]>
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 | 29 | #include <linux/mempolicy.h> |
e14808b4 | 30 | #include <linux/migrate.h> |
cbee9f88 | 31 | #include <linux/task_work.h> |
029632fb PZ |
32 | |
33 | #include <trace/events/sched.h> | |
34 | ||
35 | #include "sched.h" | |
9745512c | 36 | |
bf0f6f24 | 37 | /* |
21805085 | 38 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 39 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 40 | * |
21805085 | 41 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
42 | * 'timeslice length' - timeslices in CFS are of variable length |
43 | * and have no persistent notion like in traditional, time-slice | |
44 | * based scheduling concepts. | |
bf0f6f24 | 45 | * |
d274a4ce IM |
46 | * (to see the precise effective timeslice length of your workload, |
47 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 48 | */ |
21406928 MG |
49 | unsigned int sysctl_sched_latency = 6000000ULL; |
50 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 51 | |
1983a922 CE |
52 | /* |
53 | * The initial- and re-scaling of tunables is configurable | |
54 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
55 | * | |
56 | * Options are: | |
57 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
58 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
59 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
60 | */ | |
61 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
62 | = SCHED_TUNABLESCALING_LOG; | |
63 | ||
2bd8e6d4 | 64 | /* |
b2be5e96 | 65 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 66 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 67 | */ |
0bf377bb IM |
68 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
69 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
70 | |
71 | /* | |
b2be5e96 PZ |
72 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
73 | */ | |
0bf377bb | 74 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
75 | |
76 | /* | |
2bba22c5 | 77 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 78 | * parent will (try to) run first. |
21805085 | 79 | */ |
2bba22c5 | 80 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 81 | |
bf0f6f24 IM |
82 | /* |
83 | * SCHED_OTHER wake-up granularity. | |
172e082a | 84 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
85 | * |
86 | * This option delays the preemption effects of decoupled workloads | |
87 | * and reduces their over-scheduling. Synchronous workloads will still | |
88 | * have immediate wakeup/sleep latencies. | |
89 | */ | |
172e082a | 90 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 91 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 92 | |
da84d961 IM |
93 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
94 | ||
a7a4f8a7 PT |
95 | /* |
96 | * The exponential sliding window over which load is averaged for shares | |
97 | * distribution. | |
98 | * (default: 10msec) | |
99 | */ | |
100 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
101 | ||
ec12cb7f PT |
102 | #ifdef CONFIG_CFS_BANDWIDTH |
103 | /* | |
104 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
105 | * each time a cfs_rq requests quota. | |
106 | * | |
107 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
108 | * to consumption or the quota being specified to be smaller than the slice) | |
109 | * we will always only issue the remaining available time. | |
110 | * | |
111 | * default: 5 msec, units: microseconds | |
112 | */ | |
113 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
114 | #endif | |
115 | ||
029632fb PZ |
116 | /* |
117 | * Increase the granularity value when there are more CPUs, | |
118 | * because with more CPUs the 'effective latency' as visible | |
119 | * to users decreases. But the relationship is not linear, | |
120 | * so pick a second-best guess by going with the log2 of the | |
121 | * number of CPUs. | |
122 | * | |
123 | * This idea comes from the SD scheduler of Con Kolivas: | |
124 | */ | |
125 | static int get_update_sysctl_factor(void) | |
126 | { | |
127 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
128 | unsigned int factor; | |
129 | ||
130 | switch (sysctl_sched_tunable_scaling) { | |
131 | case SCHED_TUNABLESCALING_NONE: | |
132 | factor = 1; | |
133 | break; | |
134 | case SCHED_TUNABLESCALING_LINEAR: | |
135 | factor = cpus; | |
136 | break; | |
137 | case SCHED_TUNABLESCALING_LOG: | |
138 | default: | |
139 | factor = 1 + ilog2(cpus); | |
140 | break; | |
141 | } | |
142 | ||
143 | return factor; | |
144 | } | |
145 | ||
146 | static void update_sysctl(void) | |
147 | { | |
148 | unsigned int factor = get_update_sysctl_factor(); | |
149 | ||
150 | #define SET_SYSCTL(name) \ | |
151 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
152 | SET_SYSCTL(sched_min_granularity); | |
153 | SET_SYSCTL(sched_latency); | |
154 | SET_SYSCTL(sched_wakeup_granularity); | |
155 | #undef SET_SYSCTL | |
156 | } | |
157 | ||
158 | void sched_init_granularity(void) | |
159 | { | |
160 | update_sysctl(); | |
161 | } | |
162 | ||
163 | #if BITS_PER_LONG == 32 | |
164 | # define WMULT_CONST (~0UL) | |
165 | #else | |
166 | # define WMULT_CONST (1UL << 32) | |
167 | #endif | |
168 | ||
169 | #define WMULT_SHIFT 32 | |
170 | ||
171 | /* | |
172 | * Shift right and round: | |
173 | */ | |
174 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
175 | ||
176 | /* | |
177 | * delta *= weight / lw | |
178 | */ | |
179 | static unsigned long | |
180 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
181 | struct load_weight *lw) | |
182 | { | |
183 | u64 tmp; | |
184 | ||
185 | /* | |
186 | * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched | |
187 | * entities since MIN_SHARES = 2. Treat weight as 1 if less than | |
188 | * 2^SCHED_LOAD_RESOLUTION. | |
189 | */ | |
190 | if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) | |
191 | tmp = (u64)delta_exec * scale_load_down(weight); | |
192 | else | |
193 | tmp = (u64)delta_exec; | |
194 | ||
195 | if (!lw->inv_weight) { | |
196 | unsigned long w = scale_load_down(lw->weight); | |
197 | ||
198 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
199 | lw->inv_weight = 1; | |
200 | else if (unlikely(!w)) | |
201 | lw->inv_weight = WMULT_CONST; | |
202 | else | |
203 | lw->inv_weight = WMULT_CONST / w; | |
204 | } | |
205 | ||
206 | /* | |
207 | * Check whether we'd overflow the 64-bit multiplication: | |
208 | */ | |
209 | if (unlikely(tmp > WMULT_CONST)) | |
210 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
211 | WMULT_SHIFT/2); | |
212 | else | |
213 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
214 | ||
215 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
216 | } | |
217 | ||
218 | ||
219 | const struct sched_class fair_sched_class; | |
a4c2f00f | 220 | |
bf0f6f24 IM |
221 | /************************************************************** |
222 | * CFS operations on generic schedulable entities: | |
223 | */ | |
224 | ||
62160e3f | 225 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 226 | |
62160e3f | 227 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
228 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
229 | { | |
62160e3f | 230 | return cfs_rq->rq; |
bf0f6f24 IM |
231 | } |
232 | ||
62160e3f IM |
233 | /* An entity is a task if it doesn't "own" a runqueue */ |
234 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 235 | |
8f48894f PZ |
236 | static inline struct task_struct *task_of(struct sched_entity *se) |
237 | { | |
238 | #ifdef CONFIG_SCHED_DEBUG | |
239 | WARN_ON_ONCE(!entity_is_task(se)); | |
240 | #endif | |
241 | return container_of(se, struct task_struct, se); | |
242 | } | |
243 | ||
b758149c PZ |
244 | /* Walk up scheduling entities hierarchy */ |
245 | #define for_each_sched_entity(se) \ | |
246 | for (; se; se = se->parent) | |
247 | ||
248 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
249 | { | |
250 | return p->se.cfs_rq; | |
251 | } | |
252 | ||
253 | /* runqueue on which this entity is (to be) queued */ | |
254 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
255 | { | |
256 | return se->cfs_rq; | |
257 | } | |
258 | ||
259 | /* runqueue "owned" by this group */ | |
260 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
261 | { | |
262 | return grp->my_q; | |
263 | } | |
264 | ||
3d4b47b4 PZ |
265 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
266 | { | |
267 | if (!cfs_rq->on_list) { | |
67e86250 PT |
268 | /* |
269 | * Ensure we either appear before our parent (if already | |
270 | * enqueued) or force our parent to appear after us when it is | |
271 | * enqueued. The fact that we always enqueue bottom-up | |
272 | * reduces this to two cases. | |
273 | */ | |
274 | if (cfs_rq->tg->parent && | |
275 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
276 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
277 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
278 | } else { | |
279 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 280 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 281 | } |
3d4b47b4 PZ |
282 | |
283 | cfs_rq->on_list = 1; | |
284 | } | |
285 | } | |
286 | ||
287 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
288 | { | |
289 | if (cfs_rq->on_list) { | |
290 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
291 | cfs_rq->on_list = 0; | |
292 | } | |
293 | } | |
294 | ||
b758149c PZ |
295 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
296 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
297 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
298 | ||
299 | /* Do the two (enqueued) entities belong to the same group ? */ | |
300 | static inline int | |
301 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
302 | { | |
303 | if (se->cfs_rq == pse->cfs_rq) | |
304 | return 1; | |
305 | ||
306 | return 0; | |
307 | } | |
308 | ||
309 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
310 | { | |
311 | return se->parent; | |
312 | } | |
313 | ||
464b7527 PZ |
314 | /* return depth at which a sched entity is present in the hierarchy */ |
315 | static inline int depth_se(struct sched_entity *se) | |
316 | { | |
317 | int depth = 0; | |
318 | ||
319 | for_each_sched_entity(se) | |
320 | depth++; | |
321 | ||
322 | return depth; | |
323 | } | |
324 | ||
325 | static void | |
326 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
327 | { | |
328 | int se_depth, pse_depth; | |
329 | ||
330 | /* | |
331 | * preemption test can be made between sibling entities who are in the | |
332 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
333 | * both tasks until we find their ancestors who are siblings of common | |
334 | * parent. | |
335 | */ | |
336 | ||
337 | /* First walk up until both entities are at same depth */ | |
338 | se_depth = depth_se(*se); | |
339 | pse_depth = depth_se(*pse); | |
340 | ||
341 | while (se_depth > pse_depth) { | |
342 | se_depth--; | |
343 | *se = parent_entity(*se); | |
344 | } | |
345 | ||
346 | while (pse_depth > se_depth) { | |
347 | pse_depth--; | |
348 | *pse = parent_entity(*pse); | |
349 | } | |
350 | ||
351 | while (!is_same_group(*se, *pse)) { | |
352 | *se = parent_entity(*se); | |
353 | *pse = parent_entity(*pse); | |
354 | } | |
355 | } | |
356 | ||
8f48894f PZ |
357 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
358 | ||
359 | static inline struct task_struct *task_of(struct sched_entity *se) | |
360 | { | |
361 | return container_of(se, struct task_struct, se); | |
362 | } | |
bf0f6f24 | 363 | |
62160e3f IM |
364 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
365 | { | |
366 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
367 | } |
368 | ||
369 | #define entity_is_task(se) 1 | |
370 | ||
b758149c PZ |
371 | #define for_each_sched_entity(se) \ |
372 | for (; se; se = NULL) | |
bf0f6f24 | 373 | |
b758149c | 374 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 375 | { |
b758149c | 376 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
377 | } |
378 | ||
b758149c PZ |
379 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
380 | { | |
381 | struct task_struct *p = task_of(se); | |
382 | struct rq *rq = task_rq(p); | |
383 | ||
384 | return &rq->cfs; | |
385 | } | |
386 | ||
387 | /* runqueue "owned" by this group */ | |
388 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
389 | { | |
390 | return NULL; | |
391 | } | |
392 | ||
3d4b47b4 PZ |
393 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
394 | { | |
395 | } | |
396 | ||
397 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
398 | { | |
399 | } | |
400 | ||
b758149c PZ |
401 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
402 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
403 | ||
404 | static inline int | |
405 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
406 | { | |
407 | return 1; | |
408 | } | |
409 | ||
410 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
411 | { | |
412 | return NULL; | |
413 | } | |
414 | ||
464b7527 PZ |
415 | static inline void |
416 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
417 | { | |
418 | } | |
419 | ||
b758149c PZ |
420 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
421 | ||
6c16a6dc PZ |
422 | static __always_inline |
423 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
424 | |
425 | /************************************************************** | |
426 | * Scheduling class tree data structure manipulation methods: | |
427 | */ | |
428 | ||
0702e3eb | 429 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 430 | { |
368059a9 PZ |
431 | s64 delta = (s64)(vruntime - min_vruntime); |
432 | if (delta > 0) | |
02e0431a PZ |
433 | min_vruntime = vruntime; |
434 | ||
435 | return min_vruntime; | |
436 | } | |
437 | ||
0702e3eb | 438 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
439 | { |
440 | s64 delta = (s64)(vruntime - min_vruntime); | |
441 | if (delta < 0) | |
442 | min_vruntime = vruntime; | |
443 | ||
444 | return min_vruntime; | |
445 | } | |
446 | ||
54fdc581 FC |
447 | static inline int entity_before(struct sched_entity *a, |
448 | struct sched_entity *b) | |
449 | { | |
450 | return (s64)(a->vruntime - b->vruntime) < 0; | |
451 | } | |
452 | ||
1af5f730 PZ |
453 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
454 | { | |
455 | u64 vruntime = cfs_rq->min_vruntime; | |
456 | ||
457 | if (cfs_rq->curr) | |
458 | vruntime = cfs_rq->curr->vruntime; | |
459 | ||
460 | if (cfs_rq->rb_leftmost) { | |
461 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
462 | struct sched_entity, | |
463 | run_node); | |
464 | ||
e17036da | 465 | if (!cfs_rq->curr) |
1af5f730 PZ |
466 | vruntime = se->vruntime; |
467 | else | |
468 | vruntime = min_vruntime(vruntime, se->vruntime); | |
469 | } | |
470 | ||
471 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
3fe1698b PZ |
472 | #ifndef CONFIG_64BIT |
473 | smp_wmb(); | |
474 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
475 | #endif | |
1af5f730 PZ |
476 | } |
477 | ||
bf0f6f24 IM |
478 | /* |
479 | * Enqueue an entity into the rb-tree: | |
480 | */ | |
0702e3eb | 481 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
482 | { |
483 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
484 | struct rb_node *parent = NULL; | |
485 | struct sched_entity *entry; | |
bf0f6f24 IM |
486 | int leftmost = 1; |
487 | ||
488 | /* | |
489 | * Find the right place in the rbtree: | |
490 | */ | |
491 | while (*link) { | |
492 | parent = *link; | |
493 | entry = rb_entry(parent, struct sched_entity, run_node); | |
494 | /* | |
495 | * We dont care about collisions. Nodes with | |
496 | * the same key stay together. | |
497 | */ | |
2bd2d6f2 | 498 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
499 | link = &parent->rb_left; |
500 | } else { | |
501 | link = &parent->rb_right; | |
502 | leftmost = 0; | |
503 | } | |
504 | } | |
505 | ||
506 | /* | |
507 | * Maintain a cache of leftmost tree entries (it is frequently | |
508 | * used): | |
509 | */ | |
1af5f730 | 510 | if (leftmost) |
57cb499d | 511 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
512 | |
513 | rb_link_node(&se->run_node, parent, link); | |
514 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
515 | } |
516 | ||
0702e3eb | 517 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 518 | { |
3fe69747 PZ |
519 | if (cfs_rq->rb_leftmost == &se->run_node) { |
520 | struct rb_node *next_node; | |
3fe69747 PZ |
521 | |
522 | next_node = rb_next(&se->run_node); | |
523 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 524 | } |
e9acbff6 | 525 | |
bf0f6f24 | 526 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
527 | } |
528 | ||
029632fb | 529 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 530 | { |
f4b6755f PZ |
531 | struct rb_node *left = cfs_rq->rb_leftmost; |
532 | ||
533 | if (!left) | |
534 | return NULL; | |
535 | ||
536 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
537 | } |
538 | ||
ac53db59 RR |
539 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
540 | { | |
541 | struct rb_node *next = rb_next(&se->run_node); | |
542 | ||
543 | if (!next) | |
544 | return NULL; | |
545 | ||
546 | return rb_entry(next, struct sched_entity, run_node); | |
547 | } | |
548 | ||
549 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 550 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 551 | { |
7eee3e67 | 552 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 553 | |
70eee74b BS |
554 | if (!last) |
555 | return NULL; | |
7eee3e67 IM |
556 | |
557 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
558 | } |
559 | ||
bf0f6f24 IM |
560 | /************************************************************** |
561 | * Scheduling class statistics methods: | |
562 | */ | |
563 | ||
acb4a848 | 564 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 565 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
566 | loff_t *ppos) |
567 | { | |
8d65af78 | 568 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 569 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
570 | |
571 | if (ret || !write) | |
572 | return ret; | |
573 | ||
574 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
575 | sysctl_sched_min_granularity); | |
576 | ||
acb4a848 CE |
577 | #define WRT_SYSCTL(name) \ |
578 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
579 | WRT_SYSCTL(sched_min_granularity); | |
580 | WRT_SYSCTL(sched_latency); | |
581 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
582 | #undef WRT_SYSCTL |
583 | ||
b2be5e96 PZ |
584 | return 0; |
585 | } | |
586 | #endif | |
647e7cac | 587 | |
a7be37ac | 588 | /* |
f9c0b095 | 589 | * delta /= w |
a7be37ac PZ |
590 | */ |
591 | static inline unsigned long | |
592 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
593 | { | |
f9c0b095 PZ |
594 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
595 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
596 | |
597 | return delta; | |
598 | } | |
599 | ||
647e7cac IM |
600 | /* |
601 | * The idea is to set a period in which each task runs once. | |
602 | * | |
532b1858 | 603 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
604 | * this period because otherwise the slices get too small. |
605 | * | |
606 | * p = (nr <= nl) ? l : l*nr/nl | |
607 | */ | |
4d78e7b6 PZ |
608 | static u64 __sched_period(unsigned long nr_running) |
609 | { | |
610 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 611 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
612 | |
613 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 614 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 615 | period *= nr_running; |
4d78e7b6 PZ |
616 | } |
617 | ||
618 | return period; | |
619 | } | |
620 | ||
647e7cac IM |
621 | /* |
622 | * We calculate the wall-time slice from the period by taking a part | |
623 | * proportional to the weight. | |
624 | * | |
f9c0b095 | 625 | * s = p*P[w/rw] |
647e7cac | 626 | */ |
6d0f0ebd | 627 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 628 | { |
0a582440 | 629 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 630 | |
0a582440 | 631 | for_each_sched_entity(se) { |
6272d68c | 632 | struct load_weight *load; |
3104bf03 | 633 | struct load_weight lw; |
6272d68c LM |
634 | |
635 | cfs_rq = cfs_rq_of(se); | |
636 | load = &cfs_rq->load; | |
f9c0b095 | 637 | |
0a582440 | 638 | if (unlikely(!se->on_rq)) { |
3104bf03 | 639 | lw = cfs_rq->load; |
0a582440 MG |
640 | |
641 | update_load_add(&lw, se->load.weight); | |
642 | load = &lw; | |
643 | } | |
644 | slice = calc_delta_mine(slice, se->load.weight, load); | |
645 | } | |
646 | return slice; | |
bf0f6f24 IM |
647 | } |
648 | ||
647e7cac | 649 | /* |
ac884dec | 650 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 651 | * |
f9c0b095 | 652 | * vs = s/w |
647e7cac | 653 | */ |
f9c0b095 | 654 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 655 | { |
f9c0b095 | 656 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
657 | } |
658 | ||
d6b55918 | 659 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); |
6d5ab293 | 660 | static void update_cfs_shares(struct cfs_rq *cfs_rq); |
3b3d190e | 661 | |
bf0f6f24 IM |
662 | /* |
663 | * Update the current task's runtime statistics. Skip current tasks that | |
664 | * are not in our scheduling class. | |
665 | */ | |
666 | static inline void | |
8ebc91d9 IM |
667 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
668 | unsigned long delta_exec) | |
bf0f6f24 | 669 | { |
bbdba7c0 | 670 | unsigned long delta_exec_weighted; |
bf0f6f24 | 671 | |
41acab88 LDM |
672 | schedstat_set(curr->statistics.exec_max, |
673 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
674 | |
675 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 676 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 677 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 678 | |
e9acbff6 | 679 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 680 | update_min_vruntime(cfs_rq); |
3b3d190e | 681 | |
70caf8a6 | 682 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
3b3d190e | 683 | cfs_rq->load_unacc_exec_time += delta_exec; |
3b3d190e | 684 | #endif |
bf0f6f24 IM |
685 | } |
686 | ||
b7cc0896 | 687 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 688 | { |
429d43bc | 689 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 690 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
691 | unsigned long delta_exec; |
692 | ||
693 | if (unlikely(!curr)) | |
694 | return; | |
695 | ||
696 | /* | |
697 | * Get the amount of time the current task was running | |
698 | * since the last time we changed load (this cannot | |
699 | * overflow on 32 bits): | |
700 | */ | |
8ebc91d9 | 701 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
702 | if (!delta_exec) |
703 | return; | |
bf0f6f24 | 704 | |
8ebc91d9 IM |
705 | __update_curr(cfs_rq, curr, delta_exec); |
706 | curr->exec_start = now; | |
d842de87 SV |
707 | |
708 | if (entity_is_task(curr)) { | |
709 | struct task_struct *curtask = task_of(curr); | |
710 | ||
f977bb49 | 711 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 712 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 713 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 714 | } |
ec12cb7f PT |
715 | |
716 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
717 | } |
718 | ||
719 | static inline void | |
5870db5b | 720 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 721 | { |
41acab88 | 722 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
723 | } |
724 | ||
bf0f6f24 IM |
725 | /* |
726 | * Task is being enqueued - update stats: | |
727 | */ | |
d2417e5a | 728 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 729 | { |
bf0f6f24 IM |
730 | /* |
731 | * Are we enqueueing a waiting task? (for current tasks | |
732 | * a dequeue/enqueue event is a NOP) | |
733 | */ | |
429d43bc | 734 | if (se != cfs_rq->curr) |
5870db5b | 735 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
736 | } |
737 | ||
bf0f6f24 | 738 | static void |
9ef0a961 | 739 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 740 | { |
41acab88 LDM |
741 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
742 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
743 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
744 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
745 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
746 | #ifdef CONFIG_SCHEDSTATS |
747 | if (entity_is_task(se)) { | |
748 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 749 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
750 | } |
751 | #endif | |
41acab88 | 752 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
753 | } |
754 | ||
755 | static inline void | |
19b6a2e3 | 756 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 757 | { |
bf0f6f24 IM |
758 | /* |
759 | * Mark the end of the wait period if dequeueing a | |
760 | * waiting task: | |
761 | */ | |
429d43bc | 762 | if (se != cfs_rq->curr) |
9ef0a961 | 763 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
764 | } |
765 | ||
766 | /* | |
767 | * We are picking a new current task - update its stats: | |
768 | */ | |
769 | static inline void | |
79303e9e | 770 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
771 | { |
772 | /* | |
773 | * We are starting a new run period: | |
774 | */ | |
