Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
3436ae12 | 25 | #include <linux/cpumask.h> |
029632fb PZ |
26 | #include <linux/slab.h> |
27 | #include <linux/profile.h> | |
28 | #include <linux/interrupt.h> | |
cbee9f88 | 29 | #include <linux/mempolicy.h> |
e14808b4 | 30 | #include <linux/migrate.h> |
cbee9f88 | 31 | #include <linux/task_work.h> |
029632fb PZ |
32 | |
33 | #include <trace/events/sched.h> | |
34 | ||
35 | #include "sched.h" | |
9745512c | 36 | |
bf0f6f24 | 37 | /* |
21805085 | 38 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 39 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 40 | * |
21805085 | 41 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
42 | * 'timeslice length' - timeslices in CFS are of variable length |
43 | * and have no persistent notion like in traditional, time-slice | |
44 | * based scheduling concepts. | |
bf0f6f24 | 45 | * |
d274a4ce IM |
46 | * (to see the precise effective timeslice length of your workload, |
47 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 48 | */ |
21406928 MG |
49 | unsigned int sysctl_sched_latency = 6000000ULL; |
50 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 51 | |
1983a922 CE |
52 | /* |
53 | * The initial- and re-scaling of tunables is configurable | |
54 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
55 | * | |
56 | * Options are: | |
57 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
58 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
59 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
60 | */ | |
61 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
62 | = SCHED_TUNABLESCALING_LOG; | |
63 | ||
2bd8e6d4 | 64 | /* |
b2be5e96 | 65 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 66 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 67 | */ |
0bf377bb IM |
68 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
69 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
70 | |
71 | /* | |
b2be5e96 PZ |
72 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
73 | */ | |
0bf377bb | 74 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
75 | |
76 | /* | |
2bba22c5 | 77 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 78 | * parent will (try to) run first. |
21805085 | 79 | */ |
2bba22c5 | 80 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 81 | |
bf0f6f24 IM |
82 | /* |
83 | * SCHED_OTHER wake-up granularity. | |
172e082a | 84 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
85 | * |
86 | * This option delays the preemption effects of decoupled workloads | |
87 | * and reduces their over-scheduling. Synchronous workloads will still | |
88 | * have immediate wakeup/sleep latencies. | |
89 | */ | |
172e082a | 90 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 91 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 92 | |
da84d961 IM |
93 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
94 | ||
a7a4f8a7 PT |
95 | /* |
96 | * The exponential sliding window over which load is averaged for shares | |
97 | * distribution. | |
98 | * (default: 10msec) | |
99 | */ | |
100 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
101 | ||
ec12cb7f PT |
102 | #ifdef CONFIG_CFS_BANDWIDTH |
103 | /* | |
104 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
105 | * each time a cfs_rq requests quota. | |
106 | * | |
107 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
108 | * to consumption or the quota being specified to be smaller than the slice) | |
109 | * we will always only issue the remaining available time. | |
110 | * | |
111 | * default: 5 msec, units: microseconds | |
112 | */ | |
113 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
114 | #endif | |
115 | ||
8527632d PG |
116 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
117 | { | |
118 | lw->weight += inc; | |
119 | lw->inv_weight = 0; | |
120 | } | |
121 | ||
122 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
123 | { | |
124 | lw->weight -= dec; | |
125 | lw->inv_weight = 0; | |
126 | } | |
127 | ||
128 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
129 | { | |
130 | lw->weight = w; | |
131 | lw->inv_weight = 0; | |
132 | } | |
133 | ||
029632fb PZ |
134 | /* |
135 | * Increase the granularity value when there are more CPUs, | |
136 | * because with more CPUs the 'effective latency' as visible | |
137 | * to users decreases. But the relationship is not linear, | |
138 | * so pick a second-best guess by going with the log2 of the | |
139 | * number of CPUs. | |
140 | * | |
141 | * This idea comes from the SD scheduler of Con Kolivas: | |
142 | */ | |
143 | static int get_update_sysctl_factor(void) | |
144 | { | |
145 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
146 | unsigned int factor; | |
147 | ||
148 | switch (sysctl_sched_tunable_scaling) { | |
149 | case SCHED_TUNABLESCALING_NONE: | |
150 | factor = 1; | |
151 | break; | |
152 | case SCHED_TUNABLESCALING_LINEAR: | |
153 | factor = cpus; | |
154 | break; | |
155 | case SCHED_TUNABLESCALING_LOG: | |
156 | default: | |
157 | factor = 1 + ilog2(cpus); | |
158 | break; | |
159 | } | |
160 | ||
161 | return factor; | |
162 | } | |
163 | ||
164 | static void update_sysctl(void) | |
165 | { | |
166 | unsigned int factor = get_update_sysctl_factor(); | |
167 | ||
168 | #define SET_SYSCTL(name) \ | |
169 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
170 | SET_SYSCTL(sched_min_granularity); | |
171 | SET_SYSCTL(sched_latency); | |
172 | SET_SYSCTL(sched_wakeup_granularity); | |
173 | #undef SET_SYSCTL | |
174 | } | |
175 | ||
176 | void sched_init_granularity(void) | |
177 | { | |
178 | update_sysctl(); | |
179 | } | |
180 | ||
181 | #if BITS_PER_LONG == 32 | |
182 | # define WMULT_CONST (~0UL) | |
183 | #else | |
184 | # define WMULT_CONST (1UL << 32) | |
185 | #endif | |
186 | ||
187 | #define WMULT_SHIFT 32 | |
188 | ||
189 | /* | |
190 | * Shift right and round: | |
191 | */ | |
192 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
193 | ||
194 | /* | |
195 | * delta *= weight / lw | |
196 | */ | |
197 | static unsigned long | |
198 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
199 | struct load_weight *lw) | |
200 | { | |
201 | u64 tmp; | |
202 | ||
203 | /* | |
204 | * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched | |
205 | * entities since MIN_SHARES = 2. Treat weight as 1 if less than | |
206 | * 2^SCHED_LOAD_RESOLUTION. | |
207 | */ | |
208 | if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) | |
209 | tmp = (u64)delta_exec * scale_load_down(weight); | |
210 | else | |
211 | tmp = (u64)delta_exec; | |
212 | ||
213 | if (!lw->inv_weight) { | |
214 | unsigned long w = scale_load_down(lw->weight); | |
215 | ||
216 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
217 | lw->inv_weight = 1; | |
218 | else if (unlikely(!w)) | |
219 | lw->inv_weight = WMULT_CONST; | |
220 | else | |
221 | lw->inv_weight = WMULT_CONST / w; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Check whether we'd overflow the 64-bit multiplication: | |
226 | */ | |
227 | if (unlikely(tmp > WMULT_CONST)) | |
228 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
229 | WMULT_SHIFT/2); | |
230 | else | |
231 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
232 | ||
233 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
234 | } | |
235 | ||
236 | ||
237 | const struct sched_class fair_sched_class; | |
a4c2f00f | 238 | |
bf0f6f24 IM |
239 | /************************************************************** |
240 | * CFS operations on generic schedulable entities: | |
241 | */ | |
242 | ||
62160e3f | 243 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 244 | |
62160e3f | 245 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
246 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
247 | { | |
62160e3f | 248 | return cfs_rq->rq; |
bf0f6f24 IM |
249 | } |
250 | ||
62160e3f IM |
251 | /* An entity is a task if it doesn't "own" a runqueue */ |
252 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 253 | |
8f48894f PZ |
254 | static inline struct task_struct *task_of(struct sched_entity *se) |
255 | { | |
256 | #ifdef CONFIG_SCHED_DEBUG | |
257 | WARN_ON_ONCE(!entity_is_task(se)); | |
258 | #endif | |
259 | return container_of(se, struct task_struct, se); | |
260 | } | |
261 | ||
b758149c PZ |
262 | /* Walk up scheduling entities hierarchy */ |
263 | #define for_each_sched_entity(se) \ | |
264 | for (; se; se = se->parent) | |
265 | ||
266 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
267 | { | |
268 | return p->se.cfs_rq; | |
269 | } | |
270 | ||
271 | /* runqueue on which this entity is (to be) queued */ | |
272 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
273 | { | |
274 | return se->cfs_rq; | |
275 | } | |
276 | ||
277 | /* runqueue "owned" by this group */ | |
278 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
279 | { | |
280 | return grp->my_q; | |
281 | } | |
282 | ||
aff3e498 PT |
283 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
284 | int force_update); | |
9ee474f5 | 285 | |
3d4b47b4 PZ |
286 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
287 | { | |
288 | if (!cfs_rq->on_list) { | |
67e86250 PT |
289 | /* |
290 | * Ensure we either appear before our parent (if already | |
291 | * enqueued) or force our parent to appear after us when it is | |
292 | * enqueued. The fact that we always enqueue bottom-up | |
293 | * reduces this to two cases. | |
294 | */ | |
295 | if (cfs_rq->tg->parent && | |
296 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
297 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
298 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
299 | } else { | |
300 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 302 | } |
3d4b47b4 PZ |
303 | |
304 | cfs_rq->on_list = 1; | |
9ee474f5 | 305 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 306 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
307 | } |
308 | } | |
309 | ||
310 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
311 | { | |
312 | if (cfs_rq->on_list) { | |
313 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
314 | cfs_rq->on_list = 0; | |
315 | } | |
316 | } | |
317 | ||
b758149c PZ |
318 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
319 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
320 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
321 | ||
322 | /* Do the two (enqueued) entities belong to the same group ? */ | |
323 | static inline int | |
324 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
325 | { | |
326 | if (se->cfs_rq == pse->cfs_rq) | |
327 | return 1; | |
328 | ||
329 | return 0; | |
330 | } | |
331 | ||
332 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
333 | { | |
334 | return se->parent; | |
335 | } | |
336 | ||
464b7527 PZ |
337 | /* return depth at which a sched entity is present in the hierarchy */ |
338 | static inline int depth_se(struct sched_entity *se) | |
339 | { | |
340 | int depth = 0; | |
341 | ||
342 | for_each_sched_entity(se) | |
343 | depth++; | |
344 | ||
345 | return depth; | |
346 | } | |
347 | ||
348 | static void | |
349 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
350 | { | |
351 | int se_depth, pse_depth; | |
352 | ||
353 | /* | |
354 | * preemption test can be made between sibling entities who are in the | |
355 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
356 | * both tasks until we find their ancestors who are siblings of common | |
357 | * parent. | |
358 | */ | |
359 | ||
360 | /* First walk up until both entities are at same depth */ | |
361 | se_depth = depth_se(*se); | |
362 | pse_depth = depth_se(*pse); | |
363 | ||
364 | while (se_depth > pse_depth) { | |
365 | se_depth--; | |
366 | *se = parent_entity(*se); | |
367 | } | |
368 | ||
369 | while (pse_depth > se_depth) { | |
370 | pse_depth--; | |
371 | *pse = parent_entity(*pse); | |
372 | } | |
373 | ||
374 | while (!is_same_group(*se, *pse)) { | |
375 | *se = parent_entity(*se); | |
376 | *pse = parent_entity(*pse); | |
377 | } | |
378 | } | |
379 | ||
8f48894f PZ |
380 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
381 | ||
382 | static inline struct task_struct *task_of(struct sched_entity *se) | |
383 | { | |
384 | return container_of(se, struct task_struct, se); | |
385 | } | |
bf0f6f24 | 386 | |
62160e3f IM |
387 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
388 | { | |
389 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
390 | } |
391 | ||
392 | #define entity_is_task(se) 1 | |
393 | ||
b758149c PZ |
394 | #define for_each_sched_entity(se) \ |
395 | for (; se; se = NULL) | |
bf0f6f24 | 396 | |
b758149c | 397 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 398 | { |
b758149c | 399 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
400 | } |
401 | ||
b758149c PZ |
402 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
403 | { | |
404 | struct task_struct *p = task_of(se); | |
405 | struct rq *rq = task_rq(p); | |
406 | ||
407 | return &rq->cfs; | |
408 | } | |
409 | ||
410 | /* runqueue "owned" by this group */ | |
411 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
412 | { | |
413 | return NULL; | |
414 | } | |
415 | ||
3d4b47b4 PZ |
416 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
417 | { | |
418 | } | |
419 | ||
420 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
421 | { | |
422 | } | |
423 | ||
b758149c PZ |
424 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
425 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
426 | ||
427 | static inline int | |
428 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
429 | { | |
430 | return 1; | |
431 | } | |
432 | ||
433 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
434 | { | |
435 | return NULL; | |
436 | } | |
437 | ||
464b7527 PZ |
438 | static inline void |
439 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
440 | { | |
441 | } | |
442 | ||
b758149c PZ |
443 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
444 | ||
6c16a6dc PZ |
445 | static __always_inline |
446 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
447 | |
448 | /************************************************************** | |
449 | * Scheduling class tree data structure manipulation methods: | |
450 | */ | |
451 | ||
1bf08230 | 452 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 453 | { |
1bf08230 | 454 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 455 | if (delta > 0) |
1bf08230 | 456 | max_vruntime = vruntime; |
02e0431a | 457 | |
1bf08230 | 458 | return max_vruntime; |
02e0431a PZ |
459 | } |
460 | ||
0702e3eb | 461 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
462 | { |
463 | s64 delta = (s64)(vruntime - min_vruntime); | |
464 | if (delta < 0) | |
465 | min_vruntime = vruntime; | |
466 | ||
467 | return min_vruntime; | |
468 | } | |
469 | ||
54fdc581 FC |
470 | static inline int entity_before(struct sched_entity *a, |
471 | struct sched_entity *b) | |
472 | { | |
473 | return (s64)(a->vruntime - b->vruntime) < 0; | |
474 | } | |
475 | ||
1af5f730 PZ |
476 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
477 | { | |
478 | u64 vruntime = cfs_rq->min_vruntime; | |
479 | ||
480 | if (cfs_rq->curr) | |
481 | vruntime = cfs_rq->curr->vruntime; | |
482 | ||
483 | if (cfs_rq->rb_leftmost) { | |
484 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
485 | struct sched_entity, | |
486 | run_node); | |
487 | ||
e17036da | 488 | if (!cfs_rq->curr) |
1af5f730 PZ |
489 | vruntime = se->vruntime; |
490 | else | |
491 | vruntime = min_vruntime(vruntime, se->vruntime); | |
492 | } | |
493 | ||
1bf08230 | 494 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 495 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
496 | #ifndef CONFIG_64BIT |
497 | smp_wmb(); | |
498 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
499 | #endif | |
1af5f730 PZ |
500 | } |
501 | ||
bf0f6f24 IM |
502 | /* |
503 | * Enqueue an entity into the rb-tree: | |
504 | */ | |
0702e3eb | 505 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
506 | { |
507 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
508 | struct rb_node *parent = NULL; | |
509 | struct sched_entity *entry; | |
bf0f6f24 IM |
510 | int leftmost = 1; |
511 | ||
512 | /* | |
513 | * Find the right place in the rbtree: | |
514 | */ | |
515 | while (*link) { | |
516 | parent = *link; | |
517 | entry = rb_entry(parent, struct sched_entity, run_node); | |
518 | /* | |
519 | * We dont care about collisions. Nodes with | |
520 | * the same key stay together. | |
521 | */ | |
2bd2d6f2 | 522 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
523 | link = &parent->rb_left; |
524 | } else { | |
525 | link = &parent->rb_right; | |
526 | leftmost = 0; | |
527 | } | |
528 | } | |
529 | ||
530 | /* | |
531 | * Maintain a cache of leftmost tree entries (it is frequently | |
532 | * used): | |
533 | */ | |
1af5f730 | 534 | if (leftmost) |
57cb499d | 535 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
536 | |
537 | rb_link_node(&se->run_node, parent, link); | |
538 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
539 | } |
540 | ||
0702e3eb | 541 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 542 | { |
3fe69747 PZ |
543 | if (cfs_rq->rb_leftmost == &se->run_node) { |
544 | struct rb_node *next_node; | |
3fe69747 PZ |
545 | |
546 | next_node = rb_next(&se->run_node); | |
547 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 548 | } |
e9acbff6 | 549 | |
bf0f6f24 | 550 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
551 | } |
552 | ||
029632fb | 553 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 554 | { |
f4b6755f PZ |
555 | struct rb_node *left = cfs_rq->rb_leftmost; |
556 | ||
557 | if (!left) | |
558 | return NULL; | |
559 | ||
560 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
561 | } |
562 | ||
ac53db59 RR |
563 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
564 | { | |
565 | struct rb_node *next = rb_next(&se->run_node); | |
566 | ||
567 | if (!next) | |
568 | return NULL; | |
569 | ||
570 | return rb_entry(next, struct sched_entity, run_node); | |
571 | } | |
572 | ||
573 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 574 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 575 | { |
7eee3e67 | 576 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 577 | |
70eee74b BS |
578 | if (!last) |
579 | return NULL; | |
7eee3e67 IM |
580 | |
581 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
582 | } |
583 | ||
bf0f6f24 IM |
584 | /************************************************************** |
585 | * Scheduling class statistics methods: | |
586 | */ | |
587 | ||
acb4a848 | 588 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 589 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
590 | loff_t *ppos) |
591 | { | |
8d65af78 | 592 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 593 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
594 | |
595 | if (ret || !write) | |
596 | return ret; | |
597 | ||
598 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
599 | sysctl_sched_min_granularity); | |
600 | ||
acb4a848 CE |
601 | #define WRT_SYSCTL(name) \ |
602 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
603 | WRT_SYSCTL(sched_min_granularity); | |
604 | WRT_SYSCTL(sched_latency); | |
605 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
606 | #undef WRT_SYSCTL |
607 | ||
b2be5e96 PZ |
608 | return 0; |
609 | } | |
610 | #endif | |
647e7cac | 611 | |
a7be37ac | 612 | /* |
f9c0b095 | 613 | * delta /= w |
a7be37ac PZ |
614 | */ |
615 | static inline unsigned long | |
616 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
617 | { | |
f9c0b095 PZ |
618 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
619 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
620 | |
621 | return delta; | |
622 | } | |
623 | ||
647e7cac IM |
624 | /* |
625 | * The idea is to set a period in which each task runs once. | |
626 | * | |
532b1858 | 627 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
628 | * this period because otherwise the slices get too small. |
629 | * | |
630 | * p = (nr <= nl) ? l : l*nr/nl | |
631 | */ | |
4d78e7b6 PZ |
632 | static u64 __sched_period(unsigned long nr_running) |
633 | { | |
634 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 635 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
636 | |
637 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 638 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 639 | period *= nr_running; |
4d78e7b6 PZ |
640 | } |
641 | ||
642 | return period; | |
643 | } | |
644 | ||
647e7cac IM |
645 | /* |
646 | * We calculate the wall-time slice from the period by taking a part | |
647 | * proportional to the weight. | |
648 | * | |
f9c0b095 | 649 | * s = p*P[w/rw] |
647e7cac | 650 | */ |
6d0f0ebd | 651 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 652 | { |
0a582440 | 653 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 654 | |
0a582440 | 655 | for_each_sched_entity(se) { |
6272d68c | 656 | struct load_weight *load; |
3104bf03 | 657 | struct load_weight lw; |
6272d68c LM |
658 | |
659 | cfs_rq = cfs_rq_of(se); | |
660 | load = &cfs_rq->load; | |
f9c0b095 | 661 | |
0a582440 | 662 | if (unlikely(!se->on_rq)) { |
3104bf03 | 663 | lw = cfs_rq->load; |
0a582440 MG |
664 | |
665 | update_load_add(&lw, se->load.weight); | |
666 | load = &lw; | |
667 | } | |
668 | slice = calc_delta_mine(slice, se->load.weight, load); | |
669 | } | |
670 | return slice; | |
bf0f6f24 IM |
671 | } |
672 | ||
647e7cac | 673 | /* |
660cc00f | 674 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 675 | * |
f9c0b095 | 676 | * vs = s/w |
647e7cac | 677 | */ |
f9c0b095 | 678 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 679 | { |
f9c0b095 | 680 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
681 | } |
682 | ||
a75cdaa9 | 683 | #ifdef CONFIG_SMP |
fb13c7ee MG |
684 | static unsigned long task_h_load(struct task_struct *p); |
685 | ||
a75cdaa9 AS |
686 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
687 | ||
688 | /* Give new task start runnable values to heavy its load in infant time */ | |
689 | void init_task_runnable_average(struct task_struct *p) | |
690 | { | |
691 | u32 slice; | |
692 | ||
693 | p->se.avg.decay_count = 0; | |
694 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | |
695 | p->se.avg.runnable_avg_sum = slice; | |
696 | p->se.avg.runnable_avg_period = slice; | |
697 | __update_task_entity_contrib(&p->se); | |
698 | } | |
699 | #else | |
700 | void init_task_runnable_average(struct task_struct *p) | |
701 | { | |
702 | } | |
703 | #endif | |
704 | ||
bf0f6f24 IM |
705 | /* |
706 | * Update the current task's runtime statistics. Skip current tasks that | |
707 | * are not in our scheduling class. | |
708 | */ | |
709 | static inline void | |
8ebc91d9 IM |
710 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
711 | unsigned long delta_exec) | |
bf0f6f24 | 712 | { |
bbdba7c0 | 713 | unsigned long delta_exec_weighted; |
bf0f6f24 | 714 | |
41acab88 LDM |
715 | schedstat_set(curr->statistics.exec_max, |
716 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
717 | |
718 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 719 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 720 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 721 | |
e9acbff6 | 722 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 723 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
724 | } |
725 | ||
b7cc0896 | 726 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 727 | { |
429d43bc | 728 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 729 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
730 | unsigned long delta_exec; |
731 | ||
732 | if (unlikely(!curr)) | |
733 | return; | |
734 | ||
735 | /* | |
736 | * Get the amount of time the current task was running | |
737 | * since the last time we changed load (this cannot | |
738 | * overflow on 32 bits): | |
739 | */ | |
8ebc91d9 | 740 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
741 | if (!delta_exec) |
742 | return; | |
bf0f6f24 | 743 | |
8ebc91d9 IM |
744 | __update_curr(cfs_rq, curr, delta_exec); |
745 | curr->exec_start = now; | |
d842de87 SV |
746 | |
747 | if (entity_is_task(curr)) { | |
748 | struct task_struct *curtask = task_of(curr); | |
749 | ||
f977bb49 | 750 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 751 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 752 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 753 | } |
ec12cb7f PT |
754 | |
755 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
756 | } |
757 | ||
758 | static inline void | |
5870db5b | 759 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 760 | { |
78becc27 | 761 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
762 | } |
763 | ||
bf0f6f24 IM |
764 | /* |
765 | * Task is being enqueued - update stats: | |
766 | */ | |
d2417e5a | 767 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 768 | { |
bf0f6f24 IM |
769 | /* |
770 | * Are we enqueueing a waiting task? (for current tasks | |
771 | * a dequeue/enqueue event is a NOP) | |
772 | */ | |
429d43bc | 773 | if (se != cfs_rq->curr) |
5870db5b | 774 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
775 | } |
776 | ||
bf0f6f24 | 777 | static void |
9ef0a961 | 778 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 779 | { |
41acab88 | 780 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 781 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
782 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
783 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 784 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
785 | #ifdef CONFIG_SCHEDSTATS |
786 | if (entity_is_task(se)) { | |
787 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 788 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
789 | } |
790 | #endif | |
41acab88 | 791 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
792 | } |
793 | ||
794 | static inline void | |
19b6a2e3 | 795 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 796 | { |
bf0f6f24 IM |
797 | /* |
798 | * Mark the end of the wait period if dequeueing a | |
799 | * waiting task: | |
800 | */ | |
429d43bc | 801 | if (se != cfs_rq->curr) |
9ef0a961 | 802 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
803 | } |
804 | ||
805 | /* | |
806 | * We are picking a new current task - update its stats: | |
807 | */ | |
808 | static inline void | |
79303e9e | 809 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
810 | { |
811 | /* | |
812 | * We are starting a new run period: | |
813 | */ | |
78becc27 | 814 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
815 | } |
816 | ||
bf0f6f24 IM |
817 | /************************************************** |
818 | * Scheduling class queueing methods: | |
819 | */ | |
820 | ||
cbee9f88 PZ |
821 | #ifdef CONFIG_NUMA_BALANCING |
822 | /* | |
598f0ec0 MG |
823 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
824 | * calculated based on the tasks virtual memory size and | |
825 | * numa_balancing_scan_size. | |
cbee9f88 | 826 | */ |
598f0ec0 MG |
827 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
828 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
829 | unsigned int sysctl_numa_balancing_scan_period_reset = 60000; | |
6e5fb223 PZ |
830 | |
831 | /* Portion of address space to scan in MB */ | |
832 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 833 | |
4b96a29b PZ |
834 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
835 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
836 | ||
598f0ec0 MG |
837 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
838 | { | |
839 | unsigned long rss = 0; | |
840 | unsigned long nr_scan_pages; | |
841 | ||
842 | /* | |
843 | * Calculations based on RSS as non-present and empty pages are skipped | |
844 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
845 | * on resident pages | |
846 | */ | |
847 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
848 | rss = get_mm_rss(p->mm); | |
849 | if (!rss) | |
850 | rss = nr_scan_pages; | |
851 | ||
852 | rss = round_up(rss, nr_scan_pages); | |
853 | return rss / nr_scan_pages; | |
854 | } | |
855 | ||
856 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
857 | #define MAX_SCAN_WINDOW 2560 | |
858 | ||
859 | static unsigned int task_scan_min(struct task_struct *p) | |
860 | { | |
861 | unsigned int scan, floor; | |
862 | unsigned int windows = 1; | |
863 | ||
864 | if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW) | |
865 | windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size; | |
866 | floor = 1000 / windows; | |
867 | ||
868 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
869 | return max_t(unsigned int, floor, scan); | |
870 | } | |
871 | ||
872 | static unsigned int task_scan_max(struct task_struct *p) | |
873 | { | |
874 | unsigned int smin = task_scan_min(p); | |
875 | unsigned int smax; | |
876 | ||
877 | /* Watch for min being lower than max due to floor calculations */ | |
878 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
879 | return max(smin, smax); | |
880 | } | |
881 | ||
3a7053b3 MG |
882 | /* |
883 | * Once a preferred node is selected the scheduler balancer will prefer moving | |
884 | * a task to that node for sysctl_numa_balancing_settle_count number of PTE | |
885 | * scans. This will give the process the chance to accumulate more faults on | |
886 | * the preferred node but still allow the scheduler to move the task again if | |
887 | * the nodes CPUs are overloaded. | |
888 | */ | |
6fe6b2d6 | 889 | unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4; |
3a7053b3 | 890 | |
0ec8aa00 PZ |
891 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
892 | { | |
893 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
894 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
895 | } | |
896 | ||
897 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
898 | { | |
899 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
900 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
901 | } | |
902 | ||
8c8a743c PZ |
903 | struct numa_group { |
904 | atomic_t refcount; | |
905 | ||
906 | spinlock_t lock; /* nr_tasks, tasks */ | |
907 | int nr_tasks; | |
e29cf08b | 908 | pid_t gid; |
8c8a743c PZ |
909 | struct list_head task_list; |
910 | ||
911 | struct rcu_head rcu; | |
83e1d2cd | 912 | atomic_long_t total_faults; |
8c8a743c PZ |
913 | atomic_long_t faults[0]; |
914 | }; | |
915 | ||
e29cf08b MG |
916 | pid_t task_numa_group_id(struct task_struct *p) |
917 | { | |
918 | return p->numa_group ? p->numa_group->gid : 0; | |
919 | } | |
920 | ||
ac8e895b MG |
921 | static inline int task_faults_idx(int nid, int priv) |
922 | { | |
923 | return 2 * nid + priv; | |
924 | } | |
925 | ||
926 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
927 | { | |
928 | if (!p->numa_faults) | |
929 | return 0; | |
930 | ||
931 | return p->numa_faults[task_faults_idx(nid, 0)] + | |
932 | p->numa_faults[task_faults_idx(nid, 1)]; | |
933 | } | |
934 | ||
83e1d2cd MG |
935 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
936 | { | |
937 | if (!p->numa_group) | |
938 | return 0; | |
939 | ||
940 | return atomic_long_read(&p->numa_group->faults[2*nid]) + | |
941 | atomic_long_read(&p->numa_group->faults[2*nid+1]); | |
942 | } | |
943 | ||
944 | /* | |
945 | * These return the fraction of accesses done by a particular task, or | |
946 | * task group, on a particular numa node. The group weight is given a | |
947 | * larger multiplier, in order to group tasks together that are almost | |
948 | * evenly spread out between numa nodes. | |
949 | */ | |
950 | static inline unsigned long task_weight(struct task_struct *p, int nid) | |
951 | { | |
952 | unsigned long total_faults; | |
953 | ||
954 | if (!p->numa_faults) | |
955 | return 0; | |
956 | ||
957 | total_faults = p->total_numa_faults; | |
958 | ||
959 | if (!total_faults) | |
960 | return 0; | |
961 | ||
962 | return 1000 * task_faults(p, nid) / total_faults; | |
963 | } | |
964 | ||
965 | static inline unsigned long group_weight(struct task_struct *p, int nid) | |
966 | { | |
967 | unsigned long total_faults; | |
968 | ||
969 | if (!p->numa_group) | |
970 | return 0; | |
971 | ||
972 | total_faults = atomic_long_read(&p->numa_group->total_faults); | |
973 | ||
974 | if (!total_faults) | |
975 | return 0; | |
976 | ||
ca28aa53 | 977 | return 1000 * group_faults(p, nid) / total_faults; |
83e1d2cd MG |
978 | } |
979 | ||
e6628d5b | 980 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
981 | static unsigned long source_load(int cpu, int type); |
982 | static unsigned long target_load(int cpu, int type); | |
983 | static unsigned long power_of(int cpu); | |
984 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
985 | ||
fb13c7ee | 986 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 987 | struct numa_stats { |
fb13c7ee | 988 | unsigned long nr_running; |
58d081b5 | 989 | unsigned long load; |
fb13c7ee MG |
990 | |
991 | /* Total compute capacity of CPUs on a node */ | |
992 | unsigned long power; | |
993 | ||
994 | /* Approximate capacity in terms of runnable tasks on a node */ | |
995 | unsigned long capacity; | |
996 | int has_capacity; | |
58d081b5 | 997 | }; |
e6628d5b | 998 | |
fb13c7ee MG |
999 | /* |
1000 | * XXX borrowed from update_sg_lb_stats | |
1001 | */ | |
1002 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1003 | { | |
1004 | int cpu; | |
1005 | ||
1006 | memset(ns, 0, sizeof(*ns)); | |
1007 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1008 | struct rq *rq = cpu_rq(cpu); | |
1009 | ||
1010 | ns->nr_running += rq->nr_running; | |
1011 | ns->load += weighted_cpuload(cpu); | |
1012 | ns->power += power_of(cpu); | |
