]>
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 | ||
aff3e498 PT |
262 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
263 | int force_update); | |
9ee474f5 | 264 | |
3d4b47b4 PZ |
265 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
266 | { | |
267 | if (!cfs_rq->on_list) { | |
67e86250 PT |
268 | /* |
269 | * Ensure we either appear before our parent (if already | |
270 | * enqueued) or force our parent to appear after us when it is | |
271 | * enqueued. The fact that we always enqueue bottom-up | |
272 | * reduces this to two cases. | |
273 | */ | |
274 | if (cfs_rq->tg->parent && | |
275 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
276 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
277 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
278 | } else { | |
279 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 280 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 281 | } |
3d4b47b4 PZ |
282 | |
283 | cfs_rq->on_list = 1; | |
9ee474f5 | 284 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 285 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
286 | } |
287 | } | |
288 | ||
289 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
290 | { | |
291 | if (cfs_rq->on_list) { | |
292 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
293 | cfs_rq->on_list = 0; | |
294 | } | |
295 | } | |
296 | ||
b758149c PZ |
297 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
298 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
299 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
300 | ||
301 | /* Do the two (enqueued) entities belong to the same group ? */ | |
302 | static inline int | |
303 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
304 | { | |
305 | if (se->cfs_rq == pse->cfs_rq) | |
306 | return 1; | |
307 | ||
308 | return 0; | |
309 | } | |
310 | ||
311 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
312 | { | |
313 | return se->parent; | |
314 | } | |
315 | ||
464b7527 PZ |
316 | /* return depth at which a sched entity is present in the hierarchy */ |
317 | static inline int depth_se(struct sched_entity *se) | |
318 | { | |
319 | int depth = 0; | |
320 | ||
321 | for_each_sched_entity(se) | |
322 | depth++; | |
323 | ||
324 | return depth; | |
325 | } | |
326 | ||
327 | static void | |
328 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
329 | { | |
330 | int se_depth, pse_depth; | |
331 | ||
332 | /* | |
333 | * preemption test can be made between sibling entities who are in the | |
334 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
335 | * both tasks until we find their ancestors who are siblings of common | |
336 | * parent. | |
337 | */ | |
338 | ||
339 | /* First walk up until both entities are at same depth */ | |
340 | se_depth = depth_se(*se); | |
341 | pse_depth = depth_se(*pse); | |
342 | ||
343 | while (se_depth > pse_depth) { | |
344 | se_depth--; | |
345 | *se = parent_entity(*se); | |
346 | } | |
347 | ||
348 | while (pse_depth > se_depth) { | |
349 | pse_depth--; | |
350 | *pse = parent_entity(*pse); | |
351 | } | |
352 | ||
353 | while (!is_same_group(*se, *pse)) { | |
354 | *se = parent_entity(*se); | |
355 | *pse = parent_entity(*pse); | |
356 | } | |
357 | } | |
358 | ||
8f48894f PZ |
359 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
360 | ||
361 | static inline struct task_struct *task_of(struct sched_entity *se) | |
362 | { | |
363 | return container_of(se, struct task_struct, se); | |
364 | } | |
bf0f6f24 | 365 | |
62160e3f IM |
366 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
367 | { | |
368 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
369 | } |
370 | ||
371 | #define entity_is_task(se) 1 | |
372 | ||
b758149c PZ |
373 | #define for_each_sched_entity(se) \ |
374 | for (; se; se = NULL) | |
bf0f6f24 | 375 | |
b758149c | 376 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 377 | { |
b758149c | 378 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
379 | } |
380 | ||
b758149c PZ |
381 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
382 | { | |
383 | struct task_struct *p = task_of(se); | |
384 | struct rq *rq = task_rq(p); | |
385 | ||
386 | return &rq->cfs; | |
387 | } | |
388 | ||
389 | /* runqueue "owned" by this group */ | |
390 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
391 | { | |
392 | return NULL; | |
393 | } | |
394 | ||
3d4b47b4 PZ |
395 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
396 | { | |
397 | } | |
398 | ||
399 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
400 | { | |
401 | } | |
402 | ||
b758149c PZ |
403 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
404 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
405 | ||
406 | static inline int | |
407 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
408 | { | |
409 | return 1; | |
410 | } | |
411 | ||
412 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
413 | { | |
414 | return NULL; | |
415 | } | |
416 | ||
464b7527 PZ |
417 | static inline void |
418 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
419 | { | |
420 | } | |
421 | ||
b758149c PZ |
422 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
423 | ||
6c16a6dc PZ |
424 | static __always_inline |
425 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
426 | |
427 | /************************************************************** | |
428 | * Scheduling class tree data structure manipulation methods: | |
429 | */ | |
430 | ||
0702e3eb | 431 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
02e0431a | 432 | { |
368059a9 PZ |
433 | s64 delta = (s64)(vruntime - min_vruntime); |
434 | if (delta > 0) | |
02e0431a PZ |
435 | min_vruntime = vruntime; |
436 | ||
437 | return min_vruntime; | |
438 | } | |
439 | ||
0702e3eb | 440 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
441 | { |
442 | s64 delta = (s64)(vruntime - min_vruntime); | |
443 | if (delta < 0) | |
444 | min_vruntime = vruntime; | |
445 | ||
446 | return min_vruntime; | |
447 | } | |
448 | ||
54fdc581 FC |
449 | static inline int entity_before(struct sched_entity *a, |
450 | struct sched_entity *b) | |
451 | { | |
452 | return (s64)(a->vruntime - b->vruntime) < 0; | |
453 | } | |
454 | ||
1af5f730 PZ |
455 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
456 | { | |
457 | u64 vruntime = cfs_rq->min_vruntime; | |
458 | ||
459 | if (cfs_rq->curr) | |
460 | vruntime = cfs_rq->curr->vruntime; | |
461 | ||
462 | if (cfs_rq->rb_leftmost) { | |
463 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
464 | struct sched_entity, | |
465 | run_node); | |
466 | ||
e17036da | 467 | if (!cfs_rq->curr) |
1af5f730 PZ |
468 | vruntime = se->vruntime; |
469 | else | |
470 | vruntime = min_vruntime(vruntime, se->vruntime); | |
471 | } | |
472 | ||
473 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | |
3fe1698b PZ |
474 | #ifndef CONFIG_64BIT |
475 | smp_wmb(); | |
476 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
477 | #endif | |
1af5f730 PZ |
478 | } |
479 | ||
bf0f6f24 IM |
480 | /* |
481 | * Enqueue an entity into the rb-tree: | |
482 | */ | |
0702e3eb | 483 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
484 | { |
485 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
486 | struct rb_node *parent = NULL; | |
487 | struct sched_entity *entry; | |
bf0f6f24 IM |
488 | int leftmost = 1; |
489 | ||
490 | /* | |
491 | * Find the right place in the rbtree: | |
492 | */ | |
493 | while (*link) { | |
494 | parent = *link; | |
495 | entry = rb_entry(parent, struct sched_entity, run_node); | |
496 | /* | |
497 | * We dont care about collisions. Nodes with | |
498 | * the same key stay together. | |
499 | */ | |
2bd2d6f2 | 500 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
501 | link = &parent->rb_left; |
502 | } else { | |
503 | link = &parent->rb_right; | |
504 | leftmost = 0; | |
505 | } | |
506 | } | |
507 | ||
508 | /* | |
509 | * Maintain a cache of leftmost tree entries (it is frequently | |
510 | * used): | |
511 | */ | |
1af5f730 | 512 | if (leftmost) |
57cb499d | 513 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
514 | |
515 | rb_link_node(&se->run_node, parent, link); | |
516 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
517 | } |
518 | ||
0702e3eb | 519 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 520 | { |
3fe69747 PZ |
521 | if (cfs_rq->rb_leftmost == &se->run_node) { |
522 | struct rb_node *next_node; | |
3fe69747 PZ |
523 | |
524 | next_node = rb_next(&se->run_node); | |
525 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 526 | } |
e9acbff6 | 527 | |
bf0f6f24 | 528 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
529 | } |
530 | ||
029632fb | 531 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 532 | { |
f4b6755f PZ |
533 | struct rb_node *left = cfs_rq->rb_leftmost; |
534 | ||
535 | if (!left) | |
536 | return NULL; | |
537 | ||
538 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
539 | } |
540 | ||
ac53db59 RR |
541 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
542 | { | |
543 | struct rb_node *next = rb_next(&se->run_node); | |
544 | ||
545 | if (!next) | |
546 | return NULL; | |
547 | ||
548 | return rb_entry(next, struct sched_entity, run_node); | |
549 | } | |
550 | ||
551 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 552 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 553 | { |
7eee3e67 | 554 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 555 | |
70eee74b BS |
556 | if (!last) |
557 | return NULL; | |
7eee3e67 IM |
558 | |
559 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
560 | } |
561 | ||
bf0f6f24 IM |
562 | /************************************************************** |
563 | * Scheduling class statistics methods: | |
564 | */ | |
565 | ||
acb4a848 | 566 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 567 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
568 | loff_t *ppos) |
569 | { | |
8d65af78 | 570 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 571 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
572 | |
573 | if (ret || !write) | |
574 | return ret; | |
575 | ||
576 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
577 | sysctl_sched_min_granularity); | |
578 | ||
acb4a848 CE |
579 | #define WRT_SYSCTL(name) \ |
580 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
581 | WRT_SYSCTL(sched_min_granularity); | |
582 | WRT_SYSCTL(sched_latency); | |
583 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
584 | #undef WRT_SYSCTL |
585 | ||
b2be5e96 PZ |
586 | return 0; |
587 | } | |
588 | #endif | |
647e7cac | 589 | |
a7be37ac | 590 | /* |
f9c0b095 | 591 | * delta /= w |
a7be37ac PZ |
592 | */ |
593 | static inline unsigned long | |
594 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
595 | { | |
f9c0b095 PZ |
596 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
597 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
598 | |
599 | return delta; | |
600 | } | |
601 | ||
647e7cac IM |
602 | /* |
603 | * The idea is to set a period in which each task runs once. | |
604 | * | |
532b1858 | 605 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
606 | * this period because otherwise the slices get too small. |
607 | * | |
608 | * p = (nr <= nl) ? l : l*nr/nl | |
609 | */ | |
4d78e7b6 PZ |
610 | static u64 __sched_period(unsigned long nr_running) |
611 | { | |
612 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 613 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
614 | |
615 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 616 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 617 | period *= nr_running; |
4d78e7b6 PZ |
618 | } |
619 | ||
620 | return period; | |
621 | } | |
622 | ||
647e7cac IM |
623 | /* |
624 | * We calculate the wall-time slice from the period by taking a part | |
625 | * proportional to the weight. | |
626 | * | |
f9c0b095 | 627 | * s = p*P[w/rw] |
647e7cac | 628 | */ |
6d0f0ebd | 629 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 630 | { |
0a582440 | 631 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 632 | |
0a582440 | 633 | for_each_sched_entity(se) { |
6272d68c | 634 | struct load_weight *load; |
3104bf03 | 635 | struct load_weight lw; |
6272d68c LM |
636 | |
637 | cfs_rq = cfs_rq_of(se); | |
638 | load = &cfs_rq->load; | |
f9c0b095 | 639 | |
0a582440 | 640 | if (unlikely(!se->on_rq)) { |
3104bf03 | 641 | lw = cfs_rq->load; |
0a582440 MG |
642 | |
643 | update_load_add(&lw, se->load.weight); | |
644 | load = &lw; | |
645 | } | |
646 | slice = calc_delta_mine(slice, se->load.weight, load); | |
647 | } | |
648 | return slice; | |
bf0f6f24 IM |
649 | } |
650 | ||
647e7cac | 651 | /* |
ac884dec | 652 | * We calculate the vruntime slice of a to be inserted task |
647e7cac | 653 | * |
f9c0b095 | 654 | * vs = s/w |
647e7cac | 655 | */ |
f9c0b095 | 656 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 657 | { |
f9c0b095 | 658 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
659 | } |
660 | ||
bf0f6f24 IM |
661 | /* |
662 | * Update the current task's runtime statistics. Skip current tasks that | |
663 | * are not in our scheduling class. | |
664 | */ | |
665 | static inline void | |
8ebc91d9 IM |
666 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
667 | unsigned long delta_exec) | |
bf0f6f24 | 668 | { |
bbdba7c0 | 669 | unsigned long delta_exec_weighted; |
bf0f6f24 | 670 | |
41acab88 LDM |
671 | schedstat_set(curr->statistics.exec_max, |
672 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
673 | |
674 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 675 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 676 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 677 | |
e9acbff6 | 678 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 679 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
680 | } |
681 | ||
b7cc0896 | 682 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 683 | { |
429d43bc | 684 | struct sched_entity *curr = cfs_rq->curr; |
305e6835 | 685 | u64 now = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
686 | unsigned long delta_exec; |
687 | ||
688 | if (unlikely(!curr)) | |
689 | return; | |
690 | ||
691 | /* | |
692 | * Get the amount of time the current task was running | |
693 | * since the last time we changed load (this cannot | |
694 | * overflow on 32 bits): | |
695 | */ | |
8ebc91d9 | 696 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
697 | if (!delta_exec) |
698 | return; | |
bf0f6f24 | 699 | |
8ebc91d9 IM |
700 | __update_curr(cfs_rq, curr, delta_exec); |
701 | curr->exec_start = now; | |
d842de87 SV |
702 | |
703 | if (entity_is_task(curr)) { | |
704 | struct task_struct *curtask = task_of(curr); | |
705 | ||
f977bb49 | 706 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 707 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 708 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 709 | } |
ec12cb7f PT |
710 | |
711 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
712 | } |
713 | ||
714 | static inline void | |
5870db5b | 715 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 716 | { |
41acab88 | 717 | schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); |
bf0f6f24 IM |
718 | } |
719 | ||
bf0f6f24 IM |
720 | /* |
721 | * Task is being enqueued - update stats: | |
722 | */ | |
d2417e5a | 723 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 724 | { |
bf0f6f24 IM |
725 | /* |
726 | * Are we enqueueing a waiting task? (for current tasks | |
727 | * a dequeue/enqueue event is a NOP) | |
728 | */ | |
429d43bc | 729 | if (se != cfs_rq->curr) |
5870db5b | 730 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
731 | } |
732 | ||
bf0f6f24 | 733 | static void |
9ef0a961 | 734 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 735 | { |
41acab88 LDM |
736 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
737 | rq_of(cfs_rq)->clock - se->statistics.wait_start)); | |
738 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | |
739 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
740 | rq_of(cfs_rq)->clock - se->statistics.wait_start); | |
768d0c27 PZ |
741 | #ifdef CONFIG_SCHEDSTATS |
742 | if (entity_is_task(se)) { | |
743 | trace_sched_stat_wait(task_of(se), | |
41acab88 | 744 | rq_of(cfs_rq)->clock - se->statistics.wait_start); |
768d0c27 PZ |
745 | } |
746 | #endif | |
41acab88 | 747 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
748 | } |
749 | ||
750 | static inline void | |
19b6a2e3 | 751 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 752 | { |
bf0f6f24 IM |
753 | /* |
754 | * Mark the end of the wait period if dequeueing a | |
755 | * waiting task: | |
756 | */ | |
429d43bc | 757 | if (se != cfs_rq->curr) |
9ef0a961 | 758 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
759 | } |
760 | ||
761 | /* | |
762 | * We are picking a new current task - update its stats: | |
763 | */ | |
764 | static inline void | |
79303e9e | 765 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
766 | { |
767 | /* | |
768 | * We are starting a new run period: | |
769 | */ | |
305e6835 | 770 | se->exec_start = rq_of(cfs_rq)->clock_task; |
bf0f6f24 IM |
771 | } |
772 | ||
bf0f6f24 IM |
773 | /************************************************** |
774 | * Scheduling class queueing methods: | |
775 | */ | |
776 | ||
30cfdcfc DA |
777 | static void |
778 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
779 | { | |
780 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 781 | if (!parent_entity(se)) |
029632fb | 782 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
783 | #ifdef CONFIG_SMP |
784 | if (entity_is_task(se)) | |
eb95308e | 785 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 786 | #endif |
30cfdcfc | 787 | cfs_rq->nr_running++; |
30cfdcfc DA |
788 | } |
789 | ||
790 | static void | |
791 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
792 | { | |
793 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 794 | if (!parent_entity(se)) |
029632fb | 795 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 796 | if (entity_is_task(se)) |
b87f1724 | 797 | list_del_init(&se->group_node); |
30cfdcfc | 798 | cfs_rq->nr_running--; |
30cfdcfc DA |
799 | } |
800 | ||
3ff6dcac YZ |
801 | #ifdef CONFIG_FAIR_GROUP_SCHED |
802 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
803 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
804 | { | |
805 | long tg_weight; | |
806 | ||
807 | /* | |
808 | * Use this CPU's actual weight instead of the last load_contribution | |
809 | * to gain a more accurate current total weight. See | |
810 | * update_cfs_rq_load_contribution(). | |
811 | */ | |
82958366 PT |
812 | tg_weight = atomic64_read(&tg->load_avg); |
813 | tg_weight -= cfs_rq->tg_load_contrib; | |
cf5f0acf PZ |
814 | tg_weight += cfs_rq->load.weight; |
815 | ||
816 | return tg_weight; | |
817 | } | |
818 | ||
6d5ab293 | 819 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 820 | { |
cf5f0acf | 821 | long tg_weight, load, shares; |
3ff6dcac | 822 | |
cf5f0acf | 823 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 824 | load = cfs_rq->load.weight; |
3ff6dcac | 825 | |
3ff6dcac | 826 | shares = (tg->shares * load); |
cf5f0acf PZ |
827 | if (tg_weight) |
828 | shares /= tg_weight; | |
3ff6dcac YZ |
829 | |
830 | if (shares < MIN_SHARES) | |
831 | shares = MIN_SHARES; | |
832 | if (shares > tg->shares) | |
833 | shares = tg->shares; | |
834 | ||
835 | return shares; | |
836 | } | |
3ff6dcac | 837 | # else /* CONFIG_SMP */ |
6d5ab293 | 838 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
839 | { |
840 | return tg->shares; | |
841 | } | |
3ff6dcac | 842 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
843 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
844 | unsigned long weight) | |
845 | { | |
19e5eebb PT |
846 | if (se->on_rq) { |
847 | /* commit outstanding execution time */ | |
848 | if (cfs_rq->curr == se) | |
849 | update_curr(cfs_rq); | |
2069dd75 | 850 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 851 | } |
2069dd75 PZ |
852 | |
853 | update_load_set(&se->load, weight); | |
854 | ||
855 | if (se->on_rq) | |
856 | account_entity_enqueue(cfs_rq, se); | |
857 | } | |
858 | ||
82958366 PT |
859 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
860 | ||
6d5ab293 | 861 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
862 | { |
863 | struct task_group *tg; | |
864 | struct sched_entity *se; | |
3ff6dcac | 865 | long shares; |
2069dd75 | 866 | |
2069dd75 PZ |
867 | tg = cfs_rq->tg; |
868 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 869 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 870 | return; |
3ff6dcac YZ |
871 | #ifndef CONFIG_SMP |
872 | if (likely(se->load.weight == tg->shares)) | |
873 | return; | |
874 | #endif | |
6d5ab293 | 875 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
876 | |
877 | reweight_entity(cfs_rq_of(se), se, shares); | |
878 | } | |
879 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 880 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
881 | { |
882 | } | |
883 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
884 | ||
9d85f21c | 885 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
886 | /* |
887 | * We choose a half-life close to 1 scheduling period. | |
888 | * Note: The tables below are dependent on this value. | |
889 | */ | |
890 | #define LOAD_AVG_PERIOD 32 | |
891 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
892 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
893 | ||
894 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
895 | static const u32 runnable_avg_yN_inv[] = { | |
896 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
897 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
898 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
899 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
900 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
901 | 0x85aac367, 0x82cd8698, | |
902 | }; | |
903 | ||
904 | /* | |
905 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
906 | * over-estimates when re-combining. | |
907 | */ | |
908 | static const u32 runnable_avg_yN_sum[] = { | |
909 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
910 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
911 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
912 | }; | |
913 | ||
9d85f21c PT |
914 | /* |
915 | * Approximate: | |
916 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
917 | */ | |
918 | static __always_inline u64 decay_load(u64 val, u64 n) | |
919 | { | |
5b51f2f8 PT |
920 | unsigned int local_n; |
921 | ||
922 | if (!n) | |
923 | return val; | |
924 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
925 | return 0; | |
926 | ||
927 | /* after bounds checking we can collapse to 32-bit */ | |
928 | local_n = n; | |
929 | ||
930 | /* | |
931 | * As y^PERIOD = 1/2, we can combine | |
932 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
933 | * With a look-up table which covers k^n (n<PERIOD) | |
934 | * | |
935 | * To achieve constant time decay_load. | |
936 | */ | |
937 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
938 | val >>= local_n / LOAD_AVG_PERIOD; | |
939 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
940 | } |
941 | ||
5b51f2f8 PT |
942 | val *= runnable_avg_yN_inv[local_n]; |
943 | /* We don't use SRR here since we always want to round down. */ | |
944 | return val >> 32; | |
945 | } | |
946 | ||
947 | /* | |
948 | * For updates fully spanning n periods, the contribution to runnable | |
949 | * average will be: \Sum 1024*y^n | |
950 | * | |
951 | * We can compute this reasonably efficiently by combining: | |
952 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
953 | */ | |
954 | static u32 __compute_runnable_contrib(u64 n) | |
955 | { | |
956 | u32 contrib = 0; | |
957 | ||
958 | if (likely(n <= LOAD_AVG_PERIOD)) | |
959 | return runnable_avg_yN_sum[n]; | |
960 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
961 | return LOAD_AVG_MAX; | |
962 | ||
963 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
964 | do { | |
965 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
966 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
967 | ||
968 | n -= LOAD_AVG_PERIOD; | |
969 | } while (n > LOAD_AVG_PERIOD); | |
970 | ||
971 | contrib = decay_load(contrib, n); | |
972 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
973 | } |
974 | ||
975 | /* | |
976 | * We can represent the historical contribution to runnable average as the | |
977 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
978 | * history into segments of approximately 1ms (1024us); label the segment that | |
979 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
980 | * | |
981 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
982 | * p0 p1 p2 | |
983 | * (now) (~1ms ago) (~2ms ago) | |
984 | * | |
985 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
986 | * | |
987 | * We then designate the fractions u_i as our co-efficients, yielding the | |
988 | * following representation of historical load: | |
989 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
990 | * | |
991 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
992 | * y^32 = 0.5 | |
993 | * | |
994 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
995 | * approximately half as much as the contribution to load within the last ms | |
996 | * (u_0). | |
997 | * | |
998 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
999 | * sum again by y is sufficient to update: | |
1000 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1001 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1002 | */ | |
