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 ? */ | |
fed14d45 | 325 | static inline struct cfs_rq * |
b758149c PZ |
326 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
327 | { | |
328 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 329 | return se->cfs_rq; |
b758149c | 330 | |
fed14d45 | 331 | return NULL; |
b758149c PZ |
332 | } |
333 | ||
334 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
335 | { | |
336 | return se->parent; | |
337 | } | |
338 | ||
464b7527 PZ |
339 | static void |
340 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
341 | { | |
342 | int se_depth, pse_depth; | |
343 | ||
344 | /* | |
345 | * preemption test can be made between sibling entities who are in the | |
346 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
347 | * both tasks until we find their ancestors who are siblings of common | |
348 | * parent. | |
349 | */ | |
350 | ||
351 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
352 | se_depth = (*se)->depth; |
353 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
354 | |
355 | while (se_depth > pse_depth) { | |
356 | se_depth--; | |
357 | *se = parent_entity(*se); | |
358 | } | |
359 | ||
360 | while (pse_depth > se_depth) { | |
361 | pse_depth--; | |
362 | *pse = parent_entity(*pse); | |
363 | } | |
364 | ||
365 | while (!is_same_group(*se, *pse)) { | |
366 | *se = parent_entity(*se); | |
367 | *pse = parent_entity(*pse); | |
368 | } | |
369 | } | |
370 | ||
8f48894f PZ |
371 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
372 | ||
373 | static inline struct task_struct *task_of(struct sched_entity *se) | |
374 | { | |
375 | return container_of(se, struct task_struct, se); | |
376 | } | |
bf0f6f24 | 377 | |
62160e3f IM |
378 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
379 | { | |
380 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
381 | } |
382 | ||
383 | #define entity_is_task(se) 1 | |
384 | ||
b758149c PZ |
385 | #define for_each_sched_entity(se) \ |
386 | for (; se; se = NULL) | |
bf0f6f24 | 387 | |
b758149c | 388 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 389 | { |
b758149c | 390 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
391 | } |
392 | ||
b758149c PZ |
393 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
394 | { | |
395 | struct task_struct *p = task_of(se); | |
396 | struct rq *rq = task_rq(p); | |
397 | ||
398 | return &rq->cfs; | |
399 | } | |
400 | ||
401 | /* runqueue "owned" by this group */ | |
402 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
403 | { | |
404 | return NULL; | |
405 | } | |
406 | ||
3d4b47b4 PZ |
407 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
408 | { | |
409 | } | |
410 | ||
411 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
412 | { | |
413 | } | |
414 | ||
b758149c PZ |
415 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
416 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
417 | ||
b758149c PZ |
418 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
419 | { | |
420 | return NULL; | |
421 | } | |
422 | ||
464b7527 PZ |
423 | static inline void |
424 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
425 | { | |
426 | } | |
427 | ||
b758149c PZ |
428 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
429 | ||
6c16a6dc | 430 | static __always_inline |
9dbdb155 | 431 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
432 | |
433 | /************************************************************** | |
434 | * Scheduling class tree data structure manipulation methods: | |
435 | */ | |
436 | ||
1bf08230 | 437 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 438 | { |
1bf08230 | 439 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 440 | if (delta > 0) |
1bf08230 | 441 | max_vruntime = vruntime; |
02e0431a | 442 | |
1bf08230 | 443 | return max_vruntime; |
02e0431a PZ |
444 | } |
445 | ||
0702e3eb | 446 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
447 | { |
448 | s64 delta = (s64)(vruntime - min_vruntime); | |
449 | if (delta < 0) | |
450 | min_vruntime = vruntime; | |
451 | ||
452 | return min_vruntime; | |
453 | } | |
454 | ||
54fdc581 FC |
455 | static inline int entity_before(struct sched_entity *a, |
456 | struct sched_entity *b) | |
457 | { | |
458 | return (s64)(a->vruntime - b->vruntime) < 0; | |
459 | } | |
460 | ||
1af5f730 PZ |
461 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
462 | { | |
463 | u64 vruntime = cfs_rq->min_vruntime; | |
464 | ||
465 | if (cfs_rq->curr) | |
466 | vruntime = cfs_rq->curr->vruntime; | |
467 | ||
468 | if (cfs_rq->rb_leftmost) { | |
469 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
470 | struct sched_entity, | |
471 | run_node); | |
472 | ||
e17036da | 473 | if (!cfs_rq->curr) |
1af5f730 PZ |
474 | vruntime = se->vruntime; |
475 | else | |
476 | vruntime = min_vruntime(vruntime, se->vruntime); | |
477 | } | |
478 | ||
1bf08230 | 479 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 480 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
481 | #ifndef CONFIG_64BIT |
482 | smp_wmb(); | |
483 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
484 | #endif | |
1af5f730 PZ |
485 | } |
486 | ||
bf0f6f24 IM |
487 | /* |
488 | * Enqueue an entity into the rb-tree: | |
489 | */ | |
0702e3eb | 490 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
491 | { |
492 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
493 | struct rb_node *parent = NULL; | |
494 | struct sched_entity *entry; | |
bf0f6f24 IM |
495 | int leftmost = 1; |
496 | ||
497 | /* | |
498 | * Find the right place in the rbtree: | |
499 | */ | |
500 | while (*link) { | |
501 | parent = *link; | |
502 | entry = rb_entry(parent, struct sched_entity, run_node); | |
503 | /* | |
504 | * We dont care about collisions. Nodes with | |
505 | * the same key stay together. | |
506 | */ | |
2bd2d6f2 | 507 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
508 | link = &parent->rb_left; |
509 | } else { | |
510 | link = &parent->rb_right; | |
511 | leftmost = 0; | |
512 | } | |
513 | } | |
514 | ||
515 | /* | |
516 | * Maintain a cache of leftmost tree entries (it is frequently | |
517 | * used): | |
518 | */ | |
1af5f730 | 519 | if (leftmost) |
57cb499d | 520 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
521 | |
522 | rb_link_node(&se->run_node, parent, link); | |
523 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
524 | } |
525 | ||
0702e3eb | 526 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 527 | { |
3fe69747 PZ |
528 | if (cfs_rq->rb_leftmost == &se->run_node) { |
529 | struct rb_node *next_node; | |
3fe69747 PZ |
530 | |
531 | next_node = rb_next(&se->run_node); | |
532 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 533 | } |
e9acbff6 | 534 | |
bf0f6f24 | 535 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
536 | } |
537 | ||
029632fb | 538 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 539 | { |
f4b6755f PZ |
540 | struct rb_node *left = cfs_rq->rb_leftmost; |
541 | ||
542 | if (!left) | |
543 | return NULL; | |
544 | ||
545 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
546 | } |
547 | ||
ac53db59 RR |
548 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
549 | { | |
550 | struct rb_node *next = rb_next(&se->run_node); | |
551 | ||
552 | if (!next) | |
553 | return NULL; | |
554 | ||
555 | return rb_entry(next, struct sched_entity, run_node); | |
556 | } | |
557 | ||
558 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 559 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 560 | { |
7eee3e67 | 561 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 562 | |
70eee74b BS |
563 | if (!last) |
564 | return NULL; | |
7eee3e67 IM |
565 | |
566 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
567 | } |
568 | ||
bf0f6f24 IM |
569 | /************************************************************** |
570 | * Scheduling class statistics methods: | |
571 | */ | |
572 | ||
acb4a848 | 573 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 574 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
575 | loff_t *ppos) |
576 | { | |
8d65af78 | 577 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 578 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
579 | |
580 | if (ret || !write) | |
581 | return ret; | |
582 | ||
583 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
584 | sysctl_sched_min_granularity); | |
585 | ||
acb4a848 CE |
586 | #define WRT_SYSCTL(name) \ |
587 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
588 | WRT_SYSCTL(sched_min_granularity); | |
589 | WRT_SYSCTL(sched_latency); | |
590 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
591 | #undef WRT_SYSCTL |
592 | ||
b2be5e96 PZ |
593 | return 0; |
594 | } | |
595 | #endif | |
647e7cac | 596 | |
a7be37ac | 597 | /* |
f9c0b095 | 598 | * delta /= w |
a7be37ac | 599 | */ |
9dbdb155 | 600 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 601 | { |
f9c0b095 | 602 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 603 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
604 | |
605 | return delta; | |
606 | } | |
607 | ||
647e7cac IM |
608 | /* |
609 | * The idea is to set a period in which each task runs once. | |
610 | * | |
532b1858 | 611 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
612 | * this period because otherwise the slices get too small. |
613 | * | |
614 | * p = (nr <= nl) ? l : l*nr/nl | |
615 | */ | |
4d78e7b6 PZ |
616 | static u64 __sched_period(unsigned long nr_running) |
617 | { | |
618 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 619 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
620 | |
621 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 622 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 623 | period *= nr_running; |
4d78e7b6 PZ |
624 | } |
625 | ||
626 | return period; | |
627 | } | |
628 | ||
647e7cac IM |
629 | /* |
630 | * We calculate the wall-time slice from the period by taking a part | |
631 | * proportional to the weight. | |
632 | * | |
f9c0b095 | 633 | * s = p*P[w/rw] |
647e7cac | 634 | */ |
6d0f0ebd | 635 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 636 | { |
0a582440 | 637 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 638 | |
0a582440 | 639 | for_each_sched_entity(se) { |
6272d68c | 640 | struct load_weight *load; |
3104bf03 | 641 | struct load_weight lw; |
6272d68c LM |
642 | |
643 | cfs_rq = cfs_rq_of(se); | |
644 | load = &cfs_rq->load; | |
f9c0b095 | 645 | |
0a582440 | 646 | if (unlikely(!se->on_rq)) { |
3104bf03 | 647 | lw = cfs_rq->load; |
0a582440 MG |
648 | |
649 | update_load_add(&lw, se->load.weight); | |
650 | load = &lw; | |
651 | } | |
9dbdb155 | 652 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
653 | } |
654 | return slice; | |
bf0f6f24 IM |
655 | } |
656 | ||
647e7cac | 657 | /* |
660cc00f | 658 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 659 | * |
f9c0b095 | 660 | * vs = s/w |
647e7cac | 661 | */ |
f9c0b095 | 662 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 663 | { |
f9c0b095 | 664 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
665 | } |
666 | ||
a75cdaa9 | 667 | #ifdef CONFIG_SMP |
fb13c7ee MG |
668 | static unsigned long task_h_load(struct task_struct *p); |
669 | ||
a75cdaa9 AS |
670 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
671 | ||
672 | /* Give new task start runnable values to heavy its load in infant time */ | |
673 | void init_task_runnable_average(struct task_struct *p) | |
674 | { | |
675 | u32 slice; | |
676 | ||
677 | p->se.avg.decay_count = 0; | |
678 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | |
679 | p->se.avg.runnable_avg_sum = slice; | |
680 | p->se.avg.runnable_avg_period = slice; | |
681 | __update_task_entity_contrib(&p->se); | |
682 | } | |
683 | #else | |
684 | void init_task_runnable_average(struct task_struct *p) | |
685 | { | |
686 | } | |
687 | #endif | |
688 | ||
bf0f6f24 | 689 | /* |
9dbdb155 | 690 | * Update the current task's runtime statistics. |
bf0f6f24 | 691 | */ |
b7cc0896 | 692 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 693 | { |
429d43bc | 694 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 695 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 696 | u64 delta_exec; |
bf0f6f24 IM |
697 | |
698 | if (unlikely(!curr)) | |
699 | return; | |
700 | ||
9dbdb155 PZ |
701 | delta_exec = now - curr->exec_start; |
702 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 703 | return; |
bf0f6f24 | 704 | |
8ebc91d9 | 705 | curr->exec_start = now; |
d842de87 | 706 | |
9dbdb155 PZ |
707 | schedstat_set(curr->statistics.exec_max, |
708 | max(delta_exec, curr->statistics.exec_max)); | |
709 | ||
710 | curr->sum_exec_runtime += delta_exec; | |
711 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
712 | ||
713 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
714 | update_min_vruntime(cfs_rq); | |
715 | ||
d842de87 SV |
716 | if (entity_is_task(curr)) { |
717 | struct task_struct *curtask = task_of(curr); | |
718 | ||
f977bb49 | 719 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 720 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 721 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 722 | } |
ec12cb7f PT |
723 | |
724 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
725 | } |
726 | ||
727 | static inline void | |
5870db5b | 728 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 729 | { |
78becc27 | 730 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
731 | } |
732 | ||
bf0f6f24 IM |
733 | /* |
734 | * Task is being enqueued - update stats: | |
735 | */ | |
d2417e5a | 736 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 737 | { |
bf0f6f24 IM |
738 | /* |
739 | * Are we enqueueing a waiting task? (for current tasks | |
740 | * a dequeue/enqueue event is a NOP) | |
741 | */ | |
429d43bc | 742 | if (se != cfs_rq->curr) |
5870db5b | 743 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
744 | } |
745 | ||
bf0f6f24 | 746 | static void |
9ef0a961 | 747 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 748 | { |
41acab88 | 749 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 750 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
751 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
752 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 753 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
754 | #ifdef CONFIG_SCHEDSTATS |
755 | if (entity_is_task(se)) { | |
756 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 757 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
758 | } |
759 | #endif | |
41acab88 | 760 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
761 | } |
762 | ||
763 | static inline void | |
19b6a2e3 | 764 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 765 | { |
bf0f6f24 IM |
766 | /* |
767 | * Mark the end of the wait period if dequeueing a | |
768 | * waiting task: | |
769 | */ | |
429d43bc | 770 | if (se != cfs_rq->curr) |
9ef0a961 | 771 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
772 | } |
773 | ||
774 | /* | |
775 | * We are picking a new current task - update its stats: | |
776 | */ | |
777 | static inline void | |
79303e9e | 778 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
779 | { |
780 | /* | |
781 | * We are starting a new run period: | |
782 | */ | |
78becc27 | 783 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
784 | } |
785 | ||
bf0f6f24 IM |
786 | /************************************************** |
787 | * Scheduling class queueing methods: | |
788 | */ | |
789 | ||
cbee9f88 PZ |
790 | #ifdef CONFIG_NUMA_BALANCING |
791 | /* | |
598f0ec0 MG |
792 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
793 | * calculated based on the tasks virtual memory size and | |
794 | * numa_balancing_scan_size. | |
cbee9f88 | 795 | */ |
598f0ec0 MG |
796 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
797 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
798 | |
799 | /* Portion of address space to scan in MB */ | |
800 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 801 | |
4b96a29b PZ |
802 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
803 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
804 | ||
598f0ec0 MG |
805 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
806 | { | |
807 | unsigned long rss = 0; | |
808 | unsigned long nr_scan_pages; | |
809 | ||
810 | /* | |
811 | * Calculations based on RSS as non-present and empty pages are skipped | |
812 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
813 | * on resident pages | |
814 | */ | |
815 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
816 | rss = get_mm_rss(p->mm); | |
817 | if (!rss) | |
818 | rss = nr_scan_pages; | |
819 | ||
820 | rss = round_up(rss, nr_scan_pages); | |
821 | return rss / nr_scan_pages; | |
822 | } | |
823 | ||
824 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
825 | #define MAX_SCAN_WINDOW 2560 | |
826 | ||
827 | static unsigned int task_scan_min(struct task_struct *p) | |
828 | { | |
829 | unsigned int scan, floor; | |
830 | unsigned int windows = 1; | |
831 | ||
832 | if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW) | |
833 | windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size; | |
834 | floor = 1000 / windows; | |
835 | ||
836 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
837 | return max_t(unsigned int, floor, scan); | |
838 | } | |
839 | ||
840 | static unsigned int task_scan_max(struct task_struct *p) | |
841 | { | |
842 | unsigned int smin = task_scan_min(p); | |
843 | unsigned int smax; | |
844 | ||
845 | /* Watch for min being lower than max due to floor calculations */ | |
846 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
847 | return max(smin, smax); | |
848 | } | |
849 | ||
0ec8aa00 PZ |
850 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
851 | { | |
852 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
853 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
854 | } | |
855 | ||
856 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
857 | { | |
858 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
859 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
860 | } | |
861 | ||
8c8a743c PZ |
862 | struct numa_group { |
863 | atomic_t refcount; | |
864 | ||
865 | spinlock_t lock; /* nr_tasks, tasks */ | |
866 | int nr_tasks; | |
e29cf08b | 867 | pid_t gid; |
8c8a743c PZ |
868 | struct list_head task_list; |
869 | ||
870 | struct rcu_head rcu; | |
20e07dea | 871 | nodemask_t active_nodes; |
989348b5 | 872 | unsigned long total_faults; |
7e2703e6 RR |
873 | /* |
874 | * Faults_cpu is used to decide whether memory should move | |
875 | * towards the CPU. As a consequence, these stats are weighted | |
876 | * more by CPU use than by memory faults. | |
877 | */ | |
50ec8a40 | 878 | unsigned long *faults_cpu; |
989348b5 | 879 | unsigned long faults[0]; |
8c8a743c PZ |
880 | }; |
881 | ||
be1e4e76 RR |
882 | /* Shared or private faults. */ |
883 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
884 | ||
885 | /* Memory and CPU locality */ | |
886 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
887 | ||
888 | /* Averaged statistics, and temporary buffers. */ | |
889 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
890 | ||
e29cf08b MG |
891 | pid_t task_numa_group_id(struct task_struct *p) |
892 | { | |
893 | return p->numa_group ? p->numa_group->gid : 0; | |
894 | } | |
895 | ||
ac8e895b MG |
896 | static inline int task_faults_idx(int nid, int priv) |
897 | { | |
be1e4e76 | 898 | return NR_NUMA_HINT_FAULT_TYPES * nid + priv; |
ac8e895b MG |
899 | } |
900 | ||
901 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
902 | { | |
ff1df896 | 903 | if (!p->numa_faults_memory) |
ac8e895b MG |
904 | return 0; |
905 | ||
ff1df896 RR |
906 | return p->numa_faults_memory[task_faults_idx(nid, 0)] + |
907 | p->numa_faults_memory[task_faults_idx(nid, 1)]; | |
ac8e895b MG |
908 | } |
909 | ||
83e1d2cd MG |
910 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
911 | { | |
912 | if (!p->numa_group) | |
913 | return 0; | |
914 | ||
82897b4f WL |
915 | return p->numa_group->faults[task_faults_idx(nid, 0)] + |
916 | p->numa_group->faults[task_faults_idx(nid, 1)]; | |
83e1d2cd MG |
917 | } |
918 | ||
20e07dea RR |
919 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
920 | { | |
921 | return group->faults_cpu[task_faults_idx(nid, 0)] + | |
922 | group->faults_cpu[task_faults_idx(nid, 1)]; | |
923 | } | |
924 | ||
83e1d2cd MG |
925 | /* |
926 | * These return the fraction of accesses done by a particular task, or | |
927 | * task group, on a particular numa node. The group weight is given a | |
928 | * larger multiplier, in order to group tasks together that are almost | |
929 | * evenly spread out between numa nodes. | |
930 | */ | |
931 | static inline unsigned long task_weight(struct task_struct *p, int nid) | |
932 | { | |
933 | unsigned long total_faults; | |
934 | ||
ff1df896 | 935 | if (!p->numa_faults_memory) |
83e1d2cd MG |
936 | return 0; |
937 | ||
938 | total_faults = p->total_numa_faults; | |
939 | ||
940 | if (!total_faults) | |
941 | return 0; | |
942 | ||
943 | return 1000 * task_faults(p, nid) / total_faults; | |
944 | } | |
945 | ||
946 | static inline unsigned long group_weight(struct task_struct *p, int nid) | |
947 | { | |
989348b5 | 948 | if (!p->numa_group || !p->numa_group->total_faults) |
83e1d2cd MG |
949 | return 0; |
950 | ||
989348b5 | 951 | return 1000 * group_faults(p, nid) / p->numa_group->total_faults; |
83e1d2cd MG |
952 | } |
953 | ||
10f39042 RR |
954 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
955 | int src_nid, int dst_cpu) | |
956 | { | |
957 | struct numa_group *ng = p->numa_group; | |
958 | int dst_nid = cpu_to_node(dst_cpu); | |
959 | int last_cpupid, this_cpupid; | |
960 | ||
961 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
962 | ||
963 | /* | |
964 | * Multi-stage node selection is used in conjunction with a periodic | |
965 | * migration fault to build a temporal task<->page relation. By using | |
966 | * a two-stage filter we remove short/unlikely relations. | |
967 | * | |
968 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
969 | * a task's usage of a particular page (n_p) per total usage of this | |
970 | * page (n_t) (in a given time-span) to a probability. | |
971 | * | |
972 | * Our periodic faults will sample this probability and getting the | |
973 | * same result twice in a row, given these samples are fully | |
974 | * independent, is then given by P(n)^2, provided our sample period | |
975 | * is sufficiently short compared to the usage pattern. | |
976 | * | |
977 | * This quadric squishes small probabilities, making it less likely we | |
978 | * act on an unlikely task<->page relation. | |
979 | */ | |
980 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
981 | if (!cpupid_pid_unset(last_cpupid) && | |
982 | cpupid_to_nid(last_cpupid) != dst_nid) | |
983 | return false; | |
984 | ||
985 | /* Always allow migrate on private faults */ | |
986 | if (cpupid_match_pid(p, last_cpupid)) | |
987 | return true; | |
988 | ||
989 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
990 | if (!ng) | |
991 | return true; | |
992 | ||
993 | /* | |
994 | * Do not migrate if the destination is not a node that | |
995 | * is actively used by this numa group. | |
996 | */ | |
997 | if (!node_isset(dst_nid, ng->active_nodes)) | |
998 | return false; | |
999 | ||
1000 | /* | |
1001 | * Source is a node that is not actively used by this | |
1002 | * numa group, while the destination is. Migrate. | |
1003 | */ | |
1004 | if (!node_isset(src_nid, ng->active_nodes)) | |
1005 | return true; | |
1006 | ||
1007 | /* | |
1008 | * Both source and destination are nodes in active | |
1009 | * use by this numa group. Maximize memory bandwidth | |
1010 | * by migrating from more heavily used groups, to less | |
1011 | * heavily used ones, spreading the load around. | |
1012 | * Use a 1/4 hysteresis to avoid spurious page movement. | |
1013 | */ | |
1014 | return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); | |
1015 | } | |
1016 | ||
e6628d5b | 1017 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
1018 | static unsigned long source_load(int cpu, int type); |
1019 | static unsigned long target_load(int cpu, int type); | |
1020 | static unsigned long power_of(int cpu); | |
1021 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
1022 | ||
fb13c7ee | 1023 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1024 | struct numa_stats { |
fb13c7ee | 1025 | unsigned long nr_running; |
58d081b5 | 1026 | unsigned long load; |
fb13c7ee MG |
1027 | |
1028 | /* Total compute capacity of CPUs on a node */ | |
1029 | unsigned long power; | |
1030 | ||
1031 | /* Approximate capacity in terms of runnable tasks on a node */ | |
1032 | unsigned long capacity; | |
1033 | int has_capacity; | |
58d081b5 | 1034 | }; |
e6628d5b | 1035 | |
fb13c7ee MG |
1036 | /* |
1037 | * XXX borrowed from update_sg_lb_stats | |
1038 | */ | |
1039 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1040 | { | |
5eca82a9 | 1041 | int cpu, cpus = 0; |
fb13c7ee MG |
1042 | |
1043 | memset(ns, 0, sizeof(*ns)); | |
1044 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1045 | struct rq *rq = cpu_rq(cpu); | |
1046 | ||
1047 | ns->nr_running += rq->nr_running; | |
1048 | ns->load += weighted_cpuload(cpu); | |
1049 | ns->power += power_of(cpu); | |
5eca82a9 PZ |
1050 | |
1051 | cpus++; | |
fb13c7ee MG |
1052 | } |
1053 | ||
5eca82a9 PZ |
1054 | /* |
1055 | * If we raced with hotplug and there are no CPUs left in our mask | |
1056 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1057 | * not find this node attractive. | |
1058 | * | |
1059 | * We'll either bail at !has_capacity, or we'll detect a huge imbalance | |
1060 | * and bail there. | |
1061 | */ | |
1062 | if (!cpus) | |
1063 | return; | |
1064 | ||
fb13c7ee MG |
1065 | ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power; |
1066 | ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE); | |
1067 | ns->has_capacity = (ns->nr_running < ns->capacity); | |
1068 | } | |
1069 | ||
58d081b5 MG |
1070 | struct task_numa_env { |
1071 | struct task_struct *p; | |
e6628d5b | 1072 | |
58d081b5 MG |
1073 | int src_cpu, src_nid; |
1074 | int dst_cpu, dst_nid; | |
e6628d5b | 1075 | |
58d081b5 | 1076 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1077 | |
40ea2b42 | 1078 | int imbalance_pct; |
fb13c7ee MG |
1079 | |
1080 | struct task_struct *best_task; | |
1081 | long best_imp; | |
58d081b5 MG |
1082 | int best_cpu; |
1083 | }; | |
1084 | ||
fb13c7ee MG |
1085 | static void task_numa_assign(struct task_numa_env *env, |
1086 | struct task_struct *p, long imp) | |
1087 | { | |
1088 | if (env->best_task) | |
1089 | put_task_struct(env->best_task); | |
1090 | if (p) | |
1091 | get_task_struct(p); | |
1092 | ||
1093 | env->best_task = p; | |
1094 | env->best_imp = imp; | |
1095 | env->best_cpu = env->dst_cpu; | |
1096 | } | |
1097 | ||
e63da036 RR |
1098 | static bool load_too_imbalanced(long orig_src_load, long orig_dst_load, |
1099 | long src_load, long dst_load, | |
1100 | struct task_numa_env *env) | |
1101 | { | |
1102 | long imb, old_imb; | |
1103 | ||
1104 | /* We care about the slope of the imbalance, not the direction. */ | |
1105 | if (dst_load < src_load) | |
1106 | swap(dst_load, src_load); | |
1107 | ||
1108 | /* Is the difference below the threshold? */ | |
1109 | imb = dst_load * 100 - src_load * env->imbalance_pct; | |
1110 | if (imb <= 0) | |
1111 | return false; | |
1112 | ||
1113 | /* | |
1114 | * The imbalance is above the allowed threshold. | |
1115 | * Compare it with the old imbalance. | |
1116 | */ | |
1117 | if (orig_dst_load < orig_src_load) | |
1118 | swap(orig_dst_load, orig_src_load); | |
1119 | ||
1120 | old_imb = orig_dst_load * 100 - orig_src_load * env->imbalance_pct; | |
1121 | ||
1122 | /* Would this change make things worse? */ | |
1123 | return (old_imb > imb); | |
1124 | } | |
1125 | ||
fb13c7ee MG |
1126 | /* |
1127 | * This checks if the overall compute and NUMA accesses of the system would | |
1128 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1129 | * into account that it might be best if task running on the dst_cpu should | |
1130 | * be exchanged with the source task | |
1131 | */ | |
887c290e RR |
1132 | static void task_numa_compare(struct task_numa_env *env, |
1133 | long taskimp, long groupimp) | |
fb13c7ee MG |
1134 | { |
1135 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1136 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1137 | struct task_struct *cur; | |
e63da036 RR |
1138 | long orig_src_load, src_load; |
1139 | long orig_dst_load, dst_load; | |
fb13c7ee | 1140 | long load; |
887c290e | 1141 | long imp = (groupimp > 0) ? groupimp : taskimp; |
fb13c7ee MG |
1142 | |
1143 | rcu_read_lock(); | |
1144 | cur = ACCESS_ONCE(dst_rq->curr); | |
1145 | if (cur->pid == 0) /* idle */ | |
1146 | cur = NULL; | |
1147 | ||
1148 | /* | |
1149 | * "imp" is the fault differential for the source task between the | |
1150 | * source and destination node. Calculate the total differential for | |
1151 | * the source task and potential destination task. The more negative | |
1152 | * the value is, the more rmeote accesses that would be expected to | |
1153 | * be incurred if the tasks were swapped. | |
1154 | */ | |
1155 | if (cur) { | |
1156 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1157 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1158 | goto unlock; | |
1159 | ||
887c290e RR |
1160 | /* |
1161 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1162 | * in any group then look only at task weights. |
887c290e | 1163 | */ |
ca28aa53 | 1164 | if (cur->numa_group == env->p->numa_group) { |
887c290e RR |
1165 | imp = taskimp + task_weight(cur, env->src_nid) - |
1166 | task_weight(cur, env->dst_nid); | |
ca28aa53 RR |
1167 | /* |
1168 | * Add some hysteresis to prevent swapping the | |
1169 | * tasks within a group over tiny differences. | |
1170 | */ | |
1171 | if (cur->numa_group) | |
1172 | imp -= imp/16; | |
887c290e | 1173 | } else { |
ca28aa53 RR |
1174 | /* |
1175 | * Compare the group weights. If a task is all by | |
1176 | * itself (not part of a group), use the task weight | |
1177 | * instead. | |
1178 | */ | |
1179 | if (env->p->numa_group) | |
1180 | imp = groupimp; | |
1181 | else | |
1182 | imp = taskimp; | |
1183 | ||
1184 | if (cur->numa_group) | |
1185 | imp += group_weight(cur, env->src_nid) - | |
1186 | group_weight(cur, env->dst_nid); | |
1187 | else | |
1188 | imp += task_weight(cur, env->src_nid) - | |
1189 | task_weight(cur, env->dst_nid); | |
887c290e | 1190 | } |
fb13c7ee MG |
1191 | } |
1192 | ||
1193 | if (imp < env->best_imp) | |
1194 | goto unlock; | |
1195 | ||
1196 | if (!cur) { | |
1197 | /* Is there capacity at our destination? */ | |
1198 | if (env->src_stats.has_capacity && | |
1199 | !env->dst_stats.has_capacity) | |
1200 | goto unlock; | |
1201 | ||
1202 | goto balance; | |
1203 | } | |
1204 | ||
1205 | /* Balance doesn't matter much if we're running a task per cpu */ | |
1206 | if (src_rq->nr_running == 1 && dst_rq->nr_running == 1) | |
1207 | goto assign; | |
1208 | ||
1209 | /* | |
1210 | * In the overloaded case, try and keep the load balanced. | |
1211 | */ | |
1212 | balance: | |
e63da036 RR |
1213 | orig_dst_load = env->dst_stats.load; |
1214 | orig_src_load = env->src_stats.load; | |
fb13c7ee MG |
1215 | |
1216 | /* XXX missing power terms */ | |
1217 | load = task_h_load(env->p); | |
e63da036 RR |
1218 | dst_load = orig_dst_load + load; |
1219 | src_load = orig_src_load - load; | |
fb13c7ee MG |
1220 | |
1221 | if (cur) { | |
1222 | load = task_h_load(cur); | |
1223 | dst_load -= load; | |
1224 | src_load += load; | |
1225 | } | |
1226 | ||
e63da036 RR |
1227 | if (load_too_imbalanced(orig_src_load, orig_dst_load, |
1228 | src_load, dst_load, env)) | |
fb13c7ee MG |
1229 | goto unlock; |
1230 | ||
1231 | assign: | |
1232 | task_numa_assign(env, cur, imp); | |
1233 | unlock: | |
1234 | rcu_read_unlock(); | |
1235 | } | |
1236 | ||
887c290e RR |
1237 | static void task_numa_find_cpu(struct task_numa_env *env, |
1238 | long taskimp, long groupimp) | |
2c8a50aa MG |
1239 | { |
1240 | int cpu; | |
1241 | ||
1242 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1243 | /* Skip this CPU if the source task cannot migrate */ | |
1244 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1245 | continue; | |
1246 | ||
1247 | env->dst_cpu = cpu; | |
887c290e | 1248 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1249 | } |
1250 | } | |
1251 | ||
58d081b5 MG |
1252 | static int task_numa_migrate(struct task_struct *p) |
1253 | { | |
58d081b5 MG |
1254 | struct task_numa_env env = { |
1255 | .p = p, | |
fb13c7ee | 1256 | |
58d081b5 | 1257 | .src_cpu = task_cpu(p), |
b32e86b4 | 1258 | .src_nid = task_node(p), |
fb13c7ee MG |
1259 | |
1260 | .imbalance_pct = 112, | |
1261 | ||
1262 | .best_task = NULL, | |
1263 | .best_imp = 0, | |
1264 | .best_cpu = -1 | |
58d081b5 MG |
1265 | }; |
1266 | struct sched_domain *sd; | |
887c290e | 1267 | unsigned long taskweight, groupweight; |
2c8a50aa | 1268 | int nid, ret; |
887c290e | 1269 | long taskimp, groupimp; |
e6628d5b | 1270 | |
58d081b5 | 1271 | /* |
fb13c7ee MG |
1272 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1273 | * imbalance and would be the first to start moving tasks about. | |
1274 | * | |
1275 | * And we want to avoid any moving of tasks about, as that would create | |
1276 | * random movement of tasks -- counter the numa conditions we're trying | |
1277 | * to satisfy here. | |
58d081b5 MG |
1278 | */ |
1279 | rcu_read_lock(); | |
fb13c7ee | 1280 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1281 | if (sd) |
1282 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1283 | rcu_read_unlock(); |
1284 | ||
46a73e8a RR |
1285 | /* |
1286 | * Cpusets can break the scheduler domain tree into smaller | |
1287 | * balance domains, some of which do not cross NUMA boundaries. | |
1288 | * Tasks that are "trapped" in such domains cannot be migrated | |
1289 | * elsewhere, so there is no point in (re)trying. | |
1290 | */ | |
1291 | if (unlikely(!sd)) { | |
de1b301a | 1292 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1293 | return -EINVAL; |
1294 | } | |
1295 | ||
887c290e RR |
1296 | taskweight = task_weight(p, env.src_nid); |
1297 | groupweight = group_weight(p, env.src_nid); | |
fb13c7ee | 1298 | update_numa_stats(&env.src_stats, env.src_nid); |
2c8a50aa | 1299 | env.dst_nid = p->numa_preferred_nid; |
887c290e RR |
1300 | taskimp = task_weight(p, env.dst_nid) - taskweight; |
1301 | groupimp = group_weight(p, env.dst_nid) - groupweight; | |
