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> | |
29 | ||
30 | #include <trace/events/sched.h> | |
31 | ||
32 | #include "sched.h" | |
9745512c | 33 | |
bf0f6f24 | 34 | /* |
21805085 | 35 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 36 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 37 | * |
21805085 | 38 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
39 | * 'timeslice length' - timeslices in CFS are of variable length |
40 | * and have no persistent notion like in traditional, time-slice | |
41 | * based scheduling concepts. | |
bf0f6f24 | 42 | * |
d274a4ce IM |
43 | * (to see the precise effective timeslice length of your workload, |
44 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 45 | */ |
21406928 MG |
46 | unsigned int sysctl_sched_latency = 6000000ULL; |
47 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 48 | |
1983a922 CE |
49 | /* |
50 | * The initial- and re-scaling of tunables is configurable | |
51 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
52 | * | |
53 | * Options are: | |
54 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
55 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
56 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
57 | */ | |
58 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
59 | = SCHED_TUNABLESCALING_LOG; | |
60 | ||
2bd8e6d4 | 61 | /* |
b2be5e96 | 62 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 63 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 64 | */ |
0bf377bb IM |
65 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
66 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
67 | |
68 | /* | |
b2be5e96 PZ |
69 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
70 | */ | |
0bf377bb | 71 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
72 | |
73 | /* | |
2bba22c5 | 74 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 75 | * parent will (try to) run first. |
21805085 | 76 | */ |
2bba22c5 | 77 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 78 | |
bf0f6f24 IM |
79 | /* |
80 | * SCHED_OTHER wake-up granularity. | |
172e082a | 81 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
82 | * |
83 | * This option delays the preemption effects of decoupled workloads | |
84 | * and reduces their over-scheduling. Synchronous workloads will still | |
85 | * have immediate wakeup/sleep latencies. | |
86 | */ | |
172e082a | 87 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 88 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 89 | |
da84d961 IM |
90 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
91 | ||
a7a4f8a7 PT |
92 | /* |
93 | * The exponential sliding window over which load is averaged for shares | |
94 | * distribution. | |
95 | * (default: 10msec) | |
96 | */ | |
97 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
98 | ||
ec12cb7f PT |
99 | #ifdef CONFIG_CFS_BANDWIDTH |
100 | /* | |
101 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
102 | * each time a cfs_rq requests quota. | |
103 | * | |
104 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
105 | * to consumption or the quota being specified to be smaller than the slice) | |
106 | * we will always only issue the remaining available time. | |
107 | * | |
108 | * default: 5 msec, units: microseconds | |
109 | */ | |
110 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
111 | #endif | |
112 | ||
029632fb PZ |
113 | /* |
114 | * Increase the granularity value when there are more CPUs, | |
115 | * because with more CPUs the 'effective latency' as visible | |
116 | * to users decreases. But the relationship is not linear, | |
117 | * so pick a second-best guess by going with the log2 of the | |
118 | * number of CPUs. | |
119 | * | |
120 | * This idea comes from the SD scheduler of Con Kolivas: | |
121 | */ | |
122 | static int get_update_sysctl_factor(void) | |
123 | { | |
124 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
125 | unsigned int factor; | |
126 | ||
127 | switch (sysctl_sched_tunable_scaling) { | |
128 | case SCHED_TUNABLESCALING_NONE: | |
129 | factor = 1; | |
130 | break; | |
131 | case SCHED_TUNABLESCALING_LINEAR: | |
132 | factor = cpus; | |
133 | break; | |
134 | case SCHED_TUNABLESCALING_LOG: | |
135 | default: | |
136 | factor = 1 + ilog2(cpus); | |
137 | break; | |
138 | } | |
139 | ||
140 | return factor; | |
141 | } | |
142 | ||
143 | static void update_sysctl(void) | |
144 | { | |
145 | unsigned int factor = get_update_sysctl_factor(); | |
146 | ||
147 | #define SET_SYSCTL(name) \ | |
148 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
149 | SET_SYSCTL(sched_min_granularity); | |
150 | SET_SYSCTL(sched_latency); | |
151 | SET_SYSCTL(sched_wakeup_granularity); | |
152 | #undef SET_SYSCTL | |
153 | } | |
154 | ||
155 | void sched_init_granularity(void) | |
156 | { | |
157 | update_sysctl(); | |
158 | } | |
159 | ||
160 | #if BITS_PER_LONG == 32 | |
161 | # define WMULT_CONST (~0UL) | |
162 | #else | |
163 | # define WMULT_CONST (1UL << 32) | |
164 | #endif | |
165 | ||
166 | #define WMULT_SHIFT 32 | |
167 | ||
168 | /* | |
169 | * Shift right and round: | |
170 | */ | |
171 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
172 | ||
173 | /* | |
174 | * delta *= weight / lw | |
175 | */ | |
176 | static unsigned long | |
177 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
178 | struct load_weight *lw) | |
179 | { | |
180 | u64 tmp; | |
181 | ||
182 | /* | |
183 | * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched | |
184 | * entities since MIN_SHARES = 2. Treat weight as 1 if less than | |
185 | * 2^SCHED_LOAD_RESOLUTION. | |
186 | */ | |
187 | if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) | |
188 | tmp = (u64)delta_exec * scale_load_down(weight); | |
189 | else | |
190 | tmp = (u64)delta_exec; | |
191 | ||
192 | if (!lw->inv_weight) { | |
193 | unsigned long w = scale_load_down(lw->weight); | |
194 | ||
195 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
196 | lw->inv_weight = 1; | |
197 | else if (unlikely(!w)) | |
198 | lw->inv_weight = WMULT_CONST; | |
199 | else | |
200 | lw->inv_weight = WMULT_CONST / w; | |
201 | } | |
202 | ||
203 | /* | |
204 | * Check whether we'd overflow the 64-bit multiplication: | |
205 | */ | |
206 | if (unlikely(tmp > WMULT_CONST)) | |
207 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
208 | WMULT_SHIFT/2); | |
209 | else | |
210 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
211 | ||
212 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
213 | } | |
214 | ||
215 | ||
216 | const struct sched_class fair_sched_class; | |
a4c2f00f | 217 | |
bf0f6f24 IM |
218 | /************************************************************** |
219 | * CFS operations on generic schedulable entities: | |
220 | */ | |
221 | ||
62160e3f | 222 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 223 | |
62160e3f | 224 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
225 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
226 | { | |
62160e3f | 227 | return cfs_rq->rq; |
bf0f6f24 IM |
228 | } |
229 | ||
62160e3f IM |
230 | /* An entity is a task if it doesn't "own" a runqueue */ |
231 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 232 | |
8f48894f PZ |
233 | static inline struct task_struct *task_of(struct sched_entity *se) |
234 | { | |
235 | #ifdef CONFIG_SCHED_DEBUG | |
236 | WARN_ON_ONCE(!entity_is_task(se)); | |
237 | #endif | |
238 | return container_of(se, struct task_struct, se); | |
239 | } | |
240 | ||
b758149c PZ |
241 | /* Walk up scheduling entities hierarchy */ |
242 | #define for_each_sched_entity(se) \ | |
243 | for (; se; se = se->parent) | |
244 | ||
245 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
246 | { | |
247 | return p->se.cfs_rq; | |
248 | } | |
249 | ||
250 | /* runqueue on which this entity is (to be) queued */ | |
251 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
252 | { | |
253 | return se->cfs_rq; | |
254 | } | |
255 | ||
256 | /* runqueue "owned" by this group */ | |
257 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
258 | { | |
259 | return grp->my_q; | |
260 | } | |
261 | ||
9ee474f5 PT |
262 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq); |
263 | ||
3d4b47b4 PZ |
264 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
265 | { | |
266 | if (!cfs_rq->on_list) { | |
67e86250 PT |
267 | /* |
268 | * Ensure we either appear before our parent (if already | |
269 | * enqueued) or force our parent to appear after us when it is | |
270 | * enqueued. The fact that we always enqueue bottom-up | |
271 | * reduces this to two cases. | |
272 | */ | |
273 | if (cfs_rq->tg->parent && | |
274 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
275 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
276 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
277 | } else { | |
278 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 279 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 280 | } |
3d4b47b4 PZ |
281 | |
282 | cfs_rq->on_list = 1; | |
9ee474f5 PT |
283 | /* We should have no load, but we need to update last_decay. */ |
284 | update_cfs_rq_blocked_load(cfs_rq); | |
3d4b47b4 PZ |
285 | } |
286 | } | |
287 | ||
288 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
289 | { | |
290 | if (cfs_rq->on_list) { | |
291 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
292 | cfs_rq->on_list = 0; | |
293 | } | |
294 | } | |
295 | ||
b758149c PZ |
296 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
297 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
298 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
299 | ||
300 | /* Do the two (enqueued) entities belong to the same group ? */ | |
301 | static inline int | |
302 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
303 | { | |
304 | if (se->cfs_rq == pse->cfs_rq) | |
305 | return 1; | |
306 | ||
307 | return 0; | |
308 | } | |
309 | ||
310 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
311 | { | |
312 | return se->parent; | |
313 | } | |
314 | ||
464b7527 PZ |
315 | /* return depth at which a sched entity is present in the hierarchy */ |
316 | static inline int depth_se(struct sched_entity *se) | |
317 | { | |
318 | int depth = 0; | |
319 | ||
320 | for_each_sched_entity(se) | |
321 | depth++; | |
322 | ||
323 | return depth; | |
324 | } | |
325 | ||
326 | static void | |
327 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
328 | { | |
329 | int se_depth, pse_depth; | |
330 | ||
331 | /* | |
332 | * preemption test can be made between sibling entities who are in the | |
333 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
334 | * both tasks until we find their ancestors who are siblings of common | |
335 | * parent. | |
336 | */ | |
337 | ||
338 | /* First walk up until both entities are at same depth */ | |
339 | se_depth = depth_se(*se); | |
340 | pse_depth = depth_se(*pse); | |
341 | ||
342 | while (se_depth > pse_depth) { | |
343 | se_depth--; | |
344 | *se = parent_entity(*se); | |
345 | } | |
346 | ||
347 | while (pse_depth > se_depth) { | |
348 | pse_depth--; | |
349 | *pse = parent_entity(*pse); | |
350 | } | |
351 | ||
352 | while (!is_same_group(*se, *pse)) { | |
353 | *se = parent_entity(*se); | |
354 | *pse = parent_entity(*pse); | |
355 | } | |
356 | } | |
357 | ||
8f48894f PZ |
358 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
359 | ||
360 | static inline struct task_struct *task_of(struct sched_entity *se) | |
361 | { | |
362 | return container_of(se, struct task_struct, se); | |
363 | } | |
bf0f6f24 | 364 | |
62160e3f IM |
365 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
366 | { | |
367 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
368 | } |
369 | ||
370 | #define entity_is_task(se) 1 | |
371 | ||
b758149c PZ |
372 | #define for_each_sched_entity(se) \ |
373 | for (; se; se = NULL) | |
bf0f6f24 | 374 | |
b758149c | 375 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 376 | { |
b758149c | 377 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
378 | } |
379 | ||
b758149c PZ |
380 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
381 | { | |
382 | struct task_struct *p = task_of(se); | |
383 | struct rq *rq = task_rq(p); | |
384 | ||
385 | return &rq->cfs; | |
386 | } | |
387 | ||
388 | /* runqueue "owned" by this group */ | |
389 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
390 | { | |
391 | return NULL; | |
392 | } | |
393 | ||
3d4b47b4 PZ |
394 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
395 | { | |
396 | } | |
397 | ||
398 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
399 | { | |
400 | } | |
401 | ||
b758149c PZ |
402 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
403 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
404 | ||
405 | static inline int | |
406 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
407 | { | |
408 | return 1; | |
409 | } | |
410 | ||
411 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
412 | { | |
413 | return NULL; | |
414 | } | |
415 | ||
464b7527 PZ |
416 | static inline void |
417 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
418 | { | |
419 | } | |
420 | ||
b758149c PZ |
421 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
422 | ||
6c16a6dc PZ |
423 | static __always_inline |
424 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
425 | |
426 | /************************************************************** | |
427 | * Scheduling class tree data structure manipulation methods: | |
428 | */ | |
429 | ||
0702e3eb | 430 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 431 | { |
368059a9 PZ |
432 | s64 delta = (s64)(vruntime - min_vruntime); |
433 | if (delta > 0) | |
02e0431a PZ |
434 | min_vruntime = vruntime; |
435 | ||
436 | return min_vruntime; | |
437 | } | |
438 | ||
0702e3eb | 439 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
440 | { |
441 | s64 delta = (s64)(vruntime - min_vruntime); | |
442 | if (delta < 0) | |
443 | min_vruntime = vruntime; | |
444 | ||
445 | return min_vruntime; | |
446 | } | |
447 | ||
54fdc581 FC |
448 | static inline int entity_before(struct sched_entity *a, |
449 | struct sched_entity *b) | |
450 | { | |
451 | return (s64)(a->vruntime - b->vruntime) < 0; | |
452 | } | |
453 | ||
1af5f730 PZ |
454 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
455 | { | |
456 | u64 vruntime = cfs_rq->min_vruntime; | |
457 | ||
458 | if (cfs_rq->curr) | |
459 | vruntime = cfs_rq->curr->vruntime; | |
460 | ||
461 | if (cfs_rq->rb_leftmost) { | |
462 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
463 | struct sched_entity, | |
464 | run_node); | |
465 | ||
e17036da | 466 | if (!cfs_rq->curr) |
1af5f730 PZ |
467 | vruntime = se->vruntime; |
468 | else | |
469 | vruntime = min_vruntime(vruntime, se->vruntime); | |
470 | } | |
471 | ||
472 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
3fe1698b PZ |
473 | #ifndef CONFIG_64BIT |
474 | smp_wmb(); | |
475 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
476 | #endif | |
1af5f730 PZ |
477 | } |
478 | ||
bf0f6f24 IM |
479 | /* |
480 | * Enqueue an entity into the rb-tree: | |
481 | */ | |
0702e3eb | 482 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
483 | { |
484 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
485 | struct rb_node *parent = NULL; | |
486 | struct sched_entity *entry; | |
bf0f6f24 IM |
487 | int leftmost = 1; |
488 | ||
489 | /* | |
490 | * Find the right place in the rbtree: | |
491 | */ | |
492 | while (*link) { | |
493 | parent = *link; | |
494 | entry = rb_entry(parent, struct sched_entity, run_node); | |
495 | /* | |
496 | * We dont care about collisions. Nodes with | |
497 | * the same key stay together. | |
498 | */ | |
2bd2d6f2 | 499 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
500 | link = &parent->rb_left; |
501 | } else { | |
502 | link = &parent->rb_right; | |
503 | leftmost = 0; | |
504 | } | |
505 | } | |
506 | ||
507 | /* | |
508 | * Maintain a cache of leftmost tree entries (it is frequently | |
509 | * used): | |
510 | */ | |
1af5f730 | 511 | if (leftmost) |
57cb499d | 512 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
513 | |
514 | rb_link_node(&se->run_node, parent, link); | |
515 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
516 | } |
517 | ||
0702e3eb | 518 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 519 | { |
3fe69747 PZ |
520 | if (cfs_rq->rb_leftmost == &se->run_node) { |
521 | struct rb_node *next_node; | |
3fe69747 PZ |
522 | |
523 | next_node = rb_next(&se->run_node); | |
524 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 525 | } |
e9acbff6 | 526 | |
bf0f6f24 | 527 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
528 | } |
529 | ||
029632fb | 530 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 531 | { |
f4b6755f PZ |
532 | struct rb_node *left = cfs_rq->rb_leftmost; |
533 | ||
534 | if (!left) | |
535 | return NULL; | |
536 | ||
537 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
538 | } |
539 | ||
ac53db59 RR |
540 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
541 | { | |
542 | struct rb_node *next = rb_next(&se->run_node); | |
543 | ||
544 | if (!next) | |
545 | return NULL; | |
546 | ||
547 | return rb_entry(next, struct sched_entity, run_node); | |
548 | } | |
549 | ||
550 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 551 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 552 | { |
7eee3e67 | 553 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 554 | |
70eee74b BS |
555 | if (!last) |
556 | return NULL; | |
7eee3e67 IM |
557 | |
558 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
559 | } |
560 | ||
bf0f6f24 IM |
561 | /************************************************************** |
562 | * Scheduling class statistics methods: | |
563 | */ | |
564 | ||
acb4a848 | 565 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 566 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
567 | loff_t *ppos) |
568 | { | |
8d65af78 | 569 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 570 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
571 | |
572 | if (ret || !write) | |
573 | return ret; | |
574 | ||
575 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
576 | sysctl_sched_min_granularity); | |
577 | ||
acb4a848 CE |
578 | #define WRT_SYSCTL(name) \ |
579 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
580 | WRT_SYSCTL(sched_min_granularity); | |
581 | WRT_SYSCTL(sched_latency); | |
582 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
583 | #undef WRT_SYSCTL |
584 | ||
b2be5e96 PZ |
585 | return 0; |
586 | } | |
587 | #endif | |
647e7cac | 588 | |
a7be37ac | 589 | /* |
f9c0b095 | 590 | * delta /= w |
a7be37ac PZ |
591 | */ |
592 | static inline unsigned long | |
593 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
594 | { | |
f9c0b095 PZ |
595 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
596 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
597 | |
598 | return delta; | |
599 | } | |
600 | ||
647e7cac IM |
601 | /* |
602 | * The idea is to set a period in which each task runs once. | |
603 | * | |
532b1858 | 604 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
605 | * this period because otherwise the slices get too small. |
606 | * | |
607 | * p = (nr <= nl) ? l : l*nr/nl | |
608 | */ | |
4d78e7b6 PZ |
609 | static u64 __sched_period(unsigned long nr_running) |
610 | { | |
611 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 612 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
613 | |
614 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 615 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 616 | period *= nr_running; |
4d78e7b6 PZ |
617 | } |
618 | ||
619 | return period; | |
620 | } | |
621 | ||
647e7cac IM |
622 | /* |
623 | * We calculate the wall-time slice from the period by taking a part | |
624 | * proportional to the weight. | |
625 | * | |
f9c0b095 | 626 | * s = p*P[w/rw] |
647e7cac | 627 | */ |
6d0f0ebd | 628 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 629 | { |
0a582440 | 630 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 631 | |
0a582440 | 632 | for_each_sched_entity(se) { |
6272d68c | 633 | struct load_weight *load; |
3104bf03 | 634 | struct load_weight lw; |
6272d68c LM |
635 | |
636 | cfs_rq = cfs_rq_of(se); | |
637 | load = &cfs_rq->load; | |
f9c0b095 | 638 | |
0a582440 | 639 | if (unlikely(!se->on_rq)) { |
3104bf03 | 640 | lw = cfs_rq->load; |
0a582440 MG |
641 | |
642 | update_load_add(&lw, se->load.weight); | |
643 | load = &lw; | |
644 | } | |
645 | slice = calc_delta_mine(slice, se->load.weight, load); | |
646 | } | |
647 | return slice; | |
bf0f6f24 IM |
648 | } |
649 | ||
647e7cac | 650 | /* |
ac884dec | 651 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 652 | * |
f9c0b095 | 653 | * vs = s/w |
647e7cac | 654 | */ |
f9c0b095 | 655 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 656 | { |
f9c0b095 | 657 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
658 | } |
659 | ||
d6b55918 | 660 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); |
6d5ab293 | 661 | static void update_cfs_shares(struct cfs_rq *cfs_rq); |
3b3d190e | 662 | |
bf0f6f24 IM |
663 | /* |
664 | * Update the current task's runtime statistics. Skip current tasks that | |
665 | * are not in our scheduling class. | |
666 | */ | |
667 | static inline void | |
8ebc91d9 IM |
668 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
669 | unsigned long delta_exec) | |
bf0f6f24 | 670 | { |
bbdba7c0 | 671 | unsigned long delta_exec_weighted; |
bf0f6f24 | 672 | |
41acab88 LDM |
673 | schedstat_set(curr->statistics.exec_max, |
674 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
675 | |
676 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 677 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 678 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 679 | |
e9acbff6 | 680 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 681 | update_min_vruntime(cfs_rq); |
3b3d190e | 682 | |
70caf8a6 | 683 | #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED |
3b3d190e | 684 | cfs_rq->load_unacc_exec_time += delta_exec; |
3b3d190e | 685 | #endif |
bf0f6f24 IM |
686 | } |
687 | ||
b7cc0896 | 688 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 689 | { |
429d43bc | 690 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 691 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
692 | unsigned long delta_exec; |
693 | ||
694 | if (unlikely(!curr)) | |
695 | return; | |
696 | ||
697 | /* | |
698 | * Get the amount of time the current task was running | |
699 | * since the last time we changed load (this cannot | |
700 | * overflow on 32 bits): | |
701 | */ | |
8ebc91d9 | 702 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
703 | if (!delta_exec) |
704 | return; | |
bf0f6f24 | 705 | |
8ebc91d9 IM |
706 | __update_curr(cfs_rq, curr, delta_exec); |
707 | curr->exec_start = now; | |
d842de87 SV |
708 | |
709 | if (entity_is_task(curr)) { | |
710 | struct task_struct *curtask = task_of(curr); | |
711 | ||
f977bb49 | 712 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 713 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 714 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 715 | } |
ec12cb7f PT |
716 | |
717 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
718 | } |
719 | ||
720 | static inline void | |
5870db5b | 721 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 722 | { |
41acab88 | 723 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
724 | } |
725 | ||
bf0f6f24 IM |
726 | /* |
727 | * Task is being enqueued - update stats: | |
728 | */ | |
d2417e5a | 729 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 730 | { |
bf0f6f24 IM |
731 | /* |
732 | * Are we enqueueing a waiting task? (for current tasks | |
733 | * a dequeue/enqueue event is a NOP) | |
734 | */ | |
429d43bc | 735 | if (se != cfs_rq->curr) |
5870db5b | 736 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
737 | } |
738 | ||
bf0f6f24 | 739 | static void |
9ef0a961 | 740 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 741 | { |
41acab88 LDM |
742 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
743 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
744 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
745 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
746 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
747 | #ifdef CONFIG_SCHEDSTATS |
748 | if (entity_is_task(se)) { | |
749 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 750 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
751 | } |
752 | #endif | |
41acab88 | 753 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
754 | } |
755 | ||
756 | static inline void | |
19b6a2e3 | 757 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 758 | { |
bf0f6f24 IM |
759 | /* |
760 | * Mark the end of the wait period if dequeueing a | |
761 | * waiting task: | |
762 | */ | |
429d43bc | 763 | if (se != cfs_rq->curr) |
9ef0a961 | 764 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
765 | } |
766 | ||
767 | /* | |
768 | * We are picking a new current task - update its stats: | |
769 | */ | |
770 | static inline void | |
79303e9e | 771 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
772 | { |
773 | /* | |
774 | * We are starting a new run period: | |
775 | */ | |
305e6835 | 776 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
777 | } |
778 | ||
bf0f6f24 IM |
779 | /************************************************** |
780 | * Scheduling class queueing methods: | |
781 | */ | |
782 | ||
30cfdcfc DA |
783 | static void |