1003 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1004 | struct sched_avg *sa, | |
1005 | int runnable) | |
1006 | { | |
5b51f2f8 PT |
1007 | u64 delta, periods; |
1008 | u32 runnable_contrib; | |
9d85f21c PT |
1009 | int delta_w, decayed = 0; |
1010 | ||
1011 | delta = now - sa->last_runnable_update; | |
1012 | /* | |
1013 | * This should only happen when time goes backwards, which it | |
1014 | * unfortunately does during sched clock init when we swap over to TSC. | |
1015 | */ | |
1016 | if ((s64)delta < 0) { | |
1017 | sa->last_runnable_update = now; | |
1018 | return 0; | |
1019 | } | |
1020 | ||
1021 | /* | |
1022 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1023 | * approximation of 1us and fast to compute. | |
1024 | */ | |
1025 | delta >>= 10; | |
1026 | if (!delta) | |
1027 | return 0; | |
1028 | sa->last_runnable_update = now; | |
1029 | ||
1030 | /* delta_w is the amount already accumulated against our next period */ | |
1031 | delta_w = sa->runnable_avg_period % 1024; | |
1032 | if (delta + delta_w >= 1024) { | |
1033 | /* period roll-over */ | |
1034 | decayed = 1; | |
1035 | ||
1036 | /* | |
1037 | * Now that we know we're crossing a period boundary, figure | |
1038 | * out how much from delta we need to complete the current | |
1039 | * period and accrue it. | |
1040 | */ | |
1041 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
1042 | if (runnable) |
1043 | sa->runnable_avg_sum += delta_w; | |
1044 | sa->runnable_avg_period += delta_w; | |
1045 | ||
1046 | delta -= delta_w; | |
1047 | ||
1048 | /* Figure out how many additional periods this update spans */ | |
1049 | periods = delta / 1024; | |
1050 | delta %= 1024; | |
1051 | ||
1052 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
1053 | periods + 1); | |
1054 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
1055 | periods + 1); | |
1056 | ||
1057 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
1058 | runnable_contrib = __compute_runnable_contrib(periods); | |
1059 | if (runnable) | |
1060 | sa->runnable_avg_sum += runnable_contrib; | |
1061 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
1062 | } |
1063 | ||
1064 | /* Remainder of delta accrued against u_0` */ | |
1065 | if (runnable) | |
1066 | sa->runnable_avg_sum += delta; | |
1067 | sa->runnable_avg_period += delta; | |
1068 | ||
1069 | return decayed; | |
1070 | } | |
1071 | ||
9ee474f5 | 1072 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 1073 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
1074 | { |
1075 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1076 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1077 | ||
1078 | decays -= se->avg.decay_count; | |
1079 | if (!decays) | |
aff3e498 | 1080 | return 0; |
9ee474f5 PT |
1081 | |
1082 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1083 | se->avg.decay_count = 0; | |
aff3e498 PT |
1084 | |
1085 | return decays; | |
9ee474f5 PT |
1086 | } |
1087 | ||
c566e8e9 PT |
1088 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1089 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1090 | int force_update) | |
1091 | { | |
1092 | struct task_group *tg = cfs_rq->tg; | |
1093 | s64 tg_contrib; | |
1094 | ||
1095 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
1096 | tg_contrib -= cfs_rq->tg_load_contrib; | |
1097 | ||
1098 | if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) { | |
1099 | atomic64_add(tg_contrib, &tg->load_avg); | |
1100 | cfs_rq->tg_load_contrib += tg_contrib; | |
1101 | } | |
1102 | } | |
8165e145 | 1103 | |
bb17f655 PT |
1104 | /* |
1105 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
1106 | * representation for computing load contributions. | |
1107 | */ | |
1108 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
1109 | struct cfs_rq *cfs_rq) | |
1110 | { | |
1111 | struct task_group *tg = cfs_rq->tg; | |
1112 | long contrib; | |
1113 | ||
1114 | /* The fraction of a cpu used by this cfs_rq */ | |
1115 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
1116 | sa->runnable_avg_period + 1); | |
1117 | contrib -= cfs_rq->tg_runnable_contrib; | |
1118 | ||
1119 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
1120 | atomic_add(contrib, &tg->runnable_avg); | |
1121 | cfs_rq->tg_runnable_contrib += contrib; | |
1122 | } | |
1123 | } | |
1124 | ||
8165e145 PT |
1125 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
1126 | { | |
1127 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
1128 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
1129 | int runnable_avg; |
1130 | ||
8165e145 PT |
1131 | u64 contrib; |
1132 | ||
1133 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
1134 | se->avg.load_avg_contrib = div64_u64(contrib, | |
1135 | atomic64_read(&tg->load_avg) + 1); | |
bb17f655 PT |
1136 | |
1137 | /* | |
1138 | * For group entities we need to compute a correction term in the case | |
1139 | * that they are consuming <1 cpu so that we would contribute the same | |
1140 | * load as a task of equal weight. | |
1141 | * | |
1142 | * Explicitly co-ordinating this measurement would be expensive, but | |
1143 | * fortunately the sum of each cpus contribution forms a usable | |
1144 | * lower-bound on the true value. | |
1145 | * | |
1146 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
1147 | * (and the sum represents true value) or they are disjoint and we are | |
1148 | * understating by the aggregate of their overlap. | |
1149 | * | |
1150 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
1151 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
1152 | * cpus that overlap for this interval and w_i is the interval width. | |
1153 | * | |
1154 | * On a small machine; the first term is well-bounded which bounds the | |
1155 | * total error since w_i is a subset of the period. Whereas on a | |
1156 | * larger machine, while this first term can be larger, if w_i is the | |
1157 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
1158 | * our upper bound of 1-cpu. | |
1159 | */ | |
1160 | runnable_avg = atomic_read(&tg->runnable_avg); | |
1161 | if (runnable_avg < NICE_0_LOAD) { | |
1162 | se->avg.load_avg_contrib *= runnable_avg; | |
1163 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
1164 | } | |
8165e145 | 1165 | } |
c566e8e9 PT |
1166 | #else |
1167 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1168 | int force_update) {} | |
bb17f655 PT |
1169 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
1170 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 1171 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
1172 | #endif |
1173 | ||
8165e145 PT |
1174 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
1175 | { | |
1176 | u32 contrib; | |
1177 | ||
1178 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
1179 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
1180 | contrib /= (se->avg.runnable_avg_period + 1); | |
1181 | se->avg.load_avg_contrib = scale_load(contrib); | |
1182 | } | |
1183 | ||
2dac754e PT |
1184 | /* Compute the current contribution to load_avg by se, return any delta */ |
1185 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
1186 | { | |
1187 | long old_contrib = se->avg.load_avg_contrib; | |
1188 | ||
8165e145 PT |
1189 | if (entity_is_task(se)) { |
1190 | __update_task_entity_contrib(se); | |
1191 | } else { | |
bb17f655 | 1192 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
1193 | __update_group_entity_contrib(se); |
1194 | } | |
2dac754e PT |
1195 | |
1196 | return se->avg.load_avg_contrib - old_contrib; | |
1197 | } | |
1198 | ||
9ee474f5 PT |
1199 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
1200 | long load_contrib) | |
1201 | { | |
1202 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
1203 | cfs_rq->blocked_load_avg -= load_contrib; | |
1204 | else | |
1205 | cfs_rq->blocked_load_avg = 0; | |
1206 | } | |
1207 | ||
f1b17280 PT |
1208 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
1209 | ||
9d85f21c | 1210 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
1211 | static inline void update_entity_load_avg(struct sched_entity *se, |
1212 | int update_cfs_rq) | |
9d85f21c | 1213 | { |
2dac754e PT |
1214 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
1215 | long contrib_delta; | |
f1b17280 | 1216 | u64 now; |
2dac754e | 1217 | |
f1b17280 PT |
1218 | /* |
1219 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
1220 | * case they are the parent of a throttled hierarchy. | |
1221 | */ | |
1222 | if (entity_is_task(se)) | |
1223 | now = cfs_rq_clock_task(cfs_rq); | |
1224 | else | |
1225 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
1226 | ||
1227 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
1228 | return; |
1229 | ||
1230 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
1231 | |
1232 | if (!update_cfs_rq) | |
1233 | return; | |
1234 | ||
2dac754e PT |
1235 | if (se->on_rq) |
1236 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
1237 | else |
1238 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
1239 | } | |
1240 | ||
1241 | /* | |
1242 | * Decay the load contributed by all blocked children and account this so that | |
1243 | * their contribution may appropriately discounted when they wake up. | |
1244 | */ | |
aff3e498 | 1245 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 1246 | { |
f1b17280 | 1247 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
1248 | u64 decays; |
1249 | ||
1250 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 1251 | if (!decays && !force_update) |
9ee474f5 PT |
1252 | return; |
1253 | ||
aff3e498 PT |
1254 | if (atomic64_read(&cfs_rq->removed_load)) { |
1255 | u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0); | |
1256 | subtract_blocked_load_contrib(cfs_rq, removed_load); | |
1257 | } | |
9ee474f5 | 1258 | |
aff3e498 PT |
1259 | if (decays) { |
1260 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
1261 | decays); | |
1262 | atomic64_add(decays, &cfs_rq->decay_counter); | |
1263 | cfs_rq->last_decay = now; | |
1264 | } | |
c566e8e9 PT |
1265 | |
1266 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
f269ae04 | 1267 | update_cfs_shares(cfs_rq); |
9d85f21c | 1268 | } |
18bf2805 BS |
1269 | |
1270 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
1271 | { | |
1272 | __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable); | |
bb17f655 | 1273 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 1274 | } |
2dac754e PT |
1275 | |
1276 | /* Add the load generated by se into cfs_rq's child load-average */ | |
1277 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
1278 | struct sched_entity *se, |
1279 | int wakeup) | |
2dac754e | 1280 | { |
aff3e498 PT |
1281 | /* |
1282 | * We track migrations using entity decay_count <= 0, on a wake-up | |
1283 | * migration we use a negative decay count to track the remote decays | |
1284 | * accumulated while sleeping. | |
1285 | */ | |
1286 | if (unlikely(se->avg.decay_count <= 0)) { | |
9ee474f5 | 1287 | se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task; |
aff3e498 PT |
1288 | if (se->avg.decay_count) { |
1289 | /* | |
1290 | * In a wake-up migration we have to approximate the | |
1291 | * time sleeping. This is because we can't synchronize | |
1292 | * clock_task between the two cpus, and it is not | |
1293 | * guaranteed to be read-safe. Instead, we can | |
1294 | * approximate this using our carried decays, which are | |
1295 | * explicitly atomically readable. | |
1296 | */ | |
1297 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
1298 | << 20; | |
1299 | update_entity_load_avg(se, 0); | |
1300 | /* Indicate that we're now synchronized and on-rq */ | |
1301 | se->avg.decay_count = 0; | |
1302 | } | |
9ee474f5 PT |
1303 | wakeup = 0; |
1304 | } else { | |
1305 | __synchronize_entity_decay(se); | |
1306 | } | |
1307 | ||
aff3e498 PT |
1308 | /* migrated tasks did not contribute to our blocked load */ |
1309 | if (wakeup) { | |
9ee474f5 | 1310 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
1311 | update_entity_load_avg(se, 0); |
1312 | } | |
9ee474f5 | 1313 | |
2dac754e | 1314 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
1315 | /* we force update consideration on load-balancer moves */ |
1316 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
1317 | } |
1318 | ||
9ee474f5 PT |
1319 | /* |
1320 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
1321 | * transitioning to a blocked state we track its projected decay using | |
1322 | * blocked_load_avg. | |
1323 | */ | |
2dac754e | 1324 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1325 | struct sched_entity *se, |
1326 | int sleep) | |
2dac754e | 1327 | { |
9ee474f5 | 1328 | update_entity_load_avg(se, 1); |
aff3e498 PT |
1329 | /* we force update consideration on load-balancer moves */ |
1330 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 1331 | |
2dac754e | 1332 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
1333 | if (sleep) { |
1334 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
1335 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
1336 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 1337 | } |
9d85f21c | 1338 | #else |
9ee474f5 PT |
1339 | static inline void update_entity_load_avg(struct sched_entity *se, |
1340 | int update_cfs_rq) {} | |
18bf2805 | 1341 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 1342 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1343 | struct sched_entity *se, |
1344 | int wakeup) {} | |
2dac754e | 1345 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1346 | struct sched_entity *se, |
1347 | int sleep) {} | |
aff3e498 PT |
1348 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
1349 | int force_update) {} | |
9d85f21c PT |
1350 | #endif |
1351 | ||
2396af69 | 1352 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1353 | { |
bf0f6f24 | 1354 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1355 | struct task_struct *tsk = NULL; |
1356 | ||
1357 | if (entity_is_task(se)) | |
1358 | tsk = task_of(se); | |
1359 | ||
41acab88 LDM |
1360 | if (se->statistics.sleep_start) { |
1361 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; | |
bf0f6f24 IM |
1362 | |
1363 | if ((s64)delta < 0) | |
1364 | delta = 0; | |
1365 | ||
41acab88 LDM |
1366 | if (unlikely(delta > se->statistics.sleep_max)) |
1367 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1368 | |
8c79a045 | 1369 | se->statistics.sleep_start = 0; |
41acab88 | 1370 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1371 | |
768d0c27 | 1372 | if (tsk) { |
e414314c | 1373 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1374 | trace_sched_stat_sleep(tsk, delta); |
1375 | } | |
bf0f6f24 | 1376 | } |
41acab88 LDM |
1377 | if (se->statistics.block_start) { |
1378 | u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; | |
bf0f6f24 IM |
1379 | |
1380 | if ((s64)delta < 0) | |
1381 | delta = 0; | |
1382 | ||
41acab88 LDM |
1383 | if (unlikely(delta > se->statistics.block_max)) |
1384 | se->statistics.block_max = delta; | |
bf0f6f24 | 1385 | |
8c79a045 | 1386 | se->statistics.block_start = 0; |
41acab88 | 1387 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1388 | |
e414314c | 1389 | if (tsk) { |
8f0dfc34 | 1390 | if (tsk->in_iowait) { |
41acab88 LDM |
1391 | se->statistics.iowait_sum += delta; |
1392 | se->statistics.iowait_count++; | |
768d0c27 | 1393 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1394 | } |
1395 | ||
b781a602 AV |
1396 | trace_sched_stat_blocked(tsk, delta); |
1397 | ||
e414314c PZ |
1398 | /* |
1399 | * Blocking time is in units of nanosecs, so shift by | |
1400 | * 20 to get a milliseconds-range estimation of the | |
1401 | * amount of time that the task spent sleeping: | |
1402 | */ | |
1403 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1404 | profile_hits(SLEEP_PROFILING, | |
1405 | (void *)get_wchan(tsk), | |
1406 | delta >> 20); | |
1407 | } | |
1408 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1409 | } |
bf0f6f24 IM |
1410 | } |
1411 | #endif | |
1412 | } | |
1413 | ||
ddc97297 PZ |
1414 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1415 | { | |
1416 | #ifdef CONFIG_SCHED_DEBUG | |
1417 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
1418 | ||
1419 | if (d < 0) | |
1420 | d = -d; | |
1421 | ||
1422 | if (d > 3*sysctl_sched_latency) | |
1423 | schedstat_inc(cfs_rq, nr_spread_over); | |
1424 | #endif | |
1425 | } | |
1426 | ||
aeb73b04 PZ |
1427 | static void |
1428 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
1429 | { | |
1af5f730 | 1430 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 1431 | |
2cb8600e PZ |
1432 | /* |
1433 | * The 'current' period is already promised to the current tasks, | |
1434 | * however the extra weight of the new task will slow them down a | |
1435 | * little, place the new task so that it fits in the slot that | |
1436 | * stays open at the end. | |
1437 | */ | |
94dfb5e7 | 1438 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 1439 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 1440 | |
a2e7a7eb | 1441 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 1442 | if (!initial) { |
a2e7a7eb | 1443 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 1444 | |
a2e7a7eb MG |
1445 | /* |
1446 | * Halve their sleep time's effect, to allow | |
1447 | * for a gentler effect of sleepers: | |
1448 | */ | |
1449 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
1450 | thresh >>= 1; | |
51e0304c | 1451 | |
a2e7a7eb | 1452 | vruntime -= thresh; |
aeb73b04 PZ |
1453 | } |
1454 | ||
b5d9d734 MG |
1455 | /* ensure we never gain time by being placed backwards. */ |
1456 | vruntime = max_vruntime(se->vruntime, vruntime); | |
1457 | ||
67e9fb2a | 1458 | se->vruntime = vruntime; |
aeb73b04 PZ |
1459 | } |
1460 | ||
d3d9dc33 PT |
1461 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
1462 | ||
bf0f6f24 | 1463 | static void |
88ec22d3 | 1464 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1465 | { |
88ec22d3 PZ |
1466 | /* |
1467 | * Update the normalized vruntime before updating min_vruntime | |
1468 | * through callig update_curr(). | |
1469 | */ | |
371fd7e7 | 1470 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
1471 | se->vruntime += cfs_rq->min_vruntime; |
1472 | ||
bf0f6f24 | 1473 | /* |
a2a2d680 | 1474 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1475 | */ |
b7cc0896 | 1476 | update_curr(cfs_rq); |
a992241d | 1477 | account_entity_enqueue(cfs_rq, se); |
f269ae04 | 1478 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
bf0f6f24 | 1479 | |
88ec22d3 | 1480 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 1481 | place_entity(cfs_rq, se, 0); |
2396af69 | 1482 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 1483 | } |
bf0f6f24 | 1484 | |
d2417e5a | 1485 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 1486 | check_spread(cfs_rq, se); |
83b699ed SV |
1487 | if (se != cfs_rq->curr) |
1488 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1489 | se->on_rq = 1; |
3d4b47b4 | 1490 | |
d3d9dc33 | 1491 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1492 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1493 | check_enqueue_throttle(cfs_rq); |
1494 | } | |
bf0f6f24 IM |
1495 | } |
1496 | ||
2c13c919 | 1497 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1498 | { |
2c13c919 RR |
1499 | for_each_sched_entity(se) { |
1500 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1501 | if (cfs_rq->last == se) | |
1502 | cfs_rq->last = NULL; | |
1503 | else | |
1504 | break; | |
1505 | } | |
1506 | } | |
2002c695 | 1507 | |
2c13c919 RR |
1508 | static void __clear_buddies_next(struct sched_entity *se) |
1509 | { | |
1510 | for_each_sched_entity(se) { | |
1511 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1512 | if (cfs_rq->next == se) | |
1513 | cfs_rq->next = NULL; | |
1514 | else | |
1515 | break; | |
1516 | } | |
2002c695 PZ |
1517 | } |
1518 | ||
ac53db59 RR |
1519 | static void __clear_buddies_skip(struct sched_entity *se) |
1520 | { | |
1521 | for_each_sched_entity(se) { | |
1522 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1523 | if (cfs_rq->skip == se) | |
1524 | cfs_rq->skip = NULL; | |
1525 | else | |
1526 | break; | |
1527 | } | |
1528 | } | |
1529 | ||
a571bbea PZ |
1530 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1531 | { | |
2c13c919 RR |
1532 | if (cfs_rq->last == se) |
1533 | __clear_buddies_last(se); | |
1534 | ||
1535 | if (cfs_rq->next == se) | |
1536 | __clear_buddies_next(se); | |
ac53db59 RR |
1537 | |
1538 | if (cfs_rq->skip == se) | |
1539 | __clear_buddies_skip(se); | |
a571bbea PZ |
1540 | } |
1541 | ||
6c16a6dc | 1542 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 1543 | |
bf0f6f24 | 1544 | static void |
371fd7e7 | 1545 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1546 | { |
a2a2d680 DA |
1547 | /* |
1548 | * Update run-time statistics of the 'current'. | |
1549 | */ | |
1550 | update_curr(cfs_rq); | |
1551 | ||
19b6a2e3 | 1552 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1553 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1554 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1555 | if (entity_is_task(se)) { |
1556 | struct task_struct *tsk = task_of(se); | |
1557 | ||
1558 | if (tsk->state & TASK_INTERRUPTIBLE) | |
41acab88 | 1559 | se->statistics.sleep_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1560 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
41acab88 | 1561 | se->statistics.block_start = rq_of(cfs_rq)->clock; |
bf0f6f24 | 1562 | } |
db36cc7d | 1563 | #endif |
67e9fb2a PZ |
1564 | } |
1565 | ||
2002c695 | 1566 | clear_buddies(cfs_rq, se); |
4793241b | 1567 | |
83b699ed | 1568 | if (se != cfs_rq->curr) |
30cfdcfc DA |
1569 | __dequeue_entity(cfs_rq, se); |
1570 | account_entity_dequeue(cfs_rq, se); | |
f269ae04 | 1571 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
88ec22d3 PZ |
1572 | |
1573 | /* | |
1574 | * Normalize the entity after updating the min_vruntime because the | |
1575 | * update can refer to the ->curr item and we need to reflect this | |
1576 | * movement in our normalized position. | |
1577 | */ | |
371fd7e7 | 1578 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1579 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 1580 | |
d8b4986d PT |
1581 | /* return excess runtime on last dequeue */ |
1582 | return_cfs_rq_runtime(cfs_rq); | |
1583 | ||
1e876231 | 1584 | update_min_vruntime(cfs_rq); |
f269ae04 | 1585 | se->on_rq = 0; |
bf0f6f24 IM |
1586 | } |
1587 | ||
1588 | /* | |
1589 | * Preempt the current task with a newly woken task if needed: | |
1590 | */ | |
7c92e54f | 1591 | static void |
2e09bf55 | 1592 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1593 | { |
11697830 | 1594 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
1595 | struct sched_entity *se; |
1596 | s64 delta; | |
11697830 | 1597 | |
6d0f0ebd | 1598 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1599 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1600 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1601 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1602 | /* |
1603 | * The current task ran long enough, ensure it doesn't get | |
1604 | * re-elected due to buddy favours. | |
1605 | */ | |
1606 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1607 | return; |
1608 | } | |
1609 | ||
1610 | /* | |
1611 | * Ensure that a task that missed wakeup preemption by a | |
1612 | * narrow margin doesn't have to wait for a full slice. | |
1613 | * This also mitigates buddy induced latencies under load. | |
1614 | */ | |
f685ceac MG |
1615 | if (delta_exec < sysctl_sched_min_granularity) |
1616 | return; | |
1617 | ||
f4cfb33e WX |
1618 | se = __pick_first_entity(cfs_rq); |
1619 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 1620 | |
f4cfb33e WX |
1621 | if (delta < 0) |
1622 | return; | |
d7d82944 | 1623 | |
f4cfb33e WX |
1624 | if (delta > ideal_runtime) |
1625 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
1626 | } |
1627 | ||
83b699ed | 1628 | static void |
8494f412 | 1629 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1630 | { |
83b699ed SV |
1631 | /* 'current' is not kept within the tree. */ |
1632 | if (se->on_rq) { | |
1633 | /* | |
1634 | * Any task has to be enqueued before it get to execute on | |
1635 | * a CPU. So account for the time it spent waiting on the | |
1636 | * runqueue. | |
1637 | */ | |
1638 | update_stats_wait_end(cfs_rq, se); | |
1639 | __dequeue_entity(cfs_rq, se); | |
1640 | } | |
1641 | ||
79303e9e | 1642 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1643 | cfs_rq->curr = se; |
eba1ed4b IM |
1644 | #ifdef CONFIG_SCHEDSTATS |
1645 | /* | |
1646 | * Track our maximum slice length, if the CPU's load is at | |
1647 | * least twice that of our own weight (i.e. dont track it | |
1648 | * when there are only lesser-weight tasks around): | |
1649 | */ | |
495eca49 | 1650 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1651 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1652 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1653 | } | |
1654 | #endif | |
4a55b450 | 1655 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1656 | } |
1657 | ||
3f3a4904 PZ |
1658 | static int |
1659 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1660 | ||
ac53db59 RR |
1661 | /* |
1662 | * Pick the next process, keeping these things in mind, in this order: | |
1663 | * 1) keep things fair between processes/task groups | |
1664 | * 2) pick the "next" process, since someone really wants that to run | |
1665 | * 3) pick the "last" process, for cache locality | |
