]>
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 AS |
683 | #ifdef CONFIG_SMP |
684 | static inline void __update_task_entity_contrib(struct sched_entity *se); | |
685 | ||
686 | /* Give new task start runnable values to heavy its load in infant time */ | |
687 | void init_task_runnable_average(struct task_struct *p) | |
688 | { | |
689 | u32 slice; | |
690 | ||
691 | p->se.avg.decay_count = 0; | |
692 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | |
693 | p->se.avg.runnable_avg_sum = slice; | |
694 | p->se.avg.runnable_avg_period = slice; | |
695 | __update_task_entity_contrib(&p->se); | |
696 | } | |
697 | #else | |
698 | void init_task_runnable_average(struct task_struct *p) | |
699 | { | |
700 | } | |
701 | #endif | |
702 | ||
bf0f6f24 IM |
703 | /* |
704 | * Update the current task's runtime statistics. Skip current tasks that | |
705 | * are not in our scheduling class. | |
706 | */ | |
707 | static inline void | |
8ebc91d9 IM |
708 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
709 | unsigned long delta_exec) | |
bf0f6f24 | 710 | { |
bbdba7c0 | 711 | unsigned long delta_exec_weighted; |
bf0f6f24 | 712 | |
41acab88 LDM |
713 | schedstat_set(curr->statistics.exec_max, |
714 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
715 | |
716 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 717 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 718 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 719 | |
e9acbff6 | 720 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 721 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
722 | } |
723 | ||
b7cc0896 | 724 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 725 | { |
429d43bc | 726 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 727 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
728 | unsigned long delta_exec; |
729 | ||
730 | if (unlikely(!curr)) | |
731 | return; | |
732 | ||
733 | /* | |
734 | * Get the amount of time the current task was running | |
735 | * since the last time we changed load (this cannot | |
736 | * overflow on 32 bits): | |
737 | */ | |
8ebc91d9 | 738 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
739 | if (!delta_exec) |
740 | return; | |
bf0f6f24 | 741 | |
8ebc91d9 IM |
742 | __update_curr(cfs_rq, curr, delta_exec); |
743 | curr->exec_start = now; | |
d842de87 SV |
744 | |
745 | if (entity_is_task(curr)) { | |
746 | struct task_struct *curtask = task_of(curr); | |
747 | ||
f977bb49 | 748 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 749 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 750 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 751 | } |
ec12cb7f PT |
752 | |
753 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
754 | } |
755 | ||
756 | static inline void | |
5870db5b | 757 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 758 | { |
78becc27 | 759 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
760 | } |
761 | ||
bf0f6f24 IM |
762 | /* |
763 | * Task is being enqueued - update stats: | |
764 | */ | |
d2417e5a | 765 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 766 | { |
bf0f6f24 IM |
767 | /* |
768 | * Are we enqueueing a waiting task? (for current tasks | |
769 | * a dequeue/enqueue event is a NOP) | |
770 | */ | |
429d43bc | 771 | if (se != cfs_rq->curr) |
5870db5b | 772 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
773 | } |
774 | ||
bf0f6f24 | 775 | static void |
9ef0a961 | 776 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 777 | { |
41acab88 | 778 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 779 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
780 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
781 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 782 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
783 | #ifdef CONFIG_SCHEDSTATS |
784 | if (entity_is_task(se)) { | |
785 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 786 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
787 | } |
788 | #endif | |
41acab88 | 789 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
790 | } |
791 | ||
792 | static inline void | |
19b6a2e3 | 793 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 794 | { |
bf0f6f24 IM |
795 | /* |
796 | * Mark the end of the wait period if dequeueing a | |
797 | * waiting task: | |
798 | */ | |
429d43bc | 799 | if (se != cfs_rq->curr) |
9ef0a961 | 800 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
801 | } |
802 | ||
803 | /* | |
804 | * We are picking a new current task - update its stats: | |
805 | */ | |
806 | static inline void | |
79303e9e | 807 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
808 | { |
809 | /* | |
810 | * We are starting a new run period: | |
811 | */ | |
78becc27 | 812 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
813 | } |
814 | ||
bf0f6f24 IM |
815 | /************************************************** |
816 | * Scheduling class queueing methods: | |
817 | */ | |
818 | ||
cbee9f88 PZ |
819 | #ifdef CONFIG_NUMA_BALANCING |
820 | /* | |
598f0ec0 MG |
821 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
822 | * calculated based on the tasks virtual memory size and | |
823 | * numa_balancing_scan_size. | |
cbee9f88 | 824 | */ |
598f0ec0 MG |
825 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
826 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
827 | unsigned int sysctl_numa_balancing_scan_period_reset = 60000; | |
6e5fb223 PZ |
828 | |
829 | /* Portion of address space to scan in MB */ | |
830 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 831 | |
4b96a29b PZ |
832 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
833 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
834 | ||
598f0ec0 MG |
835 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
836 | { | |
837 | unsigned long rss = 0; | |
838 | unsigned long nr_scan_pages; | |
839 | ||
840 | /* | |
841 | * Calculations based on RSS as non-present and empty pages are skipped | |
842 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
843 | * on resident pages | |
844 | */ | |
845 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
846 | rss = get_mm_rss(p->mm); | |
847 | if (!rss) | |
848 | rss = nr_scan_pages; | |
849 | ||
850 | rss = round_up(rss, nr_scan_pages); | |
851 | return rss / nr_scan_pages; | |
852 | } | |
853 | ||
854 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
855 | #define MAX_SCAN_WINDOW 2560 | |
856 | ||
857 | static unsigned int task_scan_min(struct task_struct *p) | |
858 | { | |
859 | unsigned int scan, floor; | |
860 | unsigned int windows = 1; | |
861 | ||
862 | if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW) | |
863 | windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size; | |
864 | floor = 1000 / windows; | |
865 | ||
866 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
867 | return max_t(unsigned int, floor, scan); | |
868 | } | |
869 | ||
870 | static unsigned int task_scan_max(struct task_struct *p) | |
871 | { | |
872 | unsigned int smin = task_scan_min(p); | |
873 | unsigned int smax; | |
874 | ||
875 | /* Watch for min being lower than max due to floor calculations */ | |
876 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
877 | return max(smin, smax); | |
878 | } | |
879 | ||
3a7053b3 MG |
880 | /* |
881 | * Once a preferred node is selected the scheduler balancer will prefer moving | |
882 | * a task to that node for sysctl_numa_balancing_settle_count number of PTE | |
883 | * scans. This will give the process the chance to accumulate more faults on | |
884 | * the preferred node but still allow the scheduler to move the task again if | |
885 | * the nodes CPUs are overloaded. | |
886 | */ | |
887 | unsigned int sysctl_numa_balancing_settle_count __read_mostly = 3; | |
888 | ||
cbee9f88 PZ |
889 | static void task_numa_placement(struct task_struct *p) |
890 | { | |
688b7585 MG |
891 | int seq, nid, max_nid = -1; |
892 | unsigned long max_faults = 0; | |
cbee9f88 | 893 | |
2832bc19 HD |
894 | if (!p->mm) /* for example, ksmd faulting in a user's mm */ |
895 | return; | |
896 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); | |
cbee9f88 PZ |
897 | if (p->numa_scan_seq == seq) |
898 | return; | |
899 | p->numa_scan_seq = seq; | |
3a7053b3 | 900 | p->numa_migrate_seq++; |
598f0ec0 | 901 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 902 | |
688b7585 MG |
903 | /* Find the node with the highest number of faults */ |
904 | for_each_online_node(nid) { | |
745d6147 MG |
905 | unsigned long faults; |
906 | ||
907 | /* Decay existing window and copy faults since last scan */ | |
688b7585 | 908 | p->numa_faults[nid] >>= 1; |
745d6147 MG |
909 | p->numa_faults[nid] += p->numa_faults_buffer[nid]; |
910 | p->numa_faults_buffer[nid] = 0; | |
911 | ||
912 | faults = p->numa_faults[nid]; | |
688b7585 MG |
913 | if (faults > max_faults) { |
914 | max_faults = faults; | |
915 | max_nid = nid; | |
916 | } | |
917 | } | |
918 | ||
919 | /* Update the tasks preferred node if necessary */ | |
3a7053b3 | 920 | if (max_faults && max_nid != p->numa_preferred_nid) { |
688b7585 | 921 | p->numa_preferred_nid = max_nid; |
3a7053b3 MG |
922 | p->numa_migrate_seq = 0; |
923 | } | |
cbee9f88 PZ |
924 | } |
925 | ||
926 | /* | |
927 | * Got a PROT_NONE fault for a page on @node. | |
928 | */ | |
b8593bfd | 929 | void task_numa_fault(int node, int pages, bool migrated) |
cbee9f88 PZ |
930 | { |
931 | struct task_struct *p = current; | |
932 | ||
10e84b97 | 933 | if (!numabalancing_enabled) |
1a687c2e MG |
934 | return; |
935 | ||
f809ca9a MG |
936 | /* Allocate buffer to track faults on a per-node basis */ |
937 | if (unlikely(!p->numa_faults)) { | |
938 | int size = sizeof(*p->numa_faults) * nr_node_ids; | |
939 | ||
745d6147 MG |
940 | /* numa_faults and numa_faults_buffer share the allocation */ |
941 | p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN); | |
f809ca9a MG |
942 | if (!p->numa_faults) |
943 | return; | |
745d6147 MG |
944 | |
945 | BUG_ON(p->numa_faults_buffer); | |
946 | p->numa_faults_buffer = p->numa_faults + nr_node_ids; | |
f809ca9a | 947 | } |
cbee9f88 | 948 | |
fb003b80 | 949 | /* |
b8593bfd MG |
950 | * If pages are properly placed (did not migrate) then scan slower. |
951 | * This is reset periodically in case of phase changes | |
fb003b80 | 952 | */ |
598f0ec0 MG |
953 | if (!migrated) { |
954 | /* Initialise if necessary */ | |
955 | if (!p->numa_scan_period_max) | |
956 | p->numa_scan_period_max = task_scan_max(p); | |
957 | ||
958 | p->numa_scan_period = min(p->numa_scan_period_max, | |
959 | p->numa_scan_period + 10); | |
960 | } | |
fb003b80 | 961 | |
cbee9f88 | 962 | task_numa_placement(p); |
f809ca9a | 963 | |
745d6147 | 964 | p->numa_faults_buffer[node] += pages; |
cbee9f88 PZ |
965 | } |
966 | ||
6e5fb223 PZ |
967 | static void reset_ptenuma_scan(struct task_struct *p) |
968 | { | |
969 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
970 | p->mm->numa_scan_offset = 0; | |
971 | } | |
972 | ||
cbee9f88 PZ |
973 | /* |
974 | * The expensive part of numa migration is done from task_work context. | |
975 | * Triggered from task_tick_numa(). | |
976 | */ | |
977 | void task_numa_work(struct callback_head *work) | |
978 | { | |
979 | unsigned long migrate, next_scan, now = jiffies; | |
980 | struct task_struct *p = current; | |
981 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 982 | struct vm_area_struct *vma; |
9f40604c | 983 | unsigned long start, end; |
598f0ec0 | 984 | unsigned long nr_pte_updates = 0; |
9f40604c | 985 | long pages; |
cbee9f88 PZ |
986 | |
987 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
988 | ||
989 | work->next = work; /* protect against double add */ | |
990 | /* | |
991 | * Who cares about NUMA placement when they're dying. | |
992 | * | |
993 | * NOTE: make sure not to dereference p->mm before this check, | |
994 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
995 | * without p->mm even though we still had it when we enqueued this | |
996 | * work. | |
997 | */ | |
998 | if (p->flags & PF_EXITING) | |
999 | return; | |
1000 | ||
7e8d16b6 MG |
1001 | if (!mm->numa_next_reset || !mm->numa_next_scan) { |
1002 | mm->numa_next_scan = now + | |
1003 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
1004 | mm->numa_next_reset = now + | |
1005 | msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
1006 | } | |
1007 | ||
b8593bfd MG |
1008 | /* |
1009 | * Reset the scan period if enough time has gone by. Objective is that | |
1010 | * scanning will be reduced if pages are properly placed. As tasks | |
1011 | * can enter different phases this needs to be re-examined. Lacking | |
1012 | * proper tracking of reference behaviour, this blunt hammer is used. | |
1013 | */ | |
1014 | migrate = mm->numa_next_reset; | |
1015 | if (time_after(now, migrate)) { | |
598f0ec0 | 1016 | p->numa_scan_period = task_scan_min(p); |
b8593bfd MG |
1017 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); |
1018 | xchg(&mm->numa_next_reset, next_scan); | |
1019 | } | |
1020 | ||
cbee9f88 PZ |
1021 | /* |
1022 | * Enforce maximal scan/migration frequency.. | |
1023 | */ | |
1024 | migrate = mm->numa_next_scan; | |
1025 | if (time_before(now, migrate)) | |
1026 | return; | |
1027 | ||
598f0ec0 MG |
1028 | if (p->numa_scan_period == 0) { |
1029 | p->numa_scan_period_max = task_scan_max(p); | |
1030 | p->numa_scan_period = task_scan_min(p); | |
1031 | } | |
cbee9f88 | 1032 | |
fb003b80 | 1033 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1034 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1035 | return; | |
1036 | ||
19a78d11 PZ |
1037 | /* |
1038 | * Delay this task enough that another task of this mm will likely win | |
1039 | * the next time around. | |
1040 | */ | |
1041 | p->node_stamp += 2 * TICK_NSEC; | |
1042 | ||
9f40604c MG |
1043 | start = mm->numa_scan_offset; |
1044 | pages = sysctl_numa_balancing_scan_size; | |
1045 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1046 | if (!pages) | |
1047 | return; | |
cbee9f88 | 1048 | |
6e5fb223 | 1049 | down_read(&mm->mmap_sem); |
9f40604c | 1050 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1051 | if (!vma) { |
1052 | reset_ptenuma_scan(p); | |
9f40604c | 1053 | start = 0; |
6e5fb223 PZ |
1054 | vma = mm->mmap; |
1055 | } | |
9f40604c | 1056 | for (; vma; vma = vma->vm_next) { |
6e5fb223 PZ |
1057 | if (!vma_migratable(vma)) |
1058 | continue; | |
1059 | ||
1060 | /* Skip small VMAs. They are not likely to be of relevance */ | |
221392c3 | 1061 | if (vma->vm_end - vma->vm_start < HPAGE_SIZE) |
6e5fb223 PZ |
1062 | continue; |
1063 | ||
9f40604c MG |
1064 | do { |
1065 | start = max(start, vma->vm_start); | |
1066 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1067 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1068 | nr_pte_updates += change_prot_numa(vma, start, end); |
1069 | ||
1070 | /* | |
1071 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1072 | * at least one PTE is updated so that unused virtual | |
1073 | * address space is quickly skipped. | |
1074 | */ | |
1075 | if (nr_pte_updates) | |
1076 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1077 | |
9f40604c MG |
1078 | start = end; |
1079 | if (pages <= 0) | |
1080 | goto out; | |
1081 | } while (end != vma->vm_end); | |
cbee9f88 | 1082 | } |
6e5fb223 | 1083 | |
9f40604c | 1084 | out: |
f307cd1a MG |
1085 | /* |
1086 | * If the whole process was scanned without updates then no NUMA | |
1087 | * hinting faults are being recorded and scan rate should be lower. | |
1088 | */ | |
1089 | if (mm->numa_scan_offset == 0 && !nr_pte_updates) { | |
1090 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1091 | p->numa_scan_period << 1); | |
1092 | ||
1093 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); | |
1094 | mm->numa_next_scan = next_scan; | |
1095 | } | |
1096 | ||
6e5fb223 | 1097 | /* |
c69307d5 PZ |
1098 | * It is possible to reach the end of the VMA list but the last few |
1099 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1100 | * would find the !migratable VMA on the next scan but not reset the | |
1101 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1102 | */ |
1103 | if (vma) | |
9f40604c | 1104 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1105 | else |
1106 | reset_ptenuma_scan(p); | |
1107 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1108 | } |
1109 | ||
1110 | /* | |
1111 | * Drive the periodic memory faults.. | |
1112 | */ | |
1113 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1114 | { | |
1115 | struct callback_head *work = &curr->numa_work; | |
1116 | u64 period, now; | |
1117 | ||
1118 | /* | |
1119 | * We don't care about NUMA placement if we don't have memory. | |
1120 | */ | |
1121 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1122 | return; | |
1123 | ||
1124 | /* | |
1125 | * Using runtime rather than walltime has the dual advantage that | |
1126 | * we (mostly) drive the selection from busy threads and that the | |
1127 | * task needs to have done some actual work before we bother with | |
1128 | * NUMA placement. | |
1129 | */ | |
1130 | now = curr->se.sum_exec_runtime; | |
1131 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1132 | ||
1133 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1134 | if (!curr->node_stamp) |
598f0ec0 | 1135 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1136 | curr->node_stamp += period; |
cbee9f88 PZ |
1137 | |
1138 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1139 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1140 | task_work_add(curr, work, true); | |
1141 | } | |
1142 | } | |
1143 | } | |
1144 | #else | |
1145 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1146 | { | |
1147 | } | |
1148 | #endif /* CONFIG_NUMA_BALANCING */ | |
1149 | ||
30cfdcfc DA |
1150 | static void |
1151 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1152 | { | |
1153 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1154 | if (!parent_entity(se)) |
029632fb | 1155 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
1156 | #ifdef CONFIG_SMP |
1157 | if (entity_is_task(se)) | |
eb95308e | 1158 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 1159 | #endif |
30cfdcfc | 1160 | cfs_rq->nr_running++; |
30cfdcfc DA |
1161 | } |
1162 | ||
1163 | static void | |
1164 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1165 | { | |
1166 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1167 | if (!parent_entity(se)) |
029632fb | 1168 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1169 | if (entity_is_task(se)) |
b87f1724 | 1170 | list_del_init(&se->group_node); |
30cfdcfc | 1171 | cfs_rq->nr_running--; |
30cfdcfc DA |
1172 | } |
1173 | ||
3ff6dcac YZ |
1174 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1175 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1176 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1177 | { | |
1178 | long tg_weight; | |
1179 | ||
1180 | /* | |
1181 | * Use this CPU's actual weight instead of the last load_contribution | |
1182 | * to gain a more accurate current total weight. See | |
1183 | * update_cfs_rq_load_contribution(). | |
1184 | */ | |
bf5b986e | 1185 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 1186 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
1187 | tg_weight += cfs_rq->load.weight; |
1188 | ||
1189 | return tg_weight; | |
1190 | } | |
1191 | ||
6d5ab293 | 1192 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1193 | { |
cf5f0acf | 1194 | long tg_weight, load, shares; |
3ff6dcac | 1195 | |
cf5f0acf | 1196 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1197 | load = cfs_rq->load.weight; |
3ff6dcac | 1198 | |
3ff6dcac | 1199 | shares = (tg->shares * load); |
cf5f0acf PZ |
1200 | if (tg_weight) |
1201 | shares /= tg_weight; | |
3ff6dcac YZ |
1202 | |
1203 | if (shares < MIN_SHARES) | |
1204 | shares = MIN_SHARES; | |
1205 | if (shares > tg->shares) | |
1206 | shares = tg->shares; | |
1207 | ||
1208 | return shares; | |
1209 | } | |
3ff6dcac | 1210 | # else /* CONFIG_SMP */ |
6d5ab293 | 1211 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1212 | { |
1213 | return tg->shares; | |
1214 | } | |
3ff6dcac | 1215 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1216 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1217 | unsigned long weight) | |
1218 | { | |
19e5eebb PT |
1219 | if (se->on_rq) { |
1220 | /* commit outstanding execution time */ | |
1221 | if (cfs_rq->curr == se) | |
1222 | update_curr(cfs_rq); | |
2069dd75 | 1223 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1224 | } |
2069dd75 PZ |
1225 | |
1226 | update_load_set(&se->load, weight); | |
1227 | ||
1228 | if (se->on_rq) | |
1229 | account_entity_enqueue(cfs_rq, se); | |
1230 | } | |
1231 | ||
82958366 PT |
1232 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1233 | ||
6d5ab293 | 1234 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1235 | { |
1236 | struct task_group *tg; | |
1237 | struct sched_entity *se; | |
3ff6dcac | 1238 | long shares; |
2069dd75 | 1239 | |
2069dd75 PZ |
1240 | tg = cfs_rq->tg; |
1241 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1242 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1243 | return; |
3ff6dcac YZ |
1244 | #ifndef CONFIG_SMP |
1245 | if (likely(se->load.weight == tg->shares)) | |
1246 | return; | |
1247 | #endif | |
6d5ab293 | 1248 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1249 | |
1250 | reweight_entity(cfs_rq_of(se), se, shares); | |
1251 | } | |
1252 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1253 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1254 | { |
1255 | } | |
1256 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1257 | ||
141965c7 | 1258 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
1259 | /* |
1260 | * We choose a half-life close to 1 scheduling period. | |
1261 | * Note: The tables below are dependent on this value. | |
1262 | */ | |
1263 | #define LOAD_AVG_PERIOD 32 | |
1264 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1265 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1266 | ||
1267 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1268 | static const u32 runnable_avg_yN_inv[] = { | |
1269 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1270 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1271 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1272 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1273 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1274 | 0x85aac367, 0x82cd8698, | |
1275 | }; | |
1276 | ||
1277 | /* | |
1278 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1279 | * over-estimates when re-combining. | |
1280 | */ | |
1281 | static const u32 runnable_avg_yN_sum[] = { | |
1282 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1283 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1284 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1285 | }; | |
1286 | ||
9d85f21c PT |
1287 | /* |
1288 | * Approximate: | |
1289 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1290 | */ | |
1291 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1292 | { | |
5b51f2f8 PT |
1293 | unsigned int local_n; |
1294 | ||
1295 | if (!n) | |
1296 | return val; | |
1297 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1298 | return 0; | |
1299 | ||
1300 | /* after bounds checking we can collapse to 32-bit */ | |
1301 | local_n = n; | |
1302 | ||
1303 | /* | |
1304 | * As y^PERIOD = 1/2, we can combine | |
1305 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
1306 | * With a look-up table which covers k^n (n<PERIOD) | |
1307 | * | |
1308 | * To achieve constant time decay_load. | |
1309 | */ | |
1310 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
1311 | val >>= local_n / LOAD_AVG_PERIOD; | |
1312 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
1313 | } |
1314 | ||
5b51f2f8 PT |
1315 | val *= runnable_avg_yN_inv[local_n]; |
1316 | /* We don't use SRR here since we always want to round down. */ | |
1317 | return val >> 32; | |
1318 | } | |
1319 | ||
1320 | /* | |
1321 | * For updates fully spanning n periods, the contribution to runnable | |
1322 | * average will be: \Sum 1024*y^n | |
1323 | * | |
1324 | * We can compute this reasonably efficiently by combining: | |
1325 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
1326 | */ | |
1327 | static u32 __compute_runnable_contrib(u64 n) | |
1328 | { | |
1329 | u32 contrib = 0; | |
1330 | ||
1331 | if (likely(n <= LOAD_AVG_PERIOD)) | |
1332 | return runnable_avg_yN_sum[n]; | |
1333 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
1334 | return LOAD_AVG_MAX; | |
1335 | ||
1336 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
1337 | do { | |
1338 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
1339 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
1340 | ||
1341 | n -= LOAD_AVG_PERIOD; | |
1342 | } while (n > LOAD_AVG_PERIOD); | |
1343 | ||
1344 | contrib = decay_load(contrib, n); | |
1345 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
1346 | } |
1347 | ||
1348 | /* | |
1349 | * We can represent the historical contribution to runnable average as the | |
1350 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
1351 | * history into segments of approximately 1ms (1024us); label the segment that | |
1352 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
1353 | * | |
1354 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
1355 | * p0 p1 p2 | |
1356 | * (now) (~1ms ago) (~2ms ago) | |
1357 | * | |
1358 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
1359 | * | |
1360 | * We then designate the fractions u_i as our co-efficients, yielding the | |
1361 | * following representation of historical load: | |
1362 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
1363 | * | |
1364 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
1365 | * y^32 = 0.5 | |
1366 | * | |
1367 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
1368 | * approximately half as much as the contribution to load within the last ms | |
1369 | * (u_0). | |
1370 | * | |
1371 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
1372 | * sum again by y is sufficient to update: | |
1373 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1374 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1375 | */ | |
1376 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1377 | struct sched_avg *sa, | |
1378 | int runnable) | |
1379 | { | |
5b51f2f8 PT |
1380 | u64 delta, periods; |
1381 | u32 runnable_contrib; | |
9d85f21c PT |
1382 | int delta_w, decayed = 0; |
1383 | ||
1384 | delta = now - sa->last_runnable_update; | |
1385 | /* | |
1386 | * This should only happen when time goes backwards, which it | |
1387 | * unfortunately does during sched clock init when we swap over to TSC. | |
1388 | */ | |
1389 | if ((s64)delta < 0) { | |
1390 | sa->last_runnable_update = now; | |
1391 | return 0; | |
1392 | } | |
1393 | ||
1394 | /* | |
1395 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1396 | * approximation of 1us and fast to compute. | |
1397 | */ | |
1398 | delta >>= 10; | |
1399 | if (!delta) | |
1400 | return 0; | |
1401 | sa->last_runnable_update = now; | |
1402 | ||
1403 | /* delta_w is the amount already accumulated against our next period */ | |
