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: |
172e082a | 27 | * (default: 5ms * (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 | */ |
172e082a | 37 | unsigned int sysctl_sched_latency = 5000000ULL; |
2bd8e6d4 IM |
38 | |
39 | /* | |
b2be5e96 | 40 | * Minimal preemption granularity for CPU-bound tasks: |
172e082a | 41 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 42 | */ |
172e082a | 43 | unsigned int sysctl_sched_min_granularity = 1000000ULL; |
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 | /* | |
2bba22c5 | 51 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 52 | * parent will (try to) run first. |
21805085 | 53 | */ |
2bba22c5 | 54 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
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_OTHER wake-up granularity. | |
172e082a | 66 | * (default: 1 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 | */ | |
172e082a | 72 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 73 | |
da84d961 IM |
74 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
75 | ||
a4c2f00f PZ |
76 | static const struct sched_class fair_sched_class; |
77 | ||
bf0f6f24 IM |
78 | /************************************************************** |
79 | * CFS operations on generic schedulable entities: | |
80 | */ | |
81 | ||
62160e3f | 82 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 83 | |
62160e3f | 84 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
85 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
86 | { | |
62160e3f | 87 | return cfs_rq->rq; |
bf0f6f24 IM |
88 | } |
89 | ||
62160e3f IM |
90 | /* An entity is a task if it doesn't "own" a runqueue */ |
91 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 92 | |
8f48894f PZ |
93 | static inline struct task_struct *task_of(struct sched_entity *se) |
94 | { | |
95 | #ifdef CONFIG_SCHED_DEBUG | |
96 | WARN_ON_ONCE(!entity_is_task(se)); | |
97 | #endif | |
98 | return container_of(se, struct task_struct, se); | |
99 | } | |
100 | ||
b758149c PZ |
101 | /* Walk up scheduling entities hierarchy */ |
102 | #define for_each_sched_entity(se) \ | |
103 | for (; se; se = se->parent) | |
104 | ||
105 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
106 | { | |
107 | return p->se.cfs_rq; | |
108 | } | |
109 | ||
110 | /* runqueue on which this entity is (to be) queued */ | |
111 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
112 | { | |
113 | return se->cfs_rq; | |
114 | } | |
115 | ||
116 | /* runqueue "owned" by this group */ | |
117 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
118 | { | |
119 | return grp->my_q; | |
120 | } | |
121 | ||
122 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
123 | * another cpu ('this_cpu') | |
124 | */ | |
125 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
126 | { | |
127 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
128 | } | |
129 | ||
130 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
131 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
132 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
133 | ||
134 | /* Do the two (enqueued) entities belong to the same group ? */ | |
135 | static inline int | |
136 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
137 | { | |
138 | if (se->cfs_rq == pse->cfs_rq) | |
139 | return 1; | |
140 | ||
141 | return 0; | |
142 | } | |
143 | ||
144 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
145 | { | |
146 | return se->parent; | |
147 | } | |
148 | ||
464b7527 PZ |
149 | /* return depth at which a sched entity is present in the hierarchy */ |
150 | static inline int depth_se(struct sched_entity *se) | |
151 | { | |
152 | int depth = 0; | |
153 | ||
154 | for_each_sched_entity(se) | |
155 | depth++; | |
156 | ||
157 | return depth; | |
158 | } | |
159 | ||
160 | static void | |
161 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
162 | { | |
163 | int se_depth, pse_depth; | |
164 | ||
165 | /* | |
166 | * preemption test can be made between sibling entities who are in the | |
167 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
168 | * both tasks until we find their ancestors who are siblings of common | |
169 | * parent. | |
170 | */ | |
171 | ||
172 | /* First walk up until both entities are at same depth */ | |
173 | se_depth = depth_se(*se); | |
174 | pse_depth = depth_se(*pse); | |
175 | ||
176 | while (se_depth > pse_depth) { | |
177 | se_depth--; | |
178 | *se = parent_entity(*se); | |
179 | } | |
180 | ||
181 | while (pse_depth > se_depth) { | |
182 | pse_depth--; | |
183 | *pse = parent_entity(*pse); | |
184 | } | |
185 | ||
186 | while (!is_same_group(*se, *pse)) { | |
187 | *se = parent_entity(*se); | |
188 | *pse = parent_entity(*pse); | |
189 | } | |
190 | } | |
191 | ||
8f48894f PZ |
192 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
193 | ||
194 | static inline struct task_struct *task_of(struct sched_entity *se) | |
195 | { | |
196 | return container_of(se, struct task_struct, se); | |
197 | } | |
bf0f6f24 | 198 | |
62160e3f IM |
199 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
200 | { | |
201 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
202 | } |
203 | ||
204 | #define entity_is_task(se) 1 | |
205 | ||
b758149c PZ |
206 | #define for_each_sched_entity(se) \ |
207 | for (; se; se = NULL) | |
bf0f6f24 | 208 | |
b758149c | 209 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 210 | { |
b758149c | 211 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
212 | } |
213 | ||
b758149c PZ |
214 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
215 | { | |
216 | struct task_struct *p = task_of(se); | |
217 | struct rq *rq = task_rq(p); | |
218 | ||
219 | return &rq->cfs; | |
220 | } | |
221 | ||
222 | /* runqueue "owned" by this group */ | |
223 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
224 | { | |
225 | return NULL; | |
226 | } | |
227 | ||
228 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
229 | { | |
230 | return &cpu_rq(this_cpu)->cfs; | |
231 | } | |
232 | ||
233 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
234 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
235 | ||
236 | static inline int | |
237 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
238 | { | |
239 | return 1; | |
240 | } | |
241 | ||
242 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
243 | { | |
244 | return NULL; | |
245 | } | |
246 | ||
464b7527 PZ |
247 | static inline void |
248 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
249 | { | |
250 | } | |
251 | ||
b758149c PZ |
252 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
253 | ||
bf0f6f24 IM |
254 | |
255 | /************************************************************** | |
256 | * Scheduling class tree data structure manipulation methods: | |
257 | */ | |
258 | ||
0702e3eb | 259 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 260 | { |
368059a9 PZ |
261 | s64 delta = (s64)(vruntime - min_vruntime); |
262 | if (delta > 0) | |
02e0431a PZ |
263 | min_vruntime = vruntime; |
264 | ||
265 | return min_vruntime; | |
266 | } | |
267 | ||
0702e3eb | 268 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
269 | { |
270 | s64 delta = (s64)(vruntime - min_vruntime); | |
271 | if (delta < 0) | |
272 | min_vruntime = vruntime; | |
273 | ||
274 | return min_vruntime; | |
275 | } | |
276 | ||
54fdc581 FC |
277 | static inline int entity_before(struct sched_entity *a, |
278 | struct sched_entity *b) | |
279 | { | |
280 | return (s64)(a->vruntime - b->vruntime) < 0; | |
281 | } | |
282 | ||
0702e3eb | 283 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 284 | { |
30cfdcfc | 285 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
286 | } |
287 | ||
1af5f730 PZ |
288 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
289 | { | |
290 | u64 vruntime = cfs_rq->min_vruntime; | |
291 | ||
292 | if (cfs_rq->curr) | |
293 | vruntime = cfs_rq->curr->vruntime; | |
294 | ||
295 | if (cfs_rq->rb_leftmost) { | |
296 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
297 | struct sched_entity, | |
298 | run_node); | |
299 | ||
e17036da | 300 | if (!cfs_rq->curr) |
1af5f730 PZ |
301 | vruntime = se->vruntime; |
302 | else | |
303 | vruntime = min_vruntime(vruntime, se->vruntime); | |
304 | } | |
305 | ||
306 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
307 | } | |
308 | ||
bf0f6f24 IM |
309 | /* |
310 | * Enqueue an entity into the rb-tree: | |
311 | */ | |
0702e3eb | 312 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
313 | { |
314 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
315 | struct rb_node *parent = NULL; | |