1013 | } | |
1014 | ||
1015 | ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power; | |
1016 | ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE); | |
1017 | ns->has_capacity = (ns->nr_running < ns->capacity); | |
1018 | } | |
1019 | ||
58d081b5 MG |
1020 | struct task_numa_env { |
1021 | struct task_struct *p; | |
e6628d5b | 1022 | |
58d081b5 MG |
1023 | int src_cpu, src_nid; |
1024 | int dst_cpu, dst_nid; | |
e6628d5b | 1025 | |
58d081b5 | 1026 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1027 | |
fb13c7ee MG |
1028 | int imbalance_pct, idx; |
1029 | ||
1030 | struct task_struct *best_task; | |
1031 | long best_imp; | |
58d081b5 MG |
1032 | int best_cpu; |
1033 | }; | |
1034 | ||
fb13c7ee MG |
1035 | static void task_numa_assign(struct task_numa_env *env, |
1036 | struct task_struct *p, long imp) | |
1037 | { | |
1038 | if (env->best_task) | |
1039 | put_task_struct(env->best_task); | |
1040 | if (p) | |
1041 | get_task_struct(p); | |
1042 | ||
1043 | env->best_task = p; | |
1044 | env->best_imp = imp; | |
1045 | env->best_cpu = env->dst_cpu; | |
1046 | } | |
1047 | ||
1048 | /* | |
1049 | * This checks if the overall compute and NUMA accesses of the system would | |
1050 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1051 | * into account that it might be best if task running on the dst_cpu should | |
1052 | * be exchanged with the source task | |
1053 | */ | |
887c290e RR |
1054 | static void task_numa_compare(struct task_numa_env *env, |
1055 | long taskimp, long groupimp) | |
fb13c7ee MG |
1056 | { |
1057 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1058 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1059 | struct task_struct *cur; | |
1060 | long dst_load, src_load; | |
1061 | long load; | |
887c290e | 1062 | long imp = (groupimp > 0) ? groupimp : taskimp; |
fb13c7ee MG |
1063 | |
1064 | rcu_read_lock(); | |
1065 | cur = ACCESS_ONCE(dst_rq->curr); | |
1066 | if (cur->pid == 0) /* idle */ | |
1067 | cur = NULL; | |
1068 | ||
1069 | /* | |
1070 | * "imp" is the fault differential for the source task between the | |
1071 | * source and destination node. Calculate the total differential for | |
1072 | * the source task and potential destination task. The more negative | |
1073 | * the value is, the more rmeote accesses that would be expected to | |
1074 | * be incurred if the tasks were swapped. | |
1075 | */ | |
1076 | if (cur) { | |
1077 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1078 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1079 | goto unlock; | |
1080 | ||
887c290e RR |
1081 | /* |
1082 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1083 | * in any group then look only at task weights. |
887c290e | 1084 | */ |
ca28aa53 | 1085 | if (cur->numa_group == env->p->numa_group) { |
887c290e RR |
1086 | imp = taskimp + task_weight(cur, env->src_nid) - |
1087 | task_weight(cur, env->dst_nid); | |
ca28aa53 RR |
1088 | /* |
1089 | * Add some hysteresis to prevent swapping the | |
1090 | * tasks within a group over tiny differences. | |
1091 | */ | |
1092 | if (cur->numa_group) | |
1093 | imp -= imp/16; | |
887c290e | 1094 | } else { |
ca28aa53 RR |
1095 | /* |
1096 | * Compare the group weights. If a task is all by | |
1097 | * itself (not part of a group), use the task weight | |
1098 | * instead. | |
1099 | */ | |
1100 | if (env->p->numa_group) | |
1101 | imp = groupimp; | |
1102 | else | |
1103 | imp = taskimp; | |
1104 | ||
1105 | if (cur->numa_group) | |
1106 | imp += group_weight(cur, env->src_nid) - | |
1107 | group_weight(cur, env->dst_nid); | |
1108 | else | |
1109 | imp += task_weight(cur, env->src_nid) - | |
1110 | task_weight(cur, env->dst_nid); | |
887c290e | 1111 | } |
fb13c7ee MG |
1112 | } |
1113 | ||
1114 | if (imp < env->best_imp) | |
1115 | goto unlock; | |
1116 | ||
1117 | if (!cur) { | |
1118 | /* Is there capacity at our destination? */ | |
1119 | if (env->src_stats.has_capacity && | |
1120 | !env->dst_stats.has_capacity) | |
1121 | goto unlock; | |
1122 | ||
1123 | goto balance; | |
1124 | } | |
1125 | ||
1126 | /* Balance doesn't matter much if we're running a task per cpu */ | |
1127 | if (src_rq->nr_running == 1 && dst_rq->nr_running == 1) | |
1128 | goto assign; | |
1129 | ||
1130 | /* | |
1131 | * In the overloaded case, try and keep the load balanced. | |
1132 | */ | |
1133 | balance: | |
1134 | dst_load = env->dst_stats.load; | |
1135 | src_load = env->src_stats.load; | |
1136 | ||
1137 | /* XXX missing power terms */ | |
1138 | load = task_h_load(env->p); | |
1139 | dst_load += load; | |
1140 | src_load -= load; | |
1141 | ||
1142 | if (cur) { | |
1143 | load = task_h_load(cur); | |
1144 | dst_load -= load; | |
1145 | src_load += load; | |
1146 | } | |
1147 | ||
1148 | /* make src_load the smaller */ | |
1149 | if (dst_load < src_load) | |
1150 | swap(dst_load, src_load); | |
1151 | ||
1152 | if (src_load * env->imbalance_pct < dst_load * 100) | |
1153 | goto unlock; | |
1154 | ||
1155 | assign: | |
1156 | task_numa_assign(env, cur, imp); | |
1157 | unlock: | |
1158 | rcu_read_unlock(); | |
1159 | } | |
1160 | ||
887c290e RR |
1161 | static void task_numa_find_cpu(struct task_numa_env *env, |
1162 | long taskimp, long groupimp) | |
2c8a50aa MG |
1163 | { |
1164 | int cpu; | |
1165 | ||
1166 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1167 | /* Skip this CPU if the source task cannot migrate */ | |
1168 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1169 | continue; | |
1170 | ||
1171 | env->dst_cpu = cpu; | |
887c290e | 1172 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1173 | } |
1174 | } | |
1175 | ||
58d081b5 MG |
1176 | static int task_numa_migrate(struct task_struct *p) |
1177 | { | |
58d081b5 MG |
1178 | struct task_numa_env env = { |
1179 | .p = p, | |
fb13c7ee | 1180 | |
58d081b5 | 1181 | .src_cpu = task_cpu(p), |
b32e86b4 | 1182 | .src_nid = task_node(p), |
fb13c7ee MG |
1183 | |
1184 | .imbalance_pct = 112, | |
1185 | ||
1186 | .best_task = NULL, | |
1187 | .best_imp = 0, | |
1188 | .best_cpu = -1 | |
58d081b5 MG |
1189 | }; |
1190 | struct sched_domain *sd; | |
887c290e | 1191 | unsigned long taskweight, groupweight; |
2c8a50aa | 1192 | int nid, ret; |
887c290e | 1193 | long taskimp, groupimp; |
e6628d5b | 1194 | |
58d081b5 | 1195 | /* |
fb13c7ee MG |
1196 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1197 | * imbalance and would be the first to start moving tasks about. | |
1198 | * | |
1199 | * And we want to avoid any moving of tasks about, as that would create | |
1200 | * random movement of tasks -- counter the numa conditions we're trying | |
1201 | * to satisfy here. | |
58d081b5 MG |
1202 | */ |
1203 | rcu_read_lock(); | |
fb13c7ee MG |
1204 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
1205 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1206 | rcu_read_unlock(); |
1207 | ||
887c290e RR |
1208 | taskweight = task_weight(p, env.src_nid); |
1209 | groupweight = group_weight(p, env.src_nid); | |
fb13c7ee | 1210 | update_numa_stats(&env.src_stats, env.src_nid); |
2c8a50aa | 1211 | env.dst_nid = p->numa_preferred_nid; |
887c290e RR |
1212 | taskimp = task_weight(p, env.dst_nid) - taskweight; |
1213 | groupimp = group_weight(p, env.dst_nid) - groupweight; | |
2c8a50aa | 1214 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1215 | |
e1dda8a7 RR |
1216 | /* If the preferred nid has capacity, try to use it. */ |
1217 | if (env.dst_stats.has_capacity) | |
887c290e | 1218 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 RR |
1219 | |
1220 | /* No space available on the preferred nid. Look elsewhere. */ | |
1221 | if (env.best_cpu == -1) { | |
2c8a50aa MG |
1222 | for_each_online_node(nid) { |
1223 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1224 | continue; | |
58d081b5 | 1225 | |
83e1d2cd | 1226 | /* Only consider nodes where both task and groups benefit */ |
887c290e RR |
1227 | taskimp = task_weight(p, nid) - taskweight; |
1228 | groupimp = group_weight(p, nid) - groupweight; | |
1229 | if (taskimp < 0 && groupimp < 0) | |
fb13c7ee MG |
1230 | continue; |
1231 | ||
2c8a50aa MG |
1232 | env.dst_nid = nid; |
1233 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
887c290e | 1234 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1235 | } |
1236 | } | |
1237 | ||
fb13c7ee MG |
1238 | /* No better CPU than the current one was found. */ |
1239 | if (env.best_cpu == -1) | |
1240 | return -EAGAIN; | |
1241 | ||
0ec8aa00 PZ |
1242 | sched_setnuma(p, env.dst_nid); |
1243 | ||
04bb2f94 RR |
1244 | /* |
1245 | * Reset the scan period if the task is being rescheduled on an | |
1246 | * alternative node to recheck if the tasks is now properly placed. | |
1247 | */ | |
1248 | p->numa_scan_period = task_scan_min(p); | |
1249 | ||
fb13c7ee MG |
1250 | if (env.best_task == NULL) { |
1251 | int ret = migrate_task_to(p, env.best_cpu); | |
1252 | return ret; | |
1253 | } | |
1254 | ||
1255 | ret = migrate_swap(p, env.best_task); | |
1256 | put_task_struct(env.best_task); | |
1257 | return ret; | |
e6628d5b MG |
1258 | } |
1259 | ||
6b9a7460 MG |
1260 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1261 | static void numa_migrate_preferred(struct task_struct *p) | |
1262 | { | |
1263 | /* Success if task is already running on preferred CPU */ | |
1264 | p->numa_migrate_retry = 0; | |
06ea5e03 RR |
1265 | if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) { |
1266 | /* | |
1267 | * If migration is temporarily disabled due to a task migration | |
1268 | * then re-enable it now as the task is running on its | |
1269 | * preferred node and memory should migrate locally | |
1270 | */ | |
1271 | if (!p->numa_migrate_seq) | |
1272 | p->numa_migrate_seq++; | |
6b9a7460 | 1273 | return; |
06ea5e03 | 1274 | } |
6b9a7460 MG |
1275 | |
1276 | /* This task has no NUMA fault statistics yet */ | |
1277 | if (unlikely(p->numa_preferred_nid == -1)) | |
1278 | return; | |
1279 | ||
1280 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
1281 | if (task_numa_migrate(p) != 0) | |
1282 | p->numa_migrate_retry = jiffies + HZ*5; | |
1283 | } | |
1284 | ||
04bb2f94 RR |
1285 | /* |
1286 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1287 | * increments. The more local the fault statistics are, the higher the scan | |
1288 | * period will be for the next scan window. If local/remote ratio is below | |
1289 | * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the | |
1290 | * scan period will decrease | |
1291 | */ | |
1292 | #define NUMA_PERIOD_SLOTS 10 | |
1293 | #define NUMA_PERIOD_THRESHOLD 3 | |
1294 | ||
1295 | /* | |
1296 | * Increase the scan period (slow down scanning) if the majority of | |
1297 | * our memory is already on our local node, or if the majority of | |
1298 | * the page accesses are shared with other processes. | |
1299 | * Otherwise, decrease the scan period. | |
1300 | */ | |
1301 | static void update_task_scan_period(struct task_struct *p, | |
1302 | unsigned long shared, unsigned long private) | |
1303 | { | |
1304 | unsigned int period_slot; | |
1305 | int ratio; | |
1306 | int diff; | |
1307 | ||
1308 | unsigned long remote = p->numa_faults_locality[0]; | |
1309 | unsigned long local = p->numa_faults_locality[1]; | |
1310 | ||
1311 | /* | |
1312 | * If there were no record hinting faults then either the task is | |
1313 | * completely idle or all activity is areas that are not of interest | |
1314 | * to automatic numa balancing. Scan slower | |
1315 | */ | |
1316 | if (local + shared == 0) { | |
1317 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1318 | p->numa_scan_period << 1); | |
1319 | ||
1320 | p->mm->numa_next_scan = jiffies + | |
1321 | msecs_to_jiffies(p->numa_scan_period); | |
1322 | ||
1323 | return; | |
1324 | } | |
1325 | ||
1326 | /* | |
1327 | * Prepare to scale scan period relative to the current period. | |
1328 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1329 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1330 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1331 | */ | |
1332 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1333 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1334 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1335 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1336 | if (!slot) | |
1337 | slot = 1; | |
1338 | diff = slot * period_slot; | |
1339 | } else { | |
1340 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1341 | ||
1342 | /* | |
1343 | * Scale scan rate increases based on sharing. There is an | |
1344 | * inverse relationship between the degree of sharing and | |
1345 | * the adjustment made to the scanning period. Broadly | |
1346 | * speaking the intent is that there is little point | |
1347 | * scanning faster if shared accesses dominate as it may | |
1348 | * simply bounce migrations uselessly | |
1349 | */ | |
1350 | period_slot = DIV_ROUND_UP(diff, NUMA_PERIOD_SLOTS); | |
1351 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared)); | |
1352 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; | |
1353 | } | |
1354 | ||
1355 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1356 | task_scan_min(p), task_scan_max(p)); | |
1357 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1358 | } | |
1359 | ||
cbee9f88 PZ |
1360 | static void task_numa_placement(struct task_struct *p) |
1361 | { | |
83e1d2cd MG |
1362 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1363 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 1364 | unsigned long fault_types[2] = { 0, 0 }; |
7dbd13ed | 1365 | spinlock_t *group_lock = NULL; |
cbee9f88 | 1366 | |
2832bc19 | 1367 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1368 | if (p->numa_scan_seq == seq) |
1369 | return; | |
1370 | p->numa_scan_seq = seq; | |
3a7053b3 | 1371 | p->numa_migrate_seq++; |
598f0ec0 | 1372 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1373 | |
7dbd13ed MG |
1374 | /* If the task is part of a group prevent parallel updates to group stats */ |
1375 | if (p->numa_group) { | |
1376 | group_lock = &p->numa_group->lock; | |
1377 | spin_lock(group_lock); | |
1378 | } | |
1379 | ||
688b7585 MG |
1380 | /* Find the node with the highest number of faults */ |
1381 | for_each_online_node(nid) { | |
83e1d2cd | 1382 | unsigned long faults = 0, group_faults = 0; |
ac8e895b | 1383 | int priv, i; |
745d6147 | 1384 | |
ac8e895b | 1385 | for (priv = 0; priv < 2; priv++) { |
8c8a743c PZ |
1386 | long diff; |
1387 | ||
ac8e895b | 1388 | i = task_faults_idx(nid, priv); |
8c8a743c | 1389 | diff = -p->numa_faults[i]; |
745d6147 | 1390 | |
ac8e895b MG |
1391 | /* Decay existing window, copy faults since last scan */ |
1392 | p->numa_faults[i] >>= 1; | |
1393 | p->numa_faults[i] += p->numa_faults_buffer[i]; | |
04bb2f94 | 1394 | fault_types[priv] += p->numa_faults_buffer[i]; |
ac8e895b | 1395 | p->numa_faults_buffer[i] = 0; |
fb13c7ee MG |
1396 | |
1397 | faults += p->numa_faults[i]; | |
8c8a743c | 1398 | diff += p->numa_faults[i]; |
83e1d2cd | 1399 | p->total_numa_faults += diff; |
8c8a743c PZ |
1400 | if (p->numa_group) { |
1401 | /* safe because we can only change our own group */ | |
1402 | atomic_long_add(diff, &p->numa_group->faults[i]); | |
83e1d2cd MG |
1403 | atomic_long_add(diff, &p->numa_group->total_faults); |
1404 | group_faults += atomic_long_read(&p->numa_group->faults[i]); | |
8c8a743c | 1405 | } |
ac8e895b MG |
1406 | } |
1407 | ||
688b7585 MG |
1408 | if (faults > max_faults) { |
1409 | max_faults = faults; | |
1410 | max_nid = nid; | |
1411 | } | |
83e1d2cd MG |
1412 | |
1413 | if (group_faults > max_group_faults) { | |
1414 | max_group_faults = group_faults; | |
1415 | max_group_nid = nid; | |
1416 | } | |
1417 | } | |
1418 | ||
04bb2f94 RR |
1419 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1420 | ||
7dbd13ed MG |
1421 | if (p->numa_group) { |
1422 | /* | |
1423 | * If the preferred task and group nids are different, | |
1424 | * iterate over the nodes again to find the best place. | |
1425 | */ | |
1426 | if (max_nid != max_group_nid) { | |
1427 | unsigned long weight, max_weight = 0; | |
1428 | ||
1429 | for_each_online_node(nid) { | |
1430 | weight = task_weight(p, nid) + group_weight(p, nid); | |
1431 | if (weight > max_weight) { | |
1432 | max_weight = weight; | |
1433 | max_nid = nid; | |
1434 | } | |
83e1d2cd MG |
1435 | } |
1436 | } | |
7dbd13ed MG |
1437 | |
1438 | spin_unlock(group_lock); | |
688b7585 MG |
1439 | } |
1440 | ||
6b9a7460 | 1441 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1442 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1443 | /* Update the preferred nid and migrate task if possible */ |
0ec8aa00 | 1444 | sched_setnuma(p, max_nid); |
6b9a7460 | 1445 | numa_migrate_preferred(p); |
3a7053b3 | 1446 | } |
cbee9f88 PZ |
1447 | } |
1448 | ||
8c8a743c PZ |
1449 | static inline int get_numa_group(struct numa_group *grp) |
1450 | { | |
1451 | return atomic_inc_not_zero(&grp->refcount); | |
1452 | } | |
1453 | ||
1454 | static inline void put_numa_group(struct numa_group *grp) | |
1455 | { | |
1456 | if (atomic_dec_and_test(&grp->refcount)) | |
1457 | kfree_rcu(grp, rcu); | |
1458 | } | |
1459 | ||
1460 | static void double_lock(spinlock_t *l1, spinlock_t *l2) | |
1461 | { | |
1462 | if (l1 > l2) | |
1463 | swap(l1, l2); | |
1464 | ||
1465 | spin_lock(l1); | |
1466 | spin_lock_nested(l2, SINGLE_DEPTH_NESTING); | |
1467 | } | |
1468 | ||
3e6a9418 MG |
1469 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1470 | int *priv) | |
8c8a743c PZ |
1471 | { |
1472 | struct numa_group *grp, *my_grp; | |
1473 | struct task_struct *tsk; | |
1474 | bool join = false; | |
1475 | int cpu = cpupid_to_cpu(cpupid); | |
1476 | int i; | |
1477 | ||
1478 | if (unlikely(!p->numa_group)) { | |
1479 | unsigned int size = sizeof(struct numa_group) + | |
1480 | 2*nr_node_ids*sizeof(atomic_long_t); | |
1481 | ||
1482 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1483 | if (!grp) | |
1484 | return; | |
1485 | ||
1486 | atomic_set(&grp->refcount, 1); | |
1487 | spin_lock_init(&grp->lock); | |
1488 | INIT_LIST_HEAD(&grp->task_list); | |
e29cf08b | 1489 | grp->gid = p->pid; |
8c8a743c PZ |
1490 | |
1491 | for (i = 0; i < 2*nr_node_ids; i++) | |
1492 | atomic_long_set(&grp->faults[i], p->numa_faults[i]); | |
1493 | ||
83e1d2cd MG |
1494 | atomic_long_set(&grp->total_faults, p->total_numa_faults); |
1495 | ||
8c8a743c PZ |
1496 | list_add(&p->numa_entry, &grp->task_list); |
1497 | grp->nr_tasks++; | |
1498 | rcu_assign_pointer(p->numa_group, grp); | |
1499 | } | |
1500 | ||
1501 | rcu_read_lock(); | |
1502 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1503 | ||
1504 | if (!cpupid_match_pid(tsk, cpupid)) | |
1505 | goto unlock; | |
1506 | ||
1507 | grp = rcu_dereference(tsk->numa_group); | |
1508 | if (!grp) | |
1509 | goto unlock; | |
1510 | ||
1511 | my_grp = p->numa_group; | |
1512 | if (grp == my_grp) | |
1513 | goto unlock; | |
1514 | ||
1515 | /* | |
1516 | * Only join the other group if its bigger; if we're the bigger group, | |
1517 | * the other task will join us. | |
1518 | */ | |
1519 | if (my_grp->nr_tasks > grp->nr_tasks) | |
1520 | goto unlock; | |
1521 | ||
1522 | /* | |
1523 | * Tie-break on the grp address. | |
1524 | */ | |
1525 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
1526 | goto unlock; | |
1527 | ||
dabe1d99 RR |
1528 | /* Always join threads in the same process. */ |
1529 | if (tsk->mm == current->mm) | |
1530 | join = true; | |
1531 | ||
1532 | /* Simple filter to avoid false positives due to PID collisions */ | |
1533 | if (flags & TNF_SHARED) | |
1534 | join = true; | |
8c8a743c | 1535 | |
3e6a9418 MG |
1536 | /* Update priv based on whether false sharing was detected */ |
1537 | *priv = !join; | |
1538 | ||
dabe1d99 RR |
1539 | if (join && !get_numa_group(grp)) |
1540 | join = false; | |
8c8a743c PZ |
1541 | |
1542 | unlock: | |
1543 | rcu_read_unlock(); | |
1544 | ||
1545 | if (!join) | |
1546 | return; | |
1547 | ||
1548 | for (i = 0; i < 2*nr_node_ids; i++) { | |
1549 | atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]); | |
1550 | atomic_long_add(p->numa_faults[i], &grp->faults[i]); | |
1551 | } | |
83e1d2cd MG |
1552 | atomic_long_sub(p->total_numa_faults, &my_grp->total_faults); |
1553 | atomic_long_add(p->total_numa_faults, &grp->total_faults); | |
8c8a743c PZ |
1554 | |
1555 | double_lock(&my_grp->lock, &grp->lock); | |
1556 | ||
1557 | list_move(&p->numa_entry, &grp->task_list); | |
1558 | my_grp->nr_tasks--; | |
1559 | grp->nr_tasks++; | |
1560 | ||
1561 | spin_unlock(&my_grp->lock); | |
1562 | spin_unlock(&grp->lock); | |
1563 | ||
1564 | rcu_assign_pointer(p->numa_group, grp); | |
1565 | ||
1566 | put_numa_group(my_grp); | |
1567 | } | |
1568 | ||
1569 | void task_numa_free(struct task_struct *p) | |
1570 | { | |
1571 | struct numa_group *grp = p->numa_group; | |
1572 | int i; | |
82727018 | 1573 | void *numa_faults = p->numa_faults; |
8c8a743c PZ |
1574 | |
1575 | if (grp) { | |
1576 | for (i = 0; i < 2*nr_node_ids; i++) | |
1577 | atomic_long_sub(p->numa_faults[i], &grp->faults[i]); | |
1578 | ||
83e1d2cd MG |
1579 | atomic_long_sub(p->total_numa_faults, &grp->total_faults); |
1580 | ||
8c8a743c PZ |
1581 | spin_lock(&grp->lock); |
1582 | list_del(&p->numa_entry); | |
1583 | grp->nr_tasks--; | |
1584 | spin_unlock(&grp->lock); | |
1585 | rcu_assign_pointer(p->numa_group, NULL); | |
1586 | put_numa_group(grp); | |
1587 | } | |
1588 | ||
82727018 RR |
1589 | p->numa_faults = NULL; |
1590 | p->numa_faults_buffer = NULL; | |
1591 | kfree(numa_faults); | |
8c8a743c PZ |
1592 | } |
1593 | ||
cbee9f88 PZ |
1594 | /* |
1595 | * Got a PROT_NONE fault for a page on @node. | |
1596 | */ | |
6688cc05 | 1597 | void task_numa_fault(int last_cpupid, int node, int pages, int flags) |
cbee9f88 PZ |
1598 | { |
1599 | struct task_struct *p = current; | |
6688cc05 | 1600 | bool migrated = flags & TNF_MIGRATED; |
ac8e895b | 1601 | int priv; |
cbee9f88 | 1602 | |
10e84b97 | 1603 | if (!numabalancing_enabled) |
1a687c2e MG |
1604 | return; |
1605 | ||
9ff1d9ff MG |
1606 | /* for example, ksmd faulting in a user's mm */ |
1607 | if (!p->mm) | |
1608 | return; | |
1609 | ||
82727018 RR |
1610 | /* Do not worry about placement if exiting */ |
1611 | if (p->state == TASK_DEAD) | |
1612 | return; | |
1613 | ||
f809ca9a MG |
1614 | /* Allocate buffer to track faults on a per-node basis */ |
1615 | if (unlikely(!p->numa_faults)) { | |
ac8e895b | 1616 | int size = sizeof(*p->numa_faults) * 2 * nr_node_ids; |
f809ca9a | 1617 | |
745d6147 MG |
1618 | /* numa_faults and numa_faults_buffer share the allocation */ |
1619 | p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN); | |
f809ca9a MG |
1620 | if (!p->numa_faults) |
1621 | return; | |
745d6147 MG |
1622 | |
1623 | BUG_ON(p->numa_faults_buffer); | |
ac8e895b | 1624 | p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids); |
83e1d2cd | 1625 | p->total_numa_faults = 0; |
04bb2f94 | 1626 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 1627 | } |
cbee9f88 | 1628 | |
8c8a743c PZ |
1629 | /* |
1630 | * First accesses are treated as private, otherwise consider accesses | |
1631 | * to be private if the accessing pid has not changed | |
1632 | */ | |
1633 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
1634 | priv = 1; | |
1635 | } else { | |
1636 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 1637 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 1638 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
1639 | } |
1640 | ||
cbee9f88 | 1641 | task_numa_placement(p); |
f809ca9a | 1642 | |
6b9a7460 MG |
1643 | /* Retry task to preferred node migration if it previously failed */ |
1644 | if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry)) | |
1645 | numa_migrate_preferred(p); | |
1646 | ||
b32e86b4 IM |
1647 | if (migrated) |
1648 | p->numa_pages_migrated += pages; | |
1649 | ||
ac8e895b | 1650 | p->numa_faults_buffer[task_faults_idx(node, priv)] += pages; |
04bb2f94 | 1651 | p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages; |
cbee9f88 PZ |
1652 | } |
1653 | ||
6e5fb223 PZ |
1654 | static void reset_ptenuma_scan(struct task_struct *p) |
1655 | { | |
1656 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1657 | p->mm->numa_scan_offset = 0; | |
1658 | } | |
1659 | ||
cbee9f88 PZ |
1660 | /* |
1661 | * The expensive part of numa migration is done from task_work context. | |
1662 | * Triggered from task_tick_numa(). | |
1663 | */ | |
1664 | void task_numa_work(struct callback_head *work) | |
1665 | { | |
1666 | unsigned long migrate, next_scan, now = jiffies; | |
1667 | struct task_struct *p = current; | |
1668 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1669 | struct vm_area_struct *vma; |
9f40604c | 1670 | unsigned long start, end; |
598f0ec0 | 1671 | unsigned long nr_pte_updates = 0; |
9f40604c | 1672 | long pages; |
cbee9f88 PZ |
1673 | |
1674 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1675 | ||
1676 | work->next = work; /* protect against double add */ | |
1677 | /* | |
1678 | * Who cares about NUMA placement when they're dying. | |
1679 | * | |
1680 | * NOTE: make sure not to dereference p->mm before this check, | |
1681 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1682 | * without p->mm even though we still had it when we enqueued this | |
1683 | * work. | |
1684 | */ | |
1685 | if (p->flags & PF_EXITING) | |
1686 | return; | |
1687 | ||
7e8d16b6 MG |
1688 | if (!mm->numa_next_reset || !mm->numa_next_scan) { |
1689 | mm->numa_next_scan = now + | |
1690 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
1691 | mm->numa_next_reset = now + | |
1692 | msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
1693 | } | |
1694 | ||
b8593bfd MG |
1695 | /* |
1696 | * Reset the scan period if enough time has gone by. Objective is that | |
1697 | * scanning will be reduced if pages are properly placed. As tasks | |
1698 | * can enter different phases this needs to be re-examined. Lacking | |
1699 | * proper tracking of reference behaviour, this blunt hammer is used. | |
1700 | */ | |
1701 | migrate = mm->numa_next_reset; | |
1702 | if (time_after(now, migrate)) { | |
598f0ec0 | 1703 | p->numa_scan_period = task_scan_min(p); |
b8593bfd MG |
1704 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); |
1705 | xchg(&mm->numa_next_reset, next_scan); | |
1706 | } | |
1707 | ||
cbee9f88 PZ |
1708 | /* |
1709 | * Enforce maximal scan/migration frequency.. | |
1710 | */ | |
1711 | migrate = mm->numa_next_scan; | |
1712 | if (time_before(now, migrate)) | |
1713 | return; | |
1714 | ||
598f0ec0 MG |
1715 | if (p->numa_scan_period == 0) { |
1716 | p->numa_scan_period_max = task_scan_max(p); | |
1717 | p->numa_scan_period = task_scan_min(p); | |
1718 | } | |
cbee9f88 | 1719 | |
fb003b80 | 1720 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1721 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1722 | return; | |
1723 | ||
19a78d11 PZ |
1724 | /* |
1725 | * Delay this task enough that another task of this mm will likely win | |
1726 | * the next time around. | |
1727 | */ | |
1728 | p->node_stamp += 2 * TICK_NSEC; | |
1729 | ||
9f40604c MG |
1730 | start = mm->numa_scan_offset; |
1731 | pages = sysctl_numa_balancing_scan_size; | |
1732 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1733 | if (!pages) | |
1734 | return; | |
cbee9f88 | 1735 | |
6e5fb223 | 1736 | down_read(&mm->mmap_sem); |
9f40604c | 1737 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1738 | if (!vma) { |
1739 | reset_ptenuma_scan(p); | |
9f40604c | 1740 | start = 0; |
6e5fb223 PZ |
1741 | vma = mm->mmap; |
1742 | } | |
9f40604c | 1743 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1744 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1745 | continue; |
1746 | ||
4591ce4f MG |
1747 | /* |
1748 | * Shared library pages mapped by multiple processes are not | |
1749 | * migrated as it is expected they are cache replicated. Avoid | |
1750 | * hinting faults in read-only file-backed mappings or the vdso | |
1751 | * as migrating the pages will be of marginal benefit. | |
1752 | */ | |
1753 | if (!vma->vm_mm || | |
1754 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1755 | continue; | |
1756 | ||
9f40604c MG |
1757 | do { |
1758 | start = max(start, vma->vm_start); | |
1759 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1760 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1761 | nr_pte_updates += change_prot_numa(vma, start, end); |
1762 | ||
1763 | /* | |
1764 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1765 | * at least one PTE is updated so that unused virtual | |
1766 | * address space is quickly skipped. | |
1767 | */ | |
1768 | if (nr_pte_updates) | |
1769 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1770 | |
9f40604c MG |
1771 | start = end; |
1772 | if (pages <= 0) | |
1773 | goto out; | |
1774 | } while (end != vma->vm_end); | |
cbee9f88 | 1775 | } |
6e5fb223 | 1776 | |
9f40604c | 1777 | out: |
6e5fb223 | 1778 | /* |
c69307d5 PZ |
1779 | * It is possible to reach the end of the VMA list but the last few |
1780 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1781 | * would find the !migratable VMA on the next scan but not reset the | |
1782 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1783 | */ |
1784 | if (vma) | |
9f40604c | 1785 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1786 | else |
1787 | reset_ptenuma_scan(p); | |
1788 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1789 | } |
1790 | ||
1791 | /* | |
1792 | * Drive the periodic memory faults.. | |
1793 | */ | |
1794 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1795 | { | |
1796 | struct callback_head *work = &curr->numa_work; | |
1797 | u64 period, now; | |
1798 | ||
1799 | /* | |
1800 | * We don't care about NUMA placement if we don't have memory. | |
1801 | */ | |
1802 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1803 | return; | |
1804 | ||
1805 | /* | |
1806 | * Using runtime rather than walltime has the dual advantage that | |
1807 | * we (mostly) drive the selection from busy threads and that the | |
1808 | * task needs to have done some actual work before we bother with | |
1809 | * NUMA placement. | |
1810 | */ | |
1811 | now = curr->se.sum_exec_runtime; | |
1812 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1813 | ||
1814 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1815 | if (!curr->node_stamp) |
598f0ec0 | 1816 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1817 | curr->node_stamp += period; |
cbee9f88 PZ |
1818 | |
1819 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1820 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1821 | task_work_add(curr, work, true); | |
1822 | } | |
1823 | } | |
1824 | } | |
1825 | #else | |
1826 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1827 | { | |
1828 | } | |
0ec8aa00 PZ |
1829 | |
1830 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
1831 | { | |
1832 | } | |
1833 | ||
1834 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1835 | { | |
1836 | } | |
cbee9f88 PZ |
1837 | #endif /* CONFIG_NUMA_BALANCING */ |
1838 | ||
30cfdcfc DA |
1839 | static void |
1840 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1841 | { | |
1842 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1843 | if (!parent_entity(se)) |
029632fb | 1844 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1845 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
1846 | if (entity_is_task(se)) { |
1847 | struct rq *rq = rq_of(cfs_rq); | |
1848 | ||
1849 | account_numa_enqueue(rq, task_of(se)); | |
1850 | list_add(&se->group_node, &rq->cfs_tasks); | |
1851 | } | |
367456c7 | 1852 | #endif |
30cfdcfc | 1853 | cfs_rq->nr_running++; |
30cfdcfc DA |
1854 | } |
1855 | ||
1856 | static void | |
1857 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1858 | { | |
1859 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1860 | if (!parent_entity(se)) |
029632fb | 1861 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
0ec8aa00 PZ |
1862 | if (entity_is_task(se)) { |
1863 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 1864 | list_del_init(&se->group_node); |
0ec8aa00 | 1865 | } |
30cfdcfc | 1866 | cfs_rq->nr_running--; |
30cfdcfc DA |
1867 | } |
1868 | ||
3ff6dcac YZ |
1869 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1870 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1871 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1872 | { | |
1873 | long tg_weight; | |
1874 | ||
1875 | /* | |
1876 | * Use this CPU's actual weight instead of the last load_contribution | |
1877 | * to gain a more accurate current total weight. See | |
1878 | * update_cfs_rq_load_contribution(). | |
1879 | */ | |
bf5b986e | 1880 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 1881 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
1882 | tg_weight += cfs_rq->load.weight; |
1883 | ||
1884 | return tg_weight; | |
1885 | } | |
1886 | ||
6d5ab293 | 1887 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1888 | { |
cf5f0acf | 1889 | long tg_weight, load, shares; |
3ff6dcac | 1890 | |
cf5f0acf | 1891 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1892 | load = cfs_rq->load.weight; |
3ff6dcac | 1893 | |
3ff6dcac | 1894 | shares = (tg->shares * load); |
cf5f0acf PZ |
1895 | if (tg_weight) |
1896 | shares /= tg_weight; | |
3ff6dcac YZ |
1897 | |
1898 | if (shares < MIN_SHARES) | |
1899 | shares = MIN_SHARES; | |
1900 | if (shares > tg->shares) | |
1901 | shares = tg->shares; | |
1902 | ||
1903 | return shares; | |
1904 | } | |
3ff6dcac | 1905 | # else /* CONFIG_SMP */ |
6d5ab293 | 1906 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1907 | { |
1908 | return tg->shares; | |
1909 | } | |
