]>
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_OTHER wake-up granularity. | |
103638d9 | 66 | * (default: 5 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 | */ | |
103638d9 | 72 | unsigned int sysctl_sched_wakeup_granularity = 5000000UL; |
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 | ||
b758149c PZ |
82 | static inline struct task_struct *task_of(struct sched_entity *se) |
83 | { | |
84 | return container_of(se, struct task_struct, se); | |
85 | } | |
86 | ||
62160e3f | 87 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 88 | |
62160e3f | 89 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
90 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
91 | { | |
62160e3f | 92 | return cfs_rq->rq; |
bf0f6f24 IM |
93 | } |
94 | ||
62160e3f IM |
95 | /* An entity is a task if it doesn't "own" a runqueue */ |
96 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 97 | |
b758149c PZ |
98 | /* Walk up scheduling entities hierarchy */ |
99 | #define for_each_sched_entity(se) \ | |
100 | for (; se; se = se->parent) | |
101 | ||
102 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
103 | { | |
104 | return p->se.cfs_rq; | |
105 | } | |
106 | ||
107 | /* runqueue on which this entity is (to be) queued */ | |
108 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
109 | { | |
110 | return se->cfs_rq; | |
111 | } | |
112 | ||
113 | /* runqueue "owned" by this group */ | |
114 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
115 | { | |
116 | return grp->my_q; | |
117 | } | |
118 | ||
119 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on | |
120 | * another cpu ('this_cpu') | |
121 | */ | |
122 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
123 | { | |
124 | return cfs_rq->tg->cfs_rq[this_cpu]; | |
125 | } | |
126 | ||
127 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | |
128 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
129 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
130 | ||
131 | /* Do the two (enqueued) entities belong to the same group ? */ | |
132 | static inline int | |
133 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
134 | { | |
135 | if (se->cfs_rq == pse->cfs_rq) | |
136 | return 1; | |
137 | ||
138 | return 0; | |
139 | } | |
140 | ||
141 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
142 | { | |
143 | return se->parent; | |
144 | } | |
145 | ||
464b7527 PZ |
146 | /* return depth at which a sched entity is present in the hierarchy */ |
147 | static inline int depth_se(struct sched_entity *se) | |
148 | { | |
149 | int depth = 0; | |
150 | ||
151 | for_each_sched_entity(se) | |
152 | depth++; | |
153 | ||
154 | return depth; | |
155 | } | |
156 | ||
157 | static void | |
158 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
159 | { | |
160 | int se_depth, pse_depth; | |
161 | ||
162 | /* | |
163 | * preemption test can be made between sibling entities who are in the | |
164 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
165 | * both tasks until we find their ancestors who are siblings of common | |
166 | * parent. | |
167 | */ | |
168 | ||
169 | /* First walk up until both entities are at same depth */ | |
170 | se_depth = depth_se(*se); | |
171 | pse_depth = depth_se(*pse); | |
172 | ||
173 | while (se_depth > pse_depth) { | |
174 | se_depth--; | |
175 | *se = parent_entity(*se); | |
176 | } | |
177 | ||
178 | while (pse_depth > se_depth) { | |
179 | pse_depth--; | |
180 | *pse = parent_entity(*pse); | |
181 | } | |
182 | ||
183 | while (!is_same_group(*se, *pse)) { | |
184 | *se = parent_entity(*se); | |
185 | *pse = parent_entity(*pse); | |
186 | } | |
187 | } | |
188 | ||
62160e3f | 189 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
bf0f6f24 | 190 | |
62160e3f IM |
191 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
192 | { | |
193 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
194 | } |
195 | ||
196 | #define entity_is_task(se) 1 | |
197 | ||
b758149c PZ |
198 | #define for_each_sched_entity(se) \ |
199 | for (; se; se = NULL) | |
bf0f6f24 | 200 | |
b758149c | 201 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 202 | { |
b758149c | 203 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
204 | } |
205 | ||
b758149c PZ |
206 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
207 | { | |
208 | struct task_struct *p = task_of(se); | |
209 | struct rq *rq = task_rq(p); | |
210 | ||
211 | return &rq->cfs; | |
212 | } | |
213 | ||
214 | /* runqueue "owned" by this group */ | |
215 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
216 | { | |
217 | return NULL; | |
218 | } | |
219 | ||
220 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) | |
221 | { | |
222 | return &cpu_rq(this_cpu)->cfs; | |
223 | } | |
224 | ||
225 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
226 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
227 | ||
228 | static inline int | |
229 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
230 | { | |
231 | return 1; | |
232 | } | |
233 | ||
234 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
235 | { | |
236 | return NULL; | |
237 | } | |
238 | ||
464b7527 PZ |
239 | static inline void |
240 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
241 | { | |
242 | } | |
243 | ||
b758149c PZ |
244 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
245 | ||
bf0f6f24 IM |
246 | |
247 | /************************************************************** | |
248 | * Scheduling class tree data structure manipulation methods: | |
249 | */ | |
250 | ||
0702e3eb | 251 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 252 | { |
368059a9 PZ |
253 | s64 delta = (s64)(vruntime - min_vruntime); |
254 | if (delta > 0) | |
02e0431a PZ |
255 | min_vruntime = vruntime; |
256 | ||
257 | return min_vruntime; | |
258 | } | |
259 | ||
0702e3eb | 260 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
261 | { |
262 | s64 delta = (s64)(vruntime - min_vruntime); | |
263 | if (delta < 0) | |
264 | min_vruntime = vruntime; | |
265 | ||
266 | return min_vruntime; | |
267 | } | |
268 | ||
54fdc581 FC |
269 | static inline int entity_before(struct sched_entity *a, |
270 | struct sched_entity *b) | |
271 | { | |
272 | return (s64)(a->vruntime - b->vruntime) < 0; | |
273 | } | |
274 | ||
0702e3eb | 275 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
9014623c | 276 | { |
30cfdcfc | 277 | return se->vruntime - cfs_rq->min_vruntime; |
9014623c PZ |
278 | } |
279 | ||
1af5f730 PZ |
280 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
281 | { | |
282 | u64 vruntime = cfs_rq->min_vruntime; | |
283 | ||
284 | if (cfs_rq->curr) | |
285 | vruntime = cfs_rq->curr->vruntime; | |
286 | ||
287 | if (cfs_rq->rb_leftmost) { | |
288 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
289 | struct sched_entity, | |
290 | run_node); | |
291 | ||
e17036da | 292 | if (!cfs_rq->curr) |
1af5f730 PZ |
293 | vruntime = se->vruntime; |
294 | else | |
295 | vruntime = min_vruntime(vruntime, se->vruntime); | |
296 | } | |
297 | ||
298 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
299 | } | |
300 | ||
bf0f6f24 IM |
301 | /* |
302 | * Enqueue an entity into the rb-tree: | |
303 | */ | |
0702e3eb | 304 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
305 | { |
306 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
307 | struct rb_node *parent = NULL; | |
308 | struct sched_entity *entry; | |
9014623c | 309 | s64 key = entity_key(cfs_rq, se); |
bf0f6f24 IM |
310 | int leftmost = 1; |
311 | ||
312 | /* | |
313 | * Find the right place in the rbtree: | |
314 | */ | |
315 | while (*link) { | |
316 | parent = *link; | |
317 | entry = rb_entry(parent, struct sched_entity, run_node); | |
318 | /* | |
319 | * We dont care about collisions. Nodes with | |
320 | * the same key stay together. | |
321 | */ | |
9014623c | 322 | if (key < entity_key(cfs_rq, entry)) { |
bf0f6f24 IM |
323 | link = &parent->rb_left; |
