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> |
029632fb PZ |
26 | #include <linux/slab.h> |
27 | #include <linux/profile.h> | |
28 | #include <linux/interrupt.h> | |
cbee9f88 PZ |
29 | #include <linux/mempolicy.h> |
30 | #include <linux/task_work.h> | |
029632fb PZ |
31 | |
32 | #include <trace/events/sched.h> | |
33 | ||
34 | #include "sched.h" | |
9745512c | 35 | |
bf0f6f24 | 36 | /* |
21805085 | 37 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 38 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 39 | * |
21805085 | 40 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
41 | * 'timeslice length' - timeslices in CFS are of variable length |
42 | * and have no persistent notion like in traditional, time-slice | |
43 | * based scheduling concepts. | |
bf0f6f24 | 44 | * |
d274a4ce IM |
45 | * (to see the precise effective timeslice length of your workload, |
46 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 47 | */ |
21406928 MG |
48 | unsigned int sysctl_sched_latency = 6000000ULL; |
49 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 50 | |
1983a922 CE |
51 | /* |
52 | * The initial- and re-scaling of tunables is configurable | |
53 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
54 | * | |
55 | * Options are: | |
56 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
57 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
58 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
59 | */ | |
60 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
61 | = SCHED_TUNABLESCALING_LOG; | |
62 | ||
2bd8e6d4 | 63 | /* |
b2be5e96 | 64 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 65 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 66 | */ |
0bf377bb IM |
67 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
68 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
69 | |
70 | /* | |
b2be5e96 PZ |
71 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
72 | */ | |
0bf377bb | 73 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
74 | |
75 | /* | |
2bba22c5 | 76 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 77 | * parent will (try to) run first. |
21805085 | 78 | */ |
2bba22c5 | 79 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 80 | |
bf0f6f24 IM |
81 | /* |
82 | * SCHED_OTHER wake-up granularity. | |
172e082a | 83 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
84 | * |
85 | * This option delays the preemption effects of decoupled workloads | |
86 | * and reduces their over-scheduling. Synchronous workloads will still | |
87 | * have immediate wakeup/sleep latencies. | |
88 | */ | |
172e082a | 89 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 90 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 91 | |
da84d961 IM |
92 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
93 | ||
a7a4f8a7 PT |
94 | /* |
95 | * The exponential sliding window over which load is averaged for shares | |
96 | * distribution. | |
97 | * (default: 10msec) | |
98 | */ | |
99 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
100 | ||
ec12cb7f PT |
101 | #ifdef CONFIG_CFS_BANDWIDTH |
102 | /* | |
103 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
104 | * each time a cfs_rq requests quota. | |
105 | * | |
106 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
107 | * to consumption or the quota being specified to be smaller than the slice) | |
108 | * we will always only issue the remaining available time. | |
109 | * | |
110 | * default: 5 msec, units: microseconds | |
111 | */ | |
112 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
113 | #endif | |
114 | ||
029632fb PZ |
115 | /* |
116 | * Increase the granularity value when there are more CPUs, | |
117 | * because with more CPUs the 'effective latency' as visible | |
118 | * to users decreases. But the relationship is not linear, | |
119 | * so pick a second-best guess by going with the log2 of the | |
120 | * number of CPUs. | |
121 | * | |
122 | * This idea comes from the SD scheduler of Con Kolivas: | |
123 | */ | |
124 | static int get_update_sysctl_factor(void) | |
125 | { | |
126 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
127 | unsigned int factor; | |
128 | ||
129 | switch (sysctl_sched_tunable_scaling) { | |
130 | case SCHED_TUNABLESCALING_NONE: | |
131 | factor = 1; | |
132 | break; | |
133 | case SCHED_TUNABLESCALING_LINEAR: | |
134 | factor = cpus; | |
135 | break; | |
136 | case SCHED_TUNABLESCALING_LOG: | |
137 | default: | |
138 | factor = 1 + ilog2(cpus); | |
139 | break; | |
140 | } | |
141 | ||
142 | return factor; | |
143 | } | |
144 | ||
145 | static void update_sysctl(void) | |
146 | { | |
147 | unsigned int factor = get_update_sysctl_factor(); | |
148 | ||
149 | #define SET_SYSCTL(name) \ | |
150 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
151 | SET_SYSCTL(sched_min_granularity); | |
152 | SET_SYSCTL(sched_latency); | |
153 | SET_SYSCTL(sched_wakeup_granularity); | |
154 | #undef SET_SYSCTL | |
155 | } | |
156 | ||
157 | void sched_init_granularity(void) | |
158 | { | |
159 | update_sysctl(); | |
160 | } | |
161 | ||
162 | #if BITS_PER_LONG == 32 | |
163 | # define WMULT_CONST (~0UL) | |
164 | #else | |
165 | # define WMULT_CONST (1UL << 32) | |
166 | #endif | |
167 | ||
168 | #define WMULT_SHIFT 32 | |
169 | ||
170 | /* | |
171 | * Shift right and round: | |
172 | */ | |
173 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
174 | ||
175 | /* | |
176 | * delta *= weight / lw | |
177 | */ | |
178 | static unsigned long | |
179 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
180 | struct load_weight *lw) | |
181 | { | |
182 | u64 tmp; | |
183 | ||
184 | /* | |
185 | * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched | |
186 | * entities since MIN_SHARES = 2. Treat weight as 1 if less than | |
187 | * 2^SCHED_LOAD_RESOLUTION. | |
188 | */ | |
189 | if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) | |
190 | tmp = (u64)delta_exec * scale_load_down(weight); | |
191 | else | |
192 | tmp = (u64)delta_exec; | |
193 | ||
194 | if (!lw->inv_weight) { | |
195 | unsigned long w = scale_load_down(lw->weight); | |
196 | ||
197 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
198 | lw->inv_weight = 1; | |
199 | else if (unlikely(!w)) | |
200 | lw->inv_weight = WMULT_CONST; | |
201 | else | |
202 | lw->inv_weight = WMULT_CONST / w; | |
203 | } | |
204 | ||
205 | /* | |
206 | * Check whether we'd overflow the 64-bit multiplication: | |
207 | */ | |
208 | if (unlikely(tmp > WMULT_CONST)) | |
209 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
210 | WMULT_SHIFT/2); | |
211 | else | |
212 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
213 | ||
214 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
215 | } | |
216 | ||
217 | ||
218 | const struct sched_class fair_sched_class; | |
a4c2f00f | 219 | |
bf0f6f24 IM |
220 | /************************************************************** |
221 | * CFS operations on generic schedulable entities: | |
222 | */ | |
223 | ||
62160e3f | 224 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 225 | |
62160e3f | 226 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
227 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
228 | { | |
62160e3f | 229 | return cfs_rq->rq; |
bf0f6f24 IM |
230 | } |
231 | ||
62160e3f IM |
232 | /* An entity is a task if it doesn't "own" a runqueue */ |
233 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 234 | |
8f48894f PZ |
235 | static inline struct task_struct *task_of(struct sched_entity *se) |
236 | { | |
237 | #ifdef CONFIG_SCHED_DEBUG | |
238 | WARN_ON_ONCE(!entity_is_task(se)); | |
239 | #endif | |
240 | return container_of(se, struct task_struct, se); | |
241 | } | |
242 | ||
b758149c PZ |
243 | /* Walk up scheduling entities hierarchy */ |
244 | #define for_each_sched_entity(se) \ | |
245 | for (; se; se = se->parent) | |
246 | ||
247 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
248 | { | |
249 | return p->se.cfs_rq; | |
250 | } | |
251 | ||
252 | /* runqueue on which this entity is (to be) queued */ | |
253 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
254 | { | |
255 | return se->cfs_rq; | |
256 | } | |
257 | ||
258 | /* runqueue "owned" by this group */ | |
259 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
260 | { | |
261 | return grp->my_q; | |
262 | } | |
263 | ||
3d4b47b4 PZ |
264 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
265 | { | |
266 | if (!cfs_rq->on_list) { | |
67e86250 PT |
267 | /* |
268 | * Ensure we either appear before our parent (if already | |
269 | * enqueued) or force our parent to appear after us when it is | |
270 | * enqueued. The fact that we always enqueue bottom-up | |
271 | * reduces this to two cases. | |
272 | */ | |
273 | if (cfs_rq->tg->parent && | |
274 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
275 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
276 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
277 | } else { | |
278 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 279 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 280 | } |
3d4b47b4 PZ |
281 | |
282 | cfs_rq->on_list = 1; | |
283 | } | |
284 | } | |
285 | ||
286 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
287 | { | |
288 | if (cfs_rq->on_list) { | |
289 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
290 | cfs_rq->on_list = 0; | |
291 | } | |
292 | } | |
293 | ||
b758149c PZ |
294 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
295 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
296 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
297 | ||
298 | /* Do the two (enqueued) entities belong to the same group ? */ | |
299 | static inline int | |
300 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
301 | { | |
302 | if (se->cfs_rq == pse->cfs_rq) | |
303 | return 1; | |
304 | ||
305 | return 0; | |
306 | } | |
307 | ||
308 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
309 | { | |
310 | return se->parent; | |
311 | } | |
312 | ||
464b7527 PZ |
313 | /* return depth at which a sched entity is present in the hierarchy */ |
314 | static inline int depth_se(struct sched_entity *se) | |
315 | { | |
316 | int depth = 0; | |
317 | ||
318 | for_each_sched_entity(se) | |
319 | depth++; | |
320 | ||
321 | return depth; | |
322 | } | |
323 | ||
324 | static void | |
325 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
326 | { | |
327 | int se_depth, pse_depth; | |
328 | ||
329 | /* | |
330 | * preemption test can be made between sibling entities who are in the | |
331 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
332 | * both tasks until we find their ancestors who are siblings of common | |
333 | * parent. | |
334 | */ | |
335 | ||
336 | /* First walk up until both entities are at same depth */ | |
337 | se_depth = depth_se(*se); | |
338 | pse_depth = depth_se(*pse); | |
339 | ||
340 | while (se_depth > pse_depth) { | |
341 | se_depth--; | |
342 | *se = parent_entity(*se); | |
343 | } | |
344 | ||
345 | while (pse_depth > se_depth) { | |
346 | pse_depth--; | |
347 | *pse = parent_entity(*pse); | |
348 | } | |
349 | ||
350 | while (!is_same_group(*se, *pse)) { | |
351 | *se = parent_entity(*se); | |
352 | *pse = parent_entity(*pse); | |
353 | } | |
354 | } | |
355 | ||
8f48894f PZ |
356 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
357 | ||
358 | static inline struct task_struct *task_of(struct sched_entity *se) | |
359 | { | |
360 | return container_of(se, struct task_struct, se); | |
361 | } | |
bf0f6f24 | 362 | |
62160e3f IM |
363 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
364 | { | |
365 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
366 | } |
367 | ||
368 | #define entity_is_task(se) 1 | |
369 | ||
b758149c PZ |
370 | #define for_each_sched_entity(se) \ |
371 | for (; se; se = NULL) | |
bf0f6f24 | 372 | |
b758149c | 373 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 374 | { |
b758149c | 375 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
376 | } |
377 | ||
b758149c PZ |
378 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
379 | { | |
380 | struct task_struct *p = task_of(se); | |
381 | struct rq *rq = task_rq(p); | |
382 | ||
383 | return &rq->cfs; | |
384 | } | |
385 | ||
386 | /* runqueue "owned" by this group */ | |
387 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
388 | { | |
389 | return NULL; | |
390 | } | |
391 | ||
3d4b47b4 PZ |
392 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
393 | { | |
394 | } | |
395 | ||
396 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
397 | { | |
398 | } | |
399 | ||
b758149c PZ |
400 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
401 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
402 | ||
403 | static inline int | |
404 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
405 | { | |
406 | return 1; | |
407 | } | |
408 | ||
409 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
410 | { | |
411 | return NULL; | |
412 | } | |
413 | ||
464b7527 PZ |
414 | static inline void |
415 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
416 | { | |
417 | } | |
418 | ||
b758149c PZ |
419 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
420 | ||
6c16a6dc PZ |
421 | static __always_inline |
422 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
423 | |
424 | /************************************************************** | |
425 | * Scheduling class tree data structure manipulation methods: | |
426 | */ | |
427 | ||
0702e3eb | 428 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 429 | { |
368059a9 PZ |
430 | s64 delta = (s64)(vruntime - min_vruntime); |
431 | if (delta > 0) | |
02e0431a PZ |
432 | min_vruntime = vruntime; |
433 | ||
434 | return min_vruntime; | |
435 | } | |
436 | ||
0702e3eb | 437 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
438 | { |
439 | s64 delta = (s64)(vruntime - min_vruntime); | |
440 | if (delta < 0) | |
441 | min_vruntime = vruntime; | |
442 | ||
443 | return min_vruntime; | |
444 | } | |
445 | ||
54fdc581 FC |
446 | static inline int entity_before(struct sched_entity *a, |
447 | struct sched_entity *b) | |
448 | { | |
449 | return (s64)(a->vruntime - b->vruntime) < 0; | |
450 | } | |
451 | ||
1af5f730 PZ |
452 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
453 | { | |
454 | u64 vruntime = cfs_rq->min_vruntime; | |
455 | ||
456 | if (cfs_rq->curr) | |
457 | vruntime = cfs_rq->curr->vruntime; | |
458 | ||
459 | if (cfs_rq->rb_leftmost) { | |
460 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
461 | struct sched_entity, | |
462 | run_node); | |
463 | ||
e17036da | 464 | if (!cfs_rq->curr) |
1af5f730 PZ |
465 | vruntime = se->vruntime; |
466 | else | |
467 | vruntime = min_vruntime(vruntime, se->vruntime); | |
468 | } | |
469 | ||
470 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
3fe1698b PZ |
471 | #ifndef CONFIG_64BIT |
472 | smp_wmb(); | |
473 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
474 | #endif | |
1af5f730 PZ |
475 | } |
476 | ||
bf0f6f24 IM |
477 | /* |
478 | * Enqueue an entity into the rb-tree: | |
479 | */ | |
0702e3eb | 480 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
481 | { |
482 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
483 | struct rb_node *parent = NULL; | |
484 | struct sched_entity *entry; | |
bf0f6f24 IM |
485 | int leftmost = 1; |
486 | ||
487 | /* | |
488 | * Find the right place in the rbtree: | |
489 | */ | |
490 | while (*link) { | |
491 | parent = *link; | |
492 | entry = rb_entry(parent, struct sched_entity, run_node); | |
493 | /* | |
494 | * We dont care about collisions. Nodes with | |
495 | * the same key stay together. | |
496 | */ | |
2bd2d6f2 | 497 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
498 | link = &parent->rb_left; |
499 | } else { | |
500 | link = &parent->rb_right; | |
501 | leftmost = 0; | |
502 | } | |
503 | } | |
504 | ||
505 | /* | |
506 | * Maintain a cache of leftmost tree entries (it is frequently | |
507 | * used): | |
508 | */ | |
1af5f730 | 509 | if (leftmost) |
57cb499d | 510 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
511 | |
512 | rb_link_node(&se->run_node, parent, link); | |
513 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
514 | } |
515 | ||
0702e3eb | 516 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 517 | { |
3fe69747 PZ |
518 | if (cfs_rq->rb_leftmost == &se->run_node) { |
519 | struct rb_node *next_node; | |
3fe69747 PZ |
520 | |
521 | next_node = rb_next(&se->run_node); | |
522 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 523 | } |
e9acbff6 | 524 | |
bf0f6f24 | 525 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
526 | } |
527 | ||
029632fb | 528 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 529 | { |
f4b6755f PZ |
530 | struct rb_node *left = cfs_rq->rb_leftmost; |
531 | ||
532 | if (!left) | |
533 | return NULL; | |
534 | ||
535 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
536 | } |
537 | ||
ac53db59 RR |
538 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
539 | { | |
540 | struct rb_node *next = rb_next(&se->run_node); | |
541 | ||
542 | if (!next) | |
543 | return NULL; | |
544 | ||
545 | return rb_entry(next, struct sched_entity, run_node); | |
546 | } | |
547 | ||
548 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 549 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 550 | { |
7eee3e67 | 551 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 552 | |
70eee74b BS |
553 | if (!last) |
554 | return NULL; | |
7eee3e67 IM |
555 | |
556 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
557 | } |
558 | ||
bf0f6f24 IM |
559 | /************************************************************** |
560 | * Scheduling class statistics methods: | |
561 | */ | |
562 | ||
acb4a848 | 563 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 564 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
565 | loff_t *ppos) |
566 | { | |
8d65af78 | 567 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 568 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
569 | |
570 | if (ret || !write) | |
571 | return ret; | |
572 | ||
573 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
574 | sysctl_sched_min_granularity); | |
575 | ||
acb4a848 CE |
576 | #define WRT_SYSCTL(name) \ |
577 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
578 | WRT_SYSCTL(sched_min_granularity); | |
579 | WRT_SYSCTL(sched_latency); | |
580 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
581 | #undef WRT_SYSCTL |
582 | ||
b2be5e96 PZ |
583 | return 0; |
584 | } | |
585 | #endif | |
647e7cac | 586 | |
a7be37ac | 587 | /* |
f9c0b095 | 588 | * delta /= w |
a7be37ac PZ |
589 | */ |
590 | static inline unsigned long | |
591 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
592 | { | |
f9c0b095 PZ |
593 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
594 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
595 | |
596 | return delta; | |
597 | } | |
598 | ||
647e7cac IM |
599 | /* |
600 | * The idea is to set a period in which each task runs once. | |
601 | * | |
532b1858 | 602 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
603 | * this period because otherwise the slices get too small. |
604 | * | |
605 | * p = (nr <= nl) ? l : l*nr/nl | |
606 | */ | |
4d78e7b6 PZ |
607 | static u64 __sched_period(unsigned long nr_running) |
608 | { | |
609 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 610 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
611 | |
612 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 613 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 614 | period *= nr_running; |
4d78e7b6 PZ |
615 | } |
616 | ||
617 | return period; | |
618 | } | |
619 | ||
647e7cac IM |
620 | /* |
621 | * We calculate the wall-time slice from the period by taking a part | |
622 | * proportional to the weight. | |
623 | * | |
f9c0b095 | 624 | * s = p*P[w/rw] |
647e7cac | 625 | */ |
6d0f0ebd | 626 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 627 | { |
0a582440 | 628 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 629 | |
0a582440 | 630 | for_each_sched_entity(se) { |
6272d68c | 631 | struct load_weight *load; |
3104bf03 | 632 | struct load_weight lw; |
6272d68c LM |
633 | |
634 | cfs_rq = cfs_rq_of(se); | |
635 | load = &cfs_rq->load; | |
f9c0b095 | 636 | |
0a582440 | 637 | if (unlikely(!se->on_rq)) { |
3104bf03 | 638 | lw = cfs_rq->load; |
0a582440 MG |
639 | |
640 | update_load_add(&lw, se->load.weight); | |
641 | load = &lw; | |
642 | } | |
643 | slice = calc_delta_mine(slice, se->load.weight, load); | |
644 | } | |
645 | return slice; | |
bf0f6f24 IM |
646 | } |
647 | ||
647e7cac | 648 | /* |
ac884dec | 649 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 650 | * |
f9c0b095 | 651 | * vs = s/w |
647e7cac | 652 | */ |
f9c0b095 | 653 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 654 | { |
f9c0b095 | 655 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
656 | } |
657 | ||
d6b55918 | 658 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); |
6d5ab293 | 659 | static void update_cfs_shares(struct cfs_rq *cfs_rq); |
3b3d190e | 660 | |
bf0f6f24 IM |
661 | /* |
662 | * Update the current task's runtime statistics. Skip current tasks that | |
663 | * are not in our scheduling class. | |
664 | */ | |
665 | static inline void | |
8ebc91d9 IM |
666 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
667 | unsigned long delta_exec) | |
bf0f6f24 | 668 | { |
bbdba7c0 | 669 | unsigned long delta_exec_weighted; |
bf0f6f24 | 670 | |
41acab88 LDM |
671 | schedstat_set(curr->statistics.exec_max, |
672 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
673 | |
674 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 675 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 676 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 677 | |
e9acbff6 | 678 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 679 | update_min_vruntime(cfs_rq); |
3b3d190e | 680 | |
70caf8a6 | 681 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
3b3d190e | 682 | cfs_rq->load_unacc_exec_time += delta_exec; |
3b3d190e | 683 | #endif |
bf0f6f24 IM |
684 | } |
685 | ||
b7cc0896 | 686 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 687 | { |
429d43bc | 688 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 689 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
690 | unsigned long delta_exec; |
691 | ||
692 | if (unlikely(!curr)) | |
693 | return; | |
694 | ||
695 | /* | |
696 | * Get the amount of time the current task was running | |
697 | * since the last time we changed load (this cannot | |
698 | * overflow on 32 bits): | |
699 | */ | |
8ebc91d9 | 700 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
701 | if (!delta_exec) |
702 | return; | |
bf0f6f24 | 703 | |
8ebc91d9 IM |
704 | __update_curr(cfs_rq, curr, delta_exec); |
705 | curr->exec_start = now; | |
d842de87 SV |
706 | |
707 | if (entity_is_task(curr)) { | |
708 | struct task_struct *curtask = task_of(curr); | |
709 | ||
f977bb49 | 710 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 711 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 712 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 713 | } |
ec12cb7f PT |
714 | |
715 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
716 | } |
717 | ||
718 | static inline void | |
5870db5b | 719 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 720 | { |
41acab88 | 721 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
722 | } |
723 | ||
bf0f6f24 IM |
724 | /* |
725 | * Task is being enqueued - update stats: | |
726 | */ | |
d2417e5a | 727 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 728 | { |
bf0f6f24 IM |
729 | /* |
730 | * Are we enqueueing a waiting task? (for current tasks | |
731 | * a dequeue/enqueue event is a NOP) | |
732 | */ | |
429d43bc | 733 | if (se != cfs_rq->curr) |
5870db5b | 734 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
735 | } |
736 | ||
bf0f6f24 | 737 | static void |
9ef0a961 | 738 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 739 | { |
41acab88 LDM |
740 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
741 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
742 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
743 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
744 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
745 | #ifdef CONFIG_SCHEDSTATS |
746 | if (entity_is_task(se)) { | |
747 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 748 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
749 | } |
750 | #endif | |
41acab88 | 751 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
752 | } |
753 | ||
754 | static inline void | |
19b6a2e3 | 755 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 756 | { |