2c8a50aa | 1302 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1303 | |
e1dda8a7 RR |
1304 | /* If the preferred nid has capacity, try to use it. */ |
1305 | if (env.dst_stats.has_capacity) | |
887c290e | 1306 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 RR |
1307 | |
1308 | /* No space available on the preferred nid. Look elsewhere. */ | |
1309 | if (env.best_cpu == -1) { | |
2c8a50aa MG |
1310 | for_each_online_node(nid) { |
1311 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1312 | continue; | |
58d081b5 | 1313 | |
83e1d2cd | 1314 | /* Only consider nodes where both task and groups benefit */ |
887c290e RR |
1315 | taskimp = task_weight(p, nid) - taskweight; |
1316 | groupimp = group_weight(p, nid) - groupweight; | |
1317 | if (taskimp < 0 && groupimp < 0) | |
fb13c7ee MG |
1318 | continue; |
1319 | ||
2c8a50aa MG |
1320 | env.dst_nid = nid; |
1321 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
887c290e | 1322 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1323 | } |
1324 | } | |
1325 | ||
fb13c7ee MG |
1326 | /* No better CPU than the current one was found. */ |
1327 | if (env.best_cpu == -1) | |
1328 | return -EAGAIN; | |
1329 | ||
68d1b02a RR |
1330 | /* |
1331 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1332 | * and is migrating into one of the workload's active nodes, remember | |
1333 | * this node as the task's preferred numa node, so the workload can | |
1334 | * settle down. | |
1335 | * A task that migrated to a second choice node will be better off | |
1336 | * trying for a better one later. Do not set the preferred node here. | |
1337 | */ | |
1338 | if (p->numa_group && node_isset(env.dst_nid, p->numa_group->active_nodes)) | |
1339 | sched_setnuma(p, env.dst_nid); | |
0ec8aa00 | 1340 | |
04bb2f94 RR |
1341 | /* |
1342 | * Reset the scan period if the task is being rescheduled on an | |
1343 | * alternative node to recheck if the tasks is now properly placed. | |
1344 | */ | |
1345 | p->numa_scan_period = task_scan_min(p); | |
1346 | ||
fb13c7ee | 1347 | if (env.best_task == NULL) { |
286549dc MG |
1348 | ret = migrate_task_to(p, env.best_cpu); |
1349 | if (ret != 0) | |
1350 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1351 | return ret; |
1352 | } | |
1353 | ||
1354 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1355 | if (ret != 0) |
1356 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1357 | put_task_struct(env.best_task); |
1358 | return ret; | |
e6628d5b MG |
1359 | } |
1360 | ||
6b9a7460 MG |
1361 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1362 | static void numa_migrate_preferred(struct task_struct *p) | |
1363 | { | |
5085e2a3 RR |
1364 | unsigned long interval = HZ; |
1365 | ||
2739d3ee | 1366 | /* This task has no NUMA fault statistics yet */ |
ff1df896 | 1367 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory)) |
6b9a7460 MG |
1368 | return; |
1369 | ||
2739d3ee | 1370 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 RR |
1371 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1372 | p->numa_migrate_retry = jiffies + interval; | |
2739d3ee RR |
1373 | |
1374 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1375 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1376 | return; |
1377 | ||
1378 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1379 | task_numa_migrate(p); |
6b9a7460 MG |
1380 | } |
1381 | ||
20e07dea RR |
1382 | /* |
1383 | * Find the nodes on which the workload is actively running. We do this by | |
1384 | * tracking the nodes from which NUMA hinting faults are triggered. This can | |
1385 | * be different from the set of nodes where the workload's memory is currently | |
1386 | * located. | |
1387 | * | |
1388 | * The bitmask is used to make smarter decisions on when to do NUMA page | |
1389 | * migrations, To prevent flip-flopping, and excessive page migrations, nodes | |
1390 | * are added when they cause over 6/16 of the maximum number of faults, but | |
1391 | * only removed when they drop below 3/16. | |
1392 | */ | |
1393 | static void update_numa_active_node_mask(struct numa_group *numa_group) | |
1394 | { | |
1395 | unsigned long faults, max_faults = 0; | |
1396 | int nid; | |
1397 | ||
1398 | for_each_online_node(nid) { | |
1399 | faults = group_faults_cpu(numa_group, nid); | |
1400 | if (faults > max_faults) | |
1401 | max_faults = faults; | |
1402 | } | |
1403 | ||
1404 | for_each_online_node(nid) { | |
1405 | faults = group_faults_cpu(numa_group, nid); | |
1406 | if (!node_isset(nid, numa_group->active_nodes)) { | |
1407 | if (faults > max_faults * 6 / 16) | |
1408 | node_set(nid, numa_group->active_nodes); | |
1409 | } else if (faults < max_faults * 3 / 16) | |
1410 | node_clear(nid, numa_group->active_nodes); | |
1411 | } | |
1412 | } | |
1413 | ||
04bb2f94 RR |
1414 | /* |
1415 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1416 | * increments. The more local the fault statistics are, the higher the scan | |
1417 | * period will be for the next scan window. If local/remote ratio is below | |
1418 | * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the | |
1419 | * scan period will decrease | |
1420 | */ | |
1421 | #define NUMA_PERIOD_SLOTS 10 | |
1422 | #define NUMA_PERIOD_THRESHOLD 3 | |
1423 | ||
1424 | /* | |
1425 | * Increase the scan period (slow down scanning) if the majority of | |
1426 | * our memory is already on our local node, or if the majority of | |
1427 | * the page accesses are shared with other processes. | |
1428 | * Otherwise, decrease the scan period. | |
1429 | */ | |
1430 | static void update_task_scan_period(struct task_struct *p, | |
1431 | unsigned long shared, unsigned long private) | |
1432 | { | |
1433 | unsigned int period_slot; | |
1434 | int ratio; | |
1435 | int diff; | |
1436 | ||
1437 | unsigned long remote = p->numa_faults_locality[0]; | |
1438 | unsigned long local = p->numa_faults_locality[1]; | |
1439 | ||
1440 | /* | |
1441 | * If there were no record hinting faults then either the task is | |
1442 | * completely idle or all activity is areas that are not of interest | |
1443 | * to automatic numa balancing. Scan slower | |
1444 | */ | |
1445 | if (local + shared == 0) { | |
1446 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1447 | p->numa_scan_period << 1); | |
1448 | ||
1449 | p->mm->numa_next_scan = jiffies + | |
1450 | msecs_to_jiffies(p->numa_scan_period); | |
1451 | ||
1452 | return; | |
1453 | } | |
1454 | ||
1455 | /* | |
1456 | * Prepare to scale scan period relative to the current period. | |
1457 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1458 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1459 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1460 | */ | |
1461 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1462 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1463 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1464 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1465 | if (!slot) | |
1466 | slot = 1; | |
1467 | diff = slot * period_slot; | |
1468 | } else { | |
1469 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1470 | ||
1471 | /* | |
1472 | * Scale scan rate increases based on sharing. There is an | |
1473 | * inverse relationship between the degree of sharing and | |
1474 | * the adjustment made to the scanning period. Broadly | |
1475 | * speaking the intent is that there is little point | |
1476 | * scanning faster if shared accesses dominate as it may | |
1477 | * simply bounce migrations uselessly | |
1478 | */ | |
04bb2f94 RR |
1479 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared)); |
1480 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; | |
1481 | } | |
1482 | ||
1483 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1484 | task_scan_min(p), task_scan_max(p)); | |
1485 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1486 | } | |
1487 | ||
7e2703e6 RR |
1488 | /* |
1489 | * Get the fraction of time the task has been running since the last | |
1490 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1491 | * decays those on a 32ms period, which is orders of magnitude off | |
1492 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1493 | * stats only if the task is so new there are no NUMA statistics yet. | |
1494 | */ | |
1495 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1496 | { | |
1497 | u64 runtime, delta, now; | |
1498 | /* Use the start of this time slice to avoid calculations. */ | |
1499 | now = p->se.exec_start; | |
1500 | runtime = p->se.sum_exec_runtime; | |
1501 | ||
1502 | if (p->last_task_numa_placement) { | |
1503 | delta = runtime - p->last_sum_exec_runtime; | |
1504 | *period = now - p->last_task_numa_placement; | |
1505 | } else { | |
1506 | delta = p->se.avg.runnable_avg_sum; | |
1507 | *period = p->se.avg.runnable_avg_period; | |
1508 | } | |
1509 | ||
1510 | p->last_sum_exec_runtime = runtime; | |
1511 | p->last_task_numa_placement = now; | |
1512 | ||
1513 | return delta; | |
1514 | } | |
1515 | ||
cbee9f88 PZ |
1516 | static void task_numa_placement(struct task_struct *p) |
1517 | { | |
83e1d2cd MG |
1518 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1519 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 1520 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
1521 | unsigned long total_faults; |
1522 | u64 runtime, period; | |
7dbd13ed | 1523 | spinlock_t *group_lock = NULL; |
cbee9f88 | 1524 | |
2832bc19 | 1525 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1526 | if (p->numa_scan_seq == seq) |
1527 | return; | |
1528 | p->numa_scan_seq = seq; | |
598f0ec0 | 1529 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1530 | |
7e2703e6 RR |
1531 | total_faults = p->numa_faults_locality[0] + |
1532 | p->numa_faults_locality[1]; | |
1533 | runtime = numa_get_avg_runtime(p, &period); | |
1534 | ||
7dbd13ed MG |
1535 | /* If the task is part of a group prevent parallel updates to group stats */ |
1536 | if (p->numa_group) { | |
1537 | group_lock = &p->numa_group->lock; | |
60e69eed | 1538 | spin_lock_irq(group_lock); |
7dbd13ed MG |
1539 | } |
1540 | ||
688b7585 MG |
1541 | /* Find the node with the highest number of faults */ |
1542 | for_each_online_node(nid) { | |
83e1d2cd | 1543 | unsigned long faults = 0, group_faults = 0; |
ac8e895b | 1544 | int priv, i; |
745d6147 | 1545 | |
be1e4e76 | 1546 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 1547 | long diff, f_diff, f_weight; |
8c8a743c | 1548 | |
ac8e895b | 1549 | i = task_faults_idx(nid, priv); |
745d6147 | 1550 | |
ac8e895b | 1551 | /* Decay existing window, copy faults since last scan */ |
35664fd4 | 1552 | diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2; |
ff1df896 RR |
1553 | fault_types[priv] += p->numa_faults_buffer_memory[i]; |
1554 | p->numa_faults_buffer_memory[i] = 0; | |
fb13c7ee | 1555 | |
7e2703e6 RR |
1556 | /* |
1557 | * Normalize the faults_from, so all tasks in a group | |
1558 | * count according to CPU use, instead of by the raw | |
1559 | * number of faults. Tasks with little runtime have | |
1560 | * little over-all impact on throughput, and thus their | |
1561 | * faults are less important. | |
1562 | */ | |
1563 | f_weight = div64_u64(runtime << 16, period + 1); | |
1564 | f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) / | |
1565 | (total_faults + 1); | |
35664fd4 | 1566 | f_diff = f_weight - p->numa_faults_cpu[i] / 2; |
50ec8a40 RR |
1567 | p->numa_faults_buffer_cpu[i] = 0; |
1568 | ||
35664fd4 RR |
1569 | p->numa_faults_memory[i] += diff; |
1570 | p->numa_faults_cpu[i] += f_diff; | |
ff1df896 | 1571 | faults += p->numa_faults_memory[i]; |
83e1d2cd | 1572 | p->total_numa_faults += diff; |
8c8a743c PZ |
1573 | if (p->numa_group) { |
1574 | /* safe because we can only change our own group */ | |
989348b5 | 1575 | p->numa_group->faults[i] += diff; |
50ec8a40 | 1576 | p->numa_group->faults_cpu[i] += f_diff; |
989348b5 MG |
1577 | p->numa_group->total_faults += diff; |
1578 | group_faults += p->numa_group->faults[i]; | |
8c8a743c | 1579 | } |
ac8e895b MG |
1580 | } |
1581 | ||
688b7585 MG |
1582 | if (faults > max_faults) { |
1583 | max_faults = faults; | |
1584 | max_nid = nid; | |
1585 | } | |
83e1d2cd MG |
1586 | |
1587 | if (group_faults > max_group_faults) { | |
1588 | max_group_faults = group_faults; | |
1589 | max_group_nid = nid; | |
1590 | } | |
1591 | } | |
1592 | ||
04bb2f94 RR |
1593 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1594 | ||
7dbd13ed | 1595 | if (p->numa_group) { |
20e07dea | 1596 | update_numa_active_node_mask(p->numa_group); |
7dbd13ed MG |
1597 | /* |
1598 | * If the preferred task and group nids are different, | |
1599 | * iterate over the nodes again to find the best place. | |
1600 | */ | |
1601 | if (max_nid != max_group_nid) { | |
1602 | unsigned long weight, max_weight = 0; | |
1603 | ||
1604 | for_each_online_node(nid) { | |
1605 | weight = task_weight(p, nid) + group_weight(p, nid); | |
1606 | if (weight > max_weight) { | |
1607 | max_weight = weight; | |
1608 | max_nid = nid; | |
1609 | } | |
83e1d2cd MG |
1610 | } |
1611 | } | |
7dbd13ed | 1612 | |
60e69eed | 1613 | spin_unlock_irq(group_lock); |
688b7585 MG |
1614 | } |
1615 | ||
6b9a7460 | 1616 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1617 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1618 | /* Update the preferred nid and migrate task if possible */ |
0ec8aa00 | 1619 | sched_setnuma(p, max_nid); |
6b9a7460 | 1620 | numa_migrate_preferred(p); |
3a7053b3 | 1621 | } |
cbee9f88 PZ |
1622 | } |
1623 | ||
8c8a743c PZ |
1624 | static inline int get_numa_group(struct numa_group *grp) |
1625 | { | |
1626 | return atomic_inc_not_zero(&grp->refcount); | |
1627 | } | |
1628 | ||
1629 | static inline void put_numa_group(struct numa_group *grp) | |
1630 | { | |
1631 | if (atomic_dec_and_test(&grp->refcount)) | |
1632 | kfree_rcu(grp, rcu); | |
1633 | } | |
1634 | ||
3e6a9418 MG |
1635 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1636 | int *priv) | |
8c8a743c PZ |
1637 | { |
1638 | struct numa_group *grp, *my_grp; | |
1639 | struct task_struct *tsk; | |
1640 | bool join = false; | |
1641 | int cpu = cpupid_to_cpu(cpupid); | |
1642 | int i; | |
1643 | ||
1644 | if (unlikely(!p->numa_group)) { | |
1645 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 1646 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
1647 | |
1648 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1649 | if (!grp) | |
1650 | return; | |
1651 | ||
1652 | atomic_set(&grp->refcount, 1); | |
1653 | spin_lock_init(&grp->lock); | |
1654 | INIT_LIST_HEAD(&grp->task_list); | |
e29cf08b | 1655 | grp->gid = p->pid; |
50ec8a40 | 1656 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
1657 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
1658 | nr_node_ids; | |
8c8a743c | 1659 | |
20e07dea RR |
1660 | node_set(task_node(current), grp->active_nodes); |
1661 | ||
be1e4e76 | 1662 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
ff1df896 | 1663 | grp->faults[i] = p->numa_faults_memory[i]; |
8c8a743c | 1664 | |
989348b5 | 1665 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 1666 | |
8c8a743c PZ |
1667 | list_add(&p->numa_entry, &grp->task_list); |
1668 | grp->nr_tasks++; | |
1669 | rcu_assign_pointer(p->numa_group, grp); | |
1670 | } | |
1671 | ||
1672 | rcu_read_lock(); | |
1673 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1674 | ||
1675 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 1676 | goto no_join; |
8c8a743c PZ |
1677 | |
1678 | grp = rcu_dereference(tsk->numa_group); | |
1679 | if (!grp) | |
3354781a | 1680 | goto no_join; |
8c8a743c PZ |
1681 | |
1682 | my_grp = p->numa_group; | |
1683 | if (grp == my_grp) | |
3354781a | 1684 | goto no_join; |
8c8a743c PZ |
1685 | |
1686 | /* | |
1687 | * Only join the other group if its bigger; if we're the bigger group, | |
1688 | * the other task will join us. | |
1689 | */ | |
1690 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 1691 | goto no_join; |
8c8a743c PZ |
1692 | |
1693 | /* | |
1694 | * Tie-break on the grp address. | |
1695 | */ | |
1696 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 1697 | goto no_join; |
8c8a743c | 1698 | |
dabe1d99 RR |
1699 | /* Always join threads in the same process. */ |
1700 | if (tsk->mm == current->mm) | |
1701 | join = true; | |
1702 | ||
1703 | /* Simple filter to avoid false positives due to PID collisions */ | |
1704 | if (flags & TNF_SHARED) | |
1705 | join = true; | |
8c8a743c | 1706 | |
3e6a9418 MG |
1707 | /* Update priv based on whether false sharing was detected */ |
1708 | *priv = !join; | |
1709 | ||
dabe1d99 | 1710 | if (join && !get_numa_group(grp)) |
3354781a | 1711 | goto no_join; |
8c8a743c | 1712 | |
8c8a743c PZ |
1713 | rcu_read_unlock(); |
1714 | ||
1715 | if (!join) | |
1716 | return; | |
1717 | ||
60e69eed MG |
1718 | BUG_ON(irqs_disabled()); |
1719 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 1720 | |
be1e4e76 | 1721 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
ff1df896 RR |
1722 | my_grp->faults[i] -= p->numa_faults_memory[i]; |
1723 | grp->faults[i] += p->numa_faults_memory[i]; | |
8c8a743c | 1724 | } |
989348b5 MG |
1725 | my_grp->total_faults -= p->total_numa_faults; |
1726 | grp->total_faults += p->total_numa_faults; | |
8c8a743c PZ |
1727 | |
1728 | list_move(&p->numa_entry, &grp->task_list); | |
1729 | my_grp->nr_tasks--; | |
1730 | grp->nr_tasks++; | |
1731 | ||
1732 | spin_unlock(&my_grp->lock); | |
60e69eed | 1733 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
1734 | |
1735 | rcu_assign_pointer(p->numa_group, grp); | |
1736 | ||
1737 | put_numa_group(my_grp); | |
3354781a PZ |
1738 | return; |
1739 | ||
1740 | no_join: | |
1741 | rcu_read_unlock(); | |
1742 | return; | |
8c8a743c PZ |
1743 | } |
1744 | ||
1745 | void task_numa_free(struct task_struct *p) | |
1746 | { | |
1747 | struct numa_group *grp = p->numa_group; | |
1748 | int i; | |
ff1df896 | 1749 | void *numa_faults = p->numa_faults_memory; |
8c8a743c PZ |
1750 | |
1751 | if (grp) { | |
60e69eed | 1752 | spin_lock_irq(&grp->lock); |
be1e4e76 | 1753 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
ff1df896 | 1754 | grp->faults[i] -= p->numa_faults_memory[i]; |
989348b5 | 1755 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 1756 | |
8c8a743c PZ |
1757 | list_del(&p->numa_entry); |
1758 | grp->nr_tasks--; | |
60e69eed | 1759 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
1760 | rcu_assign_pointer(p->numa_group, NULL); |
1761 | put_numa_group(grp); | |
1762 | } | |
1763 | ||
ff1df896 RR |
1764 | p->numa_faults_memory = NULL; |
1765 | p->numa_faults_buffer_memory = NULL; | |
50ec8a40 RR |
1766 | p->numa_faults_cpu= NULL; |
1767 | p->numa_faults_buffer_cpu = NULL; | |
82727018 | 1768 | kfree(numa_faults); |
8c8a743c PZ |
1769 | } |
1770 | ||
cbee9f88 PZ |
1771 | /* |
1772 | * Got a PROT_NONE fault for a page on @node. | |
1773 | */ | |
58b46da3 | 1774 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
1775 | { |
1776 | struct task_struct *p = current; | |
6688cc05 | 1777 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 1778 | int cpu_node = task_node(current); |
792568ec | 1779 | int local = !!(flags & TNF_FAULT_LOCAL); |
ac8e895b | 1780 | int priv; |
cbee9f88 | 1781 | |
10e84b97 | 1782 | if (!numabalancing_enabled) |
1a687c2e MG |
1783 | return; |
1784 | ||
9ff1d9ff MG |
1785 | /* for example, ksmd faulting in a user's mm */ |
1786 | if (!p->mm) | |
1787 | return; | |
1788 | ||
82727018 RR |
1789 | /* Do not worry about placement if exiting */ |
1790 | if (p->state == TASK_DEAD) | |
1791 | return; | |
1792 | ||
f809ca9a | 1793 | /* Allocate buffer to track faults on a per-node basis */ |
ff1df896 | 1794 | if (unlikely(!p->numa_faults_memory)) { |
be1e4e76 RR |
1795 | int size = sizeof(*p->numa_faults_memory) * |
1796 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; | |
f809ca9a | 1797 | |
be1e4e76 | 1798 | p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
ff1df896 | 1799 | if (!p->numa_faults_memory) |
f809ca9a | 1800 | return; |
745d6147 | 1801 | |
ff1df896 | 1802 | BUG_ON(p->numa_faults_buffer_memory); |
be1e4e76 RR |
1803 | /* |
1804 | * The averaged statistics, shared & private, memory & cpu, | |
1805 | * occupy the first half of the array. The second half of the | |
1806 | * array is for current counters, which are averaged into the | |
1807 | * first set by task_numa_placement. | |
1808 | */ | |
50ec8a40 RR |
1809 | p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids); |
1810 | p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids); | |
1811 | p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids); | |
83e1d2cd | 1812 | p->total_numa_faults = 0; |
04bb2f94 | 1813 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 1814 | } |
cbee9f88 | 1815 | |
8c8a743c PZ |
1816 | /* |
1817 | * First accesses are treated as private, otherwise consider accesses | |
1818 | * to be private if the accessing pid has not changed | |
1819 | */ | |
1820 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
1821 | priv = 1; | |
1822 | } else { | |
1823 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 1824 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 1825 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
1826 | } |
1827 | ||
792568ec RR |
1828 | /* |
1829 | * If a workload spans multiple NUMA nodes, a shared fault that | |
1830 | * occurs wholly within the set of nodes that the workload is | |
1831 | * actively using should be counted as local. This allows the | |
1832 | * scan rate to slow down when a workload has settled down. | |
1833 | */ | |
1834 | if (!priv && !local && p->numa_group && | |
1835 | node_isset(cpu_node, p->numa_group->active_nodes) && | |
1836 | node_isset(mem_node, p->numa_group->active_nodes)) | |
1837 | local = 1; | |
1838 | ||
cbee9f88 | 1839 | task_numa_placement(p); |
f809ca9a | 1840 | |
2739d3ee RR |
1841 | /* |
1842 | * Retry task to preferred node migration periodically, in case it | |
1843 | * case it previously failed, or the scheduler moved us. | |
1844 | */ | |
1845 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
1846 | numa_migrate_preferred(p); |
1847 | ||
b32e86b4 IM |
1848 | if (migrated) |
1849 | p->numa_pages_migrated += pages; | |
1850 | ||
58b46da3 RR |
1851 | p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages; |
1852 | p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages; | |
792568ec | 1853 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
1854 | } |
1855 | ||
6e5fb223 PZ |
1856 | static void reset_ptenuma_scan(struct task_struct *p) |
1857 | { | |
1858 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1859 | p->mm->numa_scan_offset = 0; | |
1860 | } | |
1861 | ||
cbee9f88 PZ |
1862 | /* |
1863 | * The expensive part of numa migration is done from task_work context. | |
1864 | * Triggered from task_tick_numa(). | |
1865 | */ | |
1866 | void task_numa_work(struct callback_head *work) | |
1867 | { | |
1868 | unsigned long migrate, next_scan, now = jiffies; | |
1869 | struct task_struct *p = current; | |
1870 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1871 | struct vm_area_struct *vma; |
9f40604c | 1872 | unsigned long start, end; |
598f0ec0 | 1873 | unsigned long nr_pte_updates = 0; |
9f40604c | 1874 | long pages; |
cbee9f88 PZ |
1875 | |
1876 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1877 | ||
1878 | work->next = work; /* protect against double add */ | |
1879 | /* | |
1880 | * Who cares about NUMA placement when they're dying. | |
1881 | * | |
1882 | * NOTE: make sure not to dereference p->mm before this check, | |
1883 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1884 | * without p->mm even though we still had it when we enqueued this | |
1885 | * work. | |
1886 | */ | |
1887 | if (p->flags & PF_EXITING) | |
1888 | return; | |
1889 | ||
930aa174 | 1890 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
1891 | mm->numa_next_scan = now + |
1892 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
1893 | } |
1894 | ||
cbee9f88 PZ |
1895 | /* |
1896 | * Enforce maximal scan/migration frequency.. | |
1897 | */ | |
1898 | migrate = mm->numa_next_scan; | |
1899 | if (time_before(now, migrate)) | |
1900 | return; | |
1901 | ||
598f0ec0 MG |
1902 | if (p->numa_scan_period == 0) { |
1903 | p->numa_scan_period_max = task_scan_max(p); | |
1904 | p->numa_scan_period = task_scan_min(p); | |
1905 | } | |
cbee9f88 | 1906 | |
fb003b80 | 1907 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1908 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1909 | return; | |
1910 | ||
19a78d11 PZ |
1911 | /* |
1912 | * Delay this task enough that another task of this mm will likely win | |
1913 | * the next time around. | |
1914 | */ | |
1915 | p->node_stamp += 2 * TICK_NSEC; | |
1916 | ||
9f40604c MG |
1917 | start = mm->numa_scan_offset; |
1918 | pages = sysctl_numa_balancing_scan_size; | |
1919 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1920 | if (!pages) | |
1921 | return; | |
cbee9f88 | 1922 | |
6e5fb223 | 1923 | down_read(&mm->mmap_sem); |
9f40604c | 1924 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1925 | if (!vma) { |
1926 | reset_ptenuma_scan(p); | |
9f40604c | 1927 | start = 0; |
6e5fb223 PZ |
1928 | vma = mm->mmap; |
1929 | } | |
9f40604c | 1930 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1931 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1932 | continue; |
1933 | ||
4591ce4f MG |
1934 | /* |
1935 | * Shared library pages mapped by multiple processes are not | |
1936 | * migrated as it is expected they are cache replicated. Avoid | |
1937 | * hinting faults in read-only file-backed mappings or the vdso | |
1938 | * as migrating the pages will be of marginal benefit. | |
1939 | */ | |
1940 | if (!vma->vm_mm || | |
1941 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1942 | continue; | |
1943 | ||
3c67f474 MG |
1944 | /* |
1945 | * Skip inaccessible VMAs to avoid any confusion between | |
1946 | * PROT_NONE and NUMA hinting ptes | |
1947 | */ | |
1948 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
1949 | continue; | |
4591ce4f | 1950 | |
9f40604c MG |
1951 | do { |
1952 | start = max(start, vma->vm_start); | |
1953 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1954 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1955 | nr_pte_updates += change_prot_numa(vma, start, end); |
1956 | ||
1957 | /* | |
1958 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1959 | * at least one PTE is updated so that unused virtual | |
1960 | * address space is quickly skipped. | |
1961 | */ | |
1962 | if (nr_pte_updates) | |
1963 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1964 | |
9f40604c MG |
1965 | start = end; |
1966 | if (pages <= 0) | |
1967 | goto out; | |
3cf1962c RR |
1968 | |
1969 | cond_resched(); | |
9f40604c | 1970 | } while (end != vma->vm_end); |
cbee9f88 | 1971 | } |
6e5fb223 | 1972 | |
9f40604c | 1973 | out: |
6e5fb223 | 1974 | /* |
c69307d5 PZ |
1975 | * It is possible to reach the end of the VMA list but the last few |
1976 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1977 | * would find the !migratable VMA on the next scan but not reset the | |
1978 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1979 | */ |
1980 | if (vma) | |
9f40604c | 1981 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1982 | else |
1983 | reset_ptenuma_scan(p); | |
1984 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1985 | } |
1986 | ||
1987 | /* | |
1988 | * Drive the periodic memory faults.. | |
1989 | */ | |
1990 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1991 | { | |
1992 | struct callback_head *work = &curr->numa_work; | |
1993 | u64 period, now; | |
1994 | ||
1995 | /* | |
1996 | * We don't care about NUMA placement if we don't have memory. | |
1997 | */ | |
1998 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1999 | return; | |
2000 | ||
2001 | /* | |
2002 | * Using runtime rather than walltime has the dual advantage that | |
2003 | * we (mostly) drive the selection from busy threads and that the | |
2004 | * task needs to have done some actual work before we bother with | |
2005 | * NUMA placement. | |
2006 | */ | |
2007 | now = curr->se.sum_exec_runtime; | |
2008 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2009 | ||
2010 | if (now - curr->node_stamp > period) { | |
4b96a29b | 2011 | if (!curr->node_stamp) |
598f0ec0 | 2012 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 2013 | curr->node_stamp += period; |
cbee9f88 PZ |
2014 | |
2015 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2016 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2017 | task_work_add(curr, work, true); | |
2018 | } | |
2019 | } | |
2020 | } | |
2021 | #else | |
2022 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2023 | { | |
2024 | } | |
0ec8aa00 PZ |
2025 | |
2026 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2027 | { | |
2028 | } | |
2029 | ||
2030 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2031 | { | |
2032 | } | |
cbee9f88 PZ |
2033 | #endif /* CONFIG_NUMA_BALANCING */ |
2034 | ||
30cfdcfc DA |
2035 | static void |
2036 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2037 | { | |
2038 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2039 | if (!parent_entity(se)) |
029632fb | 2040 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2041 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2042 | if (entity_is_task(se)) { |
2043 | struct rq *rq = rq_of(cfs_rq); | |
2044 | ||
2045 | account_numa_enqueue(rq, task_of(se)); | |
2046 | list_add(&se->group_node, &rq->cfs_tasks); | |
2047 | } | |
367456c7 | 2048 | #endif |
30cfdcfc | 2049 | cfs_rq->nr_running++; |
30cfdcfc DA |
2050 | } |
2051 | ||
2052 | static void | |
2053 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2054 | { | |
2055 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2056 | if (!parent_entity(se)) |
029632fb | 2057 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
0ec8aa00 PZ |
2058 | if (entity_is_task(se)) { |
2059 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2060 | list_del_init(&se->group_node); |
0ec8aa00 | 2061 | } |
30cfdcfc | 2062 | cfs_rq->nr_running--; |
30cfdcfc DA |
2063 | } |
2064 | ||
3ff6dcac YZ |
2065 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2066 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
2067 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
2068 | { | |
2069 | long tg_weight; | |
2070 | ||
2071 | /* | |
2072 | * Use this CPU's actual weight instead of the last load_contribution | |
2073 | * to gain a more accurate current total weight. See | |
2074 | * update_cfs_rq_load_contribution(). | |
2075 | */ | |
bf5b986e | 2076 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 2077 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
2078 | tg_weight += cfs_rq->load.weight; |
2079 | ||
2080 | return tg_weight; | |
2081 | } | |
2082 | ||
6d5ab293 | 2083 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 2084 | { |
cf5f0acf | 2085 | long tg_weight, load, shares; |
3ff6dcac | 2086 | |
cf5f0acf | 2087 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 2088 | load = cfs_rq->load.weight; |
3ff6dcac | 2089 | |
3ff6dcac | 2090 | shares = (tg->shares * load); |
cf5f0acf PZ |
2091 | if (tg_weight) |
2092 | shares /= tg_weight; | |
3ff6dcac YZ |
2093 | |
2094 | if (shares < MIN_SHARES) | |
2095 | shares = MIN_SHARES; | |
2096 | if (shares > tg->shares) | |
2097 | shares = tg->shares; | |
2098 | ||
2099 | return shares; | |
2100 | } | |
3ff6dcac | 2101 | # else /* CONFIG_SMP */ |
6d5ab293 | 2102 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
2103 | { |
2104 | return tg->shares; | |
2105 | } | |
3ff6dcac | 2106 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
2107 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2108 | unsigned long weight) | |
2109 | { | |
19e5eebb PT |
2110 | if (se->on_rq) { |
2111 | /* commit outstanding execution time */ | |
2112 | if (cfs_rq->curr == se) | |
2113 | update_curr(cfs_rq); | |
2069dd75 | 2114 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 2115 | } |
2069dd75 PZ |
2116 | |
2117 | update_load_set(&se->load, weight); | |
2118 | ||
2119 | if (se->on_rq) | |
2120 | account_entity_enqueue(cfs_rq, se); | |
2121 | } | |
2122 | ||
82958366 PT |
2123 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2124 | ||
6d5ab293 | 2125 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2126 | { |
2127 | struct task_group *tg; | |
2128 | struct sched_entity *se; | |
3ff6dcac | 2129 | long shares; |
2069dd75 | 2130 | |
2069dd75 PZ |
2131 | tg = cfs_rq->tg; |
2132 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 2133 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 2134 | return; |
3ff6dcac YZ |
2135 | #ifndef CONFIG_SMP |
2136 | if (likely(se->load.weight == tg->shares)) | |
2137 | return; | |
2138 | #endif | |
6d5ab293 | 2139 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
2140 | |
2141 | reweight_entity(cfs_rq_of(se), se, shares); | |
2142 | } | |
2143 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 2144 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2145 | { |
2146 | } | |
2147 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
2148 | ||
141965c7 | 2149 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
2150 | /* |
2151 | * We choose a half-life close to 1 scheduling period. | |
2152 | * Note: The tables below are dependent on this value. | |
2153 | */ | |
2154 | #define LOAD_AVG_PERIOD 32 | |
2155 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
2156 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
2157 | ||
2158 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
2159 | static const u32 runnable_avg_yN_inv[] = { | |
2160 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
2161 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
2162 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
2163 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
2164 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
2165 | 0x85aac367, 0x82cd8698, | |
2166 | }; | |
2167 | ||
2168 | /* | |
2169 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
2170 | * over-estimates when re-combining. | |
2171 | */ | |
2172 | static const u32 runnable_avg_yN_sum[] = { | |
2173 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
2174 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
2175 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
2176 | }; | |
2177 | ||
9d85f21c PT |
2178 | /* |
2179 | * Approximate: | |
2180 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
2181 | */ | |
2182 | static __always_inline u64 decay_load(u64 val, u64 n) | |
2183 | { | |
5b51f2f8 PT |
2184 | unsigned int local_n; |
2185 | ||
