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 | 708 | #ifdef CONFIG_FAIR_GROUP_SCHED |
64660c86 PT |
709 | /* we need this in update_cfs_load and load-balance functions below */ |
710 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); | |
3ff6dcac | 711 | # ifdef CONFIG_SMP |
d6b55918 PT |
712 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
713 | int global_update) | |
714 | { | |
715 | struct task_group *tg = cfs_rq->tg; | |
716 | long load_avg; | |
717 | ||
718 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
719 | load_avg -= cfs_rq->load_contribution; | |
720 | ||
721 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
722 | atomic_add(load_avg, &tg->load_weight); | |
723 | cfs_rq->load_contribution += load_avg; | |
724 | } | |
725 | } | |
726 | ||
727 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 728 | { |
a7a4f8a7 | 729 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 730 | u64 now, delta; |
e33078ba | 731 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 732 | |
64660c86 | 733 | if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq)) |
2069dd75 PZ |
734 | return; |
735 | ||
05ca62c6 | 736 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
737 | delta = now - cfs_rq->load_stamp; |
738 | ||
e33078ba PT |
739 | /* truncate load history at 4 idle periods */ |
740 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
741 | now - cfs_rq->load_last > 4 * period) { | |
742 | cfs_rq->load_period = 0; | |
743 | cfs_rq->load_avg = 0; | |
f07333bf | 744 | delta = period - 1; |
e33078ba PT |
745 | } |
746 | ||
2069dd75 | 747 | cfs_rq->load_stamp = now; |
3b3d190e | 748 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 749 | cfs_rq->load_period += delta; |
e33078ba PT |
750 | if (load) { |
751 | cfs_rq->load_last = now; | |
752 | cfs_rq->load_avg += delta * load; | |
753 | } | |
2069dd75 | 754 | |
d6b55918 PT |
755 | /* consider updating load contribution on each fold or truncate */ |
756 | if (global_update || cfs_rq->load_period > period | |
757 | || !cfs_rq->load_period) | |
758 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
759 | ||
2069dd75 PZ |
760 | while (cfs_rq->load_period > period) { |
761 | /* | |
762 | * Inline assembly required to prevent the compiler | |
763 | * optimising this loop into a divmod call. | |
764 | * See __iter_div_u64_rem() for another example of this. | |
765 | */ | |
766 | asm("" : "+rm" (cfs_rq->load_period)); | |
767 | cfs_rq->load_period /= 2; | |
768 | cfs_rq->load_avg /= 2; | |
769 | } | |
3d4b47b4 | 770 | |
e33078ba PT |
771 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
772 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
773 | } |
774 | ||
6d5ab293 | 775 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
776 | { |
777 | long load_weight, load, shares; | |
778 | ||
6d5ab293 | 779 | load = cfs_rq->load.weight; |
3ff6dcac YZ |
780 | |
781 | load_weight = atomic_read(&tg->load_weight); | |
3ff6dcac | 782 | load_weight += load; |
6d5ab293 | 783 | load_weight -= cfs_rq->load_contribution; |
3ff6dcac YZ |
784 | |
785 | shares = (tg->shares * load); | |
786 | if (load_weight) | |
787 | shares /= load_weight; | |
788 | ||
789 | if (shares < MIN_SHARES) | |
790 | shares = MIN_SHARES; | |
791 | if (shares > tg->shares) | |
792 | shares = tg->shares; | |
793 | ||
794 | return shares; | |
795 | } | |
796 | ||
797 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
798 | { | |
799 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
800 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 801 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
802 | } |
803 | } | |
804 | # else /* CONFIG_SMP */ | |
805 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
806 | { | |
807 | } | |
808 | ||
6d5ab293 | 809 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
810 | { |
811 | return tg->shares; | |
812 | } | |
813 | ||
814 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
815 | { | |
816 | } | |
817 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
818 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
819 | unsigned long weight) | |
820 | { | |
19e5eebb PT |
821 | if (se->on_rq) { |
822 | /* commit outstanding execution time */ | |
823 | if (cfs_rq->curr == se) | |
824 | update_curr(cfs_rq); | |
2069dd75 | 825 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 826 | } |
2069dd75 PZ |
827 | |
828 | update_load_set(&se->load, weight); | |
829 | ||
830 | if (se->on_rq) | |
831 | account_entity_enqueue(cfs_rq, se); | |
832 | } | |
833 | ||
6d5ab293 | 834 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
835 | { |
836 | struct task_group *tg; | |
837 | struct sched_entity *se; | |
3ff6dcac | 838 | long shares; |
2069dd75 | 839 | |
2069dd75 PZ |
840 | tg = cfs_rq->tg; |
841 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 842 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 843 | return; |
3ff6dcac YZ |
844 | #ifndef CONFIG_SMP |
845 | if (likely(se->load.weight == tg->shares)) | |
846 | return; | |
847 | #endif | |
6d5ab293 | 848 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
849 | |
850 | reweight_entity(cfs_rq_of(se), se, shares); | |
851 | } | |
852 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 853 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
854 | { |
855 | } | |
856 | ||
6d5ab293 | 857 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
858 | { |
859 | } | |
43365bd7 PT |
860 | |
861 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
862 | { | |
863 | } | |
2069dd75 PZ |
864 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
865 | ||
2396af69 | 866 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 867 | { |
bf0f6f24 | 868 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
869 | struct task_struct *tsk = NULL; |
870 | ||
871 | if (entity_is_task(se)) | |
872 | tsk = task_of(se); | |
873 | ||
41acab88 LDM |
874 | if (se->statistics.sleep_start) { |
875 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
876 | |
877 | if ((s64)delta < 0) | |
878 | delta = 0; | |
879 | ||
41acab88 LDM |
880 | if (unlikely(delta > se->statistics.sleep_max)) |
881 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 882 | |
41acab88 LDM |
883 | se->statistics.sleep_start = 0; |
884 | se->statistics.sum_sleep_runtime += delta; | |
9745512c | 885 | |
768d0c27 | 886 | if (tsk) { |
e414314c | 887 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
888 | trace_sched_stat_sleep(tsk, delta); |
889 | } | |
bf0f6f24 | 890 | } |
41acab88 LDM |
891 | if (se->statistics.block_start) { |
892 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
893 | |
894 | if ((s64)delta < 0) | |
895 | delta = 0; | |
896 | ||
41acab88 LDM |
897 | if (unlikely(delta > se->statistics.block_max)) |
898 | se->statistics.block_max = delta; | |
bf0f6f24 | 899 | |
41acab88 LDM |
900 | se->statistics.block_start = 0; |
901 | se->statistics.sum_sleep_runtime += delta; | |
30084fbd | 902 | |
e414314c | 903 | if (tsk) { |
8f0dfc34 | 904 | if (tsk->in_iowait) { |
41acab88 LDM |
905 | se->statistics.iowait_sum += delta; |
906 | se->statistics.iowait_count++; | |
768d0c27 | 907 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
908 | } |
909 | ||
e414314c PZ |
910 | /* |
911 | * Blocking time is in units of nanosecs, so shift by | |
912 | * 20 to get a milliseconds-range estimation of the | |
913 | * amount of time that the task spent sleeping: | |
914 | */ | |
915 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
916 | profile_hits(SLEEP_PROFILING, | |
917 | (void *)get_wchan(tsk), | |
918 | delta >> 20); | |
919 | } | |
920 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 921 | } |
bf0f6f24 IM |
922 | } |
923 | #endif | |
924 | } | |
925 | ||
ddc97297 PZ |
926 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
927 | { | |
928 | #ifdef CONFIG_SCHED_DEBUG | |
929 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
930 | ||
931 | if (d < 0) | |
932 | d = -d; | |
933 | ||
934 | if (d > 3*sysctl_sched_latency) | |
935 | schedstat_inc(cfs_rq, nr_spread_over); | |
936 | #endif | |
937 | } | |
938 | ||
aeb73b04 PZ |
939 | static void |
940 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
941 | { | |
1af5f730 | 942 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 943 | |
2cb8600e PZ |
944 | /* |
945 | * The 'current' period is already promised to the current tasks, | |
946 | * however the extra weight of the new task will slow them down a | |
947 | * little, place the new task so that it fits in the slot that | |
948 | * stays open at the end. | |
949 | */ | |
94dfb5e7 | 950 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 951 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 952 | |
a2e7a7eb | 953 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 954 | if (!initial) { |
a2e7a7eb | 955 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 956 | |
a2e7a7eb MG |
957 | /* |
958 | * Halve their sleep time's effect, to allow | |
959 | * for a gentler effect of sleepers: | |
960 | */ | |
961 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
962 | thresh >>= 1; | |
51e0304c | 963 | |
a2e7a7eb | 964 | vruntime -= thresh; |
aeb73b04 PZ |
965 | } |
966 | ||
b5d9d734 MG |
967 | /* ensure we never gain time by being placed backwards. */ |
968 | vruntime = max_vruntime(se->vruntime, vruntime); | |
969 | ||
67e9fb2a | 970 | se->vruntime = vruntime; |
aeb73b04 PZ |
971 | } |
972 | ||
d3d9dc33 PT |
973 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
974 | ||
bf0f6f24 | 975 | static void |
88ec22d3 | 976 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 977 | { |
88ec22d3 PZ |
978 | /* |
979 | * Update the normalized vruntime before updating min_vruntime | |
980 | * through callig update_curr(). | |
981 | */ | |
371fd7e7 | 982 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
983 | se->vruntime += cfs_rq->min_vruntime; |
984 | ||
bf0f6f24 | 985 | /* |
a2a2d680 | 986 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 987 | */ |
b7cc0896 | 988 | update_curr(cfs_rq); |
d6b55918 | 989 | update_cfs_load(cfs_rq, 0); |
a992241d | 990 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 991 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 992 | |
88ec22d3 | 993 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 994 | place_entity(cfs_rq, se, 0); |
2396af69 | 995 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 996 | } |
bf0f6f24 | 997 | |
d2417e5a | 998 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 999 | check_spread(cfs_rq, se); |
83b699ed SV |
1000 | if (se != cfs_rq->curr) |
1001 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1002 | se->on_rq = 1; |
3d4b47b4 | 1003 | |
d3d9dc33 | 1004 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1005 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1006 | check_enqueue_throttle(cfs_rq); |
1007 | } | |
bf0f6f24 IM |
1008 | } |
1009 | ||
2c13c919 | 1010 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1011 | { |
2c13c919 RR |
1012 | for_each_sched_entity(se) { |
1013 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1014 | if (cfs_rq->last == se) | |
1015 | cfs_rq->last = NULL; | |
1016 | else | |
1017 | break; | |
1018 | } | |
1019 | } | |
2002c695 | 1020 | |
2c13c919 RR |
1021 | static void __clear_buddies_next(struct sched_entity *se) |
1022 | { | |
1023 | for_each_sched_entity(se) { | |
1024 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1025 | if (cfs_rq->next == se) | |
1026 | cfs_rq->next = NULL; | |
1027 | else | |
1028 | break; | |
1029 | } | |
2002c695 PZ |
1030 | } |
1031 | ||
ac53db59 RR |
1032 | static void __clear_buddies_skip(struct sched_entity *se) |
1033 | { | |
1034 | for_each_sched_entity(se) { | |
1035 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1036 | if (cfs_rq->skip == se) | |
1037 | cfs_rq->skip = NULL; | |
1038 | else | |
1039 | break; | |
1040 | } | |
1041 | } | |
1042 | ||
a571bbea PZ |
1043 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1044 | { | |
2c13c919 RR |
1045 | if (cfs_rq->last == se) |
1046 | __clear_buddies_last(se); | |
1047 | ||
1048 | if (cfs_rq->next == se) | |
1049 | __clear_buddies_next(se); | |
ac53db59 RR |
1050 | |
1051 | if (cfs_rq->skip == se) | |
1052 | __clear_buddies_skip(se); | |
a571bbea PZ |
1053 | } |
1054 | ||
bf0f6f24 | 1055 | static void |
371fd7e7 | 1056 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1057 | { |
a2a2d680 DA |
1058 | /* |
1059 | * Update run-time statistics of the 'current'. | |
1060 | */ | |
1061 | update_curr(cfs_rq); | |
1062 | ||
19b6a2e3 | 1063 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1064 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1065 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1066 | if (entity_is_task(se)) { |
1067 | struct task_struct *tsk = task_of(se); | |
1068 | ||
1069 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1070 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1071 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1072 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1073 | } |
db36cc7d | 1074 | #endif |
67e9fb2a PZ |
1075 | } |
1076 | ||
2002c695 | 1077 | clear_buddies(cfs_rq, se); |
4793241b | 1078 | |
83b699ed | 1079 | if (se != cfs_rq->curr) |
30cfdcfc | 1080 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1081 | se->on_rq = 0; |
d6b55918 | 1082 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1083 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1084 | |
1085 | /* | |
1086 | * Normalize the entity after updating the min_vruntime because the | |
1087 | * update can refer to the ->curr item and we need to reflect this | |
1088 | * movement in our normalized position. | |
1089 | */ | |
371fd7e7 | 1090 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1091 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 PZ |
1092 | |
1093 | update_min_vruntime(cfs_rq); | |
1094 | update_cfs_shares(cfs_rq); | |
bf0f6f24 IM |
1095 | } |
1096 | ||
1097 | /* | |
1098 | * Preempt the current task with a newly woken task if needed: | |
1099 | */ | |
7c92e54f | 1100 | static void |
2e09bf55 | 1101 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1102 | { |
11697830 PZ |
1103 | unsigned long ideal_runtime, delta_exec; |
1104 | ||
6d0f0ebd | 1105 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1106 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1107 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1108 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1109 | /* |
1110 | * The current task ran long enough, ensure it doesn't get | |
1111 | * re-elected due to buddy favours. | |
1112 | */ | |
1113 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1114 | return; |
1115 | } | |
1116 | ||
1117 | /* | |
1118 | * Ensure that a task that missed wakeup preemption by a | |
1119 | * narrow margin doesn't have to wait for a full slice. | |
1120 | * This also mitigates buddy induced latencies under load. | |
1121 | */ | |
f685ceac MG |
1122 | if (delta_exec < sysctl_sched_min_granularity) |
1123 | return; | |
1124 | ||
1125 | if (cfs_rq->nr_running > 1) { | |
ac53db59 | 1126 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac MG |
1127 | s64 delta = curr->vruntime - se->vruntime; |
1128 | ||
d7d82944 MG |
1129 | if (delta < 0) |
1130 | return; | |
1131 | ||
f685ceac MG |
1132 | if (delta > ideal_runtime) |
1133 | resched_task(rq_of(cfs_rq)->curr); | |
a9f3e2b5 | 1134 | } |
bf0f6f24 IM |
1135 | } |
1136 | ||
83b699ed | 1137 | static void |
8494f412 | 1138 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1139 | { |
83b699ed SV |
1140 | /* 'current' is not kept within the tree. */ |
1141 | if (se->on_rq) { | |
1142 | /* | |
1143 | * Any task has to be enqueued before it get to execute on | |
1144 | * a CPU. So account for the time it spent waiting on the | |
1145 | * runqueue. | |
1146 | */ | |
1147 | update_stats_wait_end(cfs_rq, se); | |
1148 | __dequeue_entity(cfs_rq, se); | |
1149 | } | |
1150 | ||
79303e9e | 1151 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1152 | cfs_rq->curr = se; |
eba1ed4b IM |
1153 | #ifdef CONFIG_SCHEDSTATS |
1154 | /* | |
1155 | * Track our maximum slice length, if the CPU's load is at | |
1156 | * least twice that of our own weight (i.e. dont track it | |
1157 | * when there are only lesser-weight tasks around): | |
1158 | */ | |
495eca49 | 1159 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1160 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1161 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1162 | } | |
1163 | #endif | |
4a55b450 | 1164 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1165 | } |
1166 | ||
3f3a4904 PZ |
1167 | static int |
1168 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1169 | ||
ac53db59 RR |
1170 | /* |
1171 | * Pick the next process, keeping these things in mind, in this order: | |
1172 | * 1) keep things fair between processes/task groups | |
1173 | * 2) pick the "next" process, since someone really wants that to run | |
1174 | * 3) pick the "last" process, for cache locality | |
1175 | * 4) do not run the "skip" process, if something else is available | |
1176 | */ | |
f4b6755f | 1177 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1178 | { |
ac53db59 | 1179 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1180 | struct sched_entity *left = se; |
f4b6755f | 1181 | |
ac53db59 RR |
1182 | /* |
1183 | * Avoid running the skip buddy, if running something else can | |
1184 | * be done without getting too unfair. | |
1185 | */ | |
1186 | if (cfs_rq->skip == se) { | |
1187 | struct sched_entity *second = __pick_next_entity(se); | |
1188 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1189 | se = second; | |
1190 | } | |
aa2ac252 | 1191 | |
f685ceac MG |
1192 | /* |
1193 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1194 | */ | |
1195 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1196 | se = cfs_rq->last; | |
1197 | ||
ac53db59 RR |
1198 | /* |
1199 | * Someone really wants this to run. If it's not unfair, run it. | |
1200 | */ | |
1201 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1202 | se = cfs_rq->next; | |
1203 | ||
f685ceac | 1204 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1205 | |
1206 | return se; | |
aa2ac252 PZ |
1207 | } |
1208 | ||
d3d9dc33 PT |
1209 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
1210 | ||
ab6cde26 | 1211 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1212 | { |
1213 | /* | |
1214 | * If still on the runqueue then deactivate_task() | |
1215 | * was not called and update_curr() has to be done: | |
1216 | */ | |
1217 | if (prev->on_rq) | |
b7cc0896 | 1218 | update_curr(cfs_rq); |
bf0f6f24 | 1219 | |
d3d9dc33 PT |
1220 | /* throttle cfs_rqs exceeding runtime */ |
1221 | check_cfs_rq_runtime(cfs_rq); | |
1222 | ||
ddc97297 | 1223 | check_spread(cfs_rq, prev); |
30cfdcfc | 1224 | if (prev->on_rq) { |
5870db5b | 1225 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1226 | /* Put 'current' back into the tree. */ |
1227 | __enqueue_entity(cfs_rq, prev); | |
1228 | } | |
429d43bc | 1229 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1230 | } |
1231 | ||
8f4d37ec PZ |
1232 | static void |
1233 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1234 | { |
bf0f6f24 | 1235 | /* |
30cfdcfc | 1236 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1237 | */ |
30cfdcfc | 1238 | update_curr(cfs_rq); |
bf0f6f24 | 1239 | |
43365bd7 PT |
1240 | /* |
1241 | * Update share accounting for long-running entities. | |
1242 | */ | |
1243 | update_entity_shares_tick(cfs_rq); | |
1244 | ||
8f4d37ec PZ |