bf0f6f24 IM |
757 | /* |
758 | * Mark the end of the wait period if dequeueing a | |
759 | * waiting task: | |
760 | */ | |
429d43bc | 761 | if (se != cfs_rq->curr) |
9ef0a961 | 762 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
763 | } |
764 | ||
765 | /* | |
766 | * We are picking a new current task - update its stats: | |
767 | */ | |
768 | static inline void | |
79303e9e | 769 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
770 | { |
771 | /* | |
772 | * We are starting a new run period: | |
773 | */ | |
305e6835 | 774 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
775 | } |
776 | ||
bf0f6f24 IM |
777 | /************************************************** |
778 | * Scheduling class queueing methods: | |
779 | */ | |
780 | ||
cbee9f88 PZ |
781 | #ifdef CONFIG_NUMA_BALANCING |
782 | /* | |
783 | * numa task sample period in ms: 5s | |
784 | */ | |
785 | unsigned int sysctl_numa_balancing_scan_period_min = 5000; | |
786 | unsigned int sysctl_numa_balancing_scan_period_max = 5000*16; | |
787 | ||
788 | static void task_numa_placement(struct task_struct *p) | |
789 | { | |
790 | int seq = ACCESS_ONCE(p->mm->numa_scan_seq); | |
791 | ||
792 | if (p->numa_scan_seq == seq) | |
793 | return; | |
794 | p->numa_scan_seq = seq; | |
795 | ||
796 | /* FIXME: Scheduling placement policy hints go here */ | |
797 | } | |
798 | ||
799 | /* | |
800 | * Got a PROT_NONE fault for a page on @node. | |
801 | */ | |
802 | void task_numa_fault(int node, int pages) | |
803 | { | |
804 | struct task_struct *p = current; | |
805 | ||
806 | /* FIXME: Allocate task-specific structure for placement policy here */ | |
807 | ||
808 | task_numa_placement(p); | |
809 | } | |
810 | ||
811 | /* | |
812 | * The expensive part of numa migration is done from task_work context. | |
813 | * Triggered from task_tick_numa(). | |
814 | */ | |
815 | void task_numa_work(struct callback_head *work) | |
816 | { | |
817 | unsigned long migrate, next_scan, now = jiffies; | |
818 | struct task_struct *p = current; | |
819 | struct mm_struct *mm = p->mm; | |
820 | ||
821 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
822 | ||
823 | work->next = work; /* protect against double add */ | |
824 | /* | |
825 | * Who cares about NUMA placement when they're dying. | |
826 | * | |
827 | * NOTE: make sure not to dereference p->mm before this check, | |
828 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
829 | * without p->mm even though we still had it when we enqueued this | |
830 | * work. | |
831 | */ | |
832 | if (p->flags & PF_EXITING) | |
833 | return; | |
834 | ||
835 | /* | |
836 | * Enforce maximal scan/migration frequency.. | |
837 | */ | |
838 | migrate = mm->numa_next_scan; | |
839 | if (time_before(now, migrate)) | |
840 | return; | |
841 | ||
842 | if (p->numa_scan_period == 0) | |
843 | p->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
844 | ||
845 | next_scan = now + 2*msecs_to_jiffies(p->numa_scan_period); | |
846 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) | |
847 | return; | |
848 | ||
849 | ACCESS_ONCE(mm->numa_scan_seq)++; | |
850 | { | |
851 | struct vm_area_struct *vma; | |
852 | ||
853 | down_read(&mm->mmap_sem); | |
854 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
855 | if (!vma_migratable(vma)) | |
856 | continue; | |
857 | change_prot_numa(vma, vma->vm_start, vma->vm_end); | |
858 | } | |
859 | up_read(&mm->mmap_sem); | |
860 | } | |
861 | } | |
862 | ||
863 | /* | |
864 | * Drive the periodic memory faults.. | |
865 | */ | |
866 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
867 | { | |
868 | struct callback_head *work = &curr->numa_work; | |
869 | u64 period, now; | |
870 | ||
871 | /* | |
872 | * We don't care about NUMA placement if we don't have memory. | |
873 | */ | |
874 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
875 | return; | |
876 | ||
877 | /* | |
878 | * Using runtime rather than walltime has the dual advantage that | |
879 | * we (mostly) drive the selection from busy threads and that the | |
880 | * task needs to have done some actual work before we bother with | |
881 | * NUMA placement. | |
882 | */ | |
883 | now = curr->se.sum_exec_runtime; | |
884 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
885 | ||
886 | if (now - curr->node_stamp > period) { | |
887 | curr->node_stamp = now; | |
888 | ||
889 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
890 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
891 | task_work_add(curr, work, true); | |
892 | } | |
893 | } | |
894 | } | |
895 | #else | |
896 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
897 | { | |
898 | } | |
899 | #endif /* CONFIG_NUMA_BALANCING */ | |
900 | ||
30cfdcfc DA |
901 | static void |
902 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
903 | { | |
904 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 905 | if (!parent_entity(se)) |
029632fb | 906 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
907 | #ifdef CONFIG_SMP |
908 | if (entity_is_task(se)) | |
eb95308e | 909 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 910 | #endif |
30cfdcfc | 911 | cfs_rq->nr_running++; |
30cfdcfc DA |
912 | } |
913 | ||
914 | static void | |
915 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
916 | { | |
917 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 918 | if (!parent_entity(se)) |
029632fb | 919 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 920 | if (entity_is_task(se)) |
b87f1724 | 921 | list_del_init(&se->group_node); |
30cfdcfc | 922 | cfs_rq->nr_running--; |
30cfdcfc DA |
923 | } |
924 | ||
3ff6dcac | 925 | #ifdef CONFIG_FAIR_GROUP_SCHED |
64660c86 PT |
926 | /* we need this in update_cfs_load and load-balance functions below */ |
927 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); | |
3ff6dcac | 928 | # ifdef CONFIG_SMP |
d6b55918 PT |
929 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
930 | int global_update) | |
931 | { | |
932 | struct task_group *tg = cfs_rq->tg; | |
933 | long load_avg; | |
934 | ||
935 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
936 | load_avg -= cfs_rq->load_contribution; | |
937 | ||
938 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
939 | atomic_add(load_avg, &tg->load_weight); | |
940 | cfs_rq->load_contribution += load_avg; | |
941 | } | |
942 | } | |
943 | ||
944 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 945 | { |
a7a4f8a7 | 946 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 947 | u64 now, delta; |
e33078ba | 948 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 949 | |
64660c86 | 950 | if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq)) |
2069dd75 PZ |
951 | return; |
952 | ||
05ca62c6 | 953 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
954 | delta = now - cfs_rq->load_stamp; |
955 | ||
e33078ba PT |
956 | /* truncate load history at 4 idle periods */ |
957 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
958 | now - cfs_rq->load_last > 4 * period) { | |
959 | cfs_rq->load_period = 0; | |
960 | cfs_rq->load_avg = 0; | |
f07333bf | 961 | delta = period - 1; |
e33078ba PT |
962 | } |
963 | ||
2069dd75 | 964 | cfs_rq->load_stamp = now; |
3b3d190e | 965 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 966 | cfs_rq->load_period += delta; |
e33078ba PT |
967 | if (load) { |
968 | cfs_rq->load_last = now; | |
969 | cfs_rq->load_avg += delta * load; | |
970 | } | |
2069dd75 | 971 | |
d6b55918 PT |
972 | /* consider updating load contribution on each fold or truncate */ |
973 | if (global_update || cfs_rq->load_period > period | |
974 | || !cfs_rq->load_period) | |
975 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
976 | ||
2069dd75 PZ |
977 | while (cfs_rq->load_period > period) { |
978 | /* | |
979 | * Inline assembly required to prevent the compiler | |
980 | * optimising this loop into a divmod call. | |
981 | * See __iter_div_u64_rem() for another example of this. | |
982 | */ | |
983 | asm("" : "+rm" (cfs_rq->load_period)); | |
984 | cfs_rq->load_period /= 2; | |
985 | cfs_rq->load_avg /= 2; | |
986 | } | |
3d4b47b4 | 987 | |
e33078ba PT |
988 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
989 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
990 | } |
991 | ||
cf5f0acf PZ |
992 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
993 | { | |
994 | long tg_weight; | |
995 | ||
996 | /* | |
997 | * Use this CPU's actual weight instead of the last load_contribution | |
998 | * to gain a more accurate current total weight. See | |
999 | * update_cfs_rq_load_contribution(). | |
1000 | */ | |
1001 | tg_weight = atomic_read(&tg->load_weight); | |
1002 | tg_weight -= cfs_rq->load_contribution; | |
1003 | tg_weight += cfs_rq->load.weight; | |
1004 | ||
1005 | return tg_weight; | |
1006 | } | |
1007 | ||
6d5ab293 | 1008 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1009 | { |
cf5f0acf | 1010 | long tg_weight, load, shares; |
3ff6dcac | 1011 | |
cf5f0acf | 1012 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1013 | load = cfs_rq->load.weight; |
3ff6dcac | 1014 | |
3ff6dcac | 1015 | shares = (tg->shares * load); |
cf5f0acf PZ |
1016 | if (tg_weight) |
1017 | shares /= tg_weight; | |
3ff6dcac YZ |
1018 | |
1019 | if (shares < MIN_SHARES) | |
1020 | shares = MIN_SHARES; | |
1021 | if (shares > tg->shares) | |
1022 | shares = tg->shares; | |
1023 | ||
1024 | return shares; | |
1025 | } | |
1026 | ||
1027 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
1028 | { | |
1029 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
1030 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 1031 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
1032 | } |
1033 | } | |
1034 | # else /* CONFIG_SMP */ | |
1035 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
1036 | { | |
1037 | } | |
1038 | ||
6d5ab293 | 1039 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1040 | { |
1041 | return tg->shares; | |
1042 | } | |
1043 | ||
1044 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
1045 | { | |
1046 | } | |
1047 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
1048 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1049 | unsigned long weight) | |
1050 | { | |
19e5eebb PT |
1051 | if (se->on_rq) { |
1052 | /* commit outstanding execution time */ | |
1053 | if (cfs_rq->curr == se) | |
1054 | update_curr(cfs_rq); | |
2069dd75 | 1055 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1056 | } |
2069dd75 PZ |
1057 | |
1058 | update_load_set(&se->load, weight); | |
1059 | ||
1060 | if (se->on_rq) | |
1061 | account_entity_enqueue(cfs_rq, se); | |
1062 | } | |
1063 | ||
6d5ab293 | 1064 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1065 | { |
1066 | struct task_group *tg; | |
1067 | struct sched_entity *se; | |
3ff6dcac | 1068 | long shares; |
2069dd75 | 1069 | |
2069dd75 PZ |
1070 | tg = cfs_rq->tg; |
1071 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1072 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1073 | return; |
3ff6dcac YZ |
1074 | #ifndef CONFIG_SMP |
1075 | if (likely(se->load.weight == tg->shares)) | |
1076 | return; | |
1077 | #endif | |
6d5ab293 | 1078 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1079 | |
1080 | reweight_entity(cfs_rq_of(se), se, shares); | |
1081 | } | |
1082 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 1083 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
1084 | { |
1085 | } | |
1086 | ||
6d5ab293 | 1087 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1088 | { |
1089 | } | |
43365bd7 PT |
1090 | |
1091 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
1092 | { | |
1093 | } | |
2069dd75 PZ |
1094 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
1095 | ||
2396af69 | 1096 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1097 | { |
bf0f6f24 | 1098 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1099 | struct task_struct *tsk = NULL; |
1100 | ||
1101 | if (entity_is_task(se)) | |
1102 | tsk = task_of(se); | |
1103 | ||
41acab88 LDM |
1104 | if (se->statistics.sleep_start) { |
1105 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
1106 | |
1107 | if ((s64)delta < 0) | |
1108 | delta = 0; | |
1109 | ||
41acab88 LDM |
1110 | if (unlikely(delta > se->statistics.sleep_max)) |
1111 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1112 | |
8c79a045 | 1113 | se->statistics.sleep_start = 0; |
41acab88 | 1114 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1115 | |
768d0c27 | 1116 | if (tsk) { |
e414314c | 1117 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1118 | trace_sched_stat_sleep(tsk, delta); |
1119 | } | |
bf0f6f24 | 1120 | } |
41acab88 LDM |
1121 | if (se->statistics.block_start) { |
1122 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
1123 | |
1124 | if ((s64)delta < 0) | |
1125 | delta = 0; | |
1126 | ||
41acab88 LDM |
1127 | if (unlikely(delta > se->statistics.block_max)) |
1128 | se->statistics.block_max = delta; | |
bf0f6f24 | 1129 | |
8c79a045 | 1130 | se->statistics.block_start = 0; |
41acab88 | 1131 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1132 | |
e414314c | 1133 | if (tsk) { |
8f0dfc34 | 1134 | if (tsk->in_iowait) { |
41acab88 LDM |
1135 | se->statistics.iowait_sum += delta; |
1136 | se->statistics.iowait_count++; | |
768d0c27 | 1137 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1138 | } |
1139 | ||
b781a602 AV |
1140 | trace_sched_stat_blocked(tsk, delta); |
1141 | ||
e414314c PZ |
1142 | /* |
1143 | * Blocking time is in units of nanosecs, so shift by | |
1144 | * 20 to get a milliseconds-range estimation of the | |
1145 | * amount of time that the task spent sleeping: | |
1146 | */ | |
1147 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1148 | profile_hits(SLEEP_PROFILING, | |
1149 | (void *)get_wchan(tsk), | |
1150 | delta >> 20); | |
1151 | } | |
1152 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1153 | } |
bf0f6f24 IM |
1154 | } |
1155 | #endif | |
1156 | } | |
1157 | ||
ddc97297 PZ |
1158 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1159 | { | |
1160 | #ifdef CONFIG_SCHED_DEBUG | |
1161 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
1162 | ||
1163 | if (d < 0) | |
1164 | d = -d; | |
1165 | ||
1166 | if (d > 3*sysctl_sched_latency) | |
1167 | schedstat_inc(cfs_rq, nr_spread_over); | |
1168 | #endif | |
1169 | } | |
1170 | ||
aeb73b04 PZ |
1171 | static void |
1172 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
1173 | { | |
1af5f730 | 1174 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 1175 | |
2cb8600e PZ |
1176 | /* |
1177 | * The 'current' period is already promised to the current tasks, | |
1178 | * however the extra weight of the new task will slow them down a | |
1179 | * little, place the new task so that it fits in the slot that | |
1180 | * stays open at the end. | |
1181 | */ | |
94dfb5e7 | 1182 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 1183 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 1184 | |
a2e7a7eb | 1185 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 1186 | if (!initial) { |
a2e7a7eb | 1187 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 1188 | |
a2e7a7eb MG |
1189 | /* |
1190 | * Halve their sleep time's effect, to allow | |
1191 | * for a gentler effect of sleepers: | |
1192 | */ | |
1193 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
1194 | thresh >>= 1; | |
51e0304c | 1195 | |
a2e7a7eb | 1196 | vruntime -= thresh; |
aeb73b04 PZ |
1197 | } |
1198 | ||
b5d9d734 MG |
1199 | /* ensure we never gain time by being placed backwards. */ |
1200 | vruntime = max_vruntime(se->vruntime, vruntime); | |
1201 | ||
67e9fb2a | 1202 | se->vruntime = vruntime; |
aeb73b04 PZ |
1203 | } |
1204 | ||
d3d9dc33 PT |
1205 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
1206 | ||
bf0f6f24 | 1207 | static void |
88ec22d3 | 1208 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1209 | { |
88ec22d3 PZ |
1210 | /* |
1211 | * Update the normalized vruntime before updating min_vruntime | |
1212 | * through callig update_curr(). | |
1213 | */ | |
371fd7e7 | 1214 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
1215 | se->vruntime += cfs_rq->min_vruntime; |
1216 | ||
bf0f6f24 | 1217 | /* |
a2a2d680 | 1218 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1219 | */ |
b7cc0896 | 1220 | update_curr(cfs_rq); |
d6b55918 | 1221 | update_cfs_load(cfs_rq, 0); |
a992241d | 1222 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 1223 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 1224 | |
88ec22d3 | 1225 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 1226 | place_entity(cfs_rq, se, 0); |
2396af69 | 1227 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 1228 | } |
bf0f6f24 | 1229 | |
d2417e5a | 1230 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 1231 | check_spread(cfs_rq, se); |
83b699ed SV |
1232 | if (se != cfs_rq->curr) |
1233 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1234 | se->on_rq = 1; |
3d4b47b4 | 1235 | |
d3d9dc33 | 1236 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1237 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1238 | check_enqueue_throttle(cfs_rq); |
1239 | } | |
bf0f6f24 IM |
1240 | } |
1241 | ||
2c13c919 | 1242 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1243 | { |
2c13c919 RR |
1244 | for_each_sched_entity(se) { |
1245 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1246 | if (cfs_rq->last == se) | |
1247 | cfs_rq->last = NULL; | |
1248 | else | |
1249 | break; | |
1250 | } | |
1251 | } | |
2002c695 | 1252 | |
2c13c919 RR |
1253 | static void __clear_buddies_next(struct sched_entity *se) |
1254 | { | |
1255 | for_each_sched_entity(se) { | |
1256 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1257 | if (cfs_rq->next == se) | |
1258 | cfs_rq->next = NULL; | |
1259 | else | |
1260 | break; | |
1261 | } | |
2002c695 PZ |
1262 | } |
1263 | ||
ac53db59 RR |
1264 | static void __clear_buddies_skip(struct sched_entity *se) |
1265 | { | |
1266 | for_each_sched_entity(se) { | |
1267 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1268 | if (cfs_rq->skip == se) | |
1269 | cfs_rq->skip = NULL; | |
1270 | else | |
1271 | break; | |
1272 | } | |
1273 | } | |
1274 | ||
a571bbea PZ |
1275 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1276 | { | |
2c13c919 RR |
1277 | if (cfs_rq->last == se) |
1278 | __clear_buddies_last(se); | |
1279 | ||
1280 | if (cfs_rq->next == se) | |
1281 | __clear_buddies_next(se); | |
ac53db59 RR |
1282 | |
1283 | if (cfs_rq->skip == se) | |
1284 | __clear_buddies_skip(se); | |
a571bbea PZ |
1285 | } |
1286 | ||
6c16a6dc | 1287 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 1288 | |
bf0f6f24 | 1289 | static void |
371fd7e7 | 1290 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1291 | { |
a2a2d680 DA |
1292 | /* |
1293 | * Update run-time statistics of the 'current'. | |
1294 | */ | |
1295 | update_curr(cfs_rq); | |
1296 | ||
19b6a2e3 | 1297 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1298 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1299 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1300 | if (entity_is_task(se)) { |
1301 | struct task_struct *tsk = task_of(se); | |
1302 | ||
1303 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1304 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1305 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1306 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1307 | } |
db36cc7d | 1308 | #endif |
67e9fb2a PZ |
1309 | } |
1310 | ||
2002c695 | 1311 | clear_buddies(cfs_rq, se); |
4793241b | 1312 | |
83b699ed | 1313 | if (se != cfs_rq->curr) |
30cfdcfc | 1314 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1315 | se->on_rq = 0; |
d6b55918 | 1316 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1317 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1318 | |
1319 | /* | |
1320 | * Normalize the entity after updating the min_vruntime because the | |
1321 | * update can refer to the ->curr item and we need to reflect this | |
1322 | * movement in our normalized position. | |
1323 | */ | |
371fd7e7 | 1324 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1325 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 1326 | |
d8b4986d PT |
1327 | /* return excess runtime on last dequeue */ |
1328 | return_cfs_rq_runtime(cfs_rq); | |
1329 | ||
1e876231 PZ |
1330 | update_min_vruntime(cfs_rq); |
1331 | update_cfs_shares(cfs_rq); | |
bf0f6f24 IM |
1332 | } |
1333 | ||
1334 | /* | |
1335 | * Preempt the current task with a newly woken task if needed: | |
1336 | */ | |
7c92e54f | 1337 | static void |
2e09bf55 | 1338 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1339 | { |
11697830 | 1340 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
1341 | struct sched_entity *se; |
1342 | s64 delta; | |
11697830 | 1343 | |
6d0f0ebd | 1344 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1345 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1346 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1347 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1348 | /* |
1349 | * The current task ran long enough, ensure it doesn't get | |
1350 | * re-elected due to buddy favours. | |
1351 | */ | |
1352 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1353 | return; |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * Ensure that a task that missed wakeup preemption by a | |
1358 | * narrow margin doesn't have to wait for a full slice. | |
1359 | * This also mitigates buddy induced latencies under load. | |
1360 | */ | |
f685ceac MG |
1361 | if (delta_exec < sysctl_sched_min_granularity) |
1362 | return; | |
1363 | ||
f4cfb33e WX |
1364 | se = __pick_first_entity(cfs_rq); |
1365 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 1366 | |
f4cfb33e WX |
1367 | if (delta < 0) |
1368 | return; | |
d7d82944 | 1369 | |
f4cfb33e WX |
1370 | if (delta > ideal_runtime) |
1371 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
1372 | } |
1373 | ||
83b699ed | 1374 | static void |
8494f412 | 1375 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1376 | { |
83b699ed SV |
1377 | /* 'current' is not kept within the tree. */ |
1378 | if (se->on_rq) { | |
1379 | /* | |
1380 | * Any task has to be enqueued before it get to execute on | |
1381 | * a CPU. So account for the time it spent waiting on the | |
1382 | * runqueue. | |
1383 | */ | |
1384 | update_stats_wait_end(cfs_rq, se); | |
1385 | __dequeue_entity(cfs_rq, se); | |
1386 | } | |
1387 | ||
79303e9e | 1388 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1389 | cfs_rq->curr = se; |
eba1ed4b IM |
1390 | #ifdef CONFIG_SCHEDSTATS |
1391 | /* | |
1392 | * Track our maximum slice length, if the CPU's load is at | |
1393 | * least twice that of our own weight (i.e. dont track it | |
1394 | * when there are only lesser-weight tasks around): | |
1395 | */ | |
495eca49 | 1396 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1397 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1398 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1399 | } | |
1400 | #endif | |
4a55b450 | 1401 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1402 | } |
1403 | ||
3f3a4904 PZ |
1404 | static int |
1405 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1406 | ||
ac53db59 RR |
1407 | /* |
1408 | * Pick the next process, keeping these things in mind, in this order: | |
1409 | * 1) keep things fair between processes/task groups | |
1410 | * 2) pick the "next" process, since someone really wants that to run | |
1411 | * 3) pick the "last" process, for cache locality | |
1412 | * 4) do not run the "skip" process, if something else is available | |
1413 | */ | |
f4b6755f | 1414 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1415 | { |
ac53db59 | 1416 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1417 | struct sched_entity *left = se; |
f4b6755f | 1418 | |
ac53db59 RR |
1419 | /* |
1420 | * Avoid running the skip buddy, if running something else can | |
1421 | * be done without getting too unfair. | |
1422 | */ | |
1423 | if (cfs_rq->skip == se) { | |
1424 | struct sched_entity *second = __pick_next_entity(se); | |
1425 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1426 | se = second; | |
1427 | } | |
aa2ac252 | 1428 | |
f685ceac MG |
1429 | /* |
1430 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1431 | */ | |
1432 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1433 | se = cfs_rq->last; | |
1434 | ||
ac53db59 RR |
1435 | /* |
1436 | * Someone really wants this to run. If it's not unfair, run it. | |
1437 | */ | |
1438 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1439 | se = cfs_rq->next; | |
1440 | ||
f685ceac | 1441 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1442 | |
1443 | return se; | |
aa2ac252 PZ |
1444 | } |
1445 | ||
