]>
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 | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
3436ae12 | 25 | #include <linux/cpumask.h> |
9745512c | 26 | |
bf0f6f24 | 27 | /* |
21805085 | 28 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 29 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 30 | * |
21805085 | 31 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
32 | * 'timeslice length' - timeslices in CFS are of variable length |
33 | * and have no persistent notion like in traditional, time-slice | |
34 | * based scheduling concepts. | |
bf0f6f24 | 35 | * |
d274a4ce IM |
36 | * (to see the precise effective timeslice length of your workload, |
37 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 38 | */ |
21406928 MG |
39 | unsigned int sysctl_sched_latency = 6000000ULL; |
40 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 41 | |
1983a922 CE |
42 | /* |
43 | * The initial- and re-scaling of tunables is configurable | |
44 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
45 | * | |
46 | * Options are: | |
47 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
48 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
49 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
50 | */ | |
51 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
52 | = SCHED_TUNABLESCALING_LOG; | |
53 | ||
2bd8e6d4 | 54 | /* |
b2be5e96 | 55 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 56 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 57 | */ |
0bf377bb IM |
58 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
59 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
60 | |
61 | /* | |
b2be5e96 PZ |
62 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
63 | */ | |
0bf377bb | 64 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
65 | |
66 | /* | |
2bba22c5 | 67 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 68 | * parent will (try to) run first. |
21805085 | 69 | */ |
2bba22c5 | 70 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 71 | |
bf0f6f24 IM |
72 | /* |
73 | * SCHED_OTHER wake-up granularity. | |
172e082a | 74 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
75 | * |
76 | * This option delays the preemption effects of decoupled workloads | |
77 | * and reduces their over-scheduling. Synchronous workloads will still | |
78 | * have immediate wakeup/sleep latencies. | |
79 | */ | |
172e082a | 80 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 81 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 82 | |
da84d961 IM |
83 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
84 | ||
a7a4f8a7 PT |
85 | /* |
86 | * The exponential sliding window over which load is averaged for shares | |
87 | * distribution. | |
88 | * (default: 10msec) | |
89 | */ | |
90 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
91 | ||
ec12cb7f PT |
92 | #ifdef CONFIG_CFS_BANDWIDTH |
93 | /* | |
94 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
95 | * each time a cfs_rq requests quota. | |
96 | * | |
97 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
98 | * to consumption or the quota being specified to be smaller than the slice) | |
99 | * we will always only issue the remaining available time. | |
100 | * | |
101 | * default: 5 msec, units: microseconds | |
102 | */ | |
103 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
104 | #endif | |
105 | ||
a4c2f00f PZ |
106 | static const struct sched_class fair_sched_class; |
107 | ||
bf0f6f24 IM |
108 | /************************************************************** |
109 | * CFS operations on generic schedulable entities: | |
110 | */ | |
111 | ||
62160e3f | 112 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 113 | |
62160e3f | 114 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
115 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
116 | { | |
62160e3f | 117 | return cfs_rq->rq; |
bf0f6f24 IM |
118 | } |
119 | ||
62160e3f IM |
120 | /* An entity is a task if it doesn't "own" a runqueue */ |
121 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 122 | |
8f48894f PZ |
123 | static inline struct task_struct *task_of(struct sched_entity *se) |
124 | { | |
125 | #ifdef CONFIG_SCHED_DEBUG | |
126 | WARN_ON_ONCE(!entity_is_task(se)); | |
127 | #endif | |
128 | return container_of(se, struct task_struct, se); | |
129 | } | |
130 | ||
b758149c PZ |
131 | /* Walk up scheduling entities hierarchy */ |
132 | #define for_each_sched_entity(se) \ | |
133 | for (; se; se = se->parent) | |
134 | ||
135 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
136 | { | |
137 | return p->se.cfs_rq; | |
138 | } | |
139 | ||
140 | /* runqueue on which this entity is (to be) queued */ | |
141 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
142 | { | |
143 | return se->cfs_rq; | |
144 | } | |
145 | ||
146 | /* runqueue "owned" by this group */ | |
147 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
148 | { | |
149 | return grp->my_q; | |
150 | } | |
151 | ||
3d4b47b4 PZ |
152 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
153 | { | |
154 | if (!cfs_rq->on_list) { | |
67e86250 PT |
155 | /* |
156 | * Ensure we either appear before our parent (if already | |
157 | * enqueued) or force our parent to appear after us when it is | |
158 | * enqueued. The fact that we always enqueue bottom-up | |
159 | * reduces this to two cases. | |
160 | */ | |
161 | if (cfs_rq->tg->parent && | |
162 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
163 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
164 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
165 | } else { | |
166 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 167 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 168 | } |
3d4b47b4 PZ |
169 | |
170 | cfs_rq->on_list = 1; | |
171 | } | |
172 | } | |
173 | ||
174 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
175 | { | |
176 | if (cfs_rq->on_list) { | |
177 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
178 | cfs_rq->on_list = 0; | |
179 | } | |
180 | } | |
181 | ||
b758149c PZ |
182 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
183 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
184 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
185 | ||
186 | /* Do the two (enqueued) entities belong to the same group ? */ | |
187 | static inline int | |
188 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
189 | { | |
190 | if (se->cfs_rq == pse->cfs_rq) | |
191 | return 1; | |
192 | ||
193 | return 0; | |
194 | } | |
195 | ||
196 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
197 | { | |
198 | return se->parent; | |
199 | } | |
200 | ||
464b7527 PZ |
201 | /* return depth at which a sched entity is present in the hierarchy */ |
202 | static inline int depth_se(struct sched_entity *se) | |
203 | { | |
204 | int depth = 0; | |
205 | ||
206 | for_each_sched_entity(se) | |
207 | depth++; | |
208 | ||
209 | return depth; | |
210 | } | |
211 | ||
212 | static void | |
213 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
214 | { | |
215 | int se_depth, pse_depth; | |
216 | ||
217 | /* | |
218 | * preemption test can be made between sibling entities who are in the | |
219 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
220 | * both tasks until we find their ancestors who are siblings of common | |
221 | * parent. | |
222 | */ | |
223 | ||
224 | /* First walk up until both entities are at same depth */ | |
225 | se_depth = depth_se(*se); | |
226 | pse_depth = depth_se(*pse); | |
227 | ||
228 | while (se_depth > pse_depth) { | |
229 | se_depth--; | |
230 | *se = parent_entity(*se); | |
231 | } | |
232 | ||
233 | while (pse_depth > se_depth) { | |
234 | pse_depth--; | |
235 | *pse = parent_entity(*pse); | |
236 | } | |
237 | ||
238 | while (!is_same_group(*se, *pse)) { | |
239 | *se = parent_entity(*se); | |
240 | *pse = parent_entity(*pse); | |
241 | } | |
242 | } | |
243 | ||
8f48894f PZ |
244 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
245 | ||
246 | static inline struct task_struct *task_of(struct sched_entity *se) | |
247 | { | |
248 | return container_of(se, struct task_struct, se); | |
249 | } | |
bf0f6f24 | 250 | |
62160e3f IM |
251 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
252 | { | |
253 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
254 | } |
255 | ||
256 | #define entity_is_task(se) 1 | |
257 | ||
b758149c PZ |
258 | #define for_each_sched_entity(se) \ |
259 | for (; se; se = NULL) | |
bf0f6f24 | 260 | |
b758149c | 261 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 262 | { |
b758149c | 263 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
264 | } |
265 | ||
b758149c PZ |
266 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
267 | { | |
268 | struct task_struct *p = task_of(se); | |
269 | struct rq *rq = task_rq(p); | |
270 | ||
271 | return &rq->cfs; | |
272 | } | |
273 | ||
274 | /* runqueue "owned" by this group */ | |
275 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
276 | { | |
277 | return NULL; | |
278 | } | |
279 | ||
3d4b47b4 PZ |
280 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
281 | { | |
282 | } | |
283 | ||
284 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
285 | { | |
286 | } | |
287 | ||
b758149c PZ |
288 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
289 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
290 | ||
291 | static inline int | |
292 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
293 | { | |
294 | return 1; | |
295 | } | |
296 | ||
297 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
298 | { | |
299 | return NULL; | |
300 | } | |
301 | ||
464b7527 PZ |
302 | static inline void |
303 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
304 | { | |
305 | } | |
306 | ||
b758149c PZ |
307 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
308 | ||
ec12cb7f PT |
309 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, |
310 | unsigned long delta_exec); | |
bf0f6f24 IM |
311 | |
312 | /************************************************************** | |
313 | * Scheduling class tree data structure manipulation methods: | |
314 | */ | |
315 | ||
0702e3eb | 316 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 317 | { |
368059a9 PZ |
318 | s64 delta = (s64)(vruntime - min_vruntime); |
319 | if (delta > 0) | |
02e0431a PZ |
320 | min_vruntime = vruntime; |
321 | ||
322 | return min_vruntime; | |
323 | } | |
324 | ||
0702e3eb | 325 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
326 | { |
327 | s64 delta = (s64)(vruntime - min_vruntime); | |
328 | if (delta < 0) | |
329 | min_vruntime = vruntime; | |
330 | ||
331 | return min_vruntime; | |
332 | } | |
333 | ||
54fdc581 FC |
334 | static inline int entity_before(struct sched_entity *a, |
335 | struct sched_entity *b) | |
336 | { | |
337 | return (s64)(a->vruntime - b->vruntime) < 0; | |
338 | } | |
339 | ||
1af5f730 PZ |
340 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
341 | { | |
342 | u64 vruntime = cfs_rq->min_vruntime; | |
343 | ||
344 | if (cfs_rq->curr) | |
345 | vruntime = cfs_rq->curr->vruntime; | |
346 | ||
347 | if (cfs_rq->rb_leftmost) { | |
348 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
349 | struct sched_entity, | |
350 | run_node); | |
351 | ||
e17036da | 352 | if (!cfs_rq->curr) |
1af5f730 PZ |
353 | vruntime = se->vruntime; |
354 | else | |
355 | vruntime = min_vruntime(vruntime, se->vruntime); | |
356 | } | |
357 | ||
358 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
3fe1698b PZ |
359 | #ifndef CONFIG_64BIT |
360 | smp_wmb(); | |
361 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
362 | #endif | |
1af5f730 PZ |
363 | } |
364 | ||
bf0f6f24 IM |
365 | /* |
366 | * Enqueue an entity into the rb-tree: | |
367 | */ | |
0702e3eb | 368 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
369 | { |
370 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
371 | struct rb_node *parent = NULL; | |
372 | struct sched_entity *entry; | |
bf0f6f24 IM |
373 | int leftmost = 1; |
374 | ||
375 | /* | |
376 | * Find the right place in the rbtree: | |
377 | */ | |
378 | while (*link) { | |
379 | parent = *link; | |
380 | entry = rb_entry(parent, struct sched_entity, run_node); | |
381 | /* | |
382 | * We dont care about collisions. Nodes with | |
383 | * the same key stay together. | |
384 | */ | |
2bd2d6f2 | 385 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
386 | link = &parent->rb_left; |
387 | } else { | |
388 | link = &parent->rb_right; | |
389 | leftmost = 0; | |
390 | } | |
391 | } | |
392 | ||
393 | /* | |
394 | * Maintain a cache of leftmost tree entries (it is frequently | |
395 | * used): | |
396 | */ | |
1af5f730 | 397 | if (leftmost) |
57cb499d | 398 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
399 | |
400 | rb_link_node(&se->run_node, parent, link); | |
401 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
402 | } |
403 | ||
0702e3eb | 404 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 405 | { |
3fe69747 PZ |
406 | if (cfs_rq->rb_leftmost == &se->run_node) { |
407 | struct rb_node *next_node; | |
3fe69747 PZ |
408 | |
409 | next_node = rb_next(&se->run_node); | |
410 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 411 | } |
e9acbff6 | 412 | |
bf0f6f24 | 413 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
414 | } |
415 | ||
ac53db59 | 416 | static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 417 | { |
f4b6755f PZ |
418 | struct rb_node *left = cfs_rq->rb_leftmost; |
419 | ||
420 | if (!left) | |
421 | return NULL; | |
422 | ||
423 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
424 | } |
425 | ||
ac53db59 RR |
426 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
427 | { | |
428 | struct rb_node *next = rb_next(&se->run_node); | |
429 | ||
430 | if (!next) | |
431 | return NULL; | |
432 | ||
433 | return rb_entry(next, struct sched_entity, run_node); | |
434 | } | |
435 | ||
436 | #ifdef CONFIG_SCHED_DEBUG | |
f4b6755f | 437 | static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 438 | { |
7eee3e67 | 439 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 440 | |
70eee74b BS |
441 | if (!last) |
442 | return NULL; | |
7eee3e67 IM |
443 | |
444 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
445 | } |
446 | ||
bf0f6f24 IM |
447 | /************************************************************** |
448 | * Scheduling class statistics methods: | |
449 | */ | |
450 | ||
acb4a848 | 451 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 452 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
453 | loff_t *ppos) |
454 | { | |
8d65af78 | 455 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 456 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
457 | |
458 | if (ret || !write) | |
459 | return ret; | |
460 | ||
461 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
462 | sysctl_sched_min_granularity); | |
463 | ||
acb4a848 CE |
464 | #define WRT_SYSCTL(name) \ |
465 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
466 | WRT_SYSCTL(sched_min_granularity); | |
467 | WRT_SYSCTL(sched_latency); | |
468 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
469 | #undef WRT_SYSCTL |
470 | ||
b2be5e96 PZ |
471 | return 0; |
472 | } | |
473 | #endif | |
647e7cac | 474 | |
a7be37ac | 475 | /* |
f9c0b095 | 476 | * delta /= w |
a7be37ac PZ |
477 | */ |
478 | static inline unsigned long | |
479 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
480 | { | |
f9c0b095 PZ |
481 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
482 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
483 | |
484 | return delta; | |
485 | } | |
486 | ||
647e7cac IM |
487 | /* |
488 | * The idea is to set a period in which each task runs once. | |
489 | * | |
490 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch | |
491 | * this period because otherwise the slices get too small. | |
492 | * | |
493 | * p = (nr <= nl) ? l : l*nr/nl | |
494 | */ | |
4d78e7b6 PZ |
495 | static u64 __sched_period(unsigned long nr_running) |
496 | { | |
497 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 498 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
499 | |
500 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 501 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 502 | period *= nr_running; |
4d78e7b6 PZ |
503 | } |
504 | ||
505 | return period; | |
506 | } | |
507 | ||
647e7cac IM |
508 | /* |
509 | * We calculate the wall-time slice from the period by taking a part | |
510 | * proportional to the weight. | |
511 | * | |
f9c0b095 | 512 | * s = p*P[w/rw] |
647e7cac | 513 | */ |
6d0f0ebd | 514 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 515 | { |
0a582440 | 516 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 517 | |
0a582440 | 518 | for_each_sched_entity(se) { |
6272d68c | 519 | struct load_weight *load; |
3104bf03 | 520 | struct load_weight lw; |
6272d68c LM |
521 | |
522 | cfs_rq = cfs_rq_of(se); | |
523 | load = &cfs_rq->load; | |
f9c0b095 | 524 | |
0a582440 | 525 | if (unlikely(!se->on_rq)) { |
3104bf03 | 526 | lw = cfs_rq->load; |
0a582440 MG |
527 | |
528 | update_load_add(&lw, se->load.weight); | |
529 | load = &lw; | |
530 | } | |
531 | slice = calc_delta_mine(slice, se->load.weight, load); | |
532 | } | |
533 | return slice; | |
bf0f6f24 IM |
534 | } |
535 | ||
647e7cac | 536 | /* |
ac884dec | 537 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 538 | * |
f9c0b095 | 539 | * vs = s/w |
647e7cac | 540 | */ |
f9c0b095 | 541 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 542 | { |
f9c0b095 | 543 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
544 | } |
545 | ||
d6b55918 | 546 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); |
6d5ab293 | 547 | static void update_cfs_shares(struct cfs_rq *cfs_rq); |
3b3d190e | 548 | |
bf0f6f24 IM |
549 | /* |
550 | * Update the current task's runtime statistics. Skip current tasks that | |
551 | * are not in our scheduling class. | |
552 | */ | |
553 | static inline void | |
8ebc91d9 IM |
554 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
555 | unsigned long delta_exec) | |
bf0f6f24 | 556 | { |
bbdba7c0 | 557 | unsigned long delta_exec_weighted; |
bf0f6f24 | 558 | |
41acab88 LDM |
559 | schedstat_set(curr->statistics.exec_max, |
560 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
561 | |
562 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 563 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 564 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 565 | |
e9acbff6 | 566 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 567 | update_min_vruntime(cfs_rq); |
3b3d190e | 568 | |
70caf8a6 | 569 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
3b3d190e | 570 | cfs_rq->load_unacc_exec_time += delta_exec; |
3b3d190e | 571 | #endif |
bf0f6f24 IM |
572 | } |
573 | ||
b7cc0896 | 574 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 575 | { |
429d43bc | 576 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 577 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
578 | unsigned long delta_exec; |
579 | ||
580 | if (unlikely(!curr)) | |
581 | return; | |
582 | ||
583 | /* | |
584 | * Get the amount of time the current task was running | |
585 | * since the last time we changed load (this cannot | |
586 | * overflow on 32 bits): | |
587 | */ | |
8ebc91d9 | 588 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
589 | if (!delta_exec) |
590 | return; | |
bf0f6f24 | 591 | |
8ebc91d9 IM |
592 | __update_curr(cfs_rq, curr, delta_exec); |
593 | curr->exec_start = now; | |
d842de87 SV |
594 | |
595 | if (entity_is_task(curr)) { | |
596 | struct task_struct *curtask = task_of(curr); | |
597 | ||
f977bb49 | 598 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 599 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 600 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 601 | } |
ec12cb7f PT |
602 | |
603 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
604 | } |
605 | ||
606 | static inline void | |
5870db5b | 607 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 608 | { |
41acab88 | 609 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
610 | } |
611 | ||
bf0f6f24 IM |
612 | /* |