1245 | #ifdef CONFIG_SCHED_HRTICK |
1246 | /* | |
1247 | * queued ticks are scheduled to match the slice, so don't bother | |
1248 | * validating it and just reschedule. | |
1249 | */ | |
983ed7a6 HH |
1250 | if (queued) { |
1251 | resched_task(rq_of(cfs_rq)->curr); | |
1252 | return; | |
1253 | } | |
8f4d37ec PZ |
1254 | /* |
1255 | * don't let the period tick interfere with the hrtick preemption | |
1256 | */ | |
1257 | if (!sched_feat(DOUBLE_TICK) && | |
1258 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1259 | return; | |
1260 | #endif | |
1261 | ||
2c2efaed | 1262 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1263 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1264 | } |
1265 | ||
ab84d31e PT |
1266 | |
1267 | /************************************************** | |
1268 | * CFS bandwidth control machinery | |
1269 | */ | |
1270 | ||
1271 | #ifdef CONFIG_CFS_BANDWIDTH | |
1272 | /* | |
1273 | * default period for cfs group bandwidth. | |
1274 | * default: 0.1s, units: nanoseconds | |
1275 | */ | |
1276 | static inline u64 default_cfs_period(void) | |
1277 | { | |
1278 | return 100000000ULL; | |
1279 | } | |
ec12cb7f PT |
1280 | |
1281 | static inline u64 sched_cfs_bandwidth_slice(void) | |
1282 | { | |
1283 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
1284 | } | |
1285 | ||
a9cf55b2 PT |
1286 | /* |
1287 | * Replenish runtime according to assigned quota and update expiration time. | |
1288 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
1289 | * additional synchronization around rq->lock. | |
1290 | * | |
1291 | * requires cfs_b->lock | |
1292 | */ | |
1293 | static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) | |
1294 | { | |
1295 | u64 now; | |
1296 | ||
1297 | if (cfs_b->quota == RUNTIME_INF) | |
1298 | return; | |
1299 | ||
1300 | now = sched_clock_cpu(smp_processor_id()); | |
1301 | cfs_b->runtime = cfs_b->quota; | |
1302 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
1303 | } | |
1304 | ||
85dac906 PT |
1305 | /* returns 0 on failure to allocate runtime */ |
1306 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
1307 | { |
1308 | struct task_group *tg = cfs_rq->tg; | |
1309 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 1310 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
1311 | |
1312 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
1313 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
1314 | ||
1315 | raw_spin_lock(&cfs_b->lock); | |
1316 | if (cfs_b->quota == RUNTIME_INF) | |
1317 | amount = min_amount; | |
58088ad0 | 1318 | else { |
a9cf55b2 PT |
1319 | /* |
1320 | * If the bandwidth pool has become inactive, then at least one | |
1321 | * period must have elapsed since the last consumption. | |
1322 | * Refresh the global state and ensure bandwidth timer becomes | |
1323 | * active. | |
1324 | */ | |
1325 | if (!cfs_b->timer_active) { | |
1326 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 1327 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 1328 | } |
58088ad0 PT |
1329 | |
1330 | if (cfs_b->runtime > 0) { | |
1331 | amount = min(cfs_b->runtime, min_amount); | |
1332 | cfs_b->runtime -= amount; | |
1333 | cfs_b->idle = 0; | |
1334 | } | |
ec12cb7f | 1335 | } |
a9cf55b2 | 1336 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
1337 | raw_spin_unlock(&cfs_b->lock); |
1338 | ||
1339 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
1340 | /* |
1341 | * we may have advanced our local expiration to account for allowed | |
1342 | * spread between our sched_clock and the one on which runtime was | |
1343 | * issued. | |
1344 | */ | |
1345 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
1346 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
1347 | |
1348 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
1349 | } |
1350 | ||
a9cf55b2 PT |
1351 | /* |
1352 | * Note: This depends on the synchronization provided by sched_clock and the | |
1353 | * fact that rq->clock snapshots this value. | |
1354 | */ | |
1355 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 1356 | { |
a9cf55b2 PT |
1357 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
1358 | struct rq *rq = rq_of(cfs_rq); | |
1359 | ||
1360 | /* if the deadline is ahead of our clock, nothing to do */ | |
1361 | if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0)) | |
ec12cb7f PT |
1362 | return; |
1363 | ||
a9cf55b2 PT |
1364 | if (cfs_rq->runtime_remaining < 0) |
1365 | return; | |
1366 | ||
1367 | /* | |
1368 | * If the local deadline has passed we have to consider the | |
1369 | * possibility that our sched_clock is 'fast' and the global deadline | |
1370 | * has not truly expired. | |
1371 | * | |
1372 | * Fortunately we can check determine whether this the case by checking | |
1373 | * whether the global deadline has advanced. | |
1374 | */ | |
1375 | ||
1376 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
1377 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
1378 | cfs_rq->runtime_expires += TICK_NSEC; | |
1379 | } else { | |
1380 | /* global deadline is ahead, expiration has passed */ | |
1381 | cfs_rq->runtime_remaining = 0; | |
1382 | } | |
1383 | } | |
1384 | ||
1385 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1386 | unsigned long delta_exec) | |
1387 | { | |
1388 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 1389 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
1390 | expire_cfs_rq_runtime(cfs_rq); |
1391 | ||
1392 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
1393 | return; |
1394 | ||
85dac906 PT |
1395 | /* |
1396 | * if we're unable to extend our runtime we resched so that the active | |
1397 | * hierarchy can be throttled | |
1398 | */ | |
1399 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
1400 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
1401 | } |
1402 | ||
1403 | static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1404 | unsigned long delta_exec) | |
1405 | { | |
1406 | if (!cfs_rq->runtime_enabled) | |
1407 | return; | |
1408 | ||
1409 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
1410 | } | |
1411 | ||
85dac906 PT |
1412 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
1413 | { | |
1414 | return cfs_rq->throttled; | |
1415 | } | |
1416 | ||
64660c86 PT |
1417 | /* check whether cfs_rq, or any parent, is throttled */ |
1418 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
1419 | { | |
1420 | return cfs_rq->throttle_count; | |
1421 | } | |
1422 | ||
1423 | /* | |
1424 | * Ensure that neither of the group entities corresponding to src_cpu or | |
1425 | * dest_cpu are members of a throttled hierarchy when performing group | |
1426 | * load-balance operations. | |
1427 | */ | |
1428 | static inline int throttled_lb_pair(struct task_group *tg, | |
1429 | int src_cpu, int dest_cpu) | |
1430 | { | |
1431 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
1432 | ||
1433 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
1434 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
1435 | ||
1436 | return throttled_hierarchy(src_cfs_rq) || | |
1437 | throttled_hierarchy(dest_cfs_rq); | |
1438 | } | |
1439 | ||
1440 | /* updated child weight may affect parent so we have to do this bottom up */ | |
1441 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
1442 | { | |
1443 | struct rq *rq = data; | |
1444 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1445 | ||
1446 | cfs_rq->throttle_count--; | |
1447 | #ifdef CONFIG_SMP | |
1448 | if (!cfs_rq->throttle_count) { | |
1449 | u64 delta = rq->clock_task - cfs_rq->load_stamp; | |
1450 | ||
1451 | /* leaving throttled state, advance shares averaging windows */ | |
1452 | cfs_rq->load_stamp += delta; | |
1453 | cfs_rq->load_last += delta; | |
1454 | ||
1455 | /* update entity weight now that we are on_rq again */ | |
1456 | update_cfs_shares(cfs_rq); | |
1457 | } | |
1458 | #endif | |
1459 | ||
1460 | return 0; | |
1461 | } | |
1462 | ||
1463 | static int tg_throttle_down(struct task_group *tg, void *data) | |
1464 | { | |
1465 | struct rq *rq = data; | |
1466 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1467 | ||
1468 | /* group is entering throttled state, record last load */ | |
1469 | if (!cfs_rq->throttle_count) | |
1470 | update_cfs_load(cfs_rq, 0); | |
1471 | cfs_rq->throttle_count++; | |
1472 | ||
1473 | return 0; | |
1474 | } | |
1475 | ||
d3d9dc33 | 1476 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
1477 | { |
1478 | struct rq *rq = rq_of(cfs_rq); | |
1479 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1480 | struct sched_entity *se; | |
1481 | long task_delta, dequeue = 1; | |
1482 | ||
1483 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1484 | ||
1485 | /* account load preceding throttle */ | |
64660c86 PT |
1486 | rcu_read_lock(); |
1487 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
1488 | rcu_read_unlock(); | |
85dac906 PT |
1489 | |
1490 | task_delta = cfs_rq->h_nr_running; | |
1491 | for_each_sched_entity(se) { | |
1492 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
1493 | /* throttled entity or throttle-on-deactivate */ | |
1494 | if (!se->on_rq) | |
1495 | break; | |
1496 | ||
1497 | if (dequeue) | |
1498 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
1499 | qcfs_rq->h_nr_running -= task_delta; | |
1500 | ||
1501 | if (qcfs_rq->load.weight) | |
1502 | dequeue = 0; | |
1503 | } | |
1504 | ||
1505 | if (!se) | |
1506 | rq->nr_running -= task_delta; | |
1507 | ||
1508 | cfs_rq->throttled = 1; | |
e8da1b18 | 1509 | cfs_rq->throttled_timestamp = rq->clock; |
85dac906 PT |
1510 | raw_spin_lock(&cfs_b->lock); |
1511 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
1512 | raw_spin_unlock(&cfs_b->lock); | |
1513 | } | |
1514 | ||
671fd9da PT |
1515 | static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
1516 | { | |
1517 | struct rq *rq = rq_of(cfs_rq); | |
1518 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1519 | struct sched_entity *se; | |
1520 | int enqueue = 1; | |
1521 | long task_delta; | |
1522 | ||
1523 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1524 | ||
1525 | cfs_rq->throttled = 0; | |
1526 | raw_spin_lock(&cfs_b->lock); | |
e8da1b18 | 1527 | cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp; |
671fd9da PT |
1528 | list_del_rcu(&cfs_rq->throttled_list); |
1529 | raw_spin_unlock(&cfs_b->lock); | |
e8da1b18 | 1530 | cfs_rq->throttled_timestamp = 0; |
671fd9da | 1531 | |
64660c86 PT |
1532 | update_rq_clock(rq); |
1533 | /* update hierarchical throttle state */ | |
1534 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
1535 | ||
671fd9da PT |
1536 | if (!cfs_rq->load.weight) |
1537 | return; | |
1538 | ||
1539 | task_delta = cfs_rq->h_nr_running; | |
1540 | for_each_sched_entity(se) { | |
1541 | if (se->on_rq) | |
1542 | enqueue = 0; | |
1543 | ||
1544 | cfs_rq = cfs_rq_of(se); | |
1545 | if (enqueue) | |
1546 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
1547 | cfs_rq->h_nr_running += task_delta; | |
1548 | ||
1549 | if (cfs_rq_throttled(cfs_rq)) | |
1550 | break; | |
1551 | } | |
1552 | ||
1553 | if (!se) | |
1554 | rq->nr_running += task_delta; | |
1555 | ||
1556 | /* determine whether we need to wake up potentially idle cpu */ | |
1557 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
1558 | resched_task(rq->curr); | |
1559 | } | |
1560 | ||
1561 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
1562 | u64 remaining, u64 expires) | |
1563 | { | |
1564 | struct cfs_rq *cfs_rq; | |
1565 | u64 runtime = remaining; | |
1566 | ||
1567 | rcu_read_lock(); | |
1568 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
1569 | throttled_list) { | |
1570 | struct rq *rq = rq_of(cfs_rq); | |
1571 | ||
1572 | raw_spin_lock(&rq->lock); | |
1573 | if (!cfs_rq_throttled(cfs_rq)) | |
1574 | goto next; | |
1575 | ||
1576 | runtime = -cfs_rq->runtime_remaining + 1; | |
1577 | if (runtime > remaining) | |
1578 | runtime = remaining; | |
1579 | remaining -= runtime; | |
1580 | ||
1581 | cfs_rq->runtime_remaining += runtime; | |
1582 | cfs_rq->runtime_expires = expires; | |
1583 | ||
1584 | /* we check whether we're throttled above */ | |
1585 | if (cfs_rq->runtime_remaining > 0) | |
1586 | unthrottle_cfs_rq(cfs_rq); | |
1587 | ||
1588 | next: | |
1589 | raw_spin_unlock(&rq->lock); | |
1590 | ||
1591 | if (!remaining) | |
1592 | break; | |
1593 | } | |
1594 | rcu_read_unlock(); | |
1595 | ||
1596 | return remaining; | |
1597 | } | |
1598 | ||
58088ad0 PT |
1599 | /* |
1600 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
1601 | * cfs_rqs as appropriate. If there has been no activity within the last | |
1602 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
1603 | * used to track this state. | |
1604 | */ | |
1605 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
1606 | { | |
671fd9da PT |
1607 | u64 runtime, runtime_expires; |
1608 | int idle = 1, throttled; | |
58088ad0 PT |
1609 | |
1610 | raw_spin_lock(&cfs_b->lock); | |
1611 | /* no need to continue the timer with no bandwidth constraint */ | |
1612 | if (cfs_b->quota == RUNTIME_INF) | |
1613 | goto out_unlock; | |
1614 | ||
671fd9da PT |
1615 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
1616 | /* idle depends on !throttled (for the case of a large deficit) */ | |
1617 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 1618 | cfs_b->nr_periods += overrun; |
671fd9da | 1619 | |
a9cf55b2 PT |
1620 | /* if we're going inactive then everything else can be deferred */ |
1621 | if (idle) | |
1622 | goto out_unlock; | |
1623 | ||
1624 | __refill_cfs_bandwidth_runtime(cfs_b); | |
1625 | ||
671fd9da PT |
1626 | if (!throttled) { |
1627 | /* mark as potentially idle for the upcoming period */ | |
1628 | cfs_b->idle = 1; | |
1629 | goto out_unlock; | |
1630 | } | |
1631 | ||
e8da1b18 NR |
1632 | /* account preceding periods in which throttling occurred */ |
1633 | cfs_b->nr_throttled += overrun; | |
1634 | ||
671fd9da PT |
1635 | /* |
1636 | * There are throttled entities so we must first use the new bandwidth | |
1637 | * to unthrottle them before making it generally available. This | |
1638 | * ensures that all existing debts will be paid before a new cfs_rq is | |
1639 | * allowed to run. | |
1640 | */ | |
1641 | runtime = cfs_b->runtime; | |
1642 | runtime_expires = cfs_b->runtime_expires; | |
1643 | cfs_b->runtime = 0; | |
1644 | ||
1645 | /* | |
1646 | * This check is repeated as we are holding onto the new bandwidth | |
1647 | * while we unthrottle. This can potentially race with an unthrottled | |
1648 | * group trying to acquire new bandwidth from the global pool. | |
1649 | */ | |
1650 | while (throttled && runtime > 0) { | |
1651 | raw_spin_unlock(&cfs_b->lock); | |
1652 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
1653 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
1654 | runtime_expires); | |
1655 | raw_spin_lock(&cfs_b->lock); | |
1656 | ||
1657 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
1658 | } | |
58088ad0 | 1659 | |
671fd9da PT |
1660 | /* return (any) remaining runtime */ |
1661 | cfs_b->runtime = runtime; | |
1662 | /* | |
1663 | * While we are ensured activity in the period following an | |
1664 | * unthrottle, this also covers the case in which the new bandwidth is | |
1665 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
1666 | * timer to remain active while there are any throttled entities.) | |
1667 | */ | |
1668 | cfs_b->idle = 0; | |
58088ad0 PT |
1669 | out_unlock: |
1670 | if (idle) | |
1671 | cfs_b->timer_active = 0; | |
1672 | raw_spin_unlock(&cfs_b->lock); | |
1673 | ||
1674 | return idle; | |
1675 | } | |
d3d9dc33 PT |
1676 | |
1677 | /* | |
1678 | * When a group wakes up we want to make sure that its quota is not already | |
1679 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
1680 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
1681 | */ | |
1682 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
1683 | { | |
1684 | /* an active group must be handled by the update_curr()->put() path */ | |
1685 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
1686 | return; | |
1687 | ||
1688 | /* ensure the group is not already throttled */ | |
1689 | if (cfs_rq_throttled(cfs_rq)) | |
1690 | return; | |
1691 | ||
1692 | /* update runtime allocation */ | |
1693 | account_cfs_rq_runtime(cfs_rq, 0); | |
1694 | if (cfs_rq->runtime_remaining <= 0) | |
1695 | throttle_cfs_rq(cfs_rq); | |
1696 | } | |
1697 | ||
1698 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
1699 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
1700 | { | |
1701 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) | |
1702 | return; | |
1703 | ||
1704 | /* | |
1705 | * it's possible for a throttled entity to be forced into a running | |
1706 | * state (e.g. set_curr_task), in this case we're finished. | |
1707 | */ | |
1708 | if (cfs_rq_throttled(cfs_rq)) | |
1709 | return; | |
1710 | ||
1711 | throttle_cfs_rq(cfs_rq); | |
1712 | } | |
ec12cb7f PT |
1713 | #else |
1714 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1715 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
1716 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
1717 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
85dac906 PT |
1718 | |
1719 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
1720 | { | |
1721 | return 0; | |
1722 | } | |
64660c86 PT |
1723 | |
1724 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
1725 | { | |
1726 | return 0; | |
1727 | } | |
1728 | ||
1729 | static inline int throttled_lb_pair(struct task_group *tg, | |
1730 | int src_cpu, int dest_cpu) | |
1731 | { | |
1732 | return 0; | |
1733 | } | |
ab84d31e PT |
1734 | #endif |
1735 | ||
bf0f6f24 IM |
1736 | /************************************************** |
1737 | * CFS operations on tasks: | |
1738 | */ | |
1739 | ||
8f4d37ec PZ |
1740 | #ifdef CONFIG_SCHED_HRTICK |
1741 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1742 | { | |
8f4d37ec PZ |
1743 | struct sched_entity *se = &p->se; |
1744 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1745 | ||
1746 | WARN_ON(task_rq(p) != rq); | |
1747 | ||
1748 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { | |
1749 | u64 slice = sched_slice(cfs_rq, se); | |
1750 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
1751 | s64 delta = slice - ran; | |
1752 | ||
1753 | if (delta < 0) { | |
1754 | if (rq->curr == p) | |
1755 | resched_task(p); | |
1756 | return; | |
1757 | } | |
1758 | ||
1759 | /* | |
1760 | * Don't schedule slices shorter than 10000ns, that just | |
1761 | * doesn't make sense. Rely on vruntime for fairness. | |
1762 | */ | |
31656519 | 1763 | if (rq->curr != p) |
157124c1 | 1764 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 1765 | |
31656519 | 1766 | hrtick_start(rq, delta); |
8f4d37ec PZ |
1767 | } |
1768 | } | |
a4c2f00f PZ |
1769 | |
1770 | /* | |
1771 | * called from enqueue/dequeue and updates the hrtick when the | |
1772 | * current task is from our class and nr_running is low enough | |
1773 | * to matter. | |
1774 | */ | |
1775 | static void hrtick_update(struct rq *rq) | |
1776 | { | |
1777 | struct task_struct *curr = rq->curr; | |
1778 | ||
1779 | if (curr->sched_class != &fair_sched_class) | |
1780 | return; | |
1781 | ||
1782 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
1783 | hrtick_start_fair(rq, curr); | |
1784 | } | |
55e12e5e | 1785 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
1786 | static inline void |
1787 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
1788 | { | |
1789 | } | |
a4c2f00f PZ |
1790 | |