3ff6dcac | 1910 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1911 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1912 | unsigned long weight) | |
1913 | { | |
19e5eebb PT |
1914 | if (se->on_rq) { |
1915 | /* commit outstanding execution time */ | |
1916 | if (cfs_rq->curr == se) | |
1917 | update_curr(cfs_rq); | |
2069dd75 | 1918 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1919 | } |
2069dd75 PZ |
1920 | |
1921 | update_load_set(&se->load, weight); | |
1922 | ||
1923 | if (se->on_rq) | |
1924 | account_entity_enqueue(cfs_rq, se); | |
1925 | } | |
1926 | ||
82958366 PT |
1927 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1928 | ||
6d5ab293 | 1929 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1930 | { |
1931 | struct task_group *tg; | |
1932 | struct sched_entity *se; | |
3ff6dcac | 1933 | long shares; |
2069dd75 | 1934 | |
2069dd75 PZ |
1935 | tg = cfs_rq->tg; |
1936 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1937 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1938 | return; |
3ff6dcac YZ |
1939 | #ifndef CONFIG_SMP |
1940 | if (likely(se->load.weight == tg->shares)) | |
1941 | return; | |
1942 | #endif | |
6d5ab293 | 1943 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1944 | |
1945 | reweight_entity(cfs_rq_of(se), se, shares); | |
1946 | } | |
1947 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1948 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1949 | { |
1950 | } | |
1951 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1952 | ||
141965c7 | 1953 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
1954 | /* |
1955 | * We choose a half-life close to 1 scheduling period. | |
1956 | * Note: The tables below are dependent on this value. | |
1957 | */ | |
1958 | #define LOAD_AVG_PERIOD 32 | |
1959 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1960 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1961 | ||
1962 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1963 | static const u32 runnable_avg_yN_inv[] = { | |
1964 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1965 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1966 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1967 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1968 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1969 | 0x85aac367, 0x82cd8698, | |
1970 | }; | |
1971 | ||
1972 | /* | |
1973 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1974 | * over-estimates when re-combining. | |
1975 | */ | |
1976 | static const u32 runnable_avg_yN_sum[] = { | |
1977 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1978 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1979 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1980 | }; | |
1981 | ||
9d85f21c PT |
1982 | /* |
1983 | * Approximate: | |
1984 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1985 | */ | |
1986 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1987 | { | |
5b51f2f8 PT |
1988 | unsigned int local_n; |
1989 | ||
1990 | if (!n) | |
1991 | return val; | |
1992 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1993 | return 0; | |
1994 | ||
1995 | /* after bounds checking we can collapse to 32-bit */ | |
1996 | local_n = n; | |
1997 | ||
1998 | /* | |
1999 | * As y^PERIOD = 1/2, we can combine | |
2000 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
2001 | * With a look-up table which covers k^n (n<PERIOD) | |
2002 | * | |
2003 | * To achieve constant time decay_load. | |
2004 | */ | |
2005 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2006 | val >>= local_n / LOAD_AVG_PERIOD; | |
2007 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2008 | } |
2009 | ||
5b51f2f8 PT |
2010 | val *= runnable_avg_yN_inv[local_n]; |
2011 | /* We don't use SRR here since we always want to round down. */ | |
2012 | return val >> 32; | |
2013 | } | |
2014 | ||
2015 | /* | |
2016 | * For updates fully spanning n periods, the contribution to runnable | |
2017 | * average will be: \Sum 1024*y^n | |
2018 | * | |
2019 | * We can compute this reasonably efficiently by combining: | |
2020 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2021 | */ | |
2022 | static u32 __compute_runnable_contrib(u64 n) | |
2023 | { | |
2024 | u32 contrib = 0; | |
2025 | ||
2026 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2027 | return runnable_avg_yN_sum[n]; | |
2028 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2029 | return LOAD_AVG_MAX; | |
2030 | ||
2031 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
2032 | do { | |
2033 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
2034 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
2035 | ||
2036 | n -= LOAD_AVG_PERIOD; | |
2037 | } while (n > LOAD_AVG_PERIOD); | |
2038 | ||
2039 | contrib = decay_load(contrib, n); | |
2040 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2041 | } |
2042 | ||
2043 | /* | |
2044 | * We can represent the historical contribution to runnable average as the | |
2045 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2046 | * history into segments of approximately 1ms (1024us); label the segment that | |
2047 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2048 | * | |
2049 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2050 | * p0 p1 p2 | |
2051 | * (now) (~1ms ago) (~2ms ago) | |
2052 | * | |
2053 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2054 | * | |
2055 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2056 | * following representation of historical load: | |
2057 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2058 | * | |
2059 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2060 | * y^32 = 0.5 | |
2061 | * | |
2062 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2063 | * approximately half as much as the contribution to load within the last ms | |
2064 | * (u_0). | |
2065 | * | |
2066 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2067 | * sum again by y is sufficient to update: | |
2068 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2069 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2070 | */ | |
2071 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
2072 | struct sched_avg *sa, | |
2073 | int runnable) | |
2074 | { | |
5b51f2f8 PT |
2075 | u64 delta, periods; |
2076 | u32 runnable_contrib; | |
9d85f21c PT |
2077 | int delta_w, decayed = 0; |
2078 | ||
2079 | delta = now - sa->last_runnable_update; | |
2080 | /* | |
2081 | * This should only happen when time goes backwards, which it | |
2082 | * unfortunately does during sched clock init when we swap over to TSC. | |
2083 | */ | |
2084 | if ((s64)delta < 0) { | |
2085 | sa->last_runnable_update = now; | |
2086 | return 0; | |
2087 | } | |
2088 | ||
2089 | /* | |
2090 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2091 | * approximation of 1us and fast to compute. | |
2092 | */ | |
2093 | delta >>= 10; | |
2094 | if (!delta) | |
2095 | return 0; | |
2096 | sa->last_runnable_update = now; | |
2097 | ||
2098 | /* delta_w is the amount already accumulated against our next period */ | |
2099 | delta_w = sa->runnable_avg_period % 1024; | |
2100 | if (delta + delta_w >= 1024) { | |
2101 | /* period roll-over */ | |
2102 | decayed = 1; | |
2103 | ||
2104 | /* | |
2105 | * Now that we know we're crossing a period boundary, figure | |
2106 | * out how much from delta we need to complete the current | |
2107 | * period and accrue it. | |
2108 | */ | |
2109 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
2110 | if (runnable) |
2111 | sa->runnable_avg_sum += delta_w; | |
2112 | sa->runnable_avg_period += delta_w; | |
2113 | ||
2114 | delta -= delta_w; | |
2115 | ||
2116 | /* Figure out how many additional periods this update spans */ | |
2117 | periods = delta / 1024; | |
2118 | delta %= 1024; | |
2119 | ||
2120 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
2121 | periods + 1); | |
2122 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
2123 | periods + 1); | |
2124 | ||
2125 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
2126 | runnable_contrib = __compute_runnable_contrib(periods); | |
2127 | if (runnable) | |
2128 | sa->runnable_avg_sum += runnable_contrib; | |
2129 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
2130 | } |
2131 | ||
2132 | /* Remainder of delta accrued against u_0` */ | |
2133 | if (runnable) | |
2134 | sa->runnable_avg_sum += delta; | |
2135 | sa->runnable_avg_period += delta; | |
2136 | ||
2137 | return decayed; | |
2138 | } | |
2139 | ||
9ee474f5 | 2140 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 2141 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
2142 | { |
2143 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2144 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
2145 | ||
2146 | decays -= se->avg.decay_count; | |
2147 | if (!decays) | |
aff3e498 | 2148 | return 0; |
9ee474f5 PT |
2149 | |
2150 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
2151 | se->avg.decay_count = 0; | |
aff3e498 PT |
2152 | |
2153 | return decays; | |
9ee474f5 PT |
2154 | } |
2155 | ||
c566e8e9 PT |
2156 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2157 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2158 | int force_update) | |
2159 | { | |
2160 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 2161 | long tg_contrib; |
c566e8e9 PT |
2162 | |
2163 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
2164 | tg_contrib -= cfs_rq->tg_load_contrib; | |
2165 | ||
bf5b986e AS |
2166 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
2167 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
2168 | cfs_rq->tg_load_contrib += tg_contrib; |
2169 | } | |
2170 | } | |
8165e145 | 2171 | |
bb17f655 PT |
2172 | /* |
2173 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
2174 | * representation for computing load contributions. | |
2175 | */ | |
2176 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
2177 | struct cfs_rq *cfs_rq) | |
2178 | { | |
2179 | struct task_group *tg = cfs_rq->tg; | |
2180 | long contrib; | |
2181 | ||
2182 | /* The fraction of a cpu used by this cfs_rq */ | |
2183 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
2184 | sa->runnable_avg_period + 1); | |
2185 | contrib -= cfs_rq->tg_runnable_contrib; | |
2186 | ||
2187 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
2188 | atomic_add(contrib, &tg->runnable_avg); | |
2189 | cfs_rq->tg_runnable_contrib += contrib; | |
2190 | } | |
2191 | } | |
2192 | ||
8165e145 PT |
2193 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
2194 | { | |
2195 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
2196 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
2197 | int runnable_avg; |
2198 | ||
8165e145 PT |
2199 | u64 contrib; |
2200 | ||
2201 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
2202 | se->avg.load_avg_contrib = div_u64(contrib, |
2203 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
2204 | |
2205 | /* | |
2206 | * For group entities we need to compute a correction term in the case | |
2207 | * that they are consuming <1 cpu so that we would contribute the same | |
2208 | * load as a task of equal weight. | |
2209 | * | |
2210 | * Explicitly co-ordinating this measurement would be expensive, but | |
2211 | * fortunately the sum of each cpus contribution forms a usable | |
2212 | * lower-bound on the true value. | |
2213 | * | |
2214 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
2215 | * (and the sum represents true value) or they are disjoint and we are | |
2216 | * understating by the aggregate of their overlap. | |
2217 | * | |
2218 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
2219 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
2220 | * cpus that overlap for this interval and w_i is the interval width. | |
2221 | * | |
2222 | * On a small machine; the first term is well-bounded which bounds the | |
2223 | * total error since w_i is a subset of the period. Whereas on a | |
2224 | * larger machine, while this first term can be larger, if w_i is the | |
2225 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
2226 | * our upper bound of 1-cpu. | |
2227 | */ | |
2228 | runnable_avg = atomic_read(&tg->runnable_avg); | |
2229 | if (runnable_avg < NICE_0_LOAD) { | |
2230 | se->avg.load_avg_contrib *= runnable_avg; | |
2231 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
2232 | } | |
8165e145 | 2233 | } |
c566e8e9 PT |
2234 | #else |
2235 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2236 | int force_update) {} | |
bb17f655 PT |
2237 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2238 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 2239 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
2240 | #endif |
2241 | ||
8165e145 PT |
2242 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2243 | { | |
2244 | u32 contrib; | |
2245 | ||
2246 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2247 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2248 | contrib /= (se->avg.runnable_avg_period + 1); | |
2249 | se->avg.load_avg_contrib = scale_load(contrib); | |
2250 | } | |
2251 | ||
2dac754e PT |
2252 | /* Compute the current contribution to load_avg by se, return any delta */ |
2253 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2254 | { | |
2255 | long old_contrib = se->avg.load_avg_contrib; | |
2256 | ||
8165e145 PT |
2257 | if (entity_is_task(se)) { |
2258 | __update_task_entity_contrib(se); | |
2259 | } else { | |
bb17f655 | 2260 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2261 | __update_group_entity_contrib(se); |
2262 | } | |
2dac754e PT |
2263 | |
2264 | return se->avg.load_avg_contrib - old_contrib; | |
2265 | } | |
2266 | ||
9ee474f5 PT |
2267 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2268 | long load_contrib) | |
2269 | { | |
2270 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2271 | cfs_rq->blocked_load_avg -= load_contrib; | |
2272 | else | |
2273 | cfs_rq->blocked_load_avg = 0; | |
2274 | } | |
2275 | ||
f1b17280 PT |
2276 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2277 | ||
9d85f21c | 2278 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2279 | static inline void update_entity_load_avg(struct sched_entity *se, |
2280 | int update_cfs_rq) | |
9d85f21c | 2281 | { |
2dac754e PT |
2282 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2283 | long contrib_delta; | |
f1b17280 | 2284 | u64 now; |
2dac754e | 2285 | |
f1b17280 PT |
2286 | /* |
2287 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2288 | * case they are the parent of a throttled hierarchy. | |
2289 | */ | |
2290 | if (entity_is_task(se)) | |
2291 | now = cfs_rq_clock_task(cfs_rq); | |
2292 | else | |
2293 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2294 | ||
2295 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2296 | return; |
2297 | ||
2298 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2299 | |
2300 | if (!update_cfs_rq) | |
2301 | return; | |
2302 | ||
2dac754e PT |
2303 | if (se->on_rq) |
2304 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2305 | else |
2306 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2307 | } | |
2308 | ||
2309 | /* | |
2310 | * Decay the load contributed by all blocked children and account this so that | |
2311 | * their contribution may appropriately discounted when they wake up. | |
2312 | */ | |
aff3e498 | 2313 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2314 | { |
f1b17280 | 2315 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2316 | u64 decays; |
2317 | ||
2318 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2319 | if (!decays && !force_update) |
9ee474f5 PT |
2320 | return; |
2321 | ||
2509940f AS |
2322 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2323 | unsigned long removed_load; | |
2324 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2325 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2326 | } | |
9ee474f5 | 2327 | |
aff3e498 PT |
2328 | if (decays) { |
2329 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2330 | decays); | |
2331 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2332 | cfs_rq->last_decay = now; | |
2333 | } | |
c566e8e9 PT |
2334 | |
2335 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2336 | } |
18bf2805 BS |
2337 | |
2338 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2339 | { | |
78becc27 | 2340 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 2341 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 2342 | } |
2dac754e PT |
2343 | |
2344 | /* Add the load generated by se into cfs_rq's child load-average */ | |
2345 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2346 | struct sched_entity *se, |
2347 | int wakeup) | |
2dac754e | 2348 | { |
aff3e498 PT |
2349 | /* |
2350 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2351 | * migration we use a negative decay count to track the remote decays | |
2352 | * accumulated while sleeping. | |
a75cdaa9 AS |
2353 | * |
2354 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2355 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2356 | * constructed load_avg_contrib. | |
aff3e498 PT |
2357 | */ |
2358 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2359 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2360 | if (se->avg.decay_count) { |
2361 | /* | |
2362 | * In a wake-up migration we have to approximate the | |
2363 | * time sleeping. This is because we can't synchronize | |
2364 | * clock_task between the two cpus, and it is not | |
2365 | * guaranteed to be read-safe. Instead, we can | |
2366 | * approximate this using our carried decays, which are | |
2367 | * explicitly atomically readable. | |
2368 | */ | |
2369 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2370 | << 20; | |
2371 | update_entity_load_avg(se, 0); | |
2372 | /* Indicate that we're now synchronized and on-rq */ | |
2373 | se->avg.decay_count = 0; | |
2374 | } | |
9ee474f5 PT |
2375 | wakeup = 0; |
2376 | } else { | |
282cf499 AS |
2377 | /* |
2378 | * Task re-woke on same cpu (or else migrate_task_rq_fair() | |
2379 | * would have made count negative); we must be careful to avoid | |
2380 | * double-accounting blocked time after synchronizing decays. | |
2381 | */ | |
2382 | se->avg.last_runnable_update += __synchronize_entity_decay(se) | |
2383 | << 20; | |
9ee474f5 PT |
2384 | } |
2385 | ||
aff3e498 PT |
2386 | /* migrated tasks did not contribute to our blocked load */ |
2387 | if (wakeup) { | |
9ee474f5 | 2388 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2389 | update_entity_load_avg(se, 0); |
2390 | } | |
9ee474f5 | 2391 | |
2dac754e | 2392 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2393 | /* we force update consideration on load-balancer moves */ |
2394 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2395 | } |
2396 | ||
9ee474f5 PT |
2397 | /* |
2398 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2399 | * transitioning to a blocked state we track its projected decay using | |
2400 | * blocked_load_avg. | |
2401 | */ | |
2dac754e | 2402 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2403 | struct sched_entity *se, |
2404 | int sleep) | |
2dac754e | 2405 | { |
9ee474f5 | 2406 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2407 | /* we force update consideration on load-balancer moves */ |
2408 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2409 | |
2dac754e | 2410 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2411 | if (sleep) { |
2412 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2413 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2414 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2415 | } |
642dbc39 VG |
2416 | |
2417 | /* | |
2418 | * Update the rq's load with the elapsed running time before entering | |
2419 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2420 | * be the only way to update the runnable statistic. | |
2421 | */ | |
2422 | void idle_enter_fair(struct rq *this_rq) | |
2423 | { | |
2424 | update_rq_runnable_avg(this_rq, 1); | |
2425 | } | |
2426 | ||
2427 | /* | |
2428 | * Update the rq's load with the elapsed idle time before a task is | |
2429 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2430 | * be the only way to update the runnable statistic. | |
2431 | */ | |
2432 | void idle_exit_fair(struct rq *this_rq) | |
2433 | { | |
2434 | update_rq_runnable_avg(this_rq, 0); | |
2435 | } | |
2436 | ||
9d85f21c | 2437 | #else |
9ee474f5 PT |
2438 | static inline void update_entity_load_avg(struct sched_entity *se, |
2439 | int update_cfs_rq) {} | |
18bf2805 | 2440 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2441 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2442 | struct sched_entity *se, |
2443 | int wakeup) {} | |
2dac754e | 2444 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2445 | struct sched_entity *se, |
2446 | int sleep) {} | |
aff3e498 PT |
2447 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2448 | int force_update) {} | |
9d85f21c PT |
2449 | #endif |
2450 | ||
2396af69 | 2451 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2452 | { |
bf0f6f24 | 2453 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2454 | struct task_struct *tsk = NULL; |
2455 | ||
2456 | if (entity_is_task(se)) | |
2457 | tsk = task_of(se); | |
2458 | ||
41acab88 | 2459 | if (se->statistics.sleep_start) { |
78becc27 | 2460 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2461 | |
2462 | if ((s64)delta < 0) | |
2463 | delta = 0; | |
2464 | ||
41acab88 LDM |
2465 | if (unlikely(delta > se->statistics.sleep_max)) |
2466 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2467 | |
8c79a045 | 2468 | se->statistics.sleep_start = 0; |
41acab88 | 2469 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2470 | |
768d0c27 | 2471 | if (tsk) { |
e414314c | 2472 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2473 | trace_sched_stat_sleep(tsk, delta); |
2474 | } | |
bf0f6f24 | 2475 | } |
41acab88 | 2476 | if (se->statistics.block_start) { |
78becc27 | 2477 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2478 | |
2479 | if ((s64)delta < 0) | |
2480 | delta = 0; | |
2481 | ||
41acab88 LDM |
2482 | if (unlikely(delta > se->statistics.block_max)) |
2483 | se->statistics.block_max = delta; | |
bf0f6f24 | 2484 | |
8c79a045 | 2485 | se->statistics.block_start = 0; |
41acab88 | 2486 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2487 | |
e414314c | 2488 | if (tsk) { |
8f0dfc34 | 2489 | if (tsk->in_iowait) { |
41acab88 LDM |
2490 | se->statistics.iowait_sum += delta; |
2491 | se->statistics.iowait_count++; | |
768d0c27 | 2492 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2493 | } |
2494 | ||
b781a602 AV |
2495 | trace_sched_stat_blocked(tsk, delta); |
2496 | ||
e414314c PZ |
2497 | /* |
2498 | * Blocking time is in units of nanosecs, so shift by | |
2499 | * 20 to get a milliseconds-range estimation of the | |
2500 | * amount of time that the task spent sleeping: | |
2501 | */ | |
2502 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2503 | profile_hits(SLEEP_PROFILING, | |
2504 | (void *)get_wchan(tsk), | |
2505 | delta >> 20); | |
2506 | } | |
2507 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2508 | } |
bf0f6f24 IM |
2509 | } |
2510 | #endif | |
2511 | } | |
2512 | ||
ddc97297 PZ |
2513 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2514 | { | |
2515 | #ifdef CONFIG_SCHED_DEBUG | |
2516 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2517 | ||
2518 | if (d < 0) | |
2519 | d = -d; | |
2520 | ||
2521 | if (d > 3*sysctl_sched_latency) | |
2522 | schedstat_inc(cfs_rq, nr_spread_over); | |
2523 | #endif | |
2524 | } | |
2525 | ||
aeb73b04 PZ |
2526 | static void |
2527 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2528 | { | |
1af5f730 | 2529 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2530 | |
2cb8600e PZ |
2531 | /* |
2532 | * The 'current' period is already promised to the current tasks, | |
2533 | * however the extra weight of the new task will slow them down a | |
2534 | * little, place the new task so that it fits in the slot that | |
2535 | * stays open at the end. | |
2536 | */ | |
94dfb5e7 | 2537 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2538 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2539 | |
a2e7a7eb | 2540 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2541 | if (!initial) { |
a2e7a7eb | 2542 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2543 | |
a2e7a7eb MG |
2544 | /* |
2545 | * Halve their sleep time's effect, to allow | |
2546 | * for a gentler effect of sleepers: | |
2547 | */ | |
2548 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2549 | thresh >>= 1; | |
51e0304c | 2550 | |
a2e7a7eb | 2551 | vruntime -= thresh; |
aeb73b04 PZ |
2552 | } |
2553 | ||
b5d9d734 | 2554 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2555 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2556 | } |
2557 | ||
d3d9dc33 PT |
2558 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2559 | ||
bf0f6f24 | 2560 | static void |
88ec22d3 | 2561 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2562 | { |
88ec22d3 PZ |
2563 | /* |
2564 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2565 | * through calling update_curr(). |
88ec22d3 | 2566 | */ |
371fd7e7 | 2567 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2568 | se->vruntime += cfs_rq->min_vruntime; |
2569 | ||
bf0f6f24 | 2570 | /* |
a2a2d680 | 2571 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2572 | */ |
b7cc0896 | 2573 | update_curr(cfs_rq); |
f269ae04 | 2574 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2575 | account_entity_enqueue(cfs_rq, se); |
2576 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2577 | |
88ec22d3 | 2578 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2579 | place_entity(cfs_rq, se, 0); |
2396af69 | 2580 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2581 | } |
bf0f6f24 | 2582 | |
d2417e5a | 2583 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2584 | check_spread(cfs_rq, se); |
83b699ed SV |
2585 | if (se != cfs_rq->curr) |
2586 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2587 | se->on_rq = 1; |
3d4b47b4 | 2588 | |
d3d9dc33 | 2589 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2590 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2591 | check_enqueue_throttle(cfs_rq); |
2592 | } | |
bf0f6f24 IM |
2593 | } |
2594 | ||
2c13c919 | 2595 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2596 | { |
2c13c919 RR |
2597 | for_each_sched_entity(se) { |
2598 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2599 | if (cfs_rq->last == se) | |
2600 | cfs_rq->last = NULL; | |
2601 | else | |
2602 | break; | |
2603 | } | |
2604 | } | |
2002c695 | 2605 | |
2c13c919 RR |
2606 | static void __clear_buddies_next(struct sched_entity *se) |
2607 | { | |
2608 | for_each_sched_entity(se) { | |
2609 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2610 | if (cfs_rq->next == se) | |
2611 | cfs_rq->next = NULL; | |
2612 | else | |
2613 | break; | |
2614 | } | |
2002c695 PZ |
2615 | } |
2616 | ||
ac53db59 RR |
2617 | static void __clear_buddies_skip(struct sched_entity *se) |
2618 | { | |
2619 | for_each_sched_entity(se) { | |
2620 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2621 | if (cfs_rq->skip == se) | |
2622 | cfs_rq->skip = NULL; | |
2623 | else | |
2624 | break; | |
2625 | } | |
2626 | } | |
2627 | ||
a571bbea PZ |
2628 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2629 | { | |
2c13c919 RR |
2630 | if (cfs_rq->last == se) |
2631 | __clear_buddies_last(se); | |
2632 | ||
2633 | if (cfs_rq->next == se) | |
2634 | __clear_buddies_next(se); | |
ac53db59 RR |
2635 | |
2636 | if (cfs_rq->skip == se) | |
2637 | __clear_buddies_skip(se); | |
a571bbea PZ |
2638 | } |
2639 | ||
6c16a6dc | 2640 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2641 | |
bf0f6f24 | 2642 | static void |
371fd7e7 | 2643 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2644 | { |
a2a2d680 DA |
2645 | /* |
2646 | * Update run-time statistics of the 'current'. | |
2647 | */ | |
2648 | update_curr(cfs_rq); | |
17bc14b7 | 2649 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2650 | |
19b6a2e3 | 2651 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2652 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2653 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2654 | if (entity_is_task(se)) { |
2655 | struct task_struct *tsk = task_of(se); | |
2656 | ||
2657 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2658 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2659 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2660 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2661 | } |
db36cc7d | 2662 | #endif |
67e9fb2a PZ |
2663 | } |
2664 | ||
2002c695 | 2665 | clear_buddies(cfs_rq, se); |
4793241b | 2666 | |
83b699ed | 2667 | if (se != cfs_rq->curr) |
30cfdcfc | 2668 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2669 | se->on_rq = 0; |
30cfdcfc | 2670 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2671 | |
2672 | /* | |
2673 | * Normalize the entity after updating the min_vruntime because the | |
2674 | * update can refer to the ->curr item and we need to reflect this | |
2675 | * movement in our normalized position. | |
2676 | */ | |
371fd7e7 | 2677 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2678 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2679 | |
d8b4986d PT |
2680 | /* return excess runtime on last dequeue */ |
2681 | return_cfs_rq_runtime(cfs_rq); | |
2682 | ||
1e876231 | 2683 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2684 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2685 | } |
2686 | ||
2687 | /* | |
2688 | * Preempt the current task with a newly woken task if needed: | |
2689 | */ | |
7c92e54f | 2690 | static void |
2e09bf55 | 2691 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2692 | { |
11697830 | 2693 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2694 | struct sched_entity *se; |
2695 | s64 delta; | |
11697830 | 2696 | |
6d0f0ebd | 2697 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2698 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2699 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2700 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2701 | /* |
2702 | * The current task ran long enough, ensure it doesn't get | |
2703 | * re-elected due to buddy favours. | |
2704 | */ | |
2705 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2706 | return; |
2707 | } | |
2708 | ||
2709 | /* | |
2710 | * Ensure that a task that missed wakeup preemption by a | |
2711 | * narrow margin doesn't have to wait for a full slice. | |
2712 | * This also mitigates buddy induced latencies under load. | |
2713 | */ | |
f685ceac MG |
2714 | if (delta_exec < sysctl_sched_min_granularity) |
2715 | return; | |
2716 | ||
f4cfb33e WX |
2717 | se = __pick_first_entity(cfs_rq); |
2718 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2719 | |
f4cfb33e WX |
2720 | if (delta < 0) |
2721 | return; | |
d7d82944 | 2722 | |
f4cfb33e WX |
2723 | if (delta > ideal_runtime) |
2724 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2725 | } |
2726 | ||
83b699ed | 2727 | static void |
8494f412 | 2728 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2729 | { |
83b699ed SV |
2730 | /* 'current' is not kept within the tree. */ |
2731 | if (se->on_rq) { | |
2732 | /* | |
2733 | * Any task has to be enqueued before it get to execute on | |
2734 | * a CPU. So account for the time it spent waiting on the | |
2735 | * runqueue. | |
2736 | */ | |
2737 | update_stats_wait_end(cfs_rq, se); | |
2738 | __dequeue_entity(cfs_rq, se); | |
2739 | } | |
2740 | ||
79303e9e | 2741 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2742 | cfs_rq->curr = se; |
eba1ed4b IM |
2743 | #ifdef CONFIG_SCHEDSTATS |
2744 | /* | |
2745 | * Track our maximum slice length, if the CPU's load is at | |
2746 | * least twice that of our own weight (i.e. dont track it | |
2747 | * when there are only lesser-weight tasks around): | |
2748 | */ | |
495eca49 | 2749 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2750 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2751 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2752 | } | |
2753 | #endif | |
4a55b450 | 2754 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2755 | } |
2756 | ||
3f3a4904 PZ |
2757 | static int |
2758 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2759 | ||
ac53db59 RR |
2760 | /* |
2761 | * Pick the next process, keeping these things in mind, in this order: | |
2762 | * 1) keep things fair between processes/task groups | |
2763 | * 2) pick the "next" process, since someone really wants that to run | |
2764 | * 3) pick the "last" process, for cache locality | |
2765 | * 4) do not run the "skip" process, if something else is available | |
2766 | */ | |
f4b6755f | 2767 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 2768 | { |
ac53db59 | 2769 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 2770 | struct sched_entity *left = se; |
f4b6755f | 2771 | |
ac53db59 RR |
2772 | /* |
2773 | * Avoid running the skip buddy, if running something else can | |
2774 | * be done without getting too unfair. | |
2775 | */ | |
2776 | if (cfs_rq->skip == se) { | |
2777 | struct sched_entity *second = __pick_next_entity(se); | |
2778 | if (second && wakeup_preempt_entity(second, left) < 1) | |
2779 | se = second; | |
2780 | } | |
aa2ac252 | 2781 | |
f685ceac MG |
2782 | /* |
2783 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2784 | */ | |
2785 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2786 | se = cfs_rq->last; | |
2787 | ||
ac53db59 RR |
2788 | /* |