784 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
785 | { | |
786 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 787 | if (!parent_entity(se)) |
029632fb | 788 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
789 | #ifdef CONFIG_SMP |
790 | if (entity_is_task(se)) | |
eb95308e | 791 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 792 | #endif |
30cfdcfc | 793 | cfs_rq->nr_running++; |
30cfdcfc DA |
794 | } |
795 | ||
796 | static void | |
797 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
798 | { | |
799 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 800 | if (!parent_entity(se)) |
029632fb | 801 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 802 | if (entity_is_task(se)) |
b87f1724 | 803 | list_del_init(&se->group_node); |
30cfdcfc | 804 | cfs_rq->nr_running--; |
30cfdcfc DA |
805 | } |
806 | ||
3ff6dcac | 807 | #ifdef CONFIG_FAIR_GROUP_SCHED |
64660c86 PT |
808 | /* we need this in update_cfs_load and load-balance functions below */ |
809 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); | |
3ff6dcac | 810 | # ifdef CONFIG_SMP |
d6b55918 PT |
811 | static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, |
812 | int global_update) | |
813 | { | |
814 | struct task_group *tg = cfs_rq->tg; | |
815 | long load_avg; | |
816 | ||
817 | load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); | |
818 | load_avg -= cfs_rq->load_contribution; | |
819 | ||
820 | if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { | |
821 | atomic_add(load_avg, &tg->load_weight); | |
822 | cfs_rq->load_contribution += load_avg; | |
823 | } | |
824 | } | |
825 | ||
826 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
2069dd75 | 827 | { |
a7a4f8a7 | 828 | u64 period = sysctl_sched_shares_window; |
2069dd75 | 829 | u64 now, delta; |
e33078ba | 830 | unsigned long load = cfs_rq->load.weight; |
2069dd75 | 831 | |
64660c86 | 832 | if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq)) |
2069dd75 PZ |
833 | return; |
834 | ||
05ca62c6 | 835 | now = rq_of(cfs_rq)->clock_task; |
2069dd75 PZ |
836 | delta = now - cfs_rq->load_stamp; |
837 | ||
e33078ba PT |
838 | /* truncate load history at 4 idle periods */ |
839 | if (cfs_rq->load_stamp > cfs_rq->load_last && | |
840 | now - cfs_rq->load_last > 4 * period) { | |
841 | cfs_rq->load_period = 0; | |
842 | cfs_rq->load_avg = 0; | |
f07333bf | 843 | delta = period - 1; |
e33078ba PT |
844 | } |
845 | ||
2069dd75 | 846 | cfs_rq->load_stamp = now; |
3b3d190e | 847 | cfs_rq->load_unacc_exec_time = 0; |
2069dd75 | 848 | cfs_rq->load_period += delta; |
e33078ba PT |
849 | if (load) { |
850 | cfs_rq->load_last = now; | |
851 | cfs_rq->load_avg += delta * load; | |
852 | } | |
2069dd75 | 853 | |
d6b55918 PT |
854 | /* consider updating load contribution on each fold or truncate */ |
855 | if (global_update || cfs_rq->load_period > period | |
856 | || !cfs_rq->load_period) | |
857 | update_cfs_rq_load_contribution(cfs_rq, global_update); | |
858 | ||
2069dd75 PZ |
859 | while (cfs_rq->load_period > period) { |
860 | /* | |
861 | * Inline assembly required to prevent the compiler | |
862 | * optimising this loop into a divmod call. | |
863 | * See __iter_div_u64_rem() for another example of this. | |
864 | */ | |
865 | asm("" : "+rm" (cfs_rq->load_period)); | |
866 | cfs_rq->load_period /= 2; | |
867 | cfs_rq->load_avg /= 2; | |
868 | } | |
3d4b47b4 | 869 | |
e33078ba PT |
870 | if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) |
871 | list_del_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
872 | } |
873 | ||
cf5f0acf PZ |
874 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
875 | { | |
876 | long tg_weight; | |
877 | ||
878 | /* | |
879 | * Use this CPU's actual weight instead of the last load_contribution | |
880 | * to gain a more accurate current total weight. See | |
881 | * update_cfs_rq_load_contribution(). | |
882 | */ | |
883 | tg_weight = atomic_read(&tg->load_weight); | |
884 | tg_weight -= cfs_rq->load_contribution; | |
885 | tg_weight += cfs_rq->load.weight; | |
886 | ||
887 | return tg_weight; | |
888 | } | |
889 | ||
6d5ab293 | 890 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 891 | { |
cf5f0acf | 892 | long tg_weight, load, shares; |
3ff6dcac | 893 | |
cf5f0acf | 894 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 895 | load = cfs_rq->load.weight; |
3ff6dcac | 896 | |
3ff6dcac | 897 | shares = (tg->shares * load); |
cf5f0acf PZ |
898 | if (tg_weight) |
899 | shares /= tg_weight; | |
3ff6dcac YZ |
900 | |
901 | if (shares < MIN_SHARES) | |
902 | shares = MIN_SHARES; | |
903 | if (shares > tg->shares) | |
904 | shares = tg->shares; | |
905 | ||
906 | return shares; | |
907 | } | |
908 | ||
909 | static void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
910 | { | |
911 | if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { | |
912 | update_cfs_load(cfs_rq, 0); | |
6d5ab293 | 913 | update_cfs_shares(cfs_rq); |
3ff6dcac YZ |
914 | } |
915 | } | |
916 | # else /* CONFIG_SMP */ | |
917 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) | |
918 | { | |
919 | } | |
920 | ||
6d5ab293 | 921 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
922 | { |
923 | return tg->shares; | |
924 | } | |
925 | ||
926 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
927 | { | |
928 | } | |
929 | # endif /* CONFIG_SMP */ | |
2069dd75 PZ |
930 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
931 | unsigned long weight) | |
932 | { | |
19e5eebb PT |
933 | if (se->on_rq) { |
934 | /* commit outstanding execution time */ | |
935 | if (cfs_rq->curr == se) | |
936 | update_curr(cfs_rq); | |
2069dd75 | 937 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 938 | } |
2069dd75 PZ |
939 | |
940 | update_load_set(&se->load, weight); | |
941 | ||
942 | if (se->on_rq) | |
943 | account_entity_enqueue(cfs_rq, se); | |
944 | } | |
945 | ||
6d5ab293 | 946 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
947 | { |
948 | struct task_group *tg; | |
949 | struct sched_entity *se; | |
3ff6dcac | 950 | long shares; |
2069dd75 | 951 | |
2069dd75 PZ |
952 | tg = cfs_rq->tg; |
953 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 954 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 955 | return; |
3ff6dcac YZ |
956 | #ifndef CONFIG_SMP |
957 | if (likely(se->load.weight == tg->shares)) | |
958 | return; | |
959 | #endif | |
6d5ab293 | 960 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
961 | |
962 | reweight_entity(cfs_rq_of(se), se, shares); | |
963 | } | |
964 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
d6b55918 | 965 | static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) |
2069dd75 PZ |
966 | { |
967 | } | |
968 | ||
6d5ab293 | 969 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
970 | { |
971 | } | |
43365bd7 PT |
972 | |
973 | static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) | |
974 | { | |
975 | } | |
2069dd75 PZ |
976 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
977 | ||
9d85f21c PT |
978 | #ifdef CONFIG_SMP |
979 | /* | |
980 | * Approximate: | |
981 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
982 | */ | |
983 | static __always_inline u64 decay_load(u64 val, u64 n) | |
984 | { | |
985 | for (; n && val; n--) { | |
986 | val *= 4008; | |
987 | val >>= 12; | |
988 | } | |
989 | ||
990 | return val; | |
991 | } | |
992 | ||
993 | /* | |
994 | * We can represent the historical contribution to runnable average as the | |
995 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
996 | * history into segments of approximately 1ms (1024us); label the segment that | |
997 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
998 | * | |
999 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
1000 | * p0 p1 p2 | |
1001 | * (now) (~1ms ago) (~2ms ago) | |
1002 | * | |
1003 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
1004 | * | |
1005 | * We then designate the fractions u_i as our co-efficients, yielding the | |
1006 | * following representation of historical load: | |
1007 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
1008 | * | |
1009 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
1010 | * y^32 = 0.5 | |
1011 | * | |
1012 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
1013 | * approximately half as much as the contribution to load within the last ms | |
1014 | * (u_0). | |
1015 | * | |
1016 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
1017 | * sum again by y is sufficient to update: | |
1018 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1019 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1020 | */ | |
1021 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1022 | struct sched_avg *sa, | |
1023 | int runnable) | |
1024 | { | |
1025 | u64 delta; | |
1026 | int delta_w, decayed = 0; | |
1027 | ||
1028 | delta = now - sa->last_runnable_update; | |
1029 | /* | |
1030 | * This should only happen when time goes backwards, which it | |
1031 | * unfortunately does during sched clock init when we swap over to TSC. | |
1032 | */ | |
1033 | if ((s64)delta < 0) { | |
1034 | sa->last_runnable_update = now; | |
1035 | return 0; | |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1040 | * approximation of 1us and fast to compute. | |
1041 | */ | |
1042 | delta >>= 10; | |
1043 | if (!delta) | |
1044 | return 0; | |
1045 | sa->last_runnable_update = now; | |
1046 | ||
1047 | /* delta_w is the amount already accumulated against our next period */ | |
1048 | delta_w = sa->runnable_avg_period % 1024; | |
1049 | if (delta + delta_w >= 1024) { | |
1050 | /* period roll-over */ | |
1051 | decayed = 1; | |
1052 | ||
1053 | /* | |
1054 | * Now that we know we're crossing a period boundary, figure | |
1055 | * out how much from delta we need to complete the current | |
1056 | * period and accrue it. | |
1057 | */ | |
1058 | delta_w = 1024 - delta_w; | |
1059 | BUG_ON(delta_w > delta); | |
1060 | do { | |
1061 | if (runnable) | |
1062 | sa->runnable_avg_sum += delta_w; | |
1063 | sa->runnable_avg_period += delta_w; | |
1064 | ||
1065 | /* | |
1066 | * Remainder of delta initiates a new period, roll over | |
1067 | * the previous. | |
1068 | */ | |
1069 | sa->runnable_avg_sum = | |
1070 | decay_load(sa->runnable_avg_sum, 1); | |
1071 | sa->runnable_avg_period = | |
1072 | decay_load(sa->runnable_avg_period, 1); | |
1073 | ||
1074 | delta -= delta_w; | |
1075 | /* New period is empty */ | |
1076 | delta_w = 1024; | |
1077 | } while (delta >= 1024); | |
1078 | } | |
1079 | ||
1080 | /* Remainder of delta accrued against u_0` */ | |
1081 | if (runnable) | |
1082 | sa->runnable_avg_sum += delta; | |
1083 | sa->runnable_avg_period += delta; | |
1084 | ||
1085 | return decayed; | |
1086 | } | |
1087 | ||
9ee474f5 PT |
1088 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
1089 | static inline void __synchronize_entity_decay(struct sched_entity *se) | |
1090 | { | |
1091 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1092 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1093 | ||
1094 | decays -= se->avg.decay_count; | |
1095 | if (!decays) | |
1096 | return; | |
1097 | ||
1098 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1099 | se->avg.decay_count = 0; | |
1100 | } | |
1101 | ||
2dac754e PT |
1102 | /* Compute the current contribution to load_avg by se, return any delta */ |
1103 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
1104 | { | |
1105 | long old_contrib = se->avg.load_avg_contrib; | |
1106 | ||
1107 | if (!entity_is_task(se)) | |
1108 | return 0; | |
1109 | ||
1110 | se->avg.load_avg_contrib = div64_u64(se->avg.runnable_avg_sum * | |
1111 | se->load.weight, | |
1112 | se->avg.runnable_avg_period + 1); | |
1113 | ||
1114 | return se->avg.load_avg_contrib - old_contrib; | |
1115 | } | |
1116 | ||
9ee474f5 PT |
1117 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
1118 | long load_contrib) | |
1119 | { | |
1120 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
1121 | cfs_rq->blocked_load_avg -= load_contrib; | |
1122 | else | |
1123 | cfs_rq->blocked_load_avg = 0; | |
1124 | } | |
1125 | ||
9d85f21c | 1126 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
1127 | static inline void update_entity_load_avg(struct sched_entity *se, |
1128 | int update_cfs_rq) | |
9d85f21c | 1129 | { |
2dac754e PT |
1130 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
1131 | long contrib_delta; | |
1132 | ||
1133 | if (!__update_entity_runnable_avg(rq_of(cfs_rq)->clock_task, &se->avg, | |
1134 | se->on_rq)) | |
1135 | return; | |
1136 | ||
1137 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
1138 | |
1139 | if (!update_cfs_rq) | |
1140 | return; | |
1141 | ||
2dac754e PT |
1142 | if (se->on_rq) |
1143 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
1144 | else |
1145 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
1146 | } | |
1147 | ||
1148 | /* | |
1149 | * Decay the load contributed by all blocked children and account this so that | |
1150 | * their contribution may appropriately discounted when they wake up. | |
1151 | */ | |
1152 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq) | |
1153 | { | |
1154 | u64 now = rq_of(cfs_rq)->clock_task >> 20; | |
1155 | u64 decays; | |
1156 | ||
1157 | decays = now - cfs_rq->last_decay; | |
1158 | if (!decays) | |
1159 | return; | |
1160 | ||
1161 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
1162 | decays); | |
1163 | atomic64_add(decays, &cfs_rq->decay_counter); | |
1164 | ||
1165 | cfs_rq->last_decay = now; | |
9d85f21c | 1166 | } |
18bf2805 BS |
1167 | |
1168 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
1169 | { | |
1170 | __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable); | |
1171 | } | |
2dac754e PT |
1172 | |
1173 | /* Add the load generated by se into cfs_rq's child load-average */ | |
1174 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
1175 | struct sched_entity *se, |
1176 | int wakeup) | |
2dac754e | 1177 | { |
9ee474f5 PT |
1178 | /* we track migrations using entity decay_count == 0 */ |
1179 | if (unlikely(!se->avg.decay_count)) { | |
1180 | se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task; | |
1181 | wakeup = 0; | |
1182 | } else { | |
1183 | __synchronize_entity_decay(se); | |
1184 | } | |
1185 | ||
1186 | if (wakeup) | |
1187 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
1188 | ||
1189 | update_entity_load_avg(se, 0); | |
2dac754e | 1190 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
9ee474f5 | 1191 | update_cfs_rq_blocked_load(cfs_rq); |
2dac754e PT |
1192 | } |
1193 | ||
9ee474f5 PT |
1194 | /* |
1195 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
1196 | * transitioning to a blocked state we track its projected decay using | |
1197 | * blocked_load_avg. | |
1198 | */ | |
2dac754e | 1199 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1200 | struct sched_entity *se, |
1201 | int sleep) | |
2dac754e | 1202 | { |
9ee474f5 PT |
1203 | update_entity_load_avg(se, 1); |
1204 | ||
2dac754e | 1205 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
1206 | if (sleep) { |
1207 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
1208 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
1209 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 1210 | } |
9d85f21c | 1211 | #else |
9ee474f5 PT |
1212 | static inline void update_entity_load_avg(struct sched_entity *se, |
1213 | int update_cfs_rq) {} | |
18bf2805 | 1214 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 1215 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1216 | struct sched_entity *se, |
1217 | int wakeup) {} | |
2dac754e | 1218 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1219 | struct sched_entity *se, |
1220 | int sleep) {} | |
1221 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq) {} | |
9d85f21c PT |
1222 | #endif |
1223 | ||
2396af69 | 1224 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1225 | { |
bf0f6f24 | 1226 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1227 | struct task_struct *tsk = NULL; |
1228 | ||
1229 | if (entity_is_task(se)) | |
1230 | tsk = task_of(se); | |
1231 | ||
41acab88 LDM |
1232 | if (se->statistics.sleep_start) { |
1233 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
1234 | |
1235 | if ((s64)delta < 0) | |
1236 | delta = 0; | |
1237 | ||
41acab88 LDM |
1238 | if (unlikely(delta > se->statistics.sleep_max)) |
1239 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1240 | |
8c79a045 | 1241 | se->statistics.sleep_start = 0; |
41acab88 | 1242 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1243 | |
768d0c27 | 1244 | if (tsk) { |
e414314c | 1245 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1246 | trace_sched_stat_sleep(tsk, delta); |
1247 | } | |
bf0f6f24 | 1248 | } |
41acab88 LDM |
1249 | if (se->statistics.block_start) { |
1250 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
1251 | |
1252 | if ((s64)delta < 0) | |
1253 | delta = 0; | |
1254 | ||
41acab88 LDM |
1255 | if (unlikely(delta > se->statistics.block_max)) |
1256 | se->statistics.block_max = delta; | |
bf0f6f24 | 1257 | |
8c79a045 | 1258 | se->statistics.block_start = 0; |
41acab88 | 1259 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1260 | |
e414314c | 1261 | if (tsk) { |
8f0dfc34 | 1262 | if (tsk->in_iowait) { |
41acab88 LDM |
1263 | se->statistics.iowait_sum += delta; |
1264 | se->statistics.iowait_count++; | |
768d0c27 | 1265 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1266 | } |
1267 | ||
b781a602 AV |
1268 | trace_sched_stat_blocked(tsk, delta); |
1269 | ||
e414314c PZ |
1270 | /* |
1271 | * Blocking time is in units of nanosecs, so shift by | |
1272 | * 20 to get a milliseconds-range estimation of the | |
1273 | * amount of time that the task spent sleeping: | |
1274 | */ | |
1275 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1276 | profile_hits(SLEEP_PROFILING, | |
1277 | (void *)get_wchan(tsk), | |
1278 | delta >> 20); | |
1279 | } | |
1280 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1281 | } |
bf0f6f24 IM |
1282 | } |
1283 | #endif | |
1284 | } | |
1285 | ||
ddc97297 PZ |
1286 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1287 | { | |
1288 | #ifdef CONFIG_SCHED_DEBUG | |
1289 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
1290 | ||
1291 | if (d < 0) | |
1292 | d = -d; | |
1293 | ||
1294 | if (d > 3*sysctl_sched_latency) | |
1295 | schedstat_inc(cfs_rq, nr_spread_over); | |
1296 | #endif | |
1297 | } | |
1298 | ||
aeb73b04 PZ |
1299 | static void |
1300 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
1301 | { | |
1af5f730 | 1302 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 1303 | |
2cb8600e PZ |
1304 | /* |
1305 | * The 'current' period is already promised to the current tasks, | |
1306 | * however the extra weight of the new task will slow them down a | |
1307 | * little, place the new task so that it fits in the slot that | |
1308 | * stays open at the end. | |
1309 | */ | |
94dfb5e7 | 1310 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 1311 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 1312 | |
a2e7a7eb | 1313 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 1314 | if (!initial) { |
a2e7a7eb | 1315 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 1316 | |
a2e7a7eb MG |
1317 | /* |
1318 | * Halve their sleep time's effect, to allow | |
1319 | * for a gentler effect of sleepers: | |
1320 | */ | |
1321 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
1322 | thresh >>= 1; | |
51e0304c | 1323 | |
a2e7a7eb | 1324 | vruntime -= thresh; |
aeb73b04 PZ |
1325 | } |
1326 | ||
b5d9d734 MG |
1327 | /* ensure we never gain time by being placed backwards. */ |
1328 | vruntime = max_vruntime(se->vruntime, vruntime); | |
1329 | ||
67e9fb2a | 1330 | se->vruntime = vruntime; |
aeb73b04 PZ |
1331 | } |
1332 | ||
d3d9dc33 PT |
1333 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
1334 | ||
bf0f6f24 | 1335 | static void |
88ec22d3 | 1336 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1337 | { |
88ec22d3 PZ |
1338 | /* |
1339 | * Update the normalized vruntime before updating min_vruntime | |
1340 | * through callig update_curr(). | |
1341 | */ | |
371fd7e7 | 1342 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
1343 | se->vruntime += cfs_rq->min_vruntime; |
1344 | ||
bf0f6f24 | 1345 | /* |
a2a2d680 | 1346 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1347 | */ |
b7cc0896 | 1348 | update_curr(cfs_rq); |
d6b55918 | 1349 | update_cfs_load(cfs_rq, 0); |
9ee474f5 | 1350 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
a992241d | 1351 | account_entity_enqueue(cfs_rq, se); |
6d5ab293 | 1352 | update_cfs_shares(cfs_rq); |
bf0f6f24 | 1353 | |
88ec22d3 | 1354 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 1355 | place_entity(cfs_rq, se, 0); |
2396af69 | 1356 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 1357 | } |
bf0f6f24 | 1358 | |
d2417e5a | 1359 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 1360 | check_spread(cfs_rq, se); |
83b699ed SV |
1361 | if (se != cfs_rq->curr) |
1362 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1363 | se->on_rq = 1; |
3d4b47b4 | 1364 | |
d3d9dc33 | 1365 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1366 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1367 | check_enqueue_throttle(cfs_rq); |
1368 | } | |
bf0f6f24 IM |
1369 | } |
1370 | ||
2c13c919 | 1371 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1372 | { |
2c13c919 RR |
1373 | for_each_sched_entity(se) { |
1374 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1375 | if (cfs_rq->last == se) | |
1376 | cfs_rq->last = NULL; | |
1377 | else | |
1378 | break; | |
1379 | } | |
1380 | } | |
2002c695 | 1381 | |
2c13c919 RR |
1382 | static void __clear_buddies_next(struct sched_entity *se) |
1383 | { | |
1384 | for_each_sched_entity(se) { | |
1385 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1386 | if (cfs_rq->next == se) | |
1387 | cfs_rq->next = NULL; | |
1388 | else | |
1389 | break; | |
1390 | } | |
2002c695 PZ |
1391 | } |
1392 | ||
ac53db59 RR |
1393 | static void __clear_buddies_skip(struct sched_entity *se) |
1394 | { | |
1395 | for_each_sched_entity(se) { | |
1396 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1397 | if (cfs_rq->skip == se) | |
1398 | cfs_rq->skip = NULL; | |
1399 | else | |
1400 | break; | |
1401 | } | |
1402 | } | |
1403 | ||
a571bbea PZ |
1404 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1405 | { | |
2c13c919 RR |
1406 | if (cfs_rq->last == se) |
1407 | __clear_buddies_last(se); | |
1408 | ||
1409 | if (cfs_rq->next == se) | |
1410 | __clear_buddies_next(se); | |
ac53db59 RR |
1411 | |
1412 | if (cfs_rq->skip == se) | |
1413 | __clear_buddies_skip(se); | |
a571bbea PZ |
1414 | } |
1415 | ||
6c16a6dc | 1416 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 1417 | |
bf0f6f24 | 1418 | static void |
371fd7e7 | 1419 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1420 | { |
a2a2d680 DA |
1421 | /* |
1422 | * Update run-time statistics of the 'current'. | |
1423 | */ | |
1424 | update_curr(cfs_rq); | |
9ee474f5 | 1425 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 1426 | |
19b6a2e3 | 1427 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1428 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1429 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1430 | if (entity_is_task(se)) { |
1431 | struct task_struct *tsk = task_of(se); | |
1432 | ||
1433 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1434 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1435 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1436 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1437 | } |
db36cc7d | 1438 | #endif |
67e9fb2a PZ |
1439 | } |
1440 | ||
2002c695 | 1441 | clear_buddies(cfs_rq, se); |
4793241b | 1442 | |
83b699ed | 1443 | if (se != cfs_rq->curr) |
30cfdcfc | 1444 | __dequeue_entity(cfs_rq, se); |
2069dd75 | 1445 | se->on_rq = 0; |
d6b55918 | 1446 | update_cfs_load(cfs_rq, 0); |
30cfdcfc | 1447 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1448 | |
1449 | /* | |
1450 | * Normalize the entity after updating the min_vruntime because the | |
1451 | * update can refer to the ->curr item and we need to reflect this | |
1452 | * movement in our normalized position. | |
1453 | */ | |
371fd7e7 | 1454 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1455 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 1456 | |
d8b4986d PT |
1457 | /* return excess runtime on last dequeue */ |
1458 | return_cfs_rq_runtime(cfs_rq); | |
1459 | ||
1e876231 PZ |
1460 | update_min_vruntime(cfs_rq); |
1461 | update_cfs_shares(cfs_rq); | |
bf0f6f24 IM |
1462 | } |
1463 | ||
1464 | /* | |
1465 | * Preempt the current task with a newly woken task if needed: | |
1466 | */ | |
7c92e54f | 1467 | static void |
2e09bf55 | 1468 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1469 | { |
11697830 | 1470 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
1471 | struct sched_entity *se; |
1472 | s64 delta; | |
11697830 | 1473 | |
6d0f0ebd | 1474 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1475 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1476 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1477 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1478 | /* |
1479 | * The current task ran long enough, ensure it doesn't get | |
1480 | * re-elected due to buddy favours. | |
1481 | */ | |
1482 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1483 | return; |
1484 | } | |
1485 | ||
1486 | /* | |
1487 | * Ensure that a task that missed wakeup preemption by a | |
