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. | |
74e3cd7f | 76 | * (default: 5 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 | */ | |
74e3cd7f | 82 | unsigned int sysctl_sched_wakeup_granularity = 5000000UL; |
bf0f6f24 | 83 | |
da84d961 IM |
84 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
85 | ||
bf0f6f24 IM |
86 | /************************************************************** |
87 | * CFS operations on generic schedulable entities: | |
88 | */ | |
89 | ||
62160e3f | 90 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 91 | |
62160e3f | 92 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
93 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
94 | { | |
62160e3f | 95 | return cfs_rq->rq; |
bf0f6f24 IM |
96 | } |
97 | ||
62160e3f IM |
98 | /* An entity is a task if it doesn't "own" a runqueue */ |
99 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 100 | |
62160e3f | 101 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 102 | |
62160e3f IM |
103 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
104 | { | |
105 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
106 | } |
107 | ||
108 | #define entity_is_task(se) 1 | |
109 | ||
bf0f6f24 IM |
110 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
111 | ||
112 | static inline struct task_struct *task_of(struct sched_entity *se) | |
113 | { | |
114 | return container_of(se, struct task_struct, se); | |
115 | } | |
116 | ||
117 | ||
118 | /************************************************************** | |
119 | * Scheduling class tree data structure manipulation methods: | |
120 | */ | |
121 | ||
0702e3eb | 122 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 123 | { |
368059a9 PZ |
124 | s64 delta = (s64)(vruntime - min_vruntime); |
125 | if (delta > 0) | |
02e0431a PZ |
126 | min_vruntime = vruntime; |
127 | ||
128 | return min_vruntime; | |
129 | } | |
130 | ||
0702e3eb | 131 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
132 | { |
133 | s64 delta = (s64)(vruntime - min_vruntime); | |
134 | if (delta < 0) | |
135 | min_vruntime = vruntime; | |
136 | ||
137 | return min_vruntime; | |
138 | } | |
139 | ||
0702e3eb | 140 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 141 | { |
30cfdcfc | 142 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
143 | } |
144 | ||
bf0f6f24 IM |
145 | /* |
146 | * Enqueue an entity into the rb-tree: | |
147 | */ | |
0702e3eb | 148 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
149 | { |
150 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
151 | struct rb_node *parent = NULL; | |
152 | struct sched_entity *entry; | |
9014623c | 153 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
154 | int leftmost = 1; |
155 | ||
156 | /* | |
157 | * Find the right place in the rbtree: | |
158 | */ | |
159 | while (*link) { | |
160 | parent = *link; | |
161 | entry = rb_entry(parent, struct sched_entity, run_node); | |
162 | /* | |
163 | * We dont care about collisions. Nodes with | |
164 | * the same key stay together. | |
165 | */ | |
9014623c | 166 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
167 | link = &parent->rb_left; |
168 | } else { | |
169 | link = &parent->rb_right; | |
170 | leftmost = 0; | |
171 | } | |
172 | } | |
173 | ||
174 | /* | |
175 | * Maintain a cache of leftmost tree entries (it is frequently | |
176 | * used): | |
177 | */ | |
3fe69747 | 178 | if (leftmost) { |
57cb499d | 179 | cfs_rq->rb_leftmost = &se->run_node; |
3fe69747 PZ |
180 | /* |
181 | * maintain cfs_rq->min_vruntime to be a monotonic increasing | |
182 | * value tracking the leftmost vruntime in the tree. | |
183 | */ | |
184 | cfs_rq->min_vruntime = | |
185 | max_vruntime(cfs_rq->min_vruntime, se->vruntime); | |
186 | } | |
bf0f6f24 IM |
187 | |
188 | rb_link_node(&se->run_node, parent, link); | |
189 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
190 | } |
191 | ||
0702e3eb | 192 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 193 | { |
3fe69747 PZ |
194 | if (cfs_rq->rb_leftmost == &se->run_node) { |
195 | struct rb_node *next_node; | |
196 | struct sched_entity *next; | |
197 | ||
198 | next_node = rb_next(&se->run_node); | |
199 | cfs_rq->rb_leftmost = next_node; | |
200 | ||
201 | if (next_node) { | |
202 | next = rb_entry(next_node, | |
203 | struct sched_entity, run_node); | |
204 | cfs_rq->min_vruntime = | |
205 | max_vruntime(cfs_rq->min_vruntime, | |
206 | next->vruntime); | |
207 | } | |
208 | } | |
e9acbff6 | 209 | |
aa2ac252 PZ |
210 | if (cfs_rq->next == se) |
211 | cfs_rq->next = NULL; | |
212 | ||
bf0f6f24 | 213 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
214 | } |
215 | ||
216 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) | |
217 | { | |
218 | return cfs_rq->rb_leftmost; | |
219 | } | |
220 | ||
221 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) | |
222 | { | |
223 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); | |
224 | } | |
225 | ||
aeb73b04 PZ |
226 | static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
227 | { | |
7eee3e67 | 228 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 229 | |
70eee74b BS |
230 | if (!last) |
231 | return NULL; | |
7eee3e67 IM |
232 | |
233 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
234 | } |
235 | ||
bf0f6f24 IM |
236 | /************************************************************** |
237 | * Scheduling class statistics methods: | |
238 | */ | |
239 | ||
b2be5e96 PZ |
240 | #ifdef CONFIG_SCHED_DEBUG |