2789 | * Someone really wants this to run. If it's not unfair, run it. | |
2790 | */ | |
2791 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2792 | se = cfs_rq->next; | |
2793 | ||
f685ceac | 2794 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2795 | |
2796 | return se; | |
aa2ac252 PZ |
2797 | } |
2798 | ||
d3d9dc33 PT |
2799 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
2800 | ||
ab6cde26 | 2801 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2802 | { |
2803 | /* | |
2804 | * If still on the runqueue then deactivate_task() | |
2805 | * was not called and update_curr() has to be done: | |
2806 | */ | |
2807 | if (prev->on_rq) | |
b7cc0896 | 2808 | update_curr(cfs_rq); |
bf0f6f24 | 2809 | |
d3d9dc33 PT |
2810 | /* throttle cfs_rqs exceeding runtime */ |
2811 | check_cfs_rq_runtime(cfs_rq); | |
2812 | ||
ddc97297 | 2813 | check_spread(cfs_rq, prev); |
30cfdcfc | 2814 | if (prev->on_rq) { |
5870db5b | 2815 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2816 | /* Put 'current' back into the tree. */ |
2817 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2818 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2819 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2820 | } |
429d43bc | 2821 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2822 | } |
2823 | ||
8f4d37ec PZ |
2824 | static void |
2825 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2826 | { |
bf0f6f24 | 2827 | /* |
30cfdcfc | 2828 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2829 | */ |
30cfdcfc | 2830 | update_curr(cfs_rq); |
bf0f6f24 | 2831 | |
9d85f21c PT |
2832 | /* |
2833 | * Ensure that runnable average is periodically updated. | |
2834 | */ | |
9ee474f5 | 2835 | update_entity_load_avg(curr, 1); |
aff3e498 | 2836 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2837 | update_cfs_shares(cfs_rq); |
9d85f21c | 2838 | |
8f4d37ec PZ |
2839 | #ifdef CONFIG_SCHED_HRTICK |
2840 | /* | |
2841 | * queued ticks are scheduled to match the slice, so don't bother | |
2842 | * validating it and just reschedule. | |
2843 | */ | |
983ed7a6 HH |
2844 | if (queued) { |
2845 | resched_task(rq_of(cfs_rq)->curr); | |
2846 | return; | |
2847 | } | |
8f4d37ec PZ |
2848 | /* |
2849 | * don't let the period tick interfere with the hrtick preemption | |
2850 | */ | |
2851 | if (!sched_feat(DOUBLE_TICK) && | |
2852 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2853 | return; | |
2854 | #endif | |
2855 | ||
2c2efaed | 2856 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2857 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2858 | } |
2859 | ||
ab84d31e PT |
2860 | |
2861 | /************************************************** | |
2862 | * CFS bandwidth control machinery | |
2863 | */ | |
2864 | ||
2865 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2866 | |
2867 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2868 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2869 | |
2870 | static inline bool cfs_bandwidth_used(void) | |
2871 | { | |
c5905afb | 2872 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2873 | } |
2874 | ||
2875 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
2876 | { | |
2877 | /* only need to count groups transitioning between enabled/!enabled */ | |
2878 | if (enabled && !was_enabled) | |
c5905afb | 2879 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 2880 | else if (!enabled && was_enabled) |
c5905afb | 2881 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
2882 | } |
2883 | #else /* HAVE_JUMP_LABEL */ | |
2884 | static bool cfs_bandwidth_used(void) | |
2885 | { | |
2886 | return true; | |
2887 | } | |
2888 | ||
2889 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
2890 | #endif /* HAVE_JUMP_LABEL */ | |
2891 | ||
ab84d31e PT |
2892 | /* |
2893 | * default period for cfs group bandwidth. | |
2894 | * default: 0.1s, units: nanoseconds | |
2895 | */ | |
2896 | static inline u64 default_cfs_period(void) | |
2897 | { | |
2898 | return 100000000ULL; | |
2899 | } | |
ec12cb7f PT |
2900 | |
2901 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2902 | { | |
2903 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2904 | } | |
2905 | ||
a9cf55b2 PT |
2906 | /* |
2907 | * Replenish runtime according to assigned quota and update expiration time. | |
2908 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2909 | * additional synchronization around rq->lock. | |
2910 | * | |
2911 | * requires cfs_b->lock | |
2912 | */ | |
029632fb | 2913 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2914 | { |
2915 | u64 now; | |
2916 | ||
2917 | if (cfs_b->quota == RUNTIME_INF) | |
2918 | return; | |
2919 | ||
2920 | now = sched_clock_cpu(smp_processor_id()); | |
2921 | cfs_b->runtime = cfs_b->quota; | |
2922 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2923 | } | |
2924 | ||
029632fb PZ |
2925 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2926 | { | |
2927 | return &tg->cfs_bandwidth; | |
2928 | } | |
2929 | ||
f1b17280 PT |
2930 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2931 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2932 | { | |
2933 | if (unlikely(cfs_rq->throttle_count)) | |
2934 | return cfs_rq->throttled_clock_task; | |
2935 | ||
78becc27 | 2936 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2937 | } |
2938 | ||
85dac906 PT |
2939 | /* returns 0 on failure to allocate runtime */ |
2940 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2941 | { |
2942 | struct task_group *tg = cfs_rq->tg; | |
2943 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2944 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2945 | |
2946 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2947 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2948 | ||
2949 | raw_spin_lock(&cfs_b->lock); | |
2950 | if (cfs_b->quota == RUNTIME_INF) | |
2951 | amount = min_amount; | |
58088ad0 | 2952 | else { |
a9cf55b2 PT |
2953 | /* |
2954 | * If the bandwidth pool has become inactive, then at least one | |
2955 | * period must have elapsed since the last consumption. | |
2956 | * Refresh the global state and ensure bandwidth timer becomes | |
2957 | * active. | |
2958 | */ | |
2959 | if (!cfs_b->timer_active) { | |
2960 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2961 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2962 | } |
58088ad0 PT |
2963 | |
2964 | if (cfs_b->runtime > 0) { | |
2965 | amount = min(cfs_b->runtime, min_amount); | |
2966 | cfs_b->runtime -= amount; | |
2967 | cfs_b->idle = 0; | |
2968 | } | |
ec12cb7f | 2969 | } |
a9cf55b2 | 2970 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2971 | raw_spin_unlock(&cfs_b->lock); |
2972 | ||
2973 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2974 | /* |
2975 | * we may have advanced our local expiration to account for allowed | |
2976 | * spread between our sched_clock and the one on which runtime was | |
2977 | * issued. | |
2978 | */ | |
2979 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2980 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2981 | |
2982 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2983 | } |
2984 | ||
a9cf55b2 PT |
2985 | /* |
2986 | * Note: This depends on the synchronization provided by sched_clock and the | |
2987 | * fact that rq->clock snapshots this value. | |
2988 | */ | |
2989 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2990 | { |
a9cf55b2 | 2991 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2992 | |
2993 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2994 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2995 | return; |
2996 | ||
a9cf55b2 PT |
2997 | if (cfs_rq->runtime_remaining < 0) |
2998 | return; | |
2999 | ||
3000 | /* | |
3001 | * If the local deadline has passed we have to consider the | |
3002 | * possibility that our sched_clock is 'fast' and the global deadline | |
3003 | * has not truly expired. | |
3004 | * | |
3005 | * Fortunately we can check determine whether this the case by checking | |
3006 | * whether the global deadline has advanced. | |
3007 | */ | |
3008 | ||
3009 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
3010 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
3011 | cfs_rq->runtime_expires += TICK_NSEC; | |
3012 | } else { | |
3013 | /* global deadline is ahead, expiration has passed */ | |
3014 | cfs_rq->runtime_remaining = 0; | |
3015 | } | |
3016 | } | |
3017 | ||
3018 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
3019 | unsigned long delta_exec) | |
3020 | { | |
3021 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3022 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3023 | expire_cfs_rq_runtime(cfs_rq); |
3024 | ||
3025 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3026 | return; |
3027 | ||
85dac906 PT |
3028 | /* |
3029 | * if we're unable to extend our runtime we resched so that the active | |
3030 | * hierarchy can be throttled | |
3031 | */ | |
3032 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
3033 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
3034 | } |
3035 | ||
6c16a6dc PZ |
3036 | static __always_inline |
3037 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 3038 | { |
56f570e5 | 3039 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3040 | return; |
3041 | ||
3042 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3043 | } | |
3044 | ||
85dac906 PT |
3045 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3046 | { | |
56f570e5 | 3047 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3048 | } |
3049 | ||
64660c86 PT |
3050 | /* check whether cfs_rq, or any parent, is throttled */ |
3051 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3052 | { | |
56f570e5 | 3053 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3054 | } |
3055 | ||
3056 | /* | |
3057 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3058 | * dest_cpu are members of a throttled hierarchy when performing group | |
3059 | * load-balance operations. | |
3060 | */ | |
3061 | static inline int throttled_lb_pair(struct task_group *tg, | |
3062 | int src_cpu, int dest_cpu) | |
3063 | { | |
3064 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3065 | ||
3066 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3067 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3068 | ||
3069 | return throttled_hierarchy(src_cfs_rq) || | |
3070 | throttled_hierarchy(dest_cfs_rq); | |
3071 | } | |
3072 | ||
3073 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3074 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3075 | { | |
3076 | struct rq *rq = data; | |
3077 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3078 | ||
3079 | cfs_rq->throttle_count--; | |
3080 | #ifdef CONFIG_SMP | |
3081 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 3082 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3083 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3084 | cfs_rq->throttled_clock_task; |
64660c86 PT |
3085 | } |
3086 | #endif | |
3087 | ||
3088 | return 0; | |
3089 | } | |
3090 | ||
3091 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3092 | { | |
3093 | struct rq *rq = data; | |
3094 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3095 | ||
82958366 PT |
3096 | /* group is entering throttled state, stop time */ |
3097 | if (!cfs_rq->throttle_count) | |
78becc27 | 3098 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3099 | cfs_rq->throttle_count++; |
3100 | ||
3101 | return 0; | |
3102 | } | |
3103 | ||
d3d9dc33 | 3104 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3105 | { |
3106 | struct rq *rq = rq_of(cfs_rq); | |
3107 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3108 | struct sched_entity *se; | |
3109 | long task_delta, dequeue = 1; | |
3110 | ||
3111 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3112 | ||
f1b17280 | 3113 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3114 | rcu_read_lock(); |
3115 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3116 | rcu_read_unlock(); | |
85dac906 PT |
3117 | |
3118 | task_delta = cfs_rq->h_nr_running; | |
3119 | for_each_sched_entity(se) { | |
3120 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3121 | /* throttled entity or throttle-on-deactivate */ | |
3122 | if (!se->on_rq) | |
3123 | break; | |
3124 | ||
3125 | if (dequeue) | |
3126 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3127 | qcfs_rq->h_nr_running -= task_delta; | |
3128 | ||
3129 | if (qcfs_rq->load.weight) | |
3130 | dequeue = 0; | |
3131 | } | |
3132 | ||
3133 | if (!se) | |
3134 | rq->nr_running -= task_delta; | |
3135 | ||
3136 | cfs_rq->throttled = 1; | |
78becc27 | 3137 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
3138 | raw_spin_lock(&cfs_b->lock); |
3139 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
3140 | raw_spin_unlock(&cfs_b->lock); | |
3141 | } | |
3142 | ||
029632fb | 3143 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3144 | { |
3145 | struct rq *rq = rq_of(cfs_rq); | |
3146 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3147 | struct sched_entity *se; | |
3148 | int enqueue = 1; | |
3149 | long task_delta; | |
3150 | ||
22b958d8 | 3151 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3152 | |
3153 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3154 | |
3155 | update_rq_clock(rq); | |
3156 | ||
671fd9da | 3157 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3158 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3159 | list_del_rcu(&cfs_rq->throttled_list); |
3160 | raw_spin_unlock(&cfs_b->lock); | |
3161 | ||
64660c86 PT |
3162 | /* update hierarchical throttle state */ |
3163 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3164 | ||
671fd9da PT |
3165 | if (!cfs_rq->load.weight) |
3166 | return; | |
3167 | ||
3168 | task_delta = cfs_rq->h_nr_running; | |
3169 | for_each_sched_entity(se) { | |
3170 | if (se->on_rq) | |
3171 | enqueue = 0; | |
3172 | ||
3173 | cfs_rq = cfs_rq_of(se); | |
3174 | if (enqueue) | |
3175 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3176 | cfs_rq->h_nr_running += task_delta; | |
3177 | ||
3178 | if (cfs_rq_throttled(cfs_rq)) | |
3179 | break; | |
3180 | } | |
3181 | ||
3182 | if (!se) | |
3183 | rq->nr_running += task_delta; | |
3184 | ||
3185 | /* determine whether we need to wake up potentially idle cpu */ | |
3186 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
3187 | resched_task(rq->curr); | |
3188 | } | |
3189 | ||
3190 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
3191 | u64 remaining, u64 expires) | |
3192 | { | |
3193 | struct cfs_rq *cfs_rq; | |
3194 | u64 runtime = remaining; | |
3195 | ||
3196 | rcu_read_lock(); | |
3197 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
3198 | throttled_list) { | |
3199 | struct rq *rq = rq_of(cfs_rq); | |
3200 | ||
3201 | raw_spin_lock(&rq->lock); | |
3202 | if (!cfs_rq_throttled(cfs_rq)) | |
3203 | goto next; | |
3204 | ||
3205 | runtime = -cfs_rq->runtime_remaining + 1; | |
3206 | if (runtime > remaining) | |
3207 | runtime = remaining; | |
3208 | remaining -= runtime; | |
3209 | ||
3210 | cfs_rq->runtime_remaining += runtime; | |
3211 | cfs_rq->runtime_expires = expires; | |
3212 | ||
3213 | /* we check whether we're throttled above */ | |
3214 | if (cfs_rq->runtime_remaining > 0) | |
3215 | unthrottle_cfs_rq(cfs_rq); | |
3216 | ||
3217 | next: | |
3218 | raw_spin_unlock(&rq->lock); | |
3219 | ||
3220 | if (!remaining) | |
3221 | break; | |
3222 | } | |
3223 | rcu_read_unlock(); | |
3224 | ||
3225 | return remaining; | |
3226 | } | |
3227 | ||
58088ad0 PT |
3228 | /* |
3229 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
3230 | * cfs_rqs as appropriate. If there has been no activity within the last | |
3231 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
3232 | * used to track this state. | |
3233 | */ | |
3234 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3235 | { | |
671fd9da PT |
3236 | u64 runtime, runtime_expires; |
3237 | int idle = 1, throttled; | |
58088ad0 PT |
3238 | |
3239 | raw_spin_lock(&cfs_b->lock); | |
3240 | /* no need to continue the timer with no bandwidth constraint */ | |
3241 | if (cfs_b->quota == RUNTIME_INF) | |
3242 | goto out_unlock; | |
3243 | ||
671fd9da PT |
3244 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3245 | /* idle depends on !throttled (for the case of a large deficit) */ | |
3246 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 3247 | cfs_b->nr_periods += overrun; |
671fd9da | 3248 | |
a9cf55b2 PT |
3249 | /* if we're going inactive then everything else can be deferred */ |
3250 | if (idle) | |
3251 | goto out_unlock; | |
3252 | ||
3253 | __refill_cfs_bandwidth_runtime(cfs_b); | |
3254 | ||
671fd9da PT |
3255 | if (!throttled) { |
3256 | /* mark as potentially idle for the upcoming period */ | |
3257 | cfs_b->idle = 1; | |
3258 | goto out_unlock; | |
3259 | } | |
3260 | ||
e8da1b18 NR |
3261 | /* account preceding periods in which throttling occurred */ |
3262 | cfs_b->nr_throttled += overrun; | |
3263 | ||
671fd9da PT |
3264 | /* |
3265 | * There are throttled entities so we must first use the new bandwidth | |
3266 | * to unthrottle them before making it generally available. This | |
3267 | * ensures that all existing debts will be paid before a new cfs_rq is | |
3268 | * allowed to run. | |
3269 | */ | |
3270 | runtime = cfs_b->runtime; | |
3271 | runtime_expires = cfs_b->runtime_expires; | |
3272 | cfs_b->runtime = 0; | |
3273 | ||
3274 | /* | |
3275 | * This check is repeated as we are holding onto the new bandwidth | |
3276 | * while we unthrottle. This can potentially race with an unthrottled | |
3277 | * group trying to acquire new bandwidth from the global pool. | |
3278 | */ | |
3279 | while (throttled && runtime > 0) { | |
3280 | raw_spin_unlock(&cfs_b->lock); | |
3281 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3282 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3283 | runtime_expires); | |
3284 | raw_spin_lock(&cfs_b->lock); | |
3285 | ||
3286 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
3287 | } | |
58088ad0 | 3288 | |
671fd9da PT |
3289 | /* return (any) remaining runtime */ |
3290 | cfs_b->runtime = runtime; | |
3291 | /* | |
3292 | * While we are ensured activity in the period following an | |
3293 | * unthrottle, this also covers the case in which the new bandwidth is | |
3294 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3295 | * timer to remain active while there are any throttled entities.) | |
3296 | */ | |
3297 | cfs_b->idle = 0; | |
58088ad0 PT |
3298 | out_unlock: |
3299 | if (idle) | |
3300 | cfs_b->timer_active = 0; | |
3301 | raw_spin_unlock(&cfs_b->lock); | |
3302 | ||
3303 | return idle; | |
3304 | } | |
d3d9dc33 | 3305 | |
d8b4986d PT |
3306 | /* a cfs_rq won't donate quota below this amount */ |
3307 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3308 | /* minimum remaining period time to redistribute slack quota */ | |
3309 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3310 | /* how long we wait to gather additional slack before distributing */ | |
3311 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3312 | ||
3313 | /* are we near the end of the current quota period? */ | |
3314 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
3315 | { | |
3316 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3317 | u64 remaining; | |
3318 | ||
3319 | /* if the call-back is running a quota refresh is already occurring */ | |
3320 | if (hrtimer_callback_running(refresh_timer)) | |
3321 | return 1; | |
3322 | ||
3323 | /* is a quota refresh about to occur? */ | |
3324 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3325 | if (remaining < min_expire) | |
3326 | return 1; | |
3327 | ||
3328 | return 0; | |
3329 | } | |
3330 | ||
3331 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3332 | { | |
3333 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3334 | ||
3335 | /* if there's a quota refresh soon don't bother with slack */ | |
3336 | if (runtime_refresh_within(cfs_b, min_left)) | |
3337 | return; | |
3338 | ||
3339 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3340 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3341 | } | |
3342 | ||
3343 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3344 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3345 | { | |
3346 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3347 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3348 | ||
3349 | if (slack_runtime <= 0) | |
3350 | return; | |
3351 | ||
3352 | raw_spin_lock(&cfs_b->lock); | |
3353 | if (cfs_b->quota != RUNTIME_INF && | |
3354 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3355 | cfs_b->runtime += slack_runtime; | |
3356 | ||
3357 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3358 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3359 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3360 | start_cfs_slack_bandwidth(cfs_b); | |
3361 | } | |
3362 | raw_spin_unlock(&cfs_b->lock); | |
3363 | ||
3364 | /* even if it's not valid for return we don't want to try again */ | |
3365 | cfs_rq->runtime_remaining -= slack_runtime; | |
3366 | } | |
3367 | ||
3368 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3369 | { | |
56f570e5 PT |
3370 | if (!cfs_bandwidth_used()) |
3371 | return; | |
3372 | ||
fccfdc6f | 3373 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3374 | return; |
3375 | ||
3376 | __return_cfs_rq_runtime(cfs_rq); | |
3377 | } | |
3378 | ||
3379 | /* | |
3380 | * This is done with a timer (instead of inline with bandwidth return) since | |
3381 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3382 | */ | |
3383 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3384 | { | |
3385 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3386 | u64 expires; | |
3387 | ||
3388 | /* confirm we're still not at a refresh boundary */ | |
3389 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
3390 | return; | |
3391 | ||
3392 | raw_spin_lock(&cfs_b->lock); | |
3393 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
3394 | runtime = cfs_b->runtime; | |
3395 | cfs_b->runtime = 0; | |
3396 | } | |
3397 | expires = cfs_b->runtime_expires; | |
3398 | raw_spin_unlock(&cfs_b->lock); | |
3399 | ||
3400 | if (!runtime) | |
3401 | return; | |
3402 | ||
3403 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3404 | ||
3405 | raw_spin_lock(&cfs_b->lock); | |
3406 | if (expires == cfs_b->runtime_expires) | |
3407 | cfs_b->runtime = runtime; | |
3408 | raw_spin_unlock(&cfs_b->lock); | |
3409 | } | |
3410 | ||
d3d9dc33 PT |
3411 | /* |
3412 | * When a group wakes up we want to make sure that its quota is not already | |
3413 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3414 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3415 | */ | |
3416 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3417 | { | |
56f570e5 PT |
3418 | if (!cfs_bandwidth_used()) |
3419 | return; | |
3420 | ||
d3d9dc33 PT |
3421 | /* an active group must be handled by the update_curr()->put() path */ |
3422 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3423 | return; | |
3424 | ||
3425 | /* ensure the group is not already throttled */ | |
3426 | if (cfs_rq_throttled(cfs_rq)) | |
3427 | return; | |
3428 | ||
3429 | /* update runtime allocation */ | |
3430 | account_cfs_rq_runtime(cfs_rq, 0); | |
3431 | if (cfs_rq->runtime_remaining <= 0) | |
3432 | throttle_cfs_rq(cfs_rq); | |
3433 | } | |
3434 | ||
3435 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
3436 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3437 | { | |
56f570e5 PT |
3438 | if (!cfs_bandwidth_used()) |
3439 | return; | |
3440 | ||
d3d9dc33 PT |
3441 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
3442 | return; | |
3443 | ||
3444 | /* | |
3445 | * it's possible for a throttled entity to be forced into a running | |
3446 | * state (e.g. set_curr_task), in this case we're finished. | |
3447 | */ | |
3448 | if (cfs_rq_throttled(cfs_rq)) | |
3449 | return; | |
3450 | ||
3451 | throttle_cfs_rq(cfs_rq); | |
3452 | } | |
029632fb | 3453 | |
029632fb PZ |
3454 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3455 | { | |
3456 | struct cfs_bandwidth *cfs_b = | |
3457 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3458 | do_sched_cfs_slack_timer(cfs_b); | |
3459 | ||
3460 | return HRTIMER_NORESTART; | |
3461 | } | |
3462 | ||
3463 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3464 | { | |
3465 | struct cfs_bandwidth *cfs_b = | |
3466 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3467 | ktime_t now; | |
3468 | int overrun; | |
3469 | int idle = 0; | |
3470 | ||
3471 | for (;;) { | |
3472 | now = hrtimer_cb_get_time(timer); | |
3473 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3474 | ||
3475 | if (!overrun) | |
3476 | break; | |
3477 | ||
3478 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3479 | } | |
3480 | ||
3481 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3482 | } | |
3483 | ||
3484 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3485 | { | |
3486 | raw_spin_lock_init(&cfs_b->lock); | |
3487 | cfs_b->runtime = 0; | |
3488 | cfs_b->quota = RUNTIME_INF; | |
3489 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3490 | ||
3491 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3492 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3493 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3494 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3495 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3496 | } | |
3497 | ||
3498 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3499 | { | |
3500 | cfs_rq->runtime_enabled = 0; | |
3501 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3502 | } | |
3503 | ||
3504 | /* requires cfs_b->lock, may release to reprogram timer */ | |
3505 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3506 | { | |
3507 | /* | |
3508 | * The timer may be active because we're trying to set a new bandwidth | |
3509 | * period or because we're racing with the tear-down path | |
3510 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3511 | * terminates). In either case we ensure that it's re-programmed | |
3512 | */ | |
3513 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
3514 | raw_spin_unlock(&cfs_b->lock); | |
3515 | /* ensure cfs_b->lock is available while we wait */ | |
3516 | hrtimer_cancel(&cfs_b->period_timer); | |
3517 | ||
3518 | raw_spin_lock(&cfs_b->lock); | |
3519 | /* if someone else restarted the timer then we're done */ | |
3520 | if (cfs_b->timer_active) | |
3521 | return; | |
3522 | } | |
3523 | ||
3524 | cfs_b->timer_active = 1; | |
3525 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3526 | } | |
3527 | ||
3528 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3529 | { | |
3530 | hrtimer_cancel(&cfs_b->period_timer); | |
3531 | hrtimer_cancel(&cfs_b->slack_timer); | |
3532 | } | |
3533 | ||
38dc3348 | 3534 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3535 | { |
3536 | struct cfs_rq *cfs_rq; | |
3537 | ||
3538 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3539 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3540 | ||
3541 | if (!cfs_rq->runtime_enabled) | |
3542 | continue; | |
3543 | ||
3544 | /* | |
3545 | * clock_task is not advancing so we just need to make sure | |
3546 | * there's some valid quota amount | |
3547 | */ | |
3548 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3549 | if (cfs_rq_throttled(cfs_rq)) | |
3550 | unthrottle_cfs_rq(cfs_rq); | |
3551 | } | |
3552 | } | |
3553 | ||
3554 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3555 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3556 | { | |
78becc27 | 3557 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3558 | } |
3559 | ||
3560 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
3561 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
3562 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
3563 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 3564 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3565 | |
3566 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3567 | { | |
3568 | return 0; | |
3569 | } | |
64660c86 PT |
3570 | |
3571 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3572 | { | |
3573 | return 0; | |
3574 | } | |
3575 | ||
3576 | static inline int throttled_lb_pair(struct task_group *tg, | |
3577 | int src_cpu, int dest_cpu) | |
3578 | { | |
3579 | return 0; | |
3580 | } | |
029632fb PZ |
3581 | |
3582 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3583 | ||
3584 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3585 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3586 | #endif |
3587 | ||
029632fb PZ |
3588 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3589 | { | |
3590 | return NULL; | |
3591 | } | |
3592 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3593 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3594 | |
3595 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3596 | ||
bf0f6f24 IM |
3597 | /************************************************** |
3598 | * CFS operations on tasks: | |
3599 | */ | |
3600 | ||
8f4d37ec PZ |
3601 | #ifdef CONFIG_SCHED_HRTICK |
3602 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3603 | { | |
8f4d37ec PZ |
3604 | struct sched_entity *se = &p->se; |
3605 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3606 | ||
3607 | WARN_ON(task_rq(p) != rq); | |
3608 | ||
b39e66ea | 3609 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3610 | u64 slice = sched_slice(cfs_rq, se); |
3611 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3612 | s64 delta = slice - ran; | |
3613 | ||
3614 | if (delta < 0) { | |
3615 | if (rq->curr == p) | |
3616 | resched_task(p); | |
3617 | return; | |
3618 | } | |
3619 | ||
3620 | /* | |
3621 | * Don't schedule slices shorter than 10000ns, that just | |
3622 | * doesn't make sense. Rely on vruntime for fairness. | |
3623 | */ | |
31656519 | 3624 | if (rq->curr != p) |
157124c1 | 3625 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3626 | |
31656519 | 3627 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3628 | } |
3629 | } | |
a4c2f00f PZ |
3630 | |
3631 | /* | |
3632 | * called from enqueue/dequeue and updates the hrtick when the | |
3633 | * current task is from our class and nr_running is low enough | |
3634 | * to matter. | |
3635 | */ | |
3636 | static void hrtick_update(struct rq *rq) | |
3637 | { | |
3638 | struct task_struct *curr = rq->curr; | |
3639 | ||
b39e66ea | 3640 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3641 | return; |
3642 | ||
3643 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3644 | hrtick_start_fair(rq, curr); | |
3645 | } | |
55e12e5e | 3646 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3647 | static inline void |
3648 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3649 | { | |
3650 | } | |
a4c2f00f PZ |
3651 | |
3652 | static inline void hrtick_update(struct rq *rq) | |
3653 | { | |
3654 | } | |
8f4d37ec PZ |
3655 | #endif |
3656 | ||
bf0f6f24 IM |
3657 | /* |
3658 | * The enqueue_task method is called before nr_running is | |
3659 | * increased. Here we update the fair scheduling stats and | |
3660 | * then put the task into the rbtree: | |
3661 | */ | |
ea87bb78 | 3662 | static void |
371fd7e7 | 3663 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3664 | { |
3665 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3666 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3667 | |
3668 | for_each_sched_entity(se) { | |
62fb1851 | 3669 | if (se->on_rq) |
bf0f6f24 IM |
3670 | break; |
3671 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3672 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3673 | |
3674 | /* | |
3675 | * end evaluation on encountering a throttled cfs_rq | |
3676 | * | |
3677 | * note: in the case of encountering a throttled cfs_rq we will | |
3678 | * post the final h_nr_running increment below. | |
3679 | */ | |
3680 | if (cfs_rq_throttled(cfs_rq)) | |
3681 | break; | |
953bfcd1 | 3682 | cfs_rq->h_nr_running++; |
85dac906 | 3683 | |
88ec22d3 | 3684 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3685 | } |
8f4d37ec | 3686 | |
2069dd75 | 3687 | for_each_sched_entity(se) { |
0f317143 | 3688 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3689 | cfs_rq->h_nr_running++; |
2069dd75 | 3690 | |
85dac906 PT |
3691 | if (cfs_rq_throttled(cfs_rq)) |
3692 | break; | |
3693 | ||
17bc14b7 | 3694 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3695 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3696 | } |
3697 | ||
18bf2805 BS |
3698 | if (!se) { |
3699 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3700 | inc_nr_running(rq); |
18bf2805 | 3701 | } |
a4c2f00f | 3702 | hrtick_update(rq); |
bf0f6f24 IM |
3703 | } |
3704 | ||
2f36825b VP |
3705 | static void set_next_buddy(struct sched_entity *se); |
3706 | ||
bf0f6f24 IM |
3707 | /* |
3708 | * The dequeue_task method is called before nr_running is | |
3709 | * decreased. We remove the task from the rbtree and | |
3710 | * update the fair scheduling stats: | |
3711 | */ | |
371fd7e7 | 3712 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3713 | { |
3714 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3715 | struct sched_entity *se = &p->se; |
2f36825b | 3716 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3717 | |
3718 | for_each_sched_entity(se) { | |
3719 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3720 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3721 | |
3722 | /* | |
3723 | * end evaluation on encountering a throttled cfs_rq | |
3724 | * | |
3725 | * note: in the case of encountering a throttled cfs_rq we will | |