1488 | * narrow margin doesn't have to wait for a full slice. | |
1489 | * This also mitigates buddy induced latencies under load. | |
1490 | */ | |
f685ceac MG |
1491 | if (delta_exec < sysctl_sched_min_granularity) |
1492 | return; | |
1493 | ||
f4cfb33e WX |
1494 | se = __pick_first_entity(cfs_rq); |
1495 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 1496 | |
f4cfb33e WX |
1497 | if (delta < 0) |
1498 | return; | |
d7d82944 | 1499 | |
f4cfb33e WX |
1500 | if (delta > ideal_runtime) |
1501 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
1502 | } |
1503 | ||
83b699ed | 1504 | static void |
8494f412 | 1505 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1506 | { |
83b699ed SV |
1507 | /* 'current' is not kept within the tree. */ |
1508 | if (se->on_rq) { | |
1509 | /* | |
1510 | * Any task has to be enqueued before it get to execute on | |
1511 | * a CPU. So account for the time it spent waiting on the | |
1512 | * runqueue. | |
1513 | */ | |
1514 | update_stats_wait_end(cfs_rq, se); | |
1515 | __dequeue_entity(cfs_rq, se); | |
1516 | } | |
1517 | ||
79303e9e | 1518 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1519 | cfs_rq->curr = se; |
eba1ed4b IM |
1520 | #ifdef CONFIG_SCHEDSTATS |
1521 | /* | |
1522 | * Track our maximum slice length, if the CPU's load is at | |
1523 | * least twice that of our own weight (i.e. dont track it | |
1524 | * when there are only lesser-weight tasks around): | |
1525 | */ | |
495eca49 | 1526 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1527 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1528 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1529 | } | |
1530 | #endif | |
4a55b450 | 1531 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1532 | } |
1533 | ||
3f3a4904 PZ |
1534 | static int |
1535 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1536 | ||
ac53db59 RR |
1537 | /* |
1538 | * Pick the next process, keeping these things in mind, in this order: | |
1539 | * 1) keep things fair between processes/task groups | |
1540 | * 2) pick the "next" process, since someone really wants that to run | |
1541 | * 3) pick the "last" process, for cache locality | |
1542 | * 4) do not run the "skip" process, if something else is available | |
1543 | */ | |
f4b6755f | 1544 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1545 | { |
ac53db59 | 1546 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1547 | struct sched_entity *left = se; |
f4b6755f | 1548 | |
ac53db59 RR |
1549 | /* |
1550 | * Avoid running the skip buddy, if running something else can | |
1551 | * be done without getting too unfair. | |
1552 | */ | |
1553 | if (cfs_rq->skip == se) { | |
1554 | struct sched_entity *second = __pick_next_entity(se); | |
1555 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1556 | se = second; | |
1557 | } | |
aa2ac252 | 1558 | |
f685ceac MG |
1559 | /* |
1560 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1561 | */ | |
1562 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1563 | se = cfs_rq->last; | |
1564 | ||
ac53db59 RR |
1565 | /* |
1566 | * Someone really wants this to run. If it's not unfair, run it. | |
1567 | */ | |
1568 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1569 | se = cfs_rq->next; | |
1570 | ||
f685ceac | 1571 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1572 | |
1573 | return se; | |
aa2ac252 PZ |
1574 | } |
1575 | ||
d3d9dc33 PT |
1576 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
1577 | ||
ab6cde26 | 1578 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1579 | { |
1580 | /* | |
1581 | * If still on the runqueue then deactivate_task() | |
1582 | * was not called and update_curr() has to be done: | |
1583 | */ | |
1584 | if (prev->on_rq) | |
b7cc0896 | 1585 | update_curr(cfs_rq); |
bf0f6f24 | 1586 | |
d3d9dc33 PT |
1587 | /* throttle cfs_rqs exceeding runtime */ |
1588 | check_cfs_rq_runtime(cfs_rq); | |
1589 | ||
ddc97297 | 1590 | check_spread(cfs_rq, prev); |
30cfdcfc | 1591 | if (prev->on_rq) { |
5870db5b | 1592 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1593 | /* Put 'current' back into the tree. */ |
1594 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 1595 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 1596 | update_entity_load_avg(prev, 1); |
30cfdcfc | 1597 | } |
429d43bc | 1598 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1599 | } |
1600 | ||
8f4d37ec PZ |
1601 | static void |
1602 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1603 | { |
bf0f6f24 | 1604 | /* |
30cfdcfc | 1605 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1606 | */ |
30cfdcfc | 1607 | update_curr(cfs_rq); |
bf0f6f24 | 1608 | |
9d85f21c PT |
1609 | /* |
1610 | * Ensure that runnable average is periodically updated. | |
1611 | */ | |
9ee474f5 PT |
1612 | update_entity_load_avg(curr, 1); |
1613 | update_cfs_rq_blocked_load(cfs_rq); | |
9d85f21c | 1614 | |
43365bd7 PT |
1615 | /* |
1616 | * Update share accounting for long-running entities. | |
1617 | */ | |
1618 | update_entity_shares_tick(cfs_rq); | |
1619 | ||
8f4d37ec PZ |
1620 | #ifdef CONFIG_SCHED_HRTICK |
1621 | /* | |
1622 | * queued ticks are scheduled to match the slice, so don't bother | |
1623 | * validating it and just reschedule. | |
1624 | */ | |
983ed7a6 HH |
1625 | if (queued) { |
1626 | resched_task(rq_of(cfs_rq)->curr); | |
1627 | return; | |
1628 | } | |
8f4d37ec PZ |
1629 | /* |
1630 | * don't let the period tick interfere with the hrtick preemption | |
1631 | */ | |
1632 | if (!sched_feat(DOUBLE_TICK) && | |
1633 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1634 | return; | |
1635 | #endif | |
1636 | ||
2c2efaed | 1637 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1638 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1639 | } |
1640 | ||
ab84d31e PT |
1641 | |
1642 | /************************************************** | |
1643 | * CFS bandwidth control machinery | |
1644 | */ | |
1645 | ||
1646 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
1647 | |
1648 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 1649 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
1650 | |
1651 | static inline bool cfs_bandwidth_used(void) | |
1652 | { | |
c5905afb | 1653 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
1654 | } |
1655 | ||
1656 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
1657 | { | |
1658 | /* only need to count groups transitioning between enabled/!enabled */ | |
1659 | if (enabled && !was_enabled) | |
c5905afb | 1660 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 1661 | else if (!enabled && was_enabled) |
c5905afb | 1662 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
1663 | } |
1664 | #else /* HAVE_JUMP_LABEL */ | |
1665 | static bool cfs_bandwidth_used(void) | |
1666 | { | |
1667 | return true; | |
1668 | } | |
1669 | ||
1670 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
1671 | #endif /* HAVE_JUMP_LABEL */ | |
1672 | ||
ab84d31e PT |
1673 | /* |
1674 | * default period for cfs group bandwidth. | |
1675 | * default: 0.1s, units: nanoseconds | |
1676 | */ | |
1677 | static inline u64 default_cfs_period(void) | |
1678 | { | |
1679 | return 100000000ULL; | |
1680 | } | |
ec12cb7f PT |
1681 | |
1682 | static inline u64 sched_cfs_bandwidth_slice(void) | |
1683 | { | |
1684 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
1685 | } | |
1686 | ||
a9cf55b2 PT |
1687 | /* |
1688 | * Replenish runtime according to assigned quota and update expiration time. | |
1689 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
1690 | * additional synchronization around rq->lock. | |
1691 | * | |
1692 | * requires cfs_b->lock | |
1693 | */ | |
029632fb | 1694 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
1695 | { |
1696 | u64 now; | |
1697 | ||
1698 | if (cfs_b->quota == RUNTIME_INF) | |
1699 | return; | |
1700 | ||
1701 | now = sched_clock_cpu(smp_processor_id()); | |
1702 | cfs_b->runtime = cfs_b->quota; | |
1703 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
1704 | } | |
1705 | ||
029632fb PZ |
1706 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
1707 | { | |
1708 | return &tg->cfs_bandwidth; | |
1709 | } | |
1710 | ||
85dac906 PT |
1711 | /* returns 0 on failure to allocate runtime */ |
1712 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
1713 | { |
1714 | struct task_group *tg = cfs_rq->tg; | |
1715 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 1716 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
1717 | |
1718 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
1719 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
1720 | ||
1721 | raw_spin_lock(&cfs_b->lock); | |
1722 | if (cfs_b->quota == RUNTIME_INF) | |
1723 | amount = min_amount; | |
58088ad0 | 1724 | else { |
a9cf55b2 PT |
1725 | /* |
1726 | * If the bandwidth pool has become inactive, then at least one | |
1727 | * period must have elapsed since the last consumption. | |
1728 | * Refresh the global state and ensure bandwidth timer becomes | |
1729 | * active. | |
1730 | */ | |
1731 | if (!cfs_b->timer_active) { | |
1732 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 1733 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 1734 | } |
58088ad0 PT |
1735 | |
1736 | if (cfs_b->runtime > 0) { | |
1737 | amount = min(cfs_b->runtime, min_amount); | |
1738 | cfs_b->runtime -= amount; | |
1739 | cfs_b->idle = 0; | |
1740 | } | |
ec12cb7f | 1741 | } |
a9cf55b2 | 1742 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
1743 | raw_spin_unlock(&cfs_b->lock); |
1744 | ||
1745 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
1746 | /* |
1747 | * we may have advanced our local expiration to account for allowed | |
1748 | * spread between our sched_clock and the one on which runtime was | |
1749 | * issued. | |
1750 | */ | |
1751 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
1752 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
1753 | |
1754 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
1755 | } |
1756 | ||
a9cf55b2 PT |
1757 | /* |
1758 | * Note: This depends on the synchronization provided by sched_clock and the | |
1759 | * fact that rq->clock snapshots this value. | |
1760 | */ | |
1761 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 1762 | { |
a9cf55b2 PT |
1763 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
1764 | struct rq *rq = rq_of(cfs_rq); | |
1765 | ||
1766 | /* if the deadline is ahead of our clock, nothing to do */ | |
1767 | if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0)) | |
ec12cb7f PT |
1768 | return; |
1769 | ||
a9cf55b2 PT |
1770 | if (cfs_rq->runtime_remaining < 0) |
1771 | return; | |
1772 | ||
1773 | /* | |
1774 | * If the local deadline has passed we have to consider the | |
1775 | * possibility that our sched_clock is 'fast' and the global deadline | |
1776 | * has not truly expired. | |
1777 | * | |
1778 | * Fortunately we can check determine whether this the case by checking | |
1779 | * whether the global deadline has advanced. | |
1780 | */ | |
1781 | ||
1782 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
1783 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
1784 | cfs_rq->runtime_expires += TICK_NSEC; | |
1785 | } else { | |
1786 | /* global deadline is ahead, expiration has passed */ | |
1787 | cfs_rq->runtime_remaining = 0; | |
1788 | } | |
1789 | } | |
1790 | ||
1791 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1792 | unsigned long delta_exec) | |
1793 | { | |
1794 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 1795 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
1796 | expire_cfs_rq_runtime(cfs_rq); |
1797 | ||
1798 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
1799 | return; |
1800 | ||
85dac906 PT |
1801 | /* |
1802 | * if we're unable to extend our runtime we resched so that the active | |
1803 | * hierarchy can be throttled | |
1804 | */ | |
1805 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
1806 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
1807 | } |
1808 | ||
6c16a6dc PZ |
1809 | static __always_inline |
1810 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 1811 | { |
56f570e5 | 1812 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
1813 | return; |
1814 | ||
1815 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
1816 | } | |
1817 | ||
85dac906 PT |
1818 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
1819 | { | |
56f570e5 | 1820 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
1821 | } |
1822 | ||
64660c86 PT |
1823 | /* check whether cfs_rq, or any parent, is throttled */ |
1824 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
1825 | { | |
56f570e5 | 1826 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
1827 | } |
1828 | ||
1829 | /* | |
1830 | * Ensure that neither of the group entities corresponding to src_cpu or | |
1831 | * dest_cpu are members of a throttled hierarchy when performing group | |
1832 | * load-balance operations. | |
1833 | */ | |
1834 | static inline int throttled_lb_pair(struct task_group *tg, | |
1835 | int src_cpu, int dest_cpu) | |
1836 | { | |
1837 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
1838 | ||
1839 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
1840 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
1841 | ||
1842 | return throttled_hierarchy(src_cfs_rq) || | |
1843 | throttled_hierarchy(dest_cfs_rq); | |
1844 | } | |
1845 | ||
1846 | /* updated child weight may affect parent so we have to do this bottom up */ | |
1847 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
1848 | { | |
1849 | struct rq *rq = data; | |
1850 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1851 | ||
1852 | cfs_rq->throttle_count--; | |
1853 | #ifdef CONFIG_SMP | |
1854 | if (!cfs_rq->throttle_count) { | |
1855 | u64 delta = rq->clock_task - cfs_rq->load_stamp; | |
1856 | ||
1857 | /* leaving throttled state, advance shares averaging windows */ | |
1858 | cfs_rq->load_stamp += delta; | |
1859 | cfs_rq->load_last += delta; | |
1860 | ||
1861 | /* update entity weight now that we are on_rq again */ | |
1862 | update_cfs_shares(cfs_rq); | |
1863 | } | |
1864 | #endif | |
1865 | ||
1866 | return 0; | |
1867 | } | |
1868 | ||
1869 | static int tg_throttle_down(struct task_group *tg, void *data) | |
1870 | { | |
1871 | struct rq *rq = data; | |
1872 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1873 | ||
1874 | /* group is entering throttled state, record last load */ | |
1875 | if (!cfs_rq->throttle_count) | |
1876 | update_cfs_load(cfs_rq, 0); | |
1877 | cfs_rq->throttle_count++; | |
1878 | ||
1879 | return 0; | |
1880 | } | |
1881 | ||
d3d9dc33 | 1882 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
1883 | { |
1884 | struct rq *rq = rq_of(cfs_rq); | |
1885 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1886 | struct sched_entity *se; | |
1887 | long task_delta, dequeue = 1; | |
1888 | ||
1889 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1890 | ||
1891 | /* account load preceding throttle */ | |
64660c86 PT |
1892 | rcu_read_lock(); |
1893 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
1894 | rcu_read_unlock(); | |
85dac906 PT |
1895 | |
1896 | task_delta = cfs_rq->h_nr_running; | |
1897 | for_each_sched_entity(se) { | |
1898 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
1899 | /* throttled entity or throttle-on-deactivate */ | |
1900 | if (!se->on_rq) | |
1901 | break; | |
1902 | ||
1903 | if (dequeue) | |
1904 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
1905 | qcfs_rq->h_nr_running -= task_delta; | |
1906 | ||
1907 | if (qcfs_rq->load.weight) | |
1908 | dequeue = 0; | |
1909 | } | |
1910 | ||
1911 | if (!se) | |
1912 | rq->nr_running -= task_delta; | |
1913 | ||
1914 | cfs_rq->throttled = 1; | |
e8da1b18 | 1915 | cfs_rq->throttled_timestamp = rq->clock; |
85dac906 PT |
1916 | raw_spin_lock(&cfs_b->lock); |
1917 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
1918 | raw_spin_unlock(&cfs_b->lock); | |
1919 | } | |
1920 | ||
029632fb | 1921 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
1922 | { |
1923 | struct rq *rq = rq_of(cfs_rq); | |
1924 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
1925 | struct sched_entity *se; | |
1926 | int enqueue = 1; | |
1927 | long task_delta; | |
1928 | ||
1929 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
1930 | ||
1931 | cfs_rq->throttled = 0; | |
1932 | raw_spin_lock(&cfs_b->lock); | |
e8da1b18 | 1933 | cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp; |
671fd9da PT |
1934 | list_del_rcu(&cfs_rq->throttled_list); |
1935 | raw_spin_unlock(&cfs_b->lock); | |
e8da1b18 | 1936 | cfs_rq->throttled_timestamp = 0; |
671fd9da | 1937 | |
64660c86 PT |
1938 | update_rq_clock(rq); |
1939 | /* update hierarchical throttle state */ | |
1940 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
1941 | ||
671fd9da PT |
1942 | if (!cfs_rq->load.weight) |
1943 | return; | |
1944 | ||
1945 | task_delta = cfs_rq->h_nr_running; | |
1946 | for_each_sched_entity(se) { | |
1947 | if (se->on_rq) | |
1948 | enqueue = 0; | |
1949 | ||
1950 | cfs_rq = cfs_rq_of(se); | |
1951 | if (enqueue) | |
1952 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
1953 | cfs_rq->h_nr_running += task_delta; | |
1954 | ||
1955 | if (cfs_rq_throttled(cfs_rq)) | |
1956 | break; | |
1957 | } | |
1958 | ||
1959 | if (!se) | |
1960 | rq->nr_running += task_delta; | |
1961 | ||
1962 | /* determine whether we need to wake up potentially idle cpu */ | |
1963 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
1964 | resched_task(rq->curr); | |
1965 | } | |
1966 | ||
1967 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
1968 | u64 remaining, u64 expires) | |
1969 | { | |
1970 | struct cfs_rq *cfs_rq; | |
1971 | u64 runtime = remaining; | |
1972 | ||
1973 | rcu_read_lock(); | |
1974 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
1975 | throttled_list) { | |
1976 | struct rq *rq = rq_of(cfs_rq); | |
1977 | ||
1978 | raw_spin_lock(&rq->lock); | |
1979 | if (!cfs_rq_throttled(cfs_rq)) | |
1980 | goto next; | |
1981 | ||
1982 | runtime = -cfs_rq->runtime_remaining + 1; | |
1983 | if (runtime > remaining) | |
1984 | runtime = remaining; | |
1985 | remaining -= runtime; | |
1986 | ||
1987 | cfs_rq->runtime_remaining += runtime; | |
1988 | cfs_rq->runtime_expires = expires; | |
1989 | ||
1990 | /* we check whether we're throttled above */ | |
1991 | if (cfs_rq->runtime_remaining > 0) | |
1992 | unthrottle_cfs_rq(cfs_rq); | |
1993 | ||
1994 | next: | |
1995 | raw_spin_unlock(&rq->lock); | |
1996 | ||
1997 | if (!remaining) | |
1998 | break; | |
1999 | } | |
2000 | rcu_read_unlock(); | |
2001 | ||
2002 | return remaining; | |
2003 | } | |
2004 | ||
58088ad0 PT |
2005 | /* |
2006 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2007 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2008 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2009 | * used to track this state. | |
2010 | */ | |
2011 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
2012 | { | |
671fd9da PT |
2013 | u64 runtime, runtime_expires; |
2014 | int idle = 1, throttled; | |
58088ad0 PT |
2015 | |
2016 | raw_spin_lock(&cfs_b->lock); | |
2017 | /* no need to continue the timer with no bandwidth constraint */ | |
2018 | if (cfs_b->quota == RUNTIME_INF) | |
2019 | goto out_unlock; | |
2020 | ||
671fd9da PT |
2021 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
2022 | /* idle depends on !throttled (for the case of a large deficit) */ | |
2023 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 2024 | cfs_b->nr_periods += overrun; |
671fd9da | 2025 | |
a9cf55b2 PT |
2026 | /* if we're going inactive then everything else can be deferred */ |
2027 | if (idle) | |
2028 | goto out_unlock; | |
2029 | ||
2030 | __refill_cfs_bandwidth_runtime(cfs_b); | |
2031 | ||
671fd9da PT |
2032 | if (!throttled) { |
2033 | /* mark as potentially idle for the upcoming period */ | |
2034 | cfs_b->idle = 1; | |
2035 | goto out_unlock; | |
2036 | } | |
2037 | ||
e8da1b18 NR |
2038 | /* account preceding periods in which throttling occurred */ |
2039 | cfs_b->nr_throttled += overrun; | |
2040 | ||
671fd9da PT |
2041 | /* |
2042 | * There are throttled entities so we must first use the new bandwidth | |
2043 | * to unthrottle them before making it generally available. This | |
2044 | * ensures that all existing debts will be paid before a new cfs_rq is | |
2045 | * allowed to run. | |
2046 | */ | |
2047 | runtime = cfs_b->runtime; | |
2048 | runtime_expires = cfs_b->runtime_expires; | |
2049 | cfs_b->runtime = 0; | |
2050 | ||
2051 | /* | |
2052 | * This check is repeated as we are holding onto the new bandwidth | |
2053 | * while we unthrottle. This can potentially race with an unthrottled | |
2054 | * group trying to acquire new bandwidth from the global pool. | |
2055 | */ | |
2056 | while (throttled && runtime > 0) { | |
2057 | raw_spin_unlock(&cfs_b->lock); | |
2058 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
2059 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
2060 | runtime_expires); | |
2061 | raw_spin_lock(&cfs_b->lock); | |
2062 | ||
2063 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
2064 | } | |
58088ad0 | 2065 | |
671fd9da PT |
2066 | /* return (any) remaining runtime */ |
2067 | cfs_b->runtime = runtime; | |
2068 | /* | |
2069 | * While we are ensured activity in the period following an | |
2070 | * unthrottle, this also covers the case in which the new bandwidth is | |
2071 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
2072 | * timer to remain active while there are any throttled entities.) | |
2073 | */ | |
2074 | cfs_b->idle = 0; | |
58088ad0 PT |
2075 | out_unlock: |
2076 | if (idle) | |
2077 | cfs_b->timer_active = 0; | |
2078 | raw_spin_unlock(&cfs_b->lock); | |
2079 | ||
2080 | return idle; | |
2081 | } | |
d3d9dc33 | 2082 | |
d8b4986d PT |
2083 | /* a cfs_rq won't donate quota below this amount */ |
2084 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
2085 | /* minimum remaining period time to redistribute slack quota */ | |
2086 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
2087 | /* how long we wait to gather additional slack before distributing */ | |
2088 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
2089 | ||
2090 | /* are we near the end of the current quota period? */ | |
2091 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
2092 | { | |
2093 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
2094 | u64 remaining; | |
2095 | ||
2096 | /* if the call-back is running a quota refresh is already occurring */ | |
2097 | if (hrtimer_callback_running(refresh_timer)) | |
2098 | return 1; | |
2099 | ||
2100 | /* is a quota refresh about to occur? */ | |
2101 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
2102 | if (remaining < min_expire) | |
2103 | return 1; | |
2104 | ||
2105 | return 0; | |
2106 | } | |
2107 | ||
2108 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
2109 | { | |
2110 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
2111 | ||
2112 | /* if there's a quota refresh soon don't bother with slack */ | |
2113 | if (runtime_refresh_within(cfs_b, min_left)) | |
2114 | return; | |
2115 | ||
2116 | start_bandwidth_timer(&cfs_b->slack_timer, | |
2117 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
2118 | } | |
2119 | ||
2120 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
2121 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2122 | { | |
2123 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2124 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
2125 | ||
2126 | if (slack_runtime <= 0) | |
2127 | return; | |
2128 | ||
2129 | raw_spin_lock(&cfs_b->lock); | |
2130 | if (cfs_b->quota != RUNTIME_INF && | |
2131 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
2132 | cfs_b->runtime += slack_runtime; | |
2133 | ||
2134 | /* we are under rq->lock, defer unthrottling using a timer */ | |
2135 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
2136 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
2137 | start_cfs_slack_bandwidth(cfs_b); | |
2138 | } | |
2139 | raw_spin_unlock(&cfs_b->lock); | |
2140 | ||
2141 | /* even if it's not valid for return we don't want to try again */ | |
2142 | cfs_rq->runtime_remaining -= slack_runtime; | |
2143 | } | |
2144 | ||
2145 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2146 | { | |
56f570e5 PT |
2147 | if (!cfs_bandwidth_used()) |
2148 | return; | |
2149 | ||
fccfdc6f | 2150 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2151 | return; |
2152 | ||
2153 | __return_cfs_rq_runtime(cfs_rq); | |
2154 | } | |
2155 | ||
2156 | /* | |
2157 | * This is done with a timer (instead of inline with bandwidth return) since | |
2158 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2159 | */ | |
2160 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2161 | { | |
2162 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2163 | u64 expires; | |
2164 | ||
2165 | /* confirm we're still not at a refresh boundary */ | |
2166 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2167 | return; | |
2168 | ||
2169 | raw_spin_lock(&cfs_b->lock); | |
2170 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2171 | runtime = cfs_b->runtime; | |
2172 | cfs_b->runtime = 0; | |
2173 | } | |
2174 | expires = cfs_b->runtime_expires; | |
2175 | raw_spin_unlock(&cfs_b->lock); | |
2176 | ||
2177 | if (!runtime) | |
2178 | return; | |
2179 | ||
2180 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2181 | ||
2182 | raw_spin_lock(&cfs_b->lock); | |
