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 | ||
9dbdb155 | 181 | #define WMULT_CONST (~0U) |
029632fb PZ |
182 | #define WMULT_SHIFT 32 |
183 | ||
9dbdb155 PZ |
184 | static void __update_inv_weight(struct load_weight *lw) |
185 | { | |
186 | unsigned long w; | |
187 | ||
188 | if (likely(lw->inv_weight)) | |
189 | return; | |
190 | ||
191 | w = scale_load_down(lw->weight); | |
192 | ||
193 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
194 | lw->inv_weight = 1; | |
195 | else if (unlikely(!w)) | |
196 | lw->inv_weight = WMULT_CONST; | |
197 | else | |
198 | lw->inv_weight = WMULT_CONST / w; | |
199 | } | |
029632fb PZ |
200 | |
201 | /* | |
9dbdb155 PZ |
202 | * delta_exec * weight / lw.weight |
203 | * OR | |
204 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
205 | * | |
206 | * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case | |
207 | * we're guaranteed shift stays positive because inv_weight is guaranteed to | |
208 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
209 | * | |
210 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
211 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 212 | */ |
9dbdb155 | 213 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 214 | { |
9dbdb155 PZ |
215 | u64 fact = scale_load_down(weight); |
216 | int shift = WMULT_SHIFT; | |
029632fb | 217 | |
9dbdb155 | 218 | __update_inv_weight(lw); |
029632fb | 219 | |
9dbdb155 PZ |
220 | if (unlikely(fact >> 32)) { |
221 | while (fact >> 32) { | |
222 | fact >>= 1; | |
223 | shift--; | |
224 | } | |
029632fb PZ |
225 | } |
226 | ||
9dbdb155 PZ |
227 | /* hint to use a 32x32->64 mul */ |
228 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 229 | |
9dbdb155 PZ |
230 | while (fact >> 32) { |
231 | fact >>= 1; | |
232 | shift--; | |
233 | } | |
029632fb | 234 | |
9dbdb155 | 235 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
236 | } |
237 | ||
238 | ||
239 | const struct sched_class fair_sched_class; | |
a4c2f00f | 240 | |
bf0f6f24 IM |
241 | /************************************************************** |
242 | * CFS operations on generic schedulable entities: | |
243 | */ | |
244 | ||
62160e3f | 245 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 246 | |
62160e3f | 247 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
248 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
249 | { | |
62160e3f | 250 | return cfs_rq->rq; |
bf0f6f24 IM |
251 | } |
252 | ||
62160e3f IM |
253 | /* An entity is a task if it doesn't "own" a runqueue */ |
254 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 255 | |
8f48894f PZ |
256 | static inline struct task_struct *task_of(struct sched_entity *se) |
257 | { | |
258 | #ifdef CONFIG_SCHED_DEBUG | |
259 | WARN_ON_ONCE(!entity_is_task(se)); | |
260 | #endif | |
261 | return container_of(se, struct task_struct, se); | |
262 | } | |
263 | ||
b758149c PZ |
264 | /* Walk up scheduling entities hierarchy */ |
265 | #define for_each_sched_entity(se) \ | |
266 | for (; se; se = se->parent) | |
267 | ||
268 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
269 | { | |
270 | return p->se.cfs_rq; | |
271 | } | |
272 | ||
273 | /* runqueue on which this entity is (to be) queued */ | |
274 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
275 | { | |
276 | return se->cfs_rq; | |
277 | } | |
278 | ||
279 | /* runqueue "owned" by this group */ | |
280 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
281 | { | |
282 | return grp->my_q; | |
283 | } | |
284 | ||
aff3e498 PT |
285 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
286 | int force_update); | |
9ee474f5 | 287 | |
3d4b47b4 PZ |
288 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
289 | { | |
290 | if (!cfs_rq->on_list) { | |
67e86250 PT |
291 | /* |
292 | * Ensure we either appear before our parent (if already | |
293 | * enqueued) or force our parent to appear after us when it is | |
294 | * enqueued. The fact that we always enqueue bottom-up | |
295 | * reduces this to two cases. | |
296 | */ | |
297 | if (cfs_rq->tg->parent && | |
298 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
299 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
300 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
301 | } else { | |
302 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 303 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 304 | } |
3d4b47b4 PZ |
305 | |
306 | cfs_rq->on_list = 1; | |
9ee474f5 | 307 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 308 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
309 | } |
310 | } | |
311 | ||
312 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
313 | { | |
314 | if (cfs_rq->on_list) { | |
315 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
316 | cfs_rq->on_list = 0; | |
317 | } | |
318 | } | |
319 | ||
b758149c PZ |
320 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
321 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
322 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
323 | ||
324 | /* Do the two (enqueued) entities belong to the same group ? */ | |
325 | static inline int | |
326 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
327 | { | |
328 | if (se->cfs_rq == pse->cfs_rq) | |
329 | return 1; | |
330 | ||
331 | return 0; | |
332 | } | |
333 | ||
334 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
335 | { | |
336 | return se->parent; | |
337 | } | |
338 | ||
464b7527 PZ |
339 | /* return depth at which a sched entity is present in the hierarchy */ |
340 | static inline int depth_se(struct sched_entity *se) | |
341 | { | |
342 | int depth = 0; | |
343 | ||
344 | for_each_sched_entity(se) | |
345 | depth++; | |
346 | ||
347 | return depth; | |
348 | } | |
349 | ||
350 | static void | |
351 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
352 | { | |
353 | int se_depth, pse_depth; | |
354 | ||
355 | /* | |
356 | * preemption test can be made between sibling entities who are in the | |
357 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
358 | * both tasks until we find their ancestors who are siblings of common | |
359 | * parent. | |
360 | */ | |
361 | ||
362 | /* First walk up until both entities are at same depth */ | |
363 | se_depth = depth_se(*se); | |
364 | pse_depth = depth_se(*pse); | |
365 | ||
366 | while (se_depth > pse_depth) { | |
367 | se_depth--; | |
368 | *se = parent_entity(*se); | |
369 | } | |
370 | ||
371 | while (pse_depth > se_depth) { | |
372 | pse_depth--; | |
373 | *pse = parent_entity(*pse); | |
374 | } | |
375 | ||
376 | while (!is_same_group(*se, *pse)) { | |
377 | *se = parent_entity(*se); | |
378 | *pse = parent_entity(*pse); | |
379 | } | |
380 | } | |
381 | ||
8f48894f PZ |
382 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
383 | ||
384 | static inline struct task_struct *task_of(struct sched_entity *se) | |
385 | { | |
386 | return container_of(se, struct task_struct, se); | |
387 | } | |
bf0f6f24 | 388 | |
62160e3f IM |
389 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
390 | { | |
391 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
392 | } |
393 | ||
394 | #define entity_is_task(se) 1 | |
395 | ||
b758149c PZ |
396 | #define for_each_sched_entity(se) \ |
397 | for (; se; se = NULL) | |
bf0f6f24 | 398 | |
b758149c | 399 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 400 | { |
b758149c | 401 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
402 | } |
403 | ||
b758149c PZ |
404 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
405 | { | |
406 | struct task_struct *p = task_of(se); | |
407 | struct rq *rq = task_rq(p); | |
408 | ||
409 | return &rq->cfs; | |
410 | } | |
411 | ||
412 | /* runqueue "owned" by this group */ | |
413 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
414 | { | |
415 | return NULL; | |
416 | } | |
417 | ||
3d4b47b4 PZ |
418 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
419 | { | |
420 | } | |
421 | ||
422 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
423 | { | |
424 | } | |
425 | ||
b758149c PZ |
426 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
427 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
428 | ||
429 | static inline int | |
430 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
431 | { | |
432 | return 1; | |
433 | } | |
434 | ||
435 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
436 | { | |
437 | return NULL; | |
438 | } | |
439 | ||
464b7527 PZ |
440 | static inline void |
441 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
442 | { | |
443 | } | |
444 | ||
b758149c PZ |
445 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
446 | ||
6c16a6dc | 447 | static __always_inline |
9dbdb155 | 448 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
449 | |
450 | /************************************************************** | |
451 | * Scheduling class tree data structure manipulation methods: | |
452 | */ | |
453 | ||
1bf08230 | 454 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 455 | { |
1bf08230 | 456 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 457 | if (delta > 0) |
1bf08230 | 458 | max_vruntime = vruntime; |
02e0431a | 459 | |
1bf08230 | 460 | return max_vruntime; |
02e0431a PZ |
461 | } |
462 | ||
0702e3eb | 463 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
464 | { |
465 | s64 delta = (s64)(vruntime - min_vruntime); | |
466 | if (delta < 0) | |
467 | min_vruntime = vruntime; | |
468 | ||
469 | return min_vruntime; | |
470 | } | |
471 | ||
54fdc581 FC |
472 | static inline int entity_before(struct sched_entity *a, |
473 | struct sched_entity *b) | |
474 | { | |
475 | return (s64)(a->vruntime - b->vruntime) < 0; | |
476 | } | |
477 | ||
1af5f730 PZ |
478 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
479 | { | |
480 | u64 vruntime = cfs_rq->min_vruntime; | |
481 | ||
482 | if (cfs_rq->curr) | |
483 | vruntime = cfs_rq->curr->vruntime; | |
484 | ||
485 | if (cfs_rq->rb_leftmost) { | |
486 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
487 | struct sched_entity, | |
488 | run_node); | |
489 | ||
e17036da | 490 | if (!cfs_rq->curr) |
1af5f730 PZ |
491 | vruntime = se->vruntime; |
492 | else | |
493 | vruntime = min_vruntime(vruntime, se->vruntime); | |
494 | } | |
495 | ||
1bf08230 | 496 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 497 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
498 | #ifndef CONFIG_64BIT |
499 | smp_wmb(); | |
500 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
501 | #endif | |
1af5f730 PZ |
502 | } |
503 | ||
bf0f6f24 IM |
504 | /* |
505 | * Enqueue an entity into the rb-tree: | |
506 | */ | |
0702e3eb | 507 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
508 | { |
509 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
510 | struct rb_node *parent = NULL; | |
511 | struct sched_entity *entry; | |
bf0f6f24 IM |
512 | int leftmost = 1; |
513 | ||
514 | /* | |
515 | * Find the right place in the rbtree: | |
516 | */ | |
517 | while (*link) { | |
518 | parent = *link; | |
519 | entry = rb_entry(parent, struct sched_entity, run_node); | |
520 | /* | |
521 | * We dont care about collisions. Nodes with | |
522 | * the same key stay together. | |
523 | */ | |
2bd2d6f2 | 524 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
525 | link = &parent->rb_left; |
526 | } else { | |
527 | link = &parent->rb_right; | |
528 | leftmost = 0; | |
529 | } | |
530 | } | |
531 | ||
532 | /* | |
533 | * Maintain a cache of leftmost tree entries (it is frequently | |
534 | * used): | |
535 | */ | |
1af5f730 | 536 | if (leftmost) |
57cb499d | 537 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
538 | |
539 | rb_link_node(&se->run_node, parent, link); | |
540 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
541 | } |
542 | ||
0702e3eb | 543 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 544 | { |
3fe69747 PZ |
545 | if (cfs_rq->rb_leftmost == &se->run_node) { |
546 | struct rb_node *next_node; | |
3fe69747 PZ |
547 | |
548 | next_node = rb_next(&se->run_node); | |
549 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 550 | } |
e9acbff6 | 551 | |
bf0f6f24 | 552 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
553 | } |
554 | ||
029632fb | 555 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 556 | { |
f4b6755f PZ |
557 | struct rb_node *left = cfs_rq->rb_leftmost; |
558 | ||
559 | if (!left) | |
560 | return NULL; | |
561 | ||
562 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
563 | } |
564 | ||
ac53db59 RR |
565 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
566 | { | |
567 | struct rb_node *next = rb_next(&se->run_node); | |
568 | ||
569 | if (!next) | |
570 | return NULL; | |
571 | ||
572 | return rb_entry(next, struct sched_entity, run_node); | |
573 | } | |
574 | ||
575 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 576 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 577 | { |
7eee3e67 | 578 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 579 | |
70eee74b BS |
580 | if (!last) |
581 | return NULL; | |
7eee3e67 IM |
582 | |
583 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
584 | } |
585 | ||
bf0f6f24 IM |
586 | /************************************************************** |
587 | * Scheduling class statistics methods: | |
588 | */ | |
589 | ||
acb4a848 | 590 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 591 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
592 | loff_t *ppos) |
593 | { | |
8d65af78 | 594 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 595 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
596 | |
597 | if (ret || !write) | |
598 | return ret; | |
599 | ||
600 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
601 | sysctl_sched_min_granularity); | |
602 | ||
acb4a848 CE |
603 | #define WRT_SYSCTL(name) \ |
604 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
605 | WRT_SYSCTL(sched_min_granularity); | |
606 | WRT_SYSCTL(sched_latency); | |
607 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
608 | #undef WRT_SYSCTL |
609 | ||
b2be5e96 PZ |
610 | return 0; |
611 | } | |
612 | #endif | |
647e7cac | 613 | |
a7be37ac | 614 | /* |
f9c0b095 | 615 | * delta /= w |
a7be37ac | 616 | */ |
9dbdb155 | 617 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 618 | { |
f9c0b095 | 619 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 620 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
621 | |
622 | return delta; | |
623 | } | |
624 | ||
647e7cac IM |
625 | /* |
626 | * The idea is to set a period in which each task runs once. | |
627 | * | |
532b1858 | 628 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
629 | * this period because otherwise the slices get too small. |
630 | * | |
631 | * p = (nr <= nl) ? l : l*nr/nl | |
632 | */ | |
4d78e7b6 PZ |
633 | static u64 __sched_period(unsigned long nr_running) |
634 | { | |
635 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 636 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
637 | |
638 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 639 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 640 | period *= nr_running; |
4d78e7b6 PZ |
641 | } |
642 | ||
643 | return period; | |
644 | } | |
645 | ||
647e7cac IM |
646 | /* |
647 | * We calculate the wall-time slice from the period by taking a part | |
648 | * proportional to the weight. | |
649 | * | |
f9c0b095 | 650 | * s = p*P[w/rw] |
647e7cac | 651 | */ |
6d0f0ebd | 652 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 653 | { |
0a582440 | 654 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 655 | |
0a582440 | 656 | for_each_sched_entity(se) { |
6272d68c | 657 | struct load_weight *load; |
3104bf03 | 658 | struct load_weight lw; |
6272d68c LM |
659 | |
660 | cfs_rq = cfs_rq_of(se); | |
661 | load = &cfs_rq->load; | |
f9c0b095 | 662 | |
0a582440 | 663 | if (unlikely(!se->on_rq)) { |
3104bf03 | 664 | lw = cfs_rq->load; |
0a582440 MG |
665 | |
666 | update_load_add(&lw, se->load.weight); | |
667 | load = &lw; | |
668 | } | |
9dbdb155 | 669 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
670 | } |
671 | return slice; | |
bf0f6f24 IM |
672 | } |
673 | ||
647e7cac | 674 | /* |
660cc00f | 675 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 676 | * |
f9c0b095 | 677 | * vs = s/w |
647e7cac | 678 | */ |
f9c0b095 | 679 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 680 | { |
f9c0b095 | 681 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
682 | } |
683 | ||
a75cdaa9 | 684 | #ifdef CONFIG_SMP |
fb13c7ee MG |
685 | static unsigned long task_h_load(struct task_struct *p); |
686 | ||
a75cdaa9 AS |
687 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
688 | ||
689 | /* Give new task start runnable values to heavy its load in infant time */ | |
690 | void init_task_runnable_average(struct task_struct *p) | |
691 | { | |
692 | u32 slice; | |
693 | ||
694 | p->se.avg.decay_count = 0; | |
695 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | |
696 | p->se.avg.runnable_avg_sum = slice; | |
697 | p->se.avg.runnable_avg_period = slice; | |
698 | __update_task_entity_contrib(&p->se); | |
699 | } | |
700 | #else | |
701 | void init_task_runnable_average(struct task_struct *p) | |
702 | { | |
703 | } | |
704 | #endif | |
705 | ||
bf0f6f24 | 706 | /* |
9dbdb155 | 707 | * Update the current task's runtime statistics. |
bf0f6f24 | 708 | */ |
b7cc0896 | 709 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 710 | { |
429d43bc | 711 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 712 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 713 | u64 delta_exec; |
bf0f6f24 IM |
714 | |
715 | if (unlikely(!curr)) | |
716 | return; | |
717 | ||
9dbdb155 PZ |
718 | delta_exec = now - curr->exec_start; |
719 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 720 | return; |
bf0f6f24 | 721 | |
8ebc91d9 | 722 | curr->exec_start = now; |
d842de87 | 723 | |
9dbdb155 PZ |
724 | schedstat_set(curr->statistics.exec_max, |
725 | max(delta_exec, curr->statistics.exec_max)); | |
726 | ||
727 | curr->sum_exec_runtime += delta_exec; | |
728 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
729 | ||
730 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
731 | update_min_vruntime(cfs_rq); | |
732 | ||
d842de87 SV |
733 | if (entity_is_task(curr)) { |
734 | struct task_struct *curtask = task_of(curr); | |
735 | ||
f977bb49 | 736 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 737 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 738 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 739 | } |
ec12cb7f PT |
740 | |
741 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
742 | } |
743 | ||
744 | static inline void | |
5870db5b | 745 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 746 | { |
78becc27 | 747 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
748 | } |
749 | ||
bf0f6f24 IM |
750 | /* |
751 | * Task is being enqueued - update stats: | |
752 | */ | |
d2417e5a | 753 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 754 | { |
bf0f6f24 IM |
755 | /* |
756 | * Are we enqueueing a waiting task? (for current tasks | |
757 | * a dequeue/enqueue event is a NOP) | |
758 | */ | |
429d43bc | 759 | if (se != cfs_rq->curr) |
5870db5b | 760 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
761 | } |
762 | ||
bf0f6f24 | 763 | static void |
9ef0a961 | 764 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 765 | { |
41acab88 | 766 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 767 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
768 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
769 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 770 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
771 | #ifdef CONFIG_SCHEDSTATS |
772 | if (entity_is_task(se)) { | |
773 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 774 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
775 | } |
776 | #endif | |
41acab88 | 777 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
778 | } |
779 | ||
780 | static inline void | |
19b6a2e3 | 781 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 782 | { |
bf0f6f24 IM |
783 | /* |
784 | * Mark the end of the wait period if dequeueing a | |
785 | * waiting task: | |
786 | */ | |
429d43bc | 787 | if (se != cfs_rq->curr) |
9ef0a961 | 788 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
789 | } |
790 | ||
791 | /* | |
792 | * We are picking a new current task - update its stats: | |
793 | */ | |
794 | static inline void | |
79303e9e | 795 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
796 | { |
797 | /* | |
798 | * We are starting a new run period: | |
799 | */ | |
78becc27 | 800 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
801 | } |
802 | ||
bf0f6f24 IM |
803 | /************************************************** |
804 | * Scheduling class queueing methods: | |
805 | */ | |
806 | ||
cbee9f88 PZ |
807 | #ifdef CONFIG_NUMA_BALANCING |
808 | /* | |
598f0ec0 MG |
809 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
810 | * calculated based on the tasks virtual memory size and | |
811 | * numa_balancing_scan_size. | |
cbee9f88 | 812 | */ |
598f0ec0 MG |
813 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
814 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
815 | |
816 | /* Portion of address space to scan in MB */ | |
817 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 818 | |
4b96a29b PZ |
819 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
820 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
821 | ||
598f0ec0 MG |
822 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
823 | { | |
824 | unsigned long rss = 0; | |
825 | unsigned long nr_scan_pages; | |
826 | ||
827 | /* | |
828 | * Calculations based on RSS as non-present and empty pages are skipped | |
829 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
830 | * on resident pages | |
831 | */ | |
832 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
833 | rss = get_mm_rss(p->mm); | |
834 | if (!rss) | |
835 | rss = nr_scan_pages; | |
836 | ||
837 | rss = round_up(rss, nr_scan_pages); | |
838 | return rss / nr_scan_pages; | |
839 | } | |
840 | ||
841 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
842 | #define MAX_SCAN_WINDOW 2560 | |
843 | ||
844 | static unsigned int task_scan_min(struct task_struct *p) | |
845 | { | |
846 | unsigned int scan, floor; | |
847 | unsigned int windows = 1; | |
848 | ||
849 | if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW) | |
850 | windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size; | |
851 | floor = 1000 / windows; | |
852 | ||
853 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
854 | return max_t(unsigned int, floor, scan); | |
855 | } | |
856 | ||
857 | static unsigned int task_scan_max(struct task_struct *p) | |
858 | { | |
859 | unsigned int smin = task_scan_min(p); | |
860 | unsigned int smax; | |
861 | ||
862 | /* Watch for min being lower than max due to floor calculations */ | |
863 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
864 | return max(smin, smax); | |
865 | } | |
866 | ||
0ec8aa00 PZ |
867 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
868 | { | |
869 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
870 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
871 | } | |
872 | ||
873 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
874 | { | |
875 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
876 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
877 | } | |
878 | ||
8c8a743c PZ |
879 | struct numa_group { |
880 | atomic_t refcount; | |
881 | ||
882 | spinlock_t lock; /* nr_tasks, tasks */ | |
883 | int nr_tasks; | |
e29cf08b | 884 | pid_t gid; |
8c8a743c PZ |
885 | struct list_head task_list; |
886 | ||
887 | struct rcu_head rcu; | |
989348b5 | 888 | unsigned long total_faults; |
50ec8a40 | 889 | unsigned long *faults_cpu; |
989348b5 | 890 | unsigned long faults[0]; |
8c8a743c PZ |
891 | }; |
892 | ||
e29cf08b MG |
893 | pid_t task_numa_group_id(struct task_struct *p) |
894 | { | |
895 | return p->numa_group ? p->numa_group->gid : 0; | |
896 | } | |
897 | ||
ac8e895b MG |
898 | static inline int task_faults_idx(int nid, int priv) |
899 | { | |
900 | return 2 * nid + priv; | |
901 | } | |
902 | ||
903 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
904 | { | |
ff1df896 | 905 | if (!p->numa_faults_memory) |
ac8e895b MG |
906 | return 0; |
907 | ||
ff1df896 RR |
908 | return p->numa_faults_memory[task_faults_idx(nid, 0)] + |
909 | p->numa_faults_memory[task_faults_idx(nid, 1)]; | |
ac8e895b MG |
910 | } |
911 | ||
83e1d2cd MG |
912 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
913 | { | |
914 | if (!p->numa_group) | |
915 | return 0; | |
916 | ||
82897b4f WL |
917 | return p->numa_group->faults[task_faults_idx(nid, 0)] + |
918 | p->numa_group->faults[task_faults_idx(nid, 1)]; | |
83e1d2cd MG |
919 | } |
920 | ||
921 | /* | |
922 | * These return the fraction of accesses done by a particular task, or | |
923 | * task group, on a particular numa node. The group weight is given a | |
924 | * larger multiplier, in order to group tasks together that are almost | |
925 | * evenly spread out between numa nodes. | |
926 | */ | |
927 | static inline unsigned long task_weight(struct task_struct *p, int nid) | |
928 | { | |
929 | unsigned long total_faults; | |
930 | ||
ff1df896 | 931 | if (!p->numa_faults_memory) |
83e1d2cd MG |
932 | return 0; |
933 | ||
934 | total_faults = p->total_numa_faults; | |
935 | ||
936 | if (!total_faults) | |
937 | return 0; | |
938 | ||
939 | return 1000 * task_faults(p, nid) / total_faults; | |
940 | } | |
941 | ||
942 | static inline unsigned long group_weight(struct task_struct *p, int nid) | |
943 | { | |
989348b5 | 944 | if (!p->numa_group || !p->numa_group->total_faults) |
83e1d2cd MG |
945 | return 0; |
946 | ||
989348b5 | 947 | return 1000 * group_faults(p, nid) / p->numa_group->total_faults; |
83e1d2cd MG |
948 | } |
949 | ||
e6628d5b | 950 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
951 | static unsigned long source_load(int cpu, int type); |
952 | static unsigned long target_load(int cpu, int type); | |
953 | static unsigned long power_of(int cpu); | |
954 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
955 | ||
fb13c7ee | 956 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 957 | struct numa_stats { |
fb13c7ee | 958 | unsigned long nr_running; |
58d081b5 | 959 | unsigned long load; |
fb13c7ee MG |
960 | |
961 | /* Total compute capacity of CPUs on a node */ | |
962 | unsigned long power; | |
963 | ||
964 | /* Approximate capacity in terms of runnable tasks on a node */ | |
965 | unsigned long capacity; | |
966 | int has_capacity; | |
58d081b5 | 967 | }; |
e6628d5b | 968 | |
fb13c7ee MG |
969 | /* |
970 | * XXX borrowed from update_sg_lb_stats | |
971 | */ | |
972 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
973 | { | |
5eca82a9 | 974 | int cpu, cpus = 0; |
fb13c7ee MG |
975 | |
976 | memset(ns, 0, sizeof(*ns)); | |
977 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
978 | struct rq *rq = cpu_rq(cpu); | |
979 | ||
980 | ns->nr_running += rq->nr_running; | |
981 | ns->load += weighted_cpuload(cpu); | |
982 | ns->power += power_of(cpu); | |
5eca82a9 PZ |
983 | |
984 | cpus++; | |
fb13c7ee MG |
985 | } |
986 | ||
5eca82a9 PZ |
987 | /* |
988 | * If we raced with hotplug and there are no CPUs left in our mask | |
989 | * the @ns structure is NULL'ed and task_numa_compare() will | |
990 | * not find this node attractive. | |
991 | * | |
992 | * We'll either bail at !has_capacity, or we'll detect a huge imbalance | |
993 | * and bail there. | |
994 | */ | |
995 | if (!cpus) | |
996 | return; | |
997 | ||
fb13c7ee MG |
998 | ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power; |
999 | ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE); | |
1000 | ns->has_capacity = (ns->nr_running < ns->capacity); | |
1001 | } | |
1002 | ||
58d081b5 MG |
1003 | struct task_numa_env { |
1004 | struct task_struct *p; | |
e6628d5b | 1005 | |
58d081b5 MG |
1006 | int src_cpu, src_nid; |
1007 | int dst_cpu, dst_nid; | |
e6628d5b | 1008 | |
58d081b5 | 1009 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1010 | |
40ea2b42 | 1011 | int imbalance_pct; |
fb13c7ee MG |
1012 | |
1013 | struct task_struct *best_task; | |
1014 | long best_imp; | |
58d081b5 MG |
1015 | int best_cpu; |
1016 | }; | |
1017 | ||
fb13c7ee MG |
1018 | static void task_numa_assign(struct task_numa_env *env, |
1019 | struct task_struct *p, long imp) | |
1020 | { | |
1021 | if (env->best_task) | |
1022 | put_task_struct(env->best_task); | |
1023 | if (p) | |
1024 | get_task_struct(p); | |
1025 | ||
1026 | env->best_task = p; | |
1027 | env->best_imp = imp; | |
1028 | env->best_cpu = env->dst_cpu; | |
1029 | } | |
1030 | ||
1031 | /* | |
1032 | * This checks if the overall compute and NUMA accesses of the system would | |
1033 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1034 | * into account that it might be best if task running on the dst_cpu should | |
1035 | * be exchanged with the source task | |
1036 | */ | |
887c290e RR |
1037 | static void task_numa_compare(struct task_numa_env *env, |
1038 | long taskimp, long groupimp) | |
fb13c7ee MG |
1039 | { |
1040 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1041 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1042 | struct task_struct *cur; | |
1043 | long dst_load, src_load; | |
1044 | long load; | |
887c290e | 1045 | long imp = (groupimp > 0) ? groupimp : taskimp; |
fb13c7ee MG |
1046 | |
1047 | rcu_read_lock(); | |
1048 | cur = ACCESS_ONCE(dst_rq->curr); | |
1049 | if (cur->pid == 0) /* idle */ | |
1050 | cur = NULL; | |
1051 | ||
1052 | /* | |
1053 | * "imp" is the fault differential for the source task between the | |
1054 | * source and destination node. Calculate the total differential for | |
1055 | * the source task and potential destination task. The more negative | |
1056 | * the value is, the more rmeote accesses that would be expected to | |
1057 | * be incurred if the tasks were swapped. | |
1058 | */ | |
1059 | if (cur) { | |
1060 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1061 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1062 | goto unlock; | |
1063 | ||
887c290e RR |
1064 | /* |
1065 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1066 | * in any group then look only at task weights. |
887c290e | 1067 | */ |
ca28aa53 | 1068 | if (cur->numa_group == env->p->numa_group) { |
887c290e RR |
1069 | imp = taskimp + task_weight(cur, env->src_nid) - |
1070 | task_weight(cur, env->dst_nid); | |
ca28aa53 RR |
1071 | /* |
1072 | * Add some hysteresis to prevent swapping the | |
1073 | * tasks within a group over tiny differences. | |
1074 | */ | |
1075 | if (cur->numa_group) | |
1076 | imp -= imp/16; | |
887c290e | 1077 | } else { |
ca28aa53 RR |
1078 | /* |
1079 | * Compare the group weights. If a task is all by | |
1080 | * itself (not part of a group), use the task weight | |
1081 | * instead. | |
1082 | */ | |
1083 | if (env->p->numa_group) | |
1084 | imp = groupimp; | |
1085 | else | |
1086 | imp = taskimp; | |
1087 | ||
1088 | if (cur->numa_group) | |
1089 | imp += group_weight(cur, env->src_nid) - | |
1090 | group_weight(cur, env->dst_nid); | |
1091 | else | |
1092 | imp += task_weight(cur, env->src_nid) - | |
1093 | task_weight(cur, env->dst_nid); | |
887c290e | 1094 | } |
fb13c7ee MG |
1095 | } |
1096 | ||
1097 | if (imp < env->best_imp) | |
1098 | goto unlock; | |
1099 | ||
1100 | if (!cur) { | |
1101 | /* Is there capacity at our destination? */ | |
1102 | if (env->src_stats.has_capacity && | |
1103 | !env->dst_stats.has_capacity) | |
1104 | goto unlock; | |
1105 | ||
1106 | goto balance; | |
1107 | } | |
1108 | ||
1109 | /* Balance doesn't matter much if we're running a task per cpu */ | |
1110 | if (src_rq->nr_running == 1 && dst_rq->nr_running == 1) | |
1111 | goto assign; | |
1112 | ||
1113 | /* | |
1114 | * In the overloaded case, try and keep the load balanced. | |
1115 | */ | |
1116 | balance: | |
1117 | dst_load = env->dst_stats.load; | |
1118 | src_load = env->src_stats.load; | |
1119 | ||
1120 | /* XXX missing power terms */ | |
1121 | load = task_h_load(env->p); | |
1122 | dst_load += load; | |
1123 | src_load -= load; | |
1124 | ||
1125 | if (cur) { | |
1126 | load = task_h_load(cur); | |
1127 | dst_load -= load; | |
1128 | src_load += load; | |
1129 | } | |
1130 | ||
1131 | /* make src_load the smaller */ | |
1132 | if (dst_load < src_load) | |
1133 | swap(dst_load, src_load); | |
1134 | ||
1135 | if (src_load * env->imbalance_pct < dst_load * 100) | |
1136 | goto unlock; | |
1137 | ||
1138 | assign: | |
1139 | task_numa_assign(env, cur, imp); | |
1140 | unlock: | |
1141 | rcu_read_unlock(); | |
1142 | } | |
1143 | ||
887c290e RR |
1144 | static void task_numa_find_cpu(struct task_numa_env *env, |
1145 | long taskimp, long groupimp) | |
2c8a50aa MG |
1146 | { |
1147 | int cpu; | |
1148 | ||
1149 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1150 | /* Skip this CPU if the source task cannot migrate */ | |
1151 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1152 | continue; | |