1666 | * 4) do not run the "skip" process, if something else is available | |
1667 | */ | |
f4b6755f | 1668 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1669 | { |
ac53db59 | 1670 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1671 | struct sched_entity *left = se; |
f4b6755f | 1672 | |
ac53db59 RR |
1673 | /* |
1674 | * Avoid running the skip buddy, if running something else can | |
1675 | * be done without getting too unfair. | |
1676 | */ | |
1677 | if (cfs_rq->skip == se) { | |
1678 | struct sched_entity *second = __pick_next_entity(se); | |
1679 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1680 | se = second; | |
1681 | } | |
aa2ac252 | 1682 | |
f685ceac MG |
1683 | /* |
1684 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1685 | */ | |
1686 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1687 | se = cfs_rq->last; | |
1688 | ||
ac53db59 RR |
1689 | /* |
1690 | * Someone really wants this to run. If it's not unfair, run it. | |
1691 | */ | |
1692 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1693 | se = cfs_rq->next; | |
1694 | ||
f685ceac | 1695 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1696 | |
1697 | return se; | |
aa2ac252 PZ |
1698 | } |
1699 | ||
d3d9dc33 PT |
1700 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
1701 | ||
ab6cde26 | 1702 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1703 | { |
1704 | /* | |
1705 | * If still on the runqueue then deactivate_task() | |
1706 | * was not called and update_curr() has to be done: | |
1707 | */ | |
1708 | if (prev->on_rq) | |
b7cc0896 | 1709 | update_curr(cfs_rq); |
bf0f6f24 | 1710 | |
d3d9dc33 PT |
1711 | /* throttle cfs_rqs exceeding runtime */ |
1712 | check_cfs_rq_runtime(cfs_rq); | |
1713 | ||
ddc97297 | 1714 | check_spread(cfs_rq, prev); |
30cfdcfc | 1715 | if (prev->on_rq) { |
5870db5b | 1716 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1717 | /* Put 'current' back into the tree. */ |
1718 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 1719 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 1720 | update_entity_load_avg(prev, 1); |
30cfdcfc | 1721 | } |
429d43bc | 1722 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1723 | } |
1724 | ||
8f4d37ec PZ |
1725 | static void |
1726 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1727 | { |
bf0f6f24 | 1728 | /* |
30cfdcfc | 1729 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1730 | */ |
30cfdcfc | 1731 | update_curr(cfs_rq); |
bf0f6f24 | 1732 | |
9d85f21c PT |
1733 | /* |
1734 | * Ensure that runnable average is periodically updated. | |
1735 | */ | |
9ee474f5 | 1736 | update_entity_load_avg(curr, 1); |
aff3e498 | 1737 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9d85f21c | 1738 | |
8f4d37ec PZ |
1739 | #ifdef CONFIG_SCHED_HRTICK |
1740 | /* | |
1741 | * queued ticks are scheduled to match the slice, so don't bother | |
1742 | * validating it and just reschedule. | |
1743 | */ | |
983ed7a6 HH |
1744 | if (queued) { |
1745 | resched_task(rq_of(cfs_rq)->curr); | |
1746 | return; | |
1747 | } | |
8f4d37ec PZ |
1748 | /* |
1749 | * don't let the period tick interfere with the hrtick preemption | |
1750 | */ | |
1751 | if (!sched_feat(DOUBLE_TICK) && | |
1752 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
1753 | return; | |
1754 | #endif | |
1755 | ||
2c2efaed | 1756 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 1757 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
1758 | } |
1759 | ||
ab84d31e PT |
1760 | |
1761 | /************************************************** | |
1762 | * CFS bandwidth control machinery | |
1763 | */ | |
1764 | ||
1765 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
1766 | |
1767 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 1768 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
1769 | |
1770 | static inline bool cfs_bandwidth_used(void) | |
1771 | { | |
c5905afb | 1772 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
1773 | } |
1774 | ||
1775 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
1776 | { | |
1777 | /* only need to count groups transitioning between enabled/!enabled */ | |
1778 | if (enabled && !was_enabled) | |
c5905afb | 1779 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 1780 | else if (!enabled && was_enabled) |
c5905afb | 1781 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
1782 | } |
1783 | #else /* HAVE_JUMP_LABEL */ | |
1784 | static bool cfs_bandwidth_used(void) | |
1785 | { | |
1786 | return true; | |
1787 | } | |
1788 | ||
1789 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
1790 | #endif /* HAVE_JUMP_LABEL */ | |
1791 | ||
ab84d31e PT |
1792 | /* |
1793 | * default period for cfs group bandwidth. | |
1794 | * default: 0.1s, units: nanoseconds | |
1795 | */ | |
1796 | static inline u64 default_cfs_period(void) | |
1797 | { | |
1798 | return 100000000ULL; | |
1799 | } | |
ec12cb7f PT |
1800 | |
1801 | static inline u64 sched_cfs_bandwidth_slice(void) | |
1802 | { | |
1803 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
1804 | } | |
1805 | ||
a9cf55b2 PT |
1806 | /* |
1807 | * Replenish runtime according to assigned quota and update expiration time. | |
1808 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
1809 | * additional synchronization around rq->lock. | |
1810 | * | |
1811 | * requires cfs_b->lock | |
1812 | */ | |
029632fb | 1813 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
1814 | { |
1815 | u64 now; | |
1816 | ||
1817 | if (cfs_b->quota == RUNTIME_INF) | |
1818 | return; | |
1819 | ||
1820 | now = sched_clock_cpu(smp_processor_id()); | |
1821 | cfs_b->runtime = cfs_b->quota; | |
1822 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
1823 | } | |
1824 | ||
029632fb PZ |
1825 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
1826 | { | |
1827 | return &tg->cfs_bandwidth; | |
1828 | } | |
1829 | ||
f1b17280 PT |
1830 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
1831 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
1832 | { | |
1833 | if (unlikely(cfs_rq->throttle_count)) | |
1834 | return cfs_rq->throttled_clock_task; | |
1835 | ||
1836 | return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time; | |
1837 | } | |
1838 | ||
85dac906 PT |
1839 | /* returns 0 on failure to allocate runtime */ |
1840 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
1841 | { |
1842 | struct task_group *tg = cfs_rq->tg; | |
1843 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 1844 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
1845 | |
1846 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
1847 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
1848 | ||
1849 | raw_spin_lock(&cfs_b->lock); | |
1850 | if (cfs_b->quota == RUNTIME_INF) | |
1851 | amount = min_amount; | |
58088ad0 | 1852 | else { |
a9cf55b2 PT |
1853 | /* |
1854 | * If the bandwidth pool has become inactive, then at least one | |
1855 | * period must have elapsed since the last consumption. | |
1856 | * Refresh the global state and ensure bandwidth timer becomes | |
1857 | * active. | |
1858 | */ | |
1859 | if (!cfs_b->timer_active) { | |
1860 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 1861 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 1862 | } |
58088ad0 PT |
1863 | |
1864 | if (cfs_b->runtime > 0) { | |
1865 | amount = min(cfs_b->runtime, min_amount); | |
1866 | cfs_b->runtime -= amount; | |
1867 | cfs_b->idle = 0; | |
1868 | } | |
ec12cb7f | 1869 | } |
a9cf55b2 | 1870 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
1871 | raw_spin_unlock(&cfs_b->lock); |
1872 | ||
1873 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
1874 | /* |
1875 | * we may have advanced our local expiration to account for allowed | |
1876 | * spread between our sched_clock and the one on which runtime was | |
1877 | * issued. | |
1878 | */ | |
1879 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
1880 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
1881 | |
1882 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
1883 | } |
1884 | ||
a9cf55b2 PT |
1885 | /* |
1886 | * Note: This depends on the synchronization provided by sched_clock and the | |
1887 | * fact that rq->clock snapshots this value. | |
1888 | */ | |
1889 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 1890 | { |
a9cf55b2 PT |
1891 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
1892 | struct rq *rq = rq_of(cfs_rq); | |
1893 | ||
1894 | /* if the deadline is ahead of our clock, nothing to do */ | |
1895 | if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0)) | |
ec12cb7f PT |
1896 | return; |
1897 | ||
a9cf55b2 PT |
1898 | if (cfs_rq->runtime_remaining < 0) |
1899 | return; | |
1900 | ||
1901 | /* | |
1902 | * If the local deadline has passed we have to consider the | |
1903 | * possibility that our sched_clock is 'fast' and the global deadline | |
1904 | * has not truly expired. | |
1905 | * | |
1906 | * Fortunately we can check determine whether this the case by checking | |
1907 | * whether the global deadline has advanced. | |
1908 | */ | |
1909 | ||
1910 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
1911 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
1912 | cfs_rq->runtime_expires += TICK_NSEC; | |
1913 | } else { | |
1914 | /* global deadline is ahead, expiration has passed */ | |
1915 | cfs_rq->runtime_remaining = 0; | |
1916 | } | |
1917 | } | |
1918 | ||
1919 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
1920 | unsigned long delta_exec) | |
1921 | { | |
1922 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 1923 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
1924 | expire_cfs_rq_runtime(cfs_rq); |
1925 | ||
1926 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
1927 | return; |
1928 | ||
85dac906 PT |
1929 | /* |
1930 | * if we're unable to extend our runtime we resched so that the active | |
1931 | * hierarchy can be throttled | |
1932 | */ | |
1933 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
1934 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
1935 | } |
1936 | ||
6c16a6dc PZ |
1937 | static __always_inline |
1938 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 1939 | { |
56f570e5 | 1940 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
1941 | return; |
1942 | ||
1943 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
1944 | } | |
1945 | ||
85dac906 PT |
1946 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
1947 | { | |
56f570e5 | 1948 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
1949 | } |
1950 | ||
64660c86 PT |
1951 | /* check whether cfs_rq, or any parent, is throttled */ |
1952 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
1953 | { | |
56f570e5 | 1954 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
1955 | } |
1956 | ||
1957 | /* | |
1958 | * Ensure that neither of the group entities corresponding to src_cpu or | |
1959 | * dest_cpu are members of a throttled hierarchy when performing group | |
1960 | * load-balance operations. | |
1961 | */ | |
1962 | static inline int throttled_lb_pair(struct task_group *tg, | |
1963 | int src_cpu, int dest_cpu) | |
1964 | { | |
1965 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
1966 | ||
1967 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
1968 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
1969 | ||
1970 | return throttled_hierarchy(src_cfs_rq) || | |
1971 | throttled_hierarchy(dest_cfs_rq); | |
1972 | } | |
1973 | ||
1974 | /* updated child weight may affect parent so we have to do this bottom up */ | |
1975 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
1976 | { | |
1977 | struct rq *rq = data; | |
1978 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1979 | ||
1980 | cfs_rq->throttle_count--; | |
1981 | #ifdef CONFIG_SMP | |
1982 | if (!cfs_rq->throttle_count) { | |
f1b17280 PT |
1983 | /* adjust cfs_rq_clock_task() */ |
1984 | cfs_rq->throttled_clock_task_time += rq->clock_task - | |
1985 | cfs_rq->throttled_clock_task; | |
64660c86 PT |
1986 | } |
1987 | #endif | |
1988 | ||
1989 | return 0; | |
1990 | } | |
1991 | ||
1992 | static int tg_throttle_down(struct task_group *tg, void *data) | |
1993 | { | |
1994 | struct rq *rq = data; | |
1995 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
1996 | ||
82958366 PT |
1997 | /* group is entering throttled state, stop time */ |
1998 | if (!cfs_rq->throttle_count) | |
f1b17280 | 1999 | cfs_rq->throttled_clock_task = rq->clock_task; |
64660c86 PT |
2000 | cfs_rq->throttle_count++; |
2001 | ||
2002 | return 0; | |
2003 | } | |
2004 | ||
d3d9dc33 | 2005 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
2006 | { |
2007 | struct rq *rq = rq_of(cfs_rq); | |
2008 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2009 | struct sched_entity *se; | |
2010 | long task_delta, dequeue = 1; | |
2011 | ||
2012 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2013 | ||
f1b17280 | 2014 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
2015 | rcu_read_lock(); |
2016 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
2017 | rcu_read_unlock(); | |
85dac906 PT |
2018 | |
2019 | task_delta = cfs_rq->h_nr_running; | |
2020 | for_each_sched_entity(se) { | |
2021 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
2022 | /* throttled entity or throttle-on-deactivate */ | |
2023 | if (!se->on_rq) | |
2024 | break; | |
2025 | ||
2026 | if (dequeue) | |
2027 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
2028 | qcfs_rq->h_nr_running -= task_delta; | |
2029 | ||
2030 | if (qcfs_rq->load.weight) | |
2031 | dequeue = 0; | |
2032 | } | |
2033 | ||
2034 | if (!se) | |
2035 | rq->nr_running -= task_delta; | |
2036 | ||
2037 | cfs_rq->throttled = 1; | |
f1b17280 | 2038 | cfs_rq->throttled_clock = rq->clock; |
85dac906 PT |
2039 | raw_spin_lock(&cfs_b->lock); |
2040 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
2041 | raw_spin_unlock(&cfs_b->lock); | |
2042 | } | |
2043 | ||
029632fb | 2044 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
2045 | { |
2046 | struct rq *rq = rq_of(cfs_rq); | |
2047 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2048 | struct sched_entity *se; | |
2049 | int enqueue = 1; | |
2050 | long task_delta; | |
2051 | ||
2052 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2053 | ||
2054 | cfs_rq->throttled = 0; | |
2055 | raw_spin_lock(&cfs_b->lock); | |
f1b17280 | 2056 | cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock; |
671fd9da PT |
2057 | list_del_rcu(&cfs_rq->throttled_list); |
2058 | raw_spin_unlock(&cfs_b->lock); | |
2059 | ||
64660c86 PT |
2060 | update_rq_clock(rq); |
2061 | /* update hierarchical throttle state */ | |
2062 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
2063 | ||
671fd9da PT |
2064 | if (!cfs_rq->load.weight) |
2065 | return; | |
2066 | ||
2067 | task_delta = cfs_rq->h_nr_running; | |
2068 | for_each_sched_entity(se) { | |
2069 | if (se->on_rq) | |
2070 | enqueue = 0; | |
2071 | ||
2072 | cfs_rq = cfs_rq_of(se); | |
2073 | if (enqueue) | |
2074 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
2075 | cfs_rq->h_nr_running += task_delta; | |
2076 | ||
2077 | if (cfs_rq_throttled(cfs_rq)) | |
2078 | break; | |
2079 | } | |
2080 | ||
2081 | if (!se) | |
2082 | rq->nr_running += task_delta; | |
2083 | ||
2084 | /* determine whether we need to wake up potentially idle cpu */ | |
2085 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
2086 | resched_task(rq->curr); | |
2087 | } | |
2088 | ||
2089 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
2090 | u64 remaining, u64 expires) | |
2091 | { | |
2092 | struct cfs_rq *cfs_rq; | |
2093 | u64 runtime = remaining; | |
2094 | ||
2095 | rcu_read_lock(); | |
2096 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
2097 | throttled_list) { | |
2098 | struct rq *rq = rq_of(cfs_rq); | |
2099 | ||
2100 | raw_spin_lock(&rq->lock); | |
2101 | if (!cfs_rq_throttled(cfs_rq)) | |
2102 | goto next; | |
2103 | ||
2104 | runtime = -cfs_rq->runtime_remaining + 1; | |
2105 | if (runtime > remaining) | |
2106 | runtime = remaining; | |
2107 | remaining -= runtime; | |
2108 | ||
2109 | cfs_rq->runtime_remaining += runtime; | |
2110 | cfs_rq->runtime_expires = expires; | |
2111 | ||
2112 | /* we check whether we're throttled above */ | |
2113 | if (cfs_rq->runtime_remaining > 0) | |
2114 | unthrottle_cfs_rq(cfs_rq); | |
2115 | ||
2116 | next: | |
2117 | raw_spin_unlock(&rq->lock); | |
2118 | ||
2119 | if (!remaining) | |
2120 | break; | |
2121 | } | |
2122 | rcu_read_unlock(); | |
2123 | ||
2124 | return remaining; | |
2125 | } | |
2126 | ||
58088ad0 PT |
2127 | /* |
2128 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2129 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2130 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2131 | * used to track this state. | |
2132 | */ | |
2133 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
2134 | { | |
671fd9da PT |
2135 | u64 runtime, runtime_expires; |
2136 | int idle = 1, throttled; | |
58088ad0 PT |
2137 | |
2138 | raw_spin_lock(&cfs_b->lock); | |
2139 | /* no need to continue the timer with no bandwidth constraint */ | |
2140 | if (cfs_b->quota == RUNTIME_INF) | |
2141 | goto out_unlock; | |
2142 | ||
671fd9da PT |
2143 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
2144 | /* idle depends on !throttled (for the case of a large deficit) */ | |
2145 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 2146 | cfs_b->nr_periods += overrun; |
671fd9da | 2147 | |
a9cf55b2 PT |
2148 | /* if we're going inactive then everything else can be deferred */ |
2149 | if (idle) | |
2150 | goto out_unlock; | |
2151 | ||
2152 | __refill_cfs_bandwidth_runtime(cfs_b); | |
2153 | ||
671fd9da PT |
2154 | if (!throttled) { |
2155 | /* mark as potentially idle for the upcoming period */ | |
2156 | cfs_b->idle = 1; | |
2157 | goto out_unlock; | |
2158 | } | |
2159 | ||
e8da1b18 NR |
2160 | /* account preceding periods in which throttling occurred */ |
2161 | cfs_b->nr_throttled += overrun; | |
2162 | ||
671fd9da PT |
2163 | /* |
2164 | * There are throttled entities so we must first use the new bandwidth | |
2165 | * to unthrottle them before making it generally available. This | |
2166 | * ensures that all existing debts will be paid before a new cfs_rq is | |
2167 | * allowed to run. | |
2168 | */ | |
2169 | runtime = cfs_b->runtime; | |
2170 | runtime_expires = cfs_b->runtime_expires; | |
2171 | cfs_b->runtime = 0; | |
2172 | ||
2173 | /* | |
2174 | * This check is repeated as we are holding onto the new bandwidth | |
2175 | * while we unthrottle. This can potentially race with an unthrottled | |
2176 | * group trying to acquire new bandwidth from the global pool. | |
2177 | */ | |
2178 | while (throttled && runtime > 0) { | |
2179 | raw_spin_unlock(&cfs_b->lock); | |
2180 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
2181 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
2182 | runtime_expires); | |
2183 | raw_spin_lock(&cfs_b->lock); | |
2184 | ||
2185 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
2186 | } | |
58088ad0 | 2187 | |
671fd9da PT |
2188 | /* return (any) remaining runtime */ |
2189 | cfs_b->runtime = runtime; | |
2190 | /* | |
2191 | * While we are ensured activity in the period following an | |
2192 | * unthrottle, this also covers the case in which the new bandwidth is | |
2193 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
2194 | * timer to remain active while there are any throttled entities.) | |
2195 | */ | |
2196 | cfs_b->idle = 0; | |
58088ad0 PT |
2197 | out_unlock: |
2198 | if (idle) | |
2199 | cfs_b->timer_active = 0; | |
2200 | raw_spin_unlock(&cfs_b->lock); | |
2201 | ||
2202 | return idle; | |
2203 | } | |
d3d9dc33 | 2204 | |
d8b4986d PT |
2205 | /* a cfs_rq won't donate quota below this amount */ |
2206 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
2207 | /* minimum remaining period time to redistribute slack quota */ | |
2208 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
2209 | /* how long we wait to gather additional slack before distributing */ | |
2210 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
2211 | ||
2212 | /* are we near the end of the current quota period? */ | |
2213 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
2214 | { | |
2215 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
2216 | u64 remaining; | |
2217 | ||
2218 | /* if the call-back is running a quota refresh is already occurring */ | |
2219 | if (hrtimer_callback_running(refresh_timer)) | |
2220 | return 1; | |
2221 | ||
2222 | /* is a quota refresh about to occur? */ | |
2223 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
2224 | if (remaining < min_expire) | |
2225 | return 1; | |
2226 | ||
2227 | return 0; | |
2228 | } | |
2229 | ||
2230 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
2231 | { | |
2232 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
2233 | ||
2234 | /* if there's a quota refresh soon don't bother with slack */ | |
2235 | if (runtime_refresh_within(cfs_b, min_left)) | |
2236 | return; | |
2237 | ||
2238 | start_bandwidth_timer(&cfs_b->slack_timer, | |
2239 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
2240 | } | |
2241 | ||
2242 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
2243 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2244 | { | |
2245 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2246 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
2247 | ||
2248 | if (slack_runtime <= 0) | |
2249 | return; | |
2250 | ||
2251 | raw_spin_lock(&cfs_b->lock); | |
2252 | if (cfs_b->quota != RUNTIME_INF && | |
2253 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
2254 | cfs_b->runtime += slack_runtime; | |
2255 | ||
2256 | /* we are under rq->lock, defer unthrottling using a timer */ | |
2257 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
2258 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
2259 | start_cfs_slack_bandwidth(cfs_b); | |
2260 | } | |
2261 | raw_spin_unlock(&cfs_b->lock); | |
2262 | ||
2263 | /* even if it's not valid for return we don't want to try again */ | |
2264 | cfs_rq->runtime_remaining -= slack_runtime; | |
2265 | } | |
2266 | ||
2267 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2268 | { | |
56f570e5 PT |
2269 | if (!cfs_bandwidth_used()) |
2270 | return; | |
2271 | ||
fccfdc6f | 2272 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2273 | return; |
2274 | ||
2275 | __return_cfs_rq_runtime(cfs_rq); | |
2276 | } | |
2277 | ||
2278 | /* | |
2279 | * This is done with a timer (instead of inline with bandwidth return) since | |
2280 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2281 | */ | |
2282 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2283 | { | |
2284 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2285 | u64 expires; | |
2286 | ||
2287 | /* confirm we're still not at a refresh boundary */ | |
2288 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2289 | return; | |
2290 | ||
2291 | raw_spin_lock(&cfs_b->lock); | |
2292 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2293 | runtime = cfs_b->runtime; | |
2294 | cfs_b->runtime = 0; | |
2295 | } | |
2296 | expires = cfs_b->runtime_expires; | |
2297 | raw_spin_unlock(&cfs_b->lock); | |
2298 | ||
2299 | if (!runtime) | |
2300 | return; | |
2301 | ||
2302 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2303 | ||
2304 | raw_spin_lock(&cfs_b->lock); | |
2305 | if (expires == cfs_b->runtime_expires) | |
2306 | cfs_b->runtime = runtime; | |
2307 | raw_spin_unlock(&cfs_b->lock); | |
2308 | } | |
2309 | ||
d3d9dc33 PT |
2310 | /* |
2311 | * When a group wakes up we want to make sure that its quota is not already | |
2312 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2313 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2314 | */ | |
2315 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2316 | { | |
56f570e5 PT |
2317 | if (!cfs_bandwidth_used()) |
2318 | return; | |
2319 | ||
d3d9dc33 PT |
2320 | /* an active group must be handled by the update_curr()->put() path */ |
2321 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2322 | return; | |
2323 | ||
2324 | /* ensure the group is not already throttled */ | |
2325 | if (cfs_rq_throttled(cfs_rq)) | |
2326 | return; | |
2327 | ||
2328 | /* update runtime allocation */ | |
2329 | account_cfs_rq_runtime(cfs_rq, 0); | |
2330 | if (cfs_rq->runtime_remaining <= 0) | |
2331 | throttle_cfs_rq(cfs_rq); | |
2332 | } | |
2333 | ||
2334 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2335 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2336 | { | |
56f570e5 PT |
2337 | if (!cfs_bandwidth_used()) |
2338 | return; | |
2339 | ||
d3d9dc33 PT |
2340 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2341 | return; | |
2342 | ||
2343 | /* | |
2344 | * it's possible for a throttled entity to be forced into a running | |
2345 | * state (e.g. set_curr_task), in this case we're finished. | |
2346 | */ | |
2347 | if (cfs_rq_throttled(cfs_rq)) | |
2348 | return; | |
2349 | ||
2350 | throttle_cfs_rq(cfs_rq); | |
2351 | } | |
029632fb PZ |
2352 | |
2353 | static inline u64 default_cfs_period(void); | |
2354 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun); | |
2355 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b); | |
2356 | ||
2357 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) | |
2358 | { | |
2359 | struct cfs_bandwidth *cfs_b = | |
2360 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2361 | do_sched_cfs_slack_timer(cfs_b); | |
2362 | ||
2363 | return HRTIMER_NORESTART; | |
2364 | } | |
2365 | ||