324 | } else { | |
325 | link = &parent->rb_right; | |
326 | leftmost = 0; | |
327 | } | |
328 | } | |
329 | ||
330 | /* | |
331 | * Maintain a cache of leftmost tree entries (it is frequently | |
332 | * used): | |
333 | */ | |
1af5f730 | 334 | if (leftmost) |
57cb499d | 335 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
336 | |
337 | rb_link_node(&se->run_node, parent, link); | |
338 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
339 | } |
340 | ||
0702e3eb | 341 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 342 | { |
3fe69747 PZ |
343 | if (cfs_rq->rb_leftmost == &se->run_node) { |
344 | struct rb_node *next_node; | |
3fe69747 PZ |
345 | |
346 | next_node = rb_next(&se->run_node); | |
347 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 348 | } |
e9acbff6 | 349 | |
bf0f6f24 | 350 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
351 | } |
352 | ||
bf0f6f24 IM |
353 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
354 | { | |
f4b6755f PZ |
355 | struct rb_node *left = cfs_rq->rb_leftmost; |
356 | ||
357 | if (!left) | |
358 | return NULL; | |
359 | ||
360 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
361 | } |
362 | ||
f4b6755f | 363 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 364 | { |
7eee3e67 | 365 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 366 | |
70eee74b BS |
367 | if (!last) |
368 | return NULL; | |
7eee3e67 IM |
369 | |
370 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
371 | } |
372 | ||
bf0f6f24 IM |
373 | /************************************************************** |
374 | * Scheduling class statistics methods: | |
375 | */ | |
376 | ||
b2be5e96 PZ |
377 | #ifdef CONFIG_SCHED_DEBUG |
378 | int sched_nr_latency_handler(struct ctl_table *table, int write, | |
379 | struct file *filp, void __user *buffer, size_t *lenp, | |
380 | loff_t *ppos) | |
381 | { | |
382 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); | |
383 | ||
384 | if (ret || !write) | |
385 | return ret; | |
386 | ||
387 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
388 | sysctl_sched_min_granularity); | |
389 | ||
390 | return 0; | |
391 | } | |
392 | #endif | |
647e7cac | 393 | |
a7be37ac | 394 | /* |
f9c0b095 | 395 | * delta /= w |
a7be37ac PZ |
396 | */ |
397 | static inline unsigned long | |
398 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
399 | { | |
f9c0b095 PZ |
400 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
401 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
402 | |
403 | return delta; | |
404 | } | |
405 | ||
647e7cac IM |
406 | /* |
407 | * The idea is to set a period in which each task runs once. | |
408 | * | |
409 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
410 | * this period because otherwise the slices get too small. | |
411 | * | |
412 | * p = (nr <= nl) ? l : l*nr/nl | |
413 | */ | |
4d78e7b6 PZ |
414 | static u64 __sched_period(unsigned long nr_running) |
415 | { | |
416 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 417 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
418 | |
419 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 420 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 421 | period *= nr_running; |
4d78e7b6 PZ |
422 | } |
423 | ||
424 | return period; | |
425 | } | |
426 | ||
647e7cac IM |
427 | /* |
428 | * We calculate the wall-time slice from the period by taking a part | |
429 | * proportional to the weight. | |
430 | * | |
f9c0b095 | 431 | * s = p*P[w/rw] |
647e7cac | 432 | */ |
6d0f0ebd | 433 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 434 | { |
0a582440 | 435 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 436 | |
0a582440 | 437 | for_each_sched_entity(se) { |
6272d68c | 438 | struct load_weight *load; |
3104bf03 | 439 | struct load_weight lw; |
6272d68c LM |
440 | |
441 | cfs_rq = cfs_rq_of(se); | |
442 | load = &cfs_rq->load; | |
f9c0b095 | 443 | |
0a582440 | 444 | if (unlikely(!se->on_rq)) { |
3104bf03 | 445 | lw = cfs_rq->load; |
0a582440 MG |
446 | |
447 | update_load_add(&lw, se->load.weight); | |
448 | load = &lw; | |
449 | } | |
450 | slice = calc_delta_mine(slice, se->load.weight, load); | |
451 | } | |
452 | return slice; | |
bf0f6f24 IM |
453 | } |
454 | ||
647e7cac | 455 | /* |
ac884dec | 456 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 457 | * |
f9c0b095 | 458 | * vs = s/w |
647e7cac | 459 | */ |
f9c0b095 | 460 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 461 | { |
f9c0b095 | 462 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
463 | } |
464 | ||
bf0f6f24 IM |
465 | /* |
466 | * Update the current task's runtime statistics. Skip current tasks that | |
467 | * are not in our scheduling class. | |
468 | */ | |
469 | static inline void | |
8ebc91d9 IM |
470 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
471 | unsigned long delta_exec) | |
bf0f6f24 | 472 | { |
bbdba7c0 | 473 | unsigned long delta_exec_weighted; |
bf0f6f24 | 474 | |
8179ca23 | 475 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
bf0f6f24 IM |
476 | |
477 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 478 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 479 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
e9acbff6 | 480 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 481 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
482 | } |
483 | ||
b7cc0896 | 484 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 485 | { |
429d43bc | 486 | struct sched_entity *curr = cfs_rq->curr; |
8ebc91d9 | 487 | u64 now = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
488 | unsigned long delta_exec; |
489 | ||
490 | if (unlikely(!curr)) | |
491 | return; | |
492 | ||
493 | /* | |
494 | * Get the amount of time the current task was running | |
495 | * since the last time we changed load (this cannot | |
496 | * overflow on 32 bits): | |
497 | */ | |
8ebc91d9 | 498 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
499 | if (!delta_exec) |
500 | return; | |
bf0f6f24 | 501 | |
8ebc91d9 IM |
502 | __update_curr(cfs_rq, curr, delta_exec); |
503 | curr->exec_start = now; | |
d842de87 SV |
504 | |
505 | if (entity_is_task(curr)) { | |
506 | struct task_struct *curtask = task_of(curr); | |
507 | ||
508 | cpuacct_charge(curtask, delta_exec); | |
f06febc9 | 509 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 510 | } |
bf0f6f24 IM |
511 | } |
512 | ||
513 | static inline void | |
5870db5b | 514 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 515 | { |
d281918d | 516 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
517 | } |
518 | ||
bf0f6f24 IM |
519 | /* |
520 | * Task is being enqueued - update stats: | |
521 | */ | |
d2417e5a | 522 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 523 | { |
bf0f6f24 IM |
524 | /* |
525 | * Are we enqueueing a waiting task? (for current tasks | |
526 | * a dequeue/enqueue event is a NOP) | |
527 | */ | |
429d43bc | 528 | if (se != cfs_rq->curr) |
5870db5b | 529 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
530 | } |
531 | ||
bf0f6f24 | 532 | static void |
9ef0a961 | 533 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 534 | { |
bbdba7c0 IM |
535 | schedstat_set(se->wait_max, max(se->wait_max, |
536 | rq_of(cfs_rq)->clock - se->wait_start)); | |
6d082592 AV |
537 | schedstat_set(se->wait_count, se->wait_count + 1); |
538 | schedstat_set(se->wait_sum, se->wait_sum + | |
539 | rq_of(cfs_rq)->clock - se->wait_start); | |
6cfb0d5d | 540 | schedstat_set(se->wait_start, 0); |
bf0f6f24 IM |
541 | } |
542 | ||
543 | static inline void | |
19b6a2e3 | 544 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 545 | { |
bf0f6f24 IM |
546 | /* |
547 | * Mark the end of the wait period if dequeueing a | |
548 | * waiting task: | |
549 | */ | |
429d43bc | 550 | if (se != cfs_rq->curr) |
9ef0a961 | 551 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
552 | } |
553 | ||
554 | /* | |