1404 | delta_w = sa->runnable_avg_period % 1024; | |
1405 | if (delta + delta_w >= 1024) { | |
1406 | /* period roll-over */ | |
1407 | decayed = 1; | |
1408 | ||
1409 | /* | |
1410 | * Now that we know we're crossing a period boundary, figure | |
1411 | * out how much from delta we need to complete the current | |
1412 | * period and accrue it. | |
1413 | */ | |
1414 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
1415 | if (runnable) |
1416 | sa->runnable_avg_sum += delta_w; | |
1417 | sa->runnable_avg_period += delta_w; | |
1418 | ||
1419 | delta -= delta_w; | |
1420 | ||
1421 | /* Figure out how many additional periods this update spans */ | |
1422 | periods = delta / 1024; | |
1423 | delta %= 1024; | |
1424 | ||
1425 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
1426 | periods + 1); | |
1427 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
1428 | periods + 1); | |
1429 | ||
1430 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
1431 | runnable_contrib = __compute_runnable_contrib(periods); | |
1432 | if (runnable) | |
1433 | sa->runnable_avg_sum += runnable_contrib; | |
1434 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
1435 | } |
1436 | ||
1437 | /* Remainder of delta accrued against u_0` */ | |
1438 | if (runnable) | |
1439 | sa->runnable_avg_sum += delta; | |
1440 | sa->runnable_avg_period += delta; | |
1441 | ||
1442 | return decayed; | |
1443 | } | |
1444 | ||
9ee474f5 | 1445 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 1446 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
1447 | { |
1448 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1449 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1450 | ||
1451 | decays -= se->avg.decay_count; | |
1452 | if (!decays) | |
aff3e498 | 1453 | return 0; |
9ee474f5 PT |
1454 | |
1455 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1456 | se->avg.decay_count = 0; | |
aff3e498 PT |
1457 | |
1458 | return decays; | |
9ee474f5 PT |
1459 | } |
1460 | ||
c566e8e9 PT |
1461 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1462 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1463 | int force_update) | |
1464 | { | |
1465 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 1466 | long tg_contrib; |
c566e8e9 PT |
1467 | |
1468 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
1469 | tg_contrib -= cfs_rq->tg_load_contrib; | |
1470 | ||
bf5b986e AS |
1471 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
1472 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
1473 | cfs_rq->tg_load_contrib += tg_contrib; |
1474 | } | |
1475 | } | |
8165e145 | 1476 | |
bb17f655 PT |
1477 | /* |
1478 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
1479 | * representation for computing load contributions. | |
1480 | */ | |
1481 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
1482 | struct cfs_rq *cfs_rq) | |
1483 | { | |
1484 | struct task_group *tg = cfs_rq->tg; | |
1485 | long contrib; | |
1486 | ||
1487 | /* The fraction of a cpu used by this cfs_rq */ | |
1488 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
1489 | sa->runnable_avg_period + 1); | |
1490 | contrib -= cfs_rq->tg_runnable_contrib; | |
1491 | ||
1492 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
1493 | atomic_add(contrib, &tg->runnable_avg); | |
1494 | cfs_rq->tg_runnable_contrib += contrib; | |
1495 | } | |
1496 | } | |
1497 | ||
8165e145 PT |
1498 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
1499 | { | |
1500 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
1501 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
1502 | int runnable_avg; |
1503 | ||
8165e145 PT |
1504 | u64 contrib; |
1505 | ||
1506 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
1507 | se->avg.load_avg_contrib = div_u64(contrib, |
1508 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
1509 | |
1510 | /* | |
1511 | * For group entities we need to compute a correction term in the case | |
1512 | * that they are consuming <1 cpu so that we would contribute the same | |
1513 | * load as a task of equal weight. | |
1514 | * | |
1515 | * Explicitly co-ordinating this measurement would be expensive, but | |
1516 | * fortunately the sum of each cpus contribution forms a usable | |
1517 | * lower-bound on the true value. | |
1518 | * | |
1519 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
1520 | * (and the sum represents true value) or they are disjoint and we are | |
1521 | * understating by the aggregate of their overlap. | |
1522 | * | |
1523 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
1524 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
1525 | * cpus that overlap for this interval and w_i is the interval width. | |
1526 | * | |
1527 | * On a small machine; the first term is well-bounded which bounds the | |
1528 | * total error since w_i is a subset of the period. Whereas on a | |
1529 | * larger machine, while this first term can be larger, if w_i is the | |
1530 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
1531 | * our upper bound of 1-cpu. | |
1532 | */ | |
1533 | runnable_avg = atomic_read(&tg->runnable_avg); | |
1534 | if (runnable_avg < NICE_0_LOAD) { | |
1535 | se->avg.load_avg_contrib *= runnable_avg; | |
1536 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
1537 | } | |
8165e145 | 1538 | } |
c566e8e9 PT |
1539 | #else |
1540 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1541 | int force_update) {} | |
bb17f655 PT |
1542 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
1543 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 1544 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
1545 | #endif |
1546 | ||
8165e145 PT |
1547 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
1548 | { | |
1549 | u32 contrib; | |
1550 | ||
1551 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
1552 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
1553 | contrib /= (se->avg.runnable_avg_period + 1); | |
1554 | se->avg.load_avg_contrib = scale_load(contrib); | |
1555 | } | |
1556 | ||
2dac754e PT |
1557 | /* Compute the current contribution to load_avg by se, return any delta */ |
1558 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
1559 | { | |
1560 | long old_contrib = se->avg.load_avg_contrib; | |
1561 | ||
8165e145 PT |
1562 | if (entity_is_task(se)) { |
1563 | __update_task_entity_contrib(se); | |
1564 | } else { | |
bb17f655 | 1565 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
1566 | __update_group_entity_contrib(se); |
1567 | } | |
2dac754e PT |
1568 | |
1569 | return se->avg.load_avg_contrib - old_contrib; | |
1570 | } | |
1571 | ||
9ee474f5 PT |
1572 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
1573 | long load_contrib) | |
1574 | { | |
1575 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
1576 | cfs_rq->blocked_load_avg -= load_contrib; | |
1577 | else | |
1578 | cfs_rq->blocked_load_avg = 0; | |
1579 | } | |
1580 | ||
f1b17280 PT |
1581 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
1582 | ||
9d85f21c | 1583 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
1584 | static inline void update_entity_load_avg(struct sched_entity *se, |
1585 | int update_cfs_rq) | |
9d85f21c | 1586 | { |
2dac754e PT |
1587 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
1588 | long contrib_delta; | |
f1b17280 | 1589 | u64 now; |
2dac754e | 1590 | |
f1b17280 PT |
1591 | /* |
1592 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
1593 | * case they are the parent of a throttled hierarchy. | |
1594 | */ | |
1595 | if (entity_is_task(se)) | |
1596 | now = cfs_rq_clock_task(cfs_rq); | |
1597 | else | |
1598 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
1599 | ||
1600 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
1601 | return; |
1602 | ||
1603 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
1604 | |
1605 | if (!update_cfs_rq) | |
1606 | return; | |
1607 | ||
2dac754e PT |
1608 | if (se->on_rq) |
1609 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
1610 | else |
1611 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
1612 | } | |
1613 | ||
1614 | /* | |
1615 | * Decay the load contributed by all blocked children and account this so that | |
1616 | * their contribution may appropriately discounted when they wake up. | |
1617 | */ | |
aff3e498 | 1618 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 1619 | { |
f1b17280 | 1620 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
1621 | u64 decays; |
1622 | ||
1623 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 1624 | if (!decays && !force_update) |
9ee474f5 PT |
1625 | return; |
1626 | ||
2509940f AS |
1627 | if (atomic_long_read(&cfs_rq->removed_load)) { |
1628 | unsigned long removed_load; | |
1629 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
1630 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
1631 | } | |
9ee474f5 | 1632 | |
aff3e498 PT |
1633 | if (decays) { |
1634 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
1635 | decays); | |
1636 | atomic64_add(decays, &cfs_rq->decay_counter); | |
1637 | cfs_rq->last_decay = now; | |
1638 | } | |
c566e8e9 PT |
1639 | |
1640 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 1641 | } |
18bf2805 BS |
1642 | |
1643 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
1644 | { | |
78becc27 | 1645 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 1646 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 1647 | } |
2dac754e PT |
1648 | |
1649 | /* Add the load generated by se into cfs_rq's child load-average */ | |
1650 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
1651 | struct sched_entity *se, |
1652 | int wakeup) | |
2dac754e | 1653 | { |
aff3e498 PT |
1654 | /* |
1655 | * We track migrations using entity decay_count <= 0, on a wake-up | |
1656 | * migration we use a negative decay count to track the remote decays | |
1657 | * accumulated while sleeping. | |
a75cdaa9 AS |
1658 | * |
1659 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
1660 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
1661 | * constructed load_avg_contrib. | |
aff3e498 PT |
1662 | */ |
1663 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 1664 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
1665 | if (se->avg.decay_count) { |
1666 | /* | |
1667 | * In a wake-up migration we have to approximate the | |
1668 | * time sleeping. This is because we can't synchronize | |
1669 | * clock_task between the two cpus, and it is not | |
1670 | * guaranteed to be read-safe. Instead, we can | |
1671 | * approximate this using our carried decays, which are | |
1672 | * explicitly atomically readable. | |
1673 | */ | |
1674 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
1675 | << 20; | |
1676 | update_entity_load_avg(se, 0); | |
1677 | /* Indicate that we're now synchronized and on-rq */ | |
1678 | se->avg.decay_count = 0; | |
1679 | } | |
9ee474f5 PT |
1680 | wakeup = 0; |
1681 | } else { | |
282cf499 AS |
1682 | /* |
1683 | * Task re-woke on same cpu (or else migrate_task_rq_fair() | |
1684 | * would have made count negative); we must be careful to avoid | |
1685 | * double-accounting blocked time after synchronizing decays. | |
1686 | */ | |
1687 | se->avg.last_runnable_update += __synchronize_entity_decay(se) | |
1688 | << 20; | |
9ee474f5 PT |
1689 | } |
1690 | ||
aff3e498 PT |
1691 | /* migrated tasks did not contribute to our blocked load */ |
1692 | if (wakeup) { | |
9ee474f5 | 1693 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
1694 | update_entity_load_avg(se, 0); |
1695 | } | |
9ee474f5 | 1696 | |
2dac754e | 1697 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
1698 | /* we force update consideration on load-balancer moves */ |
1699 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
1700 | } |
1701 | ||
9ee474f5 PT |
1702 | /* |
1703 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
1704 | * transitioning to a blocked state we track its projected decay using | |
1705 | * blocked_load_avg. | |
1706 | */ | |
2dac754e | 1707 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1708 | struct sched_entity *se, |
1709 | int sleep) | |
2dac754e | 1710 | { |
9ee474f5 | 1711 | update_entity_load_avg(se, 1); |
aff3e498 PT |
1712 | /* we force update consideration on load-balancer moves */ |
1713 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 1714 | |
2dac754e | 1715 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
1716 | if (sleep) { |
1717 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
1718 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
1719 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 1720 | } |
642dbc39 VG |
1721 | |
1722 | /* | |
1723 | * Update the rq's load with the elapsed running time before entering | |
1724 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
1725 | * be the only way to update the runnable statistic. | |
1726 | */ | |
1727 | void idle_enter_fair(struct rq *this_rq) | |
1728 | { | |
1729 | update_rq_runnable_avg(this_rq, 1); | |
1730 | } | |
1731 | ||
1732 | /* | |
1733 | * Update the rq's load with the elapsed idle time before a task is | |
1734 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
1735 | * be the only way to update the runnable statistic. | |
1736 | */ | |
1737 | void idle_exit_fair(struct rq *this_rq) | |
1738 | { | |
1739 | update_rq_runnable_avg(this_rq, 0); | |
1740 | } | |
1741 | ||
9d85f21c | 1742 | #else |
9ee474f5 PT |
1743 | static inline void update_entity_load_avg(struct sched_entity *se, |
1744 | int update_cfs_rq) {} | |
18bf2805 | 1745 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 1746 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1747 | struct sched_entity *se, |
1748 | int wakeup) {} | |
2dac754e | 1749 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1750 | struct sched_entity *se, |
1751 | int sleep) {} | |
aff3e498 PT |
1752 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
1753 | int force_update) {} | |
9d85f21c PT |
1754 | #endif |
1755 | ||
2396af69 | 1756 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1757 | { |
bf0f6f24 | 1758 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1759 | struct task_struct *tsk = NULL; |
1760 | ||
1761 | if (entity_is_task(se)) | |
1762 | tsk = task_of(se); | |
1763 | ||
41acab88 | 1764 | if (se->statistics.sleep_start) { |
78becc27 | 1765 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
1766 | |
1767 | if ((s64)delta < 0) | |
1768 | delta = 0; | |
1769 | ||
41acab88 LDM |
1770 | if (unlikely(delta > se->statistics.sleep_max)) |
1771 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1772 | |
8c79a045 | 1773 | se->statistics.sleep_start = 0; |
41acab88 | 1774 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1775 | |
768d0c27 | 1776 | if (tsk) { |
e414314c | 1777 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1778 | trace_sched_stat_sleep(tsk, delta); |
1779 | } | |
bf0f6f24 | 1780 | } |
41acab88 | 1781 | if (se->statistics.block_start) { |
78becc27 | 1782 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
1783 | |
1784 | if ((s64)delta < 0) | |
1785 | delta = 0; | |
1786 | ||
41acab88 LDM |
1787 | if (unlikely(delta > se->statistics.block_max)) |
1788 | se->statistics.block_max = delta; | |
bf0f6f24 | 1789 | |
8c79a045 | 1790 | se->statistics.block_start = 0; |
41acab88 | 1791 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1792 | |
e414314c | 1793 | if (tsk) { |
8f0dfc34 | 1794 | if (tsk->in_iowait) { |
41acab88 LDM |
1795 | se->statistics.iowait_sum += delta; |
1796 | se->statistics.iowait_count++; | |
768d0c27 | 1797 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1798 | } |
1799 | ||
b781a602 AV |
1800 | trace_sched_stat_blocked(tsk, delta); |
1801 | ||
e414314c PZ |
1802 | /* |
1803 | * Blocking time is in units of nanosecs, so shift by | |
1804 | * 20 to get a milliseconds-range estimation of the | |
1805 | * amount of time that the task spent sleeping: | |
1806 | */ | |
1807 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1808 | profile_hits(SLEEP_PROFILING, | |
1809 | (void *)get_wchan(tsk), | |
1810 | delta >> 20); | |
1811 | } | |
1812 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1813 | } |
bf0f6f24 IM |
1814 | } |
1815 | #endif | |
1816 | } | |
1817 | ||
ddc97297 PZ |
1818 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1819 | { | |
1820 | #ifdef CONFIG_SCHED_DEBUG | |
1821 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
1822 | ||
1823 | if (d < 0) | |
1824 | d = -d; | |
1825 | ||
1826 | if (d > 3*sysctl_sched_latency) | |
1827 | schedstat_inc(cfs_rq, nr_spread_over); | |
1828 | #endif | |
1829 | } | |
1830 | ||
aeb73b04 PZ |
1831 | static void |
1832 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
1833 | { | |
1af5f730 | 1834 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 1835 | |
2cb8600e PZ |
1836 | /* |
1837 | * The 'current' period is already promised to the current tasks, | |
1838 | * however the extra weight of the new task will slow them down a | |
1839 | * little, place the new task so that it fits in the slot that | |
1840 | * stays open at the end. | |
1841 | */ | |
94dfb5e7 | 1842 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 1843 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 1844 | |
a2e7a7eb | 1845 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 1846 | if (!initial) { |
a2e7a7eb | 1847 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 1848 | |
a2e7a7eb MG |
1849 | /* |
1850 | * Halve their sleep time's effect, to allow | |
1851 | * for a gentler effect of sleepers: | |
1852 | */ | |
1853 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
1854 | thresh >>= 1; | |
51e0304c | 1855 | |
a2e7a7eb | 1856 | vruntime -= thresh; |
aeb73b04 PZ |
1857 | } |
1858 | ||
b5d9d734 | 1859 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 1860 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
1861 | } |
1862 | ||
d3d9dc33 PT |
1863 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
1864 | ||
bf0f6f24 | 1865 | static void |
88ec22d3 | 1866 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1867 | { |
88ec22d3 PZ |
1868 | /* |
1869 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 1870 | * through calling update_curr(). |
88ec22d3 | 1871 | */ |
371fd7e7 | 1872 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
1873 | se->vruntime += cfs_rq->min_vruntime; |
1874 | ||
bf0f6f24 | 1875 | /* |
a2a2d680 | 1876 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 1877 | */ |
b7cc0896 | 1878 | update_curr(cfs_rq); |
f269ae04 | 1879 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
1880 | account_entity_enqueue(cfs_rq, se); |
1881 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 1882 | |
88ec22d3 | 1883 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 1884 | place_entity(cfs_rq, se, 0); |
2396af69 | 1885 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 1886 | } |
bf0f6f24 | 1887 | |
d2417e5a | 1888 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 1889 | check_spread(cfs_rq, se); |
83b699ed SV |
1890 | if (se != cfs_rq->curr) |
1891 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 1892 | se->on_rq = 1; |
3d4b47b4 | 1893 | |
d3d9dc33 | 1894 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 1895 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
1896 | check_enqueue_throttle(cfs_rq); |
1897 | } | |
bf0f6f24 IM |
1898 | } |
1899 | ||
2c13c919 | 1900 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 1901 | { |
2c13c919 RR |
1902 | for_each_sched_entity(se) { |
1903 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1904 | if (cfs_rq->last == se) | |
1905 | cfs_rq->last = NULL; | |
1906 | else | |
1907 | break; | |
1908 | } | |
1909 | } | |
2002c695 | 1910 | |
2c13c919 RR |
1911 | static void __clear_buddies_next(struct sched_entity *se) |
1912 | { | |
1913 | for_each_sched_entity(se) { | |
1914 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1915 | if (cfs_rq->next == se) | |
1916 | cfs_rq->next = NULL; | |
1917 | else | |
1918 | break; | |
1919 | } | |
2002c695 PZ |
1920 | } |
1921 | ||
ac53db59 RR |
1922 | static void __clear_buddies_skip(struct sched_entity *se) |
1923 | { | |
1924 | for_each_sched_entity(se) { | |
1925 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1926 | if (cfs_rq->skip == se) | |
1927 | cfs_rq->skip = NULL; | |
1928 | else | |
1929 | break; | |
1930 | } | |
1931 | } | |
1932 | ||
a571bbea PZ |
1933 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1934 | { | |
2c13c919 RR |
1935 | if (cfs_rq->last == se) |
1936 | __clear_buddies_last(se); | |
1937 | ||
1938 | if (cfs_rq->next == se) | |
1939 | __clear_buddies_next(se); | |
ac53db59 RR |
1940 | |
1941 | if (cfs_rq->skip == se) | |
1942 | __clear_buddies_skip(se); | |
a571bbea PZ |
1943 | } |
1944 | ||
6c16a6dc | 1945 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 1946 | |
bf0f6f24 | 1947 | static void |
371fd7e7 | 1948 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1949 | { |
a2a2d680 DA |
1950 | /* |
1951 | * Update run-time statistics of the 'current'. | |
1952 | */ | |
1953 | update_curr(cfs_rq); | |
17bc14b7 | 1954 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 1955 | |
19b6a2e3 | 1956 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 1957 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 1958 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
1959 | if (entity_is_task(se)) { |
1960 | struct task_struct *tsk = task_of(se); | |
1961 | ||
1962 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 1963 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 1964 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 1965 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 1966 | } |
db36cc7d | 1967 | #endif |
67e9fb2a PZ |
1968 | } |
1969 | ||
2002c695 | 1970 | clear_buddies(cfs_rq, se); |
4793241b | 1971 | |
83b699ed | 1972 | if (se != cfs_rq->curr) |
30cfdcfc | 1973 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 1974 | se->on_rq = 0; |
30cfdcfc | 1975 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
1976 | |
1977 | /* | |
1978 | * Normalize the entity after updating the min_vruntime because the | |
1979 | * update can refer to the ->curr item and we need to reflect this | |
1980 | * movement in our normalized position. | |
1981 | */ | |
371fd7e7 | 1982 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 1983 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 1984 | |
d8b4986d PT |
1985 | /* return excess runtime on last dequeue */ |
1986 | return_cfs_rq_runtime(cfs_rq); | |
1987 | ||
1e876231 | 1988 | update_min_vruntime(cfs_rq); |
17bc14b7 | 1989 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
1990 | } |
1991 | ||
1992 | /* | |
1993 | * Preempt the current task with a newly woken task if needed: | |
1994 | */ | |
7c92e54f | 1995 | static void |
2e09bf55 | 1996 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 1997 | { |
11697830 | 1998 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
1999 | struct sched_entity *se; |
2000 | s64 delta; | |
11697830 | 2001 | |
6d0f0ebd | 2002 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2003 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2004 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2005 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2006 | /* |
2007 | * The current task ran long enough, ensure it doesn't get | |
2008 | * re-elected due to buddy favours. | |
2009 | */ | |
2010 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2011 | return; |
2012 | } | |
2013 | ||
2014 | /* | |
2015 | * Ensure that a task that missed wakeup preemption by a | |
2016 | * narrow margin doesn't have to wait for a full slice. | |
2017 | * This also mitigates buddy induced latencies under load. | |
2018 | */ | |
f685ceac MG |
2019 | if (delta_exec < sysctl_sched_min_granularity) |
2020 | return; | |
2021 | ||
f4cfb33e WX |
2022 | se = __pick_first_entity(cfs_rq); |
2023 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2024 | |
f4cfb33e WX |
2025 | if (delta < 0) |
2026 | return; | |
d7d82944 | 2027 | |
f4cfb33e WX |
2028 | if (delta > ideal_runtime) |
2029 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2030 | } |
2031 | ||
83b699ed | 2032 | static void |
8494f412 | 2033 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2034 | { |
83b699ed SV |
2035 | /* 'current' is not kept within the tree. */ |
2036 | if (se->on_rq) { | |
2037 | /* | |
2038 | * Any task has to be enqueued before it get to execute on | |
2039 | * a CPU. So account for the time it spent waiting on the | |
2040 | * runqueue. | |
2041 | */ | |
2042 | update_stats_wait_end(cfs_rq, se); | |
2043 | __dequeue_entity(cfs_rq, se); | |
2044 | } | |
2045 | ||
79303e9e | 2046 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2047 | cfs_rq->curr = se; |
eba1ed4b IM |
2048 | #ifdef CONFIG_SCHEDSTATS |
2049 | /* | |
2050 | * Track our maximum slice length, if the CPU's load is at | |
2051 | * least twice that of our own weight (i.e. dont track it | |
2052 | * when there are only lesser-weight tasks around): | |
2053 | */ | |
495eca49 | 2054 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2055 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2056 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2057 | } | |
2058 | #endif | |
4a55b450 | 2059 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2060 | } |
2061 | ||
3f3a4904 PZ |
2062 | static int |
2063 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2064 | ||
ac53db59 RR |
2065 | /* |
2066 | * Pick the next process, keeping these things in mind, in this order: | |
2067 | * 1) keep things fair between processes/task groups | |
2068 | * 2) pick the "next" process, since someone really wants that to run | |
2069 | * 3) pick the "last" process, for cache locality | |
2070 | * 4) do not run the "skip" process, if something else is available | |
2071 | */ | |
f4b6755f | 2072 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 2073 | { |
ac53db59 | 2074 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 2075 | struct sched_entity *left = se; |
f4b6755f | 2076 | |
ac53db59 RR |
2077 | /* |
2078 | * Avoid running the skip buddy, if running something else can | |
2079 | * be done without getting too unfair. | |
2080 | */ | |
2081 | if (cfs_rq->skip == se) { | |
2082 | struct sched_entity *second = __pick_next_entity(se); | |
2083 | if (second && wakeup_preempt_entity(second, left) < 1) | |
2084 | se = second; | |
2085 | } | |
aa2ac252 | 2086 | |
f685ceac MG |
2087 | /* |
2088 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2089 | */ | |
2090 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2091 | se = cfs_rq->last; | |
2092 | ||
ac53db59 RR |
2093 | /* |
2094 | * Someone really wants this to run. If it's not unfair, run it. | |
2095 | */ | |
2096 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2097 | se = cfs_rq->next; | |
2098 | ||
f685ceac | 2099 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2100 | |
2101 | return se; | |
aa2ac252 PZ |
2102 | } |
2103 | ||
d3d9dc33 PT |
2104 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
2105 | ||
ab6cde26 | 2106 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2107 | { |
2108 | /* | |
2109 | * If still on the runqueue then deactivate_task() | |