2366 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2367 | { | |
2368 | struct cfs_bandwidth *cfs_b = | |
2369 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2370 | ktime_t now; | |
2371 | int overrun; | |
2372 | int idle = 0; | |
2373 | ||
2374 | for (;;) { | |
2375 | now = hrtimer_cb_get_time(timer); | |
2376 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2377 | ||
2378 | if (!overrun) | |
2379 | break; | |
2380 | ||
2381 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2382 | } | |
2383 | ||
2384 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2385 | } | |
2386 | ||
2387 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2388 | { | |
2389 | raw_spin_lock_init(&cfs_b->lock); | |
2390 | cfs_b->runtime = 0; | |
2391 | cfs_b->quota = RUNTIME_INF; | |
2392 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2393 | ||
2394 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2395 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2396 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2397 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2398 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2399 | } | |
2400 | ||
2401 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2402 | { | |
2403 | cfs_rq->runtime_enabled = 0; | |
2404 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2405 | } | |
2406 | ||
2407 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2408 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2409 | { | |
2410 | /* | |
2411 | * The timer may be active because we're trying to set a new bandwidth | |
2412 | * period or because we're racing with the tear-down path | |
2413 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2414 | * terminates). In either case we ensure that it's re-programmed | |
2415 | */ | |
2416 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2417 | raw_spin_unlock(&cfs_b->lock); | |
2418 | /* ensure cfs_b->lock is available while we wait */ | |
2419 | hrtimer_cancel(&cfs_b->period_timer); | |
2420 | ||
2421 | raw_spin_lock(&cfs_b->lock); | |
2422 | /* if someone else restarted the timer then we're done */ | |
2423 | if (cfs_b->timer_active) | |
2424 | return; | |
2425 | } | |
2426 | ||
2427 | cfs_b->timer_active = 1; | |
2428 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
2429 | } | |
2430 | ||
2431 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2432 | { | |
2433 | hrtimer_cancel(&cfs_b->period_timer); | |
2434 | hrtimer_cancel(&cfs_b->slack_timer); | |
2435 | } | |
2436 | ||
a4c96ae3 | 2437 | static void unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
2438 | { |
2439 | struct cfs_rq *cfs_rq; | |
2440 | ||
2441 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2442 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2443 | ||
2444 | if (!cfs_rq->runtime_enabled) | |
2445 | continue; | |
2446 | ||
2447 | /* | |
2448 | * clock_task is not advancing so we just need to make sure | |
2449 | * there's some valid quota amount | |
2450 | */ | |
2451 | cfs_rq->runtime_remaining = cfs_b->quota; | |
2452 | if (cfs_rq_throttled(cfs_rq)) | |
2453 | unthrottle_cfs_rq(cfs_rq); | |
2454 | } | |
2455 | } | |
2456 | ||
2457 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
2458 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
2459 | { | |
2460 | return rq_of(cfs_rq)->clock_task; | |
2461 | } | |
2462 | ||
2463 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2464 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
2465 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
2466 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 2467 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
2468 | |
2469 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
2470 | { | |
2471 | return 0; | |
2472 | } | |
64660c86 PT |
2473 | |
2474 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2475 | { | |
2476 | return 0; | |
2477 | } | |
2478 | ||
2479 | static inline int throttled_lb_pair(struct task_group *tg, | |
2480 | int src_cpu, int dest_cpu) | |
2481 | { | |
2482 | return 0; | |
2483 | } | |
029632fb PZ |
2484 | |
2485 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
2486 | ||
2487 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2488 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
2489 | #endif |
2490 | ||
029632fb PZ |
2491 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2492 | { | |
2493 | return NULL; | |
2494 | } | |
2495 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 2496 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
2497 | |
2498 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
2499 | ||
bf0f6f24 IM |
2500 | /************************************************** |
2501 | * CFS operations on tasks: | |
2502 | */ | |
2503 | ||
8f4d37ec PZ |
2504 | #ifdef CONFIG_SCHED_HRTICK |
2505 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2506 | { | |
8f4d37ec PZ |
2507 | struct sched_entity *se = &p->se; |
2508 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2509 | ||
2510 | WARN_ON(task_rq(p) != rq); | |
2511 | ||
b39e66ea | 2512 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
2513 | u64 slice = sched_slice(cfs_rq, se); |
2514 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
2515 | s64 delta = slice - ran; | |
2516 | ||
2517 | if (delta < 0) { | |
2518 | if (rq->curr == p) | |
2519 | resched_task(p); | |
2520 | return; | |
2521 | } | |
2522 | ||
2523 | /* | |
2524 | * Don't schedule slices shorter than 10000ns, that just | |
2525 | * doesn't make sense. Rely on vruntime for fairness. | |
2526 | */ | |
31656519 | 2527 | if (rq->curr != p) |
157124c1 | 2528 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 2529 | |
31656519 | 2530 | hrtick_start(rq, delta); |
8f4d37ec PZ |
2531 | } |
2532 | } | |
a4c2f00f PZ |
2533 | |
2534 | /* | |
2535 | * called from enqueue/dequeue and updates the hrtick when the | |
2536 | * current task is from our class and nr_running is low enough | |
2537 | * to matter. | |
2538 | */ | |
2539 | static void hrtick_update(struct rq *rq) | |
2540 | { | |
2541 | struct task_struct *curr = rq->curr; | |
2542 | ||
b39e66ea | 2543 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
2544 | return; |
2545 | ||
2546 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
2547 | hrtick_start_fair(rq, curr); | |
2548 | } | |
55e12e5e | 2549 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
2550 | static inline void |
2551 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2552 | { | |
2553 | } | |
a4c2f00f PZ |
2554 | |
2555 | static inline void hrtick_update(struct rq *rq) | |
2556 | { | |
2557 | } | |
8f4d37ec PZ |
2558 | #endif |
2559 | ||
bf0f6f24 IM |
2560 | /* |
2561 | * The enqueue_task method is called before nr_running is | |
2562 | * increased. Here we update the fair scheduling stats and | |
2563 | * then put the task into the rbtree: | |
2564 | */ | |
ea87bb78 | 2565 | static void |
371fd7e7 | 2566 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2567 | { |
2568 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2569 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
2570 | |
2571 | for_each_sched_entity(se) { | |
62fb1851 | 2572 | if (se->on_rq) |
bf0f6f24 IM |
2573 | break; |
2574 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 2575 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
2576 | |
2577 | /* | |
2578 | * end evaluation on encountering a throttled cfs_rq | |
2579 | * | |
2580 | * note: in the case of encountering a throttled cfs_rq we will | |
2581 | * post the final h_nr_running increment below. | |
2582 | */ | |
2583 | if (cfs_rq_throttled(cfs_rq)) | |
2584 | break; | |
953bfcd1 | 2585 | cfs_rq->h_nr_running++; |
85dac906 | 2586 | |
88ec22d3 | 2587 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 2588 | } |
8f4d37ec | 2589 | |
2069dd75 | 2590 | for_each_sched_entity(se) { |
0f317143 | 2591 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2592 | cfs_rq->h_nr_running++; |
2069dd75 | 2593 | |
85dac906 PT |
2594 | if (cfs_rq_throttled(cfs_rq)) |
2595 | break; | |
2596 | ||
9ee474f5 | 2597 | update_entity_load_avg(se, 1); |
f269ae04 | 2598 | update_cfs_rq_blocked_load(cfs_rq, 0); |
2069dd75 PZ |
2599 | } |
2600 | ||
18bf2805 BS |
2601 | if (!se) { |
2602 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 2603 | inc_nr_running(rq); |
18bf2805 | 2604 | } |
a4c2f00f | 2605 | hrtick_update(rq); |
bf0f6f24 IM |
2606 | } |
2607 | ||
2f36825b VP |
2608 | static void set_next_buddy(struct sched_entity *se); |
2609 | ||
bf0f6f24 IM |
2610 | /* |
2611 | * The dequeue_task method is called before nr_running is | |
2612 | * decreased. We remove the task from the rbtree and | |
2613 | * update the fair scheduling stats: | |
2614 | */ | |
371fd7e7 | 2615 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2616 | { |
2617 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2618 | struct sched_entity *se = &p->se; |
2f36825b | 2619 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
2620 | |
2621 | for_each_sched_entity(se) { | |
2622 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 2623 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
2624 | |
2625 | /* | |
2626 | * end evaluation on encountering a throttled cfs_rq | |
2627 | * | |
2628 | * note: in the case of encountering a throttled cfs_rq we will | |
2629 | * post the final h_nr_running decrement below. | |
2630 | */ | |
2631 | if (cfs_rq_throttled(cfs_rq)) | |
2632 | break; | |
953bfcd1 | 2633 | cfs_rq->h_nr_running--; |
2069dd75 | 2634 | |
bf0f6f24 | 2635 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
2636 | if (cfs_rq->load.weight) { |
2637 | /* | |
2638 | * Bias pick_next to pick a task from this cfs_rq, as | |
2639 | * p is sleeping when it is within its sched_slice. | |
2640 | */ | |
2641 | if (task_sleep && parent_entity(se)) | |
2642 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
2643 | |
2644 | /* avoid re-evaluating load for this entity */ | |
2645 | se = parent_entity(se); | |
bf0f6f24 | 2646 | break; |
2f36825b | 2647 | } |
371fd7e7 | 2648 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 2649 | } |
8f4d37ec | 2650 | |
2069dd75 | 2651 | for_each_sched_entity(se) { |
0f317143 | 2652 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2653 | cfs_rq->h_nr_running--; |
2069dd75 | 2654 | |
85dac906 PT |
2655 | if (cfs_rq_throttled(cfs_rq)) |
2656 | break; | |
2657 | ||
9ee474f5 | 2658 | update_entity_load_avg(se, 1); |
f269ae04 | 2659 | update_cfs_rq_blocked_load(cfs_rq, 0); |
2069dd75 PZ |
2660 | } |
2661 | ||
18bf2805 | 2662 | if (!se) { |
85dac906 | 2663 | dec_nr_running(rq); |
18bf2805 BS |
2664 | update_rq_runnable_avg(rq, 1); |
2665 | } | |
a4c2f00f | 2666 | hrtick_update(rq); |
bf0f6f24 IM |
2667 | } |
2668 | ||
e7693a36 | 2669 | #ifdef CONFIG_SMP |
029632fb PZ |
2670 | /* Used instead of source_load when we know the type == 0 */ |
2671 | static unsigned long weighted_cpuload(const int cpu) | |
2672 | { | |
2673 | return cpu_rq(cpu)->load.weight; | |
2674 | } | |
2675 | ||
2676 | /* | |
2677 | * Return a low guess at the load of a migration-source cpu weighted | |
2678 | * according to the scheduling class and "nice" value. | |
2679 | * | |
2680 | * We want to under-estimate the load of migration sources, to | |
2681 | * balance conservatively. | |
2682 | */ | |
2683 | static unsigned long source_load(int cpu, int type) | |
2684 | { | |
2685 | struct rq *rq = cpu_rq(cpu); | |
2686 | unsigned long total = weighted_cpuload(cpu); | |
2687 | ||
2688 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2689 | return total; | |
2690 | ||
2691 | return min(rq->cpu_load[type-1], total); | |
2692 | } | |
2693 | ||
2694 | /* | |
2695 | * Return a high guess at the load of a migration-target cpu weighted | |
2696 | * according to the scheduling class and "nice" value. | |
2697 | */ | |
2698 | static unsigned long target_load(int cpu, int type) | |
2699 | { | |
2700 | struct rq *rq = cpu_rq(cpu); | |
2701 | unsigned long total = weighted_cpuload(cpu); | |
2702 | ||
2703 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2704 | return total; | |
2705 | ||
2706 | return max(rq->cpu_load[type-1], total); | |
2707 | } | |
2708 | ||
2709 | static unsigned long power_of(int cpu) | |
2710 | { | |
2711 | return cpu_rq(cpu)->cpu_power; | |
2712 | } | |
2713 | ||
2714 | static unsigned long cpu_avg_load_per_task(int cpu) | |
2715 | { | |
2716 | struct rq *rq = cpu_rq(cpu); | |
2717 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
2718 | ||
2719 | if (nr_running) | |
2720 | return rq->load.weight / nr_running; | |
2721 | ||
2722 | return 0; | |
2723 | } | |
2724 | ||
098fb9db | 2725 | |
74f8e4b2 | 2726 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
2727 | { |
2728 | struct sched_entity *se = &p->se; | |
2729 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
2730 | u64 min_vruntime; |
2731 | ||
2732 | #ifndef CONFIG_64BIT | |
2733 | u64 min_vruntime_copy; | |
88ec22d3 | 2734 | |
3fe1698b PZ |
2735 | do { |
2736 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
2737 | smp_rmb(); | |
2738 | min_vruntime = cfs_rq->min_vruntime; | |
2739 | } while (min_vruntime != min_vruntime_copy); | |
2740 | #else | |
2741 | min_vruntime = cfs_rq->min_vruntime; | |
2742 | #endif | |
88ec22d3 | 2743 | |
3fe1698b | 2744 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
2745 | } |
2746 | ||
bb3469ac | 2747 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
2748 | /* |
2749 | * effective_load() calculates the load change as seen from the root_task_group | |
2750 | * | |
2751 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
2752 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
2753 | * can calculate the shift in shares. | |
cf5f0acf PZ |
2754 | * |
2755 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
2756 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
2757 | * total group weight. | |
2758 | * | |
2759 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
2760 | * distribution (s_i) using: | |
2761 | * | |
2762 | * s_i = rw_i / \Sum rw_j (1) | |
2763 | * | |
2764 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
2765 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
2766 | * shares distribution (s_i): | |
2767 | * | |
2768 | * rw_i = { 2, 4, 1, 0 } | |
2769 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
2770 | * | |
2771 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
2772 | * task used to run on and the CPU the waker is running on), we need to | |
2773 | * compute the effect of waking a task on either CPU and, in case of a sync | |
2774 | * wakeup, compute the effect of the current task going to sleep. | |
2775 | * | |
2776 | * So for a change of @wl to the local @cpu with an overall group weight change | |
2777 | * of @wl we can compute the new shares distribution (s'_i) using: | |
2778 | * | |
2779 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
2780 | * | |
2781 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
2782 | * differences in waking a task to CPU 0. The additional task changes the | |
2783 | * weight and shares distributions like: | |
2784 | * | |
2785 | * rw'_i = { 3, 4, 1, 0 } | |
2786 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
2787 | * | |
2788 | * We can then compute the difference in effective weight by using: | |
2789 | * | |
2790 | * dw_i = S * (s'_i - s_i) (3) | |
2791 | * | |
2792 | * Where 'S' is the group weight as seen by its parent. | |
2793 | * | |
2794 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
2795 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
2796 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 2797 | */ |
2069dd75 | 2798 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 2799 | { |
4be9daaa | 2800 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 2801 | |
cf5f0acf | 2802 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
2803 | return wl; |
2804 | ||
4be9daaa | 2805 | for_each_sched_entity(se) { |
cf5f0acf | 2806 | long w, W; |
4be9daaa | 2807 | |
977dda7c | 2808 | tg = se->my_q->tg; |
bb3469ac | 2809 | |
cf5f0acf PZ |
2810 | /* |
2811 | * W = @wg + \Sum rw_j | |
2812 | */ | |
2813 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 2814 | |
cf5f0acf PZ |
2815 | /* |
2816 | * w = rw_i + @wl | |
2817 | */ | |
2818 | w = se->my_q->load.weight + wl; | |
940959e9 | 2819 | |
cf5f0acf PZ |
2820 | /* |
2821 | * wl = S * s'_i; see (2) | |
2822 | */ | |
2823 | if (W > 0 && w < W) | |
2824 | wl = (w * tg->shares) / W; | |
977dda7c PT |
2825 | else |
2826 | wl = tg->shares; | |
940959e9 | 2827 | |
cf5f0acf PZ |
2828 | /* |
2829 | * Per the above, wl is the new se->load.weight value; since | |
2830 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
2831 | * calc_cfs_shares(). | |
2832 | */ | |
977dda7c PT |
2833 | if (wl < MIN_SHARES) |
2834 | wl = MIN_SHARES; | |
cf5f0acf PZ |
2835 | |
2836 | /* | |
2837 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
2838 | */ | |
977dda7c | 2839 | wl -= se->load.weight; |
cf5f0acf PZ |
2840 | |
2841 | /* | |
2842 | * Recursively apply this logic to all parent groups to compute | |
2843 | * the final effective load change on the root group. Since | |
2844 | * only the @tg group gets extra weight, all parent groups can | |
2845 | * only redistribute existing shares. @wl is the shift in shares | |
2846 | * resulting from this level per the above. | |
2847 | */ | |
4be9daaa | 2848 | wg = 0; |
4be9daaa | 2849 | } |
bb3469ac | 2850 | |
4be9daaa | 2851 | return wl; |
bb3469ac PZ |
2852 | } |
2853 | #else | |
4be9daaa | 2854 | |
83378269 PZ |
2855 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
2856 | unsigned long wl, unsigned long wg) | |
4be9daaa | 2857 | { |
83378269 | 2858 | return wl; |
bb3469ac | 2859 | } |
4be9daaa | 2860 | |
bb3469ac PZ |
2861 | #endif |
2862 | ||
c88d5910 | 2863 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 2864 | { |
e37b6a7b | 2865 | s64 this_load, load; |
c88d5910 | 2866 | int idx, this_cpu, prev_cpu; |
098fb9db | 2867 | unsigned long tl_per_task; |
c88d5910 | 2868 | struct task_group *tg; |
83378269 | 2869 | unsigned long weight; |
b3137bc8 | 2870 | int balanced; |
098fb9db | 2871 | |
c88d5910 PZ |
2872 | idx = sd->wake_idx; |
2873 | this_cpu = smp_processor_id(); | |
2874 | prev_cpu = task_cpu(p); | |
2875 | load = source_load(prev_cpu, idx); | |
2876 | this_load = target_load(this_cpu, idx); | |
098fb9db | 2877 | |
b3137bc8 MG |
2878 | /* |
2879 | * If sync wakeup then subtract the (maximum possible) | |
2880 | * effect of the currently running task from the load | |
2881 | * of the current CPU: | |
2882 | */ | |
83378269 PZ |
2883 | if (sync) { |
2884 | tg = task_group(current); | |
2885 | weight = current->se.load.weight; | |
2886 | ||
c88d5910 | 2887 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
2888 | load += effective_load(tg, prev_cpu, 0, -weight); |
2889 | } | |
b3137bc8 | 2890 | |
83378269 PZ |
2891 | tg = task_group(p); |
2892 | weight = p->se.load.weight; | |
b3137bc8 | 2893 | |
71a29aa7 PZ |
2894 | /* |
2895 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
2896 | * due to the sync cause above having dropped this_load to 0, we'll |
2897 | * always have an imbalance, but there's really nothing you can do | |
2898 | * about that, so that's good too. | |
71a29aa7 PZ |
2899 | * |
2900 | * Otherwise check if either cpus are near enough in load to allow this | |
2901 | * task to be woken on this_cpu. | |
2902 | */ | |
e37b6a7b PT |
2903 | if (this_load > 0) { |
2904 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
2905 | |
2906 | this_eff_load = 100; | |
2907 | this_eff_load *= power_of(prev_cpu); | |
2908 | this_eff_load *= this_load + | |
2909 | effective_load(tg, this_cpu, weight, weight); | |
2910 | ||
2911 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
2912 | prev_eff_load *= power_of(this_cpu); | |
2913 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
2914 | ||
2915 | balanced = this_eff_load <= prev_eff_load; | |
2916 | } else | |
2917 | balanced = true; | |
b3137bc8 | 2918 | |
098fb9db | 2919 | /* |
4ae7d5ce IM |
2920 | * If the currently running task will sleep within |
2921 | * a reasonable amount of time then attract this newly | |
2922 | * woken task: | |
098fb9db | 2923 | */ |
2fb7635c PZ |
2924 | if (sync && balanced) |
2925 | return 1; | |
098fb9db | 2926 | |
41acab88 | 2927 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
2928 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
2929 | ||
c88d5910 PZ |
2930 | if (balanced || |
2931 | (this_load <= load && | |
2932 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
2933 | /* |
2934 | * This domain has SD_WAKE_AFFINE and | |
2935 | * p is cache cold in this domain, and | |
2936 | * there is no bad imbalance. | |
2937 | */ | |
c88d5910 | 2938 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 2939 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
2940 | |
2941 | return 1; | |
2942 | } | |
2943 | return 0; | |
2944 | } | |
2945 | ||
aaee1203 PZ |
2946 | /* |
2947 | * find_idlest_group finds and returns the least busy CPU group within the | |
2948 | * domain. | |
2949 | */ | |
2950 | static struct sched_group * | |
78e7ed53 | 2951 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 2952 | int this_cpu, int load_idx) |
e7693a36 | 2953 | { |
b3bd3de6 | 2954 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 2955 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 2956 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 2957 | |
aaee1203 PZ |
2958 | do { |
2959 | unsigned long load, avg_load; | |
2960 | int local_group; | |
2961 | int i; | |
e7693a36 | 2962 | |
aaee1203 PZ |
2963 | /* Skip over this group if it has no CPUs allowed */ |
2964 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 2965 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
2966 | continue; |
2967 | ||
2968 | local_group = cpumask_test_cpu(this_cpu, | |
2969 | sched_group_cpus(group)); | |
2970 | ||
2971 | /* Tally up the load of all CPUs in the group */ | |
2972 | avg_load = 0; | |
2973 | ||
2974 | for_each_cpu(i, sched_group_cpus(group)) { | |
2975 | /* Bias balancing toward cpus of our domain */ | |
2976 | if (local_group) | |
2977 | load = source_load(i, load_idx); | |
2978 | else | |
2979 | load = target_load(i, load_idx); | |
2980 | ||
2981 | avg_load += load; | |
2982 | } | |
2983 | ||
2984 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 2985 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
2986 | |
2987 | if (local_group) { | |
2988 | this_load = avg_load; | |
aaee1203 PZ |
2989 | } else if (avg_load < min_load) { |
2990 | min_load = avg_load; | |
2991 | idlest = group; | |
2992 | } | |
2993 | } while (group = group->next, group != sd->groups); | |
2994 | ||
2995 | if (!idlest || 100*this_load < imbalance*min_load) | |
2996 | return NULL; | |
2997 | return idlest; | |
2998 | } | |
2999 | ||
3000 | /* | |
3001 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
3002 | */ | |
3003 | static int | |
3004 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
3005 | { | |
3006 | unsigned long load, min_load = ULONG_MAX; | |
3007 | int idlest = -1; | |
3008 | int i; | |
3009 | ||
3010 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 3011 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
3012 | load = weighted_cpuload(i); |
3013 | ||
3014 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
3015 | min_load = load; | |
3016 | idlest = i; | |
e7693a36 GH |
3017 | } |
3018 | } | |
3019 | ||
aaee1203 PZ |
3020 | return idlest; |
3021 | } | |
e7693a36 | 3022 | |
a50bde51 PZ |
3023 | /* |
3024 | * Try and locate an idle CPU in the sched_domain. | |
3025 | */ | |
99bd5e2f | 3026 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 PZ |
3027 | { |
3028 | int cpu = smp_processor_id(); | |
3029 | int prev_cpu = task_cpu(p); | |
99bd5e2f | 3030 | struct sched_domain *sd; |
37407ea7 LT |
3031 | struct sched_group *sg; |
3032 | int i; | |
a50bde51 PZ |
3033 | |
3034 | /* | |
99bd5e2f SS |
3035 | * If the task is going to be woken-up on this cpu and if it is |
3036 | * already idle, then it is the right target. | |
a50bde51 | 3037 | */ |
99bd5e2f SS |
3038 | if (target == cpu && idle_cpu(cpu)) |
3039 | return cpu; | |
3040 | ||
3041 | /* | |
3042 | * If the task is going to be woken-up on the cpu where it previously | |
3043 | * ran and if it is currently idle, then it the right target. | |
3044 | */ | |
3045 | if (target == prev_cpu && idle_cpu(prev_cpu)) | |
fe3bcfe1 | 3046 | return prev_cpu; |
a50bde51 PZ |
3047 | |
3048 | /* | |
37407ea7 | 3049 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 3050 | */ |
518cd623 | 3051 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 3052 | for_each_lower_domain(sd) { |
37407ea7 LT |
3053 | sg = sd->groups; |
3054 | do { | |
3055 | if (!cpumask_intersects(sched_group_cpus(sg), | |
3056 | tsk_cpus_allowed(p))) | |
3057 | goto next; | |
3058 | ||
3059 | for_each_cpu(i, sched_group_cpus(sg)) { | |
3060 | if (!idle_cpu(i)) | |
3061 | goto next; | |
3062 | } | |
970e1789 | 3063 | |
37407ea7 LT |
3064 | target = cpumask_first_and(sched_group_cpus(sg), |
3065 | tsk_cpus_allowed(p)); | |
3066 | goto done; | |
3067 | next: | |
3068 | sg = sg->next; | |
3069 | } while (sg != sd->groups); | |
3070 | } | |
3071 | done: | |
a50bde51 PZ |
3072 | return target; |
3073 | } | |
3074 | ||
aaee1203 PZ |
3075 | /* |
3076 | * sched_balance_self: balance the current task (running on cpu) in domains | |
3077 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
3078 | * SD_BALANCE_EXEC. | |
3079 | * | |
3080 | * Balance, ie. select the least loaded group. | |