1153 | ||
1154 | env->dst_cpu = cpu; | |
887c290e | 1155 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1156 | } |
1157 | } | |
1158 | ||
58d081b5 MG |
1159 | static int task_numa_migrate(struct task_struct *p) |
1160 | { | |
58d081b5 MG |
1161 | struct task_numa_env env = { |
1162 | .p = p, | |
fb13c7ee | 1163 | |
58d081b5 | 1164 | .src_cpu = task_cpu(p), |
b32e86b4 | 1165 | .src_nid = task_node(p), |
fb13c7ee MG |
1166 | |
1167 | .imbalance_pct = 112, | |
1168 | ||
1169 | .best_task = NULL, | |
1170 | .best_imp = 0, | |
1171 | .best_cpu = -1 | |
58d081b5 MG |
1172 | }; |
1173 | struct sched_domain *sd; | |
887c290e | 1174 | unsigned long taskweight, groupweight; |
2c8a50aa | 1175 | int nid, ret; |
887c290e | 1176 | long taskimp, groupimp; |
e6628d5b | 1177 | |
58d081b5 | 1178 | /* |
fb13c7ee MG |
1179 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1180 | * imbalance and would be the first to start moving tasks about. | |
1181 | * | |
1182 | * And we want to avoid any moving of tasks about, as that would create | |
1183 | * random movement of tasks -- counter the numa conditions we're trying | |
1184 | * to satisfy here. | |
58d081b5 MG |
1185 | */ |
1186 | rcu_read_lock(); | |
fb13c7ee | 1187 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1188 | if (sd) |
1189 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1190 | rcu_read_unlock(); |
1191 | ||
46a73e8a RR |
1192 | /* |
1193 | * Cpusets can break the scheduler domain tree into smaller | |
1194 | * balance domains, some of which do not cross NUMA boundaries. | |
1195 | * Tasks that are "trapped" in such domains cannot be migrated | |
1196 | * elsewhere, so there is no point in (re)trying. | |
1197 | */ | |
1198 | if (unlikely(!sd)) { | |
de1b301a | 1199 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1200 | return -EINVAL; |
1201 | } | |
1202 | ||
887c290e RR |
1203 | taskweight = task_weight(p, env.src_nid); |
1204 | groupweight = group_weight(p, env.src_nid); | |
fb13c7ee | 1205 | update_numa_stats(&env.src_stats, env.src_nid); |
2c8a50aa | 1206 | env.dst_nid = p->numa_preferred_nid; |
887c290e RR |
1207 | taskimp = task_weight(p, env.dst_nid) - taskweight; |
1208 | groupimp = group_weight(p, env.dst_nid) - groupweight; | |
2c8a50aa | 1209 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1210 | |
e1dda8a7 RR |
1211 | /* If the preferred nid has capacity, try to use it. */ |
1212 | if (env.dst_stats.has_capacity) | |
887c290e | 1213 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 RR |
1214 | |
1215 | /* No space available on the preferred nid. Look elsewhere. */ | |
1216 | if (env.best_cpu == -1) { | |
2c8a50aa MG |
1217 | for_each_online_node(nid) { |
1218 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1219 | continue; | |
58d081b5 | 1220 | |
83e1d2cd | 1221 | /* Only consider nodes where both task and groups benefit */ |
887c290e RR |
1222 | taskimp = task_weight(p, nid) - taskweight; |
1223 | groupimp = group_weight(p, nid) - groupweight; | |
1224 | if (taskimp < 0 && groupimp < 0) | |
fb13c7ee MG |
1225 | continue; |
1226 | ||
2c8a50aa MG |
1227 | env.dst_nid = nid; |
1228 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
887c290e | 1229 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1230 | } |
1231 | } | |
1232 | ||
fb13c7ee MG |
1233 | /* No better CPU than the current one was found. */ |
1234 | if (env.best_cpu == -1) | |
1235 | return -EAGAIN; | |
1236 | ||
0ec8aa00 PZ |
1237 | sched_setnuma(p, env.dst_nid); |
1238 | ||
04bb2f94 RR |
1239 | /* |
1240 | * Reset the scan period if the task is being rescheduled on an | |
1241 | * alternative node to recheck if the tasks is now properly placed. | |
1242 | */ | |
1243 | p->numa_scan_period = task_scan_min(p); | |
1244 | ||
fb13c7ee MG |
1245 | if (env.best_task == NULL) { |
1246 | int ret = migrate_task_to(p, env.best_cpu); | |
1247 | return ret; | |
1248 | } | |
1249 | ||
1250 | ret = migrate_swap(p, env.best_task); | |
1251 | put_task_struct(env.best_task); | |
1252 | return ret; | |
e6628d5b MG |
1253 | } |
1254 | ||
6b9a7460 MG |
1255 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1256 | static void numa_migrate_preferred(struct task_struct *p) | |
1257 | { | |
2739d3ee | 1258 | /* This task has no NUMA fault statistics yet */ |
ff1df896 | 1259 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory)) |
6b9a7460 MG |
1260 | return; |
1261 | ||
2739d3ee RR |
1262 | /* Periodically retry migrating the task to the preferred node */ |
1263 | p->numa_migrate_retry = jiffies + HZ; | |
1264 | ||
1265 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1266 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1267 | return; |
1268 | ||
1269 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1270 | task_numa_migrate(p); |
6b9a7460 MG |
1271 | } |
1272 | ||
04bb2f94 RR |
1273 | /* |
1274 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1275 | * increments. The more local the fault statistics are, the higher the scan | |
1276 | * period will be for the next scan window. If local/remote ratio is below | |
1277 | * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the | |
1278 | * scan period will decrease | |
1279 | */ | |
1280 | #define NUMA_PERIOD_SLOTS 10 | |
1281 | #define NUMA_PERIOD_THRESHOLD 3 | |
1282 | ||
1283 | /* | |
1284 | * Increase the scan period (slow down scanning) if the majority of | |
1285 | * our memory is already on our local node, or if the majority of | |
1286 | * the page accesses are shared with other processes. | |
1287 | * Otherwise, decrease the scan period. | |
1288 | */ | |
1289 | static void update_task_scan_period(struct task_struct *p, | |
1290 | unsigned long shared, unsigned long private) | |
1291 | { | |
1292 | unsigned int period_slot; | |
1293 | int ratio; | |
1294 | int diff; | |
1295 | ||
1296 | unsigned long remote = p->numa_faults_locality[0]; | |
1297 | unsigned long local = p->numa_faults_locality[1]; | |
1298 | ||
1299 | /* | |
1300 | * If there were no record hinting faults then either the task is | |
1301 | * completely idle or all activity is areas that are not of interest | |
1302 | * to automatic numa balancing. Scan slower | |
1303 | */ | |
1304 | if (local + shared == 0) { | |
1305 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1306 | p->numa_scan_period << 1); | |
1307 | ||
1308 | p->mm->numa_next_scan = jiffies + | |
1309 | msecs_to_jiffies(p->numa_scan_period); | |
1310 | ||
1311 | return; | |
1312 | } | |
1313 | ||
1314 | /* | |
1315 | * Prepare to scale scan period relative to the current period. | |
1316 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1317 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1318 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1319 | */ | |
1320 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1321 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1322 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1323 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1324 | if (!slot) | |
1325 | slot = 1; | |
1326 | diff = slot * period_slot; | |
1327 | } else { | |
1328 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1329 | ||
1330 | /* | |
1331 | * Scale scan rate increases based on sharing. There is an | |
1332 | * inverse relationship between the degree of sharing and | |
1333 | * the adjustment made to the scanning period. Broadly | |
1334 | * speaking the intent is that there is little point | |
1335 | * scanning faster if shared accesses dominate as it may | |
1336 | * simply bounce migrations uselessly | |
1337 | */ | |
04bb2f94 RR |
1338 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared)); |
1339 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; | |
1340 | } | |
1341 | ||
1342 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1343 | task_scan_min(p), task_scan_max(p)); | |
1344 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1345 | } | |
1346 | ||
cbee9f88 PZ |
1347 | static void task_numa_placement(struct task_struct *p) |
1348 | { | |
83e1d2cd MG |
1349 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1350 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 1351 | unsigned long fault_types[2] = { 0, 0 }; |
7dbd13ed | 1352 | spinlock_t *group_lock = NULL; |
cbee9f88 | 1353 | |
2832bc19 | 1354 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1355 | if (p->numa_scan_seq == seq) |
1356 | return; | |
1357 | p->numa_scan_seq = seq; | |
598f0ec0 | 1358 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1359 | |
7dbd13ed MG |
1360 | /* If the task is part of a group prevent parallel updates to group stats */ |
1361 | if (p->numa_group) { | |
1362 | group_lock = &p->numa_group->lock; | |
1363 | spin_lock(group_lock); | |
1364 | } | |
1365 | ||
688b7585 MG |
1366 | /* Find the node with the highest number of faults */ |
1367 | for_each_online_node(nid) { | |
83e1d2cd | 1368 | unsigned long faults = 0, group_faults = 0; |
ac8e895b | 1369 | int priv, i; |
745d6147 | 1370 | |
ac8e895b | 1371 | for (priv = 0; priv < 2; priv++) { |
50ec8a40 | 1372 | long diff, f_diff; |
8c8a743c | 1373 | |
ac8e895b | 1374 | i = task_faults_idx(nid, priv); |
ff1df896 | 1375 | diff = -p->numa_faults_memory[i]; |
50ec8a40 | 1376 | f_diff = -p->numa_faults_cpu[i]; |
745d6147 | 1377 | |
ac8e895b | 1378 | /* Decay existing window, copy faults since last scan */ |
ff1df896 RR |
1379 | p->numa_faults_memory[i] >>= 1; |
1380 | p->numa_faults_memory[i] += p->numa_faults_buffer_memory[i]; | |
1381 | fault_types[priv] += p->numa_faults_buffer_memory[i]; | |
1382 | p->numa_faults_buffer_memory[i] = 0; | |
fb13c7ee | 1383 | |
50ec8a40 RR |
1384 | p->numa_faults_cpu[i] >>= 1; |
1385 | p->numa_faults_cpu[i] += p->numa_faults_buffer_cpu[i]; | |
1386 | p->numa_faults_buffer_cpu[i] = 0; | |
1387 | ||
ff1df896 RR |
1388 | faults += p->numa_faults_memory[i]; |
1389 | diff += p->numa_faults_memory[i]; | |
50ec8a40 | 1390 | f_diff += p->numa_faults_cpu[i]; |
83e1d2cd | 1391 | p->total_numa_faults += diff; |
8c8a743c PZ |
1392 | if (p->numa_group) { |
1393 | /* safe because we can only change our own group */ | |
989348b5 | 1394 | p->numa_group->faults[i] += diff; |
50ec8a40 | 1395 | p->numa_group->faults_cpu[i] += f_diff; |
989348b5 MG |
1396 | p->numa_group->total_faults += diff; |
1397 | group_faults += p->numa_group->faults[i]; | |
8c8a743c | 1398 | } |
ac8e895b MG |
1399 | } |
1400 | ||
688b7585 MG |
1401 | if (faults > max_faults) { |
1402 | max_faults = faults; | |
1403 | max_nid = nid; | |
1404 | } | |
83e1d2cd MG |
1405 | |
1406 | if (group_faults > max_group_faults) { | |
1407 | max_group_faults = group_faults; | |
1408 | max_group_nid = nid; | |
1409 | } | |
1410 | } | |
1411 | ||
04bb2f94 RR |
1412 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1413 | ||
7dbd13ed MG |
1414 | if (p->numa_group) { |
1415 | /* | |
1416 | * If the preferred task and group nids are different, | |
1417 | * iterate over the nodes again to find the best place. | |
1418 | */ | |
1419 | if (max_nid != max_group_nid) { | |
1420 | unsigned long weight, max_weight = 0; | |
1421 | ||
1422 | for_each_online_node(nid) { | |
1423 | weight = task_weight(p, nid) + group_weight(p, nid); | |
1424 | if (weight > max_weight) { | |
1425 | max_weight = weight; | |
1426 | max_nid = nid; | |
1427 | } | |
83e1d2cd MG |
1428 | } |
1429 | } | |
7dbd13ed MG |
1430 | |
1431 | spin_unlock(group_lock); | |
688b7585 MG |
1432 | } |
1433 | ||
6b9a7460 | 1434 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1435 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1436 | /* Update the preferred nid and migrate task if possible */ |
0ec8aa00 | 1437 | sched_setnuma(p, max_nid); |
6b9a7460 | 1438 | numa_migrate_preferred(p); |
3a7053b3 | 1439 | } |
cbee9f88 PZ |
1440 | } |
1441 | ||
8c8a743c PZ |
1442 | static inline int get_numa_group(struct numa_group *grp) |
1443 | { | |
1444 | return atomic_inc_not_zero(&grp->refcount); | |
1445 | } | |
1446 | ||
1447 | static inline void put_numa_group(struct numa_group *grp) | |
1448 | { | |
1449 | if (atomic_dec_and_test(&grp->refcount)) | |
1450 | kfree_rcu(grp, rcu); | |
1451 | } | |
1452 | ||
3e6a9418 MG |
1453 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1454 | int *priv) | |
8c8a743c PZ |
1455 | { |
1456 | struct numa_group *grp, *my_grp; | |
1457 | struct task_struct *tsk; | |
1458 | bool join = false; | |
1459 | int cpu = cpupid_to_cpu(cpupid); | |
1460 | int i; | |
1461 | ||
1462 | if (unlikely(!p->numa_group)) { | |
1463 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 1464 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
1465 | |
1466 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1467 | if (!grp) | |
1468 | return; | |
1469 | ||
1470 | atomic_set(&grp->refcount, 1); | |
1471 | spin_lock_init(&grp->lock); | |
1472 | INIT_LIST_HEAD(&grp->task_list); | |
e29cf08b | 1473 | grp->gid = p->pid; |
50ec8a40 RR |
1474 | /* Second half of the array tracks nids where faults happen */ |
1475 | grp->faults_cpu = grp->faults + 2 * nr_node_ids; | |
8c8a743c | 1476 | |
50ec8a40 | 1477 | for (i = 0; i < 4*nr_node_ids; i++) |
ff1df896 | 1478 | grp->faults[i] = p->numa_faults_memory[i]; |
8c8a743c | 1479 | |
989348b5 | 1480 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 1481 | |
8c8a743c PZ |
1482 | list_add(&p->numa_entry, &grp->task_list); |
1483 | grp->nr_tasks++; | |
1484 | rcu_assign_pointer(p->numa_group, grp); | |
1485 | } | |
1486 | ||
1487 | rcu_read_lock(); | |
1488 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1489 | ||
1490 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 1491 | goto no_join; |
8c8a743c PZ |
1492 | |
1493 | grp = rcu_dereference(tsk->numa_group); | |
1494 | if (!grp) | |
3354781a | 1495 | goto no_join; |
8c8a743c PZ |
1496 | |
1497 | my_grp = p->numa_group; | |
1498 | if (grp == my_grp) | |
3354781a | 1499 | goto no_join; |
8c8a743c PZ |
1500 | |
1501 | /* | |
1502 | * Only join the other group if its bigger; if we're the bigger group, | |
1503 | * the other task will join us. | |
1504 | */ | |
1505 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 1506 | goto no_join; |
8c8a743c PZ |
1507 | |
1508 | /* | |
1509 | * Tie-break on the grp address. | |
1510 | */ | |
1511 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 1512 | goto no_join; |
8c8a743c | 1513 | |
dabe1d99 RR |
1514 | /* Always join threads in the same process. */ |
1515 | if (tsk->mm == current->mm) | |
1516 | join = true; | |
1517 | ||
1518 | /* Simple filter to avoid false positives due to PID collisions */ | |
1519 | if (flags & TNF_SHARED) | |
1520 | join = true; | |
8c8a743c | 1521 | |
3e6a9418 MG |
1522 | /* Update priv based on whether false sharing was detected */ |
1523 | *priv = !join; | |
1524 | ||
dabe1d99 | 1525 | if (join && !get_numa_group(grp)) |
3354781a | 1526 | goto no_join; |
8c8a743c | 1527 | |
8c8a743c PZ |
1528 | rcu_read_unlock(); |
1529 | ||
1530 | if (!join) | |
1531 | return; | |
1532 | ||
989348b5 MG |
1533 | double_lock(&my_grp->lock, &grp->lock); |
1534 | ||
50ec8a40 | 1535 | for (i = 0; i < 4*nr_node_ids; i++) { |
ff1df896 RR |
1536 | my_grp->faults[i] -= p->numa_faults_memory[i]; |
1537 | grp->faults[i] += p->numa_faults_memory[i]; | |
8c8a743c | 1538 | } |
989348b5 MG |
1539 | my_grp->total_faults -= p->total_numa_faults; |
1540 | grp->total_faults += p->total_numa_faults; | |
8c8a743c PZ |
1541 | |
1542 | list_move(&p->numa_entry, &grp->task_list); | |
1543 | my_grp->nr_tasks--; | |
1544 | grp->nr_tasks++; | |
1545 | ||
1546 | spin_unlock(&my_grp->lock); | |
1547 | spin_unlock(&grp->lock); | |
1548 | ||
1549 | rcu_assign_pointer(p->numa_group, grp); | |
1550 | ||
1551 | put_numa_group(my_grp); | |
3354781a PZ |
1552 | return; |
1553 | ||
1554 | no_join: | |
1555 | rcu_read_unlock(); | |
1556 | return; | |
8c8a743c PZ |
1557 | } |
1558 | ||
1559 | void task_numa_free(struct task_struct *p) | |
1560 | { | |
1561 | struct numa_group *grp = p->numa_group; | |
1562 | int i; | |
ff1df896 | 1563 | void *numa_faults = p->numa_faults_memory; |
8c8a743c PZ |
1564 | |
1565 | if (grp) { | |
989348b5 | 1566 | spin_lock(&grp->lock); |
50ec8a40 | 1567 | for (i = 0; i < 4*nr_node_ids; i++) |
ff1df896 | 1568 | grp->faults[i] -= p->numa_faults_memory[i]; |
989348b5 | 1569 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 1570 | |
8c8a743c PZ |
1571 | list_del(&p->numa_entry); |
1572 | grp->nr_tasks--; | |
1573 | spin_unlock(&grp->lock); | |
1574 | rcu_assign_pointer(p->numa_group, NULL); | |
1575 | put_numa_group(grp); | |
1576 | } | |
1577 | ||
ff1df896 RR |
1578 | p->numa_faults_memory = NULL; |
1579 | p->numa_faults_buffer_memory = NULL; | |
50ec8a40 RR |
1580 | p->numa_faults_cpu= NULL; |
1581 | p->numa_faults_buffer_cpu = NULL; | |
82727018 | 1582 | kfree(numa_faults); |
8c8a743c PZ |
1583 | } |
1584 | ||
cbee9f88 PZ |
1585 | /* |
1586 | * Got a PROT_NONE fault for a page on @node. | |
1587 | */ | |
6688cc05 | 1588 | void task_numa_fault(int last_cpupid, int node, int pages, int flags) |
cbee9f88 PZ |
1589 | { |
1590 | struct task_struct *p = current; | |
6688cc05 | 1591 | bool migrated = flags & TNF_MIGRATED; |
50ec8a40 | 1592 | int this_node = task_node(current); |
ac8e895b | 1593 | int priv; |
cbee9f88 | 1594 | |
10e84b97 | 1595 | if (!numabalancing_enabled) |
1a687c2e MG |
1596 | return; |
1597 | ||
9ff1d9ff MG |
1598 | /* for example, ksmd faulting in a user's mm */ |
1599 | if (!p->mm) | |
1600 | return; | |
1601 | ||
82727018 RR |
1602 | /* Do not worry about placement if exiting */ |
1603 | if (p->state == TASK_DEAD) | |
1604 | return; | |
1605 | ||
f809ca9a | 1606 | /* Allocate buffer to track faults on a per-node basis */ |
ff1df896 | 1607 | if (unlikely(!p->numa_faults_memory)) { |
50ec8a40 | 1608 | int size = sizeof(*p->numa_faults_memory) * 4 * nr_node_ids; |
f809ca9a | 1609 | |
745d6147 | 1610 | /* numa_faults and numa_faults_buffer share the allocation */ |
ff1df896 RR |
1611 | p->numa_faults_memory = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN); |
1612 | if (!p->numa_faults_memory) | |
f809ca9a | 1613 | return; |
745d6147 | 1614 | |
ff1df896 | 1615 | BUG_ON(p->numa_faults_buffer_memory); |
50ec8a40 RR |
1616 | p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids); |
1617 | p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids); | |
1618 | p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids); | |
83e1d2cd | 1619 | p->total_numa_faults = 0; |
04bb2f94 | 1620 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 1621 | } |
cbee9f88 | 1622 | |
8c8a743c PZ |
1623 | /* |
1624 | * First accesses are treated as private, otherwise consider accesses | |
1625 | * to be private if the accessing pid has not changed | |
1626 | */ | |
1627 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
1628 | priv = 1; | |
1629 | } else { | |
1630 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 1631 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 1632 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
1633 | } |
1634 | ||
cbee9f88 | 1635 | task_numa_placement(p); |
f809ca9a | 1636 | |
2739d3ee RR |
1637 | /* |
1638 | * Retry task to preferred node migration periodically, in case it | |
1639 | * case it previously failed, or the scheduler moved us. | |
1640 | */ | |
1641 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
1642 | numa_migrate_preferred(p); |
1643 | ||
b32e86b4 IM |
1644 | if (migrated) |
1645 | p->numa_pages_migrated += pages; | |
1646 | ||
ff1df896 | 1647 | p->numa_faults_buffer_memory[task_faults_idx(node, priv)] += pages; |
50ec8a40 | 1648 | p->numa_faults_buffer_cpu[task_faults_idx(this_node, priv)] += pages; |
04bb2f94 | 1649 | p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages; |
cbee9f88 PZ |
1650 | } |
1651 | ||
6e5fb223 PZ |
1652 | static void reset_ptenuma_scan(struct task_struct *p) |
1653 | { | |
1654 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1655 | p->mm->numa_scan_offset = 0; | |
1656 | } | |
1657 | ||
cbee9f88 PZ |
1658 | /* |
1659 | * The expensive part of numa migration is done from task_work context. | |
1660 | * Triggered from task_tick_numa(). | |
1661 | */ | |
1662 | void task_numa_work(struct callback_head *work) | |
1663 | { | |
1664 | unsigned long migrate, next_scan, now = jiffies; | |
1665 | struct task_struct *p = current; | |
1666 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1667 | struct vm_area_struct *vma; |
9f40604c | 1668 | unsigned long start, end; |
598f0ec0 | 1669 | unsigned long nr_pte_updates = 0; |
9f40604c | 1670 | long pages; |
cbee9f88 PZ |
1671 | |
1672 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1673 | ||
1674 | work->next = work; /* protect against double add */ | |
1675 | /* | |
1676 | * Who cares about NUMA placement when they're dying. | |
1677 | * | |
1678 | * NOTE: make sure not to dereference p->mm before this check, | |
1679 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1680 | * without p->mm even though we still had it when we enqueued this | |
1681 | * work. | |
1682 | */ | |
1683 | if (p->flags & PF_EXITING) | |
1684 | return; | |
1685 | ||
930aa174 | 1686 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
1687 | mm->numa_next_scan = now + |
1688 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
1689 | } |
1690 | ||
cbee9f88 PZ |
1691 | /* |
1692 | * Enforce maximal scan/migration frequency.. | |
1693 | */ | |
1694 | migrate = mm->numa_next_scan; | |
1695 | if (time_before(now, migrate)) | |
1696 | return; | |
1697 | ||
598f0ec0 MG |
1698 | if (p->numa_scan_period == 0) { |
1699 | p->numa_scan_period_max = task_scan_max(p); | |
1700 | p->numa_scan_period = task_scan_min(p); | |
1701 | } | |
cbee9f88 | 1702 | |
fb003b80 | 1703 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1704 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1705 | return; | |
1706 | ||
19a78d11 PZ |
1707 | /* |
1708 | * Delay this task enough that another task of this mm will likely win | |
1709 | * the next time around. | |
1710 | */ | |
1711 | p->node_stamp += 2 * TICK_NSEC; | |
1712 | ||
9f40604c MG |
1713 | start = mm->numa_scan_offset; |
1714 | pages = sysctl_numa_balancing_scan_size; | |
1715 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1716 | if (!pages) | |
1717 | return; | |
cbee9f88 | 1718 | |
6e5fb223 | 1719 | down_read(&mm->mmap_sem); |
9f40604c | 1720 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1721 | if (!vma) { |
1722 | reset_ptenuma_scan(p); | |
9f40604c | 1723 | start = 0; |
6e5fb223 PZ |
1724 | vma = mm->mmap; |
1725 | } | |
9f40604c | 1726 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1727 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1728 | continue; |
1729 | ||
4591ce4f MG |
1730 | /* |
1731 | * Shared library pages mapped by multiple processes are not | |
1732 | * migrated as it is expected they are cache replicated. Avoid | |
1733 | * hinting faults in read-only file-backed mappings or the vdso | |
1734 | * as migrating the pages will be of marginal benefit. | |
1735 | */ | |
1736 | if (!vma->vm_mm || | |
1737 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1738 | continue; | |
1739 | ||
3c67f474 MG |
1740 | /* |
1741 | * Skip inaccessible VMAs to avoid any confusion between | |
1742 | * PROT_NONE and NUMA hinting ptes | |
1743 | */ | |
1744 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
1745 | continue; | |
4591ce4f | 1746 | |
9f40604c MG |
1747 | do { |
1748 | start = max(start, vma->vm_start); | |
1749 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1750 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1751 | nr_pte_updates += change_prot_numa(vma, start, end); |
1752 | ||
1753 | /* | |
1754 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1755 | * at least one PTE is updated so that unused virtual | |
1756 | * address space is quickly skipped. | |
1757 | */ | |
1758 | if (nr_pte_updates) | |
1759 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1760 | |
9f40604c MG |
1761 | start = end; |
1762 | if (pages <= 0) | |
1763 | goto out; | |
1764 | } while (end != vma->vm_end); | |
cbee9f88 | 1765 | } |
6e5fb223 | 1766 | |
9f40604c | 1767 | out: |
6e5fb223 | 1768 | /* |
c69307d5 PZ |
1769 | * It is possible to reach the end of the VMA list but the last few |
1770 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1771 | * would find the !migratable VMA on the next scan but not reset the | |
1772 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1773 | */ |
1774 | if (vma) | |
9f40604c | 1775 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1776 | else |
1777 | reset_ptenuma_scan(p); | |
1778 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1779 | } |
1780 | ||
1781 | /* | |
1782 | * Drive the periodic memory faults.. | |
1783 | */ | |
1784 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1785 | { | |
1786 | struct callback_head *work = &curr->numa_work; | |
1787 | u64 period, now; | |
1788 | ||
1789 | /* | |
1790 | * We don't care about NUMA placement if we don't have memory. | |
1791 | */ | |
1792 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1793 | return; | |
1794 | ||
1795 | /* | |
1796 | * Using runtime rather than walltime has the dual advantage that | |
1797 | * we (mostly) drive the selection from busy threads and that the | |
1798 | * task needs to have done some actual work before we bother with | |
1799 | * NUMA placement. | |
1800 | */ | |
1801 | now = curr->se.sum_exec_runtime; | |
1802 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1803 | ||
1804 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1805 | if (!curr->node_stamp) |
598f0ec0 | 1806 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1807 | curr->node_stamp += period; |
cbee9f88 PZ |
1808 | |
1809 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1810 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1811 | task_work_add(curr, work, true); | |
1812 | } | |
1813 | } | |
1814 | } | |
1815 | #else | |
1816 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1817 | { | |
1818 | } | |
0ec8aa00 PZ |
1819 | |
1820 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
1821 | { | |
1822 | } | |
1823 | ||
1824 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1825 | { | |
1826 | } | |
cbee9f88 PZ |
1827 | #endif /* CONFIG_NUMA_BALANCING */ |
1828 | ||
30cfdcfc DA |
1829 | static void |
1830 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1831 | { | |
1832 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1833 | if (!parent_entity(se)) |
029632fb | 1834 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1835 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
1836 | if (entity_is_task(se)) { |
1837 | struct rq *rq = rq_of(cfs_rq); | |
1838 | ||
1839 | account_numa_enqueue(rq, task_of(se)); | |
1840 | list_add(&se->group_node, &rq->cfs_tasks); | |
1841 | } | |
367456c7 | 1842 | #endif |
30cfdcfc | 1843 | cfs_rq->nr_running++; |
30cfdcfc DA |
1844 | } |
1845 | ||
1846 | static void | |
1847 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1848 | { | |
1849 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1850 | if (!parent_entity(se)) |
029632fb | 1851 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
0ec8aa00 PZ |
1852 | if (entity_is_task(se)) { |
1853 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 1854 | list_del_init(&se->group_node); |
0ec8aa00 | 1855 | } |
30cfdcfc | 1856 | cfs_rq->nr_running--; |
30cfdcfc DA |
1857 | } |
1858 | ||
3ff6dcac YZ |
1859 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1860 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1861 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1862 | { | |
1863 | long tg_weight; | |
1864 | ||
1865 | /* | |
1866 | * Use this CPU's actual weight instead of the last load_contribution | |
1867 | * to gain a more accurate current total weight. See | |
1868 | * update_cfs_rq_load_contribution(). | |
1869 | */ | |
bf5b986e | 1870 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 1871 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
1872 | tg_weight += cfs_rq->load.weight; |
1873 | ||
1874 | return tg_weight; | |
1875 | } | |
1876 | ||
6d5ab293 | 1877 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1878 | { |
cf5f0acf | 1879 | long tg_weight, load, shares; |
3ff6dcac | 1880 | |
cf5f0acf | 1881 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1882 | load = cfs_rq->load.weight; |
3ff6dcac | 1883 | |
3ff6dcac | 1884 | shares = (tg->shares * load); |
cf5f0acf PZ |
1885 | if (tg_weight) |
1886 | shares /= tg_weight; | |
3ff6dcac YZ |
1887 | |
1888 | if (shares < MIN_SHARES) | |
1889 | shares = MIN_SHARES; | |
1890 | if (shares > tg->shares) | |
1891 | shares = tg->shares; | |
1892 | ||
1893 | return shares; | |
1894 | } | |
3ff6dcac | 1895 | # else /* CONFIG_SMP */ |
6d5ab293 | 1896 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1897 | { |
1898 | return tg->shares; | |
1899 | } | |
3ff6dcac | 1900 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1901 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1902 | unsigned long weight) | |
1903 | { | |
19e5eebb PT |
1904 | if (se->on_rq) { |
1905 | /* commit outstanding execution time */ | |
1906 | if (cfs_rq->curr == se) | |
1907 | update_curr(cfs_rq); | |
2069dd75 | 1908 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1909 | } |
2069dd75 PZ |
1910 | |
1911 | update_load_set(&se->load, weight); | |
1912 | ||
1913 | if (se->on_rq) | |
1914 | account_entity_enqueue(cfs_rq, se); | |
1915 | } | |
1916 | ||
82958366 PT |
1917 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1918 | ||
6d5ab293 | 1919 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1920 | { |
1921 | struct task_group *tg; | |
1922 | struct sched_entity *se; | |
3ff6dcac | 1923 | long shares; |
2069dd75 | 1924 | |
2069dd75 PZ |
1925 | tg = cfs_rq->tg; |
1926 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1927 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1928 | return; |
3ff6dcac YZ |
1929 | #ifndef CONFIG_SMP |
1930 | if (likely(se->load.weight == tg->shares)) | |
1931 | return; | |
1932 | #endif | |
6d5ab293 | 1933 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1934 | |
1935 | reweight_entity(cfs_rq_of(se), se, shares); | |
1936 | } | |
1937 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1938 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1939 | { |
1940 | } | |
1941 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1942 | ||
141965c7 | 1943 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
1944 | /* |
1945 | * We choose a half-life close to 1 scheduling period. | |
1946 | * Note: The tables below are dependent on this value. | |
1947 | */ | |
1948 | #define LOAD_AVG_PERIOD 32 | |
1949 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1950 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1951 | ||
1952 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1953 | static const u32 runnable_avg_yN_inv[] = { | |
1954 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1955 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1956 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1957 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1958 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1959 | 0x85aac367, 0x82cd8698, | |
1960 | }; | |
1961 | ||
1962 | /* | |
1963 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1964 | * over-estimates when re-combining. | |
1965 | */ | |
1966 | static const u32 runnable_avg_yN_sum[] = { | |
1967 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1968 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1969 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1970 | }; | |
1971 | ||
9d85f21c PT |
1972 | /* |
1973 | * Approximate: | |
1974 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1975 | */ | |
1976 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1977 | { | |
5b51f2f8 PT |
1978 | unsigned int local_n; |
1979 | ||
1980 | if (!n) | |
1981 | return val; | |
1982 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1983 | return 0; | |
1984 | ||
1985 | /* after bounds checking we can collapse to 32-bit */ | |
1986 | local_n = n; | |
1987 | ||
1988 | /* | |
1989 | * As y^PERIOD = 1/2, we can combine | |
1990 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
1991 | * With a look-up table which covers k^n (n<PERIOD) | |
1992 | * | |
1993 | * To achieve constant time decay_load. | |
1994 | */ | |
1995 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
1996 | val >>= local_n / LOAD_AVG_PERIOD; | |
1997 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
1998 | } |
1999 | ||
5b51f2f8 PT |
2000 | val *= runnable_avg_yN_inv[local_n]; |
2001 | /* We don't use SRR here since we always want to round down. */ | |
2002 | return val >> 32; | |
2003 | } | |
2004 | ||
2005 | /* | |
2006 | * For updates fully spanning n periods, the contribution to runnable | |
2007 | * average will be: \Sum 1024*y^n | |
2008 | * | |
2009 | * We can compute this reasonably efficiently by combining: | |
2010 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2011 | */ | |
2012 | static u32 __compute_runnable_contrib(u64 n) | |