555 | * We are picking a new current task - update its stats: | |
556 | */ | |
557 | static inline void | |
79303e9e | 558 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
559 | { |
560 | /* | |
561 | * We are starting a new run period: | |
562 | */ | |
d281918d | 563 | se->exec_start = rq_of(cfs_rq)->clock; |
bf0f6f24 IM |
564 | } |
565 | ||
bf0f6f24 IM |
566 | /************************************************** |
567 | * Scheduling class queueing methods: | |
568 | */ | |
569 | ||
c09595f6 PZ |
570 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
571 | static void | |
572 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
573 | { | |
574 | cfs_rq->task_weight += weight; | |
575 | } | |
576 | #else | |
577 | static inline void | |
578 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
579 | { | |
580 | } | |
581 | #endif | |
582 | ||
30cfdcfc DA |
583 | static void |
584 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
585 | { | |
586 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
587 | if (!parent_entity(se)) |
588 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 589 | if (entity_is_task(se)) { |
c09595f6 | 590 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
591 | list_add(&se->group_node, &cfs_rq->tasks); |
592 | } | |
30cfdcfc DA |
593 | cfs_rq->nr_running++; |
594 | se->on_rq = 1; | |
595 | } | |
596 | ||
597 | static void | |
598 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
599 | { | |
600 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
601 | if (!parent_entity(se)) |
602 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 603 | if (entity_is_task(se)) { |
c09595f6 | 604 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
605 | list_del_init(&se->group_node); |
606 | } | |
30cfdcfc DA |
607 | cfs_rq->nr_running--; |
608 | se->on_rq = 0; | |
609 | } | |
610 | ||
2396af69 | 611 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 612 | { |
bf0f6f24 IM |
613 | #ifdef CONFIG_SCHEDSTATS |
614 | if (se->sleep_start) { | |
d281918d | 615 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
9745512c | 616 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
617 | |
618 | if ((s64)delta < 0) | |
619 | delta = 0; | |
620 | ||
621 | if (unlikely(delta > se->sleep_max)) | |
622 | se->sleep_max = delta; | |
623 | ||
624 | se->sleep_start = 0; | |
625 | se->sum_sleep_runtime += delta; | |
9745512c AV |
626 | |
627 | account_scheduler_latency(tsk, delta >> 10, 1); | |
bf0f6f24 IM |
628 | } |
629 | if (se->block_start) { | |
d281918d | 630 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
9745512c | 631 | struct task_struct *tsk = task_of(se); |
bf0f6f24 IM |
632 | |
633 | if ((s64)delta < 0) | |
634 | delta = 0; | |
635 | ||
636 | if (unlikely(delta > se->block_max)) | |
637 | se->block_max = delta; | |
638 | ||
639 | se->block_start = 0; | |
640 | se->sum_sleep_runtime += delta; | |
30084fbd IM |
641 | |
642 | /* | |
643 | * Blocking time is in units of nanosecs, so shift by 20 to | |
644 | * get a milliseconds-range estimation of the amount of | |
645 | * time that the task spent sleeping: | |
646 | */ | |
647 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
e22f5bbf | 648 | |
30084fbd IM |
649 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
650 | delta >> 20); | |
651 | } | |
9745512c | 652 | account_scheduler_latency(tsk, delta >> 10, 0); |
bf0f6f24 IM |
653 | } |
654 | #endif | |
655 | } | |
656 | ||
ddc97297 PZ |
657 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
658 | { | |
659 | #ifdef CONFIG_SCHED_DEBUG | |
660 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
661 | ||
662 | if (d < 0) | |
663 | d = -d; | |
664 | ||
665 | if (d > 3*sysctl_sched_latency) | |
666 | schedstat_inc(cfs_rq, nr_spread_over); | |
667 | #endif | |
668 | } | |
669 | ||
aeb73b04 PZ |
670 | static void |
671 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
672 | { | |
1af5f730 | 673 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 674 | |
2cb8600e PZ |
675 | /* |
676 | * The 'current' period is already promised to the current tasks, | |
677 | * however the extra weight of the new task will slow them down a | |
678 | * little, place the new task so that it fits in the slot that | |
679 | * stays open at the end. | |
680 | */ | |
94dfb5e7 | 681 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 682 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 683 | |
8465e792 | 684 | if (!initial) { |
2cb8600e | 685 | /* sleeps upto a single latency don't count. */ |
a7be37ac PZ |
686 | if (sched_feat(NEW_FAIR_SLEEPERS)) { |
687 | unsigned long thresh = sysctl_sched_latency; | |
688 | ||
689 | /* | |
6bc912b7 PZ |
690 | * Convert the sleeper threshold into virtual time. |
691 | * SCHED_IDLE is a special sub-class. We care about | |
692 | * fairness only relative to other SCHED_IDLE tasks, | |
693 | * all of which have the same weight. | |
a7be37ac | 694 | */ |
6bc912b7 | 695 | if (sched_feat(NORMALIZED_SLEEPER) && |
d07387b4 PT |
696 | (!entity_is_task(se) || |
697 | task_of(se)->policy != SCHED_IDLE)) | |
a7be37ac PZ |
698 | thresh = calc_delta_fair(thresh, se); |
699 | ||
700 | vruntime -= thresh; | |
701 | } | |
94359f05 | 702 | |
2cb8600e PZ |
703 | /* ensure we never gain time by being placed backwards. */ |
704 | vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
705 | } |
706 | ||
67e9fb2a | 707 | se->vruntime = vruntime; |
aeb73b04 PZ |
708 | } |
709 | ||
bf0f6f24 | 710 | static void |
83b699ed | 711 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
bf0f6f24 IM |
712 | { |
713 | /* | |
a2a2d680 | 714 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 715 | */ |
b7cc0896 | 716 | update_curr(cfs_rq); |
a992241d | 717 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 718 | |
e9acbff6 | 719 | if (wakeup) { |
aeb73b04 | 720 | place_entity(cfs_rq, se, 0); |
2396af69 | 721 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 722 | } |
bf0f6f24 | 723 | |
d2417e5a | 724 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 725 | check_spread(cfs_rq, se); |
83b699ed SV |
726 | if (se != cfs_rq->curr) |
727 | __enqueue_entity(cfs_rq, se); | |
bf0f6f24 IM |
728 | } |
729 | ||
a571bbea | 730 | static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2002c695 PZ |
731 | { |
732 | if (cfs_rq->last == se) | |
733 | cfs_rq->last = NULL; | |
734 | ||
735 | if (cfs_rq->next == se) | |
736 | cfs_rq->next = NULL; | |
737 | } | |
738 | ||
a571bbea PZ |
739 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
740 | { | |
741 | for_each_sched_entity(se) | |
742 | __clear_buddies(cfs_rq_of(se), se); | |
743 | } | |
744 | ||
bf0f6f24 | 745 | static void |
525c2716 | 746 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
bf0f6f24 | 747 | { |
a2a2d680 DA |
748 | /* |
749 | * Update run-time statistics of the 'current'. | |
750 | */ | |
751 | update_curr(cfs_rq); | |
752 | ||
19b6a2e3 | 753 | update_stats_dequeue(cfs_rq, se); |
db36cc7d | 754 | if (sleep) { |
67e9fb2a | 755 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
756 | if (entity_is_task(se)) { |
757 | struct task_struct *tsk = task_of(se); | |
758 | ||
759 | if (tsk->state & TASK_INTERRUPTIBLE) | |
d281918d | 760 | se->sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 761 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
d281918d | 762 | se->block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 763 | } |
db36cc7d | 764 | #endif |
67e9fb2a PZ |
765 | } |
766 | ||
2002c695 | 767 | clear_buddies(cfs_rq, se); |
4793241b | 768 | |
83b699ed | 769 | if (se != cfs_rq->curr) |
30cfdcfc DA |
770 | __dequeue_entity(cfs_rq, se); |
771 | account_entity_dequeue(cfs_rq, se); | |
1af5f730 | 772 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
773 | } |
774 | ||
775 | /* | |
776 | * Preempt the current task with a newly woken task if needed: | |
777 | */ | |
7c92e54f | 778 | static void |
2e09bf55 | 779 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 780 | { |
