]>
Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
3436ae12 | 25 | #include <linux/cpumask.h> |
029632fb PZ |
26 | #include <linux/slab.h> |
27 | #include <linux/profile.h> | |
28 | #include <linux/interrupt.h> | |
cbee9f88 | 29 | #include <linux/mempolicy.h> |
e14808b4 | 30 | #include <linux/migrate.h> |
cbee9f88 | 31 | #include <linux/task_work.h> |
029632fb PZ |
32 | |
33 | #include <trace/events/sched.h> | |
34 | ||
35 | #include "sched.h" | |
9745512c | 36 | |
bf0f6f24 | 37 | /* |
21805085 | 38 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 39 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 40 | * |
21805085 | 41 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
42 | * 'timeslice length' - timeslices in CFS are of variable length |
43 | * and have no persistent notion like in traditional, time-slice | |
44 | * based scheduling concepts. | |
bf0f6f24 | 45 | * |
d274a4ce IM |
46 | * (to see the precise effective timeslice length of your workload, |
47 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 48 | */ |
21406928 MG |
49 | unsigned int sysctl_sched_latency = 6000000ULL; |
50 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 51 | |
1983a922 CE |
52 | /* |
53 | * The initial- and re-scaling of tunables is configurable | |
54 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
55 | * | |
56 | * Options are: | |
57 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
58 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
59 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
60 | */ | |
61 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
62 | = SCHED_TUNABLESCALING_LOG; | |
63 | ||
2bd8e6d4 | 64 | /* |
b2be5e96 | 65 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 66 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 67 | */ |
0bf377bb IM |
68 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
69 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
70 | |
71 | /* | |
b2be5e96 PZ |
72 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
73 | */ | |
0bf377bb | 74 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
75 | |
76 | /* | |
2bba22c5 | 77 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 78 | * parent will (try to) run first. |
21805085 | 79 | */ |
2bba22c5 | 80 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 81 | |
bf0f6f24 IM |
82 | /* |
83 | * SCHED_OTHER wake-up granularity. | |
172e082a | 84 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
85 | * |
86 | * This option delays the preemption effects of decoupled workloads | |
87 | * and reduces their over-scheduling. Synchronous workloads will still | |
88 | * have immediate wakeup/sleep latencies. | |
89 | */ | |
172e082a | 90 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 91 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 92 | |
da84d961 IM |
93 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
94 | ||
a7a4f8a7 PT |
95 | /* |
96 | * The exponential sliding window over which load is averaged for shares | |
97 | * distribution. | |
98 | * (default: 10msec) | |
99 | */ | |
100 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
101 | ||
ec12cb7f PT |
102 | #ifdef CONFIG_CFS_BANDWIDTH |
103 | /* | |
104 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
105 | * each time a cfs_rq requests quota. | |
106 | * | |
107 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
108 | * to consumption or the quota being specified to be smaller than the slice) | |
109 | * we will always only issue the remaining available time. | |
110 | * | |
111 | * default: 5 msec, units: microseconds | |
112 | */ | |
113 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
114 | #endif | |
115 | ||
8527632d PG |
116 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
117 | { | |
118 | lw->weight += inc; | |
119 | lw->inv_weight = 0; | |
120 | } | |
121 | ||
122 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
123 | { | |
124 | lw->weight -= dec; | |
125 | lw->inv_weight = 0; | |
126 | } | |
127 | ||
128 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
129 | { | |
130 | lw->weight = w; | |
131 | lw->inv_weight = 0; | |
132 | } | |
133 | ||
029632fb PZ |
134 | /* |
135 | * Increase the granularity value when there are more CPUs, | |
136 | * because with more CPUs the 'effective latency' as visible | |
137 | * to users decreases. But the relationship is not linear, | |
138 | * so pick a second-best guess by going with the log2 of the | |
139 | * number of CPUs. | |
140 | * | |
141 | * This idea comes from the SD scheduler of Con Kolivas: | |
142 | */ | |
143 | static int get_update_sysctl_factor(void) | |
144 | { | |
145 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
146 | unsigned int factor; | |
147 | ||
148 | switch (sysctl_sched_tunable_scaling) { | |
149 | case SCHED_TUNABLESCALING_NONE: | |
150 | factor = 1; | |
151 | break; | |
152 | case SCHED_TUNABLESCALING_LINEAR: | |
153 | factor = cpus; | |
154 | break; | |
155 | case SCHED_TUNABLESCALING_LOG: | |
156 | default: | |
157 | factor = 1 + ilog2(cpus); | |
158 | break; | |
159 | } | |
160 | ||
161 | return factor; | |
162 | } | |
163 | ||
164 | static void update_sysctl(void) | |
165 | { | |
166 | unsigned int factor = get_update_sysctl_factor(); | |
167 | ||
168 | #define SET_SYSCTL(name) \ | |
169 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
170 | SET_SYSCTL(sched_min_granularity); | |
171 | SET_SYSCTL(sched_latency); | |
172 | SET_SYSCTL(sched_wakeup_granularity); | |
173 | #undef SET_SYSCTL | |
174 | } | |
175 | ||
176 | void sched_init_granularity(void) | |
177 | { | |
178 | update_sysctl(); | |
179 | } | |
180 | ||
181 | #if BITS_PER_LONG == 32 | |
182 | # define WMULT_CONST (~0UL) | |
183 | #else | |
184 | # define WMULT_CONST (1UL << 32) | |
185 | #endif | |
186 | ||
187 | #define WMULT_SHIFT 32 | |
188 | ||
189 | /* | |
190 | * Shift right and round: | |
191 | */ | |
192 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
193 | ||
194 | /* | |
195 | * delta *= weight / lw | |
196 | */ | |
197 | static unsigned long | |
198 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
199 | struct load_weight *lw) | |
200 | { | |
201 | u64 tmp; | |
202 | ||
203 | /* | |
204 | * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched | |
205 | * entities since MIN_SHARES = 2. Treat weight as 1 if less than | |
206 | * 2^SCHED_LOAD_RESOLUTION. | |
207 | */ | |
208 | if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) | |
209 | tmp = (u64)delta_exec * scale_load_down(weight); | |
210 | else | |
211 | tmp = (u64)delta_exec; | |
212 | ||
213 | if (!lw->inv_weight) { | |
214 | unsigned long w = scale_load_down(lw->weight); | |
215 | ||
216 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
217 | lw->inv_weight = 1; | |
218 | else if (unlikely(!w)) | |
219 | lw->inv_weight = WMULT_CONST; | |
220 | else | |
221 | lw->inv_weight = WMULT_CONST / w; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Check whether we'd overflow the 64-bit multiplication: | |
226 | */ | |
227 | if (unlikely(tmp > WMULT_CONST)) | |
228 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
229 | WMULT_SHIFT/2); | |
230 | else | |
231 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
232 | ||
233 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
234 | } | |
235 | ||
236 | ||
237 | const struct sched_class fair_sched_class; | |
a4c2f00f | 238 | |
bf0f6f24 IM |
239 | /************************************************************** |
240 | * CFS operations on generic schedulable entities: | |
241 | */ | |
242 | ||
62160e3f | 243 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 244 | |
62160e3f | 245 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
246 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
247 | { | |
62160e3f | 248 | return cfs_rq->rq; |
bf0f6f24 IM |
249 | } |
250 | ||
62160e3f IM |
251 | /* An entity is a task if it doesn't "own" a runqueue */ |
252 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 253 | |
8f48894f PZ |
254 | static inline struct task_struct *task_of(struct sched_entity *se) |
255 | { | |
256 | #ifdef CONFIG_SCHED_DEBUG | |
257 | WARN_ON_ONCE(!entity_is_task(se)); | |
258 | #endif | |
259 | return container_of(se, struct task_struct, se); | |
260 | } | |
261 | ||
b758149c PZ |
262 | /* Walk up scheduling entities hierarchy */ |
263 | #define for_each_sched_entity(se) \ | |
264 | for (; se; se = se->parent) | |
265 | ||
266 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
267 | { | |
268 | return p->se.cfs_rq; | |
269 | } | |
270 | ||
271 | /* runqueue on which this entity is (to be) queued */ | |
272 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
273 | { | |
274 | return se->cfs_rq; | |
275 | } | |
276 | ||
277 | /* runqueue "owned" by this group */ | |
278 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
279 | { | |
280 | return grp->my_q; | |
281 | } | |
282 | ||
aff3e498 PT |
283 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
284 | int force_update); | |
9ee474f5 | 285 | |
3d4b47b4 PZ |
286 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
287 | { | |
288 | if (!cfs_rq->on_list) { | |
67e86250 PT |
289 | /* |
290 | * Ensure we either appear before our parent (if already | |
291 | * enqueued) or force our parent to appear after us when it is | |
292 | * enqueued. The fact that we always enqueue bottom-up | |
293 | * reduces this to two cases. | |
294 | */ | |
295 | if (cfs_rq->tg->parent && | |
296 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
297 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
298 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
299 | } else { | |
300 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 302 | } |
3d4b47b4 PZ |
303 | |
304 | cfs_rq->on_list = 1; | |
9ee474f5 | 305 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 306 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
307 | } |
308 | } | |
309 | ||
310 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
311 | { | |
312 | if (cfs_rq->on_list) { | |
313 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
314 | cfs_rq->on_list = 0; | |
315 | } | |
316 | } | |
317 | ||
b758149c PZ |
318 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
319 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
320 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
321 | ||
322 | /* Do the two (enqueued) entities belong to the same group ? */ | |
323 | static inline int | |
324 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
325 | { | |
326 | if (se->cfs_rq == pse->cfs_rq) | |
327 | return 1; | |
328 | ||
329 | return 0; | |
330 | } | |
331 | ||
332 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
333 | { | |
334 | return se->parent; | |
335 | } | |
336 | ||
464b7527 PZ |
337 | /* return depth at which a sched entity is present in the hierarchy */ |
338 | static inline int depth_se(struct sched_entity *se) | |
339 | { | |
340 | int depth = 0; | |
341 | ||
342 | for_each_sched_entity(se) | |
343 | depth++; | |
344 | ||
345 | return depth; | |
346 | } | |
347 | ||
348 | static void | |
349 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
350 | { | |
351 | int se_depth, pse_depth; | |
352 | ||
353 | /* | |
354 | * preemption test can be made between sibling entities who are in the | |
355 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
356 | * both tasks until we find their ancestors who are siblings of common | |
357 | * parent. | |
358 | */ | |
359 | ||
360 | /* First walk up until both entities are at same depth */ | |
361 | se_depth = depth_se(*se); | |
362 | pse_depth = depth_se(*pse); | |
363 | ||
364 | while (se_depth > pse_depth) { | |
365 | se_depth--; | |
366 | *se = parent_entity(*se); | |
367 | } | |
368 | ||
369 | while (pse_depth > se_depth) { | |
370 | pse_depth--; | |
371 | *pse = parent_entity(*pse); | |
372 | } | |
373 | ||
374 | while (!is_same_group(*se, *pse)) { | |
375 | *se = parent_entity(*se); | |
376 | *pse = parent_entity(*pse); | |
377 | } | |
378 | } | |
379 | ||
8f48894f PZ |
380 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
381 | ||
382 | static inline struct task_struct *task_of(struct sched_entity *se) | |
383 | { | |
384 | return container_of(se, struct task_struct, se); | |
385 | } | |
bf0f6f24 | 386 | |
62160e3f IM |
387 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
388 | { | |
389 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
390 | } |
391 | ||
392 | #define entity_is_task(se) 1 | |
393 | ||
b758149c PZ |
394 | #define for_each_sched_entity(se) \ |
395 | for (; se; se = NULL) | |
bf0f6f24 | 396 | |
b758149c | 397 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 398 | { |
b758149c | 399 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
400 | } |
401 | ||
b758149c PZ |
402 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
403 | { | |
404 | struct task_struct *p = task_of(se); | |
405 | struct rq *rq = task_rq(p); | |
406 | ||
407 | return &rq->cfs; | |
408 | } | |
409 | ||
410 | /* runqueue "owned" by this group */ | |
411 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
412 | { | |
413 | return NULL; | |
414 | } | |
415 | ||
3d4b47b4 PZ |
416 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
417 | { | |
418 | } | |
419 | ||
420 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
421 | { | |
422 | } | |
423 | ||
b758149c PZ |
424 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
425 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
426 | ||
427 | static inline int | |
428 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
429 | { | |
430 | return 1; | |
431 | } | |
432 | ||
433 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
434 | { | |
435 | return NULL; | |
436 | } | |
437 | ||
464b7527 PZ |
438 | static inline void |
439 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
440 | { | |
441 | } | |
442 | ||
b758149c PZ |
443 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
444 | ||
6c16a6dc PZ |
445 | static __always_inline |
446 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
447 | |
448 | /************************************************************** | |
449 | * Scheduling class tree data structure manipulation methods: | |
450 | */ | |
451 | ||
1bf08230 | 452 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 453 | { |
1bf08230 | 454 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 455 | if (delta > 0) |
1bf08230 | 456 | max_vruntime = vruntime; |
02e0431a | 457 | |
1bf08230 | 458 | return max_vruntime; |
02e0431a PZ |
459 | } |
460 | ||
0702e3eb | 461 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
462 | { |
463 | s64 delta = (s64)(vruntime - min_vruntime); | |
464 | if (delta < 0) | |
465 | min_vruntime = vruntime; | |
466 | ||
467 | return min_vruntime; | |
468 | } | |
469 | ||
54fdc581 FC |
470 | static inline int entity_before(struct sched_entity *a, |
471 | struct sched_entity *b) | |
472 | { | |
473 | return (s64)(a->vruntime - b->vruntime) < 0; | |
474 | } | |
475 | ||
1af5f730 PZ |
476 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
477 | { | |
478 | u64 vruntime = cfs_rq->min_vruntime; | |
479 | ||
480 | if (cfs_rq->curr) | |
481 | vruntime = cfs_rq->curr->vruntime; | |
482 | ||
483 | if (cfs_rq->rb_leftmost) { | |
484 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
485 | struct sched_entity, | |
486 | run_node); | |
487 | ||
e17036da | 488 | if (!cfs_rq->curr) |
1af5f730 PZ |
489 | vruntime = se->vruntime; |
490 | else | |
491 | vruntime = min_vruntime(vruntime, se->vruntime); | |
492 | } | |
493 | ||
1bf08230 | 494 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 495 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
496 | #ifndef CONFIG_64BIT |
497 | smp_wmb(); | |
498 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
499 | #endif | |
1af5f730 PZ |
500 | } |
501 | ||
bf0f6f24 IM |
502 | /* |
503 | * Enqueue an entity into the rb-tree: | |
504 | */ | |
0702e3eb | 505 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
506 | { |
507 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
508 | struct rb_node *parent = NULL; | |
509 | struct sched_entity *entry; | |
bf0f6f24 IM |
510 | int leftmost = 1; |
511 | ||
512 | /* | |
513 | * Find the right place in the rbtree: | |
514 | */ | |
515 | while (*link) { | |
516 | parent = *link; | |
517 | entry = rb_entry(parent, struct sched_entity, run_node); | |
518 | /* | |
519 | * We dont care about collisions. Nodes with | |
520 | * the same key stay together. | |
521 | */ | |
2bd2d6f2 | 522 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
523 | link = &parent->rb_left; |
524 | } else { | |
525 | link = &parent->rb_right; | |
526 | leftmost = 0; | |
527 | } | |
528 | } | |
529 | ||
530 | /* | |
531 | * Maintain a cache of leftmost tree entries (it is frequently | |
532 | * used): | |
533 | */ | |
1af5f730 | 534 | if (leftmost) |
57cb499d | 535 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
536 | |
537 | rb_link_node(&se->run_node, parent, link); | |
538 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
539 | } |
540 | ||
0702e3eb | 541 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 542 | { |
3fe69747 PZ |
543 | if (cfs_rq->rb_leftmost == &se->run_node) { |
544 | struct rb_node *next_node; | |
3fe69747 PZ |
545 | |
546 | next_node = rb_next(&se->run_node); | |
547 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 548 | } |
e9acbff6 | 549 | |
bf0f6f24 | 550 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
551 | } |
552 | ||
029632fb | 553 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 554 | { |
f4b6755f PZ |
555 | struct rb_node *left = cfs_rq->rb_leftmost; |
556 | ||
557 | if (!left) | |
558 | return NULL; | |
559 | ||
560 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
561 | } |
562 | ||
ac53db59 RR |
563 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
564 | { | |
565 | struct rb_node *next = rb_next(&se->run_node); | |
566 | ||
567 | if (!next) | |
568 | return NULL; | |
569 | ||
570 | return rb_entry(next, struct sched_entity, run_node); | |
571 | } | |
572 | ||
573 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 574 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 575 | { |
7eee3e67 | 576 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 577 | |
70eee74b BS |
578 | if (!last) |
579 | return NULL; | |
7eee3e67 IM |
580 | |
581 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
582 | } |
583 | ||
bf0f6f24 IM |
584 | /************************************************************** |
585 | * Scheduling class statistics methods: | |
586 | */ | |
587 | ||
acb4a848 | 588 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 589 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
590 | loff_t *ppos) |
591 | { | |
8d65af78 | 592 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 593 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
594 | |
595 | if (ret || !write) | |
596 | return ret; | |
597 | ||
598 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
599 | sysctl_sched_min_granularity); | |
600 | ||
acb4a848 CE |
601 | #define WRT_SYSCTL(name) \ |
602 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
603 | WRT_SYSCTL(sched_min_granularity); | |
604 | WRT_SYSCTL(sched_latency); | |
605 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
606 | #undef WRT_SYSCTL |
607 | ||
b2be5e96 PZ |
608 | return 0; |
609 | } | |
610 | #endif | |
647e7cac | 611 | |
a7be37ac | 612 | /* |
f9c0b095 | 613 | * delta /= w |
a7be37ac PZ |
614 | */ |
615 | static inline unsigned long | |
616 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
617 | { | |
f9c0b095 PZ |
618 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
619 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
620 | |
621 | return delta; | |
622 | } | |
623 | ||
647e7cac IM |
624 | /* |
625 | * The idea is to set a period in which each task runs once. | |
626 | * | |
532b1858 | 627 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
628 | * this period because otherwise the slices get too small. |
629 | * | |
630 | * p = (nr <= nl) ? l : l*nr/nl | |
631 | */ | |
4d78e7b6 PZ |
632 | static u64 __sched_period(unsigned long nr_running) |
633 | { | |
634 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 635 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
636 | |
637 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 638 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 639 | period *= nr_running; |
4d78e7b6 PZ |
640 | } |
641 | ||
642 | return period; | |
643 | } | |
644 | ||
647e7cac IM |
645 | /* |
646 | * We calculate the wall-time slice from the period by taking a part | |
647 | * proportional to the weight. | |
648 | * | |
f9c0b095 | 649 | * s = p*P[w/rw] |
647e7cac | 650 | */ |
6d0f0ebd | 651 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 652 | { |
0a582440 | 653 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 654 | |
0a582440 | 655 | for_each_sched_entity(se) { |
6272d68c | 656 | struct load_weight *load; |
3104bf03 | 657 | struct load_weight lw; |
6272d68c LM |
658 | |
659 | cfs_rq = cfs_rq_of(se); | |
660 | load = &cfs_rq->load; | |
f9c0b095 | 661 | |
0a582440 | 662 | if (unlikely(!se->on_rq)) { |
3104bf03 | 663 | lw = cfs_rq->load; |
0a582440 MG |
664 | |
665 | update_load_add(&lw, se->load.weight); | |
666 | load = &lw; | |
667 | } | |
668 | slice = calc_delta_mine(slice, se->load.weight, load); | |
669 | } | |
670 | return slice; | |
bf0f6f24 IM |
671 | } |
672 | ||
647e7cac | 673 | /* |
660cc00f | 674 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 675 | * |
f9c0b095 | 676 | * vs = s/w |
647e7cac | 677 | */ |
f9c0b095 | 678 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 679 | { |
f9c0b095 | 680 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
681 | } |
682 | ||
bf0f6f24 IM |
683 | /* |
684 | * Update the current task's runtime statistics. Skip current tasks that | |
685 | * are not in our scheduling class. | |
686 | */ | |
687 | static inline void | |
8ebc91d9 IM |
688 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
689 | unsigned long delta_exec) | |
bf0f6f24 | 690 | { |
bbdba7c0 | 691 | unsigned long delta_exec_weighted; |
bf0f6f24 | 692 | |
41acab88 LDM |
693 | schedstat_set(curr->statistics.exec_max, |
694 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
695 | |
696 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 697 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 698 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 699 | |
e9acbff6 | 700 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 701 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
702 | } |
703 | ||
b7cc0896 | 704 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 705 | { |
429d43bc | 706 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 707 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
708 | unsigned long delta_exec; |
709 | ||
710 | if (unlikely(!curr)) | |
711 | return; | |
712 | ||
713 | /* | |
714 | * Get the amount of time the current task was running | |
715 | * since the last time we changed load (this cannot | |
716 | * overflow on 32 bits): | |
717 | */ | |
8ebc91d9 | 718 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
719 | if (!delta_exec) |
720 | return; | |
bf0f6f24 | 721 | |
8ebc91d9 IM |
722 | __update_curr(cfs_rq, curr, delta_exec); |
723 | curr->exec_start = now; | |
d842de87 SV |
724 | |
725 | if (entity_is_task(curr)) { | |
726 | struct task_struct *curtask = task_of(curr); | |
727 | ||
f977bb49 | 728 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 729 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 730 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 731 | } |
ec12cb7f PT |
732 | |
733 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
734 | } |
735 | ||
736 | static inline void | |
5870db5b | 737 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 738 | { |
78becc27 | 739 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
740 | } |
741 | ||
bf0f6f24 IM |
742 | /* |
743 | * Task is being enqueued - update stats: | |
744 | */ | |
d2417e5a | 745 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 746 | { |
bf0f6f24 IM |
747 | /* |
748 | * Are we enqueueing a waiting task? (for current tasks | |
749 | * a dequeue/enqueue event is a NOP) | |
750 | */ | |
429d43bc | 751 | if (se != cfs_rq->curr) |
5870db5b | 752 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
753 | } |
754 | ||
bf0f6f24 | 755 | static void |
9ef0a961 | 756 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 757 | { |
41acab88 | 758 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 759 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
760 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
761 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 762 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
763 | #ifdef CONFIG_SCHEDSTATS |
764 | if (entity_is_task(se)) { | |
765 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 766 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
767 | } |
768 | #endif | |
41acab88 | 769 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
770 | } |
771 | ||
772 | static inline void | |
19b6a2e3 | 773 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 774 | { |
bf0f6f24 IM |
775 | /* |
776 | * Mark the end of the wait period if dequeueing a | |
777 | * waiting task: | |
778 | */ | |
429d43bc | 779 | if (se != cfs_rq->curr) |
9ef0a961 | 780 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
781 | } |
782 | ||
783 | /* | |
784 | * We are picking a new current task - update its stats: | |
785 | */ | |
786 | static inline void | |
79303e9e | 787 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
788 | { |
789 | /* | |
790 | * We are starting a new run period: | |
791 | */ | |
78becc27 | 792 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
793 | } |
794 | ||
bf0f6f24 IM |
795 | /************************************************** |
796 | * Scheduling class queueing methods: | |
797 | */ | |
798 | ||
cbee9f88 PZ |
799 | #ifdef CONFIG_NUMA_BALANCING |
800 | /* | |
6e5fb223 | 801 | * numa task sample period in ms |
cbee9f88 | 802 | */ |
6e5fb223 | 803 | unsigned int sysctl_numa_balancing_scan_period_min = 100; |
b8593bfd MG |
804 | unsigned int sysctl_numa_balancing_scan_period_max = 100*50; |
805 | unsigned int sysctl_numa_balancing_scan_period_reset = 100*600; | |
6e5fb223 PZ |
806 | |
807 | /* Portion of address space to scan in MB */ | |
808 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 809 | |
4b96a29b PZ |
810 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
811 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
812 | ||
cbee9f88 PZ |
813 | static void task_numa_placement(struct task_struct *p) |
814 | { | |
2832bc19 | 815 | int seq; |
cbee9f88 | 816 | |
2832bc19 HD |
817 | if (!p->mm) /* for example, ksmd faulting in a user's mm */ |
818 | return; | |
819 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); | |
cbee9f88 PZ |
820 | if (p->numa_scan_seq == seq) |
821 | return; | |
822 | p->numa_scan_seq = seq; | |
823 | ||
824 | /* FIXME: Scheduling placement policy hints go here */ | |
825 | } | |
826 | ||
827 | /* | |
828 | * Got a PROT_NONE fault for a page on @node. | |
829 | */ | |
b8593bfd | 830 | void task_numa_fault(int node, int pages, bool migrated) |
cbee9f88 PZ |
831 | { |
832 | struct task_struct *p = current; | |
833 | ||
1a687c2e MG |
834 | if (!sched_feat_numa(NUMA)) |
835 | return; | |
836 | ||
cbee9f88 PZ |
837 | /* FIXME: Allocate task-specific structure for placement policy here */ |
838 | ||
fb003b80 | 839 | /* |
b8593bfd MG |
840 | * If pages are properly placed (did not migrate) then scan slower. |
841 | * This is reset periodically in case of phase changes | |
fb003b80 | 842 | */ |
b8593bfd MG |
843 | if (!migrated) |
844 | p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max, | |
845 | p->numa_scan_period + jiffies_to_msecs(10)); | |
fb003b80 | 846 | |
cbee9f88 PZ |
847 | task_numa_placement(p); |
848 | } | |
849 | ||
6e5fb223 PZ |
850 | static void reset_ptenuma_scan(struct task_struct *p) |
851 | { | |
852 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
853 | p->mm->numa_scan_offset = 0; | |
854 | } | |
855 | ||
cbee9f88 PZ |
856 | /* |
857 | * The expensive part of numa migration is done from task_work context. | |
858 | * Triggered from task_tick_numa(). | |
859 | */ | |
860 | void task_numa_work(struct callback_head *work) | |
861 | { | |
862 | unsigned long migrate, next_scan, now = jiffies; | |