613 | * Task is being enqueued - update stats: | |
614 | */ | |
d2417e5a | 615 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 616 | { |
bf0f6f24 IM |
617 | /* |
618 | * Are we enqueueing a waiting task? (for current tasks | |
619 | * a dequeue/enqueue event is a NOP) | |
620 | */ | |
429d43bc | 621 | if (se != cfs_rq->curr) |
5870db5b | 622 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
623 | } |
624 | ||
bf0f6f24 | 625 | static void |
9ef0a961 | 626 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 627 | { |
41acab88 LDM |
628 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
629 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
630 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
631 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
632 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
633 | #ifdef CONFIG_SCHEDSTATS |
634 | if (entity_is_task(se)) { | |
635 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 636 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
637 | } |
638 | #endif | |
41acab88 | 639 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
640 | } |
641 | ||
642 | static inline void | |
19b6a2e3 | 643 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 644 | { |
bf0f6f24 IM |
645 | /* |
646 | * Mark the end of the wait period if dequeueing a | |
647 | * waiting task: | |
648 | */ | |
429d43bc | 649 | if (se != cfs_rq->curr) |
9ef0a961 | 650 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
651 | } |
652 | ||
653 | /* | |
654 | * We are picking a new current task - update its stats: | |
655 | */ | |
656 | static inline void | |
79303e9e | 657 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
658 | { |
659 | /* | |
660 | * We are starting a new run period: | |
661 | */ | |
305e6835 | 662 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
663 | } |
664 | ||
bf0f6f24 IM |
665 | /************************************************** |
666 | * Scheduling class queueing methods: | |
667 | */ | |
668 | ||
c09595f6 PZ |
669 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
670 | static void | |
671 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
672 | { | |
673 | cfs_rq->task_weight += weight; | |
674 | } | |
675 | #else | |
676 | static inline void | |
677 | add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) | |
678 | { | |
679 | } | |
680 | #endif | |
681 | ||
30cfdcfc DA |
682 | static void |
683 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
684 | { | |
685 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
686 | if (!parent_entity(se)) |
687 | inc_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 688 | if (entity_is_task(se)) { |
c09595f6 | 689 | add_cfs_task_weight(cfs_rq, se->load.weight); |
b87f1724 BR |
690 | list_add(&se->group_node, &cfs_rq->tasks); |
691 | } | |
30cfdcfc | 692 | cfs_rq->nr_running++; |
30cfdcfc DA |
693 | } |
694 | ||
695 | static void | |
696 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
697 | { | |
698 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 PZ |
699 | if (!parent_entity(se)) |
700 | dec_cpu_load(rq_of(cfs_rq), se->load.weight); | |
b87f1724 | 701 | if (entity_is_task(se)) { |
c09595f6 | 702 | add_cfs_task_weight(cfs_rq, -se->load.weight); |
b87f1724 BR |
703 | list_del_init(&se->group_node); |
704 | } | |
30cfdcfc | 705 | cfs_rq->nr_running--; |
30cfdcfc DA |
706 | } |
707 | ||
3ff6dcac YZ |
708 | #ifdef CONFIG_FAIR_GROUP_SCHED |
709 | # ifdef CONFIG_SMP | |
d6b55918 PT |
710 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
711 | int global_update) | |
712 | { | |
713 | struct task_group *tg = cfs_rq->tg; | |
714 | long load_avg; | |
715 | ||
716 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
717 | load_avg -= cfs_rq->load_contribution; | |
718 | ||
719 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
720 | atomic_add(load_avg, &tg->load_weight); | |
721 | cfs_rq->load_contribution += load_avg; | |
722 | } | |
723 | } | |
724 | ||
725 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 726 | { |
a7a4f8a7 | 727 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 728 | u64 now, delta; |
e33078ba | 729 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 730 | |
b815f196 | 731 | if (cfs_rq->tg == &root_task_group) |
2069dd75 PZ |
732 | return; |
733 | ||
05ca62c6 | 734 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
735 | delta = now - cfs_rq->load_stamp; |
736 | ||
e33078ba PT |
737 | /* truncate load history at 4 idle periods */ |
738 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
739 | now - cfs_rq->load_last > 4 * period) { | |
740 | cfs_rq->load_period = 0; | |
741 | cfs_rq->load_avg = 0; | |
f07333bf | 742 | delta = period - 1; |
e33078ba PT |
743 | } |
744 | ||
2069dd75 | 745 | cfs_rq->load_stamp = now; |
3b3d190e | 746 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 747 | cfs_rq->load_period += delta; |
e33078ba PT |
748 | if (load) { |
749 | cfs_rq->load_last = now; | |
750 | cfs_rq->load_avg += delta * load; | |
751 | } | |
2069dd75 | 752 | |
d6b55918 PT |
753 | /* consider updating load contribution on each fold or truncate */ |
754 | if (global_update || cfs_rq->load_period > period | |
755 | || !cfs_rq->load_period) | |
756 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
757 | ||
2069dd75 PZ |
758 | while (cfs_rq->load_period > period) { |
759 | /* | |
760 | * Inline assembly required to prevent the compiler | |
761 | * optimising this loop into a divmod call. | |
762 | * See __iter_div_u64_rem() for another example of this. | |
763 | */ | |
764 | asm("" : "+rm" (cfs_rq->load_period)); | |
765 | cfs_rq->load_period /= 2; | |
766 | cfs_rq->load_avg /= 2; | |
767 | } | |
3d4b47b4 | 768 | |
e33078ba PT |
769 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
770 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
771 | } |
772 | ||
6d5ab293 | 773 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
774 | { |
775 | long load_weight, load, shares; | |
776 | ||
6d5ab293 | 777 | load = cfs_rq->load.weight; |
3ff6dcac YZ |
778 | |
779 | load_weight = atomic_read(&tg->load_weight); | |
3ff6dcac | 780 | load_weight += load; |
6d5ab293 | 781 | load_weight -= cfs_rq->load_contribution; |
3ff6dcac YZ |
782 | |
783 | shares = (tg->shares * load); | |
784 | if (load_weight) | |
785 | shares /= load_weight; | |
786 | ||
787 | if (shares < MIN_SHARES) | |
788 | shares = MIN_SHARES; | |
789 | if (shares > tg->shares) | |
790 | shares = tg->shares; | |
791 | ||
792 | return shares; | |
793 | } | |
794 | ||
795 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
796 | { | |
797 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
798 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 799 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
800 | } |
801 | } | |
802 | # else /* CONFIG_SMP */ | |
803 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
804 | { | |
805 | } | |
806 | ||
6d5ab293 | 807 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
808 | { |
809 | return tg->shares; | |
810 | } | |
811 | ||
812 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
813 | { | |
814 | } | |
815 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
816 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
817 | unsigned long weight) | |
818 | { | |
19e5eebb PT |
819 | if (se->on_rq) { |
820 | /* commit outstanding execution time */ | |
821 | if (cfs_rq->curr == se) | |
822 | update_curr(cfs_rq); | |
2069dd75 | 823 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 824 | } |
2069dd75 PZ |
825 | |
826 | update_load_set(&se->load, weight); | |
827 | ||
828 | if (se->on_rq) | |
829 | account_entity_enqueue(cfs_rq, se); | |
830 | } | |
831 | ||
6d5ab293 | 832 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
833 | { |
834 | struct task_group *tg; | |
835 | struct sched_entity *se; | |
3ff6dcac | 836 | long shares; |
2069dd75 | 837 | |
2069dd75 PZ |
838 | tg = cfs_rq->tg; |
839 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
840 | if (!se) | |
841 | return; | |
3ff6dcac YZ |
842 | #ifndef CONFIG_SMP |
843 | if (likely(se->load.weight == tg->shares)) | |
844 | return; | |
845 | #endif | |
6d5ab293 | 846 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
847 | |
848 | reweight_entity(cfs_rq_of(se), se, shares); | |
849 | } | |
850 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 851 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
852 | { |
853 | } | |
854 | ||
6d5ab293 | 855 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
856 | { |
857 | } | |
43365bd7 PT |
858 | |
859 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
860 | { | |
861 | } | |
2069dd75 PZ |
862 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
863 | ||
2396af69 | 864 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 865 | { |
bf0f6f24 | 866 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
867 | struct task_struct *tsk = NULL; |
868 | ||
869 | if (entity_is_task(se)) | |
870 | tsk = task_of(se); | |
871 | ||
41acab88 LDM |
872 | if (se->statistics.sleep_start) { |
873 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
874 | |
875 | if ((s64)delta < 0) | |
876 | delta = 0; | |
877 | ||
41acab88 LDM |
878 | if (unlikely(delta > se->statistics.sleep_max)) |
879 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 880 | |
41acab88 LDM |
881 | se->statistics.sleep_start = 0; |
882 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 883 | |
768d0c27 | 884 | if (tsk) { |
e414314c | 885 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
886 | trace_sched_stat_sleep(tsk, delta); |
887 | } | |
bf0f6f24 | 888 | } |
41acab88 LDM |
889 | if (se->statistics.block_start) { |
890 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
891 | |
892 | if ((s64)delta < 0) | |
893 | delta = 0; | |
894 | ||
41acab88 LDM |
895 | if (unlikely(delta > se->statistics.block_max)) |
896 | se->statistics.block_max = delta; | |
bf0f6f24 | 897 | |
41acab88 LDM |
898 | se->statistics.block_start = 0; |
899 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 900 | |
e414314c | 901 | if (tsk) { |
8f0dfc34 | 902 | if (tsk->in_iowait) { |
41acab88 LDM |
903 | se->statistics.iowait_sum += delta; |
904 | se->statistics.iowait_count++; | |
768d0c27 | 905 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
906 | } |
907 | ||
e414314c PZ |
908 | /* |
909 | * Blocking time is in units of nanosecs, so shift by | |
910 | * 20 to get a milliseconds-range estimation of the | |
911 | * amount of time that the task spent sleeping: | |
912 | */ | |
913 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
914 | profile_hits(SLEEP_PROFILING, | |
915 | (void *)get_wchan(tsk), | |
916 | delta >> 20); | |
917 | } | |
918 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 919 | } |
bf0f6f24 IM |
920 | } |
921 | #endif | |
922 | } | |
923 | ||
ddc97297 PZ |
924 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
925 | { | |
926 | #ifdef CONFIG_SCHED_DEBUG | |
927 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
928 | ||
929 | if (d < 0) | |
930 | d = -d; | |
931 | ||
932 | if (d > 3*sysctl_sched_latency) | |
933 | schedstat_inc(cfs_rq, nr_spread_over); | |
934 | #endif | |
935 | } | |
936 | ||
aeb73b04 PZ |
937 | static void |
938 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
939 | { | |
1af5f730 | 940 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 941 | |
2cb8600e PZ |
942 | /* |
943 | * The 'current' period is already promised to the current tasks, | |
944 | * however the extra weight of the new task will slow them down a | |
945 | * little, place the new task so that it fits in the slot that | |
946 | * stays open at the end. | |
947 | */ | |
94dfb5e7 | 948 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 949 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 950 | |
a2e7a7eb | 951 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 952 | if (!initial) { |
a2e7a7eb | 953 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 954 | |
a2e7a7eb MG |
955 | /* |
956 | * Halve their sleep time's effect, to allow | |
957 | * for a gentler effect of sleepers: | |
958 | */ | |
959 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
960 | thresh >>= 1; | |
51e0304c | 961 | |
a2e7a7eb | 962 | vruntime -= thresh; |
aeb73b04 PZ |
963 | } |
964 | ||
b5d9d734 MG |
965 | /* ensure we never gain time by being placed backwards. */ |
966 | vruntime = max_vruntime(se->vruntime, vruntime); | |
967 | ||
67e9fb2a | 968 | se->vruntime = vruntime; |
aeb73b04 PZ |
969 | } |
970 | ||
bf0f6f24 | 971 | static void |
88ec22d3 | 972 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 973 | { |
88ec22d3 PZ |
974 | /* |
975 | * Update the normalized vruntime before updating min_vruntime | |
976 | * through callig update_curr(). | |
977 | */ | |
371fd7e7 | 978 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
979 | se->vruntime += cfs_rq->min_vruntime; |
980 | ||
bf0f6f24 | 981 | /* |
a2a2d680 | 982 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 983 | */ |
b7cc0896 | 984 | update_curr(cfs_rq); |
d6b55918 | 985 | update_cfs_load(cfs_rq, 0); |
a992241d | 986 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 987 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 988 | |
88ec22d3 | 989 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 990 | place_entity(cfs_rq, se, 0); |
2396af69 | 991 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 992 | } |
bf0f6f24 | 993 | |
d2417e5a | 994 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 995 | check_spread(cfs_rq, se); |
83b699ed SV |
996 | if (se != cfs_rq->curr) |
997 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 998 | se->on_rq = 1; |
3d4b47b4 PZ |
999 | |
1000 | if (cfs_rq->nr_running == 1) | |
1001 | list_add_leaf_cfs_rq(cfs_rq); | |
bf0f6f24 IM |
1002 | } |
1003 | ||
2c13c919 | 1004 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1005 | { |
2c13c919 RR |
1006 | for_each_sched_entity(se) { |
1007 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1008 | if (cfs_rq->last == se) | |
1009 | cfs_rq->last = NULL; | |
1010 | else | |
1011 | break; | |
1012 | } | |
1013 | } | |
2002c695 | 1014 | |
2c13c919 RR |
1015 | static void __clear_buddies_next(struct sched_entity *se) |
1016 | { | |
1017 | for_each_sched_entity(se) { | |
1018 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1019 | if (cfs_rq->next == se) | |
1020 | cfs_rq->next = NULL; | |
1021 | else | |
1022 | break; | |
1023 | } | |
2002c695 PZ |
1024 | } |
1025 | ||
ac53db59 RR |
1026 | static void __clear_buddies_skip(struct sched_entity *se) |
1027 | { | |
1028 | for_each_sched_entity(se) { | |
1029 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1030 | if (cfs_rq->skip == se) | |
1031 | cfs_rq->skip = NULL; | |
1032 | else | |
1033 | break; | |
1034 | } | |
1035 | } | |
1036 | ||
a571bbea PZ |
1037 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1038 | { | |
2c13c919 RR |
1039 | if (cfs_rq->last == se) |
1040 | __clear_buddies_last(se); | |
1041 | ||
1042 | if (cfs_rq->next == se) | |
1043 | __clear_buddies_next(se); | |
ac53db59 RR |
1044 | |
1045 | if (cfs_rq->skip == se) | |
1046 | __clear_buddies_skip(se); | |
a571bbea PZ |
1047 | } |
1048 | ||
bf0f6f24 | 1049 | static void |
371fd7e7 | 1050 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1051 | { |
a2a2d680 DA |
1052 | /* |
1053 | * Update run-time statistics of the 'current'. | |
1054 | */ | |
1055 | update_curr(cfs_rq); | |
1056 | ||
19b6a2e3 | 1057 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1058 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1059 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1060 | if (entity_is_task(se)) { |
1061 | struct task_struct *tsk = task_of(se); | |
1062 | ||
1063 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1064 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1065 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1066 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1067 | } |
db36cc7d | 1068 | #endif |
67e9fb2a PZ |
1069 | } |
1070 | ||
2002c695 | 1071 | clear_buddies(cfs_rq, se); |
4793241b | 1072 | |
83b699ed | 1073 | if (se != cfs_rq->curr) |
30cfdcfc | 1074 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1075 | se->on_rq = 0; |
d6b55918 | 1076 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1077 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1078 | |
1079 | /* | |
1080 | * Normalize the entity after updating the min_vruntime because the | |
1081 | * update can refer to the ->curr item and we need to reflect this | |
1082 | * movement in our normalized position. | |
1083 | */ | |
371fd7e7 | 1084 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1085 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 PZ |
1086 | |
1087 | update_min_vruntime(cfs_rq); | |
1088 | update_cfs_shares(cfs_rq); | |
bf0f6f24 IM |
1089 | } |
1090 | ||
1091 | /* | |
1092 | * Preempt the current task with a newly woken task if needed: | |
1093 | */ | |
7c92e54f | 1094 | static void |
2e09bf55 | 1095 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1096 | { |
11697830 PZ |
1097 | unsigned long ideal_runtime, delta_exec; |
1098 | ||
6d0f0ebd | 1099 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1100 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1101 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1102 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1103 | /* |
1104 | * The current task ran long enough, ensure it doesn't get | |
1105 | * re-elected due to buddy favours. | |
1106 | */ | |
1107 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1108 | return; |
1109 | } | |
1110 | ||
1111 | /* | |
1112 | * Ensure that a task that missed wakeup preemption by a | |
1113 | * narrow margin doesn't have to wait for a full slice. | |
1114 | * This also mitigates buddy induced latencies under load. | |
1115 | */ | |
f685ceac MG |
1116 | if (delta_exec < sysctl_sched_min_granularity) |
1117 | return; | |
1118 | ||
1119 | if (cfs_rq->nr_running > 1) { | |
ac53db59 | 1120 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac MG |
1121 | s64 delta = curr->vruntime - se->vruntime; |
1122 | ||
d7d82944 MG |
1123 | if (delta < 0) |
1124 | return; | |
1125 | ||
f685ceac MG |
1126 | if (delta > ideal_runtime) |
1127 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 1128 | } |
bf0f6f24 IM |
1129 | } |
1130 | ||
83b699ed | 1131 | static void |
8494f412 | 1132 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1133 | { |
83b699ed SV |
1134 | /* 'current' is not kept within the tree. */ |
1135 | if (se->on_rq) { | |
1136 | /* | |
1137 | * Any task has to be enqueued before it get to execute on | |
1138 | * a CPU. So account for the time it spent waiting on the | |
1139 | * runqueue. | |
1140 | */ | |
1141 | update_stats_wait_end(cfs_rq, se); | |
1142 | __dequeue_entity(cfs_rq, se); | |
1143 | } | |
1144 | ||
79303e9e | 1145 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1146 | cfs_rq->curr = se; |
eba1ed4b IM |
1147 | #ifdef CONFIG_SCHEDSTATS |
1148 | /* | |
1149 | * Track our maximum slice length, if the CPU's load is at | |
1150 | * least twice that of our own weight (i.e. dont track it | |
1151 | * when there are only lesser-weight tasks around): | |
1152 | */ | |
495eca49 | 1153 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1154 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1155 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1156 | } | |
1157 | #endif | |
4a55b450 | 1158 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1159 | } |
1160 | ||
3f3a4904 PZ |
1161 | static int |
1162 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1163 | ||
ac53db59 RR |
1164 | /* |
1165 | * Pick the next process, keeping these things in mind, in this order: | |
1166 | * 1) keep things fair between processes/task groups | |
1167 | * 2) pick the "next" process, since someone really wants that to run | |
1168 | * 3) pick the "last" process, for cache locality | |
1169 | * 4) do not run the "skip" process, if something else is available | |
1170 | */ | |
f4b6755f | 1171 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1172 | { |
ac53db59 | 1173 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1174 | struct sched_entity *left = se; |