2183 | if (expires == cfs_b->runtime_expires) | |
2184 | cfs_b->runtime = runtime; | |
2185 | raw_spin_unlock(&cfs_b->lock); | |
2186 | } | |
2187 | ||
d3d9dc33 PT |
2188 | /* |
2189 | * When a group wakes up we want to make sure that its quota is not already | |
2190 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2191 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2192 | */ | |
2193 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2194 | { | |
56f570e5 PT |
2195 | if (!cfs_bandwidth_used()) |
2196 | return; | |
2197 | ||
d3d9dc33 PT |
2198 | /* an active group must be handled by the update_curr()->put() path */ |
2199 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2200 | return; | |
2201 | ||
2202 | /* ensure the group is not already throttled */ | |
2203 | if (cfs_rq_throttled(cfs_rq)) | |
2204 | return; | |
2205 | ||
2206 | /* update runtime allocation */ | |
2207 | account_cfs_rq_runtime(cfs_rq, 0); | |
2208 | if (cfs_rq->runtime_remaining <= 0) | |
2209 | throttle_cfs_rq(cfs_rq); | |
2210 | } | |
2211 | ||
2212 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2213 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2214 | { | |
56f570e5 PT |
2215 | if (!cfs_bandwidth_used()) |
2216 | return; | |
2217 | ||
d3d9dc33 PT |
2218 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2219 | return; | |
2220 | ||
2221 | /* | |
2222 | * it's possible for a throttled entity to be forced into a running | |
2223 | * state (e.g. set_curr_task), in this case we're finished. | |
2224 | */ | |
2225 | if (cfs_rq_throttled(cfs_rq)) | |
2226 | return; | |
2227 | ||
2228 | throttle_cfs_rq(cfs_rq); | |
2229 | } | |
029632fb PZ |
2230 | |
2231 | static inline u64 default_cfs_period(void); | |
2232 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun); | |
2233 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b); | |
2234 | ||
2235 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) | |
2236 | { | |
2237 | struct cfs_bandwidth *cfs_b = | |
2238 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2239 | do_sched_cfs_slack_timer(cfs_b); | |
2240 | ||
2241 | return HRTIMER_NORESTART; | |
2242 | } | |
2243 | ||
2244 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2245 | { | |
2246 | struct cfs_bandwidth *cfs_b = | |
2247 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2248 | ktime_t now; | |
2249 | int overrun; | |
2250 | int idle = 0; | |
2251 | ||
2252 | for (;;) { | |
2253 | now = hrtimer_cb_get_time(timer); | |
2254 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2255 | ||
2256 | if (!overrun) | |
2257 | break; | |
2258 | ||
2259 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2260 | } | |
2261 | ||
2262 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2263 | } | |
2264 | ||
2265 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2266 | { | |
2267 | raw_spin_lock_init(&cfs_b->lock); | |
2268 | cfs_b->runtime = 0; | |
2269 | cfs_b->quota = RUNTIME_INF; | |
2270 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2271 | ||
2272 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2273 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2274 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2275 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2276 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2277 | } | |
2278 | ||
2279 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2280 | { | |
2281 | cfs_rq->runtime_enabled = 0; | |
2282 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2283 | } | |
2284 | ||
2285 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2286 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2287 | { | |
2288 | /* | |
2289 | * The timer may be active because we're trying to set a new bandwidth | |
2290 | * period or because we're racing with the tear-down path | |
2291 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2292 | * terminates). In either case we ensure that it's re-programmed | |
2293 | */ | |
2294 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2295 | raw_spin_unlock(&cfs_b->lock); | |
2296 | /* ensure cfs_b->lock is available while we wait */ | |
2297 | hrtimer_cancel(&cfs_b->period_timer); | |
2298 | ||
2299 | raw_spin_lock(&cfs_b->lock); | |
2300 | /* if someone else restarted the timer then we're done */ | |
2301 | if (cfs_b->timer_active) | |
2302 | return; | |
2303 | } | |
2304 | ||
2305 | cfs_b->timer_active = 1; | |
2306 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
2307 | } | |
2308 | ||
2309 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2310 | { | |
2311 | hrtimer_cancel(&cfs_b->period_timer); | |
2312 | hrtimer_cancel(&cfs_b->slack_timer); | |
2313 | } | |
2314 | ||
a4c96ae3 | 2315 | static void unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
2316 | { |
2317 | struct cfs_rq *cfs_rq; | |
2318 | ||
2319 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2320 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2321 | ||
2322 | if (!cfs_rq->runtime_enabled) | |
2323 | continue; | |
2324 | ||
2325 | /* | |
2326 | * clock_task is not advancing so we just need to make sure | |
2327 | * there's some valid quota amount | |
2328 | */ | |
2329 | cfs_rq->runtime_remaining = cfs_b->quota; | |
2330 | if (cfs_rq_throttled(cfs_rq)) | |
2331 | unthrottle_cfs_rq(cfs_rq); | |
2332 | } | |
2333 | } | |
2334 | ||
2335 | #else /* CONFIG_CFS_BANDWIDTH */ | |
6c16a6dc PZ |
2336 | static __always_inline |
2337 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {} | |
d3d9dc33 PT |
2338 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
2339 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 2340 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
2341 | |
2342 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
2343 | { | |
2344 | return 0; | |
2345 | } | |
64660c86 PT |
2346 | |
2347 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2348 | { | |
2349 | return 0; | |
2350 | } | |
2351 | ||
2352 | static inline int throttled_lb_pair(struct task_group *tg, | |
2353 | int src_cpu, int dest_cpu) | |
2354 | { | |
2355 | return 0; | |
2356 | } | |
029632fb PZ |
2357 | |
2358 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
2359 | ||
2360 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2361 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
2362 | #endif |
2363 | ||
029632fb PZ |
2364 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2365 | { | |
2366 | return NULL; | |
2367 | } | |
2368 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 2369 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
2370 | |
2371 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
2372 | ||
bf0f6f24 IM |
2373 | /************************************************** |
2374 | * CFS operations on tasks: | |
2375 | */ | |
2376 | ||
8f4d37ec PZ |
2377 | #ifdef CONFIG_SCHED_HRTICK |
2378 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2379 | { | |
8f4d37ec PZ |
2380 | struct sched_entity *se = &p->se; |
2381 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2382 | ||
2383 | WARN_ON(task_rq(p) != rq); | |
2384 | ||
b39e66ea | 2385 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
2386 | u64 slice = sched_slice(cfs_rq, se); |
2387 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
2388 | s64 delta = slice - ran; | |
2389 | ||
2390 | if (delta < 0) { | |
2391 | if (rq->curr == p) | |
2392 | resched_task(p); | |
2393 | return; | |
2394 | } | |
2395 | ||
2396 | /* | |
2397 | * Don't schedule slices shorter than 10000ns, that just | |
2398 | * doesn't make sense. Rely on vruntime for fairness. | |
2399 | */ | |
31656519 | 2400 | if (rq->curr != p) |
157124c1 | 2401 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 2402 | |
31656519 | 2403 | hrtick_start(rq, delta); |
8f4d37ec PZ |
2404 | } |
2405 | } | |
a4c2f00f PZ |
2406 | |
2407 | /* | |
2408 | * called from enqueue/dequeue and updates the hrtick when the | |
2409 | * current task is from our class and nr_running is low enough | |
2410 | * to matter. | |
2411 | */ | |
2412 | static void hrtick_update(struct rq *rq) | |
2413 | { | |
2414 | struct task_struct *curr = rq->curr; | |
2415 | ||
b39e66ea | 2416 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
2417 | return; |
2418 | ||
2419 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
2420 | hrtick_start_fair(rq, curr); | |
2421 | } | |
55e12e5e | 2422 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
2423 | static inline void |
2424 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2425 | { | |
2426 | } | |
a4c2f00f PZ |
2427 | |
2428 | static inline void hrtick_update(struct rq *rq) | |
2429 | { | |
2430 | } | |
8f4d37ec PZ |
2431 | #endif |
2432 | ||
bf0f6f24 IM |
2433 | /* |
2434 | * The enqueue_task method is called before nr_running is | |
2435 | * increased. Here we update the fair scheduling stats and | |
2436 | * then put the task into the rbtree: | |
2437 | */ | |
ea87bb78 | 2438 | static void |
371fd7e7 | 2439 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2440 | { |
2441 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2442 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
2443 | |
2444 | for_each_sched_entity(se) { | |
62fb1851 | 2445 | if (se->on_rq) |
bf0f6f24 IM |
2446 | break; |
2447 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 2448 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
2449 | |
2450 | /* | |
2451 | * end evaluation on encountering a throttled cfs_rq | |
2452 | * | |
2453 | * note: in the case of encountering a throttled cfs_rq we will | |
2454 | * post the final h_nr_running increment below. | |
2455 | */ | |
2456 | if (cfs_rq_throttled(cfs_rq)) | |
2457 | break; | |
953bfcd1 | 2458 | cfs_rq->h_nr_running++; |
85dac906 | 2459 | |
88ec22d3 | 2460 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 2461 | } |
8f4d37ec | 2462 | |
2069dd75 | 2463 | for_each_sched_entity(se) { |
0f317143 | 2464 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2465 | cfs_rq->h_nr_running++; |
2069dd75 | 2466 | |
85dac906 PT |
2467 | if (cfs_rq_throttled(cfs_rq)) |
2468 | break; | |
2469 | ||
d6b55918 | 2470 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 2471 | update_cfs_shares(cfs_rq); |
9ee474f5 | 2472 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
2473 | } |
2474 | ||
18bf2805 BS |
2475 | if (!se) { |
2476 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 2477 | inc_nr_running(rq); |
18bf2805 | 2478 | } |
a4c2f00f | 2479 | hrtick_update(rq); |
bf0f6f24 IM |
2480 | } |
2481 | ||
2f36825b VP |
2482 | static void set_next_buddy(struct sched_entity *se); |
2483 | ||
bf0f6f24 IM |
2484 | /* |
2485 | * The dequeue_task method is called before nr_running is | |
2486 | * decreased. We remove the task from the rbtree and | |
2487 | * update the fair scheduling stats: | |
2488 | */ | |
371fd7e7 | 2489 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2490 | { |
2491 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2492 | struct sched_entity *se = &p->se; |
2f36825b | 2493 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
2494 | |
2495 | for_each_sched_entity(se) { | |
2496 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 2497 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
2498 | |
2499 | /* | |
2500 | * end evaluation on encountering a throttled cfs_rq | |
2501 | * | |
2502 | * note: in the case of encountering a throttled cfs_rq we will | |
2503 | * post the final h_nr_running decrement below. | |
2504 | */ | |
2505 | if (cfs_rq_throttled(cfs_rq)) | |
2506 | break; | |
953bfcd1 | 2507 | cfs_rq->h_nr_running--; |
2069dd75 | 2508 | |
bf0f6f24 | 2509 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
2510 | if (cfs_rq->load.weight) { |
2511 | /* | |
2512 | * Bias pick_next to pick a task from this cfs_rq, as | |
2513 | * p is sleeping when it is within its sched_slice. | |
2514 | */ | |
2515 | if (task_sleep && parent_entity(se)) | |
2516 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
2517 | |
2518 | /* avoid re-evaluating load for this entity */ | |
2519 | se = parent_entity(se); | |
bf0f6f24 | 2520 | break; |
2f36825b | 2521 | } |
371fd7e7 | 2522 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 2523 | } |
8f4d37ec | 2524 | |
2069dd75 | 2525 | for_each_sched_entity(se) { |
0f317143 | 2526 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2527 | cfs_rq->h_nr_running--; |
2069dd75 | 2528 | |
85dac906 PT |
2529 | if (cfs_rq_throttled(cfs_rq)) |
2530 | break; | |
2531 | ||
d6b55918 | 2532 | update_cfs_load(cfs_rq, 0); |
6d5ab293 | 2533 | update_cfs_shares(cfs_rq); |
9ee474f5 | 2534 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
2535 | } |
2536 | ||
18bf2805 | 2537 | if (!se) { |
85dac906 | 2538 | dec_nr_running(rq); |
18bf2805 BS |
2539 | update_rq_runnable_avg(rq, 1); |
2540 | } | |
a4c2f00f | 2541 | hrtick_update(rq); |
bf0f6f24 IM |
2542 | } |
2543 | ||
e7693a36 | 2544 | #ifdef CONFIG_SMP |
029632fb PZ |
2545 | /* Used instead of source_load when we know the type == 0 */ |
2546 | static unsigned long weighted_cpuload(const int cpu) | |
2547 | { | |
2548 | return cpu_rq(cpu)->load.weight; | |
2549 | } | |
2550 | ||
2551 | /* | |
2552 | * Return a low guess at the load of a migration-source cpu weighted | |
2553 | * according to the scheduling class and "nice" value. | |
2554 | * | |
2555 | * We want to under-estimate the load of migration sources, to | |
2556 | * balance conservatively. | |
2557 | */ | |
2558 | static unsigned long source_load(int cpu, int type) | |
2559 | { | |
2560 | struct rq *rq = cpu_rq(cpu); | |
2561 | unsigned long total = weighted_cpuload(cpu); | |
2562 | ||
2563 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2564 | return total; | |
2565 | ||
2566 | return min(rq->cpu_load[type-1], total); | |
2567 | } | |
2568 | ||
2569 | /* | |
2570 | * Return a high guess at the load of a migration-target cpu weighted | |
2571 | * according to the scheduling class and "nice" value. | |
2572 | */ | |
2573 | static unsigned long target_load(int cpu, int type) | |
2574 | { | |
2575 | struct rq *rq = cpu_rq(cpu); | |
2576 | unsigned long total = weighted_cpuload(cpu); | |
2577 | ||
2578 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2579 | return total; | |
2580 | ||
2581 | return max(rq->cpu_load[type-1], total); | |
2582 | } | |
2583 | ||
2584 | static unsigned long power_of(int cpu) | |
2585 | { | |
2586 | return cpu_rq(cpu)->cpu_power; | |
2587 | } | |
2588 | ||
2589 | static unsigned long cpu_avg_load_per_task(int cpu) | |
2590 | { | |
2591 | struct rq *rq = cpu_rq(cpu); | |
2592 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
2593 | ||
2594 | if (nr_running) | |
2595 | return rq->load.weight / nr_running; | |
2596 | ||
2597 | return 0; | |
2598 | } | |
2599 | ||
098fb9db | 2600 | |
74f8e4b2 | 2601 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
2602 | { |
2603 | struct sched_entity *se = &p->se; | |
2604 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
2605 | u64 min_vruntime; |
2606 | ||
2607 | #ifndef CONFIG_64BIT | |
2608 | u64 min_vruntime_copy; | |
88ec22d3 | 2609 | |
3fe1698b PZ |
2610 | do { |
2611 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
2612 | smp_rmb(); | |
2613 | min_vruntime = cfs_rq->min_vruntime; | |
2614 | } while (min_vruntime != min_vruntime_copy); | |
2615 | #else | |
2616 | min_vruntime = cfs_rq->min_vruntime; | |
2617 | #endif | |
88ec22d3 | 2618 | |
3fe1698b | 2619 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
2620 | } |
2621 | ||
bb3469ac | 2622 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
2623 | /* |
2624 | * effective_load() calculates the load change as seen from the root_task_group | |
2625 | * | |
2626 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
2627 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
2628 | * can calculate the shift in shares. | |
cf5f0acf PZ |
2629 | * |
2630 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
2631 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
2632 | * total group weight. | |
2633 | * | |
2634 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
2635 | * distribution (s_i) using: | |
2636 | * | |
2637 | * s_i = rw_i / \Sum rw_j (1) | |
2638 | * | |
2639 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
2640 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
2641 | * shares distribution (s_i): | |
2642 | * | |
2643 | * rw_i = { 2, 4, 1, 0 } | |
2644 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
2645 | * | |
2646 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
2647 | * task used to run on and the CPU the waker is running on), we need to | |
2648 | * compute the effect of waking a task on either CPU and, in case of a sync | |
2649 | * wakeup, compute the effect of the current task going to sleep. | |
2650 | * | |
2651 | * So for a change of @wl to the local @cpu with an overall group weight change | |
2652 | * of @wl we can compute the new shares distribution (s'_i) using: | |
2653 | * | |
2654 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
2655 | * | |
2656 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
2657 | * differences in waking a task to CPU 0. The additional task changes the | |
2658 | * weight and shares distributions like: | |
2659 | * | |
2660 | * rw'_i = { 3, 4, 1, 0 } | |
2661 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
2662 | * | |
2663 | * We can then compute the difference in effective weight by using: | |
2664 | * | |
2665 | * dw_i = S * (s'_i - s_i) (3) | |
2666 | * | |
2667 | * Where 'S' is the group weight as seen by its parent. | |
2668 | * | |
2669 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
2670 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
2671 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 2672 | */ |
2069dd75 | 2673 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 2674 | { |
4be9daaa | 2675 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 2676 | |
cf5f0acf | 2677 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
2678 | return wl; |
2679 | ||
4be9daaa | 2680 | for_each_sched_entity(se) { |
cf5f0acf | 2681 | long w, W; |
4be9daaa | 2682 | |
977dda7c | 2683 | tg = se->my_q->tg; |
bb3469ac | 2684 | |
cf5f0acf PZ |
2685 | /* |
2686 | * W = @wg + \Sum rw_j | |
2687 | */ | |
2688 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 2689 | |
cf5f0acf PZ |
2690 | /* |
2691 | * w = rw_i + @wl | |
2692 | */ | |
2693 | w = se->my_q->load.weight + wl; | |
940959e9 | 2694 | |
cf5f0acf PZ |
2695 | /* |
2696 | * wl = S * s'_i; see (2) | |
2697 | */ | |
2698 | if (W > 0 && w < W) | |
2699 | wl = (w * tg->shares) / W; | |
977dda7c PT |
2700 | else |
2701 | wl = tg->shares; | |
940959e9 | 2702 | |
cf5f0acf PZ |
2703 | /* |
2704 | * Per the above, wl is the new se->load.weight value; since | |
2705 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
2706 | * calc_cfs_shares(). | |
2707 | */ | |
977dda7c PT |
2708 | if (wl < MIN_SHARES) |
2709 | wl = MIN_SHARES; | |
cf5f0acf PZ |
2710 | |
2711 | /* | |
2712 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
2713 | */ | |
977dda7c | 2714 | wl -= se->load.weight; |
cf5f0acf PZ |
2715 | |
2716 | /* | |
2717 | * Recursively apply this logic to all parent groups to compute | |
2718 | * the final effective load change on the root group. Since | |
2719 | * only the @tg group gets extra weight, all parent groups can | |
2720 | * only redistribute existing shares. @wl is the shift in shares | |
2721 | * resulting from this level per the above. | |
2722 | */ | |
4be9daaa | 2723 | wg = 0; |
4be9daaa | 2724 | } |
bb3469ac | 2725 | |
4be9daaa | 2726 | return wl; |
bb3469ac PZ |
2727 | } |
2728 | #else | |
4be9daaa | 2729 | |
83378269 PZ |
2730 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
2731 | unsigned long wl, unsigned long wg) | |
4be9daaa | 2732 | { |
83378269 | 2733 | return wl; |
bb3469ac | 2734 | } |
4be9daaa | 2735 | |
bb3469ac PZ |
2736 | #endif |
2737 | ||
c88d5910 | 2738 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 2739 | { |
e37b6a7b | 2740 | s64 this_load, load; |
c88d5910 | 2741 | int idx, this_cpu, prev_cpu; |
098fb9db | 2742 | unsigned long tl_per_task; |
c88d5910 | 2743 | struct task_group *tg; |
83378269 | 2744 | unsigned long weight; |
b3137bc8 | 2745 | int balanced; |
098fb9db | 2746 | |
c88d5910 PZ |
2747 | idx = sd->wake_idx; |
2748 | this_cpu = smp_processor_id(); | |
2749 | prev_cpu = task_cpu(p); | |
2750 | load = source_load(prev_cpu, idx); | |
2751 | this_load = target_load(this_cpu, idx); | |
098fb9db | 2752 | |
b3137bc8 MG |
2753 | /* |
2754 | * If sync wakeup then subtract the (maximum possible) | |
2755 | * effect of the currently running task from the load | |
2756 | * of the current CPU: | |
2757 | */ | |
83378269 PZ |
2758 | if (sync) { |
2759 | tg = task_group(current); | |
2760 | weight = current->se.load.weight; | |
2761 | ||
c88d5910 | 2762 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
2763 | load += effective_load(tg, prev_cpu, 0, -weight); |
2764 | } | |
b3137bc8 | 2765 | |
83378269 PZ |
2766 | tg = task_group(p); |
2767 | weight = p->se.load.weight; | |
b3137bc8 | 2768 | |
71a29aa7 PZ |
2769 | /* |
2770 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
2771 | * due to the sync cause above having dropped this_load to 0, we'll |
2772 | * always have an imbalance, but there's really nothing you can do | |
2773 | * about that, so that's good too. | |
71a29aa7 PZ |
2774 | * |
2775 | * Otherwise check if either cpus are near enough in load to allow this | |
2776 | * task to be woken on this_cpu. | |
2777 | */ | |
e37b6a7b PT |
2778 | if (this_load > 0) { |
2779 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
2780 | |
2781 | this_eff_load = 100; | |
2782 | this_eff_load *= power_of(prev_cpu); | |
2783 | this_eff_load *= this_load + | |
2784 | effective_load(tg, this_cpu, weight, weight); | |
2785 | ||
2786 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
2787 | prev_eff_load *= power_of(this_cpu); | |
2788 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
2789 | ||
2790 | balanced = this_eff_load <= prev_eff_load; | |
2791 | } else | |
2792 | balanced = true; | |
b3137bc8 | 2793 | |
098fb9db | 2794 | /* |
4ae7d5ce IM |
2795 | * If the currently running task will sleep within |
2796 | * a reasonable amount of time then attract this newly | |
2797 | * woken task: | |
098fb9db | 2798 | */ |
2fb7635c PZ |
2799 | if (sync && balanced) |
2800 | return 1; | |
098fb9db | 2801 | |
41acab88 | 2802 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
2803 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
2804 | ||
c88d5910 PZ |
2805 | if (balanced || |
2806 | (this_load <= load && | |
2807 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
2808 | /* |
2809 | * This domain has SD_WAKE_AFFINE and | |
2810 | * p is cache cold in this domain, and | |
2811 | * there is no bad imbalance. | |
2812 | */ | |
c88d5910 | 2813 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 2814 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
2815 | |
2816 | return 1; | |
2817 | } | |
2818 | return 0; | |
2819 | } | |
2820 | ||
aaee1203 PZ |
2821 | /* |
2822 | * find_idlest_group finds and returns the least busy CPU group within the | |
2823 | * domain. | |
2824 | */ | |
2825 | static struct sched_group * | |
78e7ed53 | 2826 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 2827 | int this_cpu, int load_idx) |
e7693a36 | 2828 | { |
b3bd3de6 | 2829 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 2830 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 2831 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 2832 | |
aaee1203 PZ |
2833 | do { |
2834 | unsigned long load, avg_load; | |
2835 | int local_group; | |
2836 | int i; | |
e7693a36 | 2837 | |
aaee1203 PZ |
2838 | /* Skip over this group if it has no CPUs allowed */ |
2839 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 2840 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
2841 | continue; |
2842 | ||
2843 | local_group = cpumask_test_cpu(this_cpu, | |
2844 | sched_group_cpus(group)); | |
2845 | ||
2846 | /* Tally up the load of all CPUs in the group */ | |
2847 | avg_load = 0; | |
2848 | ||
2849 | for_each_cpu(i, sched_group_cpus(group)) { | |
2850 | /* Bias balancing toward cpus of our domain */ | |
2851 | if (local_group) | |
2852 | load = source_load(i, load_idx); | |
2853 | else | |
2854 | load = target_load(i, load_idx); | |
2855 | ||
2856 | avg_load += load; | |
2857 | } | |
2858 | ||
2859 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2860 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
2861 | |
2862 | if (local_group) { | |
2863 | this_load = avg_load; | |
aaee1203 PZ |
2864 | } else if (avg_load < min_load) { |
2865 | min_load = avg_load; | |
2866 | idlest = group; | |
2867 | } | |
2868 | } while (group = group->next, group != sd->groups); | |
2869 | ||
2870 | if (!idlest || 100*this_load < imbalance*min_load) | |
2871 | return NULL; | |
2872 | return idlest; | |
2873 | } | |
2874 | ||
2875 | /* | |
2876 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
2877 | */ | |
2878 | static int | |
2879 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
2880 | { | |
2881 | unsigned long load, min_load = ULONG_MAX; | |
2882 | int idlest = -1; | |
2883 | int i; | |
2884 | ||
2885 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 2886 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
2887 | load = weighted_cpuload(i); |
2888 | ||
2889 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2890 | min_load = load; | |
2891 | idlest = i; | |