863 | struct task_struct *p = current; | |
864 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 865 | struct vm_area_struct *vma; |
9f40604c MG |
866 | unsigned long start, end; |
867 | long pages; | |
cbee9f88 PZ |
868 | |
869 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
870 | ||
871 | work->next = work; /* protect against double add */ | |
872 | /* | |
873 | * Who cares about NUMA placement when they're dying. | |
874 | * | |
875 | * NOTE: make sure not to dereference p->mm before this check, | |
876 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
877 | * without p->mm even though we still had it when we enqueued this | |
878 | * work. | |
879 | */ | |
880 | if (p->flags & PF_EXITING) | |
881 | return; | |
882 | ||
5bca2303 MG |
883 | /* |
884 | * We do not care about task placement until a task runs on a node | |
885 | * other than the first one used by the address space. This is | |
886 | * largely because migrations are driven by what CPU the task | |
887 | * is running on. If it's never scheduled on another node, it'll | |
888 | * not migrate so why bother trapping the fault. | |
889 | */ | |
890 | if (mm->first_nid == NUMA_PTE_SCAN_INIT) | |
891 | mm->first_nid = numa_node_id(); | |
892 | if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) { | |
893 | /* Are we running on a new node yet? */ | |
894 | if (numa_node_id() == mm->first_nid && | |
895 | !sched_feat_numa(NUMA_FORCE)) | |
896 | return; | |
897 | ||
898 | mm->first_nid = NUMA_PTE_SCAN_ACTIVE; | |
899 | } | |
900 | ||
b8593bfd MG |
901 | /* |
902 | * Reset the scan period if enough time has gone by. Objective is that | |
903 | * scanning will be reduced if pages are properly placed. As tasks | |
904 | * can enter different phases this needs to be re-examined. Lacking | |
905 | * proper tracking of reference behaviour, this blunt hammer is used. | |
906 | */ | |
907 | migrate = mm->numa_next_reset; | |
908 | if (time_after(now, migrate)) { | |
909 | p->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
910 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
911 | xchg(&mm->numa_next_reset, next_scan); | |
912 | } | |
913 | ||
cbee9f88 PZ |
914 | /* |
915 | * Enforce maximal scan/migration frequency.. | |
916 | */ | |
917 | migrate = mm->numa_next_scan; | |
918 | if (time_before(now, migrate)) | |
919 | return; | |
920 | ||
921 | if (p->numa_scan_period == 0) | |
922 | p->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
923 | ||
fb003b80 | 924 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
925 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
926 | return; | |
927 | ||
e14808b4 MG |
928 | /* |
929 | * Do not set pte_numa if the current running node is rate-limited. | |
930 | * This loses statistics on the fault but if we are unwilling to | |
931 | * migrate to this node, it is less likely we can do useful work | |
932 | */ | |
933 | if (migrate_ratelimited(numa_node_id())) | |
934 | return; | |
935 | ||
9f40604c MG |
936 | start = mm->numa_scan_offset; |
937 | pages = sysctl_numa_balancing_scan_size; | |
938 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
939 | if (!pages) | |
940 | return; | |
cbee9f88 | 941 | |
6e5fb223 | 942 | down_read(&mm->mmap_sem); |
9f40604c | 943 | vma = find_vma(mm, start); |
6e5fb223 PZ |
944 | if (!vma) { |
945 | reset_ptenuma_scan(p); | |
9f40604c | 946 | start = 0; |
6e5fb223 PZ |
947 | vma = mm->mmap; |
948 | } | |
9f40604c | 949 | for (; vma; vma = vma->vm_next) { |
6e5fb223 PZ |
950 | if (!vma_migratable(vma)) |
951 | continue; | |
952 | ||
953 | /* Skip small VMAs. They are not likely to be of relevance */ | |
221392c3 | 954 | if (vma->vm_end - vma->vm_start < HPAGE_SIZE) |
6e5fb223 PZ |
955 | continue; |
956 | ||
9f40604c MG |
957 | do { |
958 | start = max(start, vma->vm_start); | |
959 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
960 | end = min(end, vma->vm_end); | |
961 | pages -= change_prot_numa(vma, start, end); | |
6e5fb223 | 962 | |
9f40604c MG |
963 | start = end; |
964 | if (pages <= 0) | |
965 | goto out; | |
966 | } while (end != vma->vm_end); | |
cbee9f88 | 967 | } |
6e5fb223 | 968 | |
9f40604c | 969 | out: |
6e5fb223 PZ |
970 | /* |
971 | * It is possible to reach the end of the VMA list but the last few VMAs are | |
972 | * not guaranteed to the vma_migratable. If they are not, we would find the | |
973 | * !migratable VMA on the next scan but not reset the scanner to the start | |
974 | * so check it now. | |
975 | */ | |
976 | if (vma) | |
9f40604c | 977 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
978 | else |
979 | reset_ptenuma_scan(p); | |
980 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
981 | } |
982 | ||
983 | /* | |
984 | * Drive the periodic memory faults.. | |
985 | */ | |
986 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
987 | { | |
988 | struct callback_head *work = &curr->numa_work; | |
989 | u64 period, now; | |
990 | ||
991 | /* | |
992 | * We don't care about NUMA placement if we don't have memory. | |
993 | */ | |
994 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
995 | return; | |
996 | ||
997 | /* | |
998 | * Using runtime rather than walltime has the dual advantage that | |
999 | * we (mostly) drive the selection from busy threads and that the | |
1000 | * task needs to have done some actual work before we bother with | |
1001 | * NUMA placement. | |
1002 | */ | |
1003 | now = curr->se.sum_exec_runtime; | |
1004 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1005 | ||
1006 | if (now - curr->node_stamp > period) { | |
4b96a29b PZ |
1007 | if (!curr->node_stamp) |
1008 | curr->numa_scan_period = sysctl_numa_balancing_scan_period_min; | |
cbee9f88 PZ |
1009 | curr->node_stamp = now; |
1010 | ||
1011 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1012 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1013 | task_work_add(curr, work, true); | |
1014 | } | |
1015 | } | |
1016 | } | |
1017 | #else | |
1018 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1019 | { | |
1020 | } | |
1021 | #endif /* CONFIG_NUMA_BALANCING */ | |
1022 | ||
30cfdcfc DA |
1023 | static void |
1024 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1025 | { | |
1026 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1027 | if (!parent_entity(se)) |
029632fb | 1028 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
1029 | #ifdef CONFIG_SMP |
1030 | if (entity_is_task(se)) | |
eb95308e | 1031 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 1032 | #endif |
30cfdcfc | 1033 | cfs_rq->nr_running++; |
30cfdcfc DA |
1034 | } |
1035 | ||
1036 | static void | |
1037 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1038 | { | |
1039 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1040 | if (!parent_entity(se)) |
029632fb | 1041 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1042 | if (entity_is_task(se)) |
b87f1724 | 1043 | list_del_init(&se->group_node); |
30cfdcfc | 1044 | cfs_rq->nr_running--; |
30cfdcfc DA |
1045 | } |
1046 | ||
3ff6dcac YZ |
1047 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1048 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1049 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1050 | { | |
1051 | long tg_weight; | |
1052 | ||
1053 | /* | |
1054 | * Use this CPU's actual weight instead of the last load_contribution | |
1055 | * to gain a more accurate current total weight. See | |
1056 | * update_cfs_rq_load_contribution(). | |
1057 | */ | |
82958366 PT |
1058 | tg_weight = atomic64_read(&tg->load_avg); |
1059 | tg_weight -= cfs_rq->tg_load_contrib; | |
cf5f0acf PZ |
1060 | tg_weight += cfs_rq->load.weight; |
1061 | ||
1062 | return tg_weight; | |
1063 | } | |
1064 | ||
6d5ab293 | 1065 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1066 | { |
cf5f0acf | 1067 | long tg_weight, load, shares; |
3ff6dcac | 1068 | |
cf5f0acf | 1069 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1070 | load = cfs_rq->load.weight; |
3ff6dcac | 1071 | |
3ff6dcac | 1072 | shares = (tg->shares * load); |
cf5f0acf PZ |
1073 | if (tg_weight) |
1074 | shares /= tg_weight; | |
3ff6dcac YZ |
1075 | |
1076 | if (shares < MIN_SHARES) | |
1077 | shares = MIN_SHARES; | |
1078 | if (shares > tg->shares) | |
1079 | shares = tg->shares; | |
1080 | ||
1081 | return shares; | |
1082 | } | |
3ff6dcac | 1083 | # else /* CONFIG_SMP */ |
6d5ab293 | 1084 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1085 | { |
1086 | return tg->shares; | |
1087 | } | |
3ff6dcac | 1088 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1089 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1090 | unsigned long weight) | |
1091 | { | |
19e5eebb PT |
1092 | if (se->on_rq) { |
1093 | /* commit outstanding execution time */ | |
1094 | if (cfs_rq->curr == se) | |
1095 | update_curr(cfs_rq); | |
2069dd75 | 1096 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1097 | } |
2069dd75 PZ |
1098 | |
1099 | update_load_set(&se->load, weight); | |
1100 | ||
1101 | if (se->on_rq) | |
1102 | account_entity_enqueue(cfs_rq, se); | |
1103 | } | |
1104 | ||
82958366 PT |
1105 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1106 | ||
6d5ab293 | 1107 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1108 | { |
1109 | struct task_group *tg; | |
1110 | struct sched_entity *se; | |
3ff6dcac | 1111 | long shares; |
2069dd75 | 1112 | |
2069dd75 PZ |
1113 | tg = cfs_rq->tg; |
1114 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1115 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1116 | return; |
3ff6dcac YZ |
1117 | #ifndef CONFIG_SMP |
1118 | if (likely(se->load.weight == tg->shares)) | |
1119 | return; | |
1120 | #endif | |
6d5ab293 | 1121 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1122 | |
1123 | reweight_entity(cfs_rq_of(se), se, shares); | |
1124 | } | |
1125 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1126 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1127 | { |
1128 | } | |
1129 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1130 | ||
f4e26b12 PT |
1131 | /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */ |
1132 | #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED) | |
5b51f2f8 PT |
1133 | /* |
1134 | * We choose a half-life close to 1 scheduling period. | |
1135 | * Note: The tables below are dependent on this value. | |
1136 | */ | |
1137 | #define LOAD_AVG_PERIOD 32 | |
1138 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1139 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1140 | ||
1141 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1142 | static const u32 runnable_avg_yN_inv[] = { | |
1143 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1144 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1145 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1146 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1147 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1148 | 0x85aac367, 0x82cd8698, | |
1149 | }; | |
1150 | ||
1151 | /* | |
1152 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1153 | * over-estimates when re-combining. | |
1154 | */ | |
1155 | static const u32 runnable_avg_yN_sum[] = { | |
1156 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1157 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1158 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1159 | }; | |
1160 | ||
9d85f21c PT |
1161 | /* |
1162 | * Approximate: | |
1163 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1164 | */ | |
1165 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1166 | { | |
5b51f2f8 PT |
1167 | unsigned int local_n; |
1168 | ||
1169 | if (!n) | |
1170 | return val; | |
1171 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1172 | return 0; | |
1173 | ||
1174 | /* after bounds checking we can collapse to 32-bit */ | |
1175 | local_n = n; | |
1176 | ||
1177 | /* | |
1178 | * As y^PERIOD = 1/2, we can combine | |
1179 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
1180 | * With a look-up table which covers k^n (n<PERIOD) | |
1181 | * | |
1182 | * To achieve constant time decay_load. | |
1183 | */ | |
1184 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
1185 | val >>= local_n / LOAD_AVG_PERIOD; | |
1186 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
1187 | } |
1188 | ||
5b51f2f8 PT |
1189 | val *= runnable_avg_yN_inv[local_n]; |
1190 | /* We don't use SRR here since we always want to round down. */ | |
1191 | return val >> 32; | |
1192 | } | |
1193 | ||
1194 | /* | |
1195 | * For updates fully spanning n periods, the contribution to runnable | |
1196 | * average will be: \Sum 1024*y^n | |
1197 | * | |
1198 | * We can compute this reasonably efficiently by combining: | |
1199 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
1200 | */ | |
1201 | static u32 __compute_runnable_contrib(u64 n) | |
1202 | { | |
1203 | u32 contrib = 0; | |
1204 | ||
1205 | if (likely(n <= LOAD_AVG_PERIOD)) | |
1206 | return runnable_avg_yN_sum[n]; | |
1207 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
1208 | return LOAD_AVG_MAX; | |
1209 | ||
1210 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
1211 | do { | |
1212 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
1213 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
1214 | ||
1215 | n -= LOAD_AVG_PERIOD; | |
1216 | } while (n > LOAD_AVG_PERIOD); | |
1217 | ||
1218 | contrib = decay_load(contrib, n); | |
1219 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
1220 | } |
1221 | ||
1222 | /* | |
1223 | * We can represent the historical contribution to runnable average as the | |
1224 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
1225 | * history into segments of approximately 1ms (1024us); label the segment that | |
1226 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
1227 | * | |
1228 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
1229 | * p0 p1 p2 | |
1230 | * (now) (~1ms ago) (~2ms ago) | |
1231 | * | |
1232 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
1233 | * | |
1234 | * We then designate the fractions u_i as our co-efficients, yielding the | |
1235 | * following representation of historical load: | |
1236 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
1237 | * | |
1238 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
1239 | * y^32 = 0.5 | |
1240 | * | |
1241 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
1242 | * approximately half as much as the contribution to load within the last ms | |
1243 | * (u_0). | |
1244 | * | |
1245 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
1246 | * sum again by y is sufficient to update: | |
1247 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1248 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1249 | */ | |
1250 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1251 | struct sched_avg *sa, | |
1252 | int runnable) | |
1253 | { | |
5b51f2f8 PT |
1254 | u64 delta, periods; |
1255 | u32 runnable_contrib; | |
9d85f21c PT |
1256 | int delta_w, decayed = 0; |
1257 | ||
1258 | delta = now - sa->last_runnable_update; | |
1259 | /* | |
1260 | * This should only happen when time goes backwards, which it | |
1261 | * unfortunately does during sched clock init when we swap over to TSC. | |
1262 | */ | |
1263 | if ((s64)delta < 0) { | |
1264 | sa->last_runnable_update = now; | |
1265 | return 0; | |
1266 | } | |
1267 | ||
1268 | /* | |
1269 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1270 | * approximation of 1us and fast to compute. | |
1271 | */ | |
1272 | delta >>= 10; | |
1273 | if (!delta) | |
1274 | return 0; | |
1275 | sa->last_runnable_update = now; | |
1276 | ||
1277 | /* delta_w is the amount already accumulated against our next period */ | |
1278 | delta_w = sa->runnable_avg_period % 1024; | |
1279 | if (delta + delta_w >= 1024) { | |
1280 | /* period roll-over */ | |
1281 | decayed = 1; | |
1282 | ||
1283 | /* | |
1284 | * Now that we know we're crossing a period boundary, figure | |
1285 | * out how much from delta we need to complete the current | |
1286 | * period and accrue it. | |
1287 | */ | |
1288 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
1289 | if (runnable) |
1290 | sa->runnable_avg_sum += delta_w; | |
1291 | sa->runnable_avg_period += delta_w; | |
1292 | ||
1293 | delta -= delta_w; | |
1294 | ||
1295 | /* Figure out how many additional periods this update spans */ | |
1296 | periods = delta / 1024; | |
1297 | delta %= 1024; | |
1298 | ||
1299 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
1300 | periods + 1); | |
1301 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
1302 | periods + 1); | |
1303 | ||
1304 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
1305 | runnable_contrib = __compute_runnable_contrib(periods); | |
1306 | if (runnable) | |
1307 | sa->runnable_avg_sum += runnable_contrib; | |
1308 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
1309 | } |
1310 | ||
1311 | /* Remainder of delta accrued against u_0` */ | |
1312 | if (runnable) | |
1313 | sa->runnable_avg_sum += delta; | |
1314 | sa->runnable_avg_period += delta; | |
1315 | ||
1316 | return decayed; | |
1317 | } | |
1318 | ||
9ee474f5 | 1319 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 1320 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
1321 | { |
1322 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1323 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1324 | ||
1325 | decays -= se->avg.decay_count; | |
1326 | if (!decays) | |
aff3e498 | 1327 | return 0; |
9ee474f5 PT |
1328 | |
1329 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1330 | se->avg.decay_count = 0; | |
aff3e498 PT |
1331 | |
1332 | return decays; | |
9ee474f5 PT |
1333 | } |
1334 | ||
c566e8e9 PT |
1335 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1336 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1337 | int force_update) | |
1338 | { | |
1339 | struct task_group *tg = cfs_rq->tg; | |
1340 | s64 tg_contrib; | |
1341 | ||
1342 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
1343 | tg_contrib -= cfs_rq->tg_load_contrib; | |
1344 | ||
1345 | if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) { | |
1346 | atomic64_add(tg_contrib, &tg->load_avg); | |
1347 | cfs_rq->tg_load_contrib += tg_contrib; | |
1348 | } | |
1349 | } | |
8165e145 | 1350 | |
bb17f655 PT |
1351 | /* |
1352 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
1353 | * representation for computing load contributions. | |
1354 | */ | |
1355 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
1356 | struct cfs_rq *cfs_rq) | |
1357 | { | |
1358 | struct task_group *tg = cfs_rq->tg; | |
1359 | long contrib; | |
1360 | ||
1361 | /* The fraction of a cpu used by this cfs_rq */ | |
1362 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
1363 | sa->runnable_avg_period + 1); | |
1364 | contrib -= cfs_rq->tg_runnable_contrib; | |
1365 | ||
1366 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
1367 | atomic_add(contrib, &tg->runnable_avg); | |
1368 | cfs_rq->tg_runnable_contrib += contrib; | |
1369 | } | |
1370 | } | |
1371 | ||
8165e145 PT |
1372 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
1373 | { | |
1374 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
1375 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
1376 | int runnable_avg; |
1377 | ||
8165e145 PT |
1378 | u64 contrib; |
1379 | ||
1380 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
1381 | se->avg.load_avg_contrib = div64_u64(contrib, | |
1382 | atomic64_read(&tg->load_avg) + 1); | |
bb17f655 PT |
1383 | |
1384 | /* | |
1385 | * For group entities we need to compute a correction term in the case | |
1386 | * that they are consuming <1 cpu so that we would contribute the same | |
1387 | * load as a task of equal weight. | |
1388 | * | |
1389 | * Explicitly co-ordinating this measurement would be expensive, but | |
1390 | * fortunately the sum of each cpus contribution forms a usable | |
1391 | * lower-bound on the true value. | |
1392 | * | |
1393 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
1394 | * (and the sum represents true value) or they are disjoint and we are | |
1395 | * understating by the aggregate of their overlap. | |
1396 | * | |
1397 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
1398 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
1399 | * cpus that overlap for this interval and w_i is the interval width. | |
1400 | * | |
1401 | * On a small machine; the first term is well-bounded which bounds the | |
1402 | * total error since w_i is a subset of the period. Whereas on a | |
1403 | * larger machine, while this first term can be larger, if w_i is the | |
1404 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
1405 | * our upper bound of 1-cpu. | |
1406 | */ | |
1407 | runnable_avg = atomic_read(&tg->runnable_avg); | |
1408 | if (runnable_avg < NICE_0_LOAD) { | |
1409 | se->avg.load_avg_contrib *= runnable_avg; | |
1410 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
1411 | } | |
8165e145 | 1412 | } |
c566e8e9 PT |
1413 | #else |
1414 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1415 | int force_update) {} | |
bb17f655 PT |
1416 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
1417 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 1418 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
1419 | #endif |
1420 | ||
8165e145 PT |
1421 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
1422 | { | |
1423 | u32 contrib; | |
1424 | ||
1425 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
1426 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
1427 | contrib /= (se->avg.runnable_avg_period + 1); | |
1428 | se->avg.load_avg_contrib = scale_load(contrib); | |
1429 | } | |
1430 | ||
2dac754e PT |
1431 | /* Compute the current contribution to load_avg by se, return any delta */ |
1432 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
1433 | { | |
1434 | long old_contrib = se->avg.load_avg_contrib; | |
1435 | ||
8165e145 PT |
1436 | if (entity_is_task(se)) { |
1437 | __update_task_entity_contrib(se); | |
1438 | } else { | |
bb17f655 | 1439 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
1440 | __update_group_entity_contrib(se); |
1441 | } | |
2dac754e PT |
1442 | |
1443 | return se->avg.load_avg_contrib - old_contrib; | |
1444 | } | |
1445 | ||
9ee474f5 PT |
1446 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
1447 | long load_contrib) | |
1448 | { | |
1449 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
1450 | cfs_rq->blocked_load_avg -= load_contrib; | |
1451 | else | |
1452 | cfs_rq->blocked_load_avg = 0; | |
1453 | } | |
1454 | ||
f1b17280 PT |
1455 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
1456 | ||
9d85f21c | 1457 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
1458 | static inline void update_entity_load_avg(struct sched_entity *se, |
1459 | int update_cfs_rq) | |
9d85f21c | 1460 | { |
2dac754e PT |
1461 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
1462 | long contrib_delta; | |
f1b17280 | 1463 | u64 now; |
2dac754e | 1464 | |
f1b17280 PT |
1465 | /* |
1466 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
1467 | * case they are the parent of a throttled hierarchy. | |
1468 | */ | |
1469 | if (entity_is_task(se)) | |
1470 | now = cfs_rq_clock_task(cfs_rq); | |
1471 | else | |
1472 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
1473 | ||
1474 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
1475 | return; |
1476 | ||
1477 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
1478 | |
1479 | if (!update_cfs_rq) | |
1480 | return; | |
1481 | ||
2dac754e PT |
1482 | if (se->on_rq) |
1483 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
1484 | else |
1485 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
1486 | } | |
1487 | ||
1488 | /* | |
1489 | * Decay the load contributed by all blocked children and account this so that | |
1490 | * their contribution may appropriately discounted when they wake up. | |
1491 | */ | |
aff3e498 | 1492 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 1493 | { |
f1b17280 | 1494 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
1495 | u64 decays; |
1496 | ||
1497 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 1498 | if (!decays && !force_update) |
9ee474f5 PT |
1499 | return; |
1500 | ||
aff3e498 PT |
1501 | if (atomic64_read(&cfs_rq->removed_load)) { |
1502 | u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0); | |
1503 | subtract_blocked_load_contrib(cfs_rq, removed_load); | |
1504 | } | |
9ee474f5 | 1505 | |
aff3e498 PT |
1506 | if (decays) { |
1507 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
1508 | decays); | |
1509 | atomic64_add(decays, &cfs_rq->decay_counter); | |
1510 | cfs_rq->last_decay = now; | |
1511 | } | |
c566e8e9 PT |
1512 | |
1513 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 1514 | } |
18bf2805 BS |
1515 | |
1516 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
1517 | { | |
78becc27 | 1518 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 1519 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 1520 | } |
2dac754e PT |
1521 | |
1522 | /* Add the load generated by se into cfs_rq's child load-average */ | |
1523 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
1524 | struct sched_entity *se, |
1525 | int wakeup) | |
2dac754e | 1526 | { |
aff3e498 PT |
1527 | /* |
1528 | * We track migrations using entity decay_count <= 0, on a wake-up | |
1529 | * migration we use a negative decay count to track the remote decays | |
1530 | * accumulated while sleeping. | |
1531 | */ | |
1532 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 1533 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
1534 | if (se->avg.decay_count) { |
1535 | /* | |
1536 | * In a wake-up migration we have to approximate the | |
1537 | * time sleeping. This is because we can't synchronize | |
1538 | * clock_task between the two cpus, and it is not | |
1539 | * guaranteed to be read-safe. Instead, we can | |
1540 | * approximate this using our carried decays, which are | |
1541 | * explicitly atomically readable. | |
1542 | */ | |
1543 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
1544 | << 20; | |
1545 | update_entity_load_avg(se, 0); | |
1546 | /* Indicate that we're now synchronized and on-rq */ | |
1547 | se->avg.decay_count = 0; | |
1548 | } | |
9ee474f5 PT |
1549 | wakeup = 0; |
1550 | } else { | |
1551 | __synchronize_entity_decay(se); | |
1552 | } | |
1553 | ||
aff3e498 PT |
1554 | /* migrated tasks did not contribute to our blocked load */ |
1555 | if (wakeup) { | |
9ee474f5 | 1556 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
1557 | update_entity_load_avg(se, 0); |
1558 | } | |
9ee474f5 | 1559 | |
2dac754e | 1560 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
1561 | /* we force update consideration on load-balancer moves */ |
1562 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
1563 | } |
1564 | ||
9ee474f5 PT |
1565 | /* |
1566 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
1567 | * transitioning to a blocked state we track its projected decay using | |
1568 | * blocked_load_avg. | |
1569 | */ | |
2dac754e | 1570 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1571 | struct sched_entity *se, |
1572 | int sleep) | |
2dac754e | 1573 | { |
9ee474f5 | 1574 | update_entity_load_avg(se, 1); |
aff3e498 PT |
1575 | /* we force update consideration on load-balancer moves */ |
1576 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 1577 | |
2dac754e | 1578 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
1579 | if (sleep) { |
1580 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
1581 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
1582 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 1583 | } |
642dbc39 VG |
1584 | |
1585 | /* | |
1586 | * Update the rq's load with the elapsed running time before entering | |
1587 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
1588 | * be the only way to update the runnable statistic. | |
1589 | */ | |
1590 | void idle_enter_fair(struct rq *this_rq) | |
1591 | { | |
1592 | update_rq_runnable_avg(this_rq, 1); | |
1593 | } | |
1594 | ||
1595 | /* | |
1596 | * Update the rq's load with the elapsed idle time before a task is | |
1597 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
1598 | * be the only way to update the runnable statistic. | |
1599 | */ | |
1600 | void idle_exit_fair(struct rq *this_rq) | |
1601 | { | |
1602 | update_rq_runnable_avg(this_rq, 0); | |
1603 | } | |
1604 | ||
9d85f21c | 1605 | #else |