1791 | static inline void hrtick_update(struct rq *rq) | |
1792 | { | |
1793 | } | |
8f4d37ec PZ |
1794 | #endif |
1795 | ||
bf0f6f24 IM |
1796 | /* |
1797 | * The enqueue_task method is called before nr_running is | |
1798 | * increased. Here we update the fair scheduling stats and | |
1799 | * then put the task into the rbtree: | |
1800 | */ | |
ea87bb78 | 1801 | static void |
371fd7e7 | 1802 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1803 | { |
1804 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1805 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
1806 | |
1807 | for_each_sched_entity(se) { | |
62fb1851 | 1808 | if (se->on_rq) |
bf0f6f24 IM |
1809 | break; |
1810 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 1811 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
1812 | |
1813 | /* | |
1814 | * end evaluation on encountering a throttled cfs_rq | |
1815 | * | |
1816 | * note: in the case of encountering a throttled cfs_rq we will | |
1817 | * post the final h_nr_running increment below. | |
1818 | */ | |
1819 | if (cfs_rq_throttled(cfs_rq)) | |
1820 | break; | |
953bfcd1 | 1821 | cfs_rq->h_nr_running++; |
85dac906 | 1822 | |
88ec22d3 | 1823 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 1824 | } |
8f4d37ec | 1825 | |
2069dd75 | 1826 | for_each_sched_entity(se) { |
0f317143 | 1827 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 1828 | cfs_rq->h_nr_running++; |
2069dd75 | 1829 | |
85dac906 PT |
1830 | if (cfs_rq_throttled(cfs_rq)) |
1831 | break; | |
1832 | ||
d6b55918 | 1833 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1834 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1835 | } |
1836 | ||
85dac906 PT |
1837 | if (!se) |
1838 | inc_nr_running(rq); | |
a4c2f00f | 1839 | hrtick_update(rq); |
bf0f6f24 IM |
1840 | } |
1841 | ||
2f36825b VP |
1842 | static void set_next_buddy(struct sched_entity *se); |
1843 | ||
bf0f6f24 IM |
1844 | /* |
1845 | * The dequeue_task method is called before nr_running is | |
1846 | * decreased. We remove the task from the rbtree and | |
1847 | * update the fair scheduling stats: | |
1848 | */ | |
371fd7e7 | 1849 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
1850 | { |
1851 | struct cfs_rq *cfs_rq; | |
62fb1851 | 1852 | struct sched_entity *se = &p->se; |
2f36825b | 1853 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
1854 | |
1855 | for_each_sched_entity(se) { | |
1856 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 1857 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
1858 | |
1859 | /* | |
1860 | * end evaluation on encountering a throttled cfs_rq | |
1861 | * | |
1862 | * note: in the case of encountering a throttled cfs_rq we will | |
1863 | * post the final h_nr_running decrement below. | |
1864 | */ | |
1865 | if (cfs_rq_throttled(cfs_rq)) | |
1866 | break; | |
953bfcd1 | 1867 | cfs_rq->h_nr_running--; |
2069dd75 | 1868 | |
bf0f6f24 | 1869 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
1870 | if (cfs_rq->load.weight) { |
1871 | /* | |
1872 | * Bias pick_next to pick a task from this cfs_rq, as | |
1873 | * p is sleeping when it is within its sched_slice. | |
1874 | */ | |
1875 | if (task_sleep && parent_entity(se)) | |
1876 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
1877 | |
1878 | /* avoid re-evaluating load for this entity */ | |
1879 | se = parent_entity(se); | |
bf0f6f24 | 1880 | break; |
2f36825b | 1881 | } |
371fd7e7 | 1882 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 1883 | } |
8f4d37ec | 1884 | |
2069dd75 | 1885 | for_each_sched_entity(se) { |
0f317143 | 1886 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 1887 | cfs_rq->h_nr_running--; |
2069dd75 | 1888 | |
85dac906 PT |
1889 | if (cfs_rq_throttled(cfs_rq)) |
1890 | break; | |
1891 | ||
d6b55918 | 1892 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 1893 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
1894 | } |
1895 | ||
85dac906 PT |
1896 | if (!se) |
1897 | dec_nr_running(rq); | |
a4c2f00f | 1898 | hrtick_update(rq); |
bf0f6f24 IM |
1899 | } |
1900 | ||
e7693a36 | 1901 | #ifdef CONFIG_SMP |
098fb9db | 1902 | |
74f8e4b2 | 1903 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
1904 | { |
1905 | struct sched_entity *se = &p->se; | |
1906 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
1907 | u64 min_vruntime; |
1908 | ||
1909 | #ifndef CONFIG_64BIT | |
1910 | u64 min_vruntime_copy; | |
88ec22d3 | 1911 | |
3fe1698b PZ |
1912 | do { |
1913 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
1914 | smp_rmb(); | |
1915 | min_vruntime = cfs_rq->min_vruntime; | |
1916 | } while (min_vruntime != min_vruntime_copy); | |
1917 | #else | |
1918 | min_vruntime = cfs_rq->min_vruntime; | |
1919 | #endif | |
88ec22d3 | 1920 | |
3fe1698b | 1921 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
1922 | } |
1923 | ||
bb3469ac | 1924 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
1925 | /* |
1926 | * effective_load() calculates the load change as seen from the root_task_group | |
1927 | * | |
1928 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
1929 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
1930 | * can calculate the shift in shares. | |
f5bfb7d9 | 1931 | */ |
2069dd75 | 1932 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 1933 | { |
4be9daaa | 1934 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 PZ |
1935 | |
1936 | if (!tg->parent) | |
1937 | return wl; | |
1938 | ||
4be9daaa | 1939 | for_each_sched_entity(se) { |
977dda7c | 1940 | long lw, w; |
4be9daaa | 1941 | |
977dda7c PT |
1942 | tg = se->my_q->tg; |
1943 | w = se->my_q->load.weight; | |
bb3469ac | 1944 | |
977dda7c PT |
1945 | /* use this cpu's instantaneous contribution */ |
1946 | lw = atomic_read(&tg->load_weight); | |
1947 | lw -= se->my_q->load_contribution; | |
1948 | lw += w + wg; | |
4be9daaa | 1949 | |
977dda7c | 1950 | wl += w; |
940959e9 | 1951 | |
977dda7c PT |
1952 | if (lw > 0 && wl < lw) |
1953 | wl = (wl * tg->shares) / lw; | |
1954 | else | |
1955 | wl = tg->shares; | |
940959e9 | 1956 | |
977dda7c PT |
1957 | /* zero point is MIN_SHARES */ |
1958 | if (wl < MIN_SHARES) | |
1959 | wl = MIN_SHARES; | |
1960 | wl -= se->load.weight; | |
4be9daaa | 1961 | wg = 0; |
4be9daaa | 1962 | } |
bb3469ac | 1963 | |
4be9daaa | 1964 | return wl; |
bb3469ac PZ |
1965 | } |
1966 | #else | |
4be9daaa | 1967 | |
83378269 PZ |
1968 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
1969 | unsigned long wl, unsigned long wg) | |
4be9daaa | 1970 | { |
83378269 | 1971 | return wl; |
bb3469ac | 1972 | } |
4be9daaa | 1973 | |
bb3469ac PZ |
1974 | #endif |
1975 | ||
c88d5910 | 1976 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 1977 | { |
e37b6a7b | 1978 | s64 this_load, load; |
c88d5910 | 1979 | int idx, this_cpu, prev_cpu; |
098fb9db | 1980 | unsigned long tl_per_task; |
c88d5910 | 1981 | struct task_group *tg; |
83378269 | 1982 | unsigned long weight; |
b3137bc8 | 1983 | int balanced; |
098fb9db | 1984 | |
c88d5910 PZ |
1985 | idx = sd->wake_idx; |
1986 | this_cpu = smp_processor_id(); | |
1987 | prev_cpu = task_cpu(p); | |
1988 | load = source_load(prev_cpu, idx); | |
1989 | this_load = target_load(this_cpu, idx); | |
098fb9db | 1990 | |
b3137bc8 MG |
1991 | /* |
1992 | * If sync wakeup then subtract the (maximum possible) | |
1993 | * effect of the currently running task from the load | |
1994 | * of the current CPU: | |
1995 | */ | |
83378269 PZ |
1996 | if (sync) { |
1997 | tg = task_group(current); | |
1998 | weight = current->se.load.weight; | |
1999 | ||
c88d5910 | 2000 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
2001 | load += effective_load(tg, prev_cpu, 0, -weight); |
2002 | } | |
b3137bc8 | 2003 | |
83378269 PZ |
2004 | tg = task_group(p); |
2005 | weight = p->se.load.weight; | |
b3137bc8 | 2006 | |
71a29aa7 PZ |
2007 | /* |
2008 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
2009 | * due to the sync cause above having dropped this_load to 0, we'll |
2010 | * always have an imbalance, but there's really nothing you can do | |
2011 | * about that, so that's good too. | |
71a29aa7 PZ |
2012 | * |
2013 | * Otherwise check if either cpus are near enough in load to allow this | |
2014 | * task to be woken on this_cpu. | |
2015 | */ | |
e37b6a7b PT |
2016 | if (this_load > 0) { |
2017 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
2018 | |
2019 | this_eff_load = 100; | |
2020 | this_eff_load *= power_of(prev_cpu); | |
2021 | this_eff_load *= this_load + | |
2022 | effective_load(tg, this_cpu, weight, weight); | |
2023 | ||
2024 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
2025 | prev_eff_load *= power_of(this_cpu); | |
2026 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
2027 | ||
2028 | balanced = this_eff_load <= prev_eff_load; | |
2029 | } else | |
2030 | balanced = true; | |
b3137bc8 | 2031 | |
098fb9db | 2032 | /* |
4ae7d5ce IM |
2033 | * If the currently running task will sleep within |
2034 | * a reasonable amount of time then attract this newly | |
2035 | * woken task: | |
098fb9db | 2036 | */ |
2fb7635c PZ |
2037 | if (sync && balanced) |
2038 | return 1; | |
098fb9db | 2039 | |
41acab88 | 2040 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
2041 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
2042 | ||
c88d5910 PZ |
2043 | if (balanced || |
2044 | (this_load <= load && | |
2045 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
2046 | /* |
2047 | * This domain has SD_WAKE_AFFINE and | |
2048 | * p is cache cold in this domain, and | |
2049 | * there is no bad imbalance. | |
2050 | */ | |
c88d5910 | 2051 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 2052 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
2053 | |
2054 | return 1; | |
2055 | } | |
2056 | return 0; | |
2057 | } | |
2058 | ||
aaee1203 PZ |
2059 | /* |
2060 | * find_idlest_group finds and returns the least busy CPU group within the | |
2061 | * domain. | |
2062 | */ | |
2063 | static struct sched_group * | |
78e7ed53 | 2064 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 2065 | int this_cpu, int load_idx) |
e7693a36 | 2066 | { |
b3bd3de6 | 2067 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 2068 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 2069 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 2070 | |
aaee1203 PZ |
2071 | do { |
2072 | unsigned long load, avg_load; | |
2073 | int local_group; | |
2074 | int i; | |
e7693a36 | 2075 | |
aaee1203 PZ |
2076 | /* Skip over this group if it has no CPUs allowed */ |
2077 | if (!cpumask_intersects(sched_group_cpus(group), | |
2078 | &p->cpus_allowed)) | |
2079 | continue; | |
2080 | ||
2081 | local_group = cpumask_test_cpu(this_cpu, | |
2082 | sched_group_cpus(group)); | |
2083 | ||
2084 | /* Tally up the load of all CPUs in the group */ | |
2085 | avg_load = 0; | |
2086 | ||
2087 | for_each_cpu(i, sched_group_cpus(group)) { | |
2088 | /* Bias balancing toward cpus of our domain */ | |
2089 | if (local_group) | |
2090 | load = source_load(i, load_idx); | |
2091 | else | |
2092 | load = target_load(i, load_idx); | |
2093 | ||
2094 | avg_load += load; | |
2095 | } | |
2096 | ||
2097 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2098 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
2099 | |
2100 | if (local_group) { | |
2101 | this_load = avg_load; | |
aaee1203 PZ |
2102 | } else if (avg_load < min_load) { |
2103 | min_load = avg_load; | |
2104 | idlest = group; | |
2105 | } | |
2106 | } while (group = group->next, group != sd->groups); | |
2107 | ||
2108 | if (!idlest || 100*this_load < imbalance*min_load) | |
2109 | return NULL; | |
2110 | return idlest; | |
2111 | } | |
2112 | ||
2113 | /* | |
2114 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
2115 | */ | |
2116 | static int | |
2117 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
2118 | { | |
2119 | unsigned long load, min_load = ULONG_MAX; | |
2120 | int idlest = -1; | |
2121 | int i; | |
2122 | ||
2123 | /* Traverse only the allowed CPUs */ | |
2124 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { | |
2125 | load = weighted_cpuload(i); | |
2126 | ||
2127 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2128 | min_load = load; | |
2129 | idlest = i; | |
e7693a36 GH |
2130 | } |
2131 | } | |
2132 | ||
aaee1203 PZ |
2133 | return idlest; |
2134 | } | |
e7693a36 | 2135 | |
a50bde51 PZ |
2136 | /* |
2137 | * Try and locate an idle CPU in the sched_domain. | |
2138 | */ | |
99bd5e2f | 2139 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
2140 | { |
2141 | int cpu = smp_processor_id(); | |
2142 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 2143 | struct sched_domain *sd; |
a50bde51 PZ |
2144 | int i; |
2145 | ||
2146 | /* | |
99bd5e2f SS |
2147 | * If the task is going to be woken-up on this cpu and if it is |
2148 | * already idle, then it is the right target. | |
a50bde51 | 2149 | */ |
99bd5e2f SS |
2150 | if (target == cpu && idle_cpu(cpu)) |
2151 | return cpu; | |
2152 | ||
2153 | /* | |
2154 | * If the task is going to be woken-up on the cpu where it previously | |
2155 | * ran and if it is currently idle, then it the right target. | |
2156 | */ | |
2157 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 2158 | return prev_cpu; |
a50bde51 PZ |
2159 | |
2160 | /* | |
99bd5e2f | 2161 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 2162 | */ |
dce840a0 | 2163 | rcu_read_lock(); |
99bd5e2f SS |
2164 | for_each_domain(target, sd) { |
2165 | if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) | |
fe3bcfe1 | 2166 | break; |
99bd5e2f SS |
2167 | |
2168 | for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) { | |
2169 | if (idle_cpu(i)) { | |
2170 | target = i; | |
2171 | break; | |
2172 | } | |
a50bde51 | 2173 | } |
99bd5e2f SS |
2174 | |
2175 | /* | |
2176 | * Lets stop looking for an idle sibling when we reached | |
2177 | * the domain that spans the current cpu and prev_cpu. | |
2178 | */ | |
2179 | if (cpumask_test_cpu(cpu, sched_domain_span(sd)) && | |
2180 | cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
2181 | break; | |
a50bde51 | 2182 | } |
dce840a0 | 2183 | rcu_read_unlock(); |
a50bde51 PZ |
2184 | |
2185 | return target; | |
2186 | } | |
2187 | ||
aaee1203 PZ |
2188 | /* |
2189 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2190 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2191 | * SD_BALANCE_EXEC. | |
2192 | * | |
2193 | * Balance, ie. select the least loaded group. | |
2194 | * | |
2195 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2196 | * | |
2197 | * preempt must be disabled. | |
2198 | */ | |
0017d735 | 2199 | static int |
7608dec2 | 2200 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 2201 | { |
29cd8bae | 2202 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
2203 | int cpu = smp_processor_id(); |
2204 | int prev_cpu = task_cpu(p); | |
2205 | int new_cpu = cpu; | |
99bd5e2f | 2206 | int want_affine = 0; |
29cd8bae | 2207 | int want_sd = 1; |
5158f4e4 | 2208 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 2209 | |
0763a660 | 2210 | if (sd_flag & SD_BALANCE_WAKE) { |
beac4c7e | 2211 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) |
c88d5910 PZ |
2212 | want_affine = 1; |
2213 | new_cpu = prev_cpu; | |
2214 | } | |
aaee1203 | 2215 | |
dce840a0 | 2216 | rcu_read_lock(); |
aaee1203 | 2217 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
2218 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
2219 | continue; | |
2220 | ||
aaee1203 | 2221 | /* |
ae154be1 PZ |
2222 | * If power savings logic is enabled for a domain, see if we |
2223 | * are not overloaded, if so, don't balance wider. | |
aaee1203 | 2224 | */ |
59abf026 | 2225 | if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { |
ae154be1 PZ |
2226 | unsigned long power = 0; |
2227 | unsigned long nr_running = 0; | |
2228 | unsigned long capacity; | |
2229 | int i; | |
2230 | ||
2231 | for_each_cpu(i, sched_domain_span(tmp)) { | |
2232 | power += power_of(i); | |
2233 | nr_running += cpu_rq(i)->cfs.nr_running; | |
2234 | } | |
2235 | ||
1399fa78 | 2236 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); |
ae154be1 | 2237 | |
59abf026 PZ |
2238 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
2239 | nr_running /= 2; | |
2240 | ||
2241 | if (nr_running < capacity) | |
29cd8bae | 2242 | want_sd = 0; |
ae154be1 | 2243 | } |
aaee1203 | 2244 | |
fe3bcfe1 | 2245 | /* |
99bd5e2f SS |
2246 | * If both cpu and prev_cpu are part of this domain, |
2247 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 2248 | */ |
99bd5e2f SS |
2249 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
2250 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
2251 | affine_sd = tmp; | |
2252 | want_affine = 0; | |
c88d5910 PZ |
2253 | } |
2254 | ||
29cd8bae PZ |
2255 | if (!want_sd && !want_affine) |
2256 | break; | |
2257 | ||
0763a660 | 2258 | if (!(tmp->flags & sd_flag)) |
c88d5910 PZ |
2259 | continue; |
2260 | ||
29cd8bae PZ |
2261 | if (want_sd) |
2262 | sd = tmp; | |
2263 | } | |
2264 | ||
8b911acd | 2265 | if (affine_sd) { |
99bd5e2f | 2266 | if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
2267 | prev_cpu = cpu; |
2268 | ||
2269 | new_cpu = select_idle_sibling(p, prev_cpu); | |
2270 | goto unlock; | |
8b911acd | 2271 | } |
e7693a36 | 2272 | |
aaee1203 | 2273 | while (sd) { |
5158f4e4 | 2274 | int load_idx = sd->forkexec_idx; |
aaee1203 | 2275 | struct sched_group *group; |
c88d5910 | 2276 | int weight; |
098fb9db | 2277 | |
0763a660 | 2278 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
2279 | sd = sd->child; |
2280 | continue; | |
2281 | } | |
098fb9db | 2282 | |
5158f4e4 PZ |
2283 | if (sd_flag & SD_BALANCE_WAKE) |
2284 | load_idx = sd->wake_idx; | |
098fb9db | 2285 | |
5158f4e4 | 2286 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
2287 | if (!group) { |
2288 | sd = sd->child; | |
2289 | continue; | |
2290 | } | |
4ae7d5ce | 2291 | |
d7c33c49 | 2292 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
2293 | if (new_cpu == -1 || new_cpu == cpu) { |
2294 | /* Now try balancing at a lower domain level of cpu */ | |
2295 | sd = sd->child; | |
2296 | continue; | |
e7693a36 | 2297 | } |
aaee1203 PZ |
2298 | |
2299 | /* Now try balancing at a lower domain level of new_cpu */ | |
2300 | cpu = new_cpu; | |
669c55e9 | 2301 | weight = sd->span_weight; |
aaee1203 PZ |
2302 | sd = NULL; |
2303 | for_each_domain(cpu, tmp) { | |
669c55e9 | 2304 | if (weight <= tmp->span_weight) |
aaee1203 | 2305 | break; |
0763a660 | 2306 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
2307 | sd = tmp; |
2308 | } | |
2309 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 2310 | } |
dce840a0 PZ |
2311 | unlock: |
2312 | rcu_read_unlock(); | |
e7693a36 | 2313 | |
c88d5910 | 2314 | return new_cpu; |
e7693a36 GH |
2315 | } |
2316 | #endif /* CONFIG_SMP */ | |
2317 | ||
e52fb7c0 PZ |
2318 | static unsigned long |
2319 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
2320 | { |
2321 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
2322 | ||
2323 | /* | |
e52fb7c0 PZ |
2324 | * Since its curr running now, convert the gran from real-time |
2325 | * to virtual-time in his units. | |
13814d42 MG |
2326 | * |
2327 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
2328 | * they get preempted easier. That is, if 'se' < 'curr' then | |
2329 | * the resulting gran will be larger, therefore penalizing the | |
2330 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
2331 | * be smaller, again penalizing the lighter task. | |
2332 | * | |
2333 | * This is especially important for buddies when the leftmost | |