305e6835 | 775 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
776 | } |
777 | ||
bf0f6f24 IM |
778 | /************************************************** |
779 | * Scheduling class queueing methods: | |
780 | */ | |
781 | ||
cbee9f88 PZ |
782 | #ifdef CONFIG_NUMA_BALANCING |
783 | /* | |
6e5fb223 | 784 | * numa task sample period in ms |
cbee9f88 | 785 | */ |
6e5fb223 | 786 | unsigned int sysctl_numa_balancing_scan_period_min = 100; |
b8593bfd MG |
787 | unsigned int sysctl_numa_balancing_scan_period_max = 100*50; |
788 | unsigned int sysctl_numa_balancing_scan_period_reset = 100*600; | |
6e5fb223 PZ |
789 | |
790 | /* Portion of address space to scan in MB */ | |
791 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 792 | |
4b96a29b PZ |
793 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
794 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
795 | ||
cbee9f88 PZ |
796 | static void task_numa_placement(struct task_struct *p) |
797 | { | |
798 | int seq = ACCESS_ONCE(p->mm->numa_scan_seq); | |
799 | ||
800 | if (p->numa_scan_seq == seq) | |
801 | return; | |
802 | p->numa_scan_seq = seq; | |
803 | ||
804 | /* FIXME: Scheduling placement policy hints go here */ | |
805 | } | |
806 | ||
807 | /* | |
808 | * Got a PROT_NONE fault for a page on @node. | |
809 | */ | |
b8593bfd | 810 | void task_numa_fault(int node, int pages, bool migrated) |
cbee9f88 PZ |
811 | { |
812 | struct task_struct *p = current; | |
813 | ||
1a687c2e MG |
814 | if (!sched_feat_numa(NUMA)) |
815 | return; | |
816 | ||
cbee9f88 PZ |
817 | /* FIXME: Allocate task-specific structure for placement policy here */ |
818 | ||
fb003b80 | 819 | /* |
b8593bfd MG |
820 | * If pages are properly placed (did not migrate) then scan slower. |
821 | * This is reset periodically in case of phase changes | |
fb003b80 | 822 | */ |
b8593bfd MG |
823 | if (!migrated) |
824 | p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max, | |
825 | p->numa_scan_period + jiffies_to_msecs(10)); | |
fb003b80 | 826 | |
cbee9f88 PZ |
827 | task_numa_placement(p); |
828 | } | |
829 | ||
6e5fb223 PZ |
830 | static void reset_ptenuma_scan(struct task_struct *p) |
831 | { | |
832 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
833 | p->mm->numa_scan_offset = 0; | |
834 | } | |
835 | ||
cbee9f88 PZ |
836 | /* |
837 | * The expensive part of numa migration is done from task_work context. | |
838 | * Triggered from task_tick_numa(). | |
839 | */ | |
840 | void task_numa_work(struct callback_head *work) | |
841 | { | |
842 | unsigned long migrate, next_scan, now = jiffies; | |
843 | struct task_struct *p = current; | |
844 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 845 | struct vm_area_struct *vma; |
9f40604c MG |
846 | unsigned long start, end; |
847 | long pages; | |
cbee9f88 PZ |
848 | |
849 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
850 | ||
851 | work->next = work; /* protect against double add */ | |
852 | /* | |
853 | * Who cares about NUMA placement when they're dying. | |
854 | * | |
855 | * NOTE: make sure not to dereference p->mm before this check, | |
856 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
857 | * without p->mm even though we still had it when we enqueued this | |
858 | * work. | |
859 | */ | |
860 | if (p->flags & PF_EXITING) | |
861 | return; | |
862 | ||
b8593bfd MG |
863 | /* |
864 | * Reset the scan period if enough time has gone by. Objective is that | |
865 | * scanning will be reduced if pages are properly placed. As tasks | |
866 | * can enter different phases this needs to be re-examined. Lacking | |
867 | * proper tracking of reference behaviour, this blunt hammer is used. | |
868 | */ | |
869 | migrate = mm->numa_next_reset; | |
870 | if (time_after(now, migrate)) { | |
871 | p->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
872 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
873 | xchg(&mm->numa_next_reset, next_scan); | |
874 | } | |
875 | ||
cbee9f88 PZ |
876 | /* |
877 | * Enforce maximal scan/migration frequency.. | |
878 | */ | |
879 | migrate = mm->numa_next_scan; | |
880 | if (time_before(now, migrate)) | |
881 | return; | |
882 | ||
883 | if (p->numa_scan_period == 0) | |
884 | p->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
885 | ||
fb003b80 | 886 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
887 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
888 | return; | |
889 | ||
e14808b4 MG |
890 | /* |
891 | * Do not set pte_numa if the current running node is rate-limited. | |
892 | * This loses statistics on the fault but if we are unwilling to | |
893 | * migrate to this node, it is less likely we can do useful work | |
894 | */ | |
895 | if (migrate_ratelimited(numa_node_id())) | |
896 | return; | |
897 | ||
9f40604c MG |
898 | start = mm->numa_scan_offset; |
899 | pages = sysctl_numa_balancing_scan_size; | |
900 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
901 | if (!pages) | |
902 | return; | |
cbee9f88 | 903 | |
6e5fb223 | 904 | down_read(&mm->mmap_sem); |
9f40604c | 905 | vma = find_vma(mm, start); |
6e5fb223 PZ |
906 | if (!vma) { |
907 | reset_ptenuma_scan(p); | |
9f40604c | 908 | start = 0; |
6e5fb223 PZ |
909 | vma = mm->mmap; |
910 | } | |
9f40604c | 911 | for (; vma; vma = vma->vm_next) { |
6e5fb223 PZ |
912 | if (!vma_migratable(vma)) |
913 | continue; | |
914 | ||
915 | /* Skip small VMAs. They are not likely to be of relevance */ | |
916 | if (((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) < HPAGE_PMD_NR) | |
917 | continue; | |
918 | ||
9f40604c MG |
919 | do { |
920 | start = max(start, vma->vm_start); | |
921 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
922 | end = min(end, vma->vm_end); | |
923 | pages -= change_prot_numa(vma, start, end); | |
6e5fb223 | 924 | |
9f40604c MG |
925 | start = end; |
926 | if (pages <= 0) | |
927 | goto out; | |
928 | } while (end != vma->vm_end); | |
cbee9f88 | 929 | } |
6e5fb223 | 930 | |
9f40604c | 931 | out: |
6e5fb223 PZ |
932 | /* |
933 | * It is possible to reach the end of the VMA list but the last few VMAs are | |
934 | * not guaranteed to the vma_migratable. If they are not, we would find the | |
935 | * !migratable VMA on the next scan but not reset the scanner to the start | |
936 | * so check it now. | |
937 | */ | |
938 | if (vma) | |
9f40604c | 939 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
940 | else |
941 | reset_ptenuma_scan(p); | |
942 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
943 | } |
944 | ||
945 | /* | |
946 | * Drive the periodic memory faults.. | |
947 | */ | |
948 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
949 | { | |
950 | struct callback_head *work = &curr->numa_work; | |
951 | u64 period, now; | |
952 | ||
953 | /* | |
954 | * We don't care about NUMA placement if we don't have memory. | |
955 | */ | |
956 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
957 | return; | |
958 | ||
959 | /* | |
960 | * Using runtime rather than walltime has the dual advantage that | |
961 | * we (mostly) drive the selection from busy threads and that the | |
962 | * task needs to have done some actual work before we bother with | |
963 | * NUMA placement. | |
964 | */ | |
965 | now = curr->se.sum_exec_runtime; | |
966 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
967 | ||
968 | if (now - curr->node_stamp > period) { | |
4b96a29b PZ |
969 | if (!curr->node_stamp) |
970 | curr->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
cbee9f88 PZ |
971 | curr->node_stamp = now; |
972 | ||
973 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
974 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
975 | task_work_add(curr, work, true); | |
976 | } | |
977 | } | |
978 | } | |
979 | #else | |
980 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
981 | { | |
982 | } | |
983 | #endif /* CONFIG_NUMA_BALANCING */ | |
984 | ||
30cfdcfc DA |
985 | static void |
986 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
987 | { | |
988 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 989 | if (!parent_entity(se)) |
029632fb | 990 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
991 | #ifdef CONFIG_SMP |
992 | if (entity_is_task(se)) | |
eb95308e | 993 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 994 | #endif |
30cfdcfc | 995 | cfs_rq->nr_running++; |
30cfdcfc DA |
996 | } |
997 | ||
998 | static void | |
999 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1000 | { | |
1001 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1002 | if (!parent_entity(se)) |
029632fb | 1003 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1004 | if (entity_is_task(se)) |
b87f1724 | 1005 | list_del_init(&se->group_node); |
30cfdcfc | 1006 | cfs_rq->nr_running--; |
30cfdcfc DA |
1007 | } |
1008 | ||
3ff6dcac | 1009 | #ifdef CONFIG_FAIR_GROUP_SCHED |
64660c86 PT |
1010 | /* we need this in update_cfs_load and load-balance functions below */ |
1011 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); | |
3ff6dcac | 1012 | # ifdef CONFIG_SMP |
d6b55918 PT |
1013 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
1014 | int global_update) | |
1015 | { | |
1016 | struct task_group *tg = cfs_rq->tg; | |
1017 | long load_avg; | |
1018 | ||
1019 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
1020 | load_avg -= cfs_rq->load_contribution; | |
1021 | ||
1022 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
1023 | atomic_add(load_avg, &tg->load_weight); | |
1024 | cfs_rq->load_contribution += load_avg; | |
1025 | } | |
1026 | } | |
1027 | ||
1028 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 1029 | { |
a7a4f8a7 | 1030 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 1031 | u64 now, delta; |
e33078ba | 1032 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 1033 | |
64660c86 | 1034 | if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq)) |
2069dd75 PZ |
1035 | return; |
1036 | ||
05ca62c6 | 1037 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
1038 | delta = now - cfs_rq->load_stamp; |
1039 | ||
e33078ba PT |
1040 | /* truncate load history at 4 idle periods */ |
1041 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
1042 | now - cfs_rq->load_last > 4 * period) { | |
1043 | cfs_rq->load_period = 0; | |
1044 | cfs_rq->load_avg = 0; | |
f07333bf | 1045 | delta = period - 1; |
e33078ba PT |
1046 | } |
1047 | ||
2069dd75 | 1048 | cfs_rq->load_stamp = now; |
3b3d190e | 1049 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 1050 | cfs_rq->load_period += delta; |
e33078ba PT |
1051 | if (load) { |
1052 | cfs_rq->load_last = now; | |
1053 | cfs_rq->load_avg += delta * load; | |
1054 | } | |
2069dd75 | 1055 | |
d6b55918 PT |
1056 | /* consider updating load contribution on each fold or truncate */ |
1057 | if (global_update || cfs_rq->load_period > period | |
1058 | || !cfs_rq->load_period) | |
1059 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
1060 | ||
2069dd75 PZ |
1061 | while (cfs_rq->load_period > period) { |
1062 | /* | |
1063 | * Inline assembly required to prevent the compiler | |
1064 | * optimising this loop into a divmod call. | |
1065 | * See __iter_div_u64_rem() for another example of this. | |
1066 | */ | |
1067 | asm("" : "+rm" (cfs_rq->load_period)); | |
1068 | cfs_rq->load_period /= 2; | |
1069 | cfs_rq->load_avg /= 2; | |
1070 | } | |
3d4b47b4 | 1071 | |
e33078ba PT |
1072 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
1073 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
1074 | } |
1075 | ||
cf5f0acf PZ |
1076 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1077 | { | |
1078 | long tg_weight; | |
1079 | ||
1080 | /* | |
1081 | * Use this CPU's actual weight instead of the last load_contribution | |
1082 | * to gain a more accurate current total weight. See | |
1083 | * update_cfs_rq_load_contribution(). | |
1084 | */ | |
1085 | tg_weight = atomic_read(&tg->load_weight); | |
1086 | tg_weight -= cfs_rq->load_contribution; | |
1087 | tg_weight += cfs_rq->load.weight; | |
1088 | ||
1089 | return tg_weight; | |
1090 | } | |
1091 | ||
6d5ab293 | 1092 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1093 | { |
cf5f0acf | 1094 | long tg_weight, load, shares; |
3ff6dcac | 1095 | |
cf5f0acf | 1096 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1097 | load = cfs_rq->load.weight; |
3ff6dcac | 1098 | |
3ff6dcac | 1099 | shares = (tg->shares * load); |
cf5f0acf PZ |
1100 | if (tg_weight) |
1101 | shares /= tg_weight; | |
3ff6dcac YZ |
1102 | |
1103 | if (shares < MIN_SHARES) | |
1104 | shares = MIN_SHARES; | |
1105 | if (shares > tg->shares) | |
1106 | shares = tg->shares; | |
1107 | ||
1108 | return shares; | |
1109 | } | |
1110 | ||
1111 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
1112 | { | |
1113 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
1114 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 1115 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
1116 | } |
1117 | } | |
1118 | # else /* CONFIG_SMP */ | |
1119 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
1120 | { | |
1121 | } | |
1122 | ||
6d5ab293 | 1123 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1124 | { |
1125 | return tg->shares; | |
1126 | } | |
1127 | ||
1128 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
1129 | { | |
1130 | } | |
1131 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
1132 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1133 | unsigned long weight) | |
1134 | { | |
19e5eebb PT |
1135 | if (se->on_rq) { |
1136 | /* commit outstanding execution time */ | |
1137 | if (cfs_rq->curr == se) | |
1138 | update_curr(cfs_rq); | |
2069dd75 | 1139 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1140 | } |
2069dd75 PZ |
1141 | |
1142 | update_load_set(&se->load, weight); | |
1143 | ||
1144 | if (se->on_rq) | |
1145 | account_entity_enqueue(cfs_rq, se); | |
1146 | } | |
1147 | ||
6d5ab293 | 1148 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1149 | { |
1150 | struct task_group *tg; | |
1151 | struct sched_entity *se; | |
3ff6dcac | 1152 | long shares; |
2069dd75 | 1153 | |
2069dd75 PZ |
1154 | tg = cfs_rq->tg; |
1155 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1156 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1157 | return; |
3ff6dcac YZ |
1158 | #ifndef CONFIG_SMP |
1159 | if (likely(se->load.weight == tg->shares)) | |
1160 | return; | |
1161 | #endif | |
6d5ab293 | 1162 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1163 | |
1164 | reweight_entity(cfs_rq_of(se), se, shares); | |
1165 | } | |
1166 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 1167 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
1168 | { |
1169 | } | |
1170 | ||
6d5ab293 | 1171 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1172 | { |
1173 | } | |
43365bd7 PT |
1174 | |
1175 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
1176 | { | |
1177 | } | |
2069dd75 PZ |
1178 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
1179 | ||
2396af69 | 1180 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1181 | { |
bf0f6f24 | 1182 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1183 | struct task_struct *tsk = NULL; |
1184 | ||
1185 | if (entity_is_task(se)) | |
1186 | tsk = task_of(se); | |
1187 | ||
41acab88 LDM |
1188 | if (se->statistics.sleep_start) { |
1189 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
1190 | |
1191 | if ((s64)delta < 0) | |
1192 | delta = 0; | |
1193 | ||
41acab88 LDM |
1194 | if (unlikely(delta > se->statistics.sleep_max)) |
1195 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1196 | |
8c79a045 | 1197 | se->statistics.sleep_start = 0; |
41acab88 | 1198 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1199 | |
768d0c27 | 1200 | if (tsk) { |
e414314c | 1201 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1202 | trace_sched_stat_sleep(tsk, delta); |
1203 | } | |
bf0f6f24 | 1204 | } |
41acab88 LDM |
1205 | if (se->statistics.block_start) { |
1206 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
1207 | |
1208 | if ((s64)delta < 0) | |
1209 | delta = 0; | |
1210 | ||
41acab88 LDM |
1211 | if (unlikely(delta > se->statistics.block_max)) |
1212 | se->statistics.block_max = delta; | |
bf0f6f24 | 1213 | |
8c79a045 | 1214 | se->statistics.block_start = 0; |
41acab88 | 1215 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1216 | |
e414314c | 1217 | if (tsk) { |
8f0dfc34 | 1218 | if (tsk->in_iowait) { |
41acab88 LDM |
1219 | se->statistics.iowait_sum += delta; |
1220 | se->statistics.iowait_count++; | |
768d0c27 | 1221 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1222 | } |
1223 | ||
b781a602 AV |
1224 | trace_sched_stat_blocked(tsk, delta); |
1225 | ||
e414314c PZ |
1226 | /* |
1227 | * Blocking time is in units of nanosecs, so shift by | |
1228 | * 20 to get a milliseconds-range estimation of the | |
1229 | * amount of time that the task spent sleeping: | |
1230 | */ | |
1231 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1232 | profile_hits(SLEEP_PROFILING, | |
1233 | (void *)get_wchan(tsk), | |
1234 | delta >> 20); | |
1235 | } | |
1236 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1237 | } |
bf0f6f24 IM |
1238 | } |
1239 | #endif | |
1240 | } | |
1241 | ||
ddc97297 PZ |
1242 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1243 | { | |
1244 | #ifdef CONFIG_SCHED_DEBUG | |
1245 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
1246 | ||
1247 | if (d < 0) | |
1248 | d = -d; | |
1249 | ||
1250 | if (d > 3*sysctl_sched_latency) | |
1251 | schedstat_inc(cfs_rq, nr_spread_over); | |
1252 | #endif | |
1253 | } | |
1254 | ||
aeb73b04 PZ |
1255 | static void |
1256 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
1257 | { | |
1af5f730 | 1258 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 1259 | |
2cb8600e PZ |
1260 | /* |
1261 | * The 'current' period is already promised to the current tasks, | |
1262 | * however the extra weight of the new task will slow them down a | |
1263 | * little, place the new task so that it fits in the slot that | |
1264 | * stays open at the end. | |
1265 | */ | |
94dfb5e7 | 1266 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 1267 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 1268 | |
a2e7a7eb | 1269 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 1270 | if (!initial) { |
a2e7a7eb | 1271 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 1272 | |
a2e7a7eb MG |
1273 | /* |
1274 | * Halve their sleep time's effect, to allow | |
1275 | * for a gentler effect of sleepers: | |
1276 | */ | |
1277 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
1278 | thresh >>= 1; | |
51e0304c | 1279 | |
a2e7a7eb | 1280 | vruntime -= thresh; |
aeb73b04 PZ |
1281 | } |
1282 | ||
b5d9d734 MG |
1283 | /* ensure we never gain time by being placed backwards. */ |
1284 | vruntime = max_vruntime(se->vruntime, vruntime); | |
1285 | ||
67e9fb2a | 1286 | se->vruntime = vruntime; |
aeb73b04 PZ |
1287 | } |
1288 | ||
d3d9dc33 PT |
1289 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
1290 | ||
bf0f6f24 | 1291 | static void |
88ec22d3 | 1292 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1293 | { |
88ec22d3 PZ |
1294 | /* |
1295 | * Update the normalized vruntime before updating min_vruntime | |
1296 | * through callig update_curr(). | |
1297 | */ | |
371fd7e7 | 1298 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
1299 | se->vruntime += cfs_rq->min_vruntime; |
1300 | ||
bf0f6f24 | 1301 | /* |
a2a2d680 | 1302 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1303 | */ |
b7cc0896 | 1304 | update_curr(cfs_rq); |
d6b55918 | 1305 | update_cfs_load(cfs_rq, 0); |
a992241d | 1306 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 1307 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 1308 | |
88ec22d3 | 1309 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 1310 | place_entity(cfs_rq, se, 0); |
2396af69 | 1311 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 1312 | } |
bf0f6f24 | 1313 | |
d2417e5a | 1314 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 1315 | check_spread(cfs_rq, se); |
83b699ed SV |
1316 | if (se != cfs_rq->curr) |
1317 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1318 | se->on_rq = 1; |
3d4b47b4 | 1319 | |
d3d9dc33 | 1320 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1321 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1322 | check_enqueue_throttle(cfs_rq); |
1323 | } | |
bf0f6f24 IM |
1324 | } |
1325 | ||
2c13c919 | 1326 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1327 | { |
2c13c919 RR |
1328 | for_each_sched_entity(se) { |
1329 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1330 | if (cfs_rq->last == se) | |
1331 | cfs_rq->last = NULL; | |
1332 | else | |
1333 | break; | |
1334 | } | |
1335 | } | |
2002c695 | 1336 | |
2c13c919 RR |
1337 | static void __clear_buddies_next(struct sched_entity *se) |
1338 | { | |
1339 | for_each_sched_entity(se) { | |
1340 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1341 | if (cfs_rq->next == se) | |
1342 | cfs_rq->next = NULL; | |
1343 | else | |
1344 | break; | |
1345 | } | |
2002c695 PZ |
1346 | } |
1347 | ||
ac53db59 RR |
1348 | static void __clear_buddies_skip(struct sched_entity *se) |
1349 | { | |
1350 | for_each_sched_entity(se) { | |
1351 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1352 | if (cfs_rq->skip == se) | |
1353 | cfs_rq->skip = NULL; | |
1354 | else | |
1355 | break; | |
1356 | } | |
1357 | } | |
1358 | ||
a571bbea PZ |
1359 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1360 | { | |
2c13c919 RR |
1361 | if (cfs_rq->last == se) |
1362 | __clear_buddies_last(se); | |
1363 | ||
1364 | if (cfs_rq->next == se) | |
1365 | __clear_buddies_next(se); | |
ac53db59 RR |
1366 | |
1367 | if (cfs_rq->skip == se) | |
1368 | __clear_buddies_skip(se); | |
a571bbea PZ |
1369 | } |
1370 | ||
6c16a6dc | 1371 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 1372 | |
bf0f6f24 | 1373 | static void |
371fd7e7 | 1374 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1375 | { |
a2a2d680 DA |
1376 | /* |
1377 | * Update run-time statistics of the 'current'. | |
1378 | */ | |
1379 | update_curr(cfs_rq); | |
1380 | ||
19b6a2e3 | 1381 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1382 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1383 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1384 | if (entity_is_task(se)) { |
1385 | struct task_struct *tsk = task_of(se); | |
1386 | ||
1387 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1388 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1389 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1390 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1391 | } |
db36cc7d | 1392 | #endif |
67e9fb2a PZ |
1393 | } |
1394 | ||
2002c695 | 1395 | clear_buddies(cfs_rq, se); |
4793241b | 1396 | |
83b699ed | 1397 | if (se != cfs_rq->curr) |
30cfdcfc | 1398 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1399 | se->on_rq = 0; |
d6b55918 | 1400 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1401 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1402 | |
1403 | /* | |
1404 | * Normalize the entity after updating the min_vruntime because the | |
1405 | * update can refer to the ->curr item and we need to reflect this | |
1406 | * movement in our normalized position. | |
1407 | */ | |
371fd7e7 | 1408 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1409 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 1410 | |
d8b4986d PT |
1411 | /* return excess runtime on last dequeue */ |
1412 | return_cfs_rq_runtime(cfs_rq); | |
1413 | ||
1e876231 PZ |
1414 | update_min_vruntime(cfs_rq); |
1415 | update_cfs_shares(cfs_rq); | |
bf0f6f24 IM |
1416 | } |
1417 | ||
1418 | /* | |
1419 | * Preempt the current task with a newly woken task if needed: | |
1420 | */ | |
7c92e54f | 1421 | static void |
2e09bf55 | 1422 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1423 | { |
11697830 | 1424 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
1425 | struct sched_entity *se; |
1426 | s64 delta; | |
11697830 | 1427 | |
6d0f0ebd | 1428 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1429 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1430 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1431 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1432 | /* |
1433 | * The current task ran long enough, ensure it doesn't get | |
1434 | * re-elected due to buddy favours. | |
1435 | */ | |
1436 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1437 | return; |
1438 | } | |
1439 | ||
1440 | /* | |
1441 | * Ensure that a task that missed wakeup preemption by a | |
1442 | * narrow margin doesn't have to wait for a full slice. | |
1443 | * This also mitigates buddy induced latencies under load. | |
1444 | */ | |
f685ceac MG |
1445 | if (delta_exec < sysctl_sched_min_granularity) |
1446 | return; | |
1447 | ||
f4cfb33e WX |
1448 | se = __pick_first_entity(cfs_rq); |
1449 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 1450 | |
f4cfb33e WX |
1451 | if (delta < 0) |
1452 | return; | |
d7d82944 | 1453 | |
f4cfb33e WX |
1454 | if (delta > ideal_runtime) |
1455 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
1456 | } |
1457 | ||
83b699ed | 1458 | static void |