d3d9dc33 PT |
1446 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
1447 | ||
ab6cde26 | 1448 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1449 | { |
1450 | /* | |
1451 | * If still on the runqueue then deactivate_task() | |
1452 | * was not called and update_curr() has to be done: | |
1453 | */ | |
1454 | if (prev->on_rq) | |
b7cc0896 | 1455 | update_curr(cfs_rq); |
bf0f6f24 | 1456 | |
d3d9dc33 PT |
1457 | /* throttle cfs_rqs exceeding runtime */ |
1458 | check_cfs_rq_runtime(cfs_rq); | |
1459 | ||
ddc97297 | 1460 | check_spread(cfs_rq, prev); |
30cfdcfc | 1461 | if (prev->on_rq) { |
5870db5b | 1462 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1463 | /* Put 'current' back into the tree. */ |
1464 | __enqueue_entity(cfs_rq, prev); | |
1465 | } | |
429d43bc | 1466 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1467 | } |
1468 | ||
8f4d37ec PZ |
1469 | static void |
1470 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1471 | { |
bf0f6f24 | 1472 | /* |
30cfdcfc | 1473 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1474 | */ |
30cfdcfc | 1475 | update_curr(cfs_rq); |
bf0f6f24 | 1476 | |
43365bd7 PT |
1477 | /* |
1478 | * Update share accounting for long-running entities. | |
1479 | */ | |
1480 | update_entity_shares_tick(cfs_rq); | |
1481 | ||
8f4d37ec PZ |
1482 | #ifdef CONFIG_SCHED_HRTICK |
1483 | /* | |
1484 | * queued ticks are scheduled to match the slice, so don't bother | |
1485 | * validating it and just reschedule. | |
1486 | */ | |
983ed7a6 HH |
1487 | if (queued) { |
1488 | resched_task(rq_of(cfs_rq)->curr); | |
1489 | return; | |
1490 | } | |
8f4d37ec PZ |
1491 | /* |
1492 | * don't let the period tick interfere with the hrtick preemption | |
1493 | */ | |
1494 | if (!sched_feat(DOUBLE_TICK) && | |
1495 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1496 | return; | |
1497 | #endif | |
1498 | ||
2c2efaed | 1499 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1500 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1501 | } |
1502 | ||
ab84d31e PT |
1503 | |
1504 | /************************************************** | |
1505 | * CFS bandwidth control machinery | |
1506 | */ | |
1507 | ||
1508 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
1509 | |
1510 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 1511 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
1512 | |
1513 | static inline bool cfs_bandwidth_used(void) | |
1514 | { | |
c5905afb | 1515 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
1516 | } |
1517 | ||
1518 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
1519 | { | |
1520 | /* only need to count groups transitioning between enabled/!enabled */ | |
1521 | if (enabled && !was_enabled) | |
c5905afb | 1522 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 1523 | else if (!enabled && was_enabled) |
c5905afb | 1524 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
1525 | } |
1526 | #else /* HAVE_JUMP_LABEL */ | |
1527 | static bool cfs_bandwidth_used(void) | |
1528 | { | |
1529 | return true; | |
1530 | } | |
1531 | ||
1532 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
1533 | #endif /* HAVE_JUMP_LABEL */ | |
1534 | ||
ab84d31e PT |
1535 | /* |
1536 | * default period for cfs group bandwidth. | |
1537 | * default: 0.1s, units: nanoseconds | |
1538 | */ | |
1539 | static inline u64 default_cfs_period(void) | |
1540 | { | |
1541 | return 100000000ULL; | |
1542 | } | |
ec12cb7f PT |
1543 | |
1544 | static inline u64 sched_cfs_bandwidth_slice(void) | |
1545 | { | |
1546 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
1547 | } | |
1548 | ||
a9cf55b2 PT |
1549 | /* |
1550 | * Replenish runtime according to assigned quota and update expiration time. | |
1551 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
1552 | * additional synchronization around rq->lock. | |
1553 | * | |
1554 | * requires cfs_b->lock | |
1555 | */ | |
029632fb | 1556 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
1557 | { |
1558 | u64 now; | |
1559 | ||
1560 | if (cfs_b->quota == RUNTIME_INF) | |
1561 | return; | |
1562 | ||
1563 | now = sched_clock_cpu(smp_processor_id()); | |
1564 | cfs_b->runtime = cfs_b->quota; | |
1565 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
1566 | } | |
1567 | ||
029632fb PZ |
1568 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
1569 | { | |
1570 | return &tg->cfs_bandwidth; | |
1571 | } | |
1572 | ||
85dac906 PT |
1573 | /* returns 0 on failure to allocate runtime */ |
1574 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
1575 | { |
1576 | struct task_group *tg = cfs_rq->tg; | |
1577 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 1578 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
1579 | |
1580 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
1581 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
1582 | ||
1583 | raw_spin_lock(&cfs_b->lock); | |
1584 | if (cfs_b->quota == RUNTIME_INF) | |
1585 | amount = min_amount; | |
58088ad0 | 1586 | else { |
a9cf55b2 PT |
1587 | /* |
1588 | * If the bandwidth pool has become inactive, then at least one | |
1589 | * period must have elapsed since the last consumption. | |
1590 | * Refresh the global state and ensure bandwidth timer becomes | |
1591 | * active. | |
1592 | */ | |
1593 | if (!cfs_b->timer_active) { | |
1594 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 1595 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 1596 | } |
58088ad0 PT |
1597 | |
1598 | if (cfs_b->runtime > 0) { | |
1599 | amount = min(cfs_b->runtime, min_amount); | |
1600 | cfs_b->runtime -= amount; | |
1601 | cfs_b->idle = 0; | |
1602 | } | |
ec12cb7f | 1603 | } |
a9cf55b2 | 1604 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
1605 | raw_spin_unlock(&cfs_b->lock); |
1606 | ||
1607 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
1608 | /* |
1609 | * we may have advanced our local expiration to account for allowed | |
1610 | * spread between our sched_clock and the one on which runtime was | |
1611 | * issued. | |
1612 | */ | |
1613 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
1614 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
1615 | |
1616 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
1617 | } |
1618 | ||
a9cf55b2 PT |
1619 | /* |
1620 | * Note: This depends on the synchronization provided by sched_clock and the | |
1621 | * fact that rq->clock snapshots this value. | |
1622 | */ | |
1623 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 1624 | { |
a9cf55b2 PT |
1625 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
1626 | struct rq *rq = rq_of(cfs_rq); | |
1627 | ||
1628 | /* if the deadline is ahead of our clock, nothing to do */ | |
1629 | if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0)) | |
ec12cb7f PT |
1630 | return; |
1631 | ||
a9cf55b2 PT |
1632 | if (cfs_rq->runtime_remaining < 0) |
1633 | return; | |
1634 | ||
1635 | /* | |
1636 | * If the local deadline has passed we have to consider the | |
1637 | * possibility that our sched_clock is 'fast' and the global deadline | |
1638 | * has not truly expired. | |
1639 | * | |
1640 | * Fortunately we can check determine whether this the case by checking | |
1641 | * whether the global deadline has advanced. | |
1642 | */ | |
1643 | ||
1644 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
1645 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
1646 | cfs_rq->runtime_expires += TICK_NSEC; | |
1647 | } else { | |
1648 | /* global deadline is ahead, expiration has passed */ | |
1649 | cfs_rq->runtime_remaining = 0; | |
1650 | } | |
1651 | } | |
1652 | ||
1653 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1654 | unsigned long delta_exec) | |
1655 | { | |
1656 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 1657 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
1658 | expire_cfs_rq_runtime(cfs_rq); |
1659 | ||
1660 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
1661 | return; |
1662 | ||
85dac906 PT |
1663 | /* |
1664 | * if we're unable to extend our runtime we resched so that the active | |
1665 | * hierarchy can be throttled | |
1666 | */ | |
1667 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
1668 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
1669 | } |
1670 | ||
6c16a6dc PZ |
1671 | static __always_inline |
1672 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 1673 | { |
56f570e5 | 1674 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
1675 | return; |
1676 | ||
1677 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
1678 | } | |
1679 | ||
85dac906 PT |
1680 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
1681 | { | |
56f570e5 | 1682 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
1683 | } |
1684 | ||
64660c86 PT |
1685 | /* check whether cfs_rq, or any parent, is throttled */ |
1686 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
1687 | { | |
56f570e5 | 1688 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
1689 | } |
1690 | ||
1691 | /* | |
1692 | * Ensure that neither of the group entities corresponding to src_cpu or | |
1693 | * dest_cpu are members of a throttled hierarchy when performing group | |
1694 | * load-balance operations. | |
1695 | */ | |
1696 | static inline int throttled_lb_pair(struct task_group *tg, | |
1697 | int src_cpu, int dest_cpu) | |
1698 | { | |
1699 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
1700 | ||
1701 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
1702 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
1703 | ||
1704 | return throttled_hierarchy(src_cfs_rq) || | |
1705 | throttled_hierarchy(dest_cfs_rq); | |
1706 | } | |
1707 | ||
1708 | /* updated child weight may affect parent so we have to do this bottom up */ | |
1709 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
1710 | { | |
1711 | struct rq *rq = data; | |
1712 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1713 | ||
1714 | cfs_rq->throttle_count--; | |
1715 | #ifdef CONFIG_SMP | |
1716 | if (!cfs_rq->throttle_count) { | |
1717 | u64 delta = rq->clock_task - cfs_rq->load_stamp; | |
1718 | ||
1719 | /* leaving throttled state, advance shares averaging windows */ | |
1720 | cfs_rq->load_stamp += delta; | |
1721 | cfs_rq->load_last += delta; | |
1722 | ||
1723 | /* update entity weight now that we are on_rq again */ | |
1724 | update_cfs_shares(cfs_rq); | |
1725 | } | |
1726 | #endif | |
1727 | ||
1728 | return 0; | |
1729 | } | |
1730 | ||
1731 | static int tg_throttle_down(struct task_group *tg, void *data) | |
1732 | { | |
1733 | struct rq *rq = data; | |
1734 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1735 | ||
1736 | /* group is entering throttled state, record last load */ | |
1737 | if (!cfs_rq->throttle_count) | |
1738 | update_cfs_load(cfs_rq, 0); | |
1739 | cfs_rq->throttle_count++; | |
1740 | ||
1741 | return 0; | |
1742 | } | |
1743 | ||
d3d9dc33 | 1744 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
1745 | { |
1746 | struct rq *rq = rq_of(cfs_rq); | |
1747 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1748 | struct sched_entity *se; | |
1749 | long task_delta, dequeue = 1; | |
1750 | ||
1751 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1752 | ||
1753 | /* account load preceding throttle */ | |
64660c86 PT |
1754 | rcu_read_lock(); |
1755 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
1756 | rcu_read_unlock(); | |
85dac906 PT |
1757 | |
1758 | task_delta = cfs_rq->h_nr_running; | |
1759 | for_each_sched_entity(se) { | |
1760 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
1761 | /* throttled entity or throttle-on-deactivate */ | |
1762 | if (!se->on_rq) | |
1763 | break; | |
1764 | ||
1765 | if (dequeue) | |
1766 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
1767 | qcfs_rq->h_nr_running -= task_delta; | |
1768 | ||
1769 | if (qcfs_rq->load.weight) | |
1770 | dequeue = 0; | |
1771 | } | |
1772 | ||
1773 | if (!se) | |
1774 | rq->nr_running -= task_delta; | |
1775 | ||
1776 | cfs_rq->throttled = 1; | |
e8da1b18 | 1777 | cfs_rq->throttled_timestamp = rq->clock; |
85dac906 PT |
1778 | raw_spin_lock(&cfs_b->lock); |
1779 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
1780 | raw_spin_unlock(&cfs_b->lock); | |
1781 | } | |
1782 | ||
029632fb | 1783 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
1784 | { |
1785 | struct rq *rq = rq_of(cfs_rq); | |
1786 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1787 | struct sched_entity *se; | |
1788 | int enqueue = 1; | |
1789 | long task_delta; | |
1790 | ||
1791 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1792 | ||
1793 | cfs_rq->throttled = 0; | |
1794 | raw_spin_lock(&cfs_b->lock); | |
e8da1b18 | 1795 | cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp; |
671fd9da PT |
1796 | list_del_rcu(&cfs_rq->throttled_list); |
1797 | raw_spin_unlock(&cfs_b->lock); | |
e8da1b18 | 1798 | cfs_rq->throttled_timestamp = 0; |
671fd9da | 1799 | |
64660c86 PT |
1800 | update_rq_clock(rq); |
1801 | /* update hierarchical throttle state */ | |
1802 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
1803 | ||
671fd9da PT |
1804 | if (!cfs_rq->load.weight) |
1805 | return; | |
1806 | ||
1807 | task_delta = cfs_rq->h_nr_running; | |
1808 | for_each_sched_entity(se) { | |
1809 | if (se->on_rq) | |
1810 | enqueue = 0; | |
1811 | ||
1812 | cfs_rq = cfs_rq_of(se); | |
1813 | if (enqueue) | |
1814 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
1815 | cfs_rq->h_nr_running += task_delta; | |
1816 | ||
1817 | if (cfs_rq_throttled(cfs_rq)) | |
1818 | break; | |
1819 | } | |
1820 | ||
1821 | if (!se) | |
1822 | rq->nr_running += task_delta; | |
1823 | ||
1824 | /* determine whether we need to wake up potentially idle cpu */ | |
1825 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
1826 | resched_task(rq->curr); | |
1827 | } | |
1828 | ||
1829 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
1830 | u64 remaining, u64 expires) | |
1831 | { | |
1832 | struct cfs_rq *cfs_rq; | |
1833 | u64 runtime = remaining; | |
1834 | ||
1835 | rcu_read_lock(); | |
1836 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
1837 | throttled_list) { | |
1838 | struct rq *rq = rq_of(cfs_rq); | |
1839 | ||
1840 | raw_spin_lock(&rq->lock); | |
1841 | if (!cfs_rq_throttled(cfs_rq)) | |
1842 | goto next; | |
1843 | ||
1844 | runtime = -cfs_rq->runtime_remaining + 1; | |
1845 | if (runtime > remaining) | |
1846 | runtime = remaining; | |
1847 | remaining -= runtime; | |
1848 | ||
1849 | cfs_rq->runtime_remaining += runtime; | |
1850 | cfs_rq->runtime_expires = expires; | |
1851 | ||
1852 | /* we check whether we're throttled above */ | |
1853 | if (cfs_rq->runtime_remaining > 0) | |
1854 | unthrottle_cfs_rq(cfs_rq); | |
1855 | ||
1856 | next: | |
1857 | raw_spin_unlock(&rq->lock); | |
1858 | ||
1859 | if (!remaining) | |
1860 | break; | |
1861 | } | |
1862 | rcu_read_unlock(); | |
1863 | ||
1864 | return remaining; | |
1865 | } | |
1866 | ||
58088ad0 PT |
1867 | /* |
1868 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
1869 | * cfs_rqs as appropriate. If there has been no activity within the last | |
1870 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
1871 | * used to track this state. | |
1872 | */ | |
1873 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
1874 | { | |
671fd9da PT |
1875 | u64 runtime, runtime_expires; |
1876 | int idle = 1, throttled; | |
58088ad0 PT |
1877 | |
1878 | raw_spin_lock(&cfs_b->lock); | |
1879 | /* no need to continue the timer with no bandwidth constraint */ | |
1880 | if (cfs_b->quota == RUNTIME_INF) | |
1881 | goto out_unlock; | |
1882 | ||
671fd9da PT |
1883 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
1884 | /* idle depends on !throttled (for the case of a large deficit) */ | |
1885 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 1886 | cfs_b->nr_periods += overrun; |
671fd9da | 1887 | |
a9cf55b2 PT |
1888 | /* if we're going inactive then everything else can be deferred */ |
1889 | if (idle) | |
1890 | goto out_unlock; | |
1891 | ||
1892 | __refill_cfs_bandwidth_runtime(cfs_b); | |
1893 | ||
671fd9da PT |
1894 | if (!throttled) { |
1895 | /* mark as potentially idle for the upcoming period */ | |
1896 | cfs_b->idle = 1; | |
1897 | goto out_unlock; | |
1898 | } | |
1899 | ||
e8da1b18 NR |
1900 | /* account preceding periods in which throttling occurred */ |
1901 | cfs_b->nr_throttled += overrun; | |
1902 | ||
671fd9da PT |
1903 | /* |
1904 | * There are throttled entities so we must first use the new bandwidth | |
1905 | * to unthrottle them before making it generally available. This | |
1906 | * ensures that all existing debts will be paid before a new cfs_rq is | |
1907 | * allowed to run. | |
1908 | */ | |
1909 | runtime = cfs_b->runtime; | |
1910 | runtime_expires = cfs_b->runtime_expires; | |
1911 | cfs_b->runtime = 0; | |
1912 | ||
1913 | /* | |
1914 | * This check is repeated as we are holding onto the new bandwidth | |
1915 | * while we unthrottle. This can potentially race with an unthrottled | |
1916 | * group trying to acquire new bandwidth from the global pool. | |
1917 | */ | |
1918 | while (throttled && runtime > 0) { | |
1919 | raw_spin_unlock(&cfs_b->lock); | |
1920 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
1921 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
1922 | runtime_expires); | |
1923 | raw_spin_lock(&cfs_b->lock); | |
1924 | ||
1925 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
1926 | } | |
58088ad0 | 1927 | |
671fd9da PT |
1928 | /* return (any) remaining runtime */ |
1929 | cfs_b->runtime = runtime; | |
1930 | /* | |
1931 | * While we are ensured activity in the period following an | |
1932 | * unthrottle, this also covers the case in which the new bandwidth is | |
1933 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
1934 | * timer to remain active while there are any throttled entities.) | |
1935 | */ | |
1936 | cfs_b->idle = 0; | |
58088ad0 PT |
1937 | out_unlock: |
1938 | if (idle) | |
1939 | cfs_b->timer_active = 0; | |
1940 | raw_spin_unlock(&cfs_b->lock); | |
1941 | ||
1942 | return idle; | |
1943 | } | |
d3d9dc33 | 1944 | |
d8b4986d PT |
1945 | /* a cfs_rq won't donate quota below this amount */ |
1946 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
1947 | /* minimum remaining period time to redistribute slack quota */ | |
1948 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
1949 | /* how long we wait to gather additional slack before distributing */ | |
1950 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
1951 | ||
1952 | /* are we near the end of the current quota period? */ | |
1953 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
1954 | { | |
1955 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
1956 | u64 remaining; | |
1957 | ||
1958 | /* if the call-back is running a quota refresh is already occurring */ | |
1959 | if (hrtimer_callback_running(refresh_timer)) | |
1960 | return 1; | |
1961 | ||
1962 | /* is a quota refresh about to occur? */ | |
1963 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
1964 | if (remaining < min_expire) | |
1965 | return 1; | |
1966 | ||
1967 | return 0; | |
1968 | } | |
1969 | ||
1970 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
1971 | { | |
1972 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
1973 | ||
1974 | /* if there's a quota refresh soon don't bother with slack */ | |
1975 | if (runtime_refresh_within(cfs_b, min_left)) | |
1976 | return; | |
1977 | ||
1978 | start_bandwidth_timer(&cfs_b->slack_timer, | |
1979 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
1980 | } | |
1981 | ||
1982 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
1983 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
1984 | { | |
1985 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1986 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
1987 | ||
1988 | if (slack_runtime <= 0) | |
1989 | return; | |
1990 | ||
1991 | raw_spin_lock(&cfs_b->lock); | |
1992 | if (cfs_b->quota != RUNTIME_INF && | |
1993 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
1994 | cfs_b->runtime += slack_runtime; | |
1995 | ||
1996 | /* we are under rq->lock, defer unthrottling using a timer */ | |
1997 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
1998 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
1999 | start_cfs_slack_bandwidth(cfs_b); | |
2000 | } | |
2001 | raw_spin_unlock(&cfs_b->lock); | |
2002 | ||
2003 | /* even if it's not valid for return we don't want to try again */ | |
2004 | cfs_rq->runtime_remaining -= slack_runtime; | |
2005 | } | |
2006 | ||
2007 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2008 | { | |
56f570e5 PT |
2009 | if (!cfs_bandwidth_used()) |
2010 | return; | |
2011 | ||
fccfdc6f | 2012 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2013 | return; |
2014 | ||
2015 | __return_cfs_rq_runtime(cfs_rq); | |
2016 | } | |
2017 | ||
2018 | /* | |
2019 | * This is done with a timer (instead of inline with bandwidth return) since | |
2020 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2021 | */ | |
2022 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2023 | { | |
2024 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2025 | u64 expires; | |
2026 | ||
2027 | /* confirm we're still not at a refresh boundary */ | |
2028 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2029 | return; | |
2030 | ||
2031 | raw_spin_lock(&cfs_b->lock); | |
2032 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2033 | runtime = cfs_b->runtime; | |
2034 | cfs_b->runtime = 0; | |
2035 | } | |
2036 | expires = cfs_b->runtime_expires; | |
2037 | raw_spin_unlock(&cfs_b->lock); | |
2038 | ||
2039 | if (!runtime) | |
2040 | return; | |
2041 | ||
2042 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2043 | ||
2044 | raw_spin_lock(&cfs_b->lock); | |
2045 | if (expires == cfs_b->runtime_expires) | |
2046 | cfs_b->runtime = runtime; | |
2047 | raw_spin_unlock(&cfs_b->lock); | |
2048 | } | |
2049 | ||
d3d9dc33 PT |
2050 | /* |
2051 | * When a group wakes up we want to make sure that its quota is not already | |
2052 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2053 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2054 | */ | |
2055 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2056 | { | |
56f570e5 PT |
2057 | if (!cfs_bandwidth_used()) |
2058 | return; | |
2059 | ||
d3d9dc33 PT |
2060 | /* an active group must be handled by the update_curr()->put() path */ |
2061 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2062 | return; | |
2063 | ||
2064 | /* ensure the group is not already throttled */ | |
2065 | if (cfs_rq_throttled(cfs_rq)) | |
2066 | return; | |
2067 | ||
2068 | /* update runtime allocation */ | |
2069 | account_cfs_rq_runtime(cfs_rq, 0); | |
2070 | if (cfs_rq->runtime_remaining <= 0) | |
2071 | throttle_cfs_rq(cfs_rq); | |
2072 | } | |
2073 | ||
2074 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2075 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2076 | { | |
56f570e5 PT |
2077 | if (!cfs_bandwidth_used()) |
2078 | return; | |
2079 | ||
d3d9dc33 PT |
2080 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2081 | return; | |
2082 | ||
2083 | /* | |
2084 | * it's possible for a throttled entity to be forced into a running | |
2085 | * state (e.g. set_curr_task), in this case we're finished. | |
2086 | */ | |
2087 | if (cfs_rq_throttled(cfs_rq)) | |
2088 | return; | |
2089 | ||
2090 | throttle_cfs_rq(cfs_rq); | |
2091 | } | |
029632fb PZ |
2092 | |
2093 | static inline u64 default_cfs_period(void); | |
2094 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun); | |
2095 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b); | |
2096 | ||
2097 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) | |
2098 | { | |
2099 | struct cfs_bandwidth *cfs_b = | |
2100 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2101 | do_sched_cfs_slack_timer(cfs_b); | |
2102 | ||
2103 | return HRTIMER_NORESTART; | |
2104 | } | |
2105 | ||
2106 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2107 | { | |
2108 | struct cfs_bandwidth *cfs_b = | |
2109 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2110 | ktime_t now; | |
2111 | int overrun; | |
2112 | int idle = 0; | |
2113 | ||
2114 | for (;;) { | |
2115 | now = hrtimer_cb_get_time(timer); | |
2116 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2117 | ||
2118 | if (!overrun) | |
2119 | break; | |
2120 | ||
2121 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2122 | } | |
2123 | ||
2124 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2125 | } | |
2126 | ||
2127 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2128 | { | |
2129 | raw_spin_lock_init(&cfs_b->lock); | |
2130 | cfs_b->runtime = 0; | |
2131 | cfs_b->quota = RUNTIME_INF; | |