f4b6755f | 1175 | |
ac53db59 RR |
1176 | /* |
1177 | * Avoid running the skip buddy, if running something else can | |
1178 | * be done without getting too unfair. | |
1179 | */ | |
1180 | if (cfs_rq->skip == se) { | |
1181 | struct sched_entity *second = __pick_next_entity(se); | |
1182 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1183 | se = second; | |
1184 | } | |
aa2ac252 | 1185 | |
f685ceac MG |
1186 | /* |
1187 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1188 | */ | |
1189 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1190 | se = cfs_rq->last; | |
1191 | ||
ac53db59 RR |
1192 | /* |
1193 | * Someone really wants this to run. If it's not unfair, run it. | |
1194 | */ | |
1195 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1196 | se = cfs_rq->next; | |
1197 | ||
f685ceac | 1198 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1199 | |
1200 | return se; | |
aa2ac252 PZ |
1201 | } |
1202 | ||
ab6cde26 | 1203 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1204 | { |
1205 | /* | |
1206 | * If still on the runqueue then deactivate_task() | |
1207 | * was not called and update_curr() has to be done: | |
1208 | */ | |
1209 | if (prev->on_rq) | |
b7cc0896 | 1210 | update_curr(cfs_rq); |
bf0f6f24 | 1211 | |
ddc97297 | 1212 | check_spread(cfs_rq, prev); |
30cfdcfc | 1213 | if (prev->on_rq) { |
5870db5b | 1214 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1215 | /* Put 'current' back into the tree. */ |
1216 | __enqueue_entity(cfs_rq, prev); | |
1217 | } | |
429d43bc | 1218 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1219 | } |
1220 | ||
8f4d37ec PZ |
1221 | static void |
1222 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1223 | { |
bf0f6f24 | 1224 | /* |
30cfdcfc | 1225 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1226 | */ |
30cfdcfc | 1227 | update_curr(cfs_rq); |
bf0f6f24 | 1228 | |
43365bd7 PT |
1229 | /* |
1230 | * Update share accounting for long-running entities. | |
1231 | */ | |
1232 | update_entity_shares_tick(cfs_rq); | |
1233 | ||
8f4d37ec PZ |
1234 | #ifdef CONFIG_SCHED_HRTICK |
1235 | /* | |
1236 | * queued ticks are scheduled to match the slice, so don't bother | |
1237 | * validating it and just reschedule. | |
1238 | */ | |
983ed7a6 HH |
1239 | if (queued) { |
1240 | resched_task(rq_of(cfs_rq)->curr); | |
1241 | return; | |
1242 | } | |
8f4d37ec PZ |
1243 | /* |
1244 | * don't let the period tick interfere with the hrtick preemption | |
1245 | */ | |
1246 | if (!sched_feat(DOUBLE_TICK) && | |
1247 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1248 | return; | |
1249 | #endif | |
1250 | ||
2c2efaed | 1251 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1252 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1253 | } |
1254 | ||
ab84d31e PT |
1255 | |
1256 | /************************************************** | |
1257 | * CFS bandwidth control machinery | |
1258 | */ | |
1259 | ||
1260 | #ifdef CONFIG_CFS_BANDWIDTH | |
1261 | /* | |
1262 | * default period for cfs group bandwidth. | |
1263 | * default: 0.1s, units: nanoseconds | |
1264 | */ | |
1265 | static inline u64 default_cfs_period(void) | |
1266 | { | |
1267 | return 100000000ULL; | |
1268 | } | |
ec12cb7f PT |
1269 | |
1270 | static inline u64 sched_cfs_bandwidth_slice(void) | |
1271 | { | |
1272 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
1273 | } | |
1274 | ||
1275 | static void assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
1276 | { | |
1277 | struct task_group *tg = cfs_rq->tg; | |
1278 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
1279 | u64 amount = 0, min_amount; | |
1280 | ||
1281 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
1282 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
1283 | ||
1284 | raw_spin_lock(&cfs_b->lock); | |
1285 | if (cfs_b->quota == RUNTIME_INF) | |
1286 | amount = min_amount; | |
58088ad0 PT |
1287 | else { |
1288 | /* ensure bandwidth timer remains active under consumption */ | |
1289 | if (!cfs_b->timer_active) | |
1290 | __start_cfs_bandwidth(cfs_b); | |
1291 | ||
1292 | if (cfs_b->runtime > 0) { | |
1293 | amount = min(cfs_b->runtime, min_amount); | |
1294 | cfs_b->runtime -= amount; | |
1295 | cfs_b->idle = 0; | |
1296 | } | |
ec12cb7f PT |
1297 | } |
1298 | raw_spin_unlock(&cfs_b->lock); | |
1299 | ||
1300 | cfs_rq->runtime_remaining += amount; | |
1301 | } | |
1302 | ||
1303 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1304 | unsigned long delta_exec) | |
1305 | { | |
1306 | if (!cfs_rq->runtime_enabled) | |
1307 | return; | |
1308 | ||
1309 | cfs_rq->runtime_remaining -= delta_exec; | |
1310 | if (cfs_rq->runtime_remaining > 0) | |
1311 | return; | |
1312 | ||
1313 | assign_cfs_rq_runtime(cfs_rq); | |
1314 | } | |
1315 | ||
1316 | static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1317 | unsigned long delta_exec) | |
1318 | { | |
1319 | if (!cfs_rq->runtime_enabled) | |
1320 | return; | |
1321 | ||
1322 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
1323 | } | |
1324 | ||
58088ad0 PT |
1325 | /* |
1326 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
1327 | * cfs_rqs as appropriate. If there has been no activity within the last | |
1328 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
1329 | * used to track this state. | |
1330 | */ | |
1331 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
1332 | { | |
1333 | int idle = 1; | |
1334 | ||
1335 | raw_spin_lock(&cfs_b->lock); | |
1336 | /* no need to continue the timer with no bandwidth constraint */ | |
1337 | if (cfs_b->quota == RUNTIME_INF) | |
1338 | goto out_unlock; | |
1339 | ||
1340 | idle = cfs_b->idle; | |
1341 | cfs_b->runtime = cfs_b->quota; | |
1342 | ||
1343 | /* mark as potentially idle for the upcoming period */ | |
1344 | cfs_b->idle = 1; | |
1345 | out_unlock: | |
1346 | if (idle) | |
1347 | cfs_b->timer_active = 0; | |
1348 | raw_spin_unlock(&cfs_b->lock); | |
1349 | ||
1350 | return idle; | |
1351 | } | |
ec12cb7f PT |
1352 | #else |
1353 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1354 | unsigned long delta_exec) {} | |
ab84d31e PT |
1355 | #endif |
1356 | ||
bf0f6f24 IM |
1357 | /************************************************** |
1358 | * CFS operations on tasks: | |
1359 | */ | |
1360 | ||
8f4d37ec PZ |
1361 | #ifdef CONFIG_SCHED_HRTICK |
1362 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1363 | { | |
8f4d37ec PZ |
1364 | struct sched_entity *se = &p->se; |
1365 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1366 | ||
1367 | WARN_ON(task_rq(p) != rq); | |
1368 | ||
1369 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1370 | u64 slice = sched_slice(cfs_rq, se); | |
1371 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1372 | s64 delta = slice - ran; | |
1373 | ||
1374 | if (delta < 0) { | |
1375 | if (rq->curr == p) | |
1376 | resched_task(p); | |
1377 | return; | |
1378 | } | |
1379 | ||
1380 | /* | |
1381 | * Don't schedule slices shorter than 10000ns, that just | |
1382 | * doesn't make sense. Rely on vruntime for fairness. | |
1383 | */ | |
31656519 | 1384 | if (rq->curr != p) |
157124c1 | 1385 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1386 | |
31656519 | 1387 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1388 | } |
1389 | } | |
a4c2f00f PZ |
1390 | |
1391 | /* | |
1392 | * called from enqueue/dequeue and updates the hrtick when the | |
1393 | * current task is from our class and nr_running is low enough | |
1394 | * to matter. | |
1395 | */ | |
1396 | static void hrtick_update(struct rq *rq) | |
1397 | { | |
1398 | struct task_struct *curr = rq->curr; | |
1399 | ||
1400 | if (curr->sched_class != &fair_sched_class) | |
1401 | return; | |
1402 | ||
1403 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1404 | hrtick_start_fair(rq, curr); | |
1405 | } | |
55e12e5e | 1406 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1407 | static inline void |
1408 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1409 | { | |
1410 | } | |
a4c2f00f PZ |
1411 | |
1412 | static inline void hrtick_update(struct rq *rq) | |
1413 | { | |
1414 | } | |
8f4d37ec PZ |
1415 | #endif |
1416 | ||
bf0f6f24 IM |
1417 | /* |
1418 | * The enqueue_task method is called before nr_running is | |
1419 | * increased. Here we update the fair scheduling stats and | |
1420 | * then put the task into the rbtree: | |
1421 | */ | |
ea87bb78 | 1422 | static void |
371fd7e7 | 1423 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1424 | { |
1425 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1426 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1427 | |
1428 | for_each_sched_entity(se) { | |
62fb1851 | 1429 | if (se->on_rq) |
bf0f6f24 IM |
1430 | break; |
1431 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 1432 | enqueue_entity(cfs_rq, se, flags); |
953bfcd1 | 1433 | cfs_rq->h_nr_running++; |
88ec22d3 | 1434 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 1435 | } |
8f4d37ec | 1436 | |
2069dd75 | 1437 | for_each_sched_entity(se) { |
0f317143 | 1438 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 1439 | cfs_rq->h_nr_running++; |
2069dd75 | 1440 | |
d6b55918 | 1441 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1442 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1443 | } |
1444 | ||
953bfcd1 | 1445 | inc_nr_running(rq); |
a4c2f00f | 1446 | hrtick_update(rq); |
bf0f6f24 IM |
1447 | } |
1448 | ||
2f36825b VP |
1449 | static void set_next_buddy(struct sched_entity *se); |
1450 | ||
bf0f6f24 IM |
1451 | /* |
1452 | * The dequeue_task method is called before nr_running is | |
1453 | * decreased. We remove the task from the rbtree and | |
1454 | * update the fair scheduling stats: | |
1455 | */ | |
371fd7e7 | 1456 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1457 | { |
1458 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1459 | struct sched_entity *se = &p->se; |
2f36825b | 1460 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
1461 | |
1462 | for_each_sched_entity(se) { | |
1463 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1464 | dequeue_entity(cfs_rq, se, flags); |
953bfcd1 | 1465 | cfs_rq->h_nr_running--; |
2069dd75 | 1466 | |
bf0f6f24 | 1467 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
1468 | if (cfs_rq->load.weight) { |
1469 | /* | |
1470 | * Bias pick_next to pick a task from this cfs_rq, as | |
1471 | * p is sleeping when it is within its sched_slice. | |
1472 | */ | |
1473 | if (task_sleep && parent_entity(se)) | |
1474 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
1475 | |
1476 | /* avoid re-evaluating load for this entity */ | |
1477 | se = parent_entity(se); | |
bf0f6f24 | 1478 | break; |
2f36825b | 1479 | } |
371fd7e7 | 1480 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1481 | } |
8f4d37ec | 1482 | |
2069dd75 | 1483 | for_each_sched_entity(se) { |
0f317143 | 1484 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 1485 | cfs_rq->h_nr_running--; |
2069dd75 | 1486 | |
d6b55918 | 1487 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1488 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1489 | } |
1490 | ||
953bfcd1 | 1491 | dec_nr_running(rq); |
a4c2f00f | 1492 | hrtick_update(rq); |
bf0f6f24 IM |
1493 | } |
1494 | ||
e7693a36 | 1495 | #ifdef CONFIG_SMP |
098fb9db | 1496 | |
74f8e4b2 | 1497 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
1498 | { |
1499 | struct sched_entity *se = &p->se; | |
1500 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
1501 | u64 min_vruntime; |
1502 | ||
1503 | #ifndef CONFIG_64BIT | |
1504 | u64 min_vruntime_copy; | |
88ec22d3 | 1505 | |
3fe1698b PZ |
1506 | do { |
1507 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
1508 | smp_rmb(); | |
1509 | min_vruntime = cfs_rq->min_vruntime; | |
1510 | } while (min_vruntime != min_vruntime_copy); | |
1511 | #else | |
1512 | min_vruntime = cfs_rq->min_vruntime; | |
1513 | #endif | |
88ec22d3 | 1514 | |
3fe1698b | 1515 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
1516 | } |
1517 | ||
bb3469ac | 1518 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1519 | /* |
1520 | * effective_load() calculates the load change as seen from the root_task_group | |
1521 | * | |
1522 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1523 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1524 | * can calculate the shift in shares. | |
f5bfb7d9 | 1525 | */ |
2069dd75 | 1526 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 1527 | { |
4be9daaa | 1528 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1529 | |
1530 | if (!tg->parent) | |
1531 | return wl; | |
1532 | ||
4be9daaa | 1533 | for_each_sched_entity(se) { |
977dda7c | 1534 | long lw, w; |
4be9daaa | 1535 | |
977dda7c PT |
1536 | tg = se->my_q->tg; |
1537 | w = se->my_q->load.weight; | |
bb3469ac | 1538 | |
977dda7c PT |
1539 | /* use this cpu's instantaneous contribution */ |
1540 | lw = atomic_read(&tg->load_weight); | |
1541 | lw -= se->my_q->load_contribution; | |
1542 | lw += w + wg; | |
4be9daaa | 1543 | |
977dda7c | 1544 | wl += w; |
940959e9 | 1545 | |
977dda7c PT |
1546 | if (lw > 0 && wl < lw) |
1547 | wl = (wl * tg->shares) / lw; | |
1548 | else | |
1549 | wl = tg->shares; | |
940959e9 | 1550 | |
977dda7c PT |
1551 | /* zero point is MIN_SHARES */ |
1552 | if (wl < MIN_SHARES) | |
1553 | wl = MIN_SHARES; | |
1554 | wl -= se->load.weight; | |
4be9daaa | 1555 | wg = 0; |
4be9daaa | 1556 | } |
bb3469ac | 1557 | |
4be9daaa | 1558 | return wl; |
bb3469ac | 1559 | } |
4be9daaa | 1560 | |
bb3469ac | 1561 | #else |
4be9daaa | 1562 | |
83378269 PZ |
1563 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1564 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1565 | { |
83378269 | 1566 | return wl; |
bb3469ac | 1567 | } |
4be9daaa | 1568 | |
bb3469ac PZ |
1569 | #endif |
1570 | ||
c88d5910 | 1571 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1572 | { |
e37b6a7b | 1573 | s64 this_load, load; |
c88d5910 | 1574 | int idx, this_cpu, prev_cpu; |
098fb9db | 1575 | unsigned long tl_per_task; |
c88d5910 | 1576 | struct task_group *tg; |
83378269 | 1577 | unsigned long weight; |
b3137bc8 | 1578 | int balanced; |
098fb9db | 1579 | |
c88d5910 PZ |
1580 | idx = sd->wake_idx; |
1581 | this_cpu = smp_processor_id(); | |
1582 | prev_cpu = task_cpu(p); | |
1583 | load = source_load(prev_cpu, idx); | |
1584 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1585 | |
b3137bc8 MG |
1586 | /* |
1587 | * If sync wakeup then subtract the (maximum possible) | |
1588 | * effect of the currently running task from the load | |
1589 | * of the current CPU: | |
1590 | */ | |
83378269 PZ |
1591 | if (sync) { |
1592 | tg = task_group(current); | |
1593 | weight = current->se.load.weight; | |
1594 | ||
c88d5910 | 1595 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
1596 | load += effective_load(tg, prev_cpu, 0, -weight); |
1597 | } | |
b3137bc8 | 1598 | |
83378269 PZ |
1599 | tg = task_group(p); |
1600 | weight = p->se.load.weight; | |
b3137bc8 | 1601 | |
71a29aa7 PZ |
1602 | /* |
1603 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
1604 | * due to the sync cause above having dropped this_load to 0, we'll |
1605 | * always have an imbalance, but there's really nothing you can do | |
1606 | * about that, so that's good too. | |
71a29aa7 PZ |
1607 | * |
1608 | * Otherwise check if either cpus are near enough in load to allow this | |
1609 | * task to be woken on this_cpu. | |
1610 | */ | |
e37b6a7b PT |
1611 | if (this_load > 0) { |
1612 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
1613 | |
1614 | this_eff_load = 100; | |
1615 | this_eff_load *= power_of(prev_cpu); | |
1616 | this_eff_load *= this_load + | |
1617 | effective_load(tg, this_cpu, weight, weight); | |
1618 | ||
1619 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
1620 | prev_eff_load *= power_of(this_cpu); | |
1621 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
1622 | ||
1623 | balanced = this_eff_load <= prev_eff_load; | |
1624 | } else | |
1625 | balanced = true; | |
b3137bc8 | 1626 | |
098fb9db | 1627 | /* |
4ae7d5ce IM |
1628 | * If the currently running task will sleep within |
1629 | * a reasonable amount of time then attract this newly | |
1630 | * woken task: | |
098fb9db | 1631 | */ |
2fb7635c PZ |
1632 | if (sync && balanced) |
1633 | return 1; | |
098fb9db | 1634 | |
41acab88 | 1635 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
1636 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
1637 | ||
c88d5910 PZ |
1638 | if (balanced || |
1639 | (this_load <= load && | |
1640 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
1641 | /* |
1642 | * This domain has SD_WAKE_AFFINE and | |
1643 | * p is cache cold in this domain, and | |
1644 | * there is no bad imbalance. | |
1645 | */ | |
c88d5910 | 1646 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 1647 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
1648 | |
1649 | return 1; | |
1650 | } | |
1651 | return 0; | |
1652 | } | |
1653 | ||
aaee1203 PZ |
1654 | /* |
1655 | * find_idlest_group finds and returns the least busy CPU group within the | |
1656 | * domain. | |
1657 | */ | |
1658 | static struct sched_group * | |
78e7ed53 | 1659 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 1660 | int this_cpu, int load_idx) |
e7693a36 | 1661 | { |
b3bd3de6 | 1662 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 1663 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 1664 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 1665 | |
aaee1203 PZ |
1666 | do { |
1667 | unsigned long load, avg_load; | |
1668 | int local_group; | |
1669 | int i; | |
e7693a36 | 1670 | |
aaee1203 PZ |
1671 | /* Skip over this group if it has no CPUs allowed */ |
1672 | if (!cpumask_intersects(sched_group_cpus(group), | |
1673 | &p->cpus_allowed)) | |
1674 | continue; | |
1675 | ||
1676 | local_group = cpumask_test_cpu(this_cpu, | |
1677 | sched_group_cpus(group)); | |
1678 | ||
1679 | /* Tally up the load of all CPUs in the group */ | |
1680 | avg_load = 0; | |
1681 | ||
1682 | for_each_cpu(i, sched_group_cpus(group)) { | |
1683 | /* Bias balancing toward cpus of our domain */ | |
1684 | if (local_group) | |
1685 | load = source_load(i, load_idx); | |
1686 | else | |
1687 | load = target_load(i, load_idx); | |
1688 | ||
1689 | avg_load += load; | |
1690 | } | |
1691 | ||
1692 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 1693 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
1694 | |
1695 | if (local_group) { | |
1696 | this_load = avg_load; | |
aaee1203 PZ |
1697 | } else if (avg_load < min_load) { |
1698 | min_load = avg_load; | |
1699 | idlest = group; | |
1700 | } | |
1701 | } while (group = group->next, group != sd->groups); | |
1702 | ||
1703 | if (!idlest || 100*this_load < imbalance*min_load) | |
1704 | return NULL; | |
1705 | return idlest; | |
1706 | } | |
1707 | ||
1708 | /* | |
1709 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
1710 | */ | |
1711 | static int | |
1712 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
1713 | { | |
1714 | unsigned long load, min_load = ULONG_MAX; | |
1715 | int idlest = -1; | |
1716 | int i; | |
1717 | ||
1718 | /* Traverse only the allowed CPUs */ | |
1719 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
1720 | load = weighted_cpuload(i); | |
1721 | ||
1722 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
1723 | min_load = load; | |
1724 | idlest = i; | |
e7693a36 GH |
1725 | } |
1726 | } | |
1727 | ||
aaee1203 PZ |
1728 | return idlest; |
1729 | } | |
e7693a36 | 1730 | |
a50bde51 PZ |
1731 | /* |
1732 | * Try and locate an idle CPU in the sched_domain. | |
1733 | */ | |
99bd5e2f | 1734 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
1735 | { |
1736 | int cpu = smp_processor_id(); | |
1737 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 1738 | struct sched_domain *sd; |
a50bde51 PZ |
1739 | int i; |
1740 | ||
1741 | /* | |
99bd5e2f SS |
1742 | * If the task is going to be woken-up on this cpu and if it is |
1743 | * already idle, then it is the right target. | |
a50bde51 | 1744 | */ |
99bd5e2f SS |
1745 | if (target == cpu && idle_cpu(cpu)) |
1746 | return cpu; | |
1747 | ||
1748 | /* | |