3726 | * post the final h_nr_running decrement below. | |
3727 | */ | |
3728 | if (cfs_rq_throttled(cfs_rq)) | |
3729 | break; | |
953bfcd1 | 3730 | cfs_rq->h_nr_running--; |
2069dd75 | 3731 | |
bf0f6f24 | 3732 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3733 | if (cfs_rq->load.weight) { |
3734 | /* | |
3735 | * Bias pick_next to pick a task from this cfs_rq, as | |
3736 | * p is sleeping when it is within its sched_slice. | |
3737 | */ | |
3738 | if (task_sleep && parent_entity(se)) | |
3739 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3740 | |
3741 | /* avoid re-evaluating load for this entity */ | |
3742 | se = parent_entity(se); | |
bf0f6f24 | 3743 | break; |
2f36825b | 3744 | } |
371fd7e7 | 3745 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3746 | } |
8f4d37ec | 3747 | |
2069dd75 | 3748 | for_each_sched_entity(se) { |
0f317143 | 3749 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3750 | cfs_rq->h_nr_running--; |
2069dd75 | 3751 | |
85dac906 PT |
3752 | if (cfs_rq_throttled(cfs_rq)) |
3753 | break; | |
3754 | ||
17bc14b7 | 3755 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3756 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3757 | } |
3758 | ||
18bf2805 | 3759 | if (!se) { |
85dac906 | 3760 | dec_nr_running(rq); |
18bf2805 BS |
3761 | update_rq_runnable_avg(rq, 1); |
3762 | } | |
a4c2f00f | 3763 | hrtick_update(rq); |
bf0f6f24 IM |
3764 | } |
3765 | ||
e7693a36 | 3766 | #ifdef CONFIG_SMP |
029632fb PZ |
3767 | /* Used instead of source_load when we know the type == 0 */ |
3768 | static unsigned long weighted_cpuload(const int cpu) | |
3769 | { | |
b92486cb | 3770 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3771 | } |
3772 | ||
3773 | /* | |
3774 | * Return a low guess at the load of a migration-source cpu weighted | |
3775 | * according to the scheduling class and "nice" value. | |
3776 | * | |
3777 | * We want to under-estimate the load of migration sources, to | |
3778 | * balance conservatively. | |
3779 | */ | |
3780 | static unsigned long source_load(int cpu, int type) | |
3781 | { | |
3782 | struct rq *rq = cpu_rq(cpu); | |
3783 | unsigned long total = weighted_cpuload(cpu); | |
3784 | ||
3785 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3786 | return total; | |
3787 | ||
3788 | return min(rq->cpu_load[type-1], total); | |
3789 | } | |
3790 | ||
3791 | /* | |
3792 | * Return a high guess at the load of a migration-target cpu weighted | |
3793 | * according to the scheduling class and "nice" value. | |
3794 | */ | |
3795 | static unsigned long target_load(int cpu, int type) | |
3796 | { | |
3797 | struct rq *rq = cpu_rq(cpu); | |
3798 | unsigned long total = weighted_cpuload(cpu); | |
3799 | ||
3800 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3801 | return total; | |
3802 | ||
3803 | return max(rq->cpu_load[type-1], total); | |
3804 | } | |
3805 | ||
3806 | static unsigned long power_of(int cpu) | |
3807 | { | |
3808 | return cpu_rq(cpu)->cpu_power; | |
3809 | } | |
3810 | ||
3811 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3812 | { | |
3813 | struct rq *rq = cpu_rq(cpu); | |
3814 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3815 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3816 | |
3817 | if (nr_running) | |
b92486cb | 3818 | return load_avg / nr_running; |
029632fb PZ |
3819 | |
3820 | return 0; | |
3821 | } | |
3822 | ||
62470419 MW |
3823 | static void record_wakee(struct task_struct *p) |
3824 | { | |
3825 | /* | |
3826 | * Rough decay (wiping) for cost saving, don't worry | |
3827 | * about the boundary, really active task won't care | |
3828 | * about the loss. | |
3829 | */ | |
3830 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
3831 | current->wakee_flips = 0; | |
3832 | current->wakee_flip_decay_ts = jiffies; | |
3833 | } | |
3834 | ||
3835 | if (current->last_wakee != p) { | |
3836 | current->last_wakee = p; | |
3837 | current->wakee_flips++; | |
3838 | } | |
3839 | } | |
098fb9db | 3840 | |
74f8e4b2 | 3841 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
3842 | { |
3843 | struct sched_entity *se = &p->se; | |
3844 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
3845 | u64 min_vruntime; |
3846 | ||
3847 | #ifndef CONFIG_64BIT | |
3848 | u64 min_vruntime_copy; | |
88ec22d3 | 3849 | |
3fe1698b PZ |
3850 | do { |
3851 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3852 | smp_rmb(); | |
3853 | min_vruntime = cfs_rq->min_vruntime; | |
3854 | } while (min_vruntime != min_vruntime_copy); | |
3855 | #else | |
3856 | min_vruntime = cfs_rq->min_vruntime; | |
3857 | #endif | |
88ec22d3 | 3858 | |
3fe1698b | 3859 | se->vruntime -= min_vruntime; |
62470419 | 3860 | record_wakee(p); |
88ec22d3 PZ |
3861 | } |
3862 | ||
bb3469ac | 3863 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3864 | /* |
3865 | * effective_load() calculates the load change as seen from the root_task_group | |
3866 | * | |
3867 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3868 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3869 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3870 | * |
3871 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3872 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3873 | * total group weight. | |
3874 | * | |
3875 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3876 | * distribution (s_i) using: | |
3877 | * | |
3878 | * s_i = rw_i / \Sum rw_j (1) | |
3879 | * | |
3880 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3881 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3882 | * shares distribution (s_i): | |
3883 | * | |
3884 | * rw_i = { 2, 4, 1, 0 } | |
3885 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3886 | * | |
3887 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3888 | * task used to run on and the CPU the waker is running on), we need to | |
3889 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3890 | * wakeup, compute the effect of the current task going to sleep. | |
3891 | * | |
3892 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3893 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3894 | * | |
3895 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3896 | * | |
3897 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3898 | * differences in waking a task to CPU 0. The additional task changes the | |
3899 | * weight and shares distributions like: | |
3900 | * | |
3901 | * rw'_i = { 3, 4, 1, 0 } | |
3902 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3903 | * | |
3904 | * We can then compute the difference in effective weight by using: | |
3905 | * | |
3906 | * dw_i = S * (s'_i - s_i) (3) | |
3907 | * | |
3908 | * Where 'S' is the group weight as seen by its parent. | |
3909 | * | |
3910 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3911 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3912 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3913 | */ |
2069dd75 | 3914 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3915 | { |
4be9daaa | 3916 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3917 | |
58d081b5 | 3918 | if (!tg->parent || !wl) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3919 | return wl; |
3920 | ||
4be9daaa | 3921 | for_each_sched_entity(se) { |
cf5f0acf | 3922 | long w, W; |
4be9daaa | 3923 | |
977dda7c | 3924 | tg = se->my_q->tg; |
bb3469ac | 3925 | |
cf5f0acf PZ |
3926 | /* |
3927 | * W = @wg + \Sum rw_j | |
3928 | */ | |
3929 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3930 | |
cf5f0acf PZ |
3931 | /* |
3932 | * w = rw_i + @wl | |
3933 | */ | |
3934 | w = se->my_q->load.weight + wl; | |
940959e9 | 3935 | |
cf5f0acf PZ |
3936 | /* |
3937 | * wl = S * s'_i; see (2) | |
3938 | */ | |
3939 | if (W > 0 && w < W) | |
3940 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3941 | else |
3942 | wl = tg->shares; | |
940959e9 | 3943 | |
cf5f0acf PZ |
3944 | /* |
3945 | * Per the above, wl is the new se->load.weight value; since | |
3946 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3947 | * calc_cfs_shares(). | |
3948 | */ | |
977dda7c PT |
3949 | if (wl < MIN_SHARES) |
3950 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3951 | |
3952 | /* | |
3953 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3954 | */ | |
977dda7c | 3955 | wl -= se->load.weight; |
cf5f0acf PZ |
3956 | |
3957 | /* | |
3958 | * Recursively apply this logic to all parent groups to compute | |
3959 | * the final effective load change on the root group. Since | |
3960 | * only the @tg group gets extra weight, all parent groups can | |
3961 | * only redistribute existing shares. @wl is the shift in shares | |
3962 | * resulting from this level per the above. | |
3963 | */ | |
4be9daaa | 3964 | wg = 0; |
4be9daaa | 3965 | } |
bb3469ac | 3966 | |
4be9daaa | 3967 | return wl; |
bb3469ac PZ |
3968 | } |
3969 | #else | |
4be9daaa | 3970 | |
58d081b5 | 3971 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 3972 | { |
83378269 | 3973 | return wl; |
bb3469ac | 3974 | } |
4be9daaa | 3975 | |
bb3469ac PZ |
3976 | #endif |
3977 | ||
62470419 MW |
3978 | static int wake_wide(struct task_struct *p) |
3979 | { | |
7d9ffa89 | 3980 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
3981 | |
3982 | /* | |
3983 | * Yeah, it's the switching-frequency, could means many wakee or | |
3984 | * rapidly switch, use factor here will just help to automatically | |
3985 | * adjust the loose-degree, so bigger node will lead to more pull. | |
3986 | */ | |
3987 | if (p->wakee_flips > factor) { | |
3988 | /* | |
3989 | * wakee is somewhat hot, it needs certain amount of cpu | |
3990 | * resource, so if waker is far more hot, prefer to leave | |
3991 | * it alone. | |
3992 | */ | |
3993 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
3994 | return 1; | |
3995 | } | |
3996 | ||
3997 | return 0; | |
3998 | } | |
3999 | ||
c88d5910 | 4000 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 4001 | { |
e37b6a7b | 4002 | s64 this_load, load; |
c88d5910 | 4003 | int idx, this_cpu, prev_cpu; |
098fb9db | 4004 | unsigned long tl_per_task; |
c88d5910 | 4005 | struct task_group *tg; |
83378269 | 4006 | unsigned long weight; |
b3137bc8 | 4007 | int balanced; |
098fb9db | 4008 | |
62470419 MW |
4009 | /* |
4010 | * If we wake multiple tasks be careful to not bounce | |
4011 | * ourselves around too much. | |
4012 | */ | |
4013 | if (wake_wide(p)) | |
4014 | return 0; | |
4015 | ||
c88d5910 PZ |
4016 | idx = sd->wake_idx; |
4017 | this_cpu = smp_processor_id(); | |
4018 | prev_cpu = task_cpu(p); | |
4019 | load = source_load(prev_cpu, idx); | |
4020 | this_load = target_load(this_cpu, idx); | |
098fb9db | 4021 | |
b3137bc8 MG |
4022 | /* |
4023 | * If sync wakeup then subtract the (maximum possible) | |
4024 | * effect of the currently running task from the load | |
4025 | * of the current CPU: | |
4026 | */ | |
83378269 PZ |
4027 | if (sync) { |
4028 | tg = task_group(current); | |
4029 | weight = current->se.load.weight; | |
4030 | ||
c88d5910 | 4031 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
4032 | load += effective_load(tg, prev_cpu, 0, -weight); |
4033 | } | |
b3137bc8 | 4034 | |
83378269 PZ |
4035 | tg = task_group(p); |
4036 | weight = p->se.load.weight; | |
b3137bc8 | 4037 | |
71a29aa7 PZ |
4038 | /* |
4039 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
4040 | * due to the sync cause above having dropped this_load to 0, we'll |
4041 | * always have an imbalance, but there's really nothing you can do | |
4042 | * about that, so that's good too. | |
71a29aa7 PZ |
4043 | * |
4044 | * Otherwise check if either cpus are near enough in load to allow this | |
4045 | * task to be woken on this_cpu. | |
4046 | */ | |
e37b6a7b PT |
4047 | if (this_load > 0) { |
4048 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
4049 | |
4050 | this_eff_load = 100; | |
4051 | this_eff_load *= power_of(prev_cpu); | |
4052 | this_eff_load *= this_load + | |
4053 | effective_load(tg, this_cpu, weight, weight); | |
4054 | ||
4055 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
4056 | prev_eff_load *= power_of(this_cpu); | |
4057 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
4058 | ||
4059 | balanced = this_eff_load <= prev_eff_load; | |
4060 | } else | |
4061 | balanced = true; | |
b3137bc8 | 4062 | |
098fb9db | 4063 | /* |
4ae7d5ce IM |
4064 | * If the currently running task will sleep within |
4065 | * a reasonable amount of time then attract this newly | |
4066 | * woken task: | |
098fb9db | 4067 | */ |
2fb7635c PZ |
4068 | if (sync && balanced) |
4069 | return 1; | |
098fb9db | 4070 | |
41acab88 | 4071 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
4072 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
4073 | ||
c88d5910 PZ |
4074 | if (balanced || |
4075 | (this_load <= load && | |
4076 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
4077 | /* |
4078 | * This domain has SD_WAKE_AFFINE and | |
4079 | * p is cache cold in this domain, and | |
4080 | * there is no bad imbalance. | |
4081 | */ | |
c88d5910 | 4082 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 4083 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
4084 | |
4085 | return 1; | |
4086 | } | |
4087 | return 0; | |
4088 | } | |
4089 | ||
aaee1203 PZ |
4090 | /* |
4091 | * find_idlest_group finds and returns the least busy CPU group within the | |
4092 | * domain. | |
4093 | */ | |
4094 | static struct sched_group * | |
78e7ed53 | 4095 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 4096 | int this_cpu, int load_idx) |
e7693a36 | 4097 | { |
b3bd3de6 | 4098 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 4099 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 4100 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 4101 | |
aaee1203 PZ |
4102 | do { |
4103 | unsigned long load, avg_load; | |
4104 | int local_group; | |
4105 | int i; | |
e7693a36 | 4106 | |
aaee1203 PZ |
4107 | /* Skip over this group if it has no CPUs allowed */ |
4108 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 4109 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
4110 | continue; |
4111 | ||
4112 | local_group = cpumask_test_cpu(this_cpu, | |
4113 | sched_group_cpus(group)); | |
4114 | ||
4115 | /* Tally up the load of all CPUs in the group */ | |
4116 | avg_load = 0; | |
4117 | ||
4118 | for_each_cpu(i, sched_group_cpus(group)) { | |
4119 | /* Bias balancing toward cpus of our domain */ | |
4120 | if (local_group) | |
4121 | load = source_load(i, load_idx); | |
4122 | else | |
4123 | load = target_load(i, load_idx); | |
4124 | ||
4125 | avg_load += load; | |
4126 | } | |
4127 | ||
4128 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 4129 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
4130 | |
4131 | if (local_group) { | |
4132 | this_load = avg_load; | |
aaee1203 PZ |
4133 | } else if (avg_load < min_load) { |
4134 | min_load = avg_load; | |
4135 | idlest = group; | |
4136 | } | |
4137 | } while (group = group->next, group != sd->groups); | |
4138 | ||
4139 | if (!idlest || 100*this_load < imbalance*min_load) | |
4140 | return NULL; | |
4141 | return idlest; | |
4142 | } | |
4143 | ||
4144 | /* | |
4145 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
4146 | */ | |
4147 | static int | |
4148 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
4149 | { | |
4150 | unsigned long load, min_load = ULONG_MAX; | |
4151 | int idlest = -1; | |
4152 | int i; | |
4153 | ||
4154 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 4155 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
4156 | load = weighted_cpuload(i); |
4157 | ||
4158 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
4159 | min_load = load; | |
4160 | idlest = i; | |
e7693a36 GH |
4161 | } |
4162 | } | |
4163 | ||
aaee1203 PZ |
4164 | return idlest; |
4165 | } | |
e7693a36 | 4166 | |
a50bde51 PZ |
4167 | /* |
4168 | * Try and locate an idle CPU in the sched_domain. | |
4169 | */ | |
99bd5e2f | 4170 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 4171 | { |
99bd5e2f | 4172 | struct sched_domain *sd; |
37407ea7 | 4173 | struct sched_group *sg; |
e0a79f52 | 4174 | int i = task_cpu(p); |
a50bde51 | 4175 | |
e0a79f52 MG |
4176 | if (idle_cpu(target)) |
4177 | return target; | |
99bd5e2f SS |
4178 | |
4179 | /* | |
e0a79f52 | 4180 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 4181 | */ |
e0a79f52 MG |
4182 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4183 | return i; | |
a50bde51 PZ |
4184 | |
4185 | /* | |
37407ea7 | 4186 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 4187 | */ |
518cd623 | 4188 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 4189 | for_each_lower_domain(sd) { |
37407ea7 LT |
4190 | sg = sd->groups; |
4191 | do { | |
4192 | if (!cpumask_intersects(sched_group_cpus(sg), | |
4193 | tsk_cpus_allowed(p))) | |
4194 | goto next; | |
4195 | ||
4196 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 4197 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
4198 | goto next; |
4199 | } | |
970e1789 | 4200 | |
37407ea7 LT |
4201 | target = cpumask_first_and(sched_group_cpus(sg), |
4202 | tsk_cpus_allowed(p)); | |
4203 | goto done; | |
4204 | next: | |
4205 | sg = sg->next; | |
4206 | } while (sg != sd->groups); | |
4207 | } | |
4208 | done: | |
a50bde51 PZ |
4209 | return target; |
4210 | } | |
4211 | ||
aaee1203 PZ |
4212 | /* |
4213 | * sched_balance_self: balance the current task (running on cpu) in domains | |
4214 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
4215 | * SD_BALANCE_EXEC. | |
4216 | * | |
4217 | * Balance, ie. select the least loaded group. | |
4218 | * | |
4219 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
4220 | * | |
4221 | * preempt must be disabled. | |
4222 | */ | |
0017d735 | 4223 | static int |
ac66f547 | 4224 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 4225 | { |
29cd8bae | 4226 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 4227 | int cpu = smp_processor_id(); |
c88d5910 | 4228 | int new_cpu = cpu; |
99bd5e2f | 4229 | int want_affine = 0; |
5158f4e4 | 4230 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 4231 | |
29baa747 | 4232 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
4233 | return prev_cpu; |
4234 | ||
0763a660 | 4235 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 4236 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
4237 | want_affine = 1; |
4238 | new_cpu = prev_cpu; | |
4239 | } | |
aaee1203 | 4240 | |
dce840a0 | 4241 | rcu_read_lock(); |
aaee1203 | 4242 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4243 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4244 | continue; | |
4245 | ||
fe3bcfe1 | 4246 | /* |
99bd5e2f SS |
4247 | * If both cpu and prev_cpu are part of this domain, |
4248 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4249 | */ |
99bd5e2f SS |
4250 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4251 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4252 | affine_sd = tmp; | |
29cd8bae | 4253 | break; |
f03542a7 | 4254 | } |
29cd8bae | 4255 | |
f03542a7 | 4256 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4257 | sd = tmp; |
4258 | } | |
4259 | ||
8b911acd | 4260 | if (affine_sd) { |
f03542a7 | 4261 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
4262 | prev_cpu = cpu; |
4263 | ||
4264 | new_cpu = select_idle_sibling(p, prev_cpu); | |
4265 | goto unlock; | |
8b911acd | 4266 | } |
e7693a36 | 4267 | |
aaee1203 | 4268 | while (sd) { |
5158f4e4 | 4269 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4270 | struct sched_group *group; |
c88d5910 | 4271 | int weight; |
098fb9db | 4272 | |
0763a660 | 4273 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4274 | sd = sd->child; |
4275 | continue; | |
4276 | } | |
098fb9db | 4277 | |
5158f4e4 PZ |
4278 | if (sd_flag & SD_BALANCE_WAKE) |
4279 | load_idx = sd->wake_idx; | |
098fb9db | 4280 | |
5158f4e4 | 4281 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
4282 | if (!group) { |
4283 | sd = sd->child; | |
4284 | continue; | |
4285 | } | |
4ae7d5ce | 4286 | |
d7c33c49 | 4287 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4288 | if (new_cpu == -1 || new_cpu == cpu) { |
4289 | /* Now try balancing at a lower domain level of cpu */ | |
4290 | sd = sd->child; | |
4291 | continue; | |
e7693a36 | 4292 | } |
aaee1203 PZ |
4293 | |
4294 | /* Now try balancing at a lower domain level of new_cpu */ | |
4295 | cpu = new_cpu; | |
669c55e9 | 4296 | weight = sd->span_weight; |
aaee1203 PZ |
4297 | sd = NULL; |
4298 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4299 | if (weight <= tmp->span_weight) |
aaee1203 | 4300 | break; |
0763a660 | 4301 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4302 | sd = tmp; |
4303 | } | |
4304 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4305 | } |
dce840a0 PZ |
4306 | unlock: |
4307 | rcu_read_unlock(); | |
e7693a36 | 4308 | |
c88d5910 | 4309 | return new_cpu; |
e7693a36 | 4310 | } |
0a74bef8 PT |
4311 | |
4312 | /* | |
4313 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4314 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4315 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4316 | * other assumptions, including the state of rq->lock, should be made. | |
4317 | */ | |
4318 | static void | |
4319 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4320 | { | |
aff3e498 PT |
4321 | struct sched_entity *se = &p->se; |
4322 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4323 | ||
4324 | /* | |
4325 | * Load tracking: accumulate removed load so that it can be processed | |
4326 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4327 | * to blocked load iff they have a positive decay-count. It can never | |
4328 | * be negative here since on-rq tasks have decay-count == 0. | |
4329 | */ | |
4330 | if (se->avg.decay_count) { | |
4331 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4332 | atomic_long_add(se->avg.load_avg_contrib, |
4333 | &cfs_rq->removed_load); | |
aff3e498 | 4334 | } |
0a74bef8 | 4335 | } |
e7693a36 GH |
4336 | #endif /* CONFIG_SMP */ |
4337 | ||
e52fb7c0 PZ |
4338 | static unsigned long |
4339 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4340 | { |
4341 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4342 | ||
4343 | /* | |
e52fb7c0 PZ |
4344 | * Since its curr running now, convert the gran from real-time |
4345 | * to virtual-time in his units. | |
13814d42 MG |
4346 | * |
4347 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4348 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4349 | * the resulting gran will be larger, therefore penalizing the | |
4350 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4351 | * be smaller, again penalizing the lighter task. | |
4352 | * | |
4353 | * This is especially important for buddies when the leftmost | |
4354 | * task is higher priority than the buddy. | |
0bbd3336 | 4355 | */ |
f4ad9bd2 | 4356 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4357 | } |
4358 | ||
464b7527 PZ |
4359 | /* |
4360 | * Should 'se' preempt 'curr'. | |
4361 | * | |
4362 | * |s1 | |
4363 | * |s2 | |
4364 | * |s3 | |
4365 | * g | |
4366 | * |<--->|c | |
4367 | * | |
4368 | * w(c, s1) = -1 | |
4369 | * w(c, s2) = 0 | |
4370 | * w(c, s3) = 1 | |
4371 | * | |
4372 | */ | |
4373 | static int | |
4374 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4375 | { | |
4376 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4377 | ||
4378 | if (vdiff <= 0) | |
4379 | return -1; | |
4380 | ||
e52fb7c0 | 4381 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4382 | if (vdiff > gran) |
4383 | return 1; | |
4384 | ||
4385 | return 0; | |
4386 | } | |
4387 | ||
02479099 PZ |
4388 | static void set_last_buddy(struct sched_entity *se) |
4389 | { | |
69c80f3e VP |
4390 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4391 | return; | |
4392 | ||
4393 | for_each_sched_entity(se) | |
4394 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4395 | } |
4396 | ||
4397 | static void set_next_buddy(struct sched_entity *se) | |
4398 | { | |
69c80f3e VP |
4399 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4400 | return; | |
4401 | ||
4402 | for_each_sched_entity(se) | |
4403 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4404 | } |
4405 | ||
ac53db59 RR |
4406 | static void set_skip_buddy(struct sched_entity *se) |
4407 | { | |
69c80f3e VP |
4408 | for_each_sched_entity(se) |
4409 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4410 | } |
4411 | ||
bf0f6f24 IM |
4412 | /* |
4413 | * Preempt the current task with a newly woken task if needed: | |
4414 | */ | |
5a9b86f6 | 4415 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4416 | { |
4417 | struct task_struct *curr = rq->curr; | |
8651a86c | 4418 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4419 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4420 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4421 | int next_buddy_marked = 0; |
bf0f6f24 | 4422 | |
4ae7d5ce IM |
4423 | if (unlikely(se == pse)) |
4424 | return; | |
4425 | ||
5238cdd3 | 4426 | /* |
ddcdf6e7 | 4427 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
4428 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4429 | * lead to a throttle). This both saves work and prevents false | |
4430 | * next-buddy nomination below. | |
4431 | */ | |
4432 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4433 | return; | |
4434 | ||
2f36825b | 4435 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4436 | set_next_buddy(pse); |
2f36825b VP |
4437 | next_buddy_marked = 1; |
4438 | } | |
57fdc26d | 4439 | |
aec0a514 BR |
4440 | /* |
4441 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4442 | * wake up path. | |
5238cdd3 PT |
4443 | * |
4444 | * Note: this also catches the edge-case of curr being in a throttled | |
4445 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4446 | * enqueue of curr) will have resulted in resched being set. This | |
4447 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4448 | * below. | |
aec0a514 BR |
4449 | */ |
4450 | if (test_tsk_need_resched(curr)) | |
4451 | return; | |
4452 | ||
a2f5c9ab DH |
4453 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4454 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4455 | likely(p->policy != SCHED_IDLE)) | |
4456 | goto preempt; | |
4457 | ||
91c234b4 | 4458 | /* |
a2f5c9ab DH |
4459 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4460 | * is driven by the tick): | |
91c234b4 | 4461 | */ |
8ed92e51 | 4462 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4463 | return; |
bf0f6f24 | 4464 | |
464b7527 | 4465 | find_matching_se(&se, &pse); |
9bbd7374 | 4466 | update_curr(cfs_rq_of(se)); |
002f128b | 4467 | BUG_ON(!pse); |
2f36825b VP |
4468 | if (wakeup_preempt_entity(se, pse) == 1) { |
4469 | /* | |
4470 | * Bias pick_next to pick the sched entity that is | |
4471 | * triggering this preemption. | |
4472 | */ | |
4473 | if (!next_buddy_marked) | |
4474 | set_next_buddy(pse); | |
3a7e73a2 | 4475 | goto preempt; |
2f36825b | 4476 | } |
464b7527 | 4477 | |
3a7e73a2 | 4478 | return; |
a65ac745 | 4479 | |
3a7e73a2 PZ |
4480 | preempt: |
4481 | resched_task(curr); | |
4482 | /* | |
4483 | * Only set the backward buddy when the current task is still | |
4484 | * on the rq. This can happen when a wakeup gets interleaved | |
4485 | * with schedule on the ->pre_schedule() or idle_balance() | |
4486 | * point, either of which can * drop the rq lock. | |
4487 | * | |
4488 | * Also, during early boot the idle thread is in the fair class, | |
4489 | * for obvious reasons its a bad idea to schedule back to it. | |
4490 | */ | |
4491 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
4492 | return; | |
4493 | ||
4494 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
4495 | set_last_buddy(se); | |
bf0f6f24 IM |
4496 | } |
4497 | ||
fb8d4724 | 4498 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 4499 | { |
8f4d37ec | 4500 | struct task_struct *p; |
bf0f6f24 IM |
4501 | struct cfs_rq *cfs_rq = &rq->cfs; |
4502 | struct sched_entity *se; | |
4503 | ||
36ace27e | 4504 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
4505 | return NULL; |
4506 | ||
4507 | do { | |
9948f4b2 | 4508 | se = pick_next_entity(cfs_rq); |
f4b6755f | 4509 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
4510 | cfs_rq = group_cfs_rq(se); |
4511 | } while (cfs_rq); | |
4512 | ||
8f4d37ec | 4513 | p = task_of(se); |
b39e66ea MG |
4514 | if (hrtick_enabled(rq)) |
4515 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4516 | |
4517 | return p; | |
bf0f6f24 IM |
4518 | } |
4519 | ||
4520 | /* | |
4521 | * Account for a descheduled task: | |
4522 | */ | |
31ee529c | 4523 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4524 | { |
4525 | struct sched_entity *se = &prev->se; | |
4526 | struct cfs_rq *cfs_rq; | |
4527 | ||
4528 | for_each_sched_entity(se) { | |
4529 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4530 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4531 | } |
4532 | } | |
4533 | ||
ac53db59 RR |
4534 | /* |
4535 | * sched_yield() is very simple | |
4536 | * | |
4537 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4538 | */ | |
4539 | static void yield_task_fair(struct rq *rq) | |
4540 | { | |
4541 | struct task_struct *curr = rq->curr; | |
4542 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4543 | struct sched_entity *se = &curr->se; | |
4544 | ||
4545 | /* | |
4546 | * Are we the only task in the tree? | |
4547 | */ | |
4548 | if (unlikely(rq->nr_running == 1)) | |
4549 | return; | |
4550 | ||
4551 | clear_buddies(cfs_rq, se); | |
4552 | ||
4553 | if (curr->policy != SCHED_BATCH) { | |
4554 | update_rq_clock(rq); | |
4555 | /* | |
4556 | * Update run-time statistics of the 'current'. | |
4557 | */ | |
4558 | update_curr(cfs_rq); | |
916671c0 MG |
4559 | /* |
4560 | * Tell update_rq_clock() that we've just updated, | |
4561 | * so we don't do microscopic update in schedule() | |
4562 | * and double the fastpath cost. | |
4563 | */ | |
4564 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4565 | } |
4566 | ||
4567 | set_skip_buddy(se); | |
4568 | } | |
4569 | ||
d95f4122 MG |
4570 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4571 | { | |
4572 | struct sched_entity *se = &p->se; | |
4573 | ||
5238cdd3 PT |
4574 | /* throttled hierarchies are not runnable */ |
4575 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4576 | return false; |
4577 | ||
4578 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4579 | set_next_buddy(se); | |
4580 | ||
d95f4122 MG |
4581 | yield_task_fair(rq); |
4582 | ||
4583 | return true; | |
4584 | } | |
4585 | ||
681f3e68 | 4586 | #ifdef CONFIG_SMP |
bf0f6f24 | 4587 | /************************************************** |
e9c84cb8 PZ |
4588 | * Fair scheduling class load-balancing methods. |
4589 | * | |
4590 | * BASICS | |
4591 | * | |
4592 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4593 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4594 | * time to each task. This is expressed in the following equation: | |
4595 | * | |
4596 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4597 | * | |
4598 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4599 | * W_i,0 is defined as: | |
4600 | * | |
4601 | * W_i,0 = \Sum_j w_i,j (2) | |
4602 | * | |
4603 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4604 | * is derived from the nice value as per prio_to_weight[]. | |
4605 | * | |
4606 | * The weight average is an exponential decay average of the instantaneous | |
4607 | * weight: | |
4608 | * | |
4609 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4610 | * | |
4611 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4612 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4613 | * can also include other factors [XXX]. | |
4614 | * | |
4615 | * To achieve this balance we define a measure of imbalance which follows | |
4616 | * directly from (1): | |
4617 | * | |
4618 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4619 | * | |
4620 | * We them move tasks around to minimize the imbalance. In the continuous | |
4621 | * function space it is obvious this converges, in the discrete case we get | |
4622 | * a few fun cases generally called infeasible weight scenarios. | |
4623 | * | |
4624 | * [XXX expand on: | |
4625 | * - infeasible weights; | |
4626 | * - local vs global optima in the discrete case. ] | |
4627 | * | |
4628 | * | |
4629 | * SCHED DOMAINS | |
4630 | * | |
4631 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4632 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4633 | * topology where each level pairs two lower groups (or better). This results | |
4634 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4635 | * tree to only the first of the previous level and we decrease the frequency | |
4636 | * of load-balance at each level inv. proportional to the number of cpus in | |
4637 | * the groups. | |
4638 | * | |
4639 | * This yields: | |
4640 | * | |
4641 | * log_2 n 1 n | |
4642 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4643 | * i = 0 2^i 2^i | |
4644 | * `- size of each group | |
4645 | * | | `- number of cpus doing load-balance | |
4646 | * | `- freq | |
4647 | * `- sum over all levels | |
4648 | * | |
4649 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4650 | * this makes (5) the runtime complexity of the balancer. | |
4651 | * | |
4652 | * An important property here is that each CPU is still (indirectly) connected | |
4653 | * to every other cpu in at most O(log n) steps: | |
4654 | * | |
4655 | * The adjacency matrix of the resulting graph is given by: | |
4656 | * | |
4657 | * log_2 n | |
4658 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4659 | * k = 0 | |
4660 | * | |
4661 | * And you'll find that: | |