2334 | * task is higher priority than the buddy. | |
0bbd3336 | 2335 | */ |
f4ad9bd2 | 2336 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
2337 | } |
2338 | ||
464b7527 PZ |
2339 | /* |
2340 | * Should 'se' preempt 'curr'. | |
2341 | * | |
2342 | * |s1 | |
2343 | * |s2 | |
2344 | * |s3 | |
2345 | * g | |
2346 | * |<--->|c | |
2347 | * | |
2348 | * w(c, s1) = -1 | |
2349 | * w(c, s2) = 0 | |
2350 | * w(c, s3) = 1 | |
2351 | * | |
2352 | */ | |
2353 | static int | |
2354 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
2355 | { | |
2356 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
2357 | ||
2358 | if (vdiff <= 0) | |
2359 | return -1; | |
2360 | ||
e52fb7c0 | 2361 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
2362 | if (vdiff > gran) |
2363 | return 1; | |
2364 | ||
2365 | return 0; | |
2366 | } | |
2367 | ||
02479099 PZ |
2368 | static void set_last_buddy(struct sched_entity *se) |
2369 | { | |
69c80f3e VP |
2370 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
2371 | return; | |
2372 | ||
2373 | for_each_sched_entity(se) | |
2374 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
2375 | } |
2376 | ||
2377 | static void set_next_buddy(struct sched_entity *se) | |
2378 | { | |
69c80f3e VP |
2379 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
2380 | return; | |
2381 | ||
2382 | for_each_sched_entity(se) | |
2383 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
2384 | } |
2385 | ||
ac53db59 RR |
2386 | static void set_skip_buddy(struct sched_entity *se) |
2387 | { | |
69c80f3e VP |
2388 | for_each_sched_entity(se) |
2389 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
2390 | } |
2391 | ||
bf0f6f24 IM |
2392 | /* |
2393 | * Preempt the current task with a newly woken task if needed: | |
2394 | */ | |
5a9b86f6 | 2395 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
2396 | { |
2397 | struct task_struct *curr = rq->curr; | |
8651a86c | 2398 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 2399 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 2400 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 2401 | int next_buddy_marked = 0; |
bf0f6f24 | 2402 | |
4ae7d5ce IM |
2403 | if (unlikely(se == pse)) |
2404 | return; | |
2405 | ||
5238cdd3 PT |
2406 | /* |
2407 | * This is possible from callers such as pull_task(), in which we | |
2408 | * unconditionally check_prempt_curr() after an enqueue (which may have | |
2409 | * lead to a throttle). This both saves work and prevents false | |
2410 | * next-buddy nomination below. | |
2411 | */ | |
2412 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
2413 | return; | |
2414 | ||
2f36825b | 2415 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 2416 | set_next_buddy(pse); |
2f36825b VP |
2417 | next_buddy_marked = 1; |
2418 | } | |
57fdc26d | 2419 | |
aec0a514 BR |
2420 | /* |
2421 | * We can come here with TIF_NEED_RESCHED already set from new task | |
2422 | * wake up path. | |
5238cdd3 PT |
2423 | * |
2424 | * Note: this also catches the edge-case of curr being in a throttled | |
2425 | * group (e.g. via set_curr_task), since update_curr() (in the | |
2426 | * enqueue of curr) will have resulted in resched being set. This | |
2427 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
2428 | * below. | |
aec0a514 BR |
2429 | */ |
2430 | if (test_tsk_need_resched(curr)) | |
2431 | return; | |
2432 | ||
a2f5c9ab DH |
2433 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
2434 | if (unlikely(curr->policy == SCHED_IDLE) && | |
2435 | likely(p->policy != SCHED_IDLE)) | |
2436 | goto preempt; | |
2437 | ||
91c234b4 | 2438 | /* |
a2f5c9ab DH |
2439 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
2440 | * is driven by the tick): | |
91c234b4 | 2441 | */ |
6bc912b7 | 2442 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 2443 | return; |
bf0f6f24 | 2444 | |
464b7527 | 2445 | find_matching_se(&se, &pse); |
9bbd7374 | 2446 | update_curr(cfs_rq_of(se)); |
002f128b | 2447 | BUG_ON(!pse); |
2f36825b VP |
2448 | if (wakeup_preempt_entity(se, pse) == 1) { |
2449 | /* | |
2450 | * Bias pick_next to pick the sched entity that is | |
2451 | * triggering this preemption. | |
2452 | */ | |
2453 | if (!next_buddy_marked) | |
2454 | set_next_buddy(pse); | |
3a7e73a2 | 2455 | goto preempt; |
2f36825b | 2456 | } |
464b7527 | 2457 | |
3a7e73a2 | 2458 | return; |
a65ac745 | 2459 | |
3a7e73a2 PZ |
2460 | preempt: |
2461 | resched_task(curr); | |
2462 | /* | |
2463 | * Only set the backward buddy when the current task is still | |
2464 | * on the rq. This can happen when a wakeup gets interleaved | |
2465 | * with schedule on the ->pre_schedule() or idle_balance() | |
2466 | * point, either of which can * drop the rq lock. | |
2467 | * | |
2468 | * Also, during early boot the idle thread is in the fair class, | |
2469 | * for obvious reasons its a bad idea to schedule back to it. | |
2470 | */ | |
2471 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
2472 | return; | |
2473 | ||
2474 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
2475 | set_last_buddy(se); | |
bf0f6f24 IM |
2476 | } |
2477 | ||
fb8d4724 | 2478 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 2479 | { |
8f4d37ec | 2480 | struct task_struct *p; |
bf0f6f24 IM |
2481 | struct cfs_rq *cfs_rq = &rq->cfs; |
2482 | struct sched_entity *se; | |
2483 | ||
36ace27e | 2484 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
2485 | return NULL; |
2486 | ||
2487 | do { | |
9948f4b2 | 2488 | se = pick_next_entity(cfs_rq); |
f4b6755f | 2489 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
2490 | cfs_rq = group_cfs_rq(se); |
2491 | } while (cfs_rq); | |
2492 | ||
8f4d37ec PZ |
2493 | p = task_of(se); |
2494 | hrtick_start_fair(rq, p); | |
2495 | ||
2496 | return p; | |
bf0f6f24 IM |
2497 | } |
2498 | ||
2499 | /* | |
2500 | * Account for a descheduled task: | |
2501 | */ | |
31ee529c | 2502 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
2503 | { |
2504 | struct sched_entity *se = &prev->se; | |
2505 | struct cfs_rq *cfs_rq; | |
2506 | ||
2507 | for_each_sched_entity(se) { | |
2508 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 2509 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
2510 | } |
2511 | } | |
2512 | ||
ac53db59 RR |
2513 | /* |
2514 | * sched_yield() is very simple | |
2515 | * | |
2516 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
2517 | */ | |
2518 | static void yield_task_fair(struct rq *rq) | |
2519 | { | |
2520 | struct task_struct *curr = rq->curr; | |
2521 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
2522 | struct sched_entity *se = &curr->se; | |
2523 | ||
2524 | /* | |
2525 | * Are we the only task in the tree? | |
2526 | */ | |
2527 | if (unlikely(rq->nr_running == 1)) | |
2528 | return; | |
2529 | ||
2530 | clear_buddies(cfs_rq, se); | |
2531 | ||
2532 | if (curr->policy != SCHED_BATCH) { | |
2533 | update_rq_clock(rq); | |
2534 | /* | |
2535 | * Update run-time statistics of the 'current'. | |
2536 | */ | |
2537 | update_curr(cfs_rq); | |
2538 | } | |
2539 | ||
2540 | set_skip_buddy(se); | |
2541 | } | |
2542 | ||
d95f4122 MG |
2543 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
2544 | { | |
2545 | struct sched_entity *se = &p->se; | |
2546 | ||
5238cdd3 PT |
2547 | /* throttled hierarchies are not runnable */ |
2548 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
2549 | return false; |
2550 | ||
2551 | /* Tell the scheduler that we'd really like pse to run next. */ | |
2552 | set_next_buddy(se); | |
2553 | ||
d95f4122 MG |
2554 | yield_task_fair(rq); |
2555 | ||
2556 | return true; | |
2557 | } | |
2558 | ||
681f3e68 | 2559 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
2560 | /************************************************** |
2561 | * Fair scheduling class load-balancing methods: | |
2562 | */ | |
2563 | ||
1e3c88bd PZ |
2564 | /* |
2565 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
2566 | * Both runqueues must be locked. | |
2567 | */ | |
2568 | static void pull_task(struct rq *src_rq, struct task_struct *p, | |
2569 | struct rq *this_rq, int this_cpu) | |
2570 | { | |
2571 | deactivate_task(src_rq, p, 0); | |
2572 | set_task_cpu(p, this_cpu); | |
2573 | activate_task(this_rq, p, 0); | |
2574 | check_preempt_curr(this_rq, p, 0); | |
2575 | } | |
2576 | ||
2577 | /* | |
2578 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
2579 | */ | |
2580 | static | |
2581 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, | |
2582 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2583 | int *all_pinned) | |
2584 | { | |
2585 | int tsk_cache_hot = 0; | |
2586 | /* | |
2587 | * We do not migrate tasks that are: | |
2588 | * 1) running (obviously), or | |
2589 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
2590 | * 3) are cache-hot on their current CPU. | |
2591 | */ | |
2592 | if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) { | |
41acab88 | 2593 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
1e3c88bd PZ |
2594 | return 0; |
2595 | } | |
2596 | *all_pinned = 0; | |
2597 | ||
2598 | if (task_running(rq, p)) { | |
41acab88 | 2599 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
2600 | return 0; |
2601 | } | |
2602 | ||
2603 | /* | |
2604 | * Aggressive migration if: | |
2605 | * 1) task is cache cold, or | |
2606 | * 2) too many balance attempts have failed. | |
2607 | */ | |
2608 | ||
305e6835 | 2609 | tsk_cache_hot = task_hot(p, rq->clock_task, sd); |
1e3c88bd PZ |
2610 | if (!tsk_cache_hot || |
2611 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
2612 | #ifdef CONFIG_SCHEDSTATS | |
2613 | if (tsk_cache_hot) { | |
2614 | schedstat_inc(sd, lb_hot_gained[idle]); | |
41acab88 | 2615 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
2616 | } |
2617 | #endif | |
2618 | return 1; | |
2619 | } | |
2620 | ||
2621 | if (tsk_cache_hot) { | |
41acab88 | 2622 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
2623 | return 0; |
2624 | } | |
2625 | return 1; | |
2626 | } | |
2627 | ||
897c395f PZ |
2628 | /* |
2629 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
2630 | * part of active balancing operations within "domain". | |
2631 | * Returns 1 if successful and 0 otherwise. | |
2632 | * | |
2633 | * Called with both runqueues locked. | |
2634 | */ | |
2635 | static int | |
2636 | move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2637 | struct sched_domain *sd, enum cpu_idle_type idle) | |
2638 | { | |
2639 | struct task_struct *p, *n; | |
2640 | struct cfs_rq *cfs_rq; | |
2641 | int pinned = 0; | |
2642 | ||
2643 | for_each_leaf_cfs_rq(busiest, cfs_rq) { | |
2644 | list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { | |
64660c86 PT |
2645 | if (throttled_lb_pair(task_group(p), |
2646 | busiest->cpu, this_cpu)) | |
2647 | break; | |
897c395f PZ |
2648 | |
2649 | if (!can_migrate_task(p, busiest, this_cpu, | |
2650 | sd, idle, &pinned)) | |
2651 | continue; | |
2652 | ||
2653 | pull_task(busiest, p, this_rq, this_cpu); | |
2654 | /* | |
2655 | * Right now, this is only the second place pull_task() | |
2656 | * is called, so we can safely collect pull_task() | |
2657 | * stats here rather than inside pull_task(). | |
2658 | */ | |
2659 | schedstat_inc(sd, lb_gained[idle]); | |
2660 | return 1; | |
2661 | } | |
2662 | } | |
2663 | ||
2664 | return 0; | |
2665 | } | |
2666 | ||
1e3c88bd PZ |
2667 | static unsigned long |
2668 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2669 | unsigned long max_load_move, struct sched_domain *sd, | |
2670 | enum cpu_idle_type idle, int *all_pinned, | |
931aeeda | 2671 | struct cfs_rq *busiest_cfs_rq) |
1e3c88bd | 2672 | { |
b30aef17 | 2673 | int loops = 0, pulled = 0; |
1e3c88bd | 2674 | long rem_load_move = max_load_move; |
ee00e66f | 2675 | struct task_struct *p, *n; |
1e3c88bd PZ |
2676 | |
2677 | if (max_load_move == 0) | |
2678 | goto out; | |
2679 | ||
ee00e66f PZ |
2680 | list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { |
2681 | if (loops++ > sysctl_sched_nr_migrate) | |
2682 | break; | |
1e3c88bd | 2683 | |
ee00e66f | 2684 | if ((p->se.load.weight >> 1) > rem_load_move || |
b30aef17 KC |
2685 | !can_migrate_task(p, busiest, this_cpu, sd, idle, |
2686 | all_pinned)) | |
ee00e66f | 2687 | continue; |
1e3c88bd | 2688 | |
ee00e66f PZ |
2689 | pull_task(busiest, p, this_rq, this_cpu); |
2690 | pulled++; | |
2691 | rem_load_move -= p->se.load.weight; | |
1e3c88bd PZ |
2692 | |
2693 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
2694 | /* |
2695 | * NEWIDLE balancing is a source of latency, so preemptible | |
2696 | * kernels will stop after the first task is pulled to minimize | |
2697 | * the critical section. | |
2698 | */ | |
2699 | if (idle == CPU_NEWLY_IDLE) | |
2700 | break; | |
1e3c88bd PZ |
2701 | #endif |
2702 | ||
ee00e66f PZ |
2703 | /* |
2704 | * We only want to steal up to the prescribed amount of | |
2705 | * weighted load. | |
2706 | */ | |
2707 | if (rem_load_move <= 0) | |
2708 | break; | |
1e3c88bd PZ |
2709 | } |
2710 | out: | |
2711 | /* | |
2712 | * Right now, this is one of only two places pull_task() is called, | |
2713 | * so we can safely collect pull_task() stats here rather than | |
2714 | * inside pull_task(). | |
2715 | */ | |
2716 | schedstat_add(sd, lb_gained[idle], pulled); | |
2717 | ||
1e3c88bd PZ |
2718 | return max_load_move - rem_load_move; |
2719 | } | |
2720 | ||
230059de | 2721 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
2722 | /* |
2723 | * update tg->load_weight by folding this cpu's load_avg | |
2724 | */ | |
67e86250 | 2725 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
2726 | { |
2727 | struct cfs_rq *cfs_rq; | |
2728 | unsigned long flags; | |
2729 | struct rq *rq; | |
9e3081ca PZ |
2730 | |
2731 | if (!tg->se[cpu]) | |
2732 | return 0; | |
2733 | ||
2734 | rq = cpu_rq(cpu); | |
2735 | cfs_rq = tg->cfs_rq[cpu]; | |
2736 | ||
2737 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2738 | ||
2739 | update_rq_clock(rq); | |
d6b55918 | 2740 | update_cfs_load(cfs_rq, 1); |
9e3081ca PZ |
2741 | |
2742 | /* | |
2743 | * We need to update shares after updating tg->load_weight in | |
2744 | * order to adjust the weight of groups with long running tasks. | |
2745 | */ | |
6d5ab293 | 2746 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
2747 | |
2748 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2749 | ||
2750 | return 0; | |
2751 | } | |
2752 | ||
2753 | static void update_shares(int cpu) | |
2754 | { | |
2755 | struct cfs_rq *cfs_rq; | |
2756 | struct rq *rq = cpu_rq(cpu); | |
2757 | ||
2758 | rcu_read_lock(); | |
9763b67f PZ |
2759 | /* |
2760 | * Iterates the task_group tree in a bottom up fashion, see | |
2761 | * list_add_leaf_cfs_rq() for details. | |
2762 | */ | |
64660c86 PT |
2763 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
2764 | /* throttled entities do not contribute to load */ | |
2765 | if (throttled_hierarchy(cfs_rq)) | |
2766 | continue; | |
2767 | ||
67e86250 | 2768 | update_shares_cpu(cfs_rq->tg, cpu); |
64660c86 | 2769 | } |
9e3081ca PZ |
2770 | rcu_read_unlock(); |
2771 | } | |
2772 | ||
9763b67f PZ |
2773 | /* |
2774 | * Compute the cpu's hierarchical load factor for each task group. | |
2775 | * This needs to be done in a top-down fashion because the load of a child | |
2776 | * group is a fraction of its parents load. | |
2777 | */ | |
2778 | static int tg_load_down(struct task_group *tg, void *data) | |
2779 | { | |
2780 | unsigned long load; | |
2781 | long cpu = (long)data; | |
2782 | ||
2783 | if (!tg->parent) { | |
2784 | load = cpu_rq(cpu)->load.weight; | |
2785 | } else { | |
2786 | load = tg->parent->cfs_rq[cpu]->h_load; | |
2787 | load *= tg->se[cpu]->load.weight; | |
2788 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
2789 | } | |
2790 | ||
2791 | tg->cfs_rq[cpu]->h_load = load; | |
2792 | ||
2793 | return 0; | |
2794 | } | |
2795 | ||
2796 | static void update_h_load(long cpu) | |
2797 | { | |
2798 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); | |
2799 | } | |
2800 | ||
230059de PZ |
2801 | static unsigned long |
2802 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2803 | unsigned long max_load_move, | |
2804 | struct sched_domain *sd, enum cpu_idle_type idle, | |
931aeeda | 2805 | int *all_pinned) |
230059de PZ |
2806 | { |
2807 | long rem_load_move = max_load_move; | |
9763b67f | 2808 | struct cfs_rq *busiest_cfs_rq; |
230059de PZ |
2809 | |
2810 | rcu_read_lock(); | |
9763b67f | 2811 | update_h_load(cpu_of(busiest)); |
230059de | 2812 | |
9763b67f | 2813 | for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) { |
230059de PZ |
2814 | unsigned long busiest_h_load = busiest_cfs_rq->h_load; |
2815 | unsigned long busiest_weight = busiest_cfs_rq->load.weight; | |
2816 | u64 rem_load, moved_load; | |
2817 | ||
2818 | /* | |
64660c86 | 2819 | * empty group or part of a throttled hierarchy |
230059de | 2820 | */ |
64660c86 PT |
2821 | if (!busiest_cfs_rq->task_weight || |
2822 | throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu)) | |
230059de PZ |
2823 | continue; |
2824 | ||
2825 | rem_load = (u64)rem_load_move * busiest_weight; | |
2826 | rem_load = div_u64(rem_load, busiest_h_load + 1); | |
2827 | ||
2828 | moved_load = balance_tasks(this_rq, this_cpu, busiest, | |
931aeeda | 2829 | rem_load, sd, idle, all_pinned, |
230059de PZ |
2830 | busiest_cfs_rq); |
2831 | ||
2832 | if (!moved_load) | |
2833 | continue; | |
2834 | ||
2835 | moved_load *= busiest_h_load; | |
2836 | moved_load = div_u64(moved_load, busiest_weight + 1); | |
2837 | ||
2838 | rem_load_move -= moved_load; | |
2839 | if (rem_load_move < 0) | |
2840 | break; | |
2841 | } | |
2842 | rcu_read_unlock(); | |
2843 | ||
2844 | return max_load_move - rem_load_move; | |
2845 | } | |
2846 | #else | |
9e3081ca PZ |
2847 | static inline void update_shares(int cpu) |
2848 | { | |
2849 | } | |
2850 | ||
230059de PZ |
2851 | static unsigned long |
2852 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2853 | unsigned long max_load_move, | |
2854 | struct sched_domain *sd, enum cpu_idle_type idle, | |
931aeeda | 2855 | int *all_pinned) |
230059de PZ |
2856 | { |
2857 | return balance_tasks(this_rq, this_cpu, busiest, | |
2858 | max_load_move, sd, idle, all_pinned, | |
931aeeda | 2859 | &busiest->cfs); |
230059de PZ |
2860 | } |
2861 | #endif | |
2862 | ||
1e3c88bd PZ |
2863 | /* |
2864 | * move_tasks tries to move up to max_load_move weighted load from busiest to | |
2865 | * this_rq, as part of a balancing operation within domain "sd". | |
2866 | * Returns 1 if successful and 0 otherwise. | |
2867 | * | |
2868 | * Called with both runqueues locked. | |
2869 | */ | |
2870 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
2871 | unsigned long max_load_move, | |
2872 | struct sched_domain *sd, enum cpu_idle_type idle, | |
2873 | int *all_pinned) | |
2874 | { | |
3d45fd80 | 2875 | unsigned long total_load_moved = 0, load_moved; |
1e3c88bd PZ |
2876 | |
2877 | do { | |
3d45fd80 | 2878 | load_moved = load_balance_fair(this_rq, this_cpu, busiest, |
1e3c88bd | 2879 | max_load_move - total_load_moved, |
931aeeda | 2880 | sd, idle, all_pinned); |
3d45fd80 PZ |
2881 | |
2882 | total_load_moved += load_moved; | |
1e3c88bd PZ |
2883 | |
2884 | #ifdef CONFIG_PREEMPT | |
2885 | /* | |
2886 | * NEWIDLE balancing is a source of latency, so preemptible | |
2887 | * kernels will stop after the first task is pulled to minimize | |
2888 | * the critical section. | |
2889 | */ | |
2890 | if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) | |
2891 | break; | |
baa8c110 PZ |
2892 | |