1749 | * If the task is going to be woken-up on the cpu where it previously | |
1750 | * ran and if it is currently idle, then it the right target. | |
1751 | */ | |
1752 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 1753 | return prev_cpu; |
a50bde51 PZ |
1754 | |
1755 | /* | |
99bd5e2f | 1756 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 1757 | */ |
dce840a0 | 1758 | rcu_read_lock(); |
99bd5e2f SS |
1759 | for_each_domain(target, sd) { |
1760 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 1761 | break; |
99bd5e2f SS |
1762 | |
1763 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
1764 | if (idle_cpu(i)) { | |
1765 | target = i; | |
1766 | break; | |
1767 | } | |
a50bde51 | 1768 | } |
99bd5e2f SS |
1769 | |
1770 | /* | |
1771 | * Lets stop looking for an idle sibling when we reached | |
1772 | * the domain that spans the current cpu and prev_cpu. | |
1773 | */ | |
1774 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
1775 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
1776 | break; | |
a50bde51 | 1777 | } |
dce840a0 | 1778 | rcu_read_unlock(); |
a50bde51 PZ |
1779 | |
1780 | return target; | |
1781 | } | |
1782 | ||
aaee1203 PZ |
1783 | /* |
1784 | * sched_balance_self: balance the current task (running on cpu) in domains | |
1785 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
1786 | * SD_BALANCE_EXEC. | |
1787 | * | |
1788 | * Balance, ie. select the least loaded group. | |
1789 | * | |
1790 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
1791 | * | |
1792 | * preempt must be disabled. | |
1793 | */ | |
0017d735 | 1794 | static int |
7608dec2 | 1795 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 1796 | { |
29cd8bae | 1797 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
1798 | int cpu = smp_processor_id(); |
1799 | int prev_cpu = task_cpu(p); | |
1800 | int new_cpu = cpu; | |
99bd5e2f | 1801 | int want_affine = 0; |
29cd8bae | 1802 | int want_sd = 1; |
5158f4e4 | 1803 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 1804 | |
0763a660 | 1805 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 1806 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
1807 | want_affine = 1; |
1808 | new_cpu = prev_cpu; | |
1809 | } | |
aaee1203 | 1810 | |
dce840a0 | 1811 | rcu_read_lock(); |
aaee1203 | 1812 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
1813 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
1814 | continue; | |
1815 | ||
aaee1203 | 1816 | /* |
ae154be1 PZ |
1817 | * If power savings logic is enabled for a domain, see if we |
1818 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 1819 | */ |
59abf026 | 1820 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
1821 | unsigned long power = 0; |
1822 | unsigned long nr_running = 0; | |
1823 | unsigned long capacity; | |
1824 | int i; | |
1825 | ||
1826 | for_each_cpu(i, sched_domain_span(tmp)) { | |
1827 | power += power_of(i); | |
1828 | nr_running += cpu_rq(i)->cfs.nr_running; | |
1829 | } | |
1830 | ||
1399fa78 | 1831 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); |
ae154be1 | 1832 | |
59abf026 PZ |
1833 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
1834 | nr_running /= 2; | |
1835 | ||
1836 | if (nr_running < capacity) | |
29cd8bae | 1837 | want_sd = 0; |
ae154be1 | 1838 | } |
aaee1203 | 1839 | |
fe3bcfe1 | 1840 | /* |
99bd5e2f SS |
1841 | * If both cpu and prev_cpu are part of this domain, |
1842 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 1843 | */ |
99bd5e2f SS |
1844 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
1845 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
1846 | affine_sd = tmp; | |
1847 | want_affine = 0; | |
c88d5910 PZ |
1848 | } |
1849 | ||
29cd8bae PZ |
1850 | if (!want_sd && !want_affine) |
1851 | break; | |
1852 | ||
0763a660 | 1853 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
1854 | continue; |
1855 | ||
29cd8bae PZ |
1856 | if (want_sd) |
1857 | sd = tmp; | |
1858 | } | |
1859 | ||
8b911acd | 1860 | if (affine_sd) { |
99bd5e2f | 1861 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
1862 | prev_cpu = cpu; |
1863 | ||
1864 | new_cpu = select_idle_sibling(p, prev_cpu); | |
1865 | goto unlock; | |
8b911acd | 1866 | } |
e7693a36 | 1867 | |
aaee1203 | 1868 | while (sd) { |
5158f4e4 | 1869 | int load_idx = sd->forkexec_idx; |
aaee1203 | 1870 | struct sched_group *group; |
c88d5910 | 1871 | int weight; |
098fb9db | 1872 | |
0763a660 | 1873 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
1874 | sd = sd->child; |
1875 | continue; | |
1876 | } | |
098fb9db | 1877 | |
5158f4e4 PZ |
1878 | if (sd_flag & SD_BALANCE_WAKE) |
1879 | load_idx = sd->wake_idx; | |
098fb9db | 1880 | |
5158f4e4 | 1881 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
1882 | if (!group) { |
1883 | sd = sd->child; | |
1884 | continue; | |
1885 | } | |
4ae7d5ce | 1886 | |
d7c33c49 | 1887 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
1888 | if (new_cpu == -1 || new_cpu == cpu) { |
1889 | /* Now try balancing at a lower domain level of cpu */ | |
1890 | sd = sd->child; | |
1891 | continue; | |
e7693a36 | 1892 | } |
aaee1203 PZ |
1893 | |
1894 | /* Now try balancing at a lower domain level of new_cpu */ | |
1895 | cpu = new_cpu; | |
669c55e9 | 1896 | weight = sd->span_weight; |
aaee1203 PZ |
1897 | sd = NULL; |
1898 | for_each_domain(cpu, tmp) { | |
669c55e9 | 1899 | if (weight <= tmp->span_weight) |
aaee1203 | 1900 | break; |
0763a660 | 1901 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
1902 | sd = tmp; |
1903 | } | |
1904 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 1905 | } |
dce840a0 PZ |
1906 | unlock: |
1907 | rcu_read_unlock(); | |
e7693a36 | 1908 | |
c88d5910 | 1909 | return new_cpu; |
e7693a36 GH |
1910 | } |
1911 | #endif /* CONFIG_SMP */ | |
1912 | ||
e52fb7c0 PZ |
1913 | static unsigned long |
1914 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
1915 | { |
1916 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
1917 | ||
1918 | /* | |
e52fb7c0 PZ |
1919 | * Since its curr running now, convert the gran from real-time |
1920 | * to virtual-time in his units. | |
13814d42 MG |
1921 | * |
1922 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
1923 | * they get preempted easier. That is, if 'se' < 'curr' then | |
1924 | * the resulting gran will be larger, therefore penalizing the | |
1925 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
1926 | * be smaller, again penalizing the lighter task. | |
1927 | * | |
1928 | * This is especially important for buddies when the leftmost | |
1929 | * task is higher priority than the buddy. | |
0bbd3336 | 1930 | */ |
f4ad9bd2 | 1931 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
1932 | } |
1933 | ||
464b7527 PZ |
1934 | /* |
1935 | * Should 'se' preempt 'curr'. | |
1936 | * | |
1937 | * |s1 | |
1938 | * |s2 | |
1939 | * |s3 | |
1940 | * g | |
1941 | * |<--->|c | |
1942 | * | |
1943 | * w(c, s1) = -1 | |
1944 | * w(c, s2) = 0 | |
1945 | * w(c, s3) = 1 | |
1946 | * | |
1947 | */ | |
1948 | static int | |
1949 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
1950 | { | |
1951 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
1952 | ||
1953 | if (vdiff <= 0) | |
1954 | return -1; | |
1955 | ||
e52fb7c0 | 1956 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
1957 | if (vdiff > gran) |
1958 | return 1; | |
1959 | ||
1960 | return 0; | |
1961 | } | |
1962 | ||
02479099 PZ |
1963 | static void set_last_buddy(struct sched_entity *se) |
1964 | { | |
69c80f3e VP |
1965 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
1966 | return; | |
1967 | ||
1968 | for_each_sched_entity(se) | |
1969 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
1970 | } |
1971 | ||
1972 | static void set_next_buddy(struct sched_entity *se) | |
1973 | { | |
69c80f3e VP |
1974 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
1975 | return; | |
1976 | ||
1977 | for_each_sched_entity(se) | |
1978 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
1979 | } |
1980 | ||
ac53db59 RR |
1981 | static void set_skip_buddy(struct sched_entity *se) |
1982 | { | |
69c80f3e VP |
1983 | for_each_sched_entity(se) |
1984 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
1985 | } |
1986 | ||
bf0f6f24 IM |
1987 | /* |
1988 | * Preempt the current task with a newly woken task if needed: | |
1989 | */ | |
5a9b86f6 | 1990 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
1991 | { |
1992 | struct task_struct *curr = rq->curr; | |
8651a86c | 1993 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 1994 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 1995 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 1996 | int next_buddy_marked = 0; |
bf0f6f24 | 1997 | |
4ae7d5ce IM |
1998 | if (unlikely(se == pse)) |
1999 | return; | |
2000 | ||
2f36825b | 2001 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 2002 | set_next_buddy(pse); |
2f36825b VP |
2003 | next_buddy_marked = 1; |
2004 | } | |
57fdc26d | 2005 | |
aec0a514 BR |
2006 | /* |
2007 | * We can come here with TIF_NEED_RESCHED already set from new task | |
2008 | * wake up path. | |
2009 | */ | |
2010 | if (test_tsk_need_resched(curr)) | |
2011 | return; | |
2012 | ||
a2f5c9ab DH |
2013 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
2014 | if (unlikely(curr->policy == SCHED_IDLE) && | |
2015 | likely(p->policy != SCHED_IDLE)) | |
2016 | goto preempt; | |
2017 | ||
91c234b4 | 2018 | /* |
a2f5c9ab DH |
2019 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
2020 | * is driven by the tick): | |
91c234b4 | 2021 | */ |
6bc912b7 | 2022 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 2023 | return; |
bf0f6f24 | 2024 | |
464b7527 | 2025 | find_matching_se(&se, &pse); |
9bbd7374 | 2026 | update_curr(cfs_rq_of(se)); |
002f128b | 2027 | BUG_ON(!pse); |
2f36825b VP |
2028 | if (wakeup_preempt_entity(se, pse) == 1) { |
2029 | /* | |
2030 | * Bias pick_next to pick the sched entity that is | |
2031 | * triggering this preemption. | |
2032 | */ | |
2033 | if (!next_buddy_marked) | |
2034 | set_next_buddy(pse); | |
3a7e73a2 | 2035 | goto preempt; |
2f36825b | 2036 | } |
464b7527 | 2037 | |
3a7e73a2 | 2038 | return; |
a65ac745 | 2039 | |
3a7e73a2 PZ |
2040 | preempt: |
2041 | resched_task(curr); | |
2042 | /* | |
2043 | * Only set the backward buddy when the current task is still | |
2044 | * on the rq. This can happen when a wakeup gets interleaved | |
2045 | * with schedule on the ->pre_schedule() or idle_balance() | |
2046 | * point, either of which can * drop the rq lock. | |
2047 | * | |
2048 | * Also, during early boot the idle thread is in the fair class, | |
2049 | * for obvious reasons its a bad idea to schedule back to it. | |
2050 | */ | |
2051 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
2052 | return; | |
2053 | ||
2054 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
2055 | set_last_buddy(se); | |
bf0f6f24 IM |
2056 | } |
2057 | ||
fb8d4724 | 2058 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 2059 | { |
8f4d37ec | 2060 | struct task_struct *p; |
bf0f6f24 IM |
2061 | struct cfs_rq *cfs_rq = &rq->cfs; |
2062 | struct sched_entity *se; | |
2063 | ||
36ace27e | 2064 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
2065 | return NULL; |
2066 | ||
2067 | do { | |
9948f4b2 | 2068 | se = pick_next_entity(cfs_rq); |
f4b6755f | 2069 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
2070 | cfs_rq = group_cfs_rq(se); |
2071 | } while (cfs_rq); | |
2072 | ||
8f4d37ec PZ |
2073 | p = task_of(se); |
2074 | hrtick_start_fair(rq, p); | |
2075 | ||
2076 | return p; | |
bf0f6f24 IM |
2077 | } |
2078 | ||
2079 | /* | |
2080 | * Account for a descheduled task: | |
2081 | */ | |
31ee529c | 2082 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
2083 | { |
2084 | struct sched_entity *se = &prev->se; | |
2085 | struct cfs_rq *cfs_rq; | |
2086 | ||
2087 | for_each_sched_entity(se) { | |
2088 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 2089 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
2090 | } |
2091 | } | |
2092 | ||
ac53db59 RR |
2093 | /* |
2094 | * sched_yield() is very simple | |
2095 | * | |
2096 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
2097 | */ | |
2098 | static void yield_task_fair(struct rq *rq) | |
2099 | { | |
2100 | struct task_struct *curr = rq->curr; | |
2101 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
2102 | struct sched_entity *se = &curr->se; | |
2103 | ||
2104 | /* | |
2105 | * Are we the only task in the tree? | |
2106 | */ | |
2107 | if (unlikely(rq->nr_running == 1)) | |
2108 | return; | |
2109 | ||
2110 | clear_buddies(cfs_rq, se); | |
2111 | ||
2112 | if (curr->policy != SCHED_BATCH) { | |
2113 | update_rq_clock(rq); | |
2114 | /* | |
2115 | * Update run-time statistics of the 'current'. | |
2116 | */ | |
2117 | update_curr(cfs_rq); | |
2118 | } | |
2119 | ||
2120 | set_skip_buddy(se); | |
2121 | } | |
2122 | ||
d95f4122 MG |
2123 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
2124 | { | |
2125 | struct sched_entity *se = &p->se; | |
2126 | ||
2127 | if (!se->on_rq) | |
2128 | return false; | |
2129 | ||
2130 | /* Tell the scheduler that we'd really like pse to run next. */ | |
2131 | set_next_buddy(se); | |
2132 | ||
d95f4122 MG |
2133 | yield_task_fair(rq); |
2134 | ||
2135 | return true; | |
2136 | } | |
2137 | ||
681f3e68 | 2138 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
2139 | /************************************************** |
2140 | * Fair scheduling class load-balancing methods: | |
2141 | */ | |
2142 | ||
1e3c88bd PZ |
2143 | /* |
2144 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
2145 | * Both runqueues must be locked. | |
2146 | */ | |
2147 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
2148 | struct rq *this_rq, int this_cpu) | |
2149 | { | |
2150 | deactivate_task(src_rq, p, 0); | |
2151 | set_task_cpu(p, this_cpu); | |
2152 | activate_task(this_rq, p, 0); | |
2153 | check_preempt_curr(this_rq, p, 0); | |
2154 | } | |
2155 | ||
2156 | /* | |
2157 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2158 | */ | |
2159 | static | |
2160 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
2161 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2162 | int *all_pinned) | |
2163 | { | |
2164 | int tsk_cache_hot = 0; | |
2165 | /* | |
2166 | * We do not migrate tasks that are: | |
2167 | * 1) running (obviously), or | |
2168 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2169 | * 3) are cache-hot on their current CPU. | |
2170 | */ | |
2171 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 2172 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
2173 | return 0; |
2174 | } | |
2175 | *all_pinned = 0; | |
2176 | ||
2177 | if (task_running(rq, p)) { | |
41acab88 | 2178 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
2179 | return 0; |
2180 | } | |
2181 | ||
2182 | /* | |
2183 | * Aggressive migration if: | |
2184 | * 1) task is cache cold, or | |
2185 | * 2) too many balance attempts have failed. | |
2186 | */ | |
2187 | ||
305e6835 | 2188 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
2189 | if (!tsk_cache_hot || |
2190 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
2191 | #ifdef CONFIG_SCHEDSTATS | |
2192 | if (tsk_cache_hot) { | |
2193 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 2194 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
2195 | } |
2196 | #endif | |
2197 | return 1; | |
2198 | } | |
2199 | ||
2200 | if (tsk_cache_hot) { | |
41acab88 | 2201 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
2202 | return 0; |
2203 | } | |
2204 | return 1; | |
2205 | } | |
2206 | ||
897c395f PZ |
2207 | /* |
2208 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
2209 | * part of active balancing operations within "domain". | |
2210 | * Returns 1 if successful and 0 otherwise. | |
2211 | * | |
2212 | * Called with both runqueues locked. | |
2213 | */ | |
2214 | static int | |
2215 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2216 | struct sched_domain *sd, enum cpu_idle_type idle) | |
2217 | { | |
2218 | struct task_struct *p, *n; | |
2219 | struct cfs_rq *cfs_rq; | |
2220 | int pinned = 0; | |
2221 | ||
2222 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
2223 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
2224 | ||
2225 | if (!can_migrate_task(p, busiest, this_cpu, | |
2226 | sd, idle, &pinned)) | |
2227 | continue; | |
2228 | ||
2229 | pull_task(busiest, p, this_rq, this_cpu); | |
2230 | /* | |
2231 | * Right now, this is only the second place pull_task() | |
2232 | * is called, so we can safely collect pull_task() | |
2233 | * stats here rather than inside pull_task(). | |
2234 | */ | |
2235 | schedstat_inc(sd, lb_gained[idle]); | |
2236 | return 1; | |
2237 | } | |
2238 | } | |
2239 | ||
2240 | return 0; | |
2241 | } | |
2242 | ||
1e3c88bd PZ |
2243 | static unsigned long |
2244 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2245 | unsigned long max_load_move, struct sched_domain *sd, | |
2246 | enum cpu_idle_type idle, int *all_pinned, | |
931aeeda | 2247 | struct cfs_rq *busiest_cfs_rq) |
1e3c88bd | 2248 | { |
b30aef17 | 2249 | int loops = 0, pulled = 0; |
1e3c88bd | 2250 | long rem_load_move = max_load_move; |
ee00e66f | 2251 | struct task_struct *p, *n; |
1e3c88bd PZ |
2252 | |
2253 | if (max_load_move == 0) | |
2254 | goto out; | |
2255 | ||
ee00e66f PZ |
2256 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
2257 | if (loops++ > sysctl_sched_nr_migrate) | |
2258 | break; | |
1e3c88bd | 2259 | |
ee00e66f | 2260 | if ((p->se.load.weight >> 1) > rem_load_move || |
b30aef17 KC |
2261 | !can_migrate_task(p, busiest, this_cpu, sd, idle, |
2262 | all_pinned)) | |
ee00e66f | 2263 | continue; |
1e3c88bd | 2264 | |
ee00e66f PZ |
2265 | pull_task(busiest, p, this_rq, this_cpu); |
2266 | pulled++; | |
2267 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
2268 | |
2269 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
2270 | /* |
2271 | * NEWIDLE balancing is a source of latency, so preemptible | |
2272 | * kernels will stop after the first task is pulled to minimize | |
2273 | * the critical section. | |
2274 | */ | |
2275 | if (idle == CPU_NEWLY_IDLE) | |
2276 | break; | |
1e3c88bd PZ |
2277 | #endif |
2278 | ||
ee00e66f PZ |
2279 | /* |
2280 | * We only want to steal up to the prescribed amount of | |
2281 | * weighted load. | |
2282 | */ | |
2283 | if (rem_load_move <= 0) | |
2284 | break; | |
1e3c88bd PZ |
2285 | } |
2286 | out: | |
2287 | /* | |
2288 | * Right now, this is one of only two places pull_task() is called, | |
2289 | * so we can safely collect pull_task() stats here rather than | |
2290 | * inside pull_task(). | |
2291 | */ | |
2292 | schedstat_add(sd, lb_gained[idle], pulled); | |
2293 | ||
1e3c88bd PZ |
2294 | return max_load_move - rem_load_move; |
2295 | } | |
2296 | ||
230059de | 2297 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
2298 | /* |
2299 | * update tg->load_weight by folding this cpu's load_avg | |
2300 | */ | |
67e86250 | 2301 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
2302 | { |
2303 | struct cfs_rq *cfs_rq; | |
2304 | unsigned long flags; | |
2305 | struct rq *rq; | |
9e3081ca PZ |
2306 | |
2307 | if (!tg->se[cpu]) | |
2308 | return 0; | |
2309 | ||
2310 | rq = cpu_rq(cpu); | |
2311 | cfs_rq = tg->cfs_rq[cpu]; | |
2312 | ||
2313 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2314 | ||
2315 | update_rq_clock(rq); | |
d6b55918 | 2316 | update_cfs_load(cfs_rq, 1); |
9e3081ca PZ |
2317 | |
2318 | /* | |
2319 | * We need to update shares after updating tg->load_weight in | |
2320 | * order to adjust the weight of groups with long running tasks. | |
2321 | */ | |
6d5ab293 | 2322 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
2323 | |
2324 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2325 | ||
2326 | return 0; | |
2327 | } | |
2328 | ||
2329 | static void update_shares(int cpu) | |
2330 | { | |
2331 | struct cfs_rq *cfs_rq; | |
2332 | struct rq *rq = cpu_rq(cpu); | |
2333 | ||
2334 | rcu_read_lock(); | |
9763b67f PZ |
2335 | /* |
2336 | * Iterates the task_group tree in a bottom up fashion, see | |
2337 | * list_add_leaf_cfs_rq() for details. | |
2338 | */ | |
67e86250 PT |