8494f412 | 1459 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1460 | { |
83b699ed SV |
1461 | /* 'current' is not kept within the tree. */ |
1462 | if (se->on_rq) { | |
1463 | /* | |
1464 | * Any task has to be enqueued before it get to execute on | |
1465 | * a CPU. So account for the time it spent waiting on the | |
1466 | * runqueue. | |
1467 | */ | |
1468 | update_stats_wait_end(cfs_rq, se); | |
1469 | __dequeue_entity(cfs_rq, se); | |
1470 | } | |
1471 | ||
79303e9e | 1472 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1473 | cfs_rq->curr = se; |
eba1ed4b IM |
1474 | #ifdef CONFIG_SCHEDSTATS |
1475 | /* | |
1476 | * Track our maximum slice length, if the CPU's load is at | |
1477 | * least twice that of our own weight (i.e. dont track it | |
1478 | * when there are only lesser-weight tasks around): | |
1479 | */ | |
495eca49 | 1480 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1481 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1482 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1483 | } | |
1484 | #endif | |
4a55b450 | 1485 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1486 | } |
1487 | ||
3f3a4904 PZ |
1488 | static int |
1489 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1490 | ||
ac53db59 RR |
1491 | /* |
1492 | * Pick the next process, keeping these things in mind, in this order: | |
1493 | * 1) keep things fair between processes/task groups | |
1494 | * 2) pick the "next" process, since someone really wants that to run | |
1495 | * 3) pick the "last" process, for cache locality | |
1496 | * 4) do not run the "skip" process, if something else is available | |
1497 | */ | |
f4b6755f | 1498 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1499 | { |
ac53db59 | 1500 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1501 | struct sched_entity *left = se; |
f4b6755f | 1502 | |
ac53db59 RR |
1503 | /* |
1504 | * Avoid running the skip buddy, if running something else can | |
1505 | * be done without getting too unfair. | |
1506 | */ | |
1507 | if (cfs_rq->skip == se) { | |
1508 | struct sched_entity *second = __pick_next_entity(se); | |
1509 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1510 | se = second; | |
1511 | } | |
aa2ac252 | 1512 | |
f685ceac MG |
1513 | /* |
1514 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1515 | */ | |
1516 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1517 | se = cfs_rq->last; | |
1518 | ||
ac53db59 RR |
1519 | /* |
1520 | * Someone really wants this to run. If it's not unfair, run it. | |
1521 | */ | |
1522 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1523 | se = cfs_rq->next; | |
1524 | ||
f685ceac | 1525 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1526 | |
1527 | return se; | |
aa2ac252 PZ |
1528 | } |
1529 | ||
d3d9dc33 PT |
1530 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
1531 | ||
ab6cde26 | 1532 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1533 | { |
1534 | /* | |
1535 | * If still on the runqueue then deactivate_task() | |
1536 | * was not called and update_curr() has to be done: | |
1537 | */ | |
1538 | if (prev->on_rq) | |
b7cc0896 | 1539 | update_curr(cfs_rq); |
bf0f6f24 | 1540 | |
d3d9dc33 PT |
1541 | /* throttle cfs_rqs exceeding runtime */ |
1542 | check_cfs_rq_runtime(cfs_rq); | |
1543 | ||
ddc97297 | 1544 | check_spread(cfs_rq, prev); |
30cfdcfc | 1545 | if (prev->on_rq) { |
5870db5b | 1546 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1547 | /* Put 'current' back into the tree. */ |
1548 | __enqueue_entity(cfs_rq, prev); | |
1549 | } | |
429d43bc | 1550 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1551 | } |
1552 | ||
8f4d37ec PZ |
1553 | static void |
1554 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1555 | { |
bf0f6f24 | 1556 | /* |
30cfdcfc | 1557 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1558 | */ |
30cfdcfc | 1559 | update_curr(cfs_rq); |
bf0f6f24 | 1560 | |
43365bd7 PT |
1561 | /* |
1562 | * Update share accounting for long-running entities. | |
1563 | */ | |
1564 | update_entity_shares_tick(cfs_rq); | |
1565 | ||
8f4d37ec PZ |
1566 | #ifdef CONFIG_SCHED_HRTICK |
1567 | /* | |
1568 | * queued ticks are scheduled to match the slice, so don't bother | |
1569 | * validating it and just reschedule. | |
1570 | */ | |
983ed7a6 HH |
1571 | if (queued) { |
1572 | resched_task(rq_of(cfs_rq)->curr); | |
1573 | return; | |
1574 | } | |
8f4d37ec PZ |
1575 | /* |
1576 | * don't let the period tick interfere with the hrtick preemption | |
1577 | */ | |
1578 | if (!sched_feat(DOUBLE_TICK) && | |
1579 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1580 | return; | |
1581 | #endif | |
1582 | ||
2c2efaed | 1583 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1584 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1585 | } |
1586 | ||
ab84d31e PT |
1587 | |
1588 | /************************************************** | |
1589 | * CFS bandwidth control machinery | |
1590 | */ | |
1591 | ||
1592 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
1593 | |
1594 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 1595 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
1596 | |
1597 | static inline bool cfs_bandwidth_used(void) | |
1598 | { | |
c5905afb | 1599 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
1600 | } |
1601 | ||
1602 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
1603 | { | |
1604 | /* only need to count groups transitioning between enabled/!enabled */ | |
1605 | if (enabled && !was_enabled) | |
c5905afb | 1606 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 1607 | else if (!enabled && was_enabled) |
c5905afb | 1608 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
1609 | } |
1610 | #else /* HAVE_JUMP_LABEL */ | |
1611 | static bool cfs_bandwidth_used(void) | |
1612 | { | |
1613 | return true; | |
1614 | } | |
1615 | ||
1616 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
1617 | #endif /* HAVE_JUMP_LABEL */ | |
1618 | ||
ab84d31e PT |
1619 | /* |
1620 | * default period for cfs group bandwidth. | |
1621 | * default: 0.1s, units: nanoseconds | |
1622 | */ | |
1623 | static inline u64 default_cfs_period(void) | |
1624 | { | |
1625 | return 100000000ULL; | |
1626 | } | |
ec12cb7f PT |
1627 | |
1628 | static inline u64 sched_cfs_bandwidth_slice(void) | |
1629 | { | |
1630 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
1631 | } | |
1632 | ||
a9cf55b2 PT |
1633 | /* |
1634 | * Replenish runtime according to assigned quota and update expiration time. | |
1635 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
1636 | * additional synchronization around rq->lock. | |
1637 | * | |
1638 | * requires cfs_b->lock | |
1639 | */ | |
029632fb | 1640 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
1641 | { |
1642 | u64 now; | |
1643 | ||
1644 | if (cfs_b->quota == RUNTIME_INF) | |
1645 | return; | |
1646 | ||
1647 | now = sched_clock_cpu(smp_processor_id()); | |
1648 | cfs_b->runtime = cfs_b->quota; | |
1649 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
1650 | } | |
1651 | ||
029632fb PZ |
1652 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
1653 | { | |
1654 | return &tg->cfs_bandwidth; | |
1655 | } | |
1656 | ||
85dac906 PT |
1657 | /* returns 0 on failure to allocate runtime */ |
1658 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
1659 | { |
1660 | struct task_group *tg = cfs_rq->tg; | |
1661 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 1662 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
1663 | |
1664 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
1665 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
1666 | ||
1667 | raw_spin_lock(&cfs_b->lock); | |
1668 | if (cfs_b->quota == RUNTIME_INF) | |
1669 | amount = min_amount; | |
58088ad0 | 1670 | else { |
a9cf55b2 PT |
1671 | /* |
1672 | * If the bandwidth pool has become inactive, then at least one | |
1673 | * period must have elapsed since the last consumption. | |
1674 | * Refresh the global state and ensure bandwidth timer becomes | |
1675 | * active. | |
1676 | */ | |
1677 | if (!cfs_b->timer_active) { | |
1678 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 1679 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 1680 | } |
58088ad0 PT |
1681 | |
1682 | if (cfs_b->runtime > 0) { | |
1683 | amount = min(cfs_b->runtime, min_amount); | |
1684 | cfs_b->runtime -= amount; | |
1685 | cfs_b->idle = 0; | |
1686 | } | |
ec12cb7f | 1687 | } |
a9cf55b2 | 1688 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
1689 | raw_spin_unlock(&cfs_b->lock); |
1690 | ||
1691 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
1692 | /* |
1693 | * we may have advanced our local expiration to account for allowed | |
1694 | * spread between our sched_clock and the one on which runtime was | |
1695 | * issued. | |
1696 | */ | |
1697 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
1698 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
1699 | |
1700 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
1701 | } |
1702 | ||
a9cf55b2 PT |
1703 | /* |
1704 | * Note: This depends on the synchronization provided by sched_clock and the | |
1705 | * fact that rq->clock snapshots this value. | |
1706 | */ | |
1707 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 1708 | { |
a9cf55b2 PT |
1709 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
1710 | struct rq *rq = rq_of(cfs_rq); | |
1711 | ||
1712 | /* if the deadline is ahead of our clock, nothing to do */ | |
1713 | if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0)) | |
ec12cb7f PT |
1714 | return; |
1715 | ||
a9cf55b2 PT |
1716 | if (cfs_rq->runtime_remaining < 0) |
1717 | return; | |
1718 | ||
1719 | /* | |
1720 | * If the local deadline has passed we have to consider the | |
1721 | * possibility that our sched_clock is 'fast' and the global deadline | |
1722 | * has not truly expired. | |
1723 | * | |
1724 | * Fortunately we can check determine whether this the case by checking | |
1725 | * whether the global deadline has advanced. | |
1726 | */ | |
1727 | ||
1728 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
1729 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
1730 | cfs_rq->runtime_expires += TICK_NSEC; | |
1731 | } else { | |
1732 | /* global deadline is ahead, expiration has passed */ | |
1733 | cfs_rq->runtime_remaining = 0; | |
1734 | } | |
1735 | } | |
1736 | ||
1737 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1738 | unsigned long delta_exec) | |
1739 | { | |
1740 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 1741 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
1742 | expire_cfs_rq_runtime(cfs_rq); |
1743 | ||
1744 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
1745 | return; |
1746 | ||
85dac906 PT |
1747 | /* |
1748 | * if we're unable to extend our runtime we resched so that the active | |
1749 | * hierarchy can be throttled | |
1750 | */ | |
1751 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
1752 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
1753 | } |
1754 | ||
6c16a6dc PZ |
1755 | static __always_inline |
1756 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 1757 | { |
56f570e5 | 1758 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
1759 | return; |
1760 | ||
1761 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
1762 | } | |
1763 | ||
85dac906 PT |
1764 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
1765 | { | |
56f570e5 | 1766 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
1767 | } |
1768 | ||
64660c86 PT |
1769 | /* check whether cfs_rq, or any parent, is throttled */ |
1770 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
1771 | { | |
56f570e5 | 1772 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
1773 | } |
1774 | ||
1775 | /* | |
1776 | * Ensure that neither of the group entities corresponding to src_cpu or | |
1777 | * dest_cpu are members of a throttled hierarchy when performing group | |
1778 | * load-balance operations. | |
1779 | */ | |
1780 | static inline int throttled_lb_pair(struct task_group *tg, | |
1781 | int src_cpu, int dest_cpu) | |
1782 | { | |
1783 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
1784 | ||
1785 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
1786 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
1787 | ||
1788 | return throttled_hierarchy(src_cfs_rq) || | |
1789 | throttled_hierarchy(dest_cfs_rq); | |
1790 | } | |
1791 | ||
1792 | /* updated child weight may affect parent so we have to do this bottom up */ | |
1793 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
1794 | { | |
1795 | struct rq *rq = data; | |
1796 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1797 | ||
1798 | cfs_rq->throttle_count--; | |
1799 | #ifdef CONFIG_SMP | |
1800 | if (!cfs_rq->throttle_count) { | |
1801 | u64 delta = rq->clock_task - cfs_rq->load_stamp; | |
1802 | ||
1803 | /* leaving throttled state, advance shares averaging windows */ | |
1804 | cfs_rq->load_stamp += delta; | |
1805 | cfs_rq->load_last += delta; | |
1806 | ||
1807 | /* update entity weight now that we are on_rq again */ | |
1808 | update_cfs_shares(cfs_rq); | |
1809 | } | |
1810 | #endif | |
1811 | ||
1812 | return 0; | |
1813 | } | |
1814 | ||
1815 | static int tg_throttle_down(struct task_group *tg, void *data) | |
1816 | { | |
1817 | struct rq *rq = data; | |
1818 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1819 | ||
1820 | /* group is entering throttled state, record last load */ | |
1821 | if (!cfs_rq->throttle_count) | |
1822 | update_cfs_load(cfs_rq, 0); | |
1823 | cfs_rq->throttle_count++; | |
1824 | ||
1825 | return 0; | |
1826 | } | |
1827 | ||
d3d9dc33 | 1828 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
1829 | { |
1830 | struct rq *rq = rq_of(cfs_rq); | |
1831 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1832 | struct sched_entity *se; | |
1833 | long task_delta, dequeue = 1; | |
1834 | ||
1835 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1836 | ||
1837 | /* account load preceding throttle */ | |
64660c86 PT |
1838 | rcu_read_lock(); |
1839 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
1840 | rcu_read_unlock(); | |
85dac906 PT |
1841 | |
1842 | task_delta = cfs_rq->h_nr_running; | |
1843 | for_each_sched_entity(se) { | |
1844 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
1845 | /* throttled entity or throttle-on-deactivate */ | |
1846 | if (!se->on_rq) | |
1847 | break; | |
1848 | ||
1849 | if (dequeue) | |
1850 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
1851 | qcfs_rq->h_nr_running -= task_delta; | |
1852 | ||
1853 | if (qcfs_rq->load.weight) | |
1854 | dequeue = 0; | |
1855 | } | |
1856 | ||
1857 | if (!se) | |
1858 | rq->nr_running -= task_delta; | |
1859 | ||
1860 | cfs_rq->throttled = 1; | |
e8da1b18 | 1861 | cfs_rq->throttled_timestamp = rq->clock; |
85dac906 PT |
1862 | raw_spin_lock(&cfs_b->lock); |
1863 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
1864 | raw_spin_unlock(&cfs_b->lock); | |
1865 | } | |
1866 | ||
029632fb | 1867 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
1868 | { |
1869 | struct rq *rq = rq_of(cfs_rq); | |
1870 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1871 | struct sched_entity *se; | |
1872 | int enqueue = 1; | |
1873 | long task_delta; | |
1874 | ||
1875 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1876 | ||
1877 | cfs_rq->throttled = 0; | |
1878 | raw_spin_lock(&cfs_b->lock); | |
e8da1b18 | 1879 | cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp; |
671fd9da PT |
1880 | list_del_rcu(&cfs_rq->throttled_list); |
1881 | raw_spin_unlock(&cfs_b->lock); | |
e8da1b18 | 1882 | cfs_rq->throttled_timestamp = 0; |
671fd9da | 1883 | |
64660c86 PT |
1884 | update_rq_clock(rq); |
1885 | /* update hierarchical throttle state */ | |
1886 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
1887 | ||
671fd9da PT |
1888 | if (!cfs_rq->load.weight) |
1889 | return; | |
1890 | ||
1891 | task_delta = cfs_rq->h_nr_running; | |
1892 | for_each_sched_entity(se) { | |
1893 | if (se->on_rq) | |
1894 | enqueue = 0; | |
1895 | ||
1896 | cfs_rq = cfs_rq_of(se); | |
1897 | if (enqueue) | |
1898 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
1899 | cfs_rq->h_nr_running += task_delta; | |
1900 | ||
1901 | if (cfs_rq_throttled(cfs_rq)) | |
1902 | break; | |
1903 | } | |
1904 | ||
1905 | if (!se) | |
1906 | rq->nr_running += task_delta; | |
1907 | ||
1908 | /* determine whether we need to wake up potentially idle cpu */ | |
1909 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
1910 | resched_task(rq->curr); | |
1911 | } | |
1912 | ||
1913 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
1914 | u64 remaining, u64 expires) | |
1915 | { | |
1916 | struct cfs_rq *cfs_rq; | |
1917 | u64 runtime = remaining; | |
1918 | ||
1919 | rcu_read_lock(); | |
1920 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
1921 | throttled_list) { | |
1922 | struct rq *rq = rq_of(cfs_rq); | |
1923 | ||
1924 | raw_spin_lock(&rq->lock); | |
1925 | if (!cfs_rq_throttled(cfs_rq)) | |
1926 | goto next; | |
1927 | ||
1928 | runtime = -cfs_rq->runtime_remaining + 1; | |
1929 | if (runtime > remaining) | |
1930 | runtime = remaining; | |
1931 | remaining -= runtime; | |
1932 | ||
1933 | cfs_rq->runtime_remaining += runtime; | |
1934 | cfs_rq->runtime_expires = expires; | |
1935 | ||
1936 | /* we check whether we're throttled above */ | |
1937 | if (cfs_rq->runtime_remaining > 0) | |
1938 | unthrottle_cfs_rq(cfs_rq); | |
1939 | ||
1940 | next: | |
1941 | raw_spin_unlock(&rq->lock); | |
1942 | ||
1943 | if (!remaining) | |
1944 | break; | |
1945 | } | |
1946 | rcu_read_unlock(); | |
1947 | ||
1948 | return remaining; | |
1949 | } | |
1950 | ||
58088ad0 PT |
1951 | /* |
1952 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
1953 | * cfs_rqs as appropriate. If there has been no activity within the last | |
1954 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
1955 | * used to track this state. | |
1956 | */ | |
1957 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
1958 | { | |
671fd9da PT |
1959 | u64 runtime, runtime_expires; |
1960 | int idle = 1, throttled; | |
58088ad0 PT |
1961 | |
1962 | raw_spin_lock(&cfs_b->lock); | |
1963 | /* no need to continue the timer with no bandwidth constraint */ | |
1964 | if (cfs_b->quota == RUNTIME_INF) | |
1965 | goto out_unlock; | |
1966 | ||
671fd9da PT |
1967 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
1968 | /* idle depends on !throttled (for the case of a large deficit) */ | |
1969 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 1970 | cfs_b->nr_periods += overrun; |
671fd9da | 1971 | |
a9cf55b2 PT |
1972 | /* if we're going inactive then everything else can be deferred */ |
1973 | if (idle) | |
1974 | goto out_unlock; | |
1975 | ||
1976 | __refill_cfs_bandwidth_runtime(cfs_b); | |
1977 | ||
671fd9da PT |
1978 | if (!throttled) { |
1979 | /* mark as potentially idle for the upcoming period */ | |
1980 | cfs_b->idle = 1; | |
1981 | goto out_unlock; | |
1982 | } | |
1983 | ||
e8da1b18 NR |
1984 | /* account preceding periods in which throttling occurred */ |
1985 | cfs_b->nr_throttled += overrun; | |
1986 | ||
671fd9da PT |
1987 | /* |
1988 | * There are throttled entities so we must first use the new bandwidth | |
1989 | * to unthrottle them before making it generally available. This | |
1990 | * ensures that all existing debts will be paid before a new cfs_rq is | |
1991 | * allowed to run. | |
1992 | */ | |
1993 | runtime = cfs_b->runtime; | |
1994 | runtime_expires = cfs_b->runtime_expires; | |
1995 | cfs_b->runtime = 0; | |
1996 | ||
1997 | /* | |
1998 | * This check is repeated as we are holding onto the new bandwidth | |
1999 | * while we unthrottle. This can potentially race with an unthrottled | |
2000 | * group trying to acquire new bandwidth from the global pool. | |
2001 | */ | |
2002 | while (throttled && runtime > 0) { | |
2003 | raw_spin_unlock(&cfs_b->lock); | |
2004 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
2005 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
2006 | runtime_expires); | |
2007 | raw_spin_lock(&cfs_b->lock); | |
2008 | ||
2009 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
2010 | } | |
58088ad0 | 2011 | |
671fd9da PT |
2012 | /* return (any) remaining runtime */ |
2013 | cfs_b->runtime = runtime; | |
2014 | /* | |
2015 | * While we are ensured activity in the period following an | |
2016 | * unthrottle, this also covers the case in which the new bandwidth is | |
2017 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
2018 | * timer to remain active while there are any throttled entities.) | |
2019 | */ | |
2020 | cfs_b->idle = 0; | |
58088ad0 PT |
2021 | out_unlock: |
2022 | if (idle) | |
2023 | cfs_b->timer_active = 0; | |
2024 | raw_spin_unlock(&cfs_b->lock); | |
2025 | ||
2026 | return idle; | |
2027 | } | |
d3d9dc33 | 2028 | |
d8b4986d PT |
2029 | /* a cfs_rq won't donate quota below this amount */ |
2030 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
2031 | /* minimum remaining period time to redistribute slack quota */ | |
2032 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
2033 | /* how long we wait to gather additional slack before distributing */ | |
2034 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
2035 | ||
2036 | /* are we near the end of the current quota period? */ | |
2037 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
2038 | { | |
2039 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
2040 | u64 remaining; | |
2041 | ||
2042 | /* if the call-back is running a quota refresh is already occurring */ | |
2043 | if (hrtimer_callback_running(refresh_timer)) | |
2044 | return 1; | |
2045 | ||
2046 | /* is a quota refresh about to occur? */ | |
2047 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
2048 | if (remaining < min_expire) | |
2049 | return 1; | |
2050 | ||
2051 | return 0; | |
2052 | } | |
2053 | ||
2054 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
2055 | { | |
2056 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
2057 | ||
2058 | /* if there's a quota refresh soon don't bother with slack */ | |
2059 | if (runtime_refresh_within(cfs_b, min_left)) | |
2060 | return; | |
2061 | ||
2062 | start_bandwidth_timer(&cfs_b->slack_timer, | |
2063 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
2064 | } | |
2065 | ||
2066 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
2067 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2068 | { | |
2069 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2070 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
2071 | ||
2072 | if (slack_runtime <= 0) | |
2073 | return; | |
2074 | ||
2075 | raw_spin_lock(&cfs_b->lock); | |
2076 | if (cfs_b->quota != RUNTIME_INF && | |
2077 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
2078 | cfs_b->runtime += slack_runtime; | |
2079 | ||
2080 | /* we are under rq->lock, defer unthrottling using a timer */ | |
2081 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
2082 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
2083 | start_cfs_slack_bandwidth(cfs_b); | |
2084 | } | |
2085 | raw_spin_unlock(&cfs_b->lock); | |
2086 | ||
2087 | /* even if it's not valid for return we don't want to try again */ | |
2088 | cfs_rq->runtime_remaining -= slack_runtime; | |
2089 | } | |
2090 | ||
2091 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2092 | { | |
56f570e5 PT |
2093 | if (!cfs_bandwidth_used()) |
2094 | return; | |
2095 | ||
fccfdc6f | 2096 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2097 | return; |
2098 | ||
2099 | __return_cfs_rq_runtime(cfs_rq); | |
2100 | } | |
2101 | ||
2102 | /* | |
2103 | * This is done with a timer (instead of inline with bandwidth return) since | |
2104 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2105 | */ | |
2106 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2107 | { | |
2108 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2109 | u64 expires; | |
2110 | ||
2111 | /* confirm we're still not at a refresh boundary */ | |
2112 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2113 | return; | |
2114 | ||
2115 | raw_spin_lock(&cfs_b->lock); | |
2116 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2117 | runtime = cfs_b->runtime; | |
2118 | cfs_b->runtime = 0; | |
2119 | } | |
2120 | expires = cfs_b->runtime_expires; | |
2121 | raw_spin_unlock(&cfs_b->lock); | |
2122 | ||
2123 | if (!runtime) | |
2124 | return; | |
2125 | ||
2126 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2127 | ||
2128 | raw_spin_lock(&cfs_b->lock); | |
2129 | if (expires == cfs_b->runtime_expires) | |
2130 | cfs_b->runtime = runtime; | |
2131 | raw_spin_unlock(&cfs_b->lock); | |
2132 | } | |
2133 | ||
d3d9dc33 PT |
2134 | /* |
2135 | * When a group wakes up we want to make sure that its quota is not already | |
2136 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2137 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2138 | */ | |
2139 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2140 | { | |
56f570e5 PT |
2141 | if (!cfs_bandwidth_used()) |
2142 | return; | |
2143 | ||
d3d9dc33 PT |
2144 | /* an active group must be handled by the update_curr()->put() path */ |