316 | struct sched_entity *entry; | |
9014623c | 317 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
318 | int leftmost = 1; |
319 | ||
320 | /* | |
321 | * Find the right place in the rbtree: | |
322 | */ | |
323 | while (*link) { | |
324 | parent = *link; | |
325 | entry = rb_entry(parent, struct sched_entity, run_node); | |
326 | /* | |
327 | * We dont care about collisions. Nodes with | |
328 | * the same key stay together. | |
329 | */ | |
9014623c | 330 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
331 | link = &parent->rb_left; |
332 | } else { | |
333 | link = &parent->rb_right; | |
334 | leftmost = 0; | |
335 | } | |
336 | } | |
337 | ||
338 | /* | |
339 | * Maintain a cache of leftmost tree entries (it is frequently | |
340 | * used): | |
341 | */ | |
1af5f730 | 342 | if (leftmost) |
57cb499d | 343 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
344 | |
345 | rb_link_node(&se->run_node, parent, link); | |
346 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
347 | } |
348 | ||
0702e3eb | 349 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 350 | { |
3fe69747 PZ |
351 | if (cfs_rq->rb_leftmost == &se->run_node) { |
352 | struct rb_node *next_node; | |
3fe69747 PZ |
353 | |
354 | next_node = rb_next(&se->run_node); | |
355 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 356 | } |
e9acbff6 | 357 | |
bf0f6f24 | 358 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
359 | } |
360 | ||
bf0f6f24 IM |
361 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
362 | { | |
f4b6755f PZ |
363 | struct rb_node *left = cfs_rq->rb_leftmost; |
364 | ||
365 | if (!left) | |
366 | return NULL; | |
367 | ||
368 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
369 | } |
370 | ||
f4b6755f | 371 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 372 | { |
7eee3e67 | 373 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 374 | |
70eee74b BS |
375 | if (!last) |
376 | return NULL; | |
7eee3e67 IM |
377 | |
378 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
379 | } |
380 | ||
bf0f6f24 IM |
381 | /************************************************************** |
382 | * Scheduling class statistics methods: | |
383 | */ | |
384 | ||
b2be5e96 PZ |
385 | #ifdef CONFIG_SCHED_DEBUG |
386 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
8d65af78 | 387 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
388 | loff_t *ppos) |
389 | { | |
8d65af78 | 390 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
b2be5e96 PZ |
391 | |
392 | if (ret || !write) | |
393 | return ret; | |
394 | ||
395 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
396 | sysctl_sched_min_granularity); | |
397 | ||
398 | return 0; | |
399 | } | |
400 | #endif | |
647e7cac | 401 | |
a7be37ac | 402 | /* |
f9c0b095 | 403 | * delta /= w |
a7be37ac PZ |
404 | */ |
405 | static inline unsigned long | |
406 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
407 | { | |
f9c0b095 PZ |
408 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
409 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
410 | |
411 | return delta; | |
412 | } | |
413 | ||
647e7cac IM |
414 | /* |
415 | * The idea is to set a period in which each task runs once. | |
416 | * | |
417 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
418 | * this period because otherwise the slices get too small. | |
419 | * | |
420 | * p = (nr <= nl) ? l : l*nr/nl | |
421 | */ | |
4d78e7b6 PZ |
422 | static u64 __sched_period(unsigned long nr_running) |
423 | { | |
424 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 425 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
426 | |
427 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 428 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 429 | period *= nr_running; |
4d78e7b6 PZ |
430 | } |
431 | ||
432 | return period; | |
433 | } | |
434 | ||
647e7cac IM |
435 | /* |
436 | * We calculate the wall-time slice from the period by taking a part | |
437 | * proportional to the weight. | |
438 | * | |
f9c0b095 | 439 | * s = p*P[w/rw] |
647e7cac | 440 | */ |
6d0f0ebd | 441 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 442 | { |
0a582440 | 443 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 444 | |
0a582440 | 445 | for_each_sched_entity(se) { |
6272d68c | 446 | struct load_weight *load; |
3104bf03 | 447 | struct load_weight lw; |
6272d68c LM |
448 | |
449 | cfs_rq = cfs_rq_of(se); | |
450 | load = &cfs_rq->load; | |
f9c0b095 | 451 | |
0a582440 | 452 | if (unlikely(!se->on_rq)) { |
3104bf03 | 453 | lw = cfs_rq->load; |
0a582440 MG |
454 | |
455 | update_load_add(&lw, se->load.weight); | |
456 | load = &lw; | |
457 | } | |
458 | slice = calc_delta_mine(slice, se->load.weight, load); | |
459 | } | |
460 | return slice; | |
bf0f6f24 IM |
461 | } |
462 | ||
647e7cac | 463 | /* |
ac884dec | 464 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 465 | * |
f9c0b095 | 466 | * vs = s/w |
647e7cac | 467 | */ |
f9c0b095 | 468 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 469 | { |
f9c0b095 | 470 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
471 | } |
472 | ||
bf0f6f24 IM |
473 | /* |
474 | * Update the current task's runtime statistics. Skip current tasks that | |
475 | * are not in our scheduling class. | |
476 | */ | |
477 | static inline void | |
8ebc91d9 IM |
478 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
479 | unsigned long delta_exec) | |
bf0f6f24 | 480 | { |
bbdba7c0 | 481 | unsigned long delta_exec_weighted; |
bf0f6f24 | 482 | |
8179ca23 | 483 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
484 | |
485 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 486 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 487 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
e9acbff6 | 488 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 489 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
490 | } |
491 | ||
b7cc0896 | 492 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 493 | { |
429d43bc | 494 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 495 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
496 | unsigned long delta_exec; |
497 | ||
498 | if (unlikely(!curr)) | |
499 | return; | |
500 | ||
501 | /* | |
502 | * Get the amount of time the current task was running | |
503 | * since the last time we changed load (this cannot | |
504 | * overflow on 32 bits): | |
505 | */ | |
8ebc91d9 | 506 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
507 | if (!delta_exec) |
508 | return; | |
bf0f6f24 | 509 | |
8ebc91d9 IM |
510 | __update_curr(cfs_rq, curr, delta_exec); |
511 | curr->exec_start = now; | |
d842de87 SV |
512 | |
513 | if (entity_is_task(curr)) { | |
514 | struct task_struct *curtask = task_of(curr); | |
515 | ||
f977bb49 | 516 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 517 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 518 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 519 | } |
bf0f6f24 IM |
520 | } |
521 | ||
522 | static inline void | |
5870db5b | 523 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 524 | { |
d281918d | 525 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
526 | } |
527 | ||
bf0f6f24 IM |
528 | /* |
529 | * Task is being enqueued - update stats: | |
530 | */ | |
d2417e5a | 531 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 532 | { |
bf0f6f24 IM |
533 | /* |
534 | * Are we enqueueing a waiting task? (for current tasks | |
535 | * a dequeue/enqueue event is a NOP) | |
536 | */ | |
429d43bc | 537 | if (se != cfs_rq->curr) |
5870db5b | 538 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
539 | } |
540 | ||
bf0f6f24 | 541 | static void |
9ef0a961 | 542 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 543 | { |
bbdba7c0 IM |
544 | schedstat_set(se->wait_max, max(se->wait_max, |
545 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
546 | schedstat_set(se->wait_count, se->wait_count + 1); |
547 | schedstat_set(se->wait_sum, se->wait_sum + | |
548 | rq_of(cfs_rq)->clock - se->wait_start); | |
768d0c27 PZ |
549 | #ifdef CONFIG_SCHEDSTATS |
550 | if (entity_is_task(se)) { | |
551 | trace_sched_stat_wait(task_of(se), | |
552 | rq_of(cfs_rq)->clock - se->wait_start); | |
553 | } | |
554 | #endif | |
e1f84508 | 555 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
556 | } |
557 | ||
558 | static inline void | |
19b6a2e3 | 559 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 560 | { |
bf0f6f24 IM |
561 | /* |
562 | * Mark the end of the wait period if dequeueing a | |
563 | * waiting task: | |
564 | */ | |
429d43bc | 565 | if (se != cfs_rq->curr) |