2339 | for_each_leaf_cfs_rq(rq, cfs_rq) |
2340 | update_shares_cpu(cfs_rq->tg, cpu); | |
9e3081ca PZ |
2341 | rcu_read_unlock(); |
2342 | } | |
2343 | ||
9763b67f PZ |
2344 | /* |
2345 | * Compute the cpu's hierarchical load factor for each task group. | |
2346 | * This needs to be done in a top-down fashion because the load of a child | |
2347 | * group is a fraction of its parents load. | |
2348 | */ | |
2349 | static int tg_load_down(struct task_group *tg, void *data) | |
2350 | { | |
2351 | unsigned long load; | |
2352 | long cpu = (long)data; | |
2353 | ||
2354 | if (!tg->parent) { | |
2355 | load = cpu_rq(cpu)->load.weight; | |
2356 | } else { | |
2357 | load = tg->parent->cfs_rq[cpu]->h_load; | |
2358 | load *= tg->se[cpu]->load.weight; | |
2359 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
2360 | } | |
2361 | ||
2362 | tg->cfs_rq[cpu]->h_load = load; | |
2363 | ||
2364 | return 0; | |
2365 | } | |
2366 | ||
2367 | static void update_h_load(long cpu) | |
2368 | { | |
2369 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); | |
2370 | } | |
2371 | ||
230059de PZ |
2372 | static unsigned long |
2373 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2374 | unsigned long max_load_move, | |
2375 | struct sched_domain *sd, enum cpu_idle_type idle, | |
931aeeda | 2376 | int *all_pinned) |
230059de PZ |
2377 | { |
2378 | long rem_load_move = max_load_move; | |
9763b67f | 2379 | struct cfs_rq *busiest_cfs_rq; |
230059de PZ |
2380 | |
2381 | rcu_read_lock(); | |
9763b67f | 2382 | update_h_load(cpu_of(busiest)); |
230059de | 2383 | |
9763b67f | 2384 | for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) { |
230059de PZ |
2385 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
2386 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
2387 | u64 rem_load, moved_load; | |
2388 | ||
2389 | /* | |
2390 | * empty group | |
2391 | */ | |
2392 | if (!busiest_cfs_rq->task_weight) | |
2393 | continue; | |
2394 | ||
2395 | rem_load = (u64)rem_load_move * busiest_weight; | |
2396 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
2397 | ||
2398 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
931aeeda | 2399 | rem_load, sd, idle, all_pinned, |
230059de PZ |
2400 | busiest_cfs_rq); |
2401 | ||
2402 | if (!moved_load) | |
2403 | continue; | |
2404 | ||
2405 | moved_load *= busiest_h_load; | |
2406 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
2407 | ||
2408 | rem_load_move -= moved_load; | |
2409 | if (rem_load_move < 0) | |
2410 | break; | |
2411 | } | |
2412 | rcu_read_unlock(); | |
2413 | ||
2414 | return max_load_move - rem_load_move; | |
2415 | } | |
2416 | #else | |
9e3081ca PZ |
2417 | static inline void update_shares(int cpu) |
2418 | { | |
2419 | } | |
2420 | ||
230059de PZ |
2421 | static unsigned long |
2422 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2423 | unsigned long max_load_move, | |
2424 | struct sched_domain *sd, enum cpu_idle_type idle, | |
931aeeda | 2425 | int *all_pinned) |
230059de PZ |
2426 | { |
2427 | return balance_tasks(this_rq, this_cpu, busiest, | |
2428 | max_load_move, sd, idle, all_pinned, | |
931aeeda | 2429 | &busiest->cfs); |
230059de PZ |
2430 | } |
2431 | #endif | |
2432 | ||
1e3c88bd PZ |
2433 | /* |
2434 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2435 | * this_rq, as part of a balancing operation within domain "sd". | |
2436 | * Returns 1 if successful and 0 otherwise. | |
2437 | * | |
2438 | * Called with both runqueues locked. | |
2439 | */ | |
2440 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2441 | unsigned long max_load_move, | |
2442 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2443 | int *all_pinned) | |
2444 | { | |
3d45fd80 | 2445 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2446 | |
2447 | do { | |
3d45fd80 | 2448 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd | 2449 | max_load_move - total_load_moved, |
931aeeda | 2450 | sd, idle, all_pinned); |
3d45fd80 PZ |
2451 | |
2452 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2453 | |
2454 | #ifdef CONFIG_PREEMPT | |
2455 | /* | |
2456 | * NEWIDLE balancing is a source of latency, so preemptible | |
2457 | * kernels will stop after the first task is pulled to minimize | |
2458 | * the critical section. | |
2459 | */ | |
2460 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2461 | break; | |
baa8c110 PZ |
2462 | |
2463 | if (raw_spin_is_contended(&this_rq->lock) || | |
2464 | raw_spin_is_contended(&busiest->lock)) | |
2465 | break; | |
1e3c88bd | 2466 | #endif |
3d45fd80 | 2467 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2468 | |
2469 | return total_load_moved > 0; | |
2470 | } | |
2471 | ||
1e3c88bd PZ |
2472 | /********** Helpers for find_busiest_group ************************/ |
2473 | /* | |
2474 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2475 | * during load balancing. | |
2476 | */ | |
2477 | struct sd_lb_stats { | |
2478 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2479 | struct sched_group *this; /* Local group in this sd */ | |
2480 | unsigned long total_load; /* Total load of all groups in sd */ | |
2481 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2482 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2483 | ||
2484 | /** Statistics of this group */ | |
2485 | unsigned long this_load; | |
2486 | unsigned long this_load_per_task; | |
2487 | unsigned long this_nr_running; | |
fab47622 | 2488 | unsigned long this_has_capacity; |
aae6d3dd | 2489 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
2490 | |
2491 | /* Statistics of the busiest group */ | |
aae6d3dd | 2492 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
2493 | unsigned long max_load; |
2494 | unsigned long busiest_load_per_task; | |
2495 | unsigned long busiest_nr_running; | |
dd5feea1 | 2496 | unsigned long busiest_group_capacity; |
fab47622 | 2497 | unsigned long busiest_has_capacity; |
aae6d3dd | 2498 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
2499 | |
2500 | int group_imb; /* Is there imbalance in this sd */ | |
2501 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2502 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2503 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2504 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2505 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2506 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2507 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2508 | #endif | |
2509 | }; | |
2510 | ||
2511 | /* | |
2512 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2513 | */ | |
2514 | struct sg_lb_stats { | |
2515 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2516 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2517 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2518 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2519 | unsigned long group_capacity; | |
aae6d3dd SS |
2520 | unsigned long idle_cpus; |
2521 | unsigned long group_weight; | |
1e3c88bd | 2522 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 2523 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2524 | }; |
2525 | ||
2526 | /** | |
2527 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2528 | * @group: The group whose first cpu is to be returned. | |
2529 | */ | |
2530 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2531 | { | |
2532 | return cpumask_first(sched_group_cpus(group)); | |
2533 | } | |
2534 | ||
2535 | /** | |
2536 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2537 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2538 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2539 | */ | |
2540 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2541 | enum cpu_idle_type idle) | |
2542 | { | |
2543 | int load_idx; | |
2544 | ||
2545 | switch (idle) { | |
2546 | case CPU_NOT_IDLE: | |
2547 | load_idx = sd->busy_idx; | |
2548 | break; | |
2549 | ||
2550 | case CPU_NEWLY_IDLE: | |
2551 | load_idx = sd->newidle_idx; | |
2552 | break; | |
2553 | default: | |
2554 | load_idx = sd->idle_idx; | |
2555 | break; | |
2556 | } | |
2557 | ||
2558 | return load_idx; | |
2559 | } | |
2560 | ||
2561 | ||
2562 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2563 | /** | |
2564 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2565 | * the given sched_domain, during load balancing. | |
2566 | * | |
2567 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2568 | * @sds: Variable containing the statistics for sd. | |
2569 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
2570 | */ | |
2571 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2572 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2573 | { | |
2574 | /* | |
2575 | * Busy processors will not participate in power savings | |
2576 | * balance. | |
2577 | */ | |
2578 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
2579 | sds->power_savings_balance = 0; | |
2580 | else { | |
2581 | sds->power_savings_balance = 1; | |
2582 | sds->min_nr_running = ULONG_MAX; | |
2583 | sds->leader_nr_running = 0; | |
2584 | } | |
2585 | } | |
2586 | ||
2587 | /** | |
2588 | * update_sd_power_savings_stats - Update the power saving stats for a | |
2589 | * sched_domain while performing load balancing. | |
2590 | * | |
2591 | * @group: sched_group belonging to the sched_domain under consideration. | |
2592 | * @sds: Variable containing the statistics of the sched_domain | |
2593 | * @local_group: Does group contain the CPU for which we're performing | |
2594 | * load balancing ? | |
2595 | * @sgs: Variable containing the statistics of the group. | |
2596 | */ | |
2597 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2598 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2599 | { | |
2600 | ||
2601 | if (!sds->power_savings_balance) | |
2602 | return; | |
2603 | ||
2604 | /* | |
2605 | * If the local group is idle or completely loaded | |
2606 | * no need to do power savings balance at this domain | |
2607 | */ | |
2608 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
2609 | !sds->this_nr_running)) | |
2610 | sds->power_savings_balance = 0; | |
2611 | ||
2612 | /* | |
2613 | * If a group is already running at full capacity or idle, | |
2614 | * don't include that group in power savings calculations | |
2615 | */ | |
2616 | if (!sds->power_savings_balance || | |
2617 | sgs->sum_nr_running >= sgs->group_capacity || | |
2618 | !sgs->sum_nr_running) | |
2619 | return; | |
2620 | ||
2621 | /* | |
2622 | * Calculate the group which has the least non-idle load. | |
2623 | * This is the group from where we need to pick up the load | |
2624 | * for saving power | |
2625 | */ | |
2626 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
2627 | (sgs->sum_nr_running == sds->min_nr_running && | |
2628 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
2629 | sds->group_min = group; | |
2630 | sds->min_nr_running = sgs->sum_nr_running; | |
2631 | sds->min_load_per_task = sgs->sum_weighted_load / | |
2632 | sgs->sum_nr_running; | |
2633 | } | |
2634 | ||
2635 | /* | |
2636 | * Calculate the group which is almost near its | |
2637 | * capacity but still has some space to pick up some load | |
2638 | * from other group and save more power | |
2639 | */ | |
2640 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
2641 | return; | |
2642 | ||
2643 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
2644 | (sgs->sum_nr_running == sds->leader_nr_running && | |
2645 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
2646 | sds->group_leader = group; | |
2647 | sds->leader_nr_running = sgs->sum_nr_running; | |
2648 | } | |
2649 | } | |
2650 | ||
2651 | /** | |
2652 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
2653 | * @sds: Variable containing the statistics of the sched_domain | |
2654 | * under consideration. | |
2655 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
2656 | * @imbalance: Variable to store the imbalance. | |
2657 | * | |
2658 | * Description: | |
2659 | * Check if we have potential to perform some power-savings balance. | |
2660 | * If yes, set the busiest group to be the least loaded group in the | |
2661 | * sched_domain, so that it's CPUs can be put to idle. | |
2662 | * | |
2663 | * Returns 1 if there is potential to perform power-savings balance. | |
2664 | * Else returns 0. | |
2665 | */ | |
2666 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2667 | int this_cpu, unsigned long *imbalance) | |
2668 | { | |
2669 | if (!sds->power_savings_balance) | |
2670 | return 0; | |
2671 | ||
2672 | if (sds->this != sds->group_leader || | |
2673 | sds->group_leader == sds->group_min) | |
2674 | return 0; | |
2675 | ||
2676 | *imbalance = sds->min_load_per_task; | |
2677 | sds->busiest = sds->group_min; | |
2678 | ||
2679 | return 1; | |
2680 | ||
2681 | } | |
2682 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2683 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
2684 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
2685 | { | |
2686 | return; | |
2687 | } | |
2688 | ||
2689 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
2690 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
2691 | { | |
2692 | return; | |
2693 | } | |
2694 | ||
2695 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
2696 | int this_cpu, unsigned long *imbalance) | |
2697 | { | |
2698 | return 0; | |
2699 | } | |
2700 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
2701 | ||
2702 | ||
2703 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
2704 | { | |
1399fa78 | 2705 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
2706 | } |
2707 | ||
2708 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
2709 | { | |
2710 | return default_scale_freq_power(sd, cpu); | |
2711 | } | |
2712 | ||
2713 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
2714 | { | |
669c55e9 | 2715 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
2716 | unsigned long smt_gain = sd->smt_gain; |
2717 | ||
2718 | smt_gain /= weight; | |
2719 | ||
2720 | return smt_gain; | |
2721 | } | |
2722 | ||
2723 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
2724 | { | |
2725 | return default_scale_smt_power(sd, cpu); | |
2726 | } | |
2727 | ||
2728 | unsigned long scale_rt_power(int cpu) | |
2729 | { | |
2730 | struct rq *rq = cpu_rq(cpu); | |
2731 | u64 total, available; | |
2732 | ||
1e3c88bd | 2733 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
2734 | |
2735 | if (unlikely(total < rq->rt_avg)) { | |
2736 | /* Ensures that power won't end up being negative */ | |
2737 | available = 0; | |
2738 | } else { | |
2739 | available = total - rq->rt_avg; | |
2740 | } | |
1e3c88bd | 2741 | |
1399fa78 NR |
2742 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
2743 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 2744 | |
1399fa78 | 2745 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
2746 | |
2747 | return div_u64(available, total); | |
2748 | } | |
2749 | ||
2750 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
2751 | { | |
669c55e9 | 2752 | unsigned long weight = sd->span_weight; |
1399fa78 | 2753 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
2754 | struct sched_group *sdg = sd->groups; |
2755 | ||
1e3c88bd PZ |
2756 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
2757 | if (sched_feat(ARCH_POWER)) | |
2758 | power *= arch_scale_smt_power(sd, cpu); | |
2759 | else | |
2760 | power *= default_scale_smt_power(sd, cpu); | |
2761 | ||
1399fa78 | 2762 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
2763 | } |
2764 | ||
9c3f75cb | 2765 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
2766 | |
2767 | if (sched_feat(ARCH_POWER)) | |
2768 | power *= arch_scale_freq_power(sd, cpu); | |
2769 | else | |
2770 | power *= default_scale_freq_power(sd, cpu); | |
2771 | ||
1399fa78 | 2772 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 2773 | |
1e3c88bd | 2774 | power *= scale_rt_power(cpu); |
1399fa78 | 2775 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
2776 | |
2777 | if (!power) | |
2778 | power = 1; | |
2779 | ||
e51fd5e2 | 2780 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 2781 | sdg->sgp->power = power; |
1e3c88bd PZ |
2782 | } |
2783 | ||
2784 | static void update_group_power(struct sched_domain *sd, int cpu) | |
2785 | { | |
2786 | struct sched_domain *child = sd->child; | |
2787 | struct sched_group *group, *sdg = sd->groups; | |
2788 | unsigned long power; | |
2789 | ||
2790 | if (!child) { | |
2791 | update_cpu_power(sd, cpu); | |
2792 | return; | |
2793 | } | |
2794 | ||
2795 | power = 0; | |
2796 | ||
2797 | group = child->groups; | |
2798 | do { | |
9c3f75cb | 2799 | power += group->sgp->power; |
1e3c88bd PZ |
2800 | group = group->next; |
2801 | } while (group != child->groups); | |
2802 | ||
9c3f75cb | 2803 | sdg->sgp->power = power; |
1e3c88bd PZ |
2804 | } |
2805 | ||
9d5efe05 SV |
2806 | /* |
2807 | * Try and fix up capacity for tiny siblings, this is needed when | |
2808 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
2809 | * which on its own isn't powerful enough. | |
2810 | * | |
2811 | * See update_sd_pick_busiest() and check_asym_packing(). | |
2812 | */ | |
2813 | static inline int | |
2814 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
2815 | { | |
2816 | /* | |
1399fa78 | 2817 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 2818 | */ |
a6c75f2f | 2819 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
2820 | return 0; |
2821 | ||
2822 | /* | |
2823 | * If ~90% of the cpu_power is still there, we're good. | |
2824 | */ | |
9c3f75cb | 2825 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
2826 | return 1; |
2827 | ||
2828 | return 0; | |
2829 | } | |
2830 | ||
1e3c88bd PZ |
2831 | /** |
2832 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
2833 | * @sd: The sched_domain whose statistics are to be updated. | |
2834 | * @group: sched_group whose statistics are to be updated. | |
2835 | * @this_cpu: Cpu for which load balance is currently performed. | |
2836 | * @idle: Idle status of this_cpu | |
2837 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
1e3c88bd PZ |
2838 | * @local_group: Does group contain this_cpu. |
2839 | * @cpus: Set of cpus considered for load balancing. | |
2840 | * @balance: Should we balance. | |
2841 | * @sgs: variable to hold the statistics for this group. | |
2842 | */ | |
2843 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
2844 | struct sched_group *group, int this_cpu, | |
46e49b38 | 2845 | enum cpu_idle_type idle, int load_idx, |
1e3c88bd PZ |
2846 | int local_group, const struct cpumask *cpus, |
2847 | int *balance, struct sg_lb_stats *sgs) | |
2848 | { | |
2582f0eb | 2849 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
2850 | int i; |
2851 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 2852 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 2853 | |
871e35bc | 2854 | if (local_group) |
1e3c88bd | 2855 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
2856 | |
2857 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
2858 | max_cpu_load = 0; |
2859 | min_cpu_load = ~0UL; | |
2582f0eb | 2860 | max_nr_running = 0; |
1e3c88bd PZ |
2861 | |
2862 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
2863 | struct rq *rq = cpu_rq(i); | |
2864 | ||
1e3c88bd PZ |
2865 | /* Bias balancing toward cpus of our domain */ |
2866 | if (local_group) { | |
2867 | if (idle_cpu(i) && !first_idle_cpu) { | |
2868 | first_idle_cpu = 1; | |
2869 | balance_cpu = i; | |
2870 | } | |
2871 | ||
2872 | load = target_load(i, load_idx); | |
2873 | } else { | |
2874 | load = source_load(i, load_idx); | |
2582f0eb | 2875 | if (load > max_cpu_load) { |
1e3c88bd | 2876 | max_cpu_load = load; |
2582f0eb NR |
2877 | max_nr_running = rq->nr_running; |
2878 | } | |
1e3c88bd PZ |
2879 | if (min_cpu_load > load) |
2880 | min_cpu_load = load; | |
2881 | } | |
2882 | ||
2883 | sgs->group_load += load; | |
2884 | sgs->sum_nr_running += rq->nr_running; | |
2885 | sgs->sum_weighted_load += weighted_cpuload(i); | |
aae6d3dd SS |
2886 | if (idle_cpu(i)) |
2887 | sgs->idle_cpus++; | |
1e3c88bd PZ |
2888 | } |
2889 | ||
2890 | /* | |
2891 | * First idle cpu or the first cpu(busiest) in this sched group | |