2186 | if (!n) | |
2187 | return val; | |
2188 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
2189 | return 0; | |
2190 | ||
2191 | /* after bounds checking we can collapse to 32-bit */ | |
2192 | local_n = n; | |
2193 | ||
2194 | /* | |
2195 | * As y^PERIOD = 1/2, we can combine | |
2196 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
2197 | * With a look-up table which covers k^n (n<PERIOD) | |
2198 | * | |
2199 | * To achieve constant time decay_load. | |
2200 | */ | |
2201 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2202 | val >>= local_n / LOAD_AVG_PERIOD; | |
2203 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2204 | } |
2205 | ||
5b51f2f8 PT |
2206 | val *= runnable_avg_yN_inv[local_n]; |
2207 | /* We don't use SRR here since we always want to round down. */ | |
2208 | return val >> 32; | |
2209 | } | |
2210 | ||
2211 | /* | |
2212 | * For updates fully spanning n periods, the contribution to runnable | |
2213 | * average will be: \Sum 1024*y^n | |
2214 | * | |
2215 | * We can compute this reasonably efficiently by combining: | |
2216 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2217 | */ | |
2218 | static u32 __compute_runnable_contrib(u64 n) | |
2219 | { | |
2220 | u32 contrib = 0; | |
2221 | ||
2222 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2223 | return runnable_avg_yN_sum[n]; | |
2224 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2225 | return LOAD_AVG_MAX; | |
2226 | ||
2227 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
2228 | do { | |
2229 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
2230 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
2231 | ||
2232 | n -= LOAD_AVG_PERIOD; | |
2233 | } while (n > LOAD_AVG_PERIOD); | |
2234 | ||
2235 | contrib = decay_load(contrib, n); | |
2236 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2237 | } |
2238 | ||
2239 | /* | |
2240 | * We can represent the historical contribution to runnable average as the | |
2241 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2242 | * history into segments of approximately 1ms (1024us); label the segment that | |
2243 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2244 | * | |
2245 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2246 | * p0 p1 p2 | |
2247 | * (now) (~1ms ago) (~2ms ago) | |
2248 | * | |
2249 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2250 | * | |
2251 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2252 | * following representation of historical load: | |
2253 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2254 | * | |
2255 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2256 | * y^32 = 0.5 | |
2257 | * | |
2258 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2259 | * approximately half as much as the contribution to load within the last ms | |
2260 | * (u_0). | |
2261 | * | |
2262 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2263 | * sum again by y is sufficient to update: | |
2264 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2265 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2266 | */ | |
2267 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
2268 | struct sched_avg *sa, | |
2269 | int runnable) | |
2270 | { | |
5b51f2f8 PT |
2271 | u64 delta, periods; |
2272 | u32 runnable_contrib; | |
9d85f21c PT |
2273 | int delta_w, decayed = 0; |
2274 | ||
2275 | delta = now - sa->last_runnable_update; | |
2276 | /* | |
2277 | * This should only happen when time goes backwards, which it | |
2278 | * unfortunately does during sched clock init when we swap over to TSC. | |
2279 | */ | |
2280 | if ((s64)delta < 0) { | |
2281 | sa->last_runnable_update = now; | |
2282 | return 0; | |
2283 | } | |
2284 | ||
2285 | /* | |
2286 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2287 | * approximation of 1us and fast to compute. | |
2288 | */ | |
2289 | delta >>= 10; | |
2290 | if (!delta) | |
2291 | return 0; | |
2292 | sa->last_runnable_update = now; | |
2293 | ||
2294 | /* delta_w is the amount already accumulated against our next period */ | |
2295 | delta_w = sa->runnable_avg_period % 1024; | |
2296 | if (delta + delta_w >= 1024) { | |
2297 | /* period roll-over */ | |
2298 | decayed = 1; | |
2299 | ||
2300 | /* | |
2301 | * Now that we know we're crossing a period boundary, figure | |
2302 | * out how much from delta we need to complete the current | |
2303 | * period and accrue it. | |
2304 | */ | |
2305 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
2306 | if (runnable) |
2307 | sa->runnable_avg_sum += delta_w; | |
2308 | sa->runnable_avg_period += delta_w; | |
2309 | ||
2310 | delta -= delta_w; | |
2311 | ||
2312 | /* Figure out how many additional periods this update spans */ | |
2313 | periods = delta / 1024; | |
2314 | delta %= 1024; | |
2315 | ||
2316 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
2317 | periods + 1); | |
2318 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
2319 | periods + 1); | |
2320 | ||
2321 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
2322 | runnable_contrib = __compute_runnable_contrib(periods); | |
2323 | if (runnable) | |
2324 | sa->runnable_avg_sum += runnable_contrib; | |
2325 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
2326 | } |
2327 | ||
2328 | /* Remainder of delta accrued against u_0` */ | |
2329 | if (runnable) | |
2330 | sa->runnable_avg_sum += delta; | |
2331 | sa->runnable_avg_period += delta; | |
2332 | ||
2333 | return decayed; | |
2334 | } | |
2335 | ||
9ee474f5 | 2336 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 2337 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
2338 | { |
2339 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2340 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
2341 | ||
2342 | decays -= se->avg.decay_count; | |
2343 | if (!decays) | |
aff3e498 | 2344 | return 0; |
9ee474f5 PT |
2345 | |
2346 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
2347 | se->avg.decay_count = 0; | |
aff3e498 PT |
2348 | |
2349 | return decays; | |
9ee474f5 PT |
2350 | } |
2351 | ||
c566e8e9 PT |
2352 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2353 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2354 | int force_update) | |
2355 | { | |
2356 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 2357 | long tg_contrib; |
c566e8e9 PT |
2358 | |
2359 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
2360 | tg_contrib -= cfs_rq->tg_load_contrib; | |
2361 | ||
bf5b986e AS |
2362 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
2363 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
2364 | cfs_rq->tg_load_contrib += tg_contrib; |
2365 | } | |
2366 | } | |
8165e145 | 2367 | |
bb17f655 PT |
2368 | /* |
2369 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
2370 | * representation for computing load contributions. | |
2371 | */ | |
2372 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
2373 | struct cfs_rq *cfs_rq) | |
2374 | { | |
2375 | struct task_group *tg = cfs_rq->tg; | |
2376 | long contrib; | |
2377 | ||
2378 | /* The fraction of a cpu used by this cfs_rq */ | |
85b088e9 | 2379 | contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, |
bb17f655 PT |
2380 | sa->runnable_avg_period + 1); |
2381 | contrib -= cfs_rq->tg_runnable_contrib; | |
2382 | ||
2383 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
2384 | atomic_add(contrib, &tg->runnable_avg); | |
2385 | cfs_rq->tg_runnable_contrib += contrib; | |
2386 | } | |
2387 | } | |
2388 | ||
8165e145 PT |
2389 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
2390 | { | |
2391 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
2392 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
2393 | int runnable_avg; |
2394 | ||
8165e145 PT |
2395 | u64 contrib; |
2396 | ||
2397 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
2398 | se->avg.load_avg_contrib = div_u64(contrib, |
2399 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
2400 | |
2401 | /* | |
2402 | * For group entities we need to compute a correction term in the case | |
2403 | * that they are consuming <1 cpu so that we would contribute the same | |
2404 | * load as a task of equal weight. | |
2405 | * | |
2406 | * Explicitly co-ordinating this measurement would be expensive, but | |
2407 | * fortunately the sum of each cpus contribution forms a usable | |
2408 | * lower-bound on the true value. | |
2409 | * | |
2410 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
2411 | * (and the sum represents true value) or they are disjoint and we are | |
2412 | * understating by the aggregate of their overlap. | |
2413 | * | |
2414 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
2415 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
2416 | * cpus that overlap for this interval and w_i is the interval width. | |
2417 | * | |
2418 | * On a small machine; the first term is well-bounded which bounds the | |
2419 | * total error since w_i is a subset of the period. Whereas on a | |
2420 | * larger machine, while this first term can be larger, if w_i is the | |
2421 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
2422 | * our upper bound of 1-cpu. | |
2423 | */ | |
2424 | runnable_avg = atomic_read(&tg->runnable_avg); | |
2425 | if (runnable_avg < NICE_0_LOAD) { | |
2426 | se->avg.load_avg_contrib *= runnable_avg; | |
2427 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
2428 | } | |
8165e145 | 2429 | } |
f5f9739d DE |
2430 | |
2431 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2432 | { | |
2433 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); | |
2434 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); | |
2435 | } | |
6e83125c | 2436 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 PT |
2437 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, |
2438 | int force_update) {} | |
bb17f655 PT |
2439 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2440 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 2441 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
f5f9739d | 2442 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
6e83125c | 2443 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 2444 | |
8165e145 PT |
2445 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2446 | { | |
2447 | u32 contrib; | |
2448 | ||
2449 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2450 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2451 | contrib /= (se->avg.runnable_avg_period + 1); | |
2452 | se->avg.load_avg_contrib = scale_load(contrib); | |
2453 | } | |
2454 | ||
2dac754e PT |
2455 | /* Compute the current contribution to load_avg by se, return any delta */ |
2456 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2457 | { | |
2458 | long old_contrib = se->avg.load_avg_contrib; | |
2459 | ||
8165e145 PT |
2460 | if (entity_is_task(se)) { |
2461 | __update_task_entity_contrib(se); | |
2462 | } else { | |
bb17f655 | 2463 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2464 | __update_group_entity_contrib(se); |
2465 | } | |
2dac754e PT |
2466 | |
2467 | return se->avg.load_avg_contrib - old_contrib; | |
2468 | } | |
2469 | ||
9ee474f5 PT |
2470 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2471 | long load_contrib) | |
2472 | { | |
2473 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2474 | cfs_rq->blocked_load_avg -= load_contrib; | |
2475 | else | |
2476 | cfs_rq->blocked_load_avg = 0; | |
2477 | } | |
2478 | ||
f1b17280 PT |
2479 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2480 | ||
9d85f21c | 2481 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2482 | static inline void update_entity_load_avg(struct sched_entity *se, |
2483 | int update_cfs_rq) | |
9d85f21c | 2484 | { |
2dac754e PT |
2485 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2486 | long contrib_delta; | |
f1b17280 | 2487 | u64 now; |
2dac754e | 2488 | |
f1b17280 PT |
2489 | /* |
2490 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2491 | * case they are the parent of a throttled hierarchy. | |
2492 | */ | |
2493 | if (entity_is_task(se)) | |
2494 | now = cfs_rq_clock_task(cfs_rq); | |
2495 | else | |
2496 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2497 | ||
2498 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2499 | return; |
2500 | ||
2501 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2502 | |
2503 | if (!update_cfs_rq) | |
2504 | return; | |
2505 | ||
2dac754e PT |
2506 | if (se->on_rq) |
2507 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2508 | else |
2509 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2510 | } | |
2511 | ||
2512 | /* | |
2513 | * Decay the load contributed by all blocked children and account this so that | |
2514 | * their contribution may appropriately discounted when they wake up. | |
2515 | */ | |
aff3e498 | 2516 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2517 | { |
f1b17280 | 2518 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2519 | u64 decays; |
2520 | ||
2521 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2522 | if (!decays && !force_update) |
9ee474f5 PT |
2523 | return; |
2524 | ||
2509940f AS |
2525 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2526 | unsigned long removed_load; | |
2527 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2528 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2529 | } | |
9ee474f5 | 2530 | |
aff3e498 PT |
2531 | if (decays) { |
2532 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2533 | decays); | |
2534 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2535 | cfs_rq->last_decay = now; | |
2536 | } | |
c566e8e9 PT |
2537 | |
2538 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2539 | } |
18bf2805 | 2540 | |
2dac754e PT |
2541 | /* Add the load generated by se into cfs_rq's child load-average */ |
2542 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2543 | struct sched_entity *se, |
2544 | int wakeup) | |
2dac754e | 2545 | { |
aff3e498 PT |
2546 | /* |
2547 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2548 | * migration we use a negative decay count to track the remote decays | |
2549 | * accumulated while sleeping. | |
a75cdaa9 AS |
2550 | * |
2551 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2552 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2553 | * constructed load_avg_contrib. | |
aff3e498 PT |
2554 | */ |
2555 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2556 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2557 | if (se->avg.decay_count) { |
2558 | /* | |
2559 | * In a wake-up migration we have to approximate the | |
2560 | * time sleeping. This is because we can't synchronize | |
2561 | * clock_task between the two cpus, and it is not | |
2562 | * guaranteed to be read-safe. Instead, we can | |
2563 | * approximate this using our carried decays, which are | |
2564 | * explicitly atomically readable. | |
2565 | */ | |
2566 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2567 | << 20; | |
2568 | update_entity_load_avg(se, 0); | |
2569 | /* Indicate that we're now synchronized and on-rq */ | |
2570 | se->avg.decay_count = 0; | |
2571 | } | |
9ee474f5 PT |
2572 | wakeup = 0; |
2573 | } else { | |
9390675a | 2574 | __synchronize_entity_decay(se); |
9ee474f5 PT |
2575 | } |
2576 | ||
aff3e498 PT |
2577 | /* migrated tasks did not contribute to our blocked load */ |
2578 | if (wakeup) { | |
9ee474f5 | 2579 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2580 | update_entity_load_avg(se, 0); |
2581 | } | |
9ee474f5 | 2582 | |
2dac754e | 2583 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2584 | /* we force update consideration on load-balancer moves */ |
2585 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2586 | } |
2587 | ||
9ee474f5 PT |
2588 | /* |
2589 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2590 | * transitioning to a blocked state we track its projected decay using | |
2591 | * blocked_load_avg. | |
2592 | */ | |
2dac754e | 2593 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2594 | struct sched_entity *se, |
2595 | int sleep) | |
2dac754e | 2596 | { |
9ee474f5 | 2597 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2598 | /* we force update consideration on load-balancer moves */ |
2599 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2600 | |
2dac754e | 2601 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2602 | if (sleep) { |
2603 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2604 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2605 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2606 | } |
642dbc39 VG |
2607 | |
2608 | /* | |
2609 | * Update the rq's load with the elapsed running time before entering | |
2610 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2611 | * be the only way to update the runnable statistic. | |
2612 | */ | |
2613 | void idle_enter_fair(struct rq *this_rq) | |
2614 | { | |
2615 | update_rq_runnable_avg(this_rq, 1); | |
2616 | } | |
2617 | ||
2618 | /* | |
2619 | * Update the rq's load with the elapsed idle time before a task is | |
2620 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2621 | * be the only way to update the runnable statistic. | |
2622 | */ | |
2623 | void idle_exit_fair(struct rq *this_rq) | |
2624 | { | |
2625 | update_rq_runnable_avg(this_rq, 0); | |
2626 | } | |
2627 | ||
6e83125c PZ |
2628 | static int idle_balance(struct rq *this_rq); |
2629 | ||
38033c37 PZ |
2630 | #else /* CONFIG_SMP */ |
2631 | ||
9ee474f5 PT |
2632 | static inline void update_entity_load_avg(struct sched_entity *se, |
2633 | int update_cfs_rq) {} | |
18bf2805 | 2634 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2635 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2636 | struct sched_entity *se, |
2637 | int wakeup) {} | |
2dac754e | 2638 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2639 | struct sched_entity *se, |
2640 | int sleep) {} | |
aff3e498 PT |
2641 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2642 | int force_update) {} | |
6e83125c PZ |
2643 | |
2644 | static inline int idle_balance(struct rq *rq) | |
2645 | { | |
2646 | return 0; | |
2647 | } | |
2648 | ||
38033c37 | 2649 | #endif /* CONFIG_SMP */ |
9d85f21c | 2650 | |
2396af69 | 2651 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2652 | { |
bf0f6f24 | 2653 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2654 | struct task_struct *tsk = NULL; |
2655 | ||
2656 | if (entity_is_task(se)) | |
2657 | tsk = task_of(se); | |
2658 | ||
41acab88 | 2659 | if (se->statistics.sleep_start) { |
78becc27 | 2660 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2661 | |
2662 | if ((s64)delta < 0) | |
2663 | delta = 0; | |
2664 | ||
41acab88 LDM |
2665 | if (unlikely(delta > se->statistics.sleep_max)) |
2666 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2667 | |
8c79a045 | 2668 | se->statistics.sleep_start = 0; |
41acab88 | 2669 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2670 | |
768d0c27 | 2671 | if (tsk) { |
e414314c | 2672 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2673 | trace_sched_stat_sleep(tsk, delta); |
2674 | } | |
bf0f6f24 | 2675 | } |
41acab88 | 2676 | if (se->statistics.block_start) { |
78becc27 | 2677 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2678 | |
2679 | if ((s64)delta < 0) | |
2680 | delta = 0; | |
2681 | ||
41acab88 LDM |
2682 | if (unlikely(delta > se->statistics.block_max)) |
2683 | se->statistics.block_max = delta; | |
bf0f6f24 | 2684 | |
8c79a045 | 2685 | se->statistics.block_start = 0; |
41acab88 | 2686 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2687 | |
e414314c | 2688 | if (tsk) { |
8f0dfc34 | 2689 | if (tsk->in_iowait) { |
41acab88 LDM |
2690 | se->statistics.iowait_sum += delta; |
2691 | se->statistics.iowait_count++; | |
768d0c27 | 2692 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2693 | } |
2694 | ||
b781a602 AV |
2695 | trace_sched_stat_blocked(tsk, delta); |
2696 | ||
e414314c PZ |
2697 | /* |
2698 | * Blocking time is in units of nanosecs, so shift by | |
2699 | * 20 to get a milliseconds-range estimation of the | |
2700 | * amount of time that the task spent sleeping: | |
2701 | */ | |
2702 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2703 | profile_hits(SLEEP_PROFILING, | |
2704 | (void *)get_wchan(tsk), | |
2705 | delta >> 20); | |
2706 | } | |
2707 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2708 | } |
bf0f6f24 IM |
2709 | } |
2710 | #endif | |
2711 | } | |
2712 | ||
ddc97297 PZ |
2713 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2714 | { | |
2715 | #ifdef CONFIG_SCHED_DEBUG | |
2716 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2717 | ||
2718 | if (d < 0) | |
2719 | d = -d; | |
2720 | ||
2721 | if (d > 3*sysctl_sched_latency) | |
2722 | schedstat_inc(cfs_rq, nr_spread_over); | |
2723 | #endif | |
2724 | } | |
2725 | ||
aeb73b04 PZ |
2726 | static void |
2727 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2728 | { | |
1af5f730 | 2729 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2730 | |
2cb8600e PZ |
2731 | /* |
2732 | * The 'current' period is already promised to the current tasks, | |
2733 | * however the extra weight of the new task will slow them down a | |
2734 | * little, place the new task so that it fits in the slot that | |
2735 | * stays open at the end. | |
2736 | */ | |
94dfb5e7 | 2737 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2738 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2739 | |
a2e7a7eb | 2740 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2741 | if (!initial) { |
a2e7a7eb | 2742 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2743 | |
a2e7a7eb MG |
2744 | /* |
2745 | * Halve their sleep time's effect, to allow | |
2746 | * for a gentler effect of sleepers: | |
2747 | */ | |
2748 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2749 | thresh >>= 1; | |
51e0304c | 2750 | |
a2e7a7eb | 2751 | vruntime -= thresh; |
aeb73b04 PZ |
2752 | } |
2753 | ||
b5d9d734 | 2754 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2755 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2756 | } |
2757 | ||
d3d9dc33 PT |
2758 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2759 | ||
bf0f6f24 | 2760 | static void |
88ec22d3 | 2761 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2762 | { |
88ec22d3 PZ |
2763 | /* |
2764 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2765 | * through calling update_curr(). |
88ec22d3 | 2766 | */ |
371fd7e7 | 2767 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2768 | se->vruntime += cfs_rq->min_vruntime; |
2769 | ||
bf0f6f24 | 2770 | /* |
a2a2d680 | 2771 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2772 | */ |
b7cc0896 | 2773 | update_curr(cfs_rq); |
f269ae04 | 2774 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2775 | account_entity_enqueue(cfs_rq, se); |
2776 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2777 | |
88ec22d3 | 2778 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2779 | place_entity(cfs_rq, se, 0); |
2396af69 | 2780 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2781 | } |
bf0f6f24 | 2782 | |
d2417e5a | 2783 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2784 | check_spread(cfs_rq, se); |
83b699ed SV |
2785 | if (se != cfs_rq->curr) |
2786 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2787 | se->on_rq = 1; |
3d4b47b4 | 2788 | |
d3d9dc33 | 2789 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2790 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2791 | check_enqueue_throttle(cfs_rq); |
2792 | } | |
bf0f6f24 IM |
2793 | } |
2794 | ||
2c13c919 | 2795 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2796 | { |
2c13c919 RR |
2797 | for_each_sched_entity(se) { |
2798 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 2799 | if (cfs_rq->last != se) |
2c13c919 | 2800 | break; |
f1044799 PZ |
2801 | |
2802 | cfs_rq->last = NULL; | |
2c13c919 RR |
2803 | } |
2804 | } | |
2002c695 | 2805 | |
2c13c919 RR |
2806 | static void __clear_buddies_next(struct sched_entity *se) |
2807 | { | |
2808 | for_each_sched_entity(se) { | |
2809 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 2810 | if (cfs_rq->next != se) |
2c13c919 | 2811 | break; |
f1044799 PZ |
2812 | |
2813 | cfs_rq->next = NULL; | |
2c13c919 | 2814 | } |
2002c695 PZ |
2815 | } |
2816 | ||
ac53db59 RR |
2817 | static void __clear_buddies_skip(struct sched_entity *se) |
2818 | { | |
2819 | for_each_sched_entity(se) { | |
2820 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 2821 | if (cfs_rq->skip != se) |
ac53db59 | 2822 | break; |
f1044799 PZ |
2823 | |
2824 | cfs_rq->skip = NULL; | |
ac53db59 RR |
2825 | } |
2826 | } | |
2827 | ||
a571bbea PZ |
2828 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2829 | { | |
2c13c919 RR |
2830 | if (cfs_rq->last == se) |
2831 | __clear_buddies_last(se); | |
2832 | ||
2833 | if (cfs_rq->next == se) | |
2834 | __clear_buddies_next(se); | |
ac53db59 RR |
2835 | |
2836 | if (cfs_rq->skip == se) | |
2837 | __clear_buddies_skip(se); | |
a571bbea PZ |
2838 | } |
2839 | ||
6c16a6dc | 2840 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2841 | |
bf0f6f24 | 2842 | static void |
371fd7e7 | 2843 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2844 | { |
a2a2d680 DA |
2845 | /* |
2846 | * Update run-time statistics of the 'current'. | |
2847 | */ | |
2848 | update_curr(cfs_rq); | |
17bc14b7 | 2849 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2850 | |
19b6a2e3 | 2851 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2852 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2853 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2854 | if (entity_is_task(se)) { |
2855 | struct task_struct *tsk = task_of(se); | |
2856 | ||
2857 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2858 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2859 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2860 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2861 | } |
db36cc7d | 2862 | #endif |
67e9fb2a PZ |
2863 | } |
2864 | ||
2002c695 | 2865 | clear_buddies(cfs_rq, se); |
4793241b | 2866 | |
83b699ed | 2867 | if (se != cfs_rq->curr) |
30cfdcfc | 2868 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2869 | se->on_rq = 0; |
30cfdcfc | 2870 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2871 | |
2872 | /* | |
2873 | * Normalize the entity after updating the min_vruntime because the | |
2874 | * update can refer to the ->curr item and we need to reflect this | |
2875 | * movement in our normalized position. | |
2876 | */ | |
371fd7e7 | 2877 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2878 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2879 | |
d8b4986d PT |
2880 | /* return excess runtime on last dequeue */ |
2881 | return_cfs_rq_runtime(cfs_rq); | |
2882 | ||
1e876231 | 2883 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2884 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2885 | } |
2886 | ||
2887 | /* | |
2888 | * Preempt the current task with a newly woken task if needed: | |
2889 | */ | |
7c92e54f | 2890 | static void |
2e09bf55 | 2891 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2892 | { |
11697830 | 2893 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2894 | struct sched_entity *se; |
2895 | s64 delta; | |
11697830 | 2896 | |
6d0f0ebd | 2897 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2898 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2899 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2900 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2901 | /* |
2902 | * The current task ran long enough, ensure it doesn't get | |
2903 | * re-elected due to buddy favours. | |
2904 | */ | |
2905 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2906 | return; |
2907 | } | |
2908 | ||
2909 | /* | |
2910 | * Ensure that a task that missed wakeup preemption by a | |
2911 | * narrow margin doesn't have to wait for a full slice. | |
2912 | * This also mitigates buddy induced latencies under load. | |
2913 | */ | |
f685ceac MG |
2914 | if (delta_exec < sysctl_sched_min_granularity) |
2915 | return; | |
2916 | ||
f4cfb33e WX |
2917 | se = __pick_first_entity(cfs_rq); |
2918 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2919 | |
f4cfb33e WX |
2920 | if (delta < 0) |
2921 | return; | |
d7d82944 | 2922 | |
f4cfb33e WX |
2923 | if (delta > ideal_runtime) |
2924 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2925 | } |
2926 | ||
83b699ed | 2927 | static void |
8494f412 | 2928 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2929 | { |
83b699ed SV |
2930 | /* 'current' is not kept within the tree. */ |
2931 | if (se->on_rq) { | |
2932 | /* | |
2933 | * Any task has to be enqueued before it get to execute on | |
2934 | * a CPU. So account for the time it spent waiting on the | |
2935 | * runqueue. | |
2936 | */ | |
2937 | update_stats_wait_end(cfs_rq, se); | |
2938 | __dequeue_entity(cfs_rq, se); | |
2939 | } | |
2940 | ||
79303e9e | 2941 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2942 | cfs_rq->curr = se; |
eba1ed4b IM |
2943 | #ifdef CONFIG_SCHEDSTATS |
2944 | /* | |
2945 | * Track our maximum slice length, if the CPU's load is at | |
2946 | * least twice that of our own weight (i.e. dont track it | |
2947 | * when there are only lesser-weight tasks around): | |
2948 | */ | |
495eca49 | 2949 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2950 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2951 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2952 | } | |
2953 | #endif | |
4a55b450 | 2954 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2955 | } |
2956 | ||
3f3a4904 PZ |
2957 | static int |
2958 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2959 | ||
ac53db59 RR |
2960 | /* |
2961 | * Pick the next process, keeping these things in mind, in this order: | |
2962 | * 1) keep things fair between processes/task groups | |
2963 | * 2) pick the "next" process, since someone really wants that to run | |
2964 | * 3) pick the "last" process, for cache locality | |
2965 | * 4) do not run the "skip" process, if something else is available | |
2966 | */ | |
678d5718 PZ |
2967 | static struct sched_entity * |
2968 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 2969 | { |
678d5718 PZ |
2970 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
2971 | struct sched_entity *se; | |
2972 | ||
2973 | /* | |
2974 | * If curr is set we have to see if its left of the leftmost entity | |
2975 | * still in the tree, provided there was anything in the tree at all. | |
2976 | */ | |
2977 | if (!left || (curr && entity_before(curr, left))) | |
2978 | left = curr; | |
2979 | ||
2980 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 2981 | |
ac53db59 RR |
2982 | /* |
2983 | * Avoid running the skip buddy, if running something else can | |
2984 | * be done without getting too unfair. | |
2985 | */ | |
2986 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
2987 | struct sched_entity *second; |
2988 | ||
2989 | if (se == curr) { | |
2990 | second = __pick_first_entity(cfs_rq); | |
2991 | } else { | |
2992 | second = __pick_next_entity(se); | |
2993 | if (!second || (curr && entity_before(curr, second))) | |
2994 | second = curr; | |
2995 | } | |
2996 | ||
ac53db59 RR |
2997 | if (second && wakeup_preempt_entity(second, left) < 1) |
2998 | se = second; | |
2999 | } | |
aa2ac252 | 3000 | |
f685ceac MG |
3001 | /* |
3002 | * Prefer last buddy, try to return the CPU to a preempted task. | |
3003 | */ | |
3004 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
3005 | se = cfs_rq->last; | |
3006 | ||
ac53db59 RR |
3007 | /* |
3008 | * Someone really wants this to run. If it's not unfair, run it. | |
3009 | */ | |
3010 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
3011 | se = cfs_rq->next; | |
3012 | ||
f685ceac | 3013 | clear_buddies(cfs_rq, se); |
4793241b PZ |
3014 | |
3015 | return se; | |
aa2ac252 PZ |
3016 | } |
3017 | ||
678d5718 | 3018 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 3019 | |
ab6cde26 | 3020 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
3021 | { |
3022 | /* | |
3023 | * If still on the runqueue then deactivate_task() | |
3024 | * was not called and update_curr() has to be done: | |
3025 | */ | |
3026 | if (prev->on_rq) | |
b7cc0896 | 3027 | update_curr(cfs_rq); |
bf0f6f24 | 3028 | |
d3d9dc33 PT |
3029 | /* throttle cfs_rqs exceeding runtime */ |
3030 | check_cfs_rq_runtime(cfs_rq); | |
3031 | ||
ddc97297 | 3032 | check_spread(cfs_rq, prev); |
30cfdcfc | 3033 | if (prev->on_rq) { |
5870db5b | 3034 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
3035 | /* Put 'current' back into the tree. */ |
3036 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 3037 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 3038 | update_entity_load_avg(prev, 1); |
30cfdcfc | 3039 | } |
429d43bc | 3040 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
3041 | } |
3042 | ||
8f4d37ec PZ |
3043 | static void |
3044 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 3045 | { |
bf0f6f24 | 3046 | /* |
30cfdcfc | 3047 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3048 | */ |
30cfdcfc | 3049 | update_curr(cfs_rq); |
bf0f6f24 | 3050 | |
9d85f21c PT |
3051 | /* |
3052 | * Ensure that runnable average is periodically updated. | |
3053 | */ | |
9ee474f5 | 3054 | update_entity_load_avg(curr, 1); |
aff3e498 | 3055 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 3056 | update_cfs_shares(cfs_rq); |
9d85f21c | 3057 | |
8f4d37ec PZ |
3058 | #ifdef CONFIG_SCHED_HRTICK |
3059 | /* | |
3060 | * queued ticks are scheduled to match the slice, so don't bother | |
3061 | * validating it and just reschedule. | |
3062 | */ | |
983ed7a6 HH |
3063 | if (queued) { |
3064 | resched_task(rq_of(cfs_rq)->curr); | |
3065 | return; | |
3066 | } | |
8f4d37ec PZ |
3067 | /* |
3068 | * don't let the period tick interfere with the hrtick preemption | |
3069 | */ | |
3070 | if (!sched_feat(DOUBLE_TICK) && | |
3071 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
3072 | return; | |
3073 | #endif | |
3074 | ||
2c2efaed | 3075 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 3076 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
3077 | } |
3078 | ||
ab84d31e PT |
3079 | |
3080 | /************************************************** | |
3081 | * CFS bandwidth control machinery | |
3082 | */ | |
3083 | ||
3084 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
3085 | |
3086 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 3087 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
3088 | |
3089 | static inline bool cfs_bandwidth_used(void) | |
3090 | { | |
c5905afb | 3091 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
3092 | } |
3093 | ||