2110 | * was not called and update_curr() has to be done: | |
2111 | */ | |
2112 | if (prev->on_rq) | |
b7cc0896 | 2113 | update_curr(cfs_rq); |
bf0f6f24 | 2114 | |
d3d9dc33 PT |
2115 | /* throttle cfs_rqs exceeding runtime */ |
2116 | check_cfs_rq_runtime(cfs_rq); | |
2117 | ||
ddc97297 | 2118 | check_spread(cfs_rq, prev); |
30cfdcfc | 2119 | if (prev->on_rq) { |
5870db5b | 2120 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2121 | /* Put 'current' back into the tree. */ |
2122 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2123 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2124 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2125 | } |
429d43bc | 2126 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2127 | } |
2128 | ||
8f4d37ec PZ |
2129 | static void |
2130 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2131 | { |
bf0f6f24 | 2132 | /* |
30cfdcfc | 2133 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2134 | */ |
30cfdcfc | 2135 | update_curr(cfs_rq); |
bf0f6f24 | 2136 | |
9d85f21c PT |
2137 | /* |
2138 | * Ensure that runnable average is periodically updated. | |
2139 | */ | |
9ee474f5 | 2140 | update_entity_load_avg(curr, 1); |
aff3e498 | 2141 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2142 | update_cfs_shares(cfs_rq); |
9d85f21c | 2143 | |
8f4d37ec PZ |
2144 | #ifdef CONFIG_SCHED_HRTICK |
2145 | /* | |
2146 | * queued ticks are scheduled to match the slice, so don't bother | |
2147 | * validating it and just reschedule. | |
2148 | */ | |
983ed7a6 HH |
2149 | if (queued) { |
2150 | resched_task(rq_of(cfs_rq)->curr); | |
2151 | return; | |
2152 | } | |
8f4d37ec PZ |
2153 | /* |
2154 | * don't let the period tick interfere with the hrtick preemption | |
2155 | */ | |
2156 | if (!sched_feat(DOUBLE_TICK) && | |
2157 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2158 | return; | |
2159 | #endif | |
2160 | ||
2c2efaed | 2161 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2162 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2163 | } |
2164 | ||
ab84d31e PT |
2165 | |
2166 | /************************************************** | |
2167 | * CFS bandwidth control machinery | |
2168 | */ | |
2169 | ||
2170 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2171 | |
2172 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2173 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2174 | |
2175 | static inline bool cfs_bandwidth_used(void) | |
2176 | { | |
c5905afb | 2177 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2178 | } |
2179 | ||
2180 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
2181 | { | |
2182 | /* only need to count groups transitioning between enabled/!enabled */ | |
2183 | if (enabled && !was_enabled) | |
c5905afb | 2184 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 2185 | else if (!enabled && was_enabled) |
c5905afb | 2186 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
2187 | } |
2188 | #else /* HAVE_JUMP_LABEL */ | |
2189 | static bool cfs_bandwidth_used(void) | |
2190 | { | |
2191 | return true; | |
2192 | } | |
2193 | ||
2194 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
2195 | #endif /* HAVE_JUMP_LABEL */ | |
2196 | ||
ab84d31e PT |
2197 | /* |
2198 | * default period for cfs group bandwidth. | |
2199 | * default: 0.1s, units: nanoseconds | |
2200 | */ | |
2201 | static inline u64 default_cfs_period(void) | |
2202 | { | |
2203 | return 100000000ULL; | |
2204 | } | |
ec12cb7f PT |
2205 | |
2206 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2207 | { | |
2208 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2209 | } | |
2210 | ||
a9cf55b2 PT |
2211 | /* |
2212 | * Replenish runtime according to assigned quota and update expiration time. | |
2213 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2214 | * additional synchronization around rq->lock. | |
2215 | * | |
2216 | * requires cfs_b->lock | |
2217 | */ | |
029632fb | 2218 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2219 | { |
2220 | u64 now; | |
2221 | ||
2222 | if (cfs_b->quota == RUNTIME_INF) | |
2223 | return; | |
2224 | ||
2225 | now = sched_clock_cpu(smp_processor_id()); | |
2226 | cfs_b->runtime = cfs_b->quota; | |
2227 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2228 | } | |
2229 | ||
029632fb PZ |
2230 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2231 | { | |
2232 | return &tg->cfs_bandwidth; | |
2233 | } | |
2234 | ||
f1b17280 PT |
2235 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2236 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2237 | { | |
2238 | if (unlikely(cfs_rq->throttle_count)) | |
2239 | return cfs_rq->throttled_clock_task; | |
2240 | ||
78becc27 | 2241 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2242 | } |
2243 | ||
85dac906 PT |
2244 | /* returns 0 on failure to allocate runtime */ |
2245 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2246 | { |
2247 | struct task_group *tg = cfs_rq->tg; | |
2248 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2249 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2250 | |
2251 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2252 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2253 | ||
2254 | raw_spin_lock(&cfs_b->lock); | |
2255 | if (cfs_b->quota == RUNTIME_INF) | |
2256 | amount = min_amount; | |
58088ad0 | 2257 | else { |
a9cf55b2 PT |
2258 | /* |
2259 | * If the bandwidth pool has become inactive, then at least one | |
2260 | * period must have elapsed since the last consumption. | |
2261 | * Refresh the global state and ensure bandwidth timer becomes | |
2262 | * active. | |
2263 | */ | |
2264 | if (!cfs_b->timer_active) { | |
2265 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2266 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2267 | } |
58088ad0 PT |
2268 | |
2269 | if (cfs_b->runtime > 0) { | |
2270 | amount = min(cfs_b->runtime, min_amount); | |
2271 | cfs_b->runtime -= amount; | |
2272 | cfs_b->idle = 0; | |
2273 | } | |
ec12cb7f | 2274 | } |
a9cf55b2 | 2275 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2276 | raw_spin_unlock(&cfs_b->lock); |
2277 | ||
2278 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2279 | /* |
2280 | * we may have advanced our local expiration to account for allowed | |
2281 | * spread between our sched_clock and the one on which runtime was | |
2282 | * issued. | |
2283 | */ | |
2284 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2285 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2286 | |
2287 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2288 | } |
2289 | ||
a9cf55b2 PT |
2290 | /* |
2291 | * Note: This depends on the synchronization provided by sched_clock and the | |
2292 | * fact that rq->clock snapshots this value. | |
2293 | */ | |
2294 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2295 | { |
a9cf55b2 | 2296 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2297 | |
2298 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2299 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2300 | return; |
2301 | ||
a9cf55b2 PT |
2302 | if (cfs_rq->runtime_remaining < 0) |
2303 | return; | |
2304 | ||
2305 | /* | |
2306 | * If the local deadline has passed we have to consider the | |
2307 | * possibility that our sched_clock is 'fast' and the global deadline | |
2308 | * has not truly expired. | |
2309 | * | |
2310 | * Fortunately we can check determine whether this the case by checking | |
2311 | * whether the global deadline has advanced. | |
2312 | */ | |
2313 | ||
2314 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
2315 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
2316 | cfs_rq->runtime_expires += TICK_NSEC; | |
2317 | } else { | |
2318 | /* global deadline is ahead, expiration has passed */ | |
2319 | cfs_rq->runtime_remaining = 0; | |
2320 | } | |
2321 | } | |
2322 | ||
2323 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2324 | unsigned long delta_exec) | |
2325 | { | |
2326 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 2327 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
2328 | expire_cfs_rq_runtime(cfs_rq); |
2329 | ||
2330 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
2331 | return; |
2332 | ||
85dac906 PT |
2333 | /* |
2334 | * if we're unable to extend our runtime we resched so that the active | |
2335 | * hierarchy can be throttled | |
2336 | */ | |
2337 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
2338 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
2339 | } |
2340 | ||
6c16a6dc PZ |
2341 | static __always_inline |
2342 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 2343 | { |
56f570e5 | 2344 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
2345 | return; |
2346 | ||
2347 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
2348 | } | |
2349 | ||
85dac906 PT |
2350 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
2351 | { | |
56f570e5 | 2352 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
2353 | } |
2354 | ||
64660c86 PT |
2355 | /* check whether cfs_rq, or any parent, is throttled */ |
2356 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2357 | { | |
56f570e5 | 2358 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
2359 | } |
2360 | ||
2361 | /* | |
2362 | * Ensure that neither of the group entities corresponding to src_cpu or | |
2363 | * dest_cpu are members of a throttled hierarchy when performing group | |
2364 | * load-balance operations. | |
2365 | */ | |
2366 | static inline int throttled_lb_pair(struct task_group *tg, | |
2367 | int src_cpu, int dest_cpu) | |
2368 | { | |
2369 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
2370 | ||
2371 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
2372 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
2373 | ||
2374 | return throttled_hierarchy(src_cfs_rq) || | |
2375 | throttled_hierarchy(dest_cfs_rq); | |
2376 | } | |
2377 | ||
2378 | /* updated child weight may affect parent so we have to do this bottom up */ | |
2379 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
2380 | { | |
2381 | struct rq *rq = data; | |
2382 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2383 | ||
2384 | cfs_rq->throttle_count--; | |
2385 | #ifdef CONFIG_SMP | |
2386 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 2387 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 2388 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 2389 | cfs_rq->throttled_clock_task; |
64660c86 PT |
2390 | } |
2391 | #endif | |
2392 | ||
2393 | return 0; | |
2394 | } | |
2395 | ||
2396 | static int tg_throttle_down(struct task_group *tg, void *data) | |
2397 | { | |
2398 | struct rq *rq = data; | |
2399 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2400 | ||
82958366 PT |
2401 | /* group is entering throttled state, stop time */ |
2402 | if (!cfs_rq->throttle_count) | |
78becc27 | 2403 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
2404 | cfs_rq->throttle_count++; |
2405 | ||
2406 | return 0; | |
2407 | } | |
2408 | ||
d3d9dc33 | 2409 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
2410 | { |
2411 | struct rq *rq = rq_of(cfs_rq); | |
2412 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2413 | struct sched_entity *se; | |
2414 | long task_delta, dequeue = 1; | |
2415 | ||
2416 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2417 | ||
f1b17280 | 2418 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
2419 | rcu_read_lock(); |
2420 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
2421 | rcu_read_unlock(); | |
85dac906 PT |
2422 | |
2423 | task_delta = cfs_rq->h_nr_running; | |
2424 | for_each_sched_entity(se) { | |
2425 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
2426 | /* throttled entity or throttle-on-deactivate */ | |
2427 | if (!se->on_rq) | |
2428 | break; | |
2429 | ||
2430 | if (dequeue) | |
2431 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
2432 | qcfs_rq->h_nr_running -= task_delta; | |
2433 | ||
2434 | if (qcfs_rq->load.weight) | |
2435 | dequeue = 0; | |
2436 | } | |
2437 | ||
2438 | if (!se) | |
2439 | rq->nr_running -= task_delta; | |
2440 | ||
2441 | cfs_rq->throttled = 1; | |
78becc27 | 2442 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
2443 | raw_spin_lock(&cfs_b->lock); |
2444 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
2445 | raw_spin_unlock(&cfs_b->lock); | |
2446 | } | |
2447 | ||
029632fb | 2448 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
2449 | { |
2450 | struct rq *rq = rq_of(cfs_rq); | |
2451 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2452 | struct sched_entity *se; | |
2453 | int enqueue = 1; | |
2454 | long task_delta; | |
2455 | ||
22b958d8 | 2456 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
2457 | |
2458 | cfs_rq->throttled = 0; | |
1a55af2e FW |
2459 | |
2460 | update_rq_clock(rq); | |
2461 | ||
671fd9da | 2462 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 2463 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
2464 | list_del_rcu(&cfs_rq->throttled_list); |
2465 | raw_spin_unlock(&cfs_b->lock); | |
2466 | ||
64660c86 PT |
2467 | /* update hierarchical throttle state */ |
2468 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
2469 | ||
671fd9da PT |
2470 | if (!cfs_rq->load.weight) |
2471 | return; | |
2472 | ||
2473 | task_delta = cfs_rq->h_nr_running; | |
2474 | for_each_sched_entity(se) { | |
2475 | if (se->on_rq) | |
2476 | enqueue = 0; | |
2477 | ||
2478 | cfs_rq = cfs_rq_of(se); | |
2479 | if (enqueue) | |
2480 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
2481 | cfs_rq->h_nr_running += task_delta; | |
2482 | ||
2483 | if (cfs_rq_throttled(cfs_rq)) | |
2484 | break; | |
2485 | } | |
2486 | ||
2487 | if (!se) | |
2488 | rq->nr_running += task_delta; | |
2489 | ||
2490 | /* determine whether we need to wake up potentially idle cpu */ | |
2491 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
2492 | resched_task(rq->curr); | |
2493 | } | |
2494 | ||
2495 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
2496 | u64 remaining, u64 expires) | |
2497 | { | |
2498 | struct cfs_rq *cfs_rq; | |
2499 | u64 runtime = remaining; | |
2500 | ||
2501 | rcu_read_lock(); | |
2502 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
2503 | throttled_list) { | |
2504 | struct rq *rq = rq_of(cfs_rq); | |
2505 | ||
2506 | raw_spin_lock(&rq->lock); | |
2507 | if (!cfs_rq_throttled(cfs_rq)) | |
2508 | goto next; | |
2509 | ||
2510 | runtime = -cfs_rq->runtime_remaining + 1; | |
2511 | if (runtime > remaining) | |
2512 | runtime = remaining; | |
2513 | remaining -= runtime; | |
2514 | ||
2515 | cfs_rq->runtime_remaining += runtime; | |
2516 | cfs_rq->runtime_expires = expires; | |
2517 | ||
2518 | /* we check whether we're throttled above */ | |
2519 | if (cfs_rq->runtime_remaining > 0) | |
2520 | unthrottle_cfs_rq(cfs_rq); | |
2521 | ||
2522 | next: | |
2523 | raw_spin_unlock(&rq->lock); | |
2524 | ||
2525 | if (!remaining) | |
2526 | break; | |
2527 | } | |
2528 | rcu_read_unlock(); | |
2529 | ||
2530 | return remaining; | |
2531 | } | |
2532 | ||
58088ad0 PT |
2533 | /* |
2534 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2535 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2536 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2537 | * used to track this state. | |
2538 | */ | |
2539 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
2540 | { | |
671fd9da PT |
2541 | u64 runtime, runtime_expires; |
2542 | int idle = 1, throttled; | |
58088ad0 PT |
2543 | |
2544 | raw_spin_lock(&cfs_b->lock); | |
2545 | /* no need to continue the timer with no bandwidth constraint */ | |
2546 | if (cfs_b->quota == RUNTIME_INF) | |
2547 | goto out_unlock; | |
2548 | ||
671fd9da PT |
2549 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
2550 | /* idle depends on !throttled (for the case of a large deficit) */ | |
2551 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 2552 | cfs_b->nr_periods += overrun; |
671fd9da | 2553 | |
a9cf55b2 PT |
2554 | /* if we're going inactive then everything else can be deferred */ |
2555 | if (idle) | |
2556 | goto out_unlock; | |
2557 | ||
2558 | __refill_cfs_bandwidth_runtime(cfs_b); | |
2559 | ||
671fd9da PT |
2560 | if (!throttled) { |
2561 | /* mark as potentially idle for the upcoming period */ | |
2562 | cfs_b->idle = 1; | |
2563 | goto out_unlock; | |
2564 | } | |
2565 | ||
e8da1b18 NR |
2566 | /* account preceding periods in which throttling occurred */ |
2567 | cfs_b->nr_throttled += overrun; | |
2568 | ||
671fd9da PT |
2569 | /* |
2570 | * There are throttled entities so we must first use the new bandwidth | |
2571 | * to unthrottle them before making it generally available. This | |
2572 | * ensures that all existing debts will be paid before a new cfs_rq is | |
2573 | * allowed to run. | |
2574 | */ | |
2575 | runtime = cfs_b->runtime; | |
2576 | runtime_expires = cfs_b->runtime_expires; | |
2577 | cfs_b->runtime = 0; | |
2578 | ||
2579 | /* | |
2580 | * This check is repeated as we are holding onto the new bandwidth | |
2581 | * while we unthrottle. This can potentially race with an unthrottled | |
2582 | * group trying to acquire new bandwidth from the global pool. | |
2583 | */ | |
2584 | while (throttled && runtime > 0) { | |
2585 | raw_spin_unlock(&cfs_b->lock); | |
2586 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
2587 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
2588 | runtime_expires); | |
2589 | raw_spin_lock(&cfs_b->lock); | |
2590 | ||
2591 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
2592 | } | |
58088ad0 | 2593 | |
671fd9da PT |
2594 | /* return (any) remaining runtime */ |
2595 | cfs_b->runtime = runtime; | |
2596 | /* | |
2597 | * While we are ensured activity in the period following an | |
2598 | * unthrottle, this also covers the case in which the new bandwidth is | |
2599 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
2600 | * timer to remain active while there are any throttled entities.) | |
2601 | */ | |
2602 | cfs_b->idle = 0; | |
58088ad0 PT |
2603 | out_unlock: |
2604 | if (idle) | |
2605 | cfs_b->timer_active = 0; | |
2606 | raw_spin_unlock(&cfs_b->lock); | |
2607 | ||
2608 | return idle; | |
2609 | } | |
d3d9dc33 | 2610 | |
d8b4986d PT |
2611 | /* a cfs_rq won't donate quota below this amount */ |
2612 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
2613 | /* minimum remaining period time to redistribute slack quota */ | |
2614 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
2615 | /* how long we wait to gather additional slack before distributing */ | |
2616 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
2617 | ||
2618 | /* are we near the end of the current quota period? */ | |
2619 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
2620 | { | |
2621 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
2622 | u64 remaining; | |
2623 | ||
2624 | /* if the call-back is running a quota refresh is already occurring */ | |
2625 | if (hrtimer_callback_running(refresh_timer)) | |
2626 | return 1; | |
2627 | ||
2628 | /* is a quota refresh about to occur? */ | |
2629 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
2630 | if (remaining < min_expire) | |
2631 | return 1; | |
2632 | ||
2633 | return 0; | |
2634 | } | |
2635 | ||
2636 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
2637 | { | |
2638 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
2639 | ||
2640 | /* if there's a quota refresh soon don't bother with slack */ | |
2641 | if (runtime_refresh_within(cfs_b, min_left)) | |
2642 | return; | |
2643 | ||
2644 | start_bandwidth_timer(&cfs_b->slack_timer, | |
2645 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
2646 | } | |
2647 | ||
2648 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
2649 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2650 | { | |
2651 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2652 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
2653 | ||
2654 | if (slack_runtime <= 0) | |
2655 | return; | |
2656 | ||
2657 | raw_spin_lock(&cfs_b->lock); | |
2658 | if (cfs_b->quota != RUNTIME_INF && | |
2659 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
2660 | cfs_b->runtime += slack_runtime; | |
2661 | ||
2662 | /* we are under rq->lock, defer unthrottling using a timer */ | |
2663 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
2664 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
2665 | start_cfs_slack_bandwidth(cfs_b); | |
2666 | } | |
2667 | raw_spin_unlock(&cfs_b->lock); | |
2668 | ||
2669 | /* even if it's not valid for return we don't want to try again */ | |
2670 | cfs_rq->runtime_remaining -= slack_runtime; | |
2671 | } | |
2672 | ||
2673 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2674 | { | |
56f570e5 PT |
2675 | if (!cfs_bandwidth_used()) |
2676 | return; | |
2677 | ||
fccfdc6f | 2678 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2679 | return; |
2680 | ||
2681 | __return_cfs_rq_runtime(cfs_rq); | |
2682 | } | |
2683 | ||
2684 | /* | |
2685 | * This is done with a timer (instead of inline with bandwidth return) since | |
2686 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2687 | */ | |
2688 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2689 | { | |
2690 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2691 | u64 expires; | |
2692 | ||
2693 | /* confirm we're still not at a refresh boundary */ | |
2694 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2695 | return; | |
2696 | ||
2697 | raw_spin_lock(&cfs_b->lock); | |
2698 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2699 | runtime = cfs_b->runtime; | |
2700 | cfs_b->runtime = 0; | |
2701 | } | |
2702 | expires = cfs_b->runtime_expires; | |
2703 | raw_spin_unlock(&cfs_b->lock); | |
2704 | ||
2705 | if (!runtime) | |
2706 | return; | |
2707 | ||
2708 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2709 | ||
2710 | raw_spin_lock(&cfs_b->lock); | |
2711 | if (expires == cfs_b->runtime_expires) | |
2712 | cfs_b->runtime = runtime; | |
2713 | raw_spin_unlock(&cfs_b->lock); | |
2714 | } | |
2715 | ||
d3d9dc33 PT |
2716 | /* |
2717 | * When a group wakes up we want to make sure that its quota is not already | |
2718 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2719 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2720 | */ | |
2721 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2722 | { | |
56f570e5 PT |
2723 | if (!cfs_bandwidth_used()) |
2724 | return; | |
2725 | ||
d3d9dc33 PT |
2726 | /* an active group must be handled by the update_curr()->put() path */ |
2727 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2728 | return; | |
2729 | ||
2730 | /* ensure the group is not already throttled */ | |
2731 | if (cfs_rq_throttled(cfs_rq)) | |
2732 | return; | |
2733 | ||
2734 | /* update runtime allocation */ | |
2735 | account_cfs_rq_runtime(cfs_rq, 0); | |
2736 | if (cfs_rq->runtime_remaining <= 0) | |
2737 | throttle_cfs_rq(cfs_rq); | |
2738 | } | |
2739 | ||
2740 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2741 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2742 | { | |
56f570e5 PT |
2743 | if (!cfs_bandwidth_used()) |
2744 | return; | |
2745 | ||
d3d9dc33 PT |
2746 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2747 | return; | |
2748 | ||
2749 | /* | |
2750 | * it's possible for a throttled entity to be forced into a running | |
2751 | * state (e.g. set_curr_task), in this case we're finished. | |
2752 | */ | |
2753 | if (cfs_rq_throttled(cfs_rq)) | |
2754 | return; | |
2755 | ||
2756 | throttle_cfs_rq(cfs_rq); | |
2757 | } | |
029632fb | 2758 | |
029632fb PZ |
2759 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
2760 | { | |
2761 | struct cfs_bandwidth *cfs_b = | |
2762 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2763 | do_sched_cfs_slack_timer(cfs_b); | |
2764 | ||
2765 | return HRTIMER_NORESTART; | |
2766 | } | |
2767 | ||
2768 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2769 | { | |
2770 | struct cfs_bandwidth *cfs_b = | |
2771 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2772 | ktime_t now; | |
2773 | int overrun; | |
2774 | int idle = 0; | |
2775 | ||
2776 | for (;;) { | |
2777 | now = hrtimer_cb_get_time(timer); | |
2778 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2779 | ||
2780 | if (!overrun) | |
2781 | break; | |
2782 | ||
2783 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2784 | } | |
2785 | ||
2786 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2787 | } | |
2788 | ||
2789 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2790 | { | |
2791 | raw_spin_lock_init(&cfs_b->lock); | |
2792 | cfs_b->runtime = 0; | |
2793 | cfs_b->quota = RUNTIME_INF; | |
2794 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2795 | ||
2796 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2797 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2798 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2799 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2800 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2801 | } | |
2802 | ||
2803 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2804 | { | |
2805 | cfs_rq->runtime_enabled = 0; | |
2806 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2807 | } | |
2808 | ||
2809 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2810 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2811 | { | |
2812 | /* | |
2813 | * The timer may be active because we're trying to set a new bandwidth | |
2814 | * period or because we're racing with the tear-down path | |
2815 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2816 | * terminates). In either case we ensure that it's re-programmed | |
2817 | */ | |
2818 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2819 | raw_spin_unlock(&cfs_b->lock); | |
2820 | /* ensure cfs_b->lock is available while we wait */ | |
2821 | hrtimer_cancel(&cfs_b->period_timer); | |
2822 | ||
2823 | raw_spin_lock(&cfs_b->lock); | |
2824 | /* if someone else restarted the timer then we're done */ | |
2825 | if (cfs_b->timer_active) | |
2826 | return; | |
2827 | } | |
2828 | ||
2829 | cfs_b->timer_active = 1; | |
2830 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
2831 | } | |
2832 | ||
2833 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2834 | { | |
2835 | hrtimer_cancel(&cfs_b->period_timer); | |
2836 | hrtimer_cancel(&cfs_b->slack_timer); | |
2837 | } | |
2838 | ||
38dc3348 | 2839 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
2840 | { |
2841 | struct cfs_rq *cfs_rq; | |
2842 | ||
2843 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
2844 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2845 | ||
2846 | if (!cfs_rq->runtime_enabled) | |
2847 | continue; | |
2848 | ||
2849 | /* | |
2850 | * clock_task is not advancing so we just need to make sure | |
2851 | * there's some valid quota amount | |
2852 | */ | |
2853 | cfs_rq->runtime_remaining = cfs_b->quota; | |
2854 | if (cfs_rq_throttled(cfs_rq)) | |
2855 | unthrottle_cfs_rq(cfs_rq); | |
2856 | } | |
2857 | } | |
2858 | ||
2859 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
2860 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
2861 | { | |
78becc27 | 2862 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
2863 | } |
2864 | ||
2865 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2866 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