2145 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2146 | return; | |
2147 | ||
2148 | /* ensure the group is not already throttled */ | |
2149 | if (cfs_rq_throttled(cfs_rq)) | |
2150 | return; | |
2151 | ||
2152 | /* update runtime allocation */ | |
2153 | account_cfs_rq_runtime(cfs_rq, 0); | |
2154 | if (cfs_rq->runtime_remaining <= 0) | |
2155 | throttle_cfs_rq(cfs_rq); | |
2156 | } | |
2157 | ||
2158 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2159 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2160 | { | |
56f570e5 PT |
2161 | if (!cfs_bandwidth_used()) |
2162 | return; | |
2163 | ||
d3d9dc33 PT |
2164 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2165 | return; | |
2166 | ||
2167 | /* | |
2168 | * it's possible for a throttled entity to be forced into a running | |
2169 | * state (e.g. set_curr_task), in this case we're finished. | |
2170 | */ | |
2171 | if (cfs_rq_throttled(cfs_rq)) | |
2172 | return; | |
2173 | ||
2174 | throttle_cfs_rq(cfs_rq); | |
2175 | } | |
029632fb PZ |
2176 | |
2177 | static inline u64 default_cfs_period(void); | |
2178 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun); | |
2179 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b); | |
2180 | ||
2181 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) | |
2182 | { | |
2183 | struct cfs_bandwidth *cfs_b = | |
2184 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2185 | do_sched_cfs_slack_timer(cfs_b); | |
2186 | ||
2187 | return HRTIMER_NORESTART; | |
2188 | } | |
2189 | ||
2190 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2191 | { | |
2192 | struct cfs_bandwidth *cfs_b = | |
2193 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2194 | ktime_t now; | |
2195 | int overrun; | |
2196 | int idle = 0; | |
2197 | ||
2198 | for (;;) { | |
2199 | now = hrtimer_cb_get_time(timer); | |
2200 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2201 | ||
2202 | if (!overrun) | |
2203 | break; | |
2204 | ||
2205 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2206 | } | |
2207 | ||
2208 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2209 | } | |
2210 | ||
2211 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2212 | { | |
2213 | raw_spin_lock_init(&cfs_b->lock); | |
2214 | cfs_b->runtime = 0; | |
2215 | cfs_b->quota = RUNTIME_INF; | |
2216 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2217 | ||
2218 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2219 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2220 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2221 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2222 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2223 | } | |
2224 | ||
2225 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2226 | { | |
2227 | cfs_rq->runtime_enabled = 0; | |
2228 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2229 | } | |
2230 | ||
2231 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2232 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2233 | { | |
2234 | /* | |
2235 | * The timer may be active because we're trying to set a new bandwidth | |
2236 | * period or because we're racing with the tear-down path | |
2237 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2238 | * terminates). In either case we ensure that it's re-programmed | |
2239 | */ | |
2240 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2241 | raw_spin_unlock(&cfs_b->lock); | |
2242 | /* ensure cfs_b->lock is available while we wait */ | |
2243 | hrtimer_cancel(&cfs_b->period_timer); | |
2244 | ||
2245 | raw_spin_lock(&cfs_b->lock); | |
2246 | /* if someone else restarted the timer then we're done */ | |
2247 | if (cfs_b->timer_active) | |
2248 | return; | |
2249 | } | |
2250 | ||
2251 | cfs_b->timer_active = 1; | |
2252 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
2253 | } | |
2254 | ||
2255 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2256 | { | |
2257 | hrtimer_cancel(&cfs_b->period_timer); | |
2258 | hrtimer_cancel(&cfs_b->slack_timer); | |
2259 | } | |
2260 | ||
a4c96ae3 | 2261 | static void unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
2262 | { |
2263 | struct cfs_rq *cfs_rq; | |
2264 | ||
2265 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2266 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2267 | ||
2268 | if (!cfs_rq->runtime_enabled) | |
2269 | continue; | |
2270 | ||
2271 | /* | |
2272 | * clock_task is not advancing so we just need to make sure | |
2273 | * there's some valid quota amount | |
2274 | */ | |
2275 | cfs_rq->runtime_remaining = cfs_b->quota; | |
2276 | if (cfs_rq_throttled(cfs_rq)) | |
2277 | unthrottle_cfs_rq(cfs_rq); | |
2278 | } | |
2279 | } | |
2280 | ||
2281 | #else /* CONFIG_CFS_BANDWIDTH */ | |
6c16a6dc PZ |
2282 | static __always_inline |
2283 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {} | |
d3d9dc33 PT |
2284 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
2285 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 2286 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
2287 | |
2288 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
2289 | { | |
2290 | return 0; | |
2291 | } | |
64660c86 PT |
2292 | |
2293 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2294 | { | |
2295 | return 0; | |
2296 | } | |
2297 | ||
2298 | static inline int throttled_lb_pair(struct task_group *tg, | |
2299 | int src_cpu, int dest_cpu) | |
2300 | { | |
2301 | return 0; | |
2302 | } | |
029632fb PZ |
2303 | |
2304 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
2305 | ||
2306 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2307 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
2308 | #endif |
2309 | ||
029632fb PZ |
2310 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2311 | { | |
2312 | return NULL; | |
2313 | } | |
2314 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 2315 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
2316 | |
2317 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
2318 | ||
bf0f6f24 IM |
2319 | /************************************************** |
2320 | * CFS operations on tasks: | |
2321 | */ | |
2322 | ||
8f4d37ec PZ |
2323 | #ifdef CONFIG_SCHED_HRTICK |
2324 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2325 | { | |
8f4d37ec PZ |
2326 | struct sched_entity *se = &p->se; |
2327 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2328 | ||
2329 | WARN_ON(task_rq(p) != rq); | |
2330 | ||
b39e66ea | 2331 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
2332 | u64 slice = sched_slice(cfs_rq, se); |
2333 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
2334 | s64 delta = slice - ran; | |
2335 | ||
2336 | if (delta < 0) { | |
2337 | if (rq->curr == p) | |
2338 | resched_task(p); | |
2339 | return; | |
2340 | } | |
2341 | ||
2342 | /* | |
2343 | * Don't schedule slices shorter than 10000ns, that just | |
2344 | * doesn't make sense. Rely on vruntime for fairness. | |
2345 | */ | |
31656519 | 2346 | if (rq->curr != p) |
157124c1 | 2347 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 2348 | |
31656519 | 2349 | hrtick_start(rq, delta); |
8f4d37ec PZ |
2350 | } |
2351 | } | |
a4c2f00f PZ |
2352 | |
2353 | /* | |
2354 | * called from enqueue/dequeue and updates the hrtick when the | |
2355 | * current task is from our class and nr_running is low enough | |
2356 | * to matter. | |
2357 | */ | |
2358 | static void hrtick_update(struct rq *rq) | |
2359 | { | |
2360 | struct task_struct *curr = rq->curr; | |
2361 | ||
b39e66ea | 2362 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
2363 | return; |
2364 | ||
2365 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
2366 | hrtick_start_fair(rq, curr); | |
2367 | } | |
55e12e5e | 2368 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
2369 | static inline void |
2370 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2371 | { | |
2372 | } | |
a4c2f00f PZ |
2373 | |
2374 | static inline void hrtick_update(struct rq *rq) | |
2375 | { | |
2376 | } | |
8f4d37ec PZ |
2377 | #endif |
2378 | ||
bf0f6f24 IM |
2379 | /* |
2380 | * The enqueue_task method is called before nr_running is | |
2381 | * increased. Here we update the fair scheduling stats and | |
2382 | * then put the task into the rbtree: | |
2383 | */ | |
ea87bb78 | 2384 | static void |
371fd7e7 | 2385 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2386 | { |
2387 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2388 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
2389 | |
2390 | for_each_sched_entity(se) { | |
62fb1851 | 2391 | if (se->on_rq) |
bf0f6f24 IM |
2392 | break; |
2393 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 2394 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
2395 | |
2396 | /* | |
2397 | * end evaluation on encountering a throttled cfs_rq | |
2398 | * | |
2399 | * note: in the case of encountering a throttled cfs_rq we will | |
2400 | * post the final h_nr_running increment below. | |
2401 | */ | |
2402 | if (cfs_rq_throttled(cfs_rq)) | |
2403 | break; | |
953bfcd1 | 2404 | cfs_rq->h_nr_running++; |
85dac906 | 2405 | |
88ec22d3 | 2406 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 2407 | } |
8f4d37ec | 2408 | |
2069dd75 | 2409 | for_each_sched_entity(se) { |
0f317143 | 2410 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2411 | cfs_rq->h_nr_running++; |
2069dd75 | 2412 | |
85dac906 PT |
2413 | if (cfs_rq_throttled(cfs_rq)) |
2414 | break; | |
2415 | ||
d6b55918 | 2416 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 2417 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
2418 | } |
2419 | ||
85dac906 PT |
2420 | if (!se) |
2421 | inc_nr_running(rq); | |
a4c2f00f | 2422 | hrtick_update(rq); |
bf0f6f24 IM |
2423 | } |
2424 | ||
2f36825b VP |
2425 | static void set_next_buddy(struct sched_entity *se); |
2426 | ||
bf0f6f24 IM |
2427 | /* |
2428 | * The dequeue_task method is called before nr_running is | |
2429 | * decreased. We remove the task from the rbtree and | |
2430 | * update the fair scheduling stats: | |
2431 | */ | |
371fd7e7 | 2432 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2433 | { |
2434 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2435 | struct sched_entity *se = &p->se; |
2f36825b | 2436 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
2437 | |
2438 | for_each_sched_entity(se) { | |
2439 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 2440 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
2441 | |
2442 | /* | |
2443 | * end evaluation on encountering a throttled cfs_rq | |
2444 | * | |
2445 | * note: in the case of encountering a throttled cfs_rq we will | |
2446 | * post the final h_nr_running decrement below. | |
2447 | */ | |
2448 | if (cfs_rq_throttled(cfs_rq)) | |
2449 | break; | |
953bfcd1 | 2450 | cfs_rq->h_nr_running--; |
2069dd75 | 2451 | |
bf0f6f24 | 2452 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
2453 | if (cfs_rq->load.weight) { |
2454 | /* | |
2455 | * Bias pick_next to pick a task from this cfs_rq, as | |
2456 | * p is sleeping when it is within its sched_slice. | |
2457 | */ | |
2458 | if (task_sleep && parent_entity(se)) | |
2459 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
2460 | |
2461 | /* avoid re-evaluating load for this entity */ | |
2462 | se = parent_entity(se); | |
bf0f6f24 | 2463 | break; |
2f36825b | 2464 | } |
371fd7e7 | 2465 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 2466 | } |
8f4d37ec | 2467 | |
2069dd75 | 2468 | for_each_sched_entity(se) { |
0f317143 | 2469 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2470 | cfs_rq->h_nr_running--; |
2069dd75 | 2471 | |
85dac906 PT |
2472 | if (cfs_rq_throttled(cfs_rq)) |
2473 | break; | |
2474 | ||
d6b55918 | 2475 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 2476 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
2477 | } |
2478 | ||
85dac906 PT |
2479 | if (!se) |
2480 | dec_nr_running(rq); | |
a4c2f00f | 2481 | hrtick_update(rq); |
bf0f6f24 IM |
2482 | } |
2483 | ||
e7693a36 | 2484 | #ifdef CONFIG_SMP |
029632fb PZ |
2485 | /* Used instead of source_load when we know the type == 0 */ |
2486 | static unsigned long weighted_cpuload(const int cpu) | |
2487 | { | |
2488 | return cpu_rq(cpu)->load.weight; | |
2489 | } | |
2490 | ||
2491 | /* | |
2492 | * Return a low guess at the load of a migration-source cpu weighted | |
2493 | * according to the scheduling class and "nice" value. | |
2494 | * | |
2495 | * We want to under-estimate the load of migration sources, to | |
2496 | * balance conservatively. | |
2497 | */ | |
2498 | static unsigned long source_load(int cpu, int type) | |
2499 | { | |
2500 | struct rq *rq = cpu_rq(cpu); | |
2501 | unsigned long total = weighted_cpuload(cpu); | |
2502 | ||
2503 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2504 | return total; | |
2505 | ||
2506 | return min(rq->cpu_load[type-1], total); | |
2507 | } | |
2508 | ||
2509 | /* | |
2510 | * Return a high guess at the load of a migration-target cpu weighted | |
2511 | * according to the scheduling class and "nice" value. | |
2512 | */ | |
2513 | static unsigned long target_load(int cpu, int type) | |
2514 | { | |
2515 | struct rq *rq = cpu_rq(cpu); | |
2516 | unsigned long total = weighted_cpuload(cpu); | |
2517 | ||
2518 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2519 | return total; | |
2520 | ||
2521 | return max(rq->cpu_load[type-1], total); | |
2522 | } | |
2523 | ||
2524 | static unsigned long power_of(int cpu) | |
2525 | { | |
2526 | return cpu_rq(cpu)->cpu_power; | |
2527 | } | |
2528 | ||
2529 | static unsigned long cpu_avg_load_per_task(int cpu) | |
2530 | { | |
2531 | struct rq *rq = cpu_rq(cpu); | |
2532 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
2533 | ||
2534 | if (nr_running) | |
2535 | return rq->load.weight / nr_running; | |
2536 | ||
2537 | return 0; | |
2538 | } | |
2539 | ||
098fb9db | 2540 | |
74f8e4b2 | 2541 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
2542 | { |
2543 | struct sched_entity *se = &p->se; | |
2544 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
2545 | u64 min_vruntime; |
2546 | ||
2547 | #ifndef CONFIG_64BIT | |
2548 | u64 min_vruntime_copy; | |
88ec22d3 | 2549 | |
3fe1698b PZ |
2550 | do { |
2551 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
2552 | smp_rmb(); | |
2553 | min_vruntime = cfs_rq->min_vruntime; | |
2554 | } while (min_vruntime != min_vruntime_copy); | |
2555 | #else | |
2556 | min_vruntime = cfs_rq->min_vruntime; | |
2557 | #endif | |
88ec22d3 | 2558 | |
3fe1698b | 2559 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
2560 | } |
2561 | ||
bb3469ac | 2562 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
2563 | /* |
2564 | * effective_load() calculates the load change as seen from the root_task_group | |
2565 | * | |
2566 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
2567 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
2568 | * can calculate the shift in shares. | |
cf5f0acf PZ |
2569 | * |
2570 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
2571 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
2572 | * total group weight. | |
2573 | * | |
2574 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
2575 | * distribution (s_i) using: | |
2576 | * | |
2577 | * s_i = rw_i / \Sum rw_j (1) | |
2578 | * | |
2579 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
2580 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
2581 | * shares distribution (s_i): | |
2582 | * | |
2583 | * rw_i = { 2, 4, 1, 0 } | |
2584 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
2585 | * | |
2586 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
2587 | * task used to run on and the CPU the waker is running on), we need to | |
2588 | * compute the effect of waking a task on either CPU and, in case of a sync | |
2589 | * wakeup, compute the effect of the current task going to sleep. | |
2590 | * | |
2591 | * So for a change of @wl to the local @cpu with an overall group weight change | |
2592 | * of @wl we can compute the new shares distribution (s'_i) using: | |
2593 | * | |
2594 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
2595 | * | |
2596 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
2597 | * differences in waking a task to CPU 0. The additional task changes the | |
2598 | * weight and shares distributions like: | |
2599 | * | |
2600 | * rw'_i = { 3, 4, 1, 0 } | |
2601 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
2602 | * | |
2603 | * We can then compute the difference in effective weight by using: | |
2604 | * | |
2605 | * dw_i = S * (s'_i - s_i) (3) | |
2606 | * | |
2607 | * Where 'S' is the group weight as seen by its parent. | |
2608 | * | |
2609 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
2610 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
2611 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 2612 | */ |
2069dd75 | 2613 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 2614 | { |
4be9daaa | 2615 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 2616 | |
cf5f0acf | 2617 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
2618 | return wl; |
2619 | ||
4be9daaa | 2620 | for_each_sched_entity(se) { |
cf5f0acf | 2621 | long w, W; |
4be9daaa | 2622 | |
977dda7c | 2623 | tg = se->my_q->tg; |
bb3469ac | 2624 | |
cf5f0acf PZ |
2625 | /* |
2626 | * W = @wg + \Sum rw_j | |
2627 | */ | |
2628 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 2629 | |
cf5f0acf PZ |
2630 | /* |
2631 | * w = rw_i + @wl | |
2632 | */ | |
2633 | w = se->my_q->load.weight + wl; | |
940959e9 | 2634 | |
cf5f0acf PZ |
2635 | /* |
2636 | * wl = S * s'_i; see (2) | |
2637 | */ | |
2638 | if (W > 0 && w < W) | |
2639 | wl = (w * tg->shares) / W; | |
977dda7c PT |
2640 | else |
2641 | wl = tg->shares; | |
940959e9 | 2642 | |
cf5f0acf PZ |
2643 | /* |
2644 | * Per the above, wl is the new se->load.weight value; since | |
2645 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
2646 | * calc_cfs_shares(). | |
2647 | */ | |
977dda7c PT |
2648 | if (wl < MIN_SHARES) |
2649 | wl = MIN_SHARES; | |
cf5f0acf PZ |
2650 | |
2651 | /* | |
2652 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
2653 | */ | |
977dda7c | 2654 | wl -= se->load.weight; |
cf5f0acf PZ |
2655 | |
2656 | /* | |
2657 | * Recursively apply this logic to all parent groups to compute | |
2658 | * the final effective load change on the root group. Since | |
2659 | * only the @tg group gets extra weight, all parent groups can | |
2660 | * only redistribute existing shares. @wl is the shift in shares | |
2661 | * resulting from this level per the above. | |
2662 | */ | |
4be9daaa | 2663 | wg = 0; |
4be9daaa | 2664 | } |
bb3469ac | 2665 | |
4be9daaa | 2666 | return wl; |
bb3469ac PZ |
2667 | } |
2668 | #else | |
4be9daaa | 2669 | |
83378269 PZ |
2670 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
2671 | unsigned long wl, unsigned long wg) | |
4be9daaa | 2672 | { |
83378269 | 2673 | return wl; |
bb3469ac | 2674 | } |
4be9daaa | 2675 | |
bb3469ac PZ |
2676 | #endif |
2677 | ||
c88d5910 | 2678 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 2679 | { |
e37b6a7b | 2680 | s64 this_load, load; |
c88d5910 | 2681 | int idx, this_cpu, prev_cpu; |
098fb9db | 2682 | unsigned long tl_per_task; |
c88d5910 | 2683 | struct task_group *tg; |
83378269 | 2684 | unsigned long weight; |
b3137bc8 | 2685 | int balanced; |
098fb9db | 2686 | |
c88d5910 PZ |
2687 | idx = sd->wake_idx; |
2688 | this_cpu = smp_processor_id(); | |
2689 | prev_cpu = task_cpu(p); | |
2690 | load = source_load(prev_cpu, idx); | |
2691 | this_load = target_load(this_cpu, idx); | |
098fb9db | 2692 | |
b3137bc8 MG |
2693 | /* |
2694 | * If sync wakeup then subtract the (maximum possible) | |
2695 | * effect of the currently running task from the load | |
2696 | * of the current CPU: | |
2697 | */ | |
83378269 PZ |
2698 | if (sync) { |
2699 | tg = task_group(current); | |
2700 | weight = current->se.load.weight; | |
2701 | ||
c88d5910 | 2702 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
2703 | load += effective_load(tg, prev_cpu, 0, -weight); |
2704 | } | |
b3137bc8 | 2705 | |
83378269 PZ |
2706 | tg = task_group(p); |
2707 | weight = p->se.load.weight; | |
b3137bc8 | 2708 | |
71a29aa7 PZ |
2709 | /* |
2710 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
2711 | * due to the sync cause above having dropped this_load to 0, we'll |
2712 | * always have an imbalance, but there's really nothing you can do | |
2713 | * about that, so that's good too. | |
71a29aa7 PZ |
2714 | * |
2715 | * Otherwise check if either cpus are near enough in load to allow this | |
2716 | * task to be woken on this_cpu. | |
2717 | */ | |
e37b6a7b PT |
2718 | if (this_load > 0) { |
2719 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
2720 | |
2721 | this_eff_load = 100; | |
2722 | this_eff_load *= power_of(prev_cpu); | |
2723 | this_eff_load *= this_load + | |
2724 | effective_load(tg, this_cpu, weight, weight); | |
2725 | ||
2726 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
2727 | prev_eff_load *= power_of(this_cpu); | |
2728 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
2729 | ||
2730 | balanced = this_eff_load <= prev_eff_load; | |
2731 | } else | |
2732 | balanced = true; | |
b3137bc8 | 2733 | |
098fb9db | 2734 | /* |
4ae7d5ce IM |
2735 | * If the currently running task will sleep within |
2736 | * a reasonable amount of time then attract this newly | |
2737 | * woken task: | |
098fb9db | 2738 | */ |
2fb7635c PZ |
2739 | if (sync && balanced) |
2740 | return 1; | |
098fb9db | 2741 | |
41acab88 | 2742 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
2743 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
2744 | ||
c88d5910 PZ |
2745 | if (balanced || |
2746 | (this_load <= load && | |
2747 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
2748 | /* |
2749 | * This domain has SD_WAKE_AFFINE and | |
2750 | * p is cache cold in this domain, and | |
2751 | * there is no bad imbalance. | |
2752 | */ | |
c88d5910 | 2753 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 2754 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
2755 | |
2756 | return 1; | |
2757 | } | |
2758 | return 0; | |
2759 | } | |
2760 | ||
aaee1203 PZ |
2761 | /* |
2762 | * find_idlest_group finds and returns the least busy CPU group within the | |
2763 | * domain. | |
2764 | */ | |
2765 | static struct sched_group * | |
78e7ed53 | 2766 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 2767 | int this_cpu, int load_idx) |
e7693a36 | 2768 | { |
b3bd3de6 | 2769 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 2770 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 2771 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 2772 | |
aaee1203 PZ |
2773 | do { |
2774 | unsigned long load, avg_load; | |
2775 | int local_group; | |
2776 | int i; | |
e7693a36 | 2777 | |
aaee1203 PZ |
2778 | /* Skip over this group if it has no CPUs allowed */ |
2779 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 2780 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
2781 | continue; |
2782 | ||
2783 | local_group = cpumask_test_cpu(this_cpu, | |
2784 | sched_group_cpus(group)); | |
2785 | ||
2786 | /* Tally up the load of all CPUs in the group */ | |
2787 | avg_load = 0; | |
2788 | ||
2789 | for_each_cpu(i, sched_group_cpus(group)) { | |
2790 | /* Bias balancing toward cpus of our domain */ | |
2791 | if (local_group) | |
2792 | load = source_load(i, load_idx); | |
2793 | else | |
2794 | load = target_load(i, load_idx); | |
2795 | ||
2796 | avg_load += load; | |
2797 | } | |
2798 | ||
2799 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2800 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
2801 | |
2802 | if (local_group) { | |
2803 | this_load = avg_load; | |
aaee1203 PZ |
2804 | } else if (avg_load < min_load) { |
2805 | min_load = avg_load; | |
2806 | idlest = group; | |
2807 | } | |
2808 | } while (group = group->next, group != sd->groups); | |
2809 | ||
2810 | if (!idlest || 100*this_load < imbalance*min_load) | |
2811 | return NULL; | |
2812 | return idlest; | |
2813 | } | |
2814 | ||
2815 | /* | |
2816 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
2817 | */ | |
2818 | static int | |
2819 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
2820 | { | |
2821 | unsigned long load, min_load = ULONG_MAX; | |
2822 | int idlest = -1; | |
2823 | int i; | |
2824 | ||
2825 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 2826 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
2827 | load = weighted_cpuload(i); |
2828 | ||
2829 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2830 | min_load = load; | |
2831 | idlest = i; | |
e7693a36 GH |
2832 | } |
2833 | } | |
2834 | ||
aaee1203 PZ |
2835 | return idlest; |
2836 | } | |
e7693a36 | 2837 | |
a50bde51 PZ |
2838 | /* |
2839 | * Try and locate an idle CPU in the sched_domain. | |
2840 | */ | |
99bd5e2f | 2841 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
2842 | { |
2843 | int cpu = smp_processor_id(); | |
2844 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 2845 | struct sched_domain *sd; |
37407ea7 LT |
2846 | struct sched_group *sg; |
2847 | int i; | |
a50bde51 PZ |
2848 | |
2849 | /* | |