2892 | * is eligible for doing load balancing at this and above | |
2893 | * domains. In the newly idle case, we will allow all the cpu's | |
2894 | * to do the newly idle load balance. | |
2895 | */ | |
bbc8cb5b PZ |
2896 | if (idle != CPU_NEWLY_IDLE && local_group) { |
2897 | if (balance_cpu != this_cpu) { | |
2898 | *balance = 0; | |
2899 | return; | |
2900 | } | |
2901 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
2902 | } |
2903 | ||
2904 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2905 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 2906 | |
1e3c88bd PZ |
2907 | /* |
2908 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 2909 | * than the average weight of a task. |
1e3c88bd PZ |
2910 | * |
2911 | * APZ: with cgroup the avg task weight can vary wildly and | |
2912 | * might not be a suitable number - should we keep a | |
2913 | * normalized nr_running number somewhere that negates | |
2914 | * the hierarchy? | |
2915 | */ | |
dd5feea1 SS |
2916 | if (sgs->sum_nr_running) |
2917 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 2918 | |
866ab43e | 2919 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
2920 | sgs->group_imb = 1; |
2921 | ||
9c3f75cb | 2922 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 2923 | SCHED_POWER_SCALE); |
9d5efe05 SV |
2924 | if (!sgs->group_capacity) |
2925 | sgs->group_capacity = fix_small_capacity(sd, group); | |
aae6d3dd | 2926 | sgs->group_weight = group->group_weight; |
fab47622 NR |
2927 | |
2928 | if (sgs->group_capacity > sgs->sum_nr_running) | |
2929 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
2930 | } |
2931 | ||
532cb4c4 MN |
2932 | /** |
2933 | * update_sd_pick_busiest - return 1 on busiest group | |
2934 | * @sd: sched_domain whose statistics are to be checked | |
2935 | * @sds: sched_domain statistics | |
2936 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
2937 | * @sgs: sched_group statistics |
2938 | * @this_cpu: the current cpu | |
532cb4c4 MN |
2939 | * |
2940 | * Determine if @sg is a busier group than the previously selected | |
2941 | * busiest group. | |
2942 | */ | |
2943 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
2944 | struct sd_lb_stats *sds, | |
2945 | struct sched_group *sg, | |
2946 | struct sg_lb_stats *sgs, | |
2947 | int this_cpu) | |
2948 | { | |
2949 | if (sgs->avg_load <= sds->max_load) | |
2950 | return false; | |
2951 | ||
2952 | if (sgs->sum_nr_running > sgs->group_capacity) | |
2953 | return true; | |
2954 | ||
2955 | if (sgs->group_imb) | |
2956 | return true; | |
2957 | ||
2958 | /* | |
2959 | * ASYM_PACKING needs to move all the work to the lowest | |
2960 | * numbered CPUs in the group, therefore mark all groups | |
2961 | * higher than ourself as busy. | |
2962 | */ | |
2963 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
2964 | this_cpu < group_first_cpu(sg)) { | |
2965 | if (!sds->busiest) | |
2966 | return true; | |
2967 | ||
2968 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
2969 | return true; | |
2970 | } | |
2971 | ||
2972 | return false; | |
2973 | } | |
2974 | ||
1e3c88bd PZ |
2975 | /** |
2976 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
2977 | * @sd: sched_domain whose statistics are to be updated. | |
2978 | * @this_cpu: Cpu for which load balance is currently performed. | |
2979 | * @idle: Idle status of this_cpu | |
1e3c88bd PZ |
2980 | * @cpus: Set of cpus considered for load balancing. |
2981 | * @balance: Should we balance. | |
2982 | * @sds: variable to hold the statistics for this sched_domain. | |
2983 | */ | |
2984 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
46e49b38 VP |
2985 | enum cpu_idle_type idle, const struct cpumask *cpus, |
2986 | int *balance, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
2987 | { |
2988 | struct sched_domain *child = sd->child; | |
532cb4c4 | 2989 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
2990 | struct sg_lb_stats sgs; |
2991 | int load_idx, prefer_sibling = 0; | |
2992 | ||
2993 | if (child && child->flags & SD_PREFER_SIBLING) | |
2994 | prefer_sibling = 1; | |
2995 | ||
2996 | init_sd_power_savings_stats(sd, sds, idle); | |
2997 | load_idx = get_sd_load_idx(sd, idle); | |
2998 | ||
2999 | do { | |
3000 | int local_group; | |
3001 | ||
532cb4c4 | 3002 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 3003 | memset(&sgs, 0, sizeof(sgs)); |
46e49b38 | 3004 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, |
1e3c88bd PZ |
3005 | local_group, cpus, balance, &sgs); |
3006 | ||
8f190fb3 | 3007 | if (local_group && !(*balance)) |
1e3c88bd PZ |
3008 | return; |
3009 | ||
3010 | sds->total_load += sgs.group_load; | |
9c3f75cb | 3011 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
3012 | |
3013 | /* | |
3014 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 3015 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
3016 | * and move all the excess tasks away. We lower the capacity |
3017 | * of a group only if the local group has the capacity to fit | |
3018 | * these excess tasks, i.e. nr_running < group_capacity. The | |
3019 | * extra check prevents the case where you always pull from the | |
3020 | * heaviest group when it is already under-utilized (possible | |
3021 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 3022 | */ |
75dd321d | 3023 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
3024 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
3025 | ||
3026 | if (local_group) { | |
3027 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 3028 | sds->this = sg; |
1e3c88bd PZ |
3029 | sds->this_nr_running = sgs.sum_nr_running; |
3030 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 3031 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 3032 | sds->this_idle_cpus = sgs.idle_cpus; |
532cb4c4 | 3033 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 3034 | sds->max_load = sgs.avg_load; |
532cb4c4 | 3035 | sds->busiest = sg; |
1e3c88bd | 3036 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 3037 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 3038 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 3039 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 3040 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 3041 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
3042 | sds->group_imb = sgs.group_imb; |
3043 | } | |
3044 | ||
532cb4c4 MN |
3045 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
3046 | sg = sg->next; | |
3047 | } while (sg != sd->groups); | |
3048 | } | |
3049 | ||
2ec57d44 | 3050 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
3051 | { |
3052 | return 0*SD_ASYM_PACKING; | |
3053 | } | |
3054 | ||
3055 | /** | |
3056 | * check_asym_packing - Check to see if the group is packed into the | |
3057 | * sched doman. | |
3058 | * | |
3059 | * This is primarily intended to used at the sibling level. Some | |
3060 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
3061 | * case of POWER7, it can move to lower SMT modes only when higher | |
3062 | * threads are idle. When in lower SMT modes, the threads will | |
3063 | * perform better since they share less core resources. Hence when we | |
3064 | * have idle threads, we want them to be the higher ones. | |
3065 | * | |
3066 | * This packing function is run on idle threads. It checks to see if | |
3067 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
3068 | * CPU number than the packing function is being run on. Here we are | |
3069 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
3070 | * number. | |
3071 | * | |
b6b12294 MN |
3072 | * Returns 1 when packing is required and a task should be moved to |
3073 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
3074 | * | |
532cb4c4 MN |
3075 | * @sd: The sched_domain whose packing is to be checked. |
3076 | * @sds: Statistics of the sched_domain which is to be packed | |
3077 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
3078 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
3079 | */ |
3080 | static int check_asym_packing(struct sched_domain *sd, | |
3081 | struct sd_lb_stats *sds, | |
3082 | int this_cpu, unsigned long *imbalance) | |
3083 | { | |
3084 | int busiest_cpu; | |
3085 | ||
3086 | if (!(sd->flags & SD_ASYM_PACKING)) | |
3087 | return 0; | |
3088 | ||
3089 | if (!sds->busiest) | |
3090 | return 0; | |
3091 | ||
3092 | busiest_cpu = group_first_cpu(sds->busiest); | |
3093 | if (this_cpu > busiest_cpu) | |
3094 | return 0; | |
3095 | ||
9c3f75cb | 3096 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power, |
1399fa78 | 3097 | SCHED_POWER_SCALE); |
532cb4c4 | 3098 | return 1; |
1e3c88bd PZ |
3099 | } |
3100 | ||
3101 | /** | |
3102 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
3103 | * amongst the groups of a sched_domain, during | |
3104 | * load balancing. | |
3105 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
3106 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
3107 | * @imbalance: Variable to store the imbalance. | |
3108 | */ | |
3109 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
3110 | int this_cpu, unsigned long *imbalance) | |
3111 | { | |
3112 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
3113 | unsigned int imbn = 2; | |
dd5feea1 | 3114 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
3115 | |
3116 | if (sds->this_nr_running) { | |
3117 | sds->this_load_per_task /= sds->this_nr_running; | |
3118 | if (sds->busiest_load_per_task > | |
3119 | sds->this_load_per_task) | |
3120 | imbn = 1; | |
3121 | } else | |
3122 | sds->this_load_per_task = | |
3123 | cpu_avg_load_per_task(this_cpu); | |
3124 | ||
dd5feea1 | 3125 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 3126 | * SCHED_POWER_SCALE; |
9c3f75cb | 3127 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
3128 | |
3129 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
3130 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
3131 | *imbalance = sds->busiest_load_per_task; |
3132 | return; | |
3133 | } | |
3134 | ||
3135 | /* | |
3136 | * OK, we don't have enough imbalance to justify moving tasks, | |
3137 | * however we may be able to increase total CPU power used by | |
3138 | * moving them. | |
3139 | */ | |
3140 | ||
9c3f75cb | 3141 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 3142 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 3143 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 3144 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 3145 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
3146 | |
3147 | /* Amount of load we'd subtract */ | |
1399fa78 | 3148 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 3149 | sds->busiest->sgp->power; |
1e3c88bd | 3150 | if (sds->max_load > tmp) |
9c3f75cb | 3151 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
3152 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
3153 | ||
3154 | /* Amount of load we'd add */ | |
9c3f75cb | 3155 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 3156 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
3157 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
3158 | sds->this->sgp->power; | |
1e3c88bd | 3159 | else |
1399fa78 | 3160 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
3161 | sds->this->sgp->power; |
3162 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 3163 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 3164 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
3165 | |
3166 | /* Move if we gain throughput */ | |
3167 | if (pwr_move > pwr_now) | |
3168 | *imbalance = sds->busiest_load_per_task; | |
3169 | } | |
3170 | ||
3171 | /** | |
3172 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
3173 | * groups of a given sched_domain during load balance. | |
3174 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
3175 | * @this_cpu: Cpu for which currently load balance is being performed. | |
3176 | * @imbalance: The variable to store the imbalance. | |
3177 | */ | |
3178 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
3179 | unsigned long *imbalance) | |
3180 | { | |
dd5feea1 SS |
3181 | unsigned long max_pull, load_above_capacity = ~0UL; |
3182 | ||
3183 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
3184 | if (sds->group_imb) { | |
3185 | sds->busiest_load_per_task = | |
3186 | min(sds->busiest_load_per_task, sds->avg_load); | |
3187 | } | |
3188 | ||
1e3c88bd PZ |
3189 | /* |
3190 | * In the presence of smp nice balancing, certain scenarios can have | |
3191 | * max load less than avg load(as we skip the groups at or below | |
3192 | * its cpu_power, while calculating max_load..) | |
3193 | */ | |
3194 | if (sds->max_load < sds->avg_load) { | |
3195 | *imbalance = 0; | |
3196 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3197 | } | |
3198 | ||
dd5feea1 SS |
3199 | if (!sds->group_imb) { |
3200 | /* | |
3201 | * Don't want to pull so many tasks that a group would go idle. | |
3202 | */ | |
3203 | load_above_capacity = (sds->busiest_nr_running - | |
3204 | sds->busiest_group_capacity); | |
3205 | ||
1399fa78 | 3206 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 3207 | |
9c3f75cb | 3208 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
3209 | } |
3210 | ||
3211 | /* | |
3212 | * We're trying to get all the cpus to the average_load, so we don't | |
3213 | * want to push ourselves above the average load, nor do we wish to | |
3214 | * reduce the max loaded cpu below the average load. At the same time, | |
3215 | * we also don't want to reduce the group load below the group capacity | |
3216 | * (so that we can implement power-savings policies etc). Thus we look | |
3217 | * for the minimum possible imbalance. | |
3218 | * Be careful of negative numbers as they'll appear as very large values | |
3219 | * with unsigned longs. | |
3220 | */ | |
3221 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
3222 | |
3223 | /* How much load to actually move to equalise the imbalance */ | |
9c3f75cb PZ |
3224 | *imbalance = min(max_pull * sds->busiest->sgp->power, |
3225 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) | |
1399fa78 | 3226 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
3227 | |
3228 | /* | |
3229 | * if *imbalance is less than the average load per runnable task | |
25985edc | 3230 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
3231 | * a think about bumping its value to force at least one task to be |
3232 | * moved | |
3233 | */ | |
3234 | if (*imbalance < sds->busiest_load_per_task) | |
3235 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3236 | ||
3237 | } | |
fab47622 | 3238 | |
1e3c88bd PZ |
3239 | /******* find_busiest_group() helpers end here *********************/ |
3240 | ||
3241 | /** | |
3242 | * find_busiest_group - Returns the busiest group within the sched_domain | |
3243 | * if there is an imbalance. If there isn't an imbalance, and | |
3244 | * the user has opted for power-savings, it returns a group whose | |
3245 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
3246 | * such a group exists. | |
3247 | * | |
3248 | * Also calculates the amount of weighted load which should be moved | |
3249 | * to restore balance. | |
3250 | * | |
3251 | * @sd: The sched_domain whose busiest group is to be returned. | |
3252 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
3253 | * @imbalance: Variable which stores amount of weighted load which should | |
3254 | * be moved to restore balance/put a group to idle. | |
3255 | * @idle: The idle status of this_cpu. | |
1e3c88bd PZ |
3256 | * @cpus: The set of CPUs under consideration for load-balancing. |
3257 | * @balance: Pointer to a variable indicating if this_cpu | |
3258 | * is the appropriate cpu to perform load balancing at this_level. | |
3259 | * | |
3260 | * Returns: - the busiest group if imbalance exists. | |
3261 | * - If no imbalance and user has opted for power-savings balance, | |
3262 | * return the least loaded group whose CPUs can be | |
3263 | * put to idle by rebalancing its tasks onto our group. | |
3264 | */ | |
3265 | static struct sched_group * | |
3266 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
3267 | unsigned long *imbalance, enum cpu_idle_type idle, | |
46e49b38 | 3268 | const struct cpumask *cpus, int *balance) |
1e3c88bd PZ |
3269 | { |
3270 | struct sd_lb_stats sds; | |
3271 | ||
3272 | memset(&sds, 0, sizeof(sds)); | |
3273 | ||
3274 | /* | |
3275 | * Compute the various statistics relavent for load balancing at | |
3276 | * this level. | |
3277 | */ | |
46e49b38 | 3278 | update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds); |
1e3c88bd | 3279 | |
cc57aa8f PZ |
3280 | /* |
3281 | * this_cpu is not the appropriate cpu to perform load balancing at | |
3282 | * this level. | |
1e3c88bd | 3283 | */ |
8f190fb3 | 3284 | if (!(*balance)) |
1e3c88bd PZ |
3285 | goto ret; |
3286 | ||
532cb4c4 MN |
3287 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
3288 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
3289 | return sds.busiest; | |
3290 | ||
cc57aa8f | 3291 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
3292 | if (!sds.busiest || sds.busiest_nr_running == 0) |
3293 | goto out_balanced; | |
3294 | ||
1399fa78 | 3295 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 3296 | |
866ab43e PZ |
3297 | /* |
3298 | * If the busiest group is imbalanced the below checks don't | |
3299 | * work because they assumes all things are equal, which typically | |
3300 | * isn't true due to cpus_allowed constraints and the like. | |
3301 | */ | |
3302 | if (sds.group_imb) | |
3303 | goto force_balance; | |
3304 | ||
cc57aa8f | 3305 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
fab47622 NR |
3306 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
3307 | !sds.busiest_has_capacity) | |
3308 | goto force_balance; | |
3309 | ||
cc57aa8f PZ |
3310 | /* |
3311 | * If the local group is more busy than the selected busiest group | |
3312 | * don't try and pull any tasks. | |
3313 | */ | |
1e3c88bd PZ |
3314 | if (sds.this_load >= sds.max_load) |
3315 | goto out_balanced; | |
3316 | ||
cc57aa8f PZ |
3317 | /* |
3318 | * Don't pull any tasks if this group is already above the domain | |
3319 | * average load. | |
3320 | */ | |
1e3c88bd PZ |
3321 | if (sds.this_load >= sds.avg_load) |
3322 | goto out_balanced; | |
3323 | ||
c186fafe | 3324 | if (idle == CPU_IDLE) { |
aae6d3dd SS |
3325 | /* |
3326 | * This cpu is idle. If the busiest group load doesn't | |
3327 | * have more tasks than the number of available cpu's and | |
3328 | * there is no imbalance between this and busiest group | |
3329 | * wrt to idle cpu's, it is balanced. | |
3330 | */ | |
c186fafe | 3331 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
3332 | sds.busiest_nr_running <= sds.busiest_group_weight) |
3333 | goto out_balanced; | |
c186fafe PZ |
3334 | } else { |
3335 | /* | |
3336 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
3337 | * imbalance_pct to be conservative. | |
3338 | */ | |
3339 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
3340 | goto out_balanced; | |
aae6d3dd | 3341 | } |
1e3c88bd | 3342 | |
fab47622 | 3343 | force_balance: |
1e3c88bd PZ |
3344 | /* Looks like there is an imbalance. Compute it */ |
3345 | calculate_imbalance(&sds, this_cpu, imbalance); | |
3346 | return sds.busiest; | |
3347 | ||
3348 | out_balanced: | |
3349 | /* | |
3350 | * There is no obvious imbalance. But check if we can do some balancing | |
3351 | * to save power. | |
3352 | */ | |
3353 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
3354 | return sds.busiest; | |
3355 | ret: | |
3356 | *imbalance = 0; | |
3357 | return NULL; | |
3358 | } | |
3359 | ||
3360 | /* | |
3361 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3362 | */ | |
3363 | static struct rq * | |
9d5efe05 SV |
3364 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
3365 | enum cpu_idle_type idle, unsigned long imbalance, | |
3366 | const struct cpumask *cpus) | |
1e3c88bd PZ |
3367 | { |
3368 | struct rq *busiest = NULL, *rq; | |
3369 | unsigned long max_load = 0; | |
3370 | int i; | |
3371 | ||
3372 | for_each_cpu(i, sched_group_cpus(group)) { | |
3373 | unsigned long power = power_of(i); | |
1399fa78 NR |
3374 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