2867 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
2868 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 2869 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
2870 | |
2871 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
2872 | { | |
2873 | return 0; | |
2874 | } | |
64660c86 PT |
2875 | |
2876 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2877 | { | |
2878 | return 0; | |
2879 | } | |
2880 | ||
2881 | static inline int throttled_lb_pair(struct task_group *tg, | |
2882 | int src_cpu, int dest_cpu) | |
2883 | { | |
2884 | return 0; | |
2885 | } | |
029632fb PZ |
2886 | |
2887 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
2888 | ||
2889 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2890 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
2891 | #endif |
2892 | ||
029632fb PZ |
2893 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2894 | { | |
2895 | return NULL; | |
2896 | } | |
2897 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 2898 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
2899 | |
2900 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
2901 | ||
bf0f6f24 IM |
2902 | /************************************************** |
2903 | * CFS operations on tasks: | |
2904 | */ | |
2905 | ||
8f4d37ec PZ |
2906 | #ifdef CONFIG_SCHED_HRTICK |
2907 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2908 | { | |
8f4d37ec PZ |
2909 | struct sched_entity *se = &p->se; |
2910 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2911 | ||
2912 | WARN_ON(task_rq(p) != rq); | |
2913 | ||
b39e66ea | 2914 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
2915 | u64 slice = sched_slice(cfs_rq, se); |
2916 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
2917 | s64 delta = slice - ran; | |
2918 | ||
2919 | if (delta < 0) { | |
2920 | if (rq->curr == p) | |
2921 | resched_task(p); | |
2922 | return; | |
2923 | } | |
2924 | ||
2925 | /* | |
2926 | * Don't schedule slices shorter than 10000ns, that just | |
2927 | * doesn't make sense. Rely on vruntime for fairness. | |
2928 | */ | |
31656519 | 2929 | if (rq->curr != p) |
157124c1 | 2930 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 2931 | |
31656519 | 2932 | hrtick_start(rq, delta); |
8f4d37ec PZ |
2933 | } |
2934 | } | |
a4c2f00f PZ |
2935 | |
2936 | /* | |
2937 | * called from enqueue/dequeue and updates the hrtick when the | |
2938 | * current task is from our class and nr_running is low enough | |
2939 | * to matter. | |
2940 | */ | |
2941 | static void hrtick_update(struct rq *rq) | |
2942 | { | |
2943 | struct task_struct *curr = rq->curr; | |
2944 | ||
b39e66ea | 2945 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
2946 | return; |
2947 | ||
2948 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
2949 | hrtick_start_fair(rq, curr); | |
2950 | } | |
55e12e5e | 2951 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
2952 | static inline void |
2953 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
2954 | { | |
2955 | } | |
a4c2f00f PZ |
2956 | |
2957 | static inline void hrtick_update(struct rq *rq) | |
2958 | { | |
2959 | } | |
8f4d37ec PZ |
2960 | #endif |
2961 | ||
bf0f6f24 IM |
2962 | /* |
2963 | * The enqueue_task method is called before nr_running is | |
2964 | * increased. Here we update the fair scheduling stats and | |
2965 | * then put the task into the rbtree: | |
2966 | */ | |
ea87bb78 | 2967 | static void |
371fd7e7 | 2968 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
2969 | { |
2970 | struct cfs_rq *cfs_rq; | |
62fb1851 | 2971 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
2972 | |
2973 | for_each_sched_entity(se) { | |
62fb1851 | 2974 | if (se->on_rq) |
bf0f6f24 IM |
2975 | break; |
2976 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 2977 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
2978 | |
2979 | /* | |
2980 | * end evaluation on encountering a throttled cfs_rq | |
2981 | * | |
2982 | * note: in the case of encountering a throttled cfs_rq we will | |
2983 | * post the final h_nr_running increment below. | |
2984 | */ | |
2985 | if (cfs_rq_throttled(cfs_rq)) | |
2986 | break; | |
953bfcd1 | 2987 | cfs_rq->h_nr_running++; |
85dac906 | 2988 | |
88ec22d3 | 2989 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 2990 | } |
8f4d37ec | 2991 | |
2069dd75 | 2992 | for_each_sched_entity(se) { |
0f317143 | 2993 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 2994 | cfs_rq->h_nr_running++; |
2069dd75 | 2995 | |
85dac906 PT |
2996 | if (cfs_rq_throttled(cfs_rq)) |
2997 | break; | |
2998 | ||
17bc14b7 | 2999 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3000 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3001 | } |
3002 | ||
18bf2805 BS |
3003 | if (!se) { |
3004 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3005 | inc_nr_running(rq); |
18bf2805 | 3006 | } |
a4c2f00f | 3007 | hrtick_update(rq); |
bf0f6f24 IM |
3008 | } |
3009 | ||
2f36825b VP |
3010 | static void set_next_buddy(struct sched_entity *se); |
3011 | ||
bf0f6f24 IM |
3012 | /* |
3013 | * The dequeue_task method is called before nr_running is | |
3014 | * decreased. We remove the task from the rbtree and | |
3015 | * update the fair scheduling stats: | |
3016 | */ | |
371fd7e7 | 3017 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3018 | { |
3019 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3020 | struct sched_entity *se = &p->se; |
2f36825b | 3021 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3022 | |
3023 | for_each_sched_entity(se) { | |
3024 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3025 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3026 | |
3027 | /* | |
3028 | * end evaluation on encountering a throttled cfs_rq | |
3029 | * | |
3030 | * note: in the case of encountering a throttled cfs_rq we will | |
3031 | * post the final h_nr_running decrement below. | |
3032 | */ | |
3033 | if (cfs_rq_throttled(cfs_rq)) | |
3034 | break; | |
953bfcd1 | 3035 | cfs_rq->h_nr_running--; |
2069dd75 | 3036 | |
bf0f6f24 | 3037 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3038 | if (cfs_rq->load.weight) { |
3039 | /* | |
3040 | * Bias pick_next to pick a task from this cfs_rq, as | |
3041 | * p is sleeping when it is within its sched_slice. | |
3042 | */ | |
3043 | if (task_sleep && parent_entity(se)) | |
3044 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3045 | |
3046 | /* avoid re-evaluating load for this entity */ | |
3047 | se = parent_entity(se); | |
bf0f6f24 | 3048 | break; |
2f36825b | 3049 | } |
371fd7e7 | 3050 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3051 | } |
8f4d37ec | 3052 | |
2069dd75 | 3053 | for_each_sched_entity(se) { |
0f317143 | 3054 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3055 | cfs_rq->h_nr_running--; |
2069dd75 | 3056 | |
85dac906 PT |
3057 | if (cfs_rq_throttled(cfs_rq)) |
3058 | break; | |
3059 | ||
17bc14b7 | 3060 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3061 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3062 | } |
3063 | ||
18bf2805 | 3064 | if (!se) { |
85dac906 | 3065 | dec_nr_running(rq); |
18bf2805 BS |
3066 | update_rq_runnable_avg(rq, 1); |
3067 | } | |
a4c2f00f | 3068 | hrtick_update(rq); |
bf0f6f24 IM |
3069 | } |
3070 | ||
e7693a36 | 3071 | #ifdef CONFIG_SMP |
029632fb PZ |
3072 | /* Used instead of source_load when we know the type == 0 */ |
3073 | static unsigned long weighted_cpuload(const int cpu) | |
3074 | { | |
b92486cb | 3075 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3076 | } |
3077 | ||
3078 | /* | |
3079 | * Return a low guess at the load of a migration-source cpu weighted | |
3080 | * according to the scheduling class and "nice" value. | |
3081 | * | |
3082 | * We want to under-estimate the load of migration sources, to | |
3083 | * balance conservatively. | |
3084 | */ | |
3085 | static unsigned long source_load(int cpu, int type) | |
3086 | { | |
3087 | struct rq *rq = cpu_rq(cpu); | |
3088 | unsigned long total = weighted_cpuload(cpu); | |
3089 | ||
3090 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3091 | return total; | |
3092 | ||
3093 | return min(rq->cpu_load[type-1], total); | |
3094 | } | |
3095 | ||
3096 | /* | |
3097 | * Return a high guess at the load of a migration-target cpu weighted | |
3098 | * according to the scheduling class and "nice" value. | |
3099 | */ | |
3100 | static unsigned long target_load(int cpu, int type) | |
3101 | { | |
3102 | struct rq *rq = cpu_rq(cpu); | |
3103 | unsigned long total = weighted_cpuload(cpu); | |
3104 | ||
3105 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3106 | return total; | |
3107 | ||
3108 | return max(rq->cpu_load[type-1], total); | |
3109 | } | |
3110 | ||
3111 | static unsigned long power_of(int cpu) | |
3112 | { | |
3113 | return cpu_rq(cpu)->cpu_power; | |
3114 | } | |
3115 | ||
3116 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3117 | { | |
3118 | struct rq *rq = cpu_rq(cpu); | |
3119 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3120 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3121 | |
3122 | if (nr_running) | |
b92486cb | 3123 | return load_avg / nr_running; |
029632fb PZ |
3124 | |
3125 | return 0; | |
3126 | } | |
3127 | ||
62470419 MW |
3128 | static void record_wakee(struct task_struct *p) |
3129 | { | |
3130 | /* | |
3131 | * Rough decay (wiping) for cost saving, don't worry | |
3132 | * about the boundary, really active task won't care | |
3133 | * about the loss. | |
3134 | */ | |
3135 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
3136 | current->wakee_flips = 0; | |
3137 | current->wakee_flip_decay_ts = jiffies; | |
3138 | } | |
3139 | ||
3140 | if (current->last_wakee != p) { | |
3141 | current->last_wakee = p; | |
3142 | current->wakee_flips++; | |
3143 | } | |
3144 | } | |
098fb9db | 3145 | |
74f8e4b2 | 3146 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
3147 | { |
3148 | struct sched_entity *se = &p->se; | |
3149 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
3150 | u64 min_vruntime; |
3151 | ||
3152 | #ifndef CONFIG_64BIT | |
3153 | u64 min_vruntime_copy; | |
88ec22d3 | 3154 | |
3fe1698b PZ |
3155 | do { |
3156 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3157 | smp_rmb(); | |
3158 | min_vruntime = cfs_rq->min_vruntime; | |
3159 | } while (min_vruntime != min_vruntime_copy); | |
3160 | #else | |
3161 | min_vruntime = cfs_rq->min_vruntime; | |
3162 | #endif | |
88ec22d3 | 3163 | |
3fe1698b | 3164 | se->vruntime -= min_vruntime; |
62470419 | 3165 | record_wakee(p); |
88ec22d3 PZ |
3166 | } |
3167 | ||
bb3469ac | 3168 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3169 | /* |
3170 | * effective_load() calculates the load change as seen from the root_task_group | |
3171 | * | |
3172 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3173 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3174 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3175 | * |
3176 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3177 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3178 | * total group weight. | |
3179 | * | |
3180 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3181 | * distribution (s_i) using: | |
3182 | * | |
3183 | * s_i = rw_i / \Sum rw_j (1) | |
3184 | * | |
3185 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3186 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3187 | * shares distribution (s_i): | |
3188 | * | |
3189 | * rw_i = { 2, 4, 1, 0 } | |
3190 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3191 | * | |
3192 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3193 | * task used to run on and the CPU the waker is running on), we need to | |
3194 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3195 | * wakeup, compute the effect of the current task going to sleep. | |
3196 | * | |
3197 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3198 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3199 | * | |
3200 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3201 | * | |
3202 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3203 | * differences in waking a task to CPU 0. The additional task changes the | |
3204 | * weight and shares distributions like: | |
3205 | * | |
3206 | * rw'_i = { 3, 4, 1, 0 } | |
3207 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3208 | * | |
3209 | * We can then compute the difference in effective weight by using: | |
3210 | * | |
3211 | * dw_i = S * (s'_i - s_i) (3) | |
3212 | * | |
3213 | * Where 'S' is the group weight as seen by its parent. | |
3214 | * | |
3215 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3216 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3217 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3218 | */ |
2069dd75 | 3219 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3220 | { |
4be9daaa | 3221 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3222 | |
cf5f0acf | 3223 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3224 | return wl; |
3225 | ||
4be9daaa | 3226 | for_each_sched_entity(se) { |
cf5f0acf | 3227 | long w, W; |
4be9daaa | 3228 | |
977dda7c | 3229 | tg = se->my_q->tg; |
bb3469ac | 3230 | |
cf5f0acf PZ |
3231 | /* |
3232 | * W = @wg + \Sum rw_j | |
3233 | */ | |
3234 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3235 | |
cf5f0acf PZ |
3236 | /* |
3237 | * w = rw_i + @wl | |
3238 | */ | |
3239 | w = se->my_q->load.weight + wl; | |
940959e9 | 3240 | |
cf5f0acf PZ |
3241 | /* |
3242 | * wl = S * s'_i; see (2) | |
3243 | */ | |
3244 | if (W > 0 && w < W) | |
3245 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3246 | else |
3247 | wl = tg->shares; | |
940959e9 | 3248 | |
cf5f0acf PZ |
3249 | /* |
3250 | * Per the above, wl is the new se->load.weight value; since | |
3251 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3252 | * calc_cfs_shares(). | |
3253 | */ | |
977dda7c PT |
3254 | if (wl < MIN_SHARES) |
3255 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3256 | |
3257 | /* | |
3258 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3259 | */ | |
977dda7c | 3260 | wl -= se->load.weight; |
cf5f0acf PZ |
3261 | |
3262 | /* | |
3263 | * Recursively apply this logic to all parent groups to compute | |
3264 | * the final effective load change on the root group. Since | |
3265 | * only the @tg group gets extra weight, all parent groups can | |
3266 | * only redistribute existing shares. @wl is the shift in shares | |
3267 | * resulting from this level per the above. | |
3268 | */ | |
4be9daaa | 3269 | wg = 0; |
4be9daaa | 3270 | } |
bb3469ac | 3271 | |
4be9daaa | 3272 | return wl; |
bb3469ac PZ |
3273 | } |
3274 | #else | |
4be9daaa | 3275 | |
83378269 PZ |
3276 | static inline unsigned long effective_load(struct task_group *tg, int cpu, |
3277 | unsigned long wl, unsigned long wg) | |
4be9daaa | 3278 | { |
83378269 | 3279 | return wl; |
bb3469ac | 3280 | } |
4be9daaa | 3281 | |
bb3469ac PZ |
3282 | #endif |
3283 | ||
62470419 MW |
3284 | static int wake_wide(struct task_struct *p) |
3285 | { | |
7d9ffa89 | 3286 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
3287 | |
3288 | /* | |
3289 | * Yeah, it's the switching-frequency, could means many wakee or | |
3290 | * rapidly switch, use factor here will just help to automatically | |
3291 | * adjust the loose-degree, so bigger node will lead to more pull. | |
3292 | */ | |
3293 | if (p->wakee_flips > factor) { | |
3294 | /* | |
3295 | * wakee is somewhat hot, it needs certain amount of cpu | |
3296 | * resource, so if waker is far more hot, prefer to leave | |
3297 | * it alone. | |
3298 | */ | |
3299 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
3300 | return 1; | |
3301 | } | |
3302 | ||
3303 | return 0; | |
3304 | } | |
3305 | ||
c88d5910 | 3306 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 3307 | { |
e37b6a7b | 3308 | s64 this_load, load; |
c88d5910 | 3309 | int idx, this_cpu, prev_cpu; |
098fb9db | 3310 | unsigned long tl_per_task; |
c88d5910 | 3311 | struct task_group *tg; |
83378269 | 3312 | unsigned long weight; |
b3137bc8 | 3313 | int balanced; |
098fb9db | 3314 | |
62470419 MW |
3315 | /* |
3316 | * If we wake multiple tasks be careful to not bounce | |
3317 | * ourselves around too much. | |
3318 | */ | |
3319 | if (wake_wide(p)) | |
3320 | return 0; | |
3321 | ||
c88d5910 PZ |
3322 | idx = sd->wake_idx; |
3323 | this_cpu = smp_processor_id(); | |
3324 | prev_cpu = task_cpu(p); | |
3325 | load = source_load(prev_cpu, idx); | |
3326 | this_load = target_load(this_cpu, idx); | |
098fb9db | 3327 | |
b3137bc8 MG |
3328 | /* |
3329 | * If sync wakeup then subtract the (maximum possible) | |
3330 | * effect of the currently running task from the load | |
3331 | * of the current CPU: | |
3332 | */ | |
83378269 PZ |
3333 | if (sync) { |
3334 | tg = task_group(current); | |
3335 | weight = current->se.load.weight; | |
3336 | ||
c88d5910 | 3337 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
3338 | load += effective_load(tg, prev_cpu, 0, -weight); |
3339 | } | |
b3137bc8 | 3340 | |
83378269 PZ |
3341 | tg = task_group(p); |
3342 | weight = p->se.load.weight; | |
b3137bc8 | 3343 | |
71a29aa7 PZ |
3344 | /* |
3345 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
3346 | * due to the sync cause above having dropped this_load to 0, we'll |
3347 | * always have an imbalance, but there's really nothing you can do | |
3348 | * about that, so that's good too. | |
71a29aa7 PZ |
3349 | * |
3350 | * Otherwise check if either cpus are near enough in load to allow this | |
3351 | * task to be woken on this_cpu. | |
3352 | */ | |
e37b6a7b PT |
3353 | if (this_load > 0) { |
3354 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
3355 | |
3356 | this_eff_load = 100; | |
3357 | this_eff_load *= power_of(prev_cpu); | |
3358 | this_eff_load *= this_load + | |
3359 | effective_load(tg, this_cpu, weight, weight); | |
3360 | ||
3361 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
3362 | prev_eff_load *= power_of(this_cpu); | |
3363 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
3364 | ||
3365 | balanced = this_eff_load <= prev_eff_load; | |
3366 | } else | |
3367 | balanced = true; | |
b3137bc8 | 3368 | |
098fb9db | 3369 | /* |
4ae7d5ce IM |
3370 | * If the currently running task will sleep within |
3371 | * a reasonable amount of time then attract this newly | |
3372 | * woken task: | |
098fb9db | 3373 | */ |
2fb7635c PZ |
3374 | if (sync && balanced) |
3375 | return 1; | |
098fb9db | 3376 | |
41acab88 | 3377 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
3378 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
3379 | ||
c88d5910 PZ |
3380 | if (balanced || |
3381 | (this_load <= load && | |
3382 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
3383 | /* |
3384 | * This domain has SD_WAKE_AFFINE and | |
3385 | * p is cache cold in this domain, and | |
3386 | * there is no bad imbalance. | |
3387 | */ | |
c88d5910 | 3388 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 3389 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
3390 | |
3391 | return 1; | |
3392 | } | |
3393 | return 0; | |
3394 | } | |
3395 | ||
aaee1203 PZ |
3396 | /* |
3397 | * find_idlest_group finds and returns the least busy CPU group within the | |
3398 | * domain. | |
3399 | */ | |
3400 | static struct sched_group * | |
78e7ed53 | 3401 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 3402 | int this_cpu, int load_idx) |
e7693a36 | 3403 | { |
b3bd3de6 | 3404 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 3405 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 3406 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 3407 | |
aaee1203 PZ |
3408 | do { |
3409 | unsigned long load, avg_load; | |
3410 | int local_group; | |
3411 | int i; | |
e7693a36 | 3412 | |
aaee1203 PZ |
3413 | /* Skip over this group if it has no CPUs allowed */ |
3414 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 3415 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
3416 | continue; |
3417 | ||
3418 | local_group = cpumask_test_cpu(this_cpu, | |
3419 | sched_group_cpus(group)); | |
3420 | ||
3421 | /* Tally up the load of all CPUs in the group */ | |
3422 | avg_load = 0; | |
3423 | ||
3424 | for_each_cpu(i, sched_group_cpus(group)) { | |
3425 | /* Bias balancing toward cpus of our domain */ | |
3426 | if (local_group) | |
3427 | load = source_load(i, load_idx); | |
3428 | else | |
3429 | load = target_load(i, load_idx); | |
3430 | ||
3431 | avg_load += load; | |
3432 | } | |
3433 | ||
3434 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3435 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
3436 | |
3437 | if (local_group) { | |
3438 | this_load = avg_load; | |
aaee1203 PZ |
3439 | } else if (avg_load < min_load) { |
3440 | min_load = avg_load; | |
3441 | idlest = group; | |
3442 | } | |
3443 | } while (group = group->next, group != sd->groups); | |
3444 | ||
3445 | if (!idlest || 100*this_load < imbalance*min_load) | |
3446 | return NULL; | |
3447 | return idlest; | |
3448 | } | |
3449 | ||
3450 | /* | |
3451 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
3452 | */ | |
3453 | static int | |
3454 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
3455 | { | |
3456 | unsigned long load, min_load = ULONG_MAX; | |
3457 | int idlest = -1; | |
3458 | int i; | |
3459 | ||
3460 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 3461 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
3462 | load = weighted_cpuload(i); |
3463 | ||
3464 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
3465 | min_load = load; | |
3466 | idlest = i; | |
e7693a36 GH |
3467 | } |
3468 | } | |
3469 | ||
aaee1203 PZ |
3470 | return idlest; |
3471 | } | |
e7693a36 | 3472 | |
a50bde51 PZ |
3473 | /* |
3474 | * Try and locate an idle CPU in the sched_domain. | |
3475 | */ | |
99bd5e2f | 3476 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 3477 | { |
99bd5e2f | 3478 | struct sched_domain *sd; |
37407ea7 | 3479 | struct sched_group *sg; |
e0a79f52 | 3480 | int i = task_cpu(p); |
a50bde51 | 3481 | |
e0a79f52 MG |
3482 | if (idle_cpu(target)) |
3483 | return target; | |
99bd5e2f SS |
3484 | |
3485 | /* | |
e0a79f52 | 3486 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 3487 | */ |
e0a79f52 MG |
3488 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
3489 | return i; | |
a50bde51 PZ |
3490 | |
3491 | /* | |
37407ea7 | 3492 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 3493 | */ |
518cd623 | 3494 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 3495 | for_each_lower_domain(sd) { |
37407ea7 LT |
3496 | sg = sd->groups; |
3497 | do { | |
3498 | if (!cpumask_intersects(sched_group_cpus(sg), | |
3499 | tsk_cpus_allowed(p))) | |
3500 | goto next; | |
3501 | ||
3502 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 3503 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
3504 | goto next; |
3505 | } | |
970e1789 | 3506 | |
37407ea7 LT |
3507 | target = cpumask_first_and(sched_group_cpus(sg), |
3508 | tsk_cpus_allowed(p)); | |
3509 | goto done; | |
3510 | next: | |
3511 | sg = sg->next; | |
3512 | } while (sg != sd->groups); | |
3513 | } | |
3514 | done: | |
a50bde51 PZ |
3515 | return target; |
3516 | } | |
3517 | ||
aaee1203 PZ |
3518 | /* |
3519 | * sched_balance_self: balance the current task (running on cpu) in domains | |
3520 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
3521 | * SD_BALANCE_EXEC. | |
3522 | * | |
3523 | * Balance, ie. select the least loaded group. | |
3524 | * | |
3525 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
3526 | * | |
3527 | * preempt must be disabled. | |
3528 | */ | |
0017d735 | 3529 | static int |
7608dec2 | 3530 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 3531 | { |
29cd8bae | 3532 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
3533 | int cpu = smp_processor_id(); |
3534 | int prev_cpu = task_cpu(p); | |
3535 | int new_cpu = cpu; | |
99bd5e2f | 3536 | int want_affine = 0; |
5158f4e4 | 3537 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 3538 | |
29baa747 | 3539 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
3540 | return prev_cpu; |
3541 | ||
0763a660 | 3542 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 3543 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
3544 | want_affine = 1; |
3545 | new_cpu = prev_cpu; | |
3546 | } | |
aaee1203 | 3547 | |
dce840a0 | 3548 | rcu_read_lock(); |
aaee1203 | 3549 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
3550 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
3551 | continue; | |
3552 | ||
fe3bcfe1 | 3553 | /* |
99bd5e2f SS |
3554 | * If both cpu and prev_cpu are part of this domain, |
3555 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 3556 | */ |
99bd5e2f SS |
3557 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
3558 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
3559 | affine_sd = tmp; | |
29cd8bae | 3560 | break; |
f03542a7 | 3561 | } |
29cd8bae | 3562 | |
f03542a7 | 3563 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
3564 | sd = tmp; |
3565 | } | |
3566 | ||
8b911acd | 3567 | if (affine_sd) { |
f03542a7 | 3568 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
3569 | prev_cpu = cpu; |
3570 | ||
3571 | new_cpu = select_idle_sibling(p, prev_cpu); | |
3572 | goto unlock; | |
8b911acd | 3573 | } |
e7693a36 | 3574 | |
aaee1203 | 3575 | while (sd) { |
5158f4e4 | 3576 | int load_idx = sd->forkexec_idx; |
aaee1203 | 3577 | struct sched_group *group; |
c88d5910 | 3578 | int weight; |
098fb9db | 3579 | |
0763a660 | 3580 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
3581 | sd = sd->child; |
3582 | continue; | |
3583 | } | |
098fb9db | 3584 | |
5158f4e4 PZ |
3585 | if (sd_flag & SD_BALANCE_WAKE) |
3586 | load_idx = sd->wake_idx; | |
098fb9db | 3587 | |
5158f4e4 | 3588 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
3589 | if (!group) { |
3590 | sd = sd->child; | |
3591 | continue; | |
3592 | } | |
4ae7d5ce | 3593 | |
d7c33c49 | 3594 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
3595 | if (new_cpu == -1 || new_cpu == cpu) { |
3596 | /* Now try balancing at a lower domain level of cpu */ | |
3597 | sd = sd->child; | |
3598 | continue; | |
e7693a36 | 3599 | } |
aaee1203 PZ |
3600 | |
3601 | /* Now try balancing at a lower domain level of new_cpu */ | |
3602 | cpu = new_cpu; | |
669c55e9 | 3603 | weight = sd->span_weight; |
aaee1203 PZ |
3604 | sd = NULL; |
3605 | for_each_domain(cpu, tmp) { | |
669c55e9 | 3606 | if (weight <= tmp->span_weight) |
aaee1203 | 3607 | break; |
0763a660 | 3608 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
3609 | sd = tmp; |
3610 | } | |
3611 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 3612 | } |
dce840a0 PZ |
3613 | unlock: |
3614 | rcu_read_unlock(); | |
e7693a36 | 3615 | |
c88d5910 | 3616 | return new_cpu; |
e7693a36 | 3617 | } |
0a74bef8 PT |
3618 | |
3619 | /* | |
3620 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
3621 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
3622 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
3623 | * other assumptions, including the state of rq->lock, should be made. | |
3624 | */ | |
3625 | static void | |
3626 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
3627 | { | |
aff3e498 PT |