9ee474f5 PT |
1606 | static inline void update_entity_load_avg(struct sched_entity *se, |
1607 | int update_cfs_rq) {} | |
18bf2805 | 1608 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 1609 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1610 | struct sched_entity *se, |
1611 | int wakeup) {} | |
2dac754e | 1612 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1613 | struct sched_entity *se, |
1614 | int sleep) {} | |
aff3e498 PT |
1615 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
1616 | int force_update) {} | |
9d85f21c PT |
1617 | #endif |
1618 | ||
2396af69 | 1619 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1620 | { |
bf0f6f24 | 1621 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1622 | struct task_struct *tsk = NULL; |
1623 | ||
1624 | if (entity_is_task(se)) | |
1625 | tsk = task_of(se); | |
1626 | ||
41acab88 | 1627 | if (se->statistics.sleep_start) { |
78becc27 | 1628 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
1629 | |
1630 | if ((s64)delta < 0) | |
1631 | delta = 0; | |
1632 | ||
41acab88 LDM |
1633 | if (unlikely(delta > se->statistics.sleep_max)) |
1634 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1635 | |
8c79a045 | 1636 | se->statistics.sleep_start = 0; |
41acab88 | 1637 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1638 | |
768d0c27 | 1639 | if (tsk) { |
e414314c | 1640 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1641 | trace_sched_stat_sleep(tsk, delta); |
1642 | } | |
bf0f6f24 | 1643 | } |
41acab88 | 1644 | if (se->statistics.block_start) { |
78becc27 | 1645 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
1646 | |
1647 | if ((s64)delta < 0) | |
1648 | delta = 0; | |
1649 | ||
41acab88 LDM |
1650 | if (unlikely(delta > se->statistics.block_max)) |
1651 | se->statistics.block_max = delta; | |
bf0f6f24 | 1652 | |
8c79a045 | 1653 | se->statistics.block_start = 0; |
41acab88 | 1654 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1655 | |
e414314c | 1656 | if (tsk) { |
8f0dfc34 | 1657 | if (tsk->in_iowait) { |
41acab88 LDM |
1658 | se->statistics.iowait_sum += delta; |
1659 | se->statistics.iowait_count++; | |
768d0c27 | 1660 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1661 | } |
1662 | ||
b781a602 AV |
1663 | trace_sched_stat_blocked(tsk, delta); |
1664 | ||
e414314c PZ |
1665 | /* |
1666 | * Blocking time is in units of nanosecs, so shift by | |
1667 | * 20 to get a milliseconds-range estimation of the | |
1668 | * amount of time that the task spent sleeping: | |
1669 | */ | |
1670 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1671 | profile_hits(SLEEP_PROFILING, | |
1672 | (void *)get_wchan(tsk), | |
1673 | delta >> 20); | |
1674 | } | |
1675 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1676 | } |
bf0f6f24 IM |
1677 | } |
1678 | #endif | |
1679 | } | |
1680 | ||
ddc97297 PZ |
1681 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1682 | { | |
1683 | #ifdef CONFIG_SCHED_DEBUG | |
1684 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
1685 | ||
1686 | if (d < 0) | |
1687 | d = -d; | |
1688 | ||
1689 | if (d > 3*sysctl_sched_latency) | |
1690 | schedstat_inc(cfs_rq, nr_spread_over); | |
1691 | #endif | |
1692 | } | |
1693 | ||
aeb73b04 PZ |
1694 | static void |
1695 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
1696 | { | |
1af5f730 | 1697 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 1698 | |
2cb8600e PZ |
1699 | /* |
1700 | * The 'current' period is already promised to the current tasks, | |
1701 | * however the extra weight of the new task will slow them down a | |
1702 | * little, place the new task so that it fits in the slot that | |
1703 | * stays open at the end. | |
1704 | */ | |
94dfb5e7 | 1705 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 1706 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 1707 | |
a2e7a7eb | 1708 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 1709 | if (!initial) { |
a2e7a7eb | 1710 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 1711 | |
a2e7a7eb MG |
1712 | /* |
1713 | * Halve their sleep time's effect, to allow | |
1714 | * for a gentler effect of sleepers: | |
1715 | */ | |
1716 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
1717 | thresh >>= 1; | |
51e0304c | 1718 | |
a2e7a7eb | 1719 | vruntime -= thresh; |
aeb73b04 PZ |
1720 | } |
1721 | ||
b5d9d734 | 1722 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 1723 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
1724 | } |
1725 | ||
d3d9dc33 PT |
1726 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
1727 | ||
bf0f6f24 | 1728 | static void |
88ec22d3 | 1729 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1730 | { |
88ec22d3 PZ |
1731 | /* |
1732 | * Update the normalized vruntime before updating min_vruntime | |
1733 | * through callig update_curr(). | |
1734 | */ | |
371fd7e7 | 1735 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
1736 | se->vruntime += cfs_rq->min_vruntime; |
1737 | ||
bf0f6f24 | 1738 | /* |
a2a2d680 | 1739 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1740 | */ |
b7cc0896 | 1741 | update_curr(cfs_rq); |
f269ae04 | 1742 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
1743 | account_entity_enqueue(cfs_rq, se); |
1744 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 1745 | |
88ec22d3 | 1746 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 1747 | place_entity(cfs_rq, se, 0); |
2396af69 | 1748 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 1749 | } |
bf0f6f24 | 1750 | |
d2417e5a | 1751 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 1752 | check_spread(cfs_rq, se); |
83b699ed SV |
1753 | if (se != cfs_rq->curr) |
1754 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1755 | se->on_rq = 1; |
3d4b47b4 | 1756 | |
d3d9dc33 | 1757 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1758 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1759 | check_enqueue_throttle(cfs_rq); |
1760 | } | |
bf0f6f24 IM |
1761 | } |
1762 | ||
2c13c919 | 1763 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1764 | { |
2c13c919 RR |
1765 | for_each_sched_entity(se) { |
1766 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1767 | if (cfs_rq->last == se) | |
1768 | cfs_rq->last = NULL; | |
1769 | else | |
1770 | break; | |
1771 | } | |
1772 | } | |
2002c695 | 1773 | |
2c13c919 RR |
1774 | static void __clear_buddies_next(struct sched_entity *se) |
1775 | { | |
1776 | for_each_sched_entity(se) { | |
1777 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1778 | if (cfs_rq->next == se) | |
1779 | cfs_rq->next = NULL; | |
1780 | else | |
1781 | break; | |
1782 | } | |
2002c695 PZ |
1783 | } |
1784 | ||
ac53db59 RR |
1785 | static void __clear_buddies_skip(struct sched_entity *se) |
1786 | { | |
1787 | for_each_sched_entity(se) { | |
1788 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1789 | if (cfs_rq->skip == se) | |
1790 | cfs_rq->skip = NULL; | |
1791 | else | |
1792 | break; | |
1793 | } | |
1794 | } | |
1795 | ||
a571bbea PZ |
1796 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1797 | { | |
2c13c919 RR |
1798 | if (cfs_rq->last == se) |
1799 | __clear_buddies_last(se); | |
1800 | ||
1801 | if (cfs_rq->next == se) | |
1802 | __clear_buddies_next(se); | |
ac53db59 RR |
1803 | |
1804 | if (cfs_rq->skip == se) | |
1805 | __clear_buddies_skip(se); | |
a571bbea PZ |
1806 | } |
1807 | ||
6c16a6dc | 1808 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 1809 | |
bf0f6f24 | 1810 | static void |
371fd7e7 | 1811 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1812 | { |
a2a2d680 DA |
1813 | /* |
1814 | * Update run-time statistics of the 'current'. | |
1815 | */ | |
1816 | update_curr(cfs_rq); | |
17bc14b7 | 1817 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 1818 | |
19b6a2e3 | 1819 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1820 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1821 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1822 | if (entity_is_task(se)) { |
1823 | struct task_struct *tsk = task_of(se); | |
1824 | ||
1825 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 1826 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 1827 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 1828 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 1829 | } |
db36cc7d | 1830 | #endif |
67e9fb2a PZ |
1831 | } |
1832 | ||
2002c695 | 1833 | clear_buddies(cfs_rq, se); |
4793241b | 1834 | |
83b699ed | 1835 | if (se != cfs_rq->curr) |
30cfdcfc | 1836 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 1837 | se->on_rq = 0; |
30cfdcfc | 1838 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1839 | |
1840 | /* | |
1841 | * Normalize the entity after updating the min_vruntime because the | |
1842 | * update can refer to the ->curr item and we need to reflect this | |
1843 | * movement in our normalized position. | |
1844 | */ | |
371fd7e7 | 1845 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1846 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 1847 | |
d8b4986d PT |
1848 | /* return excess runtime on last dequeue */ |
1849 | return_cfs_rq_runtime(cfs_rq); | |
1850 | ||
1e876231 | 1851 | update_min_vruntime(cfs_rq); |
17bc14b7 | 1852 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
1853 | } |
1854 | ||
1855 | /* | |
1856 | * Preempt the current task with a newly woken task if needed: | |
1857 | */ | |
7c92e54f | 1858 | static void |
2e09bf55 | 1859 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1860 | { |
11697830 | 1861 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
1862 | struct sched_entity *se; |
1863 | s64 delta; | |
11697830 | 1864 | |
6d0f0ebd | 1865 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 1866 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 1867 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 1868 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
1869 | /* |
1870 | * The current task ran long enough, ensure it doesn't get | |
1871 | * re-elected due to buddy favours. | |
1872 | */ | |
1873 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
1874 | return; |
1875 | } | |
1876 | ||
1877 | /* | |
1878 | * Ensure that a task that missed wakeup preemption by a | |
1879 | * narrow margin doesn't have to wait for a full slice. | |
1880 | * This also mitigates buddy induced latencies under load. | |
1881 | */ | |
f685ceac MG |
1882 | if (delta_exec < sysctl_sched_min_granularity) |
1883 | return; | |
1884 | ||
f4cfb33e WX |
1885 | se = __pick_first_entity(cfs_rq); |
1886 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 1887 | |
f4cfb33e WX |
1888 | if (delta < 0) |
1889 | return; | |
d7d82944 | 1890 | |
f4cfb33e WX |
1891 | if (delta > ideal_runtime) |
1892 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
1893 | } |
1894 | ||
83b699ed | 1895 | static void |
8494f412 | 1896 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1897 | { |
83b699ed SV |
1898 | /* 'current' is not kept within the tree. */ |
1899 | if (se->on_rq) { | |
1900 | /* | |
1901 | * Any task has to be enqueued before it get to execute on | |
1902 | * a CPU. So account for the time it spent waiting on the | |
1903 | * runqueue. | |
1904 | */ | |
1905 | update_stats_wait_end(cfs_rq, se); | |
1906 | __dequeue_entity(cfs_rq, se); | |
1907 | } | |
1908 | ||
79303e9e | 1909 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 1910 | cfs_rq->curr = se; |
eba1ed4b IM |
1911 | #ifdef CONFIG_SCHEDSTATS |
1912 | /* | |
1913 | * Track our maximum slice length, if the CPU's load is at | |
1914 | * least twice that of our own weight (i.e. dont track it | |
1915 | * when there are only lesser-weight tasks around): | |
1916 | */ | |
495eca49 | 1917 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 1918 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
1919 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
1920 | } | |
1921 | #endif | |
4a55b450 | 1922 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
1923 | } |
1924 | ||
3f3a4904 PZ |
1925 | static int |
1926 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
1927 | ||
ac53db59 RR |
1928 | /* |
1929 | * Pick the next process, keeping these things in mind, in this order: | |
1930 | * 1) keep things fair between processes/task groups | |
1931 | * 2) pick the "next" process, since someone really wants that to run | |
1932 | * 3) pick the "last" process, for cache locality | |
1933 | * 4) do not run the "skip" process, if something else is available | |
1934 | */ | |
f4b6755f | 1935 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 1936 | { |
ac53db59 | 1937 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 1938 | struct sched_entity *left = se; |
f4b6755f | 1939 | |
ac53db59 RR |
1940 | /* |
1941 | * Avoid running the skip buddy, if running something else can | |
1942 | * be done without getting too unfair. | |
1943 | */ | |
1944 | if (cfs_rq->skip == se) { | |
1945 | struct sched_entity *second = __pick_next_entity(se); | |
1946 | if (second && wakeup_preempt_entity(second, left) < 1) | |
1947 | se = second; | |
1948 | } | |
aa2ac252 | 1949 | |
f685ceac MG |
1950 | /* |
1951 | * Prefer last buddy, try to return the CPU to a preempted task. | |
1952 | */ | |
1953 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
1954 | se = cfs_rq->last; | |
1955 | ||
ac53db59 RR |
1956 | /* |
1957 | * Someone really wants this to run. If it's not unfair, run it. | |
1958 | */ | |
1959 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
1960 | se = cfs_rq->next; | |
1961 | ||
f685ceac | 1962 | clear_buddies(cfs_rq, se); |
4793241b PZ |
1963 | |
1964 | return se; | |
aa2ac252 PZ |
1965 | } |
1966 | ||
d3d9dc33 PT |
1967 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
1968 | ||
ab6cde26 | 1969 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
1970 | { |
1971 | /* | |
1972 | * If still on the runqueue then deactivate_task() | |
1973 | * was not called and update_curr() has to be done: | |
1974 | */ | |
1975 | if (prev->on_rq) | |
b7cc0896 | 1976 | update_curr(cfs_rq); |
bf0f6f24 | 1977 | |
d3d9dc33 PT |
1978 | /* throttle cfs_rqs exceeding runtime */ |
1979 | check_cfs_rq_runtime(cfs_rq); | |
1980 | ||
ddc97297 | 1981 | check_spread(cfs_rq, prev); |
30cfdcfc | 1982 | if (prev->on_rq) { |
5870db5b | 1983 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
1984 | /* Put 'current' back into the tree. */ |
1985 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 1986 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 1987 | update_entity_load_avg(prev, 1); |
30cfdcfc | 1988 | } |
429d43bc | 1989 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
1990 | } |
1991 | ||
8f4d37ec PZ |
1992 | static void |
1993 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 1994 | { |
bf0f6f24 | 1995 | /* |
30cfdcfc | 1996 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1997 | */ |
30cfdcfc | 1998 | update_curr(cfs_rq); |
bf0f6f24 | 1999 | |
9d85f21c PT |
2000 | /* |
2001 | * Ensure that runnable average is periodically updated. | |
2002 | */ | |
9ee474f5 | 2003 | update_entity_load_avg(curr, 1); |
aff3e498 | 2004 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9d85f21c | 2005 | |
8f4d37ec PZ |
2006 | #ifdef CONFIG_SCHED_HRTICK |
2007 | /* | |
2008 | * queued ticks are scheduled to match the slice, so don't bother | |
2009 | * validating it and just reschedule. | |
2010 | */ | |
983ed7a6 HH |
2011 | if (queued) { |
2012 | resched_task(rq_of(cfs_rq)->curr); | |
2013 | return; | |
2014 | } | |
8f4d37ec PZ |
2015 | /* |
2016 | * don't let the period tick interfere with the hrtick preemption | |
2017 | */ | |
2018 | if (!sched_feat(DOUBLE_TICK) && | |
2019 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2020 | return; | |
2021 | #endif | |
2022 | ||
2c2efaed | 2023 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2024 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2025 | } |
2026 | ||
ab84d31e PT |
2027 | |
2028 | /************************************************** | |
2029 | * CFS bandwidth control machinery | |
2030 | */ | |
2031 | ||
2032 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2033 | |
2034 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2035 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2036 | |
2037 | static inline bool cfs_bandwidth_used(void) | |
2038 | { | |
c5905afb | 2039 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2040 | } |
2041 | ||
2042 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
2043 | { | |
2044 | /* only need to count groups transitioning between enabled/!enabled */ | |
2045 | if (enabled && !was_enabled) | |
c5905afb | 2046 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 2047 | else if (!enabled && was_enabled) |
c5905afb | 2048 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
2049 | } |
2050 | #else /* HAVE_JUMP_LABEL */ | |
2051 | static bool cfs_bandwidth_used(void) | |
2052 | { | |
2053 | return true; | |
2054 | } | |
2055 | ||
2056 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
2057 | #endif /* HAVE_JUMP_LABEL */ | |
2058 | ||
ab84d31e PT |
2059 | /* |
2060 | * default period for cfs group bandwidth. | |
2061 | * default: 0.1s, units: nanoseconds | |
2062 | */ | |
2063 | static inline u64 default_cfs_period(void) | |
2064 | { | |
2065 | return 100000000ULL; | |
2066 | } | |
ec12cb7f PT |
2067 | |
2068 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2069 | { | |
2070 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2071 | } | |
2072 | ||
a9cf55b2 PT |
2073 | /* |
2074 | * Replenish runtime according to assigned quota and update expiration time. | |
2075 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2076 | * additional synchronization around rq->lock. | |
2077 | * | |
2078 | * requires cfs_b->lock | |
2079 | */ | |
029632fb | 2080 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2081 | { |
2082 | u64 now; | |
2083 | ||
2084 | if (cfs_b->quota == RUNTIME_INF) | |
2085 | return; | |
2086 | ||
2087 | now = sched_clock_cpu(smp_processor_id()); | |
2088 | cfs_b->runtime = cfs_b->quota; | |
2089 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2090 | } | |
2091 | ||
029632fb PZ |
2092 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2093 | { | |
2094 | return &tg->cfs_bandwidth; | |
2095 | } | |
2096 | ||
f1b17280 PT |
2097 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2098 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2099 | { | |
2100 | if (unlikely(cfs_rq->throttle_count)) | |
2101 | return cfs_rq->throttled_clock_task; | |
2102 | ||
78becc27 | 2103 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2104 | } |
2105 | ||
85dac906 PT |
2106 | /* returns 0 on failure to allocate runtime */ |
2107 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2108 | { |
2109 | struct task_group *tg = cfs_rq->tg; | |
2110 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2111 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2112 | |
2113 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2114 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2115 | ||
2116 | raw_spin_lock(&cfs_b->lock); | |
2117 | if (cfs_b->quota == RUNTIME_INF) | |
2118 | amount = min_amount; | |
58088ad0 | 2119 | else { |
a9cf55b2 PT |
2120 | /* |
2121 | * If the bandwidth pool has become inactive, then at least one | |
2122 | * period must have elapsed since the last consumption. | |
2123 | * Refresh the global state and ensure bandwidth timer becomes | |
2124 | * active. | |
2125 | */ | |
2126 | if (!cfs_b->timer_active) { | |
2127 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2128 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2129 | } |
58088ad0 PT |
2130 | |
2131 | if (cfs_b->runtime > 0) { | |
2132 | amount = min(cfs_b->runtime, min_amount); | |
2133 | cfs_b->runtime -= amount; | |
2134 | cfs_b->idle = 0; | |
2135 | } | |
ec12cb7f | 2136 | } |
a9cf55b2 | 2137 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2138 | raw_spin_unlock(&cfs_b->lock); |
2139 | ||
2140 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2141 | /* |
2142 | * we may have advanced our local expiration to account for allowed | |
2143 | * spread between our sched_clock and the one on which runtime was | |
2144 | * issued. | |
2145 | */ | |
2146 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2147 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2148 | |
2149 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2150 | } |
2151 | ||
a9cf55b2 PT |
2152 | /* |
2153 | * Note: This depends on the synchronization provided by sched_clock and the | |
2154 | * fact that rq->clock snapshots this value. | |
2155 | */ | |
2156 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2157 | { |
a9cf55b2 | 2158 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2159 | |
2160 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2161 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2162 | return; |
2163 | ||
a9cf55b2 PT |
2164 | if (cfs_rq->runtime_remaining < 0) |
2165 | return; | |
2166 | ||
2167 | /* | |
2168 | * If the local deadline has passed we have to consider the | |
2169 | * possibility that our sched_clock is 'fast' and the global deadline | |
2170 | * has not truly expired. | |
2171 | * | |
2172 | * Fortunately we can check determine whether this the case by checking | |
2173 | * whether the global deadline has advanced. | |
2174 | */ | |
2175 | ||
2176 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
2177 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
2178 | cfs_rq->runtime_expires += TICK_NSEC; | |
2179 | } else { | |
2180 | /* global deadline is ahead, expiration has passed */ | |
2181 | cfs_rq->runtime_remaining = 0; | |
2182 | } | |
2183 | } | |
2184 | ||
2185 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2186 | unsigned long delta_exec) | |
2187 | { | |
2188 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 2189 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
2190 | expire_cfs_rq_runtime(cfs_rq); |
2191 | ||
2192 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
2193 | return; |
2194 | ||
85dac906 PT |
2195 | /* |
2196 | * if we're unable to extend our runtime we resched so that the active | |
2197 | * hierarchy can be throttled | |
2198 | */ | |
2199 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
2200 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
2201 | } |
2202 | ||
6c16a6dc PZ |
2203 | static __always_inline |
2204 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 2205 | { |
56f570e5 | 2206 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
2207 | return; |
2208 | ||
2209 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
2210 | } | |
2211 | ||
85dac906 PT |
2212 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
2213 | { | |
56f570e5 | 2214 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
2215 | } |
2216 | ||
64660c86 PT |
2217 | /* check whether cfs_rq, or any parent, is throttled */ |
2218 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2219 | { | |
56f570e5 | 2220 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
2221 | } |
2222 | ||
2223 | /* | |
2224 | * Ensure that neither of the group entities corresponding to src_cpu or | |
2225 | * dest_cpu are members of a throttled hierarchy when performing group | |
2226 | * load-balance operations. | |
2227 | */ | |
2228 | static inline int throttled_lb_pair(struct task_group *tg, | |
2229 | int src_cpu, int dest_cpu) | |
2230 | { | |
2231 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
2232 | ||
2233 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
2234 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
2235 | ||
2236 | return throttled_hierarchy(src_cfs_rq) || | |
2237 | throttled_hierarchy(dest_cfs_rq); | |
2238 | } | |
2239 | ||
2240 | /* updated child weight may affect parent so we have to do this bottom up */ | |
2241 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
2242 | { | |
2243 | struct rq *rq = data; | |
2244 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2245 | ||
2246 | cfs_rq->throttle_count--; | |
2247 | #ifdef CONFIG_SMP | |
2248 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 2249 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 2250 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 2251 | cfs_rq->throttled_clock_task; |
64660c86 PT |
2252 | } |
2253 | #endif | |
2254 | ||
2255 | return 0; | |
2256 | } | |
2257 | ||
2258 | static int tg_throttle_down(struct task_group *tg, void *data) | |
2259 | { | |
2260 | struct rq *rq = data; | |
2261 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2262 | ||
82958366 PT |
2263 | /* group is entering throttled state, stop time */ |
2264 | if (!cfs_rq->throttle_count) | |
78becc27 | 2265 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
2266 | cfs_rq->throttle_count++; |
2267 | ||
2268 | return 0; | |
2269 | } | |
2270 | ||
d3d9dc33 | 2271 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
2272 | { |
2273 | struct rq *rq = rq_of(cfs_rq); | |
2274 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2275 | struct sched_entity *se; | |
2276 | long task_delta, dequeue = 1; | |
2277 | ||
2278 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2279 | ||
f1b17280 | 2280 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
2281 | rcu_read_lock(); |
2282 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
2283 | rcu_read_unlock(); | |
85dac906 PT |
2284 | |
2285 | task_delta = cfs_rq->h_nr_running; | |
2286 | for_each_sched_entity(se) { | |
2287 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
2288 | /* throttled entity or throttle-on-deactivate */ | |
2289 | if (!se->on_rq) | |
2290 | break; | |
2291 | ||
2292 | if (dequeue) | |
2293 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
2294 | qcfs_rq->h_nr_running -= task_delta; | |
2295 | ||
2296 | if (qcfs_rq->load.weight) | |
2297 | dequeue = 0; | |
2298 | } | |
2299 | ||
2300 | if (!se) | |
2301 | rq->nr_running -= task_delta; | |
2302 | ||
2303 | cfs_rq->throttled = 1; | |
78becc27 | 2304 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
2305 | raw_spin_lock(&cfs_b->lock); |
2306 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
2307 | raw_spin_unlock(&cfs_b->lock); | |
2308 | } | |
2309 | ||
029632fb | 2310 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
2311 | { |
2312 | struct rq *rq = rq_of(cfs_rq); | |
2313 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2314 | struct sched_entity *se; | |
2315 | int enqueue = 1; | |
2316 | long task_delta; | |
2317 | ||
2318 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2319 | ||
2320 | cfs_rq->throttled = 0; | |
1a55af2e FW |
2321 | |
2322 | update_rq_clock(rq); | |
2323 | ||
671fd9da | 2324 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 2325 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
2326 | list_del_rcu(&cfs_rq->throttled_list); |
2327 | raw_spin_unlock(&cfs_b->lock); | |
2328 | ||
64660c86 PT |
2329 | /* update hierarchical throttle state */ |
2330 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
2331 | ||
671fd9da PT |
2332 | if (!cfs_rq->load.weight) |
2333 | return; | |
2334 | ||
2335 | task_delta = cfs_rq->h_nr_running; | |
2336 | for_each_sched_entity(se) { | |
2337 | if (se->on_rq) | |
2338 | enqueue = 0; | |
2339 | ||
2340 | cfs_rq = cfs_rq_of(se); | |
2341 | if (enqueue) | |
2342 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
2343 | cfs_rq->h_nr_running += task_delta; | |
2344 | ||
2345 | if (cfs_rq_throttled(cfs_rq)) | |
2346 | break; | |
2347 | } | |
2348 | ||
2349 | if (!se) | |
2350 | rq->nr_running += task_delta; | |
2351 | ||
2352 | /* determine whether we need to wake up potentially idle cpu */ | |
2353 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
2354 | resched_task(rq->curr); | |
2355 | } | |
2356 | ||
2357 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
2358 | u64 remaining, u64 expires) | |
2359 | { | |
2360 | struct cfs_rq *cfs_rq; | |
2361 | u64 runtime = remaining; | |
2362 | ||
2363 | rcu_read_lock(); | |
2364 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
2365 | throttled_list) { | |
2366 | struct rq *rq = rq_of(cfs_rq); | |
2367 | ||
2368 | raw_spin_lock(&rq->lock); | |
2369 | if (!cfs_rq_throttled(cfs_rq)) | |
2370 | goto next; | |
2371 | ||
2372 | runtime = -cfs_rq->runtime_remaining + 1; | |
2373 | if (runtime > remaining) | |
2374 | runtime = remaining; | |
2375 | remaining -= runtime; | |
2376 | ||
2377 | cfs_rq->runtime_remaining += runtime; | |
2378 | cfs_rq->runtime_expires = expires; | |
2379 | ||