241 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
242 | struct file *filp, void __user *buffer, size_t *lenp, | |
243 | loff_t *ppos) | |
244 | { | |
245 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
246 | ||
247 | if (ret || !write) | |
248 | return ret; | |
249 | ||
250 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
251 | sysctl_sched_min_granularity); | |
252 | ||
253 | return 0; | |
254 | } | |
255 | #endif | |
647e7cac IM |
256 | |
257 | /* | |
258 | * The idea is to set a period in which each task runs once. | |
259 | * | |
260 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
261 | * this period because otherwise the slices get too small. | |
262 | * | |
263 | * p = (nr <= nl) ? l : l*nr/nl | |
264 | */ | |
4d78e7b6 PZ |
265 | static u64 __sched_period(unsigned long nr_running) |
266 | { | |
267 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 268 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
269 | |
270 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 271 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 272 | period *= nr_running; |
4d78e7b6 PZ |
273 | } |
274 | ||
275 | return period; | |
276 | } | |
277 | ||
647e7cac IM |
278 | /* |
279 | * We calculate the wall-time slice from the period by taking a part | |
280 | * proportional to the weight. | |
281 | * | |
282 | * s = p*w/rw | |
283 | */ | |
6d0f0ebd | 284 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 285 | { |
6a6029b8 IM |
286 | return calc_delta_mine(__sched_period(cfs_rq->nr_running), |
287 | se->load.weight, &cfs_rq->load); | |
bf0f6f24 IM |
288 | } |
289 | ||
647e7cac IM |
290 | /* |
291 | * We calculate the vruntime slice. | |
292 | * | |
293 | * vs = s/w = p/rw | |
294 | */ | |
295 | static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running) | |
67e9fb2a | 296 | { |
647e7cac | 297 | u64 vslice = __sched_period(nr_running); |
67e9fb2a | 298 | |
10b77724 | 299 | vslice *= NICE_0_LOAD; |
647e7cac | 300 | do_div(vslice, rq_weight); |
67e9fb2a | 301 | |
647e7cac IM |
302 | return vslice; |
303 | } | |
5f6d858e | 304 | |
647e7cac IM |
305 | static u64 sched_vslice(struct cfs_rq *cfs_rq) |
306 | { | |
307 | return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running); | |
308 | } | |
309 | ||
310 | static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
311 | { | |
312 | return __sched_vslice(cfs_rq->load.weight + se->load.weight, | |
313 | cfs_rq->nr_running + 1); | |
67e9fb2a PZ |
314 | } |
315 | ||
bf0f6f24 IM |
316 | /* |
317 | * Update the current task's runtime statistics. Skip current tasks that | |
318 | * are not in our scheduling class. | |
319 | */ | |
320 | static inline void | |
8ebc91d9 IM |
321 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
322 | unsigned long delta_exec) | |
bf0f6f24 | 323 | { |
bbdba7c0 | 324 | unsigned long delta_exec_weighted; |
bf0f6f24 | 325 | |
8179ca23 | 326 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
327 | |
328 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 329 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
e9acbff6 IM |
330 | delta_exec_weighted = delta_exec; |
331 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { | |
332 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, | |
333 | &curr->load); | |
334 | } | |
335 | curr->vruntime += delta_exec_weighted; | |
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 | |
3fe69747 PZ |
501 | if (first_fair(cfs_rq)) { |
502 | vruntime = min_vruntime(cfs_rq->min_vruntime, | |
503 | __pick_next_entity(cfs_rq)->vruntime); | |
504 | } else | |
505 | vruntime = cfs_rq->min_vruntime; | |
94dfb5e7 | 506 | |
06877c33 | 507 | if (sched_feat(TREE_AVG)) { |
94dfb5e7 PZ |
508 | struct sched_entity *last = __pick_last_entity(cfs_rq); |
509 | if (last) { | |
67e9fb2a PZ |
510 | vruntime += last->vruntime; |
511 | vruntime >>= 1; | |
94dfb5e7 | 512 | } |
67e9fb2a | 513 | } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running) |
647e7cac | 514 | vruntime += sched_vslice(cfs_rq)/2; |
94dfb5e7 | 515 | |
2cb8600e PZ |
516 | /* |
517 | * The 'current' period is already promised to the current tasks, | |
518 | * however the extra weight of the new task will slow them down a | |
519 | * little, place the new task so that it fits in the slot that | |
520 | * stays open at the end. | |
521 | */ | |
94dfb5e7 | 522 | if (initial && sched_feat(START_DEBIT)) |
647e7cac | 523 | vruntime += sched_vslice_add(cfs_rq, se); |
aeb73b04 | 524 | |
8465e792 | 525 | if (!initial) { |
2cb8600e | 526 | /* sleeps upto a single latency don't count. */ |
e22ecef1 IM |
527 | if (sched_feat(NEW_FAIR_SLEEPERS)) { |
528 | vruntime -= calc_delta_fair(sysctl_sched_latency, | |
529 | &cfs_rq->load); | |
530 | } | |
94359f05 | 531 | |
2cb8600e PZ |
532 | /* ensure we never gain time by being placed backwards. */ |
533 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
534 | } |
535 | ||
67e9fb2a | 536 | se->vruntime = vruntime; |
aeb73b04 PZ |
537 | } |
538 | ||
bf0f6f24 | 539 | static void |
83b699ed | 540 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
541 | { |
542 | /* | |
a2a2d680 | 543 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 544 | */ |
b7cc0896 | 545 | update_curr(cfs_rq); |
bf0f6f24 | 546 | |
e9acbff6 | 547 | if (wakeup) { |
aeb73b04 | 548 | place_entity(cfs_rq, se, 0); |
2396af69 | 549 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 550 | } |
bf0f6f24 | 551 | |