2013 | { | |
2014 | u32 contrib = 0; | |
2015 | ||
2016 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2017 | return runnable_avg_yN_sum[n]; | |
2018 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2019 | return LOAD_AVG_MAX; | |
2020 | ||
2021 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
2022 | do { | |
2023 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
2024 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
2025 | ||
2026 | n -= LOAD_AVG_PERIOD; | |
2027 | } while (n > LOAD_AVG_PERIOD); | |
2028 | ||
2029 | contrib = decay_load(contrib, n); | |
2030 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2031 | } |
2032 | ||
2033 | /* | |
2034 | * We can represent the historical contribution to runnable average as the | |
2035 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2036 | * history into segments of approximately 1ms (1024us); label the segment that | |
2037 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2038 | * | |
2039 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2040 | * p0 p1 p2 | |
2041 | * (now) (~1ms ago) (~2ms ago) | |
2042 | * | |
2043 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2044 | * | |
2045 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2046 | * following representation of historical load: | |
2047 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2048 | * | |
2049 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2050 | * y^32 = 0.5 | |
2051 | * | |
2052 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2053 | * approximately half as much as the contribution to load within the last ms | |
2054 | * (u_0). | |
2055 | * | |
2056 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2057 | * sum again by y is sufficient to update: | |
2058 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2059 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2060 | */ | |
2061 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
2062 | struct sched_avg *sa, | |
2063 | int runnable) | |
2064 | { | |
5b51f2f8 PT |
2065 | u64 delta, periods; |
2066 | u32 runnable_contrib; | |
9d85f21c PT |
2067 | int delta_w, decayed = 0; |
2068 | ||
2069 | delta = now - sa->last_runnable_update; | |
2070 | /* | |
2071 | * This should only happen when time goes backwards, which it | |
2072 | * unfortunately does during sched clock init when we swap over to TSC. | |
2073 | */ | |
2074 | if ((s64)delta < 0) { | |
2075 | sa->last_runnable_update = now; | |
2076 | return 0; | |
2077 | } | |
2078 | ||
2079 | /* | |
2080 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2081 | * approximation of 1us and fast to compute. | |
2082 | */ | |
2083 | delta >>= 10; | |
2084 | if (!delta) | |
2085 | return 0; | |
2086 | sa->last_runnable_update = now; | |
2087 | ||
2088 | /* delta_w is the amount already accumulated against our next period */ | |
2089 | delta_w = sa->runnable_avg_period % 1024; | |
2090 | if (delta + delta_w >= 1024) { | |
2091 | /* period roll-over */ | |
2092 | decayed = 1; | |
2093 | ||
2094 | /* | |
2095 | * Now that we know we're crossing a period boundary, figure | |
2096 | * out how much from delta we need to complete the current | |
2097 | * period and accrue it. | |
2098 | */ | |
2099 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
2100 | if (runnable) |
2101 | sa->runnable_avg_sum += delta_w; | |
2102 | sa->runnable_avg_period += delta_w; | |
2103 | ||
2104 | delta -= delta_w; | |
2105 | ||
2106 | /* Figure out how many additional periods this update spans */ | |
2107 | periods = delta / 1024; | |
2108 | delta %= 1024; | |
2109 | ||
2110 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
2111 | periods + 1); | |
2112 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
2113 | periods + 1); | |
2114 | ||
2115 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
2116 | runnable_contrib = __compute_runnable_contrib(periods); | |
2117 | if (runnable) | |
2118 | sa->runnable_avg_sum += runnable_contrib; | |
2119 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
2120 | } |
2121 | ||
2122 | /* Remainder of delta accrued against u_0` */ | |
2123 | if (runnable) | |
2124 | sa->runnable_avg_sum += delta; | |
2125 | sa->runnable_avg_period += delta; | |
2126 | ||
2127 | return decayed; | |
2128 | } | |
2129 | ||
9ee474f5 | 2130 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 2131 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
2132 | { |
2133 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2134 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
2135 | ||
2136 | decays -= se->avg.decay_count; | |
2137 | if (!decays) | |
aff3e498 | 2138 | return 0; |
9ee474f5 PT |
2139 | |
2140 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
2141 | se->avg.decay_count = 0; | |
aff3e498 PT |
2142 | |
2143 | return decays; | |
9ee474f5 PT |
2144 | } |
2145 | ||
c566e8e9 PT |
2146 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2147 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2148 | int force_update) | |
2149 | { | |
2150 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 2151 | long tg_contrib; |
c566e8e9 PT |
2152 | |
2153 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
2154 | tg_contrib -= cfs_rq->tg_load_contrib; | |
2155 | ||
bf5b986e AS |
2156 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
2157 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
2158 | cfs_rq->tg_load_contrib += tg_contrib; |
2159 | } | |
2160 | } | |
8165e145 | 2161 | |
bb17f655 PT |
2162 | /* |
2163 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
2164 | * representation for computing load contributions. | |
2165 | */ | |
2166 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
2167 | struct cfs_rq *cfs_rq) | |
2168 | { | |
2169 | struct task_group *tg = cfs_rq->tg; | |
2170 | long contrib; | |
2171 | ||
2172 | /* The fraction of a cpu used by this cfs_rq */ | |
85b088e9 | 2173 | contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, |
bb17f655 PT |
2174 | sa->runnable_avg_period + 1); |
2175 | contrib -= cfs_rq->tg_runnable_contrib; | |
2176 | ||
2177 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
2178 | atomic_add(contrib, &tg->runnable_avg); | |
2179 | cfs_rq->tg_runnable_contrib += contrib; | |
2180 | } | |
2181 | } | |
2182 | ||
8165e145 PT |
2183 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
2184 | { | |
2185 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
2186 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
2187 | int runnable_avg; |
2188 | ||
8165e145 PT |
2189 | u64 contrib; |
2190 | ||
2191 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
2192 | se->avg.load_avg_contrib = div_u64(contrib, |
2193 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
2194 | |
2195 | /* | |
2196 | * For group entities we need to compute a correction term in the case | |
2197 | * that they are consuming <1 cpu so that we would contribute the same | |
2198 | * load as a task of equal weight. | |
2199 | * | |
2200 | * Explicitly co-ordinating this measurement would be expensive, but | |
2201 | * fortunately the sum of each cpus contribution forms a usable | |
2202 | * lower-bound on the true value. | |
2203 | * | |
2204 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
2205 | * (and the sum represents true value) or they are disjoint and we are | |
2206 | * understating by the aggregate of their overlap. | |
2207 | * | |
2208 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
2209 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
2210 | * cpus that overlap for this interval and w_i is the interval width. | |
2211 | * | |
2212 | * On a small machine; the first term is well-bounded which bounds the | |
2213 | * total error since w_i is a subset of the period. Whereas on a | |
2214 | * larger machine, while this first term can be larger, if w_i is the | |
2215 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
2216 | * our upper bound of 1-cpu. | |
2217 | */ | |
2218 | runnable_avg = atomic_read(&tg->runnable_avg); | |
2219 | if (runnable_avg < NICE_0_LOAD) { | |
2220 | se->avg.load_avg_contrib *= runnable_avg; | |
2221 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
2222 | } | |
8165e145 | 2223 | } |
c566e8e9 PT |
2224 | #else |
2225 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2226 | int force_update) {} | |
bb17f655 PT |
2227 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2228 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 2229 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
2230 | #endif |
2231 | ||
8165e145 PT |
2232 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2233 | { | |
2234 | u32 contrib; | |
2235 | ||
2236 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2237 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2238 | contrib /= (se->avg.runnable_avg_period + 1); | |
2239 | se->avg.load_avg_contrib = scale_load(contrib); | |
2240 | } | |
2241 | ||
2dac754e PT |
2242 | /* Compute the current contribution to load_avg by se, return any delta */ |
2243 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2244 | { | |
2245 | long old_contrib = se->avg.load_avg_contrib; | |
2246 | ||
8165e145 PT |
2247 | if (entity_is_task(se)) { |
2248 | __update_task_entity_contrib(se); | |
2249 | } else { | |
bb17f655 | 2250 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2251 | __update_group_entity_contrib(se); |
2252 | } | |
2dac754e PT |
2253 | |
2254 | return se->avg.load_avg_contrib - old_contrib; | |
2255 | } | |
2256 | ||
9ee474f5 PT |
2257 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2258 | long load_contrib) | |
2259 | { | |
2260 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2261 | cfs_rq->blocked_load_avg -= load_contrib; | |
2262 | else | |
2263 | cfs_rq->blocked_load_avg = 0; | |
2264 | } | |
2265 | ||
f1b17280 PT |
2266 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2267 | ||
9d85f21c | 2268 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2269 | static inline void update_entity_load_avg(struct sched_entity *se, |
2270 | int update_cfs_rq) | |
9d85f21c | 2271 | { |
2dac754e PT |
2272 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2273 | long contrib_delta; | |
f1b17280 | 2274 | u64 now; |
2dac754e | 2275 | |
f1b17280 PT |
2276 | /* |
2277 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2278 | * case they are the parent of a throttled hierarchy. | |
2279 | */ | |
2280 | if (entity_is_task(se)) | |
2281 | now = cfs_rq_clock_task(cfs_rq); | |
2282 | else | |
2283 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2284 | ||
2285 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2286 | return; |
2287 | ||
2288 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2289 | |
2290 | if (!update_cfs_rq) | |
2291 | return; | |
2292 | ||
2dac754e PT |
2293 | if (se->on_rq) |
2294 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2295 | else |
2296 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2297 | } | |
2298 | ||
2299 | /* | |
2300 | * Decay the load contributed by all blocked children and account this so that | |
2301 | * their contribution may appropriately discounted when they wake up. | |
2302 | */ | |
aff3e498 | 2303 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2304 | { |
f1b17280 | 2305 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2306 | u64 decays; |
2307 | ||
2308 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2309 | if (!decays && !force_update) |
9ee474f5 PT |
2310 | return; |
2311 | ||
2509940f AS |
2312 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2313 | unsigned long removed_load; | |
2314 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2315 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2316 | } | |
9ee474f5 | 2317 | |
aff3e498 PT |
2318 | if (decays) { |
2319 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2320 | decays); | |
2321 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2322 | cfs_rq->last_decay = now; | |
2323 | } | |
c566e8e9 PT |
2324 | |
2325 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2326 | } |
18bf2805 BS |
2327 | |
2328 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2329 | { | |
78becc27 | 2330 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 2331 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 2332 | } |
2dac754e PT |
2333 | |
2334 | /* Add the load generated by se into cfs_rq's child load-average */ | |
2335 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2336 | struct sched_entity *se, |
2337 | int wakeup) | |
2dac754e | 2338 | { |
aff3e498 PT |
2339 | /* |
2340 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2341 | * migration we use a negative decay count to track the remote decays | |
2342 | * accumulated while sleeping. | |
a75cdaa9 AS |
2343 | * |
2344 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2345 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2346 | * constructed load_avg_contrib. | |
aff3e498 PT |
2347 | */ |
2348 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2349 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2350 | if (se->avg.decay_count) { |
2351 | /* | |
2352 | * In a wake-up migration we have to approximate the | |
2353 | * time sleeping. This is because we can't synchronize | |
2354 | * clock_task between the two cpus, and it is not | |
2355 | * guaranteed to be read-safe. Instead, we can | |
2356 | * approximate this using our carried decays, which are | |
2357 | * explicitly atomically readable. | |
2358 | */ | |
2359 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2360 | << 20; | |
2361 | update_entity_load_avg(se, 0); | |
2362 | /* Indicate that we're now synchronized and on-rq */ | |
2363 | se->avg.decay_count = 0; | |
2364 | } | |
9ee474f5 PT |
2365 | wakeup = 0; |
2366 | } else { | |
9390675a | 2367 | __synchronize_entity_decay(se); |
9ee474f5 PT |
2368 | } |
2369 | ||
aff3e498 PT |
2370 | /* migrated tasks did not contribute to our blocked load */ |
2371 | if (wakeup) { | |
9ee474f5 | 2372 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2373 | update_entity_load_avg(se, 0); |
2374 | } | |
9ee474f5 | 2375 | |
2dac754e | 2376 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2377 | /* we force update consideration on load-balancer moves */ |
2378 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2379 | } |
2380 | ||
9ee474f5 PT |
2381 | /* |
2382 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2383 | * transitioning to a blocked state we track its projected decay using | |
2384 | * blocked_load_avg. | |
2385 | */ | |
2dac754e | 2386 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2387 | struct sched_entity *se, |
2388 | int sleep) | |
2dac754e | 2389 | { |
9ee474f5 | 2390 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2391 | /* we force update consideration on load-balancer moves */ |
2392 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2393 | |
2dac754e | 2394 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2395 | if (sleep) { |
2396 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2397 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2398 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2399 | } |
642dbc39 VG |
2400 | |
2401 | /* | |
2402 | * Update the rq's load with the elapsed running time before entering | |
2403 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2404 | * be the only way to update the runnable statistic. | |
2405 | */ | |
2406 | void idle_enter_fair(struct rq *this_rq) | |
2407 | { | |
2408 | update_rq_runnable_avg(this_rq, 1); | |
2409 | } | |
2410 | ||
2411 | /* | |
2412 | * Update the rq's load with the elapsed idle time before a task is | |
2413 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2414 | * be the only way to update the runnable statistic. | |
2415 | */ | |
2416 | void idle_exit_fair(struct rq *this_rq) | |
2417 | { | |
2418 | update_rq_runnable_avg(this_rq, 0); | |
2419 | } | |
2420 | ||
9d85f21c | 2421 | #else |
9ee474f5 PT |
2422 | static inline void update_entity_load_avg(struct sched_entity *se, |
2423 | int update_cfs_rq) {} | |
18bf2805 | 2424 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2425 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2426 | struct sched_entity *se, |
2427 | int wakeup) {} | |
2dac754e | 2428 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2429 | struct sched_entity *se, |
2430 | int sleep) {} | |
aff3e498 PT |
2431 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2432 | int force_update) {} | |
9d85f21c PT |
2433 | #endif |
2434 | ||
2396af69 | 2435 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2436 | { |
bf0f6f24 | 2437 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2438 | struct task_struct *tsk = NULL; |
2439 | ||
2440 | if (entity_is_task(se)) | |
2441 | tsk = task_of(se); | |
2442 | ||
41acab88 | 2443 | if (se->statistics.sleep_start) { |
78becc27 | 2444 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2445 | |
2446 | if ((s64)delta < 0) | |
2447 | delta = 0; | |
2448 | ||
41acab88 LDM |
2449 | if (unlikely(delta > se->statistics.sleep_max)) |
2450 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2451 | |
8c79a045 | 2452 | se->statistics.sleep_start = 0; |
41acab88 | 2453 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2454 | |
768d0c27 | 2455 | if (tsk) { |
e414314c | 2456 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2457 | trace_sched_stat_sleep(tsk, delta); |
2458 | } | |
bf0f6f24 | 2459 | } |
41acab88 | 2460 | if (se->statistics.block_start) { |
78becc27 | 2461 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2462 | |
2463 | if ((s64)delta < 0) | |
2464 | delta = 0; | |
2465 | ||
41acab88 LDM |
2466 | if (unlikely(delta > se->statistics.block_max)) |
2467 | se->statistics.block_max = delta; | |
bf0f6f24 | 2468 | |
8c79a045 | 2469 | se->statistics.block_start = 0; |
41acab88 | 2470 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2471 | |
e414314c | 2472 | if (tsk) { |
8f0dfc34 | 2473 | if (tsk->in_iowait) { |
41acab88 LDM |
2474 | se->statistics.iowait_sum += delta; |
2475 | se->statistics.iowait_count++; | |
768d0c27 | 2476 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2477 | } |
2478 | ||
b781a602 AV |
2479 | trace_sched_stat_blocked(tsk, delta); |
2480 | ||
e414314c PZ |
2481 | /* |
2482 | * Blocking time is in units of nanosecs, so shift by | |
2483 | * 20 to get a milliseconds-range estimation of the | |
2484 | * amount of time that the task spent sleeping: | |
2485 | */ | |
2486 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2487 | profile_hits(SLEEP_PROFILING, | |
2488 | (void *)get_wchan(tsk), | |
2489 | delta >> 20); | |
2490 | } | |
2491 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2492 | } |
bf0f6f24 IM |
2493 | } |
2494 | #endif | |
2495 | } | |
2496 | ||
ddc97297 PZ |
2497 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2498 | { | |
2499 | #ifdef CONFIG_SCHED_DEBUG | |
2500 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2501 | ||
2502 | if (d < 0) | |
2503 | d = -d; | |
2504 | ||
2505 | if (d > 3*sysctl_sched_latency) | |
2506 | schedstat_inc(cfs_rq, nr_spread_over); | |
2507 | #endif | |
2508 | } | |
2509 | ||
aeb73b04 PZ |
2510 | static void |
2511 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2512 | { | |
1af5f730 | 2513 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2514 | |
2cb8600e PZ |
2515 | /* |
2516 | * The 'current' period is already promised to the current tasks, | |
2517 | * however the extra weight of the new task will slow them down a | |
2518 | * little, place the new task so that it fits in the slot that | |
2519 | * stays open at the end. | |
2520 | */ | |
94dfb5e7 | 2521 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2522 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2523 | |
a2e7a7eb | 2524 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2525 | if (!initial) { |
a2e7a7eb | 2526 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2527 | |
a2e7a7eb MG |
2528 | /* |
2529 | * Halve their sleep time's effect, to allow | |
2530 | * for a gentler effect of sleepers: | |
2531 | */ | |
2532 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2533 | thresh >>= 1; | |
51e0304c | 2534 | |
a2e7a7eb | 2535 | vruntime -= thresh; |
aeb73b04 PZ |
2536 | } |
2537 | ||
b5d9d734 | 2538 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2539 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2540 | } |
2541 | ||
d3d9dc33 PT |
2542 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2543 | ||
bf0f6f24 | 2544 | static void |
88ec22d3 | 2545 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2546 | { |
88ec22d3 PZ |
2547 | /* |
2548 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2549 | * through calling update_curr(). |
88ec22d3 | 2550 | */ |
371fd7e7 | 2551 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2552 | se->vruntime += cfs_rq->min_vruntime; |
2553 | ||
bf0f6f24 | 2554 | /* |
a2a2d680 | 2555 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2556 | */ |
b7cc0896 | 2557 | update_curr(cfs_rq); |
f269ae04 | 2558 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2559 | account_entity_enqueue(cfs_rq, se); |
2560 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2561 | |
88ec22d3 | 2562 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2563 | place_entity(cfs_rq, se, 0); |
2396af69 | 2564 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2565 | } |
bf0f6f24 | 2566 | |
d2417e5a | 2567 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2568 | check_spread(cfs_rq, se); |
83b699ed SV |
2569 | if (se != cfs_rq->curr) |
2570 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2571 | se->on_rq = 1; |
3d4b47b4 | 2572 | |
d3d9dc33 | 2573 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2574 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2575 | check_enqueue_throttle(cfs_rq); |
2576 | } | |
bf0f6f24 IM |
2577 | } |
2578 | ||
2c13c919 | 2579 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2580 | { |
2c13c919 RR |
2581 | for_each_sched_entity(se) { |
2582 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2583 | if (cfs_rq->last == se) | |
2584 | cfs_rq->last = NULL; | |
2585 | else | |
2586 | break; | |
2587 | } | |
2588 | } | |
2002c695 | 2589 | |
2c13c919 RR |
2590 | static void __clear_buddies_next(struct sched_entity *se) |
2591 | { | |
2592 | for_each_sched_entity(se) { | |
2593 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2594 | if (cfs_rq->next == se) | |
2595 | cfs_rq->next = NULL; | |
2596 | else | |
2597 | break; | |
2598 | } | |
2002c695 PZ |
2599 | } |
2600 | ||
ac53db59 RR |
2601 | static void __clear_buddies_skip(struct sched_entity *se) |
2602 | { | |
2603 | for_each_sched_entity(se) { | |
2604 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2605 | if (cfs_rq->skip == se) | |
2606 | cfs_rq->skip = NULL; | |
2607 | else | |
2608 | break; | |
2609 | } | |
2610 | } | |
2611 | ||
a571bbea PZ |
2612 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2613 | { | |
2c13c919 RR |
2614 | if (cfs_rq->last == se) |
2615 | __clear_buddies_last(se); | |
2616 | ||
2617 | if (cfs_rq->next == se) | |
2618 | __clear_buddies_next(se); | |
ac53db59 RR |
2619 | |
2620 | if (cfs_rq->skip == se) | |
2621 | __clear_buddies_skip(se); | |
a571bbea PZ |
2622 | } |
2623 | ||
6c16a6dc | 2624 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2625 | |
bf0f6f24 | 2626 | static void |
371fd7e7 | 2627 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2628 | { |
a2a2d680 DA |
2629 | /* |
2630 | * Update run-time statistics of the 'current'. | |
2631 | */ | |
2632 | update_curr(cfs_rq); | |
17bc14b7 | 2633 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2634 | |
19b6a2e3 | 2635 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2636 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2637 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2638 | if (entity_is_task(se)) { |
2639 | struct task_struct *tsk = task_of(se); | |
2640 | ||
2641 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2642 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2643 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2644 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2645 | } |
db36cc7d | 2646 | #endif |
67e9fb2a PZ |
2647 | } |
2648 | ||
2002c695 | 2649 | clear_buddies(cfs_rq, se); |
4793241b | 2650 | |
83b699ed | 2651 | if (se != cfs_rq->curr) |
30cfdcfc | 2652 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2653 | se->on_rq = 0; |
30cfdcfc | 2654 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2655 | |
2656 | /* | |
2657 | * Normalize the entity after updating the min_vruntime because the | |
2658 | * update can refer to the ->curr item and we need to reflect this | |
2659 | * movement in our normalized position. | |
2660 | */ | |
371fd7e7 | 2661 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2662 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2663 | |
d8b4986d PT |
2664 | /* return excess runtime on last dequeue */ |
2665 | return_cfs_rq_runtime(cfs_rq); | |
2666 | ||
1e876231 | 2667 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2668 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2669 | } |
2670 | ||
2671 | /* | |
2672 | * Preempt the current task with a newly woken task if needed: | |
2673 | */ | |
7c92e54f | 2674 | static void |
2e09bf55 | 2675 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2676 | { |
11697830 | 2677 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2678 | struct sched_entity *se; |
2679 | s64 delta; | |
11697830 | 2680 | |
6d0f0ebd | 2681 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2682 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2683 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2684 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2685 | /* |
2686 | * The current task ran long enough, ensure it doesn't get | |
2687 | * re-elected due to buddy favours. | |
2688 | */ | |
2689 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2690 | return; |
2691 | } | |
2692 | ||
2693 | /* | |
2694 | * Ensure that a task that missed wakeup preemption by a | |
2695 | * narrow margin doesn't have to wait for a full slice. | |
2696 | * This also mitigates buddy induced latencies under load. | |
2697 | */ | |
f685ceac MG |
2698 | if (delta_exec < sysctl_sched_min_granularity) |
2699 | return; | |
2700 | ||
f4cfb33e WX |
2701 | se = __pick_first_entity(cfs_rq); |
2702 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2703 | |
f4cfb33e WX |
2704 | if (delta < 0) |
2705 | return; | |
d7d82944 | 2706 | |
f4cfb33e WX |
2707 | if (delta > ideal_runtime) |
2708 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2709 | } |
2710 | ||
83b699ed | 2711 | static void |
8494f412 | 2712 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2713 | { |
83b699ed SV |
2714 | /* 'current' is not kept within the tree. */ |
2715 | if (se->on_rq) { | |
2716 | /* | |
2717 | * Any task has to be enqueued before it get to execute on | |
2718 | * a CPU. So account for the time it spent waiting on the | |
2719 | * runqueue. | |
2720 | */ | |
2721 | update_stats_wait_end(cfs_rq, se); | |
2722 | __dequeue_entity(cfs_rq, se); | |
2723 | } | |
2724 | ||
79303e9e | 2725 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2726 | cfs_rq->curr = se; |
eba1ed4b IM |
2727 | #ifdef CONFIG_SCHEDSTATS |
2728 | /* | |
2729 | * Track our maximum slice length, if the CPU's load is at | |
2730 | * least twice that of our own weight (i.e. dont track it | |
2731 | * when there are only lesser-weight tasks around): | |
2732 | */ | |
495eca49 | 2733 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2734 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2735 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2736 | } | |
2737 | #endif | |
4a55b450 | 2738 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2739 | } |
2740 | ||
3f3a4904 PZ |
2741 | static int |
2742 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2743 | ||
ac53db59 RR |
2744 | /* |
2745 | * Pick the next process, keeping these things in mind, in this order: | |
2746 | * 1) keep things fair between processes/task groups | |
2747 | * 2) pick the "next" process, since someone really wants that to run | |
2748 | * 3) pick the "last" process, for cache locality | |
2749 | * 4) do not run the "skip" process, if something else is available | |
2750 | */ | |
f4b6755f | 2751 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 2752 | { |
ac53db59 | 2753 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 2754 | struct sched_entity *left = se; |
f4b6755f | 2755 | |
ac53db59 RR |
2756 | /* |
2757 | * Avoid running the skip buddy, if running something else can | |
2758 | * be done without getting too unfair. | |
2759 | */ | |
2760 | if (cfs_rq->skip == se) { | |
2761 | struct sched_entity *second = __pick_next_entity(se); | |
2762 | if (second && wakeup_preempt_entity(second, left) < 1) | |
2763 | se = second; | |
2764 | } | |
aa2ac252 | 2765 | |
f685ceac MG |
2766 | /* |
2767 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2768 | */ | |
2769 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2770 | se = cfs_rq->last; | |
2771 | ||
ac53db59 RR |
2772 | /* |
2773 | * Someone really wants this to run. If it's not unfair, run it. | |
2774 | */ | |
2775 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2776 | se = cfs_rq->next; | |
2777 | ||
f685ceac | 2778 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2779 | |
2780 | return se; | |
aa2ac252 PZ |
2781 | } |
2782 | ||
d3d9dc33 PT |
2783 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
2784 | ||
ab6cde26 | 2785 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2786 | { |
2787 | /* | |
2788 | * If still on the runqueue then deactivate_task() | |
2789 | * was not called and update_curr() has to be done: | |
2790 | */ | |
2791 | if (prev->on_rq) | |
b7cc0896 | 2792 | update_curr(cfs_rq); |
bf0f6f24 | 2793 | |
d3d9dc33 PT |
2794 | /* throttle cfs_rqs exceeding runtime */ |
2795 | check_cfs_rq_runtime(cfs_rq); | |
2796 | ||
ddc97297 | 2797 | check_spread(cfs_rq, prev); |
30cfdcfc | 2798 | if (prev->on_rq) { |
5870db5b | 2799 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2800 | /* Put 'current' back into the tree. */ |
2801 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2802 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2803 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2804 | } |
429d43bc | 2805 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2806 | } |
2807 | ||
8f4d37ec PZ |
2808 | static void |
2809 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2810 | { |
bf0f6f24 | 2811 | /* |
30cfdcfc | 2812 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2813 | */ |
30cfdcfc | 2814 | update_curr(cfs_rq); |
bf0f6f24 | 2815 | |
9d85f21c PT |
2816 | /* |
2817 | * Ensure that runnable average is periodically updated. | |
2818 | */ | |
9ee474f5 | 2819 | update_entity_load_avg(curr, 1); |
aff3e498 | 2820 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2821 | update_cfs_shares(cfs_rq); |
9d85f21c | 2822 | |
8f4d37ec PZ |
2823 | #ifdef CONFIG_SCHED_HRTICK |
2824 | /* | |
2825 | * queued ticks are scheduled to match the slice, so don't bother | |
2826 | * validating it and just reschedule. | |
2827 | */ | |
983ed7a6 HH |
2828 | if (queued) { |
2829 | resched_task(rq_of(cfs_rq)->curr); | |
2830 | return; | |
2831 | } | |
8f4d37ec PZ |
2832 | /* |
2833 | * don't let the period tick interfere with the hrtick preemption | |
2834 | */ | |
2835 | if (!sched_feat(DOUBLE_TICK) && | |
2836 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2837 | return; | |
2838 | #endif | |
2839 | ||
2c2efaed | 2840 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2841 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2842 | } |
2843 | ||
ab84d31e PT |
2844 | |
2845 | /************************************************** | |
2846 | * CFS bandwidth control machinery | |
2847 | */ | |
2848 | ||
2849 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2850 | |
2851 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2852 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2853 | |
2854 | static inline bool cfs_bandwidth_used(void) | |
2855 | { | |
c5905afb | 2856 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2857 | } |
2858 | ||
1ee14e6c | 2859 | void cfs_bandwidth_usage_inc(void) |
029632fb | 2860 | { |
1ee14e6c BS |
2861 | static_key_slow_inc(&__cfs_bandwidth_used); |
2862 | } | |
2863 | ||
2864 | void cfs_bandwidth_usage_dec(void) | |
2865 | { | |
2866 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
2867 | } |
2868 | #else /* HAVE_JUMP_LABEL */ | |
2869 | static bool cfs_bandwidth_used(void) | |
2870 | { | |
2871 | return true; | |
2872 | } | |
2873 | ||
1ee14e6c BS |
2874 | void cfs_bandwidth_usage_inc(void) {} |
2875 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
2876 | #endif /* HAVE_JUMP_LABEL */ |
2877 | ||
ab84d31e PT |
2878 | /* |
2879 | * default period for cfs group bandwidth. | |
2880 | * default: 0.1s, units: nanoseconds | |
2881 | */ | |
2882 | static inline u64 default_cfs_period(void) | |
2883 | { | |