3375 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
3376 | unsigned long wl; |
3377 | ||
9d5efe05 SV |
3378 | if (!capacity) |
3379 | capacity = fix_small_capacity(sd, group); | |
3380 | ||
1e3c88bd PZ |
3381 | if (!cpumask_test_cpu(i, cpus)) |
3382 | continue; | |
3383 | ||
3384 | rq = cpu_rq(i); | |
6e40f5bb | 3385 | wl = weighted_cpuload(i); |
1e3c88bd | 3386 | |
6e40f5bb TG |
3387 | /* |
3388 | * When comparing with imbalance, use weighted_cpuload() | |
3389 | * which is not scaled with the cpu power. | |
3390 | */ | |
1e3c88bd PZ |
3391 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
3392 | continue; | |
3393 | ||
6e40f5bb TG |
3394 | /* |
3395 | * For the load comparisons with the other cpu's, consider | |
3396 | * the weighted_cpuload() scaled with the cpu power, so that | |
3397 | * the load can be moved away from the cpu that is potentially | |
3398 | * running at a lower capacity. | |
3399 | */ | |
1399fa78 | 3400 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 3401 | |
1e3c88bd PZ |
3402 | if (wl > max_load) { |
3403 | max_load = wl; | |
3404 | busiest = rq; | |
3405 | } | |
3406 | } | |
3407 | ||
3408 | return busiest; | |
3409 | } | |
3410 | ||
3411 | /* | |
3412 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3413 | * so long as it is large enough. | |
3414 | */ | |
3415 | #define MAX_PINNED_INTERVAL 512 | |
3416 | ||
3417 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
3418 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
3419 | ||
46e49b38 | 3420 | static int need_active_balance(struct sched_domain *sd, int idle, |
532cb4c4 | 3421 | int busiest_cpu, int this_cpu) |
1af3ed3d PZ |
3422 | { |
3423 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
3424 | |
3425 | /* | |
3426 | * ASYM_PACKING needs to force migrate tasks from busy but | |
3427 | * higher numbered CPUs in order to pack all tasks in the | |
3428 | * lowest numbered CPUs. | |
3429 | */ | |
3430 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
3431 | return 1; | |
3432 | ||
1af3ed3d PZ |
3433 | /* |
3434 | * The only task running in a non-idle cpu can be moved to this | |
3435 | * cpu in an attempt to completely freeup the other CPU | |
3436 | * package. | |
3437 | * | |
3438 | * The package power saving logic comes from | |
3439 | * find_busiest_group(). If there are no imbalance, then | |
3440 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3441 | * f_b_g() will select a group from which a running task may be | |
3442 | * pulled to this cpu in order to make the other package idle. | |
3443 | * If there is no opportunity to make a package idle and if | |
3444 | * there are no imbalance, then f_b_g() will return NULL and no | |
3445 | * action will be taken in load_balance_newidle(). | |
3446 | * | |
3447 | * Under normal task pull operation due to imbalance, there | |
3448 | * will be more than one task in the source run queue and | |
3449 | * move_tasks() will succeed. ld_moved will be true and this | |
3450 | * active balance code will not be triggered. | |
3451 | */ | |
1af3ed3d PZ |
3452 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) |
3453 | return 0; | |
3454 | } | |
3455 | ||
3456 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3457 | } | |
3458 | ||
969c7921 TH |
3459 | static int active_load_balance_cpu_stop(void *data); |
3460 | ||
1e3c88bd PZ |
3461 | /* |
3462 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3463 | * tasks if there is an imbalance. | |
3464 | */ | |
3465 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3466 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3467 | int *balance) | |
3468 | { | |
46e49b38 | 3469 | int ld_moved, all_pinned = 0, active_balance = 0; |
1e3c88bd PZ |
3470 | struct sched_group *group; |
3471 | unsigned long imbalance; | |
3472 | struct rq *busiest; | |
3473 | unsigned long flags; | |
3474 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3475 | ||
3476 | cpumask_copy(cpus, cpu_active_mask); | |
3477 | ||
1e3c88bd PZ |
3478 | schedstat_inc(sd, lb_count[idle]); |
3479 | ||
3480 | redo: | |
46e49b38 | 3481 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, |
1e3c88bd PZ |
3482 | cpus, balance); |
3483 | ||
3484 | if (*balance == 0) | |
3485 | goto out_balanced; | |
3486 | ||
3487 | if (!group) { | |
3488 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3489 | goto out_balanced; | |
3490 | } | |
3491 | ||
9d5efe05 | 3492 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3493 | if (!busiest) { |
3494 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3495 | goto out_balanced; | |
3496 | } | |
3497 | ||
3498 | BUG_ON(busiest == this_rq); | |
3499 | ||
3500 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3501 | ||
3502 | ld_moved = 0; | |
3503 | if (busiest->nr_running > 1) { | |
3504 | /* | |
3505 | * Attempt to move tasks. If find_busiest_group has found | |
3506 | * an imbalance but busiest->nr_running <= 1, the group is | |
3507 | * still unbalanced. ld_moved simply stays zero, so it is | |
3508 | * correctly treated as an imbalance. | |
3509 | */ | |
b30aef17 | 3510 | all_pinned = 1; |
1e3c88bd PZ |
3511 | local_irq_save(flags); |
3512 | double_rq_lock(this_rq, busiest); | |
3513 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3514 | imbalance, sd, idle, &all_pinned); | |
3515 | double_rq_unlock(this_rq, busiest); | |
3516 | local_irq_restore(flags); | |
3517 | ||
3518 | /* | |
3519 | * some other cpu did the load balance for us. | |
3520 | */ | |
3521 | if (ld_moved && this_cpu != smp_processor_id()) | |
3522 | resched_cpu(this_cpu); | |
3523 | ||
3524 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3525 | if (unlikely(all_pinned)) { | |
3526 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3527 | if (!cpumask_empty(cpus)) | |
3528 | goto redo; | |
3529 | goto out_balanced; | |
3530 | } | |
3531 | } | |
3532 | ||
3533 | if (!ld_moved) { | |
3534 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3535 | /* |
3536 | * Increment the failure counter only on periodic balance. | |
3537 | * We do not want newidle balance, which can be very | |
3538 | * frequent, pollute the failure counter causing | |
3539 | * excessive cache_hot migrations and active balances. | |
3540 | */ | |
3541 | if (idle != CPU_NEWLY_IDLE) | |
3542 | sd->nr_balance_failed++; | |
1e3c88bd | 3543 | |
46e49b38 | 3544 | if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) { |
1e3c88bd PZ |
3545 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3546 | ||
969c7921 TH |
3547 | /* don't kick the active_load_balance_cpu_stop, |
3548 | * if the curr task on busiest cpu can't be | |
3549 | * moved to this_cpu | |
1e3c88bd PZ |
3550 | */ |
3551 | if (!cpumask_test_cpu(this_cpu, | |
3552 | &busiest->curr->cpus_allowed)) { | |
3553 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3554 | flags); | |
3555 | all_pinned = 1; | |
3556 | goto out_one_pinned; | |
3557 | } | |
3558 | ||
969c7921 TH |
3559 | /* |
3560 | * ->active_balance synchronizes accesses to | |
3561 | * ->active_balance_work. Once set, it's cleared | |
3562 | * only after active load balance is finished. | |
3563 | */ | |
1e3c88bd PZ |
3564 | if (!busiest->active_balance) { |
3565 | busiest->active_balance = 1; | |
3566 | busiest->push_cpu = this_cpu; | |
3567 | active_balance = 1; | |
3568 | } | |
3569 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 3570 | |
1e3c88bd | 3571 | if (active_balance) |
969c7921 TH |
3572 | stop_one_cpu_nowait(cpu_of(busiest), |
3573 | active_load_balance_cpu_stop, busiest, | |
3574 | &busiest->active_balance_work); | |
1e3c88bd PZ |
3575 | |
3576 | /* | |
3577 | * We've kicked active balancing, reset the failure | |
3578 | * counter. | |
3579 | */ | |
3580 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
3581 | } | |
3582 | } else | |
3583 | sd->nr_balance_failed = 0; | |
3584 | ||
3585 | if (likely(!active_balance)) { | |
3586 | /* We were unbalanced, so reset the balancing interval */ | |
3587 | sd->balance_interval = sd->min_interval; | |
3588 | } else { | |
3589 | /* | |
3590 | * If we've begun active balancing, start to back off. This | |
3591 | * case may not be covered by the all_pinned logic if there | |
3592 | * is only 1 task on the busy runqueue (because we don't call | |
3593 | * move_tasks). | |
3594 | */ | |
3595 | if (sd->balance_interval < sd->max_interval) | |
3596 | sd->balance_interval *= 2; | |
3597 | } | |
3598 | ||
1e3c88bd PZ |
3599 | goto out; |
3600 | ||
3601 | out_balanced: | |
3602 | schedstat_inc(sd, lb_balanced[idle]); | |
3603 | ||
3604 | sd->nr_balance_failed = 0; | |
3605 | ||
3606 | out_one_pinned: | |
3607 | /* tune up the balancing interval */ | |
3608 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
3609 | (sd->balance_interval < sd->max_interval)) | |
3610 | sd->balance_interval *= 2; | |
3611 | ||
46e49b38 | 3612 | ld_moved = 0; |
1e3c88bd | 3613 | out: |
1e3c88bd PZ |
3614 | return ld_moved; |
3615 | } | |
3616 | ||
1e3c88bd PZ |
3617 | /* |
3618 | * idle_balance is called by schedule() if this_cpu is about to become | |
3619 | * idle. Attempts to pull tasks from other CPUs. | |
3620 | */ | |
3621 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
3622 | { | |
3623 | struct sched_domain *sd; | |
3624 | int pulled_task = 0; | |
3625 | unsigned long next_balance = jiffies + HZ; | |
3626 | ||
3627 | this_rq->idle_stamp = this_rq->clock; | |
3628 | ||
3629 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
3630 | return; | |
3631 | ||
f492e12e PZ |
3632 | /* |
3633 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
3634 | */ | |
3635 | raw_spin_unlock(&this_rq->lock); | |
3636 | ||
c66eaf61 | 3637 | update_shares(this_cpu); |
dce840a0 | 3638 | rcu_read_lock(); |
1e3c88bd PZ |
3639 | for_each_domain(this_cpu, sd) { |
3640 | unsigned long interval; | |
f492e12e | 3641 | int balance = 1; |
1e3c88bd PZ |
3642 | |
3643 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3644 | continue; | |
3645 | ||
f492e12e | 3646 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 3647 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
3648 | pulled_task = load_balance(this_cpu, this_rq, |
3649 | sd, CPU_NEWLY_IDLE, &balance); | |
3650 | } | |
1e3c88bd PZ |
3651 | |
3652 | interval = msecs_to_jiffies(sd->balance_interval); | |
3653 | if (time_after(next_balance, sd->last_balance + interval)) | |
3654 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
3655 | if (pulled_task) { |
3656 | this_rq->idle_stamp = 0; | |
1e3c88bd | 3657 | break; |
d5ad140b | 3658 | } |
1e3c88bd | 3659 | } |
dce840a0 | 3660 | rcu_read_unlock(); |
f492e12e PZ |
3661 | |
3662 | raw_spin_lock(&this_rq->lock); | |
3663 | ||
1e3c88bd PZ |
3664 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
3665 | /* | |
3666 | * We are going idle. next_balance may be set based on | |
3667 | * a busy processor. So reset next_balance. | |
3668 | */ | |
3669 | this_rq->next_balance = next_balance; | |
3670 | } | |
3671 | } | |
3672 | ||
3673 | /* | |
969c7921 TH |
3674 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
3675 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
3676 | * least 1 task to be running on each physical CPU where possible, and | |
3677 | * avoids physical / logical imbalances. | |
1e3c88bd | 3678 | */ |
969c7921 | 3679 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 3680 | { |
969c7921 TH |
3681 | struct rq *busiest_rq = data; |
3682 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 3683 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 3684 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 3685 | struct sched_domain *sd; |
969c7921 TH |
3686 | |
3687 | raw_spin_lock_irq(&busiest_rq->lock); | |
3688 | ||
3689 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
3690 | if (unlikely(busiest_cpu != smp_processor_id() || | |
3691 | !busiest_rq->active_balance)) | |
3692 | goto out_unlock; | |
1e3c88bd PZ |
3693 | |
3694 | /* Is there any task to move? */ | |
3695 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 3696 | goto out_unlock; |
1e3c88bd PZ |
3697 | |
3698 | /* | |
3699 | * This condition is "impossible", if it occurs | |
3700 | * we need to fix it. Originally reported by | |
3701 | * Bjorn Helgaas on a 128-cpu setup. | |
3702 | */ | |
3703 | BUG_ON(busiest_rq == target_rq); | |
3704 | ||
3705 | /* move a task from busiest_rq to target_rq */ | |
3706 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
3707 | |
3708 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 3709 | rcu_read_lock(); |
1e3c88bd PZ |
3710 | for_each_domain(target_cpu, sd) { |
3711 | if ((sd->flags & SD_LOAD_BALANCE) && | |
3712 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
3713 | break; | |
3714 | } | |
3715 | ||
3716 | if (likely(sd)) { | |
3717 | schedstat_inc(sd, alb_count); | |
3718 | ||
3719 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
3720 | sd, CPU_IDLE)) | |
3721 | schedstat_inc(sd, alb_pushed); | |
3722 | else | |
3723 | schedstat_inc(sd, alb_failed); | |
3724 | } | |
dce840a0 | 3725 | rcu_read_unlock(); |
1e3c88bd | 3726 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
3727 | out_unlock: |
3728 | busiest_rq->active_balance = 0; | |
3729 | raw_spin_unlock_irq(&busiest_rq->lock); | |
3730 | return 0; | |
1e3c88bd PZ |
3731 | } |
3732 | ||
3733 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
3734 | |
3735 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
3736 | ||
3737 | static void trigger_sched_softirq(void *data) | |
3738 | { | |
3739 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
3740 | } | |
3741 | ||
3742 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
3743 | { | |
3744 | csd->func = trigger_sched_softirq; | |
3745 | csd->info = NULL; | |
3746 | csd->flags = 0; | |
3747 | csd->priv = 0; | |
3748 | } | |
3749 | ||
3750 | /* | |
3751 | * idle load balancing details | |
3752 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
3753 | * entering idle. | |
3754 | * - This idle load balancer CPU will also go into tickless mode when | |
3755 | * it is idle, just like all other idle CPUs | |
3756 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
3757 | * needed, they will kick the idle load balancer, which then does idle | |
3758 | * load balancing for all the idle CPUs. | |
3759 | */ | |
1e3c88bd PZ |
3760 | static struct { |
3761 | atomic_t load_balancer; | |
83cd4fe2 VP |
3762 | atomic_t first_pick_cpu; |
3763 | atomic_t second_pick_cpu; | |
3764 | cpumask_var_t idle_cpus_mask; | |
3765 | cpumask_var_t grp_idle_mask; | |
3766 | unsigned long next_balance; /* in jiffy units */ | |
3767 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
3768 | |
3769 | int get_nohz_load_balancer(void) | |
3770 | { | |
3771 | return atomic_read(&nohz.load_balancer); | |
3772 | } | |
3773 | ||
3774 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3775 | /** | |
3776 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
3777 | * @cpu: The cpu whose lowest level of sched domain is to | |
3778 | * be returned. | |
3779 | * @flag: The flag to check for the lowest sched_domain | |
3780 | * for the given cpu. | |
3781 | * | |
3782 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
3783 | */ | |
3784 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
3785 | { | |
3786 | struct sched_domain *sd; | |
3787 | ||
3788 | for_each_domain(cpu, sd) | |
08354716 | 3789 | if (sd->flags & flag) |
1e3c88bd PZ |
3790 | break; |
3791 | ||
3792 | return sd; | |
3793 | } | |
3794 | ||
3795 | /** | |
3796 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
3797 | * @cpu: The cpu whose domains we're iterating over. | |
3798 | * @sd: variable holding the value of the power_savings_sd | |
3799 | * for cpu. | |
3800 | * @flag: The flag to filter the sched_domains to be iterated. | |
3801 | * | |
3802 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
3803 | * set, starting from the lowest sched_domain to the highest. | |
3804 | */ | |
3805 | #define for_each_flag_domain(cpu, sd, flag) \ | |
3806 | for (sd = lowest_flag_domain(cpu, flag); \ | |
3807 | (sd && (sd->flags & flag)); sd = sd->parent) | |
3808 | ||
3809 | /** | |
3810 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
3811 | * @ilb_group: group to be checked for semi-idleness | |
3812 | * | |
3813 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
3814 | * | |
3815 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
3816 | * and atleast one non-idle CPU. This helper function checks if the given | |
3817 | * sched_group is semi-idle or not. | |
3818 | */ | |
3819 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
3820 | { | |
83cd4fe2 | 3821 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
3822 | sched_group_cpus(ilb_group)); |
3823 | ||
3824 | /* | |
3825 | * A sched_group is semi-idle when it has atleast one busy cpu | |
3826 | * and atleast one idle cpu. | |
3827 | */ | |
83cd4fe2 | 3828 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
3829 | return 0; |
3830 | ||
83cd4fe2 | 3831 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
3832 | return 0; |
3833 | ||
3834 | return 1; | |
3835 | } | |
3836 | /** | |
3837 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
3838 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
3839 | * | |
3840 | * Returns: Returns the id of the idle load balancer if it exists, | |
3841 | * Else, returns >= nr_cpu_ids. | |
3842 | * | |
3843 | * This algorithm picks the idle load balancer such that it belongs to a | |
3844 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
3845 | * completely idle packages/cores just for the purpose of idle load balancing | |
3846 | * when there are other idle cpu's which are better suited for that job. | |
3847 | */ | |
3848 | static int find_new_ilb(int cpu) | |
3849 | { | |
3850 | struct sched_domain *sd; | |
3851 | struct sched_group *ilb_group; | |
dce840a0 | 3852 | int ilb = nr_cpu_ids; |
1e3c88bd PZ |
3853 | |
3854 | /* | |
3855 | * Have idle load balancer selection from semi-idle packages only | |
3856 | * when power-aware load balancing is enabled | |
3857 | */ | |
3858 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
3859 | goto out_done; | |
3860 | ||
3861 | /* | |
3862 | * Optimize for the case when we have no idle CPUs or only one | |
3863 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
3864 | */ | |
83cd4fe2 | 3865 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
3866 | goto out_done; |
3867 | ||
dce840a0 | 3868 | rcu_read_lock(); |
1e3c88bd PZ |
3869 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { |
3870 | ilb_group = sd->groups; | |
3871 | ||
3872 | do { | |
dce840a0 PZ |
3873 | if (is_semi_idle_group(ilb_group)) { |
3874 | ilb = cpumask_first(nohz.grp_idle_mask); | |
3875 | goto unlock; | |
3876 | } | |
1e3c88bd PZ |
3877 | |
3878 | ilb_group = ilb_group->next; | |
3879 | ||
3880 | } while (ilb_group != sd->groups); | |
3881 | } | |
dce840a0 PZ |
3882 | unlock: |
3883 | rcu_read_unlock(); | |
1e3c88bd PZ |
3884 | |
3885 | out_done: | |
dce840a0 | 3886 | return ilb; |
1e3c88bd PZ |
3887 | } |
3888 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
3889 | static inline int find_new_ilb(int call_cpu) | |
3890 | { | |
83cd4fe2 | 3891 | return nr_cpu_ids; |
1e3c88bd PZ |
3892 | } |
3893 | #endif | |
3894 | ||
83cd4fe2 VP |
3895 | /* |
3896 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
3897 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
3898 | * CPU (if there is one). | |
3899 | */ | |
3900 | static void nohz_balancer_kick(int cpu) | |
3901 | { | |
3902 | int ilb_cpu; | |
3903 | ||
3904 | nohz.next_balance++; | |
3905 | ||
3906 | ilb_cpu = get_nohz_load_balancer(); | |
3907 | ||
3908 | if (ilb_cpu >= nr_cpu_ids) { | |
3909 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
3910 | if (ilb_cpu >= nr_cpu_ids) | |
3911 | return; | |
3912 | } | |
3913 | ||
3914 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
3915 | struct call_single_data *cp; | |
3916 | ||
3917 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
3918 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
3919 | __smp_call_function_single(ilb_cpu, cp, 0); | |
3920 | } | |
3921 | return; | |
3922 | } | |
3923 | ||