d2417e5a | 552 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 553 | check_spread(cfs_rq, se); |
83b699ed SV |
554 | if (se != cfs_rq->curr) |
555 | __enqueue_entity(cfs_rq, se); | |
30cfdcfc | 556 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 IM |
557 | } |
558 | ||
4ae7d5ce IM |
559 | static void update_avg(u64 *avg, u64 sample) |
560 | { | |
561 | s64 diff = sample - *avg; | |
562 | *avg += diff >> 3; | |
563 | } | |
564 | ||
565 | static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
566 | { | |
567 | if (!se->last_wakeup) | |
568 | return; | |
569 | ||
570 | update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup); | |
571 | se->last_wakeup = 0; | |
572 | } | |
573 | ||
bf0f6f24 | 574 | static void |
525c2716 | 575 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 576 | { |
a2a2d680 DA |
577 | /* |
578 | * Update run-time statistics of the 'current'. | |
579 | */ | |
580 | update_curr(cfs_rq); | |
581 | ||
19b6a2e3 | 582 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 583 | if (sleep) { |
4ae7d5ce | 584 | update_avg_stats(cfs_rq, se); |
67e9fb2a | 585 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
586 | if (entity_is_task(se)) { |
587 | struct task_struct *tsk = task_of(se); | |
588 | ||
589 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 590 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 591 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 592 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 593 | } |
db36cc7d | 594 | #endif |
67e9fb2a PZ |
595 | } |
596 | ||
83b699ed | 597 | if (se != cfs_rq->curr) |
30cfdcfc DA |
598 | __dequeue_entity(cfs_rq, se); |
599 | account_entity_dequeue(cfs_rq, se); | |
bf0f6f24 IM |
600 | } |
601 | ||
602 | /* | |
603 | * Preempt the current task with a newly woken task if needed: | |
604 | */ | |
7c92e54f | 605 | static void |
2e09bf55 | 606 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 607 | { |
11697830 PZ |
608 | unsigned long ideal_runtime, delta_exec; |
609 | ||
6d0f0ebd | 610 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 611 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
3e3e13f3 | 612 | if (delta_exec > ideal_runtime) |
bf0f6f24 IM |
613 | resched_task(rq_of(cfs_rq)->curr); |
614 | } | |
615 | ||
83b699ed | 616 | static void |
8494f412 | 617 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 618 | { |
83b699ed SV |
619 | /* 'current' is not kept within the tree. */ |
620 | if (se->on_rq) { | |
621 | /* | |
622 | * Any task has to be enqueued before it get to execute on | |
623 | * a CPU. So account for the time it spent waiting on the | |
624 | * runqueue. | |
625 | */ | |
626 | update_stats_wait_end(cfs_rq, se); | |
627 | __dequeue_entity(cfs_rq, se); | |
628 | } | |
629 | ||
79303e9e | 630 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 631 | cfs_rq->curr = se; |
eba1ed4b IM |
632 | #ifdef CONFIG_SCHEDSTATS |
633 | /* | |
634 | * Track our maximum slice length, if the CPU's load is at | |
635 | * least twice that of our own weight (i.e. dont track it | |
636 | * when there are only lesser-weight tasks around): | |
637 | */ | |
495eca49 | 638 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
639 | se->slice_max = max(se->slice_max, |
640 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
641 | } | |
642 | #endif | |
4a55b450 | 643 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
644 | } |
645 | ||
aa2ac252 PZ |
646 | static struct sched_entity * |
647 | pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
648 | { | |
649 | s64 diff, gran; | |
650 | ||
651 | if (!cfs_rq->next) | |
652 | return se; | |
653 | ||
654 | diff = cfs_rq->next->vruntime - se->vruntime; | |
655 | if (diff < 0) | |
656 | return se; | |
657 | ||
658 | gran = calc_delta_fair(sysctl_sched_wakeup_granularity, &cfs_rq->load); | |
659 | if (diff > gran) | |
660 | return se; | |
661 | ||
662 | return cfs_rq->next; | |
663 | } | |
664 | ||
9948f4b2 | 665 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 666 | { |
08ec3df5 | 667 | struct sched_entity *se = NULL; |
bf0f6f24 | 668 | |
08ec3df5 DA |
669 | if (first_fair(cfs_rq)) { |
670 | se = __pick_next_entity(cfs_rq); | |
aa2ac252 | 671 | se = pick_next(cfs_rq, se); |
08ec3df5 DA |
672 | set_next_entity(cfs_rq, se); |
673 | } | |
bf0f6f24 IM |
674 | |
675 | return se; | |
676 | } | |
677 | ||
ab6cde26 | 678 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
679 | { |
680 | /* | |
681 | * If still on the runqueue then deactivate_task() | |
682 | * was not called and update_curr() has to be done: | |
683 | */ | |
684 | if (prev->on_rq) | |
b7cc0896 | 685 | update_curr(cfs_rq); |
bf0f6f24 | 686 | |
ddc97297 | 687 | check_spread(cfs_rq, prev); |
30cfdcfc | 688 | if (prev->on_rq) { |
5870db5b | 689 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
690 | /* Put 'current' back into the tree. */ |
691 | __enqueue_entity(cfs_rq, prev); | |
692 | } | |
429d43bc | 693 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
694 | } |
695 | ||
8f4d37ec PZ |
696 | static void |
697 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 698 | { |
bf0f6f24 | 699 | /* |
30cfdcfc | 700 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 701 | */ |
30cfdcfc | 702 | update_curr(cfs_rq); |
bf0f6f24 | 703 | |
8f4d37ec PZ |
704 | #ifdef CONFIG_SCHED_HRTICK |
705 | /* | |
706 | * queued ticks are scheduled to match the slice, so don't bother | |
707 | * validating it and just reschedule. | |
708 | */ | |
709 | if (queued) | |