2893 | if (raw_spin_is_contended(&this_rq->lock) || | |
2894 | raw_spin_is_contended(&busiest->lock)) | |
2895 | break; | |
1e3c88bd | 2896 | #endif |
3d45fd80 | 2897 | } while (load_moved && max_load_move > total_load_moved); |
1e3c88bd PZ |
2898 | |
2899 | return total_load_moved > 0; | |
2900 | } | |
2901 | ||
1e3c88bd PZ |
2902 | /********** Helpers for find_busiest_group ************************/ |
2903 | /* | |
2904 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
2905 | * during load balancing. | |
2906 | */ | |
2907 | struct sd_lb_stats { | |
2908 | struct sched_group *busiest; /* Busiest group in this sd */ | |
2909 | struct sched_group *this; /* Local group in this sd */ | |
2910 | unsigned long total_load; /* Total load of all groups in sd */ | |
2911 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
2912 | unsigned long avg_load; /* Average load across all groups in sd */ | |
2913 | ||
2914 | /** Statistics of this group */ | |
2915 | unsigned long this_load; | |
2916 | unsigned long this_load_per_task; | |
2917 | unsigned long this_nr_running; | |
fab47622 | 2918 | unsigned long this_has_capacity; |
aae6d3dd | 2919 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
2920 | |
2921 | /* Statistics of the busiest group */ | |
aae6d3dd | 2922 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
2923 | unsigned long max_load; |
2924 | unsigned long busiest_load_per_task; | |
2925 | unsigned long busiest_nr_running; | |
dd5feea1 | 2926 | unsigned long busiest_group_capacity; |
fab47622 | 2927 | unsigned long busiest_has_capacity; |
aae6d3dd | 2928 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
2929 | |
2930 | int group_imb; /* Is there imbalance in this sd */ | |
2931 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2932 | int power_savings_balance; /* Is powersave balance needed for this sd */ | |
2933 | struct sched_group *group_min; /* Least loaded group in sd */ | |
2934 | struct sched_group *group_leader; /* Group which relieves group_min */ | |
2935 | unsigned long min_load_per_task; /* load_per_task in group_min */ | |
2936 | unsigned long leader_nr_running; /* Nr running of group_leader */ | |
2937 | unsigned long min_nr_running; /* Nr running of group_min */ | |
2938 | #endif | |
2939 | }; | |
2940 | ||
2941 | /* | |
2942 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
2943 | */ | |
2944 | struct sg_lb_stats { | |
2945 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
2946 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
2947 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
2948 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
2949 | unsigned long group_capacity; | |
aae6d3dd SS |
2950 | unsigned long idle_cpus; |
2951 | unsigned long group_weight; | |
1e3c88bd | 2952 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 2953 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
2954 | }; |
2955 | ||
2956 | /** | |
2957 | * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. | |
2958 | * @group: The group whose first cpu is to be returned. | |
2959 | */ | |
2960 | static inline unsigned int group_first_cpu(struct sched_group *group) | |
2961 | { | |
2962 | return cpumask_first(sched_group_cpus(group)); | |
2963 | } | |
2964 | ||
2965 | /** | |
2966 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
2967 | * @sd: The sched_domain whose load_idx is to be obtained. | |
2968 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
2969 | */ | |
2970 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
2971 | enum cpu_idle_type idle) | |
2972 | { | |
2973 | int load_idx; | |
2974 | ||
2975 | switch (idle) { | |
2976 | case CPU_NOT_IDLE: | |
2977 | load_idx = sd->busy_idx; | |
2978 | break; | |
2979 | ||
2980 | case CPU_NEWLY_IDLE: | |
2981 | load_idx = sd->newidle_idx; | |
2982 | break; | |
2983 | default: | |
2984 | load_idx = sd->idle_idx; | |
2985 | break; | |
2986 | } | |
2987 | ||
2988 | return load_idx; | |
2989 | } | |
2990 | ||
2991 | ||
2992 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
2993 | /** | |
2994 | * init_sd_power_savings_stats - Initialize power savings statistics for | |
2995 | * the given sched_domain, during load balancing. | |
2996 | * | |
2997 | * @sd: Sched domain whose power-savings statistics are to be initialized. | |
2998 | * @sds: Variable containing the statistics for sd. | |
2999 | * @idle: Idle status of the CPU at which we're performing load-balancing. | |
3000 | */ | |
3001 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
3002 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
3003 | { | |
3004 | /* | |
3005 | * Busy processors will not participate in power savings | |
3006 | * balance. | |
3007 | */ | |
3008 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3009 | sds->power_savings_balance = 0; | |
3010 | else { | |
3011 | sds->power_savings_balance = 1; | |
3012 | sds->min_nr_running = ULONG_MAX; | |
3013 | sds->leader_nr_running = 0; | |
3014 | } | |
3015 | } | |
3016 | ||
3017 | /** | |
3018 | * update_sd_power_savings_stats - Update the power saving stats for a | |
3019 | * sched_domain while performing load balancing. | |
3020 | * | |
3021 | * @group: sched_group belonging to the sched_domain under consideration. | |
3022 | * @sds: Variable containing the statistics of the sched_domain | |
3023 | * @local_group: Does group contain the CPU for which we're performing | |
3024 | * load balancing ? | |
3025 | * @sgs: Variable containing the statistics of the group. | |
3026 | */ | |
3027 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
3028 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
3029 | { | |
3030 | ||
3031 | if (!sds->power_savings_balance) | |
3032 | return; | |
3033 | ||
3034 | /* | |
3035 | * If the local group is idle or completely loaded | |
3036 | * no need to do power savings balance at this domain | |
3037 | */ | |
3038 | if (local_group && (sds->this_nr_running >= sgs->group_capacity || | |
3039 | !sds->this_nr_running)) | |
3040 | sds->power_savings_balance = 0; | |
3041 | ||
3042 | /* | |
3043 | * If a group is already running at full capacity or idle, | |
3044 | * don't include that group in power savings calculations | |
3045 | */ | |
3046 | if (!sds->power_savings_balance || | |
3047 | sgs->sum_nr_running >= sgs->group_capacity || | |
3048 | !sgs->sum_nr_running) | |
3049 | return; | |
3050 | ||
3051 | /* | |
3052 | * Calculate the group which has the least non-idle load. | |
3053 | * This is the group from where we need to pick up the load | |
3054 | * for saving power | |
3055 | */ | |
3056 | if ((sgs->sum_nr_running < sds->min_nr_running) || | |
3057 | (sgs->sum_nr_running == sds->min_nr_running && | |
3058 | group_first_cpu(group) > group_first_cpu(sds->group_min))) { | |
3059 | sds->group_min = group; | |
3060 | sds->min_nr_running = sgs->sum_nr_running; | |
3061 | sds->min_load_per_task = sgs->sum_weighted_load / | |
3062 | sgs->sum_nr_running; | |
3063 | } | |
3064 | ||
3065 | /* | |
3066 | * Calculate the group which is almost near its | |
3067 | * capacity but still has some space to pick up some load | |
3068 | * from other group and save more power | |
3069 | */ | |
3070 | if (sgs->sum_nr_running + 1 > sgs->group_capacity) | |
3071 | return; | |
3072 | ||
3073 | if (sgs->sum_nr_running > sds->leader_nr_running || | |
3074 | (sgs->sum_nr_running == sds->leader_nr_running && | |
3075 | group_first_cpu(group) < group_first_cpu(sds->group_leader))) { | |
3076 | sds->group_leader = group; | |
3077 | sds->leader_nr_running = sgs->sum_nr_running; | |
3078 | } | |
3079 | } | |
3080 | ||
3081 | /** | |
3082 | * check_power_save_busiest_group - see if there is potential for some power-savings balance | |
3083 | * @sds: Variable containing the statistics of the sched_domain | |
3084 | * under consideration. | |
3085 | * @this_cpu: Cpu at which we're currently performing load-balancing. | |
3086 | * @imbalance: Variable to store the imbalance. | |
3087 | * | |
3088 | * Description: | |
3089 | * Check if we have potential to perform some power-savings balance. | |
3090 | * If yes, set the busiest group to be the least loaded group in the | |
3091 | * sched_domain, so that it's CPUs can be put to idle. | |
3092 | * | |
3093 | * Returns 1 if there is potential to perform power-savings balance. | |
3094 | * Else returns 0. | |
3095 | */ | |
3096 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
3097 | int this_cpu, unsigned long *imbalance) | |
3098 | { | |
3099 | if (!sds->power_savings_balance) | |
3100 | return 0; | |
3101 | ||
3102 | if (sds->this != sds->group_leader || | |
3103 | sds->group_leader == sds->group_min) | |
3104 | return 0; | |
3105 | ||
3106 | *imbalance = sds->min_load_per_task; | |
3107 | sds->busiest = sds->group_min; | |
3108 | ||
3109 | return 1; | |
3110 | ||
3111 | } | |
3112 | #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
3113 | static inline void init_sd_power_savings_stats(struct sched_domain *sd, | |
3114 | struct sd_lb_stats *sds, enum cpu_idle_type idle) | |
3115 | { | |
3116 | return; | |
3117 | } | |
3118 | ||
3119 | static inline void update_sd_power_savings_stats(struct sched_group *group, | |
3120 | struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) | |
3121 | { | |
3122 | return; | |
3123 | } | |
3124 | ||
3125 | static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, | |
3126 | int this_cpu, unsigned long *imbalance) | |
3127 | { | |
3128 | return 0; | |
3129 | } | |
3130 | #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ | |
3131 | ||
3132 | ||
3133 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) | |
3134 | { | |
1399fa78 | 3135 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
3136 | } |
3137 | ||
3138 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
3139 | { | |
3140 | return default_scale_freq_power(sd, cpu); | |
3141 | } | |
3142 | ||
3143 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
3144 | { | |
669c55e9 | 3145 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
3146 | unsigned long smt_gain = sd->smt_gain; |
3147 | ||
3148 | smt_gain /= weight; | |
3149 | ||
3150 | return smt_gain; | |
3151 | } | |
3152 | ||
3153 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
3154 | { | |
3155 | return default_scale_smt_power(sd, cpu); | |
3156 | } | |
3157 | ||
3158 | unsigned long scale_rt_power(int cpu) | |
3159 | { | |
3160 | struct rq *rq = cpu_rq(cpu); | |
3161 | u64 total, available; | |
3162 | ||
1e3c88bd | 3163 | total = sched_avg_period() + (rq->clock - rq->age_stamp); |
aa483808 VP |
3164 | |
3165 | if (unlikely(total < rq->rt_avg)) { | |
3166 | /* Ensures that power won't end up being negative */ | |
3167 | available = 0; | |
3168 | } else { | |
3169 | available = total - rq->rt_avg; | |
3170 | } | |
1e3c88bd | 3171 | |
1399fa78 NR |
3172 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
3173 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 3174 | |
1399fa78 | 3175 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3176 | |
3177 | return div_u64(available, total); | |
3178 | } | |
3179 | ||
3180 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
3181 | { | |
669c55e9 | 3182 | unsigned long weight = sd->span_weight; |
1399fa78 | 3183 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
3184 | struct sched_group *sdg = sd->groups; |
3185 | ||
1e3c88bd PZ |
3186 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
3187 | if (sched_feat(ARCH_POWER)) | |
3188 | power *= arch_scale_smt_power(sd, cpu); | |
3189 | else | |
3190 | power *= default_scale_smt_power(sd, cpu); | |
3191 | ||
1399fa78 | 3192 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3193 | } |
3194 | ||
9c3f75cb | 3195 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
3196 | |
3197 | if (sched_feat(ARCH_POWER)) | |
3198 | power *= arch_scale_freq_power(sd, cpu); | |
3199 | else | |
3200 | power *= default_scale_freq_power(sd, cpu); | |
3201 | ||
1399fa78 | 3202 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 3203 | |
1e3c88bd | 3204 | power *= scale_rt_power(cpu); |
1399fa78 | 3205 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3206 | |
3207 | if (!power) | |
3208 | power = 1; | |
3209 | ||
e51fd5e2 | 3210 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 3211 | sdg->sgp->power = power; |
1e3c88bd PZ |
3212 | } |
3213 | ||
3214 | static void update_group_power(struct sched_domain *sd, int cpu) | |
3215 | { | |
3216 | struct sched_domain *child = sd->child; | |
3217 | struct sched_group *group, *sdg = sd->groups; | |
3218 | unsigned long power; | |
3219 | ||
3220 | if (!child) { | |
3221 | update_cpu_power(sd, cpu); | |
3222 | return; | |
3223 | } | |
3224 | ||
3225 | power = 0; | |
3226 | ||
3227 | group = child->groups; | |
3228 | do { | |
9c3f75cb | 3229 | power += group->sgp->power; |
1e3c88bd PZ |
3230 | group = group->next; |
3231 | } while (group != child->groups); | |
3232 | ||
9c3f75cb | 3233 | sdg->sgp->power = power; |
1e3c88bd PZ |
3234 | } |
3235 | ||
9d5efe05 SV |
3236 | /* |
3237 | * Try and fix up capacity for tiny siblings, this is needed when | |
3238 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
3239 | * which on its own isn't powerful enough. | |
3240 | * | |
3241 | * See update_sd_pick_busiest() and check_asym_packing(). | |
3242 | */ | |
3243 | static inline int | |
3244 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
3245 | { | |
3246 | /* | |
1399fa78 | 3247 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 3248 | */ |
a6c75f2f | 3249 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
3250 | return 0; |
3251 | ||
3252 | /* | |
3253 | * If ~90% of the cpu_power is still there, we're good. | |
3254 | */ | |
9c3f75cb | 3255 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
3256 | return 1; |
3257 | ||
3258 | return 0; | |
3259 | } | |
3260 | ||
1e3c88bd PZ |
3261 | /** |
3262 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
3263 | * @sd: The sched_domain whose statistics are to be updated. | |
3264 | * @group: sched_group whose statistics are to be updated. | |
3265 | * @this_cpu: Cpu for which load balance is currently performed. | |
3266 | * @idle: Idle status of this_cpu | |
3267 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | |
1e3c88bd PZ |
3268 | * @local_group: Does group contain this_cpu. |
3269 | * @cpus: Set of cpus considered for load balancing. | |
3270 | * @balance: Should we balance. | |
3271 | * @sgs: variable to hold the statistics for this group. | |
3272 | */ | |
3273 | static inline void update_sg_lb_stats(struct sched_domain *sd, | |
3274 | struct sched_group *group, int this_cpu, | |
46e49b38 | 3275 | enum cpu_idle_type idle, int load_idx, |
1e3c88bd PZ |
3276 | int local_group, const struct cpumask *cpus, |
3277 | int *balance, struct sg_lb_stats *sgs) | |
3278 | { | |
2582f0eb | 3279 | unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; |
1e3c88bd PZ |
3280 | int i; |
3281 | unsigned int balance_cpu = -1, first_idle_cpu = 0; | |
dd5feea1 | 3282 | unsigned long avg_load_per_task = 0; |
1e3c88bd | 3283 | |
871e35bc | 3284 | if (local_group) |
1e3c88bd | 3285 | balance_cpu = group_first_cpu(group); |
1e3c88bd PZ |
3286 | |
3287 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
3288 | max_cpu_load = 0; |
3289 | min_cpu_load = ~0UL; | |
2582f0eb | 3290 | max_nr_running = 0; |
1e3c88bd PZ |
3291 | |
3292 | for_each_cpu_and(i, sched_group_cpus(group), cpus) { | |
3293 | struct rq *rq = cpu_rq(i); | |
3294 | ||
1e3c88bd PZ |
3295 | /* Bias balancing toward cpus of our domain */ |
3296 | if (local_group) { | |
3297 | if (idle_cpu(i) && !first_idle_cpu) { | |
3298 | first_idle_cpu = 1; | |
3299 | balance_cpu = i; | |
3300 | } | |
3301 | ||
3302 | load = target_load(i, load_idx); | |
3303 | } else { | |
3304 | load = source_load(i, load_idx); | |
2582f0eb | 3305 | if (load > max_cpu_load) { |
1e3c88bd | 3306 | max_cpu_load = load; |
2582f0eb NR |
3307 | max_nr_running = rq->nr_running; |
3308 | } | |
1e3c88bd PZ |
3309 | if (min_cpu_load > load) |
3310 | min_cpu_load = load; | |
3311 | } | |
3312 | ||
3313 | sgs->group_load += load; | |
3314 | sgs->sum_nr_running += rq->nr_running; | |
3315 | sgs->sum_weighted_load += weighted_cpuload(i); | |
aae6d3dd SS |
3316 | if (idle_cpu(i)) |
3317 | sgs->idle_cpus++; | |
1e3c88bd PZ |
3318 | } |
3319 | ||
3320 | /* | |
3321 | * First idle cpu or the first cpu(busiest) in this sched group | |
3322 | * is eligible for doing load balancing at this and above | |
3323 | * domains. In the newly idle case, we will allow all the cpu's | |
3324 | * to do the newly idle load balance. | |
3325 | */ | |
bbc8cb5b PZ |
3326 | if (idle != CPU_NEWLY_IDLE && local_group) { |
3327 | if (balance_cpu != this_cpu) { | |
3328 | *balance = 0; | |
3329 | return; | |
3330 | } | |
3331 | update_group_power(sd, this_cpu); | |
1e3c88bd PZ |
3332 | } |
3333 | ||
3334 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3335 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 3336 | |
1e3c88bd PZ |
3337 | /* |
3338 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 3339 | * than the average weight of a task. |
1e3c88bd PZ |
3340 | * |
3341 | * APZ: with cgroup the avg task weight can vary wildly and | |
3342 | * might not be a suitable number - should we keep a | |
3343 | * normalized nr_running number somewhere that negates | |
3344 | * the hierarchy? | |
3345 | */ | |
dd5feea1 SS |
3346 | if (sgs->sum_nr_running) |
3347 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 3348 | |
866ab43e | 3349 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1) |
1e3c88bd PZ |
3350 | sgs->group_imb = 1; |
3351 | ||
9c3f75cb | 3352 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 3353 | SCHED_POWER_SCALE); |
9d5efe05 SV |
3354 | if (!sgs->group_capacity) |
3355 | sgs->group_capacity = fix_small_capacity(sd, group); | |
aae6d3dd | 3356 | sgs->group_weight = group->group_weight; |
fab47622 NR |
3357 | |
3358 | if (sgs->group_capacity > sgs->sum_nr_running) | |
3359 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
3360 | } |
3361 | ||
532cb4c4 MN |
3362 | /** |
3363 | * update_sd_pick_busiest - return 1 on busiest group | |
3364 | * @sd: sched_domain whose statistics are to be checked | |
3365 | * @sds: sched_domain statistics | |
3366 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 MN |
3367 | * @sgs: sched_group statistics |
3368 | * @this_cpu: the current cpu | |
532cb4c4 MN |
3369 | * |
3370 | * Determine if @sg is a busier group than the previously selected | |
3371 | * busiest group. | |
3372 | */ | |
3373 | static bool update_sd_pick_busiest(struct sched_domain *sd, | |
3374 | struct sd_lb_stats *sds, | |
3375 | struct sched_group *sg, | |
3376 | struct sg_lb_stats *sgs, | |
3377 | int this_cpu) | |
3378 | { | |
3379 | if (sgs->avg_load <= sds->max_load) | |
3380 | return false; | |
3381 | ||
3382 | if (sgs->sum_nr_running > sgs->group_capacity) | |
3383 | return true; | |
3384 | ||
3385 | if (sgs->group_imb) | |
3386 | return true; | |
3387 | ||
3388 | /* | |
3389 | * ASYM_PACKING needs to move all the work to the lowest | |
3390 | * numbered CPUs in the group, therefore mark all groups | |
3391 | * higher than ourself as busy. | |
3392 | */ | |
3393 | if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && | |
3394 | this_cpu < group_first_cpu(sg)) { | |
3395 | if (!sds->busiest) | |
3396 | return true; | |
3397 | ||
3398 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
3399 | return true; | |
3400 | } | |
3401 | ||
3402 | return false; | |
3403 | } | |
3404 | ||
1e3c88bd PZ |
3405 | /** |