9ef0a961 | 566 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
567 | } |
568 | ||
569 | /* | |
570 | * We are picking a new current task - update its stats: | |
571 | */ | |
572 | static inline void | |
79303e9e | 573 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
574 | { |
575 | /* | |
576 | * We are starting a new run period: | |
577 | */ | |
d281918d | 578 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
579 | } |
580 | ||
bf0f6f24 IM |
581 | /************************************************** |
582 | * Scheduling class queueing methods: | |
583 | */ | |
584 | ||
c09595f6 PZ |
585 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
586 | static void | |
587 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
588 | { | |
589 | cfs_rq->task_weight += weight; | |
590 | } | |
591 | #else | |
592 | static inline void | |
593 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
594 | { | |
595 | } | |
596 | #endif | |
597 | ||
30cfdcfc DA |
598 | static void |
599 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
600 | { | |
601 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
602 | if (!parent_entity(se)) |
603 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 604 | if (entity_is_task(se)) { |
c09595f6 | 605 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
606 | list_add(&se->group_node, &cfs_rq->tasks); |
607 | } | |
30cfdcfc DA |
608 | cfs_rq->nr_running++; |
609 | se->on_rq = 1; | |
610 | } | |
611 | ||
612 | static void | |
613 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
614 | { | |
615 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
616 | if (!parent_entity(se)) |
617 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 618 | if (entity_is_task(se)) { |
c09595f6 | 619 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
620 | list_del_init(&se->group_node); |
621 | } | |
30cfdcfc DA |
622 | cfs_rq->nr_running--; |
623 | se->on_rq = 0; | |
624 | } | |
625 | ||
2396af69 | 626 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 627 | { |
bf0f6f24 | 628 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
629 | struct task_struct *tsk = NULL; |
630 | ||
631 | if (entity_is_task(se)) | |
632 | tsk = task_of(se); | |
633 | ||
bf0f6f24 | 634 | if (se->sleep_start) { |
d281918d | 635 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
bf0f6f24 IM |
636 | |
637 | if ((s64)delta < 0) | |
638 | delta = 0; | |
639 | ||
640 | if (unlikely(delta > se->sleep_max)) | |
641 | se->sleep_max = delta; | |
642 | ||
643 | se->sleep_start = 0; | |
644 | se->sum_sleep_runtime += delta; | |
9745512c | 645 | |
768d0c27 | 646 | if (tsk) { |
e414314c | 647 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
648 | trace_sched_stat_sleep(tsk, delta); |
649 | } | |
bf0f6f24 IM |
650 | } |
651 | if (se->block_start) { | |
d281918d | 652 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
bf0f6f24 IM |
653 | |
654 | if ((s64)delta < 0) | |
655 | delta = 0; | |
656 | ||
657 | if (unlikely(delta > se->block_max)) | |
658 | se->block_max = delta; | |
659 | ||
660 | se->block_start = 0; | |
661 | se->sum_sleep_runtime += delta; | |
30084fbd | 662 | |
e414314c | 663 | if (tsk) { |
8f0dfc34 AV |
664 | if (tsk->in_iowait) { |
665 | se->iowait_sum += delta; | |
666 | se->iowait_count++; | |
768d0c27 | 667 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
668 | } |
669 | ||
e414314c PZ |
670 | /* |
671 | * Blocking time is in units of nanosecs, so shift by | |
672 | * 20 to get a milliseconds-range estimation of the | |
673 | * amount of time that the task spent sleeping: | |
674 | */ | |
675 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
676 | profile_hits(SLEEP_PROFILING, | |
677 | (void *)get_wchan(tsk), | |
678 | delta >> 20); | |
679 | } | |
680 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 681 | } |
bf0f6f24 IM |
682 | } |
683 | #endif | |
684 | } | |
685 | ||
ddc97297 PZ |
686 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
687 | { | |
688 | #ifdef CONFIG_SCHED_DEBUG | |
689 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
690 | ||
691 | if (d < 0) | |
692 | d = -d; | |
693 | ||
694 | if (d > 3*sysctl_sched_latency) | |
695 | schedstat_inc(cfs_rq, nr_spread_over); | |
696 | #endif | |
697 | } | |
698 | ||
aeb73b04 PZ |
699 | static void |
700 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
701 | { | |
1af5f730 | 702 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 703 | |
2cb8600e PZ |
704 | /* |
705 | * The 'current' period is already promised to the current tasks, | |
706 | * however the extra weight of the new task will slow them down a | |
707 | * little, place the new task so that it fits in the slot that | |
708 | * stays open at the end. | |
709 | */ | |
94dfb5e7 | 710 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 711 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 712 | |
a2e7a7eb MG |
713 | /* sleeps up to a single latency don't count. */ |
714 | if (!initial && sched_feat(FAIR_SLEEPERS)) { | |
715 | unsigned long thresh = sysctl_sched_latency; | |
a7be37ac | 716 | |
a2e7a7eb MG |
717 | /* |
718 | * Convert the sleeper threshold into virtual time. | |
719 | * SCHED_IDLE is a special sub-class. We care about | |
720 | * fairness only relative to other SCHED_IDLE tasks, | |
721 | * all of which have the same weight. | |
722 | */ | |
723 | if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) || | |
724 | task_of(se)->policy != SCHED_IDLE)) | |
725 | thresh = calc_delta_fair(thresh, se); | |
a7be37ac | 726 | |
a2e7a7eb MG |
727 | /* |
728 | * Halve their sleep time's effect, to allow | |
729 | * for a gentler effect of sleepers: | |
730 | */ | |
731 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
732 | thresh >>= 1; | |
51e0304c | 733 | |
a2e7a7eb | 734 | vruntime -= thresh; |
aeb73b04 PZ |
735 | } |
736 | ||
b5d9d734 MG |
737 | /* ensure we never gain time by being placed backwards. */ |
738 | vruntime = max_vruntime(se->vruntime, vruntime); | |
739 | ||
67e9fb2a | 740 | se->vruntime = vruntime; |
aeb73b04 PZ |
741 | } |
742 | ||
bf0f6f24 | 743 | static void |
83b699ed | 744 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
745 | { |
746 | /* | |
a2a2d680 | 747 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 748 | */ |
b7cc0896 | 749 | update_curr(cfs_rq); |
a992241d | 750 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 751 | |
e9acbff6 | 752 | if (wakeup) { |
aeb73b04 | 753 | place_entity(cfs_rq, se, 0); |
2396af69 | 754 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 755 | } |
bf0f6f24 | 756 | |
d2417e5a | 757 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 758 | check_spread(cfs_rq, se); |
83b699ed SV |
759 | if (se != cfs_rq->curr) |
760 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
761 | } |
762 | ||
a571bbea | 763 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 | 764 | { |
de69a80b | 765 | if (!se || cfs_rq->last == se) |
2002c695 PZ |
766 | cfs_rq->last = NULL; |
767 | ||
de69a80b | 768 | if (!se || cfs_rq->next == se) |
2002c695 PZ |
769 | cfs_rq->next = NULL; |
770 | } | |
771 | ||
a571bbea PZ |
772 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
773 | { | |
774 | for_each_sched_entity(se) | |
775 | __clear_buddies(cfs_rq_of(se), se); | |
776 | } | |
777 | ||
bf0f6f24 | 778 | static void |
525c2716 | 779 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 780 | { |
a2a2d680 DA |
781 | /* |
782 | * Update run-time statistics of the 'current'. | |
783 | */ | |
784 | update_curr(cfs_rq); | |
785 | ||
19b6a2e3 | 786 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 787 | if (sleep) { |
67e9fb2a | 788 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
789 | if (entity_is_task(se)) { |
790 | struct task_struct *tsk = task_of(se); | |
791 | ||
792 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 793 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 794 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 795 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 796 | } |
db36cc7d | 797 | #endif |
67e9fb2a PZ |
798 | } |
799 | ||
2002c695 | 800 | clear_buddies(cfs_rq, se); |
4793241b | 801 | |
83b699ed | 802 | if (se != cfs_rq->curr) |
30cfdcfc DA |
803 | __dequeue_entity(cfs_rq, se); |
804 | account_entity_dequeue(cfs_rq, se); | |
1af5f730 | 805 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
806 | } |
807 | ||
808 | /* | |
809 | * Preempt the current task with a newly woken task if needed: | |
810 | */ | |
7c92e54f | 811 | static void |
2e09bf55 | 812 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 813 | { |
11697830 PZ |