1ee14e6c | 3094 | void cfs_bandwidth_usage_inc(void) |
029632fb | 3095 | { |
1ee14e6c BS |
3096 | static_key_slow_inc(&__cfs_bandwidth_used); |
3097 | } | |
3098 | ||
3099 | void cfs_bandwidth_usage_dec(void) | |
3100 | { | |
3101 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
3102 | } |
3103 | #else /* HAVE_JUMP_LABEL */ | |
3104 | static bool cfs_bandwidth_used(void) | |
3105 | { | |
3106 | return true; | |
3107 | } | |
3108 | ||
1ee14e6c BS |
3109 | void cfs_bandwidth_usage_inc(void) {} |
3110 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
3111 | #endif /* HAVE_JUMP_LABEL */ |
3112 | ||
ab84d31e PT |
3113 | /* |
3114 | * default period for cfs group bandwidth. | |
3115 | * default: 0.1s, units: nanoseconds | |
3116 | */ | |
3117 | static inline u64 default_cfs_period(void) | |
3118 | { | |
3119 | return 100000000ULL; | |
3120 | } | |
ec12cb7f PT |
3121 | |
3122 | static inline u64 sched_cfs_bandwidth_slice(void) | |
3123 | { | |
3124 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
3125 | } | |
3126 | ||
a9cf55b2 PT |
3127 | /* |
3128 | * Replenish runtime according to assigned quota and update expiration time. | |
3129 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
3130 | * additional synchronization around rq->lock. | |
3131 | * | |
3132 | * requires cfs_b->lock | |
3133 | */ | |
029632fb | 3134 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
3135 | { |
3136 | u64 now; | |
3137 | ||
3138 | if (cfs_b->quota == RUNTIME_INF) | |
3139 | return; | |
3140 | ||
3141 | now = sched_clock_cpu(smp_processor_id()); | |
3142 | cfs_b->runtime = cfs_b->quota; | |
3143 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
3144 | } | |
3145 | ||
029632fb PZ |
3146 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3147 | { | |
3148 | return &tg->cfs_bandwidth; | |
3149 | } | |
3150 | ||
f1b17280 PT |
3151 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
3152 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
3153 | { | |
3154 | if (unlikely(cfs_rq->throttle_count)) | |
3155 | return cfs_rq->throttled_clock_task; | |
3156 | ||
78becc27 | 3157 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
3158 | } |
3159 | ||
85dac906 PT |
3160 | /* returns 0 on failure to allocate runtime */ |
3161 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
3162 | { |
3163 | struct task_group *tg = cfs_rq->tg; | |
3164 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 3165 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
3166 | |
3167 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
3168 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
3169 | ||
3170 | raw_spin_lock(&cfs_b->lock); | |
3171 | if (cfs_b->quota == RUNTIME_INF) | |
3172 | amount = min_amount; | |
58088ad0 | 3173 | else { |
a9cf55b2 PT |
3174 | /* |
3175 | * If the bandwidth pool has become inactive, then at least one | |
3176 | * period must have elapsed since the last consumption. | |
3177 | * Refresh the global state and ensure bandwidth timer becomes | |
3178 | * active. | |
3179 | */ | |
3180 | if (!cfs_b->timer_active) { | |
3181 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 3182 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 3183 | } |
58088ad0 PT |
3184 | |
3185 | if (cfs_b->runtime > 0) { | |
3186 | amount = min(cfs_b->runtime, min_amount); | |
3187 | cfs_b->runtime -= amount; | |
3188 | cfs_b->idle = 0; | |
3189 | } | |
ec12cb7f | 3190 | } |
a9cf55b2 | 3191 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
3192 | raw_spin_unlock(&cfs_b->lock); |
3193 | ||
3194 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
3195 | /* |
3196 | * we may have advanced our local expiration to account for allowed | |
3197 | * spread between our sched_clock and the one on which runtime was | |
3198 | * issued. | |
3199 | */ | |
3200 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
3201 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
3202 | |
3203 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
3204 | } |
3205 | ||
a9cf55b2 PT |
3206 | /* |
3207 | * Note: This depends on the synchronization provided by sched_clock and the | |
3208 | * fact that rq->clock snapshots this value. | |
3209 | */ | |
3210 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 3211 | { |
a9cf55b2 | 3212 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
3213 | |
3214 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 3215 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
3216 | return; |
3217 | ||
a9cf55b2 PT |
3218 | if (cfs_rq->runtime_remaining < 0) |
3219 | return; | |
3220 | ||
3221 | /* | |
3222 | * If the local deadline has passed we have to consider the | |
3223 | * possibility that our sched_clock is 'fast' and the global deadline | |
3224 | * has not truly expired. | |
3225 | * | |
3226 | * Fortunately we can check determine whether this the case by checking | |
3227 | * whether the global deadline has advanced. | |
3228 | */ | |
3229 | ||
3230 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
3231 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
3232 | cfs_rq->runtime_expires += TICK_NSEC; | |
3233 | } else { | |
3234 | /* global deadline is ahead, expiration has passed */ | |
3235 | cfs_rq->runtime_remaining = 0; | |
3236 | } | |
3237 | } | |
3238 | ||
9dbdb155 | 3239 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
3240 | { |
3241 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3242 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3243 | expire_cfs_rq_runtime(cfs_rq); |
3244 | ||
3245 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3246 | return; |
3247 | ||
85dac906 PT |
3248 | /* |
3249 | * if we're unable to extend our runtime we resched so that the active | |
3250 | * hierarchy can be throttled | |
3251 | */ | |
3252 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
3253 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
3254 | } |
3255 | ||
6c16a6dc | 3256 | static __always_inline |
9dbdb155 | 3257 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 3258 | { |
56f570e5 | 3259 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3260 | return; |
3261 | ||
3262 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3263 | } | |
3264 | ||
85dac906 PT |
3265 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3266 | { | |
56f570e5 | 3267 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3268 | } |
3269 | ||
64660c86 PT |
3270 | /* check whether cfs_rq, or any parent, is throttled */ |
3271 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3272 | { | |
56f570e5 | 3273 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3274 | } |
3275 | ||
3276 | /* | |
3277 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3278 | * dest_cpu are members of a throttled hierarchy when performing group | |
3279 | * load-balance operations. | |
3280 | */ | |
3281 | static inline int throttled_lb_pair(struct task_group *tg, | |
3282 | int src_cpu, int dest_cpu) | |
3283 | { | |
3284 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3285 | ||
3286 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3287 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3288 | ||
3289 | return throttled_hierarchy(src_cfs_rq) || | |
3290 | throttled_hierarchy(dest_cfs_rq); | |
3291 | } | |
3292 | ||
3293 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3294 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3295 | { | |
3296 | struct rq *rq = data; | |
3297 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3298 | ||
3299 | cfs_rq->throttle_count--; | |
3300 | #ifdef CONFIG_SMP | |
3301 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 3302 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3303 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3304 | cfs_rq->throttled_clock_task; |
64660c86 PT |
3305 | } |
3306 | #endif | |
3307 | ||
3308 | return 0; | |
3309 | } | |
3310 | ||
3311 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3312 | { | |
3313 | struct rq *rq = data; | |
3314 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3315 | ||
82958366 PT |
3316 | /* group is entering throttled state, stop time */ |
3317 | if (!cfs_rq->throttle_count) | |
78becc27 | 3318 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3319 | cfs_rq->throttle_count++; |
3320 | ||
3321 | return 0; | |
3322 | } | |
3323 | ||
d3d9dc33 | 3324 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3325 | { |
3326 | struct rq *rq = rq_of(cfs_rq); | |
3327 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3328 | struct sched_entity *se; | |
3329 | long task_delta, dequeue = 1; | |
3330 | ||
3331 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3332 | ||
f1b17280 | 3333 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3334 | rcu_read_lock(); |
3335 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3336 | rcu_read_unlock(); | |
85dac906 PT |
3337 | |
3338 | task_delta = cfs_rq->h_nr_running; | |
3339 | for_each_sched_entity(se) { | |
3340 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3341 | /* throttled entity or throttle-on-deactivate */ | |
3342 | if (!se->on_rq) | |
3343 | break; | |
3344 | ||
3345 | if (dequeue) | |
3346 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3347 | qcfs_rq->h_nr_running -= task_delta; | |
3348 | ||
3349 | if (qcfs_rq->load.weight) | |
3350 | dequeue = 0; | |
3351 | } | |
3352 | ||
3353 | if (!se) | |
72465447 | 3354 | sub_nr_running(rq, task_delta); |
85dac906 PT |
3355 | |
3356 | cfs_rq->throttled = 1; | |
78becc27 | 3357 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
3358 | raw_spin_lock(&cfs_b->lock); |
3359 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
f9f9ffc2 BS |
3360 | if (!cfs_b->timer_active) |
3361 | __start_cfs_bandwidth(cfs_b); | |
85dac906 PT |
3362 | raw_spin_unlock(&cfs_b->lock); |
3363 | } | |
3364 | ||
029632fb | 3365 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3366 | { |
3367 | struct rq *rq = rq_of(cfs_rq); | |
3368 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3369 | struct sched_entity *se; | |
3370 | int enqueue = 1; | |
3371 | long task_delta; | |
3372 | ||
22b958d8 | 3373 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3374 | |
3375 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3376 | |
3377 | update_rq_clock(rq); | |
3378 | ||
671fd9da | 3379 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3380 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3381 | list_del_rcu(&cfs_rq->throttled_list); |
3382 | raw_spin_unlock(&cfs_b->lock); | |
3383 | ||
64660c86 PT |
3384 | /* update hierarchical throttle state */ |
3385 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3386 | ||
671fd9da PT |
3387 | if (!cfs_rq->load.weight) |
3388 | return; | |
3389 | ||
3390 | task_delta = cfs_rq->h_nr_running; | |
3391 | for_each_sched_entity(se) { | |
3392 | if (se->on_rq) | |
3393 | enqueue = 0; | |
3394 | ||
3395 | cfs_rq = cfs_rq_of(se); | |
3396 | if (enqueue) | |
3397 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3398 | cfs_rq->h_nr_running += task_delta; | |
3399 | ||
3400 | if (cfs_rq_throttled(cfs_rq)) | |
3401 | break; | |
3402 | } | |
3403 | ||
3404 | if (!se) | |
72465447 | 3405 | add_nr_running(rq, task_delta); |
671fd9da PT |
3406 | |
3407 | /* determine whether we need to wake up potentially idle cpu */ | |
3408 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
3409 | resched_task(rq->curr); | |
3410 | } | |
3411 | ||
3412 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
3413 | u64 remaining, u64 expires) | |
3414 | { | |
3415 | struct cfs_rq *cfs_rq; | |
3416 | u64 runtime = remaining; | |
3417 | ||
3418 | rcu_read_lock(); | |
3419 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
3420 | throttled_list) { | |
3421 | struct rq *rq = rq_of(cfs_rq); | |
3422 | ||
3423 | raw_spin_lock(&rq->lock); | |
3424 | if (!cfs_rq_throttled(cfs_rq)) | |
3425 | goto next; | |
3426 | ||
3427 | runtime = -cfs_rq->runtime_remaining + 1; | |
3428 | if (runtime > remaining) | |
3429 | runtime = remaining; | |
3430 | remaining -= runtime; | |
3431 | ||
3432 | cfs_rq->runtime_remaining += runtime; | |
3433 | cfs_rq->runtime_expires = expires; | |
3434 | ||
3435 | /* we check whether we're throttled above */ | |
3436 | if (cfs_rq->runtime_remaining > 0) | |
3437 | unthrottle_cfs_rq(cfs_rq); | |
3438 | ||
3439 | next: | |
3440 | raw_spin_unlock(&rq->lock); | |
3441 | ||
3442 | if (!remaining) | |
3443 | break; | |
3444 | } | |
3445 | rcu_read_unlock(); | |
3446 | ||
3447 | return remaining; | |
3448 | } | |
3449 | ||
58088ad0 PT |
3450 | /* |
3451 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
3452 | * cfs_rqs as appropriate. If there has been no activity within the last | |
3453 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
3454 | * used to track this state. | |
3455 | */ | |
3456 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3457 | { | |
671fd9da PT |
3458 | u64 runtime, runtime_expires; |
3459 | int idle = 1, throttled; | |
58088ad0 PT |
3460 | |
3461 | raw_spin_lock(&cfs_b->lock); | |
3462 | /* no need to continue the timer with no bandwidth constraint */ | |
3463 | if (cfs_b->quota == RUNTIME_INF) | |
3464 | goto out_unlock; | |
3465 | ||
671fd9da PT |
3466 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3467 | /* idle depends on !throttled (for the case of a large deficit) */ | |
3468 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 3469 | cfs_b->nr_periods += overrun; |
671fd9da | 3470 | |
a9cf55b2 PT |
3471 | /* if we're going inactive then everything else can be deferred */ |
3472 | if (idle) | |
3473 | goto out_unlock; | |
3474 | ||
927b54fc BS |
3475 | /* |
3476 | * if we have relooped after returning idle once, we need to update our | |
3477 | * status as actually running, so that other cpus doing | |
3478 | * __start_cfs_bandwidth will stop trying to cancel us. | |
3479 | */ | |
3480 | cfs_b->timer_active = 1; | |
3481 | ||
a9cf55b2 PT |
3482 | __refill_cfs_bandwidth_runtime(cfs_b); |
3483 | ||
671fd9da PT |
3484 | if (!throttled) { |
3485 | /* mark as potentially idle for the upcoming period */ | |
3486 | cfs_b->idle = 1; | |
3487 | goto out_unlock; | |
3488 | } | |
3489 | ||
e8da1b18 NR |
3490 | /* account preceding periods in which throttling occurred */ |
3491 | cfs_b->nr_throttled += overrun; | |
3492 | ||
671fd9da PT |
3493 | /* |
3494 | * There are throttled entities so we must first use the new bandwidth | |
3495 | * to unthrottle them before making it generally available. This | |
3496 | * ensures that all existing debts will be paid before a new cfs_rq is | |
3497 | * allowed to run. | |
3498 | */ | |
3499 | runtime = cfs_b->runtime; | |
3500 | runtime_expires = cfs_b->runtime_expires; | |
3501 | cfs_b->runtime = 0; | |
3502 | ||
3503 | /* | |
3504 | * This check is repeated as we are holding onto the new bandwidth | |
3505 | * while we unthrottle. This can potentially race with an unthrottled | |
3506 | * group trying to acquire new bandwidth from the global pool. | |
3507 | */ | |
3508 | while (throttled && runtime > 0) { | |
3509 | raw_spin_unlock(&cfs_b->lock); | |
3510 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3511 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3512 | runtime_expires); | |
3513 | raw_spin_lock(&cfs_b->lock); | |
3514 | ||
3515 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
3516 | } | |
58088ad0 | 3517 | |
671fd9da PT |
3518 | /* return (any) remaining runtime */ |
3519 | cfs_b->runtime = runtime; | |
3520 | /* | |
3521 | * While we are ensured activity in the period following an | |
3522 | * unthrottle, this also covers the case in which the new bandwidth is | |
3523 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3524 | * timer to remain active while there are any throttled entities.) | |
3525 | */ | |
3526 | cfs_b->idle = 0; | |
58088ad0 PT |
3527 | out_unlock: |
3528 | if (idle) | |
3529 | cfs_b->timer_active = 0; | |
3530 | raw_spin_unlock(&cfs_b->lock); | |
3531 | ||
3532 | return idle; | |
3533 | } | |
d3d9dc33 | 3534 | |
d8b4986d PT |
3535 | /* a cfs_rq won't donate quota below this amount */ |
3536 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3537 | /* minimum remaining period time to redistribute slack quota */ | |
3538 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3539 | /* how long we wait to gather additional slack before distributing */ | |
3540 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3541 | ||
db06e78c BS |
3542 | /* |
3543 | * Are we near the end of the current quota period? | |
3544 | * | |
3545 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
3546 | * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of | |
3547 | * migrate_hrtimers, base is never cleared, so we are fine. | |
3548 | */ | |
d8b4986d PT |
3549 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
3550 | { | |
3551 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3552 | u64 remaining; | |
3553 | ||
3554 | /* if the call-back is running a quota refresh is already occurring */ | |
3555 | if (hrtimer_callback_running(refresh_timer)) | |
3556 | return 1; | |
3557 | ||
3558 | /* is a quota refresh about to occur? */ | |
3559 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3560 | if (remaining < min_expire) | |
3561 | return 1; | |
3562 | ||
3563 | return 0; | |
3564 | } | |
3565 | ||
3566 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3567 | { | |
3568 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3569 | ||
3570 | /* if there's a quota refresh soon don't bother with slack */ | |
3571 | if (runtime_refresh_within(cfs_b, min_left)) | |
3572 | return; | |
3573 | ||
3574 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3575 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3576 | } | |
3577 | ||
3578 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3579 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3580 | { | |
3581 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3582 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3583 | ||
3584 | if (slack_runtime <= 0) | |
3585 | return; | |
3586 | ||
3587 | raw_spin_lock(&cfs_b->lock); | |
3588 | if (cfs_b->quota != RUNTIME_INF && | |
3589 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3590 | cfs_b->runtime += slack_runtime; | |
3591 | ||
3592 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3593 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3594 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3595 | start_cfs_slack_bandwidth(cfs_b); | |
3596 | } | |
3597 | raw_spin_unlock(&cfs_b->lock); | |
3598 | ||
3599 | /* even if it's not valid for return we don't want to try again */ | |
3600 | cfs_rq->runtime_remaining -= slack_runtime; | |
3601 | } | |
3602 | ||
3603 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3604 | { | |
56f570e5 PT |
3605 | if (!cfs_bandwidth_used()) |
3606 | return; | |
3607 | ||
fccfdc6f | 3608 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3609 | return; |
3610 | ||
3611 | __return_cfs_rq_runtime(cfs_rq); | |
3612 | } | |
3613 | ||
3614 | /* | |
3615 | * This is done with a timer (instead of inline with bandwidth return) since | |
3616 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3617 | */ | |
3618 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3619 | { | |
3620 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3621 | u64 expires; | |
3622 | ||
3623 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
3624 | raw_spin_lock(&cfs_b->lock); |
3625 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
3626 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 3627 | return; |
db06e78c | 3628 | } |
d8b4986d | 3629 | |
d8b4986d PT |
3630 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { |
3631 | runtime = cfs_b->runtime; | |
3632 | cfs_b->runtime = 0; | |
3633 | } | |
3634 | expires = cfs_b->runtime_expires; | |
3635 | raw_spin_unlock(&cfs_b->lock); | |
3636 | ||
3637 | if (!runtime) | |
3638 | return; | |
3639 | ||
3640 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3641 | ||
3642 | raw_spin_lock(&cfs_b->lock); | |
3643 | if (expires == cfs_b->runtime_expires) | |
3644 | cfs_b->runtime = runtime; | |
3645 | raw_spin_unlock(&cfs_b->lock); | |
3646 | } | |
3647 | ||
d3d9dc33 PT |
3648 | /* |
3649 | * When a group wakes up we want to make sure that its quota is not already | |
3650 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3651 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3652 | */ | |
3653 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3654 | { | |
56f570e5 PT |
3655 | if (!cfs_bandwidth_used()) |
3656 | return; | |
3657 | ||
d3d9dc33 PT |
3658 | /* an active group must be handled by the update_curr()->put() path */ |
3659 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3660 | return; | |
3661 | ||
3662 | /* ensure the group is not already throttled */ | |
3663 | if (cfs_rq_throttled(cfs_rq)) | |
3664 | return; | |
3665 | ||
3666 | /* update runtime allocation */ | |
3667 | account_cfs_rq_runtime(cfs_rq, 0); | |
3668 | if (cfs_rq->runtime_remaining <= 0) | |
3669 | throttle_cfs_rq(cfs_rq); | |
3670 | } | |
3671 | ||
3672 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
678d5718 | 3673 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 3674 | { |
56f570e5 | 3675 | if (!cfs_bandwidth_used()) |
678d5718 | 3676 | return false; |
56f570e5 | 3677 | |
d3d9dc33 | 3678 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 3679 | return false; |
d3d9dc33 PT |
3680 | |
3681 | /* | |
3682 | * it's possible for a throttled entity to be forced into a running | |
3683 | * state (e.g. set_curr_task), in this case we're finished. | |
3684 | */ | |
3685 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 3686 | return true; |
d3d9dc33 PT |
3687 | |
3688 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 3689 | return true; |
d3d9dc33 | 3690 | } |
029632fb | 3691 | |
029632fb PZ |
3692 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3693 | { | |
3694 | struct cfs_bandwidth *cfs_b = | |
3695 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3696 | do_sched_cfs_slack_timer(cfs_b); | |
3697 | ||
3698 | return HRTIMER_NORESTART; | |
3699 | } | |
3700 | ||
3701 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3702 | { | |
3703 | struct cfs_bandwidth *cfs_b = | |
3704 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3705 | ktime_t now; | |
3706 | int overrun; | |
3707 | int idle = 0; | |
3708 | ||
3709 | for (;;) { | |
3710 | now = hrtimer_cb_get_time(timer); | |
3711 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3712 | ||
3713 | if (!overrun) | |
3714 | break; | |
3715 | ||
3716 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3717 | } | |
3718 | ||
3719 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3720 | } | |
3721 | ||
3722 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3723 | { | |
3724 | raw_spin_lock_init(&cfs_b->lock); | |
3725 | cfs_b->runtime = 0; | |
3726 | cfs_b->quota = RUNTIME_INF; | |
3727 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3728 | ||
3729 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3730 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3731 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3732 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3733 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3734 | } | |
3735 | ||
3736 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3737 | { | |
3738 | cfs_rq->runtime_enabled = 0; | |
3739 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3740 | } | |
3741 | ||
3742 | /* requires cfs_b->lock, may release to reprogram timer */ | |
3743 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3744 | { | |
3745 | /* | |
3746 | * The timer may be active because we're trying to set a new bandwidth | |
3747 | * period or because we're racing with the tear-down path | |
3748 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3749 | * terminates). In either case we ensure that it's re-programmed | |
3750 | */ | |
927b54fc BS |
3751 | while (unlikely(hrtimer_active(&cfs_b->period_timer)) && |
3752 | hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { | |
3753 | /* bounce the lock to allow do_sched_cfs_period_timer to run */ | |
029632fb | 3754 | raw_spin_unlock(&cfs_b->lock); |
927b54fc | 3755 | cpu_relax(); |
029632fb PZ |
3756 | raw_spin_lock(&cfs_b->lock); |
3757 | /* if someone else restarted the timer then we're done */ | |
3758 | if (cfs_b->timer_active) | |
3759 | return; | |
3760 | } | |
3761 | ||
3762 | cfs_b->timer_active = 1; | |
3763 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3764 | } | |
3765 | ||
3766 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3767 | { | |
3768 | hrtimer_cancel(&cfs_b->period_timer); | |
3769 | hrtimer_cancel(&cfs_b->slack_timer); | |
3770 | } | |
3771 | ||
38dc3348 | 3772 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3773 | { |
3774 | struct cfs_rq *cfs_rq; | |
3775 | ||
3776 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3777 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3778 | ||
3779 | if (!cfs_rq->runtime_enabled) | |
3780 | continue; | |
3781 | ||
3782 | /* | |
3783 | * clock_task is not advancing so we just need to make sure | |
3784 | * there's some valid quota amount | |
3785 | */ | |
3786 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3787 | if (cfs_rq_throttled(cfs_rq)) | |
3788 | unthrottle_cfs_rq(cfs_rq); | |
3789 | } | |
3790 | } | |
3791 | ||
3792 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3793 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3794 | { | |
78becc27 | 3795 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3796 | } |
3797 | ||
9dbdb155 | 3798 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 3799 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 3800 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
6c16a6dc | 3801 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3802 | |
3803 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3804 | { | |
3805 | return 0; | |
3806 | } | |
64660c86 PT |
3807 | |
3808 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3809 | { | |
3810 | return 0; | |
3811 | } | |
3812 | ||
3813 | static inline int throttled_lb_pair(struct task_group *tg, | |
3814 | int src_cpu, int dest_cpu) | |
3815 | { | |
3816 | return 0; | |
3817 | } | |
029632fb PZ |
3818 | |
3819 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3820 | ||
3821 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3822 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3823 | #endif |
3824 | ||
029632fb PZ |
3825 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3826 | { | |
3827 | return NULL; | |
3828 | } | |
3829 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3830 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3831 | |
3832 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3833 | ||
bf0f6f24 IM |
3834 | /************************************************** |
3835 | * CFS operations on tasks: | |
3836 | */ | |
3837 | ||
8f4d37ec PZ |
3838 | #ifdef CONFIG_SCHED_HRTICK |
3839 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3840 | { | |
8f4d37ec PZ |
3841 | struct sched_entity *se = &p->se; |
3842 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3843 | ||
3844 | WARN_ON(task_rq(p) != rq); | |
3845 | ||
b39e66ea | 3846 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3847 | u64 slice = sched_slice(cfs_rq, se); |
3848 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3849 | s64 delta = slice - ran; | |
3850 | ||
3851 | if (delta < 0) { | |
3852 | if (rq->curr == p) | |
3853 | resched_task(p); | |
3854 | return; | |
3855 | } | |
3856 | ||
3857 | /* | |
3858 | * Don't schedule slices shorter than 10000ns, that just | |
3859 | * doesn't make sense. Rely on vruntime for fairness. | |
3860 | */ | |
31656519 | 3861 | if (rq->curr != p) |
157124c1 | 3862 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3863 | |
31656519 | 3864 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3865 | } |
3866 | } | |
a4c2f00f PZ |
3867 | |
3868 | /* | |
3869 | * called from enqueue/dequeue and updates the hrtick when the | |
3870 | * current task is from our class and nr_running is low enough | |
3871 | * to matter. | |
3872 | */ | |
3873 | static void hrtick_update(struct rq *rq) | |
3874 | { | |
3875 | struct task_struct *curr = rq->curr; | |
3876 | ||
b39e66ea | 3877 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3878 | return; |
3879 | ||
3880 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3881 | hrtick_start_fair(rq, curr); | |
3882 | } | |
55e12e5e | 3883 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3884 | static inline void |
3885 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3886 | { | |
3887 | } | |
a4c2f00f PZ |
3888 | |
3889 | static inline void hrtick_update(struct rq *rq) | |
3890 | { | |
3891 | } | |
8f4d37ec PZ |
3892 | #endif |
3893 | ||
bf0f6f24 IM |
3894 | /* |
3895 | * The enqueue_task method is called before nr_running is | |
3896 | * increased. Here we update the fair scheduling stats and | |
3897 | * then put the task into the rbtree: | |
3898 | */ | |
ea87bb78 | 3899 | static void |
371fd7e7 | 3900 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3901 | { |
3902 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3903 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3904 | |
3905 | for_each_sched_entity(se) { | |
62fb1851 | 3906 | if (se->on_rq) |
bf0f6f24 IM |
3907 | break; |
3908 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3909 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3910 | |
3911 | /* | |
3912 | * end evaluation on encountering a throttled cfs_rq | |
3913 | * | |
3914 | * note: in the case of encountering a throttled cfs_rq we will | |
3915 | * post the final h_nr_running increment below. | |
3916 | */ | |
3917 | if (cfs_rq_throttled(cfs_rq)) | |
3918 | break; | |
953bfcd1 | 3919 | cfs_rq->h_nr_running++; |
85dac906 | 3920 | |
88ec22d3 | 3921 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3922 | } |
8f4d37ec | 3923 | |
2069dd75 | 3924 | for_each_sched_entity(se) { |
0f317143 | 3925 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3926 | cfs_rq->h_nr_running++; |
2069dd75 | 3927 | |
85dac906 PT |
3928 | if (cfs_rq_throttled(cfs_rq)) |
3929 | break; | |
3930 | ||
17bc14b7 | 3931 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3932 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3933 | } |
3934 | ||
18bf2805 BS |
3935 | if (!se) { |
3936 | update_rq_runnable_avg(rq, rq->nr_running); | |
72465447 | 3937 | add_nr_running(rq, 1); |
18bf2805 | 3938 | } |
a4c2f00f | 3939 | hrtick_update(rq); |
bf0f6f24 IM |
3940 | } |
3941 | ||
2f36825b VP |
3942 | static void set_next_buddy(struct sched_entity *se); |
3943 | ||
bf0f6f24 IM |
3944 | /* |
3945 | * The dequeue_task method is called before nr_running is | |
3946 | * decreased. We remove the task from the rbtree and | |
3947 | * update the fair scheduling stats: | |
3948 | */ | |
371fd7e7 | 3949 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3950 | { |
3951 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3952 | struct sched_entity *se = &p->se; |
2f36825b | 3953 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3954 | |
3955 | for_each_sched_entity(se) { | |
3956 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3957 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3958 | |
3959 | /* | |
3960 | * end evaluation on encountering a throttled cfs_rq | |
3961 | * | |
3962 | * note: in the case of encountering a throttled cfs_rq we will | |
3963 | * post the final h_nr_running decrement below. | |
3964 | */ | |
3965 | if (cfs_rq_throttled(cfs_rq)) | |
3966 | break; | |
953bfcd1 | 3967 | cfs_rq->h_nr_running--; |
2069dd75 | 3968 | |
bf0f6f24 | 3969 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3970 | if (cfs_rq->load.weight) { |
3971 | /* | |
3972 | * Bias pick_next to pick a task from this cfs_rq, as | |
3973 | * p is sleeping when it is within its sched_slice. | |
3974 | */ | |
3975 | if (task_sleep && parent_entity(se)) | |
3976 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3977 | |
3978 | /* avoid re-evaluating load for this entity */ | |
3979 | se = parent_entity(se); | |
bf0f6f24 | 3980 | break; |
2f36825b | 3981 | } |
371fd7e7 | 3982 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3983 | } |
8f4d37ec | 3984 | |
2069dd75 | 3985 | for_each_sched_entity(se) { |
0f317143 | 3986 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3987 | cfs_rq->h_nr_running--; |
2069dd75 | 3988 | |
85dac906 PT |
3989 | if (cfs_rq_throttled(cfs_rq)) |
3990 | break; | |
3991 | ||
17bc14b7 | 3992 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3993 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3994 | } |
3995 | ||
18bf2805 | 3996 | if (!se) { |
72465447 | 3997 | sub_nr_running(rq, 1); |
18bf2805 BS |
3998 | update_rq_runnable_avg(rq, 1); |
3999 | } | |
a4c2f00f | 4000 | hrtick_update(rq); |
bf0f6f24 IM |
4001 | } |
4002 | ||
e7693a36 | 4003 | #ifdef CONFIG_SMP |
029632fb PZ |
4004 | /* Used instead of source_load when we know the type == 0 */ |
4005 | static unsigned long weighted_cpuload(const int cpu) | |
4006 | { | |
b92486cb | 4007 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
4008 | } |
4009 | ||
4010 | /* | |
4011 | * Return a low guess at the load of a migration-source cpu weighted | |
4012 | * according to the scheduling class and "nice" value. | |
4013 | * | |
4014 | * We want to under-estimate the load of migration sources, to | |
4015 | * balance conservatively. | |
4016 | */ | |
4017 | static unsigned long source_load(int cpu, int type) | |
4018 | { | |
4019 | struct rq *rq = cpu_rq(cpu); | |
4020 | unsigned long total = weighted_cpuload(cpu); | |
4021 | ||
4022 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4023 | return total; | |
4024 | ||
4025 | return min(rq->cpu_load[type-1], total); | |
4026 | } | |
4027 | ||
4028 | /* | |
4029 | * Return a high guess at the load of a migration-target cpu weighted | |
4030 | * according to the scheduling class and "nice" value. | |
4031 | */ | |
4032 | static unsigned long target_load(int cpu, int type) | |
4033 | { | |
4034 | struct rq *rq = cpu_rq(cpu); | |
4035 | unsigned long total = weighted_cpuload(cpu); | |
4036 | ||
4037 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4038 | return total; | |
4039 | ||
4040 | return max(rq->cpu_load[type-1], total); | |
4041 | } | |
4042 | ||
4043 | static unsigned long power_of(int cpu) | |
4044 | { | |
4045 | return cpu_rq(cpu)->cpu_power; | |
4046 | } | |
4047 | ||
4048 | static unsigned long cpu_avg_load_per_task(int cpu) | |
4049 | { | |
4050 | struct rq *rq = cpu_rq(cpu); | |