11697830 PZ |
781 | unsigned long ideal_runtime, delta_exec; |
782 | ||
6d0f0ebd | 783 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 784 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 785 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 786 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
787 | /* |
788 | * The current task ran long enough, ensure it doesn't get | |
789 | * re-elected due to buddy favours. | |
790 | */ | |
791 | clear_buddies(cfs_rq, curr); | |
792 | } | |
bf0f6f24 IM |
793 | } |
794 | ||
83b699ed | 795 | static void |
8494f412 | 796 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 797 | { |
83b699ed SV |
798 | /* 'current' is not kept within the tree. */ |
799 | if (se->on_rq) { | |
800 | /* | |
801 | * Any task has to be enqueued before it get to execute on | |
802 | * a CPU. So account for the time it spent waiting on the | |
803 | * runqueue. | |
804 | */ | |
805 | update_stats_wait_end(cfs_rq, se); | |
806 | __dequeue_entity(cfs_rq, se); | |
807 | } | |
808 | ||
79303e9e | 809 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 810 | cfs_rq->curr = se; |
eba1ed4b IM |
811 | #ifdef CONFIG_SCHEDSTATS |
812 | /* | |
813 | * Track our maximum slice length, if the CPU's load is at | |
814 | * least twice that of our own weight (i.e. dont track it | |
815 | * when there are only lesser-weight tasks around): | |
816 | */ | |
495eca49 | 817 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
eba1ed4b IM |
818 | se->slice_max = max(se->slice_max, |
819 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | |
820 | } | |
821 | #endif | |
4a55b450 | 822 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
823 | } |
824 | ||
3f3a4904 PZ |
825 | static int |
826 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
827 | ||
f4b6755f | 828 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 829 | { |
f4b6755f PZ |
830 | struct sched_entity *se = __pick_next_entity(cfs_rq); |
831 | ||
4793241b PZ |
832 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1) |
833 | return cfs_rq->next; | |
aa2ac252 | 834 | |
4793241b PZ |
835 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1) |
836 | return cfs_rq->last; | |
837 | ||
838 | return se; | |
aa2ac252 PZ |
839 | } |
840 | ||
ab6cde26 | 841 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
842 | { |
843 | /* | |
844 | * If still on the runqueue then deactivate_task() | |
845 | * was not called and update_curr() has to be done: | |
846 | */ | |
847 | if (prev->on_rq) | |
b7cc0896 | 848 | update_curr(cfs_rq); |
bf0f6f24 | 849 | |
ddc97297 | 850 | check_spread(cfs_rq, prev); |
30cfdcfc | 851 | if (prev->on_rq) { |
5870db5b | 852 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
853 | /* Put 'current' back into the tree. */ |
854 | __enqueue_entity(cfs_rq, prev); | |
855 | } | |
429d43bc | 856 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
857 | } |
858 | ||
8f4d37ec PZ |
859 | static void |
860 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 861 | { |
bf0f6f24 | 862 | /* |
30cfdcfc | 863 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 864 | */ |
30cfdcfc | 865 | update_curr(cfs_rq); |
bf0f6f24 | 866 | |
8f4d37ec PZ |
867 | #ifdef CONFIG_SCHED_HRTICK |
868 | /* | |
869 | * queued ticks are scheduled to match the slice, so don't bother | |
870 | * validating it and just reschedule. | |
871 | */ | |
983ed7a6 HH |
872 | if (queued) { |
873 | resched_task(rq_of(cfs_rq)->curr); | |
874 | return; | |
875 | } | |
8f4d37ec PZ |
876 | /* |
877 | * don't let the period tick interfere with the hrtick preemption | |
878 | */ | |
879 | if (!sched_feat(DOUBLE_TICK) && | |
880 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
881 | return; | |
882 | #endif | |
883 | ||
ce6c1311 | 884 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
2e09bf55 | 885 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
886 | } |
887 | ||
888 | /************************************************** | |
889 | * CFS operations on tasks: | |
890 | */ | |
891 | ||
8f4d37ec PZ |
892 | #ifdef CONFIG_SCHED_HRTICK |
893 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
894 | { | |
8f4d37ec PZ |
895 | struct sched_entity *se = &p->se; |
896 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
897 | ||
898 | WARN_ON(task_rq(p) != rq); | |
899 | ||
900 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
901 | u64 slice = sched_slice(cfs_rq, se); | |
902 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
903 | s64 delta = slice - ran; | |
904 | ||
905 | if (delta < 0) { | |
906 | if (rq->curr == p) | |
907 | resched_task(p); | |
908 | return; | |
909 | } | |
910 | ||
911 | /* | |
912 | * Don't schedule slices shorter than 10000ns, that just | |
913 | * doesn't make sense. Rely on vruntime for fairness. | |
914 | */ | |
31656519 | 915 | if (rq->curr != p) |
157124c1 | 916 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 917 | |
31656519 | 918 | hrtick_start(rq, delta); |
8f4d37ec PZ |
919 | } |
920 | } | |
a4c2f00f PZ |
921 | |
922 | /* | |
923 | * called from enqueue/dequeue and updates the hrtick when the | |
924 | * current task is from our class and nr_running is low enough | |
925 | * to matter. | |
926 | */ | |
927 | static void hrtick_update(struct rq *rq) | |
928 | { | |
929 | struct task_struct *curr = rq->curr; | |
930 | ||
931 | if (curr->sched_class != &fair_sched_class) | |
932 | return; | |
933 | ||
934 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
935 | hrtick_start_fair(rq, curr); | |
936 | } | |
55e12e5e | 937 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
938 | static inline void |
939 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
940 | { | |
941 | } | |
a4c2f00f PZ |
942 | |
943 | static inline void hrtick_update(struct rq *rq) | |
944 | { | |
945 | } | |
8f4d37ec PZ |
946 | #endif |
947 | ||
bf0f6f24 IM |
948 | /* |
949 | * The enqueue_task method is called before nr_running is | |
950 | * increased. Here we update the fair scheduling stats and | |
951 | * then put the task into the rbtree: | |
952 | */ | |
fd390f6a | 953 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
bf0f6f24 IM |
954 | { |
955 | struct cfs_rq *cfs_rq; | |
62fb1851 | 956 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
957 | |
958 | for_each_sched_entity(se) { | |
62fb1851 | 959 | if (se->on_rq) |
bf0f6f24 IM |
960 | break; |
961 | cfs_rq = cfs_rq_of(se); | |
83b699ed | 962 | enqueue_entity(cfs_rq, se, wakeup); |
b9fa3df3 | 963 | wakeup = 1; |
bf0f6f24 | 964 | } |
8f4d37ec | 965 | |
a4c2f00f | 966 | hrtick_update(rq); |
bf0f6f24 IM |
967 | } |
968 | ||
969 | /* | |
970 | * The dequeue_task method is called before nr_running is | |
971 | * decreased. We remove the task from the rbtree and | |
972 | * update the fair scheduling stats: | |
973 | */ | |
f02231e5 | 974 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
bf0f6f24 IM |
975 | { |
976 | struct cfs_rq *cfs_rq; | |
62fb1851 | 977 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
978 | |
979 | for_each_sched_entity(se) { | |
980 | cfs_rq = cfs_rq_of(se); | |
525c2716 | 981 | dequeue_entity(cfs_rq, se, sleep); |
bf0f6f24 | 982 | /* Don't dequeue parent if it has other entities besides us */ |
62fb1851 | 983 | if (cfs_rq->load.weight) |
bf0f6f24 | 984 | break; |
b9fa3df3 | 985 | sleep = 1; |
bf0f6f24 | 986 | } |
8f4d37ec | 987 | |
a4c2f00f | 988 | hrtick_update(rq); |
bf0f6f24 IM |
989 | } |
990 | ||
991 | /* | |
1799e35d IM |
992 | * sched_yield() support is very simple - we dequeue and enqueue. |
993 | * | |
994 | * If compat_yield is turned on then we requeue to the end of the tree. | |
bf0f6f24 | 995 | */ |
4530d7ab | 996 | static void yield_task_fair(struct rq *rq) |
bf0f6f24 | 997 | { |
db292ca3 IM |
998 | struct task_struct *curr = rq->curr; |
999 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
1000 | struct sched_entity *rightmost, *se = &curr->se; | |