710 | return resched_task(rq_of(cfs_rq)->curr); | |
711 | /* | |
712 | * don't let the period tick interfere with the hrtick preemption | |
713 | */ | |
714 | if (!sched_feat(DOUBLE_TICK) && | |
715 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
716 | return; | |
717 | #endif | |
718 | ||
ce6c1311 | 719 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 720 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
721 | } |
722 | ||
723 | /************************************************** | |
724 | * CFS operations on tasks: | |
725 | */ | |
726 | ||
727 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
728 | ||
729 | /* Walk up scheduling entities hierarchy */ | |
730 | #define for_each_sched_entity(se) \ | |
731 | for (; se; se = se->parent) | |
732 | ||
733 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
734 | { | |
735 | return p->se.cfs_rq; | |
736 | } | |
737 | ||
738 | /* runqueue on which this entity is (to be) queued */ | |
739 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
740 | { | |
741 | return se->cfs_rq; | |
742 | } | |
743 | ||
744 | /* runqueue "owned" by this group */ | |
745 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
746 | { | |
747 | return grp->my_q; | |
748 | } | |
749 | ||
750 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
751 | * another cpu ('this_cpu') | |
752 | */ | |
753 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
754 | { | |
29f59db3 | 755 | return cfs_rq->tg->cfs_rq[this_cpu]; |
bf0f6f24 IM |
756 | } |
757 | ||
758 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
759 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
ec2c507f | 760 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
bf0f6f24 | 761 | |
fad095a7 SV |
762 | /* Do the two (enqueued) entities belong to the same group ? */ |
763 | static inline int | |
764 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 | 765 | { |
fad095a7 | 766 | if (se->cfs_rq == pse->cfs_rq) |
bf0f6f24 IM |
767 | return 1; |
768 | ||
769 | return 0; | |
770 | } | |
771 | ||
fad095a7 SV |
772 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
773 | { | |
774 | return se->parent; | |
775 | } | |
776 | ||
bf0f6f24 IM |
777 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
778 | ||
779 | #define for_each_sched_entity(se) \ | |
780 | for (; se; se = NULL) | |
781 | ||
782 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
783 | { | |
784 | return &task_rq(p)->cfs; | |
785 | } | |
786 | ||
787 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
788 | { | |
789 | struct task_struct *p = task_of(se); | |
790 | struct rq *rq = task_rq(p); | |
791 | ||
792 | return &rq->cfs; | |
793 | } | |
794 | ||
795 | /* runqueue "owned" by this group */ | |
796 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
797 | { | |
798 | return NULL; | |
799 | } | |
800 | ||
801 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
802 | { | |
803 | return &cpu_rq(this_cpu)->cfs; | |
804 | } | |
805 | ||
806 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
807 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
808 | ||
fad095a7 SV |
809 | static inline int |
810 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
bf0f6f24 IM |
811 | { |
812 | return 1; | |
813 | } | |
814 | ||
fad095a7 SV |
815 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
816 | { | |
817 | return NULL; | |
818 | } | |
819 | ||
bf0f6f24 IM |
820 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
821 | ||
8f4d37ec PZ |
822 | #ifdef CONFIG_SCHED_HRTICK |
823 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
824 | { | |
825 | int requeue = rq->curr == p; | |
826 | struct sched_entity *se = &p->se; | |
827 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
828 | ||
829 | WARN_ON(task_rq(p) != rq); | |
830 | ||
831 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
832 | u64 slice = sched_slice(cfs_rq, se); | |
833 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
834 | s64 delta = slice - ran; | |
835 | ||
836 | if (delta < 0) { | |
837 | if (rq->curr == p) | |
838 | resched_task(p); | |
839 | return; | |
840 | } | |
841 | ||
842 | /* | |
843 | * Don't schedule slices shorter than 10000ns, that just | |
844 | * doesn't make sense. Rely on vruntime for fairness. | |
845 | */ | |
846 | if (!requeue) | |
847 | delta = max(10000LL, delta); | |
848 | ||
849 | hrtick_start(rq, delta, requeue); | |
850 | } | |
851 | } | |
852 | #else | |
853 | static inline void | |
854 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
855 | { | |
856 | } | |
857 | #endif | |
858 | ||
bf0f6f24 IM |
859 | /* |
860 | * The enqueue_task method is called before nr_running is | |
861 | * increased. Here we update the fair scheduling stats and | |
862 | * then put the task into the rbtree: | |
863 | */ | |
fd390f6a | 864 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
865 | { |
866 | struct cfs_rq *cfs_rq; | |
62fb1851 | 867 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
868 | |
869 | for_each_sched_entity(se) { | |
62fb1851 | 870 | if (se->on_rq) |
bf0f6f24 IM |
871 | break; |
872 | cfs_rq = cfs_rq_of(se); | |
83b699ed | 873 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 874 | wakeup = 1; |
bf0f6f24 | 875 | } |
8f4d37ec PZ |
876 | |
877 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
878 | } |
879 | ||
880 | /* | |
881 | * The dequeue_task method is called before nr_running is | |
882 | * decreased. We remove the task from the rbtree and | |