2132 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2133 | ||
2134 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2135 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2136 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2137 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2138 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2139 | } | |
2140 | ||
2141 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2142 | { | |
2143 | cfs_rq->runtime_enabled = 0; | |
2144 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2145 | } | |
2146 | ||
2147 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2148 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2149 | { | |
2150 | /* | |
2151 | * The timer may be active because we're trying to set a new bandwidth | |
2152 | * period or because we're racing with the tear-down path | |
2153 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2154 | * terminates). In either case we ensure that it's re-programmed | |
2155 | */ | |
2156 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2157 | raw_spin_unlock(&cfs_b->lock); | |
2158 | /* ensure cfs_b->lock is available while we wait */ | |
2159 | hrtimer_cancel(&cfs_b->period_timer); | |
2160 | ||
2161 | raw_spin_lock(&cfs_b->lock); | |
2162 | /* if someone else restarted the timer then we're done */ | |
2163 | if (cfs_b->timer_active) | |
2164 | return; | |
2165 | } | |
2166 | ||
2167 | cfs_b->timer_active = 1; | |
2168 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
2169 | } | |
2170 | ||
2171 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2172 | { | |
2173 | hrtimer_cancel(&cfs_b->period_timer); | |
2174 | hrtimer_cancel(&cfs_b->slack_timer); | |
2175 | } | |
2176 | ||
a4c96ae3 | 2177 | static void unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
2178 | { |
2179 | struct cfs_rq *cfs_rq; | |
2180 | ||
2181 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2182 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2183 | ||
2184 | if (!cfs_rq->runtime_enabled) | |
2185 | continue; | |
2186 | ||
2187 | /* | |
2188 | * clock_task is not advancing so we just need to make sure | |
2189 | * there's some valid quota amount | |
2190 | */ | |
2191 | cfs_rq->runtime_remaining = cfs_b->quota; | |
2192 | if (cfs_rq_throttled(cfs_rq)) | |
2193 | unthrottle_cfs_rq(cfs_rq); | |
2194 | } | |
2195 | } | |
2196 | ||
2197 | #else /* CONFIG_CFS_BANDWIDTH */ | |
6c16a6dc PZ |
2198 | static __always_inline |
2199 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {} | |
d3d9dc33 PT |
2200 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
2201 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 2202 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
2203 | |
2204 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
2205 | { | |
2206 | return 0; | |
2207 | } | |
64660c86 PT |
2208 | |
2209 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2210 | { | |
2211 | return 0; | |
2212 | } | |
2213 | ||
2214 | static inline int throttled_lb_pair(struct task_group *tg, | |
2215 | int src_cpu, int dest_cpu) | |
2216 | { | |
2217 | return 0; | |
2218 | } | |
029632fb PZ |
2219 | |
2220 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
2221 | ||
2222 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2223 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
2224 | #endif |
2225 | ||
029632fb PZ |
2226 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2227 | { | |
2228 | return NULL; | |
2229 | } | |
2230 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 2231 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
2232 | |
2233 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
2234 | ||
bf0f6f24 IM |
2235 | /************************************************** |
2236 | * CFS operations on tasks: | |
2237 | */ | |
2238 | ||
8f4d37ec PZ |
2239 | #ifdef CONFIG_SCHED_HRTICK |
2240 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2241 | { | |
8f4d37ec PZ |
2242 | struct sched_entity *se = &p->se; |
2243 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2244 | ||
2245 | WARN_ON(task_rq(p) != rq); | |
2246 | ||
b39e66ea | 2247 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
2248 | u64 slice = sched_slice(cfs_rq, se); |
2249 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
2250 | s64 delta = slice - ran; | |
2251 | ||
2252 | if (delta < 0) { | |
2253 | if (rq->curr == p) | |
2254 | resched_task(p); | |
2255 | return; | |
2256 | } | |
2257 | ||
2258 | /* | |
2259 | * Don't schedule slices shorter than 10000ns, that just | |
2260 | * doesn't make sense. Rely on vruntime for fairness. | |
2261 | */ | |
31656519 | 2262 | if (rq->curr != p) |
157124c1 | 2263 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 2264 | |
31656519 | 2265 | hrtick_start(rq, delta); |
8f4d37ec PZ |
2266 | } |
2267 | } | |
a4c2f00f PZ |
2268 | |
2269 | /* | |
2270 | * called from enqueue/dequeue and updates the hrtick when the | |
2271 | * current task is from our class and nr_running is low enough | |
2272 | * to matter. | |
2273 | */ | |
2274 | static void hrtick_update(struct rq *rq) | |
2275 | { | |
2276 | struct task_struct *curr = rq->curr; | |
2277 | ||
b39e66ea | 2278 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
2279 | return; |
2280 | ||
2281 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
2282 | hrtick_start_fair(rq, curr); | |
2283 | } | |
55e12e5e | 2284 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
2285 | static inline void |
2286 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2287 | { | |
2288 | } | |
a4c2f00f PZ |
2289 | |
2290 | static inline void hrtick_update(struct rq *rq) | |
2291 | { | |
2292 | } | |
8f4d37ec PZ |
2293 | #endif |
2294 | ||
bf0f6f24 IM |
2295 | /* |
2296 | * The enqueue_task method is called before nr_running is | |
2297 | * increased. Here we update the fair scheduling stats and | |
2298 | * then put the task into the rbtree: | |
2299 | */ | |
ea87bb78 | 2300 | static void |
371fd7e7 | 2301 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2302 | { |
2303 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2304 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
2305 | |
2306 | for_each_sched_entity(se) { | |
62fb1851 | 2307 | if (se->on_rq) |
bf0f6f24 IM |
2308 | break; |
2309 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 2310 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
2311 | |
2312 | /* | |
2313 | * end evaluation on encountering a throttled cfs_rq | |
2314 | * | |
2315 | * note: in the case of encountering a throttled cfs_rq we will | |
2316 | * post the final h_nr_running increment below. | |
2317 | */ | |
2318 | if (cfs_rq_throttled(cfs_rq)) | |
2319 | break; | |
953bfcd1 | 2320 | cfs_rq->h_nr_running++; |
85dac906 | 2321 | |
88ec22d3 | 2322 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 2323 | } |
8f4d37ec | 2324 | |
2069dd75 | 2325 | for_each_sched_entity(se) { |
0f317143 | 2326 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2327 | cfs_rq->h_nr_running++; |
2069dd75 | 2328 | |
85dac906 PT |
2329 | if (cfs_rq_throttled(cfs_rq)) |
2330 | break; | |
2331 | ||
d6b55918 | 2332 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 2333 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
2334 | } |
2335 | ||
85dac906 PT |
2336 | if (!se) |
2337 | inc_nr_running(rq); | |
a4c2f00f | 2338 | hrtick_update(rq); |
bf0f6f24 IM |
2339 | } |
2340 | ||
2f36825b VP |
2341 | static void set_next_buddy(struct sched_entity *se); |
2342 | ||
bf0f6f24 IM |
2343 | /* |
2344 | * The dequeue_task method is called before nr_running is | |
2345 | * decreased. We remove the task from the rbtree and | |
2346 | * update the fair scheduling stats: | |
2347 | */ | |
371fd7e7 | 2348 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2349 | { |
2350 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2351 | struct sched_entity *se = &p->se; |
2f36825b | 2352 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
2353 | |
2354 | for_each_sched_entity(se) { | |
2355 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 2356 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
2357 | |
2358 | /* | |
2359 | * end evaluation on encountering a throttled cfs_rq | |
2360 | * | |
2361 | * note: in the case of encountering a throttled cfs_rq we will | |
2362 | * post the final h_nr_running decrement below. | |
2363 | */ | |
2364 | if (cfs_rq_throttled(cfs_rq)) | |
2365 | break; | |
953bfcd1 | 2366 | cfs_rq->h_nr_running--; |
2069dd75 | 2367 | |
bf0f6f24 | 2368 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
2369 | if (cfs_rq->load.weight) { |
2370 | /* | |
2371 | * Bias pick_next to pick a task from this cfs_rq, as | |
2372 | * p is sleeping when it is within its sched_slice. | |
2373 | */ | |
2374 | if (task_sleep && parent_entity(se)) | |
2375 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
2376 | |
2377 | /* avoid re-evaluating load for this entity */ | |
2378 | se = parent_entity(se); | |
bf0f6f24 | 2379 | break; |
2f36825b | 2380 | } |
371fd7e7 | 2381 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 2382 | } |
8f4d37ec | 2383 | |
2069dd75 | 2384 | for_each_sched_entity(se) { |
0f317143 | 2385 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2386 | cfs_rq->h_nr_running--; |
2069dd75 | 2387 | |
85dac906 PT |
2388 | if (cfs_rq_throttled(cfs_rq)) |
2389 | break; | |
2390 | ||
d6b55918 | 2391 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 2392 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
2393 | } |
2394 | ||
85dac906 PT |
2395 | if (!se) |
2396 | dec_nr_running(rq); | |
a4c2f00f | 2397 | hrtick_update(rq); |
bf0f6f24 IM |
2398 | } |
2399 | ||
e7693a36 | 2400 | #ifdef CONFIG_SMP |
029632fb PZ |
2401 | /* Used instead of source_load when we know the type == 0 */ |
2402 | static unsigned long weighted_cpuload(const int cpu) | |
2403 | { | |
2404 | return cpu_rq(cpu)->load.weight; | |
2405 | } | |
2406 | ||
2407 | /* | |
2408 | * Return a low guess at the load of a migration-source cpu weighted | |
2409 | * according to the scheduling class and "nice" value. | |
2410 | * | |
2411 | * We want to under-estimate the load of migration sources, to | |
2412 | * balance conservatively. | |
2413 | */ | |
2414 | static unsigned long source_load(int cpu, int type) | |
2415 | { | |
2416 | struct rq *rq = cpu_rq(cpu); | |
2417 | unsigned long total = weighted_cpuload(cpu); | |
2418 | ||
2419 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2420 | return total; | |
2421 | ||
2422 | return min(rq->cpu_load[type-1], total); | |
2423 | } | |
2424 | ||
2425 | /* | |
2426 | * Return a high guess at the load of a migration-target cpu weighted | |
2427 | * according to the scheduling class and "nice" value. | |
2428 | */ | |
2429 | static unsigned long target_load(int cpu, int type) | |
2430 | { | |
2431 | struct rq *rq = cpu_rq(cpu); | |
2432 | unsigned long total = weighted_cpuload(cpu); | |
2433 | ||
2434 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2435 | return total; | |
2436 | ||
2437 | return max(rq->cpu_load[type-1], total); | |
2438 | } | |
2439 | ||
2440 | static unsigned long power_of(int cpu) | |
2441 | { | |
2442 | return cpu_rq(cpu)->cpu_power; | |
2443 | } | |
2444 | ||
2445 | static unsigned long cpu_avg_load_per_task(int cpu) | |
2446 | { | |
2447 | struct rq *rq = cpu_rq(cpu); | |
2448 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
2449 | ||
2450 | if (nr_running) | |
2451 | return rq->load.weight / nr_running; | |
2452 | ||
2453 | return 0; | |
2454 | } | |
2455 | ||
098fb9db | 2456 | |
74f8e4b2 | 2457 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
2458 | { |
2459 | struct sched_entity *se = &p->se; | |
2460 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
2461 | u64 min_vruntime; |
2462 | ||
2463 | #ifndef CONFIG_64BIT | |
2464 | u64 min_vruntime_copy; | |
88ec22d3 | 2465 | |
3fe1698b PZ |
2466 | do { |
2467 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
2468 | smp_rmb(); | |
2469 | min_vruntime = cfs_rq->min_vruntime; | |
2470 | } while (min_vruntime != min_vruntime_copy); | |
2471 | #else | |
2472 | min_vruntime = cfs_rq->min_vruntime; | |
2473 | #endif | |
88ec22d3 | 2474 | |
3fe1698b | 2475 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
2476 | } |
2477 | ||
bb3469ac | 2478 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
2479 | /* |
2480 | * effective_load() calculates the load change as seen from the root_task_group | |
2481 | * | |
2482 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
2483 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
2484 | * can calculate the shift in shares. | |
cf5f0acf PZ |
2485 | * |
2486 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
2487 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
2488 | * total group weight. | |
2489 | * | |
2490 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
2491 | * distribution (s_i) using: | |
2492 | * | |
2493 | * s_i = rw_i / \Sum rw_j (1) | |
2494 | * | |
2495 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
2496 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
2497 | * shares distribution (s_i): | |
2498 | * | |
2499 | * rw_i = { 2, 4, 1, 0 } | |
2500 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
2501 | * | |
2502 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
2503 | * task used to run on and the CPU the waker is running on), we need to | |
2504 | * compute the effect of waking a task on either CPU and, in case of a sync | |
2505 | * wakeup, compute the effect of the current task going to sleep. | |
2506 | * | |
2507 | * So for a change of @wl to the local @cpu with an overall group weight change | |
2508 | * of @wl we can compute the new shares distribution (s'_i) using: | |
2509 | * | |
2510 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
2511 | * | |
2512 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
2513 | * differences in waking a task to CPU 0. The additional task changes the | |
2514 | * weight and shares distributions like: | |
2515 | * | |
2516 | * rw'_i = { 3, 4, 1, 0 } | |
2517 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
2518 | * | |
2519 | * We can then compute the difference in effective weight by using: | |
2520 | * | |
2521 | * dw_i = S * (s'_i - s_i) (3) | |
2522 | * | |
2523 | * Where 'S' is the group weight as seen by its parent. | |
2524 | * | |
2525 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
2526 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
2527 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 2528 | */ |
2069dd75 | 2529 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 2530 | { |
4be9daaa | 2531 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 2532 | |
cf5f0acf | 2533 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
2534 | return wl; |
2535 | ||
4be9daaa | 2536 | for_each_sched_entity(se) { |
cf5f0acf | 2537 | long w, W; |
4be9daaa | 2538 | |
977dda7c | 2539 | tg = se->my_q->tg; |
bb3469ac | 2540 | |
cf5f0acf PZ |
2541 | /* |
2542 | * W = @wg + \Sum rw_j | |
2543 | */ | |
2544 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 2545 | |
cf5f0acf PZ |
2546 | /* |
2547 | * w = rw_i + @wl | |
2548 | */ | |
2549 | w = se->my_q->load.weight + wl; | |
940959e9 | 2550 | |
cf5f0acf PZ |
2551 | /* |
2552 | * wl = S * s'_i; see (2) | |
2553 | */ | |
2554 | if (W > 0 && w < W) | |
2555 | wl = (w * tg->shares) / W; | |
977dda7c PT |
2556 | else |
2557 | wl = tg->shares; | |
940959e9 | 2558 | |
cf5f0acf PZ |
2559 | /* |
2560 | * Per the above, wl is the new se->load.weight value; since | |
2561 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
2562 | * calc_cfs_shares(). | |
2563 | */ | |
977dda7c PT |
2564 | if (wl < MIN_SHARES) |
2565 | wl = MIN_SHARES; | |
cf5f0acf PZ |
2566 | |
2567 | /* | |
2568 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
2569 | */ | |
977dda7c | 2570 | wl -= se->load.weight; |
cf5f0acf PZ |
2571 | |
2572 | /* | |
2573 | * Recursively apply this logic to all parent groups to compute | |
2574 | * the final effective load change on the root group. Since | |
2575 | * only the @tg group gets extra weight, all parent groups can | |
2576 | * only redistribute existing shares. @wl is the shift in shares | |
2577 | * resulting from this level per the above. | |
2578 | */ | |
4be9daaa | 2579 | wg = 0; |
4be9daaa | 2580 | } |
bb3469ac | 2581 | |
4be9daaa | 2582 | return wl; |
bb3469ac PZ |
2583 | } |
2584 | #else | |
4be9daaa | 2585 | |
83378269 PZ |
2586 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
2587 | unsigned long wl, unsigned long wg) | |
4be9daaa | 2588 | { |
83378269 | 2589 | return wl; |
bb3469ac | 2590 | } |
4be9daaa | 2591 | |
bb3469ac PZ |
2592 | #endif |
2593 | ||
c88d5910 | 2594 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 2595 | { |
e37b6a7b | 2596 | s64 this_load, load; |
c88d5910 | 2597 | int idx, this_cpu, prev_cpu; |
098fb9db | 2598 | unsigned long tl_per_task; |
c88d5910 | 2599 | struct task_group *tg; |
83378269 | 2600 | unsigned long weight; |
b3137bc8 | 2601 | int balanced; |
098fb9db | 2602 | |
c88d5910 PZ |
2603 | idx = sd->wake_idx; |
2604 | this_cpu = smp_processor_id(); | |
2605 | prev_cpu = task_cpu(p); | |
2606 | load = source_load(prev_cpu, idx); | |
2607 | this_load = target_load(this_cpu, idx); | |
098fb9db | 2608 | |
b3137bc8 MG |
2609 | /* |
2610 | * If sync wakeup then subtract the (maximum possible) | |
2611 | * effect of the currently running task from the load | |
2612 | * of the current CPU: | |
2613 | */ | |
83378269 PZ |
2614 | if (sync) { |
2615 | tg = task_group(current); | |
2616 | weight = current->se.load.weight; | |
2617 | ||
c88d5910 | 2618 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
2619 | load += effective_load(tg, prev_cpu, 0, -weight); |
2620 | } | |
b3137bc8 | 2621 | |
83378269 PZ |
2622 | tg = task_group(p); |
2623 | weight = p->se.load.weight; | |
b3137bc8 | 2624 | |
71a29aa7 PZ |
2625 | /* |
2626 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
2627 | * due to the sync cause above having dropped this_load to 0, we'll |
2628 | * always have an imbalance, but there's really nothing you can do | |
2629 | * about that, so that's good too. | |
71a29aa7 PZ |
2630 | * |
2631 | * Otherwise check if either cpus are near enough in load to allow this | |
2632 | * task to be woken on this_cpu. | |
2633 | */ | |
e37b6a7b PT |
2634 | if (this_load > 0) { |
2635 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
2636 | |
2637 | this_eff_load = 100; | |
2638 | this_eff_load *= power_of(prev_cpu); | |
2639 | this_eff_load *= this_load + | |
2640 | effective_load(tg, this_cpu, weight, weight); | |
2641 | ||
2642 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
2643 | prev_eff_load *= power_of(this_cpu); | |
2644 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
2645 | ||
2646 | balanced = this_eff_load <= prev_eff_load; | |
2647 | } else | |
2648 | balanced = true; | |
b3137bc8 | 2649 | |
098fb9db | 2650 | /* |
4ae7d5ce IM |
2651 | * If the currently running task will sleep within |
2652 | * a reasonable amount of time then attract this newly | |
2653 | * woken task: | |
098fb9db | 2654 | */ |
2fb7635c PZ |
2655 | if (sync && balanced) |
2656 | return 1; | |
098fb9db | 2657 | |
41acab88 | 2658 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
2659 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
2660 | ||
c88d5910 PZ |
2661 | if (balanced || |
2662 | (this_load <= load && | |
2663 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
2664 | /* |
2665 | * This domain has SD_WAKE_AFFINE and | |
2666 | * p is cache cold in this domain, and | |
2667 | * there is no bad imbalance. | |
2668 | */ | |
c88d5910 | 2669 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 2670 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
2671 | |
2672 | return 1; | |
2673 | } | |
2674 | return 0; | |
2675 | } | |
2676 | ||
aaee1203 PZ |
2677 | /* |
2678 | * find_idlest_group finds and returns the least busy CPU group within the | |
2679 | * domain. | |
2680 | */ | |
2681 | static struct sched_group * | |
78e7ed53 | 2682 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 2683 | int this_cpu, int load_idx) |
e7693a36 | 2684 | { |
b3bd3de6 | 2685 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 2686 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 2687 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 2688 | |
aaee1203 PZ |
2689 | do { |
2690 | unsigned long load, avg_load; | |
2691 | int local_group; | |
2692 | int i; | |
e7693a36 | 2693 | |
aaee1203 PZ |
2694 | /* Skip over this group if it has no CPUs allowed */ |
2695 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 2696 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
2697 | continue; |
2698 | ||
2699 | local_group = cpumask_test_cpu(this_cpu, | |
2700 | sched_group_cpus(group)); | |
2701 | ||
2702 | /* Tally up the load of all CPUs in the group */ | |
2703 | avg_load = 0; | |
2704 | ||
2705 | for_each_cpu(i, sched_group_cpus(group)) { | |
2706 | /* Bias balancing toward cpus of our domain */ | |
2707 | if (local_group) | |
2708 | load = source_load(i, load_idx); | |
2709 | else | |
2710 | load = target_load(i, load_idx); | |
2711 | ||
2712 | avg_load += load; | |
2713 | } | |
2714 | ||
2715 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2716 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
2717 | |
2718 | if (local_group) { | |
2719 | this_load = avg_load; | |
aaee1203 PZ |
2720 | } else if (avg_load < min_load) { |
2721 | min_load = avg_load; | |
2722 | idlest = group; | |
2723 | } | |
2724 | } while (group = group->next, group != sd->groups); | |
2725 | ||
2726 | if (!idlest || 100*this_load < imbalance*min_load) | |
2727 | return NULL; | |
2728 | return idlest; | |
2729 | } | |
2730 | ||
2731 | /* | |
2732 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
2733 | */ | |
2734 | static int | |
2735 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
2736 | { | |
2737 | unsigned long load, min_load = ULONG_MAX; | |
2738 | int idlest = -1; | |
2739 | int i; | |
2740 | ||
2741 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 2742 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
2743 | load = weighted_cpuload(i); |
2744 | ||
2745 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2746 | min_load = load; | |
2747 | idlest = i; | |
e7693a36 GH |
2748 | } |
2749 | } | |
2750 | ||
aaee1203 PZ |
2751 | return idlest; |
2752 | } | |
e7693a36 | 2753 | |
a50bde51 PZ |
2754 | /* |
2755 | * Try and locate an idle CPU in the sched_domain. | |
2756 | */ | |
99bd5e2f | 2757 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
2758 | { |
2759 | int cpu = smp_processor_id(); | |
2760 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 2761 | struct sched_domain *sd; |
37407ea7 LT |
2762 | struct sched_group *sg; |
2763 | int i; | |
a50bde51 PZ |
2764 | |
2765 | /* | |
99bd5e2f SS |
2766 | * If the task is going to be woken-up on this cpu and if it is |
2767 | * already idle, then it is the right target. | |
a50bde51 | 2768 | */ |
99bd5e2f SS |
2769 | if (target == cpu && idle_cpu(cpu)) |
2770 | return cpu; | |
2771 | ||
2772 | /* | |
2773 | * If the task is going to be woken-up on the cpu where it previously | |
2774 | * ran and if it is currently idle, then it the right target. | |
2775 | */ | |
2776 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 2777 | return prev_cpu; |
a50bde51 PZ |
2778 | |
2779 | /* | |
37407ea7 | 2780 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 2781 | */ |
518cd623 | 2782 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 2783 | for_each_lower_domain(sd) { |
37407ea7 LT |
2784 | sg = sd->groups; |
2785 | do { | |
2786 | if (!cpumask_intersects(sched_group_cpus(sg), | |
2787 | tsk_cpus_allowed(p))) | |
2788 | goto next; | |
2789 | ||
2790 | for_each_cpu(i, sched_group_cpus(sg)) { | |
2791 | if (!idle_cpu(i)) | |
2792 | goto next; | |
2793 | } | |
970e1789 | 2794 | |
37407ea7 LT |
2795 | target = cpumask_first_and(sched_group_cpus(sg), |
2796 | tsk_cpus_allowed(p)); | |
2797 | goto done; | |
2798 | next: | |
2799 | sg = sg->next; | |
2800 | } while (sg != sd->groups); | |
2801 | } | |
2802 | done: | |
a50bde51 PZ |
2803 | return target; |
2804 | } | |
2805 | ||
aaee1203 PZ |
2806 | /* |
2807 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2808 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2809 | * SD_BALANCE_EXEC. | |
2810 | * | |
2811 | * Balance, ie. select the least loaded group. | |
2812 | * | |
2813 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2814 | * | |
2815 | * preempt must be disabled. | |
2816 | */ | |
0017d735 | 2817 | static int |
7608dec2 | 2818 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 2819 | { |