2380 | /* we check whether we're throttled above */ | |
2381 | if (cfs_rq->runtime_remaining > 0) | |
2382 | unthrottle_cfs_rq(cfs_rq); | |
2383 | ||
2384 | next: | |
2385 | raw_spin_unlock(&rq->lock); | |
2386 | ||
2387 | if (!remaining) | |
2388 | break; | |
2389 | } | |
2390 | rcu_read_unlock(); | |
2391 | ||
2392 | return remaining; | |
2393 | } | |
2394 | ||
58088ad0 PT |
2395 | /* |
2396 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2397 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2398 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2399 | * used to track this state. | |
2400 | */ | |
2401 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
2402 | { | |
671fd9da PT |
2403 | u64 runtime, runtime_expires; |
2404 | int idle = 1, throttled; | |
58088ad0 PT |
2405 | |
2406 | raw_spin_lock(&cfs_b->lock); | |
2407 | /* no need to continue the timer with no bandwidth constraint */ | |
2408 | if (cfs_b->quota == RUNTIME_INF) | |
2409 | goto out_unlock; | |
2410 | ||
671fd9da PT |
2411 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
2412 | /* idle depends on !throttled (for the case of a large deficit) */ | |
2413 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 2414 | cfs_b->nr_periods += overrun; |
671fd9da | 2415 | |
a9cf55b2 PT |
2416 | /* if we're going inactive then everything else can be deferred */ |
2417 | if (idle) | |
2418 | goto out_unlock; | |
2419 | ||
2420 | __refill_cfs_bandwidth_runtime(cfs_b); | |
2421 | ||
671fd9da PT |
2422 | if (!throttled) { |
2423 | /* mark as potentially idle for the upcoming period */ | |
2424 | cfs_b->idle = 1; | |
2425 | goto out_unlock; | |
2426 | } | |
2427 | ||
e8da1b18 NR |
2428 | /* account preceding periods in which throttling occurred */ |
2429 | cfs_b->nr_throttled += overrun; | |
2430 | ||
671fd9da PT |
2431 | /* |
2432 | * There are throttled entities so we must first use the new bandwidth | |
2433 | * to unthrottle them before making it generally available. This | |
2434 | * ensures that all existing debts will be paid before a new cfs_rq is | |
2435 | * allowed to run. | |
2436 | */ | |
2437 | runtime = cfs_b->runtime; | |
2438 | runtime_expires = cfs_b->runtime_expires; | |
2439 | cfs_b->runtime = 0; | |
2440 | ||
2441 | /* | |
2442 | * This check is repeated as we are holding onto the new bandwidth | |
2443 | * while we unthrottle. This can potentially race with an unthrottled | |
2444 | * group trying to acquire new bandwidth from the global pool. | |
2445 | */ | |
2446 | while (throttled && runtime > 0) { | |
2447 | raw_spin_unlock(&cfs_b->lock); | |
2448 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
2449 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
2450 | runtime_expires); | |
2451 | raw_spin_lock(&cfs_b->lock); | |
2452 | ||
2453 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
2454 | } | |
58088ad0 | 2455 | |
671fd9da PT |
2456 | /* return (any) remaining runtime */ |
2457 | cfs_b->runtime = runtime; | |
2458 | /* | |
2459 | * While we are ensured activity in the period following an | |
2460 | * unthrottle, this also covers the case in which the new bandwidth is | |
2461 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
2462 | * timer to remain active while there are any throttled entities.) | |
2463 | */ | |
2464 | cfs_b->idle = 0; | |
58088ad0 PT |
2465 | out_unlock: |
2466 | if (idle) | |
2467 | cfs_b->timer_active = 0; | |
2468 | raw_spin_unlock(&cfs_b->lock); | |
2469 | ||
2470 | return idle; | |
2471 | } | |
d3d9dc33 | 2472 | |
d8b4986d PT |
2473 | /* a cfs_rq won't donate quota below this amount */ |
2474 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
2475 | /* minimum remaining period time to redistribute slack quota */ | |
2476 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
2477 | /* how long we wait to gather additional slack before distributing */ | |
2478 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
2479 | ||
2480 | /* are we near the end of the current quota period? */ | |
2481 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
2482 | { | |
2483 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
2484 | u64 remaining; | |
2485 | ||
2486 | /* if the call-back is running a quota refresh is already occurring */ | |
2487 | if (hrtimer_callback_running(refresh_timer)) | |
2488 | return 1; | |
2489 | ||
2490 | /* is a quota refresh about to occur? */ | |
2491 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
2492 | if (remaining < min_expire) | |
2493 | return 1; | |
2494 | ||
2495 | return 0; | |
2496 | } | |
2497 | ||
2498 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
2499 | { | |
2500 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
2501 | ||
2502 | /* if there's a quota refresh soon don't bother with slack */ | |
2503 | if (runtime_refresh_within(cfs_b, min_left)) | |
2504 | return; | |
2505 | ||
2506 | start_bandwidth_timer(&cfs_b->slack_timer, | |
2507 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
2508 | } | |
2509 | ||
2510 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
2511 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2512 | { | |
2513 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2514 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
2515 | ||
2516 | if (slack_runtime <= 0) | |
2517 | return; | |
2518 | ||
2519 | raw_spin_lock(&cfs_b->lock); | |
2520 | if (cfs_b->quota != RUNTIME_INF && | |
2521 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
2522 | cfs_b->runtime += slack_runtime; | |
2523 | ||
2524 | /* we are under rq->lock, defer unthrottling using a timer */ | |
2525 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
2526 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
2527 | start_cfs_slack_bandwidth(cfs_b); | |
2528 | } | |
2529 | raw_spin_unlock(&cfs_b->lock); | |
2530 | ||
2531 | /* even if it's not valid for return we don't want to try again */ | |
2532 | cfs_rq->runtime_remaining -= slack_runtime; | |
2533 | } | |
2534 | ||
2535 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2536 | { | |
56f570e5 PT |
2537 | if (!cfs_bandwidth_used()) |
2538 | return; | |
2539 | ||
fccfdc6f | 2540 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2541 | return; |
2542 | ||
2543 | __return_cfs_rq_runtime(cfs_rq); | |
2544 | } | |
2545 | ||
2546 | /* | |
2547 | * This is done with a timer (instead of inline with bandwidth return) since | |
2548 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2549 | */ | |
2550 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2551 | { | |
2552 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2553 | u64 expires; | |
2554 | ||
2555 | /* confirm we're still not at a refresh boundary */ | |
2556 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2557 | return; | |
2558 | ||
2559 | raw_spin_lock(&cfs_b->lock); | |
2560 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2561 | runtime = cfs_b->runtime; | |
2562 | cfs_b->runtime = 0; | |
2563 | } | |
2564 | expires = cfs_b->runtime_expires; | |
2565 | raw_spin_unlock(&cfs_b->lock); | |
2566 | ||
2567 | if (!runtime) | |
2568 | return; | |
2569 | ||
2570 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2571 | ||
2572 | raw_spin_lock(&cfs_b->lock); | |
2573 | if (expires == cfs_b->runtime_expires) | |
2574 | cfs_b->runtime = runtime; | |
2575 | raw_spin_unlock(&cfs_b->lock); | |
2576 | } | |
2577 | ||
d3d9dc33 PT |
2578 | /* |
2579 | * When a group wakes up we want to make sure that its quota is not already | |
2580 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2581 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2582 | */ | |
2583 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2584 | { | |
56f570e5 PT |
2585 | if (!cfs_bandwidth_used()) |
2586 | return; | |
2587 | ||
d3d9dc33 PT |
2588 | /* an active group must be handled by the update_curr()->put() path */ |
2589 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2590 | return; | |
2591 | ||
2592 | /* ensure the group is not already throttled */ | |
2593 | if (cfs_rq_throttled(cfs_rq)) | |
2594 | return; | |
2595 | ||
2596 | /* update runtime allocation */ | |
2597 | account_cfs_rq_runtime(cfs_rq, 0); | |
2598 | if (cfs_rq->runtime_remaining <= 0) | |
2599 | throttle_cfs_rq(cfs_rq); | |
2600 | } | |
2601 | ||
2602 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2603 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2604 | { | |
56f570e5 PT |
2605 | if (!cfs_bandwidth_used()) |
2606 | return; | |
2607 | ||
d3d9dc33 PT |
2608 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2609 | return; | |
2610 | ||
2611 | /* | |
2612 | * it's possible for a throttled entity to be forced into a running | |
2613 | * state (e.g. set_curr_task), in this case we're finished. | |
2614 | */ | |
2615 | if (cfs_rq_throttled(cfs_rq)) | |
2616 | return; | |
2617 | ||
2618 | throttle_cfs_rq(cfs_rq); | |
2619 | } | |
029632fb PZ |
2620 | |
2621 | static inline u64 default_cfs_period(void); | |
2622 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun); | |
2623 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b); | |
2624 | ||
2625 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) | |
2626 | { | |
2627 | struct cfs_bandwidth *cfs_b = | |
2628 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2629 | do_sched_cfs_slack_timer(cfs_b); | |
2630 | ||
2631 | return HRTIMER_NORESTART; | |
2632 | } | |
2633 | ||
2634 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2635 | { | |
2636 | struct cfs_bandwidth *cfs_b = | |
2637 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2638 | ktime_t now; | |
2639 | int overrun; | |
2640 | int idle = 0; | |
2641 | ||
2642 | for (;;) { | |
2643 | now = hrtimer_cb_get_time(timer); | |
2644 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2645 | ||
2646 | if (!overrun) | |
2647 | break; | |
2648 | ||
2649 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2650 | } | |
2651 | ||
2652 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2653 | } | |
2654 | ||
2655 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2656 | { | |
2657 | raw_spin_lock_init(&cfs_b->lock); | |
2658 | cfs_b->runtime = 0; | |
2659 | cfs_b->quota = RUNTIME_INF; | |
2660 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2661 | ||
2662 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2663 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2664 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2665 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2666 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2667 | } | |
2668 | ||
2669 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2670 | { | |
2671 | cfs_rq->runtime_enabled = 0; | |
2672 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2673 | } | |
2674 | ||
2675 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2676 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2677 | { | |
2678 | /* | |
2679 | * The timer may be active because we're trying to set a new bandwidth | |
2680 | * period or because we're racing with the tear-down path | |
2681 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2682 | * terminates). In either case we ensure that it's re-programmed | |
2683 | */ | |
2684 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2685 | raw_spin_unlock(&cfs_b->lock); | |
2686 | /* ensure cfs_b->lock is available while we wait */ | |
2687 | hrtimer_cancel(&cfs_b->period_timer); | |
2688 | ||
2689 | raw_spin_lock(&cfs_b->lock); | |
2690 | /* if someone else restarted the timer then we're done */ | |
2691 | if (cfs_b->timer_active) | |
2692 | return; | |
2693 | } | |
2694 | ||
2695 | cfs_b->timer_active = 1; | |
2696 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
2697 | } | |
2698 | ||
2699 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2700 | { | |
2701 | hrtimer_cancel(&cfs_b->period_timer); | |
2702 | hrtimer_cancel(&cfs_b->slack_timer); | |
2703 | } | |
2704 | ||
38dc3348 | 2705 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
2706 | { |
2707 | struct cfs_rq *cfs_rq; | |
2708 | ||
2709 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2710 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2711 | ||
2712 | if (!cfs_rq->runtime_enabled) | |
2713 | continue; | |
2714 | ||
2715 | /* | |
2716 | * clock_task is not advancing so we just need to make sure | |
2717 | * there's some valid quota amount | |
2718 | */ | |
2719 | cfs_rq->runtime_remaining = cfs_b->quota; | |
2720 | if (cfs_rq_throttled(cfs_rq)) | |
2721 | unthrottle_cfs_rq(cfs_rq); | |
2722 | } | |
2723 | } | |
2724 | ||
2725 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
2726 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
2727 | { | |
78becc27 | 2728 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
2729 | } |
2730 | ||
2731 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2732 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
2733 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
2734 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 2735 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
2736 | |
2737 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
2738 | { | |
2739 | return 0; | |
2740 | } | |
64660c86 PT |
2741 | |
2742 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2743 | { | |
2744 | return 0; | |
2745 | } | |
2746 | ||
2747 | static inline int throttled_lb_pair(struct task_group *tg, | |
2748 | int src_cpu, int dest_cpu) | |
2749 | { | |
2750 | return 0; | |
2751 | } | |
029632fb PZ |
2752 | |
2753 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
2754 | ||
2755 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2756 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
2757 | #endif |
2758 | ||
029632fb PZ |
2759 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2760 | { | |
2761 | return NULL; | |
2762 | } | |
2763 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 2764 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
2765 | |
2766 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
2767 | ||
bf0f6f24 IM |
2768 | /************************************************** |
2769 | * CFS operations on tasks: | |
2770 | */ | |
2771 | ||
8f4d37ec PZ |
2772 | #ifdef CONFIG_SCHED_HRTICK |
2773 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2774 | { | |
8f4d37ec PZ |
2775 | struct sched_entity *se = &p->se; |
2776 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2777 | ||
2778 | WARN_ON(task_rq(p) != rq); | |
2779 | ||
b39e66ea | 2780 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
2781 | u64 slice = sched_slice(cfs_rq, se); |
2782 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
2783 | s64 delta = slice - ran; | |
2784 | ||
2785 | if (delta < 0) { | |
2786 | if (rq->curr == p) | |
2787 | resched_task(p); | |
2788 | return; | |
2789 | } | |
2790 | ||
2791 | /* | |
2792 | * Don't schedule slices shorter than 10000ns, that just | |
2793 | * doesn't make sense. Rely on vruntime for fairness. | |
2794 | */ | |
31656519 | 2795 | if (rq->curr != p) |
157124c1 | 2796 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 2797 | |
31656519 | 2798 | hrtick_start(rq, delta); |
8f4d37ec PZ |
2799 | } |
2800 | } | |
a4c2f00f PZ |
2801 | |
2802 | /* | |
2803 | * called from enqueue/dequeue and updates the hrtick when the | |
2804 | * current task is from our class and nr_running is low enough | |
2805 | * to matter. | |
2806 | */ | |
2807 | static void hrtick_update(struct rq *rq) | |
2808 | { | |
2809 | struct task_struct *curr = rq->curr; | |
2810 | ||
b39e66ea | 2811 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
2812 | return; |
2813 | ||
2814 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
2815 | hrtick_start_fair(rq, curr); | |
2816 | } | |
55e12e5e | 2817 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
2818 | static inline void |
2819 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2820 | { | |
2821 | } | |
a4c2f00f PZ |
2822 | |
2823 | static inline void hrtick_update(struct rq *rq) | |
2824 | { | |
2825 | } | |
8f4d37ec PZ |
2826 | #endif |
2827 | ||
bf0f6f24 IM |
2828 | /* |
2829 | * The enqueue_task method is called before nr_running is | |
2830 | * increased. Here we update the fair scheduling stats and | |
2831 | * then put the task into the rbtree: | |
2832 | */ | |
ea87bb78 | 2833 | static void |
371fd7e7 | 2834 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2835 | { |
2836 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2837 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
2838 | |
2839 | for_each_sched_entity(se) { | |
62fb1851 | 2840 | if (se->on_rq) |
bf0f6f24 IM |
2841 | break; |
2842 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 2843 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
2844 | |
2845 | /* | |
2846 | * end evaluation on encountering a throttled cfs_rq | |
2847 | * | |
2848 | * note: in the case of encountering a throttled cfs_rq we will | |
2849 | * post the final h_nr_running increment below. | |
2850 | */ | |
2851 | if (cfs_rq_throttled(cfs_rq)) | |
2852 | break; | |
953bfcd1 | 2853 | cfs_rq->h_nr_running++; |
85dac906 | 2854 | |
88ec22d3 | 2855 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 2856 | } |
8f4d37ec | 2857 | |
2069dd75 | 2858 | for_each_sched_entity(se) { |
0f317143 | 2859 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2860 | cfs_rq->h_nr_running++; |
2069dd75 | 2861 | |
85dac906 PT |
2862 | if (cfs_rq_throttled(cfs_rq)) |
2863 | break; | |
2864 | ||
17bc14b7 | 2865 | update_cfs_shares(cfs_rq); |
9ee474f5 | 2866 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
2867 | } |
2868 | ||
18bf2805 BS |
2869 | if (!se) { |
2870 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 2871 | inc_nr_running(rq); |
18bf2805 | 2872 | } |
a4c2f00f | 2873 | hrtick_update(rq); |
bf0f6f24 IM |
2874 | } |
2875 | ||
2f36825b VP |
2876 | static void set_next_buddy(struct sched_entity *se); |
2877 | ||
bf0f6f24 IM |
2878 | /* |
2879 | * The dequeue_task method is called before nr_running is | |
2880 | * decreased. We remove the task from the rbtree and | |
2881 | * update the fair scheduling stats: | |
2882 | */ | |
371fd7e7 | 2883 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2884 | { |
2885 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2886 | struct sched_entity *se = &p->se; |
2f36825b | 2887 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
2888 | |
2889 | for_each_sched_entity(se) { | |
2890 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 2891 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
2892 | |
2893 | /* | |
2894 | * end evaluation on encountering a throttled cfs_rq | |
2895 | * | |
2896 | * note: in the case of encountering a throttled cfs_rq we will | |
2897 | * post the final h_nr_running decrement below. | |
2898 | */ | |
2899 | if (cfs_rq_throttled(cfs_rq)) | |
2900 | break; | |
953bfcd1 | 2901 | cfs_rq->h_nr_running--; |
2069dd75 | 2902 | |
bf0f6f24 | 2903 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
2904 | if (cfs_rq->load.weight) { |
2905 | /* | |
2906 | * Bias pick_next to pick a task from this cfs_rq, as | |
2907 | * p is sleeping when it is within its sched_slice. | |
2908 | */ | |
2909 | if (task_sleep && parent_entity(se)) | |
2910 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
2911 | |
2912 | /* avoid re-evaluating load for this entity */ | |
2913 | se = parent_entity(se); | |
bf0f6f24 | 2914 | break; |
2f36825b | 2915 | } |
371fd7e7 | 2916 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 2917 | } |
8f4d37ec | 2918 | |
2069dd75 | 2919 | for_each_sched_entity(se) { |
0f317143 | 2920 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2921 | cfs_rq->h_nr_running--; |
2069dd75 | 2922 | |
85dac906 PT |
2923 | if (cfs_rq_throttled(cfs_rq)) |
2924 | break; | |
2925 | ||
17bc14b7 | 2926 | update_cfs_shares(cfs_rq); |
9ee474f5 | 2927 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
2928 | } |
2929 | ||
18bf2805 | 2930 | if (!se) { |
85dac906 | 2931 | dec_nr_running(rq); |
18bf2805 BS |
2932 | update_rq_runnable_avg(rq, 1); |
2933 | } | |
a4c2f00f | 2934 | hrtick_update(rq); |
bf0f6f24 IM |
2935 | } |
2936 | ||
e7693a36 | 2937 | #ifdef CONFIG_SMP |
029632fb PZ |
2938 | /* Used instead of source_load when we know the type == 0 */ |
2939 | static unsigned long weighted_cpuload(const int cpu) | |
2940 | { | |
2941 | return cpu_rq(cpu)->load.weight; | |
2942 | } | |
2943 | ||
2944 | /* | |
2945 | * Return a low guess at the load of a migration-source cpu weighted | |
2946 | * according to the scheduling class and "nice" value. | |
2947 | * | |
2948 | * We want to under-estimate the load of migration sources, to | |
2949 | * balance conservatively. | |
2950 | */ | |
2951 | static unsigned long source_load(int cpu, int type) | |
2952 | { | |
2953 | struct rq *rq = cpu_rq(cpu); | |
2954 | unsigned long total = weighted_cpuload(cpu); | |
2955 | ||
2956 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2957 | return total; | |
2958 | ||
2959 | return min(rq->cpu_load[type-1], total); | |
2960 | } | |
2961 | ||
2962 | /* | |
2963 | * Return a high guess at the load of a migration-target cpu weighted | |
2964 | * according to the scheduling class and "nice" value. | |
2965 | */ | |
2966 | static unsigned long target_load(int cpu, int type) | |
2967 | { | |
2968 | struct rq *rq = cpu_rq(cpu); | |
2969 | unsigned long total = weighted_cpuload(cpu); | |
2970 | ||
2971 | if (type == 0 || !sched_feat(LB_BIAS)) | |
2972 | return total; | |
2973 | ||
2974 | return max(rq->cpu_load[type-1], total); | |
2975 | } | |
2976 | ||
2977 | static unsigned long power_of(int cpu) | |
2978 | { | |
2979 | return cpu_rq(cpu)->cpu_power; | |
2980 | } | |
2981 | ||
2982 | static unsigned long cpu_avg_load_per_task(int cpu) | |
2983 | { | |
2984 | struct rq *rq = cpu_rq(cpu); | |
2985 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
2986 | ||
2987 | if (nr_running) | |
2988 | return rq->load.weight / nr_running; | |
2989 | ||
2990 | return 0; | |
2991 | } | |
2992 | ||
098fb9db | 2993 | |
74f8e4b2 | 2994 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
2995 | { |
2996 | struct sched_entity *se = &p->se; | |
2997 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
2998 | u64 min_vruntime; |
2999 | ||
3000 | #ifndef CONFIG_64BIT | |
3001 | u64 min_vruntime_copy; | |
88ec22d3 | 3002 | |
3fe1698b PZ |
3003 | do { |
3004 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3005 | smp_rmb(); | |
3006 | min_vruntime = cfs_rq->min_vruntime; | |
3007 | } while (min_vruntime != min_vruntime_copy); | |
3008 | #else | |
3009 | min_vruntime = cfs_rq->min_vruntime; | |
3010 | #endif | |
88ec22d3 | 3011 | |
3fe1698b | 3012 | se->vruntime -= min_vruntime; |
88ec22d3 PZ |
3013 | } |
3014 | ||
bb3469ac | 3015 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3016 | /* |
3017 | * effective_load() calculates the load change as seen from the root_task_group | |
3018 | * | |
3019 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3020 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3021 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3022 | * |
3023 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3024 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3025 | * total group weight. | |
3026 | * | |
3027 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3028 | * distribution (s_i) using: | |
3029 | * | |
3030 | * s_i = rw_i / \Sum rw_j (1) | |
3031 | * | |
3032 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3033 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3034 | * shares distribution (s_i): | |
3035 | * | |
3036 | * rw_i = { 2, 4, 1, 0 } | |
3037 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3038 | * | |
3039 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3040 | * task used to run on and the CPU the waker is running on), we need to | |
3041 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3042 | * wakeup, compute the effect of the current task going to sleep. | |
3043 | * | |
3044 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3045 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3046 | * | |
3047 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3048 | * | |
3049 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3050 | * differences in waking a task to CPU 0. The additional task changes the | |
3051 | * weight and shares distributions like: | |
3052 | * | |
3053 | * rw'_i = { 3, 4, 1, 0 } | |
3054 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3055 | * | |
3056 | * We can then compute the difference in effective weight by using: | |
3057 | * | |
3058 | * dw_i = S * (s'_i - s_i) (3) | |
3059 | * | |
3060 | * Where 'S' is the group weight as seen by its parent. | |
3061 | * | |
3062 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3063 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3064 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3065 | */ |
2069dd75 | 3066 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3067 | { |
4be9daaa | 3068 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3069 | |
cf5f0acf | 3070 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3071 | return wl; |
3072 | ||
4be9daaa | 3073 | for_each_sched_entity(se) { |
cf5f0acf | 3074 | long w, W; |
4be9daaa | 3075 | |
977dda7c | 3076 | tg = se->my_q->tg; |
bb3469ac | 3077 | |
cf5f0acf PZ |
3078 | /* |
3079 | * W = @wg + \Sum rw_j | |
3080 | */ | |
3081 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3082 | |
cf5f0acf PZ |
3083 | /* |
3084 | * w = rw_i + @wl | |
3085 | */ | |
3086 | w = se->my_q->load.weight + wl; | |
940959e9 | 3087 | |
cf5f0acf PZ |
3088 | /* |
3089 | * wl = S * s'_i; see (2) | |
3090 | */ | |
3091 | if (W > 0 && w < W) | |
3092 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3093 | else |
3094 | wl = tg->shares; | |
940959e9 | 3095 | |
cf5f0acf PZ |
3096 | /* |
3097 | * Per the above, wl is the new se->load.weight value; since | |
3098 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3099 | * calc_cfs_shares(). | |
3100 | */ | |
977dda7c PT |
3101 | if (wl < MIN_SHARES) |
3102 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3103 | |
3104 | /* | |
3105 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3106 | */ | |
977dda7c | 3107 | wl -= se->load.weight; |
cf5f0acf PZ |
3108 | |
3109 | /* | |
3110 | * Recursively apply this logic to all parent groups to compute | |
3111 | * the final effective load change on the root group. Since | |
3112 | * only the @tg group gets extra weight, all parent groups can | |
3113 | * only redistribute existing shares. @wl is the shift in shares | |
3114 | * resulting from this level per the above. | |
3115 | */ | |
4be9daaa | 3116 | wg = 0; |
4be9daaa | 3117 | } |
bb3469ac | 3118 | |
4be9daaa | 3119 | return wl; |
bb3469ac PZ |
3120 | } |
3121 | #else | |
4be9daaa | 3122 | |
83378269 PZ |
3123 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
3124 | unsigned long wl, unsigned long wg) | |
4be9daaa | 3125 | { |
83378269 | 3126 | return wl; |
bb3469ac | 3127 | } |
4be9daaa | 3128 | |
bb3469ac PZ |
3129 | #endif |
3130 | ||
c88d5910 | 3131 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 3132 | { |
e37b6a7b | 3133 | s64 this_load, load; |
c88d5910 | 3134 | int idx, this_cpu, prev_cpu; |
098fb9db | 3135 | unsigned long tl_per_task; |
c88d5910 | 3136 | struct task_group *tg; |
83378269 | 3137 | unsigned long weight; |
b3137bc8 | 3138 | int balanced; |
098fb9db | 3139 | |
c88d5910 PZ |
3140 | idx = sd->wake_idx; |
3141 | this_cpu = smp_processor_id(); | |
3142 | prev_cpu = task_cpu(p); | |
3143 | load = source_load(prev_cpu, idx); | |
3144 | this_load = target_load(this_cpu, idx); | |
098fb9db | 3145 | |
b3137bc8 MG |
3146 | /* |
3147 | * If sync wakeup then subtract the (maximum possible) | |
3148 | * effect of the currently running task from the load | |
3149 | * of the current CPU: | |
3150 | */ | |
83378269 PZ |
3151 | if (sync) { |
3152 | tg = task_group(current); | |
3153 | weight = current->se.load.weight; | |
3154 | ||
c88d5910 | 3155 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
3156 | load += effective_load(tg, prev_cpu, 0, -weight); |
3157 | } | |
b3137bc8 | 3158 | |
83378269 PZ |
3159 | tg = task_group(p); |
3160 | weight = p->se.load.weight; | |
b3137bc8 | 3161 | |
71a29aa7 PZ |
3162 | /* |
3163 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
3164 | * due to the sync cause above having dropped this_load to 0, we'll |
3165 | * always have an imbalance, but there's really nothing you can do | |