bf0f6f24 IM |
1001 | |
1002 | /* | |
1799e35d IM |
1003 | * Are we the only task in the tree? |
1004 | */ | |
1005 | if (unlikely(cfs_rq->nr_running == 1)) | |
1006 | return; | |
1007 | ||
2002c695 PZ |
1008 | clear_buddies(cfs_rq, se); |
1009 | ||
db292ca3 | 1010 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
3e51f33f | 1011 | update_rq_clock(rq); |
1799e35d | 1012 | /* |
a2a2d680 | 1013 | * Update run-time statistics of the 'current'. |
1799e35d | 1014 | */ |
2b1e315d | 1015 | update_curr(cfs_rq); |
1799e35d IM |
1016 | |
1017 | return; | |
1018 | } | |
1019 | /* | |
1020 | * Find the rightmost entry in the rbtree: | |
bf0f6f24 | 1021 | */ |
2b1e315d | 1022 | rightmost = __pick_last_entity(cfs_rq); |
1799e35d IM |
1023 | /* |
1024 | * Already in the rightmost position? | |
1025 | */ | |
54fdc581 | 1026 | if (unlikely(!rightmost || entity_before(rightmost, se))) |
1799e35d IM |
1027 | return; |
1028 | ||
1029 | /* | |
1030 | * Minimally necessary key value to be last in the tree: | |
2b1e315d DA |
1031 | * Upon rescheduling, sched_class::put_prev_task() will place |
1032 | * 'current' within the tree based on its new key value. | |
1799e35d | 1033 | */ |
30cfdcfc | 1034 | se->vruntime = rightmost->vruntime + 1; |
bf0f6f24 IM |
1035 | } |
1036 | ||
e7693a36 GH |
1037 | /* |
1038 | * wake_idle() will wake a task on an idle cpu if task->cpu is | |
1039 | * not idle and an idle cpu is available. The span of cpus to | |
1040 | * search starts with cpus closest then further out as needed, | |
1041 | * so we always favor a closer, idle cpu. | |
e761b772 | 1042 | * Domains may include CPUs that are not usable for migration, |
96f874e2 | 1043 | * hence we need to mask them out (cpu_active_mask) |
e7693a36 GH |
1044 | * |
1045 | * Returns the CPU we should wake onto. | |
1046 | */ | |
1047 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) | |
1048 | static int wake_idle(int cpu, struct task_struct *p) | |
1049 | { | |
e7693a36 GH |
1050 | struct sched_domain *sd; |
1051 | int i; | |
7eb52dfa VS |
1052 | unsigned int chosen_wakeup_cpu; |
1053 | int this_cpu; | |
1054 | ||
1055 | /* | |
1056 | * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu | |
1057 | * are idle and this is not a kernel thread and this task's affinity | |
1058 | * allows it to be moved to preferred cpu, then just move! | |
1059 | */ | |
1060 | ||
1061 | this_cpu = smp_processor_id(); | |
1062 | chosen_wakeup_cpu = | |
1063 | cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu; | |
1064 | ||
1065 | if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP && | |
1066 | idle_cpu(cpu) && idle_cpu(this_cpu) && | |
1067 | p->mm && !(p->flags & PF_KTHREAD) && | |
1068 | cpu_isset(chosen_wakeup_cpu, p->cpus_allowed)) | |
1069 | return chosen_wakeup_cpu; | |
e7693a36 GH |
1070 | |
1071 | /* | |
1072 | * If it is idle, then it is the best cpu to run this task. | |
1073 | * | |
1074 | * This cpu is also the best, if it has more than one task already. | |
1075 | * Siblings must be also busy(in most cases) as they didn't already | |
1076 | * pickup the extra load from this cpu and hence we need not check | |
1077 | * sibling runqueue info. This will avoid the checks and cache miss | |
1078 | * penalities associated with that. | |
1079 | */ | |
104f6454 | 1080 | if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1) |
e7693a36 GH |
1081 | return cpu; |
1082 | ||
1083 | for_each_domain(cpu, sd) { | |
1d3504fc HS |
1084 | if ((sd->flags & SD_WAKE_IDLE) |
1085 | || ((sd->flags & SD_WAKE_IDLE_FAR) | |
1086 | && !task_hot(p, task_rq(p)->clock, sd))) { | |
758b2cdc RR |
1087 | for_each_cpu_and(i, sched_domain_span(sd), |
1088 | &p->cpus_allowed) { | |
1089 | if (cpu_active(i) && idle_cpu(i)) { | |
e7693a36 GH |
1090 | if (i != task_cpu(p)) { |
1091 | schedstat_inc(p, | |
1092 | se.nr_wakeups_idle); | |
1093 | } | |
1094 | return i; | |
1095 | } | |
1096 | } | |
1097 | } else { | |
1098 | break; | |
1099 | } | |
1100 | } | |
1101 | return cpu; | |
1102 | } | |
55e12e5e | 1103 | #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/ |
e7693a36 GH |
1104 | static inline int wake_idle(int cpu, struct task_struct *p) |
1105 | { | |
1106 | return cpu; | |
1107 | } | |
1108 | #endif | |
1109 | ||
1110 | #ifdef CONFIG_SMP | |
098fb9db | 1111 | |
bb3469ac | 1112 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1113 | /* |
1114 | * effective_load() calculates the load change as seen from the root_task_group | |
1115 | * | |
1116 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1117 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1118 | * can calculate the shift in shares. | |
1119 | * | |
1120 | * The problem is that perfectly aligning the shares is rather expensive, hence | |
1121 | * we try to avoid doing that too often - see update_shares(), which ratelimits | |
1122 | * this change. | |
1123 | * | |
1124 | * We compensate this by not only taking the current delta into account, but | |
1125 | * also considering the delta between when the shares were last adjusted and | |
1126 | * now. | |
1127 | * | |
1128 | * We still saw a performance dip, some tracing learned us that between | |
1129 | * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased | |
1130 | * significantly. Therefore try to bias the error in direction of failing | |
1131 | * the affine wakeup. | |
1132 | * | |
1133 | */ | |
f1d239f7 PZ |
1134 | static long effective_load(struct task_group *tg, int cpu, |
1135 | long wl, long wg) | |
bb3469ac | 1136 | { |
4be9daaa | 1137 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1138 | |
1139 | if (!tg->parent) | |
1140 | return wl; | |
1141 | ||
f5bfb7d9 PZ |
1142 | /* |
1143 | * By not taking the decrease of shares on the other cpu into | |
1144 | * account our error leans towards reducing the affine wakeups. | |
1145 | */ | |
1146 | if (!wl && sched_feat(ASYM_EFF_LOAD)) | |
1147 | return wl; | |
1148 | ||
4be9daaa | 1149 | for_each_sched_entity(se) { |
cb5ef42a | 1150 | long S, rw, s, a, b; |
940959e9 PZ |
1151 | long more_w; |
1152 | ||
1153 | /* | |
1154 | * Instead of using this increment, also add the difference | |
1155 | * between when the shares were last updated and now. | |
1156 | */ | |
1157 | more_w = se->my_q->load.weight - se->my_q->rq_weight; | |
1158 | wl += more_w; | |
1159 | wg += more_w; | |
4be9daaa PZ |
1160 | |
1161 | S = se->my_q->tg->shares; | |
1162 | s = se->my_q->shares; | |
f1d239f7 | 1163 | rw = se->my_q->rq_weight; |
bb3469ac | 1164 | |
cb5ef42a PZ |
1165 | a = S*(rw + wl); |
1166 | b = S*rw + s*wg; | |
4be9daaa | 1167 | |
940959e9 PZ |
1168 | wl = s*(a-b); |
1169 | ||
1170 | if (likely(b)) | |
1171 | wl /= b; | |
1172 | ||
83378269 PZ |
1173 | /* |
1174 | * Assume the group is already running and will | |
1175 | * thus already be accounted for in the weight. | |
1176 | * | |
1177 | * That is, moving shares between CPUs, does not | |
1178 | * alter the group weight. | |
1179 | */ | |
4be9daaa | 1180 | wg = 0; |
4be9daaa | 1181 | } |
bb3469ac | 1182 | |
4be9daaa | 1183 | return wl; |
bb3469ac | 1184 | } |
4be9daaa | 1185 | |
bb3469ac | 1186 | #else |
4be9daaa | 1187 | |
83378269 PZ |
1188 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1189 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1190 | { |
83378269 | 1191 | return wl; |
bb3469ac | 1192 | } |
4be9daaa | 1193 | |
bb3469ac PZ |
1194 | #endif |
1195 | ||
098fb9db | 1196 | static int |
64b9e029 | 1197 | wake_affine(struct sched_domain *this_sd, struct rq *this_rq, |
4ae7d5ce IM |
1198 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, |
1199 | int idx, unsigned long load, unsigned long this_load, | |
098fb9db IM |
1200 | unsigned int imbalance) |
1201 | { | |
fc631c82 PZ |
1202 | struct task_struct *curr = this_rq->curr; |
1203 | struct task_group *tg; | |
098fb9db IM |
1204 | unsigned long tl = this_load; |
1205 | unsigned long tl_per_task; | |