e7693a36 GH |
2892 | } |
2893 | } | |
2894 | ||
aaee1203 PZ |
2895 | return idlest; |
2896 | } | |
e7693a36 | 2897 | |
a50bde51 PZ |
2898 | /* |
2899 | * Try and locate an idle CPU in the sched_domain. | |
2900 | */ | |
99bd5e2f | 2901 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
2902 | { |
2903 | int cpu = smp_processor_id(); | |
2904 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 2905 | struct sched_domain *sd; |
37407ea7 LT |
2906 | struct sched_group *sg; |
2907 | int i; | |
a50bde51 PZ |
2908 | |
2909 | /* | |
99bd5e2f SS |
2910 | * If the task is going to be woken-up on this cpu and if it is |
2911 | * already idle, then it is the right target. | |
a50bde51 | 2912 | */ |
99bd5e2f SS |
2913 | if (target == cpu && idle_cpu(cpu)) |
2914 | return cpu; | |
2915 | ||
2916 | /* | |
2917 | * If the task is going to be woken-up on the cpu where it previously | |
2918 | * ran and if it is currently idle, then it the right target. | |
2919 | */ | |
2920 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 2921 | return prev_cpu; |
a50bde51 PZ |
2922 | |
2923 | /* | |
37407ea7 | 2924 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 2925 | */ |
518cd623 | 2926 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 2927 | for_each_lower_domain(sd) { |
37407ea7 LT |
2928 | sg = sd->groups; |
2929 | do { | |
2930 | if (!cpumask_intersects(sched_group_cpus(sg), | |
2931 | tsk_cpus_allowed(p))) | |
2932 | goto next; | |
2933 | ||
2934 | for_each_cpu(i, sched_group_cpus(sg)) { | |
2935 | if (!idle_cpu(i)) | |
2936 | goto next; | |
2937 | } | |
970e1789 | 2938 | |
37407ea7 LT |
2939 | target = cpumask_first_and(sched_group_cpus(sg), |
2940 | tsk_cpus_allowed(p)); | |
2941 | goto done; | |
2942 | next: | |
2943 | sg = sg->next; | |
2944 | } while (sg != sd->groups); | |
2945 | } | |
2946 | done: | |
a50bde51 PZ |
2947 | return target; |
2948 | } | |
2949 | ||
aaee1203 PZ |
2950 | /* |
2951 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2952 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2953 | * SD_BALANCE_EXEC. | |
2954 | * | |
2955 | * Balance, ie. select the least loaded group. | |
2956 | * | |
2957 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2958 | * | |
2959 | * preempt must be disabled. | |
2960 | */ | |
0017d735 | 2961 | static int |
7608dec2 | 2962 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 2963 | { |
29cd8bae | 2964 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
2965 | int cpu = smp_processor_id(); |
2966 | int prev_cpu = task_cpu(p); | |
2967 | int new_cpu = cpu; | |
99bd5e2f | 2968 | int want_affine = 0; |
5158f4e4 | 2969 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 2970 | |
29baa747 | 2971 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
2972 | return prev_cpu; |
2973 | ||
0763a660 | 2974 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 2975 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
2976 | want_affine = 1; |
2977 | new_cpu = prev_cpu; | |
2978 | } | |
aaee1203 | 2979 | |
dce840a0 | 2980 | rcu_read_lock(); |
aaee1203 | 2981 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
2982 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
2983 | continue; | |
2984 | ||
fe3bcfe1 | 2985 | /* |
99bd5e2f SS |
2986 | * If both cpu and prev_cpu are part of this domain, |
2987 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 2988 | */ |
99bd5e2f SS |
2989 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
2990 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
2991 | affine_sd = tmp; | |
29cd8bae | 2992 | break; |
f03542a7 | 2993 | } |
29cd8bae | 2994 | |
f03542a7 | 2995 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
2996 | sd = tmp; |
2997 | } | |
2998 | ||
8b911acd | 2999 | if (affine_sd) { |
f03542a7 | 3000 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
3001 | prev_cpu = cpu; |
3002 | ||
3003 | new_cpu = select_idle_sibling(p, prev_cpu); | |
3004 | goto unlock; | |
8b911acd | 3005 | } |
e7693a36 | 3006 | |
aaee1203 | 3007 | while (sd) { |
5158f4e4 | 3008 | int load_idx = sd->forkexec_idx; |
aaee1203 | 3009 | struct sched_group *group; |
c88d5910 | 3010 | int weight; |
098fb9db | 3011 | |
0763a660 | 3012 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
3013 | sd = sd->child; |
3014 | continue; | |
3015 | } | |
098fb9db | 3016 | |
5158f4e4 PZ |
3017 | if (sd_flag & SD_BALANCE_WAKE) |
3018 | load_idx = sd->wake_idx; | |
098fb9db | 3019 | |
5158f4e4 | 3020 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
3021 | if (!group) { |
3022 | sd = sd->child; | |
3023 | continue; | |
3024 | } | |
4ae7d5ce | 3025 | |
d7c33c49 | 3026 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
3027 | if (new_cpu == -1 || new_cpu == cpu) { |
3028 | /* Now try balancing at a lower domain level of cpu */ | |
3029 | sd = sd->child; | |
3030 | continue; | |
e7693a36 | 3031 | } |
aaee1203 PZ |
3032 | |
3033 | /* Now try balancing at a lower domain level of new_cpu */ | |
3034 | cpu = new_cpu; | |
669c55e9 | 3035 | weight = sd->span_weight; |
aaee1203 PZ |
3036 | sd = NULL; |
3037 | for_each_domain(cpu, tmp) { | |
669c55e9 | 3038 | if (weight <= tmp->span_weight) |
aaee1203 | 3039 | break; |
0763a660 | 3040 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
3041 | sd = tmp; |
3042 | } | |
3043 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 3044 | } |
dce840a0 PZ |
3045 | unlock: |
3046 | rcu_read_unlock(); | |
e7693a36 | 3047 | |
c88d5910 | 3048 | return new_cpu; |
e7693a36 GH |
3049 | } |
3050 | #endif /* CONFIG_SMP */ | |
3051 | ||
e52fb7c0 PZ |
3052 | static unsigned long |
3053 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
3054 | { |
3055 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
3056 | ||
3057 | /* | |
e52fb7c0 PZ |
3058 | * Since its curr running now, convert the gran from real-time |
3059 | * to virtual-time in his units. | |
13814d42 MG |
3060 | * |
3061 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
3062 | * they get preempted easier. That is, if 'se' < 'curr' then | |
3063 | * the resulting gran will be larger, therefore penalizing the | |
3064 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
3065 | * be smaller, again penalizing the lighter task. | |
3066 | * | |
3067 | * This is especially important for buddies when the leftmost | |
3068 | * task is higher priority than the buddy. | |
0bbd3336 | 3069 | */ |
f4ad9bd2 | 3070 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
3071 | } |
3072 | ||
464b7527 PZ |
3073 | /* |
3074 | * Should 'se' preempt 'curr'. | |
3075 | * | |
3076 | * |s1 | |
3077 | * |s2 | |
3078 | * |s3 | |
3079 | * g | |
3080 | * |<--->|c | |
3081 | * | |
3082 | * w(c, s1) = -1 | |
3083 | * w(c, s2) = 0 | |
3084 | * w(c, s3) = 1 | |
3085 | * | |
3086 | */ | |
3087 | static int | |
3088 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
3089 | { | |
3090 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
3091 | ||
3092 | if (vdiff <= 0) | |
3093 | return -1; | |
3094 | ||
e52fb7c0 | 3095 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
3096 | if (vdiff > gran) |
3097 | return 1; | |
3098 | ||
3099 | return 0; | |
3100 | } | |
3101 | ||
02479099 PZ |
3102 | static void set_last_buddy(struct sched_entity *se) |
3103 | { | |
69c80f3e VP |
3104 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3105 | return; | |
3106 | ||
3107 | for_each_sched_entity(se) | |
3108 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
3109 | } |
3110 | ||
3111 | static void set_next_buddy(struct sched_entity *se) | |
3112 | { | |
69c80f3e VP |
3113 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3114 | return; | |
3115 | ||
3116 | for_each_sched_entity(se) | |
3117 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
3118 | } |
3119 | ||
ac53db59 RR |
3120 | static void set_skip_buddy(struct sched_entity *se) |
3121 | { | |
69c80f3e VP |
3122 | for_each_sched_entity(se) |
3123 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
3124 | } |
3125 | ||
bf0f6f24 IM |
3126 | /* |
3127 | * Preempt the current task with a newly woken task if needed: | |
3128 | */ | |
5a9b86f6 | 3129 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
3130 | { |
3131 | struct task_struct *curr = rq->curr; | |
8651a86c | 3132 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 3133 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 3134 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 3135 | int next_buddy_marked = 0; |
bf0f6f24 | 3136 | |
4ae7d5ce IM |
3137 | if (unlikely(se == pse)) |
3138 | return; | |
3139 | ||
5238cdd3 | 3140 | /* |
ddcdf6e7 | 3141 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
3142 | * unconditionally check_prempt_curr() after an enqueue (which may have |
3143 | * lead to a throttle). This both saves work and prevents false | |
3144 | * next-buddy nomination below. | |
3145 | */ | |
3146 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3147 | return; | |
3148 | ||
2f36825b | 3149 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3150 | set_next_buddy(pse); |
2f36825b VP |
3151 | next_buddy_marked = 1; |
3152 | } | |
57fdc26d | 3153 | |
aec0a514 BR |
3154 | /* |
3155 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3156 | * wake up path. | |
5238cdd3 PT |
3157 | * |
3158 | * Note: this also catches the edge-case of curr being in a throttled | |
3159 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3160 | * enqueue of curr) will have resulted in resched being set. This | |
3161 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3162 | * below. | |
aec0a514 BR |
3163 | */ |
3164 | if (test_tsk_need_resched(curr)) | |
3165 | return; | |
3166 | ||
a2f5c9ab DH |
3167 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3168 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3169 | likely(p->policy != SCHED_IDLE)) | |
3170 | goto preempt; | |
3171 | ||
91c234b4 | 3172 | /* |
a2f5c9ab DH |
3173 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3174 | * is driven by the tick): | |
91c234b4 | 3175 | */ |
6bc912b7 | 3176 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 3177 | return; |
bf0f6f24 | 3178 | |
464b7527 | 3179 | find_matching_se(&se, &pse); |
9bbd7374 | 3180 | update_curr(cfs_rq_of(se)); |
002f128b | 3181 | BUG_ON(!pse); |
2f36825b VP |
3182 | if (wakeup_preempt_entity(se, pse) == 1) { |
3183 | /* | |
3184 | * Bias pick_next to pick the sched entity that is | |
3185 | * triggering this preemption. | |
3186 | */ | |
3187 | if (!next_buddy_marked) | |
3188 | set_next_buddy(pse); | |
3a7e73a2 | 3189 | goto preempt; |
2f36825b | 3190 | } |
464b7527 | 3191 | |
3a7e73a2 | 3192 | return; |
a65ac745 | 3193 | |
3a7e73a2 PZ |
3194 | preempt: |
3195 | resched_task(curr); | |
3196 | /* | |
3197 | * Only set the backward buddy when the current task is still | |
3198 | * on the rq. This can happen when a wakeup gets interleaved | |
3199 | * with schedule on the ->pre_schedule() or idle_balance() | |
3200 | * point, either of which can * drop the rq lock. | |
3201 | * | |
3202 | * Also, during early boot the idle thread is in the fair class, | |
3203 | * for obvious reasons its a bad idea to schedule back to it. | |
3204 | */ | |
3205 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3206 | return; | |
3207 | ||
3208 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3209 | set_last_buddy(se); | |
bf0f6f24 IM |
3210 | } |
3211 | ||
fb8d4724 | 3212 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3213 | { |
8f4d37ec | 3214 | struct task_struct *p; |
bf0f6f24 IM |
3215 | struct cfs_rq *cfs_rq = &rq->cfs; |
3216 | struct sched_entity *se; | |
3217 | ||
36ace27e | 3218 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3219 | return NULL; |
3220 | ||
3221 | do { | |
9948f4b2 | 3222 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3223 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3224 | cfs_rq = group_cfs_rq(se); |
3225 | } while (cfs_rq); | |
3226 | ||
8f4d37ec | 3227 | p = task_of(se); |
b39e66ea MG |
3228 | if (hrtick_enabled(rq)) |
3229 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
3230 | |
3231 | return p; | |
bf0f6f24 IM |
3232 | } |
3233 | ||
3234 | /* | |
3235 | * Account for a descheduled task: | |
3236 | */ | |
31ee529c | 3237 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
3238 | { |
3239 | struct sched_entity *se = &prev->se; | |
3240 | struct cfs_rq *cfs_rq; | |
3241 | ||
3242 | for_each_sched_entity(se) { | |
3243 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 3244 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
3245 | } |
3246 | } | |
3247 | ||
ac53db59 RR |
3248 | /* |
3249 | * sched_yield() is very simple | |
3250 | * | |
3251 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
3252 | */ | |
3253 | static void yield_task_fair(struct rq *rq) | |
3254 | { | |
3255 | struct task_struct *curr = rq->curr; | |
3256 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
3257 | struct sched_entity *se = &curr->se; | |
3258 | ||
3259 | /* | |
3260 | * Are we the only task in the tree? | |
3261 | */ | |
3262 | if (unlikely(rq->nr_running == 1)) | |
3263 | return; | |
3264 | ||
3265 | clear_buddies(cfs_rq, se); | |
3266 | ||
3267 | if (curr->policy != SCHED_BATCH) { | |
3268 | update_rq_clock(rq); | |
3269 | /* | |
3270 | * Update run-time statistics of the 'current'. | |
3271 | */ | |
3272 | update_curr(cfs_rq); | |
916671c0 MG |
3273 | /* |
3274 | * Tell update_rq_clock() that we've just updated, | |
3275 | * so we don't do microscopic update in schedule() | |
3276 | * and double the fastpath cost. | |
3277 | */ | |
3278 | rq->skip_clock_update = 1; | |
ac53db59 RR |
3279 | } |
3280 | ||
3281 | set_skip_buddy(se); | |
3282 | } | |
3283 | ||
d95f4122 MG |
3284 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
3285 | { | |
3286 | struct sched_entity *se = &p->se; | |
3287 | ||
5238cdd3 PT |
3288 | /* throttled hierarchies are not runnable */ |
3289 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
3290 | return false; |
3291 | ||
3292 | /* Tell the scheduler that we'd really like pse to run next. */ | |
3293 | set_next_buddy(se); | |
3294 | ||
d95f4122 MG |
3295 | yield_task_fair(rq); |
3296 | ||
3297 | return true; | |
3298 | } | |
3299 | ||
681f3e68 | 3300 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
3301 | /************************************************** |
3302 | * Fair scheduling class load-balancing methods: | |
3303 | */ | |
3304 | ||
ed387b78 HS |
3305 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3306 | ||
ddcdf6e7 | 3307 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 3308 | #define LBF_NEED_BREAK 0x02 |
88b8dac0 | 3309 | #define LBF_SOME_PINNED 0x04 |
ddcdf6e7 PZ |
3310 | |
3311 | struct lb_env { | |
3312 | struct sched_domain *sd; | |
3313 | ||
ddcdf6e7 | 3314 | struct rq *src_rq; |
85c1e7da | 3315 | int src_cpu; |
ddcdf6e7 PZ |
3316 | |
3317 | int dst_cpu; | |
3318 | struct rq *dst_rq; | |
3319 | ||
88b8dac0 SV |
3320 | struct cpumask *dst_grpmask; |
3321 | int new_dst_cpu; | |
ddcdf6e7 | 3322 | enum cpu_idle_type idle; |
bd939f45 | 3323 | long imbalance; |
b9403130 MW |
3324 | /* The set of CPUs under consideration for load-balancing */ |
3325 | struct cpumask *cpus; | |
3326 | ||
ddcdf6e7 | 3327 | unsigned int flags; |
367456c7 PZ |
3328 | |
3329 | unsigned int loop; | |
3330 | unsigned int loop_break; | |
3331 | unsigned int loop_max; | |
ddcdf6e7 PZ |
3332 | }; |
3333 | ||
1e3c88bd | 3334 | /* |
ddcdf6e7 | 3335 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
3336 | * Both runqueues must be locked. |
3337 | */ | |
ddcdf6e7 | 3338 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 3339 | { |
ddcdf6e7 PZ |
3340 | deactivate_task(env->src_rq, p, 0); |
3341 | set_task_cpu(p, env->dst_cpu); | |
3342 | activate_task(env->dst_rq, p, 0); | |
3343 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
3344 | } |
3345 | ||
029632fb PZ |
3346 | /* |
3347 | * Is this task likely cache-hot: | |
3348 | */ | |
3349 | static int | |
3350 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
3351 | { | |
3352 | s64 delta; | |
3353 | ||
3354 | if (p->sched_class != &fair_sched_class) | |
3355 | return 0; | |
3356 | ||
3357 | if (unlikely(p->policy == SCHED_IDLE)) | |
3358 | return 0; | |
3359 | ||
3360 | /* | |
3361 | * Buddy candidates are cache hot: | |
3362 | */ | |
3363 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
3364 | (&p->se == cfs_rq_of(&p->se)->next || | |
3365 | &p->se == cfs_rq_of(&p->se)->last)) | |
3366 | return 1; | |
3367 | ||
3368 | if (sysctl_sched_migration_cost == -1) | |
3369 | return 1; | |
3370 | if (sysctl_sched_migration_cost == 0) | |
3371 | return 0; | |
3372 | ||
3373 | delta = now - p->se.exec_start; | |
3374 | ||
3375 | return delta < (s64)sysctl_sched_migration_cost; | |
3376 | } | |
3377 | ||
1e3c88bd PZ |
3378 | /* |
3379 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3380 | */ | |
3381 | static | |
8e45cb54 | 3382 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
3383 | { |
3384 | int tsk_cache_hot = 0; | |
3385 | /* | |
3386 | * We do not migrate tasks that are: | |
3387 | * 1) running (obviously), or | |
3388 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3389 | * 3) are cache-hot on their current CPU. | |
3390 | */ | |
ddcdf6e7 | 3391 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
88b8dac0 SV |
3392 | int new_dst_cpu; |
3393 | ||
41acab88 | 3394 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 SV |
3395 | |
3396 | /* | |
3397 | * Remember if this task can be migrated to any other cpu in | |
3398 | * our sched_group. We may want to revisit it if we couldn't | |
3399 | * meet load balance goals by pulling other tasks on src_cpu. | |
3400 | * | |
3401 | * Also avoid computing new_dst_cpu if we have already computed | |
3402 | * one in current iteration. | |
3403 | */ | |
3404 | if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED)) | |
3405 | return 0; | |
3406 | ||
3407 | new_dst_cpu = cpumask_first_and(env->dst_grpmask, | |
3408 | tsk_cpus_allowed(p)); | |
3409 | if (new_dst_cpu < nr_cpu_ids) { | |
3410 | env->flags |= LBF_SOME_PINNED; | |
3411 | env->new_dst_cpu = new_dst_cpu; | |
3412 | } | |
1e3c88bd PZ |
3413 | return 0; |
3414 | } | |
88b8dac0 SV |
3415 | |
3416 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 3417 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 3418 | |
ddcdf6e7 | 3419 | if (task_running(env->src_rq, p)) { |
41acab88 | 3420 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
3421 | return 0; |
3422 | } | |
3423 | ||
3424 | /* | |
3425 | * Aggressive migration if: | |
3426 | * 1) task is cache cold, or | |
3427 | * 2) too many balance attempts have failed. | |
3428 | */ | |
3429 | ||
ddcdf6e7 | 3430 | tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd); |
1e3c88bd | 3431 | if (!tsk_cache_hot || |
8e45cb54 | 3432 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
1e3c88bd PZ |
3433 | #ifdef CONFIG_SCHEDSTATS |
3434 | if (tsk_cache_hot) { | |
8e45cb54 | 3435 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 3436 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
3437 | } |
3438 | #endif | |
3439 | return 1; | |
3440 | } | |
3441 | ||
3442 | if (tsk_cache_hot) { | |
41acab88 | 3443 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
3444 | return 0; |
3445 | } | |
3446 | return 1; | |
3447 | } | |
3448 | ||
897c395f PZ |
3449 | /* |
3450 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3451 | * part of active balancing operations within "domain". | |
3452 | * Returns 1 if successful and 0 otherwise. | |
3453 | * | |
3454 | * Called with both runqueues locked. | |
3455 | */ | |
8e45cb54 | 3456 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
3457 | { |
3458 | struct task_struct *p, *n; | |
897c395f | 3459 | |
367456c7 PZ |
3460 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
3461 | if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu)) | |
3462 | continue; | |
897c395f | 3463 | |
367456c7 PZ |
3464 | if (!can_migrate_task(p, env)) |
3465 | continue; | |
897c395f | 3466 | |
367456c7 PZ |
3467 | move_task(p, env); |
3468 | /* | |
3469 | * Right now, this is only the second place move_task() | |
3470 | * is called, so we can safely collect move_task() | |
3471 | * stats here rather than inside move_task(). | |
3472 | */ | |
3473 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
3474 | return 1; | |
897c395f | 3475 | } |
897c395f PZ |
3476 | return 0; |
3477 | } | |
3478 | ||
367456c7 PZ |
3479 | static unsigned long task_h_load(struct task_struct *p); |
3480 | ||
eb95308e PZ |
3481 | static const unsigned int sched_nr_migrate_break = 32; |
3482 | ||
5d6523eb | 3483 | /* |
bd939f45 | 3484 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
3485 | * this_rq, as part of a balancing operation within domain "sd". |
3486 | * Returns 1 if successful and 0 otherwise. | |
3487 | * | |
3488 | * Called with both runqueues locked. | |
3489 | */ | |
3490 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 3491 | { |
5d6523eb PZ |
3492 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
3493 | struct task_struct *p; | |
367456c7 PZ |
3494 | unsigned long load; |
3495 | int pulled = 0; | |
1e3c88bd | 3496 | |
bd939f45 | 3497 | if (env->imbalance <= 0) |
5d6523eb | 3498 | return 0; |
1e3c88bd | 3499 | |
5d6523eb PZ |
3500 | while (!list_empty(tasks)) { |
3501 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 3502 | |
367456c7 PZ |
3503 | env->loop++; |
3504 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 3505 | if (env->loop > env->loop_max) |
367456c7 | 3506 | break; |
5d6523eb PZ |
3507 | |
3508 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 3509 | if (env->loop > env->loop_break) { |
eb95308e | 3510 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 3511 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 3512 | break; |
a195f004 | 3513 | } |
1e3c88bd | 3514 | |
5d6523eb | 3515 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
367456c7 PZ |
3516 | goto next; |
3517 | ||
3518 | load = task_h_load(p); | |
5d6523eb | 3519 | |
eb95308e | 3520 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
3521 | goto next; |
3522 | ||
bd939f45 | 3523 | if ((load / 2) > env->imbalance) |
367456c7 | 3524 | goto next; |
1e3c88bd | 3525 | |
367456c7 PZ |
3526 | if (!can_migrate_task(p, env)) |
3527 | goto next; | |
1e3c88bd | 3528 | |
ddcdf6e7 | 3529 | move_task(p, env); |
ee00e66f | 3530 | pulled++; |
bd939f45 | 3531 | env->imbalance -= load; |
1e3c88bd PZ |
3532 | |
3533 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
3534 | /* |
3535 | * NEWIDLE balancing is a source of latency, so preemptible | |
3536 | * kernels will stop after the first task is pulled to minimize | |
3537 | * the critical section. | |
3538 | */ | |
5d6523eb | 3539 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 3540 | break; |
1e3c88bd PZ |
3541 | #endif |
3542 | ||
ee00e66f PZ |
3543 | /* |
3544 | * We only want to steal up to the prescribed amount of | |
3545 | * weighted load. | |
3546 | */ | |
bd939f45 | 3547 | if (env->imbalance <= 0) |
ee00e66f | 3548 | break; |
367456c7 PZ |
3549 | |
3550 | continue; | |
3551 | next: | |
5d6523eb | 3552 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 3553 | } |
5d6523eb | 3554 | |
1e3c88bd | 3555 | /* |
ddcdf6e7 PZ |
3556 | * Right now, this is one of only two places move_task() is called, |
3557 | * so we can safely collect move_task() stats here rather than | |
3558 | * inside move_task(). | |
1e3c88bd | 3559 | */ |
8e45cb54 | 3560 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 3561 | |
5d6523eb | 3562 | return pulled; |
1e3c88bd PZ |
3563 | } |
3564 | ||
230059de | 3565 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
3566 | /* |
3567 | * update tg->load_weight by folding this cpu's load_avg | |
3568 | */ | |
67e86250 | 3569 | static int update_shares_cpu(struct task_group *tg, int cpu) |
9e3081ca PZ |
3570 | { |
3571 | struct cfs_rq *cfs_rq; | |
3572 | unsigned long flags; | |
3573 | struct rq *rq; | |
9e3081ca PZ |
3574 | |
3575 | if (!tg->se[cpu]) | |
3576 | return 0; | |
3577 | ||
3578 | rq = cpu_rq(cpu); | |
3579 | cfs_rq = tg->cfs_rq[cpu]; | |
3580 | ||
3581 | raw_spin_lock_irqsave(&rq->lock, flags); | |
3582 | ||
3583 | update_rq_clock(rq); | |
d6b55918 | 3584 | update_cfs_load(cfs_rq, 1); |
9ee474f5 | 3585 | update_cfs_rq_blocked_load(cfs_rq); |
9e3081ca PZ |
3586 | |
3587 | /* | |
3588 | * We need to update shares after updating tg->load_weight in | |
3589 | * order to adjust the weight of groups with long running tasks. | |
3590 | */ | |
6d5ab293 | 3591 | update_cfs_shares(cfs_rq); |
9e3081ca PZ |
3592 | |
3593 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
3594 | ||
3595 | return 0; | |
3596 | } | |
3597 | ||
3598 | static void update_shares(int cpu) | |
3599 | { | |
3600 | struct cfs_rq *cfs_rq; | |
3601 | struct rq *rq = cpu_rq(cpu); | |
3602 | ||
3603 | rcu_read_lock(); | |
9763b67f PZ |
3604 | /* |
3605 | * Iterates the task_group tree in a bottom up fashion, see | |
3606 | * list_add_leaf_cfs_rq() for details. | |
3607 | */ | |
64660c86 PT |
3608 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
3609 | /* throttled entities do not contribute to load */ | |
3610 | if (throttled_hierarchy(cfs_rq)) | |
3611 | continue; | |
3612 | ||
67e86250 | 3613 | update_shares_cpu(cfs_rq->tg, cpu); |
64660c86 | 3614 | } |
9e3081ca PZ |
3615 | rcu_read_unlock(); |
3616 | } | |
3617 | ||
9763b67f PZ |
3618 | /* |
3619 | * Compute the cpu's hierarchical load factor for each task group. | |
3620 | * This needs to be done in a top-down fashion because the load of a child | |