3166 | * about that, so that's good too. | |
71a29aa7 PZ |
3167 | * |
3168 | * Otherwise check if either cpus are near enough in load to allow this | |
3169 | * task to be woken on this_cpu. | |
3170 | */ | |
e37b6a7b PT |
3171 | if (this_load > 0) { |
3172 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
3173 | |
3174 | this_eff_load = 100; | |
3175 | this_eff_load *= power_of(prev_cpu); | |
3176 | this_eff_load *= this_load + | |
3177 | effective_load(tg, this_cpu, weight, weight); | |
3178 | ||
3179 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
3180 | prev_eff_load *= power_of(this_cpu); | |
3181 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
3182 | ||
3183 | balanced = this_eff_load <= prev_eff_load; | |
3184 | } else | |
3185 | balanced = true; | |
b3137bc8 | 3186 | |
098fb9db | 3187 | /* |
4ae7d5ce IM |
3188 | * If the currently running task will sleep within |
3189 | * a reasonable amount of time then attract this newly | |
3190 | * woken task: | |
098fb9db | 3191 | */ |
2fb7635c PZ |
3192 | if (sync && balanced) |
3193 | return 1; | |
098fb9db | 3194 | |
41acab88 | 3195 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
3196 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
3197 | ||
c88d5910 PZ |
3198 | if (balanced || |
3199 | (this_load <= load && | |
3200 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
3201 | /* |
3202 | * This domain has SD_WAKE_AFFINE and | |
3203 | * p is cache cold in this domain, and | |
3204 | * there is no bad imbalance. | |
3205 | */ | |
c88d5910 | 3206 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 3207 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
3208 | |
3209 | return 1; | |
3210 | } | |
3211 | return 0; | |
3212 | } | |
3213 | ||
aaee1203 PZ |
3214 | /* |
3215 | * find_idlest_group finds and returns the least busy CPU group within the | |
3216 | * domain. | |
3217 | */ | |
3218 | static struct sched_group * | |
78e7ed53 | 3219 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 3220 | int this_cpu, int load_idx) |
e7693a36 | 3221 | { |
b3bd3de6 | 3222 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 3223 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 3224 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 3225 | |
aaee1203 PZ |
3226 | do { |
3227 | unsigned long load, avg_load; | |
3228 | int local_group; | |
3229 | int i; | |
e7693a36 | 3230 | |
aaee1203 PZ |
3231 | /* Skip over this group if it has no CPUs allowed */ |
3232 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 3233 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
3234 | continue; |
3235 | ||
3236 | local_group = cpumask_test_cpu(this_cpu, | |
3237 | sched_group_cpus(group)); | |
3238 | ||
3239 | /* Tally up the load of all CPUs in the group */ | |
3240 | avg_load = 0; | |
3241 | ||
3242 | for_each_cpu(i, sched_group_cpus(group)) { | |
3243 | /* Bias balancing toward cpus of our domain */ | |
3244 | if (local_group) | |
3245 | load = source_load(i, load_idx); | |
3246 | else | |
3247 | load = target_load(i, load_idx); | |
3248 | ||
3249 | avg_load += load; | |
3250 | } | |
3251 | ||
3252 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3253 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
3254 | |
3255 | if (local_group) { | |
3256 | this_load = avg_load; | |
aaee1203 PZ |
3257 | } else if (avg_load < min_load) { |
3258 | min_load = avg_load; | |
3259 | idlest = group; | |
3260 | } | |
3261 | } while (group = group->next, group != sd->groups); | |
3262 | ||
3263 | if (!idlest || 100*this_load < imbalance*min_load) | |
3264 | return NULL; | |
3265 | return idlest; | |
3266 | } | |
3267 | ||
3268 | /* | |
3269 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
3270 | */ | |
3271 | static int | |
3272 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
3273 | { | |
3274 | unsigned long load, min_load = ULONG_MAX; | |
3275 | int idlest = -1; | |
3276 | int i; | |
3277 | ||
3278 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 3279 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
3280 | load = weighted_cpuload(i); |
3281 | ||
3282 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
3283 | min_load = load; | |
3284 | idlest = i; | |
e7693a36 GH |
3285 | } |
3286 | } | |
3287 | ||
aaee1203 PZ |
3288 | return idlest; |
3289 | } | |
e7693a36 | 3290 | |
a50bde51 PZ |
3291 | /* |
3292 | * Try and locate an idle CPU in the sched_domain. | |
3293 | */ | |
99bd5e2f | 3294 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 3295 | { |
99bd5e2f | 3296 | struct sched_domain *sd; |
37407ea7 | 3297 | struct sched_group *sg; |
e0a79f52 | 3298 | int i = task_cpu(p); |
a50bde51 | 3299 | |
e0a79f52 MG |
3300 | if (idle_cpu(target)) |
3301 | return target; | |
99bd5e2f SS |
3302 | |
3303 | /* | |
e0a79f52 | 3304 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 3305 | */ |
e0a79f52 MG |
3306 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
3307 | return i; | |
a50bde51 PZ |
3308 | |
3309 | /* | |
37407ea7 | 3310 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 3311 | */ |
518cd623 | 3312 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 3313 | for_each_lower_domain(sd) { |
37407ea7 LT |
3314 | sg = sd->groups; |
3315 | do { | |
3316 | if (!cpumask_intersects(sched_group_cpus(sg), | |
3317 | tsk_cpus_allowed(p))) | |
3318 | goto next; | |
3319 | ||
3320 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 3321 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
3322 | goto next; |
3323 | } | |
970e1789 | 3324 | |
37407ea7 LT |
3325 | target = cpumask_first_and(sched_group_cpus(sg), |
3326 | tsk_cpus_allowed(p)); | |
3327 | goto done; | |
3328 | next: | |
3329 | sg = sg->next; | |
3330 | } while (sg != sd->groups); | |
3331 | } | |
3332 | done: | |
a50bde51 PZ |
3333 | return target; |
3334 | } | |
3335 | ||
aaee1203 PZ |
3336 | /* |
3337 | * sched_balance_self: balance the current task (running on cpu) in domains | |
3338 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
3339 | * SD_BALANCE_EXEC. | |
3340 | * | |
3341 | * Balance, ie. select the least loaded group. | |
3342 | * | |
3343 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
3344 | * | |
3345 | * preempt must be disabled. | |
3346 | */ | |
0017d735 | 3347 | static int |
7608dec2 | 3348 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 3349 | { |
29cd8bae | 3350 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
3351 | int cpu = smp_processor_id(); |
3352 | int prev_cpu = task_cpu(p); | |
3353 | int new_cpu = cpu; | |
99bd5e2f | 3354 | int want_affine = 0; |
5158f4e4 | 3355 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 3356 | |
29baa747 | 3357 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
3358 | return prev_cpu; |
3359 | ||
0763a660 | 3360 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 3361 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
3362 | want_affine = 1; |
3363 | new_cpu = prev_cpu; | |
3364 | } | |
aaee1203 | 3365 | |
dce840a0 | 3366 | rcu_read_lock(); |
aaee1203 | 3367 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
3368 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
3369 | continue; | |
3370 | ||
fe3bcfe1 | 3371 | /* |
99bd5e2f SS |
3372 | * If both cpu and prev_cpu are part of this domain, |
3373 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 3374 | */ |
99bd5e2f SS |
3375 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
3376 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
3377 | affine_sd = tmp; | |
29cd8bae | 3378 | break; |
f03542a7 | 3379 | } |
29cd8bae | 3380 | |
f03542a7 | 3381 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
3382 | sd = tmp; |
3383 | } | |
3384 | ||
8b911acd | 3385 | if (affine_sd) { |
f03542a7 | 3386 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
3387 | prev_cpu = cpu; |
3388 | ||
3389 | new_cpu = select_idle_sibling(p, prev_cpu); | |
3390 | goto unlock; | |
8b911acd | 3391 | } |
e7693a36 | 3392 | |
aaee1203 | 3393 | while (sd) { |
5158f4e4 | 3394 | int load_idx = sd->forkexec_idx; |
aaee1203 | 3395 | struct sched_group *group; |
c88d5910 | 3396 | int weight; |
098fb9db | 3397 | |
0763a660 | 3398 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
3399 | sd = sd->child; |
3400 | continue; | |
3401 | } | |
098fb9db | 3402 | |
5158f4e4 PZ |
3403 | if (sd_flag & SD_BALANCE_WAKE) |
3404 | load_idx = sd->wake_idx; | |
098fb9db | 3405 | |
5158f4e4 | 3406 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
3407 | if (!group) { |
3408 | sd = sd->child; | |
3409 | continue; | |
3410 | } | |
4ae7d5ce | 3411 | |
d7c33c49 | 3412 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
3413 | if (new_cpu == -1 || new_cpu == cpu) { |
3414 | /* Now try balancing at a lower domain level of cpu */ | |
3415 | sd = sd->child; | |
3416 | continue; | |
e7693a36 | 3417 | } |
aaee1203 PZ |
3418 | |
3419 | /* Now try balancing at a lower domain level of new_cpu */ | |
3420 | cpu = new_cpu; | |
669c55e9 | 3421 | weight = sd->span_weight; |
aaee1203 PZ |
3422 | sd = NULL; |
3423 | for_each_domain(cpu, tmp) { | |
669c55e9 | 3424 | if (weight <= tmp->span_weight) |
aaee1203 | 3425 | break; |
0763a660 | 3426 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
3427 | sd = tmp; |
3428 | } | |
3429 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 3430 | } |
dce840a0 PZ |
3431 | unlock: |
3432 | rcu_read_unlock(); | |
e7693a36 | 3433 | |
c88d5910 | 3434 | return new_cpu; |
e7693a36 | 3435 | } |
0a74bef8 | 3436 | |
f4e26b12 PT |
3437 | /* |
3438 | * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be | |
3439 | * removed when useful for applications beyond shares distribution (e.g. | |
3440 | * load-balance). | |
3441 | */ | |
3442 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
0a74bef8 PT |
3443 | /* |
3444 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
3445 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
3446 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
3447 | * other assumptions, including the state of rq->lock, should be made. | |
3448 | */ | |
3449 | static void | |
3450 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
3451 | { | |
aff3e498 PT |
3452 | struct sched_entity *se = &p->se; |
3453 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3454 | ||
3455 | /* | |
3456 | * Load tracking: accumulate removed load so that it can be processed | |
3457 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
3458 | * to blocked load iff they have a positive decay-count. It can never | |
3459 | * be negative here since on-rq tasks have decay-count == 0. | |
3460 | */ | |
3461 | if (se->avg.decay_count) { | |
3462 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
3463 | atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load); | |
3464 | } | |
0a74bef8 | 3465 | } |
f4e26b12 | 3466 | #endif |
e7693a36 GH |
3467 | #endif /* CONFIG_SMP */ |
3468 | ||
e52fb7c0 PZ |
3469 | static unsigned long |
3470 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
3471 | { |
3472 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
3473 | ||
3474 | /* | |
e52fb7c0 PZ |
3475 | * Since its curr running now, convert the gran from real-time |
3476 | * to virtual-time in his units. | |
13814d42 MG |
3477 | * |
3478 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
3479 | * they get preempted easier. That is, if 'se' < 'curr' then | |
3480 | * the resulting gran will be larger, therefore penalizing the | |
3481 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
3482 | * be smaller, again penalizing the lighter task. | |
3483 | * | |
3484 | * This is especially important for buddies when the leftmost | |
3485 | * task is higher priority than the buddy. | |
0bbd3336 | 3486 | */ |
f4ad9bd2 | 3487 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
3488 | } |
3489 | ||
464b7527 PZ |
3490 | /* |
3491 | * Should 'se' preempt 'curr'. | |
3492 | * | |
3493 | * |s1 | |
3494 | * |s2 | |
3495 | * |s3 | |
3496 | * g | |
3497 | * |<--->|c | |
3498 | * | |
3499 | * w(c, s1) = -1 | |
3500 | * w(c, s2) = 0 | |
3501 | * w(c, s3) = 1 | |
3502 | * | |
3503 | */ | |
3504 | static int | |
3505 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
3506 | { | |
3507 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
3508 | ||
3509 | if (vdiff <= 0) | |
3510 | return -1; | |
3511 | ||
e52fb7c0 | 3512 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
3513 | if (vdiff > gran) |
3514 | return 1; | |
3515 | ||
3516 | return 0; | |
3517 | } | |
3518 | ||
02479099 PZ |
3519 | static void set_last_buddy(struct sched_entity *se) |
3520 | { | |
69c80f3e VP |
3521 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3522 | return; | |
3523 | ||
3524 | for_each_sched_entity(se) | |
3525 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
3526 | } |
3527 | ||
3528 | static void set_next_buddy(struct sched_entity *se) | |
3529 | { | |
69c80f3e VP |
3530 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3531 | return; | |
3532 | ||
3533 | for_each_sched_entity(se) | |
3534 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
3535 | } |
3536 | ||
ac53db59 RR |
3537 | static void set_skip_buddy(struct sched_entity *se) |
3538 | { | |
69c80f3e VP |
3539 | for_each_sched_entity(se) |
3540 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
3541 | } |
3542 | ||
bf0f6f24 IM |
3543 | /* |
3544 | * Preempt the current task with a newly woken task if needed: | |
3545 | */ | |
5a9b86f6 | 3546 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
3547 | { |
3548 | struct task_struct *curr = rq->curr; | |
8651a86c | 3549 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 3550 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 3551 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 3552 | int next_buddy_marked = 0; |
bf0f6f24 | 3553 | |
4ae7d5ce IM |
3554 | if (unlikely(se == pse)) |
3555 | return; | |
3556 | ||
5238cdd3 | 3557 | /* |
ddcdf6e7 | 3558 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
3559 | * unconditionally check_prempt_curr() after an enqueue (which may have |
3560 | * lead to a throttle). This both saves work and prevents false | |
3561 | * next-buddy nomination below. | |
3562 | */ | |
3563 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3564 | return; | |
3565 | ||
2f36825b | 3566 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3567 | set_next_buddy(pse); |
2f36825b VP |
3568 | next_buddy_marked = 1; |
3569 | } | |
57fdc26d | 3570 | |
aec0a514 BR |
3571 | /* |
3572 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3573 | * wake up path. | |
5238cdd3 PT |
3574 | * |
3575 | * Note: this also catches the edge-case of curr being in a throttled | |
3576 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3577 | * enqueue of curr) will have resulted in resched being set. This | |
3578 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3579 | * below. | |
aec0a514 BR |
3580 | */ |
3581 | if (test_tsk_need_resched(curr)) | |
3582 | return; | |
3583 | ||
a2f5c9ab DH |
3584 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3585 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3586 | likely(p->policy != SCHED_IDLE)) | |
3587 | goto preempt; | |
3588 | ||
91c234b4 | 3589 | /* |
a2f5c9ab DH |
3590 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3591 | * is driven by the tick): | |
91c234b4 | 3592 | */ |
8ed92e51 | 3593 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 3594 | return; |
bf0f6f24 | 3595 | |
464b7527 | 3596 | find_matching_se(&se, &pse); |
9bbd7374 | 3597 | update_curr(cfs_rq_of(se)); |
002f128b | 3598 | BUG_ON(!pse); |
2f36825b VP |
3599 | if (wakeup_preempt_entity(se, pse) == 1) { |
3600 | /* | |
3601 | * Bias pick_next to pick the sched entity that is | |
3602 | * triggering this preemption. | |
3603 | */ | |
3604 | if (!next_buddy_marked) | |
3605 | set_next_buddy(pse); | |
3a7e73a2 | 3606 | goto preempt; |
2f36825b | 3607 | } |
464b7527 | 3608 | |
3a7e73a2 | 3609 | return; |
a65ac745 | 3610 | |
3a7e73a2 PZ |
3611 | preempt: |
3612 | resched_task(curr); | |
3613 | /* | |
3614 | * Only set the backward buddy when the current task is still | |
3615 | * on the rq. This can happen when a wakeup gets interleaved | |
3616 | * with schedule on the ->pre_schedule() or idle_balance() | |
3617 | * point, either of which can * drop the rq lock. | |
3618 | * | |
3619 | * Also, during early boot the idle thread is in the fair class, | |
3620 | * for obvious reasons its a bad idea to schedule back to it. | |
3621 | */ | |
3622 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3623 | return; | |
3624 | ||
3625 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3626 | set_last_buddy(se); | |
bf0f6f24 IM |
3627 | } |
3628 | ||
fb8d4724 | 3629 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3630 | { |
8f4d37ec | 3631 | struct task_struct *p; |
bf0f6f24 IM |
3632 | struct cfs_rq *cfs_rq = &rq->cfs; |
3633 | struct sched_entity *se; | |
3634 | ||
36ace27e | 3635 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3636 | return NULL; |
3637 | ||
3638 | do { | |
9948f4b2 | 3639 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3640 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3641 | cfs_rq = group_cfs_rq(se); |
3642 | } while (cfs_rq); | |
3643 | ||
8f4d37ec | 3644 | p = task_of(se); |
b39e66ea MG |
3645 | if (hrtick_enabled(rq)) |
3646 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
3647 | |
3648 | return p; | |
bf0f6f24 IM |
3649 | } |
3650 | ||
3651 | /* | |
3652 | * Account for a descheduled task: | |
3653 | */ | |
31ee529c | 3654 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
3655 | { |
3656 | struct sched_entity *se = &prev->se; | |
3657 | struct cfs_rq *cfs_rq; | |
3658 | ||
3659 | for_each_sched_entity(se) { | |
3660 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 3661 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
3662 | } |
3663 | } | |
3664 | ||
ac53db59 RR |
3665 | /* |
3666 | * sched_yield() is very simple | |
3667 | * | |
3668 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
3669 | */ | |
3670 | static void yield_task_fair(struct rq *rq) | |
3671 | { | |
3672 | struct task_struct *curr = rq->curr; | |
3673 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
3674 | struct sched_entity *se = &curr->se; | |
3675 | ||
3676 | /* | |
3677 | * Are we the only task in the tree? | |
3678 | */ | |
3679 | if (unlikely(rq->nr_running == 1)) | |
3680 | return; | |
3681 | ||
3682 | clear_buddies(cfs_rq, se); | |
3683 | ||
3684 | if (curr->policy != SCHED_BATCH) { | |
3685 | update_rq_clock(rq); | |
3686 | /* | |
3687 | * Update run-time statistics of the 'current'. | |
3688 | */ | |
3689 | update_curr(cfs_rq); | |
916671c0 MG |
3690 | /* |
3691 | * Tell update_rq_clock() that we've just updated, | |
3692 | * so we don't do microscopic update in schedule() | |
3693 | * and double the fastpath cost. | |
3694 | */ | |
3695 | rq->skip_clock_update = 1; | |
ac53db59 RR |
3696 | } |
3697 | ||
3698 | set_skip_buddy(se); | |
3699 | } | |
3700 | ||
d95f4122 MG |
3701 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
3702 | { | |
3703 | struct sched_entity *se = &p->se; | |
3704 | ||
5238cdd3 PT |
3705 | /* throttled hierarchies are not runnable */ |
3706 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
3707 | return false; |
3708 | ||
3709 | /* Tell the scheduler that we'd really like pse to run next. */ | |
3710 | set_next_buddy(se); | |
3711 | ||
d95f4122 MG |
3712 | yield_task_fair(rq); |
3713 | ||
3714 | return true; | |
3715 | } | |
3716 | ||
681f3e68 | 3717 | #ifdef CONFIG_SMP |
bf0f6f24 | 3718 | /************************************************** |
e9c84cb8 PZ |
3719 | * Fair scheduling class load-balancing methods. |
3720 | * | |
3721 | * BASICS | |
3722 | * | |
3723 | * The purpose of load-balancing is to achieve the same basic fairness the | |
3724 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
3725 | * time to each task. This is expressed in the following equation: | |
3726 | * | |
3727 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
3728 | * | |
3729 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
3730 | * W_i,0 is defined as: | |
3731 | * | |
3732 | * W_i,0 = \Sum_j w_i,j (2) | |
3733 | * | |
3734 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
3735 | * is derived from the nice value as per prio_to_weight[]. | |
3736 | * | |
3737 | * The weight average is an exponential decay average of the instantaneous | |
3738 | * weight: | |
3739 | * | |
3740 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
3741 | * | |
3742 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
3743 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
3744 | * can also include other factors [XXX]. | |
3745 | * | |
3746 | * To achieve this balance we define a measure of imbalance which follows | |
3747 | * directly from (1): | |
3748 | * | |
3749 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
3750 | * | |
3751 | * We them move tasks around to minimize the imbalance. In the continuous | |
3752 | * function space it is obvious this converges, in the discrete case we get | |
3753 | * a few fun cases generally called infeasible weight scenarios. | |
3754 | * | |
3755 | * [XXX expand on: | |
3756 | * - infeasible weights; | |
3757 | * - local vs global optima in the discrete case. ] | |
3758 | * | |
3759 | * | |
3760 | * SCHED DOMAINS | |
3761 | * | |
3762 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
3763 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
3764 | * topology where each level pairs two lower groups (or better). This results | |
3765 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
3766 | * tree to only the first of the previous level and we decrease the frequency | |
3767 | * of load-balance at each level inv. proportional to the number of cpus in | |
3768 | * the groups. | |
3769 | * | |
3770 | * This yields: | |
3771 | * | |
3772 | * log_2 n 1 n | |
3773 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
3774 | * i = 0 2^i 2^i | |
3775 | * `- size of each group | |
3776 | * | | `- number of cpus doing load-balance | |
3777 | * | `- freq | |
3778 | * `- sum over all levels | |
3779 | * | |
3780 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
3781 | * this makes (5) the runtime complexity of the balancer. | |
3782 | * | |
3783 | * An important property here is that each CPU is still (indirectly) connected | |
3784 | * to every other cpu in at most O(log n) steps: | |
3785 | * | |
3786 | * The adjacency matrix of the resulting graph is given by: | |
3787 | * | |
3788 | * log_2 n | |
3789 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
3790 | * k = 0 | |
3791 | * | |
3792 | * And you'll find that: | |
3793 | * | |
3794 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
3795 | * | |
3796 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
3797 | * The task movement gives a factor of O(m), giving a convergence complexity | |
3798 | * of: | |
3799 | * | |
3800 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
3801 | * | |
3802 | * | |
3803 | * WORK CONSERVING | |
3804 | * | |
3805 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
3806 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
3807 | * tree itself instead of relying on other CPUs to bring it work. | |
3808 | * | |
3809 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
3810 | * time. | |
3811 | * | |
3812 | * [XXX more?] | |
3813 | * | |
3814 | * | |
3815 | * CGROUPS | |
3816 | * | |
3817 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
3818 | * | |
3819 | * s_k,i | |
3820 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
3821 | * S_k | |
3822 | * | |
3823 | * Where | |
3824 | * | |
3825 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
3826 | * | |
3827 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
3828 | * | |
3829 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
3830 | * property. | |
3831 | * | |
3832 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
3833 | * rewrite all of this once again.] | |
3834 | */ | |
bf0f6f24 | 3835 | |
ed387b78 HS |
3836 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
3837 | ||
ddcdf6e7 | 3838 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 3839 | #define LBF_NEED_BREAK 0x02 |
88b8dac0 | 3840 | #define LBF_SOME_PINNED 0x04 |
ddcdf6e7 PZ |
3841 | |
3842 | struct lb_env { | |
3843 | struct sched_domain *sd; | |
3844 | ||
ddcdf6e7 | 3845 | struct rq *src_rq; |
85c1e7da | 3846 | int src_cpu; |
ddcdf6e7 PZ |
3847 | |
3848 | int dst_cpu; | |
3849 | struct rq *dst_rq; | |
3850 | ||
88b8dac0 SV |
3851 | struct cpumask *dst_grpmask; |
3852 | int new_dst_cpu; | |
ddcdf6e7 | 3853 | enum cpu_idle_type idle; |
bd939f45 | 3854 | long imbalance; |
b9403130 MW |
3855 | /* The set of CPUs under consideration for load-balancing */ |
3856 | struct cpumask *cpus; | |
3857 | ||
ddcdf6e7 | 3858 | unsigned int flags; |
367456c7 PZ |
3859 | |
3860 | unsigned int loop; | |
3861 | unsigned int loop_break; | |
3862 | unsigned int loop_max; | |
ddcdf6e7 PZ |
3863 | }; |
3864 | ||
1e3c88bd | 3865 | /* |
ddcdf6e7 | 3866 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
3867 | * Both runqueues must be locked. |
3868 | */ | |
ddcdf6e7 | 3869 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 3870 | { |
ddcdf6e7 PZ |
3871 | deactivate_task(env->src_rq, p, 0); |
3872 | set_task_cpu(p, env->dst_cpu); | |
3873 | activate_task(env->dst_rq, p, 0); | |
3874 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
3875 | } |
3876 | ||
029632fb PZ |
3877 | /* |
3878 | * Is this task likely cache-hot: | |
3879 | */ | |
3880 | static int | |
3881 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
3882 | { | |
3883 | s64 delta; | |
3884 | ||
3885 | if (p->sched_class != &fair_sched_class) | |
3886 | return 0; | |
3887 | ||
3888 | if (unlikely(p->policy == SCHED_IDLE)) | |
3889 | return 0; | |
3890 | ||
3891 | /* | |
3892 | * Buddy candidates are cache hot: | |
3893 | */ | |
3894 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
3895 | (&p->se == cfs_rq_of(&p->se)->next || | |
3896 | &p->se == cfs_rq_of(&p->se)->last)) | |
3897 | return 1; | |
3898 | ||
3899 | if (sysctl_sched_migration_cost == -1) | |
3900 | return 1; | |
3901 | if (sysctl_sched_migration_cost == 0) | |
3902 | return 0; | |
3903 | ||
3904 | delta = now - p->se.exec_start; | |
3905 | ||
3906 | return delta < (s64)sysctl_sched_migration_cost; | |
3907 | } | |
3908 | ||
1e3c88bd PZ |
3909 | /* |
3910 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3911 | */ | |
3912 | static | |
8e45cb54 | 3913 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
3914 | { |
3915 | int tsk_cache_hot = 0; | |
3916 | /* | |
3917 | * We do not migrate tasks that are: | |
d3198084 | 3918 | * 1) throttled_lb_pair, or |
1e3c88bd | 3919 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
3920 | * 3) running (obviously), or |
3921 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 3922 | */ |
d3198084 JK |
3923 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
3924 | return 0; | |
3925 | ||
ddcdf6e7 | 3926 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 3927 | int cpu; |
88b8dac0 | 3928 | |
41acab88 | 3929 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 SV |
3930 | |
3931 | /* | |
3932 | * Remember if this task can be migrated to any other cpu in | |
3933 | * our sched_group. We may want to revisit it if we couldn't | |
3934 | * meet load balance goals by pulling other tasks on src_cpu. | |
3935 | * | |
3936 | * Also avoid computing new_dst_cpu if we have already computed | |
3937 | * one in current iteration. | |
3938 | */ | |
3939 | if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED)) | |
3940 | return 0; | |
3941 | ||
e02e60c1 JK |
3942 | /* Prevent to re-select dst_cpu via env's cpus */ |
3943 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
3944 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
3945 | env->flags |= LBF_SOME_PINNED; | |
3946 | env->new_dst_cpu = cpu; | |
3947 | break; | |
3948 | } | |
88b8dac0 | 3949 | } |
e02e60c1 | 3950 | |
1e3c88bd PZ |
3951 | return 0; |
3952 | } | |
88b8dac0 SV |
3953 | |
3954 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 3955 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 3956 | |
ddcdf6e7 | 3957 | if (task_running(env->src_rq, p)) { |
41acab88 | 3958 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
3959 | return 0; |
3960 | } | |
3961 | ||
3962 | /* | |
3963 | * Aggressive migration if: | |
3964 | * 1) task is cache cold, or | |
3965 | * 2) too many balance attempts have failed. | |
