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