3628 | struct sched_entity *se = &p->se; |
3629 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3630 | ||
3631 | /* | |
3632 | * Load tracking: accumulate removed load so that it can be processed | |
3633 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
3634 | * to blocked load iff they have a positive decay-count. It can never | |
3635 | * be negative here since on-rq tasks have decay-count == 0. | |
3636 | */ | |
3637 | if (se->avg.decay_count) { | |
3638 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
3639 | atomic_long_add(se->avg.load_avg_contrib, |
3640 | &cfs_rq->removed_load); | |
aff3e498 | 3641 | } |
0a74bef8 | 3642 | } |
e7693a36 GH |
3643 | #endif /* CONFIG_SMP */ |
3644 | ||
e52fb7c0 PZ |
3645 | static unsigned long |
3646 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
3647 | { |
3648 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
3649 | ||
3650 | /* | |
e52fb7c0 PZ |
3651 | * Since its curr running now, convert the gran from real-time |
3652 | * to virtual-time in his units. | |
13814d42 MG |
3653 | * |
3654 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
3655 | * they get preempted easier. That is, if 'se' < 'curr' then | |
3656 | * the resulting gran will be larger, therefore penalizing the | |
3657 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
3658 | * be smaller, again penalizing the lighter task. | |
3659 | * | |
3660 | * This is especially important for buddies when the leftmost | |
3661 | * task is higher priority than the buddy. | |
0bbd3336 | 3662 | */ |
f4ad9bd2 | 3663 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
3664 | } |
3665 | ||
464b7527 PZ |
3666 | /* |
3667 | * Should 'se' preempt 'curr'. | |
3668 | * | |
3669 | * |s1 | |
3670 | * |s2 | |
3671 | * |s3 | |
3672 | * g | |
3673 | * |<--->|c | |
3674 | * | |
3675 | * w(c, s1) = -1 | |
3676 | * w(c, s2) = 0 | |
3677 | * w(c, s3) = 1 | |
3678 | * | |
3679 | */ | |
3680 | static int | |
3681 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
3682 | { | |
3683 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
3684 | ||
3685 | if (vdiff <= 0) | |
3686 | return -1; | |
3687 | ||
e52fb7c0 | 3688 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
3689 | if (vdiff > gran) |
3690 | return 1; | |
3691 | ||
3692 | return 0; | |
3693 | } | |
3694 | ||
02479099 PZ |
3695 | static void set_last_buddy(struct sched_entity *se) |
3696 | { | |
69c80f3e VP |
3697 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3698 | return; | |
3699 | ||
3700 | for_each_sched_entity(se) | |
3701 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
3702 | } |
3703 | ||
3704 | static void set_next_buddy(struct sched_entity *se) | |
3705 | { | |
69c80f3e VP |
3706 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3707 | return; | |
3708 | ||
3709 | for_each_sched_entity(se) | |
3710 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
3711 | } |
3712 | ||
ac53db59 RR |
3713 | static void set_skip_buddy(struct sched_entity *se) |
3714 | { | |
69c80f3e VP |
3715 | for_each_sched_entity(se) |
3716 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
3717 | } |
3718 | ||
bf0f6f24 IM |
3719 | /* |
3720 | * Preempt the current task with a newly woken task if needed: | |
3721 | */ | |
5a9b86f6 | 3722 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
3723 | { |
3724 | struct task_struct *curr = rq->curr; | |
8651a86c | 3725 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 3726 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 3727 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 3728 | int next_buddy_marked = 0; |
bf0f6f24 | 3729 | |
4ae7d5ce IM |
3730 | if (unlikely(se == pse)) |
3731 | return; | |
3732 | ||
5238cdd3 | 3733 | /* |
ddcdf6e7 | 3734 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
3735 | * unconditionally check_prempt_curr() after an enqueue (which may have |
3736 | * lead to a throttle). This both saves work and prevents false | |
3737 | * next-buddy nomination below. | |
3738 | */ | |
3739 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3740 | return; | |
3741 | ||
2f36825b | 3742 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3743 | set_next_buddy(pse); |
2f36825b VP |
3744 | next_buddy_marked = 1; |
3745 | } | |
57fdc26d | 3746 | |
aec0a514 BR |
3747 | /* |
3748 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3749 | * wake up path. | |
5238cdd3 PT |
3750 | * |
3751 | * Note: this also catches the edge-case of curr being in a throttled | |
3752 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3753 | * enqueue of curr) will have resulted in resched being set. This | |
3754 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3755 | * below. | |
aec0a514 BR |
3756 | */ |
3757 | if (test_tsk_need_resched(curr)) | |
3758 | return; | |
3759 | ||
a2f5c9ab DH |
3760 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3761 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3762 | likely(p->policy != SCHED_IDLE)) | |
3763 | goto preempt; | |
3764 | ||
91c234b4 | 3765 | /* |
a2f5c9ab DH |
3766 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3767 | * is driven by the tick): | |
91c234b4 | 3768 | */ |
8ed92e51 | 3769 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 3770 | return; |
bf0f6f24 | 3771 | |
464b7527 | 3772 | find_matching_se(&se, &pse); |
9bbd7374 | 3773 | update_curr(cfs_rq_of(se)); |
002f128b | 3774 | BUG_ON(!pse); |
2f36825b VP |
3775 | if (wakeup_preempt_entity(se, pse) == 1) { |
3776 | /* | |
3777 | * Bias pick_next to pick the sched entity that is | |
3778 | * triggering this preemption. | |
3779 | */ | |
3780 | if (!next_buddy_marked) | |
3781 | set_next_buddy(pse); | |
3a7e73a2 | 3782 | goto preempt; |
2f36825b | 3783 | } |
464b7527 | 3784 | |
3a7e73a2 | 3785 | return; |
a65ac745 | 3786 | |
3a7e73a2 PZ |
3787 | preempt: |
3788 | resched_task(curr); | |
3789 | /* | |
3790 | * Only set the backward buddy when the current task is still | |
3791 | * on the rq. This can happen when a wakeup gets interleaved | |
3792 | * with schedule on the ->pre_schedule() or idle_balance() | |
3793 | * point, either of which can * drop the rq lock. | |
3794 | * | |
3795 | * Also, during early boot the idle thread is in the fair class, | |
3796 | * for obvious reasons its a bad idea to schedule back to it. | |
3797 | */ | |
3798 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3799 | return; | |
3800 | ||
3801 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3802 | set_last_buddy(se); | |
bf0f6f24 IM |
3803 | } |
3804 | ||
fb8d4724 | 3805 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3806 | { |
8f4d37ec | 3807 | struct task_struct *p; |
bf0f6f24 IM |
3808 | struct cfs_rq *cfs_rq = &rq->cfs; |
3809 | struct sched_entity *se; | |
3810 | ||
36ace27e | 3811 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3812 | return NULL; |
3813 | ||
3814 | do { | |
9948f4b2 | 3815 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3816 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3817 | cfs_rq = group_cfs_rq(se); |
3818 | } while (cfs_rq); | |
3819 | ||
8f4d37ec | 3820 | p = task_of(se); |
b39e66ea MG |
3821 | if (hrtick_enabled(rq)) |
3822 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
3823 | |
3824 | return p; | |
bf0f6f24 IM |
3825 | } |
3826 | ||
3827 | /* | |
3828 | * Account for a descheduled task: | |
3829 | */ | |
31ee529c | 3830 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
3831 | { |
3832 | struct sched_entity *se = &prev->se; | |
3833 | struct cfs_rq *cfs_rq; | |
3834 | ||
3835 | for_each_sched_entity(se) { | |
3836 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 3837 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
3838 | } |
3839 | } | |
3840 | ||
ac53db59 RR |
3841 | /* |
3842 | * sched_yield() is very simple | |
3843 | * | |
3844 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
3845 | */ | |
3846 | static void yield_task_fair(struct rq *rq) | |
3847 | { | |
3848 | struct task_struct *curr = rq->curr; | |
3849 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
3850 | struct sched_entity *se = &curr->se; | |
3851 | ||
3852 | /* | |
3853 | * Are we the only task in the tree? | |
3854 | */ | |
3855 | if (unlikely(rq->nr_running == 1)) | |
3856 | return; | |
3857 | ||
3858 | clear_buddies(cfs_rq, se); | |
3859 | ||
3860 | if (curr->policy != SCHED_BATCH) { | |
3861 | update_rq_clock(rq); | |
3862 | /* | |
3863 | * Update run-time statistics of the 'current'. | |
3864 | */ | |
3865 | update_curr(cfs_rq); | |
916671c0 MG |
3866 | /* |
3867 | * Tell update_rq_clock() that we've just updated, | |
3868 | * so we don't do microscopic update in schedule() | |
3869 | * and double the fastpath cost. | |
3870 | */ | |
3871 | rq->skip_clock_update = 1; | |
ac53db59 RR |
3872 | } |
3873 | ||
3874 | set_skip_buddy(se); | |
3875 | } | |
3876 | ||
d95f4122 MG |
3877 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
3878 | { | |
3879 | struct sched_entity *se = &p->se; | |
3880 | ||
5238cdd3 PT |
3881 | /* throttled hierarchies are not runnable */ |
3882 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
3883 | return false; |
3884 | ||
3885 | /* Tell the scheduler that we'd really like pse to run next. */ | |
3886 | set_next_buddy(se); | |
3887 | ||
d95f4122 MG |
3888 | yield_task_fair(rq); |
3889 | ||
3890 | return true; | |
3891 | } | |
3892 | ||
681f3e68 | 3893 | #ifdef CONFIG_SMP |
bf0f6f24 | 3894 | /************************************************** |
e9c84cb8 PZ |
3895 | * Fair scheduling class load-balancing methods. |
3896 | * | |
3897 | * BASICS | |
3898 | * | |
3899 | * The purpose of load-balancing is to achieve the same basic fairness the | |
3900 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
3901 | * time to each task. This is expressed in the following equation: | |
3902 | * | |
3903 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
3904 | * | |
3905 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
3906 | * W_i,0 is defined as: | |
3907 | * | |
3908 | * W_i,0 = \Sum_j w_i,j (2) | |
3909 | * | |
3910 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
3911 | * is derived from the nice value as per prio_to_weight[]. | |
3912 | * | |
3913 | * The weight average is an exponential decay average of the instantaneous | |
3914 | * weight: | |
3915 | * | |
3916 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
3917 | * | |
3918 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
3919 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
3920 | * can also include other factors [XXX]. | |
3921 | * | |
3922 | * To achieve this balance we define a measure of imbalance which follows | |
3923 | * directly from (1): | |
3924 | * | |
3925 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
3926 | * | |
3927 | * We them move tasks around to minimize the imbalance. In the continuous | |
3928 | * function space it is obvious this converges, in the discrete case we get | |
3929 | * a few fun cases generally called infeasible weight scenarios. | |
3930 | * | |
3931 | * [XXX expand on: | |
3932 | * - infeasible weights; | |
3933 | * - local vs global optima in the discrete case. ] | |
3934 | * | |
3935 | * | |
3936 | * SCHED DOMAINS | |
3937 | * | |
3938 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
3939 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
3940 | * topology where each level pairs two lower groups (or better). This results | |
3941 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
3942 | * tree to only the first of the previous level and we decrease the frequency | |
3943 | * of load-balance at each level inv. proportional to the number of cpus in | |
3944 | * the groups. | |
3945 | * | |
3946 | * This yields: | |
3947 | * | |
3948 | * log_2 n 1 n | |
3949 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
3950 | * i = 0 2^i 2^i | |
3951 | * `- size of each group | |
3952 | * | | `- number of cpus doing load-balance | |
3953 | * | `- freq | |
3954 | * `- sum over all levels | |
3955 | * | |
3956 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
3957 | * this makes (5) the runtime complexity of the balancer. | |
3958 | * | |
3959 | * An important property here is that each CPU is still (indirectly) connected | |
3960 | * to every other cpu in at most O(log n) steps: | |
3961 | * | |
3962 | * The adjacency matrix of the resulting graph is given by: | |
3963 | * | |
3964 | * log_2 n | |
3965 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
3966 | * k = 0 | |
3967 | * | |
3968 | * And you'll find that: | |
3969 | * | |
3970 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
3971 | * | |
3972 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
3973 | * The task movement gives a factor of O(m), giving a convergence complexity | |
3974 | * of: | |
3975 | * | |
3976 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
3977 | * | |
3978 | * | |
3979 | * WORK CONSERVING | |
3980 | * | |
3981 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
3982 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
3983 | * tree itself instead of relying on other CPUs to bring it work. | |
3984 | * | |
3985 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
3986 | * time. | |
3987 | * | |
3988 | * [XXX more?] | |
3989 | * | |
3990 | * | |
3991 | * CGROUPS | |
3992 | * | |
3993 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
3994 | * | |
3995 | * s_k,i | |
3996 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
3997 | * S_k | |
3998 | * | |
3999 | * Where | |
4000 | * | |
4001 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4002 | * | |
4003 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4004 | * | |
4005 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4006 | * property. | |
4007 | * | |
4008 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4009 | * rewrite all of this once again.] | |
4010 | */ | |
bf0f6f24 | 4011 | |
ed387b78 HS |
4012 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4013 | ||
ddcdf6e7 | 4014 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4015 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4016 | #define LBF_DST_PINNED 0x04 |
4017 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4018 | |
4019 | struct lb_env { | |
4020 | struct sched_domain *sd; | |
4021 | ||
ddcdf6e7 | 4022 | struct rq *src_rq; |
85c1e7da | 4023 | int src_cpu; |
ddcdf6e7 PZ |
4024 | |
4025 | int dst_cpu; | |
4026 | struct rq *dst_rq; | |
4027 | ||
88b8dac0 SV |
4028 | struct cpumask *dst_grpmask; |
4029 | int new_dst_cpu; | |
ddcdf6e7 | 4030 | enum cpu_idle_type idle; |
bd939f45 | 4031 | long imbalance; |
b9403130 MW |
4032 | /* The set of CPUs under consideration for load-balancing */ |
4033 | struct cpumask *cpus; | |
4034 | ||
ddcdf6e7 | 4035 | unsigned int flags; |
367456c7 PZ |
4036 | |
4037 | unsigned int loop; | |
4038 | unsigned int loop_break; | |
4039 | unsigned int loop_max; | |
ddcdf6e7 PZ |
4040 | }; |
4041 | ||
1e3c88bd | 4042 | /* |
ddcdf6e7 | 4043 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
4044 | * Both runqueues must be locked. |
4045 | */ | |
ddcdf6e7 | 4046 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 4047 | { |
ddcdf6e7 PZ |
4048 | deactivate_task(env->src_rq, p, 0); |
4049 | set_task_cpu(p, env->dst_cpu); | |
4050 | activate_task(env->dst_rq, p, 0); | |
4051 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
4052 | } |
4053 | ||
029632fb PZ |
4054 | /* |
4055 | * Is this task likely cache-hot: | |
4056 | */ | |
4057 | static int | |
4058 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
4059 | { | |
4060 | s64 delta; | |
4061 | ||
4062 | if (p->sched_class != &fair_sched_class) | |
4063 | return 0; | |
4064 | ||
4065 | if (unlikely(p->policy == SCHED_IDLE)) | |
4066 | return 0; | |
4067 | ||
4068 | /* | |
4069 | * Buddy candidates are cache hot: | |
4070 | */ | |
4071 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
4072 | (&p->se == cfs_rq_of(&p->se)->next || | |
4073 | &p->se == cfs_rq_of(&p->se)->last)) | |
4074 | return 1; | |
4075 | ||
4076 | if (sysctl_sched_migration_cost == -1) | |
4077 | return 1; | |
4078 | if (sysctl_sched_migration_cost == 0) | |
4079 | return 0; | |
4080 | ||
4081 | delta = now - p->se.exec_start; | |
4082 | ||
4083 | return delta < (s64)sysctl_sched_migration_cost; | |
4084 | } | |
4085 | ||
3a7053b3 MG |
4086 | #ifdef CONFIG_NUMA_BALANCING |
4087 | /* Returns true if the destination node has incurred more faults */ | |
4088 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
4089 | { | |
4090 | int src_nid, dst_nid; | |
4091 | ||
4092 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || | |
4093 | !(env->sd->flags & SD_NUMA)) { | |
4094 | return false; | |
4095 | } | |
4096 | ||
4097 | src_nid = cpu_to_node(env->src_cpu); | |
4098 | dst_nid = cpu_to_node(env->dst_cpu); | |
4099 | ||
4100 | if (src_nid == dst_nid || | |
4101 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4102 | return false; | |
4103 | ||
4104 | if (dst_nid == p->numa_preferred_nid || | |
4105 | p->numa_faults[dst_nid] > p->numa_faults[src_nid]) | |
4106 | return true; | |
4107 | ||
4108 | return false; | |
4109 | } | |
4110 | #else | |
4111 | static inline bool migrate_improves_locality(struct task_struct *p, | |
4112 | struct lb_env *env) | |
4113 | { | |
4114 | return false; | |
4115 | } | |
4116 | #endif | |
4117 | ||
1e3c88bd PZ |
4118 | /* |
4119 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
4120 | */ | |
4121 | static | |
8e45cb54 | 4122 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
4123 | { |
4124 | int tsk_cache_hot = 0; | |
4125 | /* | |
4126 | * We do not migrate tasks that are: | |
d3198084 | 4127 | * 1) throttled_lb_pair, or |
1e3c88bd | 4128 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
4129 | * 3) running (obviously), or |
4130 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 4131 | */ |
d3198084 JK |
4132 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
4133 | return 0; | |
4134 | ||
ddcdf6e7 | 4135 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 4136 | int cpu; |
88b8dac0 | 4137 | |
41acab88 | 4138 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 4139 | |
6263322c PZ |
4140 | env->flags |= LBF_SOME_PINNED; |
4141 | ||
88b8dac0 SV |
4142 | /* |
4143 | * Remember if this task can be migrated to any other cpu in | |
4144 | * our sched_group. We may want to revisit it if we couldn't | |
4145 | * meet load balance goals by pulling other tasks on src_cpu. | |
4146 | * | |
4147 | * Also avoid computing new_dst_cpu if we have already computed | |
4148 | * one in current iteration. | |
4149 | */ | |
6263322c | 4150 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
4151 | return 0; |
4152 | ||
e02e60c1 JK |
4153 | /* Prevent to re-select dst_cpu via env's cpus */ |
4154 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
4155 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 4156 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
4157 | env->new_dst_cpu = cpu; |
4158 | break; | |
4159 | } | |
88b8dac0 | 4160 | } |
e02e60c1 | 4161 | |
1e3c88bd PZ |
4162 | return 0; |
4163 | } | |
88b8dac0 SV |
4164 | |
4165 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 4166 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 4167 | |
ddcdf6e7 | 4168 | if (task_running(env->src_rq, p)) { |
41acab88 | 4169 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
4170 | return 0; |
4171 | } | |
4172 | ||
4173 | /* | |
4174 | * Aggressive migration if: | |
3a7053b3 MG |
4175 | * 1) destination numa is preferred |
4176 | * 2) task is cache cold, or | |
4177 | * 3) too many balance attempts have failed. | |
1e3c88bd | 4178 | */ |
78becc27 | 4179 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
3a7053b3 MG |
4180 | |
4181 | if (migrate_improves_locality(p, env)) { | |
4182 | #ifdef CONFIG_SCHEDSTATS | |
4183 | if (tsk_cache_hot) { | |
4184 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
4185 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
4186 | } | |
4187 | #endif | |
4188 | return 1; | |
4189 | } | |
4190 | ||
1e3c88bd | 4191 | if (!tsk_cache_hot || |
8e45cb54 | 4192 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 4193 | |
1e3c88bd | 4194 | if (tsk_cache_hot) { |
8e45cb54 | 4195 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 4196 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 4197 | } |
4e2dcb73 | 4198 | |
1e3c88bd PZ |
4199 | return 1; |
4200 | } | |
4201 | ||
4e2dcb73 ZH |
4202 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
4203 | return 0; | |
1e3c88bd PZ |
4204 | } |
4205 | ||
897c395f PZ |
4206 | /* |
4207 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
4208 | * part of active balancing operations within "domain". | |
4209 | * Returns 1 if successful and 0 otherwise. | |
4210 | * | |
4211 | * Called with both runqueues locked. | |
4212 | */ | |
8e45cb54 | 4213 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
4214 | { |
4215 | struct task_struct *p, *n; | |
897c395f | 4216 | |
367456c7 | 4217 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
4218 | if (!can_migrate_task(p, env)) |
4219 | continue; | |
897c395f | 4220 | |
367456c7 PZ |
4221 | move_task(p, env); |
4222 | /* | |
4223 | * Right now, this is only the second place move_task() | |
4224 | * is called, so we can safely collect move_task() | |
4225 | * stats here rather than inside move_task(). | |
4226 | */ | |
4227 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4228 | return 1; | |
897c395f | 4229 | } |
897c395f PZ |
4230 | return 0; |
4231 | } | |
4232 | ||
367456c7 PZ |
4233 | static unsigned long task_h_load(struct task_struct *p); |
4234 | ||
eb95308e PZ |
4235 | static const unsigned int sched_nr_migrate_break = 32; |
4236 | ||
5d6523eb | 4237 | /* |
bd939f45 | 4238 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4239 | * this_rq, as part of a balancing operation within domain "sd". |
4240 | * Returns 1 if successful and 0 otherwise. | |
4241 | * | |
4242 | * Called with both runqueues locked. | |
4243 | */ | |
4244 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4245 | { |
5d6523eb PZ |
4246 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4247 | struct task_struct *p; | |
367456c7 PZ |
4248 | unsigned long load; |
4249 | int pulled = 0; | |
1e3c88bd | 4250 | |
bd939f45 | 4251 | if (env->imbalance <= 0) |
5d6523eb | 4252 | return 0; |
1e3c88bd | 4253 | |
5d6523eb PZ |
4254 | while (!list_empty(tasks)) { |
4255 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 4256 | |
367456c7 PZ |
4257 | env->loop++; |
4258 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 4259 | if (env->loop > env->loop_max) |
367456c7 | 4260 | break; |
5d6523eb PZ |
4261 | |
4262 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 4263 | if (env->loop > env->loop_break) { |
eb95308e | 4264 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 4265 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 4266 | break; |
a195f004 | 4267 | } |
1e3c88bd | 4268 | |
d3198084 | 4269 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
4270 | goto next; |
4271 | ||
4272 | load = task_h_load(p); | |
5d6523eb | 4273 | |
eb95308e | 4274 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
4275 | goto next; |
4276 | ||
bd939f45 | 4277 | if ((load / 2) > env->imbalance) |
367456c7 | 4278 | goto next; |
1e3c88bd | 4279 | |
ddcdf6e7 | 4280 | move_task(p, env); |
ee00e66f | 4281 | pulled++; |
bd939f45 | 4282 | env->imbalance -= load; |
1e3c88bd PZ |
4283 | |
4284 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
4285 | /* |
4286 | * NEWIDLE balancing is a source of latency, so preemptible | |
4287 | * kernels will stop after the first task is pulled to minimize | |
4288 | * the critical section. | |
4289 | */ | |
5d6523eb | 4290 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 4291 | break; |
1e3c88bd PZ |
4292 | #endif |
4293 | ||
ee00e66f PZ |
4294 | /* |
4295 | * We only want to steal up to the prescribed amount of | |
4296 | * weighted load. | |
4297 | */ | |
bd939f45 | 4298 | if (env->imbalance <= 0) |
ee00e66f | 4299 | break; |
367456c7 PZ |
4300 | |
4301 | continue; | |
4302 | next: | |
5d6523eb | 4303 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 4304 | } |
5d6523eb | 4305 | |
1e3c88bd | 4306 | /* |
ddcdf6e7 PZ |
4307 | * Right now, this is one of only two places move_task() is called, |
4308 | * so we can safely collect move_task() stats here rather than | |
4309 | * inside move_task(). | |
1e3c88bd | 4310 | */ |
8e45cb54 | 4311 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 4312 | |
5d6523eb | 4313 | return pulled; |
1e3c88bd PZ |
4314 | } |
4315 | ||
230059de | 4316 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
4317 | /* |
4318 | * update tg->load_weight by folding this cpu's load_avg | |
4319 | */ | |
48a16753 | 4320 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 4321 | { |
48a16753 PT |
4322 | struct sched_entity *se = tg->se[cpu]; |
4323 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 4324 | |
48a16753 PT |
4325 | /* throttled entities do not contribute to load */ |
4326 | if (throttled_hierarchy(cfs_rq)) | |
4327 | return; | |
9e3081ca | 4328 | |
aff3e498 | 4329 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 4330 | |
82958366 PT |
4331 | if (se) { |
4332 | update_entity_load_avg(se, 1); | |
4333 | /* | |
4334 | * We pivot on our runnable average having decayed to zero for | |
4335 | * list removal. This generally implies that all our children | |
4336 | * have also been removed (modulo rounding error or bandwidth | |
4337 | * control); however, such cases are rare and we can fix these | |
4338 | * at enqueue. | |
4339 | * | |
4340 | * TODO: fix up out-of-order children on enqueue. | |
4341 | */ | |
4342 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
4343 | list_del_leaf_cfs_rq(cfs_rq); | |
4344 | } else { | |
48a16753 | 4345 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
4346 | update_rq_runnable_avg(rq, rq->nr_running); |
4347 | } | |
9e3081ca PZ |
4348 | } |
4349 | ||
48a16753 | 4350 | static void update_blocked_averages(int cpu) |
9e3081ca | 4351 | { |
9e3081ca | 4352 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
4353 | struct cfs_rq *cfs_rq; |
4354 | unsigned long flags; | |
9e3081ca | 4355 | |
48a16753 PT |
4356 | raw_spin_lock_irqsave(&rq->lock, flags); |
4357 | update_rq_clock(rq); | |
9763b67f PZ |
4358 | /* |
4359 | * Iterates the task_group tree in a bottom up fashion, see | |
4360 | * list_add_leaf_cfs_rq() for details. | |
4361 | */ | |
64660c86 | 4362 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
4363 | /* |
4364 | * Note: We may want to consider periodically releasing | |
4365 | * rq->lock about these updates so that creating many task | |
4366 | * groups does not result in continually extending hold time. | |
4367 | */ | |
4368 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 4369 | } |
48a16753 PT |
4370 | |
4371 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
4372 | } |
4373 | ||
9763b67f | 4374 | /* |
68520796 | 4375 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
4376 | * This needs to be done in a top-down fashion because the load of a child |
4377 | * group is a fraction of its parents load. | |
4378 | */ | |
68520796 | 4379 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 4380 | { |