4051 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 4052 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
4053 | |
4054 | if (nr_running) | |
b92486cb | 4055 | return load_avg / nr_running; |
029632fb PZ |
4056 | |
4057 | return 0; | |
4058 | } | |
4059 | ||
62470419 MW |
4060 | static void record_wakee(struct task_struct *p) |
4061 | { | |
4062 | /* | |
4063 | * Rough decay (wiping) for cost saving, don't worry | |
4064 | * about the boundary, really active task won't care | |
4065 | * about the loss. | |
4066 | */ | |
4067 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
4068 | current->wakee_flips = 0; | |
4069 | current->wakee_flip_decay_ts = jiffies; | |
4070 | } | |
4071 | ||
4072 | if (current->last_wakee != p) { | |
4073 | current->last_wakee = p; | |
4074 | current->wakee_flips++; | |
4075 | } | |
4076 | } | |
098fb9db | 4077 | |
74f8e4b2 | 4078 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
4079 | { |
4080 | struct sched_entity *se = &p->se; | |
4081 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
4082 | u64 min_vruntime; |
4083 | ||
4084 | #ifndef CONFIG_64BIT | |
4085 | u64 min_vruntime_copy; | |
88ec22d3 | 4086 | |
3fe1698b PZ |
4087 | do { |
4088 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
4089 | smp_rmb(); | |
4090 | min_vruntime = cfs_rq->min_vruntime; | |
4091 | } while (min_vruntime != min_vruntime_copy); | |
4092 | #else | |
4093 | min_vruntime = cfs_rq->min_vruntime; | |
4094 | #endif | |
88ec22d3 | 4095 | |
3fe1698b | 4096 | se->vruntime -= min_vruntime; |
62470419 | 4097 | record_wakee(p); |
88ec22d3 PZ |
4098 | } |
4099 | ||
bb3469ac | 4100 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
4101 | /* |
4102 | * effective_load() calculates the load change as seen from the root_task_group | |
4103 | * | |
4104 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
4105 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
4106 | * can calculate the shift in shares. | |
cf5f0acf PZ |
4107 | * |
4108 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
4109 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
4110 | * total group weight. | |
4111 | * | |
4112 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
4113 | * distribution (s_i) using: | |
4114 | * | |
4115 | * s_i = rw_i / \Sum rw_j (1) | |
4116 | * | |
4117 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
4118 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
4119 | * shares distribution (s_i): | |
4120 | * | |
4121 | * rw_i = { 2, 4, 1, 0 } | |
4122 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
4123 | * | |
4124 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
4125 | * task used to run on and the CPU the waker is running on), we need to | |
4126 | * compute the effect of waking a task on either CPU and, in case of a sync | |
4127 | * wakeup, compute the effect of the current task going to sleep. | |
4128 | * | |
4129 | * So for a change of @wl to the local @cpu with an overall group weight change | |
4130 | * of @wl we can compute the new shares distribution (s'_i) using: | |
4131 | * | |
4132 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
4133 | * | |
4134 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
4135 | * differences in waking a task to CPU 0. The additional task changes the | |
4136 | * weight and shares distributions like: | |
4137 | * | |
4138 | * rw'_i = { 3, 4, 1, 0 } | |
4139 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
4140 | * | |
4141 | * We can then compute the difference in effective weight by using: | |
4142 | * | |
4143 | * dw_i = S * (s'_i - s_i) (3) | |
4144 | * | |
4145 | * Where 'S' is the group weight as seen by its parent. | |
4146 | * | |
4147 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
4148 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
4149 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 4150 | */ |
2069dd75 | 4151 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 4152 | { |
4be9daaa | 4153 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 4154 | |
9722c2da | 4155 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
4156 | return wl; |
4157 | ||
4be9daaa | 4158 | for_each_sched_entity(se) { |
cf5f0acf | 4159 | long w, W; |
4be9daaa | 4160 | |
977dda7c | 4161 | tg = se->my_q->tg; |
bb3469ac | 4162 | |
cf5f0acf PZ |
4163 | /* |
4164 | * W = @wg + \Sum rw_j | |
4165 | */ | |
4166 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 4167 | |
cf5f0acf PZ |
4168 | /* |
4169 | * w = rw_i + @wl | |
4170 | */ | |
4171 | w = se->my_q->load.weight + wl; | |
940959e9 | 4172 | |
cf5f0acf PZ |
4173 | /* |
4174 | * wl = S * s'_i; see (2) | |
4175 | */ | |
4176 | if (W > 0 && w < W) | |
4177 | wl = (w * tg->shares) / W; | |
977dda7c PT |
4178 | else |
4179 | wl = tg->shares; | |
940959e9 | 4180 | |
cf5f0acf PZ |
4181 | /* |
4182 | * Per the above, wl is the new se->load.weight value; since | |
4183 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
4184 | * calc_cfs_shares(). | |
4185 | */ | |
977dda7c PT |
4186 | if (wl < MIN_SHARES) |
4187 | wl = MIN_SHARES; | |
cf5f0acf PZ |
4188 | |
4189 | /* | |
4190 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
4191 | */ | |
977dda7c | 4192 | wl -= se->load.weight; |
cf5f0acf PZ |
4193 | |
4194 | /* | |
4195 | * Recursively apply this logic to all parent groups to compute | |
4196 | * the final effective load change on the root group. Since | |
4197 | * only the @tg group gets extra weight, all parent groups can | |
4198 | * only redistribute existing shares. @wl is the shift in shares | |
4199 | * resulting from this level per the above. | |
4200 | */ | |
4be9daaa | 4201 | wg = 0; |
4be9daaa | 4202 | } |
bb3469ac | 4203 | |
4be9daaa | 4204 | return wl; |
bb3469ac PZ |
4205 | } |
4206 | #else | |
4be9daaa | 4207 | |
58d081b5 | 4208 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 4209 | { |
83378269 | 4210 | return wl; |
bb3469ac | 4211 | } |
4be9daaa | 4212 | |
bb3469ac PZ |
4213 | #endif |
4214 | ||
62470419 MW |
4215 | static int wake_wide(struct task_struct *p) |
4216 | { | |
7d9ffa89 | 4217 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
4218 | |
4219 | /* | |
4220 | * Yeah, it's the switching-frequency, could means many wakee or | |
4221 | * rapidly switch, use factor here will just help to automatically | |
4222 | * adjust the loose-degree, so bigger node will lead to more pull. | |
4223 | */ | |
4224 | if (p->wakee_flips > factor) { | |
4225 | /* | |
4226 | * wakee is somewhat hot, it needs certain amount of cpu | |
4227 | * resource, so if waker is far more hot, prefer to leave | |
4228 | * it alone. | |
4229 | */ | |
4230 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
4231 | return 1; | |
4232 | } | |
4233 | ||
4234 | return 0; | |
4235 | } | |
4236 | ||
c88d5910 | 4237 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 4238 | { |
e37b6a7b | 4239 | s64 this_load, load; |
c88d5910 | 4240 | int idx, this_cpu, prev_cpu; |
098fb9db | 4241 | unsigned long tl_per_task; |
c88d5910 | 4242 | struct task_group *tg; |
83378269 | 4243 | unsigned long weight; |
b3137bc8 | 4244 | int balanced; |
098fb9db | 4245 | |
62470419 MW |
4246 | /* |
4247 | * If we wake multiple tasks be careful to not bounce | |
4248 | * ourselves around too much. | |
4249 | */ | |
4250 | if (wake_wide(p)) | |
4251 | return 0; | |
4252 | ||
c88d5910 PZ |
4253 | idx = sd->wake_idx; |
4254 | this_cpu = smp_processor_id(); | |
4255 | prev_cpu = task_cpu(p); | |
4256 | load = source_load(prev_cpu, idx); | |
4257 | this_load = target_load(this_cpu, idx); | |
098fb9db | 4258 | |
b3137bc8 MG |
4259 | /* |
4260 | * If sync wakeup then subtract the (maximum possible) | |
4261 | * effect of the currently running task from the load | |
4262 | * of the current CPU: | |
4263 | */ | |
83378269 PZ |
4264 | if (sync) { |
4265 | tg = task_group(current); | |
4266 | weight = current->se.load.weight; | |
4267 | ||
c88d5910 | 4268 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
4269 | load += effective_load(tg, prev_cpu, 0, -weight); |
4270 | } | |
b3137bc8 | 4271 | |
83378269 PZ |
4272 | tg = task_group(p); |
4273 | weight = p->se.load.weight; | |
b3137bc8 | 4274 | |
71a29aa7 PZ |
4275 | /* |
4276 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
4277 | * due to the sync cause above having dropped this_load to 0, we'll |
4278 | * always have an imbalance, but there's really nothing you can do | |
4279 | * about that, so that's good too. | |
71a29aa7 PZ |
4280 | * |
4281 | * Otherwise check if either cpus are near enough in load to allow this | |
4282 | * task to be woken on this_cpu. | |
4283 | */ | |
e37b6a7b PT |
4284 | if (this_load > 0) { |
4285 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
4286 | |
4287 | this_eff_load = 100; | |
4288 | this_eff_load *= power_of(prev_cpu); | |
4289 | this_eff_load *= this_load + | |
4290 | effective_load(tg, this_cpu, weight, weight); | |
4291 | ||
4292 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
4293 | prev_eff_load *= power_of(this_cpu); | |
4294 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
4295 | ||
4296 | balanced = this_eff_load <= prev_eff_load; | |
4297 | } else | |
4298 | balanced = true; | |
b3137bc8 | 4299 | |
098fb9db | 4300 | /* |
4ae7d5ce IM |
4301 | * If the currently running task will sleep within |
4302 | * a reasonable amount of time then attract this newly | |
4303 | * woken task: | |
098fb9db | 4304 | */ |
2fb7635c PZ |
4305 | if (sync && balanced) |
4306 | return 1; | |
098fb9db | 4307 | |
41acab88 | 4308 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
4309 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
4310 | ||
c88d5910 PZ |
4311 | if (balanced || |
4312 | (this_load <= load && | |
4313 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
4314 | /* |
4315 | * This domain has SD_WAKE_AFFINE and | |
4316 | * p is cache cold in this domain, and | |
4317 | * there is no bad imbalance. | |
4318 | */ | |
c88d5910 | 4319 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 4320 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
4321 | |
4322 | return 1; | |
4323 | } | |
4324 | return 0; | |
4325 | } | |
4326 | ||
aaee1203 PZ |
4327 | /* |
4328 | * find_idlest_group finds and returns the least busy CPU group within the | |
4329 | * domain. | |
4330 | */ | |
4331 | static struct sched_group * | |
78e7ed53 | 4332 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 4333 | int this_cpu, int sd_flag) |
e7693a36 | 4334 | { |
b3bd3de6 | 4335 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 4336 | unsigned long min_load = ULONG_MAX, this_load = 0; |
c44f2a02 | 4337 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4338 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 4339 | |
c44f2a02 VG |
4340 | if (sd_flag & SD_BALANCE_WAKE) |
4341 | load_idx = sd->wake_idx; | |
4342 | ||
aaee1203 PZ |
4343 | do { |
4344 | unsigned long load, avg_load; | |
4345 | int local_group; | |
4346 | int i; | |
e7693a36 | 4347 | |
aaee1203 PZ |
4348 | /* Skip over this group if it has no CPUs allowed */ |
4349 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 4350 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
4351 | continue; |
4352 | ||
4353 | local_group = cpumask_test_cpu(this_cpu, | |
4354 | sched_group_cpus(group)); | |
4355 | ||
4356 | /* Tally up the load of all CPUs in the group */ | |
4357 | avg_load = 0; | |
4358 | ||
4359 | for_each_cpu(i, sched_group_cpus(group)) { | |
4360 | /* Bias balancing toward cpus of our domain */ | |
4361 | if (local_group) | |
4362 | load = source_load(i, load_idx); | |
4363 | else | |
4364 | load = target_load(i, load_idx); | |
4365 | ||
4366 | avg_load += load; | |
4367 | } | |
4368 | ||
4369 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 4370 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
4371 | |
4372 | if (local_group) { | |
4373 | this_load = avg_load; | |
aaee1203 PZ |
4374 | } else if (avg_load < min_load) { |
4375 | min_load = avg_load; | |
4376 | idlest = group; | |
4377 | } | |
4378 | } while (group = group->next, group != sd->groups); | |
4379 | ||
4380 | if (!idlest || 100*this_load < imbalance*min_load) | |
4381 | return NULL; | |
4382 | return idlest; | |
4383 | } | |
4384 | ||
4385 | /* | |
4386 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
4387 | */ | |
4388 | static int | |
4389 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
4390 | { | |
4391 | unsigned long load, min_load = ULONG_MAX; | |
4392 | int idlest = -1; | |
4393 | int i; | |
4394 | ||
4395 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 4396 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
4397 | load = weighted_cpuload(i); |
4398 | ||
4399 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
4400 | min_load = load; | |
4401 | idlest = i; | |
e7693a36 GH |
4402 | } |
4403 | } | |
4404 | ||
aaee1203 PZ |
4405 | return idlest; |
4406 | } | |
e7693a36 | 4407 | |
a50bde51 PZ |
4408 | /* |
4409 | * Try and locate an idle CPU in the sched_domain. | |
4410 | */ | |
99bd5e2f | 4411 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 4412 | { |
99bd5e2f | 4413 | struct sched_domain *sd; |
37407ea7 | 4414 | struct sched_group *sg; |
e0a79f52 | 4415 | int i = task_cpu(p); |
a50bde51 | 4416 | |
e0a79f52 MG |
4417 | if (idle_cpu(target)) |
4418 | return target; | |
99bd5e2f SS |
4419 | |
4420 | /* | |
e0a79f52 | 4421 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 4422 | */ |
e0a79f52 MG |
4423 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4424 | return i; | |
a50bde51 PZ |
4425 | |
4426 | /* | |
37407ea7 | 4427 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 4428 | */ |
518cd623 | 4429 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 4430 | for_each_lower_domain(sd) { |
37407ea7 LT |
4431 | sg = sd->groups; |
4432 | do { | |
4433 | if (!cpumask_intersects(sched_group_cpus(sg), | |
4434 | tsk_cpus_allowed(p))) | |
4435 | goto next; | |
4436 | ||
4437 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 4438 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
4439 | goto next; |
4440 | } | |
970e1789 | 4441 | |
37407ea7 LT |
4442 | target = cpumask_first_and(sched_group_cpus(sg), |
4443 | tsk_cpus_allowed(p)); | |
4444 | goto done; | |
4445 | next: | |
4446 | sg = sg->next; | |
4447 | } while (sg != sd->groups); | |
4448 | } | |
4449 | done: | |
a50bde51 PZ |
4450 | return target; |
4451 | } | |
4452 | ||
aaee1203 | 4453 | /* |
de91b9cb MR |
4454 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
4455 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
4456 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 4457 | * |
de91b9cb MR |
4458 | * Balances load by selecting the idlest cpu in the idlest group, or under |
4459 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 4460 | * |
de91b9cb | 4461 | * Returns the target cpu number. |
aaee1203 PZ |
4462 | * |
4463 | * preempt must be disabled. | |
4464 | */ | |
0017d735 | 4465 | static int |
ac66f547 | 4466 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 4467 | { |
29cd8bae | 4468 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 4469 | int cpu = smp_processor_id(); |
c88d5910 | 4470 | int new_cpu = cpu; |
99bd5e2f | 4471 | int want_affine = 0; |
5158f4e4 | 4472 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 4473 | |
29baa747 | 4474 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
4475 | return prev_cpu; |
4476 | ||
0763a660 | 4477 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 4478 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
4479 | want_affine = 1; |
4480 | new_cpu = prev_cpu; | |
4481 | } | |
aaee1203 | 4482 | |
dce840a0 | 4483 | rcu_read_lock(); |
aaee1203 | 4484 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4485 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4486 | continue; | |
4487 | ||
fe3bcfe1 | 4488 | /* |
99bd5e2f SS |
4489 | * If both cpu and prev_cpu are part of this domain, |
4490 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4491 | */ |
99bd5e2f SS |
4492 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4493 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4494 | affine_sd = tmp; | |
29cd8bae | 4495 | break; |
f03542a7 | 4496 | } |
29cd8bae | 4497 | |
f03542a7 | 4498 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4499 | sd = tmp; |
4500 | } | |
4501 | ||
8bf21433 RR |
4502 | if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
4503 | prev_cpu = cpu; | |
dce840a0 | 4504 | |
8bf21433 | 4505 | if (sd_flag & SD_BALANCE_WAKE) { |
dce840a0 PZ |
4506 | new_cpu = select_idle_sibling(p, prev_cpu); |
4507 | goto unlock; | |
8b911acd | 4508 | } |
e7693a36 | 4509 | |
aaee1203 PZ |
4510 | while (sd) { |
4511 | struct sched_group *group; | |
c88d5910 | 4512 | int weight; |
098fb9db | 4513 | |
0763a660 | 4514 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4515 | sd = sd->child; |
4516 | continue; | |
4517 | } | |
098fb9db | 4518 | |
c44f2a02 | 4519 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
4520 | if (!group) { |
4521 | sd = sd->child; | |
4522 | continue; | |
4523 | } | |
4ae7d5ce | 4524 | |
d7c33c49 | 4525 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4526 | if (new_cpu == -1 || new_cpu == cpu) { |
4527 | /* Now try balancing at a lower domain level of cpu */ | |
4528 | sd = sd->child; | |
4529 | continue; | |
e7693a36 | 4530 | } |
aaee1203 PZ |
4531 | |
4532 | /* Now try balancing at a lower domain level of new_cpu */ | |
4533 | cpu = new_cpu; | |
669c55e9 | 4534 | weight = sd->span_weight; |
aaee1203 PZ |
4535 | sd = NULL; |
4536 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4537 | if (weight <= tmp->span_weight) |
aaee1203 | 4538 | break; |
0763a660 | 4539 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4540 | sd = tmp; |
4541 | } | |
4542 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4543 | } |
dce840a0 PZ |
4544 | unlock: |
4545 | rcu_read_unlock(); | |
e7693a36 | 4546 | |
c88d5910 | 4547 | return new_cpu; |
e7693a36 | 4548 | } |
0a74bef8 PT |
4549 | |
4550 | /* | |
4551 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4552 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4553 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4554 | * other assumptions, including the state of rq->lock, should be made. | |
4555 | */ | |
4556 | static void | |
4557 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4558 | { | |
aff3e498 PT |
4559 | struct sched_entity *se = &p->se; |
4560 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4561 | ||
4562 | /* | |
4563 | * Load tracking: accumulate removed load so that it can be processed | |
4564 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4565 | * to blocked load iff they have a positive decay-count. It can never | |
4566 | * be negative here since on-rq tasks have decay-count == 0. | |
4567 | */ | |
4568 | if (se->avg.decay_count) { | |
4569 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4570 | atomic_long_add(se->avg.load_avg_contrib, |
4571 | &cfs_rq->removed_load); | |
aff3e498 | 4572 | } |
3944a927 BS |
4573 | |
4574 | /* We have migrated, no longer consider this task hot */ | |
4575 | se->exec_start = 0; | |
0a74bef8 | 4576 | } |
e7693a36 GH |
4577 | #endif /* CONFIG_SMP */ |
4578 | ||
e52fb7c0 PZ |
4579 | static unsigned long |
4580 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4581 | { |
4582 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4583 | ||
4584 | /* | |
e52fb7c0 PZ |
4585 | * Since its curr running now, convert the gran from real-time |
4586 | * to virtual-time in his units. | |
13814d42 MG |
4587 | * |
4588 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4589 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4590 | * the resulting gran will be larger, therefore penalizing the | |
4591 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4592 | * be smaller, again penalizing the lighter task. | |
4593 | * | |
4594 | * This is especially important for buddies when the leftmost | |
4595 | * task is higher priority than the buddy. | |
0bbd3336 | 4596 | */ |
f4ad9bd2 | 4597 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4598 | } |
4599 | ||
464b7527 PZ |
4600 | /* |
4601 | * Should 'se' preempt 'curr'. | |
4602 | * | |
4603 | * |s1 | |
4604 | * |s2 | |
4605 | * |s3 | |
4606 | * g | |
4607 | * |<--->|c | |
4608 | * | |
4609 | * w(c, s1) = -1 | |
4610 | * w(c, s2) = 0 | |
4611 | * w(c, s3) = 1 | |
4612 | * | |
4613 | */ | |
4614 | static int | |
4615 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4616 | { | |
4617 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4618 | ||
4619 | if (vdiff <= 0) | |
4620 | return -1; | |
4621 | ||
e52fb7c0 | 4622 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4623 | if (vdiff > gran) |
4624 | return 1; | |
4625 | ||
4626 | return 0; | |
4627 | } | |
4628 | ||
02479099 PZ |
4629 | static void set_last_buddy(struct sched_entity *se) |
4630 | { | |
69c80f3e VP |
4631 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4632 | return; | |
4633 | ||
4634 | for_each_sched_entity(se) | |
4635 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4636 | } |
4637 | ||
4638 | static void set_next_buddy(struct sched_entity *se) | |
4639 | { | |
69c80f3e VP |
4640 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4641 | return; | |
4642 | ||
4643 | for_each_sched_entity(se) | |
4644 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4645 | } |
4646 | ||
ac53db59 RR |
4647 | static void set_skip_buddy(struct sched_entity *se) |
4648 | { | |
69c80f3e VP |
4649 | for_each_sched_entity(se) |
4650 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4651 | } |
4652 | ||
bf0f6f24 IM |
4653 | /* |
4654 | * Preempt the current task with a newly woken task if needed: | |
4655 | */ | |
5a9b86f6 | 4656 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4657 | { |
4658 | struct task_struct *curr = rq->curr; | |
8651a86c | 4659 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4660 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4661 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4662 | int next_buddy_marked = 0; |
bf0f6f24 | 4663 | |
4ae7d5ce IM |
4664 | if (unlikely(se == pse)) |
4665 | return; | |
4666 | ||
5238cdd3 | 4667 | /* |
ddcdf6e7 | 4668 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
4669 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4670 | * lead to a throttle). This both saves work and prevents false | |
4671 | * next-buddy nomination below. | |
4672 | */ | |
4673 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4674 | return; | |
4675 | ||
2f36825b | 4676 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4677 | set_next_buddy(pse); |
2f36825b VP |
4678 | next_buddy_marked = 1; |
4679 | } | |
57fdc26d | 4680 | |
aec0a514 BR |
4681 | /* |
4682 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4683 | * wake up path. | |
5238cdd3 PT |
4684 | * |
4685 | * Note: this also catches the edge-case of curr being in a throttled | |
4686 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4687 | * enqueue of curr) will have resulted in resched being set. This | |
4688 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4689 | * below. | |
aec0a514 BR |
4690 | */ |
4691 | if (test_tsk_need_resched(curr)) | |
4692 | return; | |
4693 | ||
a2f5c9ab DH |
4694 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4695 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4696 | likely(p->policy != SCHED_IDLE)) | |
4697 | goto preempt; | |
4698 | ||
91c234b4 | 4699 | /* |
a2f5c9ab DH |
4700 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4701 | * is driven by the tick): | |
91c234b4 | 4702 | */ |
8ed92e51 | 4703 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4704 | return; |
bf0f6f24 | 4705 | |
464b7527 | 4706 | find_matching_se(&se, &pse); |
9bbd7374 | 4707 | update_curr(cfs_rq_of(se)); |
002f128b | 4708 | BUG_ON(!pse); |
2f36825b VP |
4709 | if (wakeup_preempt_entity(se, pse) == 1) { |
4710 | /* | |
4711 | * Bias pick_next to pick the sched entity that is | |
4712 | * triggering this preemption. | |
4713 | */ | |
4714 | if (!next_buddy_marked) | |
4715 | set_next_buddy(pse); | |
3a7e73a2 | 4716 | goto preempt; |
2f36825b | 4717 | } |
464b7527 | 4718 | |
3a7e73a2 | 4719 | return; |
a65ac745 | 4720 | |
3a7e73a2 PZ |
4721 | preempt: |
4722 | resched_task(curr); | |
4723 | /* | |
4724 | * Only set the backward buddy when the current task is still | |
4725 | * on the rq. This can happen when a wakeup gets interleaved | |
4726 | * with schedule on the ->pre_schedule() or idle_balance() | |
4727 | * point, either of which can * drop the rq lock. | |
4728 | * | |
4729 | * Also, during early boot the idle thread is in the fair class, | |
4730 | * for obvious reasons its a bad idea to schedule back to it. | |
4731 | */ | |
4732 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
4733 | return; | |
4734 | ||
4735 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
4736 | set_last_buddy(se); | |
bf0f6f24 IM |
4737 | } |
4738 | ||
606dba2e PZ |
4739 | static struct task_struct * |
4740 | pick_next_task_fair(struct rq *rq, struct task_struct *prev) | |
bf0f6f24 IM |
4741 | { |
4742 | struct cfs_rq *cfs_rq = &rq->cfs; | |
4743 | struct sched_entity *se; | |
678d5718 | 4744 | struct task_struct *p; |
37e117c0 | 4745 | int new_tasks; |
678d5718 | 4746 | |
6e83125c | 4747 | again: |
678d5718 PZ |
4748 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4749 | if (!cfs_rq->nr_running) | |
38033c37 | 4750 | goto idle; |
678d5718 | 4751 | |
3f1d2a31 | 4752 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
4753 | goto simple; |
4754 | ||
4755 | /* | |
4756 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
4757 | * likely that a next task is from the same cgroup as the current. | |
4758 | * | |
4759 | * Therefore attempt to avoid putting and setting the entire cgroup | |
4760 | * hierarchy, only change the part that actually changes. | |
4761 | */ | |
4762 | ||
4763 | do { | |
4764 | struct sched_entity *curr = cfs_rq->curr; | |
4765 | ||
4766 | /* | |
4767 | * Since we got here without doing put_prev_entity() we also | |
4768 | * have to consider cfs_rq->curr. If it is still a runnable | |
4769 | * entity, update_curr() will update its vruntime, otherwise | |
4770 | * forget we've ever seen it. | |
4771 | */ | |
4772 | if (curr && curr->on_rq) | |
4773 | update_curr(cfs_rq); | |
4774 | else | |
4775 | curr = NULL; | |
4776 | ||
4777 | /* | |
4778 | * This call to check_cfs_rq_runtime() will do the throttle and | |
4779 | * dequeue its entity in the parent(s). Therefore the 'simple' | |
4780 | * nr_running test will indeed be correct. | |
4781 | */ | |
4782 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
4783 | goto simple; | |
4784 | ||
4785 | se = pick_next_entity(cfs_rq, curr); | |
4786 | cfs_rq = group_cfs_rq(se); | |
4787 | } while (cfs_rq); | |
4788 | ||
4789 | p = task_of(se); | |
4790 | ||
4791 | /* | |
4792 | * Since we haven't yet done put_prev_entity and if the selected task | |
4793 | * is a different task than we started out with, try and touch the | |
4794 | * least amount of cfs_rqs. | |
4795 | */ | |
4796 | if (prev != p) { | |
4797 | struct sched_entity *pse = &prev->se; | |
4798 | ||
4799 | while (!(cfs_rq = is_same_group(se, pse))) { | |
4800 | int se_depth = se->depth; | |
4801 | int pse_depth = pse->depth; | |
4802 | ||
4803 | if (se_depth <= pse_depth) { | |
4804 | put_prev_entity(cfs_rq_of(pse), pse); | |
4805 | pse = parent_entity(pse); | |
4806 | } | |
4807 | if (se_depth >= pse_depth) { | |
4808 | set_next_entity(cfs_rq_of(se), se); | |
4809 | se = parent_entity(se); | |
4810 | } | |
4811 | } | |
4812 | ||
4813 | put_prev_entity(cfs_rq, pse); | |
4814 | set_next_entity(cfs_rq, se); | |
4815 | } | |
4816 | ||
4817 | if (hrtick_enabled(rq)) | |
4818 | hrtick_start_fair(rq, p); | |
4819 | ||
4820 | return p; | |
4821 | simple: | |
4822 | cfs_rq = &rq->cfs; | |
4823 | #endif | |
bf0f6f24 | 4824 | |
36ace27e | 4825 | if (!cfs_rq->nr_running) |
38033c37 | 4826 | goto idle; |
bf0f6f24 | 4827 | |
3f1d2a31 | 4828 | put_prev_task(rq, prev); |
606dba2e | 4829 | |
bf0f6f24 | 4830 | do { |
678d5718 | 4831 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 4832 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
4833 | cfs_rq = group_cfs_rq(se); |
4834 | } while (cfs_rq); | |
4835 | ||
8f4d37ec | 4836 | p = task_of(se); |
678d5718 | 4837 | |
b39e66ea MG |
4838 | if (hrtick_enabled(rq)) |
4839 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4840 | |
4841 | return p; | |
38033c37 PZ |
4842 | |
4843 | idle: | |
e4aa358b | 4844 | new_tasks = idle_balance(rq); |
37e117c0 PZ |
4845 | /* |
4846 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
4847 | * possible for any higher priority task to appear. In that case we | |
4848 | * must re-start the pick_next_entity() loop. | |
4849 | */ | |
e4aa358b | 4850 | if (new_tasks < 0) |
37e117c0 PZ |
4851 | return RETRY_TASK; |
4852 | ||
e4aa358b | 4853 | if (new_tasks > 0) |
38033c37 | 4854 | goto again; |
38033c37 PZ |
4855 | |
4856 | return NULL; | |
bf0f6f24 IM |
4857 | } |
4858 | ||
4859 | /* | |
4860 | * Account for a descheduled task: | |
4861 | */ | |
31ee529c | 4862 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4863 | { |
4864 | struct sched_entity *se = &prev->se; | |
4865 | struct cfs_rq *cfs_rq; | |
4866 | ||
4867 | for_each_sched_entity(se) { | |
4868 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4869 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4870 | } |
4871 | } | |
4872 | ||
ac53db59 RR |
4873 | /* |
4874 | * sched_yield() is very simple | |
4875 | * | |
4876 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4877 | */ | |
4878 | static void yield_task_fair(struct rq *rq) | |
4879 | { | |
4880 | struct task_struct *curr = rq->curr; | |
4881 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4882 | struct sched_entity *se = &curr->se; | |
4883 | ||
4884 | /* | |
4885 | * Are we the only task in the tree? | |
4886 | */ | |
4887 | if (unlikely(rq->nr_running == 1)) | |
4888 | return; | |
4889 | ||
4890 | clear_buddies(cfs_rq, se); | |
4891 | ||
4892 | if (curr->policy != SCHED_BATCH) { | |
4893 | update_rq_clock(rq); | |
4894 | /* | |
4895 | * Update run-time statistics of the 'current'. | |
4896 | */ | |
4897 | update_curr(cfs_rq); | |
916671c0 MG |
4898 | /* |
4899 | * Tell update_rq_clock() that we've just updated, | |
4900 | * so we don't do microscopic update in schedule() | |
4901 | * and double the fastpath cost. | |
4902 | */ | |
4903 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4904 | } |
4905 | ||
4906 | set_skip_buddy(se); | |
4907 | } | |
4908 | ||
d95f4122 MG |
4909 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4910 | { | |
4911 | struct sched_entity *se = &p->se; | |
4912 | ||
5238cdd3 PT |
4913 | /* throttled hierarchies are not runnable */ |
4914 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4915 | return false; |
4916 | ||
4917 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4918 | set_next_buddy(se); | |
4919 | ||
d95f4122 MG |
4920 | yield_task_fair(rq); |
4921 | ||
4922 | return true; | |
4923 | } | |
4924 | ||
681f3e68 | 4925 | #ifdef CONFIG_SMP |
bf0f6f24 | 4926 | /************************************************** |
e9c84cb8 PZ |
4927 | * Fair scheduling class load-balancing methods. |
4928 | * | |
4929 | * BASICS | |
4930 | * | |
4931 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4932 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4933 | * time to each task. This is expressed in the following equation: | |
4934 | * | |
4935 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4936 | * | |
4937 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4938 | * W_i,0 is defined as: | |
4939 | * | |
4940 | * W_i,0 = \Sum_j w_i,j (2) | |
4941 | * | |
4942 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4943 | * is derived from the nice value as per prio_to_weight[]. | |
4944 | * | |
4945 | * The weight average is an exponential decay average of the instantaneous | |
4946 | * weight: | |
4947 | * | |
4948 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4949 | * | |
4950 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4951 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4952 | * can also include other factors [XXX]. | |
4953 | * | |
4954 | * To achieve this balance we define a measure of imbalance which follows | |
4955 | * directly from (1): | |
4956 | * | |
4957 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4958 | * | |
4959 | * We them move tasks around to minimize the imbalance. In the continuous | |
4960 | * function space it is obvious this converges, in the discrete case we get | |
4961 | * a few fun cases generally called infeasible weight scenarios. | |
4962 | * | |
4963 | * [XXX expand on: | |
4964 | * - infeasible weights; | |
4965 | * - local vs global optima in the discrete case. ] | |
4966 | * | |
4967 | * | |
4968 | * SCHED DOMAINS | |
4969 | * | |
4970 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4971 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4972 | * topology where each level pairs two lower groups (or better). This results | |
4973 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4974 | * tree to only the first of the previous level and we decrease the frequency | |
4975 | * of load-balance at each level inv. proportional to the number of cpus in | |
4976 | * the groups. | |
4977 | * | |
4978 | * This yields: | |
4979 | * | |
4980 | * log_2 n 1 n | |
4981 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4982 | * i = 0 2^i 2^i | |
4983 | * `- size of each group | |
4984 | * | | `- number of cpus doing load-balance | |
4985 | * | `- freq | |
4986 | * `- sum over all levels | |
4987 | * | |
4988 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4989 | * this makes (5) the runtime complexity of the balancer. | |
4990 | * | |
4991 | * An important property here is that each CPU is still (indirectly) connected | |
4992 | * to every other cpu in at most O(log n) steps: | |
4993 | * | |
4994 | * The adjacency matrix of the resulting graph is given by: | |
4995 | * | |
4996 | * log_2 n | |
4997 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4998 | * k = 0 | |
4999 | * | |
5000 | * And you'll find that: | |
5001 | * | |
5002 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
5003 | * | |
5004 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
5005 | * The task movement gives a factor of O(m), giving a convergence complexity | |
5006 | * of: | |
5007 | * | |
5008 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
5009 | * | |
5010 | * | |
5011 | * WORK CONSERVING | |
5012 | * | |
5013 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