99bd5e2f SS |
2850 | * If the task is going to be woken-up on this cpu and if it is |
2851 | * already idle, then it is the right target. | |
a50bde51 | 2852 | */ |
99bd5e2f SS |
2853 | if (target == cpu && idle_cpu(cpu)) |
2854 | return cpu; | |
2855 | ||
2856 | /* | |
2857 | * If the task is going to be woken-up on the cpu where it previously | |
2858 | * ran and if it is currently idle, then it the right target. | |
2859 | */ | |
2860 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 2861 | return prev_cpu; |
a50bde51 PZ |
2862 | |
2863 | /* | |
37407ea7 | 2864 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 2865 | */ |
518cd623 | 2866 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 2867 | for_each_lower_domain(sd) { |
37407ea7 LT |
2868 | sg = sd->groups; |
2869 | do { | |
2870 | if (!cpumask_intersects(sched_group_cpus(sg), | |
2871 | tsk_cpus_allowed(p))) | |
2872 | goto next; | |
2873 | ||
2874 | for_each_cpu(i, sched_group_cpus(sg)) { | |
2875 | if (!idle_cpu(i)) | |
2876 | goto next; | |
2877 | } | |
970e1789 | 2878 | |
37407ea7 LT |
2879 | target = cpumask_first_and(sched_group_cpus(sg), |
2880 | tsk_cpus_allowed(p)); | |
2881 | goto done; | |
2882 | next: | |
2883 | sg = sg->next; | |
2884 | } while (sg != sd->groups); | |
2885 | } | |
2886 | done: | |
a50bde51 PZ |
2887 | return target; |
2888 | } | |
2889 | ||
aaee1203 PZ |
2890 | /* |
2891 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2892 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2893 | * SD_BALANCE_EXEC. | |
2894 | * | |
2895 | * Balance, ie. select the least loaded group. | |
2896 | * | |
2897 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2898 | * | |
2899 | * preempt must be disabled. | |
2900 | */ | |
0017d735 | 2901 | static int |
7608dec2 | 2902 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 2903 | { |
29cd8bae | 2904 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
2905 | int cpu = smp_processor_id(); |
2906 | int prev_cpu = task_cpu(p); | |
2907 | int new_cpu = cpu; | |
99bd5e2f | 2908 | int want_affine = 0; |
5158f4e4 | 2909 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 2910 | |
29baa747 | 2911 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
2912 | return prev_cpu; |
2913 | ||
0763a660 | 2914 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 2915 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
2916 | want_affine = 1; |
2917 | new_cpu = prev_cpu; | |
2918 | } | |
aaee1203 | 2919 | |
dce840a0 | 2920 | rcu_read_lock(); |
aaee1203 | 2921 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
2922 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
2923 | continue; | |
2924 | ||
fe3bcfe1 | 2925 | /* |
99bd5e2f SS |
2926 | * If both cpu and prev_cpu are part of this domain, |
2927 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 2928 | */ |
99bd5e2f SS |
2929 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
2930 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
2931 | affine_sd = tmp; | |
29cd8bae | 2932 | break; |
f03542a7 | 2933 | } |
29cd8bae | 2934 | |
f03542a7 | 2935 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
2936 | sd = tmp; |
2937 | } | |
2938 | ||
8b911acd | 2939 | if (affine_sd) { |
f03542a7 | 2940 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
2941 | prev_cpu = cpu; |
2942 | ||
2943 | new_cpu = select_idle_sibling(p, prev_cpu); | |
2944 | goto unlock; | |
8b911acd | 2945 | } |
e7693a36 | 2946 | |
aaee1203 | 2947 | while (sd) { |
5158f4e4 | 2948 | int load_idx = sd->forkexec_idx; |
aaee1203 | 2949 | struct sched_group *group; |
c88d5910 | 2950 | int weight; |
098fb9db | 2951 | |
0763a660 | 2952 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
2953 | sd = sd->child; |
2954 | continue; | |
2955 | } | |
098fb9db | 2956 | |
5158f4e4 PZ |
2957 | if (sd_flag & SD_BALANCE_WAKE) |
2958 | load_idx = sd->wake_idx; | |
098fb9db | 2959 | |
5158f4e4 | 2960 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
2961 | if (!group) { |
2962 | sd = sd->child; | |
2963 | continue; | |
2964 | } | |
4ae7d5ce | 2965 | |
d7c33c49 | 2966 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
2967 | if (new_cpu == -1 || new_cpu == cpu) { |
2968 | /* Now try balancing at a lower domain level of cpu */ | |
2969 | sd = sd->child; | |
2970 | continue; | |
e7693a36 | 2971 | } |
aaee1203 PZ |
2972 | |
2973 | /* Now try balancing at a lower domain level of new_cpu */ | |
2974 | cpu = new_cpu; | |
669c55e9 | 2975 | weight = sd->span_weight; |
aaee1203 PZ |
2976 | sd = NULL; |
2977 | for_each_domain(cpu, tmp) { | |
669c55e9 | 2978 | if (weight <= tmp->span_weight) |
aaee1203 | 2979 | break; |
0763a660 | 2980 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
2981 | sd = tmp; |
2982 | } | |
2983 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 2984 | } |
dce840a0 PZ |
2985 | unlock: |
2986 | rcu_read_unlock(); | |
e7693a36 | 2987 | |
c88d5910 | 2988 | return new_cpu; |
e7693a36 GH |
2989 | } |
2990 | #endif /* CONFIG_SMP */ | |
2991 | ||
e52fb7c0 PZ |
2992 | static unsigned long |
2993 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
2994 | { |
2995 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
2996 | ||
2997 | /* | |
e52fb7c0 PZ |
2998 | * Since its curr running now, convert the gran from real-time |
2999 | * to virtual-time in his units. | |
13814d42 MG |
3000 | * |
3001 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
3002 | * they get preempted easier. That is, if 'se' < 'curr' then | |
3003 | * the resulting gran will be larger, therefore penalizing the | |
3004 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
3005 | * be smaller, again penalizing the lighter task. | |
3006 | * | |
3007 | * This is especially important for buddies when the leftmost | |
3008 | * task is higher priority than the buddy. | |
0bbd3336 | 3009 | */ |
f4ad9bd2 | 3010 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
3011 | } |
3012 | ||
464b7527 PZ |
3013 | /* |
3014 | * Should 'se' preempt 'curr'. | |
3015 | * | |
3016 | * |s1 | |
3017 | * |s2 | |
3018 | * |s3 | |
3019 | * g | |
3020 | * |<--->|c | |
3021 | * | |
3022 | * w(c, s1) = -1 | |
3023 | * w(c, s2) = 0 | |
3024 | * w(c, s3) = 1 | |
3025 | * | |
3026 | */ | |
3027 | static int | |
3028 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
3029 | { | |
3030 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
3031 | ||
3032 | if (vdiff <= 0) | |
3033 | return -1; | |
3034 | ||
e52fb7c0 | 3035 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
3036 | if (vdiff > gran) |
3037 | return 1; | |
3038 | ||
3039 | return 0; | |
3040 | } | |
3041 | ||
02479099 PZ |
3042 | static void set_last_buddy(struct sched_entity *se) |
3043 | { | |
69c80f3e VP |
3044 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3045 | return; | |
3046 | ||
3047 | for_each_sched_entity(se) | |
3048 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
3049 | } |
3050 | ||
3051 | static void set_next_buddy(struct sched_entity *se) | |
3052 | { | |
69c80f3e VP |
3053 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3054 | return; | |
3055 | ||
3056 | for_each_sched_entity(se) | |
3057 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
3058 | } |
3059 | ||
ac53db59 RR |
3060 | static void set_skip_buddy(struct sched_entity *se) |
3061 | { | |
69c80f3e VP |
3062 | for_each_sched_entity(se) |
3063 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
3064 | } |
3065 | ||
bf0f6f24 IM |
3066 | /* |
3067 | * Preempt the current task with a newly woken task if needed: | |
3068 | */ | |
5a9b86f6 | 3069 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
3070 | { |
3071 | struct task_struct *curr = rq->curr; | |
8651a86c | 3072 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 3073 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 3074 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 3075 | int next_buddy_marked = 0; |
bf0f6f24 | 3076 | |
4ae7d5ce IM |
3077 | if (unlikely(se == pse)) |
3078 | return; | |
3079 | ||
5238cdd3 | 3080 | /* |
ddcdf6e7 | 3081 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
3082 | * unconditionally check_prempt_curr() after an enqueue (which may have |
3083 | * lead to a throttle). This both saves work and prevents false | |
3084 | * next-buddy nomination below. | |
3085 | */ | |
3086 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3087 | return; | |
3088 | ||
2f36825b | 3089 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3090 | set_next_buddy(pse); |
2f36825b VP |
3091 | next_buddy_marked = 1; |
3092 | } | |
57fdc26d | 3093 | |
aec0a514 BR |
3094 | /* |
3095 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3096 | * wake up path. | |
5238cdd3 PT |
3097 | * |
3098 | * Note: this also catches the edge-case of curr being in a throttled | |
3099 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3100 | * enqueue of curr) will have resulted in resched being set. This | |
3101 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3102 | * below. | |
aec0a514 BR |
3103 | */ |
3104 | if (test_tsk_need_resched(curr)) | |
3105 | return; | |
3106 | ||
a2f5c9ab DH |
3107 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3108 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3109 | likely(p->policy != SCHED_IDLE)) | |
3110 | goto preempt; | |
3111 | ||
91c234b4 | 3112 | /* |
a2f5c9ab DH |
3113 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3114 | * is driven by the tick): | |
91c234b4 | 3115 | */ |
6bc912b7 | 3116 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 3117 | return; |
bf0f6f24 | 3118 | |
464b7527 | 3119 | find_matching_se(&se, &pse); |
9bbd7374 | 3120 | update_curr(cfs_rq_of(se)); |
002f128b | 3121 | BUG_ON(!pse); |
2f36825b VP |
3122 | if (wakeup_preempt_entity(se, pse) == 1) { |
3123 | /* | |
3124 | * Bias pick_next to pick the sched entity that is | |
3125 | * triggering this preemption. | |
3126 | */ | |
3127 | if (!next_buddy_marked) | |
3128 | set_next_buddy(pse); | |
3a7e73a2 | 3129 | goto preempt; |
2f36825b | 3130 | } |
464b7527 | 3131 | |
3a7e73a2 | 3132 | return; |
a65ac745 | 3133 | |
3a7e73a2 PZ |
3134 | preempt: |
3135 | resched_task(curr); | |
3136 | /* | |
3137 | * Only set the backward buddy when the current task is still | |
3138 | * on the rq. This can happen when a wakeup gets interleaved | |
3139 | * with schedule on the ->pre_schedule() or idle_balance() | |
3140 | * point, either of which can * drop the rq lock. | |
3141 | * | |
3142 | * Also, during early boot the idle thread is in the fair class, | |
3143 | * for obvious reasons its a bad idea to schedule back to it. | |
3144 | */ | |
3145 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3146 | return; | |
3147 | ||
3148 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3149 | set_last_buddy(se); | |
bf0f6f24 IM |
3150 | } |
3151 | ||
fb8d4724 | 3152 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3153 | { |
8f4d37ec | 3154 | struct task_struct *p; |
bf0f6f24 IM |
3155 | struct cfs_rq *cfs_rq = &rq->cfs; |
3156 | struct sched_entity *se; | |
3157 | ||
36ace27e | 3158 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3159 | return NULL; |
3160 | ||
3161 | do { | |
9948f4b2 | 3162 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3163 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3164 | cfs_rq = group_cfs_rq(se); |
3165 | } while (cfs_rq); | |
3166 | ||
8f4d37ec | 3167 | p = task_of(se); |
b39e66ea MG |
3168 | if (hrtick_enabled(rq)) |
3169 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
3170 | |
3171 | return p; | |
bf0f6f24 IM |
3172 | } |
3173 | ||
3174 | /* | |
3175 | * Account for a descheduled task: | |
3176 | */ | |
31ee529c | 3177 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
3178 | { |
3179 | struct sched_entity *se = &prev->se; | |
3180 | struct cfs_rq *cfs_rq; | |
3181 | ||
3182 | for_each_sched_entity(se) { | |
3183 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 3184 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
3185 | } |
3186 | } | |
3187 | ||
ac53db59 RR |
3188 | /* |
3189 | * sched_yield() is very simple | |
3190 | * | |
3191 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
3192 | */ | |
3193 | static void yield_task_fair(struct rq *rq) | |
3194 | { | |
3195 | struct task_struct *curr = rq->curr; | |
3196 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
3197 | struct sched_entity *se = &curr->se; | |
3198 | ||
3199 | /* | |
3200 | * Are we the only task in the tree? | |
3201 | */ | |
3202 | if (unlikely(rq->nr_running == 1)) | |
3203 | return; | |
3204 | ||
3205 | clear_buddies(cfs_rq, se); | |
3206 | ||
3207 | if (curr->policy != SCHED_BATCH) { | |
3208 | update_rq_clock(rq); | |
3209 | /* | |
3210 | * Update run-time statistics of the 'current'. | |
3211 | */ | |
3212 | update_curr(cfs_rq); | |
916671c0 MG |
3213 | /* |
3214 | * Tell update_rq_clock() that we've just updated, | |
3215 | * so we don't do microscopic update in schedule() | |
3216 | * and double the fastpath cost. | |
3217 | */ | |
3218 | rq->skip_clock_update = 1; | |
ac53db59 RR |
3219 | } |
3220 | ||
3221 | set_skip_buddy(se); | |
3222 | } | |
3223 | ||
d95f4122 MG |
3224 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
3225 | { | |
3226 | struct sched_entity *se = &p->se; | |
3227 | ||
5238cdd3 PT |
3228 | /* throttled hierarchies are not runnable */ |
3229 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
3230 | return false; |
3231 | ||
3232 | /* Tell the scheduler that we'd really like pse to run next. */ | |
3233 | set_next_buddy(se); | |
3234 | ||
d95f4122 MG |
3235 | yield_task_fair(rq); |
3236 | ||
3237 | return true; | |
3238 | } | |
3239 | ||
681f3e68 | 3240 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
3241 | /************************************************** |
3242 | * Fair scheduling class load-balancing methods: | |
3243 | */ | |
3244 | ||
ed387b78 HS |
3245 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3246 | ||
ddcdf6e7 | 3247 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 3248 | #define LBF_NEED_BREAK 0x02 |
88b8dac0 | 3249 | #define LBF_SOME_PINNED 0x04 |
ddcdf6e7 PZ |
3250 | |
3251 | struct lb_env { | |
3252 | struct sched_domain *sd; | |
3253 | ||
ddcdf6e7 | 3254 | struct rq *src_rq; |
85c1e7da | 3255 | int src_cpu; |
ddcdf6e7 PZ |
3256 | |
3257 | int dst_cpu; | |
3258 | struct rq *dst_rq; | |
3259 | ||
88b8dac0 SV |
3260 | struct cpumask *dst_grpmask; |
3261 | int new_dst_cpu; | |
ddcdf6e7 | 3262 | enum cpu_idle_type idle; |
bd939f45 | 3263 | long imbalance; |
b9403130 MW |
3264 | /* The set of CPUs under consideration for load-balancing */ |
3265 | struct cpumask *cpus; | |
3266 | ||
ddcdf6e7 | 3267 | unsigned int flags; |
367456c7 PZ |
3268 | |
3269 | unsigned int loop; | |
3270 | unsigned int loop_break; | |
3271 | unsigned int loop_max; | |
ddcdf6e7 PZ |
3272 | }; |
3273 | ||
1e3c88bd | 3274 | /* |
ddcdf6e7 | 3275 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
3276 | * Both runqueues must be locked. |
3277 | */ | |
ddcdf6e7 | 3278 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 3279 | { |
ddcdf6e7 PZ |
3280 | deactivate_task(env->src_rq, p, 0); |
3281 | set_task_cpu(p, env->dst_cpu); | |
3282 | activate_task(env->dst_rq, p, 0); | |
3283 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
3284 | } |
3285 | ||
029632fb PZ |
3286 | /* |
3287 | * Is this task likely cache-hot: | |
3288 | */ | |
3289 | static int | |
3290 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
3291 | { | |
3292 | s64 delta; | |
3293 | ||
3294 | if (p->sched_class != &fair_sched_class) | |
3295 | return 0; | |
3296 | ||
3297 | if (unlikely(p->policy == SCHED_IDLE)) | |
3298 | return 0; | |
3299 | ||
3300 | /* | |
3301 | * Buddy candidates are cache hot: | |
3302 | */ | |
3303 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
3304 | (&p->se == cfs_rq_of(&p->se)->next || | |
3305 | &p->se == cfs_rq_of(&p->se)->last)) | |
3306 | return 1; | |
3307 | ||
3308 | if (sysctl_sched_migration_cost == -1) | |
3309 | return 1; | |
3310 | if (sysctl_sched_migration_cost == 0) | |
3311 | return 0; | |
3312 | ||
3313 | delta = now - p->se.exec_start; | |
3314 | ||
3315 | return delta < (s64)sysctl_sched_migration_cost; | |
3316 | } | |
3317 | ||
1e3c88bd PZ |
3318 | /* |
3319 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3320 | */ | |
3321 | static | |
8e45cb54 | 3322 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
3323 | { |
3324 | int tsk_cache_hot = 0; | |
3325 | /* | |
3326 | * We do not migrate tasks that are: | |
3327 | * 1) running (obviously), or | |
3328 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3329 | * 3) are cache-hot on their current CPU. | |
3330 | */ | |
ddcdf6e7 | 3331 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
88b8dac0 SV |
3332 | int new_dst_cpu; |
3333 | ||
41acab88 | 3334 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 SV |
3335 | |
3336 | /* | |
3337 | * Remember if this task can be migrated to any other cpu in | |
3338 | * our sched_group. We may want to revisit it if we couldn't | |
3339 | * meet load balance goals by pulling other tasks on src_cpu. | |
3340 | * | |
3341 | * Also avoid computing new_dst_cpu if we have already computed | |
3342 | * one in current iteration. | |
3343 | */ | |
3344 | if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED)) | |
3345 | return 0; | |
3346 | ||
3347 | new_dst_cpu = cpumask_first_and(env->dst_grpmask, | |
3348 | tsk_cpus_allowed(p)); | |
3349 | if (new_dst_cpu < nr_cpu_ids) { | |
3350 | env->flags |= LBF_SOME_PINNED; | |
3351 | env->new_dst_cpu = new_dst_cpu; | |
3352 | } | |
1e3c88bd PZ |
3353 | return 0; |
3354 | } | |
88b8dac0 SV |
3355 | |
3356 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 3357 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 3358 | |
ddcdf6e7 | 3359 | if (task_running(env->src_rq, p)) { |
41acab88 | 3360 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
3361 | return 0; |
3362 | } | |
3363 | ||
3364 | /* | |
3365 | * Aggressive migration if: | |
3366 | * 1) task is cache cold, or | |
3367 | * 2) too many balance attempts have failed. | |
3368 | */ | |
3369 | ||
ddcdf6e7 | 3370 | tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd); |
1e3c88bd | 3371 | if (!tsk_cache_hot || |
8e45cb54 | 3372 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
1e3c88bd PZ |
3373 | #ifdef CONFIG_SCHEDSTATS |
3374 | if (tsk_cache_hot) { | |
8e45cb54 | 3375 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 3376 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
3377 | } |
3378 | #endif | |
3379 | return 1; | |
3380 | } | |
3381 | ||
3382 | if (tsk_cache_hot) { | |
41acab88 | 3383 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
3384 | return 0; |
3385 | } | |
3386 | return 1; | |
3387 | } | |
3388 | ||
897c395f PZ |
3389 | /* |
3390 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3391 | * part of active balancing operations within "domain". | |
3392 | * Returns 1 if successful and 0 otherwise. | |
3393 | * | |
3394 | * Called with both runqueues locked. | |
3395 | */ | |
8e45cb54 | 3396 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
3397 | { |
3398 | struct task_struct *p, *n; | |
897c395f | 3399 | |
367456c7 PZ |
3400 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
3401 | if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu)) | |
3402 | continue; | |
897c395f | 3403 | |
367456c7 PZ |
3404 | if (!can_migrate_task(p, env)) |
3405 | continue; | |
897c395f | 3406 | |
367456c7 PZ |
3407 | move_task(p, env); |
3408 | /* | |
3409 | * Right now, this is only the second place move_task() | |
3410 | * is called, so we can safely collect move_task() | |
3411 | * stats here rather than inside move_task(). | |
3412 | */ | |
3413 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
3414 | return 1; | |
897c395f | 3415 | } |
897c395f PZ |
3416 | return 0; |
3417 | } | |
3418 | ||
367456c7 PZ |
3419 | static unsigned long task_h_load(struct task_struct *p); |
3420 | ||
eb95308e PZ |
3421 | static const unsigned int sched_nr_migrate_break = 32; |
3422 | ||
5d6523eb | 3423 | /* |
bd939f45 | 3424 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
3425 | * this_rq, as part of a balancing operation within domain "sd". |
3426 | * Returns 1 if successful and 0 otherwise. | |
3427 | * | |
3428 | * Called with both runqueues locked. | |
3429 | */ | |
3430 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 3431 | { |
5d6523eb PZ |
3432 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
3433 | struct task_struct *p; | |
367456c7 PZ |
3434 | unsigned long load; |
3435 | int pulled = 0; | |
1e3c88bd | 3436 | |
bd939f45 | 3437 | if (env->imbalance <= 0) |
5d6523eb | 3438 | return 0; |
1e3c88bd | 3439 | |
5d6523eb PZ |
3440 | while (!list_empty(tasks)) { |
3441 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 3442 | |
367456c7 PZ |
3443 | env->loop++; |
3444 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 3445 | if (env->loop > env->loop_max) |
367456c7 | 3446 | break; |
5d6523eb PZ |
3447 | |
3448 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 3449 | if (env->loop > env->loop_break) { |
eb95308e | 3450 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 3451 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 3452 | break; |
a195f004 | 3453 | } |
1e3c88bd | 3454 | |
5d6523eb | 3455 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
367456c7 PZ |
3456 | goto next; |
3457 | ||
3458 | load = task_h_load(p); | |
5d6523eb | 3459 | |
eb95308e | 3460 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
3461 | goto next; |
3462 | ||
bd939f45 | 3463 | if ((load / 2) > env->imbalance) |
367456c7 | 3464 | goto next; |
1e3c88bd | 3465 | |
367456c7 PZ |
3466 | if (!can_migrate_task(p, env)) |
3467 | goto next; | |
1e3c88bd | 3468 | |
ddcdf6e7 | 3469 | move_task(p, env); |
ee00e66f | 3470 | pulled++; |
bd939f45 | 3471 | env->imbalance -= load; |
1e3c88bd PZ |
3472 | |
3473 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
3474 | /* |
3475 | * NEWIDLE balancing is a source of latency, so preemptible | |
3476 | * kernels will stop after the first task is pulled to minimize | |
3477 | * the critical section. | |
3478 | */ | |
5d6523eb | 3479 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 3480 | break; |
1e3c88bd PZ |
3481 | #endif |
3482 | ||
ee00e66f PZ |
3483 | /* |
3484 | * We only want to steal up to the prescribed amount of | |
3485 | * weighted load. | |
3486 | */ | |
bd939f45 | 3487 | if (env->imbalance <= 0) |
ee00e66f | 3488 | break; |
367456c7 PZ |
3489 | |
3490 | continue; | |
3491 | next: | |
5d6523eb | 3492 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 3493 | } |
5d6523eb | 3494 | |
1e3c88bd | 3495 | /* |
ddcdf6e7 PZ |
3496 | * Right now, this is one of only two places move_task() is called, |
3497 | * so we can safely collect move_task() stats here rather than | |
3498 | * inside move_task(). | |
1e3c88bd | 3499 | */ |
8e45cb54 | 3500 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 3501 | |
5d6523eb | 3502 | return pulled; |
1e3c88bd PZ |
3503 | } |
3504 | ||
230059de | 3505 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
3506 | /* |
3507 | * update tg->load_weight by folding this cpu's load_avg | |
3508 | */ | |
67e86250 | 3509 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
3510 | { |
3511 | struct cfs_rq *cfs_rq; | |
3512 | unsigned long flags; | |
3513 | struct rq *rq; | |
9e3081ca PZ |
3514 | |
3515 | if (!tg->se[cpu]) | |
3516 | return 0; | |
3517 | ||
3518 | rq = cpu_rq(cpu); | |
3519 | cfs_rq = tg->cfs_rq[cpu]; | |
3520 | ||
3521 | raw_spin_lock_irqsave(&rq->lock, flags); | |
3522 | ||
3523 | update_rq_clock(rq); | |
d6b55918 | 3524 | update_cfs_load(cfs_rq, 1); |
9e3081ca PZ |
3525 | |
3526 | /* | |
3527 | * We need to update shares after updating tg->load_weight in | |
3528 | * order to adjust the weight of groups with long running tasks. | |
3529 | */ | |
6d5ab293 | 3530 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
3531 | |
3532 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
3533 | ||
3534 | return 0; | |
3535 | } | |
3536 | ||
3537 | static void update_shares(int cpu) | |
3538 | { | |
3539 | struct cfs_rq *cfs_rq; | |
3540 | struct rq *rq = cpu_rq(cpu); | |
3541 | ||
3542 | rcu_read_lock(); | |
9763b67f PZ |
3543 | /* |
3544 | * Iterates the task_group tree in a bottom up fashion, see | |
3545 | * list_add_leaf_cfs_rq() for details. | |
3546 | */ | |
64660c86 PT |
3547 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
3548 | /* throttled entities do not contribute to load */ | |
3549 | if (throttled_hierarchy(cfs_rq)) | |
3550 | continue; | |
3551 | ||
67e86250 | 3552 | update_shares_cpu(cfs_rq->tg, cpu); |
64660c86 | 3553 | } |
9e3081ca PZ |
3554 | rcu_read_unlock(); |
3555 | } | |
3556 | ||
9763b67f PZ |
3557 | /* |
3558 | * Compute the cpu's hierarchical load factor for each task group. | |
3559 | * This needs to be done in a top-down fashion because the load of a child | |
3560 | * group is a fraction of its parents load. | |
3561 | */ | |
3562 | static int tg_load_down(struct task_group *tg, void *data) | |
3563 | { | |
3564 | unsigned long load; | |
3565 | long cpu = (long)data; | |
3566 | ||
3567 | if (!tg->parent) { | |