68520796 VD |
4381 | struct rq *rq = rq_of(cfs_rq); |
4382 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 4383 | unsigned long now = jiffies; |
68520796 | 4384 | unsigned long load; |
a35b6466 | 4385 | |
68520796 | 4386 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
4387 | return; |
4388 | ||
68520796 VD |
4389 | cfs_rq->h_load_next = NULL; |
4390 | for_each_sched_entity(se) { | |
4391 | cfs_rq = cfs_rq_of(se); | |
4392 | cfs_rq->h_load_next = se; | |
4393 | if (cfs_rq->last_h_load_update == now) | |
4394 | break; | |
4395 | } | |
a35b6466 | 4396 | |
68520796 | 4397 | if (!se) { |
7e3115ef | 4398 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
4399 | cfs_rq->last_h_load_update = now; |
4400 | } | |
4401 | ||
4402 | while ((se = cfs_rq->h_load_next) != NULL) { | |
4403 | load = cfs_rq->h_load; | |
4404 | load = div64_ul(load * se->avg.load_avg_contrib, | |
4405 | cfs_rq->runnable_load_avg + 1); | |
4406 | cfs_rq = group_cfs_rq(se); | |
4407 | cfs_rq->h_load = load; | |
4408 | cfs_rq->last_h_load_update = now; | |
4409 | } | |
9763b67f PZ |
4410 | } |
4411 | ||
367456c7 | 4412 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 4413 | { |
367456c7 | 4414 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 4415 | |
68520796 | 4416 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
4417 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
4418 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
4419 | } |
4420 | #else | |
48a16753 | 4421 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
4422 | { |
4423 | } | |
4424 | ||
367456c7 | 4425 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 4426 | { |
a003a25b | 4427 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 4428 | } |
230059de | 4429 | #endif |
1e3c88bd | 4430 | |
1e3c88bd | 4431 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
4432 | /* |
4433 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
4434 | */ | |
4435 | struct sg_lb_stats { | |
4436 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
4437 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 4438 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 4439 | unsigned long load_per_task; |
3ae11c90 | 4440 | unsigned long group_power; |
147c5fc2 PZ |
4441 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
4442 | unsigned int group_capacity; | |
4443 | unsigned int idle_cpus; | |
4444 | unsigned int group_weight; | |
1e3c88bd | 4445 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 4446 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
4447 | }; |
4448 | ||
56cf515b JK |
4449 | /* |
4450 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
4451 | * during load balancing. | |
4452 | */ | |
4453 | struct sd_lb_stats { | |
4454 | struct sched_group *busiest; /* Busiest group in this sd */ | |
4455 | struct sched_group *local; /* Local group in this sd */ | |
4456 | unsigned long total_load; /* Total load of all groups in sd */ | |
4457 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
4458 | unsigned long avg_load; /* Average load across all groups in sd */ | |
4459 | ||
56cf515b | 4460 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 4461 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
4462 | }; |
4463 | ||
147c5fc2 PZ |
4464 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
4465 | { | |
4466 | /* | |
4467 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
4468 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
4469 | * We must however clear busiest_stat::avg_load because | |
4470 | * update_sd_pick_busiest() reads this before assignment. | |
4471 | */ | |
4472 | *sds = (struct sd_lb_stats){ | |
4473 | .busiest = NULL, | |
4474 | .local = NULL, | |
4475 | .total_load = 0UL, | |
4476 | .total_pwr = 0UL, | |
4477 | .busiest_stat = { | |
4478 | .avg_load = 0UL, | |
4479 | }, | |
4480 | }; | |
4481 | } | |
4482 | ||
1e3c88bd PZ |
4483 | /** |
4484 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
4485 | * @sd: The sched_domain whose load_idx is to be obtained. | |
4486 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
e69f6186 YB |
4487 | * |
4488 | * Return: The load index. | |
1e3c88bd PZ |
4489 | */ |
4490 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
4491 | enum cpu_idle_type idle) | |
4492 | { | |
4493 | int load_idx; | |
4494 | ||
4495 | switch (idle) { | |
4496 | case CPU_NOT_IDLE: | |
4497 | load_idx = sd->busy_idx; | |
4498 | break; | |
4499 | ||
4500 | case CPU_NEWLY_IDLE: | |
4501 | load_idx = sd->newidle_idx; | |
4502 | break; | |
4503 | default: | |
4504 | load_idx = sd->idle_idx; | |
4505 | break; | |
4506 | } | |
4507 | ||
4508 | return load_idx; | |
4509 | } | |
4510 | ||
15f803c9 | 4511 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 4512 | { |
1399fa78 | 4513 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
4514 | } |
4515 | ||
4516 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
4517 | { | |
4518 | return default_scale_freq_power(sd, cpu); | |
4519 | } | |
4520 | ||
15f803c9 | 4521 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 4522 | { |
669c55e9 | 4523 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
4524 | unsigned long smt_gain = sd->smt_gain; |
4525 | ||
4526 | smt_gain /= weight; | |
4527 | ||
4528 | return smt_gain; | |
4529 | } | |
4530 | ||
4531 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
4532 | { | |
4533 | return default_scale_smt_power(sd, cpu); | |
4534 | } | |
4535 | ||
15f803c9 | 4536 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
4537 | { |
4538 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 4539 | u64 total, available, age_stamp, avg; |
1e3c88bd | 4540 | |
b654f7de PZ |
4541 | /* |
4542 | * Since we're reading these variables without serialization make sure | |
4543 | * we read them once before doing sanity checks on them. | |
4544 | */ | |
4545 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
4546 | avg = ACCESS_ONCE(rq->rt_avg); | |
4547 | ||
78becc27 | 4548 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 4549 | |
b654f7de | 4550 | if (unlikely(total < avg)) { |
aa483808 VP |
4551 | /* Ensures that power won't end up being negative */ |
4552 | available = 0; | |
4553 | } else { | |
b654f7de | 4554 | available = total - avg; |
aa483808 | 4555 | } |
1e3c88bd | 4556 | |
1399fa78 NR |
4557 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
4558 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 4559 | |
1399fa78 | 4560 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4561 | |
4562 | return div_u64(available, total); | |
4563 | } | |
4564 | ||
4565 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
4566 | { | |
669c55e9 | 4567 | unsigned long weight = sd->span_weight; |
1399fa78 | 4568 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
4569 | struct sched_group *sdg = sd->groups; |
4570 | ||
1e3c88bd PZ |
4571 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
4572 | if (sched_feat(ARCH_POWER)) | |
4573 | power *= arch_scale_smt_power(sd, cpu); | |
4574 | else | |
4575 | power *= default_scale_smt_power(sd, cpu); | |
4576 | ||
1399fa78 | 4577 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4578 | } |
4579 | ||
9c3f75cb | 4580 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
4581 | |
4582 | if (sched_feat(ARCH_POWER)) | |
4583 | power *= arch_scale_freq_power(sd, cpu); | |
4584 | else | |
4585 | power *= default_scale_freq_power(sd, cpu); | |
4586 | ||
1399fa78 | 4587 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 4588 | |
1e3c88bd | 4589 | power *= scale_rt_power(cpu); |
1399fa78 | 4590 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4591 | |
4592 | if (!power) | |
4593 | power = 1; | |
4594 | ||
e51fd5e2 | 4595 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 4596 | sdg->sgp->power = power; |
1e3c88bd PZ |
4597 | } |
4598 | ||
029632fb | 4599 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
4600 | { |
4601 | struct sched_domain *child = sd->child; | |
4602 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 4603 | unsigned long power, power_orig; |
4ec4412e VG |
4604 | unsigned long interval; |
4605 | ||
4606 | interval = msecs_to_jiffies(sd->balance_interval); | |
4607 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
4608 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
4609 | |
4610 | if (!child) { | |
4611 | update_cpu_power(sd, cpu); | |
4612 | return; | |
4613 | } | |
4614 | ||
863bffc8 | 4615 | power_orig = power = 0; |
1e3c88bd | 4616 | |
74a5ce20 PZ |
4617 | if (child->flags & SD_OVERLAP) { |
4618 | /* | |
4619 | * SD_OVERLAP domains cannot assume that child groups | |
4620 | * span the current group. | |
4621 | */ | |
4622 | ||
863bffc8 PZ |
4623 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
4624 | struct sched_group *sg = cpu_rq(cpu)->sd->groups; | |
4625 | ||
4626 | power_orig += sg->sgp->power_orig; | |
4627 | power += sg->sgp->power; | |
4628 | } | |
74a5ce20 PZ |
4629 | } else { |
4630 | /* | |
4631 | * !SD_OVERLAP domains can assume that child groups | |
4632 | * span the current group. | |
4633 | */ | |
4634 | ||
4635 | group = child->groups; | |
4636 | do { | |
863bffc8 | 4637 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
4638 | power += group->sgp->power; |
4639 | group = group->next; | |
4640 | } while (group != child->groups); | |
4641 | } | |
1e3c88bd | 4642 | |
863bffc8 PZ |
4643 | sdg->sgp->power_orig = power_orig; |
4644 | sdg->sgp->power = power; | |
1e3c88bd PZ |
4645 | } |
4646 | ||
9d5efe05 SV |
4647 | /* |
4648 | * Try and fix up capacity for tiny siblings, this is needed when | |
4649 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
4650 | * which on its own isn't powerful enough. | |
4651 | * | |
4652 | * See update_sd_pick_busiest() and check_asym_packing(). | |
4653 | */ | |
4654 | static inline int | |
4655 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
4656 | { | |
4657 | /* | |
1399fa78 | 4658 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 4659 | */ |
a6c75f2f | 4660 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
4661 | return 0; |
4662 | ||
4663 | /* | |
4664 | * If ~90% of the cpu_power is still there, we're good. | |
4665 | */ | |
9c3f75cb | 4666 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
4667 | return 1; |
4668 | ||
4669 | return 0; | |
4670 | } | |
4671 | ||
30ce5dab PZ |
4672 | /* |
4673 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
4674 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
4675 | * | |
4676 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
4677 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
4678 | * Something like: | |
4679 | * | |
4680 | * { 0 1 2 3 } { 4 5 6 7 } | |
4681 | * * * * * | |
4682 | * | |
4683 | * If we were to balance group-wise we'd place two tasks in the first group and | |
4684 | * two tasks in the second group. Clearly this is undesired as it will overload | |
4685 | * cpu 3 and leave one of the cpus in the second group unused. | |
4686 | * | |
4687 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
4688 | * by noticing the lower domain failed to reach balance and had difficulty |
4689 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
4690 | * |
4691 | * When this is so detected; this group becomes a candidate for busiest; see | |
4692 | * update_sd_pick_busiest(). And calculcate_imbalance() and | |
6263322c | 4693 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
4694 | * to create an effective group imbalance. |
4695 | * | |
4696 | * This is a somewhat tricky proposition since the next run might not find the | |
4697 | * group imbalance and decide the groups need to be balanced again. A most | |
4698 | * subtle and fragile situation. | |
4699 | */ | |
4700 | ||
6263322c | 4701 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 4702 | { |
6263322c | 4703 | return group->sgp->imbalance; |
30ce5dab PZ |
4704 | } |
4705 | ||
b37d9316 PZ |
4706 | /* |
4707 | * Compute the group capacity. | |
4708 | * | |
c61037e9 PZ |
4709 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
4710 | * first dividing out the smt factor and computing the actual number of cores | |
4711 | * and limit power unit capacity with that. | |
b37d9316 PZ |
4712 | */ |
4713 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
4714 | { | |
c61037e9 PZ |
4715 | unsigned int capacity, smt, cpus; |
4716 | unsigned int power, power_orig; | |
4717 | ||
4718 | power = group->sgp->power; | |
4719 | power_orig = group->sgp->power_orig; | |
4720 | cpus = group->group_weight; | |
b37d9316 | 4721 | |
c61037e9 PZ |
4722 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
4723 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
4724 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 4725 | |
c61037e9 | 4726 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
4727 | if (!capacity) |
4728 | capacity = fix_small_capacity(env->sd, group); | |
4729 | ||
4730 | return capacity; | |
4731 | } | |
4732 | ||
1e3c88bd PZ |
4733 | /** |
4734 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 4735 | * @env: The load balancing environment. |
1e3c88bd | 4736 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 4737 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 4738 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
4739 | * @sgs: variable to hold the statistics for this group. |
4740 | */ | |
bd939f45 PZ |
4741 | static inline void update_sg_lb_stats(struct lb_env *env, |
4742 | struct sched_group *group, int load_idx, | |
23f0d209 | 4743 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 4744 | { |
30ce5dab PZ |
4745 | unsigned long nr_running; |
4746 | unsigned long load; | |
bd939f45 | 4747 | int i; |
1e3c88bd | 4748 | |
b72ff13c PZ |
4749 | memset(sgs, 0, sizeof(*sgs)); |
4750 | ||
b9403130 | 4751 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
4752 | struct rq *rq = cpu_rq(i); |
4753 | ||
e44bc5c5 PZ |
4754 | nr_running = rq->nr_running; |
4755 | ||
1e3c88bd | 4756 | /* Bias balancing toward cpus of our domain */ |
6263322c | 4757 | if (local_group) |
04f733b4 | 4758 | load = target_load(i, load_idx); |
6263322c | 4759 | else |
1e3c88bd | 4760 | load = source_load(i, load_idx); |
1e3c88bd PZ |
4761 | |
4762 | sgs->group_load += load; | |
e44bc5c5 | 4763 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 4764 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
4765 | if (idle_cpu(i)) |
4766 | sgs->idle_cpus++; | |
1e3c88bd PZ |
4767 | } |
4768 | ||
1e3c88bd | 4769 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
4770 | sgs->group_power = group->sgp->power; |
4771 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 4772 | |
dd5feea1 | 4773 | if (sgs->sum_nr_running) |
38d0f770 | 4774 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 4775 | |
aae6d3dd | 4776 | sgs->group_weight = group->group_weight; |
fab47622 | 4777 | |
b37d9316 PZ |
4778 | sgs->group_imb = sg_imbalanced(group); |
4779 | sgs->group_capacity = sg_capacity(env, group); | |
4780 | ||
fab47622 NR |
4781 | if (sgs->group_capacity > sgs->sum_nr_running) |
4782 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
4783 | } |
4784 | ||
532cb4c4 MN |
4785 | /** |
4786 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 4787 | * @env: The load balancing environment. |
532cb4c4 MN |
4788 | * @sds: sched_domain statistics |
4789 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 4790 | * @sgs: sched_group statistics |
532cb4c4 MN |
4791 | * |
4792 | * Determine if @sg is a busier group than the previously selected | |
4793 | * busiest group. | |
e69f6186 YB |
4794 | * |
4795 | * Return: %true if @sg is a busier group than the previously selected | |
4796 | * busiest group. %false otherwise. | |
532cb4c4 | 4797 | */ |
bd939f45 | 4798 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
4799 | struct sd_lb_stats *sds, |
4800 | struct sched_group *sg, | |
bd939f45 | 4801 | struct sg_lb_stats *sgs) |
532cb4c4 | 4802 | { |
56cf515b | 4803 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
4804 | return false; |
4805 | ||
4806 | if (sgs->sum_nr_running > sgs->group_capacity) | |
4807 | return true; | |
4808 | ||
4809 | if (sgs->group_imb) | |
4810 | return true; | |
4811 | ||
4812 | /* | |
4813 | * ASYM_PACKING needs to move all the work to the lowest | |
4814 | * numbered CPUs in the group, therefore mark all groups | |
4815 | * higher than ourself as busy. | |
4816 | */ | |
bd939f45 PZ |
4817 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
4818 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
4819 | if (!sds->busiest) |
4820 | return true; | |
4821 | ||
4822 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
4823 | return true; | |
4824 | } | |
4825 | ||
4826 | return false; | |
4827 | } | |
4828 | ||
1e3c88bd | 4829 | /** |
461819ac | 4830 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 4831 | * @env: The load balancing environment. |
1e3c88bd PZ |
4832 | * @balance: Should we balance. |
4833 | * @sds: variable to hold the statistics for this sched_domain. | |
4834 | */ | |
bd939f45 | 4835 | static inline void update_sd_lb_stats(struct lb_env *env, |
23f0d209 | 4836 | struct sd_lb_stats *sds) |
1e3c88bd | 4837 | { |
bd939f45 PZ |
4838 | struct sched_domain *child = env->sd->child; |
4839 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 4840 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
4841 | int load_idx, prefer_sibling = 0; |
4842 | ||
4843 | if (child && child->flags & SD_PREFER_SIBLING) | |
4844 | prefer_sibling = 1; | |
4845 | ||
bd939f45 | 4846 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
4847 | |
4848 | do { | |
56cf515b | 4849 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
4850 | int local_group; |
4851 | ||
bd939f45 | 4852 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
4853 | if (local_group) { |
4854 | sds->local = sg; | |
4855 | sgs = &sds->local_stat; | |
b72ff13c PZ |
4856 | |
4857 | if (env->idle != CPU_NEWLY_IDLE || | |
4858 | time_after_eq(jiffies, sg->sgp->next_update)) | |
4859 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 4860 | } |
1e3c88bd | 4861 | |
56cf515b | 4862 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 4863 | |
b72ff13c PZ |
4864 | if (local_group) |
4865 | goto next_group; | |
4866 | ||
1e3c88bd PZ |
4867 | /* |
4868 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 4869 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
4870 | * and move all the excess tasks away. We lower the capacity |
4871 | * of a group only if the local group has the capacity to fit | |
4872 | * these excess tasks, i.e. nr_running < group_capacity. The | |
4873 | * extra check prevents the case where you always pull from the | |
4874 | * heaviest group when it is already under-utilized (possible | |
4875 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 4876 | */ |
b72ff13c PZ |
4877 | if (prefer_sibling && sds->local && |
4878 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 4879 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 4880 | |
b72ff13c | 4881 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 4882 | sds->busiest = sg; |
56cf515b | 4883 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
4884 | } |
4885 | ||
b72ff13c PZ |
4886 | next_group: |
4887 | /* Now, start updating sd_lb_stats */ | |
4888 | sds->total_load += sgs->group_load; | |
4889 | sds->total_pwr += sgs->group_power; | |
4890 | ||
532cb4c4 | 4891 | sg = sg->next; |
bd939f45 | 4892 | } while (sg != env->sd->groups); |
532cb4c4 MN |
4893 | } |
4894 | ||
532cb4c4 MN |
4895 | /** |
4896 | * check_asym_packing - Check to see if the group is packed into the | |
4897 | * sched doman. | |
4898 | * | |
4899 | * This is primarily intended to used at the sibling level. Some | |
4900 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
4901 | * case of POWER7, it can move to lower SMT modes only when higher | |
4902 | * threads are idle. When in lower SMT modes, the threads will | |
4903 | * perform better since they share less core resources. Hence when we | |
4904 | * have idle threads, we want them to be the higher ones. | |
4905 | * | |
4906 | * This packing function is run on idle threads. It checks to see if | |
4907 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
4908 | * CPU number than the packing function is being run on. Here we are | |
4909 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
4910 | * number. | |
4911 | * | |
e69f6186 | 4912 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
4913 | * this CPU. The amount of the imbalance is returned in *imbalance. |
4914 | * | |
cd96891d | 4915 | * @env: The load balancing environment. |
532cb4c4 | 4916 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 4917 | */ |
bd939f45 | 4918 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
4919 | { |
4920 | int busiest_cpu; | |
4921 | ||
bd939f45 | 4922 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
4923 | return 0; |
4924 | ||
4925 | if (!sds->busiest) | |
4926 | return 0; | |
4927 | ||
4928 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 4929 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
4930 | return 0; |
4931 | ||
bd939f45 | 4932 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
4933 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
4934 | SCHED_POWER_SCALE); | |
bd939f45 | 4935 | |
532cb4c4 | 4936 | return 1; |
1e3c88bd PZ |
4937 | } |
4938 | ||
4939 | /** | |
4940 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
4941 | * amongst the groups of a sched_domain, during | |
4942 | * load balancing. | |
cd96891d | 4943 | * @env: The load balancing environment. |
1e3c88bd | 4944 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 4945 | */ |
bd939f45 PZ |
4946 | static inline |
4947 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
4948 | { |
4949 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
4950 | unsigned int imbn = 2; | |
dd5feea1 | 4951 | unsigned long scaled_busy_load_per_task; |
56cf515b | 4952 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 4953 | |
56cf515b JK |
4954 | local = &sds->local_stat; |
4955 | busiest = &sds->busiest_stat; | |
1e3c88bd | 4956 | |
56cf515b JK |
4957 | if (!local->sum_nr_running) |
4958 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
4959 | else if (busiest->load_per_task > local->load_per_task) | |
4960 | imbn = 1; | |
dd5feea1 | 4961 | |
56cf515b JK |
4962 | scaled_busy_load_per_task = |
4963 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 4964 | busiest->group_power; |
56cf515b | 4965 | |
3029ede3 VD |
4966 | if (busiest->avg_load + scaled_busy_load_per_task >= |
4967 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 4968 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
4969 | return; |
4970 | } | |
4971 | ||
4972 | /* | |
4973 | * OK, we don't have enough imbalance to justify moving tasks, | |
4974 | * however we may be able to increase total CPU power used by | |
4975 | * moving them. | |
4976 | */ | |
4977 | ||
3ae11c90 | 4978 | pwr_now += busiest->group_power * |
56cf515b | 4979 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 4980 | pwr_now += local->group_power * |
56cf515b | 4981 | min(local->load_per_task, local->avg_load); |
1399fa78 | 4982 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
4983 | |
4984 | /* Amount of load we'd subtract */ | |
56cf515b | 4985 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 4986 | busiest->group_power; |
56cf515b | 4987 | if (busiest->avg_load > tmp) { |
3ae11c90 | 4988 | pwr_move += busiest->group_power * |
56cf515b JK |
4989 | min(busiest->load_per_task, |
4990 | busiest->avg_load - tmp); | |
4991 | } | |
1e3c88bd PZ |
4992 | |
4993 | /* Amount of load we'd add */ | |
3ae11c90 | 4994 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 4995 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
4996 | tmp = (busiest->avg_load * busiest->group_power) / |
4997 | local->group_power; | |
56cf515b JK |
4998 | } else { |
4999 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5000 | local->group_power; |
56cf515b | 5001 | } |
3ae11c90 PZ |
5002 | pwr_move += local->group_power * |
5003 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 5004 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5005 | |
5006 | /* Move if we gain throughput */ | |
5007 | if (pwr_move > pwr_now) | |
56cf515b | 5008 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5009 | } |
5010 | ||
5011 | /** | |
5012 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
5013 | * groups of a given sched_domain during load balance. | |
bd939f45 | 5014 | * @env: load balance environment |
1e3c88bd | 5015 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5016 | */ |
bd939f45 | 5017 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5018 | { |
dd5feea1 | 5019 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
5020 | struct sg_lb_stats *local, *busiest; |
5021 | ||
5022 | local = &sds->local_stat; | |
56cf515b | 5023 | busiest = &sds->busiest_stat; |
dd5feea1 | 5024 | |
56cf515b | 5025 | if (busiest->group_imb) { |
30ce5dab PZ |
5026 | /* |
5027 | * In the group_imb case we cannot rely on group-wide averages | |
5028 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
5029 | */ | |
56cf515b JK |
5030 | busiest->load_per_task = |
5031 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
5032 | } |
5033 | ||
1e3c88bd PZ |
5034 | /* |
5035 | * In the presence of smp nice balancing, certain scenarios can have | |
5036 | * max load less than avg load(as we skip the groups at or below | |
5037 | * its cpu_power, while calculating max_load..) | |
5038 | */ | |
b1885550 VD |
5039 | if (busiest->avg_load <= sds->avg_load || |
5040 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
5041 | env->imbalance = 0; |
5042 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
5043 | } |
5044 | ||
56cf515b | 5045 | if (!busiest->group_imb) { |
dd5feea1 SS |
5046 | /* |
5047 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
5048 | * Except of course for the group_imb case, since then we might |
5049 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 5050 | */ |
56cf515b JK |
5051 | load_above_capacity = |
5052 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 5053 | |
1399fa78 | 5054 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 5055 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
5056 | } |
5057 | ||
5058 | /* | |
5059 | * We're trying to get all the cpus to the average_load, so we don't | |
5060 | * want to push ourselves above the average load, nor do we wish to | |
5061 | * reduce the max loaded cpu below the average load. At the same time, | |
5062 | * we also don't want to reduce the group load below the group capacity | |
5063 | * (so that we can implement power-savings policies etc). Thus we look | |
5064 | * for the minimum possible imbalance. | |
dd5feea1 | 5065 | */ |
30ce5dab | 5066 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
5067 | |
5068 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 5069 | env->imbalance = min( |
3ae11c90 PZ |
5070 | max_pull * busiest->group_power, |
5071 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 5072 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
5073 | |
5074 | /* | |
5075 | * if *imbalance is less than the average load per runnable task | |