4662 | * | |
4663 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
4664 | * | |
4665 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
4666 | * The task movement gives a factor of O(m), giving a convergence complexity | |
4667 | * of: | |
4668 | * | |
4669 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
4670 | * | |
4671 | * | |
4672 | * WORK CONSERVING | |
4673 | * | |
4674 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
4675 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
4676 | * tree itself instead of relying on other CPUs to bring it work. | |
4677 | * | |
4678 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
4679 | * time. | |
4680 | * | |
4681 | * [XXX more?] | |
4682 | * | |
4683 | * | |
4684 | * CGROUPS | |
4685 | * | |
4686 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
4687 | * | |
4688 | * s_k,i | |
4689 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
4690 | * S_k | |
4691 | * | |
4692 | * Where | |
4693 | * | |
4694 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4695 | * | |
4696 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4697 | * | |
4698 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4699 | * property. | |
4700 | * | |
4701 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4702 | * rewrite all of this once again.] | |
4703 | */ | |
bf0f6f24 | 4704 | |
ed387b78 HS |
4705 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4706 | ||
0ec8aa00 PZ |
4707 | enum fbq_type { regular, remote, all }; |
4708 | ||
ddcdf6e7 | 4709 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4710 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4711 | #define LBF_DST_PINNED 0x04 |
4712 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4713 | |
4714 | struct lb_env { | |
4715 | struct sched_domain *sd; | |
4716 | ||
ddcdf6e7 | 4717 | struct rq *src_rq; |
85c1e7da | 4718 | int src_cpu; |
ddcdf6e7 PZ |
4719 | |
4720 | int dst_cpu; | |
4721 | struct rq *dst_rq; | |
4722 | ||
88b8dac0 SV |
4723 | struct cpumask *dst_grpmask; |
4724 | int new_dst_cpu; | |
ddcdf6e7 | 4725 | enum cpu_idle_type idle; |
bd939f45 | 4726 | long imbalance; |
b9403130 MW |
4727 | /* The set of CPUs under consideration for load-balancing */ |
4728 | struct cpumask *cpus; | |
4729 | ||
ddcdf6e7 | 4730 | unsigned int flags; |
367456c7 PZ |
4731 | |
4732 | unsigned int loop; | |
4733 | unsigned int loop_break; | |
4734 | unsigned int loop_max; | |
0ec8aa00 PZ |
4735 | |
4736 | enum fbq_type fbq_type; | |
ddcdf6e7 PZ |
4737 | }; |
4738 | ||
1e3c88bd | 4739 | /* |
ddcdf6e7 | 4740 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
4741 | * Both runqueues must be locked. |
4742 | */ | |
ddcdf6e7 | 4743 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 4744 | { |
ddcdf6e7 PZ |
4745 | deactivate_task(env->src_rq, p, 0); |
4746 | set_task_cpu(p, env->dst_cpu); | |
4747 | activate_task(env->dst_rq, p, 0); | |
4748 | check_preempt_curr(env->dst_rq, p, 0); | |
6fe6b2d6 RR |
4749 | #ifdef CONFIG_NUMA_BALANCING |
4750 | if (p->numa_preferred_nid != -1) { | |
4751 | int src_nid = cpu_to_node(env->src_cpu); | |
4752 | int dst_nid = cpu_to_node(env->dst_cpu); | |
4753 | ||
4754 | /* | |
4755 | * If the load balancer has moved the task then limit | |
4756 | * migrations from taking place in the short term in | |
4757 | * case this is a short-lived migration. | |
4758 | */ | |
4759 | if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid) | |
4760 | p->numa_migrate_seq = 0; | |
4761 | } | |
4762 | #endif | |
1e3c88bd PZ |
4763 | } |
4764 | ||
029632fb PZ |
4765 | /* |
4766 | * Is this task likely cache-hot: | |
4767 | */ | |
4768 | static int | |
4769 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
4770 | { | |
4771 | s64 delta; | |
4772 | ||
4773 | if (p->sched_class != &fair_sched_class) | |
4774 | return 0; | |
4775 | ||
4776 | if (unlikely(p->policy == SCHED_IDLE)) | |
4777 | return 0; | |
4778 | ||
4779 | /* | |
4780 | * Buddy candidates are cache hot: | |
4781 | */ | |
4782 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
4783 | (&p->se == cfs_rq_of(&p->se)->next || | |
4784 | &p->se == cfs_rq_of(&p->se)->last)) | |
4785 | return 1; | |
4786 | ||
4787 | if (sysctl_sched_migration_cost == -1) | |
4788 | return 1; | |
4789 | if (sysctl_sched_migration_cost == 0) | |
4790 | return 0; | |
4791 | ||
4792 | delta = now - p->se.exec_start; | |
4793 | ||
4794 | return delta < (s64)sysctl_sched_migration_cost; | |
4795 | } | |
4796 | ||
3a7053b3 MG |
4797 | #ifdef CONFIG_NUMA_BALANCING |
4798 | /* Returns true if the destination node has incurred more faults */ | |
4799 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
4800 | { | |
4801 | int src_nid, dst_nid; | |
4802 | ||
4803 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || | |
4804 | !(env->sd->flags & SD_NUMA)) { | |
4805 | return false; | |
4806 | } | |
4807 | ||
4808 | src_nid = cpu_to_node(env->src_cpu); | |
4809 | dst_nid = cpu_to_node(env->dst_cpu); | |
4810 | ||
83e1d2cd | 4811 | if (src_nid == dst_nid) |
3a7053b3 MG |
4812 | return false; |
4813 | ||
83e1d2cd MG |
4814 | /* Always encourage migration to the preferred node. */ |
4815 | if (dst_nid == p->numa_preferred_nid) | |
4816 | return true; | |
4817 | ||
887c290e RR |
4818 | /* If both task and group weight improve, this move is a winner. */ |
4819 | if (task_weight(p, dst_nid) > task_weight(p, src_nid) && | |
4820 | group_weight(p, dst_nid) > group_weight(p, src_nid)) | |
3a7053b3 MG |
4821 | return true; |
4822 | ||
4823 | return false; | |
4824 | } | |
7a0f3083 MG |
4825 | |
4826 | ||
4827 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
4828 | { | |
4829 | int src_nid, dst_nid; | |
4830 | ||
4831 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
4832 | return false; | |
4833 | ||
4834 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) | |
4835 | return false; | |
4836 | ||
4837 | src_nid = cpu_to_node(env->src_cpu); | |
4838 | dst_nid = cpu_to_node(env->dst_cpu); | |
4839 | ||
83e1d2cd | 4840 | if (src_nid == dst_nid) |
7a0f3083 MG |
4841 | return false; |
4842 | ||
83e1d2cd MG |
4843 | /* Migrating away from the preferred node is always bad. */ |
4844 | if (src_nid == p->numa_preferred_nid) | |
4845 | return true; | |
4846 | ||
887c290e RR |
4847 | /* If either task or group weight get worse, don't do it. */ |
4848 | if (task_weight(p, dst_nid) < task_weight(p, src_nid) || | |
4849 | group_weight(p, dst_nid) < group_weight(p, src_nid)) | |
7a0f3083 MG |
4850 | return true; |
4851 | ||
4852 | return false; | |
4853 | } | |
4854 | ||
3a7053b3 MG |
4855 | #else |
4856 | static inline bool migrate_improves_locality(struct task_struct *p, | |
4857 | struct lb_env *env) | |
4858 | { | |
4859 | return false; | |
4860 | } | |
7a0f3083 MG |
4861 | |
4862 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
4863 | struct lb_env *env) | |
4864 | { | |
4865 | return false; | |
4866 | } | |
3a7053b3 MG |
4867 | #endif |
4868 | ||
1e3c88bd PZ |
4869 | /* |
4870 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
4871 | */ | |
4872 | static | |
8e45cb54 | 4873 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
4874 | { |
4875 | int tsk_cache_hot = 0; | |
4876 | /* | |
4877 | * We do not migrate tasks that are: | |
d3198084 | 4878 | * 1) throttled_lb_pair, or |
1e3c88bd | 4879 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
4880 | * 3) running (obviously), or |
4881 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 4882 | */ |
d3198084 JK |
4883 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
4884 | return 0; | |
4885 | ||
ddcdf6e7 | 4886 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 4887 | int cpu; |
88b8dac0 | 4888 | |
41acab88 | 4889 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 4890 | |
6263322c PZ |
4891 | env->flags |= LBF_SOME_PINNED; |
4892 | ||
88b8dac0 SV |
4893 | /* |
4894 | * Remember if this task can be migrated to any other cpu in | |
4895 | * our sched_group. We may want to revisit it if we couldn't | |
4896 | * meet load balance goals by pulling other tasks on src_cpu. | |
4897 | * | |
4898 | * Also avoid computing new_dst_cpu if we have already computed | |
4899 | * one in current iteration. | |
4900 | */ | |
6263322c | 4901 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
4902 | return 0; |
4903 | ||
e02e60c1 JK |
4904 | /* Prevent to re-select dst_cpu via env's cpus */ |
4905 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
4906 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 4907 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
4908 | env->new_dst_cpu = cpu; |
4909 | break; | |
4910 | } | |
88b8dac0 | 4911 | } |
e02e60c1 | 4912 | |
1e3c88bd PZ |
4913 | return 0; |
4914 | } | |
88b8dac0 SV |
4915 | |
4916 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 4917 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 4918 | |
ddcdf6e7 | 4919 | if (task_running(env->src_rq, p)) { |
41acab88 | 4920 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
4921 | return 0; |
4922 | } | |
4923 | ||
4924 | /* | |
4925 | * Aggressive migration if: | |
3a7053b3 MG |
4926 | * 1) destination numa is preferred |
4927 | * 2) task is cache cold, or | |
4928 | * 3) too many balance attempts have failed. | |
1e3c88bd | 4929 | */ |
78becc27 | 4930 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
7a0f3083 MG |
4931 | if (!tsk_cache_hot) |
4932 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
4933 | |
4934 | if (migrate_improves_locality(p, env)) { | |
4935 | #ifdef CONFIG_SCHEDSTATS | |
4936 | if (tsk_cache_hot) { | |
4937 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
4938 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
4939 | } | |
4940 | #endif | |
4941 | return 1; | |
4942 | } | |
4943 | ||
1e3c88bd | 4944 | if (!tsk_cache_hot || |
8e45cb54 | 4945 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 4946 | |
1e3c88bd | 4947 | if (tsk_cache_hot) { |
8e45cb54 | 4948 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 4949 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 4950 | } |
4e2dcb73 | 4951 | |
1e3c88bd PZ |
4952 | return 1; |
4953 | } | |
4954 | ||
4e2dcb73 ZH |
4955 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
4956 | return 0; | |
1e3c88bd PZ |
4957 | } |
4958 | ||
897c395f PZ |
4959 | /* |
4960 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
4961 | * part of active balancing operations within "domain". | |
4962 | * Returns 1 if successful and 0 otherwise. | |
4963 | * | |
4964 | * Called with both runqueues locked. | |
4965 | */ | |
8e45cb54 | 4966 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
4967 | { |
4968 | struct task_struct *p, *n; | |
897c395f | 4969 | |
367456c7 | 4970 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
4971 | if (!can_migrate_task(p, env)) |
4972 | continue; | |
897c395f | 4973 | |
367456c7 PZ |
4974 | move_task(p, env); |
4975 | /* | |
4976 | * Right now, this is only the second place move_task() | |
4977 | * is called, so we can safely collect move_task() | |
4978 | * stats here rather than inside move_task(). | |
4979 | */ | |
4980 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4981 | return 1; | |
897c395f | 4982 | } |
897c395f PZ |
4983 | return 0; |
4984 | } | |
4985 | ||
eb95308e PZ |
4986 | static const unsigned int sched_nr_migrate_break = 32; |
4987 | ||
5d6523eb | 4988 | /* |
bd939f45 | 4989 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4990 | * this_rq, as part of a balancing operation within domain "sd". |
4991 | * Returns 1 if successful and 0 otherwise. | |
4992 | * | |
4993 | * Called with both runqueues locked. | |
4994 | */ | |
4995 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4996 | { |
5d6523eb PZ |
4997 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4998 | struct task_struct *p; | |
367456c7 PZ |
4999 | unsigned long load; |
5000 | int pulled = 0; | |
1e3c88bd | 5001 | |
bd939f45 | 5002 | if (env->imbalance <= 0) |
5d6523eb | 5003 | return 0; |
1e3c88bd | 5004 | |
5d6523eb PZ |
5005 | while (!list_empty(tasks)) { |
5006 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 5007 | |
367456c7 PZ |
5008 | env->loop++; |
5009 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 5010 | if (env->loop > env->loop_max) |
367456c7 | 5011 | break; |
5d6523eb PZ |
5012 | |
5013 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 5014 | if (env->loop > env->loop_break) { |
eb95308e | 5015 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 5016 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 5017 | break; |
a195f004 | 5018 | } |
1e3c88bd | 5019 | |
d3198084 | 5020 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
5021 | goto next; |
5022 | ||
5023 | load = task_h_load(p); | |
5d6523eb | 5024 | |
eb95308e | 5025 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
5026 | goto next; |
5027 | ||
bd939f45 | 5028 | if ((load / 2) > env->imbalance) |
367456c7 | 5029 | goto next; |
1e3c88bd | 5030 | |
ddcdf6e7 | 5031 | move_task(p, env); |
ee00e66f | 5032 | pulled++; |
bd939f45 | 5033 | env->imbalance -= load; |
1e3c88bd PZ |
5034 | |
5035 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
5036 | /* |
5037 | * NEWIDLE balancing is a source of latency, so preemptible | |
5038 | * kernels will stop after the first task is pulled to minimize | |
5039 | * the critical section. | |
5040 | */ | |
5d6523eb | 5041 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 5042 | break; |
1e3c88bd PZ |
5043 | #endif |
5044 | ||
ee00e66f PZ |
5045 | /* |
5046 | * We only want to steal up to the prescribed amount of | |
5047 | * weighted load. | |
5048 | */ | |
bd939f45 | 5049 | if (env->imbalance <= 0) |
ee00e66f | 5050 | break; |
367456c7 PZ |
5051 | |
5052 | continue; | |
5053 | next: | |
5d6523eb | 5054 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 5055 | } |
5d6523eb | 5056 | |
1e3c88bd | 5057 | /* |
ddcdf6e7 PZ |
5058 | * Right now, this is one of only two places move_task() is called, |
5059 | * so we can safely collect move_task() stats here rather than | |
5060 | * inside move_task(). | |
1e3c88bd | 5061 | */ |
8e45cb54 | 5062 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 5063 | |
5d6523eb | 5064 | return pulled; |
1e3c88bd PZ |
5065 | } |
5066 | ||
230059de | 5067 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
5068 | /* |
5069 | * update tg->load_weight by folding this cpu's load_avg | |
5070 | */ | |
48a16753 | 5071 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 5072 | { |
48a16753 PT |
5073 | struct sched_entity *se = tg->se[cpu]; |
5074 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 5075 | |
48a16753 PT |
5076 | /* throttled entities do not contribute to load */ |
5077 | if (throttled_hierarchy(cfs_rq)) | |
5078 | return; | |
9e3081ca | 5079 | |
aff3e498 | 5080 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 5081 | |
82958366 PT |
5082 | if (se) { |
5083 | update_entity_load_avg(se, 1); | |
5084 | /* | |
5085 | * We pivot on our runnable average having decayed to zero for | |
5086 | * list removal. This generally implies that all our children | |
5087 | * have also been removed (modulo rounding error or bandwidth | |
5088 | * control); however, such cases are rare and we can fix these | |
5089 | * at enqueue. | |
5090 | * | |
5091 | * TODO: fix up out-of-order children on enqueue. | |
5092 | */ | |
5093 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
5094 | list_del_leaf_cfs_rq(cfs_rq); | |
5095 | } else { | |
48a16753 | 5096 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
5097 | update_rq_runnable_avg(rq, rq->nr_running); |
5098 | } | |
9e3081ca PZ |
5099 | } |
5100 | ||
48a16753 | 5101 | static void update_blocked_averages(int cpu) |
9e3081ca | 5102 | { |
9e3081ca | 5103 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
5104 | struct cfs_rq *cfs_rq; |
5105 | unsigned long flags; | |
9e3081ca | 5106 | |
48a16753 PT |
5107 | raw_spin_lock_irqsave(&rq->lock, flags); |
5108 | update_rq_clock(rq); | |
9763b67f PZ |
5109 | /* |
5110 | * Iterates the task_group tree in a bottom up fashion, see | |
5111 | * list_add_leaf_cfs_rq() for details. | |
5112 | */ | |
64660c86 | 5113 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
5114 | /* |
5115 | * Note: We may want to consider periodically releasing | |
5116 | * rq->lock about these updates so that creating many task | |
5117 | * groups does not result in continually extending hold time. | |
5118 | */ | |
5119 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 5120 | } |
48a16753 PT |
5121 | |
5122 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
5123 | } |
5124 | ||
9763b67f | 5125 | /* |
68520796 | 5126 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
5127 | * This needs to be done in a top-down fashion because the load of a child |
5128 | * group is a fraction of its parents load. | |
5129 | */ | |
68520796 | 5130 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 5131 | { |
68520796 VD |
5132 | struct rq *rq = rq_of(cfs_rq); |
5133 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 5134 | unsigned long now = jiffies; |
68520796 | 5135 | unsigned long load; |
a35b6466 | 5136 | |
68520796 | 5137 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
5138 | return; |
5139 | ||
68520796 VD |
5140 | cfs_rq->h_load_next = NULL; |
5141 | for_each_sched_entity(se) { | |
5142 | cfs_rq = cfs_rq_of(se); | |
5143 | cfs_rq->h_load_next = se; | |
5144 | if (cfs_rq->last_h_load_update == now) | |
5145 | break; | |
5146 | } | |
a35b6466 | 5147 | |
68520796 | 5148 | if (!se) { |
7e3115ef | 5149 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
5150 | cfs_rq->last_h_load_update = now; |
5151 | } | |
5152 | ||
5153 | while ((se = cfs_rq->h_load_next) != NULL) { | |
5154 | load = cfs_rq->h_load; | |
5155 | load = div64_ul(load * se->avg.load_avg_contrib, | |
5156 | cfs_rq->runnable_load_avg + 1); | |
5157 | cfs_rq = group_cfs_rq(se); | |
5158 | cfs_rq->h_load = load; | |
5159 | cfs_rq->last_h_load_update = now; | |
5160 | } | |
9763b67f PZ |
5161 | } |
5162 | ||
367456c7 | 5163 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 5164 | { |
367456c7 | 5165 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 5166 | |
68520796 | 5167 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
5168 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
5169 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
5170 | } |
5171 | #else | |
48a16753 | 5172 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
5173 | { |
5174 | } | |
5175 | ||
367456c7 | 5176 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 5177 | { |
a003a25b | 5178 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 5179 | } |
230059de | 5180 | #endif |
1e3c88bd | 5181 | |
1e3c88bd | 5182 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
5183 | /* |
5184 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
5185 | */ | |
5186 | struct sg_lb_stats { | |
5187 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
5188 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 5189 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 5190 | unsigned long load_per_task; |
3ae11c90 | 5191 | unsigned long group_power; |
147c5fc2 PZ |
5192 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
5193 | unsigned int group_capacity; | |
5194 | unsigned int idle_cpus; | |
5195 | unsigned int group_weight; | |
1e3c88bd | 5196 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 5197 | int group_has_capacity; /* Is there extra capacity in the group? */ |
0ec8aa00 PZ |
5198 | #ifdef CONFIG_NUMA_BALANCING |
5199 | unsigned int nr_numa_running; | |
5200 | unsigned int nr_preferred_running; | |
5201 | #endif | |
1e3c88bd PZ |
5202 | }; |
5203 | ||
56cf515b JK |
5204 | /* |
5205 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
5206 | * during load balancing. | |
5207 | */ | |
5208 | struct sd_lb_stats { | |
5209 | struct sched_group *busiest; /* Busiest group in this sd */ | |
5210 | struct sched_group *local; /* Local group in this sd */ | |
5211 | unsigned long total_load; /* Total load of all groups in sd */ | |
5212 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
5213 | unsigned long avg_load; /* Average load across all groups in sd */ | |
5214 | ||
56cf515b | 5215 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 5216 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
5217 | }; |
5218 | ||
147c5fc2 PZ |
5219 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
5220 | { | |
5221 | /* | |
5222 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
5223 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
5224 | * We must however clear busiest_stat::avg_load because | |
5225 | * update_sd_pick_busiest() reads this before assignment. | |
5226 | */ | |
5227 | *sds = (struct sd_lb_stats){ | |
5228 | .busiest = NULL, | |
5229 | .local = NULL, | |
5230 | .total_load = 0UL, | |
5231 | .total_pwr = 0UL, | |
5232 | .busiest_stat = { | |
5233 | .avg_load = 0UL, | |
5234 | }, | |
5235 | }; | |
5236 | } | |
5237 | ||
1e3c88bd PZ |
5238 | /** |
5239 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
5240 | * @sd: The sched_domain whose load_idx is to be obtained. | |
5241 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
e69f6186 YB |
5242 | * |
5243 | * Return: The load index. | |
1e3c88bd PZ |
5244 | */ |
5245 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
5246 | enum cpu_idle_type idle) | |
5247 | { | |
5248 | int load_idx; | |
5249 | ||
5250 | switch (idle) { | |
5251 | case CPU_NOT_IDLE: | |
5252 | load_idx = sd->busy_idx; | |
5253 | break; | |
5254 | ||
5255 | case CPU_NEWLY_IDLE: | |
5256 | load_idx = sd->newidle_idx; | |
5257 | break; | |
5258 | default: | |
5259 | load_idx = sd->idle_idx; | |
5260 | break; | |
5261 | } | |
5262 | ||
5263 | return load_idx; | |
5264 | } | |
5265 | ||
15f803c9 | 5266 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5267 | { |
1399fa78 | 5268 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
5269 | } |
5270 | ||
5271 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
5272 | { | |
5273 | return default_scale_freq_power(sd, cpu); | |
5274 | } | |
5275 | ||
15f803c9 | 5276 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5277 | { |
669c55e9 | 5278 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
5279 | unsigned long smt_gain = sd->smt_gain; |
5280 | ||
5281 | smt_gain /= weight; | |
5282 | ||
5283 | return smt_gain; | |
5284 | } | |
5285 | ||
5286 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
5287 | { | |
5288 | return default_scale_smt_power(sd, cpu); | |
5289 | } | |
5290 | ||
15f803c9 | 5291 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
5292 | { |
5293 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5294 | u64 total, available, age_stamp, avg; |
1e3c88bd | 5295 | |
b654f7de PZ |
5296 | /* |
5297 | * Since we're reading these variables without serialization make sure | |
5298 | * we read them once before doing sanity checks on them. | |
5299 | */ | |
5300 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5301 | avg = ACCESS_ONCE(rq->rt_avg); | |
5302 | ||
78becc27 | 5303 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 5304 | |
b654f7de | 5305 | if (unlikely(total < avg)) { |
aa483808 VP |
5306 | /* Ensures that power won't end up being negative */ |
5307 | available = 0; | |
5308 | } else { | |
b654f7de | 5309 | available = total - avg; |
aa483808 | 5310 | } |
1e3c88bd | 5311 | |
1399fa78 NR |
5312 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
5313 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 5314 | |
1399fa78 | 5315 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5316 | |
5317 | return div_u64(available, total); | |
5318 | } | |
5319 | ||
5320 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
5321 | { | |
669c55e9 | 5322 | unsigned long weight = sd->span_weight; |
1399fa78 | 5323 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
5324 | struct sched_group *sdg = sd->groups; |
5325 | ||
1e3c88bd PZ |
5326 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
5327 | if (sched_feat(ARCH_POWER)) | |
5328 | power *= arch_scale_smt_power(sd, cpu); | |
5329 | else | |
5330 | power *= default_scale_smt_power(sd, cpu); | |
5331 | ||
1399fa78 | 5332 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5333 | } |
5334 | ||
9c3f75cb | 5335 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
5336 | |
5337 | if (sched_feat(ARCH_POWER)) | |
5338 | power *= arch_scale_freq_power(sd, cpu); | |
5339 | else | |
5340 | power *= default_scale_freq_power(sd, cpu); | |
5341 | ||
1399fa78 | 5342 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 5343 | |
1e3c88bd | 5344 | power *= scale_rt_power(cpu); |
1399fa78 | 5345 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5346 | |
5347 | if (!power) | |
5348 | power = 1; | |
5349 | ||
e51fd5e2 | 5350 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 5351 | sdg->sgp->power = power; |
1e3c88bd PZ |
5352 | } |
5353 | ||
029632fb | 5354 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
5355 | { |
5356 | struct sched_domain *child = sd->child; | |
5357 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 5358 | unsigned long power, power_orig; |
4ec4412e VG |
5359 | unsigned long interval; |
5360 | ||
5361 | interval = msecs_to_jiffies(sd->balance_interval); | |
5362 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
5363 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
5364 | |
5365 | if (!child) { | |
5366 | update_cpu_power(sd, cpu); | |
5367 | return; | |
5368 | } | |
5369 | ||
863bffc8 | 5370 | power_orig = power = 0; |
1e3c88bd | 5371 | |
74a5ce20 PZ |
5372 | if (child->flags & SD_OVERLAP) { |
5373 | /* | |
5374 | * SD_OVERLAP domains cannot assume that child groups | |
5375 | * span the current group. | |
5376 | */ | |
5377 | ||
863bffc8 PZ |
5378 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
5379 | struct sched_group *sg = cpu_rq(cpu)->sd->groups; | |
5380 | ||
5381 | power_orig += sg->sgp->power_orig; | |
5382 | power += sg->sgp->power; | |
5383 | } | |
74a5ce20 PZ |
5384 | } else { |
5385 | /* | |
5386 | * !SD_OVERLAP domains can assume that child groups | |
5387 | * span the current group. | |
5388 | */ | |
5389 | ||
5390 | group = child->groups; | |
5391 | do { | |
863bffc8 | 5392 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
5393 | power += group->sgp->power; |
5394 | group = group->next; | |
5395 | } while (group != child->groups); | |
5396 | } | |
1e3c88bd | 5397 | |
863bffc8 PZ |
5398 | sdg->sgp->power_orig = power_orig; |
5399 | sdg->sgp->power = power; | |
1e3c88bd PZ |
5400 | } |
5401 | ||
9d5efe05 SV |
5402 | /* |
5403 | * Try and fix up capacity for tiny siblings, this is needed when | |
5404 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
5405 | * which on its own isn't powerful enough. | |
5406 | * | |
5407 | * See update_sd_pick_busiest() and check_asym_packing(). | |
5408 | */ | |
5409 | static inline int | |
5410 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
5411 | { | |
5412 | /* | |
1399fa78 | 5413 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 5414 | */ |
a6c75f2f | 5415 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
5416 | return 0; |
5417 | ||
5418 | /* | |
5419 | * If ~90% of the cpu_power is still there, we're good. | |
5420 | */ | |
9c3f75cb | 5421 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
5422 | return 1; |
5423 | ||
5424 | return 0; | |
5425 | } | |
5426 | ||
30ce5dab PZ |
5427 | /* |
5428 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
5429 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
5430 | * | |
5431 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
5432 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
5433 | * Something like: | |
5434 | * | |
5435 | * { 0 1 2 3 } { 4 5 6 7 } | |
5436 | * * * * * | |
5437 | * | |
5438 | * If we were to balance group-wise we'd place two tasks in the first group and | |
5439 | * two tasks in the second group. Clearly this is undesired as it will overload | |
5440 | * cpu 3 and leave one of the cpus in the second group unused. | |
5441 | * | |
5442 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
5443 | * by noticing the lower domain failed to reach balance and had difficulty |
5444 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
5445 | * |
5446 | * When this is so detected; this group becomes a candidate for busiest; see | |
5447 | * update_sd_pick_busiest(). And calculcate_imbalance() and | |
6263322c | 5448 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
5449 | * to create an effective group imbalance. |
5450 | * | |
5451 | * This is a somewhat tricky proposition since the next run might not find the | |
5452 | * group imbalance and decide the groups need to be balanced again. A most | |
5453 | * subtle and fragile situation. | |
5454 | */ | |
5455 | ||
6263322c | 5456 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 5457 | { |
6263322c | 5458 | return group->sgp->imbalance; |
30ce5dab PZ |
5459 | } |
5460 | ||
b37d9316 PZ |
5461 | /* |
5462 | * Compute the group capacity. | |
5463 | * | |
c61037e9 PZ |
5464 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
5465 | * first dividing out the smt factor and computing the actual number of cores | |
5466 | * and limit power unit capacity with that. | |
b37d9316 PZ |
5467 | */ |
5468 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
5469 | { | |
c61037e9 PZ |
5470 | unsigned int capacity, smt, cpus; |
5471 | unsigned int power, power_orig; | |
5472 | ||
5473 | power = group->sgp->power; | |
5474 | power_orig = group->sgp->power_orig; | |
5475 | cpus = group->group_weight; | |
b37d9316 | 5476 | |
c61037e9 PZ |
5477 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
5478 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
5479 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 5480 | |
c61037e9 | 5481 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
5482 | if (!capacity) |
5483 | capacity = fix_small_capacity(env->sd, group); | |
5484 | ||
5485 | return capacity; | |
5486 | } | |
5487 | ||
1e3c88bd PZ |
5488 | /** |
5489 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 5490 | * @env: The load balancing environment. |
1e3c88bd | 5491 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 5492 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 5493 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
5494 | * @sgs: variable to hold the statistics for this group. |
5495 | */ | |
bd939f45 PZ |
5496 | static inline void update_sg_lb_stats(struct lb_env *env, |
5497 | struct sched_group *group, int load_idx, | |
23f0d209 | 5498 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 5499 | { |
30ce5dab PZ |
5500 | unsigned long nr_running; |
5501 | unsigned long load; | |
bd939f45 | 5502 | int i; |
1e3c88bd | 5503 | |
b72ff13c PZ |
5504 | memset(sgs, 0, sizeof(*sgs)); |
5505 | ||
b9403130 | 5506 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
5507 | struct rq *rq = cpu_rq(i); |
5508 | ||
e44bc5c5 PZ |
5509 | nr_running = rq->nr_running; |
5510 | ||
1e3c88bd | 5511 | /* Bias balancing toward cpus of our domain */ |
6263322c | 5512 | if (local_group) |
04f733b4 | 5513 | load = target_load(i, load_idx); |
6263322c | 5514 | else |
1e3c88bd | 5515 | load = source_load(i, load_idx); |
1e3c88bd PZ |
5516 | |
5517 | sgs->group_load += load; | |
e44bc5c5 | 5518 | sgs->sum_nr_running += nr_running; |
0ec8aa00 PZ |
5519 | #ifdef CONFIG_NUMA_BALANCING |
5520 | sgs->nr_numa_running += rq->nr_numa_running; | |
5521 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
5522 | #endif | |
1e3c88bd | 5523 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
5524 | if (idle_cpu(i)) |
5525 | sgs->idle_cpus++; | |
1e3c88bd PZ |
5526 | } |
5527 | ||
1e3c88bd | 5528 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
5529 | sgs->group_power = group->sgp->power; |
5530 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 5531 | |
dd5feea1 | 5532 | if (sgs->sum_nr_running) |
38d0f770 | 5533 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5534 | |
aae6d3dd | 5535 | sgs->group_weight = group->group_weight; |
fab47622 | 5536 | |
b37d9316 PZ |
5537 | sgs->group_imb = sg_imbalanced(group); |
5538 | sgs->group_capacity = sg_capacity(env, group); | |
5539 | ||
fab47622 NR |
5540 | if (sgs->group_capacity > sgs->sum_nr_running) |
5541 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5542 | } |
5543 | ||
532cb4c4 MN |
5544 | /** |
5545 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5546 | * @env: The load balancing environment. |
532cb4c4 MN |
5547 | * @sds: sched_domain statistics |
5548 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5549 | * @sgs: sched_group statistics |
532cb4c4 MN |
5550 | * |
5551 | * Determine if @sg is a busier group than the previously selected | |
5552 | * busiest group. | |
e69f6186 YB |
5553 | * |
5554 | * Return: %true if @sg is a busier group than the previously selected | |
5555 | * busiest group. %false otherwise. | |
532cb4c4 | 5556 | */ |
bd939f45 | 5557 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5558 | struct sd_lb_stats *sds, |
5559 | struct sched_group *sg, | |
bd939f45 | 5560 | struct sg_lb_stats *sgs) |
532cb4c4 | 5561 | { |
56cf515b | 5562 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5563 | return false; |
5564 | ||
5565 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5566 | return true; | |
5567 | ||
5568 | if (sgs->group_imb) | |
5569 | return true; | |
5570 | ||
5571 | /* | |
5572 | * ASYM_PACKING needs to move all the work to the lowest | |
5573 | * numbered CPUs in the group, therefore mark all groups | |
5574 | * higher than ourself as busy. | |
5575 | */ | |
bd939f45 PZ |