5014 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
5015 | * tree itself instead of relying on other CPUs to bring it work. | |
5016 | * | |
5017 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
5018 | * time. | |
5019 | * | |
5020 | * [XXX more?] | |
5021 | * | |
5022 | * | |
5023 | * CGROUPS | |
5024 | * | |
5025 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
5026 | * | |
5027 | * s_k,i | |
5028 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
5029 | * S_k | |
5030 | * | |
5031 | * Where | |
5032 | * | |
5033 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
5034 | * | |
5035 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
5036 | * | |
5037 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
5038 | * property. | |
5039 | * | |
5040 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
5041 | * rewrite all of this once again.] | |
5042 | */ | |
bf0f6f24 | 5043 | |
ed387b78 HS |
5044 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
5045 | ||
0ec8aa00 PZ |
5046 | enum fbq_type { regular, remote, all }; |
5047 | ||
ddcdf6e7 | 5048 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 5049 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
5050 | #define LBF_DST_PINNED 0x04 |
5051 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
5052 | |
5053 | struct lb_env { | |
5054 | struct sched_domain *sd; | |
5055 | ||
ddcdf6e7 | 5056 | struct rq *src_rq; |
85c1e7da | 5057 | int src_cpu; |
ddcdf6e7 PZ |
5058 | |
5059 | int dst_cpu; | |
5060 | struct rq *dst_rq; | |
5061 | ||
88b8dac0 SV |
5062 | struct cpumask *dst_grpmask; |
5063 | int new_dst_cpu; | |
ddcdf6e7 | 5064 | enum cpu_idle_type idle; |
bd939f45 | 5065 | long imbalance; |
b9403130 MW |
5066 | /* The set of CPUs under consideration for load-balancing */ |
5067 | struct cpumask *cpus; | |
5068 | ||
ddcdf6e7 | 5069 | unsigned int flags; |
367456c7 PZ |
5070 | |
5071 | unsigned int loop; | |
5072 | unsigned int loop_break; | |
5073 | unsigned int loop_max; | |
0ec8aa00 PZ |
5074 | |
5075 | enum fbq_type fbq_type; | |
ddcdf6e7 PZ |
5076 | }; |
5077 | ||
1e3c88bd | 5078 | /* |
ddcdf6e7 | 5079 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
5080 | * Both runqueues must be locked. |
5081 | */ | |
ddcdf6e7 | 5082 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 5083 | { |
ddcdf6e7 PZ |
5084 | deactivate_task(env->src_rq, p, 0); |
5085 | set_task_cpu(p, env->dst_cpu); | |
5086 | activate_task(env->dst_rq, p, 0); | |
5087 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
5088 | } |
5089 | ||
029632fb PZ |
5090 | /* |
5091 | * Is this task likely cache-hot: | |
5092 | */ | |
5093 | static int | |
6037dd1a | 5094 | task_hot(struct task_struct *p, u64 now) |
029632fb PZ |
5095 | { |
5096 | s64 delta; | |
5097 | ||
5098 | if (p->sched_class != &fair_sched_class) | |
5099 | return 0; | |
5100 | ||
5101 | if (unlikely(p->policy == SCHED_IDLE)) | |
5102 | return 0; | |
5103 | ||
5104 | /* | |
5105 | * Buddy candidates are cache hot: | |
5106 | */ | |
5107 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
5108 | (&p->se == cfs_rq_of(&p->se)->next || | |
5109 | &p->se == cfs_rq_of(&p->se)->last)) | |
5110 | return 1; | |
5111 | ||
5112 | if (sysctl_sched_migration_cost == -1) | |
5113 | return 1; | |
5114 | if (sysctl_sched_migration_cost == 0) | |
5115 | return 0; | |
5116 | ||
5117 | delta = now - p->se.exec_start; | |
5118 | ||
5119 | return delta < (s64)sysctl_sched_migration_cost; | |
5120 | } | |
5121 | ||
3a7053b3 MG |
5122 | #ifdef CONFIG_NUMA_BALANCING |
5123 | /* Returns true if the destination node has incurred more faults */ | |
5124 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
5125 | { | |
5126 | int src_nid, dst_nid; | |
5127 | ||
ff1df896 | 5128 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory || |
3a7053b3 MG |
5129 | !(env->sd->flags & SD_NUMA)) { |
5130 | return false; | |
5131 | } | |
5132 | ||
5133 | src_nid = cpu_to_node(env->src_cpu); | |
5134 | dst_nid = cpu_to_node(env->dst_cpu); | |
5135 | ||
83e1d2cd | 5136 | if (src_nid == dst_nid) |
3a7053b3 MG |
5137 | return false; |
5138 | ||
83e1d2cd MG |
5139 | /* Always encourage migration to the preferred node. */ |
5140 | if (dst_nid == p->numa_preferred_nid) | |
5141 | return true; | |
5142 | ||
887c290e RR |
5143 | /* If both task and group weight improve, this move is a winner. */ |
5144 | if (task_weight(p, dst_nid) > task_weight(p, src_nid) && | |
5145 | group_weight(p, dst_nid) > group_weight(p, src_nid)) | |
3a7053b3 MG |
5146 | return true; |
5147 | ||
5148 | return false; | |
5149 | } | |
7a0f3083 MG |
5150 | |
5151 | ||
5152 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
5153 | { | |
5154 | int src_nid, dst_nid; | |
5155 | ||
5156 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
5157 | return false; | |
5158 | ||
ff1df896 | 5159 | if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA)) |
7a0f3083 MG |
5160 | return false; |
5161 | ||
5162 | src_nid = cpu_to_node(env->src_cpu); | |
5163 | dst_nid = cpu_to_node(env->dst_cpu); | |
5164 | ||
83e1d2cd | 5165 | if (src_nid == dst_nid) |
7a0f3083 MG |
5166 | return false; |
5167 | ||
83e1d2cd MG |
5168 | /* Migrating away from the preferred node is always bad. */ |
5169 | if (src_nid == p->numa_preferred_nid) | |
5170 | return true; | |
5171 | ||
887c290e RR |
5172 | /* If either task or group weight get worse, don't do it. */ |
5173 | if (task_weight(p, dst_nid) < task_weight(p, src_nid) || | |
5174 | group_weight(p, dst_nid) < group_weight(p, src_nid)) | |
7a0f3083 MG |
5175 | return true; |
5176 | ||
5177 | return false; | |
5178 | } | |
5179 | ||
3a7053b3 MG |
5180 | #else |
5181 | static inline bool migrate_improves_locality(struct task_struct *p, | |
5182 | struct lb_env *env) | |
5183 | { | |
5184 | return false; | |
5185 | } | |
7a0f3083 MG |
5186 | |
5187 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
5188 | struct lb_env *env) | |
5189 | { | |
5190 | return false; | |
5191 | } | |
3a7053b3 MG |
5192 | #endif |
5193 | ||
1e3c88bd PZ |
5194 | /* |
5195 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
5196 | */ | |
5197 | static | |
8e45cb54 | 5198 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
5199 | { |
5200 | int tsk_cache_hot = 0; | |
5201 | /* | |
5202 | * We do not migrate tasks that are: | |
d3198084 | 5203 | * 1) throttled_lb_pair, or |
1e3c88bd | 5204 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
5205 | * 3) running (obviously), or |
5206 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 5207 | */ |
d3198084 JK |
5208 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
5209 | return 0; | |
5210 | ||
ddcdf6e7 | 5211 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 5212 | int cpu; |
88b8dac0 | 5213 | |
41acab88 | 5214 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 5215 | |
6263322c PZ |
5216 | env->flags |= LBF_SOME_PINNED; |
5217 | ||
88b8dac0 SV |
5218 | /* |
5219 | * Remember if this task can be migrated to any other cpu in | |
5220 | * our sched_group. We may want to revisit it if we couldn't | |
5221 | * meet load balance goals by pulling other tasks on src_cpu. | |
5222 | * | |
5223 | * Also avoid computing new_dst_cpu if we have already computed | |
5224 | * one in current iteration. | |
5225 | */ | |
6263322c | 5226 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
5227 | return 0; |
5228 | ||
e02e60c1 JK |
5229 | /* Prevent to re-select dst_cpu via env's cpus */ |
5230 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
5231 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 5232 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
5233 | env->new_dst_cpu = cpu; |
5234 | break; | |
5235 | } | |
88b8dac0 | 5236 | } |
e02e60c1 | 5237 | |
1e3c88bd PZ |
5238 | return 0; |
5239 | } | |
88b8dac0 SV |
5240 | |
5241 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 5242 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 5243 | |
ddcdf6e7 | 5244 | if (task_running(env->src_rq, p)) { |
41acab88 | 5245 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
5246 | return 0; |
5247 | } | |
5248 | ||
5249 | /* | |
5250 | * Aggressive migration if: | |
3a7053b3 MG |
5251 | * 1) destination numa is preferred |
5252 | * 2) task is cache cold, or | |
5253 | * 3) too many balance attempts have failed. | |
1e3c88bd | 5254 | */ |
6037dd1a | 5255 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq)); |
7a0f3083 MG |
5256 | if (!tsk_cache_hot) |
5257 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
5258 | |
5259 | if (migrate_improves_locality(p, env)) { | |
5260 | #ifdef CONFIG_SCHEDSTATS | |
5261 | if (tsk_cache_hot) { | |
5262 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
5263 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
5264 | } | |
5265 | #endif | |
5266 | return 1; | |
5267 | } | |
5268 | ||
1e3c88bd | 5269 | if (!tsk_cache_hot || |
8e45cb54 | 5270 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 5271 | |
1e3c88bd | 5272 | if (tsk_cache_hot) { |
8e45cb54 | 5273 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 5274 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 5275 | } |
4e2dcb73 | 5276 | |
1e3c88bd PZ |
5277 | return 1; |
5278 | } | |
5279 | ||
4e2dcb73 ZH |
5280 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
5281 | return 0; | |
1e3c88bd PZ |
5282 | } |
5283 | ||
897c395f PZ |
5284 | /* |
5285 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
5286 | * part of active balancing operations within "domain". | |
5287 | * Returns 1 if successful and 0 otherwise. | |
5288 | * | |
5289 | * Called with both runqueues locked. | |
5290 | */ | |
8e45cb54 | 5291 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
5292 | { |
5293 | struct task_struct *p, *n; | |
897c395f | 5294 | |
367456c7 | 5295 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
5296 | if (!can_migrate_task(p, env)) |
5297 | continue; | |
897c395f | 5298 | |
367456c7 PZ |
5299 | move_task(p, env); |
5300 | /* | |
5301 | * Right now, this is only the second place move_task() | |
5302 | * is called, so we can safely collect move_task() | |
5303 | * stats here rather than inside move_task(). | |
5304 | */ | |
5305 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
5306 | return 1; | |
897c395f | 5307 | } |
897c395f PZ |
5308 | return 0; |
5309 | } | |
5310 | ||
eb95308e PZ |
5311 | static const unsigned int sched_nr_migrate_break = 32; |
5312 | ||
5d6523eb | 5313 | /* |
bd939f45 | 5314 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
5315 | * this_rq, as part of a balancing operation within domain "sd". |
5316 | * Returns 1 if successful and 0 otherwise. | |
5317 | * | |
5318 | * Called with both runqueues locked. | |
5319 | */ | |
5320 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 5321 | { |
5d6523eb PZ |
5322 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
5323 | struct task_struct *p; | |
367456c7 PZ |
5324 | unsigned long load; |
5325 | int pulled = 0; | |
1e3c88bd | 5326 | |
bd939f45 | 5327 | if (env->imbalance <= 0) |
5d6523eb | 5328 | return 0; |
1e3c88bd | 5329 | |
5d6523eb PZ |
5330 | while (!list_empty(tasks)) { |
5331 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 5332 | |
367456c7 PZ |
5333 | env->loop++; |
5334 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 5335 | if (env->loop > env->loop_max) |
367456c7 | 5336 | break; |
5d6523eb PZ |
5337 | |
5338 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 5339 | if (env->loop > env->loop_break) { |
eb95308e | 5340 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 5341 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 5342 | break; |
a195f004 | 5343 | } |
1e3c88bd | 5344 | |
d3198084 | 5345 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
5346 | goto next; |
5347 | ||
5348 | load = task_h_load(p); | |
5d6523eb | 5349 | |
eb95308e | 5350 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
5351 | goto next; |
5352 | ||
bd939f45 | 5353 | if ((load / 2) > env->imbalance) |
367456c7 | 5354 | goto next; |
1e3c88bd | 5355 | |
ddcdf6e7 | 5356 | move_task(p, env); |
ee00e66f | 5357 | pulled++; |
bd939f45 | 5358 | env->imbalance -= load; |
1e3c88bd PZ |
5359 | |
5360 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
5361 | /* |
5362 | * NEWIDLE balancing is a source of latency, so preemptible | |
5363 | * kernels will stop after the first task is pulled to minimize | |
5364 | * the critical section. | |
5365 | */ | |
5d6523eb | 5366 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 5367 | break; |
1e3c88bd PZ |
5368 | #endif |
5369 | ||
ee00e66f PZ |
5370 | /* |
5371 | * We only want to steal up to the prescribed amount of | |
5372 | * weighted load. | |
5373 | */ | |
bd939f45 | 5374 | if (env->imbalance <= 0) |
ee00e66f | 5375 | break; |
367456c7 PZ |
5376 | |
5377 | continue; | |
5378 | next: | |
5d6523eb | 5379 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 5380 | } |
5d6523eb | 5381 | |
1e3c88bd | 5382 | /* |
ddcdf6e7 PZ |
5383 | * Right now, this is one of only two places move_task() is called, |
5384 | * so we can safely collect move_task() stats here rather than | |
5385 | * inside move_task(). | |
1e3c88bd | 5386 | */ |
8e45cb54 | 5387 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 5388 | |
5d6523eb | 5389 | return pulled; |
1e3c88bd PZ |
5390 | } |
5391 | ||
230059de | 5392 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
5393 | /* |
5394 | * update tg->load_weight by folding this cpu's load_avg | |
5395 | */ | |
48a16753 | 5396 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 5397 | { |
48a16753 PT |
5398 | struct sched_entity *se = tg->se[cpu]; |
5399 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 5400 | |
48a16753 PT |
5401 | /* throttled entities do not contribute to load */ |
5402 | if (throttled_hierarchy(cfs_rq)) | |
5403 | return; | |
9e3081ca | 5404 | |
aff3e498 | 5405 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 5406 | |
82958366 PT |
5407 | if (se) { |
5408 | update_entity_load_avg(se, 1); | |
5409 | /* | |
5410 | * We pivot on our runnable average having decayed to zero for | |
5411 | * list removal. This generally implies that all our children | |
5412 | * have also been removed (modulo rounding error or bandwidth | |
5413 | * control); however, such cases are rare and we can fix these | |
5414 | * at enqueue. | |
5415 | * | |
5416 | * TODO: fix up out-of-order children on enqueue. | |
5417 | */ | |
5418 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
5419 | list_del_leaf_cfs_rq(cfs_rq); | |
5420 | } else { | |
48a16753 | 5421 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
5422 | update_rq_runnable_avg(rq, rq->nr_running); |
5423 | } | |
9e3081ca PZ |
5424 | } |
5425 | ||
48a16753 | 5426 | static void update_blocked_averages(int cpu) |
9e3081ca | 5427 | { |
9e3081ca | 5428 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
5429 | struct cfs_rq *cfs_rq; |
5430 | unsigned long flags; | |
9e3081ca | 5431 | |
48a16753 PT |
5432 | raw_spin_lock_irqsave(&rq->lock, flags); |
5433 | update_rq_clock(rq); | |
9763b67f PZ |
5434 | /* |
5435 | * Iterates the task_group tree in a bottom up fashion, see | |
5436 | * list_add_leaf_cfs_rq() for details. | |
5437 | */ | |
64660c86 | 5438 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
5439 | /* |
5440 | * Note: We may want to consider periodically releasing | |
5441 | * rq->lock about these updates so that creating many task | |
5442 | * groups does not result in continually extending hold time. | |
5443 | */ | |
5444 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 5445 | } |
48a16753 PT |
5446 | |
5447 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
5448 | } |
5449 | ||
9763b67f | 5450 | /* |
68520796 | 5451 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
5452 | * This needs to be done in a top-down fashion because the load of a child |
5453 | * group is a fraction of its parents load. | |
5454 | */ | |
68520796 | 5455 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 5456 | { |
68520796 VD |
5457 | struct rq *rq = rq_of(cfs_rq); |
5458 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 5459 | unsigned long now = jiffies; |
68520796 | 5460 | unsigned long load; |
a35b6466 | 5461 | |
68520796 | 5462 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
5463 | return; |
5464 | ||
68520796 VD |
5465 | cfs_rq->h_load_next = NULL; |
5466 | for_each_sched_entity(se) { | |
5467 | cfs_rq = cfs_rq_of(se); | |
5468 | cfs_rq->h_load_next = se; | |
5469 | if (cfs_rq->last_h_load_update == now) | |
5470 | break; | |
5471 | } | |
a35b6466 | 5472 | |
68520796 | 5473 | if (!se) { |
7e3115ef | 5474 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
5475 | cfs_rq->last_h_load_update = now; |
5476 | } | |
5477 | ||
5478 | while ((se = cfs_rq->h_load_next) != NULL) { | |
5479 | load = cfs_rq->h_load; | |
5480 | load = div64_ul(load * se->avg.load_avg_contrib, | |
5481 | cfs_rq->runnable_load_avg + 1); | |
5482 | cfs_rq = group_cfs_rq(se); | |
5483 | cfs_rq->h_load = load; | |
5484 | cfs_rq->last_h_load_update = now; | |
5485 | } | |
9763b67f PZ |
5486 | } |
5487 | ||
367456c7 | 5488 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 5489 | { |
367456c7 | 5490 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 5491 | |
68520796 | 5492 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
5493 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
5494 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
5495 | } |
5496 | #else | |
48a16753 | 5497 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
5498 | { |
5499 | } | |
5500 | ||
367456c7 | 5501 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 5502 | { |
a003a25b | 5503 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 5504 | } |
230059de | 5505 | #endif |
1e3c88bd | 5506 | |
1e3c88bd | 5507 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
5508 | /* |
5509 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
5510 | */ | |
5511 | struct sg_lb_stats { | |
5512 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
5513 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 5514 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 5515 | unsigned long load_per_task; |
3ae11c90 | 5516 | unsigned long group_power; |
147c5fc2 PZ |
5517 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
5518 | unsigned int group_capacity; | |
5519 | unsigned int idle_cpus; | |
5520 | unsigned int group_weight; | |
1e3c88bd | 5521 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 5522 | int group_has_capacity; /* Is there extra capacity in the group? */ |
0ec8aa00 PZ |
5523 | #ifdef CONFIG_NUMA_BALANCING |
5524 | unsigned int nr_numa_running; | |
5525 | unsigned int nr_preferred_running; | |
5526 | #endif | |
1e3c88bd PZ |
5527 | }; |
5528 | ||
56cf515b JK |
5529 | /* |
5530 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
5531 | * during load balancing. | |
5532 | */ | |
5533 | struct sd_lb_stats { | |
5534 | struct sched_group *busiest; /* Busiest group in this sd */ | |
5535 | struct sched_group *local; /* Local group in this sd */ | |
5536 | unsigned long total_load; /* Total load of all groups in sd */ | |
5537 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
5538 | unsigned long avg_load; /* Average load across all groups in sd */ | |
5539 | ||
56cf515b | 5540 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 5541 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
5542 | }; |
5543 | ||
147c5fc2 PZ |
5544 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
5545 | { | |
5546 | /* | |
5547 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
5548 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
5549 | * We must however clear busiest_stat::avg_load because | |
5550 | * update_sd_pick_busiest() reads this before assignment. | |
5551 | */ | |
5552 | *sds = (struct sd_lb_stats){ | |
5553 | .busiest = NULL, | |
5554 | .local = NULL, | |
5555 | .total_load = 0UL, | |
5556 | .total_pwr = 0UL, | |
5557 | .busiest_stat = { | |
5558 | .avg_load = 0UL, | |
5559 | }, | |
5560 | }; | |
5561 | } | |
5562 | ||
1e3c88bd PZ |
5563 | /** |
5564 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
5565 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 5566 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
5567 | * |
5568 | * Return: The load index. | |
1e3c88bd PZ |
5569 | */ |
5570 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
5571 | enum cpu_idle_type idle) | |
5572 | { | |
5573 | int load_idx; | |
5574 | ||
5575 | switch (idle) { | |
5576 | case CPU_NOT_IDLE: | |
5577 | load_idx = sd->busy_idx; | |
5578 | break; | |
5579 | ||
5580 | case CPU_NEWLY_IDLE: | |
5581 | load_idx = sd->newidle_idx; | |
5582 | break; | |
5583 | default: | |
5584 | load_idx = sd->idle_idx; | |
5585 | break; | |
5586 | } | |
5587 | ||
5588 | return load_idx; | |
5589 | } | |
5590 | ||
15f803c9 | 5591 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5592 | { |
1399fa78 | 5593 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
5594 | } |
5595 | ||
5596 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
5597 | { | |
5598 | return default_scale_freq_power(sd, cpu); | |
5599 | } | |
5600 | ||
15f803c9 | 5601 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5602 | { |
669c55e9 | 5603 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
5604 | unsigned long smt_gain = sd->smt_gain; |
5605 | ||
5606 | smt_gain /= weight; | |
5607 | ||
5608 | return smt_gain; | |
5609 | } | |
5610 | ||
5611 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
5612 | { | |
5613 | return default_scale_smt_power(sd, cpu); | |
5614 | } | |
5615 | ||
15f803c9 | 5616 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
5617 | { |
5618 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5619 | u64 total, available, age_stamp, avg; |
cadefd3d | 5620 | s64 delta; |
1e3c88bd | 5621 | |
b654f7de PZ |
5622 | /* |
5623 | * Since we're reading these variables without serialization make sure | |
5624 | * we read them once before doing sanity checks on them. | |
5625 | */ | |
5626 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5627 | avg = ACCESS_ONCE(rq->rt_avg); | |
5628 | ||
cadefd3d PZ |
5629 | delta = rq_clock(rq) - age_stamp; |
5630 | if (unlikely(delta < 0)) | |
5631 | delta = 0; | |
5632 | ||
5633 | total = sched_avg_period() + delta; | |
aa483808 | 5634 | |
b654f7de | 5635 | if (unlikely(total < avg)) { |
aa483808 VP |
5636 | /* Ensures that power won't end up being negative */ |
5637 | available = 0; | |
5638 | } else { | |
b654f7de | 5639 | available = total - avg; |
aa483808 | 5640 | } |
1e3c88bd | 5641 | |
1399fa78 NR |
5642 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
5643 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 5644 | |
1399fa78 | 5645 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5646 | |
5647 | return div_u64(available, total); | |
5648 | } | |
5649 | ||
5650 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
5651 | { | |
669c55e9 | 5652 | unsigned long weight = sd->span_weight; |
1399fa78 | 5653 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
5654 | struct sched_group *sdg = sd->groups; |
5655 | ||
1e3c88bd PZ |
5656 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
5657 | if (sched_feat(ARCH_POWER)) | |
5658 | power *= arch_scale_smt_power(sd, cpu); | |
5659 | else | |
5660 | power *= default_scale_smt_power(sd, cpu); | |
5661 | ||
1399fa78 | 5662 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5663 | } |
5664 | ||
9c3f75cb | 5665 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
5666 | |
5667 | if (sched_feat(ARCH_POWER)) | |
5668 | power *= arch_scale_freq_power(sd, cpu); | |
5669 | else | |
5670 | power *= default_scale_freq_power(sd, cpu); | |
5671 | ||
1399fa78 | 5672 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 5673 | |
1e3c88bd | 5674 | power *= scale_rt_power(cpu); |
1399fa78 | 5675 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5676 | |
5677 | if (!power) | |
5678 | power = 1; | |
5679 | ||
e51fd5e2 | 5680 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 5681 | sdg->sgp->power = power; |
1e3c88bd PZ |
5682 | } |
5683 | ||
029632fb | 5684 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
5685 | { |
5686 | struct sched_domain *child = sd->child; | |
5687 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 5688 | unsigned long power, power_orig; |
4ec4412e VG |
5689 | unsigned long interval; |
5690 | ||
5691 | interval = msecs_to_jiffies(sd->balance_interval); | |
5692 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
5693 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
5694 | |
5695 | if (!child) { | |
5696 | update_cpu_power(sd, cpu); | |
5697 | return; | |
5698 | } | |
5699 | ||
863bffc8 | 5700 | power_orig = power = 0; |
1e3c88bd | 5701 | |
74a5ce20 PZ |
5702 | if (child->flags & SD_OVERLAP) { |
5703 | /* | |
5704 | * SD_OVERLAP domains cannot assume that child groups | |
5705 | * span the current group. | |
5706 | */ | |
5707 | ||
863bffc8 | 5708 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
9abf24d4 SD |
5709 | struct sched_group_power *sgp; |
5710 | struct rq *rq = cpu_rq(cpu); | |
863bffc8 | 5711 | |
9abf24d4 SD |
5712 | /* |
5713 | * build_sched_domains() -> init_sched_groups_power() | |
5714 | * gets here before we've attached the domains to the | |
5715 | * runqueues. | |
5716 | * | |
5717 | * Use power_of(), which is set irrespective of domains | |
5718 | * in update_cpu_power(). | |
5719 | * | |
5720 | * This avoids power/power_orig from being 0 and | |
5721 | * causing divide-by-zero issues on boot. | |
5722 | * | |
5723 | * Runtime updates will correct power_orig. | |
5724 | */ | |
5725 | if (unlikely(!rq->sd)) { | |
5726 | power_orig += power_of(cpu); | |
5727 | power += power_of(cpu); | |
5728 | continue; | |
5729 | } | |
863bffc8 | 5730 | |
9abf24d4 SD |
5731 | sgp = rq->sd->groups->sgp; |
5732 | power_orig += sgp->power_orig; | |
5733 | power += sgp->power; | |
863bffc8 | 5734 | } |
74a5ce20 PZ |
5735 | } else { |
5736 | /* | |
5737 | * !SD_OVERLAP domains can assume that child groups | |
5738 | * span the current group. | |
5739 | */ | |
5740 | ||
5741 | group = child->groups; | |
5742 | do { | |
863bffc8 | 5743 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
5744 | power += group->sgp->power; |
5745 | group = group->next; | |
5746 | } while (group != child->groups); | |
5747 | } | |
1e3c88bd | 5748 | |
863bffc8 PZ |
5749 | sdg->sgp->power_orig = power_orig; |
5750 | sdg->sgp->power = power; | |
1e3c88bd PZ |
5751 | } |
5752 | ||
9d5efe05 SV |
5753 | /* |
5754 | * Try and fix up capacity for tiny siblings, this is needed when | |
5755 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
5756 | * which on its own isn't powerful enough. | |
5757 | * | |
5758 | * See update_sd_pick_busiest() and check_asym_packing(). | |
5759 | */ | |
5760 | static inline int | |
5761 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
5762 | { | |
5763 | /* | |
1399fa78 | 5764 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 5765 | */ |
a6c75f2f | 5766 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
5767 | return 0; |
5768 | ||
5769 | /* | |
5770 | * If ~90% of the cpu_power is still there, we're good. | |
5771 | */ | |
9c3f75cb | 5772 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
5773 | return 1; |
5774 | ||
5775 | return 0; | |
5776 | } | |
5777 | ||
30ce5dab PZ |
5778 | /* |
5779 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
5780 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
5781 | * | |
5782 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
5783 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
5784 | * Something like: | |
5785 | * | |
5786 | * { 0 1 2 3 } { 4 5 6 7 } | |
5787 | * * * * * | |
5788 | * | |
5789 | * If we were to balance group-wise we'd place two tasks in the first group and | |
5790 | * two tasks in the second group. Clearly this is undesired as it will overload | |
5791 | * cpu 3 and leave one of the cpus in the second group unused. | |
5792 | * | |
5793 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
5794 | * by noticing the lower domain failed to reach balance and had difficulty |
5795 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
5796 | * |
5797 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 5798 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 5799 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
5800 | * to create an effective group imbalance. |
5801 | * | |
5802 | * This is a somewhat tricky proposition since the next run might not find the | |
5803 | * group imbalance and decide the groups need to be balanced again. A most | |
5804 | * subtle and fragile situation. | |
5805 | */ | |
5806 | ||
6263322c | 5807 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 5808 | { |
6263322c | 5809 | return group->sgp->imbalance; |
30ce5dab PZ |
5810 | } |
5811 | ||
b37d9316 PZ |
5812 | /* |
5813 | * Compute the group capacity. | |
5814 | * | |
c61037e9 PZ |
5815 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
5816 | * first dividing out the smt factor and computing the actual number of cores | |
5817 | * and limit power unit capacity with that. | |
b37d9316 PZ |
5818 | */ |
5819 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
5820 | { | |
c61037e9 PZ |
5821 | unsigned int capacity, smt, cpus; |
5822 | unsigned int power, power_orig; | |
5823 | ||
5824 | power = group->sgp->power; | |
5825 | power_orig = group->sgp->power_orig; | |
5826 | cpus = group->group_weight; | |
b37d9316 | 5827 | |
c61037e9 PZ |
5828 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
5829 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
5830 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 5831 | |
c61037e9 | 5832 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
5833 | if (!capacity) |
5834 | capacity = fix_small_capacity(env->sd, group); | |
5835 | ||
5836 | return capacity; | |
5837 | } | |
5838 | ||
1e3c88bd PZ |
5839 | /** |
5840 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 5841 | * @env: The load balancing environment. |
1e3c88bd | 5842 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 5843 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 5844 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
5845 | * @sgs: variable to hold the statistics for this group. |
5846 | */ | |
bd939f45 PZ |
5847 | static inline void update_sg_lb_stats(struct lb_env *env, |
5848 | struct sched_group *group, int load_idx, | |
23f0d209 | 5849 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 5850 | { |
30ce5dab | 5851 | unsigned long load; |
bd939f45 | 5852 | int i; |
1e3c88bd | 5853 | |
b72ff13c PZ |
5854 | memset(sgs, 0, sizeof(*sgs)); |
5855 | ||
b9403130 | 5856 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
5857 | struct rq *rq = cpu_rq(i); |
5858 | ||
1e3c88bd | 5859 | /* Bias balancing toward cpus of our domain */ |
6263322c | 5860 | if (local_group) |
04f733b4 | 5861 | load = target_load(i, load_idx); |
6263322c | 5862 | else |
1e3c88bd | 5863 | load = source_load(i, load_idx); |
1e3c88bd PZ |
5864 | |
5865 | sgs->group_load += load; | |
380c9077 | 5866 | sgs->sum_nr_running += rq->nr_running; |
0ec8aa00 PZ |
5867 | #ifdef CONFIG_NUMA_BALANCING |
5868 | sgs->nr_numa_running += rq->nr_numa_running; | |
5869 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
5870 | #endif | |
1e3c88bd | 5871 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
5872 | if (idle_cpu(i)) |
5873 | sgs->idle_cpus++; | |
1e3c88bd PZ |
5874 | } |
5875 | ||
1e3c88bd | 5876 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
5877 | sgs->group_power = group->sgp->power; |
5878 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 5879 | |
dd5feea1 | 5880 | if (sgs->sum_nr_running) |
38d0f770 | 5881 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5882 | |
aae6d3dd | 5883 | sgs->group_weight = group->group_weight; |
fab47622 | 5884 | |
b37d9316 PZ |
5885 | sgs->group_imb = sg_imbalanced(group); |
5886 | sgs->group_capacity = sg_capacity(env, group); | |
5887 | ||
fab47622 NR |
5888 | if (sgs->group_capacity > sgs->sum_nr_running) |
5889 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5890 | } |
5891 | ||
532cb4c4 MN |
5892 | /** |
5893 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5894 | * @env: The load balancing environment. |
532cb4c4 MN |
5895 | * @sds: sched_domain statistics |
5896 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5897 | * @sgs: sched_group statistics |
532cb4c4 MN |
5898 | * |
5899 | * Determine if @sg is a busier group than the previously selected | |
5900 | * busiest group. | |
e69f6186 YB |
5901 | * |
5902 | * Return: %true if @sg is a busier group than the previously selected | |
5903 | * busiest group. %false otherwise. | |
532cb4c4 | 5904 | */ |
bd939f45 | 5905 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5906 | struct sd_lb_stats *sds, |
5907 | struct sched_group *sg, | |
bd939f45 | 5908 | struct sg_lb_stats *sgs) |
532cb4c4 | 5909 | { |
56cf515b | 5910 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5911 | return false; |
5912 | ||
5913 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5914 | return true; | |
5915 | ||
5916 | if (sgs->group_imb) | |
5917 | return true; | |
5918 | ||
5919 | /* | |
5920 | * ASYM_PACKING needs to move all the work to the lowest | |
5921 | * numbered CPUs in the group, therefore mark all groups | |
5922 | * higher than ourself as busy. | |
5923 | */ | |
bd939f45 PZ |
5924 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5925 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5926 | if (!sds->busiest) |
5927 | return true; | |
5928 | ||
5929 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5930 | return true; | |
5931 | } | |
5932 | ||
5933 | return false; | |
5934 | } | |
5935 | ||
0ec8aa00 PZ |
5936 | #ifdef CONFIG_NUMA_BALANCING |
5937 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5938 | { | |
5939 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
5940 | return regular; | |
5941 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
5942 | return remote; | |
5943 | return all; | |
5944 | } | |
5945 | ||
5946 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5947 | { | |
5948 | if (rq->nr_running > rq->nr_numa_running) | |
5949 | return regular; | |
5950 | if (rq->nr_running > rq->nr_preferred_running) | |
5951 | return remote; | |
5952 | return all; | |
5953 | } | |
5954 | #else | |
5955 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5956 | { | |
5957 | return all; | |
5958 | } | |
5959 | ||
5960 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5961 | { | |
5962 | return regular; | |
5963 | } | |
5964 | #endif /* CONFIG_NUMA_BALANCING */ | |
5965 | ||