814 | unsigned long ideal_runtime, delta_exec; |
815 | ||
6d0f0ebd | 816 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 817 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 818 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 819 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
820 | /* |
821 | * The current task ran long enough, ensure it doesn't get | |
822 | * re-elected due to buddy favours. | |
823 | */ | |
824 | clear_buddies(cfs_rq, curr); | |
825 | } | |
bf0f6f24 IM |
826 | } |
827 | ||
83b699ed | 828 | static void |
8494f412 | 829 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 830 | { |
83b699ed SV |
831 | /* 'current' is not kept within the tree. */ |
832 | if (se->on_rq) { | |
833 | /* | |
834 | * Any task has to be enqueued before it get to execute on | |
835 | * a CPU. So account for the time it spent waiting on the | |
836 | * runqueue. | |
837 | */ | |
838 | update_stats_wait_end(cfs_rq, se); | |
839 | __dequeue_entity(cfs_rq, se); | |
840 | } | |
841 | ||
79303e9e | 842 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 843 | cfs_rq->curr = se; |
eba1ed4b IM |
844 | #ifdef CONFIG_SCHEDSTATS |
845 | /* | |
846 | * Track our maximum slice length, if the CPU's load is at | |
847 | * least twice that of our own weight (i.e. dont track it | |
848 | * when there are only lesser-weight tasks around): | |
849 | */ | |
495eca49 | 850 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
851 | se->slice_max = max(se->slice_max, |
852 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
853 | } | |
854 | #endif | |
4a55b450 | 855 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
856 | } |
857 | ||
3f3a4904 PZ |
858 | static int |
859 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
860 | ||
f4b6755f | 861 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 862 | { |
f4b6755f | 863 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
92f6a5e3 | 864 | struct sched_entity *buddy; |
f4b6755f | 865 | |
92f6a5e3 PZ |
866 | if (cfs_rq->next) { |
867 | buddy = cfs_rq->next; | |
868 | cfs_rq->next = NULL; | |
869 | if (wakeup_preempt_entity(buddy, se) < 1) | |
870 | return buddy; | |
871 | } | |
aa2ac252 | 872 | |
92f6a5e3 PZ |
873 | if (cfs_rq->last) { |
874 | buddy = cfs_rq->last; | |
875 | cfs_rq->last = NULL; | |
876 | if (wakeup_preempt_entity(buddy, se) < 1) | |
877 | return buddy; | |
878 | } | |
4793241b PZ |
879 | |
880 | return se; | |
aa2ac252 PZ |
881 | } |
882 | ||
ab6cde26 | 883 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
884 | { |
885 | /* | |
886 | * If still on the runqueue then deactivate_task() | |
887 | * was not called and update_curr() has to be done: | |
888 | */ | |
889 | if (prev->on_rq) | |
b7cc0896 | 890 | update_curr(cfs_rq); |
bf0f6f24 | 891 | |
ddc97297 | 892 | check_spread(cfs_rq, prev); |
30cfdcfc | 893 | if (prev->on_rq) { |
5870db5b | 894 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
895 | /* Put 'current' back into the tree. */ |
896 | __enqueue_entity(cfs_rq, prev); | |
897 | } | |
429d43bc | 898 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
899 | } |
900 | ||
8f4d37ec PZ |
901 | static void |
902 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 903 | { |
bf0f6f24 | 904 | /* |
30cfdcfc | 905 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 906 | */ |
30cfdcfc | 907 | update_curr(cfs_rq); |
bf0f6f24 | 908 | |
8f4d37ec PZ |
909 | #ifdef CONFIG_SCHED_HRTICK |
910 | /* | |
911 | * queued ticks are scheduled to match the slice, so don't bother | |
912 | * validating it and just reschedule. | |
913 | */ | |
983ed7a6 HH |
914 | if (queued) { |
915 | resched_task(rq_of(cfs_rq)->curr); | |
916 | return; | |
917 | } | |
8f4d37ec PZ |
918 | /* |
919 | * don't let the period tick interfere with the hrtick preemption | |
920 | */ | |
921 | if (!sched_feat(DOUBLE_TICK) && | |
922 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
923 | return; | |
924 | #endif | |
925 | ||
ce6c1311 | 926 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 927 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
928 | } |
929 | ||
930 | /************************************************** | |
931 | * CFS operations on tasks: | |
932 | */ | |
933 | ||
8f4d37ec PZ |
934 | #ifdef CONFIG_SCHED_HRTICK |
935 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
936 | { | |
8f4d37ec PZ |
937 | struct sched_entity *se = &p->se; |
938 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
939 | ||
940 | WARN_ON(task_rq(p) != rq); | |
941 | ||
942 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
943 | u64 slice = sched_slice(cfs_rq, se); | |
944 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
945 | s64 delta = slice - ran; | |
946 | ||
947 | if (delta < 0) { | |
948 | if (rq->curr == p) | |
949 | resched_task(p); | |
950 | return; | |
951 | } | |
952 | ||
953 | /* | |
954 | * Don't schedule slices shorter than 10000ns, that just | |
955 | * doesn't make sense. Rely on vruntime for fairness. | |
956 | */ | |
31656519 | 957 | if (rq->curr != p) |
157124c1 | 958 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 959 | |
31656519 | 960 | hrtick_start(rq, delta); |
8f4d37ec PZ |
961 | } |
962 | } | |
a4c2f00f PZ |
963 | |
964 | /* | |
965 | * called from enqueue/dequeue and updates the hrtick when the | |
966 | * current task is from our class and nr_running is low enough | |
967 | * to matter. | |
968 | */ | |
969 | static void hrtick_update(struct rq *rq) | |
970 | { | |
971 | struct task_struct *curr = rq->curr; | |
972 | ||
973 | if (curr->sched_class != &fair_sched_class) | |
974 | return; | |
975 | ||
976 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
977 | hrtick_start_fair(rq, curr); | |
978 | } | |
55e12e5e | 979 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
980 | static inline void |
981 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
982 | { | |
983 | } | |
a4c2f00f PZ |
984 | |
985 | static inline void hrtick_update(struct rq *rq) | |
986 | { | |
987 | } | |
8f4d37ec PZ |
988 | #endif |
989 | ||
bf0f6f24 IM |
990 | /* |
991 | * The enqueue_task method is called before nr_running is | |
992 | * increased. Here we update the fair scheduling stats and | |
993 | * then put the task into the rbtree: | |
994 | */ | |
fd390f6a | 995 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
996 | { |
997 | struct cfs_rq *cfs_rq; | |
62fb1851 | 998 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
999 | |
1000 | for_each_sched_entity(se) { | |
62fb1851 | 1001 | if (se->on_rq) |
bf0f6f24 IM |
1002 | break; |
1003 | cfs_rq = cfs_rq_of(se); | |
83b699ed | 1004 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 1005 | wakeup = 1; |
bf0f6f24 | 1006 | } |
8f4d37ec | 1007 | |
a4c2f00f | 1008 | hrtick_update(rq); |
bf0f6f24 IM |
1009 | } |
1010 | ||
1011 | /* | |
1012 | * The dequeue_task method is called before nr_running is | |
1013 | * decreased. We remove the task from the rbtree and | |
1014 | * update the fair scheduling stats: | |
1015 | */ | |
f02231e5 | 1016 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
1017 | { |
1018 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1019 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1020 | |
1021 | for_each_sched_entity(se) { | |
1022 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 1023 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 1024 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 1025 | if (cfs_rq->load.weight) |
bf0f6f24 | 1026 | break; |
b9fa3df3 | 1027 | sleep = 1; |
bf0f6f24 | 1028 | } |
8f4d37ec | 1029 | |
a4c2f00f | 1030 | hrtick_update(rq); |
bf0f6f24 IM |
1031 | } |
1032 | ||
1033 | /* | |
1799e35d IM |
1034 | * sched_yield() support is very simple - we dequeue and enqueue. |
1035 | * | |
1036 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 1037 | */ |
4530d7ab | 1038 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 1039 | { |
db292ca3 IM |
1040 | struct task_struct *curr = rq->curr; |
1041 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1042 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
1043 | |
1044 | /* | |
1799e35d IM |
1045 | * Are we the only task in the tree? |
1046 | */ | |
1047 | if (unlikely(cfs_rq->nr_running == 1)) | |
1048 | return; | |
1049 | ||
2002c695 PZ |
1050 | clear_buddies(cfs_rq, se); |
1051 | ||
db292ca3 | 1052 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1053 | update_rq_clock(rq); |
1799e35d | 1054 | /* |
a2a2d680 | 1055 | * Update run-time statistics of the 'current'. |
1799e35d | 1056 | */ |
2b1e315d | 1057 | update_curr(cfs_rq); |
1799e35d IM |
1058 | |
1059 | return; | |