29cd8bae | 2820 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
2821 | int cpu = smp_processor_id(); |
2822 | int prev_cpu = task_cpu(p); | |
2823 | int new_cpu = cpu; | |
99bd5e2f | 2824 | int want_affine = 0; |
5158f4e4 | 2825 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 2826 | |
29baa747 | 2827 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
2828 | return prev_cpu; |
2829 | ||
0763a660 | 2830 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 2831 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
2832 | want_affine = 1; |
2833 | new_cpu = prev_cpu; | |
2834 | } | |
aaee1203 | 2835 | |
dce840a0 | 2836 | rcu_read_lock(); |
aaee1203 | 2837 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
2838 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
2839 | continue; | |
2840 | ||
fe3bcfe1 | 2841 | /* |
99bd5e2f SS |
2842 | * If both cpu and prev_cpu are part of this domain, |
2843 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 2844 | */ |
99bd5e2f SS |
2845 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
2846 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
2847 | affine_sd = tmp; | |
29cd8bae | 2848 | break; |
f03542a7 | 2849 | } |
29cd8bae | 2850 | |
f03542a7 | 2851 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
2852 | sd = tmp; |
2853 | } | |
2854 | ||
8b911acd | 2855 | if (affine_sd) { |
f03542a7 | 2856 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
2857 | prev_cpu = cpu; |
2858 | ||
2859 | new_cpu = select_idle_sibling(p, prev_cpu); | |
2860 | goto unlock; | |
8b911acd | 2861 | } |
e7693a36 | 2862 | |
aaee1203 | 2863 | while (sd) { |
5158f4e4 | 2864 | int load_idx = sd->forkexec_idx; |
aaee1203 | 2865 | struct sched_group *group; |
c88d5910 | 2866 | int weight; |
098fb9db | 2867 | |
0763a660 | 2868 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
2869 | sd = sd->child; |
2870 | continue; | |
2871 | } | |
098fb9db | 2872 | |
5158f4e4 PZ |
2873 | if (sd_flag & SD_BALANCE_WAKE) |
2874 | load_idx = sd->wake_idx; | |
098fb9db | 2875 | |
5158f4e4 | 2876 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
2877 | if (!group) { |
2878 | sd = sd->child; | |
2879 | continue; | |
2880 | } | |
4ae7d5ce | 2881 | |
d7c33c49 | 2882 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
2883 | if (new_cpu == -1 || new_cpu == cpu) { |
2884 | /* Now try balancing at a lower domain level of cpu */ | |
2885 | sd = sd->child; | |
2886 | continue; | |
e7693a36 | 2887 | } |
aaee1203 PZ |
2888 | |
2889 | /* Now try balancing at a lower domain level of new_cpu */ | |
2890 | cpu = new_cpu; | |
669c55e9 | 2891 | weight = sd->span_weight; |
aaee1203 PZ |
2892 | sd = NULL; |
2893 | for_each_domain(cpu, tmp) { | |
669c55e9 | 2894 | if (weight <= tmp->span_weight) |
aaee1203 | 2895 | break; |
0763a660 | 2896 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
2897 | sd = tmp; |
2898 | } | |
2899 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 2900 | } |
dce840a0 PZ |
2901 | unlock: |
2902 | rcu_read_unlock(); | |
e7693a36 | 2903 | |
c88d5910 | 2904 | return new_cpu; |
e7693a36 GH |
2905 | } |
2906 | #endif /* CONFIG_SMP */ | |
2907 | ||
e52fb7c0 PZ |
2908 | static unsigned long |
2909 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
2910 | { |
2911 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
2912 | ||
2913 | /* | |
e52fb7c0 PZ |
2914 | * Since its curr running now, convert the gran from real-time |
2915 | * to virtual-time in his units. | |
13814d42 MG |
2916 | * |
2917 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
2918 | * they get preempted easier. That is, if 'se' < 'curr' then | |
2919 | * the resulting gran will be larger, therefore penalizing the | |
2920 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
2921 | * be smaller, again penalizing the lighter task. | |
2922 | * | |
2923 | * This is especially important for buddies when the leftmost | |
2924 | * task is higher priority than the buddy. | |
0bbd3336 | 2925 | */ |
f4ad9bd2 | 2926 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
2927 | } |
2928 | ||
464b7527 PZ |
2929 | /* |
2930 | * Should 'se' preempt 'curr'. | |
2931 | * | |
2932 | * |s1 | |
2933 | * |s2 | |
2934 | * |s3 | |
2935 | * g | |
2936 | * |<--->|c | |
2937 | * | |
2938 | * w(c, s1) = -1 | |
2939 | * w(c, s2) = 0 | |
2940 | * w(c, s3) = 1 | |
2941 | * | |
2942 | */ | |
2943 | static int | |
2944 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
2945 | { | |
2946 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
2947 | ||
2948 | if (vdiff <= 0) | |
2949 | return -1; | |
2950 | ||
e52fb7c0 | 2951 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
2952 | if (vdiff > gran) |
2953 | return 1; | |
2954 | ||
2955 | return 0; | |
2956 | } | |
2957 | ||
02479099 PZ |
2958 | static void set_last_buddy(struct sched_entity *se) |
2959 | { | |
69c80f3e VP |
2960 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
2961 | return; | |
2962 | ||
2963 | for_each_sched_entity(se) | |
2964 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
2965 | } |
2966 | ||
2967 | static void set_next_buddy(struct sched_entity *se) | |
2968 | { | |
69c80f3e VP |
2969 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
2970 | return; | |
2971 | ||
2972 | for_each_sched_entity(se) | |
2973 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
2974 | } |
2975 | ||
ac53db59 RR |
2976 | static void set_skip_buddy(struct sched_entity *se) |
2977 | { | |
69c80f3e VP |
2978 | for_each_sched_entity(se) |
2979 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
2980 | } |
2981 | ||
bf0f6f24 IM |
2982 | /* |
2983 | * Preempt the current task with a newly woken task if needed: | |
2984 | */ | |
5a9b86f6 | 2985 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
2986 | { |
2987 | struct task_struct *curr = rq->curr; | |
8651a86c | 2988 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 2989 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 2990 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 2991 | int next_buddy_marked = 0; |
bf0f6f24 | 2992 | |
4ae7d5ce IM |
2993 | if (unlikely(se == pse)) |
2994 | return; | |
2995 | ||
5238cdd3 | 2996 | /* |
ddcdf6e7 | 2997 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
2998 | * unconditionally check_prempt_curr() after an enqueue (which may have |
2999 | * lead to a throttle). This both saves work and prevents false | |
3000 | * next-buddy nomination below. | |
3001 | */ | |
3002 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3003 | return; | |
3004 | ||
2f36825b | 3005 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3006 | set_next_buddy(pse); |
2f36825b VP |
3007 | next_buddy_marked = 1; |
3008 | } | |
57fdc26d | 3009 | |
aec0a514 BR |
3010 | /* |
3011 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3012 | * wake up path. | |
5238cdd3 PT |
3013 | * |
3014 | * Note: this also catches the edge-case of curr being in a throttled | |
3015 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3016 | * enqueue of curr) will have resulted in resched being set. This | |
3017 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3018 | * below. | |
aec0a514 BR |
3019 | */ |
3020 | if (test_tsk_need_resched(curr)) | |
3021 | return; | |
3022 | ||
a2f5c9ab DH |
3023 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3024 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3025 | likely(p->policy != SCHED_IDLE)) | |
3026 | goto preempt; | |
3027 | ||
91c234b4 | 3028 | /* |
a2f5c9ab DH |
3029 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3030 | * is driven by the tick): | |
91c234b4 | 3031 | */ |
6bc912b7 | 3032 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 3033 | return; |
bf0f6f24 | 3034 | |
464b7527 | 3035 | find_matching_se(&se, &pse); |
9bbd7374 | 3036 | update_curr(cfs_rq_of(se)); |
002f128b | 3037 | BUG_ON(!pse); |
2f36825b VP |
3038 | if (wakeup_preempt_entity(se, pse) == 1) { |
3039 | /* | |
3040 | * Bias pick_next to pick the sched entity that is | |
3041 | * triggering this preemption. | |
3042 | */ | |
3043 | if (!next_buddy_marked) | |
3044 | set_next_buddy(pse); | |
3a7e73a2 | 3045 | goto preempt; |
2f36825b | 3046 | } |
464b7527 | 3047 | |
3a7e73a2 | 3048 | return; |
a65ac745 | 3049 | |
3a7e73a2 PZ |
3050 | preempt: |
3051 | resched_task(curr); | |
3052 | /* | |
3053 | * Only set the backward buddy when the current task is still | |
3054 | * on the rq. This can happen when a wakeup gets interleaved | |
3055 | * with schedule on the ->pre_schedule() or idle_balance() | |
3056 | * point, either of which can * drop the rq lock. | |
3057 | * | |
3058 | * Also, during early boot the idle thread is in the fair class, | |
3059 | * for obvious reasons its a bad idea to schedule back to it. | |
3060 | */ | |
3061 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3062 | return; | |
3063 | ||
3064 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3065 | set_last_buddy(se); | |
bf0f6f24 IM |
3066 | } |
3067 | ||
fb8d4724 | 3068 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3069 | { |
8f4d37ec | 3070 | struct task_struct *p; |
bf0f6f24 IM |
3071 | struct cfs_rq *cfs_rq = &rq->cfs; |
3072 | struct sched_entity *se; | |
3073 | ||
36ace27e | 3074 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3075 | return NULL; |
3076 | ||
3077 | do { | |
9948f4b2 | 3078 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3079 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3080 | cfs_rq = group_cfs_rq(se); |
3081 | } while (cfs_rq); | |
3082 | ||
8f4d37ec | 3083 | p = task_of(se); |
b39e66ea MG |
3084 | if (hrtick_enabled(rq)) |
3085 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
3086 | |
3087 | return p; | |
bf0f6f24 IM |
3088 | } |
3089 | ||
3090 | /* | |
3091 | * Account for a descheduled task: | |
3092 | */ | |
31ee529c | 3093 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
3094 | { |
3095 | struct sched_entity *se = &prev->se; | |
3096 | struct cfs_rq *cfs_rq; | |
3097 | ||
3098 | for_each_sched_entity(se) { | |
3099 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 3100 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
3101 | } |
3102 | } | |
3103 | ||
ac53db59 RR |
3104 | /* |
3105 | * sched_yield() is very simple | |
3106 | * | |
3107 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
3108 | */ | |
3109 | static void yield_task_fair(struct rq *rq) | |
3110 | { | |
3111 | struct task_struct *curr = rq->curr; | |
3112 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
3113 | struct sched_entity *se = &curr->se; | |
3114 | ||
3115 | /* | |
3116 | * Are we the only task in the tree? | |
3117 | */ | |
3118 | if (unlikely(rq->nr_running == 1)) | |
3119 | return; | |
3120 | ||
3121 | clear_buddies(cfs_rq, se); | |
3122 | ||
3123 | if (curr->policy != SCHED_BATCH) { | |
3124 | update_rq_clock(rq); | |
3125 | /* | |
3126 | * Update run-time statistics of the 'current'. | |
3127 | */ | |
3128 | update_curr(cfs_rq); | |
916671c0 MG |
3129 | /* |
3130 | * Tell update_rq_clock() that we've just updated, | |
3131 | * so we don't do microscopic update in schedule() | |
3132 | * and double the fastpath cost. | |
3133 | */ | |
3134 | rq->skip_clock_update = 1; | |
ac53db59 RR |
3135 | } |
3136 | ||
3137 | set_skip_buddy(se); | |
3138 | } | |
3139 | ||
d95f4122 MG |
3140 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
3141 | { | |
3142 | struct sched_entity *se = &p->se; | |
3143 | ||
5238cdd3 PT |
3144 | /* throttled hierarchies are not runnable */ |
3145 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
3146 | return false; |
3147 | ||
3148 | /* Tell the scheduler that we'd really like pse to run next. */ | |
3149 | set_next_buddy(se); | |
3150 | ||
d95f4122 MG |
3151 | yield_task_fair(rq); |
3152 | ||
3153 | return true; | |
3154 | } | |
3155 | ||
681f3e68 | 3156 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
3157 | /************************************************** |
3158 | * Fair scheduling class load-balancing methods: | |
3159 | */ | |
3160 | ||
ed387b78 HS |
3161 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3162 | ||
ddcdf6e7 | 3163 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 3164 | #define LBF_NEED_BREAK 0x02 |
88b8dac0 | 3165 | #define LBF_SOME_PINNED 0x04 |
ddcdf6e7 PZ |
3166 | |
3167 | struct lb_env { | |
3168 | struct sched_domain *sd; | |
3169 | ||
ddcdf6e7 | 3170 | struct rq *src_rq; |
85c1e7da | 3171 | int src_cpu; |
ddcdf6e7 PZ |
3172 | |
3173 | int dst_cpu; | |
3174 | struct rq *dst_rq; | |
3175 | ||
88b8dac0 SV |
3176 | struct cpumask *dst_grpmask; |
3177 | int new_dst_cpu; | |
ddcdf6e7 | 3178 | enum cpu_idle_type idle; |
bd939f45 | 3179 | long imbalance; |
b9403130 MW |
3180 | /* The set of CPUs under consideration for load-balancing */ |
3181 | struct cpumask *cpus; | |
3182 | ||
ddcdf6e7 | 3183 | unsigned int flags; |
367456c7 PZ |
3184 | |
3185 | unsigned int loop; | |
3186 | unsigned int loop_break; | |
3187 | unsigned int loop_max; | |
ddcdf6e7 PZ |
3188 | }; |
3189 | ||
1e3c88bd | 3190 | /* |
ddcdf6e7 | 3191 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
3192 | * Both runqueues must be locked. |
3193 | */ | |
ddcdf6e7 | 3194 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 3195 | { |
ddcdf6e7 PZ |
3196 | deactivate_task(env->src_rq, p, 0); |
3197 | set_task_cpu(p, env->dst_cpu); | |
3198 | activate_task(env->dst_rq, p, 0); | |
3199 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
3200 | } |
3201 | ||
029632fb PZ |
3202 | /* |
3203 | * Is this task likely cache-hot: | |
3204 | */ | |
3205 | static int | |
3206 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
3207 | { | |
3208 | s64 delta; | |
3209 | ||
3210 | if (p->sched_class != &fair_sched_class) | |
3211 | return 0; | |
3212 | ||
3213 | if (unlikely(p->policy == SCHED_IDLE)) | |
3214 | return 0; | |
3215 | ||
3216 | /* | |
3217 | * Buddy candidates are cache hot: | |
3218 | */ | |
3219 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
3220 | (&p->se == cfs_rq_of(&p->se)->next || | |
3221 | &p->se == cfs_rq_of(&p->se)->last)) | |
3222 | return 1; | |
3223 | ||
3224 | if (sysctl_sched_migration_cost == -1) | |
3225 | return 1; | |
3226 | if (sysctl_sched_migration_cost == 0) | |
3227 | return 0; | |
3228 | ||
3229 | delta = now - p->se.exec_start; | |
3230 | ||
3231 | return delta < (s64)sysctl_sched_migration_cost; | |
3232 | } | |
3233 | ||
1e3c88bd PZ |
3234 | /* |
3235 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3236 | */ | |
3237 | static | |
8e45cb54 | 3238 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
3239 | { |
3240 | int tsk_cache_hot = 0; | |
3241 | /* | |
3242 | * We do not migrate tasks that are: | |
3243 | * 1) running (obviously), or | |
3244 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3245 | * 3) are cache-hot on their current CPU. | |
3246 | */ | |
ddcdf6e7 | 3247 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
88b8dac0 SV |
3248 | int new_dst_cpu; |
3249 | ||
41acab88 | 3250 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 SV |
3251 | |
3252 | /* | |
3253 | * Remember if this task can be migrated to any other cpu in | |
3254 | * our sched_group. We may want to revisit it if we couldn't | |
3255 | * meet load balance goals by pulling other tasks on src_cpu. | |
3256 | * | |
3257 | * Also avoid computing new_dst_cpu if we have already computed | |
3258 | * one in current iteration. | |
3259 | */ | |
3260 | if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED)) | |
3261 | return 0; | |
3262 | ||
3263 | new_dst_cpu = cpumask_first_and(env->dst_grpmask, | |
3264 | tsk_cpus_allowed(p)); | |
3265 | if (new_dst_cpu < nr_cpu_ids) { | |
3266 | env->flags |= LBF_SOME_PINNED; | |
3267 | env->new_dst_cpu = new_dst_cpu; | |
3268 | } | |
1e3c88bd PZ |
3269 | return 0; |
3270 | } | |
88b8dac0 SV |
3271 | |
3272 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 3273 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 3274 | |
ddcdf6e7 | 3275 | if (task_running(env->src_rq, p)) { |
41acab88 | 3276 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
3277 | return 0; |
3278 | } | |
3279 | ||
3280 | /* | |
3281 | * Aggressive migration if: | |
3282 | * 1) task is cache cold, or | |
3283 | * 2) too many balance attempts have failed. | |
3284 | */ | |
3285 | ||
ddcdf6e7 | 3286 | tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd); |
1e3c88bd | 3287 | if (!tsk_cache_hot || |
8e45cb54 | 3288 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
1e3c88bd PZ |
3289 | #ifdef CONFIG_SCHEDSTATS |
3290 | if (tsk_cache_hot) { | |
8e45cb54 | 3291 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 3292 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
3293 | } |
3294 | #endif | |
3295 | return 1; | |
3296 | } | |
3297 | ||
3298 | if (tsk_cache_hot) { | |
41acab88 | 3299 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
3300 | return 0; |
3301 | } | |
3302 | return 1; | |
3303 | } | |
3304 | ||
897c395f PZ |
3305 | /* |
3306 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3307 | * part of active balancing operations within "domain". | |
3308 | * Returns 1 if successful and 0 otherwise. | |
3309 | * | |
3310 | * Called with both runqueues locked. | |
3311 | */ | |
8e45cb54 | 3312 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
3313 | { |
3314 | struct task_struct *p, *n; | |
897c395f | 3315 | |
367456c7 PZ |
3316 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
3317 | if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu)) | |
3318 | continue; | |
897c395f | 3319 | |
367456c7 PZ |
3320 | if (!can_migrate_task(p, env)) |
3321 | continue; | |
897c395f | 3322 | |
367456c7 PZ |
3323 | move_task(p, env); |
3324 | /* | |
3325 | * Right now, this is only the second place move_task() | |
3326 | * is called, so we can safely collect move_task() | |
3327 | * stats here rather than inside move_task(). | |
3328 | */ | |
3329 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
3330 | return 1; | |
897c395f | 3331 | } |
897c395f PZ |
3332 | return 0; |
3333 | } | |
3334 | ||
367456c7 PZ |
3335 | static unsigned long task_h_load(struct task_struct *p); |
3336 | ||
eb95308e PZ |
3337 | static const unsigned int sched_nr_migrate_break = 32; |
3338 | ||
5d6523eb | 3339 | /* |
bd939f45 | 3340 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
3341 | * this_rq, as part of a balancing operation within domain "sd". |
3342 | * Returns 1 if successful and 0 otherwise. | |
3343 | * | |
3344 | * Called with both runqueues locked. | |
3345 | */ | |
3346 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 3347 | { |
5d6523eb PZ |
3348 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
3349 | struct task_struct *p; | |
367456c7 PZ |
3350 | unsigned long load; |
3351 | int pulled = 0; | |
1e3c88bd | 3352 | |
bd939f45 | 3353 | if (env->imbalance <= 0) |
5d6523eb | 3354 | return 0; |
1e3c88bd | 3355 | |
5d6523eb PZ |
3356 | while (!list_empty(tasks)) { |
3357 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 3358 | |
367456c7 PZ |
3359 | env->loop++; |
3360 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 3361 | if (env->loop > env->loop_max) |
367456c7 | 3362 | break; |
5d6523eb PZ |
3363 | |
3364 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 3365 | if (env->loop > env->loop_break) { |
eb95308e | 3366 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 3367 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 3368 | break; |
a195f004 | 3369 | } |
1e3c88bd | 3370 | |
5d6523eb | 3371 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
367456c7 PZ |
3372 | goto next; |
3373 | ||
3374 | load = task_h_load(p); | |
5d6523eb | 3375 | |
eb95308e | 3376 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
3377 | goto next; |
3378 | ||
bd939f45 | 3379 | if ((load / 2) > env->imbalance) |
367456c7 | 3380 | goto next; |
1e3c88bd | 3381 | |
367456c7 PZ |
3382 | if (!can_migrate_task(p, env)) |
3383 | goto next; | |
1e3c88bd | 3384 | |
ddcdf6e7 | 3385 | move_task(p, env); |
ee00e66f | 3386 | pulled++; |
bd939f45 | 3387 | env->imbalance -= load; |
1e3c88bd PZ |
3388 | |
3389 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
3390 | /* |
3391 | * NEWIDLE balancing is a source of latency, so preemptible | |
3392 | * kernels will stop after the first task is pulled to minimize | |
3393 | * the critical section. | |
3394 | */ | |
5d6523eb | 3395 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 3396 | break; |
1e3c88bd PZ |
3397 | #endif |
3398 | ||
ee00e66f PZ |
3399 | /* |
3400 | * We only want to steal up to the prescribed amount of | |
3401 | * weighted load. | |
3402 | */ | |
bd939f45 | 3403 | if (env->imbalance <= 0) |
ee00e66f | 3404 | break; |
367456c7 PZ |
3405 | |
3406 | continue; | |
3407 | next: | |
5d6523eb | 3408 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 3409 | } |
5d6523eb | 3410 | |
1e3c88bd | 3411 | /* |
ddcdf6e7 PZ |
3412 | * Right now, this is one of only two places move_task() is called, |
3413 | * so we can safely collect move_task() stats here rather than | |
3414 | * inside move_task(). | |
1e3c88bd | 3415 | */ |
8e45cb54 | 3416 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 3417 | |
5d6523eb | 3418 | return pulled; |
1e3c88bd PZ |
3419 | } |
3420 | ||
230059de | 3421 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
3422 | /* |
3423 | * update tg->load_weight by folding this cpu's load_avg | |
3424 | */ | |
67e86250 | 3425 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
3426 | { |
3427 | struct cfs_rq *cfs_rq; | |
3428 | unsigned long flags; | |
3429 | struct rq *rq; | |
9e3081ca PZ |
3430 | |
3431 | if (!tg->se[cpu]) | |
3432 | return 0; | |
3433 | ||
3434 | rq = cpu_rq(cpu); | |
3435 | cfs_rq = tg->cfs_rq[cpu]; | |
3436 | ||
3437 | raw_spin_lock_irqsave(&rq->lock, flags); | |
3438 | ||
3439 | update_rq_clock(rq); | |
d6b55918 | 3440 | update_cfs_load(cfs_rq, 1); |
9e3081ca PZ |
3441 | |
3442 | /* | |
3443 | * We need to update shares after updating tg->load_weight in | |
3444 | * order to adjust the weight of groups with long running tasks. | |
3445 | */ | |
6d5ab293 | 3446 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
3447 | |
3448 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
3449 | ||
3450 | return 0; | |
3451 | } | |
3452 | ||
3453 | static void update_shares(int cpu) | |
3454 | { | |
3455 | struct cfs_rq *cfs_rq; | |
3456 | struct rq *rq = cpu_rq(cpu); | |
3457 | ||
3458 | rcu_read_lock(); | |
9763b67f PZ |
3459 | /* |
3460 | * Iterates the task_group tree in a bottom up fashion, see | |
3461 | * list_add_leaf_cfs_rq() for details. | |
3462 | */ | |
64660c86 PT |
3463 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
3464 | /* throttled entities do not contribute to load */ | |
3465 | if (throttled_hierarchy(cfs_rq)) | |
3466 | continue; | |
3467 | ||
67e86250 | 3468 | update_shares_cpu(cfs_rq->tg, cpu); |
64660c86 | 3469 | } |
9e3081ca PZ |
3470 | rcu_read_unlock(); |
3471 | } | |
3472 | ||
9763b67f PZ |
3473 | /* |
3474 | * Compute the cpu's hierarchical load factor for each task group. | |
3475 | * This needs to be done in a top-down fashion because the load of a child | |
3476 | * group is a fraction of its parents load. | |
3477 | */ | |
3478 | static int tg_load_down(struct task_group *tg, void *data) | |
3479 | { | |
3480 | unsigned long load; | |
3481 | long cpu = (long)data; | |
3482 | ||
3483 | if (!tg->parent) { | |
3484 | load = cpu_rq(cpu)->load.weight; | |
3485 | } else { | |
3486 | load = tg->parent->cfs_rq[cpu]->h_load; | |
3487 | load *= tg->se[cpu]->load.weight; | |
3488 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
3489 | } | |
3490 | ||
3491 | tg->cfs_rq[cpu]->h_load = load; | |
3492 | ||
3493 | return 0; | |
3494 | } | |
3495 | ||
3496 | static void update_h_load(long cpu) | |
3497 | { | |
a35b6466 PZ |
3498 | struct rq *rq = cpu_rq(cpu); |
3499 | unsigned long now = jiffies; | |
3500 | ||
3501 | if (rq->h_load_throttle == now) | |
3502 | return; | |
3503 | ||
3504 | rq->h_load_throttle = now; | |
3505 | ||
367456c7 | 3506 | rcu_read_lock(); |
9763b67f | 3507 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
367456c7 | 3508 | rcu_read_unlock(); |
9763b67f PZ |
3509 | } |
3510 | ||
367456c7 | 3511 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 3512 | { |
367456c7 PZ |
3513 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
3514 | unsigned long load; | |
230059de | 3515 | |
367456c7 PZ |
3516 | load = p->se.load.weight; |
3517 | load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1); | |
230059de | 3518 | |
367456c7 | 3519 | return load; |
230059de PZ |
3520 | } |
3521 | #else | |
9e3081ca PZ |
3522 | static inline void update_shares(int cpu) |
3523 | { | |
3524 | } | |
3525 | ||
367456c7 | 3526 | static inline void update_h_load(long cpu) |