1e3c88bd PZ |
3924 | /* |
3925 | * This routine will try to nominate the ilb (idle load balancing) | |
3926 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 3927 | * load balancing on behalf of all those cpus. |
1e3c88bd | 3928 | * |
83cd4fe2 VP |
3929 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
3930 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
3931 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 3932 | * |
83cd4fe2 VP |
3933 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
3934 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
3935 | * behalf of all idle CPUs). | |
1e3c88bd | 3936 | */ |
83cd4fe2 | 3937 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
3938 | { |
3939 | int cpu = smp_processor_id(); | |
3940 | ||
3941 | if (stop_tick) { | |
1e3c88bd PZ |
3942 | if (!cpu_active(cpu)) { |
3943 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 3944 | return; |
1e3c88bd PZ |
3945 | |
3946 | /* | |
3947 | * If we are going offline and still the leader, | |
3948 | * give up! | |
3949 | */ | |
83cd4fe2 VP |
3950 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3951 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3952 | BUG(); |
3953 | ||
83cd4fe2 | 3954 | return; |
1e3c88bd PZ |
3955 | } |
3956 | ||
83cd4fe2 | 3957 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 3958 | |
83cd4fe2 VP |
3959 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
3960 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
3961 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
3962 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 3963 | |
83cd4fe2 | 3964 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
3965 | int new_ilb; |
3966 | ||
83cd4fe2 VP |
3967 | /* make me the ilb owner */ |
3968 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
3969 | cpu) != nr_cpu_ids) | |
3970 | return; | |
3971 | ||
1e3c88bd PZ |
3972 | /* |
3973 | * Check to see if there is a more power-efficient | |
3974 | * ilb. | |
3975 | */ | |
3976 | new_ilb = find_new_ilb(cpu); | |
3977 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 3978 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 3979 | resched_cpu(new_ilb); |
83cd4fe2 | 3980 | return; |
1e3c88bd | 3981 | } |
83cd4fe2 | 3982 | return; |
1e3c88bd PZ |
3983 | } |
3984 | } else { | |
83cd4fe2 VP |
3985 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
3986 | return; | |
1e3c88bd | 3987 | |
83cd4fe2 | 3988 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
3989 | |
3990 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
3991 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
3992 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
3993 | BUG(); |
3994 | } | |
83cd4fe2 | 3995 | return; |
1e3c88bd PZ |
3996 | } |
3997 | #endif | |
3998 | ||
3999 | static DEFINE_SPINLOCK(balancing); | |
4000 | ||
49c022e6 PZ |
4001 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4002 | ||
4003 | /* | |
4004 | * Scale the max load_balance interval with the number of CPUs in the system. | |
4005 | * This trades load-balance latency on larger machines for less cross talk. | |
4006 | */ | |
4007 | static void update_max_interval(void) | |
4008 | { | |
4009 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
4010 | } | |
4011 | ||
1e3c88bd PZ |
4012 | /* |
4013 | * It checks each scheduling domain to see if it is due to be balanced, | |
4014 | * and initiates a balancing operation if so. | |
4015 | * | |
4016 | * Balancing parameters are set up in arch_init_sched_domains. | |
4017 | */ | |
4018 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4019 | { | |
4020 | int balance = 1; | |
4021 | struct rq *rq = cpu_rq(cpu); | |
4022 | unsigned long interval; | |
4023 | struct sched_domain *sd; | |
4024 | /* Earliest time when we have to do rebalance again */ | |
4025 | unsigned long next_balance = jiffies + 60*HZ; | |
4026 | int update_next_balance = 0; | |
4027 | int need_serialize; | |
4028 | ||
2069dd75 PZ |
4029 | update_shares(cpu); |
4030 | ||
dce840a0 | 4031 | rcu_read_lock(); |
1e3c88bd PZ |
4032 | for_each_domain(cpu, sd) { |
4033 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4034 | continue; | |
4035 | ||
4036 | interval = sd->balance_interval; | |
4037 | if (idle != CPU_IDLE) | |
4038 | interval *= sd->busy_factor; | |
4039 | ||
4040 | /* scale ms to jiffies */ | |
4041 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 4042 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
4043 | |
4044 | need_serialize = sd->flags & SD_SERIALIZE; | |
4045 | ||
4046 | if (need_serialize) { | |
4047 | if (!spin_trylock(&balancing)) | |
4048 | goto out; | |
4049 | } | |
4050 | ||
4051 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4052 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
4053 | /* | |
4054 | * We've pulled tasks over so either we're no | |
c186fafe | 4055 | * longer idle. |
1e3c88bd PZ |
4056 | */ |
4057 | idle = CPU_NOT_IDLE; | |
4058 | } | |
4059 | sd->last_balance = jiffies; | |
4060 | } | |
4061 | if (need_serialize) | |
4062 | spin_unlock(&balancing); | |
4063 | out: | |
4064 | if (time_after(next_balance, sd->last_balance + interval)) { | |
4065 | next_balance = sd->last_balance + interval; | |
4066 | update_next_balance = 1; | |
4067 | } | |
4068 | ||
4069 | /* | |
4070 | * Stop the load balance at this level. There is another | |
4071 | * CPU in our sched group which is doing load balancing more | |
4072 | * actively. | |
4073 | */ | |
4074 | if (!balance) | |
4075 | break; | |
4076 | } | |
dce840a0 | 4077 | rcu_read_unlock(); |
1e3c88bd PZ |
4078 | |
4079 | /* | |
4080 | * next_balance will be updated only when there is a need. | |
4081 | * When the cpu is attached to null domain for ex, it will not be | |
4082 | * updated. | |
4083 | */ | |
4084 | if (likely(update_next_balance)) | |
4085 | rq->next_balance = next_balance; | |
4086 | } | |
4087 | ||
83cd4fe2 | 4088 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 4089 | /* |
83cd4fe2 | 4090 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
4091 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
4092 | */ | |
83cd4fe2 VP |
4093 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
4094 | { | |
4095 | struct rq *this_rq = cpu_rq(this_cpu); | |
4096 | struct rq *rq; | |
4097 | int balance_cpu; | |
4098 | ||
4099 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
4100 | return; | |
4101 | ||
4102 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
4103 | if (balance_cpu == this_cpu) | |
4104 | continue; | |
4105 | ||
4106 | /* | |
4107 | * If this cpu gets work to do, stop the load balancing | |
4108 | * work being done for other cpus. Next load | |
4109 | * balancing owner will pick it up. | |
4110 | */ | |
4111 | if (need_resched()) { | |
4112 | this_rq->nohz_balance_kick = 0; | |
4113 | break; | |
4114 | } | |
4115 | ||
4116 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 4117 | update_rq_clock(this_rq); |
83cd4fe2 VP |
4118 | update_cpu_load(this_rq); |
4119 | raw_spin_unlock_irq(&this_rq->lock); | |
4120 | ||
4121 | rebalance_domains(balance_cpu, CPU_IDLE); | |
4122 | ||
4123 | rq = cpu_rq(balance_cpu); | |
4124 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
4125 | this_rq->next_balance = rq->next_balance; | |
4126 | } | |
4127 | nohz.next_balance = this_rq->next_balance; | |
4128 | this_rq->nohz_balance_kick = 0; | |
4129 | } | |
4130 | ||
4131 | /* | |
4132 | * Current heuristic for kicking the idle load balancer | |
4133 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
4134 | * idle load balancer when it has more than one process active. This | |
4135 | * eliminates the need for idle load balancing altogether when we have | |
4136 | * only one running process in the system (common case). | |
4137 | * - If there are more than one busy CPU, idle load balancer may have | |
4138 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
4139 | * SMT or core siblings and can run better if they move to different | |
4140 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
4141 | * which will kick idle load balancer as soon as it has any load. | |
4142 | */ | |
4143 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
4144 | { | |
4145 | unsigned long now = jiffies; | |
4146 | int ret; | |
4147 | int first_pick_cpu, second_pick_cpu; | |
4148 | ||
4149 | if (time_before(now, nohz.next_balance)) | |
4150 | return 0; | |
4151 | ||
f6c3f168 | 4152 | if (rq->idle_at_tick) |
83cd4fe2 VP |
4153 | return 0; |
4154 | ||
4155 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
4156 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
4157 | ||
4158 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
4159 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
4160 | return 0; | |
4161 | ||
4162 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
4163 | if (ret == nr_cpu_ids || ret == cpu) { | |
4164 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
4165 | if (rq->nr_running > 1) | |
4166 | return 1; | |
4167 | } else { | |
4168 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
4169 | if (ret == nr_cpu_ids || ret == cpu) { | |
4170 | if (rq->nr_running) | |
4171 | return 1; | |
4172 | } | |
4173 | } | |
4174 | return 0; | |
4175 | } | |
4176 | #else | |
4177 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
4178 | #endif | |
4179 | ||
4180 | /* | |
4181 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4182 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
4183 | */ | |
1e3c88bd PZ |
4184 | static void run_rebalance_domains(struct softirq_action *h) |
4185 | { | |
4186 | int this_cpu = smp_processor_id(); | |
4187 | struct rq *this_rq = cpu_rq(this_cpu); | |
4188 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4189 | CPU_IDLE : CPU_NOT_IDLE; | |
4190 | ||
4191 | rebalance_domains(this_cpu, idle); | |
4192 | ||
1e3c88bd | 4193 | /* |
83cd4fe2 | 4194 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
4195 | * balancing on behalf of the other idle cpus whose ticks are |
4196 | * stopped. | |
4197 | */ | |
83cd4fe2 | 4198 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
4199 | } |
4200 | ||
4201 | static inline int on_null_domain(int cpu) | |
4202 | { | |
90a6501f | 4203 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
4204 | } |
4205 | ||
4206 | /* | |
4207 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
4208 | */ |
4209 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
4210 | { | |
1e3c88bd PZ |
4211 | /* Don't need to rebalance while attached to NULL domain */ |
4212 | if (time_after_eq(jiffies, rq->next_balance) && | |
4213 | likely(!on_null_domain(cpu))) | |
4214 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
4215 | #ifdef CONFIG_NO_HZ |
4216 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
4217 | nohz_balancer_kick(cpu); | |
4218 | #endif | |
1e3c88bd PZ |
4219 | } |
4220 | ||
0bcdcf28 CE |
4221 | static void rq_online_fair(struct rq *rq) |
4222 | { | |
4223 | update_sysctl(); | |
4224 | } | |
4225 | ||
4226 | static void rq_offline_fair(struct rq *rq) | |
4227 | { | |
4228 | update_sysctl(); | |
4229 | } | |
4230 | ||
1e3c88bd PZ |
4231 | #else /* CONFIG_SMP */ |
4232 | ||
4233 | /* | |
4234 | * on UP we do not need to balance between CPUs: | |
4235 | */ | |
4236 | static inline void idle_balance(int cpu, struct rq *rq) | |
4237 | { | |
4238 | } | |
4239 | ||
55e12e5e | 4240 | #endif /* CONFIG_SMP */ |
e1d1484f | 4241 | |
bf0f6f24 IM |
4242 | /* |
4243 | * scheduler tick hitting a task of our scheduling class: | |
4244 | */ | |
8f4d37ec | 4245 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
4246 | { |
4247 | struct cfs_rq *cfs_rq; | |
4248 | struct sched_entity *se = &curr->se; | |
4249 | ||
4250 | for_each_sched_entity(se) { | |
4251 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 4252 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
4253 | } |
4254 | } | |
4255 | ||
4256 | /* | |
cd29fe6f PZ |
4257 | * called on fork with the child task as argument from the parent's context |
4258 | * - child not yet on the tasklist | |
4259 | * - preemption disabled | |
bf0f6f24 | 4260 | */ |
cd29fe6f | 4261 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 4262 | { |
cd29fe6f | 4263 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 4264 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 4265 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
4266 | struct rq *rq = this_rq(); |
4267 | unsigned long flags; | |
4268 | ||
05fa785c | 4269 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 4270 | |
861d034e PZ |
4271 | update_rq_clock(rq); |
4272 | ||
b0a0f667 PM |
4273 | if (unlikely(task_cpu(p) != this_cpu)) { |
4274 | rcu_read_lock(); | |
cd29fe6f | 4275 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
4276 | rcu_read_unlock(); |
4277 | } | |
bf0f6f24 | 4278 | |
7109c442 | 4279 | update_curr(cfs_rq); |
cd29fe6f | 4280 | |
b5d9d734 MG |
4281 | if (curr) |
4282 | se->vruntime = curr->vruntime; | |
aeb73b04 | 4283 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 4284 | |
cd29fe6f | 4285 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 4286 | /* |
edcb60a3 IM |
4287 | * Upon rescheduling, sched_class::put_prev_task() will place |
4288 | * 'current' within the tree based on its new key value. | |
4289 | */ | |
4d78e7b6 | 4290 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 4291 | resched_task(rq->curr); |
4d78e7b6 | 4292 | } |
bf0f6f24 | 4293 | |
88ec22d3 PZ |
4294 | se->vruntime -= cfs_rq->min_vruntime; |
4295 | ||
05fa785c | 4296 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
4297 | } |
4298 | ||
cb469845 SR |
4299 | /* |
4300 | * Priority of the task has changed. Check to see if we preempt | |
4301 | * the current task. | |
4302 | */ | |
da7a735e PZ |
4303 | static void |
4304 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 4305 | { |
da7a735e PZ |
4306 | if (!p->se.on_rq) |
4307 | return; | |
4308 | ||
cb469845 SR |
4309 | /* |
4310 | * Reschedule if we are currently running on this runqueue and | |
4311 | * our priority decreased, or if we are not currently running on | |
4312 | * this runqueue and our priority is higher than the current's | |
4313 | */ | |
da7a735e | 4314 | if (rq->curr == p) { |
cb469845 SR |
4315 | if (p->prio > oldprio) |
4316 | resched_task(rq->curr); | |
4317 | } else | |
15afe09b | 4318 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4319 | } |
4320 | ||
da7a735e PZ |
4321 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
4322 | { | |
4323 | struct sched_entity *se = &p->se; | |
4324 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4325 | ||
4326 | /* | |
4327 | * Ensure the task's vruntime is normalized, so that when its | |
4328 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
4329 | * do the right thing. | |
4330 | * | |
4331 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
4332 | * have normalized the vruntime, if it was !on_rq, then only when | |
4333 | * the task is sleeping will it still have non-normalized vruntime. | |
4334 | */ | |
4335 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
4336 | /* | |
4337 | * Fix up our vruntime so that the current sleep doesn't | |
4338 | * cause 'unlimited' sleep bonus. | |
4339 | */ | |
4340 | place_entity(cfs_rq, se, 0); | |
4341 | se->vruntime -= cfs_rq->min_vruntime; | |
4342 | } | |
4343 | } | |
4344 | ||
cb469845 SR |
4345 | /* |
4346 | * We switched to the sched_fair class. | |
4347 | */ | |
da7a735e | 4348 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 4349 | { |
da7a735e PZ |
4350 | if (!p->se.on_rq) |
4351 | return; | |
4352 | ||
cb469845 SR |
4353 | /* |
4354 | * We were most likely switched from sched_rt, so | |
4355 | * kick off the schedule if running, otherwise just see | |
4356 | * if we can still preempt the current task. | |
4357 | */ | |
da7a735e | 4358 | if (rq->curr == p) |
cb469845 SR |
4359 | resched_task(rq->curr); |
4360 | else | |
15afe09b | 4361 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4362 | } |
4363 | ||
83b699ed SV |
4364 | /* Account for a task changing its policy or group. |
4365 | * | |
4366 | * This routine is mostly called to set cfs_rq->curr field when a task | |
4367 | * migrates between groups/classes. | |
4368 | */ | |
4369 | static void set_curr_task_fair(struct rq *rq) | |
4370 | { | |
4371 | struct sched_entity *se = &rq->curr->se; | |
4372 | ||
ec12cb7f PT |
4373 | for_each_sched_entity(se) { |
4374 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4375 | ||
4376 | set_next_entity(cfs_rq, se); | |
4377 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
4378 | account_cfs_rq_runtime(cfs_rq, 0); | |
4379 | } | |
83b699ed SV |
4380 | } |
4381 | ||
810b3817 | 4382 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4383 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 4384 | { |
b2b5ce02 PZ |
4385 | /* |
4386 | * If the task was not on the rq at the time of this cgroup movement | |
4387 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
4388 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
4389 | * bonus in place_entity()). | |
4390 | * | |
4391 | * If it was on the rq, we've just 'preempted' it, which does convert | |
4392 | * ->vruntime to a relative base. | |
4393 | * | |
4394 | * Make sure both cases convert their relative position when migrating | |
4395 | * to another cgroup's rq. This does somewhat interfere with the | |
4396 | * fair sleeper stuff for the first placement, but who cares. | |
4397 | */ | |
4398 | if (!on_rq) | |
4399 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
4400 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 4401 | if (!on_rq) |
b2b5ce02 | 4402 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
4403 | } |
4404 | #endif | |
4405 | ||
6d686f45 | 4406 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
4407 | { |
4408 | struct sched_entity *se = &task->se; | |
0d721cea PW |
4409 | unsigned int rr_interval = 0; |
4410 | ||
4411 | /* | |
4412 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
4413 | * idle runqueue: | |
4414 | */ | |
0d721cea PW |
4415 | if (rq->cfs.load.weight) |
4416 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
4417 | |
4418 | return rr_interval; | |
4419 | } | |
4420 | ||
bf0f6f24 IM |
4421 | /* |
4422 | * All the scheduling class methods: | |
4423 | */ | |
5522d5d5 IM |
4424 | static const struct sched_class fair_sched_class = { |
4425 | .next = &idle_sched_class, | |
bf0f6f24 IM |
4426 | .enqueue_task = enqueue_task_fair, |
4427 | .dequeue_task = dequeue_task_fair, | |
4428 | .yield_task = yield_task_fair, | |
d95f4122 | 4429 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 4430 | |
2e09bf55 | 4431 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
4432 | |
4433 | .pick_next_task = pick_next_task_fair, | |
4434 | .put_prev_task = put_prev_task_fair, | |
4435 | ||
681f3e68 | 4436 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
4437 | .select_task_rq = select_task_rq_fair, |
4438 | ||
0bcdcf28 CE |
4439 | .rq_online = rq_online_fair, |
4440 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
4441 | |
4442 | .task_waking = task_waking_fair, | |
681f3e68 | 4443 | #endif |
bf0f6f24 | 4444 | |
83b699ed | 4445 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 4446 | .task_tick = task_tick_fair, |
cd29fe6f | 4447 | .task_fork = task_fork_fair, |
cb469845 SR |
4448 | |
4449 | .prio_changed = prio_changed_fair, | |
da7a735e | 4450 | .switched_from = switched_from_fair, |
cb469845 | 4451 | .switched_to = switched_to_fair, |
810b3817 | 4452 | |
0d721cea PW |
4453 | .get_rr_interval = get_rr_interval_fair, |
4454 | ||
810b3817 | 4455 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4456 | .task_move_group = task_move_group_fair, |
810b3817 | 4457 | #endif |
bf0f6f24 IM |
4458 | }; |
4459 | ||
4460 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 4461 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 4462 | { |
bf0f6f24 IM |
4463 | struct cfs_rq *cfs_rq; |
4464 | ||
5973e5b9 | 4465 | rcu_read_lock(); |
c3b64f1e | 4466 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 4467 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 4468 | rcu_read_unlock(); |
bf0f6f24 IM |
4469 | } |
4470 | #endif |