83378269 | 1206 | unsigned long weight; |
b3137bc8 | 1207 | int balanced; |
098fb9db | 1208 | |
b3137bc8 | 1209 | if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS)) |
098fb9db IM |
1210 | return 0; |
1211 | ||
fc631c82 PZ |
1212 | if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost || |
1213 | p->se.avg_overlap > sysctl_sched_migration_cost)) | |
1214 | sync = 0; | |
1215 | ||
b3137bc8 MG |
1216 | /* |
1217 | * If sync wakeup then subtract the (maximum possible) | |
1218 | * effect of the currently running task from the load | |
1219 | * of the current CPU: | |
1220 | */ | |
83378269 PZ |
1221 | if (sync) { |
1222 | tg = task_group(current); | |
1223 | weight = current->se.load.weight; | |
1224 | ||
1225 | tl += effective_load(tg, this_cpu, -weight, -weight); | |
1226 | load += effective_load(tg, prev_cpu, 0, -weight); | |
1227 | } | |
b3137bc8 | 1228 | |
83378269 PZ |
1229 | tg = task_group(p); |
1230 | weight = p->se.load.weight; | |
b3137bc8 | 1231 | |
83378269 PZ |
1232 | balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <= |
1233 | imbalance*(load + effective_load(tg, prev_cpu, 0, weight)); | |
b3137bc8 | 1234 | |
098fb9db | 1235 | /* |
4ae7d5ce IM |
1236 | * If the currently running task will sleep within |
1237 | * a reasonable amount of time then attract this newly | |
1238 | * woken task: | |
098fb9db | 1239 | */ |
2fb7635c PZ |
1240 | if (sync && balanced) |
1241 | return 1; | |
098fb9db IM |
1242 | |
1243 | schedstat_inc(p, se.nr_wakeups_affine_attempts); | |
1244 | tl_per_task = cpu_avg_load_per_task(this_cpu); | |
1245 | ||
64b9e029 AA |
1246 | if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <= |
1247 | tl_per_task)) { | |
098fb9db IM |
1248 | /* |
1249 | * This domain has SD_WAKE_AFFINE and | |
1250 | * p is cache cold in this domain, and | |
1251 | * there is no bad imbalance. | |
1252 | */ | |
1253 | schedstat_inc(this_sd, ttwu_move_affine); | |
1254 | schedstat_inc(p, se.nr_wakeups_affine); | |
1255 | ||
1256 | return 1; | |
1257 | } | |
1258 | return 0; | |
1259 | } | |
1260 | ||
e7693a36 GH |
1261 | static int select_task_rq_fair(struct task_struct *p, int sync) |
1262 | { | |
e7693a36 | 1263 | struct sched_domain *sd, *this_sd = NULL; |
ac192d39 | 1264 | int prev_cpu, this_cpu, new_cpu; |
098fb9db | 1265 | unsigned long load, this_load; |
64b9e029 | 1266 | struct rq *this_rq; |
098fb9db | 1267 | unsigned int imbalance; |
098fb9db | 1268 | int idx; |
e7693a36 | 1269 | |
ac192d39 | 1270 | prev_cpu = task_cpu(p); |
ac192d39 | 1271 | this_cpu = smp_processor_id(); |
4ae7d5ce | 1272 | this_rq = cpu_rq(this_cpu); |
ac192d39 | 1273 | new_cpu = prev_cpu; |
e7693a36 | 1274 | |
64b9e029 AA |
1275 | if (prev_cpu == this_cpu) |
1276 | goto out; | |
ac192d39 IM |
1277 | /* |
1278 | * 'this_sd' is the first domain that both | |
1279 | * this_cpu and prev_cpu are present in: | |
1280 | */ | |
e7693a36 | 1281 | for_each_domain(this_cpu, sd) { |
758b2cdc | 1282 | if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) { |
e7693a36 GH |
1283 | this_sd = sd; |
1284 | break; | |
1285 | } | |
1286 | } | |
1287 | ||
96f874e2 | 1288 | if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed))) |
f4827386 | 1289 | goto out; |
e7693a36 GH |
1290 | |
1291 | /* | |
1292 | * Check for affine wakeup and passive balancing possibilities. | |
1293 | */ | |
098fb9db | 1294 | if (!this_sd) |
f4827386 | 1295 | goto out; |
e7693a36 | 1296 | |
098fb9db IM |
1297 | idx = this_sd->wake_idx; |
1298 | ||
1299 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; | |
1300 | ||
ac192d39 | 1301 | load = source_load(prev_cpu, idx); |
098fb9db IM |
1302 | this_load = target_load(this_cpu, idx); |
1303 | ||
64b9e029 | 1304 | if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, |
4ae7d5ce IM |
1305 | load, this_load, imbalance)) |
1306 | return this_cpu; | |
1307 | ||
098fb9db IM |
1308 | /* |
1309 | * Start passive balancing when half the imbalance_pct | |
1310 | * limit is reached. | |
1311 | */ | |
1312 | if (this_sd->flags & SD_WAKE_BALANCE) { | |
1313 | if (imbalance*this_load <= 100*load) { | |
1314 | schedstat_inc(this_sd, ttwu_move_balance); | |
1315 | schedstat_inc(p, se.nr_wakeups_passive); | |
4ae7d5ce | 1316 | return this_cpu; |
e7693a36 GH |
1317 | } |
1318 | } | |
1319 | ||
f4827386 | 1320 | out: |
e7693a36 GH |
1321 | return wake_idle(new_cpu, p); |
1322 | } | |
1323 | #endif /* CONFIG_SMP */ | |
1324 | ||
e52fb7c0 PZ |
1325 | /* |
1326 | * Adaptive granularity | |
1327 | * | |
1328 | * se->avg_wakeup gives the average time a task runs until it does a wakeup, | |
1329 | * with the limit of wakeup_gran -- when it never does a wakeup. | |
1330 | * | |
1331 | * So the smaller avg_wakeup is the faster we want this task to preempt, | |
1332 | * but we don't want to treat the preemptee unfairly and therefore allow it | |
1333 | * to run for at least the amount of time we'd like to run. | |
1334 | * | |
1335 | * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one | |
1336 | * | |
1337 | * NOTE: we use *nr_running to scale with load, this nicely matches the | |
1338 | * degrading latency on load. | |
1339 | */ | |
1340 | static unsigned long | |
1341 | adaptive_gran(struct sched_entity *curr, struct sched_entity *se) | |
1342 | { | |
1343 | u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | |
1344 | u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running; | |
1345 | u64 gran = 0; | |
1346 | ||
1347 | if (this_run < expected_wakeup) | |
1348 | gran = expected_wakeup - this_run; | |
1349 | ||
1350 | return min_t(s64, gran, sysctl_sched_wakeup_granularity); | |
1351 | } | |
1352 | ||
1353 | static unsigned long | |
1354 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1355 | { |
1356 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1357 | ||
e52fb7c0 PZ |
1358 | if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN)) |
1359 | gran = adaptive_gran(curr, se); | |
1360 | ||
0bbd3336 | 1361 | /* |
e52fb7c0 PZ |
1362 | * Since its curr running now, convert the gran from real-time |
1363 | * to virtual-time in his units. | |
0bbd3336 | 1364 | */ |
e52fb7c0 PZ |
1365 | if (sched_feat(ASYM_GRAN)) { |
1366 | /* | |
1367 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1368 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1369 | * the resulting gran will be larger, therefore penalizing the | |
1370 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1371 | * be smaller, again penalizing the lighter task. | |
1372 | * | |
1373 | * This is especially important for buddies when the leftmost | |
1374 | * task is higher priority than the buddy. | |
1375 | */ | |
1376 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
1377 | gran = calc_delta_fair(gran, se); | |
1378 | } else { | |
1379 | if (unlikely(curr->load.weight != NICE_0_LOAD)) | |
1380 | gran = calc_delta_fair(gran, curr); | |
1381 | } | |
0bbd3336 PZ |
1382 | |
1383 | return gran; | |
1384 | } | |
1385 | ||
464b7527 PZ |
1386 | /* |
1387 | * Should 'se' preempt 'curr'. | |
1388 | * | |
1389 | * |s1 | |
1390 | * |s2 | |
1391 | * |s3 | |
1392 | * g | |
1393 | * |<--->|c | |
1394 | * | |
1395 | * w(c, s1) = -1 | |
1396 | * w(c, s2) = 0 | |
1397 | * w(c, s3) = 1 | |
1398 | * | |
1399 | */ | |
1400 | static int | |
1401 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1402 | { | |
1403 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1404 | ||
1405 | if (vdiff <= 0) | |
1406 | return -1; | |
1407 | ||
e52fb7c0 | 1408 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1409 | if (vdiff > gran) |
1410 | return 1; | |
1411 | ||
1412 | return 0; | |
1413 | } | |
1414 | ||
02479099 PZ |
1415 | static void set_last_buddy(struct sched_entity *se) |
1416 | { | |
6bc912b7 PZ |
1417 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1418 | for_each_sched_entity(se) | |
1419 | cfs_rq_of(se)->last = se; | |
1420 | } | |
02479099 PZ |
1421 | } |
1422 | ||