3081 | * | |
3082 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
3083 | * | |
3084 | * preempt must be disabled. | |
3085 | */ | |
0017d735 | 3086 | static int |
7608dec2 | 3087 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 3088 | { |
29cd8bae | 3089 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
3090 | int cpu = smp_processor_id(); |
3091 | int prev_cpu = task_cpu(p); | |
3092 | int new_cpu = cpu; | |
99bd5e2f | 3093 | int want_affine = 0; |
5158f4e4 | 3094 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 3095 | |
29baa747 | 3096 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
3097 | return prev_cpu; |
3098 | ||
0763a660 | 3099 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 3100 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
3101 | want_affine = 1; |
3102 | new_cpu = prev_cpu; | |
3103 | } | |
aaee1203 | 3104 | |
dce840a0 | 3105 | rcu_read_lock(); |
aaee1203 | 3106 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
3107 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
3108 | continue; | |
3109 | ||
fe3bcfe1 | 3110 | /* |
99bd5e2f SS |
3111 | * If both cpu and prev_cpu are part of this domain, |
3112 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 3113 | */ |
99bd5e2f SS |
3114 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
3115 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
3116 | affine_sd = tmp; | |
29cd8bae | 3117 | break; |
f03542a7 | 3118 | } |
29cd8bae | 3119 | |
f03542a7 | 3120 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
3121 | sd = tmp; |
3122 | } | |
3123 | ||
8b911acd | 3124 | if (affine_sd) { |
f03542a7 | 3125 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
3126 | prev_cpu = cpu; |
3127 | ||
3128 | new_cpu = select_idle_sibling(p, prev_cpu); | |
3129 | goto unlock; | |
8b911acd | 3130 | } |
e7693a36 | 3131 | |
aaee1203 | 3132 | while (sd) { |
5158f4e4 | 3133 | int load_idx = sd->forkexec_idx; |
aaee1203 | 3134 | struct sched_group *group; |
c88d5910 | 3135 | int weight; |
098fb9db | 3136 | |
0763a660 | 3137 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
3138 | sd = sd->child; |
3139 | continue; | |
3140 | } | |
098fb9db | 3141 | |
5158f4e4 PZ |
3142 | if (sd_flag & SD_BALANCE_WAKE) |
3143 | load_idx = sd->wake_idx; | |
098fb9db | 3144 | |
5158f4e4 | 3145 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
3146 | if (!group) { |
3147 | sd = sd->child; | |
3148 | continue; | |
3149 | } | |
4ae7d5ce | 3150 | |
d7c33c49 | 3151 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
3152 | if (new_cpu == -1 || new_cpu == cpu) { |
3153 | /* Now try balancing at a lower domain level of cpu */ | |
3154 | sd = sd->child; | |
3155 | continue; | |
e7693a36 | 3156 | } |
aaee1203 PZ |
3157 | |
3158 | /* Now try balancing at a lower domain level of new_cpu */ | |
3159 | cpu = new_cpu; | |
669c55e9 | 3160 | weight = sd->span_weight; |
aaee1203 PZ |
3161 | sd = NULL; |
3162 | for_each_domain(cpu, tmp) { | |
669c55e9 | 3163 | if (weight <= tmp->span_weight) |
aaee1203 | 3164 | break; |
0763a660 | 3165 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
3166 | sd = tmp; |
3167 | } | |
3168 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 3169 | } |
dce840a0 PZ |
3170 | unlock: |
3171 | rcu_read_unlock(); | |
e7693a36 | 3172 | |
c88d5910 | 3173 | return new_cpu; |
e7693a36 | 3174 | } |
0a74bef8 PT |
3175 | |
3176 | /* | |
3177 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
3178 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
3179 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
3180 | * other assumptions, including the state of rq->lock, should be made. | |
3181 | */ | |
3182 | static void | |
3183 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
3184 | { | |
aff3e498 PT |
3185 | struct sched_entity *se = &p->se; |
3186 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3187 | ||
3188 | /* | |
3189 | * Load tracking: accumulate removed load so that it can be processed | |
3190 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
3191 | * to blocked load iff they have a positive decay-count. It can never | |
3192 | * be negative here since on-rq tasks have decay-count == 0. | |
3193 | */ | |
3194 | if (se->avg.decay_count) { | |
3195 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
3196 | atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load); | |
3197 | } | |
0a74bef8 | 3198 | } |
e7693a36 GH |
3199 | #endif /* CONFIG_SMP */ |
3200 | ||
e52fb7c0 PZ |
3201 | static unsigned long |
3202 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
3203 | { |
3204 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
3205 | ||
3206 | /* | |
e52fb7c0 PZ |
3207 | * Since its curr running now, convert the gran from real-time |
3208 | * to virtual-time in his units. | |
13814d42 MG |
3209 | * |
3210 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
3211 | * they get preempted easier. That is, if 'se' < 'curr' then | |
3212 | * the resulting gran will be larger, therefore penalizing the | |
3213 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
3214 | * be smaller, again penalizing the lighter task. | |
3215 | * | |
3216 | * This is especially important for buddies when the leftmost | |
3217 | * task is higher priority than the buddy. | |
0bbd3336 | 3218 | */ |
f4ad9bd2 | 3219 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
3220 | } |
3221 | ||
464b7527 PZ |
3222 | /* |
3223 | * Should 'se' preempt 'curr'. | |
3224 | * | |
3225 | * |s1 | |
3226 | * |s2 | |
3227 | * |s3 | |
3228 | * g | |
3229 | * |<--->|c | |
3230 | * | |
3231 | * w(c, s1) = -1 | |
3232 | * w(c, s2) = 0 | |
3233 | * w(c, s3) = 1 | |
3234 | * | |
3235 | */ | |
3236 | static int | |
3237 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
3238 | { | |
3239 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
3240 | ||
3241 | if (vdiff <= 0) | |
3242 | return -1; | |
3243 | ||
e52fb7c0 | 3244 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
3245 | if (vdiff > gran) |
3246 | return 1; | |
3247 | ||
3248 | return 0; | |
3249 | } | |
3250 | ||
02479099 PZ |
3251 | static void set_last_buddy(struct sched_entity *se) |
3252 | { | |
69c80f3e VP |
3253 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3254 | return; | |
3255 | ||
3256 | for_each_sched_entity(se) | |
3257 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
3258 | } |
3259 | ||
3260 | static void set_next_buddy(struct sched_entity *se) | |
3261 | { | |
69c80f3e VP |
3262 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3263 | return; | |
3264 | ||
3265 | for_each_sched_entity(se) | |
3266 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
3267 | } |
3268 | ||
ac53db59 RR |
3269 | static void set_skip_buddy(struct sched_entity *se) |
3270 | { | |
69c80f3e VP |
3271 | for_each_sched_entity(se) |
3272 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
3273 | } |
3274 | ||
bf0f6f24 IM |
3275 | /* |
3276 | * Preempt the current task with a newly woken task if needed: | |
3277 | */ | |
5a9b86f6 | 3278 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
3279 | { |
3280 | struct task_struct *curr = rq->curr; | |
8651a86c | 3281 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 3282 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 3283 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 3284 | int next_buddy_marked = 0; |
bf0f6f24 | 3285 | |
4ae7d5ce IM |
3286 | if (unlikely(se == pse)) |
3287 | return; | |
3288 | ||
5238cdd3 | 3289 | /* |
ddcdf6e7 | 3290 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
3291 | * unconditionally check_prempt_curr() after an enqueue (which may have |
3292 | * lead to a throttle). This both saves work and prevents false | |
3293 | * next-buddy nomination below. | |
3294 | */ | |
3295 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3296 | return; | |
3297 | ||
2f36825b | 3298 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3299 | set_next_buddy(pse); |
2f36825b VP |
3300 | next_buddy_marked = 1; |
3301 | } | |
57fdc26d | 3302 | |
aec0a514 BR |
3303 | /* |
3304 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3305 | * wake up path. | |
5238cdd3 PT |
3306 | * |
3307 | * Note: this also catches the edge-case of curr being in a throttled | |
3308 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3309 | * enqueue of curr) will have resulted in resched being set. This | |
3310 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3311 | * below. | |
aec0a514 BR |
3312 | */ |
3313 | if (test_tsk_need_resched(curr)) | |
3314 | return; | |
3315 | ||
a2f5c9ab DH |
3316 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3317 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3318 | likely(p->policy != SCHED_IDLE)) | |
3319 | goto preempt; | |
3320 | ||
91c234b4 | 3321 | /* |
a2f5c9ab DH |
3322 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3323 | * is driven by the tick): | |
91c234b4 | 3324 | */ |
6bc912b7 | 3325 | if (unlikely(p->policy != SCHED_NORMAL)) |
91c234b4 | 3326 | return; |
bf0f6f24 | 3327 | |
464b7527 | 3328 | find_matching_se(&se, &pse); |
9bbd7374 | 3329 | update_curr(cfs_rq_of(se)); |
002f128b | 3330 | BUG_ON(!pse); |
2f36825b VP |
3331 | if (wakeup_preempt_entity(se, pse) == 1) { |
3332 | /* | |
3333 | * Bias pick_next to pick the sched entity that is | |
3334 | * triggering this preemption. | |
3335 | */ | |
3336 | if (!next_buddy_marked) | |
3337 | set_next_buddy(pse); | |
3a7e73a2 | 3338 | goto preempt; |
2f36825b | 3339 | } |
464b7527 | 3340 | |
3a7e73a2 | 3341 | return; |
a65ac745 | 3342 | |
3a7e73a2 PZ |
3343 | preempt: |
3344 | resched_task(curr); | |
3345 | /* | |
3346 | * Only set the backward buddy when the current task is still | |
3347 | * on the rq. This can happen when a wakeup gets interleaved | |
3348 | * with schedule on the ->pre_schedule() or idle_balance() | |
3349 | * point, either of which can * drop the rq lock. | |
3350 | * | |
3351 | * Also, during early boot the idle thread is in the fair class, | |
3352 | * for obvious reasons its a bad idea to schedule back to it. | |
3353 | */ | |
3354 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3355 | return; | |
3356 | ||
3357 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3358 | set_last_buddy(se); | |
bf0f6f24 IM |
3359 | } |
3360 | ||
fb8d4724 | 3361 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3362 | { |
8f4d37ec | 3363 | struct task_struct *p; |
bf0f6f24 IM |
3364 | struct cfs_rq *cfs_rq = &rq->cfs; |
3365 | struct sched_entity *se; | |
3366 | ||
36ace27e | 3367 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3368 | return NULL; |
3369 | ||
3370 | do { | |
9948f4b2 | 3371 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3372 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3373 | cfs_rq = group_cfs_rq(se); |
3374 | } while (cfs_rq); | |
3375 | ||
8f4d37ec | 3376 | p = task_of(se); |
b39e66ea MG |
3377 | if (hrtick_enabled(rq)) |
3378 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
3379 | |
3380 | return p; | |
bf0f6f24 IM |
3381 | } |
3382 | ||
3383 | /* | |
3384 | * Account for a descheduled task: | |
3385 | */ | |
31ee529c | 3386 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
3387 | { |
3388 | struct sched_entity *se = &prev->se; | |
3389 | struct cfs_rq *cfs_rq; | |
3390 | ||
3391 | for_each_sched_entity(se) { | |
3392 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 3393 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
3394 | } |
3395 | } | |
3396 | ||
ac53db59 RR |
3397 | /* |
3398 | * sched_yield() is very simple | |
3399 | * | |
3400 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
3401 | */ | |
3402 | static void yield_task_fair(struct rq *rq) | |
3403 | { | |
3404 | struct task_struct *curr = rq->curr; | |
3405 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
3406 | struct sched_entity *se = &curr->se; | |
3407 | ||
3408 | /* | |
3409 | * Are we the only task in the tree? | |
3410 | */ | |
3411 | if (unlikely(rq->nr_running == 1)) | |
3412 | return; | |
3413 | ||
3414 | clear_buddies(cfs_rq, se); | |
3415 | ||
3416 | if (curr->policy != SCHED_BATCH) { | |
3417 | update_rq_clock(rq); | |
3418 | /* | |
3419 | * Update run-time statistics of the 'current'. | |
3420 | */ | |
3421 | update_curr(cfs_rq); | |
916671c0 MG |
3422 | /* |
3423 | * Tell update_rq_clock() that we've just updated, | |
3424 | * so we don't do microscopic update in schedule() | |
3425 | * and double the fastpath cost. | |
3426 | */ | |
3427 | rq->skip_clock_update = 1; | |
ac53db59 RR |
3428 | } |
3429 | ||
3430 | set_skip_buddy(se); | |
3431 | } | |
3432 | ||
d95f4122 MG |
3433 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
3434 | { | |
3435 | struct sched_entity *se = &p->se; | |
3436 | ||
5238cdd3 PT |
3437 | /* throttled hierarchies are not runnable */ |
3438 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
3439 | return false; |
3440 | ||
3441 | /* Tell the scheduler that we'd really like pse to run next. */ | |
3442 | set_next_buddy(se); | |
3443 | ||
d95f4122 MG |
3444 | yield_task_fair(rq); |
3445 | ||
3446 | return true; | |
3447 | } | |
3448 | ||
681f3e68 | 3449 | #ifdef CONFIG_SMP |
bf0f6f24 IM |
3450 | /************************************************** |
3451 | * Fair scheduling class load-balancing methods: | |
3452 | */ | |
3453 | ||
ed387b78 HS |
3454 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3455 | ||
ddcdf6e7 | 3456 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 3457 | #define LBF_NEED_BREAK 0x02 |
88b8dac0 | 3458 | #define LBF_SOME_PINNED 0x04 |
ddcdf6e7 PZ |
3459 | |
3460 | struct lb_env { | |
3461 | struct sched_domain *sd; | |
3462 | ||
ddcdf6e7 | 3463 | struct rq *src_rq; |
85c1e7da | 3464 | int src_cpu; |
ddcdf6e7 PZ |
3465 | |
3466 | int dst_cpu; | |
3467 | struct rq *dst_rq; | |
3468 | ||
88b8dac0 SV |
3469 | struct cpumask *dst_grpmask; |
3470 | int new_dst_cpu; | |
ddcdf6e7 | 3471 | enum cpu_idle_type idle; |
bd939f45 | 3472 | long imbalance; |
b9403130 MW |
3473 | /* The set of CPUs under consideration for load-balancing */ |
3474 | struct cpumask *cpus; | |
3475 | ||
ddcdf6e7 | 3476 | unsigned int flags; |
367456c7 PZ |
3477 | |
3478 | unsigned int loop; | |
3479 | unsigned int loop_break; | |
3480 | unsigned int loop_max; | |
ddcdf6e7 PZ |
3481 | }; |
3482 | ||
1e3c88bd | 3483 | /* |
ddcdf6e7 | 3484 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
3485 | * Both runqueues must be locked. |
3486 | */ | |
ddcdf6e7 | 3487 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 3488 | { |
ddcdf6e7 PZ |
3489 | deactivate_task(env->src_rq, p, 0); |
3490 | set_task_cpu(p, env->dst_cpu); | |
3491 | activate_task(env->dst_rq, p, 0); | |
3492 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
3493 | } |
3494 | ||
029632fb PZ |
3495 | /* |
3496 | * Is this task likely cache-hot: | |
3497 | */ | |
3498 | static int | |
3499 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
3500 | { | |
3501 | s64 delta; | |
3502 | ||
3503 | if (p->sched_class != &fair_sched_class) | |
3504 | return 0; | |
3505 | ||
3506 | if (unlikely(p->policy == SCHED_IDLE)) | |
3507 | return 0; | |
3508 | ||
3509 | /* | |
3510 | * Buddy candidates are cache hot: | |
3511 | */ | |
3512 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
3513 | (&p->se == cfs_rq_of(&p->se)->next || | |
3514 | &p->se == cfs_rq_of(&p->se)->last)) | |
3515 | return 1; | |
3516 | ||
3517 | if (sysctl_sched_migration_cost == -1) | |
3518 | return 1; | |
3519 | if (sysctl_sched_migration_cost == 0) | |
3520 | return 0; | |
3521 | ||
3522 | delta = now - p->se.exec_start; | |
3523 | ||
3524 | return delta < (s64)sysctl_sched_migration_cost; | |
3525 | } | |
3526 | ||
1e3c88bd PZ |
3527 | /* |
3528 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3529 | */ | |
3530 | static | |
8e45cb54 | 3531 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
3532 | { |
3533 | int tsk_cache_hot = 0; | |
3534 | /* | |
3535 | * We do not migrate tasks that are: | |
3536 | * 1) running (obviously), or | |
3537 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3538 | * 3) are cache-hot on their current CPU. | |
3539 | */ | |
ddcdf6e7 | 3540 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
88b8dac0 SV |
3541 | int new_dst_cpu; |
3542 | ||
41acab88 | 3543 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 SV |
3544 | |
3545 | /* | |
3546 | * Remember if this task can be migrated to any other cpu in | |
3547 | * our sched_group. We may want to revisit it if we couldn't | |
3548 | * meet load balance goals by pulling other tasks on src_cpu. | |
3549 | * | |
3550 | * Also avoid computing new_dst_cpu if we have already computed | |
3551 | * one in current iteration. | |
3552 | */ | |
3553 | if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED)) | |
3554 | return 0; | |
3555 | ||
3556 | new_dst_cpu = cpumask_first_and(env->dst_grpmask, | |
3557 | tsk_cpus_allowed(p)); | |
3558 | if (new_dst_cpu < nr_cpu_ids) { | |
3559 | env->flags |= LBF_SOME_PINNED; | |
3560 | env->new_dst_cpu = new_dst_cpu; | |
3561 | } | |
1e3c88bd PZ |
3562 | return 0; |
3563 | } | |
88b8dac0 SV |
3564 | |
3565 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 3566 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 3567 | |
ddcdf6e7 | 3568 | if (task_running(env->src_rq, p)) { |
41acab88 | 3569 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
3570 | return 0; |
3571 | } | |
3572 | ||
3573 | /* | |
3574 | * Aggressive migration if: | |
3575 | * 1) task is cache cold, or | |
3576 | * 2) too many balance attempts have failed. | |
3577 | */ | |
3578 | ||
ddcdf6e7 | 3579 | tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd); |
1e3c88bd | 3580 | if (!tsk_cache_hot || |
8e45cb54 | 3581 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
1e3c88bd PZ |
3582 | #ifdef CONFIG_SCHEDSTATS |
3583 | if (tsk_cache_hot) { | |
8e45cb54 | 3584 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 3585 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd PZ |
3586 | } |
3587 | #endif | |
3588 | return 1; | |
3589 | } | |
3590 | ||
3591 | if (tsk_cache_hot) { | |
41acab88 | 3592 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
1e3c88bd PZ |
3593 | return 0; |
3594 | } | |
3595 | return 1; | |
3596 | } | |
3597 | ||
897c395f PZ |
3598 | /* |
3599 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3600 | * part of active balancing operations within "domain". | |
3601 | * Returns 1 if successful and 0 otherwise. | |
3602 | * | |
3603 | * Called with both runqueues locked. | |
3604 | */ | |
8e45cb54 | 3605 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
3606 | { |
3607 | struct task_struct *p, *n; | |
897c395f | 3608 | |
367456c7 PZ |
3609 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
3610 | if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu)) | |
3611 | continue; | |
897c395f | 3612 | |
367456c7 PZ |
3613 | if (!can_migrate_task(p, env)) |
3614 | continue; | |
897c395f | 3615 | |
367456c7 PZ |
3616 | move_task(p, env); |
3617 | /* | |
3618 | * Right now, this is only the second place move_task() | |
3619 | * is called, so we can safely collect move_task() | |
3620 | * stats here rather than inside move_task(). | |
3621 | */ | |
3622 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
3623 | return 1; | |
897c395f | 3624 | } |
897c395f PZ |
3625 | return 0; |
3626 | } | |
3627 | ||
367456c7 PZ |
3628 | static unsigned long task_h_load(struct task_struct *p); |
3629 | ||
eb95308e PZ |
3630 | static const unsigned int sched_nr_migrate_break = 32; |
3631 | ||
5d6523eb | 3632 | /* |
bd939f45 | 3633 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
3634 | * this_rq, as part of a balancing operation within domain "sd". |
3635 | * Returns 1 if successful and 0 otherwise. | |
3636 | * | |
3637 | * Called with both runqueues locked. | |
3638 | */ | |
3639 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 3640 | { |
5d6523eb PZ |
3641 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
3642 | struct task_struct *p; | |
367456c7 PZ |
3643 | unsigned long load; |
3644 | int pulled = 0; | |
1e3c88bd | 3645 | |
bd939f45 | 3646 | if (env->imbalance <= 0) |
5d6523eb | 3647 | return 0; |
1e3c88bd | 3648 | |
5d6523eb PZ |
3649 | while (!list_empty(tasks)) { |
3650 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 3651 | |
367456c7 PZ |
3652 | env->loop++; |
3653 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 3654 | if (env->loop > env->loop_max) |
367456c7 | 3655 | break; |
5d6523eb PZ |
3656 | |
3657 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 3658 | if (env->loop > env->loop_break) { |
eb95308e | 3659 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 3660 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 3661 | break; |
a195f004 | 3662 | } |
1e3c88bd | 3663 | |
5d6523eb | 3664 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
367456c7 PZ |
3665 | goto next; |
3666 | ||
3667 | load = task_h_load(p); | |
5d6523eb | 3668 | |
eb95308e | 3669 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
3670 | goto next; |
3671 | ||
bd939f45 | 3672 | if ((load / 2) > env->imbalance) |
367456c7 | 3673 | goto next; |
1e3c88bd | 3674 | |
367456c7 PZ |
3675 | if (!can_migrate_task(p, env)) |
3676 | goto next; | |
1e3c88bd | 3677 | |
ddcdf6e7 | 3678 | move_task(p, env); |
ee00e66f | 3679 | pulled++; |
bd939f45 | 3680 | env->imbalance -= load; |
1e3c88bd PZ |
3681 | |
3682 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
3683 | /* |
3684 | * NEWIDLE balancing is a source of latency, so preemptible | |
3685 | * kernels will stop after the first task is pulled to minimize | |
3686 | * the critical section. | |
3687 | */ | |
5d6523eb | 3688 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 3689 | break; |
1e3c88bd PZ |
3690 | #endif |
3691 | ||
ee00e66f PZ |
3692 | /* |
3693 | * We only want to steal up to the prescribed amount of | |
3694 | * weighted load. | |
3695 | */ | |
bd939f45 | 3696 | if (env->imbalance <= 0) |
ee00e66f | 3697 | break; |
367456c7 PZ |
3698 | |
3699 | continue; | |
3700 | next: | |
5d6523eb | 3701 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 3702 | } |
5d6523eb | 3703 | |
1e3c88bd | 3704 | /* |
ddcdf6e7 PZ |
3705 | * Right now, this is one of only two places move_task() is called, |
3706 | * so we can safely collect move_task() stats here rather than | |
3707 | * inside move_task(). | |
1e3c88bd | 3708 | */ |
8e45cb54 | 3709 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 3710 | |
5d6523eb | 3711 | return pulled; |
1e3c88bd PZ |
3712 | } |
3713 | ||
230059de | 3714 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
3715 | /* |
3716 | * update tg->load_weight by folding this cpu's load_avg | |
3717 | */ | |
48a16753 | 3718 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 3719 | { |
48a16753 PT |
3720 | struct sched_entity *se = tg->se[cpu]; |
3721 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 3722 | |
48a16753 PT |
3723 | /* throttled entities do not contribute to load */ |
3724 | if (throttled_hierarchy(cfs_rq)) | |
3725 | return; | |
9e3081ca | 3726 | |
aff3e498 | 3727 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 3728 | |
82958366 PT |
3729 | if (se) { |
3730 | update_entity_load_avg(se, 1); | |
3731 | /* | |
3732 | * We pivot on our runnable average having decayed to zero for | |
3733 | * list removal. This generally implies that all our children | |
3734 | * have also been removed (modulo rounding error or bandwidth | |
3735 | * control); however, such cases are rare and we can fix these | |
3736 | * at enqueue. | |
3737 | * | |
3738 | * TODO: fix up out-of-order children on enqueue. | |
3739 | */ | |
3740 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
3741 | list_del_leaf_cfs_rq(cfs_rq); | |
3742 | } else { | |
48a16753 | 3743 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
3744 | update_rq_runnable_avg(rq, rq->nr_running); |
3745 | } | |
9e3081ca PZ |
3746 | } |
3747 | ||
48a16753 | 3748 | static void update_blocked_averages(int cpu) |
9e3081ca | 3749 | { |
9e3081ca | 3750 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
3751 | struct cfs_rq *cfs_rq; |
3752 | unsigned long flags; | |
9e3081ca | 3753 | |
48a16753 PT |
3754 | raw_spin_lock_irqsave(&rq->lock, flags); |
3755 | update_rq_clock(rq); | |
9763b67f PZ |
3756 | /* |
3757 | * Iterates the task_group tree in a bottom up fashion, see | |
3758 | * list_add_leaf_cfs_rq() for details. | |
3759 | */ | |
64660c86 | 3760 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
3761 | /* |
3762 | * Note: We may want to consider periodically releasing | |
3763 | * rq->lock about these updates so that creating many task | |
3764 | * groups does not result in continually extending hold time. | |
3765 | */ | |
3766 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 3767 | } |
48a16753 PT |
3768 | |
3769 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
3770 | } |
3771 | ||
9763b67f PZ |
3772 | /* |
3773 | * Compute the cpu's hierarchical load factor for each task group. | |
3774 | * This needs to be done in a top-down fashion because the load of a child | |
3775 | * group is a fraction of its parents load. | |
3776 | */ | |
3777 | static int tg_load_down(struct task_group *tg, void *data) | |
3778 | { | |
3779 | unsigned long load; | |
3780 | long cpu = (long)data; | |
3781 | ||
3782 | if (!tg->parent) { | |
3783 | load = cpu_rq(cpu)->load.weight; | |
3784 | } else { | |
3785 | load = tg->parent->cfs_rq[cpu]->h_load; | |
3786 | load *= tg->se[cpu]->load.weight; | |
3787 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
3788 | } | |
3789 | ||