3966 | */ | |
3967 | ||
78becc27 | 3968 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
1e3c88bd | 3969 | if (!tsk_cache_hot || |
8e45cb54 | 3970 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 3971 | |
1e3c88bd | 3972 | if (tsk_cache_hot) { |
8e45cb54 | 3973 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 3974 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 3975 | } |
4e2dcb73 | 3976 | |
1e3c88bd PZ |
3977 | return 1; |
3978 | } | |
3979 | ||
4e2dcb73 ZH |
3980 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
3981 | return 0; | |
1e3c88bd PZ |
3982 | } |
3983 | ||
897c395f PZ |
3984 | /* |
3985 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3986 | * part of active balancing operations within "domain". | |
3987 | * Returns 1 if successful and 0 otherwise. | |
3988 | * | |
3989 | * Called with both runqueues locked. | |
3990 | */ | |
8e45cb54 | 3991 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
3992 | { |
3993 | struct task_struct *p, *n; | |
897c395f | 3994 | |
367456c7 | 3995 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
3996 | if (!can_migrate_task(p, env)) |
3997 | continue; | |
897c395f | 3998 | |
367456c7 PZ |
3999 | move_task(p, env); |
4000 | /* | |
4001 | * Right now, this is only the second place move_task() | |
4002 | * is called, so we can safely collect move_task() | |
4003 | * stats here rather than inside move_task(). | |
4004 | */ | |
4005 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4006 | return 1; | |
897c395f | 4007 | } |
897c395f PZ |
4008 | return 0; |
4009 | } | |
4010 | ||
367456c7 PZ |
4011 | static unsigned long task_h_load(struct task_struct *p); |
4012 | ||
eb95308e PZ |
4013 | static const unsigned int sched_nr_migrate_break = 32; |
4014 | ||
5d6523eb | 4015 | /* |
bd939f45 | 4016 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4017 | * this_rq, as part of a balancing operation within domain "sd". |
4018 | * Returns 1 if successful and 0 otherwise. | |
4019 | * | |
4020 | * Called with both runqueues locked. | |
4021 | */ | |
4022 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4023 | { |
5d6523eb PZ |
4024 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4025 | struct task_struct *p; | |
367456c7 PZ |
4026 | unsigned long load; |
4027 | int pulled = 0; | |
1e3c88bd | 4028 | |
bd939f45 | 4029 | if (env->imbalance <= 0) |
5d6523eb | 4030 | return 0; |
1e3c88bd | 4031 | |
5d6523eb PZ |
4032 | while (!list_empty(tasks)) { |
4033 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 4034 | |
367456c7 PZ |
4035 | env->loop++; |
4036 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 4037 | if (env->loop > env->loop_max) |
367456c7 | 4038 | break; |
5d6523eb PZ |
4039 | |
4040 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 4041 | if (env->loop > env->loop_break) { |
eb95308e | 4042 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 4043 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 4044 | break; |
a195f004 | 4045 | } |
1e3c88bd | 4046 | |
d3198084 | 4047 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
4048 | goto next; |
4049 | ||
4050 | load = task_h_load(p); | |
5d6523eb | 4051 | |
eb95308e | 4052 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
4053 | goto next; |
4054 | ||
bd939f45 | 4055 | if ((load / 2) > env->imbalance) |
367456c7 | 4056 | goto next; |
1e3c88bd | 4057 | |
ddcdf6e7 | 4058 | move_task(p, env); |
ee00e66f | 4059 | pulled++; |
bd939f45 | 4060 | env->imbalance -= load; |
1e3c88bd PZ |
4061 | |
4062 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
4063 | /* |
4064 | * NEWIDLE balancing is a source of latency, so preemptible | |
4065 | * kernels will stop after the first task is pulled to minimize | |
4066 | * the critical section. | |
4067 | */ | |
5d6523eb | 4068 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 4069 | break; |
1e3c88bd PZ |
4070 | #endif |
4071 | ||
ee00e66f PZ |
4072 | /* |
4073 | * We only want to steal up to the prescribed amount of | |
4074 | * weighted load. | |
4075 | */ | |
bd939f45 | 4076 | if (env->imbalance <= 0) |
ee00e66f | 4077 | break; |
367456c7 PZ |
4078 | |
4079 | continue; | |
4080 | next: | |
5d6523eb | 4081 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 4082 | } |
5d6523eb | 4083 | |
1e3c88bd | 4084 | /* |
ddcdf6e7 PZ |
4085 | * Right now, this is one of only two places move_task() is called, |
4086 | * so we can safely collect move_task() stats here rather than | |
4087 | * inside move_task(). | |
1e3c88bd | 4088 | */ |
8e45cb54 | 4089 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 4090 | |
5d6523eb | 4091 | return pulled; |
1e3c88bd PZ |
4092 | } |
4093 | ||
230059de | 4094 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
4095 | /* |
4096 | * update tg->load_weight by folding this cpu's load_avg | |
4097 | */ | |
48a16753 | 4098 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 4099 | { |
48a16753 PT |
4100 | struct sched_entity *se = tg->se[cpu]; |
4101 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 4102 | |
48a16753 PT |
4103 | /* throttled entities do not contribute to load */ |
4104 | if (throttled_hierarchy(cfs_rq)) | |
4105 | return; | |
9e3081ca | 4106 | |
aff3e498 | 4107 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 4108 | |
82958366 PT |
4109 | if (se) { |
4110 | update_entity_load_avg(se, 1); | |
4111 | /* | |
4112 | * We pivot on our runnable average having decayed to zero for | |
4113 | * list removal. This generally implies that all our children | |
4114 | * have also been removed (modulo rounding error or bandwidth | |
4115 | * control); however, such cases are rare and we can fix these | |
4116 | * at enqueue. | |
4117 | * | |
4118 | * TODO: fix up out-of-order children on enqueue. | |
4119 | */ | |
4120 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
4121 | list_del_leaf_cfs_rq(cfs_rq); | |
4122 | } else { | |
48a16753 | 4123 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
4124 | update_rq_runnable_avg(rq, rq->nr_running); |
4125 | } | |
9e3081ca PZ |
4126 | } |
4127 | ||
48a16753 | 4128 | static void update_blocked_averages(int cpu) |
9e3081ca | 4129 | { |
9e3081ca | 4130 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
4131 | struct cfs_rq *cfs_rq; |
4132 | unsigned long flags; | |
9e3081ca | 4133 | |
48a16753 PT |
4134 | raw_spin_lock_irqsave(&rq->lock, flags); |
4135 | update_rq_clock(rq); | |
9763b67f PZ |
4136 | /* |
4137 | * Iterates the task_group tree in a bottom up fashion, see | |
4138 | * list_add_leaf_cfs_rq() for details. | |
4139 | */ | |
64660c86 | 4140 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
4141 | /* |
4142 | * Note: We may want to consider periodically releasing | |
4143 | * rq->lock about these updates so that creating many task | |
4144 | * groups does not result in continually extending hold time. | |
4145 | */ | |
4146 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 4147 | } |
48a16753 PT |
4148 | |
4149 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
4150 | } |
4151 | ||
9763b67f PZ |
4152 | /* |
4153 | * Compute the cpu's hierarchical load factor for each task group. | |
4154 | * This needs to be done in a top-down fashion because the load of a child | |
4155 | * group is a fraction of its parents load. | |
4156 | */ | |
4157 | static int tg_load_down(struct task_group *tg, void *data) | |
4158 | { | |
4159 | unsigned long load; | |
4160 | long cpu = (long)data; | |
4161 | ||
4162 | if (!tg->parent) { | |
4163 | load = cpu_rq(cpu)->load.weight; | |
4164 | } else { | |
4165 | load = tg->parent->cfs_rq[cpu]->h_load; | |
4166 | load *= tg->se[cpu]->load.weight; | |
4167 | load /= tg->parent->cfs_rq[cpu]->load.weight + 1; | |
4168 | } | |
4169 | ||
4170 | tg->cfs_rq[cpu]->h_load = load; | |
4171 | ||
4172 | return 0; | |
4173 | } | |
4174 | ||
4175 | static void update_h_load(long cpu) | |
4176 | { | |
a35b6466 PZ |
4177 | struct rq *rq = cpu_rq(cpu); |
4178 | unsigned long now = jiffies; | |
4179 | ||
4180 | if (rq->h_load_throttle == now) | |
4181 | return; | |
4182 | ||
4183 | rq->h_load_throttle = now; | |
4184 | ||
367456c7 | 4185 | rcu_read_lock(); |
9763b67f | 4186 | walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); |
367456c7 | 4187 | rcu_read_unlock(); |
9763b67f PZ |
4188 | } |
4189 | ||
367456c7 | 4190 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 4191 | { |
367456c7 PZ |
4192 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
4193 | unsigned long load; | |
230059de | 4194 | |
367456c7 PZ |
4195 | load = p->se.load.weight; |
4196 | load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1); | |
230059de | 4197 | |
367456c7 | 4198 | return load; |
230059de PZ |
4199 | } |
4200 | #else | |
48a16753 | 4201 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
4202 | { |
4203 | } | |
4204 | ||
367456c7 | 4205 | static inline void update_h_load(long cpu) |
230059de | 4206 | { |
230059de | 4207 | } |
230059de | 4208 | |
367456c7 | 4209 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 4210 | { |
367456c7 | 4211 | return p->se.load.weight; |
1e3c88bd | 4212 | } |
230059de | 4213 | #endif |
1e3c88bd | 4214 | |
1e3c88bd PZ |
4215 | /********** Helpers for find_busiest_group ************************/ |
4216 | /* | |
4217 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
4218 | * during load balancing. | |
4219 | */ | |
4220 | struct sd_lb_stats { | |
4221 | struct sched_group *busiest; /* Busiest group in this sd */ | |
4222 | struct sched_group *this; /* Local group in this sd */ | |
4223 | unsigned long total_load; /* Total load of all groups in sd */ | |
4224 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
4225 | unsigned long avg_load; /* Average load across all groups in sd */ | |
4226 | ||
4227 | /** Statistics of this group */ | |
4228 | unsigned long this_load; | |
4229 | unsigned long this_load_per_task; | |
4230 | unsigned long this_nr_running; | |
fab47622 | 4231 | unsigned long this_has_capacity; |
aae6d3dd | 4232 | unsigned int this_idle_cpus; |
1e3c88bd PZ |
4233 | |
4234 | /* Statistics of the busiest group */ | |
aae6d3dd | 4235 | unsigned int busiest_idle_cpus; |
1e3c88bd PZ |
4236 | unsigned long max_load; |
4237 | unsigned long busiest_load_per_task; | |
4238 | unsigned long busiest_nr_running; | |
dd5feea1 | 4239 | unsigned long busiest_group_capacity; |
fab47622 | 4240 | unsigned long busiest_has_capacity; |
aae6d3dd | 4241 | unsigned int busiest_group_weight; |
1e3c88bd PZ |
4242 | |
4243 | int group_imb; /* Is there imbalance in this sd */ | |
1e3c88bd PZ |
4244 | }; |
4245 | ||
4246 | /* | |
4247 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
4248 | */ | |
4249 | struct sg_lb_stats { | |
4250 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
4251 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
4252 | unsigned long sum_nr_running; /* Nr tasks running in the group */ | |
4253 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | |
4254 | unsigned long group_capacity; | |
aae6d3dd SS |
4255 | unsigned long idle_cpus; |
4256 | unsigned long group_weight; | |
1e3c88bd | 4257 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 4258 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
4259 | }; |
4260 | ||
1e3c88bd PZ |
4261 | /** |
4262 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
4263 | * @sd: The sched_domain whose load_idx is to be obtained. | |
4264 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
4265 | */ | |
4266 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
4267 | enum cpu_idle_type idle) | |
4268 | { | |
4269 | int load_idx; | |
4270 | ||
4271 | switch (idle) { | |
4272 | case CPU_NOT_IDLE: | |
4273 | load_idx = sd->busy_idx; | |
4274 | break; | |
4275 | ||
4276 | case CPU_NEWLY_IDLE: | |
4277 | load_idx = sd->newidle_idx; | |
4278 | break; | |
4279 | default: | |
4280 | load_idx = sd->idle_idx; | |
4281 | break; | |
4282 | } | |
4283 | ||
4284 | return load_idx; | |
4285 | } | |
4286 | ||
15f803c9 | 4287 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 4288 | { |
1399fa78 | 4289 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
4290 | } |
4291 | ||
4292 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
4293 | { | |
4294 | return default_scale_freq_power(sd, cpu); | |
4295 | } | |
4296 | ||
15f803c9 | 4297 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 4298 | { |
669c55e9 | 4299 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
4300 | unsigned long smt_gain = sd->smt_gain; |
4301 | ||
4302 | smt_gain /= weight; | |
4303 | ||
4304 | return smt_gain; | |
4305 | } | |
4306 | ||
4307 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
4308 | { | |
4309 | return default_scale_smt_power(sd, cpu); | |
4310 | } | |
4311 | ||
15f803c9 | 4312 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
4313 | { |
4314 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 4315 | u64 total, available, age_stamp, avg; |
1e3c88bd | 4316 | |
b654f7de PZ |
4317 | /* |
4318 | * Since we're reading these variables without serialization make sure | |
4319 | * we read them once before doing sanity checks on them. | |
4320 | */ | |
4321 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
4322 | avg = ACCESS_ONCE(rq->rt_avg); | |
4323 | ||
78becc27 | 4324 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 4325 | |
b654f7de | 4326 | if (unlikely(total < avg)) { |
aa483808 VP |
4327 | /* Ensures that power won't end up being negative */ |
4328 | available = 0; | |
4329 | } else { | |
b654f7de | 4330 | available = total - avg; |
aa483808 | 4331 | } |
1e3c88bd | 4332 | |
1399fa78 NR |
4333 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
4334 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 4335 | |
1399fa78 | 4336 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4337 | |
4338 | return div_u64(available, total); | |
4339 | } | |
4340 | ||
4341 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
4342 | { | |
669c55e9 | 4343 | unsigned long weight = sd->span_weight; |
1399fa78 | 4344 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
4345 | struct sched_group *sdg = sd->groups; |
4346 | ||
1e3c88bd PZ |
4347 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
4348 | if (sched_feat(ARCH_POWER)) | |
4349 | power *= arch_scale_smt_power(sd, cpu); | |
4350 | else | |
4351 | power *= default_scale_smt_power(sd, cpu); | |
4352 | ||
1399fa78 | 4353 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4354 | } |
4355 | ||
9c3f75cb | 4356 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
4357 | |
4358 | if (sched_feat(ARCH_POWER)) | |
4359 | power *= arch_scale_freq_power(sd, cpu); | |
4360 | else | |
4361 | power *= default_scale_freq_power(sd, cpu); | |
4362 | ||
1399fa78 | 4363 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 4364 | |
1e3c88bd | 4365 | power *= scale_rt_power(cpu); |
1399fa78 | 4366 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4367 | |
4368 | if (!power) | |
4369 | power = 1; | |
4370 | ||
e51fd5e2 | 4371 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 4372 | sdg->sgp->power = power; |
1e3c88bd PZ |
4373 | } |
4374 | ||
029632fb | 4375 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
4376 | { |
4377 | struct sched_domain *child = sd->child; | |
4378 | struct sched_group *group, *sdg = sd->groups; | |
4379 | unsigned long power; | |
4ec4412e VG |
4380 | unsigned long interval; |
4381 | ||
4382 | interval = msecs_to_jiffies(sd->balance_interval); | |
4383 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
4384 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
4385 | |
4386 | if (!child) { | |
4387 | update_cpu_power(sd, cpu); | |
4388 | return; | |
4389 | } | |
4390 | ||
4391 | power = 0; | |
4392 | ||
74a5ce20 PZ |
4393 | if (child->flags & SD_OVERLAP) { |
4394 | /* | |
4395 | * SD_OVERLAP domains cannot assume that child groups | |
4396 | * span the current group. | |
4397 | */ | |
4398 | ||
4399 | for_each_cpu(cpu, sched_group_cpus(sdg)) | |
4400 | power += power_of(cpu); | |
4401 | } else { | |
4402 | /* | |
4403 | * !SD_OVERLAP domains can assume that child groups | |
4404 | * span the current group. | |
4405 | */ | |
4406 | ||
4407 | group = child->groups; | |
4408 | do { | |
4409 | power += group->sgp->power; | |
4410 | group = group->next; | |
4411 | } while (group != child->groups); | |
4412 | } | |
1e3c88bd | 4413 | |
c3decf0d | 4414 | sdg->sgp->power_orig = sdg->sgp->power = power; |
1e3c88bd PZ |
4415 | } |
4416 | ||
9d5efe05 SV |
4417 | /* |
4418 | * Try and fix up capacity for tiny siblings, this is needed when | |
4419 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
4420 | * which on its own isn't powerful enough. | |
4421 | * | |
4422 | * See update_sd_pick_busiest() and check_asym_packing(). | |
4423 | */ | |
4424 | static inline int | |
4425 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
4426 | { | |
4427 | /* | |
1399fa78 | 4428 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 4429 | */ |
a6c75f2f | 4430 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
4431 | return 0; |
4432 | ||
4433 | /* | |
4434 | * If ~90% of the cpu_power is still there, we're good. | |
4435 | */ | |
9c3f75cb | 4436 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
4437 | return 1; |
4438 | ||
4439 | return 0; | |
4440 | } | |
4441 | ||
1e3c88bd PZ |
4442 | /** |
4443 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 4444 | * @env: The load balancing environment. |
1e3c88bd | 4445 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 4446 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 4447 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
4448 | * @balance: Should we balance. |
4449 | * @sgs: variable to hold the statistics for this group. | |
4450 | */ | |
bd939f45 PZ |
4451 | static inline void update_sg_lb_stats(struct lb_env *env, |
4452 | struct sched_group *group, int load_idx, | |
b9403130 | 4453 | int local_group, int *balance, struct sg_lb_stats *sgs) |
1e3c88bd | 4454 | { |
e44bc5c5 PZ |
4455 | unsigned long nr_running, max_nr_running, min_nr_running; |
4456 | unsigned long load, max_cpu_load, min_cpu_load; | |
04f733b4 | 4457 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
dd5feea1 | 4458 | unsigned long avg_load_per_task = 0; |
bd939f45 | 4459 | int i; |
1e3c88bd | 4460 | |
871e35bc | 4461 | if (local_group) |
c1174876 | 4462 | balance_cpu = group_balance_cpu(group); |
1e3c88bd PZ |
4463 | |
4464 | /* Tally up the load of all CPUs in the group */ | |
1e3c88bd PZ |
4465 | max_cpu_load = 0; |
4466 | min_cpu_load = ~0UL; | |
2582f0eb | 4467 | max_nr_running = 0; |
e44bc5c5 | 4468 | min_nr_running = ~0UL; |
1e3c88bd | 4469 | |
b9403130 | 4470 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
4471 | struct rq *rq = cpu_rq(i); |
4472 | ||
e44bc5c5 PZ |
4473 | nr_running = rq->nr_running; |
4474 | ||
1e3c88bd PZ |
4475 | /* Bias balancing toward cpus of our domain */ |
4476 | if (local_group) { | |
c1174876 PZ |
4477 | if (idle_cpu(i) && !first_idle_cpu && |
4478 | cpumask_test_cpu(i, sched_group_mask(group))) { | |
04f733b4 | 4479 | first_idle_cpu = 1; |
1e3c88bd PZ |
4480 | balance_cpu = i; |
4481 | } | |
04f733b4 PZ |
4482 | |
4483 | load = target_load(i, load_idx); | |
1e3c88bd PZ |
4484 | } else { |
4485 | load = source_load(i, load_idx); | |
e44bc5c5 | 4486 | if (load > max_cpu_load) |
1e3c88bd PZ |
4487 | max_cpu_load = load; |
4488 | if (min_cpu_load > load) | |
4489 | min_cpu_load = load; | |
e44bc5c5 PZ |
4490 | |
4491 | if (nr_running > max_nr_running) | |
4492 | max_nr_running = nr_running; | |
4493 | if (min_nr_running > nr_running) | |
4494 | min_nr_running = nr_running; | |
1e3c88bd PZ |
4495 | } |
4496 | ||
4497 | sgs->group_load += load; | |
e44bc5c5 | 4498 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 4499 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
4500 | if (idle_cpu(i)) |
4501 | sgs->idle_cpus++; | |
1e3c88bd PZ |
4502 | } |
4503 | ||
4504 | /* | |
4505 | * First idle cpu or the first cpu(busiest) in this sched group | |
4506 | * is eligible for doing load balancing at this and above | |
4507 | * domains. In the newly idle case, we will allow all the cpu's | |
4508 | * to do the newly idle load balance. | |
4509 | */ | |
4ec4412e | 4510 | if (local_group) { |
bd939f45 | 4511 | if (env->idle != CPU_NEWLY_IDLE) { |
04f733b4 | 4512 | if (balance_cpu != env->dst_cpu) { |
4ec4412e VG |
4513 | *balance = 0; |
4514 | return; | |
4515 | } | |
bd939f45 | 4516 | update_group_power(env->sd, env->dst_cpu); |
4ec4412e | 4517 | } else if (time_after_eq(jiffies, group->sgp->next_update)) |
bd939f45 | 4518 | update_group_power(env->sd, env->dst_cpu); |
1e3c88bd PZ |
4519 | } |
4520 | ||
4521 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 4522 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; |
1e3c88bd | 4523 | |
1e3c88bd PZ |
4524 | /* |
4525 | * Consider the group unbalanced when the imbalance is larger | |
866ab43e | 4526 | * than the average weight of a task. |
1e3c88bd PZ |
4527 | * |
4528 | * APZ: with cgroup the avg task weight can vary wildly and | |
4529 | * might not be a suitable number - should we keep a | |
4530 | * normalized nr_running number somewhere that negates | |
4531 | * the hierarchy? | |
4532 | */ | |
dd5feea1 SS |
4533 | if (sgs->sum_nr_running) |
4534 | avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | |
1e3c88bd | 4535 | |
e44bc5c5 PZ |
4536 | if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && |
4537 | (max_nr_running - min_nr_running) > 1) | |
1e3c88bd PZ |
4538 | sgs->group_imb = 1; |
4539 | ||
9c3f75cb | 4540 | sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, |
1399fa78 | 4541 | SCHED_POWER_SCALE); |
9d5efe05 | 4542 | if (!sgs->group_capacity) |
bd939f45 | 4543 | sgs->group_capacity = fix_small_capacity(env->sd, group); |
aae6d3dd | 4544 | sgs->group_weight = group->group_weight; |
fab47622 NR |
4545 | |
4546 | if (sgs->group_capacity > sgs->sum_nr_running) | |
4547 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
4548 | } |
4549 | ||
532cb4c4 MN |
4550 | /** |
4551 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 4552 | * @env: The load balancing environment. |
532cb4c4 MN |
4553 | * @sds: sched_domain statistics |
4554 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 4555 | * @sgs: sched_group statistics |
532cb4c4 MN |
4556 | * |
4557 | * Determine if @sg is a busier group than the previously selected | |
4558 | * busiest group. | |
4559 | */ | |
bd939f45 | 4560 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
4561 | struct sd_lb_stats *sds, |
4562 | struct sched_group *sg, | |
bd939f45 | 4563 | struct sg_lb_stats *sgs) |
532cb4c4 MN |
4564 | { |
4565 | if (sgs->avg_load <= sds->max_load) | |
4566 | return false; | |
4567 | ||
4568 | if (sgs->sum_nr_running > sgs->group_capacity) | |
4569 | return true; | |
4570 | ||
4571 | if (sgs->group_imb) | |
4572 | return true; | |
4573 | ||
4574 | /* | |
4575 | * ASYM_PACKING needs to move all the work to the lowest | |
4576 | * numbered CPUs in the group, therefore mark all groups | |
4577 | * higher than ourself as busy. | |
4578 | */ | |
bd939f45 PZ |
4579 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
4580 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
4581 | if (!sds->busiest) |
4582 | return true; | |
4583 | ||
4584 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
4585 | return true; | |
4586 | } | |
4587 | ||
4588 | return false; | |
4589 | } | |
4590 | ||
1e3c88bd | 4591 | /** |
461819ac | 4592 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 4593 | * @env: The load balancing environment. |
1e3c88bd PZ |
4594 | * @balance: Should we balance. |
4595 | * @sds: variable to hold the statistics for this sched_domain. | |
4596 | */ | |
bd939f45 | 4597 | static inline void update_sd_lb_stats(struct lb_env *env, |
b9403130 | 4598 | int *balance, struct sd_lb_stats *sds) |
1e3c88bd | 4599 | { |
bd939f45 PZ |
4600 | struct sched_domain *child = env->sd->child; |
4601 | struct sched_group *sg = env->sd->groups; | |
1e3c88bd PZ |
4602 | struct sg_lb_stats sgs; |
4603 | int load_idx, prefer_sibling = 0; | |
4604 | ||
4605 | if (child && child->flags & SD_PREFER_SIBLING) | |
4606 | prefer_sibling = 1; | |
4607 | ||
bd939f45 | 4608 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
4609 | |
4610 | do { | |
4611 | int local_group; | |
4612 | ||
bd939f45 | 4613 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
1e3c88bd | 4614 | memset(&sgs, 0, sizeof(sgs)); |
b9403130 | 4615 | update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs); |
1e3c88bd | 4616 | |
8f190fb3 | 4617 | if (local_group && !(*balance)) |
1e3c88bd PZ |
4618 | return; |
4619 | ||
4620 | sds->total_load += sgs.group_load; | |
9c3f75cb | 4621 | sds->total_pwr += sg->sgp->power; |
1e3c88bd PZ |
4622 | |
4623 | /* | |
4624 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 4625 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
4626 | * and move all the excess tasks away. We lower the capacity |
4627 | * of a group only if the local group has the capacity to fit | |
4628 | * these excess tasks, i.e. nr_running < group_capacity. The | |
4629 | * extra check prevents the case where you always pull from the | |
4630 | * heaviest group when it is already under-utilized (possible | |
4631 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 4632 | */ |
75dd321d | 4633 | if (prefer_sibling && !local_group && sds->this_has_capacity) |
1e3c88bd PZ |
4634 | sgs.group_capacity = min(sgs.group_capacity, 1UL); |
4635 | ||
4636 | if (local_group) { | |
4637 | sds->this_load = sgs.avg_load; | |
532cb4c4 | 4638 | sds->this = sg; |
1e3c88bd PZ |
4639 | sds->this_nr_running = sgs.sum_nr_running; |
4640 | sds->this_load_per_task = sgs.sum_weighted_load; | |
fab47622 | 4641 | sds->this_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4642 | sds->this_idle_cpus = sgs.idle_cpus; |
bd939f45 | 4643 | } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) { |
1e3c88bd | 4644 | sds->max_load = sgs.avg_load; |
532cb4c4 | 4645 | sds->busiest = sg; |
1e3c88bd | 4646 | sds->busiest_nr_running = sgs.sum_nr_running; |
aae6d3dd | 4647 | sds->busiest_idle_cpus = sgs.idle_cpus; |
dd5feea1 | 4648 | sds->busiest_group_capacity = sgs.group_capacity; |
1e3c88bd | 4649 | sds->busiest_load_per_task = sgs.sum_weighted_load; |
fab47622 | 4650 | sds->busiest_has_capacity = sgs.group_has_capacity; |
aae6d3dd | 4651 | sds->busiest_group_weight = sgs.group_weight; |
1e3c88bd PZ |
4652 | sds->group_imb = sgs.group_imb; |
4653 | } | |
4654 | ||
532cb4c4 | 4655 | sg = sg->next; |
bd939f45 | 4656 | } while (sg != env->sd->groups); |
532cb4c4 MN |
4657 | } |
4658 | ||
532cb4c4 MN |
4659 | /** |
4660 | * check_asym_packing - Check to see if the group is packed into the | |
4661 | * sched doman. | |
4662 | * | |
4663 | * This is primarily intended to used at the sibling level. Some | |
4664 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
4665 | * case of POWER7, it can move to lower SMT modes only when higher | |
4666 | * threads are idle. When in lower SMT modes, the threads will | |
4667 | * perform better since they share less core resources. Hence when we | |
4668 | * have idle threads, we want them to be the higher ones. | |
4669 | * | |
4670 | * This packing function is run on idle threads. It checks to see if | |
4671 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
4672 | * CPU number than the packing function is being run on. Here we are | |
4673 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
4674 | * number. | |
4675 | * | |
b6b12294 MN |
4676 | * Returns 1 when packing is required and a task should be moved to |
4677 | * this CPU. The amount of the imbalance is returned in *imbalance. | |
4678 | * | |
cd96891d | 4679 | * @env: The load balancing environment. |
532cb4c4 | 4680 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 4681 | */ |
bd939f45 | 4682 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
4683 | { |
4684 | int busiest_cpu; | |
4685 | ||
bd939f45 | 4686 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
4687 | return 0; |
4688 | ||
4689 | if (!sds->busiest) | |
4690 | return 0; | |
4691 | ||
4692 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 4693 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
4694 | return 0; |
4695 | ||
bd939f45 PZ |
4696 | env->imbalance = DIV_ROUND_CLOSEST( |
4697 | sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE); | |
4698 | ||
532cb4c4 | 4699 | return 1; |
1e3c88bd PZ |
4700 | } |