1423 | static void set_next_buddy(struct sched_entity *se) | |
1424 | { | |
6bc912b7 PZ |
1425 | if (likely(task_of(se)->policy != SCHED_IDLE)) { |
1426 | for_each_sched_entity(se) | |
1427 | cfs_rq_of(se)->next = se; | |
1428 | } | |
02479099 PZ |
1429 | } |
1430 | ||
bf0f6f24 IM |
1431 | /* |
1432 | * Preempt the current task with a newly woken task if needed: | |
1433 | */ | |
15afe09b | 1434 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync) |
bf0f6f24 IM |
1435 | { |
1436 | struct task_struct *curr = rq->curr; | |
8651a86c | 1437 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1438 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
bf0f6f24 | 1439 | |
03e89e45 | 1440 | update_curr(cfs_rq); |
4793241b | 1441 | |
03e89e45 | 1442 | if (unlikely(rt_prio(p->prio))) { |
bf0f6f24 IM |
1443 | resched_task(curr); |
1444 | return; | |
1445 | } | |
aa2ac252 | 1446 | |
d95f98d0 PZ |
1447 | if (unlikely(p->sched_class != &fair_sched_class)) |
1448 | return; | |
1449 | ||
4ae7d5ce IM |
1450 | if (unlikely(se == pse)) |
1451 | return; | |
1452 | ||
4793241b PZ |
1453 | /* |
1454 | * Only set the backward buddy when the current task is still on the | |
1455 | * rq. This can happen when a wakeup gets interleaved with schedule on | |
1456 | * the ->pre_schedule() or idle_balance() point, either of which can | |
1457 | * drop the rq lock. | |
1458 | * | |
1459 | * Also, during early boot the idle thread is in the fair class, for | |
1460 | * obvious reasons its a bad idea to schedule back to the idle thread. | |
1461 | */ | |
1462 | if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle)) | |
02479099 PZ |
1463 | set_last_buddy(se); |
1464 | set_next_buddy(pse); | |
57fdc26d | 1465 | |
aec0a514 BR |
1466 | /* |
1467 | * We can come here with TIF_NEED_RESCHED already set from new task | |
1468 | * wake up path. | |
1469 | */ | |
1470 | if (test_tsk_need_resched(curr)) | |
1471 | return; | |
1472 | ||
91c234b4 | 1473 | /* |
6bc912b7 | 1474 | * Batch and idle tasks do not preempt (their preemption is driven by |
91c234b4 IM |
1475 | * the tick): |
1476 | */ | |
6bc912b7 | 1477 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 1478 | return; |
bf0f6f24 | 1479 | |
6bc912b7 PZ |
1480 | /* Idle tasks are by definition preempted by everybody. */ |
1481 | if (unlikely(curr->policy == SCHED_IDLE)) { | |
1482 | resched_task(curr); | |
91c234b4 | 1483 | return; |
6bc912b7 | 1484 | } |
bf0f6f24 | 1485 | |
77d9cc44 IM |
1486 | if (!sched_feat(WAKEUP_PREEMPT)) |
1487 | return; | |
8651a86c | 1488 | |
fc631c82 PZ |
1489 | if (sched_feat(WAKEUP_OVERLAP) && (sync || |
1490 | (se->avg_overlap < sysctl_sched_migration_cost && | |
1491 | pse->avg_overlap < sysctl_sched_migration_cost))) { | |
15afe09b PZ |
1492 | resched_task(curr); |
1493 | return; | |
1494 | } | |
1495 | ||
464b7527 PZ |
1496 | find_matching_se(&se, &pse); |
1497 | ||
002f128b | 1498 | BUG_ON(!pse); |
464b7527 | 1499 | |
002f128b PT |
1500 | if (wakeup_preempt_entity(se, pse) == 1) |
1501 | resched_task(curr); | |
bf0f6f24 IM |
1502 | } |
1503 | ||
fb8d4724 | 1504 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 1505 | { |
8f4d37ec | 1506 | struct task_struct *p; |
bf0f6f24 IM |
1507 | struct cfs_rq *cfs_rq = &rq->cfs; |
1508 | struct sched_entity *se; | |
1509 | ||
1510 | if (unlikely(!cfs_rq->nr_running)) | |
1511 | return NULL; | |
1512 | ||
1513 | do { | |
9948f4b2 | 1514 | se = pick_next_entity(cfs_rq); |
a9f3e2b5 MG |
1515 | /* |
1516 | * If se was a buddy, clear it so that it will have to earn | |
1517 | * the favour again. | |
1518 | */ | |
a571bbea | 1519 | __clear_buddies(cfs_rq, se); |
f4b6755f | 1520 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
1521 | cfs_rq = group_cfs_rq(se); |
1522 | } while (cfs_rq); | |
1523 | ||
8f4d37ec PZ |
1524 | p = task_of(se); |
1525 | hrtick_start_fair(rq, p); | |
1526 | ||
1527 | return p; | |
bf0f6f24 IM |
1528 | } |
1529 | ||
1530 | /* | |
1531 | * Account for a descheduled task: | |
1532 | */ | |
31ee529c | 1533 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
1534 | { |
1535 | struct sched_entity *se = &prev->se; | |
1536 | struct cfs_rq *cfs_rq; | |
1537 | ||
1538 | for_each_sched_entity(se) { | |
1539 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 1540 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
1541 | } |
1542 | } | |
1543 | ||
681f3e68 | 1544 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
1545 | /************************************************** |
1546 | * Fair scheduling class load-balancing methods: | |
1547 | */ | |
1548 | ||
1549 | /* | |
1550 | * Load-balancing iterator. Note: while the runqueue stays locked | |
1551 | * during the whole iteration, the current task might be | |
1552 | * dequeued so the iterator has to be dequeue-safe. Here we | |
1553 | * achieve that by always pre-iterating before returning | |
1554 | * the current task: | |
1555 | */ | |
a9957449 | 1556 | static struct task_struct * |
4a55bd5e | 1557 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next) |
bf0f6f24 | 1558 | { |
354d60c2 DG |
1559 | struct task_struct *p = NULL; |
1560 | struct sched_entity *se; | |
bf0f6f24 | 1561 | |
77ae6513 MG |
1562 | if (next == &cfs_rq->tasks) |
1563 | return NULL; | |
1564 | ||
b87f1724 BR |
1565 | se = list_entry(next, struct sched_entity, group_node); |
1566 | p = task_of(se); | |
1567 | cfs_rq->balance_iterator = next->next; | |
77ae6513 | 1568 | |
bf0f6f24 IM |
1569 | return p; |
1570 | } | |
1571 | ||
1572 | static struct task_struct *load_balance_start_fair(void *arg) | |
1573 | { | |
1574 | struct cfs_rq *cfs_rq = arg; | |
1575 | ||
4a55bd5e | 1576 | return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next); |
bf0f6f24 IM |
1577 | } |
1578 | ||
1579 | static struct task_struct *load_balance_next_fair(void *arg) | |
1580 | { | |
1581 | struct cfs_rq *cfs_rq = arg; | |
1582 | ||
4a55bd5e | 1583 | return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator); |
bf0f6f24 IM |
1584 | } |
1585 | ||
c09595f6 PZ |
1586 | static unsigned long |
1587 | __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1588 | unsigned long max_load_move, struct sched_domain *sd, | |
1589 | enum cpu_idle_type idle, int *all_pinned, int *this_best_prio, | |
1590 | struct cfs_rq *cfs_rq) | |
62fb1851 | 1591 | { |
c09595f6 | 1592 | struct rq_iterator cfs_rq_iterator; |
62fb1851 | 1593 | |
c09595f6 PZ |
1594 | cfs_rq_iterator.start = load_balance_start_fair; |
1595 | cfs_rq_iterator.next = load_balance_next_fair; | |
1596 | cfs_rq_iterator.arg = cfs_rq; | |
62fb1851 | 1597 | |
c09595f6 PZ |
1598 | return balance_tasks(this_rq, this_cpu, busiest, |
1599 | max_load_move, sd, idle, all_pinned, | |
1600 | this_best_prio, &cfs_rq_iterator); | |
62fb1851 | 1601 | } |
62fb1851 | 1602 | |
c09595f6 | 1603 | #ifdef CONFIG_FAIR_GROUP_SCHED |
43010659 | 1604 | static unsigned long |
bf0f6f24 | 1605 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f | 1606 | unsigned long max_load_move, |
a4ac01c3 PW |
1607 | struct sched_domain *sd, enum cpu_idle_type idle, |
1608 | int *all_pinned, int *this_best_prio) | |
bf0f6f24 | 1609 | { |
bf0f6f24 | 1610 | long rem_load_move = max_load_move; |
c09595f6 PZ |
1611 | int busiest_cpu = cpu_of(busiest); |
1612 | struct task_group *tg; | |
18d95a28 | 1613 | |
c09595f6 | 1614 | rcu_read_lock(); |
c8cba857 | 1615 | update_h_load(busiest_cpu); |
18d95a28 | 1616 | |
caea8a03 | 1617 | list_for_each_entry_rcu(tg, &task_groups, list) { |
c8cba857 | 1618 | struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu]; |
42a3ac7d PZ |
1619 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
1620 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
243e0e7b | 1621 | u64 rem_load, moved_load; |
18d95a28 | 1622 | |
c09595f6 PZ |
1623 | /* |
1624 | * empty group | |
1625 | */ | |
c8cba857 | 1626 | if (!busiest_cfs_rq->task_weight) |
bf0f6f24 IM |
1627 | continue; |
1628 | ||