3621 | * group is a fraction of its parents load. | |
3622 | */ | |
3623 | static int tg_load_down(struct task_group *tg, void *data) | |
3624 | { | |
3625 | unsigned long load; | |
3626 | long cpu = (long)data; | |
3627 | ||
3628 | if (!tg->parent) { | |
3629 | load = cpu_rq(cpu)->load.weight; | |
3630 | } else { | |
3631 | load = tg->parent->cfs_rq[cpu]->h_load; | |
3632 | load *= tg->se[cpu]->load.weight; | |
3633 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
3634 | } | |
3635 | ||
3636 | tg->cfs_rq[cpu]->h_load = load; | |
3637 | ||
3638 | return 0; | |
3639 | } | |
3640 | ||
3641 | static void update_h_load(long cpu) | |
3642 | { | |
a35b6466 PZ |
3643 | struct rq *rq = cpu_rq(cpu); |
3644 | unsigned long now = jiffies; | |
3645 | ||
3646 | if (rq->h_load_throttle == now) | |
3647 | return; | |
3648 | ||
3649 | rq->h_load_throttle = now; | |
3650 | ||
367456c7 | 3651 | rcu_read_lock(); |
9763b67f | 3652 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
367456c7 | 3653 | rcu_read_unlock(); |
9763b67f PZ |
3654 | } |
3655 | ||
367456c7 | 3656 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 3657 | { |
367456c7 PZ |
3658 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
3659 | unsigned long load; | |
230059de | 3660 | |
367456c7 PZ |
3661 | load = p->se.load.weight; |
3662 | load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1); | |
230059de | 3663 | |
367456c7 | 3664 | return load; |
230059de PZ |
3665 | } |
3666 | #else | |
9e3081ca PZ |
3667 | static inline void update_shares(int cpu) |
3668 | { | |
3669 | } | |
3670 | ||
367456c7 | 3671 | static inline void update_h_load(long cpu) |
230059de | 3672 | { |
230059de | 3673 | } |
230059de | 3674 | |
367456c7 | 3675 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 3676 | { |
367456c7 | 3677 | return p->se.load.weight; |
1e3c88bd | 3678 | } |
230059de | 3679 | #endif |
1e3c88bd | 3680 | |
1e3c88bd PZ |
3681 | /********** Helpers for find_busiest_group ************************/ |
3682 | /* | |
3683 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
3684 | * during load balancing. | |
3685 | */ | |
3686 | struct sd_lb_stats { | |
3687 | struct sched_group *busiest; /* Busiest group in this sd */ | |
3688 | struct sched_group *this; /* Local group in this sd */ | |
3689 | unsigned long total_load; /* Total load of all groups in sd */ | |
3690 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
3691 | unsigned long avg_load; /* Average load across all groups in sd */ | |
3692 | ||
3693 | /** Statistics of this group */ | |
3694 | unsigned long this_load; | |
3695 | unsigned long this_load_per_task; | |
3696 | unsigned long this_nr_running; | |
fab47622 | 3697 | unsigned long this_has_capacity; |
aae6d3dd | 3698 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
3699 | |
3700 | /* Statistics of the busiest group */ | |
aae6d3dd | 3701 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
3702 | unsigned long max_load; |
3703 | unsigned long busiest_load_per_task; | |
3704 | unsigned long busiest_nr_running; | |
dd5feea1 | 3705 | unsigned long busiest_group_capacity; |
fab47622 | 3706 | unsigned long busiest_has_capacity; |
aae6d3dd | 3707 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
3708 | |
3709 | int group_imb; /* Is there imbalance in this sd */ | |
1e3c88bd PZ |
3710 | }; |
3711 | ||
3712 | /* | |
3713 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
3714 | */ | |
3715 | struct sg_lb_stats { | |
3716 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
3717 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
3718 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
3719 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
3720 | unsigned long group_capacity; | |
aae6d3dd SS |
3721 | unsigned long idle_cpus; |
3722 | unsigned long group_weight; | |
1e3c88bd | 3723 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 3724 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
3725 | }; |
3726 | ||
1e3c88bd PZ |
3727 | /** |
3728 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
3729 | * @sd: The sched_domain whose load_idx is to be obtained. | |
3730 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
3731 | */ | |
3732 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
3733 | enum cpu_idle_type idle) | |
3734 | { | |
3735 | int load_idx; | |
3736 | ||
3737 | switch (idle) { | |
3738 | case CPU_NOT_IDLE: | |
3739 | load_idx = sd->busy_idx; | |
3740 | break; | |
3741 | ||
3742 | case CPU_NEWLY_IDLE: | |
3743 | load_idx = sd->newidle_idx; | |
3744 | break; | |
3745 | default: | |
3746 | load_idx = sd->idle_idx; | |
3747 | break; | |
3748 | } | |
3749 | ||
3750 | return load_idx; | |
3751 | } | |
3752 | ||
1e3c88bd PZ |
3753 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
3754 | { | |
1399fa78 | 3755 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
3756 | } |
3757 | ||
3758 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
3759 | { | |
3760 | return default_scale_freq_power(sd, cpu); | |
3761 | } | |
3762 | ||
3763 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
3764 | { | |
669c55e9 | 3765 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
3766 | unsigned long smt_gain = sd->smt_gain; |
3767 | ||
3768 | smt_gain /= weight; | |
3769 | ||
3770 | return smt_gain; | |
3771 | } | |
3772 | ||
3773 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
3774 | { | |
3775 | return default_scale_smt_power(sd, cpu); | |
3776 | } | |
3777 | ||
3778 | unsigned long scale_rt_power(int cpu) | |
3779 | { | |
3780 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 3781 | u64 total, available, age_stamp, avg; |
1e3c88bd | 3782 | |
b654f7de PZ |
3783 | /* |
3784 | * Since we're reading these variables without serialization make sure | |
3785 | * we read them once before doing sanity checks on them. | |
3786 | */ | |
3787 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
3788 | avg = ACCESS_ONCE(rq->rt_avg); | |
3789 | ||
3790 | total = sched_avg_period() + (rq->clock - age_stamp); | |
aa483808 | 3791 | |
b654f7de | 3792 | if (unlikely(total < avg)) { |
aa483808 VP |
3793 | /* Ensures that power won't end up being negative */ |
3794 | available = 0; | |
3795 | } else { | |
b654f7de | 3796 | available = total - avg; |
aa483808 | 3797 | } |
1e3c88bd | 3798 | |
1399fa78 NR |
3799 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
3800 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 3801 | |
1399fa78 | 3802 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3803 | |
3804 | return div_u64(available, total); | |
3805 | } | |
3806 | ||
3807 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
3808 | { | |
669c55e9 | 3809 | unsigned long weight = sd->span_weight; |
1399fa78 | 3810 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
3811 | struct sched_group *sdg = sd->groups; |
3812 | ||
1e3c88bd PZ |
3813 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
3814 | if (sched_feat(ARCH_POWER)) | |
3815 | power *= arch_scale_smt_power(sd, cpu); | |
3816 | else | |
3817 | power *= default_scale_smt_power(sd, cpu); | |
3818 | ||
1399fa78 | 3819 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3820 | } |
3821 | ||
9c3f75cb | 3822 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
3823 | |
3824 | if (sched_feat(ARCH_POWER)) | |
3825 | power *= arch_scale_freq_power(sd, cpu); | |
3826 | else | |
3827 | power *= default_scale_freq_power(sd, cpu); | |
3828 | ||
1399fa78 | 3829 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 3830 | |
1e3c88bd | 3831 | power *= scale_rt_power(cpu); |
1399fa78 | 3832 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3833 | |
3834 | if (!power) | |
3835 | power = 1; | |
3836 | ||
e51fd5e2 | 3837 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 3838 | sdg->sgp->power = power; |
1e3c88bd PZ |
3839 | } |
3840 | ||
029632fb | 3841 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
3842 | { |
3843 | struct sched_domain *child = sd->child; | |
3844 | struct sched_group *group, *sdg = sd->groups; | |
3845 | unsigned long power; | |
4ec4412e VG |
3846 | unsigned long interval; |
3847 | ||
3848 | interval = msecs_to_jiffies(sd->balance_interval); | |
3849 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
3850 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
3851 | |
3852 | if (!child) { | |
3853 | update_cpu_power(sd, cpu); | |
3854 | return; | |
3855 | } | |
3856 | ||
3857 | power = 0; | |
3858 | ||
74a5ce20 PZ |
3859 | if (child->flags & SD_OVERLAP) { |
3860 | /* | |
3861 | * SD_OVERLAP domains cannot assume that child groups | |
3862 | * span the current group. | |
3863 | */ | |
3864 | ||
3865 | for_each_cpu(cpu, sched_group_cpus(sdg)) | |
3866 | power += power_of(cpu); | |
3867 | } else { | |
3868 | /* | |
3869 | * !SD_OVERLAP domains can assume that child groups | |
3870 | * span the current group. | |
3871 | */ | |
3872 | ||
3873 | group = child->groups; | |
3874 | do { | |
3875 | power += group->sgp->power; | |
3876 | group = group->next; | |
3877 | } while (group != child->groups); | |
3878 | } | |
1e3c88bd | 3879 | |
c3decf0d | 3880 | sdg->sgp->power_orig = sdg->sgp->power = power; |
1e3c88bd PZ |
3881 | } |
3882 | ||
9d5efe05 SV |
3883 | /* |
3884 | * Try and fix up capacity for tiny siblings, this is needed when | |
3885 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
3886 | * which on its own isn't powerful enough. | |
3887 | * | |
3888 | * See update_sd_pick_busiest() and check_asym_packing(). | |
3889 | */ | |
3890 | static inline int | |
3891 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
3892 | { | |
3893 | /* | |
1399fa78 | 3894 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 3895 | */ |
a6c75f2f | 3896 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
3897 | return 0; |
3898 | ||
3899 | /* | |
3900 | * If ~90% of the cpu_power is still there, we're good. | |
3901 | */ | |
9c3f75cb | 3902 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
3903 | return 1; |
3904 | ||
3905 | return 0; | |
3906 | } | |
3907 | ||
1e3c88bd PZ |
3908 | /** |
3909 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 3910 | * @env: The load balancing environment. |
1e3c88bd | 3911 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 3912 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 3913 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
3914 | * @balance: Should we balance. |
3915 | * @sgs: variable to hold the statistics for this group. | |
3916 | */ | |
bd939f45 PZ |
3917 | static inline void update_sg_lb_stats(struct lb_env *env, |
3918 | struct sched_group *group, int load_idx, | |
b9403130 | 3919 | int local_group, int *balance, struct sg_lb_stats *sgs) |
1e3c88bd | 3920 | { |
e44bc5c5 PZ |
3921 | unsigned long nr_running, max_nr_running, min_nr_running; |
3922 | unsigned long load, max_cpu_load, min_cpu_load; | |
04f733b4 | 3923 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
dd5feea1 | 3924 | unsigned long avg_load_per_task = 0; |
bd939f45 | 3925 | int i; |
1e3c88bd | 3926 | |
871e35bc | 3927 | if (local_group) |
c1174876 | 3928 | balance_cpu = group_balance_cpu(group); |
1e3c88bd PZ |
3929 | |
3930 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
3931 | max_cpu_load = 0; |
3932 | min_cpu_load = ~0UL; | |
2582f0eb | 3933 | max_nr_running = 0; |
e44bc5c5 | 3934 | min_nr_running = ~0UL; |
1e3c88bd | 3935 | |
b9403130 | 3936 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
3937 | struct rq *rq = cpu_rq(i); |
3938 | ||
e44bc5c5 PZ |
3939 | nr_running = rq->nr_running; |
3940 | ||
1e3c88bd PZ |
3941 | /* Bias balancing toward cpus of our domain */ |
3942 | if (local_group) { | |
c1174876 PZ |
3943 | if (idle_cpu(i) && !first_idle_cpu && |
3944 | cpumask_test_cpu(i, sched_group_mask(group))) { | |
04f733b4 | 3945 | first_idle_cpu = 1; |
1e3c88bd PZ |
3946 | balance_cpu = i; |
3947 | } | |
04f733b4 PZ |
3948 | |
3949 | load = target_load(i, load_idx); | |
1e3c88bd PZ |
3950 | } else { |
3951 | load = source_load(i, load_idx); | |
e44bc5c5 | 3952 | if (load > max_cpu_load) |
1e3c88bd PZ |
3953 | max_cpu_load = load; |
3954 | if (min_cpu_load > load) | |
3955 | min_cpu_load = load; | |
e44bc5c5 PZ |
3956 | |
3957 | if (nr_running > max_nr_running) | |
3958 | max_nr_running = nr_running; | |
3959 | if (min_nr_running > nr_running) | |
3960 | min_nr_running = nr_running; | |
1e3c88bd PZ |
3961 | } |
3962 | ||
3963 | sgs->group_load += load; | |
e44bc5c5 | 3964 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 3965 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
3966 | if (idle_cpu(i)) |
3967 | sgs->idle_cpus++; | |
1e3c88bd PZ |
3968 | } |
3969 | ||
3970 | /* | |
3971 | * First idle cpu or the first cpu(busiest) in this sched group | |
3972 | * is eligible for doing load balancing at this and above | |
3973 | * domains. In the newly idle case, we will allow all the cpu's | |
3974 | * to do the newly idle load balance. | |
3975 | */ | |
4ec4412e | 3976 | if (local_group) { |
bd939f45 | 3977 | if (env->idle != CPU_NEWLY_IDLE) { |
04f733b4 | 3978 | if (balance_cpu != env->dst_cpu) { |
4ec4412e VG |
3979 | *balance = 0; |
3980 | return; | |
3981 | } | |
bd939f45 | 3982 | update_group_power(env->sd, env->dst_cpu); |
4ec4412e | 3983 | } else if (time_after_eq(jiffies, group->sgp->next_update)) |
bd939f45 | 3984 | update_group_power(env->sd, env->dst_cpu); |
1e3c88bd PZ |
3985 | } |
3986 | ||
3987 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3988 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 3989 | |
1e3c88bd PZ |
3990 | /* |
3991 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 3992 | * than the average weight of a task. |
1e3c88bd PZ |
3993 | * |
3994 | * APZ: with cgroup the avg task weight can vary wildly and | |
3995 | * might not be a suitable number - should we keep a | |
3996 | * normalized nr_running number somewhere that negates | |
3997 | * the hierarchy? | |
3998 | */ | |
dd5feea1 SS |
3999 | if (sgs->sum_nr_running) |
4000 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 4001 | |
e44bc5c5 PZ |
4002 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && |
4003 | (max_nr_running - min_nr_running) > 1) | |
1e3c88bd PZ |
4004 | sgs->group_imb = 1; |
4005 | ||
9c3f75cb | 4006 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 4007 | SCHED_POWER_SCALE); |
9d5efe05 | 4008 | if (!sgs->group_capacity) |
bd939f45 | 4009 | sgs->group_capacity = fix_small_capacity(env->sd, group); |
aae6d3dd | 4010 | sgs->group_weight = group->group_weight; |
fab47622 NR |
4011 | |
4012 | if (sgs->group_capacity > sgs->sum_nr_running) | |
4013 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
4014 | } |
4015 | ||
532cb4c4 MN |
4016 | /** |
4017 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 4018 | * @env: The load balancing environment. |
532cb4c4 MN |
4019 | * @sds: sched_domain statistics |
4020 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 4021 | * @sgs: sched_group statistics |
532cb4c4 MN |
4022 | * |
4023 | * Determine if @sg is a busier group than the previously selected | |
4024 | * busiest group. | |
4025 | */ | |
bd939f45 | 4026 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
4027 | struct sd_lb_stats *sds, |
4028 | struct sched_group *sg, | |
bd939f45 | 4029 | struct sg_lb_stats *sgs) |
532cb4c4 MN |
4030 | { |
4031 | if (sgs->avg_load <= sds->max_load) | |
4032 | return false; | |
4033 | ||
4034 | if (sgs->sum_nr_running > sgs->group_capacity) | |
4035 | return true; | |
4036 | ||
4037 | if (sgs->group_imb) | |
4038 | return true; | |
4039 | ||
4040 | /* | |
4041 | * ASYM_PACKING needs to move all the work to the lowest | |
4042 | * numbered CPUs in the group, therefore mark all groups | |
4043 | * higher than ourself as busy. | |
4044 | */ | |
bd939f45 PZ |
4045 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
4046 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
4047 | if (!sds->busiest) |
4048 | return true; | |
4049 | ||
4050 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
4051 | return true; | |
4052 | } | |
4053 | ||
4054 | return false; | |
4055 | } | |
4056 | ||
1e3c88bd | 4057 | /** |
461819ac | 4058 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 4059 | * @env: The load balancing environment. |
1e3c88bd PZ |
4060 | * @balance: Should we balance. |
4061 | * @sds: variable to hold the statistics for this sched_domain. | |
4062 | */ | |
bd939f45 | 4063 | static inline void update_sd_lb_stats(struct lb_env *env, |
b9403130 | 4064 | int *balance, struct sd_lb_stats *sds) |
1e3c88bd | 4065 | { |
bd939f45 PZ |
4066 | struct sched_domain *child = env->sd->child; |
4067 | struct sched_group *sg = env->sd->groups; | |
1e3c88bd PZ |
4068 | struct sg_lb_stats sgs; |
4069 | int load_idx, prefer_sibling = 0; | |
4070 | ||
4071 | if (child && child->flags & SD_PREFER_SIBLING) | |
4072 | prefer_sibling = 1; | |
4073 | ||
bd939f45 | 4074 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
4075 | |
4076 | do { | |
4077 | int local_group; | |
4078 | ||
bd939f45 | 4079 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
1e3c88bd | 4080 | memset(&sgs, 0, sizeof(sgs)); |
b9403130 | 4081 | update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs); |
1e3c88bd | 4082 | |
8f190fb3 | 4083 | if (local_group && !(*balance)) |
1e3c88bd PZ |
4084 | return; |
4085 | ||
4086 | sds->total_load += sgs.group_load; | |
9c3f75cb | 4087 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
4088 | |
4089 | /* | |
4090 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 4091 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
4092 | * and move all the excess tasks away. We lower the capacity |
4093 | * of a group only if the local group has the capacity to fit | |
4094 | * these excess tasks, i.e. nr_running < group_capacity. The | |
4095 | * extra check prevents the case where you always pull from the | |
4096 | * heaviest group when it is already under-utilized (possible | |
4097 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 4098 | */ |
75dd321d | 4099 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
4100 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
4101 | ||
4102 | if (local_group) { | |
4103 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 4104 | sds->this = sg; |
1e3c88bd PZ |
4105 | sds->this_nr_running = sgs.sum_nr_running; |
4106 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 4107 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4108 | sds->this_idle_cpus = sgs.idle_cpus; |
bd939f45 | 4109 | } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) { |
1e3c88bd | 4110 | sds->max_load = sgs.avg_load; |
532cb4c4 | 4111 | sds->busiest = sg; |
1e3c88bd | 4112 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 4113 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 4114 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 4115 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 4116 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4117 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
4118 | sds->group_imb = sgs.group_imb; |
4119 | } | |
4120 | ||
532cb4c4 | 4121 | sg = sg->next; |
bd939f45 | 4122 | } while (sg != env->sd->groups); |
532cb4c4 MN |
4123 | } |
4124 | ||
532cb4c4 MN |
4125 | /** |
4126 | * check_asym_packing - Check to see if the group is packed into the | |
4127 | * sched doman. | |
4128 | * | |
4129 | * This is primarily intended to used at the sibling level. Some | |
4130 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
4131 | * case of POWER7, it can move to lower SMT modes only when higher | |
4132 | * threads are idle. When in lower SMT modes, the threads will | |
4133 | * perform better since they share less core resources. Hence when we | |
4134 | * have idle threads, we want them to be the higher ones. | |
4135 | * | |
4136 | * This packing function is run on idle threads. It checks to see if | |
4137 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
4138 | * CPU number than the packing function is being run on. Here we are | |
4139 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
4140 | * number. | |
4141 | * | |
b6b12294 MN |
4142 | * Returns 1 when packing is required and a task should be moved to |
4143 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
4144 | * | |
cd96891d | 4145 | * @env: The load balancing environment. |
532cb4c4 | 4146 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 4147 | */ |
bd939f45 | 4148 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
4149 | { |
4150 | int busiest_cpu; | |
4151 | ||
bd939f45 | 4152 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
4153 | return 0; |
4154 | ||
4155 | if (!sds->busiest) | |
4156 | return 0; | |
4157 | ||
4158 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 4159 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
4160 | return 0; |
4161 | ||
bd939f45 PZ |
4162 | env->imbalance = DIV_ROUND_CLOSEST( |
4163 | sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE); | |
4164 | ||
532cb4c4 | 4165 | return 1; |
1e3c88bd PZ |
4166 | } |
4167 | ||
4168 | /** | |
4169 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
4170 | * amongst the groups of a sched_domain, during | |
4171 | * load balancing. | |
cd96891d | 4172 | * @env: The load balancing environment. |
1e3c88bd | 4173 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4174 | */ |
bd939f45 PZ |
4175 | static inline |
4176 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
4177 | { |
4178 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
4179 | unsigned int imbn = 2; | |
dd5feea1 | 4180 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
4181 | |
4182 | if (sds->this_nr_running) { | |
4183 | sds->this_load_per_task /= sds->this_nr_running; | |
4184 | if (sds->busiest_load_per_task > | |
4185 | sds->this_load_per_task) | |
4186 | imbn = 1; | |
bd939f45 | 4187 | } else { |
1e3c88bd | 4188 | sds->this_load_per_task = |
bd939f45 PZ |
4189 | cpu_avg_load_per_task(env->dst_cpu); |
4190 | } | |
1e3c88bd | 4191 | |
dd5feea1 | 4192 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 4193 | * SCHED_POWER_SCALE; |
9c3f75cb | 4194 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
4195 | |
4196 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
4197 | (scaled_busy_load_per_task * imbn)) { | |
bd939f45 | 4198 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4199 | return; |
4200 | } | |
4201 | ||
4202 | /* | |
4203 | * OK, we don't have enough imbalance to justify moving tasks, | |
4204 | * however we may be able to increase total CPU power used by | |
4205 | * moving them. | |
4206 | */ | |
4207 | ||
9c3f75cb | 4208 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 4209 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 4210 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 4211 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 4212 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4213 | |
4214 | /* Amount of load we'd subtract */ | |
1399fa78 | 4215 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 4216 | sds->busiest->sgp->power; |
1e3c88bd | 4217 | if (sds->max_load > tmp) |
9c3f75cb | 4218 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
4219 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
4220 | ||
4221 | /* Amount of load we'd add */ | |
9c3f75cb | 4222 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 4223 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
4224 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
4225 | sds->this->sgp->power; | |
1e3c88bd | 4226 | else |
1399fa78 | 4227 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
4228 | sds->this->sgp->power; |
4229 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 4230 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 4231 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4232 | |
4233 | /* Move if we gain throughput */ | |
4234 | if (pwr_move > pwr_now) | |
bd939f45 | 4235 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4236 | } |
4237 | ||
4238 | /** | |
4239 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
4240 | * groups of a given sched_domain during load balance. | |
bd939f45 | 4241 | * @env: load balance environment |
1e3c88bd | 4242 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4243 | */ |
bd939f45 | 4244 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 4245 | { |
dd5feea1 SS |
4246 | unsigned long max_pull, load_above_capacity = ~0UL; |
4247 | ||
4248 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
4249 | if (sds->group_imb) { | |
4250 | sds->busiest_load_per_task = | |
4251 | min(sds->busiest_load_per_task, sds->avg_load); | |
4252 | } | |
4253 | ||
1e3c88bd PZ |
4254 | /* |
4255 | * In the presence of smp nice balancing, certain scenarios can have | |
4256 | * max load less than avg load(as we skip the groups at or below | |
4257 | * its cpu_power, while calculating max_load..) | |
4258 | */ | |
4259 | if (sds->max_load < sds->avg_load) { | |
bd939f45 PZ |
4260 | env->imbalance = 0; |
4261 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4262 | } |
4263 | ||
dd5feea1 SS |
4264 | if (!sds->group_imb) { |
4265 | /* | |
4266 | * Don't want to pull so many tasks that a group would go idle. | |
4267 | */ | |
4268 | load_above_capacity = (sds->busiest_nr_running - | |
4269 | sds->busiest_group_capacity); | |
4270 | ||
1399fa78 | 4271 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 4272 | |
9c3f75cb | 4273 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
4274 | } |
4275 | ||
4276 | /* | |
4277 | * We're trying to get all the cpus to the average_load, so we don't | |