3568 | load = cpu_rq(cpu)->load.weight; | |
3569 | } else { | |
3570 | load = tg->parent->cfs_rq[cpu]->h_load; | |
3571 | load *= tg->se[cpu]->load.weight; | |
3572 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
3573 | } | |
3574 | ||
3575 | tg->cfs_rq[cpu]->h_load = load; | |
3576 | ||
3577 | return 0; | |
3578 | } | |
3579 | ||
3580 | static void update_h_load(long cpu) | |
3581 | { | |
a35b6466 PZ |
3582 | struct rq *rq = cpu_rq(cpu); |
3583 | unsigned long now = jiffies; | |
3584 | ||
3585 | if (rq->h_load_throttle == now) | |
3586 | return; | |
3587 | ||
3588 | rq->h_load_throttle = now; | |
3589 | ||
367456c7 | 3590 | rcu_read_lock(); |
9763b67f | 3591 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
367456c7 | 3592 | rcu_read_unlock(); |
9763b67f PZ |
3593 | } |
3594 | ||
367456c7 | 3595 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 3596 | { |
367456c7 PZ |
3597 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
3598 | unsigned long load; | |
230059de | 3599 | |
367456c7 PZ |
3600 | load = p->se.load.weight; |
3601 | load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1); | |
230059de | 3602 | |
367456c7 | 3603 | return load; |
230059de PZ |
3604 | } |
3605 | #else | |
9e3081ca PZ |
3606 | static inline void update_shares(int cpu) |
3607 | { | |
3608 | } | |
3609 | ||
367456c7 | 3610 | static inline void update_h_load(long cpu) |
230059de | 3611 | { |
230059de | 3612 | } |
230059de | 3613 | |
367456c7 | 3614 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 3615 | { |
367456c7 | 3616 | return p->se.load.weight; |
1e3c88bd | 3617 | } |
230059de | 3618 | #endif |
1e3c88bd | 3619 | |
1e3c88bd PZ |
3620 | /********** Helpers for find_busiest_group ************************/ |
3621 | /* | |
3622 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
3623 | * during load balancing. | |
3624 | */ | |
3625 | struct sd_lb_stats { | |
3626 | struct sched_group *busiest; /* Busiest group in this sd */ | |
3627 | struct sched_group *this; /* Local group in this sd */ | |
3628 | unsigned long total_load; /* Total load of all groups in sd */ | |
3629 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
3630 | unsigned long avg_load; /* Average load across all groups in sd */ | |
3631 | ||
3632 | /** Statistics of this group */ | |
3633 | unsigned long this_load; | |
3634 | unsigned long this_load_per_task; | |
3635 | unsigned long this_nr_running; | |
fab47622 | 3636 | unsigned long this_has_capacity; |
aae6d3dd | 3637 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
3638 | |
3639 | /* Statistics of the busiest group */ | |
aae6d3dd | 3640 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
3641 | unsigned long max_load; |
3642 | unsigned long busiest_load_per_task; | |
3643 | unsigned long busiest_nr_running; | |
dd5feea1 | 3644 | unsigned long busiest_group_capacity; |
fab47622 | 3645 | unsigned long busiest_has_capacity; |
aae6d3dd | 3646 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
3647 | |
3648 | int group_imb; /* Is there imbalance in this sd */ | |
1e3c88bd PZ |
3649 | }; |
3650 | ||
3651 | /* | |
3652 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
3653 | */ | |
3654 | struct sg_lb_stats { | |
3655 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
3656 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
3657 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
3658 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
3659 | unsigned long group_capacity; | |
aae6d3dd SS |
3660 | unsigned long idle_cpus; |
3661 | unsigned long group_weight; | |
1e3c88bd | 3662 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 3663 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
3664 | }; |
3665 | ||
1e3c88bd PZ |
3666 | /** |
3667 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
3668 | * @sd: The sched_domain whose load_idx is to be obtained. | |
3669 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
3670 | */ | |
3671 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
3672 | enum cpu_idle_type idle) | |
3673 | { | |
3674 | int load_idx; | |
3675 | ||
3676 | switch (idle) { | |
3677 | case CPU_NOT_IDLE: | |
3678 | load_idx = sd->busy_idx; | |
3679 | break; | |
3680 | ||
3681 | case CPU_NEWLY_IDLE: | |
3682 | load_idx = sd->newidle_idx; | |
3683 | break; | |
3684 | default: | |
3685 | load_idx = sd->idle_idx; | |
3686 | break; | |
3687 | } | |
3688 | ||
3689 | return load_idx; | |
3690 | } | |
3691 | ||
1e3c88bd PZ |
3692 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
3693 | { | |
1399fa78 | 3694 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
3695 | } |
3696 | ||
3697 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
3698 | { | |
3699 | return default_scale_freq_power(sd, cpu); | |
3700 | } | |
3701 | ||
3702 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
3703 | { | |
669c55e9 | 3704 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
3705 | unsigned long smt_gain = sd->smt_gain; |
3706 | ||
3707 | smt_gain /= weight; | |
3708 | ||
3709 | return smt_gain; | |
3710 | } | |
3711 | ||
3712 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
3713 | { | |
3714 | return default_scale_smt_power(sd, cpu); | |
3715 | } | |
3716 | ||
3717 | unsigned long scale_rt_power(int cpu) | |
3718 | { | |
3719 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 3720 | u64 total, available, age_stamp, avg; |
1e3c88bd | 3721 | |
b654f7de PZ |
3722 | /* |
3723 | * Since we're reading these variables without serialization make sure | |
3724 | * we read them once before doing sanity checks on them. | |
3725 | */ | |
3726 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
3727 | avg = ACCESS_ONCE(rq->rt_avg); | |
3728 | ||
3729 | total = sched_avg_period() + (rq->clock - age_stamp); | |
aa483808 | 3730 | |
b654f7de | 3731 | if (unlikely(total < avg)) { |
aa483808 VP |
3732 | /* Ensures that power won't end up being negative */ |
3733 | available = 0; | |
3734 | } else { | |
b654f7de | 3735 | available = total - avg; |
aa483808 | 3736 | } |
1e3c88bd | 3737 | |
1399fa78 NR |
3738 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
3739 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 3740 | |
1399fa78 | 3741 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3742 | |
3743 | return div_u64(available, total); | |
3744 | } | |
3745 | ||
3746 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
3747 | { | |
669c55e9 | 3748 | unsigned long weight = sd->span_weight; |
1399fa78 | 3749 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
3750 | struct sched_group *sdg = sd->groups; |
3751 | ||
1e3c88bd PZ |
3752 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
3753 | if (sched_feat(ARCH_POWER)) | |
3754 | power *= arch_scale_smt_power(sd, cpu); | |
3755 | else | |
3756 | power *= default_scale_smt_power(sd, cpu); | |
3757 | ||
1399fa78 | 3758 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3759 | } |
3760 | ||
9c3f75cb | 3761 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
3762 | |
3763 | if (sched_feat(ARCH_POWER)) | |
3764 | power *= arch_scale_freq_power(sd, cpu); | |
3765 | else | |
3766 | power *= default_scale_freq_power(sd, cpu); | |
3767 | ||
1399fa78 | 3768 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 3769 | |
1e3c88bd | 3770 | power *= scale_rt_power(cpu); |
1399fa78 | 3771 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3772 | |
3773 | if (!power) | |
3774 | power = 1; | |
3775 | ||
e51fd5e2 | 3776 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 3777 | sdg->sgp->power = power; |
1e3c88bd PZ |
3778 | } |
3779 | ||
029632fb | 3780 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
3781 | { |
3782 | struct sched_domain *child = sd->child; | |
3783 | struct sched_group *group, *sdg = sd->groups; | |
3784 | unsigned long power; | |
4ec4412e VG |
3785 | unsigned long interval; |
3786 | ||
3787 | interval = msecs_to_jiffies(sd->balance_interval); | |
3788 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
3789 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
3790 | |
3791 | if (!child) { | |
3792 | update_cpu_power(sd, cpu); | |
3793 | return; | |
3794 | } | |
3795 | ||
3796 | power = 0; | |
3797 | ||
74a5ce20 PZ |
3798 | if (child->flags & SD_OVERLAP) { |
3799 | /* | |
3800 | * SD_OVERLAP domains cannot assume that child groups | |
3801 | * span the current group. | |
3802 | */ | |
3803 | ||
3804 | for_each_cpu(cpu, sched_group_cpus(sdg)) | |
3805 | power += power_of(cpu); | |
3806 | } else { | |
3807 | /* | |
3808 | * !SD_OVERLAP domains can assume that child groups | |
3809 | * span the current group. | |
3810 | */ | |
3811 | ||
3812 | group = child->groups; | |
3813 | do { | |
3814 | power += group->sgp->power; | |
3815 | group = group->next; | |
3816 | } while (group != child->groups); | |
3817 | } | |
1e3c88bd | 3818 | |
c3decf0d | 3819 | sdg->sgp->power_orig = sdg->sgp->power = power; |
1e3c88bd PZ |
3820 | } |
3821 | ||
9d5efe05 SV |
3822 | /* |
3823 | * Try and fix up capacity for tiny siblings, this is needed when | |
3824 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
3825 | * which on its own isn't powerful enough. | |
3826 | * | |
3827 | * See update_sd_pick_busiest() and check_asym_packing(). | |
3828 | */ | |
3829 | static inline int | |
3830 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
3831 | { | |
3832 | /* | |
1399fa78 | 3833 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 3834 | */ |
a6c75f2f | 3835 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
3836 | return 0; |
3837 | ||
3838 | /* | |
3839 | * If ~90% of the cpu_power is still there, we're good. | |
3840 | */ | |
9c3f75cb | 3841 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
3842 | return 1; |
3843 | ||
3844 | return 0; | |
3845 | } | |
3846 | ||
1e3c88bd PZ |
3847 | /** |
3848 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 3849 | * @env: The load balancing environment. |
1e3c88bd | 3850 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 3851 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 3852 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
3853 | * @balance: Should we balance. |
3854 | * @sgs: variable to hold the statistics for this group. | |
3855 | */ | |
bd939f45 PZ |
3856 | static inline void update_sg_lb_stats(struct lb_env *env, |
3857 | struct sched_group *group, int load_idx, | |
b9403130 | 3858 | int local_group, int *balance, struct sg_lb_stats *sgs) |
1e3c88bd | 3859 | { |
e44bc5c5 PZ |
3860 | unsigned long nr_running, max_nr_running, min_nr_running; |
3861 | unsigned long load, max_cpu_load, min_cpu_load; | |
04f733b4 | 3862 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
dd5feea1 | 3863 | unsigned long avg_load_per_task = 0; |
bd939f45 | 3864 | int i; |
1e3c88bd | 3865 | |
871e35bc | 3866 | if (local_group) |
c1174876 | 3867 | balance_cpu = group_balance_cpu(group); |
1e3c88bd PZ |
3868 | |
3869 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
3870 | max_cpu_load = 0; |
3871 | min_cpu_load = ~0UL; | |
2582f0eb | 3872 | max_nr_running = 0; |
e44bc5c5 | 3873 | min_nr_running = ~0UL; |
1e3c88bd | 3874 | |
b9403130 | 3875 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
3876 | struct rq *rq = cpu_rq(i); |
3877 | ||
e44bc5c5 PZ |
3878 | nr_running = rq->nr_running; |
3879 | ||
1e3c88bd PZ |
3880 | /* Bias balancing toward cpus of our domain */ |
3881 | if (local_group) { | |
c1174876 PZ |
3882 | if (idle_cpu(i) && !first_idle_cpu && |
3883 | cpumask_test_cpu(i, sched_group_mask(group))) { | |
04f733b4 | 3884 | first_idle_cpu = 1; |
1e3c88bd PZ |
3885 | balance_cpu = i; |
3886 | } | |
04f733b4 PZ |
3887 | |
3888 | load = target_load(i, load_idx); | |
1e3c88bd PZ |
3889 | } else { |
3890 | load = source_load(i, load_idx); | |
e44bc5c5 | 3891 | if (load > max_cpu_load) |
1e3c88bd PZ |
3892 | max_cpu_load = load; |
3893 | if (min_cpu_load > load) | |
3894 | min_cpu_load = load; | |
e44bc5c5 PZ |
3895 | |
3896 | if (nr_running > max_nr_running) | |
3897 | max_nr_running = nr_running; | |
3898 | if (min_nr_running > nr_running) | |
3899 | min_nr_running = nr_running; | |
1e3c88bd PZ |
3900 | } |
3901 | ||
3902 | sgs->group_load += load; | |
e44bc5c5 | 3903 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 3904 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
3905 | if (idle_cpu(i)) |
3906 | sgs->idle_cpus++; | |
1e3c88bd PZ |
3907 | } |
3908 | ||
3909 | /* | |
3910 | * First idle cpu or the first cpu(busiest) in this sched group | |
3911 | * is eligible for doing load balancing at this and above | |
3912 | * domains. In the newly idle case, we will allow all the cpu's | |
3913 | * to do the newly idle load balance. | |
3914 | */ | |
4ec4412e | 3915 | if (local_group) { |
bd939f45 | 3916 | if (env->idle != CPU_NEWLY_IDLE) { |
04f733b4 | 3917 | if (balance_cpu != env->dst_cpu) { |
4ec4412e VG |
3918 | *balance = 0; |
3919 | return; | |
3920 | } | |
bd939f45 | 3921 | update_group_power(env->sd, env->dst_cpu); |
4ec4412e | 3922 | } else if (time_after_eq(jiffies, group->sgp->next_update)) |
bd939f45 | 3923 | update_group_power(env->sd, env->dst_cpu); |
1e3c88bd PZ |
3924 | } |
3925 | ||
3926 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3927 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 3928 | |
1e3c88bd PZ |
3929 | /* |
3930 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 3931 | * than the average weight of a task. |
1e3c88bd PZ |
3932 | * |
3933 | * APZ: with cgroup the avg task weight can vary wildly and | |
3934 | * might not be a suitable number - should we keep a | |
3935 | * normalized nr_running number somewhere that negates | |
3936 | * the hierarchy? | |
3937 | */ | |
dd5feea1 SS |
3938 | if (sgs->sum_nr_running) |
3939 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 3940 | |
e44bc5c5 PZ |
3941 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && |
3942 | (max_nr_running - min_nr_running) > 1) | |
1e3c88bd PZ |
3943 | sgs->group_imb = 1; |
3944 | ||
9c3f75cb | 3945 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 3946 | SCHED_POWER_SCALE); |
9d5efe05 | 3947 | if (!sgs->group_capacity) |
bd939f45 | 3948 | sgs->group_capacity = fix_small_capacity(env->sd, group); |
aae6d3dd | 3949 | sgs->group_weight = group->group_weight; |
fab47622 NR |
3950 | |
3951 | if (sgs->group_capacity > sgs->sum_nr_running) | |
3952 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
3953 | } |
3954 | ||
532cb4c4 MN |
3955 | /** |
3956 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 3957 | * @env: The load balancing environment. |
532cb4c4 MN |
3958 | * @sds: sched_domain statistics |
3959 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 3960 | * @sgs: sched_group statistics |
532cb4c4 MN |
3961 | * |
3962 | * Determine if @sg is a busier group than the previously selected | |
3963 | * busiest group. | |
3964 | */ | |
bd939f45 | 3965 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
3966 | struct sd_lb_stats *sds, |
3967 | struct sched_group *sg, | |
bd939f45 | 3968 | struct sg_lb_stats *sgs) |
532cb4c4 MN |
3969 | { |
3970 | if (sgs->avg_load <= sds->max_load) | |
3971 | return false; | |
3972 | ||
3973 | if (sgs->sum_nr_running > sgs->group_capacity) | |
3974 | return true; | |
3975 | ||
3976 | if (sgs->group_imb) | |
3977 | return true; | |
3978 | ||
3979 | /* | |
3980 | * ASYM_PACKING needs to move all the work to the lowest | |
3981 | * numbered CPUs in the group, therefore mark all groups | |
3982 | * higher than ourself as busy. | |
3983 | */ | |
bd939f45 PZ |
3984 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
3985 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
3986 | if (!sds->busiest) |
3987 | return true; | |
3988 | ||
3989 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
3990 | return true; | |
3991 | } | |
3992 | ||
3993 | return false; | |
3994 | } | |
3995 | ||
1e3c88bd | 3996 | /** |
461819ac | 3997 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 3998 | * @env: The load balancing environment. |
1e3c88bd PZ |
3999 | * @balance: Should we balance. |
4000 | * @sds: variable to hold the statistics for this sched_domain. | |
4001 | */ | |
bd939f45 | 4002 | static inline void update_sd_lb_stats(struct lb_env *env, |
b9403130 | 4003 | int *balance, struct sd_lb_stats *sds) |
1e3c88bd | 4004 | { |
bd939f45 PZ |
4005 | struct sched_domain *child = env->sd->child; |
4006 | struct sched_group *sg = env->sd->groups; | |
1e3c88bd PZ |
4007 | struct sg_lb_stats sgs; |
4008 | int load_idx, prefer_sibling = 0; | |
4009 | ||
4010 | if (child && child->flags & SD_PREFER_SIBLING) | |
4011 | prefer_sibling = 1; | |
4012 | ||
bd939f45 | 4013 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
4014 | |
4015 | do { | |
4016 | int local_group; | |
4017 | ||
bd939f45 | 4018 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
1e3c88bd | 4019 | memset(&sgs, 0, sizeof(sgs)); |
b9403130 | 4020 | update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs); |
1e3c88bd | 4021 | |
8f190fb3 | 4022 | if (local_group && !(*balance)) |
1e3c88bd PZ |
4023 | return; |
4024 | ||
4025 | sds->total_load += sgs.group_load; | |
9c3f75cb | 4026 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
4027 | |
4028 | /* | |
4029 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 4030 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
4031 | * and move all the excess tasks away. We lower the capacity |
4032 | * of a group only if the local group has the capacity to fit | |
4033 | * these excess tasks, i.e. nr_running < group_capacity. The | |
4034 | * extra check prevents the case where you always pull from the | |
4035 | * heaviest group when it is already under-utilized (possible | |
4036 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 4037 | */ |
75dd321d | 4038 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
4039 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
4040 | ||
4041 | if (local_group) { | |
4042 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 4043 | sds->this = sg; |
1e3c88bd PZ |
4044 | sds->this_nr_running = sgs.sum_nr_running; |
4045 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 4046 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4047 | sds->this_idle_cpus = sgs.idle_cpus; |
bd939f45 | 4048 | } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) { |
1e3c88bd | 4049 | sds->max_load = sgs.avg_load; |
532cb4c4 | 4050 | sds->busiest = sg; |
1e3c88bd | 4051 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 4052 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 4053 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 4054 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 4055 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4056 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
4057 | sds->group_imb = sgs.group_imb; |
4058 | } | |
4059 | ||
532cb4c4 | 4060 | sg = sg->next; |
bd939f45 | 4061 | } while (sg != env->sd->groups); |
532cb4c4 MN |
4062 | } |
4063 | ||
532cb4c4 MN |
4064 | /** |
4065 | * check_asym_packing - Check to see if the group is packed into the | |
4066 | * sched doman. | |
4067 | * | |
4068 | * This is primarily intended to used at the sibling level. Some | |
4069 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
4070 | * case of POWER7, it can move to lower SMT modes only when higher | |
4071 | * threads are idle. When in lower SMT modes, the threads will | |
4072 | * perform better since they share less core resources. Hence when we | |
4073 | * have idle threads, we want them to be the higher ones. | |
4074 | * | |
4075 | * This packing function is run on idle threads. It checks to see if | |
4076 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
4077 | * CPU number than the packing function is being run on. Here we are | |
4078 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
4079 | * number. | |
4080 | * | |
b6b12294 MN |
4081 | * Returns 1 when packing is required and a task should be moved to |
4082 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
4083 | * | |
cd96891d | 4084 | * @env: The load balancing environment. |
532cb4c4 | 4085 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 4086 | */ |
bd939f45 | 4087 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
4088 | { |
4089 | int busiest_cpu; | |
4090 | ||
bd939f45 | 4091 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
4092 | return 0; |
4093 | ||
4094 | if (!sds->busiest) | |
4095 | return 0; | |
4096 | ||
4097 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 4098 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
4099 | return 0; |
4100 | ||
bd939f45 PZ |
4101 | env->imbalance = DIV_ROUND_CLOSEST( |
4102 | sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE); | |
4103 | ||
532cb4c4 | 4104 | return 1; |
1e3c88bd PZ |
4105 | } |
4106 | ||
4107 | /** | |
4108 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
4109 | * amongst the groups of a sched_domain, during | |
4110 | * load balancing. | |
cd96891d | 4111 | * @env: The load balancing environment. |
1e3c88bd | 4112 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4113 | */ |
bd939f45 PZ |
4114 | static inline |
4115 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
4116 | { |
4117 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
4118 | unsigned int imbn = 2; | |
dd5feea1 | 4119 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
4120 | |
4121 | if (sds->this_nr_running) { | |
4122 | sds->this_load_per_task /= sds->this_nr_running; | |
4123 | if (sds->busiest_load_per_task > | |
4124 | sds->this_load_per_task) | |
4125 | imbn = 1; | |
bd939f45 | 4126 | } else { |
1e3c88bd | 4127 | sds->this_load_per_task = |
bd939f45 PZ |
4128 | cpu_avg_load_per_task(env->dst_cpu); |
4129 | } | |
1e3c88bd | 4130 | |
dd5feea1 | 4131 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 4132 | * SCHED_POWER_SCALE; |
9c3f75cb | 4133 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
4134 | |
4135 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
4136 | (scaled_busy_load_per_task * imbn)) { | |
bd939f45 | 4137 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4138 | return; |
4139 | } | |
4140 | ||
4141 | /* | |
4142 | * OK, we don't have enough imbalance to justify moving tasks, | |
4143 | * however we may be able to increase total CPU power used by | |
4144 | * moving them. | |
4145 | */ | |
4146 | ||
9c3f75cb | 4147 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 4148 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 4149 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 4150 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 4151 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4152 | |
4153 | /* Amount of load we'd subtract */ | |
1399fa78 | 4154 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 4155 | sds->busiest->sgp->power; |
1e3c88bd | 4156 | if (sds->max_load > tmp) |
9c3f75cb | 4157 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
4158 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
4159 | ||
4160 | /* Amount of load we'd add */ | |
9c3f75cb | 4161 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 4162 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
4163 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
4164 | sds->this->sgp->power; | |
1e3c88bd | 4165 | else |
1399fa78 | 4166 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
4167 | sds->this->sgp->power; |
4168 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 4169 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 4170 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4171 | |
4172 | /* Move if we gain throughput */ | |
4173 | if (pwr_move > pwr_now) | |
bd939f45 | 4174 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4175 | } |
4176 | ||
4177 | /** | |
4178 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
4179 | * groups of a given sched_domain during load balance. | |
bd939f45 | 4180 | * @env: load balance environment |
1e3c88bd | 4181 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4182 | */ |
bd939f45 | 4183 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 4184 | { |
dd5feea1 SS |
4185 | unsigned long max_pull, load_above_capacity = ~0UL; |
4186 | ||
4187 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
4188 | if (sds->group_imb) { | |
4189 | sds->busiest_load_per_task = | |
4190 | min(sds->busiest_load_per_task, sds->avg_load); | |
4191 | } | |
4192 | ||
1e3c88bd PZ |
4193 | /* |
4194 | * In the presence of smp nice balancing, certain scenarios can have | |
4195 | * max load less than avg load(as we skip the groups at or below | |
4196 | * its cpu_power, while calculating max_load..) | |
4197 | */ | |
4198 | if (sds->max_load < sds->avg_load) { | |
bd939f45 PZ |
4199 | env->imbalance = 0; |
4200 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4201 | } |
4202 | ||
dd5feea1 SS |
4203 | if (!sds->group_imb) { |
4204 | /* | |
4205 | * Don't want to pull so many tasks that a group would go idle. | |
4206 | */ | |
4207 | load_above_capacity = (sds->busiest_nr_running - | |
4208 | sds->busiest_group_capacity); | |
4209 | ||
1399fa78 | 4210 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 4211 | |
9c3f75cb | 4212 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