5576 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5577 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5578 | if (!sds->busiest) |
5579 | return true; | |
5580 | ||
5581 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5582 | return true; | |
5583 | } | |
5584 | ||
5585 | return false; | |
5586 | } | |
5587 | ||
0ec8aa00 PZ |
5588 | #ifdef CONFIG_NUMA_BALANCING |
5589 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5590 | { | |
5591 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
5592 | return regular; | |
5593 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
5594 | return remote; | |
5595 | return all; | |
5596 | } | |
5597 | ||
5598 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5599 | { | |
5600 | if (rq->nr_running > rq->nr_numa_running) | |
5601 | return regular; | |
5602 | if (rq->nr_running > rq->nr_preferred_running) | |
5603 | return remote; | |
5604 | return all; | |
5605 | } | |
5606 | #else | |
5607 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5608 | { | |
5609 | return all; | |
5610 | } | |
5611 | ||
5612 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5613 | { | |
5614 | return regular; | |
5615 | } | |
5616 | #endif /* CONFIG_NUMA_BALANCING */ | |
5617 | ||
1e3c88bd | 5618 | /** |
461819ac | 5619 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5620 | * @env: The load balancing environment. |
1e3c88bd PZ |
5621 | * @balance: Should we balance. |
5622 | * @sds: variable to hold the statistics for this sched_domain. | |
5623 | */ | |
0ec8aa00 | 5624 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5625 | { |
bd939f45 PZ |
5626 | struct sched_domain *child = env->sd->child; |
5627 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5628 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5629 | int load_idx, prefer_sibling = 0; |
5630 | ||
5631 | if (child && child->flags & SD_PREFER_SIBLING) | |
5632 | prefer_sibling = 1; | |
5633 | ||
bd939f45 | 5634 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5635 | |
5636 | do { | |
56cf515b | 5637 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5638 | int local_group; |
5639 | ||
bd939f45 | 5640 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5641 | if (local_group) { |
5642 | sds->local = sg; | |
5643 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5644 | |
5645 | if (env->idle != CPU_NEWLY_IDLE || | |
5646 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5647 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5648 | } |
1e3c88bd | 5649 | |
56cf515b | 5650 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5651 | |
b72ff13c PZ |
5652 | if (local_group) |
5653 | goto next_group; | |
5654 | ||
1e3c88bd PZ |
5655 | /* |
5656 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 5657 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
5658 | * and move all the excess tasks away. We lower the capacity |
5659 | * of a group only if the local group has the capacity to fit | |
5660 | * these excess tasks, i.e. nr_running < group_capacity. The | |
5661 | * extra check prevents the case where you always pull from the | |
5662 | * heaviest group when it is already under-utilized (possible | |
5663 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 5664 | */ |
b72ff13c PZ |
5665 | if (prefer_sibling && sds->local && |
5666 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 5667 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 5668 | |
b72ff13c | 5669 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 5670 | sds->busiest = sg; |
56cf515b | 5671 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
5672 | } |
5673 | ||
b72ff13c PZ |
5674 | next_group: |
5675 | /* Now, start updating sd_lb_stats */ | |
5676 | sds->total_load += sgs->group_load; | |
5677 | sds->total_pwr += sgs->group_power; | |
5678 | ||
532cb4c4 | 5679 | sg = sg->next; |
bd939f45 | 5680 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
5681 | |
5682 | if (env->sd->flags & SD_NUMA) | |
5683 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
532cb4c4 MN |
5684 | } |
5685 | ||
532cb4c4 MN |
5686 | /** |
5687 | * check_asym_packing - Check to see if the group is packed into the | |
5688 | * sched doman. | |
5689 | * | |
5690 | * This is primarily intended to used at the sibling level. Some | |
5691 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
5692 | * case of POWER7, it can move to lower SMT modes only when higher | |
5693 | * threads are idle. When in lower SMT modes, the threads will | |
5694 | * perform better since they share less core resources. Hence when we | |
5695 | * have idle threads, we want them to be the higher ones. | |
5696 | * | |
5697 | * This packing function is run on idle threads. It checks to see if | |
5698 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
5699 | * CPU number than the packing function is being run on. Here we are | |
5700 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
5701 | * number. | |
5702 | * | |
e69f6186 | 5703 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
5704 | * this CPU. The amount of the imbalance is returned in *imbalance. |
5705 | * | |
cd96891d | 5706 | * @env: The load balancing environment. |
532cb4c4 | 5707 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 5708 | */ |
bd939f45 | 5709 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
5710 | { |
5711 | int busiest_cpu; | |
5712 | ||
bd939f45 | 5713 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
5714 | return 0; |
5715 | ||
5716 | if (!sds->busiest) | |
5717 | return 0; | |
5718 | ||
5719 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 5720 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
5721 | return 0; |
5722 | ||
bd939f45 | 5723 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
5724 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
5725 | SCHED_POWER_SCALE); | |
bd939f45 | 5726 | |
532cb4c4 | 5727 | return 1; |
1e3c88bd PZ |
5728 | } |
5729 | ||
5730 | /** | |
5731 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
5732 | * amongst the groups of a sched_domain, during | |
5733 | * load balancing. | |
cd96891d | 5734 | * @env: The load balancing environment. |
1e3c88bd | 5735 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5736 | */ |
bd939f45 PZ |
5737 | static inline |
5738 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
5739 | { |
5740 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
5741 | unsigned int imbn = 2; | |
dd5feea1 | 5742 | unsigned long scaled_busy_load_per_task; |
56cf515b | 5743 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 5744 | |
56cf515b JK |
5745 | local = &sds->local_stat; |
5746 | busiest = &sds->busiest_stat; | |
1e3c88bd | 5747 | |
56cf515b JK |
5748 | if (!local->sum_nr_running) |
5749 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
5750 | else if (busiest->load_per_task > local->load_per_task) | |
5751 | imbn = 1; | |
dd5feea1 | 5752 | |
56cf515b JK |
5753 | scaled_busy_load_per_task = |
5754 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5755 | busiest->group_power; |
56cf515b | 5756 | |
3029ede3 VD |
5757 | if (busiest->avg_load + scaled_busy_load_per_task >= |
5758 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 5759 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5760 | return; |
5761 | } | |
5762 | ||
5763 | /* | |
5764 | * OK, we don't have enough imbalance to justify moving tasks, | |
5765 | * however we may be able to increase total CPU power used by | |
5766 | * moving them. | |
5767 | */ | |
5768 | ||
3ae11c90 | 5769 | pwr_now += busiest->group_power * |
56cf515b | 5770 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 5771 | pwr_now += local->group_power * |
56cf515b | 5772 | min(local->load_per_task, local->avg_load); |
1399fa78 | 5773 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5774 | |
5775 | /* Amount of load we'd subtract */ | |
56cf515b | 5776 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 5777 | busiest->group_power; |
56cf515b | 5778 | if (busiest->avg_load > tmp) { |
3ae11c90 | 5779 | pwr_move += busiest->group_power * |
56cf515b JK |
5780 | min(busiest->load_per_task, |
5781 | busiest->avg_load - tmp); | |
5782 | } | |
1e3c88bd PZ |
5783 | |
5784 | /* Amount of load we'd add */ | |
3ae11c90 | 5785 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 5786 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
5787 | tmp = (busiest->avg_load * busiest->group_power) / |
5788 | local->group_power; | |
56cf515b JK |
5789 | } else { |
5790 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5791 | local->group_power; |
56cf515b | 5792 | } |
3ae11c90 PZ |
5793 | pwr_move += local->group_power * |
5794 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 5795 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5796 | |
5797 | /* Move if we gain throughput */ | |
5798 | if (pwr_move > pwr_now) | |
56cf515b | 5799 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5800 | } |
5801 | ||
5802 | /** | |
5803 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
5804 | * groups of a given sched_domain during load balance. | |
bd939f45 | 5805 | * @env: load balance environment |
1e3c88bd | 5806 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5807 | */ |
bd939f45 | 5808 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5809 | { |
dd5feea1 | 5810 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
5811 | struct sg_lb_stats *local, *busiest; |
5812 | ||
5813 | local = &sds->local_stat; | |
56cf515b | 5814 | busiest = &sds->busiest_stat; |
dd5feea1 | 5815 | |
56cf515b | 5816 | if (busiest->group_imb) { |
30ce5dab PZ |
5817 | /* |
5818 | * In the group_imb case we cannot rely on group-wide averages | |
5819 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
5820 | */ | |
56cf515b JK |
5821 | busiest->load_per_task = |
5822 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
5823 | } |
5824 | ||
1e3c88bd PZ |
5825 | /* |
5826 | * In the presence of smp nice balancing, certain scenarios can have | |
5827 | * max load less than avg load(as we skip the groups at or below | |
5828 | * its cpu_power, while calculating max_load..) | |
5829 | */ | |
b1885550 VD |
5830 | if (busiest->avg_load <= sds->avg_load || |
5831 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
5832 | env->imbalance = 0; |
5833 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
5834 | } |
5835 | ||
56cf515b | 5836 | if (!busiest->group_imb) { |
dd5feea1 SS |
5837 | /* |
5838 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
5839 | * Except of course for the group_imb case, since then we might |
5840 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 5841 | */ |
56cf515b JK |
5842 | load_above_capacity = |
5843 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 5844 | |
1399fa78 | 5845 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 5846 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
5847 | } |
5848 | ||
5849 | /* | |
5850 | * We're trying to get all the cpus to the average_load, so we don't | |
5851 | * want to push ourselves above the average load, nor do we wish to | |
5852 | * reduce the max loaded cpu below the average load. At the same time, | |
5853 | * we also don't want to reduce the group load below the group capacity | |
5854 | * (so that we can implement power-savings policies etc). Thus we look | |
5855 | * for the minimum possible imbalance. | |
dd5feea1 | 5856 | */ |
30ce5dab | 5857 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
5858 | |
5859 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 5860 | env->imbalance = min( |
3ae11c90 PZ |
5861 | max_pull * busiest->group_power, |
5862 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 5863 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
5864 | |
5865 | /* | |
5866 | * if *imbalance is less than the average load per runnable task | |
25985edc | 5867 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
5868 | * a think about bumping its value to force at least one task to be |
5869 | * moved | |
5870 | */ | |
56cf515b | 5871 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 5872 | return fix_small_imbalance(env, sds); |
1e3c88bd | 5873 | } |
fab47622 | 5874 | |
1e3c88bd PZ |
5875 | /******* find_busiest_group() helpers end here *********************/ |
5876 | ||
5877 | /** | |
5878 | * find_busiest_group - Returns the busiest group within the sched_domain | |
5879 | * if there is an imbalance. If there isn't an imbalance, and | |
5880 | * the user has opted for power-savings, it returns a group whose | |
5881 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
5882 | * such a group exists. | |
5883 | * | |
5884 | * Also calculates the amount of weighted load which should be moved | |
5885 | * to restore balance. | |
5886 | * | |
cd96891d | 5887 | * @env: The load balancing environment. |
1e3c88bd | 5888 | * |
e69f6186 | 5889 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
5890 | * - If no imbalance and user has opted for power-savings balance, |
5891 | * return the least loaded group whose CPUs can be | |
5892 | * put to idle by rebalancing its tasks onto our group. | |
5893 | */ | |
56cf515b | 5894 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 5895 | { |
56cf515b | 5896 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
5897 | struct sd_lb_stats sds; |
5898 | ||
147c5fc2 | 5899 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
5900 | |
5901 | /* | |
5902 | * Compute the various statistics relavent for load balancing at | |
5903 | * this level. | |
5904 | */ | |
23f0d209 | 5905 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
5906 | local = &sds.local_stat; |
5907 | busiest = &sds.busiest_stat; | |
1e3c88bd | 5908 | |
bd939f45 PZ |
5909 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
5910 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
5911 | return sds.busiest; |
5912 | ||
cc57aa8f | 5913 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 5914 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
5915 | goto out_balanced; |
5916 | ||
1399fa78 | 5917 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 5918 | |
866ab43e PZ |
5919 | /* |
5920 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 5921 | * work because they assume all things are equal, which typically |
866ab43e PZ |
5922 | * isn't true due to cpus_allowed constraints and the like. |
5923 | */ | |
56cf515b | 5924 | if (busiest->group_imb) |
866ab43e PZ |
5925 | goto force_balance; |
5926 | ||
cc57aa8f | 5927 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
5928 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
5929 | !busiest->group_has_capacity) | |
fab47622 NR |
5930 | goto force_balance; |
5931 | ||
cc57aa8f PZ |
5932 | /* |
5933 | * If the local group is more busy than the selected busiest group | |
5934 | * don't try and pull any tasks. | |
5935 | */ | |
56cf515b | 5936 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
5937 | goto out_balanced; |
5938 | ||
cc57aa8f PZ |
5939 | /* |
5940 | * Don't pull any tasks if this group is already above the domain | |
5941 | * average load. | |
5942 | */ | |
56cf515b | 5943 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
5944 | goto out_balanced; |
5945 | ||
bd939f45 | 5946 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
5947 | /* |
5948 | * This cpu is idle. If the busiest group load doesn't | |
5949 | * have more tasks than the number of available cpu's and | |
5950 | * there is no imbalance between this and busiest group | |
5951 | * wrt to idle cpu's, it is balanced. | |
5952 | */ | |
56cf515b JK |
5953 | if ((local->idle_cpus < busiest->idle_cpus) && |
5954 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 5955 | goto out_balanced; |
c186fafe PZ |
5956 | } else { |
5957 | /* | |
5958 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
5959 | * imbalance_pct to be conservative. | |
5960 | */ | |
56cf515b JK |
5961 | if (100 * busiest->avg_load <= |
5962 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 5963 | goto out_balanced; |
aae6d3dd | 5964 | } |
1e3c88bd | 5965 | |
fab47622 | 5966 | force_balance: |
1e3c88bd | 5967 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 5968 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
5969 | return sds.busiest; |
5970 | ||
5971 | out_balanced: | |
bd939f45 | 5972 | env->imbalance = 0; |
1e3c88bd PZ |
5973 | return NULL; |
5974 | } | |
5975 | ||
5976 | /* | |
5977 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
5978 | */ | |
bd939f45 | 5979 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 5980 | struct sched_group *group) |
1e3c88bd PZ |
5981 | { |
5982 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 5983 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
5984 | int i; |
5985 | ||
6906a408 | 5986 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
0ec8aa00 PZ |
5987 | unsigned long power, capacity, wl; |
5988 | enum fbq_type rt; | |
5989 | ||
5990 | rq = cpu_rq(i); | |
5991 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 5992 | |
0ec8aa00 PZ |
5993 | /* |
5994 | * We classify groups/runqueues into three groups: | |
5995 | * - regular: there are !numa tasks | |
5996 | * - remote: there are numa tasks that run on the 'wrong' node | |
5997 | * - all: there is no distinction | |
5998 | * | |
5999 | * In order to avoid migrating ideally placed numa tasks, | |
6000 | * ignore those when there's better options. | |
6001 | * | |
6002 | * If we ignore the actual busiest queue to migrate another | |
6003 | * task, the next balance pass can still reduce the busiest | |
6004 | * queue by moving tasks around inside the node. | |
6005 | * | |
6006 | * If we cannot move enough load due to this classification | |
6007 | * the next pass will adjust the group classification and | |
6008 | * allow migration of more tasks. | |
6009 | * | |
6010 | * Both cases only affect the total convergence complexity. | |
6011 | */ | |
6012 | if (rt > env->fbq_type) | |
6013 | continue; | |
6014 | ||
6015 | power = power_of(i); | |
6016 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); | |
9d5efe05 | 6017 | if (!capacity) |
bd939f45 | 6018 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 6019 | |
6e40f5bb | 6020 | wl = weighted_cpuload(i); |
1e3c88bd | 6021 | |
6e40f5bb TG |
6022 | /* |
6023 | * When comparing with imbalance, use weighted_cpuload() | |
6024 | * which is not scaled with the cpu power. | |
6025 | */ | |
bd939f45 | 6026 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
6027 | continue; |
6028 | ||
6e40f5bb TG |
6029 | /* |
6030 | * For the load comparisons with the other cpu's, consider | |
6031 | * the weighted_cpuload() scaled with the cpu power, so that | |
6032 | * the load can be moved away from the cpu that is potentially | |
6033 | * running at a lower capacity. | |
95a79b80 JK |
6034 | * |
6035 | * Thus we're looking for max(wl_i / power_i), crosswise | |
6036 | * multiplication to rid ourselves of the division works out | |
6037 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
6038 | * previous maximum. | |
6e40f5bb | 6039 | */ |
95a79b80 JK |
6040 | if (wl * busiest_power > busiest_load * power) { |
6041 | busiest_load = wl; | |
6042 | busiest_power = power; | |
1e3c88bd PZ |
6043 | busiest = rq; |
6044 | } | |
6045 | } | |
6046 | ||
6047 | return busiest; | |
6048 | } | |
6049 | ||
6050 | /* | |
6051 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
6052 | * so long as it is large enough. | |
6053 | */ | |
6054 | #define MAX_PINNED_INTERVAL 512 | |
6055 | ||
6056 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 6057 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 6058 | |
bd939f45 | 6059 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 6060 | { |
bd939f45 PZ |
6061 | struct sched_domain *sd = env->sd; |
6062 | ||
6063 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
6064 | |
6065 | /* | |
6066 | * ASYM_PACKING needs to force migrate tasks from busy but | |
6067 | * higher numbered CPUs in order to pack all tasks in the | |
6068 | * lowest numbered CPUs. | |
6069 | */ | |
bd939f45 | 6070 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 6071 | return 1; |
1af3ed3d PZ |
6072 | } |
6073 | ||
6074 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
6075 | } | |
6076 | ||
969c7921 TH |
6077 | static int active_load_balance_cpu_stop(void *data); |
6078 | ||
23f0d209 JK |
6079 | static int should_we_balance(struct lb_env *env) |
6080 | { | |
6081 | struct sched_group *sg = env->sd->groups; | |
6082 | struct cpumask *sg_cpus, *sg_mask; | |
6083 | int cpu, balance_cpu = -1; | |
6084 | ||
6085 | /* | |
6086 | * In the newly idle case, we will allow all the cpu's | |
6087 | * to do the newly idle load balance. | |
6088 | */ | |
6089 | if (env->idle == CPU_NEWLY_IDLE) | |
6090 | return 1; | |
6091 | ||
6092 | sg_cpus = sched_group_cpus(sg); | |
6093 | sg_mask = sched_group_mask(sg); | |
6094 | /* Try to find first idle cpu */ | |
6095 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
6096 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
6097 | continue; | |
6098 | ||
6099 | balance_cpu = cpu; | |
6100 | break; | |
6101 | } | |
6102 | ||
6103 | if (balance_cpu == -1) | |
6104 | balance_cpu = group_balance_cpu(sg); | |
6105 | ||
6106 | /* | |
6107 | * First idle cpu or the first cpu(busiest) in this sched group | |
6108 | * is eligible for doing load balancing at this and above domains. | |
6109 | */ | |
b0cff9d8 | 6110 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
6111 | } |
6112 | ||
1e3c88bd PZ |
6113 | /* |
6114 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
6115 | * tasks if there is an imbalance. | |
6116 | */ | |
6117 | static int load_balance(int this_cpu, struct rq *this_rq, | |
6118 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 6119 | int *continue_balancing) |
1e3c88bd | 6120 | { |
88b8dac0 | 6121 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 6122 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 6123 | struct sched_group *group; |
1e3c88bd PZ |
6124 | struct rq *busiest; |
6125 | unsigned long flags; | |
e6252c3e | 6126 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 6127 | |
8e45cb54 PZ |
6128 | struct lb_env env = { |
6129 | .sd = sd, | |
ddcdf6e7 PZ |
6130 | .dst_cpu = this_cpu, |
6131 | .dst_rq = this_rq, | |
88b8dac0 | 6132 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 6133 | .idle = idle, |
eb95308e | 6134 | .loop_break = sched_nr_migrate_break, |
b9403130 | 6135 | .cpus = cpus, |
0ec8aa00 | 6136 | .fbq_type = all, |
8e45cb54 PZ |
6137 | }; |
6138 | ||
cfc03118 JK |
6139 | /* |
6140 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
6141 | * other cpus in our group | |
6142 | */ | |
e02e60c1 | 6143 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 6144 | env.dst_grpmask = NULL; |
cfc03118 | 6145 | |
1e3c88bd PZ |
6146 | cpumask_copy(cpus, cpu_active_mask); |
6147 | ||
1e3c88bd PZ |
6148 | schedstat_inc(sd, lb_count[idle]); |
6149 | ||
6150 | redo: | |
23f0d209 JK |
6151 | if (!should_we_balance(&env)) { |
6152 | *continue_balancing = 0; | |
1e3c88bd | 6153 | goto out_balanced; |
23f0d209 | 6154 | } |
1e3c88bd | 6155 | |
23f0d209 | 6156 | group = find_busiest_group(&env); |
1e3c88bd PZ |
6157 | if (!group) { |
6158 | schedstat_inc(sd, lb_nobusyg[idle]); | |
6159 | goto out_balanced; | |
6160 | } | |
6161 | ||
b9403130 | 6162 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
6163 | if (!busiest) { |
6164 | schedstat_inc(sd, lb_nobusyq[idle]); | |
6165 | goto out_balanced; | |
6166 | } | |
6167 | ||
78feefc5 | 6168 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 6169 | |
bd939f45 | 6170 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
6171 | |
6172 | ld_moved = 0; | |
6173 | if (busiest->nr_running > 1) { | |
6174 | /* | |
6175 | * Attempt to move tasks. If find_busiest_group has found | |
6176 | * an imbalance but busiest->nr_running <= 1, the group is | |
6177 | * still unbalanced. ld_moved simply stays zero, so it is | |
6178 | * correctly treated as an imbalance. | |
6179 | */ | |
8e45cb54 | 6180 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
6181 | env.src_cpu = busiest->cpu; |
6182 | env.src_rq = busiest; | |
6183 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 6184 | |
5d6523eb | 6185 | more_balance: |
1e3c88bd | 6186 | local_irq_save(flags); |
78feefc5 | 6187 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
6188 | |
6189 | /* | |
6190 | * cur_ld_moved - load moved in current iteration | |
6191 | * ld_moved - cumulative load moved across iterations | |
6192 | */ | |
6193 | cur_ld_moved = move_tasks(&env); | |
6194 | ld_moved += cur_ld_moved; | |
78feefc5 | 6195 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
6196 | local_irq_restore(flags); |
6197 | ||
6198 | /* | |
6199 | * some other cpu did the load balance for us. | |
6200 | */ | |
88b8dac0 SV |
6201 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
6202 | resched_cpu(env.dst_cpu); | |
6203 | ||
f1cd0858 JK |
6204 | if (env.flags & LBF_NEED_BREAK) { |
6205 | env.flags &= ~LBF_NEED_BREAK; | |
6206 | goto more_balance; | |
6207 | } | |
6208 | ||
88b8dac0 SV |
6209 | /* |
6210 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
6211 | * us and move them to an alternate dst_cpu in our sched_group | |
6212 | * where they can run. The upper limit on how many times we | |
6213 | * iterate on same src_cpu is dependent on number of cpus in our | |
6214 | * sched_group. | |
6215 | * | |
6216 | * This changes load balance semantics a bit on who can move | |
6217 | * load to a given_cpu. In addition to the given_cpu itself | |
6218 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
6219 | * nohz-idle), we now have balance_cpu in a position to move | |
6220 | * load to given_cpu. In rare situations, this may cause | |
6221 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
6222 | * _independently_ and at _same_ time to move some load to | |
6223 | * given_cpu) causing exceess load to be moved to given_cpu. | |
6224 | * This however should not happen so much in practice and | |
6225 | * moreover subsequent load balance cycles should correct the | |
6226 | * excess load moved. | |
6227 | */ | |
6263322c | 6228 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 6229 | |
7aff2e3a VD |
6230 | /* Prevent to re-select dst_cpu via env's cpus */ |
6231 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
6232 | ||
78feefc5 | 6233 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 6234 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 6235 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
6236 | env.loop = 0; |
6237 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 6238 | |
88b8dac0 SV |
6239 | /* |
6240 | * Go back to "more_balance" rather than "redo" since we | |
6241 | * need to continue with same src_cpu. | |
6242 | */ | |
6243 | goto more_balance; | |
6244 | } | |
1e3c88bd | 6245 | |
6263322c PZ |
6246 | /* |
6247 | * We failed to reach balance because of affinity. | |
6248 | */ | |
6249 | if (sd_parent) { | |
6250 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
6251 | ||
6252 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
6253 | *group_imbalance = 1; | |
6254 | } else if (*group_imbalance) | |
6255 | *group_imbalance = 0; | |
6256 | } | |
6257 | ||
1e3c88bd | 6258 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 6259 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 6260 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
6261 | if (!cpumask_empty(cpus)) { |
6262 | env.loop = 0; | |
6263 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 6264 | goto redo; |
bbf18b19 | 6265 | } |
1e3c88bd PZ |
6266 | goto out_balanced; |
6267 | } | |
6268 | } | |
6269 | ||
6270 | if (!ld_moved) { | |
6271 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
6272 | /* |
6273 | * Increment the failure counter only on periodic balance. | |
6274 | * We do not want newidle balance, which can be very | |
6275 | * frequent, pollute the failure counter causing | |
6276 | * excessive cache_hot migrations and active balances. | |
6277 | */ | |
6278 | if (idle != CPU_NEWLY_IDLE) | |
6279 | sd->nr_balance_failed++; | |
1e3c88bd | 6280 | |
bd939f45 | 6281 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
6282 | raw_spin_lock_irqsave(&busiest->lock, flags); |
6283 | ||
969c7921 TH |
6284 | /* don't kick the active_load_balance_cpu_stop, |
6285 | * if the curr task on busiest cpu can't be | |
6286 | * moved to this_cpu | |
1e3c88bd PZ |
6287 | */ |
6288 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 6289 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
6290 | raw_spin_unlock_irqrestore(&busiest->lock, |
6291 | flags); | |
8e45cb54 | 6292 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
6293 | goto out_one_pinned; |
6294 | } | |
6295 | ||
969c7921 TH |
6296 | /* |
6297 | * ->active_balance synchronizes accesses to | |
6298 | * ->active_balance_work. Once set, it's cleared | |
6299 | * only after active load balance is finished. | |
6300 | */ | |
1e3c88bd PZ |
6301 | if (!busiest->active_balance) { |
6302 | busiest->active_balance = 1; | |
6303 | busiest->push_cpu = this_cpu; | |
6304 | active_balance = 1; | |
6305 | } | |
6306 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 6307 | |
bd939f45 | 6308 | if (active_balance) { |
969c7921 TH |
6309 | stop_one_cpu_nowait(cpu_of(busiest), |
6310 | active_load_balance_cpu_stop, busiest, | |
6311 | &busiest->active_balance_work); | |
bd939f45 | 6312 | } |
1e3c88bd PZ |
6313 | |
6314 | /* | |
6315 | * We've kicked active balancing, reset the failure | |
6316 | * counter. | |
6317 | */ | |
6318 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
6319 | } | |
6320 | } else | |
6321 | sd->nr_balance_failed = 0; | |
6322 | ||
6323 | if (likely(!active_balance)) { | |
6324 | /* We were unbalanced, so reset the balancing interval */ | |
6325 | sd->balance_interval = sd->min_interval; | |
6326 | } else { | |
6327 | /* | |
6328 | * If we've begun active balancing, start to back off. This | |
6329 | * case may not be covered by the all_pinned logic if there | |
6330 | * is only 1 task on the busy runqueue (because we don't call | |
6331 | * move_tasks). | |
6332 | */ | |
6333 | if (sd->balance_interval < sd->max_interval) | |
6334 | sd->balance_interval *= 2; | |
6335 | } | |
6336 | ||
1e3c88bd PZ |
6337 | goto out; |
6338 | ||
6339 | out_balanced: | |
6340 | schedstat_inc(sd, lb_balanced[idle]); | |
6341 | ||
6342 | sd->nr_balance_failed = 0; | |
6343 | ||
6344 | out_one_pinned: | |
6345 | /* tune up the balancing interval */ | |
8e45cb54 | 6346 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 6347 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
6348 | (sd->balance_interval < sd->max_interval)) |
6349 | sd->balance_interval *= 2; | |
6350 | ||
46e49b38 | 6351 | ld_moved = 0; |
1e3c88bd | 6352 | out: |
1e3c88bd PZ |
6353 | return ld_moved; |
6354 | } | |
6355 | ||
1e3c88bd PZ |
6356 | /* |
6357 | * idle_balance is called by schedule() if this_cpu is about to become | |
6358 | * idle. Attempts to pull tasks from other CPUs. | |
6359 | */ | |
029632fb | 6360 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
6361 | { |
6362 | struct sched_domain *sd; | |
6363 | int pulled_task = 0; | |
6364 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 6365 | u64 curr_cost = 0; |
1e3c88bd | 6366 | |
78becc27 | 6367 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
6368 | |
6369 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
6370 | return; | |
6371 | ||
f492e12e PZ |
6372 | /* |
6373 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
6374 | */ | |
6375 | raw_spin_unlock(&this_rq->lock); | |
6376 | ||
48a16753 | 6377 | update_blocked_averages(this_cpu); |
dce840a0 | 6378 | rcu_read_lock(); |
1e3c88bd PZ |
6379 | for_each_domain(this_cpu, sd) { |
6380 | unsigned long interval; | |
23f0d209 | 6381 | int continue_balancing = 1; |
9bd721c5 | 6382 | u64 t0, domain_cost; |
1e3c88bd PZ |
6383 | |
6384 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
6385 | continue; | |
6386 | ||
9bd721c5 JL |
6387 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
6388 | break; | |
6389 | ||
f492e12e | 6390 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
6391 | t0 = sched_clock_cpu(this_cpu); |
6392 | ||
1e3c88bd | 6393 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 6394 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
6395 | sd, CPU_NEWLY_IDLE, |
6396 | &continue_balancing); | |
9bd721c5 JL |
6397 | |
6398 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
6399 | if (domain_cost > sd->max_newidle_lb_cost) | |
6400 | sd->max_newidle_lb_cost = domain_cost; | |
6401 | ||
6402 | curr_cost += domain_cost; | |
f492e12e | 6403 | } |
1e3c88bd PZ |
6404 | |
6405 | interval = msecs_to_jiffies(sd->balance_interval); | |
6406 | if (time_after(next_balance, sd->last_balance + interval)) | |
6407 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
6408 | if (pulled_task) { |
6409 | this_rq->idle_stamp = 0; | |
1e3c88bd | 6410 | break; |
d5ad140b | 6411 | } |
1e3c88bd | 6412 | } |
dce840a0 | 6413 | rcu_read_unlock(); |
f492e12e PZ |
6414 | |
6415 | raw_spin_lock(&this_rq->lock); | |
6416 | ||
1e3c88bd PZ |
6417 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
6418 | /* | |
6419 | * We are going idle. next_balance may be set based on | |
6420 | * a busy processor. So reset next_balance. | |
6421 | */ | |
6422 | this_rq->next_balance = next_balance; | |
6423 | } | |
9bd721c5 JL |
6424 | |
6425 | if (curr_cost > this_rq->max_idle_balance_cost) | |
6426 | this_rq->max_idle_balance_cost = curr_cost; | |
1e3c88bd PZ |
6427 | } |
6428 | ||
6429 | /* | |
969c7921 TH |
6430 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