2884 | return 100000000ULL; | |
2885 | } | |
ec12cb7f PT |
2886 | |
2887 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2888 | { | |
2889 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2890 | } | |
2891 | ||
a9cf55b2 PT |
2892 | /* |
2893 | * Replenish runtime according to assigned quota and update expiration time. | |
2894 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2895 | * additional synchronization around rq->lock. | |
2896 | * | |
2897 | * requires cfs_b->lock | |
2898 | */ | |
029632fb | 2899 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2900 | { |
2901 | u64 now; | |
2902 | ||
2903 | if (cfs_b->quota == RUNTIME_INF) | |
2904 | return; | |
2905 | ||
2906 | now = sched_clock_cpu(smp_processor_id()); | |
2907 | cfs_b->runtime = cfs_b->quota; | |
2908 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2909 | } | |
2910 | ||
029632fb PZ |
2911 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2912 | { | |
2913 | return &tg->cfs_bandwidth; | |
2914 | } | |
2915 | ||
f1b17280 PT |
2916 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2917 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2918 | { | |
2919 | if (unlikely(cfs_rq->throttle_count)) | |
2920 | return cfs_rq->throttled_clock_task; | |
2921 | ||
78becc27 | 2922 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2923 | } |
2924 | ||
85dac906 PT |
2925 | /* returns 0 on failure to allocate runtime */ |
2926 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2927 | { |
2928 | struct task_group *tg = cfs_rq->tg; | |
2929 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2930 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2931 | |
2932 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2933 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2934 | ||
2935 | raw_spin_lock(&cfs_b->lock); | |
2936 | if (cfs_b->quota == RUNTIME_INF) | |
2937 | amount = min_amount; | |
58088ad0 | 2938 | else { |
a9cf55b2 PT |
2939 | /* |
2940 | * If the bandwidth pool has become inactive, then at least one | |
2941 | * period must have elapsed since the last consumption. | |
2942 | * Refresh the global state and ensure bandwidth timer becomes | |
2943 | * active. | |
2944 | */ | |
2945 | if (!cfs_b->timer_active) { | |
2946 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2947 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2948 | } |
58088ad0 PT |
2949 | |
2950 | if (cfs_b->runtime > 0) { | |
2951 | amount = min(cfs_b->runtime, min_amount); | |
2952 | cfs_b->runtime -= amount; | |
2953 | cfs_b->idle = 0; | |
2954 | } | |
ec12cb7f | 2955 | } |
a9cf55b2 | 2956 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2957 | raw_spin_unlock(&cfs_b->lock); |
2958 | ||
2959 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2960 | /* |
2961 | * we may have advanced our local expiration to account for allowed | |
2962 | * spread between our sched_clock and the one on which runtime was | |
2963 | * issued. | |
2964 | */ | |
2965 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2966 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2967 | |
2968 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2969 | } |
2970 | ||
a9cf55b2 PT |
2971 | /* |
2972 | * Note: This depends on the synchronization provided by sched_clock and the | |
2973 | * fact that rq->clock snapshots this value. | |
2974 | */ | |
2975 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2976 | { |
a9cf55b2 | 2977 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2978 | |
2979 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2980 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2981 | return; |
2982 | ||
a9cf55b2 PT |
2983 | if (cfs_rq->runtime_remaining < 0) |
2984 | return; | |
2985 | ||
2986 | /* | |
2987 | * If the local deadline has passed we have to consider the | |
2988 | * possibility that our sched_clock is 'fast' and the global deadline | |
2989 | * has not truly expired. | |
2990 | * | |
2991 | * Fortunately we can check determine whether this the case by checking | |
2992 | * whether the global deadline has advanced. | |
2993 | */ | |
2994 | ||
2995 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
2996 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
2997 | cfs_rq->runtime_expires += TICK_NSEC; | |
2998 | } else { | |
2999 | /* global deadline is ahead, expiration has passed */ | |
3000 | cfs_rq->runtime_remaining = 0; | |
3001 | } | |
3002 | } | |
3003 | ||
9dbdb155 | 3004 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
3005 | { |
3006 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3007 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3008 | expire_cfs_rq_runtime(cfs_rq); |
3009 | ||
3010 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3011 | return; |
3012 | ||
85dac906 PT |
3013 | /* |
3014 | * if we're unable to extend our runtime we resched so that the active | |
3015 | * hierarchy can be throttled | |
3016 | */ | |
3017 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
3018 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
3019 | } |
3020 | ||
6c16a6dc | 3021 | static __always_inline |
9dbdb155 | 3022 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 3023 | { |
56f570e5 | 3024 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3025 | return; |
3026 | ||
3027 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3028 | } | |
3029 | ||
85dac906 PT |
3030 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3031 | { | |
56f570e5 | 3032 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3033 | } |
3034 | ||
64660c86 PT |
3035 | /* check whether cfs_rq, or any parent, is throttled */ |
3036 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3037 | { | |
56f570e5 | 3038 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3039 | } |
3040 | ||
3041 | /* | |
3042 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3043 | * dest_cpu are members of a throttled hierarchy when performing group | |
3044 | * load-balance operations. | |
3045 | */ | |
3046 | static inline int throttled_lb_pair(struct task_group *tg, | |
3047 | int src_cpu, int dest_cpu) | |
3048 | { | |
3049 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3050 | ||
3051 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3052 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3053 | ||
3054 | return throttled_hierarchy(src_cfs_rq) || | |
3055 | throttled_hierarchy(dest_cfs_rq); | |
3056 | } | |
3057 | ||
3058 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3059 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3060 | { | |
3061 | struct rq *rq = data; | |
3062 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3063 | ||
3064 | cfs_rq->throttle_count--; | |
3065 | #ifdef CONFIG_SMP | |
3066 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 3067 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3068 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3069 | cfs_rq->throttled_clock_task; |
64660c86 PT |
3070 | } |
3071 | #endif | |
3072 | ||
3073 | return 0; | |
3074 | } | |
3075 | ||
3076 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3077 | { | |
3078 | struct rq *rq = data; | |
3079 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3080 | ||
82958366 PT |
3081 | /* group is entering throttled state, stop time */ |
3082 | if (!cfs_rq->throttle_count) | |
78becc27 | 3083 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3084 | cfs_rq->throttle_count++; |
3085 | ||
3086 | return 0; | |
3087 | } | |
3088 | ||
d3d9dc33 | 3089 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3090 | { |
3091 | struct rq *rq = rq_of(cfs_rq); | |
3092 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3093 | struct sched_entity *se; | |
3094 | long task_delta, dequeue = 1; | |
3095 | ||
3096 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3097 | ||
f1b17280 | 3098 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3099 | rcu_read_lock(); |
3100 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3101 | rcu_read_unlock(); | |
85dac906 PT |
3102 | |
3103 | task_delta = cfs_rq->h_nr_running; | |
3104 | for_each_sched_entity(se) { | |
3105 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3106 | /* throttled entity or throttle-on-deactivate */ | |
3107 | if (!se->on_rq) | |
3108 | break; | |
3109 | ||
3110 | if (dequeue) | |
3111 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3112 | qcfs_rq->h_nr_running -= task_delta; | |
3113 | ||
3114 | if (qcfs_rq->load.weight) | |
3115 | dequeue = 0; | |
3116 | } | |
3117 | ||
3118 | if (!se) | |
3119 | rq->nr_running -= task_delta; | |
3120 | ||
3121 | cfs_rq->throttled = 1; | |
78becc27 | 3122 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
3123 | raw_spin_lock(&cfs_b->lock); |
3124 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
f9f9ffc2 BS |
3125 | if (!cfs_b->timer_active) |
3126 | __start_cfs_bandwidth(cfs_b); | |
85dac906 PT |
3127 | raw_spin_unlock(&cfs_b->lock); |
3128 | } | |
3129 | ||
029632fb | 3130 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3131 | { |
3132 | struct rq *rq = rq_of(cfs_rq); | |
3133 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3134 | struct sched_entity *se; | |
3135 | int enqueue = 1; | |
3136 | long task_delta; | |
3137 | ||
22b958d8 | 3138 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3139 | |
3140 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3141 | |
3142 | update_rq_clock(rq); | |
3143 | ||
671fd9da | 3144 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3145 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3146 | list_del_rcu(&cfs_rq->throttled_list); |
3147 | raw_spin_unlock(&cfs_b->lock); | |
3148 | ||
64660c86 PT |
3149 | /* update hierarchical throttle state */ |
3150 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3151 | ||
671fd9da PT |
3152 | if (!cfs_rq->load.weight) |
3153 | return; | |
3154 | ||
3155 | task_delta = cfs_rq->h_nr_running; | |
3156 | for_each_sched_entity(se) { | |
3157 | if (se->on_rq) | |
3158 | enqueue = 0; | |
3159 | ||
3160 | cfs_rq = cfs_rq_of(se); | |
3161 | if (enqueue) | |
3162 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3163 | cfs_rq->h_nr_running += task_delta; | |
3164 | ||
3165 | if (cfs_rq_throttled(cfs_rq)) | |
3166 | break; | |
3167 | } | |
3168 | ||
3169 | if (!se) | |
3170 | rq->nr_running += task_delta; | |
3171 | ||
3172 | /* determine whether we need to wake up potentially idle cpu */ | |
3173 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
3174 | resched_task(rq->curr); | |
3175 | } | |
3176 | ||
3177 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
3178 | u64 remaining, u64 expires) | |
3179 | { | |
3180 | struct cfs_rq *cfs_rq; | |
3181 | u64 runtime = remaining; | |
3182 | ||
3183 | rcu_read_lock(); | |
3184 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
3185 | throttled_list) { | |
3186 | struct rq *rq = rq_of(cfs_rq); | |
3187 | ||
3188 | raw_spin_lock(&rq->lock); | |
3189 | if (!cfs_rq_throttled(cfs_rq)) | |
3190 | goto next; | |
3191 | ||
3192 | runtime = -cfs_rq->runtime_remaining + 1; | |
3193 | if (runtime > remaining) | |
3194 | runtime = remaining; | |
3195 | remaining -= runtime; | |
3196 | ||
3197 | cfs_rq->runtime_remaining += runtime; | |
3198 | cfs_rq->runtime_expires = expires; | |
3199 | ||
3200 | /* we check whether we're throttled above */ | |
3201 | if (cfs_rq->runtime_remaining > 0) | |
3202 | unthrottle_cfs_rq(cfs_rq); | |
3203 | ||
3204 | next: | |
3205 | raw_spin_unlock(&rq->lock); | |
3206 | ||
3207 | if (!remaining) | |
3208 | break; | |
3209 | } | |
3210 | rcu_read_unlock(); | |
3211 | ||
3212 | return remaining; | |
3213 | } | |
3214 | ||
58088ad0 PT |
3215 | /* |
3216 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
3217 | * cfs_rqs as appropriate. If there has been no activity within the last | |
3218 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
3219 | * used to track this state. | |
3220 | */ | |
3221 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3222 | { | |
671fd9da PT |
3223 | u64 runtime, runtime_expires; |
3224 | int idle = 1, throttled; | |
58088ad0 PT |
3225 | |
3226 | raw_spin_lock(&cfs_b->lock); | |
3227 | /* no need to continue the timer with no bandwidth constraint */ | |
3228 | if (cfs_b->quota == RUNTIME_INF) | |
3229 | goto out_unlock; | |
3230 | ||
671fd9da PT |
3231 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3232 | /* idle depends on !throttled (for the case of a large deficit) */ | |
3233 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 3234 | cfs_b->nr_periods += overrun; |
671fd9da | 3235 | |
a9cf55b2 PT |
3236 | /* if we're going inactive then everything else can be deferred */ |
3237 | if (idle) | |
3238 | goto out_unlock; | |
3239 | ||
927b54fc BS |
3240 | /* |
3241 | * if we have relooped after returning idle once, we need to update our | |
3242 | * status as actually running, so that other cpus doing | |
3243 | * __start_cfs_bandwidth will stop trying to cancel us. | |
3244 | */ | |
3245 | cfs_b->timer_active = 1; | |
3246 | ||
a9cf55b2 PT |
3247 | __refill_cfs_bandwidth_runtime(cfs_b); |
3248 | ||
671fd9da PT |
3249 | if (!throttled) { |
3250 | /* mark as potentially idle for the upcoming period */ | |
3251 | cfs_b->idle = 1; | |
3252 | goto out_unlock; | |
3253 | } | |
3254 | ||
e8da1b18 NR |
3255 | /* account preceding periods in which throttling occurred */ |
3256 | cfs_b->nr_throttled += overrun; | |
3257 | ||
671fd9da PT |
3258 | /* |
3259 | * There are throttled entities so we must first use the new bandwidth | |
3260 | * to unthrottle them before making it generally available. This | |
3261 | * ensures that all existing debts will be paid before a new cfs_rq is | |
3262 | * allowed to run. | |
3263 | */ | |
3264 | runtime = cfs_b->runtime; | |
3265 | runtime_expires = cfs_b->runtime_expires; | |
3266 | cfs_b->runtime = 0; | |
3267 | ||
3268 | /* | |
3269 | * This check is repeated as we are holding onto the new bandwidth | |
3270 | * while we unthrottle. This can potentially race with an unthrottled | |
3271 | * group trying to acquire new bandwidth from the global pool. | |
3272 | */ | |
3273 | while (throttled && runtime > 0) { | |
3274 | raw_spin_unlock(&cfs_b->lock); | |
3275 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3276 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3277 | runtime_expires); | |
3278 | raw_spin_lock(&cfs_b->lock); | |
3279 | ||
3280 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
3281 | } | |
58088ad0 | 3282 | |
671fd9da PT |
3283 | /* return (any) remaining runtime */ |
3284 | cfs_b->runtime = runtime; | |
3285 | /* | |
3286 | * While we are ensured activity in the period following an | |
3287 | * unthrottle, this also covers the case in which the new bandwidth is | |
3288 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3289 | * timer to remain active while there are any throttled entities.) | |
3290 | */ | |
3291 | cfs_b->idle = 0; | |
58088ad0 PT |
3292 | out_unlock: |
3293 | if (idle) | |
3294 | cfs_b->timer_active = 0; | |
3295 | raw_spin_unlock(&cfs_b->lock); | |
3296 | ||
3297 | return idle; | |
3298 | } | |
d3d9dc33 | 3299 | |
d8b4986d PT |
3300 | /* a cfs_rq won't donate quota below this amount */ |
3301 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3302 | /* minimum remaining period time to redistribute slack quota */ | |
3303 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3304 | /* how long we wait to gather additional slack before distributing */ | |
3305 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3306 | ||
db06e78c BS |
3307 | /* |
3308 | * Are we near the end of the current quota period? | |
3309 | * | |
3310 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
3311 | * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of | |
3312 | * migrate_hrtimers, base is never cleared, so we are fine. | |
3313 | */ | |
d8b4986d PT |
3314 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
3315 | { | |
3316 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3317 | u64 remaining; | |
3318 | ||
3319 | /* if the call-back is running a quota refresh is already occurring */ | |
3320 | if (hrtimer_callback_running(refresh_timer)) | |
3321 | return 1; | |
3322 | ||
3323 | /* is a quota refresh about to occur? */ | |
3324 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3325 | if (remaining < min_expire) | |
3326 | return 1; | |
3327 | ||
3328 | return 0; | |
3329 | } | |
3330 | ||
3331 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3332 | { | |
3333 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3334 | ||
3335 | /* if there's a quota refresh soon don't bother with slack */ | |
3336 | if (runtime_refresh_within(cfs_b, min_left)) | |
3337 | return; | |
3338 | ||
3339 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3340 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3341 | } | |
3342 | ||
3343 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3344 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3345 | { | |
3346 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3347 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3348 | ||
3349 | if (slack_runtime <= 0) | |
3350 | return; | |
3351 | ||
3352 | raw_spin_lock(&cfs_b->lock); | |
3353 | if (cfs_b->quota != RUNTIME_INF && | |
3354 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3355 | cfs_b->runtime += slack_runtime; | |
3356 | ||
3357 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3358 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3359 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3360 | start_cfs_slack_bandwidth(cfs_b); | |
3361 | } | |
3362 | raw_spin_unlock(&cfs_b->lock); | |
3363 | ||
3364 | /* even if it's not valid for return we don't want to try again */ | |
3365 | cfs_rq->runtime_remaining -= slack_runtime; | |
3366 | } | |
3367 | ||
3368 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3369 | { | |
56f570e5 PT |
3370 | if (!cfs_bandwidth_used()) |
3371 | return; | |
3372 | ||
fccfdc6f | 3373 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3374 | return; |
3375 | ||
3376 | __return_cfs_rq_runtime(cfs_rq); | |
3377 | } | |
3378 | ||
3379 | /* | |
3380 | * This is done with a timer (instead of inline with bandwidth return) since | |
3381 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3382 | */ | |
3383 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3384 | { | |
3385 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3386 | u64 expires; | |
3387 | ||
3388 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
3389 | raw_spin_lock(&cfs_b->lock); |
3390 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
3391 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 3392 | return; |
db06e78c | 3393 | } |
d8b4986d | 3394 | |
d8b4986d PT |
3395 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { |
3396 | runtime = cfs_b->runtime; | |
3397 | cfs_b->runtime = 0; | |
3398 | } | |
3399 | expires = cfs_b->runtime_expires; | |
3400 | raw_spin_unlock(&cfs_b->lock); | |
3401 | ||
3402 | if (!runtime) | |
3403 | return; | |
3404 | ||
3405 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3406 | ||
3407 | raw_spin_lock(&cfs_b->lock); | |
3408 | if (expires == cfs_b->runtime_expires) | |
3409 | cfs_b->runtime = runtime; | |
3410 | raw_spin_unlock(&cfs_b->lock); | |
3411 | } | |
3412 | ||
d3d9dc33 PT |
3413 | /* |
3414 | * When a group wakes up we want to make sure that its quota is not already | |
3415 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3416 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3417 | */ | |
3418 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3419 | { | |
56f570e5 PT |
3420 | if (!cfs_bandwidth_used()) |
3421 | return; | |
3422 | ||
d3d9dc33 PT |
3423 | /* an active group must be handled by the update_curr()->put() path */ |
3424 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3425 | return; | |
3426 | ||
3427 | /* ensure the group is not already throttled */ | |
3428 | if (cfs_rq_throttled(cfs_rq)) | |
3429 | return; | |
3430 | ||
3431 | /* update runtime allocation */ | |
3432 | account_cfs_rq_runtime(cfs_rq, 0); | |
3433 | if (cfs_rq->runtime_remaining <= 0) | |
3434 | throttle_cfs_rq(cfs_rq); | |
3435 | } | |
3436 | ||
3437 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
3438 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3439 | { | |
56f570e5 PT |
3440 | if (!cfs_bandwidth_used()) |
3441 | return; | |
3442 | ||
d3d9dc33 PT |
3443 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
3444 | return; | |
3445 | ||
3446 | /* | |
3447 | * it's possible for a throttled entity to be forced into a running | |
3448 | * state (e.g. set_curr_task), in this case we're finished. | |
3449 | */ | |
3450 | if (cfs_rq_throttled(cfs_rq)) | |
3451 | return; | |
3452 | ||
3453 | throttle_cfs_rq(cfs_rq); | |
3454 | } | |
029632fb | 3455 | |
029632fb PZ |
3456 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3457 | { | |
3458 | struct cfs_bandwidth *cfs_b = | |
3459 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3460 | do_sched_cfs_slack_timer(cfs_b); | |
3461 | ||
3462 | return HRTIMER_NORESTART; | |
3463 | } | |
3464 | ||
3465 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3466 | { | |
3467 | struct cfs_bandwidth *cfs_b = | |
3468 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3469 | ktime_t now; | |
3470 | int overrun; | |
3471 | int idle = 0; | |
3472 | ||
3473 | for (;;) { | |
3474 | now = hrtimer_cb_get_time(timer); | |
3475 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3476 | ||
3477 | if (!overrun) | |
3478 | break; | |
3479 | ||
3480 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3481 | } | |
3482 | ||
3483 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3484 | } | |
3485 | ||
3486 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3487 | { | |
3488 | raw_spin_lock_init(&cfs_b->lock); | |
3489 | cfs_b->runtime = 0; | |
3490 | cfs_b->quota = RUNTIME_INF; | |
3491 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3492 | ||
3493 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3494 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3495 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3496 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3497 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3498 | } | |
3499 | ||
3500 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3501 | { | |
3502 | cfs_rq->runtime_enabled = 0; | |
3503 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3504 | } | |
3505 | ||
3506 | /* requires cfs_b->lock, may release to reprogram timer */ | |
3507 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3508 | { | |
3509 | /* | |
3510 | * The timer may be active because we're trying to set a new bandwidth | |
3511 | * period or because we're racing with the tear-down path | |
3512 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3513 | * terminates). In either case we ensure that it's re-programmed | |
3514 | */ | |
927b54fc BS |
3515 | while (unlikely(hrtimer_active(&cfs_b->period_timer)) && |
3516 | hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { | |
3517 | /* bounce the lock to allow do_sched_cfs_period_timer to run */ | |
029632fb | 3518 | raw_spin_unlock(&cfs_b->lock); |
927b54fc | 3519 | cpu_relax(); |
029632fb PZ |
3520 | raw_spin_lock(&cfs_b->lock); |
3521 | /* if someone else restarted the timer then we're done */ | |
3522 | if (cfs_b->timer_active) | |
3523 | return; | |
3524 | } | |
3525 | ||
3526 | cfs_b->timer_active = 1; | |
3527 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3528 | } | |
3529 | ||
3530 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3531 | { | |
3532 | hrtimer_cancel(&cfs_b->period_timer); | |
3533 | hrtimer_cancel(&cfs_b->slack_timer); | |
3534 | } | |
3535 | ||
38dc3348 | 3536 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3537 | { |
3538 | struct cfs_rq *cfs_rq; | |
3539 | ||
3540 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3541 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3542 | ||
3543 | if (!cfs_rq->runtime_enabled) | |
3544 | continue; | |
3545 | ||
3546 | /* | |
3547 | * clock_task is not advancing so we just need to make sure | |
3548 | * there's some valid quota amount | |
3549 | */ | |
3550 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3551 | if (cfs_rq_throttled(cfs_rq)) | |
3552 | unthrottle_cfs_rq(cfs_rq); | |
3553 | } | |
3554 | } | |
3555 | ||
3556 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3557 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3558 | { | |
78becc27 | 3559 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3560 | } |
3561 | ||
9dbdb155 | 3562 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
d3d9dc33 PT |
3563 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
3564 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 3565 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3566 | |
3567 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3568 | { | |
3569 | return 0; | |
3570 | } | |
64660c86 PT |
3571 | |
3572 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3573 | { | |
3574 | return 0; | |
3575 | } | |
3576 | ||
3577 | static inline int throttled_lb_pair(struct task_group *tg, | |
3578 | int src_cpu, int dest_cpu) | |
3579 | { | |
3580 | return 0; | |
3581 | } | |
029632fb PZ |
3582 | |
3583 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3584 | ||
3585 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3586 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3587 | #endif |
3588 | ||
029632fb PZ |
3589 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3590 | { | |
3591 | return NULL; | |
3592 | } | |
3593 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3594 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3595 | |
3596 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3597 | ||
bf0f6f24 IM |
3598 | /************************************************** |
3599 | * CFS operations on tasks: | |
3600 | */ | |
3601 | ||
8f4d37ec PZ |
3602 | #ifdef CONFIG_SCHED_HRTICK |
3603 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3604 | { | |
8f4d37ec PZ |
3605 | struct sched_entity *se = &p->se; |
3606 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3607 | ||
3608 | WARN_ON(task_rq(p) != rq); | |
3609 | ||
b39e66ea | 3610 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3611 | u64 slice = sched_slice(cfs_rq, se); |
3612 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3613 | s64 delta = slice - ran; | |
3614 | ||
3615 | if (delta < 0) { | |
3616 | if (rq->curr == p) | |
3617 | resched_task(p); | |
3618 | return; | |
3619 | } | |
3620 | ||
3621 | /* | |
3622 | * Don't schedule slices shorter than 10000ns, that just | |
3623 | * doesn't make sense. Rely on vruntime for fairness. | |
3624 | */ | |
31656519 | 3625 | if (rq->curr != p) |
157124c1 | 3626 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3627 | |
31656519 | 3628 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3629 | } |
3630 | } | |
a4c2f00f PZ |
3631 | |
3632 | /* | |
3633 | * called from enqueue/dequeue and updates the hrtick when the | |
3634 | * current task is from our class and nr_running is low enough | |
3635 | * to matter. | |
3636 | */ | |
3637 | static void hrtick_update(struct rq *rq) | |
3638 | { | |
3639 | struct task_struct *curr = rq->curr; | |
3640 | ||
b39e66ea | 3641 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3642 | return; |
3643 | ||
3644 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3645 | hrtick_start_fair(rq, curr); | |
3646 | } | |
55e12e5e | 3647 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3648 | static inline void |
3649 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3650 | { | |
3651 | } | |
a4c2f00f PZ |
3652 | |
3653 | static inline void hrtick_update(struct rq *rq) | |
3654 | { | |
3655 | } | |
8f4d37ec PZ |
3656 | #endif |
3657 | ||
bf0f6f24 IM |
3658 | /* |
3659 | * The enqueue_task method is called before nr_running is | |
3660 | * increased. Here we update the fair scheduling stats and | |
3661 | * then put the task into the rbtree: | |
3662 | */ | |
ea87bb78 | 3663 | static void |
371fd7e7 | 3664 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3665 | { |
3666 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3667 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3668 | |
3669 | for_each_sched_entity(se) { | |
62fb1851 | 3670 | if (se->on_rq) |
bf0f6f24 IM |
3671 | break; |
3672 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3673 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3674 | |
3675 | /* | |
3676 | * end evaluation on encountering a throttled cfs_rq | |
3677 | * | |
3678 | * note: in the case of encountering a throttled cfs_rq we will | |
3679 | * post the final h_nr_running increment below. | |
3680 | */ | |
3681 | if (cfs_rq_throttled(cfs_rq)) | |
3682 | break; | |
953bfcd1 | 3683 | cfs_rq->h_nr_running++; |
85dac906 | 3684 | |
88ec22d3 | 3685 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3686 | } |
8f4d37ec | 3687 | |
2069dd75 | 3688 | for_each_sched_entity(se) { |
0f317143 | 3689 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3690 | cfs_rq->h_nr_running++; |
2069dd75 | 3691 | |
85dac906 PT |
3692 | if (cfs_rq_throttled(cfs_rq)) |
3693 | break; | |
3694 | ||
17bc14b7 | 3695 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3696 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3697 | } |
3698 | ||
18bf2805 BS |
3699 | if (!se) { |
3700 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3701 | inc_nr_running(rq); |
18bf2805 | 3702 | } |
a4c2f00f | 3703 | hrtick_update(rq); |
bf0f6f24 IM |
3704 | } |
3705 | ||
2f36825b VP |
3706 | static void set_next_buddy(struct sched_entity *se); |
3707 | ||
bf0f6f24 IM |
3708 | /* |
3709 | * The dequeue_task method is called before nr_running is | |
3710 | * decreased. We remove the task from the rbtree and | |
3711 | * update the fair scheduling stats: | |
3712 | */ | |
371fd7e7 | 3713 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3714 | { |
3715 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3716 | struct sched_entity *se = &p->se; |
2f36825b | 3717 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3718 | |
3719 | for_each_sched_entity(se) { | |
3720 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3721 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3722 | |
3723 | /* | |
3724 | * end evaluation on encountering a throttled cfs_rq | |
3725 | * | |
3726 | * note: in the case of encountering a throttled cfs_rq we will | |
3727 | * post the final h_nr_running decrement below. | |
3728 | */ | |
3729 | if (cfs_rq_throttled(cfs_rq)) | |
3730 | break; | |
953bfcd1 | 3731 | cfs_rq->h_nr_running--; |
2069dd75 | 3732 | |
bf0f6f24 | 3733 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3734 | if (cfs_rq->load.weight) { |
3735 | /* | |
3736 | * Bias pick_next to pick a task from this cfs_rq, as | |
3737 | * p is sleeping when it is within its sched_slice. | |
3738 | */ | |
3739 | if (task_sleep && parent_entity(se)) | |
3740 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3741 | |
3742 | /* avoid re-evaluating load for this entity */ | |
3743 | se = parent_entity(se); | |
bf0f6f24 | 3744 | break; |
2f36825b | 3745 | } |
371fd7e7 | 3746 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3747 | } |
8f4d37ec | 3748 | |
2069dd75 | 3749 | for_each_sched_entity(se) { |
0f317143 | 3750 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3751 | cfs_rq->h_nr_running--; |
2069dd75 | 3752 | |
85dac906 PT |
3753 | if (cfs_rq_throttled(cfs_rq)) |
3754 | break; | |
3755 | ||
17bc14b7 | 3756 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3757 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3758 | } |
3759 | ||
18bf2805 | 3760 | if (!se) { |
85dac906 | 3761 | dec_nr_running(rq); |
18bf2805 BS |
3762 | update_rq_runnable_avg(rq, 1); |
3763 | } | |
a4c2f00f | 3764 | hrtick_update(rq); |
bf0f6f24 IM |
3765 | } |
3766 | ||
e7693a36 | 3767 | #ifdef CONFIG_SMP |
029632fb PZ |
3768 | /* Used instead of source_load when we know the type == 0 */ |
3769 | static unsigned long weighted_cpuload(const int cpu) | |
3770 | { | |
b92486cb | 3771 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3772 | } |
3773 | ||
3774 | /* | |
3775 | * Return a low guess at the load of a migration-source cpu weighted | |
3776 | * according to the scheduling class and "nice" value. | |
3777 | * | |
3778 | * We want to under-estimate the load of migration sources, to | |
3779 | * balance conservatively. | |
3780 | */ | |
3781 | static unsigned long source_load(int cpu, int type) | |
3782 | { | |
3783 | struct rq *rq = cpu_rq(cpu); | |
3784 | unsigned long total = weighted_cpuload(cpu); | |
3785 | ||
3786 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3787 | return total; | |
3788 | ||
3789 | return min(rq->cpu_load[type-1], total); | |
3790 | } | |
3791 | ||
3792 | /* | |
3793 | * Return a high guess at the load of a migration-target cpu weighted | |
3794 | * according to the scheduling class and "nice" value. | |
3795 | */ | |
3796 | static unsigned long target_load(int cpu, int type) | |