3790 | tg->cfs_rq[cpu]->h_load = load; | |
3791 | ||
3792 | return 0; | |
3793 | } | |
3794 | ||
3795 | static void update_h_load(long cpu) | |
3796 | { | |
a35b6466 PZ |
3797 | struct rq *rq = cpu_rq(cpu); |
3798 | unsigned long now = jiffies; | |
3799 | ||
3800 | if (rq->h_load_throttle == now) | |
3801 | return; | |
3802 | ||
3803 | rq->h_load_throttle = now; | |
3804 | ||
367456c7 | 3805 | rcu_read_lock(); |
9763b67f | 3806 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
367456c7 | 3807 | rcu_read_unlock(); |
9763b67f PZ |
3808 | } |
3809 | ||
367456c7 | 3810 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 3811 | { |
367456c7 PZ |
3812 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
3813 | unsigned long load; | |
230059de | 3814 | |
367456c7 PZ |
3815 | load = p->se.load.weight; |
3816 | load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1); | |
230059de | 3817 | |
367456c7 | 3818 | return load; |
230059de PZ |
3819 | } |
3820 | #else | |
48a16753 | 3821 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
3822 | { |
3823 | } | |
3824 | ||
367456c7 | 3825 | static inline void update_h_load(long cpu) |
230059de | 3826 | { |
230059de | 3827 | } |
230059de | 3828 | |
367456c7 | 3829 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 3830 | { |
367456c7 | 3831 | return p->se.load.weight; |
1e3c88bd | 3832 | } |
230059de | 3833 | #endif |
1e3c88bd | 3834 | |
1e3c88bd PZ |
3835 | /********** Helpers for find_busiest_group ************************/ |
3836 | /* | |
3837 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
3838 | * during load balancing. | |
3839 | */ | |
3840 | struct sd_lb_stats { | |
3841 | struct sched_group *busiest; /* Busiest group in this sd */ | |
3842 | struct sched_group *this; /* Local group in this sd */ | |
3843 | unsigned long total_load; /* Total load of all groups in sd */ | |
3844 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
3845 | unsigned long avg_load; /* Average load across all groups in sd */ | |
3846 | ||
3847 | /** Statistics of this group */ | |
3848 | unsigned long this_load; | |
3849 | unsigned long this_load_per_task; | |
3850 | unsigned long this_nr_running; | |
fab47622 | 3851 | unsigned long this_has_capacity; |
aae6d3dd | 3852 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
3853 | |
3854 | /* Statistics of the busiest group */ | |
aae6d3dd | 3855 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
3856 | unsigned long max_load; |
3857 | unsigned long busiest_load_per_task; | |
3858 | unsigned long busiest_nr_running; | |
dd5feea1 | 3859 | unsigned long busiest_group_capacity; |
fab47622 | 3860 | unsigned long busiest_has_capacity; |
aae6d3dd | 3861 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
3862 | |
3863 | int group_imb; /* Is there imbalance in this sd */ | |
1e3c88bd PZ |
3864 | }; |
3865 | ||
3866 | /* | |
3867 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
3868 | */ | |
3869 | struct sg_lb_stats { | |
3870 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
3871 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
3872 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
3873 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
3874 | unsigned long group_capacity; | |
aae6d3dd SS |
3875 | unsigned long idle_cpus; |
3876 | unsigned long group_weight; | |
1e3c88bd | 3877 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 3878 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
3879 | }; |
3880 | ||
1e3c88bd PZ |
3881 | /** |
3882 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
3883 | * @sd: The sched_domain whose load_idx is to be obtained. | |
3884 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
3885 | */ | |
3886 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
3887 | enum cpu_idle_type idle) | |
3888 | { | |
3889 | int load_idx; | |
3890 | ||
3891 | switch (idle) { | |
3892 | case CPU_NOT_IDLE: | |
3893 | load_idx = sd->busy_idx; | |
3894 | break; | |
3895 | ||
3896 | case CPU_NEWLY_IDLE: | |
3897 | load_idx = sd->newidle_idx; | |
3898 | break; | |
3899 | default: | |
3900 | load_idx = sd->idle_idx; | |
3901 | break; | |
3902 | } | |
3903 | ||
3904 | return load_idx; | |
3905 | } | |
3906 | ||
1e3c88bd PZ |
3907 | unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
3908 | { | |
1399fa78 | 3909 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
3910 | } |
3911 | ||
3912 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
3913 | { | |
3914 | return default_scale_freq_power(sd, cpu); | |
3915 | } | |
3916 | ||
3917 | unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) | |
3918 | { | |
669c55e9 | 3919 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
3920 | unsigned long smt_gain = sd->smt_gain; |
3921 | ||
3922 | smt_gain /= weight; | |
3923 | ||
3924 | return smt_gain; | |
3925 | } | |
3926 | ||
3927 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
3928 | { | |
3929 | return default_scale_smt_power(sd, cpu); | |
3930 | } | |
3931 | ||
3932 | unsigned long scale_rt_power(int cpu) | |
3933 | { | |
3934 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 3935 | u64 total, available, age_stamp, avg; |
1e3c88bd | 3936 | |
b654f7de PZ |
3937 | /* |
3938 | * Since we're reading these variables without serialization make sure | |
3939 | * we read them once before doing sanity checks on them. | |
3940 | */ | |
3941 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
3942 | avg = ACCESS_ONCE(rq->rt_avg); | |
3943 | ||
3944 | total = sched_avg_period() + (rq->clock - age_stamp); | |
aa483808 | 3945 | |
b654f7de | 3946 | if (unlikely(total < avg)) { |
aa483808 VP |
3947 | /* Ensures that power won't end up being negative */ |
3948 | available = 0; | |
3949 | } else { | |
b654f7de | 3950 | available = total - avg; |
aa483808 | 3951 | } |
1e3c88bd | 3952 | |
1399fa78 NR |
3953 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
3954 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 3955 | |
1399fa78 | 3956 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3957 | |
3958 | return div_u64(available, total); | |
3959 | } | |
3960 | ||
3961 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
3962 | { | |
669c55e9 | 3963 | unsigned long weight = sd->span_weight; |
1399fa78 | 3964 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
3965 | struct sched_group *sdg = sd->groups; |
3966 | ||
1e3c88bd PZ |
3967 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
3968 | if (sched_feat(ARCH_POWER)) | |
3969 | power *= arch_scale_smt_power(sd, cpu); | |
3970 | else | |
3971 | power *= default_scale_smt_power(sd, cpu); | |
3972 | ||
1399fa78 | 3973 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3974 | } |
3975 | ||
9c3f75cb | 3976 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
3977 | |
3978 | if (sched_feat(ARCH_POWER)) | |
3979 | power *= arch_scale_freq_power(sd, cpu); | |
3980 | else | |
3981 | power *= default_scale_freq_power(sd, cpu); | |
3982 | ||
1399fa78 | 3983 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 3984 | |
1e3c88bd | 3985 | power *= scale_rt_power(cpu); |
1399fa78 | 3986 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
3987 | |
3988 | if (!power) | |
3989 | power = 1; | |
3990 | ||
e51fd5e2 | 3991 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 3992 | sdg->sgp->power = power; |
1e3c88bd PZ |
3993 | } |
3994 | ||
029632fb | 3995 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
3996 | { |
3997 | struct sched_domain *child = sd->child; | |
3998 | struct sched_group *group, *sdg = sd->groups; | |
3999 | unsigned long power; | |
4ec4412e VG |
4000 | unsigned long interval; |
4001 | ||
4002 | interval = msecs_to_jiffies(sd->balance_interval); | |
4003 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
4004 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
4005 | |
4006 | if (!child) { | |
4007 | update_cpu_power(sd, cpu); | |
4008 | return; | |
4009 | } | |
4010 | ||
4011 | power = 0; | |
4012 | ||
74a5ce20 PZ |
4013 | if (child->flags & SD_OVERLAP) { |
4014 | /* | |
4015 | * SD_OVERLAP domains cannot assume that child groups | |
4016 | * span the current group. | |
4017 | */ | |
4018 | ||
4019 | for_each_cpu(cpu, sched_group_cpus(sdg)) | |
4020 | power += power_of(cpu); | |
4021 | } else { | |
4022 | /* | |
4023 | * !SD_OVERLAP domains can assume that child groups | |
4024 | * span the current group. | |
4025 | */ | |
4026 | ||
4027 | group = child->groups; | |
4028 | do { | |
4029 | power += group->sgp->power; | |
4030 | group = group->next; | |
4031 | } while (group != child->groups); | |
4032 | } | |
1e3c88bd | 4033 | |
c3decf0d | 4034 | sdg->sgp->power_orig = sdg->sgp->power = power; |
1e3c88bd PZ |
4035 | } |
4036 | ||
9d5efe05 SV |
4037 | /* |
4038 | * Try and fix up capacity for tiny siblings, this is needed when | |
4039 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
4040 | * which on its own isn't powerful enough. | |
4041 | * | |
4042 | * See update_sd_pick_busiest() and check_asym_packing(). | |
4043 | */ | |
4044 | static inline int | |
4045 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
4046 | { | |
4047 | /* | |
1399fa78 | 4048 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 4049 | */ |
a6c75f2f | 4050 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
4051 | return 0; |
4052 | ||
4053 | /* | |
4054 | * If ~90% of the cpu_power is still there, we're good. | |
4055 | */ | |
9c3f75cb | 4056 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
4057 | return 1; |
4058 | ||
4059 | return 0; | |
4060 | } | |
4061 | ||
1e3c88bd PZ |
4062 | /** |
4063 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 4064 | * @env: The load balancing environment. |
1e3c88bd | 4065 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 4066 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 4067 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
4068 | * @balance: Should we balance. |
4069 | * @sgs: variable to hold the statistics for this group. | |
4070 | */ | |
bd939f45 PZ |
4071 | static inline void update_sg_lb_stats(struct lb_env *env, |
4072 | struct sched_group *group, int load_idx, | |
b9403130 | 4073 | int local_group, int *balance, struct sg_lb_stats *sgs) |
1e3c88bd | 4074 | { |
e44bc5c5 PZ |
4075 | unsigned long nr_running, max_nr_running, min_nr_running; |
4076 | unsigned long load, max_cpu_load, min_cpu_load; | |
04f733b4 | 4077 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
dd5feea1 | 4078 | unsigned long avg_load_per_task = 0; |
bd939f45 | 4079 | int i; |
1e3c88bd | 4080 | |
871e35bc | 4081 | if (local_group) |
c1174876 | 4082 | balance_cpu = group_balance_cpu(group); |
1e3c88bd PZ |
4083 | |
4084 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
4085 | max_cpu_load = 0; |
4086 | min_cpu_load = ~0UL; | |
2582f0eb | 4087 | max_nr_running = 0; |
e44bc5c5 | 4088 | min_nr_running = ~0UL; |
1e3c88bd | 4089 | |
b9403130 | 4090 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
4091 | struct rq *rq = cpu_rq(i); |
4092 | ||
e44bc5c5 PZ |
4093 | nr_running = rq->nr_running; |
4094 | ||
1e3c88bd PZ |
4095 | /* Bias balancing toward cpus of our domain */ |
4096 | if (local_group) { | |
c1174876 PZ |
4097 | if (idle_cpu(i) && !first_idle_cpu && |
4098 | cpumask_test_cpu(i, sched_group_mask(group))) { | |
04f733b4 | 4099 | first_idle_cpu = 1; |
1e3c88bd PZ |
4100 | balance_cpu = i; |
4101 | } | |
04f733b4 PZ |
4102 | |
4103 | load = target_load(i, load_idx); | |
1e3c88bd PZ |
4104 | } else { |
4105 | load = source_load(i, load_idx); | |
e44bc5c5 | 4106 | if (load > max_cpu_load) |
1e3c88bd PZ |
4107 | max_cpu_load = load; |
4108 | if (min_cpu_load > load) | |
4109 | min_cpu_load = load; | |
e44bc5c5 PZ |
4110 | |
4111 | if (nr_running > max_nr_running) | |
4112 | max_nr_running = nr_running; | |
4113 | if (min_nr_running > nr_running) | |
4114 | min_nr_running = nr_running; | |
1e3c88bd PZ |
4115 | } |
4116 | ||
4117 | sgs->group_load += load; | |
e44bc5c5 | 4118 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 4119 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
4120 | if (idle_cpu(i)) |
4121 | sgs->idle_cpus++; | |
1e3c88bd PZ |
4122 | } |
4123 | ||
4124 | /* | |
4125 | * First idle cpu or the first cpu(busiest) in this sched group | |
4126 | * is eligible for doing load balancing at this and above | |
4127 | * domains. In the newly idle case, we will allow all the cpu's | |
4128 | * to do the newly idle load balance. | |
4129 | */ | |
4ec4412e | 4130 | if (local_group) { |
bd939f45 | 4131 | if (env->idle != CPU_NEWLY_IDLE) { |
04f733b4 | 4132 | if (balance_cpu != env->dst_cpu) { |
4ec4412e VG |
4133 | *balance = 0; |
4134 | return; | |
4135 | } | |
bd939f45 | 4136 | update_group_power(env->sd, env->dst_cpu); |
4ec4412e | 4137 | } else if (time_after_eq(jiffies, group->sgp->next_update)) |
bd939f45 | 4138 | update_group_power(env->sd, env->dst_cpu); |
1e3c88bd PZ |
4139 | } |
4140 | ||
4141 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 4142 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 4143 | |
1e3c88bd PZ |
4144 | /* |
4145 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 4146 | * than the average weight of a task. |
1e3c88bd PZ |
4147 | * |
4148 | * APZ: with cgroup the avg task weight can vary wildly and | |
4149 | * might not be a suitable number - should we keep a | |
4150 | * normalized nr_running number somewhere that negates | |
4151 | * the hierarchy? | |
4152 | */ | |
dd5feea1 SS |
4153 | if (sgs->sum_nr_running) |
4154 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 4155 | |
e44bc5c5 PZ |
4156 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && |
4157 | (max_nr_running - min_nr_running) > 1) | |
1e3c88bd PZ |
4158 | sgs->group_imb = 1; |
4159 | ||
9c3f75cb | 4160 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 4161 | SCHED_POWER_SCALE); |
9d5efe05 | 4162 | if (!sgs->group_capacity) |
bd939f45 | 4163 | sgs->group_capacity = fix_small_capacity(env->sd, group); |
aae6d3dd | 4164 | sgs->group_weight = group->group_weight; |
fab47622 NR |
4165 | |
4166 | if (sgs->group_capacity > sgs->sum_nr_running) | |
4167 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
4168 | } |
4169 | ||
532cb4c4 MN |
4170 | /** |
4171 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 4172 | * @env: The load balancing environment. |
532cb4c4 MN |
4173 | * @sds: sched_domain statistics |
4174 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 4175 | * @sgs: sched_group statistics |
532cb4c4 MN |
4176 | * |
4177 | * Determine if @sg is a busier group than the previously selected | |
4178 | * busiest group. | |
4179 | */ | |
bd939f45 | 4180 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
4181 | struct sd_lb_stats *sds, |
4182 | struct sched_group *sg, | |
bd939f45 | 4183 | struct sg_lb_stats *sgs) |
532cb4c4 MN |
4184 | { |
4185 | if (sgs->avg_load <= sds->max_load) | |
4186 | return false; | |
4187 | ||
4188 | if (sgs->sum_nr_running > sgs->group_capacity) | |
4189 | return true; | |
4190 | ||
4191 | if (sgs->group_imb) | |
4192 | return true; | |
4193 | ||
4194 | /* | |
4195 | * ASYM_PACKING needs to move all the work to the lowest | |
4196 | * numbered CPUs in the group, therefore mark all groups | |
4197 | * higher than ourself as busy. | |
4198 | */ | |
bd939f45 PZ |
4199 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
4200 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
4201 | if (!sds->busiest) |
4202 | return true; | |
4203 | ||
4204 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
4205 | return true; | |
4206 | } | |
4207 | ||
4208 | return false; | |
4209 | } | |
4210 | ||
1e3c88bd | 4211 | /** |
461819ac | 4212 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 4213 | * @env: The load balancing environment. |
1e3c88bd PZ |
4214 | * @balance: Should we balance. |
4215 | * @sds: variable to hold the statistics for this sched_domain. | |
4216 | */ | |
bd939f45 | 4217 | static inline void update_sd_lb_stats(struct lb_env *env, |
b9403130 | 4218 | int *balance, struct sd_lb_stats *sds) |
1e3c88bd | 4219 | { |
bd939f45 PZ |
4220 | struct sched_domain *child = env->sd->child; |
4221 | struct sched_group *sg = env->sd->groups; | |
1e3c88bd PZ |
4222 | struct sg_lb_stats sgs; |
4223 | int load_idx, prefer_sibling = 0; | |
4224 | ||
4225 | if (child && child->flags & SD_PREFER_SIBLING) | |
4226 | prefer_sibling = 1; | |
4227 | ||
bd939f45 | 4228 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
4229 | |
4230 | do { | |
4231 | int local_group; | |
4232 | ||
bd939f45 | 4233 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
1e3c88bd | 4234 | memset(&sgs, 0, sizeof(sgs)); |
b9403130 | 4235 | update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs); |
1e3c88bd | 4236 | |
8f190fb3 | 4237 | if (local_group && !(*balance)) |
1e3c88bd PZ |
4238 | return; |
4239 | ||
4240 | sds->total_load += sgs.group_load; | |
9c3f75cb | 4241 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
4242 | |
4243 | /* | |
4244 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 4245 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
4246 | * and move all the excess tasks away. We lower the capacity |
4247 | * of a group only if the local group has the capacity to fit | |
4248 | * these excess tasks, i.e. nr_running < group_capacity. The | |
4249 | * extra check prevents the case where you always pull from the | |
4250 | * heaviest group when it is already under-utilized (possible | |
4251 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 4252 | */ |
75dd321d | 4253 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
4254 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
4255 | ||
4256 | if (local_group) { | |
4257 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 4258 | sds->this = sg; |
1e3c88bd PZ |
4259 | sds->this_nr_running = sgs.sum_nr_running; |
4260 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 4261 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4262 | sds->this_idle_cpus = sgs.idle_cpus; |
bd939f45 | 4263 | } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) { |
1e3c88bd | 4264 | sds->max_load = sgs.avg_load; |
532cb4c4 | 4265 | sds->busiest = sg; |
1e3c88bd | 4266 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 4267 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 4268 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 4269 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 4270 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4271 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
4272 | sds->group_imb = sgs.group_imb; |
4273 | } | |
4274 | ||
532cb4c4 | 4275 | sg = sg->next; |
bd939f45 | 4276 | } while (sg != env->sd->groups); |
532cb4c4 MN |
4277 | } |
4278 | ||
532cb4c4 MN |
4279 | /** |
4280 | * check_asym_packing - Check to see if the group is packed into the | |
4281 | * sched doman. | |
4282 | * | |
4283 | * This is primarily intended to used at the sibling level. Some | |
4284 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
4285 | * case of POWER7, it can move to lower SMT modes only when higher | |
4286 | * threads are idle. When in lower SMT modes, the threads will | |
4287 | * perform better since they share less core resources. Hence when we | |
4288 | * have idle threads, we want them to be the higher ones. | |
4289 | * | |
4290 | * This packing function is run on idle threads. It checks to see if | |
4291 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
4292 | * CPU number than the packing function is being run on. Here we are | |
4293 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
4294 | * number. | |
4295 | * | |
b6b12294 MN |
4296 | * Returns 1 when packing is required and a task should be moved to |
4297 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
4298 | * | |
cd96891d | 4299 | * @env: The load balancing environment. |
532cb4c4 | 4300 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 4301 | */ |
bd939f45 | 4302 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
4303 | { |
4304 | int busiest_cpu; | |
4305 | ||
bd939f45 | 4306 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
4307 | return 0; |
4308 | ||
4309 | if (!sds->busiest) | |
4310 | return 0; | |
4311 | ||
4312 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 4313 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
4314 | return 0; |
4315 | ||
bd939f45 PZ |
4316 | env->imbalance = DIV_ROUND_CLOSEST( |
4317 | sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE); | |
4318 | ||
532cb4c4 | 4319 | return 1; |
1e3c88bd PZ |
4320 | } |
4321 | ||
4322 | /** | |
4323 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
4324 | * amongst the groups of a sched_domain, during | |
4325 | * load balancing. | |
cd96891d | 4326 | * @env: The load balancing environment. |
1e3c88bd | 4327 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4328 | */ |
bd939f45 PZ |
4329 | static inline |
4330 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
4331 | { |
4332 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
4333 | unsigned int imbn = 2; | |
dd5feea1 | 4334 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
4335 | |
4336 | if (sds->this_nr_running) { | |
4337 | sds->this_load_per_task /= sds->this_nr_running; | |
4338 | if (sds->busiest_load_per_task > | |
4339 | sds->this_load_per_task) | |
4340 | imbn = 1; | |
bd939f45 | 4341 | } else { |
1e3c88bd | 4342 | sds->this_load_per_task = |
bd939f45 PZ |
4343 | cpu_avg_load_per_task(env->dst_cpu); |
4344 | } | |
1e3c88bd | 4345 | |
dd5feea1 | 4346 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 4347 | * SCHED_POWER_SCALE; |
9c3f75cb | 4348 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
4349 | |
4350 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
4351 | (scaled_busy_load_per_task * imbn)) { | |
bd939f45 | 4352 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4353 | return; |
4354 | } | |
4355 | ||
4356 | /* | |
4357 | * OK, we don't have enough imbalance to justify moving tasks, | |
4358 | * however we may be able to increase total CPU power used by | |
4359 | * moving them. | |
4360 | */ | |
4361 | ||
9c3f75cb | 4362 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 4363 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 4364 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 4365 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 4366 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4367 | |
4368 | /* Amount of load we'd subtract */ | |
1399fa78 | 4369 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 4370 | sds->busiest->sgp->power; |
1e3c88bd | 4371 | if (sds->max_load > tmp) |
9c3f75cb | 4372 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
4373 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
4374 | ||
4375 | /* Amount of load we'd add */ | |
9c3f75cb | 4376 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 4377 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
4378 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
4379 | sds->this->sgp->power; | |
1e3c88bd | 4380 | else |
1399fa78 | 4381 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
4382 | sds->this->sgp->power; |
4383 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 4384 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 4385 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4386 | |
4387 | /* Move if we gain throughput */ | |
4388 | if (pwr_move > pwr_now) | |
bd939f45 | 4389 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4390 | } |
4391 | ||
4392 | /** | |
4393 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
4394 | * groups of a given sched_domain during load balance. | |
bd939f45 | 4395 | * @env: load balance environment |
1e3c88bd | 4396 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4397 | */ |
bd939f45 | 4398 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 4399 | { |
dd5feea1 SS |
4400 | unsigned long max_pull, load_above_capacity = ~0UL; |
4401 | ||
4402 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
4403 | if (sds->group_imb) { | |
4404 | sds->busiest_load_per_task = | |
4405 | min(sds->busiest_load_per_task, sds->avg_load); | |
4406 | } | |
4407 | ||
1e3c88bd PZ |
4408 | /* |
4409 | * In the presence of smp nice balancing, certain scenarios can have | |
4410 | * max load less than avg load(as we skip the groups at or below | |
4411 | * its cpu_power, while calculating max_load..) | |
4412 | */ | |
4413 | if (sds->max_load < sds->avg_load) { | |
bd939f45 PZ |
4414 | env->imbalance = 0; |
4415 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4416 | } |
4417 | ||
dd5feea1 SS |
4418 | if (!sds->group_imb) { |
4419 | /* | |
4420 | * Don't want to pull so many tasks that a group would go idle. | |
4421 | */ | |
4422 | load_above_capacity = (sds->busiest_nr_running - | |
4423 | sds->busiest_group_capacity); | |
4424 | ||
1399fa78 | 4425 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 4426 | |
9c3f75cb | 4427 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
4428 | } |
4429 | ||
4430 | /* | |
4431 | * We're trying to get all the cpus to the average_load, so we don't | |
4432 | * want to push ourselves above the average load, nor do we wish to | |
4433 | * reduce the max loaded cpu below the average load. At the same time, | |
4434 | * we also don't want to reduce the group load below the group capacity | |
4435 | * (so that we can implement power-savings policies etc). Thus we look | |
4436 | * for the minimum possible imbalance. | |
4437 | * Be careful of negative numbers as they'll appear as very large values | |
4438 | * with unsigned longs. | |