3406 | * update_sd_lb_stats - Update sched_group's statistics for load balancing. | |
3407 | * @sd: sched_domain whose statistics are to be updated. | |
3408 | * @this_cpu: Cpu for which load balance is currently performed. | |
3409 | * @idle: Idle status of this_cpu | |
1e3c88bd PZ |
3410 | * @cpus: Set of cpus considered for load balancing. |
3411 | * @balance: Should we balance. | |
3412 | * @sds: variable to hold the statistics for this sched_domain. | |
3413 | */ | |
3414 | static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, | |
46e49b38 VP |
3415 | enum cpu_idle_type idle, const struct cpumask *cpus, |
3416 | int *balance, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
3417 | { |
3418 | struct sched_domain *child = sd->child; | |
532cb4c4 | 3419 | struct sched_group *sg = sd->groups; |
1e3c88bd PZ |
3420 | struct sg_lb_stats sgs; |
3421 | int load_idx, prefer_sibling = 0; | |
3422 | ||
3423 | if (child && child->flags & SD_PREFER_SIBLING) | |
3424 | prefer_sibling = 1; | |
3425 | ||
3426 | init_sd_power_savings_stats(sd, sds, idle); | |
3427 | load_idx = get_sd_load_idx(sd, idle); | |
3428 | ||
3429 | do { | |
3430 | int local_group; | |
3431 | ||
532cb4c4 | 3432 | local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); |
1e3c88bd | 3433 | memset(&sgs, 0, sizeof(sgs)); |
46e49b38 | 3434 | update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, |
1e3c88bd PZ |
3435 | local_group, cpus, balance, &sgs); |
3436 | ||
8f190fb3 | 3437 | if (local_group && !(*balance)) |
1e3c88bd PZ |
3438 | return; |
3439 | ||
3440 | sds->total_load += sgs.group_load; | |
9c3f75cb | 3441 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
3442 | |
3443 | /* | |
3444 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 3445 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
3446 | * and move all the excess tasks away. We lower the capacity |
3447 | * of a group only if the local group has the capacity to fit | |
3448 | * these excess tasks, i.e. nr_running < group_capacity. The | |
3449 | * extra check prevents the case where you always pull from the | |
3450 | * heaviest group when it is already under-utilized (possible | |
3451 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 3452 | */ |
75dd321d | 3453 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
3454 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
3455 | ||
3456 | if (local_group) { | |
3457 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 3458 | sds->this = sg; |
1e3c88bd PZ |
3459 | sds->this_nr_running = sgs.sum_nr_running; |
3460 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 3461 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 3462 | sds->this_idle_cpus = sgs.idle_cpus; |
532cb4c4 | 3463 | } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { |
1e3c88bd | 3464 | sds->max_load = sgs.avg_load; |
532cb4c4 | 3465 | sds->busiest = sg; |
1e3c88bd | 3466 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 3467 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 3468 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 3469 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 3470 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 3471 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
3472 | sds->group_imb = sgs.group_imb; |
3473 | } | |
3474 | ||
532cb4c4 MN |
3475 | update_sd_power_savings_stats(sg, sds, local_group, &sgs); |
3476 | sg = sg->next; | |
3477 | } while (sg != sd->groups); | |
3478 | } | |
3479 | ||
2ec57d44 | 3480 | int __weak arch_sd_sibling_asym_packing(void) |
532cb4c4 MN |
3481 | { |
3482 | return 0*SD_ASYM_PACKING; | |
3483 | } | |
3484 | ||
3485 | /** | |
3486 | * check_asym_packing - Check to see if the group is packed into the | |
3487 | * sched doman. | |
3488 | * | |
3489 | * This is primarily intended to used at the sibling level. Some | |
3490 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
3491 | * case of POWER7, it can move to lower SMT modes only when higher | |
3492 | * threads are idle. When in lower SMT modes, the threads will | |
3493 | * perform better since they share less core resources. Hence when we | |
3494 | * have idle threads, we want them to be the higher ones. | |
3495 | * | |
3496 | * This packing function is run on idle threads. It checks to see if | |
3497 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
3498 | * CPU number than the packing function is being run on. Here we are | |
3499 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
3500 | * number. | |
3501 | * | |
b6b12294 MN |
3502 | * Returns 1 when packing is required and a task should be moved to |
3503 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
3504 | * | |
532cb4c4 MN |
3505 | * @sd: The sched_domain whose packing is to be checked. |
3506 | * @sds: Statistics of the sched_domain which is to be packed | |
3507 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
3508 | * @imbalance: returns amount of imbalanced due to packing. | |
532cb4c4 MN |
3509 | */ |
3510 | static int check_asym_packing(struct sched_domain *sd, | |
3511 | struct sd_lb_stats *sds, | |
3512 | int this_cpu, unsigned long *imbalance) | |
3513 | { | |
3514 | int busiest_cpu; | |
3515 | ||
3516 | if (!(sd->flags & SD_ASYM_PACKING)) | |
3517 | return 0; | |
3518 | ||
3519 | if (!sds->busiest) | |
3520 | return 0; | |
3521 | ||
3522 | busiest_cpu = group_first_cpu(sds->busiest); | |
3523 | if (this_cpu > busiest_cpu) | |
3524 | return 0; | |
3525 | ||
9c3f75cb | 3526 | *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power, |
1399fa78 | 3527 | SCHED_POWER_SCALE); |
532cb4c4 | 3528 | return 1; |
1e3c88bd PZ |
3529 | } |
3530 | ||
3531 | /** | |
3532 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
3533 | * amongst the groups of a sched_domain, during | |
3534 | * load balancing. | |
3535 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | |
3536 | * @this_cpu: The cpu at whose sched_domain we're performing load-balance. | |
3537 | * @imbalance: Variable to store the imbalance. | |
3538 | */ | |
3539 | static inline void fix_small_imbalance(struct sd_lb_stats *sds, | |
3540 | int this_cpu, unsigned long *imbalance) | |
3541 | { | |
3542 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
3543 | unsigned int imbn = 2; | |
dd5feea1 | 3544 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
3545 | |
3546 | if (sds->this_nr_running) { | |
3547 | sds->this_load_per_task /= sds->this_nr_running; | |
3548 | if (sds->busiest_load_per_task > | |
3549 | sds->this_load_per_task) | |
3550 | imbn = 1; | |
3551 | } else | |
3552 | sds->this_load_per_task = | |
3553 | cpu_avg_load_per_task(this_cpu); | |
3554 | ||
dd5feea1 | 3555 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 3556 | * SCHED_POWER_SCALE; |
9c3f75cb | 3557 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
3558 | |
3559 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
3560 | (scaled_busy_load_per_task * imbn)) { | |
1e3c88bd PZ |
3561 | *imbalance = sds->busiest_load_per_task; |
3562 | return; | |
3563 | } | |
3564 | ||
3565 | /* | |
3566 | * OK, we don't have enough imbalance to justify moving tasks, | |
3567 | * however we may be able to increase total CPU power used by | |
3568 | * moving them. | |
3569 | */ | |
3570 | ||
9c3f75cb | 3571 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 3572 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 3573 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 3574 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 3575 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
3576 | |
3577 | /* Amount of load we'd subtract */ | |
1399fa78 | 3578 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 3579 | sds->busiest->sgp->power; |
1e3c88bd | 3580 | if (sds->max_load > tmp) |
9c3f75cb | 3581 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
3582 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
3583 | ||
3584 | /* Amount of load we'd add */ | |
9c3f75cb | 3585 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 3586 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
3587 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
3588 | sds->this->sgp->power; | |
1e3c88bd | 3589 | else |
1399fa78 | 3590 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
3591 | sds->this->sgp->power; |
3592 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 3593 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 3594 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
3595 | |
3596 | /* Move if we gain throughput */ | |
3597 | if (pwr_move > pwr_now) | |
3598 | *imbalance = sds->busiest_load_per_task; | |
3599 | } | |
3600 | ||
3601 | /** | |
3602 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
3603 | * groups of a given sched_domain during load balance. | |
3604 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | |
3605 | * @this_cpu: Cpu for which currently load balance is being performed. | |
3606 | * @imbalance: The variable to store the imbalance. | |
3607 | */ | |
3608 | static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, | |
3609 | unsigned long *imbalance) | |
3610 | { | |
dd5feea1 SS |
3611 | unsigned long max_pull, load_above_capacity = ~0UL; |
3612 | ||
3613 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
3614 | if (sds->group_imb) { | |
3615 | sds->busiest_load_per_task = | |
3616 | min(sds->busiest_load_per_task, sds->avg_load); | |
3617 | } | |
3618 | ||
1e3c88bd PZ |
3619 | /* |
3620 | * In the presence of smp nice balancing, certain scenarios can have | |
3621 | * max load less than avg load(as we skip the groups at or below | |
3622 | * its cpu_power, while calculating max_load..) | |
3623 | */ | |
3624 | if (sds->max_load < sds->avg_load) { | |
3625 | *imbalance = 0; | |
3626 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3627 | } | |
3628 | ||
dd5feea1 SS |
3629 | if (!sds->group_imb) { |
3630 | /* | |
3631 | * Don't want to pull so many tasks that a group would go idle. | |
3632 | */ | |
3633 | load_above_capacity = (sds->busiest_nr_running - | |
3634 | sds->busiest_group_capacity); | |
3635 | ||
1399fa78 | 3636 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 3637 | |
9c3f75cb | 3638 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
3639 | } |
3640 | ||
3641 | /* | |
3642 | * We're trying to get all the cpus to the average_load, so we don't | |
3643 | * want to push ourselves above the average load, nor do we wish to | |
3644 | * reduce the max loaded cpu below the average load. At the same time, | |
3645 | * we also don't want to reduce the group load below the group capacity | |
3646 | * (so that we can implement power-savings policies etc). Thus we look | |
3647 | * for the minimum possible imbalance. | |
3648 | * Be careful of negative numbers as they'll appear as very large values | |
3649 | * with unsigned longs. | |
3650 | */ | |
3651 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
3652 | |
3653 | /* How much load to actually move to equalise the imbalance */ | |
9c3f75cb PZ |
3654 | *imbalance = min(max_pull * sds->busiest->sgp->power, |
3655 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) | |
1399fa78 | 3656 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
3657 | |
3658 | /* | |
3659 | * if *imbalance is less than the average load per runnable task | |
25985edc | 3660 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
3661 | * a think about bumping its value to force at least one task to be |
3662 | * moved | |
3663 | */ | |
3664 | if (*imbalance < sds->busiest_load_per_task) | |
3665 | return fix_small_imbalance(sds, this_cpu, imbalance); | |
3666 | ||
3667 | } | |
fab47622 | 3668 | |
1e3c88bd PZ |
3669 | /******* find_busiest_group() helpers end here *********************/ |
3670 | ||
3671 | /** | |
3672 | * find_busiest_group - Returns the busiest group within the sched_domain | |
3673 | * if there is an imbalance. If there isn't an imbalance, and | |
3674 | * the user has opted for power-savings, it returns a group whose | |
3675 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
3676 | * such a group exists. | |
3677 | * | |
3678 | * Also calculates the amount of weighted load which should be moved | |
3679 | * to restore balance. | |
3680 | * | |
3681 | * @sd: The sched_domain whose busiest group is to be returned. | |
3682 | * @this_cpu: The cpu for which load balancing is currently being performed. | |
3683 | * @imbalance: Variable which stores amount of weighted load which should | |
3684 | * be moved to restore balance/put a group to idle. | |
3685 | * @idle: The idle status of this_cpu. | |
1e3c88bd PZ |
3686 | * @cpus: The set of CPUs under consideration for load-balancing. |
3687 | * @balance: Pointer to a variable indicating if this_cpu | |
3688 | * is the appropriate cpu to perform load balancing at this_level. | |
3689 | * | |
3690 | * Returns: - the busiest group if imbalance exists. | |
3691 | * - If no imbalance and user has opted for power-savings balance, | |
3692 | * return the least loaded group whose CPUs can be | |
3693 | * put to idle by rebalancing its tasks onto our group. | |
3694 | */ | |
3695 | static struct sched_group * | |
3696 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
3697 | unsigned long *imbalance, enum cpu_idle_type idle, | |
46e49b38 | 3698 | const struct cpumask *cpus, int *balance) |
1e3c88bd PZ |
3699 | { |
3700 | struct sd_lb_stats sds; | |
3701 | ||
3702 | memset(&sds, 0, sizeof(sds)); | |
3703 | ||
3704 | /* | |
3705 | * Compute the various statistics relavent for load balancing at | |
3706 | * this level. | |
3707 | */ | |
46e49b38 | 3708 | update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds); |
1e3c88bd | 3709 | |
cc57aa8f PZ |
3710 | /* |
3711 | * this_cpu is not the appropriate cpu to perform load balancing at | |
3712 | * this level. | |
1e3c88bd | 3713 | */ |
8f190fb3 | 3714 | if (!(*balance)) |
1e3c88bd PZ |
3715 | goto ret; |
3716 | ||
532cb4c4 MN |
3717 | if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && |
3718 | check_asym_packing(sd, &sds, this_cpu, imbalance)) | |
3719 | return sds.busiest; | |
3720 | ||
cc57aa8f | 3721 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
3722 | if (!sds.busiest || sds.busiest_nr_running == 0) |
3723 | goto out_balanced; | |
3724 | ||
1399fa78 | 3725 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 3726 | |
866ab43e PZ |
3727 | /* |
3728 | * If the busiest group is imbalanced the below checks don't | |
3729 | * work because they assumes all things are equal, which typically | |
3730 | * isn't true due to cpus_allowed constraints and the like. | |
3731 | */ | |
3732 | if (sds.group_imb) | |
3733 | goto force_balance; | |
3734 | ||
cc57aa8f | 3735 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
fab47622 NR |
3736 | if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
3737 | !sds.busiest_has_capacity) | |
3738 | goto force_balance; | |
3739 | ||
cc57aa8f PZ |
3740 | /* |
3741 | * If the local group is more busy than the selected busiest group | |
3742 | * don't try and pull any tasks. | |
3743 | */ | |
1e3c88bd PZ |
3744 | if (sds.this_load >= sds.max_load) |
3745 | goto out_balanced; | |
3746 | ||
cc57aa8f PZ |
3747 | /* |
3748 | * Don't pull any tasks if this group is already above the domain | |
3749 | * average load. | |
3750 | */ | |
1e3c88bd PZ |
3751 | if (sds.this_load >= sds.avg_load) |
3752 | goto out_balanced; | |
3753 | ||
c186fafe | 3754 | if (idle == CPU_IDLE) { |
aae6d3dd SS |
3755 | /* |
3756 | * This cpu is idle. If the busiest group load doesn't | |
3757 | * have more tasks than the number of available cpu's and | |
3758 | * there is no imbalance between this and busiest group | |
3759 | * wrt to idle cpu's, it is balanced. | |
3760 | */ | |
c186fafe | 3761 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
3762 | sds.busiest_nr_running <= sds.busiest_group_weight) |
3763 | goto out_balanced; | |
c186fafe PZ |
3764 | } else { |
3765 | /* | |
3766 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
3767 | * imbalance_pct to be conservative. | |
3768 | */ | |
3769 | if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) | |
3770 | goto out_balanced; | |
aae6d3dd | 3771 | } |
1e3c88bd | 3772 | |
fab47622 | 3773 | force_balance: |
1e3c88bd PZ |
3774 | /* Looks like there is an imbalance. Compute it */ |
3775 | calculate_imbalance(&sds, this_cpu, imbalance); | |
3776 | return sds.busiest; | |
3777 | ||
3778 | out_balanced: | |
3779 | /* | |
3780 | * There is no obvious imbalance. But check if we can do some balancing | |
3781 | * to save power. | |
3782 | */ | |
3783 | if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) | |
3784 | return sds.busiest; | |
3785 | ret: | |
3786 | *imbalance = 0; | |
3787 | return NULL; | |
3788 | } | |
3789 | ||
3790 | /* | |
3791 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3792 | */ | |
3793 | static struct rq * | |
9d5efe05 SV |
3794 | find_busiest_queue(struct sched_domain *sd, struct sched_group *group, |
3795 | enum cpu_idle_type idle, unsigned long imbalance, | |
3796 | const struct cpumask *cpus) | |
1e3c88bd PZ |
3797 | { |
3798 | struct rq *busiest = NULL, *rq; | |
3799 | unsigned long max_load = 0; | |
3800 | int i; | |
3801 | ||
3802 | for_each_cpu(i, sched_group_cpus(group)) { | |
3803 | unsigned long power = power_of(i); | |
1399fa78 NR |
3804 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
3805 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
3806 | unsigned long wl; |
3807 | ||
9d5efe05 SV |
3808 | if (!capacity) |
3809 | capacity = fix_small_capacity(sd, group); | |
3810 | ||
1e3c88bd PZ |
3811 | if (!cpumask_test_cpu(i, cpus)) |
3812 | continue; | |
3813 | ||
3814 | rq = cpu_rq(i); | |
6e40f5bb | 3815 | wl = weighted_cpuload(i); |
1e3c88bd | 3816 | |
6e40f5bb TG |
3817 | /* |
3818 | * When comparing with imbalance, use weighted_cpuload() | |
3819 | * which is not scaled with the cpu power. | |
3820 | */ | |
1e3c88bd PZ |
3821 | if (capacity && rq->nr_running == 1 && wl > imbalance) |
3822 | continue; | |
3823 | ||
6e40f5bb TG |
3824 | /* |
3825 | * For the load comparisons with the other cpu's, consider | |
3826 | * the weighted_cpuload() scaled with the cpu power, so that | |
3827 | * the load can be moved away from the cpu that is potentially | |
3828 | * running at a lower capacity. | |
3829 | */ | |
1399fa78 | 3830 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 3831 | |
1e3c88bd PZ |
3832 | if (wl > max_load) { |
3833 | max_load = wl; | |
3834 | busiest = rq; | |
3835 | } | |
3836 | } | |
3837 | ||
3838 | return busiest; | |
3839 | } | |
3840 | ||
3841 | /* | |
3842 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3843 | * so long as it is large enough. | |
3844 | */ | |
3845 | #define MAX_PINNED_INTERVAL 512 | |
3846 | ||
3847 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
3848 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); | |
3849 | ||
46e49b38 | 3850 | static int need_active_balance(struct sched_domain *sd, int idle, |
532cb4c4 | 3851 | int busiest_cpu, int this_cpu) |
1af3ed3d PZ |
3852 | { |
3853 | if (idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
3854 | |
3855 | /* | |
3856 | * ASYM_PACKING needs to force migrate tasks from busy but | |
3857 | * higher numbered CPUs in order to pack all tasks in the | |
3858 | * lowest numbered CPUs. | |
3859 | */ | |
3860 | if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) | |
3861 | return 1; | |
3862 | ||
1af3ed3d PZ |
3863 | /* |
3864 | * The only task running in a non-idle cpu can be moved to this | |