4701 | ||
4702 | /** | |
4703 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
4704 | * amongst the groups of a sched_domain, during | |
4705 | * load balancing. | |
cd96891d | 4706 | * @env: The load balancing environment. |
1e3c88bd | 4707 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4708 | */ |
bd939f45 PZ |
4709 | static inline |
4710 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
4711 | { |
4712 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
4713 | unsigned int imbn = 2; | |
dd5feea1 | 4714 | unsigned long scaled_busy_load_per_task; |
1e3c88bd PZ |
4715 | |
4716 | if (sds->this_nr_running) { | |
4717 | sds->this_load_per_task /= sds->this_nr_running; | |
4718 | if (sds->busiest_load_per_task > | |
4719 | sds->this_load_per_task) | |
4720 | imbn = 1; | |
bd939f45 | 4721 | } else { |
1e3c88bd | 4722 | sds->this_load_per_task = |
bd939f45 PZ |
4723 | cpu_avg_load_per_task(env->dst_cpu); |
4724 | } | |
1e3c88bd | 4725 | |
dd5feea1 | 4726 | scaled_busy_load_per_task = sds->busiest_load_per_task |
1399fa78 | 4727 | * SCHED_POWER_SCALE; |
9c3f75cb | 4728 | scaled_busy_load_per_task /= sds->busiest->sgp->power; |
dd5feea1 SS |
4729 | |
4730 | if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= | |
4731 | (scaled_busy_load_per_task * imbn)) { | |
bd939f45 | 4732 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4733 | return; |
4734 | } | |
4735 | ||
4736 | /* | |
4737 | * OK, we don't have enough imbalance to justify moving tasks, | |
4738 | * however we may be able to increase total CPU power used by | |
4739 | * moving them. | |
4740 | */ | |
4741 | ||
9c3f75cb | 4742 | pwr_now += sds->busiest->sgp->power * |
1e3c88bd | 4743 | min(sds->busiest_load_per_task, sds->max_load); |
9c3f75cb | 4744 | pwr_now += sds->this->sgp->power * |
1e3c88bd | 4745 | min(sds->this_load_per_task, sds->this_load); |
1399fa78 | 4746 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4747 | |
4748 | /* Amount of load we'd subtract */ | |
1399fa78 | 4749 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb | 4750 | sds->busiest->sgp->power; |
1e3c88bd | 4751 | if (sds->max_load > tmp) |
9c3f75cb | 4752 | pwr_move += sds->busiest->sgp->power * |
1e3c88bd PZ |
4753 | min(sds->busiest_load_per_task, sds->max_load - tmp); |
4754 | ||
4755 | /* Amount of load we'd add */ | |
9c3f75cb | 4756 | if (sds->max_load * sds->busiest->sgp->power < |
1399fa78 | 4757 | sds->busiest_load_per_task * SCHED_POWER_SCALE) |
9c3f75cb PZ |
4758 | tmp = (sds->max_load * sds->busiest->sgp->power) / |
4759 | sds->this->sgp->power; | |
1e3c88bd | 4760 | else |
1399fa78 | 4761 | tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / |
9c3f75cb PZ |
4762 | sds->this->sgp->power; |
4763 | pwr_move += sds->this->sgp->power * | |
1e3c88bd | 4764 | min(sds->this_load_per_task, sds->this_load + tmp); |
1399fa78 | 4765 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4766 | |
4767 | /* Move if we gain throughput */ | |
4768 | if (pwr_move > pwr_now) | |
bd939f45 | 4769 | env->imbalance = sds->busiest_load_per_task; |
1e3c88bd PZ |
4770 | } |
4771 | ||
4772 | /** | |
4773 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
4774 | * groups of a given sched_domain during load balance. | |
bd939f45 | 4775 | * @env: load balance environment |
1e3c88bd | 4776 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4777 | */ |
bd939f45 | 4778 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 4779 | { |
dd5feea1 SS |
4780 | unsigned long max_pull, load_above_capacity = ~0UL; |
4781 | ||
4782 | sds->busiest_load_per_task /= sds->busiest_nr_running; | |
4783 | if (sds->group_imb) { | |
4784 | sds->busiest_load_per_task = | |
4785 | min(sds->busiest_load_per_task, sds->avg_load); | |
4786 | } | |
4787 | ||
1e3c88bd PZ |
4788 | /* |
4789 | * In the presence of smp nice balancing, certain scenarios can have | |
4790 | * max load less than avg load(as we skip the groups at or below | |
4791 | * its cpu_power, while calculating max_load..) | |
4792 | */ | |
4793 | if (sds->max_load < sds->avg_load) { | |
bd939f45 PZ |
4794 | env->imbalance = 0; |
4795 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4796 | } |
4797 | ||
dd5feea1 SS |
4798 | if (!sds->group_imb) { |
4799 | /* | |
4800 | * Don't want to pull so many tasks that a group would go idle. | |
4801 | */ | |
4802 | load_above_capacity = (sds->busiest_nr_running - | |
4803 | sds->busiest_group_capacity); | |
4804 | ||
1399fa78 | 4805 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
dd5feea1 | 4806 | |
9c3f75cb | 4807 | load_above_capacity /= sds->busiest->sgp->power; |
dd5feea1 SS |
4808 | } |
4809 | ||
4810 | /* | |
4811 | * We're trying to get all the cpus to the average_load, so we don't | |
4812 | * want to push ourselves above the average load, nor do we wish to | |
4813 | * reduce the max loaded cpu below the average load. At the same time, | |
4814 | * we also don't want to reduce the group load below the group capacity | |
4815 | * (so that we can implement power-savings policies etc). Thus we look | |
4816 | * for the minimum possible imbalance. | |
4817 | * Be careful of negative numbers as they'll appear as very large values | |
4818 | * with unsigned longs. | |
4819 | */ | |
4820 | max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); | |
1e3c88bd PZ |
4821 | |
4822 | /* How much load to actually move to equalise the imbalance */ | |
bd939f45 | 4823 | env->imbalance = min(max_pull * sds->busiest->sgp->power, |
9c3f75cb | 4824 | (sds->avg_load - sds->this_load) * sds->this->sgp->power) |
1399fa78 | 4825 | / SCHED_POWER_SCALE; |
1e3c88bd PZ |
4826 | |
4827 | /* | |
4828 | * if *imbalance is less than the average load per runnable task | |
25985edc | 4829 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
4830 | * a think about bumping its value to force at least one task to be |
4831 | * moved | |
4832 | */ | |
bd939f45 PZ |
4833 | if (env->imbalance < sds->busiest_load_per_task) |
4834 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
4835 | |
4836 | } | |
fab47622 | 4837 | |
1e3c88bd PZ |
4838 | /******* find_busiest_group() helpers end here *********************/ |
4839 | ||
4840 | /** | |
4841 | * find_busiest_group - Returns the busiest group within the sched_domain | |
4842 | * if there is an imbalance. If there isn't an imbalance, and | |
4843 | * the user has opted for power-savings, it returns a group whose | |
4844 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
4845 | * such a group exists. | |
4846 | * | |
4847 | * Also calculates the amount of weighted load which should be moved | |
4848 | * to restore balance. | |
4849 | * | |
cd96891d | 4850 | * @env: The load balancing environment. |
1e3c88bd PZ |
4851 | * @balance: Pointer to a variable indicating if this_cpu |
4852 | * is the appropriate cpu to perform load balancing at this_level. | |
4853 | * | |
4854 | * Returns: - the busiest group if imbalance exists. | |
4855 | * - If no imbalance and user has opted for power-savings balance, | |
4856 | * return the least loaded group whose CPUs can be | |
4857 | * put to idle by rebalancing its tasks onto our group. | |
4858 | */ | |
4859 | static struct sched_group * | |
b9403130 | 4860 | find_busiest_group(struct lb_env *env, int *balance) |
1e3c88bd PZ |
4861 | { |
4862 | struct sd_lb_stats sds; | |
4863 | ||
4864 | memset(&sds, 0, sizeof(sds)); | |
4865 | ||
4866 | /* | |
4867 | * Compute the various statistics relavent for load balancing at | |
4868 | * this level. | |
4869 | */ | |
b9403130 | 4870 | update_sd_lb_stats(env, balance, &sds); |
1e3c88bd | 4871 | |
cc57aa8f PZ |
4872 | /* |
4873 | * this_cpu is not the appropriate cpu to perform load balancing at | |
4874 | * this level. | |
1e3c88bd | 4875 | */ |
8f190fb3 | 4876 | if (!(*balance)) |
1e3c88bd PZ |
4877 | goto ret; |
4878 | ||
bd939f45 PZ |
4879 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
4880 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
4881 | return sds.busiest; |
4882 | ||
cc57aa8f | 4883 | /* There is no busy sibling group to pull tasks from */ |
1e3c88bd PZ |
4884 | if (!sds.busiest || sds.busiest_nr_running == 0) |
4885 | goto out_balanced; | |
4886 | ||
1399fa78 | 4887 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 4888 | |
866ab43e PZ |
4889 | /* |
4890 | * If the busiest group is imbalanced the below checks don't | |
4891 | * work because they assumes all things are equal, which typically | |
4892 | * isn't true due to cpus_allowed constraints and the like. | |
4893 | */ | |
4894 | if (sds.group_imb) | |
4895 | goto force_balance; | |
4896 | ||
cc57aa8f | 4897 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
bd939f45 | 4898 | if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity && |
fab47622 NR |
4899 | !sds.busiest_has_capacity) |
4900 | goto force_balance; | |
4901 | ||
cc57aa8f PZ |
4902 | /* |
4903 | * If the local group is more busy than the selected busiest group | |
4904 | * don't try and pull any tasks. | |
4905 | */ | |
1e3c88bd PZ |
4906 | if (sds.this_load >= sds.max_load) |
4907 | goto out_balanced; | |
4908 | ||
cc57aa8f PZ |
4909 | /* |
4910 | * Don't pull any tasks if this group is already above the domain | |
4911 | * average load. | |
4912 | */ | |
1e3c88bd PZ |
4913 | if (sds.this_load >= sds.avg_load) |
4914 | goto out_balanced; | |
4915 | ||
bd939f45 | 4916 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
4917 | /* |
4918 | * This cpu is idle. If the busiest group load doesn't | |
4919 | * have more tasks than the number of available cpu's and | |
4920 | * there is no imbalance between this and busiest group | |
4921 | * wrt to idle cpu's, it is balanced. | |
4922 | */ | |
c186fafe | 4923 | if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && |
aae6d3dd SS |
4924 | sds.busiest_nr_running <= sds.busiest_group_weight) |
4925 | goto out_balanced; | |
c186fafe PZ |
4926 | } else { |
4927 | /* | |
4928 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
4929 | * imbalance_pct to be conservative. | |
4930 | */ | |
bd939f45 | 4931 | if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load) |
c186fafe | 4932 | goto out_balanced; |
aae6d3dd | 4933 | } |
1e3c88bd | 4934 | |
fab47622 | 4935 | force_balance: |
1e3c88bd | 4936 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 4937 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
4938 | return sds.busiest; |
4939 | ||
4940 | out_balanced: | |
1e3c88bd | 4941 | ret: |
bd939f45 | 4942 | env->imbalance = 0; |
1e3c88bd PZ |
4943 | return NULL; |
4944 | } | |
4945 | ||
4946 | /* | |
4947 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
4948 | */ | |
bd939f45 | 4949 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 4950 | struct sched_group *group) |
1e3c88bd PZ |
4951 | { |
4952 | struct rq *busiest = NULL, *rq; | |
4953 | unsigned long max_load = 0; | |
4954 | int i; | |
4955 | ||
4956 | for_each_cpu(i, sched_group_cpus(group)) { | |
4957 | unsigned long power = power_of(i); | |
1399fa78 NR |
4958 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
4959 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
4960 | unsigned long wl; |
4961 | ||
9d5efe05 | 4962 | if (!capacity) |
bd939f45 | 4963 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 4964 | |
b9403130 | 4965 | if (!cpumask_test_cpu(i, env->cpus)) |
1e3c88bd PZ |
4966 | continue; |
4967 | ||
4968 | rq = cpu_rq(i); | |
6e40f5bb | 4969 | wl = weighted_cpuload(i); |
1e3c88bd | 4970 | |
6e40f5bb TG |
4971 | /* |
4972 | * When comparing with imbalance, use weighted_cpuload() | |
4973 | * which is not scaled with the cpu power. | |
4974 | */ | |
bd939f45 | 4975 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
4976 | continue; |
4977 | ||
6e40f5bb TG |
4978 | /* |
4979 | * For the load comparisons with the other cpu's, consider | |
4980 | * the weighted_cpuload() scaled with the cpu power, so that | |
4981 | * the load can be moved away from the cpu that is potentially | |
4982 | * running at a lower capacity. | |
4983 | */ | |
1399fa78 | 4984 | wl = (wl * SCHED_POWER_SCALE) / power; |
6e40f5bb | 4985 | |
1e3c88bd PZ |
4986 | if (wl > max_load) { |
4987 | max_load = wl; | |
4988 | busiest = rq; | |
4989 | } | |
4990 | } | |
4991 | ||
4992 | return busiest; | |
4993 | } | |
4994 | ||
4995 | /* | |
4996 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
4997 | * so long as it is large enough. | |
4998 | */ | |
4999 | #define MAX_PINNED_INTERVAL 512 | |
5000 | ||
5001 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 5002 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 5003 | |
bd939f45 | 5004 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 5005 | { |
bd939f45 PZ |
5006 | struct sched_domain *sd = env->sd; |
5007 | ||
5008 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
5009 | |
5010 | /* | |
5011 | * ASYM_PACKING needs to force migrate tasks from busy but | |
5012 | * higher numbered CPUs in order to pack all tasks in the | |
5013 | * lowest numbered CPUs. | |
5014 | */ | |
bd939f45 | 5015 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 5016 | return 1; |
1af3ed3d PZ |
5017 | } |
5018 | ||
5019 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
5020 | } | |
5021 | ||
969c7921 TH |
5022 | static int active_load_balance_cpu_stop(void *data); |
5023 | ||
1e3c88bd PZ |
5024 | /* |
5025 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
5026 | * tasks if there is an imbalance. | |
5027 | */ | |
5028 | static int load_balance(int this_cpu, struct rq *this_rq, | |
5029 | struct sched_domain *sd, enum cpu_idle_type idle, | |
5030 | int *balance) | |
5031 | { | |
88b8dac0 | 5032 | int ld_moved, cur_ld_moved, active_balance = 0; |
1e3c88bd | 5033 | struct sched_group *group; |
1e3c88bd PZ |
5034 | struct rq *busiest; |
5035 | unsigned long flags; | |
e6252c3e | 5036 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 5037 | |
8e45cb54 PZ |
5038 | struct lb_env env = { |
5039 | .sd = sd, | |
ddcdf6e7 PZ |
5040 | .dst_cpu = this_cpu, |
5041 | .dst_rq = this_rq, | |
88b8dac0 | 5042 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 5043 | .idle = idle, |
eb95308e | 5044 | .loop_break = sched_nr_migrate_break, |
b9403130 | 5045 | .cpus = cpus, |
8e45cb54 PZ |
5046 | }; |
5047 | ||
cfc03118 JK |
5048 | /* |
5049 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
5050 | * other cpus in our group | |
5051 | */ | |
e02e60c1 | 5052 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 5053 | env.dst_grpmask = NULL; |
cfc03118 | 5054 | |
1e3c88bd PZ |
5055 | cpumask_copy(cpus, cpu_active_mask); |
5056 | ||
1e3c88bd PZ |
5057 | schedstat_inc(sd, lb_count[idle]); |
5058 | ||
5059 | redo: | |
b9403130 | 5060 | group = find_busiest_group(&env, balance); |
1e3c88bd PZ |
5061 | |
5062 | if (*balance == 0) | |
5063 | goto out_balanced; | |
5064 | ||
5065 | if (!group) { | |
5066 | schedstat_inc(sd, lb_nobusyg[idle]); | |
5067 | goto out_balanced; | |
5068 | } | |
5069 | ||
b9403130 | 5070 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
5071 | if (!busiest) { |
5072 | schedstat_inc(sd, lb_nobusyq[idle]); | |
5073 | goto out_balanced; | |
5074 | } | |
5075 | ||
78feefc5 | 5076 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 5077 | |
bd939f45 | 5078 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
5079 | |
5080 | ld_moved = 0; | |
5081 | if (busiest->nr_running > 1) { | |
5082 | /* | |
5083 | * Attempt to move tasks. If find_busiest_group has found | |
5084 | * an imbalance but busiest->nr_running <= 1, the group is | |
5085 | * still unbalanced. ld_moved simply stays zero, so it is | |
5086 | * correctly treated as an imbalance. | |
5087 | */ | |
8e45cb54 | 5088 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
5089 | env.src_cpu = busiest->cpu; |
5090 | env.src_rq = busiest; | |
5091 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 5092 | |
a35b6466 | 5093 | update_h_load(env.src_cpu); |
5d6523eb | 5094 | more_balance: |
1e3c88bd | 5095 | local_irq_save(flags); |
78feefc5 | 5096 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
5097 | |
5098 | /* | |
5099 | * cur_ld_moved - load moved in current iteration | |
5100 | * ld_moved - cumulative load moved across iterations | |
5101 | */ | |
5102 | cur_ld_moved = move_tasks(&env); | |
5103 | ld_moved += cur_ld_moved; | |
78feefc5 | 5104 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
5105 | local_irq_restore(flags); |
5106 | ||
5107 | /* | |
5108 | * some other cpu did the load balance for us. | |
5109 | */ | |
88b8dac0 SV |
5110 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
5111 | resched_cpu(env.dst_cpu); | |
5112 | ||
f1cd0858 JK |
5113 | if (env.flags & LBF_NEED_BREAK) { |
5114 | env.flags &= ~LBF_NEED_BREAK; | |
5115 | goto more_balance; | |
5116 | } | |
5117 | ||
88b8dac0 SV |
5118 | /* |
5119 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
5120 | * us and move them to an alternate dst_cpu in our sched_group | |
5121 | * where they can run. The upper limit on how many times we | |
5122 | * iterate on same src_cpu is dependent on number of cpus in our | |
5123 | * sched_group. | |
5124 | * | |
5125 | * This changes load balance semantics a bit on who can move | |
5126 | * load to a given_cpu. In addition to the given_cpu itself | |
5127 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
5128 | * nohz-idle), we now have balance_cpu in a position to move | |
5129 | * load to given_cpu. In rare situations, this may cause | |
5130 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
5131 | * _independently_ and at _same_ time to move some load to | |
5132 | * given_cpu) causing exceess load to be moved to given_cpu. | |
5133 | * This however should not happen so much in practice and | |
5134 | * moreover subsequent load balance cycles should correct the | |
5135 | * excess load moved. | |
5136 | */ | |
e02e60c1 | 5137 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { |
88b8dac0 | 5138 | |
78feefc5 | 5139 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 SV |
5140 | env.dst_cpu = env.new_dst_cpu; |
5141 | env.flags &= ~LBF_SOME_PINNED; | |
5142 | env.loop = 0; | |
5143 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 JK |
5144 | |
5145 | /* Prevent to re-select dst_cpu via env's cpus */ | |
5146 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
5147 | ||
88b8dac0 SV |
5148 | /* |
5149 | * Go back to "more_balance" rather than "redo" since we | |
5150 | * need to continue with same src_cpu. | |
5151 | */ | |
5152 | goto more_balance; | |
5153 | } | |
1e3c88bd PZ |
5154 | |
5155 | /* All tasks on this runqueue were pinned by CPU affinity */ | |
8e45cb54 | 5156 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 5157 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
5158 | if (!cpumask_empty(cpus)) { |
5159 | env.loop = 0; | |
5160 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 5161 | goto redo; |
bbf18b19 | 5162 | } |
1e3c88bd PZ |
5163 | goto out_balanced; |
5164 | } | |
5165 | } | |
5166 | ||
5167 | if (!ld_moved) { | |
5168 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
5169 | /* |
5170 | * Increment the failure counter only on periodic balance. | |
5171 | * We do not want newidle balance, which can be very | |
5172 | * frequent, pollute the failure counter causing | |
5173 | * excessive cache_hot migrations and active balances. | |
5174 | */ | |
5175 | if (idle != CPU_NEWLY_IDLE) | |
5176 | sd->nr_balance_failed++; | |
1e3c88bd | 5177 | |
bd939f45 | 5178 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
5179 | raw_spin_lock_irqsave(&busiest->lock, flags); |
5180 | ||
969c7921 TH |
5181 | /* don't kick the active_load_balance_cpu_stop, |
5182 | * if the curr task on busiest cpu can't be | |
5183 | * moved to this_cpu | |
1e3c88bd PZ |
5184 | */ |
5185 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 5186 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
5187 | raw_spin_unlock_irqrestore(&busiest->lock, |
5188 | flags); | |
8e45cb54 | 5189 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
5190 | goto out_one_pinned; |
5191 | } | |
5192 | ||
969c7921 TH |
5193 | /* |
5194 | * ->active_balance synchronizes accesses to | |
5195 | * ->active_balance_work. Once set, it's cleared | |
5196 | * only after active load balance is finished. | |
5197 | */ | |
1e3c88bd PZ |
5198 | if (!busiest->active_balance) { |
5199 | busiest->active_balance = 1; | |
5200 | busiest->push_cpu = this_cpu; | |
5201 | active_balance = 1; | |
5202 | } | |
5203 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 5204 | |
bd939f45 | 5205 | if (active_balance) { |
969c7921 TH |
5206 | stop_one_cpu_nowait(cpu_of(busiest), |
5207 | active_load_balance_cpu_stop, busiest, | |
5208 | &busiest->active_balance_work); | |
bd939f45 | 5209 | } |
1e3c88bd PZ |
5210 | |
5211 | /* | |
5212 | * We've kicked active balancing, reset the failure | |
5213 | * counter. | |
5214 | */ | |
5215 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
5216 | } | |
5217 | } else | |
5218 | sd->nr_balance_failed = 0; | |
5219 | ||
5220 | if (likely(!active_balance)) { | |
5221 | /* We were unbalanced, so reset the balancing interval */ | |
5222 | sd->balance_interval = sd->min_interval; | |
5223 | } else { | |
5224 | /* | |
5225 | * If we've begun active balancing, start to back off. This | |
5226 | * case may not be covered by the all_pinned logic if there | |
5227 | * is only 1 task on the busy runqueue (because we don't call | |
5228 | * move_tasks). | |
5229 | */ | |
5230 | if (sd->balance_interval < sd->max_interval) | |
5231 | sd->balance_interval *= 2; | |
5232 | } | |
5233 | ||
1e3c88bd PZ |
5234 | goto out; |
5235 | ||
5236 | out_balanced: | |
5237 | schedstat_inc(sd, lb_balanced[idle]); | |
5238 | ||
5239 | sd->nr_balance_failed = 0; | |
5240 | ||
5241 | out_one_pinned: | |
5242 | /* tune up the balancing interval */ | |
8e45cb54 | 5243 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 5244 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
5245 | (sd->balance_interval < sd->max_interval)) |
5246 | sd->balance_interval *= 2; | |
5247 | ||
46e49b38 | 5248 | ld_moved = 0; |
1e3c88bd | 5249 | out: |
1e3c88bd PZ |
5250 | return ld_moved; |
5251 | } | |
5252 | ||
1e3c88bd PZ |
5253 | /* |
5254 | * idle_balance is called by schedule() if this_cpu is about to become | |
5255 | * idle. Attempts to pull tasks from other CPUs. | |
5256 | */ | |
029632fb | 5257 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
5258 | { |
5259 | struct sched_domain *sd; | |
5260 | int pulled_task = 0; | |
5261 | unsigned long next_balance = jiffies + HZ; | |
5262 | ||
78becc27 | 5263 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
5264 | |
5265 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
5266 | return; | |
5267 | ||
f492e12e PZ |
5268 | /* |
5269 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
5270 | */ | |
5271 | raw_spin_unlock(&this_rq->lock); | |
5272 | ||
48a16753 | 5273 | update_blocked_averages(this_cpu); |
dce840a0 | 5274 | rcu_read_lock(); |
1e3c88bd PZ |
5275 | for_each_domain(this_cpu, sd) { |
5276 | unsigned long interval; | |
f492e12e | 5277 | int balance = 1; |
1e3c88bd PZ |
5278 | |
5279 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
5280 | continue; | |
5281 | ||
f492e12e | 5282 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
1e3c88bd | 5283 | /* If we've pulled tasks over stop searching: */ |
f492e12e PZ |
5284 | pulled_task = load_balance(this_cpu, this_rq, |
5285 | sd, CPU_NEWLY_IDLE, &balance); | |
5286 | } | |
1e3c88bd PZ |
5287 | |
5288 | interval = msecs_to_jiffies(sd->balance_interval); | |
5289 | if (time_after(next_balance, sd->last_balance + interval)) | |
5290 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
5291 | if (pulled_task) { |
5292 | this_rq->idle_stamp = 0; | |
1e3c88bd | 5293 | break; |
d5ad140b | 5294 | } |
1e3c88bd | 5295 | } |
dce840a0 | 5296 | rcu_read_unlock(); |
f492e12e PZ |
5297 | |
5298 | raw_spin_lock(&this_rq->lock); | |
5299 | ||
1e3c88bd PZ |
5300 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
5301 | /* | |
5302 | * We are going idle. next_balance may be set based on | |
5303 | * a busy processor. So reset next_balance. | |
5304 | */ | |
5305 | this_rq->next_balance = next_balance; | |
5306 | } | |
5307 | } | |
5308 | ||
5309 | /* | |
969c7921 TH |
5310 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
5311 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
5312 | * least 1 task to be running on each physical CPU where possible, and | |
5313 | * avoids physical / logical imbalances. | |
1e3c88bd | 5314 | */ |
969c7921 | 5315 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 5316 | { |
969c7921 TH |
5317 | struct rq *busiest_rq = data; |
5318 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 5319 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 5320 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 5321 | struct sched_domain *sd; |
969c7921 TH |
5322 | |
5323 | raw_spin_lock_irq(&busiest_rq->lock); | |
5324 | ||
5325 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
5326 | if (unlikely(busiest_cpu != smp_processor_id() || | |
5327 | !busiest_rq->active_balance)) | |
5328 | goto out_unlock; | |
1e3c88bd PZ |
5329 | |
5330 | /* Is there any task to move? */ | |
5331 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 5332 | goto out_unlock; |
1e3c88bd PZ |
5333 | |
5334 | /* | |
5335 | * This condition is "impossible", if it occurs | |
5336 | * we need to fix it. Originally reported by | |
5337 | * Bjorn Helgaas on a 128-cpu setup. | |
5338 | */ | |
5339 | BUG_ON(busiest_rq == target_rq); | |
5340 | ||
5341 | /* move a task from busiest_rq to target_rq */ | |
5342 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
5343 | |
5344 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 5345 | rcu_read_lock(); |
1e3c88bd PZ |
5346 | for_each_domain(target_cpu, sd) { |
5347 | if ((sd->flags & SD_LOAD_BALANCE) && | |
5348 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
5349 | break; | |
5350 | } | |
5351 | ||
5352 | if (likely(sd)) { | |
8e45cb54 PZ |
5353 | struct lb_env env = { |
5354 | .sd = sd, | |
ddcdf6e7 PZ |
5355 | .dst_cpu = target_cpu, |
5356 | .dst_rq = target_rq, | |
5357 | .src_cpu = busiest_rq->cpu, | |
5358 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
5359 | .idle = CPU_IDLE, |
5360 | }; | |
5361 | ||
1e3c88bd PZ |
5362 | schedstat_inc(sd, alb_count); |
5363 | ||
8e45cb54 | 5364 | if (move_one_task(&env)) |
1e3c88bd PZ |
5365 | schedstat_inc(sd, alb_pushed); |
5366 | else | |
5367 | schedstat_inc(sd, alb_failed); | |
5368 | } | |
dce840a0 | 5369 | rcu_read_unlock(); |
1e3c88bd | 5370 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
5371 | out_unlock: |
5372 | busiest_rq->active_balance = 0; | |
5373 | raw_spin_unlock_irq(&busiest_rq->lock); | |
5374 | return 0; | |
1e3c88bd PZ |
5375 | } |
5376 | ||
3451d024 | 5377 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
5378 | /* |
5379 | * idle load balancing details | |
83cd4fe2 VP |
5380 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
5381 | * needed, they will kick the idle load balancer, which then does idle | |
5382 | * load balancing for all the idle CPUs. | |
5383 | */ | |
1e3c88bd | 5384 | static struct { |
83cd4fe2 | 5385 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 5386 | atomic_t nr_cpus; |
83cd4fe2 VP |
5387 | unsigned long next_balance; /* in jiffy units */ |
5388 | } nohz ____cacheline_aligned; | |
1e3c88bd | 5389 | |
8e7fbcbc | 5390 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 5391 | { |
0b005cf5 | 5392 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 5393 | |
786d6dc7 SS |
5394 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
5395 | return ilb; | |
5396 | ||
5397 | return nr_cpu_ids; | |
1e3c88bd | 5398 | } |
1e3c88bd | 5399 | |
83cd4fe2 VP |
5400 | /* |
5401 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
5402 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
5403 | * CPU (if there is one). | |
5404 | */ | |
5405 | static void nohz_balancer_kick(int cpu) | |
5406 | { | |
5407 | int ilb_cpu; | |
5408 | ||
5409 | nohz.next_balance++; | |
5410 | ||
0b005cf5 | 5411 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 5412 | |
0b005cf5 SS |
5413 | if (ilb_cpu >= nr_cpu_ids) |
5414 | return; | |
83cd4fe2 | 5415 | |
cd490c5b | 5416 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
5417 | return; |
5418 | /* | |
5419 | * Use smp_send_reschedule() instead of resched_cpu(). | |
5420 | * This way we generate a sched IPI on the target cpu which | |