4439 | */ | |
4440 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
4441 | |
4442 | /* How much load to actually move to equalise the imbalance */ | |
bd939f45 | 4443 | env->imbalance = min(max_pull * sds->busiest->sgp->power, |
9c3f75cb | 4444 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) |
1399fa78 | 4445 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
4446 | |
4447 | /* | |
4448 | * if *imbalance is less than the average load per runnable task | |
25985edc | 4449 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
4450 | * a think about bumping its value to force at least one task to be |
4451 | * moved | |
4452 | */ | |
bd939f45 PZ |
4453 | if (env->imbalance < sds->busiest_load_per_task) |
4454 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4455 | |
4456 | } | |
fab47622 | 4457 | |
1e3c88bd PZ |
4458 | /******* find_busiest_group() helpers end here *********************/ |
4459 | ||
4460 | /** | |
4461 | * find_busiest_group - Returns the busiest group within the sched_domain | |
4462 | * if there is an imbalance. If there isn't an imbalance, and | |
4463 | * the user has opted for power-savings, it returns a group whose | |
4464 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
4465 | * such a group exists. | |
4466 | * | |
4467 | * Also calculates the amount of weighted load which should be moved | |
4468 | * to restore balance. | |
4469 | * | |
cd96891d | 4470 | * @env: The load balancing environment. |
1e3c88bd PZ |
4471 | * @balance: Pointer to a variable indicating if this_cpu |
4472 | * is the appropriate cpu to perform load balancing at this_level. | |
4473 | * | |
4474 | * Returns: - the busiest group if imbalance exists. | |
4475 | * - If no imbalance and user has opted for power-savings balance, | |
4476 | * return the least loaded group whose CPUs can be | |
4477 | * put to idle by rebalancing its tasks onto our group. | |
4478 | */ | |
4479 | static struct sched_group * | |
b9403130 | 4480 | find_busiest_group(struct lb_env *env, int *balance) |
1e3c88bd PZ |
4481 | { |
4482 | struct sd_lb_stats sds; | |
4483 | ||
4484 | memset(&sds, 0, sizeof(sds)); | |
4485 | ||
4486 | /* | |
4487 | * Compute the various statistics relavent for load balancing at | |
4488 | * this level. | |
4489 | */ | |
b9403130 | 4490 | update_sd_lb_stats(env, balance, &sds); |
1e3c88bd | 4491 | |
cc57aa8f PZ |
4492 | /* |
4493 | * this_cpu is not the appropriate cpu to perform load balancing at | |
4494 | * this level. | |
1e3c88bd | 4495 | */ |
8f190fb3 | 4496 | if (!(*balance)) |
1e3c88bd PZ |
4497 | goto ret; |
4498 | ||
bd939f45 PZ |
4499 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
4500 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
4501 | return sds.busiest; |
4502 | ||
cc57aa8f | 4503 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
4504 | if (!sds.busiest || sds.busiest_nr_running == 0) |
4505 | goto out_balanced; | |
4506 | ||
1399fa78 | 4507 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 4508 | |
866ab43e PZ |
4509 | /* |
4510 | * If the busiest group is imbalanced the below checks don't | |
4511 | * work because they assumes all things are equal, which typically | |
4512 | * isn't true due to cpus_allowed constraints and the like. | |
4513 | */ | |
4514 | if (sds.group_imb) | |
4515 | goto force_balance; | |
4516 | ||
cc57aa8f | 4517 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
bd939f45 | 4518 | if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
fab47622 NR |
4519 | !sds.busiest_has_capacity) |
4520 | goto force_balance; | |
4521 | ||
cc57aa8f PZ |
4522 | /* |
4523 | * If the local group is more busy than the selected busiest group | |
4524 | * don't try and pull any tasks. | |
4525 | */ | |
1e3c88bd PZ |
4526 | if (sds.this_load >= sds.max_load) |
4527 | goto out_balanced; | |
4528 | ||
cc57aa8f PZ |
4529 | /* |
4530 | * Don't pull any tasks if this group is already above the domain | |
4531 | * average load. | |
4532 | */ | |
1e3c88bd PZ |
4533 | if (sds.this_load >= sds.avg_load) |
4534 | goto out_balanced; | |
4535 | ||
bd939f45 | 4536 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
4537 | /* |
4538 | * This cpu is idle. If the busiest group load doesn't | |
4539 | * have more tasks than the number of available cpu's and | |
4540 | * there is no imbalance between this and busiest group | |
4541 | * wrt to idle cpu's, it is balanced. | |
4542 | */ | |
c186fafe | 4543 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
4544 | sds.busiest_nr_running <= sds.busiest_group_weight) |
4545 | goto out_balanced; | |
c186fafe PZ |
4546 | } else { |
4547 | /* | |
4548 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
4549 | * imbalance_pct to be conservative. | |
4550 | */ | |
bd939f45 | 4551 | if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load) |
c186fafe | 4552 | goto out_balanced; |
aae6d3dd | 4553 | } |
1e3c88bd | 4554 | |
fab47622 | 4555 | force_balance: |
1e3c88bd | 4556 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 4557 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
4558 | return sds.busiest; |
4559 | ||
4560 | out_balanced: | |
1e3c88bd | 4561 | ret: |
bd939f45 | 4562 | env->imbalance = 0; |
1e3c88bd PZ |
4563 | return NULL; |
4564 | } | |
4565 | ||
4566 | /* | |
4567 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
4568 | */ | |
bd939f45 | 4569 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 4570 | struct sched_group *group) |
1e3c88bd PZ |
4571 | { |
4572 | struct rq *busiest = NULL, *rq; | |
4573 | unsigned long max_load = 0; | |
4574 | int i; | |
4575 | ||
4576 | for_each_cpu(i, sched_group_cpus(group)) { | |
4577 | unsigned long power = power_of(i); | |
1399fa78 NR |
4578 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
4579 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
4580 | unsigned long wl; |
4581 | ||
9d5efe05 | 4582 | if (!capacity) |
bd939f45 | 4583 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 4584 | |
b9403130 | 4585 | if (!cpumask_test_cpu(i, env->cpus)) |
1e3c88bd PZ |
4586 | continue; |
4587 | ||
4588 | rq = cpu_rq(i); | |
6e40f5bb | 4589 | wl = weighted_cpuload(i); |
1e3c88bd | 4590 | |
6e40f5bb TG |
4591 | /* |
4592 | * When comparing with imbalance, use weighted_cpuload() | |
4593 | * which is not scaled with the cpu power. | |
4594 | */ | |
bd939f45 | 4595 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
4596 | continue; |
4597 | ||
6e40f5bb TG |
4598 | /* |
4599 | * For the load comparisons with the other cpu's, consider | |
4600 | * the weighted_cpuload() scaled with the cpu power, so that | |
4601 | * the load can be moved away from the cpu that is potentially | |
4602 | * running at a lower capacity. | |
4603 | */ | |
1399fa78 | 4604 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 4605 | |
1e3c88bd PZ |
4606 | if (wl > max_load) { |
4607 | max_load = wl; | |
4608 | busiest = rq; | |
4609 | } | |
4610 | } | |
4611 | ||
4612 | return busiest; | |
4613 | } | |
4614 | ||
4615 | /* | |
4616 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
4617 | * so long as it is large enough. | |
4618 | */ | |
4619 | #define MAX_PINNED_INTERVAL 512 | |
4620 | ||
4621 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
029632fb | 4622 | DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); |
1e3c88bd | 4623 | |
bd939f45 | 4624 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 4625 | { |
bd939f45 PZ |
4626 | struct sched_domain *sd = env->sd; |
4627 | ||
4628 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
4629 | |
4630 | /* | |
4631 | * ASYM_PACKING needs to force migrate tasks from busy but | |
4632 | * higher numbered CPUs in order to pack all tasks in the | |
4633 | * lowest numbered CPUs. | |
4634 | */ | |
bd939f45 | 4635 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 4636 | return 1; |
1af3ed3d PZ |
4637 | } |
4638 | ||
4639 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
4640 | } | |
4641 | ||
969c7921 TH |
4642 | static int active_load_balance_cpu_stop(void *data); |
4643 | ||
1e3c88bd PZ |
4644 | /* |
4645 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
4646 | * tasks if there is an imbalance. | |
4647 | */ | |
4648 | static int load_balance(int this_cpu, struct rq *this_rq, | |
4649 | struct sched_domain *sd, enum cpu_idle_type idle, | |
4650 | int *balance) | |
4651 | { | |
88b8dac0 SV |
4652 | int ld_moved, cur_ld_moved, active_balance = 0; |
4653 | int lb_iterations, max_lb_iterations; | |
1e3c88bd | 4654 | struct sched_group *group; |
1e3c88bd PZ |
4655 | struct rq *busiest; |
4656 | unsigned long flags; | |
4657 | struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); | |
4658 | ||
8e45cb54 PZ |
4659 | struct lb_env env = { |
4660 | .sd = sd, | |
ddcdf6e7 PZ |
4661 | .dst_cpu = this_cpu, |
4662 | .dst_rq = this_rq, | |
88b8dac0 | 4663 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 4664 | .idle = idle, |
eb95308e | 4665 | .loop_break = sched_nr_migrate_break, |
b9403130 | 4666 | .cpus = cpus, |
8e45cb54 PZ |
4667 | }; |
4668 | ||
1e3c88bd | 4669 | cpumask_copy(cpus, cpu_active_mask); |
88b8dac0 | 4670 | max_lb_iterations = cpumask_weight(env.dst_grpmask); |
1e3c88bd | 4671 | |
1e3c88bd PZ |
4672 | schedstat_inc(sd, lb_count[idle]); |
4673 | ||
4674 | redo: | |
b9403130 | 4675 | group = find_busiest_group(&env, balance); |
1e3c88bd PZ |
4676 | |
4677 | if (*balance == 0) | |
4678 | goto out_balanced; | |
4679 | ||
4680 | if (!group) { | |
4681 | schedstat_inc(sd, lb_nobusyg[idle]); | |
4682 | goto out_balanced; | |
4683 | } | |
4684 | ||
b9403130 | 4685 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
4686 | if (!busiest) { |
4687 | schedstat_inc(sd, lb_nobusyq[idle]); | |
4688 | goto out_balanced; | |
4689 | } | |
4690 | ||
78feefc5 | 4691 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 4692 | |
bd939f45 | 4693 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
4694 | |
4695 | ld_moved = 0; | |
88b8dac0 | 4696 | lb_iterations = 1; |
1e3c88bd PZ |
4697 | if (busiest->nr_running > 1) { |
4698 | /* | |
4699 | * Attempt to move tasks. If find_busiest_group has found | |
4700 | * an imbalance but busiest->nr_running <= 1, the group is | |
4701 | * still unbalanced. ld_moved simply stays zero, so it is | |
4702 | * correctly treated as an imbalance. | |
4703 | */ | |
8e45cb54 | 4704 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
4705 | env.src_cpu = busiest->cpu; |
4706 | env.src_rq = busiest; | |
4707 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 4708 | |
a35b6466 | 4709 | update_h_load(env.src_cpu); |
5d6523eb | 4710 | more_balance: |
1e3c88bd | 4711 | local_irq_save(flags); |
78feefc5 | 4712 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
4713 | |
4714 | /* | |
4715 | * cur_ld_moved - load moved in current iteration | |
4716 | * ld_moved - cumulative load moved across iterations | |
4717 | */ | |
4718 | cur_ld_moved = move_tasks(&env); | |
4719 | ld_moved += cur_ld_moved; | |
78feefc5 | 4720 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
4721 | local_irq_restore(flags); |
4722 | ||
5d6523eb PZ |
4723 | if (env.flags & LBF_NEED_BREAK) { |
4724 | env.flags &= ~LBF_NEED_BREAK; | |
4725 | goto more_balance; | |
4726 | } | |
4727 | ||
1e3c88bd PZ |
4728 | /* |
4729 | * some other cpu did the load balance for us. | |
4730 | */ | |
88b8dac0 SV |
4731 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
4732 | resched_cpu(env.dst_cpu); | |
4733 | ||
4734 | /* | |
4735 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
4736 | * us and move them to an alternate dst_cpu in our sched_group | |
4737 | * where they can run. The upper limit on how many times we | |
4738 | * iterate on same src_cpu is dependent on number of cpus in our | |
4739 | * sched_group. | |
4740 | * | |
4741 | * This changes load balance semantics a bit on who can move | |
4742 | * load to a given_cpu. In addition to the given_cpu itself | |
4743 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
4744 | * nohz-idle), we now have balance_cpu in a position to move | |
4745 | * load to given_cpu. In rare situations, this may cause | |
4746 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
4747 | * _independently_ and at _same_ time to move some load to | |
4748 | * given_cpu) causing exceess load to be moved to given_cpu. | |
4749 | * This however should not happen so much in practice and | |
4750 | * moreover subsequent load balance cycles should correct the | |
4751 | * excess load moved. | |
4752 | */ | |
4753 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 && | |
4754 | lb_iterations++ < max_lb_iterations) { | |
4755 | ||
78feefc5 | 4756 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 SV |
4757 | env.dst_cpu = env.new_dst_cpu; |
4758 | env.flags &= ~LBF_SOME_PINNED; | |
4759 | env.loop = 0; | |
4760 | env.loop_break = sched_nr_migrate_break; | |
4761 | /* | |
4762 | * Go back to "more_balance" rather than "redo" since we | |
4763 | * need to continue with same src_cpu. | |
4764 | */ | |
4765 | goto more_balance; | |
4766 | } | |
1e3c88bd PZ |
4767 | |
4768 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
8e45cb54 | 4769 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 4770 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
4771 | if (!cpumask_empty(cpus)) { |
4772 | env.loop = 0; | |
4773 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 4774 | goto redo; |
bbf18b19 | 4775 | } |
1e3c88bd PZ |
4776 | goto out_balanced; |
4777 | } | |
4778 | } | |
4779 | ||
4780 | if (!ld_moved) { | |
4781 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
4782 | /* |
4783 | * Increment the failure counter only on periodic balance. | |
4784 | * We do not want newidle balance, which can be very | |
4785 | * frequent, pollute the failure counter causing | |
4786 | * excessive cache_hot migrations and active balances. | |
4787 | */ | |
4788 | if (idle != CPU_NEWLY_IDLE) | |
4789 | sd->nr_balance_failed++; | |
1e3c88bd | 4790 | |
bd939f45 | 4791 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
4792 | raw_spin_lock_irqsave(&busiest->lock, flags); |
4793 | ||
969c7921 TH |
4794 | /* don't kick the active_load_balance_cpu_stop, |
4795 | * if the curr task on busiest cpu can't be | |
4796 | * moved to this_cpu | |
1e3c88bd PZ |
4797 | */ |
4798 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 4799 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
4800 | raw_spin_unlock_irqrestore(&busiest->lock, |
4801 | flags); | |
8e45cb54 | 4802 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
4803 | goto out_one_pinned; |
4804 | } | |
4805 | ||
969c7921 TH |
4806 | /* |
4807 | * ->active_balance synchronizes accesses to | |
4808 | * ->active_balance_work. Once set, it's cleared | |
4809 | * only after active load balance is finished. | |
4810 | */ | |
1e3c88bd PZ |
4811 | if (!busiest->active_balance) { |
4812 | busiest->active_balance = 1; | |
4813 | busiest->push_cpu = this_cpu; | |
4814 | active_balance = 1; | |
4815 | } | |
4816 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 4817 | |
bd939f45 | 4818 | if (active_balance) { |
969c7921 TH |
4819 | stop_one_cpu_nowait(cpu_of(busiest), |
4820 | active_load_balance_cpu_stop, busiest, | |
4821 | &busiest->active_balance_work); | |
bd939f45 | 4822 | } |
1e3c88bd PZ |
4823 | |
4824 | /* | |
4825 | * We've kicked active balancing, reset the failure | |
4826 | * counter. | |
4827 | */ | |
4828 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
4829 | } | |
4830 | } else | |
4831 | sd->nr_balance_failed = 0; | |
4832 | ||
4833 | if (likely(!active_balance)) { | |
4834 | /* We were unbalanced, so reset the balancing interval */ | |
4835 | sd->balance_interval = sd->min_interval; | |
4836 | } else { | |
4837 | /* | |
4838 | * If we've begun active balancing, start to back off. This | |
4839 | * case may not be covered by the all_pinned logic if there | |
4840 | * is only 1 task on the busy runqueue (because we don't call | |
4841 | * move_tasks). | |
4842 | */ | |
4843 | if (sd->balance_interval < sd->max_interval) | |
4844 | sd->balance_interval *= 2; | |
4845 | } | |
4846 | ||
1e3c88bd PZ |
4847 | goto out; |
4848 | ||
4849 | out_balanced: | |
4850 | schedstat_inc(sd, lb_balanced[idle]); | |
4851 | ||
4852 | sd->nr_balance_failed = 0; | |
4853 | ||
4854 | out_one_pinned: | |
4855 | /* tune up the balancing interval */ | |
8e45cb54 | 4856 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 4857 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
4858 | (sd->balance_interval < sd->max_interval)) |
4859 | sd->balance_interval *= 2; | |
4860 | ||
46e49b38 | 4861 | ld_moved = 0; |
1e3c88bd | 4862 | out: |
1e3c88bd PZ |
4863 | return ld_moved; |
4864 | } | |
4865 | ||
1e3c88bd PZ |
4866 | /* |
4867 | * idle_balance is called by schedule() if this_cpu is about to become | |
4868 | * idle. Attempts to pull tasks from other CPUs. | |
4869 | */ | |
029632fb | 4870 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
4871 | { |
4872 | struct sched_domain *sd; | |
4873 | int pulled_task = 0; | |
4874 | unsigned long next_balance = jiffies + HZ; | |
4875 | ||
4876 | this_rq->idle_stamp = this_rq->clock; | |
4877 | ||
4878 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
4879 | return; | |
4880 | ||
18bf2805 BS |
4881 | update_rq_runnable_avg(this_rq, 1); |
4882 | ||
f492e12e PZ |
4883 | /* |
4884 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
4885 | */ | |
4886 | raw_spin_unlock(&this_rq->lock); | |
4887 | ||
48a16753 | 4888 | update_blocked_averages(this_cpu); |
dce840a0 | 4889 | rcu_read_lock(); |
1e3c88bd PZ |
4890 | for_each_domain(this_cpu, sd) { |
4891 | unsigned long interval; | |
f492e12e | 4892 | int balance = 1; |
1e3c88bd PZ |
4893 | |
4894 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
4895 | continue; | |
4896 | ||
f492e12e | 4897 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 4898 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
4899 | pulled_task = load_balance(this_cpu, this_rq, |
4900 | sd, CPU_NEWLY_IDLE, &balance); | |
4901 | } | |
1e3c88bd PZ |
4902 | |
4903 | interval = msecs_to_jiffies(sd->balance_interval); | |
4904 | if (time_after(next_balance, sd->last_balance + interval)) | |
4905 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
4906 | if (pulled_task) { |
4907 | this_rq->idle_stamp = 0; | |
1e3c88bd | 4908 | break; |
d5ad140b | 4909 | } |
1e3c88bd | 4910 | } |
dce840a0 | 4911 | rcu_read_unlock(); |
f492e12e PZ |
4912 | |
4913 | raw_spin_lock(&this_rq->lock); | |
4914 | ||
1e3c88bd PZ |
4915 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
4916 | /* | |
4917 | * We are going idle. next_balance may be set based on | |
4918 | * a busy processor. So reset next_balance. | |
4919 | */ | |
4920 | this_rq->next_balance = next_balance; | |
4921 | } | |
4922 | } | |
4923 | ||
4924 | /* | |
969c7921 TH |
4925 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
4926 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
4927 | * least 1 task to be running on each physical CPU where possible, and | |
4928 | * avoids physical / logical imbalances. | |
1e3c88bd | 4929 | */ |
969c7921 | 4930 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 4931 | { |
969c7921 TH |
4932 | struct rq *busiest_rq = data; |
4933 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 4934 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 4935 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 4936 | struct sched_domain *sd; |
969c7921 TH |
4937 | |
4938 | raw_spin_lock_irq(&busiest_rq->lock); | |
4939 | ||
4940 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
4941 | if (unlikely(busiest_cpu != smp_processor_id() || | |
4942 | !busiest_rq->active_balance)) | |
4943 | goto out_unlock; | |
1e3c88bd PZ |
4944 | |
4945 | /* Is there any task to move? */ | |
4946 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 4947 | goto out_unlock; |
1e3c88bd PZ |
4948 | |
4949 | /* | |
4950 | * This condition is "impossible", if it occurs | |
4951 | * we need to fix it. Originally reported by | |
4952 | * Bjorn Helgaas on a 128-cpu setup. | |
4953 | */ | |
4954 | BUG_ON(busiest_rq == target_rq); | |
4955 | ||
4956 | /* move a task from busiest_rq to target_rq */ | |
4957 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
4958 | |
4959 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 4960 | rcu_read_lock(); |
1e3c88bd PZ |
4961 | for_each_domain(target_cpu, sd) { |
4962 | if ((sd->flags & SD_LOAD_BALANCE) && | |
4963 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
4964 | break; | |
4965 | } | |
4966 | ||
4967 | if (likely(sd)) { | |
8e45cb54 PZ |
4968 | struct lb_env env = { |
4969 | .sd = sd, | |
ddcdf6e7 PZ |
4970 | .dst_cpu = target_cpu, |
4971 | .dst_rq = target_rq, | |
4972 | .src_cpu = busiest_rq->cpu, | |
4973 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
4974 | .idle = CPU_IDLE, |
4975 | }; | |
4976 | ||
1e3c88bd PZ |
4977 | schedstat_inc(sd, alb_count); |
4978 | ||
8e45cb54 | 4979 | if (move_one_task(&env)) |
1e3c88bd PZ |
4980 | schedstat_inc(sd, alb_pushed); |
4981 | else | |
4982 | schedstat_inc(sd, alb_failed); | |
4983 | } | |
dce840a0 | 4984 | rcu_read_unlock(); |
1e3c88bd | 4985 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
4986 | out_unlock: |
4987 | busiest_rq->active_balance = 0; | |
4988 | raw_spin_unlock_irq(&busiest_rq->lock); | |
4989 | return 0; | |
1e3c88bd PZ |
4990 | } |
4991 | ||
4992 | #ifdef CONFIG_NO_HZ | |
83cd4fe2 VP |
4993 | /* |
4994 | * idle load balancing details | |
83cd4fe2 VP |
4995 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
4996 | * needed, they will kick the idle load balancer, which then does idle | |
4997 | * load balancing for all the idle CPUs. | |
4998 | */ | |
1e3c88bd | 4999 | static struct { |
83cd4fe2 | 5000 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 5001 | atomic_t nr_cpus; |
83cd4fe2 VP |
5002 | unsigned long next_balance; /* in jiffy units */ |
5003 | } nohz ____cacheline_aligned; | |
1e3c88bd | 5004 | |
8e7fbcbc | 5005 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 5006 | { |
0b005cf5 | 5007 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 5008 | |
786d6dc7 SS |
5009 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
5010 | return ilb; | |
5011 | ||
5012 | return nr_cpu_ids; | |
1e3c88bd | 5013 | } |
1e3c88bd | 5014 | |
83cd4fe2 VP |
5015 | /* |
5016 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
5017 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
5018 | * CPU (if there is one). | |
5019 | */ | |
5020 | static void nohz_balancer_kick(int cpu) | |
5021 | { | |
5022 | int ilb_cpu; | |
5023 | ||
5024 | nohz.next_balance++; | |
5025 | ||
0b005cf5 | 5026 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 5027 | |
0b005cf5 SS |
5028 | if (ilb_cpu >= nr_cpu_ids) |
5029 | return; | |
83cd4fe2 | 5030 | |
cd490c5b | 5031 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
5032 | return; |
5033 | /* | |
5034 | * Use smp_send_reschedule() instead of resched_cpu(). | |
5035 | * This way we generate a sched IPI on the target cpu which | |
5036 | * is idle. And the softirq performing nohz idle load balance | |
5037 | * will be run before returning from the IPI. | |
5038 | */ | |
5039 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
5040 | return; |
5041 | } | |
5042 | ||
c1cc017c | 5043 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
5044 | { |
5045 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
5046 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
5047 | atomic_dec(&nohz.nr_cpus); | |
5048 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
5049 | } | |
5050 | } | |
5051 | ||
69e1e811 SS |
5052 | static inline void set_cpu_sd_state_busy(void) |
5053 | { | |
5054 | struct sched_domain *sd; | |
5055 | int cpu = smp_processor_id(); | |
5056 | ||
5057 | if (!test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
5058 | return; | |
5059 | clear_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
5060 | ||
5061 | rcu_read_lock(); | |
5062 | for_each_domain(cpu, sd) | |
5063 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); | |
5064 | rcu_read_unlock(); | |
5065 | } | |
5066 | ||
5067 | void set_cpu_sd_state_idle(void) | |
5068 | { | |
5069 | struct sched_domain *sd; | |
5070 | int cpu = smp_processor_id(); | |
5071 | ||
5072 | if (test_bit(NOHZ_IDLE, nohz_flags(cpu))) | |
5073 | return; | |
5074 | set_bit(NOHZ_IDLE, nohz_flags(cpu)); | |
5075 | ||
5076 | rcu_read_lock(); | |
5077 | for_each_domain(cpu, sd) | |
5078 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); | |
5079 | rcu_read_unlock(); | |
5080 | } | |
5081 | ||
1e3c88bd | 5082 | /* |
c1cc017c | 5083 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 5084 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 5085 | */ |
c1cc017c | 5086 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 5087 | { |
71325960 SS |
5088 | /* |
5089 | * If this cpu is going down, then nothing needs to be done. | |
5090 | */ | |
5091 | if (!cpu_active(cpu)) | |
5092 | return; | |
5093 | ||
c1cc017c AS |
5094 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
5095 | return; | |
1e3c88bd | 5096 | |
c1cc017c AS |
5097 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
5098 | atomic_inc(&nohz.nr_cpus); | |
5099 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 5100 | } |
71325960 SS |
5101 | |
5102 | static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb, | |
5103 | unsigned long action, void *hcpu) | |
5104 | { | |
5105 | switch (action & ~CPU_TASKS_FROZEN) { | |
5106 | case CPU_DYING: | |
c1cc017c | 5107 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