3797 | { | |
3798 | struct rq *rq = cpu_rq(cpu); | |
3799 | unsigned long total = weighted_cpuload(cpu); | |
3800 | ||
3801 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3802 | return total; | |
3803 | ||
3804 | return max(rq->cpu_load[type-1], total); | |
3805 | } | |
3806 | ||
3807 | static unsigned long power_of(int cpu) | |
3808 | { | |
3809 | return cpu_rq(cpu)->cpu_power; | |
3810 | } | |
3811 | ||
3812 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3813 | { | |
3814 | struct rq *rq = cpu_rq(cpu); | |
3815 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3816 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3817 | |
3818 | if (nr_running) | |
b92486cb | 3819 | return load_avg / nr_running; |
029632fb PZ |
3820 | |
3821 | return 0; | |
3822 | } | |
3823 | ||
62470419 MW |
3824 | static void record_wakee(struct task_struct *p) |
3825 | { | |
3826 | /* | |
3827 | * Rough decay (wiping) for cost saving, don't worry | |
3828 | * about the boundary, really active task won't care | |
3829 | * about the loss. | |
3830 | */ | |
3831 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
3832 | current->wakee_flips = 0; | |
3833 | current->wakee_flip_decay_ts = jiffies; | |
3834 | } | |
3835 | ||
3836 | if (current->last_wakee != p) { | |
3837 | current->last_wakee = p; | |
3838 | current->wakee_flips++; | |
3839 | } | |
3840 | } | |
098fb9db | 3841 | |
74f8e4b2 | 3842 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
3843 | { |
3844 | struct sched_entity *se = &p->se; | |
3845 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
3846 | u64 min_vruntime; |
3847 | ||
3848 | #ifndef CONFIG_64BIT | |
3849 | u64 min_vruntime_copy; | |
88ec22d3 | 3850 | |
3fe1698b PZ |
3851 | do { |
3852 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3853 | smp_rmb(); | |
3854 | min_vruntime = cfs_rq->min_vruntime; | |
3855 | } while (min_vruntime != min_vruntime_copy); | |
3856 | #else | |
3857 | min_vruntime = cfs_rq->min_vruntime; | |
3858 | #endif | |
88ec22d3 | 3859 | |
3fe1698b | 3860 | se->vruntime -= min_vruntime; |
62470419 | 3861 | record_wakee(p); |
88ec22d3 PZ |
3862 | } |
3863 | ||
bb3469ac | 3864 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3865 | /* |
3866 | * effective_load() calculates the load change as seen from the root_task_group | |
3867 | * | |
3868 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3869 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3870 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3871 | * |
3872 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3873 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3874 | * total group weight. | |
3875 | * | |
3876 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3877 | * distribution (s_i) using: | |
3878 | * | |
3879 | * s_i = rw_i / \Sum rw_j (1) | |
3880 | * | |
3881 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3882 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3883 | * shares distribution (s_i): | |
3884 | * | |
3885 | * rw_i = { 2, 4, 1, 0 } | |
3886 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3887 | * | |
3888 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3889 | * task used to run on and the CPU the waker is running on), we need to | |
3890 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3891 | * wakeup, compute the effect of the current task going to sleep. | |
3892 | * | |
3893 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3894 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3895 | * | |
3896 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3897 | * | |
3898 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3899 | * differences in waking a task to CPU 0. The additional task changes the | |
3900 | * weight and shares distributions like: | |
3901 | * | |
3902 | * rw'_i = { 3, 4, 1, 0 } | |
3903 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3904 | * | |
3905 | * We can then compute the difference in effective weight by using: | |
3906 | * | |
3907 | * dw_i = S * (s'_i - s_i) (3) | |
3908 | * | |
3909 | * Where 'S' is the group weight as seen by its parent. | |
3910 | * | |
3911 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3912 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3913 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3914 | */ |
2069dd75 | 3915 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3916 | { |
4be9daaa | 3917 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3918 | |
9722c2da | 3919 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3920 | return wl; |
3921 | ||
4be9daaa | 3922 | for_each_sched_entity(se) { |
cf5f0acf | 3923 | long w, W; |
4be9daaa | 3924 | |
977dda7c | 3925 | tg = se->my_q->tg; |
bb3469ac | 3926 | |
cf5f0acf PZ |
3927 | /* |
3928 | * W = @wg + \Sum rw_j | |
3929 | */ | |
3930 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3931 | |
cf5f0acf PZ |
3932 | /* |
3933 | * w = rw_i + @wl | |
3934 | */ | |
3935 | w = se->my_q->load.weight + wl; | |
940959e9 | 3936 | |
cf5f0acf PZ |
3937 | /* |
3938 | * wl = S * s'_i; see (2) | |
3939 | */ | |
3940 | if (W > 0 && w < W) | |
3941 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3942 | else |
3943 | wl = tg->shares; | |
940959e9 | 3944 | |
cf5f0acf PZ |
3945 | /* |
3946 | * Per the above, wl is the new se->load.weight value; since | |
3947 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3948 | * calc_cfs_shares(). | |
3949 | */ | |
977dda7c PT |
3950 | if (wl < MIN_SHARES) |
3951 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3952 | |
3953 | /* | |
3954 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3955 | */ | |
977dda7c | 3956 | wl -= se->load.weight; |
cf5f0acf PZ |
3957 | |
3958 | /* | |
3959 | * Recursively apply this logic to all parent groups to compute | |
3960 | * the final effective load change on the root group. Since | |
3961 | * only the @tg group gets extra weight, all parent groups can | |
3962 | * only redistribute existing shares. @wl is the shift in shares | |
3963 | * resulting from this level per the above. | |
3964 | */ | |
4be9daaa | 3965 | wg = 0; |
4be9daaa | 3966 | } |
bb3469ac | 3967 | |
4be9daaa | 3968 | return wl; |
bb3469ac PZ |
3969 | } |
3970 | #else | |
4be9daaa | 3971 | |
58d081b5 | 3972 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 3973 | { |
83378269 | 3974 | return wl; |
bb3469ac | 3975 | } |
4be9daaa | 3976 | |
bb3469ac PZ |
3977 | #endif |
3978 | ||
62470419 MW |
3979 | static int wake_wide(struct task_struct *p) |
3980 | { | |
7d9ffa89 | 3981 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
3982 | |
3983 | /* | |
3984 | * Yeah, it's the switching-frequency, could means many wakee or | |
3985 | * rapidly switch, use factor here will just help to automatically | |
3986 | * adjust the loose-degree, so bigger node will lead to more pull. | |
3987 | */ | |
3988 | if (p->wakee_flips > factor) { | |
3989 | /* | |
3990 | * wakee is somewhat hot, it needs certain amount of cpu | |
3991 | * resource, so if waker is far more hot, prefer to leave | |
3992 | * it alone. | |
3993 | */ | |
3994 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
3995 | return 1; | |
3996 | } | |
3997 | ||
3998 | return 0; | |
3999 | } | |
4000 | ||
c88d5910 | 4001 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 4002 | { |
e37b6a7b | 4003 | s64 this_load, load; |
c88d5910 | 4004 | int idx, this_cpu, prev_cpu; |
098fb9db | 4005 | unsigned long tl_per_task; |
c88d5910 | 4006 | struct task_group *tg; |
83378269 | 4007 | unsigned long weight; |
b3137bc8 | 4008 | int balanced; |
098fb9db | 4009 | |
62470419 MW |
4010 | /* |
4011 | * If we wake multiple tasks be careful to not bounce | |
4012 | * ourselves around too much. | |
4013 | */ | |
4014 | if (wake_wide(p)) | |
4015 | return 0; | |
4016 | ||
c88d5910 PZ |
4017 | idx = sd->wake_idx; |
4018 | this_cpu = smp_processor_id(); | |
4019 | prev_cpu = task_cpu(p); | |
4020 | load = source_load(prev_cpu, idx); | |
4021 | this_load = target_load(this_cpu, idx); | |
098fb9db | 4022 | |
b3137bc8 MG |
4023 | /* |
4024 | * If sync wakeup then subtract the (maximum possible) | |
4025 | * effect of the currently running task from the load | |
4026 | * of the current CPU: | |
4027 | */ | |
83378269 PZ |
4028 | if (sync) { |
4029 | tg = task_group(current); | |
4030 | weight = current->se.load.weight; | |
4031 | ||
c88d5910 | 4032 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
4033 | load += effective_load(tg, prev_cpu, 0, -weight); |
4034 | } | |
b3137bc8 | 4035 | |
83378269 PZ |
4036 | tg = task_group(p); |
4037 | weight = p->se.load.weight; | |
b3137bc8 | 4038 | |
71a29aa7 PZ |
4039 | /* |
4040 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
4041 | * due to the sync cause above having dropped this_load to 0, we'll |
4042 | * always have an imbalance, but there's really nothing you can do | |
4043 | * about that, so that's good too. | |
71a29aa7 PZ |
4044 | * |
4045 | * Otherwise check if either cpus are near enough in load to allow this | |
4046 | * task to be woken on this_cpu. | |
4047 | */ | |
e37b6a7b PT |
4048 | if (this_load > 0) { |
4049 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
4050 | |
4051 | this_eff_load = 100; | |
4052 | this_eff_load *= power_of(prev_cpu); | |
4053 | this_eff_load *= this_load + | |
4054 | effective_load(tg, this_cpu, weight, weight); | |
4055 | ||
4056 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
4057 | prev_eff_load *= power_of(this_cpu); | |
4058 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
4059 | ||
4060 | balanced = this_eff_load <= prev_eff_load; | |
4061 | } else | |
4062 | balanced = true; | |
b3137bc8 | 4063 | |
098fb9db | 4064 | /* |
4ae7d5ce IM |
4065 | * If the currently running task will sleep within |
4066 | * a reasonable amount of time then attract this newly | |
4067 | * woken task: | |
098fb9db | 4068 | */ |
2fb7635c PZ |
4069 | if (sync && balanced) |
4070 | return 1; | |
098fb9db | 4071 | |
41acab88 | 4072 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
4073 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
4074 | ||
c88d5910 PZ |
4075 | if (balanced || |
4076 | (this_load <= load && | |
4077 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
4078 | /* |
4079 | * This domain has SD_WAKE_AFFINE and | |
4080 | * p is cache cold in this domain, and | |
4081 | * there is no bad imbalance. | |
4082 | */ | |
c88d5910 | 4083 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 4084 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
4085 | |
4086 | return 1; | |
4087 | } | |
4088 | return 0; | |
4089 | } | |
4090 | ||
aaee1203 PZ |
4091 | /* |
4092 | * find_idlest_group finds and returns the least busy CPU group within the | |
4093 | * domain. | |
4094 | */ | |
4095 | static struct sched_group * | |
78e7ed53 | 4096 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 4097 | int this_cpu, int sd_flag) |
e7693a36 | 4098 | { |
b3bd3de6 | 4099 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 4100 | unsigned long min_load = ULONG_MAX, this_load = 0; |
c44f2a02 | 4101 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4102 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 4103 | |
c44f2a02 VG |
4104 | if (sd_flag & SD_BALANCE_WAKE) |
4105 | load_idx = sd->wake_idx; | |
4106 | ||
aaee1203 PZ |
4107 | do { |
4108 | unsigned long load, avg_load; | |
4109 | int local_group; | |
4110 | int i; | |
e7693a36 | 4111 | |
aaee1203 PZ |
4112 | /* Skip over this group if it has no CPUs allowed */ |
4113 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 4114 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
4115 | continue; |
4116 | ||
4117 | local_group = cpumask_test_cpu(this_cpu, | |
4118 | sched_group_cpus(group)); | |
4119 | ||
4120 | /* Tally up the load of all CPUs in the group */ | |
4121 | avg_load = 0; | |
4122 | ||
4123 | for_each_cpu(i, sched_group_cpus(group)) { | |
4124 | /* Bias balancing toward cpus of our domain */ | |
4125 | if (local_group) | |
4126 | load = source_load(i, load_idx); | |
4127 | else | |
4128 | load = target_load(i, load_idx); | |
4129 | ||
4130 | avg_load += load; | |
4131 | } | |
4132 | ||
4133 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 4134 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
4135 | |
4136 | if (local_group) { | |
4137 | this_load = avg_load; | |
aaee1203 PZ |
4138 | } else if (avg_load < min_load) { |
4139 | min_load = avg_load; | |
4140 | idlest = group; | |
4141 | } | |
4142 | } while (group = group->next, group != sd->groups); | |
4143 | ||
4144 | if (!idlest || 100*this_load < imbalance*min_load) | |
4145 | return NULL; | |
4146 | return idlest; | |
4147 | } | |
4148 | ||
4149 | /* | |
4150 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
4151 | */ | |
4152 | static int | |
4153 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
4154 | { | |
4155 | unsigned long load, min_load = ULONG_MAX; | |
4156 | int idlest = -1; | |
4157 | int i; | |
4158 | ||
4159 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 4160 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
4161 | load = weighted_cpuload(i); |
4162 | ||
4163 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
4164 | min_load = load; | |
4165 | idlest = i; | |
e7693a36 GH |
4166 | } |
4167 | } | |
4168 | ||
aaee1203 PZ |
4169 | return idlest; |
4170 | } | |
e7693a36 | 4171 | |
a50bde51 PZ |
4172 | /* |
4173 | * Try and locate an idle CPU in the sched_domain. | |
4174 | */ | |
99bd5e2f | 4175 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 4176 | { |
99bd5e2f | 4177 | struct sched_domain *sd; |
37407ea7 | 4178 | struct sched_group *sg; |
e0a79f52 | 4179 | int i = task_cpu(p); |
a50bde51 | 4180 | |
e0a79f52 MG |
4181 | if (idle_cpu(target)) |
4182 | return target; | |
99bd5e2f SS |
4183 | |
4184 | /* | |
e0a79f52 | 4185 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 4186 | */ |
e0a79f52 MG |
4187 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4188 | return i; | |
a50bde51 PZ |
4189 | |
4190 | /* | |
37407ea7 | 4191 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 4192 | */ |
518cd623 | 4193 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 4194 | for_each_lower_domain(sd) { |
37407ea7 LT |
4195 | sg = sd->groups; |
4196 | do { | |
4197 | if (!cpumask_intersects(sched_group_cpus(sg), | |
4198 | tsk_cpus_allowed(p))) | |
4199 | goto next; | |
4200 | ||
4201 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 4202 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
4203 | goto next; |
4204 | } | |
970e1789 | 4205 | |
37407ea7 LT |
4206 | target = cpumask_first_and(sched_group_cpus(sg), |
4207 | tsk_cpus_allowed(p)); | |
4208 | goto done; | |
4209 | next: | |
4210 | sg = sg->next; | |
4211 | } while (sg != sd->groups); | |
4212 | } | |
4213 | done: | |
a50bde51 PZ |
4214 | return target; |
4215 | } | |
4216 | ||
aaee1203 PZ |
4217 | /* |
4218 | * sched_balance_self: balance the current task (running on cpu) in domains | |
4219 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
4220 | * SD_BALANCE_EXEC. | |
4221 | * | |
4222 | * Balance, ie. select the least loaded group. | |
4223 | * | |
4224 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
4225 | * | |
4226 | * preempt must be disabled. | |
4227 | */ | |
0017d735 | 4228 | static int |
ac66f547 | 4229 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 4230 | { |
29cd8bae | 4231 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 4232 | int cpu = smp_processor_id(); |
c88d5910 | 4233 | int new_cpu = cpu; |
99bd5e2f | 4234 | int want_affine = 0; |
5158f4e4 | 4235 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 4236 | |
29baa747 | 4237 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
4238 | return prev_cpu; |
4239 | ||
0763a660 | 4240 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 4241 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
4242 | want_affine = 1; |
4243 | new_cpu = prev_cpu; | |
4244 | } | |
aaee1203 | 4245 | |
dce840a0 | 4246 | rcu_read_lock(); |
aaee1203 | 4247 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4248 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4249 | continue; | |
4250 | ||
fe3bcfe1 | 4251 | /* |
99bd5e2f SS |
4252 | * If both cpu and prev_cpu are part of this domain, |
4253 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4254 | */ |
99bd5e2f SS |
4255 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4256 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4257 | affine_sd = tmp; | |
29cd8bae | 4258 | break; |
f03542a7 | 4259 | } |
29cd8bae | 4260 | |
f03542a7 | 4261 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4262 | sd = tmp; |
4263 | } | |
4264 | ||
8b911acd | 4265 | if (affine_sd) { |
f03542a7 | 4266 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
4267 | prev_cpu = cpu; |
4268 | ||
4269 | new_cpu = select_idle_sibling(p, prev_cpu); | |
4270 | goto unlock; | |
8b911acd | 4271 | } |
e7693a36 | 4272 | |
aaee1203 PZ |
4273 | while (sd) { |
4274 | struct sched_group *group; | |
c88d5910 | 4275 | int weight; |
098fb9db | 4276 | |
0763a660 | 4277 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4278 | sd = sd->child; |
4279 | continue; | |
4280 | } | |
098fb9db | 4281 | |
c44f2a02 | 4282 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
4283 | if (!group) { |
4284 | sd = sd->child; | |
4285 | continue; | |
4286 | } | |
4ae7d5ce | 4287 | |
d7c33c49 | 4288 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4289 | if (new_cpu == -1 || new_cpu == cpu) { |
4290 | /* Now try balancing at a lower domain level of cpu */ | |
4291 | sd = sd->child; | |
4292 | continue; | |
e7693a36 | 4293 | } |
aaee1203 PZ |
4294 | |
4295 | /* Now try balancing at a lower domain level of new_cpu */ | |
4296 | cpu = new_cpu; | |
669c55e9 | 4297 | weight = sd->span_weight; |
aaee1203 PZ |
4298 | sd = NULL; |
4299 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4300 | if (weight <= tmp->span_weight) |
aaee1203 | 4301 | break; |
0763a660 | 4302 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4303 | sd = tmp; |
4304 | } | |
4305 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4306 | } |
dce840a0 PZ |
4307 | unlock: |
4308 | rcu_read_unlock(); | |
e7693a36 | 4309 | |
c88d5910 | 4310 | return new_cpu; |
e7693a36 | 4311 | } |
0a74bef8 PT |
4312 | |
4313 | /* | |
4314 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4315 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4316 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4317 | * other assumptions, including the state of rq->lock, should be made. | |
4318 | */ | |
4319 | static void | |
4320 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4321 | { | |
aff3e498 PT |
4322 | struct sched_entity *se = &p->se; |
4323 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4324 | ||
4325 | /* | |
4326 | * Load tracking: accumulate removed load so that it can be processed | |
4327 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4328 | * to blocked load iff they have a positive decay-count. It can never | |
4329 | * be negative here since on-rq tasks have decay-count == 0. | |
4330 | */ | |
4331 | if (se->avg.decay_count) { | |
4332 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4333 | atomic_long_add(se->avg.load_avg_contrib, |
4334 | &cfs_rq->removed_load); | |
aff3e498 | 4335 | } |
0a74bef8 | 4336 | } |
e7693a36 GH |
4337 | #endif /* CONFIG_SMP */ |
4338 | ||
e52fb7c0 PZ |
4339 | static unsigned long |
4340 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4341 | { |
4342 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4343 | ||
4344 | /* | |
e52fb7c0 PZ |
4345 | * Since its curr running now, convert the gran from real-time |
4346 | * to virtual-time in his units. | |
13814d42 MG |
4347 | * |
4348 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4349 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4350 | * the resulting gran will be larger, therefore penalizing the | |
4351 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4352 | * be smaller, again penalizing the lighter task. | |
4353 | * | |
4354 | * This is especially important for buddies when the leftmost | |
4355 | * task is higher priority than the buddy. | |
0bbd3336 | 4356 | */ |
f4ad9bd2 | 4357 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4358 | } |
4359 | ||
464b7527 PZ |
4360 | /* |
4361 | * Should 'se' preempt 'curr'. | |
4362 | * | |
4363 | * |s1 | |
4364 | * |s2 | |
4365 | * |s3 | |
4366 | * g | |
4367 | * |<--->|c | |
4368 | * | |
4369 | * w(c, s1) = -1 | |
4370 | * w(c, s2) = 0 | |
4371 | * w(c, s3) = 1 | |
4372 | * | |
4373 | */ | |
4374 | static int | |
4375 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4376 | { | |
4377 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4378 | ||
4379 | if (vdiff <= 0) | |
4380 | return -1; | |
4381 | ||
e52fb7c0 | 4382 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4383 | if (vdiff > gran) |
4384 | return 1; | |
4385 | ||
4386 | return 0; | |
4387 | } | |
4388 | ||
02479099 PZ |
4389 | static void set_last_buddy(struct sched_entity *se) |
4390 | { | |
69c80f3e VP |
4391 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4392 | return; | |
4393 | ||
4394 | for_each_sched_entity(se) | |
4395 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4396 | } |
4397 | ||
4398 | static void set_next_buddy(struct sched_entity *se) | |
4399 | { | |
69c80f3e VP |
4400 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4401 | return; | |
4402 | ||
4403 | for_each_sched_entity(se) | |
4404 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4405 | } |
4406 | ||
ac53db59 RR |
4407 | static void set_skip_buddy(struct sched_entity *se) |
4408 | { | |
69c80f3e VP |
4409 | for_each_sched_entity(se) |
4410 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4411 | } |
4412 | ||
bf0f6f24 IM |
4413 | /* |
4414 | * Preempt the current task with a newly woken task if needed: | |
4415 | */ | |
5a9b86f6 | 4416 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4417 | { |
4418 | struct task_struct *curr = rq->curr; | |
8651a86c | 4419 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4420 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4421 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4422 | int next_buddy_marked = 0; |
bf0f6f24 | 4423 | |
4ae7d5ce IM |
4424 | if (unlikely(se == pse)) |
4425 | return; | |
4426 | ||
5238cdd3 | 4427 | /* |
ddcdf6e7 | 4428 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
4429 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4430 | * lead to a throttle). This both saves work and prevents false | |
4431 | * next-buddy nomination below. | |
4432 | */ | |
4433 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4434 | return; | |
4435 | ||
2f36825b | 4436 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4437 | set_next_buddy(pse); |
2f36825b VP |
4438 | next_buddy_marked = 1; |
4439 | } | |
57fdc26d | 4440 | |
aec0a514 BR |
4441 | /* |
4442 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4443 | * wake up path. | |
5238cdd3 PT |
4444 | * |
4445 | * Note: this also catches the edge-case of curr being in a throttled | |
4446 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4447 | * enqueue of curr) will have resulted in resched being set. This | |
4448 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4449 | * below. | |
aec0a514 BR |
4450 | */ |
4451 | if (test_tsk_need_resched(curr)) | |
4452 | return; | |
4453 | ||
a2f5c9ab DH |
4454 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4455 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4456 | likely(p->policy != SCHED_IDLE)) | |
4457 | goto preempt; | |
4458 | ||
91c234b4 | 4459 | /* |
a2f5c9ab DH |
4460 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4461 | * is driven by the tick): | |
91c234b4 | 4462 | */ |
8ed92e51 | 4463 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4464 | return; |
bf0f6f24 | 4465 | |
464b7527 | 4466 | find_matching_se(&se, &pse); |
9bbd7374 | 4467 | update_curr(cfs_rq_of(se)); |
002f128b | 4468 | BUG_ON(!pse); |
2f36825b VP |
4469 | if (wakeup_preempt_entity(se, pse) == 1) { |
4470 | /* | |
4471 | * Bias pick_next to pick the sched entity that is | |
4472 | * triggering this preemption. | |
4473 | */ | |
4474 | if (!next_buddy_marked) | |
4475 | set_next_buddy(pse); | |
3a7e73a2 | 4476 | goto preempt; |
2f36825b | 4477 | } |
464b7527 | 4478 | |
3a7e73a2 | 4479 | return; |
a65ac745 | 4480 | |
3a7e73a2 PZ |
4481 | preempt: |
4482 | resched_task(curr); | |
4483 | /* | |
4484 | * Only set the backward buddy when the current task is still | |
4485 | * on the rq. This can happen when a wakeup gets interleaved | |
4486 | * with schedule on the ->pre_schedule() or idle_balance() | |
4487 | * point, either of which can * drop the rq lock. | |
4488 | * | |
4489 | * Also, during early boot the idle thread is in the fair class, | |
4490 | * for obvious reasons its a bad idea to schedule back to it. | |
4491 | */ | |
4492 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
4493 | return; | |
4494 | ||
4495 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
4496 | set_last_buddy(se); | |
bf0f6f24 IM |
4497 | } |
4498 | ||
fb8d4724 | 4499 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 4500 | { |
8f4d37ec | 4501 | struct task_struct *p; |
bf0f6f24 IM |
4502 | struct cfs_rq *cfs_rq = &rq->cfs; |
4503 | struct sched_entity *se; | |
4504 | ||
36ace27e | 4505 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
4506 | return NULL; |
4507 | ||
4508 | do { | |
9948f4b2 | 4509 | se = pick_next_entity(cfs_rq); |
f4b6755f | 4510 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
4511 | cfs_rq = group_cfs_rq(se); |
4512 | } while (cfs_rq); | |
4513 | ||
8f4d37ec | 4514 | p = task_of(se); |
b39e66ea MG |
4515 | if (hrtick_enabled(rq)) |
4516 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4517 | |
4518 | return p; | |
bf0f6f24 IM |
4519 | } |
4520 | ||
4521 | /* | |
4522 | * Account for a descheduled task: | |
4523 | */ | |
31ee529c | 4524 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4525 | { |
4526 | struct sched_entity *se = &prev->se; | |
4527 | struct cfs_rq *cfs_rq; | |
4528 | ||
4529 | for_each_sched_entity(se) { | |
4530 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4531 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4532 | } |
4533 | } | |
4534 | ||
ac53db59 RR |
4535 | /* |
4536 | * sched_yield() is very simple | |
4537 | * | |
4538 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4539 | */ | |
4540 | static void yield_task_fair(struct rq *rq) | |
4541 | { | |
4542 | struct task_struct *curr = rq->curr; | |
4543 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4544 | struct sched_entity *se = &curr->se; | |
4545 | ||
4546 | /* | |
4547 | * Are we the only task in the tree? | |
4548 | */ | |
4549 | if (unlikely(rq->nr_running == 1)) | |
4550 | return; | |
4551 | ||
4552 | clear_buddies(cfs_rq, se); | |
4553 | ||
4554 | if (curr->policy != SCHED_BATCH) { | |
4555 | update_rq_clock(rq); | |
4556 | /* | |
4557 | * Update run-time statistics of the 'current'. | |
4558 | */ | |
4559 | update_curr(cfs_rq); | |
916671c0 MG |
4560 | /* |
4561 | * Tell update_rq_clock() that we've just updated, | |
4562 | * so we don't do microscopic update in schedule() | |
4563 | * and double the fastpath cost. | |
4564 | */ | |
4565 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4566 | } |
4567 | ||
4568 | set_skip_buddy(se); | |
4569 | } | |
4570 | ||
d95f4122 MG |
4571 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4572 | { | |
4573 | struct sched_entity *se = &p->se; | |
4574 | ||
5238cdd3 PT |
4575 | /* throttled hierarchies are not runnable */ |
4576 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4577 | return false; |
4578 | ||
4579 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4580 | set_next_buddy(se); | |
4581 | ||
d95f4122 MG |
4582 | yield_task_fair(rq); |
4583 | ||
4584 | return true; | |
4585 | } | |
4586 | ||
681f3e68 | 4587 | #ifdef CONFIG_SMP |
bf0f6f24 | 4588 | /************************************************** |
e9c84cb8 PZ |
4589 | * Fair scheduling class load-balancing methods. |
4590 | * | |
4591 | * BASICS | |
4592 | * | |
4593 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4594 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4595 | * time to each task. This is expressed in the following equation: | |
4596 | * | |
4597 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4598 | * | |
4599 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4600 | * W_i,0 is defined as: | |
4601 | * | |
4602 | * W_i,0 = \Sum_j w_i,j (2) | |
4603 | * | |
4604 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4605 | * is derived from the nice value as per prio_to_weight[]. | |
4606 | * | |
4607 | * The weight average is an exponential decay average of the instantaneous | |
4608 | * weight: | |
4609 | * | |
4610 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4611 | * | |
4612 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4613 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4614 | * can also include other factors [XXX]. | |
4615 | * | |
4616 | * To achieve this balance we define a measure of imbalance which follows | |
4617 | * directly from (1): | |
4618 | * | |
4619 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4620 | * | |
4621 | * We them move tasks around to minimize the imbalance. In the continuous | |
4622 | * function space it is obvious this converges, in the discrete case we get | |
4623 | * a few fun cases generally called infeasible weight scenarios. | |
4624 | * | |
4625 | * [XXX expand on: | |
4626 | * - infeasible weights; | |
4627 | * - local vs global optima in the discrete case. ] | |
4628 | * | |
4629 | * | |
4630 | * SCHED DOMAINS | |
4631 | * | |
4632 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4633 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4634 | * topology where each level pairs two lower groups (or better). This results | |
4635 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4636 | * tree to only the first of the previous level and we decrease the frequency | |
4637 | * of load-balance at each level inv. proportional to the number of cpus in | |
4638 | * the groups. | |
4639 | * | |
4640 | * This yields: | |
4641 | * | |
4642 | * log_2 n 1 n | |
4643 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4644 | * i = 0 2^i 2^i | |
4645 | * `- size of each group | |
4646 | * | | `- number of cpus doing load-balance | |
4647 | * | `- freq | |
4648 | * `- sum over all levels | |
4649 | * | |
4650 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4651 | * this makes (5) the runtime complexity of the balancer. | |
4652 | * | |
4653 | * An important property here is that each CPU is still (indirectly) connected | |
4654 | * to every other cpu in at most O(log n) steps: | |
4655 | * | |
4656 | * The adjacency matrix of the resulting graph is given by: | |
4657 | * | |
4658 | * log_2 n | |
4659 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4660 | * k = 0 | |
4661 | * | |
4662 | * And you'll find that: | |
4663 | * | |
4664 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
4665 | * | |
4666 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
4667 | * The task movement gives a factor of O(m), giving a convergence complexity | |
4668 | * of: | |
4669 | * | |
4670 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
4671 | * | |
4672 | * | |
4673 | * WORK CONSERVING | |
4674 | * | |
4675 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
4676 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
4677 | * tree itself instead of relying on other CPUs to bring it work. | |
4678 | * | |
4679 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
4680 | * time. | |
4681 | * | |
4682 | * [XXX more?] | |
4683 | * | |
4684 | * | |
4685 | * CGROUPS | |
4686 | * | |
4687 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
4688 | * | |
4689 | * s_k,i | |
4690 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
4691 | * S_k | |
4692 | * | |
4693 | * Where | |
4694 | * | |
4695 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4696 | * | |
4697 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4698 | * | |
4699 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4700 | * property. | |
4701 | * | |
4702 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4703 | * rewrite all of this once again.] | |
4704 | */ | |
bf0f6f24 | 4705 | |
ed387b78 HS |
4706 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4707 | ||
0ec8aa00 PZ |
4708 | enum fbq_type { regular, remote, all }; |
4709 | ||
ddcdf6e7 | 4710 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4711 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4712 | #define LBF_DST_PINNED 0x04 |
4713 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4714 | |
4715 | struct lb_env { | |
4716 | struct sched_domain *sd; | |
4717 | ||
ddcdf6e7 | 4718 | struct rq *src_rq; |
85c1e7da | 4719 | int src_cpu; |
ddcdf6e7 PZ |
4720 | |
4721 | int dst_cpu; | |
4722 | struct rq *dst_rq; | |
4723 | ||
88b8dac0 SV |