6431 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
6432 | * least 1 task to be running on each physical CPU where possible, and | |
6433 | * avoids physical / logical imbalances. | |
1e3c88bd | 6434 | */ |
969c7921 | 6435 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 6436 | { |
969c7921 TH |
6437 | struct rq *busiest_rq = data; |
6438 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 6439 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 6440 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 6441 | struct sched_domain *sd; |
969c7921 TH |
6442 | |
6443 | raw_spin_lock_irq(&busiest_rq->lock); | |
6444 | ||
6445 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
6446 | if (unlikely(busiest_cpu != smp_processor_id() || | |
6447 | !busiest_rq->active_balance)) | |
6448 | goto out_unlock; | |
1e3c88bd PZ |
6449 | |
6450 | /* Is there any task to move? */ | |
6451 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 6452 | goto out_unlock; |
1e3c88bd PZ |
6453 | |
6454 | /* | |
6455 | * This condition is "impossible", if it occurs | |
6456 | * we need to fix it. Originally reported by | |
6457 | * Bjorn Helgaas on a 128-cpu setup. | |
6458 | */ | |
6459 | BUG_ON(busiest_rq == target_rq); | |
6460 | ||
6461 | /* move a task from busiest_rq to target_rq */ | |
6462 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
6463 | |
6464 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 6465 | rcu_read_lock(); |
1e3c88bd PZ |
6466 | for_each_domain(target_cpu, sd) { |
6467 | if ((sd->flags & SD_LOAD_BALANCE) && | |
6468 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
6469 | break; | |
6470 | } | |
6471 | ||
6472 | if (likely(sd)) { | |
8e45cb54 PZ |
6473 | struct lb_env env = { |
6474 | .sd = sd, | |
ddcdf6e7 PZ |
6475 | .dst_cpu = target_cpu, |
6476 | .dst_rq = target_rq, | |
6477 | .src_cpu = busiest_rq->cpu, | |
6478 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
6479 | .idle = CPU_IDLE, |
6480 | }; | |
6481 | ||
1e3c88bd PZ |
6482 | schedstat_inc(sd, alb_count); |
6483 | ||
8e45cb54 | 6484 | if (move_one_task(&env)) |
1e3c88bd PZ |
6485 | schedstat_inc(sd, alb_pushed); |
6486 | else | |
6487 | schedstat_inc(sd, alb_failed); | |
6488 | } | |
dce840a0 | 6489 | rcu_read_unlock(); |
1e3c88bd | 6490 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
6491 | out_unlock: |
6492 | busiest_rq->active_balance = 0; | |
6493 | raw_spin_unlock_irq(&busiest_rq->lock); | |
6494 | return 0; | |
1e3c88bd PZ |
6495 | } |
6496 | ||
3451d024 | 6497 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
6498 | /* |
6499 | * idle load balancing details | |
83cd4fe2 VP |
6500 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
6501 | * needed, they will kick the idle load balancer, which then does idle | |
6502 | * load balancing for all the idle CPUs. | |
6503 | */ | |
1e3c88bd | 6504 | static struct { |
83cd4fe2 | 6505 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 6506 | atomic_t nr_cpus; |
83cd4fe2 VP |
6507 | unsigned long next_balance; /* in jiffy units */ |
6508 | } nohz ____cacheline_aligned; | |
1e3c88bd | 6509 | |
8e7fbcbc | 6510 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 6511 | { |
0b005cf5 | 6512 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 6513 | |
786d6dc7 SS |
6514 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
6515 | return ilb; | |
6516 | ||
6517 | return nr_cpu_ids; | |
1e3c88bd | 6518 | } |
1e3c88bd | 6519 | |
83cd4fe2 VP |
6520 | /* |
6521 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
6522 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
6523 | * CPU (if there is one). | |
6524 | */ | |
6525 | static void nohz_balancer_kick(int cpu) | |
6526 | { | |
6527 | int ilb_cpu; | |
6528 | ||
6529 | nohz.next_balance++; | |
6530 | ||
0b005cf5 | 6531 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 6532 | |
0b005cf5 SS |
6533 | if (ilb_cpu >= nr_cpu_ids) |
6534 | return; | |
83cd4fe2 | 6535 | |
cd490c5b | 6536 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
6537 | return; |
6538 | /* | |
6539 | * Use smp_send_reschedule() instead of resched_cpu(). | |
6540 | * This way we generate a sched IPI on the target cpu which | |
6541 | * is idle. And the softirq performing nohz idle load balance | |
6542 | * will be run before returning from the IPI. | |
6543 | */ | |
6544 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
6545 | return; |
6546 | } | |
6547 | ||
c1cc017c | 6548 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
6549 | { |
6550 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
6551 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
6552 | atomic_dec(&nohz.nr_cpus); | |
6553 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
6554 | } | |
6555 | } | |
6556 | ||
69e1e811 SS |
6557 | static inline void set_cpu_sd_state_busy(void) |
6558 | { | |
6559 | struct sched_domain *sd; | |
69e1e811 | 6560 | |
69e1e811 | 6561 | rcu_read_lock(); |
424c93fe | 6562 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
6563 | |
6564 | if (!sd || !sd->nohz_idle) | |
6565 | goto unlock; | |
6566 | sd->nohz_idle = 0; | |
6567 | ||
6568 | for (; sd; sd = sd->parent) | |
69e1e811 | 6569 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6570 | unlock: |
69e1e811 SS |
6571 | rcu_read_unlock(); |
6572 | } | |
6573 | ||
6574 | void set_cpu_sd_state_idle(void) | |
6575 | { | |
6576 | struct sched_domain *sd; | |
69e1e811 | 6577 | |
69e1e811 | 6578 | rcu_read_lock(); |
424c93fe | 6579 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
6580 | |
6581 | if (!sd || sd->nohz_idle) | |
6582 | goto unlock; | |
6583 | sd->nohz_idle = 1; | |
6584 | ||
6585 | for (; sd; sd = sd->parent) | |
69e1e811 | 6586 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6587 | unlock: |
69e1e811 SS |
6588 | rcu_read_unlock(); |
6589 | } | |
6590 | ||
1e3c88bd | 6591 | /* |
c1cc017c | 6592 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 6593 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 6594 | */ |
c1cc017c | 6595 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 6596 | { |
71325960 SS |
6597 | /* |
6598 | * If this cpu is going down, then nothing needs to be done. | |
6599 | */ | |
6600 | if (!cpu_active(cpu)) | |
6601 | return; | |
6602 | ||
c1cc017c AS |
6603 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
6604 | return; | |
1e3c88bd | 6605 | |
c1cc017c AS |
6606 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
6607 | atomic_inc(&nohz.nr_cpus); | |
6608 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 6609 | } |
71325960 | 6610 | |
0db0628d | 6611 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
6612 | unsigned long action, void *hcpu) |
6613 | { | |
6614 | switch (action & ~CPU_TASKS_FROZEN) { | |
6615 | case CPU_DYING: | |
c1cc017c | 6616 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
6617 | return NOTIFY_OK; |
6618 | default: | |
6619 | return NOTIFY_DONE; | |
6620 | } | |
6621 | } | |
1e3c88bd PZ |
6622 | #endif |
6623 | ||
6624 | static DEFINE_SPINLOCK(balancing); | |
6625 | ||
49c022e6 PZ |
6626 | /* |
6627 | * Scale the max load_balance interval with the number of CPUs in the system. | |
6628 | * This trades load-balance latency on larger machines for less cross talk. | |
6629 | */ | |
029632fb | 6630 | void update_max_interval(void) |
49c022e6 PZ |
6631 | { |
6632 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
6633 | } | |
6634 | ||
1e3c88bd PZ |
6635 | /* |
6636 | * It checks each scheduling domain to see if it is due to be balanced, | |
6637 | * and initiates a balancing operation if so. | |
6638 | * | |
b9b0853a | 6639 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd PZ |
6640 | */ |
6641 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
6642 | { | |
23f0d209 | 6643 | int continue_balancing = 1; |
1e3c88bd PZ |
6644 | struct rq *rq = cpu_rq(cpu); |
6645 | unsigned long interval; | |
04f733b4 | 6646 | struct sched_domain *sd; |
1e3c88bd PZ |
6647 | /* Earliest time when we have to do rebalance again */ |
6648 | unsigned long next_balance = jiffies + 60*HZ; | |
6649 | int update_next_balance = 0; | |
f48627e6 JL |
6650 | int need_serialize, need_decay = 0; |
6651 | u64 max_cost = 0; | |
1e3c88bd | 6652 | |
48a16753 | 6653 | update_blocked_averages(cpu); |
2069dd75 | 6654 | |
dce840a0 | 6655 | rcu_read_lock(); |
1e3c88bd | 6656 | for_each_domain(cpu, sd) { |
f48627e6 JL |
6657 | /* |
6658 | * Decay the newidle max times here because this is a regular | |
6659 | * visit to all the domains. Decay ~1% per second. | |
6660 | */ | |
6661 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
6662 | sd->max_newidle_lb_cost = | |
6663 | (sd->max_newidle_lb_cost * 253) / 256; | |
6664 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
6665 | need_decay = 1; | |
6666 | } | |
6667 | max_cost += sd->max_newidle_lb_cost; | |
6668 | ||
1e3c88bd PZ |
6669 | if (!(sd->flags & SD_LOAD_BALANCE)) |
6670 | continue; | |
6671 | ||
f48627e6 JL |
6672 | /* |
6673 | * Stop the load balance at this level. There is another | |
6674 | * CPU in our sched group which is doing load balancing more | |
6675 | * actively. | |
6676 | */ | |
6677 | if (!continue_balancing) { | |
6678 | if (need_decay) | |
6679 | continue; | |
6680 | break; | |
6681 | } | |
6682 | ||
1e3c88bd PZ |
6683 | interval = sd->balance_interval; |
6684 | if (idle != CPU_IDLE) | |
6685 | interval *= sd->busy_factor; | |
6686 | ||
6687 | /* scale ms to jiffies */ | |
6688 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 6689 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
6690 | |
6691 | need_serialize = sd->flags & SD_SERIALIZE; | |
6692 | ||
6693 | if (need_serialize) { | |
6694 | if (!spin_trylock(&balancing)) | |
6695 | goto out; | |
6696 | } | |
6697 | ||
6698 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 6699 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 6700 | /* |
6263322c | 6701 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
6702 | * env->dst_cpu, so we can't know our idle |
6703 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 6704 | */ |
de5eb2dd | 6705 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
6706 | } |
6707 | sd->last_balance = jiffies; | |
6708 | } | |
6709 | if (need_serialize) | |
6710 | spin_unlock(&balancing); | |
6711 | out: | |
6712 | if (time_after(next_balance, sd->last_balance + interval)) { | |
6713 | next_balance = sd->last_balance + interval; | |
6714 | update_next_balance = 1; | |
6715 | } | |
f48627e6 JL |
6716 | } |
6717 | if (need_decay) { | |
1e3c88bd | 6718 | /* |
f48627e6 JL |
6719 | * Ensure the rq-wide value also decays but keep it at a |
6720 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 6721 | */ |
f48627e6 JL |
6722 | rq->max_idle_balance_cost = |
6723 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 6724 | } |
dce840a0 | 6725 | rcu_read_unlock(); |
1e3c88bd PZ |
6726 | |
6727 | /* | |
6728 | * next_balance will be updated only when there is a need. | |
6729 | * When the cpu is attached to null domain for ex, it will not be | |
6730 | * updated. | |
6731 | */ | |
6732 | if (likely(update_next_balance)) | |
6733 | rq->next_balance = next_balance; | |
6734 | } | |
6735 | ||
3451d024 | 6736 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 6737 | /* |
3451d024 | 6738 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
6739 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
6740 | */ | |
83cd4fe2 VP |
6741 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
6742 | { | |
6743 | struct rq *this_rq = cpu_rq(this_cpu); | |
6744 | struct rq *rq; | |
6745 | int balance_cpu; | |
6746 | ||
1c792db7 SS |
6747 | if (idle != CPU_IDLE || |
6748 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
6749 | goto end; | |
83cd4fe2 VP |
6750 | |
6751 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 6752 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
6753 | continue; |
6754 | ||
6755 | /* | |
6756 | * If this cpu gets work to do, stop the load balancing | |
6757 | * work being done for other cpus. Next load | |
6758 | * balancing owner will pick it up. | |
6759 | */ | |
1c792db7 | 6760 | if (need_resched()) |
83cd4fe2 | 6761 | break; |
83cd4fe2 | 6762 | |
5ed4f1d9 VG |
6763 | rq = cpu_rq(balance_cpu); |
6764 | ||
6765 | raw_spin_lock_irq(&rq->lock); | |
6766 | update_rq_clock(rq); | |
6767 | update_idle_cpu_load(rq); | |
6768 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
6769 | |
6770 | rebalance_domains(balance_cpu, CPU_IDLE); | |
6771 | ||
83cd4fe2 VP |
6772 | if (time_after(this_rq->next_balance, rq->next_balance)) |
6773 | this_rq->next_balance = rq->next_balance; | |
6774 | } | |
6775 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
6776 | end: |
6777 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
6778 | } |
6779 | ||
6780 | /* | |
0b005cf5 SS |
6781 | * Current heuristic for kicking the idle load balancer in the presence |
6782 | * of an idle cpu is the system. | |
6783 | * - This rq has more than one task. | |
6784 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
6785 | * busy cpu's exceeding the group's power. | |
6786 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
6787 | * domain span are idle. | |
83cd4fe2 VP |
6788 | */ |
6789 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
6790 | { | |
6791 | unsigned long now = jiffies; | |
0b005cf5 | 6792 | struct sched_domain *sd; |
83cd4fe2 | 6793 | |
1c792db7 | 6794 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
6795 | return 0; |
6796 | ||
1c792db7 SS |
6797 | /* |
6798 | * We may be recently in ticked or tickless idle mode. At the first | |
6799 | * busy tick after returning from idle, we will update the busy stats. | |
6800 | */ | |
69e1e811 | 6801 | set_cpu_sd_state_busy(); |
c1cc017c | 6802 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
6803 | |
6804 | /* | |
6805 | * None are in tickless mode and hence no need for NOHZ idle load | |
6806 | * balancing. | |
6807 | */ | |
6808 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
6809 | return 0; | |
1c792db7 SS |
6810 | |
6811 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
6812 | return 0; |
6813 | ||
0b005cf5 SS |
6814 | if (rq->nr_running >= 2) |
6815 | goto need_kick; | |
83cd4fe2 | 6816 | |
067491b7 | 6817 | rcu_read_lock(); |
0b005cf5 SS |
6818 | for_each_domain(cpu, sd) { |
6819 | struct sched_group *sg = sd->groups; | |
6820 | struct sched_group_power *sgp = sg->sgp; | |
6821 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 6822 | |
0b005cf5 | 6823 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 6824 | goto need_kick_unlock; |
0b005cf5 SS |
6825 | |
6826 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
6827 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
6828 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 6829 | goto need_kick_unlock; |
0b005cf5 SS |
6830 | |
6831 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
6832 | break; | |
83cd4fe2 | 6833 | } |
067491b7 | 6834 | rcu_read_unlock(); |
83cd4fe2 | 6835 | return 0; |
067491b7 PZ |
6836 | |
6837 | need_kick_unlock: | |
6838 | rcu_read_unlock(); | |
0b005cf5 SS |
6839 | need_kick: |
6840 | return 1; | |
83cd4fe2 VP |
6841 | } |
6842 | #else | |
6843 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
6844 | #endif | |
6845 | ||
6846 | /* | |
6847 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
6848 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
6849 | */ | |
1e3c88bd PZ |
6850 | static void run_rebalance_domains(struct softirq_action *h) |
6851 | { | |
6852 | int this_cpu = smp_processor_id(); | |
6853 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 6854 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
6855 | CPU_IDLE : CPU_NOT_IDLE; |
6856 | ||
6857 | rebalance_domains(this_cpu, idle); | |
6858 | ||
1e3c88bd | 6859 | /* |
83cd4fe2 | 6860 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
6861 | * balancing on behalf of the other idle cpus whose ticks are |
6862 | * stopped. | |
6863 | */ | |
83cd4fe2 | 6864 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
6865 | } |
6866 | ||
6867 | static inline int on_null_domain(int cpu) | |
6868 | { | |
90a6501f | 6869 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
6870 | } |
6871 | ||
6872 | /* | |
6873 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 6874 | */ |
029632fb | 6875 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 6876 | { |
1e3c88bd PZ |
6877 | /* Don't need to rebalance while attached to NULL domain */ |
6878 | if (time_after_eq(jiffies, rq->next_balance) && | |
6879 | likely(!on_null_domain(cpu))) | |
6880 | raise_softirq(SCHED_SOFTIRQ); | |
3451d024 | 6881 | #ifdef CONFIG_NO_HZ_COMMON |
1c792db7 | 6882 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
6883 | nohz_balancer_kick(cpu); |
6884 | #endif | |
1e3c88bd PZ |
6885 | } |
6886 | ||
0bcdcf28 CE |
6887 | static void rq_online_fair(struct rq *rq) |
6888 | { | |
6889 | update_sysctl(); | |
6890 | } | |
6891 | ||
6892 | static void rq_offline_fair(struct rq *rq) | |
6893 | { | |
6894 | update_sysctl(); | |
a4c96ae3 PB |
6895 | |
6896 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
6897 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
6898 | } |
6899 | ||
55e12e5e | 6900 | #endif /* CONFIG_SMP */ |
e1d1484f | 6901 | |
bf0f6f24 IM |
6902 | /* |
6903 | * scheduler tick hitting a task of our scheduling class: | |
6904 | */ | |
8f4d37ec | 6905 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
6906 | { |
6907 | struct cfs_rq *cfs_rq; | |
6908 | struct sched_entity *se = &curr->se; | |
6909 | ||
6910 | for_each_sched_entity(se) { | |
6911 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 6912 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 6913 | } |
18bf2805 | 6914 | |
10e84b97 | 6915 | if (numabalancing_enabled) |
cbee9f88 | 6916 | task_tick_numa(rq, curr); |
3d59eebc | 6917 | |
18bf2805 | 6918 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
6919 | } |
6920 | ||
6921 | /* | |
cd29fe6f PZ |
6922 | * called on fork with the child task as argument from the parent's context |
6923 | * - child not yet on the tasklist | |
6924 | * - preemption disabled | |
bf0f6f24 | 6925 | */ |
cd29fe6f | 6926 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 6927 | { |
4fc420c9 DN |
6928 | struct cfs_rq *cfs_rq; |
6929 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 6930 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
6931 | struct rq *rq = this_rq(); |
6932 | unsigned long flags; | |
6933 | ||
05fa785c | 6934 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 6935 | |
861d034e PZ |
6936 | update_rq_clock(rq); |
6937 | ||
4fc420c9 DN |
6938 | cfs_rq = task_cfs_rq(current); |
6939 | curr = cfs_rq->curr; | |
6940 | ||
6c9a27f5 DN |
6941 | /* |
6942 | * Not only the cpu but also the task_group of the parent might have | |
6943 | * been changed after parent->se.parent,cfs_rq were copied to | |
6944 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
6945 | * of child point to valid ones. | |
6946 | */ | |
6947 | rcu_read_lock(); | |
6948 | __set_task_cpu(p, this_cpu); | |
6949 | rcu_read_unlock(); | |
bf0f6f24 | 6950 | |
7109c442 | 6951 | update_curr(cfs_rq); |
cd29fe6f | 6952 | |
b5d9d734 MG |
6953 | if (curr) |
6954 | se->vruntime = curr->vruntime; | |
aeb73b04 | 6955 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 6956 | |
cd29fe6f | 6957 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 6958 | /* |
edcb60a3 IM |
6959 | * Upon rescheduling, sched_class::put_prev_task() will place |
6960 | * 'current' within the tree based on its new key value. | |
6961 | */ | |
4d78e7b6 | 6962 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 6963 | resched_task(rq->curr); |
4d78e7b6 | 6964 | } |
bf0f6f24 | 6965 | |
88ec22d3 PZ |
6966 | se->vruntime -= cfs_rq->min_vruntime; |
6967 | ||
05fa785c | 6968 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
6969 | } |
6970 | ||
cb469845 SR |
6971 | /* |
6972 | * Priority of the task has changed. Check to see if we preempt | |
6973 | * the current task. | |
6974 | */ | |
da7a735e PZ |
6975 | static void |
6976 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 6977 | { |
da7a735e PZ |
6978 | if (!p->se.on_rq) |
6979 | return; | |
6980 | ||
cb469845 SR |
6981 | /* |
6982 | * Reschedule if we are currently running on this runqueue and | |
6983 | * our priority decreased, or if we are not currently running on | |
6984 | * this runqueue and our priority is higher than the current's | |
6985 | */ | |
da7a735e | 6986 | if (rq->curr == p) { |
cb469845 SR |
6987 | if (p->prio > oldprio) |
6988 | resched_task(rq->curr); | |
6989 | } else | |
15afe09b | 6990 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6991 | } |
6992 | ||
da7a735e PZ |
6993 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
6994 | { | |
6995 | struct sched_entity *se = &p->se; | |
6996 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6997 | ||
6998 | /* | |
6999 | * Ensure the task's vruntime is normalized, so that when its | |
7000 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
7001 | * do the right thing. | |
7002 | * | |
7003 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
7004 | * have normalized the vruntime, if it was !on_rq, then only when | |
7005 | * the task is sleeping will it still have non-normalized vruntime. | |
7006 | */ | |
7007 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
7008 | /* | |
7009 | * Fix up our vruntime so that the current sleep doesn't | |
7010 | * cause 'unlimited' sleep bonus. | |
7011 | */ | |
7012 | place_entity(cfs_rq, se, 0); | |
7013 | se->vruntime -= cfs_rq->min_vruntime; | |
7014 | } | |
9ee474f5 | 7015 | |
141965c7 | 7016 | #ifdef CONFIG_SMP |
9ee474f5 PT |
7017 | /* |
7018 | * Remove our load from contribution when we leave sched_fair | |
7019 | * and ensure we don't carry in an old decay_count if we | |
7020 | * switch back. | |
7021 | */ | |
87e3c8ae KT |
7022 | if (se->avg.decay_count) { |
7023 | __synchronize_entity_decay(se); | |
7024 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
7025 | } |
7026 | #endif | |
da7a735e PZ |
7027 | } |
7028 | ||
cb469845 SR |
7029 | /* |
7030 | * We switched to the sched_fair class. | |
7031 | */ | |
da7a735e | 7032 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 7033 | { |
da7a735e PZ |
7034 | if (!p->se.on_rq) |
7035 | return; | |
7036 | ||
cb469845 SR |
7037 | /* |
7038 | * We were most likely switched from sched_rt, so | |
7039 | * kick off the schedule if running, otherwise just see | |
7040 | * if we can still preempt the current task. | |
7041 | */ | |
da7a735e | 7042 | if (rq->curr == p) |
cb469845 SR |
7043 | resched_task(rq->curr); |
7044 | else | |
15afe09b | 7045 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7046 | } |
7047 | ||
83b699ed SV |
7048 | /* Account for a task changing its policy or group. |
7049 | * | |
7050 | * This routine is mostly called to set cfs_rq->curr field when a task | |
7051 | * migrates between groups/classes. | |
7052 | */ | |
7053 | static void set_curr_task_fair(struct rq *rq) | |
7054 | { | |
7055 | struct sched_entity *se = &rq->curr->se; | |
7056 | ||
ec12cb7f PT |
7057 | for_each_sched_entity(se) { |
7058 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7059 | ||
7060 | set_next_entity(cfs_rq, se); | |
7061 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
7062 | account_cfs_rq_runtime(cfs_rq, 0); | |
7063 | } | |
83b699ed SV |
7064 | } |
7065 | ||
029632fb PZ |
7066 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
7067 | { | |
7068 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
7069 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
7070 | #ifndef CONFIG_64BIT | |
7071 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
7072 | #endif | |
141965c7 | 7073 | #ifdef CONFIG_SMP |
9ee474f5 | 7074 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 7075 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 7076 | #endif |
029632fb PZ |
7077 | } |
7078 | ||
810b3817 | 7079 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7080 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 7081 | { |
aff3e498 | 7082 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
7083 | /* |
7084 | * If the task was not on the rq at the time of this cgroup movement | |
7085 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
7086 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
7087 | * bonus in place_entity()). | |
7088 | * | |
7089 | * If it was on the rq, we've just 'preempted' it, which does convert | |
7090 | * ->vruntime to a relative base. | |
7091 | * | |
7092 | * Make sure both cases convert their relative position when migrating | |
7093 | * to another cgroup's rq. This does somewhat interfere with the | |
7094 | * fair sleeper stuff for the first placement, but who cares. | |
7095 | */ | |
7ceff013 DN |
7096 | /* |
7097 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
7098 | * But there are some cases where it has already been normalized: | |
7099 | * | |
7100 | * - Moving a forked child which is waiting for being woken up by | |
7101 | * wake_up_new_task(). | |
62af3783 DN |
7102 | * - Moving a task which has been woken up by try_to_wake_up() and |
7103 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
7104 | * |
7105 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
7106 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
7107 | */ | |
62af3783 | 7108 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
7109 | on_rq = 1; |
7110 | ||
b2b5ce02 PZ |
7111 | if (!on_rq) |
7112 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
7113 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
7114 | if (!on_rq) { |
7115 | cfs_rq = cfs_rq_of(&p->se); | |
7116 | p->se.vruntime += cfs_rq->min_vruntime; | |
7117 | #ifdef CONFIG_SMP | |
7118 | /* | |
7119 | * migrate_task_rq_fair() will have removed our previous | |
7120 | * contribution, but we must synchronize for ongoing future | |
7121 | * decay. | |
7122 | */ | |
7123 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
7124 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
7125 | #endif | |
7126 | } | |
810b3817 | 7127 | } |
029632fb PZ |
7128 | |
7129 | void free_fair_sched_group(struct task_group *tg) | |
7130 | { | |
7131 | int i; | |
7132 | ||
7133 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7134 | ||
7135 | for_each_possible_cpu(i) { | |
7136 | if (tg->cfs_rq) | |
7137 | kfree(tg->cfs_rq[i]); | |
7138 | if (tg->se) | |
7139 | kfree(tg->se[i]); | |
7140 | } | |
7141 | ||
7142 | kfree(tg->cfs_rq); | |
7143 | kfree(tg->se); | |
7144 | } | |
7145 | ||
7146 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7147 | { | |
7148 | struct cfs_rq *cfs_rq; | |
7149 | struct sched_entity *se; | |
7150 | int i; | |
7151 | ||
7152 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
7153 | if (!tg->cfs_rq) | |
7154 | goto err; | |
7155 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
7156 | if (!tg->se) | |
7157 | goto err; | |
7158 | ||
7159 | tg->shares = NICE_0_LOAD; | |
7160 | ||
7161 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7162 | ||
7163 | for_each_possible_cpu(i) { | |
7164 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
7165 | GFP_KERNEL, cpu_to_node(i)); | |
7166 | if (!cfs_rq) | |
7167 | goto err; | |
7168 | ||
7169 | se = kzalloc_node(sizeof(struct sched_entity), | |
7170 | GFP_KERNEL, cpu_to_node(i)); | |
7171 | if (!se) | |
7172 | goto err_free_rq; | |
7173 | ||
7174 | init_cfs_rq(cfs_rq); | |
7175 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
7176 | } | |
7177 | ||
7178 | return 1; | |
7179 | ||
7180 | err_free_rq: | |
7181 | kfree(cfs_rq); | |
7182 | err: | |
7183 | return 0; | |
7184 | } | |
7185 | ||
7186 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
7187 | { | |
7188 | struct rq *rq = cpu_rq(cpu); | |
7189 | unsigned long flags; | |
7190 | ||
7191 | /* | |
7192 | * Only empty task groups can be destroyed; so we can speculatively | |
7193 | * check on_list without danger of it being re-added. | |
7194 | */ | |
7195 | if (!tg->cfs_rq[cpu]->on_list) | |
7196 | return; | |
7197 | ||
7198 | raw_spin_lock_irqsave(&rq->lock, flags); | |
7199 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
7200 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7201 | } | |
7202 | ||
7203 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
7204 | struct sched_entity *se, int cpu, | |
7205 | struct sched_entity *parent) | |
7206 | { | |
7207 | struct rq *rq = cpu_rq(cpu); | |
7208 | ||
7209 | cfs_rq->tg = tg; | |
7210 | cfs_rq->rq = rq; | |
029632fb PZ |
7211 | init_cfs_rq_runtime(cfs_rq); |
7212 | ||
7213 | tg->cfs_rq[cpu] = cfs_rq; | |
7214 | tg->se[cpu] = se; | |
7215 | ||
7216 | /* se could be NULL for root_task_group */ | |
7217 | if (!se) | |
7218 | return; | |
7219 | ||
7220 | if (!parent) | |
7221 | se->cfs_rq = &rq->cfs; | |
7222 | else | |
7223 | se->cfs_rq = parent->my_q; | |
7224 | ||
7225 | se->my_q = cfs_rq; | |
7226 | update_load_set(&se->load, 0); | |
7227 | se->parent = parent; | |
7228 | } | |
7229 | ||
7230 | static DEFINE_MUTEX(shares_mutex); | |
7231 | ||
7232 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
7233 | { | |
7234 | int i; | |
7235 | unsigned long flags; | |
7236 | ||
7237 | /* | |
7238 | * We can't change the weight of the root cgroup. | |
7239 | */ | |
7240 | if (!tg->se[0]) | |
7241 | return -EINVAL; | |
7242 | ||
7243 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
7244 | ||
7245 | mutex_lock(&shares_mutex); | |
7246 | if (tg->shares == shares) | |
7247 | goto done; | |
7248 | ||
7249 | tg->shares = shares; | |
7250 | for_each_possible_cpu(i) { | |
7251 | struct rq *rq = cpu_rq(i); | |
7252 | struct sched_entity *se; | |
7253 | ||
7254 | se = tg->se[i]; | |
7255 | /* Propagate contribution to hierarchy */ | |
7256 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
7257 | |
7258 | /* Possible calls to update_curr() need rq clock */ | |
7259 | update_rq_clock(rq); | |
17bc14b7 | 7260 | for_each_sched_entity(se) |
029632fb PZ |
7261 | update_cfs_shares(group_cfs_rq(se)); |
7262 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7263 | } | |
7264 | ||
7265 | done: | |
7266 | mutex_unlock(&shares_mutex); | |
7267 | return 0; | |
7268 | } | |
7269 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
7270 | ||
7271 | void free_fair_sched_group(struct task_group *tg) { } | |
7272 | ||
7273 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7274 | { | |
7275 | return 1; | |
7276 | } | |
7277 | ||
7278 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
7279 | ||
7280 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
7281 | ||
810b3817 | 7282 | |
6d686f45 | 7283 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
7284 | { |
7285 | struct sched_entity *se = &task->se; | |
0d721cea PW |
7286 | unsigned int rr_interval = 0; |
7287 | ||
7288 | /* | |
7289 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
7290 | * idle runqueue: | |
7291 | */ | |
0d721cea | 7292 | if (rq->cfs.load.weight) |
a59f4e07 | 7293 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
7294 | |
7295 | return rr_interval; | |
7296 | } | |
7297 | ||
bf0f6f24 IM |
7298 | /* |
7299 | * All the scheduling class methods: | |
7300 | */ | |
029632fb | 7301 | const struct sched_class fair_sched_class = { |
5522d5d5 | 7302 | .next = &idle_sched_class, |
bf0f6f24 IM |
7303 | .enqueue_task = enqueue_task_fair, |
7304 | .dequeue_task = dequeue_task_fair, | |
7305 | .yield_task = yield_task_fair, | |
d95f4122 | 7306 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 7307 | |
2e09bf55 | 7308 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
7309 | |
7310 | .pick_next_task = pick_next_task_fair, | |
7311 | .put_prev_task = put_prev_task_fair, | |
7312 | ||
681f3e68 | 7313 | #ifdef CONFIG_SMP |
4ce72a2c | 7314 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 7315 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 7316 | |
0bcdcf28 CE |
7317 | .rq_online = rq_online_fair, |
7318 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
7319 | |
7320 | .task_waking = task_waking_fair, | |
681f3e68 | 7321 | #endif |
bf0f6f24 | 7322 | |
83b699ed | 7323 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 7324 | .task_tick = task_tick_fair, |
cd29fe6f | 7325 | .task_fork = task_fork_fair, |
cb469845 SR |
7326 | |
7327 | .prio_changed = prio_changed_fair, | |
da7a735e | 7328 | .switched_from = switched_from_fair, |
cb469845 | 7329 | .switched_to = switched_to_fair, |
810b3817 | 7330 | |
0d721cea PW |
7331 | .get_rr_interval = get_rr_interval_fair, |
7332 | ||
810b3817 | 7333 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7334 | .task_move_group = task_move_group_fair, |
810b3817 | 7335 | #endif |
bf0f6f24 IM |
7336 | }; |
7337 | ||
7338 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 7339 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 7340 | { |
bf0f6f24 IM |
7341 | struct cfs_rq *cfs_rq; |
7342 | ||
5973e5b9 | 7343 | rcu_read_lock(); |
c3b64f1e | 7344 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 7345 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 7346 | rcu_read_unlock(); |
bf0f6f24 IM |
7347 | } |
7348 | #endif | |
029632fb PZ |
7349 | |
7350 | __init void init_sched_fair_class(void) | |
7351 | { | |
7352 | #ifdef CONFIG_SMP | |
7353 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
7354 | ||
3451d024 | 7355 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 7356 | nohz.next_balance = jiffies; |
029632fb | 7357 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 7358 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
7359 | #endif |
7360 | #endif /* SMP */ | |
7361 | ||
7362 | } |