4278 | * want to push ourselves above the average load, nor do we wish to | |
4279 | * reduce the max loaded cpu below the average load. At the same time, | |
4280 | * we also don't want to reduce the group load below the group capacity | |
4281 | * (so that we can implement power-savings policies etc). Thus we look | |
4282 | * for the minimum possible imbalance. | |
4283 | * Be careful of negative numbers as they'll appear as very large values | |
4284 | * with unsigned longs. | |
4285 | */ | |
4286 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
4287 | |
4288 | /* How much load to actually move to equalise the imbalance */ | |
bd939f45 | 4289 | env->imbalance = min(max_pull * sds->busiest->sgp->power, |
9c3f75cb | 4290 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) |
1399fa78 | 4291 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
4292 | |
4293 | /* | |
4294 | * if *imbalance is less than the average load per runnable task | |
25985edc | 4295 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
4296 | * a think about bumping its value to force at least one task to be |
4297 | * moved | |
4298 | */ | |
bd939f45 PZ |
4299 | if (env->imbalance < sds->busiest_load_per_task) |
4300 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4301 | |
4302 | } | |
fab47622 | 4303 | |
1e3c88bd PZ |
4304 | /******* find_busiest_group() helpers end here *********************/ |
4305 | ||
4306 | /** | |
4307 | * find_busiest_group - Returns the busiest group within the sched_domain | |
4308 | * if there is an imbalance. If there isn't an imbalance, and | |
4309 | * the user has opted for power-savings, it returns a group whose | |
4310 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
4311 | * such a group exists. | |
4312 | * | |
4313 | * Also calculates the amount of weighted load which should be moved | |
4314 | * to restore balance. | |
4315 | * | |
cd96891d | 4316 | * @env: The load balancing environment. |
1e3c88bd PZ |
4317 | * @balance: Pointer to a variable indicating if this_cpu |
4318 | * is the appropriate cpu to perform load balancing at this_level. | |
4319 | * | |
4320 | * Returns: - the busiest group if imbalance exists. | |
4321 | * - If no imbalance and user has opted for power-savings balance, | |
4322 | * return the least loaded group whose CPUs can be | |
4323 | * put to idle by rebalancing its tasks onto our group. | |
4324 | */ | |
4325 | static struct sched_group * | |
b9403130 | 4326 | find_busiest_group(struct lb_env *env, int *balance) |
1e3c88bd PZ |
4327 | { |
4328 | struct sd_lb_stats sds; | |
4329 | ||
4330 | memset(&sds, 0, sizeof(sds)); | |
4331 | ||
4332 | /* | |
4333 | * Compute the various statistics relavent for load balancing at | |
4334 | * this level. | |
4335 | */ | |
b9403130 | 4336 | update_sd_lb_stats(env, balance, &sds); |
1e3c88bd | 4337 | |
cc57aa8f PZ |
4338 | /* |
4339 | * this_cpu is not the appropriate cpu to perform load balancing at | |
4340 | * this level. | |
1e3c88bd | 4341 | */ |
8f190fb3 | 4342 | if (!(*balance)) |
1e3c88bd PZ |
4343 | goto ret; |
4344 | ||
bd939f45 PZ |
4345 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
4346 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
4347 | return sds.busiest; |
4348 | ||
cc57aa8f | 4349 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
4350 | if (!sds.busiest || sds.busiest_nr_running == 0) |
4351 | goto out_balanced; | |
4352 | ||
1399fa78 | 4353 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 4354 | |
866ab43e PZ |
4355 | /* |
4356 | * If the busiest group is imbalanced the below checks don't | |
4357 | * work because they assumes all things are equal, which typically | |
4358 | * isn't true due to cpus_allowed constraints and the like. | |
4359 | */ | |
4360 | if (sds.group_imb) | |
4361 | goto force_balance; | |
4362 | ||
cc57aa8f | 4363 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
bd939f45 | 4364 | if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
fab47622 NR |
4365 | !sds.busiest_has_capacity) |
4366 | goto force_balance; | |
4367 | ||
cc57aa8f PZ |
4368 | /* |
4369 | * If the local group is more busy than the selected busiest group | |
4370 | * don't try and pull any tasks. | |
4371 | */ | |
1e3c88bd PZ |
4372 | if (sds.this_load >= sds.max_load) |
4373 | goto out_balanced; | |
4374 | ||
cc57aa8f PZ |
4375 | /* |
4376 | * Don't pull any tasks if this group is already above the domain | |
4377 | * average load. | |
4378 | */ | |
1e3c88bd PZ |
4379 | if (sds.this_load >= sds.avg_load) |
4380 | goto out_balanced; | |
4381 | ||
bd939f45 | 4382 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
4383 | /* |
4384 | * This cpu is idle. If the busiest group load doesn't | |
4385 | * have more tasks than the number of available cpu's and | |
4386 | * there is no imbalance between this and busiest group | |
4387 | * wrt to idle cpu's, it is balanced. | |
4388 | */ | |
c186fafe | 4389 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
4390 | sds.busiest_nr_running <= sds.busiest_group_weight) |
4391 | goto out_balanced; | |
c186fafe PZ |
4392 | } else { |
4393 | /* | |
4394 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
4395 | * imbalance_pct to be conservative. | |
4396 | */ | |
bd939f45 | 4397 | if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load) |
c186fafe | 4398 | goto out_balanced; |
aae6d3dd | 4399 | } |
1e3c88bd | 4400 | |
fab47622 | 4401 | force_balance: |
1e3c88bd | 4402 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 4403 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
4404 | return sds.busiest; |
4405 | ||
4406 | out_balanced: | |
1e3c88bd | 4407 | ret: |
bd939f45 | 4408 | env->imbalance = 0; |
1e3c88bd PZ |
4409 | return NULL; |
4410 | } | |
4411 | ||
4412 | /* | |
4413 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
4414 | */ | |
bd939f45 | 4415 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 4416 | struct sched_group *group) |
1e3c88bd PZ |
4417 | { |
4418 | struct rq *busiest = NULL, *rq; | |
4419 | unsigned long max_load = 0; | |
4420 | int i; | |
4421 | ||
4422 | for_each_cpu(i, sched_group_cpus(group)) { | |
4423 | unsigned long power = power_of(i); | |
1399fa78 NR |
4424 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
4425 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
4426 | unsigned long wl; |
4427 | ||
9d5efe05 | 4428 | if (!capacity) |
bd939f45 | 4429 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 4430 | |
b9403130 | 4431 | if (!cpumask_test_cpu(i, env->cpus)) |
1e3c88bd PZ |
4432 | continue; |
4433 | ||
4434 | rq = cpu_rq(i); | |
6e40f5bb | 4435 | wl = weighted_cpuload(i); |
1e3c88bd | 4436 | |
6e40f5bb TG |
4437 | /* |
4438 | * When comparing with imbalance, use weighted_cpuload() | |
4439 | * which is not scaled with the cpu power. | |
4440 | */ | |
bd939f45 | 4441 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
4442 | continue; |
4443 | ||
6e40f5bb TG |
4444 | /* |
4445 | * For the load comparisons with the other cpu's, consider | |
4446 | * the weighted_cpuload() scaled with the cpu power, so that | |
4447 | * the load can be moved away from the cpu that is potentially | |
4448 | * running at a lower capacity. | |
4449 | */ | |
1399fa78 | 4450 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 4451 | |
1e3c88bd PZ |
4452 | if (wl > max_load) { |
4453 | max_load = wl; | |
4454 | busiest = rq; | |
4455 | } | |
4456 | } | |
4457 | ||
4458 | return busiest; | |
4459 | } | |
4460 | ||
4461 | /* | |
4462 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
4463 | * so long as it is large enough. | |
4464 | */ | |
4465 | #define MAX_PINNED_INTERVAL 512 | |
4466 | ||
4467 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
029632fb | 4468 | DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); |
1e3c88bd | 4469 | |
bd939f45 | 4470 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 4471 | { |
bd939f45 PZ |
4472 | struct sched_domain *sd = env->sd; |
4473 | ||
4474 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
4475 | |
4476 | /* | |
4477 | * ASYM_PACKING needs to force migrate tasks from busy but | |
4478 | * higher numbered CPUs in order to pack all tasks in the | |
4479 | * lowest numbered CPUs. | |
4480 | */ | |
bd939f45 | 4481 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 4482 | return 1; |
1af3ed3d PZ |
4483 | } |
4484 | ||
4485 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
4486 | } | |
4487 | ||
969c7921 TH |
4488 | static int active_load_balance_cpu_stop(void *data); |
4489 | ||
1e3c88bd PZ |
4490 | /* |
4491 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
4492 | * tasks if there is an imbalance. | |
4493 | */ | |
4494 | static int load_balance(int this_cpu, struct rq *this_rq, | |
4495 | struct sched_domain *sd, enum cpu_idle_type idle, | |
4496 | int *balance) | |
4497 | { | |
88b8dac0 SV |
4498 | int ld_moved, cur_ld_moved, active_balance = 0; |
4499 | int lb_iterations, max_lb_iterations; | |
1e3c88bd | 4500 | struct sched_group *group; |
1e3c88bd PZ |
4501 | struct rq *busiest; |
4502 | unsigned long flags; | |
4503 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
4504 | ||
8e45cb54 PZ |
4505 | struct lb_env env = { |
4506 | .sd = sd, | |
ddcdf6e7 PZ |
4507 | .dst_cpu = this_cpu, |
4508 | .dst_rq = this_rq, | |
88b8dac0 | 4509 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 4510 | .idle = idle, |
eb95308e | 4511 | .loop_break = sched_nr_migrate_break, |
b9403130 | 4512 | .cpus = cpus, |
8e45cb54 PZ |
4513 | }; |
4514 | ||
1e3c88bd | 4515 | cpumask_copy(cpus, cpu_active_mask); |
88b8dac0 | 4516 | max_lb_iterations = cpumask_weight(env.dst_grpmask); |
1e3c88bd | 4517 | |
1e3c88bd PZ |
4518 | schedstat_inc(sd, lb_count[idle]); |
4519 | ||
4520 | redo: | |
b9403130 | 4521 | group = find_busiest_group(&env, balance); |
1e3c88bd PZ |
4522 | |
4523 | if (*balance == 0) | |
4524 | goto out_balanced; | |
4525 | ||
4526 | if (!group) { | |
4527 | schedstat_inc(sd, lb_nobusyg[idle]); | |
4528 | goto out_balanced; | |
4529 | } | |
4530 | ||
b9403130 | 4531 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
4532 | if (!busiest) { |
4533 | schedstat_inc(sd, lb_nobusyq[idle]); | |
4534 | goto out_balanced; | |
4535 | } | |
4536 | ||
78feefc5 | 4537 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 4538 | |
bd939f45 | 4539 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
4540 | |
4541 | ld_moved = 0; | |
88b8dac0 | 4542 | lb_iterations = 1; |
1e3c88bd PZ |
4543 | if (busiest->nr_running > 1) { |
4544 | /* | |
4545 | * Attempt to move tasks. If find_busiest_group has found | |
4546 | * an imbalance but busiest->nr_running <= 1, the group is | |
4547 | * still unbalanced. ld_moved simply stays zero, so it is | |
4548 | * correctly treated as an imbalance. | |
4549 | */ | |
8e45cb54 | 4550 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
4551 | env.src_cpu = busiest->cpu; |
4552 | env.src_rq = busiest; | |
4553 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 4554 | |
a35b6466 | 4555 | update_h_load(env.src_cpu); |
5d6523eb | 4556 | more_balance: |
1e3c88bd | 4557 | local_irq_save(flags); |
78feefc5 | 4558 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
4559 | |
4560 | /* | |
4561 | * cur_ld_moved - load moved in current iteration | |
4562 | * ld_moved - cumulative load moved across iterations | |
4563 | */ | |
4564 | cur_ld_moved = move_tasks(&env); | |
4565 | ld_moved += cur_ld_moved; | |
78feefc5 | 4566 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
4567 | local_irq_restore(flags); |
4568 | ||
5d6523eb PZ |
4569 | if (env.flags & LBF_NEED_BREAK) { |
4570 | env.flags &= ~LBF_NEED_BREAK; | |
4571 | goto more_balance; | |
4572 | } | |
4573 | ||
1e3c88bd PZ |
4574 | /* |
4575 | * some other cpu did the load balance for us. | |
4576 | */ | |
88b8dac0 SV |
4577 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
4578 | resched_cpu(env.dst_cpu); | |
4579 | ||
4580 | /* | |
4581 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
4582 | * us and move them to an alternate dst_cpu in our sched_group | |
4583 | * where they can run. The upper limit on how many times we | |
4584 | * iterate on same src_cpu is dependent on number of cpus in our | |
4585 | * sched_group. | |
4586 | * | |
4587 | * This changes load balance semantics a bit on who can move | |
4588 | * load to a given_cpu. In addition to the given_cpu itself | |
4589 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
4590 | * nohz-idle), we now have balance_cpu in a position to move | |
4591 | * load to given_cpu. In rare situations, this may cause | |
4592 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
4593 | * _independently_ and at _same_ time to move some load to | |
4594 | * given_cpu) causing exceess load to be moved to given_cpu. | |
4595 | * This however should not happen so much in practice and | |
4596 | * moreover subsequent load balance cycles should correct the | |
4597 | * excess load moved. | |
4598 | */ | |
4599 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 && | |
4600 | lb_iterations++ < max_lb_iterations) { | |
4601 | ||
78feefc5 | 4602 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 SV |
4603 | env.dst_cpu = env.new_dst_cpu; |
4604 | env.flags &= ~LBF_SOME_PINNED; | |
4605 | env.loop = 0; | |
4606 | env.loop_break = sched_nr_migrate_break; | |
4607 | /* | |
4608 | * Go back to "more_balance" rather than "redo" since we | |
4609 | * need to continue with same src_cpu. | |
4610 | */ | |
4611 | goto more_balance; | |
4612 | } | |
1e3c88bd PZ |
4613 | |
4614 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
8e45cb54 | 4615 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 4616 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
4617 | if (!cpumask_empty(cpus)) { |
4618 | env.loop = 0; | |
4619 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 4620 | goto redo; |
bbf18b19 | 4621 | } |
1e3c88bd PZ |
4622 | goto out_balanced; |
4623 | } | |
4624 | } | |
4625 | ||
4626 | if (!ld_moved) { | |
4627 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
4628 | /* |
4629 | * Increment the failure counter only on periodic balance. | |
4630 | * We do not want newidle balance, which can be very | |
4631 | * frequent, pollute the failure counter causing | |
4632 | * excessive cache_hot migrations and active balances. | |
4633 | */ | |
4634 | if (idle != CPU_NEWLY_IDLE) | |
4635 | sd->nr_balance_failed++; | |
1e3c88bd | 4636 | |
bd939f45 | 4637 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
4638 | raw_spin_lock_irqsave(&busiest->lock, flags); |
4639 | ||
969c7921 TH |
4640 | /* don't kick the active_load_balance_cpu_stop, |
4641 | * if the curr task on busiest cpu can't be | |
4642 | * moved to this_cpu | |
1e3c88bd PZ |
4643 | */ |
4644 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 4645 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
4646 | raw_spin_unlock_irqrestore(&busiest->lock, |
4647 | flags); | |
8e45cb54 | 4648 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
4649 | goto out_one_pinned; |
4650 | } | |
4651 | ||
969c7921 TH |
4652 | /* |
4653 | * ->active_balance synchronizes accesses to | |
4654 | * ->active_balance_work. Once set, it's cleared | |
4655 | * only after active load balance is finished. | |
4656 | */ | |
1e3c88bd PZ |
4657 | if (!busiest->active_balance) { |
4658 | busiest->active_balance = 1; | |
4659 | busiest->push_cpu = this_cpu; | |
4660 | active_balance = 1; | |
4661 | } | |
4662 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 4663 | |
bd939f45 | 4664 | if (active_balance) { |
969c7921 TH |
4665 | stop_one_cpu_nowait(cpu_of(busiest), |
4666 | active_load_balance_cpu_stop, busiest, | |
4667 | &busiest->active_balance_work); | |
bd939f45 | 4668 | } |
1e3c88bd PZ |
4669 | |
4670 | /* | |
4671 | * We've kicked active balancing, reset the failure | |
4672 | * counter. | |
4673 | */ | |
4674 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
4675 | } | |
4676 | } else | |
4677 | sd->nr_balance_failed = 0; | |
4678 | ||
4679 | if (likely(!active_balance)) { | |
4680 | /* We were unbalanced, so reset the balancing interval */ | |
4681 | sd->balance_interval = sd->min_interval; | |
4682 | } else { | |
4683 | /* | |
4684 | * If we've begun active balancing, start to back off. This | |
4685 | * case may not be covered by the all_pinned logic if there | |
4686 | * is only 1 task on the busy runqueue (because we don't call | |
4687 | * move_tasks). | |
4688 | */ | |
4689 | if (sd->balance_interval < sd->max_interval) | |
4690 | sd->balance_interval *= 2; | |
4691 | } | |
4692 | ||
1e3c88bd PZ |
4693 | goto out; |
4694 | ||
4695 | out_balanced: | |
4696 | schedstat_inc(sd, lb_balanced[idle]); | |
4697 | ||
4698 | sd->nr_balance_failed = 0; | |
4699 | ||
4700 | out_one_pinned: | |
4701 | /* tune up the balancing interval */ | |
8e45cb54 | 4702 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 4703 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
4704 | (sd->balance_interval < sd->max_interval)) |
4705 | sd->balance_interval *= 2; | |
4706 | ||
46e49b38 | 4707 | ld_moved = 0; |
1e3c88bd | 4708 | out: |
1e3c88bd PZ |
4709 | return ld_moved; |
4710 | } | |
4711 | ||
1e3c88bd PZ |
4712 | /* |
4713 | * idle_balance is called by schedule() if this_cpu is about to become | |
4714 | * idle. Attempts to pull tasks from other CPUs. | |
4715 | */ | |
029632fb | 4716 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
4717 | { |
4718 | struct sched_domain *sd; | |
4719 | int pulled_task = 0; | |
4720 | unsigned long next_balance = jiffies + HZ; | |
4721 | ||
4722 | this_rq->idle_stamp = this_rq->clock; | |
4723 | ||
4724 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
4725 | return; | |
4726 | ||
18bf2805 BS |
4727 | update_rq_runnable_avg(this_rq, 1); |
4728 | ||
f492e12e PZ |
4729 | /* |
4730 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
4731 | */ | |
4732 | raw_spin_unlock(&this_rq->lock); | |
4733 | ||
c66eaf61 | 4734 | update_shares(this_cpu); |
dce840a0 | 4735 | rcu_read_lock(); |
1e3c88bd PZ |
4736 | for_each_domain(this_cpu, sd) { |
4737 | unsigned long interval; | |
f492e12e | 4738 | int balance = 1; |
1e3c88bd PZ |
4739 | |
4740 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4741 | continue; | |
4742 | ||
f492e12e | 4743 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 4744 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
4745 | pulled_task = load_balance(this_cpu, this_rq, |
4746 | sd, CPU_NEWLY_IDLE, &balance); | |
4747 | } | |
1e3c88bd PZ |
4748 | |
4749 | interval = msecs_to_jiffies(sd->balance_interval); | |
4750 | if (time_after(next_balance, sd->last_balance + interval)) | |
4751 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
4752 | if (pulled_task) { |
4753 | this_rq->idle_stamp = 0; | |
1e3c88bd | 4754 | break; |
d5ad140b | 4755 | } |
1e3c88bd | 4756 | } |
dce840a0 | 4757 | rcu_read_unlock(); |
f492e12e PZ |
4758 | |
4759 | raw_spin_lock(&this_rq->lock); | |
4760 | ||
1e3c88bd PZ |
4761 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
4762 | /* | |
4763 | * We are going idle. next_balance may be set based on | |
4764 | * a busy processor. So reset next_balance. | |
4765 | */ | |
4766 | this_rq->next_balance = next_balance; | |
4767 | } | |
4768 | } | |
4769 | ||
4770 | /* | |
969c7921 TH |
4771 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
4772 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
4773 | * least 1 task to be running on each physical CPU where possible, and | |
4774 | * avoids physical / logical imbalances. | |
1e3c88bd | 4775 | */ |
969c7921 | 4776 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 4777 | { |
969c7921 TH |
4778 | struct rq *busiest_rq = data; |
4779 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 4780 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 4781 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 4782 | struct sched_domain *sd; |
969c7921 TH |
4783 | |
4784 | raw_spin_lock_irq(&busiest_rq->lock); | |
4785 | ||
4786 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
4787 | if (unlikely(busiest_cpu != smp_processor_id() || | |
4788 | !busiest_rq->active_balance)) | |
4789 | goto out_unlock; | |
1e3c88bd PZ |
4790 | |
4791 | /* Is there any task to move? */ | |
4792 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 4793 | goto out_unlock; |
1e3c88bd PZ |
4794 | |
4795 | /* | |
4796 | * This condition is "impossible", if it occurs | |
4797 | * we need to fix it. Originally reported by | |
4798 | * Bjorn Helgaas on a 128-cpu setup. | |
4799 | */ | |
4800 | BUG_ON(busiest_rq == target_rq); | |
4801 | ||
4802 | /* move a task from busiest_rq to target_rq */ | |
4803 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
4804 | |
4805 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 4806 | rcu_read_lock(); |
1e3c88bd PZ |
4807 | for_each_domain(target_cpu, sd) { |
4808 | if ((sd->flags & SD_LOAD_BALANCE) && | |
4809 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
4810 | break; | |
4811 | } | |
4812 | ||
4813 | if (likely(sd)) { | |
8e45cb54 PZ |
4814 | struct lb_env env = { |
4815 | .sd = sd, | |
ddcdf6e7 PZ |
4816 | .dst_cpu = target_cpu, |
4817 | .dst_rq = target_rq, | |
4818 | .src_cpu = busiest_rq->cpu, | |
4819 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
4820 | .idle = CPU_IDLE, |
4821 | }; | |
4822 | ||
1e3c88bd PZ |
4823 | schedstat_inc(sd, alb_count); |
4824 | ||
8e45cb54 | 4825 | if (move_one_task(&env)) |
1e3c88bd PZ |
4826 | schedstat_inc(sd, alb_pushed); |
4827 | else | |
4828 | schedstat_inc(sd, alb_failed); | |
4829 | } | |
dce840a0 | 4830 | rcu_read_unlock(); |
1e3c88bd | 4831 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
4832 | out_unlock: |
4833 | busiest_rq->active_balance = 0; | |
4834 | raw_spin_unlock_irq(&busiest_rq->lock); | |
4835 | return 0; | |
1e3c88bd PZ |
4836 | } |
4837 | ||
4838 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
4839 | /* |
4840 | * idle load balancing details | |
83cd4fe2 VP |
4841 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
4842 | * needed, they will kick the idle load balancer, which then does idle | |
4843 | * load balancing for all the idle CPUs. | |
4844 | */ | |
1e3c88bd | 4845 | static struct { |
83cd4fe2 | 4846 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 4847 | atomic_t nr_cpus; |
83cd4fe2 VP |
4848 | unsigned long next_balance; /* in jiffy units */ |
4849 | } nohz ____cacheline_aligned; | |
1e3c88bd | 4850 | |
8e7fbcbc | 4851 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 4852 | { |
0b005cf5 | 4853 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 4854 | |
786d6dc7 SS |
4855 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
4856 | return ilb; | |
4857 | ||
4858 | return nr_cpu_ids; | |
1e3c88bd | 4859 | } |
1e3c88bd | 4860 | |
83cd4fe2 VP |
4861 | /* |
4862 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
4863 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
4864 | * CPU (if there is one). | |
4865 | */ | |
4866 | static void nohz_balancer_kick(int cpu) | |
4867 | { | |
4868 | int ilb_cpu; | |
4869 | ||
4870 | nohz.next_balance++; | |
4871 | ||
0b005cf5 | 4872 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 4873 | |
0b005cf5 SS |
4874 | if (ilb_cpu >= nr_cpu_ids) |
4875 | return; | |
83cd4fe2 | 4876 | |
cd490c5b | 4877 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
4878 | return; |
4879 | /* | |
4880 | * Use smp_send_reschedule() instead of resched_cpu(). | |
4881 | * This way we generate a sched IPI on the target cpu which | |
4882 | * is idle. And the softirq performing nohz idle load balance | |
4883 | * will be run before returning from the IPI. | |
4884 | */ | |
4885 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
4886 | return; |
4887 | } | |
4888 | ||
c1cc017c | 4889 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
4890 | { |
4891 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
4892 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
4893 | atomic_dec(&nohz.nr_cpus); | |
4894 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
4895 | } | |
4896 | } | |
4897 | ||
69e1e811 SS |
4898 | static inline void set_cpu_sd_state_busy(void) |
4899 | { | |
4900 | struct sched_domain *sd; | |
4901 | int cpu = smp_processor_id(); | |
4902 | ||
4903 | if (!test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
4904 | return; | |
4905 | clear_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
4906 | ||
4907 | rcu_read_lock(); | |
4908 | for_each_domain(cpu, sd) | |
4909 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); | |
4910 | rcu_read_unlock(); | |
4911 | } | |
4912 | ||
4913 | void set_cpu_sd_state_idle(void) | |
4914 | { | |
4915 | struct sched_domain *sd; | |
4916 | int cpu = smp_processor_id(); | |
4917 | ||
4918 | if (test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
4919 | return; | |
4920 | set_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
4921 | ||
4922 | rcu_read_lock(); | |
4923 | for_each_domain(cpu, sd) | |
4924 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); | |
4925 | rcu_read_unlock(); | |
4926 | } | |
4927 | ||
1e3c88bd | 4928 | /* |
c1cc017c | 4929 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 4930 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 4931 | */ |
c1cc017c | 4932 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 4933 | { |
71325960 SS |
4934 | /* |
4935 | * If this cpu is going down, then nothing needs to be done. | |
4936 | */ | |
4937 | if (!cpu_active(cpu)) | |
4938 | return; | |