883 | * update the fair scheduling stats: | |
884 | */ | |
f02231e5 | 885 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
886 | { |
887 | struct cfs_rq *cfs_rq; | |
62fb1851 | 888 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
889 | |
890 | for_each_sched_entity(se) { | |
891 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 892 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 893 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 894 | if (cfs_rq->load.weight) |
bf0f6f24 | 895 | break; |
b9fa3df3 | 896 | sleep = 1; |
bf0f6f24 | 897 | } |
8f4d37ec PZ |
898 | |
899 | hrtick_start_fair(rq, rq->curr); | |
bf0f6f24 IM |
900 | } |
901 | ||
902 | /* | |
1799e35d IM |
903 | * sched_yield() support is very simple - we dequeue and enqueue. |
904 | * | |
905 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 906 | */ |
4530d7ab | 907 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 908 | { |
db292ca3 IM |
909 | struct task_struct *curr = rq->curr; |
910 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
911 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
912 | |
913 | /* | |
1799e35d IM |
914 | * Are we the only task in the tree? |
915 | */ | |
916 | if (unlikely(cfs_rq->nr_running == 1)) | |
917 | return; | |
918 | ||
db292ca3 | 919 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
1799e35d IM |
920 | __update_rq_clock(rq); |
921 | /* | |
a2a2d680 | 922 | * Update run-time statistics of the 'current'. |
1799e35d | 923 | */ |
2b1e315d | 924 | update_curr(cfs_rq); |
1799e35d IM |
925 | |
926 | return; | |
927 | } | |
928 | /* | |
929 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 930 | */ |
2b1e315d | 931 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
932 | /* |
933 | * Already in the rightmost position? | |
934 | */ | |
2b1e315d | 935 | if (unlikely(rightmost->vruntime < se->vruntime)) |
1799e35d IM |
936 | return; |
937 | ||
938 | /* | |
939 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
940 | * Upon rescheduling, sched_class::put_prev_task() will place |
941 | * 'current' within the tree based on its new key value. | |
1799e35d | 942 | */ |
30cfdcfc | 943 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
944 | } |
945 | ||
e7693a36 GH |
946 | /* |
947 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
948 | * not idle and an idle cpu is available. The span of cpus to | |
949 | * search starts with cpus closest then further out as needed, | |
950 | * so we always favor a closer, idle cpu. | |
951 | * | |
952 | * Returns the CPU we should wake onto. | |
953 | */ | |
954 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
955 | static int wake_idle(int cpu, struct task_struct *p) | |
956 | { | |
957 | cpumask_t tmp; | |
958 | struct sched_domain *sd; | |
959 | int i; | |
960 | ||
961 | /* | |
962 | * If it is idle, then it is the best cpu to run this task. | |
963 | * | |
964 | * This cpu is also the best, if it has more than one task already. | |
965 | * Siblings must be also busy(in most cases) as they didn't already | |
966 | * pickup the extra load from this cpu and hence we need not check | |
967 | * sibling runqueue info. This will avoid the checks and cache miss | |
968 | * penalities associated with that. | |
969 | */ | |
970 | if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1) | |
971 | return cpu; | |
972 | ||
973 | for_each_domain(cpu, sd) { | |
974 | if (sd->flags & SD_WAKE_IDLE) { | |
975 | cpus_and(tmp, sd->span, p->cpus_allowed); | |
976 | for_each_cpu_mask(i, tmp) { | |
977 | if (idle_cpu(i)) { | |
978 | if (i != task_cpu(p)) { | |
979 | schedstat_inc(p, | |
980 | se.nr_wakeups_idle); | |
981 | } | |
982 | return i; | |
983 | } | |
984 | } | |
985 | } else { | |
986 | break; | |
987 | } | |
988 | } | |
989 | return cpu; | |
990 | } | |
991 | #else | |
992 | static inline int wake_idle(int cpu, struct task_struct *p) | |
993 | { | |
994 | return cpu; | |
995 | } | |
996 | #endif | |
997 | ||
998 | #ifdef CONFIG_SMP | |
098fb9db | 999 | |
4ae7d5ce IM |
1000 | static const struct sched_class fair_sched_class; |
1001 | ||
098fb9db | 1002 | static int |
4ae7d5ce IM |
1003 | wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq, |
1004 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, | |
1005 | int idx, unsigned long load, unsigned long this_load, | |
098fb9db IM |
1006 | unsigned int imbalance) |
1007 | { | |
4ae7d5ce | 1008 | struct task_struct *curr = this_rq->curr; |
098fb9db IM |
1009 | unsigned long tl = this_load; |
1010 | unsigned long tl_per_task; | |
1011 | ||
1012 | if (!(this_sd->flags & SD_WAKE_AFFINE)) | |
1013 | return 0; | |
1014 | ||
1015 | /* | |
4ae7d5ce IM |
1016 | * If the currently running task will sleep within |
1017 | * a reasonable amount of time then attract this newly | |
1018 | * woken task: | |
098fb9db | 1019 | */ |
4ae7d5ce IM |
1020 | if (sync && curr->sched_class == &fair_sched_class) { |
1021 | if (curr->se.avg_overlap < sysctl_sched_migration_cost && | |
1022 | p->se.avg_overlap < sysctl_sched_migration_cost) | |
1023 | return 1; | |
1024 | } | |
098fb9db IM |
1025 | |
1026 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1027 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1028 | ||
1029 | /* | |
1030 | * If sync wakeup then subtract the (maximum possible) | |
1031 | * effect of the currently running task from the load | |
1032 | * of the current CPU: | |
1033 | */ | |
1034 | if (sync) | |
1035 | tl -= current->se.load.weight; | |