1060 | } | |
1061 | /* | |
1062 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1063 | */ |
2b1e315d | 1064 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1065 | /* |
1066 | * Already in the rightmost position? | |
1067 | */ | |
54fdc581 | 1068 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
1799e35d IM |
1069 | return; |
1070 | ||
1071 | /* | |
1072 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1073 | * Upon rescheduling, sched_class::put_prev_task() will place |
1074 | * 'current' within the tree based on its new key value. | |
1799e35d | 1075 | */ |
30cfdcfc | 1076 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1077 | } |
1078 | ||
e7693a36 | 1079 | #ifdef CONFIG_SMP |
098fb9db | 1080 | |
bb3469ac | 1081 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1082 | /* |
1083 | * effective_load() calculates the load change as seen from the root_task_group | |
1084 | * | |
1085 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1086 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1087 | * can calculate the shift in shares. | |
1088 | * | |
1089 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1090 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1091 | * this change. | |
1092 | * | |
1093 | * We compensate this by not only taking the current delta into account, but | |
1094 | * also considering the delta between when the shares were last adjusted and | |
1095 | * now. | |
1096 | * | |
1097 | * We still saw a performance dip, some tracing learned us that between | |
1098 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1099 | * significantly. Therefore try to bias the error in direction of failing | |
1100 | * the affine wakeup. | |
1101 | * | |
1102 | */ | |
f1d239f7 PZ |
1103 | static long effective_load(struct task_group *tg, int cpu, |
1104 | long wl, long wg) | |
bb3469ac | 1105 | { |
4be9daaa | 1106 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1107 | |
1108 | if (!tg->parent) | |
1109 | return wl; | |
1110 | ||
f5bfb7d9 PZ |
1111 | /* |
1112 | * By not taking the decrease of shares on the other cpu into | |
1113 | * account our error leans towards reducing the affine wakeups. | |
1114 | */ | |
1115 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1116 | return wl; | |
1117 | ||
4be9daaa | 1118 | for_each_sched_entity(se) { |
cb5ef42a | 1119 | long S, rw, s, a, b; |
940959e9 PZ |
1120 | long more_w; |
1121 | ||
1122 | /* | |
1123 | * Instead of using this increment, also add the difference | |
1124 | * between when the shares were last updated and now. | |
1125 | */ | |
1126 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1127 | wl += more_w; | |
1128 | wg += more_w; | |
4be9daaa PZ |
1129 | |
1130 | S = se->my_q->tg->shares; | |
1131 | s = se->my_q->shares; | |
f1d239f7 | 1132 | rw = se->my_q->rq_weight; |
bb3469ac | 1133 | |
cb5ef42a PZ |
1134 | a = S*(rw + wl); |
1135 | b = S*rw + s*wg; | |
4be9daaa | 1136 | |
940959e9 PZ |
1137 | wl = s*(a-b); |
1138 | ||
1139 | if (likely(b)) | |
1140 | wl /= b; | |
1141 | ||
83378269 PZ |
1142 | /* |
1143 | * Assume the group is already running and will | |
1144 | * thus already be accounted for in the weight. | |
1145 | * | |
1146 | * That is, moving shares between CPUs, does not | |
1147 | * alter the group weight. | |
1148 | */ | |
4be9daaa | 1149 | wg = 0; |
4be9daaa | 1150 | } |
bb3469ac | 1151 | |
4be9daaa | 1152 | return wl; |
bb3469ac | 1153 | } |
4be9daaa | 1154 | |
bb3469ac | 1155 | #else |
4be9daaa | 1156 | |
83378269 PZ |
1157 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1158 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1159 | { |
83378269 | 1160 | return wl; |
bb3469ac | 1161 | } |
4be9daaa | 1162 | |
bb3469ac PZ |
1163 | #endif |
1164 | ||
c88d5910 | 1165 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1166 | { |
c88d5910 PZ |
1167 | struct task_struct *curr = current; |
1168 | unsigned long this_load, load; | |
1169 | int idx, this_cpu, prev_cpu; | |
098fb9db | 1170 | unsigned long tl_per_task; |
c88d5910 PZ |
1171 | unsigned int imbalance; |
1172 | struct task_group *tg; | |
83378269 | 1173 | unsigned long weight; |
b3137bc8 | 1174 | int balanced; |
098fb9db | 1175 | |
c88d5910 PZ |
1176 | idx = sd->wake_idx; |
1177 | this_cpu = smp_processor_id(); | |
1178 | prev_cpu = task_cpu(p); | |
1179 | load = source_load(prev_cpu, idx); | |
1180 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1181 | |
e69b0f1b PZ |
1182 | if (sync) { |
1183 | if (sched_feat(SYNC_LESS) && | |
1184 | (curr->se.avg_overlap > sysctl_sched_migration_cost || | |
1185 | p->se.avg_overlap > sysctl_sched_migration_cost)) | |
1186 | sync = 0; | |
1187 | } else { | |
1188 | if (sched_feat(SYNC_MORE) && | |
1189 | (curr->se.avg_overlap < sysctl_sched_migration_cost && | |
1190 | p->se.avg_overlap < sysctl_sched_migration_cost)) | |
1191 | sync = 1; | |
1192 | } | |
fc631c82 | 1193 | |
b3137bc8 MG |
1194 | /* |
1195 | * If sync wakeup then subtract the (maximum possible) | |
1196 | * effect of the currently running task from the load | |
1197 | * of the current CPU: | |
1198 | */ | |
83378269 PZ |
1199 | if (sync) { |
1200 | tg = task_group(current); | |
1201 | weight = current->se.load.weight; | |
1202 | ||
c88d5910 | 1203 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1204 | load += effective_load(tg, prev_cpu, 0, -weight); |
1205 | } | |
b3137bc8 | 1206 | |
83378269 PZ |
1207 | tg = task_group(p); |
1208 | weight = p->se.load.weight; | |
b3137bc8 | 1209 | |
c88d5910 PZ |
1210 | imbalance = 100 + (sd->imbalance_pct - 100) / 2; |
1211 | ||
71a29aa7 PZ |
1212 | /* |
1213 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1214 | * due to the sync cause above having dropped this_load to 0, we'll |
1215 | * always have an imbalance, but there's really nothing you can do | |
1216 | * about that, so that's good too. | |
71a29aa7 PZ |
1217 | * |
1218 | * Otherwise check if either cpus are near enough in load to allow this | |
1219 | * task to be woken on this_cpu. | |
1220 | */ | |
c88d5910 PZ |
1221 | balanced = !this_load || |
1222 | 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <= | |
83378269 | 1223 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); |
b3137bc8 | 1224 | |
098fb9db | 1225 | /* |
4ae7d5ce IM |
1226 | * If the currently running task will sleep within |
1227 | * a reasonable amount of time then attract this newly | |
1228 | * woken task: | |
098fb9db | 1229 | */ |
2fb7635c PZ |
1230 | if (sync && balanced) |
1231 | return 1; | |
098fb9db IM |
1232 | |
1233 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1234 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1235 | ||
c88d5910 PZ |
1236 | if (balanced || |
1237 | (this_load <= load && | |
1238 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1239 | /* |
1240 | * This domain has SD_WAKE_AFFINE and | |
1241 | * p is cache cold in this domain, and | |
1242 | * there is no bad imbalance. | |
1243 | */ | |
c88d5910 | 1244 | schedstat_inc(sd, ttwu_move_affine); |
098fb9db IM |
1245 | schedstat_inc(p, se.nr_wakeups_affine); |
1246 | ||
1247 | return 1; | |
1248 | } | |
1249 | return 0; | |
1250 | } | |
1251 | ||
aaee1203 PZ |
1252 | /* |
1253 | * find_idlest_group finds and returns the least busy CPU group within the | |
1254 | * domain. | |
1255 | */ | |
1256 | static struct sched_group * | |
78e7ed53 | 1257 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1258 | int this_cpu, int load_idx) |
e7693a36 | 1259 | { |
aaee1203 PZ |
1260 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; |
1261 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
aaee1203 | 1262 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1263 | |
aaee1203 PZ |
1264 | do { |
1265 | unsigned long load, avg_load; | |
1266 | int local_group; | |
1267 | int i; | |
e7693a36 | 1268 | |
aaee1203 PZ |
1269 | /* Skip over this group if it has no CPUs allowed */ |
1270 | if (!cpumask_intersects(sched_group_cpus(group), | |
1271 | &p->cpus_allowed)) | |
1272 | continue; | |
1273 | ||
1274 | local_group = cpumask_test_cpu(this_cpu, | |
1275 | sched_group_cpus(group)); | |
1276 | ||
1277 | /* Tally up the load of all CPUs in the group */ | |
1278 | avg_load = 0; | |
1279 | ||
1280 | for_each_cpu(i, sched_group_cpus(group)) { | |
1281 | /* Bias balancing toward cpus of our domain */ | |
1282 | if (local_group) | |
1283 | load = source_load(i, load_idx); | |
1284 | else | |
1285 | load = target_load(i, load_idx); | |
1286 | ||
1287 | avg_load += load; | |
1288 | } | |
1289 | ||
1290 | /* Adjust by relative CPU power of the group */ | |