4939 | ||
c1cc017c AS |
4940 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
4941 | return; | |
1e3c88bd | 4942 | |
c1cc017c AS |
4943 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
4944 | atomic_inc(&nohz.nr_cpus); | |
4945 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 4946 | } |
71325960 SS |
4947 | |
4948 | static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb, | |
4949 | unsigned long action, void *hcpu) | |
4950 | { | |
4951 | switch (action & ~CPU_TASKS_FROZEN) { | |
4952 | case CPU_DYING: | |
c1cc017c | 4953 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
4954 | return NOTIFY_OK; |
4955 | default: | |
4956 | return NOTIFY_DONE; | |
4957 | } | |
4958 | } | |
1e3c88bd PZ |
4959 | #endif |
4960 | ||
4961 | static DEFINE_SPINLOCK(balancing); | |
4962 | ||
49c022e6 PZ |
4963 | /* |
4964 | * Scale the max load_balance interval with the number of CPUs in the system. | |
4965 | * This trades load-balance latency on larger machines for less cross talk. | |
4966 | */ | |
029632fb | 4967 | void update_max_interval(void) |
49c022e6 PZ |
4968 | { |
4969 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
4970 | } | |
4971 | ||
1e3c88bd PZ |
4972 | /* |
4973 | * It checks each scheduling domain to see if it is due to be balanced, | |
4974 | * and initiates a balancing operation if so. | |
4975 | * | |
4976 | * Balancing parameters are set up in arch_init_sched_domains. | |
4977 | */ | |
4978 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
4979 | { | |
4980 | int balance = 1; | |
4981 | struct rq *rq = cpu_rq(cpu); | |
4982 | unsigned long interval; | |
04f733b4 | 4983 | struct sched_domain *sd; |
1e3c88bd PZ |
4984 | /* Earliest time when we have to do rebalance again */ |
4985 | unsigned long next_balance = jiffies + 60*HZ; | |
4986 | int update_next_balance = 0; | |
4987 | int need_serialize; | |
4988 | ||
2069dd75 PZ |
4989 | update_shares(cpu); |
4990 | ||
dce840a0 | 4991 | rcu_read_lock(); |
1e3c88bd PZ |
4992 | for_each_domain(cpu, sd) { |
4993 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4994 | continue; | |
4995 | ||
4996 | interval = sd->balance_interval; | |
4997 | if (idle != CPU_IDLE) | |
4998 | interval *= sd->busy_factor; | |
4999 | ||
5000 | /* scale ms to jiffies */ | |
5001 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 5002 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
5003 | |
5004 | need_serialize = sd->flags & SD_SERIALIZE; | |
5005 | ||
5006 | if (need_serialize) { | |
5007 | if (!spin_trylock(&balancing)) | |
5008 | goto out; | |
5009 | } | |
5010 | ||
5011 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
5012 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
5013 | /* | |
5014 | * We've pulled tasks over so either we're no | |
c186fafe | 5015 | * longer idle. |
1e3c88bd PZ |
5016 | */ |
5017 | idle = CPU_NOT_IDLE; | |
5018 | } | |
5019 | sd->last_balance = jiffies; | |
5020 | } | |
5021 | if (need_serialize) | |
5022 | spin_unlock(&balancing); | |
5023 | out: | |
5024 | if (time_after(next_balance, sd->last_balance + interval)) { | |
5025 | next_balance = sd->last_balance + interval; | |
5026 | update_next_balance = 1; | |
5027 | } | |
5028 | ||
5029 | /* | |
5030 | * Stop the load balance at this level. There is another | |
5031 | * CPU in our sched group which is doing load balancing more | |
5032 | * actively. | |
5033 | */ | |
5034 | if (!balance) | |
5035 | break; | |
5036 | } | |
dce840a0 | 5037 | rcu_read_unlock(); |
1e3c88bd PZ |
5038 | |
5039 | /* | |
5040 | * next_balance will be updated only when there is a need. | |
5041 | * When the cpu is attached to null domain for ex, it will not be | |
5042 | * updated. | |
5043 | */ | |
5044 | if (likely(update_next_balance)) | |
5045 | rq->next_balance = next_balance; | |
5046 | } | |
5047 | ||
83cd4fe2 | 5048 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 5049 | /* |
83cd4fe2 | 5050 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
5051 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
5052 | */ | |
83cd4fe2 VP |
5053 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
5054 | { | |
5055 | struct rq *this_rq = cpu_rq(this_cpu); | |
5056 | struct rq *rq; | |
5057 | int balance_cpu; | |
5058 | ||
1c792db7 SS |
5059 | if (idle != CPU_IDLE || |
5060 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
5061 | goto end; | |
83cd4fe2 VP |
5062 | |
5063 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 5064 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
5065 | continue; |
5066 | ||
5067 | /* | |
5068 | * If this cpu gets work to do, stop the load balancing | |
5069 | * work being done for other cpus. Next load | |
5070 | * balancing owner will pick it up. | |
5071 | */ | |
1c792db7 | 5072 | if (need_resched()) |
83cd4fe2 | 5073 | break; |
83cd4fe2 | 5074 | |
5ed4f1d9 VG |
5075 | rq = cpu_rq(balance_cpu); |
5076 | ||
5077 | raw_spin_lock_irq(&rq->lock); | |
5078 | update_rq_clock(rq); | |
5079 | update_idle_cpu_load(rq); | |
5080 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
5081 | |
5082 | rebalance_domains(balance_cpu, CPU_IDLE); | |
5083 | ||
83cd4fe2 VP |
5084 | if (time_after(this_rq->next_balance, rq->next_balance)) |
5085 | this_rq->next_balance = rq->next_balance; | |
5086 | } | |
5087 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
5088 | end: |
5089 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
5090 | } |
5091 | ||
5092 | /* | |
0b005cf5 SS |
5093 | * Current heuristic for kicking the idle load balancer in the presence |
5094 | * of an idle cpu is the system. | |
5095 | * - This rq has more than one task. | |
5096 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
5097 | * busy cpu's exceeding the group's power. | |
5098 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
5099 | * domain span are idle. | |
83cd4fe2 VP |
5100 | */ |
5101 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
5102 | { | |
5103 | unsigned long now = jiffies; | |
0b005cf5 | 5104 | struct sched_domain *sd; |
83cd4fe2 | 5105 | |
1c792db7 | 5106 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
5107 | return 0; |
5108 | ||
1c792db7 SS |
5109 | /* |
5110 | * We may be recently in ticked or tickless idle mode. At the first | |
5111 | * busy tick after returning from idle, we will update the busy stats. | |
5112 | */ | |
69e1e811 | 5113 | set_cpu_sd_state_busy(); |
c1cc017c | 5114 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
5115 | |
5116 | /* | |
5117 | * None are in tickless mode and hence no need for NOHZ idle load | |
5118 | * balancing. | |
5119 | */ | |
5120 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
5121 | return 0; | |
1c792db7 SS |
5122 | |
5123 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
5124 | return 0; |
5125 | ||
0b005cf5 SS |
5126 | if (rq->nr_running >= 2) |
5127 | goto need_kick; | |
83cd4fe2 | 5128 | |
067491b7 | 5129 | rcu_read_lock(); |
0b005cf5 SS |
5130 | for_each_domain(cpu, sd) { |
5131 | struct sched_group *sg = sd->groups; | |
5132 | struct sched_group_power *sgp = sg->sgp; | |
5133 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 5134 | |
0b005cf5 | 5135 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 5136 | goto need_kick_unlock; |
0b005cf5 SS |
5137 | |
5138 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
5139 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
5140 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 5141 | goto need_kick_unlock; |
0b005cf5 SS |
5142 | |
5143 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
5144 | break; | |
83cd4fe2 | 5145 | } |
067491b7 | 5146 | rcu_read_unlock(); |
83cd4fe2 | 5147 | return 0; |
067491b7 PZ |
5148 | |
5149 | need_kick_unlock: | |
5150 | rcu_read_unlock(); | |
0b005cf5 SS |
5151 | need_kick: |
5152 | return 1; | |
83cd4fe2 VP |
5153 | } |
5154 | #else | |
5155 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
5156 | #endif | |
5157 | ||
5158 | /* | |
5159 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
5160 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
5161 | */ | |
1e3c88bd PZ |
5162 | static void run_rebalance_domains(struct softirq_action *h) |
5163 | { | |
5164 | int this_cpu = smp_processor_id(); | |
5165 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 5166 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
5167 | CPU_IDLE : CPU_NOT_IDLE; |
5168 | ||
5169 | rebalance_domains(this_cpu, idle); | |
5170 | ||
1e3c88bd | 5171 | /* |
83cd4fe2 | 5172 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
5173 | * balancing on behalf of the other idle cpus whose ticks are |
5174 | * stopped. | |
5175 | */ | |
83cd4fe2 | 5176 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
5177 | } |
5178 | ||
5179 | static inline int on_null_domain(int cpu) | |
5180 | { | |
90a6501f | 5181 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
5182 | } |
5183 | ||
5184 | /* | |
5185 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 5186 | */ |
029632fb | 5187 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 5188 | { |
1e3c88bd PZ |
5189 | /* Don't need to rebalance while attached to NULL domain */ |
5190 | if (time_after_eq(jiffies, rq->next_balance) && | |
5191 | likely(!on_null_domain(cpu))) | |
5192 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 | 5193 | #ifdef CONFIG_NO_HZ |
1c792db7 | 5194 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
5195 | nohz_balancer_kick(cpu); |
5196 | #endif | |
1e3c88bd PZ |
5197 | } |
5198 | ||
0bcdcf28 CE |
5199 | static void rq_online_fair(struct rq *rq) |
5200 | { | |
5201 | update_sysctl(); | |
5202 | } | |
5203 | ||
5204 | static void rq_offline_fair(struct rq *rq) | |
5205 | { | |
5206 | update_sysctl(); | |
a4c96ae3 PB |
5207 | |
5208 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
5209 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
5210 | } |
5211 | ||
55e12e5e | 5212 | #endif /* CONFIG_SMP */ |
e1d1484f | 5213 | |
bf0f6f24 IM |
5214 | /* |
5215 | * scheduler tick hitting a task of our scheduling class: | |
5216 | */ | |
8f4d37ec | 5217 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
5218 | { |
5219 | struct cfs_rq *cfs_rq; | |
5220 | struct sched_entity *se = &curr->se; | |
5221 | ||
5222 | for_each_sched_entity(se) { | |
5223 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 5224 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 5225 | } |
18bf2805 BS |
5226 | |
5227 | update_rq_runnable_avg(rq, 1); | |
bf0f6f24 IM |
5228 | } |
5229 | ||
5230 | /* | |
cd29fe6f PZ |
5231 | * called on fork with the child task as argument from the parent's context |
5232 | * - child not yet on the tasklist | |
5233 | * - preemption disabled | |
bf0f6f24 | 5234 | */ |
cd29fe6f | 5235 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 5236 | { |
4fc420c9 DN |
5237 | struct cfs_rq *cfs_rq; |
5238 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 5239 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
5240 | struct rq *rq = this_rq(); |
5241 | unsigned long flags; | |
5242 | ||
05fa785c | 5243 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 5244 | |
861d034e PZ |
5245 | update_rq_clock(rq); |
5246 | ||
4fc420c9 DN |
5247 | cfs_rq = task_cfs_rq(current); |
5248 | curr = cfs_rq->curr; | |
5249 | ||
b0a0f667 PM |
5250 | if (unlikely(task_cpu(p) != this_cpu)) { |
5251 | rcu_read_lock(); | |
cd29fe6f | 5252 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
5253 | rcu_read_unlock(); |
5254 | } | |
bf0f6f24 | 5255 | |
7109c442 | 5256 | update_curr(cfs_rq); |
cd29fe6f | 5257 | |
b5d9d734 MG |
5258 | if (curr) |
5259 | se->vruntime = curr->vruntime; | |
aeb73b04 | 5260 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 5261 | |
cd29fe6f | 5262 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 5263 | /* |
edcb60a3 IM |
5264 | * Upon rescheduling, sched_class::put_prev_task() will place |
5265 | * 'current' within the tree based on its new key value. | |
5266 | */ | |
4d78e7b6 | 5267 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 5268 | resched_task(rq->curr); |
4d78e7b6 | 5269 | } |
bf0f6f24 | 5270 | |
88ec22d3 PZ |
5271 | se->vruntime -= cfs_rq->min_vruntime; |
5272 | ||
05fa785c | 5273 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
5274 | } |
5275 | ||
cb469845 SR |
5276 | /* |
5277 | * Priority of the task has changed. Check to see if we preempt | |
5278 | * the current task. | |
5279 | */ | |
da7a735e PZ |
5280 | static void |
5281 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 5282 | { |
da7a735e PZ |
5283 | if (!p->se.on_rq) |
5284 | return; | |
5285 | ||
cb469845 SR |
5286 | /* |
5287 | * Reschedule if we are currently running on this runqueue and | |
5288 | * our priority decreased, or if we are not currently running on | |
5289 | * this runqueue and our priority is higher than the current's | |
5290 | */ | |
da7a735e | 5291 | if (rq->curr == p) { |
cb469845 SR |
5292 | if (p->prio > oldprio) |
5293 | resched_task(rq->curr); | |
5294 | } else | |
15afe09b | 5295 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5296 | } |
5297 | ||
da7a735e PZ |
5298 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
5299 | { | |
5300 | struct sched_entity *se = &p->se; | |
5301 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5302 | ||
5303 | /* | |
5304 | * Ensure the task's vruntime is normalized, so that when its | |
5305 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
5306 | * do the right thing. | |
5307 | * | |
5308 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
5309 | * have normalized the vruntime, if it was !on_rq, then only when | |
5310 | * the task is sleeping will it still have non-normalized vruntime. | |
5311 | */ | |
5312 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
5313 | /* | |
5314 | * Fix up our vruntime so that the current sleep doesn't | |
5315 | * cause 'unlimited' sleep bonus. | |
5316 | */ | |
5317 | place_entity(cfs_rq, se, 0); | |
5318 | se->vruntime -= cfs_rq->min_vruntime; | |
5319 | } | |
9ee474f5 PT |
5320 | |
5321 | #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP) | |
5322 | /* | |
5323 | * Remove our load from contribution when we leave sched_fair | |
5324 | * and ensure we don't carry in an old decay_count if we | |
5325 | * switch back. | |
5326 | */ | |
5327 | if (p->se.avg.decay_count) { | |
5328 | struct cfs_rq *cfs_rq = cfs_rq_of(&p->se); | |
5329 | __synchronize_entity_decay(&p->se); | |
5330 | subtract_blocked_load_contrib(cfs_rq, | |
5331 | p->se.avg.load_avg_contrib); | |
5332 | } | |
5333 | #endif | |
da7a735e PZ |
5334 | } |
5335 | ||
cb469845 SR |
5336 | /* |
5337 | * We switched to the sched_fair class. | |
5338 | */ | |
da7a735e | 5339 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 5340 | { |
da7a735e PZ |
5341 | if (!p->se.on_rq) |
5342 | return; | |
5343 | ||
cb469845 SR |
5344 | /* |
5345 | * We were most likely switched from sched_rt, so | |
5346 | * kick off the schedule if running, otherwise just see | |
5347 | * if we can still preempt the current task. | |
5348 | */ | |
da7a735e | 5349 | if (rq->curr == p) |
cb469845 SR |
5350 | resched_task(rq->curr); |
5351 | else | |
15afe09b | 5352 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5353 | } |
5354 | ||
83b699ed SV |
5355 | /* Account for a task changing its policy or group. |
5356 | * | |
5357 | * This routine is mostly called to set cfs_rq->curr field when a task | |
5358 | * migrates between groups/classes. | |
5359 | */ | |
5360 | static void set_curr_task_fair(struct rq *rq) | |
5361 | { | |
5362 | struct sched_entity *se = &rq->curr->se; | |
5363 | ||
ec12cb7f PT |
5364 | for_each_sched_entity(se) { |
5365 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5366 | ||
5367 | set_next_entity(cfs_rq, se); | |
5368 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
5369 | account_cfs_rq_runtime(cfs_rq, 0); | |
5370 | } | |
83b699ed SV |
5371 | } |
5372 | ||
029632fb PZ |
5373 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
5374 | { | |
5375 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
5376 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
5377 | #ifndef CONFIG_64BIT | |
5378 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
5379 | #endif | |
9ee474f5 PT |
5380 | #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP) |
5381 | atomic64_set(&cfs_rq->decay_counter, 1); | |
5382 | #endif | |
029632fb PZ |
5383 | } |
5384 | ||
810b3817 | 5385 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5386 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 5387 | { |
b2b5ce02 PZ |
5388 | /* |
5389 | * If the task was not on the rq at the time of this cgroup movement | |
5390 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
5391 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
5392 | * bonus in place_entity()). | |
5393 | * | |
5394 | * If it was on the rq, we've just 'preempted' it, which does convert | |
5395 | * ->vruntime to a relative base. | |
5396 | * | |
5397 | * Make sure both cases convert their relative position when migrating | |
5398 | * to another cgroup's rq. This does somewhat interfere with the | |
5399 | * fair sleeper stuff for the first placement, but who cares. | |
5400 | */ | |
7ceff013 DN |
5401 | /* |
5402 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
5403 | * But there are some cases where it has already been normalized: | |
5404 | * | |
5405 | * - Moving a forked child which is waiting for being woken up by | |
5406 | * wake_up_new_task(). | |
62af3783 DN |
5407 | * - Moving a task which has been woken up by try_to_wake_up() and |
5408 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
5409 | * |
5410 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
5411 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
5412 | */ | |
62af3783 | 5413 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
5414 | on_rq = 1; |
5415 | ||
b2b5ce02 PZ |
5416 | if (!on_rq) |
5417 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
5418 | set_task_rq(p, task_cpu(p)); | |
88ec22d3 | 5419 | if (!on_rq) |
b2b5ce02 | 5420 | p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; |
810b3817 | 5421 | } |
029632fb PZ |
5422 | |
5423 | void free_fair_sched_group(struct task_group *tg) | |
5424 | { | |
5425 | int i; | |
5426 | ||
5427 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5428 | ||
5429 | for_each_possible_cpu(i) { | |
5430 | if (tg->cfs_rq) | |
5431 | kfree(tg->cfs_rq[i]); | |
5432 | if (tg->se) | |
5433 | kfree(tg->se[i]); | |
5434 | } | |
5435 | ||
5436 | kfree(tg->cfs_rq); | |
5437 | kfree(tg->se); | |
5438 | } | |
5439 | ||
5440 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5441 | { | |
5442 | struct cfs_rq *cfs_rq; | |
5443 | struct sched_entity *se; | |
5444 | int i; | |
5445 | ||
5446 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
5447 | if (!tg->cfs_rq) | |
5448 | goto err; | |
5449 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
5450 | if (!tg->se) | |
5451 | goto err; | |
5452 | ||
5453 | tg->shares = NICE_0_LOAD; | |
5454 | ||
5455 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5456 | ||
5457 | for_each_possible_cpu(i) { | |
5458 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
5459 | GFP_KERNEL, cpu_to_node(i)); | |
5460 | if (!cfs_rq) | |
5461 | goto err; | |
5462 | ||
5463 | se = kzalloc_node(sizeof(struct sched_entity), | |
5464 | GFP_KERNEL, cpu_to_node(i)); | |
5465 | if (!se) | |
5466 | goto err_free_rq; | |
5467 | ||
5468 | init_cfs_rq(cfs_rq); | |
5469 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
5470 | } | |
5471 | ||
5472 | return 1; | |
5473 | ||
5474 | err_free_rq: | |
5475 | kfree(cfs_rq); | |
5476 | err: | |
5477 | return 0; | |
5478 | } | |
5479 | ||
5480 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
5481 | { | |
5482 | struct rq *rq = cpu_rq(cpu); | |
5483 | unsigned long flags; | |
5484 | ||
5485 | /* | |
5486 | * Only empty task groups can be destroyed; so we can speculatively | |
5487 | * check on_list without danger of it being re-added. | |
5488 | */ | |
5489 | if (!tg->cfs_rq[cpu]->on_list) | |
5490 | return; | |
5491 | ||
5492 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5493 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
5494 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5495 | } | |
5496 | ||
5497 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
5498 | struct sched_entity *se, int cpu, | |
5499 | struct sched_entity *parent) | |
5500 | { | |
5501 | struct rq *rq = cpu_rq(cpu); | |
5502 | ||
5503 | cfs_rq->tg = tg; | |
5504 | cfs_rq->rq = rq; | |
5505 | #ifdef CONFIG_SMP | |
5506 | /* allow initial update_cfs_load() to truncate */ | |
5507 | cfs_rq->load_stamp = 1; | |
810b3817 | 5508 | #endif |
029632fb PZ |
5509 | init_cfs_rq_runtime(cfs_rq); |
5510 | ||
5511 | tg->cfs_rq[cpu] = cfs_rq; | |
5512 | tg->se[cpu] = se; | |
5513 | ||
5514 | /* se could be NULL for root_task_group */ | |
5515 | if (!se) | |
5516 | return; | |
5517 | ||
5518 | if (!parent) | |
5519 | se->cfs_rq = &rq->cfs; | |
5520 | else | |
5521 | se->cfs_rq = parent->my_q; | |
5522 | ||
5523 | se->my_q = cfs_rq; | |
5524 | update_load_set(&se->load, 0); | |
5525 | se->parent = parent; | |
5526 | } | |
5527 | ||
5528 | static DEFINE_MUTEX(shares_mutex); | |
5529 | ||
5530 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
5531 | { | |
5532 | int i; | |
5533 | unsigned long flags; | |
5534 | ||
5535 | /* | |
5536 | * We can't change the weight of the root cgroup. | |
5537 | */ | |
5538 | if (!tg->se[0]) | |
5539 | return -EINVAL; | |
5540 | ||
5541 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
5542 | ||
5543 | mutex_lock(&shares_mutex); | |
5544 | if (tg->shares == shares) | |
5545 | goto done; | |
5546 | ||
5547 | tg->shares = shares; | |
5548 | for_each_possible_cpu(i) { | |
5549 | struct rq *rq = cpu_rq(i); | |
5550 | struct sched_entity *se; | |
5551 | ||
5552 | se = tg->se[i]; | |
5553 | /* Propagate contribution to hierarchy */ | |
5554 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5555 | for_each_sched_entity(se) | |
5556 | update_cfs_shares(group_cfs_rq(se)); | |
5557 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5558 | } | |
5559 | ||
5560 | done: | |
5561 | mutex_unlock(&shares_mutex); | |
5562 | return 0; | |
5563 | } | |
5564 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
5565 | ||
5566 | void free_fair_sched_group(struct task_group *tg) { } | |
5567 | ||
5568 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5569 | { | |
5570 | return 1; | |
5571 | } | |
5572 | ||
5573 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
5574 | ||
5575 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
5576 | ||
810b3817 | 5577 | |
6d686f45 | 5578 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
5579 | { |
5580 | struct sched_entity *se = &task->se; | |
0d721cea PW |
5581 | unsigned int rr_interval = 0; |
5582 | ||
5583 | /* | |
5584 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
5585 | * idle runqueue: | |
5586 | */ | |
0d721cea PW |
5587 | if (rq->cfs.load.weight) |
5588 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
5589 | |
5590 | return rr_interval; | |
5591 | } | |
5592 | ||
bf0f6f24 IM |
5593 | /* |
5594 | * All the scheduling class methods: | |
5595 | */ | |
029632fb | 5596 | const struct sched_class fair_sched_class = { |
5522d5d5 | 5597 | .next = &idle_sched_class, |
bf0f6f24 IM |
5598 | .enqueue_task = enqueue_task_fair, |
5599 | .dequeue_task = dequeue_task_fair, | |
5600 | .yield_task = yield_task_fair, | |
d95f4122 | 5601 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 5602 | |
2e09bf55 | 5603 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
5604 | |
5605 | .pick_next_task = pick_next_task_fair, | |
5606 | .put_prev_task = put_prev_task_fair, | |
5607 | ||
681f3e68 | 5608 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
5609 | .select_task_rq = select_task_rq_fair, |
5610 | ||
0bcdcf28 CE |
5611 | .rq_online = rq_online_fair, |
5612 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
5613 | |
5614 | .task_waking = task_waking_fair, | |
681f3e68 | 5615 | #endif |
bf0f6f24 | 5616 | |
83b699ed | 5617 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 5618 | .task_tick = task_tick_fair, |
cd29fe6f | 5619 | .task_fork = task_fork_fair, |
cb469845 SR |
5620 | |
5621 | .prio_changed = prio_changed_fair, | |
da7a735e | 5622 | .switched_from = switched_from_fair, |
cb469845 | 5623 | .switched_to = switched_to_fair, |
810b3817 | 5624 | |
0d721cea PW |
5625 | .get_rr_interval = get_rr_interval_fair, |
5626 | ||
810b3817 | 5627 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5628 | .task_move_group = task_move_group_fair, |
810b3817 | 5629 | #endif |
bf0f6f24 IM |
5630 | }; |
5631 | ||
5632 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 5633 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 5634 | { |
bf0f6f24 IM |
5635 | struct cfs_rq *cfs_rq; |
5636 | ||
5973e5b9 | 5637 | rcu_read_lock(); |
c3b64f1e | 5638 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 5639 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 5640 | rcu_read_unlock(); |
bf0f6f24 IM |
5641 | } |
5642 | #endif | |
029632fb PZ |
5643 | |
5644 | __init void init_sched_fair_class(void) | |
5645 | { | |
5646 | #ifdef CONFIG_SMP | |
5647 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
5648 | ||
5649 | #ifdef CONFIG_NO_HZ | |
554cecaf | 5650 | nohz.next_balance = jiffies; |
029632fb | 5651 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 5652 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
5653 | #endif |
5654 | #endif /* SMP */ | |
5655 | ||
5656 | } |