230059de | 3527 | { |
230059de | 3528 | } |
230059de | 3529 | |
367456c7 | 3530 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 3531 | { |
367456c7 | 3532 | return p->se.load.weight; |
1e3c88bd | 3533 | } |
230059de | 3534 | #endif |
1e3c88bd | 3535 | |
1e3c88bd PZ |
3536 | /********** Helpers for find_busiest_group ************************/ |
3537 | /* | |
3538 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
3539 | * during load balancing. | |
3540 | */ | |
3541 | struct sd_lb_stats { | |
3542 | struct sched_group *busiest; /* Busiest group in this sd */ | |
3543 | struct sched_group *this; /* Local group in this sd */ | |
3544 | unsigned long total_load; /* Total load of all groups in sd */ | |
3545 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
3546 | unsigned long avg_load; /* Average load across all groups in sd */ | |
3547 | ||
3548 | /** Statistics of this group */ | |
3549 | unsigned long this_load; | |
3550 | unsigned long this_load_per_task; | |
3551 | unsigned long this_nr_running; | |
fab47622 | 3552 | unsigned long this_has_capacity; |
aae6d3dd | 3553 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
3554 | |
3555 | /* Statistics of the busiest group */ | |
aae6d3dd | 3556 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
3557 | unsigned long max_load; |
3558 | unsigned long busiest_load_per_task; | |
3559 | unsigned long busiest_nr_running; | |
dd5feea1 | 3560 | unsigned long busiest_group_capacity; |
fab47622 | 3561 | unsigned long busiest_has_capacity; |
aae6d3dd | 3562 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
3563 | |
3564 | int group_imb; /* Is there imbalance in this sd */ | |
1e3c88bd PZ |
3565 | }; |
3566 | ||
3567 | /* | |
3568 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
3569 | */ | |
3570 | struct sg_lb_stats { | |
3571 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
3572 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
3573 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
3574 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
3575 | unsigned long group_capacity; | |
aae6d3dd SS |
3576 | unsigned long idle_cpus; |
3577 | unsigned long group_weight; | |
1e3c88bd | 3578 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 3579 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
3580 | }; |
3581 | ||
1e3c88bd PZ |
3582 | /** |
3583 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
3584 | * @sd: The sched_domain whose load_idx is to be obtained. | |
3585 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
3586 | */ | |
3587 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
3588 | enum cpu_idle_type idle) | |
3589 | { | |
3590 | int load_idx; | |
3591 | ||
3592 | switch (idle) { | |
3593 | case CPU_NOT_IDLE: | |
3594 | load_idx = sd->busy_idx; | |
3595 | break; | |
3596 | ||
3597 | case CPU_NEWLY_IDLE: | |
3598 | load_idx = sd->newidle_idx; | |
3599 | break; | |
3600 | default: | |
3601 | load_idx = sd->idle_idx; | |
3602 | break; | |
3603 | } | |
3604 | ||
3605 | return load_idx; | |
3606 | } | |
3607 | ||
1e3c88bd PZ |
3608 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
3609 | { | |
1399fa78 | 3610 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
3611 | } |
3612 | ||
3613 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
3614 | { | |
3615 | return default_scale_freq_power(sd, cpu); | |
3616 | } | |
3617 | ||
3618 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
3619 | { | |
669c55e9 | 3620 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
3621 | unsigned long smt_gain = sd->smt_gain; |
3622 | ||
3623 | smt_gain /= weight; | |
3624 | ||
3625 | return smt_gain; | |
3626 | } | |
3627 | ||
3628 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
3629 | { | |
3630 | return default_scale_smt_power(sd, cpu); | |
3631 | } | |
3632 | ||
3633 | unsigned long scale_rt_power(int cpu) | |
3634 | { | |
3635 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 3636 | u64 total, available, age_stamp, avg; |
1e3c88bd | 3637 | |
b654f7de PZ |
3638 | /* |
3639 | * Since we're reading these variables without serialization make sure | |
3640 | * we read them once before doing sanity checks on them. | |
3641 | */ | |
3642 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
3643 | avg = ACCESS_ONCE(rq->rt_avg); | |
3644 | ||
3645 | total = sched_avg_period() + (rq->clock - age_stamp); | |
aa483808 | 3646 | |
b654f7de | 3647 | if (unlikely(total < avg)) { |
aa483808 VP |
3648 | /* Ensures that power won't end up being negative */ |
3649 | available = 0; | |
3650 | } else { | |
b654f7de | 3651 | available = total - avg; |
aa483808 | 3652 | } |
1e3c88bd | 3653 | |
1399fa78 NR |
3654 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
3655 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 3656 | |
1399fa78 | 3657 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3658 | |
3659 | return div_u64(available, total); | |
3660 | } | |
3661 | ||
3662 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
3663 | { | |
669c55e9 | 3664 | unsigned long weight = sd->span_weight; |
1399fa78 | 3665 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
3666 | struct sched_group *sdg = sd->groups; |
3667 | ||
1e3c88bd PZ |
3668 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
3669 | if (sched_feat(ARCH_POWER)) | |
3670 | power *= arch_scale_smt_power(sd, cpu); | |
3671 | else | |
3672 | power *= default_scale_smt_power(sd, cpu); | |
3673 | ||
1399fa78 | 3674 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3675 | } |
3676 | ||
9c3f75cb | 3677 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
3678 | |
3679 | if (sched_feat(ARCH_POWER)) | |
3680 | power *= arch_scale_freq_power(sd, cpu); | |
3681 | else | |
3682 | power *= default_scale_freq_power(sd, cpu); | |
3683 | ||
1399fa78 | 3684 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 3685 | |
1e3c88bd | 3686 | power *= scale_rt_power(cpu); |
1399fa78 | 3687 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3688 | |
3689 | if (!power) | |
3690 | power = 1; | |
3691 | ||
e51fd5e2 | 3692 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 3693 | sdg->sgp->power = power; |
1e3c88bd PZ |
3694 | } |
3695 | ||
029632fb | 3696 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
3697 | { |
3698 | struct sched_domain *child = sd->child; | |
3699 | struct sched_group *group, *sdg = sd->groups; | |
3700 | unsigned long power; | |
4ec4412e VG |
3701 | unsigned long interval; |
3702 | ||
3703 | interval = msecs_to_jiffies(sd->balance_interval); | |
3704 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
3705 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
3706 | |
3707 | if (!child) { | |
3708 | update_cpu_power(sd, cpu); | |
3709 | return; | |
3710 | } | |
3711 | ||
3712 | power = 0; | |
3713 | ||
74a5ce20 PZ |
3714 | if (child->flags & SD_OVERLAP) { |
3715 | /* | |
3716 | * SD_OVERLAP domains cannot assume that child groups | |
3717 | * span the current group. | |
3718 | */ | |
3719 | ||
3720 | for_each_cpu(cpu, sched_group_cpus(sdg)) | |
3721 | power += power_of(cpu); | |
3722 | } else { | |
3723 | /* | |
3724 | * !SD_OVERLAP domains can assume that child groups | |
3725 | * span the current group. | |
3726 | */ | |
3727 | ||
3728 | group = child->groups; | |
3729 | do { | |
3730 | power += group->sgp->power; | |
3731 | group = group->next; | |
3732 | } while (group != child->groups); | |
3733 | } | |
1e3c88bd | 3734 | |
c3decf0d | 3735 | sdg->sgp->power_orig = sdg->sgp->power = power; |
1e3c88bd PZ |
3736 | } |
3737 | ||
9d5efe05 SV |
3738 | /* |
3739 | * Try and fix up capacity for tiny siblings, this is needed when | |
3740 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
3741 | * which on its own isn't powerful enough. | |
3742 | * | |
3743 | * See update_sd_pick_busiest() and check_asym_packing(). | |
3744 | */ | |
3745 | static inline int | |
3746 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
3747 | { | |
3748 | /* | |
1399fa78 | 3749 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 3750 | */ |
a6c75f2f | 3751 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
3752 | return 0; |
3753 | ||
3754 | /* | |
3755 | * If ~90% of the cpu_power is still there, we're good. | |
3756 | */ | |
9c3f75cb | 3757 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
3758 | return 1; |
3759 | ||
3760 | return 0; | |
3761 | } | |
3762 | ||
1e3c88bd PZ |
3763 | /** |
3764 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 3765 | * @env: The load balancing environment. |
1e3c88bd | 3766 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 3767 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 3768 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
3769 | * @balance: Should we balance. |
3770 | * @sgs: variable to hold the statistics for this group. | |
3771 | */ | |
bd939f45 PZ |
3772 | static inline void update_sg_lb_stats(struct lb_env *env, |
3773 | struct sched_group *group, int load_idx, | |
b9403130 | 3774 | int local_group, int *balance, struct sg_lb_stats *sgs) |
1e3c88bd | 3775 | { |
e44bc5c5 PZ |
3776 | unsigned long nr_running, max_nr_running, min_nr_running; |
3777 | unsigned long load, max_cpu_load, min_cpu_load; | |
04f733b4 | 3778 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
dd5feea1 | 3779 | unsigned long avg_load_per_task = 0; |
bd939f45 | 3780 | int i; |
1e3c88bd | 3781 | |
871e35bc | 3782 | if (local_group) |
c1174876 | 3783 | balance_cpu = group_balance_cpu(group); |
1e3c88bd PZ |
3784 | |
3785 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
3786 | max_cpu_load = 0; |
3787 | min_cpu_load = ~0UL; | |
2582f0eb | 3788 | max_nr_running = 0; |
e44bc5c5 | 3789 | min_nr_running = ~0UL; |
1e3c88bd | 3790 | |
b9403130 | 3791 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
3792 | struct rq *rq = cpu_rq(i); |
3793 | ||
e44bc5c5 PZ |
3794 | nr_running = rq->nr_running; |
3795 | ||
1e3c88bd PZ |
3796 | /* Bias balancing toward cpus of our domain */ |
3797 | if (local_group) { | |
c1174876 PZ |
3798 | if (idle_cpu(i) && !first_idle_cpu && |
3799 | cpumask_test_cpu(i, sched_group_mask(group))) { | |
04f733b4 | 3800 | first_idle_cpu = 1; |
1e3c88bd PZ |
3801 | balance_cpu = i; |
3802 | } | |
04f733b4 PZ |
3803 | |
3804 | load = target_load(i, load_idx); | |
1e3c88bd PZ |
3805 | } else { |
3806 | load = source_load(i, load_idx); | |
e44bc5c5 | 3807 | if (load > max_cpu_load) |
1e3c88bd PZ |
3808 | max_cpu_load = load; |
3809 | if (min_cpu_load > load) | |
3810 | min_cpu_load = load; | |
e44bc5c5 PZ |
3811 | |
3812 | if (nr_running > max_nr_running) | |
3813 | max_nr_running = nr_running; | |
3814 | if (min_nr_running > nr_running) | |
3815 | min_nr_running = nr_running; | |
1e3c88bd PZ |
3816 | } |
3817 | ||
3818 | sgs->group_load += load; | |
e44bc5c5 | 3819 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 3820 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
3821 | if (idle_cpu(i)) |
3822 | sgs->idle_cpus++; | |
1e3c88bd PZ |
3823 | } |
3824 | ||
3825 | /* | |
3826 | * First idle cpu or the first cpu(busiest) in this sched group | |
3827 | * is eligible for doing load balancing at this and above | |
3828 | * domains. In the newly idle case, we will allow all the cpu's | |
3829 | * to do the newly idle load balance. | |
3830 | */ | |
4ec4412e | 3831 | if (local_group) { |
bd939f45 | 3832 | if (env->idle != CPU_NEWLY_IDLE) { |
04f733b4 | 3833 | if (balance_cpu != env->dst_cpu) { |
4ec4412e VG |
3834 | *balance = 0; |
3835 | return; | |
3836 | } | |
bd939f45 | 3837 | update_group_power(env->sd, env->dst_cpu); |
4ec4412e | 3838 | } else if (time_after_eq(jiffies, group->sgp->next_update)) |
bd939f45 | 3839 | update_group_power(env->sd, env->dst_cpu); |
1e3c88bd PZ |
3840 | } |
3841 | ||
3842 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3843 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 3844 | |
1e3c88bd PZ |
3845 | /* |
3846 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 3847 | * than the average weight of a task. |
1e3c88bd PZ |
3848 | * |
3849 | * APZ: with cgroup the avg task weight can vary wildly and | |
3850 | * might not be a suitable number - should we keep a | |
3851 | * normalized nr_running number somewhere that negates | |
3852 | * the hierarchy? | |
3853 | */ | |
dd5feea1 SS |
3854 | if (sgs->sum_nr_running) |
3855 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 3856 | |
e44bc5c5 PZ |
3857 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && |
3858 | (max_nr_running - min_nr_running) > 1) | |
1e3c88bd PZ |
3859 | sgs->group_imb = 1; |
3860 | ||
9c3f75cb | 3861 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 3862 | SCHED_POWER_SCALE); |
9d5efe05 | 3863 | if (!sgs->group_capacity) |
bd939f45 | 3864 | sgs->group_capacity = fix_small_capacity(env->sd, group); |
aae6d3dd | 3865 | sgs->group_weight = group->group_weight; |
fab47622 NR |
3866 | |
3867 | if (sgs->group_capacity > sgs->sum_nr_running) | |
3868 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
3869 | } |
3870 | ||
532cb4c4 MN |
3871 | /** |
3872 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 3873 | * @env: The load balancing environment. |
532cb4c4 MN |
3874 | * @sds: sched_domain statistics |
3875 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 3876 | * @sgs: sched_group statistics |
532cb4c4 MN |
3877 | * |
3878 | * Determine if @sg is a busier group than the previously selected | |
3879 | * busiest group. | |
3880 | */ | |
bd939f45 | 3881 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
3882 | struct sd_lb_stats *sds, |
3883 | struct sched_group *sg, | |
bd939f45 | 3884 | struct sg_lb_stats *sgs) |
532cb4c4 MN |
3885 | { |
3886 | if (sgs->avg_load <= sds->max_load) | |
3887 | return false; | |
3888 | ||
3889 | if (sgs->sum_nr_running > sgs->group_capacity) | |
3890 | return true; | |
3891 | ||
3892 | if (sgs->group_imb) | |
3893 | return true; | |
3894 | ||
3895 | /* | |
3896 | * ASYM_PACKING needs to move all the work to the lowest | |
3897 | * numbered CPUs in the group, therefore mark all groups | |
3898 | * higher than ourself as busy. | |
3899 | */ | |
bd939f45 PZ |
3900 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
3901 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
3902 | if (!sds->busiest) |
3903 | return true; | |
3904 | ||
3905 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
3906 | return true; | |
3907 | } | |
3908 | ||
3909 | return false; | |
3910 | } | |
3911 | ||
1e3c88bd | 3912 | /** |
461819ac | 3913 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 3914 | * @env: The load balancing environment. |
1e3c88bd PZ |
3915 | * @balance: Should we balance. |
3916 | * @sds: variable to hold the statistics for this sched_domain. | |
3917 | */ | |
bd939f45 | 3918 | static inline void update_sd_lb_stats(struct lb_env *env, |
b9403130 | 3919 | int *balance, struct sd_lb_stats *sds) |
1e3c88bd | 3920 | { |
bd939f45 PZ |
3921 | struct sched_domain *child = env->sd->child; |
3922 | struct sched_group *sg = env->sd->groups; | |
1e3c88bd PZ |
3923 | struct sg_lb_stats sgs; |
3924 | int load_idx, prefer_sibling = 0; | |
3925 | ||
3926 | if (child && child->flags & SD_PREFER_SIBLING) | |
3927 | prefer_sibling = 1; | |
3928 | ||
bd939f45 | 3929 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
3930 | |
3931 | do { | |
3932 | int local_group; | |
3933 | ||
bd939f45 | 3934 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
1e3c88bd | 3935 | memset(&sgs, 0, sizeof(sgs)); |
b9403130 | 3936 | update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs); |
1e3c88bd | 3937 | |
8f190fb3 | 3938 | if (local_group && !(*balance)) |
1e3c88bd PZ |
3939 | return; |
3940 | ||
3941 | sds->total_load += sgs.group_load; | |
9c3f75cb | 3942 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
3943 | |
3944 | /* | |
3945 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 3946 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
3947 | * and move all the excess tasks away. We lower the capacity |
3948 | * of a group only if the local group has the capacity to fit | |
3949 | * these excess tasks, i.e. nr_running < group_capacity. The | |
3950 | * extra check prevents the case where you always pull from the | |
3951 | * heaviest group when it is already under-utilized (possible | |
3952 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 3953 | */ |
75dd321d | 3954 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
3955 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
3956 | ||
3957 | if (local_group) { | |
3958 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 3959 | sds->this = sg; |
1e3c88bd PZ |
3960 | sds->this_nr_running = sgs.sum_nr_running; |
3961 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 3962 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 3963 | sds->this_idle_cpus = sgs.idle_cpus; |
bd939f45 | 3964 | } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) { |
1e3c88bd | 3965 | sds->max_load = sgs.avg_load; |
532cb4c4 | 3966 | sds->busiest = sg; |
1e3c88bd | 3967 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 3968 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 3969 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 3970 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 3971 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 3972 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
3973 | sds->group_imb = sgs.group_imb; |
3974 | } | |
3975 | ||
532cb4c4 | 3976 | sg = sg->next; |
bd939f45 | 3977 | } while (sg != env->sd->groups); |
532cb4c4 MN |
3978 | } |
3979 | ||
532cb4c4 MN |
3980 | /** |
3981 | * check_asym_packing - Check to see if the group is packed into the | |
3982 | * sched doman. | |
3983 | * | |
3984 | * This is primarily intended to used at the sibling level. Some | |
3985 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
3986 | * case of POWER7, it can move to lower SMT modes only when higher | |
3987 | * threads are idle. When in lower SMT modes, the threads will | |
3988 | * perform better since they share less core resources. Hence when we | |
3989 | * have idle threads, we want them to be the higher ones. | |
3990 | * | |
3991 | * This packing function is run on idle threads. It checks to see if | |
3992 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
3993 | * CPU number than the packing function is being run on. Here we are | |
3994 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
3995 | * number. | |
3996 | * | |
b6b12294 MN |
3997 | * Returns 1 when packing is required and a task should be moved to |
3998 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
3999 | * | |
cd96891d | 4000 | * @env: The load balancing environment. |
532cb4c4 | 4001 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 4002 | */ |
bd939f45 | 4003 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
4004 | { |
4005 | int busiest_cpu; | |
4006 | ||
bd939f45 | 4007 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
4008 | return 0; |
4009 | ||
4010 | if (!sds->busiest) | |
4011 | return 0; | |
4012 | ||
4013 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 4014 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
4015 | return 0; |
4016 | ||
bd939f45 PZ |
4017 | env->imbalance = DIV_ROUND_CLOSEST( |
4018 | sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE); | |
4019 | ||
532cb4c4 | 4020 | return 1; |
1e3c88bd PZ |
4021 | } |
4022 | ||
4023 | /** | |
4024 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
4025 | * amongst the groups of a sched_domain, during | |
4026 | * load balancing. | |
cd96891d | 4027 | * @env: The load balancing environment. |
1e3c88bd | 4028 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4029 | */ |
bd939f45 PZ |
4030 | static inline |
4031 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
4032 | { |
4033 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
4034 | unsigned int imbn = 2; | |
dd5feea1 | 4035 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
4036 | |
4037 | if (sds->this_nr_running) { | |
4038 | sds->this_load_per_task /= sds->this_nr_running; | |
4039 | if (sds->busiest_load_per_task > | |
4040 | sds->this_load_per_task) | |
4041 | imbn = 1; | |
bd939f45 | 4042 | } else { |
1e3c88bd | 4043 | sds->this_load_per_task = |
bd939f45 PZ |
4044 | cpu_avg_load_per_task(env->dst_cpu); |
4045 | } | |
1e3c88bd | 4046 | |
dd5feea1 | 4047 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 4048 | * SCHED_POWER_SCALE; |
9c3f75cb | 4049 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
4050 | |
4051 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
4052 | (scaled_busy_load_per_task * imbn)) { | |
bd939f45 | 4053 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4054 | return; |
4055 | } | |
4056 | ||
4057 | /* | |
4058 | * OK, we don't have enough imbalance to justify moving tasks, | |
4059 | * however we may be able to increase total CPU power used by | |
4060 | * moving them. | |
4061 | */ | |
4062 | ||
9c3f75cb | 4063 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 4064 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 4065 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 4066 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 4067 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4068 | |
4069 | /* Amount of load we'd subtract */ | |
1399fa78 | 4070 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 4071 | sds->busiest->sgp->power; |
1e3c88bd | 4072 | if (sds->max_load > tmp) |
9c3f75cb | 4073 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
4074 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
4075 | ||
4076 | /* Amount of load we'd add */ | |
9c3f75cb | 4077 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 4078 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
4079 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
4080 | sds->this->sgp->power; | |
1e3c88bd | 4081 | else |
1399fa78 | 4082 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
4083 | sds->this->sgp->power; |
4084 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 4085 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 4086 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4087 | |
4088 | /* Move if we gain throughput */ | |
4089 | if (pwr_move > pwr_now) | |
bd939f45 | 4090 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4091 | } |
4092 | ||
4093 | /** | |
4094 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
4095 | * groups of a given sched_domain during load balance. | |
bd939f45 | 4096 | * @env: load balance environment |
1e3c88bd | 4097 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4098 | */ |
bd939f45 | 4099 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 4100 | { |
dd5feea1 SS |
4101 | unsigned long max_pull, load_above_capacity = ~0UL; |
4102 | ||
4103 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
4104 | if (sds->group_imb) { | |
4105 | sds->busiest_load_per_task = | |
4106 | min(sds->busiest_load_per_task, sds->avg_load); | |
4107 | } | |
4108 | ||
1e3c88bd PZ |
4109 | /* |
4110 | * In the presence of smp nice balancing, certain scenarios can have | |
4111 | * max load less than avg load(as we skip the groups at or below | |
4112 | * its cpu_power, while calculating max_load..) | |
4113 | */ | |
4114 | if (sds->max_load < sds->avg_load) { | |
bd939f45 PZ |
4115 | env->imbalance = 0; |
4116 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4117 | } |
4118 | ||
dd5feea1 SS |
4119 | if (!sds->group_imb) { |
4120 | /* | |
4121 | * Don't want to pull so many tasks that a group would go idle. | |
4122 | */ | |
4123 | load_above_capacity = (sds->busiest_nr_running - | |
4124 | sds->busiest_group_capacity); | |
4125 | ||
1399fa78 | 4126 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 4127 | |
9c3f75cb | 4128 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
4129 | } |
4130 | ||
4131 | /* | |
4132 | * We're trying to get all the cpus to the average_load, so we don't | |
4133 | * want to push ourselves above the average load, nor do we wish to | |
4134 | * reduce the max loaded cpu below the average load. At the same time, | |
4135 | * we also don't want to reduce the group load below the group capacity | |
4136 | * (so that we can implement power-savings policies etc). Thus we look | |
4137 | * for the minimum possible imbalance. | |
4138 | * Be careful of negative numbers as they'll appear as very large values | |
4139 | * with unsigned longs. | |
4140 | */ | |
4141 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
4142 | |
4143 | /* How much load to actually move to equalise the imbalance */ | |
bd939f45 | 4144 | env->imbalance = min(max_pull * sds->busiest->sgp->power, |
9c3f75cb | 4145 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) |
