700 SV |
1234 | /* Disable priority-based load balance */ |
1235 | *this_best_prio = 0; | |
1236 | thisload = this_cfs_rq->load.weight; | |
a4ac01c3 | 1237 | #else |
e56f31aa | 1238 | # define maxload rem_load_move |
a4ac01c3 | 1239 | #endif |
e1d1484f PW |
1240 | /* |
1241 | * pass busy_cfs_rq argument into | |
bf0f6f24 IM |
1242 | * load_balance_[start|next]_fair iterators |
1243 | */ | |
1244 | cfs_rq_iterator.arg = busy_cfs_rq; | |
6b2d7700 | 1245 | load_moved = balance_tasks(this_rq, this_cpu, busiest, |
e1d1484f PW |
1246 | maxload, sd, idle, all_pinned, |
1247 | this_best_prio, | |
1248 | &cfs_rq_iterator); | |
bf0f6f24 | 1249 | |
6b2d7700 SV |
1250 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1251 | /* | |
1252 | * load_moved holds the task load that was moved. The | |
1253 | * effective (group) weight moved would be: | |
1254 | * load_moved_eff = load_moved/task_load * group_weight; | |
1255 | */ | |
1256 | load_moved = (group_weight * load_moved) / task_load; | |
1257 | ||
1258 | /* Adjust shares on both cpus to reflect load_moved */ | |
1259 | group_weight -= load_moved; | |
1260 | set_se_shares(se, group_weight); | |
1261 | ||
1262 | se = busy_cfs_rq->tg->se[this_cpu]; | |
1263 | if (!thisload) | |
1264 | group_weight = load_moved; | |
1265 | else | |
1266 | group_weight = se->load.weight + load_moved; | |
1267 | set_se_shares(se, group_weight); | |
1268 | #endif | |
1269 | ||
1270 | rem_load_move -= load_moved; | |
1271 | ||
e1d1484f | 1272 | if (rem_load_move <= 0) |
bf0f6f24 IM |
1273 | break; |
1274 | } | |
1275 | ||
43010659 | 1276 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1277 | } |
1278 | ||
e1d1484f PW |
1279 | static int |
1280 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1281 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1282 | { | |
1283 | struct cfs_rq *busy_cfs_rq; | |
1284 | struct rq_iterator cfs_rq_iterator; | |
1285 | ||
1286 | cfs_rq_iterator.start = load_balance_start_fair; | |
1287 | cfs_rq_iterator.next = load_balance_next_fair; | |
1288 | ||
1289 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1290 | /* | |
1291 | * pass busy_cfs_rq argument into | |
1292 | * load_balance_[start|next]_fair iterators | |
1293 | */ | |
1294 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1295 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1296 | &cfs_rq_iterator)) | |
1297 | return 1; | |
1298 | } | |
1299 | ||
1300 | return 0; | |
1301 | } | |
681f3e68 | 1302 | #endif |
e1d1484f | 1303 | |
bf0f6f24 IM |
1304 | /* |
1305 | * scheduler tick hitting a task of our scheduling class: | |
1306 | */ | |
8f4d37ec | 1307 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1308 | { |
1309 | struct cfs_rq *cfs_rq; | |
1310 | struct sched_entity *se = &curr->se; | |
1311 | ||
1312 | for_each_sched_entity(se) { | |
1313 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1314 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1315 | } |
1316 | } | |
1317 | ||
8eb172d9 | 1318 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
4d78e7b6 | 1319 | |
bf0f6f24 IM |
1320 | /* |
1321 | * Share the fairness runtime between parent and child, thus the | |
1322 | * total amount of pressure for CPU stays equal - new tasks | |
1323 | * get a chance to run but frequent forkers are not allowed to | |
1324 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1325 | * the child is not running yet. | |
1326 | */ | |
ee0827d8 | 1327 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1328 | { |
1329 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1330 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1331 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1332 | |
1333 | sched_info_queued(p); | |
1334 | ||
7109c442 | 1335 | update_curr(cfs_rq); |
aeb73b04 | 1336 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1337 | |