1291 | avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power; | |
1292 | ||
1293 | if (local_group) { | |
1294 | this_load = avg_load; | |
1295 | this = group; | |
1296 | } else if (avg_load < min_load) { | |
1297 | min_load = avg_load; | |
1298 | idlest = group; | |
1299 | } | |
1300 | } while (group = group->next, group != sd->groups); | |
1301 | ||
1302 | if (!idlest || 100*this_load < imbalance*min_load) | |
1303 | return NULL; | |
1304 | return idlest; | |
1305 | } | |
1306 | ||
1307 | /* | |
1308 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1309 | */ | |
1310 | static int | |
1311 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1312 | { | |
1313 | unsigned long load, min_load = ULONG_MAX; | |
1314 | int idlest = -1; | |
1315 | int i; | |
1316 | ||
1317 | /* Traverse only the allowed CPUs */ | |
1318 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1319 | load = weighted_cpuload(i); | |
1320 | ||
1321 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1322 | min_load = load; | |
1323 | idlest = i; | |
e7693a36 GH |
1324 | } |
1325 | } | |
1326 | ||
aaee1203 PZ |
1327 | return idlest; |
1328 | } | |
e7693a36 | 1329 | |
aaee1203 PZ |
1330 | /* |
1331 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1332 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1333 | * SD_BALANCE_EXEC. | |
1334 | * | |
1335 | * Balance, ie. select the least loaded group. | |
1336 | * | |
1337 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1338 | * | |
1339 | * preempt must be disabled. | |
1340 | */ | |
5158f4e4 | 1341 | static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 1342 | { |
29cd8bae | 1343 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1344 | int cpu = smp_processor_id(); |
1345 | int prev_cpu = task_cpu(p); | |
1346 | int new_cpu = cpu; | |
1347 | int want_affine = 0; | |
29cd8bae | 1348 | int want_sd = 1; |
5158f4e4 | 1349 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1350 | |
0763a660 | 1351 | if (sd_flag & SD_BALANCE_WAKE) { |
3f04e8cd MG |
1352 | if (sched_feat(AFFINE_WAKEUPS) && |
1353 | cpumask_test_cpu(cpu, &p->cpus_allowed)) | |
c88d5910 PZ |
1354 | want_affine = 1; |
1355 | new_cpu = prev_cpu; | |
1356 | } | |
aaee1203 | 1357 | |
83f54960 | 1358 | rcu_read_lock(); |
aaee1203 PZ |
1359 | for_each_domain(cpu, tmp) { |
1360 | /* | |
ae154be1 PZ |
1361 | * If power savings logic is enabled for a domain, see if we |
1362 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1363 | */ |
59abf026 | 1364 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1365 | unsigned long power = 0; |
1366 | unsigned long nr_running = 0; | |
1367 | unsigned long capacity; | |
1368 | int i; | |
1369 | ||
1370 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1371 | power += power_of(i); | |
1372 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1373 | } | |
1374 | ||
1375 | capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE); | |
1376 | ||
59abf026 PZ |
1377 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1378 | nr_running /= 2; | |
1379 | ||
1380 | if (nr_running < capacity) | |
29cd8bae | 1381 | want_sd = 0; |
ae154be1 | 1382 | } |
aaee1203 | 1383 | |
c88d5910 PZ |
1384 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1385 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1386 | ||
29cd8bae | 1387 | affine_sd = tmp; |
c88d5910 PZ |
1388 | want_affine = 0; |
1389 | } | |
1390 | ||
29cd8bae PZ |
1391 | if (!want_sd && !want_affine) |
1392 | break; | |
1393 | ||
0763a660 | 1394 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1395 | continue; |
1396 | ||
29cd8bae PZ |
1397 | if (want_sd) |
1398 | sd = tmp; | |
1399 | } | |
1400 | ||
1401 | if (sched_feat(LB_SHARES_UPDATE)) { | |
1402 | /* | |
1403 | * Pick the largest domain to update shares over | |
1404 | */ | |
1405 | tmp = sd; | |
1406 | if (affine_sd && (!tmp || | |
1407 | cpumask_weight(sched_domain_span(affine_sd)) > | |
1408 | cpumask_weight(sched_domain_span(sd)))) | |
1409 | tmp = affine_sd; | |
1410 | ||
1411 | if (tmp) | |
1412 | update_shares(tmp); | |
c88d5910 | 1413 | } |
aaee1203 | 1414 | |
29cd8bae PZ |
1415 | if (affine_sd && wake_affine(affine_sd, p, sync)) { |
1416 | new_cpu = cpu; | |
f4827386 | 1417 | goto out; |
29cd8bae | 1418 | } |
e7693a36 | 1419 | |
aaee1203 | 1420 | while (sd) { |
5158f4e4 | 1421 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1422 | struct sched_group *group; |
c88d5910 | 1423 | int weight; |
098fb9db | 1424 | |
0763a660 | 1425 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1426 | sd = sd->child; |
1427 | continue; | |
1428 | } | |
098fb9db | 1429 | |
5158f4e4 PZ |
1430 | if (sd_flag & SD_BALANCE_WAKE) |
1431 | load_idx = sd->wake_idx; | |
098fb9db | 1432 | |
5158f4e4 | 1433 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1434 | if (!group) { |
1435 | sd = sd->child; | |
1436 | continue; | |
1437 | } | |
4ae7d5ce | 1438 | |
d7c33c49 | 1439 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1440 | if (new_cpu == -1 || new_cpu == cpu) { |
1441 | /* Now try balancing at a lower domain level of cpu */ | |
1442 | sd = sd->child; | |
1443 | continue; | |
e7693a36 | 1444 | } |
aaee1203 PZ |
1445 | |
1446 | /* Now try balancing at a lower domain level of new_cpu */ | |
1447 | cpu = new_cpu; | |
1448 | weight = cpumask_weight(sched_domain_span(sd)); | |
1449 | sd = NULL; | |
1450 | for_each_domain(cpu, tmp) { | |
1451 | if (weight <= cpumask_weight(sched_domain_span(tmp))) | |
1452 | break; | |
0763a660 | 1453 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1454 | sd = tmp; |
1455 | } | |
1456 | /* while loop will break here if sd == NULL */ | |
e7693a36 GH |
1457 | } |
1458 | ||
f4827386 | 1459 | out: |
83f54960 | 1460 | rcu_read_unlock(); |
c88d5910 | 1461 | return new_cpu; |
e7693a36 GH |
1462 | } |
1463 | #endif /* CONFIG_SMP */ | |
1464 | ||
e52fb7c0 PZ |
1465 | /* |
1466 | * Adaptive granularity | |
1467 | * | |
1468 | * se->avg_wakeup gives the average time a task runs until it does a wakeup, | |
1469 | * with the limit of wakeup_gran -- when it never does a wakeup. | |
1470 | * | |
1471 | * So the smaller avg_wakeup is the faster we want this task to preempt, | |
1472 | * but we don't want to treat the preemptee unfairly and therefore allow it | |
1473 | * to run for at least the amount of time we'd like to run. | |
1474 | * | |
1475 | * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one | |
1476 | * | |
1477 | * NOTE: we use *nr_running to scale with load, this nicely matches the | |
1478 | * degrading latency on load. | |
1479 | */ | |
1480 | static unsigned long | |
1481 | adaptive_gran(struct sched_entity *curr, struct sched_entity *se) | |
1482 | { | |
1483 | u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
1484 | u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running; | |
1485 | u64 gran = 0; | |
1486 | ||
1487 | if (this_run < expected_wakeup) | |
1488 | gran = expected_wakeup - this_run; | |
1489 | ||
1490 | return min_t(s64, gran, sysctl_sched_wakeup_granularity); | |
1491 | } | |
1492 | ||
1493 | static unsigned long | |
1494 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1495 | { |
1496 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1497 | ||
e52fb7c0 PZ |
1498 | if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN)) |
1499 | gran = adaptive_gran(curr, se); | |
1500 | ||
0bbd3336 | 1501 | /* |
e52fb7c0 PZ |
1502 | * Since its curr running now, convert the gran from real-time |
1503 | * to virtual-time in his units. | |
0bbd3336 | 1504 | */ |
e52fb7c0 PZ |
1505 | if (sched_feat(ASYM_GRAN)) { |
1506 | /* | |
1507 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1508 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1509 | * the resulting gran will be larger, therefore penalizing the | |
1510 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1511 | * be smaller, again penalizing the lighter task. | |
1512 | * | |
1513 | * This is especially important for buddies when the leftmost | |
1514 | * task is higher priority than the buddy. | |
1515 | */ | |
1516 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1517 | gran = calc_delta_fair(gran, se); | |
1518 | } else { | |
1519 | if (unlikely(curr->load.weight != NICE_0_LOAD)) | |
1520 | gran = calc_delta_fair(gran, curr); | |
1521 | } | |
0bbd3336 PZ |
1522 | |
1523 | return gran; | |
1524 | } | |
1525 | ||
464b7527 PZ |
1526 | /* |
1527 | * Should 'se' preempt 'curr'. | |
1528 | * | |
1529 | * |s1 | |
1530 | * |s2 | |
1531 | * |s3 | |
1532 | * g | |
1533 | * |<--->|c | |
1534 | * | |
1535 | * w(c, s1) = -1 | |
1536 | * w(c, s2) = 0 | |
1537 | * w(c, s3) = 1 | |