4724 | struct cpumask *dst_grpmask; |
4725 | int new_dst_cpu; | |
ddcdf6e7 | 4726 | enum cpu_idle_type idle; |
bd939f45 | 4727 | long imbalance; |
b9403130 MW |
4728 | /* The set of CPUs under consideration for load-balancing */ |
4729 | struct cpumask *cpus; | |
4730 | ||
ddcdf6e7 | 4731 | unsigned int flags; |
367456c7 PZ |
4732 | |
4733 | unsigned int loop; | |
4734 | unsigned int loop_break; | |
4735 | unsigned int loop_max; | |
0ec8aa00 PZ |
4736 | |
4737 | enum fbq_type fbq_type; | |
ddcdf6e7 PZ |
4738 | }; |
4739 | ||
1e3c88bd | 4740 | /* |
ddcdf6e7 | 4741 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
4742 | * Both runqueues must be locked. |
4743 | */ | |
ddcdf6e7 | 4744 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 4745 | { |
ddcdf6e7 PZ |
4746 | deactivate_task(env->src_rq, p, 0); |
4747 | set_task_cpu(p, env->dst_cpu); | |
4748 | activate_task(env->dst_rq, p, 0); | |
4749 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
4750 | } |
4751 | ||
029632fb PZ |
4752 | /* |
4753 | * Is this task likely cache-hot: | |
4754 | */ | |
4755 | static int | |
4756 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
4757 | { | |
4758 | s64 delta; | |
4759 | ||
4760 | if (p->sched_class != &fair_sched_class) | |
4761 | return 0; | |
4762 | ||
4763 | if (unlikely(p->policy == SCHED_IDLE)) | |
4764 | return 0; | |
4765 | ||
4766 | /* | |
4767 | * Buddy candidates are cache hot: | |
4768 | */ | |
4769 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
4770 | (&p->se == cfs_rq_of(&p->se)->next || | |
4771 | &p->se == cfs_rq_of(&p->se)->last)) | |
4772 | return 1; | |
4773 | ||
4774 | if (sysctl_sched_migration_cost == -1) | |
4775 | return 1; | |
4776 | if (sysctl_sched_migration_cost == 0) | |
4777 | return 0; | |
4778 | ||
4779 | delta = now - p->se.exec_start; | |
4780 | ||
4781 | return delta < (s64)sysctl_sched_migration_cost; | |
4782 | } | |
4783 | ||
3a7053b3 MG |
4784 | #ifdef CONFIG_NUMA_BALANCING |
4785 | /* Returns true if the destination node has incurred more faults */ | |
4786 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
4787 | { | |
4788 | int src_nid, dst_nid; | |
4789 | ||
ff1df896 | 4790 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory || |
3a7053b3 MG |
4791 | !(env->sd->flags & SD_NUMA)) { |
4792 | return false; | |
4793 | } | |
4794 | ||
4795 | src_nid = cpu_to_node(env->src_cpu); | |
4796 | dst_nid = cpu_to_node(env->dst_cpu); | |
4797 | ||
83e1d2cd | 4798 | if (src_nid == dst_nid) |
3a7053b3 MG |
4799 | return false; |
4800 | ||
83e1d2cd MG |
4801 | /* Always encourage migration to the preferred node. */ |
4802 | if (dst_nid == p->numa_preferred_nid) | |
4803 | return true; | |
4804 | ||
887c290e RR |
4805 | /* If both task and group weight improve, this move is a winner. */ |
4806 | if (task_weight(p, dst_nid) > task_weight(p, src_nid) && | |
4807 | group_weight(p, dst_nid) > group_weight(p, src_nid)) | |
3a7053b3 MG |
4808 | return true; |
4809 | ||
4810 | return false; | |
4811 | } | |
7a0f3083 MG |
4812 | |
4813 | ||
4814 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
4815 | { | |
4816 | int src_nid, dst_nid; | |
4817 | ||
4818 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
4819 | return false; | |
4820 | ||
ff1df896 | 4821 | if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA)) |
7a0f3083 MG |
4822 | return false; |
4823 | ||
4824 | src_nid = cpu_to_node(env->src_cpu); | |
4825 | dst_nid = cpu_to_node(env->dst_cpu); | |
4826 | ||
83e1d2cd | 4827 | if (src_nid == dst_nid) |
7a0f3083 MG |
4828 | return false; |
4829 | ||
83e1d2cd MG |
4830 | /* Migrating away from the preferred node is always bad. */ |
4831 | if (src_nid == p->numa_preferred_nid) | |
4832 | return true; | |
4833 | ||
887c290e RR |
4834 | /* If either task or group weight get worse, don't do it. */ |
4835 | if (task_weight(p, dst_nid) < task_weight(p, src_nid) || | |
4836 | group_weight(p, dst_nid) < group_weight(p, src_nid)) | |
7a0f3083 MG |
4837 | return true; |
4838 | ||
4839 | return false; | |
4840 | } | |
4841 | ||
3a7053b3 MG |
4842 | #else |
4843 | static inline bool migrate_improves_locality(struct task_struct *p, | |
4844 | struct lb_env *env) | |
4845 | { | |
4846 | return false; | |
4847 | } | |
7a0f3083 MG |
4848 | |
4849 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
4850 | struct lb_env *env) | |
4851 | { | |
4852 | return false; | |
4853 | } | |
3a7053b3 MG |
4854 | #endif |
4855 | ||
1e3c88bd PZ |
4856 | /* |
4857 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
4858 | */ | |
4859 | static | |
8e45cb54 | 4860 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
4861 | { |
4862 | int tsk_cache_hot = 0; | |
4863 | /* | |
4864 | * We do not migrate tasks that are: | |
d3198084 | 4865 | * 1) throttled_lb_pair, or |
1e3c88bd | 4866 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
4867 | * 3) running (obviously), or |
4868 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 4869 | */ |
d3198084 JK |
4870 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
4871 | return 0; | |
4872 | ||
ddcdf6e7 | 4873 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 4874 | int cpu; |
88b8dac0 | 4875 | |
41acab88 | 4876 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 4877 | |
6263322c PZ |
4878 | env->flags |= LBF_SOME_PINNED; |
4879 | ||
88b8dac0 SV |
4880 | /* |
4881 | * Remember if this task can be migrated to any other cpu in | |
4882 | * our sched_group. We may want to revisit it if we couldn't | |
4883 | * meet load balance goals by pulling other tasks on src_cpu. | |
4884 | * | |
4885 | * Also avoid computing new_dst_cpu if we have already computed | |
4886 | * one in current iteration. | |
4887 | */ | |
6263322c | 4888 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
4889 | return 0; |
4890 | ||
e02e60c1 JK |
4891 | /* Prevent to re-select dst_cpu via env's cpus */ |
4892 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
4893 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 4894 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
4895 | env->new_dst_cpu = cpu; |
4896 | break; | |
4897 | } | |
88b8dac0 | 4898 | } |
e02e60c1 | 4899 | |
1e3c88bd PZ |
4900 | return 0; |
4901 | } | |
88b8dac0 SV |
4902 | |
4903 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 4904 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 4905 | |
ddcdf6e7 | 4906 | if (task_running(env->src_rq, p)) { |
41acab88 | 4907 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
4908 | return 0; |
4909 | } | |
4910 | ||
4911 | /* | |
4912 | * Aggressive migration if: | |
3a7053b3 MG |
4913 | * 1) destination numa is preferred |
4914 | * 2) task is cache cold, or | |
4915 | * 3) too many balance attempts have failed. | |
1e3c88bd | 4916 | */ |
78becc27 | 4917 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
7a0f3083 MG |
4918 | if (!tsk_cache_hot) |
4919 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
4920 | |
4921 | if (migrate_improves_locality(p, env)) { | |
4922 | #ifdef CONFIG_SCHEDSTATS | |
4923 | if (tsk_cache_hot) { | |
4924 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
4925 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
4926 | } | |
4927 | #endif | |
4928 | return 1; | |
4929 | } | |
4930 | ||
1e3c88bd | 4931 | if (!tsk_cache_hot || |
8e45cb54 | 4932 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 4933 | |
1e3c88bd | 4934 | if (tsk_cache_hot) { |
8e45cb54 | 4935 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 4936 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 4937 | } |
4e2dcb73 | 4938 | |
1e3c88bd PZ |
4939 | return 1; |
4940 | } | |
4941 | ||
4e2dcb73 ZH |
4942 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
4943 | return 0; | |
1e3c88bd PZ |
4944 | } |
4945 | ||
897c395f PZ |
4946 | /* |
4947 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
4948 | * part of active balancing operations within "domain". | |
4949 | * Returns 1 if successful and 0 otherwise. | |
4950 | * | |
4951 | * Called with both runqueues locked. | |
4952 | */ | |
8e45cb54 | 4953 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
4954 | { |
4955 | struct task_struct *p, *n; | |
897c395f | 4956 | |
367456c7 | 4957 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
4958 | if (!can_migrate_task(p, env)) |
4959 | continue; | |
897c395f | 4960 | |
367456c7 PZ |
4961 | move_task(p, env); |
4962 | /* | |
4963 | * Right now, this is only the second place move_task() | |
4964 | * is called, so we can safely collect move_task() | |
4965 | * stats here rather than inside move_task(). | |
4966 | */ | |
4967 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4968 | return 1; | |
897c395f | 4969 | } |
897c395f PZ |
4970 | return 0; |
4971 | } | |
4972 | ||
eb95308e PZ |
4973 | static const unsigned int sched_nr_migrate_break = 32; |
4974 | ||
5d6523eb | 4975 | /* |
bd939f45 | 4976 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4977 | * this_rq, as part of a balancing operation within domain "sd". |
4978 | * Returns 1 if successful and 0 otherwise. | |
4979 | * | |
4980 | * Called with both runqueues locked. | |
4981 | */ | |
4982 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4983 | { |
5d6523eb PZ |
4984 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4985 | struct task_struct *p; | |
367456c7 PZ |
4986 | unsigned long load; |
4987 | int pulled = 0; | |
1e3c88bd | 4988 | |
bd939f45 | 4989 | if (env->imbalance <= 0) |
5d6523eb | 4990 | return 0; |
1e3c88bd | 4991 | |
5d6523eb PZ |
4992 | while (!list_empty(tasks)) { |
4993 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 4994 | |
367456c7 PZ |
4995 | env->loop++; |
4996 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 4997 | if (env->loop > env->loop_max) |
367456c7 | 4998 | break; |
5d6523eb PZ |
4999 | |
5000 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 5001 | if (env->loop > env->loop_break) { |
eb95308e | 5002 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 5003 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 5004 | break; |
a195f004 | 5005 | } |
1e3c88bd | 5006 | |
d3198084 | 5007 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
5008 | goto next; |
5009 | ||
5010 | load = task_h_load(p); | |
5d6523eb | 5011 | |
eb95308e | 5012 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
5013 | goto next; |
5014 | ||
bd939f45 | 5015 | if ((load / 2) > env->imbalance) |
367456c7 | 5016 | goto next; |
1e3c88bd | 5017 | |
ddcdf6e7 | 5018 | move_task(p, env); |
ee00e66f | 5019 | pulled++; |
bd939f45 | 5020 | env->imbalance -= load; |
1e3c88bd PZ |
5021 | |
5022 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
5023 | /* |
5024 | * NEWIDLE balancing is a source of latency, so preemptible | |
5025 | * kernels will stop after the first task is pulled to minimize | |
5026 | * the critical section. | |
5027 | */ | |
5d6523eb | 5028 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 5029 | break; |
1e3c88bd PZ |
5030 | #endif |
5031 | ||
ee00e66f PZ |
5032 | /* |
5033 | * We only want to steal up to the prescribed amount of | |
5034 | * weighted load. | |
5035 | */ | |
bd939f45 | 5036 | if (env->imbalance <= 0) |
ee00e66f | 5037 | break; |
367456c7 PZ |
5038 | |
5039 | continue; | |
5040 | next: | |
5d6523eb | 5041 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 5042 | } |
5d6523eb | 5043 | |
1e3c88bd | 5044 | /* |
ddcdf6e7 PZ |
5045 | * Right now, this is one of only two places move_task() is called, |
5046 | * so we can safely collect move_task() stats here rather than | |
5047 | * inside move_task(). | |
1e3c88bd | 5048 | */ |
8e45cb54 | 5049 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 5050 | |
5d6523eb | 5051 | return pulled; |
1e3c88bd PZ |
5052 | } |
5053 | ||
230059de | 5054 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
5055 | /* |
5056 | * update tg->load_weight by folding this cpu's load_avg | |
5057 | */ | |
48a16753 | 5058 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 5059 | { |
48a16753 PT |
5060 | struct sched_entity *se = tg->se[cpu]; |
5061 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 5062 | |
48a16753 PT |
5063 | /* throttled entities do not contribute to load */ |
5064 | if (throttled_hierarchy(cfs_rq)) | |
5065 | return; | |
9e3081ca | 5066 | |
aff3e498 | 5067 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 5068 | |
82958366 PT |
5069 | if (se) { |
5070 | update_entity_load_avg(se, 1); | |
5071 | /* | |
5072 | * We pivot on our runnable average having decayed to zero for | |
5073 | * list removal. This generally implies that all our children | |
5074 | * have also been removed (modulo rounding error or bandwidth | |
5075 | * control); however, such cases are rare and we can fix these | |
5076 | * at enqueue. | |
5077 | * | |
5078 | * TODO: fix up out-of-order children on enqueue. | |
5079 | */ | |
5080 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
5081 | list_del_leaf_cfs_rq(cfs_rq); | |
5082 | } else { | |
48a16753 | 5083 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
5084 | update_rq_runnable_avg(rq, rq->nr_running); |
5085 | } | |
9e3081ca PZ |
5086 | } |
5087 | ||
48a16753 | 5088 | static void update_blocked_averages(int cpu) |
9e3081ca | 5089 | { |
9e3081ca | 5090 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
5091 | struct cfs_rq *cfs_rq; |
5092 | unsigned long flags; | |
9e3081ca | 5093 | |
48a16753 PT |
5094 | raw_spin_lock_irqsave(&rq->lock, flags); |
5095 | update_rq_clock(rq); | |
9763b67f PZ |
5096 | /* |
5097 | * Iterates the task_group tree in a bottom up fashion, see | |
5098 | * list_add_leaf_cfs_rq() for details. | |
5099 | */ | |
64660c86 | 5100 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
5101 | /* |
5102 | * Note: We may want to consider periodically releasing | |
5103 | * rq->lock about these updates so that creating many task | |
5104 | * groups does not result in continually extending hold time. | |
5105 | */ | |
5106 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 5107 | } |
48a16753 PT |
5108 | |
5109 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
5110 | } |
5111 | ||
9763b67f | 5112 | /* |
68520796 | 5113 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
5114 | * This needs to be done in a top-down fashion because the load of a child |
5115 | * group is a fraction of its parents load. | |
5116 | */ | |
68520796 | 5117 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 5118 | { |
68520796 VD |
5119 | struct rq *rq = rq_of(cfs_rq); |
5120 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 5121 | unsigned long now = jiffies; |
68520796 | 5122 | unsigned long load; |
a35b6466 | 5123 | |
68520796 | 5124 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
5125 | return; |
5126 | ||
68520796 VD |
5127 | cfs_rq->h_load_next = NULL; |
5128 | for_each_sched_entity(se) { | |
5129 | cfs_rq = cfs_rq_of(se); | |
5130 | cfs_rq->h_load_next = se; | |
5131 | if (cfs_rq->last_h_load_update == now) | |
5132 | break; | |
5133 | } | |
a35b6466 | 5134 | |
68520796 | 5135 | if (!se) { |
7e3115ef | 5136 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
5137 | cfs_rq->last_h_load_update = now; |
5138 | } | |
5139 | ||
5140 | while ((se = cfs_rq->h_load_next) != NULL) { | |
5141 | load = cfs_rq->h_load; | |
5142 | load = div64_ul(load * se->avg.load_avg_contrib, | |
5143 | cfs_rq->runnable_load_avg + 1); | |
5144 | cfs_rq = group_cfs_rq(se); | |
5145 | cfs_rq->h_load = load; | |
5146 | cfs_rq->last_h_load_update = now; | |
5147 | } | |
9763b67f PZ |
5148 | } |
5149 | ||
367456c7 | 5150 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 5151 | { |
367456c7 | 5152 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 5153 | |
68520796 | 5154 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
5155 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
5156 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
5157 | } |
5158 | #else | |
48a16753 | 5159 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
5160 | { |
5161 | } | |
5162 | ||
367456c7 | 5163 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 5164 | { |
a003a25b | 5165 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 5166 | } |
230059de | 5167 | #endif |
1e3c88bd | 5168 | |
1e3c88bd | 5169 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
5170 | /* |
5171 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
5172 | */ | |
5173 | struct sg_lb_stats { | |
5174 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
5175 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 5176 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 5177 | unsigned long load_per_task; |
3ae11c90 | 5178 | unsigned long group_power; |
147c5fc2 PZ |
5179 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
5180 | unsigned int group_capacity; | |
5181 | unsigned int idle_cpus; | |
5182 | unsigned int group_weight; | |
1e3c88bd | 5183 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 5184 | int group_has_capacity; /* Is there extra capacity in the group? */ |
0ec8aa00 PZ |
5185 | #ifdef CONFIG_NUMA_BALANCING |
5186 | unsigned int nr_numa_running; | |
5187 | unsigned int nr_preferred_running; | |
5188 | #endif | |
1e3c88bd PZ |
5189 | }; |
5190 | ||
56cf515b JK |
5191 | /* |
5192 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
5193 | * during load balancing. | |
5194 | */ | |
5195 | struct sd_lb_stats { | |
5196 | struct sched_group *busiest; /* Busiest group in this sd */ | |
5197 | struct sched_group *local; /* Local group in this sd */ | |
5198 | unsigned long total_load; /* Total load of all groups in sd */ | |
5199 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
5200 | unsigned long avg_load; /* Average load across all groups in sd */ | |
5201 | ||
56cf515b | 5202 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 5203 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
5204 | }; |
5205 | ||
147c5fc2 PZ |
5206 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
5207 | { | |
5208 | /* | |
5209 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
5210 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
5211 | * We must however clear busiest_stat::avg_load because | |
5212 | * update_sd_pick_busiest() reads this before assignment. | |
5213 | */ | |
5214 | *sds = (struct sd_lb_stats){ | |
5215 | .busiest = NULL, | |
5216 | .local = NULL, | |
5217 | .total_load = 0UL, | |
5218 | .total_pwr = 0UL, | |
5219 | .busiest_stat = { | |
5220 | .avg_load = 0UL, | |
5221 | }, | |
5222 | }; | |
5223 | } | |
5224 | ||
1e3c88bd PZ |
5225 | /** |
5226 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
5227 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 5228 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
5229 | * |
5230 | * Return: The load index. | |
1e3c88bd PZ |
5231 | */ |
5232 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
5233 | enum cpu_idle_type idle) | |
5234 | { | |
5235 | int load_idx; | |
5236 | ||
5237 | switch (idle) { | |
5238 | case CPU_NOT_IDLE: | |
5239 | load_idx = sd->busy_idx; | |
5240 | break; | |
5241 | ||
5242 | case CPU_NEWLY_IDLE: | |
5243 | load_idx = sd->newidle_idx; | |
5244 | break; | |
5245 | default: | |
5246 | load_idx = sd->idle_idx; | |
5247 | break; | |
5248 | } | |
5249 | ||
5250 | return load_idx; | |
5251 | } | |
5252 | ||
15f803c9 | 5253 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5254 | { |
1399fa78 | 5255 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
5256 | } |
5257 | ||
5258 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
5259 | { | |
5260 | return default_scale_freq_power(sd, cpu); | |
5261 | } | |
5262 | ||
15f803c9 | 5263 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5264 | { |
669c55e9 | 5265 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
5266 | unsigned long smt_gain = sd->smt_gain; |
5267 | ||
5268 | smt_gain /= weight; | |
5269 | ||
5270 | return smt_gain; | |
5271 | } | |
5272 | ||
5273 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
5274 | { | |
5275 | return default_scale_smt_power(sd, cpu); | |
5276 | } | |
5277 | ||
15f803c9 | 5278 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
5279 | { |
5280 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5281 | u64 total, available, age_stamp, avg; |
1e3c88bd | 5282 | |
b654f7de PZ |
5283 | /* |
5284 | * Since we're reading these variables without serialization make sure | |
5285 | * we read them once before doing sanity checks on them. | |
5286 | */ | |
5287 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5288 | avg = ACCESS_ONCE(rq->rt_avg); | |
5289 | ||
78becc27 | 5290 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 5291 | |
b654f7de | 5292 | if (unlikely(total < avg)) { |
aa483808 VP |
5293 | /* Ensures that power won't end up being negative */ |
5294 | available = 0; | |
5295 | } else { | |
b654f7de | 5296 | available = total - avg; |
aa483808 | 5297 | } |
1e3c88bd | 5298 | |
1399fa78 NR |
5299 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
5300 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 5301 | |
1399fa78 | 5302 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5303 | |
5304 | return div_u64(available, total); | |
5305 | } | |
5306 | ||
5307 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
5308 | { | |
669c55e9 | 5309 | unsigned long weight = sd->span_weight; |
1399fa78 | 5310 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
5311 | struct sched_group *sdg = sd->groups; |
5312 | ||
1e3c88bd PZ |
5313 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
5314 | if (sched_feat(ARCH_POWER)) | |
5315 | power *= arch_scale_smt_power(sd, cpu); | |
5316 | else | |
5317 | power *= default_scale_smt_power(sd, cpu); | |
5318 | ||
1399fa78 | 5319 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5320 | } |
5321 | ||
9c3f75cb | 5322 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
5323 | |
5324 | if (sched_feat(ARCH_POWER)) | |
5325 | power *= arch_scale_freq_power(sd, cpu); | |
5326 | else | |
5327 | power *= default_scale_freq_power(sd, cpu); | |
5328 | ||
1399fa78 | 5329 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 5330 | |
1e3c88bd | 5331 | power *= scale_rt_power(cpu); |
1399fa78 | 5332 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5333 | |
5334 | if (!power) | |
5335 | power = 1; | |
5336 | ||
e51fd5e2 | 5337 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 5338 | sdg->sgp->power = power; |
1e3c88bd PZ |
5339 | } |
5340 | ||
029632fb | 5341 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
5342 | { |
5343 | struct sched_domain *child = sd->child; | |
5344 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 5345 | unsigned long power, power_orig; |
4ec4412e VG |
5346 | unsigned long interval; |
5347 | ||
5348 | interval = msecs_to_jiffies(sd->balance_interval); | |
5349 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
5350 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
5351 | |
5352 | if (!child) { | |
5353 | update_cpu_power(sd, cpu); | |
5354 | return; | |
5355 | } | |
5356 | ||
863bffc8 | 5357 | power_orig = power = 0; |
1e3c88bd | 5358 | |
74a5ce20 PZ |
5359 | if (child->flags & SD_OVERLAP) { |
5360 | /* | |
5361 | * SD_OVERLAP domains cannot assume that child groups | |
5362 | * span the current group. | |
5363 | */ | |
5364 | ||
863bffc8 | 5365 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
9abf24d4 SD |
5366 | struct sched_group_power *sgp; |
5367 | struct rq *rq = cpu_rq(cpu); | |
863bffc8 | 5368 | |
9abf24d4 SD |
5369 | /* |
5370 | * build_sched_domains() -> init_sched_groups_power() | |
5371 | * gets here before we've attached the domains to the | |
5372 | * runqueues. | |
5373 | * | |
5374 | * Use power_of(), which is set irrespective of domains | |
5375 | * in update_cpu_power(). | |
5376 | * | |
5377 | * This avoids power/power_orig from being 0 and | |
5378 | * causing divide-by-zero issues on boot. | |
5379 | * | |
5380 | * Runtime updates will correct power_orig. | |
5381 | */ | |
5382 | if (unlikely(!rq->sd)) { | |
5383 | power_orig += power_of(cpu); | |
5384 | power += power_of(cpu); | |
5385 | continue; | |
5386 | } | |
863bffc8 | 5387 | |
9abf24d4 SD |
5388 | sgp = rq->sd->groups->sgp; |
5389 | power_orig += sgp->power_orig; | |
5390 | power += sgp->power; | |
863bffc8 | 5391 | } |
74a5ce20 PZ |
5392 | } else { |
5393 | /* | |
5394 | * !SD_OVERLAP domains can assume that child groups | |
5395 | * span the current group. | |
5396 | */ | |
5397 | ||
5398 | group = child->groups; | |
5399 | do { | |
863bffc8 | 5400 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
5401 | power += group->sgp->power; |
5402 | group = group->next; | |
5403 | } while (group != child->groups); | |
5404 | } | |
1e3c88bd | 5405 | |
863bffc8 PZ |
5406 | sdg->sgp->power_orig = power_orig; |
5407 | sdg->sgp->power = power; | |
1e3c88bd PZ |
5408 | } |
5409 | ||
9d5efe05 SV |
5410 | /* |
5411 | * Try and fix up capacity for tiny siblings, this is needed when | |
5412 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
5413 | * which on its own isn't powerful enough. | |
5414 | * | |
5415 | * See update_sd_pick_busiest() and check_asym_packing(). | |
5416 | */ | |
5417 | static inline int | |
5418 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
5419 | { | |
5420 | /* | |
1399fa78 | 5421 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 5422 | */ |
a6c75f2f | 5423 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
5424 | return 0; |
5425 | ||
5426 | /* | |
5427 | * If ~90% of the cpu_power is still there, we're good. | |
5428 | */ | |
9c3f75cb | 5429 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
5430 | return 1; |
5431 | ||
5432 | return 0; | |
5433 | } | |
5434 | ||
30ce5dab PZ |
5435 | /* |
5436 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
5437 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
5438 | * | |
5439 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
5440 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
5441 | * Something like: | |
5442 | * | |
5443 | * { 0 1 2 3 } { 4 5 6 7 } | |
5444 | * * * * * | |
5445 | * | |
5446 | * If we were to balance group-wise we'd place two tasks in the first group and | |
5447 | * two tasks in the second group. Clearly this is undesired as it will overload | |
5448 | * cpu 3 and leave one of the cpus in the second group unused. | |
5449 | * | |
5450 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
5451 | * by noticing the lower domain failed to reach balance and had difficulty |
5452 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
5453 | * |
5454 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 5455 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 5456 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
5457 | * to create an effective group imbalance. |
5458 | * | |
5459 | * This is a somewhat tricky proposition since the next run might not find the | |
5460 | * group imbalance and decide the groups need to be balanced again. A most | |
5461 | * subtle and fragile situation. | |
5462 | */ | |
5463 | ||
6263322c | 5464 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 5465 | { |
6263322c | 5466 | return group->sgp->imbalance; |
30ce5dab PZ |
5467 | } |
5468 | ||
b37d9316 PZ |
5469 | /* |
5470 | * Compute the group capacity. | |
5471 | * | |
c61037e9 PZ |
5472 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
5473 | * first dividing out the smt factor and computing the actual number of cores | |
5474 | * and limit power unit capacity with that. | |
b37d9316 PZ |
5475 | */ |
5476 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
5477 | { | |
c61037e9 PZ |
5478 | unsigned int capacity, smt, cpus; |
5479 | unsigned int power, power_orig; | |
5480 | ||
5481 | power = group->sgp->power; | |
5482 | power_orig = group->sgp->power_orig; | |
5483 | cpus = group->group_weight; | |
b37d9316 | 5484 | |
c61037e9 PZ |
5485 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
5486 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
5487 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 5488 | |
c61037e9 | 5489 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
5490 | if (!capacity) |
5491 | capacity = fix_small_capacity(env->sd, group); | |
5492 | ||
5493 | return capacity; | |
5494 | } | |
5495 | ||
1e3c88bd PZ |
5496 | /** |
5497 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 5498 | * @env: The load balancing environment. |
1e3c88bd | 5499 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 5500 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 5501 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
5502 | * @sgs: variable to hold the statistics for this group. |
5503 | */ | |
bd939f45 PZ |
5504 | static inline void update_sg_lb_stats(struct lb_env *env, |
5505 | struct sched_group *group, int load_idx, | |
23f0d209 | 5506 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 5507 | { |
30ce5dab | 5508 | unsigned long load; |
bd939f45 | 5509 | int i; |
1e3c88bd | 5510 | |
b72ff13c PZ |
5511 | memset(sgs, 0, sizeof(*sgs)); |
5512 | ||
b9403130 | 5513 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
5514 | struct rq *rq = cpu_rq(i); |
5515 | ||
1e3c88bd | 5516 | /* Bias balancing toward cpus of our domain */ |
6263322c | 5517 | if (local_group) |
04f733b4 | 5518 | load = target_load(i, load_idx); |
6263322c | 5519 | else |
1e3c88bd | 5520 | load = source_load(i, load_idx); |
1e3c88bd PZ |
5521 | |
5522 | sgs->group_load += load; | |
380c9077 | 5523 | sgs->sum_nr_running += rq->nr_running; |
0ec8aa00 PZ |
5524 | #ifdef CONFIG_NUMA_BALANCING |
5525 | sgs->nr_numa_running += rq->nr_numa_running; | |
5526 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
5527 | #endif | |
1e3c88bd | 5528 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
5529 | if (idle_cpu(i)) |
5530 | sgs->idle_cpus++; | |
1e3c88bd PZ |
5531 | } |
5532 | ||
1e3c88bd | 5533 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
5534 | sgs->group_power = group->sgp->power; |
5535 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 5536 | |
dd5feea1 | 5537 | if (sgs->sum_nr_running) |
38d0f770 | 5538 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5539 | |
aae6d3dd | 5540 | sgs->group_weight = group->group_weight; |
fab47622 | 5541 | |
b37d9316 PZ |
5542 | sgs->group_imb = sg_imbalanced(group); |
5543 | sgs->group_capacity = sg_capacity(env, group); | |
5544 | ||
fab47622 NR |
5545 | if (sgs->group_capacity > sgs->sum_nr_running) |
5546 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5547 | } |
5548 | ||
532cb4c4 MN |
5549 | /** |
5550 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5551 | * @env: The load balancing environment. |
532cb4c4 MN |
5552 | * @sds: sched_domain statistics |
5553 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5554 | * @sgs: sched_group statistics |
532cb4c4 MN |
5555 | * |
5556 | * Determine if @sg is a busier group than the previously selected | |
5557 | * busiest group. | |
e69f6186 YB |
5558 | * |
5559 | * Return: %true if @sg is a busier group than the previously selected | |
5560 | * busiest group. %false otherwise. | |
532cb4c4 | 5561 | */ |
bd939f45 | 5562 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5563 | struct sd_lb_stats *sds, |
5564 | struct sched_group *sg, | |
bd939f45 | 5565 | struct sg_lb_stats *sgs) |
532cb4c4 | 5566 | { |
56cf515b | 5567 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5568 | return false; |
5569 | ||
5570 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5571 | return true; | |
5572 | ||
5573 | if (sgs->group_imb) | |
5574 | return true; | |
5575 | ||
5576 | /* | |
5577 | * ASYM_PACKING needs to move all the work to the lowest | |
5578 | * numbered CPUs in the group, therefore mark all groups | |
5579 | * higher than ourself as busy. | |
5580 | */ | |
bd939f45 PZ |
5581 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5582 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5583 | if (!sds->busiest) |
5584 | return true; | |
5585 | ||
5586 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5587 | return true; | |
5588 | } | |
5589 | ||
5590 | return false; | |
5591 | } | |
5592 | ||
0ec8aa00 PZ |
5593 | #ifdef CONFIG_NUMA_BALANCING |
5594 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5595 | { | |
5596 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
5597 | return regular; | |
5598 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
5599 | return remote; | |
5600 | return all; | |
5601 | } | |
5602 | ||
5603 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5604 | { | |
5605 | if (rq->nr_running > rq->nr_numa_running) | |
5606 | return regular; | |
5607 | if (rq->nr_running > rq->nr_preferred_running) | |
5608 | return remote; | |
5609 | return all; | |
5610 | } | |
5611 | #else | |
5612 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5613 | { | |
5614 | return all; | |
5615 | } | |
5616 | ||
5617 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5618 | { | |
5619 | return regular; | |
5620 | } | |
5621 | #endif /* CONFIG_NUMA_BALANCING */ | |
5622 | ||
1e3c88bd | 5623 | /** |
461819ac | 5624 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5625 | * @env: The load balancing environment. |