3865 | * cpu in an attempt to completely freeup the other CPU | |
3866 | * package. | |
3867 | * | |
3868 | * The package power saving logic comes from | |
3869 | * find_busiest_group(). If there are no imbalance, then | |
3870 | * f_b_g() will return NULL. However when sched_mc={1,2} then | |
3871 | * f_b_g() will select a group from which a running task may be | |
3872 | * pulled to this cpu in order to make the other package idle. | |
3873 | * If there is no opportunity to make a package idle and if | |
3874 | * there are no imbalance, then f_b_g() will return NULL and no | |
3875 | * action will be taken in load_balance_newidle(). | |
3876 | * | |
3877 | * Under normal task pull operation due to imbalance, there | |
3878 | * will be more than one task in the source run queue and | |
3879 | * move_tasks() will succeed. ld_moved will be true and this | |
3880 | * active balance code will not be triggered. | |
3881 | */ | |
1af3ed3d PZ |
3882 | if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) |
3883 | return 0; | |
3884 | } | |
3885 | ||
3886 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
3887 | } | |
3888 | ||
969c7921 TH |
3889 | static int active_load_balance_cpu_stop(void *data); |
3890 | ||
1e3c88bd PZ |
3891 | /* |
3892 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3893 | * tasks if there is an imbalance. | |
3894 | */ | |
3895 | static int load_balance(int this_cpu, struct rq *this_rq, | |
3896 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3897 | int *balance) | |
3898 | { | |
46e49b38 | 3899 | int ld_moved, all_pinned = 0, active_balance = 0; |
1e3c88bd PZ |
3900 | struct sched_group *group; |
3901 | unsigned long imbalance; | |
3902 | struct rq *busiest; | |
3903 | unsigned long flags; | |
3904 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
3905 | ||
3906 | cpumask_copy(cpus, cpu_active_mask); | |
3907 | ||
1e3c88bd PZ |
3908 | schedstat_inc(sd, lb_count[idle]); |
3909 | ||
3910 | redo: | |
46e49b38 | 3911 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, |
1e3c88bd PZ |
3912 | cpus, balance); |
3913 | ||
3914 | if (*balance == 0) | |
3915 | goto out_balanced; | |
3916 | ||
3917 | if (!group) { | |
3918 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3919 | goto out_balanced; | |
3920 | } | |
3921 | ||
9d5efe05 | 3922 | busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); |
1e3c88bd PZ |
3923 | if (!busiest) { |
3924 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3925 | goto out_balanced; | |
3926 | } | |
3927 | ||
3928 | BUG_ON(busiest == this_rq); | |
3929 | ||
3930 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3931 | ||
3932 | ld_moved = 0; | |
3933 | if (busiest->nr_running > 1) { | |
3934 | /* | |
3935 | * Attempt to move tasks. If find_busiest_group has found | |
3936 | * an imbalance but busiest->nr_running <= 1, the group is | |
3937 | * still unbalanced. ld_moved simply stays zero, so it is | |
3938 | * correctly treated as an imbalance. | |
3939 | */ | |
b30aef17 | 3940 | all_pinned = 1; |
1e3c88bd PZ |
3941 | local_irq_save(flags); |
3942 | double_rq_lock(this_rq, busiest); | |
3943 | ld_moved = move_tasks(this_rq, this_cpu, busiest, | |
3944 | imbalance, sd, idle, &all_pinned); | |
3945 | double_rq_unlock(this_rq, busiest); | |
3946 | local_irq_restore(flags); | |
3947 | ||
3948 | /* | |
3949 | * some other cpu did the load balance for us. | |
3950 | */ | |
3951 | if (ld_moved && this_cpu != smp_processor_id()) | |
3952 | resched_cpu(this_cpu); | |
3953 | ||
3954 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
3955 | if (unlikely(all_pinned)) { | |
3956 | cpumask_clear_cpu(cpu_of(busiest), cpus); | |
3957 | if (!cpumask_empty(cpus)) | |
3958 | goto redo; | |
3959 | goto out_balanced; | |
3960 | } | |
3961 | } | |
3962 | ||
3963 | if (!ld_moved) { | |
3964 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
3965 | /* |
3966 | * Increment the failure counter only on periodic balance. | |
3967 | * We do not want newidle balance, which can be very | |
3968 | * frequent, pollute the failure counter causing | |
3969 | * excessive cache_hot migrations and active balances. | |
3970 | */ | |
3971 | if (idle != CPU_NEWLY_IDLE) | |
3972 | sd->nr_balance_failed++; | |
1e3c88bd | 3973 | |
46e49b38 | 3974 | if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) { |
1e3c88bd PZ |
3975 | raw_spin_lock_irqsave(&busiest->lock, flags); |
3976 | ||
969c7921 TH |
3977 | /* don't kick the active_load_balance_cpu_stop, |
3978 | * if the curr task on busiest cpu can't be | |
3979 | * moved to this_cpu | |
1e3c88bd PZ |
3980 | */ |
3981 | if (!cpumask_test_cpu(this_cpu, | |
3982 | &busiest->curr->cpus_allowed)) { | |
3983 | raw_spin_unlock_irqrestore(&busiest->lock, | |
3984 | flags); | |
3985 | all_pinned = 1; | |
3986 | goto out_one_pinned; | |
3987 | } | |
3988 | ||
969c7921 TH |
3989 | /* |
3990 | * ->active_balance synchronizes accesses to | |
3991 | * ->active_balance_work. Once set, it's cleared | |
3992 | * only after active load balance is finished. | |
3993 | */ | |
1e3c88bd PZ |
3994 | if (!busiest->active_balance) { |
3995 | busiest->active_balance = 1; | |
3996 | busiest->push_cpu = this_cpu; | |
3997 | active_balance = 1; | |
3998 | } | |
3999 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 4000 | |
1e3c88bd | 4001 | if (active_balance) |
969c7921 TH |
4002 | stop_one_cpu_nowait(cpu_of(busiest), |
4003 | active_load_balance_cpu_stop, busiest, | |
4004 | &busiest->active_balance_work); | |
1e3c88bd PZ |
4005 | |
4006 | /* | |
4007 | * We've kicked active balancing, reset the failure | |
4008 | * counter. | |
4009 | */ | |
4010 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
4011 | } | |
4012 | } else | |
4013 | sd->nr_balance_failed = 0; | |
4014 | ||
4015 | if (likely(!active_balance)) { | |
4016 | /* We were unbalanced, so reset the balancing interval */ | |
4017 | sd->balance_interval = sd->min_interval; | |
4018 | } else { | |
4019 | /* | |
4020 | * If we've begun active balancing, start to back off. This | |
4021 | * case may not be covered by the all_pinned logic if there | |
4022 | * is only 1 task on the busy runqueue (because we don't call | |
4023 | * move_tasks). | |
4024 | */ | |
4025 | if (sd->balance_interval < sd->max_interval) | |
4026 | sd->balance_interval *= 2; | |
4027 | } | |
4028 | ||
1e3c88bd PZ |
4029 | goto out; |
4030 | ||
4031 | out_balanced: | |
4032 | schedstat_inc(sd, lb_balanced[idle]); | |
4033 | ||
4034 | sd->nr_balance_failed = 0; | |
4035 | ||
4036 | out_one_pinned: | |
4037 | /* tune up the balancing interval */ | |
4038 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || | |
4039 | (sd->balance_interval < sd->max_interval)) | |
4040 | sd->balance_interval *= 2; | |
4041 | ||
46e49b38 | 4042 | ld_moved = 0; |
1e3c88bd | 4043 | out: |
1e3c88bd PZ |
4044 | return ld_moved; |
4045 | } | |
4046 | ||
1e3c88bd PZ |
4047 | /* |
4048 | * idle_balance is called by schedule() if this_cpu is about to become | |
4049 | * idle. Attempts to pull tasks from other CPUs. | |
4050 | */ | |
4051 | static void idle_balance(int this_cpu, struct rq *this_rq) | |
4052 | { | |
4053 | struct sched_domain *sd; | |
4054 | int pulled_task = 0; | |
4055 | unsigned long next_balance = jiffies + HZ; | |
4056 | ||
4057 | this_rq->idle_stamp = this_rq->clock; | |
4058 | ||
4059 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
4060 | return; | |
4061 | ||
f492e12e PZ |
4062 | /* |
4063 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
4064 | */ | |
4065 | raw_spin_unlock(&this_rq->lock); | |
4066 | ||
c66eaf61 | 4067 | update_shares(this_cpu); |
dce840a0 | 4068 | rcu_read_lock(); |
1e3c88bd PZ |
4069 | for_each_domain(this_cpu, sd) { |
4070 | unsigned long interval; | |
f492e12e | 4071 | int balance = 1; |
1e3c88bd PZ |
4072 | |
4073 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4074 | continue; | |
4075 | ||
f492e12e | 4076 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 4077 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
4078 | pulled_task = load_balance(this_cpu, this_rq, |
4079 | sd, CPU_NEWLY_IDLE, &balance); | |
4080 | } | |
1e3c88bd PZ |
4081 | |
4082 | interval = msecs_to_jiffies(sd->balance_interval); | |
4083 | if (time_after(next_balance, sd->last_balance + interval)) | |
4084 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
4085 | if (pulled_task) { |
4086 | this_rq->idle_stamp = 0; | |
1e3c88bd | 4087 | break; |
d5ad140b | 4088 | } |
1e3c88bd | 4089 | } |
dce840a0 | 4090 | rcu_read_unlock(); |
f492e12e PZ |
4091 | |
4092 | raw_spin_lock(&this_rq->lock); | |
4093 | ||
1e3c88bd PZ |
4094 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
4095 | /* | |
4096 | * We are going idle. next_balance may be set based on | |
4097 | * a busy processor. So reset next_balance. | |
4098 | */ | |
4099 | this_rq->next_balance = next_balance; | |
4100 | } | |
4101 | } | |
4102 | ||
4103 | /* | |
969c7921 TH |
4104 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
4105 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
4106 | * least 1 task to be running on each physical CPU where possible, and | |
4107 | * avoids physical / logical imbalances. | |
1e3c88bd | 4108 | */ |
969c7921 | 4109 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 4110 | { |
969c7921 TH |
4111 | struct rq *busiest_rq = data; |
4112 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 4113 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 4114 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 4115 | struct sched_domain *sd; |
969c7921 TH |
4116 | |
4117 | raw_spin_lock_irq(&busiest_rq->lock); | |
4118 | ||
4119 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
4120 | if (unlikely(busiest_cpu != smp_processor_id() || | |
4121 | !busiest_rq->active_balance)) | |
4122 | goto out_unlock; | |
1e3c88bd PZ |
4123 | |
4124 | /* Is there any task to move? */ | |
4125 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 4126 | goto out_unlock; |
1e3c88bd PZ |
4127 | |
4128 | /* | |
4129 | * This condition is "impossible", if it occurs | |
4130 | * we need to fix it. Originally reported by | |
4131 | * Bjorn Helgaas on a 128-cpu setup. | |
4132 | */ | |
4133 | BUG_ON(busiest_rq == target_rq); | |
4134 | ||
4135 | /* move a task from busiest_rq to target_rq */ | |
4136 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
4137 | |
4138 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 4139 | rcu_read_lock(); |
1e3c88bd PZ |
4140 | for_each_domain(target_cpu, sd) { |
4141 | if ((sd->flags & SD_LOAD_BALANCE) && | |
4142 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
4143 | break; | |
4144 | } | |
4145 | ||
4146 | if (likely(sd)) { | |
4147 | schedstat_inc(sd, alb_count); | |
4148 | ||
4149 | if (move_one_task(target_rq, target_cpu, busiest_rq, | |
4150 | sd, CPU_IDLE)) | |
4151 | schedstat_inc(sd, alb_pushed); | |
4152 | else | |
4153 | schedstat_inc(sd, alb_failed); | |
4154 | } | |
dce840a0 | 4155 | rcu_read_unlock(); |
1e3c88bd | 4156 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
4157 | out_unlock: |
4158 | busiest_rq->active_balance = 0; | |
4159 | raw_spin_unlock_irq(&busiest_rq->lock); | |
4160 | return 0; | |
1e3c88bd PZ |
4161 | } |
4162 | ||
4163 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
4164 | |
4165 | static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb); | |
4166 | ||
4167 | static void trigger_sched_softirq(void *data) | |
4168 | { | |
4169 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
4170 | } | |
4171 | ||
4172 | static inline void init_sched_softirq_csd(struct call_single_data *csd) | |
4173 | { | |
4174 | csd->func = trigger_sched_softirq; | |
4175 | csd->info = NULL; | |
4176 | csd->flags = 0; | |
4177 | csd->priv = 0; | |
4178 | } | |
4179 | ||
4180 | /* | |
4181 | * idle load balancing details | |
4182 | * - One of the idle CPUs nominates itself as idle load_balancer, while | |
4183 | * entering idle. | |
4184 | * - This idle load balancer CPU will also go into tickless mode when | |
4185 | * it is idle, just like all other idle CPUs | |
4186 | * - When one of the busy CPUs notice that there may be an idle rebalancing | |
4187 | * needed, they will kick the idle load balancer, which then does idle | |
4188 | * load balancing for all the idle CPUs. | |
4189 | */ | |
1e3c88bd PZ |
4190 | static struct { |
4191 | atomic_t load_balancer; | |
83cd4fe2 VP |
4192 | atomic_t first_pick_cpu; |
4193 | atomic_t second_pick_cpu; | |
4194 | cpumask_var_t idle_cpus_mask; | |
4195 | cpumask_var_t grp_idle_mask; | |
4196 | unsigned long next_balance; /* in jiffy units */ | |
4197 | } nohz ____cacheline_aligned; | |
1e3c88bd PZ |
4198 | |
4199 | int get_nohz_load_balancer(void) | |
4200 | { | |
4201 | return atomic_read(&nohz.load_balancer); | |
4202 | } | |
4203 | ||
4204 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
4205 | /** | |
4206 | * lowest_flag_domain - Return lowest sched_domain containing flag. | |
4207 | * @cpu: The cpu whose lowest level of sched domain is to | |
4208 | * be returned. | |
4209 | * @flag: The flag to check for the lowest sched_domain | |
4210 | * for the given cpu. | |
4211 | * | |
4212 | * Returns the lowest sched_domain of a cpu which contains the given flag. | |
4213 | */ | |
4214 | static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) | |
4215 | { | |
4216 | struct sched_domain *sd; | |
4217 | ||
4218 | for_each_domain(cpu, sd) | |
08354716 | 4219 | if (sd->flags & flag) |
1e3c88bd PZ |
4220 | break; |
4221 | ||
4222 | return sd; | |
4223 | } | |
4224 | ||
4225 | /** | |
4226 | * for_each_flag_domain - Iterates over sched_domains containing the flag. | |
4227 | * @cpu: The cpu whose domains we're iterating over. | |
4228 | * @sd: variable holding the value of the power_savings_sd | |
4229 | * for cpu. | |
4230 | * @flag: The flag to filter the sched_domains to be iterated. | |
4231 | * | |
4232 | * Iterates over all the scheduler domains for a given cpu that has the 'flag' | |
4233 | * set, starting from the lowest sched_domain to the highest. | |
4234 | */ | |
4235 | #define for_each_flag_domain(cpu, sd, flag) \ | |
4236 | for (sd = lowest_flag_domain(cpu, flag); \ | |
4237 | (sd && (sd->flags & flag)); sd = sd->parent) | |
4238 | ||
4239 | /** | |
4240 | * is_semi_idle_group - Checks if the given sched_group is semi-idle. | |
4241 | * @ilb_group: group to be checked for semi-idleness | |
4242 | * | |
4243 | * Returns: 1 if the group is semi-idle. 0 otherwise. | |
4244 | * | |
4245 | * We define a sched_group to be semi idle if it has atleast one idle-CPU | |
4246 | * and atleast one non-idle CPU. This helper function checks if the given | |
4247 | * sched_group is semi-idle or not. | |
4248 | */ | |
4249 | static inline int is_semi_idle_group(struct sched_group *ilb_group) | |
4250 | { | |
83cd4fe2 | 4251 | cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask, |
1e3c88bd PZ |
4252 | sched_group_cpus(ilb_group)); |
4253 | ||
4254 | /* | |
4255 | * A sched_group is semi-idle when it has atleast one busy cpu | |
4256 | * and atleast one idle cpu. | |
4257 | */ | |
83cd4fe2 | 4258 | if (cpumask_empty(nohz.grp_idle_mask)) |
1e3c88bd PZ |
4259 | return 0; |
4260 | ||
83cd4fe2 | 4261 | if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group))) |
1e3c88bd PZ |
4262 | return 0; |
4263 | ||
4264 | return 1; | |
4265 | } | |
4266 | /** | |
4267 | * find_new_ilb - Finds the optimum idle load balancer for nomination. | |
4268 | * @cpu: The cpu which is nominating a new idle_load_balancer. | |
4269 | * | |
4270 | * Returns: Returns the id of the idle load balancer if it exists, | |
4271 | * Else, returns >= nr_cpu_ids. | |
4272 | * | |
4273 | * This algorithm picks the idle load balancer such that it belongs to a | |
4274 | * semi-idle powersavings sched_domain. The idea is to try and avoid | |
4275 | * completely idle packages/cores just for the purpose of idle load balancing | |
4276 | * when there are other idle cpu's which are better suited for that job. | |
4277 | */ | |
4278 | static int find_new_ilb(int cpu) | |
4279 | { | |
4280 | struct sched_domain *sd; | |
4281 | struct sched_group *ilb_group; | |
dce840a0 | 4282 | int ilb = nr_cpu_ids; |
1e3c88bd PZ |
4283 | |
4284 | /* | |
4285 | * Have idle load balancer selection from semi-idle packages only | |
4286 | * when power-aware load balancing is enabled | |
4287 | */ | |
4288 | if (!(sched_smt_power_savings || sched_mc_power_savings)) | |
4289 | goto out_done; | |
4290 | ||
4291 | /* | |
4292 | * Optimize for the case when we have no idle CPUs or only one | |
4293 | * idle CPU. Don't walk the sched_domain hierarchy in such cases | |
4294 | */ | |
83cd4fe2 | 4295 | if (cpumask_weight(nohz.idle_cpus_mask) < 2) |
1e3c88bd PZ |
4296 | goto out_done; |
4297 | ||
dce840a0 | 4298 | rcu_read_lock(); |
1e3c88bd PZ |
4299 | for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { |
4300 | ilb_group = sd->groups; | |
4301 | ||
4302 | do { | |
dce840a0 PZ |
4303 | if (is_semi_idle_group(ilb_group)) { |
4304 | ilb = cpumask_first(nohz.grp_idle_mask); | |
4305 | goto unlock; | |
4306 | } | |
1e3c88bd PZ |
4307 | |
4308 | ilb_group = ilb_group->next; | |
4309 | ||
4310 | } while (ilb_group != sd->groups); | |
4311 | } | |
dce840a0 PZ |
4312 | unlock: |
4313 | rcu_read_unlock(); | |
1e3c88bd PZ |
4314 | |
4315 | out_done: | |
dce840a0 | 4316 | return ilb; |
1e3c88bd PZ |
4317 | } |
4318 | #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ | |
4319 | static inline int find_new_ilb(int call_cpu) | |
4320 | { | |
83cd4fe2 | 4321 | return nr_cpu_ids; |
1e3c88bd PZ |
4322 | } |
4323 | #endif | |
4324 | ||
83cd4fe2 VP |
4325 | /* |
4326 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
4327 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
4328 | * CPU (if there is one). | |
4329 | */ | |
4330 | static void nohz_balancer_kick(int cpu) | |
4331 | { | |
4332 | int ilb_cpu; | |
4333 | ||
4334 | nohz.next_balance++; | |
4335 | ||
4336 | ilb_cpu = get_nohz_load_balancer(); | |
4337 | ||
4338 | if (ilb_cpu >= nr_cpu_ids) { | |
4339 | ilb_cpu = cpumask_first(nohz.idle_cpus_mask); | |
4340 | if (ilb_cpu >= nr_cpu_ids) | |
4341 | return; | |
4342 | } | |
4343 | ||
4344 | if (!cpu_rq(ilb_cpu)->nohz_balance_kick) { | |
4345 | struct call_single_data *cp; | |
4346 | ||
4347 | cpu_rq(ilb_cpu)->nohz_balance_kick = 1; | |
4348 | cp = &per_cpu(remote_sched_softirq_cb, cpu); | |
4349 | __smp_call_function_single(ilb_cpu, cp, 0); | |
4350 | } | |
4351 | return; | |
4352 | } | |
4353 | ||
1e3c88bd PZ |
4354 | /* |
4355 | * This routine will try to nominate the ilb (idle load balancing) | |
4356 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
83cd4fe2 | 4357 | * load balancing on behalf of all those cpus. |
1e3c88bd | 4358 | * |
83cd4fe2 VP |
4359 | * When the ilb owner becomes busy, we will not have new ilb owner until some |
4360 | * idle CPU wakes up and goes back to idle or some busy CPU tries to kick | |
4361 | * idle load balancing by kicking one of the idle CPUs. | |
1e3c88bd | 4362 | * |
83cd4fe2 VP |
4363 | * Ticks are stopped for the ilb owner as well, with busy CPU kicking this |
4364 | * ilb owner CPU in future (when there is a need for idle load balancing on | |
4365 | * behalf of all idle CPUs). | |
1e3c88bd | 4366 | */ |
83cd4fe2 | 4367 | void select_nohz_load_balancer(int stop_tick) |
1e3c88bd PZ |
4368 | { |
4369 | int cpu = smp_processor_id(); | |
4370 | ||
4371 | if (stop_tick) { | |
1e3c88bd PZ |
4372 | if (!cpu_active(cpu)) { |
4373 | if (atomic_read(&nohz.load_balancer) != cpu) | |
83cd4fe2 | 4374 | return; |