1399fa78 | 4146 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
4147 | |
4148 | /* | |
4149 | * if *imbalance is less than the average load per runnable task | |
25985edc | 4150 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
4151 | * a think about bumping its value to force at least one task to be |
4152 | * moved | |
4153 | */ | |
bd939f45 PZ |
4154 | if (env->imbalance < sds->busiest_load_per_task) |
4155 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4156 | |
4157 | } | |
fab47622 | 4158 | |
1e3c88bd PZ |
4159 | /******* find_busiest_group() helpers end here *********************/ |
4160 | ||
4161 | /** | |
4162 | * find_busiest_group - Returns the busiest group within the sched_domain | |
4163 | * if there is an imbalance. If there isn't an imbalance, and | |
4164 | * the user has opted for power-savings, it returns a group whose | |
4165 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
4166 | * such a group exists. | |
4167 | * | |
4168 | * Also calculates the amount of weighted load which should be moved | |
4169 | * to restore balance. | |
4170 | * | |
cd96891d | 4171 | * @env: The load balancing environment. |
1e3c88bd PZ |
4172 | * @balance: Pointer to a variable indicating if this_cpu |
4173 | * is the appropriate cpu to perform load balancing at this_level. | |
4174 | * | |
4175 | * Returns: - the busiest group if imbalance exists. | |
4176 | * - If no imbalance and user has opted for power-savings balance, | |
4177 | * return the least loaded group whose CPUs can be | |
4178 | * put to idle by rebalancing its tasks onto our group. | |
4179 | */ | |
4180 | static struct sched_group * | |
b9403130 | 4181 | find_busiest_group(struct lb_env *env, int *balance) |
1e3c88bd PZ |
4182 | { |
4183 | struct sd_lb_stats sds; | |
4184 | ||
4185 | memset(&sds, 0, sizeof(sds)); | |
4186 | ||
4187 | /* | |
4188 | * Compute the various statistics relavent for load balancing at | |
4189 | * this level. | |
4190 | */ | |
b9403130 | 4191 | update_sd_lb_stats(env, balance, &sds); |
1e3c88bd | 4192 | |
cc57aa8f PZ |
4193 | /* |
4194 | * this_cpu is not the appropriate cpu to perform load balancing at | |
4195 | * this level. | |
1e3c88bd | 4196 | */ |
8f190fb3 | 4197 | if (!(*balance)) |
1e3c88bd PZ |
4198 | goto ret; |
4199 | ||
bd939f45 PZ |
4200 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
4201 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
4202 | return sds.busiest; |
4203 | ||
cc57aa8f | 4204 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
4205 | if (!sds.busiest || sds.busiest_nr_running == 0) |
4206 | goto out_balanced; | |
4207 | ||
1399fa78 | 4208 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 4209 | |
866ab43e PZ |
4210 | /* |
4211 | * If the busiest group is imbalanced the below checks don't | |
4212 | * work because they assumes all things are equal, which typically | |
4213 | * isn't true due to cpus_allowed constraints and the like. | |
4214 | */ | |
4215 | if (sds.group_imb) | |
4216 | goto force_balance; | |
4217 | ||
cc57aa8f | 4218 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
bd939f45 | 4219 | if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
fab47622 NR |
4220 | !sds.busiest_has_capacity) |
4221 | goto force_balance; | |
4222 | ||
cc57aa8f PZ |
4223 | /* |
4224 | * If the local group is more busy than the selected busiest group | |
4225 | * don't try and pull any tasks. | |
4226 | */ | |
1e3c88bd PZ |
4227 | if (sds.this_load >= sds.max_load) |
4228 | goto out_balanced; | |
4229 | ||
cc57aa8f PZ |
4230 | /* |
4231 | * Don't pull any tasks if this group is already above the domain | |
4232 | * average load. | |
4233 | */ | |
1e3c88bd PZ |
4234 | if (sds.this_load >= sds.avg_load) |
4235 | goto out_balanced; | |
4236 | ||
bd939f45 | 4237 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
4238 | /* |
4239 | * This cpu is idle. If the busiest group load doesn't | |
4240 | * have more tasks than the number of available cpu's and | |
4241 | * there is no imbalance between this and busiest group | |
4242 | * wrt to idle cpu's, it is balanced. | |
4243 | */ | |
c186fafe | 4244 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
4245 | sds.busiest_nr_running <= sds.busiest_group_weight) |
4246 | goto out_balanced; | |
c186fafe PZ |
4247 | } else { |
4248 | /* | |
4249 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
4250 | * imbalance_pct to be conservative. | |
4251 | */ | |
bd939f45 | 4252 | if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load) |
c186fafe | 4253 | goto out_balanced; |
aae6d3dd | 4254 | } |
1e3c88bd | 4255 | |
fab47622 | 4256 | force_balance: |
1e3c88bd | 4257 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 4258 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
4259 | return sds.busiest; |
4260 | ||
4261 | out_balanced: | |
1e3c88bd | 4262 | ret: |
bd939f45 | 4263 | env->imbalance = 0; |
1e3c88bd PZ |
4264 | return NULL; |
4265 | } | |
4266 | ||
4267 | /* | |
4268 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
4269 | */ | |
bd939f45 | 4270 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 4271 | struct sched_group *group) |
1e3c88bd PZ |
4272 | { |
4273 | struct rq *busiest = NULL, *rq; | |
4274 | unsigned long max_load = 0; | |
4275 | int i; | |
4276 | ||
4277 | for_each_cpu(i, sched_group_cpus(group)) { | |
4278 | unsigned long power = power_of(i); | |
1399fa78 NR |
4279 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
4280 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
4281 | unsigned long wl; |
4282 | ||
9d5efe05 | 4283 | if (!capacity) |
bd939f45 | 4284 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 4285 | |
b9403130 | 4286 | if (!cpumask_test_cpu(i, env->cpus)) |
1e3c88bd PZ |
4287 | continue; |
4288 | ||
4289 | rq = cpu_rq(i); | |
6e40f5bb | 4290 | wl = weighted_cpuload(i); |
1e3c88bd | 4291 | |
6e40f5bb TG |
4292 | /* |
4293 | * When comparing with imbalance, use weighted_cpuload() | |
4294 | * which is not scaled with the cpu power. | |
4295 | */ | |
bd939f45 | 4296 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
4297 | continue; |
4298 | ||
6e40f5bb TG |
4299 | /* |
4300 | * For the load comparisons with the other cpu's, consider | |
4301 | * the weighted_cpuload() scaled with the cpu power, so that | |
4302 | * the load can be moved away from the cpu that is potentially | |
4303 | * running at a lower capacity. | |
4304 | */ | |
1399fa78 | 4305 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 4306 | |
1e3c88bd PZ |
4307 | if (wl > max_load) { |
4308 | max_load = wl; | |
4309 | busiest = rq; | |
4310 | } | |
4311 | } | |
4312 | ||
4313 | return busiest; | |
4314 | } | |
4315 | ||
4316 | /* | |
4317 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
4318 | * so long as it is large enough. | |
4319 | */ | |
4320 | #define MAX_PINNED_INTERVAL 512 | |
4321 | ||
4322 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
029632fb | 4323 | DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); |
1e3c88bd | 4324 | |
bd939f45 | 4325 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 4326 | { |
bd939f45 PZ |
4327 | struct sched_domain *sd = env->sd; |
4328 | ||
4329 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
4330 | |
4331 | /* | |
4332 | * ASYM_PACKING needs to force migrate tasks from busy but | |
4333 | * higher numbered CPUs in order to pack all tasks in the | |
4334 | * lowest numbered CPUs. | |
4335 | */ | |
bd939f45 | 4336 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 4337 | return 1; |
1af3ed3d PZ |
4338 | } |
4339 | ||
4340 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
4341 | } | |
4342 | ||
969c7921 TH |
4343 | static int active_load_balance_cpu_stop(void *data); |
4344 | ||
1e3c88bd PZ |
4345 | /* |
4346 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
4347 | * tasks if there is an imbalance. | |
4348 | */ | |
4349 | static int load_balance(int this_cpu, struct rq *this_rq, | |
4350 | struct sched_domain *sd, enum cpu_idle_type idle, | |
4351 | int *balance) | |
4352 | { | |
88b8dac0 SV |
4353 | int ld_moved, cur_ld_moved, active_balance = 0; |
4354 | int lb_iterations, max_lb_iterations; | |
1e3c88bd | 4355 | struct sched_group *group; |
1e3c88bd PZ |
4356 | struct rq *busiest; |
4357 | unsigned long flags; | |
4358 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
4359 | ||
8e45cb54 PZ |
4360 | struct lb_env env = { |
4361 | .sd = sd, | |
ddcdf6e7 PZ |
4362 | .dst_cpu = this_cpu, |
4363 | .dst_rq = this_rq, | |
88b8dac0 | 4364 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 4365 | .idle = idle, |
eb95308e | 4366 | .loop_break = sched_nr_migrate_break, |
b9403130 | 4367 | .cpus = cpus, |
8e45cb54 PZ |
4368 | }; |
4369 | ||
1e3c88bd | 4370 | cpumask_copy(cpus, cpu_active_mask); |
88b8dac0 | 4371 | max_lb_iterations = cpumask_weight(env.dst_grpmask); |
1e3c88bd | 4372 | |
1e3c88bd PZ |
4373 | schedstat_inc(sd, lb_count[idle]); |
4374 | ||
4375 | redo: | |
b9403130 | 4376 | group = find_busiest_group(&env, balance); |
1e3c88bd PZ |
4377 | |
4378 | if (*balance == 0) | |
4379 | goto out_balanced; | |
4380 | ||
4381 | if (!group) { | |
4382 | schedstat_inc(sd, lb_nobusyg[idle]); | |
4383 | goto out_balanced; | |
4384 | } | |
4385 | ||
b9403130 | 4386 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
4387 | if (!busiest) { |
4388 | schedstat_inc(sd, lb_nobusyq[idle]); | |
4389 | goto out_balanced; | |
4390 | } | |
4391 | ||
78feefc5 | 4392 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 4393 | |
bd939f45 | 4394 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
4395 | |
4396 | ld_moved = 0; | |
88b8dac0 | 4397 | lb_iterations = 1; |
1e3c88bd PZ |
4398 | if (busiest->nr_running > 1) { |
4399 | /* | |
4400 | * Attempt to move tasks. If find_busiest_group has found | |
4401 | * an imbalance but busiest->nr_running <= 1, the group is | |
4402 | * still unbalanced. ld_moved simply stays zero, so it is | |
4403 | * correctly treated as an imbalance. | |
4404 | */ | |
8e45cb54 | 4405 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
4406 | env.src_cpu = busiest->cpu; |
4407 | env.src_rq = busiest; | |
4408 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 4409 | |
a35b6466 | 4410 | update_h_load(env.src_cpu); |
5d6523eb | 4411 | more_balance: |
1e3c88bd | 4412 | local_irq_save(flags); |
78feefc5 | 4413 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
4414 | |
4415 | /* | |
4416 | * cur_ld_moved - load moved in current iteration | |
4417 | * ld_moved - cumulative load moved across iterations | |
4418 | */ | |
4419 | cur_ld_moved = move_tasks(&env); | |
4420 | ld_moved += cur_ld_moved; | |
78feefc5 | 4421 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
4422 | local_irq_restore(flags); |
4423 | ||
5d6523eb PZ |
4424 | if (env.flags & LBF_NEED_BREAK) { |
4425 | env.flags &= ~LBF_NEED_BREAK; | |
4426 | goto more_balance; | |
4427 | } | |
4428 | ||
1e3c88bd PZ |
4429 | /* |
4430 | * some other cpu did the load balance for us. | |
4431 | */ | |
88b8dac0 SV |
4432 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
4433 | resched_cpu(env.dst_cpu); | |
4434 | ||
4435 | /* | |
4436 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
4437 | * us and move them to an alternate dst_cpu in our sched_group | |
4438 | * where they can run. The upper limit on how many times we | |
4439 | * iterate on same src_cpu is dependent on number of cpus in our | |
4440 | * sched_group. | |
4441 | * | |
4442 | * This changes load balance semantics a bit on who can move | |
4443 | * load to a given_cpu. In addition to the given_cpu itself | |
4444 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
4445 | * nohz-idle), we now have balance_cpu in a position to move | |
4446 | * load to given_cpu. In rare situations, this may cause | |
4447 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
4448 | * _independently_ and at _same_ time to move some load to | |
4449 | * given_cpu) causing exceess load to be moved to given_cpu. | |
4450 | * This however should not happen so much in practice and | |
4451 | * moreover subsequent load balance cycles should correct the | |
4452 | * excess load moved. | |
4453 | */ | |
4454 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 && | |
4455 | lb_iterations++ < max_lb_iterations) { | |
4456 | ||
78feefc5 | 4457 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 SV |
4458 | env.dst_cpu = env.new_dst_cpu; |
4459 | env.flags &= ~LBF_SOME_PINNED; | |
4460 | env.loop = 0; | |
4461 | env.loop_break = sched_nr_migrate_break; | |
4462 | /* | |
4463 | * Go back to "more_balance" rather than "redo" since we | |
4464 | * need to continue with same src_cpu. | |
4465 | */ | |
4466 | goto more_balance; | |
4467 | } | |
1e3c88bd PZ |
4468 | |
4469 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
8e45cb54 | 4470 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 4471 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
4472 | if (!cpumask_empty(cpus)) { |
4473 | env.loop = 0; | |
4474 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 4475 | goto redo; |
bbf18b19 | 4476 | } |
1e3c88bd PZ |
4477 | goto out_balanced; |
4478 | } | |
4479 | } | |
4480 | ||
4481 | if (!ld_moved) { | |
4482 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
4483 | /* |
4484 | * Increment the failure counter only on periodic balance. | |
4485 | * We do not want newidle balance, which can be very | |
4486 | * frequent, pollute the failure counter causing | |
4487 | * excessive cache_hot migrations and active balances. | |
4488 | */ | |
4489 | if (idle != CPU_NEWLY_IDLE) | |
4490 | sd->nr_balance_failed++; | |
1e3c88bd | 4491 | |
bd939f45 | 4492 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
4493 | raw_spin_lock_irqsave(&busiest->lock, flags); |
4494 | ||
969c7921 TH |
4495 | /* don't kick the active_load_balance_cpu_stop, |
4496 | * if the curr task on busiest cpu can't be | |
4497 | * moved to this_cpu | |
1e3c88bd PZ |
4498 | */ |
4499 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 4500 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
4501 | raw_spin_unlock_irqrestore(&busiest->lock, |
4502 | flags); | |
8e45cb54 | 4503 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
4504 | goto out_one_pinned; |
4505 | } | |
4506 | ||
969c7921 TH |
4507 | /* |
4508 | * ->active_balance synchronizes accesses to | |
4509 | * ->active_balance_work. Once set, it's cleared | |
4510 | * only after active load balance is finished. | |
4511 | */ | |
1e3c88bd PZ |
4512 | if (!busiest->active_balance) { |
4513 | busiest->active_balance = 1; | |
4514 | busiest->push_cpu = this_cpu; | |
4515 | active_balance = 1; | |
4516 | } | |
4517 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 4518 | |
bd939f45 | 4519 | if (active_balance) { |
969c7921 TH |
4520 | stop_one_cpu_nowait(cpu_of(busiest), |
4521 | active_load_balance_cpu_stop, busiest, | |
4522 | &busiest->active_balance_work); | |
bd939f45 | 4523 | } |
1e3c88bd PZ |
4524 | |
4525 | /* | |
4526 | * We've kicked active balancing, reset the failure | |
4527 | * counter. | |
4528 | */ | |
4529 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
4530 | } | |
4531 | } else | |
4532 | sd->nr_balance_failed = 0; | |
4533 | ||
4534 | if (likely(!active_balance)) { | |
4535 | /* We were unbalanced, so reset the balancing interval */ | |
4536 | sd->balance_interval = sd->min_interval; | |
4537 | } else { | |
4538 | /* | |
4539 | * If we've begun active balancing, start to back off. This | |
4540 | * case may not be covered by the all_pinned logic if there | |
4541 | * is only 1 task on the busy runqueue (because we don't call | |
4542 | * move_tasks). | |
4543 | */ | |
4544 | if (sd->balance_interval < sd->max_interval) | |
4545 | sd->balance_interval *= 2; | |
4546 | } | |
4547 | ||
1e3c88bd PZ |
4548 | goto out; |
4549 | ||
4550 | out_balanced: | |
4551 | schedstat_inc(sd, lb_balanced[idle]); | |
4552 | ||
4553 | sd->nr_balance_failed = 0; | |
4554 | ||
4555 | out_one_pinned: | |
4556 | /* tune up the balancing interval */ | |
8e45cb54 | 4557 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 4558 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
4559 | (sd->balance_interval < sd->max_interval)) |
4560 | sd->balance_interval *= 2; | |
4561 | ||
46e49b38 | 4562 | ld_moved = 0; |
1e3c88bd | 4563 | out: |
1e3c88bd PZ |
4564 | return ld_moved; |
4565 | } | |
4566 | ||
1e3c88bd PZ |
4567 | /* |
4568 | * idle_balance is called by schedule() if this_cpu is about to become | |
4569 | * idle. Attempts to pull tasks from other CPUs. | |
4570 | */ | |
029632fb | 4571 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
4572 | { |
4573 | struct sched_domain *sd; | |
4574 | int pulled_task = 0; | |
4575 | unsigned long next_balance = jiffies + HZ; | |
4576 | ||
4577 | this_rq->idle_stamp = this_rq->clock; | |
4578 | ||
4579 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
4580 | return; | |
4581 | ||
f492e12e PZ |
4582 | /* |
4583 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
4584 | */ | |
4585 | raw_spin_unlock(&this_rq->lock); | |
4586 | ||
c66eaf61 | 4587 | update_shares(this_cpu); |
dce840a0 | 4588 | rcu_read_lock(); |
1e3c88bd PZ |
4589 | for_each_domain(this_cpu, sd) { |
4590 | unsigned long interval; | |
f492e12e | 4591 | int balance = 1; |
1e3c88bd PZ |
4592 | |
4593 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4594 | continue; | |
4595 | ||
f492e12e | 4596 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 4597 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
4598 | pulled_task = load_balance(this_cpu, this_rq, |
4599 | sd, CPU_NEWLY_IDLE, &balance); | |
4600 | } | |
1e3c88bd PZ |
4601 | |
4602 | interval = msecs_to_jiffies(sd->balance_interval); | |
4603 | if (time_after(next_balance, sd->last_balance + interval)) | |
4604 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
4605 | if (pulled_task) { |
4606 | this_rq->idle_stamp = 0; | |
1e3c88bd | 4607 | break; |
d5ad140b | 4608 | } |
1e3c88bd | 4609 | } |
dce840a0 | 4610 | rcu_read_unlock(); |
f492e12e PZ |
4611 | |
4612 | raw_spin_lock(&this_rq->lock); | |
4613 | ||
1e3c88bd PZ |
4614 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
4615 | /* | |
4616 | * We are going idle. next_balance may be set based on | |
4617 | * a busy processor. So reset next_balance. | |
4618 | */ | |
4619 | this_rq->next_balance = next_balance; | |
4620 | } | |
4621 | } | |
4622 | ||
4623 | /* | |
969c7921 TH |
4624 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
4625 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
4626 | * least 1 task to be running on each physical CPU where possible, and | |
4627 | * avoids physical / logical imbalances. | |
1e3c88bd | 4628 | */ |
969c7921 | 4629 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 4630 | { |
969c7921 TH |
4631 | struct rq *busiest_rq = data; |
4632 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 4633 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 4634 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 4635 | struct sched_domain *sd; |
969c7921 TH |
4636 | |
4637 | raw_spin_lock_irq(&busiest_rq->lock); | |
4638 | ||
4639 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
4640 | if (unlikely(busiest_cpu != smp_processor_id() || | |
4641 | !busiest_rq->active_balance)) | |
4642 | goto out_unlock; | |
1e3c88bd PZ |
4643 | |
4644 | /* Is there any task to move? */ | |
4645 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 4646 | goto out_unlock; |
1e3c88bd PZ |
4647 | |
4648 | /* | |
4649 | * This condition is "impossible", if it occurs | |
4650 | * we need to fix it. Originally reported by | |
4651 | * Bjorn Helgaas on a 128-cpu setup. | |
4652 | */ | |
4653 | BUG_ON(busiest_rq == target_rq); | |
4654 | ||
4655 | /* move a task from busiest_rq to target_rq */ | |
4656 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
4657 | |
4658 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 4659 | rcu_read_lock(); |
1e3c88bd PZ |
4660 | for_each_domain(target_cpu, sd) { |
4661 | if ((sd->flags & SD_LOAD_BALANCE) && | |
4662 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
4663 | break; | |
4664 | } | |
4665 | ||
4666 | if (likely(sd)) { | |
8e45cb54 PZ |
4667 | struct lb_env env = { |
4668 | .sd = sd, | |
ddcdf6e7 PZ |
4669 | .dst_cpu = target_cpu, |
4670 | .dst_rq = target_rq, | |
4671 | .src_cpu = busiest_rq->cpu, | |
4672 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
4673 | .idle = CPU_IDLE, |
4674 | }; | |
4675 | ||
1e3c88bd PZ |
4676 | schedstat_inc(sd, alb_count); |
4677 | ||
8e45cb54 | 4678 | if (move_one_task(&env)) |
1e3c88bd PZ |
4679 | schedstat_inc(sd, alb_pushed); |
4680 | else | |
4681 | schedstat_inc(sd, alb_failed); | |
4682 | } | |
dce840a0 | 4683 | rcu_read_unlock(); |
1e3c88bd | 4684 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
4685 | out_unlock: |
4686 | busiest_rq->active_balance = 0; | |
4687 | raw_spin_unlock_irq(&busiest_rq->lock); | |
4688 | return 0; | |
1e3c88bd PZ |
4689 | } |
4690 | ||
4691 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
4692 | /* |
4693 | * idle load balancing details | |
83cd4fe2 VP |
4694 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
4695 | * needed, they will kick the idle load balancer, which then does idle | |
4696 | * load balancing for all the idle CPUs. | |
4697 | */ | |
1e3c88bd | 4698 | static struct { |
83cd4fe2 | 4699 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 4700 | atomic_t nr_cpus; |
83cd4fe2 VP |
4701 | unsigned long next_balance; /* in jiffy units */ |
4702 | } nohz ____cacheline_aligned; | |
1e3c88bd | 4703 | |
8e7fbcbc | 4704 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 4705 | { |
0b005cf5 | 4706 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 4707 | |
786d6dc7 SS |
4708 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
4709 | return ilb; | |
4710 | ||
4711 | return nr_cpu_ids; | |
1e3c88bd | 4712 | } |
1e3c88bd | 4713 | |
83cd4fe2 VP |
4714 | /* |
4715 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
4716 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
4717 | * CPU (if there is one). | |
4718 | */ | |
4719 | static void nohz_balancer_kick(int cpu) | |
4720 | { | |
4721 | int ilb_cpu; | |
4722 | ||
4723 | nohz.next_balance++; | |
4724 | ||
0b005cf5 | 4725 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 4726 | |
0b005cf5 SS |
4727 | if (ilb_cpu >= nr_cpu_ids) |
4728 | return; | |
83cd4fe2 | 4729 | |
cd490c5b | 4730 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
4731 | return; |
4732 | /* | |
4733 | * Use smp_send_reschedule() instead of resched_cpu(). | |
4734 | * This way we generate a sched IPI on the target cpu which | |
4735 | * is idle. And the softirq performing nohz idle load balance | |
4736 | * will be run before returning from the IPI. | |
4737 | */ | |
4738 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
4739 | return; |
4740 | } | |
4741 | ||
c1cc017c | 4742 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
4743 | { |
4744 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
4745 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
4746 | atomic_dec(&nohz.nr_cpus); | |
4747 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
4748 | } | |
4749 | } | |
4750 | ||
69e1e811 SS |
4751 | static inline void set_cpu_sd_state_busy(void) |
4752 | { | |
4753 | struct sched_domain *sd; | |
4754 | int cpu = smp_processor_id(); | |
4755 | ||
4756 | if (!test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
4757 | return; | |
4758 | clear_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
4759 | ||
4760 | rcu_read_lock(); | |
4761 | for_each_domain(cpu, sd) | |
4762 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); | |
4763 | rcu_read_unlock(); | |
4764 | } | |
4765 | ||
4766 | void set_cpu_sd_state_idle(void) | |
4767 | { | |
4768 | struct sched_domain *sd; | |
4769 | int cpu = smp_processor_id(); | |
4770 | ||
4771 | if (test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
4772 | return; | |
4773 | set_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
4774 | ||
4775 | rcu_read_lock(); | |
4776 | for_each_domain(cpu, sd) | |
4777 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); | |
4778 | rcu_read_unlock(); | |
4779 | } | |
4780 | ||
1e3c88bd | 4781 | /* |
c1cc017c | 4782 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 4783 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 4784 | */ |
c1cc017c | 4785 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 4786 | { |
71325960 SS |
4787 | /* |
4788 | * If this cpu is going down, then nothing needs to be done. | |
4789 | */ | |
4790 | if (!cpu_active(cpu)) | |
4791 | return; | |
4792 | ||
c1cc017c AS |
4793 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
4794 | return; | |