1e3c88bd PZ |
5626 | * @sds: variable to hold the statistics for this sched_domain. |
5627 | */ | |
0ec8aa00 | 5628 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5629 | { |
bd939f45 PZ |
5630 | struct sched_domain *child = env->sd->child; |
5631 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5632 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5633 | int load_idx, prefer_sibling = 0; |
5634 | ||
5635 | if (child && child->flags & SD_PREFER_SIBLING) | |
5636 | prefer_sibling = 1; | |
5637 | ||
bd939f45 | 5638 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5639 | |
5640 | do { | |
56cf515b | 5641 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5642 | int local_group; |
5643 | ||
bd939f45 | 5644 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5645 | if (local_group) { |
5646 | sds->local = sg; | |
5647 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5648 | |
5649 | if (env->idle != CPU_NEWLY_IDLE || | |
5650 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5651 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5652 | } |
1e3c88bd | 5653 | |
56cf515b | 5654 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5655 | |
b72ff13c PZ |
5656 | if (local_group) |
5657 | goto next_group; | |
5658 | ||
1e3c88bd PZ |
5659 | /* |
5660 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 5661 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
5662 | * and move all the excess tasks away. We lower the capacity |
5663 | * of a group only if the local group has the capacity to fit | |
5664 | * these excess tasks, i.e. nr_running < group_capacity. The | |
5665 | * extra check prevents the case where you always pull from the | |
5666 | * heaviest group when it is already under-utilized (possible | |
5667 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 5668 | */ |
b72ff13c PZ |
5669 | if (prefer_sibling && sds->local && |
5670 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 5671 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 5672 | |
b72ff13c | 5673 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 5674 | sds->busiest = sg; |
56cf515b | 5675 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
5676 | } |
5677 | ||
b72ff13c PZ |
5678 | next_group: |
5679 | /* Now, start updating sd_lb_stats */ | |
5680 | sds->total_load += sgs->group_load; | |
5681 | sds->total_pwr += sgs->group_power; | |
5682 | ||
532cb4c4 | 5683 | sg = sg->next; |
bd939f45 | 5684 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
5685 | |
5686 | if (env->sd->flags & SD_NUMA) | |
5687 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
532cb4c4 MN |
5688 | } |
5689 | ||
532cb4c4 MN |
5690 | /** |
5691 | * check_asym_packing - Check to see if the group is packed into the | |
5692 | * sched doman. | |
5693 | * | |
5694 | * This is primarily intended to used at the sibling level. Some | |
5695 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
5696 | * case of POWER7, it can move to lower SMT modes only when higher | |
5697 | * threads are idle. When in lower SMT modes, the threads will | |
5698 | * perform better since they share less core resources. Hence when we | |
5699 | * have idle threads, we want them to be the higher ones. | |
5700 | * | |
5701 | * This packing function is run on idle threads. It checks to see if | |
5702 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
5703 | * CPU number than the packing function is being run on. Here we are | |
5704 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
5705 | * number. | |
5706 | * | |
e69f6186 | 5707 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
5708 | * this CPU. The amount of the imbalance is returned in *imbalance. |
5709 | * | |
cd96891d | 5710 | * @env: The load balancing environment. |
532cb4c4 | 5711 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 5712 | */ |
bd939f45 | 5713 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
5714 | { |
5715 | int busiest_cpu; | |
5716 | ||
bd939f45 | 5717 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
5718 | return 0; |
5719 | ||
5720 | if (!sds->busiest) | |
5721 | return 0; | |
5722 | ||
5723 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 5724 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
5725 | return 0; |
5726 | ||
bd939f45 | 5727 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
5728 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
5729 | SCHED_POWER_SCALE); | |
bd939f45 | 5730 | |
532cb4c4 | 5731 | return 1; |
1e3c88bd PZ |
5732 | } |
5733 | ||
5734 | /** | |
5735 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
5736 | * amongst the groups of a sched_domain, during | |
5737 | * load balancing. | |
cd96891d | 5738 | * @env: The load balancing environment. |
1e3c88bd | 5739 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5740 | */ |
bd939f45 PZ |
5741 | static inline |
5742 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
5743 | { |
5744 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
5745 | unsigned int imbn = 2; | |
dd5feea1 | 5746 | unsigned long scaled_busy_load_per_task; |
56cf515b | 5747 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 5748 | |
56cf515b JK |
5749 | local = &sds->local_stat; |
5750 | busiest = &sds->busiest_stat; | |
1e3c88bd | 5751 | |
56cf515b JK |
5752 | if (!local->sum_nr_running) |
5753 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
5754 | else if (busiest->load_per_task > local->load_per_task) | |
5755 | imbn = 1; | |
dd5feea1 | 5756 | |
56cf515b JK |
5757 | scaled_busy_load_per_task = |
5758 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5759 | busiest->group_power; |
56cf515b | 5760 | |
3029ede3 VD |
5761 | if (busiest->avg_load + scaled_busy_load_per_task >= |
5762 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 5763 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5764 | return; |
5765 | } | |
5766 | ||
5767 | /* | |
5768 | * OK, we don't have enough imbalance to justify moving tasks, | |
5769 | * however we may be able to increase total CPU power used by | |
5770 | * moving them. | |
5771 | */ | |
5772 | ||
3ae11c90 | 5773 | pwr_now += busiest->group_power * |
56cf515b | 5774 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 5775 | pwr_now += local->group_power * |
56cf515b | 5776 | min(local->load_per_task, local->avg_load); |
1399fa78 | 5777 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5778 | |
5779 | /* Amount of load we'd subtract */ | |
56cf515b | 5780 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 5781 | busiest->group_power; |
56cf515b | 5782 | if (busiest->avg_load > tmp) { |
3ae11c90 | 5783 | pwr_move += busiest->group_power * |
56cf515b JK |
5784 | min(busiest->load_per_task, |
5785 | busiest->avg_load - tmp); | |
5786 | } | |
1e3c88bd PZ |
5787 | |
5788 | /* Amount of load we'd add */ | |
3ae11c90 | 5789 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 5790 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
5791 | tmp = (busiest->avg_load * busiest->group_power) / |
5792 | local->group_power; | |
56cf515b JK |
5793 | } else { |
5794 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5795 | local->group_power; |
56cf515b | 5796 | } |
3ae11c90 PZ |
5797 | pwr_move += local->group_power * |
5798 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 5799 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5800 | |
5801 | /* Move if we gain throughput */ | |
5802 | if (pwr_move > pwr_now) | |
56cf515b | 5803 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5804 | } |
5805 | ||
5806 | /** | |
5807 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
5808 | * groups of a given sched_domain during load balance. | |
bd939f45 | 5809 | * @env: load balance environment |
1e3c88bd | 5810 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5811 | */ |
bd939f45 | 5812 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5813 | { |
dd5feea1 | 5814 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
5815 | struct sg_lb_stats *local, *busiest; |
5816 | ||
5817 | local = &sds->local_stat; | |
56cf515b | 5818 | busiest = &sds->busiest_stat; |
dd5feea1 | 5819 | |
56cf515b | 5820 | if (busiest->group_imb) { |
30ce5dab PZ |
5821 | /* |
5822 | * In the group_imb case we cannot rely on group-wide averages | |
5823 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
5824 | */ | |
56cf515b JK |
5825 | busiest->load_per_task = |
5826 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
5827 | } |
5828 | ||
1e3c88bd PZ |
5829 | /* |
5830 | * In the presence of smp nice balancing, certain scenarios can have | |
5831 | * max load less than avg load(as we skip the groups at or below | |
5832 | * its cpu_power, while calculating max_load..) | |
5833 | */ | |
b1885550 VD |
5834 | if (busiest->avg_load <= sds->avg_load || |
5835 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
5836 | env->imbalance = 0; |
5837 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
5838 | } |
5839 | ||
56cf515b | 5840 | if (!busiest->group_imb) { |
dd5feea1 SS |
5841 | /* |
5842 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
5843 | * Except of course for the group_imb case, since then we might |
5844 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 5845 | */ |
56cf515b JK |
5846 | load_above_capacity = |
5847 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 5848 | |
1399fa78 | 5849 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 5850 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
5851 | } |
5852 | ||
5853 | /* | |
5854 | * We're trying to get all the cpus to the average_load, so we don't | |
5855 | * want to push ourselves above the average load, nor do we wish to | |
5856 | * reduce the max loaded cpu below the average load. At the same time, | |
5857 | * we also don't want to reduce the group load below the group capacity | |
5858 | * (so that we can implement power-savings policies etc). Thus we look | |
5859 | * for the minimum possible imbalance. | |
dd5feea1 | 5860 | */ |
30ce5dab | 5861 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
5862 | |
5863 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 5864 | env->imbalance = min( |
3ae11c90 PZ |
5865 | max_pull * busiest->group_power, |
5866 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 5867 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
5868 | |
5869 | /* | |
5870 | * if *imbalance is less than the average load per runnable task | |
25985edc | 5871 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
5872 | * a think about bumping its value to force at least one task to be |
5873 | * moved | |
5874 | */ | |
56cf515b | 5875 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 5876 | return fix_small_imbalance(env, sds); |
1e3c88bd | 5877 | } |
fab47622 | 5878 | |
1e3c88bd PZ |
5879 | /******* find_busiest_group() helpers end here *********************/ |
5880 | ||
5881 | /** | |
5882 | * find_busiest_group - Returns the busiest group within the sched_domain | |
5883 | * if there is an imbalance. If there isn't an imbalance, and | |
5884 | * the user has opted for power-savings, it returns a group whose | |
5885 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
5886 | * such a group exists. | |
5887 | * | |
5888 | * Also calculates the amount of weighted load which should be moved | |
5889 | * to restore balance. | |
5890 | * | |
cd96891d | 5891 | * @env: The load balancing environment. |
1e3c88bd | 5892 | * |
e69f6186 | 5893 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
5894 | * - If no imbalance and user has opted for power-savings balance, |
5895 | * return the least loaded group whose CPUs can be | |
5896 | * put to idle by rebalancing its tasks onto our group. | |
5897 | */ | |
56cf515b | 5898 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 5899 | { |
56cf515b | 5900 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
5901 | struct sd_lb_stats sds; |
5902 | ||
147c5fc2 | 5903 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
5904 | |
5905 | /* | |
5906 | * Compute the various statistics relavent for load balancing at | |
5907 | * this level. | |
5908 | */ | |
23f0d209 | 5909 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
5910 | local = &sds.local_stat; |
5911 | busiest = &sds.busiest_stat; | |
1e3c88bd | 5912 | |
bd939f45 PZ |
5913 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
5914 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
5915 | return sds.busiest; |
5916 | ||
cc57aa8f | 5917 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 5918 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
5919 | goto out_balanced; |
5920 | ||
1399fa78 | 5921 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 5922 | |
866ab43e PZ |
5923 | /* |
5924 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 5925 | * work because they assume all things are equal, which typically |
866ab43e PZ |
5926 | * isn't true due to cpus_allowed constraints and the like. |
5927 | */ | |
56cf515b | 5928 | if (busiest->group_imb) |
866ab43e PZ |
5929 | goto force_balance; |
5930 | ||
cc57aa8f | 5931 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
5932 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
5933 | !busiest->group_has_capacity) | |
fab47622 NR |
5934 | goto force_balance; |
5935 | ||
cc57aa8f PZ |
5936 | /* |
5937 | * If the local group is more busy than the selected busiest group | |
5938 | * don't try and pull any tasks. | |
5939 | */ | |
56cf515b | 5940 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
5941 | goto out_balanced; |
5942 | ||
cc57aa8f PZ |
5943 | /* |
5944 | * Don't pull any tasks if this group is already above the domain | |
5945 | * average load. | |
5946 | */ | |
56cf515b | 5947 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
5948 | goto out_balanced; |
5949 | ||
bd939f45 | 5950 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
5951 | /* |
5952 | * This cpu is idle. If the busiest group load doesn't | |
5953 | * have more tasks than the number of available cpu's and | |
5954 | * there is no imbalance between this and busiest group | |
5955 | * wrt to idle cpu's, it is balanced. | |
5956 | */ | |
56cf515b JK |
5957 | if ((local->idle_cpus < busiest->idle_cpus) && |
5958 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 5959 | goto out_balanced; |
c186fafe PZ |
5960 | } else { |
5961 | /* | |
5962 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
5963 | * imbalance_pct to be conservative. | |
5964 | */ | |
56cf515b JK |
5965 | if (100 * busiest->avg_load <= |
5966 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 5967 | goto out_balanced; |
aae6d3dd | 5968 | } |
1e3c88bd | 5969 | |
fab47622 | 5970 | force_balance: |
1e3c88bd | 5971 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 5972 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
5973 | return sds.busiest; |
5974 | ||
5975 | out_balanced: | |
bd939f45 | 5976 | env->imbalance = 0; |
1e3c88bd PZ |
5977 | return NULL; |
5978 | } | |
5979 | ||
5980 | /* | |
5981 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
5982 | */ | |
bd939f45 | 5983 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 5984 | struct sched_group *group) |
1e3c88bd PZ |
5985 | { |
5986 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 5987 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
5988 | int i; |
5989 | ||
6906a408 | 5990 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
0ec8aa00 PZ |
5991 | unsigned long power, capacity, wl; |
5992 | enum fbq_type rt; | |
5993 | ||
5994 | rq = cpu_rq(i); | |
5995 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 5996 | |
0ec8aa00 PZ |
5997 | /* |
5998 | * We classify groups/runqueues into three groups: | |
5999 | * - regular: there are !numa tasks | |
6000 | * - remote: there are numa tasks that run on the 'wrong' node | |
6001 | * - all: there is no distinction | |
6002 | * | |
6003 | * In order to avoid migrating ideally placed numa tasks, | |
6004 | * ignore those when there's better options. | |
6005 | * | |
6006 | * If we ignore the actual busiest queue to migrate another | |
6007 | * task, the next balance pass can still reduce the busiest | |
6008 | * queue by moving tasks around inside the node. | |
6009 | * | |
6010 | * If we cannot move enough load due to this classification | |
6011 | * the next pass will adjust the group classification and | |
6012 | * allow migration of more tasks. | |
6013 | * | |
6014 | * Both cases only affect the total convergence complexity. | |
6015 | */ | |
6016 | if (rt > env->fbq_type) | |
6017 | continue; | |
6018 | ||
6019 | power = power_of(i); | |
6020 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); | |
9d5efe05 | 6021 | if (!capacity) |
bd939f45 | 6022 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 6023 | |
6e40f5bb | 6024 | wl = weighted_cpuload(i); |
1e3c88bd | 6025 | |
6e40f5bb TG |
6026 | /* |
6027 | * When comparing with imbalance, use weighted_cpuload() | |
6028 | * which is not scaled with the cpu power. | |
6029 | */ | |
bd939f45 | 6030 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
6031 | continue; |
6032 | ||
6e40f5bb TG |
6033 | /* |
6034 | * For the load comparisons with the other cpu's, consider | |
6035 | * the weighted_cpuload() scaled with the cpu power, so that | |
6036 | * the load can be moved away from the cpu that is potentially | |
6037 | * running at a lower capacity. | |
95a79b80 JK |
6038 | * |
6039 | * Thus we're looking for max(wl_i / power_i), crosswise | |
6040 | * multiplication to rid ourselves of the division works out | |
6041 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
6042 | * previous maximum. | |
6e40f5bb | 6043 | */ |
95a79b80 JK |
6044 | if (wl * busiest_power > busiest_load * power) { |
6045 | busiest_load = wl; | |
6046 | busiest_power = power; | |
1e3c88bd PZ |
6047 | busiest = rq; |
6048 | } | |
6049 | } | |
6050 | ||
6051 | return busiest; | |
6052 | } | |
6053 | ||
6054 | /* | |
6055 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
6056 | * so long as it is large enough. | |
6057 | */ | |
6058 | #define MAX_PINNED_INTERVAL 512 | |
6059 | ||
6060 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 6061 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 6062 | |
bd939f45 | 6063 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 6064 | { |
bd939f45 PZ |
6065 | struct sched_domain *sd = env->sd; |
6066 | ||
6067 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
6068 | |
6069 | /* | |
6070 | * ASYM_PACKING needs to force migrate tasks from busy but | |
6071 | * higher numbered CPUs in order to pack all tasks in the | |
6072 | * lowest numbered CPUs. | |
6073 | */ | |
bd939f45 | 6074 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 6075 | return 1; |
1af3ed3d PZ |
6076 | } |
6077 | ||
6078 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
6079 | } | |
6080 | ||
969c7921 TH |
6081 | static int active_load_balance_cpu_stop(void *data); |
6082 | ||
23f0d209 JK |
6083 | static int should_we_balance(struct lb_env *env) |
6084 | { | |
6085 | struct sched_group *sg = env->sd->groups; | |
6086 | struct cpumask *sg_cpus, *sg_mask; | |
6087 | int cpu, balance_cpu = -1; | |
6088 | ||
6089 | /* | |
6090 | * In the newly idle case, we will allow all the cpu's | |
6091 | * to do the newly idle load balance. | |
6092 | */ | |
6093 | if (env->idle == CPU_NEWLY_IDLE) | |
6094 | return 1; | |
6095 | ||
6096 | sg_cpus = sched_group_cpus(sg); | |
6097 | sg_mask = sched_group_mask(sg); | |
6098 | /* Try to find first idle cpu */ | |
6099 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
6100 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
6101 | continue; | |
6102 | ||
6103 | balance_cpu = cpu; | |
6104 | break; | |
6105 | } | |
6106 | ||
6107 | if (balance_cpu == -1) | |
6108 | balance_cpu = group_balance_cpu(sg); | |
6109 | ||
6110 | /* | |
6111 | * First idle cpu or the first cpu(busiest) in this sched group | |
6112 | * is eligible for doing load balancing at this and above domains. | |
6113 | */ | |
b0cff9d8 | 6114 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
6115 | } |
6116 | ||
1e3c88bd PZ |
6117 | /* |
6118 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
6119 | * tasks if there is an imbalance. | |
6120 | */ | |
6121 | static int load_balance(int this_cpu, struct rq *this_rq, | |
6122 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 6123 | int *continue_balancing) |
1e3c88bd | 6124 | { |
88b8dac0 | 6125 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 6126 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 6127 | struct sched_group *group; |
1e3c88bd PZ |
6128 | struct rq *busiest; |
6129 | unsigned long flags; | |
e6252c3e | 6130 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 6131 | |
8e45cb54 PZ |
6132 | struct lb_env env = { |
6133 | .sd = sd, | |
ddcdf6e7 PZ |
6134 | .dst_cpu = this_cpu, |
6135 | .dst_rq = this_rq, | |
88b8dac0 | 6136 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 6137 | .idle = idle, |
eb95308e | 6138 | .loop_break = sched_nr_migrate_break, |
b9403130 | 6139 | .cpus = cpus, |
0ec8aa00 | 6140 | .fbq_type = all, |
8e45cb54 PZ |
6141 | }; |
6142 | ||
cfc03118 JK |
6143 | /* |
6144 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
6145 | * other cpus in our group | |
6146 | */ | |
e02e60c1 | 6147 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 6148 | env.dst_grpmask = NULL; |
cfc03118 | 6149 | |
1e3c88bd PZ |
6150 | cpumask_copy(cpus, cpu_active_mask); |
6151 | ||
1e3c88bd PZ |
6152 | schedstat_inc(sd, lb_count[idle]); |
6153 | ||
6154 | redo: | |
23f0d209 JK |
6155 | if (!should_we_balance(&env)) { |
6156 | *continue_balancing = 0; | |
1e3c88bd | 6157 | goto out_balanced; |
23f0d209 | 6158 | } |
1e3c88bd | 6159 | |
23f0d209 | 6160 | group = find_busiest_group(&env); |
1e3c88bd PZ |
6161 | if (!group) { |
6162 | schedstat_inc(sd, lb_nobusyg[idle]); | |
6163 | goto out_balanced; | |
6164 | } | |
6165 | ||
b9403130 | 6166 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
6167 | if (!busiest) { |
6168 | schedstat_inc(sd, lb_nobusyq[idle]); | |
6169 | goto out_balanced; | |
6170 | } | |
6171 | ||
78feefc5 | 6172 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 6173 | |
bd939f45 | 6174 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
6175 | |
6176 | ld_moved = 0; | |
6177 | if (busiest->nr_running > 1) { | |
6178 | /* | |
6179 | * Attempt to move tasks. If find_busiest_group has found | |
6180 | * an imbalance but busiest->nr_running <= 1, the group is | |
6181 | * still unbalanced. ld_moved simply stays zero, so it is | |
6182 | * correctly treated as an imbalance. | |
6183 | */ | |
8e45cb54 | 6184 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
6185 | env.src_cpu = busiest->cpu; |
6186 | env.src_rq = busiest; | |
6187 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 6188 | |
5d6523eb | 6189 | more_balance: |
1e3c88bd | 6190 | local_irq_save(flags); |
78feefc5 | 6191 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
6192 | |
6193 | /* | |
6194 | * cur_ld_moved - load moved in current iteration | |
6195 | * ld_moved - cumulative load moved across iterations | |
6196 | */ | |
6197 | cur_ld_moved = move_tasks(&env); | |
6198 | ld_moved += cur_ld_moved; | |
78feefc5 | 6199 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
6200 | local_irq_restore(flags); |
6201 | ||
6202 | /* | |
6203 | * some other cpu did the load balance for us. | |
6204 | */ | |
88b8dac0 SV |
6205 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
6206 | resched_cpu(env.dst_cpu); | |
6207 | ||
f1cd0858 JK |
6208 | if (env.flags & LBF_NEED_BREAK) { |
6209 | env.flags &= ~LBF_NEED_BREAK; | |
6210 | goto more_balance; | |
6211 | } | |
6212 | ||
88b8dac0 SV |
6213 | /* |
6214 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
6215 | * us and move them to an alternate dst_cpu in our sched_group | |
6216 | * where they can run. The upper limit on how many times we | |
6217 | * iterate on same src_cpu is dependent on number of cpus in our | |
6218 | * sched_group. | |
6219 | * | |
6220 | * This changes load balance semantics a bit on who can move | |
6221 | * load to a given_cpu. In addition to the given_cpu itself | |
6222 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
6223 | * nohz-idle), we now have balance_cpu in a position to move | |
6224 | * load to given_cpu. In rare situations, this may cause | |
6225 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
6226 | * _independently_ and at _same_ time to move some load to | |
6227 | * given_cpu) causing exceess load to be moved to given_cpu. | |
6228 | * This however should not happen so much in practice and | |
6229 | * moreover subsequent load balance cycles should correct the | |
6230 | * excess load moved. | |
6231 | */ | |
6263322c | 6232 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 6233 | |
7aff2e3a VD |
6234 | /* Prevent to re-select dst_cpu via env's cpus */ |
6235 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
6236 | ||
78feefc5 | 6237 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 6238 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 6239 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
6240 | env.loop = 0; |
6241 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 6242 | |
88b8dac0 SV |
6243 | /* |
6244 | * Go back to "more_balance" rather than "redo" since we | |
6245 | * need to continue with same src_cpu. | |
6246 | */ | |
6247 | goto more_balance; | |
6248 | } | |
1e3c88bd | 6249 | |
6263322c PZ |
6250 | /* |
6251 | * We failed to reach balance because of affinity. | |
6252 | */ | |
6253 | if (sd_parent) { | |
6254 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
6255 | ||
6256 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
6257 | *group_imbalance = 1; | |
6258 | } else if (*group_imbalance) | |
6259 | *group_imbalance = 0; | |
6260 | } | |
6261 | ||
1e3c88bd | 6262 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 6263 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 6264 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
6265 | if (!cpumask_empty(cpus)) { |
6266 | env.loop = 0; | |
6267 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 6268 | goto redo; |
bbf18b19 | 6269 | } |
1e3c88bd PZ |
6270 | goto out_balanced; |
6271 | } | |
6272 | } | |
6273 | ||
6274 | if (!ld_moved) { | |
6275 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
6276 | /* |
6277 | * Increment the failure counter only on periodic balance. | |
6278 | * We do not want newidle balance, which can be very | |
6279 | * frequent, pollute the failure counter causing | |
6280 | * excessive cache_hot migrations and active balances. | |
6281 | */ | |
6282 | if (idle != CPU_NEWLY_IDLE) | |
6283 | sd->nr_balance_failed++; | |
1e3c88bd | 6284 | |
bd939f45 | 6285 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
6286 | raw_spin_lock_irqsave(&busiest->lock, flags); |
6287 | ||
969c7921 TH |
6288 | /* don't kick the active_load_balance_cpu_stop, |
6289 | * if the curr task on busiest cpu can't be | |
6290 | * moved to this_cpu | |
1e3c88bd PZ |
6291 | */ |
6292 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 6293 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
6294 | raw_spin_unlock_irqrestore(&busiest->lock, |
6295 | flags); | |
8e45cb54 | 6296 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
6297 | goto out_one_pinned; |
6298 | } | |
6299 | ||
969c7921 TH |
6300 | /* |
6301 | * ->active_balance synchronizes accesses to | |
6302 | * ->active_balance_work. Once set, it's cleared | |
6303 | * only after active load balance is finished. | |
6304 | */ | |
1e3c88bd PZ |
6305 | if (!busiest->active_balance) { |
6306 | busiest->active_balance = 1; | |
6307 | busiest->push_cpu = this_cpu; | |
6308 | active_balance = 1; | |
6309 | } | |
6310 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 6311 | |
bd939f45 | 6312 | if (active_balance) { |
969c7921 TH |
6313 | stop_one_cpu_nowait(cpu_of(busiest), |
6314 | active_load_balance_cpu_stop, busiest, | |
6315 | &busiest->active_balance_work); | |
bd939f45 | 6316 | } |
1e3c88bd PZ |
6317 | |
6318 | /* | |
6319 | * We've kicked active balancing, reset the failure | |
6320 | * counter. | |
6321 | */ | |
6322 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
6323 | } | |
6324 | } else | |
6325 | sd->nr_balance_failed = 0; | |
6326 | ||
6327 | if (likely(!active_balance)) { | |
6328 | /* We were unbalanced, so reset the balancing interval */ | |
6329 | sd->balance_interval = sd->min_interval; | |
6330 | } else { | |
6331 | /* | |
6332 | * If we've begun active balancing, start to back off. This | |
6333 | * case may not be covered by the all_pinned logic if there | |
6334 | * is only 1 task on the busy runqueue (because we don't call | |
6335 | * move_tasks). | |
6336 | */ | |
6337 | if (sd->balance_interval < sd->max_interval) | |
6338 | sd->balance_interval *= 2; | |
6339 | } | |
6340 | ||
1e3c88bd PZ |
6341 | goto out; |
6342 | ||
6343 | out_balanced: | |
6344 | schedstat_inc(sd, lb_balanced[idle]); | |
6345 | ||
6346 | sd->nr_balance_failed = 0; | |
6347 | ||
6348 | out_one_pinned: | |
6349 | /* tune up the balancing interval */ | |
8e45cb54 | 6350 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 6351 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
6352 | (sd->balance_interval < sd->max_interval)) |
6353 | sd->balance_interval *= 2; | |
6354 | ||
46e49b38 | 6355 | ld_moved = 0; |
1e3c88bd | 6356 | out: |
1e3c88bd PZ |
6357 | return ld_moved; |
6358 | } | |
6359 | ||
1e3c88bd PZ |
6360 | /* |
6361 | * idle_balance is called by schedule() if this_cpu is about to become | |
6362 | * idle. Attempts to pull tasks from other CPUs. | |
6363 | */ | |
029632fb | 6364 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
6365 | { |
6366 | struct sched_domain *sd; | |
6367 | int pulled_task = 0; | |
6368 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 6369 | u64 curr_cost = 0; |
1e3c88bd | 6370 | |
78becc27 | 6371 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
6372 | |
6373 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
6374 | return; | |
6375 | ||
f492e12e PZ |
6376 | /* |
6377 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
6378 | */ | |
6379 | raw_spin_unlock(&this_rq->lock); | |
6380 | ||
48a16753 | 6381 | update_blocked_averages(this_cpu); |
dce840a0 | 6382 | rcu_read_lock(); |
1e3c88bd PZ |
6383 | for_each_domain(this_cpu, sd) { |
6384 | unsigned long interval; | |
23f0d209 | 6385 | int continue_balancing = 1; |
9bd721c5 | 6386 | u64 t0, domain_cost; |
1e3c88bd PZ |
6387 | |
6388 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
6389 | continue; | |
6390 | ||
9bd721c5 JL |
6391 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
6392 | break; | |
6393 | ||
f492e12e | 6394 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
6395 | t0 = sched_clock_cpu(this_cpu); |
6396 | ||
1e3c88bd | 6397 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 6398 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
6399 | sd, CPU_NEWLY_IDLE, |
6400 | &continue_balancing); | |
9bd721c5 JL |
6401 | |
6402 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
6403 | if (domain_cost > sd->max_newidle_lb_cost) | |
6404 | sd->max_newidle_lb_cost = domain_cost; | |
6405 | ||
6406 | curr_cost += domain_cost; | |
f492e12e | 6407 | } |
1e3c88bd PZ |
6408 | |
6409 | interval = msecs_to_jiffies(sd->balance_interval); | |
6410 | if (time_after(next_balance, sd->last_balance + interval)) | |
6411 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
6412 | if (pulled_task) { |
6413 | this_rq->idle_stamp = 0; | |
1e3c88bd | 6414 | break; |
d5ad140b | 6415 | } |
1e3c88bd | 6416 | } |
dce840a0 | 6417 | rcu_read_unlock(); |
f492e12e PZ |
6418 | |
6419 | raw_spin_lock(&this_rq->lock); | |
6420 | ||
1e3c88bd PZ |
6421 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
6422 | /* | |
6423 | * We are going idle. next_balance may be set based on | |
6424 | * a busy processor. So reset next_balance. | |
6425 | */ | |
6426 | this_rq->next_balance = next_balance; | |
6427 | } | |
9bd721c5 JL |
6428 | |
6429 | if (curr_cost > this_rq->max_idle_balance_cost) | |
6430 | this_rq->max_idle_balance_cost = curr_cost; | |
1e3c88bd PZ |
6431 | } |
6432 | ||
6433 | /* | |
969c7921 TH |
6434 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
6435 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
6436 | * least 1 task to be running on each physical CPU where possible, and | |
6437 | * avoids physical / logical imbalances. | |
1e3c88bd | 6438 | */ |
969c7921 | 6439 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 6440 | { |
969c7921 TH |
6441 | struct rq *busiest_rq = data; |
6442 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 6443 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 6444 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 6445 | struct sched_domain *sd; |
969c7921 TH |
6446 | |
6447 | raw_spin_lock_irq(&busiest_rq->lock); | |
6448 | ||
6449 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
6450 | if (unlikely(busiest_cpu != smp_processor_id() || | |
6451 | !busiest_rq->active_balance)) | |