3c90e6e9 | 1338 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1339 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1340 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1341 | /* |
edcb60a3 IM |
1342 | * Upon rescheduling, sched_class::put_prev_task() will place |
1343 | * 'current' within the tree based on its new key value. | |
1344 | */ | |
4d78e7b6 | 1345 | swap(curr->vruntime, se->vruntime); |
4d78e7b6 | 1346 | } |
bf0f6f24 | 1347 | |
b9dca1e0 | 1348 | enqueue_task_fair(rq, p, 0); |
bb61c210 | 1349 | resched_task(rq->curr); |
bf0f6f24 IM |
1350 | } |
1351 | ||
cb469845 SR |
1352 | /* |
1353 | * Priority of the task has changed. Check to see if we preempt | |
1354 | * the current task. | |
1355 | */ | |
1356 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1357 | int oldprio, int running) | |
1358 | { | |
1359 | /* | |
1360 | * Reschedule if we are currently running on this runqueue and | |
1361 | * our priority decreased, or if we are not currently running on | |
1362 | * this runqueue and our priority is higher than the current's | |
1363 | */ | |
1364 | if (running) { | |
1365 | if (p->prio > oldprio) | |
1366 | resched_task(rq->curr); | |
1367 | } else | |
1368 | check_preempt_curr(rq, p); | |
1369 | } | |
1370 | ||
1371 | /* | |
1372 | * We switched to the sched_fair class. | |
1373 | */ | |
1374 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1375 | int running) | |
1376 | { | |
1377 | /* | |
1378 | * We were most likely switched from sched_rt, so | |
1379 | * kick off the schedule if running, otherwise just see | |
1380 | * if we can still preempt the current task. | |
1381 | */ | |
1382 | if (running) | |
1383 | resched_task(rq->curr); | |
1384 | else | |
1385 | check_preempt_curr(rq, p); | |
1386 | } | |
1387 | ||
83b699ed SV |
1388 | /* Account for a task changing its policy or group. |
1389 | * | |
1390 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1391 | * migrates between groups/classes. | |
1392 | */ | |
1393 | static void set_curr_task_fair(struct rq *rq) | |
1394 | { | |
1395 | struct sched_entity *se = &rq->curr->se; | |
1396 | ||
1397 | for_each_sched_entity(se) | |
1398 | set_next_entity(cfs_rq_of(se), se); | |
1399 | } | |
1400 | ||
bf0f6f24 IM |
1401 | /* |
1402 | * All the scheduling class methods: | |
1403 | */ | |
5522d5d5 IM |
1404 | static const struct sched_class fair_sched_class = { |
1405 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1406 | .enqueue_task = enqueue_task_fair, |
1407 | .dequeue_task = dequeue_task_fair, | |
1408 | .yield_task = yield_task_fair, | |
e7693a36 GH |
1409 | #ifdef CONFIG_SMP |
1410 | .select_task_rq = select_task_rq_fair, | |
1411 | #endif /* CONFIG_SMP */ | |
bf0f6f24 | 1412 | |
2e09bf55 | 1413 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1414 | |
1415 | .pick_next_task = pick_next_task_fair, | |
1416 | .put_prev_task = put_prev_task_fair, | |
1417 | ||
681f3e68 | 1418 | #ifdef CONFIG_SMP |
bf0f6f24 | 1419 | .load_balance = load_balance_fair, |
e1d1484f | 1420 | .move_one_task = move_one_task_fair, |
681f3e68 | 1421 | #endif |
bf0f6f24 | 1422 | |
83b699ed | 1423 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1424 | .task_tick = task_tick_fair, |
1425 | .task_new = task_new_fair, | |
cb469845 SR |
1426 | |
1427 | .prio_changed = prio_changed_fair, | |
1428 | .switched_to = switched_to_fair, | |
bf0f6f24 IM |
1429 | }; |
1430 | ||
1431 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1432 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1433 | { |
bf0f6f24 IM |
1434 | struct cfs_rq *cfs_rq; |
1435 | ||
75c28ace SV |
1436 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1437 | print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs); | |
1438 | #endif | |
5973e5b9 | 1439 | rcu_read_lock(); |
c3b64f1e | 1440 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1441 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1442 | rcu_read_unlock(); |
bf0f6f24 IM |
1443 | } |
1444 | #endif |