4213 | } |
4214 | ||
4215 | /* | |
4216 | * We're trying to get all the cpus to the average_load, so we don't | |
4217 | * want to push ourselves above the average load, nor do we wish to | |
4218 | * reduce the max loaded cpu below the average load. At the same time, | |
4219 | * we also don't want to reduce the group load below the group capacity | |
4220 | * (so that we can implement power-savings policies etc). Thus we look | |
4221 | * for the minimum possible imbalance. | |
4222 | * Be careful of negative numbers as they'll appear as very large values | |
4223 | * with unsigned longs. | |
4224 | */ | |
4225 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
4226 | |
4227 | /* How much load to actually move to equalise the imbalance */ | |
bd939f45 | 4228 | env->imbalance = min(max_pull * sds->busiest->sgp->power, |
9c3f75cb | 4229 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) |
1399fa78 | 4230 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
4231 | |
4232 | /* | |
4233 | * if *imbalance is less than the average load per runnable task | |
25985edc | 4234 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
4235 | * a think about bumping its value to force at least one task to be |
4236 | * moved | |
4237 | */ | |
bd939f45 PZ |
4238 | if (env->imbalance < sds->busiest_load_per_task) |
4239 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4240 | |
4241 | } | |
fab47622 | 4242 | |
1e3c88bd PZ |
4243 | /******* find_busiest_group() helpers end here *********************/ |
4244 | ||
4245 | /** | |
4246 | * find_busiest_group - Returns the busiest group within the sched_domain | |
4247 | * if there is an imbalance. If there isn't an imbalance, and | |
4248 | * the user has opted for power-savings, it returns a group whose | |
4249 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
4250 | * such a group exists. | |
4251 | * | |
4252 | * Also calculates the amount of weighted load which should be moved | |
4253 | * to restore balance. | |
4254 | * | |
cd96891d | 4255 | * @env: The load balancing environment. |
1e3c88bd PZ |
4256 | * @balance: Pointer to a variable indicating if this_cpu |
4257 | * is the appropriate cpu to perform load balancing at this_level. | |
4258 | * | |
4259 | * Returns: - the busiest group if imbalance exists. | |
4260 | * - If no imbalance and user has opted for power-savings balance, | |
4261 | * return the least loaded group whose CPUs can be | |
4262 | * put to idle by rebalancing its tasks onto our group. | |
4263 | */ | |
4264 | static struct sched_group * | |
b9403130 | 4265 | find_busiest_group(struct lb_env *env, int *balance) |
1e3c88bd PZ |
4266 | { |
4267 | struct sd_lb_stats sds; | |
4268 | ||
4269 | memset(&sds, 0, sizeof(sds)); | |
4270 | ||
4271 | /* | |
4272 | * Compute the various statistics relavent for load balancing at | |
4273 | * this level. | |
4274 | */ | |
b9403130 | 4275 | update_sd_lb_stats(env, balance, &sds); |
1e3c88bd | 4276 | |
cc57aa8f PZ |
4277 | /* |
4278 | * this_cpu is not the appropriate cpu to perform load balancing at | |
4279 | * this level. | |
1e3c88bd | 4280 | */ |
8f190fb3 | 4281 | if (!(*balance)) |
1e3c88bd PZ |
4282 | goto ret; |
4283 | ||
bd939f45 PZ |
4284 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
4285 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
4286 | return sds.busiest; |
4287 | ||
cc57aa8f | 4288 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
4289 | if (!sds.busiest || sds.busiest_nr_running == 0) |
4290 | goto out_balanced; | |
4291 | ||
1399fa78 | 4292 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 4293 | |
866ab43e PZ |
4294 | /* |
4295 | * If the busiest group is imbalanced the below checks don't | |
4296 | * work because they assumes all things are equal, which typically | |
4297 | * isn't true due to cpus_allowed constraints and the like. | |
4298 | */ | |
4299 | if (sds.group_imb) | |
4300 | goto force_balance; | |
4301 | ||
cc57aa8f | 4302 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
bd939f45 | 4303 | if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
fab47622 NR |
4304 | !sds.busiest_has_capacity) |
4305 | goto force_balance; | |
4306 | ||
cc57aa8f PZ |
4307 | /* |
4308 | * If the local group is more busy than the selected busiest group | |
4309 | * don't try and pull any tasks. | |
4310 | */ | |
1e3c88bd PZ |
4311 | if (sds.this_load >= sds.max_load) |
4312 | goto out_balanced; | |
4313 | ||
cc57aa8f PZ |
4314 | /* |
4315 | * Don't pull any tasks if this group is already above the domain | |
4316 | * average load. | |
4317 | */ | |
1e3c88bd PZ |
4318 | if (sds.this_load >= sds.avg_load) |
4319 | goto out_balanced; | |
4320 | ||
bd939f45 | 4321 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
4322 | /* |
4323 | * This cpu is idle. If the busiest group load doesn't | |
4324 | * have more tasks than the number of available cpu's and | |
4325 | * there is no imbalance between this and busiest group | |
4326 | * wrt to idle cpu's, it is balanced. | |
4327 | */ | |
c186fafe | 4328 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
4329 | sds.busiest_nr_running <= sds.busiest_group_weight) |
4330 | goto out_balanced; | |
c186fafe PZ |
4331 | } else { |
4332 | /* | |
4333 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
4334 | * imbalance_pct to be conservative. | |
4335 | */ | |
bd939f45 | 4336 | if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load) |
c186fafe | 4337 | goto out_balanced; |
aae6d3dd | 4338 | } |
1e3c88bd | 4339 | |
fab47622 | 4340 | force_balance: |
1e3c88bd | 4341 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 4342 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
4343 | return sds.busiest; |
4344 | ||
4345 | out_balanced: | |
1e3c88bd | 4346 | ret: |
bd939f45 | 4347 | env->imbalance = 0; |
1e3c88bd PZ |
4348 | return NULL; |
4349 | } | |
4350 | ||
4351 | /* | |
4352 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
4353 | */ | |
bd939f45 | 4354 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 4355 | struct sched_group *group) |
1e3c88bd PZ |
4356 | { |
4357 | struct rq *busiest = NULL, *rq; | |
4358 | unsigned long max_load = 0; | |
4359 | int i; | |
4360 | ||
4361 | for_each_cpu(i, sched_group_cpus(group)) { | |
4362 | unsigned long power = power_of(i); | |
1399fa78 NR |
4363 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
4364 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
4365 | unsigned long wl; |
4366 | ||
9d5efe05 | 4367 | if (!capacity) |
bd939f45 | 4368 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 4369 | |
b9403130 | 4370 | if (!cpumask_test_cpu(i, env->cpus)) |
1e3c88bd PZ |
4371 | continue; |
4372 | ||
4373 | rq = cpu_rq(i); | |
6e40f5bb | 4374 | wl = weighted_cpuload(i); |
1e3c88bd | 4375 | |
6e40f5bb TG |
4376 | /* |
4377 | * When comparing with imbalance, use weighted_cpuload() | |
4378 | * which is not scaled with the cpu power. | |
4379 | */ | |
bd939f45 | 4380 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
4381 | continue; |
4382 | ||
6e40f5bb TG |
4383 | /* |
4384 | * For the load comparisons with the other cpu's, consider | |
4385 | * the weighted_cpuload() scaled with the cpu power, so that | |
4386 | * the load can be moved away from the cpu that is potentially | |
4387 | * running at a lower capacity. | |
4388 | */ | |
1399fa78 | 4389 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 4390 | |
1e3c88bd PZ |
4391 | if (wl > max_load) { |
4392 | max_load = wl; | |
4393 | busiest = rq; | |
4394 | } | |
4395 | } | |
4396 | ||
4397 | return busiest; | |
4398 | } | |
4399 | ||
4400 | /* | |
4401 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
4402 | * so long as it is large enough. | |
4403 | */ | |
4404 | #define MAX_PINNED_INTERVAL 512 | |
4405 | ||
4406 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
029632fb | 4407 | DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); |
1e3c88bd | 4408 | |
bd939f45 | 4409 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 4410 | { |
bd939f45 PZ |
4411 | struct sched_domain *sd = env->sd; |
4412 | ||
4413 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
4414 | |
4415 | /* | |
4416 | * ASYM_PACKING needs to force migrate tasks from busy but | |
4417 | * higher numbered CPUs in order to pack all tasks in the | |
4418 | * lowest numbered CPUs. | |
4419 | */ | |
bd939f45 | 4420 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 4421 | return 1; |
1af3ed3d PZ |
4422 | } |
4423 | ||
4424 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
4425 | } | |
4426 | ||
969c7921 TH |
4427 | static int active_load_balance_cpu_stop(void *data); |
4428 | ||
1e3c88bd PZ |
4429 | /* |
4430 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
4431 | * tasks if there is an imbalance. | |
4432 | */ | |
4433 | static int load_balance(int this_cpu, struct rq *this_rq, | |
4434 | struct sched_domain *sd, enum cpu_idle_type idle, | |
4435 | int *balance) | |
4436 | { | |
88b8dac0 SV |
4437 | int ld_moved, cur_ld_moved, active_balance = 0; |
4438 | int lb_iterations, max_lb_iterations; | |
1e3c88bd | 4439 | struct sched_group *group; |
1e3c88bd PZ |
4440 | struct rq *busiest; |
4441 | unsigned long flags; | |
4442 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
4443 | ||
8e45cb54 PZ |
4444 | struct lb_env env = { |
4445 | .sd = sd, | |
ddcdf6e7 PZ |
4446 | .dst_cpu = this_cpu, |
4447 | .dst_rq = this_rq, | |
88b8dac0 | 4448 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 4449 | .idle = idle, |
eb95308e | 4450 | .loop_break = sched_nr_migrate_break, |
b9403130 | 4451 | .cpus = cpus, |
8e45cb54 PZ |
4452 | }; |
4453 | ||
1e3c88bd | 4454 | cpumask_copy(cpus, cpu_active_mask); |
88b8dac0 | 4455 | max_lb_iterations = cpumask_weight(env.dst_grpmask); |
1e3c88bd | 4456 | |
1e3c88bd PZ |
4457 | schedstat_inc(sd, lb_count[idle]); |
4458 | ||
4459 | redo: | |
b9403130 | 4460 | group = find_busiest_group(&env, balance); |
1e3c88bd PZ |
4461 | |
4462 | if (*balance == 0) | |
4463 | goto out_balanced; | |
4464 | ||
4465 | if (!group) { | |
4466 | schedstat_inc(sd, lb_nobusyg[idle]); | |
4467 | goto out_balanced; | |
4468 | } | |
4469 | ||
b9403130 | 4470 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
4471 | if (!busiest) { |
4472 | schedstat_inc(sd, lb_nobusyq[idle]); | |
4473 | goto out_balanced; | |
4474 | } | |
4475 | ||
78feefc5 | 4476 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 4477 | |
bd939f45 | 4478 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
4479 | |
4480 | ld_moved = 0; | |
88b8dac0 | 4481 | lb_iterations = 1; |
1e3c88bd PZ |
4482 | if (busiest->nr_running > 1) { |
4483 | /* | |
4484 | * Attempt to move tasks. If find_busiest_group has found | |
4485 | * an imbalance but busiest->nr_running <= 1, the group is | |
4486 | * still unbalanced. ld_moved simply stays zero, so it is | |
4487 | * correctly treated as an imbalance. | |
4488 | */ | |
8e45cb54 | 4489 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
4490 | env.src_cpu = busiest->cpu; |
4491 | env.src_rq = busiest; | |
4492 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 4493 | |
a35b6466 | 4494 | update_h_load(env.src_cpu); |
5d6523eb | 4495 | more_balance: |
1e3c88bd | 4496 | local_irq_save(flags); |
78feefc5 | 4497 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
4498 | |
4499 | /* | |
4500 | * cur_ld_moved - load moved in current iteration | |
4501 | * ld_moved - cumulative load moved across iterations | |
4502 | */ | |
4503 | cur_ld_moved = move_tasks(&env); | |
4504 | ld_moved += cur_ld_moved; | |
78feefc5 | 4505 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
4506 | local_irq_restore(flags); |
4507 | ||
5d6523eb PZ |
4508 | if (env.flags & LBF_NEED_BREAK) { |
4509 | env.flags &= ~LBF_NEED_BREAK; | |
4510 | goto more_balance; | |
4511 | } | |
4512 | ||
1e3c88bd PZ |
4513 | /* |
4514 | * some other cpu did the load balance for us. | |
4515 | */ | |
88b8dac0 SV |
4516 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
4517 | resched_cpu(env.dst_cpu); | |
4518 | ||
4519 | /* | |
4520 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
4521 | * us and move them to an alternate dst_cpu in our sched_group | |
4522 | * where they can run. The upper limit on how many times we | |
4523 | * iterate on same src_cpu is dependent on number of cpus in our | |
4524 | * sched_group. | |
4525 | * | |
4526 | * This changes load balance semantics a bit on who can move | |
4527 | * load to a given_cpu. In addition to the given_cpu itself | |
4528 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
4529 | * nohz-idle), we now have balance_cpu in a position to move | |
4530 | * load to given_cpu. In rare situations, this may cause | |
4531 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
4532 | * _independently_ and at _same_ time to move some load to | |
4533 | * given_cpu) causing exceess load to be moved to given_cpu. | |
4534 | * This however should not happen so much in practice and | |
4535 | * moreover subsequent load balance cycles should correct the | |
4536 | * excess load moved. | |
4537 | */ | |
4538 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 && | |
4539 | lb_iterations++ < max_lb_iterations) { | |
4540 | ||
78feefc5 | 4541 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 SV |
4542 | env.dst_cpu = env.new_dst_cpu; |
4543 | env.flags &= ~LBF_SOME_PINNED; | |
4544 | env.loop = 0; | |
4545 | env.loop_break = sched_nr_migrate_break; | |
4546 | /* | |
4547 | * Go back to "more_balance" rather than "redo" since we | |
4548 | * need to continue with same src_cpu. | |
4549 | */ | |
4550 | goto more_balance; | |
4551 | } | |
1e3c88bd PZ |
4552 | |
4553 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
8e45cb54 | 4554 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 4555 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
4556 | if (!cpumask_empty(cpus)) { |
4557 | env.loop = 0; | |
4558 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 4559 | goto redo; |
bbf18b19 | 4560 | } |
1e3c88bd PZ |
4561 | goto out_balanced; |
4562 | } | |
4563 | } | |
4564 | ||
4565 | if (!ld_moved) { | |
4566 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
4567 | /* |
4568 | * Increment the failure counter only on periodic balance. | |
4569 | * We do not want newidle balance, which can be very | |
4570 | * frequent, pollute the failure counter causing | |
4571 | * excessive cache_hot migrations and active balances. | |
4572 | */ | |
4573 | if (idle != CPU_NEWLY_IDLE) | |
4574 | sd->nr_balance_failed++; | |
1e3c88bd | 4575 | |
bd939f45 | 4576 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
4577 | raw_spin_lock_irqsave(&busiest->lock, flags); |
4578 | ||
969c7921 TH |
4579 | /* don't kick the active_load_balance_cpu_stop, |
4580 | * if the curr task on busiest cpu can't be | |
4581 | * moved to this_cpu | |
1e3c88bd PZ |
4582 | */ |
4583 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 4584 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
4585 | raw_spin_unlock_irqrestore(&busiest->lock, |
4586 | flags); | |
8e45cb54 | 4587 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
4588 | goto out_one_pinned; |
4589 | } | |
4590 | ||
969c7921 TH |
4591 | /* |
4592 | * ->active_balance synchronizes accesses to | |
4593 | * ->active_balance_work. Once set, it's cleared | |
4594 | * only after active load balance is finished. | |
4595 | */ | |
1e3c88bd PZ |
4596 | if (!busiest->active_balance) { |
4597 | busiest->active_balance = 1; | |
4598 | busiest->push_cpu = this_cpu; | |
4599 | active_balance = 1; | |
4600 | } | |
4601 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 4602 | |
bd939f45 | 4603 | if (active_balance) { |
969c7921 TH |
4604 | stop_one_cpu_nowait(cpu_of(busiest), |
4605 | active_load_balance_cpu_stop, busiest, | |
4606 | &busiest->active_balance_work); | |
bd939f45 | 4607 | } |
1e3c88bd PZ |
4608 | |
4609 | /* | |
4610 | * We've kicked active balancing, reset the failure | |
4611 | * counter. | |
4612 | */ | |
4613 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
4614 | } | |
4615 | } else | |
4616 | sd->nr_balance_failed = 0; | |
4617 | ||
4618 | if (likely(!active_balance)) { | |
4619 | /* We were unbalanced, so reset the balancing interval */ | |
4620 | sd->balance_interval = sd->min_interval; | |
4621 | } else { | |
4622 | /* | |
4623 | * If we've begun active balancing, start to back off. This | |
4624 | * case may not be covered by the all_pinned logic if there | |
4625 | * is only 1 task on the busy runqueue (because we don't call | |
4626 | * move_tasks). | |
4627 | */ | |
4628 | if (sd->balance_interval < sd->max_interval) | |
4629 | sd->balance_interval *= 2; | |
4630 | } | |
4631 | ||
1e3c88bd PZ |
4632 | goto out; |
4633 | ||
4634 | out_balanced: | |
4635 | schedstat_inc(sd, lb_balanced[idle]); | |
4636 | ||
4637 | sd->nr_balance_failed = 0; | |
4638 | ||
4639 | out_one_pinned: | |
4640 | /* tune up the balancing interval */ | |
8e45cb54 | 4641 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 4642 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
4643 | (sd->balance_interval < sd->max_interval)) |
4644 | sd->balance_interval *= 2; | |
4645 | ||
46e49b38 | 4646 | ld_moved = 0; |
1e3c88bd | 4647 | out: |
1e3c88bd PZ |
4648 | return ld_moved; |
4649 | } | |
4650 | ||
1e3c88bd PZ |
4651 | /* |
4652 | * idle_balance is called by schedule() if this_cpu is about to become | |
4653 | * idle. Attempts to pull tasks from other CPUs. | |
4654 | */ | |
029632fb | 4655 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
4656 | { |
4657 | struct sched_domain *sd; | |
4658 | int pulled_task = 0; | |
4659 | unsigned long next_balance = jiffies + HZ; | |
4660 | ||
4661 | this_rq->idle_stamp = this_rq->clock; | |
4662 | ||
4663 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
4664 | return; | |
4665 | ||
f492e12e PZ |
4666 | /* |
4667 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
4668 | */ | |
4669 | raw_spin_unlock(&this_rq->lock); | |
4670 | ||
c66eaf61 | 4671 | update_shares(this_cpu); |
dce840a0 | 4672 | rcu_read_lock(); |
1e3c88bd PZ |
4673 | for_each_domain(this_cpu, sd) { |
4674 | unsigned long interval; | |
f492e12e | 4675 | int balance = 1; |
1e3c88bd PZ |
4676 | |
4677 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4678 | continue; | |
4679 | ||
f492e12e | 4680 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 4681 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
4682 | pulled_task = load_balance(this_cpu, this_rq, |
4683 | sd, CPU_NEWLY_IDLE, &balance); | |
4684 | } | |
1e3c88bd PZ |
4685 | |
4686 | interval = msecs_to_jiffies(sd->balance_interval); | |
4687 | if (time_after(next_balance, sd->last_balance + interval)) | |
4688 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
4689 | if (pulled_task) { |
4690 | this_rq->idle_stamp = 0; | |
1e3c88bd | 4691 | break; |
d5ad140b | 4692 | } |
1e3c88bd | 4693 | } |
dce840a0 | 4694 | rcu_read_unlock(); |
f492e12e PZ |
4695 | |
4696 | raw_spin_lock(&this_rq->lock); | |
4697 | ||
1e3c88bd PZ |
4698 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
4699 | /* | |
4700 | * We are going idle. next_balance may be set based on | |
4701 | * a busy processor. So reset next_balance. | |
4702 | */ | |
4703 | this_rq->next_balance = next_balance; | |
4704 | } | |
4705 | } | |
4706 | ||
4707 | /* | |
969c7921 TH |
4708 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
4709 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
4710 | * least 1 task to be running on each physical CPU where possible, and | |
4711 | * avoids physical / logical imbalances. | |
1e3c88bd | 4712 | */ |
969c7921 | 4713 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 4714 | { |
969c7921 TH |
4715 | struct rq *busiest_rq = data; |
4716 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 4717 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 4718 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 4719 | struct sched_domain *sd; |
969c7921 TH |
4720 | |
4721 | raw_spin_lock_irq(&busiest_rq->lock); | |
4722 | ||
4723 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
4724 | if (unlikely(busiest_cpu != smp_processor_id() || | |
4725 | !busiest_rq->active_balance)) | |
4726 | goto out_unlock; | |
1e3c88bd PZ |
4727 | |
4728 | /* Is there any task to move? */ | |
4729 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 4730 | goto out_unlock; |
1e3c88bd PZ |
4731 | |
4732 | /* | |
4733 | * This condition is "impossible", if it occurs | |
4734 | * we need to fix it. Originally reported by | |
4735 | * Bjorn Helgaas on a 128-cpu setup. | |
4736 | */ | |
4737 | BUG_ON(busiest_rq == target_rq); | |
4738 | ||
4739 | /* move a task from busiest_rq to target_rq */ | |
4740 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
4741 | |
4742 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 4743 | rcu_read_lock(); |
1e3c88bd PZ |
4744 | for_each_domain(target_cpu, sd) { |
4745 | if ((sd->flags & SD_LOAD_BALANCE) && | |
4746 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
4747 | break; | |
4748 | } | |
4749 | ||
4750 | if (likely(sd)) { | |
8e45cb54 PZ |
4751 | struct lb_env env = { |
4752 | .sd = sd, | |
ddcdf6e7 PZ |
4753 | .dst_cpu = target_cpu, |
4754 | .dst_rq = target_rq, | |
4755 | .src_cpu = busiest_rq->cpu, | |
4756 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
4757 | .idle = CPU_IDLE, |
4758 | }; | |
4759 | ||
1e3c88bd PZ |
4760 | schedstat_inc(sd, alb_count); |
4761 | ||
8e45cb54 | 4762 | if (move_one_task(&env)) |
1e3c88bd PZ |
4763 | schedstat_inc(sd, alb_pushed); |
4764 | else | |
4765 | schedstat_inc(sd, alb_failed); | |
4766 | } | |
dce840a0 | 4767 | rcu_read_unlock(); |
1e3c88bd | 4768 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
4769 | out_unlock: |
4770 | busiest_rq->active_balance = 0; | |
4771 | raw_spin_unlock_irq(&busiest_rq->lock); | |
4772 | return 0; | |
1e3c88bd PZ |
4773 | } |
4774 | ||
4775 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
4776 | /* |
4777 | * idle load balancing details | |
83cd4fe2 VP |
4778 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
4779 | * needed, they will kick the idle load balancer, which then does idle | |
4780 | * load balancing for all the idle CPUs. | |
4781 | */ | |
1e3c88bd | 4782 | static struct { |
83cd4fe2 | 4783 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 4784 | atomic_t nr_cpus; |
83cd4fe2 VP |
4785 | unsigned long next_balance; /* in jiffy units */ |
4786 | } nohz ____cacheline_aligned; | |
1e3c88bd | 4787 | |
8e7fbcbc | 4788 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 4789 | { |
0b005cf5 | 4790 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 4791 | |
786d6dc7 SS |
4792 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
4793 | return ilb; | |
4794 | ||
4795 | return nr_cpu_ids; | |
1e3c88bd | 4796 | } |
1e3c88bd | 4797 | |
83cd4fe2 VP |
4798 | /* |
4799 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
4800 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
4801 | * CPU (if there is one). | |
4802 | */ | |
4803 | static void nohz_balancer_kick(int cpu) | |
4804 | { | |
4805 | int ilb_cpu; | |
4806 | ||
4807 | nohz.next_balance++; | |
4808 | ||
0b005cf5 | 4809 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 4810 | |
0b005cf5 SS |
4811 | if (ilb_cpu >= nr_cpu_ids) |
4812 | return; | |
83cd4fe2 | 4813 | |
cd490c5b | 4814 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
4815 | return; |
4816 | /* | |
4817 | * Use smp_send_reschedule() instead of resched_cpu(). | |
4818 | * This way we generate a sched IPI on the target cpu which | |
4819 | * is idle. And the softirq performing nohz idle load balance | |
4820 | * will be run before returning from the IPI. | |
4821 | */ | |
4822 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
4823 | return; |
4824 | } | |
4825 | ||
c1cc017c | 4826 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
4827 | { |
4828 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
4829 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
4830 | atomic_dec(&nohz.nr_cpus); | |
4831 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
4832 | } | |
4833 | } | |
4834 | ||
69e1e811 SS |
4835 | static inline void set_cpu_sd_state_busy(void) |
4836 | { | |
4837 | struct sched_domain *sd; | |
4838 | int cpu = smp_processor_id(); | |
4839 | ||
4840 | if (!test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
4841 | return; | |
4842 | clear_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
4843 | ||
4844 | rcu_read_lock(); | |
4845 | for_each_domain(cpu, sd) | |
4846 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); | |
4847 | rcu_read_unlock(); | |
4848 | } | |
4849 | ||
4850 | void set_cpu_sd_state_idle(void) | |
4851 | { | |
4852 | struct sched_domain *sd; | |
4853 | int cpu = smp_processor_id(); | |
4854 | ||
4855 | if (test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
4856 | return; | |
4857 | set_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
4858 | ||
4859 | rcu_read_lock(); | |
4860 | for_each_domain(cpu, sd) | |
4861 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); | |
4862 | rcu_read_unlock(); | |
4863 | } | |
4864 | ||
1e3c88bd | 4865 | /* |