25985edc | 5076 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
5077 | * a think about bumping its value to force at least one task to be |
5078 | * moved | |
5079 | */ | |
56cf515b | 5080 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 5081 | return fix_small_imbalance(env, sds); |
1e3c88bd | 5082 | } |
fab47622 | 5083 | |
1e3c88bd PZ |
5084 | /******* find_busiest_group() helpers end here *********************/ |
5085 | ||
5086 | /** | |
5087 | * find_busiest_group - Returns the busiest group within the sched_domain | |
5088 | * if there is an imbalance. If there isn't an imbalance, and | |
5089 | * the user has opted for power-savings, it returns a group whose | |
5090 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
5091 | * such a group exists. | |
5092 | * | |
5093 | * Also calculates the amount of weighted load which should be moved | |
5094 | * to restore balance. | |
5095 | * | |
cd96891d | 5096 | * @env: The load balancing environment. |
1e3c88bd | 5097 | * |
e69f6186 | 5098 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
5099 | * - If no imbalance and user has opted for power-savings balance, |
5100 | * return the least loaded group whose CPUs can be | |
5101 | * put to idle by rebalancing its tasks onto our group. | |
5102 | */ | |
56cf515b | 5103 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 5104 | { |
56cf515b | 5105 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
5106 | struct sd_lb_stats sds; |
5107 | ||
147c5fc2 | 5108 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
5109 | |
5110 | /* | |
5111 | * Compute the various statistics relavent for load balancing at | |
5112 | * this level. | |
5113 | */ | |
23f0d209 | 5114 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
5115 | local = &sds.local_stat; |
5116 | busiest = &sds.busiest_stat; | |
1e3c88bd | 5117 | |
bd939f45 PZ |
5118 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
5119 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
5120 | return sds.busiest; |
5121 | ||
cc57aa8f | 5122 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 5123 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
5124 | goto out_balanced; |
5125 | ||
1399fa78 | 5126 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 5127 | |
866ab43e PZ |
5128 | /* |
5129 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 5130 | * work because they assume all things are equal, which typically |
866ab43e PZ |
5131 | * isn't true due to cpus_allowed constraints and the like. |
5132 | */ | |
56cf515b | 5133 | if (busiest->group_imb) |
866ab43e PZ |
5134 | goto force_balance; |
5135 | ||
cc57aa8f | 5136 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
5137 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
5138 | !busiest->group_has_capacity) | |
fab47622 NR |
5139 | goto force_balance; |
5140 | ||
cc57aa8f PZ |
5141 | /* |
5142 | * If the local group is more busy than the selected busiest group | |
5143 | * don't try and pull any tasks. | |
5144 | */ | |
56cf515b | 5145 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
5146 | goto out_balanced; |
5147 | ||
cc57aa8f PZ |
5148 | /* |
5149 | * Don't pull any tasks if this group is already above the domain | |
5150 | * average load. | |
5151 | */ | |
56cf515b | 5152 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
5153 | goto out_balanced; |
5154 | ||
bd939f45 | 5155 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
5156 | /* |
5157 | * This cpu is idle. If the busiest group load doesn't | |
5158 | * have more tasks than the number of available cpu's and | |
5159 | * there is no imbalance between this and busiest group | |
5160 | * wrt to idle cpu's, it is balanced. | |
5161 | */ | |
56cf515b JK |
5162 | if ((local->idle_cpus < busiest->idle_cpus) && |
5163 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 5164 | goto out_balanced; |
c186fafe PZ |
5165 | } else { |
5166 | /* | |
5167 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
5168 | * imbalance_pct to be conservative. | |
5169 | */ | |
56cf515b JK |
5170 | if (100 * busiest->avg_load <= |
5171 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 5172 | goto out_balanced; |
aae6d3dd | 5173 | } |
1e3c88bd | 5174 | |
fab47622 | 5175 | force_balance: |
1e3c88bd | 5176 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 5177 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
5178 | return sds.busiest; |
5179 | ||
5180 | out_balanced: | |
bd939f45 | 5181 | env->imbalance = 0; |
1e3c88bd PZ |
5182 | return NULL; |
5183 | } | |
5184 | ||
5185 | /* | |
5186 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
5187 | */ | |
bd939f45 | 5188 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 5189 | struct sched_group *group) |
1e3c88bd PZ |
5190 | { |
5191 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 5192 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
5193 | int i; |
5194 | ||
6906a408 | 5195 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd | 5196 | unsigned long power = power_of(i); |
1399fa78 NR |
5197 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
5198 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
5199 | unsigned long wl; |
5200 | ||
9d5efe05 | 5201 | if (!capacity) |
bd939f45 | 5202 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 5203 | |
1e3c88bd | 5204 | rq = cpu_rq(i); |
6e40f5bb | 5205 | wl = weighted_cpuload(i); |
1e3c88bd | 5206 | |
6e40f5bb TG |
5207 | /* |
5208 | * When comparing with imbalance, use weighted_cpuload() | |
5209 | * which is not scaled with the cpu power. | |
5210 | */ | |
bd939f45 | 5211 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
5212 | continue; |
5213 | ||
6e40f5bb TG |
5214 | /* |
5215 | * For the load comparisons with the other cpu's, consider | |
5216 | * the weighted_cpuload() scaled with the cpu power, so that | |
5217 | * the load can be moved away from the cpu that is potentially | |
5218 | * running at a lower capacity. | |
95a79b80 JK |
5219 | * |
5220 | * Thus we're looking for max(wl_i / power_i), crosswise | |
5221 | * multiplication to rid ourselves of the division works out | |
5222 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
5223 | * previous maximum. | |
6e40f5bb | 5224 | */ |
95a79b80 JK |
5225 | if (wl * busiest_power > busiest_load * power) { |
5226 | busiest_load = wl; | |
5227 | busiest_power = power; | |
1e3c88bd PZ |
5228 | busiest = rq; |
5229 | } | |
5230 | } | |
5231 | ||
5232 | return busiest; | |
5233 | } | |
5234 | ||
5235 | /* | |
5236 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
5237 | * so long as it is large enough. | |
5238 | */ | |
5239 | #define MAX_PINNED_INTERVAL 512 | |
5240 | ||
5241 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 5242 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 5243 | |
bd939f45 | 5244 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 5245 | { |
bd939f45 PZ |
5246 | struct sched_domain *sd = env->sd; |
5247 | ||
5248 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
5249 | |
5250 | /* | |
5251 | * ASYM_PACKING needs to force migrate tasks from busy but | |
5252 | * higher numbered CPUs in order to pack all tasks in the | |
5253 | * lowest numbered CPUs. | |
5254 | */ | |
bd939f45 | 5255 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 5256 | return 1; |
1af3ed3d PZ |
5257 | } |
5258 | ||
5259 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
5260 | } | |
5261 | ||
969c7921 TH |
5262 | static int active_load_balance_cpu_stop(void *data); |
5263 | ||
23f0d209 JK |
5264 | static int should_we_balance(struct lb_env *env) |
5265 | { | |
5266 | struct sched_group *sg = env->sd->groups; | |
5267 | struct cpumask *sg_cpus, *sg_mask; | |
5268 | int cpu, balance_cpu = -1; | |
5269 | ||
5270 | /* | |
5271 | * In the newly idle case, we will allow all the cpu's | |
5272 | * to do the newly idle load balance. | |
5273 | */ | |
5274 | if (env->idle == CPU_NEWLY_IDLE) | |
5275 | return 1; | |
5276 | ||
5277 | sg_cpus = sched_group_cpus(sg); | |
5278 | sg_mask = sched_group_mask(sg); | |
5279 | /* Try to find first idle cpu */ | |
5280 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
5281 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
5282 | continue; | |
5283 | ||
5284 | balance_cpu = cpu; | |
5285 | break; | |
5286 | } | |
5287 | ||
5288 | if (balance_cpu == -1) | |
5289 | balance_cpu = group_balance_cpu(sg); | |
5290 | ||
5291 | /* | |
5292 | * First idle cpu or the first cpu(busiest) in this sched group | |
5293 | * is eligible for doing load balancing at this and above domains. | |
5294 | */ | |
b0cff9d8 | 5295 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
5296 | } |
5297 | ||
1e3c88bd PZ |
5298 | /* |
5299 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
5300 | * tasks if there is an imbalance. | |
5301 | */ | |
5302 | static int load_balance(int this_cpu, struct rq *this_rq, | |
5303 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 5304 | int *continue_balancing) |
1e3c88bd | 5305 | { |
88b8dac0 | 5306 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 5307 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 5308 | struct sched_group *group; |
1e3c88bd PZ |
5309 | struct rq *busiest; |
5310 | unsigned long flags; | |
e6252c3e | 5311 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 5312 | |
8e45cb54 PZ |
5313 | struct lb_env env = { |
5314 | .sd = sd, | |
ddcdf6e7 PZ |
5315 | .dst_cpu = this_cpu, |
5316 | .dst_rq = this_rq, | |
88b8dac0 | 5317 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 5318 | .idle = idle, |
eb95308e | 5319 | .loop_break = sched_nr_migrate_break, |
b9403130 | 5320 | .cpus = cpus, |
8e45cb54 PZ |
5321 | }; |
5322 | ||
cfc03118 JK |
5323 | /* |
5324 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
5325 | * other cpus in our group | |
5326 | */ | |
e02e60c1 | 5327 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 5328 | env.dst_grpmask = NULL; |
cfc03118 | 5329 | |
1e3c88bd PZ |
5330 | cpumask_copy(cpus, cpu_active_mask); |
5331 | ||
1e3c88bd PZ |
5332 | schedstat_inc(sd, lb_count[idle]); |
5333 | ||
5334 | redo: | |
23f0d209 JK |
5335 | if (!should_we_balance(&env)) { |
5336 | *continue_balancing = 0; | |
1e3c88bd | 5337 | goto out_balanced; |
23f0d209 | 5338 | } |
1e3c88bd | 5339 | |
23f0d209 | 5340 | group = find_busiest_group(&env); |
1e3c88bd PZ |
5341 | if (!group) { |
5342 | schedstat_inc(sd, lb_nobusyg[idle]); | |
5343 | goto out_balanced; | |
5344 | } | |
5345 | ||
b9403130 | 5346 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
5347 | if (!busiest) { |
5348 | schedstat_inc(sd, lb_nobusyq[idle]); | |
5349 | goto out_balanced; | |
5350 | } | |
5351 | ||
78feefc5 | 5352 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 5353 | |
bd939f45 | 5354 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
5355 | |
5356 | ld_moved = 0; | |
5357 | if (busiest->nr_running > 1) { | |
5358 | /* | |
5359 | * Attempt to move tasks. If find_busiest_group has found | |
5360 | * an imbalance but busiest->nr_running <= 1, the group is | |
5361 | * still unbalanced. ld_moved simply stays zero, so it is | |
5362 | * correctly treated as an imbalance. | |
5363 | */ | |
8e45cb54 | 5364 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
5365 | env.src_cpu = busiest->cpu; |
5366 | env.src_rq = busiest; | |
5367 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 5368 | |
5d6523eb | 5369 | more_balance: |
1e3c88bd | 5370 | local_irq_save(flags); |
78feefc5 | 5371 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
5372 | |
5373 | /* | |
5374 | * cur_ld_moved - load moved in current iteration | |
5375 | * ld_moved - cumulative load moved across iterations | |
5376 | */ | |
5377 | cur_ld_moved = move_tasks(&env); | |
5378 | ld_moved += cur_ld_moved; | |
78feefc5 | 5379 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
5380 | local_irq_restore(flags); |
5381 | ||
5382 | /* | |
5383 | * some other cpu did the load balance for us. | |
5384 | */ | |
88b8dac0 SV |
5385 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
5386 | resched_cpu(env.dst_cpu); | |
5387 | ||
f1cd0858 JK |
5388 | if (env.flags & LBF_NEED_BREAK) { |
5389 | env.flags &= ~LBF_NEED_BREAK; | |
5390 | goto more_balance; | |
5391 | } | |
5392 | ||
88b8dac0 SV |
5393 | /* |
5394 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
5395 | * us and move them to an alternate dst_cpu in our sched_group | |
5396 | * where they can run. The upper limit on how many times we | |
5397 | * iterate on same src_cpu is dependent on number of cpus in our | |
5398 | * sched_group. | |
5399 | * | |
5400 | * This changes load balance semantics a bit on who can move | |
5401 | * load to a given_cpu. In addition to the given_cpu itself | |
5402 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
5403 | * nohz-idle), we now have balance_cpu in a position to move | |
5404 | * load to given_cpu. In rare situations, this may cause | |
5405 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
5406 | * _independently_ and at _same_ time to move some load to | |
5407 | * given_cpu) causing exceess load to be moved to given_cpu. | |
5408 | * This however should not happen so much in practice and | |
5409 | * moreover subsequent load balance cycles should correct the | |
5410 | * excess load moved. | |
5411 | */ | |
6263322c | 5412 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 5413 | |
7aff2e3a VD |
5414 | /* Prevent to re-select dst_cpu via env's cpus */ |
5415 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
5416 | ||
78feefc5 | 5417 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 5418 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 5419 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
5420 | env.loop = 0; |
5421 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 5422 | |
88b8dac0 SV |
5423 | /* |
5424 | * Go back to "more_balance" rather than "redo" since we | |
5425 | * need to continue with same src_cpu. | |
5426 | */ | |
5427 | goto more_balance; | |
5428 | } | |
1e3c88bd | 5429 | |
6263322c PZ |
5430 | /* |
5431 | * We failed to reach balance because of affinity. | |
5432 | */ | |
5433 | if (sd_parent) { | |
5434 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
5435 | ||
5436 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
5437 | *group_imbalance = 1; | |
5438 | } else if (*group_imbalance) | |
5439 | *group_imbalance = 0; | |
5440 | } | |
5441 | ||
1e3c88bd | 5442 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 5443 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 5444 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
5445 | if (!cpumask_empty(cpus)) { |
5446 | env.loop = 0; | |
5447 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 5448 | goto redo; |
bbf18b19 | 5449 | } |
1e3c88bd PZ |
5450 | goto out_balanced; |
5451 | } | |
5452 | } | |
5453 | ||
5454 | if (!ld_moved) { | |
5455 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
5456 | /* |
5457 | * Increment the failure counter only on periodic balance. | |
5458 | * We do not want newidle balance, which can be very | |
5459 | * frequent, pollute the failure counter causing | |
5460 | * excessive cache_hot migrations and active balances. | |
5461 | */ | |
5462 | if (idle != CPU_NEWLY_IDLE) | |
5463 | sd->nr_balance_failed++; | |
1e3c88bd | 5464 | |
bd939f45 | 5465 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
5466 | raw_spin_lock_irqsave(&busiest->lock, flags); |
5467 | ||
969c7921 TH |
5468 | /* don't kick the active_load_balance_cpu_stop, |
5469 | * if the curr task on busiest cpu can't be | |
5470 | * moved to this_cpu | |
1e3c88bd PZ |
5471 | */ |
5472 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 5473 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
5474 | raw_spin_unlock_irqrestore(&busiest->lock, |
5475 | flags); | |
8e45cb54 | 5476 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
5477 | goto out_one_pinned; |
5478 | } | |
5479 | ||
969c7921 TH |
5480 | /* |
5481 | * ->active_balance synchronizes accesses to | |
5482 | * ->active_balance_work. Once set, it's cleared | |
5483 | * only after active load balance is finished. | |
5484 | */ | |
1e3c88bd PZ |
5485 | if (!busiest->active_balance) { |
5486 | busiest->active_balance = 1; | |
5487 | busiest->push_cpu = this_cpu; | |
5488 | active_balance = 1; | |
5489 | } | |
5490 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 5491 | |
bd939f45 | 5492 | if (active_balance) { |
969c7921 TH |
5493 | stop_one_cpu_nowait(cpu_of(busiest), |
5494 | active_load_balance_cpu_stop, busiest, | |
5495 | &busiest->active_balance_work); | |
bd939f45 | 5496 | } |
1e3c88bd PZ |
5497 | |
5498 | /* | |
5499 | * We've kicked active balancing, reset the failure | |
5500 | * counter. | |
5501 | */ | |
5502 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
5503 | } | |
5504 | } else | |
5505 | sd->nr_balance_failed = 0; | |
5506 | ||
5507 | if (likely(!active_balance)) { | |
5508 | /* We were unbalanced, so reset the balancing interval */ | |
5509 | sd->balance_interval = sd->min_interval; | |
5510 | } else { | |
5511 | /* | |
5512 | * If we've begun active balancing, start to back off. This | |
5513 | * case may not be covered by the all_pinned logic if there | |
5514 | * is only 1 task on the busy runqueue (because we don't call | |
5515 | * move_tasks). | |
5516 | */ | |
5517 | if (sd->balance_interval < sd->max_interval) | |
5518 | sd->balance_interval *= 2; | |
5519 | } | |
5520 | ||
1e3c88bd PZ |
5521 | goto out; |
5522 | ||
5523 | out_balanced: | |
5524 | schedstat_inc(sd, lb_balanced[idle]); | |
5525 | ||
5526 | sd->nr_balance_failed = 0; | |
5527 | ||
5528 | out_one_pinned: | |
5529 | /* tune up the balancing interval */ | |
8e45cb54 | 5530 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 5531 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
5532 | (sd->balance_interval < sd->max_interval)) |
5533 | sd->balance_interval *= 2; | |
5534 | ||
46e49b38 | 5535 | ld_moved = 0; |
1e3c88bd | 5536 | out: |
1e3c88bd PZ |
5537 | return ld_moved; |
5538 | } | |
5539 | ||
1e3c88bd PZ |
5540 | /* |
5541 | * idle_balance is called by schedule() if this_cpu is about to become | |
5542 | * idle. Attempts to pull tasks from other CPUs. | |
5543 | */ | |
029632fb | 5544 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
5545 | { |
5546 | struct sched_domain *sd; | |
5547 | int pulled_task = 0; | |
5548 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 5549 | u64 curr_cost = 0; |
1e3c88bd | 5550 | |
78becc27 | 5551 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
5552 | |
5553 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
5554 | return; | |
5555 | ||
f492e12e PZ |
5556 | /* |
5557 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
5558 | */ | |
5559 | raw_spin_unlock(&this_rq->lock); | |
5560 | ||
48a16753 | 5561 | update_blocked_averages(this_cpu); |
dce840a0 | 5562 | rcu_read_lock(); |
1e3c88bd PZ |
5563 | for_each_domain(this_cpu, sd) { |
5564 | unsigned long interval; | |
23f0d209 | 5565 | int continue_balancing = 1; |
9bd721c5 | 5566 | u64 t0, domain_cost; |
1e3c88bd PZ |
5567 | |
5568 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
5569 | continue; | |
5570 | ||
9bd721c5 JL |
5571 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
5572 | break; | |
5573 | ||
f492e12e | 5574 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
5575 | t0 = sched_clock_cpu(this_cpu); |
5576 | ||
1e3c88bd | 5577 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 5578 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
5579 | sd, CPU_NEWLY_IDLE, |
5580 | &continue_balancing); | |
9bd721c5 JL |
5581 | |
5582 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
5583 | if (domain_cost > sd->max_newidle_lb_cost) | |
5584 | sd->max_newidle_lb_cost = domain_cost; | |
5585 | ||
5586 | curr_cost += domain_cost; | |
f492e12e | 5587 | } |
1e3c88bd PZ |
5588 | |
5589 | interval = msecs_to_jiffies(sd->balance_interval); | |
5590 | if (time_after(next_balance, sd->last_balance + interval)) | |
5591 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
5592 | if (pulled_task) { |
5593 | this_rq->idle_stamp = 0; | |
1e3c88bd | 5594 | break; |
d5ad140b | 5595 | } |
1e3c88bd | 5596 | } |
dce840a0 | 5597 | rcu_read_unlock(); |
f492e12e PZ |
5598 | |
5599 | raw_spin_lock(&this_rq->lock); | |
5600 | ||
1e3c88bd PZ |
5601 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
5602 | /* | |
5603 | * We are going idle. next_balance may be set based on | |
5604 | * a busy processor. So reset next_balance. | |
5605 | */ | |
5606 | this_rq->next_balance = next_balance; | |
5607 | } | |
9bd721c5 JL |
5608 | |
5609 | if (curr_cost > this_rq->max_idle_balance_cost) | |
5610 | this_rq->max_idle_balance_cost = curr_cost; | |
1e3c88bd PZ |
5611 | } |
5612 | ||
5613 | /* | |
969c7921 TH |
5614 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
5615 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
5616 | * least 1 task to be running on each physical CPU where possible, and | |
5617 | * avoids physical / logical imbalances. | |
1e3c88bd | 5618 | */ |
969c7921 | 5619 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 5620 | { |
969c7921 TH |
5621 | struct rq *busiest_rq = data; |
5622 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 5623 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 5624 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 5625 | struct sched_domain *sd; |
969c7921 TH |
5626 | |
5627 | raw_spin_lock_irq(&busiest_rq->lock); | |
5628 | ||
5629 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
5630 | if (unlikely(busiest_cpu != smp_processor_id() || | |
5631 | !busiest_rq->active_balance)) | |
5632 | goto out_unlock; | |
1e3c88bd PZ |
5633 | |
5634 | /* Is there any task to move? */ | |
5635 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 5636 | goto out_unlock; |
1e3c88bd PZ |
5637 | |
5638 | /* | |
5639 | * This condition is "impossible", if it occurs | |
5640 | * we need to fix it. Originally reported by | |
5641 | * Bjorn Helgaas on a 128-cpu setup. | |
5642 | */ | |
5643 | BUG_ON(busiest_rq == target_rq); | |
5644 | ||
5645 | /* move a task from busiest_rq to target_rq */ | |
5646 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
5647 | |
5648 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 5649 | rcu_read_lock(); |
1e3c88bd PZ |
5650 | for_each_domain(target_cpu, sd) { |
5651 | if ((sd->flags & SD_LOAD_BALANCE) && | |
5652 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
5653 | break; | |
5654 | } | |
5655 | ||
5656 | if (likely(sd)) { | |
8e45cb54 PZ |
5657 | struct lb_env env = { |
5658 | .sd = sd, | |
ddcdf6e7 PZ |
5659 | .dst_cpu = target_cpu, |
5660 | .dst_rq = target_rq, | |
5661 | .src_cpu = busiest_rq->cpu, | |
5662 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
5663 | .idle = CPU_IDLE, |
5664 | }; | |
5665 | ||
1e3c88bd PZ |
5666 | schedstat_inc(sd, alb_count); |
5667 | ||
8e45cb54 | 5668 | if (move_one_task(&env)) |
1e3c88bd PZ |
5669 | schedstat_inc(sd, alb_pushed); |
5670 | else | |
5671 | schedstat_inc(sd, alb_failed); | |
5672 | } | |
dce840a0 | 5673 | rcu_read_unlock(); |
1e3c88bd | 5674 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
5675 | out_unlock: |
5676 | busiest_rq->active_balance = 0; | |
5677 | raw_spin_unlock_irq(&busiest_rq->lock); | |
5678 | return 0; | |
1e3c88bd PZ |
5679 | } |
5680 | ||
3451d024 | 5681 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
5682 | /* |
5683 | * idle load balancing details | |
83cd4fe2 VP |
5684 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
5685 | * needed, they will kick the idle load balancer, which then does idle | |
5686 | * load balancing for all the idle CPUs. | |
5687 | */ | |
1e3c88bd | 5688 | static struct { |
83cd4fe2 | 5689 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 5690 | atomic_t nr_cpus; |
83cd4fe2 VP |
5691 | unsigned long next_balance; /* in jiffy units */ |
5692 | } nohz ____cacheline_aligned; | |
1e3c88bd | 5693 | |
8e7fbcbc | 5694 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 5695 | { |
0b005cf5 | 5696 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 5697 | |
786d6dc7 SS |
5698 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
5699 | return ilb; | |
5700 | ||
5701 | return nr_cpu_ids; | |
1e3c88bd | 5702 | } |
1e3c88bd | 5703 | |
83cd4fe2 VP |
5704 | /* |
5705 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
5706 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
5707 | * CPU (if there is one). | |
5708 | */ | |
5709 | static void nohz_balancer_kick(int cpu) | |
5710 | { | |
5711 | int ilb_cpu; | |
5712 | ||
5713 | nohz.next_balance++; | |
5714 | ||
0b005cf5 | 5715 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 5716 | |
0b005cf5 SS |
5717 | if (ilb_cpu >= nr_cpu_ids) |
5718 | return; | |
83cd4fe2 | 5719 | |
cd490c5b | 5720 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
5721 | return; |
5722 | /* | |
5723 | * Use smp_send_reschedule() instead of resched_cpu(). | |
5724 | * This way we generate a sched IPI on the target cpu which | |
5725 | * is idle. And the softirq performing nohz idle load balance | |
5726 | * will be run before returning from the IPI. | |
5727 | */ | |
5728 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
5729 | return; |
5730 | } | |
5731 | ||
c1cc017c | 5732 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
5733 | { |
5734 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
5735 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
5736 | atomic_dec(&nohz.nr_cpus); | |
5737 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
5738 | } | |
5739 | } | |
5740 | ||
69e1e811 SS |
5741 | static inline void set_cpu_sd_state_busy(void) |
5742 | { | |
5743 | struct sched_domain *sd; | |
69e1e811 | 5744 | |
69e1e811 | 5745 | rcu_read_lock(); |
424c93fe | 5746 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
5747 | |
5748 | if (!sd || !sd->nohz_idle) | |
5749 | goto unlock; | |
5750 | sd->nohz_idle = 0; | |
5751 | ||
5752 | for (; sd; sd = sd->parent) | |
69e1e811 | 5753 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 5754 | unlock: |
69e1e811 SS |
5755 | rcu_read_unlock(); |
5756 | } | |
5757 | ||
5758 | void set_cpu_sd_state_idle(void) | |
5759 | { | |
5760 | struct sched_domain *sd; | |
69e1e811 | 5761 | |
69e1e811 | 5762 | rcu_read_lock(); |
424c93fe | 5763 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
5764 | |
5765 | if (!sd || sd->nohz_idle) | |
5766 | goto unlock; | |
5767 | sd->nohz_idle = 1; | |
5768 | ||
5769 | for (; sd; sd = sd->parent) | |
69e1e811 | 5770 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 5771 | unlock: |
69e1e811 SS |
5772 | rcu_read_unlock(); |
5773 | } | |
5774 | ||
1e3c88bd | 5775 | /* |
c1cc017c | 5776 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 5777 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 5778 | */ |
c1cc017c | 5779 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 5780 | { |
71325960 SS |
5781 | /* |
5782 | * If this cpu is going down, then nothing needs to be done. | |
5783 | */ | |
5784 | if (!cpu_active(cpu)) | |
5785 | return; | |
5786 | ||
c1cc017c AS |
5787 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
5788 | return; | |
1e3c88bd | 5789 | |
c1cc017c AS |