243e0e7b SV |
1629 | rem_load = (u64)rem_load_move * busiest_weight; |
1630 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
bf0f6f24 | 1631 | |
c09595f6 | 1632 | moved_load = __load_balance_fair(this_rq, this_cpu, busiest, |
53fecd8a | 1633 | rem_load, sd, idle, all_pinned, this_best_prio, |
c09595f6 | 1634 | tg->cfs_rq[busiest_cpu]); |
bf0f6f24 | 1635 | |
c09595f6 | 1636 | if (!moved_load) |
bf0f6f24 IM |
1637 | continue; |
1638 | ||
42a3ac7d | 1639 | moved_load *= busiest_h_load; |
243e0e7b | 1640 | moved_load = div_u64(moved_load, busiest_weight + 1); |
bf0f6f24 | 1641 | |
c09595f6 PZ |
1642 | rem_load_move -= moved_load; |
1643 | if (rem_load_move < 0) | |
bf0f6f24 IM |
1644 | break; |
1645 | } | |
c09595f6 | 1646 | rcu_read_unlock(); |
bf0f6f24 | 1647 | |
43010659 | 1648 | return max_load_move - rem_load_move; |
bf0f6f24 | 1649 | } |
c09595f6 PZ |
1650 | #else |
1651 | static unsigned long | |
1652 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1653 | unsigned long max_load_move, | |
1654 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1655 | int *all_pinned, int *this_best_prio) | |
1656 | { | |
1657 | return __load_balance_fair(this_rq, this_cpu, busiest, | |
1658 | max_load_move, sd, idle, all_pinned, | |
1659 | this_best_prio, &busiest->cfs); | |
1660 | } | |
1661 | #endif | |
bf0f6f24 | 1662 | |
e1d1484f PW |
1663 | static int |
1664 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1665 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1666 | { | |
1667 | struct cfs_rq *busy_cfs_rq; | |
1668 | struct rq_iterator cfs_rq_iterator; | |
1669 | ||
1670 | cfs_rq_iterator.start = load_balance_start_fair; | |
1671 | cfs_rq_iterator.next = load_balance_next_fair; | |
1672 | ||
1673 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { | |
1674 | /* | |
1675 | * pass busy_cfs_rq argument into | |
1676 | * load_balance_[start|next]_fair iterators | |
1677 | */ | |
1678 | cfs_rq_iterator.arg = busy_cfs_rq; | |
1679 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, | |
1680 | &cfs_rq_iterator)) | |
1681 | return 1; | |
1682 | } | |
1683 | ||
1684 | return 0; | |
1685 | } | |
55e12e5e | 1686 | #endif /* CONFIG_SMP */ |
e1d1484f | 1687 | |
bf0f6f24 IM |
1688 | /* |
1689 | * scheduler tick hitting a task of our scheduling class: | |
1690 | */ | |
8f4d37ec | 1691 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
1692 | { |
1693 | struct cfs_rq *cfs_rq; | |
1694 | struct sched_entity *se = &curr->se; | |
1695 | ||
1696 | for_each_sched_entity(se) { | |
1697 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 1698 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
1699 | } |
1700 | } | |
1701 | ||
1702 | /* | |
1703 | * Share the fairness runtime between parent and child, thus the | |
1704 | * total amount of pressure for CPU stays equal - new tasks | |
1705 | * get a chance to run but frequent forkers are not allowed to | |
1706 | * monopolize the CPU. Note: the parent runqueue is locked, | |
1707 | * the child is not running yet. | |
1708 | */ | |
ee0827d8 | 1709 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
bf0f6f24 IM |
1710 | { |
1711 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
429d43bc | 1712 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 1713 | int this_cpu = smp_processor_id(); |
bf0f6f24 IM |
1714 | |
1715 | sched_info_queued(p); | |
1716 | ||
7109c442 | 1717 | update_curr(cfs_rq); |
aeb73b04 | 1718 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 1719 | |
3c90e6e9 | 1720 | /* 'curr' will be NULL if the child belongs to a different group */ |
00bf7bfc | 1721 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
54fdc581 | 1722 | curr && entity_before(curr, se)) { |
87fefa38 | 1723 | /* |
edcb60a3 IM |
1724 | * Upon rescheduling, sched_class::put_prev_task() will place |
1725 | * 'current' within the tree based on its new key value. | |
1726 | */ | |
4d78e7b6 | 1727 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 1728 | resched_task(rq->curr); |
4d78e7b6 | 1729 | } |
bf0f6f24 | 1730 | |
b9dca1e0 | 1731 | enqueue_task_fair(rq, p, 0); |
bf0f6f24 IM |
1732 | } |
1733 | ||
cb469845 SR |
1734 | /* |
1735 | * Priority of the task has changed. Check to see if we preempt | |
1736 | * the current task. | |
1737 | */ | |
1738 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, | |
1739 | int oldprio, int running) | |
1740 | { | |
1741 | /* | |
1742 | * Reschedule if we are currently running on this runqueue and | |
1743 | * our priority decreased, or if we are not currently running on | |
1744 | * this runqueue and our priority is higher than the current's | |
1745 | */ | |
1746 | if (running) { | |
1747 | if (p->prio > oldprio) | |
1748 | resched_task(rq->curr); | |
1749 | } else | |
15afe09b | 1750 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1751 | } |
1752 | ||
1753 | /* | |
1754 | * We switched to the sched_fair class. | |
1755 | */ | |
1756 | static void switched_to_fair(struct rq *rq, struct task_struct *p, | |
1757 | int running) | |
1758 | { | |
1759 | /* | |
1760 | * We were most likely switched from sched_rt, so | |
1761 | * kick off the schedule if running, otherwise just see | |
1762 | * if we can still preempt the current task. | |
1763 | */ | |
1764 | if (running) | |
1765 | resched_task(rq->curr); | |
1766 | else | |
15afe09b | 1767 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
1768 | } |
1769 | ||
83b699ed SV |
1770 | /* Account for a task changing its policy or group. |
1771 | * | |
1772 | * This routine is mostly called to set cfs_rq->curr field when a task | |
1773 | * migrates between groups/classes. | |
1774 | */ | |
1775 | static void set_curr_task_fair(struct rq *rq) | |
1776 | { | |
1777 | struct sched_entity *se = &rq->curr->se; | |
1778 | ||
1779 | for_each_sched_entity(se) | |
1780 | set_next_entity(cfs_rq_of(se), se); | |
1781 | } | |
1782 | ||
810b3817 PZ |
1783 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1784 | static void moved_group_fair(struct task_struct *p) | |
1785 | { | |
1786 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
1787 | ||
1788 | update_curr(cfs_rq); | |
1789 | place_entity(cfs_rq, &p->se, 1); | |
1790 | } | |
1791 | #endif | |
1792 | ||
bf0f6f24 IM |
1793 | /* |
1794 | * All the scheduling class methods: | |
1795 | */ | |
5522d5d5 IM |
1796 | static const struct sched_class fair_sched_class = { |
1797 | .next = &idle_sched_class, | |
bf0f6f24 IM |
1798 | .enqueue_task = enqueue_task_fair, |
1799 | .dequeue_task = dequeue_task_fair, | |
1800 | .yield_task = yield_task_fair, | |
1801 | ||
2e09bf55 | 1802 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
1803 | |
1804 | .pick_next_task = pick_next_task_fair, | |
1805 | .put_prev_task = put_prev_task_fair, | |
1806 | ||
681f3e68 | 1807 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
1808 | .select_task_rq = select_task_rq_fair, |
1809 | ||
bf0f6f24 | 1810 | .load_balance = load_balance_fair, |
e1d1484f | 1811 | .move_one_task = move_one_task_fair, |
681f3e68 | 1812 | #endif |
bf0f6f24 | 1813 | |
83b699ed | 1814 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 IM |
1815 | .task_tick = task_tick_fair, |
1816 | .task_new = task_new_fair, | |
cb469845 SR |
1817 | |
1818 | .prio_changed = prio_changed_fair, | |
1819 | .switched_to = switched_to_fair, | |
810b3817 PZ |
1820 | |
1821 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1822 | .moved_group = moved_group_fair, | |
1823 | #endif | |
bf0f6f24 IM |
1824 | }; |
1825 | ||
1826 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 1827 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 1828 | { |
bf0f6f24 IM |
1829 | struct cfs_rq *cfs_rq; |
1830 | ||
5973e5b9 | 1831 | rcu_read_lock(); |
c3b64f1e | 1832 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 1833 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 1834 | rcu_read_unlock(); |
bf0f6f24 IM |
1835 | } |
1836 | #endif |