c1cc017c | 4866 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 4867 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 4868 | */ |
c1cc017c | 4869 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 4870 | { |
71325960 SS |
4871 | /* |
4872 | * If this cpu is going down, then nothing needs to be done. | |
4873 | */ | |
4874 | if (!cpu_active(cpu)) | |
4875 | return; | |
4876 | ||
c1cc017c AS |
4877 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
4878 | return; | |
1e3c88bd | 4879 | |
c1cc017c AS |
4880 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
4881 | atomic_inc(&nohz.nr_cpus); | |
4882 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 4883 | } |
71325960 SS |
4884 | |
4885 | static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb, | |
4886 | unsigned long action, void *hcpu) | |
4887 | { | |
4888 | switch (action & ~CPU_TASKS_FROZEN) { | |
4889 | case CPU_DYING: | |
c1cc017c | 4890 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
4891 | return NOTIFY_OK; |
4892 | default: | |
4893 | return NOTIFY_DONE; | |
4894 | } | |
4895 | } | |
1e3c88bd PZ |
4896 | #endif |
4897 | ||
4898 | static DEFINE_SPINLOCK(balancing); | |
4899 | ||
49c022e6 PZ |
4900 | /* |
4901 | * Scale the max load_balance interval with the number of CPUs in the system. | |
4902 | * This trades load-balance latency on larger machines for less cross talk. | |
4903 | */ | |
029632fb | 4904 | void update_max_interval(void) |
49c022e6 PZ |
4905 | { |
4906 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
4907 | } | |
4908 | ||
1e3c88bd PZ |
4909 | /* |
4910 | * It checks each scheduling domain to see if it is due to be balanced, | |
4911 | * and initiates a balancing operation if so. | |
4912 | * | |
4913 | * Balancing parameters are set up in arch_init_sched_domains. | |
4914 | */ | |
4915 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4916 | { | |
4917 | int balance = 1; | |
4918 | struct rq *rq = cpu_rq(cpu); | |
4919 | unsigned long interval; | |
04f733b4 | 4920 | struct sched_domain *sd; |
1e3c88bd PZ |
4921 | /* Earliest time when we have to do rebalance again */ |
4922 | unsigned long next_balance = jiffies + 60*HZ; | |
4923 | int update_next_balance = 0; | |
4924 | int need_serialize; | |
4925 | ||
2069dd75 PZ |
4926 | update_shares(cpu); |
4927 | ||
dce840a0 | 4928 | rcu_read_lock(); |
1e3c88bd PZ |
4929 | for_each_domain(cpu, sd) { |
4930 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4931 | continue; | |
4932 | ||
4933 | interval = sd->balance_interval; | |
4934 | if (idle != CPU_IDLE) | |
4935 | interval *= sd->busy_factor; | |
4936 | ||
4937 | /* scale ms to jiffies */ | |
4938 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 4939 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
4940 | |
4941 | need_serialize = sd->flags & SD_SERIALIZE; | |
4942 | ||
4943 | if (need_serialize) { | |
4944 | if (!spin_trylock(&balancing)) | |
4945 | goto out; | |
4946 | } | |
4947 | ||
4948 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4949 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
4950 | /* | |
4951 | * We've pulled tasks over so either we're no | |
c186fafe | 4952 | * longer idle. |
1e3c88bd PZ |
4953 | */ |
4954 | idle = CPU_NOT_IDLE; | |
4955 | } | |
4956 | sd->last_balance = jiffies; | |
4957 | } | |
4958 | if (need_serialize) | |
4959 | spin_unlock(&balancing); | |
4960 | out: | |
4961 | if (time_after(next_balance, sd->last_balance + interval)) { | |
4962 | next_balance = sd->last_balance + interval; | |
4963 | update_next_balance = 1; | |
4964 | } | |
4965 | ||
4966 | /* | |
4967 | * Stop the load balance at this level. There is another | |
4968 | * CPU in our sched group which is doing load balancing more | |
4969 | * actively. | |
4970 | */ | |
4971 | if (!balance) | |
4972 | break; | |
4973 | } | |
dce840a0 | 4974 | rcu_read_unlock(); |
1e3c88bd PZ |
4975 | |
4976 | /* | |
4977 | * next_balance will be updated only when there is a need. | |
4978 | * When the cpu is attached to null domain for ex, it will not be | |
4979 | * updated. | |
4980 | */ | |
4981 | if (likely(update_next_balance)) | |
4982 | rq->next_balance = next_balance; | |
4983 | } | |
4984 | ||
83cd4fe2 | 4985 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 4986 | /* |
83cd4fe2 | 4987 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
4988 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
4989 | */ | |
83cd4fe2 VP |
4990 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
4991 | { | |
4992 | struct rq *this_rq = cpu_rq(this_cpu); | |
4993 | struct rq *rq; | |
4994 | int balance_cpu; | |
4995 | ||
1c792db7 SS |
4996 | if (idle != CPU_IDLE || |
4997 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
4998 | goto end; | |
83cd4fe2 VP |
4999 | |
5000 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 5001 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
5002 | continue; |
5003 | ||
5004 | /* | |
5005 | * If this cpu gets work to do, stop the load balancing | |
5006 | * work being done for other cpus. Next load | |
5007 | * balancing owner will pick it up. | |
5008 | */ | |
1c792db7 | 5009 | if (need_resched()) |
83cd4fe2 | 5010 | break; |
83cd4fe2 | 5011 | |
5ed4f1d9 VG |
5012 | rq = cpu_rq(balance_cpu); |
5013 | ||
5014 | raw_spin_lock_irq(&rq->lock); | |
5015 | update_rq_clock(rq); | |
5016 | update_idle_cpu_load(rq); | |
5017 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
5018 | |
5019 | rebalance_domains(balance_cpu, CPU_IDLE); | |
5020 | ||
83cd4fe2 VP |
5021 | if (time_after(this_rq->next_balance, rq->next_balance)) |
5022 | this_rq->next_balance = rq->next_balance; | |
5023 | } | |
5024 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
5025 | end: |
5026 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
5027 | } |
5028 | ||
5029 | /* | |
0b005cf5 SS |
5030 | * Current heuristic for kicking the idle load balancer in the presence |
5031 | * of an idle cpu is the system. | |
5032 | * - This rq has more than one task. | |
5033 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
5034 | * busy cpu's exceeding the group's power. | |
5035 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
5036 | * domain span are idle. | |
83cd4fe2 VP |
5037 | */ |
5038 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
5039 | { | |
5040 | unsigned long now = jiffies; | |
0b005cf5 | 5041 | struct sched_domain *sd; |
83cd4fe2 | 5042 | |
1c792db7 | 5043 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
5044 | return 0; |
5045 | ||
1c792db7 SS |
5046 | /* |
5047 | * We may be recently in ticked or tickless idle mode. At the first | |
5048 | * busy tick after returning from idle, we will update the busy stats. | |
5049 | */ | |
69e1e811 | 5050 | set_cpu_sd_state_busy(); |
c1cc017c | 5051 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
5052 | |
5053 | /* | |
5054 | * None are in tickless mode and hence no need for NOHZ idle load | |
5055 | * balancing. | |
5056 | */ | |
5057 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
5058 | return 0; | |
1c792db7 SS |
5059 | |
5060 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
5061 | return 0; |
5062 | ||
0b005cf5 SS |
5063 | if (rq->nr_running >= 2) |
5064 | goto need_kick; | |
83cd4fe2 | 5065 | |
067491b7 | 5066 | rcu_read_lock(); |
0b005cf5 SS |
5067 | for_each_domain(cpu, sd) { |
5068 | struct sched_group *sg = sd->groups; | |
5069 | struct sched_group_power *sgp = sg->sgp; | |
5070 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 5071 | |
0b005cf5 | 5072 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 5073 | goto need_kick_unlock; |
0b005cf5 SS |
5074 | |
5075 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
5076 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
5077 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 5078 | goto need_kick_unlock; |
0b005cf5 SS |
5079 | |
5080 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
5081 | break; | |
83cd4fe2 | 5082 | } |
067491b7 | 5083 | rcu_read_unlock(); |
83cd4fe2 | 5084 | return 0; |
067491b7 PZ |
5085 | |
5086 | need_kick_unlock: | |
5087 | rcu_read_unlock(); | |
0b005cf5 SS |
5088 | need_kick: |
5089 | return 1; | |
83cd4fe2 VP |
5090 | } |
5091 | #else | |
5092 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
5093 | #endif | |
5094 | ||
5095 | /* | |
5096 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
5097 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
5098 | */ | |
1e3c88bd PZ |
5099 | static void run_rebalance_domains(struct softirq_action *h) |
5100 | { | |
5101 | int this_cpu = smp_processor_id(); | |
5102 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 5103 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
5104 | CPU_IDLE : CPU_NOT_IDLE; |
5105 | ||
5106 | rebalance_domains(this_cpu, idle); | |
5107 | ||
1e3c88bd | 5108 | /* |
83cd4fe2 | 5109 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
5110 | * balancing on behalf of the other idle cpus whose ticks are |
5111 | * stopped. | |
5112 | */ | |
83cd4fe2 | 5113 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
5114 | } |
5115 | ||
5116 | static inline int on_null_domain(int cpu) | |
5117 | { | |
90a6501f | 5118 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
5119 | } |
5120 | ||
5121 | /* | |
5122 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 5123 | */ |
029632fb | 5124 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 5125 | { |
1e3c88bd PZ |
5126 | /* Don't need to rebalance while attached to NULL domain */ |
5127 | if (time_after_eq(jiffies, rq->next_balance) && | |
5128 | likely(!on_null_domain(cpu))) | |
5129 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 | 5130 | #ifdef CONFIG_NO_HZ |
1c792db7 | 5131 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
5132 | nohz_balancer_kick(cpu); |
5133 | #endif | |
1e3c88bd PZ |
5134 | } |
5135 | ||
0bcdcf28 CE |
5136 | static void rq_online_fair(struct rq *rq) |
5137 | { | |
5138 | update_sysctl(); | |
5139 | } | |
5140 | ||
5141 | static void rq_offline_fair(struct rq *rq) | |
5142 | { | |
5143 | update_sysctl(); | |
a4c96ae3 PB |
5144 | |
5145 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
5146 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
5147 | } |
5148 | ||
55e12e5e | 5149 | #endif /* CONFIG_SMP */ |
e1d1484f | 5150 | |
bf0f6f24 IM |
5151 | /* |
5152 | * scheduler tick hitting a task of our scheduling class: | |
5153 | */ | |
8f4d37ec | 5154 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
5155 | { |
5156 | struct cfs_rq *cfs_rq; | |
5157 | struct sched_entity *se = &curr->se; | |
5158 | ||
5159 | for_each_sched_entity(se) { | |
5160 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 5161 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 5162 | } |
cbee9f88 PZ |
5163 | |
5164 | if (sched_feat_numa(NUMA)) | |
5165 | task_tick_numa(rq, curr); | |
bf0f6f24 IM |
5166 | } |
5167 | ||
5168 | /* | |
cd29fe6f PZ |
5169 | * called on fork with the child task as argument from the parent's context |
5170 | * - child not yet on the tasklist | |
5171 | * - preemption disabled | |
bf0f6f24 | 5172 | */ |
cd29fe6f | 5173 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 5174 | { |
4fc420c9 DN |
5175 | struct cfs_rq *cfs_rq; |
5176 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 5177 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
5178 | struct rq *rq = this_rq(); |
5179 | unsigned long flags; | |
5180 | ||
05fa785c | 5181 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 5182 | |
861d034e PZ |
5183 | update_rq_clock(rq); |
5184 | ||
4fc420c9 DN |
5185 | cfs_rq = task_cfs_rq(current); |
5186 | curr = cfs_rq->curr; | |
5187 | ||
b0a0f667 PM |
5188 | if (unlikely(task_cpu(p) != this_cpu)) { |
5189 | rcu_read_lock(); | |
cd29fe6f | 5190 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
5191 | rcu_read_unlock(); |
5192 | } | |
bf0f6f24 | 5193 | |
7109c442 | 5194 | update_curr(cfs_rq); |
cd29fe6f | 5195 | |
b5d9d734 MG |
5196 | if (curr) |
5197 | se->vruntime = curr->vruntime; | |
aeb73b04 | 5198 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 5199 | |
cd29fe6f | 5200 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 5201 | /* |
edcb60a3 IM |
5202 | * Upon rescheduling, sched_class::put_prev_task() will place |
5203 | * 'current' within the tree based on its new key value. | |
5204 | */ | |
4d78e7b6 | 5205 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 5206 | resched_task(rq->curr); |
4d78e7b6 | 5207 | } |
bf0f6f24 | 5208 | |
88ec22d3 PZ |
5209 | se->vruntime -= cfs_rq->min_vruntime; |
5210 | ||
05fa785c | 5211 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
5212 | } |
5213 | ||
cb469845 SR |
5214 | /* |
5215 | * Priority of the task has changed. Check to see if we preempt | |
5216 | * the current task. | |
5217 | */ | |
da7a735e PZ |
5218 | static void |
5219 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 5220 | { |
da7a735e PZ |
5221 | if (!p->se.on_rq) |
5222 | return; | |
5223 | ||
cb469845 SR |
5224 | /* |
5225 | * Reschedule if we are currently running on this runqueue and | |
5226 | * our priority decreased, or if we are not currently running on | |
5227 | * this runqueue and our priority is higher than the current's | |
5228 | */ | |
da7a735e | 5229 | if (rq->curr == p) { |
cb469845 SR |
5230 | if (p->prio > oldprio) |
5231 | resched_task(rq->curr); | |
5232 | } else | |
15afe09b | 5233 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5234 | } |
5235 | ||
da7a735e PZ |
5236 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
5237 | { | |
5238 | struct sched_entity *se = &p->se; | |
5239 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5240 | ||
5241 | /* | |
5242 | * Ensure the task's vruntime is normalized, so that when its | |
5243 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
5244 | * do the right thing. | |
5245 | * | |
5246 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
5247 | * have normalized the vruntime, if it was !on_rq, then only when | |
5248 | * the task is sleeping will it still have non-normalized vruntime. | |
5249 | */ | |
5250 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
5251 | /* | |
5252 | * Fix up our vruntime so that the current sleep doesn't | |
5253 | * cause 'unlimited' sleep bonus. | |
5254 | */ | |
5255 | place_entity(cfs_rq, se, 0); | |
5256 | se->vruntime -= cfs_rq->min_vruntime; | |
5257 | } | |
5258 | } | |
5259 | ||
cb469845 SR |
5260 | /* |
5261 | * We switched to the sched_fair class. | |
5262 | */ | |
da7a735e | 5263 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 5264 | { |
da7a735e PZ |
5265 | if (!p->se.on_rq) |
5266 | return; | |
5267 | ||
cb469845 SR |
5268 | /* |
5269 | * We were most likely switched from sched_rt, so | |
5270 | * kick off the schedule if running, otherwise just see | |
5271 | * if we can still preempt the current task. | |
5272 | */ | |
da7a735e | 5273 | if (rq->curr == p) |
cb469845 SR |
5274 | resched_task(rq->curr); |
5275 | else | |
15afe09b | 5276 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5277 | } |
5278 | ||
83b699ed SV |
5279 | /* Account for a task changing its policy or group. |
5280 | * | |
5281 | * This routine is mostly called to set cfs_rq->curr field when a task | |
5282 | * migrates between groups/classes. | |
5283 | */ | |
5284 | static void set_curr_task_fair(struct rq *rq) | |
5285 | { | |
5286 | struct sched_entity *se = &rq->curr->se; | |
5287 | ||
ec12cb7f PT |
5288 | for_each_sched_entity(se) { |
5289 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5290 | ||
5291 | set_next_entity(cfs_rq, se); | |
5292 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
5293 | account_cfs_rq_runtime(cfs_rq, 0); | |
5294 | } | |
83b699ed SV |
5295 | } |
5296 | ||
029632fb PZ |
5297 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
5298 | { | |
5299 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
5300 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
5301 | #ifndef CONFIG_64BIT | |
5302 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
5303 | #endif | |
5304 | } | |
5305 | ||
810b3817 | 5306 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5307 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 5308 | { |
b2b5ce02 PZ |
5309 | /* |
5310 | * If the task was not on the rq at the time of this cgroup movement | |
5311 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
5312 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
5313 | * bonus in place_entity()). | |
5314 | * | |
5315 | * If it was on the rq, we've just 'preempted' it, which does convert | |
5316 | * ->vruntime to a relative base. | |
5317 | * | |
5318 | * Make sure both cases convert their relative position when migrating | |
5319 | * to another cgroup's rq. This does somewhat interfere with the | |
5320 | * fair sleeper stuff for the first placement, but who cares. | |
5321 | */ | |
7ceff013 DN |
5322 | /* |
5323 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
5324 | * But there are some cases where it has already been normalized: | |
5325 | * | |
5326 | * - Moving a forked child which is waiting for being woken up by | |
5327 | * wake_up_new_task(). | |
62af3783 DN |
5328 | * - Moving a task which has been woken up by try_to_wake_up() and |
5329 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
5330 | * |
5331 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
5332 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
5333 | */ | |
62af3783 | 5334 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
5335 | on_rq = 1; |
5336 | ||
b2b5ce02 PZ |
5337 | if (!on_rq) |
5338 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
5339 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 5340 | if (!on_rq) |
b2b5ce02 | 5341 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 | 5342 | } |
029632fb PZ |
5343 | |
5344 | void free_fair_sched_group(struct task_group *tg) | |
5345 | { | |
5346 | int i; | |
5347 | ||
5348 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5349 | ||
5350 | for_each_possible_cpu(i) { | |
5351 | if (tg->cfs_rq) | |
5352 | kfree(tg->cfs_rq[i]); | |
5353 | if (tg->se) | |
5354 | kfree(tg->se[i]); | |
5355 | } | |
5356 | ||
5357 | kfree(tg->cfs_rq); | |
5358 | kfree(tg->se); | |
5359 | } | |
5360 | ||
5361 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5362 | { | |
5363 | struct cfs_rq *cfs_rq; | |
5364 | struct sched_entity *se; | |
5365 | int i; | |
5366 | ||
5367 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
5368 | if (!tg->cfs_rq) | |
5369 | goto err; | |
5370 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
5371 | if (!tg->se) | |
5372 | goto err; | |
5373 | ||
5374 | tg->shares = NICE_0_LOAD; | |
5375 | ||
5376 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5377 | ||
5378 | for_each_possible_cpu(i) { | |
5379 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
5380 | GFP_KERNEL, cpu_to_node(i)); | |
5381 | if (!cfs_rq) | |
5382 | goto err; | |
5383 | ||
5384 | se = kzalloc_node(sizeof(struct sched_entity), | |
5385 | GFP_KERNEL, cpu_to_node(i)); | |
5386 | if (!se) | |
5387 | goto err_free_rq; | |
5388 | ||
5389 | init_cfs_rq(cfs_rq); | |
5390 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
5391 | } | |
5392 | ||
5393 | return 1; | |
5394 | ||
5395 | err_free_rq: | |
5396 | kfree(cfs_rq); | |
5397 | err: | |
5398 | return 0; | |
5399 | } | |
5400 | ||
5401 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
5402 | { | |
5403 | struct rq *rq = cpu_rq(cpu); | |
5404 | unsigned long flags; | |
5405 | ||
5406 | /* | |
5407 | * Only empty task groups can be destroyed; so we can speculatively | |
5408 | * check on_list without danger of it being re-added. | |
5409 | */ | |
5410 | if (!tg->cfs_rq[cpu]->on_list) | |
5411 | return; | |
5412 | ||
5413 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5414 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
5415 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5416 | } | |
5417 | ||
5418 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
5419 | struct sched_entity *se, int cpu, | |
5420 | struct sched_entity *parent) | |
5421 | { | |
5422 | struct rq *rq = cpu_rq(cpu); | |
5423 | ||
5424 | cfs_rq->tg = tg; | |
5425 | cfs_rq->rq = rq; | |
5426 | #ifdef CONFIG_SMP | |
5427 | /* allow initial update_cfs_load() to truncate */ | |
5428 | cfs_rq->load_stamp = 1; | |
810b3817 | 5429 | #endif |
029632fb PZ |
5430 | init_cfs_rq_runtime(cfs_rq); |
5431 | ||
5432 | tg->cfs_rq[cpu] = cfs_rq; | |
5433 | tg->se[cpu] = se; | |
5434 | ||
5435 | /* se could be NULL for root_task_group */ | |
5436 | if (!se) | |
5437 | return; | |
5438 | ||
5439 | if (!parent) | |
5440 | se->cfs_rq = &rq->cfs; | |
5441 | else | |
5442 | se->cfs_rq = parent->my_q; | |
5443 | ||
5444 | se->my_q = cfs_rq; | |
5445 | update_load_set(&se->load, 0); | |
5446 | se->parent = parent; | |
5447 | } | |
5448 | ||
5449 | static DEFINE_MUTEX(shares_mutex); | |
5450 | ||
5451 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
5452 | { | |
5453 | int i; | |
5454 | unsigned long flags; | |
5455 | ||
5456 | /* | |
5457 | * We can't change the weight of the root cgroup. | |
5458 | */ | |
5459 | if (!tg->se[0]) | |
5460 | return -EINVAL; | |
5461 | ||
5462 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
5463 | ||
5464 | mutex_lock(&shares_mutex); | |
5465 | if (tg->shares == shares) | |
5466 | goto done; | |
5467 | ||
5468 | tg->shares = shares; | |
5469 | for_each_possible_cpu(i) { | |
5470 | struct rq *rq = cpu_rq(i); | |
5471 | struct sched_entity *se; | |
5472 | ||
5473 | se = tg->se[i]; | |
5474 | /* Propagate contribution to hierarchy */ | |
5475 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5476 | for_each_sched_entity(se) | |
5477 | update_cfs_shares(group_cfs_rq(se)); | |
5478 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5479 | } | |
5480 | ||
5481 | done: | |
5482 | mutex_unlock(&shares_mutex); | |
5483 | return 0; | |
5484 | } | |
5485 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
5486 | ||
5487 | void free_fair_sched_group(struct task_group *tg) { } | |
5488 | ||
5489 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5490 | { | |
5491 | return 1; | |
5492 | } | |
5493 | ||
5494 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
5495 | ||
5496 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
5497 | ||
810b3817 | 5498 | |
6d686f45 | 5499 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
5500 | { |
5501 | struct sched_entity *se = &task->se; | |
0d721cea PW |
5502 | unsigned int rr_interval = 0; |
5503 | ||
5504 | /* | |
5505 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
5506 | * idle runqueue: | |
5507 | */ | |
0d721cea PW |
5508 | if (rq->cfs.load.weight) |
5509 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
5510 | |
5511 | return rr_interval; | |
5512 | } | |
5513 | ||
bf0f6f24 IM |
5514 | /* |
5515 | * All the scheduling class methods: | |
5516 | */ | |
029632fb | 5517 | const struct sched_class fair_sched_class = { |
5522d5d5 | 5518 | .next = &idle_sched_class, |
bf0f6f24 IM |
5519 | .enqueue_task = enqueue_task_fair, |
5520 | .dequeue_task = dequeue_task_fair, | |
5521 | .yield_task = yield_task_fair, | |
d95f4122 | 5522 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 5523 | |
2e09bf55 | 5524 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
5525 | |
5526 | .pick_next_task = pick_next_task_fair, | |
5527 | .put_prev_task = put_prev_task_fair, | |
5528 | ||
681f3e68 | 5529 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
5530 | .select_task_rq = select_task_rq_fair, |
5531 | ||
0bcdcf28 CE |
5532 | .rq_online = rq_online_fair, |
5533 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
5534 | |
5535 | .task_waking = task_waking_fair, | |
681f3e68 | 5536 | #endif |
bf0f6f24 | 5537 | |
83b699ed | 5538 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 5539 | .task_tick = task_tick_fair, |
cd29fe6f | 5540 | .task_fork = task_fork_fair, |
cb469845 SR |
5541 | |
5542 | .prio_changed = prio_changed_fair, | |
da7a735e | 5543 | .switched_from = switched_from_fair, |
cb469845 | 5544 | .switched_to = switched_to_fair, |
810b3817 | 5545 | |
0d721cea PW |
5546 | .get_rr_interval = get_rr_interval_fair, |
5547 | ||
810b3817 | 5548 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5549 | .task_move_group = task_move_group_fair, |
810b3817 | 5550 | #endif |
bf0f6f24 IM |
5551 | }; |
5552 | ||
5553 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 5554 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 5555 | { |
bf0f6f24 IM |
5556 | struct cfs_rq *cfs_rq; |
5557 | ||
5973e5b9 | 5558 | rcu_read_lock(); |
c3b64f1e | 5559 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 5560 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 5561 | rcu_read_unlock(); |
bf0f6f24 IM |
5562 | } |
5563 | #endif | |
029632fb PZ |
5564 | |
5565 | __init void init_sched_fair_class(void) | |
5566 | { | |
5567 | #ifdef CONFIG_SMP | |
5568 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
5569 | ||
5570 | #ifdef CONFIG_NO_HZ | |
554cecaf | 5571 | nohz.next_balance = jiffies; |
029632fb | 5572 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 5573 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
5574 | #endif |
5575 | #endif /* SMP */ | |
5576 | ||
5577 | } |