1e3c88bd | 5966 | /** |
461819ac | 5967 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5968 | * @env: The load balancing environment. |
1e3c88bd PZ |
5969 | * @sds: variable to hold the statistics for this sched_domain. |
5970 | */ | |
0ec8aa00 | 5971 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5972 | { |
bd939f45 PZ |
5973 | struct sched_domain *child = env->sd->child; |
5974 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5975 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5976 | int load_idx, prefer_sibling = 0; |
5977 | ||
5978 | if (child && child->flags & SD_PREFER_SIBLING) | |
5979 | prefer_sibling = 1; | |
5980 | ||
bd939f45 | 5981 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5982 | |
5983 | do { | |
56cf515b | 5984 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5985 | int local_group; |
5986 | ||
bd939f45 | 5987 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5988 | if (local_group) { |
5989 | sds->local = sg; | |
5990 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5991 | |
5992 | if (env->idle != CPU_NEWLY_IDLE || | |
5993 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5994 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5995 | } |
1e3c88bd | 5996 | |
56cf515b | 5997 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5998 | |
b72ff13c PZ |
5999 | if (local_group) |
6000 | goto next_group; | |
6001 | ||
1e3c88bd PZ |
6002 | /* |
6003 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 6004 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
6005 | * and move all the excess tasks away. We lower the capacity |
6006 | * of a group only if the local group has the capacity to fit | |
6007 | * these excess tasks, i.e. nr_running < group_capacity. The | |
6008 | * extra check prevents the case where you always pull from the | |
6009 | * heaviest group when it is already under-utilized (possible | |
6010 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 6011 | */ |
b72ff13c PZ |
6012 | if (prefer_sibling && sds->local && |
6013 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 6014 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 6015 | |
b72ff13c | 6016 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 6017 | sds->busiest = sg; |
56cf515b | 6018 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
6019 | } |
6020 | ||
b72ff13c PZ |
6021 | next_group: |
6022 | /* Now, start updating sd_lb_stats */ | |
6023 | sds->total_load += sgs->group_load; | |
6024 | sds->total_pwr += sgs->group_power; | |
6025 | ||
532cb4c4 | 6026 | sg = sg->next; |
bd939f45 | 6027 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
6028 | |
6029 | if (env->sd->flags & SD_NUMA) | |
6030 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
532cb4c4 MN |
6031 | } |
6032 | ||
532cb4c4 MN |
6033 | /** |
6034 | * check_asym_packing - Check to see if the group is packed into the | |
6035 | * sched doman. | |
6036 | * | |
6037 | * This is primarily intended to used at the sibling level. Some | |
6038 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
6039 | * case of POWER7, it can move to lower SMT modes only when higher | |
6040 | * threads are idle. When in lower SMT modes, the threads will | |
6041 | * perform better since they share less core resources. Hence when we | |
6042 | * have idle threads, we want them to be the higher ones. | |
6043 | * | |
6044 | * This packing function is run on idle threads. It checks to see if | |
6045 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
6046 | * CPU number than the packing function is being run on. Here we are | |
6047 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
6048 | * number. | |
6049 | * | |
e69f6186 | 6050 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
6051 | * this CPU. The amount of the imbalance is returned in *imbalance. |
6052 | * | |
cd96891d | 6053 | * @env: The load balancing environment. |
532cb4c4 | 6054 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 6055 | */ |
bd939f45 | 6056 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
6057 | { |
6058 | int busiest_cpu; | |
6059 | ||
bd939f45 | 6060 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
6061 | return 0; |
6062 | ||
6063 | if (!sds->busiest) | |
6064 | return 0; | |
6065 | ||
6066 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 6067 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
6068 | return 0; |
6069 | ||
bd939f45 | 6070 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
6071 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
6072 | SCHED_POWER_SCALE); | |
bd939f45 | 6073 | |
532cb4c4 | 6074 | return 1; |
1e3c88bd PZ |
6075 | } |
6076 | ||
6077 | /** | |
6078 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
6079 | * amongst the groups of a sched_domain, during | |
6080 | * load balancing. | |
cd96891d | 6081 | * @env: The load balancing environment. |
1e3c88bd | 6082 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6083 | */ |
bd939f45 PZ |
6084 | static inline |
6085 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
6086 | { |
6087 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
6088 | unsigned int imbn = 2; | |
dd5feea1 | 6089 | unsigned long scaled_busy_load_per_task; |
56cf515b | 6090 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 6091 | |
56cf515b JK |
6092 | local = &sds->local_stat; |
6093 | busiest = &sds->busiest_stat; | |
1e3c88bd | 6094 | |
56cf515b JK |
6095 | if (!local->sum_nr_running) |
6096 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
6097 | else if (busiest->load_per_task > local->load_per_task) | |
6098 | imbn = 1; | |
dd5feea1 | 6099 | |
56cf515b JK |
6100 | scaled_busy_load_per_task = |
6101 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 6102 | busiest->group_power; |
56cf515b | 6103 | |
3029ede3 VD |
6104 | if (busiest->avg_load + scaled_busy_load_per_task >= |
6105 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 6106 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6107 | return; |
6108 | } | |
6109 | ||
6110 | /* | |
6111 | * OK, we don't have enough imbalance to justify moving tasks, | |
6112 | * however we may be able to increase total CPU power used by | |
6113 | * moving them. | |
6114 | */ | |
6115 | ||
3ae11c90 | 6116 | pwr_now += busiest->group_power * |
56cf515b | 6117 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 6118 | pwr_now += local->group_power * |
56cf515b | 6119 | min(local->load_per_task, local->avg_load); |
1399fa78 | 6120 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
6121 | |
6122 | /* Amount of load we'd subtract */ | |
a2cd4260 | 6123 | if (busiest->avg_load > scaled_busy_load_per_task) { |
3ae11c90 | 6124 | pwr_move += busiest->group_power * |
56cf515b | 6125 | min(busiest->load_per_task, |
a2cd4260 | 6126 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 6127 | } |
1e3c88bd PZ |
6128 | |
6129 | /* Amount of load we'd add */ | |
3ae11c90 | 6130 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 6131 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
6132 | tmp = (busiest->avg_load * busiest->group_power) / |
6133 | local->group_power; | |
56cf515b JK |
6134 | } else { |
6135 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 6136 | local->group_power; |
56cf515b | 6137 | } |
3ae11c90 PZ |
6138 | pwr_move += local->group_power * |
6139 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 6140 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
6141 | |
6142 | /* Move if we gain throughput */ | |
6143 | if (pwr_move > pwr_now) | |
56cf515b | 6144 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6145 | } |
6146 | ||
6147 | /** | |
6148 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
6149 | * groups of a given sched_domain during load balance. | |
bd939f45 | 6150 | * @env: load balance environment |
1e3c88bd | 6151 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6152 | */ |
bd939f45 | 6153 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6154 | { |
dd5feea1 | 6155 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
6156 | struct sg_lb_stats *local, *busiest; |
6157 | ||
6158 | local = &sds->local_stat; | |
56cf515b | 6159 | busiest = &sds->busiest_stat; |
dd5feea1 | 6160 | |
56cf515b | 6161 | if (busiest->group_imb) { |
30ce5dab PZ |
6162 | /* |
6163 | * In the group_imb case we cannot rely on group-wide averages | |
6164 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
6165 | */ | |
56cf515b JK |
6166 | busiest->load_per_task = |
6167 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
6168 | } |
6169 | ||
1e3c88bd PZ |
6170 | /* |
6171 | * In the presence of smp nice balancing, certain scenarios can have | |
6172 | * max load less than avg load(as we skip the groups at or below | |
6173 | * its cpu_power, while calculating max_load..) | |
6174 | */ | |
b1885550 VD |
6175 | if (busiest->avg_load <= sds->avg_load || |
6176 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
6177 | env->imbalance = 0; |
6178 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
6179 | } |
6180 | ||
56cf515b | 6181 | if (!busiest->group_imb) { |
dd5feea1 SS |
6182 | /* |
6183 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
6184 | * Except of course for the group_imb case, since then we might |
6185 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 6186 | */ |
56cf515b JK |
6187 | load_above_capacity = |
6188 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 6189 | |
1399fa78 | 6190 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 6191 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
6192 | } |
6193 | ||
6194 | /* | |
6195 | * We're trying to get all the cpus to the average_load, so we don't | |
6196 | * want to push ourselves above the average load, nor do we wish to | |
6197 | * reduce the max loaded cpu below the average load. At the same time, | |
6198 | * we also don't want to reduce the group load below the group capacity | |
6199 | * (so that we can implement power-savings policies etc). Thus we look | |
6200 | * for the minimum possible imbalance. | |
dd5feea1 | 6201 | */ |
30ce5dab | 6202 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
6203 | |
6204 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 6205 | env->imbalance = min( |
3ae11c90 PZ |
6206 | max_pull * busiest->group_power, |
6207 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 6208 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
6209 | |
6210 | /* | |
6211 | * if *imbalance is less than the average load per runnable task | |
25985edc | 6212 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
6213 | * a think about bumping its value to force at least one task to be |
6214 | * moved | |
6215 | */ | |
56cf515b | 6216 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 6217 | return fix_small_imbalance(env, sds); |
1e3c88bd | 6218 | } |
fab47622 | 6219 | |
1e3c88bd PZ |
6220 | /******* find_busiest_group() helpers end here *********************/ |
6221 | ||
6222 | /** | |
6223 | * find_busiest_group - Returns the busiest group within the sched_domain | |
6224 | * if there is an imbalance. If there isn't an imbalance, and | |
6225 | * the user has opted for power-savings, it returns a group whose | |
6226 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
6227 | * such a group exists. | |
6228 | * | |
6229 | * Also calculates the amount of weighted load which should be moved | |
6230 | * to restore balance. | |
6231 | * | |
cd96891d | 6232 | * @env: The load balancing environment. |
1e3c88bd | 6233 | * |
e69f6186 | 6234 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
6235 | * - If no imbalance and user has opted for power-savings balance, |
6236 | * return the least loaded group whose CPUs can be | |
6237 | * put to idle by rebalancing its tasks onto our group. | |
6238 | */ | |
56cf515b | 6239 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 6240 | { |
56cf515b | 6241 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
6242 | struct sd_lb_stats sds; |
6243 | ||
147c5fc2 | 6244 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
6245 | |
6246 | /* | |
6247 | * Compute the various statistics relavent for load balancing at | |
6248 | * this level. | |
6249 | */ | |
23f0d209 | 6250 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
6251 | local = &sds.local_stat; |
6252 | busiest = &sds.busiest_stat; | |
1e3c88bd | 6253 | |
bd939f45 PZ |
6254 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
6255 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
6256 | return sds.busiest; |
6257 | ||
cc57aa8f | 6258 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 6259 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
6260 | goto out_balanced; |
6261 | ||
1399fa78 | 6262 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 6263 | |
866ab43e PZ |
6264 | /* |
6265 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 6266 | * work because they assume all things are equal, which typically |
866ab43e PZ |
6267 | * isn't true due to cpus_allowed constraints and the like. |
6268 | */ | |
56cf515b | 6269 | if (busiest->group_imb) |
866ab43e PZ |
6270 | goto force_balance; |
6271 | ||
cc57aa8f | 6272 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
6273 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
6274 | !busiest->group_has_capacity) | |
fab47622 NR |
6275 | goto force_balance; |
6276 | ||
cc57aa8f PZ |
6277 | /* |
6278 | * If the local group is more busy than the selected busiest group | |
6279 | * don't try and pull any tasks. | |
6280 | */ | |
56cf515b | 6281 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
6282 | goto out_balanced; |
6283 | ||
cc57aa8f PZ |
6284 | /* |
6285 | * Don't pull any tasks if this group is already above the domain | |
6286 | * average load. | |
6287 | */ | |
56cf515b | 6288 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
6289 | goto out_balanced; |
6290 | ||
bd939f45 | 6291 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
6292 | /* |
6293 | * This cpu is idle. If the busiest group load doesn't | |
6294 | * have more tasks than the number of available cpu's and | |
6295 | * there is no imbalance between this and busiest group | |
6296 | * wrt to idle cpu's, it is balanced. | |
6297 | */ | |
56cf515b JK |
6298 | if ((local->idle_cpus < busiest->idle_cpus) && |
6299 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 6300 | goto out_balanced; |
c186fafe PZ |
6301 | } else { |
6302 | /* | |
6303 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
6304 | * imbalance_pct to be conservative. | |
6305 | */ | |
56cf515b JK |
6306 | if (100 * busiest->avg_load <= |
6307 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 6308 | goto out_balanced; |
aae6d3dd | 6309 | } |
1e3c88bd | 6310 | |
fab47622 | 6311 | force_balance: |
1e3c88bd | 6312 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 6313 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
6314 | return sds.busiest; |
6315 | ||
6316 | out_balanced: | |
bd939f45 | 6317 | env->imbalance = 0; |
1e3c88bd PZ |
6318 | return NULL; |
6319 | } | |
6320 | ||
6321 | /* | |
6322 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
6323 | */ | |
bd939f45 | 6324 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 6325 | struct sched_group *group) |
1e3c88bd PZ |
6326 | { |
6327 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 6328 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
6329 | int i; |
6330 | ||
6906a408 | 6331 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
0ec8aa00 PZ |
6332 | unsigned long power, capacity, wl; |
6333 | enum fbq_type rt; | |
6334 | ||
6335 | rq = cpu_rq(i); | |
6336 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 6337 | |
0ec8aa00 PZ |
6338 | /* |
6339 | * We classify groups/runqueues into three groups: | |
6340 | * - regular: there are !numa tasks | |
6341 | * - remote: there are numa tasks that run on the 'wrong' node | |
6342 | * - all: there is no distinction | |
6343 | * | |
6344 | * In order to avoid migrating ideally placed numa tasks, | |
6345 | * ignore those when there's better options. | |
6346 | * | |
6347 | * If we ignore the actual busiest queue to migrate another | |
6348 | * task, the next balance pass can still reduce the busiest | |
6349 | * queue by moving tasks around inside the node. | |
6350 | * | |
6351 | * If we cannot move enough load due to this classification | |
6352 | * the next pass will adjust the group classification and | |
6353 | * allow migration of more tasks. | |
6354 | * | |
6355 | * Both cases only affect the total convergence complexity. | |
6356 | */ | |
6357 | if (rt > env->fbq_type) | |
6358 | continue; | |
6359 | ||
6360 | power = power_of(i); | |
6361 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); | |
9d5efe05 | 6362 | if (!capacity) |
bd939f45 | 6363 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 6364 | |
6e40f5bb | 6365 | wl = weighted_cpuload(i); |
1e3c88bd | 6366 | |
6e40f5bb TG |
6367 | /* |
6368 | * When comparing with imbalance, use weighted_cpuload() | |
6369 | * which is not scaled with the cpu power. | |
6370 | */ | |
bd939f45 | 6371 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
6372 | continue; |
6373 | ||
6e40f5bb TG |
6374 | /* |
6375 | * For the load comparisons with the other cpu's, consider | |
6376 | * the weighted_cpuload() scaled with the cpu power, so that | |
6377 | * the load can be moved away from the cpu that is potentially | |
6378 | * running at a lower capacity. | |
95a79b80 JK |
6379 | * |
6380 | * Thus we're looking for max(wl_i / power_i), crosswise | |
6381 | * multiplication to rid ourselves of the division works out | |
6382 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
6383 | * previous maximum. | |
6e40f5bb | 6384 | */ |
95a79b80 JK |
6385 | if (wl * busiest_power > busiest_load * power) { |
6386 | busiest_load = wl; | |
6387 | busiest_power = power; | |
1e3c88bd PZ |
6388 | busiest = rq; |
6389 | } | |
6390 | } | |
6391 | ||
6392 | return busiest; | |
6393 | } | |
6394 | ||
6395 | /* | |
6396 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
6397 | * so long as it is large enough. | |
6398 | */ | |
6399 | #define MAX_PINNED_INTERVAL 512 | |
6400 | ||
6401 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 6402 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 6403 | |
bd939f45 | 6404 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 6405 | { |
bd939f45 PZ |
6406 | struct sched_domain *sd = env->sd; |
6407 | ||
6408 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
6409 | |
6410 | /* | |
6411 | * ASYM_PACKING needs to force migrate tasks from busy but | |
6412 | * higher numbered CPUs in order to pack all tasks in the | |
6413 | * lowest numbered CPUs. | |
6414 | */ | |
bd939f45 | 6415 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 6416 | return 1; |
1af3ed3d PZ |
6417 | } |
6418 | ||
6419 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
6420 | } | |
6421 | ||
969c7921 TH |
6422 | static int active_load_balance_cpu_stop(void *data); |
6423 | ||
23f0d209 JK |
6424 | static int should_we_balance(struct lb_env *env) |
6425 | { | |
6426 | struct sched_group *sg = env->sd->groups; | |
6427 | struct cpumask *sg_cpus, *sg_mask; | |
6428 | int cpu, balance_cpu = -1; | |
6429 | ||
6430 | /* | |
6431 | * In the newly idle case, we will allow all the cpu's | |
6432 | * to do the newly idle load balance. | |
6433 | */ | |
6434 | if (env->idle == CPU_NEWLY_IDLE) | |
6435 | return 1; | |
6436 | ||
6437 | sg_cpus = sched_group_cpus(sg); | |
6438 | sg_mask = sched_group_mask(sg); | |
6439 | /* Try to find first idle cpu */ | |
6440 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
6441 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
6442 | continue; | |
6443 | ||
6444 | balance_cpu = cpu; | |
6445 | break; | |
6446 | } | |
6447 | ||
6448 | if (balance_cpu == -1) | |
6449 | balance_cpu = group_balance_cpu(sg); | |
6450 | ||
6451 | /* | |
6452 | * First idle cpu or the first cpu(busiest) in this sched group | |
6453 | * is eligible for doing load balancing at this and above domains. | |
6454 | */ | |
b0cff9d8 | 6455 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
6456 | } |
6457 | ||
1e3c88bd PZ |
6458 | /* |
6459 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
6460 | * tasks if there is an imbalance. | |
6461 | */ | |
6462 | static int load_balance(int this_cpu, struct rq *this_rq, | |
6463 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 6464 | int *continue_balancing) |
1e3c88bd | 6465 | { |
88b8dac0 | 6466 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 6467 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 6468 | struct sched_group *group; |
1e3c88bd PZ |
6469 | struct rq *busiest; |
6470 | unsigned long flags; | |
e6252c3e | 6471 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 6472 | |
8e45cb54 PZ |
6473 | struct lb_env env = { |
6474 | .sd = sd, | |
ddcdf6e7 PZ |
6475 | .dst_cpu = this_cpu, |
6476 | .dst_rq = this_rq, | |
88b8dac0 | 6477 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 6478 | .idle = idle, |
eb95308e | 6479 | .loop_break = sched_nr_migrate_break, |
b9403130 | 6480 | .cpus = cpus, |
0ec8aa00 | 6481 | .fbq_type = all, |
8e45cb54 PZ |
6482 | }; |
6483 | ||
cfc03118 JK |
6484 | /* |
6485 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
6486 | * other cpus in our group | |
6487 | */ | |
e02e60c1 | 6488 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 6489 | env.dst_grpmask = NULL; |
cfc03118 | 6490 | |
1e3c88bd PZ |
6491 | cpumask_copy(cpus, cpu_active_mask); |
6492 | ||
1e3c88bd PZ |
6493 | schedstat_inc(sd, lb_count[idle]); |
6494 | ||
6495 | redo: | |
23f0d209 JK |
6496 | if (!should_we_balance(&env)) { |
6497 | *continue_balancing = 0; | |
1e3c88bd | 6498 | goto out_balanced; |
23f0d209 | 6499 | } |
1e3c88bd | 6500 | |
23f0d209 | 6501 | group = find_busiest_group(&env); |
1e3c88bd PZ |
6502 | if (!group) { |
6503 | schedstat_inc(sd, lb_nobusyg[idle]); | |
6504 | goto out_balanced; | |
6505 | } | |
6506 | ||
b9403130 | 6507 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
6508 | if (!busiest) { |
6509 | schedstat_inc(sd, lb_nobusyq[idle]); | |
6510 | goto out_balanced; | |
6511 | } | |
6512 | ||
78feefc5 | 6513 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 6514 | |
bd939f45 | 6515 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
6516 | |
6517 | ld_moved = 0; | |
6518 | if (busiest->nr_running > 1) { | |
6519 | /* | |
6520 | * Attempt to move tasks. If find_busiest_group has found | |
6521 | * an imbalance but busiest->nr_running <= 1, the group is | |
6522 | * still unbalanced. ld_moved simply stays zero, so it is | |
6523 | * correctly treated as an imbalance. | |
6524 | */ | |
8e45cb54 | 6525 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
6526 | env.src_cpu = busiest->cpu; |
6527 | env.src_rq = busiest; | |
6528 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 6529 | |
5d6523eb | 6530 | more_balance: |
1e3c88bd | 6531 | local_irq_save(flags); |
78feefc5 | 6532 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
6533 | |
6534 | /* | |
6535 | * cur_ld_moved - load moved in current iteration | |
6536 | * ld_moved - cumulative load moved across iterations | |
6537 | */ | |
6538 | cur_ld_moved = move_tasks(&env); | |
6539 | ld_moved += cur_ld_moved; | |
78feefc5 | 6540 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
6541 | local_irq_restore(flags); |
6542 | ||
6543 | /* | |
6544 | * some other cpu did the load balance for us. | |
6545 | */ | |
88b8dac0 SV |
6546 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
6547 | resched_cpu(env.dst_cpu); | |
6548 | ||
f1cd0858 JK |
6549 | if (env.flags & LBF_NEED_BREAK) { |
6550 | env.flags &= ~LBF_NEED_BREAK; | |
6551 | goto more_balance; | |
6552 | } | |
6553 | ||
88b8dac0 SV |
6554 | /* |
6555 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
6556 | * us and move them to an alternate dst_cpu in our sched_group | |
6557 | * where they can run. The upper limit on how many times we | |
6558 | * iterate on same src_cpu is dependent on number of cpus in our | |
6559 | * sched_group. | |
6560 | * | |
6561 | * This changes load balance semantics a bit on who can move | |
6562 | * load to a given_cpu. In addition to the given_cpu itself | |
6563 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
6564 | * nohz-idle), we now have balance_cpu in a position to move | |
6565 | * load to given_cpu. In rare situations, this may cause | |
6566 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
6567 | * _independently_ and at _same_ time to move some load to | |
6568 | * given_cpu) causing exceess load to be moved to given_cpu. | |
6569 | * This however should not happen so much in practice and | |
6570 | * moreover subsequent load balance cycles should correct the | |
6571 | * excess load moved. | |
6572 | */ | |
6263322c | 6573 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 6574 | |
7aff2e3a VD |
6575 | /* Prevent to re-select dst_cpu via env's cpus */ |
6576 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
6577 | ||
78feefc5 | 6578 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 6579 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 6580 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
6581 | env.loop = 0; |
6582 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 6583 | |
88b8dac0 SV |
6584 | /* |
6585 | * Go back to "more_balance" rather than "redo" since we | |
6586 | * need to continue with same src_cpu. | |
6587 | */ | |
6588 | goto more_balance; | |
6589 | } | |
1e3c88bd | 6590 | |
6263322c PZ |
6591 | /* |
6592 | * We failed to reach balance because of affinity. | |
6593 | */ | |
6594 | if (sd_parent) { | |
6595 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
6596 | ||
6597 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
6598 | *group_imbalance = 1; | |
6599 | } else if (*group_imbalance) | |
6600 | *group_imbalance = 0; | |
6601 | } | |
6602 | ||
1e3c88bd | 6603 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 6604 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 6605 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
6606 | if (!cpumask_empty(cpus)) { |
6607 | env.loop = 0; | |
6608 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 6609 | goto redo; |
bbf18b19 | 6610 | } |
1e3c88bd PZ |
6611 | goto out_balanced; |
6612 | } | |
6613 | } | |
6614 | ||
6615 | if (!ld_moved) { | |
6616 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
6617 | /* |
6618 | * Increment the failure counter only on periodic balance. | |
6619 | * We do not want newidle balance, which can be very | |
6620 | * frequent, pollute the failure counter causing | |
6621 | * excessive cache_hot migrations and active balances. | |
6622 | */ | |
6623 | if (idle != CPU_NEWLY_IDLE) | |
6624 | sd->nr_balance_failed++; | |
1e3c88bd | 6625 | |
bd939f45 | 6626 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
6627 | raw_spin_lock_irqsave(&busiest->lock, flags); |
6628 | ||
969c7921 TH |
6629 | /* don't kick the active_load_balance_cpu_stop, |
6630 | * if the curr task on busiest cpu can't be | |
6631 | * moved to this_cpu | |
1e3c88bd PZ |
6632 | */ |
6633 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 6634 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
6635 | raw_spin_unlock_irqrestore(&busiest->lock, |
6636 | flags); | |
8e45cb54 | 6637 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
6638 | goto out_one_pinned; |
6639 | } | |
6640 | ||
969c7921 TH |
6641 | /* |
6642 | * ->active_balance synchronizes accesses to | |
6643 | * ->active_balance_work. Once set, it's cleared | |
6644 | * only after active load balance is finished. | |
6645 | */ | |
1e3c88bd PZ |
6646 | if (!busiest->active_balance) { |
6647 | busiest->active_balance = 1; | |
6648 | busiest->push_cpu = this_cpu; | |
6649 | active_balance = 1; | |
6650 | } | |
6651 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 6652 | |
bd939f45 | 6653 | if (active_balance) { |
969c7921 TH |
6654 | stop_one_cpu_nowait(cpu_of(busiest), |
6655 | active_load_balance_cpu_stop, busiest, | |
6656 | &busiest->active_balance_work); | |
bd939f45 | 6657 | } |
1e3c88bd PZ |
6658 | |
6659 | /* | |
6660 | * We've kicked active balancing, reset the failure | |
6661 | * counter. | |
6662 | */ | |
6663 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
6664 | } | |
6665 | } else | |
6666 | sd->nr_balance_failed = 0; | |
6667 | ||
6668 | if (likely(!active_balance)) { | |
6669 | /* We were unbalanced, so reset the balancing interval */ | |
6670 | sd->balance_interval = sd->min_interval; | |
6671 | } else { | |
6672 | /* | |
6673 | * If we've begun active balancing, start to back off. This | |
6674 | * case may not be covered by the all_pinned logic if there | |
6675 | * is only 1 task on the busy runqueue (because we don't call | |
6676 | * move_tasks). | |
6677 | */ | |
6678 | if (sd->balance_interval < sd->max_interval) | |
6679 | sd->balance_interval *= 2; | |
6680 | } | |
6681 | ||
1e3c88bd PZ |
6682 | goto out; |
6683 | ||
6684 | out_balanced: | |
6685 | schedstat_inc(sd, lb_balanced[idle]); | |
6686 | ||
6687 | sd->nr_balance_failed = 0; | |
6688 | ||
6689 | out_one_pinned: | |
6690 | /* tune up the balancing interval */ | |
8e45cb54 | 6691 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 6692 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
6693 | (sd->balance_interval < sd->max_interval)) |
6694 | sd->balance_interval *= 2; | |
6695 | ||
46e49b38 | 6696 | ld_moved = 0; |
1e3c88bd | 6697 | out: |
1e3c88bd PZ |
6698 | return ld_moved; |
6699 | } | |
6700 | ||
52a08ef1 JL |
6701 | static inline unsigned long |
6702 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
6703 | { | |
6704 | unsigned long interval = sd->balance_interval; | |
6705 | ||
6706 | if (cpu_busy) | |
6707 | interval *= sd->busy_factor; | |
6708 | ||
6709 | /* scale ms to jiffies */ | |
6710 | interval = msecs_to_jiffies(interval); | |
6711 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
6712 | ||
6713 | return interval; | |
6714 | } | |
6715 | ||
6716 | static inline void | |
6717 | update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) | |
6718 | { | |
6719 | unsigned long interval, next; | |
6720 | ||
6721 | interval = get_sd_balance_interval(sd, cpu_busy); | |
6722 | next = sd->last_balance + interval; | |
6723 | ||
6724 | if (time_after(*next_balance, next)) | |
6725 | *next_balance = next; | |
6726 | } | |
6727 | ||
1e3c88bd PZ |
6728 | /* |
6729 | * idle_balance is called by schedule() if this_cpu is about to become | |
6730 | * idle. Attempts to pull tasks from other CPUs. | |
6731 | */ | |
6e83125c | 6732 | static int idle_balance(struct rq *this_rq) |
1e3c88bd | 6733 | { |
52a08ef1 JL |
6734 | unsigned long next_balance = jiffies + HZ; |
6735 | int this_cpu = this_rq->cpu; | |
1e3c88bd PZ |
6736 | struct sched_domain *sd; |
6737 | int pulled_task = 0; | |
9bd721c5 | 6738 | u64 curr_cost = 0; |
1e3c88bd | 6739 | |
6e83125c | 6740 | idle_enter_fair(this_rq); |
0e5b5337 | 6741 | |
6e83125c PZ |
6742 | /* |
6743 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
6744 | * measure the duration of idle_balance() as idle time. | |
6745 | */ | |
6746 | this_rq->idle_stamp = rq_clock(this_rq); | |
6747 | ||
52a08ef1 JL |
6748 | if (this_rq->avg_idle < sysctl_sched_migration_cost) { |
6749 | rcu_read_lock(); | |
6750 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
6751 | if (sd) | |
6752 | update_next_balance(sd, 0, &next_balance); | |
6753 | rcu_read_unlock(); | |
6754 | ||
6e83125c | 6755 | goto out; |
52a08ef1 | 6756 | } |
1e3c88bd | 6757 | |
f492e12e PZ |
6758 | /* |
6759 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
6760 | */ | |
6761 | raw_spin_unlock(&this_rq->lock); | |
6762 | ||
48a16753 | 6763 | update_blocked_averages(this_cpu); |
dce840a0 | 6764 | rcu_read_lock(); |
1e3c88bd | 6765 | for_each_domain(this_cpu, sd) { |
23f0d209 | 6766 | int continue_balancing = 1; |
9bd721c5 | 6767 | u64 t0, domain_cost; |
1e3c88bd PZ |
6768 | |
6769 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
6770 | continue; | |
6771 | ||
52a08ef1 JL |
6772 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
6773 | update_next_balance(sd, 0, &next_balance); | |
9bd721c5 | 6774 | break; |
52a08ef1 | 6775 | } |
9bd721c5 | 6776 | |
f492e12e | 6777 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
6778 | t0 = sched_clock_cpu(this_cpu); |
6779 | ||
f492e12e | 6780 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
6781 | sd, CPU_NEWLY_IDLE, |
6782 | &continue_balancing); | |
9bd721c5 JL |
6783 | |
6784 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
6785 | if (domain_cost > sd->max_newidle_lb_cost) | |
6786 | sd->max_newidle_lb_cost = domain_cost; | |
6787 | ||
6788 | curr_cost += domain_cost; | |
f492e12e | 6789 | } |
1e3c88bd | 6790 | |
52a08ef1 | 6791 | update_next_balance(sd, 0, &next_balance); |
39a4d9ca JL |
6792 | |
6793 | /* | |
6794 | * Stop searching for tasks to pull if there are | |
6795 | * now runnable tasks on this rq. | |
6796 | */ | |
6797 | if (pulled_task || this_rq->nr_running > 0) | |
1e3c88bd | 6798 | break; |
1e3c88bd | 6799 | } |
dce840a0 | 6800 | rcu_read_unlock(); |
f492e12e PZ |
6801 | |
6802 | raw_spin_lock(&this_rq->lock); | |
6803 | ||
0e5b5337 JL |
6804 | if (curr_cost > this_rq->max_idle_balance_cost) |
6805 | this_rq->max_idle_balance_cost = curr_cost; | |
6806 | ||
e5fc6611 | 6807 | /* |
0e5b5337 JL |
6808 | * While browsing the domains, we released the rq lock, a task could |
6809 | * have been enqueued in the meantime. Since we're not going idle, | |
6810 | * pretend we pulled a task. | |
e5fc6611 | 6811 | */ |
0e5b5337 | 6812 | if (this_rq->cfs.h_nr_running && !pulled_task) |
6e83125c | 6813 | pulled_task = 1; |
e5fc6611 | 6814 | |
52a08ef1 JL |
6815 | out: |
6816 | /* Move the next balance forward */ | |
6817 | if (time_after(this_rq->next_balance, next_balance)) | |
1e3c88bd | 6818 | this_rq->next_balance = next_balance; |
9bd721c5 | 6819 | |
e4aa358b | 6820 | /* Is there a task of a high priority class? */ |
46383648 | 6821 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
e4aa358b KT |
6822 | pulled_task = -1; |
6823 | ||
6824 | if (pulled_task) { | |
6825 | idle_exit_fair(this_rq); | |
6e83125c | 6826 | this_rq->idle_stamp = 0; |
e4aa358b | 6827 | } |
6e83125c | 6828 | |
3c4017c1 | 6829 | return pulled_task; |
1e3c88bd PZ |
6830 | } |
6831 | ||
6832 | /* | |
969c7921 TH |
6833 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
6834 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
6835 | * least 1 task to be running on each physical CPU where possible, and | |
6836 | * avoids physical / logical imbalances. | |
1e3c88bd | 6837 | */ |