1538 | * | |
1539 | */ | |
1540 | static int | |
1541 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1542 | { | |
1543 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1544 | ||
1545 | if (vdiff <= 0) | |
1546 | return -1; | |
1547 | ||
e52fb7c0 | 1548 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1549 | if (vdiff > gran) |
1550 | return 1; | |
1551 | ||
1552 | return 0; | |
1553 | } | |
1554 | ||
02479099 PZ |
1555 | static void set_last_buddy(struct sched_entity *se) |
1556 | { | |
6bc912b7 PZ |
1557 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1558 | for_each_sched_entity(se) | |
1559 | cfs_rq_of(se)->last = se; | |
1560 | } | |
02479099 PZ |
1561 | } |
1562 | ||
1563 | static void set_next_buddy(struct sched_entity *se) | |
1564 | { | |
6bc912b7 PZ |
1565 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1566 | for_each_sched_entity(se) | |
1567 | cfs_rq_of(se)->next = se; | |
1568 | } | |
02479099 PZ |
1569 | } |
1570 | ||
bf0f6f24 IM |
1571 | /* |
1572 | * Preempt the current task with a newly woken task if needed: | |
1573 | */ | |
5a9b86f6 | 1574 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1575 | { |
1576 | struct task_struct *curr = rq->curr; | |
8651a86c | 1577 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1578 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
5a9b86f6 | 1579 | int sync = wake_flags & WF_SYNC; |
bf0f6f24 | 1580 | |
03e89e45 | 1581 | update_curr(cfs_rq); |
4793241b | 1582 | |
03e89e45 | 1583 | if (unlikely(rt_prio(p->prio))) { |
bf0f6f24 IM |
1584 | resched_task(curr); |
1585 | return; | |
1586 | } | |
aa2ac252 | 1587 | |
d95f98d0 PZ |
1588 | if (unlikely(p->sched_class != &fair_sched_class)) |
1589 | return; | |
1590 | ||
4ae7d5ce IM |
1591 | if (unlikely(se == pse)) |
1592 | return; | |
1593 | ||
4793241b PZ |
1594 | /* |
1595 | * Only set the backward buddy when the current task is still on the | |
1596 | * rq. This can happen when a wakeup gets interleaved with schedule on | |
1597 | * the ->pre_schedule() or idle_balance() point, either of which can | |
1598 | * drop the rq lock. | |
1599 | * | |
1600 | * Also, during early boot the idle thread is in the fair class, for | |
1601 | * obvious reasons its a bad idea to schedule back to the idle thread. | |
1602 | */ | |
1603 | if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle)) | |
02479099 | 1604 | set_last_buddy(se); |
5a9b86f6 | 1605 | if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) |
3cb63d52 | 1606 | set_next_buddy(pse); |
57fdc26d | 1607 | |
aec0a514 BR |
1608 | /* |
1609 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1610 | * wake up path. | |
1611 | */ | |
1612 | if (test_tsk_need_resched(curr)) | |
1613 | return; | |
1614 | ||
91c234b4 | 1615 | /* |
6bc912b7 | 1616 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1617 | * the tick): |
1618 | */ | |
6bc912b7 | 1619 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1620 | return; |
bf0f6f24 | 1621 | |
6bc912b7 PZ |
1622 | /* Idle tasks are by definition preempted by everybody. */ |
1623 | if (unlikely(curr->policy == SCHED_IDLE)) { | |
1624 | resched_task(curr); | |
91c234b4 | 1625 | return; |
6bc912b7 | 1626 | } |
bf0f6f24 | 1627 | |
e6b1b2c9 PZ |
1628 | if ((sched_feat(WAKEUP_SYNC) && sync) || |
1629 | (sched_feat(WAKEUP_OVERLAP) && | |
1630 | (se->avg_overlap < sysctl_sched_migration_cost && | |
1631 | pse->avg_overlap < sysctl_sched_migration_cost))) { | |
15afe09b PZ |
1632 | resched_task(curr); |
1633 | return; | |
1634 | } | |
1635 | ||
ad4b78bb PZ |
1636 | if (sched_feat(WAKEUP_RUNNING)) { |
1637 | if (pse->avg_running < se->avg_running) { | |
1638 | set_next_buddy(pse); | |
1639 | resched_task(curr); | |
1640 | return; | |
1641 | } | |
1642 | } | |
1643 | ||
1644 | if (!sched_feat(WAKEUP_PREEMPT)) | |
1645 | return; | |
1646 | ||
464b7527 PZ |
1647 | find_matching_se(&se, &pse); |
1648 | ||
002f128b | 1649 | BUG_ON(!pse); |
464b7527 | 1650 | |
002f128b PT |
1651 | if (wakeup_preempt_entity(se, pse) == 1) |
1652 | resched_task(curr); | |
bf0f6f24 IM |
1653 | } |
1654 | ||
fb8d4724 | 1655 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1656 | { |
8f4d37ec | 1657 | struct task_struct *p; |
bf0f6f24 IM |
1658 | struct cfs_rq *cfs_rq = &rq->cfs; |
1659 | struct sched_entity *se; | |
1660 | ||
1661 | if (unlikely(!cfs_rq->nr_running)) | |
1662 | return NULL; | |
1663 | ||
1664 | do { | |
9948f4b2 | 1665 | se = pick_next_entity(cfs_rq); |
f4b6755f | 1666 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1667 | cfs_rq = group_cfs_rq(se); |
1668 | } while (cfs_rq); | |
1669 | ||
8f4d37ec PZ |
1670 | p = task_of(se); |
1671 | hrtick_start_fair(rq, p); | |
1672 | ||
1673 | return p; | |
bf0f6f24 IM |
1674 | } |
1675 | ||
1676 | /* | |
1677 | * Account for a descheduled task: | |
1678 | */ | |
31ee529c | 1679 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1680 | { |
1681 | struct sched_entity *se = &prev->se; | |
1682 | struct cfs_rq *cfs_rq; | |
1683 | ||
1684 | for_each_sched_entity(se) { | |
1685 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1686 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1687 | } |
1688 | } | |
1689 | ||
681f3e68 | 1690 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1691 | /************************************************** |
1692 | * Fair scheduling class load-balancing methods: | |
1693 | */ | |
1694 | ||
1695 | /* | |
1696 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1697 | * during the whole iteration, the current task might be | |
1698 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1699 | * achieve that by always pre-iterating before returning | |
1700 | * the current task: | |
1701 | */ | |
a9957449 | 1702 | static struct task_struct * |
4a55bd5e | 1703 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) |
bf0f6f24 | 1704 | { |
354d60c2 DG |
1705 | struct task_struct *p = NULL; |
1706 | struct sched_entity *se; | |
bf0f6f24 | 1707 | |
77ae6513 MG |
1708 | if (next == &cfs_rq->tasks) |
1709 | return NULL; | |
1710 | ||
b87f1724 BR |
1711 | se = list_entry(next, struct sched_entity, group_node); |
1712 | p = task_of(se); | |
1713 | cfs_rq->balance_iterator = next->next; | |
77ae6513 | 1714 | |
bf0f6f24 IM |
1715 | return p; |
1716 | } | |
1717 | ||
1718 | static struct task_struct *load_balance_start_fair(void *arg) | |
1719 | { | |
1720 | struct cfs_rq *cfs_rq = arg; | |
1721 | ||
4a55bd5e | 1722 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); |
bf0f6f24 IM |
1723 | } |
1724 | ||
1725 | static struct task_struct *load_balance_next_fair(void *arg) | |
1726 | { | |
1727 | struct cfs_rq *cfs_rq = arg; | |
1728 | ||
4a55bd5e | 1729 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); |
bf0f6f24 IM |
1730 | } |
1731 | ||
c09595f6 PZ |
1732 | static unsigned long |
1733 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1734 | unsigned long max_load_move, struct sched_domain *sd, | |
1735 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, | |
1736 | struct cfs_rq *cfs_rq) | |
62fb1851 | 1737 | { |
c09595f6 | 1738 | struct rq_iterator cfs_rq_iterator; |
62fb1851 | 1739 | |
c09595f6 PZ |
1740 | cfs_rq_iterator.start = load_balance_start_fair; |
1741 | cfs_rq_iterator.next = load_balance_next_fair; | |
1742 | cfs_rq_iterator.arg = cfs_rq; | |
62fb1851 | 1743 | |
c09595f6 PZ |
1744 | return balance_tasks(this_rq, this_cpu, busiest, |
1745 | max_load_move, sd, idle, all_pinned, | |
1746 | this_best_prio, &cfs_rq_iterator); | |
62fb1851 | 1747 | } |
62fb1851 | 1748 | |
c09595f6 | 1749 | #ifdef CONFIG_FAIR_GROUP_SCHED |
43010659 | 1750 | static unsigned long |
bf0f6f24 | 1751 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1752 | unsigned long max_load_move, |
a4ac01c3 PW |
1753 | struct sched_domain *sd, enum cpu_idle_type idle, |
1754 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 | 1755 | { |
bf0f6f24 | 1756 | long rem_load_move = max_load_move; |
c09595f6 PZ |
1757 | int busiest_cpu = cpu_of(busiest); |
1758 | struct task_group *tg; | |
18d95a28 | 1759 | |
c09595f6 | 1760 | rcu_read_lock(); |
c8cba857 | 1761 | update_h_load(busiest_cpu); |
18d95a28 | 1762 | |
caea8a03 | 1763 | list_for_each_entry_rcu(tg, &task_groups, list) { |
c8cba857 | 1764 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; |
42a3ac7d PZ |
1765 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
1766 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
243e0e7b | 1767 | u64 rem_load, moved_load; |
18d95a28 | 1768 | |
c09595f6 PZ |
1769 | /* |
1770 | * empty group | |
1771 | */ | |
c8cba857 | 1772 | if (!busiest_cfs_rq->task_weight) |