5790 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
5791 | atomic_inc(&nohz.nr_cpus); | |
5792 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 5793 | } |
71325960 | 5794 | |
0db0628d | 5795 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
5796 | unsigned long action, void *hcpu) |
5797 | { | |
5798 | switch (action & ~CPU_TASKS_FROZEN) { | |
5799 | case CPU_DYING: | |
c1cc017c | 5800 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
5801 | return NOTIFY_OK; |
5802 | default: | |
5803 | return NOTIFY_DONE; | |
5804 | } | |
5805 | } | |
1e3c88bd PZ |
5806 | #endif |
5807 | ||
5808 | static DEFINE_SPINLOCK(balancing); | |
5809 | ||
49c022e6 PZ |
5810 | /* |
5811 | * Scale the max load_balance interval with the number of CPUs in the system. | |
5812 | * This trades load-balance latency on larger machines for less cross talk. | |
5813 | */ | |
029632fb | 5814 | void update_max_interval(void) |
49c022e6 PZ |
5815 | { |
5816 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
5817 | } | |
5818 | ||
1e3c88bd PZ |
5819 | /* |
5820 | * It checks each scheduling domain to see if it is due to be balanced, | |
5821 | * and initiates a balancing operation if so. | |
5822 | * | |
b9b0853a | 5823 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd PZ |
5824 | */ |
5825 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
5826 | { | |
23f0d209 | 5827 | int continue_balancing = 1; |
1e3c88bd PZ |
5828 | struct rq *rq = cpu_rq(cpu); |
5829 | unsigned long interval; | |
04f733b4 | 5830 | struct sched_domain *sd; |
1e3c88bd PZ |
5831 | /* Earliest time when we have to do rebalance again */ |
5832 | unsigned long next_balance = jiffies + 60*HZ; | |
5833 | int update_next_balance = 0; | |
f48627e6 JL |
5834 | int need_serialize, need_decay = 0; |
5835 | u64 max_cost = 0; | |
1e3c88bd | 5836 | |
48a16753 | 5837 | update_blocked_averages(cpu); |
2069dd75 | 5838 | |
dce840a0 | 5839 | rcu_read_lock(); |
1e3c88bd | 5840 | for_each_domain(cpu, sd) { |
f48627e6 JL |
5841 | /* |
5842 | * Decay the newidle max times here because this is a regular | |
5843 | * visit to all the domains. Decay ~1% per second. | |
5844 | */ | |
5845 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
5846 | sd->max_newidle_lb_cost = | |
5847 | (sd->max_newidle_lb_cost * 253) / 256; | |
5848 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
5849 | need_decay = 1; | |
5850 | } | |
5851 | max_cost += sd->max_newidle_lb_cost; | |
5852 | ||
1e3c88bd PZ |
5853 | if (!(sd->flags & SD_LOAD_BALANCE)) |
5854 | continue; | |
5855 | ||
f48627e6 JL |
5856 | /* |
5857 | * Stop the load balance at this level. There is another | |
5858 | * CPU in our sched group which is doing load balancing more | |
5859 | * actively. | |
5860 | */ | |
5861 | if (!continue_balancing) { | |
5862 | if (need_decay) | |
5863 | continue; | |
5864 | break; | |
5865 | } | |
5866 | ||
1e3c88bd PZ |
5867 | interval = sd->balance_interval; |
5868 | if (idle != CPU_IDLE) | |
5869 | interval *= sd->busy_factor; | |
5870 | ||
5871 | /* scale ms to jiffies */ | |
5872 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 5873 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
5874 | |
5875 | need_serialize = sd->flags & SD_SERIALIZE; | |
5876 | ||
5877 | if (need_serialize) { | |
5878 | if (!spin_trylock(&balancing)) | |
5879 | goto out; | |
5880 | } | |
5881 | ||
5882 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 5883 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 5884 | /* |
6263322c | 5885 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
5886 | * env->dst_cpu, so we can't know our idle |
5887 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 5888 | */ |
de5eb2dd | 5889 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
5890 | } |
5891 | sd->last_balance = jiffies; | |
5892 | } | |
5893 | if (need_serialize) | |
5894 | spin_unlock(&balancing); | |
5895 | out: | |
5896 | if (time_after(next_balance, sd->last_balance + interval)) { | |
5897 | next_balance = sd->last_balance + interval; | |
5898 | update_next_balance = 1; | |
5899 | } | |
f48627e6 JL |
5900 | } |
5901 | if (need_decay) { | |
1e3c88bd | 5902 | /* |
f48627e6 JL |
5903 | * Ensure the rq-wide value also decays but keep it at a |
5904 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 5905 | */ |
f48627e6 JL |
5906 | rq->max_idle_balance_cost = |
5907 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 5908 | } |
dce840a0 | 5909 | rcu_read_unlock(); |
1e3c88bd PZ |
5910 | |
5911 | /* | |
5912 | * next_balance will be updated only when there is a need. | |
5913 | * When the cpu is attached to null domain for ex, it will not be | |
5914 | * updated. | |
5915 | */ | |
5916 | if (likely(update_next_balance)) | |
5917 | rq->next_balance = next_balance; | |
5918 | } | |
5919 | ||
3451d024 | 5920 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 5921 | /* |
3451d024 | 5922 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
5923 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
5924 | */ | |
83cd4fe2 VP |
5925 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
5926 | { | |
5927 | struct rq *this_rq = cpu_rq(this_cpu); | |
5928 | struct rq *rq; | |
5929 | int balance_cpu; | |
5930 | ||
1c792db7 SS |
5931 | if (idle != CPU_IDLE || |
5932 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
5933 | goto end; | |
83cd4fe2 VP |
5934 | |
5935 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 5936 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
5937 | continue; |
5938 | ||
5939 | /* | |
5940 | * If this cpu gets work to do, stop the load balancing | |
5941 | * work being done for other cpus. Next load | |
5942 | * balancing owner will pick it up. | |
5943 | */ | |
1c792db7 | 5944 | if (need_resched()) |
83cd4fe2 | 5945 | break; |
83cd4fe2 | 5946 | |
5ed4f1d9 VG |
5947 | rq = cpu_rq(balance_cpu); |
5948 | ||
5949 | raw_spin_lock_irq(&rq->lock); | |
5950 | update_rq_clock(rq); | |
5951 | update_idle_cpu_load(rq); | |
5952 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
5953 | |
5954 | rebalance_domains(balance_cpu, CPU_IDLE); | |
5955 | ||
83cd4fe2 VP |
5956 | if (time_after(this_rq->next_balance, rq->next_balance)) |
5957 | this_rq->next_balance = rq->next_balance; | |
5958 | } | |
5959 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
5960 | end: |
5961 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
5962 | } |
5963 | ||
5964 | /* | |
0b005cf5 SS |
5965 | * Current heuristic for kicking the idle load balancer in the presence |
5966 | * of an idle cpu is the system. | |
5967 | * - This rq has more than one task. | |
5968 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
5969 | * busy cpu's exceeding the group's power. | |
5970 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
5971 | * domain span are idle. | |
83cd4fe2 VP |
5972 | */ |
5973 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
5974 | { | |
5975 | unsigned long now = jiffies; | |
0b005cf5 | 5976 | struct sched_domain *sd; |
83cd4fe2 | 5977 | |
1c792db7 | 5978 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
5979 | return 0; |
5980 | ||
1c792db7 SS |
5981 | /* |
5982 | * We may be recently in ticked or tickless idle mode. At the first | |
5983 | * busy tick after returning from idle, we will update the busy stats. | |
5984 | */ | |
69e1e811 | 5985 | set_cpu_sd_state_busy(); |
c1cc017c | 5986 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
5987 | |
5988 | /* | |
5989 | * None are in tickless mode and hence no need for NOHZ idle load | |
5990 | * balancing. | |
5991 | */ | |
5992 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
5993 | return 0; | |
1c792db7 SS |
5994 | |
5995 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
5996 | return 0; |
5997 | ||
0b005cf5 SS |
5998 | if (rq->nr_running >= 2) |
5999 | goto need_kick; | |
83cd4fe2 | 6000 | |
067491b7 | 6001 | rcu_read_lock(); |
0b005cf5 SS |
6002 | for_each_domain(cpu, sd) { |
6003 | struct sched_group *sg = sd->groups; | |
6004 | struct sched_group_power *sgp = sg->sgp; | |
6005 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 6006 | |
0b005cf5 | 6007 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 6008 | goto need_kick_unlock; |
0b005cf5 SS |
6009 | |
6010 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
6011 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
6012 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 6013 | goto need_kick_unlock; |
0b005cf5 SS |
6014 | |
6015 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
6016 | break; | |
83cd4fe2 | 6017 | } |
067491b7 | 6018 | rcu_read_unlock(); |
83cd4fe2 | 6019 | return 0; |
067491b7 PZ |
6020 | |
6021 | need_kick_unlock: | |
6022 | rcu_read_unlock(); | |
0b005cf5 SS |
6023 | need_kick: |
6024 | return 1; | |
83cd4fe2 VP |
6025 | } |
6026 | #else | |
6027 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
6028 | #endif | |
6029 | ||
6030 | /* | |
6031 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
6032 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
6033 | */ | |
1e3c88bd PZ |
6034 | static void run_rebalance_domains(struct softirq_action *h) |
6035 | { | |
6036 | int this_cpu = smp_processor_id(); | |
6037 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 6038 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
6039 | CPU_IDLE : CPU_NOT_IDLE; |
6040 | ||
6041 | rebalance_domains(this_cpu, idle); | |
6042 | ||
1e3c88bd | 6043 | /* |
83cd4fe2 | 6044 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
6045 | * balancing on behalf of the other idle cpus whose ticks are |
6046 | * stopped. | |
6047 | */ | |
83cd4fe2 | 6048 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
6049 | } |
6050 | ||
6051 | static inline int on_null_domain(int cpu) | |
6052 | { | |
90a6501f | 6053 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
6054 | } |
6055 | ||
6056 | /* | |
6057 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 6058 | */ |
029632fb | 6059 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 6060 | { |
1e3c88bd PZ |
6061 | /* Don't need to rebalance while attached to NULL domain */ |
6062 | if (time_after_eq(jiffies, rq->next_balance) && | |
6063 | likely(!on_null_domain(cpu))) | |
6064 | raise_softirq(SCHED_SOFTIRQ); | |
3451d024 | 6065 | #ifdef CONFIG_NO_HZ_COMMON |
1c792db7 | 6066 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
6067 | nohz_balancer_kick(cpu); |
6068 | #endif | |
1e3c88bd PZ |
6069 | } |
6070 | ||
0bcdcf28 CE |
6071 | static void rq_online_fair(struct rq *rq) |
6072 | { | |
6073 | update_sysctl(); | |
6074 | } | |
6075 | ||
6076 | static void rq_offline_fair(struct rq *rq) | |
6077 | { | |
6078 | update_sysctl(); | |
a4c96ae3 PB |
6079 | |
6080 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
6081 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
6082 | } |
6083 | ||
55e12e5e | 6084 | #endif /* CONFIG_SMP */ |
e1d1484f | 6085 | |
bf0f6f24 IM |
6086 | /* |
6087 | * scheduler tick hitting a task of our scheduling class: | |
6088 | */ | |
8f4d37ec | 6089 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
6090 | { |
6091 | struct cfs_rq *cfs_rq; | |
6092 | struct sched_entity *se = &curr->se; | |
6093 | ||
6094 | for_each_sched_entity(se) { | |
6095 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 6096 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 6097 | } |
18bf2805 | 6098 | |
10e84b97 | 6099 | if (numabalancing_enabled) |
cbee9f88 | 6100 | task_tick_numa(rq, curr); |
3d59eebc | 6101 | |
18bf2805 | 6102 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
6103 | } |
6104 | ||
6105 | /* | |
cd29fe6f PZ |
6106 | * called on fork with the child task as argument from the parent's context |
6107 | * - child not yet on the tasklist | |
6108 | * - preemption disabled | |
bf0f6f24 | 6109 | */ |
cd29fe6f | 6110 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 6111 | { |
4fc420c9 DN |
6112 | struct cfs_rq *cfs_rq; |
6113 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 6114 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
6115 | struct rq *rq = this_rq(); |
6116 | unsigned long flags; | |
6117 | ||
05fa785c | 6118 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 6119 | |
861d034e PZ |
6120 | update_rq_clock(rq); |
6121 | ||
4fc420c9 DN |
6122 | cfs_rq = task_cfs_rq(current); |
6123 | curr = cfs_rq->curr; | |
6124 | ||
6c9a27f5 DN |
6125 | /* |
6126 | * Not only the cpu but also the task_group of the parent might have | |
6127 | * been changed after parent->se.parent,cfs_rq were copied to | |
6128 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
6129 | * of child point to valid ones. | |
6130 | */ | |
6131 | rcu_read_lock(); | |
6132 | __set_task_cpu(p, this_cpu); | |
6133 | rcu_read_unlock(); | |
bf0f6f24 | 6134 | |
7109c442 | 6135 | update_curr(cfs_rq); |
cd29fe6f | 6136 | |
b5d9d734 MG |
6137 | if (curr) |
6138 | se->vruntime = curr->vruntime; | |
aeb73b04 | 6139 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 6140 | |
cd29fe6f | 6141 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 6142 | /* |
edcb60a3 IM |
6143 | * Upon rescheduling, sched_class::put_prev_task() will place |
6144 | * 'current' within the tree based on its new key value. | |
6145 | */ | |
4d78e7b6 | 6146 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 6147 | resched_task(rq->curr); |
4d78e7b6 | 6148 | } |
bf0f6f24 | 6149 | |
88ec22d3 PZ |
6150 | se->vruntime -= cfs_rq->min_vruntime; |
6151 | ||
05fa785c | 6152 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
6153 | } |
6154 | ||
cb469845 SR |
6155 | /* |
6156 | * Priority of the task has changed. Check to see if we preempt | |
6157 | * the current task. | |
6158 | */ | |
da7a735e PZ |
6159 | static void |
6160 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 6161 | { |
da7a735e PZ |
6162 | if (!p->se.on_rq) |
6163 | return; | |
6164 | ||
cb469845 SR |
6165 | /* |
6166 | * Reschedule if we are currently running on this runqueue and | |
6167 | * our priority decreased, or if we are not currently running on | |
6168 | * this runqueue and our priority is higher than the current's | |
6169 | */ | |
da7a735e | 6170 | if (rq->curr == p) { |
cb469845 SR |
6171 | if (p->prio > oldprio) |
6172 | resched_task(rq->curr); | |
6173 | } else | |
15afe09b | 6174 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6175 | } |
6176 | ||
da7a735e PZ |
6177 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
6178 | { | |
6179 | struct sched_entity *se = &p->se; | |
6180 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6181 | ||
6182 | /* | |
6183 | * Ensure the task's vruntime is normalized, so that when its | |
6184 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
6185 | * do the right thing. | |
6186 | * | |
6187 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
6188 | * have normalized the vruntime, if it was !on_rq, then only when | |
6189 | * the task is sleeping will it still have non-normalized vruntime. | |
6190 | */ | |
6191 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
6192 | /* | |
6193 | * Fix up our vruntime so that the current sleep doesn't | |
6194 | * cause 'unlimited' sleep bonus. | |
6195 | */ | |
6196 | place_entity(cfs_rq, se, 0); | |
6197 | se->vruntime -= cfs_rq->min_vruntime; | |
6198 | } | |
9ee474f5 | 6199 | |
141965c7 | 6200 | #ifdef CONFIG_SMP |
9ee474f5 PT |
6201 | /* |
6202 | * Remove our load from contribution when we leave sched_fair | |
6203 | * and ensure we don't carry in an old decay_count if we | |
6204 | * switch back. | |
6205 | */ | |
87e3c8ae KT |
6206 | if (se->avg.decay_count) { |
6207 | __synchronize_entity_decay(se); | |
6208 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
6209 | } |
6210 | #endif | |
da7a735e PZ |
6211 | } |
6212 | ||
cb469845 SR |
6213 | /* |
6214 | * We switched to the sched_fair class. | |
6215 | */ | |
da7a735e | 6216 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 6217 | { |
da7a735e PZ |
6218 | if (!p->se.on_rq) |
6219 | return; | |
6220 | ||
cb469845 SR |
6221 | /* |
6222 | * We were most likely switched from sched_rt, so | |
6223 | * kick off the schedule if running, otherwise just see | |
6224 | * if we can still preempt the current task. | |
6225 | */ | |
da7a735e | 6226 | if (rq->curr == p) |
cb469845 SR |
6227 | resched_task(rq->curr); |
6228 | else | |
15afe09b | 6229 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6230 | } |
6231 | ||
83b699ed SV |
6232 | /* Account for a task changing its policy or group. |
6233 | * | |
6234 | * This routine is mostly called to set cfs_rq->curr field when a task | |
6235 | * migrates between groups/classes. | |
6236 | */ | |
6237 | static void set_curr_task_fair(struct rq *rq) | |
6238 | { | |
6239 | struct sched_entity *se = &rq->curr->se; | |
6240 | ||
ec12cb7f PT |
6241 | for_each_sched_entity(se) { |
6242 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6243 | ||
6244 | set_next_entity(cfs_rq, se); | |
6245 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
6246 | account_cfs_rq_runtime(cfs_rq, 0); | |
6247 | } | |
83b699ed SV |
6248 | } |
6249 | ||
029632fb PZ |
6250 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
6251 | { | |
6252 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
6253 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
6254 | #ifndef CONFIG_64BIT | |
6255 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
6256 | #endif | |
141965c7 | 6257 | #ifdef CONFIG_SMP |
9ee474f5 | 6258 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 6259 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 6260 | #endif |
029632fb PZ |
6261 | } |
6262 | ||
810b3817 | 6263 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6264 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 6265 | { |
aff3e498 | 6266 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
6267 | /* |
6268 | * If the task was not on the rq at the time of this cgroup movement | |
6269 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
6270 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
6271 | * bonus in place_entity()). | |
6272 | * | |
6273 | * If it was on the rq, we've just 'preempted' it, which does convert | |
6274 | * ->vruntime to a relative base. | |
6275 | * | |
6276 | * Make sure both cases convert their relative position when migrating | |
6277 | * to another cgroup's rq. This does somewhat interfere with the | |
6278 | * fair sleeper stuff for the first placement, but who cares. | |
6279 | */ | |
7ceff013 DN |
6280 | /* |
6281 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
6282 | * But there are some cases where it has already been normalized: | |
6283 | * | |
6284 | * - Moving a forked child which is waiting for being woken up by | |
6285 | * wake_up_new_task(). | |
62af3783 DN |
6286 | * - Moving a task which has been woken up by try_to_wake_up() and |
6287 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
6288 | * |
6289 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
6290 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
6291 | */ | |
62af3783 | 6292 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
6293 | on_rq = 1; |
6294 | ||
b2b5ce02 PZ |
6295 | if (!on_rq) |
6296 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
6297 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
6298 | if (!on_rq) { |
6299 | cfs_rq = cfs_rq_of(&p->se); | |
6300 | p->se.vruntime += cfs_rq->min_vruntime; | |
6301 | #ifdef CONFIG_SMP | |
6302 | /* | |
6303 | * migrate_task_rq_fair() will have removed our previous | |
6304 | * contribution, but we must synchronize for ongoing future | |
6305 | * decay. | |
6306 | */ | |
6307 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
6308 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
6309 | #endif | |
6310 | } | |
810b3817 | 6311 | } |
029632fb PZ |
6312 | |
6313 | void free_fair_sched_group(struct task_group *tg) | |
6314 | { | |
6315 | int i; | |
6316 | ||
6317 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6318 | ||
6319 | for_each_possible_cpu(i) { | |
6320 | if (tg->cfs_rq) | |
6321 | kfree(tg->cfs_rq[i]); | |
6322 | if (tg->se) | |
6323 | kfree(tg->se[i]); | |
6324 | } | |
6325 | ||
6326 | kfree(tg->cfs_rq); | |
6327 | kfree(tg->se); | |
6328 | } | |
6329 | ||
6330 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6331 | { | |
6332 | struct cfs_rq *cfs_rq; | |
6333 | struct sched_entity *se; | |
6334 | int i; | |
6335 | ||
6336 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
6337 | if (!tg->cfs_rq) | |
6338 | goto err; | |
6339 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
6340 | if (!tg->se) | |
6341 | goto err; | |
6342 | ||
6343 | tg->shares = NICE_0_LOAD; | |
6344 | ||
6345 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6346 | ||
6347 | for_each_possible_cpu(i) { | |
6348 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
6349 | GFP_KERNEL, cpu_to_node(i)); | |
6350 | if (!cfs_rq) | |
6351 | goto err; | |
6352 | ||
6353 | se = kzalloc_node(sizeof(struct sched_entity), | |
6354 | GFP_KERNEL, cpu_to_node(i)); | |
6355 | if (!se) | |
6356 | goto err_free_rq; | |
6357 | ||
6358 | init_cfs_rq(cfs_rq); | |
6359 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
6360 | } | |
6361 | ||
6362 | return 1; | |
6363 | ||
6364 | err_free_rq: | |
6365 | kfree(cfs_rq); | |
6366 | err: | |
6367 | return 0; | |
6368 | } | |
6369 | ||
6370 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
6371 | { | |
6372 | struct rq *rq = cpu_rq(cpu); | |
6373 | unsigned long flags; | |
6374 | ||
6375 | /* | |
6376 | * Only empty task groups can be destroyed; so we can speculatively | |
6377 | * check on_list without danger of it being re-added. | |
6378 | */ | |
6379 | if (!tg->cfs_rq[cpu]->on_list) | |
6380 | return; | |
6381 | ||
6382 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6383 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
6384 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6385 | } | |
6386 | ||
6387 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
6388 | struct sched_entity *se, int cpu, | |
6389 | struct sched_entity *parent) | |
6390 | { | |
6391 | struct rq *rq = cpu_rq(cpu); | |
6392 | ||
6393 | cfs_rq->tg = tg; | |
6394 | cfs_rq->rq = rq; | |
029632fb PZ |
6395 | init_cfs_rq_runtime(cfs_rq); |
6396 | ||
6397 | tg->cfs_rq[cpu] = cfs_rq; | |
6398 | tg->se[cpu] = se; | |
6399 | ||
6400 | /* se could be NULL for root_task_group */ | |
6401 | if (!se) | |
6402 | return; | |
6403 | ||
6404 | if (!parent) | |
6405 | se->cfs_rq = &rq->cfs; | |
6406 | else | |
6407 | se->cfs_rq = parent->my_q; | |
6408 | ||
6409 | se->my_q = cfs_rq; | |
6410 | update_load_set(&se->load, 0); | |
6411 | se->parent = parent; | |
6412 | } | |
6413 | ||
6414 | static DEFINE_MUTEX(shares_mutex); | |
6415 | ||
6416 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
6417 | { | |
6418 | int i; | |
6419 | unsigned long flags; | |
6420 | ||
6421 | /* | |
6422 | * We can't change the weight of the root cgroup. | |
6423 | */ | |
6424 | if (!tg->se[0]) | |
6425 | return -EINVAL; | |
6426 | ||
6427 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
6428 | ||
6429 | mutex_lock(&shares_mutex); | |
6430 | if (tg->shares == shares) | |
6431 | goto done; | |
6432 | ||
6433 | tg->shares = shares; | |
6434 | for_each_possible_cpu(i) { | |
6435 | struct rq *rq = cpu_rq(i); | |
6436 | struct sched_entity *se; | |
6437 | ||
6438 | se = tg->se[i]; | |
6439 | /* Propagate contribution to hierarchy */ | |
6440 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
6441 | |
6442 | /* Possible calls to update_curr() need rq clock */ | |
6443 | update_rq_clock(rq); | |
17bc14b7 | 6444 | for_each_sched_entity(se) |
029632fb PZ |
6445 | update_cfs_shares(group_cfs_rq(se)); |
6446 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6447 | } | |
6448 | ||
6449 | done: | |
6450 | mutex_unlock(&shares_mutex); | |
6451 | return 0; | |
6452 | } | |
6453 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6454 | ||
6455 | void free_fair_sched_group(struct task_group *tg) { } | |
6456 | ||
6457 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6458 | { | |
6459 | return 1; | |
6460 | } | |
6461 | ||
6462 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
6463 | ||
6464 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
6465 | ||
810b3817 | 6466 | |
6d686f45 | 6467 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
6468 | { |
6469 | struct sched_entity *se = &task->se; | |
0d721cea PW |
6470 | unsigned int rr_interval = 0; |
6471 | ||
6472 | /* | |
6473 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
6474 | * idle runqueue: | |
6475 | */ | |
0d721cea | 6476 | if (rq->cfs.load.weight) |
a59f4e07 | 6477 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
6478 | |
6479 | return rr_interval; | |
6480 | } | |
6481 | ||
bf0f6f24 IM |
6482 | /* |
6483 | * All the scheduling class methods: | |
6484 | */ | |
029632fb | 6485 | const struct sched_class fair_sched_class = { |
5522d5d5 | 6486 | .next = &idle_sched_class, |
bf0f6f24 IM |
6487 | .enqueue_task = enqueue_task_fair, |
6488 | .dequeue_task = dequeue_task_fair, | |
6489 | .yield_task = yield_task_fair, | |
d95f4122 | 6490 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 6491 | |
2e09bf55 | 6492 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
6493 | |
6494 | .pick_next_task = pick_next_task_fair, | |
6495 | .put_prev_task = put_prev_task_fair, | |
6496 | ||
681f3e68 | 6497 | #ifdef CONFIG_SMP |
4ce72a2c | 6498 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 6499 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 6500 | |
0bcdcf28 CE |
6501 | .rq_online = rq_online_fair, |
6502 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
6503 | |
6504 | .task_waking = task_waking_fair, | |
681f3e68 | 6505 | #endif |
bf0f6f24 | 6506 | |
83b699ed | 6507 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 6508 | .task_tick = task_tick_fair, |
cd29fe6f | 6509 | .task_fork = task_fork_fair, |
cb469845 SR |
6510 | |
6511 | .prio_changed = prio_changed_fair, | |
da7a735e | 6512 | .switched_from = switched_from_fair, |
cb469845 | 6513 | .switched_to = switched_to_fair, |
810b3817 | 6514 | |
0d721cea PW |
6515 | .get_rr_interval = get_rr_interval_fair, |
6516 | ||
810b3817 | 6517 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6518 | .task_move_group = task_move_group_fair, |
810b3817 | 6519 | #endif |
bf0f6f24 IM |
6520 | }; |
6521 | ||
6522 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 6523 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 6524 | { |
bf0f6f24 IM |
6525 | struct cfs_rq *cfs_rq; |
6526 | ||
5973e5b9 | 6527 | rcu_read_lock(); |
c3b64f1e | 6528 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 6529 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 6530 | rcu_read_unlock(); |
bf0f6f24 IM |
6531 | } |
6532 | #endif | |
029632fb PZ |
6533 | |
6534 | __init void init_sched_fair_class(void) | |
6535 | { | |
6536 | #ifdef CONFIG_SMP | |
6537 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
6538 | ||
3451d024 | 6539 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 6540 | nohz.next_balance = jiffies; |
029632fb | 6541 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 6542 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
6543 | #endif |
6544 | #endif /* SMP */ | |
6545 | ||
6546 | } |