1036 | ||
ac192d39 | 1037 | if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) || |
098fb9db IM |
1038 | 100*(tl + p->se.load.weight) <= imbalance*load) { |
1039 | /* | |
1040 | * This domain has SD_WAKE_AFFINE and | |
1041 | * p is cache cold in this domain, and | |
1042 | * there is no bad imbalance. | |
1043 | */ | |
1044 | schedstat_inc(this_sd, ttwu_move_affine); | |
1045 | schedstat_inc(p, se.nr_wakeups_affine); | |
1046 | ||
1047 | return 1; | |
1048 | } | |
1049 | return 0; | |
1050 | } | |
1051 | ||
e7693a36 GH |
1052 | static int select_task_rq_fair(struct task_struct *p, int sync) |
1053 | { | |
e7693a36 | 1054 | struct sched_domain *sd, *this_sd = NULL; |
ac192d39 | 1055 | int prev_cpu, this_cpu, new_cpu; |
098fb9db | 1056 | unsigned long load, this_load; |
4ae7d5ce | 1057 | struct rq *rq, *this_rq; |
098fb9db | 1058 | unsigned int imbalance; |
098fb9db | 1059 | int idx; |
e7693a36 | 1060 | |
ac192d39 IM |
1061 | prev_cpu = task_cpu(p); |
1062 | rq = task_rq(p); | |
1063 | this_cpu = smp_processor_id(); | |
4ae7d5ce | 1064 | this_rq = cpu_rq(this_cpu); |
ac192d39 | 1065 | new_cpu = prev_cpu; |
e7693a36 | 1066 | |
ac192d39 IM |
1067 | /* |
1068 | * 'this_sd' is the first domain that both | |
1069 | * this_cpu and prev_cpu are present in: | |
1070 | */ | |
e7693a36 | 1071 | for_each_domain(this_cpu, sd) { |
ac192d39 | 1072 | if (cpu_isset(prev_cpu, sd->span)) { |
e7693a36 GH |
1073 | this_sd = sd; |
1074 | break; | |
1075 | } | |
1076 | } | |
1077 | ||
1078 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) | |
f4827386 | 1079 | goto out; |
e7693a36 GH |
1080 | |
1081 | /* | |
1082 | * Check for affine wakeup and passive balancing possibilities. | |
1083 | */ | |
098fb9db | 1084 | if (!this_sd) |
f4827386 | 1085 | goto out; |
e7693a36 | 1086 | |
098fb9db IM |
1087 | idx = this_sd->wake_idx; |
1088 | ||
1089 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1090 | ||
ac192d39 | 1091 | load = source_load(prev_cpu, idx); |
098fb9db IM |
1092 | this_load = target_load(this_cpu, idx); |
1093 | ||
4ae7d5ce IM |
1094 | if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, |
1095 | load, this_load, imbalance)) | |
1096 | return this_cpu; | |
1097 | ||
1098 | if (prev_cpu == this_cpu) | |
f4827386 | 1099 | goto out; |
098fb9db IM |
1100 | |
1101 | /* | |
1102 | * Start passive balancing when half the imbalance_pct | |
1103 | * limit is reached. | |
1104 | */ | |
1105 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1106 | if (imbalance*this_load <= 100*load) { | |
1107 | schedstat_inc(this_sd, ttwu_move_balance); | |
1108 | schedstat_inc(p, se.nr_wakeups_passive); | |
4ae7d5ce | 1109 | return this_cpu; |
e7693a36 GH |
1110 | } |
1111 | } | |
1112 | ||
f4827386 | 1113 | out: |
e7693a36 GH |
1114 | return wake_idle(new_cpu, p); |
1115 | } | |
1116 | #endif /* CONFIG_SMP */ | |
1117 | ||
1118 | ||
bf0f6f24 IM |
1119 | /* |
1120 | * Preempt the current task with a newly woken task if needed: | |
1121 | */ | |
2e09bf55 | 1122 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1123 | { |
1124 | struct task_struct *curr = rq->curr; | |
fad095a7 | 1125 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
8651a86c | 1126 | struct sched_entity *se = &curr->se, *pse = &p->se; |
502d26b5 | 1127 | unsigned long gran; |
bf0f6f24 IM |
1128 | |
1129 | if (unlikely(rt_prio(p->prio))) { | |
a8e504d2 | 1130 | update_rq_clock(rq); |
b7cc0896 | 1131 | update_curr(cfs_rq); |
bf0f6f24 IM |
1132 | resched_task(curr); |
1133 | return; | |
1134 | } | |
aa2ac252 | 1135 | |
4ae7d5ce IM |
1136 | se->last_wakeup = se->sum_exec_runtime; |
1137 | if (unlikely(se == pse)) | |
1138 | return; | |
1139 | ||
aa2ac252 PZ |
1140 | cfs_rq_of(pse)->next = pse; |
1141 | ||
91c234b4 IM |
1142 | /* |
1143 | * Batch tasks do not preempt (their preemption is driven by | |
1144 | * the tick): | |
1145 | */ | |
1146 | if (unlikely(p->policy == SCHED_BATCH)) | |
1147 | return; | |
bf0f6f24 | 1148 | |
77d9cc44 IM |
1149 | if (!sched_feat(WAKEUP_PREEMPT)) |
1150 | return; | |
8651a86c | 1151 | |
77d9cc44 IM |
1152 | while (!is_same_group(se, pse)) { |
1153 | se = parent_entity(se); | |
1154 | pse = parent_entity(pse); | |
ce6c1311 | 1155 | } |
77d9cc44 | 1156 | |
77d9cc44 | 1157 | gran = sysctl_sched_wakeup_granularity; |
ef9884e6 PZ |
1158 | /* |
1159 | * More easily preempt - nice tasks, while not making | |
1160 | * it harder for + nice tasks. | |
1161 | */ | |
1162 | if (unlikely(se->load.weight > NICE_0_LOAD)) | |
77d9cc44 IM |
1163 | gran = calc_delta_fair(gran, &se->load); |
1164 | ||
502d26b5 | 1165 | if (pse->vruntime + gran < se->vruntime) |
77d9cc44 | 1166 | resched_task(curr); |
bf0f6f24 IM |
1167 | } |
1168 | ||
fb8d4724 | 1169 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1170 | { |
8f4d37ec | 1171 | struct task_struct *p; |
bf0f6f24 IM |
1172 | struct cfs_rq *cfs_rq = &rq->cfs; |
1173 | struct sched_entity *se; | |
1174 | ||
1175 | if (unlikely(!cfs_rq->nr_running)) | |
1176 | return NULL; | |
1177 | ||
1178 | do { | |
9948f4b2 | 1179 | se = pick_next_entity(cfs_rq); |
bf0f6f24 IM |
1180 | cfs_rq = group_cfs_rq(se); |
1181 | } while (cfs_rq); | |
1182 | ||
8f4d37ec PZ |
1183 | p = task_of(se); |
1184 | hrtick_start_fair(rq, p); | |
1185 | ||
1186 | return p; | |
bf0f6f24 IM |
1187 | } |
1188 | ||
1189 | /* | |
1190 | * Account for a descheduled task: | |
1191 | */ | |
31ee529c | 1192 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1193 | { |