1e3c88bd PZ |
4375 | |
4376 | /* | |
4377 | * If we are going offline and still the leader, | |
4378 | * give up! | |
4379 | */ | |
83cd4fe2 VP |
4380 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
4381 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
4382 | BUG(); |
4383 | ||
83cd4fe2 | 4384 | return; |
1e3c88bd PZ |
4385 | } |
4386 | ||
83cd4fe2 | 4387 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd | 4388 | |
83cd4fe2 VP |
4389 | if (atomic_read(&nohz.first_pick_cpu) == cpu) |
4390 | atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids); | |
4391 | if (atomic_read(&nohz.second_pick_cpu) == cpu) | |
4392 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
1e3c88bd | 4393 | |
83cd4fe2 | 4394 | if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) { |
1e3c88bd PZ |
4395 | int new_ilb; |
4396 | ||
83cd4fe2 VP |
4397 | /* make me the ilb owner */ |
4398 | if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids, | |
4399 | cpu) != nr_cpu_ids) | |
4400 | return; | |
4401 | ||
1e3c88bd PZ |
4402 | /* |
4403 | * Check to see if there is a more power-efficient | |
4404 | * ilb. | |
4405 | */ | |
4406 | new_ilb = find_new_ilb(cpu); | |
4407 | if (new_ilb < nr_cpu_ids && new_ilb != cpu) { | |
83cd4fe2 | 4408 | atomic_set(&nohz.load_balancer, nr_cpu_ids); |
1e3c88bd | 4409 | resched_cpu(new_ilb); |
83cd4fe2 | 4410 | return; |
1e3c88bd | 4411 | } |
83cd4fe2 | 4412 | return; |
1e3c88bd PZ |
4413 | } |
4414 | } else { | |
83cd4fe2 VP |
4415 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
4416 | return; | |
1e3c88bd | 4417 | |
83cd4fe2 | 4418 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
1e3c88bd PZ |
4419 | |
4420 | if (atomic_read(&nohz.load_balancer) == cpu) | |
83cd4fe2 VP |
4421 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, |
4422 | nr_cpu_ids) != cpu) | |
1e3c88bd PZ |
4423 | BUG(); |
4424 | } | |
83cd4fe2 | 4425 | return; |
1e3c88bd PZ |
4426 | } |
4427 | #endif | |
4428 | ||
4429 | static DEFINE_SPINLOCK(balancing); | |
4430 | ||
49c022e6 PZ |
4431 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4432 | ||
4433 | /* | |
4434 | * Scale the max load_balance interval with the number of CPUs in the system. | |
4435 | * This trades load-balance latency on larger machines for less cross talk. | |
4436 | */ | |
4437 | static void update_max_interval(void) | |
4438 | { | |
4439 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
4440 | } | |
4441 | ||
1e3c88bd PZ |
4442 | /* |
4443 | * It checks each scheduling domain to see if it is due to be balanced, | |
4444 | * and initiates a balancing operation if so. | |
4445 | * | |
4446 | * Balancing parameters are set up in arch_init_sched_domains. | |
4447 | */ | |
4448 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4449 | { | |
4450 | int balance = 1; | |
4451 | struct rq *rq = cpu_rq(cpu); | |
4452 | unsigned long interval; | |
4453 | struct sched_domain *sd; | |
4454 | /* Earliest time when we have to do rebalance again */ | |
4455 | unsigned long next_balance = jiffies + 60*HZ; | |
4456 | int update_next_balance = 0; | |
4457 | int need_serialize; | |
4458 | ||
2069dd75 PZ |
4459 | update_shares(cpu); |
4460 | ||
dce840a0 | 4461 | rcu_read_lock(); |
1e3c88bd PZ |
4462 | for_each_domain(cpu, sd) { |
4463 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4464 | continue; | |
4465 | ||
4466 | interval = sd->balance_interval; | |
4467 | if (idle != CPU_IDLE) | |
4468 | interval *= sd->busy_factor; | |
4469 | ||
4470 | /* scale ms to jiffies */ | |
4471 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 4472 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
4473 | |
4474 | need_serialize = sd->flags & SD_SERIALIZE; | |
4475 | ||
4476 | if (need_serialize) { | |
4477 | if (!spin_trylock(&balancing)) | |
4478 | goto out; | |
4479 | } | |
4480 | ||
4481 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4482 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
4483 | /* | |
4484 | * We've pulled tasks over so either we're no | |
c186fafe | 4485 | * longer idle. |
1e3c88bd PZ |
4486 | */ |
4487 | idle = CPU_NOT_IDLE; | |
4488 | } | |
4489 | sd->last_balance = jiffies; | |
4490 | } | |
4491 | if (need_serialize) | |
4492 | spin_unlock(&balancing); | |
4493 | out: | |
4494 | if (time_after(next_balance, sd->last_balance + interval)) { | |
4495 | next_balance = sd->last_balance + interval; | |
4496 | update_next_balance = 1; | |
4497 | } | |
4498 | ||
4499 | /* | |
4500 | * Stop the load balance at this level. There is another | |
4501 | * CPU in our sched group which is doing load balancing more | |
4502 | * actively. | |
4503 | */ | |
4504 | if (!balance) | |
4505 | break; | |
4506 | } | |
dce840a0 | 4507 | rcu_read_unlock(); |
1e3c88bd PZ |
4508 | |
4509 | /* | |
4510 | * next_balance will be updated only when there is a need. | |
4511 | * When the cpu is attached to null domain for ex, it will not be | |
4512 | * updated. | |
4513 | */ | |
4514 | if (likely(update_next_balance)) | |
4515 | rq->next_balance = next_balance; | |
4516 | } | |
4517 | ||
83cd4fe2 | 4518 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 4519 | /* |
83cd4fe2 | 4520 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
4521 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
4522 | */ | |
83cd4fe2 VP |
4523 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
4524 | { | |
4525 | struct rq *this_rq = cpu_rq(this_cpu); | |
4526 | struct rq *rq; | |
4527 | int balance_cpu; | |
4528 | ||
4529 | if (idle != CPU_IDLE || !this_rq->nohz_balance_kick) | |
4530 | return; | |
4531 | ||
4532 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
4533 | if (balance_cpu == this_cpu) | |
4534 | continue; | |
4535 | ||
4536 | /* | |
4537 | * If this cpu gets work to do, stop the load balancing | |
4538 | * work being done for other cpus. Next load | |
4539 | * balancing owner will pick it up. | |
4540 | */ | |
4541 | if (need_resched()) { | |
4542 | this_rq->nohz_balance_kick = 0; | |
4543 | break; | |
4544 | } | |
4545 | ||
4546 | raw_spin_lock_irq(&this_rq->lock); | |
5343bdb8 | 4547 | update_rq_clock(this_rq); |
83cd4fe2 VP |
4548 | update_cpu_load(this_rq); |
4549 | raw_spin_unlock_irq(&this_rq->lock); | |
4550 | ||
4551 | rebalance_domains(balance_cpu, CPU_IDLE); | |
4552 | ||
4553 | rq = cpu_rq(balance_cpu); | |
4554 | if (time_after(this_rq->next_balance, rq->next_balance)) | |
4555 | this_rq->next_balance = rq->next_balance; | |
4556 | } | |
4557 | nohz.next_balance = this_rq->next_balance; | |
4558 | this_rq->nohz_balance_kick = 0; | |
4559 | } | |
4560 | ||
4561 | /* | |
4562 | * Current heuristic for kicking the idle load balancer | |
4563 | * - first_pick_cpu is the one of the busy CPUs. It will kick | |
4564 | * idle load balancer when it has more than one process active. This | |
4565 | * eliminates the need for idle load balancing altogether when we have | |
4566 | * only one running process in the system (common case). | |
4567 | * - If there are more than one busy CPU, idle load balancer may have | |
4568 | * to run for active_load_balance to happen (i.e., two busy CPUs are | |
4569 | * SMT or core siblings and can run better if they move to different | |
4570 | * physical CPUs). So, second_pick_cpu is the second of the busy CPUs | |
4571 | * which will kick idle load balancer as soon as it has any load. | |
4572 | */ | |
4573 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
4574 | { | |
4575 | unsigned long now = jiffies; | |
4576 | int ret; | |
4577 | int first_pick_cpu, second_pick_cpu; | |
4578 | ||
4579 | if (time_before(now, nohz.next_balance)) | |
4580 | return 0; | |
4581 | ||
f6c3f168 | 4582 | if (rq->idle_at_tick) |
83cd4fe2 VP |
4583 | return 0; |
4584 | ||
4585 | first_pick_cpu = atomic_read(&nohz.first_pick_cpu); | |
4586 | second_pick_cpu = atomic_read(&nohz.second_pick_cpu); | |
4587 | ||
4588 | if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu && | |
4589 | second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu) | |
4590 | return 0; | |
4591 | ||
4592 | ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu); | |
4593 | if (ret == nr_cpu_ids || ret == cpu) { | |
4594 | atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids); | |
4595 | if (rq->nr_running > 1) | |
4596 | return 1; | |
4597 | } else { | |
4598 | ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu); | |
4599 | if (ret == nr_cpu_ids || ret == cpu) { | |
4600 | if (rq->nr_running) | |
4601 | return 1; | |
4602 | } | |
4603 | } | |
4604 | return 0; | |
4605 | } | |
4606 | #else | |
4607 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
4608 | #endif | |
4609 | ||
4610 | /* | |
4611 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4612 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
4613 | */ | |
1e3c88bd PZ |
4614 | static void run_rebalance_domains(struct softirq_action *h) |
4615 | { | |
4616 | int this_cpu = smp_processor_id(); | |
4617 | struct rq *this_rq = cpu_rq(this_cpu); | |
4618 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4619 | CPU_IDLE : CPU_NOT_IDLE; | |
4620 | ||
4621 | rebalance_domains(this_cpu, idle); | |
4622 | ||
1e3c88bd | 4623 | /* |
83cd4fe2 | 4624 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
4625 | * balancing on behalf of the other idle cpus whose ticks are |
4626 | * stopped. | |
4627 | */ | |
83cd4fe2 | 4628 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
4629 | } |
4630 | ||
4631 | static inline int on_null_domain(int cpu) | |
4632 | { | |
90a6501f | 4633 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
4634 | } |
4635 | ||
4636 | /* | |
4637 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd PZ |
4638 | */ |
4639 | static inline void trigger_load_balance(struct rq *rq, int cpu) | |
4640 | { | |
1e3c88bd PZ |
4641 | /* Don't need to rebalance while attached to NULL domain */ |
4642 | if (time_after_eq(jiffies, rq->next_balance) && | |
4643 | likely(!on_null_domain(cpu))) | |
4644 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 VP |
4645 | #ifdef CONFIG_NO_HZ |
4646 | else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) | |
4647 | nohz_balancer_kick(cpu); | |
4648 | #endif | |
1e3c88bd PZ |
4649 | } |
4650 | ||
0bcdcf28 CE |
4651 | static void rq_online_fair(struct rq *rq) |
4652 | { | |
4653 | update_sysctl(); | |
4654 | } | |
4655 | ||
4656 | static void rq_offline_fair(struct rq *rq) | |
4657 | { | |
4658 | update_sysctl(); | |
4659 | } | |
4660 | ||
1e3c88bd PZ |
4661 | #else /* CONFIG_SMP */ |
4662 | ||
4663 | /* | |
4664 | * on UP we do not need to balance between CPUs: | |
4665 | */ | |
4666 | static inline void idle_balance(int cpu, struct rq *rq) | |
4667 | { | |
4668 | } | |
4669 | ||
55e12e5e | 4670 | #endif /* CONFIG_SMP */ |
e1d1484f | 4671 | |
bf0f6f24 IM |
4672 | /* |
4673 | * scheduler tick hitting a task of our scheduling class: | |
4674 | */ | |
8f4d37ec | 4675 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
4676 | { |
4677 | struct cfs_rq *cfs_rq; | |
4678 | struct sched_entity *se = &curr->se; | |
4679 | ||
4680 | for_each_sched_entity(se) { | |
4681 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 4682 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 IM |
4683 | } |
4684 | } | |
4685 | ||
4686 | /* | |
cd29fe6f PZ |
4687 | * called on fork with the child task as argument from the parent's context |
4688 | * - child not yet on the tasklist | |
4689 | * - preemption disabled | |
bf0f6f24 | 4690 | */ |
cd29fe6f | 4691 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 4692 | { |
cd29fe6f | 4693 | struct cfs_rq *cfs_rq = task_cfs_rq(current); |
429d43bc | 4694 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
00bf7bfc | 4695 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
4696 | struct rq *rq = this_rq(); |
4697 | unsigned long flags; | |
4698 | ||
05fa785c | 4699 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 4700 | |
861d034e PZ |
4701 | update_rq_clock(rq); |
4702 | ||
b0a0f667 PM |
4703 | if (unlikely(task_cpu(p) != this_cpu)) { |
4704 | rcu_read_lock(); | |
cd29fe6f | 4705 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
4706 | rcu_read_unlock(); |
4707 | } | |
bf0f6f24 | 4708 | |
7109c442 | 4709 | update_curr(cfs_rq); |
cd29fe6f | 4710 | |
b5d9d734 MG |
4711 | if (curr) |
4712 | se->vruntime = curr->vruntime; | |
aeb73b04 | 4713 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 4714 | |
cd29fe6f | 4715 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 4716 | /* |
edcb60a3 IM |
4717 | * Upon rescheduling, sched_class::put_prev_task() will place |
4718 | * 'current' within the tree based on its new key value. | |
4719 | */ | |
4d78e7b6 | 4720 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 4721 | resched_task(rq->curr); |
4d78e7b6 | 4722 | } |
bf0f6f24 | 4723 | |
88ec22d3 PZ |
4724 | se->vruntime -= cfs_rq->min_vruntime; |
4725 | ||
05fa785c | 4726 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
4727 | } |
4728 | ||
cb469845 SR |
4729 | /* |
4730 | * Priority of the task has changed. Check to see if we preempt | |
4731 | * the current task. | |
4732 | */ | |
da7a735e PZ |
4733 | static void |
4734 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 4735 | { |
da7a735e PZ |
4736 | if (!p->se.on_rq) |
4737 | return; | |
4738 | ||
cb469845 SR |
4739 | /* |
4740 | * Reschedule if we are currently running on this runqueue and | |
4741 | * our priority decreased, or if we are not currently running on | |
4742 | * this runqueue and our priority is higher than the current's | |
4743 | */ | |
da7a735e | 4744 | if (rq->curr == p) { |
cb469845 SR |
4745 | if (p->prio > oldprio) |
4746 | resched_task(rq->curr); | |
4747 | } else | |
15afe09b | 4748 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4749 | } |
4750 | ||
da7a735e PZ |
4751 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
4752 | { | |
4753 | struct sched_entity *se = &p->se; | |
4754 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4755 | ||
4756 | /* | |
4757 | * Ensure the task's vruntime is normalized, so that when its | |
4758 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
4759 | * do the right thing. | |
4760 | * | |
4761 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
4762 | * have normalized the vruntime, if it was !on_rq, then only when | |
4763 | * the task is sleeping will it still have non-normalized vruntime. | |
4764 | */ | |
4765 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
4766 | /* | |
4767 | * Fix up our vruntime so that the current sleep doesn't | |
4768 | * cause 'unlimited' sleep bonus. | |
4769 | */ | |
4770 | place_entity(cfs_rq, se, 0); | |
4771 | se->vruntime -= cfs_rq->min_vruntime; | |
4772 | } | |
4773 | } | |
4774 | ||
cb469845 SR |
4775 | /* |
4776 | * We switched to the sched_fair class. | |
4777 | */ | |
da7a735e | 4778 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 4779 | { |
da7a735e PZ |
4780 | if (!p->se.on_rq) |
4781 | return; | |
4782 | ||
cb469845 SR |
4783 | /* |
4784 | * We were most likely switched from sched_rt, so | |
4785 | * kick off the schedule if running, otherwise just see | |
4786 | * if we can still preempt the current task. | |
4787 | */ | |
da7a735e | 4788 | if (rq->curr == p) |
cb469845 SR |
4789 | resched_task(rq->curr); |
4790 | else | |
15afe09b | 4791 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
4792 | } |
4793 | ||
83b699ed SV |
4794 | /* Account for a task changing its policy or group. |
4795 | * | |
4796 | * This routine is mostly called to set cfs_rq->curr field when a task | |
4797 | * migrates between groups/classes. | |
4798 | */ | |
4799 | static void set_curr_task_fair(struct rq *rq) | |
4800 | { | |
4801 | struct sched_entity *se = &rq->curr->se; | |
4802 | ||
ec12cb7f PT |
4803 | for_each_sched_entity(se) { |
4804 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4805 | ||
4806 | set_next_entity(cfs_rq, se); | |
4807 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
4808 | account_cfs_rq_runtime(cfs_rq, 0); | |
4809 | } | |
83b699ed SV |
4810 | } |
4811 | ||
810b3817 | 4812 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4813 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 4814 | { |
b2b5ce02 PZ |
4815 | /* |
4816 | * If the task was not on the rq at the time of this cgroup movement | |
4817 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
4818 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
4819 | * bonus in place_entity()). | |
4820 | * | |
4821 | * If it was on the rq, we've just 'preempted' it, which does convert | |
4822 | * ->vruntime to a relative base. | |
4823 | * | |
4824 | * Make sure both cases convert their relative position when migrating | |
4825 | * to another cgroup's rq. This does somewhat interfere with the | |
4826 | * fair sleeper stuff for the first placement, but who cares. | |
4827 | */ | |
4828 | if (!on_rq) | |
4829 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
4830 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 4831 | if (!on_rq) |
b2b5ce02 | 4832 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 PZ |
4833 | } |
4834 | #endif | |
4835 | ||
6d686f45 | 4836 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
4837 | { |
4838 | struct sched_entity *se = &task->se; | |
0d721cea PW |
4839 | unsigned int rr_interval = 0; |
4840 | ||
4841 | /* | |
4842 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
4843 | * idle runqueue: | |
4844 | */ | |
0d721cea PW |
4845 | if (rq->cfs.load.weight) |
4846 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
4847 | |
4848 | return rr_interval; | |
4849 | } | |
4850 | ||
bf0f6f24 IM |
4851 | /* |
4852 | * All the scheduling class methods: | |
4853 | */ | |
5522d5d5 IM |
4854 | static const struct sched_class fair_sched_class = { |
4855 | .next = &idle_sched_class, | |
bf0f6f24 IM |
4856 | .enqueue_task = enqueue_task_fair, |
4857 | .dequeue_task = dequeue_task_fair, | |
4858 | .yield_task = yield_task_fair, | |
d95f4122 | 4859 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 4860 | |
2e09bf55 | 4861 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
4862 | |
4863 | .pick_next_task = pick_next_task_fair, | |
4864 | .put_prev_task = put_prev_task_fair, | |
4865 | ||
681f3e68 | 4866 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
4867 | .select_task_rq = select_task_rq_fair, |
4868 | ||
0bcdcf28 CE |
4869 | .rq_online = rq_online_fair, |
4870 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
4871 | |
4872 | .task_waking = task_waking_fair, | |
681f3e68 | 4873 | #endif |
bf0f6f24 | 4874 | |
83b699ed | 4875 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 4876 | .task_tick = task_tick_fair, |
cd29fe6f | 4877 | .task_fork = task_fork_fair, |
cb469845 SR |
4878 | |
4879 | .prio_changed = prio_changed_fair, | |
da7a735e | 4880 | .switched_from = switched_from_fair, |
cb469845 | 4881 | .switched_to = switched_to_fair, |
810b3817 | 4882 | |
0d721cea PW |
4883 | .get_rr_interval = get_rr_interval_fair, |
4884 | ||
810b3817 | 4885 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 4886 | .task_move_group = task_move_group_fair, |
810b3817 | 4887 | #endif |
bf0f6f24 IM |
4888 | }; |
4889 | ||
4890 | #ifdef CONFIG_SCHED_DEBUG | |
5cef9eca | 4891 | static void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 4892 | { |
bf0f6f24 IM |
4893 | struct cfs_rq *cfs_rq; |
4894 | ||
5973e5b9 | 4895 | rcu_read_lock(); |
c3b64f1e | 4896 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 4897 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 4898 | rcu_read_unlock(); |
bf0f6f24 IM |
4899 | } |
4900 | #endif |