969c7921 | 6838 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 6839 | { |
969c7921 TH |
6840 | struct rq *busiest_rq = data; |
6841 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 6842 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 6843 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 6844 | struct sched_domain *sd; |
969c7921 TH |
6845 | |
6846 | raw_spin_lock_irq(&busiest_rq->lock); | |
6847 | ||
6848 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
6849 | if (unlikely(busiest_cpu != smp_processor_id() || | |
6850 | !busiest_rq->active_balance)) | |
6851 | goto out_unlock; | |
1e3c88bd PZ |
6852 | |
6853 | /* Is there any task to move? */ | |
6854 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 6855 | goto out_unlock; |
1e3c88bd PZ |
6856 | |
6857 | /* | |
6858 | * This condition is "impossible", if it occurs | |
6859 | * we need to fix it. Originally reported by | |
6860 | * Bjorn Helgaas on a 128-cpu setup. | |
6861 | */ | |
6862 | BUG_ON(busiest_rq == target_rq); | |
6863 | ||
6864 | /* move a task from busiest_rq to target_rq */ | |
6865 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
6866 | |
6867 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 6868 | rcu_read_lock(); |
1e3c88bd PZ |
6869 | for_each_domain(target_cpu, sd) { |
6870 | if ((sd->flags & SD_LOAD_BALANCE) && | |
6871 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
6872 | break; | |
6873 | } | |
6874 | ||
6875 | if (likely(sd)) { | |
8e45cb54 PZ |
6876 | struct lb_env env = { |
6877 | .sd = sd, | |
ddcdf6e7 PZ |
6878 | .dst_cpu = target_cpu, |
6879 | .dst_rq = target_rq, | |
6880 | .src_cpu = busiest_rq->cpu, | |
6881 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
6882 | .idle = CPU_IDLE, |
6883 | }; | |
6884 | ||
1e3c88bd PZ |
6885 | schedstat_inc(sd, alb_count); |
6886 | ||
8e45cb54 | 6887 | if (move_one_task(&env)) |
1e3c88bd PZ |
6888 | schedstat_inc(sd, alb_pushed); |
6889 | else | |
6890 | schedstat_inc(sd, alb_failed); | |
6891 | } | |
dce840a0 | 6892 | rcu_read_unlock(); |
1e3c88bd | 6893 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
6894 | out_unlock: |
6895 | busiest_rq->active_balance = 0; | |
6896 | raw_spin_unlock_irq(&busiest_rq->lock); | |
6897 | return 0; | |
1e3c88bd PZ |
6898 | } |
6899 | ||
d987fc7f MG |
6900 | static inline int on_null_domain(struct rq *rq) |
6901 | { | |
6902 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
6903 | } | |
6904 | ||
3451d024 | 6905 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
6906 | /* |
6907 | * idle load balancing details | |
83cd4fe2 VP |
6908 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
6909 | * needed, they will kick the idle load balancer, which then does idle | |
6910 | * load balancing for all the idle CPUs. | |
6911 | */ | |
1e3c88bd | 6912 | static struct { |
83cd4fe2 | 6913 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 6914 | atomic_t nr_cpus; |
83cd4fe2 VP |
6915 | unsigned long next_balance; /* in jiffy units */ |
6916 | } nohz ____cacheline_aligned; | |
1e3c88bd | 6917 | |
3dd0337d | 6918 | static inline int find_new_ilb(void) |
1e3c88bd | 6919 | { |
0b005cf5 | 6920 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 6921 | |
786d6dc7 SS |
6922 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
6923 | return ilb; | |
6924 | ||
6925 | return nr_cpu_ids; | |
1e3c88bd | 6926 | } |
1e3c88bd | 6927 | |
83cd4fe2 VP |
6928 | /* |
6929 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
6930 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
6931 | * CPU (if there is one). | |
6932 | */ | |
0aeeeeba | 6933 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
6934 | { |
6935 | int ilb_cpu; | |
6936 | ||
6937 | nohz.next_balance++; | |
6938 | ||
3dd0337d | 6939 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 6940 | |
0b005cf5 SS |
6941 | if (ilb_cpu >= nr_cpu_ids) |
6942 | return; | |
83cd4fe2 | 6943 | |
cd490c5b | 6944 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
6945 | return; |
6946 | /* | |
6947 | * Use smp_send_reschedule() instead of resched_cpu(). | |
6948 | * This way we generate a sched IPI on the target cpu which | |
6949 | * is idle. And the softirq performing nohz idle load balance | |
6950 | * will be run before returning from the IPI. | |
6951 | */ | |
6952 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
6953 | return; |
6954 | } | |
6955 | ||
c1cc017c | 6956 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
6957 | { |
6958 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
d987fc7f MG |
6959 | /* |
6960 | * Completely isolated CPUs don't ever set, so we must test. | |
6961 | */ | |
6962 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | |
6963 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
6964 | atomic_dec(&nohz.nr_cpus); | |
6965 | } | |
71325960 SS |
6966 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
6967 | } | |
6968 | } | |
6969 | ||
69e1e811 SS |
6970 | static inline void set_cpu_sd_state_busy(void) |
6971 | { | |
6972 | struct sched_domain *sd; | |
37dc6b50 | 6973 | int cpu = smp_processor_id(); |
69e1e811 | 6974 | |
69e1e811 | 6975 | rcu_read_lock(); |
37dc6b50 | 6976 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
6977 | |
6978 | if (!sd || !sd->nohz_idle) | |
6979 | goto unlock; | |
6980 | sd->nohz_idle = 0; | |
6981 | ||
37dc6b50 | 6982 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6983 | unlock: |
69e1e811 SS |
6984 | rcu_read_unlock(); |
6985 | } | |
6986 | ||
6987 | void set_cpu_sd_state_idle(void) | |
6988 | { | |
6989 | struct sched_domain *sd; | |
37dc6b50 | 6990 | int cpu = smp_processor_id(); |
69e1e811 | 6991 | |
69e1e811 | 6992 | rcu_read_lock(); |
37dc6b50 | 6993 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
6994 | |
6995 | if (!sd || sd->nohz_idle) | |
6996 | goto unlock; | |
6997 | sd->nohz_idle = 1; | |
6998 | ||
37dc6b50 | 6999 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 7000 | unlock: |
69e1e811 SS |
7001 | rcu_read_unlock(); |
7002 | } | |
7003 | ||
1e3c88bd | 7004 | /* |
c1cc017c | 7005 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 7006 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 7007 | */ |
c1cc017c | 7008 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 7009 | { |
71325960 SS |
7010 | /* |
7011 | * If this cpu is going down, then nothing needs to be done. | |
7012 | */ | |
7013 | if (!cpu_active(cpu)) | |
7014 | return; | |
7015 | ||
c1cc017c AS |
7016 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
7017 | return; | |
1e3c88bd | 7018 | |
d987fc7f MG |
7019 | /* |
7020 | * If we're a completely isolated CPU, we don't play. | |
7021 | */ | |
7022 | if (on_null_domain(cpu_rq(cpu))) | |
7023 | return; | |
7024 | ||
c1cc017c AS |
7025 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
7026 | atomic_inc(&nohz.nr_cpus); | |
7027 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 7028 | } |
71325960 | 7029 | |
0db0628d | 7030 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
7031 | unsigned long action, void *hcpu) |
7032 | { | |
7033 | switch (action & ~CPU_TASKS_FROZEN) { | |
7034 | case CPU_DYING: | |
c1cc017c | 7035 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
7036 | return NOTIFY_OK; |
7037 | default: | |
7038 | return NOTIFY_DONE; | |
7039 | } | |
7040 | } | |
1e3c88bd PZ |
7041 | #endif |
7042 | ||
7043 | static DEFINE_SPINLOCK(balancing); | |
7044 | ||
49c022e6 PZ |
7045 | /* |
7046 | * Scale the max load_balance interval with the number of CPUs in the system. | |
7047 | * This trades load-balance latency on larger machines for less cross talk. | |
7048 | */ | |
029632fb | 7049 | void update_max_interval(void) |
49c022e6 PZ |
7050 | { |
7051 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
7052 | } | |
7053 | ||
1e3c88bd PZ |
7054 | /* |
7055 | * It checks each scheduling domain to see if it is due to be balanced, | |
7056 | * and initiates a balancing operation if so. | |
7057 | * | |
b9b0853a | 7058 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 7059 | */ |
f7ed0a89 | 7060 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 7061 | { |
23f0d209 | 7062 | int continue_balancing = 1; |
f7ed0a89 | 7063 | int cpu = rq->cpu; |
1e3c88bd | 7064 | unsigned long interval; |
04f733b4 | 7065 | struct sched_domain *sd; |
1e3c88bd PZ |
7066 | /* Earliest time when we have to do rebalance again */ |
7067 | unsigned long next_balance = jiffies + 60*HZ; | |
7068 | int update_next_balance = 0; | |
f48627e6 JL |
7069 | int need_serialize, need_decay = 0; |
7070 | u64 max_cost = 0; | |
1e3c88bd | 7071 | |
48a16753 | 7072 | update_blocked_averages(cpu); |
2069dd75 | 7073 | |
dce840a0 | 7074 | rcu_read_lock(); |
1e3c88bd | 7075 | for_each_domain(cpu, sd) { |
f48627e6 JL |
7076 | /* |
7077 | * Decay the newidle max times here because this is a regular | |
7078 | * visit to all the domains. Decay ~1% per second. | |
7079 | */ | |
7080 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
7081 | sd->max_newidle_lb_cost = | |
7082 | (sd->max_newidle_lb_cost * 253) / 256; | |
7083 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
7084 | need_decay = 1; | |
7085 | } | |
7086 | max_cost += sd->max_newidle_lb_cost; | |
7087 | ||
1e3c88bd PZ |
7088 | if (!(sd->flags & SD_LOAD_BALANCE)) |
7089 | continue; | |
7090 | ||
f48627e6 JL |
7091 | /* |
7092 | * Stop the load balance at this level. There is another | |
7093 | * CPU in our sched group which is doing load balancing more | |
7094 | * actively. | |
7095 | */ | |
7096 | if (!continue_balancing) { | |
7097 | if (need_decay) | |
7098 | continue; | |
7099 | break; | |
7100 | } | |
7101 | ||
52a08ef1 | 7102 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
7103 | |
7104 | need_serialize = sd->flags & SD_SERIALIZE; | |
1e3c88bd PZ |
7105 | if (need_serialize) { |
7106 | if (!spin_trylock(&balancing)) | |
7107 | goto out; | |
7108 | } | |
7109 | ||
7110 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 7111 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 7112 | /* |
6263322c | 7113 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
7114 | * env->dst_cpu, so we can't know our idle |
7115 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 7116 | */ |
de5eb2dd | 7117 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
7118 | } |
7119 | sd->last_balance = jiffies; | |
52a08ef1 | 7120 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
7121 | } |
7122 | if (need_serialize) | |
7123 | spin_unlock(&balancing); | |
7124 | out: | |
7125 | if (time_after(next_balance, sd->last_balance + interval)) { | |
7126 | next_balance = sd->last_balance + interval; | |
7127 | update_next_balance = 1; | |
7128 | } | |
f48627e6 JL |
7129 | } |
7130 | if (need_decay) { | |
1e3c88bd | 7131 | /* |
f48627e6 JL |
7132 | * Ensure the rq-wide value also decays but keep it at a |
7133 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 7134 | */ |
f48627e6 JL |
7135 | rq->max_idle_balance_cost = |
7136 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 7137 | } |
dce840a0 | 7138 | rcu_read_unlock(); |
1e3c88bd PZ |
7139 | |
7140 | /* | |
7141 | * next_balance will be updated only when there is a need. | |
7142 | * When the cpu is attached to null domain for ex, it will not be | |
7143 | * updated. | |
7144 | */ | |
7145 | if (likely(update_next_balance)) | |
7146 | rq->next_balance = next_balance; | |
7147 | } | |
7148 | ||
3451d024 | 7149 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 7150 | /* |
3451d024 | 7151 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
7152 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
7153 | */ | |
208cb16b | 7154 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 7155 | { |
208cb16b | 7156 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
7157 | struct rq *rq; |
7158 | int balance_cpu; | |
7159 | ||
1c792db7 SS |
7160 | if (idle != CPU_IDLE || |
7161 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
7162 | goto end; | |
83cd4fe2 VP |
7163 | |
7164 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 7165 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
7166 | continue; |
7167 | ||
7168 | /* | |
7169 | * If this cpu gets work to do, stop the load balancing | |
7170 | * work being done for other cpus. Next load | |
7171 | * balancing owner will pick it up. | |
7172 | */ | |
1c792db7 | 7173 | if (need_resched()) |
83cd4fe2 | 7174 | break; |
83cd4fe2 | 7175 | |
5ed4f1d9 VG |
7176 | rq = cpu_rq(balance_cpu); |
7177 | ||
7178 | raw_spin_lock_irq(&rq->lock); | |
7179 | update_rq_clock(rq); | |
7180 | update_idle_cpu_load(rq); | |
7181 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 | 7182 | |
f7ed0a89 | 7183 | rebalance_domains(rq, CPU_IDLE); |
83cd4fe2 | 7184 | |
83cd4fe2 VP |
7185 | if (time_after(this_rq->next_balance, rq->next_balance)) |
7186 | this_rq->next_balance = rq->next_balance; | |
7187 | } | |
7188 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
7189 | end: |
7190 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
7191 | } |
7192 | ||
7193 | /* | |
0b005cf5 SS |
7194 | * Current heuristic for kicking the idle load balancer in the presence |
7195 | * of an idle cpu is the system. | |
7196 | * - This rq has more than one task. | |
7197 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
7198 | * busy cpu's exceeding the group's power. | |
7199 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
7200 | * domain span are idle. | |
83cd4fe2 | 7201 | */ |
4a725627 | 7202 | static inline int nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
7203 | { |
7204 | unsigned long now = jiffies; | |
0b005cf5 | 7205 | struct sched_domain *sd; |
37dc6b50 | 7206 | struct sched_group_power *sgp; |
4a725627 | 7207 | int nr_busy, cpu = rq->cpu; |
83cd4fe2 | 7208 | |
4a725627 | 7209 | if (unlikely(rq->idle_balance)) |
83cd4fe2 VP |
7210 | return 0; |
7211 | ||
1c792db7 SS |
7212 | /* |
7213 | * We may be recently in ticked or tickless idle mode. At the first | |
7214 | * busy tick after returning from idle, we will update the busy stats. | |
7215 | */ | |
69e1e811 | 7216 | set_cpu_sd_state_busy(); |
c1cc017c | 7217 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
7218 | |
7219 | /* | |
7220 | * None are in tickless mode and hence no need for NOHZ idle load | |
7221 | * balancing. | |
7222 | */ | |
7223 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
7224 | return 0; | |
1c792db7 SS |
7225 | |
7226 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
7227 | return 0; |
7228 | ||
0b005cf5 SS |
7229 | if (rq->nr_running >= 2) |
7230 | goto need_kick; | |
83cd4fe2 | 7231 | |
067491b7 | 7232 | rcu_read_lock(); |
37dc6b50 | 7233 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
83cd4fe2 | 7234 | |
37dc6b50 PM |
7235 | if (sd) { |
7236 | sgp = sd->groups->sgp; | |
7237 | nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
0b005cf5 | 7238 | |
37dc6b50 | 7239 | if (nr_busy > 1) |
067491b7 | 7240 | goto need_kick_unlock; |
83cd4fe2 | 7241 | } |
37dc6b50 PM |
7242 | |
7243 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); | |
7244 | ||
7245 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, | |
7246 | sched_domain_span(sd)) < cpu)) | |
7247 | goto need_kick_unlock; | |
7248 | ||
067491b7 | 7249 | rcu_read_unlock(); |
83cd4fe2 | 7250 | return 0; |
067491b7 PZ |
7251 | |
7252 | need_kick_unlock: | |
7253 | rcu_read_unlock(); | |
0b005cf5 SS |
7254 | need_kick: |
7255 | return 1; | |
83cd4fe2 VP |
7256 | } |
7257 | #else | |
208cb16b | 7258 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
7259 | #endif |
7260 | ||
7261 | /* | |
7262 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
7263 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
7264 | */ | |
1e3c88bd PZ |
7265 | static void run_rebalance_domains(struct softirq_action *h) |
7266 | { | |
208cb16b | 7267 | struct rq *this_rq = this_rq(); |
6eb57e0d | 7268 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
7269 | CPU_IDLE : CPU_NOT_IDLE; |
7270 | ||
f7ed0a89 | 7271 | rebalance_domains(this_rq, idle); |
1e3c88bd | 7272 | |
1e3c88bd | 7273 | /* |
83cd4fe2 | 7274 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
7275 | * balancing on behalf of the other idle cpus whose ticks are |
7276 | * stopped. | |
7277 | */ | |
208cb16b | 7278 | nohz_idle_balance(this_rq, idle); |
1e3c88bd PZ |
7279 | } |
7280 | ||
1e3c88bd PZ |
7281 | /* |
7282 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 7283 | */ |
7caff66f | 7284 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 7285 | { |
1e3c88bd | 7286 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
7287 | if (unlikely(on_null_domain(rq))) |
7288 | return; | |
7289 | ||
7290 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 7291 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 7292 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 7293 | if (nohz_kick_needed(rq)) |
0aeeeeba | 7294 | nohz_balancer_kick(); |
83cd4fe2 | 7295 | #endif |
1e3c88bd PZ |
7296 | } |
7297 | ||
0bcdcf28 CE |
7298 | static void rq_online_fair(struct rq *rq) |
7299 | { | |
7300 | update_sysctl(); | |
7301 | } | |
7302 | ||
7303 | static void rq_offline_fair(struct rq *rq) | |
7304 | { | |
7305 | update_sysctl(); | |
a4c96ae3 PB |
7306 | |
7307 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
7308 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
7309 | } |
7310 | ||
55e12e5e | 7311 | #endif /* CONFIG_SMP */ |
e1d1484f | 7312 | |
bf0f6f24 IM |
7313 | /* |
7314 | * scheduler tick hitting a task of our scheduling class: | |
7315 | */ | |
8f4d37ec | 7316 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
7317 | { |
7318 | struct cfs_rq *cfs_rq; | |
7319 | struct sched_entity *se = &curr->se; | |
7320 | ||
7321 | for_each_sched_entity(se) { | |
7322 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 7323 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 7324 | } |
18bf2805 | 7325 | |
10e84b97 | 7326 | if (numabalancing_enabled) |
cbee9f88 | 7327 | task_tick_numa(rq, curr); |
3d59eebc | 7328 | |
18bf2805 | 7329 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
7330 | } |
7331 | ||
7332 | /* | |
cd29fe6f PZ |
7333 | * called on fork with the child task as argument from the parent's context |
7334 | * - child not yet on the tasklist | |
7335 | * - preemption disabled | |
bf0f6f24 | 7336 | */ |
cd29fe6f | 7337 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 7338 | { |
4fc420c9 DN |
7339 | struct cfs_rq *cfs_rq; |
7340 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 7341 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
7342 | struct rq *rq = this_rq(); |
7343 | unsigned long flags; | |
7344 | ||
05fa785c | 7345 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 7346 | |
861d034e PZ |
7347 | update_rq_clock(rq); |
7348 | ||
4fc420c9 DN |
7349 | cfs_rq = task_cfs_rq(current); |
7350 | curr = cfs_rq->curr; | |
7351 | ||
6c9a27f5 DN |
7352 | /* |
7353 | * Not only the cpu but also the task_group of the parent might have | |
7354 | * been changed after parent->se.parent,cfs_rq were copied to | |
7355 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
7356 | * of child point to valid ones. | |
7357 | */ | |
7358 | rcu_read_lock(); | |
7359 | __set_task_cpu(p, this_cpu); | |
7360 | rcu_read_unlock(); | |
bf0f6f24 | 7361 | |
7109c442 | 7362 | update_curr(cfs_rq); |
cd29fe6f | 7363 | |
b5d9d734 MG |
7364 | if (curr) |
7365 | se->vruntime = curr->vruntime; | |
aeb73b04 | 7366 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 7367 | |
cd29fe6f | 7368 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 7369 | /* |
edcb60a3 IM |
7370 | * Upon rescheduling, sched_class::put_prev_task() will place |
7371 | * 'current' within the tree based on its new key value. | |
7372 | */ | |
4d78e7b6 | 7373 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 7374 | resched_task(rq->curr); |
4d78e7b6 | 7375 | } |
bf0f6f24 | 7376 | |
88ec22d3 PZ |
7377 | se->vruntime -= cfs_rq->min_vruntime; |
7378 | ||
05fa785c | 7379 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
7380 | } |
7381 | ||
cb469845 SR |
7382 | /* |
7383 | * Priority of the task has changed. Check to see if we preempt | |
7384 | * the current task. | |
7385 | */ | |
da7a735e PZ |
7386 | static void |
7387 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 7388 | { |
da7a735e PZ |
7389 | if (!p->se.on_rq) |
7390 | return; | |
7391 | ||
cb469845 SR |
7392 | /* |
7393 | * Reschedule if we are currently running on this runqueue and | |
7394 | * our priority decreased, or if we are not currently running on | |
7395 | * this runqueue and our priority is higher than the current's | |
7396 | */ | |
da7a735e | 7397 | if (rq->curr == p) { |
cb469845 SR |
7398 | if (p->prio > oldprio) |
7399 | resched_task(rq->curr); | |
7400 | } else | |
15afe09b | 7401 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7402 | } |
7403 | ||
da7a735e PZ |
7404 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
7405 | { | |
7406 | struct sched_entity *se = &p->se; | |
7407 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7408 | ||
7409 | /* | |
791c9e02 | 7410 | * Ensure the task's vruntime is normalized, so that when it's |
da7a735e PZ |
7411 | * switched back to the fair class the enqueue_entity(.flags=0) will |
7412 | * do the right thing. | |
7413 | * | |
791c9e02 GM |
7414 | * If it's on_rq, then the dequeue_entity(.flags=0) will already |
7415 | * have normalized the vruntime, if it's !on_rq, then only when | |
da7a735e PZ |
7416 | * the task is sleeping will it still have non-normalized vruntime. |
7417 | */ | |
791c9e02 | 7418 | if (!p->on_rq && p->state != TASK_RUNNING) { |
da7a735e PZ |
7419 | /* |
7420 | * Fix up our vruntime so that the current sleep doesn't | |
7421 | * cause 'unlimited' sleep bonus. | |
7422 | */ | |
7423 | place_entity(cfs_rq, se, 0); | |
7424 | se->vruntime -= cfs_rq->min_vruntime; | |
7425 | } | |
9ee474f5 | 7426 | |
141965c7 | 7427 | #ifdef CONFIG_SMP |
9ee474f5 PT |
7428 | /* |
7429 | * Remove our load from contribution when we leave sched_fair | |
7430 | * and ensure we don't carry in an old decay_count if we | |
7431 | * switch back. | |
7432 | */ | |
87e3c8ae KT |
7433 | if (se->avg.decay_count) { |
7434 | __synchronize_entity_decay(se); | |
7435 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
7436 | } |
7437 | #endif | |
da7a735e PZ |
7438 | } |
7439 | ||
cb469845 SR |
7440 | /* |
7441 | * We switched to the sched_fair class. | |
7442 | */ | |
da7a735e | 7443 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 7444 | { |
eb7a59b2 M |
7445 | struct sched_entity *se = &p->se; |
7446 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7447 | /* | |
7448 | * Since the real-depth could have been changed (only FAIR | |
7449 | * class maintain depth value), reset depth properly. | |
7450 | */ | |
7451 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
7452 | #endif | |
7453 | if (!se->on_rq) | |
da7a735e PZ |
7454 | return; |
7455 | ||
cb469845 SR |
7456 | /* |
7457 | * We were most likely switched from sched_rt, so | |
7458 | * kick off the schedule if running, otherwise just see | |
7459 | * if we can still preempt the current task. | |
7460 | */ | |
da7a735e | 7461 | if (rq->curr == p) |
cb469845 SR |
7462 | resched_task(rq->curr); |
7463 | else | |
15afe09b | 7464 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7465 | } |
7466 | ||
83b699ed SV |
7467 | /* Account for a task changing its policy or group. |
7468 | * | |
7469 | * This routine is mostly called to set cfs_rq->curr field when a task | |
7470 | * migrates between groups/classes. | |
7471 | */ | |
7472 | static void set_curr_task_fair(struct rq *rq) | |
7473 | { | |
7474 | struct sched_entity *se = &rq->curr->se; | |
7475 | ||
ec12cb7f PT |
7476 | for_each_sched_entity(se) { |
7477 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7478 | ||
7479 | set_next_entity(cfs_rq, se); | |
7480 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
7481 | account_cfs_rq_runtime(cfs_rq, 0); | |
7482 | } | |
83b699ed SV |
7483 | } |
7484 | ||
029632fb PZ |
7485 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
7486 | { | |
7487 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
7488 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
7489 | #ifndef CONFIG_64BIT | |
7490 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
7491 | #endif | |
141965c7 | 7492 | #ifdef CONFIG_SMP |
9ee474f5 | 7493 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 7494 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 7495 | #endif |
029632fb PZ |
7496 | } |
7497 | ||
810b3817 | 7498 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7499 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 7500 | { |
fed14d45 | 7501 | struct sched_entity *se = &p->se; |
aff3e498 | 7502 | struct cfs_rq *cfs_rq; |
fed14d45 | 7503 | |
b2b5ce02 PZ |
7504 | /* |
7505 | * If the task was not on the rq at the time of this cgroup movement | |
7506 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
7507 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
7508 | * bonus in place_entity()). | |
7509 | * | |
7510 | * If it was on the rq, we've just 'preempted' it, which does convert | |
7511 | * ->vruntime to a relative base. | |
7512 | * | |
7513 | * Make sure both cases convert their relative position when migrating | |
7514 | * to another cgroup's rq. This does somewhat interfere with the | |
7515 | * fair sleeper stuff for the first placement, but who cares. | |
7516 | */ | |
7ceff013 DN |
7517 | /* |
7518 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
7519 | * But there are some cases where it has already been normalized: | |
7520 | * | |
7521 | * - Moving a forked child which is waiting for being woken up by | |
7522 | * wake_up_new_task(). | |
62af3783 DN |
7523 | * - Moving a task which has been woken up by try_to_wake_up() and |
7524 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
7525 | * |
7526 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
7527 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
7528 | */ | |
fed14d45 | 7529 | if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
7530 | on_rq = 1; |
7531 | ||
b2b5ce02 | 7532 | if (!on_rq) |
fed14d45 | 7533 | se->vruntime -= cfs_rq_of(se)->min_vruntime; |
b2b5ce02 | 7534 | set_task_rq(p, task_cpu(p)); |
fed14d45 | 7535 | se->depth = se->parent ? se->parent->depth + 1 : 0; |
aff3e498 | 7536 | if (!on_rq) { |
fed14d45 PZ |
7537 | cfs_rq = cfs_rq_of(se); |
7538 | se->vruntime += cfs_rq->min_vruntime; | |
aff3e498 PT |
7539 | #ifdef CONFIG_SMP |
7540 | /* | |
7541 | * migrate_task_rq_fair() will have removed our previous | |
7542 | * contribution, but we must synchronize for ongoing future | |
7543 | * decay. | |
7544 | */ | |
fed14d45 PZ |
7545 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); |
7546 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
aff3e498 PT |
7547 | #endif |
7548 | } | |
810b3817 | 7549 | } |
029632fb PZ |
7550 | |
7551 | void free_fair_sched_group(struct task_group *tg) | |
7552 | { | |
7553 | int i; | |
7554 | ||
7555 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7556 | ||
7557 | for_each_possible_cpu(i) { | |
7558 | if (tg->cfs_rq) | |
7559 | kfree(tg->cfs_rq[i]); | |
7560 | if (tg->se) | |
7561 | kfree(tg->se[i]); | |
7562 | } | |
7563 | ||
7564 | kfree(tg->cfs_rq); | |
7565 | kfree(tg->se); | |
7566 | } | |
7567 | ||
7568 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7569 | { | |
7570 | struct cfs_rq *cfs_rq; | |
7571 | struct sched_entity *se; | |
7572 | int i; | |
7573 | ||
7574 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
7575 | if (!tg->cfs_rq) | |
7576 | goto err; | |
7577 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
7578 | if (!tg->se) | |
7579 | goto err; | |
7580 | ||
7581 | tg->shares = NICE_0_LOAD; | |
7582 | ||
7583 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7584 | ||
7585 | for_each_possible_cpu(i) { | |
7586 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
7587 | GFP_KERNEL, cpu_to_node(i)); | |
7588 | if (!cfs_rq) | |
7589 | goto err; | |
7590 | ||
7591 | se = kzalloc_node(sizeof(struct sched_entity), | |
7592 | GFP_KERNEL, cpu_to_node(i)); | |
7593 | if (!se) | |
7594 | goto err_free_rq; | |
7595 | ||
7596 | init_cfs_rq(cfs_rq); | |
7597 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
7598 | } | |
7599 | ||
7600 | return 1; | |
7601 | ||
7602 | err_free_rq: | |
7603 | kfree(cfs_rq); | |
7604 | err: | |
7605 | return 0; | |
7606 | } | |
7607 | ||
7608 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
7609 | { | |
7610 | struct rq *rq = cpu_rq(cpu); | |
7611 | unsigned long flags; | |
7612 | ||
7613 | /* | |
7614 | * Only empty task groups can be destroyed; so we can speculatively | |
7615 | * check on_list without danger of it being re-added. | |
7616 | */ | |
7617 | if (!tg->cfs_rq[cpu]->on_list) | |
7618 | return; | |
7619 | ||
7620 | raw_spin_lock_irqsave(&rq->lock, flags); | |
7621 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
7622 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7623 | } | |
7624 | ||
7625 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
7626 | struct sched_entity *se, int cpu, | |
7627 | struct sched_entity *parent) | |
7628 | { | |
7629 | struct rq *rq = cpu_rq(cpu); | |
7630 | ||
7631 | cfs_rq->tg = tg; | |
7632 | cfs_rq->rq = rq; | |
029632fb PZ |
7633 | init_cfs_rq_runtime(cfs_rq); |
7634 | ||
7635 | tg->cfs_rq[cpu] = cfs_rq; | |
7636 | tg->se[cpu] = se; | |
7637 | ||
7638 | /* se could be NULL for root_task_group */ | |
7639 | if (!se) | |
7640 | return; | |
7641 | ||
fed14d45 | 7642 | if (!parent) { |
029632fb | 7643 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
7644 | se->depth = 0; |
7645 | } else { | |
029632fb | 7646 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
7647 | se->depth = parent->depth + 1; |
7648 | } | |
029632fb PZ |
7649 | |
7650 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
7651 | /* guarantee group entities always have weight */ |
7652 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
7653 | se->parent = parent; |
7654 | } | |
7655 | ||
7656 | static DEFINE_MUTEX(shares_mutex); | |
7657 | ||
7658 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
7659 | { | |
7660 | int i; | |
7661 | unsigned long flags; | |
7662 | ||
7663 | /* | |
7664 | * We can't change the weight of the root cgroup. | |
7665 | */ | |
7666 | if (!tg->se[0]) | |
7667 | return -EINVAL; | |
7668 | ||
7669 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
7670 | ||
7671 | mutex_lock(&shares_mutex); | |
7672 | if (tg->shares == shares) | |
7673 | goto done; | |
7674 | ||
7675 | tg->shares = shares; | |
7676 | for_each_possible_cpu(i) { | |
7677 | struct rq *rq = cpu_rq(i); | |
7678 | struct sched_entity *se; | |
7679 | ||
7680 | se = tg->se[i]; | |
7681 | /* Propagate contribution to hierarchy */ | |
7682 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
7683 | |
7684 | /* Possible calls to update_curr() need rq clock */ | |
7685 | update_rq_clock(rq); | |
17bc14b7 | 7686 | for_each_sched_entity(se) |
029632fb PZ |
7687 | update_cfs_shares(group_cfs_rq(se)); |
7688 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7689 | } | |
7690 | ||
7691 | done: | |
7692 | mutex_unlock(&shares_mutex); | |
7693 | return 0; | |
7694 | } | |
7695 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
7696 | ||
7697 | void free_fair_sched_group(struct task_group *tg) { } | |
7698 | ||
7699 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7700 | { | |
7701 | return 1; | |
7702 | } | |
7703 | ||
7704 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
7705 | ||
7706 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
7707 | ||
810b3817 | 7708 | |
6d686f45 | 7709 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
7710 | { |
7711 | struct sched_entity *se = &task->se; | |
0d721cea PW |
7712 | unsigned int rr_interval = 0; |
7713 | ||
7714 | /* | |
7715 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
7716 | * idle runqueue: | |
7717 | */ | |
0d721cea | 7718 | if (rq->cfs.load.weight) |
a59f4e07 | 7719 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
7720 | |
7721 | return rr_interval; | |
7722 | } | |
7723 | ||
bf0f6f24 IM |
7724 | /* |
7725 | * All the scheduling class methods: | |
7726 | */ | |
029632fb | 7727 | const struct sched_class fair_sched_class = { |
5522d5d5 | 7728 | .next = &idle_sched_class, |
bf0f6f24 IM |
7729 | .enqueue_task = enqueue_task_fair, |
7730 | .dequeue_task = dequeue_task_fair, | |
7731 | .yield_task = yield_task_fair, | |
d95f4122 | 7732 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 7733 | |
2e09bf55 | 7734 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
7735 | |
7736 | .pick_next_task = pick_next_task_fair, | |
7737 | .put_prev_task = put_prev_task_fair, | |
7738 | ||
681f3e68 | 7739 | #ifdef CONFIG_SMP |
4ce72a2c | 7740 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 7741 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 7742 | |
0bcdcf28 CE |
7743 | .rq_online = rq_online_fair, |
7744 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
7745 | |
7746 | .task_waking = task_waking_fair, | |
681f3e68 | 7747 | #endif |
bf0f6f24 | 7748 | |
83b699ed | 7749 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 7750 | .task_tick = task_tick_fair, |
cd29fe6f | 7751 | .task_fork = task_fork_fair, |
cb469845 SR |
7752 | |
7753 | .prio_changed = prio_changed_fair, | |
da7a735e | 7754 | .switched_from = switched_from_fair, |
cb469845 | 7755 | .switched_to = switched_to_fair, |
810b3817 | 7756 | |
0d721cea PW |
7757 | .get_rr_interval = get_rr_interval_fair, |
7758 | ||
810b3817 | 7759 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7760 | .task_move_group = task_move_group_fair, |
810b3817 | 7761 | #endif |
bf0f6f24 IM |
7762 | }; |
7763 | ||
7764 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 7765 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 7766 | { |
bf0f6f24 IM |
7767 | struct cfs_rq *cfs_rq; |
7768 | ||
5973e5b9 | 7769 | rcu_read_lock(); |
c3b64f1e | 7770 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 7771 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 7772 | rcu_read_unlock(); |
bf0f6f24 IM |
7773 | } |
7774 | #endif | |
029632fb PZ |
7775 | |
7776 | __init void init_sched_fair_class(void) | |
7777 | { | |
7778 | #ifdef CONFIG_SMP | |
7779 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
7780 | ||
3451d024 | 7781 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 7782 | nohz.next_balance = jiffies; |
029632fb | 7783 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 7784 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
7785 | #endif |
7786 | #endif /* SMP */ | |
7787 | ||
7788 | } |