1e3c88bd | 4795 | |
c1cc017c AS |
4796 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
4797 | atomic_inc(&nohz.nr_cpus); | |
4798 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 4799 | } |
71325960 SS |
4800 | |
4801 | static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb, | |
4802 | unsigned long action, void *hcpu) | |
4803 | { | |
4804 | switch (action & ~CPU_TASKS_FROZEN) { | |
4805 | case CPU_DYING: | |
c1cc017c | 4806 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
4807 | return NOTIFY_OK; |
4808 | default: | |
4809 | return NOTIFY_DONE; | |
4810 | } | |
4811 | } | |
1e3c88bd PZ |
4812 | #endif |
4813 | ||
4814 | static DEFINE_SPINLOCK(balancing); | |
4815 | ||
49c022e6 PZ |
4816 | /* |
4817 | * Scale the max load_balance interval with the number of CPUs in the system. | |
4818 | * This trades load-balance latency on larger machines for less cross talk. | |
4819 | */ | |
029632fb | 4820 | void update_max_interval(void) |
49c022e6 PZ |
4821 | { |
4822 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
4823 | } | |
4824 | ||
1e3c88bd PZ |
4825 | /* |
4826 | * It checks each scheduling domain to see if it is due to be balanced, | |
4827 | * and initiates a balancing operation if so. | |
4828 | * | |
4829 | * Balancing parameters are set up in arch_init_sched_domains. | |
4830 | */ | |
4831 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4832 | { | |
4833 | int balance = 1; | |
4834 | struct rq *rq = cpu_rq(cpu); | |
4835 | unsigned long interval; | |
04f733b4 | 4836 | struct sched_domain *sd; |
1e3c88bd PZ |
4837 | /* Earliest time when we have to do rebalance again */ |
4838 | unsigned long next_balance = jiffies + 60*HZ; | |
4839 | int update_next_balance = 0; | |
4840 | int need_serialize; | |
4841 | ||
2069dd75 PZ |
4842 | update_shares(cpu); |
4843 | ||
dce840a0 | 4844 | rcu_read_lock(); |
1e3c88bd PZ |
4845 | for_each_domain(cpu, sd) { |
4846 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4847 | continue; | |
4848 | ||
4849 | interval = sd->balance_interval; | |
4850 | if (idle != CPU_IDLE) | |
4851 | interval *= sd->busy_factor; | |
4852 | ||
4853 | /* scale ms to jiffies */ | |
4854 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 4855 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
4856 | |
4857 | need_serialize = sd->flags & SD_SERIALIZE; | |
4858 | ||
4859 | if (need_serialize) { | |
4860 | if (!spin_trylock(&balancing)) | |
4861 | goto out; | |
4862 | } | |
4863 | ||
4864 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4865 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
4866 | /* | |
4867 | * We've pulled tasks over so either we're no | |
c186fafe | 4868 | * longer idle. |
1e3c88bd PZ |
4869 | */ |
4870 | idle = CPU_NOT_IDLE; | |
4871 | } | |
4872 | sd->last_balance = jiffies; | |
4873 | } | |
4874 | if (need_serialize) | |
4875 | spin_unlock(&balancing); | |
4876 | out: | |
4877 | if (time_after(next_balance, sd->last_balance + interval)) { | |
4878 | next_balance = sd->last_balance + interval; | |
4879 | update_next_balance = 1; | |
4880 | } | |
4881 | ||
4882 | /* | |
4883 | * Stop the load balance at this level. There is another | |
4884 | * CPU in our sched group which is doing load balancing more | |
4885 | * actively. | |
4886 | */ | |
4887 | if (!balance) | |
4888 | break; | |
4889 | } | |
dce840a0 | 4890 | rcu_read_unlock(); |
1e3c88bd PZ |
4891 | |
4892 | /* | |
4893 | * next_balance will be updated only when there is a need. | |
4894 | * When the cpu is attached to null domain for ex, it will not be | |
4895 | * updated. | |
4896 | */ | |
4897 | if (likely(update_next_balance)) | |
4898 | rq->next_balance = next_balance; | |
4899 | } | |
4900 | ||
83cd4fe2 | 4901 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 4902 | /* |
83cd4fe2 | 4903 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
4904 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
4905 | */ | |
83cd4fe2 VP |
4906 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
4907 | { | |
4908 | struct rq *this_rq = cpu_rq(this_cpu); | |
4909 | struct rq *rq; | |
4910 | int balance_cpu; | |
4911 | ||
1c792db7 SS |
4912 | if (idle != CPU_IDLE || |
4913 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
4914 | goto end; | |
83cd4fe2 VP |
4915 | |
4916 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 4917 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
4918 | continue; |
4919 | ||
4920 | /* | |
4921 | * If this cpu gets work to do, stop the load balancing | |
4922 | * work being done for other cpus. Next load | |
4923 | * balancing owner will pick it up. | |
4924 | */ | |
1c792db7 | 4925 | if (need_resched()) |
83cd4fe2 | 4926 | break; |
83cd4fe2 | 4927 | |
5ed4f1d9 VG |
4928 | rq = cpu_rq(balance_cpu); |
4929 | ||
4930 | raw_spin_lock_irq(&rq->lock); | |
4931 | update_rq_clock(rq); | |
4932 | update_idle_cpu_load(rq); | |
4933 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
4934 | |
4935 | rebalance_domains(balance_cpu, CPU_IDLE); | |
4936 | ||
83cd4fe2 VP |
4937 | if (time_after(this_rq->next_balance, rq->next_balance)) |
4938 | this_rq->next_balance = rq->next_balance; | |
4939 | } | |
4940 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
4941 | end: |
4942 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
4943 | } |
4944 | ||
4945 | /* | |
0b005cf5 SS |
4946 | * Current heuristic for kicking the idle load balancer in the presence |
4947 | * of an idle cpu is the system. | |
4948 | * - This rq has more than one task. | |
4949 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
4950 | * busy cpu's exceeding the group's power. | |
4951 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
4952 | * domain span are idle. | |
83cd4fe2 VP |
4953 | */ |
4954 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
4955 | { | |
4956 | unsigned long now = jiffies; | |
0b005cf5 | 4957 | struct sched_domain *sd; |
83cd4fe2 | 4958 | |
1c792db7 | 4959 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
4960 | return 0; |
4961 | ||
1c792db7 SS |
4962 | /* |
4963 | * We may be recently in ticked or tickless idle mode. At the first | |
4964 | * busy tick after returning from idle, we will update the busy stats. | |
4965 | */ | |
69e1e811 | 4966 | set_cpu_sd_state_busy(); |
c1cc017c | 4967 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
4968 | |
4969 | /* | |
4970 | * None are in tickless mode and hence no need for NOHZ idle load | |
4971 | * balancing. | |
4972 | */ | |
4973 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
4974 | return 0; | |
1c792db7 SS |
4975 | |
4976 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
4977 | return 0; |
4978 | ||
0b005cf5 SS |
4979 | if (rq->nr_running >= 2) |
4980 | goto need_kick; | |
83cd4fe2 | 4981 | |
067491b7 | 4982 | rcu_read_lock(); |
0b005cf5 SS |
4983 | for_each_domain(cpu, sd) { |
4984 | struct sched_group *sg = sd->groups; | |
4985 | struct sched_group_power *sgp = sg->sgp; | |
4986 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 4987 | |
0b005cf5 | 4988 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 4989 | goto need_kick_unlock; |
0b005cf5 SS |
4990 | |
4991 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
4992 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
4993 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 4994 | goto need_kick_unlock; |
0b005cf5 SS |
4995 | |
4996 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
4997 | break; | |
83cd4fe2 | 4998 | } |
067491b7 | 4999 | rcu_read_unlock(); |
83cd4fe2 | 5000 | return 0; |
067491b7 PZ |
5001 | |
5002 | need_kick_unlock: | |
5003 | rcu_read_unlock(); | |
0b005cf5 SS |
5004 | need_kick: |
5005 | return 1; | |
83cd4fe2 VP |
5006 | } |
5007 | #else | |
5008 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
5009 | #endif | |
5010 | ||
5011 | /* | |
5012 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
5013 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
5014 | */ | |
1e3c88bd PZ |
5015 | static void run_rebalance_domains(struct softirq_action *h) |
5016 | { | |
5017 | int this_cpu = smp_processor_id(); | |
5018 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 5019 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
5020 | CPU_IDLE : CPU_NOT_IDLE; |
5021 | ||
5022 | rebalance_domains(this_cpu, idle); | |
5023 | ||
1e3c88bd | 5024 | /* |
83cd4fe2 | 5025 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
5026 | * balancing on behalf of the other idle cpus whose ticks are |
5027 | * stopped. | |
5028 | */ | |
83cd4fe2 | 5029 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
5030 | } |
5031 | ||
5032 | static inline int on_null_domain(int cpu) | |
5033 | { | |
90a6501f | 5034 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
5035 | } |
5036 | ||
5037 | /* | |
5038 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 5039 | */ |
029632fb | 5040 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 5041 | { |
1e3c88bd PZ |
5042 | /* Don't need to rebalance while attached to NULL domain */ |
5043 | if (time_after_eq(jiffies, rq->next_balance) && | |
5044 | likely(!on_null_domain(cpu))) | |
5045 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 | 5046 | #ifdef CONFIG_NO_HZ |
1c792db7 | 5047 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
5048 | nohz_balancer_kick(cpu); |
5049 | #endif | |
1e3c88bd PZ |
5050 | } |
5051 | ||
0bcdcf28 CE |
5052 | static void rq_online_fair(struct rq *rq) |
5053 | { | |
5054 | update_sysctl(); | |
5055 | } | |
5056 | ||
5057 | static void rq_offline_fair(struct rq *rq) | |
5058 | { | |
5059 | update_sysctl(); | |
a4c96ae3 PB |
5060 | |
5061 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
5062 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
5063 | } |
5064 | ||
55e12e5e | 5065 | #endif /* CONFIG_SMP */ |
e1d1484f | 5066 | |
bf0f6f24 IM |
5067 | /* |
5068 | * scheduler tick hitting a task of our scheduling class: | |
5069 | */ | |
8f4d37ec | 5070 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
5071 | { |
5072 | struct cfs_rq *cfs_rq; | |
5073 | struct sched_entity *se = &curr->se; | |
5074 | ||
5075 | for_each_sched_entity(se) { | |
5076 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 5077 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 5078 | } |
cbee9f88 PZ |
5079 | |
5080 | if (sched_feat_numa(NUMA)) | |
5081 | task_tick_numa(rq, curr); | |
bf0f6f24 IM |
5082 | } |
5083 | ||
5084 | /* | |
cd29fe6f PZ |
5085 | * called on fork with the child task as argument from the parent's context |
5086 | * - child not yet on the tasklist | |
5087 | * - preemption disabled | |
bf0f6f24 | 5088 | */ |
cd29fe6f | 5089 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 5090 | { |
4fc420c9 DN |
5091 | struct cfs_rq *cfs_rq; |
5092 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 5093 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
5094 | struct rq *rq = this_rq(); |
5095 | unsigned long flags; | |
5096 | ||
05fa785c | 5097 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 5098 | |
861d034e PZ |
5099 | update_rq_clock(rq); |
5100 | ||
4fc420c9 DN |
5101 | cfs_rq = task_cfs_rq(current); |
5102 | curr = cfs_rq->curr; | |
5103 | ||
b0a0f667 PM |
5104 | if (unlikely(task_cpu(p) != this_cpu)) { |
5105 | rcu_read_lock(); | |
cd29fe6f | 5106 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
5107 | rcu_read_unlock(); |
5108 | } | |
bf0f6f24 | 5109 | |
7109c442 | 5110 | update_curr(cfs_rq); |
cd29fe6f | 5111 | |
b5d9d734 MG |
5112 | if (curr) |
5113 | se->vruntime = curr->vruntime; | |
aeb73b04 | 5114 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 5115 | |
cd29fe6f | 5116 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 5117 | /* |
edcb60a3 IM |
5118 | * Upon rescheduling, sched_class::put_prev_task() will place |
5119 | * 'current' within the tree based on its new key value. | |
5120 | */ | |
4d78e7b6 | 5121 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 5122 | resched_task(rq->curr); |
4d78e7b6 | 5123 | } |
bf0f6f24 | 5124 | |
88ec22d3 PZ |
5125 | se->vruntime -= cfs_rq->min_vruntime; |
5126 | ||
05fa785c | 5127 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
5128 | } |
5129 | ||
cb469845 SR |
5130 | /* |
5131 | * Priority of the task has changed. Check to see if we preempt | |
5132 | * the current task. | |
5133 | */ | |
da7a735e PZ |
5134 | static void |
5135 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 5136 | { |
da7a735e PZ |
5137 | if (!p->se.on_rq) |
5138 | return; | |
5139 | ||
cb469845 SR |
5140 | /* |
5141 | * Reschedule if we are currently running on this runqueue and | |
5142 | * our priority decreased, or if we are not currently running on | |
5143 | * this runqueue and our priority is higher than the current's | |
5144 | */ | |
da7a735e | 5145 | if (rq->curr == p) { |
cb469845 SR |
5146 | if (p->prio > oldprio) |
5147 | resched_task(rq->curr); | |
5148 | } else | |
15afe09b | 5149 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5150 | } |
5151 | ||
da7a735e PZ |
5152 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
5153 | { | |
5154 | struct sched_entity *se = &p->se; | |
5155 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5156 | ||
5157 | /* | |
5158 | * Ensure the task's vruntime is normalized, so that when its | |
5159 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
5160 | * do the right thing. | |
5161 | * | |
5162 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
5163 | * have normalized the vruntime, if it was !on_rq, then only when | |
5164 | * the task is sleeping will it still have non-normalized vruntime. | |
5165 | */ | |
5166 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
5167 | /* | |
5168 | * Fix up our vruntime so that the current sleep doesn't | |
5169 | * cause 'unlimited' sleep bonus. | |
5170 | */ | |
5171 | place_entity(cfs_rq, se, 0); | |
5172 | se->vruntime -= cfs_rq->min_vruntime; | |
5173 | } | |
5174 | } | |
5175 | ||
cb469845 SR |
5176 | /* |
5177 | * We switched to the sched_fair class. | |
5178 | */ | |
da7a735e | 5179 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 5180 | { |
da7a735e PZ |
5181 | if (!p->se.on_rq) |
5182 | return; | |
5183 | ||
cb469845 SR |
5184 | /* |
5185 | * We were most likely switched from sched_rt, so | |
5186 | * kick off the schedule if running, otherwise just see | |
5187 | * if we can still preempt the current task. | |
5188 | */ | |
da7a735e | 5189 | if (rq->curr == p) |
cb469845 SR |
5190 | resched_task(rq->curr); |
5191 | else | |
15afe09b | 5192 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5193 | } |
5194 | ||
83b699ed SV |
5195 | /* Account for a task changing its policy or group. |
5196 | * | |
5197 | * This routine is mostly called to set cfs_rq->curr field when a task | |
5198 | * migrates between groups/classes. | |
5199 | */ | |
5200 | static void set_curr_task_fair(struct rq *rq) | |
5201 | { | |
5202 | struct sched_entity *se = &rq->curr->se; | |
5203 | ||
ec12cb7f PT |
5204 | for_each_sched_entity(se) { |
5205 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5206 | ||
5207 | set_next_entity(cfs_rq, se); | |
5208 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
5209 | account_cfs_rq_runtime(cfs_rq, 0); | |
5210 | } | |
83b699ed SV |
5211 | } |
5212 | ||
029632fb PZ |
5213 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
5214 | { | |
5215 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
5216 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
5217 | #ifndef CONFIG_64BIT | |
5218 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
5219 | #endif | |
5220 | } | |
5221 | ||
810b3817 | 5222 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5223 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 5224 | { |
b2b5ce02 PZ |
5225 | /* |
5226 | * If the task was not on the rq at the time of this cgroup movement | |
5227 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
5228 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
5229 | * bonus in place_entity()). | |
5230 | * | |
5231 | * If it was on the rq, we've just 'preempted' it, which does convert | |
5232 | * ->vruntime to a relative base. | |
5233 | * | |
5234 | * Make sure both cases convert their relative position when migrating | |
5235 | * to another cgroup's rq. This does somewhat interfere with the | |
5236 | * fair sleeper stuff for the first placement, but who cares. | |
5237 | */ | |
7ceff013 DN |
5238 | /* |
5239 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
5240 | * But there are some cases where it has already been normalized: | |
5241 | * | |
5242 | * - Moving a forked child which is waiting for being woken up by | |
5243 | * wake_up_new_task(). | |
62af3783 DN |
5244 | * - Moving a task which has been woken up by try_to_wake_up() and |
5245 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
5246 | * |
5247 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
5248 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
5249 | */ | |
62af3783 | 5250 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
5251 | on_rq = 1; |
5252 | ||
b2b5ce02 PZ |
5253 | if (!on_rq) |
5254 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
5255 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 5256 | if (!on_rq) |
b2b5ce02 | 5257 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 | 5258 | } |
029632fb PZ |
5259 | |
5260 | void free_fair_sched_group(struct task_group *tg) | |
5261 | { | |
5262 | int i; | |
5263 | ||
5264 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5265 | ||
5266 | for_each_possible_cpu(i) { | |
5267 | if (tg->cfs_rq) | |
5268 | kfree(tg->cfs_rq[i]); | |
5269 | if (tg->se) | |
5270 | kfree(tg->se[i]); | |
5271 | } | |
5272 | ||
5273 | kfree(tg->cfs_rq); | |
5274 | kfree(tg->se); | |
5275 | } | |
5276 | ||
5277 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5278 | { | |
5279 | struct cfs_rq *cfs_rq; | |
5280 | struct sched_entity *se; | |
5281 | int i; | |
5282 | ||
5283 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
5284 | if (!tg->cfs_rq) | |
5285 | goto err; | |
5286 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
5287 | if (!tg->se) | |
5288 | goto err; | |
5289 | ||
5290 | tg->shares = NICE_0_LOAD; | |
5291 | ||
5292 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5293 | ||
5294 | for_each_possible_cpu(i) { | |
5295 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
5296 | GFP_KERNEL, cpu_to_node(i)); | |
5297 | if (!cfs_rq) | |
5298 | goto err; | |
5299 | ||
5300 | se = kzalloc_node(sizeof(struct sched_entity), | |
5301 | GFP_KERNEL, cpu_to_node(i)); | |
5302 | if (!se) | |
5303 | goto err_free_rq; | |
5304 | ||
5305 | init_cfs_rq(cfs_rq); | |
5306 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
5307 | } | |
5308 | ||
5309 | return 1; | |
5310 | ||
5311 | err_free_rq: | |
5312 | kfree(cfs_rq); | |
5313 | err: | |
5314 | return 0; | |
5315 | } | |
5316 | ||
5317 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
5318 | { | |
5319 | struct rq *rq = cpu_rq(cpu); | |
5320 | unsigned long flags; | |
5321 | ||
5322 | /* | |
5323 | * Only empty task groups can be destroyed; so we can speculatively | |
5324 | * check on_list without danger of it being re-added. | |
5325 | */ | |
5326 | if (!tg->cfs_rq[cpu]->on_list) | |
5327 | return; | |
5328 | ||
5329 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5330 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
5331 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5332 | } | |
5333 | ||
5334 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
5335 | struct sched_entity *se, int cpu, | |
5336 | struct sched_entity *parent) | |
5337 | { | |
5338 | struct rq *rq = cpu_rq(cpu); | |
5339 | ||
5340 | cfs_rq->tg = tg; | |
5341 | cfs_rq->rq = rq; | |
5342 | #ifdef CONFIG_SMP | |
5343 | /* allow initial update_cfs_load() to truncate */ | |
5344 | cfs_rq->load_stamp = 1; | |
810b3817 | 5345 | #endif |
029632fb PZ |
5346 | init_cfs_rq_runtime(cfs_rq); |
5347 | ||
5348 | tg->cfs_rq[cpu] = cfs_rq; | |
5349 | tg->se[cpu] = se; | |
5350 | ||
5351 | /* se could be NULL for root_task_group */ | |
5352 | if (!se) | |
5353 | return; | |
5354 | ||
5355 | if (!parent) | |
5356 | se->cfs_rq = &rq->cfs; | |
5357 | else | |
5358 | se->cfs_rq = parent->my_q; | |
5359 | ||
5360 | se->my_q = cfs_rq; | |
5361 | update_load_set(&se->load, 0); | |
5362 | se->parent = parent; | |
5363 | } | |
5364 | ||
5365 | static DEFINE_MUTEX(shares_mutex); | |
5366 | ||
5367 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
5368 | { | |
5369 | int i; | |
5370 | unsigned long flags; | |
5371 | ||
5372 | /* | |
5373 | * We can't change the weight of the root cgroup. | |
5374 | */ | |
5375 | if (!tg->se[0]) | |
5376 | return -EINVAL; | |
5377 | ||
5378 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
5379 | ||
5380 | mutex_lock(&shares_mutex); | |
5381 | if (tg->shares == shares) | |
5382 | goto done; | |
5383 | ||
5384 | tg->shares = shares; | |
5385 | for_each_possible_cpu(i) { | |
5386 | struct rq *rq = cpu_rq(i); | |
5387 | struct sched_entity *se; | |
5388 | ||
5389 | se = tg->se[i]; | |
5390 | /* Propagate contribution to hierarchy */ | |
5391 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5392 | for_each_sched_entity(se) | |
5393 | update_cfs_shares(group_cfs_rq(se)); | |
5394 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5395 | } | |
5396 | ||
5397 | done: | |
5398 | mutex_unlock(&shares_mutex); | |
5399 | return 0; | |
5400 | } | |
5401 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
5402 | ||
5403 | void free_fair_sched_group(struct task_group *tg) { } | |
5404 | ||
5405 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5406 | { | |
5407 | return 1; | |
5408 | } | |
5409 | ||
5410 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
5411 | ||
5412 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
5413 | ||
810b3817 | 5414 | |
6d686f45 | 5415 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
5416 | { |
5417 | struct sched_entity *se = &task->se; | |
0d721cea PW |
5418 | unsigned int rr_interval = 0; |
5419 | ||
5420 | /* | |
5421 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
5422 | * idle runqueue: | |
5423 | */ | |
0d721cea PW |
5424 | if (rq->cfs.load.weight) |
5425 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
5426 | |
5427 | return rr_interval; | |
5428 | } | |
5429 | ||
bf0f6f24 IM |
5430 | /* |
5431 | * All the scheduling class methods: | |
5432 | */ | |
029632fb | 5433 | const struct sched_class fair_sched_class = { |
5522d5d5 | 5434 | .next = &idle_sched_class, |
bf0f6f24 IM |
5435 | .enqueue_task = enqueue_task_fair, |
5436 | .dequeue_task = dequeue_task_fair, | |
5437 | .yield_task = yield_task_fair, | |
d95f4122 | 5438 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 5439 | |
2e09bf55 | 5440 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
5441 | |
5442 | .pick_next_task = pick_next_task_fair, | |
5443 | .put_prev_task = put_prev_task_fair, | |
5444 | ||
681f3e68 | 5445 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
5446 | .select_task_rq = select_task_rq_fair, |
5447 | ||
0bcdcf28 CE |
5448 | .rq_online = rq_online_fair, |
5449 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
5450 | |
5451 | .task_waking = task_waking_fair, | |
681f3e68 | 5452 | #endif |
bf0f6f24 | 5453 | |
83b699ed | 5454 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 5455 | .task_tick = task_tick_fair, |
cd29fe6f | 5456 | .task_fork = task_fork_fair, |
cb469845 SR |
5457 | |
5458 | .prio_changed = prio_changed_fair, | |
da7a735e | 5459 | .switched_from = switched_from_fair, |
cb469845 | 5460 | .switched_to = switched_to_fair, |
810b3817 | 5461 | |
0d721cea PW |
5462 | .get_rr_interval = get_rr_interval_fair, |
5463 | ||
810b3817 | 5464 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5465 | .task_move_group = task_move_group_fair, |
810b3817 | 5466 | #endif |
bf0f6f24 IM |
5467 | }; |
5468 | ||
5469 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 5470 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 5471 | { |
bf0f6f24 IM |
5472 | struct cfs_rq *cfs_rq; |
5473 | ||
5973e5b9 | 5474 | rcu_read_lock(); |
c3b64f1e | 5475 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 5476 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 5477 | rcu_read_unlock(); |
bf0f6f24 IM |
5478 | } |
5479 | #endif | |
029632fb PZ |
5480 | |
5481 | __init void init_sched_fair_class(void) | |
5482 | { | |
5483 | #ifdef CONFIG_SMP | |
5484 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
5485 | ||
5486 | #ifdef CONFIG_NO_HZ | |
554cecaf | 5487 | nohz.next_balance = jiffies; |
029632fb | 5488 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 5489 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
5490 | #endif |
5491 | #endif /* SMP */ | |
5492 | ||
5493 | } |