bf0f6f24 IM |
1773 | continue; |
1774 | ||
243e0e7b SV |
1775 | rem_load = (u64)rem_load_move * busiest_weight; |
1776 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
bf0f6f24 | 1777 | |
c09595f6 | 1778 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, |
53fecd8a | 1779 | rem_load, sd, idle, all_pinned, this_best_prio, |
c09595f6 | 1780 | tg->cfs_rq[busiest_cpu]); |
bf0f6f24 | 1781 | |
c09595f6 | 1782 | if (!moved_load) |
bf0f6f24 IM |
1783 | continue; |
1784 | ||
42a3ac7d | 1785 | moved_load *= busiest_h_load; |
243e0e7b | 1786 | moved_load = div_u64(moved_load, busiest_weight + 1); |
bf0f6f24 | 1787 | |
c09595f6 PZ |
1788 | rem_load_move -= moved_load; |
1789 | if (rem_load_move < 0) | |
bf0f6f24 IM |
1790 | break; |
1791 | } | |
c09595f6 | 1792 | rcu_read_unlock(); |
bf0f6f24 | 1793 | |
43010659 | 1794 | return max_load_move - rem_load_move; |
bf0f6f24 | 1795 | } |
c09595f6 PZ |
1796 | #else |
1797 | static unsigned long | |
1798 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1799 | unsigned long max_load_move, | |
1800 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1801 | int *all_pinned, int *this_best_prio) | |
1802 | { | |
1803 | return __load_balance_fair(this_rq, this_cpu, busiest, | |
1804 | max_load_move, sd, idle, all_pinned, | |
1805 | this_best_prio, &busiest->cfs); | |
1806 | } | |
1807 | #endif | |
bf0f6f24 | 1808 | |
e1d1484f PW |
1809 | static int |
1810 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1811 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1812 | { | |
1813 | struct cfs_rq *busy_cfs_rq; | |
1814 | struct rq_iterator cfs_rq_iterator; | |
1815 | ||
1816 | cfs_rq_iterator.start = load_balance_start_fair; | |
1817 | cfs_rq_iterator.next = load_balance_next_fair; | |
1818 | ||
1819 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1820 | /* | |
1821 | * pass busy_cfs_rq argument into | |
1822 | * load_balance_[start|next]_fair iterators | |
1823 | */ | |
1824 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1825 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1826 | &cfs_rq_iterator)) | |
1827 | return 1; | |
1828 | } | |
1829 | ||
1830 | return 0; | |
1831 | } | |
55e12e5e | 1832 | #endif /* CONFIG_SMP */ |
e1d1484f | 1833 | |
bf0f6f24 IM |
1834 | /* |
1835 | * scheduler tick hitting a task of our scheduling class: | |
1836 | */ | |
8f4d37ec | 1837 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1838 | { |
1839 | struct cfs_rq *cfs_rq; | |
1840 | struct sched_entity *se = &curr->se; | |
1841 | ||
1842 | for_each_sched_entity(se) { | |
1843 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1844 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1845 | } |
1846 | } | |
1847 | ||
1848 | /* | |
1849 | * Share the fairness runtime between parent and child, thus the | |
1850 | * total amount of pressure for CPU stays equal - new tasks | |
1851 | * get a chance to run but frequent forkers are not allowed to | |
1852 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1853 | * the child is not running yet. | |
1854 | */ | |
ee0827d8 | 1855 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1856 | { |
1857 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1858 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1859 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1860 | |
1861 | sched_info_queued(p); | |
1862 | ||
7109c442 | 1863 | update_curr(cfs_rq); |
b5d9d734 MG |
1864 | if (curr) |
1865 | se->vruntime = curr->vruntime; | |
aeb73b04 | 1866 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1867 | |
3c90e6e9 | 1868 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1869 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
54fdc581 | 1870 | curr && entity_before(curr, se)) { |
87fefa38 | 1871 | /* |
edcb60a3 IM |
1872 | * Upon rescheduling, sched_class::put_prev_task() will place |
1873 | * 'current' within the tree based on its new key value. | |
1874 | */ | |
4d78e7b6 | 1875 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 1876 | resched_task(rq->curr); |
4d78e7b6 | 1877 | } |
bf0f6f24 | 1878 | |
b9dca1e0 | 1879 | enqueue_task_fair(rq, p, 0); |
bf0f6f24 IM |
1880 | } |
1881 | ||
cb469845 SR |
1882 | /* |
1883 | * Priority of the task has changed. Check to see if we preempt | |
1884 | * the current task. | |
1885 | */ | |
1886 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1887 | int oldprio, int running) | |
1888 | { | |
1889 | /* | |
1890 | * Reschedule if we are currently running on this runqueue and | |
1891 | * our priority decreased, or if we are not currently running on | |
1892 | * this runqueue and our priority is higher than the current's | |
1893 | */ | |
1894 | if (running) { | |
1895 | if (p->prio > oldprio) | |
1896 | resched_task(rq->curr); | |
1897 | } else | |
15afe09b | 1898 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1899 | } |
1900 | ||
1901 | /* | |
1902 | * We switched to the sched_fair class. | |
1903 | */ | |
1904 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1905 | int running) | |
1906 | { | |
1907 | /* | |
1908 | * We were most likely switched from sched_rt, so | |
1909 | * kick off the schedule if running, otherwise just see | |
1910 | * if we can still preempt the current task. | |
1911 | */ | |
1912 | if (running) | |
1913 | resched_task(rq->curr); | |
1914 | else | |
15afe09b | 1915 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1916 | } |
1917 | ||
83b699ed SV |
1918 | /* Account for a task changing its policy or group. |
1919 | * | |
1920 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1921 | * migrates between groups/classes. | |
1922 | */ | |
1923 | static void set_curr_task_fair(struct rq *rq) | |
1924 | { | |
1925 | struct sched_entity *se = &rq->curr->se; | |
1926 | ||
1927 | for_each_sched_entity(se) | |
1928 | set_next_entity(cfs_rq_of(se), se); | |
1929 | } | |
1930 | ||
810b3817 PZ |
1931 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1932 | static void moved_group_fair(struct task_struct *p) | |
1933 | { | |
1934 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1935 | ||
1936 | update_curr(cfs_rq); | |
1937 | place_entity(cfs_rq, &p->se, 1); | |
1938 | } | |
1939 | #endif | |
1940 | ||
0d721cea PW |
1941 | unsigned int get_rr_interval_fair(struct task_struct *task) |
1942 | { | |
1943 | struct sched_entity *se = &task->se; | |
1944 | unsigned long flags; | |
1945 | struct rq *rq; | |
1946 | unsigned int rr_interval = 0; | |
1947 | ||
1948 | /* | |
1949 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
1950 | * idle runqueue: | |
1951 | */ | |
1952 | rq = task_rq_lock(task, &flags); | |
1953 | if (rq->cfs.load.weight) | |
1954 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
1955 | task_rq_unlock(rq, &flags); | |
1956 | ||
1957 | return rr_interval; | |
1958 | } | |
1959 | ||
bf0f6f24 IM |
1960 | /* |
1961 | * All the scheduling class methods: | |
1962 | */ | |
5522d5d5 IM |
1963 | static const struct sched_class fair_sched_class = { |
1964 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1965 | .enqueue_task = enqueue_task_fair, |
1966 | .dequeue_task = dequeue_task_fair, | |
1967 | .yield_task = yield_task_fair, | |
1968 | ||
2e09bf55 | 1969 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1970 | |
1971 | .pick_next_task = pick_next_task_fair, | |
1972 | .put_prev_task = put_prev_task_fair, | |
1973 | ||
681f3e68 | 1974 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
1975 | .select_task_rq = select_task_rq_fair, |
1976 | ||
bf0f6f24 | 1977 | .load_balance = load_balance_fair, |
e1d1484f | 1978 | .move_one_task = move_one_task_fair, |
681f3e68 | 1979 | #endif |
bf0f6f24 | 1980 | |
83b699ed | 1981 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1982 | .task_tick = task_tick_fair, |
1983 | .task_new = task_new_fair, | |
cb469845 SR |
1984 | |
1985 | .prio_changed = prio_changed_fair, | |
1986 | .switched_to = switched_to_fair, | |
810b3817 | 1987 | |
0d721cea PW |
1988 | .get_rr_interval = get_rr_interval_fair, |
1989 | ||
810b3817 PZ |
1990 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1991 | .moved_group = moved_group_fair, | |
1992 | #endif | |
bf0f6f24 IM |
1993 | }; |
1994 | ||
1995 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1996 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1997 | { |
bf0f6f24 IM |
1998 | struct cfs_rq *cfs_rq; |
1999 | ||
5973e5b9 | 2000 | rcu_read_lock(); |
c3b64f1e | 2001 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 2002 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 2003 | rcu_read_unlock(); |
bf0f6f24 IM |
2004 | } |
2005 | #endif |