1194 | struct sched_entity *se = &prev->se; | |
1195 | struct cfs_rq *cfs_rq; | |
1196 | ||
1197 | for_each_sched_entity(se) { | |
1198 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1199 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1200 | } |
1201 | } | |
1202 | ||
681f3e68 | 1203 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1204 | /************************************************** |
1205 | * Fair scheduling class load-balancing methods: | |
1206 | */ | |
1207 | ||
1208 | /* | |
1209 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1210 | * during the whole iteration, the current task might be | |
1211 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1212 | * achieve that by always pre-iterating before returning | |
1213 | * the current task: | |
1214 | */ | |
a9957449 | 1215 | static struct task_struct * |
bf0f6f24 IM |
1216 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) |
1217 | { | |
1218 | struct task_struct *p; | |
1219 | ||
1220 | if (!curr) | |
1221 | return NULL; | |
1222 | ||
1223 | p = rb_entry(curr, struct task_struct, se.run_node); | |
1224 | cfs_rq->rb_load_balance_curr = rb_next(curr); | |
1225 | ||
1226 | return p; | |
1227 | } | |
1228 | ||
1229 | static struct task_struct *load_balance_start_fair(void *arg) | |
1230 | { | |
1231 | struct cfs_rq *cfs_rq = arg; | |
1232 | ||
1233 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); | |
1234 | } | |
1235 | ||
1236 | static struct task_struct *load_balance_next_fair(void *arg) | |
1237 | { | |
1238 | struct cfs_rq *cfs_rq = arg; | |
1239 | ||
1240 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); | |
1241 | } | |
1242 | ||
62fb1851 PZ |
1243 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1244 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) | |
1245 | { | |
1246 | struct sched_entity *curr; | |
1247 | struct task_struct *p; | |
1248 | ||
1249 | if (!cfs_rq->nr_running || !first_fair(cfs_rq)) | |
1250 | return MAX_PRIO; | |
1251 | ||
1252 | curr = cfs_rq->curr; | |
1253 | if (!curr) | |
1254 | curr = __pick_next_entity(cfs_rq); | |
1255 | ||
1256 | p = task_of(curr); | |
1257 | ||
1258 | return p->prio; | |
1259 | } | |
1260 | #endif | |
1261 | ||
43010659 | 1262 | static unsigned long |
bf0f6f24 | 1263 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1264 | unsigned long max_load_move, |
a4ac01c3 PW |
1265 | struct sched_domain *sd, enum cpu_idle_type idle, |
1266 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 IM |
1267 | { |
1268 | struct cfs_rq *busy_cfs_rq; | |
bf0f6f24 IM |
1269 | long rem_load_move = max_load_move; |
1270 | struct rq_iterator cfs_rq_iterator; | |
1271 | ||
1272 | cfs_rq_iterator.start = load_balance_start_fair; | |
1273 | cfs_rq_iterator.next = load_balance_next_fair; | |
1274 | ||
1275 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
a4ac01c3 | 1276 | #ifdef CONFIG_FAIR_GROUP_SCHED |
62fb1851 PZ |
1277 | struct cfs_rq *this_cfs_rq; |
1278 | long imbalance; | |
1279 | unsigned long maxload; | |
bf0f6f24 | 1280 | |
62fb1851 | 1281 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); |
6b2d7700 | 1282 | |
62fb1851 PZ |
1283 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
1284 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ | |
1285 | if (imbalance <= 0) | |
bf0f6f24 IM |
1286 | continue; |
1287 | ||
62fb1851 PZ |
1288 | /* Don't pull more than imbalance/2 */ |
1289 | imbalance /= 2; | |
1290 | maxload = min(rem_load_move, imbalance); | |
bf0f6f24 | 1291 | |
62fb1851 | 1292 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
a4ac01c3 | 1293 | #else |
e56f31aa | 1294 | # define maxload rem_load_move |
a4ac01c3 | 1295 | #endif |
e1d1484f PW |
1296 | /* |
1297 | * pass busy_cfs_rq argument into | |
bf0f6f24 IM |
1298 | * load_balance_[start|next]_fair iterators |
1299 | */ | |
1300 | cfs_rq_iterator.arg = busy_cfs_rq; | |
62fb1851 | 1301 | rem_load_move -= balance_tasks(this_rq, this_cpu, busiest, |
e1d1484f PW |
1302 | maxload, sd, idle, all_pinned, |
1303 | this_best_prio, | |
1304 | &cfs_rq_iterator); | |
bf0f6f24 | 1305 | |
e1d1484f | 1306 | if (rem_load_move <= 0) |
bf0f6f24 IM |
1307 | break; |
1308 | } | |
1309 | ||
43010659 | 1310 | return max_load_move - rem_load_move; |
bf0f6f24 IM |
1311 | } |
1312 | ||
e1d1484f PW |
1313 | static int |
1314 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1315 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1316 | { | |
1317 | struct cfs_rq *busy_cfs_rq; | |
1318 | struct rq_iterator cfs_rq_iterator; | |
1319 | ||
1320 | cfs_rq_iterator.start = load_balance_start_fair; | |
1321 | cfs_rq_iterator.next = load_balance_next_fair; | |
1322 | ||
1323 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1324 | /* | |
1325 | * pass busy_cfs_rq argument into | |
1326 | * load_balance_[start|next]_fair iterators | |
1327 | */ | |
1328 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1329 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1330 | &cfs_rq_iterator)) | |
1331 | return 1; | |
1332 | } | |
1333 | ||
1334 | return 0; | |
1335 | } | |
681f3e68 | 1336 | #endif |
e1d1484f | 1337 | |
bf0f6f24 IM |
1338 | /* |
1339 | * scheduler tick hitting a task of our scheduling class: | |
1340 | */ | |
8f4d37ec | 1341 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1342 | { |
1343 | struct cfs_rq *cfs_rq; | |
1344 | struct sched_entity *se = &curr->se; | |
1345 | ||
1346 | for_each_sched_entity(se) { | |