6452 | goto out_unlock; | |
1e3c88bd PZ |
6453 | |
6454 | /* Is there any task to move? */ | |
6455 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 6456 | goto out_unlock; |
1e3c88bd PZ |
6457 | |
6458 | /* | |
6459 | * This condition is "impossible", if it occurs | |
6460 | * we need to fix it. Originally reported by | |
6461 | * Bjorn Helgaas on a 128-cpu setup. | |
6462 | */ | |
6463 | BUG_ON(busiest_rq == target_rq); | |
6464 | ||
6465 | /* move a task from busiest_rq to target_rq */ | |
6466 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
6467 | |
6468 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 6469 | rcu_read_lock(); |
1e3c88bd PZ |
6470 | for_each_domain(target_cpu, sd) { |
6471 | if ((sd->flags & SD_LOAD_BALANCE) && | |
6472 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
6473 | break; | |
6474 | } | |
6475 | ||
6476 | if (likely(sd)) { | |
8e45cb54 PZ |
6477 | struct lb_env env = { |
6478 | .sd = sd, | |
ddcdf6e7 PZ |
6479 | .dst_cpu = target_cpu, |
6480 | .dst_rq = target_rq, | |
6481 | .src_cpu = busiest_rq->cpu, | |
6482 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
6483 | .idle = CPU_IDLE, |
6484 | }; | |
6485 | ||
1e3c88bd PZ |
6486 | schedstat_inc(sd, alb_count); |
6487 | ||
8e45cb54 | 6488 | if (move_one_task(&env)) |
1e3c88bd PZ |
6489 | schedstat_inc(sd, alb_pushed); |
6490 | else | |
6491 | schedstat_inc(sd, alb_failed); | |
6492 | } | |
dce840a0 | 6493 | rcu_read_unlock(); |
1e3c88bd | 6494 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
6495 | out_unlock: |
6496 | busiest_rq->active_balance = 0; | |
6497 | raw_spin_unlock_irq(&busiest_rq->lock); | |
6498 | return 0; | |
1e3c88bd PZ |
6499 | } |
6500 | ||
3451d024 | 6501 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
6502 | /* |
6503 | * idle load balancing details | |
83cd4fe2 VP |
6504 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
6505 | * needed, they will kick the idle load balancer, which then does idle | |
6506 | * load balancing for all the idle CPUs. | |
6507 | */ | |
1e3c88bd | 6508 | static struct { |
83cd4fe2 | 6509 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 6510 | atomic_t nr_cpus; |
83cd4fe2 VP |
6511 | unsigned long next_balance; /* in jiffy units */ |
6512 | } nohz ____cacheline_aligned; | |
1e3c88bd | 6513 | |
3dd0337d | 6514 | static inline int find_new_ilb(void) |
1e3c88bd | 6515 | { |
0b005cf5 | 6516 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 6517 | |
786d6dc7 SS |
6518 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
6519 | return ilb; | |
6520 | ||
6521 | return nr_cpu_ids; | |
1e3c88bd | 6522 | } |
1e3c88bd | 6523 | |
83cd4fe2 VP |
6524 | /* |
6525 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
6526 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
6527 | * CPU (if there is one). | |
6528 | */ | |
0aeeeeba | 6529 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
6530 | { |
6531 | int ilb_cpu; | |
6532 | ||
6533 | nohz.next_balance++; | |
6534 | ||
3dd0337d | 6535 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 6536 | |
0b005cf5 SS |
6537 | if (ilb_cpu >= nr_cpu_ids) |
6538 | return; | |
83cd4fe2 | 6539 | |
cd490c5b | 6540 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
6541 | return; |
6542 | /* | |
6543 | * Use smp_send_reschedule() instead of resched_cpu(). | |
6544 | * This way we generate a sched IPI on the target cpu which | |
6545 | * is idle. And the softirq performing nohz idle load balance | |
6546 | * will be run before returning from the IPI. | |
6547 | */ | |
6548 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
6549 | return; |
6550 | } | |
6551 | ||
c1cc017c | 6552 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
6553 | { |
6554 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
6555 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
6556 | atomic_dec(&nohz.nr_cpus); | |
6557 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
6558 | } | |
6559 | } | |
6560 | ||
69e1e811 SS |
6561 | static inline void set_cpu_sd_state_busy(void) |
6562 | { | |
6563 | struct sched_domain *sd; | |
37dc6b50 | 6564 | int cpu = smp_processor_id(); |
69e1e811 | 6565 | |
69e1e811 | 6566 | rcu_read_lock(); |
37dc6b50 | 6567 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
6568 | |
6569 | if (!sd || !sd->nohz_idle) | |
6570 | goto unlock; | |
6571 | sd->nohz_idle = 0; | |
6572 | ||
37dc6b50 | 6573 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6574 | unlock: |
69e1e811 SS |
6575 | rcu_read_unlock(); |
6576 | } | |
6577 | ||
6578 | void set_cpu_sd_state_idle(void) | |
6579 | { | |
6580 | struct sched_domain *sd; | |
37dc6b50 | 6581 | int cpu = smp_processor_id(); |
69e1e811 | 6582 | |
69e1e811 | 6583 | rcu_read_lock(); |
37dc6b50 | 6584 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
6585 | |
6586 | if (!sd || sd->nohz_idle) | |
6587 | goto unlock; | |
6588 | sd->nohz_idle = 1; | |
6589 | ||
37dc6b50 | 6590 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6591 | unlock: |
69e1e811 SS |
6592 | rcu_read_unlock(); |
6593 | } | |
6594 | ||
1e3c88bd | 6595 | /* |
c1cc017c | 6596 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 6597 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 6598 | */ |
c1cc017c | 6599 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 6600 | { |
71325960 SS |
6601 | /* |
6602 | * If this cpu is going down, then nothing needs to be done. | |
6603 | */ | |
6604 | if (!cpu_active(cpu)) | |
6605 | return; | |
6606 | ||
c1cc017c AS |
6607 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
6608 | return; | |
1e3c88bd | 6609 | |
c1cc017c AS |
6610 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
6611 | atomic_inc(&nohz.nr_cpus); | |
6612 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 6613 | } |
71325960 | 6614 | |
0db0628d | 6615 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
6616 | unsigned long action, void *hcpu) |
6617 | { | |
6618 | switch (action & ~CPU_TASKS_FROZEN) { | |
6619 | case CPU_DYING: | |
c1cc017c | 6620 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
6621 | return NOTIFY_OK; |
6622 | default: | |
6623 | return NOTIFY_DONE; | |
6624 | } | |
6625 | } | |
1e3c88bd PZ |
6626 | #endif |
6627 | ||
6628 | static DEFINE_SPINLOCK(balancing); | |
6629 | ||
49c022e6 PZ |
6630 | /* |
6631 | * Scale the max load_balance interval with the number of CPUs in the system. | |
6632 | * This trades load-balance latency on larger machines for less cross talk. | |
6633 | */ | |
029632fb | 6634 | void update_max_interval(void) |
49c022e6 PZ |
6635 | { |
6636 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
6637 | } | |
6638 | ||
1e3c88bd PZ |
6639 | /* |
6640 | * It checks each scheduling domain to see if it is due to be balanced, | |
6641 | * and initiates a balancing operation if so. | |
6642 | * | |
b9b0853a | 6643 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 6644 | */ |
f7ed0a89 | 6645 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 6646 | { |
23f0d209 | 6647 | int continue_balancing = 1; |
f7ed0a89 | 6648 | int cpu = rq->cpu; |
1e3c88bd | 6649 | unsigned long interval; |
04f733b4 | 6650 | struct sched_domain *sd; |
1e3c88bd PZ |
6651 | /* Earliest time when we have to do rebalance again */ |
6652 | unsigned long next_balance = jiffies + 60*HZ; | |
6653 | int update_next_balance = 0; | |
f48627e6 JL |
6654 | int need_serialize, need_decay = 0; |
6655 | u64 max_cost = 0; | |
1e3c88bd | 6656 | |
48a16753 | 6657 | update_blocked_averages(cpu); |
2069dd75 | 6658 | |
dce840a0 | 6659 | rcu_read_lock(); |
1e3c88bd | 6660 | for_each_domain(cpu, sd) { |
f48627e6 JL |
6661 | /* |
6662 | * Decay the newidle max times here because this is a regular | |
6663 | * visit to all the domains. Decay ~1% per second. | |
6664 | */ | |
6665 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
6666 | sd->max_newidle_lb_cost = | |
6667 | (sd->max_newidle_lb_cost * 253) / 256; | |
6668 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
6669 | need_decay = 1; | |
6670 | } | |
6671 | max_cost += sd->max_newidle_lb_cost; | |
6672 | ||
1e3c88bd PZ |
6673 | if (!(sd->flags & SD_LOAD_BALANCE)) |
6674 | continue; | |
6675 | ||
f48627e6 JL |
6676 | /* |
6677 | * Stop the load balance at this level. There is another | |
6678 | * CPU in our sched group which is doing load balancing more | |
6679 | * actively. | |
6680 | */ | |
6681 | if (!continue_balancing) { | |
6682 | if (need_decay) | |
6683 | continue; | |
6684 | break; | |
6685 | } | |
6686 | ||
1e3c88bd PZ |
6687 | interval = sd->balance_interval; |
6688 | if (idle != CPU_IDLE) | |
6689 | interval *= sd->busy_factor; | |
6690 | ||
6691 | /* scale ms to jiffies */ | |
6692 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 6693 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
6694 | |
6695 | need_serialize = sd->flags & SD_SERIALIZE; | |
6696 | ||
6697 | if (need_serialize) { | |
6698 | if (!spin_trylock(&balancing)) | |
6699 | goto out; | |
6700 | } | |
6701 | ||
6702 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 6703 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 6704 | /* |
6263322c | 6705 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
6706 | * env->dst_cpu, so we can't know our idle |
6707 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 6708 | */ |
de5eb2dd | 6709 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
6710 | } |
6711 | sd->last_balance = jiffies; | |
6712 | } | |
6713 | if (need_serialize) | |
6714 | spin_unlock(&balancing); | |
6715 | out: | |
6716 | if (time_after(next_balance, sd->last_balance + interval)) { | |
6717 | next_balance = sd->last_balance + interval; | |
6718 | update_next_balance = 1; | |
6719 | } | |
f48627e6 JL |
6720 | } |
6721 | if (need_decay) { | |
1e3c88bd | 6722 | /* |
f48627e6 JL |
6723 | * Ensure the rq-wide value also decays but keep it at a |
6724 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 6725 | */ |
f48627e6 JL |
6726 | rq->max_idle_balance_cost = |
6727 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 6728 | } |
dce840a0 | 6729 | rcu_read_unlock(); |
1e3c88bd PZ |
6730 | |
6731 | /* | |
6732 | * next_balance will be updated only when there is a need. | |
6733 | * When the cpu is attached to null domain for ex, it will not be | |
6734 | * updated. | |
6735 | */ | |
6736 | if (likely(update_next_balance)) | |
6737 | rq->next_balance = next_balance; | |
6738 | } | |
6739 | ||
3451d024 | 6740 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 6741 | /* |
3451d024 | 6742 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
6743 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
6744 | */ | |
208cb16b | 6745 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 6746 | { |
208cb16b | 6747 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
6748 | struct rq *rq; |
6749 | int balance_cpu; | |
6750 | ||
1c792db7 SS |
6751 | if (idle != CPU_IDLE || |
6752 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
6753 | goto end; | |
83cd4fe2 VP |
6754 | |
6755 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 6756 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
6757 | continue; |
6758 | ||
6759 | /* | |
6760 | * If this cpu gets work to do, stop the load balancing | |
6761 | * work being done for other cpus. Next load | |
6762 | * balancing owner will pick it up. | |
6763 | */ | |
1c792db7 | 6764 | if (need_resched()) |
83cd4fe2 | 6765 | break; |
83cd4fe2 | 6766 | |
5ed4f1d9 VG |
6767 | rq = cpu_rq(balance_cpu); |
6768 | ||
6769 | raw_spin_lock_irq(&rq->lock); | |
6770 | update_rq_clock(rq); | |
6771 | update_idle_cpu_load(rq); | |
6772 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 | 6773 | |
f7ed0a89 | 6774 | rebalance_domains(rq, CPU_IDLE); |
83cd4fe2 | 6775 | |
83cd4fe2 VP |
6776 | if (time_after(this_rq->next_balance, rq->next_balance)) |
6777 | this_rq->next_balance = rq->next_balance; | |
6778 | } | |
6779 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
6780 | end: |
6781 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
6782 | } |
6783 | ||
6784 | /* | |
0b005cf5 SS |
6785 | * Current heuristic for kicking the idle load balancer in the presence |
6786 | * of an idle cpu is the system. | |
6787 | * - This rq has more than one task. | |
6788 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
6789 | * busy cpu's exceeding the group's power. | |
6790 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
6791 | * domain span are idle. | |
83cd4fe2 | 6792 | */ |
4a725627 | 6793 | static inline int nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
6794 | { |
6795 | unsigned long now = jiffies; | |
0b005cf5 | 6796 | struct sched_domain *sd; |
37dc6b50 | 6797 | struct sched_group_power *sgp; |
4a725627 | 6798 | int nr_busy, cpu = rq->cpu; |
83cd4fe2 | 6799 | |
4a725627 | 6800 | if (unlikely(rq->idle_balance)) |
83cd4fe2 VP |
6801 | return 0; |
6802 | ||
1c792db7 SS |
6803 | /* |
6804 | * We may be recently in ticked or tickless idle mode. At the first | |
6805 | * busy tick after returning from idle, we will update the busy stats. | |
6806 | */ | |
69e1e811 | 6807 | set_cpu_sd_state_busy(); |
c1cc017c | 6808 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
6809 | |
6810 | /* | |
6811 | * None are in tickless mode and hence no need for NOHZ idle load | |
6812 | * balancing. | |
6813 | */ | |
6814 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
6815 | return 0; | |
1c792db7 SS |
6816 | |
6817 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
6818 | return 0; |
6819 | ||
0b005cf5 SS |
6820 | if (rq->nr_running >= 2) |
6821 | goto need_kick; | |
83cd4fe2 | 6822 | |
067491b7 | 6823 | rcu_read_lock(); |
37dc6b50 | 6824 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
83cd4fe2 | 6825 | |
37dc6b50 PM |
6826 | if (sd) { |
6827 | sgp = sd->groups->sgp; | |
6828 | nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
0b005cf5 | 6829 | |
37dc6b50 | 6830 | if (nr_busy > 1) |
067491b7 | 6831 | goto need_kick_unlock; |
83cd4fe2 | 6832 | } |
37dc6b50 PM |
6833 | |
6834 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); | |
6835 | ||
6836 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, | |
6837 | sched_domain_span(sd)) < cpu)) | |
6838 | goto need_kick_unlock; | |
6839 | ||
067491b7 | 6840 | rcu_read_unlock(); |
83cd4fe2 | 6841 | return 0; |
067491b7 PZ |
6842 | |
6843 | need_kick_unlock: | |
6844 | rcu_read_unlock(); | |
0b005cf5 SS |
6845 | need_kick: |
6846 | return 1; | |
83cd4fe2 VP |
6847 | } |
6848 | #else | |
208cb16b | 6849 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
6850 | #endif |
6851 | ||
6852 | /* | |
6853 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
6854 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
6855 | */ | |
1e3c88bd PZ |
6856 | static void run_rebalance_domains(struct softirq_action *h) |
6857 | { | |
208cb16b | 6858 | struct rq *this_rq = this_rq(); |
6eb57e0d | 6859 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
6860 | CPU_IDLE : CPU_NOT_IDLE; |
6861 | ||
f7ed0a89 | 6862 | rebalance_domains(this_rq, idle); |
1e3c88bd | 6863 | |
1e3c88bd | 6864 | /* |
83cd4fe2 | 6865 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
6866 | * balancing on behalf of the other idle cpus whose ticks are |
6867 | * stopped. | |
6868 | */ | |
208cb16b | 6869 | nohz_idle_balance(this_rq, idle); |
1e3c88bd PZ |
6870 | } |
6871 | ||
63f609b1 | 6872 | static inline int on_null_domain(struct rq *rq) |
1e3c88bd | 6873 | { |
63f609b1 | 6874 | return !rcu_dereference_sched(rq->sd); |
1e3c88bd PZ |
6875 | } |
6876 | ||
6877 | /* | |
6878 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 6879 | */ |
7caff66f | 6880 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 6881 | { |
1e3c88bd | 6882 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
6883 | if (unlikely(on_null_domain(rq))) |
6884 | return; | |
6885 | ||
6886 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 6887 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 6888 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 6889 | if (nohz_kick_needed(rq)) |
0aeeeeba | 6890 | nohz_balancer_kick(); |
83cd4fe2 | 6891 | #endif |
1e3c88bd PZ |
6892 | } |
6893 | ||
0bcdcf28 CE |
6894 | static void rq_online_fair(struct rq *rq) |
6895 | { | |
6896 | update_sysctl(); | |
6897 | } | |
6898 | ||
6899 | static void rq_offline_fair(struct rq *rq) | |
6900 | { | |
6901 | update_sysctl(); | |
a4c96ae3 PB |
6902 | |
6903 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
6904 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
6905 | } |
6906 | ||
55e12e5e | 6907 | #endif /* CONFIG_SMP */ |
e1d1484f | 6908 | |
bf0f6f24 IM |
6909 | /* |
6910 | * scheduler tick hitting a task of our scheduling class: | |
6911 | */ | |
8f4d37ec | 6912 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
6913 | { |
6914 | struct cfs_rq *cfs_rq; | |
6915 | struct sched_entity *se = &curr->se; | |
6916 | ||
6917 | for_each_sched_entity(se) { | |
6918 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 6919 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 6920 | } |
18bf2805 | 6921 | |
10e84b97 | 6922 | if (numabalancing_enabled) |
cbee9f88 | 6923 | task_tick_numa(rq, curr); |
3d59eebc | 6924 | |
18bf2805 | 6925 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
6926 | } |
6927 | ||
6928 | /* | |
cd29fe6f PZ |
6929 | * called on fork with the child task as argument from the parent's context |
6930 | * - child not yet on the tasklist | |
6931 | * - preemption disabled | |
bf0f6f24 | 6932 | */ |
cd29fe6f | 6933 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 6934 | { |
4fc420c9 DN |
6935 | struct cfs_rq *cfs_rq; |
6936 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 6937 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
6938 | struct rq *rq = this_rq(); |
6939 | unsigned long flags; | |
6940 | ||
05fa785c | 6941 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 6942 | |
861d034e PZ |
6943 | update_rq_clock(rq); |
6944 | ||
4fc420c9 DN |
6945 | cfs_rq = task_cfs_rq(current); |
6946 | curr = cfs_rq->curr; | |
6947 | ||
6c9a27f5 DN |
6948 | /* |
6949 | * Not only the cpu but also the task_group of the parent might have | |
6950 | * been changed after parent->se.parent,cfs_rq were copied to | |
6951 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
6952 | * of child point to valid ones. | |
6953 | */ | |
6954 | rcu_read_lock(); | |
6955 | __set_task_cpu(p, this_cpu); | |
6956 | rcu_read_unlock(); | |
bf0f6f24 | 6957 | |
7109c442 | 6958 | update_curr(cfs_rq); |
cd29fe6f | 6959 | |
b5d9d734 MG |
6960 | if (curr) |
6961 | se->vruntime = curr->vruntime; | |
aeb73b04 | 6962 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 6963 | |
cd29fe6f | 6964 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 6965 | /* |
edcb60a3 IM |
6966 | * Upon rescheduling, sched_class::put_prev_task() will place |
6967 | * 'current' within the tree based on its new key value. | |
6968 | */ | |
4d78e7b6 | 6969 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 6970 | resched_task(rq->curr); |
4d78e7b6 | 6971 | } |
bf0f6f24 | 6972 | |
88ec22d3 PZ |
6973 | se->vruntime -= cfs_rq->min_vruntime; |
6974 | ||
05fa785c | 6975 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
6976 | } |
6977 | ||
cb469845 SR |
6978 | /* |
6979 | * Priority of the task has changed. Check to see if we preempt | |
6980 | * the current task. | |
6981 | */ | |
da7a735e PZ |
6982 | static void |
6983 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 6984 | { |
da7a735e PZ |
6985 | if (!p->se.on_rq) |
6986 | return; | |
6987 | ||
cb469845 SR |
6988 | /* |
6989 | * Reschedule if we are currently running on this runqueue and | |
6990 | * our priority decreased, or if we are not currently running on | |
6991 | * this runqueue and our priority is higher than the current's | |
6992 | */ | |
da7a735e | 6993 | if (rq->curr == p) { |
cb469845 SR |
6994 | if (p->prio > oldprio) |
6995 | resched_task(rq->curr); | |
6996 | } else | |
15afe09b | 6997 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6998 | } |
6999 | ||
da7a735e PZ |
7000 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
7001 | { | |
7002 | struct sched_entity *se = &p->se; | |
7003 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7004 | ||
7005 | /* | |
7006 | * Ensure the task's vruntime is normalized, so that when its | |
7007 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
7008 | * do the right thing. | |
7009 | * | |
7010 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
7011 | * have normalized the vruntime, if it was !on_rq, then only when | |
7012 | * the task is sleeping will it still have non-normalized vruntime. | |
7013 | */ | |
7014 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
7015 | /* | |
7016 | * Fix up our vruntime so that the current sleep doesn't | |
7017 | * cause 'unlimited' sleep bonus. | |
7018 | */ | |
7019 | place_entity(cfs_rq, se, 0); | |
7020 | se->vruntime -= cfs_rq->min_vruntime; | |
7021 | } | |
9ee474f5 | 7022 | |
141965c7 | 7023 | #ifdef CONFIG_SMP |
9ee474f5 PT |
7024 | /* |
7025 | * Remove our load from contribution when we leave sched_fair | |
7026 | * and ensure we don't carry in an old decay_count if we | |
7027 | * switch back. | |
7028 | */ | |
87e3c8ae KT |
7029 | if (se->avg.decay_count) { |
7030 | __synchronize_entity_decay(se); | |
7031 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
7032 | } |
7033 | #endif | |
da7a735e PZ |
7034 | } |
7035 | ||
cb469845 SR |
7036 | /* |
7037 | * We switched to the sched_fair class. | |
7038 | */ | |
da7a735e | 7039 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 7040 | { |
da7a735e PZ |
7041 | if (!p->se.on_rq) |
7042 | return; | |
7043 | ||
cb469845 SR |
7044 | /* |
7045 | * We were most likely switched from sched_rt, so | |
7046 | * kick off the schedule if running, otherwise just see | |
7047 | * if we can still preempt the current task. | |
7048 | */ | |
da7a735e | 7049 | if (rq->curr == p) |
cb469845 SR |
7050 | resched_task(rq->curr); |
7051 | else | |
15afe09b | 7052 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7053 | } |
7054 | ||
83b699ed SV |
7055 | /* Account for a task changing its policy or group. |
7056 | * | |
7057 | * This routine is mostly called to set cfs_rq->curr field when a task | |
7058 | * migrates between groups/classes. | |
7059 | */ | |
7060 | static void set_curr_task_fair(struct rq *rq) | |
7061 | { | |
7062 | struct sched_entity *se = &rq->curr->se; | |
7063 | ||
ec12cb7f PT |
7064 | for_each_sched_entity(se) { |
7065 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7066 | ||
7067 | set_next_entity(cfs_rq, se); | |
7068 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
7069 | account_cfs_rq_runtime(cfs_rq, 0); | |
7070 | } | |
83b699ed SV |
7071 | } |
7072 | ||
029632fb PZ |
7073 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
7074 | { | |
7075 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
7076 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
7077 | #ifndef CONFIG_64BIT | |
7078 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
7079 | #endif | |
141965c7 | 7080 | #ifdef CONFIG_SMP |
9ee474f5 | 7081 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 7082 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 7083 | #endif |
029632fb PZ |
7084 | } |
7085 | ||
810b3817 | 7086 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7087 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 7088 | { |
aff3e498 | 7089 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
7090 | /* |
7091 | * If the task was not on the rq at the time of this cgroup movement | |
7092 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
7093 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
7094 | * bonus in place_entity()). | |
7095 | * | |
7096 | * If it was on the rq, we've just 'preempted' it, which does convert | |
7097 | * ->vruntime to a relative base. | |
7098 | * | |
7099 | * Make sure both cases convert their relative position when migrating | |
7100 | * to another cgroup's rq. This does somewhat interfere with the | |
7101 | * fair sleeper stuff for the first placement, but who cares. | |
7102 | */ | |
7ceff013 DN |
7103 | /* |
7104 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
7105 | * But there are some cases where it has already been normalized: | |
7106 | * | |
7107 | * - Moving a forked child which is waiting for being woken up by | |
7108 | * wake_up_new_task(). | |
62af3783 DN |
7109 | * - Moving a task which has been woken up by try_to_wake_up() and |
7110 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
7111 | * |
7112 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
7113 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
7114 | */ | |
62af3783 | 7115 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
7116 | on_rq = 1; |
7117 | ||
b2b5ce02 PZ |
7118 | if (!on_rq) |
7119 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
7120 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
7121 | if (!on_rq) { |
7122 | cfs_rq = cfs_rq_of(&p->se); | |
7123 | p->se.vruntime += cfs_rq->min_vruntime; | |
7124 | #ifdef CONFIG_SMP | |
7125 | /* | |
7126 | * migrate_task_rq_fair() will have removed our previous | |
7127 | * contribution, but we must synchronize for ongoing future | |
7128 | * decay. | |
7129 | */ | |
7130 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
7131 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
7132 | #endif | |
7133 | } | |
810b3817 | 7134 | } |
029632fb PZ |
7135 | |
7136 | void free_fair_sched_group(struct task_group *tg) | |
7137 | { | |
7138 | int i; | |
7139 | ||
7140 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7141 | ||
7142 | for_each_possible_cpu(i) { | |
7143 | if (tg->cfs_rq) | |
7144 | kfree(tg->cfs_rq[i]); | |
7145 | if (tg->se) | |
7146 | kfree(tg->se[i]); | |
7147 | } | |
7148 | ||
7149 | kfree(tg->cfs_rq); | |
7150 | kfree(tg->se); | |
7151 | } | |
7152 | ||
7153 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7154 | { | |
7155 | struct cfs_rq *cfs_rq; | |
7156 | struct sched_entity *se; | |
7157 | int i; | |
7158 | ||
7159 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
7160 | if (!tg->cfs_rq) | |
7161 | goto err; | |
7162 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
7163 | if (!tg->se) | |
7164 | goto err; | |
7165 | ||
7166 | tg->shares = NICE_0_LOAD; | |
7167 | ||
7168 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7169 | ||
7170 | for_each_possible_cpu(i) { | |
7171 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
7172 | GFP_KERNEL, cpu_to_node(i)); | |
7173 | if (!cfs_rq) | |
7174 | goto err; | |
7175 | ||
7176 | se = kzalloc_node(sizeof(struct sched_entity), | |
7177 | GFP_KERNEL, cpu_to_node(i)); | |
7178 | if (!se) | |
7179 | goto err_free_rq; | |
7180 | ||
7181 | init_cfs_rq(cfs_rq); | |
7182 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
7183 | } | |
7184 | ||
7185 | return 1; | |
7186 | ||
7187 | err_free_rq: | |
7188 | kfree(cfs_rq); | |
7189 | err: | |
7190 | return 0; | |
7191 | } | |
7192 | ||
7193 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
7194 | { | |
7195 | struct rq *rq = cpu_rq(cpu); | |
7196 | unsigned long flags; | |
7197 | ||
7198 | /* | |
7199 | * Only empty task groups can be destroyed; so we can speculatively | |
7200 | * check on_list without danger of it being re-added. | |
7201 | */ | |
7202 | if (!tg->cfs_rq[cpu]->on_list) | |
7203 | return; | |
7204 | ||
7205 | raw_spin_lock_irqsave(&rq->lock, flags); | |
7206 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
7207 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7208 | } | |
7209 | ||
7210 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
7211 | struct sched_entity *se, int cpu, | |
7212 | struct sched_entity *parent) | |
7213 | { | |
7214 | struct rq *rq = cpu_rq(cpu); | |
7215 | ||
7216 | cfs_rq->tg = tg; | |
7217 | cfs_rq->rq = rq; | |
029632fb PZ |
7218 | init_cfs_rq_runtime(cfs_rq); |
7219 | ||
7220 | tg->cfs_rq[cpu] = cfs_rq; | |
7221 | tg->se[cpu] = se; | |
7222 | ||
7223 | /* se could be NULL for root_task_group */ | |
7224 | if (!se) | |
7225 | return; | |
7226 | ||
7227 | if (!parent) | |
7228 | se->cfs_rq = &rq->cfs; | |
7229 | else | |
7230 | se->cfs_rq = parent->my_q; | |
7231 | ||
7232 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
7233 | /* guarantee group entities always have weight */ |
7234 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
7235 | se->parent = parent; |
7236 | } | |
7237 | ||
7238 | static DEFINE_MUTEX(shares_mutex); | |
7239 | ||
7240 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
7241 | { | |
7242 | int i; | |
7243 | unsigned long flags; | |
7244 | ||
7245 | /* | |
7246 | * We can't change the weight of the root cgroup. | |
7247 | */ | |
7248 | if (!tg->se[0]) | |
7249 | return -EINVAL; | |
7250 | ||
7251 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
7252 | ||
7253 | mutex_lock(&shares_mutex); | |
7254 | if (tg->shares == shares) | |
7255 | goto done; | |
7256 | ||
7257 | tg->shares = shares; | |
7258 | for_each_possible_cpu(i) { | |
7259 | struct rq *rq = cpu_rq(i); | |
7260 | struct sched_entity *se; | |
7261 | ||
7262 | se = tg->se[i]; | |
7263 | /* Propagate contribution to hierarchy */ | |
7264 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
7265 | |
7266 | /* Possible calls to update_curr() need rq clock */ | |
7267 | update_rq_clock(rq); | |
17bc14b7 | 7268 | for_each_sched_entity(se) |
029632fb PZ |
7269 | update_cfs_shares(group_cfs_rq(se)); |
7270 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7271 | } | |
7272 | ||
7273 | done: | |
7274 | mutex_unlock(&shares_mutex); | |
7275 | return 0; | |
7276 | } | |
7277 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
7278 | ||
7279 | void free_fair_sched_group(struct task_group *tg) { } | |
7280 | ||
7281 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7282 | { | |
7283 | return 1; | |
7284 | } | |
7285 | ||
7286 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
7287 | ||
7288 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
7289 | ||
810b3817 | 7290 | |
6d686f45 | 7291 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
7292 | { |
7293 | struct sched_entity *se = &task->se; | |
0d721cea PW |
7294 | unsigned int rr_interval = 0; |
7295 | ||
7296 | /* | |
7297 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
7298 | * idle runqueue: | |
7299 | */ | |
0d721cea | 7300 | if (rq->cfs.load.weight) |
a59f4e07 | 7301 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
7302 | |
7303 | return rr_interval; | |
7304 | } | |
7305 | ||
bf0f6f24 IM |
7306 | /* |
7307 | * All the scheduling class methods: | |
7308 | */ | |
029632fb | 7309 | const struct sched_class fair_sched_class = { |
5522d5d5 | 7310 | .next = &idle_sched_class, |
bf0f6f24 IM |
7311 | .enqueue_task = enqueue_task_fair, |
7312 | .dequeue_task = dequeue_task_fair, | |
7313 | .yield_task = yield_task_fair, | |
d95f4122 | 7314 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 7315 | |
2e09bf55 | 7316 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
7317 | |
7318 | .pick_next_task = pick_next_task_fair, | |
7319 | .put_prev_task = put_prev_task_fair, | |
7320 | ||
681f3e68 | 7321 | #ifdef CONFIG_SMP |
4ce72a2c | 7322 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 7323 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 7324 | |
0bcdcf28 CE |
7325 | .rq_online = rq_online_fair, |
7326 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
7327 | |
7328 | .task_waking = task_waking_fair, | |
681f3e68 | 7329 | #endif |
bf0f6f24 | 7330 | |
83b699ed | 7331 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 7332 | .task_tick = task_tick_fair, |
cd29fe6f | 7333 | .task_fork = task_fork_fair, |
cb469845 SR |
7334 | |
7335 | .prio_changed = prio_changed_fair, | |
da7a735e | 7336 | .switched_from = switched_from_fair, |
cb469845 | 7337 | .switched_to = switched_to_fair, |
810b3817 | 7338 | |
0d721cea PW |
7339 | .get_rr_interval = get_rr_interval_fair, |
7340 | ||
810b3817 | 7341 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7342 | .task_move_group = task_move_group_fair, |
810b3817 | 7343 | #endif |
bf0f6f24 IM |
7344 | }; |
7345 | ||
7346 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 7347 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 7348 | { |
bf0f6f24 IM |
7349 | struct cfs_rq *cfs_rq; |
7350 | ||
5973e5b9 | 7351 | rcu_read_lock(); |
c3b64f1e | 7352 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 7353 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 7354 | rcu_read_unlock(); |
bf0f6f24 IM |
7355 | } |
7356 | #endif | |
029632fb PZ |
7357 | |
7358 | __init void init_sched_fair_class(void) | |
7359 | { | |
7360 | #ifdef CONFIG_SMP | |
7361 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
7362 | ||
3451d024 | 7363 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 7364 | nohz.next_balance = jiffies; |
029632fb | 7365 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 7366 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
7367 | #endif |
7368 | #endif /* SMP */ | |
7369 | ||
7370 | } |