5108 | return NOTIFY_OK; |
5109 | default: | |
5110 | return NOTIFY_DONE; | |
5111 | } | |
5112 | } | |
1e3c88bd PZ |
5113 | #endif |
5114 | ||
5115 | static DEFINE_SPINLOCK(balancing); | |
5116 | ||
49c022e6 PZ |
5117 | /* |
5118 | * Scale the max load_balance interval with the number of CPUs in the system. | |
5119 | * This trades load-balance latency on larger machines for less cross talk. | |
5120 | */ | |
029632fb | 5121 | void update_max_interval(void) |
49c022e6 PZ |
5122 | { |
5123 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
5124 | } | |
5125 | ||
1e3c88bd PZ |
5126 | /* |
5127 | * It checks each scheduling domain to see if it is due to be balanced, | |
5128 | * and initiates a balancing operation if so. | |
5129 | * | |
5130 | * Balancing parameters are set up in arch_init_sched_domains. | |
5131 | */ | |
5132 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
5133 | { | |
5134 | int balance = 1; | |
5135 | struct rq *rq = cpu_rq(cpu); | |
5136 | unsigned long interval; | |
04f733b4 | 5137 | struct sched_domain *sd; |
1e3c88bd PZ |
5138 | /* Earliest time when we have to do rebalance again */ |
5139 | unsigned long next_balance = jiffies + 60*HZ; | |
5140 | int update_next_balance = 0; | |
5141 | int need_serialize; | |
5142 | ||
48a16753 | 5143 | update_blocked_averages(cpu); |
2069dd75 | 5144 | |
dce840a0 | 5145 | rcu_read_lock(); |
1e3c88bd PZ |
5146 | for_each_domain(cpu, sd) { |
5147 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
5148 | continue; | |
5149 | ||
5150 | interval = sd->balance_interval; | |
5151 | if (idle != CPU_IDLE) | |
5152 | interval *= sd->busy_factor; | |
5153 | ||
5154 | /* scale ms to jiffies */ | |
5155 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 5156 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
5157 | |
5158 | need_serialize = sd->flags & SD_SERIALIZE; | |
5159 | ||
5160 | if (need_serialize) { | |
5161 | if (!spin_trylock(&balancing)) | |
5162 | goto out; | |
5163 | } | |
5164 | ||
5165 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
5166 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
5167 | /* | |
5168 | * We've pulled tasks over so either we're no | |
c186fafe | 5169 | * longer idle. |
1e3c88bd PZ |
5170 | */ |
5171 | idle = CPU_NOT_IDLE; | |
5172 | } | |
5173 | sd->last_balance = jiffies; | |
5174 | } | |
5175 | if (need_serialize) | |
5176 | spin_unlock(&balancing); | |
5177 | out: | |
5178 | if (time_after(next_balance, sd->last_balance + interval)) { | |
5179 | next_balance = sd->last_balance + interval; | |
5180 | update_next_balance = 1; | |
5181 | } | |
5182 | ||
5183 | /* | |
5184 | * Stop the load balance at this level. There is another | |
5185 | * CPU in our sched group which is doing load balancing more | |
5186 | * actively. | |
5187 | */ | |
5188 | if (!balance) | |
5189 | break; | |
5190 | } | |
dce840a0 | 5191 | rcu_read_unlock(); |
1e3c88bd PZ |
5192 | |
5193 | /* | |
5194 | * next_balance will be updated only when there is a need. | |
5195 | * When the cpu is attached to null domain for ex, it will not be | |
5196 | * updated. | |
5197 | */ | |
5198 | if (likely(update_next_balance)) | |
5199 | rq->next_balance = next_balance; | |
5200 | } | |
5201 | ||
83cd4fe2 | 5202 | #ifdef CONFIG_NO_HZ |
1e3c88bd | 5203 | /* |
83cd4fe2 | 5204 | * In CONFIG_NO_HZ case, the idle balance kickee will do the |
1e3c88bd PZ |
5205 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
5206 | */ | |
83cd4fe2 VP |
5207 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
5208 | { | |
5209 | struct rq *this_rq = cpu_rq(this_cpu); | |
5210 | struct rq *rq; | |
5211 | int balance_cpu; | |
5212 | ||
1c792db7 SS |
5213 | if (idle != CPU_IDLE || |
5214 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
5215 | goto end; | |
83cd4fe2 VP |
5216 | |
5217 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 5218 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
5219 | continue; |
5220 | ||
5221 | /* | |
5222 | * If this cpu gets work to do, stop the load balancing | |
5223 | * work being done for other cpus. Next load | |
5224 | * balancing owner will pick it up. | |
5225 | */ | |
1c792db7 | 5226 | if (need_resched()) |
83cd4fe2 | 5227 | break; |
83cd4fe2 | 5228 | |
5ed4f1d9 VG |
5229 | rq = cpu_rq(balance_cpu); |
5230 | ||
5231 | raw_spin_lock_irq(&rq->lock); | |
5232 | update_rq_clock(rq); | |
5233 | update_idle_cpu_load(rq); | |
5234 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
5235 | |
5236 | rebalance_domains(balance_cpu, CPU_IDLE); | |
5237 | ||
83cd4fe2 VP |
5238 | if (time_after(this_rq->next_balance, rq->next_balance)) |
5239 | this_rq->next_balance = rq->next_balance; | |
5240 | } | |
5241 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
5242 | end: |
5243 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
5244 | } |
5245 | ||
5246 | /* | |
0b005cf5 SS |
5247 | * Current heuristic for kicking the idle load balancer in the presence |
5248 | * of an idle cpu is the system. | |
5249 | * - This rq has more than one task. | |
5250 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
5251 | * busy cpu's exceeding the group's power. | |
5252 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
5253 | * domain span are idle. | |
83cd4fe2 VP |
5254 | */ |
5255 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
5256 | { | |
5257 | unsigned long now = jiffies; | |
0b005cf5 | 5258 | struct sched_domain *sd; |
83cd4fe2 | 5259 | |
1c792db7 | 5260 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
5261 | return 0; |
5262 | ||
1c792db7 SS |
5263 | /* |
5264 | * We may be recently in ticked or tickless idle mode. At the first | |
5265 | * busy tick after returning from idle, we will update the busy stats. | |
5266 | */ | |
69e1e811 | 5267 | set_cpu_sd_state_busy(); |
c1cc017c | 5268 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
5269 | |
5270 | /* | |
5271 | * None are in tickless mode and hence no need for NOHZ idle load | |
5272 | * balancing. | |
5273 | */ | |
5274 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
5275 | return 0; | |
1c792db7 SS |
5276 | |
5277 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
5278 | return 0; |
5279 | ||
0b005cf5 SS |
5280 | if (rq->nr_running >= 2) |
5281 | goto need_kick; | |
83cd4fe2 | 5282 | |
067491b7 | 5283 | rcu_read_lock(); |
0b005cf5 SS |
5284 | for_each_domain(cpu, sd) { |
5285 | struct sched_group *sg = sd->groups; | |
5286 | struct sched_group_power *sgp = sg->sgp; | |
5287 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 5288 | |
0b005cf5 | 5289 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 5290 | goto need_kick_unlock; |
0b005cf5 SS |
5291 | |
5292 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
5293 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
5294 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 5295 | goto need_kick_unlock; |
0b005cf5 SS |
5296 | |
5297 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
5298 | break; | |
83cd4fe2 | 5299 | } |
067491b7 | 5300 | rcu_read_unlock(); |
83cd4fe2 | 5301 | return 0; |
067491b7 PZ |
5302 | |
5303 | need_kick_unlock: | |
5304 | rcu_read_unlock(); | |
0b005cf5 SS |
5305 | need_kick: |
5306 | return 1; | |
83cd4fe2 VP |
5307 | } |
5308 | #else | |
5309 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
5310 | #endif | |
5311 | ||
5312 | /* | |
5313 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
5314 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
5315 | */ | |
1e3c88bd PZ |
5316 | static void run_rebalance_domains(struct softirq_action *h) |
5317 | { | |
5318 | int this_cpu = smp_processor_id(); | |
5319 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 5320 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
5321 | CPU_IDLE : CPU_NOT_IDLE; |
5322 | ||
5323 | rebalance_domains(this_cpu, idle); | |
5324 | ||
1e3c88bd | 5325 | /* |
83cd4fe2 | 5326 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
5327 | * balancing on behalf of the other idle cpus whose ticks are |
5328 | * stopped. | |
5329 | */ | |
83cd4fe2 | 5330 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
5331 | } |
5332 | ||
5333 | static inline int on_null_domain(int cpu) | |
5334 | { | |
90a6501f | 5335 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
5336 | } |
5337 | ||
5338 | /* | |
5339 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 5340 | */ |
029632fb | 5341 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 5342 | { |
1e3c88bd PZ |
5343 | /* Don't need to rebalance while attached to NULL domain */ |
5344 | if (time_after_eq(jiffies, rq->next_balance) && | |
5345 | likely(!on_null_domain(cpu))) | |
5346 | raise_softirq(SCHED_SOFTIRQ); | |
83cd4fe2 | 5347 | #ifdef CONFIG_NO_HZ |
1c792db7 | 5348 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
5349 | nohz_balancer_kick(cpu); |
5350 | #endif | |
1e3c88bd PZ |
5351 | } |
5352 | ||
0bcdcf28 CE |
5353 | static void rq_online_fair(struct rq *rq) |
5354 | { | |
5355 | update_sysctl(); | |
5356 | } | |
5357 | ||
5358 | static void rq_offline_fair(struct rq *rq) | |
5359 | { | |
5360 | update_sysctl(); | |
a4c96ae3 PB |
5361 | |
5362 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
5363 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
5364 | } |
5365 | ||
55e12e5e | 5366 | #endif /* CONFIG_SMP */ |
e1d1484f | 5367 | |
bf0f6f24 IM |
5368 | /* |
5369 | * scheduler tick hitting a task of our scheduling class: | |
5370 | */ | |
8f4d37ec | 5371 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
5372 | { |
5373 | struct cfs_rq *cfs_rq; | |
5374 | struct sched_entity *se = &curr->se; | |
5375 | ||
5376 | for_each_sched_entity(se) { | |
5377 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 5378 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 5379 | } |
18bf2805 BS |
5380 | |
5381 | update_rq_runnable_avg(rq, 1); | |
bf0f6f24 IM |
5382 | } |
5383 | ||
5384 | /* | |
cd29fe6f PZ |
5385 | * called on fork with the child task as argument from the parent's context |
5386 | * - child not yet on the tasklist | |
5387 | * - preemption disabled | |
bf0f6f24 | 5388 | */ |
cd29fe6f | 5389 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 5390 | { |
4fc420c9 DN |
5391 | struct cfs_rq *cfs_rq; |
5392 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 5393 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
5394 | struct rq *rq = this_rq(); |
5395 | unsigned long flags; | |
5396 | ||
05fa785c | 5397 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 5398 | |
861d034e PZ |
5399 | update_rq_clock(rq); |
5400 | ||
4fc420c9 DN |
5401 | cfs_rq = task_cfs_rq(current); |
5402 | curr = cfs_rq->curr; | |
5403 | ||
b0a0f667 PM |
5404 | if (unlikely(task_cpu(p) != this_cpu)) { |
5405 | rcu_read_lock(); | |
cd29fe6f | 5406 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
5407 | rcu_read_unlock(); |
5408 | } | |
bf0f6f24 | 5409 | |
7109c442 | 5410 | update_curr(cfs_rq); |
cd29fe6f | 5411 | |
b5d9d734 MG |
5412 | if (curr) |
5413 | se->vruntime = curr->vruntime; | |
aeb73b04 | 5414 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 5415 | |
cd29fe6f | 5416 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 5417 | /* |
edcb60a3 IM |
5418 | * Upon rescheduling, sched_class::put_prev_task() will place |
5419 | * 'current' within the tree based on its new key value. | |
5420 | */ | |
4d78e7b6 | 5421 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 5422 | resched_task(rq->curr); |
4d78e7b6 | 5423 | } |
bf0f6f24 | 5424 | |
88ec22d3 PZ |
5425 | se->vruntime -= cfs_rq->min_vruntime; |
5426 | ||
05fa785c | 5427 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
5428 | } |
5429 | ||
cb469845 SR |
5430 | /* |
5431 | * Priority of the task has changed. Check to see if we preempt | |
5432 | * the current task. | |
5433 | */ | |
da7a735e PZ |
5434 | static void |
5435 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 5436 | { |
da7a735e PZ |
5437 | if (!p->se.on_rq) |
5438 | return; | |
5439 | ||
cb469845 SR |
5440 | /* |
5441 | * Reschedule if we are currently running on this runqueue and | |
5442 | * our priority decreased, or if we are not currently running on | |
5443 | * this runqueue and our priority is higher than the current's | |
5444 | */ | |
da7a735e | 5445 | if (rq->curr == p) { |
cb469845 SR |
5446 | if (p->prio > oldprio) |
5447 | resched_task(rq->curr); | |
5448 | } else | |
15afe09b | 5449 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5450 | } |
5451 | ||
da7a735e PZ |
5452 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
5453 | { | |
5454 | struct sched_entity *se = &p->se; | |
5455 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5456 | ||
5457 | /* | |
5458 | * Ensure the task's vruntime is normalized, so that when its | |
5459 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
5460 | * do the right thing. | |
5461 | * | |
5462 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
5463 | * have normalized the vruntime, if it was !on_rq, then only when | |
5464 | * the task is sleeping will it still have non-normalized vruntime. | |
5465 | */ | |
5466 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
5467 | /* | |
5468 | * Fix up our vruntime so that the current sleep doesn't | |
5469 | * cause 'unlimited' sleep bonus. | |
5470 | */ | |
5471 | place_entity(cfs_rq, se, 0); | |
5472 | se->vruntime -= cfs_rq->min_vruntime; | |
5473 | } | |
9ee474f5 PT |
5474 | |
5475 | #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP) | |
5476 | /* | |
5477 | * Remove our load from contribution when we leave sched_fair | |
5478 | * and ensure we don't carry in an old decay_count if we | |
5479 | * switch back. | |
5480 | */ | |
5481 | if (p->se.avg.decay_count) { | |
5482 | struct cfs_rq *cfs_rq = cfs_rq_of(&p->se); | |
5483 | __synchronize_entity_decay(&p->se); | |
5484 | subtract_blocked_load_contrib(cfs_rq, | |
5485 | p->se.avg.load_avg_contrib); | |
5486 | } | |
5487 | #endif | |
da7a735e PZ |
5488 | } |
5489 | ||
cb469845 SR |
5490 | /* |
5491 | * We switched to the sched_fair class. | |
5492 | */ | |
da7a735e | 5493 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 5494 | { |
da7a735e PZ |
5495 | if (!p->se.on_rq) |
5496 | return; | |
5497 | ||
cb469845 SR |
5498 | /* |
5499 | * We were most likely switched from sched_rt, so | |
5500 | * kick off the schedule if running, otherwise just see | |
5501 | * if we can still preempt the current task. | |
5502 | */ | |
da7a735e | 5503 | if (rq->curr == p) |
cb469845 SR |
5504 | resched_task(rq->curr); |
5505 | else | |
15afe09b | 5506 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5507 | } |
5508 | ||
83b699ed SV |
5509 | /* Account for a task changing its policy or group. |
5510 | * | |
5511 | * This routine is mostly called to set cfs_rq->curr field when a task | |
5512 | * migrates between groups/classes. | |
5513 | */ | |
5514 | static void set_curr_task_fair(struct rq *rq) | |
5515 | { | |
5516 | struct sched_entity *se = &rq->curr->se; | |
5517 | ||
ec12cb7f PT |
5518 | for_each_sched_entity(se) { |
5519 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5520 | ||
5521 | set_next_entity(cfs_rq, se); | |
5522 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
5523 | account_cfs_rq_runtime(cfs_rq, 0); | |
5524 | } | |
83b699ed SV |
5525 | } |
5526 | ||
029632fb PZ |
5527 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
5528 | { | |
5529 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
5530 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
5531 | #ifndef CONFIG_64BIT | |
5532 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
5533 | #endif | |
9ee474f5 PT |
5534 | #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP) |
5535 | atomic64_set(&cfs_rq->decay_counter, 1); | |
aff3e498 | 5536 | atomic64_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 5537 | #endif |
029632fb PZ |
5538 | } |
5539 | ||
810b3817 | 5540 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5541 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 5542 | { |
aff3e498 | 5543 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
5544 | /* |
5545 | * If the task was not on the rq at the time of this cgroup movement | |
5546 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
5547 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
5548 | * bonus in place_entity()). | |
5549 | * | |
5550 | * If it was on the rq, we've just 'preempted' it, which does convert | |
5551 | * ->vruntime to a relative base. | |
5552 | * | |
5553 | * Make sure both cases convert their relative position when migrating | |
5554 | * to another cgroup's rq. This does somewhat interfere with the | |
5555 | * fair sleeper stuff for the first placement, but who cares. | |
5556 | */ | |
7ceff013 DN |
5557 | /* |
5558 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
5559 | * But there are some cases where it has already been normalized: | |
5560 | * | |
5561 | * - Moving a forked child which is waiting for being woken up by | |
5562 | * wake_up_new_task(). | |
62af3783 DN |
5563 | * - Moving a task which has been woken up by try_to_wake_up() and |
5564 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
5565 | * |
5566 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
5567 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
5568 | */ | |
62af3783 | 5569 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
5570 | on_rq = 1; |
5571 | ||
b2b5ce02 PZ |
5572 | if (!on_rq) |
5573 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
5574 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
5575 | if (!on_rq) { |
5576 | cfs_rq = cfs_rq_of(&p->se); | |
5577 | p->se.vruntime += cfs_rq->min_vruntime; | |
5578 | #ifdef CONFIG_SMP | |
5579 | /* | |
5580 | * migrate_task_rq_fair() will have removed our previous | |
5581 | * contribution, but we must synchronize for ongoing future | |
5582 | * decay. | |
5583 | */ | |
5584 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
5585 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
5586 | #endif | |
5587 | } | |
810b3817 | 5588 | } |
029632fb PZ |
5589 | |
5590 | void free_fair_sched_group(struct task_group *tg) | |
5591 | { | |
5592 | int i; | |
5593 | ||
5594 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5595 | ||
5596 | for_each_possible_cpu(i) { | |
5597 | if (tg->cfs_rq) | |
5598 | kfree(tg->cfs_rq[i]); | |
5599 | if (tg->se) | |
5600 | kfree(tg->se[i]); | |
5601 | } | |
5602 | ||
5603 | kfree(tg->cfs_rq); | |
5604 | kfree(tg->se); | |
5605 | } | |
5606 | ||
5607 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5608 | { | |
5609 | struct cfs_rq *cfs_rq; | |
5610 | struct sched_entity *se; | |
5611 | int i; | |
5612 | ||
5613 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
5614 | if (!tg->cfs_rq) | |
5615 | goto err; | |
5616 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
5617 | if (!tg->se) | |
5618 | goto err; | |
5619 | ||
5620 | tg->shares = NICE_0_LOAD; | |
5621 | ||
5622 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5623 | ||
5624 | for_each_possible_cpu(i) { | |
5625 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
5626 | GFP_KERNEL, cpu_to_node(i)); | |
5627 | if (!cfs_rq) | |
5628 | goto err; | |
5629 | ||
5630 | se = kzalloc_node(sizeof(struct sched_entity), | |
5631 | GFP_KERNEL, cpu_to_node(i)); | |
5632 | if (!se) | |
5633 | goto err_free_rq; | |
5634 | ||
5635 | init_cfs_rq(cfs_rq); | |
5636 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
5637 | } | |
5638 | ||
5639 | return 1; | |
5640 | ||
5641 | err_free_rq: | |
5642 | kfree(cfs_rq); | |
5643 | err: | |
5644 | return 0; | |
5645 | } | |
5646 | ||
5647 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
5648 | { | |
5649 | struct rq *rq = cpu_rq(cpu); | |
5650 | unsigned long flags; | |
5651 | ||
5652 | /* | |
5653 | * Only empty task groups can be destroyed; so we can speculatively | |
5654 | * check on_list without danger of it being re-added. | |
5655 | */ | |
5656 | if (!tg->cfs_rq[cpu]->on_list) | |
5657 | return; | |
5658 | ||
5659 | raw_spin_lock_irqsave(&rq->lock, flags); | |
5660 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
5661 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
5662 | } | |
5663 | ||
5664 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
5665 | struct sched_entity *se, int cpu, | |
5666 | struct sched_entity *parent) | |
5667 | { | |
5668 | struct rq *rq = cpu_rq(cpu); | |
5669 | ||
5670 | cfs_rq->tg = tg; | |
5671 | cfs_rq->rq = rq; | |
029632fb PZ |
5672 | init_cfs_rq_runtime(cfs_rq); |
5673 | ||
5674 | tg->cfs_rq[cpu] = cfs_rq; | |
5675 | tg->se[cpu] = se; | |
5676 | ||
5677 | /* se could be NULL for root_task_group */ | |
5678 | if (!se) | |
5679 | return; | |
5680 | ||
5681 | if (!parent) | |
5682 | se->cfs_rq = &rq->cfs; | |
5683 | else | |
5684 | se->cfs_rq = parent->my_q; | |
5685 | ||
5686 | se->my_q = cfs_rq; | |
5687 | update_load_set(&se->load, 0); | |
5688 | se->parent = parent; | |
5689 | } | |
5690 | ||
5691 | static DEFINE_MUTEX(shares_mutex); | |
5692 | ||
5693 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
5694 | { | |
5695 | int i; | |
5696 | unsigned long flags; | |
5697 | ||
5698 | /* | |
5699 | * We can't change the weight of the root cgroup. | |
5700 | */ | |
5701 | if (!tg->se[0]) | |
5702 | return -EINVAL; | |
5703 | ||
5704 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
5705 | ||
5706 | mutex_lock(&shares_mutex); | |
5707 | if (tg->shares == shares) | |
5708 | goto done; | |
5709 | ||
5710 | tg->shares = shares; | |
5711 | for_each_possible_cpu(i) { | |
5712 | struct rq *rq = cpu_rq(i); | |
5713 | struct sched_entity *se; | |
5714 | ||
5715 | se = tg->se[i]; | |
5716 | /* Propagate contribution to hierarchy */ | |
5717 | raw_spin_lock_irqsave(&rq->lock, flags); | |
f269ae04 | 5718 | for_each_sched_entity(se) { |
029632fb | 5719 | update_cfs_shares(group_cfs_rq(se)); |
f269ae04 PT |
5720 | /* update contribution to parent */ |
5721 | update_entity_load_avg(se, 1); | |
5722 | } | |
029632fb PZ |
5723 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
5724 | } | |
5725 | ||
5726 | done: | |
5727 | mutex_unlock(&shares_mutex); | |
5728 | return 0; | |
5729 | } | |
5730 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
5731 | ||
5732 | void free_fair_sched_group(struct task_group *tg) { } | |
5733 | ||
5734 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
5735 | { | |
5736 | return 1; | |
5737 | } | |
5738 | ||
5739 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
5740 | ||
5741 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
5742 | ||
810b3817 | 5743 | |
6d686f45 | 5744 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
5745 | { |
5746 | struct sched_entity *se = &task->se; | |
0d721cea PW |
5747 | unsigned int rr_interval = 0; |
5748 | ||
5749 | /* | |
5750 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
5751 | * idle runqueue: | |
5752 | */ | |
0d721cea PW |
5753 | if (rq->cfs.load.weight) |
5754 | rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
0d721cea PW |
5755 | |
5756 | return rr_interval; | |
5757 | } | |
5758 | ||
bf0f6f24 IM |
5759 | /* |
5760 | * All the scheduling class methods: | |
5761 | */ | |
029632fb | 5762 | const struct sched_class fair_sched_class = { |
5522d5d5 | 5763 | .next = &idle_sched_class, |
bf0f6f24 IM |
5764 | .enqueue_task = enqueue_task_fair, |
5765 | .dequeue_task = dequeue_task_fair, | |
5766 | .yield_task = yield_task_fair, | |
d95f4122 | 5767 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 5768 | |
2e09bf55 | 5769 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
5770 | |
5771 | .pick_next_task = pick_next_task_fair, | |
5772 | .put_prev_task = put_prev_task_fair, | |
5773 | ||
681f3e68 | 5774 | #ifdef CONFIG_SMP |
4ce72a2c | 5775 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 5776 | .migrate_task_rq = migrate_task_rq_fair, |
4ce72a2c | 5777 | |
0bcdcf28 CE |
5778 | .rq_online = rq_online_fair, |
5779 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
5780 | |
5781 | .task_waking = task_waking_fair, | |
681f3e68 | 5782 | #endif |
bf0f6f24 | 5783 | |
83b699ed | 5784 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 5785 | .task_tick = task_tick_fair, |
cd29fe6f | 5786 | .task_fork = task_fork_fair, |
cb469845 SR |
5787 | |
5788 | .prio_changed = prio_changed_fair, | |
da7a735e | 5789 | .switched_from = switched_from_fair, |
cb469845 | 5790 | .switched_to = switched_to_fair, |
810b3817 | 5791 | |
0d721cea PW |
5792 | .get_rr_interval = get_rr_interval_fair, |
5793 | ||
810b3817 | 5794 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5795 | .task_move_group = task_move_group_fair, |
810b3817 | 5796 | #endif |
bf0f6f24 IM |
5797 | }; |
5798 | ||
5799 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 5800 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 5801 | { |
bf0f6f24 IM |
5802 | struct cfs_rq *cfs_rq; |
5803 | ||
5973e5b9 | 5804 | rcu_read_lock(); |
c3b64f1e | 5805 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 5806 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 5807 | rcu_read_unlock(); |
bf0f6f24 IM |
5808 | } |
5809 | #endif | |
029632fb PZ |
5810 | |
5811 | __init void init_sched_fair_class(void) | |
5812 | { | |
5813 | #ifdef CONFIG_SMP | |
5814 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
5815 | ||
5816 | #ifdef CONFIG_NO_HZ | |
554cecaf | 5817 | nohz.next_balance = jiffies; |
029632fb | 5818 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 5819 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
5820 | #endif |
5821 | #endif /* SMP */ | |
5822 | ||
5823 | } |