1347 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1348 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1349 | } |
1350 | } | |
1351 | ||
8eb172d9 | 1352 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
4d78e7b6 | 1353 | |
bf0f6f24 IM |
1354 | /* |
1355 | * Share the fairness runtime between parent and child, thus the | |
1356 | * total amount of pressure for CPU stays equal - new tasks | |
1357 | * get a chance to run but frequent forkers are not allowed to | |
1358 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1359 | * the child is not running yet. | |
1360 | */ | |
ee0827d8 | 1361 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1362 | { |
1363 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1364 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1365 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1366 | |
1367 | sched_info_queued(p); | |
1368 | ||
7109c442 | 1369 | update_curr(cfs_rq); |
aeb73b04 | 1370 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1371 | |
3c90e6e9 | 1372 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1373 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
3c90e6e9 | 1374 | curr && curr->vruntime < se->vruntime) { |
87fefa38 | 1375 | /* |
edcb60a3 IM |
1376 | * Upon rescheduling, sched_class::put_prev_task() will place |
1377 | * 'current' within the tree based on its new key value. | |
1378 | */ | |
4d78e7b6 | 1379 | swap(curr->vruntime, se->vruntime); |
4d78e7b6 | 1380 | } |
bf0f6f24 | 1381 | |
b9dca1e0 | 1382 | enqueue_task_fair(rq, p, 0); |
bb61c210 | 1383 | resched_task(rq->curr); |
bf0f6f24 IM |
1384 | } |
1385 | ||
cb469845 SR |
1386 | /* |
1387 | * Priority of the task has changed. Check to see if we preempt | |
1388 | * the current task. | |
1389 | */ | |
1390 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1391 | int oldprio, int running) | |
1392 | { | |
1393 | /* | |
1394 | * Reschedule if we are currently running on this runqueue and | |
1395 | * our priority decreased, or if we are not currently running on | |
1396 | * this runqueue and our priority is higher than the current's | |
1397 | */ | |
1398 | if (running) { | |
1399 | if (p->prio > oldprio) | |
1400 | resched_task(rq->curr); | |
1401 | } else | |
1402 | check_preempt_curr(rq, p); | |
1403 | } | |
1404 | ||
1405 | /* | |
1406 | * We switched to the sched_fair class. | |
1407 | */ | |
1408 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1409 | int running) | |
1410 | { | |
1411 | /* | |
1412 | * We were most likely switched from sched_rt, so | |
1413 | * kick off the schedule if running, otherwise just see | |
1414 | * if we can still preempt the current task. | |
1415 | */ | |
1416 | if (running) | |
1417 | resched_task(rq->curr); | |
1418 | else | |
1419 | check_preempt_curr(rq, p); | |
1420 | } | |
1421 | ||
83b699ed SV |
1422 | /* Account for a task changing its policy or group. |
1423 | * | |
1424 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1425 | * migrates between groups/classes. | |
1426 | */ | |
1427 | static void set_curr_task_fair(struct rq *rq) | |
1428 | { | |
1429 | struct sched_entity *se = &rq->curr->se; | |
1430 | ||
1431 | for_each_sched_entity(se) | |
1432 | set_next_entity(cfs_rq_of(se), se); | |
1433 | } | |
1434 | ||
810b3817 PZ |
1435 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1436 | static void moved_group_fair(struct task_struct *p) | |
1437 | { | |
1438 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1439 | ||
1440 | update_curr(cfs_rq); | |
1441 | place_entity(cfs_rq, &p->se, 1); | |
1442 | } | |
1443 | #endif | |
1444 | ||
bf0f6f24 IM |
1445 | /* |
1446 | * All the scheduling class methods: | |
1447 | */ | |
5522d5d5 IM |
1448 | static const struct sched_class fair_sched_class = { |
1449 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1450 | .enqueue_task = enqueue_task_fair, |
1451 | .dequeue_task = dequeue_task_fair, | |
1452 | .yield_task = yield_task_fair, | |
e7693a36 GH |
1453 | #ifdef CONFIG_SMP |
1454 | .select_task_rq = select_task_rq_fair, | |
1455 | #endif /* CONFIG_SMP */ | |
bf0f6f24 | 1456 | |
2e09bf55 | 1457 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1458 | |
1459 | .pick_next_task = pick_next_task_fair, | |
1460 | .put_prev_task = put_prev_task_fair, | |
1461 | ||
681f3e68 | 1462 | #ifdef CONFIG_SMP |
bf0f6f24 | 1463 | .load_balance = load_balance_fair, |
e1d1484f | 1464 | .move_one_task = move_one_task_fair, |
681f3e68 | 1465 | #endif |
bf0f6f24 | 1466 | |
83b699ed | 1467 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1468 | .task_tick = task_tick_fair, |
1469 | .task_new = task_new_fair, | |
cb469845 SR |
1470 | |
1471 | .prio_changed = prio_changed_fair, | |
1472 | .switched_to = switched_to_fair, | |
810b3817 PZ |
1473 | |
1474 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1475 | .moved_group = moved_group_fair, | |
1476 | #endif | |
bf0f6f24 IM |
1477 | }; |
1478 | ||
1479 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1480 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1481 | { |
bf0f6f24 IM |
1482 | struct cfs_rq *cfs_rq; |
1483 | ||
75c28ace SV |
1484 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1485 | print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs); | |
1486 | #endif | |
5973e5b9 | 1487 | rcu_read_lock(); |
c3b64f1e | 1488 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1489 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1490 | rcu_read_unlock(); |
bf0f6f24 IM |
1491 | } |
1492 | #endif |