5421 | * is idle. And the softirq performing nohz idle load balance | |
5422 | * will be run before returning from the IPI. | |
5423 | */ | |
5424 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
5425 | return; |
5426 | } | |
5427 | ||
c1cc017c | 5428 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
5429 | { |
5430 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
5431 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
5432 | atomic_dec(&nohz.nr_cpus); | |
5433 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
5434 | } | |
5435 | } | |
5436 | ||
69e1e811 SS |
5437 | static inline void set_cpu_sd_state_busy(void) |
5438 | { | |
5439 | struct sched_domain *sd; | |
69e1e811 | 5440 | |
69e1e811 | 5441 | rcu_read_lock(); |
424c93fe | 5442 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
5443 | |
5444 | if (!sd || !sd->nohz_idle) | |
5445 | goto unlock; | |
5446 | sd->nohz_idle = 0; | |
5447 | ||
5448 | for (; sd; sd = sd->parent) | |
69e1e811 | 5449 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 5450 | unlock: |
69e1e811 SS |
5451 | rcu_read_unlock(); |
5452 | } | |
5453 | ||
5454 | void set_cpu_sd_state_idle(void) | |
5455 | { | |
5456 | struct sched_domain *sd; | |
69e1e811 | 5457 | |
69e1e811 | 5458 | rcu_read_lock(); |
424c93fe | 5459 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
5460 | |
5461 | if (!sd || sd->nohz_idle) | |
5462 | goto unlock; | |
5463 | sd->nohz_idle = 1; | |
5464 | ||
5465 | for (; sd; sd = sd->parent) | |
69e1e811 | 5466 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 5467 | unlock: |
69e1e811 SS |
5468 | rcu_read_unlock(); |
5469 | } | |
5470 | ||
1e3c88bd | 5471 | /* |
c1cc017c | 5472 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 5473 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 5474 | */ |
c1cc017c | 5475 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 5476 | { |
71325960 SS |
5477 | /* |
5478 | * If this cpu is going down, then nothing needs to be done. | |
5479 | */ | |
5480 | if (!cpu_active(cpu)) | |
5481 | return; | |
5482 | ||
c1cc017c AS |
5483 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
5484 | return; | |
1e3c88bd | 5485 | |
c1cc017c AS |
5486 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
5487 | atomic_inc(&nohz.nr_cpus); | |
5488 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 5489 | } |
71325960 SS |
5490 | |
5491 | static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb, | |
5492 | unsigned long action, void *hcpu) | |
5493 | { | |
5494 | switch (action & ~CPU_TASKS_FROZEN) { | |
5495 | case CPU_DYING: | |
c1cc017c | 5496 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
5497 | return NOTIFY_OK; |
5498 | default: | |
5499 | return NOTIFY_DONE; | |
5500 | } | |
5501 | } | |
1e3c88bd PZ |
5502 | #endif |
5503 | ||
5504 | static DEFINE_SPINLOCK(balancing); | |
5505 | ||
49c022e6 PZ |
5506 | /* |
5507 | * Scale the max load_balance interval with the number of CPUs in the system. | |
5508 | * This trades load-balance latency on larger machines for less cross talk. | |
5509 | */ | |
029632fb | 5510 | void update_max_interval(void) |
49c022e6 PZ |
5511 | { |
5512 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
5513 | } | |
5514 | ||
1e3c88bd PZ |
5515 | /* |
5516 | * It checks each scheduling domain to see if it is due to be balanced, | |
5517 | * and initiates a balancing operation if so. | |
5518 | * | |
b9b0853a | 5519 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd PZ |
5520 | */ |
5521 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
5522 | { | |
5523 | int balance = 1; | |
5524 | struct rq *rq = cpu_rq(cpu); | |
5525 | unsigned long interval; | |
04f733b4 | 5526 | struct sched_domain *sd; |
1e3c88bd PZ |
5527 | /* Earliest time when we have to do rebalance again */ |
5528 | unsigned long next_balance = jiffies + 60*HZ; | |
5529 | int update_next_balance = 0; | |
5530 | int need_serialize; | |
5531 | ||
48a16753 | 5532 | update_blocked_averages(cpu); |
2069dd75 | 5533 | |
dce840a0 | 5534 | rcu_read_lock(); |
1e3c88bd PZ |
5535 | for_each_domain(cpu, sd) { |
5536 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
5537 | continue; | |
5538 | ||
5539 | interval = sd->balance_interval; | |
5540 | if (idle != CPU_IDLE) | |
5541 | interval *= sd->busy_factor; | |
5542 | ||
5543 | /* scale ms to jiffies */ | |
5544 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 5545 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
5546 | |
5547 | need_serialize = sd->flags & SD_SERIALIZE; | |
5548 | ||
5549 | if (need_serialize) { | |
5550 | if (!spin_trylock(&balancing)) | |
5551 | goto out; | |
5552 | } | |
5553 | ||
5554 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
5555 | if (load_balance(cpu, rq, sd, idle, &balance)) { | |
5556 | /* | |
de5eb2dd JK |
5557 | * The LBF_SOME_PINNED logic could have changed |
5558 | * env->dst_cpu, so we can't know our idle | |
5559 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 5560 | */ |
de5eb2dd | 5561 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
5562 | } |
5563 | sd->last_balance = jiffies; | |
5564 | } | |
5565 | if (need_serialize) | |
5566 | spin_unlock(&balancing); | |
5567 | out: | |
5568 | if (time_after(next_balance, sd->last_balance + interval)) { | |
5569 | next_balance = sd->last_balance + interval; | |
5570 | update_next_balance = 1; | |
5571 | } | |
5572 | ||
5573 | /* | |
5574 | * Stop the load balance at this level. There is another | |
5575 | * CPU in our sched group which is doing load balancing more | |
5576 | * actively. | |
5577 | */ | |
5578 | if (!balance) | |
5579 | break; | |
5580 | } | |
dce840a0 | 5581 | rcu_read_unlock(); |
1e3c88bd PZ |
5582 | |
5583 | /* | |
5584 | * next_balance will be updated only when there is a need. | |
5585 | * When the cpu is attached to null domain for ex, it will not be | |
5586 | * updated. | |
5587 | */ | |
5588 | if (likely(update_next_balance)) | |
5589 | rq->next_balance = next_balance; | |
5590 | } | |
5591 | ||
3451d024 | 5592 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 5593 | /* |
3451d024 | 5594 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
5595 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
5596 | */ | |
83cd4fe2 VP |
5597 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
5598 | { | |
5599 | struct rq *this_rq = cpu_rq(this_cpu); | |
5600 | struct rq *rq; | |
5601 | int balance_cpu; | |
5602 | ||
1c792db7 SS |
5603 | if (idle != CPU_IDLE || |
5604 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
5605 | goto end; | |
83cd4fe2 VP |
5606 | |
5607 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 5608 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
5609 | continue; |
5610 | ||
5611 | /* | |
5612 | * If this cpu gets work to do, stop the load balancing | |
5613 | * work being done for other cpus. Next load | |
5614 | * balancing owner will pick it up. | |
5615 | */ | |
1c792db7 | 5616 | if (need_resched()) |
83cd4fe2 | 5617 | break; |
83cd4fe2 | 5618 | |
5ed4f1d9 VG |
5619 | rq = cpu_rq(balance_cpu); |
5620 | ||
5621 | raw_spin_lock_irq(&rq->lock); | |
5622 | update_rq_clock(rq); | |
5623 | update_idle_cpu_load(rq); | |
5624 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
5625 | |
5626 | rebalance_domains(balance_cpu, CPU_IDLE); | |
5627 | ||
83cd4fe2 VP |
5628 | if (time_after(this_rq->next_balance, rq->next_balance)) |
5629 | this_rq->next_balance = rq->next_balance; | |
5630 | } | |
5631 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
5632 | end: |
5633 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
5634 | } |
5635 | ||
5636 | /* | |
0b005cf5 SS |
5637 | * Current heuristic for kicking the idle load balancer in the presence |
5638 | * of an idle cpu is the system. | |
5639 | * - This rq has more than one task. | |
5640 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
5641 | * busy cpu's exceeding the group's power. | |
5642 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
5643 | * domain span are idle. | |
83cd4fe2 VP |
5644 | */ |
5645 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
5646 | { | |
5647 | unsigned long now = jiffies; | |
0b005cf5 | 5648 | struct sched_domain *sd; |
83cd4fe2 | 5649 | |
1c792db7 | 5650 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
5651 | return 0; |
5652 | ||
1c792db7 SS |
5653 | /* |
5654 | * We may be recently in ticked or tickless idle mode. At the first | |
5655 | * busy tick after returning from idle, we will update the busy stats. | |
5656 | */ | |
69e1e811 | 5657 | set_cpu_sd_state_busy(); |
c1cc017c | 5658 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
5659 | |
5660 | /* | |
5661 | * None are in tickless mode and hence no need for NOHZ idle load | |
5662 | * balancing. | |
5663 | */ | |
5664 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
5665 | return 0; | |
1c792db7 SS |
5666 | |
5667 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
5668 | return 0; |
5669 | ||
0b005cf5 SS |
5670 | if (rq->nr_running >= 2) |
5671 | goto need_kick; | |
83cd4fe2 | 5672 | |
067491b7 | 5673 | rcu_read_lock(); |
0b005cf5 SS |
5674 | for_each_domain(cpu, sd) { |
5675 | struct sched_group *sg = sd->groups; | |
5676 | struct sched_group_power *sgp = sg->sgp; | |
5677 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 5678 | |
0b005cf5 | 5679 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 5680 | goto need_kick_unlock; |
0b005cf5 SS |
5681 | |
5682 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
5683 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
5684 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 5685 | goto need_kick_unlock; |
0b005cf5 SS |
5686 | |
5687 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
5688 | break; | |
83cd4fe2 | 5689 | } |
067491b7 | 5690 | rcu_read_unlock(); |
83cd4fe2 | 5691 | return 0; |
067491b7 PZ |
5692 | |
5693 | need_kick_unlock: | |
5694 | rcu_read_unlock(); | |
0b005cf5 SS |
5695 | need_kick: |
5696 | return 1; | |
83cd4fe2 VP |
5697 | } |
5698 | #else | |
5699 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
5700 | #endif | |
5701 | ||
5702 | /* | |
5703 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
5704 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
5705 | */ | |
1e3c88bd PZ |
5706 | static void run_rebalance_domains(struct softirq_action *h) |
5707 | { | |
5708 | int this_cpu = smp_processor_id(); | |
5709 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 5710 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
5711 | CPU_IDLE : CPU_NOT_IDLE; |
5712 | ||
5713 | rebalance_domains(this_cpu, idle); | |
5714 | ||
1e3c88bd | 5715 | /* |
83cd4fe2 | 5716 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
5717 | * balancing on behalf of the other idle cpus whose ticks are |
5718 | * stopped. | |
5719 | */ | |
83cd4fe2 | 5720 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
5721 | } |
5722 | ||
5723 | static inline int on_null_domain(int cpu) | |
5724 | { | |
90a6501f | 5725 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
5726 | } |
5727 | ||
5728 | /* | |
5729 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 5730 | */ |
029632fb | 5731 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 5732 | { |
1e3c88bd PZ |
5733 | /* Don't need to rebalance while attached to NULL domain */ |
5734 | if (time_after_eq(jiffies, rq->next_balance) && | |
5735 | likely(!on_null_domain(cpu))) | |
5736 | raise_softirq(SCHED_SOFTIRQ); | |
3451d024 | 5737 | #ifdef CONFIG_NO_HZ_COMMON |
1c792db7 | 5738 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
5739 | nohz_balancer_kick(cpu); |
5740 | #endif | |
1e3c88bd PZ |
5741 | } |
5742 | ||
0bcdcf28 CE |
5743 | static void rq_online_fair(struct rq *rq) |
5744 | { | |
5745 | update_sysctl(); | |
5746 | } | |
5747 | ||
5748 | static void rq_offline_fair(struct rq *rq) | |
5749 | { | |
5750 | update_sysctl(); | |
a4c96ae3 PB |
5751 | |
5752 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
5753 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
5754 | } |
5755 | ||
55e12e5e | 5756 | #endif /* CONFIG_SMP */ |
e1d1484f | 5757 | |
bf0f6f24 IM |
5758 | /* |
5759 | * scheduler tick hitting a task of our scheduling class: | |
5760 | */ | |
8f4d37ec | 5761 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
5762 | { |
5763 | struct cfs_rq *cfs_rq; | |
5764 | struct sched_entity *se = &curr->se; | |
5765 | ||
5766 | for_each_sched_entity(se) { | |
5767 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 5768 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 5769 | } |
18bf2805 | 5770 | |
cbee9f88 PZ |
5771 | if (sched_feat_numa(NUMA)) |
5772 | task_tick_numa(rq, curr); | |
3d59eebc | 5773 | |
18bf2805 | 5774 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
5775 | } |
5776 | ||
5777 | /* | |
cd29fe6f PZ |
5778 | * called on fork with the child task as argument from the parent's context |
5779 | * - child not yet on the tasklist | |
5780 | * - preemption disabled | |
bf0f6f24 | 5781 | */ |
cd29fe6f | 5782 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 5783 | { |
4fc420c9 DN |
5784 | struct cfs_rq *cfs_rq; |
5785 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 5786 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
5787 | struct rq *rq = this_rq(); |
5788 | unsigned long flags; | |
5789 | ||
05fa785c | 5790 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 5791 | |
861d034e PZ |
5792 | update_rq_clock(rq); |
5793 | ||
4fc420c9 DN |
5794 | cfs_rq = task_cfs_rq(current); |
5795 | curr = cfs_rq->curr; | |
5796 | ||
b0a0f667 PM |
5797 | if (unlikely(task_cpu(p) != this_cpu)) { |
5798 | rcu_read_lock(); | |
cd29fe6f | 5799 | __set_task_cpu(p, this_cpu); |
b0a0f667 PM |
5800 | rcu_read_unlock(); |
5801 | } | |
bf0f6f24 | 5802 | |
7109c442 | 5803 | update_curr(cfs_rq); |
cd29fe6f | 5804 | |
b5d9d734 MG |
5805 | if (curr) |
5806 | se->vruntime = curr->vruntime; | |
aeb73b04 | 5807 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 5808 | |
cd29fe6f | 5809 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 5810 | /* |
edcb60a3 IM |
5811 | * Upon rescheduling, sched_class::put_prev_task() will place |
5812 | * 'current' within the tree based on its new key value. | |
5813 | */ | |
4d78e7b6 | 5814 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 5815 | resched_task(rq->curr); |
4d78e7b6 | 5816 | } |
bf0f6f24 | 5817 | |
88ec22d3 PZ |
5818 | se->vruntime -= cfs_rq->min_vruntime; |
5819 | ||
05fa785c | 5820 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
5821 | } |
5822 | ||
cb469845 SR |
5823 | /* |
5824 | * Priority of the task has changed. Check to see if we preempt | |
5825 | * the current task. | |
5826 | */ | |
da7a735e PZ |
5827 | static void |
5828 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 5829 | { |
da7a735e PZ |
5830 | if (!p->se.on_rq) |
5831 | return; | |
5832 | ||
cb469845 SR |
5833 | /* |
5834 | * Reschedule if we are currently running on this runqueue and | |
5835 | * our priority decreased, or if we are not currently running on | |
5836 | * this runqueue and our priority is higher than the current's | |
5837 | */ | |
da7a735e | 5838 | if (rq->curr == p) { |
cb469845 SR |
5839 | if (p->prio > oldprio) |
5840 | resched_task(rq->curr); | |
5841 | } else | |
15afe09b | 5842 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5843 | } |
5844 | ||
da7a735e PZ |
5845 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
5846 | { | |
5847 | struct sched_entity *se = &p->se; | |
5848 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5849 | ||
5850 | /* | |
5851 | * Ensure the task's vruntime is normalized, so that when its | |
5852 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
5853 | * do the right thing. | |
5854 | * | |
5855 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
5856 | * have normalized the vruntime, if it was !on_rq, then only when | |
5857 | * the task is sleeping will it still have non-normalized vruntime. | |
5858 | */ | |
5859 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
5860 | /* | |
5861 | * Fix up our vruntime so that the current sleep doesn't | |
5862 | * cause 'unlimited' sleep bonus. | |
5863 | */ | |
5864 | place_entity(cfs_rq, se, 0); | |
5865 | se->vruntime -= cfs_rq->min_vruntime; | |
5866 | } | |
9ee474f5 PT |
5867 | |
5868 | #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP) | |
5869 | /* | |
5870 | * Remove our load from contribution when we leave sched_fair | |
5871 | * and ensure we don't carry in an old decay_count if we | |
5872 | * switch back. | |
5873 | */ | |
5874 | if (p->se.avg.decay_count) { | |
5875 | struct cfs_rq *cfs_rq = cfs_rq_of(&p->se); | |
5876 | __synchronize_entity_decay(&p->se); | |
5877 | subtract_blocked_load_contrib(cfs_rq, | |
5878 | p->se.avg.load_avg_contrib); | |
5879 | } | |
5880 | #endif | |
da7a735e PZ |
5881 | } |
5882 | ||
cb469845 SR |
5883 | /* |
5884 | * We switched to the sched_fair class. | |
5885 | */ | |
da7a735e | 5886 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 5887 | { |
da7a735e PZ |
5888 | if (!p->se.on_rq) |
5889 | return; | |
5890 | ||
cb469845 SR |
5891 | /* |
5892 | * We were most likely switched from sched_rt, so | |
5893 | * kick off the schedule if running, otherwise just see | |
5894 | * if we can still preempt the current task. | |
5895 | */ | |
da7a735e | 5896 | if (rq->curr == p) |
cb469845 SR |
5897 | resched_task(rq->curr); |
5898 | else | |
15afe09b | 5899 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
5900 | } |
5901 | ||
83b699ed SV |
5902 | /* Account for a task changing its policy or group. |
5903 | * | |
5904 | * This routine is mostly called to set cfs_rq->curr field when a task | |
5905 | * migrates between groups/classes. | |
5906 | */ | |
5907 | static void set_curr_task_fair(struct rq *rq) | |
5908 | { | |
5909 | struct sched_entity *se = &rq->curr->se; | |
5910 | ||
ec12cb7f PT |
5911 | for_each_sched_entity(se) { |
5912 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5913 | ||
5914 | set_next_entity(cfs_rq, se); | |
5915 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
5916 | account_cfs_rq_runtime(cfs_rq, 0); | |
5917 | } | |
83b699ed SV |
5918 | } |
5919 | ||
029632fb PZ |
5920 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
5921 | { | |
5922 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
5923 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
5924 | #ifndef CONFIG_64BIT | |
5925 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
5926 | #endif | |
9ee474f5 PT |
5927 | #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP) |
5928 | atomic64_set(&cfs_rq->decay_counter, 1); | |
aff3e498 | 5929 | atomic64_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 5930 | #endif |
029632fb PZ |
5931 | } |
5932 | ||
810b3817 | 5933 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 5934 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 5935 | { |
aff3e498 | 5936 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
5937 | /* |
5938 | * If the task was not on the rq at the time of this cgroup movement | |
5939 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
5940 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
5941 | * bonus in place_entity()). | |
5942 | * | |
5943 | * If it was on the rq, we've just 'preempted' it, which does convert | |
5944 | * ->vruntime to a relative base. | |
5945 | * | |
5946 | * Make sure both cases convert their relative position when migrating | |
5947 | * to another cgroup's rq. This does somewhat interfere with the | |
5948 | * fair sleeper stuff for the first placement, but who cares. | |
5949 | */ | |
7ceff013 DN |
5950 | /* |
5951 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
5952 | * But there are some cases where it has already been normalized: | |
5953 | * | |
5954 | * - Moving a forked child which is waiting for being woken up by | |
5955 | * wake_up_new_task(). | |
62af3783 DN |
5956 | * - Moving a task which has been woken up by try_to_wake_up() and |
5957 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
5958 | * |
5959 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
5960 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
5961 | */ | |
62af3783 | 5962 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
5963 | on_rq = 1; |
5964 | ||
b2b5ce02 PZ |
5965 | if (!on_rq) |
5966 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
5967 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
5968 | if (!on_rq) { |
5969 | cfs_rq = cfs_rq_of(&p->se); | |
5970 | p->se.vruntime += cfs_rq->min_vruntime; | |
5971 | #ifdef CONFIG_SMP | |
5972 | /* | |
5973 | * migrate_task_rq_fair() will have removed our previous | |
5974 | * contribution, but we must synchronize for ongoing future | |
5975 | * decay. | |
5976 | */ | |
5977 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
5978 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
5979 | #endif | |
5980 | } | |
810b3817 | 5981 | } |
029632fb PZ |
5982 | |
5983 | void free_fair_sched_group(struct task_group *tg) | |
5984 | { | |
5985 | int i; | |
5986 | ||
5987 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
5988 | ||
5989 | for_each_possible_cpu(i) { | |
5990 | if (tg->cfs_rq) | |
5991 | kfree(tg->cfs_rq[i]); | |
5992 | if (tg->se) | |
5993 | kfree(tg->se[i]); | |
5994 | } | |
5995 | ||
5996 | kfree(tg->cfs_rq); | |
5997 | kfree(tg->se); | |
5998 | } | |
5999 | ||
6000 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6001 | { | |
6002 | struct cfs_rq *cfs_rq; | |
6003 | struct sched_entity *se; | |
6004 | int i; | |
6005 | ||
6006 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
6007 | if (!tg->cfs_rq) | |
6008 | goto err; | |
6009 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
6010 | if (!tg->se) | |
6011 | goto err; | |
6012 | ||
6013 | tg->shares = NICE_0_LOAD; | |
6014 | ||
6015 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6016 | ||
6017 | for_each_possible_cpu(i) { | |
6018 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
6019 | GFP_KERNEL, cpu_to_node(i)); | |
6020 | if (!cfs_rq) | |
6021 | goto err; | |
6022 | ||
6023 | se = kzalloc_node(sizeof(struct sched_entity), | |
6024 | GFP_KERNEL, cpu_to_node(i)); | |
6025 | if (!se) | |
6026 | goto err_free_rq; | |
6027 | ||
6028 | init_cfs_rq(cfs_rq); | |
6029 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
6030 | } | |
6031 | ||
6032 | return 1; | |
6033 | ||
6034 | err_free_rq: | |
6035 | kfree(cfs_rq); | |
6036 | err: | |
6037 | return 0; | |
6038 | } | |
6039 | ||
6040 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
6041 | { | |
6042 | struct rq *rq = cpu_rq(cpu); | |
6043 | unsigned long flags; | |
6044 | ||
6045 | /* | |
6046 | * Only empty task groups can be destroyed; so we can speculatively | |
6047 | * check on_list without danger of it being re-added. | |
6048 | */ | |
6049 | if (!tg->cfs_rq[cpu]->on_list) | |
6050 | return; | |
6051 | ||
6052 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6053 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
6054 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6055 | } | |
6056 | ||
6057 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
6058 | struct sched_entity *se, int cpu, | |
6059 | struct sched_entity *parent) | |
6060 | { | |
6061 | struct rq *rq = cpu_rq(cpu); | |
6062 | ||
6063 | cfs_rq->tg = tg; | |
6064 | cfs_rq->rq = rq; | |
029632fb PZ |
6065 | init_cfs_rq_runtime(cfs_rq); |
6066 | ||
6067 | tg->cfs_rq[cpu] = cfs_rq; | |
6068 | tg->se[cpu] = se; | |
6069 | ||
6070 | /* se could be NULL for root_task_group */ | |
6071 | if (!se) | |
6072 | return; | |
6073 | ||
6074 | if (!parent) | |
6075 | se->cfs_rq = &rq->cfs; | |
6076 | else | |
6077 | se->cfs_rq = parent->my_q; | |
6078 | ||
6079 | se->my_q = cfs_rq; | |
6080 | update_load_set(&se->load, 0); | |
6081 | se->parent = parent; | |
6082 | } | |
6083 | ||
6084 | static DEFINE_MUTEX(shares_mutex); | |
6085 | ||
6086 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
6087 | { | |
6088 | int i; | |
6089 | unsigned long flags; | |
6090 | ||
6091 | /* | |
6092 | * We can't change the weight of the root cgroup. | |
6093 | */ | |
6094 | if (!tg->se[0]) | |
6095 | return -EINVAL; | |
6096 | ||
6097 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
6098 | ||
6099 | mutex_lock(&shares_mutex); | |
6100 | if (tg->shares == shares) | |
6101 | goto done; | |
6102 | ||
6103 | tg->shares = shares; | |
6104 | for_each_possible_cpu(i) { | |
6105 | struct rq *rq = cpu_rq(i); | |
6106 | struct sched_entity *se; | |
6107 | ||
6108 | se = tg->se[i]; | |
6109 | /* Propagate contribution to hierarchy */ | |
6110 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
6111 | |
6112 | /* Possible calls to update_curr() need rq clock */ | |
6113 | update_rq_clock(rq); | |
17bc14b7 | 6114 | for_each_sched_entity(se) |
029632fb PZ |
6115 | update_cfs_shares(group_cfs_rq(se)); |
6116 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6117 | } | |
6118 | ||
6119 | done: | |
6120 | mutex_unlock(&shares_mutex); | |
6121 | return 0; | |
6122 | } | |
6123 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6124 | ||
6125 | void free_fair_sched_group(struct task_group *tg) { } | |
6126 | ||
6127 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6128 | { | |
6129 | return 1; | |
6130 | } | |
6131 | ||
6132 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
6133 | ||
6134 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
6135 | ||
810b3817 | 6136 | |
6d686f45 | 6137 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
6138 | { |
6139 | struct sched_entity *se = &task->se; | |
0d721cea PW |
6140 | unsigned int rr_interval = 0; |
6141 | ||
6142 | /* | |
6143 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
6144 | * idle runqueue: | |
6145 | */ | |
0d721cea | 6146 | if (rq->cfs.load.weight) |
a59f4e07 | 6147 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
6148 | |
6149 | return rr_interval; | |
6150 | } | |
6151 | ||
bf0f6f24 IM |
6152 | /* |
6153 | * All the scheduling class methods: | |
6154 | */ | |
029632fb | 6155 | const struct sched_class fair_sched_class = { |
5522d5d5 | 6156 | .next = &idle_sched_class, |
bf0f6f24 IM |
6157 | .enqueue_task = enqueue_task_fair, |
6158 | .dequeue_task = dequeue_task_fair, | |
6159 | .yield_task = yield_task_fair, | |
d95f4122 | 6160 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 6161 | |
2e09bf55 | 6162 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
6163 | |
6164 | .pick_next_task = pick_next_task_fair, | |
6165 | .put_prev_task = put_prev_task_fair, | |
6166 | ||
681f3e68 | 6167 | #ifdef CONFIG_SMP |
4ce72a2c | 6168 | .select_task_rq = select_task_rq_fair, |
f4e26b12 | 6169 | #ifdef CONFIG_FAIR_GROUP_SCHED |
0a74bef8 | 6170 | .migrate_task_rq = migrate_task_rq_fair, |
f4e26b12 | 6171 | #endif |
0bcdcf28 CE |
6172 | .rq_online = rq_online_fair, |
6173 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
6174 | |
6175 | .task_waking = task_waking_fair, | |
681f3e68 | 6176 | #endif |
bf0f6f24 | 6177 | |
83b699ed | 6178 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 6179 | .task_tick = task_tick_fair, |
cd29fe6f | 6180 | .task_fork = task_fork_fair, |
cb469845 SR |
6181 | |
6182 | .prio_changed = prio_changed_fair, | |
da7a735e | 6183 | .switched_from = switched_from_fair, |
cb469845 | 6184 | .switched_to = switched_to_fair, |
810b3817 | 6185 | |
0d721cea PW |
6186 | .get_rr_interval = get_rr_interval_fair, |
6187 | ||
810b3817 | 6188 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6189 | .task_move_group = task_move_group_fair, |
810b3817 | 6190 | #endif |
bf0f6f24 IM |
6191 | }; |
6192 | ||
6193 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 6194 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 6195 | { |
bf0f6f24 IM |
6196 | struct cfs_rq *cfs_rq; |
6197 | ||
5973e5b9 | 6198 | rcu_read_lock(); |
c3b64f1e | 6199 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 6200 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 6201 | rcu_read_unlock(); |
bf0f6f24 IM |
6202 | } |
6203 | #endif | |
029632fb PZ |
6204 | |
6205 | __init void init_sched_fair_class(void) | |
6206 | { | |
6207 | #ifdef CONFIG_SMP | |
6208 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
6209 | ||
3451d024 | 6210 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 6211 | nohz.next_balance = jiffies; |
029632fb | 6212 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 6213 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
6214 | #endif |
6215 | #endif /* SMP */ | |
6216 | ||
6217 | } |